id
stringlengths 33
33
| doc_url
stringlengths 33
33
| title
stringlengths 13
211
| publication_date
timestamp[s] | update_date
timestamp[s] | authors
stringlengths 6
16.6k
| category_primary
stringclasses 60
values | category_all
stringlengths 5
132
| abstract
stringlengths 123
1.99k
| text
stringlengths 190
2.08k
| embeddings
sequencelengths 768
768
|
|---|---|---|---|---|---|---|---|---|---|---|
http://arxiv.org/abs/2401.14688v3 | http://arxiv.org/pdf/2401.14688v3 | Taiyi-Diffusion-XL: Advancing Bilingual Text-to-Image Generation with Large Vision-Language Model Support | 2024-01-26T07:17:50 | 2024-06-18T03:17:18 | Xiaojun Wu, Dixiang Zhang, Ruyi Gan, Junyu Lu, Ziwei Wu, Renliang Sun, Jiaxing Zhang, Pingjian Zhang, Yan Song | cs.CL | cs.CL | Recent advancements in text-to-image models have significantly enhanced image
generation capabilities, yet a notable gap of open-source models persists in
bilingual or Chinese language support. To address this need, we present
Taiyi-Diffusion-XL, a new Chinese and English bilingual text-to-image model
which is developed by extending the capabilities of CLIP and
Stable-Diffusion-XL through a process of bilingual continuous pre-training.
This approach includes the efficient expansion of vocabulary by integrating the
most frequently used Chinese characters into CLIP's tokenizer and embedding
layers, coupled with an absolute position encoding expansion. Additionally, we
enrich text prompts by large vision-language model, leading to better images
captions and possess higher visual quality. These enhancements are subsequently
applied to downstream text-to-image models. Our empirical results indicate that
the developed CLIP model excels in bilingual image-text retrieval.Furthermore,
the bilingual image generation capabilities of Taiyi-Diffusion-XL surpass
previous models. This research leads to the development and open-sourcing of
the Taiyi-Diffusion-XL model, representing a notable advancement in the field
of image generation, particularly for Chinese language applications. This
contribution is a step forward in addressing the need for more diverse language
support in multimodal research. The model and demonstration are made publicly
available at
\href{https://huggingface.co/IDEA-CCNL/Taiyi-Stable-Diffusion-XL-3.5B/},
fostering further research and collaboration in this domain. | Taiyi-Diffusion-XL: Advancing Bilingual Text-to-Image Generation with Large Vision-Language Model Support: Recent advancements in text-to-image models have significantly enhanced image
generation capabilities, yet a notable gap of open-source models persists in
bilingual or Chinese language support. To address this need, we present
Taiyi-Diffusion-XL, a new Chinese and English bilingual text-to-image model
which is developed by extending the capabilities of CLIP and
Stable-Diffusion-XL through a process of bilingual continuous pre-training.
This approach includes the efficient expansion of vocabulary by integrating the
most frequently used Chinese characters into CLIP's tokenizer and embedding
layers, coupled with an absolute position encoding expansion. Additionally, we
enrich text prompts by large vision-language model, leading to better images
captions and possess higher visual quality. These enhancements are subsequently
applied to downstream text-to-image models. Our empirical results indicate that
the developed CLIP model excels in bilingual image-text retrieval.Furthermore,
the bilingual image generation capabilities of Taiyi-Diffusion-XL surpass
previous models. This research leads to the development and open-sourcing of
the Taiyi-Diffusion-XL model, representing a notable advancement in the field
of image generation, particularly for Chinese language applications. This
contribution is a step forward in addressing the need for more diverse language
support in multimodal research. The model and demonstration are made publicly
available at
\href{https://huggingface.co/IDEA-CCNL/Taiyi-Stable-Diffusion-XL-3.5B/},
fostering further research and collaboration in this domain. | [
0.0207394678,
0.0319165885,
0.0175019354,
0.047003068,
-0.0006742988,
-0.0043660449,
0.0267088469,
0.0216063708,
-0.0044651278,
-0.0032915676,
-0.0457447618,
-0.070498839,
0.011680766,
0.0494774245,
0.0211271271,
-0.0950726792,
-0.0112169683,
-0.0304486956,
-0.0690101162,
-0.0167729687,
0.0316740274,
-0.0018059944,
0.0484655052,
0.0308279451,
-0.0232641418,
-0.0085770544,
0.0043051685,
0.041063413,
0.0320700742,
-0.0247770082,
-0.0264720432,
0.0440298691,
0.0402739756,
0.0223551262,
0.0000021939,
-0.0003440599,
-0.0056341626,
0.0280155931,
0.0643933788,
-0.0129299341,
0.081268914,
-0.0436248071,
-0.0344483554,
-0.0157257468,
-0.0198097154,
0.027061319,
0.0783446729,
0.0664378703,
0.070418112,
0.0280895866,
-0.0268728659,
-0.0378379822,
-0.0201484393,
-0.0367511995,
0.0274217725,
-0.0022916687,
-0.0045014597,
0.0209600702,
-0.0049153329,
0.042503085,
-0.0216150638,
0.0488025658,
0.0302381199,
-0.0229004957,
-0.0214464013,
0.0481287315,
-0.0075450069,
-0.0304119177,
-0.018017346,
0.070950143,
-0.0523506179,
-0.002170464,
-0.0335725509,
0.0145381205,
-0.0021254346,
-0.0050412882,
-0.0526526943,
0.0262585413,
0.0141413603,
0.0151492329,
-0.0412530862,
0.0044667004,
0.019428039,
-0.0185693204,
-0.0325810872,
-0.0209716521,
0.037595544,
0.0004979681,
-0.0264671315,
0.0197312962,
0.0062878523,
-0.025495274,
0.0794978216,
0.0202610344,
0.006967817,
0.0089693889,
-0.0438944958,
-0.0476922579,
0.0010373428,
-0.0476788394,
0.0150716268,
0.0443008132,
-0.0603786595,
0.0517201945,
-0.0737034306,
0.0363485329,
-0.0186460018,
0.0153350122,
-0.0728143752,
-0.029974727,
0.0103914877,
-0.0293317325,
-0.0518458001,
-0.0009117075,
-0.0621653013,
-0.0570487529,
-0.0535272062,
0.0113079837,
-0.0089634201,
-0.011164153,
0.0273178034,
0.0180455279,
-0.0591234826,
-0.0033250318,
-0.0225478783,
0.0219397563,
0.0030278792,
-0.0161719192,
0.0015974808,
-0.0259601045,
-0.0048593064,
0.0509118661,
0.0279163066,
-0.0302973539,
-0.0149820512,
0.0852139667,
-0.0653673634,
-0.0220454987,
0.001005473,
0.0044343756,
-0.0565851182,
-0.0275561642,
0.0021520127,
-0.0759508163,
0.0138487983,
-0.0257843398,
0.0223208591,
-0.0279337913,
0.0267158486,
0.0303228907,
0.001687183,
0.0621631555,
-0.0024001598,
-0.0040054494,
0.0438562036,
0.0137974257,
-0.0010133225,
0.0671810955,
0.0163004454,
0.0525756888,
0.0491586551,
0.0186758526,
-0.0130307032,
0.0666045696,
0.0212837867,
0.0037397889,
-0.0459013954,
-0.0274866149,
0.0320246331,
-0.009197725,
0.0472873412,
0.0686694011,
-0.0409254879,
0.0044283806,
0.0780307204,
0.0303941686,
-0.0193409063,
0.0136717074,
0.0644279718,
0.0167358015,
-0.0255024228,
0.0057520624,
-0.0042835879,
0.0502733365,
0.0220557954,
0.0235483795,
0.0367762633,
0.0684030354,
-0.0600218773,
-0.0297922678,
0.0245493464,
-0.0163052082,
0.0503846034,
-0.0086575188,
0.0568938032,
0.011380028,
-0.0335089378,
0.054542236,
-0.0105116656,
-0.0572000444,
0.0308597367,
-0.0274805389,
-0.024570426,
0.0768713132,
0.0281844083,
0.0272043161,
0.0177006237,
0.0240071677,
-0.0006241702,
0.0798689127,
0.0067507513,
0.0325436108,
-0.0029047895,
-0.0673228949,
-0.0348840132,
0.0130844088,
-0.0284658987,
0.0500325114,
0.0047056815,
0.0922402665,
0.0484374091,
0.0011232774,
-0.0515039712,
0.0278678536,
-0.0202991944,
-0.00555014,
-0.0054086545,
0.0268557183,
-0.0279050171,
0.0087181814,
-0.0492451005,
0.0168583672,
0.0039412035,
-0.0419707,
-0.0153968791,
0.0582171828,
-0.0339751281,
-0.0362782739,
0.0217434578,
-0.0661645755,
0.0152119398,
-0.0377647541,
-0.020240875,
-0.0551864356,
-0.0111575844,
-0.0015280001,
-0.008867397,
-0.0763765499,
0.0142399641,
0.0106227919,
0.0288250409,
-0.0091481609,
0.0540883876,
-0.0190830622,
0.0034204389,
0.0292623434,
0.0151497209,
0.0037483836,
-0.0372591689,
-0.0478279293,
-0.0044248397,
0.0117738526,
0.0233313348,
0.0307228472,
0.0348826908,
-0.0246719792,
-0.0240410883,
-0.0670038536,
-0.0104012182,
-0.0310386345,
-0.0260356497,
-0.0227261335,
0.0055968883,
0.0087097287,
-0.0319766141,
-0.0683081672,
-0.0079093091,
0.0186349824,
-0.0036627457,
0.0195928756,
-0.0094720349,
0.0480754599,
-0.0213416647,
-0.0044977004,
0.0581516065,
0.0183389485,
0.0008230646,
-0.0707678869,
-0.0532651469,
-0.0006217888,
-0.0171137508,
0.0006011878,
-0.0046235723,
-0.03077529,
0.0098792827,
0.0009238624,
0.0760907009,
0.0630366281,
0.0327750146,
0.0067640743,
-0.0060643246,
-0.0120879216,
-0.0238993913,
0.0179668255,
-0.0082484232,
-0.0269710589,
0.0235776696,
-0.0363829844,
-0.0117127243,
0.0213806313,
-0.0135386074,
0.0082835322,
-0.0805581808,
-0.0380011722,
-0.0479555726,
-0.0543880798,
-0.0658127591,
0.0450320989,
-0.0333103687,
-0.0159842633,
-0.0498435609,
-0.0450701304,
0.0375493318,
-0.0399488918,
-0.0087699033,
-0.021716889,
0.0029054165,
-0.0246231649,
0.0509002507,
0.0154566271,
0.0570372716,
-0.0146164317,
0.011972324,
-0.0493416265,
0.0365830064,
-0.0103842104,
-0.0329137892,
0.0187105425,
-0.006737968,
-0.0028596949,
-0.0435061529,
-0.0246405434,
-0.0309505854,
-0.0147106592,
0.0049412446,
0.0032120007,
0.0273377784,
-0.0119090052,
0.0087072132,
0.0488865487,
0.0374242142,
0.0251399446,
-0.0119407168,
0.012974509,
-0.0063651009,
0.0328466222,
0.0189872812,
-0.0023968427,
-0.0447541997,
-0.0576803014,
0.007291018,
0.0555451736,
0.0045727463,
-0.02239432,
-0.0051718499,
0.0107059758,
0.0192169193,
0.0684553981,
-0.0343523808,
-0.0035131932,
0.0974783674,
-0.0323999412,
-0.0067593749,
0.0477493517,
0.022158321,
-0.0323923826,
-0.0488333032,
0.001272043,
0.0578129068,
0.0020431597,
-0.0004835159,
-0.1232561544,
-0.0168443006,
-0.0222118571,
-0.016607454,
0.0283142272,
0.0343035497,
-0.0283781588,
-0.1081110984,
0.0671325251,
0.0318568535,
-0.0227156151,
-0.0205127317,
0.0133817112,
-0.000457466,
0.0388313867,
-0.0184864625,
-0.0140686342,
0.0229996387,
-0.0009251955,
0.0042979321,
0.0553213581,
-0.0471155271,
-0.0458456352,
-0.0336507894,
-0.0540161245,
0.0621186234,
0.0406132042,
0.0128320307,
-0.0661218613,
0.0122404909,
-0.001988197,
0.0193108656,
-0.0001512011,
0.0555114299,
0.0210291483,
-0.028939357,
-0.0145818712,
0.0445359163,
0.0209114514,
-0.0015544565,
0.0047231219,
0.0052682897,
0.0099018868,
-0.0044719689,
-0.0396589935,
0.0243606772,
0.0080532683,
0.0465605557,
0.0197761487,
-0.0852190778,
-0.0277655385,
-0.0305944905,
0.0022861355,
-0.0953142494,
-0.036695227,
0.0206594057,
-0.0304096974,
-0.030138772,
-0.0851553306,
-0.0090819784,
-0.0110258814,
0.0555296354,
-0.0091322791,
0.0252802111,
-0.0446997695,
0.0267874692,
-0.0323328562,
0.0178662762,
-0.0144739775,
0.0294626784,
-0.0520148911,
0.0091013806,
-0.0044444678,
-0.0663182288,
0.0081236875,
0.0275012944,
-0.0694877356,
-0.029850211,
-0.0461818278,
0.0633808225,
0.0050339135,
0.0104812738,
-0.0135219684,
-0.0228144191,
0.0657329783,
0.0540327616,
0.0543253385,
-0.0118373502,
-0.0050256653,
-0.014583094,
-0.0292460229,
0.0177841056,
-0.0355622843,
0.0071505853,
-0.0091841854,
0.0285427421,
0.0034637169,
-0.0002266795,
0.0180255007,
-0.0307545662,
-0.0236609858,
-0.0453772694,
0.0683886483,
-0.0466653854,
0.01337111,
-0.0310248453,
-0.035775844,
-0.0767251551,
0.0065857787,
0.0324089266,
-0.0031251293,
0.0201391876,
-0.0143012628,
0.0483135916,
0.022577133,
-0.0095105693,
-0.0044582654,
-0.0031826717,
0.0253286716,
0.018596122,
0.0247706436,
-0.0153190177,
0.0323999785,
-0.0422711,
0.0330767669,
-0.0216589514,
0.064044416,
0.014147548,
-0.027287107,
-0.0058160005,
0.0426120721,
-0.0014642597,
-0.0349737518,
-0.0426150858,
0.031287048,
0.078032732,
0.0073408396,
0.0699015558,
-0.0348535739,
-0.0279059447,
-0.0059859171,
0.0101103727,
0.0301401913,
-0.0463679507,
0.0171514265,
-0.0185595583,
-0.0019726944,
-0.0415076092,
0.0882133916,
-0.0163334049,
-0.0287046228,
-0.006782386,
-0.0149804754,
-0.0027657591,
0.0073947045,
0.0266570859,
-0.0127900308,
-0.015549385,
0.0108059309,
-0.0158648323,
-0.060504742,
-0.0045329626,
0.0238226354,
-0.042316623,
0.047393728,
0.042590335,
0.0478453897,
0.0101270126,
0.0117944889,
-0.0119728083,
0.0164102409,
0.0038927358,
-0.0085197929,
0.0338754356,
-6.214000358e-33,
-0.0400745794,
-0.0212921016,
0.0076699746,
-0.0324700251,
-0.0351002291,
-0.0093456497,
-0.0372869484,
0.0136556476,
-0.007331158,
-0.009390559,
0.0106254127,
-0.012769633,
0.0317318253,
0.0714899525,
-0.0427851826,
0.0620612018,
0.0330924354,
-0.0009631758,
0.0154197486,
0.0205171425,
0.054226242,
0.0244670194,
0.0289936159,
-0.0264523309,
-0.002334459,
0.0004300577,
0.0001993383,
-0.0029063651,
0.028582992,
0.0470880866,
-0.0063498169,
0.0330022424,
0.0055657458,
-0.0199367944,
-0.0258864034,
0.0260667559,
-0.0617589504,
-0.0027722509,
-0.0335894711,
0.0315715075,
-0.0749612227,
-0.0264509432,
0.026373677,
-0.0594396926,
0.0178847369,
-0.0063592414,
0.0423205048,
-0.0456217118,
-0.0090750139,
-0.0074745812,
-0.0435357653,
0.0414163955,
0.0056590666,
-0.0025063562,
0.0332595408,
0.0692474321,
0.0288266446,
0.0064467299,
-0.0839454979,
0.0099007338,
0.0088060554,
0.0233931579,
0.0606838316,
0.0628736392,
0.0160840806,
0.0402904563,
0.0113358805,
-0.001386657,
0.0078771021,
-0.0262596197,
0.0267799385,
0.044193726,
0.0352294371,
0.0523994416,
0.0085466923,
-0.0340141393,
-0.0326173715,
0.0531358533,
-0.0206555799,
-0.0175241716,
0.0631746948,
0.0120570194,
0.0149366818,
0.0067417575,
-0.028470384,
-0.0361990295,
-0.0455312654,
-0.060726691,
0.0377983302,
0.0395887457,
-0.0234641097,
-0.004981535,
0.0823790729,
0.0017594777,
-0.0241612177,
0.0003476979,
-0.0393869542,
0.0300115086,
-0.0185244959,
0.023617316,
-0.034827929,
-0.063532047,
0.0165994186,
0.0077940896,
-0.0343201049,
0.0205506999,
0.0297086127,
-0.0036634889,
-0.0386244431,
0.015121121,
-0.0224019345,
-0.0098114247,
0.0324800871,
0.0550435223,
0.0124250166,
-0.0176906828,
-0.0527372435,
0.0263948869,
0.017476555,
0.0005873641,
-0.0372943282,
-0.0629799813,
0.0047957525,
0.0164280608,
-0.0281821117,
0.0545060858,
-0.0581953302,
0.0376826413,
0.0263935942,
-0.0675503835,
-0.0391129479,
0.0229429584,
2.753e-7,
0.0542575344,
-0.0227875337,
-0.0031556466,
0.0095247552,
0.0125309993,
0.0093502952,
-0.0442939065,
0.0272899475,
0.0268900096,
0.0078848731,
0.0034266964,
0.0467550606,
0.0622014776,
0.0055233482,
-0.0410183147,
0.0107595986,
-0.0197751876,
-0.0067745163,
-0.0216095354,
0.0168742836,
0.0525820069,
0.0920827687,
-0.0495110154,
-0.0074470495,
0.029707877,
-0.0538742132,
0.1211713254,
0.0048032436,
0.0092569096,
-0.0047544916,
-0.0139917685,
0.0172635727,
-0.0634220392,
-0.0250707902,
-0.0215303134,
0.0085748779,
0.0920400769,
0.0121094137,
-0.0199017078,
0.0031257251,
0.0152235674,
0.0233636945,
-0.0209097024,
-0.0892839283,
0.071874164,
-0.0281491876,
0.0100656087,
-0.0262634326,
-0.0082755396,
0.0093157748,
0.0279007982,
-0.0447779484,
-0.0660868436,
0.040722087,
0.0518987589,
0.0269814394,
-0.0376654044,
-0.0578378737,
-0.0623074807,
-0.0405519307,
-0.0322364159,
-0.0293542221,
-0.0957757086,
-0.0187444482,
-0.0124290669,
0.030649405,
-0.005904974,
2.316242512e-34,
0.0064194719,
-0.0199687593,
0.0039199907,
-0.0127980001,
0.0143022528,
-0.0205853675,
-0.0012897101,
0.0030927649,
-0.014813358,
-0.0387389846,
-0.0519820973
] |
http://arxiv.org/abs/2401.14681v2 | http://arxiv.org/pdf/2401.14681v2 | MasonTigers@LT-EDI-2024: An Ensemble Approach Towards Detecting Homophobia and Transphobia in Social Media Comments | 2024-01-26T06:56:17 | 2024-02-15T06:05:13 | Dhiman Goswami, Sadiya Sayara Chowdhury Puspo, Md Nishat Raihan, Al Nahian Bin Emran | cs.CL | cs.CL | In this paper, we describe our approaches and results for Task 2 of the
LT-EDI 2024 Workshop, aimed at detecting homophobia and/or transphobia across
ten languages. Our methodologies include monolingual transformers and ensemble
methods, capitalizing on the strengths of each to enhance the performance of
the models. The ensemble models worked well, placing our team, MasonTigers, in
the top five for eight of the ten languages, as measured by the macro F1 score.
Our work emphasizes the efficacy of ensemble methods in multilingual scenarios,
addressing the complexities of language-specific tasks. | MasonTigers@LT-EDI-2024: An Ensemble Approach Towards Detecting Homophobia and Transphobia in Social Media Comments: In this paper, we describe our approaches and results for Task 2 of the
LT-EDI 2024 Workshop, aimed at detecting homophobia and/or transphobia across
ten languages. Our methodologies include monolingual transformers and ensemble
methods, capitalizing on the strengths of each to enhance the performance of
the models. The ensemble models worked well, placing our team, MasonTigers, in
the top five for eight of the ten languages, as measured by the macro F1 score.
Our work emphasizes the efficacy of ensemble methods in multilingual scenarios,
addressing the complexities of language-specific tasks. | [
0.0677911565,
0.0654575452,
-0.0210932698,
0.0146943321,
0.0071906946,
0.0500793196,
0.0345072635,
0.0231266953,
0.0464746058,
-0.0595294274,
-0.0063238489,
-0.0762928203,
-0.0093965577,
-0.0048340061,
-0.0236124061,
-0.0230485331,
0.0016993406,
0.0432303697,
-0.033093486,
-0.0432739593,
0.0485415794,
0.0308825113,
0.01687406,
0.0555356964,
0.0256767757,
0.0054122815,
0.025389351,
0.0002154008,
0.0339408405,
0.0241373088,
0.0384116024,
0.0571576618,
-0.0216381475,
0.0624748245,
0.0000019244,
-0.0482284725,
-0.0263061412,
0.0235092845,
0.0504865721,
-0.0456662774,
-0.0159410425,
-0.026136646,
-0.0047564977,
0.0386802852,
-0.0265832562,
-0.0117927175,
0.0777471066,
0.0087317796,
0.0570914261,
0.0416886359,
-0.0148316538,
-0.0035902029,
-0.0525206216,
0.0034657356,
-0.0489784218,
0.0203334466,
0.0444533117,
-0.0741425604,
-0.0305667054,
-0.0175489113,
-0.0062621655,
0.0076764119,
-0.0091716601,
0.0279514324,
0.0260722525,
-0.0265961234,
-0.0138117252,
-0.0690748245,
0.0295990836,
0.0312772132,
-0.0096530383,
0.0142607428,
0.0138852783,
0.0232525002,
-0.0170949157,
0.0166868623,
-0.0097935414,
0.0302685127,
-0.0571657456,
0.010976702,
-0.0381006151,
0.0048926733,
0.0447748341,
0.0908282995,
-0.0092615206,
0.0613105558,
0.0102153961,
-0.013701464,
0.0371818691,
0.0080798855,
-0.0122934356,
-0.0015228519,
0.0735305473,
0.0183201246,
0.0028018602,
0.0217167009,
0.0191926919,
-0.0615626499,
0.0223737303,
-0.0306673832,
0.0004529133,
0.0460644923,
0.0008021648,
-0.0017599774,
-0.0246794149,
0.0467815362,
-0.0027601516,
-0.040859472,
0.0151371323,
-0.026572736,
-0.0350092649,
0.0454663336,
-0.0071428139,
0.001601738,
-0.0140711227,
0.059931308,
-0.0686496198,
0.0257269666,
-0.0447220989,
-0.0219774116,
0.0216439608,
-0.0138147064,
-0.0010478364,
0.0157494787,
0.0320602395,
0.0242368951,
-0.0236829817,
-0.0140545405,
-0.0593735427,
-0.0514198169,
0.0271201842,
0.0453339033,
0.0008822298,
-0.0432926752,
0.0047064535,
0.0147685641,
0.0291414652,
-0.033109732,
0.0718579963,
-0.0073778867,
-0.065083012,
-0.0576251298,
0.0008762605,
-0.0683188885,
0.0386282317,
0.0221350361,
-0.0007219346,
-0.0506557152,
-0.0083250608,
-0.0041684811,
-0.0255841874,
0.0859718025,
0.0242379364,
-0.0106148664,
0.044979386,
0.0395367146,
-0.0161997043,
0.0833870173,
-0.0149501944,
0.0802306533,
0.0333934948,
0.0172070321,
0.0079811784,
0.0436272733,
-0.0005603284,
0.0490774512,
-0.0113168173,
0.0281100329,
-0.0028187521,
-0.0407537185,
-0.0459973142,
0.0496173613,
0.0119242705,
-0.0100233536,
0.0259233024,
0.0585140809,
0.0000390221,
0.0252282619,
0.0340824574,
0.0045243544,
-0.0202789791,
-0.0292039812,
-0.0565930977,
-0.0325772874,
-0.0446068048,
-0.0424290188,
0.0324997716,
0.083356455,
-0.0127892168,
-0.0282192752,
-0.0312328041,
-0.007493576,
0.049580764,
-0.0373927206,
0.0379181653,
-0.0252451822,
-0.0297053084,
0.009985337,
0.0299099442,
-0.0544575788,
0.0259756092,
-0.0231046937,
0.0029555301,
0.055990018,
-0.0042424183,
-0.0425780155,
0.0331291445,
-0.0500457287,
-0.0119997207,
0.0303009171,
0.0040968708,
0.0141630126,
0.030304756,
-0.0295205805,
-0.0462916829,
0.0842720047,
0.0108307684,
0.0067436504,
0.0662774667,
0.0649490952,
-0.0023953097,
0.0342611037,
-0.0112724546,
0.0191078931,
0.0214490872,
0.0146573642,
0.0082982145,
0.0167599265,
0.0422204845,
0.0131226275,
-0.0236406978,
-0.0603527464,
0.00002567,
-0.0153036881,
-0.0920170844,
0.0277030841,
-0.0278614108,
-0.0285953879,
0.0336336643,
-0.0686435923,
-0.0480847098,
0.0608135499,
-0.0109739341,
0.0320889466,
0.0824977681,
-0.0325931683,
0.0059543825,
-0.0242386609,
-0.0196433533,
-0.0621993951,
0.0291891508,
-0.0229678135,
0.0491776988,
-0.0473122336,
0.0094704768,
-0.0383370258,
-0.0876026079,
0.0363783278,
0.0176738687,
-0.0512042493,
-0.0017492399,
-0.0070690983,
0.0256188549,
0.0178710651,
-0.01485004,
0.0126284165,
-0.0380955115,
-0.1034784615,
-0.0229084678,
-0.0218287762,
-0.0087565845,
-0.0129125779,
0.0149133904,
-0.0075859223,
0.0284365825,
-0.0096331956,
-0.0124593657,
0.044417873,
0.0437526889,
-0.0134770498,
-0.0654534101,
-0.0198093317,
-0.0164459255,
-0.010154983,
-0.0234870445,
0.0531844757,
0.0168224405,
-0.0817975253,
-0.0200974792,
-0.0347209722,
-0.0687995926,
0.0063438308,
0.0254345294,
-0.0520062111,
0.0469354875,
0.0089843739,
0.0156247737,
0.0103040813,
0.0559066609,
0.0584175661,
0.0298644733,
0.0284559783,
-0.0198340509,
-0.0050055808,
-0.0171307139,
0.0432017185,
-0.0117414007,
-0.0749862343,
-0.0031732854,
0.0249196477,
0.0314796455,
-0.0180382933,
-0.0924238712,
0.0612745434,
-0.0706525743,
-0.0573956259,
-0.0131677091,
-0.0693267956,
0.0248461552,
0.0057582972,
-0.0071890135,
-0.0314009599,
0.0501066446,
0.0176196005,
-0.0022026231,
-0.0527299047,
-0.0268315934,
-0.0094387578,
0.0448032878,
-0.004933483,
0.0673197731,
-0.0147821195,
0.0207019467,
0.0064394888,
-0.040307086,
0.023812985,
-0.029780928,
0.0584524944,
0.0013770333,
-0.0007006694,
0.0143730659,
-0.0606631301,
-0.0565747097,
-0.0406086892,
-0.0568765588,
-0.0330639109,
0.0102922935,
-0.042457059,
-0.0361446366,
0.0488217063,
-0.0219699461,
-0.0360024497,
-0.0149212005,
-0.0173406154,
0.0490612686,
-0.0090336548,
-0.0313814245,
-0.0429706052,
0.0335310772,
-0.0164573137,
-0.012514187,
0.0816436559,
0.0124311754,
0.0076913233,
-0.0422554649,
-0.0150407832,
0.0172591917,
0.0016940248,
0.0179587919,
-0.0878837779,
0.0432004854,
0.0102919089,
0.0290662199,
0.0632415116,
0.0276560802,
-0.0218822639,
0.0027184892,
0.0099800248,
0.0441695377,
-0.0042800219,
-0.0118362429,
-0.0579079948,
-0.0020408677,
-0.019954307,
0.0259657074,
-0.0376927778,
-0.0616157018,
0.0068269661,
-0.0301416125,
0.01450988,
-0.0058281696,
-0.0609389916,
-0.0445184559,
-0.0364251286,
0.0457057841,
-0.0364739001,
-0.0093283709,
-0.0401003622,
0.0086324364,
0.0308964215,
0.0017943352,
-0.0677518174,
-0.0103683108,
0.0092226742,
-0.0645581931,
0.0077455668,
0.0212590788,
0.0693716481,
-0.0193990134,
0.0109116128,
-0.0067082699,
-0.0199276265,
-0.0102185197,
-0.0678474084,
0.0330346599,
0.0103232069,
-0.0336900279,
-0.028950531,
-0.0018584626,
-0.0125401346,
0.01898803,
0.0139690693,
0.0436405651,
-0.0069316435,
-0.0331549533,
-0.0183278844,
0.0409360155,
0.0094226617,
0.08281295,
0.032055825,
-0.0541853048,
0.0113903591,
-0.0818211511,
-0.0079852547,
-0.0534729287,
-0.0419928282,
-0.0256965794,
0.0273918472,
-0.0588984713,
-0.0715547502,
-0.0427675545,
0.0300547611,
-0.0141016468,
0.0020305736,
-0.052114632,
-0.0489025638,
0.0224846918,
-0.0624586679,
0.0102440519,
0.0315990373,
0.0333606265,
0.0228015743,
0.0306085031,
-0.0222154707,
-0.0037413186,
-0.0160674639,
0.0291679576,
-0.0346030034,
-0.0556270778,
0.06410712,
0.0514733791,
0.0215202495,
0.0032947983,
-0.0214648414,
-0.0641909838,
0.0321534909,
0.0609388575,
0.0120867556,
-0.0019052775,
0.0627406612,
-0.030154774,
-0.0133563513,
0.0533785708,
0.0015861621,
0.0459947176,
0.0288740676,
-0.0359204859,
-0.0373066813,
-0.0605415106,
-0.0238393191,
-0.022566583,
-0.0129481126,
-0.0300621372,
-0.0446932204,
-0.0245733652,
-0.032520093,
0.0306677539,
-0.0245725401,
-0.0067526293,
-0.0267883521,
0.0095795039,
-0.0135700153,
-0.0365754776,
0.0077526942,
-0.0177263692,
0.0055622035,
-0.0393246002,
0.0581651814,
0.0464657955,
0.0295752678,
0.0284691658,
0.0141195515,
-0.028431192,
0.0206844416,
-0.0018198842,
0.0041262461,
0.0059074457,
0.0064825206,
0.016763581,
0.0112055345,
0.0337282792,
-0.0036895943,
-0.0104883155,
-0.028105624,
-0.0316989608,
-0.0651483387,
0.0068607009,
0.0096744774,
0.0591314025,
0.0035966423,
-0.0250033792,
-0.0214114189,
0.0226597507,
-0.0251527671,
0.0098772813,
0.0191582814,
0.0069569703,
-0.0083137853,
0.0211319588,
0.0327068493,
-0.0323805138,
0.0218146276,
0.0662363097,
-0.0368009657,
-0.0390614271,
-0.0007376226,
-0.0361647345,
-0.0009369972,
-0.0001024576,
0.0229355656,
-0.0475773886,
-0.0412602797,
0.0118104909,
-0.0826931074,
0.0219194032,
0.0130393663,
-0.0095160808,
0.0455125645,
0.0206548162,
0.0160926301,
-0.0341172516,
0.006089468,
0.014281108,
-0.0299275108,
-0.0321690813,
-5.91063796e-33,
-0.0107270516,
-0.0413753539,
-0.0522616878,
-0.0191101227,
0.0048741582,
-0.0376975536,
0.0254332144,
-0.0072244499,
0.0022185252,
0.0054667289,
0.0021523032,
0.0280741807,
0.016608024,
0.0047538872,
0.0011878404,
0.0084084021,
0.0434463806,
0.0142437816,
-0.0257027522,
0.009443393,
0.0343734212,
0.0280772168,
0.0194830019,
-0.0786088556,
0.0493549332,
-0.0064002164,
-0.0518735759,
-0.0118358918,
0.0175389573,
0.04343253,
-0.0160674416,
0.0674146041,
0.0372417606,
-0.0040486669,
0.043980699,
0.063527368,
-0.0136278896,
0.0015286373,
0.0046503185,
0.035502445,
0.0529581234,
-0.0163262654,
0.0205139574,
-0.0303025767,
-0.0248479582,
-0.0534279868,
0.0493200049,
-0.0286738016,
0.0333159193,
-0.0240646861,
-0.0422879234,
0.0321365446,
-0.0203815028,
0.0394568816,
0.0088958824,
0.1092859656,
0.0343455598,
0.0125371478,
-0.078745693,
0.0320681669,
-0.021440465,
-0.0483530574,
0.0111398948,
-0.0045974148,
0.0019970774,
0.0390982069,
0.0317514837,
0.0080086971,
-0.0551963113,
-0.0360577367,
0.0259312633,
0.0149582149,
0.0623481795,
0.0367741138,
-0.0257813223,
-0.0108414097,
-0.0443582945,
0.0374918394,
0.0730554014,
-0.0262931194,
0.0191212725,
-0.0064498535,
-0.0409514308,
-0.0439512283,
0.0037999358,
-0.0555767752,
-0.0020651103,
-0.0488107912,
0.0032703052,
-0.0371749103,
-0.0744965672,
0.0285544302,
0.0079271104,
-0.0219890047,
-0.0481821634,
0.0299717057,
-0.0118093966,
-0.0198651366,
0.0378837138,
0.0400166698,
-0.0069378936,
0.0177189615,
0.0120537579,
0.0308284387,
0.0261116009,
-0.0027797646,
0.0384140089,
0.0318351388,
-0.0819991753,
0.0129744513,
-0.0043415632,
0.0471254662,
-0.0276216753,
0.1001168638,
-0.015659865,
0.0361696221,
-0.0066534868,
0.0019963745,
0.0420346633,
0.0172123648,
-0.0389503688,
0.007276962,
-0.0210922901,
0.0203485768,
0.0050498955,
-0.0514826402,
-0.0145536661,
0.0005328345,
0.0703514963,
-0.0635733232,
0.0110339588,
-0.0602031685,
2.586e-7,
0.0366569571,
0.0126861092,
-0.0123049961,
-0.0503903553,
0.0093395021,
0.0345632769,
0.0134613048,
0.0081459694,
0.0286957417,
0.0313341767,
-0.034907382,
0.0708844587,
0.0055757887,
0.0562036745,
-0.0053530368,
-0.0594200417,
-0.0345476791,
0.0084086265,
-0.0454343297,
0.0183748584,
-0.033293277,
0.0917912498,
0.0990352258,
-0.0263257567,
-0.0396100879,
-0.0011073678,
0.0237636771,
-0.0234151185,
0.0111203361,
0.0026683134,
0.0327501744,
0.0107499054,
0.0040357406,
0.0189070236,
-0.0129653169,
-0.0091810999,
0.0189232584,
0.0050418032,
-0.0236693397,
0.0558820292,
0.0299205352,
-0.023727892,
-0.0309724845,
-0.0723310113,
0.0234907176,
-0.0284935813,
0.0024545207,
-0.0147781121,
-0.0187948439,
0.0092443675,
0.04954613,
-0.0152885653,
-0.0114127276,
0.0421797521,
0.002262417,
-0.0158734582,
0.015954569,
-0.0676279142,
0.0032555915,
-0.0489655845,
-0.0270789824,
-0.0329330042,
-0.0044309641,
0.0091656223,
0.0078297909,
-0.0012921155,
0.0264368169,
2.01307632e-34,
0.0429645702,
0.0246720519,
-0.0372173712,
-0.0181771517,
0.0353562683,
-0.0106721418,
-0.0095484518,
0.0008587633,
0.037001472,
0.0053231246,
0.0142716076
] |
http://arxiv.org/abs/2401.14680v2 | http://arxiv.org/pdf/2401.14680v2 | MaLLaM -- Malaysia Large Language Model | 2024-01-26T06:56:05 | 2024-01-29T07:18:59 | Husein Zolkepli, Aisyah Razak, Kamarul Adha, Ariff Nazhan | cs.CL | cs.CL | Addressing the gap in Large Language Model pretrained from scratch with
Malaysian context, We trained models with 1.1 billion, 3 billion, and 5 billion
parameters on a substantial 349GB dataset, equivalent to 90 billion tokens
based on our pretrained Byte Pair Encoding (BPE) tokenizer for a single epoch.
MaLLaM contributes to enhanced natural language understanding and generation
tasks in the Malay language. Although trained on a smaller dataset of 90
billion tokens, our instruction-tuned MaLLaM models perform competitively. When
compared to ChatGPT3.5 and Malaysian Mistral, MaLLaM's instruction-tuned models
demonstrate notable proficiency, underscoring the effectiveness of our approach
in capturing and understanding the nuances of the Malaysian language. MaLLaM
models mark a significant contribution to the field, providing comprehensive
language representations grounded in Malaysian context. This endeavor aims to
pave the way for enhanced natural language understanding and generation tasks
specific to the linguistic nuances present in Malaysia. We discuss the training
methodology, dataset composition, and the potential impact of MaLLaM in
advancing the capabilities of large language models within the context of the
Malay language.
All models released at
https://huggingface.co/collections/mesolitica/mallam-6577b59d1e0b436ae75f930f | MaLLaM -- Malaysia Large Language Model: Addressing the gap in Large Language Model pretrained from scratch with
Malaysian context, We trained models with 1.1 billion, 3 billion, and 5 billion
parameters on a substantial 349GB dataset, equivalent to 90 billion tokens
based on our pretrained Byte Pair Encoding (BPE) tokenizer for a single epoch.
MaLLaM contributes to enhanced natural language understanding and generation
tasks in the Malay language. Although trained on a smaller dataset of 90
billion tokens, our instruction-tuned MaLLaM models perform competitively. When
compared to ChatGPT3.5 and Malaysian Mistral, MaLLaM's instruction-tuned models
demonstrate notable proficiency, underscoring the effectiveness of our approach
in capturing and understanding the nuances of the Malaysian language. MaLLaM
models mark a significant contribution to the field, providing comprehensive
language representations grounded in Malaysian context. This endeavor aims to
pave the way for enhanced natural language understanding and generation tasks
specific to the linguistic nuances present in Malaysia. We discuss the training
methodology, dataset composition, and the potential impact of MaLLaM in
advancing the capabilities of large language models within the context of the
Malay language.
All models released at
https://huggingface.co/collections/mesolitica/mallam-6577b59d1e0b436ae75f930f | [
0.0316543989,
0.068153739,
0.0038848915,
0.0076265451,
-0.0018364746,
0.0257106069,
0.0101713641,
-0.0176849123,
-0.0410059765,
-0.0565089025,
-0.0687746853,
-0.0870334357,
-0.0165334381,
0.0422636196,
0.0490928628,
-0.0810304061,
0.062120907,
-0.024797691,
0.0332624763,
0.0161308646,
0.0426821187,
0.0444873795,
0.0077955755,
0.0478588417,
-0.0348261967,
-0.0103437044,
-0.040559534,
0.0290947631,
-0.0108032087,
-0.0308196321,
-0.0039358297,
0.0359165631,
0.0292568747,
0.0305643808,
0.0000020145,
-0.0548569597,
-0.0195879079,
-0.0362362228,
-0.0297580715,
0.066288434,
0.0566414744,
-0.0003753308,
-0.0036938116,
0.0483864844,
-0.0219276361,
-0.0335241705,
0.0349897183,
0.0574180707,
0.0803199336,
0.0995937586,
-0.004029593,
-0.0329359211,
-0.0009294276,
-0.0572583191,
-0.0247495417,
-0.0126310503,
-0.0128902495,
-0.0099652028,
-0.0142562119,
0.0237063132,
-0.0211344026,
-0.017708892,
-0.011525278,
0.0114188725,
0.0179967973,
0.013356016,
-0.0350882374,
-0.0369035639,
0.0039525931,
0.0451779887,
-0.0298624337,
-0.0423463695,
-0.0171265323,
0.0168366358,
-0.0046082051,
-0.0366805308,
-0.0327101126,
0.0436279625,
-0.0033560088,
0.0611491799,
0.0474366397,
-0.0103575634,
0.0233778749,
0.0418124236,
-0.0417117849,
-0.0058836807,
0.0025948458,
-0.010735156,
-0.0342915952,
-0.0232569464,
-0.0327015519,
-0.0584616885,
0.0383266434,
0.0483329333,
-0.0419117548,
0.0125171784,
-0.0643100962,
-0.0062323511,
0.0365895703,
-0.0451037623,
0.0220571067,
0.0593206733,
-0.0173222255,
0.0043481514,
-0.0474962816,
0.0296318792,
-0.0157835092,
0.0259881075,
-0.0581304394,
-0.0330890752,
-0.0371468887,
-0.0322550647,
-0.0601730086,
0.0609824583,
-0.0735162273,
-0.0291731041,
-0.0365376435,
0.0187649019,
0.0268591885,
-0.0315571651,
0.0160669815,
0.0073815449,
-0.0769891813,
-0.0014351385,
-0.0042781746,
-0.0031941752,
-0.036878068,
0.017615702,
0.0435538441,
-0.0627262965,
-0.0015171156,
0.0339664519,
-0.002704333,
-0.0315878317,
0.0069219572,
0.0724386945,
-0.0291546676,
-0.0342110731,
-0.0461213067,
-0.0444371924,
0.0005436221,
-0.0485614985,
-0.0230213683,
-0.0390974097,
0.0439408273,
0.0100667765,
-0.0032385318,
0.0381660201,
0.0238358732,
0.0133435195,
0.0046795318,
0.0446046405,
-0.0093104262,
0.003420624,
0.0198309962,
0.0178197529,
0.0172613338,
0.0577974096,
0.0150730284,
0.0318702795,
0.0065456303,
-0.0387481712,
-0.0383932963,
0.0182568394,
0.0280452203,
-0.0080214478,
0.0193361882,
-0.0008935803,
-0.0055872328,
0.0440980271,
0.0588547625,
0.0722331554,
-0.0494248196,
0.0158545058,
0.0311220009,
0.0010800873,
0.0438256897,
0.0517185144,
0.0727536008,
0.0074921558,
0.0032083592,
-0.0137106432,
-0.0175201874,
0.0077365348,
0.0154432794,
0.03139624,
0.0476348139,
0.0456598625,
-0.0364529416,
-0.020883413,
0.0198692139,
-0.0195398498,
-0.0201532245,
-0.0209822319,
0.0415855348,
0.0264873244,
-0.0498603843,
0.090504922,
0.0274312906,
-0.0347303189,
0.0332735106,
-0.0529711619,
0.0004894426,
0.0482305512,
0.0168213416,
0.0248074085,
-0.0439697467,
-0.0272141639,
0.0135724284,
0.0475950837,
0.0092655411,
-0.0145920143,
0.0122925211,
-0.0683139935,
0.0327838808,
-0.0166696366,
0.009769802,
0.0311332159,
0.0280836187,
0.0185756795,
0.0438874327,
-0.0169775989,
0.0096559282,
0.0535907075,
-0.0045945207,
-0.0284649767,
0.00048721,
-0.0074685281,
0.0489810333,
0.0367266983,
-0.0097590955,
0.0189030524,
0.0252394564,
-0.0508512966,
0.0208511185,
0.0535558015,
-0.0862705708,
-0.0304218736,
0.0269026309,
-0.0005489668,
0.0274156295,
0.0279421266,
-0.0409417115,
-0.0692186579,
0.018562166,
-0.0137818567,
0.0568143651,
-0.0818693712,
0.0288652331,
-0.0071012187,
0.0301119164,
-0.0152003765,
0.0404983982,
-0.0143089276,
-0.0353464484,
0.0143958945,
0.0622176938,
0.0400639698,
-0.0391366109,
-0.0034536463,
-0.0163988974,
-0.0147526693,
0.0064105364,
-0.0147501826,
0.0228914917,
-0.0194815323,
-0.0383537076,
-0.1256996989,
0.0290250182,
-0.0291922446,
0.0228266884,
0.0406253785,
0.0121440357,
-0.0251542367,
-0.0171751697,
0.0027018762,
-0.0316000134,
0.0420368128,
-0.0084302379,
0.0220501982,
-0.0152381863,
-0.0052070264,
-0.0314632133,
-0.0105984444,
0.0099426191,
0.0268557873,
-0.0112713343,
-0.0483005978,
-0.052226834,
-0.0038066881,
-0.0114107803,
-0.0263318401,
0.0103475703,
-0.0644166246,
0.0200056266,
-0.0189387202,
0.0386338271,
0.026153056,
0.0379015096,
0.0291034281,
-0.0007457754,
-0.0074181273,
-0.0324254408,
-0.0004754469,
-0.0041198279,
0.0072372141,
-0.0105706379,
-0.0171931274,
0.0243838821,
0.0712976307,
-0.0015368034,
0.0031304334,
-0.0491887741,
-0.0089544524,
-0.0108863767,
-0.0257807132,
-0.0015407949,
-0.0066214208,
-0.0405371673,
0.0161752738,
-0.0205440652,
-0.0481244214,
0.0533043854,
-0.0464199744,
0.0004577682,
-0.0541644022,
0.0108555583,
-0.0475266911,
0.0344597213,
-0.0374777988,
0.0254906844,
-0.0255416445,
-0.0661707893,
-0.0245401617,
0.0092452997,
-0.0259015374,
-0.0097277323,
0.0167950336,
-0.0407966673,
-0.018285431,
-0.0313346758,
-0.0354107916,
-0.0473277159,
-0.0257951021,
-0.0120694386,
0.0334521681,
0.0634758472,
0.0108768484,
-0.0323273726,
0.0394900069,
-0.0350564122,
-0.0300761759,
-0.0319423415,
0.0095490944,
-0.0318991803,
0.0267736893,
0.0455322154,
0.0082003092,
0.0164101347,
-0.0362870209,
0.0143381143,
0.0290590562,
0.0300745387,
-0.0023867018,
-0.0322101116,
0.0226934962,
-0.0151658999,
-0.0421858393,
-0.0079380432,
-0.0125636514,
0.0947972834,
-0.0326087773,
-0.0363440923,
0.1299720258,
0.0180731975,
-0.0038524398,
0.0134531781,
0.0021516939,
0.0315655097,
0.0081686107,
0.0103760613,
-0.0678924918,
-0.0174277145,
0.0016484089,
-0.0147022689,
0.0091243321,
0.0071524587,
-0.0187613107,
-0.1169709042,
0.012026336,
0.0374930315,
-0.048593048,
-0.0324211605,
-0.0184828322,
-0.052474048,
0.0068689464,
0.069122076,
0.0446528383,
-0.0202668272,
0.0047496934,
0.0717872232,
0.0193201136,
-0.0483579747,
0.0212637633,
-0.0593658201,
0.0536372103,
0.0175687857,
0.0003028782,
-0.0125657367,
0.0319046117,
0.0305039231,
-0.0133829499,
0.0014969168,
-0.0141683193,
-0.0025852947,
-0.0175795127,
-0.0297927503,
-0.0348781757,
0.0338104442,
-0.0326394401,
0.0185434278,
0.0024580797,
0.0419187024,
-0.012921907,
0.0151331052,
-0.0093404837,
0.0080043059,
-0.0056032599,
0.040726319,
0.0507976487,
-0.0495625809,
-0.042692516,
-0.0162887275,
-0.0076786624,
-0.0763386339,
0.0304479748,
0.0283808578,
-0.0192968454,
-0.0083229151,
-0.0604098067,
-0.0493076257,
0.020049803,
0.0548637398,
-0.0049301186,
0.0046687918,
-0.0106478278,
-0.0268876161,
-0.0623304173,
0.0422581881,
-0.0125304507,
-0.0020166549,
-0.0943658873,
-0.0482018329,
-0.0129145458,
-0.0376831219,
0.0004265656,
-0.0083727296,
-0.0852222741,
-0.0542475916,
-0.0015547518,
0.0435769148,
0.0242715944,
0.0359390117,
0.0143980458,
-0.0344108865,
0.0078413486,
0.0189838242,
-0.0296646804,
0.0150719937,
0.0210446008,
0.0178702679,
-0.0160405971,
0.0338525176,
0.0070371465,
0.0165512059,
0.0007498247,
0.0409857742,
0.0124127241,
0.0443892516,
0.0468352102,
-0.0010369292,
0.0234226082,
-0.0520768203,
0.0171194114,
-0.0562624782,
-0.0026023893,
-0.0055494686,
-0.0090587195,
-0.0347465836,
-0.0122698406,
0.0439465307,
-0.0129827159,
0.0021434573,
-0.0536022931,
-0.0149614606,
0.0298518725,
0.0238102358,
0.0181281082,
-0.0196879022,
0.0392552763,
0.0092586633,
-0.0030205429,
-0.0289007574,
-0.0185064971,
-0.0162429456,
0.0211470593,
-0.045448862,
0.0728335604,
-0.0110417949,
0.0270203464,
0.0088293981,
0.042641066,
0.0300979968,
-0.0487906337,
-0.0432166532,
0.0185793769,
0.0780782029,
-0.0340509973,
0.0241727736,
-0.0222789645,
0.0080127958,
-0.0504978038,
0.0504163876,
-0.0281428248,
-0.0044745696,
0.0155234309,
0.0038569383,
-0.0362044945,
0.0272675212,
0.0773444921,
0.0238295067,
0.0383854434,
0.0374992788,
0.0244791135,
-0.0304417796,
-0.0014613877,
0.0065441155,
0.0388976336,
-0.0157290492,
0.0088939536,
-0.028887853,
-0.0600148588,
0.0001659364,
0.0011707767,
0.02121404,
0.015699137,
0.052428294,
0.0559068099,
0.002989046,
0.0043941345,
-0.0519856401,
0.018709356,
0.0059819589,
0.0082575474,
-0.0168161616,
-6.711770053e-33,
-0.0462331101,
0.0180291869,
0.025664337,
-0.052589111,
-0.0405443199,
-0.0056619998,
-0.0331440009,
-0.0229804609,
-0.0156005202,
-0.0126541182,
-0.0055394685,
-0.0015867142,
0.0207053144,
0.0495478399,
0.0276973341,
-0.015049166,
0.0547295138,
0.0107507762,
0.0160083994,
0.062922053,
0.0572608188,
0.0605406575,
0.0556320325,
-0.0018218486,
-0.0254768971,
-0.0075447848,
-0.0236535054,
0.0013792927,
-0.0132522443,
0.0846741199,
-0.0496853292,
0.0352248959,
0.0235218331,
0.0055152504,
-0.0484389216,
0.0011134788,
-0.1007776931,
0.0047225519,
0.0237362292,
0.038023714,
0.0126606589,
-0.0393603146,
0.0450602807,
-0.0170716029,
0.0039052134,
0.0093551939,
-0.0126544228,
-0.0053552859,
-0.0321007222,
-0.0072490033,
-0.0677488074,
0.0181151796,
-0.0046359845,
-0.0083241332,
0.0306999609,
0.0212292951,
-0.0307870079,
0.004593675,
-0.0518627837,
-0.0159763899,
0.049676273,
0.0516055338,
0.0297772288,
0.0526064634,
-0.0082538491,
0.0359190181,
0.0386206098,
-0.0305817481,
0.0142293172,
0.0808977708,
0.0354201794,
0.0480881743,
-0.0005477263,
0.0089614606,
0.0413714088,
-0.0424756669,
-0.0817544088,
-0.0063097626,
0.0309522226,
0.0324174948,
0.0229203142,
0.0117369965,
-0.029289959,
-0.0168734863,
0.002163827,
-0.024346685,
-0.0361012556,
-0.0322560407,
0.03827402,
0.0234022215,
-0.0406285152,
-0.0124705005,
0.052932784,
-0.0342731327,
0.0149097042,
-0.0381715856,
-0.0088289985,
-0.0100073172,
-0.0294087548,
0.0424187221,
-0.0626035184,
-0.0961991027,
0.0410926342,
-0.0024610525,
-0.0133583816,
-0.0056309383,
0.0072342427,
0.001599887,
-0.0509037897,
0.0271332748,
-0.0508084036,
0.0179357268,
-0.0147615252,
0.0229620524,
0.0203842539,
0.0000331536,
-0.0203963313,
0.0477678888,
-0.0106190369,
-0.0055301464,
-0.0189392641,
-0.0107780835,
-0.0107677048,
0.0331455916,
-0.0215609334,
0.0486029573,
-0.044179026,
-0.0252161603,
0.0181385893,
-0.0523054004,
-0.0245956089,
0.045375336,
2.674e-7,
0.0508273058,
0.0461928248,
0.0802636892,
-0.0019157659,
0.0092054075,
0.0136488592,
-0.0194987971,
0.0259472933,
0.0024545172,
0.0167325847,
0.0130675407,
-0.0049878471,
0.0459230393,
0.0011740366,
-0.1152439639,
0.020565059,
-0.0826805308,
-0.0037649097,
-0.0842307508,
-0.0124202901,
0.0620943084,
0.1360834837,
-0.0200630836,
-0.0132382726,
0.0247428138,
-0.008081248,
0.0740597397,
0.0059320615,
-0.0250415523,
0.0323435813,
0.0317310244,
0.0363646112,
0.0014315288,
-0.0019420621,
-0.0228735656,
0.0226686671,
-0.0227336418,
0.0786025375,
0.0161601771,
-0.0264947042,
0.0377885252,
0.0044222386,
-0.0132576395,
-0.0714352652,
0.073144421,
-0.0198242348,
-0.0336388126,
-0.0354913697,
-0.0210309997,
0.0290506519,
-0.0216997564,
-0.0345490873,
-0.0414516963,
0.0650955588,
0.020825481,
-0.0212535318,
-0.0160414334,
-0.0653472021,
-0.0225290842,
0.0231153816,
-0.0567429326,
-0.0002059162,
-0.0431019962,
0.0207621939,
0.0364322253,
-0.0456421338,
0.0024967142,
2.59135319e-34,
0.0520483665,
-0.0204438139,
0.0284725092,
0.0389904343,
0.048479218,
-0.0171101596,
-0.0136649339,
0.0173573606,
-0.0125630442,
-0.0547214374,
-0.0251035597
] |
http://arxiv.org/abs/2401.14664v1 | http://arxiv.org/pdf/2401.14664v1 | UNIT-DSR: Dysarthric Speech Reconstruction System Using Speech Unit Normalization | 2024-01-26T06:08:47 | 2024-01-26T06:08:47 | Yuejiao Wang, Xixin Wu, Disong Wang, Lingwei Meng, Helen Meng | cs.SD | cs.SD, cs.CL, eess.AS | Dysarthric speech reconstruction (DSR) systems aim to automatically convert
dysarthric speech into normal-sounding speech. The technology eases
communication with speakers affected by the neuromotor disorder and enhances
their social inclusion. NED-based (Neural Encoder-Decoder) systems have
significantly improved the intelligibility of the reconstructed speech as
compared with GAN-based (Generative Adversarial Network) approaches, but the
approach is still limited by training inefficiency caused by the cascaded
pipeline and auxiliary tasks of the content encoder, which may in turn affect
the quality of reconstruction. Inspired by self-supervised speech
representation learning and discrete speech units, we propose a Unit-DSR
system, which harnesses the powerful domain-adaptation capacity of HuBERT for
training efficiency improvement and utilizes speech units to constrain the
dysarthric content restoration in a discrete linguistic space. Compared with
NED approaches, the Unit-DSR system only consists of a speech unit normalizer
and a Unit HiFi-GAN vocoder, which is considerably simpler without cascaded
sub-modules or auxiliary tasks. Results on the UASpeech corpus indicate that
Unit-DSR outperforms competitive baselines in terms of content restoration,
reaching a 28.2% relative average word error rate reduction when compared to
original dysarthric speech, and shows robustness against speed perturbation and
noise. | UNIT-DSR: Dysarthric Speech Reconstruction System Using Speech Unit Normalization: Dysarthric speech reconstruction (DSR) systems aim to automatically convert
dysarthric speech into normal-sounding speech. The technology eases
communication with speakers affected by the neuromotor disorder and enhances
their social inclusion. NED-based (Neural Encoder-Decoder) systems have
significantly improved the intelligibility of the reconstructed speech as
compared with GAN-based (Generative Adversarial Network) approaches, but the
approach is still limited by training inefficiency caused by the cascaded
pipeline and auxiliary tasks of the content encoder, which may in turn affect
the quality of reconstruction. Inspired by self-supervised speech
representation learning and discrete speech units, we propose a Unit-DSR
system, which harnesses the powerful domain-adaptation capacity of HuBERT for
training efficiency improvement and utilizes speech units to constrain the
dysarthric content restoration in a discrete linguistic space. Compared with
NED approaches, the Unit-DSR system only consists of a speech unit normalizer
and a Unit HiFi-GAN vocoder, which is considerably simpler without cascaded
sub-modules or auxiliary tasks. Results on the UASpeech corpus indicate that
Unit-DSR outperforms competitive baselines in terms of content restoration,
reaching a 28.2% relative average word error rate reduction when compared to
original dysarthric speech, and shows robustness against speed perturbation and
noise. | [
-0.0395605937,
0.1188632846,
0.009927691,
-0.0115393233,
0.0121557536,
-0.0152048981,
0.0342406705,
0.0177000165,
-0.0667736679,
0.0208046809,
-0.0404811576,
-0.0512598678,
0.0438121296,
0.0062636659,
0.0395539925,
-0.0517802164,
0.0252041947,
-0.0145917563,
-0.0215046257,
-0.0621791407,
-0.0054512057,
-0.0320989415,
-0.059115991,
0.0254722014,
-0.0278881639,
-0.0002401401,
0.0084192408,
0.0150669627,
0.0036886323,
-0.0281676576,
0.0359546952,
0.0179831125,
0.0701321289,
-0.0046580206,
0.0000023368,
-0.0050335573,
0.0089915628,
-0.029409904,
-0.0325897262,
0.0060203904,
-0.0118050659,
-0.0001635896,
-0.0030733235,
0.0483121611,
0.0051440722,
-0.0528663173,
0.0114944475,
0.0513412654,
0.0187184196,
0.0448634364,
-0.0233898051,
-0.0700214207,
-0.0003376947,
-0.0292896163,
-0.0902085751,
0.0351632908,
0.0146403434,
0.0388198122,
-0.0098807756,
-0.0373519734,
-0.0152003476,
0.0473062135,
-0.0032403071,
-0.0067903083,
0.0459697805,
-0.0128483372,
0.0308501218,
-0.0024821435,
0.0152686182,
0.0474702716,
-0.0029753298,
0.0204975996,
0.020106338,
0.0229641516,
-0.0217691008,
-0.0152508691,
-0.0329569355,
-0.0334927849,
0.0313113369,
0.0140126916,
-0.0257887766,
0.0009001228,
0.0243198089,
-0.0352315567,
0.029294908,
0.0063151238,
0.0180021785,
0.0097653419,
-0.0110467989,
-0.0280126501,
0.0198712684,
0.0009861607,
0.0853672326,
0.0306505598,
-0.0500825644,
-0.0079890592,
-0.0480584502,
-0.0365201868,
0.0022777794,
-0.0058891755,
-0.0281149764,
-0.011399175,
-0.0278700609,
-0.0433100984,
-0.003421305,
0.0286832936,
-0.0000461872,
0.0218446497,
-0.0223685298,
-0.002196868,
0.0174776558,
0.0190673061,
-0.0388897546,
0.031873934,
-0.0456721671,
-0.0031895402,
-0.0347525142,
-0.0003446978,
0.0066915415,
-0.010066309,
0.0332872309,
0.0447275601,
-0.0851271823,
-0.0373665839,
0.0206198096,
0.0764798373,
-0.0067328266,
0.063138701,
0.0268742219,
-0.0710566565,
-0.0309202783,
-0.0085592959,
0.039934542,
0.0080202613,
0.0113180494,
0.0735962018,
-0.0112654055,
0.0010583367,
-0.0774661079,
0.0070025553,
-0.0437362045,
-0.0041032233,
-0.008539998,
-0.0689639002,
-0.0133903567,
0.0194908958,
-0.0052293576,
-0.0909572095,
-0.007319761,
-0.0217026211,
0.0327238292,
0.0416194536,
-0.0133076608,
0.0204257015,
0.0963729098,
-0.0123348208,
0.007021185,
0.0032782271,
0.0283559058,
0.003055545,
-0.013960368,
-0.0256618131,
0.0108323786,
0.0325021893,
0.0236174725,
0.0242451541,
-0.0550237745,
-0.0624874085,
-0.0469021499,
0.0237986147,
0.0963549688,
0.0574326143,
-0.0130546978,
-0.04423647,
0.0764934942,
0.0526761152,
0.0618257523,
0.0388821959,
-0.0308604278,
0.0090716463,
-0.0027364066,
0.0114125563,
0.0166172236,
0.0291009564,
0.0154251726,
-0.0081414534,
-0.0014693949,
0.0622189865,
-0.0419344604,
-0.0250222944,
0.0370730013,
-0.0198194403,
0.0287890062,
0.0205177963,
0.0500617661,
0.0166195501,
-0.0316373594,
-0.0175922383,
0.0312130842,
-0.0396075659,
0.0824422389,
0.0024162121,
-0.0171331968,
0.0128017822,
0.0462519228,
-0.0035869842,
-0.0145240687,
-0.0575286448,
-0.1011082083,
0.0215377286,
-0.0624275319,
0.0087361122,
0.0579492562,
-0.0605268925,
-0.033974532,
0.0065392847,
-0.017811669,
0.0532287955,
-0.0461144336,
0.037715476,
-0.0144717544,
-0.0155034205,
0.016561104,
0.0178577397,
-0.0568358153,
-0.0395682752,
0.0446160287,
-0.0001855897,
0.0333931483,
0.0412505567,
-0.0393730439,
-0.0224171672,
-0.011964079,
-0.0227643047,
-0.0757060125,
0.0224562231,
-0.0929771438,
-0.0670764223,
0.0109467655,
-0.0334170647,
0.0087366449,
0.0902854279,
0.0359769836,
-0.0684399605,
0.0316714682,
-0.0551474243,
-0.0302129183,
-0.074129723,
0.0281824898,
0.0297414828,
0.0375036746,
0.0175602585,
-0.0072291181,
0.0035201272,
-0.0325714312,
-0.0088633103,
0.0151322074,
0.0290611219,
-0.0661623031,
-0.0523626842,
0.0106032044,
-0.034564551,
0.0457206815,
0.0320631191,
-0.061623551,
0.0008959168,
0.0197919402,
-0.0243471153,
-0.0074247736,
-0.0206919909,
0.0061226226,
-0.0036780017,
0.0443857983,
0.0119087314,
-0.0471862555,
-0.0365123861,
-0.0232549161,
0.03791425,
-0.0175867267,
0.0170848984,
0.003855265,
-0.013243163,
-0.01487872,
0.0187061597,
0.0249564853,
-0.0137483282,
-0.0458222963,
-0.0681911781,
-0.0339287333,
0.0589007996,
0.0307519399,
0.0614409707,
-0.0131055526,
-0.0659935176,
-0.0147224115,
0.0766136795,
0.0712355301,
0.0249363109,
0.0004140142,
-0.0114437621,
0.0187845398,
-0.0014580613,
-0.026096167,
-0.0088552264,
-0.0107488083,
0.0726610646,
-0.0278457906,
0.0212306939,
0.0221362207,
0.0354509391,
0.083910808,
0.0209651403,
-0.0240382012,
-0.0412217341,
0.0001274649,
-0.0525855385,
-0.0269179232,
-0.0158825647,
-0.0299543124,
-0.0166134853,
0.0012096586,
-0.0173287727,
-0.0003712359,
-0.0496331118,
-0.0465753451,
-0.0142818363,
-0.0563538261,
0.0297749676,
-0.0202127118,
-0.0180321001,
0.0211321842,
-0.0179476738,
-0.0870658234,
-0.0373944603,
0.0342454463,
0.0112433163,
-0.0407381393,
0.0377588384,
-0.0009034813,
-0.0180296917,
-0.0311430022,
-0.0322706029,
-0.0542866737,
0.0511478409,
0.0306043625,
0.0345020071,
0.0395338237,
-0.0304758567,
-0.032621447,
-0.0173805878,
0.0210993551,
0.0743088424,
0.0037788718,
0.0232632775,
-0.0088450266,
0.0923394561,
0.0103314705,
0.032427799,
-0.0355295315,
0.0007880176,
0.0112806382,
0.0051763221,
-0.022038633,
0.0355462506,
-0.0618707351,
-0.0263847765,
-0.0073015592,
0.0242886152,
0.0302038509,
0.0099918265,
0.044984255,
0.0209826436,
-0.000049442,
0.047079172,
-0.0396471433,
-0.0491196588,
0.0015184482,
0.0039953203,
-0.0088556167,
-0.0166105721,
-0.0328028873,
-0.0417719558,
-0.0105871558,
0.0377783924,
0.0365503207,
0.0469270498,
0.017827034,
0.0185028911,
-0.0278576203,
0.0321058035,
0.0384854041,
0.0117238536,
-0.0451369472,
-0.0230744127,
-0.0047815614,
-0.0192518663,
0.0820483193,
0.0164670683,
0.0408424623,
-0.018180307,
0.044592686,
-0.0478466749,
-0.0476216674,
-0.0181263648,
-0.0333849378,
0.117569156,
0.0552231967,
0.0105504775,
-0.0096316254,
-0.0505561084,
0.0172493309,
0.0497335941,
-0.0005873893,
0.0241108816,
0.0455222912,
0.00859515,
-0.0072911577,
-0.0011274306,
0.0707975924,
-0.0473181307,
0.0138370311,
-0.0043098857,
0.0021256357,
-0.0293977465,
0.0365549773,
-0.041441679,
-0.0019734795,
0.0260166582,
0.0394722372,
0.008693126,
-0.0107706226,
-0.0187744107,
0.0048276316,
0.0392364636,
0.0072511141,
-0.024997199,
0.0500598773,
0.0154635105,
-0.0333479755,
-0.0549376011,
-0.0015054916,
-0.0329456255,
0.026003303,
-0.0062006786,
0.0156946741,
0.0383074284,
0.0210236702,
-0.0254619867,
-0.0350434743,
0.0201083831,
0.0189034026,
-0.0418330282,
-0.0271677133,
0.0036259275,
-0.0043266611,
-0.0116720954,
0.000398463,
0.0194085911,
-0.0346505381,
0.0637935251,
-0.0444510765,
0.004311135,
0.0271040704,
-0.0025620882,
-0.0170136262,
0.0589397997,
0.063330479,
-0.0375003144,
0.0184766501,
0.0375959836,
0.0293006673,
-0.0075351931,
0.0354711004,
-0.0117810462,
0.084664315,
-0.0495709293,
0.0268621948,
0.031660147,
-0.0392652899,
-0.0040156357,
0.025806237,
0.040435385,
0.0251736511,
0.0538165271,
-0.0327280164,
0.007982864,
-0.0409382917,
0.0796249956,
-0.048110757,
-0.0281702932,
0.0451904275,
0.0070992536,
-0.00340197,
0.0032268558,
-0.0219839253,
-0.0129788071,
-0.0165559072,
0.0277863108,
0.0121549768,
0.0041750292,
-0.0333964676,
0.0039250306,
0.022317823,
0.018662218,
0.0260422565,
0.0096395081,
-0.0052373768,
0.0146663394,
0.0008933237,
-0.026359925,
0.0079998858,
-0.0027252315,
-0.0006524088,
-0.0160042532,
-0.0583085008,
-0.0601307899,
0.0003213049,
0.0439917818,
0.0310817473,
0.0118462192,
0.037429098,
-0.016316589,
-0.0206366666,
-0.0306479931,
0.0404488258,
0.0340495892,
-0.0248850267,
-0.002487744,
0.019652009,
0.0435811244,
-0.0398478322,
-0.0057182009,
-0.0191793013,
-0.0008155904,
-0.0185419153,
-0.0698720887,
0.0259686131,
-0.0146292616,
-0.0289363526,
0.0045207958,
0.045628354,
-0.043024797,
-0.0527688973,
-0.0142615223,
0.0472617634,
-0.0533642992,
0.0043009873,
-0.0358970277,
-0.0478128456,
-0.0258067828,
0.0296804626,
-0.0134309409,
-0.0101776989,
-0.034279108,
0.0133103803,
-6.561944482e-33,
-0.0212270673,
0.0253342241,
-0.0038244748,
-0.0308596212,
-0.0366345309,
-0.0334946886,
-0.0272984579,
0.022328021,
0.0129262693,
0.0182428192,
-0.0278770402,
0.0371981524,
0.0059871534,
0.0458177663,
-0.0233016107,
0.0425815918,
0.1026631966,
-0.0380646214,
-0.0109270029,
0.0108009856,
0.0428159125,
0.061405085,
0.0614425018,
-0.0718801767,
-0.0388079174,
0.0124559412,
-0.0166256353,
-0.0417822339,
0.0560348034,
-0.0767381936,
-0.0131941549,
0.043955259,
-0.0149840452,
-0.0571293831,
-0.0134373056,
0.0214931592,
-0.0972273797,
0.0209214259,
-0.0552832074,
0.0095340936,
-0.0603965819,
0.0253296383,
0.0601373911,
-0.0190296676,
-0.0264203735,
-0.0134657184,
0.0069756764,
-0.0419803634,
0.0127655752,
-0.0286137685,
-0.0049955542,
0.0242572557,
-0.0050273128,
-0.0303815361,
0.0263979919,
0.0019753342,
-0.0337714776,
-0.0588240921,
-0.0447486117,
0.0003534481,
0.0084541449,
0.0681472048,
0.0066661355,
-0.0372572541,
0.0031578667,
0.0393684469,
0.088502571,
-0.0312412288,
0.0152166449,
-0.01972262,
0.0345953852,
0.0393634848,
0.0220177993,
0.0239561368,
0.071725443,
-0.0643299893,
-0.015270221,
0.060959518,
-0.0056725522,
0.0572244152,
0.0036169521,
0.0748404786,
-0.0147825107,
0.0004702285,
-0.0200782642,
-0.0215366445,
0.0262718834,
-0.048140429,
0.0255976208,
-0.0254544504,
0.0008909641,
-0.029747922,
0.0386050902,
-0.0455330051,
0.0334098078,
0.0415547937,
-0.0072694854,
-0.0190665852,
0.0012297769,
-0.0260642301,
0.018020466,
-0.0068255188,
0.07372576,
0.0691245422,
0.0039392519,
-0.0321467519,
-0.0334990546,
0.0483069979,
-0.0463718548,
-0.0031965135,
-0.0110506648,
0.0440240502,
0.0136040375,
0.0467404649,
0.0006027861,
-0.0345248282,
0.0053304415,
-0.0152662676,
-0.0330566354,
0.0211315881,
-0.0307315812,
0.0364736505,
0.0238969605,
-0.0021535147,
-0.0474693924,
-0.0022747673,
-0.0048377831,
-0.0205977056,
0.0493355878,
-0.0282006375,
0.0060572992,
-0.0626120195,
3.087e-7,
-0.0643142909,
0.0287670065,
0.0636955053,
-0.0321123302,
0.0473330505,
-0.0006381619,
-0.0154033192,
0.0332233869,
-0.0265613776,
-0.0039784363,
0.0266784728,
-0.0016105026,
0.0385183841,
0.0311461911,
-0.0165832583,
0.0127069829,
-0.0680638179,
0.0044293352,
0.0100789154,
-0.002755271,
-0.0376526937,
-0.0004670964,
-0.0514846332,
-0.0633056164,
0.0462832078,
0.0001501494,
-0.0013509071,
0.0303127002,
0.0288819242,
-0.0006236754,
0.0429948643,
0.0084825018,
0.0359011479,
-0.0299917422,
-0.0486280397,
0.0674718469,
-0.0188222714,
0.0591706447,
-0.0430312268,
-0.0586688779,
0.024278732,
0.0788060129,
0.0111526744,
-0.000924635,
-0.0243734214,
-0.0211384352,
-0.0341144837,
-0.0258284658,
-0.0541527532,
0.0033927038,
0.0237021744,
0.0199431293,
-0.0019517941,
-0.0476716124,
0.00443997,
0.034102235,
-0.0144701432,
-0.0457293242,
0.0396482721,
0.0122667598,
-0.0869469419,
0.0075903712,
-0.0049008788,
0.011111238,
-0.0279459693,
-0.0220523756,
0.0049178167,
2.230911724e-34,
0.0085815182,
-0.0010517838,
0.0062733716,
-0.0230028406,
-0.0036253016,
-0.0437703617,
0.0287484862,
0.01087928,
0.0070045907,
-0.0112483706,
-0.0627452955
] |
http://arxiv.org/abs/2401.14656v1 | http://arxiv.org/pdf/2401.14656v1 | Scientific Large Language Models: A Survey on Biological & Chemical Domains | 2024-01-26T05:33:34 | 2024-01-26T05:33:34 | Qiang Zhang, Keyang Ding, Tianwen Lyv, Xinda Wang, Qingyu Yin, Yiwen Zhang, Jing Yu, Yuhao Wang, Xiaotong Li, Zhuoyi Xiang, Xiang Zhuang, Zeyuan Wang, Ming Qin, Mengyao Zhang, Jinlu Zhang, Jiyu Cui, Renjun Xu, Hongyang Chen, Xiaohui Fan, Huabin Xing, Huajun Chen | cs.CL | cs.CL | Large Language Models (LLMs) have emerged as a transformative power in
enhancing natural language comprehension, representing a significant stride
toward artificial general intelligence. The application of LLMs extends beyond
conventional linguistic boundaries, encompassing specialized linguistic systems
developed within various scientific disciplines. This growing interest has led
to the advent of scientific LLMs, a novel subclass specifically engineered for
facilitating scientific discovery. As a burgeoning area in the community of AI
for Science, scientific LLMs warrant comprehensive exploration. However, a
systematic and up-to-date survey introducing them is currently lacking. In this
paper, we endeavor to methodically delineate the concept of "scientific
language", whilst providing a thorough review of the latest advancements in
scientific LLMs. Given the expansive realm of scientific disciplines, our
analysis adopts a focused lens, concentrating on the biological and chemical
domains. This includes an in-depth examination of LLMs for textual knowledge,
small molecules, macromolecular proteins, genomic sequences, and their
combinations, analyzing them in terms of model architectures, capabilities,
datasets, and evaluation. Finally, we critically examine the prevailing
challenges and point out promising research directions along with the advances
of LLMs. By offering a comprehensive overview of technical developments in this
field, this survey aspires to be an invaluable resource for researchers
navigating the intricate landscape of scientific LLMs. | Scientific Large Language Models: A Survey on Biological & Chemical Domains: Large Language Models (LLMs) have emerged as a transformative power in
enhancing natural language comprehension, representing a significant stride
toward artificial general intelligence. The application of LLMs extends beyond
conventional linguistic boundaries, encompassing specialized linguistic systems
developed within various scientific disciplines. This growing interest has led
to the advent of scientific LLMs, a novel subclass specifically engineered for
facilitating scientific discovery. As a burgeoning area in the community of AI
for Science, scientific LLMs warrant comprehensive exploration. However, a
systematic and up-to-date survey introducing them is currently lacking. In this
paper, we endeavor to methodically delineate the concept of "scientific
language", whilst providing a thorough review of the latest advancements in
scientific LLMs. Given the expansive realm of scientific disciplines, our
analysis adopts a focused lens, concentrating on the biological and chemical
domains. This includes an in-depth examination of LLMs for textual knowledge,
small molecules, macromolecular proteins, genomic sequences, and their
combinations, analyzing them in terms of model architectures, capabilities,
datasets, and evaluation. Finally, we critically examine the prevailing
challenges and point out promising research directions along with the advances
of LLMs. By offering a comprehensive overview of technical developments in this
field, this survey aspires to be an invaluable resource for researchers
navigating the intricate landscape of scientific LLMs. | [
0.0654663891,
0.039145276,
-0.0136253303,
0.0116679035,
-0.0534514785,
0.0201037098,
0.0140019832,
0.0205815509,
0.0371799581,
-0.0387789421,
-0.0158363301,
-0.0409891345,
-0.0032051313,
0.0591907948,
0.0270602647,
0.0190822426,
0.031697955,
-0.0353493392,
0.0276739821,
-0.011656126,
-0.0019584578,
-0.0004732478,
0.0208975412,
0.0122189298,
0.0412010401,
0.002173417,
0.0181127246,
-0.0497507341,
0.0117307855,
-0.0192203913,
-0.0394502431,
0.0154491113,
0.0421721376,
0.0633221269,
0.0000016249,
-0.0672766343,
-0.0062113455,
0.0513281375,
-0.0324522071,
0.0161784217,
0.0356406532,
-0.0333843231,
0.0039998158,
-0.0146115059,
-0.0222041886,
0.0004267262,
0.0789969042,
-0.0084969169,
0.0067486265,
0.0359788053,
0.0049220803,
-0.0710180476,
0.0712909251,
-0.0061126896,
0.0231921896,
0.0078201173,
0.1006999835,
0.021344522,
0.0663546547,
0.0255235471,
-0.0068642264,
0.0495769307,
-0.006931385,
-0.0256121866,
0.0080415327,
0.0377971418,
0.0078618824,
-0.0364681333,
0.0059719686,
0.0658130199,
0.0098374179,
-0.056191802,
0.0015189967,
0.0210853238,
-0.0265017562,
0.0308604222,
-0.0641974583,
0.0317171179,
0.0096558174,
0.032436315,
0.0567539409,
-0.0179864019,
0.0506527387,
0.039633505,
-0.0398636758,
0.0798268393,
0.0250897408,
-0.0348589644,
-0.0305618793,
-0.026149001,
-0.0429214835,
-0.0373486169,
0.0388873108,
0.033560656,
0.0018278927,
0.0241573919,
0.0128356153,
-0.0291211046,
0.0582045764,
-0.0313606411,
0.0252239183,
-0.0216011181,
0.0105447676,
0.0541590638,
-0.0068022292,
0.0275431219,
-0.04804454,
0.0075571528,
-0.0255534369,
-0.0071604997,
-0.0788143054,
-0.0395208001,
-0.004462312,
0.0336621851,
-0.0261529535,
-0.0205947645,
-0.0377146304,
0.0016120196,
0.0309680663,
-0.0033640561,
-0.0016328401,
-0.0084029222,
-0.0233250931,
0.0083254557,
-0.0311438423,
0.0176891908,
0.0127189197,
-0.0022058336,
-0.0399946794,
-0.0549784303,
-0.0361123495,
0.032180272,
-0.024065394,
-0.0232336894,
0.014731694,
0.0938212276,
0.0484367572,
-0.0449197665,
0.0261466131,
-0.0242311098,
-0.0476677008,
-0.0739138052,
0.0268682353,
-0.0322101526,
0.042627871,
0.0209055897,
-0.0108855553,
-0.0371376947,
0.0154644465,
-0.0165084414,
-0.052834522,
0.0825640336,
-0.0366358906,
-0.0230799541,
0.0403240658,
-0.0095744813,
0.0339882672,
0.009804911,
0.0370656364,
-0.0516929142,
0.013773609,
-0.0225351639,
0.0054138126,
-0.000487081,
-0.0109594995,
-0.0408139713,
-0.0230165292,
-0.0043803519,
0.0191372074,
0.0611535646,
-0.0351216197,
0.0559575222,
0.0138206128,
-0.0399288647,
0.0343121476,
0.0774392262,
0.0002388111,
0.039360594,
0.0526712425,
0.0093907351,
0.0241300464,
-0.0081718285,
0.0020150703,
0.0137178451,
-0.0526911542,
0.0211006701,
0.0307082627,
0.0391446874,
0.0114436401,
-0.0369806252,
-0.0337814279,
0.0024127255,
0.0080284402,
-0.0323748477,
0.0863474384,
-0.0296601001,
-0.0347754024,
-0.0031909551,
0.0612116158,
-0.0182828102,
0.0287879612,
-0.0460130051,
0.0417603999,
0.0604149848,
0.0603788905,
-0.0537204519,
-0.0128194038,
0.0100092599,
-0.0305385999,
0.0455644242,
0.0027989552,
-0.0039422023,
-0.040734496,
-0.0258744396,
-0.000417329,
-0.0201166328,
0.0148185873,
-0.0081529943,
-0.0325761661,
-0.0073557817,
-0.0036109353,
-0.0086825443,
-0.0346041545,
0.0354570299,
-0.0427367017,
0.0073146671,
-0.011637467,
-0.0476213582,
0.0477378517,
0.0193521194,
-0.0255101416,
0.0484813228,
0.075434342,
-0.060235329,
-0.0321896821,
-0.0221368838,
0.0270286947,
-0.0396849476,
0.046483431,
-0.0004589794,
0.0107684303,
-0.011775,
-0.0091186427,
-0.0427273475,
0.0504866838,
-0.0508708581,
0.0310113076,
-0.0299763493,
0.0071856086,
-0.0276174042,
0.0254941136,
-0.0471821576,
0.0508164652,
0.0065768869,
0.025316244,
0.0136729898,
-0.0184341483,
0.0161663909,
-0.0493530296,
-0.0134139024,
0.0143536953,
0.0422684811,
-0.0251325388,
0.0170773081,
-0.0590386204,
-0.0141723445,
-0.0194587428,
-0.0190083217,
-0.0193794575,
0.0092907716,
0.0333580337,
-0.0128807724,
-0.0196064766,
0.0078228749,
0.0311679859,
-0.0106882676,
0.0069234734,
0.0192534588,
0.0004459476,
0.0439740829,
0.0212890673,
-0.0656919181,
-0.0396312438,
-0.0269104559,
-0.0675415769,
0.0804111287,
0.0275168717,
-0.055498749,
-0.047408998,
0.0081191408,
-0.0331417061,
0.013480437,
-0.0042584469,
-0.0291361269,
0.0030721384,
-0.0329190083,
0.0111613572,
0.0494658612,
-0.0130701642,
0.0297310371,
0.0747189447,
0.0111342361,
0.0039515607,
-0.0179110896,
-0.0642281771,
0.0117294304,
-0.0244109891,
0.0078439042,
-0.0130131477,
0.0690470859,
-0.0281175282,
-0.0317068659,
-0.0322366022,
0.0452864021,
-0.0328170322,
-0.0417182744,
0.0229187496,
-0.0125533724,
-0.0145109184,
0.0216136836,
-0.0001579671,
-0.0588178411,
0.1010597199,
-0.031324584,
0.0110598942,
-0.0664209872,
-0.0234874003,
-0.0251247585,
0.0302701574,
0.0097819092,
0.0192243326,
0.0352143794,
0.025574483,
0.0285375975,
0.0290396046,
-0.0463737287,
-0.0213346127,
0.026507942,
-0.06567882,
0.0351626351,
0.0407585129,
-0.0074939621,
0.0212541707,
0.0059128273,
-0.0533525236,
0.0458588898,
0.0360281989,
-0.0587658919,
-0.045569025,
0.0345896892,
0.0249901768,
-0.0118353255,
-0.0085122017,
0.019246662,
-0.0350600369,
-0.0146168321,
0.0575707331,
-0.0265612807,
0.0438029319,
-0.0210837498,
0.047654178,
0.0426957346,
-0.0099835657,
0.0124946209,
0.0057474626,
-0.0011800398,
0.0124503663,
-0.0052415146,
-0.0202661026,
-0.0013763993,
-0.0039817346,
-0.0473450869,
-0.0277487617,
0.1137035713,
0.032107912,
-0.0270085875,
0.0206430126,
-0.0000658069,
-0.0220706332,
0.0010255426,
0.0432161242,
-0.0757665411,
-0.0157893673,
-0.0258636717,
-0.0299072526,
-0.0404658653,
-0.057685338,
-0.0243175123,
-0.0618707612,
0.0454909764,
0.0306531917,
-0.0210178122,
-0.0374114737,
-0.0872841999,
-0.0243187733,
-0.0208358914,
-0.00266639,
0.0142462095,
0.0103827631,
-0.0105244555,
0.0166637562,
-0.0064311433,
-0.0975062624,
0.0122172972,
-0.0188970268,
0.037547797,
0.0591888316,
0.0148814777,
-0.0048475089,
-0.0086710742,
-0.0044669467,
-0.0206437577,
0.0182603635,
-0.0320143029,
0.0077211368,
0.0329934657,
-0.0197356381,
-0.0489796586,
0.0488066152,
-0.0520446077,
-0.0216323435,
-0.017443005,
0.1299509555,
-0.0411732756,
-0.0317322277,
-0.0430535413,
0.0046853703,
-0.0347766913,
0.1024257913,
0.0103166122,
-0.0364414155,
-0.0668293983,
-0.0105388444,
0.0144089172,
0.0069200499,
-0.0446715355,
-0.0113604562,
-0.0102137784,
0.0034524743,
-0.0341060758,
0.0142123969,
0.0497450307,
-0.0087850541,
0.0253908634,
0.042399101,
0.0380602218,
-0.0537505485,
-0.0319074877,
0.011119335,
-0.0088309916,
0.0211941544,
-0.0582868457,
-0.0238487925,
0.0913882777,
-0.0166360606,
-0.02626854,
0.0470734164,
-0.0460829847,
-0.1000972837,
-0.0221125782,
0.0284999125,
0.0239147637,
-0.0071480051,
-0.0007509029,
-0.0892222747,
0.0512652993,
0.0298869442,
-0.0052419016,
0.009318579,
-0.0409834795,
0.0099350223,
-0.006422569,
0.0142583074,
-0.0203382373,
0.0238820519,
-0.0113356952,
0.048033677,
-0.0249498282,
-0.0299178455,
0.0559359267,
0.0252893586,
0.0376773179,
0.0008677047,
-0.0136784688,
-0.0073190071,
0.0026237958,
-0.0343095697,
-0.0096752858,
-0.0176345408,
-0.0077752704,
0.031413544,
0.0019406931,
-0.0208723638,
-0.0253598299,
-0.039466694,
0.0220996533,
-0.0212750621,
0.0300018173,
-0.0023607113,
-0.020672936,
0.0404855125,
-0.0152668757,
0.0119989244,
0.0241871774,
-0.0611758679,
0.0234429147,
-0.0117884651,
0.0278030634,
0.0161788799,
0.0238329694,
-0.0372648425,
0.0171518829,
-0.0743538514,
-0.009432191,
0.0100038452,
0.0411104113,
0.0319939777,
0.0158057064,
0.0315419734,
-0.0433538109,
-0.0328716114,
-0.0400984176,
0.0183821078,
-0.0085304081,
-0.0115615912,
0.0667581037,
0.0298750158,
-0.0530547798,
-0.0227205716,
0.0576520562,
0.0540456921,
0.0738813356,
-0.0177627597,
-0.0319824032,
-0.0001647326,
-0.0462798178,
-0.0459266603,
-0.0210399162,
-0.0353942141,
0.0123337926,
-0.0787765607,
-0.0501539931,
0.0357243903,
0.0293745082,
0.0189314205,
-0.0196153335,
0.0090179676,
0.0361616015,
0.0037218949,
-0.0442442112,
-0.0166026372,
0.0318026841,
-0.0100313304,
-0.042555809,
0.0016195253,
-5.592249437e-33,
0.0146799115,
-0.0220190305,
0.0069812476,
0.021855941,
0.0082736202,
0.0121460389,
-0.052913554,
-0.0290674567,
0.0096566202,
0.0557419062,
0.020645082,
-0.0022452241,
-0.011834898,
-0.0112907942,
0.0055578495,
0.0170008149,
0.0171338916,
-0.0167163331,
0.0435434282,
0.0155548817,
0.0452686772,
0.0327794626,
0.0206293426,
-0.0486463197,
0.0163461585,
0.028367918,
0.0192872696,
0.0389874429,
-0.0728998184,
0.0716086701,
-0.0547809936,
0.0274829846,
-0.0027358206,
-0.0168157145,
-0.0247744638,
0.004510833,
-0.0341166891,
0.0376862101,
0.0344221704,
-0.0373721272,
-0.0172600951,
-0.0531315617,
0.0430432558,
-0.0137514621,
-0.0573516749,
0.0195878204,
0.0117607461,
-0.0506130941,
-0.015438146,
-0.0353259183,
-0.0625096932,
0.0158516821,
-0.0030996683,
0.025269039,
0.0511552468,
-0.0255220216,
-0.0111477757,
0.0753177181,
0.0161061734,
0.0660811886,
0.0410776064,
0.025997635,
0.0541834459,
0.001084278,
0.0315108895,
0.0439268127,
0.0179741029,
-0.0122262277,
-0.005223861,
-0.0034781052,
0.0146033391,
0.025256956,
0.0461063422,
0.0188774578,
0.0020745301,
0.0399958454,
-0.0955571905,
-0.0168915782,
-0.0166785885,
0.0287682582,
0.0087587312,
0.0340984911,
0.0055198842,
-0.0106327012,
-0.0389828943,
0.0211481526,
-0.0123427818,
-0.0053236336,
-0.0164226089,
-0.0527050197,
-0.033462435,
0.0270444322,
0.0150973992,
-0.0682984963,
-0.0362525433,
0.0244627278,
0.0273342617,
0.0212322753,
-0.0042337263,
0.0380461887,
-0.0908634663,
-0.0643051043,
0.0225336608,
-0.0264847334,
0.0116985021,
-0.0334513299,
-0.0186578892,
0.0184674859,
0.0074243154,
0.0019423319,
-0.0273570307,
-0.036994528,
0.0511304848,
0.0131347217,
0.0113674477,
0.0457215384,
0.020486977,
0.0494517013,
0.0383758545,
0.0023616068,
0.033447586,
0.0309682507,
0.0035195425,
0.0120506575,
-0.0588023216,
-0.0213196166,
-0.0171875283,
0.0108160237,
0.0107848896,
-0.0639249608,
-0.004668518,
0.0022759384,
2.578e-7,
0.0750441477,
0.0819687396,
0.0334761366,
-0.0014874812,
0.0329563208,
-0.0014582492,
-0.0120555051,
0.0157459956,
-0.0258995164,
-0.023611486,
0.0026036543,
0.0004424108,
0.0095378412,
-0.0181961823,
-0.0351528227,
0.0070598596,
-0.0463230722,
-0.0320193619,
-0.0684792399,
0.0079059759,
0.0179256909,
0.1025589556,
0.0032525191,
-0.0094628427,
0.0024430717,
-0.0490408204,
0.046783492,
-0.0333253071,
-0.0018298262,
0.0687017813,
-0.0226084497,
-0.0161544774,
0.0094503146,
0.0641215518,
0.0089400942,
0.0225024205,
0.0057345838,
0.0482373312,
-0.0233896207,
0.0164808817,
0.055835627,
-0.0011130639,
0.0043356144,
-0.0432233475,
0.027961785,
-0.0626919717,
-0.0778513402,
-0.0171345174,
0.0091993315,
-0.0375770442,
-0.0027611286,
-0.0237653684,
0.025707176,
0.0227212682,
-0.0284247883,
0.00618969,
-0.0008595072,
-0.0618056543,
0.0027647659,
0.0112413019,
-0.0671989769,
-0.0136366943,
-0.044986546,
-0.0119109228,
0.0297512412,
-0.054757271,
0.002740419,
2.579034835e-34,
0.0523631647,
-0.0231727958,
-0.0127827134,
0.0328586735,
0.0285483897,
-0.0079563335,
-0.0141432891,
0.0402658135,
-0.0003936331,
-0.1468756348,
0.002021401
] |
http://arxiv.org/abs/2401.14654v1 | http://arxiv.org/pdf/2401.14654v1 | A Korean Legal Judgment Prediction Dataset for Insurance Disputes | 2024-01-26T05:26:27 | 2024-01-26T05:26:27 | Alice Saebom Kwak, Cheonkam Jeong, Ji Weon Lim, Byeongcheol Min | cs.CL | cs.CL, cs.LG | This paper introduces a Korean legal judgment prediction (LJP) dataset for
insurance disputes. Successful LJP models on insurance disputes can benefit
insurance companies and their customers. It can save both sides' time and money
by allowing them to predict how the result would come out if they proceed to
the dispute mediation process. As is often the case with low-resource
languages, there is a limitation on the amount of data available for this
specific task. To mitigate this issue, we investigate how one can achieve a
good performance despite the limitation in data. In our experiment, we
demonstrate that Sentence Transformer Fine-tuning (SetFit, Tunstall et al.,
2022) is a good alternative to standard fine-tuning when training data are
limited. The models fine-tuned with the SetFit approach on our data show
similar performance to the Korean LJP benchmark models (Hwang et al., 2022)
despite the much smaller data size. | A Korean Legal Judgment Prediction Dataset for Insurance Disputes: This paper introduces a Korean legal judgment prediction (LJP) dataset for
insurance disputes. Successful LJP models on insurance disputes can benefit
insurance companies and their customers. It can save both sides' time and money
by allowing them to predict how the result would come out if they proceed to
the dispute mediation process. As is often the case with low-resource
languages, there is a limitation on the amount of data available for this
specific task. To mitigate this issue, we investigate how one can achieve a
good performance despite the limitation in data. In our experiment, we
demonstrate that Sentence Transformer Fine-tuning (SetFit, Tunstall et al.,
2022) is a good alternative to standard fine-tuning when training data are
limited. The models fine-tuned with the SetFit approach on our data show
similar performance to the Korean LJP benchmark models (Hwang et al., 2022)
despite the much smaller data size. | [
0.077720657,
0.0717204064,
-0.0235771239,
0.0130921975,
-0.0391221717,
0.0218029097,
0.0136270681,
0.0601784736,
-0.013338997,
-0.0544911847,
-0.040958073,
-0.0171510093,
0.0060541634,
0.0157101024,
0.0036724841,
-0.0813661143,
0.0204342194,
0.014246542,
0.0733087882,
-0.0113412458,
0.0179614872,
-0.0085788975,
0.0134145077,
0.0420606397,
-0.0516684949,
0.0134415794,
-0.015576981,
0.0139411706,
0.0278379209,
-0.004672077,
0.039626617,
0.0869990513,
-0.021836739,
0.0440203585,
0.0000017823,
-0.0596288256,
-0.0083111208,
-0.0144793242,
-0.0009209358,
0.0229773242,
0.0322964042,
-0.064501211,
-0.0266651921,
0.0278504174,
-0.0455705523,
0.0024804098,
0.0285141449,
0.0988599509,
0.0694617033,
0.1306880265,
-0.0183571018,
-0.0742407739,
0.0061521446,
-0.0559149198,
-0.0348181874,
-0.0402337275,
0.0238882378,
0.0091487784,
-0.0315396599,
-0.0813667849,
-0.0333697647,
-0.0037942531,
0.0133365579,
-0.00330341,
-0.0489736013,
-0.0061533968,
-0.095080249,
-0.0049696793,
0.0019518738,
0.0356289074,
-0.0040245159,
-0.0264080651,
-0.0144877015,
-0.0132303583,
0.0379253589,
0.0078924568,
-0.0045931409,
0.0486617237,
0.0074738329,
0.0365115665,
0.0418035276,
0.026444355,
-0.0037724236,
-0.0171333048,
-0.046919819,
0.0551587492,
0.0144236684,
-0.0458819792,
0.0156865939,
0.0203172006,
-0.0092240358,
-0.047701057,
0.0177198984,
0.0057435283,
-0.0285043605,
0.0135773662,
-0.0406733193,
-0.0229681004,
0.0281683374,
-0.0624801293,
0.0410576276,
0.015175147,
-0.0208602119,
0.0210489165,
-0.0341144167,
0.0234433208,
0.0280503854,
0.0457587838,
-0.0166195463,
-0.009978299,
-0.0243617352,
0.0137219029,
-0.0664511994,
0.0205610376,
-0.0827200189,
0.0245791506,
-0.0620943457,
-0.0264100805,
-0.019104844,
-0.0074374587,
0.0484099053,
-0.0205655135,
-0.0574325435,
-0.0170034058,
-0.0631070286,
-0.0070978873,
-0.0477095544,
0.0360564701,
0.0064129047,
-0.0430178642,
0.0310305953,
-0.0017464237,
0.0166896898,
-0.016645357,
0.0124498392,
0.0621698126,
-0.0148914885,
-0.0257429779,
-0.0472270474,
-0.029269062,
-0.0429678969,
-0.0726683885,
-0.0333693326,
-0.0704166293,
0.0029704168,
-0.0046885214,
-0.0071534682,
-0.0401726626,
0.0024034495,
0.0189377218,
-0.0396552198,
0.0599267185,
0.0418041907,
0.0006762297,
-0.0445126109,
-0.022832714,
0.0562986732,
0.0368087441,
0.0284053963,
0.0003173258,
0.0388421863,
-0.0100800619,
-0.0015939561,
0.0584115647,
0.0030827941,
0.0448000729,
-0.0138466833,
0.0044462294,
-0.0138764493,
0.0057074632,
-0.0284922905,
0.0406681262,
-0.0161892977,
-0.0069155865,
0.0342229716,
0.0189256351,
0.0403341874,
0.0562882684,
-0.0127470279,
-0.0082757669,
-0.0167197902,
-0.0019698539,
-0.0392978266,
-0.0013828279,
-0.0435693152,
0.0083040064,
-0.0182673801,
0.0458172932,
-0.032723479,
-0.0318442918,
0.0021121951,
0.0012525977,
-0.0040392005,
-0.0082592573,
0.0253390335,
0.0552959368,
-0.0338533819,
0.0653722435,
0.0376782753,
-0.0245995652,
-0.0713209286,
-0.0304585528,
-0.0315767489,
0.0429690555,
0.0325669125,
0.0149862561,
-0.0191968139,
-0.0395455398,
-0.0017522527,
0.034590207,
0.0801137611,
0.0013934902,
-0.0135551775,
-0.034839239,
-0.0178652126,
0.063845396,
0.0270744078,
0.0210859049,
0.0311140195,
0.0249974355,
0.0462306142,
-0.0173538085,
-0.0069611813,
0.0253178068,
0.0096977884,
-0.0020177858,
0.0150117483,
0.0227431916,
0.0687288269,
-0.0118978731,
-0.0053256834,
-0.0360404067,
0.0238103196,
-0.0168978311,
-0.0295572504,
0.0457233116,
-0.0000346341,
-0.0538803376,
0.0044842223,
-0.0614535362,
0.0256073643,
0.0228156745,
-0.0432744883,
-0.0425268412,
-0.0456642285,
0.020422861,
-0.0138295302,
-0.0475062579,
0.0353198349,
-0.0185412634,
0.0180020388,
-0.0331353657,
0.0217817724,
0.0199098978,
0.0027274694,
-0.0228411276,
-0.0290954579,
0.0324340947,
-0.0293684453,
-0.039759066,
0.0141800297,
-0.0224662777,
0.0124631561,
0.037587963,
0.0301972795,
-0.0233870409,
-0.0477409251,
-0.0539679155,
0.0190727301,
-0.0363596827,
0.0041646142,
0.0196768474,
-0.0021653834,
-0.0145993177,
-0.0119126076,
-0.0304763392,
-0.0011850762,
0.0349623486,
0.0128934504,
0.0231882371,
-0.0149738872,
-0.013309706,
-0.0025434899,
-0.027718816,
-0.0042228098,
0.0383071639,
0.0046880306,
-0.0443033651,
-0.0600247048,
0.0796416327,
0.0184501819,
-0.0039598714,
-0.015654942,
-0.0729428753,
-0.0234878417,
0.0371674336,
-0.0018745557,
0.0419699177,
0.0295262188,
0.0121621899,
-0.0184349101,
-0.0083917035,
-0.043401327,
0.0140943294,
-0.0187038835,
-0.0457840227,
-0.0722334534,
-0.0268284269,
-0.0127035854,
0.0574429482,
0.0255064536,
0.0049640564,
-0.0538702272,
-0.02410491,
-0.0087511381,
0.0142961387,
0.0369976871,
0.0276080538,
-0.0284079351,
0.0644656196,
0.0396392941,
-0.0162167139,
0.0613872148,
0.0267013032,
0.0389248356,
-0.0515232943,
0.0005637337,
-0.0571667068,
0.0566676073,
-0.0030207986,
-0.0120594762,
0.0088259866,
-0.0151367169,
-0.0262319855,
-0.0191534236,
0.0600558855,
-0.0018203263,
0.0586075634,
0.0203835722,
0.0102829551,
-0.0576835684,
-0.0670842752,
-0.0387173034,
-0.0907213837,
-0.0295749754,
0.0134748844,
0.0371732116,
0.0037761289,
-0.0549785607,
0.0233374219,
-0.0587319508,
0.0107739875,
-0.0044341683,
-0.0182602629,
-0.010231996,
-0.0063447668,
0.0312273875,
-0.0076088523,
0.0147168431,
-0.0401442349,
0.002854188,
0.0055535706,
-0.0138060413,
0.0077397069,
-0.0038979691,
-0.015976252,
0.0101174805,
-0.0406749174,
-0.0247082356,
0.017422365,
0.0534955114,
-0.0537462234,
0.0422771312,
0.0971077755,
-0.0052500078,
0.0220413562,
0.0116008716,
0.0203837808,
0.0490613617,
0.0434928909,
0.0332917161,
-0.0981193185,
-0.0099313213,
0.0140663711,
-0.0202674512,
-0.0401586629,
-0.0141535848,
0.0182186309,
-0.1049904451,
0.0649956837,
0.0205949917,
0.0175069682,
-0.0474245287,
0.024451267,
-0.0123370979,
-0.0357174017,
0.0115424283,
-0.0112744812,
-0.0229466166,
-0.0218996983,
-0.0011317183,
-0.0397344008,
-0.0125334719,
0.0398550257,
-0.0206932258,
0.022477245,
-0.0038746542,
0.0235921573,
-0.0120391129,
-0.0128855072,
0.0160492305,
-0.0163799226,
0.0062845033,
0.0159867425,
0.0859674066,
0.0191580728,
-0.0183731914,
-0.0140577182,
0.0267936308,
-0.0162601098,
0.0008576544,
0.0364568383,
0.0182111282,
-0.023940485,
-0.040601898,
0.0456692055,
0.001013938,
-0.0133912042,
0.0424029231,
0.0108387545,
-0.0288120825,
-0.066864118,
-0.0291225035,
-0.0093369819,
-0.0354367942,
-0.0017050131,
0.0183582902,
-0.0207211338,
-0.0111895353,
-0.0400857143,
-0.0407911949,
0.0079450924,
0.065615952,
0.0066513023,
-0.0329464823,
0.0007925293,
0.0134659465,
-0.0193909053,
0.0019780176,
-0.0409377925,
0.0242692959,
-0.0242356025,
0.0246870741,
0.0066085784,
0.0091459649,
-0.0128016295,
-0.0454696603,
-0.0368111655,
-0.050790824,
0.0109626744,
0.0062764613,
0.0253143944,
0.0418674871,
-0.0150920162,
-0.0261963867,
0.0223041214,
0.0506119542,
-0.0345249847,
0.0015691601,
0.0741200075,
-0.0393476188,
-0.0159077775,
0.0254118443,
-0.0122138383,
0.0460916683,
0.0401239879,
0.0202254709,
-0.0130216647,
-0.0803988948,
0.0127692884,
0.0689963251,
0.0380526073,
-0.0023307889,
-0.0306366775,
-0.0354812294,
0.0303989407,
-0.0116749983,
0.0003814801,
-0.0296596531,
-0.026280323,
0.0198859647,
0.0077156397,
-0.0380264074,
0.0384152085,
-0.0278485436,
-0.0189680476,
-0.048594553,
0.0681751296,
0.0166002885,
0.0128283268,
0.0360463671,
-0.0060662022,
0.0044986862,
-0.0019209652,
0.0510707386,
0.0387839153,
0.0023251651,
0.0509131961,
0.0183170363,
0.0362170748,
-0.0344881378,
-0.0024680684,
-0.0209594443,
-0.0056407959,
-0.0368394665,
0.0194502268,
0.051444713,
-0.0218098126,
0.0107197333,
-0.0594382472,
-0.0091174459,
0.0113248648,
0.0452663377,
-0.0186372809,
-0.0308323037,
0.0168248937,
0.0058087548,
-0.0325460173,
0.0156414639,
0.0645922199,
0.0193476826,
0.03740035,
0.0006458653,
-0.0253383573,
-0.041484464,
-0.0307035223,
-0.0048155985,
-0.0153006781,
-0.000958898,
-0.0652510971,
-0.0691421777,
-0.044114694,
-0.0140434448,
-0.0107172672,
0.0201999135,
-0.0115371263,
-0.0066156392,
0.0198941175,
0.0279279966,
0.0141587295,
-0.0389782786,
0.0130755166,
0.0407178737,
-0.0023923393,
-0.0084008994,
-5.687628314e-33,
-0.0115145119,
0.0078988383,
0.0009410226,
0.0571688265,
-0.0143888257,
-0.0387790725,
0.0049999682,
0.0019818421,
-0.0167329311,
0.0375085771,
-0.0529951453,
0.0130955698,
0.0331605487,
0.0177733731,
0.0363550261,
-0.0219773818,
0.0501589663,
0.0140237985,
-0.0281526912,
-0.0162070971,
-0.0140546495,
0.0319998264,
0.0693077073,
-0.0550939701,
-0.0247806422,
0.0261336565,
-0.0266890191,
-0.0230156705,
-0.0430779904,
0.0635258034,
0.0215976834,
0.0534314886,
0.020677099,
-0.082928583,
-0.0079699988,
-0.0053969263,
-0.0743638054,
-0.0139068663,
0.0151957832,
0.0510187931,
0.0287753455,
-0.0244218335,
0.0334948003,
-0.0109492149,
-0.0223097,
-0.0393358432,
0.0115617588,
-0.0469842665,
-0.0118020382,
-0.0666420981,
-0.0758412406,
0.070443511,
-0.012200335,
0.0004343583,
0.0395233184,
0.0032893126,
-0.0253787171,
0.0165339019,
-0.0281964038,
-0.0093623167,
0.0124215242,
0.03662939,
0.0118129542,
-0.0431101993,
-0.0535437129,
0.0377385989,
-0.0107289618,
0.0409157164,
0.0129364897,
-0.0341266058,
0.0327391662,
0.1074020937,
0.0113624483,
0.0960481539,
0.0567159653,
-0.0953725651,
-0.0754077807,
0.0056162667,
0.0128422426,
0.0302056801,
0.0589036271,
-0.013745917,
-0.0455802418,
-0.0381480418,
-0.0346084423,
-0.0260007046,
-0.0103977872,
0.0235241875,
0.0330361053,
0.0369466953,
0.0573768057,
0.0026014119,
0.0122766308,
-0.0633772165,
0.0185021069,
0.0194899477,
-0.0230038259,
0.0198167786,
-0.0433008522,
0.0047240732,
-0.0392407142,
-0.0436331853,
-0.0056478856,
0.0246645715,
0.0006911667,
-0.0168291107,
0.1094439328,
0.000659255,
-0.0820898935,
0.003395546,
0.0263691358,
0.0227692407,
0.013637742,
0.0373496972,
-0.0026438856,
0.0282107554,
0.0091394046,
-0.0119648855,
0.0161347352,
-0.0175416674,
-0.0282404944,
0.0003336744,
0.046641171,
0.0006920593,
-0.0015837378,
0.0110269738,
-0.0041479901,
0.0196588971,
0.032042671,
-0.0018109538,
-0.0045640068,
0.0380375385,
2.499e-7,
0.0603435263,
0.0529190041,
0.006377067,
-0.0046635759,
-0.011072848,
0.0525790825,
0.0188489053,
0.0369925238,
0.0424987413,
0.0161314476,
0.0560327396,
-0.0320142619,
0.0802017152,
0.0132508576,
-0.1038087979,
0.0034916664,
-0.0549177416,
-0.0397221632,
-0.0636637136,
-0.0006913891,
0.052051153,
0.1039533764,
0.0115245748,
-0.0311452299,
-0.0067213322,
0.0172873009,
0.0445002764,
-0.027845649,
-0.0354270451,
0.0221159216,
0.0680750683,
-0.0195023734,
0.0208688714,
0.0611292459,
-0.0262427274,
0.0697115213,
-0.0473006666,
0.0614460595,
-0.0097096199,
0.0239956826,
0.0112623833,
-0.0270316303,
-0.0461324863,
-0.0354661718,
0.046843227,
-0.0624049343,
-0.0337791219,
-0.0129451063,
-0.0299229473,
-0.0242165923,
0.0093308073,
-0.0018282373,
-0.0046729199,
0.0256126598,
0.0508872233,
-0.0061151776,
-0.021511212,
-0.0524643436,
-0.0092785154,
0.0685134903,
-0.0106687332,
-0.0027286666,
-0.0098692914,
-0.0182430595,
0.0516195893,
-0.0306788888,
0.0212271363,
2.339892448e-34,
0.0474399291,
0.0014461736,
-0.0076471795,
-0.0093856174,
0.0536500663,
-0.0053715967,
0.0147126028,
0.0539025441,
-0.0189527906,
-0.0911594182,
-0.0188905038
] |
http://arxiv.org/abs/2401.14640v1 | http://arxiv.org/pdf/2401.14640v1 | Benchmarking Large Language Models in Complex Question Answering Attribution using Knowledge Graphs | 2024-01-26T04:11:07 | 2024-01-26T04:11:07 | Nan Hu, Jiaoyan Chen, Yike Wu, Guilin Qi, Sheng Bi, Tongtong Wu, Jeff Z. Pan | cs.CL | cs.CL | The attribution of question answering is to provide citations for supporting
generated statements, and has attracted wide research attention. The current
methods for automatically evaluating the attribution, which are often based on
Large Language Models (LLMs), are still inadequate, particularly in recognizing
subtle differences between attributions, and complex relationships between
citations and statements. To compare these attribution evaluation methods and
develop new ones, we introduce a set of fine-grained categories (i.e.,
supportive, insufficient, contradictory and irrelevant) for measuring the
attribution, and develop a Complex Attributed Question Answering (CAQA)
benchmark by leveraging knowledge graphs (KGs) for automatically generating
attributions of different categories to question-answer pairs. Our analysis
reveals that existing evaluators perform poorly under fine-grained attribution
settings and exhibit weaknesses in complex citation-statement reasoning. Our
CAQA benchmark, validated with human annotations, emerges as a promising tool
for selecting and developing LLM attribution evaluators. | Benchmarking Large Language Models in Complex Question Answering Attribution using Knowledge Graphs: The attribution of question answering is to provide citations for supporting
generated statements, and has attracted wide research attention. The current
methods for automatically evaluating the attribution, which are often based on
Large Language Models (LLMs), are still inadequate, particularly in recognizing
subtle differences between attributions, and complex relationships between
citations and statements. To compare these attribution evaluation methods and
develop new ones, we introduce a set of fine-grained categories (i.e.,
supportive, insufficient, contradictory and irrelevant) for measuring the
attribution, and develop a Complex Attributed Question Answering (CAQA)
benchmark by leveraging knowledge graphs (KGs) for automatically generating
attributions of different categories to question-answer pairs. Our analysis
reveals that existing evaluators perform poorly under fine-grained attribution
settings and exhibit weaknesses in complex citation-statement reasoning. Our
CAQA benchmark, validated with human annotations, emerges as a promising tool
for selecting and developing LLM attribution evaluators. | [
0.0512949601,
-0.0111101009,
0.0086978534,
0.0563870035,
-0.0117933592,
0.0126356035,
0.0068926276,
-0.024508303,
0.0227498822,
-0.0617426783,
-0.0261584576,
0.0200486109,
-0.0238679331,
-0.064442277,
0.060037531,
0.0347861461,
-0.0247912034,
0.0061575719,
0.0672695115,
-0.0085705472,
0.0071969759,
0.0168698467,
0.0060934885,
0.0043724356,
0.0286727976,
-0.0346921794,
0.064986065,
-0.013701844,
0.0434655063,
-0.0176698416,
0.0204798114,
0.0300164558,
-0.0292907245,
0.0173128713,
0.0000014781,
-0.0274115112,
0.0456173047,
0.0190037638,
-0.0406185649,
0.0448064692,
0.0066745132,
0.0386361964,
0.0550058037,
0.0005922122,
-0.0218018908,
-0.0142650204,
-0.0098396307,
0.0121394992,
0.0797192231,
0.0179133471,
0.0057044257,
-0.0542264506,
0.0018429699,
-0.0038981147,
-0.0112180673,
-0.0434839763,
0.0291615352,
0.0674336329,
-0.0048341937,
0.0501202904,
0.0026104802,
0.0527716018,
0.0059611583,
-0.0110363429,
0.0646351576,
-0.0027564613,
-0.0048711109,
0.0383126885,
0.0301846359,
0.0831601471,
0.0310654733,
-0.0406440422,
0.003885794,
-0.0052991225,
0.011585719,
0.0676633939,
-0.0487030894,
0.0092366161,
-0.0161365401,
-0.0320945755,
0.0017142126,
-0.0164897311,
0.0404714569,
0.0381385311,
0.0294598117,
0.0715821236,
0.0130997021,
-0.0160034169,
0.0231129248,
0.0166946594,
-0.0518624485,
-0.0777557716,
-0.0147035103,
0.0449133106,
0.0260773059,
0.0117004132,
0.0191821251,
-0.0025821493,
0.0323004648,
-0.0406333916,
0.0269285403,
-0.0226073768,
0.0191532634,
0.0715044066,
-0.0570535921,
0.0750009641,
-0.040215604,
-0.0031246347,
-0.0402786992,
-0.0501021966,
-0.0677206516,
-0.0329975151,
-0.0079225879,
0.0229074694,
-0.0079496391,
-0.0154161919,
-0.0427809395,
0.0057654693,
-0.0430176631,
-0.0126017351,
0.0247380026,
0.0123009179,
0.0425224379,
0.0182964504,
-0.0546933822,
0.0453964882,
-0.0589872412,
0.0692314878,
-0.0065840245,
-0.0287857261,
-0.018255949,
0.0445072092,
0.0221576672,
-0.0615288168,
0.0146650858,
0.0286194142,
0.0163789671,
0.0476151221,
-0.0067889835,
-0.0687347203,
-0.0127213998,
-0.0486596115,
0.0086988304,
-0.0768110082,
0.0245570876,
0.0493487343,
-0.0187776908,
-0.0128535712,
0.0182372816,
0.0241868086,
-0.0551274419,
0.0131622348,
0.0053673605,
0.0216244962,
0.0324026868,
-0.0162960179,
0.0024862918,
0.0081417076,
0.0410448648,
0.003530036,
0.0151981842,
0.0196549278,
0.0342640728,
0.0076114936,
0.0399773195,
0.0289302506,
0.0422397852,
0.0405765958,
-0.0166954286,
-0.0050372556,
0.0359273851,
0.0508564003,
-0.0341912732,
-0.0099369632,
0.0531743504,
0.0369926952,
-0.0266472604,
0.0306752101,
0.0489393882,
-0.0076545924,
0.0417918824,
-0.0446185768,
-0.0110081462,
0.0340304822,
-0.0119538335,
-0.0131896846,
0.0024038807,
0.0014463793,
-0.0179410819,
0.0084991194,
-0.0074869609,
0.0147953741,
0.0034110199,
0.0172035024,
0.0063154334,
-0.0021740017,
-0.018242877,
0.0679075792,
-0.0274589118,
-0.0277812444,
0.0379843116,
-0.0340289026,
0.0082820123,
0.0332781561,
0.0109939817,
0.0103668049,
-0.0071762693,
-0.0247456934,
0.0166836269,
-0.0069530769,
0.0084154215,
-0.0439748801,
0.0032751588,
-0.0058209389,
-0.0070566386,
-0.0372416936,
0.0065003103,
-0.0287859533,
0.0252783895,
0.0164644923,
0.0114201661,
-0.0970752612,
-0.0307195261,
0.012977805,
0.0181394462,
0.0027864263,
-0.0033983469,
-0.0052533611,
0.0201755296,
0.042350933,
-0.0134582669,
0.0360618122,
0.0167927612,
-0.0390052944,
-0.032147482,
0.0361934453,
-0.0082874047,
-0.0299124867,
0.0224246327,
-0.0279647112,
-0.0207325183,
-0.0045729331,
-0.0435988419,
-0.0517467968,
0.0211568717,
-0.0475710817,
-0.0189869758,
-0.0131276064,
-0.0080786115,
-0.021951979,
-0.025493579,
-0.0436725244,
0.0929869339,
-0.0021558683,
0.0660800338,
0.0267953854,
-0.0066762292,
0.0396395773,
-0.0060464558,
-0.0050437227,
-0.0269368049,
-0.0088325599,
0.0078813899,
0.0432428941,
-0.0374481529,
-0.0467596687,
-0.0016403806,
-0.051496096,
0.0144982645,
-0.0250933263,
0.0577324256,
0.0069465134,
-0.0295302905,
0.0408154912,
0.0043752021,
-0.0526874326,
-0.0282927137,
-0.0132215619,
0.0025248572,
0.0274645127,
0.0371999927,
0.0373807475,
-0.027748825,
-0.0092428587,
-0.0339759588,
0.0485971346,
0.0513116866,
0.0456913076,
-0.0434456989,
0.0019744867,
-0.0427675173,
-0.0042516403,
-0.0039687091,
-0.0143435653,
-0.005332185,
0.0221212413,
-0.0027624171,
0.071708858,
-0.0236816742,
-0.0306337997,
0.0059840786,
-0.0000419577,
-0.0579961091,
-0.0074027628,
-0.0167382248,
-0.0345684253,
-0.0570310019,
-0.0621169955,
0.0064428365,
0.043560084,
0.0111713931,
-0.0066470392,
-0.0262609106,
0.0114373313,
-0.0194789153,
-0.0103207827,
-0.0209657568,
-0.019919578,
0.0092336256,
0.0161901377,
-0.037434835,
0.0116910674,
0.0358022526,
-0.0352942683,
0.0061189588,
-0.0009590448,
0.0073457961,
-0.0601645447,
0.0353964828,
-0.0174386501,
-0.0015227497,
-0.0146431224,
0.0246880967,
-0.0278583691,
-0.0315649062,
0.07004641,
-0.0320120826,
0.0145442486,
-0.0241977628,
0.0344701596,
0.0085559655,
-0.000784672,
-0.0096855294,
-0.0305823516,
-0.0394319445,
0.0458968915,
0.0246334448,
-0.0154853472,
0.007247916,
-0.0125252018,
0.0683302805,
-0.03646769,
-0.0239254553,
0.0113509037,
-0.0110444389,
0.087211065,
0.0509517379,
0.0289091468,
0.0431352593,
-0.0219378434,
-0.0234681126,
0.0118560307,
0.0224143472,
0.0235484689,
-0.0173215233,
-0.0327260531,
-0.0128801661,
-0.0272035599,
-0.0359226577,
-0.0279514659,
0.0931996107,
-0.0483633541,
-0.0019853115,
0.0822295472,
0.0140229557,
-0.018458508,
-0.0406732522,
-0.0281552877,
-0.0159351099,
0.0336355269,
0.0016795145,
-0.1098660752,
-0.0208171923,
-0.0556486361,
0.0144258291,
-0.0057292064,
-0.0280651245,
0.0125328396,
-0.1227260903,
-0.0012147189,
-0.0266473778,
-0.0111593036,
-0.0275139026,
-0.0533950366,
0.0277629793,
-0.0221954696,
-0.0715592355,
-0.0020704092,
-0.0023946925,
0.0143347783,
0.0442312099,
0.0122024613,
-0.0146902669,
0.0716806054,
-0.0622983947,
0.0398551673,
0.0063065961,
0.0692945346,
-0.0094571924,
-0.0097300122,
-0.0139095932,
-0.0796767846,
-0.0297564808,
0.0075651049,
-0.0086105829,
0.002898531,
0.0058920239,
-0.0349407829,
0.0188801177,
-0.0142725725,
-0.0054293685,
0.0242331736,
0.0855361223,
0.0137087107,
-0.037535388,
-0.0182353444,
-0.0039172038,
0.0098486152,
0.1005986631,
0.0116530675,
0.0204038527,
-0.0513899513,
0.0025125206,
-0.0009136673,
-0.0334401131,
-0.0401907638,
0.0577457845,
0.0120253759,
0.0149664916,
-0.071267955,
-0.0351100229,
-0.0057814559,
-0.0505938344,
0.0012425046,
0.0495722219,
0.0233026426,
-0.004226808,
-0.0451964736,
0.0067739333,
-0.064609319,
0.0478808731,
0.0186501667,
0.0141973114,
0.0186474696,
-0.0086974232,
-0.0310778227,
0.0440031439,
-0.0577865243,
-0.0111566391,
-0.0332082622,
0.0328977443,
0.0141241597,
0.0241678488,
-0.0465098396,
-0.0479776748,
0.0528298616,
0.0054121995,
-0.0179441478,
0.022933796,
0.015193833,
-0.0046093524,
-0.0158857703,
0.0281800255,
-0.0585528091,
0.0117384056,
0.0181067847,
0.0254465062,
-0.043408338,
-0.0545643531,
-0.0161368102,
0.0243807957,
-0.0279299989,
0.0189748872,
0.0141968746,
0.0073902165,
0.0041533438,
-0.0063130697,
-0.0066404063,
-0.0131093562,
-0.0051853973,
0.0239372812,
-0.0108515108,
-0.0061137807,
0.004200629,
0.0430275761,
0.0605964698,
-0.0232712943,
0.0117130326,
0.023579536,
0.0075404881,
-0.0330107026,
-0.0093198484,
-0.0399711393,
-0.0434556529,
0.0092145959,
0.0519585311,
0.005424093,
0.053365238,
0.0247866604,
-0.0315520242,
0.005236906,
-0.0278521627,
-0.0999438241,
-0.0065195551,
-0.0116479937,
-0.0369135588,
0.0346112847,
-0.0132160559,
0.0681235641,
0.0048073614,
-0.0368156694,
-0.0133066271,
0.0201037135,
0.0072836978,
0.0271389205,
0.0565984733,
0.0161923431,
-0.0294603929,
0.0293744355,
0.04143418,
0.0605738126,
0.0390152298,
-0.0273396485,
-0.064950645,
-0.0191158,
-0.0655910149,
-0.0454305261,
-0.0241340008,
0.0098976893,
0.0233667027,
-0.0744874254,
-0.0691624284,
0.0199486595,
-0.0017118853,
0.0153458212,
-0.0328637399,
0.0107119726,
0.0212985519,
0.0137720024,
-0.0386326239,
-0.0612416714,
0.0330529995,
-0.0585916229,
0.0022255569,
-0.0497241281,
-5.742042683e-33,
-0.0191988703,
-0.0321670659,
0.0028486196,
-0.0000727857,
-0.0396404415,
-0.0093305679,
-0.0548721291,
-0.0207311064,
-0.0733191818,
0.0572868995,
-0.0006046326,
0.0224029068,
-0.0087652504,
-0.0013424483,
0.024060851,
-0.0340807661,
0.0182052366,
0.0205197353,
-0.0478265397,
0.0400044546,
0.0230694041,
-0.0458330475,
0.0403518789,
-0.0970116407,
0.0709960014,
-0.0389879793,
-0.01102788,
0.021444153,
-0.0199600253,
0.069466494,
0.0069026412,
0.0233821757,
0.044010289,
-0.066597037,
0.0544034466,
0.0292800404,
-0.1132890955,
-0.0173234735,
0.0606502034,
-0.0324594416,
-0.059147086,
-0.0646316484,
0.0429077856,
0.0114762895,
0.0108596962,
-0.0272267088,
0.0784426704,
-0.0219622068,
0.0262055602,
-0.0797993243,
-0.016661264,
0.0171588883,
-0.0092038196,
0.0234897416,
-0.0363999307,
0.0185851753,
-0.0235862341,
0.0472915173,
-0.0334398411,
-0.0188345686,
-0.0205483623,
-0.0038524128,
0.0480116382,
0.0050061829,
0.022756217,
0.0373802446,
-0.0347897001,
-0.0008433699,
-0.0178258494,
-0.0410685167,
-0.0261995532,
0.0361106917,
0.1024054661,
0.097391881,
-0.0285180919,
0.032231655,
-0.0910957903,
-0.0070255422,
0.0214700568,
0.0483790673,
0.018345641,
-0.0231853742,
-0.0129346922,
-0.0138584822,
-0.095295839,
-0.0130954636,
-0.0526581444,
-0.0045828242,
0.0084786899,
-0.0223956965,
-0.0391563885,
0.0353839472,
0.037488956,
-0.0559291467,
-0.013493944,
0.0738737956,
-0.0134906787,
0.0417820103,
0.0222097598,
0.0316059403,
-0.0708369911,
0.02352852,
0.0150661571,
0.014429708,
0.0418794006,
0.0444929749,
0.0462587141,
-0.0293551255,
-0.0161126964,
-0.0094572846,
-0.0294532273,
-0.0251243711,
-0.0539960973,
-0.0358722992,
0.0050798953,
0.0848827511,
0.0122430818,
0.0105396798,
0.0462670624,
0.0487048924,
0.0100985039,
0.0067060264,
0.0180397686,
-0.0186577812,
-0.030606214,
0.0150005743,
-0.015458785,
0.03308497,
-0.059382569,
-0.027245447,
0.0030306617,
0.0027273607,
2.199e-7,
0.0732214898,
0.0076411744,
-0.0097515872,
0.0268308148,
0.0251678228,
0.0203660075,
0.0223880522,
0.0089233303,
0.0317498073,
-0.0208373405,
0.0012004691,
0.0351753272,
0.0374415666,
0.0167966727,
-0.0323821977,
0.0100383842,
-0.0656343177,
-0.0303214863,
0.0142122228,
-0.0019734891,
0.0251000486,
0.044975847,
0.0278516207,
0.0126029728,
-0.0558216311,
-0.0731858015,
0.0734371766,
0.0026179084,
0.0072828168,
0.0403390639,
0.0475527309,
0.0307799038,
0.0182281528,
0.0688474551,
0.0016469925,
0.0026807156,
0.0363379531,
0.0524663739,
0.0020241602,
-0.0414687619,
-0.0592321865,
-0.0803067312,
0.0614733398,
-0.0774556845,
0.0256730597,
-0.0424996875,
-0.0034885067,
-0.0148499683,
-0.052507814,
-0.0152539788,
-0.0164420158,
-0.0044004773,
0.0158453993,
-0.0020131462,
-0.0251441579,
0.0221667606,
0.0070550521,
-0.0280126818,
-0.0290271062,
0.0063229133,
-0.015769178,
-0.0631807074,
0.0100740967,
-0.0139966784,
0.001627643,
-0.0437167697,
0.0313661285,
1.315204265e-34,
0.0059848856,
0.0277970415,
0.0156471226,
0.015190199,
0.0513733067,
0.0112183765,
-0.0583298504,
0.0100322766,
0.0302692726,
-0.082942009,
0.012241181
] |
http://arxiv.org/abs/2401.14637v1 | http://arxiv.org/pdf/2401.14637v1 | T-Rex: Text-assisted Retrosynthesis Prediction | 2024-01-26T04:08:50 | 2024-01-26T04:08:50 | Yifeng Liu, Hanwen Xu, Tangqi Fang, Haocheng Xi, Zixuan Liu, Sheng Zhang, Hoifung Poon, Sheng Wang | cs.CL | cs.CL | As a fundamental task in computational chemistry, retrosynthesis prediction
aims to identify a set of reactants to synthesize a target molecule. Existing
template-free approaches only consider the graph structures of the target
molecule, which often cannot generalize well to rare reaction types and large
molecules. Here, we propose T-Rex, a text-assisted retrosynthesis prediction
approach that exploits pre-trained text language models, such as ChatGPT, to
assist the generation of reactants. T-Rex first exploits ChatGPT to generate a
description for the target molecule and rank candidate reaction centers based
both the description and the molecular graph. It then re-ranks these candidates
by querying the descriptions for each reactants and examines which group of
reactants can best synthesize the target molecule. We observed that T-Rex
substantially outperformed graph-based state-of-the-art approaches on two
datasets, indicating the effectiveness of considering text information. We
further found that T-Rex outperformed the variant that only use ChatGPT-based
description without the re-ranking step, demonstrate how our framework
outperformed a straightforward integration of ChatGPT and graph information.
Collectively, we show that text generated by pre-trained language models can
substantially improve retrosynthesis prediction, opening up new avenues for
exploiting ChatGPT to advance computational chemistry. And the codes can be
found at https://github.com/lauyikfung/T-Rex. | T-Rex: Text-assisted Retrosynthesis Prediction: As a fundamental task in computational chemistry, retrosynthesis prediction
aims to identify a set of reactants to synthesize a target molecule. Existing
template-free approaches only consider the graph structures of the target
molecule, which often cannot generalize well to rare reaction types and large
molecules. Here, we propose T-Rex, a text-assisted retrosynthesis prediction
approach that exploits pre-trained text language models, such as ChatGPT, to
assist the generation of reactants. T-Rex first exploits ChatGPT to generate a
description for the target molecule and rank candidate reaction centers based
both the description and the molecular graph. It then re-ranks these candidates
by querying the descriptions for each reactants and examines which group of
reactants can best synthesize the target molecule. We observed that T-Rex
substantially outperformed graph-based state-of-the-art approaches on two
datasets, indicating the effectiveness of considering text information. We
further found that T-Rex outperformed the variant that only use ChatGPT-based
description without the re-ranking step, demonstrate how our framework
outperformed a straightforward integration of ChatGPT and graph information.
Collectively, we show that text generated by pre-trained language models can
substantially improve retrosynthesis prediction, opening up new avenues for
exploiting ChatGPT to advance computational chemistry. And the codes can be
found at https://github.com/lauyikfung/T-Rex. | [
0.050845556,
0.0848415345,
0.0043107062,
0.0072067562,
-0.0313839614,
-0.0019411499,
0.0144687425,
0.0545682795,
0.0512644574,
-0.0445311815,
-0.0122443708,
-0.0407353528,
-0.0163660552,
0.0632463694,
0.0257256646,
0.0101437485,
0.0117851198,
-0.0161006209,
-0.0364077277,
-0.0104247313,
0.0139130009,
0.0384531319,
0.0182159115,
0.0083250375,
0.0020738468,
-0.014569792,
-0.0056709494,
0.0169993229,
0.0499356613,
-0.0193500202,
0.0109290853,
0.043118719,
0.0248646457,
0.0638723001,
0.0000016134,
-0.0580587611,
-0.0063177985,
0.0046183011,
0.0060108728,
0.0080038011,
0.0438224189,
-0.0009320934,
0.0024431259,
-0.005967421,
-0.0315158181,
0.0213942807,
0.0519370399,
-0.0025563871,
0.0252746064,
0.053832937,
-0.0083137508,
-0.0056733824,
0.0437582918,
0.0021350365,
0.0806366578,
0.015775485,
0.0410218835,
-0.0017676839,
0.0160869937,
0.020658996,
0.0030344792,
0.0162357427,
-0.0349990577,
-0.0338153653,
0.0049590901,
0.0231914017,
-0.0154402051,
-0.013751871,
-0.0369801745,
0.0818406269,
-0.0103675164,
-0.076313071,
-0.0563571118,
-0.0029789004,
-0.0601144955,
0.0220859982,
-0.0453953408,
-0.0220073555,
-0.0032293729,
0.0145238182,
0.0084044663,
-0.0306629725,
0.0333117656,
0.0078608235,
0.0570360534,
0.0332825668,
0.0036712731,
-0.0303612798,
-0.0132721895,
0.0140565624,
-0.0441588126,
-0.0174743067,
0.0182524696,
-0.0053992933,
0.0226994734,
0.0066364515,
0.0298328884,
-0.0366432406,
0.0069634048,
-0.0497041717,
0.0022104187,
0.0163612869,
0.0248908624,
-0.0159695167,
-0.0593436845,
0.0371336043,
-0.0209369306,
-0.0295009371,
-0.0550208241,
0.0181214716,
-0.0597857572,
-0.0335042998,
-0.034068048,
0.0100169014,
-0.0218139905,
-0.0134157008,
-0.0420647524,
0.0476296805,
0.0076212767,
-0.0427018441,
0.01653447,
0.0541403703,
-0.031146979,
0.0164148081,
-0.0158508308,
0.0048438245,
0.0423719101,
-0.0081317546,
-0.0160817523,
-0.0372112803,
-0.0332405791,
-0.0067932708,
0.0001769903,
-0.0444663279,
-0.0005272375,
0.0621637255,
-0.0283487495,
0.0328185894,
0.0235385094,
-0.0605215617,
0.0136102559,
-0.0410435498,
0.0312157162,
-0.0564164408,
0.0537646413,
0.047071036,
-0.0027641177,
-0.0471917242,
0.0428246707,
0.0019202468,
-0.0420657992,
0.1125089601,
0.0096234623,
-0.0529144518,
0.0237510782,
0.003780571,
-0.0086987549,
-0.0091151111,
0.0070221331,
0.0170874093,
-0.0000722393,
0.0158143546,
0.0244983379,
-0.0071072858,
0.0096995197,
-0.0114999916,
-0.0165126268,
-0.0119648725,
0.001015098,
0.0635659769,
-0.0224379562,
-0.0018899586,
-0.0045351973,
-0.0422118604,
0.0094418237,
0.0288219042,
0.0361289755,
0.0571890175,
0.0466492437,
-0.0091188978,
-0.028058514,
-0.0207381453,
-0.0265178308,
0.0350881778,
-0.0665078163,
-0.0257580467,
0.0118393423,
0.0123564582,
0.0133811841,
0.0139969084,
0.0161764491,
0.0104835145,
0.0050002933,
-0.0756535977,
0.0277446266,
-0.0273011811,
-0.1233057231,
0.0150467874,
0.0294276737,
-0.0402360968,
0.000628926,
-0.0336643793,
-0.0300567895,
0.0901487991,
0.0836445168,
-0.0409885719,
-0.0395694189,
0.0484746993,
-0.0451681055,
0.0178567618,
0.034115307,
-0.0065783341,
0.0034664152,
-0.0030457205,
-0.0252142418,
-0.0170041062,
-0.004037308,
-0.0169910993,
0.012779437,
0.0029194988,
0.0229546148,
-0.0353778005,
-0.023403544,
0.0144302743,
0.0143786278,
-0.050845515,
0.0213114899,
-0.0431328416,
0.0237423107,
0.0086651351,
0.0045675538,
0.0457259528,
0.0479608104,
-0.0359995253,
-0.0413317159,
-0.0162239224,
-0.0083435522,
-0.097735472,
0.0304611921,
-0.0265349746,
0.0280993264,
0.0199163463,
-0.0358039513,
-0.0268341843,
0.0898844823,
0.0191434883,
0.0361415371,
-0.0313148201,
-0.004583844,
0.0100084711,
0.0284529664,
-0.0191998761,
0.0671105161,
0.0363927037,
0.0142783308,
-0.0126494663,
0.0079026232,
0.0295436624,
-0.0062827421,
0.0319451727,
-0.0331479125,
0.0458209962,
0.0238243323,
0.067263782,
-0.0041297218,
0.0031658581,
0.0098398319,
-0.0232918337,
-0.0313828513,
-0.005449513,
0.037000753,
0.032932274,
-0.0183276609,
0.0206777193,
-0.0072466824,
0.0018938365,
-0.0005518462,
-0.0354520492,
0.0170994774,
0.0275070015,
0.0055726999,
-0.0430673212,
-0.0528290197,
-0.0419702083,
-0.0632022098,
0.090769738,
0.0364927165,
-0.0651353598,
0.0144655667,
0.0161406342,
-0.0092992159,
0.0380306542,
0.0029264588,
-0.0021199214,
0.0200565644,
-0.0212806258,
0.0309598632,
-0.0140185207,
-0.0097622098,
0.0129819736,
0.0521397628,
0.0309707429,
-0.0228857584,
-0.0207948461,
-0.0536617339,
-0.0152679952,
0.0823178142,
-0.033072494,
0.0001964974,
0.0498200282,
-0.0038896678,
0.0019826454,
-0.0379881039,
-0.0100198565,
0.0176927373,
-0.014127573,
-0.0620275624,
-0.0213135369,
0.021984579,
-0.0036709385,
0.003520817,
-0.085451141,
0.0599123165,
-0.013767614,
0.0143136485,
-0.0528431311,
0.037598595,
0.0346305519,
0.039987132,
0.0473056883,
0.0435859077,
0.0138316182,
0.0082382597,
-0.05409199,
-0.0061002211,
0.0472594947,
-0.0094172191,
-0.0141350143,
-0.0360042676,
0.0383735597,
0.0211505722,
-0.0330959186,
-0.0112825744,
-0.0372340418,
0.0107692359,
0.0227462277,
0.0238756426,
-0.0292735975,
-0.01583983,
0.0661358833,
-0.0008481009,
-0.0004457168,
-0.0358841158,
-0.012407844,
-0.0113895005,
0.0108513217,
-0.0065461835,
-0.1035256311,
-0.0242533367,
-0.0192465689,
0.0386951119,
0.0607492775,
0.0438547656,
-0.0253411084,
-0.0334685892,
0.0117070256,
0.0025209216,
0.0173692536,
-0.0187816899,
-0.0295035392,
-0.0420490466,
-0.0319538116,
-0.0096216649,
0.0557152852,
0.0475195758,
-0.0247908626,
0.0274404604,
0.038861651,
0.0044634477,
-0.0244839806,
0.0373837836,
-0.0719361454,
-0.0114534833,
-0.0496740043,
0.0118727218,
-0.0233055186,
-0.012859066,
-0.0261236355,
-0.0643631965,
0.0336865224,
0.0266910829,
-0.0543982908,
0.0062697367,
-0.0590291433,
-0.0455690548,
-0.0158210769,
-0.0497202054,
0.0472628884,
0.0025488536,
-0.0248098914,
0.0039194883,
-0.0396545269,
-0.0264778212,
0.0108261751,
0.0117986826,
-0.0516480654,
0.0399695784,
0.0666894019,
0.0357460901,
0.0263551436,
0.0351019911,
0.0041944166,
0.0495159067,
-0.0441726036,
0.0356140956,
0.0360089093,
0.0003416976,
-0.0041058222,
0.0259203501,
-0.0521448962,
-0.0719332993,
-0.0640667751,
0.1113020182,
-0.056224335,
-0.0055664405,
-0.0458337888,
0.0222321451,
0.0030533811,
0.0191297587,
-0.0075090602,
-0.0177257229,
-0.1012273952,
-0.0431679003,
0.0397605449,
-0.0645410269,
-0.0048923474,
-0.0648150593,
-0.0079626897,
-0.0348463282,
-0.0476940386,
0.0193292629,
0.0110294828,
0.0008959982,
-0.0223532617,
-0.0161463302,
-0.0211749859,
-0.056395039,
-0.059667699,
0.0195879936,
0.0241189823,
-0.0013315338,
-0.0153840734,
-0.0738896653,
0.0367081016,
-0.0301218517,
-0.0295748543,
0.0420505628,
-0.0540569164,
-0.0616786033,
-0.0365418904,
0.0265937559,
0.0205892567,
0.033566609,
-0.005331221,
-0.0151329907,
0.05971637,
0.0868006721,
0.024765458,
-0.0210145488,
-0.0181963984,
-0.0236147866,
-0.0250640921,
-0.0038919896,
-0.0215053707,
0.0240548812,
0.0198404808,
0.0059023355,
0.002994173,
-0.0074681896,
0.0610245913,
0.0538439602,
0.0668009445,
-0.0122959735,
0.0029587767,
-0.0251508467,
-0.0519368686,
0.0054372549,
0.004274406,
-0.0398457199,
-0.0078809075,
0.0157480761,
-0.0061932532,
-0.0067316871,
-0.0106048649,
-0.0295669623,
0.0085470751,
-0.0411483981,
0.000847632,
0.0361567587,
0.0335267261,
0.095652245,
-0.0190329999,
0.0068254676,
-0.0047193384,
-0.0181103833,
0.0096754646,
-0.0334328339,
-0.0391095281,
0.0092313951,
0.0554341264,
-0.0309839137,
0.0101406332,
-0.0325576514,
-0.0189091712,
-0.1094047353,
0.0057263435,
0.0579946302,
0.0170967989,
0.0417484902,
-0.0221987143,
-0.025666289,
-0.0130748497,
0.0290791728,
0.0241820943,
-0.0255250707,
0.0203939434,
0.0004358955,
-0.0315927379,
0.0129498309,
0.0994709432,
0.0132620931,
0.0626182631,
-0.0034057554,
-0.0520761572,
-0.0101864692,
-0.0192492027,
-0.0341553167,
-0.0247062799,
0.0297280047,
0.0083158081,
-0.0400754958,
-0.0324558541,
0.0369000472,
0.015787784,
-0.0235484596,
0.0388399959,
0.056541577,
0.0392305329,
0.0174022187,
-0.0422865637,
-0.0235967077,
0.017356297,
-0.0049750456,
-0.0296217632,
-0.0496540554,
-5.224620926e-33,
0.0318478346,
-0.0197893325,
-0.0150907738,
0.0305163469,
0.003727407,
-0.0121018784,
-0.1037862226,
-0.0484662466,
0.0688390136,
0.0521240011,
0.0317080617,
-0.0125330221,
-0.0002936999,
-0.0233173557,
-0.0145353032,
0.0375357531,
0.0096106706,
0.0050853579,
-0.0153941168,
0.0590450615,
0.0938848034,
0.0684615374,
0.0708526522,
-0.0246152841,
-0.0296149943,
0.0257667582,
-0.0229950976,
-0.0001966409,
-0.0482955091,
0.0667929873,
-0.0067545162,
0.0732939094,
-0.0300559383,
-0.019888822,
-0.0199573785,
-0.0137995714,
0.0053485413,
-0.0044960738,
0.008775725,
-0.0093144309,
-0.0015370502,
0.0075391717,
0.0203821193,
0.0084711127,
-0.0424798094,
0.000933129,
0.0300840717,
0.0143562695,
-0.0076938495,
-0.0239000246,
-0.066361554,
0.0129889846,
0.0043384563,
-0.0089458507,
-0.0329510421,
0.0386981964,
0.0082539693,
0.0164472889,
-0.0169197265,
0.0851716772,
0.01582231,
0.0254207999,
0.0697150379,
0.0337814093,
0.0052383002,
0.0262688249,
-0.0151332486,
0.0347085893,
0.0181323905,
0.0175646003,
-0.0034121391,
0.0487317294,
0.0399877131,
0.01004308,
0.032574147,
0.0048783435,
-0.0981939882,
0.0283468068,
-0.0184456818,
-0.0002434165,
0.0228253007,
0.0562941954,
0.0221259128,
-0.0660185665,
-0.0234407075,
0.0424948744,
-0.0295239873,
-0.0071487972,
-0.0213032383,
-0.0358968303,
-0.0522496067,
0.0210834686,
0.0036120124,
-0.0443330146,
-0.0346832648,
0.0622184537,
-0.0617693216,
0.0316564031,
-0.0075413878,
0.0128533412,
-0.0258361027,
-0.0541586541,
0.0357512236,
-0.039842438,
0.0175332837,
-0.0309035853,
-0.0188746322,
0.0641303882,
-0.0298815984,
-0.0066528902,
-0.0911070034,
-0.0027814868,
0.0359551199,
-0.0125213182,
-0.0152627509,
0.0073698787,
0.0039248583,
0.0395277664,
0.0138426377,
0.009858666,
0.0394006521,
0.006255215,
0.0073984652,
0.0465063676,
-0.0758557096,
-0.0054135504,
-0.011131363,
-0.0030636014,
0.043496687,
-0.0821838677,
-0.0391327105,
0.015443786,
2.569e-7,
0.0725038126,
0.032229878,
0.0419333763,
-0.044425562,
0.0103644822,
0.0235598572,
0.0285053775,
0.0366047844,
-0.0338346101,
0.0206289832,
0.0493820868,
-0.0226983875,
0.0359926857,
0.0054045138,
0.0055726725,
0.0583360493,
-0.1085667387,
-0.0320600495,
-0.0559846424,
-0.0234652851,
0.0054923743,
0.0509206206,
0.0066659222,
-0.0058485484,
-0.0066798758,
-0.0204310771,
0.0620937683,
-0.0346785896,
-0.0198691282,
0.0376498811,
-0.0256602503,
-0.051417239,
-0.0119895861,
0.0366248339,
-0.0519088581,
-0.0227228217,
0.0267427284,
0.0164543521,
-0.0754287243,
-0.0182264186,
0.0304328613,
-0.0090648979,
0.0240477528,
-0.0101969419,
-0.0037832407,
-0.0339498781,
0.0031522876,
-0.0054023042,
-0.060525436,
0.013833127,
0.0122069791,
-0.0120438505,
0.0202017166,
0.0555432811,
-0.0187799968,
0.0419675596,
-0.0269844774,
-0.0225325543,
-0.0634474009,
0.0234935917,
-0.0196186677,
0.0047745742,
-0.0287184641,
0.0129643502,
0.0162843186,
0.0114065958,
0.0159188043,
2.437004582e-34,
0.0698749647,
-0.0152499219,
-0.0245062988,
-0.0026569993,
0.0609224923,
0.0029056156,
0.0613062605,
0.0217888188,
0.0007084942,
-0.0408962592,
-0.0341741852
] |
http://arxiv.org/abs/2401.14630v1 | http://arxiv.org/pdf/2401.14630v1 | An Empirical Investigation of Domain Adaptation Ability for Chinese Spelling Check Models | 2024-01-26T03:49:55 | 2024-01-26T03:49:55 | Xi Wang, Ruoqing Zhao, Hongliang Dai, Piji Li | cs.CL | cs.CL, cs.AI | Chinese Spelling Check (CSC) is a meaningful task in the area of Natural
Language Processing (NLP) which aims at detecting spelling errors in Chinese
texts and then correcting these errors. However, CSC models are based on
pretrained language models, which are trained on a general corpus.
Consequently, their performance may drop when confronted with downstream tasks
involving domain-specific terms. In this paper, we conduct a thorough
evaluation about the domain adaption ability of various typical CSC models by
building three new datasets encompassing rich domain-specific terms from the
financial, medical, and legal domains. Then we conduct empirical investigations
in the corresponding domain-specific test datasets to ascertain the
cross-domain adaptation ability of several typical CSC models. We also test the
performance of the popular large language model ChatGPT. As shown in our
experiments, the performances of the CSC models drop significantly in the new
domains. | An Empirical Investigation of Domain Adaptation Ability for Chinese Spelling Check Models: Chinese Spelling Check (CSC) is a meaningful task in the area of Natural
Language Processing (NLP) which aims at detecting spelling errors in Chinese
texts and then correcting these errors. However, CSC models are based on
pretrained language models, which are trained on a general corpus.
Consequently, their performance may drop when confronted with downstream tasks
involving domain-specific terms. In this paper, we conduct a thorough
evaluation about the domain adaption ability of various typical CSC models by
building three new datasets encompassing rich domain-specific terms from the
financial, medical, and legal domains. Then we conduct empirical investigations
in the corresponding domain-specific test datasets to ascertain the
cross-domain adaptation ability of several typical CSC models. We also test the
performance of the popular large language model ChatGPT. As shown in our
experiments, the performances of the CSC models drop significantly in the new
domains. | [
0.0388706103,
-0.0198559724,
-0.0100342194,
0.000946953,
-0.0224636775,
0.0235515907,
0.0407253467,
0.0442305915,
0.0338528454,
-0.0709953457,
-0.0661701411,
-0.0622069575,
0.0057200589,
0.0200571921,
0.0022764679,
-0.0353721641,
0.0567858964,
0.0021365108,
-0.0244139414,
-0.0008929177,
-0.0069567631,
-0.0163357779,
-0.0019343465,
0.0287544038,
0.0347229391,
-0.0108523304,
0.0392897166,
-0.0117052514,
0.0148506369,
-0.0220756177,
0.0407082103,
0.0540960133,
-0.0080982745,
0.0373083726,
0.0000018844,
-0.0748542547,
-0.0606361665,
0.0133328093,
-0.0060324972,
0.0331009291,
0.0470981225,
0.0024828892,
-0.0380421206,
0.0435420834,
-0.0387253799,
-0.0155268209,
0.0348323956,
0.0766594261,
0.0536052026,
0.1012162119,
-0.0267244801,
-0.0128858397,
-0.0018002397,
-0.0375259519,
-0.0321077146,
-0.0851401165,
0.0669511259,
0.0256432798,
-0.00454186,
-0.0327505432,
-0.0573392995,
-0.0218693595,
-0.0408003926,
0.0087676411,
-0.0265003443,
-0.0205012411,
0.0010018217,
0.0239990838,
-0.0050680405,
0.0258786026,
0.0112324385,
-0.0141667351,
0.0061575952,
-0.0009163186,
0.0078017116,
0.0484117195,
-0.005476221,
-0.0152067952,
-0.0083633792,
-0.0215234142,
0.048229754,
-0.0226898678,
0.0152864279,
0.0189220905,
-0.0138074458,
0.0269062873,
0.0281913448,
-0.0673473626,
0.0168890469,
0.0277225859,
-0.0010663719,
-0.0355159976,
0.0304685719,
-0.009992633,
-0.0091275079,
0.0001533597,
-0.0545981899,
-0.0365844704,
-0.0156667605,
-0.0434337035,
0.0238004122,
0.0102462303,
-0.0279193316,
0.0298132114,
-0.0410840735,
0.0359643549,
0.0147023927,
0.0061651566,
-0.0081996433,
0.0055100317,
-0.0847325549,
0.0280621536,
-0.0350848734,
0.0329989307,
-0.0101123406,
0.0043343427,
-0.1016280055,
0.0028027582,
-0.0156509392,
-0.0377214849,
0.0416327603,
-0.0613042079,
-0.037042737,
-0.0289555043,
-0.0165027846,
0.0148385586,
0.0064937724,
0.0427636653,
0.0058012241,
-0.0245907921,
-0.002396734,
-0.0106515959,
0.0232325476,
-0.0402013175,
0.0222607497,
0.0656880736,
0.018241873,
-0.0523697957,
-0.0530506372,
0.0338397995,
-0.0379020311,
-0.0064567481,
0.0009200262,
-0.0288221091,
-0.0418958329,
-0.022256393,
0.046719335,
-0.0535998344,
-0.0209561288,
-0.0297302585,
-0.0490221754,
0.0863355845,
-0.0267204233,
-0.0177250169,
0.0040730326,
-0.0189439151,
0.0044305301,
0.0299887657,
-0.0109320059,
-0.0575904883,
0.0544778518,
-0.0176813584,
0.0384808406,
0.036190331,
-0.0370165892,
0.007476395,
0.0220240615,
-0.0033875818,
-0.0096621569,
-0.002767354,
0.0252125524,
0.0104311528,
-0.0111188963,
0.025739599,
0.0487878136,
0.0250810478,
0.0732171983,
-0.0074350717,
-0.007935619,
-0.0426551178,
0.0178404637,
0.0115949754,
-0.0454574078,
-0.0358554237,
-0.0674176514,
0.0232811198,
0.0037115996,
-0.022261072,
-0.0403089523,
-0.0398183614,
-0.0344558209,
0.0246754698,
0.012272344,
-0.0316054299,
0.0693347305,
0.0322168656,
-0.0211086869,
0.03486358,
0.0170270726,
-0.041526597,
-0.0083230594,
-0.0636117384,
0.013597928,
0.0550139435,
0.0614538454,
-0.0255223773,
-0.0308693089,
-0.0076936008,
-0.0147605799,
0.0500257276,
0.075474903,
0.004803936,
-0.0352151059,
-0.0310922712,
-0.0319495685,
0.0589277558,
0.040883787,
0.010651269,
0.0327058397,
0.0358988121,
0.006893456,
-0.0579267666,
-0.0464685373,
0.0307592805,
-0.0449282899,
-0.0257824492,
0.0300168078,
-0.0162883457,
0.0104844961,
0.0040341229,
-0.0617048405,
-0.0139407916,
-0.0301506463,
-0.0187234487,
-0.0500683598,
-0.0181973539,
0.0014783052,
-0.0169345811,
0.0447176509,
-0.0616257787,
0.0111620277,
0.0479071662,
-0.0382365137,
-0.0322140157,
-0.0049597668,
-0.0026238041,
-0.0139480112,
-0.0550284758,
0.0421086363,
-0.0161682647,
0.0344345793,
-0.0080627622,
0.0320124179,
0.0246870536,
0.0018160259,
-0.0530269593,
0.051464498,
-0.0039012078,
-0.0019795031,
-0.0392291024,
-0.0255682152,
-0.0371780023,
0.0484415069,
0.0396904647,
0.0358143933,
-0.0149582131,
0.0015298534,
-0.0381400697,
-0.0023272,
-0.0431291871,
0.0337175913,
0.0268783942,
0.0148334606,
-0.0445538126,
-0.0073631615,
-0.0141896363,
0.0410238057,
0.0059317742,
0.0268863197,
0.0474365763,
-0.0376107655,
-0.001249401,
-0.0566320717,
-0.0026413947,
-0.0124874618,
-0.0000482554,
-0.0131936427,
-0.0809399113,
-0.0606889948,
0.1032621562,
0.0162611436,
0.0265648477,
0.0217215177,
-0.0295455717,
-0.0359687209,
0.0040796152,
0.010623063,
0.0274281073,
0.0098753953,
0.0174203832,
0.0196969341,
0.0122859906,
0.0033325078,
0.009010843,
0.0182340294,
-0.0287569035,
-0.0190596897,
-0.0201388095,
0.0403962173,
0.0843672454,
-0.0222658366,
-0.0373396873,
-0.0314304531,
-0.0141734425,
-0.0544290207,
0.0120556941,
-0.0237834603,
0.0348105207,
-0.0008399436,
0.0319403224,
0.0095031438,
-0.0103390617,
0.0372440219,
0.0034558384,
0.0526036918,
-0.0675142333,
-0.0191734955,
-0.028958559,
-0.0064353058,
-0.021525912,
0.0031451436,
0.0254550055,
-0.0306041222,
-0.0270452257,
-0.0018179823,
0.0633332506,
-0.0286566131,
0.0483671725,
-0.0374902003,
0.0500778705,
-0.0148007637,
-0.0099946717,
-0.0173505489,
0.0010672867,
-0.0206835587,
0.0141212698,
0.0400675572,
-0.0079685133,
-0.0245707948,
0.0578943156,
-0.0295394156,
0.0794139877,
0.0145187862,
-0.0334679298,
0.0260327924,
-0.0090032676,
0.0395250171,
0.04296498,
0.0381145664,
-0.0133411549,
0.0379650816,
0.0112724584,
-0.0047343397,
-0.0067314189,
-0.0045199199,
-0.0012815365,
0.0288775116,
-0.0768608227,
-0.0117480895,
0.0142458351,
0.0242208,
-0.0324293338,
-0.0368461013,
0.0904548317,
0.0205322951,
-0.0207918957,
-0.0096876156,
0.0057816911,
-0.0344372988,
0.0081360461,
0.029465571,
-0.075241223,
-0.0485273153,
0.0328749269,
-0.0195210967,
0.0026643714,
-0.0370817371,
0.0087247426,
-0.1325575113,
0.0494134501,
-0.0161542837,
-0.01288985,
-0.0629496798,
-0.0205272231,
0.0238086898,
-0.0503514893,
0.0464840233,
-0.0104733147,
-0.0251190495,
-0.0455855317,
0.0342285149,
-0.0512774028,
-0.0820697621,
0.02466188,
-0.0401899032,
0.0304345246,
0.0795976147,
-0.0057946644,
-0.0288849659,
-0.0105578303,
0.0411357917,
0.0161771867,
-0.0247097742,
-0.0003073444,
0.037836127,
0.0220696181,
-0.0059684655,
0.0370692499,
0.0078384969,
-0.0543472432,
0.0135963736,
0.0353241786,
0.0056124302,
-0.0154547812,
-0.0254755337,
-0.0123215737,
0.0208102241,
-0.028047815,
0.074474968,
-0.0086264908,
-0.0762962624,
-0.0180620812,
-0.0309441406,
-0.0559281446,
-0.0262691099,
-0.0284561776,
0.0491324142,
-0.0887453407,
-0.0320395157,
-0.0895954221,
-0.0156630762,
-0.0035422978,
0.0309738629,
0.0274065044,
0.0104900254,
0.0355096236,
0.0336767696,
0.0004948128,
-0.0105680833,
-0.0456302948,
0.015342012,
-0.0802869573,
0.042887751,
-0.0071988571,
-0.0243851859,
-0.0366895311,
0.0001060092,
-0.0194389019,
-0.0229164567,
0.0534045137,
-0.0115614906,
0.0136675425,
0.0063534216,
-0.0461910106,
-0.0089818025,
0.0444539897,
0.0296460781,
0.000936501,
0.016299855,
0.0399225801,
-0.0665617213,
-0.0306163039,
-0.0012142038,
-0.0469313636,
0.0614535399,
0.0425526761,
0.0422430634,
0.0275875535,
-0.0190474279,
0.0224107429,
0.0337437354,
0.0919667333,
0.003507416,
-0.0232731961,
-0.0494266227,
0.0061619929,
0.0222694799,
-0.0154963024,
-0.001380494,
-0.0389224738,
0.049401328,
-0.0157983992,
-0.0241314583,
0.0305110924,
-0.0411911383,
-0.0247519892,
-0.0310142953,
0.0018572661,
0.0111209666,
0.0491888858,
0.0026804477,
-0.0125589482,
0.0092993798,
-0.0151914516,
0.060045585,
0.0016041347,
0.0025308863,
0.0262084156,
-0.007599548,
0.0452434495,
0.0300062168,
0.0137574207,
-0.0604270287,
0.0427312516,
-0.0386384316,
0.019870786,
0.0652777106,
-0.0452963971,
0.0373181626,
-0.0430563726,
0.0092261974,
-0.0601276569,
0.013393308,
0.0273908488,
0.035261374,
0.021822717,
-0.0144367293,
-0.0168677699,
-0.0204888582,
0.072769016,
0.0368483365,
0.0323961005,
0.0053932322,
-0.0460356437,
-0.0143340137,
-0.0214287937,
0.0007516665,
-0.0299559236,
-0.0282298401,
-0.0028352016,
-0.0542744249,
-0.0734450668,
-0.0139660714,
-0.011504963,
0.0318673328,
0.0262987241,
0.0118766818,
0.0706908107,
0.0158779044,
0.0467265211,
-0.0346909389,
-0.0303840488,
0.0106429644,
-0.0134121766,
0.011629506,
-5.877178983e-33,
-0.04111626,
0.0211793445,
0.0140170828,
-0.0091830324,
0.0174180418,
-0.007900795,
-0.0025323194,
-0.0076276008,
-0.0133857308,
0.0278839115,
-0.044326134,
0.0510515422,
0.0157719422,
0.015288515,
0.0006598582,
-0.0015631681,
0.0533105619,
0.0359090194,
-0.0122308554,
-0.0492362306,
0.0160688646,
0.0574535243,
0.0856322572,
-0.0527284816,
0.0375565104,
0.0447742939,
-0.0056670993,
-0.0276892558,
0.039182812,
0.0685954392,
-0.0004467526,
0.0501575954,
0.0326730013,
-0.0635172725,
-0.0174837727,
-0.0225690734,
-0.0900625736,
-0.0080701895,
0.0052288887,
0.0411981232,
0.0245451387,
-0.011975646,
0.0297626648,
0.0221930966,
-0.0408487692,
-0.0250416584,
0.0248181578,
-0.0574246868,
0.00281777,
-0.0162095986,
-0.0621037669,
0.0479551703,
0.0406094007,
0.0108905388,
-0.0089916782,
0.0535022877,
-0.0298826247,
0.0017049164,
-0.0437956266,
0.012667099,
0.0594203584,
0.0840551928,
0.0152101237,
-0.022736147,
-0.0584491678,
0.015327665,
-0.0483922064,
0.0114273326,
0.0319269858,
-0.0307867602,
0.0310894866,
0.0070776623,
0.0246953815,
0.0911168307,
0.0620175712,
-0.0669075698,
-0.045776967,
0.0525924601,
-0.0231970567,
-0.0128803402,
0.0055686072,
-0.0063986084,
-0.0280390494,
-0.012147597,
-0.0022287117,
0.0663366988,
-0.0398453102,
0.0018343035,
0.0410605818,
0.0167987756,
0.0050545079,
-0.0030606289,
-0.0025576702,
-0.0669324696,
-0.0222884268,
-0.0410920158,
0.005914106,
-0.0187176168,
-0.0149319656,
0.0262527633,
-0.0018466149,
-0.0794452578,
-0.0218627155,
0.0321449079,
-0.0126120625,
-0.0408814251,
0.0428493507,
-0.0060366746,
-0.092722021,
-0.0005322212,
0.0273579396,
0.0115683777,
0.0242474563,
0.0556246042,
-0.0020057873,
0.0079421997,
0.0253963247,
0.0157788321,
0.0419441722,
-0.0353986695,
0.0050980207,
-0.0145698963,
-0.0162413232,
0.0317567363,
-0.0154321361,
-0.0260025747,
-0.010746669,
0.0404775962,
0.0387716517,
-0.0553510524,
-0.0274348911,
-0.0083249435,
2.702e-7,
0.0531654432,
0.0696245581,
-0.0345307253,
0.013342062,
0.0068821413,
0.045840174,
-0.0358146392,
0.0713145584,
-0.0133107947,
0.0560205653,
0.0346903242,
-0.036258705,
0.0844755173,
-0.0025537035,
-0.0377570428,
-0.0702521056,
-0.0030095908,
0.0529253036,
0.0061393706,
0.0428498089,
0.0023737585,
0.0701959506,
-0.0133946436,
-0.0347224623,
-0.0142889405,
-0.0760290027,
0.0349916033,
-0.0258602258,
-0.0046443162,
0.0446913876,
0.0412670784,
0.0417972952,
0.0227224715,
0.0509744175,
-0.0453543887,
0.0410368554,
0.0445947349,
0.0296566654,
0.0105515737,
0.0628348291,
-0.000967096,
0.0073853428,
0.0227186941,
-0.011511649,
0.0322257318,
-0.0371138118,
0.0234377738,
-0.0509918816,
-0.0288905818,
-0.001807816,
0.0325138085,
-0.0147194965,
0.0046151252,
0.0161499679,
0.0525211655,
0.0330253132,
-0.0392255671,
-0.0170782004,
0.0201361571,
0.0230714865,
0.0107748602,
-0.0177714527,
-0.0001353776,
0.0074104783,
0.0514817275,
-0.0580538101,
0.0082943309,
2.378258104e-34,
0.043277815,
-0.010201606,
-0.0206507668,
-0.003170243,
0.0319389328,
-0.0144191124,
0.0270886756,
0.0031726202,
-0.0315006003,
-0.0778785646,
0.0103716655
] |
http://arxiv.org/abs/2401.14625v1 | http://arxiv.org/pdf/2401.14625v1 | Toward Practical Automatic Speech Recognition and Post-Processing: a Call for Explainable Error Benchmark Guideline | 2024-01-26T03:42:45 | 2024-01-26T03:42:45 | Seonmin Koo, Chanjun Park, Jinsung Kim, Jaehyung Seo, Sugyeong Eo, Hyeonseok Moon, Heuiseok Lim | cs.CL | cs.CL | Automatic speech recognition (ASR) outcomes serve as input for downstream
tasks, substantially impacting the satisfaction level of end-users. Hence, the
diagnosis and enhancement of the vulnerabilities present in the ASR model bear
significant importance. However, traditional evaluation methodologies of ASR
systems generate a singular, composite quantitative metric, which fails to
provide comprehensive insight into specific vulnerabilities. This lack of
detail extends to the post-processing stage, resulting in further obfuscation
of potential weaknesses. Despite an ASR model's ability to recognize utterances
accurately, subpar readability can negatively affect user satisfaction, giving
rise to a trade-off between recognition accuracy and user-friendliness. To
effectively address this, it is imperative to consider both the speech-level,
crucial for recognition accuracy, and the text-level, critical for
user-friendliness. Consequently, we propose the development of an Error
Explainable Benchmark (EEB) dataset. This dataset, while considering both
speech- and text-level, enables a granular understanding of the model's
shortcomings. Our proposition provides a structured pathway for a more
`real-world-centric' evaluation, a marked shift away from abstracted,
traditional methods, allowing for the detection and rectification of nuanced
system weaknesses, ultimately aiming for an improved user experience. | Toward Practical Automatic Speech Recognition and Post-Processing: a Call for Explainable Error Benchmark Guideline: Automatic speech recognition (ASR) outcomes serve as input for downstream
tasks, substantially impacting the satisfaction level of end-users. Hence, the
diagnosis and enhancement of the vulnerabilities present in the ASR model bear
significant importance. However, traditional evaluation methodologies of ASR
systems generate a singular, composite quantitative metric, which fails to
provide comprehensive insight into specific vulnerabilities. This lack of
detail extends to the post-processing stage, resulting in further obfuscation
of potential weaknesses. Despite an ASR model's ability to recognize utterances
accurately, subpar readability can negatively affect user satisfaction, giving
rise to a trade-off between recognition accuracy and user-friendliness. To
effectively address this, it is imperative to consider both the speech-level,
crucial for recognition accuracy, and the text-level, critical for
user-friendliness. Consequently, we propose the development of an Error
Explainable Benchmark (EEB) dataset. This dataset, while considering both
speech- and text-level, enables a granular understanding of the model's
shortcomings. Our proposition provides a structured pathway for a more
`real-world-centric' evaluation, a marked shift away from abstracted,
traditional methods, allowing for the detection and rectification of nuanced
system weaknesses, ultimately aiming for an improved user experience. | [
-0.0003017711,
0.0474644154,
-0.0076786089,
0.05084062,
-0.0073565184,
0.0077884737,
0.034638375,
-0.0204181094,
0.0089505445,
-0.0176298209,
-0.0198728312,
-0.0144107193,
0.0550980903,
0.0392178483,
0.009371859,
-0.0676386654,
0.0234196763,
-0.0568329021,
-0.0036277543,
0.0322189778,
0.0298509877,
0.0122440895,
-0.023706276,
0.0634135455,
-0.0101152873,
0.0051225941,
-0.0142096095,
0.0541700162,
0.0151544819,
0.002987396,
-0.0317272358,
0.001711114,
-0.0032124934,
0.0266127661,
0.0000021236,
-0.032507915,
-0.0321854167,
0.029118007,
-0.0783959404,
0.0262405723,
0.0220950656,
0.0219689347,
-0.0125214299,
0.0240603369,
0.0304299034,
0.0220420379,
0.0455378443,
0.0379944406,
0.0402530022,
0.0319934674,
-0.0127941109,
-0.0237875246,
0.0146861635,
0.0156994667,
-0.0057466738,
-0.064250201,
-0.0269837286,
-0.0284649264,
-0.008618623,
-0.0537673607,
0.0273164492,
0.0131022604,
0.0095054861,
-0.0396596752,
0.0129632112,
-0.006098568,
-0.0209222455,
-0.0484892689,
-0.0410513505,
0.053049136,
-0.0106494147,
-0.0163059179,
0.0105875125,
-0.0291505251,
-0.0222375318,
-0.0137589378,
0.0393290073,
-0.031460803,
-0.0530712083,
0.0063797426,
-0.0364662036,
0.048433166,
0.0471442156,
-0.0087424694,
-0.003757538,
0.0162665471,
-0.0106372247,
-0.0351245552,
0.0060640215,
-0.0225154087,
0.0110219903,
0.0353615582,
0.0751669109,
-0.0011207245,
-0.0245872103,
0.0081916219,
-0.0344144776,
-0.0417880528,
-0.0085815564,
-0.0338942185,
0.0503562354,
0.03981103,
-0.0588680282,
0.0376801789,
0.002760624,
0.1053106263,
0.0195194408,
-0.0101632904,
-0.005650545,
0.0219184943,
-0.0154950079,
-0.0009130197,
-0.0135757132,
0.0005415675,
-0.0237473752,
-0.0236230921,
-0.0208201874,
0.0175803602,
-0.0111744413,
0.0237775352,
0.0303378738,
-0.0076461323,
-0.0382620804,
-0.0138146374,
-0.002584894,
-0.0211689658,
-0.01236769,
0.0184243172,
0.017455874,
-0.0812120959,
0.0046051303,
0.0023390313,
0.0257014316,
-0.0100812931,
-0.0031612569,
0.0611286648,
0.0118631506,
-0.025526518,
-0.0470349528,
0.0018837064,
-0.0352989323,
-0.0204684809,
-0.0487386696,
-0.0052606831,
0.0261213463,
0.0111732064,
0.0050074696,
-0.0323729776,
-0.0163166095,
-0.0221315157,
0.0037955435,
0.1064340845,
-0.0478322916,
0.0152482083,
0.0218125675,
0.0413402282,
0.007867964,
0.0211053062,
0.0467402749,
-0.0031005442,
0.0463508368,
-0.0501510948,
0.0352346189,
-0.0018911517,
-0.030411642,
-0.0331111141,
-0.0034911379,
-0.0150797172,
-0.0612193793,
0.0203001183,
0.043599911,
0.0550370403,
0.0055880845,
-0.0241666287,
0.0642447844,
0.0473413132,
0.0133490879,
0.0046390728,
0.0021671334,
-0.0076225013,
0.0009142924,
0.0107631925,
-0.0146439783,
-0.0365220197,
-0.0112433899,
0.0329488739,
0.0345979221,
0.0090536615,
-0.0246859603,
-0.0603908189,
-0.0109124165,
-0.0400574394,
-0.0076171984,
0.0128487162,
0.0506682172,
-0.0030686166,
-0.0594752096,
0.0204850715,
-0.0212286636,
-0.0886670873,
0.0006515677,
-0.1092210114,
0.0187246501,
0.0923102573,
0.0539371148,
-0.0628961399,
-0.0006403846,
0.0144180637,
-0.0403579734,
0.0215030834,
0.0067945039,
0.0163799711,
0.0560557991,
-0.0632344782,
-0.02227116,
0.0261092559,
-0.0060171578,
0.0596698746,
-0.0355727822,
-0.0113892453,
0.0464427173,
-0.0327106491,
-0.0459684953,
0.0097949365,
0.0280657429,
-0.0309518334,
-0.0006444408,
0.0136002302,
0.0803693905,
0.0425126292,
-0.0297791082,
-0.0031360842,
0.0440499112,
-0.0015212354,
0.0415267684,
0.0426709019,
-0.0415142998,
0.0229158774,
0.0366807953,
-0.0250810795,
0.0071909782,
0.0911603719,
-0.0185822733,
-0.0645373315,
-0.0022973081,
-0.043991942,
0.0334941372,
-0.0261451416,
0.0638543516,
0.0131655084,
-0.0095755821,
-0.0002719559,
-0.015928179,
0.0021853724,
0.0141943144,
-0.0211003181,
0.0006325056,
-0.0053932359,
-0.0607374273,
-0.0462489612,
0.0144596687,
-0.0365897715,
0.0319709629,
0.0078473417,
-0.0051948419,
-0.036235489,
0.0275943242,
-0.0637083724,
-0.0552752428,
-0.0319368057,
0.0337662362,
0.0142118065,
-0.0036929077,
-0.0085609974,
-0.0532396398,
0.0128694149,
-0.0097255539,
-0.0181873869,
-0.0321587734,
0.0018365295,
-0.008834525,
-0.0480335616,
-0.0467425846,
-0.0070068841,
0.0033931409,
0.0151017355,
0.0128951687,
0.0070601413,
-0.0807777196,
0.0436175503,
0.0549049973,
-0.0133563792,
-0.0191401523,
-0.0422392227,
0.010645411,
0.03916917,
-0.0283404291,
0.0546289943,
0.0220084228,
0.0382257141,
-0.0183608308,
0.0342454799,
-0.0101297963,
-0.0086551933,
0.0144873373,
-0.0042638388,
0.0037838002,
-0.0502836742,
-0.0229663476,
0.0418555737,
0.0220979359,
0.0492861159,
-0.0386997312,
-0.0071398183,
-0.03085858,
-0.0412851535,
-0.0659950376,
-0.0766192898,
-0.0311841145,
0.0595621765,
-0.0277463365,
-0.0604145601,
0.0475995243,
-0.0900859088,
-0.0021204671,
0.0004496888,
-0.0440040417,
-0.0088347215,
0.0370711647,
-0.0175199974,
0.0097481841,
0.0238813143,
-0.0693210661,
0.0108227907,
-0.0289613884,
0.0230793525,
-0.0381721333,
0.0419259407,
-0.0469052233,
-0.0139230313,
-0.0079888906,
-0.0292613264,
-0.0060799257,
-0.013828042,
-0.0342475623,
-0.0010331114,
0.0321062766,
-0.0328559428,
-0.0085864821,
0.0566778369,
0.0210866462,
0.0690854937,
0.0210950803,
-0.0163880438,
-0.0157236196,
0.0314080417,
-0.021828128,
0.048495546,
0.0144557701,
-0.0245632622,
0.0257701576,
0.0076735262,
0.0856489837,
-0.0068050628,
-0.0557390824,
-0.0168870445,
0.0240951125,
-0.0107199969,
-0.0056992271,
0.0020006341,
0.0546823069,
-0.0525722727,
0.0073310495,
0.085089013,
0.0048433109,
-0.0107240975,
-0.0053016534,
0.0136239761,
0.0229486339,
0.012562437,
-0.0220696963,
-0.0169335995,
-0.0176705737,
0.0699619353,
0.1095248312,
-0.0136425495,
0.0432047583,
0.0157416426,
-0.0675430894,
0.0428186879,
0.0063782991,
-0.0467344485,
-0.0393210091,
-0.0490747727,
-0.0522652976,
-0.0443749167,
0.0724199489,
0.023523435,
0.0082224747,
-0.0387183689,
0.0641652867,
-0.1035409793,
-0.0762206391,
0.0243988056,
0.0100406744,
0.0643185377,
0.0705149993,
0.0299228001,
0.0065058074,
-0.0094933342,
0.0088979248,
-0.0320551544,
0.0421245508,
0.0244952459,
0.0682440549,
0.0318371728,
-0.0163209476,
-0.0032848974,
0.0224370006,
-0.0248576179,
-0.0207767077,
0.0009578413,
-0.0011197659,
-0.0218925383,
0.0356855951,
0.0010937705,
-0.0138945477,
0.01744828,
0.0548151508,
0.0363994576,
-0.0038196722,
-0.0473329239,
0.0049920972,
0.0527882054,
-0.0587221198,
-0.0220303405,
0.0069409087,
-0.0121729346,
-0.0591955632,
-0.0214680657,
-0.0468816683,
-0.0432844013,
0.0530176908,
-0.0038250859,
-0.0269309115,
-0.0244171098,
0.0522491857,
-0.0461806394,
0.0258778781,
0.0487654358,
0.0360866114,
-0.1032465994,
0.0349775776,
0.0080555007,
-0.0551297329,
-0.0387783796,
0.0181209166,
0.0069453665,
-0.0437346287,
0.049604591,
0.0370388143,
0.0390248373,
0.0278577991,
-0.0575148501,
-0.0843318552,
0.0443811379,
0.0044142762,
-0.0295165461,
-0.001859354,
0.0686544925,
-0.0144655574,
-0.033757098,
0.003244879,
-0.0273069218,
0.0110545019,
-0.0361516178,
-0.0260093231,
-0.0193906184,
-0.0815748051,
-0.0039878692,
0.0502306931,
0.0581427254,
-0.0205108803,
-0.0296530016,
-0.0201938022,
0.0540412068,
-0.0037506416,
0.0506995022,
-0.0441622473,
-0.0102381743,
-0.0015228256,
-0.0025318991,
-0.0008247743,
-0.0089130262,
-0.0474406704,
-0.0465865992,
-0.0200758483,
0.0727266297,
0.0253796354,
0.0372267328,
0.0011808978,
-0.0069630067,
-0.035960149,
0.0246164519,
0.0141134206,
0.0367085189,
0.0465653278,
0.0538795479,
-0.0262009222,
0.0152211338,
0.0302256141,
0.0487186797,
-0.0478576235,
-0.0076448959,
-0.0267033372,
0.0090792803,
-0.0079337284,
-0.0092222225,
0.0636695921,
0.015610191,
0.026442714,
-0.0560456403,
0.0388821661,
0.0111065237,
0.0089093773,
-0.0148785189,
0.0234203748,
-0.0129902102,
0.0119072748,
0.0672349483,
-0.0001505885,
0.0050429641,
-0.0252336171,
0.0032201307,
-0.0257813204,
-0.0314542539,
-0.030196419,
0.0153386584,
-0.049168244,
-0.0724054724,
-0.0554257445,
-0.0891507939,
0.0326795615,
0.0280823372,
0.0701865181,
-0.014463868,
0.0172309093,
0.032837376,
-0.0294708218,
0.0148683172,
-0.0138640692,
-0.0517564267,
-0.0459429547,
0.0302907825,
0.0204790123,
-5.939818505e-33,
-0.0002687437,
-0.0318677723,
-0.0138204535,
-0.004240185,
-0.0158851352,
0.0388021395,
-0.0130303875,
-0.0000252376,
0.0266008619,
-0.0192333721,
-0.0242601447,
-0.0047610993,
0.0413025357,
0.0183335729,
-0.0280270018,
-0.0104315979,
0.0804570019,
-0.0054920558,
-0.0259080399,
-0.0075264284,
0.0619080588,
0.0238003153,
-0.0060076024,
-0.0911449343,
0.0045118672,
0.0277051739,
-0.0497935787,
-0.0241389479,
0.0286585893,
0.0545575581,
0.0164716672,
0.0260476004,
0.0055357679,
-0.0039125797,
-0.0022500278,
0.0180699434,
-0.0531492531,
0.0067026457,
-0.0138217127,
0.0767081156,
0.0146660423,
-0.0032873,
0.0458103158,
-0.0122749507,
-0.0237105694,
-0.0177031793,
0.0193588547,
-0.038809903,
-0.0223825108,
-0.0252787061,
0.0086589446,
0.0480753593,
-0.0327952765,
-0.0288016144,
0.0486839041,
0.0627981499,
-0.0245276429,
-0.018893972,
-0.0626686662,
0.0075020483,
0.0483868755,
-0.0087866178,
0.0864494145,
0.0268268678,
-0.0450661257,
0.0248197261,
0.0695066527,
-0.034075398,
-0.0073138312,
-0.0414567553,
0.0232825652,
0.0361273661,
0.05994213,
0.0404917337,
0.040515095,
-0.0476772003,
-0.0194242913,
0.0206119828,
-0.0115783494,
0.0042705466,
0.0426239893,
0.0451904796,
-0.1083641052,
-0.0034264512,
-0.0492309332,
0.0027829793,
-0.0522019714,
-0.0074597695,
0.0658795163,
-0.0130041419,
-0.0372026041,
-0.005560664,
0.0051977765,
-0.0185815077,
0.0534309074,
-0.0170395914,
0.019163534,
0.0052009304,
-0.0144678354,
0.0092840102,
-0.0090265963,
-0.0219981521,
0.0092081912,
0.0792945549,
0.0439567566,
-0.0503361896,
0.0000610961,
0.0577586591,
-0.0518782996,
-0.00046261,
0.0054091783,
0.0108081475,
-0.0136084119,
0.0087988079,
0.0531981476,
-0.0483084433,
-0.0399180166,
0.0504853763,
-0.0260710865,
0.0436600298,
-0.0101541262,
0.0100000082,
-0.0046846499,
0.0270292293,
-0.017439831,
-0.0283635799,
0.0219101459,
-0.0352372751,
0.0304387342,
-0.0068598036,
-0.0160745494,
-0.0077318954,
2.848e-7,
-0.0312954038,
0.0057164649,
0.0518398248,
-0.0282770358,
-0.005081757,
0.0361807793,
-0.00822357,
-0.0131907454,
-0.0108459732,
0.0250755753,
0.0400767252,
-0.0425990447,
0.1246142164,
0.0317496918,
-0.0258610174,
-0.0415075235,
-0.0550930686,
0.0230003204,
-0.0140660005,
0.0371479057,
0.0730798617,
0.062519826,
-0.0024084374,
-0.0332512297,
0.0162440743,
-0.040852312,
-0.0069427453,
-0.0471982211,
0.034686815,
-0.020402154,
0.0495551527,
0.0161254425,
0.0032498259,
0.0335367359,
-0.0399700627,
-0.0374476165,
0.0168178715,
0.0297239237,
-0.0376655795,
-0.0613637231,
-0.0139805162,
0.0313158743,
-0.0060667153,
-0.0611413494,
0.0424453542,
0.0221052058,
-0.0212893691,
-0.0023558973,
-0.0541297793,
-0.0291342977,
0.0189179573,
-0.01811436,
0.0187375396,
-0.0161989424,
-0.0028691955,
-0.0375175774,
-0.0221106969,
-0.0742827505,
0.0113725532,
0.0161408111,
-0.0089652389,
-0.0158945452,
-0.0504642949,
0.0111810761,
0.0590521581,
-0.0381266363,
-0.0105778296,
2.326907827e-34,
0.0142717911,
-0.0198707655,
0.006308456,
0.0014749471,
0.0265819784,
0.0087440833,
0.0049028951,
-0.000081049,
0.0166861061,
-0.0072642136,
-0.0264966264
] |
http://arxiv.org/abs/2401.14624v3 | http://arxiv.org/pdf/2401.14624v3 | Query of CC: Unearthing Large Scale Domain-Specific Knowledge from Public Corpora | 2024-01-26T03:38:23 | 2024-03-05T02:45:41 | Zhaoye Fei, Yunfan Shao, Linyang Li, Zhiyuan Zeng, Conghui He, Hang Yan, Dahua Lin, Xipeng Qiu | cs.CL | cs.CL | Large language models have demonstrated remarkable potential in various
tasks, however, there remains a significant scarcity of open-source models and
data for specific domains. Previous works have primarily focused on manually
specifying resources and collecting high-quality data on specific domains,
which significantly consume time and effort. To address this limitation, we
propose an efficient data collection method $\textit{Query of CC}$ based on
large language models. This method bootstraps seed information through a large
language model and retrieves related data from public corpora. It not only
collects knowledge-related data for specific domains but unearths the data with
potential reasoning procedures. Through the application of this method, we have
curated a high-quality dataset called KNOWLEDGE PILE, encompassing four major
domains, including stem and humanities sciences, among others. Experimental
results demonstrate that KNOWLEDGE PILE significantly improves the performance
of large language models in mathematical and knowledge-related reasoning
ability tests. To facilitate academic sharing, we open-source our dataset and
code, providing valuable support to the academic community. | Query of CC: Unearthing Large Scale Domain-Specific Knowledge from Public Corpora: Large language models have demonstrated remarkable potential in various
tasks, however, there remains a significant scarcity of open-source models and
data for specific domains. Previous works have primarily focused on manually
specifying resources and collecting high-quality data on specific domains,
which significantly consume time and effort. To address this limitation, we
propose an efficient data collection method $\textit{Query of CC}$ based on
large language models. This method bootstraps seed information through a large
language model and retrieves related data from public corpora. It not only
collects knowledge-related data for specific domains but unearths the data with
potential reasoning procedures. Through the application of this method, we have
curated a high-quality dataset called KNOWLEDGE PILE, encompassing four major
domains, including stem and humanities sciences, among others. Experimental
results demonstrate that KNOWLEDGE PILE significantly improves the performance
of large language models in mathematical and knowledge-related reasoning
ability tests. To facilitate academic sharing, we open-source our dataset and
code, providing valuable support to the academic community. | [
-0.0014132621,
0.0987543091,
-0.0200014561,
0.0515864976,
-0.0569346696,
0.0163190551,
0.0292541701,
0.0143626379,
-0.0091600511,
-0.0092768557,
0.0053571202,
-0.0134055847,
-0.0081564775,
0.0186716691,
0.033682771,
0.0005422055,
0.0272348728,
-0.0379090607,
0.0208226778,
-0.0254209712,
-0.0441691615,
0.0035881682,
-0.0440856554,
-0.018864898,
0.0168624148,
-0.0653845593,
0.0223727804,
-0.0372459702,
-0.001855397,
-0.029644236,
0.0001556459,
0.065056771,
0.0337473489,
0.050843522,
0.0000021304,
-0.0755939856,
-0.0248581953,
0.0266091749,
-0.0426938981,
-0.0090567302,
0.0079116831,
0.0267428588,
0.010569728,
0.0086449217,
-0.0387190394,
-0.010358111,
0.0230475515,
-0.0055946498,
0.0121698985,
0.0508245938,
0.0149313044,
-0.049337402,
0.0230696481,
-0.0076185144,
0.0330799595,
-0.0191914272,
0.0536707044,
0.0629240423,
0.0351307765,
0.062714003,
-0.0202376302,
0.0261885673,
0.0081383036,
-0.0155892503,
-0.0059506567,
0.0358098447,
-0.0350756831,
-0.0062936242,
0.0444260836,
0.0590344518,
0.0536542274,
-0.0254380573,
0.0309525896,
0.0354033597,
-0.016723644,
0.0829853937,
-0.0817289501,
0.052373074,
-0.0129784606,
-0.0299545247,
0.0103827575,
0.0000309415,
0.0369085521,
0.0379783846,
-0.0099169407,
0.0733529627,
0.0286878552,
-0.0236070026,
0.0300011765,
-0.0096936263,
-0.0354351103,
-0.046888154,
-0.0060744267,
0.0090211099,
0.0020989936,
0.024750324,
0.0202525668,
-0.0133917434,
0.0433243774,
0.013639763,
-0.01569915,
0.0227537733,
0.0111840153,
0.0768440738,
0.0052038999,
0.0366385952,
-0.0201347675,
0.0116401408,
0.0043830313,
-0.0319678485,
-0.0773815513,
-0.0153864156,
-0.0050209882,
0.0309799109,
0.018046407,
0.0255162511,
-0.035657037,
0.000313492,
-0.0389264934,
0.0107674263,
0.0170995127,
-0.0228322856,
-0.0017088675,
-0.0025943392,
-0.0434437543,
0.0269794278,
-0.0571040474,
0.0213020872,
0.0038965768,
-0.0234188195,
-0.0336293243,
-0.0100946035,
0.0067058112,
-0.0308097508,
0.0350614339,
0.0356978737,
0.0563487448,
0.0024352393,
0.0044717295,
-0.0372660086,
0.0208336804,
-0.0451362059,
0.0441410728,
-0.1017239466,
0.0507506616,
0.035716489,
-0.0001673862,
-0.0629723519,
-0.0135802189,
0.0313961729,
-0.0318883248,
0.0444451533,
-0.0464267582,
-0.0022553175,
0.0314666145,
0.0124236839,
-0.0198029652,
0.0563631095,
-0.0017247243,
-0.0233295374,
0.0676040873,
-0.0188751798,
0.0035159562,
-0.0044484353,
0.0438500345,
0.0125323357,
-0.0013016678,
0.0546408892,
-0.0068132235,
0.0421449989,
-0.0170095526,
0.0269652512,
-0.0154945916,
-0.0250274129,
-0.0012335737,
0.0914256573,
0.0232003406,
0.0079259193,
0.009916489,
0.0119860172,
0.0359942205,
-0.0622153617,
-0.0180866346,
0.0173131134,
-0.0445203558,
0.0048950841,
0.0051576067,
0.0222913418,
-0.0115483087,
-0.0219964255,
-0.0326080471,
0.025522802,
-0.0391248837,
0.0389649421,
0.0766602755,
-0.019079309,
0.0336827263,
0.0384704843,
0.0642233938,
-0.0509411469,
0.0149755655,
0.0031537807,
0.0350733064,
0.0110641131,
0.0529687107,
-0.0674871951,
-0.0339459367,
0.0139912162,
0.0388028547,
0.0529362783,
0.0078320801,
-0.0514665172,
-0.0349173248,
-0.0207068976,
-0.004578358,
0.0002985045,
0.005588071,
-0.0321735069,
0.0441973619,
0.030382568,
0.0032377716,
-0.0944968164,
0.0075107901,
0.0113506969,
-0.070563592,
-0.003496326,
-0.0305417236,
-0.0274535436,
0.0182086863,
0.0135351811,
-0.0012145909,
0.0335773081,
0.033353053,
-0.0339776017,
-0.0346996859,
-0.0130516551,
-0.001620798,
-0.0213212892,
0.0593339466,
-0.0595681295,
-0.0113679776,
0.0119639616,
-0.0243742354,
-0.0140524665,
0.0556481816,
0.0168408304,
0.0460922308,
-0.0160349105,
0.0056885118,
-0.0479255579,
-0.0240952801,
-0.0076444298,
0.0929670334,
0.0230932962,
0.045111753,
0.0303701982,
0.0008353281,
0.0310645252,
0.002085095,
-0.0298252981,
-0.0199583378,
-0.0042808834,
0.0194386747,
0.0417256951,
-0.0222231131,
-0.0768803731,
0.0144146718,
-0.0404355638,
0.0211664606,
0.0009538785,
0.0154092722,
-0.0027581034,
-0.0398177728,
-0.0131171532,
0.0324515291,
-0.0181214903,
-0.008988183,
-0.0139895985,
0.0086272424,
0.0141225662,
0.0040023434,
0.0017581523,
-0.0117321126,
-0.0406708866,
-0.0521413572,
0.0607947484,
0.0279675033,
0.0368733108,
0.0038013349,
0.07622765,
-0.0562505424,
-0.0023170507,
-0.0162211321,
0.0190019235,
0.0207837615,
-0.0303283278,
0.0026457063,
0.0434167013,
-0.0137358662,
0.0075510647,
0.0199515875,
-0.0279396623,
-0.0458588488,
0.0065363413,
-0.02722813,
0.0091111632,
-0.0276225507,
-0.0291733947,
-0.0053758277,
0.0387253463,
0.0013483521,
-0.0348053649,
-0.0461761095,
0.0467503853,
-0.0586121976,
0.0204922296,
-0.0638142303,
0.0378111824,
0.0119062196,
0.0594560653,
-0.0091238897,
0.0092926556,
0.0473173037,
-0.0119627826,
0.0207502414,
-0.0427034386,
0.0367753468,
-0.0411492474,
0.0755876452,
-0.0190526396,
-0.0111435447,
0.0084064789,
0.0179480948,
-0.0348529294,
0.0226417575,
0.0186081398,
-0.0427121781,
0.0119204801,
-0.0454446822,
0.0160157438,
0.0126180835,
-0.0260486193,
-0.0239803009,
0.0029713535,
-0.0151074659,
0.0420272946,
0.0093212752,
-0.0657728985,
-0.0146668265,
-0.0354906246,
0.0503588766,
0.0025426119,
0.0008096714,
-0.0151607003,
0.004439604,
0.0019300248,
0.0346841216,
0.0244230889,
0.0206238087,
-0.0364710763,
-0.0047752946,
0.062052764,
0.032405097,
-0.0074885902,
0.0283834636,
0.0109494207,
0.0225317553,
-0.0429781154,
-0.0019363035,
0.0202598069,
-0.015057235,
-0.0375580937,
-0.0285105258,
0.1233140901,
0.0294632539,
-0.0446388684,
-0.0245640278,
-0.0357996449,
-0.001971313,
0.0646141469,
0.0152673805,
-0.1218776479,
-0.0196660347,
-0.0124262301,
0.0119130202,
0.0133633381,
-0.0704277381,
-0.0048642321,
-0.1388506293,
0.0296779461,
-0.0115342345,
-0.0281107984,
-0.0063915066,
-0.0745849609,
0.0218182281,
-0.0088143684,
-0.0614694618,
0.0138012972,
-0.0074815885,
0.0237244666,
0.0282305628,
0.0114729842,
-0.0834136233,
0.0741457865,
-0.0298598427,
0.033052329,
-0.0047015455,
0.0588262267,
-0.0538071357,
-0.0625035614,
-0.0151940454,
-0.015139671,
-0.0447991937,
-0.0051470711,
0.0211481024,
0.0154638309,
-0.0059586572,
-0.0157466605,
0.0379381962,
-0.0754954666,
-0.0167116113,
0.0545729436,
0.1379579902,
0.0123813273,
-0.0076511279,
-0.0365374424,
-0.011533998,
-0.0288098324,
0.1023873612,
-0.0031507167,
-0.0067841103,
0.0269758757,
-0.011042662,
-0.0439807922,
-0.0139173875,
-0.071327202,
0.0314450935,
0.0219956115,
-0.0198580455,
-0.0381607972,
0.0000244904,
0.0661555752,
0.0316714048,
0.0313609578,
0.0326164551,
0.0392133445,
0.00694018,
-0.0757781193,
0.0500955619,
-0.0473157912,
0.0054061459,
-0.0648845956,
-0.0176488161,
0.0128076533,
-0.0186380148,
-0.0677205324,
0.0005415173,
-0.0626555756,
-0.060557127,
-0.0097624473,
-0.00150942,
0.023849044,
0.0110739423,
-0.0177357942,
-0.0057920781,
0.047220286,
0.0064523076,
-0.0660569593,
0.0043528401,
0.0000076108,
-0.0613685958,
0.012836013,
0.0318720415,
-0.0472084209,
0.0152456537,
0.0031793376,
0.0511040166,
-0.0509904735,
0.0010187475,
0.0095548015,
0.0538426004,
0.0338587426,
0.0199643094,
-0.0142748356,
-0.0355182737,
0.0104905954,
-0.0286884364,
-0.0451731309,
-0.0097408844,
0.0106541729,
0.0091818376,
-0.0471101291,
-0.002681443,
0.005030585,
-0.0040980056,
0.0324501544,
-0.017852731,
0.0250348393,
0.0183689278,
0.0412884951,
0.0490216725,
-0.0320184864,
0.0204106569,
-0.0455292128,
-0.0471929312,
0.065932937,
-0.038853243,
0.0349027514,
0.0226583686,
-0.0048213326,
0.0127483942,
-0.024823986,
-0.0435774066,
-0.0767093301,
0.0436435863,
0.0138981696,
0.0277288649,
-0.013861239,
0.0063605416,
-0.0558994412,
-0.0164426267,
-0.0147892321,
0.001489274,
0.0316668525,
-0.0349248499,
0.0478862599,
0.0498948283,
-0.0066593061,
-0.0068540522,
0.066515632,
-0.0194224436,
0.0180290453,
0.0098977126,
-0.0151512986,
-0.0007994827,
-0.0164307579,
-0.0156316217,
0.0071353517,
-0.0002061238,
0.0246479623,
-0.0968719572,
-0.0775096342,
0.007716889,
0.0049815946,
-0.0226459205,
0.0317367278,
0.0422452465,
0.0814977065,
-0.0153209781,
0.0025443665,
-0.0321489871,
0.0540409461,
-0.071987547,
-0.0508861467,
0.0139184287,
-6.401263955e-33,
-0.0058797281,
-0.0328355432,
0.0375444777,
-0.01067161,
0.019583879,
0.0182563923,
-0.036001835,
-0.001672341,
-0.0474426448,
0.0396621078,
-0.0356346667,
0.0310391728,
-0.0042296988,
0.0035628395,
0.0265727583,
0.0285742991,
-0.0045921346,
0.0050664055,
-0.0226620231,
0.0210704375,
0.0755103827,
0.0006222477,
0.0567962937,
-0.0740204155,
0.0407397188,
-0.0143456273,
0.0358567275,
0.037215855,
-0.0203340165,
0.0439151339,
-0.0076681562,
0.0450878553,
0.0055472828,
-0.0661526471,
0.0169345737,
-0.0286572371,
-0.0742142349,
-0.031446714,
0.0296975058,
-0.0121472254,
-0.0100429589,
-0.0334217958,
0.0103416583,
0.0206326637,
0.0056248214,
-0.043129608,
0.0479329117,
-0.0346452035,
-0.0113021154,
-0.0487871505,
-0.038818121,
0.0074003078,
-0.0469618775,
0.0144609744,
-0.0044326247,
-0.0029106785,
-0.0033518795,
0.0455731265,
-0.0162622239,
0.0632404387,
0.0122904535,
0.0258163884,
0.0328804553,
-0.0310376771,
0.0438630134,
0.0723612681,
-0.0268250201,
-0.0153471408,
0.023658134,
0.0052426504,
0.0336028934,
0.0498260669,
0.0275232233,
0.0716697201,
0.0485459715,
-0.0159156248,
-0.0861904547,
0.0062950943,
0.0152569395,
0.015255223,
0.024843026,
-0.0280300211,
-0.0447544679,
-0.0222789552,
-0.032991387,
0.0373650603,
-0.0311643258,
-0.0033827613,
-0.0028230436,
-0.0462825932,
-0.0329632126,
0.0076796403,
0.0107269958,
-0.052013278,
-0.041737102,
0.0693520606,
-0.0052606952,
0.0024001501,
-0.0096723996,
0.0373826884,
-0.0866637602,
-0.013310995,
-0.021039838,
-0.0084638,
0.0273322444,
0.0238148216,
0.0231337957,
0.0082508409,
-0.0276203193,
-0.0445895046,
-0.0460825972,
-0.0196170248,
0.0171792973,
0.0173291564,
0.0011845172,
0.0868067443,
0.0178697687,
0.0016735145,
0.0559048876,
-0.0029418657,
-0.0162097383,
0.0295123644,
-0.0161009803,
-0.0082194079,
-0.0020663126,
-0.0541505963,
-0.0266923029,
-0.0026800411,
0.024279153,
-0.0749006271,
0.0077094291,
0.0054687676,
2.876e-7,
0.0478643589,
0.0479018986,
-0.0252568386,
0.0377703905,
0.0086341593,
0.0029285809,
-0.0483823679,
-0.0275367834,
0.0360128544,
-0.0262814108,
-0.0267089028,
0.044151295,
0.0269168504,
0.0645200908,
-0.0903949589,
-0.0265287478,
0.0080284374,
-0.0489328019,
-0.0603541434,
0.012108054,
0.0063404921,
0.0419903845,
0.0363018401,
-0.0210498106,
-0.0175197292,
-0.0412467979,
0.0379288755,
-0.0132034319,
0.008721929,
0.0246150848,
-0.028099712,
0.0133912796,
0.0142383529,
0.0368071496,
-0.0061436463,
0.01288193,
-0.0067362492,
0.006184753,
0.0289461501,
0.0069225864,
-0.0204349216,
-0.0668269917,
-0.0025510713,
-0.0200751442,
0.0551624224,
-0.0167250447,
-0.0376013629,
0.026446173,
-0.0179567393,
-0.0189398862,
-0.007048741,
-0.0184137691,
0.0185534079,
-0.0344693772,
0.005543259,
-0.0314269811,
0.0192691665,
-0.0396567248,
0.0021889338,
0.0168544222,
-0.0477513634,
-0.0852264017,
0.0122883907,
-0.0403627492,
0.0036021266,
-0.0327820703,
0.0267175604,
2.808456435e-34,
0.0327607244,
0.0169531964,
0.0069853831,
-0.0131890895,
0.0432618782,
-0.0388795659,
0.0086269742,
-0.0219833814,
0.0008405154,
-0.0801772997,
-0.0178282596
] |
http://arxiv.org/abs/2401.14616v1 | http://arxiv.org/pdf/2401.14616v1 | Alternative Speech: Complementary Method to Counter-Narrative for Better Discourse | 2024-01-26T03:16:54 | 2024-01-26T03:16:54 | Seungyoon Lee, Dahyun Jung, Chanjun Park, Seolhwa Lee, Heuiseok Lim | cs.CL | cs.CL, cs.AI | We introduce the concept of "Alternative Speech" as a new way to directly
combat hate speech and complement the limitations of counter-narrative. An
alternative speech provides practical alternatives to hate speech in real-world
scenarios by offering speech-level corrections to speakers while considering
the surrounding context and promoting speakers to reform. Further, an
alternative speech can combat hate speech alongside counter-narratives,
offering a useful tool to address social issues such as racial discrimination
and gender inequality. We propose the new concept and provide detailed
guidelines for constructing the necessary dataset. Through discussion, we
demonstrate that combining alternative speech and counter-narrative can be a
more effective strategy for combating hate speech by complementing specificity
and guiding capacity of counter-narrative. This paper presents another
perspective for dealing with hate speech, offering viable remedies to
complement the constraints of current approaches to mitigating harmful bias. | Alternative Speech: Complementary Method to Counter-Narrative for Better Discourse: We introduce the concept of "Alternative Speech" as a new way to directly
combat hate speech and complement the limitations of counter-narrative. An
alternative speech provides practical alternatives to hate speech in real-world
scenarios by offering speech-level corrections to speakers while considering
the surrounding context and promoting speakers to reform. Further, an
alternative speech can combat hate speech alongside counter-narratives,
offering a useful tool to address social issues such as racial discrimination
and gender inequality. We propose the new concept and provide detailed
guidelines for constructing the necessary dataset. Through discussion, we
demonstrate that combining alternative speech and counter-narrative can be a
more effective strategy for combating hate speech by complementing specificity
and guiding capacity of counter-narrative. This paper presents another
perspective for dealing with hate speech, offering viable remedies to
complement the constraints of current approaches to mitigating harmful bias. | [
0.078178212,
0.0689684302,
-0.0239558183,
0.0030208784,
-0.0373659991,
0.0205217004,
0.0269609839,
-0.0183013361,
-0.0154494382,
-0.0532112345,
0.046272736,
-0.0053686234,
0.0359474123,
0.0177452434,
0.0195032991,
-0.0765294805,
0.0177423377,
0.0299243573,
0.0618878342,
-0.037092112,
0.0478799306,
0.0109506138,
-0.0269910023,
0.0178706273,
-0.0654306859,
0.019239895,
0.0583676808,
-0.0146019161,
0.0303432066,
0.0186665319,
0.0354270935,
-0.0103776259,
0.0089149196,
0.0263599642,
0.0000019938,
-0.0510111824,
0.0431484729,
-0.0123357922,
-0.0136579666,
-0.009193127,
0.0566402674,
0.0051804865,
-0.0000677873,
0.0574009866,
-0.0042466703,
0.0162112806,
0.0044630887,
0.0823740363,
0.0619224906,
0.0686717257,
0.0012382475,
-0.0332305953,
-0.022281833,
0.0035523102,
-0.0442044325,
0.0289470516,
0.0224810019,
-0.0371453129,
-0.0656277984,
-0.0333949253,
0.0114567578,
0.0179129522,
0.0056992746,
0.0170056671,
0.0031434605,
-0.0245435853,
-0.0211866703,
0.0007186797,
-0.0198652428,
0.0241274387,
-0.009164067,
0.0024581619,
0.00555227,
-0.0125605166,
0.016509261,
-0.0183105953,
-0.0301254839,
0.0294356737,
-0.0212333519,
-0.0198627524,
-0.0377632678,
-0.0170303155,
0.0115412017,
0.061362613,
0.0512660034,
0.0399899259,
0.0076830443,
-0.0220597852,
-0.0178106297,
-0.0355138816,
0.0281527098,
-0.0104653873,
-0.0158049725,
-0.0029512409,
-0.047922533,
0.0334909484,
-0.01711246,
-0.0308717769,
0.0059039188,
-0.1045728922,
0.0313699692,
0.0341157839,
-0.0070403693,
-0.0490684919,
0.0282027014,
0.0359112546,
-0.0092853066,
0.0009632297,
0.0258377846,
0.021266209,
0.0106978007,
0.0060116611,
-0.0157314762,
-0.0278014727,
-0.0545397475,
0.0654268414,
-0.0033168292,
-0.0228341371,
-0.0548531376,
0.0257873349,
-0.0509954505,
0.0491452999,
-0.0134816524,
-0.0115456302,
-0.0077551412,
0.0332559496,
-0.0446173511,
0.0156214358,
-0.0254315175,
-0.0754310489,
0.0125161391,
0.0258811209,
-0.0274429042,
-0.0001460687,
0.0180453733,
0.0427222326,
0.0069534327,
0.0253760871,
-0.0013191083,
-0.0139196729,
-0.0603935868,
-0.0468886942,
-0.0052780202,
-0.0498886742,
-0.0236271601,
-0.014036539,
-0.0106304353,
-0.047407236,
0.0299346633,
0.0068255137,
-0.0418602787,
0.0366301574,
0.0263023991,
-0.0065330835,
0.1009786949,
0.0438240729,
0.0192184485,
0.0086831441,
0.0386212617,
-0.0438221022,
0.0407821052,
-0.0178680941,
0.0001338351,
0.030232368,
0.0481543131,
0.0058896733,
-0.0270082559,
0.046712745,
-0.0306717642,
0.0002040964,
0.0007866692,
0.0080553554,
0.0007969438,
-0.0564489216,
0.0279075094,
0.0762737244,
-0.0852197036,
-0.0101464754,
0.029553894,
0.0362505801,
-0.0237690508,
-0.002027804,
-0.0204094276,
-0.0371697843,
-0.0190103315,
-0.0368805006,
0.0081017995,
0.0452579483,
-0.0173615552,
-0.0084897336,
-0.0304692704,
0.0188985951,
0.0743786842,
-0.0194500126,
-0.0335190967,
0.031236304,
-0.0035188119,
0.0057948111,
0.0437799096,
-0.0166014209,
0.0054770689,
-0.054418996,
-0.0286692344,
0.0501735359,
-0.0087768342,
0.0415335372,
-0.0007192408,
-0.087305069,
-0.025872713,
0.04337034,
0.0206123833,
-0.0069275037,
-0.0041598212,
-0.0128901117,
-0.0046625179,
0.0192691237,
0.022126548,
0.0438703559,
0.0612394363,
-0.0083458163,
-0.001062825,
0.0608093403,
-0.0563851185,
0.0060656569,
0.0349992663,
-0.0167572163,
0.0208422057,
-0.0217049401,
0.040845558,
0.0203564391,
-0.0371676385,
-0.0306275357,
-0.0127874538,
0.0379009359,
-0.025168309,
0.0572404936,
-0.0706663653,
0.0020114498,
0.0028733823,
-0.0824416578,
-0.0341516249,
0.0484843925,
0.0188690647,
0.0359703563,
0.0841789097,
-0.0168059003,
0.0038108905,
0.0083903149,
-0.024333816,
-0.0524568371,
-0.0234049112,
-0.0063692001,
0.0177762862,
-0.0006967744,
-0.0001068172,
0.0286843646,
-0.0518271364,
0.0420634411,
-0.0025792376,
-0.0096483445,
-0.0066762073,
-0.0286418404,
0.0490684435,
0.0038052374,
-0.0265728813,
-0.0222449638,
0.0625702664,
-0.0607407428,
-0.0168878082,
-0.0158674978,
-0.0535432734,
0.002958508,
0.0004644376,
-0.0298737865,
-0.0130005106,
0.0053099617,
-0.0782739148,
0.0541595072,
-0.0085707987,
0.0037503927,
-0.0120513802,
-0.0156102162,
-0.0286411829,
0.0069665094,
-0.026772229,
0.0079367552,
0.0047623874,
-0.0558225475,
0.0104467999,
-0.0374551266,
-0.0128563261,
-0.0013379845,
-0.0211645346,
-0.0034160942,
0.0181777216,
0.0730331913,
0.0107462881,
0.0051256604,
0.0691713616,
0.0440128967,
0.0131158624,
-0.0072000106,
-0.0275239516,
-0.0271588992,
-0.0017971774,
0.0259930938,
-0.0253215171,
-0.0397501066,
-0.012821313,
0.0286906566,
0.0939963013,
-0.0184781458,
-0.0530008227,
0.0275941119,
-0.064733915,
-0.0477222018,
-0.0371200815,
-0.1188856736,
-0.0504678115,
-0.0358304158,
-0.041289635,
-0.0158662777,
0.0427698456,
0.0073960535,
0.0388240106,
-0.0087704854,
-0.0182581525,
0.0200420301,
0.0430288762,
-0.0147565743,
0.0323703773,
-0.0070993169,
-0.02724755,
0.035606727,
-0.0484249331,
0.0432928987,
-0.0126692057,
-0.0268782545,
-0.0467218012,
0.0270845685,
0.0397877879,
-0.0327242948,
-0.0657721758,
-0.0467169322,
-0.0376803726,
-0.0273878537,
-0.0141414134,
0.0018509516,
-0.0610871129,
-0.0277170353,
0.05552461,
0.0502963439,
0.0279941484,
0.0159294866,
0.0379356593,
-0.0589836314,
-0.0347704589,
0.0006716365,
0.0462696441,
-0.0109831719,
-0.0063239611,
0.0497420207,
0.0115038631,
-0.0098723741,
0.0333177298,
0.0002949563,
0.0025351385,
0.0392205045,
0.0158482846,
0.0120597733,
-0.0007205269,
0.0004723507,
-0.0013704569,
0.0480153151,
0.0028729206,
0.0274048261,
-0.006129819,
-0.0138844894,
0.0078044785,
0.0219805036,
-0.0511307158,
-0.0868403688,
0.026092831,
0.0090234494,
0.0299631283,
-0.106126979,
-0.0172132626,
-0.0198875796,
-0.0354533568,
0.0220452975,
0.0128110824,
-0.0001516235,
-0.0367113613,
0.029265549,
0.0045910794,
0.0364996865,
0.0335845426,
0.0007881849,
0.0321894959,
-0.0296346061,
0.0276173893,
-0.0893810764,
-0.0005385116,
-0.0079938453,
0.0109753683,
0.0280766841,
0.0129139591,
0.0449042767,
-0.0020732274,
0.0630731732,
0.006011134,
0.0122015905,
-0.0173311234,
-0.0349420346,
0.0291399173,
-0.0002970366,
-0.023822058,
-0.0286640543,
0.0610226281,
0.0660941303,
-0.0357003659,
0.0584415123,
0.0253770482,
0.0457037352,
0.0202804804,
-0.0273965131,
0.029020112,
-0.0082025817,
0.0709009394,
0.0474312529,
-0.0129223829,
0.0741360337,
-0.079514809,
-0.0038815225,
-0.027873151,
-0.0860969275,
-0.0133992778,
0.0041810283,
-0.0093705291,
-0.0087079238,
-0.0583746247,
0.0055764774,
0.0048254337,
0.0455337539,
-0.0390853137,
0.0135115879,
-0.0188731588,
-0.0638176575,
-0.0198539104,
-0.0000477045,
0.0479763038,
0.0487198085,
0.0098236287,
-0.0461479574,
-0.0068567027,
-0.0489298552,
0.0319627151,
0.0004913649,
-0.0468369722,
0.0850872397,
0.0268907361,
0.0186595265,
0.0565446131,
0.0309549943,
-0.056967482,
0.0458663628,
0.0177882127,
-0.039579466,
0.0331623554,
0.0346003436,
-0.0033275471,
0.0254407637,
-0.0210581012,
0.0390772969,
0.0012223644,
0.0316859186,
-0.0097288741,
-0.0498591289,
-0.0132772522,
0.0046293782,
-0.0331424698,
0.0046376069,
-0.023818722,
-0.0225918423,
-0.0464411788,
0.0007986228,
-0.0071739336,
-0.0681240708,
0.0132446252,
0.021908097,
-0.0156041272,
-0.0006749327,
-0.0273757391,
-0.0336209424,
-0.0889294147,
0.0214914065,
-0.029138241,
0.0042515108,
0.0421615988,
0.034967117,
-0.0002260284,
0.0126766535,
0.0010984911,
0.015104535,
0.0011440247,
0.0069236434,
0.0124407345,
-0.0032971979,
0.031577237,
-0.0152034638,
0.0488284118,
-0.0601462387,
-0.017148925,
-0.0161393844,
-0.0469010286,
0.0152831431,
-0.0522430874,
0.0007891402,
0.0418190882,
0.0383831412,
-0.0179546513,
0.0130314548,
-0.0265845191,
-0.0426975638,
-0.0564869717,
0.0043168687,
0.0172232781,
-0.0302612167,
0.0152906906,
-0.0303930119,
-0.0204401314,
0.0827205107,
0.0124038467,
-0.01172846,
-0.0244870167,
-0.0496333204,
0.0196045395,
0.0294069182,
0.0410679765,
-0.0048847855,
-0.0839904845,
0.0115729356,
0.0169573165,
-0.005194352,
0.0492293052,
-0.0315186568,
-0.0200526137,
0.0371418446,
0.0124571975,
0.00171207,
0.0226018038,
0.0276895221,
0.0523996986,
-0.0472627394,
0.0002036749,
-5.620689054e-33,
-0.0471648835,
-0.0521887466,
-0.0108261658,
-0.0325971805,
-0.0016274226,
-0.0347463004,
0.0353499949,
0.0134752188,
0.0031508435,
0.055429168,
0.0041111661,
0.0253272485,
0.0110350652,
0.0060922131,
-0.0250910968,
0.0048463494,
0.0387962945,
-0.0236924868,
-0.0231659655,
0.0019738874,
0.021196777,
-0.0077618351,
0.0388418362,
-0.0698603988,
0.0728208795,
0.0075077531,
-0.0402485691,
-0.0078553194,
0.0150539707,
0.0104421489,
0.0052474472,
0.0749957114,
0.00795775,
-0.0124197165,
-0.0073767263,
0.0798842311,
-0.0123901218,
0.003688009,
0.0825103819,
0.0076385927,
0.0644771978,
0.0010365276,
-0.0153816398,
-0.0092581101,
-0.0612050705,
-0.0027723399,
0.0278851148,
-0.0263021793,
0.0491757579,
0.0222767387,
-0.0775914043,
0.016925374,
-0.0745720789,
0.0187710077,
-0.0532557294,
0.0636064038,
0.0152432062,
-0.0481232442,
-0.0539618693,
0.053168796,
0.0240373202,
-0.0373502113,
-0.0259156954,
-0.0223287176,
0.0013765821,
0.0467480496,
0.025031602,
-0.0025467279,
-0.0120227849,
0.0168243106,
0.0242360886,
0.0153698921,
0.025325235,
0.1091701314,
-0.0117063997,
-0.1117640436,
-0.0218126867,
-0.0442841649,
0.0306789428,
-0.0068104197,
-0.0105782654,
-0.0316076428,
-0.0628541857,
-0.0588710085,
-0.0797395334,
-0.0753547996,
-0.0026723915,
0.0247867461,
0.0575870462,
0.0192401428,
-0.0255533252,
0.0021797528,
0.0194137134,
-0.0048415065,
0.0337425135,
0.0443260893,
-0.0181196034,
-0.0170543622,
0.0247426312,
0.0245858002,
-0.0340816826,
0.0079253325,
0.0342298485,
0.0065572327,
0.0546377152,
-0.024854416,
0.0361964777,
0.0224570502,
-0.0786792114,
0.0367170833,
0.0055498714,
0.0688712001,
-0.0421324372,
-0.0419252329,
0.0046594227,
0.0881581232,
-0.0120353,
-0.0137606161,
0.0148141924,
-0.0978782699,
-0.0224440135,
-0.0074945884,
-0.0005451147,
0.0230412763,
-0.0165021531,
-0.0112954332,
0.0528930575,
-0.006774561,
0.0607284419,
-0.0441837832,
0.0084974989,
-0.0676357001,
2.561e-7,
-0.0619373284,
0.0235437527,
0.0274694581,
-0.0051340503,
0.0096992906,
0.0542498939,
0.0548265316,
0.0164288189,
0.0063233506,
-0.0080687907,
-0.0255189929,
0.028582314,
0.0095531614,
0.0083057405,
0.0266405344,
-0.0297345929,
-0.1061932594,
-0.0262962934,
-0.023885034,
-0.0013896525,
-0.0164928101,
0.0356241055,
0.0681562051,
-0.0202155989,
-0.0456187315,
0.0214285087,
-0.0113513963,
-0.0424593352,
-0.0283900592,
-0.0180406999,
0.0017703596,
0.0325546488,
0.0181476269,
-0.0340204611,
0.0164784472,
0.0562547483,
-0.0322889872,
0.0442272872,
0.0151017802,
-0.0132596428,
-0.0300387498,
0.0117805721,
-0.0299323611,
-0.0114015955,
-0.0246444661,
0.044067651,
-0.0223006681,
-0.051705841,
-0.0476988591,
0.0021318893,
0.0234778449,
-0.0105957082,
-0.0360565297,
-0.0323332585,
0.0325889662,
0.0241181497,
0.0198879801,
-0.0486188345,
0.0020263477,
-0.018454425,
-0.0063999137,
-0.0156365335,
-0.0237982068,
-0.0083425501,
0.0170412995,
0.0222930945,
0.0501219146,
1.89442968e-34,
0.0250564143,
0.0212801546,
-0.0159522425,
-0.041591391,
0.0307135507,
0.0242688339,
-0.0000908347,
-0.022618847,
0.0652503595,
0.0348682255,
-0.0517379902
] |
http://arxiv.org/abs/2401.14589v2 | http://arxiv.org/pdf/2401.14589v2 | Enhancing Diagnostic Accuracy through Multi-Agent Conversations: Using Large Language Models to Mitigate Cognitive Bias | 2024-01-26T01:35:50 | 2024-05-12T05:28:23 | Yu He Ke, Rui Yang, Sui An Lie, Taylor Xin Yi Lim, Hairil Rizal Abdullah, Daniel Shu Wei Ting, Nan Liu | cs.CL | cs.CL, cs.AI | Background: Cognitive biases in clinical decision-making significantly
contribute to errors in diagnosis and suboptimal patient outcomes. Addressing
these biases presents a formidable challenge in the medical field.
Objective: This study explores the role of large language models (LLMs) in
mitigating these biases through the utilization of a multi-agent framework. We
simulate the clinical decision-making processes through multi-agent
conversation and evaluate its efficacy in improving diagnostic accuracy.
Methods: A total of 16 published and unpublished case reports where cognitive
biases have resulted in misdiagnoses were identified from the literature. In
the multi-agent framework, we leveraged GPT-4 to facilitate interactions among
four simulated agents to replicate clinical team dynamics. Each agent has a
distinct role: 1) To make the final diagnosis after considering the
discussions, 2) The devil's advocate and correct confirmation and anchoring
bias, 3) The tutor and facilitator of the discussion to reduce premature
closure bias, and 4) To record and summarize the findings. A total of 80
simulations were evaluated for the accuracy of initial diagnosis, top
differential diagnosis and final two differential diagnoses.
Results: In a total of 80 responses evaluating both initial and final
diagnoses, the initial diagnosis had an accuracy of 0% (0/80), but following
multi-agent discussions, the accuracy for the top differential diagnosis
increased to 71.3% (57/80), and for the final two differential diagnoses, to
80.0% (64/80).
Conclusions: The framework demonstrated an ability to re-evaluate and correct
misconceptions, even in scenarios with misleading initial investigations. The
LLM-driven multi-agent conversation framework shows promise in enhancing
diagnostic accuracy in diagnostically challenging medical scenarios. | Enhancing Diagnostic Accuracy through Multi-Agent Conversations: Using Large Language Models to Mitigate Cognitive Bias: Background: Cognitive biases in clinical decision-making significantly
contribute to errors in diagnosis and suboptimal patient outcomes. Addressing
these biases presents a formidable challenge in the medical field.
Objective: This study explores the role of large language models (LLMs) in
mitigating these biases through the utilization of a multi-agent framework. We
simulate the clinical decision-making processes through multi-agent
conversation and evaluate its efficacy in improving diagnostic accuracy.
Methods: A total of 16 published and unpublished case reports where cognitive
biases have resulted in misdiagnoses were identified from the literature. In
the multi-agent framework, we leveraged GPT-4 to facilitate interactions among
four simulated agents to replicate clinical team dynamics. Each agent has a
distinct role: 1) To make the final diagnosis after considering the
discussions, 2) The devil's advocate and correct confirmation and anchoring
bias, 3) The tutor and facilitator of the discussion to reduce premature
closure bias, and 4) To record and summarize the findings. A total of 80
simulations were evaluated for the accuracy of initial diagnosis, top
differential diagnosis and final two differential diagnoses.
Results: In a total of 80 responses evaluating both initial and final
diagnoses, the initial diagnosis had an accuracy of 0% (0/80), but following
multi-agent discussions, the accuracy for the top differential diagnosis
increased to 71.3% (57/80), and for the final two differential diagnoses, to
80.0% (64/80).
Conclusions: The framework demonstrated an ability to re-evaluate and correct
misconceptions, even in scenarios with misleading initial investigations. The
LLM-driven multi-agent conversation framework shows promise in enhancing
diagnostic accuracy in diagnostically challenging medical scenarios. | [
0.0663019046,
0.0260137394,
-0.0191106256,
-0.0146373156,
-0.0160884019,
0.0168027971,
0.0351842865,
0.0069160145,
0.0239488296,
0.0179907624,
0.001058462,
-0.022362303,
0.0000936883,
0.0041675707,
-0.0093253003,
-0.0426654965,
0.0190971456,
0.0032704254,
0.1039096937,
-0.0068930858,
-0.000418173,
0.0210529547,
-0.0378246978,
0.0417501628,
0.0019540866,
-0.0376510248,
0.0245487764,
-0.0279300977,
0.0305644833,
-0.0827682316,
-0.0185644273,
0.0225891937,
-0.0004515091,
0.0127970539,
0.0000024724,
-0.0585158505,
-0.0051804935,
-0.0129006468,
-0.0350516513,
-0.0382608324,
0.0537910238,
0.0188328959,
0.023654826,
0.0163153801,
-0.0327596143,
-0.0354465693,
0.0416825972,
0.0279826988,
0.0189174227,
0.0265041534,
-0.0097884452,
0.0184823293,
-0.0268973764,
0.0121603077,
-0.0085498262,
-0.0904617906,
0.0343904756,
0.0341228545,
0.0384605117,
-0.0400566347,
0.0220266655,
-0.0148218423,
-0.0109303612,
0.0066296081,
-0.0720993057,
0.045396734,
0.0076790792,
-0.0555515029,
0.0153562417,
0.0513809882,
0.0492940843,
0.0132154627,
0.011895786,
-0.0285486318,
0.0366813913,
0.0254653003,
-0.0386450179,
0.025672771,
-0.0353082046,
-0.0071895374,
-0.0042736544,
0.0145624513,
0.0155318966,
-0.0012996329,
-0.0163941029,
0.0165718216,
-0.020003397,
0.0004922443,
-0.0086968858,
0.0088705886,
-0.0179359466,
-0.0322089233,
-0.0072128596,
0.0301599503,
-0.1044323742,
0.0117280837,
0.0446892045,
0.012762269,
0.0723359063,
-0.080741778,
0.0122307027,
-0.0022198122,
0.0565711074,
0.0070124725,
0.0191363394,
-0.0272723902,
-0.0279362183,
0.0289720967,
-0.0140820127,
0.0150652146,
-0.0364792831,
-0.0181847773,
-0.0334160104,
0.0028080444,
0.0526165403,
-0.0096566714,
-0.05220332,
-0.0167923756,
-0.0560849532,
0.0310232546,
-0.0376845524,
0.0360941179,
-0.0086753964,
0.0132235521,
-0.0658405572,
0.0449738801,
-0.0097895898,
0.0316470303,
0.0164114796,
-0.0333270654,
0.01430263,
-0.0041563627,
-0.027093716,
0.0182077996,
0.0202455558,
0.0522270612,
-0.0152901309,
0.0409490615,
0.0296379309,
-0.0428564847,
-0.0180547237,
-0.0576909706,
-0.0432100408,
-0.054555472,
-0.0058039287,
0.0349185392,
0.0134645039,
-0.0430654027,
0.0228781421,
0.0013728504,
-0.0330985188,
0.0512029305,
-0.0693910271,
0.0257709455,
0.0372786969,
0.055383686,
-0.0375657417,
0.0165231694,
0.007438051,
-0.0299761668,
-0.0198113155,
-0.0204003304,
0.0494336188,
-0.011373084,
0.017274145,
0.0092156809,
0.0228249468,
0.0114177493,
-0.0034486305,
0.0974591747,
-0.0564756133,
0.0515712239,
-0.0066294083,
0.0447222739,
0.0550045632,
0.0302686356,
-0.034409713,
0.0069103548,
-0.0103638489,
-0.0031118314,
-0.0186714474,
-0.0103403982,
0.0181475785,
0.0172235891,
0.0271004457,
-0.0363958031,
0.0086126076,
0.0232437328,
-0.0353329964,
-0.024337003,
-0.0379021615,
0.0143150045,
0.0083467551,
0.0235293247,
-0.0301521886,
-0.0383263081,
-0.0184258837,
0.0485466309,
0.0215229373,
0.009494232,
-0.0669166297,
-0.0893410295,
0.0052782334,
0.0790556446,
0.0279565249,
-0.0063723684,
-0.0679537505,
-0.0269528609,
-0.0058198688,
0.013425773,
0.0517781861,
-0.0127719948,
-0.0397304296,
-0.0356833749,
-0.0114428839,
0.0368366241,
0.0121192643,
0.0067213643,
-0.037567582,
0.0356651619,
-0.0540452227,
0.0267636515,
-0.031690497,
0.0236784108,
-0.0287833754,
-0.0105570778,
-0.0328426473,
-0.0123880478,
0.0559762642,
-0.0061359289,
-0.0177138988,
0.0225881003,
-0.012285592,
0.0321415104,
-0.0572674647,
0.0684933215,
-0.0382418409,
0.0128070209,
0.00188825,
0.0107017951,
-0.0272454098,
-0.04637786,
-0.0340314955,
-0.0024210273,
0.0231579449,
-0.0348093659,
0.0076022125,
0.0350707062,
-0.0028160936,
0.0101852613,
0.0043422147,
-0.0142395729,
0.0054152277,
0.0128792543,
-0.0288038515,
0.0261147711,
-0.0246520471,
-0.0008939522,
-0.0393609293,
0.0175986197,
0.0444169678,
0.0350018814,
0.0191946141,
0.0041682725,
-0.0265071914,
-0.0259116925,
-0.0052085659,
-0.1207992285,
-0.035398405,
0.0298545379,
-0.0338730738,
-0.0022669625,
-0.0115088355,
-0.0133434013,
-0.0157712679,
-0.0191861391,
0.0151688233,
0.0542906485,
0.05833368,
0.0278690588,
0.0479746722,
-0.0124569125,
-0.0215629227,
0.0234469436,
-0.0493174717,
0.032137271,
0.0709615275,
0.0034287034,
-0.0290871281,
-0.0082432041,
-0.0695005804,
0.0064892494,
0.0080156587,
-0.0102188373,
0.0375976674,
-0.023813881,
-0.0364125744,
0.049672395,
-0.0041415417,
0.051852379,
0.0483695343,
-0.0224007126,
-0.0594727732,
0.0039703343,
-0.0167466309,
-0.0288875662,
0.0003870483,
-0.0494546369,
-0.0327104032,
0.0876868665,
0.0338921174,
0.0075756358,
-0.009333604,
0.0452623665,
-0.0086202519,
-0.0178010091,
0.0018889597,
-0.0776203573,
-0.0039590388,
0.0327631347,
-0.0040321001,
0.0010498076,
0.0977139249,
0.0217067003,
0.0323893577,
-0.0925293416,
-0.0108326869,
-0.0356266499,
0.0104038045,
-0.0259557925,
0.0127762873,
0.0357631184,
0.0065658903,
-0.0052390196,
-0.0498155355,
0.0364512689,
-0.0352401845,
0.047618676,
-0.0870378613,
0.023439914,
0.0114015443,
-0.0100012599,
-0.0176916178,
0.0243890472,
-0.0333992839,
-0.0052115414,
0.00598728,
0.0263157114,
-0.0918403491,
0.0089981584,
0.0032166382,
0.0410913453,
0.0280062929,
-0.0242424719,
0.0176584106,
-0.0971937701,
-0.018137373,
0.0155228898,
0.0417662598,
-0.06029322,
0.0781745687,
0.0650437325,
-0.0071310676,
-0.0113852043,
-0.0137675665,
-0.0400040522,
-0.018041959,
-0.0014616252,
0.0044958116,
0.0229464285,
0.0266173892,
-0.038787175,
-0.0114026479,
0.0800759569,
0.0077600763,
-0.0921989754,
0.0027761753,
-0.0502405502,
-0.0501653291,
-0.0286741108,
-0.0162455514,
-0.0603103228,
-0.0204577707,
-0.0115153277,
-0.0012477115,
-0.053383112,
-0.0495557413,
0.0154006314,
-0.0237535331,
0.0483785644,
0.0184338484,
0.0315281898,
0.0122890063,
-0.0187726524,
0.0220818575,
-0.0036615864,
0.0514195524,
-0.0331419371,
0.0556839928,
0.0052375994,
-0.0066758487,
-0.0990001559,
-0.0479896739,
0.0794369429,
-0.0024103576,
0.0457955189,
0.01045503,
0.0341393054,
-0.0103899166,
0.0012468925,
-0.005171794,
0.032030955,
0.0220008604,
0.0035233991,
0.0383446515,
0.0025442543,
0.0430172235,
-0.0509238914,
-0.0004544993,
-0.0022994489,
-0.00073176,
0.0285370369,
0.0917223394,
-0.0513864607,
-0.0662058368,
-0.0228661187,
0.0469619706,
0.0494935364,
0.011880354,
0.0019555816,
-0.0008722591,
-0.0124326712,
0.0359741077,
0.0424978398,
0.0521039069,
-0.0014643562,
0.0032361269,
0.0190431364,
0.019410681,
0.0030366094,
-0.0484545752,
0.0410768539,
-0.0269376859,
0.0202242397,
0.0292732082,
0.0060314671,
0.0363937169,
-0.0198287331,
0.0443935692,
0.0796411484,
0.0559858866,
0.0103830071,
-0.0144927585,
0.033608783,
0.0040681981,
0.0376500189,
0.0350828022,
0.0250945464,
0.0009450843,
0.0037772877,
0.0216975044,
0.0312748402,
-0.0199623946,
-0.0040884223,
-0.0311446581,
-0.0171111561,
-0.0087886807,
-0.0368453972,
-0.0103918808,
-0.0251733176,
0.0499408506,
0.0113010779,
-0.005048356,
-0.0009743355,
0.0357936248,
-0.0154185705,
0.0203134976,
-0.0856203362,
-0.0344079062,
0.0564787462,
-0.0101211434,
0.0261523742,
-0.023047898,
0.00698986,
-0.0108438022,
-0.0102461455,
-0.0330530033,
-0.0154901724,
0.0173796471,
-0.0263024084,
-0.0092322668,
-0.0345990546,
0.0255799349,
-0.0301169399,
-0.0527999327,
-0.0669588894,
-0.0054744128,
0.0758161619,
0.0344327949,
-0.0043575582,
0.0539320558,
-0.0036058971,
-0.0005128846,
-0.0421655141,
0.0029036722,
0.0212938376,
0.0456032567,
0.0997171029,
0.0093919318,
-0.0139968507,
0.0165585112,
-0.0030178309,
-0.0308886226,
-0.088516891,
0.0139070107,
0.0149626043,
0.000807899,
0.0661864206,
-0.0073848488,
-0.0544149764,
0.0233761109,
0.0028063231,
-0.0304734129,
-0.0215750281,
-0.0132548772,
0.0321876369,
0.0031920329,
-0.0117913578,
-0.0223699249,
0.0420703962,
-0.0416208245,
0.0388780124,
-0.0724845305,
-0.0429402478,
-0.0433887318,
-0.031388294,
0.0123606343,
0.0193113089,
-0.0465776734,
0.0140559012,
-0.0982671678,
0.0002686644,
-0.0182850063,
-0.0045614755,
-0.011877764,
-0.065073967,
-0.0072035501,
0.0617860928,
0.0099539142,
-0.0348221287,
-0.0151787456,
0.0627602935,
-0.0297079757,
0.034650024,
-0.0278307982,
-6.108698105e-33,
-0.0187056158,
-0.0715219826,
0.0365720131,
-0.013819824,
0.0074858302,
0.0009331288,
-0.0534364283,
0.0385811776,
-0.0054379981,
0.0620375089,
-0.0164247993,
0.0026226777,
0.0012019306,
0.0084537938,
0.0342783555,
-0.0077184741,
0.0373411775,
0.0095700063,
-0.0393048003,
-0.0023372411,
-0.022165738,
0.0029272041,
-0.0054268055,
-0.0737989768,
0.0245230123,
0.0014234008,
0.0074213762,
0.0026492199,
-0.0550969169,
0.0203467607,
0.0241939966,
0.0374270603,
0.0323238,
-0.0008811632,
-0.028776437,
-0.0004126981,
0.0180289112,
-0.0306475572,
0.0339600369,
-0.0274297949,
-0.0029628673,
0.0008991002,
0.0102634197,
0.0059908363,
-0.016933158,
0.024633836,
0.0048595937,
-0.0341377668,
0.0159885846,
0.0339895152,
-0.0362664163,
0.0300825629,
-0.0078288633,
0.027170537,
-0.0222295802,
-0.0033152876,
-0.0587413013,
-0.0247211177,
-0.052073434,
0.0220168065,
0.0311302766,
-0.046967525,
0.0061364882,
-0.0074907863,
-0.029954005,
0.0645091459,
0.061258547,
0.0314959139,
-0.092784144,
-0.034073554,
-0.0782934427,
0.0222004578,
0.0647824556,
0.061725758,
-0.0417430773,
-0.0026380648,
-0.0548989214,
0.0019231278,
-0.0023432304,
0.0439835638,
0.0598289818,
-0.0102989618,
-0.0844410583,
-0.0509105176,
-0.0275238585,
0.0272721667,
-0.0251465961,
-0.051402282,
0.0136943934,
0.0049150148,
0.0258995537,
-0.0069417031,
0.019610472,
0.0062336442,
0.0117811719,
-0.0655408129,
0.0569050498,
-0.0044027409,
-0.0277315527,
-0.0141161513,
-0.0665520877,
0.0054666195,
-0.0207426064,
0.0163747743,
0.0463252142,
0.0340133533,
-0.026156107,
-0.0238011517,
-0.0189828202,
0.0224128719,
-0.0000206446,
0.0243841019,
-0.0125248879,
-0.0309296362,
0.0047822678,
0.024993021,
-0.0223298129,
0.0075477613,
0.0124328947,
-0.0102120778,
-0.0118774939,
-0.012627814,
-0.0567566641,
-0.0086943721,
-0.042358879,
-0.0107361423,
-0.0324596539,
-0.0022804777,
0.0305612292,
-0.0320534669,
-0.0042508948,
-0.0154113807,
3.19e-7,
0.0458197407,
0.0651667863,
-0.0098804682,
-0.0092198141,
0.0622105598,
0.0449234508,
-0.0101534333,
0.0242270716,
0.0491576605,
-0.0308525022,
0.0419960618,
0.0417284071,
-0.0111033665,
0.0603933372,
-0.0081150504,
-0.0428227745,
-0.0350562148,
0.0192256439,
-0.0107693905,
-0.0263942704,
0.049228739,
-0.0043394724,
-0.0405644961,
-0.0181373116,
-0.0010436361,
0.0202274416,
0.0261377804,
-0.0401087813,
0.0067674187,
0.060167674,
-0.0010575232,
-0.0483541302,
0.0302452538,
0.1165544316,
0.0160022117,
-0.0328105986,
-0.0299910735,
0.0357934497,
-0.0588244796,
0.0098026879,
0.0053891139,
-0.0189234558,
0.0479359925,
-0.0555574521,
-0.0069092656,
-0.1027262956,
-0.0648498237,
0.0265087113,
-0.0210322123,
-0.0424509756,
0.0726039335,
-0.0272346102,
-0.0335069224,
0.0187488478,
-0.0234177355,
0.0818284154,
0.005562657,
-0.0465671122,
0.0144268367,
0.0075157294,
-0.0429801382,
-0.0803641006,
0.0470422246,
-0.0555078052,
-0.0037154218,
-0.0585625246,
0.0478391834,
2.477455134e-34,
0.0302935876,
0.0320372917,
-0.0050747269,
0.001579519,
0.0483792201,
-0.0002284031,
0.008969523,
0.0271236654,
0.0647533834,
-0.0244603399,
-0.0176094919
] |
http://arxiv.org/abs/2401.14569v1 | http://arxiv.org/pdf/2401.14569v1 | Detecting Structured Language Alternations in Historical Documents by Combining Language Identification with Fourier Analysis | 2024-01-25T23:54:34 | 2024-01-25T23:54:34 | Hale Sirin, Sabrina Li, Tom Lippincott | cs.CL | cs.CL | In this study, we present a generalizable workflow to identify documents in a
historic language with a nonstandard language and script combination,
Armeno-Turkish. We introduce the task of detecting distinct patterns of
multilinguality based on the frequency of structured language alternations
within a document. | Detecting Structured Language Alternations in Historical Documents by Combining Language Identification with Fourier Analysis: In this study, we present a generalizable workflow to identify documents in a
historic language with a nonstandard language and script combination,
Armeno-Turkish. We introduce the task of detecting distinct patterns of
multilinguality based on the frequency of structured language alternations
within a document. | [
0.0088725593,
0.0882289559,
-0.0574747771,
0.0604680702,
-0.0534685291,
0.0268270131,
-0.0064441743,
0.0517426059,
-0.0397639908,
0.0037268139,
0.0316820405,
-0.053972926,
0.0505057909,
-0.0325435437,
0.0190923437,
0.0268754717,
0.0525462441,
-0.0224067792,
-0.0352888703,
0.05188106,
0.0121895652,
0.0171461049,
0.0006662036,
-0.0242095795,
-0.0173598714,
0.0014243837,
-0.0132387755,
-0.0105154151,
0.0518382601,
0.0380442962,
-0.0100585166,
0.0438353419,
-0.0341873355,
0.0976433679,
0.0000012009,
-0.0608813874,
-0.0422469974,
-0.0264680739,
0.0288343988,
-0.0061993208,
0.0452804826,
-0.0169681311,
-0.0518825129,
-0.0534413047,
0.0126469983,
-0.0371348709,
0.0355155617,
0.0218772963,
-0.0145757282,
0.0542049669,
-0.009283388,
-0.0483107194,
-0.0263204258,
0.0432256348,
0.0167983752,
0.024897201,
0.0030632478,
-0.0655218959,
-0.0157317352,
0.0006253338,
0.0134475632,
0.0556241944,
-0.0665815547,
-0.0178261586,
-0.0682197437,
-0.0350001417,
-0.0718952417,
-0.0013100622,
0.0136287734,
-0.0178810954,
-0.0300875008,
-0.0566194616,
0.0173309743,
0.0295665078,
-0.0367938206,
0.0359951332,
-0.0065108771,
-0.025601266,
-0.0748388469,
-0.016155893,
0.0802144259,
0.0014566077,
-0.0076899761,
-0.0077938009,
-0.0166909993,
0.0728357136,
0.0474759229,
-0.0247637238,
-0.0016205275,
-0.0315799713,
0.0154153602,
-0.005270089,
0.0252813492,
0.0186689515,
-0.0013581041,
0.0054509523,
0.0023311623,
-0.1194388866,
0.01000016,
-0.0616781525,
0.0266848002,
0.0015943366,
-0.0823532492,
0.0354871564,
-0.0466563068,
0.0356229283,
0.0107353562,
-0.1022012085,
-0.0204695184,
0.0135630509,
-0.0136498874,
-0.0013094567,
0.0036067048,
-0.0584569797,
-0.0267489869,
0.0406414717,
-0.0061991587,
0.0091684265,
0.0366033502,
0.0284988489,
-0.0455077775,
-0.0626729429,
-0.0117684714,
0.0043379669,
0.0219344441,
0.0038032939,
0.0037594407,
-0.0256629586,
-0.0922108442,
-0.0316163935,
0.0237729214,
0.0685785413,
0.0154067464,
-0.0173508544,
0.001036189,
0.0173322335,
-0.0309270229,
0.0144770984,
-0.0325438008,
0.0021553768,
-0.0240595415,
-0.018933678,
0.0264578853,
-0.0603346191,
0.0148930689,
-0.0275923256,
0.0050015901,
-0.0033407437,
0.0008139458,
-0.0037244344,
0.0045698415,
0.057727173,
-0.0047520124,
0.0087166401,
0.0667069107,
-0.0330426358,
0.0449738279,
0.052089233,
-0.0072215125,
-0.0402603932,
0.0606873259,
-0.0455710813,
0.0080446964,
0.074704282,
-0.0446421318,
0.0024337703,
-0.0199585836,
-0.0309393406,
0.0212919991,
-0.0124051673,
-0.0198855344,
0.0462254994,
0.0244425461,
-0.0356496796,
-0.0124254636,
0.0718905255,
0.0249107368,
0.0236822851,
-0.0047506597,
-0.0028860199,
0.0230810493,
0.0456623062,
0.0690577552,
-0.0527669862,
-0.0116355494,
-0.0331377946,
-0.0570791736,
0.028711509,
0.0120401308,
0.0412845872,
-0.0181676224,
-0.0120165106,
0.0132274665,
-0.0447618105,
0.0740469247,
-0.0220883489,
-0.0168191213,
-0.0458442159,
0.0146717904,
-0.0103339693,
0.0201912485,
-0.0298364554,
0.0208651125,
0.0666991174,
0.0541975535,
-0.0729305372,
0.076049082,
-0.004320397,
-0.0242011603,
0.0442941822,
0.0065489896,
-0.0438105203,
0.0551765859,
-0.0330797024,
0.0458081327,
0.0527882092,
0.0314702205,
0.0491838641,
-0.0430499986,
0.0398305878,
-0.0037914719,
-0.0510592386,
-0.0183034856,
0.0120024364,
-0.0244681314,
-0.0415736847,
0.0227739885,
0.0212165844,
-0.0039175567,
0.0456971228,
-0.0132230213,
0.0148446308,
0.0003815678,
-0.0426255129,
0.0396298207,
0.0329408348,
0.0370015278,
0.0229800567,
0.054635521,
-0.0124741197,
-0.0197719354,
-0.0325183608,
0.0141725298,
-0.0424814783,
0.0751343668,
-0.0629047453,
0.03947936,
-0.0013346291,
0.0592303872,
-0.032310009,
0.0016923536,
-0.0641043782,
0.0306837,
0.0486586057,
0.0371979922,
0.0014874282,
-0.0859829709,
0.0462150835,
0.0082106655,
0.0081472527,
0.01422102,
0.0007713825,
-0.0058965222,
0.0141870147,
-0.0283387806,
0.0044343905,
-0.0547704175,
-0.0410724506,
-0.0496486612,
-0.044254519,
0.055184193,
-0.0012686264,
0.0139872339,
0.0069494727,
-0.0046901102,
-0.0031088491,
-0.0310710948,
0.0308126453,
-0.0180991814,
0.045188427,
-0.0256572329,
-0.0638623238,
-0.0140295411,
-0.0083884504,
0.0059581231,
0.0352459848,
0.0178702101,
-0.0816285014,
-0.0751420707,
-0.0660623908,
0.0010735553,
0.0167421754,
-0.02164956,
0.0009694676,
-0.0230337698,
-0.000761269,
0.0531257018,
0.0319231823,
-0.0163985435,
0.0050385697,
0.0053652897,
-0.0508667678,
0.0233695973,
-0.0509199984,
0.0001615836,
0.0092588915,
-0.0139489165,
-0.0428452604,
0.0493973419,
0.0220259223,
-0.0179148372,
0.0191809647,
0.0084247123,
-0.0423729829,
-0.0317967124,
0.0165149774,
0.0127081797,
-0.0151951937,
-0.0461059213,
0.0724281743,
0.0330875814,
0.0025212588,
0.0160190836,
-0.0276427232,
0.0082087358,
-0.0957667604,
-0.0105117066,
-0.0402155332,
0.0185622126,
-0.0106398799,
0.0092604766,
0.0246439967,
-0.0521174446,
-0.03366163,
0.000016152,
0.043039117,
-0.0265779272,
0.0913497284,
-0.0493646339,
0.0188088324,
-0.0192840844,
0.0264869593,
-0.0203246847,
-0.0141482065,
0.0175608918,
-0.0200768393,
0.0376246534,
-0.0430813469,
-0.0251769703,
0.0432140045,
0.0118287802,
0.0478866585,
0.0122242039,
0.0334330238,
0.0081364037,
0.0841522068,
-0.0078916084,
-0.0396144018,
-0.0141389817,
-0.00229755,
-0.0177565329,
0.008179931,
-0.0059223147,
0.0015514948,
0.0006088393,
0.0016957195,
0.0070203142,
-0.0093750972,
-0.0227767341,
-0.061255224,
-0.016530944,
-0.0011640004,
0.0204426907,
0.0553214252,
0.0335315317,
-0.0219274741,
0.0054108473,
-0.0237911083,
-0.0090343859,
-0.0247379802,
0.020471666,
-0.0541577786,
0.0347191393,
0.0186561886,
-0.0067279083,
0.0015339947,
0.0089441836,
0.0143777393,
-0.0382021368,
-0.0147828711,
0.0242889114,
-0.0420078635,
-0.0386304222,
-0.0184606872,
0.0190918818,
-0.0017763164,
-0.0599435307,
0.0307493191,
-0.0018387405,
-0.0080028065,
0.0168452412,
-0.0449080057,
-0.056121815,
-0.0509772748,
-0.0137835834,
-0.0388740338,
0.0057799714,
0.0234543849,
-0.0853212476,
-0.0331261307,
0.0269797649,
0.0008701388,
-0.0001422464,
-0.0096276114,
0.0378948934,
0.0529699102,
-0.0322285816,
-0.0243169311,
-0.0018779311,
-0.0172052272,
-0.030186031,
0.0427513756,
0.0191168599,
0.0060325889,
0.0019006109,
0.0253497083,
0.0059044305,
0.0263845213,
-0.0505219847,
0.0101444665,
-0.0113514299,
0.0159234833,
-0.0595133416,
0.0539625101,
-0.0244126078,
-0.021410536,
-0.0084641622,
0.0293340422,
-0.0710908771,
-0.0975698456,
0.0107178269,
-0.025906926,
0.0195487402,
0.054831598,
0.0036364046,
0.0062032863,
-0.0038005169,
-0.0585243106,
-0.0468035862,
-0.0021012179,
0.0247810334,
-0.0707964003,
-0.0471589603,
0.0089772046,
0.0183709897,
0.0130384304,
-0.0087205721,
-0.1368348598,
-0.0323876739,
0.0554476418,
-0.0479408391,
0.0257666241,
-0.0014009367,
0.0058209542,
-0.0053017503,
-0.0404394604,
-0.0083138598,
-0.0111582149,
0.0459876917,
-0.0114665264,
-0.0465151481,
-0.0353710949,
0.0142205022,
0.0026764965,
-0.0005732619,
0.0295150392,
-0.0413350984,
-0.0214286577,
-0.0125268269,
0.0040390803,
0.0259529948,
-0.0255338717,
-0.0369934365,
-0.0047997655,
0.0011875796,
-0.0098118559,
0.0126388893,
-0.0040943297,
0.0544136278,
-0.0204004757,
-0.023444254,
-0.0109115988,
-0.0172311384,
0.0367796123,
-0.0200083759,
-0.0108935153,
0.007804648,
0.0353215747,
0.0457089916,
0.0447429903,
-0.0079672765,
0.0294343755,
0.0007336664,
0.0871992484,
-0.0137591669,
0.0843903944,
0.0056705377,
-0.0169508047,
-0.0231100023,
0.0151853096,
-0.0152772302,
0.0110140368,
-0.0070243357,
0.0050797858,
0.026458662,
-0.0251447055,
0.0602909923,
-0.0589260086,
0.0913328975,
-0.0085019963,
-0.0172015373,
-0.0289075822,
-0.0038939614,
0.0253760237,
0.0621326044,
-0.0635547116,
0.0042496151,
-0.0291434005,
0.0377113037,
0.0732435659,
0.0024246234,
-0.0185874868,
0.0103424145,
-0.0393975191,
0.0014521693,
0.0166496281,
0.0127176717,
-0.0266699754,
-0.0308028664,
-0.0019589011,
-0.0518034883,
-0.0295363013,
0.0140418205,
0.0082051661,
0.0365026668,
-0.0035918602,
-0.0142729897,
0.0682526082,
0.0321996175,
0.0717783347,
-0.0285199303,
0.0306217335,
0.0300604608,
-0.0284542404,
-0.0405009538,
-4.515312206e-33,
0.0256374963,
-0.0007384978,
-0.0668121055,
0.0053416272,
0.0430262648,
-0.0364936106,
0.0000923904,
-0.0117900632,
0.0072822771,
-0.0111116758,
-0.0167557131,
-0.0096899131,
0.0379705355,
-0.0356299281,
0.0525093824,
0.0116815697,
0.0096328957,
0.0257762801,
-0.0095795915,
-0.0096885785,
0.0918107405,
-0.0234039221,
0.1166495383,
-0.0811108127,
0.0069092507,
-0.0144679137,
0.0071231551,
-0.00446419,
0.0178583339,
0.0223332308,
0.0010566894,
0.0054123146,
-0.0004882616,
0.0136171412,
-0.006015562,
-0.0089416811,
0.0036922572,
-0.0211731847,
0.0272254404,
0.1026872322,
0.0593894348,
-0.018563332,
-0.0299777482,
-0.0422758497,
-0.0358416028,
0.0240879543,
0.0493758023,
-0.0291522145,
0.0036165635,
-0.0314965472,
-0.0692114383,
0.0519785024,
-0.0071560252,
0.0425265171,
0.042570103,
0.0169643257,
0.0047304532,
-0.0401660018,
-0.0383202359,
0.0282898583,
0.0289416611,
0.0310315602,
-0.0015783644,
0.0313407965,
-0.0613276847,
0.0197341479,
0.0427563936,
0.0215688962,
0.0448443927,
0.0543333739,
0.0240838323,
-0.0349989906,
0.0050985985,
0.0430258848,
0.0482157543,
-0.0348465331,
-0.0028887223,
0.0049335188,
0.0618470348,
-0.0109817423,
-0.0062287515,
-0.0090127047,
-0.0164332837,
-0.0287273694,
0.0166070722,
-0.0252565779,
0.003478029,
0.0202421583,
-0.0050641703,
-0.0289947055,
0.0056795236,
-0.0034960723,
0.0227940455,
-0.0370420702,
-0.0266837385,
-0.0393418968,
-0.0181362275,
0.0547399782,
-0.0236714631,
0.0019823925,
0.0162009802,
-0.0406055227,
0.0410631187,
0.014644756,
0.0578859784,
-0.0240045823,
0.0271334592,
0.0050273677,
-0.0679529682,
0.0210968722,
0.0241109971,
-0.025896715,
0.0175856426,
0.0537852384,
0.0127677936,
-0.014143968,
0.0500940904,
0.0234290119,
0.056464307,
-0.0186690427,
-0.0168855619,
-0.0320478976,
-0.0267234128,
0.0180332679,
-0.0246688128,
0.0252422038,
0.0111527992,
-0.0078335973,
0.0452400856,
-0.0144134732,
0.0256083552,
-0.0072584781,
2.137e-7,
0.0116682211,
0.0073902872,
0.0513137244,
-0.0357821099,
0.0560659356,
-0.0155752953,
-0.0538152531,
-0.0024017273,
-0.036924459,
0.0253143366,
0.0119119044,
-0.0581607521,
0.0822888985,
0.0524966642,
0.0052143596,
-0.0322658755,
-0.0263089873,
0.0141476505,
-0.0717298761,
0.0223707128,
-0.0128685702,
0.0758116692,
0.0036980784,
-0.0243207086,
-0.0204507876,
0.0053999722,
0.0114100883,
-0.008444679,
-0.023824865,
0.0137490034,
-0.0253625084,
-0.0423984155,
0.0028743092,
0.0179212987,
-0.02712957,
0.0127530415,
0.045982182,
0.0355690122,
0.00955013,
0.0900698304,
-0.0089485263,
-0.0184405949,
-0.0210884959,
-0.0044668424,
0.0452387221,
0.1138040721,
-0.0048014312,
-0.0316378437,
-0.0272856914,
0.0158554986,
-0.0099013047,
-0.006111796,
-0.0590154007,
0.0748733878,
0.0265044756,
0.0193130877,
-0.0072082658,
-0.0249779373,
0.0010637493,
0.0262996424,
-0.0177398697,
-0.0099260435,
0.0288838875,
-0.0306891333,
-0.0033813019,
0.0522525012,
0.0061512641,
2.007763407e-34,
0.0418083668,
0.0219805222,
-0.0389336199,
0.0264234655,
-0.0140133323,
-0.0478665419,
-0.0066854577,
0.0779356509,
0.0034517916,
-0.0259885006,
0.0064244028
] |
http://arxiv.org/abs/2402.01715v1 | http://arxiv.org/pdf/2402.01715v1 | ChatGPT vs Gemini vs LLaMA on Multilingual Sentiment Analysis | 2024-01-25T23:15:45 | 2024-01-25T23:15:45 | Alessio Buscemi, Daniele Proverbio | cs.CL | cs.CL, cs.AI | Automated sentiment analysis using Large Language Model (LLM)-based models
like ChatGPT, Gemini or LLaMA2 is becoming widespread, both in academic
research and in industrial applications. However, assessment and validation of
their performance in case of ambiguous or ironic text is still poor. In this
study, we constructed nuanced and ambiguous scenarios, we translated them in 10
languages, and we predicted their associated sentiment using popular LLMs. The
results are validated against post-hoc human responses. Ambiguous scenarios are
often well-coped by ChatGPT and Gemini, but we recognise significant biases and
inconsistent performance across models and evaluated human languages. This work
provides a standardised methodology for automated sentiment analysis evaluation
and makes a call for action to further improve the algorithms and their
underlying data, to improve their performance, interpretability and
applicability. | ChatGPT vs Gemini vs LLaMA on Multilingual Sentiment Analysis: Automated sentiment analysis using Large Language Model (LLM)-based models
like ChatGPT, Gemini or LLaMA2 is becoming widespread, both in academic
research and in industrial applications. However, assessment and validation of
their performance in case of ambiguous or ironic text is still poor. In this
study, we constructed nuanced and ambiguous scenarios, we translated them in 10
languages, and we predicted their associated sentiment using popular LLMs. The
results are validated against post-hoc human responses. Ambiguous scenarios are
often well-coped by ChatGPT and Gemini, but we recognise significant biases and
inconsistent performance across models and evaluated human languages. This work
provides a standardised methodology for automated sentiment analysis evaluation
and makes a call for action to further improve the algorithms and their
underlying data, to improve their performance, interpretability and
applicability. | [
0.0602682233,
0.0537468046,
-0.0185516626,
0.0670893267,
-0.0559811033,
-0.0031801476,
0.0069856136,
-0.0022172928,
0.0296513736,
-0.0383745544,
-0.0283530448,
-0.0394149572,
-0.0332063735,
0.0289350096,
-0.0159505904,
-0.0293779876,
0.0175083876,
0.0196515508,
0.0249843318,
-0.0419341549,
0.0117103821,
0.0102129364,
0.0302733313,
0.0266414937,
-0.002906305,
-0.0076943641,
0.0266866907,
0.0264222547,
0.0241600126,
-0.012155937,
0.0257988237,
0.0176265519,
-0.0119163347,
0.0320408829,
0.0000019913,
-0.0561042503,
0.0144601557,
-0.0196004529,
0.0318330862,
0.0194372199,
0.1121899262,
-0.0086545693,
-0.0079851588,
0.0298213419,
-0.031769488,
0.0041336743,
0.0346010365,
0.0365896784,
0.064890027,
0.0560558774,
-0.0248529781,
-0.0033401544,
-0.035762649,
-0.009078349,
-0.0222025551,
-0.0229200236,
0.0302715544,
0.0008214819,
-0.0338179246,
0.0164118148,
0.0052601472,
0.0250833239,
-0.0240099356,
0.0108090769,
-0.046945069,
-0.0009103576,
-0.0473685004,
-0.0414915122,
0.0058905794,
0.0265839286,
0.0141899027,
-0.0064719627,
-0.0349790454,
-0.0054410575,
-0.0084885266,
-0.0335225947,
-0.0118870735,
0.0136750909,
-0.004359446,
0.0252058152,
0.0414810255,
0.0630055517,
0.0164315049,
0.0438191965,
-0.00110636,
0.0605182275,
-0.012151855,
-0.026067799,
0.0156321004,
0.0309135914,
-0.0456621833,
-0.0015713604,
0.0280704275,
0.0057397098,
0.0020304935,
0.0341661349,
0.024599608,
-0.0052151801,
0.0450435057,
-0.1382264644,
-0.008747994,
0.0569871143,
-0.0441741087,
0.0495329387,
-0.031663321,
0.010279173,
-0.000991757,
0.0173550136,
-0.0274505652,
-0.0270920526,
-0.0325834043,
0.0115000773,
0.0093801972,
0.0132425828,
-0.039565634,
0.0143956896,
-0.0613397956,
0.0076323031,
-0.0363637246,
0.0002491713,
0.0344209448,
0.0010108066,
0.0035988528,
0.0035237207,
-0.0242857356,
0.0625761375,
-0.0463794656,
0.0375931635,
-0.091274634,
-0.0472191423,
0.0239618514,
0.0486284643,
-0.0392260179,
0.0025479372,
0.0396392681,
0.1277959943,
0.0217466485,
-0.0141490465,
-0.0105115669,
-0.0110772057,
-0.0086788265,
-0.0733816922,
0.0063500218,
-0.0511796996,
0.0645460859,
0.0324382819,
-0.0006360738,
-0.0059154741,
0.0252773929,
0.0058828685,
-0.0403736457,
0.0328357331,
-0.0138482759,
0.0080296686,
0.0124087138,
0.0223357063,
-0.0515224487,
0.0718656257,
-0.0017812005,
0.0669516623,
0.040414799,
0.0198764205,
-0.0073502851,
0.0525782369,
-0.0291172434,
-0.0067681125,
-0.0155727249,
0.0253008157,
-0.0479836352,
-0.004590299,
0.0131085338,
0.0427713916,
0.0012980604,
-0.0275936704,
0.0142559875,
-0.0162154622,
-0.0325016156,
0.0627496243,
0.0004383512,
0.0058964505,
0.0151596684,
-0.0133542847,
-0.013166517,
0.0226251576,
-0.0211327709,
-0.0142843425,
-0.0174101219,
0.0393401943,
-0.0053027575,
-0.0540281087,
-0.0270023588,
-0.0154954782,
0.0424347855,
-0.0094461283,
0.0621856302,
-0.0485568196,
-0.0542410798,
-0.0098723825,
0.0513987578,
-0.0440399721,
-0.015248497,
-0.0351251252,
0.0013874279,
0.0782905146,
-0.0118577331,
-0.0108510843,
-0.0246830732,
-0.0188168548,
-0.036920689,
0.0444008783,
-0.0008668106,
0.0180455167,
0.0037618917,
-0.0345492773,
0.011635961,
0.0048078801,
0.0187930372,
0.03373117,
0.0061430959,
0.0425020866,
-0.0230654478,
0.0119818356,
-0.0129922675,
-0.0073215566,
0.0032566378,
0.0221859626,
0.0094450153,
-0.0151417842,
0.0409259945,
0.0334696621,
-0.0136308912,
-0.0191100966,
0.0164873339,
-0.0460559204,
-0.0617703088,
0.055186756,
-0.0091519933,
-0.0477625504,
0.0225133169,
-0.0390147716,
-0.0466004387,
0.004169988,
-0.0429973975,
-0.0046236962,
0.0238559693,
0.0188514628,
0.0308840275,
-0.0430728309,
-0.0156440623,
-0.0588607714,
0.0177007876,
-0.0054034018,
0.0526905209,
-0.0056408965,
0.0098784594,
-0.01237626,
-0.0366854332,
0.0259788446,
0.0382620245,
-0.0672691464,
0.0509594828,
0.0312136561,
0.0204032864,
0.0196357928,
0.0298034772,
-0.0152936261,
-0.0640969947,
-0.1335517019,
0.0016961356,
0.0053309076,
0.0295783337,
0.0338219889,
-0.011315682,
-0.0372649841,
0.0383567698,
-0.0183441956,
-0.0191750471,
-0.0068827248,
0.0132104624,
0.0283227693,
0.0059352838,
0.0203425735,
-0.0400109142,
0.0218791645,
-0.0191153679,
0.0129817305,
0.0296223722,
-0.0466976389,
-0.0359758325,
0.0625113919,
-0.0775040761,
-0.0210946687,
-0.0236272421,
-0.0683448836,
-0.025959786,
0.0185909048,
0.0226045381,
0.0228115153,
-0.0047416608,
0.0332645886,
0.04170046,
0.0047845268,
-0.0258587841,
0.0086897183,
-0.0330145024,
-0.0099479556,
0.0014505303,
-0.0656591952,
-0.0267977826,
0.0219806489,
0.0320630334,
-0.0061379317,
-0.0472472832,
0.0014337269,
-0.0652068481,
0.0162070692,
-0.0223527849,
-0.0073341415,
-0.0001312224,
0.0138772279,
0.0398289561,
-0.0139510706,
0.0737123713,
0.0013537471,
-0.0085827885,
-0.0704706311,
-0.0098683676,
-0.0590192601,
0.0452404469,
-0.0499661975,
0.0308545362,
-0.0542255193,
0.0282249507,
-0.010101052,
-0.0426046029,
-0.0193397142,
-0.0298902895,
0.0319307186,
-0.0228155907,
-0.024787005,
-0.0029984193,
-0.0299525205,
0.0183351319,
-0.050440263,
-0.0745114014,
-0.0052583497,
0.0386881456,
-0.013229643,
-0.0004764912,
0.0127143422,
-0.0082456963,
-0.0794332325,
-0.0232583471,
-0.0039545549,
0.0246809404,
-0.0078505697,
0.0223482121,
-0.0345351584,
0.0484441891,
0.0145262349,
-0.0088076219,
0.0347151682,
-0.0532528125,
-0.0207281541,
-0.0236527808,
-0.0308098644,
0.009997935,
0.027050443,
-0.0156630203,
-0.054557547,
0.0304435529,
-0.0269274805,
0.0560723282,
0.0954069346,
0.016749749,
-0.0475941077,
-0.0048125866,
-0.0394510292,
0.0067136497,
0.0162757616,
0.012842196,
-0.064663969,
-0.0389283039,
-0.0198500231,
0.0068914331,
-0.0277705472,
-0.0057450361,
0.0052496828,
-0.0830863416,
0.0252440069,
-0.0022958044,
0.0006130208,
0.0127580026,
-0.0222093668,
-0.0031109215,
-0.0390711576,
0.0013131291,
-0.0424900092,
0.0190916304,
0.006687718,
0.000357337,
-0.051596228,
-0.0290379357,
0.0521867834,
0.0028684186,
0.0315486565,
0.0334359445,
0.0550926886,
-0.0545600057,
-0.0028323364,
0.017685594,
-0.0060893958,
-0.0082645901,
-0.0349905603,
0.0248548351,
0.0091936225,
-0.0337427817,
-0.0497120917,
-0.0143588269,
-0.0316582099,
0.0423187092,
0.0152739221,
0.0413752012,
0.0067732753,
-0.0347629637,
-0.0279713105,
0.0418481156,
0.013837182,
0.0762611479,
-0.0281657223,
-0.0701167956,
-0.0072433804,
-0.0136465384,
-0.0128421942,
-0.0065673399,
-0.0162907112,
0.003867439,
-0.0228007436,
-0.0148954177,
-0.0613691844,
-0.0575831383,
0.0004246399,
-0.0247205161,
-0.0150240744,
0.0386360735,
-0.0100170355,
0.0148862517,
0.0042041359,
0.0110185985,
0.0161639974,
0.033314243,
0.0082692783,
0.0307510868,
0.0755245313,
-0.0247052461,
-0.0059192912,
0.031532079,
-0.0831743032,
-0.0533097722,
0.0258782972,
0.0423904397,
0.0173450708,
-0.007367447,
-0.0017308326,
-0.0197746139,
0.0031370139,
0.0699385181,
-0.000167083,
0.04200444,
0.0395912863,
0.1097697467,
0.0112065291,
0.0326368622,
-0.0100937486,
0.0525238998,
-0.0041316599,
0.01625962,
0.0101370262,
-0.0399378091,
-0.0107164597,
0.030246092,
0.0230378546,
-0.0260752048,
-0.0468230583,
0.0097001763,
-0.0042261053,
-0.0164053887,
-0.0233741179,
-0.0282253865,
-0.0152224703,
-0.0035762875,
-0.0060364115,
-0.0660840273,
-0.0344515033,
-0.044752676,
-0.0061819064,
0.0044540125,
0.0737825707,
0.0197688881,
0.0157149881,
0.0032399201,
0.0322305709,
-0.0357406884,
-0.0058478457,
0.0000958285,
-0.0044989623,
-0.0196100287,
0.0315099321,
0.0108061964,
-0.0148823867,
-0.0065857954,
0.0345422179,
0.0161068067,
-0.0130008925,
-0.0024426358,
0.0447695032,
0.0504978113,
0.0094993236,
0.0762448534,
-0.0505856276,
0.0061414395,
0.007065264,
0.033916086,
-0.034214098,
-0.0222714972,
0.0535462946,
-0.0271639097,
-0.0306705087,
0.0046412013,
0.047143463,
-0.090244174,
0.0296845827,
0.0434549712,
-0.0231728069,
-0.0273166113,
-0.019823987,
0.0495303534,
-0.0494717211,
0.0048505855,
0.0002499266,
-0.0688489899,
-0.0475612171,
0.0592921376,
-0.0314923339,
0.0157956127,
-0.0327266902,
-0.0092509268,
0.0695654526,
0.0238587726,
0.0673078969,
-0.0603061952,
0.0061047575,
-0.0483428575,
-0.0053646113,
-0.0219595172,
-6.921840445e-33,
-0.0001425091,
-0.0219327398,
-0.0487053394,
0.0364055596,
-0.0154594211,
0.0046043801,
-0.0212353859,
0.0320223421,
-0.0298562311,
0.0405090041,
-0.0067933863,
-0.0037377395,
0.0077800234,
-0.0105954427,
0.0390154645,
-0.0169687551,
0.0047652926,
0.0109573966,
0.0074786665,
0.0253726821,
0.055597607,
0.0093884217,
0.0680812299,
-0.1135406345,
0.0957876667,
-0.0399351455,
-0.059212517,
0.0050368365,
0.0084770489,
0.0276691113,
-0.0325042345,
0.0925574005,
0.0290179942,
0.0246354211,
0.0199056678,
-0.0133844046,
0.0140123861,
-0.0213183705,
0.0396519899,
0.0168248359,
0.0140266772,
-0.0191539172,
0.019171007,
-0.0072490475,
-0.0698896721,
0.0321694091,
0.0148023041,
-0.0259507876,
0.0191900153,
-0.0510419458,
-0.0481987447,
0.0164450891,
0.013102036,
0.0147378622,
0.0122855678,
0.0668237209,
0.0051047457,
0.0454168133,
-0.0501750447,
0.0161243845,
-0.0449541658,
-0.0428179316,
0.0147658279,
-0.0576389953,
0.0130863404,
0.0708006993,
-0.0318184607,
0.01388859,
-0.0036233598,
0.0362068117,
0.0148951914,
-0.0018506174,
0.0426464155,
0.0832187533,
0.0023557812,
-0.046560239,
-0.061280489,
0.0191602465,
0.0415031873,
0.0095349066,
0.0656374693,
-0.0031056749,
-0.0429284796,
-0.0215573702,
0.0216434915,
-0.0129291536,
-0.0236929413,
-0.065862067,
-0.0017846517,
-0.0362244435,
-0.0089747524,
-0.0340169296,
0.0200476367,
-0.0572363138,
-0.0198867228,
0.0206846986,
-0.0628334731,
0.0410582609,
0.0321083628,
0.0528391004,
-0.0157506689,
-0.0091070533,
-0.0500912964,
0.0328247324,
0.0188916512,
0.0007085249,
0.041486185,
0.0140764602,
-0.0649065673,
0.0161182862,
-0.0179323051,
0.0042744018,
-0.0172776524,
-0.005750359,
-0.0206940677,
0.0030283106,
-0.0105635002,
0.0360373035,
0.0505306497,
0.0687061846,
-0.0155194523,
-0.0135165835,
0.0047464524,
0.022144815,
-0.0560351349,
-0.0071326126,
0.0241984818,
-0.0339182429,
-0.0110745933,
-0.0595472902,
-0.0119957011,
-0.0254392736,
2.797e-7,
0.0713668391,
0.0791918337,
-0.0590025857,
0.0224266611,
0.0080865528,
0.0630926341,
-0.0398892649,
0.0233826786,
0.0241346881,
0.0416663624,
0.0099658193,
0.0020122288,
0.0055777649,
0.047584407,
-0.0691313595,
0.0158355031,
-0.0441542342,
-0.0130058229,
-0.0200838987,
0.0512589514,
0.0970786363,
0.0922611132,
0.0193240065,
0.0061685452,
-0.0603759512,
-0.0162522309,
0.0787822977,
-0.0663154572,
-0.0389563851,
0.0166710019,
0.0072948849,
0.0076512974,
-0.0132069094,
-0.0293074194,
0.0222153496,
-0.0031936152,
0.0267553907,
0.0683203116,
0.023810681,
-0.0106155211,
0.0282702632,
-0.006639929,
0.0183219276,
-0.0596864745,
0.0357051939,
-0.058552064,
-0.059280891,
-0.0720758885,
0.016346056,
0.0162865128,
0.0911303088,
-0.0162403733,
-0.0295647383,
0.0364644229,
-0.0023706029,
-0.0384180918,
-0.0596784167,
-0.0439887829,
-0.0293817725,
-0.0180819463,
-0.0371496342,
-0.0509016663,
0.0010324552,
0.0507229008,
0.0208730493,
-0.0384306274,
0.0513495915,
2.12021161e-34,
0.0723293126,
-0.0265517645,
-0.0346340276,
0.0206660423,
0.0444395319,
0.0165343303,
-0.0042155408,
0.0216937736,
-0.0057292907,
-0.0199369621,
0.0199629311
] |
http://arxiv.org/abs/2401.14559v1 | http://arxiv.org/pdf/2401.14559v1 | Language Modelling Approaches to Adaptive Machine Translation | 2024-01-25T23:02:54 | 2024-01-25T23:02:54 | Yasmin Moslem | cs.CL | cs.CL, cs.AI, cs.HC, cs.IR | Consistency is a key requirement of high-quality translation. It is
especially important to adhere to pre-approved terminology and adapt to
corrected translations in domain-specific projects. Machine translation (MT)
has achieved significant progress in the area of domain adaptation. However,
in-domain data scarcity is common in translation settings, due to the lack of
specialised datasets and terminology, or inconsistency and inaccuracy of
available in-domain translations. In such scenarios where there is insufficient
in-domain data to fine-tune MT models, producing translations that are
consistent with the relevant context is challenging. While real-time adaptation
can make use of smaller amounts of in-domain data to improve the translation on
the fly, it remains challenging due to supported context limitations and
efficiency constraints. Large language models (LLMs) have recently shown
interesting capabilities of in-context learning, where they learn to replicate
certain input-output text generation patterns, without further fine-tuning.
Such capabilities have opened new horizons for domain-specific data
augmentation and real-time adaptive MT. This work attempts to address two main
relevant questions: 1) in scenarios involving human interaction and continuous
feedback, can we employ language models to improve the quality of adaptive MT
at inference time? and 2) in the absence of sufficient in-domain data, can we
use pre-trained large-scale language models to improve the process of MT domain
adaptation? | Language Modelling Approaches to Adaptive Machine Translation: Consistency is a key requirement of high-quality translation. It is
especially important to adhere to pre-approved terminology and adapt to
corrected translations in domain-specific projects. Machine translation (MT)
has achieved significant progress in the area of domain adaptation. However,
in-domain data scarcity is common in translation settings, due to the lack of
specialised datasets and terminology, or inconsistency and inaccuracy of
available in-domain translations. In such scenarios where there is insufficient
in-domain data to fine-tune MT models, producing translations that are
consistent with the relevant context is challenging. While real-time adaptation
can make use of smaller amounts of in-domain data to improve the translation on
the fly, it remains challenging due to supported context limitations and
efficiency constraints. Large language models (LLMs) have recently shown
interesting capabilities of in-context learning, where they learn to replicate
certain input-output text generation patterns, without further fine-tuning.
Such capabilities have opened new horizons for domain-specific data
augmentation and real-time adaptive MT. This work attempts to address two main
relevant questions: 1) in scenarios involving human interaction and continuous
feedback, can we employ language models to improve the quality of adaptive MT
at inference time? and 2) in the absence of sufficient in-domain data, can we
use pre-trained large-scale language models to improve the process of MT domain
adaptation? | [
0.0399410129,
0.0300770067,
-0.0334335454,
0.0464994647,
-0.007147368,
0.0186129566,
0.0153377876,
-0.0208896138,
-0.0481662303,
-0.0354463346,
-0.0546255074,
-0.0523157269,
0.0164715704,
0.0241461843,
0.0386904292,
-0.0393936671,
0.0106400624,
-0.0242017824,
-0.0373640247,
0.0042482591,
0.0065258997,
0.0236543864,
0.0320940465,
0.0343307741,
0.0251073204,
0.0233371016,
-0.0031195437,
0.0460828654,
0.0208007619,
0.0097245257,
0.0012769306,
0.03464742,
0.0656624287,
0.056384448,
0.0000019644,
-0.0316182151,
-0.026302807,
-0.0116508938,
0.0127701033,
0.0013212641,
0.0532385744,
-0.0636538118,
-0.0095708109,
0.0234074816,
-0.0519161932,
-0.0101493839,
0.0646453649,
-0.0355509706,
0.0624130145,
0.0695076808,
-0.0132534076,
0.0062624421,
-0.0275823474,
-0.0771275461,
-0.0343411379,
-0.04535551,
0.0129970247,
0.0012113218,
0.0027281016,
0.0347410664,
0.0118300859,
0.0224371757,
0.0251357108,
-0.0215916317,
0.0051114853,
-0.0058147106,
-0.0084612798,
-0.0048470553,
-0.0223845858,
0.0829547048,
-0.0053022732,
-0.0279302895,
-0.0101694549,
-0.0321730189,
-0.0025449756,
0.0006630087,
-0.0402289666,
0.0134619987,
0.0056931139,
0.0105006918,
0.0015569691,
0.0128757162,
0.0349546969,
0.0203481186,
-0.03658491,
0.0057549314,
0.0228408985,
-0.0000036073,
-0.0533911325,
0.0231359415,
-0.0570410155,
-0.0335652269,
0.0299159363,
0.0394939631,
-0.0244847164,
0.037749242,
-0.0920802951,
-0.0457166284,
-0.0140674543,
-0.0638361424,
0.0399519838,
0.0551303737,
-0.0094474172,
0.0342502594,
-0.0030880147,
0.0549607761,
-0.0391316563,
-0.0133424206,
-0.088379778,
-0.0201797485,
-0.0051501105,
0.0196889341,
-0.0390936919,
-0.0142051233,
-0.076815851,
-0.0012398099,
-0.0245402809,
0.0742755607,
0.0214607455,
0.0036783521,
-0.0021938551,
-0.0422135964,
-0.1242116094,
-0.0532926023,
0.0033767817,
0.0479784608,
0.0207937304,
0.0117023066,
0.0132749975,
-0.0516789965,
-0.0184839107,
0.0177049525,
-0.0089997165,
-0.0214128643,
-0.0204959381,
0.0683434308,
-0.0089921663,
-0.0240573622,
-0.0067469864,
-0.0193366837,
-0.034941975,
-0.0359938964,
-0.0153541099,
-0.0535786301,
0.042071119,
0.0235766098,
-0.003149641,
-0.0838963091,
-0.0010977132,
-0.0076030656,
-0.0139111942,
0.0427155085,
-0.0156357698,
0.0398922712,
0.0403663777,
-0.0064436267,
0.0092342207,
0.0697259307,
0.0641182065,
0.0214474387,
0.0153531581,
0.0140852584,
-0.0110330554,
0.0792047381,
0.0237966403,
0.0654809102,
-0.0445425548,
0.0125547359,
-0.0092612375,
0.0366667025,
0.0178671498,
0.025981009,
-0.0018984427,
0.0536084883,
0.091058813,
-0.0087195486,
0.0046708607,
0.0281686895,
0.0590969063,
0.0002271107,
-0.0252813697,
-0.0091239866,
-0.0164387673,
-0.0003515937,
0.0162189305,
0.0104895094,
0.0345777534,
0.0372328982,
-0.0310077071,
-0.0503274314,
0.0064125448,
-0.006715219,
0.0038339426,
0.0039473358,
0.0662101284,
0.0386417359,
-0.0501001924,
0.0053170235,
0.0214990117,
-0.0141467666,
0.0468797386,
-0.022977639,
-0.021696629,
0.0585023016,
0.0582110509,
-0.0189047772,
-0.0419693924,
-0.0175501201,
0.0373047255,
0.0799309015,
0.0503454395,
-0.0022329953,
-0.0023790065,
-0.0334070213,
-0.0432951413,
0.0283588152,
0.0136845456,
0.0197491664,
0.0082779564,
0.0314214006,
0.0444493257,
-0.0208074115,
-0.071794726,
0.0290002618,
-0.0787137374,
-0.0245720036,
0.0389824323,
-0.0081157796,
0.0246749893,
0.0232788771,
-0.0348402597,
-0.0195756648,
0.0164834503,
-0.0292511508,
-0.0132821472,
0.0563543811,
-0.0590021126,
-0.0222150367,
0.0231415406,
0.0127507746,
0.004381754,
-0.0267855246,
-0.0415551402,
-0.0338261314,
-0.0165916272,
-0.0245575123,
0.0195483007,
-0.0385313332,
0.0478932001,
-0.0217162482,
0.0130947493,
-0.0562980436,
0.0065088812,
0.0104017789,
-0.0425956175,
0.0172437374,
0.037289653,
-0.0128495013,
-0.0111047952,
-0.0069736759,
0.0167272687,
0.0051289108,
0.0301992055,
0.0358736329,
0.0387409367,
-0.0032938642,
0.0096496968,
-0.0618400835,
0.0275522806,
-0.0282938052,
0.0316946022,
0.0467679054,
0.0586798303,
-0.0169830434,
-0.0161898155,
-0.0007902901,
0.0177730564,
0.0301170032,
0.0081990017,
0.0270928387,
0.0165632982,
0.0218583252,
-0.0215734523,
0.0236922633,
0.0350479893,
0.0194716752,
0.0357473791,
-0.0387030318,
-0.090407677,
0.0429726131,
-0.0313674547,
0.0270549245,
0.0029730094,
-0.032611914,
-0.0280117691,
-0.009037626,
0.0592358522,
0.0227524452,
-0.001643248,
0.0108637474,
0.0182804726,
0.0160125885,
-0.0545186475,
0.0113362223,
-0.0381105877,
-0.0586222969,
-0.0310067739,
-0.0074610887,
0.0516170114,
0.0418731011,
-0.0055253832,
-0.0203094594,
-0.0577199496,
0.0043771681,
-0.0272650383,
-0.0325976647,
-0.0147912502,
-0.0151450951,
-0.0074354918,
0.0400380008,
-0.0180218816,
-0.0778235942,
0.0702298358,
-0.0276132803,
0.000709753,
-0.0713845715,
-0.0059666713,
-0.021839112,
0.0296999104,
-0.0011367456,
0.0169215929,
-0.016692847,
0.0001316726,
-0.0515964851,
0.0088731535,
0.0146755846,
-0.0295163151,
0.0358819216,
-0.0923340693,
-0.0007200522,
-0.0150050689,
-0.0285958275,
-0.0484003909,
0.001007983,
-0.0130560407,
-0.0168852974,
-0.0126740793,
-0.0262941997,
-0.0333865322,
0.0229325313,
0.0181680135,
0.0869439244,
-0.0063224812,
-0.0114176264,
-0.0341897421,
0.0203478448,
0.0309094191,
0.0292270537,
-0.0147123048,
-0.0427747332,
0.0071692602,
0.0069531091,
0.0031597791,
-0.0369783938,
-0.0221336186,
-0.0350667313,
0.0451888442,
0.0102688652,
0.0009183227,
0.022302324,
0.0540337637,
-0.0485892855,
-0.0427284129,
0.1017530262,
0.0138158454,
-0.060037151,
-0.0049725394,
0.0015007954,
0.0086003402,
0.0048566181,
0.0323021263,
-0.0732157603,
-0.0308611076,
0.0141407009,
0.0269144457,
0.0344413482,
0.0089896452,
-0.0294641405,
-0.0920306742,
0.0272129308,
0.0537263118,
-0.0743866712,
-0.0548410453,
-0.0073861587,
0.0188722685,
-0.0179797933,
0.035430558,
-0.0189350508,
0.0291553196,
0.0014222222,
0.0194040444,
0.0360117778,
-0.0847555548,
-0.0294537544,
-0.0526172221,
0.0461614802,
0.0915980861,
0.0100452816,
0.0203941204,
0.0048508034,
0.0679085404,
-0.0130193746,
0.0033309085,
-0.017423762,
-0.0139853442,
-0.0144472383,
-0.009879956,
0.0237273797,
0.0374885574,
-0.0517204106,
0.0192912463,
0.0432417318,
0.0545277111,
-0.0128028831,
0.0278142318,
-0.0046448121,
-0.0400024578,
0.0348345451,
0.0514608212,
0.0076268543,
-0.0827394873,
0.0162517335,
-0.0093150847,
-0.0237908997,
-0.0378219448,
-0.0230198223,
-0.0151471244,
-0.0592372045,
-0.0316562802,
-0.0967502967,
-0.0291449744,
-0.0021182674,
0.0458803549,
0.04347004,
0.0236666799,
0.029245723,
-0.0042356839,
-0.0160199925,
0.0300633851,
-0.0501491129,
0.011383309,
-0.0472140536,
0.0133336717,
0.021108849,
-0.0817040205,
0.0112261334,
0.0414825603,
-0.0750396773,
-0.0767271817,
-0.0120521365,
0.0254618749,
0.0207724962,
0.0666107237,
-0.025488833,
-0.0570603311,
0.0450057946,
0.0149246845,
-0.024314411,
0.0142574143,
-0.0023919891,
0.0120565239,
-0.0228070654,
0.0431530215,
-0.039587751,
0.0323541164,
0.0268104449,
0.0390060842,
0.0086845625,
-0.0888797939,
0.0182850305,
-0.007138615,
0.0304049142,
-0.0154942768,
-0.0559850968,
-0.0160773229,
0.0113602513,
0.0041920417,
-0.0330272503,
-0.0615634024,
-0.0094265016,
0.0144669358,
-0.0058343005,
0.004337443,
-0.0048891627,
-0.0211866628,
0.0239933711,
-0.0253153872,
-0.0261906795,
-0.0390920043,
0.0454810448,
0.0405993834,
0.0372525007,
0.0003365793,
-0.0019824575,
-0.0202779099,
0.0162102226,
-0.0320523977,
0.0397119373,
0.0137156015,
-0.0070445333,
0.0047960007,
0.0424534343,
-0.0382725038,
-0.0058954228,
-0.0596060008,
0.0246654674,
0.0661475584,
0.0163860396,
0.1051785499,
-0.04431675,
-0.0063978764,
-0.0433997363,
0.0488552824,
0.010922662,
-0.0112936683,
0.0078520253,
-0.0193093978,
-0.0164022222,
0.0428522117,
0.0593017414,
0.0291858297,
0.0210746937,
-0.0101744691,
-0.0039818697,
-0.0006383798,
0.0034151734,
0.0128238024,
-0.0219715666,
-0.0390256718,
0.0109290192,
-0.0167321917,
-0.0542598106,
-0.0167789143,
0.0050521423,
-0.012956731,
-0.0338020734,
0.0045742677,
0.0687920377,
-0.0417731889,
0.02426227,
-0.0403624475,
0.0528952107,
-0.0038024462,
-0.0112448577,
-0.020000333,
-5.932830559e-33,
-0.045644287,
-0.0315604694,
0.0325823165,
-0.0439714864,
-0.040938016,
-0.0212222654,
-0.0359672643,
-0.0197098181,
0.0134416018,
-0.0099254679,
-0.0096894419,
0.0221037399,
0.0158749856,
0.0174597576,
-0.0101258634,
0.0269070212,
0.0623054579,
-0.004903052,
-0.0291596893,
0.0543533415,
0.0126481652,
-0.0008466448,
0.0365691185,
-0.0699001327,
0.0486988574,
-0.0395659134,
-0.0232274029,
0.0323673449,
0.0280219335,
0.0300772339,
0.0002853274,
0.0285209827,
0.0165510327,
-0.0373389609,
-0.0451397635,
0.0024425485,
-0.0690449551,
0.0415512435,
0.0043298816,
0.0403142609,
0.0004746041,
-0.0340537801,
0.0689292252,
-0.0224277806,
-0.0247583725,
0.0191649329,
0.0103586791,
-0.0708152726,
-0.0289547257,
-0.0146934865,
-0.0662243888,
0.0369075835,
0.0118125873,
-0.0261043794,
-0.0019299178,
0.0742629766,
-0.0092394426,
0.02318535,
-0.0604416952,
0.0338005014,
0.0009594625,
0.0271569341,
0.0992768556,
0.0107339527,
-0.0097541222,
-0.0192346647,
-0.0271255877,
0.0230625346,
0.0223554783,
-0.0242215116,
0.0245374069,
0.0055167186,
0.0189149529,
0.0892070308,
0.0361154713,
-0.033877492,
-0.0436860546,
0.0532284342,
-0.0240793563,
-0.0048561841,
0.0430864394,
0.0189902447,
-0.0344071835,
0.0072475611,
-0.0345320962,
0.0166742522,
-0.0199898053,
-0.033378344,
0.0181001052,
0.0284212474,
-0.0338893309,
-0.0009824589,
0.078181453,
-0.0273415949,
0.0080257049,
-0.026611235,
-0.0447214544,
-0.0269580707,
-0.015540625,
0.0506418608,
-0.0031489928,
-0.0600652099,
0.0116002467,
0.0304273497,
-0.0355871916,
-0.0284255147,
0.0316835344,
0.0299047157,
-0.0386241935,
0.0322699957,
-0.0251389649,
0.0096514132,
0.0046144244,
0.0274658352,
0.0318323188,
-0.0107879378,
-0.0158674251,
0.0693223923,
0.0427428,
-0.077114962,
-0.0120848557,
0.0022646305,
0.0021482285,
0.0128698414,
-0.0561919995,
0.0015183635,
-0.0038439408,
0.0188919939,
0.0294678994,
-0.0547529273,
-0.0344775058,
-0.036233481,
2.719e-7,
0.0287031848,
0.048978731,
-0.0115621015,
-0.0186523125,
0.038726069,
0.0307102893,
-0.0235479511,
0.0449299887,
0.0172894523,
0.0443580523,
0.0491201207,
0.0237201527,
0.0734106079,
0.0049378737,
-0.0249668192,
-0.0350033939,
-0.0257457048,
0.0707587823,
-0.0516204946,
-0.007031865,
0.0053743739,
0.078271389,
0.0053987419,
0.0043318281,
-0.003972806,
-0.0749836564,
0.0876027271,
-0.0110257389,
-0.0116240252,
0.0243207496,
-0.0051856809,
0.0310548898,
-0.0581637844,
-0.0316383354,
-0.0213054195,
0.0380908102,
0.0277233925,
0.0836307406,
0.0222155601,
0.0334375501,
0.0046261987,
-0.0070591392,
0.0143869901,
-0.0407839194,
0.0438162312,
-0.093839705,
0.0044115437,
-0.0004574451,
0.0106095178,
-0.0067940559,
0.0032727083,
-0.0310115889,
-0.031376753,
0.018632751,
0.064439714,
0.0299187452,
-0.0412122607,
-0.0282082334,
-0.0185073726,
-0.0277940836,
-0.0566544607,
-0.0380090177,
-0.0565939695,
0.017618129,
-0.0000628951,
-0.0282947235,
0.043300122,
2.513805001e-34,
0.0202512685,
-0.0216142274,
-0.0088233501,
0.0029727304,
0.008283806,
-0.0205983762,
0.018353207,
0.0148100005,
0.0354579017,
-0.0545822866,
-0.0165392645
] |
http://arxiv.org/abs/2401.14556v3 | http://arxiv.org/pdf/2401.14556v3 | Looking Right is Sometimes Right: Investigating the Capabilities of Decoder-only LLMs for Sequence Labeling | 2024-01-25T22:50:48 | 2024-06-06T12:33:19 | David Dukić, Jan Šnajder | cs.CL | cs.CL | Pre-trained language models based on masked language modeling (MLM) excel in
natural language understanding (NLU) tasks. While fine-tuned MLM-based encoders
consistently outperform causal language modeling decoders of comparable size,
recent decoder-only large language models (LLMs) perform on par with smaller
MLM-based encoders. Although their performance improves with scale, LLMs fall
short of achieving state-of-the-art results in information extraction (IE)
tasks, many of which are formulated as sequence labeling (SL). We hypothesize
that LLMs' poor SL performance stems from causal masking, which prevents the
model from attending to tokens on the right of the current token. Yet, how
exactly and to what extent LLMs' performance on SL can be improved remains
unclear. We explore techniques for improving the SL performance of open LLMs on
IE tasks by applying layer-wise removal of the causal mask (CM) during LLM
fine-tuning. This approach yields performance gains competitive with
state-of-the-art SL models, matching or outperforming the results of CM removal
from all blocks. Our findings hold for diverse SL tasks, demonstrating that
open LLMs with layer-dependent CM removal outperform strong MLM-based encoders
and even instruction-tuned LLMs. | Looking Right is Sometimes Right: Investigating the Capabilities of Decoder-only LLMs for Sequence Labeling: Pre-trained language models based on masked language modeling (MLM) excel in
natural language understanding (NLU) tasks. While fine-tuned MLM-based encoders
consistently outperform causal language modeling decoders of comparable size,
recent decoder-only large language models (LLMs) perform on par with smaller
MLM-based encoders. Although their performance improves with scale, LLMs fall
short of achieving state-of-the-art results in information extraction (IE)
tasks, many of which are formulated as sequence labeling (SL). We hypothesize
that LLMs' poor SL performance stems from causal masking, which prevents the
model from attending to tokens on the right of the current token. Yet, how
exactly and to what extent LLMs' performance on SL can be improved remains
unclear. We explore techniques for improving the SL performance of open LLMs on
IE tasks by applying layer-wise removal of the causal mask (CM) during LLM
fine-tuning. This approach yields performance gains competitive with
state-of-the-art SL models, matching or outperforming the results of CM removal
from all blocks. Our findings hold for diverse SL tasks, demonstrating that
open LLMs with layer-dependent CM removal outperform strong MLM-based encoders
and even instruction-tuned LLMs. | [
0.0420325212,
0.0617443807,
0.0084750103,
0.0407129973,
-0.0032920982,
0.0537941381,
0.0402514786,
-0.0183696244,
-0.0004388343,
-0.0490765236,
-0.0185882635,
-0.0247926731,
-0.0182759073,
0.0425016508,
0.0015713449,
-0.0077307126,
0.031649936,
-0.0533488803,
0.0203265622,
0.0225543305,
-0.0317067094,
0.0053599086,
0.0130013982,
-0.0023578622,
0.032888636,
0.0023522552,
0.009203881,
-0.0097375447,
0.013663006,
-0.0575613603,
0.0032612167,
0.0168811381,
0.0449631438,
0.0507344864,
0.000002235,
-0.0295997504,
-0.009979344,
-0.0118928468,
-0.0643912181,
0.0466609113,
-0.0210445523,
-0.0439462177,
-0.0074985721,
-0.0060903905,
-0.0244546439,
-0.062173225,
0.0344616771,
0.0459395871,
0.0386385918,
0.0925927162,
0.0010342625,
-0.0719316006,
0.0299289022,
-0.0345146284,
-0.0084694568,
-0.0371617153,
0.0397344083,
-0.0130170146,
-0.030894937,
-0.0342221297,
-0.0147107644,
0.0531687662,
0.0201607682,
0.021072356,
0.0093341442,
0.050142765,
0.0440959297,
-0.046841871,
-0.0122343153,
0.0817294344,
-0.0405185781,
0.0056796926,
0.0114284679,
-0.0060998918,
-0.0336787514,
-0.0357722118,
-0.0170549024,
0.0690211728,
-0.0023040986,
0.0268039498,
0.0309638977,
0.0151461186,
0.0611861534,
0.0126408599,
-0.0811855346,
0.0778557286,
0.0001490385,
-0.0384200513,
-0.0107513517,
-0.0126392012,
0.0264190733,
-0.0539779849,
0.0345211066,
0.0253231544,
0.013468558,
0.0269747395,
-0.0339079089,
-0.0422934927,
-0.0401705429,
-0.0192315839,
0.0082240924,
0.0450344868,
-0.0240486749,
0.0480183177,
-0.0443683229,
0.0442199372,
-0.0159828775,
0.0755454451,
-0.0335302055,
0.0045327111,
0.0074843983,
-0.0279469788,
-0.0145183001,
0.0703622624,
-0.0852964967,
-0.0176449753,
0.0124494545,
0.0231013298,
0.0278584715,
0.0148575371,
-0.0183288381,
-0.0273463465,
-0.012019827,
-0.0166622736,
0.0201078467,
-0.0101234475,
0.0062484429,
0.0077178576,
0.0082469536,
-0.059555579,
-0.0500265807,
0.0308841988,
0.0259459317,
-0.0543645546,
0.023164751,
0.0788684338,
-0.0137218283,
-0.0598109029,
-0.0270468872,
0.0024098782,
0.0204803385,
-0.0386046618,
0.0168377552,
-0.0091902902,
0.0318190046,
0.0115294708,
0.014636836,
-0.0665488988,
0.0351159871,
-0.0056429356,
-0.0341721289,
0.0758205652,
-0.0167863872,
-0.0243921187,
0.0336768962,
0.0018050936,
0.0372438692,
0.0201974772,
0.0323056504,
0.0380114205,
0.0267624948,
-0.0413139537,
0.0135906702,
0.0460325405,
0.01137398,
-0.022931762,
-0.0170870498,
-0.004721697,
0.0577873513,
-0.0166917183,
0.0738878846,
0.0440046266,
-0.0309811272,
0.0009905976,
0.0624899641,
0.0027380781,
-0.0264342465,
-0.0088823056,
0.0340959765,
0.016954381,
0.0011033966,
0.0117667913,
-0.013029729,
0.0006559847,
-0.0333051085,
0.0195172895,
0.031786032,
0.036272645,
-0.0212209709,
-0.0330531895,
-0.0255959723,
-0.0175439715,
0.0097774696,
0.0248744078,
0.0479614437,
0.0136625562,
-0.0714056492,
0.0484041721,
0.0074651311,
-0.021072438,
0.0192033015,
-0.0457195267,
0.0097762244,
0.0153543474,
0.0555641577,
-0.0201712791,
-0.0232334994,
-0.018443469,
-0.0026099472,
0.0434922874,
-0.0034795757,
0.0047837021,
0.0513895229,
-0.0548861697,
-0.0183012001,
-0.0192207396,
-0.0022053737,
0.0373852216,
0.0049402998,
-0.0372795127,
0.0272704605,
-0.019825168,
-0.0318477526,
0.0617705137,
-0.0228976868,
-0.0745356753,
0.0252381153,
-0.0021545254,
0.0422104783,
0.0199899618,
-0.0434660316,
0.0053012394,
0.0424169227,
-0.0696263537,
-0.0002711449,
-0.0182787143,
-0.0782344043,
-0.0879148692,
0.0086845374,
0.0093017835,
-0.0166008174,
0.0667253807,
-0.0564888157,
-0.0383029729,
-0.0172144342,
0.0006100821,
0.0176055487,
-0.030536484,
0.0248348787,
0.0382885672,
0.0362450182,
-0.0386663713,
0.0461382791,
-0.0250571612,
-0.0068319985,
0.0246965103,
0.0071971822,
-0.0160625558,
-0.0244813245,
-0.019267628,
0.0427159779,
0.0033435882,
-0.0376159102,
-0.0170563776,
-0.0084604267,
-0.0217016581,
-0.016762929,
-0.0333955698,
-0.0067238454,
-0.0424377434,
0.0341795161,
0.022135634,
0.0273141079,
-0.0118251219,
-0.0212134346,
-0.0096111121,
-0.0541910194,
-0.0153463511,
-0.024138188,
0.0076697767,
0.0386169627,
-0.0312831327,
-0.0276978761,
-0.0060220538,
0.0250779185,
0.0325529613,
-0.0030578838,
-0.0983623862,
-0.0522374287,
0.0040559331,
0.0051155142,
0.0162475873,
-0.0360290445,
-0.0173309203,
-0.0460900851,
-0.0107777501,
-0.0046645333,
0.0447287671,
0.0029351828,
0.0495157689,
0.0304588173,
-0.0201516934,
-0.0284194257,
0.0219981689,
-0.0213009082,
0.0073964773,
-0.0571842939,
-0.0266324487,
0.0346479379,
0.1105610728,
-0.0046891309,
0.0536583215,
-0.0181717407,
-0.0155946165,
0.0097647868,
-0.0397192314,
-0.0300291441,
-0.0191120151,
-0.0380488262,
0.0291218758,
-0.0497075357,
-0.0350491516,
0.0515352525,
-0.0647090003,
-0.02705664,
-0.0156521518,
-0.0155383265,
-0.0621955208,
0.0452768765,
0.0253871251,
0.0090219742,
-0.0097332168,
-0.0707449615,
0.0244118329,
0.0362800211,
-0.0169091932,
-0.0380393974,
-0.0202404149,
-0.0775440708,
0.0225023963,
-0.0096435091,
-0.0084288679,
-0.0360403135,
-0.0295857415,
-0.0193291046,
0.003106782,
0.0544081181,
0.0207860973,
-0.0373336375,
0.0378723331,
0.0272413958,
0.0647393912,
0.020686524,
0.0600241646,
-0.0468405373,
0.0209603477,
0.034632355,
-0.0143183498,
0.0139089515,
-0.0417852513,
0.0687824264,
0.0226204097,
0.0186276138,
-0.000560979,
0.0010350893,
0.024193529,
0.0597384423,
-0.0154329687,
-0.0008309027,
0.0241229162,
0.0491182357,
-0.0711908117,
-0.0195480604,
0.1442748457,
0.0278184302,
-0.0376391932,
-0.0274924599,
0.022859579,
0.005514835,
0.0149545111,
-0.0049689766,
-0.0815755203,
-0.0448959433,
0.0217403509,
0.0135963503,
-0.0177764762,
-0.0083789332,
-0.0120667983,
-0.1081684604,
0.0291278157,
0.0246832874,
-0.0572218709,
-0.0303156022,
-0.0533475317,
-0.0318429619,
0.0097043179,
0.0320635065,
0.0150519973,
-0.0007656245,
-0.0207091589,
0.0447942242,
-0.0397053696,
-0.0747848004,
-0.0152591867,
-0.0130185774,
0.040907573,
0.0635174438,
-0.0207520537,
-0.0086209662,
-0.0268053748,
0.0586032309,
-0.0140304314,
0.0065696565,
-0.0421613269,
0.0416133218,
-0.0697328076,
-0.0185080674,
0.0112069035,
0.0244777706,
-0.0328348875,
-0.0144249946,
0.000488285,
0.0218467284,
-0.0230440162,
0.0076205987,
0.0351224318,
-0.0275425073,
-0.0237616319,
0.0771359578,
0.0427981056,
-0.021369623,
-0.0220564455,
-0.0265900698,
-0.0014408655,
-0.0518426374,
0.0041951556,
0.0376688093,
-0.018693177,
-0.0350187048,
-0.0271561146,
-0.0330138169,
0.0497960933,
0.0142739201,
0.0106205847,
0.0201982781,
0.0157399345,
-0.0231021959,
-0.056206312,
0.0139540117,
-0.0370025486,
-0.005729185,
-0.0796468034,
-0.0235649478,
0.0047012507,
-0.0346893668,
-0.0129171293,
0.0217524283,
-0.0625326112,
-0.0573564544,
0.0120839011,
0.0310489088,
0.0094079515,
0.0674969107,
0.0072992076,
-0.052523572,
0.0242818333,
0.0290241875,
-0.0159213468,
0.0204199851,
0.0292786006,
-0.0505542271,
-0.0189680886,
0.0073025129,
-0.0344001614,
0.0328725949,
0.0620843507,
0.0400925763,
-0.032536719,
0.0116502717,
0.0238583442,
0.0344749391,
0.0956680104,
-0.0190094858,
0.0036047888,
-0.0335854068,
0.0509782583,
-0.0134832803,
-0.0242459159,
-0.0181790069,
-0.0224717446,
0.0413234942,
0.0050682919,
-0.0227185749,
0.0642829835,
0.0025912586,
0.0285524186,
-0.0348252356,
0.0335074849,
-0.0828449652,
0.0167756248,
0.007544457,
-0.0195945706,
-0.0372942612,
0.0454063788,
0.0455212183,
0.0315476209,
-0.0019844703,
0.0451879017,
0.0131087555,
0.029619569,
-0.0186056998,
0.0335002653,
-0.0427989811,
-0.0324217528,
-0.048587244,
-0.0175815914,
0.0515260808,
-0.0256795883,
0.0400907323,
0.0123646334,
-0.0116731683,
-0.0047398349,
0.0039052765,
0.0163393188,
-0.0409722924,
0.0444636419,
0.010380378,
-0.0466920063,
0.0028013082,
0.0619527884,
0.0468625389,
0.0570460446,
-0.0181489326,
0.0003703061,
-0.0326837003,
-0.0158844478,
0.0324959643,
0.0382528529,
-0.0343338028,
0.0052351095,
-0.0341262333,
-0.0342187546,
-0.0029101796,
-0.0288584065,
0.022616189,
-0.0337150395,
0.0200578757,
0.0216466039,
-0.0002452407,
-0.0520389266,
-0.038251821,
0.0295325816,
0.035711851,
-0.0217321943,
0.0107845012,
-7.157222168e-33,
-0.0590527728,
0.013991476,
0.0240194891,
0.0114053488,
-0.0130284717,
0.0120115429,
-0.0523335412,
-0.0190058872,
0.003302492,
-0.0141615188,
-0.0006495566,
-0.0067421864,
0.0264371075,
0.0574345998,
-0.0395486094,
-0.0231018737,
0.0655109286,
-0.0048612314,
-0.0215437803,
-0.0065421402,
-0.0058773914,
0.0620821752,
0.0556381978,
-0.0500451997,
0.0106963422,
-0.0242504235,
0.0198387261,
-0.0112433387,
0.0147062689,
0.0145059228,
-0.0419430658,
-0.0060536372,
0.0249335524,
-0.065245688,
-0.0209645052,
0.0120599885,
-0.1173764095,
0.0460489243,
0.0157972891,
0.0333221518,
-0.0217638928,
-0.0768244565,
0.0322024785,
0.0236499719,
-0.0171382148,
0.021009095,
0.0103746196,
-0.0201675203,
-0.0247976463,
-0.0449740179,
-0.0424009301,
0.0549986362,
-0.0117979338,
0.0087753329,
0.0732134059,
-0.008134257,
-0.0152478069,
0.0697466955,
-0.0621767081,
-0.0221039709,
0.0174651612,
0.0407428108,
0.0722983778,
0.0809214786,
0.0159965977,
0.0474101119,
0.0205728114,
-0.0115725761,
-0.02012592,
0.0443066992,
0.0012678938,
0.0207626149,
0.0102765374,
-0.0253586508,
0.0412577912,
-0.0122132264,
-0.0750790089,
-0.0030251949,
-0.0674282014,
0.0022631374,
0.0282941051,
0.0382505395,
-0.0067007351,
-0.0013444696,
-0.0452141017,
0.0090137962,
-0.0216829441,
0.0342301279,
0.0439826883,
0.0268745907,
-0.0491374396,
-0.0033724101,
0.0448972657,
-0.0181959663,
0.0231345948,
-0.0138542876,
0.0267384201,
0.0140430434,
-0.021247115,
0.0056293728,
-0.097960256,
-0.0522448979,
0.0207962114,
-0.0188373029,
0.0213964153,
-0.0100252237,
0.0335073881,
-0.0034460507,
-0.0588152334,
-0.0146774286,
-0.0570995025,
0.03088535,
0.0216618814,
0.0362501033,
0.0389439203,
0.0116786966,
-0.0147097437,
0.0531174131,
-0.0074567529,
-0.0056365025,
-0.0225581955,
-0.0249833893,
-0.0010839608,
0.0277222171,
-0.0148786521,
0.0017093362,
0.0017625353,
0.0224126652,
0.0242779329,
-0.0437402129,
-0.0030496593,
0.0441127308,
2.905e-7,
0.0388555713,
0.0401952043,
0.0588394552,
0.0177585781,
0.0132846832,
0.0281549525,
-0.0021780706,
0.0016276974,
0.014239558,
-0.0160666984,
-0.0030992175,
0.0070176697,
0.0641203001,
-0.0164520517,
-0.0602764525,
-0.0007472496,
-0.0220721569,
-0.007116315,
-0.0532275736,
0.0022982308,
0.0417997278,
0.1033003405,
-0.0233884081,
0.0033945306,
-0.0039492543,
-0.0510120355,
0.0383864641,
-0.0248003155,
0.001762421,
0.0144668613,
-0.0022452914,
0.0179764424,
-0.0281346533,
0.0218273234,
-0.0094951037,
0.0252176076,
0.0186010953,
0.0795467868,
-0.00268701,
-0.0104021402,
-0.0085094487,
0.0285317004,
-0.0039676074,
-0.0256045312,
0.0979018658,
-0.0254281648,
-0.0338793844,
0.0076763649,
-0.06273406,
-0.0111406092,
-0.039457541,
-0.0151178502,
0.0121121947,
0.0496901497,
0.0240069702,
-0.0220194496,
-0.0248377193,
-0.0865244716,
-0.0208507534,
-0.0346484259,
-0.0321538895,
-0.0102843866,
-0.036985185,
0.0055724406,
0.0174448192,
-0.0417038687,
-0.0092970626,
2.898033467e-34,
0.0147435814,
-0.0238619763,
-0.0311798602,
0.0514196381,
0.0463612974,
-0.012527124,
-0.0492774583,
0.0372374989,
-0.0301929601,
-0.1166357696,
0.0205762628
] |
http://arxiv.org/abs/2401.14530v1 | http://arxiv.org/pdf/2401.14530v1 | Relative Value Biases in Large Language Models | 2024-01-25T21:49:32 | 2024-01-25T21:49:32 | William M. Hayes, Nicolas Yax, Stefano Palminteri | cs.CL | cs.CL, cs.AI, cs.LG | Studies of reinforcement learning in humans and animals have demonstrated a
preference for options that yielded relatively better outcomes in the past,
even when those options are associated with lower absolute reward. The present
study tested whether large language models would exhibit a similar bias. We had
gpt-4-1106-preview (GPT-4 Turbo) and Llama-2-70B make repeated choices between
pairs of options with the goal of maximizing payoffs. A complete record of
previous outcomes was included in each prompt. Both models exhibited relative
value decision biases similar to those observed in humans and animals. Making
relative comparisons among outcomes more explicit magnified the bias, whereas
prompting the models to estimate expected outcomes caused the bias to
disappear. These results have implications for the potential mechanisms that
contribute to context-dependent choice in human agents. | Relative Value Biases in Large Language Models: Studies of reinforcement learning in humans and animals have demonstrated a
preference for options that yielded relatively better outcomes in the past,
even when those options are associated with lower absolute reward. The present
study tested whether large language models would exhibit a similar bias. We had
gpt-4-1106-preview (GPT-4 Turbo) and Llama-2-70B make repeated choices between
pairs of options with the goal of maximizing payoffs. A complete record of
previous outcomes was included in each prompt. Both models exhibited relative
value decision biases similar to those observed in humans and animals. Making
relative comparisons among outcomes more explicit magnified the bias, whereas
prompting the models to estimate expected outcomes caused the bias to
disappear. These results have implications for the potential mechanisms that
contribute to context-dependent choice in human agents. | [
0.0368543305,
0.0619370155,
-0.0260417145,
-0.0102201225,
-0.0284929536,
0.0351854861,
-0.0496216901,
0.0194069296,
-0.0132847438,
-0.0018916994,
-0.0526112914,
-0.0097668851,
-0.0098790983,
-0.0158281606,
0.0274224188,
-0.0712410137,
0.0355357677,
-0.04467161,
0.0727512315,
-0.0414197333,
-0.0166692566,
-0.0436272547,
-0.0085716704,
0.0237779953,
0.0080189509,
0.0079399208,
0.0120954895,
-0.042429395,
0.0273323189,
-0.0852704048,
-0.0133722434,
0.0407886952,
0.037595205,
0.013233046,
0.0000020502,
-0.0599259287,
0.0089405756,
-0.0455329381,
-0.0482793376,
0.0193894021,
0.0542700402,
0.0497346297,
-0.0128308693,
-0.0139336968,
-0.0349086225,
-0.0286299866,
0.0605025105,
0.0175384898,
-0.0199355911,
0.0223705713,
0.0099231647,
0.0082045728,
-0.0054254448,
-0.0422445349,
0.0466594882,
-0.0594113395,
0.0209108889,
0.0058618099,
0.0077137202,
-0.0132080754,
-0.0052788542,
-0.0386904702,
0.0288183466,
0.0293807369,
-0.0466918163,
0.0883952752,
0.0253899992,
-0.015400812,
-0.019900592,
0.0204655211,
0.0025980852,
-0.0226300023,
-0.0202307347,
-0.0451176949,
0.0466530584,
-0.0214466434,
-0.0281287059,
0.0543599576,
0.0073569627,
0.0393609442,
0.0445159115,
0.0459988452,
0.0251564533,
0.0195161738,
-0.0166927371,
0.0685035661,
-0.0083177891,
-0.0404762812,
0.0413126498,
0.0390275493,
-0.0188269243,
-0.0526709817,
-0.0197254736,
0.0107804239,
-0.0486741215,
0.0307828207,
0.0075834263,
0.0441331267,
-0.0009386127,
-0.0585247502,
0.0378688909,
0.0453589335,
-0.0348531865,
-0.0068547186,
-0.0139184939,
0.0426779278,
-0.0238318238,
-0.0074867727,
-0.0397351421,
0.0604135022,
0.0073286225,
-0.0147593785,
0.0236136615,
0.0313961431,
-0.0160482023,
0.0030475282,
-0.0723587573,
0.0359785557,
-0.0232714564,
0.0292470045,
-0.0378735401,
0.0248546693,
-0.0265228208,
0.0231189057,
-0.108046487,
-0.0100266114,
-0.0611476637,
0.0550748035,
0.0160933938,
-0.0208518598,
-0.0060239076,
-0.038807299,
-0.0088345883,
-0.0528361686,
0.0052798269,
0.090962708,
-0.0231608395,
-0.0257008541,
0.0031950835,
0.0003607082,
0.0033422357,
-0.036983341,
-0.0209222808,
-0.0239444561,
-0.0099279732,
0.0304308776,
-0.0030818968,
-0.004384737,
0.021199014,
-0.0106433518,
-0.0224172641,
0.0622414947,
0.0764691457,
-0.0030809806,
0.0185164902,
-0.006224555,
-0.012366306,
0.0655165985,
0.0094625251,
0.0518991053,
-0.0119207185,
0.0052745105,
-0.0158411022,
0.0039188908,
-0.0181772895,
0.0066231843,
-0.0537329949,
0.0086729089,
0.0401857309,
0.0361742638,
0.0747883469,
0.0326213464,
-0.064005971,
0.0043042805,
0.0991078168,
0.0297854692,
0.0006406159,
-0.0198081955,
-0.0138012404,
-0.0469345674,
0.012363351,
0.0371629745,
-0.0843804777,
0.0494607203,
0.0227391701,
-0.0033219836,
-0.0079949899,
-0.0157773141,
0.0033383577,
-0.0085689854,
0.0083693331,
-0.0121912016,
0.0004741481,
0.0396527797,
0.0129189212,
-0.0395095684,
-0.0429610722,
-0.0236743987,
0.046747718,
0.0084791323,
-0.0386972949,
-0.0453385711,
0.0640583932,
0.0370593965,
-0.0063697919,
-0.0084951278,
0.0045005316,
-0.000176708,
-0.0720777512,
0.0302731972,
0.0884277746,
-0.0198092181,
-0.0024607019,
-0.0126729356,
-0.0074420627,
0.0057714991,
0.0223580524,
0.0534528606,
0.0017117879,
0.0241342243,
-0.0249308441,
0.0293163843,
-0.0126155848,
0.0307004768,
0.0002097541,
-0.0483603179,
-0.0274958443,
-0.0201454926,
0.018025443,
-0.0555443503,
-0.0214494541,
0.0367491953,
0.0174366031,
0.0126745692,
0.0053814603,
0.0423707068,
-0.0791798234,
-0.0114039984,
-0.0001750757,
0.007772868,
0.0014072479,
-0.0104804011,
0.0052351165,
0.0068579228,
-0.0067869416,
-0.0306488033,
0.076687403,
0.0057649235,
0.0397852995,
-0.0406202972,
0.0043679569,
-0.0725795925,
0.0771910697,
-0.0068142056,
-0.0079949601,
0.0050036265,
-0.0105073797,
-0.0230542719,
-0.0855643451,
0.0200901423,
0.0408768579,
0.0317181088,
0.0138259511,
0.0347767211,
0.0034848626,
-0.0632068887,
-0.0443413965,
-0.0467632599,
-0.0277646463,
0.0385365672,
0.0175141376,
-0.0064087473,
-0.0366621092,
-0.0175997391,
0.0030212838,
-0.0042631924,
-0.0082970764,
0.019532213,
0.0948386639,
0.0127289519,
0.0041413712,
0.0051197023,
-0.0846616998,
-0.0109050777,
-0.0104548354,
0.031313505,
0.0058752317,
0.0539338514,
-0.0429368354,
0.0347354077,
-0.0209006183,
0.00931089,
-0.0208451133,
0.003748222,
0.0450638309,
0.0020943955,
0.0611552969,
0.083455056,
0.006572654,
0.0659226775,
0.0836241618,
0.0041764253,
-0.0801289305,
-0.0349127278,
-0.0463569127,
-0.0479512811,
-0.0298258197,
-0.0496313907,
-0.0237327814,
0.0791454986,
0.0134757468,
0.0226871502,
-0.0140355853,
0.073251918,
-0.0285091698,
-0.0557398275,
-0.0183438249,
-0.0360609703,
-0.0356179476,
0.0933043435,
-0.0137365153,
-0.0663512349,
0.0862655491,
0.0502650477,
0.0197995547,
-0.020720832,
0.0442052037,
-0.0089645963,
0.0296528395,
0.0200348161,
0.0035724635,
-0.0114449542,
-0.0250665434,
-0.0704163834,
-0.0575211346,
0.0157503597,
-0.0262800623,
0.0499019362,
-0.0196858831,
-0.0208460055,
-0.010826773,
-0.0706406981,
-0.0566314496,
-0.0286994912,
-0.0370820239,
-0.0220399722,
0.0043767998,
-0.001541807,
0.0143585913,
-0.0215126313,
0.0569821261,
-0.0222196374,
-0.0140988538,
-0.0303999763,
-0.0019046249,
-0.0099904891,
0.0473761372,
-0.0214073006,
0.0762621462,
-0.0535551757,
0.0206300113,
0.0358814187,
0.0499352068,
0.0257971194,
-0.0288599581,
0.0533165,
0.013208732,
-0.0442368276,
-0.0117293531,
0.057830289,
0.0091957152,
-0.0302890427,
-0.0036760843,
0.0650702715,
0.0183268972,
-0.0651169643,
0.0423589312,
-0.0633917153,
0.0275166165,
0.0214410163,
0.0177850537,
-0.0467858836,
-0.0096078804,
0.0096100941,
0.0052144607,
-0.0408594944,
-0.0208742842,
0.0469347164,
-0.1114885658,
-0.0073855063,
-0.0169845857,
-0.0265674815,
0.0352832563,
-0.0114159938,
-0.0056966403,
-0.0130996024,
0.0168746077,
0.0392093845,
-0.0250831656,
0.0007685568,
0.0074285916,
0.0721907616,
-0.0004970619,
0.0309044011,
-0.0257459991,
0.0519560613,
0.0709326342,
0.0073475256,
-0.0440562293,
-0.0020755234,
0.009857255,
-0.0216705352,
-0.0050731795,
0.0090454035,
-0.0026193191,
0.036391139,
0.0067467294,
-0.0386590697,
-0.0053150812,
0.0562327616,
-0.0465231165,
-0.0054534678,
0.0380027853,
-0.0153173525,
-0.0441928431,
0.0114543261,
-0.0171237756,
-0.0224271603,
0.0445224755,
0.0074083963,
-0.0415371321,
-0.0378938615,
-0.0308114141,
-0.0216276627,
-0.0032943324,
-0.0379448198,
-0.0290649552,
-0.0165365282,
-0.0037870007,
-0.0626889393,
-0.046697855,
-0.0147947939,
-0.0126180975,
0.0174618866,
-0.0218145791,
-0.0061276965,
-0.031450741,
-0.0041304561,
0.0349177346,
0.0205197595,
-0.0014981293,
-0.0432894602,
0.0147299804,
0.016274482,
-0.0856405273,
-0.0329351276,
0.0101683298,
-0.0335840918,
-0.039408803,
0.028589651,
0.0448232815,
0.0306838769,
0.0069272285,
0.0334101282,
-0.0391543843,
0.0335322171,
-0.0050912825,
0.0158653706,
0.0502112098,
0.0064210352,
0.0673207194,
0.0120044909,
0.0162318461,
-0.0274846777,
0.0186196063,
0.0289550927,
0.0192811005,
0.0021508583,
-0.00607291,
0.0360868275,
-0.0388546623,
0.0453567281,
-0.0273972508,
-0.0091852397,
-0.0139204469,
0.002404135,
-0.0022435798,
-0.0125830891,
0.032696154,
-0.0484897271,
0.042995654,
0.0453655832,
-0.0052327132,
-0.0663758442,
-0.0049723848,
0.0325390659,
0.0025912775,
0.0028965138,
0.0375702642,
0.0244643036,
0.046160236,
0.039313253,
-0.0207862183,
0.0166985597,
-0.0283480771,
0.0543247461,
0.0067064981,
0.0536575429,
0.0343344174,
-0.0374777243,
-0.014986706,
-0.0000554305,
-0.0643359497,
-0.0424039066,
0.0449355319,
-0.0395170674,
0.0228698831,
-0.0401050337,
0.0529187508,
-0.0632340536,
0.0208047181,
-0.0162288826,
-0.036968179,
-0.0242255274,
-0.0286042467,
0.0228542089,
0.0115100723,
-0.0070002899,
0.0013313197,
0.0541114695,
0.0008602095,
0.0912838206,
-0.0644483939,
-0.009719979,
0.0187073592,
-0.0325177833,
0.0081289103,
-0.0149782142,
-0.0517798066,
-0.0255677719,
-0.0602997541,
-0.0381504446,
-0.047890842,
0.0311484449,
0.0060296576,
-0.05849424,
0.0019107353,
0.0156829059,
-0.0143985664,
-0.0255172905,
0.0054606297,
0.0321182683,
-0.0140333576,
-0.0050422484,
-0.0799859911,
-6.433820006e-33,
0.0025020519,
0.0046177544,
0.0190458186,
-0.0350949951,
-0.0236894339,
-0.0190492757,
-0.0127937626,
0.0488383919,
-0.0215498544,
0.0011542684,
-0.0140566053,
-0.0260289907,
0.0238719489,
-0.0474121384,
0.0638578087,
-0.0228634868,
-0.0004226207,
0.0042429897,
-0.0330062993,
-0.0261893291,
0.013551061,
-0.002527158,
0.0513007231,
-0.0538788438,
0.0167496596,
0.0497576259,
-0.0190170053,
-0.0132923955,
-0.0424334817,
0.035179764,
-0.0203025658,
0.0106969327,
0.0108891511,
-0.0442829318,
-0.0308064613,
0.004201245,
0.0338735655,
-0.0352192596,
0.0557770431,
-0.0173470974,
0.0355874039,
-0.0199232027,
0.006692179,
-0.0161609799,
-0.0692570806,
0.0568036698,
-0.0073507638,
-0.0266700741,
-0.0351613797,
0.0555861183,
-0.0444440991,
-0.0085678091,
0.005150666,
0.0357641093,
0.0097266398,
0.0358762182,
-0.0366729237,
0.0267055072,
-0.0211714916,
0.0739224106,
0.0436490998,
0.0124490038,
0.0174679365,
0.0298545621,
0.0251852553,
0.0362601317,
0.0870787129,
-0.0045018829,
-0.0292253438,
-0.0392985344,
-0.004927678,
0.0094452575,
-0.0073122331,
0.0151580796,
0.0366626419,
-0.0228405353,
0.0052945758,
-0.0150627457,
0.0441245139,
0.0343224891,
0.0743459463,
0.0300143976,
-0.0204924233,
-0.0480949208,
-0.0708955973,
-0.033034347,
-0.0416726954,
-0.0322689787,
0.0376894288,
0.0061132163,
0.0448143519,
-0.0131536089,
0.0147105996,
0.0119208321,
-0.0345522761,
-0.0210514981,
-0.0452180877,
0.0311540812,
-0.0142230755,
-0.016924955,
-0.0350186639,
-0.0038292448,
-0.0147178294,
-0.0018795327,
0.0188571047,
0.0531161726,
-0.0291762948,
0.0041967318,
-0.0309360661,
0.0221975893,
-0.0423028804,
-0.0361288302,
0.053455811,
-0.0455734283,
0.0530715361,
-0.0345814042,
-0.0074879806,
-0.0044604479,
0.0295744706,
-0.0174268261,
-0.0548130088,
-0.0108816046,
-0.0195604153,
0.0550565682,
-0.0003721185,
0.0194436815,
-0.0456883162,
0.0604351424,
0.0122176064,
-0.0284491461,
-0.0187424943,
-0.0085648224,
2.831e-7,
-0.0023899563,
0.0831834301,
-0.0191030968,
0.0363470726,
0.0138765275,
0.0845620483,
-0.0196206346,
-0.0106392074,
0.0615996271,
0.0347415283,
0.002896182,
-0.0038445895,
0.0061147502,
0.0105148433,
0.0252520274,
-0.0293991026,
-0.0684490129,
-0.0033864016,
-0.0081885271,
-0.0254938509,
0.0072170598,
0.0367083997,
0.0272109788,
-0.0003658312,
-0.003433466,
0.0368518122,
0.0460271612,
-0.0716318712,
0.0222717933,
-0.0140987923,
-0.054658398,
-0.0137379188,
0.0057467059,
0.0058625517,
0.0260433145,
-0.0044997591,
-0.0333607495,
0.0360793136,
0.0118821766,
-0.0310593359,
0.0023058951,
0.0183437094,
0.0015736142,
-0.083883889,
0.0548195541,
-0.0335184783,
-0.0119635845,
-0.003847711,
0.021187149,
-0.0193501879,
0.0413477272,
-0.045665998,
0.0093793608,
-0.0025637536,
-0.0110044042,
0.0945899487,
-0.0109669929,
-0.0595534556,
-0.0501698032,
0.0438830853,
-0.0362946726,
-0.0557318926,
-0.0573475175,
0.0733767077,
0.0186482575,
-0.0596882999,
-0.0131838527,
2.287644939e-34,
0.0059623164,
0.0221944116,
0.0341769643,
0.0004725027,
0.0257767159,
0.0368177556,
0.0145846186,
0.0387393944,
0.0598795004,
-0.0321384966,
-0.027392162
] |
http://arxiv.org/abs/2401.14526v1 | http://arxiv.org/pdf/2401.14526v1 | MEDs for PETs: Multilingual Euphemism Disambiguation for Potentially Euphemistic Terms | 2024-01-25T21:38:30 | 2024-01-25T21:38:30 | Patrick Lee, Alain Chirino Trujillo, Diana Cuevas Plancarte, Olumide Ebenezer Ojo, Xinyi Liu, Iyanuoluwa Shode, Yuan Zhao, Jing Peng, Anna Feldman | cs.CL | cs.CL | This study investigates the computational processing of euphemisms, a
universal linguistic phenomenon, across multiple languages. We train a
multilingual transformer model (XLM-RoBERTa) to disambiguate potentially
euphemistic terms (PETs) in multilingual and cross-lingual settings. In line
with current trends, we demonstrate that zero-shot learning across languages
takes place. We also show cases where multilingual models perform better on the
task compared to monolingual models by a statistically significant margin,
indicating that multilingual data presents additional opportunities for models
to learn about cross-lingual, computational properties of euphemisms. In a
follow-up analysis, we focus on universal euphemistic "categories" such as
death and bodily functions among others. We test to see whether cross-lingual
data of the same domain is more important than within-language data of other
domains to further understand the nature of the cross-lingual transfer. | MEDs for PETs: Multilingual Euphemism Disambiguation for Potentially Euphemistic Terms: This study investigates the computational processing of euphemisms, a
universal linguistic phenomenon, across multiple languages. We train a
multilingual transformer model (XLM-RoBERTa) to disambiguate potentially
euphemistic terms (PETs) in multilingual and cross-lingual settings. In line
with current trends, we demonstrate that zero-shot learning across languages
takes place. We also show cases where multilingual models perform better on the
task compared to monolingual models by a statistically significant margin,
indicating that multilingual data presents additional opportunities for models
to learn about cross-lingual, computational properties of euphemisms. In a
follow-up analysis, we focus on universal euphemistic "categories" such as
death and bodily functions among others. We test to see whether cross-lingual
data of the same domain is more important than within-language data of other
domains to further understand the nature of the cross-lingual transfer. | [
0.0558292344,
0.0148589574,
-0.0250008795,
0.0116471564,
-0.0089904303,
0.0348493531,
0.0228604935,
0.0350498073,
-0.0063236956,
-0.0365521125,
-0.0469412319,
-0.0062183402,
-0.0275227055,
-0.0230796058,
0.0063161901,
-0.0336043201,
0.0646757334,
0.0105670597,
-0.044287961,
-0.0305901486,
-0.0511673093,
0.027846355,
-0.0050593973,
0.03676752,
0.0009676319,
0.0244520176,
0.0401511341,
-0.015166535,
-0.0035201809,
-0.0109186806,
-0.012040982,
0.0048967125,
-0.0053056255,
0.0236409474,
0.0000018797,
-0.0682652444,
-0.0164520815,
-0.0017405995,
0.0162324756,
-0.0707284138,
0.0730793253,
-0.0917771906,
-0.0055443454,
-0.0351247154,
-0.011847062,
-0.0223779157,
0.0986813679,
0.0142703634,
-0.0014588385,
0.0284507982,
-0.0133246565,
-0.0570408627,
-0.0133118685,
-0.0188519694,
0.0229998045,
-0.0346913524,
0.034899734,
-0.0293616075,
-0.0279874485,
-0.0677401572,
-0.0788321942,
0.0654739439,
0.0075801499,
0.0431842692,
-0.0387618169,
0.0341063514,
-0.0129149891,
-0.0365552157,
0.0071492796,
0.0340263583,
-0.0573248006,
-0.0147286272,
0.0054189917,
0.0458092429,
-0.0099161481,
0.0579187907,
-0.0262428615,
0.0154404659,
-0.0024043373,
0.0101877758,
0.0444935113,
-0.0098725967,
0.0076847458,
0.0365700759,
-0.0191885959,
0.0532293655,
0.0357384868,
-0.0178429931,
0.0546930693,
-0.013260643,
-0.0229542721,
0.012930695,
0.0477742888,
0.0029252225,
-0.0179730039,
0.0417305455,
-0.022985369,
-0.032234218,
0.0445760787,
-0.0844058841,
0.0188953243,
0.0030948333,
-0.0036177316,
0.0644111633,
-0.0438860282,
0.0018055709,
-0.0168202203,
0.0236382894,
-0.0359962992,
-0.0256456677,
-0.0666433796,
-0.0047980119,
-0.0320740603,
0.0297166407,
0.0264564045,
0.0095198704,
-0.0840715468,
-0.0097863702,
-0.0275704153,
-0.0340834856,
-0.002942153,
0.0198641568,
-0.0796564147,
-0.0093051754,
-0.0166915115,
0.0290873908,
-0.0354575291,
0.0130747817,
-0.0772840977,
-0.0174835771,
0.0111641604,
-0.0100066178,
-0.04988629,
-0.0667588264,
0.0262551885,
0.1087245792,
0.0066868826,
0.0037410334,
0.0284529775,
-0.0015783212,
-0.0250606686,
-0.0573786832,
0.0305234902,
-0.0656436384,
0.0249739792,
-0.0137301283,
0.0259838272,
-0.0468435548,
0.0085816644,
-0.0267677456,
-0.0992633477,
0.0888192579,
-0.0175736081,
-0.0119794905,
0.0331974253,
0.0097256741,
-0.016406782,
0.0622560754,
0.003013205,
-0.0370003134,
0.0489464439,
-0.036092829,
-0.0173786338,
0.0177776571,
-0.0103078531,
-0.0139137059,
-0.0084943725,
-0.0182173941,
-0.0137454774,
0.0062195356,
0.0207742248,
0.040921852,
-0.0299797282,
-0.0206426494,
-0.0163529646,
0.0605804846,
0.0103386808,
0.035146378,
0.0568637364,
-0.0195825491,
-0.0130165964,
0.0287238862,
-0.0223713219,
-0.0331853144,
0.0159900114,
-0.0127523281,
-0.0307613313,
0.0314769521,
-0.0052855234,
-0.0032002665,
0.0150345182,
-0.0291171446,
0.0191968698,
-0.029502701,
0.0749786571,
0.0346404165,
0.0037272656,
0.0967850536,
0.0475979522,
-0.0382402651,
0.0026019434,
-0.0313590318,
-0.0563683547,
0.0683953464,
0.0209867004,
-0.0198961981,
-0.0273056161,
0.0006512135,
0.0340724699,
0.0372270159,
-0.0139277186,
-0.0011220656,
0.035819795,
-0.0179819763,
0.0139842452,
0.0116932681,
0.0279061459,
-0.0087070605,
-0.0146589521,
0.044907935,
-0.0079794396,
-0.0107827382,
0.0141191557,
0.0252807681,
-0.0408567712,
-0.0263476819,
-0.0069118822,
0.0013494797,
-0.0083296197,
-0.01521545,
-0.022624651,
0.0071004191,
0.0112584867,
-0.0507530794,
0.0066756019,
0.0520423874,
0.0071474966,
0.0099856481,
0.0187320206,
-0.0579052158,
-0.0104389032,
-0.0157650448,
-0.012566857,
-0.0084409397,
0.0059730909,
-0.0569228418,
0.0557790883,
-0.0147806779,
0.055960156,
-0.036762815,
0.0187881049,
-0.0160710178,
0.033244729,
-0.0061554909,
0.0626250803,
-0.0030527678,
-0.0393962562,
0.0044207843,
0.0319442861,
-0.0558199473,
0.037763454,
0.0032831402,
-0.0076427157,
-0.0075324811,
0.0180911031,
-0.0508863851,
-0.012898487,
-0.0737457424,
0.0176862907,
0.0189658329,
0.0102186631,
0.0277927164,
-0.0099306004,
-0.0047248965,
-0.0250405036,
0.0186036043,
-0.0026476788,
-0.0104473038,
0.060953185,
-0.000478161,
-0.0252071582,
-0.0224119481,
0.0011763524,
-0.0025458045,
-0.0312391482,
0.0326821543,
-0.039015919,
-0.0304001532,
0.0130212614,
0.0362807289,
-0.068297036,
-0.020209631,
0.0066143367,
-0.053697288,
0.0133998794,
0.0537697263,
0.0266683735,
0.0402639546,
-0.0087931687,
0.0122731514,
0.0833774582,
-0.0215890799,
-0.0002931601,
-0.0246035978,
-0.0392495543,
-0.0427147783,
-0.0171903446,
-0.0573207773,
-0.0092138471,
0.0631579384,
0.0146077424,
-0.0036777856,
-0.0234848093,
0.0475744642,
-0.0504756197,
-0.0451756306,
-0.0408638939,
0.0041255569,
-0.0099715376,
0.0028488932,
0.0066012717,
-0.0092824819,
0.0884955823,
0.0557335839,
-0.0038416509,
-0.0441430993,
0.0108300364,
0.0139621375,
0.0077177631,
-0.0372683816,
0.04014856,
0.0020378192,
-0.0241266061,
-0.036145214,
-0.016319681,
0.0044593406,
-0.0518536605,
0.0537852757,
-0.0365626365,
-0.0152751058,
0.02933673,
-0.0284906179,
-0.0226956196,
0.0248214845,
-0.0044507715,
0.0236166026,
0.010512365,
0.0283793714,
-0.021275593,
0.0022295895,
-0.0027067917,
-0.0189071074,
0.0163070261,
-0.0273625776,
0.0048636729,
-0.0108763827,
0.037306942,
-0.0501133502,
0.0126597481,
-0.0143079469,
0.0362755507,
0.0516335294,
-0.0207452383,
-0.0353414193,
0.0427061208,
0.0318310969,
0.0240801685,
-0.0367573164,
-0.0168609619,
-0.035467498,
-0.0117249377,
-0.0362918079,
-0.0110118734,
0.0842523947,
0.0349151231,
-0.0516348444,
0.0421310253,
-0.0357891321,
-0.0357865132,
0.0026191142,
0.0162303653,
-0.0926216766,
-0.0109472889,
-0.0227052402,
-0.0172254499,
-0.0010128714,
-0.0092543606,
-0.0071378015,
-0.0730826631,
0.0069730738,
0.0034105054,
-0.0597810112,
0.0024740617,
-0.0163405016,
0.0062542893,
-0.0187212769,
-0.0004198712,
-0.0382123366,
-0.0390934423,
-0.0227340739,
-0.0268621575,
0.0161196683,
-0.0378023572,
-0.0030769638,
-0.0252948105,
0.0363394283,
0.0064659803,
0.0557085574,
-0.0212011952,
-0.039910309,
0.0184981208,
-0.0020721201,
-0.0157711115,
-0.0200060848,
0.0333886743,
0.0199407805,
0.0027362143,
0.004318893,
-0.0434721187,
-0.0126527026,
-0.0417133309,
-0.0115819843,
0.0491094701,
-0.0122518046,
-0.0674870312,
-0.0218523722,
0.0455190726,
-0.0067138318,
0.0838038325,
-0.0031029414,
-0.0264816023,
-0.0206882246,
-0.0046230187,
-0.0102618132,
-0.0160841718,
-0.0563196763,
-0.0083219269,
-0.1379217654,
-0.0955842435,
-0.0620022379,
0.0037727435,
0.04394399,
-0.0086252932,
0.0347961076,
0.0111557152,
-0.0250266232,
-0.0058655329,
-0.0459906757,
0.0255838931,
0.0318149589,
-0.0203489568,
-0.0178273749,
-0.0339706503,
0.0346782207,
0.0419255346,
-0.0223628096,
0.0307090078,
0.0351811424,
-0.0501970537,
0.047719676,
0.053370513,
0.0240669586,
-0.0037880496,
-0.0064708823,
-0.0091078263,
0.0377175584,
0.0628166795,
0.0186521113,
-0.0103127314,
0.0290826447,
0.0320523456,
0.0036889887,
0.0419238433,
-0.0118305115,
0.0476409569,
0.0045021316,
0.0360533856,
-0.0478970148,
-0.0111191561,
0.0118507389,
-0.0145756109,
0.0126566282,
-0.0633315295,
-0.0303981267,
0.0070878705,
-0.0112351533,
-0.0141457506,
-0.0241223052,
0.0284705311,
-0.0226935577,
0.0529868081,
0.0098696863,
0.0002295831,
0.0052498295,
-0.037929181,
-0.0556168221,
0.0159525685,
-0.0233202223,
0.0543870293,
-0.020955991,
-0.0009778205,
0.0534978472,
0.01138095,
0.0043971194,
0.0278652832,
0.0146122482,
-0.0407934375,
0.0467582606,
-0.0235425122,
0.0002899913,
-0.0323722214,
0.0431314521,
0.0052489648,
-0.0571050793,
0.0111383712,
-0.0261167157,
0.0624543466,
-0.0561800078,
0.0531778783,
-0.015454378,
-0.0101700025,
0.0069860029,
-0.0039180457,
-0.0189585704,
0.0191586744,
0.0282070506,
0.0063698683,
-0.0083565451,
-0.015179486,
0.0691271275,
-0.0150992293,
0.0533487424,
0.0038304741,
-0.0354278982,
0.019898532,
-0.0228883214,
-0.0101782177,
-0.0165677443,
-0.0251737069,
0.0123784058,
-0.0636981651,
-0.0537641682,
-0.016721664,
-0.0395311043,
-0.0134996483,
0.0115399109,
0.0390201882,
0.0913286209,
0.0362538509,
-0.0134868613,
-0.0670686811,
0.059154965,
0.03273274,
-0.0141115775,
0.005679749,
-5.597563039e-33,
-0.0197277255,
-0.0230883081,
-0.0059741619,
0.0269627627,
0.0044182972,
0.0086705489,
-0.0021353343,
0.0224241223,
-0.0030789042,
0.0401920639,
0.016160598,
-0.0021095979,
0.0173737835,
-0.0143788252,
0.0541628487,
0.003318846,
-0.0012900883,
0.0330667384,
-0.0156591553,
-0.020613648,
0.0590424612,
0.0659775361,
0.0995808914,
-0.0420609079,
0.0640182793,
0.0746336877,
0.0228495225,
-0.0015893563,
0.0649908707,
0.0511575714,
-0.0272588469,
0.0631839857,
0.0390335992,
0.0076673217,
0.0084957015,
0.0064949575,
-0.0604340993,
-0.0091333445,
0.0948300064,
0.0020138877,
-0.0095360894,
-0.0727369636,
0.0868210793,
-0.0184666421,
-0.035366822,
0.0203334596,
0.0208109859,
-0.0527238734,
-0.0025774515,
-0.031383086,
-0.1116038486,
-0.02252247,
-0.0016585145,
-0.0178202875,
0.0442833044,
0.0088772569,
-0.0283041969,
0.021153247,
-0.0227931552,
0.032768622,
-0.0405585282,
0.0261045583,
0.0463107452,
-0.050045494,
0.0187658593,
0.0561674573,
0.0080704484,
0.0011605192,
0.0058756801,
-0.0140485046,
0.0432853624,
0.0606543086,
0.0572041869,
0.0311741829,
0.0728774518,
-0.0247640349,
0.0038106826,
-0.0054064607,
-0.0070539038,
0.0576905347,
0.016153574,
-0.027031919,
-0.0000543503,
-0.0195957124,
-0.0201667249,
0.0433580019,
-0.0157960951,
-0.0654416829,
-0.0030370182,
-0.0364133343,
-0.0098156249,
0.0180248301,
0.0059914528,
-0.0345202535,
-0.0062144445,
-0.0453490838,
0.0365039669,
0.0202148985,
0.0017175672,
0.0091357641,
0.0026600182,
-0.0237804614,
-0.0206765458,
0.0043804678,
0.052839648,
0.0049487869,
0.0301866271,
-0.0257250629,
-0.0580355264,
0.0105086556,
-0.0187778845,
-0.0360560678,
0.0158344507,
0.036751274,
0.0098197656,
0.0421432331,
-0.0100202756,
-0.0511307344,
0.0683398098,
-0.0113364914,
-0.024125535,
0.0148373088,
-0.0212849714,
0.0450953543,
-0.036116261,
-0.0144877713,
-0.0497814529,
0.0264507905,
0.0457199067,
-0.0629309863,
-0.0267413389,
-0.0557671823,
2.76e-7,
0.09737093,
0.0776175782,
-0.0468525738,
0.0197011083,
0.0126062147,
-0.0152240805,
-0.0257201921,
0.0224421583,
-0.0088072345,
0.0086489823,
-0.0083863093,
0.0143966097,
0.0245669559,
0.0838343054,
-0.0443114229,
-0.0223365575,
-0.0070801862,
0.0277025253,
-0.0380591713,
-0.0058900253,
-0.0326244347,
0.0438591726,
-0.0026466502,
-0.0405705832,
-0.0527821556,
-0.0282896664,
0.0219749175,
-0.090746887,
0.0049224971,
0.0315118544,
-0.0057970393,
0.0572566837,
0.0293278061,
0.0435444266,
-0.020615479,
-0.0151588349,
0.0532695241,
0.0429636762,
-0.0401085466,
-0.0015984338,
0.0421710871,
-0.0614226982,
-0.0114504285,
-0.0508912131,
0.0535868928,
-0.048008956,
-0.0207429491,
-0.0242642574,
-0.0021731141,
0.034807086,
0.0844463035,
-0.0741218403,
-0.0637731925,
0.0312013309,
0.0438221171,
0.0542525053,
-0.0100240801,
-0.0321923904,
-0.0033925434,
0.0667378008,
0.0086192889,
-0.0332533419,
0.0288412441,
-0.0610421263,
-0.0220282022,
-0.0750346631,
0.0163917467,
1.861717303e-34,
0.0332537591,
0.0290056467,
0.0511759445,
0.0424598865,
0.0103089986,
-0.0262184255,
0.0109766582,
0.0321402065,
-0.0302883685,
-0.0124325082,
0.0185597222
] |
http://arxiv.org/abs/2401.14524v1 | http://arxiv.org/pdf/2401.14524v1 | Evaluating GPT-3.5's Awareness and Summarization Abilities for European Constitutional Texts with Shared Topics | 2024-01-25T21:34:53 | 2024-01-25T21:34:53 | Candida M. Greco, A. Tagarelli | cs.CL | cs.CL, cs.AI, cs.CY, cs.DL, physics.soc-ph | Constitutions are foundational legal documents that underpin the governmental
and societal structures. As such, they are a reflection of a nation's cultural
and social uniqueness, but also contribute to establish topics of universal
importance, like citizens' rights and duties (RD). In this work, using the
renowned GPT-3.5, we leverage generative large language models to understand
constitutional passages that transcend national boundaries. A key contribution
of our study is the introduction of a novel application of abstractive
summarization on a multi-source collection of constitutional texts, with a
focus on European countries' constitution passages related to RD topics. Our
results show the meaningfulness of GPT-3.5 to produce informative, coherent and
faithful summaries capturing RD topics across European countries. | Evaluating GPT-3.5's Awareness and Summarization Abilities for European Constitutional Texts with Shared Topics: Constitutions are foundational legal documents that underpin the governmental
and societal structures. As such, they are a reflection of a nation's cultural
and social uniqueness, but also contribute to establish topics of universal
importance, like citizens' rights and duties (RD). In this work, using the
renowned GPT-3.5, we leverage generative large language models to understand
constitutional passages that transcend national boundaries. A key contribution
of our study is the introduction of a novel application of abstractive
summarization on a multi-source collection of constitutional texts, with a
focus on European countries' constitution passages related to RD topics. Our
results show the meaningfulness of GPT-3.5 to produce informative, coherent and
faithful summaries capturing RD topics across European countries. | [
0.0338893682,
0.0449052975,
0.0322779194,
0.0045154123,
-0.0363793038,
0.0381342284,
-0.0258033555,
-0.037585482,
-0.0336031653,
-0.0296794102,
0.0441178493,
-0.0677491277,
-0.0085667558,
0.0074113198,
0.0094003519,
-0.1039820313,
0.0358916819,
-0.0168700684,
0.0085916985,
-0.0126221534,
0.0416807681,
0.0105445208,
0.0210019629,
0.0227069873,
0.0253062099,
0.0201326422,
-0.0216662902,
-0.0377926789,
0.0458185673,
0.0078225359,
-0.0404394567,
0.0228134003,
0.0272797868,
0.0304713864,
0.0000016145,
-0.0606833473,
-0.0272569954,
0.018840082,
0.0851898491,
0.0720433444,
0.0692442432,
-0.0367335044,
-0.0314797387,
0.0097314008,
-0.0096882321,
0.0179846976,
-0.0094276108,
0.0543157533,
0.0551267192,
0.0629563555,
-0.0134571893,
-0.0421615802,
0.0371637642,
-0.0595215037,
-0.0151218036,
-0.0303076506,
0.0059746169,
0.0397531763,
-0.0293117967,
-0.0074392706,
0.017851932,
0.0189310554,
0.0031967505,
-0.024865143,
-0.0400091149,
0.0106362281,
-0.0573168024,
0.0044263317,
0.0393905789,
0.0148144029,
0.0752086565,
-0.0065937852,
0.0635511652,
-0.002299614,
0.0045090695,
0.0616052002,
-0.081481874,
0.0597953647,
-0.0202663317,
0.0479002781,
0.0660846904,
0.0220858473,
-0.0071698297,
-0.0026141468,
-0.0629663393,
0.0809031427,
0.0179226231,
-0.0130161922,
-0.0056239041,
0.0089505184,
-0.0752554908,
-0.0434125662,
0.0172262397,
0.0174856596,
0.0192272458,
-0.0114805996,
-0.0100793615,
-0.0545916036,
0.0578610227,
-0.0200466197,
-0.0310356449,
0.0141328517,
0.0035018201,
0.0652390271,
0.0210831799,
0.0026700455,
0.0109852655,
-0.0181380883,
-0.0341511257,
-0.049237337,
-0.0406420827,
0.0208279509,
-0.0503893755,
0.0240521226,
-0.0701966658,
-0.0106990058,
-0.0205651596,
0.0170413014,
0.038956251,
0.0137263238,
-0.0182587691,
-0.028815994,
-0.0219789799,
-0.0266856775,
-0.0270756017,
-0.0084602106,
-0.0674439371,
-0.0331016034,
-0.0290160012,
0.0252386481,
0.0114306323,
-0.0060641179,
0.0263463221,
0.019109223,
-0.0133131882,
0.0000901852,
-0.0191970337,
-0.0166481361,
0.0025960193,
-0.0122730937,
-0.0293985698,
-0.0614457913,
-0.0168941449,
-0.075838089,
0.0199035686,
-0.0124467146,
0.0046628029,
-0.0037405675,
0.0135693401,
0.0198595449,
-0.0186493155,
-0.007554926,
0.0085125556,
0.0082717594,
0.0058023669,
0.001311983,
0.0330543071,
0.0151179768,
0.0223538335,
0.0336044133,
0.0322426148,
0.0238154884,
0.0249649994,
0.0287496094,
-0.0096959639,
0.05248826,
0.0597874857,
-0.0147692049,
-0.035445191,
-0.0448741615,
0.0284241308,
0.0658024251,
0.0002458525,
0.0349634625,
0.0469043665,
0.0812894404,
-0.0278729033,
0.0517130122,
0.0249977093,
0.0185876582,
0.0209107194,
0.0579767451,
-0.009934728,
-0.0154994708,
-0.0229580551,
0.0540768914,
0.0268354751,
0.0443670414,
0.0039802315,
-0.0281848554,
0.0252105836,
0.0086820554,
0.0025039727,
-0.0153065464,
0.0645441711,
-0.0190385934,
-0.0289659798,
-0.0100857997,
0.0014792698,
-0.0890728012,
0.030383328,
0.0345597193,
0.037244793,
0.0038906883,
0.0297799762,
-0.035689462,
0.0092362007,
-0.0257744491,
0.038280949,
0.0605083928,
0.0609745905,
-0.0197372641,
-0.020274736,
-0.0286598448,
-0.0469564199,
-0.0126258601,
0.0224219803,
0.0296531562,
-0.008741932,
-0.0024829081,
0.0527227186,
-0.0549158081,
-0.0045829075,
0.0069190608,
-0.0273857042,
0.0104396194,
-0.0410337672,
-0.0213884804,
0.0114937266,
0.0451451838,
-0.0042400532,
0.0112280268,
0.0064213234,
-0.0293427128,
-0.0185973328,
0.022033304,
-0.0241820086,
0.0257850718,
0.0251977053,
-0.0701917484,
0.0070378925,
-0.0047735199,
-0.0573222451,
-0.0546833239,
0.0251812004,
-0.0315176845,
0.017275501,
-0.0511518903,
0.0133106317,
0.016777847,
-0.0224425942,
-0.0055635767,
0.0740541667,
-0.0128335692,
-0.0078562628,
0.0432504192,
-0.0215195809,
0.0343021639,
-0.0385751799,
-0.0184778702,
0.0029395691,
0.0078954743,
-0.0004620226,
0.0750840679,
-0.006704736,
-0.0125628775,
-0.0172387455,
-0.06403061,
0.039601665,
-0.0314513333,
0.0489118807,
0.0481885783,
0.0334825329,
-0.0027552657,
0.0231889244,
-0.0556455478,
-0.0363578461,
0.0104925046,
-0.0354253389,
0.018669894,
-0.0032383909,
0.0071403864,
-0.0128383785,
0.0011596699,
-0.0269742776,
0.0131107206,
0.0182912592,
-0.0213176422,
0.0009721548,
0.0381589048,
-0.0277979057,
0.0107509708,
-0.0338831991,
0.0044863103,
-0.0387582146,
0.0039519263,
0.07445281,
-0.0334516875,
0.0474988967,
-0.003618472,
0.0384400785,
-0.0000876351,
-0.0123132793,
-0.0092264283,
-0.0587094799,
0.0110932617,
-0.0241935346,
-0.0231428239,
0.0626393557,
0.0390559621,
0.0231158026,
-0.0076912548,
-0.0475474112,
0.025215771,
-0.0779997706,
0.0213291645,
-0.0543319695,
-0.0114944782,
-0.0048744888,
0.0387333222,
0.0035138146,
-0.0235900115,
0.0216774438,
-0.0245264564,
0.0520217307,
-0.0443556607,
-0.0175377056,
-0.0686697289,
0.0718295574,
0.0267038513,
0.041204419,
0.0035814014,
-0.0373958834,
-0.0139679089,
-0.0334208794,
0.0653525442,
-0.0091975052,
0.1161668226,
-0.0416728407,
0.015012092,
-0.0512096882,
-0.0301609449,
0.0300305076,
-0.0362944975,
0.0180214327,
0.0536358431,
0.0274844058,
-0.0132418284,
-0.0476379544,
0.0343737304,
0.0527096465,
0.0398486927,
-0.0062553193,
-0.0073649464,
-0.0275737494,
0.0616783425,
0.0431819074,
-0.0369683765,
0.0018847629,
-0.0453450456,
-0.0352522023,
0.0395246595,
0.0256498475,
0.008788202,
0.0332058296,
-0.0050171046,
-0.0237638522,
-0.1044280902,
-0.0103324959,
-0.0249966867,
0.027153613,
-0.0221001338,
0.0168455001,
0.1125513315,
0.0264351983,
-0.0468081012,
-0.0037619257,
-0.0135692675,
0.0154969376,
0.0133996606,
0.0007868103,
-0.0873142853,
0.0587409027,
0.0129155563,
0.0046174908,
-0.0207009334,
-0.0297893882,
-0.007703328,
-0.0678212717,
0.0604230314,
0.0298097171,
-0.0852929875,
-0.038787365,
-0.0208291877,
0.0199123546,
-0.0312831327,
-0.0254160743,
0.0302889515,
-0.0283884965,
0.0298426319,
0.0006854031,
-0.1039837673,
-0.0044492446,
-0.00056144,
-0.0338465162,
-0.0312014017,
0.0512655266,
0.0604130812,
-0.0244591311,
0.0142560275,
0.0218925606,
-0.0822225064,
-0.042236384,
0.0346853845,
0.0224169493,
0.0352410711,
0.0108907754,
-0.0290111247,
0.0619761609,
0.0220012106,
-0.0432402939,
0.0346265174,
0.0460981615,
-0.0158742424,
-0.031433396,
-0.0085971728,
-0.0022123454,
0.0051540481,
0.0630404651,
0.0169884656,
0.0354048908,
-0.0369356722,
-0.0490630679,
0.0067912457,
-0.0709434375,
-0.0134762889,
-0.0202312004,
-0.0142452661,
-0.034680333,
-0.0425509885,
-0.0218776781,
-0.002328834,
0.0008755199,
0.0193542875,
0.0264298711,
0.0316299275,
-0.0288242511,
-0.0148468781,
-0.0001493741,
-0.0551430434,
-0.0429828279,
-0.0215667151,
-0.0621907115,
0.0324968062,
0.0276150294,
-0.0547088049,
0.0300042927,
-0.0406995378,
-0.0491517559,
-0.0026993912,
-0.0371003933,
0.0193547606,
0.0259873234,
-0.0030013607,
-0.0138327777,
0.0492540114,
-0.0143965716,
0.0091528157,
0.0125837354,
-0.0115793375,
0.0242810994,
-0.0140725086,
0.0269913059,
-0.0429798402,
0.0420262665,
-0.0362408943,
0.0325668342,
-0.000920328,
0.0079992991,
0.0031156379,
0.0578792058,
-0.0038533038,
-0.0416026041,
-0.0044301557,
-0.0512634069,
-0.047761213,
-0.01427841,
0.0035056763,
-0.0263540987,
0.034309268,
0.0283282418,
-0.0227834806,
-0.0283995289,
0.0704125091,
0.0061489646,
0.0032748191,
-0.0611568578,
0.0254855063,
-0.0049974336,
-0.0154466731,
0.0039315014,
-0.0441366024,
-0.0218041558,
0.0071854703,
-0.0556163155,
0.0267465338,
-0.0357220471,
0.0149001451,
0.0136466091,
-0.0025543354,
0.0203390904,
0.0218269285,
-0.0638477877,
-0.0079742093,
-0.0698977858,
0.0444999486,
0.0535476394,
-0.0358262733,
0.0672857016,
-0.0369623192,
-0.0059872437,
-0.0187272839,
0.0217582043,
0.0235462505,
-0.0139338579,
-0.0221133064,
0.0124223633,
-0.0132494261,
0.0416084826,
0.0628303438,
-0.0216501206,
0.0064191166,
0.054023739,
-0.0310974959,
-0.0265710708,
-0.026994789,
-0.0286619663,
-0.0141961342,
-0.0427386649,
-0.0523952283,
-0.0833655819,
0.0046164733,
-0.0199887492,
0.0440064631,
-0.0167237651,
0.0018570162,
0.0215753671,
0.0302709136,
-0.0252113435,
-0.0039719241,
-0.0677582473,
0.0092522651,
-0.0428992547,
-0.0060727894,
0.0031106893,
-5.314376532e-33,
-0.0630640835,
0.0081183435,
-0.000061785,
0.0278976653,
-0.044965405,
-0.0415335186,
-0.0587406233,
-0.0416911282,
0.0376009941,
0.0193178542,
-0.0050843623,
0.0341023952,
0.0171372201,
0.0152306817,
-0.0139924344,
0.0077106394,
0.0381588414,
0.0234471224,
-0.0370736606,
0.0913955942,
0.0449213758,
-0.0061267237,
0.072454527,
-0.0382879414,
0.0417675152,
0.0081674373,
-0.007902476,
-0.054744985,
0.0182706136,
0.0321091972,
-0.0235536955,
0.0130958818,
0.0204392914,
-0.0449332111,
-0.0199413132,
0.0403326377,
-0.0511714108,
-0.0145640681,
0.0322184376,
-0.0697033778,
0.0219773818,
-0.0494173802,
0.0076648486,
-0.0212251563,
-0.0047879033,
-0.0159339216,
0.0601655506,
-0.0190582946,
-0.0337330103,
-0.1161091179,
-0.0530405231,
0.0423256978,
-0.0352346785,
0.0630647764,
0.000677527,
0.0004587699,
0.0253695697,
0.0140408538,
0.0048231529,
0.0374748334,
0.0391872972,
0.0308819264,
-0.0075846631,
-0.0570217595,
0.029466806,
0.0326808393,
-0.0549161844,
-0.0076606963,
-0.0542438105,
0.0373625048,
-0.0042365729,
-0.0077900765,
0.0150341857,
0.0663062185,
0.0314731523,
0.0237358734,
-0.0540883057,
0.081053257,
-0.0033271504,
0.0244337078,
0.0313182436,
-0.0244485438,
0.0252921842,
-0.0086224154,
-0.040143542,
-0.0330542885,
0.032309413,
-0.0127160372,
0.0053324006,
-0.0099864965,
0.0028567442,
-0.0160202291,
0.0300161559,
-0.0238237083,
-0.0236796774,
0.0391546525,
-0.0145687098,
0.0180788245,
-0.0207751077,
0.0113442047,
-0.0409578569,
-0.0104365507,
0.0010528259,
0.0216868706,
-0.0032867205,
-0.0095516294,
-0.0030389116,
-0.0118418662,
-0.0453251228,
-0.0022502323,
-0.0039886413,
0.0324398652,
-0.022438379,
0.0435490124,
-0.0000285148,
-0.0076614153,
0.0438857824,
0.0040665604,
0.0170479491,
-0.0439240597,
0.0058161067,
0.0278352126,
0.0182938948,
0.0133692315,
0.017135445,
0.0301651955,
0.0061073354,
0.0187387448,
0.1099965423,
0.0225739889,
-0.0120640239,
0.0335330218,
2.318e-7,
0.0661033988,
0.0128545417,
0.0175304841,
-0.0008648239,
-0.0168742817,
0.0136832613,
-0.0020758533,
0.0528380536,
-0.0005834244,
0.0131292464,
0.0544309691,
-0.0034676874,
0.0198408235,
0.0557816997,
-0.1109692529,
0.0221747644,
-0.0460427739,
-0.0482815951,
-0.0392250232,
0.0346892662,
-0.004539114,
0.0478113778,
0.0018468577,
0.0043457393,
-0.017387569,
-0.0327622183,
0.0563630797,
-0.0510120839,
-0.0932585597,
0.025576856,
0.0624231026,
0.0006129442,
0.0134019013,
0.0013178785,
0.0145110423,
0.0075406972,
0.0281444825,
0.0644234046,
0.0324901789,
0.0433243401,
-0.0143257258,
-0.0511828624,
0.0204694979,
-0.0318807885,
0.0387203135,
-0.0342721231,
0.0000751131,
-0.0336427651,
-0.0450474657,
-0.0034682949,
-0.0103376107,
0.018186884,
-0.0202325657,
0.0300824288,
0.0060620112,
-0.0068725175,
0.0281334333,
-0.0503523722,
0.0031972914,
0.0280836523,
-0.0009475205,
-0.0370588601,
0.0051265405,
0.038719818,
-0.0144680897,
-0.0698362067,
-0.0103877224,
2.121019762e-34,
0.0880357325,
-0.0025133912,
-0.06628602,
-0.0055186264,
0.0502106547,
-0.006789783,
-0.0048061535,
-0.0112606809,
-0.0464075655,
-0.080429107,
0.0097605567
] |
http://arxiv.org/abs/2401.14523v1 | http://arxiv.org/pdf/2401.14523v1 | Empathy and the Right to Be an Exception: What LLMs Can and Cannot Do | 2024-01-25T21:30:06 | 2024-01-25T21:30:06 | William Kidder, Jason D'Cruz, Kush R. Varshney | cs.CY | cs.CY, cs.AI, cs.CL | Advances in the performance of large language models (LLMs) have led some
researchers to propose the emergence of theory of mind (ToM) in artificial
intelligence (AI). LLMs can attribute beliefs, desires, intentions, and
emotions, and they will improve in their accuracy. Rather than employing the
characteristically human method of empathy, they learn to attribute mental
states by recognizing linguistic patterns in a dataset that typically do not
include that individual. We ask whether LLMs' inability to empathize precludes
them from honoring an individual's right to be an exception, that is, from
making assessments of character and predictions of behavior that reflect
appropriate sensitivity to a person's individuality. Can LLMs seriously
consider an individual's claim that their case is different based on internal
mental states like beliefs, desires, and intentions, or are they limited to
judging that case based on its similarities to others? We propose that the
method of empathy has special significance for honoring the right to be an
exception that is distinct from the value of predictive accuracy, at which LLMs
excel. We conclude by considering whether using empathy to consider exceptional
cases has intrinsic or merely practical value and we introduce conceptual and
empirical avenues for advancing this investigation. | Empathy and the Right to Be an Exception: What LLMs Can and Cannot Do: Advances in the performance of large language models (LLMs) have led some
researchers to propose the emergence of theory of mind (ToM) in artificial
intelligence (AI). LLMs can attribute beliefs, desires, intentions, and
emotions, and they will improve in their accuracy. Rather than employing the
characteristically human method of empathy, they learn to attribute mental
states by recognizing linguistic patterns in a dataset that typically do not
include that individual. We ask whether LLMs' inability to empathize precludes
them from honoring an individual's right to be an exception, that is, from
making assessments of character and predictions of behavior that reflect
appropriate sensitivity to a person's individuality. Can LLMs seriously
consider an individual's claim that their case is different based on internal
mental states like beliefs, desires, and intentions, or are they limited to
judging that case based on its similarities to others? We propose that the
method of empathy has special significance for honoring the right to be an
exception that is distinct from the value of predictive accuracy, at which LLMs
excel. We conclude by considering whether using empathy to consider exceptional
cases has intrinsic or merely practical value and we introduce conceptual and
empirical avenues for advancing this investigation. | [
0.055776801,
0.0631502122,
-0.0438822098,
0.0279415864,
-0.0631701574,
0.0545500927,
-0.0051506758,
-0.0298877377,
0.03127921,
-0.0176424738,
0.0127482833,
-0.0268812161,
-0.0192782171,
-0.0608153604,
0.0051423595,
-0.0501722395,
0.0011027611,
0.0494436212,
0.0863244981,
-0.0287565887,
-0.0383768119,
0.0593175441,
0.0099131539,
0.0244261287,
-0.0012791946,
0.0118950577,
0.0371631943,
-0.0471898243,
0.0328160152,
-0.0027189648,
-0.0200760607,
0.0163843986,
0.0322115347,
0.0351803266,
0.0000018561,
-0.0732073635,
0.0303243082,
-0.00569478,
0.0260160845,
-0.0183313433,
0.0216018129,
-0.0455610678,
0.0482303388,
0.020972332,
-0.0619753599,
0.0146511775,
-0.0123874014,
0.0456619635,
-0.021532502,
0.0362463817,
-0.0228408035,
0.0015762091,
-0.0006360903,
-0.0027717929,
0.0303448979,
-0.0187401399,
0.0152694415,
-0.0213967673,
-0.0009028791,
-0.0853147283,
-0.0043409513,
0.0103169261,
-0.0046342122,
0.0337933227,
-0.0290831756,
0.0037271003,
-0.007568026,
-0.0281177741,
-0.004123407,
0.055139903,
0.0470591709,
0.0271248389,
-0.0150329163,
-0.0010550291,
0.0259916335,
0.0183421466,
-0.0116107296,
0.051726155,
-0.0125921471,
0.0268807542,
0.0623559989,
0.0152623635,
0.0070113353,
0.0406936072,
-0.0270662382,
0.0497395545,
-0.0389373414,
-0.0134879276,
-0.0008233205,
0.0251970999,
-0.0526339971,
-0.0255916324,
-0.0229898635,
0.0287282877,
-0.0154375052,
0.0134290149,
-0.0281686652,
0.0224639568,
0.0156357791,
-0.0396228619,
0.0684625208,
0.0615103096,
0.0090734428,
-0.0084225629,
0.0103648305,
0.0307237562,
0.0049252613,
0.0487832651,
-0.0161551218,
-0.0082421545,
-0.0555161312,
0.0051778811,
0.0271267351,
0.0524244234,
0.0152016189,
0.0152618242,
-0.0164660439,
-0.0023588543,
-0.038173262,
-0.0091040358,
-0.0837053061,
-0.0095802611,
0.0014173944,
-0.0209065974,
-0.0691172779,
-0.0155523559,
-0.058543276,
0.0303593818,
-0.021246504,
-0.0305175837,
0.0130448667,
-0.0229110979,
0.0123334033,
-0.0093783243,
0.0346625336,
0.0460965894,
0.0120808017,
-0.0548671037,
-0.0323084965,
0.013367448,
-0.0132485181,
-0.0520398393,
-0.0028494485,
-0.0037722408,
-0.0031665228,
0.0404224657,
-0.0111637795,
-0.0303742681,
0.0215654559,
0.0005411307,
-0.0706462041,
0.0480167158,
0.0334605835,
0.0270866454,
0.0025895357,
0.0123905083,
0.0280802902,
-0.0001905097,
0.0107271196,
-0.0275796056,
0.0162554421,
0.0092584547,
0.0601910762,
-0.0050601643,
-0.0411237404,
0.0017354926,
-0.0342907682,
-0.0106239412,
-0.0066736853,
0.0356728397,
0.0671826228,
0.0330876037,
0.0035449509,
-0.0003743389,
0.0387152918,
0.1329806298,
-0.017157156,
0.0529762134,
0.0023708642,
-0.0006774761,
0.0046089264,
-0.0023487331,
-0.051467482,
0.006044406,
-0.0011384926,
-0.0701794997,
-0.0395353921,
0.0154090999,
-0.0305281132,
0.0181044023,
-0.0292673949,
-0.0019153744,
0.0648887679,
0.0113297738,
-0.0144851105,
-0.0596332513,
0.0391414873,
-0.0460177623,
0.059264861,
-0.0294238441,
-0.0252082068,
-0.0888334364,
-0.0014483028,
0.0599860214,
0.0303666089,
-0.0165048558,
0.0022366315,
-0.0278535634,
-0.0609235391,
0.0593426265,
0.0685081258,
0.0022260807,
0.0022768632,
-0.0329119377,
-0.0036274088,
0.0021879666,
0.0145720635,
0.0331342556,
-0.0474811494,
0.0268488284,
-0.0087841386,
-0.0076612257,
0.001608211,
0.0460757352,
-0.0680048466,
-0.0136779854,
-0.0536721162,
0.0119032208,
0.0290707033,
0.0152823171,
0.0163242761,
-0.023332119,
-0.0155154858,
0.0060958611,
-0.016731469,
0.0726764724,
-0.042631764,
0.0302670039,
0.004825695,
-0.064179264,
-0.0067816447,
0.0536994636,
0.0214618742,
-0.0139255272,
0.0123943919,
-0.0257868767,
0.0649723783,
-0.0203249399,
0.0247839261,
-0.0882023424,
0.0170475226,
0.0063462257,
0.0260288659,
-0.0426004976,
0.0038928518,
-0.0143901659,
-0.0622674562,
0.0291503519,
-0.0174941942,
-0.0467598736,
-0.0253924485,
0.0069417469,
0.0013998415,
-0.01116772,
-0.0173331797,
-0.014817277,
0.0027291332,
-0.0957428664,
-0.0526037365,
0.0207792651,
-0.0091457339,
-0.0316038802,
-0.0104706595,
-0.044200059,
0.0428643413,
-0.0211644862,
0.0315338187,
0.0454928987,
0.0096823573,
0.0134123415,
-0.00725435,
-0.0174083002,
-0.0408096686,
0.0114288786,
-0.0504597239,
0.0089106457,
0.0115981027,
-0.0193902459,
-0.0227480251,
0.0136394287,
-0.0775636882,
0.0247594472,
-0.030449817,
-0.0521073118,
0.0000380693,
0.0298351236,
-0.0169436671,
0.0537037514,
0.0202511214,
0.0336120576,
0.043439772,
0.0411284044,
-0.0622805357,
0.0010558411,
-0.0013641067,
0.0096900631,
-0.0375397503,
-0.0473430566,
-0.0467196107,
0.0439049415,
0.0557275489,
-0.0167172756,
-0.0510199964,
0.0251546856,
-0.0092281764,
-0.0671776086,
0.027757898,
-0.0274971277,
-0.0035554881,
0.0213602707,
-0.00361774,
-0.0505888127,
0.0829406753,
0.0293697715,
0.0301759318,
-0.0118881073,
-0.0152027458,
-0.0198376812,
0.0768595934,
-0.0010212493,
0.010924031,
0.0065124431,
-0.010290511,
-0.0339641683,
-0.0121758245,
-0.038648434,
-0.0167581122,
0.0437812284,
-0.0219304897,
-0.0064858226,
0.0347293466,
-0.0168434232,
-0.0297049526,
-0.0570034534,
-0.0565584376,
-0.0183147565,
0.0125384731,
-0.0272509437,
-0.0236367136,
-0.049003467,
0.0531420261,
0.010914118,
-0.0038560315,
-0.0301518124,
0.0074760942,
-0.0904164016,
-0.0071270773,
-0.0081821289,
0.064633213,
-0.0297916811,
0.046354875,
0.0090272836,
0.0239106752,
0.0466150865,
0.0421386138,
0.0263417102,
0.0093158102,
0.0415805392,
-0.0166266803,
0.0123923784,
0.0340057649,
0.014532866,
0.0082074208,
0.1039351374,
0.0368738882,
-0.059535291,
0.0346004441,
-0.0555326752,
-0.0381841809,
0.0360897481,
-0.00811176,
-0.0540232211,
0.0295793638,
-0.0141159417,
0.0347318873,
-0.0950369015,
-0.0803530365,
0.0291607175,
-0.0523542725,
0.0317120738,
0.0149502382,
0.0444857627,
-0.0495696776,
-0.0053280392,
0.0366306081,
0.0361814909,
0.024748642,
-0.0092065493,
0.0172307752,
-0.0074693118,
-0.0050439788,
-0.0247315206,
-0.0292141195,
0.067525126,
-0.0325891562,
0.0702887848,
0.0282268133,
0.0623286366,
-0.0246089827,
-0.0054463116,
-0.0371162146,
0.0271179732,
0.0214522891,
0.0118417861,
0.0014026567,
-0.0136169903,
0.0059079286,
-0.0530712456,
0.046029564,
-0.021934079,
0.0246313419,
0.0154816061,
0.0912019983,
-0.0510709211,
-0.0427444577,
0.0081170835,
-0.0196141414,
-0.0185641311,
0.0825716928,
0.0079680607,
-0.0462310612,
0.023948174,
-0.0015322057,
-0.0258519314,
0.0238334499,
-0.0186001267,
0.0492676087,
0.0386848859,
-0.0367934294,
-0.0039284034,
-0.0797115713,
-0.0032754596,
0.0199848767,
0.0326016285,
-0.0709850937,
0.0031691855,
0.0040302989,
-0.0364650302,
0.0276740585,
-0.0092098471,
-0.0031683813,
0.0250510648,
0.0241955649,
0.0131873721,
-0.0385060646,
0.0026785345,
-0.0501981005,
0.0495944917,
-0.0588887557,
0.0479023382,
-0.0065011759,
0.0265839081,
0.0266317762,
0.0036685076,
-0.0590522736,
0.0312930346,
-0.0042135906,
-0.0192484185,
0.0595594794,
0.0006726895,
0.0380941369,
-0.0113777807,
0.0314404443,
0.0142771732,
0.0507326573,
0.0443400368,
-0.0327597857,
0.0104779135,
-0.0358535461,
0.0173245389,
0.025230607,
-0.0078731086,
-0.0042502605,
-0.0036555976,
-0.0230008662,
-0.05627653,
-0.0237734858,
-0.0196442623,
-0.0210825223,
-0.0754717663,
0.0677768365,
0.0533941202,
-0.0514528342,
-0.0895233378,
-0.0640757531,
-0.0064869476,
-0.0440533198,
0.0335132442,
-0.0004427295,
0.0318586379,
-0.009005486,
0.004121094,
-0.0028069725,
-0.0076884963,
-0.0347438864,
0.0095257657,
-0.009494354,
0.0984662399,
0.0332031287,
-0.005380556,
-0.042348329,
-0.0106899142,
-0.0528841577,
-0.0587391742,
0.0100534474,
0.0126493182,
-0.004487785,
-0.0046989615,
-0.0266793296,
-0.0316573828,
0.0181165766,
0.0067767389,
0.0156547185,
-0.0298615843,
-0.0213825554,
-0.0114899781,
0.0068231369,
-0.0249271709,
0.0024078896,
-0.0097287837,
-0.0641874298,
0.0546641313,
-0.0362516306,
-0.0393510051,
-0.0190218668,
0.0078182053,
0.0366324149,
-0.0028227232,
-0.0157372057,
0.0231992248,
-0.1093816757,
-0.0325828418,
0.050507497,
-0.0024813751,
-0.0098552341,
-0.0306550711,
-0.0290601682,
0.0749454573,
-0.0128072724,
-0.0135530001,
0.0087309545,
0.0192482099,
-0.047374282,
0.0146406218,
-0.0227138139,
-5.294252069e-33,
-0.0127057778,
-0.0002421624,
-0.033703886,
-0.027762413,
-0.0241713468,
-0.0437245667,
-0.0124131041,
0.0349128619,
0.0031504943,
0.0322601572,
-0.0135766705,
0.0404444486,
-0.0079302844,
-0.0242693909,
0.0206944775,
-0.0059008454,
-0.0042312378,
-0.002528161,
-0.0132609773,
-0.0253287163,
-0.0221899133,
0.0238507949,
0.0381829292,
-0.041185271,
0.0549539626,
0.0186152961,
0.0103953583,
0.0045159627,
-0.0111135533,
-0.0210533459,
-0.0204571225,
0.0691843629,
0.0700583234,
0.0116078146,
0.0239791498,
0.0532338955,
-0.0074479734,
-0.0144690135,
0.0539548248,
-0.0377029665,
0.0460378937,
-0.0428941362,
0.0059673018,
0.0035319934,
-0.0707813278,
0.0202728342,
0.0377792716,
-0.0111842034,
0.077017732,
0.0486294813,
-0.0559036694,
-0.0436832979,
-0.0248784479,
0.009762086,
0.0003001255,
0.0140401479,
-0.02196436,
0.0391231887,
-0.0179486293,
0.0230542459,
-0.0339550115,
-0.0666645169,
0.0366619043,
-0.0249298736,
0.0043146852,
0.0916119441,
0.0652032569,
0.0248490199,
-0.0553487241,
0.0134340348,
0.0542272627,
0.0160567686,
0.0163065065,
0.0515002757,
-0.0695922673,
-0.0719035864,
-0.0227764212,
-0.0180162154,
0.0578269735,
-0.0247617364,
0.0668549165,
-0.0152038764,
-0.064100191,
-0.0384765193,
-0.0386360064,
-0.0720191002,
-0.0214355234,
0.0226188581,
-0.0015536867,
-0.0196795166,
-0.0146514438,
-0.0414879508,
0.0237616394,
-0.0264979769,
0.0117581254,
-0.0243110191,
-0.0305839777,
0.0233668815,
-0.0065227994,
0.0196089763,
-0.030274475,
-0.0634705722,
-0.0249270387,
-0.004399824,
0.0759852752,
-0.0048955134,
0.0107929418,
0.0161132533,
-0.031572327,
-0.0154570378,
-0.0211862214,
0.02775806,
0.0271232296,
-0.0237313416,
0.0105273388,
0.023464499,
0.0062348698,
-0.0415171608,
0.0270178858,
-0.0130039938,
-0.0582963824,
-0.0271792375,
-0.0464284159,
0.0061938125,
-0.0427898802,
-0.0158201102,
0.0155820316,
0.0060741152,
0.0369300656,
-0.0291263051,
-0.0209808256,
-0.0237588808,
2.537e-7,
-0.0083718151,
0.0681897774,
-0.0177593324,
-0.0407345667,
0.0025795293,
0.0676481575,
0.0055629415,
0.0563571267,
0.0592595115,
-0.0146515341,
0.0189431794,
0.0220419858,
0.0286355615,
0.0200742893,
0.0001005053,
0.0573817752,
0.0145046767,
-0.0061662607,
-0.0236459337,
0.0599887297,
0.0368192941,
0.0387446135,
0.0546536036,
-0.0187199432,
-0.0054142433,
0.0053541674,
0.0127153005,
-0.0714062527,
-0.008946023,
0.0132030956,
0.0126337698,
0.0220070407,
0.0154931955,
0.0437196828,
0.0523778684,
0.0675455481,
-0.0267133527,
0.0543364845,
-0.0383784026,
-0.0219702031,
-0.0099174222,
0.0052810907,
-0.0163847953,
-0.0374216922,
0.0185213052,
-0.0709642991,
-0.0664011464,
-0.0025658715,
0.0495487303,
-0.0039976542,
0.0145449415,
0.0461307578,
-0.0117484154,
0.0163919851,
0.0377365053,
-0.0091633135,
0.0350164548,
-0.075371027,
0.013834876,
-0.0125106471,
-0.029350847,
-0.0384587161,
0.0107898517,
0.0337855853,
0.0486574546,
-0.0649654865,
0.036455784,
1.957889386e-34,
-0.0092695598,
0.0312508382,
0.0023067368,
-0.0179431457,
0.0437703878,
0.0344762392,
0.0039760028,
0.0399273336,
0.0435907654,
-0.0634589568,
-0.0005483768
] |
http://arxiv.org/abs/2402.01714v1 | http://arxiv.org/pdf/2402.01714v1 | TrICy: Trigger-guided Data-to-text Generation with Intent aware Attention-Copy | 2024-01-25T20:17:06 | 2024-01-25T20:17:06 | Vibhav Agarwal, Sourav Ghosh, Harichandana BSS, Himanshu Arora, Barath Raj Kandur Raja | cs.CL | cs.CL, cs.AI, cs.LG | Data-to-text (D2T) generation is a crucial task in many natural language
understanding (NLU) applications and forms the foundation of task-oriented
dialog systems. In the context of conversational AI solutions that can work
directly with local data on the user's device, architectures utilizing large
pre-trained language models (PLMs) are impractical for on-device deployment due
to a high memory footprint. To this end, we propose TrICy, a novel lightweight
framework for an enhanced D2T task that generates text sequences based on the
intent in context and may further be guided by user-provided triggers. We
leverage an attention-copy mechanism to predict out-of-vocabulary (OOV) words
accurately. Performance analyses on E2E NLG dataset (BLEU: 66.43%, ROUGE-L:
70.14%), WebNLG dataset (BLEU: Seen 64.08%, Unseen 52.35%), and our Custom
dataset related to text messaging applications, showcase our architecture's
effectiveness. Moreover, we show that by leveraging an optional trigger input,
data-to-text generation quality increases significantly and achieves the new
SOTA score of 69.29% BLEU for E2E NLG. Furthermore, our analyses show that
TrICy achieves at least 24% and 3% improvement in BLEU and METEOR respectively
over LLMs like GPT-3, ChatGPT, and Llama 2. We also demonstrate that in some
scenarios, performance improvement due to triggers is observed even when they
are absent in training. | TrICy: Trigger-guided Data-to-text Generation with Intent aware Attention-Copy: Data-to-text (D2T) generation is a crucial task in many natural language
understanding (NLU) applications and forms the foundation of task-oriented
dialog systems. In the context of conversational AI solutions that can work
directly with local data on the user's device, architectures utilizing large
pre-trained language models (PLMs) are impractical for on-device deployment due
to a high memory footprint. To this end, we propose TrICy, a novel lightweight
framework for an enhanced D2T task that generates text sequences based on the
intent in context and may further be guided by user-provided triggers. We
leverage an attention-copy mechanism to predict out-of-vocabulary (OOV) words
accurately. Performance analyses on E2E NLG dataset (BLEU: 66.43%, ROUGE-L:
70.14%), WebNLG dataset (BLEU: Seen 64.08%, Unseen 52.35%), and our Custom
dataset related to text messaging applications, showcase our architecture's
effectiveness. Moreover, we show that by leveraging an optional trigger input,
data-to-text generation quality increases significantly and achieves the new
SOTA score of 69.29% BLEU for E2E NLG. Furthermore, our analyses show that
TrICy achieves at least 24% and 3% improvement in BLEU and METEOR respectively
over LLMs like GPT-3, ChatGPT, and Llama 2. We also demonstrate that in some
scenarios, performance improvement due to triggers is observed even when they
are absent in training. | [
0.0438346863,
0.0014559096,
0.0110772541,
0.0119805345,
-0.0165159535,
-0.0201891847,
0.0263621584,
-0.0174310319,
0.0300743934,
-0.060380999,
-0.0276704784,
-0.0356974229,
-0.0027307882,
0.063237004,
0.0166985467,
-0.0162470173,
0.0349784791,
-0.0244474187,
-0.0090929149,
0.0219501443,
0.0271000657,
0.0288450737,
0.012356977,
0.0511816032,
0.0067827688,
-0.0155904461,
-0.0269618556,
0.0531918406,
-0.0072435392,
-0.0562335849,
0.0035329789,
0.029581273,
-0.0104558049,
-0.0067980187,
0.0000022003,
-0.0439853258,
-0.0158330742,
-0.003174318,
-0.0272249188,
-0.0222955663,
0.0494798534,
0.0102735059,
0.0019098665,
0.0464311764,
-0.0257732458,
0.033091601,
0.0336256996,
0.0302476138,
0.0498443209,
0.0683043748,
-0.0069904099,
0.0238349568,
0.0508968011,
-0.0382229723,
0.033839073,
-0.0058369893,
-0.0569755286,
-0.0248268619,
0.0163446926,
0.0220011864,
0.0317125916,
0.0427734479,
0.0324027091,
0.0195831917,
0.0107026668,
0.0357138328,
-0.0077557829,
-0.0401783697,
-0.0460304059,
0.0952297971,
-0.0289546363,
-0.0269889813,
0.0335804597,
-0.0035963869,
0.0001197223,
-0.0523908399,
-0.016436195,
-0.0077849878,
-0.0314491168,
-0.0077226348,
0.007720049,
0.0033418227,
0.0174942166,
-0.0173776429,
-0.0352123603,
0.0017167621,
0.0098898429,
-0.0069440478,
-0.0071693612,
-0.0159234907,
-0.0049893474,
-0.0792661011,
0.0474366099,
0.0453043729,
-0.0549942516,
0.0161476564,
-0.0957935303,
-0.0416422077,
-0.0083685266,
-0.0090338048,
0.0353274308,
0.0689949021,
-0.0371617563,
0.0041795303,
-0.02706201,
0.0287264101,
0.0178905465,
0.0296705533,
-0.0197482966,
-0.0051320046,
-0.0210182313,
0.0176460091,
-0.055755306,
0.032599628,
-0.0453641862,
-0.0534058698,
-0.0219035614,
0.014380143,
0.0181475934,
-0.0198887624,
0.0199076682,
0.0125158541,
-0.0531012528,
0.0008290664,
-0.0108189518,
-0.0089122141,
-0.0164305232,
0.0237877406,
0.0608665198,
-0.0891278982,
-0.010677061,
0.0185142998,
0.0231392123,
0.0189444348,
0.002258762,
0.0507459901,
-0.0159267075,
-0.0817757398,
-0.0424546115,
-0.0352197774,
-0.0418815538,
-0.0725023523,
-0.0179503784,
-0.0678410977,
0.0158796329,
-0.0038386674,
-0.0359129049,
-0.0265291724,
0.0495148301,
0.0752064884,
-0.0158199705,
0.0452758968,
-0.0559261329,
0.0012097569,
-0.0039667226,
0.0045969649,
-0.0146954106,
0.0400699675,
0.0487280414,
0.0422529988,
0.0165896881,
0.0299026929,
0.0046199779,
0.0378230438,
0.0612959899,
-0.0060448572,
0.0290470179,
0.0163384881,
-0.0345379971,
0.0181495771,
0.0431938767,
0.0661460161,
-0.0183264371,
0.0386819541,
0.0669486076,
0.0343685485,
0.0179355275,
0.0227804221,
0.0384600721,
0.0399466045,
-0.0428870097,
-0.0288982634,
-0.0593553931,
0.0385943875,
-0.0027205045,
0.0263391789,
0.0957979485,
0.0354857519,
-0.0244601835,
-0.0595142469,
0.002846854,
-0.0234410875,
0.0035711054,
0.0203163996,
-0.0042557484,
0.0404240191,
-0.0664393529,
0.0586863235,
0.0042015198,
-0.0742260441,
-0.0263368487,
-0.0318099633,
0.0219547059,
0.0697003677,
0.0497784838,
0.0141324522,
-0.0541383922,
-0.0210383423,
0.0061658192,
0.0359432213,
-0.0108896093,
0.0273694303,
0.0265563764,
-0.0654935315,
0.0166127365,
-0.0092636086,
-0.009267834,
-0.0023946695,
0.0090325065,
0.037939176,
0.0426316336,
-0.0010940882,
-0.0225402825,
0.0111926552,
0.011098058,
-0.0329953805,
0.0304681044,
0.0039486536,
0.026427228,
0.0183391534,
0.0064740595,
0.0116202785,
0.0065602898,
-0.0404839255,
0.0233439766,
0.0138509991,
-0.0536140464,
-0.0647290275,
0.0217882786,
-0.01780314,
0.0020556529,
0.0214241836,
-0.0119728111,
-0.0646413788,
-0.0041370369,
-0.069745414,
0.0202504341,
-0.0452510379,
0.0088425688,
0.0724900886,
0.038725432,
0.0181011055,
0.0461012684,
-0.0193648916,
-0.0204703361,
0.0579496734,
0.0306098778,
0.0139198778,
-0.0242182817,
0.0173109248,
0.0016721133,
-0.0252493564,
0.0167435296,
0.0095293438,
0.0514077805,
-0.0513653532,
-0.0117868427,
-0.0925549492,
0.0401859619,
-0.035203781,
0.0274070762,
0.0497609526,
0.039744433,
-0.006179505,
-0.02121922,
-0.0396420248,
0.0207617134,
0.0421263836,
-0.0346396938,
0.0220147409,
-0.0147980284,
-0.018152006,
-0.0353956185,
0.030236112,
0.00696842,
0.0029674978,
-0.0046177506,
-0.0035202417,
-0.0568064339,
0.0392985307,
-0.0422907062,
0.0123505201,
-0.0084309569,
-0.0272621103,
0.0123522589,
0.0012884979,
0.0592588335,
0.0414039977,
0.0144482842,
-0.0033014815,
0.009860388,
-0.0303733051,
-0.0464323498,
0.0263234284,
-0.002356072,
-0.0014680285,
0.0094335023,
-0.0216560252,
0.0348004773,
0.0461413264,
0.0009364525,
0.0179265514,
-0.0570363589,
-0.0269970316,
0.0128259491,
-0.0217195638,
0.0057480922,
0.0297688544,
-0.0561374091,
0.0291074142,
-0.0335918814,
-0.059268672,
0.0392818302,
-0.0444151387,
-0.0193429012,
-0.0909874886,
0.0116034476,
-0.0220870133,
0.0535881594,
-0.017059939,
0.0173575561,
0.0077944882,
-0.0317211971,
-0.0427198857,
0.0027183637,
-0.0299099367,
-0.0499820337,
0.0689205453,
-0.0364702269,
0.0050507248,
-0.0035450477,
-0.0340751037,
-0.0268991869,
-0.0248173326,
-0.0405207425,
-0.0166771002,
0.0119053293,
-0.0036597035,
-0.0096507668,
-0.0060182624,
-0.0111659588,
0.0780779272,
-0.0147026302,
-0.0147866318,
-0.0323262811,
0.0491355918,
0.0524001829,
0.0093904128,
-0.0029967921,
-0.0106115872,
0.0196966808,
0.0642595217,
0.0347069278,
-0.0335318372,
-0.0161589999,
-0.0220929515,
0.0377121009,
0.0168413017,
-0.0213865284,
0.0695043579,
0.062903665,
-0.0729731396,
-0.036564853,
0.109307982,
0.0195165817,
-0.0112843513,
0.0060472037,
-0.0142049156,
0.0619257949,
0.0036728282,
0.0278690495,
-0.1011221632,
-0.1021671817,
0.0014577492,
0.049720671,
0.0153716849,
0.0115787555,
-0.0459218323,
-0.1035972387,
0.014264713,
-0.0009773085,
-0.0159498136,
-0.0393196195,
0.0017730135,
0.0007073148,
0.044397153,
0.0020283321,
0.0089911064,
0.0082653649,
0.013906125,
-0.0045287865,
-0.0379868671,
-0.0115074003,
0.0438660607,
-0.0140682505,
0.0468829311,
0.0691068992,
0.0470063575,
0.0507010743,
-0.0050931028,
0.0207255855,
0.0027059175,
0.0204260796,
-0.0112919491,
0.0332190022,
-0.0175588876,
-0.0025110948,
-0.0274807718,
0.0777368098,
-0.0544893108,
0.004242532,
0.0126506966,
0.0225605741,
-0.0154456645,
-0.0080789607,
-0.0105213504,
0.0121195465,
0.0589116849,
0.04477809,
0.038147565,
-0.0512485132,
-0.0637328401,
-0.0077224351,
-0.0012877834,
-0.0757643655,
0.0087818522,
-0.0079996279,
0.0187364332,
-0.0392314941,
-0.079257153,
-0.0664269999,
0.0108960913,
0.03684536,
0.0072430349,
-0.0119468439,
0.0080169244,
0.0182161387,
-0.0237195455,
0.013948895,
-0.0185653735,
0.0189676974,
-0.0632588565,
0.0022747885,
0.0102509223,
-0.0438527465,
0.0135842217,
0.0905860513,
-0.1324204504,
-0.0365748666,
-0.0163261928,
0.0611362234,
0.0180789027,
0.0323090069,
-0.0274910033,
-0.0056026373,
0.0414551646,
0.0495983399,
-0.0731472149,
0.0058183232,
0.0313283354,
0.0283543244,
0.0111804325,
0.0606392883,
-0.0373667181,
0.0159605648,
-0.0107318573,
0.0350348949,
-0.0187373236,
-0.0429354012,
0.0161672998,
0.0192022882,
0.0460722968,
-0.0349332802,
-0.015263427,
-0.0119517343,
0.0516033843,
0.0165362973,
0.0192702301,
-0.0759542808,
0.0214156415,
-0.0196509454,
0.0057317419,
0.0022538179,
-0.0077417721,
-0.0203895252,
-0.0197534077,
-0.0088071125,
0.0180933345,
-0.0163387712,
0.0203862488,
0.0244603306,
-0.0246423874,
-0.0394891948,
0.028389154,
-0.0195686836,
0.0311624408,
-0.0500335544,
0.0792831257,
0.01802532,
-0.0238597728,
0.0175925046,
0.0515443869,
-0.0297730584,
-0.0061481576,
-0.021308206,
-0.008887467,
0.0302812066,
0.0108103864,
0.0702539831,
-0.031849578,
0.0078136614,
-0.0368603431,
0.0717124864,
-0.042079445,
-0.048601225,
0.0012201834,
-0.0424529463,
-0.0078555811,
-0.0239086282,
0.066120334,
-0.0302585941,
0.0062206709,
0.0290988646,
0.0125829885,
-0.0663213655,
-0.0237639602,
-0.0136253573,
0.0173929781,
-0.0019814235,
-0.0079416493,
-0.0326438732,
-0.0472790673,
0.0229845624,
-0.0044340929,
-0.0226025488,
-0.0340157673,
0.0763070062,
0.0411845855,
-0.0114523666,
-0.0107308486,
-0.0272188708,
-0.004432207,
-0.0448285528,
-0.0338883512,
-0.0056870426,
-6.211105704e-33,
-0.0373034328,
-0.0293768439,
0.0340304933,
-0.018111635,
-0.0231545214,
0.0150850685,
-0.0690931678,
-0.0157056749,
-0.0044397078,
-0.0344390087,
-0.0373877622,
0.0202400461,
0.0118705183,
0.0569170341,
-0.0380810574,
-0.0291225538,
0.0487359315,
-0.0288516898,
-0.0447975583,
0.0703087598,
0.0476127639,
0.0375337861,
-0.0047971406,
-0.0381297693,
0.0166035611,
-0.0050650612,
-0.0050434736,
-0.0420002826,
0.0089752181,
0.0209859088,
-0.0131266387,
0.0245800354,
0.0036637608,
-0.0378679559,
-0.0482138842,
0.0530925356,
-0.0599985942,
-0.0420012549,
-0.0012193273,
-0.0251401812,
-0.060653016,
-0.0355593674,
0.0266416296,
0.0009236306,
-0.0147905014,
0.0017961439,
0.0064520109,
-0.0083427504,
-0.0317467563,
-0.0370129757,
-0.056199275,
0.0175094511,
-0.0074897991,
-0.0668036863,
-0.0087533239,
0.0624962822,
-0.0235243402,
-0.0124370456,
-0.061223194,
0.0075585674,
-0.0240392424,
0.0195460524,
0.0628614426,
0.0225395318,
0.0057687098,
0.0542604923,
0.0229972806,
0.0029711591,
-0.0433301628,
0.0077037425,
-0.0176408105,
0.0445604585,
0.0092351688,
0.0102270143,
0.0423712507,
-0.0033361018,
-0.0443180725,
0.0153276324,
0.0103517883,
0.0009315712,
0.0250818655,
-0.0091127725,
-0.0363853201,
-0.0031982991,
-0.0099369986,
0.0227142461,
-0.0237129964,
-0.0373808406,
0.0800338984,
0.033323139,
-0.0044660196,
-0.0221908763,
0.0426516794,
0.0013335801,
0.0122563867,
0.0043338984,
-0.0072674821,
0.0065440922,
-0.0173533801,
0.0212733857,
-0.0894646049,
0.0231237952,
0.0245494526,
0.0360783152,
0.0242716614,
0.0235942211,
0.0063312268,
0.0161948111,
-0.0826356784,
0.0146722198,
-0.0472039767,
0.0717759877,
-0.0351682864,
0.006804816,
0.057858739,
-0.0321239606,
-0.0285604484,
0.0840796754,
-0.0293539111,
-0.0085368427,
-0.0505452678,
-0.0304560699,
0.0260153171,
0.0119607775,
-0.0463562906,
0.0132357394,
-0.0383866914,
-0.0090121944,
-0.0030141931,
-0.0316823497,
-0.0136129018,
0.0390920341,
2.885e-7,
-0.0036827326,
0.0219162814,
0.0188979898,
0.0372520201,
-0.0086542917,
0.0378904082,
0.0395793356,
0.0397446603,
-0.0010938296,
-0.0021771749,
-0.004706006,
0.0091560418,
0.058851853,
0.0047268211,
-0.0854470581,
0.0274980925,
-0.0271100551,
0.0436086543,
-0.0411600396,
-0.011826396,
0.0620855354,
0.0840729699,
-0.0228819046,
-0.0108928084,
0.0349546112,
-0.026055336,
0.0293474942,
-0.0104650455,
0.0121968212,
0.0013637436,
0.073654227,
0.0139560653,
-0.0570924878,
0.0058302181,
-0.0243211258,
-0.005904106,
-0.0050736745,
0.045037847,
0.0169025287,
-0.0491614491,
-0.0031880632,
-0.0174289793,
-0.0009250852,
-0.0485059172,
0.0769225135,
-0.011599795,
-0.0142939882,
-0.015352889,
0.0100365803,
-0.0099091437,
-0.0311457533,
-0.001509231,
-0.0018833023,
0.0060637677,
0.0168381445,
-0.0125834374,
-0.0009528286,
-0.0787412897,
-0.0317946039,
-0.0078142313,
-0.0114318933,
-0.0337052755,
-0.1215013415,
0.0296109729,
0.0324411914,
-0.0456090383,
-0.0148188574,
2.66174349e-34,
0.0032644593,
-0.0387475677,
0.0261507202,
0.0060923016,
0.0270195995,
-0.0063153272,
0.0691350475,
-0.0060663391,
0.0678353459,
-0.0517824292,
-0.032273639
] |
http://arxiv.org/abs/2402.01713v2 | http://arxiv.org/pdf/2402.01713v2 | Prompting Large Language Models for Zero-Shot Clinical Prediction with Structured Longitudinal Electronic Health Record Data | 2024-01-25T20:14:50 | 2024-02-10T16:31:40 | Yinghao Zhu, Zixiang Wang, Junyi Gao, Yuning Tong, Jingkun An, Weibin Liao, Ewen M. Harrison, Liantao Ma, Chengwei Pan | cs.CL | cs.CL, cs.AI, cs.LG | The inherent complexity of structured longitudinal Electronic Health Records
(EHR) data poses a significant challenge when integrated with Large Language
Models (LLMs), which are traditionally tailored for natural language
processing. Motivated by the urgent need for swift decision-making during new
disease outbreaks, where traditional predictive models often fail due to a lack
of historical data, this research investigates the adaptability of LLMs, like
GPT-4, to EHR data. We particularly focus on their zero-shot capabilities,
which enable them to make predictions in scenarios in which they haven't been
explicitly trained. In response to the longitudinal, sparse, and
knowledge-infused nature of EHR data, our prompting approach involves taking
into account specific EHR characteristics such as units and reference ranges,
and employing an in-context learning strategy that aligns with clinical
contexts. Our comprehensive experiments on the MIMIC-IV and TJH datasets
demonstrate that with our elaborately designed prompting framework, LLMs can
improve prediction performance in key tasks such as mortality, length-of-stay,
and 30-day readmission by about 35\%, surpassing ML models in few-shot
settings. Our research underscores the potential of LLMs in enhancing clinical
decision-making, especially in urgent healthcare situations like the outbreak
of emerging diseases with no labeled data. The code is publicly available at
https://github.com/yhzhu99/llm4healthcare for reproducibility. | Prompting Large Language Models for Zero-Shot Clinical Prediction with Structured Longitudinal Electronic Health Record Data: The inherent complexity of structured longitudinal Electronic Health Records
(EHR) data poses a significant challenge when integrated with Large Language
Models (LLMs), which are traditionally tailored for natural language
processing. Motivated by the urgent need for swift decision-making during new
disease outbreaks, where traditional predictive models often fail due to a lack
of historical data, this research investigates the adaptability of LLMs, like
GPT-4, to EHR data. We particularly focus on their zero-shot capabilities,
which enable them to make predictions in scenarios in which they haven't been
explicitly trained. In response to the longitudinal, sparse, and
knowledge-infused nature of EHR data, our prompting approach involves taking
into account specific EHR characteristics such as units and reference ranges,
and employing an in-context learning strategy that aligns with clinical
contexts. Our comprehensive experiments on the MIMIC-IV and TJH datasets
demonstrate that with our elaborately designed prompting framework, LLMs can
improve prediction performance in key tasks such as mortality, length-of-stay,
and 30-day readmission by about 35\%, surpassing ML models in few-shot
settings. Our research underscores the potential of LLMs in enhancing clinical
decision-making, especially in urgent healthcare situations like the outbreak
of emerging diseases with no labeled data. The code is publicly available at
https://github.com/yhzhu99/llm4healthcare for reproducibility. | [
0.0362618342,
0.0997479856,
-0.0094627328,
-0.0303263329,
-0.0228705537,
0.0276539866,
0.0069926414,
0.0187481567,
0.0395647474,
-0.0143363578,
0.0092287259,
-0.0861093104,
-0.0005072076,
0.0544272996,
-0.0162689351,
-0.0401667021,
0.0390145555,
-0.0128874481,
0.0393969081,
0.0277445465,
-0.0348574333,
0.0097512212,
-0.0097731007,
0.0462354235,
0.02719431,
0.0057402872,
-0.0082129193,
0.0015455909,
0.0036924917,
-0.1061704978,
0.0139838904,
0.036720939,
0.0231507141,
0.0070425631,
0.0000019617,
-0.0467308834,
-0.0272425413,
0.0094363829,
-0.0709720254,
-0.0690439865,
0.0378871635,
-0.0958274305,
0.0154684046,
0.0182382297,
-0.051175084,
-0.0604324788,
0.0244708043,
0.033459492,
0.027417725,
0.0653187484,
-0.0234400705,
0.0073552546,
0.0162559133,
-0.044404231,
0.0416555777,
-0.0335329324,
0.0418632329,
-0.0511025041,
0.0436437763,
-0.01859124,
-0.0074465503,
0.0426086299,
0.000687721,
0.0370967314,
-0.0019315542,
0.0725980774,
-0.0225938801,
-0.0022121484,
-0.0328640305,
0.0587369949,
-0.0039487928,
0.0060145371,
0.0374926701,
-0.0063507026,
-0.0261178277,
-0.0068066446,
-0.0430099107,
-0.020642003,
0.0072469302,
0.01847017,
0.0077649844,
0.0186164659,
0.0540444478,
-0.0210931171,
-0.0158388838,
0.0239292905,
-0.0016016498,
-0.0401283428,
0.0001783608,
-0.015851533,
0.0129630016,
-0.0481914356,
-0.0266239364,
0.0481808856,
-0.0742515177,
0.0274039134,
-0.0017633263,
0.015454581,
0.0571620874,
-0.0365028381,
0.0247593112,
0.0174101777,
0.0158573557,
0.0469194762,
0.0125782471,
-0.0236268621,
-0.0087041305,
0.0000110896,
-0.0111462269,
-0.002685219,
-0.0626815334,
-0.0003224321,
-0.0140327066,
-0.0076947166,
0.0247212593,
0.0267593171,
-0.0764757469,
-0.0110426722,
0.005669564,
0.0168282539,
0.0210335255,
0.0114687653,
-0.032087639,
-0.0134331426,
-0.0330178551,
0.0239251591,
0.0229314733,
0.0043879701,
-0.0087438421,
-0.0597220212,
0.0026998187,
0.0002232114,
0.0088955229,
-0.0336845554,
0.0195146669,
0.0533814393,
0.0484048314,
-0.0061008078,
-0.0397725254,
-0.021461919,
-0.0357436165,
-0.03872253,
-0.0183981378,
-0.0628227443,
0.0129478229,
0.0302361026,
0.0174749605,
-0.0429124124,
0.0095499037,
-0.0081675034,
-0.0726275593,
0.0616515167,
-0.0412523784,
-0.0017721757,
-0.0170053784,
0.0072290683,
0.0135662295,
0.0434007756,
0.0106132934,
-0.0452284403,
-0.0076688863,
-0.0245924853,
0.0190299675,
0.0286940746,
0.0382658355,
-0.0065433052,
-0.0209656488,
0.0044115069,
0.012885781,
0.0430925265,
-0.0445360504,
0.0221992787,
0.0182523653,
-0.0095374566,
0.0394752398,
0.0477883555,
0.0203341655,
-0.00995625,
0.0477325991,
-0.0301960018,
-0.044474408,
-0.0522835888,
-0.0122539559,
-0.0079049841,
0.013111569,
-0.0040577459,
-0.001582123,
-0.0063774213,
-0.0328691974,
-0.0410461091,
-0.0397657342,
-0.015692709,
-0.0111672347,
-0.013679754,
0.0458798818,
-0.0259781014,
-0.0395318456,
0.0105630169,
0.0460556,
0.0134803709,
-0.0242381953,
-0.0548841767,
0.0023222899,
0.0984578431,
0.0347009189,
-0.0166577473,
-0.0519199185,
-0.0488637611,
0.0441478193,
-0.0223701019,
0.0048874849,
-0.0058929785,
0.0246407147,
-0.0233306084,
-0.0076482003,
0.0541586988,
0.0172866378,
-0.0363022909,
0.0076690721,
0.0310121868,
-0.0082461648,
-0.028135268,
-0.0369770341,
0.02751562,
-0.077462092,
-0.0434934571,
0.023743771,
-0.0011467128,
0.0585586019,
-0.0160179529,
0.0103415251,
0.0175323356,
0.0261731669,
-0.0524321459,
-0.026658088,
0.0300453231,
-0.0323681943,
-0.0332102664,
0.0108021731,
0.0277258791,
0.004694676,
-0.0499679744,
-0.0402296521,
-0.0252201725,
-0.0037467116,
-0.0088324863,
0.0402870104,
0.0072975601,
0.004125461,
0.044147037,
-0.018979663,
-0.0399322584,
0.0396173894,
-0.0216944423,
0.0148119498,
0.0270407125,
-0.0400375649,
-0.0041947369,
-0.0052257706,
-0.012562043,
0.0063875252,
0.0061495253,
-0.0040813363,
-0.0227477159,
0.0024591906,
-0.0247676466,
-0.001184882,
-0.1211772561,
0.0159024205,
-0.0051121423,
0.0176060852,
0.0103754094,
0.0399095602,
-0.0311172232,
-0.0319748037,
-0.0327606574,
-0.0136302412,
0.0327725038,
0.0251629669,
0.0443126224,
0.0540963449,
-0.0128125139,
-0.0110935476,
-0.0032381765,
-0.0654769689,
0.0263894163,
0.0010534457,
-0.0271646921,
-0.0025064251,
-0.0365831703,
-0.0242183749,
-0.0170534421,
-0.0187682994,
0.0092641721,
-0.0073481598,
-0.0446800366,
0.0341876224,
0.0177259855,
-0.030098401,
0.0500056334,
0.0275372118,
-0.0001253467,
-0.0431941748,
0.0231738389,
-0.0627376661,
-0.0349968337,
0.0025820462,
-0.0237185489,
-0.0030107209,
0.0875341669,
0.0252106804,
-0.0039429446,
-0.0495371409,
0.0145529592,
-0.0388136208,
0.0389025882,
-0.0554549098,
0.0523498543,
-0.0333678909,
0.0390673429,
0.0001629789,
-0.0682205036,
0.0668561682,
0.0215953607,
0.0374415256,
-0.0410035923,
-0.0421582982,
-0.0618992448,
-0.0085815042,
-0.0027998935,
-0.0135030663,
0.0295845289,
-0.0030993067,
-0.0005155329,
0.0148526262,
-0.029830914,
-0.0428690985,
0.0448620394,
-0.0211050101,
0.0479998663,
0.0242867582,
-0.0448147096,
-0.0484268479,
-0.0343060941,
-0.0491617508,
-0.0021124459,
-0.0433037579,
0.0237579979,
-0.060980279,
0.0229495764,
-0.0213154927,
0.0788334981,
0.0181429777,
-0.0353019834,
0.0180123895,
0.0043904637,
0.0293605886,
-0.0245320573,
-0.0141584333,
-0.0374813713,
0.0647029504,
0.0362681821,
0.0252089947,
0.038438607,
-0.0164571814,
0.043936912,
0.0068669757,
-0.0172185376,
-0.0183172077,
0.0724272206,
0.0253816508,
-0.0715112314,
-0.0007637343,
0.1185816675,
0.0351387151,
-0.0577960201,
0.0167971198,
-0.0498827174,
0.0246021822,
-0.008313803,
0.0079754982,
-0.0541302189,
-0.0420773998,
-0.004490179,
0.0189352799,
-0.0171735492,
-0.06610322,
0.0287975632,
-0.0502929948,
0.0343351178,
0.0146584744,
-0.0377070755,
-0.0153229609,
-0.0614051633,
0.0252084881,
-0.0356136523,
-0.0237483904,
-0.0226512644,
-0.0276283808,
-0.0071754321,
0.0038766903,
-0.0702434406,
-0.0901482478,
0.0375971012,
-0.0102137178,
0.0901910141,
0.0573076531,
0.0281905513,
-0.0113281133,
0.0032391981,
0.0044813207,
0.0072406186,
0.0101639414,
-0.0338501967,
0.0707171187,
0.0189324822,
-0.0058596041,
-0.0561434291,
0.0329835564,
-0.0393880121,
-0.0264847372,
0.0129133286,
0.136845842,
-0.0384673886,
-0.0369205549,
-0.0406915173,
0.0100974571,
0.0503390282,
0.0522078648,
-0.0089627746,
-0.0282813534,
-0.0262472462,
0.0143442871,
0.0280847214,
0.0061522997,
-0.0297839362,
0.0375071838,
-0.0356763266,
-0.0226911809,
-0.0405764878,
-0.0090011535,
0.0703005567,
-0.0090034697,
0.0176342092,
0.0453076996,
0.0181088448,
-0.0058932919,
-0.05065668,
0.0352655053,
0.0023372245,
0.023038948,
-0.0874998197,
-0.0350637324,
0.0450123549,
0.0354391597,
0.004490234,
0.0061353166,
-0.032216467,
-0.0199214816,
-0.0009250595,
0.0421174541,
0.0241718628,
0.0334727317,
0.0227433387,
-0.0003188479,
0.0643817112,
0.0411639698,
-0.0125418343,
-0.0463722795,
-0.0086297346,
0.0292086489,
-0.00036363,
0.0413859412,
0.0059728911,
0.017822491,
0.0023492808,
0.0291940179,
-0.0343013704,
0.01019389,
0.0049856631,
-0.0012757177,
0.0413599014,
-0.0242406614,
-0.0382062607,
-0.0219568219,
0.007941504,
-0.0035668022,
0.0249633845,
-0.0231475476,
0.0048849154,
0.0314472392,
0.0011071354,
-0.0310308821,
-0.002929392,
-0.0559352562,
-0.0234046653,
-0.0758791715,
0.0552894138,
0.0175075028,
-0.0030160984,
0.0805302784,
0.0057524494,
0.0283543523,
0.0454622991,
0.0057730325,
0.0276424978,
-0.0044467086,
0.0482780896,
-0.0025266719,
0.0571654886,
-0.0039658137,
0.0276213717,
-0.0367582142,
-0.0383508168,
-0.0068796547,
0.0036671839,
0.0230343807,
-0.0158746261,
0.0131050963,
-0.0260144789,
0.0141470619,
-0.0200435575,
0.0295748804,
-0.0102566415,
0.0027487276,
0.0581242889,
0.0199746899,
-0.0448322073,
0.0192015935,
0.1079817861,
0.0453133099,
0.0709645003,
0.0315317102,
-0.0068521709,
-0.0332324393,
0.0202077255,
0.0639124885,
-0.009056448,
-0.0523502864,
0.0314318538,
-0.0710306391,
-0.0558983199,
0.0377591699,
-0.0466733836,
-0.0571273975,
-0.0175803937,
0.0321527347,
0.006767021,
-0.0015185055,
0.0067950906,
-0.0390121117,
0.0372604169,
0.0055544954,
-0.0533298329,
-0.0140367225,
-5.838573177e-33,
0.0265039392,
-0.0214903019,
0.0211744308,
-0.0375937372,
0.006596284,
0.0345043987,
-0.0372855812,
-0.0041679591,
0.0156300701,
0.0063133435,
-0.0245588552,
-0.0024162589,
0.0073867887,
0.0289127901,
0.043191202,
-0.0620665327,
0.0214927364,
0.0055977604,
-0.0148872379,
0.0249200258,
-0.005696353,
0.061356429,
0.0331328921,
-0.0565250292,
0.0378198661,
0.0035788508,
0.024421256,
0.0289334189,
0.0020292657,
0.0488753691,
-0.0357870981,
0.0499437265,
0.0520724505,
-0.0435885899,
-0.009987901,
0.0143436929,
-0.0469336361,
-0.0242443904,
0.0119654154,
-0.0451015234,
-0.0187832937,
-0.0360472426,
0.0283226091,
-0.0021980295,
-0.0132708987,
-0.0082385456,
0.002986165,
-0.036625918,
-0.0210258756,
-0.0197293274,
-0.0151744056,
0.0589596704,
-0.0148227522,
-0.0032031038,
0.0011569255,
0.029889917,
-0.0393850319,
0.0715009943,
-0.0573132597,
-0.0037529978,
0.0102547091,
0.035032142,
0.0013465584,
-0.0244030561,
-0.009594569,
0.0700123906,
0.0098355757,
-0.0069199335,
-0.0566839501,
0.0265323166,
-0.0387575552,
0.0184942819,
0.0189511105,
0.0014297531,
0.060830418,
-0.020582933,
-0.0977337882,
0.0415350124,
-0.0533009507,
0.0317787826,
0.033992473,
-0.0295193922,
-0.0349343866,
-0.018853955,
-0.0199634414,
0.0007337642,
-0.0255373102,
-0.0231374837,
0.0318840779,
0.0168495756,
0.0342733264,
0.0193182006,
0.0318463482,
-0.0104606505,
-0.0285388287,
0.0028780543,
0.0531536005,
-0.0001672029,
-0.0103394557,
0.0258061327,
-0.0486463048,
-0.0033318484,
0.02177139,
0.0133991316,
0.0470546633,
0.0024854578,
0.0386636667,
0.0032085045,
-0.0689682439,
-0.0401035063,
0.0233138744,
0.0297743846,
-0.0288141407,
0.0244168397,
0.0411469266,
0.0225666836,
0.0149103943,
0.0712431893,
0.0116929775,
-0.0199152585,
-0.0143929664,
-0.0065437132,
-0.0184928142,
-0.0282308832,
-0.035138879,
-0.0245413985,
-0.050358966,
0.0321447439,
0.0385625586,
-0.0469408743,
-0.0035558229,
-0.0018046084,
2.717e-7,
0.0807494745,
0.0636769012,
-0.0024447206,
-0.0274947956,
0.0333328545,
0.0083820885,
-0.0265391525,
-0.004250417,
0.0031571314,
-0.000765624,
0.0593765192,
-0.0101372609,
0.0517357253,
0.0416815355,
-0.0804664269,
-0.0206200238,
-0.0942940488,
-0.013003544,
-0.0657193735,
-0.0188306905,
0.0697190315,
0.0974837169,
-0.0206752028,
-0.0292586423,
0.0233498216,
0.0112683484,
0.0246572271,
-0.0508195087,
-0.0183466394,
0.0201449208,
0.0312442575,
0.009512539,
0.0098542953,
0.0324925482,
-0.0332380012,
0.0082056755,
-0.041733291,
0.1065105498,
-0.0304249041,
0.0217158366,
0.0071506137,
-0.0096193459,
-0.0008178568,
-0.0288247708,
0.109008804,
-0.0825092867,
-0.061174646,
0.0139519097,
-0.0111837396,
0.0090560429,
0.0583521239,
-0.0065942868,
-0.0513535216,
0.0355173461,
0.0292865075,
0.0091085322,
-0.0243384093,
-0.0382254608,
-0.0326598324,
0.027832808,
-0.0265745372,
-0.0367269218,
0.0189631917,
-0.0036291436,
0.000449357,
-0.0871104971,
0.0480195284,
2.674452605e-34,
0.0535750687,
0.0063501284,
-0.0016644303,
0.0213135891,
0.0191207677,
-0.013356735,
0.0144756678,
0.0242048316,
0.0383245088,
-0.0752920955,
-0.0133434879
] |
http://arxiv.org/abs/2401.14493v1 | http://arxiv.org/pdf/2401.14493v1 | K-QA: A Real-World Medical Q&A Benchmark | 2024-01-25T20:11:04 | 2024-01-25T20:11:04 | Itay Manes, Naama Ronn, David Cohen, Ran Ilan Ber, Zehavi Horowitz-Kugler, Gabriel Stanovsky | cs.CL | cs.CL, cs.HC, cs.LG | Ensuring the accuracy of responses provided by large language models (LLMs)
is crucial, particularly in clinical settings where incorrect information may
directly impact patient health. To address this challenge, we construct K-QA, a
dataset containing 1,212 patient questions originating from real-world
conversations held on K Health (an AI-driven clinical platform). We employ a
panel of in-house physicians to answer and manually decompose a subset of K-QA
into self-contained statements. Additionally, we formulate two NLI-based
evaluation metrics approximating recall and precision: (1) comprehensiveness,
measuring the percentage of essential clinical information in the generated
answer and (2) hallucination rate, measuring the number of statements from the
physician-curated response contradicted by the LLM answer. Finally, we use K-QA
along with these metrics to evaluate several state-of-the-art models, as well
as the effect of in-context learning and medically-oriented augmented retrieval
schemes developed by the authors. Our findings indicate that in-context
learning improves the comprehensiveness of the models, and augmented retrieval
is effective in reducing hallucinations. We make K-QA available to to the
community to spur research into medically accurate NLP applications. | K-QA: A Real-World Medical Q&A Benchmark: Ensuring the accuracy of responses provided by large language models (LLMs)
is crucial, particularly in clinical settings where incorrect information may
directly impact patient health. To address this challenge, we construct K-QA, a
dataset containing 1,212 patient questions originating from real-world
conversations held on K Health (an AI-driven clinical platform). We employ a
panel of in-house physicians to answer and manually decompose a subset of K-QA
into self-contained statements. Additionally, we formulate two NLI-based
evaluation metrics approximating recall and precision: (1) comprehensiveness,
measuring the percentage of essential clinical information in the generated
answer and (2) hallucination rate, measuring the number of statements from the
physician-curated response contradicted by the LLM answer. Finally, we use K-QA
along with these metrics to evaluate several state-of-the-art models, as well
as the effect of in-context learning and medically-oriented augmented retrieval
schemes developed by the authors. Our findings indicate that in-context
learning improves the comprehensiveness of the models, and augmented retrieval
is effective in reducing hallucinations. We make K-QA available to to the
community to spur research into medically accurate NLP applications. | [
0.0569497608,
0.0495808534,
-0.0033051271,
-0.037157163,
-0.0263567641,
0.0236856714,
-0.0305575728,
0.0267722663,
0.0284809954,
-0.0178584866,
0.0107204905,
-0.0523526669,
-0.0014424891,
0.0247238465,
0.0000102789,
-0.0331967957,
0.0249846391,
0.0191077311,
0.0353475474,
0.0100469543,
-0.0135422004,
0.0015791563,
-0.0430020764,
0.0537430644,
-0.0022260807,
-0.0131264636,
0.0298645366,
-0.0149238333,
0.0141630117,
-0.0682995319,
0.0130188866,
0.0212040301,
-0.0286084022,
-0.0085773896,
0.000001975,
-0.0468094945,
-0.0174461976,
0.0160453431,
-0.0703944862,
-0.0290590245,
0.0130323153,
-0.0420187637,
0.0224447343,
0.0194385089,
-0.0517155156,
-0.0100605031,
0.0374278612,
0.0468345061,
0.0609594248,
0.0517490022,
-0.0136101386,
0.0037858086,
0.0457983799,
0.0187524948,
0.0354059152,
-0.0438928753,
0.0425604284,
0.0120022278,
0.0279852487,
-0.0170104317,
0.0377941132,
0.0491695628,
-0.003828044,
0.0447366461,
0.0296538342,
0.0285478532,
-0.0676382035,
-0.0146655207,
0.0053537032,
0.0767459795,
0.0192848425,
-0.0239393748,
0.0002823361,
0.0048336829,
-0.0029550868,
0.0228409655,
-0.0624295436,
0.0024652015,
-0.0315179639,
-0.0054073813,
-0.0032337601,
-0.0002286075,
0.0436501689,
-0.0116702514,
-0.0147740925,
0.0115610566,
0.0187087487,
-0.0429594778,
0.009016132,
-0.0164168943,
-0.032347057,
-0.0250993297,
0.0410021134,
0.0360673331,
-0.0655230731,
0.0181390382,
0.0165182501,
-0.0077184327,
0.1035277396,
-0.0799640045,
0.009881299,
0.0089736506,
-0.0214476194,
0.0257896408,
-0.0037283401,
0.019533556,
-0.0078317635,
0.046572838,
-0.0524833053,
-0.0041106399,
-0.0561334193,
-0.0376086496,
-0.0203588996,
0.0415596776,
-0.0006420766,
-0.000849821,
-0.0700312182,
-0.0251053181,
-0.0375549644,
0.0391200334,
-0.0122403083,
0.0580206402,
-0.0078251148,
-0.0052476302,
-0.0194248036,
0.0132838776,
-0.0364019498,
0.0081428383,
0.0047414843,
-0.0614028983,
0.0046052639,
-0.0226553418,
0.0140154893,
0.0114185568,
0.0273310039,
0.0268124137,
0.0223855563,
-0.0223439839,
0.014159508,
-0.0451662876,
-0.0543563291,
-0.0224612392,
-0.0117252637,
-0.0543321855,
0.0396630503,
0.0191974901,
-0.0094248941,
-0.0588088483,
0.0285890959,
0.0426941141,
-0.0387572944,
0.0529057123,
-0.0244547036,
0.0348993577,
0.0439198241,
0.0434454605,
0.0074899257,
0.0315679908,
0.0049291705,
-0.0380471237,
0.0283812806,
-0.0294873007,
0.0416768417,
-0.0361906625,
0.0504758358,
-0.0361251682,
0.0004773789,
0.0497909673,
0.0184229165,
0.0646209195,
-0.0850397944,
0.0779446363,
0.0054391129,
-0.0243183691,
0.0242548734,
0.0640259236,
0.0049237688,
0.0027214226,
-0.023265183,
-0.0431740694,
0.0037873783,
-0.0057037743,
-0.0081647644,
0.0064108749,
-0.0244320761,
-0.0035765441,
0.0265464243,
-0.0208549798,
-0.0250057038,
-0.0372497477,
-0.0492075197,
-0.0016260137,
-0.0094479313,
0.02912925,
0.014787171,
-0.0007280369,
-0.0307608228,
0.0828228593,
0.0337953605,
-0.0418719165,
-0.0176080987,
-0.0403954685,
-0.0084606046,
0.073514156,
0.0531298853,
-0.0624863021,
-0.0275305305,
-0.0192882456,
0.011750116,
0.0013409933,
0.0275784973,
-0.0365561545,
0.0133363549,
-0.0209558122,
0.0263933949,
0.0078978529,
0.0309194755,
-0.0571107529,
0.0108359689,
0.017563818,
-0.0139109828,
-0.0342988931,
-0.0404659621,
0.0465135798,
-0.0556455068,
-0.0460602641,
0.0109774945,
-0.0016674049,
0.007956408,
0.0253666118,
0.0023058325,
0.0596103594,
0.0513616391,
-0.0439011231,
-0.0355746448,
-0.0193501115,
-0.0445488952,
-0.019569708,
0.0223329924,
0.0083225174,
-0.0000300991,
-0.0217417944,
-0.0316010118,
-0.0200678278,
-0.0165899917,
-0.0547849759,
0.0367045142,
0.0006267728,
-0.0143875517,
0.0354944207,
-0.0034892613,
-0.0635240749,
0.0384883806,
-0.0303777624,
0.0326889418,
0.0167398266,
-0.0269767437,
-0.0108969342,
-0.0356469527,
-0.0055733477,
0.0167131815,
-0.0228407308,
0.0168516133,
0.0011922331,
-0.0458445214,
-0.0482656211,
-0.0086868089,
-0.122950919,
-0.0036700119,
-0.0144400597,
-0.0198057499,
0.019589996,
0.0213464256,
0.0144986408,
-0.0153151127,
-0.054008916,
-0.0318499729,
0.0461660661,
0.0421270244,
0.0142710814,
0.0742026046,
0.0230367146,
-0.0431545116,
-0.0173235387,
-0.0360334851,
0.0411194488,
0.0023872671,
-0.0100783845,
-0.0122441407,
-0.0121509554,
-0.0579929128,
-0.0113396132,
0.0038134886,
0.0055983402,
0.027075978,
-0.0217486266,
-0.0131033044,
0.0588192455,
0.0121722771,
0.0010437373,
0.0246685687,
-0.0066748145,
-0.039862901,
0.0157821402,
-0.0369411893,
-0.0578016192,
0.0272012893,
-0.0556779392,
0.0309581347,
0.0541081578,
0.0390093811,
0.0051381965,
-0.0572642162,
0.0155102676,
-0.0292886756,
0.0177127849,
-0.0627000034,
-0.005341616,
-0.0313575752,
0.0822650716,
-0.0036439018,
-0.0098034358,
0.0950915441,
0.0104784397,
0.045227319,
-0.0591312684,
-0.0064018453,
-0.0261796582,
0.0153278392,
-0.0179045461,
0.0382279493,
-0.001170701,
0.006527469,
-0.0005718893,
-0.0022744064,
-0.0188987348,
-0.0870490968,
0.0736701936,
-0.0360143706,
0.0372699015,
0.02542172,
-0.0479809679,
-0.0743778348,
-0.019857958,
-0.0232779086,
-0.0166923068,
-0.0300033502,
0.0375993177,
-0.0580687262,
-0.0056933286,
-0.0016875832,
0.0312350579,
0.0138343805,
-0.0316292793,
0.035100624,
0.0104854451,
0.0580211952,
0.0149826715,
0.0358913317,
-0.0080420235,
0.044059135,
0.0959348753,
0.0552493781,
0.0436671041,
0.0301784128,
-0.0269042663,
-0.0131611703,
-0.0245106947,
0.0119081363,
0.0423449725,
0.0205171593,
-0.0482395738,
-0.0555238873,
0.0797645971,
0.0579274036,
-0.0504208617,
0.0126438271,
-0.0448108464,
0.0041366029,
-0.0175620858,
0.0133320689,
-0.0702402517,
-0.0343481749,
-0.021023497,
0.0545056649,
-0.0302984789,
-0.0206069034,
0.0291360375,
-0.0639585108,
0.005937594,
0.0125279799,
-0.0308450945,
-0.045613382,
-0.048754897,
0.0082768379,
0.0008255764,
0.0134994555,
0.0109420298,
0.0034767785,
-0.009176515,
-0.005898302,
-0.0577178895,
-0.0528528318,
0.0553417578,
0.0061241305,
0.0523081794,
0.0709898099,
0.0399190485,
0.0008090485,
-0.0395022929,
-0.0135931978,
-0.0250177737,
-0.001986925,
0.0237425342,
0.0778324082,
0.0180442985,
0.0191256832,
-0.039069321,
0.0467205644,
-0.0592394508,
-0.0373758338,
0.0027122798,
0.0702946633,
-0.0419954136,
0.0038221055,
-0.0418330953,
0.0419753268,
0.0312928483,
0.0561200231,
0.0088524707,
-0.0009650491,
-0.0257303473,
-0.0011109867,
0.0749973133,
-0.0032978938,
-0.023161836,
0.0428444631,
0.0389073975,
0.0251463894,
-0.0892597586,
-0.0249235071,
0.0378158465,
-0.0326961949,
0.0068779639,
0.0327482112,
0.0131569114,
0.0036847396,
-0.053487625,
0.0291645881,
0.0049874787,
0.0009128162,
-0.016467737,
-0.0104250023,
0.0510710813,
0.0060566617,
0.0040916707,
0.0217277631,
-0.0399088524,
-0.0087876925,
0.0363453291,
0.0650059208,
0.0175118949,
0.0146225588,
-0.0066461791,
-0.0068307989,
0.0296840072,
0.0474434644,
-0.0580845885,
-0.0338437669,
-0.0069227293,
0.0285307132,
-0.0102651017,
0.0220552161,
-0.0466478914,
0.0056270538,
-0.0104389414,
0.0498628654,
-0.0216011778,
-0.0354850292,
0.0482241027,
0.0090677384,
0.0163387377,
0.0203053169,
-0.0123509569,
0.0193026327,
0.0048314715,
-0.0202909373,
0.0113467034,
0.0063291485,
-0.0005921961,
0.0321439691,
-0.0287488196,
-0.0013883391,
-0.0066698329,
-0.0702308416,
-0.0062314128,
-0.0564889722,
-0.0070848232,
0.0270243082,
-0.0086698728,
0.0361082628,
-0.0223666131,
-0.0262724794,
0.0082325926,
0.0055002584,
0.0673956275,
0.0006317579,
0.0585643835,
-0.0078400215,
0.0108801322,
-0.0168782584,
0.0282889046,
-0.0629485548,
-0.0802674145,
-0.0092042033,
0.0239835586,
0.0194258504,
0.0001299825,
0.0125952847,
-0.0188348144,
-0.0086011272,
-0.0244196281,
-0.0116911894,
0.0108003458,
-0.0042304024,
0.0316528305,
0.0262196809,
-0.0062422459,
-0.0061392542,
0.0728379637,
-0.0377638713,
0.0469459631,
-0.0102585582,
-0.0333067812,
-0.0445825495,
-0.0380599201,
-0.0324058793,
0.0440243818,
-0.0619434528,
-0.0011152525,
-0.079244107,
-0.0257200338,
0.0219916664,
-0.0093487985,
-0.0216629636,
-0.029579293,
0.0445624031,
0.0522747263,
0.0130918315,
-0.0129784364,
-0.0284136292,
0.0640764609,
-0.0522689037,
0.0136623317,
-0.0000794233,
-6.079407725e-33,
0.0127680292,
-0.0481876135,
0.0097037833,
-0.0375037901,
0.0271825008,
-0.0019104531,
-0.0308618378,
-0.0029882216,
-0.0112795318,
0.021536218,
-0.0323330387,
0.019586755,
-0.0066032447,
0.0549014471,
0.0121667795,
-0.0531274788,
0.0183882676,
0.0144456346,
-0.0209155921,
0.0505417995,
-0.0025196262,
0.0418017097,
0.0097588077,
-0.0869349539,
0.0305325035,
-0.015743779,
0.0404264182,
0.0212742798,
-0.0304199196,
0.0499861985,
-0.04610442,
0.0564673468,
0.0379595384,
-0.0170653276,
0.0025243573,
0.0036566737,
-0.0497877337,
-0.0113565391,
-0.0132265314,
-0.0327895582,
-0.0530962683,
-0.0127081387,
0.0748633817,
-0.0171739981,
0.0017421062,
-0.0114464862,
0.0179591291,
-0.03680145,
-0.0171535872,
-0.0144391907,
-0.030284036,
0.0130914627,
-0.0099371755,
-0.0128161218,
-0.036883451,
0.0585060045,
-0.0267410167,
0.0452120304,
-0.0490044318,
-0.0158888809,
-0.0008695747,
0.002255653,
-0.0086257448,
-0.0334819071,
-0.0040614274,
0.0676950216,
0.0417276807,
0.0230878573,
-0.0590017326,
-0.0115504405,
-0.069028303,
0.0305082873,
0.0743982568,
0.0437343121,
0.0477866493,
-0.0189564638,
-0.0896532536,
0.024505727,
-0.0883450136,
0.0411772728,
0.0444099829,
-0.0148789482,
-0.0784067065,
-0.01027364,
-0.0290774181,
0.0397384688,
-0.0427508429,
-0.0197323021,
0.0501006432,
-0.0044958177,
-0.0081706475,
0.0011324172,
-0.0076300823,
-0.0232149754,
0.0265820939,
-0.0106773051,
0.0159483124,
-0.0094966311,
-0.0306903813,
0.0033411996,
-0.0414484367,
0.0388775095,
0.0011037389,
-0.0101938164,
0.0473757535,
0.0223313328,
0.0263179895,
-0.0540205054,
-0.0543737747,
-0.0043682968,
0.0154221952,
0.0220792796,
-0.0136478972,
0.0055051614,
0.045914866,
0.0469852723,
0.0025791936,
0.0584815145,
-0.0001715178,
0.0208036732,
-0.046149686,
-0.0536610149,
-0.0173228998,
-0.0133632682,
-0.0474212691,
0.0012013521,
-0.0422292575,
0.0457735583,
0.0166428182,
-0.0443251692,
0.0047433479,
-0.0103476392,
2.731e-7,
0.0938907713,
0.0584635809,
0.006062245,
-0.0309531391,
0.0161248203,
0.0181625839,
0.0115367016,
-0.0119066685,
0.0386849083,
-0.041705668,
0.0102607412,
0.011283299,
0.0399755426,
0.0586648695,
-0.0525908284,
-0.0167685356,
-0.0899888724,
-0.0068165977,
-0.0258704778,
-0.0316532701,
0.0739474669,
0.0604257584,
-0.0516999923,
-0.0371166617,
0.0082527418,
-0.0178506151,
0.0570547245,
-0.0147143034,
-0.0097515658,
0.0493570194,
0.0283469446,
0.0065834192,
0.0302005019,
0.0538415462,
-0.0319763497,
-0.0400030315,
0.0205159336,
0.0749340132,
-0.0512886532,
-0.0580354072,
-0.0169891715,
0.013734566,
0.0541401133,
-0.0713157281,
0.0621092692,
-0.0640031472,
-0.045137804,
-0.0209933165,
-0.028267663,
-0.0262465514,
0.0905555338,
-0.0060144281,
-0.0264262371,
0.0285140388,
0.0114592155,
0.0598944835,
-0.0017223733,
-0.0634393692,
-0.0379049405,
0.0473280661,
-0.0237243995,
-0.0050649322,
0.0299449824,
-0.0310171451,
-0.0415459052,
-0.0729195625,
0.0302294474,
2.366260715e-34,
0.0197362509,
0.0346519202,
0.0218051635,
0.0239269845,
0.0316824391,
-0.0130164586,
0.0152189611,
0.014082131,
0.0272891596,
-0.0654792339,
0.0113916239
] |
http://arxiv.org/abs/2401.14490v1 | http://arxiv.org/pdf/2401.14490v1 | LongHealth: A Question Answering Benchmark with Long Clinical Documents | 2024-01-25T19:57:00 | 2024-01-25T19:57:00 | Lisa Adams, Felix Busch, Tianyu Han, Jean-Baptiste Excoffier, Matthieu Ortala, Alexander Löser, Hugo JWL. Aerts, Jakob Nikolas Kather, Daniel Truhn, Keno Bressem | cs.CL | cs.CL | Background: Recent advancements in large language models (LLMs) offer
potential benefits in healthcare, particularly in processing extensive patient
records. However, existing benchmarks do not fully assess LLMs' capability in
handling real-world, lengthy clinical data.
Methods: We present the LongHealth benchmark, comprising 20 detailed
fictional patient cases across various diseases, with each case containing
5,090 to 6,754 words. The benchmark challenges LLMs with 400 multiple-choice
questions in three categories: information extraction, negation, and sorting,
challenging LLMs to extract and interpret information from large clinical
documents.
Results: We evaluated nine open-source LLMs with a minimum of 16,000 tokens
and also included OpenAI's proprietary and cost-efficient GPT-3.5 Turbo for
comparison. The highest accuracy was observed for Mixtral-8x7B-Instruct-v0.1,
particularly in tasks focused on information retrieval from single and multiple
patient documents. However, all models struggled significantly in tasks
requiring the identification of missing information, highlighting a critical
area for improvement in clinical data interpretation.
Conclusion: While LLMs show considerable potential for processing long
clinical documents, their current accuracy levels are insufficient for reliable
clinical use, especially in scenarios requiring the identification of missing
information. The LongHealth benchmark provides a more realistic assessment of
LLMs in a healthcare setting and highlights the need for further model
refinement for safe and effective clinical application.
We make the benchmark and evaluation code publicly available. | LongHealth: A Question Answering Benchmark with Long Clinical Documents: Background: Recent advancements in large language models (LLMs) offer
potential benefits in healthcare, particularly in processing extensive patient
records. However, existing benchmarks do not fully assess LLMs' capability in
handling real-world, lengthy clinical data.
Methods: We present the LongHealth benchmark, comprising 20 detailed
fictional patient cases across various diseases, with each case containing
5,090 to 6,754 words. The benchmark challenges LLMs with 400 multiple-choice
questions in three categories: information extraction, negation, and sorting,
challenging LLMs to extract and interpret information from large clinical
documents.
Results: We evaluated nine open-source LLMs with a minimum of 16,000 tokens
and also included OpenAI's proprietary and cost-efficient GPT-3.5 Turbo for
comparison. The highest accuracy was observed for Mixtral-8x7B-Instruct-v0.1,
particularly in tasks focused on information retrieval from single and multiple
patient documents. However, all models struggled significantly in tasks
requiring the identification of missing information, highlighting a critical
area for improvement in clinical data interpretation.
Conclusion: While LLMs show considerable potential for processing long
clinical documents, their current accuracy levels are insufficient for reliable
clinical use, especially in scenarios requiring the identification of missing
information. The LongHealth benchmark provides a more realistic assessment of
LLMs in a healthcare setting and highlights the need for further model
refinement for safe and effective clinical application.
We make the benchmark and evaluation code publicly available. | [
0.043986205,
0.0503443405,
0.0183532257,
0.0287200697,
-0.0453695953,
0.0239998437,
-0.00816245,
0.0393327065,
0.0093605733,
-0.0062162103,
-0.0049585975,
-0.0634963885,
-0.0168293994,
0.0093894936,
-0.0200624615,
-0.0385623425,
0.0505992733,
0.0072784475,
0.0459964052,
0.0055943821,
-0.0351315401,
0.00183392,
-0.0218770783,
0.0305228252,
0.0299658906,
-0.0336319543,
0.0084136967,
-0.0108042238,
0.0151997702,
-0.0755977035,
-0.0027635638,
0.0283032581,
-0.0368229449,
0.029583279,
0.0000019951,
-0.0588948205,
-0.021936886,
0.0231195986,
-0.0508170202,
-0.0048637823,
0.0178100951,
-0.0307490565,
0.000395181,
0.0129221473,
-0.0283214822,
-0.0475106761,
0.0068381471,
0.0416496061,
0.0554477349,
0.0786629096,
-0.0102624735,
0.0023470819,
0.0583883896,
-0.0227215849,
0.03935536,
-0.0619016998,
0.0533750765,
-0.0160181895,
0.0584471002,
-0.0253019761,
-0.0035950029,
0.0515315793,
0.0287442487,
0.0622830316,
-0.006551913,
0.0477803648,
-0.0688097626,
0.0027900813,
-0.0028645361,
0.0997730494,
-0.0210432149,
0.0072000595,
0.003224317,
0.0064764228,
-0.0238613784,
-0.0217498299,
-0.0364846624,
0.009295526,
-0.0453542247,
-0.0060281246,
0.0092524663,
-0.016373381,
0.0346862599,
0.0106516676,
-0.0518673882,
0.0080186566,
0.0301667489,
-0.0353169851,
0.0276082903,
-0.0255378354,
-0.0436012968,
-0.0316989645,
0.0577071011,
0.060125947,
-0.0402564034,
0.0157548115,
-0.0125296218,
-0.0361564457,
0.0781057775,
-0.0885454714,
0.0116438745,
0.0190378651,
-0.0275396835,
0.0747634992,
-0.0420915671,
0.0222360641,
-0.0556129739,
0.0235528369,
-0.0699604377,
-0.0252056941,
-0.005015654,
-0.0135608222,
-0.0300361235,
0.0295568891,
-0.0387899354,
-0.0283723529,
-0.0467998385,
-0.0229459424,
0.0108670937,
0.0204489306,
0.0286992714,
0.0275449716,
-0.0108312145,
-0.009897816,
-0.0274775438,
0.0412090085,
-0.0427051634,
-0.019462483,
-0.0064218058,
-0.0575139448,
-0.0093513336,
0.0011238633,
0.0237599071,
-0.0436427183,
0.0529278629,
0.044135645,
-0.0028397497,
-0.0398432277,
0.0058395495,
-0.0148322135,
-0.0195013918,
-0.0489378981,
0.0001597576,
-0.0629561022,
0.0451503843,
0.0307346098,
-0.0085800271,
-0.0463894419,
0.0291074347,
0.0321906842,
-0.0559495278,
0.0491021015,
-0.0329689346,
0.0131783662,
0.0303850472,
0.0422967337,
0.0100058783,
0.0197993591,
0.0165250245,
-0.0254308023,
0.015574405,
-0.017172901,
0.0302236006,
0.0126133962,
0.0487845093,
-0.0442962237,
0.0014174655,
0.0598370545,
0.0225478821,
0.0852856562,
-0.0332837701,
0.1042217165,
0.0018556822,
-0.052402854,
0.0017753576,
0.0603114143,
0.0188220888,
-0.0191747639,
-0.0007100292,
-0.0245321952,
0.0215987619,
-0.0335667059,
-0.0149028292,
0.0083955256,
0.0168447383,
0.0212922525,
0.0043423609,
0.0026397333,
-0.0307313781,
-0.0061133807,
-0.0393614918,
-0.0186363216,
-0.0368759185,
0.0226956494,
0.0448621549,
-0.0145781944,
-0.0639902353,
0.0635771826,
0.0282265302,
-0.032908421,
-0.0142333377,
-0.0425606109,
0.0099536432,
0.0691689923,
0.0596799515,
-0.04271736,
-0.0373958088,
-0.0014545249,
-0.0101822969,
0.014203215,
0.0189180765,
-0.0134346467,
0.0393118598,
-0.0507555865,
0.0317469947,
0.0391082689,
-0.0037845464,
-0.0287278816,
0.0548451133,
0.0350258164,
-0.0039770189,
-0.0374910161,
0.0004867875,
0.037670888,
-0.0635066256,
-0.0483725592,
-0.0058756405,
-0.0016857745,
0.0313074626,
0.0080279866,
-0.0068486775,
0.0276790578,
0.0485346019,
-0.0470474809,
-0.0111047998,
-0.0471175052,
-0.0654817671,
-0.0385602415,
0.0299781226,
0.0009611865,
-0.0265714955,
0.0236897115,
-0.0106708836,
-0.0190356411,
-0.0539957434,
-0.025276009,
0.0148253012,
-0.013230525,
0.0077808006,
0.0505388901,
0.0006376901,
-0.0671696216,
0.0418151878,
-0.0249139797,
0.034143012,
0.0049889516,
-0.0087110195,
0.0037051248,
-0.0420358777,
-0.0352389812,
0.0498579592,
-0.0176998954,
-0.0038570585,
-0.045130264,
-0.0267487261,
-0.0469187237,
-0.0383356027,
-0.1246410534,
0.0253602099,
-0.002159267,
-0.0026905856,
0.021471668,
0.046997074,
-0.0238855835,
-0.0213826261,
-0.0378717333,
-0.008308446,
0.0609861687,
0.0238334239,
0.0193532724,
0.0552716479,
-0.0107322829,
-0.033450231,
-0.0215299726,
-0.0265655257,
0.0288167484,
-0.0414316244,
-0.0113813207,
0.0135751935,
-0.0055739544,
-0.0511232205,
0.0223994274,
-0.0065588458,
-0.0021282672,
0.008299239,
-0.0474171713,
0.0238204543,
0.0575923063,
0.0058087977,
0.0001866301,
0.0169894248,
-0.0119447196,
-0.0248248093,
0.039821554,
-0.0253250301,
-0.0162199717,
-0.0008649032,
-0.0215024948,
0.0038623817,
0.0610935092,
0.0200707689,
0.0031560897,
-0.0213959441,
-0.0305108484,
-0.0293957237,
0.0194264092,
-0.0413617715,
0.0206145532,
-0.042838186,
0.0222625751,
-0.0212141275,
-0.0153781343,
0.0778225735,
-0.0214879494,
0.0260028616,
-0.0769970268,
0.0019717517,
-0.0690688118,
0.0503844991,
-0.0091643883,
0.0144837955,
-0.0366997905,
-0.0158581622,
0.02189417,
0.0010366144,
-0.0303460956,
-0.0652033761,
0.0587988235,
-0.0168519188,
0.0485495329,
-0.0212424695,
-0.0442132913,
-0.0379981436,
-0.0526018441,
-0.0303997193,
-0.0165785514,
0.0127083315,
0.0258336924,
-0.02905499,
0.0159929655,
-0.00607209,
0.0860376433,
0.0032529884,
0.0015833668,
-0.0109721525,
0.0499058105,
0.0404993296,
0.0020054511,
0.0081733549,
-0.0269593783,
0.0516566262,
0.0891470537,
0.0506134778,
0.0026699288,
-0.0005371592,
-0.0121761877,
-0.0028299286,
-0.0473352857,
0.0130255558,
0.0294470452,
0.039096877,
-0.0726169199,
-0.0063277534,
0.1137933135,
0.0531278513,
-0.0513960458,
-0.0024805476,
-0.0177350733,
0.0003439983,
-0.0167625342,
0.0295657553,
-0.0655785054,
-0.0431997366,
0.0046592318,
0.0319105871,
-0.0125559457,
-0.0459505543,
0.0192166101,
-0.0587372333,
0.0247195158,
-0.0179578513,
-0.0490911864,
-0.0096051358,
-0.0410682,
-0.0178421531,
0.0102335326,
0.004687632,
0.0354664065,
-0.018435698,
-0.0053945761,
-0.0181378946,
-0.018773675,
-0.0342606567,
0.0470941961,
-0.0283460394,
0.0470431,
0.0366327576,
0.0236147735,
-0.004617041,
-0.002210215,
0.0417215303,
-0.02323539,
-0.0005665629,
-0.0166207906,
0.0901252404,
0.0124949683,
-0.0058194473,
-0.0648027956,
0.0201017577,
-0.0511653982,
-0.0196152087,
0.0179286916,
0.0890468284,
-0.0168626569,
-0.0262166038,
-0.043175675,
-0.0072240555,
-0.0154695595,
0.0624250956,
0.0006990469,
0.0094262464,
-0.0613272712,
0.0394907966,
0.0136023397,
-0.0144735165,
0.0065317168,
0.0125166243,
-0.0016351296,
0.0206704587,
-0.0437070839,
-0.0534627177,
0.0552159213,
0.0055442066,
-0.0081623709,
0.0358222127,
0.005449615,
0.0338693075,
-0.0598009042,
0.0558685102,
0.0118416371,
0.0097302338,
-0.0812654793,
0.0088904295,
0.0472452641,
0.0260976721,
-0.0096165761,
0.0273525938,
-0.0495076366,
0.0034150225,
0.0133433277,
0.0178908985,
0.0224221591,
-0.0037129815,
-0.0042654122,
-0.0103110392,
0.0150537305,
0.0115417279,
-0.0572306886,
-0.0059058648,
-0.021506371,
-0.0292085093,
0.003715096,
0.0236925688,
-0.0090896264,
0.0259169545,
0.0007359653,
0.0565028898,
-0.0034666879,
-0.0049660713,
0.0288813654,
0.0586908944,
0.0055185286,
0.0197323728,
-0.0089172572,
-0.0163165759,
0.037364278,
-0.0234040581,
0.0119525734,
-0.0069066272,
0.0020686705,
0.0695579499,
-0.0202939827,
-0.0156434439,
0.0253929589,
-0.024204893,
-0.0309064351,
-0.0496816598,
0.0323038362,
0.0341770798,
0.0040119323,
0.0594510771,
-0.0425653681,
-0.0181451365,
0.0210301857,
0.0003624293,
0.0617260411,
-0.0094481679,
0.0786388889,
-0.0165770724,
-0.0045031598,
-0.0015495489,
0.0327061899,
-0.0673540905,
-0.0278194565,
-0.0118351467,
0.0392951332,
0.0323563106,
0.000803605,
0.0180772245,
-0.0031594639,
-0.0087730391,
-0.0257716719,
-0.0060744281,
-0.0285750683,
0.0061562522,
0.0278839543,
0.0388582647,
-0.0004085159,
0.0156222684,
0.0808285996,
0.0144820455,
0.0365004092,
0.0229412951,
-0.0180859398,
-0.0548895784,
-0.0262118895,
-0.0149881877,
0.0263000298,
-0.067664437,
0.0121498993,
-0.0513642915,
-0.0238656793,
0.0209865738,
0.0006839472,
-0.029399246,
-0.012487825,
0.0329159312,
0.0511392802,
0.0257986858,
-0.0472143926,
-0.0241836682,
0.0782377571,
-0.0655879974,
-0.021685129,
0.0146118971,
-6.05105627e-33,
0.004303888,
-0.013412524,
0.0283290856,
-0.0432479456,
-0.0153989298,
0.0367179364,
-0.0439070575,
-0.008201695,
-0.0327305384,
0.0115508558,
-0.048765216,
0.0203239322,
0.0157280136,
0.0439690053,
-0.0174785685,
-0.0492391661,
0.0022307441,
0.0081878211,
-0.0264871214,
0.013171806,
-0.0153579582,
0.0654311478,
0.0103229806,
-0.0594331026,
-0.0121425111,
0.0011144002,
0.048466444,
0.0296447556,
-0.0244037081,
0.061953716,
-0.0445965454,
0.0226792451,
0.0495153442,
-0.0207632314,
0.0157571081,
-0.0016231867,
-0.114073135,
0.0110512748,
-0.0161267594,
-0.0220907014,
-0.0587008819,
-0.0395470597,
0.0520830229,
-0.0026788416,
-0.0110442238,
-0.0099881263,
0.0006170548,
-0.0417212918,
-0.0222142767,
-0.003616188,
-0.0267521013,
0.0548152626,
0.0245688614,
-0.0114103677,
-0.022492541,
0.0059515056,
-0.0495082289,
0.0606932752,
-0.0375576839,
-0.0543391816,
0.0115628298,
0.0303799752,
0.0180618148,
-0.0232613217,
-0.0295625273,
0.0589414649,
0.0495306849,
-0.0078660585,
-0.0354399681,
-0.0100975009,
-0.0474599637,
0.0270321667,
0.0664888993,
0.0191027634,
0.0445845574,
-0.0111501347,
-0.0838891044,
0.0459405445,
-0.0325650983,
0.0489852056,
0.0383910425,
-0.0213706288,
-0.0524729118,
0.0033133698,
-0.0167654715,
0.0451311804,
-0.0357054807,
-0.0121615315,
0.0642556623,
-0.0035814433,
-0.0009835353,
0.0048738876,
0.0373948738,
-0.0108031491,
0.0270378217,
0.0127865886,
0.0492058657,
0.0395997241,
-0.0256190468,
0.0301582515,
-0.0537483804,
0.0174894743,
0.0130794179,
-0.0135955112,
0.009208506,
-0.0045250766,
-0.0041485187,
-0.0456558838,
-0.0621664934,
-0.0172418617,
0.0072096237,
0.0050779069,
-0.0220041778,
0.0449846946,
0.035637442,
0.0223357491,
0.0104488982,
0.0698457956,
-0.0181457903,
0.0145925712,
-0.0537426732,
-0.056970574,
0.0145199569,
-0.0196065586,
-0.0280621331,
0.0191443525,
-0.0479919612,
0.0524236672,
-0.0020190817,
-0.0730262175,
0.0150359953,
0.0424014553,
2.667e-7,
0.0945437029,
0.0522437729,
0.0484274961,
-0.0136950891,
0.0200311132,
-0.0190960336,
-0.0022642531,
-0.0141954757,
-0.0031711131,
-0.004905853,
0.018874418,
0.0314818285,
0.045909036,
0.0522718653,
-0.070754014,
-0.0053963251,
-0.0815213472,
-0.0292908382,
-0.0687702745,
-0.0279090311,
0.0421698987,
0.0992598459,
-0.0564901307,
-0.0312201437,
0.0221828427,
-0.0260782782,
0.0372909307,
-0.0123780696,
-0.0313847959,
0.0371377952,
0.0013374203,
-0.0086132335,
0.0310609508,
0.0205544196,
0.0050628236,
-0.0014535026,
0.0313769244,
0.1163251549,
-0.0258798897,
-0.0352124311,
0.0139067722,
-0.0179036055,
0.0121822236,
-0.0563535094,
0.0995211974,
-0.0317006186,
-0.035333313,
-0.010206419,
-0.0232176296,
-0.0168203004,
0.050774727,
0.0007063365,
-0.0048459843,
0.0179683715,
0.0197621807,
-0.0134471431,
-0.0261230636,
-0.0618216954,
-0.0305261239,
0.0215571858,
-0.0380001441,
-0.0318004377,
-0.0086870957,
-0.0166846178,
-0.0106051601,
-0.0657427609,
-0.0051840427,
2.613934161e-34,
0.0311884433,
-0.0172448866,
0.0128102507,
0.0316541456,
0.0238381717,
-0.0164675601,
-0.0059643458,
0.0110492613,
-0.0071763643,
-0.0914684385,
0.0031514459
] |
http://arxiv.org/abs/2401.14400v1 | http://arxiv.org/pdf/2401.14400v1 | Modular Adaptation of Multilingual Encoders to Written Swiss German Dialect | 2024-01-25T18:59:32 | 2024-01-25T18:59:32 | Jannis Vamvas, Noëmi Aepli, Rico Sennrich | cs.CL | cs.CL | Creating neural text encoders for written Swiss German is challenging due to
a dearth of training data combined with dialectal variation. In this paper, we
build on several existing multilingual encoders and adapt them to Swiss German
using continued pre-training. Evaluation on three diverse downstream tasks
shows that simply adding a Swiss German adapter to a modular encoder achieves
97.5% of fully monolithic adaptation performance. We further find that for the
task of retrieving Swiss German sentences given Standard German queries,
adapting a character-level model is more effective than the other adaptation
strategies. We release our code and the models trained for our experiments at
https://github.com/ZurichNLP/swiss-german-text-encoders | Modular Adaptation of Multilingual Encoders to Written Swiss German Dialect: Creating neural text encoders for written Swiss German is challenging due to
a dearth of training data combined with dialectal variation. In this paper, we
build on several existing multilingual encoders and adapt them to Swiss German
using continued pre-training. Evaluation on three diverse downstream tasks
shows that simply adding a Swiss German adapter to a modular encoder achieves
97.5% of fully monolithic adaptation performance. We further find that for the
task of retrieving Swiss German sentences given Standard German queries,
adapting a character-level model is more effective than the other adaptation
strategies. We release our code and the models trained for our experiments at
https://github.com/ZurichNLP/swiss-german-text-encoders | [
0.009175322,
0.109115012,
0.0169967376,
0.0527672805,
-0.0265726354,
0.0185369346,
0.017753154,
-0.0161004029,
-0.0059589171,
-0.009338744,
-0.0176119469,
-0.0586062931,
-0.0337603241,
0.0254776143,
0.0261673946,
-0.0357498713,
0.0391324237,
-0.0351082385,
-0.0308281295,
-0.0191291701,
0.0251415707,
-0.0097259777,
0.0256475247,
0.0139639573,
0.036477413,
0.0418818332,
-0.0153621342,
0.0400866382,
-0.0177853163,
0.0082821827,
0.0315491632,
-0.0160532482,
0.0365209468,
0.0383989848,
0.000001635,
-0.0299588479,
-0.0473453738,
-0.0328392945,
0.0079934355,
0.0250657555,
0.0250311233,
-0.0634828135,
-0.0236554556,
0.0449920334,
-0.0169627219,
0.0144445216,
0.0505280718,
0.0243331175,
0.0235178266,
0.0914464518,
-0.0389292873,
-0.0318174213,
0.0027551868,
-0.0512067676,
-0.0293377489,
0.0089749172,
-0.0178443547,
-0.0371618792,
-0.0156346541,
0.0067869341,
-0.046765577,
0.049270317,
0.0241000708,
0.0282173678,
-0.0387774631,
-0.0238875989,
0.0108586028,
0.0132028228,
-0.0087890197,
0.0381831899,
-0.0414774604,
-0.0153365228,
-0.0541714914,
-0.0068839453,
0.0145866899,
0.0047922079,
-0.0385631956,
-0.0118458774,
0.0158376507,
0.0047690547,
0.0260540023,
-0.0111440495,
-0.0108469194,
-0.020956276,
-0.0160439629,
0.0616919845,
0.0255854018,
-0.0383294709,
0.0142451711,
0.0346035585,
0.021121826,
0.0218863115,
0.0326282196,
0.0129504725,
0.0087395674,
0.0090593193,
-0.0402581245,
-0.0223824624,
-0.0142239593,
-0.0366302952,
0.0131979221,
0.0497522131,
-0.0153728165,
0.0162771139,
-0.0168410521,
0.0395488925,
-0.0234466176,
0.042743329,
-0.0603484809,
-0.0651158616,
-0.0566637665,
0.021542361,
-0.0542749166,
0.0111391945,
-0.0542577021,
0.0141301975,
-0.0033356277,
0.0137044946,
0.0530148745,
-0.0283668041,
0.0119175874,
0.0090017775,
-0.0521185361,
-0.026534522,
-0.026836209,
0.1193048954,
-0.0295528751,
-0.0031501232,
-0.0311711077,
-0.0396624319,
-0.0392680317,
0.0169144552,
0.0276314747,
-0.0534693524,
0.0283569526,
-0.0079170717,
-0.0744651258,
-0.0244421382,
-0.0544795506,
0.0324485227,
-0.0101164477,
0.0029332025,
0.0039863689,
-0.0614774935,
0.0715391561,
-0.0064576897,
0.0271377787,
-0.0742736235,
-0.0194990262,
-0.025749797,
-0.034907233,
0.0412048362,
-0.0312831253,
-0.030638596,
0.0302966703,
-0.0268607605,
0.0205382183,
0.0767302811,
0.0125882179,
-0.0093246885,
0.0039260769,
-0.0059313686,
-0.0020539628,
0.0557320304,
0.0005921622,
0.0153392414,
-0.0457631126,
-0.0132039413,
0.0011491445,
-0.0062772683,
0.0392787941,
0.0577539206,
-0.0159531664,
-0.0239415579,
0.0453490429,
0.009507414,
0.0642947257,
-0.0200779904,
0.0090907961,
-0.0199089851,
-0.0226679314,
-0.0030744823,
0.0070322128,
0.009584344,
-0.0357864164,
0.0150583154,
-0.0834634528,
0.0395853072,
-0.0483513921,
-0.0083845295,
0.0254283492,
-0.0147172716,
0.0557814725,
-0.0198873263,
0.0489642695,
0.0191816911,
-0.0250255186,
0.0224207882,
0.0273569897,
-0.0588268824,
0.0528587885,
0.0066649141,
-0.0289334208,
0.080742836,
0.048185382,
-0.0194060244,
-0.0261332076,
-0.004840604,
-0.0351467989,
0.0391906239,
0.0705820993,
-0.0032164436,
0.077335909,
-0.0886446834,
-0.0114258816,
-0.0305730104,
-0.0007693872,
0.0679003,
-0.0162633397,
0.017183736,
0.0013955294,
-0.0111475121,
0.0088928249,
0.0166853014,
-0.0669641569,
-0.0077489745,
0.0302736796,
0.0108610587,
-0.0184642393,
0.010115494,
-0.0300382674,
-0.0200560726,
0.0143938437,
-0.0901860967,
0.0160843655,
0.0440585017,
0.0216411334,
-0.0598761328,
-0.0159396976,
-0.0257841051,
0.048548799,
0.0124772638,
-0.0255701058,
-0.0634435713,
-0.0075711776,
-0.0552993938,
0.0655097142,
-0.0346303545,
0.0066213808,
-0.0179681983,
0.0399779677,
-0.0337607637,
0.0178780481,
0.0101799825,
-0.0156263933,
-0.021614369,
-0.0306687914,
0.0148304608,
-0.0253172703,
-0.0317944027,
0.029631434,
-0.0116494512,
0.0211918578,
0.0197979417,
0.0212124418,
0.0155930594,
-0.0364645608,
-0.0788644999,
-0.0103342263,
-0.0746057928,
0.0459187925,
0.0354470164,
0.0708938837,
-0.0049311933,
-0.0729646683,
-0.0349654853,
0.0211910438,
-0.0008275302,
0.0110732559,
0.0241468381,
-0.0014015683,
0.0281733461,
-0.0170016047,
-0.0309049021,
0.0149530461,
0.0593479984,
-0.0554232374,
-0.05034649,
-0.0678557903,
-0.0326104537,
-0.0285280105,
0.0407206751,
0.008765962,
-0.0633808225,
-0.040180143,
-0.0297262967,
0.0277049802,
0.0507798344,
-0.0015326829,
0.0023560778,
-0.0001276023,
0.0143432943,
-0.0435176231,
0.0244818814,
-0.0287888311,
0.0155834462,
-0.1028371602,
-0.0200209506,
0.0379971303,
0.0443230346,
-0.0392833315,
0.0126348035,
-0.0517627299,
-0.036338456,
0.0086333212,
-0.0010786369,
-0.0260961968,
-0.0144449184,
0.0114052193,
0.0250084866,
-0.0004863267,
0.0155040445,
0.062066786,
-0.0106706191,
-0.0190880895,
-0.0624557771,
-0.0448093414,
-0.0290994737,
0.0551411361,
0.0342428274,
0.0412223265,
-0.0092620514,
-0.036420919,
-0.0325428136,
-0.0310149193,
0.0007534919,
0.0160714649,
0.064622879,
-0.0114263976,
-0.0078143561,
-0.0014409169,
-0.0518493056,
-0.0072927605,
-0.0207383744,
-0.024791386,
0.0287538785,
0.03091662,
-0.0167465303,
-0.0601061918,
0.0047612465,
0.0236528125,
0.097728014,
-0.0158850178,
0.0159894302,
-0.0390052907,
0.0331911854,
0.025099244,
0.0705687478,
-0.0074606845,
-0.043053735,
0.0545166731,
0.0031351692,
0.0158583075,
-0.0193621367,
0.007495692,
0.0015836605,
0.0311140064,
-0.0347267278,
-0.0115694571,
-0.0231267586,
0.0660488531,
-0.0120345224,
-0.0523623377,
0.0902219862,
0.020906372,
-0.0572625957,
-0.0066979271,
-0.0168594662,
0.0875767097,
0.0194110516,
0.0510851704,
-0.0329385661,
-0.0267926808,
-0.0197250955,
0.0247525387,
0.0141497022,
0.0483122319,
-0.0043903408,
-0.111913383,
-0.0024669012,
0.0354458429,
-0.0639610663,
-0.0217896588,
-0.0077938866,
-0.0427778624,
-0.0394273624,
0.041431617,
0.016048314,
0.0065435017,
-0.0290782843,
0.0201030988,
-0.0271995906,
-0.0681163296,
-0.0417886041,
-0.0243076608,
0.0340083055,
-0.0094550727,
-0.0311889052,
-0.0495424382,
-0.0344847217,
0.0200768244,
-0.0233269185,
0.0204061773,
0.0025121858,
0.0846033394,
0.0075676581,
-0.0246576127,
0.0294808149,
-0.0130920159,
-0.073773399,
-0.0177015215,
0.0017964684,
0.0237866901,
-0.0130435638,
0.0277822912,
0.0008792822,
-0.0200884584,
-0.005286362,
0.025668608,
0.0120342439,
-0.0674644187,
-0.0108620655,
-0.0538022146,
-0.0300144609,
-0.1025090665,
0.0129836071,
0.0426937379,
-0.0159623288,
-0.0184228867,
-0.0383205861,
-0.0588482656,
-0.0286730602,
0.0514673367,
0.0155532034,
-0.0119049614,
0.0526605323,
-0.0144376215,
-0.0622508973,
0.0103977816,
-0.0117136631,
-0.0220577251,
-0.0347956903,
0.0017162339,
0.0530795343,
-0.0303687919,
-0.0062090205,
-0.067068845,
-0.0374809429,
-0.0563745983,
-0.0100772483,
0.0199897643,
0.0138469329,
0.0773088485,
-0.0531963892,
-0.0601275712,
0.0065476112,
-0.0064519709,
-0.0097736977,
0.0047795055,
0.0300716851,
-0.0537887625,
-0.0400117822,
0.0074985144,
-0.0162012428,
0.0433759652,
-0.0166499466,
0.0216030832,
-0.0263614636,
-0.0059005148,
0.0174160935,
0.0129164578,
-0.0229013506,
0.0194831844,
0.0047504408,
-0.0311585218,
0.0408044159,
0.0073264306,
0.0143035101,
-0.0221182089,
-0.0271927733,
0.0602629781,
-0.0194533188,
-0.0020193814,
0.0173203722,
0.0143943429,
0.0193636529,
-0.0111037288,
-0.0056821252,
0.0412102565,
0.0212168619,
-0.0047896821,
0.0218519457,
0.0230710246,
0.0848117024,
0.0244717896,
0.0628554001,
-0.0546634085,
0.0306178406,
-0.0067068865,
0.0171575099,
-0.0098235719,
0.0141454842,
-0.09446913,
0.0043688519,
-0.0927457511,
0.0049811252,
0.034495458,
0.0153940823,
0.1128448844,
0.0095964111,
0.015785329,
-0.0063831555,
-0.0152777722,
-0.0096585099,
0.0088990936,
0.0011615934,
0.0013947721,
-0.0267046448,
0.0021595939,
0.0138021857,
0.0092165116,
0.0454520062,
0.0015185599,
0.0065767537,
-0.0226353575,
-0.0419422947,
0.0583785288,
0.0159029104,
-0.0408940278,
-0.0080801118,
0.0235475805,
-0.0159139521,
0.0028837498,
0.0096297441,
0.0092195561,
0.000774404,
0.0223540012,
0.0265622567,
-0.0019454996,
0.0232302267,
-0.0178858228,
0.0244188458,
0.0099449595,
-0.041811429,
0.0056826854,
-5.705337872e-33,
-0.0330024883,
0.0522713885,
-0.0183087271,
-0.0341083072,
-0.007574602,
-0.0409908965,
-0.0051069572,
-0.0152803846,
-0.0537312627,
-0.0043377373,
0.007833194,
0.0114942938,
0.0269645862,
0.0675163269,
0.0224083085,
-0.0317330509,
0.0288501214,
0.0109876888,
0.0029236225,
-0.0077265189,
0.03046152,
0.0375660881,
0.0829294473,
-0.024001047,
-0.0283273533,
0.0174786747,
-0.0102717998,
-0.0022858025,
0.0243862364,
0.0672458783,
-0.0004424505,
0.0632293671,
0.0200968962,
-0.0035181118,
-0.0151273031,
-0.001159141,
-0.0637200549,
-0.0111613683,
-0.0423578434,
0.0903746411,
-0.003340092,
-0.0344240367,
0.0570124462,
0.007477514,
-0.0001720001,
0.019215418,
0.0229554363,
0.002900294,
-0.0043304767,
-0.0097491415,
-0.0459086932,
0.0060296827,
0.0163461305,
0.028927397,
0.0307363402,
0.1005531102,
0.0106767137,
-0.0086661894,
-0.0421822406,
0.0086652432,
0.0049157296,
0.0329857171,
-0.004848239,
-0.016452536,
-0.0496646017,
0.0031685962,
0.0297159795,
0.0020757969,
0.019389227,
-0.0001818572,
0.0918962806,
-0.0206921026,
0.0564251058,
0.0510117225,
0.0676318184,
-0.0426949225,
-0.0571338348,
0.016039798,
0.0218555611,
-0.0222964678,
-0.0103172194,
0.0263028629,
0.0263133608,
-0.00451086,
-0.0148532521,
-0.0222015418,
-0.0334021747,
-0.0220774487,
0.0436824411,
-0.0052121575,
-0.0010340714,
-0.0012558051,
0.0711087063,
-0.0489889011,
0.0190746151,
0.0044073481,
-0.0102725523,
0.0307696462,
-0.0153743792,
0.0523692295,
-0.0086288061,
-0.0677211657,
0.0129289208,
0.0217282251,
-0.0190424323,
-0.0054030982,
0.077767387,
0.0226084553,
-0.0690112263,
0.0059890002,
-0.0151026724,
0.0315692611,
-0.0193987414,
0.0774979815,
-0.0208249204,
0.0004527142,
0.0087466715,
0.0029740366,
0.0179175269,
-0.0388934761,
-0.029727852,
-0.0431467853,
0.0536121726,
0.061854139,
-0.0038042506,
0.0198675282,
-0.0130415307,
-0.0244665463,
-0.0146836238,
-0.024195984,
0.0091228094,
0.036077816,
2.414e-7,
0.0804639161,
0.0423308387,
0.0061411355,
-0.0287123248,
0.0241751391,
-0.0019662727,
-0.0318945199,
0.0520904884,
-0.0020832687,
0.0081820143,
0.0385062359,
0.0164142288,
0.0568778738,
0.0366647057,
-0.0460282043,
-0.0246701781,
-0.0226825271,
0.0143390559,
-0.0246869847,
0.0274108108,
0.0859996825,
0.1219434291,
-0.0058471193,
0.0019291708,
0.0134764677,
-0.0118998503,
0.0715285763,
0.0125680016,
-0.0342825018,
0.0381220058,
-0.0266888738,
0.0303045288,
-0.0301588029,
0.014459393,
-0.014932449,
0.0304485336,
-0.0065482901,
0.0408355556,
-0.0416735299,
-0.0812549368,
0.0025925583,
0.0281002745,
-0.0108944103,
0.0033625502,
0.0315251462,
-0.0003605617,
0.0328834616,
0.0240194257,
-0.0163199306,
0.0300475489,
0.062512733,
-0.0329486243,
0.0002076057,
0.0044314405,
0.0895240679,
0.0079806466,
-0.0574650392,
-0.0456947125,
-0.0350470953,
0.0327494666,
0.0017111995,
-0.0105979834,
-0.0163088255,
0.002928765,
-0.0231599845,
-0.0632286817,
0.0190868117,
1.944837036e-34,
0.0375412889,
0.0149611058,
0.0030161296,
0.0352488831,
0.0132156042,
-0.0311398245,
-0.0323367082,
-0.0105210505,
-0.0219985098,
-0.0552405603,
-0.0192716066
] |
http://arxiv.org/abs/2401.14447v1 | http://arxiv.org/pdf/2401.14447v1 | Wordflow: Social Prompt Engineering for Large Language Models | 2024-01-25T18:58:11 | 2024-01-25T18:58:11 | Zijie J. Wang, Aishwarya Chakravarthy, David Munechika, Duen Horng Chau | cs.HC | cs.HC, cs.AI, cs.CL, cs.LG | Large language models (LLMs) require well-crafted prompts for effective use.
Prompt engineering, the process of designing prompts, is challenging,
particularly for non-experts who are less familiar with AI technologies. While
researchers have proposed techniques and tools to assist LLM users in prompt
design, these works primarily target AI application developers rather than
non-experts. To address this research gap, we propose social prompt
engineering, a novel paradigm that leverages social computing techniques to
facilitate collaborative prompt design. To investigate social prompt
engineering, we introduce Wordflow, an open-source and social text editor that
enables everyday users to easily create, run, share, and discover LLM prompts.
Additionally, by leveraging modern web technologies, Wordflow allows users to
run LLMs locally and privately in their browsers. Two usage scenarios highlight
how social prompt engineering and our tool can enhance laypeople's interaction
with LLMs. Wordflow is publicly accessible at
https://poloclub.github.io/wordflow. | Wordflow: Social Prompt Engineering for Large Language Models: Large language models (LLMs) require well-crafted prompts for effective use.
Prompt engineering, the process of designing prompts, is challenging,
particularly for non-experts who are less familiar with AI technologies. While
researchers have proposed techniques and tools to assist LLM users in prompt
design, these works primarily target AI application developers rather than
non-experts. To address this research gap, we propose social prompt
engineering, a novel paradigm that leverages social computing techniques to
facilitate collaborative prompt design. To investigate social prompt
engineering, we introduce Wordflow, an open-source and social text editor that
enables everyday users to easily create, run, share, and discover LLM prompts.
Additionally, by leveraging modern web technologies, Wordflow allows users to
run LLMs locally and privately in their browsers. Two usage scenarios highlight
how social prompt engineering and our tool can enhance laypeople's interaction
with LLMs. Wordflow is publicly accessible at
https://poloclub.github.io/wordflow. | [
0.0566429868,
0.024682425,
-0.0346399657,
0.0136438273,
-0.037136592,
-0.0177324433,
0.0514704995,
0.0289947633,
0.0600088164,
0.0169860013,
-0.0369555838,
-0.0059511857,
0.0115886098,
0.0841127187,
0.0065964386,
-0.04593198,
0.0076994481,
-0.0417927168,
0.0149245458,
0.0018559971,
0.0145699875,
0.0333247706,
-0.0206816364,
0.0525393598,
0.0090659186,
-0.0247900579,
-0.0230804011,
0.0033399211,
-0.0298308004,
-0.0360833965,
-0.0245094635,
0.0802457333,
-0.0183697958,
0.0441610478,
0.0000020447,
-0.0723295882,
0.025807796,
-0.0251240302,
-0.0463152193,
-0.0079590594,
0.0272371192,
0.0479262732,
-0.0175494757,
0.0289770104,
-0.0502007306,
0.0165663268,
0.0628665835,
0.002857032,
0.0691818818,
0.0280321855,
-0.0066176862,
0.0250339415,
-0.0209956411,
-0.0250175986,
0.0240188017,
-0.0092316773,
-0.0258206185,
-0.0592270233,
0.0835897699,
0.0224481188,
0.0469130687,
0.0328910425,
0.0167913865,
-0.0379649103,
0.0478503294,
0.0381928012,
-0.0044686343,
-0.0806760415,
-0.0546878725,
0.0662311465,
0.0456316844,
0.0046535679,
0.0015541016,
-0.0190024115,
0.0526992306,
0.0577139184,
-0.0506987497,
0.0219431706,
0.007640087,
-0.0232265908,
-0.012902501,
0.0119119594,
-0.0196930431,
-0.0000901956,
0.0537031665,
0.0606857501,
0.02175528,
0.0184410755,
-0.0309380833,
-0.0236716885,
0.0243418142,
-0.0771970004,
0.0516020246,
0.051336728,
-0.0408614166,
0.0048622014,
-0.0179656725,
-0.013207593,
0.0324812159,
0.0115978662,
0.0362962037,
0.0223632865,
-0.0141558424,
0.0365073867,
0.0361019671,
0.0204877816,
0.0244858917,
-0.0143331503,
0.0217141025,
0.0078961374,
-0.0209632032,
0.0026435407,
-0.0698442683,
0.0049432698,
-0.0449219197,
-0.0043038297,
-0.0029338445,
0.0640351847,
-0.0232969616,
0.0573880672,
0.0368327051,
0.0192287378,
-0.0055352091,
-0.0175102614,
-0.0098180817,
-0.013095377,
-0.0549728908,
-0.0295740198,
0.0234299544,
-0.0661837533,
0.0190912411,
-0.0101390453,
0.0276739113,
-0.001252588,
-0.0284903776,
0.0338556133,
0.0073818988,
-0.0136394547,
0.0030641882,
-0.0664969161,
-0.0047380496,
0.0002579052,
-0.0605669767,
-0.0877554566,
0.0395783,
0.0474003777,
-0.0092605148,
-0.0578476451,
-0.0091716265,
0.0485289283,
0.0199837405,
0.028140517,
-0.062675193,
0.0152251814,
0.0734256133,
0.0403592885,
-0.0204600468,
0.0244957265,
0.0209826529,
0.0518615395,
0.0606979206,
-0.0317389742,
0.0501758382,
0.0622985698,
-0.033396285,
-0.0044201585,
-0.003306689,
-0.0010030229,
0.0384990424,
-0.0002555392,
-0.0580647402,
0.0380246229,
0.0082244454,
0.0142182689,
0.0686303154,
0.075450018,
-0.0166806765,
0.0355556943,
0.0279855672,
0.0238850825,
0.0048952335,
-0.0429006293,
-0.0236156005,
0.0571531169,
-0.0555906743,
-0.0173792988,
-0.0242911801,
0.0218670461,
-0.0106162271,
-0.0153071452,
0.000236972,
-0.0391992703,
0.0279812254,
0.000745862,
0.0167865139,
-0.0029488176,
-0.0175313763,
0.0023345666,
0.0029448362,
-0.0563471913,
-0.0011126093,
-0.0311631914,
0.0660615787,
0.1183416322,
0.0536340922,
-0.0310708415,
0.0199463684,
-0.0327790193,
0.0193571225,
0.0077017983,
0.0885675773,
-0.0266467296,
-0.0591739565,
-0.0113804517,
-0.0145222954,
-0.0032372749,
-0.0319746435,
-0.019700015,
0.0212232824,
0.0150404163,
0.0008609429,
0.0185402483,
-0.0333345719,
-0.0126829268,
-0.025569519,
-0.0237167627,
-0.0323667228,
0.0461865366,
-0.0139372461,
0.0586968176,
0.0355147794,
0.0075721941,
0.0114948228,
-0.0155064566,
-0.0108250389,
-0.0020900634,
-0.0122859655,
0.0286578033,
0.0236993805,
0.0096917357,
-0.0098503446,
0.0233811587,
-0.0134074222,
-0.0190739073,
0.0307174567,
-0.0145520354,
0.0352912806,
0.0070397719,
0.0203990713,
-0.0184970461,
-0.0132176112,
-0.0186482836,
0.0325067639,
0.0148103694,
-0.0161945205,
0.0328617543,
0.0099377753,
-0.013712517,
-0.0465379134,
-0.0054412428,
0.0197405145,
0.0041944971,
-0.0117924577,
0.0222055353,
0.0143667404,
-0.0452574492,
-0.0345636867,
-0.0957109705,
-0.0321552344,
-0.021033505,
-0.0020701573,
0.020057451,
0.064735584,
-0.0264165103,
-0.043278154,
-0.0393269472,
0.023593083,
0.0089692166,
0.008273216,
0.0447704531,
0.0003965613,
-0.0029310905,
-0.0364496782,
-0.0057396688,
-0.0102966148,
0.0260327458,
0.0232565831,
-0.0659487769,
-0.0756609589,
0.0179973859,
-0.0383061171,
-0.0056680446,
0.0192864183,
0.0107414378,
0.0238488745,
-0.0143375341,
0.0664209425,
0.01980724,
0.0089450385,
0.0283419657,
0.0401058942,
-0.0117351357,
-0.0411549769,
-0.0187970921,
0.0130988387,
-0.0005597758,
0.0296607129,
-0.0485418253,
0.0304656997,
0.0317274034,
0.0286074225,
0.0279001631,
-0.0235762149,
0.0633175597,
-0.0064790356,
0.0066500688,
-0.03131954,
-0.0597874038,
-0.007623605,
0.0637291148,
-0.0010516373,
-0.0688030273,
0.0622416027,
-0.0460632704,
-0.0279254224,
0.0052784453,
-0.0564435236,
-0.0355507359,
0.0835080296,
-0.0181094464,
0.0169673581,
0.0183608104,
-0.0274720173,
0.0085019348,
0.0051393588,
-0.0030262137,
-0.0233640745,
-0.0141205713,
-0.025300717,
-0.0338744372,
-0.0068801646,
-0.0164363738,
-0.0111172218,
-0.0895840526,
-0.0940565541,
-0.0147106815,
0.0099161686,
-0.0240239538,
0.0553711131,
-0.0035481637,
0.0312319025,
0.0485929772,
0.0066273836,
-0.0084032612,
-0.023321338,
0.0019883732,
0.0195546523,
-0.0124192275,
-0.0259666499,
-0.0240241773,
0.0206007492,
0.0288205519,
0.0510449,
-0.0347413681,
-0.013818102,
-0.0399715863,
0.0553019568,
-0.0136159528,
-0.0390706956,
0.0631135404,
0.0517923683,
-0.0092977295,
0.0073340544,
0.1263705343,
0.0433924384,
-0.0588879324,
0.0017285881,
-0.0426648073,
0.0338744819,
0.038691774,
0.0423523523,
-0.0774671137,
-0.0261535682,
0.008264387,
-0.0010900211,
0.0091294833,
-0.027154291,
-0.0164904706,
-0.0776614547,
0.0390362032,
-0.0040816902,
0.0012930708,
0.0478704683,
-0.0207681172,
0.029168291,
0.046661485,
-0.055515971,
-0.0016200653,
-0.0432753377,
-0.0051224898,
-0.0161215961,
-0.0376469716,
-0.0021942439,
0.0160581935,
-0.0383768491,
0.0869095027,
0.0495740063,
0.0171826929,
-0.0002663203,
-0.0695417002,
0.0155455777,
0.0624118708,
0.0331717692,
-0.0152794383,
0.0235371944,
0.0175271146,
-0.0003687658,
-0.0275382176,
0.031770166,
-0.0573942512,
-0.0423419513,
0.0292581655,
0.0629591942,
-0.0222418383,
-0.0196226835,
-0.0016603038,
-0.007351263,
0.0371055752,
0.0588301048,
0.0411636978,
-0.0775688887,
-0.0039126789,
-0.0437541045,
0.024615055,
0.0135944569,
-0.0232205149,
-0.0351668522,
0.0226378031,
-0.0572189465,
-0.0614985675,
-0.0167393219,
0.0675159097,
0.0181246568,
0.0487443246,
-0.029581096,
0.0371761695,
0.0217829719,
0.0117938556,
0.0146842618,
-0.0495901592,
0.0465457514,
-0.0623765402,
-0.0302889794,
-0.0176908802,
-0.0516402014,
0.0181481764,
0.0117941815,
-0.0159225464,
-0.0093361028,
-0.0014843113,
0.0301861484,
0.0258343723,
-0.0291893203,
-0.0543141849,
-0.0772313178,
0.0094762584,
0.0253226776,
-0.0616370887,
0.0062526292,
-0.0230931751,
0.0217682961,
0.0188840348,
-0.0062371003,
-0.0265787076,
-0.042765595,
0.0166062247,
-0.0350033455,
-0.0413535237,
-0.0720866919,
-0.0004677981,
0.056161996,
0.0134261511,
-0.0153235169,
0.0033522812,
-0.0403419584,
-0.0077889403,
-0.0042417883,
-0.0306101907,
-0.029677961,
-0.0044522313,
-0.0388480201,
-0.0705051199,
0.0443038531,
-0.0252250433,
-0.0382166989,
0.0182303172,
-0.065418385,
0.0700041801,
0.0104983244,
0.0413867906,
0.0421801284,
0.0109380195,
-0.0120048532,
0.0521574952,
-0.0266020838,
0.0230135862,
-0.0074081426,
0.0637938455,
-0.037431933,
-0.0536092855,
0.0248661786,
0.0741945803,
-0.0269724522,
-0.087622568,
0.0359266922,
0.0921484977,
0.0077323662,
0.0125701297,
-0.0197084639,
-0.0053807991,
0.0013217618,
-0.0049889195,
-0.0145552736,
0.0040862537,
-0.0719260126,
-0.0225238446,
0.0133705037,
-0.0141783161,
-0.0174274631,
0.0376250669,
-0.0624046884,
0.0010640449,
0.0240345374,
-0.0355388857,
-0.0498386882,
0.0099356398,
0.0253792945,
0.0384612903,
-0.0660202205,
-0.0229567904,
-0.0563432164,
-0.0451803505,
0.0019986029,
0.0285780709,
-0.033085607,
0.0128218401,
0.0595505908,
0.0533824749,
-0.0272010341,
-0.0088675907,
-0.02546007,
0.0472445004,
-0.0083625494,
-0.0332833603,
-0.057714466,
-5.323623632e-33,
-0.0051470879,
-0.0710460395,
-0.0008554525,
0.0392706841,
0.038380973,
-0.0147918705,
-0.0598029718,
0.0476179905,
0.021596089,
-0.0173770972,
-0.0477688201,
-0.0109727215,
0.0169551764,
-0.0221135523,
0.0165357143,
-0.013080244,
0.0546856411,
0.0121273817,
-0.0273041762,
0.0492347553,
0.0073916838,
0.0079462044,
-0.0100923628,
-0.0677682161,
0.0476013087,
-0.0348893926,
-0.0253148321,
0.0112658311,
0.0252717976,
0.0359742083,
0.0003174864,
0.0164370984,
0.00730102,
-0.0185750313,
-0.0025765162,
0.0867127106,
-0.0124990074,
-0.0445738249,
0.0015287991,
0.008095094,
-0.0564119928,
-0.0418714136,
0.0101234876,
-0.0370075516,
-0.0176422037,
0.0826624781,
0.0406233184,
0.0152502619,
0.0075153443,
0.0335399024,
-0.0490142293,
-0.0056405268,
0.0072791618,
-0.0223261081,
-0.0119020594,
0.025903808,
0.0242785942,
-0.0051336549,
-0.0352106057,
0.0477208123,
-0.0108198291,
-0.0155095756,
0.0455940962,
-0.017842656,
-0.0347671472,
0.0278488696,
0.0107436841,
0.0328022726,
-0.0084888861,
0.0140826199,
-0.0337112993,
0.0519100763,
0.0539552234,
0.032409668,
-0.0336097665,
-0.0395934507,
0.0405179076,
0.0485178791,
0.0478897877,
-0.0202924814,
0.0078616831,
0.011047842,
-0.0293698777,
0.0127179623,
0.0004823187,
-0.0170669444,
-0.0534423292,
-0.0197089929,
-0.0100204898,
-0.0216962397,
-0.0277999695,
-0.0424331203,
-0.0047513805,
-0.0065529323,
-0.0174221359,
-0.0031759702,
0.0096774036,
0.0432576239,
-0.0056830677,
-0.0295297727,
0.0103189955,
0.0007132919,
-0.0317307748,
-0.0075507946,
0.0288133267,
0.0368187949,
-0.0074491063,
0.0222908724,
-0.0639925674,
-0.0017317871,
-0.0286408793,
0.0480172336,
0.0297217779,
-0.040326383,
0.0226876326,
-0.0124794012,
-0.0291264132,
0.0335983709,
0.0061174561,
-0.0204752367,
-0.0339593105,
-0.0355252586,
-0.0272437446,
-0.0337568857,
-0.0247245207,
-0.0356474072,
-0.0303323995,
0.0805061385,
0.0706437826,
0.0425672978,
-0.0106503563,
-0.0367229581,
2.659e-7,
-0.0104138153,
0.031565506,
-0.0043786643,
0.0284260008,
0.0291429348,
0.030100219,
0.0113245826,
0.0417523943,
0.0000424609,
-0.0124129634,
-0.007851324,
0.0286461897,
0.0276311953,
0.0390419923,
-0.0941500589,
-0.0129343448,
-0.0706736147,
0.0346363895,
-0.0729321092,
0.0510137565,
0.0615833513,
0.0076802112,
0.0575367175,
0.0015757775,
-0.0062997164,
-0.0162394829,
-0.0090357605,
-0.0415065698,
0.0134571306,
0.0168173872,
0.0027151385,
-0.0329858214,
0.0220292658,
0.0006594502,
-0.0246293023,
-0.0459560677,
-0.0087344879,
0.0212806147,
-0.000598844,
0.0303323939,
0.027836917,
-0.0271888934,
-0.0022875227,
-0.0826609284,
0.0266053993,
-0.0049058557,
-0.0121232858,
-0.030088149,
0.0417824648,
-0.0693291128,
0.0266603269,
-0.02486494,
-0.0034499997,
-0.0090210438,
-0.0129892929,
-0.0400938727,
-0.0380767211,
-0.0168421082,
-0.0173397809,
-0.0512254722,
-0.0281262454,
-0.0188057534,
-0.0877602249,
0.0079892389,
-0.0328245163,
-0.0311412308,
0.025155222,
2.038882783e-34,
-0.0277654231,
0.0230441131,
-0.0028020239,
0.0081533836,
0.0646343157,
0.0213918667,
-0.020510219,
0.0147615988,
0.0469139479,
-0.0515064187,
-0.0354047716
] |
http://arxiv.org/abs/2402.01712v1 | http://arxiv.org/pdf/2402.01712v1 | Socially Aware Synthetic Data Generation for Suicidal Ideation Detection Using Large Language Models | 2024-01-25T18:25:05 | 2024-01-25T18:25:05 | Hamideh Ghanadian, Isar Nejadgholi, Hussein Al Osman | cs.CL | cs.CL, cs.AI, cs.LG | Suicidal ideation detection is a vital research area that holds great
potential for improving mental health support systems. However, the sensitivity
surrounding suicide-related data poses challenges in accessing large-scale,
annotated datasets necessary for training effective machine learning models. To
address this limitation, we introduce an innovative strategy that leverages the
capabilities of generative AI models, such as ChatGPT, Flan-T5, and Llama, to
create synthetic data for suicidal ideation detection. Our data generation
approach is grounded in social factors extracted from psychology literature and
aims to ensure coverage of essential information related to suicidal ideation.
In our study, we benchmarked against state-of-the-art NLP classification
models, specifically, those centered around the BERT family structures. When
trained on the real-world dataset, UMD, these conventional models tend to yield
F1-scores ranging from 0.75 to 0.87. Our synthetic data-driven method, informed
by social factors, offers consistent F1-scores of 0.82 for both models,
suggesting that the richness of topics in synthetic data can bridge the
performance gap across different model complexities. Most impressively, when we
combined a mere 30% of the UMD dataset with our synthetic data, we witnessed a
substantial increase in performance, achieving an F1-score of 0.88 on the UMD
test set. Such results underscore the cost-effectiveness and potential of our
approach in confronting major challenges in the field, such as data scarcity
and the quest for diversity in data representation. | Socially Aware Synthetic Data Generation for Suicidal Ideation Detection Using Large Language Models: Suicidal ideation detection is a vital research area that holds great
potential for improving mental health support systems. However, the sensitivity
surrounding suicide-related data poses challenges in accessing large-scale,
annotated datasets necessary for training effective machine learning models. To
address this limitation, we introduce an innovative strategy that leverages the
capabilities of generative AI models, such as ChatGPT, Flan-T5, and Llama, to
create synthetic data for suicidal ideation detection. Our data generation
approach is grounded in social factors extracted from psychology literature and
aims to ensure coverage of essential information related to suicidal ideation.
In our study, we benchmarked against state-of-the-art NLP classification
models, specifically, those centered around the BERT family structures. When
trained on the real-world dataset, UMD, these conventional models tend to yield
F1-scores ranging from 0.75 to 0.87. Our synthetic data-driven method, informed
by social factors, offers consistent F1-scores of 0.82 for both models,
suggesting that the richness of topics in synthetic data can bridge the
performance gap across different model complexities. Most impressively, when we
combined a mere 30% of the UMD dataset with our synthetic data, we witnessed a
substantial increase in performance, achieving an F1-score of 0.88 on the UMD
test set. Such results underscore the cost-effectiveness and potential of our
approach in confronting major challenges in the field, such as data scarcity
and the quest for diversity in data representation. | [
0.0692731589,
0.0498912036,
-0.0032276837,
0.0020353384,
-0.0066783968,
0.0331019722,
0.0249016304,
-0.0031949533,
0.0564090051,
-0.0347014926,
-0.0222215205,
-0.0201203395,
-0.0556092858,
0.0528292283,
-0.0300834402,
-0.0244384687,
-0.0007080734,
0.0424262509,
0.0648560226,
-0.0104967058,
0.0032918644,
0.0414256752,
0.0099809673,
0.0537323691,
-0.0029046137,
-0.0130779129,
0.0182231124,
0.0101356674,
0.019549448,
-0.060171511,
0.0032733353,
0.0760795772,
-0.0136329895,
0.0262448303,
0.0000020381,
-0.0293838084,
-0.0082951356,
-0.0061643491,
-0.0132540697,
-0.0505952314,
0.0469195694,
-0.047521349,
0.0141814277,
0.0303137023,
-0.0510345213,
-0.0230574254,
0.031416934,
0.0597924404,
0.0411306322,
0.0737852305,
-0.0349920467,
-0.0560302846,
0.0463871546,
-0.022442311,
0.0125642503,
0.043035388,
0.0510526597,
0.0220553335,
0.0229307171,
0.0084252246,
-0.0289755613,
0.0161380898,
0.0141310077,
0.0207019579,
0.0158848893,
0.0140063986,
-0.0610961132,
-0.0659145713,
0.0237543248,
0.0853802338,
-0.0056858645,
-0.0339985751,
0.0384352915,
-0.0298589692,
-0.0109153697,
-0.0343172289,
-0.0031636807,
-0.0167361517,
-0.032225024,
0.0278567616,
0.0685931519,
0.0275419354,
0.0126086017,
-0.0236901231,
0.0566465296,
-0.0137534412,
-0.0031514482,
0.0027486107,
0.0208105296,
-0.0112069696,
-0.0743317455,
-0.0413667634,
0.0166866127,
-0.0135232368,
-0.0501054861,
0.0087324176,
-0.0713380203,
-0.0577711128,
0.027808236,
-0.0646257252,
0.0642224923,
0.0475681648,
-0.0134635773,
-0.0184408668,
0.0194900557,
-0.0380385518,
0.0091717411,
0.006448057,
-0.0236432366,
0.0207534917,
-0.0714896843,
0.0264118556,
-0.035121575,
-0.011977843,
0.0213509258,
0.0164596327,
-0.063559629,
0.0105607072,
-0.0233430061,
-0.0123472614,
-0.0151081579,
0.0039459537,
-0.0294223148,
0.011651543,
0.0077585946,
-0.0234603137,
-0.0469222888,
0.0141807022,
-0.0065390198,
-0.053129103,
0.0348519608,
-0.0006975757,
-0.0127314348,
-0.0227767788,
0.0336713269,
0.0707133785,
0.0145393033,
-0.0167292319,
-0.0591700524,
-0.020335475,
-0.0258212145,
-0.0271442365,
0.0175037272,
-0.0348170958,
0.0400922261,
-0.0063547287,
-0.0246750638,
-0.0401086025,
0.0381572768,
-0.0334625095,
0.0017771528,
0.0817656517,
-0.035415113,
-0.0001481954,
0.0482464433,
0.0087816203,
-0.0134449229,
0.0927243158,
0.0033305848,
0.0299332608,
0.0323438235,
-0.0084654512,
0.034761399,
0.0045441403,
0.0183754209,
-0.0399724357,
-0.0027216377,
-0.0172693022,
-0.0677617192,
0.0380879305,
0.0123907281,
-0.00075545,
0.0043695383,
-0.0268082507,
0.035611894,
0.0492446348,
0.012718101,
0.0773520768,
0.0627984032,
0.0594719164,
-0.0272500999,
-0.0481384471,
-0.0375579037,
0.0232407749,
0.0005235236,
0.0016134413,
0.024872236,
0.0103320051,
-0.0097136293,
-0.0349930078,
-0.0345106572,
0.0081784287,
-0.0389136076,
0.0025365029,
-0.0152079752,
0.0005178762,
-0.0135579659,
0.0015239377,
0.0405002721,
-0.0098232171,
-0.0272359736,
-0.0191920064,
0.0189942978,
0.0792760104,
0.0398196429,
-0.0398882292,
-0.0262001567,
-0.0146797011,
-0.0057846359,
0.0168821998,
0.0144757181,
-0.014203717,
0.0170248579,
-0.0184303559,
-0.0044694208,
0.0390047655,
0.0143945022,
0.0062764357,
0.0209154096,
-0.0080318088,
-0.0160087291,
-0.0160606746,
-0.0482245013,
0.0410417914,
-0.0081093311,
-0.0444483794,
0.0060638334,
0.005424628,
0.0451246947,
-0.0021710698,
-0.0277763195,
-0.0089120716,
-0.0038213157,
-0.0342492312,
-0.0462055095,
-0.0174887422,
-0.0195353571,
-0.0431964286,
-0.0092713917,
-0.0597072691,
-0.0127267381,
0.05763584,
-0.0373368636,
-0.0353536792,
-0.0011731917,
-0.0640157312,
0.0176874325,
-0.0282821488,
-0.0188908726,
-0.0110674482,
0.0242381934,
0.0038978846,
0.0657392591,
-0.0196656398,
0.0249071885,
-0.0209418982,
-0.0242734682,
-0.007549169,
-0.0491718017,
-0.0399058871,
-0.0440732352,
-0.0166366976,
0.0022452276,
-0.0101234969,
0.0212306101,
-0.0301346909,
-0.0150136352,
-0.1326232702,
0.0299481526,
-0.0264793821,
-0.0097744158,
0.0071908711,
-0.0176732875,
-0.050176695,
-0.0052398746,
-0.0477846153,
0.015782686,
0.0122642992,
0.0381902121,
-0.0013578433,
-0.0020593537,
-0.0015506346,
-0.0392367765,
-0.0013711537,
-0.0271129888,
0.0318298787,
0.0033322938,
-0.0761828348,
-0.0050788615,
0.028559586,
-0.0405221544,
0.0087989233,
-0.021385381,
-0.0374793373,
0.0119894538,
0.0097311707,
0.0749495998,
0.0263539497,
0.0090468312,
0.0444199927,
0.0378147252,
0.0389713421,
-0.0256057642,
0.016163975,
-0.0431013517,
0.0075713401,
-0.0330870561,
-0.0447273962,
-0.0225623399,
0.0363789611,
0.015186917,
-0.0215681922,
-0.0631903857,
0.0085120816,
0.0109273531,
-0.0352423266,
0.0154358298,
-0.0051842411,
0.0070863883,
0.012240611,
-0.0188236684,
-0.0396959446,
0.0525950193,
0.0331031904,
0.0017038503,
0.0088715218,
-0.0168319587,
-0.0343851447,
0.0418225154,
0.0071659163,
-0.0044311574,
0.0022995209,
0.0129814483,
0.0210570209,
0.0428264514,
-0.0297779478,
-0.0258026496,
-0.0175837427,
-0.0213105436,
0.0042310064,
0.0275002141,
-0.0204917453,
-0.059925776,
-0.04596119,
-0.0450804085,
-0.0425260216,
-0.0064835032,
-0.0096128387,
0.0218192469,
-0.0399800763,
-0.0260979943,
0.0356582366,
-0.0027733287,
-0.0338450037,
-0.0075769713,
0.0033181361,
0.0477124453,
-0.0864906386,
0.0361586325,
0.0048488374,
0.0571768209,
0.0342632011,
-0.0323483571,
0.0527233668,
-0.0218777359,
0.0314060822,
-0.0191616062,
0.0249316655,
0.0006977597,
0.0011283942,
0.0380552635,
0.0040766601,
-0.0221349318,
0.0858626366,
0.0117234187,
-0.0498942807,
0.0305244103,
0.0015585,
0.0252589546,
0.0113630835,
0.0028593701,
-0.0400590152,
-0.0303787943,
-0.0325277895,
0.0358231775,
-0.0753096044,
-0.0594382398,
-0.0065763951,
-0.0774173066,
0.0803964809,
-0.0377390794,
-0.0021486704,
0.0361836962,
-0.0383948684,
0.0233756062,
-0.0314369574,
-0.0454912148,
-0.0623294823,
-0.003387786,
0.0100623313,
-0.0049079228,
-0.064805463,
-0.0542490967,
0.048712194,
-0.0043922439,
0.1278356463,
-0.0286185946,
0.064475812,
0.009225091,
-0.0406090319,
0.003555696,
0.061033912,
-0.0243677832,
-0.0134216305,
0.0879796222,
-0.0056285593,
-0.007206283,
-0.0478495061,
0.0912745818,
-0.0182362776,
-0.0308419317,
-0.0271300692,
0.0505418256,
-0.0371303819,
-0.0553602017,
-0.0305784419,
0.0590039007,
0.0152287697,
0.0822204426,
0.0349805318,
-0.0599983856,
0.001040524,
0.0049131527,
-0.0107145719,
-0.034195181,
-0.0464113131,
-0.0082979724,
-0.043859195,
-0.0513488054,
-0.0506137498,
-0.044622194,
0.0327489488,
-0.0117064463,
-0.0019018871,
-0.0431139097,
-0.0185233708,
0.0188866612,
-0.0939671844,
0.0586106442,
0.0128955338,
-0.0128153414,
0.0176153556,
0.0011852074,
0.0232800245,
-0.008963597,
-0.0341618732,
0.0516564324,
-0.065808773,
-0.0330368243,
0.0680739209,
0.0267290026,
0.0085835895,
-0.0050149546,
-0.0182815343,
-0.049268622,
0.0687845945,
0.0415994823,
0.0096655777,
-0.0222564563,
0.0119456593,
0.0349947177,
0.0158344917,
0.0377982631,
0.016826367,
0.0777639598,
0.0533039756,
-0.0105199078,
-0.0538147017,
0.0549763441,
-0.0134030776,
0.0159384366,
0.0115375593,
-0.0396788754,
-0.0225498695,
-0.0208238233,
-0.0418895558,
-0.016973488,
-0.0195500404,
-0.0344092585,
-0.0252493098,
0.0683596283,
0.0248967316,
-0.0251902547,
-0.0620860942,
-0.0497503988,
0.0012791797,
-0.0324980132,
-0.0200341735,
0.0560264662,
0.0449937619,
0.0240974091,
-0.0001547375,
-0.0371224843,
0.0610206202,
-0.0264966357,
0.0249291621,
-0.0177453458,
0.0518446304,
0.0021135646,
0.0048277606,
0.0474304929,
0.0456636995,
-0.0688021779,
-0.0823694021,
-0.0925070867,
0.0378294289,
0.0286347661,
0.0191861633,
0.0270349514,
-0.0385699272,
0.0222539026,
0.0075921118,
0.0555288903,
-0.0193075035,
-0.0661388859,
-0.0075535378,
0.0022920514,
-0.0272896457,
-0.022721095,
0.0454749577,
0.0268121008,
0.0205393918,
0.0249940809,
-0.0116859796,
-0.0320392847,
0.0004838575,
0.0299804155,
0.0366070978,
-0.0219190177,
0.013520523,
-0.0521559268,
-0.0475120731,
0.0464606546,
0.0022980876,
-0.0282510277,
0.0317363851,
0.031780269,
0.0428397283,
-0.0213747229,
-0.0302994549,
-0.0402621888,
0.0093395906,
-0.0099782525,
-0.0410287902,
-0.0414772034,
-5.735613831e-33,
0.0008081,
0.0051777884,
-0.010184464,
0.0219272394,
-0.020055566,
0.0035605412,
-0.0096351476,
0.0272922684,
0.0150745846,
0.0459276326,
0.0007660307,
0.0897030607,
-0.0014578335,
0.0213029534,
0.0149372946,
-0.0282173231,
0.0158524122,
-0.0163050685,
-0.0148559585,
0.0158081409,
0.0166029148,
0.049568221,
0.0679078847,
-0.0687389821,
0.0477092452,
0.0171236582,
0.0086660953,
0.0262415167,
0.0393274762,
0.010864988,
-0.0628346577,
0.0736717731,
0.0560610779,
-0.0041158246,
-0.0011253681,
0.0179893244,
-0.0237924997,
-0.0200340431,
0.0171014946,
-0.0190732647,
-0.0290146507,
-0.0134182516,
0.0180649813,
-0.0057238676,
-0.0060304459,
-0.0427678414,
0.0322866775,
-0.0306595247,
0.0485711209,
-0.0245339014,
-0.0171256065,
-0.0064930473,
-0.0292864256,
0.0049929856,
-0.0219002627,
0.0163733605,
-0.0039579407,
0.0272767413,
-0.0446561426,
-0.0022262025,
-0.0262105986,
-0.025431538,
-0.0081210351,
-0.0548037738,
0.0050471751,
0.0945164114,
0.0117019173,
0.0306902491,
-0.0475075357,
-0.0198311824,
0.0292853005,
0.0851179659,
0.0721101984,
0.0871256366,
0.0336935408,
-0.013653555,
-0.0210703909,
0.037403591,
0.0199404936,
-0.0481556915,
0.0164519586,
-0.0177926198,
-0.0388267264,
-0.0093267318,
-0.0018362612,
-0.0707357228,
-0.0160048418,
0.0096946498,
-0.0014552175,
-0.0267531425,
-0.0210520029,
0.0007937175,
-0.0135139078,
-0.0314331278,
0.0374940075,
-0.015153259,
0.0065042367,
0.0372681059,
0.0144557599,
-0.009147007,
-0.0246134046,
0.0142959496,
0.0430244692,
0.0041017616,
0.0451793224,
-0.036725428,
0.0523463935,
-0.0291461777,
-0.0392485373,
-0.0177635346,
-0.0153719755,
0.0486470386,
-0.0050251447,
0.0078614773,
0.0058457386,
0.0334203392,
-0.0153144626,
-0.0194134414,
0.0514221154,
0.0651841164,
-0.0357046872,
-0.0085380785,
-0.0434892066,
0.0221113507,
0.0148664312,
-0.0397218913,
-0.0288603473,
0.006989127,
0.041994933,
-0.0515440479,
-0.0104730045,
-0.0234659854,
2.684e-7,
0.0469455086,
0.041754134,
0.0018492137,
0.0114663457,
-0.0102120899,
0.0436915532,
0.0490372591,
0.0859296247,
0.0375020653,
-0.0142647848,
0.0115998657,
0.0406975746,
0.0455576181,
0.0672777295,
-0.063494876,
-0.0233370624,
-0.0389115512,
-0.0105460566,
-0.0723193437,
-0.0315799937,
0.0834202319,
0.0848206505,
0.0516906492,
-0.0510257669,
0.0345906317,
0.0095617129,
0.039300859,
-0.0638691485,
-0.0005791659,
0.033534348,
0.0338263176,
0.0127261626,
0.0120950006,
0.001906939,
0.0076523228,
-0.046970915,
0.0001607338,
0.0415264145,
-0.0142872483,
-0.024214847,
0.0119536715,
0.0225363597,
-0.0032648314,
-0.0505816936,
0.0553704798,
-0.0618409477,
-0.0147086689,
-0.0445115902,
-0.0239515621,
-0.0076528592,
0.039845787,
-0.0117416233,
-0.0064001819,
0.0378724746,
0.0458129868,
-0.0186533574,
0.0199080948,
-0.0564274304,
-0.0316369832,
0.0172116458,
-0.0239860397,
-0.0194876175,
-0.0214150827,
0.0048980289,
0.0385087952,
-0.0593854785,
0.0268989243,
2.25782205e-34,
0.0150287887,
0.049454771,
0.0320610926,
-0.0081163784,
-0.0026760518,
0.0235725306,
0.0021796499,
-0.0121536143,
-0.0154024875,
-0.0624646284,
0.011601097
] |
http://arxiv.org/abs/2401.14373v1 | http://arxiv.org/pdf/2401.14373v1 | TURNA: A Turkish Encoder-Decoder Language Model for Enhanced Understanding and Generation | 2024-01-25T18:24:13 | 2024-01-25T18:24:13 | Gökçe Uludoğan, Zeynep Yirmibeşoğlu Balal, Furkan Akkurt, Melikşah Türker, Onur Güngör, Susan Üsküdarlı | cs.CL | cs.CL, cs.AI, cs.LG | The recent advances in natural language processing have predominantly favored
well-resourced English-centric models, resulting in a significant gap with
low-resource languages. In this work, we introduce the language model TURNA,
which is developed for the low-resource language Turkish and is capable of both
natural language understanding and generation tasks. TURNA is pretrained with
an encoder-decoder architecture based on the unified framework UL2 with a
diverse corpus that we specifically curated for this purpose. We evaluated
TURNA with three generation tasks and five understanding tasks for Turkish. The
results show that TURNA outperforms several multilingual models in both
understanding and generation tasks, and competes with monolingual Turkish
models in understanding tasks. TURNA is made available at
https://huggingface.co/boun-tabi-LMG/TURNA . | TURNA: A Turkish Encoder-Decoder Language Model for Enhanced Understanding and Generation: The recent advances in natural language processing have predominantly favored
well-resourced English-centric models, resulting in a significant gap with
low-resource languages. In this work, we introduce the language model TURNA,
which is developed for the low-resource language Turkish and is capable of both
natural language understanding and generation tasks. TURNA is pretrained with
an encoder-decoder architecture based on the unified framework UL2 with a
diverse corpus that we specifically curated for this purpose. We evaluated
TURNA with three generation tasks and five understanding tasks for Turkish. The
results show that TURNA outperforms several multilingual models in both
understanding and generation tasks, and competes with monolingual Turkish
models in understanding tasks. TURNA is made available at
https://huggingface.co/boun-tabi-LMG/TURNA . | [
0.0370977744,
0.0295713451,
-0.0216751173,
0.0326137729,
0.0015444848,
0.0567095019,
0.0035857495,
-0.043038439,
-0.0502333082,
-0.0100889998,
0.0040937033,
-0.0627916604,
-0.0028191693,
0.0498149768,
-0.0020155581,
-0.0575201437,
0.0502244718,
-0.0360476859,
0.0040744101,
0.006787356,
0.0054786899,
0.0620631352,
-0.0486528352,
0.0278037824,
-0.0278163329,
-0.0022806078,
-0.03907438,
-0.0218453147,
-0.0076307906,
-0.0175962187,
-0.0017761738,
0.005083329,
0.0226006433,
0.0218335465,
0.0000020327,
-0.0368758924,
-0.0053119087,
-0.0354646333,
-0.0247323643,
-0.0317633152,
0.0612878725,
-0.0719505027,
-0.0257818922,
-0.0018344331,
-0.0352048054,
-0.0241113938,
0.0038630627,
0.0400620028,
0.0286536962,
0.0619952753,
0.0096770329,
-0.0730908215,
0.0091583319,
-0.0224600825,
0.0128242755,
-0.058747422,
-0.0406990834,
-0.0373695008,
-0.0376851782,
-0.0149109624,
-0.0133704199,
0.0218084119,
0.0390266553,
0.002236414,
-0.0343252867,
0.0420755967,
-0.0606582649,
-0.0960399285,
0.0203377903,
0.0679137185,
-0.0009713092,
-0.0033087656,
0.0152648725,
0.0169478133,
-0.0187392142,
0.0085408036,
-0.033148244,
0.0363247283,
-0.0229122899,
0.0288126934,
0.0535202809,
0.0317029841,
-0.0082280925,
-0.0069961934,
0.0069424142,
0.0487772375,
0.0247394107,
-0.0040603285,
-0.0231550466,
0.0488273315,
0.0233761575,
-0.0347139165,
0.0266792495,
0.073479414,
-0.0301986579,
0.0211519375,
-0.047160726,
-0.0231809244,
-0.0006178258,
-0.0389086194,
0.0169353373,
0.01938737,
0.0034878168,
0.0092377774,
-0.071091786,
0.0188139565,
0.0132212359,
-0.0440245755,
-0.0643391907,
-0.0356577672,
-0.0012565715,
-0.0287285075,
-0.0569765717,
0.0214158371,
-0.0690981895,
0.0167751275,
-0.0342045575,
0.0456220172,
0.030378798,
0.0129776215,
-0.0034990418,
0.0127435355,
-0.085110195,
0.0035954555,
-0.0356025398,
0.0346959122,
-0.044293575,
-0.0146142151,
-0.0174507238,
-0.0681783631,
0.0011077947,
0.0206205975,
0.0182778109,
0.0204476696,
-0.0173922852,
0.0427396633,
-0.0314368159,
-0.0390554108,
-0.0454636738,
-0.021873666,
-0.0107028428,
-0.0803712085,
-0.0053321365,
-0.0332846604,
0.0635386184,
0.0116640553,
-0.0123950411,
-0.0259209517,
0.0029623543,
0.025758978,
0.0156568978,
0.023743758,
-0.0170651339,
0.0223487634,
0.0221578665,
-0.0064998399,
0.0552932881,
0.0727708489,
0.0332598016,
0.0289696865,
0.0122386636,
0.0229122452,
-0.00209687,
0.0194307957,
-0.0005095866,
-0.0513549782,
0.0243287366,
0.0110764802,
-0.033466924,
0.0404094905,
0.0504196584,
0.0159158111,
-0.0277522746,
0.0313394256,
0.0264696553,
-0.002776521,
0.025417693,
0.0614120737,
0.1026717946,
0.0260718409,
-0.0117994742,
0.0117297973,
0.0061995885,
0.0295903292,
0.0223085061,
0.0303842425,
-0.0128594954,
0.0557252951,
-0.0553143993,
-0.045850981,
0.0197469667,
-0.0336549915,
0.0138685619,
-0.0483850203,
0.0685532838,
0.0348064899,
-0.0682041422,
0.0520779006,
0.0332677849,
-0.0638573617,
0.0759562179,
-0.0173873957,
0.0715562031,
0.0867419541,
0.0325069427,
0.0211315826,
-0.00485666,
0.0048485259,
-0.0040457975,
0.0520217791,
0.0288181473,
-0.051785551,
0.0251279511,
-0.0589087456,
0.0122586191,
-0.0365474708,
0.0244602561,
0.0417265855,
-0.0478830151,
0.0359369889,
0.0397858731,
-0.0379854627,
0.0128259268,
0.0415121205,
-0.0311929993,
0.0008470961,
-0.019803673,
0.0104876533,
-0.0248479415,
0.0344572663,
0.0096454732,
0.0027967538,
0.0290349945,
-0.0651566982,
0.0409820862,
0.0412559919,
-0.0558067001,
-0.0813721269,
-0.0081044063,
-0.0223942678,
-0.0391434319,
0.0453609675,
-0.048956424,
-0.0553130247,
0.0558838136,
-0.0283164512,
0.0488700941,
-0.0232728329,
0.005412824,
0.0217775609,
0.0125309061,
-0.0000616558,
0.0738110393,
-0.0044187233,
-0.0165177081,
-0.0134971701,
0.0097988509,
0.0282925535,
-0.0025901103,
0.0002282417,
0.0344873071,
-0.0236764271,
-0.0328647196,
0.0117952721,
0.0372012705,
-0.0177712999,
-0.0164116379,
-0.0831303522,
-0.0145640867,
-0.0412834659,
0.0361704081,
0.0497121662,
0.0129914414,
-0.0123104732,
0.0007632751,
-0.0325681753,
-0.0116872275,
0.0428468287,
-0.0196818858,
0.0354252346,
-0.0249419138,
-0.0016465957,
-0.0144106252,
0.0026386697,
0.0354750864,
0.0203620624,
0.0417684801,
-0.0070155701,
-0.0320011526,
0.0025479794,
-0.0018559309,
-0.0225792006,
-0.0075282729,
-0.025381837,
0.0238421857,
0.0533087552,
0.0613858551,
-0.0022358273,
0.0146699343,
0.0619159713,
0.0110228118,
0.0030982119,
-0.0144277168,
-0.0062198476,
-0.0495511442,
-0.0318755433,
-0.0356468782,
-0.0309768878,
0.0076458137,
0.0842647776,
-0.0092679868,
0.0020527262,
-0.1023647785,
-0.0246397089,
-0.032230258,
-0.0272175651,
-0.0076302974,
-0.0205747914,
-0.0132409604,
0.0693427026,
-0.0351870805,
0.0106849186,
0.021036502,
-0.0501564667,
-0.0039628209,
-0.0693382174,
0.0232546404,
-0.0412009172,
0.0176661089,
0.0165914055,
0.0237763226,
-0.0182969142,
-0.0202479083,
-0.0217264332,
0.0496133678,
0.0157860797,
-0.0319135934,
0.0482373051,
-0.0652372986,
-0.0075583737,
-0.0069636409,
-0.0139248101,
-0.0684475452,
-0.0141804172,
0.0132391481,
0.0083400598,
0.0128713744,
0.0246693622,
-0.0289757363,
0.0052735521,
-0.0055516693,
0.032070376,
-0.0239726771,
-0.0066271769,
-0.0325288363,
0.0464785844,
0.0642113537,
-0.016494859,
0.0046503888,
-0.0173314549,
0.0457466058,
0.0474508889,
-0.0135289449,
-0.0039375233,
-0.0294883084,
0.0213694982,
0.0135778049,
-0.0216273721,
-0.0029169086,
-0.0056659169,
0.0600200221,
-0.0299393684,
-0.0637629852,
0.093536064,
0.0152350664,
0.0202571582,
0.0227587372,
-0.017231442,
0.0308245625,
0.0317420438,
0.0273116101,
-0.0351532623,
-0.0112397084,
0.0114998994,
0.0256452151,
0.0206606723,
-0.0002898482,
0.0425209291,
-0.1272863746,
0.0316695459,
0.0038552552,
-0.1109395102,
-0.0356780924,
-0.0552833527,
-0.0386270434,
-0.0176621955,
0.0069552776,
0.0083217984,
0.0002707621,
0.0493431687,
0.053657975,
-0.0317917652,
-0.0821755677,
-0.0392902307,
-0.0145533793,
0.0197614189,
0.0455526449,
0.0312624238,
0.0067334897,
0.0328709744,
0.0259716213,
-0.0427469648,
0.0294029433,
-0.0042353626,
0.0310581066,
-0.0139762508,
-0.0205258727,
-0.0365363285,
0.0122939339,
-0.0750338733,
0.0022230479,
-0.0288877729,
0.0268887766,
0.0181360599,
-0.001101327,
-0.0173548926,
0.0086565847,
0.0069431826,
0.0578923486,
0.0132165151,
0.0000957297,
-0.0320158042,
-0.0161816329,
-0.011039692,
-0.1158848777,
-0.0132672777,
0.0248160642,
0.0290542785,
-0.0175930485,
-0.0825217515,
-0.0749168769,
0.0175240636,
-0.029810844,
-0.0254065786,
0.0138889933,
-0.0090869255,
-0.0388043486,
-0.0391731858,
0.0016506335,
0.0069045089,
-0.0224823691,
-0.0061555495,
-0.044863455,
0.0227688178,
-0.0228811558,
-0.0213412028,
0.0431161635,
-0.0675729886,
-0.0412930474,
0.0218054,
0.0672693253,
0.0153913796,
0.052173458,
0.0115759708,
0.0015530367,
0.0316789001,
-0.0397977866,
-0.0651376322,
0.0238282289,
-0.0169944111,
0.0085066408,
-0.0100558065,
0.0447955839,
-0.0165825579,
0.0311028473,
-0.0103244353,
0.0241658222,
-0.0176597945,
-0.0038613898,
0.0297292769,
-0.0188967846,
-0.0072012185,
-0.0338732153,
0.0127340341,
-0.047187075,
0.0264462512,
-0.0246923454,
0.0212351028,
-0.0225199759,
0.0118160071,
0.04232217,
-0.0047328263,
0.0084971869,
-0.0000982064,
-0.0125948926,
0.0187386218,
-0.021989448,
-0.0131106703,
0.0285846163,
0.020838039,
-0.0040218681,
0.025447607,
-0.0342111401,
0.0324534811,
-0.0597748421,
0.0318433829,
-0.0366053991,
0.0555253215,
-0.0386532843,
0.0195221733,
0.0409818925,
0.0385318436,
0.0083157001,
-0.0190869831,
-0.0257833395,
0.0086529376,
0.036054194,
-0.043905627,
0.0776304677,
-0.0026754886,
0.035858307,
0.0075591193,
0.0476458445,
-0.0323533528,
-0.0627414137,
0.0254184287,
-0.0269919708,
-0.0192516968,
-0.0308317542,
0.0788884386,
-0.0067594578,
0.0540575571,
-0.0138984304,
0.0570640787,
-0.0089453747,
-0.0196138732,
-0.0126606375,
0.0207474977,
-0.0317368917,
0.0306778103,
-0.034336444,
-0.0443403088,
-0.0124451546,
0.0153276119,
0.0327550173,
-0.0075294636,
0.0669767484,
0.053583011,
-0.0143681588,
-0.00389056,
-0.0481002219,
0.002284317,
-0.0519003496,
-0.0025737027,
-0.0357889086,
-6.102505087e-33,
-0.0495253876,
0.0331841782,
0.0143306013,
-0.040861126,
-0.0176497716,
-0.0327175707,
-0.0381483212,
0.012306992,
-0.004058498,
0.0188255459,
0.0133537985,
0.0018855931,
0.0091972593,
0.0428303815,
0.0014019076,
-0.0297385734,
0.0455577523,
-0.0169073623,
-0.0208016895,
0.0564122126,
0.0842137113,
0.0461141728,
0.0653160587,
-0.0249746256,
-0.0376407392,
-0.0457694903,
0.0062715677,
0.0030938068,
0.0152146956,
0.0514136441,
-0.0287901927,
0.0433050208,
0.0485468246,
-0.0527722314,
-0.0211300608,
0.0021099434,
-0.0954474136,
-0.0333961882,
0.0143718533,
0.045440793,
-0.0430511534,
-0.045349326,
0.0028022511,
-0.0041373917,
0.0046730642,
0.0416576788,
0.0181635208,
-0.008695242,
-0.0110110259,
-0.0215639714,
-0.1021002829,
0.0019401971,
-0.0410785861,
-0.0144185619,
0.018510459,
0.0889559835,
-0.0057747341,
-0.0389853418,
-0.0313787237,
0.0219497476,
-0.0368138365,
0.0371368229,
0.0730608106,
-0.0326593369,
-0.0289629288,
0.0510087647,
-0.0428363346,
-0.0090395994,
-0.0495140962,
0.0227248315,
-0.0312639326,
0.0259856656,
-0.0115846293,
0.0562886633,
0.0125510152,
-0.0544100627,
-0.0768782422,
0.0575819649,
0.0044476376,
0.0813068226,
0.0196235627,
-0.0376752727,
-0.0011093828,
-0.0128561864,
-0.0167902969,
-0.0199204199,
0.0081869848,
-0.0424444713,
0.0408931039,
0.0210173596,
-0.0297482703,
-0.0293237846,
0.0113717392,
-0.0062812031,
0.0363563932,
0.0041326871,
-0.0352913588,
0.0058393236,
-0.0249807127,
0.030541094,
-0.0320163108,
-0.0654160902,
0.0332416259,
-0.0058260341,
0.0123201925,
0.0059502036,
0.0107914545,
-0.0065515558,
-0.0297856536,
0.0399661027,
-0.0656505078,
0.0555898361,
-0.0070737312,
0.0525910445,
0.0304952543,
-0.0073350077,
-0.0587829985,
0.0315398648,
-0.0070345546,
-0.0177660994,
-0.03071592,
-0.0267144442,
0.01876463,
0.0145218093,
-0.0313122347,
0.0199029725,
-0.0448040068,
0.0186668485,
0.0616884343,
-0.03800736,
-0.0108236261,
0.0262474753,
2.669e-7,
0.0296093766,
0.0075975466,
0.0452825353,
-0.004288658,
0.0537003875,
-0.0066047087,
0.0036446152,
0.0282340124,
-0.004556166,
0.0141413463,
0.014010096,
0.027792681,
0.050593257,
0.0033411891,
-0.0399234742,
-0.0059003723,
-0.0222894046,
0.022074284,
-0.0433865599,
0.0105050746,
0.064145796,
0.083205156,
-0.0428528003,
-0.0166880302,
-0.0123098614,
0.0266372543,
0.0845780969,
-0.0077124671,
-0.0202817265,
0.0132423155,
-0.0037023723,
0.0329774618,
-0.0205174517,
0.0006190999,
-0.0307619143,
0.0178337973,
0.0130219506,
0.0823366567,
-0.000614921,
0.0036521694,
0.0783183649,
0.015933251,
-0.0339657962,
-0.0739462525,
0.0954799801,
-0.045117788,
-0.0475895591,
-0.0131773297,
-0.0145831294,
0.0420230515,
-0.0403563678,
-0.0032786767,
-0.018839743,
0.0531608239,
0.0284594316,
-0.006263304,
-0.0188329928,
-0.0657340661,
0.0145714767,
0.0062181763,
-0.0308301225,
0.0105950758,
-0.0530548133,
-0.0035077299,
0.0363122411,
0.0102823535,
-0.0449614115,
1.617442919e-34,
0.001197623,
0.0227327254,
0.0107881762,
0.0264212508,
0.0093630375,
-0.0226610433,
0.0004683479,
0.0264909193,
-0.0080241151,
-0.031012496,
-0.0103827398
] |
http://arxiv.org/abs/2401.14367v1 | http://arxiv.org/pdf/2401.14367v1 | Genie: Achieving Human Parity in Content-Grounded Datasets Generation | 2024-01-25T18:14:57 | 2024-01-25T18:14:57 | Asaf Yehudai, Boaz Carmeli, Yosi Mass, Ofir Arviv, Nathaniel Mills, Assaf Toledo, Eyal Shnarch, Leshem Choshen | cs.CL | cs.CL, cs.AI, cs.LG | The lack of high-quality data for content-grounded generation tasks has been
identified as a major obstacle to advancing these tasks. To address this gap,
we propose Genie, a novel method for automatically generating high-quality
content-grounded data. It consists of three stages: (a) Content Preparation,
(b) Generation: creating task-specific examples from the content (e.g.,
question-answer pairs or summaries). (c) Filtering mechanism aiming to ensure
the quality and faithfulness of the generated data. We showcase this
methodology by generating three large-scale synthetic data, making wishes, for
Long-Form Question-Answering (LFQA), summarization, and information extraction.
In a human evaluation, our generated data was found to be natural and of high
quality. Furthermore, we compare models trained on our data with models trained
on human-written data -- ELI5 and ASQA for LFQA and CNN-DailyMail for
Summarization. We show that our models are on par with or outperforming models
trained on human-generated data and consistently outperforming them in
faithfulness. Finally, we applied our method to create LFQA data within the
medical domain and compared a model trained on it with models trained on other
domains. | Genie: Achieving Human Parity in Content-Grounded Datasets Generation: The lack of high-quality data for content-grounded generation tasks has been
identified as a major obstacle to advancing these tasks. To address this gap,
we propose Genie, a novel method for automatically generating high-quality
content-grounded data. It consists of three stages: (a) Content Preparation,
(b) Generation: creating task-specific examples from the content (e.g.,
question-answer pairs or summaries). (c) Filtering mechanism aiming to ensure
the quality and faithfulness of the generated data. We showcase this
methodology by generating three large-scale synthetic data, making wishes, for
Long-Form Question-Answering (LFQA), summarization, and information extraction.
In a human evaluation, our generated data was found to be natural and of high
quality. Furthermore, we compare models trained on our data with models trained
on human-written data -- ELI5 and ASQA for LFQA and CNN-DailyMail for
Summarization. We show that our models are on par with or outperforming models
trained on human-generated data and consistently outperforming them in
faithfulness. Finally, we applied our method to create LFQA data within the
medical domain and compared a model trained on it with models trained on other
domains. | [
0.0615904853,
0.0484736972,
0.0087579349,
-0.0018623676,
-0.0035097189,
0.0057446728,
-0.0142412772,
-0.0219538156,
-0.0031975252,
-0.0151890172,
0.0079772286,
-0.0202545691,
-0.0071552214,
0.0401669927,
0.0192428697,
-0.0438971817,
0.0154880071,
0.0366776474,
-0.0367822573,
0.0031821916,
0.0043553449,
0.0345026441,
0.0124304686,
0.0268405024,
-0.0171053186,
-0.0102978237,
-0.0044183615,
0.009478651,
0.0281186942,
-0.0229672696,
0.0346261822,
0.0634523407,
0.0076595722,
-0.0162676573,
0.0000019426,
-0.0186357833,
-0.0081788953,
-0.0110658715,
-0.031084463,
-0.0334844217,
-0.0122395977,
-0.0057298262,
0.020752253,
0.030552702,
-0.0603761077,
0.0308901779,
0.0059930077,
0.0662884116,
0.1075103804,
0.0691396222,
-0.0115364566,
-0.034485966,
0.0067199212,
-0.0045917532,
0.0146867959,
0.0207968056,
-0.0116743902,
0.0404514372,
0.0333375856,
0.0252819825,
0.0183341168,
0.0179485902,
0.0422580987,
-0.0319873653,
0.0404770374,
0.0416843519,
-0.031625215,
-0.021736471,
0.0215255935,
0.1206506565,
0.0108781662,
-0.0320376232,
0.0153380875,
-0.0079561137,
0.0250915717,
-0.0208894052,
-0.0531027503,
0.0402620547,
-0.0273890793,
0.0231679808,
-0.0143655734,
0.0132986279,
-0.0143200438,
0.0152207986,
0.0135884117,
-0.0197854564,
0.0006673666,
-0.0232632365,
-0.0244132094,
0.020280676,
0.0022301252,
-0.0922031179,
0.0057988204,
0.0476991795,
-0.0125399912,
-0.010305576,
-0.0679150373,
-0.0487383418,
0.0320011936,
-0.0208924338,
-0.0050733062,
0.0069098892,
-0.0117828203,
0.0363189019,
-0.0303691458,
0.0175107196,
-0.0119274054,
0.0120476531,
-0.0577112883,
0.0115210097,
-0.0446275286,
-0.0325938351,
-0.0443212613,
0.0619933344,
-0.0901452526,
-0.0136505114,
-0.0190329757,
0.0165211949,
-0.0022081649,
-0.0075815106,
0.0329449363,
0.0673842058,
-0.0172592718,
0.0237451848,
-0.0166885518,
-0.0012304642,
-0.0921728387,
0.0125010563,
0.0397752598,
-0.0464188419,
0.0119520985,
0.0174063984,
0.0379567258,
-0.0118545536,
0.0190767776,
0.0111568384,
-0.0402351096,
0.0080719572,
-0.03750287,
-0.0688924193,
-0.0009950941,
-0.038529858,
0.0161233395,
-0.0655019134,
0.0381413735,
-0.0091091879,
-0.0315471403,
-0.0419969521,
0.0140868165,
0.0549314991,
-0.0200362522,
0.0482988209,
-0.0447365046,
0.0007635155,
0.0397187993,
0.0149385296,
0.0308785997,
0.0583788343,
0.0330742598,
0.0196356457,
0.0324685909,
0.0014080157,
0.0082748793,
-0.025168933,
0.0451947674,
-0.0060602482,
-0.0058903093,
0.0466067903,
-0.0169271324,
0.0341212004,
0.0124285799,
0.0783740431,
-0.0137465121,
0.0028385578,
0.0526382551,
0.028571872,
-0.0061254529,
0.0437538698,
0.0388342924,
0.0064536985,
0.0255864859,
0.0478086621,
-0.0316107124,
0.067401357,
0.018727595,
0.022518333,
0.0221868958,
0.0334808044,
-0.0440541431,
-0.0407522619,
-0.0121673215,
-0.0029893352,
0.0121144913,
-0.0272485241,
-0.0083917323,
0.0427812524,
-0.0321715772,
0.0352243073,
0.0042065466,
-0.07764294,
0.0334844887,
-0.0280115362,
0.0554105118,
0.0430000536,
0.0325903483,
-0.0106738871,
-0.0321133398,
-0.0287037063,
0.0533861704,
0.0537256822,
0.0162124578,
-0.0342134014,
-0.0171649531,
-0.0382173024,
-0.0260629021,
-0.0199375283,
-0.0266808532,
-0.0155480597,
0.0301574096,
0.0170566011,
0.0239534937,
-0.0323419571,
-0.0234473795,
0.0448564701,
-0.0504535995,
-0.0110358186,
0.0282356832,
-0.0073558595,
0.0060488922,
0.0734786466,
-0.0118110068,
0.0255589616,
0.0349821001,
-0.0462823883,
-0.048634503,
0.0216697305,
-0.0543064885,
-0.0294207856,
0.0026459473,
-0.0930543616,
-0.0160874967,
0.0465004072,
-0.0370121151,
-0.0471468642,
-0.0347899981,
-0.041027464,
-0.0036653162,
-0.0524106584,
-0.0349473059,
0.0565840602,
0.0095620025,
-0.0292844605,
0.0686094537,
-0.0000965989,
-0.0041857269,
0.0259843525,
-0.0146922572,
-0.0024550641,
-0.0388713256,
0.0304483082,
0.0150744515,
-0.0174366869,
-0.0002975314,
0.0192402937,
0.0318242498,
-0.0648707971,
-0.003771926,
-0.1019162387,
-0.0064124223,
-0.0210504234,
0.0037173452,
0.0551345646,
0.020559473,
-0.0201267656,
-0.0271238387,
-0.1139359474,
-0.0259007365,
0.0301396288,
0.0117434757,
0.0548159517,
0.0079797599,
0.0514962114,
-0.0316722617,
-0.0021437968,
-0.0192464124,
0.0271751136,
0.0318355523,
0.0055462895,
-0.036130935,
0.0086521022,
-0.0493279174,
-0.0013773509,
0.0361665115,
-0.0372264124,
0.014811622,
-0.0008264261,
0.0429489538,
0.0621395744,
0.0574115179,
0.0073065003,
0.0019441386,
0.0451587513,
-0.0149000147,
0.0180698149,
-0.0431641154,
0.0432047546,
0.0244027209,
-0.017331887,
0.0367418341,
0.0353743397,
0.0361413099,
0.017070936,
-0.0590954386,
0.0583482869,
-0.0048995758,
-0.0153493015,
-0.0441093296,
0.0162387863,
-0.031554319,
0.0382945314,
-0.044783704,
-0.0079932967,
0.0843062699,
-0.0096961716,
0.0066063064,
-0.0441671312,
0.0159902219,
-0.0389312431,
0.0349039063,
-0.0352175869,
0.0292796828,
-0.0177636016,
-0.0028865386,
-0.0535249338,
-0.0028070405,
0.0195466969,
-0.0399460718,
0.0458470732,
-0.0345502123,
0.0084233321,
-0.02314418,
-0.045314908,
-0.0666231662,
-0.0333297141,
-0.0073658004,
0.0018434478,
0.0025720487,
-0.0012621229,
0.008081519,
-0.0193360448,
-0.008535332,
0.059606418,
-0.0345351323,
-0.004753687,
-0.023192035,
0.0659070835,
0.0932031125,
0.0310271904,
0.0244943798,
-0.0164027698,
0.0692001581,
0.0572389178,
0.0080685709,
0.0320249051,
0.0337445661,
0.0001790095,
-0.027457295,
-0.0283164084,
0.0226008352,
0.0037237706,
0.0770627633,
-0.0309461989,
-0.0268434528,
0.0950682387,
0.0313683748,
-0.0149374492,
0.0151173528,
-0.0112400781,
0.0025419155,
0.0084034298,
0.0114272265,
-0.0841759294,
-0.0244482197,
-0.0081380261,
0.0396633223,
0.0056402651,
-0.0503556319,
0.0078684259,
-0.1097940281,
0.0434149057,
-0.0291199144,
-0.054463584,
-0.0658043027,
-0.0396208912,
-0.0018717867,
0.0114341117,
-0.0082562519,
0.0107189072,
-0.0105456905,
0.0184049122,
0.0181611571,
-0.0073359283,
-0.0391135626,
-0.0054236809,
-0.0232026726,
0.0447017774,
0.0528609641,
0.044385206,
0.0020798626,
-0.0378092229,
0.0015402066,
-0.0304798409,
-0.0080062626,
0.0021931478,
0.0490601994,
0.0014908756,
-0.010043147,
-0.0387539752,
0.0946229622,
-0.0496703498,
-0.0201140009,
0.0090728346,
0.0314670093,
-0.0274171997,
-0.0084277149,
-0.0555550791,
0.0320542268,
0.031776946,
0.1160356924,
0.0638504103,
-0.0429384224,
-0.0137394266,
0.0096886856,
0.0041904501,
-0.0809324309,
-0.0270432029,
0.0381983444,
-0.0133071765,
0.0622472018,
-0.1317484528,
-0.0633412451,
0.0492835119,
0.0236655101,
-0.0176874362,
-0.0249728262,
-0.0131564047,
0.0138448365,
-0.0547782145,
0.0089378729,
-0.0005928884,
0.0085326349,
-0.0637136176,
-0.0505158827,
0.010227588,
-0.061342679,
-0.0286101606,
-0.0075335358,
-0.0577667691,
-0.0064253528,
0.0069119339,
0.0460737683,
0.0175044797,
0.0165846013,
0.0130522549,
0.0221539233,
0.0247227624,
0.0735663474,
-0.0426084809,
-0.0067195217,
-0.0067345542,
0.0166419912,
0.0058854711,
0.0249710213,
-0.0413636267,
0.0345751382,
0.0010559001,
0.0418776348,
0.0039544082,
-0.0620074198,
0.0209902562,
0.0290743746,
-0.0182255059,
-0.0150422025,
-0.0059161033,
-0.0512781106,
-0.0039313314,
-0.0481917486,
-0.0128303422,
-0.0361117572,
0.0283132661,
-0.0149305789,
0.0054541882,
0.0108585274,
0.0388032869,
-0.0500094593,
0.0455000773,
-0.0784752145,
-0.0112948539,
-0.0396106951,
-0.0005797032,
-0.0036633781,
-0.0174579732,
-0.035729479,
-0.0130867884,
-0.0175312683,
0.0381299295,
-0.008305232,
0.0342422463,
0.0207010526,
-0.0122325346,
0.0080526127,
0.0520001128,
-0.0840496421,
-0.0135974167,
-0.0736850947,
-0.0046103662,
0.0142024681,
-0.0101854447,
0.0166067444,
0.0049283244,
-0.0155776665,
0.0092246877,
0.0290122163,
0.0151479021,
0.0024502808,
-0.0069819666,
0.0165052339,
-0.0161970668,
-0.0468464233,
0.0803943425,
-0.0202062298,
0.0411047563,
-0.0090799797,
-0.0054534655,
-0.0430019647,
-0.0198510326,
-0.0036563745,
-0.010540315,
-0.0042312653,
-0.0025119733,
-0.0532804169,
-0.0573990941,
0.0209050067,
0.0385947376,
0.0064862329,
0.0224091038,
0.066543743,
0.0316354446,
0.0059047076,
-0.0073277652,
-0.0432930812,
0.0413581952,
-0.027715914,
-0.0268427711,
0.0035898406,
-6.328295143e-33,
-0.0142936558,
-0.024353642,
0.0034684255,
-0.0025882677,
-0.0263874363,
0.0306031369,
-0.0494498909,
-0.0053229048,
0.0024274015,
0.0100913262,
-0.0178033337,
0.0003846951,
0.0045060907,
0.0565851294,
-0.0126630003,
-0.0325304717,
0.0464376695,
-0.0165299326,
-0.0287724733,
0.0629595146,
0.0206987355,
0.0195342712,
0.0176639892,
-0.0444517843,
0.0150518334,
-0.0293264389,
-0.0005668515,
0.0306909885,
0.029807834,
0.0721633956,
-0.0000727238,
0.0174466856,
0.0514400303,
-0.0530645028,
-0.0178668667,
0.0459500775,
-0.1047308147,
-0.0366906188,
-0.0058045485,
-0.0464345515,
-0.0827185139,
-0.0264238901,
0.0413909033,
-0.029472176,
0.0369691029,
-0.0247627888,
0.0450342782,
-0.029228434,
0.0372352041,
-0.0772417784,
-0.0322995894,
-0.0041017109,
-0.0096052215,
-0.0224142037,
-0.0661848336,
0.0459539071,
0.0002727181,
0.0251959655,
-0.0406151041,
0.0142064467,
-0.0086197723,
0.0259274356,
0.0370856859,
-0.0038597595,
0.0031093066,
0.0717888176,
-0.0091264267,
-0.0084123611,
-0.0890991017,
0.0133263608,
-0.0179929882,
0.0200200956,
0.0623829551,
0.0796959996,
0.0237467065,
-0.0054094368,
-0.0708558038,
0.0742175952,
-0.0348371267,
0.0125203533,
-0.0189342964,
-0.0425383151,
0.0035603847,
0.0217476543,
-0.0125151062,
0.0077569913,
-0.007616756,
-0.0170734618,
0.0201502144,
0.0143811032,
-0.0734652281,
-0.0178853255,
0.0080853831,
-0.0146568166,
0.0474818572,
0.0318677202,
-0.0092700729,
-0.0051536602,
0.0178967845,
-0.0169911589,
-0.0203149822,
0.0162997711,
0.0462849364,
0.0156423617,
0.0207521692,
0.0382274538,
0.041363094,
-0.0043804403,
-0.0427170396,
0.0093531115,
-0.0207798947,
0.0630433261,
-0.0027514792,
0.0112166479,
0.0036024456,
0.0135476748,
-0.0159610379,
0.0387585089,
-0.0047566337,
0.0232585203,
-0.0466568284,
-0.0338229351,
0.0341720693,
0.0288944207,
-0.0321952179,
0.0002769318,
-0.046755258,
0.0290634297,
0.0384537168,
-0.0734113008,
-0.0046012285,
0.0136406524,
2.543e-7,
0.0327546,
-0.0144397607,
0.0662598908,
0.0127410078,
0.0068333237,
0.0038882466,
0.0359941944,
0.0338587947,
-0.0155233424,
-0.0439606719,
0.016246587,
0.0423044115,
0.0399052314,
0.04486534,
-0.1109113917,
0.0051152585,
-0.1081916317,
-0.0494062789,
-0.0279525016,
-0.0213470273,
0.0789512023,
0.0420313515,
-0.0130595965,
-0.0310260542,
-0.0116331037,
0.0141585348,
0.0155548435,
0.0131743522,
0.0035409916,
0.003926625,
0.0324705057,
-0.0095166815,
0.0183615107,
-0.0064004944,
-0.030088963,
-0.0294263307,
0.0362780951,
0.0308719482,
0.0054450389,
-0.0241012741,
-0.0075071757,
-0.0283348821,
0.0420504324,
-0.0362632275,
0.0676687434,
-0.039797809,
-0.0047338442,
-0.0326942503,
-0.037660934,
-0.0090386635,
-0.0027843663,
-0.0326957256,
-0.0161434542,
0.0421812348,
0.0202278439,
0.0174618047,
0.0346874297,
-0.0520316809,
-0.0318264067,
-0.0228546429,
-0.0587402955,
-0.0266794953,
-0.0124861673,
-0.0444663987,
0.0015045694,
-0.0491803437,
0.003726095,
2.267242765e-34,
0.0215083491,
0.029044684,
0.0069490038,
-0.00280191,
0.0093498323,
0.0005540475,
0.0160125773,
-0.0195299573,
0.0154104149,
-0.044279255,
-0.0264830049
] |
http://arxiv.org/abs/2401.14360v2 | http://arxiv.org/pdf/2401.14360v2 | A Comparative Analysis of Noise Reduction Methods in Sentiment Analysis on Noisy Bangla Texts | 2024-01-25T18:06:19 | 2024-01-29T19:22:20 | Kazi Toufique Elahi, Tasnuva Binte Rahman, Shakil Shahriar, Samir Sarker, Md. Tanvir Rouf Shawon, G. M. Shahariar | cs.CL | cs.CL, 68T50 (Primary), I.2.7 | While Bangla is considered a language with limited resources, sentiment
analysis has been a subject of extensive research in the literature.
Nevertheless, there is a scarcity of exploration into sentiment analysis
specifically in the realm of noisy Bangla texts. In this paper, we introduce a
dataset (NC-SentNoB) that we annotated manually to identify ten different types
of noise found in a pre-existing sentiment analysis dataset comprising of
around 15K noisy Bangla texts. At first, given an input noisy text, we identify
the noise type, addressing this as a multi-label classification task. Then, we
introduce baseline noise reduction methods to alleviate noise prior to
conducting sentiment analysis. Finally, we assess the performance of fine-tuned
sentiment analysis models with both noisy and noise-reduced texts to make
comparisons. The experimental findings indicate that the noise reduction
methods utilized are not satisfactory, highlighting the need for more suitable
noise reduction methods in future research endeavors. We have made the
implementation and dataset presented in this paper publicly available at
https://github.com/ktoufiquee/A-Comparative-Analysis-of-Noise-Reduction-Methods-in-Sentiment-Analysis-on-Noisy-Bangla-Texts | A Comparative Analysis of Noise Reduction Methods in Sentiment Analysis on Noisy Bangla Texts: While Bangla is considered a language with limited resources, sentiment
analysis has been a subject of extensive research in the literature.
Nevertheless, there is a scarcity of exploration into sentiment analysis
specifically in the realm of noisy Bangla texts. In this paper, we introduce a
dataset (NC-SentNoB) that we annotated manually to identify ten different types
of noise found in a pre-existing sentiment analysis dataset comprising of
around 15K noisy Bangla texts. At first, given an input noisy text, we identify
the noise type, addressing this as a multi-label classification task. Then, we
introduce baseline noise reduction methods to alleviate noise prior to
conducting sentiment analysis. Finally, we assess the performance of fine-tuned
sentiment analysis models with both noisy and noise-reduced texts to make
comparisons. The experimental findings indicate that the noise reduction
methods utilized are not satisfactory, highlighting the need for more suitable
noise reduction methods in future research endeavors. We have made the
implementation and dataset presented in this paper publicly available at
https://github.com/ktoufiquee/A-Comparative-Analysis-of-Noise-Reduction-Methods-in-Sentiment-Analysis-on-Noisy-Bangla-Texts | [
0.0492431931,
0.0563232154,
-0.0204204787,
0.0472175814,
-0.0275985356,
0.0087914532,
-0.0228265245,
0.0190310702,
0.0090508573,
-0.011721136,
-0.0253180377,
-0.0168009959,
0.0161878802,
-0.0221088082,
-0.0394921526,
-0.0475040972,
0.0174382515,
0.0199589897,
0.108329393,
-0.0064626518,
0.0168049186,
-0.0009404767,
0.0219642241,
-0.0004424267,
-0.0417375639,
0.0079136789,
0.0393246859,
0.0030082553,
-0.0316199251,
-0.0254168622,
0.0314513631,
0.0693138763,
-0.0227046683,
0.0054540765,
0.0000018627,
-0.0444094017,
0.0144466814,
0.0142816557,
0.0273135807,
-0.0124040842,
0.0810115188,
0.0171332918,
-0.0056845378,
-0.0015243801,
-0.0111036673,
-0.0148903588,
0.0591324568,
0.0674351081,
0.0271677803,
0.0679435879,
-0.0291935429,
-0.0680482164,
-0.0279481113,
0.0252530538,
0.0469131507,
-0.0185497496,
0.0358216166,
-0.0274123177,
-0.0215539355,
-0.0034228186,
-0.0166727751,
-0.0011200246,
-0.0417360477,
-0.0096790818,
0.031044716,
-0.0219038408,
-0.0095970267,
-0.0516173728,
0.0315401778,
0.0110583315,
-0.018747868,
-0.0295077786,
-0.0031429736,
0.0185271911,
-0.0072202813,
-0.0631324649,
-0.0167283006,
-0.010546647,
-0.0156346168,
0.0396723747,
0.0014904736,
0.0403466113,
0.0102205276,
0.0058204345,
0.0442969538,
0.0201436728,
-0.0412171036,
-0.0597667918,
0.0196219552,
0.0364150293,
0.0237242654,
0.0192001294,
0.0227213465,
-0.0019676241,
0.0068670344,
0.0263285898,
-0.0195928812,
-0.0500785448,
0.001665448,
-0.0875108689,
-0.0406393483,
-0.0194450971,
-0.0269430261,
0.0276479144,
0.0156896841,
-0.0463055633,
0.0439628363,
-0.0178782921,
0.0157033261,
-0.0031560212,
-0.0057579167,
0.006269752,
-0.0198360104,
0.0513604023,
-0.0206025895,
0.0354976654,
-0.0682303458,
-0.0029400063,
-0.036853496,
-0.0140603678,
0.0032886048,
-0.0377299674,
0.0254586749,
-0.0138903912,
0.0107797189,
0.0010684062,
-0.0291777123,
0.0534905717,
-0.0546237752,
-0.0442415215,
0.0406936146,
0.0336937346,
0.0078963302,
-0.0130395386,
0.0058353064,
0.0473545268,
0.030885987,
-0.0244779084,
-0.0706409886,
0.0040015355,
-0.0291258153,
-0.0515021011,
-0.0732818991,
-0.0549509004,
0.0157589242,
-0.0146797514,
-0.0065306826,
-0.0169993211,
0.0070981081,
0.0023599623,
-0.0520179309,
0.0838422924,
-0.0589389354,
-0.0156816151,
0.0271707326,
0.0103740497,
0.0643114001,
0.0174407754,
-0.0139428442,
0.0150012188,
0.0424505286,
-0.0228930879,
0.0074790027,
0.0110681076,
-0.0019154109,
-0.0215566047,
-0.0259433519,
-0.0050906632,
-0.0563784428,
-0.0351845548,
-0.0238616951,
0.0251360126,
0.0108711654,
0.0294309054,
0.0512211137,
-0.0040230742,
-0.0216989573,
0.0542877652,
-0.0612165853,
0.0179417469,
0.0394759476,
0.0237490106,
0.0092618698,
-0.0093075838,
-0.0730186552,
-0.0289155524,
0.0529997349,
-0.0164234471,
-0.0183014534,
-0.0696141124,
-0.0246575214,
-0.0309709217,
-0.0012464216,
-0.031031331,
0.0809875801,
0.0134618124,
-0.030803008,
-0.0381003059,
0.0805526376,
-0.0608674549,
-0.018165838,
-0.0404889584,
0.0755135715,
-0.0071419217,
-0.0074692247,
-0.0225204546,
0.0024215837,
-0.0352170542,
-0.0025267731,
0.022187775,
0.0061335289,
0.0219037849,
-0.0049116719,
-0.0514066629,
0.0240053702,
0.0570840687,
0.009128266,
-0.0270340592,
-0.0220197532,
0.0240599457,
-0.0356557444,
0.0035461595,
-0.0668812022,
0.0025775947,
0.0750556439,
-0.0027726488,
0.0541659966,
-0.0033192632,
0.0658011362,
0.0112758074,
-0.0345291793,
-0.0510317534,
0.0230705626,
0.0109390579,
0.0039885594,
0.0349779353,
0.0274528768,
-0.0256969295,
0.0152694806,
-0.0244845115,
0.0016655908,
0.0067153126,
-0.0238848235,
-0.001027653,
0.0315484516,
-0.0159471054,
-0.0275775474,
-0.011740163,
0.0160110127,
-0.0683547482,
0.0279247854,
-0.0135035971,
0.0390696749,
0.0154890418,
0.003738493,
-0.0069553629,
0.000212693,
0.0487083383,
0.0026496488,
-0.0537623055,
0.0017264525,
-0.0007175748,
0.0395079404,
0.0325014852,
0.0481809415,
-0.0104522156,
-0.0393232554,
-0.0576330423,
-0.0411641859,
-0.0019993419,
-0.0197320152,
0.0330224782,
0.0333860628,
0.002016532,
-0.0046306835,
0.0017702514,
0.0283318032,
0.0133395242,
0.0054922379,
0.0539378673,
-0.0190041717,
0.0078731989,
-0.013885323,
0.0017204382,
0.001498956,
0.0071549402,
0.035249386,
-0.0893958956,
-0.0287874471,
0.0320904441,
-0.00833946,
0.0044806693,
0.0250710547,
-0.0623543784,
-0.0147813233,
-0.0160319861,
0.0022368608,
0.0002230824,
0.0081287939,
0.0412553437,
-0.0045624408,
0.001424121,
0.0209220126,
0.0215577651,
-0.0380997285,
0.027382629,
-0.0318408236,
-0.0403583273,
-0.0033257741,
0.0454529561,
0.0198700316,
0.0075301924,
-0.0211551264,
-0.0023479599,
-0.0306665115,
-0.0292916186,
-0.0578996614,
0.0362574272,
-0.0397585332,
0.0477388836,
0.0178710483,
-0.0149442414,
0.0494033433,
0.0191570688,
0.0109138675,
-0.0911979601,
0.0096164765,
0.0099756252,
-0.0036938898,
-0.031157285,
0.0242072511,
0.0035097611,
-0.0115078492,
-0.0063762479,
-0.0504904687,
0.026664298,
-0.0203312263,
-0.0319453105,
-0.0080830427,
0.0293147489,
0.0145391347,
-0.0192244556,
0.0221500136,
-0.0416388288,
-0.0029762965,
-0.0374807604,
0.0217111502,
-0.026470039,
0.0291771255,
0.0432491414,
-0.022386048,
-0.0584167466,
0.0123686036,
-0.0107929353,
0.0369167402,
0.0427119844,
0.0182912983,
-0.017670976,
0.0667565539,
0.0107926922,
-0.0439127572,
-0.0032666111,
-0.0228538234,
0.0041062501,
-0.0035225335,
-0.0349673703,
0.0104709277,
0.0213833712,
-0.007155932,
-0.0118238628,
0.0329422355,
-0.0099962968,
-0.0158438459,
0.0924873427,
0.0134047987,
-0.0409745947,
-0.0191045087,
-0.0240467433,
0.0085884454,
-0.0238393899,
0.0135699874,
-0.0005056713,
-0.0284198932,
-0.0087982444,
0.0144080687,
-0.0113860872,
-0.0015331976,
-0.0091735274,
-0.0825924277,
0.0228556897,
-0.0082006445,
-0.046640221,
-0.0621159934,
-0.0430350602,
0.0119453445,
-0.0244559962,
0.0389770232,
-0.0398249105,
0.0305858776,
-0.0048193815,
0.0402566977,
-0.0537197255,
-0.0611326061,
-0.0249434635,
0.0268831495,
0.0557604656,
0.0118545368,
0.0321078673,
-0.0640149638,
0.00394622,
0.0343554392,
0.0110034188,
-0.0198168829,
-0.0278107077,
0.0642651245,
0.0250713825,
-0.054856956,
-0.019496724,
-0.0303737838,
-0.0402353927,
0.0542319529,
-0.003997698,
0.0297295582,
0.0290187355,
-0.0465571247,
-0.0329178125,
0.042586945,
0.0411538966,
0.0303656179,
0.0404233299,
-0.073352553,
-0.0291470382,
-0.0150188636,
-0.0199347045,
-0.0221598148,
-0.0387562886,
0.0177216101,
-0.0480172448,
-0.120790422,
-0.0253275763,
-0.0915779173,
-0.0107556218,
0.0286342427,
-0.0080101136,
0.0710266829,
0.0112926811,
0.0016868585,
0.0017112187,
-0.0296224058,
-0.0070275343,
-0.0014669012,
-0.0232679714,
-0.0088509033,
0.0284405742,
-0.0004154028,
-0.0176963136,
-0.0502452813,
-0.0646221489,
-0.0348587818,
0.0380183272,
0.0342061892,
0.0127140516,
-0.0060663563,
-0.0403931811,
-0.100613676,
0.0348474719,
0.0925594345,
-0.0371040031,
0.0155992219,
0.0689716935,
0.0093536945,
0.007119366,
0.0511361398,
0.0072779837,
0.0596548542,
0.0020414817,
-0.0700581446,
-0.0545796379,
-0.0240561254,
-0.004516941,
0.041869238,
0.0430903807,
0.0002224818,
-0.084992297,
-0.0045016082,
-0.0078060851,
0.0464333594,
-0.0510455817,
-0.0073829489,
-0.0400308743,
0.0368618257,
0.0447541922,
-0.0586499795,
-0.0036189323,
-0.0176779944,
-0.0043008942,
0.00531402,
0.0718868375,
0.055501774,
0.07032381,
0.0161438528,
0.0328858681,
-0.0006149434,
0.0071583311,
0.0183277614,
0.0018966378,
0.0149978278,
-0.0312295668,
-0.0014642738,
0.0255906228,
0.0310583655,
0.011270687,
0.0295182317,
0.0327172428,
-0.0200964008,
-0.0040407539,
0.0077487491,
0.0038461352,
0.0702499971,
-0.0038376828,
-0.0004264561,
-0.0407470204,
0.027666904,
-0.0188226271,
0.0466799363,
0.0268332139,
-0.0357299261,
-0.0272188019,
-0.007055291,
0.059718594,
-0.0538758747,
0.0064622816,
0.0423105471,
-0.0102932882,
-0.0486095138,
0.0012141189,
0.0019654119,
0.0205810722,
0.0440009236,
0.0358369313,
-0.0264187381,
-0.031612996,
0.0207328238,
-0.0836793929,
0.0525211729,
-0.0045686206,
0.010476484,
0.0364820063,
0.0393439382,
0.0990487635,
-0.0371402279,
-0.0298728161,
0.0042122025,
-0.0259030312,
0.015899457,
-5.694436263e-33,
0.0052830218,
-0.0225753151,
-0.0559045412,
-0.0072468449,
0.0153006818,
0.0001326032,
-0.0186293125,
0.0218068119,
-0.023219876,
0.0519844033,
-0.0218699798,
0.0168413557,
0.0119339693,
0.0087368079,
0.0513547584,
-0.0278138705,
0.0513121374,
-0.003792092,
0.027732797,
0.0271876473,
0.0447912924,
0.0419687852,
0.0671208054,
-0.1220799088,
0.0636527836,
-0.0252248477,
-0.0739428923,
-0.0157677568,
0.0263092984,
0.0298264157,
-0.0436931774,
0.089783065,
0.0321589597,
0.0398008265,
-0.0323073305,
0.0059384708,
-0.0230021514,
-0.0064895102,
0.019137064,
0.0825120956,
-0.0055887941,
-0.0181451607,
0.051727958,
0.007738404,
-0.0668433532,
0.0298442394,
0.027863143,
-0.0310253277,
-0.013661582,
-0.0381379835,
-0.0445780866,
0.0494045988,
-0.0012856685,
0.0291306935,
0.0533073694,
0.0397011489,
-0.032676179,
0.0096293455,
-0.0510091335,
0.0370259061,
0.0387825593,
0.0224143118,
-0.0392816812,
-0.0287799425,
0.0008045639,
-0.0098253898,
0.0425538644,
-0.0265840124,
0.0303524826,
0.0484160557,
0.0186421312,
0.0101539269,
0.0093657756,
0.1145879552,
0.0410893671,
-0.0782109052,
-0.058899302,
0.0070232912,
0.0488242246,
-0.0238140132,
0.0709861144,
0.0045519699,
-0.0715302974,
-0.0422558598,
0.0580469556,
-0.0337758511,
-0.0296042301,
-0.0016902073,
0.0090562413,
-0.031745065,
-0.0347084142,
-0.0392056368,
-0.0054097231,
-0.0492108651,
-0.0173258502,
0.0279269703,
-0.0476584099,
0.0498414636,
0.0272657368,
-0.0007498934,
-0.0398035869,
-0.0329618827,
-0.0263473447,
0.0649159774,
-0.002287762,
0.0231182333,
0.061235141,
-0.0027575772,
-0.0783451572,
0.0013169652,
0.0053136456,
0.0209851023,
0.0165280867,
0.058507923,
-0.0281692892,
-0.0298528038,
0.0075803432,
0.0436362475,
0.0344847627,
0.0549707673,
0.0307855848,
0.0127962921,
-0.0227222089,
-0.0273230039,
-0.0204984508,
-0.0501929075,
-0.010431217,
-0.0369259529,
0.0226406585,
-0.0212086719,
0.0157664418,
-0.0115028489,
2.59e-7,
0.0180577748,
0.0729237497,
-0.0204681568,
-0.0370579511,
0.0197631605,
0.069441326,
-0.0245692804,
-0.0295325536,
-0.0199121591,
0.0337276198,
0.0344747826,
-0.0309611149,
0.0063752537,
0.0470643453,
-0.0888406932,
-0.0623961613,
-0.0415589102,
-0.0174355078,
-0.0066205193,
0.0458237715,
0.0331867673,
0.0867383778,
0.0262973756,
-0.0357737541,
-0.037467327,
-0.014680434,
0.0194104929,
-0.0610984862,
-0.0032166268,
0.0407249853,
0.0065137194,
0.0003395812,
-0.0027825208,
-0.0416829474,
0.0077705076,
-0.0233789217,
0.0092443619,
0.0428883024,
-0.0048157955,
0.0064076423,
0.0618857071,
0.0542985611,
0.0170566142,
0.0019741114,
0.0207108073,
0.0110591352,
-0.0730967,
-0.0473054983,
-0.0035143984,
-0.0255037192,
0.0518116578,
-0.0138088474,
-0.0284127872,
-0.0338810869,
-0.0090218056,
-0.0544430241,
-0.0668620691,
-0.0308697261,
0.0240709893,
0.0218308605,
-0.011704105,
-0.0308164004,
0.0214462299,
0.0332453251,
0.0378254652,
-0.0302495454,
0.0566100404,
1.807546643e-34,
0.0684181899,
-0.0177543312,
0.0242409743,
0.0079800859,
-0.0087835854,
0.0066256854,
-0.0113998903,
-0.0282836109,
0.037710458,
-0.023412358,
0.019328123
] |
http://arxiv.org/abs/2401.14295v3 | http://arxiv.org/pdf/2401.14295v3 | Demystifying Chains, Trees, and Graphs of Thoughts | 2024-01-25T16:34:00 | 2024-04-05T11:40:50 | Maciej Besta, Florim Memedi, Zhenyu Zhang, Robert Gerstenberger, Guangyuan Piao, Nils Blach, Piotr Nyczyk, Marcin Copik, Grzegorz Kwaśniewski, Jürgen Müller, Lukas Gianinazzi, Ales Kubicek, Hubert Niewiadomski, Aidan O'Mahony, Onur Mutlu, Torsten Hoefler | cs.CL | cs.CL, cs.AI, cs.LG | The field of natural language processing (NLP) has witnessed significant
progress in recent years, with a notable focus on improving large language
models' (LLM) performance through innovative prompting techniques. Among these,
prompt engineering coupled with structures has emerged as a promising paradigm,
with designs such as Chain-of-Thought, Tree of Thoughts, or Graph of Thoughts,
in which the overall LLM reasoning is guided by a structure such as a graph. As
illustrated with numerous examples, this paradigm significantly enhances the
LLM's capability to solve numerous tasks, ranging from logical or mathematical
reasoning to planning or creative writing. To facilitate the understanding of
this growing field and pave the way for future developments, we devise a
general blueprint for effective and efficient LLM reasoning schemes. For this,
we conduct an in-depth analysis of the prompt execution pipeline, clarifying
and clearly defining different concepts. We then build the first taxonomy of
structure-enhanced LLM reasoning schemes. We focus on identifying fundamental
classes of harnessed structures, and we analyze the representations of these
structures, algorithms executed with these structures, and many others. We
refer to these structures as reasoning topologies, because their representation
becomes to a degree spatial, as they are contained within the LLM context. Our
study compares existing prompting schemes using the proposed taxonomy,
discussing how certain design choices lead to different patterns in performance
and cost. We also outline theoretical underpinnings, relationships between
prompting and other parts of the LLM ecosystem such as knowledge bases, and the
associated research challenges. Our work will help to advance future prompt
engineering techniques. | Demystifying Chains, Trees, and Graphs of Thoughts: The field of natural language processing (NLP) has witnessed significant
progress in recent years, with a notable focus on improving large language
models' (LLM) performance through innovative prompting techniques. Among these,
prompt engineering coupled with structures has emerged as a promising paradigm,
with designs such as Chain-of-Thought, Tree of Thoughts, or Graph of Thoughts,
in which the overall LLM reasoning is guided by a structure such as a graph. As
illustrated with numerous examples, this paradigm significantly enhances the
LLM's capability to solve numerous tasks, ranging from logical or mathematical
reasoning to planning or creative writing. To facilitate the understanding of
this growing field and pave the way for future developments, we devise a
general blueprint for effective and efficient LLM reasoning schemes. For this,
we conduct an in-depth analysis of the prompt execution pipeline, clarifying
and clearly defining different concepts. We then build the first taxonomy of
structure-enhanced LLM reasoning schemes. We focus on identifying fundamental
classes of harnessed structures, and we analyze the representations of these
structures, algorithms executed with these structures, and many others. We
refer to these structures as reasoning topologies, because their representation
becomes to a degree spatial, as they are contained within the LLM context. Our
study compares existing prompting schemes using the proposed taxonomy,
discussing how certain design choices lead to different patterns in performance
and cost. We also outline theoretical underpinnings, relationships between
prompting and other parts of the LLM ecosystem such as knowledge bases, and the
associated research challenges. Our work will help to advance future prompt
engineering techniques. | [
0.0788698271,
-0.0025145747,
-0.0297159161,
0.0470131859,
-0.1067661867,
0.0000795602,
0.0115977805,
-0.0220089965,
0.0631885156,
-0.0499503165,
-0.0040265876,
-0.0383942649,
0.0184099283,
0.0486629084,
-0.0102748219,
-0.0649389923,
0.0276586395,
-0.0663259625,
0.0448673554,
-0.0103981765,
-0.0149949426,
0.0225795787,
-0.0280605573,
0.022347033,
-0.0167723726,
-0.0420587212,
-0.0692546144,
0.0156879388,
-0.0064604552,
-0.0617855154,
-0.0580603816,
0.0189174619,
-0.0270829424,
0.0280314796,
0.0000024588,
-0.0481925383,
-0.0301360842,
0.0006876015,
-0.038814947,
-0.0245969743,
0.0066897762,
-0.0169701055,
0.0098813036,
0.0081057539,
-0.03338762,
-0.0213442259,
0.0257778689,
0.0531284548,
0.0488529801,
0.0723982006,
-0.0129175559,
-0.0578023009,
0.0166754648,
0.0177669749,
0.0125539014,
-0.0223228391,
0.0212890618,
-0.0242081992,
0.0745248348,
0.0385422595,
-0.0205195062,
0.013892808,
0.0181903653,
-0.0251561906,
0.0388004333,
0.0278046913,
0.0311783087,
-0.0449999645,
-0.0547915287,
0.0556273758,
0.0374844968,
0.0245982688,
0.0197115522,
-0.0185155105,
-0.0032658016,
0.0275131445,
-0.0318990462,
0.0597878247,
-0.0391024686,
-0.0064731669,
0.0198961888,
0.0012200458,
0.0014967957,
0.0027235991,
-0.0487881601,
0.08443591,
-0.0110792778,
-0.0523141809,
0.0150312334,
-0.0052049528,
0.014995289,
-0.1126338989,
0.0065129306,
0.0498679318,
-0.0541673675,
0.0294925664,
0.002521679,
0.0277099311,
-0.0010986718,
0.0038361703,
0.0456274785,
0.0336374901,
0.0001690877,
0.0450281501,
-0.0178873371,
0.0513019338,
-0.0331737995,
0.0275094099,
-0.0385013409,
0.0227352902,
-0.0724378973,
-0.0257039778,
-0.0477793515,
0.0672301725,
-0.0465035848,
-0.0386446677,
0.0085172094,
0.0616083108,
-0.0041525438,
0.0296386369,
0.0068310872,
0.0225858968,
0.0605727993,
-0.0138897253,
-0.0630657747,
-0.0035906041,
-0.0338706821,
0.0134151718,
0.0258077029,
-0.0241851751,
-0.0009293749,
0.0090360492,
0.0124456538,
0.0054549109,
-0.0158310607,
0.0499370359,
0.0385438912,
-0.0066651888,
-0.0420334414,
-0.0864990875,
-0.0227553491,
-0.0492228866,
0.0284356959,
-0.0692436621,
0.0097973514,
0.0331334323,
0.0010610501,
0.016609598,
0.0035223241,
0.0306324456,
-0.038235832,
0.0708068013,
-0.0194100942,
0.0337954536,
-0.018896509,
-0.0079506719,
-0.0310865715,
0.0350614861,
0.0276207235,
0.0282896254,
0.0202835202,
0.0237149056,
-0.0200166125,
0.0141990762,
-0.0225020461,
-0.0341361202,
0.0008896447,
0.0120657524,
-0.0246425923,
0.0217049252,
-0.0399613567,
0.053289324,
-0.0162805896,
0.0316033103,
0.0180718135,
0.0989390835,
-0.0062520136,
-0.0103594027,
0.0601998046,
0.0082962569,
-0.0199890062,
0.020907212,
-0.0369890742,
0.0151198851,
-0.050903853,
0.0074662012,
-0.0254118033,
-0.0062245801,
-0.0016151138,
-0.0461711287,
-0.0390268117,
0.0434156433,
-0.002242889,
-0.0123730954,
0.0019726334,
0.0264975503,
-0.021528963,
0.0165349692,
-0.0240975413,
-0.0143871251,
-0.0408006199,
-0.0370929167,
0.0088158408,
0.0710027739,
0.0453855507,
-0.0011281777,
-0.0175628383,
0.0023170456,
0.0035321638,
0.0329057761,
0.0136959348,
-0.0211297926,
-0.0606326237,
-0.0156033887,
-0.0393760912,
-0.0015546978,
0.0071747452,
0.0124754645,
-0.0231467746,
-0.0366022885,
-0.0192927904,
0.0148902331,
-0.0123668583,
0.013391871,
0.0118344286,
0.018072838,
0.0041756891,
0.0379076861,
0.0305390842,
0.0419290587,
0.0046188063,
0.052589003,
0.0255097263,
-0.0119243739,
-0.030977603,
0.0057622991,
-0.0682412833,
-0.0504885949,
0.0225543864,
-0.0065221735,
-0.0147716133,
-0.0441950448,
-0.0195576455,
-0.0531465299,
-0.0101021994,
-0.0652557611,
0.0140288351,
-0.0161711406,
0.0393288992,
0.0141866002,
-0.0765175,
-0.0248163436,
0.1013186499,
-0.0078481045,
0.0204978026,
0.0564125553,
0.0357863605,
0.0436938033,
-0.028797308,
0.0017257613,
0.0255663339,
0.0137795666,
-0.0171495639,
0.0278981701,
0.0191514362,
-0.065430209,
-0.0119363051,
-0.092215322,
0.0012785492,
-0.0129200416,
0.0236681346,
0.0071284086,
0.0041280361,
-0.006529219,
-0.0225097984,
-0.0326192938,
0.0009694993,
0.0318076909,
0.0295504667,
0.0365618989,
-0.01811135,
-0.0544021875,
0.0421800129,
-0.0116872154,
-0.0572841689,
0.0524088182,
0.0185461827,
-0.0743441954,
-0.0572070591,
0.0368627831,
-0.0291504618,
0.0078861592,
0.0067941207,
0.0348241702,
0.0456878394,
-0.0311813541,
0.0516752005,
0.0530217998,
-0.0188640468,
-0.0133228041,
0.0406963788,
-0.0303667374,
-0.0205592606,
0.0179251209,
-0.0330702327,
0.0263065845,
0.0082880249,
-0.0139504857,
-0.0090989852,
0.042856127,
0.0320436954,
-0.0055891098,
-0.0558102392,
0.0345177539,
0.0049050762,
0.0216031764,
-0.0525501929,
-0.0656179488,
0.0054606502,
0.097215265,
-0.023250591,
-0.0396171249,
0.0909803659,
-0.0267970636,
0.0223883819,
-0.0322213136,
0.0115103545,
-0.0557759553,
0.09558402,
0.0196101908,
0.0034089279,
0.0517335162,
-0.0420044735,
-0.0364642553,
0.0234859139,
0.0059069595,
-0.0239536203,
-0.0000287129,
-0.0509481616,
0.0123690758,
-0.0093454476,
-0.046071101,
-0.0511350147,
0.0114171961,
-0.0485462919,
0.016300695,
0.0092169503,
0.0100283334,
0.0224586502,
0.0082746316,
0.0124280369,
0.0190752484,
0.0017293057,
-0.0106522832,
-0.0554814674,
0.0501139425,
0.008930305,
0.0106296297,
0.0343848728,
0.0133618526,
0.05683548,
0.0384959877,
0.0170337912,
0.024388073,
-0.0037958706,
0.0411356874,
0.0262750201,
-0.0260962192,
-0.0545143262,
0.0362079479,
0.0320316851,
-0.0638969541,
-0.0258788522,
0.1052379087,
0.0715573356,
-0.0264572166,
0.0012572918,
-0.0188757591,
0.0258359835,
0.0386162288,
0.0214304067,
-0.0921064243,
-0.0171191283,
0.0173515361,
-0.0234961137,
-0.0511904061,
-0.0575165115,
-0.0387498103,
-0.1160134301,
0.0469869561,
-0.0141373565,
-0.0123488475,
-0.0220869072,
-0.0391370654,
-0.0129529415,
0.066456683,
-0.0550714582,
0.0104933325,
-0.0220006723,
-0.0024870057,
0.0167141557,
-0.0339768417,
0.0015714547,
0.0064332932,
-0.0089938156,
0.015048068,
0.0373628289,
0.0186506137,
0.0424469262,
-0.0589387007,
0.0201804489,
0.0179830939,
0.0554523803,
-0.0524558052,
-0.0085188225,
0.0195455011,
0.0010636967,
-0.0467889272,
0.0190104302,
-0.0565244034,
-0.0255816039,
0.0249601193,
0.1389091164,
-0.0504613556,
0.0141823711,
-0.0330346338,
0.0264228098,
-0.0043352861,
0.0534935109,
0.0252432842,
-0.0124250948,
-0.0597786866,
-0.0437099263,
-0.0160359871,
-0.024744954,
-0.0503739156,
0.0418999009,
0.0044929581,
0.0403934866,
-0.0493601412,
-0.0290365331,
0.0211894251,
0.0199896377,
0.0200380329,
-0.0126971742,
0.021783689,
-0.0170919579,
-0.0217060614,
0.0081936549,
-0.0302228536,
0.0262685362,
-0.025793666,
0.0240459051,
0.0035932844,
-0.0253785588,
-0.0113960309,
-0.0020878795,
-0.0339823104,
-0.0473271422,
-0.0430489518,
0.0224382207,
0.02165002,
0.0160918366,
-0.00765952,
-0.0231414326,
0.0443530343,
0.0256114546,
-0.0247178543,
0.0126749622,
0.0115372213,
0.019587744,
0.0114376284,
0.0263919495,
-0.0173104946,
0.0271189343,
-0.0015565419,
0.0415984094,
-0.0054605231,
-0.0196777862,
0.0355176181,
0.0722213387,
0.0364491828,
0.019362757,
-0.01796427,
-0.0233058222,
0.0220064372,
-0.0446554497,
-0.0726349279,
-0.0373733528,
0.0296432357,
-0.011112975,
-0.0093609756,
0.0031552883,
-0.0312128495,
-0.0086891912,
0.0289924387,
-0.005267608,
-0.0135611454,
0.0529500432,
-0.0126946559,
0.0249942522,
-0.0097028734,
-0.0170304533,
0.0311120357,
-0.0143248029,
0.0352353491,
-0.0063824095,
0.0909341946,
0.0342433639,
-0.0159772299,
-0.0413733572,
0.0139730042,
-0.0812879801,
-0.1026576236,
0.0592944734,
0.0242790096,
0.0307265706,
0.0064830719,
-0.0056406781,
-0.035172496,
0.0057203439,
0.0164862033,
0.0134115927,
0.0663885921,
-0.075596638,
0.0145721603,
0.0290694591,
0.0086344024,
-0.0204182696,
0.0140138511,
-0.0462559164,
0.016479753,
0.0126045533,
-0.0516757518,
-0.0478832498,
-0.0259203631,
-0.0205236357,
0.041036386,
-0.0042834189,
0.0285629723,
-0.1205024794,
-0.0228278041,
0.031276226,
0.0038747683,
0.0381947979,
-0.0198927037,
0.0678186789,
0.0406031683,
0.0329804085,
-0.0363489389,
-0.03145409,
0.0163333602,
-0.034051843,
-0.0249101073,
-0.0161588136,
-6.571781166e-33,
-0.0230959766,
-0.0730117336,
0.0579234771,
0.0135028064,
-0.0244067423,
0.0197179597,
-0.023663355,
0.0240697525,
-0.0039877626,
0.0123682311,
-0.0341783278,
-0.001685633,
0.003477803,
0.0135320192,
-0.0024253665,
0.0091671934,
0.0157559365,
-0.0118277976,
-0.0262170956,
-0.0092972219,
0.0582982153,
-0.0044551031,
-0.0112307081,
-0.0337462202,
0.0498392358,
0.0014110645,
0.0013897971,
0.044668816,
-0.0389905795,
0.0205989704,
-0.0303689428,
0.0247381739,
0.0312291402,
-0.0356362611,
-0.0057107457,
0.0710389838,
-0.0604540482,
-0.051900737,
0.0434272811,
-0.001643842,
-0.0732545033,
-0.0290038381,
0.0162713993,
-0.0123613216,
0.0003202829,
-0.0147833871,
0.0198063944,
-0.0260117278,
-0.0296196491,
0.0033330433,
-0.0333923846,
0.0114405034,
-0.006639536,
-0.0143627618,
0.0252139978,
0.0311039165,
-0.0041547026,
0.0236954447,
-0.0295573287,
0.0342797823,
0.0212228373,
-0.0005022862,
0.0344030522,
-0.0012042694,
-0.005881404,
0.0546522252,
0.0230976883,
0.0055214167,
-0.0412557721,
0.0358403809,
-0.0352193415,
0.072000742,
0.0372215286,
-0.0186806247,
0.0106610134,
0.0458394624,
-0.0794863477,
-0.0059350887,
0.0033006382,
0.0059401663,
0.0288547296,
-0.0061640153,
-0.0258300621,
-0.0383827984,
-0.0499036983,
-0.0380162597,
-0.0287718531,
0.0032476739,
0.0122641344,
-0.0004076799,
-0.003691009,
-0.0112503832,
-0.0031391983,
-0.0405749567,
-0.0222810283,
0.0287944246,
-0.0035473094,
0.0386554152,
-0.0033544735,
0.050803069,
-0.0290315263,
-0.0081336936,
-0.0122222686,
-0.0082002319,
0.0127856955,
0.0247295108,
0.0135252541,
0.0254035238,
-0.0396371633,
-0.0105067948,
-0.0552220419,
0.0254094973,
-0.0350077115,
-0.0088024093,
0.0574415736,
0.025979843,
-0.0021799803,
0.0608741082,
0.0114255194,
-0.002730804,
-0.0202597268,
0.0451349691,
-0.0400355272,
0.0143455053,
-0.0063625234,
-0.0165873989,
-0.0109525956,
0.0245902706,
0.040132951,
-0.0406577438,
-0.0084831491,
0.0323727876,
3.207e-7,
0.0098539675,
0.0400315523,
0.045945622,
0.0531897098,
0.0398288965,
0.0303126778,
0.0072524538,
-0.0113664279,
0.0045475163,
-0.0707839206,
0.0017619962,
0.0307550952,
0.0026552747,
0.0290067028,
-0.1218086779,
0.019223785,
-0.0406026244,
-0.0188685823,
-0.0602381155,
0.0337782949,
0.1042383164,
0.0115396231,
-0.0207816884,
-0.014760104,
-0.0322490558,
-0.0176546164,
0.0270443298,
0.0009308303,
0.0182896722,
-0.0125402333,
0.0260340534,
-0.0455969013,
0.0044150124,
0.0476931371,
0.0455542058,
0.0249791536,
-0.0249272939,
0.0336351618,
0.0179622881,
0.0133548155,
0.0038504642,
-0.043233607,
-0.0090657584,
-0.0359374471,
0.0854386389,
-0.0337058119,
-0.017746279,
0.0015717102,
0.0192436539,
-0.0173845924,
-0.0024272408,
0.0288892929,
-0.0014304894,
0.0226167068,
0.0072813984,
-0.0136380382,
0.0126081435,
-0.0770245492,
0.0002951923,
-0.0091444915,
0.0002100091,
-0.0259723365,
-0.0287964642,
0.0380868837,
0.0084943958,
-0.0717187077,
-0.0466997512,
3.459964766e-34,
0.0096247597,
-0.0201394185,
0.0154011091,
0.0311696883,
0.0309075732,
-0.009696939,
-0.0092978403,
0.0058517605,
0.0490285233,
-0.1112100258,
-0.0548212342
] |
http://arxiv.org/abs/2401.14280v3 | http://arxiv.org/pdf/2401.14280v3 | RomanSetu: Efficiently unlocking multilingual capabilities of Large Language Models via Romanization | 2024-01-25T16:11:41 | 2024-06-23T11:40:20 | Jaavid Aktar Husain, Raj Dabre, Aswanth Kumar, Jay Gala, Thanmay Jayakumar, Ratish Puduppully, Anoop Kunchukuttan | cs.CL | cs.CL, cs.AI | This study addresses the challenge of extending Large Language Models (LLMs)
to non-English languages that use non-Roman scripts. We propose an approach
that utilizes the romanized form of text as an interface for LLMs,
hypothesizing that its frequent informal use and shared tokens with English
enhance cross-lingual alignment. Our approach involves the continual
pretraining of an English LLM like Llama 2 on romanized text of non-English,
non-Roman script languages, followed by instruction tuning on romanized data.
The results indicate that romanized text not only reduces token fertility by
2x-4x but also matches or outperforms native script representation across
various NLU, NLG, and MT tasks. Moreover, the embeddings computed on romanized
text exhibit closer alignment with their English translations than those from
the native script. Our approach presents a promising direction for leveraging
the power of English LLMs in languages traditionally underrepresented in NLP.
Our code is available on https://github.com/AI4Bharat/romansetu. | RomanSetu: Efficiently unlocking multilingual capabilities of Large Language Models via Romanization: This study addresses the challenge of extending Large Language Models (LLMs)
to non-English languages that use non-Roman scripts. We propose an approach
that utilizes the romanized form of text as an interface for LLMs,
hypothesizing that its frequent informal use and shared tokens with English
enhance cross-lingual alignment. Our approach involves the continual
pretraining of an English LLM like Llama 2 on romanized text of non-English,
non-Roman script languages, followed by instruction tuning on romanized data.
The results indicate that romanized text not only reduces token fertility by
2x-4x but also matches or outperforms native script representation across
various NLU, NLG, and MT tasks. Moreover, the embeddings computed on romanized
text exhibit closer alignment with their English translations than those from
the native script. Our approach presents a promising direction for leveraging
the power of English LLMs in languages traditionally underrepresented in NLP.
Our code is available on https://github.com/AI4Bharat/romansetu. | [
0.0035711001,
0.0916090682,
-0.0085824672,
0.0447967537,
-0.0320880897,
0.0304025877,
0.0459024124,
0.0034944923,
-0.0042519644,
-0.0124746254,
-0.0531189889,
-0.0763629526,
0.0133986808,
-0.0170152541,
0.076833576,
-0.008334266,
0.0377770998,
-0.0493200719,
-0.0191953499,
-0.0184696764,
0.0421128049,
0.0225389395,
0.0320706926,
0.0365370251,
-0.0432536937,
-0.021167377,
0.0162090398,
0.0077137598,
0.0179076102,
-0.0016152546,
-0.020690836,
0.0238208398,
0.0118493699,
0.0703036189,
0.0000017169,
-0.0412803777,
-0.0101771178,
-0.0179529581,
0.0015762356,
0.0440643318,
0.0632756054,
-0.0352045968,
-0.0305575728,
0.0204696823,
-0.0093742805,
-0.0165554788,
0.0711146742,
0.0108034983,
0.0035688835,
0.0622097626,
-0.016765058,
-0.0138252182,
0.0168423429,
-0.0550246835,
-0.0400404818,
-0.0383204371,
0.0012038442,
-0.0204611737,
-0.0193312336,
-0.0111174881,
-0.0243088193,
0.0405750275,
0.0060802111,
-0.0134195816,
-0.0353821255,
0.0309993885,
-0.0232652314,
-0.0085671498,
0.0197847039,
0.0487769023,
-0.014412174,
-0.0527161919,
-0.0328095406,
0.0247243326,
-0.0098709716,
0.0019373305,
-0.0480127409,
0.0225903764,
0.023057377,
0.0458084829,
0.0683847517,
-0.011577161,
0.0606929436,
-0.0143643655,
-0.0269127805,
0.0290630925,
0.0796014518,
-0.0463748574,
0.0016174571,
0.0303649828,
-0.0193056129,
-0.0601280108,
0.03061077,
0.0362793356,
-0.0254281592,
0.0334170312,
-0.0407397412,
-0.0287810564,
0.0277262013,
0.0001930632,
0.0656871796,
0.0677097887,
-0.0368328616,
0.0237703342,
-0.047365021,
0.0702361614,
-0.0398633108,
0.020077398,
-0.0447341278,
-0.0271603484,
-0.0069873598,
-0.0144296102,
-0.0326595269,
-0.0020817437,
-0.0150139891,
-0.0045153107,
-0.0292610656,
0.0306637995,
0.025190277,
0.0318424255,
0.0358107686,
-0.0494973548,
-0.0398883857,
-0.0063956003,
-0.0157368686,
0.0594674014,
-0.0348548107,
-0.0001261321,
-0.0492592193,
-0.058700338,
-0.0083092311,
0.0265272576,
-0.0093679335,
-0.042823635,
0.0243244916,
0.0611751713,
-0.0233493764,
-0.0325676873,
-0.0135300634,
0.042247247,
-0.0174142662,
-0.0575571358,
-0.0241042227,
-0.0683407933,
0.0847532302,
-0.0010945713,
0.0290948078,
-0.0556138642,
0.0180657227,
0.0136759467,
0.0007973242,
0.0594438799,
0.0090566576,
0.0238070581,
-0.0117859002,
-0.0406026691,
0.0263230018,
0.0538142622,
0.0337672606,
0.0034021058,
0.0508774854,
-0.0125837224,
-0.029212812,
0.0952339321,
0.0210929886,
0.0263888109,
-0.035237696,
0.0100854812,
0.0607869029,
0.039704483,
0.0449605547,
0.080047667,
0.0115222977,
0.0130207213,
0.0282699335,
0.0746431798,
0.0014087789,
0.0313979387,
0.0462302119,
-0.0115109039,
-0.004668179,
-0.0024503171,
-0.0282267481,
0.0074633346,
0.0247044042,
0.0119737899,
0.0276832562,
0.0753625035,
-0.0368741527,
-0.0117692538,
-0.0294298809,
-0.0420736931,
0.0320796333,
0.0009632354,
0.058742296,
-0.0044553317,
-0.0308655202,
0.039007578,
0.0500600561,
-0.0388724916,
0.0296037868,
-0.0257680193,
0.0223655272,
0.0776205808,
0.0362143181,
-0.0042884462,
-0.0463393852,
0.015663106,
0.0037054771,
0.0533618182,
0.0361340716,
-0.0309001468,
0.0161302779,
-0.0700695962,
0.0349661596,
0.0002759819,
-0.007458616,
0.06319464,
0.0005983551,
0.0267112888,
0.0078110308,
-0.0643185601,
-0.0455445349,
0.0623059459,
-0.0773139223,
-0.0035469225,
-0.0385022648,
0.0145975528,
0.0115225455,
0.0110289063,
-0.033487048,
0.0032953576,
0.0198563673,
-0.0881562307,
0.0242633149,
0.0258618537,
-0.0413463674,
-0.0131216003,
0.0926890224,
-0.00366479,
-0.0376338214,
0.0160650238,
-0.0129952217,
-0.0419110917,
0.0167837739,
-0.0195592884,
0.0474363677,
-0.0574236959,
0.0242447834,
-0.0114103472,
0.0037425335,
-0.0329670087,
0.0339611731,
0.0006619929,
0.0334109776,
0.0477111824,
-0.0450278521,
0.049330458,
-0.0100063141,
-0.0423623584,
0.036272347,
0.0104027009,
0.0069538467,
-0.0019346176,
0.0267718937,
0.0266389791,
-0.0619025901,
-0.1163660511,
0.0135751646,
-0.046510078,
0.0473010018,
0.0170326158,
0.0275686942,
-0.0217210259,
-0.0119422358,
-0.0577051267,
-0.0488292128,
0.0053517525,
0.0211911704,
0.0363627598,
-0.0161378775,
0.0011649751,
-0.0299015921,
-0.013962592,
0.003769333,
0.0398194231,
-0.0214083605,
-0.0099651096,
-0.0305754058,
-0.0019724118,
-0.0291187447,
0.0057100216,
-0.023689663,
-0.0394101106,
-0.0000514425,
0.0121173458,
0.0661325976,
0.0208800584,
-0.0413985774,
-0.0084369807,
0.0505811237,
0.0073600016,
-0.0418277718,
-0.0061362009,
0.0264472421,
-0.0396735258,
-0.0768461823,
-0.0560798906,
-0.0111684185,
0.0631089434,
-0.0463097729,
-0.0195663255,
-0.0711383671,
-0.0451763198,
-0.0416645035,
-0.0067679211,
0.0083011175,
0.0106728934,
-0.0152570037,
0.0159964506,
-0.0282795709,
-0.0366513096,
0.0643570647,
-0.0786055699,
-0.0324135683,
-0.0313059203,
-0.0016283707,
-0.0678545833,
0.0120589565,
-0.0019457126,
0.0282657146,
-0.0386519469,
-0.0560935959,
-0.0346640199,
0.0423463024,
-0.0301565733,
-0.0307841878,
0.0612431839,
-0.0000463271,
0.00102401,
-0.0241404288,
-0.0392640121,
-0.0214095488,
-0.0168102439,
-0.0326988101,
0.0112890508,
0.0499496274,
-0.0114562055,
-0.0051914295,
0.0199769214,
0.0077900942,
0.0341754779,
-0.0094760545,
0.048102323,
-0.0309621841,
0.0143790841,
0.0320405662,
0.0229045991,
-0.0166690685,
-0.0303896628,
0.0341797434,
0.0077284602,
0.0256190132,
-0.0432127938,
-0.028679071,
-0.0007024163,
0.0508711413,
-0.0296050441,
-0.0100888768,
-0.045307558,
0.0595699102,
-0.0033427579,
-0.0104758851,
0.1215565279,
0.0139157828,
-0.0566133074,
0.0058570411,
-0.0449301153,
0.0508178659,
0.0341470689,
0.0263872277,
-0.0839264989,
-0.0204257742,
0.021884609,
-0.0057461923,
0.0478491299,
0.0091604255,
0.0118450094,
-0.1232685968,
-0.0350433141,
0.0312993675,
-0.0247325748,
-0.0259464011,
-0.0206101257,
-0.019910099,
-0.0205929745,
0.0426929146,
-0.0022264596,
-0.0136903068,
0.0050305934,
0.0309451558,
0.0030869725,
-0.0466745757,
-0.0154392635,
-0.0529860109,
0.0350453891,
0.0231817551,
0.0371889658,
-0.029698275,
-0.0169291291,
0.0227909051,
0.0051105008,
-0.0080367131,
-0.0028821642,
0.0056424206,
-0.0117530134,
-0.0402002186,
-0.0146764442,
-0.0029605979,
-0.0413092785,
-0.0059478828,
0.0024798699,
0.0671608374,
0.0149522321,
0.0106544672,
0.0203010961,
-0.0394715369,
-0.0126545085,
0.0537755042,
0.0160480347,
-0.0567686036,
-0.0167367682,
-0.0309468694,
-0.0041480525,
-0.0420667231,
-0.0353937484,
0.0122321369,
-0.0079275249,
-0.0184300225,
-0.0461838618,
-0.057930693,
0.019865144,
0.0268568099,
0.0599887148,
0.0103736296,
0.0029145796,
-0.0436241515,
-0.0150914481,
0.0160140656,
-0.0626244321,
0.0103788646,
-0.0702695251,
-0.0231502354,
0.0180243123,
-0.0095181791,
0.0075001512,
-0.025294017,
-0.0767178088,
-0.0762970001,
-0.020878002,
0.0111528728,
0.0130863031,
0.0390008502,
-0.0152344741,
-0.0356815495,
0.0338080004,
-0.0234513693,
-0.013881593,
0.027685076,
0.0175554361,
-0.0112963039,
0.0059971474,
0.0599343665,
-0.0249412246,
-0.0012272818,
0.0044687209,
0.0660509393,
-0.0265043825,
-0.0031970656,
0.0144522851,
0.0455888361,
-0.0178636368,
-0.0669504702,
-0.014351449,
-0.0660702139,
0.0266790427,
-0.0362541489,
-0.0039491653,
-0.0123838997,
0.0034915695,
0.0584369004,
-0.009894941,
0.0037952354,
-0.0091501316,
0.0110130878,
0.0187512413,
-0.022123985,
0.0104704481,
0.0080736568,
0.0218581352,
0.017463699,
0.0474850424,
-0.0042450768,
0.0199967343,
-0.0245536808,
0.0370418802,
-0.060133893,
-0.0000733182,
-0.0014790054,
0.006185221,
0.0056818081,
0.0403248072,
0.0191926733,
-0.0271293744,
0.0160157476,
0.0168029107,
0.0986841917,
-0.0281117242,
0.0696396902,
-0.0183274727,
-0.0126042766,
-0.0159807317,
0.0155128501,
0.0086705647,
-0.0180638693,
0.0153958499,
-0.0161829684,
-0.0356976539,
-0.0024700644,
0.0771842226,
0.0344001278,
0.01121262,
0.0289937295,
0.0178980902,
-0.0241939016,
-0.0287488401,
0.025592804,
0.001999941,
-0.0547009036,
-0.0136467451,
-0.0298186783,
-0.0898768827,
-0.0053975661,
0.0051513971,
-0.0508497246,
-0.0044515845,
-0.0166505575,
0.0452864952,
-0.0035894834,
0.0458853319,
-0.0435195006,
0.0319210403,
0.0233357251,
-0.0368427821,
-0.0035775783,
-5.810397679e-33,
-0.0322413109,
0.0189089552,
0.011604242,
-0.0392024033,
-0.0162016451,
-0.0021303927,
-0.0212454908,
0.0002990333,
-0.0176836886,
0.0065698707,
-0.043332871,
-0.0178730227,
0.0175335985,
0.0146348868,
0.0288224611,
-0.0202333611,
0.030912267,
0.0138051491,
-0.0067335959,
0.0196352135,
0.0383005403,
0.0416416265,
0.0789633021,
-0.0226069298,
0.0207151677,
-0.0269606914,
0.0211601146,
0.005690149,
0.0370400958,
0.0654880479,
-0.0471804403,
0.0298150536,
0.0183859076,
0.0147721339,
-0.0259147231,
0.040665213,
-0.0666439682,
-0.0211222507,
0.05559038,
0.0448238738,
0.0183683578,
-0.0814319104,
-0.0004036763,
-0.0211845879,
-0.0365607664,
0.0247508865,
0.0085670352,
-0.0017543149,
-0.0052452441,
-0.0091004148,
-0.0937115401,
0.0149422679,
0.0084489509,
0.0001722359,
0.0578965098,
0.0167693682,
0.0156273413,
0.0389821231,
-0.0517545603,
0.0037306012,
-0.0607432276,
0.0479096621,
0.0793367624,
-0.015664909,
-0.0131064076,
0.0321259983,
-0.0159306191,
-0.0090876976,
0.0129913026,
-0.0094647342,
0.0314622112,
-0.0308966525,
0.0196386985,
0.0252675917,
0.0262810569,
-0.0405172035,
-0.0717360303,
0.0160012711,
0.0291670877,
0.023439534,
0.0257693697,
-0.0029070484,
-0.0184748452,
-0.0140730226,
-0.0180162676,
-0.0029637511,
-0.0127902078,
-0.029196566,
0.0224536266,
-0.0228693951,
-0.0007552215,
0.0101711964,
0.040376734,
-0.0379256122,
-0.0286790654,
0.0333396085,
0.0216380898,
0.0120436866,
-0.0203335322,
0.0548378676,
-0.0657541901,
-0.0584203675,
-0.0137371914,
-0.0395527482,
0.0193066001,
-0.0068660229,
0.0653742626,
-0.000018637,
-0.0839479864,
0.0203089658,
-0.0444430225,
-0.0007266206,
0.0079719285,
0.0152095547,
0.0627626553,
0.0047559151,
0.0172414687,
0.0219137538,
0.0598456264,
-0.0098530352,
-0.0380586125,
-0.0179308318,
0.0056130253,
0.0079329032,
-0.0300670508,
0.0247485433,
-0.0535506792,
0.0236848872,
0.0096887518,
-0.0441542715,
-0.0117300265,
0.0025964992,
2.52e-7,
0.0687853768,
0.0447048247,
0.0433131792,
0.0060439003,
0.0182130206,
-0.0025019955,
-0.0573083088,
0.0350120999,
0.0045685442,
0.0087255472,
0.010466856,
0.0160933547,
0.072677508,
0.0238521639,
-0.0767676309,
0.018074682,
-0.0377276279,
0.0243028812,
-0.0601227023,
0.0141799822,
-0.0292156041,
0.1354085952,
-0.0167343654,
0.0053991023,
-0.0264885705,
-0.0617966205,
0.0668072551,
-0.0369718559,
-0.0419398658,
0.0326592065,
0.0051796096,
0.0153777739,
-0.0182950888,
-0.0117537696,
-0.0031169527,
0.0432817563,
0.0391824804,
0.0598314218,
-0.0090788649,
-0.0119294571,
0.0004431943,
-0.044465594,
-0.0335454755,
-0.0506981798,
0.051345475,
0.0465035923,
-0.0268360879,
-0.0222821478,
0.0113920728,
0.0166605338,
0.0275781695,
-0.0051624789,
-0.0031975794,
0.0479465276,
0.0706044286,
-0.0119970366,
-0.0399787277,
-0.0587575659,
-0.0511625931,
0.0085131861,
-0.034185227,
-0.0290644094,
-0.0082610566,
-0.0158140976,
0.0040096645,
-0.0284224618,
0.0148261664,
2.529386959e-34,
0.0421621464,
-0.0130129969,
0.0370018259,
0.066675216,
0.0250246339,
-0.0470178127,
0.0055107926,
0.0182458237,
-0.0190024562,
-0.0662332401,
-0.0073899804
] |
http://arxiv.org/abs/2401.14267v2 | http://arxiv.org/pdf/2401.14267v2 | Transformers and Cortical Waves: Encoders for Pulling In Context Across Time | 2024-01-25T16:01:49 | 2024-07-03T03:41:07 | Lyle Muller, Patricia S. Churchland, Terrence J. Sejnowski | cs.CL | cs.CL, cs.AI | The capabilities of transformer networks such as ChatGPT and other Large
Language Models (LLMs) have captured the world's attention. The crucial
computational mechanism underlying their performance relies on transforming a
complete input sequence - for example, all the words in a sentence - into a
long "encoding vector" that allows transformers to learn long-range temporal
dependencies in naturalistic sequences. Specifically, "self-attention" applied
to this encoding vector enhances temporal context in transformers by computing
associations between pairs of words in the input sequence. We suggest that
waves of neural activity traveling across single cortical areas or multiple
regions at the whole-brain scale could implement a similar encoding principle.
By encapsulating recent input history into a single spatial pattern at each
moment in time, cortical waves may enable temporal context to be extracted from
sequences of sensory inputs, the same computational principle used in
transformers. | Transformers and Cortical Waves: Encoders for Pulling In Context Across Time: The capabilities of transformer networks such as ChatGPT and other Large
Language Models (LLMs) have captured the world's attention. The crucial
computational mechanism underlying their performance relies on transforming a
complete input sequence - for example, all the words in a sentence - into a
long "encoding vector" that allows transformers to learn long-range temporal
dependencies in naturalistic sequences. Specifically, "self-attention" applied
to this encoding vector enhances temporal context in transformers by computing
associations between pairs of words in the input sequence. We suggest that
waves of neural activity traveling across single cortical areas or multiple
regions at the whole-brain scale could implement a similar encoding principle.
By encapsulating recent input history into a single spatial pattern at each
moment in time, cortical waves may enable temporal context to be extracted from
sequences of sensory inputs, the same computational principle used in
transformers. | [
-0.0295414589,
0.0632718205,
-0.0251050126,
-0.0115686655,
-0.0423391759,
-0.0160236675,
0.0250584353,
0.0267041773,
-0.0211357474,
-0.0041099386,
-0.0617907979,
-0.0447375886,
0.0477101691,
0.0323362835,
0.030375205,
-0.0234846808,
0.0620036982,
0.0215592999,
-0.0474899784,
0.0025342775,
-0.0514025576,
-0.0650979578,
-0.0332948528,
0.0332546532,
-0.0084341858,
-0.0358890444,
-0.0278913826,
-0.0593550988,
0.016048206,
-0.0290948991,
-0.0371909887,
0.0372449718,
0.0329060666,
0.0229031686,
0.0000017169,
0.0006319135,
0.0311913472,
-0.0364382155,
-0.0072277887,
-0.0248880964,
-0.0426180176,
0.0367431566,
-0.0075995387,
-0.0363332927,
-0.0477786139,
0.0216429457,
0.0403212793,
0.0322165489,
0.0232053865,
0.0431015976,
-0.0181189738,
-0.0385851413,
0.0169920996,
-0.0322467983,
0.0457259379,
0.0201382972,
0.0237380862,
-0.0506898873,
0.0112859402,
0.0269756857,
-0.059967421,
0.0402546376,
0.0351844244,
0.0194586124,
0.0427386723,
0.0947986543,
-0.0230983254,
0.0167468376,
-0.0061912281,
0.022728106,
0.0068112444,
-0.0472011566,
-0.0133184949,
0.0464729182,
-0.0128507828,
-0.0151497843,
-0.0498715378,
-0.007701416,
-0.0046804599,
0.0487582646,
0.0090190917,
0.0123628844,
0.0127069736,
-0.0197324902,
-0.0272283256,
0.049142722,
0.0085584689,
-0.0176717434,
-0.0361341387,
-0.0140510034,
0.0409580059,
-0.0483964086,
0.0252076481,
0.0469281934,
0.0322527662,
-0.0194140244,
0.0007067291,
0.0045765014,
-0.0013796123,
-0.0404989459,
-0.005552927,
0.029545404,
-0.0545404218,
0.0308765061,
-0.0220459327,
0.0711752921,
-0.084543556,
0.0261168722,
-0.0148675181,
0.0337068625,
-0.0634564087,
-0.0096349353,
-0.0192251485,
0.0372830182,
-0.0110635748,
-0.0495132059,
-0.0586291142,
0.0345785841,
0.0166974552,
-0.0743594244,
-0.0143993134,
0.0403501578,
-0.0051188646,
0.0235961508,
0.0016537989,
0.0490418635,
0.0085158432,
-0.007896116,
-0.0063569648,
-0.0112291612,
-0.0567831621,
0.0386625454,
0.0182933677,
-0.0256007388,
-0.0295858178,
0.0620209202,
-0.0130549408,
-0.0402986892,
-0.0220227279,
0.0183366966,
-0.0278535876,
0.0194155853,
0.0485184975,
-0.0457389876,
0.0061972281,
0.021293344,
-0.0286038499,
-0.0051119118,
0.0115909921,
-0.0161776412,
-0.0431816913,
0.0420639664,
0.0348600112,
0.0092157535,
0.03755676,
0.0247650333,
0.0466684587,
-0.0251843464,
0.0568055771,
0.009660162,
-0.0442143902,
-0.0334607214,
0.0242440458,
0.0165610556,
0.0188966114,
-0.0368854515,
-0.0453389771,
-0.0662075952,
0.002091842,
0.0385926999,
0.0421941057,
0.0313534103,
-0.0399618484,
-0.0339107662,
0.0656416863,
0.072387889,
0.0543801673,
0.0614792481,
0.0396717228,
-0.0531329848,
0.002107996,
0.0081242705,
-0.0500510931,
0.079153344,
0.0276802015,
-0.0267004315,
-0.0021725285,
0.0111096548,
-0.0190600138,
-0.0152796768,
0.0106483772,
-0.0039981622,
-0.0286934339,
-0.0014833729,
0.0434360765,
-0.0060258382,
-0.0084467661,
0.0371465907,
0.0123609798,
0.0080530271,
0.0268607382,
-0.0183302946,
-0.0032585475,
0.0195794366,
0.0218055751,
-0.0134481546,
-0.0050693504,
-0.0132632032,
0.0180860125,
0.0512688085,
-0.0057025035,
0.0264473483,
0.0155438846,
-0.0050622537,
-0.0395110212,
0.0175653845,
-0.0239489265,
0.0233970229,
-0.0157838315,
0.0023784551,
-0.0364816152,
0.0367519967,
0.0282738283,
0.0255821012,
-0.0224451162,
-0.0359467231,
-0.0215199944,
0.0001845457,
0.0262181424,
-0.0094626024,
-0.0137092425,
0.0342317298,
0.035374403,
-0.0530866012,
-0.0363458283,
0.062549144,
-0.1018752977,
-0.0107402233,
0.0129126441,
-0.0184735954,
0.0134284915,
-0.000209155,
0.0317257605,
-0.0379488878,
0.0435829572,
-0.0270218216,
0.0702581108,
-0.0064307591,
-0.0039111837,
0.0169550646,
0.0726272166,
-0.0362890474,
0.0497839637,
0.009095341,
-0.0211746208,
-0.0360162519,
-0.002322153,
0.0066801929,
-0.0308329482,
-0.0135763567,
0.0042741233,
-0.0301629361,
-0.0339108706,
0.0037301704,
-0.028898498,
-0.0341617167,
-0.0300196428,
-0.077991493,
-0.0244543292,
0.0328434519,
0.0080136973,
-0.004782564,
-0.0157890301,
0.0186757762,
0.0066076736,
-0.011246033,
0.0461555012,
0.0143611208,
0.0451056324,
0.0268031526,
-0.0231406838,
-0.0320602953,
-0.0352653265,
-0.0385417007,
-0.038750995,
0.0811802968,
0.02498951,
-0.0662736446,
-0.0575061254,
-0.0102041308,
-0.020452911,
0.0101692798,
-0.0060490225,
-0.0318464078,
0.0308280475,
-0.0091754552,
0.058066994,
0.065862529,
0.0496431179,
0.0211396273,
0.043719288,
-0.0404074788,
-0.0079039754,
-0.0055258754,
-0.0628830269,
0.0202724058,
-0.0395519473,
0.0321131609,
0.0041906652,
0.1050915793,
0.056698665,
-0.0150016434,
-0.0196669623,
-0.0408545136,
-0.0155338524,
-0.0223039128,
-0.0137344729,
-0.0137387542,
-0.0634528995,
0.0227142591,
-0.01192219,
-0.0370209217,
0.0522042438,
0.024104964,
-0.0162008554,
0.0261550881,
0.0058360593,
-0.0083761308,
0.0659302548,
0.052231092,
0.0148804951,
0.0037840141,
-0.031269189,
-0.0187846571,
-0.00469193,
-0.0517870821,
-0.011214911,
0.0287013575,
-0.0182678159,
0.0073119262,
0.0040948763,
-0.0088254521,
-0.0872648656,
0.011915545,
-0.0039073536,
-0.016640529,
0.0077792997,
0.0188056007,
-0.0576307438,
0.0103921667,
0.041251719,
0.0543985665,
-0.0237232074,
-0.0410910696,
-0.0097354166,
0.0692351386,
0.0011827737,
0.0381810293,
0.0292881038,
-0.0014860885,
-0.0103642512,
0.0238056928,
0.0474845059,
0.0325802937,
-0.0657930896,
0.0223078094,
0.006813548,
-0.025772566,
-0.0141301118,
0.0198384188,
0.0100762714,
-0.04639332,
-0.0543316975,
0.0660577342,
0.0090376204,
-0.0696972013,
0.0039120126,
-0.0065390053,
-0.0323533751,
0.0245990697,
0.0094768843,
-0.0210738722,
0.0198608898,
-0.0131748728,
-0.0517571755,
0.0350322537,
-0.009263631,
0.0059770346,
-0.0382262543,
-0.0025018768,
0.0140762348,
-0.0429100096,
0.0257833041,
-0.0289864801,
0.0076758941,
0.0077140457,
0.0059517319,
0.0268252734,
-0.0084211482,
-0.0337006561,
0.0358399637,
0.0235937163,
0.014482379,
0.0337241404,
0.0003974261,
0.0766828656,
0.0488595292,
0.0335753784,
-0.0346556157,
-0.0747267827,
0.0261744075,
0.0173925608,
0.0495253764,
0.0083776461,
0.0379387476,
0.0150914704,
-0.0076745972,
-0.0202851295,
0.0069039059,
-0.0556404777,
-0.0470468141,
-0.0438884012,
0.0850804597,
-0.0486850776,
-0.0238248911,
-0.001707209,
0.027753979,
-0.0143016642,
0.0440849736,
0.0414991975,
-0.0501485281,
-0.0143562518,
0.0060527129,
0.042534627,
-0.0047214935,
-0.091400601,
0.0139671071,
0.0252786614,
-0.0391439274,
-0.0443834811,
-0.0036701134,
-0.019684257,
0.0186530557,
0.0421439335,
-0.0059166658,
0.0581831001,
-0.0433207154,
-0.0151386503,
0.0353189334,
-0.0244084951,
-0.0423617475,
-0.076754719,
-0.0032578127,
0.0338124968,
-0.0421757922,
-0.018484246,
0.0300523303,
-0.0816884488,
-0.0468147174,
0.1068646684,
-0.0007074999,
0.0183280054,
0.0238882154,
0.0086101182,
0.026864225,
0.0890229344,
0.0379479267,
-0.0218692161,
0.0656180903,
-0.0037797506,
0.0123671098,
-0.014489999,
0.0018715396,
-0.0101143429,
0.0445857272,
0.0118162371,
0.0197364129,
-0.0041849911,
0.0184619892,
0.0461800061,
-0.0741404891,
0.1231181547,
-0.0017036422,
0.0555056855,
0.0148551185,
-0.0221934933,
0.0248243865,
0.0807501972,
-0.0153970681,
-0.0533855855,
0.0130323758,
0.0374376662,
0.0114966268,
0.0008124583,
0.0130916787,
-0.0186423063,
0.0182480011,
-0.0078470288,
-0.054940287,
-0.0407076478,
-0.0400133692,
0.0398037955,
0.0157814845,
0.027200643,
-0.0133650973,
0.0028626546,
-0.0010779646,
0.0372100249,
0.0451049916,
0.0250568036,
-0.0316066965,
0.0075963801,
-0.113631472,
0.0224651638,
0.0538143553,
-0.0073753386,
0.0787312165,
-0.0097981477,
0.0178489108,
0.0021824751,
0.0109689012,
0.0069554378,
0.0183090977,
-0.0124717355,
0.0010734436,
0.000572664,
0.0126698222,
0.001417384,
-0.0186157487,
0.0497574955,
-0.0737521052,
0.0217210855,
0.0308614913,
0.0169001799,
-0.0251494292,
-0.0122750383,
-0.0354326181,
-0.0451805927,
-0.0372576416,
-0.0134916594,
-0.0165074226,
-0.0475211218,
0.0562086478,
-0.0085212635,
0.03312793,
-0.0425067432,
0.0192490071,
0.0199910104,
0.013248397,
-0.0319570936,
0.032956928,
0.0356463119,
0.0323741622,
-0.0572884083,
-0.0352235511,
-6.124719726e-33,
-0.030223351,
-0.0431061685,
0.025903644,
0.0461919941,
-0.0394379012,
0.015702229,
-0.0521391928,
0.0065450268,
0.0211122036,
-0.0081738625,
-0.0553073809,
0.0217022728,
0.0032871347,
0.0155630987,
0.0713978186,
-0.0310804751,
-0.0035239051,
-0.0316995867,
0.0432791933,
0.0045962669,
0.0672501773,
0.0311052017,
0.0900684372,
-0.0548183918,
-0.0313263983,
0.0339892209,
0.0367798284,
0.0325905718,
-0.0112845758,
0.0177358184,
-0.063998431,
0.0290240962,
0.0321898013,
-0.0587066971,
0.0030357733,
0.0155646941,
-0.0783881024,
-0.0457955226,
-0.0213819891,
-0.0291800946,
-0.0427539051,
-0.0373439863,
0.0688878521,
-0.0418334976,
-0.0058007673,
-0.0113181639,
0.0268006697,
-0.0141847748,
-0.0424712412,
-0.0208024811,
-0.0407837145,
-0.0119348932,
-0.013228395,
-0.0109746614,
0.0191520676,
0.0020665531,
-0.0314602591,
0.0349478796,
-0.0670743063,
0.0323282368,
-0.0179353654,
0.0611167066,
0.021446906,
0.0259260256,
0.0460214503,
0.0349008776,
0.0659958199,
-0.0352411158,
-0.0552764945,
-0.0391583554,
0.0455131456,
-0.0127970418,
-0.006241262,
-0.0197871644,
0.045173306,
-0.0242281295,
-0.0436471589,
-0.0232350603,
-0.0051233508,
0.0309081003,
0.0150396712,
-0.0033617592,
0.0383040048,
0.0220793914,
0.0178068783,
-0.0040003075,
-0.0342663974,
-0.0129976068,
-0.0275252573,
-0.019991735,
0.0679361299,
0.0105115073,
0.0515044518,
-0.0241179522,
0.0280423947,
-0.0004441095,
0.0287511721,
-0.0138627309,
0.0383037142,
0.0030060355,
-0.0147772292,
-0.0101623535,
0.0012700574,
0.0059953541,
0.0399109572,
-0.0281277765,
0.0194206554,
0.0424170159,
-0.0318279043,
-0.0366089083,
-0.0210653264,
-0.0278675612,
0.0263619181,
-0.0349501818,
0.0581606925,
0.0077295112,
-0.0145608103,
-0.0118787186,
0.0104059121,
0.0407384448,
-0.041062966,
0.0453978404,
-0.0239511803,
0.0553524606,
-0.0094082719,
0.0134330653,
-0.0707045123,
0.0707409382,
0.0471823476,
-0.0626078397,
-0.0481890887,
0.0237100646,
2.827e-7,
0.0055927751,
0.0597541556,
0.0392128453,
0.0136037441,
0.0800219774,
-0.0219510607,
0.0054294607,
0.0234356094,
-0.0025143351,
-0.0328659154,
0.0237882063,
0.0482076332,
0.0173793528,
0.0071380404,
-0.0324823521,
-0.0491771549,
-0.0609686673,
-0.0140020214,
-0.0191789586,
-0.0199937616,
0.031435594,
0.0474054702,
0.0403866097,
-0.0152783534,
0.0233961381,
-0.0005231158,
0.0772172064,
-0.0072203721,
0.0256547928,
-0.0031155797,
-0.035620071,
-0.0736439303,
-0.0285566188,
0.0088689588,
0.0223943908,
-0.0077815712,
-0.0357168652,
-0.0089858128,
-0.0152011774,
-0.0536825806,
-0.0035836496,
0.0344312973,
-0.0213672835,
-0.0103702527,
0.0318490267,
-0.1344875395,
-0.0566630289,
-0.0066221179,
0.0065908674,
0.0123876333,
0.0325392038,
0.032117378,
-0.0158366486,
-0.0302469227,
-0.0215333384,
0.0232597049,
-0.0018292386,
-0.0720610619,
-0.0292965155,
0.0250696503,
-0.0690379813,
0.015714895,
-0.0577418394,
-0.0459079817,
0.0694147572,
-0.0083331242,
-0.0279730819,
1.927251916e-34,
-0.0319013447,
-0.0180731025,
0.0176222362,
-0.0104545513,
-0.0258027017,
-0.0344190001,
0.0064630788,
-0.0300169792,
0.0162396897,
-0.0527378581,
-0.0307237674
] |
http://arxiv.org/abs/2401.14242v1 | http://arxiv.org/pdf/2401.14242v1 | Improving Natural Language Capability of Code Large Language Model | 2024-01-25T15:33:20 | 2024-01-25T15:33:20 | Wei Li, Daoguang Zan, Bei Guan, Ailun Yu, Xiaolin Chen, Yongji Wang | cs.CL | cs.CL | Code large language models (Code LLMs) have demonstrated remarkable
performance in code generation. Nonetheless, most existing works focus on
boosting code LLMs from the perspective of programming capabilities, while
their natural language capabilities receive less attention. To fill this gap,
we thus propose a novel framework, comprising two modules: AttentionExtractor,
which is responsible for extracting key phrases from the user's natural
language requirements, and AttentionCoder, which leverages these extracted
phrases to generate target code to solve the requirement. This framework
pioneers an innovative idea by seamlessly integrating code LLMs with
traditional natural language processing tools. To validate the effectiveness of
the framework, we craft a new code generation benchmark, called MultiNL-H,
covering five natural languages. Extensive experimental results demonstrate the
effectiveness of our proposed framework. | Improving Natural Language Capability of Code Large Language Model: Code large language models (Code LLMs) have demonstrated remarkable
performance in code generation. Nonetheless, most existing works focus on
boosting code LLMs from the perspective of programming capabilities, while
their natural language capabilities receive less attention. To fill this gap,
we thus propose a novel framework, comprising two modules: AttentionExtractor,
which is responsible for extracting key phrases from the user's natural
language requirements, and AttentionCoder, which leverages these extracted
phrases to generate target code to solve the requirement. This framework
pioneers an innovative idea by seamlessly integrating code LLMs with
traditional natural language processing tools. To validate the effectiveness of
the framework, we craft a new code generation benchmark, called MultiNL-H,
covering five natural languages. Extensive experimental results demonstrate the
effectiveness of our proposed framework. | [
0.0487618521,
0.0679789111,
-0.0073911394,
0.0430444367,
-0.029730998,
-0.0100946352,
-0.0446666703,
-0.0220953505,
0.0151680186,
-0.0209152251,
-0.0181314759,
0.0115651926,
-0.018728327,
0.0700920299,
0.0575749241,
-0.0145412087,
0.011206178,
-0.029408969,
-0.0307792258,
0.0454694368,
0.0022517301,
0.0284413695,
0.0252700634,
0.0626702458,
0.0370900333,
-0.019568352,
-0.0121520599,
0.0076090996,
0.0024755464,
-0.0576707125,
0.0291019771,
0.0425803922,
-0.0148042152,
-0.0328478999,
0.0000019832,
-0.0636213869,
0.0022977239,
-0.0142822359,
0.0299738813,
0.0392953157,
0.0296847895,
-0.0007736324,
-0.0003201706,
0.0295751262,
-0.0484923348,
0.0325252824,
0.0268651601,
-0.0113890702,
0.0864174813,
0.0856645107,
-0.0018147621,
-0.0256421138,
0.0162295271,
-0.0078656441,
0.0546728447,
-0.0463377349,
-0.0029355786,
0.017542284,
0.0495334342,
0.0050371648,
0.0162314437,
0.0142301004,
0.0448327214,
-0.0268480927,
0.0401681252,
0.0186146628,
-0.0422036313,
-0.0104427421,
-0.0155170718,
0.0575626567,
-0.0749701709,
0.0060844431,
0.0298860911,
-0.0091304677,
-0.0149699012,
-0.065609701,
-0.0220043063,
0.0231455173,
-0.0183285214,
-0.0099213636,
0.0408849306,
-0.0074017677,
0.0038606962,
0.0125990603,
-0.0225888919,
0.0320169143,
0.0195374545,
-0.0049490174,
-0.0145051433,
-0.0394410528,
0.0095849633,
-0.0637603998,
0.0171606578,
0.0757614225,
-0.030353507,
0.0037382978,
-0.0415419936,
-0.013056566,
0.0387207381,
-0.0472056381,
0.0351931937,
0.0367235988,
-0.0112888757,
0.0146765886,
-0.0444955416,
0.0642918795,
-0.0249299612,
-0.0461442545,
-0.036397431,
0.0045745652,
-0.0545238107,
-0.0077677597,
-0.0537327603,
0.0394620225,
-0.0202020835,
-0.0093969787,
-0.029900996,
0.0538552999,
0.010785237,
0.0311482158,
-0.0064656776,
0.0232410487,
-0.0495054573,
-0.0080737351,
0.0018565053,
0.0105382781,
-0.0294621363,
-0.0625725314,
0.0051332461,
-0.076024659,
0.013321979,
0.0483981706,
-0.0126931509,
-0.024502201,
0.0192123745,
0.0527564473,
0.0073988535,
0.0111754583,
-0.0542164594,
-0.0262587648,
-0.0057862531,
-0.0392439961,
-0.0056001232,
-0.0943381712,
-0.0257832836,
0.0226039402,
-0.0300624538,
-0.0654848963,
0.0215051323,
0.0072265076,
0.0282014925,
0.0238591004,
-0.054325588,
-0.0029713975,
0.0054597449,
-0.0016248948,
0.0140192332,
0.0209320001,
0.0537804887,
0.016731577,
-0.0019601926,
-0.0265825819,
0.0361669064,
0.0366519094,
-0.0263636932,
-0.0264632348,
-0.0034786507,
0.0386075154,
-0.0175319202,
0.0037837892,
0.0201126896,
0.0460200161,
0.0157985594,
0.0618498139,
-0.0037225941,
0.0406714901,
0.0504519679,
0.0666200072,
0.0810380131,
0.0508309156,
-0.0222414248,
-0.0252794493,
-0.0319218971,
0.0792558044,
-0.016155092,
0.0378554277,
0.0233005024,
0.0591417924,
-0.00065767,
-0.0738908574,
0.0025756119,
-0.0201128963,
-0.0056254454,
-0.0087400507,
0.0209750794,
0.0130260382,
-0.0588202402,
0.0415770747,
0.002549259,
-0.0623759516,
0.0243946929,
-0.0305717885,
0.0101973275,
0.0770226568,
0.0424548872,
-0.0323413387,
-0.0291338284,
-0.0155295841,
-0.0336250775,
0.01625726,
0.0358388536,
0.015968075,
-0.0014583349,
-0.0544486754,
-0.0079679145,
-0.0602011345,
-0.0131570315,
0.0158943497,
0.032567203,
-0.0186557379,
0.0558996573,
0.026413599,
-0.0767985359,
0.0253979675,
-0.0080484068,
0.0090579567,
0.0079337591,
-0.0213386621,
0.0250233039,
-0.0145015875,
-0.0304662921,
0.0424021706,
0.0084518949,
-0.0453345515,
-0.0225252211,
-0.0133579364,
-0.0261820164,
-0.0504404604,
0.0274590366,
-0.0597770661,
-0.0190678742,
0.0356771424,
-0.012006267,
-0.0252211485,
0.0425827093,
-0.0512183942,
0.0076519269,
-0.0140564004,
0.0035502478,
0.0051583825,
-0.0118604349,
-0.017578423,
0.0692336336,
0.0035853956,
0.0272746831,
0.0095623964,
0.0527289435,
-0.0092551662,
0.0083285086,
-0.0167416912,
0.0150820045,
-0.0503676124,
0.0476008952,
0.0386581384,
0.0741093382,
-0.0430433564,
-0.0555858761,
-0.0503268838,
0.0535988845,
-0.0224173199,
0.0197281595,
0.0192013048,
0.0398310609,
0.005904465,
0.006999603,
-0.0457373634,
-0.0230858922,
0.0017760467,
-0.0320175737,
0.0332665257,
-0.0236913972,
-0.0492410101,
-0.0265453942,
0.0350454859,
-0.0325033814,
0.0227320362,
0.0219366923,
-0.0407075137,
-0.0686555356,
0.0859771371,
-0.0284486376,
-0.0117007773,
-0.005820903,
-0.0441633761,
0.0062226988,
-0.0268320069,
0.0790355206,
0.0391225927,
0.0107599553,
-0.009146384,
0.0363504216,
-0.0072014951,
-0.0345014483,
0.0342496634,
0.0149349272,
-0.0363070369,
-0.0136470562,
-0.0342180245,
0.0922120959,
0.05985393,
-0.0059699193,
-0.0036744326,
-0.038435135,
0.0088263256,
0.0073605306,
0.0286506228,
0.0284470115,
0.0117747737,
-0.0298766289,
0.0386964828,
-0.019793354,
-0.0725892708,
0.059838824,
-0.0308211446,
-0.0363019221,
-0.0436378047,
-0.0197870862,
-0.0708491504,
0.0425014049,
0.0042962404,
0.0321969464,
-0.0013460717,
-0.0575636216,
0.0018657499,
0.0128728542,
-0.0363845862,
-0.0269419514,
0.0751927122,
-0.0388646238,
0.0093875527,
-0.0038764656,
-0.0087926956,
-0.0471679792,
-0.0023004534,
-0.0160077512,
0.005161365,
0.0452213846,
-0.0044013984,
0.0110502504,
-0.0104971044,
0.0107192826,
0.0422452502,
0.0101801865,
-0.0261230636,
-0.0054613217,
-0.0250541754,
0.0341057554,
-0.0261028353,
-0.0197346844,
0.0014442926,
-0.0014175618,
0.0171032138,
0.0151112573,
0.0514165089,
-0.074401699,
0.0220532212,
0.0105804838,
-0.0283567607,
-0.0061604823,
0.0752347633,
0.0429107733,
-0.0600183047,
-0.0017525604,
0.1171728224,
0.0393242165,
-0.0241043158,
0.0177749153,
-0.0279198326,
0.039188765,
0.088713415,
0.0574808419,
-0.0532916449,
-0.0483299308,
-0.0290206242,
0.0286758263,
0.0053355047,
-0.0253332462,
-0.0514008142,
-0.1054601297,
0.0247591734,
0.0192396659,
-0.0220167283,
-0.0590765961,
0.0016947725,
-0.0263813846,
0.028289929,
-0.0376750045,
0.0241824146,
-0.0074500837,
-0.0039415476,
-0.0080709476,
-0.0694008023,
-0.0282476172,
0.0161327645,
-0.0518578477,
0.0155480905,
0.1069343463,
0.0165888444,
0.01927034,
-0.0035902562,
-0.0041605881,
-0.0097246701,
0.0422274582,
-0.0261272956,
0.0217801817,
-0.0157893989,
-0.022684969,
-0.0038323079,
0.0361151882,
-0.0788506046,
0.0002260968,
0.0631606057,
0.0559585318,
0.0082982816,
-0.0141984727,
-0.0221342072,
0.0226822235,
-0.0090273861,
0.0530690365,
0.0565125905,
-0.0584353879,
-0.0346379615,
-0.0652168989,
-0.0257247575,
-0.0809896514,
-0.0300035365,
-0.0009010201,
-0.0059902305,
0.0084275706,
-0.0487089083,
-0.0152264778,
0.0403396785,
0.0320196822,
0.0192566607,
-0.0128697492,
0.03760067,
-0.0242658164,
-0.0483383313,
0.0380049981,
-0.0199762322,
0.0064651202,
-0.0399617441,
-0.007238118,
0.0217252113,
-0.0185459554,
0.0081738653,
0.036755085,
-0.0764145255,
-0.0631504208,
-0.0537721738,
0.0545784831,
0.0158451293,
0.0148509843,
-0.0291763078,
-0.0571194068,
-0.0665872991,
-0.0073774741,
-0.0522265509,
-0.0125777116,
0.0132088745,
0.0338134281,
0.0229800157,
0.034842167,
-0.0572392419,
-0.0030470102,
-0.0032511682,
0.0515382551,
-0.0029808041,
0.0031985079,
0.0442728177,
0.0354302004,
-0.0239249617,
-0.0033690459,
-0.012391449,
-0.0055661816,
-0.0178653765,
0.0195357744,
-0.0631223693,
-0.057450112,
0.0316999219,
0.0147757987,
-0.032093849,
0.0397438556,
0.0259201303,
-0.0279608443,
0.0019316445,
-0.0008656455,
0.0601300709,
-0.0013672325,
0.0418495052,
0.052773416,
-0.0212853253,
-0.0377322175,
0.0134111941,
-0.0223193839,
0.0232605413,
-0.0001607513,
0.0274463091,
-0.0050850301,
0.0206058063,
0.0106016695,
0.0041940627,
-0.0221008714,
-0.0788548887,
-0.0141917495,
-0.0185492411,
0.0527637601,
0.0498654917,
0.0301485527,
-0.0118813356,
0.0210527126,
-0.0226029381,
0.0799960122,
-0.0175367855,
-0.0700281486,
0.0078714481,
0.0201819595,
-0.018741874,
-0.0701883882,
0.085434854,
-0.0079380516,
-0.0056459564,
-0.0027553185,
0.0002583016,
-0.0773753673,
0.0051860623,
0.0398972519,
-0.0146699492,
-0.0057032262,
0.0129861431,
-0.025867844,
-0.050680384,
0.0407889597,
0.0119537488,
0.0141436988,
0.0008609006,
0.0540951975,
0.0600120388,
-0.0190178305,
-0.0378835835,
-0.0196042955,
0.0147469137,
-0.019619802,
-0.0420742705,
-0.0327155627,
-6.087489248e-33,
-0.0348506272,
-0.0358162746,
0.0203034319,
0.0128599601,
0.0016939699,
0.0018417162,
-0.0754850507,
-0.00061346,
0.0185835958,
-0.0297779907,
0.0416466184,
0.019732343,
-0.0208538864,
-0.0076163304,
0.0378187709,
0.0041228258,
0.0496127568,
0.0211999435,
-0.0463849306,
0.0329870507,
-0.0291405749,
0.0282607824,
0.0360503532,
-0.0066086561,
-0.0291401669,
-0.0157260373,
0.0126173906,
0.0140238591,
0.0349581763,
0.048531428,
-0.0023212461,
0.0188376568,
0.0066308277,
-0.0412121415,
-0.0452554338,
0.000846629,
0.008098538,
-0.026988823,
0.0232866649,
-0.0182417445,
-0.0154695427,
-0.0849569812,
0.0783041716,
-0.0165082924,
-0.0020900329,
0.009113472,
0.0420938469,
-0.0042850543,
0.0092667965,
-0.0484756678,
-0.093033731,
0.0413628444,
-0.0613438971,
0.009718216,
-0.0407932289,
0.0553856753,
-0.0219625849,
0.0404309966,
0.0190298241,
-0.0102227163,
0.0225471705,
0.0657449588,
0.0843855143,
-0.0031072204,
-0.0001118085,
0.03409563,
0.016802635,
0.0054339818,
-0.0199630037,
0.0216498096,
0.0163812041,
0.095387727,
0.0308760162,
0.0154049797,
0.0161558092,
0.0072318385,
-0.0247770529,
0.0261855349,
-0.0093564009,
0.0473252311,
0.0288297981,
0.0130187329,
-0.0262352172,
-0.0106951408,
-0.0064329696,
0.0107922507,
-0.0297389869,
-0.0325892679,
0.0568713434,
-0.0007002194,
-0.033148095,
0.015250193,
0.0164474882,
-0.0239789262,
0.0173604488,
0.0537057705,
-0.0338246375,
0.0383273475,
-0.0102586038,
0.0342234597,
-0.0265845768,
-0.0036456478,
0.0079102637,
-0.0212954525,
0.0564278662,
0.013768699,
0.0096113272,
0.005724459,
-0.0643524006,
-0.0064281099,
-0.064687103,
0.0014816186,
-0.0200603828,
0.023570152,
0.0143045504,
-0.0232239682,
-0.0134004196,
0.041813653,
-0.0035636476,
-0.0523042344,
-0.0572112426,
0.0044647027,
0.0474666171,
0.0409709476,
-0.0306018423,
-0.0218897667,
-0.0662397146,
0.0328950435,
0.0312938429,
-0.0099203512,
-0.0586827546,
0.0350285098,
2.785e-7,
0.0184085052,
0.0437255315,
-0.0054272991,
0.0463468023,
-0.003627904,
0.0175641291,
-0.0158688035,
-0.0097567644,
0.0062571084,
0.0112449499,
-0.0452490523,
0.0201667715,
0.0303169042,
0.0085674617,
-0.056318596,
0.026782766,
0.006109105,
0.0078000575,
-0.0761555061,
0.0038740775,
0.0692691952,
0.0458965637,
-0.0265017133,
0.0119827576,
0.0193587597,
-0.0610044114,
0.0077030761,
-0.007470028,
-0.0021145104,
0.0510111861,
-0.0365876444,
0.0396539532,
-0.0402671099,
-0.0291897375,
-0.0057547209,
0.0082322033,
-0.0072931685,
0.0793616176,
-0.0076071667,
0.0135265365,
0.0354288481,
-0.0082092043,
-0.0014979552,
-0.0649531037,
0.0211486258,
-0.0563704669,
-0.0164098684,
-0.0227042176,
-0.0058664861,
0.0123228803,
-0.0313635655,
-0.0349834561,
0.0183791649,
-0.0132721001,
0.0021084188,
-0.060144715,
-0.0557267629,
-0.0319176726,
-0.0317771137,
0.0177516118,
-0.0136204334,
-0.0337974429,
-0.0810074285,
0.0241083167,
0.0218042973,
-0.0295445882,
0.0243620109,
2.69289983e-34,
0.0105660837,
-0.0278625041,
0.0215635672,
0.0254471265,
0.086249575,
0.0035004292,
-0.0614685044,
0.0244319271,
-0.0025246639,
-0.0359847434,
-0.0403706245
] |
http://arxiv.org/abs/2401.14240v1 | http://arxiv.org/pdf/2401.14240v1 | Enhanced Labeling Technique for Reddit Text and Fine-Tuned Longformer Models for Classifying Depression Severity in English and Luganda | 2024-01-25T15:28:07 | 2024-01-25T15:28:07 | Richard Kimera, Daniela N. Rim, Joseph Kirabira, Ubong Godwin Udomah, Heeyoul Choi | cs.CL | cs.CL, cs.LG | Depression is a global burden and one of the most challenging mental health
conditions to control. Experts can detect its severity early using the Beck
Depression Inventory (BDI) questionnaire, administer appropriate medication to
patients, and impede its progression. Due to the fear of potential
stigmatization, many patients turn to social media platforms like Reddit for
advice and assistance at various stages of their journey. This research
extracts text from Reddit to facilitate the diagnostic process. It employs a
proposed labeling approach to categorize the text and subsequently fine-tunes
the Longformer model. The model's performance is compared against baseline
models, including Naive Bayes, Random Forest, Support Vector Machines, and
Gradient Boosting. Our findings reveal that the Longformer model outperforms
the baseline models in both English (48%) and Luganda (45%) languages on a
custom-made dataset. | Enhanced Labeling Technique for Reddit Text and Fine-Tuned Longformer Models for Classifying Depression Severity in English and Luganda: Depression is a global burden and one of the most challenging mental health
conditions to control. Experts can detect its severity early using the Beck
Depression Inventory (BDI) questionnaire, administer appropriate medication to
patients, and impede its progression. Due to the fear of potential
stigmatization, many patients turn to social media platforms like Reddit for
advice and assistance at various stages of their journey. This research
extracts text from Reddit to facilitate the diagnostic process. It employs a
proposed labeling approach to categorize the text and subsequently fine-tunes
the Longformer model. The model's performance is compared against baseline
models, including Naive Bayes, Random Forest, Support Vector Machines, and
Gradient Boosting. Our findings reveal that the Longformer model outperforms
the baseline models in both English (48%) and Luganda (45%) languages on a
custom-made dataset. | [
0.0842507854,
0.0563061871,
-0.023470385,
-0.0050104656,
-0.000132537,
0.0298308115,
0.0164741613,
0.032938052,
0.0656261146,
-0.0857269317,
0.0303806383,
-0.0180310328,
-0.0321500562,
0.0213415194,
-0.0369635783,
-0.0620087087,
-0.0057248645,
0.0361434333,
0.0545610525,
-0.0095101418,
0.0377372988,
0.0253306963,
-0.0090649519,
0.0287446585,
-0.0083633047,
-0.0079841893,
0.0633360222,
-0.0324016698,
0.0157848839,
-0.0374581367,
0.0393634811,
0.0574769787,
-0.040434964,
0.0265001412,
0.0000018878,
-0.035790544,
-0.0209122747,
-0.0162649993,
0.0189233925,
-0.0582236759,
0.043028798,
-0.0478296131,
0.0251156874,
0.0390976481,
-0.0415540189,
-0.0452459008,
0.0524968728,
0.0848066881,
0.054375995,
0.0694757774,
-0.0298435017,
-0.0497248881,
0.0038698204,
-0.0101711955,
0.0144640636,
-0.0194110032,
0.0639666915,
0.0098663131,
0.0140620647,
-0.0313212462,
-0.0105484761,
0.0196498279,
-0.0202247631,
0.0239456352,
0.011086667,
0.007788159,
0.0166639742,
-0.0564101972,
0.0204141643,
0.0775935352,
0.0107561294,
-0.0010136847,
0.0192481168,
0.0074232263,
0.0028235931,
0.002240584,
-0.0264156945,
0.0087223612,
-0.0490483716,
0.0292570423,
0.052562356,
0.007930425,
0.0257686395,
0.0089014219,
0.026106691,
0.0173606388,
-0.0188092887,
-0.0273957849,
0.0150803318,
0.0054712794,
-0.0331716537,
-0.0478605777,
0.0596174002,
0.0069189603,
0.0211373083,
-0.000337752,
-0.0199448802,
-0.0867038742,
0.0256245174,
-0.0299426951,
-0.0308261197,
0.0067748399,
-0.03585127,
0.0197991114,
-0.0038291696,
-0.009796802,
0.0456222743,
-0.015578175,
0.0107047334,
0.0011568629,
-0.0742458925,
0.0173695013,
-0.0484152995,
0.001842903,
0.0076883817,
-0.0261554159,
-0.0582110733,
0.0051321615,
-0.0070768362,
-0.0520563982,
0.0182017013,
-0.0372439399,
0.0222215112,
0.0279629957,
0.0244813841,
0.0444129184,
-0.0622716695,
-0.0058592586,
-0.0047794008,
-0.0272872411,
0.0657187924,
-0.0053847614,
0.0167600438,
-0.0411680974,
0.0149665847,
0.0091073541,
0.0468075052,
-0.0268432628,
-0.0174073149,
-0.0248699971,
-0.0589775853,
-0.0191230625,
-0.018088473,
-0.0582251027,
0.0137674296,
0.0478586182,
0.011329053,
-0.0321834721,
0.021530943,
-0.0111729093,
-0.0434111357,
0.0623971149,
-0.0850839242,
0.0206472669,
0.0517299585,
0.0328933261,
0.0046963189,
0.0504309684,
-0.0149851562,
0.0400185362,
0.0650419146,
-0.043620538,
0.0398536287,
0.014575012,
0.0006045345,
-0.0466860421,
-0.0052403174,
-0.0112469345,
-0.0312448665,
-0.0052890605,
-0.0333674401,
0.0087766144,
0.0066710152,
0.0219487064,
0.0430330597,
0.0349302627,
0.0218185317,
0.0514888316,
-0.026644025,
0.0288098082,
-0.0087862462,
-0.0282697845,
-0.0281833913,
-0.0069499975,
-0.0298694689,
-0.0289547145,
0.0172650814,
0.0280880351,
0.001227684,
-0.0557307564,
-0.0661623552,
-0.0315338001,
0.0094243046,
-0.0058006924,
0.0256093536,
-0.0036499253,
-0.0065824343,
0.0507446676,
0.0419609658,
-0.0108142616,
-0.030495787,
-0.0596000142,
0.0502226502,
0.0349108018,
-0.0301799979,
-0.0646778494,
0.0140097653,
-0.0187078808,
0.0714521483,
0.0059015411,
0.0282821245,
0.021794647,
-0.0073319068,
-0.014854583,
-0.0252929665,
0.0310832728,
0.0305155888,
-0.0365122147,
-0.0097822165,
0.0447796024,
-0.0172822345,
-0.0093290145,
-0.0338215791,
0.0418165959,
0.0757113099,
-0.0167539157,
0.017273467,
0.011753344,
0.0313889831,
0.0106945159,
0.0033877781,
-0.0131222606,
0.0074868063,
-0.0191102009,
-0.0755000785,
-0.0430132747,
0.0199998952,
-0.0316203795,
0.0089344764,
-0.0192701239,
-0.0268457234,
0.0816874132,
-0.0420256443,
-0.0058092102,
0.0395989604,
-0.0421334393,
0.0224355552,
0.018511124,
-0.034672983,
-0.0398262516,
0.0339415781,
-0.014406085,
0.0381578058,
-0.0096482914,
0.0352220237,
-0.0462423638,
-0.0952266902,
-0.0316267349,
0.001800009,
-0.048914317,
-0.0134497965,
-0.0291353278,
-0.0006486359,
0.0213288721,
-0.0083666006,
-0.0207095109,
-0.0326995403,
-0.0964764953,
0.0056088828,
-0.0341000296,
0.0187215526,
0.0225837808,
0.0044858907,
-0.016792357,
-0.0216775015,
-0.0231609326,
0.0169581752,
0.0138630215,
0.0265447348,
0.0133370562,
0.0170688257,
-0.0002400021,
-0.0383505411,
-0.0197817348,
-0.0175208114,
0.0043744622,
0.0434406996,
-0.1095296964,
-0.0299517475,
-0.0120861717,
-0.04662925,
-0.0202676039,
0.0014763011,
-0.0595518425,
-0.0262990203,
-0.0030096127,
0.033508908,
0.0448647141,
0.0354494788,
0.0402665585,
0.0198658053,
0.0498338081,
0.011338369,
-0.0124257253,
-0.0532243736,
-0.0240736883,
-0.038516514,
-0.0589416139,
-0.0142888362,
0.0615056828,
0.0353444628,
0.0093229478,
-0.0209079031,
0.0035874057,
-0.0413909182,
0.0184618589,
-0.0159838125,
0.015720984,
0.0061370917,
0.0180143956,
-0.0122589879,
-0.0574520864,
0.0514922589,
0.0306445584,
0.0022407698,
-0.0207632799,
-0.0591625683,
-0.0304231718,
0.0055034379,
0.0256216954,
0.0154175991,
0.007599066,
-0.0247608125,
0.0573232174,
-0.0012756737,
-0.0145239299,
-0.033629369,
0.0188529175,
0.007423202,
0.0005894713,
0.025937343,
-0.0314675197,
-0.0208196081,
-0.0933715925,
-0.0558280982,
-0.0383618362,
-0.0069082282,
0.004432275,
-0.0295154154,
0.0268730391,
-0.0025453088,
0.0164614469,
-0.0484385937,
-0.0310120173,
0.0375075042,
-0.0051528495,
0.0621005744,
-0.0552228987,
0.0480754413,
-0.0180519633,
0.0361017175,
0.0617185757,
-0.0471555367,
0.0408343375,
0.0096122753,
-0.0290257856,
-0.0016439718,
0.0008283149,
-0.004457816,
-0.0081491908,
0.0146126505,
-0.0169133525,
0.0071060481,
0.1095222607,
-0.0066223866,
-0.0438752808,
0.0077815899,
-0.0461093374,
-0.0102352258,
0.027710041,
-0.0118168211,
-0.0593204759,
-0.0634311438,
-0.0383923203,
0.0520328842,
-0.027336441,
-0.0230643842,
0.0423130803,
-0.0386395641,
0.0221791696,
-0.0238760076,
-0.0458661653,
0.0367775187,
-0.0535418168,
0.0091569405,
-0.0248663519,
-0.0312225521,
-0.0452974811,
-0.0173140112,
0.0083336644,
0.0126545057,
-0.1151687875,
-0.0452017784,
0.0181147866,
-0.0033335632,
0.0216833651,
0.0147264842,
0.0540412031,
-0.0057894015,
0.0063211014,
0.0402197726,
0.0218418688,
-0.0213800389,
-0.021263063,
0.0896012783,
-0.0002910748,
0.0010397038,
-0.0615274385,
0.0568176508,
-0.0575249381,
0.0394850336,
0.0324721597,
0.0750635415,
0.0069570406,
-0.0378644168,
0.0033843904,
0.028027745,
0.0535996482,
0.0484555066,
0.0515648648,
-0.04137519,
-0.0085108783,
0.0191804431,
-0.0032434978,
0.0029875902,
-0.0524406247,
0.0062432089,
-0.0374653712,
-0.0655557066,
-0.0729570985,
-0.0231624395,
0.0356953666,
-0.0334198512,
-0.0238490961,
0.0019491131,
-0.0305728838,
0.0109360134,
-0.0784835592,
0.0616972074,
0.0354017392,
0.019755451,
-0.0431107394,
-0.008021784,
0.043538738,
-0.0188081041,
-0.0104231173,
0.0139897401,
-0.0378863066,
-0.0176969934,
0.0499782227,
0.0035392467,
0.0101716556,
-0.0237135831,
-0.0036328856,
-0.0787756816,
0.0166143347,
0.0398340523,
0.014199567,
-0.0198980048,
-0.0077471975,
-0.0131264031,
0.0102121644,
0.034327589,
-0.0238571074,
0.0593517832,
0.0246203002,
-0.0618529879,
-0.035805352,
0.0024073273,
-0.0307918154,
0.0325395726,
0.0494752638,
-0.0073964628,
-0.0590218604,
-0.0007068243,
-0.0570471138,
0.0209150966,
-0.0140543915,
-0.0129783591,
-0.0163308121,
0.0460240208,
0.0123874797,
-0.0352608599,
0.002828931,
-0.0150490012,
-0.022591114,
-0.0893989727,
0.054339692,
0.015553019,
0.037419457,
0.0200239401,
0.0123202885,
-0.0111441668,
-0.0028018097,
0.0480550826,
0.0035823931,
-0.0101941125,
0.0727448985,
-0.0093012862,
0.0010883802,
0.0431540273,
0.0340231024,
-0.0334221795,
-0.0841414034,
-0.0106541561,
-0.006784989,
0.0420495011,
-0.0172659382,
0.077977784,
0.0079292841,
-0.0242008436,
-0.0054019298,
0.0437607281,
-0.0310279597,
-0.0252588522,
0.0191059411,
0.0008706572,
-0.0289934222,
-0.0080803689,
0.042930197,
-0.0115588,
0.0032153854,
0.0486576483,
-0.0549988821,
0.0032498662,
0.0389343984,
0.0088150995,
0.0286909416,
-0.0002850329,
-0.0018644341,
-0.0140276393,
-0.0583236404,
0.0339967459,
-0.0610072352,
0.0042192913,
-0.0017529157,
0.0057622343,
0.0453162342,
0.0272651929,
-0.0024956341,
-0.0751945153,
0.0401169918,
-0.0270987656,
-0.0254709274,
-0.0430950411,
-5.9821352e-33,
0.0331927985,
-0.0337765478,
-0.0487348288,
0.0106362263,
-0.0081796562,
0.04376252,
0.0098799858,
0.0232905503,
0.0312189236,
0.0444425717,
0.0459877327,
0.0453097485,
-0.0031135187,
0.037866544,
0.0269578379,
-0.0129147451,
0.0069840173,
0.0061734375,
-0.0167250559,
0.0231981538,
0.0263943654,
0.0590195991,
0.0471102968,
-0.0781854689,
0.0211119615,
0.008921396,
-0.0007245057,
0.0297097676,
0.0753019899,
0.0655128881,
-0.0078191254,
0.0850498155,
0.0723223686,
-0.02043874,
0.0164634101,
0.0620004386,
-0.0713402033,
-0.0103882784,
-0.0174644366,
0.0105162449,
-0.0104280962,
0.0182481334,
0.0380194224,
-0.0233763568,
0.0128759053,
-0.0192759596,
0.0746536031,
-0.0292085018,
0.0557183959,
-0.0181101933,
-0.0261579808,
0.0203356426,
0.0029503338,
0.0029656938,
-0.0016849126,
0.066897817,
-0.0051394561,
0.0592122525,
-0.0513223708,
-0.0133931199,
-0.0479205139,
0.0171744321,
0.0156061277,
-0.0206866283,
-0.0190128535,
0.0390670858,
0.007893851,
0.0103096617,
-0.0199507978,
-0.0127577521,
-0.029202247,
0.0272841956,
0.0786681101,
0.0397630557,
0.0754391402,
-0.0553230383,
-0.0608230568,
-0.0166764222,
0.0087382244,
-0.0277900118,
0.0374487974,
-0.0276282355,
-0.0448321551,
-0.0068017542,
0.0040625045,
0.0200723205,
-0.0210499205,
-0.0099456785,
-0.0099205282,
-0.0088942181,
-0.0379357859,
-0.0124463169,
0.0051285438,
-0.030250404,
-0.0259931106,
-0.0038699028,
-0.0046835416,
0.0549000725,
-0.0011720698,
-0.0326065719,
-0.0203967132,
0.0137424571,
-0.010494153,
-0.0130755706,
0.0626088157,
-0.0119444029,
0.0632973313,
-0.0588724539,
-0.0505974293,
-0.0151798893,
0.0093767392,
0.0197683442,
0.0084737353,
0.0878392607,
-0.0120678786,
0.0198502187,
-0.0106357122,
0.0662978366,
0.0462495647,
0.057685487,
-0.0179966092,
-0.0676074177,
-0.022545591,
-0.0327405147,
-0.0230605081,
-0.0551958717,
-0.0040184376,
0.0135971187,
0.005690882,
-0.0616268925,
-0.0030537732,
-0.0269832108,
2.594e-7,
0.060806971,
0.0485735573,
-0.0290263612,
-0.0337730385,
0.0214651488,
0.0171656646,
0.0288886018,
0.0290260576,
0.0344520286,
-0.0107956221,
-0.0108360797,
0.0240156148,
0.0421430878,
0.0389543921,
-0.0359303765,
-0.0602974668,
-0.0367442034,
-0.0158457104,
-0.0302520003,
0.0211400501,
0.0643906891,
0.0830738693,
0.0113554746,
-0.0211159531,
-0.0326977596,
-0.0406842604,
0.0344826169,
-0.0192895103,
0.0245317072,
0.049935855,
0.0300096124,
0.003453542,
0.0043191887,
-0.0175479557,
-0.0148120169,
-0.0488455035,
0.0141072115,
0.0306537673,
-0.0291268285,
0.0343828425,
0.0301935375,
0.0422441475,
0.0057745161,
-0.0031343424,
0.0590397269,
-0.0857450292,
-0.0387789793,
-0.0637229085,
-0.0035672949,
-0.0686797351,
0.0611124933,
-0.0265069213,
0.0035704894,
0.0661781132,
0.0166353248,
-0.0050522927,
0.000273323,
-0.073486127,
-0.0271309409,
0.0189800411,
-0.0413154736,
-0.0023431408,
0.0095444154,
-0.0001104291,
-0.0044850516,
-0.0351446941,
0.0421812907,
2.072150421e-34,
0.0116346516,
0.0200643335,
-0.0175370462,
-0.0113051962,
0.0363514498,
0.0150168529,
-0.0043644467,
0.0086995168,
-0.0183100253,
-0.0595582388,
0.0049458682
] |
http://arxiv.org/abs/2401.14228v1 | http://arxiv.org/pdf/2401.14228v1 | Assessing the Portability of Parameter Matrices Trained by Parameter-Efficient Finetuning Methods | 2024-01-25T15:11:07 | 2024-01-25T15:11:07 | Mohammed Sabry, Anya Belz | cs.CL | cs.CL, cs.AI, cs.LG | As the cost of training ever larger language models has grown, so has the
interest in reusing previously learnt knowledge. Transfer learning methods have
shown how reusing non-task-specific knowledge can help in subsequent
task-specific learning. In this paper, we investigate the inverse: porting
whole functional modules that encode task-specific knowledge from one model to
another. We designed a study comprising 1,440 training/testing runs to test the
portability of modules trained by parameter-efficient finetuning (PEFT)
techniques, using sentiment analysis as an example task. We test portability in
a wide range of scenarios, involving different PEFT techniques and different
pretrained host models, among other dimensions. We compare the performance of
ported modules with that of equivalent modules trained (i) from scratch, and
(ii) from parameters sampled from the same distribution as the ported module.
We find that the ported modules far outperform the two alternatives tested, but
that there are interesting performance differences between the four PEFT
techniques. We conclude that task-specific knowledge in the form of
structurally modular sets of parameters as produced by PEFT techniques is
highly portable, but that degree of success depends on type of PEFT and on
differences between originating and receiving pretrained models. | Assessing the Portability of Parameter Matrices Trained by Parameter-Efficient Finetuning Methods: As the cost of training ever larger language models has grown, so has the
interest in reusing previously learnt knowledge. Transfer learning methods have
shown how reusing non-task-specific knowledge can help in subsequent
task-specific learning. In this paper, we investigate the inverse: porting
whole functional modules that encode task-specific knowledge from one model to
another. We designed a study comprising 1,440 training/testing runs to test the
portability of modules trained by parameter-efficient finetuning (PEFT)
techniques, using sentiment analysis as an example task. We test portability in
a wide range of scenarios, involving different PEFT techniques and different
pretrained host models, among other dimensions. We compare the performance of
ported modules with that of equivalent modules trained (i) from scratch, and
(ii) from parameters sampled from the same distribution as the ported module.
We find that the ported modules far outperform the two alternatives tested, but
that there are interesting performance differences between the four PEFT
techniques. We conclude that task-specific knowledge in the form of
structurally modular sets of parameters as produced by PEFT techniques is
highly portable, but that degree of success depends on type of PEFT and on
differences between originating and receiving pretrained models. | [
0.0116558885,
0.0130957365,
-0.0210944377,
0.0221960284,
0.0033255843,
0.0432827137,
-0.002800623,
-0.0170071162,
-0.0460253134,
-0.0473533012,
-0.0596880279,
0.0118680531,
-0.0220193714,
0.0134228934,
0.0318507403,
-0.0537451915,
0.0447039306,
-0.0054271063,
-0.0283456333,
0.0013308454,
0.0130681749,
-0.0017276763,
0.0034158698,
0.0169883985,
0.0337827653,
0.0114796907,
-0.0088010998,
-0.0024843116,
0.0350408889,
0.0388270989,
0.0253987815,
0.053953886,
0.0158302449,
-0.0287727155,
0.0000018791,
0.0088023897,
-0.0514218695,
-0.0127339056,
-0.0087241149,
0.0276008397,
0.1101444513,
0.0227334611,
0.0066027325,
-0.0024242562,
-0.0571506619,
-0.0369319133,
0.0474111997,
0.0675232857,
0.0431149602,
0.0727634057,
-0.0121845724,
0.0065052067,
0.0155313862,
-0.0191327054,
-0.0225507393,
-0.0479813367,
0.0121163307,
0.000005858,
-0.0002902503,
0.0203338433,
-0.0200713817,
0.0286211148,
-0.0183046497,
0.040641807,
0.0098686228,
0.0008693479,
-0.0183945503,
0.0196198653,
0.0082021719,
0.0962843597,
-0.0491433926,
-0.0153113687,
0.0059911506,
-0.0179725755,
0.0131992325,
0.0178756751,
-0.0501437597,
-0.0184108429,
0.0309797563,
0.0265520904,
0.0527955107,
-0.0236772168,
0.0121868942,
-0.0208210442,
-0.0018083516,
0.0605488122,
0.0032474834,
0.0053608664,
0.0294804517,
0.0108597521,
0.0011872994,
-0.0125266612,
-0.0555621982,
-0.0128132794,
0.0371164083,
0.005952423,
-0.0515039898,
-0.0344555974,
-0.0053110044,
-0.0184020083,
0.0533440821,
0.0140711814,
-0.0233451929,
0.0150837302,
-0.0777738243,
-0.001183424,
-0.0164262559,
0.0516404696,
-0.0508496165,
-0.0153039694,
-0.1365680248,
0.0322430991,
-0.0412274934,
0.0639502406,
-0.0271695051,
-0.0050082952,
-0.0122187594,
0.0069477805,
-0.0368847176,
-0.0208175015,
0.0020950781,
-0.0173943099,
-0.0224671587,
-0.0113969203,
-0.000961083,
0.0851779953,
-0.0320815183,
0.0332243256,
-0.0291126091,
-0.0652479753,
0.0018206548,
-0.0001999148,
0.0343234576,
-0.046975296,
0.0024969615,
0.0214185007,
-0.0252810791,
-0.0388930142,
-0.0500821732,
-0.0072679333,
-0.0292477459,
-0.0767181069,
0.0165258031,
-0.0688244775,
0.0312130954,
0.0065197633,
0.0108931502,
-0.0290297698,
-0.0255320799,
-0.0142164035,
-0.0649656579,
0.0592530034,
-0.0051442124,
0.0024068938,
-0.0410102494,
-0.0112167187,
-0.0320208371,
0.0839532241,
-0.0194416363,
0.0451242626,
0.0310189221,
-0.0530525073,
-0.0094840815,
0.0090153441,
0.0004362446,
0.024722157,
-0.0053034513,
-0.0079835197,
0.0320140012,
-0.0479148179,
0.0132706305,
-0.0054674172,
-0.0304552801,
0.0401902646,
0.0604176149,
-0.0088228276,
0.0365062654,
0.0359247625,
0.0500054099,
-0.0148422057,
0.0168177579,
0.0403914116,
-0.0610313602,
0.0211717933,
-0.0226492714,
-0.015875401,
-0.0081867874,
0.0112044374,
-0.0555581227,
0.004486253,
0.0163440742,
-0.0511527732,
0.0247028992,
-0.0052873264,
0.0706511736,
0.0231307223,
-0.037567161,
0.051156804,
0.0585033782,
-0.0664812103,
-0.0071845022,
-0.018216541,
0.0206071176,
0.0410201363,
-0.0301117674,
-0.024636047,
-0.0262869317,
0.0111385342,
0.0264577828,
-0.0404132977,
0.0809052289,
-0.0096258139,
0.030079985,
-0.0186562557,
-0.0147634065,
-0.0531534068,
0.0377656259,
0.0114150364,
-0.0032409206,
0.0327554457,
0.0149143161,
-0.0512325503,
-0.0307014845,
0.0114119509,
-0.0066009993,
-0.0160050709,
-0.0132533601,
-0.0079898639,
0.0287328959,
-0.0137784518,
0.0038377293,
-0.0088736666,
0.0304542314,
-0.0213641971,
0.0014040272,
0.0138297779,
-0.0255747121,
-0.0247633494,
-0.0177301373,
-0.0473102406,
0.0292089246,
0.0477136895,
-0.0592042878,
-0.0408620574,
0.0524629429,
0.0020618916,
0.0276672114,
-0.0083083073,
0.0028871747,
-0.0493074842,
0.0369477831,
-0.0043053473,
0.0161597226,
0.0419173092,
-0.0071138251,
-0.0190338474,
-0.0413343422,
0.0308489576,
0.0046787015,
-0.0339726247,
-0.0017595349,
-0.00187352,
0.0123095419,
-0.0006959638,
-0.0109317247,
-0.0443063378,
-0.0339724869,
-0.1191918328,
0.0169453379,
-0.0249493737,
0.0521248206,
0.0543678217,
-0.0027026033,
0.0117573468,
0.0071335095,
-0.048637528,
0.0393425338,
-0.0326080695,
0.0436606482,
0.014979694,
0.0078999475,
-0.0070770802,
-0.0675177574,
-0.022187246,
0.0071697473,
0.0380882509,
-0.0461555012,
-0.0095970854,
-0.0769874454,
0.0054729283,
-0.0317051373,
0.0288719758,
-0.0106878458,
-0.0319043994,
-0.0032188015,
-0.0473150536,
0.0319545865,
0.0443152823,
-0.0382671915,
0.0158275757,
0.0471895263,
0.004354673,
-0.0179633368,
0.0195415262,
-0.046953626,
-0.0351698287,
-0.0427602194,
-0.020464763,
-0.000850008,
0.041072946,
-0.0198136475,
-0.0145632792,
-0.0796388313,
0.0145282336,
-0.0308438074,
-0.0054482445,
0.0173495729,
0.0214014594,
0.0297009833,
0.0098171998,
0.0180460587,
-0.0313287005,
0.0402995981,
-0.0125754206,
-0.0006793758,
-0.0085904505,
-0.0369153321,
-0.0170327183,
0.046821747,
0.0253624152,
0.0016762554,
0.0157796424,
0.015116645,
-0.0459036417,
-0.010355752,
0.0389165133,
-0.0091408649,
0.0551562048,
-0.0175710674,
0.0007794928,
-0.0273263492,
-0.0286450237,
0.0219772104,
-0.0424779691,
-0.0098096291,
0.0391933173,
0.0152770756,
-0.0122274132,
0.0085264165,
0.0261275712,
0.0094585884,
-0.0028592746,
-0.0332098156,
-0.0050037773,
-0.0024397324,
0.0222988464,
0.0404164344,
0.0354308598,
0.0399414338,
-0.0256011635,
-0.0100067314,
0.0083317701,
0.0133198993,
0.0225789398,
-0.0439281687,
0.0155120008,
0.0208545085,
-0.0287571773,
-0.0454368591,
-0.0086832093,
0.1036581695,
-0.0613262802,
0.0097896941,
0.0654755905,
-0.0070319255,
-0.0363351665,
0.0019322194,
0.0203820877,
0.018689692,
0.0014640504,
0.0389598124,
-0.0314739496,
-0.0436072722,
0.0032591936,
0.0402608439,
0.0047692247,
-0.0309445001,
0.0109557137,
-0.1075938046,
0.0352060162,
-0.0017617595,
-0.0391788557,
-0.074728176,
-0.0336465761,
0.0237248223,
-0.0266917031,
0.002046617,
-0.0112559479,
-0.0357255414,
0.0231851898,
0.019804161,
-0.0438998714,
-0.0505969971,
0.0219811574,
-0.0112214591,
0.0445312746,
0.0040578945,
0.0101474179,
-0.0048756977,
-0.0322421752,
-0.013571023,
-0.0744138286,
0.0072772871,
-0.0092334794,
0.0794608891,
0.012437419,
-0.0444105826,
0.0402839333,
0.0079895668,
-0.0341759436,
0.0159146115,
-0.0335897207,
0.0871466398,
-0.0103242453,
-0.0431828275,
0.0272662044,
-0.0080737453,
0.0078468984,
0.083713524,
0.0252040401,
-0.0189251304,
-0.008117605,
0.0268333461,
-0.0325852372,
-0.0133558651,
-0.0151927397,
0.0160068106,
-0.0541016757,
0.0027271772,
-0.0869783461,
-0.0270864554,
0.0620646439,
0.058727473,
-0.018449679,
0.0350086391,
0.001545069,
0.0000430578,
-0.051818613,
0.0368009619,
-0.0376900062,
-0.006154777,
0.003244271,
-0.0455239713,
0.0504714586,
-0.0051506427,
-0.0148832677,
-0.023437107,
-0.1100064218,
-0.0563348383,
0.0132297361,
0.0472476296,
0.0156873856,
0.013549923,
-0.0086806491,
-0.0339894556,
-0.0412278734,
0.0016050838,
-0.0434466451,
0.0244775526,
0.0496031493,
-0.0011337313,
-0.011617084,
0.0496929847,
-0.0351657607,
0.0154652083,
0.0442691073,
0.0056387642,
0.0146518406,
-0.0467126481,
-0.0010240168,
0.0204392634,
0.0442115106,
0.0303928256,
-0.0778673217,
-0.0175570343,
-0.008144442,
0.0059420685,
0.0098160598,
-0.0645736381,
-0.0225688759,
0.0175278317,
0.0515573919,
-0.0136459833,
0.0324068032,
0.0168957692,
0.0603295304,
-0.0096463393,
0.0269446969,
0.0063693146,
0.0032315974,
-0.0067031193,
0.036051143,
-0.0380760171,
-0.0030162758,
0.0089104492,
0.0112142488,
-0.056269221,
0.0248965565,
0.0287329014,
0.0383345522,
-0.0453070663,
0.0367300846,
-0.0385755077,
-0.0917439163,
-0.0041837483,
0.0319192596,
0.024868276,
-0.0314293765,
0.1062504426,
-0.0220662933,
-0.0242681187,
-0.0125565417,
0.0418379009,
-0.0185003057,
0.0622292757,
0.0355990082,
0.0150921242,
-0.0153079526,
0.0106621208,
0.0736658573,
-0.0049921311,
0.0406917185,
-0.0206198469,
0.0030970115,
-0.0209884197,
0.0134640438,
0.0121103888,
-0.0153267495,
0.0047280942,
0.015499061,
-0.0083701843,
-0.0603945516,
0.0117502976,
-0.0455473177,
-0.0169895217,
-0.0172331408,
0.0523625277,
0.0516745225,
0.0105818696,
0.0143283932,
-0.052719526,
-0.034987174,
0.0586174801,
-0.0057652066,
0.0058182045,
-6.714907154e-33,
-0.0187553838,
0.0545094125,
-0.0186566599,
0.0534244478,
-0.0353953391,
-0.0182189159,
-0.01800441,
0.0415905267,
-0.0206034575,
0.0231305547,
0.0161591582,
-0.0224464424,
0.0093555888,
0.0071563041,
-0.0052066916,
-0.027885044,
0.0448852889,
0.0079727331,
0.0034583283,
0.0564159155,
0.0529025495,
0.0436535701,
0.0918830559,
-0.0724444389,
0.0038483175,
0.0201258883,
0.0065712854,
-0.0028851698,
0.0417702757,
0.017815005,
0.0092513012,
0.0304375216,
0.0394770168,
-0.0600395575,
-0.0235149171,
0.0207793526,
-0.0823970735,
-0.0427018143,
0.0449490137,
-0.0159089379,
-0.0157951917,
-0.0427301377,
0.0629588142,
0.0012614271,
0.0099390065,
-0.0209496114,
0.0369801857,
-0.0205698423,
-0.0052091144,
-0.0199057702,
-0.0946262032,
-0.0070801689,
0.0056288899,
0.0166490637,
0.0461868048,
0.0580402389,
-0.0171329472,
0.0086396467,
-0.0101636257,
-0.023502402,
0.021924356,
0.0583665557,
0.0162938572,
-0.0297729652,
-0.0013544038,
0.0180579331,
0.0188681856,
0.0240022708,
-0.0059564197,
0.0175102316,
0.0517781638,
0.0903640017,
0.0579661839,
0.0513089299,
0.0210628696,
-0.0061457683,
-0.0538749397,
-0.0125448285,
0.0171658117,
0.0293543637,
0.0173202902,
0.0142566012,
0.0013183139,
-0.0140064135,
-0.0668136254,
0.01039996,
-0.0008797706,
-0.0415777341,
0.017420454,
-0.0014119478,
-0.0163556281,
0.0154113155,
0.0487636141,
-0.0483774841,
0.0089638447,
-0.0015069173,
-0.0530499965,
0.0338328965,
0.0195510909,
-0.0296978634,
-0.0650221184,
-0.0496628173,
0.0064112656,
-0.0112331947,
0.0360445306,
-0.0151002482,
0.0743442848,
0.0487658568,
-0.0647298172,
-0.01283562,
-0.0301859342,
0.0343378447,
-0.0040459805,
0.1000671834,
0.0199000593,
0.0304471795,
-0.0200068559,
0.0037052259,
-0.0024602278,
0.0121640805,
-0.0203459933,
0.0058501968,
0.0447623581,
0.0050925897,
-0.0132829128,
-0.0264303423,
-0.0545441955,
-0.0016010909,
0.0367815457,
-0.0362248309,
-0.0199131686,
0.0535226651,
2.791e-7,
0.0534453206,
0.0404292159,
-0.0346046798,
-0.0475215651,
0.0335952006,
0.0440065637,
-0.0387527496,
0.0120707583,
0.029796049,
-0.0110656554,
0.0041417307,
0.0426636115,
0.0637771785,
0.0336280353,
-0.0806770176,
-0.0254659969,
-0.0158438776,
0.0054641222,
0.0146449869,
0.0024394118,
0.0392075703,
0.0924083367,
0.041396495,
-0.0105912294,
0.011583712,
-0.0208882019,
0.0715966076,
-0.0640707612,
-0.0094190473,
0.028782187,
-0.0024432435,
0.024280943,
-0.0380008407,
0.0339422338,
-0.0062194359,
0.0040966365,
-0.0099815112,
0.0685316846,
0.0102944756,
-0.0331904367,
0.0083374139,
-0.0202792007,
0.0151742082,
-0.0580376163,
0.0771168321,
-0.1040646136,
-0.0211926699,
-0.0678283945,
-0.0149279134,
0.0491261221,
0.00250004,
0.0011003049,
0.0052774609,
0.0446140952,
0.0227516908,
-0.0093108881,
-0.0669844672,
-0.0645427182,
0.0206529591,
0.0797810778,
-0.0185761638,
-0.0902666524,
-0.0235349461,
-0.0481088459,
0.0138016706,
-0.0663345456,
0.0453233197,
2.283606473e-34,
0.0094200484,
-0.016109813,
0.0321195088,
-0.014926861,
-0.0005492846,
-0.0115193166,
-0.0609919019,
-0.0184154995,
0.0035130479,
-0.0776208714,
0.0080142245
] |
http://arxiv.org/abs/2401.14215v3 | http://arxiv.org/pdf/2401.14215v3 | Commonsense-augmented Memory Construction and Management in Long-term Conversations via Context-aware Persona Refinement | 2024-01-25T14:54:33 | 2024-02-12T12:27:18 | Hana Kim, Kai Tzu-iunn Ong, Seoyeon Kim, Dongha Lee, Jinyoung Yeo | cs.CL | cs.CL, cs.AI | Memorizing and utilizing speakers' personas is a common practice for response
generation in long-term conversations. Yet, human-authored datasets often
provide uninformative persona sentences that hinder response quality. This
paper presents a novel framework that leverages commonsense-based persona
expansion to address such issues in long-term conversation. While prior work
focuses on not producing personas that contradict others, we focus on
transforming contradictory personas into sentences that contain rich speaker
information, by refining them based on their contextual backgrounds with
designed strategies. As the pioneer of persona expansion in multi-session
settings, our framework facilitates better response generation via human-like
persona refinement. The supplementary video of our work is available at
https://caffeine-15bbf.web.app/. | Commonsense-augmented Memory Construction and Management in Long-term Conversations via Context-aware Persona Refinement: Memorizing and utilizing speakers' personas is a common practice for response
generation in long-term conversations. Yet, human-authored datasets often
provide uninformative persona sentences that hinder response quality. This
paper presents a novel framework that leverages commonsense-based persona
expansion to address such issues in long-term conversation. While prior work
focuses on not producing personas that contradict others, we focus on
transforming contradictory personas into sentences that contain rich speaker
information, by refining them based on their contextual backgrounds with
designed strategies. As the pioneer of persona expansion in multi-session
settings, our framework facilitates better response generation via human-like
persona refinement. The supplementary video of our work is available at
https://caffeine-15bbf.web.app/. | [
0.0507318228,
0.0545296706,
-0.0365302898,
0.0146485372,
-0.0036070244,
-0.0424277075,
-0.0020196557,
-0.038139578,
-0.0063766222,
-0.0055891629,
0.0019596373,
-0.0125489933,
0.0332579613,
-0.0032225312,
0.0316814184,
-0.0896741077,
0.0438917428,
0.000033314,
0.0010390272,
0.0011913551,
0.0154263508,
-0.007174491,
-0.0000491004,
0.0253932718,
-0.0474682376,
-0.0207248274,
-0.0088951169,
-0.0183447283,
0.0592885613,
0.0047249296,
0.0193282384,
0.0304854903,
0.006073955,
-0.0223854892,
0.0000018693,
-0.0148221403,
0.0296992213,
-0.0185447913,
0.0050942726,
-0.0101830382,
0.0150420899,
0.0376832597,
0.0069650812,
0.0024517651,
-0.059222579,
0.0058575673,
0.0151755055,
0.0445463583,
0.0756413043,
0.0517600812,
-0.0077487859,
-0.0438900627,
-0.0076925219,
-0.0169706661,
0.0227468926,
-0.04051774,
-0.0037000165,
-0.0643277541,
-0.0048255431,
-0.0165566429,
0.013395247,
0.0016801003,
0.016749138,
-0.0099126874,
0.0344041996,
0.0215628501,
-0.0284236334,
0.0098688975,
0.0140179023,
0.1059586629,
0.0979925767,
0.0141520752,
0.0074713835,
0.0350834318,
0.0665302649,
0.0127897114,
-0.0311031379,
0.0021858064,
0.0000503264,
-0.0275473353,
-0.0431844331,
0.0290107969,
-0.0257248264,
0.0153501341,
-0.0341205299,
-0.007260784,
0.0008605465,
-0.0058675222,
-0.0266288128,
-0.0178322215,
0.0151718259,
-0.0425839089,
0.0278829411,
0.0446110889,
-0.0567434877,
0.0160182957,
-0.0359843038,
0.0135279885,
0.0186179355,
-0.0418287478,
-0.0086460691,
0.0208909884,
-0.0120145986,
-0.0088457679,
-0.0089580687,
0.0251401626,
-0.0023750728,
-0.0528965332,
-0.0068203984,
-0.0094862934,
-0.0724471435,
-0.0287617799,
-0.0324803554,
0.0387885906,
-0.015252349,
0.015905425,
-0.044231236,
0.0192111842,
-0.0320345983,
0.0100901267,
-0.0319975615,
0.0409609675,
-0.0650606453,
0.005182243,
-0.0488366485,
0.0816480964,
-0.079832986,
0.0020512363,
0.0474784262,
-0.0250917468,
-0.0248240009,
0.0161619745,
0.0236383695,
-0.0099823391,
-0.0259167738,
0.0105019202,
-0.0388259105,
-0.0124908332,
-0.0139242839,
-0.0580159612,
-0.0459695943,
-0.0502490178,
-0.0276608523,
-0.0182311945,
0.0288866386,
-0.005219765,
-0.034317147,
-0.0114849266,
0.0173534416,
0.0178836137,
-0.0164242927,
0.0472064354,
0.0189416297,
0.0131605268,
0.0418536626,
0.0579745248,
0.0070415726,
0.0167167373,
0.0161125995,
-0.03731253,
0.0294380914,
-0.0186286848,
0.0436425954,
0.0056886356,
0.0310937203,
-0.0181889609,
0.0445156358,
0.0297933016,
0.0089851879,
0.0402666219,
0.0664256811,
-0.0090178261,
-0.052684281,
-0.0211003684,
0.072845526,
0.0299737044,
-0.0552552417,
0.071189709,
0.0076407273,
0.0162703004,
-0.0108542321,
0.0626806319,
-0.0349123925,
0.0077168201,
-0.0091738403,
-0.0022155556,
-0.0027149951,
0.0176027343,
-0.0350231119,
0.0041128984,
0.0636314005,
-0.0073112091,
0.0509971082,
-0.0116207451,
-0.0439174362,
0.0280453432,
-0.0233413093,
0.0932063535,
-0.0037400448,
-0.1114043295,
0.0113214841,
-0.0107406098,
-0.0069863726,
0.084171854,
0.0231444705,
-0.0048755798,
-0.0199655835,
-0.0100582903,
0.0023829755,
0.0506233573,
0.0290546715,
-0.0332052782,
0.0530731492,
-0.0396644399,
0.0241258442,
-0.0129008731,
-0.0294447746,
0.038288068,
0.0435444973,
0.0199940838,
0.0479563363,
0.0110063907,
0.0486219376,
-0.0206555277,
-0.0330483019,
0.0055553741,
0.0350080207,
0.009061371,
0.0091850432,
0.0412089154,
0.0248069465,
-0.0148949586,
0.0303708222,
-0.0162197035,
0.021332182,
0.070263356,
-0.0737527981,
-0.0633575246,
0.0405299254,
-0.0358156711,
-0.0203782544,
-0.0213979483,
0.0107819438,
-0.0574967265,
0.0422156155,
-0.0311153438,
0.0650036559,
-0.0366634391,
-0.0024023643,
0.0281979442,
0.0320348963,
-0.0254021175,
0.0659673437,
0.0049489047,
-0.0179754831,
0.0123607554,
-0.0125507135,
-0.0188968405,
-0.0876057371,
0.0267257728,
-0.0009129656,
0.0069297603,
0.0044018733,
0.0205502268,
-0.0183082558,
-0.074774757,
-0.0338468701,
-0.1204944029,
-0.0124902101,
-0.0087783746,
-0.0117252935,
-0.0113021573,
0.0378270522,
-0.0034324687,
-0.0216173436,
-0.0740432441,
0.0030746281,
0.0188596491,
0.0080565475,
0.034642186,
0.0401634313,
-0.0017571435,
-0.0034882813,
0.0052893218,
0.0279897451,
0.0186242796,
-0.0334770754,
0.0258974116,
-0.034673281,
-0.0171269476,
-0.0407474563,
0.0017346895,
-0.0239475109,
-0.0293323547,
0.0342961662,
0.0292658359,
0.0948895887,
0.0619546026,
0.0532300062,
0.0220318213,
-0.0601866171,
-0.0380105078,
-0.0195867941,
0.0374072976,
-0.0264647566,
0.0110776434,
-0.0206530411,
-0.009165789,
0.0122057311,
-0.0193456206,
-0.0117151551,
0.0184038188,
-0.0530197434,
-0.0089643216,
0.0595766082,
-0.0575923733,
0.0271315873,
-0.0038004536,
-0.0359740928,
-0.0058887787,
-0.0026511152,
0.0327691212,
0.0728473067,
-0.0026433903,
0.0002980477,
-0.0218331702,
0.0074803666,
-0.0027141548,
0.0427206643,
-0.0293557812,
0.0030941996,
-0.0343081243,
-0.025357917,
-0.0393170491,
-0.0403996222,
0.0195497405,
-0.0283433162,
0.0370192565,
-0.0356590077,
-0.0137153054,
0.0121506462,
-0.0584034882,
-0.0669671595,
-0.0289271176,
0.0001166877,
-0.0127791995,
0.0111095598,
-0.043048691,
-0.0545850284,
-0.0003614015,
0.0332684107,
0.008882164,
-0.0162484292,
0.0057215099,
-0.0024540082,
0.0231535584,
0.0127344942,
0.0887723416,
0.0283052977,
-0.0080452925,
0.0183924381,
0.039202854,
0.020967491,
-0.0418437012,
0.0339678377,
-0.0287009142,
0.0034642851,
-0.0278856084,
-0.0440284796,
0.041717656,
0.0577710792,
-0.0526684634,
-0.0433527455,
0.0620415471,
0.0303180274,
-0.0175387803,
0.0099371886,
-0.0381773338,
0.0582076982,
-0.0067667016,
0.011267012,
-0.0443595536,
-0.0238319486,
-0.0102544241,
0.0273948982,
-0.0404383428,
-0.066722855,
-0.0196537375,
-0.1189024001,
0.0099827619,
0.0200686734,
-0.1000793576,
0.0042882608,
-0.043000415,
-0.0034879909,
0.0250654239,
0.0390727557,
0.0041868212,
0.0214970671,
0.0081189713,
-0.0129186725,
-0.0339066908,
-0.0567529052,
0.0092171682,
-0.0040657013,
0.0520663634,
0.0516907312,
0.0665016323,
0.0452841446,
-0.0567809902,
-0.0358897671,
-0.0270649306,
0.0467569791,
-0.016467806,
0.0398273468,
-0.0161493178,
-0.001001445,
-0.0599182062,
0.0722154453,
-0.0257751439,
-0.0566104762,
0.0331483632,
0.0042417515,
-0.0324129425,
0.0215738434,
-0.0020501078,
0.0403594859,
0.0649875328,
0.0800041929,
0.0241393279,
-0.0112978602,
0.00399868,
0.0152407074,
0.0261319261,
-0.0558998324,
0.0234252196,
-0.0044502905,
0.0133146057,
0.0129189966,
-0.059972465,
-0.0686606765,
0.0146664334,
0.0255963989,
-0.0240401663,
-0.0101929521,
-0.0017837287,
0.0358780734,
-0.0566460267,
0.0231069997,
-0.0343478397,
0.0119790183,
-0.0386297815,
-0.025769908,
0.0214418638,
-0.0311108977,
0.015591233,
0.0417196266,
-0.0254801493,
-0.0559423491,
-0.0124790734,
0.0629237816,
0.0229656789,
0.0897922888,
0.0304968283,
-0.0420414954,
0.0301488228,
0.0552991293,
-0.0608142056,
0.0201840438,
0.0256042238,
0.0522831194,
-0.0276890285,
0.0512117222,
0.0158886537,
-0.0184737965,
-0.0158014391,
0.0210554674,
-0.0517380312,
-0.0738028958,
-0.017081501,
-0.0063299094,
-0.0131262494,
-0.0067622191,
0.0188533217,
0.022682799,
0.0090318928,
-0.0496798344,
-0.0219225399,
-0.0079122223,
-0.0195110608,
0.0024791469,
0.0015242809,
0.0352050476,
-0.001866675,
-0.0672331825,
0.0339823589,
-0.0325928181,
-0.0324478149,
0.0175125618,
-0.0136053786,
0.0101479199,
0.0054956954,
-0.013268956,
-0.02401237,
0.0106024621,
0.049105674,
-0.0033816656,
0.0819364265,
0.0190514363,
-0.0330644883,
-0.0051705083,
0.0097452169,
-0.0627860427,
0.036938373,
0.0120006269,
0.071140185,
-0.0146046793,
0.0278727412,
0.0475959368,
-0.0053965105,
0.0077166213,
0.0015515233,
0.0490410626,
-0.0041011441,
-0.0232998002,
0.0131853847,
0.0037698557,
-0.057985153,
0.0114573231,
0.0666059628,
-0.0448698513,
-0.0222366825,
-0.0118128536,
0.0114444196,
-0.0683328882,
-0.062033914,
-0.0026229415,
0.0703695714,
0.0247019418,
-0.0121529968,
-0.0837995782,
-0.0047948039,
0.0160199497,
-0.0250809845,
0.0155528793,
-0.0150993373,
0.0217725914,
0.0412339196,
-0.0093861287,
-0.0590057261,
-0.014722364,
0.0473619625,
-0.0572191738,
-0.0082629863,
-0.0193890836,
-5.687782231e-33,
-0.0636374429,
-0.0377913862,
0.019694183,
-0.0134221921,
-0.0874383822,
-0.0147716952,
-0.0365496799,
0.0149928145,
0.0132359024,
-0.0045337239,
-0.0367031284,
0.0276717767,
0.0048218011,
0.0410165116,
-0.038171418,
-0.0315530933,
0.0712529123,
0.0492347106,
-0.0502653047,
0.067452766,
0.0620345809,
0.021579491,
0.0289952811,
-0.0561778098,
0.0124318553,
-0.0260185841,
-0.0185292289,
0.0140883019,
0.0107316114,
0.0505577959,
-0.0166722871,
0.0567328446,
-0.0009407671,
-0.0187926088,
-0.0023428516,
0.0853508636,
-0.0781577751,
-0.026647266,
0.0429455489,
0.0025500606,
-0.006166738,
-0.0625961795,
0.0165436268,
0.002403293,
0.0239366796,
0.0249836277,
0.0159319341,
-0.0172869768,
-0.0161244068,
0.0103183929,
-0.0323651694,
-0.0138175255,
-0.0149728823,
-0.0297519192,
-0.0506331362,
0.0154096335,
0.0020072856,
-0.0482828021,
-0.0940842107,
0.0307612978,
-0.0241743326,
-0.0134205697,
0.0258080568,
0.0597718321,
0.0158860553,
0.0376101211,
0.0107177598,
0.0399207212,
-0.0225854926,
-0.0109843267,
0.0005588991,
0.0367166549,
0.0095532611,
0.1167205572,
-0.0088526038,
-0.0429512039,
-0.034924753,
0.033223521,
0.0219763704,
0.0053019235,
0.0082572475,
0.003503703,
-0.0175611991,
-0.0196032543,
-0.0208856408,
-0.0048159696,
0.0007565555,
-0.0039536227,
-0.0076813712,
0.0162197538,
-0.046272289,
0.0016192957,
0.0261660498,
-0.0333457664,
-0.0052756667,
0.024709776,
-0.0626610443,
-0.0016294399,
0.0080842571,
-0.0315588154,
-0.021368837,
-0.0061258916,
0.0779104307,
0.0255806055,
0.0080567859,
0.0280694421,
0.0032513526,
-0.0102121625,
-0.0332359299,
-0.0214507952,
0.0022384541,
0.071436666,
-0.0596793219,
-0.0087360861,
0.0171498954,
0.0276019517,
-0.0258858409,
0.0374478735,
0.0245347284,
-0.0055458392,
-0.0205955952,
-0.0170961097,
0.0422836915,
0.0167109333,
-0.0146716507,
0.0222737193,
0.0054558786,
-0.0519086123,
0.0518883802,
-0.0379311964,
-0.0039452994,
-0.0224814489,
2.452e-7,
-0.0578702986,
0.019642368,
0.0393316224,
-0.029885089,
0.0358093195,
0.0467832424,
0.0415425189,
0.0444490984,
0.0046171113,
-0.0411812328,
0.012623244,
0.0369289257,
0.0545138419,
0.0111336093,
-0.0792250559,
-0.0251197647,
-0.0442285016,
-0.0215108059,
-0.0416114181,
-0.0055022021,
0.090343751,
0.0138777383,
0.0326017216,
-0.0102157313,
0.0205374677,
-0.0162559766,
0.0566070564,
-0.0224040374,
-0.0406668335,
-0.0086476123,
0.0224606842,
-0.0197449233,
0.007284699,
0.0655131787,
-0.0022487747,
0.0265887734,
-0.0150416913,
0.0403453112,
-0.0124364244,
-0.0755880997,
-0.0158828218,
-0.0008519259,
0.0080440314,
-0.0072164917,
0.0309678223,
-0.0699501932,
-0.0422846377,
0.0086309118,
-0.0497017018,
0.0070191114,
-0.0002754448,
0.0057993322,
-0.0461133085,
0.0466495976,
0.0081134364,
0.0125501007,
0.0135727618,
-0.07247556,
0.013686046,
0.0170182921,
-0.0484615155,
-0.0556473322,
-0.0167472046,
-0.03538616,
0.0196462478,
-0.0641598478,
0.033756014,
1.90435044e-34,
-0.0117675988,
0.032609161,
-0.0072573996,
0.0094146589,
0.0359520093,
-0.0195271596,
-0.0152574666,
0.0173990894,
0.0676132888,
-0.0373800173,
0.0057817744
] |
http://arxiv.org/abs/2401.14212v2 | http://arxiv.org/pdf/2401.14212v2 | Explicitly Representing Syntax Improves Sentence-to-layout Prediction of Unexpected Situations | 2024-01-25T14:53:30 | 2024-04-16T14:25:39 | Wolf Nuyts, Ruben Cartuyvels, Marie-Francine Moens | cs.CL | cs.CL | Recognizing visual entities in a natural language sentence and arranging them
in a 2D spatial layout require a compositional understanding of language and
space. This task of layout prediction is valuable in text-to-image synthesis as
it allows localized and controlled in-painting of the image. In this
comparative study it is shown that we can predict layouts from language
representations that implicitly or explicitly encode sentence syntax, if the
sentences mention similar entity-relationships to the ones seen during
training. To test compositional understanding, we collect a test set of
grammatically correct sentences and layouts describing compositions of entities
and relations that unlikely have been seen during training. Performance on this
test set substantially drops, showing that current models rely on correlations
in the training data and have difficulties in understanding the structure of
the input sentences. We propose a novel structural loss function that better
enforces the syntactic structure of the input sentence and show large
performance gains in the task of 2D spatial layout prediction conditioned on
text. The loss has the potential to be used in other generation tasks where a
tree-like structure underlies the conditioning modality. Code, trained models
and the USCOCO evaluation set are available via github. | Explicitly Representing Syntax Improves Sentence-to-layout Prediction of Unexpected Situations: Recognizing visual entities in a natural language sentence and arranging them
in a 2D spatial layout require a compositional understanding of language and
space. This task of layout prediction is valuable in text-to-image synthesis as
it allows localized and controlled in-painting of the image. In this
comparative study it is shown that we can predict layouts from language
representations that implicitly or explicitly encode sentence syntax, if the
sentences mention similar entity-relationships to the ones seen during
training. To test compositional understanding, we collect a test set of
grammatically correct sentences and layouts describing compositions of entities
and relations that unlikely have been seen during training. Performance on this
test set substantially drops, showing that current models rely on correlations
in the training data and have difficulties in understanding the structure of
the input sentences. We propose a novel structural loss function that better
enforces the syntactic structure of the input sentence and show large
performance gains in the task of 2D spatial layout prediction conditioned on
text. The loss has the potential to be used in other generation tasks where a
tree-like structure underlies the conditioning modality. Code, trained models
and the USCOCO evaluation set are available via github. | [
0.0276767984,
0.0583494864,
-0.0001622885,
0.0542322546,
-0.0039433222,
0.0149771078,
-0.0025353918,
0.0005998136,
0.0228707269,
-0.0493871123,
-0.014632103,
-0.0368767455,
0.0331204087,
0.0358871147,
0.0434285551,
-0.0868773311,
0.003600627,
-0.0223465748,
-0.0137364762,
0.0242789257,
0.0444295444,
0.0229676813,
0.0220532473,
0.0152478833,
-0.0660636276,
-0.0292000007,
-0.0346295573,
0.0538579635,
0.0074150614,
-0.0371756479,
-0.0136751607,
0.0251717605,
0.0318379328,
-0.0055432911,
0.0000018329,
0.0062649343,
-0.0136586269,
-0.0115910042,
0.0282372665,
-0.0065327026,
0.0079318341,
-0.0905785188,
0.001699242,
-0.0296020247,
-0.0483915247,
0.0063886167,
0.0732267648,
0.0557524152,
0.0645494685,
0.0699373111,
-0.0268017966,
-0.0450503491,
0.0031275065,
-0.0117825875,
0.0420572609,
0.0521102846,
0.0086520324,
-0.0201608595,
-0.0475114435,
0.0057739522,
-0.0280951988,
-0.0061402577,
0.044769112,
-0.0093879532,
0.0256873723,
0.0631737858,
-0.010835343,
-0.0149968546,
-0.0058956705,
0.0586202256,
-0.0146138668,
-0.040890079,
0.0075072637,
-0.0159318149,
0.0001858678,
-0.0269247293,
-0.0262065437,
0.0331529491,
-0.0072664758,
0.0184928738,
-0.0305077266,
0.0119496593,
-0.0055414657,
-0.0197845735,
0.014139846,
-0.0085281292,
0.0136732319,
-0.0057077953,
-0.0387290865,
0.0644497722,
-0.0204321723,
-0.0716992468,
0.0107496558,
0.0037433612,
-0.0259912703,
0.0162692592,
-0.0702074915,
-0.0452462472,
-0.0323549956,
-0.0725126863,
0.0286873821,
0.0025519168,
0.0080141881,
0.0170594044,
-0.040855553,
0.0903244838,
-0.027652923,
0.0217778645,
-0.0611760132,
0.022980148,
-0.0277439896,
-0.0271546952,
-0.0482993349,
0.0247707572,
-0.0242937952,
-0.0535604358,
-0.001982312,
0.0239760242,
-0.0161757153,
0.0304602887,
0.007533154,
0.0503227897,
-0.0751211792,
0.0076972246,
-0.0284051299,
-0.0259161107,
0.0016421054,
0.0045634066,
-0.004712305,
-0.0036709087,
-0.0256847814,
0.0507017151,
0.0148969265,
-0.0190782603,
-0.038748581,
0.0708304197,
-0.0308803432,
-0.0129005136,
-0.0190267283,
-0.0370317213,
-0.0412162468,
-0.0447492562,
0.0667108148,
-0.0503000021,
0.037614122,
-0.005155256,
-0.0128024546,
-0.0622511134,
0.0032232294,
0.0101110553,
-0.0005388516,
0.0577027872,
0.0238763448,
-0.0228830613,
-0.0064948224,
-0.0049103638,
0.0366582423,
0.0540540591,
0.0393185467,
0.0659504682,
-0.0059355181,
0.0186531842,
-0.00403836,
0.0329373218,
0.0067832707,
-0.0092387274,
-0.0442928821,
0.0177006833,
-0.0029488553,
0.0356251746,
0.0704609901,
0.0320639163,
-0.0130738802,
-0.0141239278,
0.0581549667,
-0.0085514551,
0.0055294763,
0.0358249284,
0.0750322565,
0.0106748315,
-0.0189613719,
0.0366479009,
-0.0815673694,
0.0415333472,
0.0376810804,
0.0473592244,
-0.011196088,
0.0102183018,
-0.0260421745,
-0.0137286149,
-0.0162588041,
-0.0228726473,
0.0274882037,
-0.0416322947,
0.0216149595,
0.0682542026,
-0.0868644714,
0.0887963772,
-0.0132603655,
-0.0310704429,
-0.014821725,
0.0415420346,
-0.0107750315,
0.0392335765,
-0.0094084572,
0.0215241946,
0.0043752515,
0.0142935552,
0.0120189087,
0.0635617673,
-0.0086327344,
0.00910448,
0.0329257138,
-0.0198746156,
-0.0424135961,
-0.0259982478,
0.0002891368,
0.0122058066,
0.0388416983,
-0.0156524945,
0.0338658653,
-0.0565014891,
-0.0089183124,
0.0648315102,
0.0452758409,
-0.0132539365,
-0.0340975784,
0.0395998619,
0.0175488126,
-0.0249940082,
-0.0126567902,
0.0221034344,
0.0265698545,
-0.0474541783,
-0.0105433976,
0.0108241485,
-0.0932293236,
-0.1138484702,
-0.0157068092,
-0.0559557602,
-0.0206406899,
0.0044713933,
-0.0595576391,
-0.0336552039,
0.0111200968,
-0.0606585816,
0.0379686244,
-0.0217686761,
-0.0357166342,
0.0495167039,
-0.0009866159,
-0.0070719211,
0.0815184861,
-0.00894765,
-0.0183878094,
-0.0013201915,
-0.0208248347,
0.0057462323,
-0.0155621516,
0.0296022054,
-0.0378281698,
0.0004880054,
-0.0209167674,
0.0005010553,
0.0359178074,
-0.0576404035,
-0.0160060097,
-0.0785190687,
-0.0099414764,
-0.0161109194,
0.0155336596,
0.0297036208,
-0.0113999732,
0.0271105636,
-0.0490536653,
-0.0255492497,
0.0008813185,
-0.0388201065,
0.0025139856,
0.0488431416,
-0.0080402214,
-0.0017512913,
-0.0125278039,
-0.0060459897,
-0.0246289819,
0.0731558427,
0.0146263437,
-0.0291048046,
-0.0400237925,
0.0615978986,
-0.0000449963,
0.003805944,
0.0195425395,
-0.0442191698,
0.0165392868,
-0.0281612109,
0.0418296345,
0.1082016826,
-0.0105346888,
0.0130062997,
0.0007074007,
0.0122283995,
-0.0352799818,
-0.0056778789,
-0.0097379237,
0.0132852299,
-0.008376196,
-0.0352581702,
-0.0213735867,
0.0669314414,
0.0118553629,
0.0311079323,
-0.0857758895,
0.0077129351,
-0.0036629906,
-0.0018652641,
-0.0136564504,
-0.0258674677,
-0.0055547492,
0.0231717527,
-0.0688407868,
-0.051363267,
0.0347637646,
-0.0371856801,
-0.020969376,
0.0052884361,
0.0391612016,
-0.0168819334,
0.0823250115,
0.0849369317,
0.0229426194,
0.0296933353,
-0.0284145307,
-0.0942951068,
0.0164950695,
0.0127656823,
0.0496505536,
-0.0025464867,
-0.0446952283,
0.0199460015,
-0.0160772409,
-0.0418937877,
-0.023732556,
-0.0122549497,
0.0074833883,
-0.017668806,
-0.0166079216,
0.0110815614,
0.0068920995,
0.0120103229,
-0.0160773415,
0.0152288033,
-0.0212931167,
-0.017581163,
-0.0354295634,
0.0782864094,
0.0341949686,
0.004821057,
-0.0300817993,
-0.0499137081,
0.008922901,
0.0558378659,
-0.016300723,
0.0401803479,
-0.0548841022,
0.0343545377,
-0.010376947,
0.046767544,
-0.0324960388,
0.0501174629,
0.0735906512,
-0.0533339046,
-0.0408519953,
0.0688969195,
0.0310978368,
-0.0006109773,
-0.0049566021,
0.0261896048,
0.0896950737,
0.0458533987,
-0.0003464615,
-0.0341768563,
-0.0184103306,
-0.0029204867,
-0.0424690284,
0.0298807528,
-0.0073426776,
-0.0042801057,
-0.0958022773,
0.0275272485,
0.0172481406,
-0.0579164103,
-0.0297891777,
-0.0210428815,
-0.0253834985,
0.0072759087,
-0.0315192565,
-0.0245829243,
-0.0193414968,
0.0147558097,
0.0104823066,
-0.020348724,
0.0324531496,
-0.0246140901,
-0.0313719735,
-0.007860967,
0.0499155223,
0.0323541053,
0.0156113766,
-0.0019654799,
0.0214262251,
0.0178324897,
0.028424846,
-0.0021292693,
0.0574266426,
0.0305996202,
-0.0106419465,
-0.0134504708,
0.0261384323,
-0.0355376713,
-0.0190037396,
-0.0009188113,
0.0363836549,
0.0056040278,
-0.0181079656,
-0.0269286111,
0.031289719,
-0.0059569962,
0.0455493964,
0.0719661787,
-0.0126853529,
-0.0745704174,
-0.0415252745,
0.0484940931,
-0.0791171864,
0.0024939566,
0.0356536023,
-0.0127926832,
-0.0034597886,
-0.0943349525,
-0.0393057503,
0.0250708628,
0.0383826494,
-0.0040710601,
-0.0310685467,
-0.0229973141,
-0.0345701687,
-0.0456816293,
0.0280922409,
0.0029615669,
0.0170787703,
0.0100569362,
0.0106564183,
-0.0186142642,
-0.0260757972,
-0.0207539778,
-0.0049313325,
-0.0966422409,
-0.0217696205,
-0.0192020629,
0.0859563872,
0.0059010433,
0.0800695047,
0.0306900889,
0.0051808362,
0.0168165341,
0.0424479246,
-0.0239711422,
0.0162699968,
0.0185110997,
-0.0101918988,
0.0032354908,
0.0286170878,
-0.0384359621,
0.0031116055,
0.0260046739,
0.0296412837,
0.0109939771,
-0.0030247357,
0.0362631865,
-0.0233556312,
0.0254663099,
-0.0450956635,
0.0019723056,
-0.0014037855,
0.0199482813,
-0.0373835303,
-0.0208036061,
-0.0637516826,
0.0152696185,
0.0206326637,
0.0697641969,
0.0153094158,
0.0259196907,
0.0168131441,
0.0481202267,
0.0040162136,
-0.0233088341,
0.0585077517,
-0.0184051674,
0.0373215117,
0.0441603661,
-0.0046309894,
0.0245382953,
-0.0214287788,
0.0210737046,
-0.0285903085,
0.0270005986,
0.0382734761,
-0.0052298368,
-0.0167244729,
0.0173797738,
-0.018077828,
-0.0129126869,
-0.0578136258,
0.0178167075,
0.036818631,
-0.0356804878,
0.0483515076,
-0.0414683931,
-0.0051469281,
0.0281975046,
0.0469388068,
0.0183053277,
-0.0575213693,
0.0401226394,
0.0053368607,
-0.001172776,
-0.0502554998,
0.0700737908,
0.0586892106,
0.0232097022,
0.0054664616,
-0.0039763846,
-0.0250523686,
-0.0020103487,
0.0369081199,
0.0066661839,
0.008727299,
0.0057017906,
0.0070026964,
-0.0603798218,
0.005464965,
0.0368339829,
-0.0055343313,
0.007098061,
0.0826784521,
0.0147270635,
0.0054057026,
-0.0183362421,
0.0069038523,
0.0006900438,
0.0205761697,
-0.0248180367,
-0.0375535637,
-5.282968792e-33,
-0.063046068,
-0.0619768985,
0.0022193845,
-0.0110922437,
-0.0256453604,
0.0090740286,
-0.0449535027,
0.0031442018,
0.02677311,
0.0213990528,
0.0239995755,
0.0046714004,
0.0150338626,
0.0601950511,
0.0135579575,
0.0140709393,
0.0405232832,
-0.0317795314,
-0.0132990945,
0.0035064234,
-0.0040736366,
0.0105742346,
0.0420921482,
-0.016070962,
-0.0184642989,
-0.0135968896,
-0.0207938626,
0.0054510627,
0.0132149179,
0.0451400578,
-0.0005317383,
0.0019432605,
0.0115135163,
-0.0048198295,
-0.0339176282,
0.0370959826,
-0.0776024684,
-0.0218947586,
0.0100922454,
0.035871502,
-0.1016345993,
-0.0147495521,
0.0426642783,
-0.0401145741,
0.0220272355,
-0.0657643303,
0.0699313208,
-0.0029283718,
-0.0145970136,
-0.0079479329,
-0.0555784144,
-0.0394750051,
-0.0301108938,
-0.023615852,
0.0325498171,
0.0249054227,
0.0395879,
0.0128078023,
-0.0434948541,
-0.0100173494,
-0.0281454287,
0.0382100828,
0.0358768404,
0.0403712094,
-0.017171571,
0.0501722433,
0.0195305441,
-0.0002423424,
-0.0527632013,
0.0118759787,
-0.0056199487,
0.0453436375,
0.0122081507,
0.0370936058,
0.0087164706,
0.0287202019,
-0.0653207898,
0.0257681441,
-0.0657502189,
0.0508429445,
-0.0002101743,
0.0187157299,
-0.0074427249,
-0.0193129852,
-0.0120230578,
-0.0128859365,
-0.0159388967,
-0.0245362986,
0.0449065603,
0.0072109783,
0.0282685421,
-0.0138258357,
0.0445463508,
-0.0306158941,
-0.0269089937,
-0.0008977332,
-0.0606730878,
0.0401085503,
-0.0115005476,
0.0286010373,
-0.0348023623,
-0.0445948988,
0.0468075946,
0.0321284495,
0.0270755179,
0.0082059251,
0.0910254866,
-0.0059018848,
-0.048374176,
0.0082560219,
-0.0552913472,
-0.0123207271,
-0.0190659966,
0.027025545,
0.0449918769,
0.0034703494,
-0.0422136858,
0.0057578599,
0.0086186649,
-0.0094962893,
-0.0209503416,
0.0042404509,
-0.0130576482,
0.0209640712,
-0.0172398034,
-0.0133982087,
-0.0541434698,
-0.0311710313,
0.0552643426,
-0.0591159835,
-0.0366142467,
0.0128443828,
2.392e-7,
0.0418341272,
0.0233240686,
0.030777324,
0.0204065889,
0.0348502882,
0.0552338883,
-0.0264366213,
0.0312128589,
-0.0468921848,
-0.0145081906,
0.0118440287,
0.0002978328,
0.0497157685,
-0.0198108144,
-0.0955764502,
0.0952108428,
-0.0817210823,
-0.0067073926,
-0.0597524121,
-0.0121696424,
0.0839662254,
0.0712847412,
-0.052786354,
-0.0077806008,
0.0028181081,
-0.0203245878,
0.0821006969,
-0.0165577903,
0.0328935795,
-0.0135178408,
-0.03969495,
-0.0147004137,
-0.0414702557,
0.0264595002,
-0.0164554957,
-0.0398897193,
0.0387456678,
0.0113793314,
-0.02290974,
0.0082741482,
-0.003309292,
0.0016437176,
-0.0014137799,
-0.0750745833,
0.0420140028,
-0.0616638437,
-0.0282682739,
-0.0033521487,
-0.0224354956,
0.0243821815,
-0.0060793282,
0.025815025,
-0.0172836315,
0.0125182997,
0.0565616265,
-0.0175419468,
-0.0134977326,
-0.0339589305,
-0.0992086008,
0.0118280537,
-0.0265910011,
0.0281056296,
-0.0836467966,
0.0464294292,
0.0312424321,
-0.0330994651,
-0.0611779243,
2.23811576e-34,
-0.019852031,
0.0254520923,
0.0236885697,
0.0244211797,
0.0276973601,
-0.0340261906,
-0.0198752228,
0.0152867967,
-0.0177739169,
-0.043616388,
-0.0124943014
] |
http://arxiv.org/abs/2401.14440v2 | http://arxiv.org/pdf/2401.14440v2 | Semantic Sensitivities and Inconsistent Predictions: Measuring the Fragility of NLI Models | 2024-01-25T14:47:05 | 2024-01-31T10:52:52 | Erik Arakelyan, Zhaoqi Liu, Isabelle Augenstein | cs.CL | cs.CL, cs.AI, cs.CY, cs.LG | Recent studies of the emergent capabilities of transformer-based Natural
Language Understanding (NLU) models have indicated that they have an
understanding of lexical and compositional semantics. We provide evidence that
suggests these claims should be taken with a grain of salt: we find that
state-of-the-art Natural Language Inference (NLI) models are sensitive towards
minor semantics preserving surface-form variations, which lead to sizable
inconsistent model decisions during inference. Notably, this behaviour differs
from valid and in-depth comprehension of compositional semantics, however does
neither emerge when evaluating model accuracy on standard benchmarks nor when
probing for syntactic, monotonic, and logically robust reasoning. We propose a
novel framework to measure the extent of semantic sensitivity. To this end, we
evaluate NLI models on adversarially generated examples containing minor
semantics-preserving surface-form input noise. This is achieved using
conditional text generation, with the explicit condition that the NLI model
predicts the relationship between the original and adversarial inputs as a
symmetric equivalence entailment. We systematically study the effects of the
phenomenon across NLI models for $\textbf{in-}$ and $\textbf{out-of-}$ domain
settings. Our experiments show that semantic sensitivity causes performance
degradations of $12.92\%$ and $23.71\%$ average over $\textbf{in-}$ and
$\textbf{out-of-}$ domain settings, respectively. We further perform ablation
studies, analysing this phenomenon across models, datasets, and variations in
inference and show that semantic sensitivity can lead to major inconsistency
within model predictions. | Semantic Sensitivities and Inconsistent Predictions: Measuring the Fragility of NLI Models: Recent studies of the emergent capabilities of transformer-based Natural
Language Understanding (NLU) models have indicated that they have an
understanding of lexical and compositional semantics. We provide evidence that
suggests these claims should be taken with a grain of salt: we find that
state-of-the-art Natural Language Inference (NLI) models are sensitive towards
minor semantics preserving surface-form variations, which lead to sizable
inconsistent model decisions during inference. Notably, this behaviour differs
from valid and in-depth comprehension of compositional semantics, however does
neither emerge when evaluating model accuracy on standard benchmarks nor when
probing for syntactic, monotonic, and logically robust reasoning. We propose a
novel framework to measure the extent of semantic sensitivity. To this end, we
evaluate NLI models on adversarially generated examples containing minor
semantics-preserving surface-form input noise. This is achieved using
conditional text generation, with the explicit condition that the NLI model
predicts the relationship between the original and adversarial inputs as a
symmetric equivalence entailment. We systematically study the effects of the
phenomenon across NLI models for $\textbf{in-}$ and $\textbf{out-of-}$ domain
settings. Our experiments show that semantic sensitivity causes performance
degradations of $12.92\%$ and $23.71\%$ average over $\textbf{in-}$ and
$\textbf{out-of-}$ domain settings, respectively. We further perform ablation
studies, analysing this phenomenon across models, datasets, and variations in
inference and show that semantic sensitivity can lead to major inconsistency
within model predictions. | [
0.0462722406,
0.090480864,
-0.0350354277,
0.0512445979,
-0.0284533855,
0.0048539108,
0.0042240876,
-0.0246285833,
-0.0233701728,
-0.0663111806,
-0.0204595681,
-0.0260236952,
0.018776441,
0.0132740308,
0.0128356013,
-0.0190674104,
0.0574402548,
-0.017380219,
-0.0663841814,
0.0055731549,
-0.0246754549,
0.0165918563,
0.0150264371,
0.0100951325,
0.0097082388,
0.0113618337,
-0.0153352553,
0.0079423971,
-0.0258000381,
-0.0055263625,
0.0000135738,
0.0643577054,
0.0414112732,
0.0628173351,
0.0000020565,
-0.0632730126,
-0.0123262303,
0.0046347147,
-0.0150522953,
-0.020372957,
0.0416913852,
-0.0926343501,
0.0216065124,
-0.005857043,
-0.0314101391,
-0.0054584597,
0.0468265712,
0.0605206601,
0.0188315101,
0.0622658581,
-0.0107053854,
-0.0159389619,
0.0156602189,
-0.0349461623,
-0.0293729212,
-0.045147758,
0.0000460123,
-0.0156420376,
-0.0450776964,
-0.0125322277,
-0.0136723984,
0.0206539128,
0.0623255856,
0.0296600778,
-0.0477654152,
0.0401562452,
-0.0312627703,
-0.029732408,
0.0094603989,
0.0395288095,
-0.0117196208,
-0.0616821051,
-0.0223190542,
-0.0123099191,
-0.0051628835,
0.0334998481,
-0.0230821483,
0.0316701308,
-0.0190840513,
-0.0146486163,
0.0208466966,
0.030026136,
0.0417892337,
0.0487345979,
-0.0049244086,
0.0501268096,
0.0307275094,
-0.0301961079,
-0.0109676085,
0.0029128043,
-0.0431992523,
-0.0427167751,
-0.023468338,
0.0033239331,
-0.0414319523,
0.0313484706,
-0.0961445943,
-0.0524027795,
0.0182373151,
-0.0310315229,
0.0388486125,
0.0598737262,
0.0039552362,
0.0212117732,
-0.0399773158,
0.0402761623,
-0.0288137887,
0.024828257,
-0.0105976639,
-0.0344751775,
-0.0001181022,
-0.0350578167,
-0.0231416207,
0.0275687017,
-0.0348805077,
0.0244849958,
-0.0076985625,
0.0230331291,
-0.0103334524,
0.011443303,
-0.0365265049,
0.020683676,
-0.0889099687,
-0.0152919423,
-0.0274498146,
0.0049799504,
-0.0172254741,
0.0515671372,
-0.0040889499,
-0.069948636,
-0.0398905538,
0.0265912395,
-0.0244853981,
-0.0450971127,
0.0110305836,
0.0906302854,
-0.005121015,
-0.0267470386,
-0.0484807827,
-0.037725158,
-0.0248670932,
-0.0403179601,
0.0287428144,
-0.0454555489,
0.0053087343,
0.0329048932,
-0.0311233383,
-0.0862105489,
0.0080043748,
0.013410137,
-0.0413483493,
0.0952550247,
0.0387031659,
-0.0342473276,
-0.0091788303,
0.0157145895,
-0.0184428375,
0.0719271675,
0.0135506345,
0.0165051278,
0.0255779698,
0.0111734569,
-0.0316958986,
0.0057009622,
0.0346803218,
0.0257107876,
-0.0261452701,
0.0191234034,
0.0024519186,
0.0054443092,
0.049451232,
0.0411878489,
0.0061114645,
-0.0736250207,
0.0706260428,
0.0042773872,
-0.0395841636,
0.0296186134,
0.037319053,
-0.0066337665,
-0.0375959612,
0.0451019555,
-0.0904996991,
0.0263502821,
-0.0528520308,
-0.008403779,
-0.0035332902,
0.0273455884,
-0.0127965333,
-0.0529987291,
-0.0188485552,
-0.0109336842,
0.0344891138,
-0.0222700909,
0.0513112023,
0.092264615,
-0.0279115606,
0.0652477741,
0.0337617584,
-0.0490092523,
0.014031955,
-0.0438426174,
-0.0009058923,
-0.0139956875,
0.0352924615,
-0.0041594957,
0.0079955868,
0.0036769686,
-0.005617829,
0.0658767521,
0.0454742871,
0.0217836276,
0.0210036654,
-0.0199104007,
0.0062994934,
-0.0696999803,
0.0439616852,
0.0212308001,
0.0168518145,
-0.0031137713,
0.0230106991,
-0.082071729,
0.0058012125,
0.0502730012,
-0.0382253267,
-0.0087542972,
0.003087227,
0.001424973,
0.0179322343,
-0.0330184326,
-0.0075563965,
0.0189087521,
0.0089717852,
-0.0300698522,
-0.0519381799,
0.0123041812,
-0.0337614343,
-0.0381445959,
-0.0089362254,
-0.0253770743,
-0.0191648621,
0.048256997,
-0.0554307252,
-0.0019856032,
0.0514479876,
-0.0329375081,
0.0140970275,
-0.0341409966,
0.001714711,
-0.0116187604,
0.0421012826,
-0.0366950147,
0.0802858248,
-0.0111255171,
0.0107912747,
-0.0011719436,
0.0536919236,
0.0331501625,
-0.0098532243,
-0.0153837316,
0.012843729,
-0.039778173,
-0.0270228423,
0.0122608002,
-0.0180517118,
-0.0243590549,
-0.0247863848,
-0.0246307086,
-0.0059852358,
-0.048135113,
0.0139377322,
0.062671788,
0.0339166597,
0.0071188062,
0.0045768493,
0.022764327,
0.0043869852,
0.0150482031,
0.0367019027,
0.0483567007,
-0.0111064734,
0.0044626067,
-0.0393671952,
0.0159109682,
-0.0294974707,
0.0625595674,
0.0277894177,
0.0165341161,
-0.0511330925,
0.0301106814,
-0.0301048998,
0.0355660915,
-0.0252661612,
-0.0396346003,
0.0120742405,
0.0418827645,
0.0315033682,
0.0830210522,
0.004258798,
0.046174679,
0.0625553951,
0.0319072343,
-0.037054386,
0.0043198415,
-0.0301050954,
-0.0175319035,
-0.0279519334,
-0.0410525016,
0.0084244292,
0.0847433433,
0.0167526063,
-0.0165504124,
-0.0550257675,
0.0531160831,
-0.0209921319,
-0.0509640761,
0.0210595857,
0.0057640746,
0.0144631956,
0.0460245945,
-0.0336259939,
-0.0238713007,
0.0899445266,
-0.0263981018,
0.0121884458,
-0.0505650006,
-0.0261639189,
-0.035405878,
0.076637134,
0.0280327443,
0.0080837179,
0.0405892581,
0.0206756182,
-0.1056053117,
-0.0091182366,
0.0421642326,
0.015402195,
0.0435346439,
-0.0373275951,
0.0133109763,
-0.0007105077,
-0.0377390347,
-0.0630831718,
-0.0529345572,
0.0037197291,
0.0141102262,
-0.0076302961,
-0.0160418209,
-0.0203311965,
0.0140267583,
0.000050164,
0.02484034,
-0.012573055,
-0.0221914575,
-0.028605381,
0.0187564753,
0.0773065761,
0.0203719642,
0.0547231324,
0.0091644861,
0.019061042,
0.065691039,
-0.0288142078,
0.0543489642,
-0.0466734208,
0.0077314116,
0.0143763646,
0.0073480965,
-0.0169095546,
0.0185958408,
0.0448508635,
-0.0543407723,
-0.047496926,
0.0748801678,
0.0205890778,
0.0347291082,
0.036539223,
-0.0289663728,
0.0616767332,
0.0250326488,
-0.0057126069,
-0.0397576652,
-0.0415319875,
0.0186520331,
0.0150902541,
-0.024947295,
-0.0181082245,
-0.0175435301,
-0.1031528637,
0.0075975051,
-0.0502016172,
-0.0869157389,
-0.0385090932,
-0.0689085573,
0.0233205631,
-0.0118072424,
0.0428064056,
-0.0197026823,
0.0025430978,
0.0153136523,
-0.0099213403,
0.0057052062,
-0.0450575575,
0.008955786,
-0.0143107353,
0.0004572296,
0.0577284843,
0.0089152791,
-0.0185782313,
-0.0112271132,
0.0486730337,
-0.0312081352,
0.0040777856,
-0.0142711783,
0.0599999689,
0.0129433088,
0.0133148013,
-0.0237440057,
0.0327093229,
-0.0087133907,
-0.0293697212,
-0.0290099569,
0.0213605873,
-0.0040499154,
-0.0404844619,
-0.0248558037,
0.022028897,
-0.0074776798,
0.0907852948,
0.0304853637,
-0.0263428893,
-0.0467471927,
-0.0388548635,
0.0171540249,
-0.027968144,
-0.0491689779,
0.0498146601,
0.0248303618,
-0.0433598496,
-0.0435038842,
-0.0515665486,
-0.0273343753,
-0.0025136769,
-0.0092435684,
-0.0295741614,
0.0056965519,
-0.0244842824,
-0.0344178416,
0.0424963757,
-0.0145490505,
0.0064235874,
0.0170231573,
0.0194920804,
0.018949395,
-0.0351780355,
-0.0365542695,
0.0204108022,
-0.0904954821,
-0.0445738323,
0.016159961,
0.0228923038,
0.0235428121,
0.0504575148,
-0.0305728298,
-0.0140793948,
0.0103043066,
0.0539455377,
-0.005736948,
0.0438740887,
0.0456254408,
-0.0132746194,
0.0251121372,
0.0632519871,
-0.0430381373,
-0.0141903367,
0.0594861843,
0.0138807455,
0.0229510218,
-0.0118831899,
0.0536981225,
-0.0083400467,
0.0422434099,
0.0120573267,
-0.0139770489,
0.0062359781,
0.0169269536,
0.0042738616,
-0.0297818612,
-0.0324721821,
-0.0262558348,
0.0212407038,
0.0437849276,
-0.0149357561,
0.0076868865,
-0.013555388,
0.0740364119,
-0.0158563294,
-0.0174237918,
0.0280177351,
0.0189733263,
0.0036121698,
0.0113384128,
-0.0433908179,
0.0048912391,
-0.0135377347,
0.0531515814,
-0.0168903694,
0.0235375334,
0.0313953236,
0.0093384217,
-0.0237059835,
-0.0191057399,
-0.0834847316,
-0.0113934316,
-0.0011240561,
-0.019939959,
0.0039684521,
-0.0387568064,
0.0214616843,
-0.038103953,
0.0033039723,
0.0065245018,
0.012733723,
0.0271786805,
-0.03976335,
0.0479837395,
-0.0083765043,
-0.0036554022,
0.0189427044,
0.0648296103,
0.0022950002,
0.0811474621,
0.0157949254,
0.0220672768,
-0.0226984471,
-0.056818679,
0.0146395676,
0.0030487753,
0.0076383101,
0.0254218224,
-0.0710454807,
-0.0744328126,
-0.015771471,
-0.0108086439,
0.0099312076,
-0.0145763643,
0.0221385676,
0.0861157179,
-0.00696269,
-0.0228419155,
-0.0077246963,
0.0168741867,
-0.0054225302,
-0.0111898743,
-0.0760363266,
-6.708693197e-33,
-0.0339092761,
-0.0496976897,
-0.0214306396,
-0.0127539765,
0.0083634183,
-0.0288567171,
-0.0549162664,
-0.0195308588,
-0.0523782894,
0.0159296915,
0.0173712745,
0.0294577051,
-0.0125060575,
0.040216174,
-0.027589811,
-0.0731215626,
0.0467598438,
0.0215124693,
-0.0542616993,
-0.0282523222,
0.0423880629,
-0.0100220591,
0.0287624635,
-0.0788328871,
0.0339975022,
0.0048762425,
-0.0034517448,
0.0054013133,
-0.0392308682,
0.0197274275,
0.0140597466,
0.0722023472,
0.0366785042,
-0.0397269838,
0.0190577228,
-0.0037347206,
-0.0652739704,
-0.0303796045,
0.0382767171,
0.0122534893,
-0.023258036,
-0.006673377,
0.0670026615,
-0.0027348062,
-0.0107322223,
-0.0507743247,
0.0217029452,
-0.0055205096,
0.0105383685,
-0.0248847324,
-0.0906839743,
0.0008617916,
0.0052499413,
-0.0056824125,
-0.0424812362,
0.0625602379,
0.0399821959,
0.0313191302,
-0.0651213378,
-0.0187906642,
-0.0109268297,
0.0300501864,
0.0577569976,
-0.0035602078,
-0.0223699175,
0.0416881628,
0.0053245956,
0.0269546714,
-0.0155389486,
-0.0310692899,
0.0000727698,
0.0092261862,
0.0399580598,
0.0726840571,
0.0465858839,
0.0242110249,
-0.0788744017,
0.0012448542,
-0.0326470248,
0.0393715128,
0.0248509925,
-0.0282547977,
-0.0351605788,
-0.0122475438,
-0.0677320361,
-0.0246357545,
-0.0344895609,
-0.0123672597,
0.0311335046,
-0.0118421912,
-0.0145884305,
-0.0087712631,
-0.0110842437,
-0.0341928303,
0.0166441239,
0.006125039,
-0.0014347169,
-0.0395822972,
0.0156682543,
0.0595280416,
-0.0063124183,
-0.0281384718,
0.0339108445,
0.0146956351,
0.0290769972,
0.0118883802,
0.0563817807,
-0.0002082295,
-0.0387020558,
-0.0128860548,
-0.0424480997,
0.0207360107,
-0.018755585,
-0.0000434514,
0.041913338,
0.03972058,
-0.0306572281,
0.0124633005,
0.030164985,
0.0145879965,
-0.0414305516,
0.0141236773,
-0.0190607049,
0.0290789045,
-0.0379400924,
-0.0410165861,
-0.0402947478,
-0.0014807147,
0.0357531197,
-0.0797103569,
-0.0181059781,
0.009273407,
2.924e-7,
0.0671319962,
0.0349122286,
-0.0123248575,
0.051447086,
0.0125428364,
0.0529834442,
0.0156790968,
0.0318252966,
-0.0039292965,
-0.0202365126,
-0.0301966518,
0.0183790252,
0.0362604707,
0.0380857922,
-0.0282287169,
0.0308670867,
-0.0745506436,
0.0208138227,
-0.0239054598,
0.0268926583,
0.0712481216,
0.0972581357,
0.0080909142,
0.0071799648,
0.0070597702,
-0.0406454392,
0.0599414185,
-0.0159664825,
0.0223667305,
-0.0040764967,
0.0223752689,
0.0023843911,
0.0339895785,
0.0466746092,
-0.0078286026,
0.0170938242,
-0.0112342192,
0.0103246542,
-0.0008218089,
-0.0097005134,
-0.0362663083,
-0.0455935411,
-0.0013807621,
-0.0850101188,
0.0868237987,
-0.1112821549,
-0.0498683862,
0.0035581589,
-0.012968326,
0.0359085314,
0.0211861823,
0.0254383981,
-0.0292902347,
-0.0058850693,
0.0252519324,
0.0222781058,
-0.0252278112,
-0.0333566256,
-0.0422044992,
-0.0101271877,
-0.0387893096,
-0.0160484016,
-0.0320967361,
0.0078134127,
0.0839541405,
-0.0498504974,
-0.0162110347,
2.865884615e-34,
0.0188084859,
0.0238346215,
0.0005674313,
-0.0330605097,
0.0629122183,
-0.008762558,
0.0241032988,
0.0097561432,
0.0164017696,
-0.0587404929,
-0.0020809027
] |
http://arxiv.org/abs/2401.14196v2 | http://arxiv.org/pdf/2401.14196v2 | DeepSeek-Coder: When the Large Language Model Meets Programming -- The Rise of Code Intelligence | 2024-01-25T14:17:53 | 2024-01-26T09:23:11 | Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Y. Wu, Y. K. Li, Fuli Luo, Yingfei Xiong, Wenfeng Liang | cs.SE | cs.SE, cs.CL, cs.LG | The rapid development of large language models has revolutionized code
intelligence in software development. However, the predominance of
closed-source models has restricted extensive research and development. To
address this, we introduce the DeepSeek-Coder series, a range of open-source
code models with sizes from 1.3B to 33B, trained from scratch on 2 trillion
tokens. These models are pre-trained on a high-quality project-level code
corpus and employ a fill-in-the-blank task with a 16K window to enhance code
generation and infilling. Our extensive evaluations demonstrate that
DeepSeek-Coder not only achieves state-of-the-art performance among open-source
code models across multiple benchmarks but also surpasses existing
closed-source models like Codex and GPT-3.5. Furthermore, DeepSeek-Coder models
are under a permissive license that allows for both research and unrestricted
commercial use. | DeepSeek-Coder: When the Large Language Model Meets Programming -- The Rise of Code Intelligence: The rapid development of large language models has revolutionized code
intelligence in software development. However, the predominance of
closed-source models has restricted extensive research and development. To
address this, we introduce the DeepSeek-Coder series, a range of open-source
code models with sizes from 1.3B to 33B, trained from scratch on 2 trillion
tokens. These models are pre-trained on a high-quality project-level code
corpus and employ a fill-in-the-blank task with a 16K window to enhance code
generation and infilling. Our extensive evaluations demonstrate that
DeepSeek-Coder not only achieves state-of-the-art performance among open-source
code models across multiple benchmarks but also surpasses existing
closed-source models like Codex and GPT-3.5. Furthermore, DeepSeek-Coder models
are under a permissive license that allows for both research and unrestricted
commercial use. | [
-0.0038363175,
0.0979280844,
-0.0068899821,
0.0119884294,
-0.0165115036,
0.0021811209,
-0.020713916,
-0.0181309208,
-0.0010825768,
-0.0078786593,
-0.0002870272,
0.0126810968,
-0.0308602918,
0.0921939388,
0.0538735837,
-0.0083893267,
0.0165951066,
-0.053712219,
-0.0186282527,
0.0110505503,
-0.0068303039,
0.0347389877,
-0.0032221465,
0.0755218863,
-0.0231292211,
-0.0536436699,
-0.0182204209,
0.0243193861,
-0.0220151413,
-0.0054284036,
0.0514312349,
0.043381799,
0.0048984187,
0.0328972079,
0.0000021792,
-0.0089925779,
-0.0107488865,
-0.0251585897,
-0.0007794963,
0.0389182046,
-0.0152689898,
-0.032044787,
-0.0049099945,
0.0133970482,
-0.0277070627,
-0.0143416105,
0.0517708696,
-0.0068157376,
0.0644180104,
0.0575869419,
-0.0014945007,
-0.0151378447,
0.0486329086,
-0.0312117469,
0.0369686149,
-0.0423616618,
-0.0055123745,
0.0014269514,
-0.0118217515,
-0.009465036,
-0.0031594625,
0.0075454367,
0.047709398,
-0.0234261937,
0.0516474359,
0.0419003516,
-0.0539327115,
-0.0126100117,
-0.0036436138,
0.0489425771,
-0.0979641601,
-0.0379698835,
0.0096811354,
-0.0250470769,
-0.0313246548,
-0.0688745305,
-0.0383034386,
0.0213562436,
0.0237692371,
0.004173771,
0.0197582766,
-0.0197331626,
0.0232233722,
0.0006153979,
-0.0155927455,
0.0403161533,
0.01920674,
-0.0017212799,
-0.0335635915,
-0.0265964922,
-0.0018240782,
-0.0844279006,
-0.0120307375,
0.0432924107,
-0.0301056001,
-0.0144455722,
-0.0546372607,
-0.0276225619,
0.0110263629,
0.0043656412,
0.026355058,
0.0680599362,
-0.0498842672,
0.0204868466,
-0.0561383329,
0.0689903051,
-0.0393900499,
-0.0033711395,
-0.0229341798,
-0.0147286346,
-0.0502080433,
0.0035558983,
-0.0536193848,
0.0408382677,
-0.0360890888,
-0.0018925534,
-0.024390161,
0.0407602228,
0.0154326735,
0.0448834524,
0.0052606775,
0.0245115981,
-0.0468622334,
-0.0036126734,
0.0223692507,
-0.0156331565,
-0.0286833066,
-0.0487503894,
0.0104830023,
-0.0560562685,
-0.0093357107,
0.0404408276,
-0.0179315135,
-0.0473383553,
0.0318407118,
0.0586646572,
0.0230840147,
0.0205916259,
-0.023713693,
-0.0141566461,
0.0150721427,
-0.0375853218,
-0.0356989317,
-0.1117945313,
-0.0319918469,
0.0283129066,
-0.0213434231,
-0.071268335,
0.0077967602,
0.0159874521,
0.0171563737,
0.031890545,
-0.0458312109,
0.007814276,
0.002873898,
-0.0189068113,
-0.0110088186,
0.0419286601,
0.0510545895,
0.0410206802,
-0.0155632282,
-0.0255582202,
0.0439604409,
0.0497727767,
-0.0157710407,
-0.006805439,
0.0131164854,
0.0174512826,
0.0413087681,
0.0023183734,
0.0493676104,
0.0438659675,
0.0043105655,
0.004891695,
0.0105534103,
0.0586798638,
0.0055089244,
0.0520754606,
0.092614688,
0.0127847791,
0.0073862472,
-0.0509913489,
-0.0417991988,
0.0670223162,
0.0459814779,
0.0052665216,
0.0095906397,
0.0353879966,
-0.0197751597,
-0.0529258661,
-0.0137646832,
-0.0283041149,
-0.0360516682,
-0.0042372337,
0.025291834,
-0.0001396054,
-0.0653533116,
0.054724168,
0.0052675814,
-0.0462801456,
-0.0118353879,
-0.0189670902,
0.0229439475,
0.068680957,
0.0551528446,
-0.0625736862,
-0.0495941229,
-0.0369573012,
-0.0047606383,
-0.0265376139,
0.0300450474,
-0.0113146137,
0.0101194587,
-0.0692974254,
-0.0106445421,
-0.064824827,
-0.0008555763,
0.0467930287,
0.0281110238,
-0.0258173645,
0.0379427113,
-0.0152367437,
-0.049060341,
0.0351839848,
-0.0283980127,
-0.0268945638,
0.0094231563,
-0.0063739843,
0.0274793655,
-0.0357465222,
-0.0247230772,
0.0077118897,
0.0332185961,
-0.0396238789,
-0.0345811471,
-0.0129429707,
-0.0507608727,
-0.0153536992,
0.0231985431,
-0.0607928857,
-0.0475429967,
0.044642169,
-0.0377168246,
-0.0297431089,
0.0704726279,
-0.0094722919,
0.0502630584,
-0.013587337,
0.006305675,
-0.0138522759,
-0.021086406,
-0.0043108766,
0.0217712186,
0.0155994808,
0.0169002879,
0.016903583,
0.0729732662,
0.0355459265,
-0.0011065819,
0.0153866084,
-0.0066904672,
-0.0546616353,
0.0165633857,
-0.013339554,
0.0535881035,
-0.0233843829,
-0.0505959019,
-0.0885240808,
0.0573840514,
-0.0209801774,
0.0100554749,
-0.0089918654,
0.0406182595,
0.0200186726,
-0.0058846343,
-0.0390264615,
-0.0224789046,
-0.0156045854,
-0.0184504185,
0.0118844286,
-0.0322522745,
-0.023871094,
-0.03229177,
0.0050170165,
-0.0481264815,
0.0125882067,
-0.0226520579,
-0.0154045271,
-0.0313124396,
0.049760107,
0.0190471429,
0.0235317275,
-0.016947072,
-0.0441247709,
-0.0053028069,
-0.0477052256,
0.0825832337,
0.0509296879,
-0.0020823698,
0.0005629911,
0.0118588824,
0.002888558,
-0.0434327498,
0.0400224924,
0.0417392813,
-0.0237166099,
-0.0048472402,
0.001568738,
0.0239920802,
0.0867168829,
-0.0158230644,
0.0020308858,
-0.0689172223,
0.0199181009,
-0.0009538389,
0.0085839713,
0.0174154267,
-0.007876493,
-0.025768647,
0.0391965136,
-0.0241504852,
-0.0700225085,
0.0433387794,
-0.0322427452,
-0.0388959646,
-0.0149780773,
0.0057494808,
-0.0743579865,
0.0436343029,
0.0193558112,
0.015572709,
-0.0079437932,
-0.0327671878,
-0.0315026715,
0.0402622931,
-0.0496367514,
-0.0011900149,
0.0567674786,
-0.0277015623,
0.0262774006,
-0.0256632902,
-0.0462628491,
-0.0485632196,
-0.0288218949,
-0.0372186154,
0.0303087197,
0.0629881248,
-0.0184035785,
0.0010643533,
-0.0005338714,
-0.0159642193,
0.0785480887,
0.0215973873,
-0.0176847242,
-0.0297517125,
-0.0308673903,
0.0178634822,
-0.0642097667,
-0.0182183739,
-0.016854411,
-0.0386709347,
0.0221239477,
0.0541603193,
0.0447523296,
-0.0781268403,
0.0165376365,
0.0204176381,
0.002016491,
-0.0245613027,
0.03740298,
0.0590495802,
-0.0507695563,
-0.0272053927,
0.1031576842,
0.0197611134,
-0.0543360859,
0.0083884485,
0.0021337585,
0.049970489,
0.0838139504,
0.0098539731,
-0.0837301984,
-0.0516592748,
0.0140249329,
0.0118992254,
0.0174742956,
-0.0301532391,
0.0003618638,
-0.0865556523,
0.0204914603,
0.0117126666,
-0.0341544449,
-0.0005179732,
-0.0259159151,
-0.0178916845,
-0.0190026984,
-0.0517812297,
0.0178918373,
0.0089483708,
0.0240010433,
0.0152114918,
-0.0226491485,
0.0047791842,
0.0433102064,
-0.0499015562,
0.0391328707,
0.1076980755,
0.009056123,
0.0076016337,
-0.0055121016,
0.0414396524,
-0.0184995811,
-0.0081144385,
-0.0121767549,
0.0458410904,
-0.0005041487,
-0.0348170176,
0.0475239307,
0.0398989692,
-0.0967588574,
-0.0110235438,
0.0397681929,
0.0781137273,
0.0146786124,
-0.0145062963,
-0.017828688,
-0.0392771997,
-0.0056571504,
0.0465403348,
0.0474340543,
-0.0322111212,
-0.0390143618,
-0.0780087337,
0.0023147771,
-0.1009387597,
-0.0001122113,
-0.0139149381,
0.0378218815,
-0.0000042658,
-0.0352427661,
0.0017672166,
0.0202826131,
0.0350074321,
0.0234268252,
-0.0377788693,
0.0141824335,
-0.0423182435,
-0.0630884767,
0.037263874,
-0.0349053852,
0.0140779279,
-0.0658346787,
-0.0195406862,
0.0021769288,
-0.0210322887,
-0.0154145202,
0.0252382997,
-0.1080322638,
-0.0371447578,
-0.0283021051,
0.0819448158,
0.0151889501,
0.0146051859,
-0.0327839442,
-0.0383809768,
0.0000291312,
-0.0192279313,
-0.035616897,
-0.0028693511,
0.0151313227,
-0.0075656362,
0.0188702531,
0.0216095038,
-0.0642238557,
-0.0043262444,
0.0332881436,
-0.0170118902,
0.0143257156,
0.0207505282,
0.0381549224,
-0.0088927653,
0.007283479,
0.0256952252,
0.0014323558,
-0.0137927374,
-0.0041063903,
0.02190575,
-0.0307837464,
-0.0511787646,
0.0519054271,
0.0317595825,
-0.0152147878,
0.0265082549,
0.0313470773,
-0.0018027589,
0.02361333,
0.0084929047,
0.0177177619,
0.0124578262,
0.0491870679,
0.0914742723,
-0.026472738,
-0.0176615231,
0.0111910971,
-0.0289520603,
0.0091987969,
-0.018237425,
0.010391687,
0.0208304562,
0.0513413437,
0.010079273,
-0.0081645558,
-0.0414729081,
-0.0561170429,
-0.0074916482,
0.0000124337,
0.0404776856,
0.0169905685,
0.0198826455,
-0.0129498793,
0.010207097,
-0.0237511043,
0.0630183518,
0.007133293,
-0.0305536985,
-0.0116092386,
0.0274213534,
-0.0219132546,
-0.0317355208,
0.1106640771,
-0.004015117,
0.0499476008,
-0.0017007018,
0.0324960835,
-0.0410958305,
0.0297104791,
0.0351190045,
0.0114186592,
-0.0035433453,
-0.0091340356,
0.0009468158,
-0.0431144834,
0.0256596394,
0.0304433759,
0.0136951851,
0.0073238481,
0.0370460711,
0.0984835476,
-0.0100485943,
-0.0226057861,
-0.0035929794,
0.0141634075,
0.0060253511,
-0.0617822446,
-0.0377125666,
-6.219284205e-33,
0.019347053,
-0.0096811187,
0.0146215344,
0.0121058356,
-0.0091871079,
0.0142378351,
-0.0526663549,
-0.0214669462,
0.0186586268,
-0.0303060859,
0.0155978957,
0.0264600273,
-0.0136117749,
0.0005330037,
0.0760642812,
0.026242204,
0.0751783848,
0.0364076085,
-0.0398237146,
0.0539183393,
-0.0342909172,
0.0338129848,
0.014324774,
0.0028974998,
-0.0279603451,
-0.0142612504,
0.0336210839,
-0.0036069611,
0.0302952193,
0.0323677957,
-0.0089352364,
0.0062602041,
-0.0118499845,
-0.0243845992,
-0.0385427475,
-0.0286967941,
-0.024108151,
-0.0090244785,
-0.0065196478,
0.0091608437,
-0.0253741983,
-0.1027982384,
0.0681460053,
-0.0016039386,
-0.0137520619,
-0.0171119813,
0.0282917637,
0.005804677,
0.0379291475,
-0.0209035873,
-0.0799498856,
0.0259734169,
-0.0609830394,
-0.030791223,
-0.0052591749,
0.0159148909,
0.0089416746,
0.0449735485,
0.0052516568,
-0.0281025339,
-0.0221511126,
0.0637914538,
0.0668568239,
-0.000159959,
0.0069852439,
0.0292963721,
0.0399624184,
0.0260874052,
-0.0127417203,
0.0200758819,
0.0246879663,
0.0678879991,
0.0342990682,
0.0287753362,
0.0365661494,
-0.0111659206,
-0.0367239006,
0.0206381492,
-0.0111914268,
0.0258148275,
0.0100833327,
0.0514897034,
-0.0126206754,
-0.0011468717,
-0.0267682392,
-0.0579762198,
-0.0271076541,
-0.0221368782,
0.0491831452,
0.014459284,
-0.0052293092,
0.0152596682,
0.0251934715,
-0.0466771014,
0.0260601696,
0.0463984273,
-0.0368709564,
0.0256031733,
-0.0150142629,
0.0443171822,
-0.0245792456,
-0.042980697,
0.0107670706,
0.0199588556,
0.089739807,
0.0242895558,
0.0360331833,
0.02812564,
-0.0464018434,
-0.0059751179,
-0.0650741532,
0.0079517197,
0.0129840681,
0.0429464802,
-0.0044948589,
-0.0037291518,
0.0207340531,
0.035199862,
-0.0186272636,
-0.0285092667,
-0.0597389527,
-0.0259739961,
0.0095466822,
0.0276823621,
-0.0365516655,
-0.0228041708,
-0.0764801502,
0.0123662325,
0.0155013902,
-0.0402391814,
-0.0534733385,
0.0112584596,
2.737e-7,
0.0063601527,
0.0481824204,
0.0300417077,
0.012972231,
-0.0013092645,
0.0476997644,
-0.0226685144,
0.011509019,
0.0145274037,
0.0147534218,
-0.0389322229,
0.0231040139,
0.0462847613,
0.0167009626,
-0.0810087994,
0.0379683971,
-0.0250431355,
-0.0041558929,
-0.0827693939,
0.0104510132,
-0.0058231517,
0.0915373564,
-0.0028619999,
-0.019061707,
0.0599362254,
-0.0312023405,
0.0226567555,
0.0255169887,
0.004002878,
0.0407169387,
-0.0667313784,
0.0292640142,
-0.0069579198,
-0.0420037434,
-0.0007802075,
-0.0195690878,
0.0311286412,
0.0666987747,
0.0027509986,
0.0349075608,
0.0412229411,
-0.0183095541,
-0.0017445809,
-0.0941219702,
0.0535293072,
-0.0288294051,
0.01890889,
-0.0371120423,
-0.0099302381,
0.0052851271,
0.0043122643,
-0.0343560986,
0.0296495482,
0.0141025456,
0.0022944787,
-0.0523756929,
-0.0325529836,
-0.0206381753,
-0.0378543176,
-0.0052362499,
-0.0215311982,
-0.0112397708,
-0.075109303,
0.028141031,
0.0118895546,
-0.0187504534,
0.0322574787,
2.709628355e-34,
0.0354942642,
-0.0335531458,
0.0224061552,
0.0038049237,
0.0377305523,
-0.0244058426,
-0.0095364545,
0.0155599248,
-0.0194493607,
-0.0206115786,
-0.0497958772
] |
http://arxiv.org/abs/2401.14194v3 | http://arxiv.org/pdf/2401.14194v3 | Parameter-Efficient Conversational Recommender System as a Language Processing Task | 2024-01-25T14:07:34 | 2024-02-25T03:00:48 | Mathieu Ravaut, Hao Zhang, Lu Xu, Aixin Sun, Yong Liu | cs.CL | cs.CL | Conversational recommender systems (CRS) aim to recommend relevant items to
users by eliciting user preference through natural language conversation. Prior
work often utilizes external knowledge graphs for items' semantic information,
a language model for dialogue generation, and a recommendation module for
ranking relevant items. This combination of multiple components suffers from a
cumbersome training process, and leads to semantic misalignment issues between
dialogue generation and item recommendation. In this paper, we represent items
in natural language and formulate CRS as a natural language processing task.
Accordingly, we leverage the power of pre-trained language models to encode
items, understand user intent via conversation, perform item recommendation
through semantic matching, and generate dialogues. As a unified model, our
PECRS (Parameter-Efficient CRS), can be optimized in a single stage, without
relying on non-textual metadata such as a knowledge graph. Experiments on two
benchmark CRS datasets, ReDial and INSPIRED, demonstrate the effectiveness of
PECRS on recommendation and conversation. Our code is available at:
https://github.com/Ravoxsg/efficient_unified_crs. | Parameter-Efficient Conversational Recommender System as a Language Processing Task: Conversational recommender systems (CRS) aim to recommend relevant items to
users by eliciting user preference through natural language conversation. Prior
work often utilizes external knowledge graphs for items' semantic information,
a language model for dialogue generation, and a recommendation module for
ranking relevant items. This combination of multiple components suffers from a
cumbersome training process, and leads to semantic misalignment issues between
dialogue generation and item recommendation. In this paper, we represent items
in natural language and formulate CRS as a natural language processing task.
Accordingly, we leverage the power of pre-trained language models to encode
items, understand user intent via conversation, perform item recommendation
through semantic matching, and generate dialogues. As a unified model, our
PECRS (Parameter-Efficient CRS), can be optimized in a single stage, without
relying on non-textual metadata such as a knowledge graph. Experiments on two
benchmark CRS datasets, ReDial and INSPIRED, demonstrate the effectiveness of
PECRS on recommendation and conversation. Our code is available at:
https://github.com/Ravoxsg/efficient_unified_crs. | [
0.069549799,
0.0165071394,
0.0114839245,
0.0451992191,
-0.0187417697,
-0.0178784933,
-0.0072698076,
-0.0231041182,
0.0433633849,
-0.0625286624,
-0.0290770698,
0.0330089927,
0.0029810569,
0.0481046885,
-0.0319486521,
-0.0422948971,
0.0093234703,
-0.0192085505,
0.0247522034,
0.0198355392,
0.0341768228,
0.0468230918,
-0.0024641925,
0.0250695981,
0.017731946,
-0.0049736621,
-0.0375369228,
0.0049470472,
-0.0159276705,
-0.0010944579,
0.0427346863,
0.0287739486,
0.0139528075,
-0.030567145,
0.0000018742,
-0.045409888,
-0.01808035,
-0.0403344557,
-0.0075681219,
-0.0367331356,
0.0489726923,
0.0695001557,
0.024509877,
0.0347986594,
-0.0457332805,
-0.0051044039,
0.0266706981,
0.081770733,
0.0450004712,
0.0922888443,
-0.0245595574,
-0.0024348721,
-0.0000251742,
-0.0087400461,
0.0379540808,
-0.0211883858,
-0.0120505737,
0.009627901,
0.0214901138,
0.046854496,
0.0232898686,
-0.0111445282,
-0.0116662672,
-0.0252752714,
0.058580853,
0.0111085176,
-0.0244293474,
-0.024519816,
-0.0208442975,
0.1243146434,
-0.0407561064,
-0.0207305681,
-0.0052183545,
0.0004359643,
0.0006471372,
-0.0403929949,
-0.0146676842,
-0.0105976667,
-0.0045040082,
0.0246713087,
0.0120579908,
0.0242608413,
0.0206200611,
-0.0076673636,
0.0098541481,
0.0735148862,
0.0048191287,
-0.0231893733,
-0.0049009384,
-0.0209246874,
-0.0000461608,
-0.1020905152,
0.0219928417,
0.0188668519,
-0.0646146834,
0.030592056,
-0.0454081036,
0.048934292,
-0.0006770104,
-0.0376679376,
0.0030328052,
0.0537500344,
0.0475949906,
0.0310128294,
0.0042027114,
0.0176124685,
0.0410257764,
0.0014716665,
-0.0717041194,
-0.0157986376,
-0.0311913621,
0.0095977271,
-0.0484379753,
0.0309850425,
-0.0683343261,
-0.0579425283,
-0.0229099449,
0.0617742501,
0.0280568115,
0.0275061112,
0.026802225,
0.0697626024,
-0.0028116638,
0.0101747438,
-0.0904261023,
0.0092599299,
-0.0406349897,
-0.0017117521,
0.0223091766,
-0.0222221278,
0.0287684016,
0.0258841105,
0.0368024297,
-0.0439930521,
-0.0131178591,
0.044819396,
0.0065159961,
-0.0359766036,
-0.0094441734,
-0.0357671157,
-0.001503882,
-0.0539906844,
-0.0421421975,
-0.0460390002,
0.0666219369,
0.0424499549,
-0.0126452595,
-0.02793153,
0.0374871641,
0.0447797813,
-0.0247416552,
0.0849171653,
-0.0268270019,
-0.021996038,
0.0043364344,
0.029598888,
-0.0088610714,
-0.0074784481,
0.0226227716,
0.0372974053,
0.062264666,
0.0130126365,
-0.01655153,
-0.0119567979,
0.0050435876,
-0.0253058411,
0.0122422939,
0.058500044,
-0.014327487,
0.0556168631,
0.0234013628,
0.0483339019,
-0.0547872335,
-0.0148219336,
0.0605315231,
0.0177665763,
-0.0112714842,
0.0426024571,
0.0323473886,
0.0735077187,
-0.0360486992,
-0.037668217,
-0.0286201797,
0.0207734089,
-0.031763982,
-0.0349257998,
0.0482720435,
-0.0250263177,
-0.040879678,
-0.045839902,
0.0332251564,
-0.0676966533,
0.0699368715,
0.0189667474,
-0.0272402614,
0.0010964156,
-0.075891912,
0.0213724654,
0.0349582136,
-0.0649830997,
-0.0243900642,
0.0053105555,
0.0616116002,
0.1032257378,
0.0518232472,
0.0046273917,
-0.0480937883,
0.0152856698,
-0.0155245624,
-0.0074537336,
0.0464791879,
-0.0312021766,
0.0397180282,
-0.0560761541,
-0.0444063544,
-0.0286030751,
-0.0538718142,
0.0281788912,
0.0131618213,
-0.0025187025,
0.0392474122,
-0.0106943287,
-0.0270221215,
-0.0154769067,
0.0055631036,
-0.0034173608,
0.053684067,
0.0028545,
0.0281208344,
0.0044871224,
0.0335285068,
0.0076338239,
0.0122395726,
0.0073356405,
-0.0104946084,
0.0434224792,
-0.0516392738,
-0.0745495334,
0.0206682086,
-0.0044632256,
0.0178914946,
0.0126075,
0.0032453856,
-0.0613175146,
0.0180133339,
-0.0034760621,
0.0667950138,
-0.0176288933,
0.0089896005,
0.0417745598,
0.0257433001,
-0.020848833,
0.0523761809,
0.034390334,
0.0138475522,
0.0148114869,
0.0178466365,
0.0349892937,
-0.0504955351,
-0.0197803974,
-0.0005258517,
0.0198750161,
0.0184724648,
0.0441251509,
0.0256672781,
-0.0382442884,
0.002420384,
-0.1111328304,
0.0285091382,
0.0137660541,
-0.0177973974,
-0.0133919157,
0.0600873828,
0.0156559385,
-0.0459831692,
-0.0532115623,
0.0344611928,
0.0236636959,
-0.0181082152,
0.0325774774,
0.0272092652,
-0.0111655947,
-0.0559193566,
0.0194933452,
0.0223542992,
0.0340422317,
0.0201441776,
0.0154193407,
-0.0498056784,
0.0246754717,
-0.0675362721,
0.0248135384,
0.0063054697,
-0.0030677165,
-0.0111157438,
0.0004097583,
0.0389533713,
0.0202759895,
0.0051041134,
0.0265027396,
0.009245554,
-0.0184570793,
-0.0904583484,
0.012058991,
-0.0315954275,
0.0254693646,
-0.0254943669,
0.0052114283,
0.0256265868,
0.005810542,
0.0001335083,
0.0005194761,
-0.04706778,
0.0222054739,
0.0574255027,
0.0066720732,
-0.0403804518,
-0.0265080873,
0.0064955307,
0.0494400784,
-0.025658695,
-0.0161791965,
0.0524731316,
-0.0353934914,
-0.0248598736,
-0.0675921589,
-0.0315928049,
-0.0443601012,
-0.0279753711,
0.0429942645,
0.0431480259,
-0.0210804809,
-0.026883347,
-0.0032488322,
0.0030512374,
0.0171301849,
-0.023232203,
0.0032378603,
-0.0109511493,
0.0166292544,
0.0243750941,
-0.0451189354,
-0.0598632656,
-0.0460618176,
0.0094234599,
0.0029142704,
0.0015738212,
-0.0081646582,
-0.0405461602,
-0.008683566,
-0.0284939352,
0.0434184186,
0.0052526928,
-0.0181848668,
-0.0122113768,
0.0254972614,
0.0498261005,
-0.0288961288,
-0.006130801,
0.0320977531,
0.0549868494,
0.0283600688,
0.0267351065,
-0.032164488,
-0.048261717,
-0.0036120373,
0.0347744226,
-0.0234068427,
-0.0026629285,
0.0451239459,
0.0262144301,
0.0293873046,
-0.0277400538,
0.1483008564,
0.0489025265,
-0.0589003451,
0.0101262778,
0.0043318756,
0.0245137662,
0.0120611703,
0.0443821363,
-0.0532494411,
-0.0363418385,
-0.0059311208,
0.008637242,
0.0050603906,
-0.0483548529,
-0.0314854346,
-0.095071502,
0.0100058196,
-0.0113529516,
0.0252279118,
0.0001014961,
-0.0213978197,
-0.0147054,
-0.0143798916,
0.0123150889,
-0.0206603277,
-0.0059571434,
0.0131429527,
-0.0345624462,
-0.0112220282,
-0.001467496,
0.0353088789,
-0.0297478475,
0.0259997137,
0.0357303172,
0.0394810699,
0.0397672281,
0.0116960322,
0.0065248637,
0.004590733,
-0.0251931287,
-0.0168225057,
0.0190856662,
-0.041632995,
0.0065000956,
-0.0265486445,
0.0042131175,
-0.0660730526,
-0.0310505666,
-0.0076446678,
0.0239089802,
-0.0589845106,
0.0218852013,
0.000574219,
0.049219586,
0.07344836,
0.0853242651,
0.0147650493,
-0.027514372,
-0.0060701156,
-0.0709854215,
-0.0067568156,
-0.014115639,
-0.0011660955,
-0.0110135293,
0.0325274169,
0.0364647321,
-0.0763999,
-0.0542070903,
-0.0064651389,
-0.0139579065,
-0.0010639486,
0.0079290112,
0.010711832,
-0.0040912516,
0.0096815266,
0.0258139819,
-0.0044813338,
0.0002227654,
-0.0366229676,
-0.0093975775,
0.0109270308,
-0.073114574,
-0.0154736517,
0.0092746345,
-0.0412519947,
-0.0469344147,
-0.0792166814,
0.0623728037,
0.0153249716,
0.0834622458,
-0.0524601713,
-0.0496102497,
-0.0209420491,
0.001272158,
-0.0519846641,
-0.0325034931,
-0.0006985466,
0.074636288,
0.0213187691,
0.0477904715,
-0.0468057692,
0.0037102995,
-0.0405905768,
0.0217984915,
-0.0303207748,
-0.112296395,
0.027674295,
-0.0008641322,
0.0128002157,
0.007739495,
0.0447736382,
-0.0172822699,
0.0286214799,
-0.0403381586,
-0.0600840226,
-0.0691455454,
-0.0029833894,
0.0141717633,
-0.0014247486,
-0.0175521728,
-0.0280648489,
-0.0491375364,
0.0376802348,
-0.0002032037,
-0.0164340492,
0.0093634399,
-0.0490653627,
0.0308924615,
0.0035727145,
-0.0066219601,
0.0433519408,
0.0091962144,
0.0256319325,
-0.0430335179,
0.0740671307,
-0.0095112836,
-0.0458043925,
-0.0052456553,
0.0346365049,
-0.0220133793,
-0.0536417812,
0.0078243967,
0.0399787724,
0.0091584679,
-0.0333428308,
0.0168137681,
-0.0376909599,
0.0282837246,
-0.0389913507,
0.0733434781,
-0.021740593,
-0.0464089997,
0.0399128795,
0.0220228136,
-0.0254736468,
0.0299985092,
0.0551849566,
-0.0201139655,
-0.0148389535,
0.028880449,
-0.0367350653,
-0.0525293089,
-0.0201700814,
-0.0099626528,
0.0566455536,
-0.0350781567,
0.0078833234,
0.0397981815,
-0.0294479225,
-0.0077659753,
-0.0421968549,
-0.0215521045,
-0.0112293754,
0.0633782297,
0.0612009726,
-0.0156794041,
0.0247819293,
-0.0622108318,
0.0034096411,
-0.0503545627,
-0.0285654422,
-0.0562720709,
-5.85630955e-33,
-0.0120740971,
-0.0007605695,
0.0247303657,
0.0551354177,
0.0074028559,
-0.0369495638,
-0.0350468121,
-0.0033800735,
0.029604869,
-0.0051892581,
-0.0084165288,
0.0127033032,
0.0074041001,
0.0532056354,
-0.0311608687,
-0.0396168716,
0.0388323367,
0.009244943,
-0.0104239769,
0.0504701398,
0.0608768612,
0.0348577574,
0.0034600373,
-0.0721214265,
0.0096241217,
-0.0238708165,
0.0219492633,
0.0320868194,
0.0510783717,
0.0571294203,
-0.0242021587,
0.0637233257,
0.0006095006,
0.0089019071,
0.0049954001,
0.0239687506,
-0.0751761496,
-0.0132089388,
0.0416785255,
0.0265783053,
-0.0463500209,
-0.023472432,
0.0021763504,
-0.0282574203,
0.0379615501,
0.0230045505,
0.0268384367,
0.0010556834,
-0.0459163003,
-0.0077472376,
-0.0291132294,
0.0037889704,
-0.0374759883,
-0.0448500887,
-0.0099352673,
0.0688672066,
0.0030451308,
0.0231356844,
-0.0437693335,
0.0240733996,
-0.0404716358,
0.016805185,
-0.0390549637,
-0.035737399,
-0.0547039136,
0.054775171,
0.0062510511,
-0.0168309081,
0.0132311061,
0.056029819,
-0.0442966074,
0.0676245242,
-0.0054250378,
0.0395857468,
-0.0022302852,
0.002831528,
-0.0539912879,
0.0488910712,
0.0224675611,
0.0039670002,
-0.0029882686,
0.0101841073,
-0.041463796,
-0.0257695373,
-0.0298677459,
-0.0085873501,
-0.0510043241,
-0.0409715846,
0.0292262994,
0.0344098136,
0.0106345499,
-0.0044181375,
0.016347561,
-0.0433052555,
0.0016679766,
0.0358722657,
-0.0396112017,
-0.0343772061,
-0.0129867671,
0.0235653799,
-0.0455333814,
-0.0153959375,
0.0477720574,
-0.048524648,
-0.0239950791,
0.0287283175,
0.0217816066,
0.0084982337,
-0.0137732178,
0.0137910098,
-0.0144406836,
0.0059677516,
-0.0479695201,
0.058982797,
0.0169445276,
0.0486881137,
-0.0545799024,
0.0670463517,
-0.0214844868,
0.014358297,
-0.0379445218,
-0.099759087,
0.0396071635,
-0.0031874704,
-0.0226224903,
-0.0337822102,
-0.018840516,
-0.0241013169,
0.0264176633,
-0.0508447625,
0.0045075542,
-0.007044815,
2.651e-7,
-0.0121180722,
0.0204425305,
-0.0117592299,
0.0140744019,
0.0233541485,
0.035341695,
0.0382562168,
0.0644756854,
0.0340067744,
0.002368117,
0.0287578069,
0.0116151273,
0.0050564255,
0.0082679205,
-0.0578640774,
-0.0205867458,
-0.0532566532,
-0.0430630408,
-0.0445042625,
-0.0064682136,
0.0538085438,
0.0810130686,
0.020371506,
0.0076150983,
-0.0361088701,
-0.0151135605,
0.0536767729,
-0.0718082786,
-0.0448217764,
0.0106763998,
0.0451220796,
0.018663615,
-0.0254197493,
-0.0094483113,
-0.0202021934,
0.0520151556,
0.0176080912,
0.0600600466,
0.0010373726,
-0.0548320971,
-0.0039255642,
-0.0037585709,
0.000067477,
-0.0322419368,
0.0770020857,
-0.0351710431,
0.0017306914,
-0.0082101449,
-0.0270407386,
-0.0031560953,
0.0095421104,
-0.018883314,
-0.0279108323,
0.0005903436,
0.028353624,
-0.0136042796,
0.0069024083,
-0.0720989108,
-0.0108248955,
0.0034077561,
-0.0250870902,
-0.0063656955,
-0.0264107529,
-0.0077777598,
0.0204208847,
-0.0475786999,
0.0235391632,
2.341183196e-34,
0.0152353328,
0.0266786702,
-0.0093468092,
-0.0200227704,
0.0530284531,
-0.0116748754,
-0.0184291713,
-0.0001970161,
0.0471027307,
-0.0847035497,
-0.0336727388
] |
http://arxiv.org/abs/2401.14192v2 | http://arxiv.org/pdf/2401.14192v2 | How Can Large Language Models Understand Spatial-Temporal Data? | 2024-01-25T14:03:15 | 2024-05-17T03:10:20 | Lei Liu, Shuo Yu, Runze Wang, Zhenxun Ma, Yanming Shen | cs.LG | cs.LG, cs.CL | While Large Language Models (LLMs) dominate tasks like natural language
processing and computer vision, harnessing their power for spatial-temporal
forecasting remains challenging. The disparity between sequential text and
complex spatial-temporal data hinders this application. To address this issue,
this paper introduces STG-LLM, an innovative approach empowering LLMs for
spatial-temporal forecasting. We tackle the data mismatch by proposing: 1)
STG-Tokenizer: This spatial-temporal graph tokenizer transforms intricate graph
data into concise tokens capturing both spatial and temporal relationships; 2)
STG-Adapter: This minimalistic adapter, consisting of linear encoding and
decoding layers, bridges the gap between tokenized data and LLM comprehension.
By fine-tuning only a small set of parameters, it can effectively grasp the
semantics of tokens generated by STG-Tokenizer, while preserving the original
natural language understanding capabilities of LLMs. Extensive experiments on
diverse spatial-temporal benchmark datasets show that STG-LLM successfully
unlocks LLM potential for spatial-temporal forecasting. Remarkably, our
approach achieves competitive performance on par with dedicated SOTA methods. | How Can Large Language Models Understand Spatial-Temporal Data?: While Large Language Models (LLMs) dominate tasks like natural language
processing and computer vision, harnessing their power for spatial-temporal
forecasting remains challenging. The disparity between sequential text and
complex spatial-temporal data hinders this application. To address this issue,
this paper introduces STG-LLM, an innovative approach empowering LLMs for
spatial-temporal forecasting. We tackle the data mismatch by proposing: 1)
STG-Tokenizer: This spatial-temporal graph tokenizer transforms intricate graph
data into concise tokens capturing both spatial and temporal relationships; 2)
STG-Adapter: This minimalistic adapter, consisting of linear encoding and
decoding layers, bridges the gap between tokenized data and LLM comprehension.
By fine-tuning only a small set of parameters, it can effectively grasp the
semantics of tokens generated by STG-Tokenizer, while preserving the original
natural language understanding capabilities of LLMs. Extensive experiments on
diverse spatial-temporal benchmark datasets show that STG-LLM successfully
unlocks LLM potential for spatial-temporal forecasting. Remarkably, our
approach achieves competitive performance on par with dedicated SOTA methods. | [
0.018305527,
0.079513438,
-0.0072035994,
0.0236956887,
-0.0267221071,
0.0123608671,
0.027573688,
-0.0361938439,
0.0406085178,
-0.0337911211,
-0.0028216401,
-0.0289481226,
0.0234639365,
0.071711354,
-0.0016244596,
-0.0847642124,
0.0394668579,
-0.0486293621,
0.0138383023,
0.012851201,
-0.0101102805,
0.0086408779,
0.0227210559,
0.0572447404,
-0.0299602784,
-0.0097616594,
-0.0274651758,
-0.0220229067,
-0.0182590503,
-0.0808676779,
-0.017057905,
-0.006770473,
0.0532661676,
0.0305224713,
0.0000019784,
0.0160599276,
-0.0294312369,
0.0127740167,
0.0345911607,
0.0356723182,
0.053341452,
-0.0942866504,
0.0350346602,
0.0015403457,
-0.0596823394,
-0.0416160636,
0.0134380879,
0.0064126379,
0.0633189753,
0.0420447662,
-0.0260708462,
-0.0455224775,
0.0342635922,
-0.0155781228,
-0.0372408554,
-0.0103560537,
0.0099596577,
-0.0409659371,
0.0244334619,
0.0285853837,
-0.0602054223,
0.0281797461,
-0.0174617097,
0.0267090742,
0.0295405574,
0.0699889958,
-0.0165419336,
0.0421788432,
-0.0496990606,
0.0247875918,
0.0199836735,
-0.0472345985,
0.0342033058,
0.0298150089,
-0.026531985,
0.0174919311,
-0.0655867532,
-0.0216775984,
0.0208085719,
0.0630080998,
0.0129499231,
0.0119472658,
0.0067196726,
0.0141785666,
-0.0435924046,
0.1274743974,
-0.0121294167,
-0.018687658,
-0.0902978927,
-0.0039574234,
0.0054278551,
-0.0544043407,
0.0123888366,
0.0524443053,
-0.0532361008,
0.0106517551,
-0.0141394762,
0.006496185,
-0.0456556715,
-0.038689103,
0.0354823545,
0.038418442,
0.0021002248,
0.0669128299,
-0.0235518981,
0.0362180546,
-0.0043339808,
0.0254893247,
-0.0151274195,
-0.0463893451,
-0.0352150798,
-0.0270555113,
-0.0670145303,
0.0040056328,
-0.0579413995,
-0.0244512931,
0.0052569425,
0.0272179097,
-0.0091605112,
0.0026493825,
-0.0116019873,
0.0159375425,
-0.0170025676,
-0.0318147913,
-0.0292488262,
0.0174439754,
0.019951215,
-0.0090857167,
-0.028002413,
-0.002632817,
-0.0120408712,
0.0161921084,
0.0479632542,
0.0037434557,
-0.034240678,
0.0572291426,
0.0278406385,
0.0071584303,
-0.0270231832,
-0.0175917801,
-0.0356475003,
-0.0164034236,
0.0146854175,
-0.0491664745,
0.004904666,
0.034653753,
-0.0025353415,
-0.0074566337,
-0.0054477942,
-0.0189343859,
-0.0607103296,
0.0548890643,
-0.0095557701,
0.0287221838,
0.0074102823,
-0.0273593478,
0.0511208139,
0.0674627349,
0.038173072,
0.0704153627,
0.0213279333,
-0.0010234368,
-0.0317400992,
0.0263480898,
0.0180960931,
-0.0016270983,
-0.0955831632,
0.0097029712,
-0.0081074033,
0.0044929599,
0.028656235,
0.036439307,
-0.016459832,
-0.0011653093,
0.0300871097,
0.0662068427,
-0.0191099662,
0.0097483499,
0.0687771961,
0.0316295661,
-0.0310730748,
-0.0355612859,
-0.0319247246,
0.026403876,
0.0088596949,
0.0052064704,
0.0018267174,
0.0524754338,
-0.0086933384,
-0.018605398,
-0.0281944182,
-0.010163107,
0.0531448871,
-0.0332546122,
0.0493124165,
-0.0242572818,
-0.0306079835,
0.0936749876,
0.0171195883,
0.0206534229,
0.0146805774,
0.0229942594,
0.0170981083,
0.0625316873,
0.006634302,
0.0086729089,
-0.0227486081,
0.0027576988,
0.0735628158,
0.0097811446,
0.0519734956,
0.0158866905,
-0.0155173363,
-0.0351817608,
-0.0500932783,
0.0317969844,
0.0296572298,
-0.0104774684,
0.0341952927,
0.0458354466,
0.0186162908,
-0.0527885258,
-0.0219862759,
-0.000087267,
0.0281454623,
-0.0120381126,
-0.0189049169,
0.0034452977,
0.013142542,
-0.0023924096,
-0.000545396,
0.0317657739,
0.048856806,
-0.0512769856,
-0.0159723572,
0.0798184201,
-0.0588791072,
-0.0357900932,
-0.0054480117,
0.0155202793,
0.0120412158,
0.0038913963,
-0.0642168596,
-0.037431743,
0.0583255626,
-0.0383983031,
0.0853913203,
-0.0074701738,
0.0029505005,
0.0218529645,
0.0050351396,
-0.0246050414,
0.03667989,
-0.0066155442,
-0.004839798,
-0.0228019357,
0.0304321907,
0.0052899262,
0.0057864431,
0.0030391822,
-0.0082432786,
0.0014216102,
-0.0234670751,
-0.0206036735,
-0.0139395064,
-0.0284677558,
-0.0362054594,
-0.0946435332,
-0.0344550498,
-0.012580717,
0.0306638628,
0.0041201967,
-0.0394620448,
-0.0272214226,
-0.016897833,
-0.0073395171,
0.017821474,
-0.0063980622,
0.0129518565,
0.0288771559,
-0.0077081369,
-0.0091867894,
0.0037801235,
-0.0176401976,
-0.0437676609,
0.0678418204,
-0.0112320185,
-0.0698233619,
-0.0399058014,
-0.027899269,
-0.022444481,
0.0088525359,
-0.0145495813,
-0.0023469382,
0.0088983886,
-0.0573562458,
0.0109018739,
-0.007885674,
0.0153871411,
-0.0042567356,
0.0168044046,
-0.0786975846,
-0.0466556959,
0.0254534651,
-0.0326506607,
0.0251336582,
-0.0533287115,
-0.0020433217,
-0.0005721964,
0.0565261543,
0.0109170359,
0.0103330323,
-0.0401815549,
0.0109181218,
-0.0680704117,
0.018499095,
-0.0082040057,
0.0061588413,
-0.000325365,
0.0080912467,
-0.0340529457,
-0.0678849667,
0.0212266799,
-0.0371485129,
-0.0133321043,
0.0170447752,
-0.002015484,
-0.0937796384,
0.0780140981,
0.0338374712,
0.0069802743,
-0.0434005819,
-0.0441453941,
0.0278213043,
-0.0209928863,
-0.0353501141,
0.0071406798,
0.0086837821,
-0.0195821002,
0.0272959936,
0.0129309157,
-0.021010723,
-0.0196739696,
-0.0564676151,
-0.068137683,
-0.0152511029,
0.0013037731,
-0.0250525139,
-0.0099460557,
0.0668993518,
-0.0444758423,
0.0272714961,
-0.03967233,
0.002707093,
-0.0075857514,
0.0235382225,
0.0711716115,
0.0483239256,
-0.0421452671,
0.0167492069,
-0.0099453786,
0.0089412509,
-0.0090314792,
0.0003362281,
-0.0537749603,
0.0389923006,
0.0362096913,
0.005259084,
-0.0249357857,
0.0078730779,
0.0563087538,
-0.0762622207,
-0.0243312418,
0.1084637344,
0.0456509255,
-0.0079300273,
0.0107292635,
-0.0155055579,
0.0457693152,
0.0141263781,
0.0091551729,
-0.0447335877,
-0.0130703095,
-0.0140171545,
-0.0236122143,
0.0526108295,
-0.0565342866,
-0.0132586733,
-0.0633137897,
0.0655429214,
0.0345750079,
-0.0697686523,
-0.0223419946,
-0.0628270358,
-0.0370058529,
-0.0189514533,
-0.0083275037,
-0.0335776918,
-0.0211207252,
0.041162286,
0.0552289337,
-0.0948829949,
0.0081774211,
-0.016959792,
-0.0075511076,
0.0448724329,
0.0020230748,
0.05058752,
-0.0083957287,
-0.0215901248,
0.0015891831,
0.0620797165,
0.0421075039,
-0.0455811992,
0.0101455254,
0.0110317497,
-0.015865488,
0.0058270171,
0.0285294838,
-0.0189584177,
-0.0325311162,
0.0031746668,
0.1066039056,
-0.0280365013,
0.0091664279,
-0.0048399703,
-0.0002489567,
0.0075871283,
0.0378569514,
0.0165806953,
0.0100619486,
-0.0689371005,
-0.0408773683,
-0.0107931374,
-0.0707417205,
-0.0231242571,
0.0103960894,
-0.0049362802,
0.0137780383,
-0.0213587973,
-0.0213353224,
0.0153608583,
0.0107852453,
-0.0161483046,
0.0316692479,
0.0336098559,
-0.0497313589,
-0.0898125097,
0.0302191861,
-0.0494679287,
0.0242508873,
-0.0053799883,
0.0031462512,
0.0118685206,
0.0030337621,
-0.0157927237,
0.0283527896,
-0.1003764272,
-0.0462125316,
0.0138609465,
-0.0082132313,
0.0090628713,
0.0371020213,
0.0080411406,
0.0046887309,
0.0962992013,
0.0649091899,
-0.0245756619,
0.0095699066,
0.0259630922,
0.0398454294,
0.0222989861,
0.0225316081,
0.0003952016,
0.0025118967,
-0.0122318221,
0.0213634688,
-0.0073554786,
0.0248874445,
-0.0113019682,
-0.0055932407,
0.0966572389,
0.0134904468,
-0.0089547615,
0.0082092248,
-0.0205201376,
-0.0273785703,
0.0102131264,
-0.0350303464,
0.0070043062,
0.0045325677,
0.0175658613,
-0.0457882807,
0.007367007,
0.0266640652,
-0.0077586235,
-0.0253138412,
0.0141638601,
-0.0176319145,
-0.045056995,
0.0396666825,
0.0364225581,
0.0082961572,
-0.030628901,
-0.0202585831,
-0.0109576182,
-0.0234140512,
-0.0146944337,
0.011642714,
0.0596444309,
-0.0526321158,
-0.0015442406,
-0.0638216808,
-0.0477679521,
0.0003849712,
0.0193509124,
0.0319977812,
-0.029442599,
0.0213099923,
-0.0232558884,
-0.0039020579,
0.0103773549,
0.036501728,
0.0164190605,
-0.0557049178,
0.0532328412,
-0.0255994648,
-0.0643652752,
-0.002264309,
0.0617268533,
0.0442519598,
0.0241554789,
0.0480955206,
-0.0413768291,
-0.0284100324,
-0.0078521874,
0.0049007162,
0.0001526158,
-0.0132379942,
-0.0145469811,
-0.0481702164,
-0.0463411175,
0.0551617034,
-0.0386208929,
-0.0375911035,
-0.016938651,
-0.0159646776,
0.0044765202,
-0.0066733938,
0.0023902101,
-0.0342602059,
0.0358332247,
0.0248669926,
-0.0599558763,
-0.0269152857,
-6.22871167e-33,
-0.02020864,
-0.0636801794,
0.0561458245,
-0.0137638543,
-0.0278986488,
0.0537861101,
-0.0296829026,
-0.0274845473,
0.0300036818,
-0.0011232828,
-0.0140638584,
0.0086281067,
0.0215297118,
0.0422860198,
0.0591923781,
-0.0048434981,
0.0428808704,
-0.023228148,
0.0114410715,
0.0113974661,
0.0531299897,
0.0458388813,
0.0377388708,
-0.0274076089,
-0.0061561689,
0.0391533375,
-0.0021995299,
-0.0036673562,
0.0526213422,
0.0466027744,
-0.0147837596,
0.0226392895,
0.0431726053,
-0.014325642,
-0.060747657,
0.0328075662,
-0.0750022605,
-0.0454333685,
0.0052284659,
0.0164605342,
-0.0125282574,
-0.0018393634,
0.0013867602,
-0.0010952873,
0.0112556051,
-0.0022638913,
0.0089059081,
-0.0115306564,
-0.0764986947,
0.0032285713,
-0.0140293352,
0.0416017212,
-0.0100564715,
0.0274827629,
0.0605696514,
0.0128539335,
0.03640154,
0.0545821525,
-0.0611869134,
0.0369203165,
-0.0099645285,
0.025846066,
0.0255862866,
0.023529401,
0.0461117923,
0.0653094649,
0.0280552283,
-0.0398371294,
-0.0640985742,
0.0560770519,
0.0450862981,
0.0117305545,
-0.0247340798,
0.0017732683,
0.0321461223,
-0.0287998356,
-0.0342086628,
-0.0004417503,
-0.0331582204,
0.0303635839,
0.0210839082,
0.0284296349,
-0.0147740236,
-0.0297309309,
-0.0092408452,
-0.0130939269,
0.0004241572,
0.0100850221,
0.0220025238,
0.0355695449,
0.0196582545,
0.0740733221,
0.0320619084,
-0.0405569673,
-0.0620924383,
0.0354743227,
0.0101841008,
-0.0153325312,
-0.0096767619,
0.0499427356,
-0.0660986826,
-0.0502897538,
-0.0132099176,
0.0012655519,
-0.006126334,
-0.015964523,
0.035567686,
0.0523433387,
-0.0719612613,
-0.0091910921,
-0.0167949423,
-0.0244478639,
-0.0174082685,
-0.0275724456,
0.0679707155,
-0.0135428086,
-0.0007430238,
0.0486376435,
0.0216108132,
0.0369454846,
0.0206256658,
0.0021371308,
-0.0184642449,
0.0101905325,
-0.0049606184,
-0.0141962729,
-0.0428505205,
-0.0175775588,
0.0201920215,
-0.065266192,
-0.0000566991,
0.0028474601,
2.656e-7,
0.0389252678,
0.0736497715,
0.0223086383,
0.0212467518,
0.0400932506,
-0.0075305961,
-0.006027983,
0.0198775548,
-0.0375988819,
0.0587799735,
0.0485626161,
0.0286692902,
0.0291093346,
0.0293920748,
-0.0849763379,
0.0490765572,
-0.0926781446,
-0.019022882,
-0.0773625821,
-0.0200603753,
0.0819651261,
0.0915539488,
0.0343507938,
0.0136301303,
0.0020126875,
-0.0589519329,
0.0342866741,
-0.021947287,
-0.0176590066,
0.000467763,
0.0207773354,
-0.0138843926,
-0.02176941,
-0.0313956887,
-0.0126801841,
0.0067979563,
-0.0031050586,
0.0399365015,
-0.0100178374,
-0.0080764657,
0.0188107658,
-0.0373630002,
-0.0597026944,
-0.0678914785,
0.0604499653,
-0.0801649988,
-0.0204319004,
-0.0219607558,
-0.0175879356,
0.0505450107,
-0.036684189,
0.0232809316,
-0.0045840479,
0.0070317951,
0.0397137105,
-0.0315671936,
0.0098630879,
-0.0566035286,
-0.0396492183,
-0.0038555556,
-0.0143890604,
-0.006582486,
-0.0522613227,
0.0553787164,
0.0474070385,
-0.0196950715,
-0.0076111397,
2.526833473e-34,
0.0480569117,
0.0171995349,
-0.0304529537,
0.003923961,
0.0329426527,
-0.0263090525,
0.0093879495,
0.0495007969,
0.0087053068,
-0.0726410002,
-0.0151623217
] |
http://arxiv.org/abs/2401.14166v3 | http://arxiv.org/pdf/2401.14166v3 | BayesPrompt: Prompting Large-Scale Pre-Trained Language Models on Few-shot Inference via Debiased Domain Abstraction | 2024-01-25T13:20:47 | 2024-03-20T08:52:42 | Jiangmeng Li, Fei Song, Yifan Jin, Wenwen Qiang, Changwen Zheng, Fuchun Sun, Hui Xiong | cs.CL | cs.CL, cs.AI | As a novel and effective fine-tuning paradigm based on large-scale
pre-trained language models (PLMs), prompt-tuning aims to reduce the gap
between downstream tasks and pre-training objectives. While prompt-tuning has
yielded continuous advancements in various tasks, such an approach still
remains a persistent defect: prompt-tuning methods fail to generalize to
specific few-shot patterns. From the perspective of distribution analyses, we
disclose that the intrinsic issues behind the phenomenon are the
over-multitudinous conceptual knowledge contained in PLMs and the abridged
knowledge for target downstream domains, which jointly result in that PLMs
mis-locate the knowledge distributions corresponding to the target domains in
the universal knowledge embedding space. To this end, we intuitively explore to
approximate the unabridged target domains of downstream tasks in a debiased
manner, and then abstract such domains to generate discriminative prompts,
thereby providing the de-ambiguous guidance for PLMs. Guided by such an
intuition, we propose a simple yet effective approach, namely BayesPrompt, to
learn prompts that contain the domain discriminative information against the
interference from domain-irrelevant knowledge. BayesPrompt primitively
leverages known distributions to approximate the debiased factual distributions
of target domains and further uniformly samples certain representative features
from the approximated distributions to generate the ultimate prompts for PLMs.
We provide theoretical insights with the connection to domain adaptation.
Empirically, our method achieves state-of-the-art performance on benchmarks. | BayesPrompt: Prompting Large-Scale Pre-Trained Language Models on Few-shot Inference via Debiased Domain Abstraction: As a novel and effective fine-tuning paradigm based on large-scale
pre-trained language models (PLMs), prompt-tuning aims to reduce the gap
between downstream tasks and pre-training objectives. While prompt-tuning has
yielded continuous advancements in various tasks, such an approach still
remains a persistent defect: prompt-tuning methods fail to generalize to
specific few-shot patterns. From the perspective of distribution analyses, we
disclose that the intrinsic issues behind the phenomenon are the
over-multitudinous conceptual knowledge contained in PLMs and the abridged
knowledge for target downstream domains, which jointly result in that PLMs
mis-locate the knowledge distributions corresponding to the target domains in
the universal knowledge embedding space. To this end, we intuitively explore to
approximate the unabridged target domains of downstream tasks in a debiased
manner, and then abstract such domains to generate discriminative prompts,
thereby providing the de-ambiguous guidance for PLMs. Guided by such an
intuition, we propose a simple yet effective approach, namely BayesPrompt, to
learn prompts that contain the domain discriminative information against the
interference from domain-irrelevant knowledge. BayesPrompt primitively
leverages known distributions to approximate the debiased factual distributions
of target domains and further uniformly samples certain representative features
from the approximated distributions to generate the ultimate prompts for PLMs.
We provide theoretical insights with the connection to domain adaptation.
Empirically, our method achieves state-of-the-art performance on benchmarks. | [
0.033311788,
0.0298358761,
-0.001713705,
0.0229331106,
-0.0121269403,
0.03030717,
0.0422633179,
-0.0178026427,
-0.0222672764,
-0.0823524371,
-0.0765685067,
-0.0344238207,
-0.0260237101,
0.0071488051,
0.0156580191,
-0.0506796837,
0.0358896516,
-0.030767763,
0.0113368584,
-0.0147365537,
-0.012770907,
0.0150861433,
-0.0290382411,
0.0193180945,
0.0250646323,
-0.0206595268,
0.0088376543,
0.0092704305,
-0.0091726631,
0.0037344056,
-0.0008393084,
0.069225207,
0.0428186245,
0.0163769908,
0.0000021663,
-0.0749215707,
-0.0072695259,
-0.0204688981,
-0.0473926775,
0.0024476566,
0.0266486648,
-0.0317158438,
0.0214522984,
-0.0319416411,
0.0004648859,
-0.0942576677,
0.0403427184,
0.0324082263,
0.0131365592,
0.0804251581,
-0.0022158935,
-0.0264446903,
-0.0431895256,
-0.0529656485,
-0.0067690467,
-0.0859201252,
0.0494950935,
-0.029582439,
0.0073894146,
0.0337905884,
-0.0457474664,
0.0194635596,
-0.002558534,
0.039630942,
-0.0162419658,
0.0276739281,
0.0155329285,
-0.0172868688,
-0.0110653173,
0.0683478117,
-0.011996557,
-0.0251602512,
-0.0065960358,
-0.0484683104,
0.0025041574,
0.0363976061,
-0.0248788446,
0.0173621308,
0.0158969853,
-0.0124561843,
-0.0182834379,
0.0275434721,
0.0369689763,
0.0287946928,
-0.0439670235,
0.0245417152,
0.0235113855,
-0.003416595,
0.0189737249,
0.0341977179,
-0.0421756133,
-0.0520267896,
-0.0194721762,
0.0040343837,
-0.0380632952,
0.018874526,
-0.067797929,
-0.0147501696,
-0.0176679008,
0.037586581,
0.0510724224,
0.0323220417,
-0.046850618,
0.0714886859,
0.0092654191,
0.0087307375,
-0.0151786581,
0.0032393378,
-0.0381760411,
-0.0014368973,
-0.01589044,
0.019061517,
-0.0399704985,
0.0001220596,
-0.0190788582,
-0.0072342088,
-0.0067939907,
0.0230264142,
0.00130713,
-0.0240591187,
0.0135795465,
-0.0089869862,
-0.0473958477,
-0.0531989075,
-0.0041186903,
0.054452423,
-0.085857369,
0.0443617441,
0.0213168617,
-0.0269453041,
-0.0398046151,
0.0168238021,
0.0099193482,
-0.0472040586,
0.0076023005,
0.0628678203,
0.0135873109,
-0.0491332263,
-0.0534031019,
-0.0145862103,
-0.0252141431,
-0.0434361435,
0.0063043768,
-0.0285657439,
0.0322372951,
-0.0004204354,
0.0057195351,
-0.059186209,
-0.0262415409,
-0.0332474113,
-0.046493277,
0.0936098695,
0.004887328,
0.0125654787,
0.0239548851,
0.0136757707,
-0.0356273428,
0.094461754,
-0.0189808309,
0.0054402137,
0.0269141532,
0.0199818648,
0.0059224763,
0.0369332209,
0.0274813175,
0.026947422,
-0.0035575069,
-0.0048470618,
0.0103341946,
0.020132849,
0.0626623631,
-0.0066417069,
-0.0349962972,
-0.0264073536,
0.0442462452,
0.004678939,
-0.0137708643,
-0.0259470865,
0.1023519635,
0.0035018958,
-0.0131505504,
0.0179679282,
-0.0822606012,
-0.00715004,
-0.0185959656,
0.0060296799,
0.0228577293,
0.0210172832,
-0.0495176874,
-0.0027927102,
0.0293787848,
-0.0109358374,
-0.0570981875,
0.0472341739,
0.0813944861,
0.0403334834,
-0.0126994289,
0.0540849306,
0.0546126962,
0.0075810784,
0.0348756425,
-0.0363360606,
0.0297655407,
0.0340587832,
0.0169270746,
-0.0338362493,
-0.0375700407,
-0.0015747162,
0.0653008074,
0.0421702191,
0.0437600948,
-0.0214076359,
0.046041593,
-0.033625152,
-0.0638804585,
-0.0160085298,
0.0339822546,
0.0284898002,
0.0251271911,
0.0001166767,
0.0365086347,
-0.0562537163,
0.0389383808,
0.0362381488,
-0.0390970074,
-0.0435326844,
0.0117030507,
0.0409725346,
-0.0010686344,
0.0074519035,
-0.0175965633,
-0.0020288692,
-0.0093604531,
-0.0068496643,
0.0269752666,
-0.0276303608,
-0.0404260755,
-0.0462429151,
0.0082335062,
0.0016284654,
-0.017724473,
-0.0244190022,
-0.0568217114,
-0.0249612238,
0.0117916428,
0.0483972393,
-0.0004757475,
-0.0310303587,
0.0229448322,
-0.0090491865,
-0.0057964139,
-0.039853435,
0.0501609296,
0.0273674466,
-0.0072299303,
0.0253387187,
0.0207366385,
-0.0060429284,
-0.0157279875,
-0.0219702423,
-0.0345533602,
0.0110703753,
-0.0054750852,
0.0283544865,
0.0052752472,
-0.0506423824,
0.0235541202,
-0.060441263,
0.0071590971,
-0.0241574626,
0.016152408,
0.0119614415,
0.0458957516,
-0.0438684821,
-0.0391990207,
0.0088581005,
-0.0202860087,
-0.033191327,
0.0403999574,
0.0074717598,
0.0141844768,
0.026650304,
-0.0076098172,
0.0037341118,
-0.0022042631,
0.047932107,
-0.0035356327,
-0.0072787991,
-0.0508758314,
0.0738107488,
-0.0055595851,
0.0454585999,
-0.0082289483,
0.0135294115,
-0.0205172859,
-0.024339661,
0.0381381102,
0.0501236506,
-0.0304356478,
-0.0068482077,
0.050621409,
-0.0166300926,
-0.0339151025,
-0.0029081227,
-0.0509353019,
-0.0446109436,
-0.0216411352,
0.0350738876,
0.0239804536,
0.0774278343,
-0.0040249,
-0.0253414512,
-0.0818806738,
0.0382374674,
-0.0149487955,
-0.0076246788,
-0.0369244479,
-0.0051159221,
0.0044358759,
0.0325660035,
-0.0144224856,
-0.0175961889,
0.0898520723,
-0.0374579094,
0.0173138194,
-0.0253887679,
0.0183313116,
-0.0201734286,
0.0490944535,
-0.0035422575,
0.0014514908,
0.0822355598,
-0.019390557,
-0.1060577706,
0.0115872025,
0.0198692307,
0.0115897097,
0.0556780286,
-0.0474155471,
0.0152845066,
-0.0003474963,
-0.0465736501,
-0.0483219102,
-0.0815527588,
-0.034841083,
-0.0103605865,
-0.0474669896,
0.0079725031,
-0.0362782218,
0.0082513327,
-0.0047036475,
0.045648396,
-0.0064732996,
-0.0108462302,
-0.0180779975,
0.0328829065,
0.0608922243,
-0.0051460713,
0.0033373248,
-0.0143264448,
0.0076933308,
0.0052709933,
0.0009602257,
-0.0038295269,
-0.0089191943,
0.00776482,
0.0374702774,
-0.0286948085,
-0.0217459965,
0.0215394534,
0.070525825,
-0.0134200212,
-0.0216306858,
0.1295644492,
-0.0033072361,
-0.0658819377,
-0.0117389197,
0.033514224,
0.0085173761,
0.0112135503,
0.0336332023,
-0.1030348912,
-0.0652493462,
0.0020321859,
0.0007482692,
0.0128118955,
-0.0195628535,
-0.0126731517,
-0.13622877,
0.013390216,
-0.0000348186,
-0.0248778258,
-0.006354243,
-0.0057106311,
0.034688063,
-0.0049123559,
0.0001437949,
0.005705426,
-0.0000044776,
0.0162815079,
0.0203097053,
0.0471606255,
-0.0948075131,
0.0318734758,
-0.0661734641,
0.0543115623,
0.0561030582,
-0.0378941111,
0.0236440953,
-0.0044856868,
0.0095372796,
-0.0404090919,
-0.0087978039,
-0.0042334371,
0.0362170152,
-0.0566684306,
-0.0106561659,
0.0301500838,
0.0356898196,
-0.0260649361,
-0.0262060761,
0.0195885934,
0.0564827062,
-0.0203131698,
-0.0120297447,
-0.0164215565,
-0.0009363358,
0.0203431118,
0.0869138092,
0.0456889756,
-0.0221866351,
0.0168610029,
-0.0018604901,
-0.0309682079,
-0.0133887278,
-0.0218467806,
0.0168097485,
-0.0322253741,
-0.0489860326,
-0.0540277176,
-0.0251072999,
0.0519300997,
0.0372092724,
0.0149240764,
-0.0135835614,
-0.0120677957,
0.0209947396,
-0.0340908244,
0.0582939796,
-0.0672050714,
0.0087944018,
-0.040447589,
-0.0119071268,
0.0470517427,
-0.051727891,
-0.0222308412,
0.0202840939,
-0.0626209676,
-0.0583589524,
-0.0217074323,
0.0510775,
0.0171311107,
0.0485327132,
-0.0548159257,
-0.0614454299,
0.0729180872,
0.0359649658,
-0.0301321093,
0.0202417038,
0.0115963789,
-0.0168112908,
-0.010082745,
0.0449786186,
-0.0613849573,
0.0536388755,
0.0611866564,
0.0229369141,
0.0163061991,
-0.0517880358,
0.0016712551,
-0.0201712865,
0.1126122177,
-0.0418856777,
-0.0057135923,
-0.0437404551,
-0.010598137,
-0.0214226376,
-0.0389152206,
-0.0325218849,
-0.0192060582,
0.0273953471,
-0.0062900814,
-0.0099083502,
-0.0046405462,
-0.0255869944,
0.0545135625,
-0.0371866077,
-0.0152926156,
0.0389140323,
0.0092540039,
0.0350848474,
-0.0116144726,
-0.0048291762,
0.0299613997,
0.0087741632,
0.015464888,
-0.041873727,
0.0732770339,
0.0349894837,
-0.006963498,
-0.0409175232,
0.0125358161,
-0.0572603568,
-0.0189342927,
-0.0095470175,
0.0419212282,
0.0522881001,
-0.0428642742,
0.0034327165,
0.0075512091,
-0.0125284335,
-0.0188782532,
0.0236352198,
0.0315850712,
0.0032861789,
0.0605470091,
0.0249419212,
-0.0175085496,
0.0098603787,
0.0844780877,
0.0023870755,
0.0603024289,
0.0118153533,
0.0175261702,
-0.0086630173,
-0.048125267,
-0.0005693187,
0.0271424782,
-0.0452310741,
0.0268768538,
-0.0785623267,
-0.062857464,
-0.0223959628,
-0.0157154333,
-0.0550614037,
0.0162407737,
0.0556514151,
0.0601970479,
-0.0180707984,
-0.0142832939,
-0.0216251425,
0.0050605489,
-0.0178837031,
-0.0161271077,
-0.0034714248,
-6.82257362e-33,
-0.0318111591,
0.0178266168,
0.0260539334,
-0.0265579317,
-0.0296709649,
-0.0187879335,
-0.0248434264,
-0.0130951013,
-0.0283751115,
0.013776009,
-0.0141966213,
0.0078139547,
0.018905513,
0.0275209732,
-0.009299323,
-0.0288571175,
0.0604896396,
0.0034272082,
-0.0389366001,
0.0056332359,
0.0037930005,
0.0427146815,
0.045369409,
0.0040397239,
0.0187019203,
-0.009125188,
-0.0052510207,
0.0438671932,
0.0302178897,
0.0377098098,
-0.0062816022,
0.0390743054,
0.0062310649,
-0.0420927629,
-0.0156005602,
0.0245814491,
-0.0883611143,
-0.0491928533,
0.030805869,
0.0108573101,
0.0029167926,
-0.023474507,
0.0041050292,
0.0000204022,
0.0347216241,
-0.0278538279,
0.0030556249,
-0.0351646654,
-0.0035590886,
-0.0399929769,
-0.0439667925,
0.0263869967,
0.0261806902,
-0.0401073247,
0.0344086476,
0.0131816277,
0.0198024642,
0.0246791523,
-0.0155291948,
0.0236368924,
0.0359060131,
0.0244197212,
0.0528700463,
0.0480457991,
-0.0027655824,
0.0411012359,
0.0046367534,
-0.0112847211,
-0.0013765509,
-0.0098760696,
0.0363374166,
0.0088889413,
-0.0388194025,
0.0727185905,
0.0278705694,
0.0149581665,
-0.0712053105,
0.0216745306,
-0.0088839764,
-0.0002954566,
0.001951742,
-0.002157846,
-0.017238088,
-0.0391202271,
-0.0540237091,
-0.0158355329,
-0.0347991884,
-0.009024811,
0.024822019,
0.009003371,
-0.0107950727,
0.0064958795,
0.0444196798,
-0.0470667966,
-0.0017183307,
0.0032631194,
-0.0038218172,
-0.0044353236,
-0.0137214186,
0.0365787111,
-0.0316424258,
-0.0327797905,
0.0520483181,
-0.0057514608,
-0.0234558713,
0.0230102222,
0.045888409,
0.0009860802,
-0.0507758334,
-0.0188602563,
-0.0164664835,
-0.0051866667,
0.0054559745,
0.0590217076,
0.0247586872,
0.0629460812,
-0.0140923057,
0.0492541306,
0.0319945924,
-0.0728929043,
0.0293591917,
0.0613285638,
0.0117220432,
0.0338980965,
-0.0438534617,
-0.0140371146,
0.004699178,
-0.0136654256,
0.0370975658,
-0.0623601004,
-0.0346379504,
-0.003785315,
2.85e-7,
0.0498998128,
0.0004536335,
-0.0121096373,
0.0238603633,
-0.0001937858,
0.0454597883,
0.013954455,
0.0201260243,
0.0240670033,
-0.0333794355,
0.0471861362,
0.0374611691,
0.0386340544,
-0.0014297921,
-0.1113842502,
0.0111996895,
0.0265876725,
0.011874754,
-0.0385257639,
-0.0157411117,
0.0399303287,
0.1019742712,
0.03320219,
-0.0087942276,
0.018749509,
-0.0361516625,
0.0692017153,
-0.0310399737,
0.0016482185,
0.0263078827,
-0.0178490616,
-0.0032952894,
0.0031868694,
0.0761063248,
-0.0118790586,
0.0405699722,
-0.0072489842,
0.0839239806,
0.017083304,
0.0173728969,
-0.0315139294,
-0.0063555487,
0.0092349192,
-0.029482346,
0.0965431035,
-0.117070891,
-0.022144625,
-0.0259645376,
-0.0120038409,
-0.0256004985,
-0.0033653821,
-0.024323862,
-0.0207462516,
0.0074658669,
0.035039071,
0.024350876,
-0.0267393738,
-0.1016276851,
0.0124140698,
0.0117304344,
-0.0331851654,
-0.0318772756,
0.0019959796,
-0.0572092272,
0.0551246814,
-0.078265518,
0.0206448771,
2.60860288e-34,
0.0233987346,
0.0010364009,
0.0027654537,
0.0025922225,
0.00024921,
-0.0134017412,
-0.0318506733,
-0.0114822444,
0.0402766019,
-0.076763846,
-0.0004314253
] |
http://arxiv.org/abs/2401.14151v2 | http://arxiv.org/pdf/2401.14151v2 | True Knowledge Comes from Practice: Aligning LLMs with Embodied Environments via Reinforcement Learning | 2024-01-25T13:03:20 | 2024-03-11T03:15:58 | Weihao Tan, Wentao Zhang, Shanqi Liu, Longtao Zheng, Xinrun Wang, Bo An | cs.LG | cs.LG, cs.AI, cs.CL | Despite the impressive performance across numerous tasks, large language
models (LLMs) often fail in solving simple decision-making tasks due to the
misalignment of the knowledge in LLMs with environments. On the contrary,
reinforcement learning (RL) agents learn policies from scratch, which makes
them always align with environments but difficult to incorporate prior
knowledge for efficient explorations. To narrow the gap, we propose TWOSOME, a
novel general online framework that deploys LLMs as decision-making agents to
efficiently interact and align with embodied environments via RL without
requiring any prepared datasets or prior knowledge of the environments.
Firstly, we query the joint probabilities of each valid action with LLMs to
form behavior policies. Then, to enhance the stability and robustness of the
policies, we propose two normalization methods and summarize four prompt design
principles. Finally, we design a novel parameter-efficient training
architecture where the actor and critic share one frozen LLM equipped with
low-rank adapters (LoRA) updated by PPO. We conduct extensive experiments to
evaluate TWOSOME. i) TWOSOME exhibits significantly better sample efficiency
and performance compared to the conventional RL method, PPO, and prompt tuning
method, SayCan, in both classical decision-making environment, Overcooked, and
simulated household environment, VirtualHome. ii) Benefiting from LLMs'
open-vocabulary feature, TWOSOME shows superior generalization ability to
unseen tasks. iii) Under our framework, there is no significant loss of the
LLMs' original ability during online PPO finetuning. | True Knowledge Comes from Practice: Aligning LLMs with Embodied Environments via Reinforcement Learning: Despite the impressive performance across numerous tasks, large language
models (LLMs) often fail in solving simple decision-making tasks due to the
misalignment of the knowledge in LLMs with environments. On the contrary,
reinforcement learning (RL) agents learn policies from scratch, which makes
them always align with environments but difficult to incorporate prior
knowledge for efficient explorations. To narrow the gap, we propose TWOSOME, a
novel general online framework that deploys LLMs as decision-making agents to
efficiently interact and align with embodied environments via RL without
requiring any prepared datasets or prior knowledge of the environments.
Firstly, we query the joint probabilities of each valid action with LLMs to
form behavior policies. Then, to enhance the stability and robustness of the
policies, we propose two normalization methods and summarize four prompt design
principles. Finally, we design a novel parameter-efficient training
architecture where the actor and critic share one frozen LLM equipped with
low-rank adapters (LoRA) updated by PPO. We conduct extensive experiments to
evaluate TWOSOME. i) TWOSOME exhibits significantly better sample efficiency
and performance compared to the conventional RL method, PPO, and prompt tuning
method, SayCan, in both classical decision-making environment, Overcooked, and
simulated household environment, VirtualHome. ii) Benefiting from LLMs'
open-vocabulary feature, TWOSOME shows superior generalization ability to
unseen tasks. iii) Under our framework, there is no significant loss of the
LLMs' original ability during online PPO finetuning. | [
0.0273102503,
0.0273741074,
-0.0194134787,
-0.016615482,
-0.0126295751,
0.0287353266,
-0.0093047563,
-0.0000305193,
0.0064054909,
-0.0051750145,
-0.0716592371,
-0.0377860889,
-0.0052889297,
0.0441271029,
0.0178391598,
-0.0502050743,
0.0339534692,
-0.0377630368,
0.0216532685,
-0.0088488609,
0.006392586,
0.0004026551,
-0.0400703587,
0.0393451601,
0.0242206678,
0.0145895733,
-0.0572223589,
-0.0018945819,
0.0492936447,
-0.0507083647,
0.0074251508,
0.0337797441,
0.0740010738,
0.0296354871,
0.0000023394,
-0.0220391508,
-0.0011264132,
-0.0365712494,
0.0178183895,
-0.037139345,
0.0178796463,
-0.0202723201,
-0.010559544,
0.013687985,
-0.0467523076,
-0.0159448329,
0.0574918836,
0.0401285924,
0.0028827866,
0.0802095756,
0.0085929791,
0.0014288507,
-0.0018980777,
0.0007251,
0.0001742691,
-0.0253362264,
0.016140677,
-0.0267549418,
0.0464045145,
-0.0459822714,
-0.0506794527,
0.0033638435,
0.0148264896,
-0.0149356388,
-0.0371704549,
0.0663096681,
0.0127849998,
-0.0474365763,
-0.0277341641,
0.0794250816,
-0.0317756794,
0.0188398622,
0.0431896485,
-0.0355120115,
0.0124863693,
-0.0151989209,
-0.0119358944,
0.0259920098,
0.0168963913,
0.0503548868,
-0.0431502275,
-0.0248775445,
0.03456055,
0.0359964482,
-0.006273028,
0.0987972915,
-0.0209340397,
-0.0383474939,
0.0299943145,
-0.0053267791,
-0.0357787721,
-0.0286588278,
-0.0161727462,
0.0512875393,
-0.0886690766,
0.0357415974,
-0.0285711177,
0.0006749253,
-0.0223679282,
-0.0018383648,
0.0688893646,
0.0496563017,
0.0215755235,
-0.001072136,
-0.0211845376,
0.0733326748,
0.0081861997,
0.058658313,
-0.0532495417,
0.0317326188,
-0.0524999425,
0.0131642241,
-0.0264950003,
0.0293869283,
-0.0420326851,
-0.0403504483,
-0.0104161054,
0.0323391743,
-0.0357171483,
0.0378447063,
-0.0144432513,
0.0309637301,
-0.020238243,
-0.0132414857,
-0.0870267525,
0.0079353191,
-0.0230642948,
0.0458740033,
0.0235555843,
-0.0176089555,
0.0038084465,
0.0058406717,
0.0147348782,
-0.0151792653,
-0.0109095024,
0.0733179897,
-0.0225351658,
-0.0392675139,
0.0428707153,
0.0159612987,
-0.0092329578,
-0.0656080246,
0.0009062398,
-0.0431997851,
0.0349855274,
-0.0488064736,
-0.0093892664,
-0.0349589437,
0.0234896764,
-0.0089566913,
-0.0159929972,
0.0701820701,
0.0307711978,
0.0167945959,
0.0297323819,
0.0005526714,
0.0052722339,
0.0686278343,
0.0509878993,
-0.0245543886,
-0.0340521485,
-0.0094549898,
-0.0239983536,
0.0272516832,
-0.0027034984,
0.0168651175,
-0.0209690854,
-0.0099303462,
0.0343226492,
0.0873210803,
0.0224101692,
0.0305611677,
-0.0268278345,
0.0393329263,
0.0922043324,
0.0270892791,
-0.0326161906,
-0.0187361427,
0.0052548298,
0.0341734216,
-0.028572306,
-0.0176144857,
-0.0505320579,
0.0738704801,
0.0068199583,
0.0112157436,
0.0151255336,
-0.0042740204,
-0.0256272461,
-0.0385275446,
0.0028851316,
-0.0104274824,
0.0359014831,
0.0302394833,
0.0268102139,
-0.0006421693,
-0.0668021962,
-0.0259922389,
0.0195052214,
-0.0031832885,
-0.0054313457,
-0.0325699486,
0.0789166242,
0.0764022246,
0.011019716,
-0.020516688,
-0.0274471231,
-0.0429485068,
0.012376315,
0.0693280697,
0.0050456407,
0.0059362701,
0.0249309577,
-0.0124616204,
-0.0757818297,
0.0016519397,
-0.0078315213,
0.0409364551,
0.0209445804,
0.0103556281,
-0.0347068831,
0.0422719307,
-0.0024706479,
0.0398032926,
-0.0428193547,
-0.0146041075,
-0.0001866673,
-0.0217954032,
0.0207339339,
-0.0421688259,
-0.012104393,
-0.0181100313,
0.0141118076,
-0.0328328609,
-0.0220443513,
0.0147563992,
-0.0583695322,
-0.0095658153,
0.0286493711,
-0.0104402052,
0.0103481356,
0.0147605501,
-0.0388353467,
-0.03922024,
-0.0154598001,
0.0260760486,
0.0209816508,
-0.0112069268,
-0.0130545506,
-0.0059897578,
-0.0080088209,
-0.0266451575,
0.0294027869,
-0.0325984918,
0.0294012129,
0.0170264989,
0.0156621523,
-0.026183879,
-0.0814638436,
-0.0092004929,
-0.0216790196,
0.0164055582,
0.032337416,
0.0219431389,
0.0471498854,
-0.0453237146,
-0.0409621894,
-0.0557509065,
-0.0085009784,
-0.0223945621,
-0.0158709697,
0.0186726265,
-0.039618209,
-0.0349179655,
-0.0223587099,
-0.0134747084,
0.0476904735,
-0.0057483157,
0.0003684431,
0.0548361577,
0.0322188027,
-0.0182022415,
-0.0495451689,
0.020857444,
0.021961214,
-0.0234993473,
-0.0296964049,
0.0044746352,
-0.0483764671,
0.0275492053,
-0.0510356054,
0.0282860287,
-0.025189871,
-0.0469095707,
0.0247399248,
-0.0216550697,
0.0262071379,
0.0833309218,
0.0168650635,
0.0168759096,
0.1029215008,
0.0202550683,
-0.0752737671,
0.0426207744,
-0.0314002782,
-0.0360030718,
-0.0618358701,
0.0217596628,
-0.0049722739,
0.0345318094,
-0.0054790084,
0.0287079066,
-0.0582892522,
0.0308254957,
-0.001911353,
-0.0610134453,
-0.0512931161,
0.045133736,
-0.0831504092,
0.0294047706,
-0.0485994071,
-0.0392983966,
0.0640582889,
0.0243718345,
-0.0014384513,
-0.0182167031,
-0.0189719256,
0.0254742876,
0.0247681309,
0.0258449782,
0.0275015235,
-0.0265010241,
0.0001954523,
-0.0857554451,
-0.0650870726,
0.031312149,
-0.0137697728,
0.0389165729,
-0.0465130322,
0.0062928181,
-0.0069626262,
-0.0629118308,
-0.0505712703,
0.0082716895,
-0.0306432862,
0.0178250074,
-0.0510817617,
-0.0192510299,
-0.0387089178,
-0.0057174587,
0.0493496507,
0.0449533463,
-0.0051020337,
-0.0098821484,
-0.034583021,
-0.0135637866,
0.096057944,
0.0990952849,
-0.0265670307,
-0.021884989,
0.0676425248,
0.0159282796,
0.0136613017,
-0.028941825,
-0.0307315253,
0.0320110843,
0.0045345058,
-0.0262762122,
0.0216766577,
0.0608228259,
0.0341657922,
-0.0364590436,
-0.0353422575,
0.1038801745,
0.0062934188,
-0.0075419727,
0.0007433203,
-0.0254946463,
0.0545226969,
0.0284773875,
0.0338608474,
-0.0394304842,
-0.0095003368,
-0.0040802569,
0.0245899614,
-0.0076818708,
-0.0318516381,
0.0047493801,
-0.1277047098,
0.0239273421,
0.0151260281,
-0.0669052899,
-0.027129421,
-0.0142826214,
0.0208648406,
-0.0097733662,
0.0562891178,
0.0390016548,
-0.0046846857,
-0.0054759388,
-0.0059779412,
0.0211281981,
-0.0201721154,
0.0225658715,
-0.0467675813,
0.0748032331,
0.074808307,
0.0396288969,
-0.0266167875,
-0.0610500835,
-0.0107380971,
0.0133965556,
0.0171990953,
-0.0024405355,
-0.0296500865,
0.0003446192,
-0.0170023907,
-0.0107010258,
0.0643653274,
0.0068599326,
-0.0298250709,
-0.0011382272,
0.0931359604,
-0.0277702417,
-0.0140146408,
0.0181232002,
-0.0056904778,
0.0455208942,
0.0530823097,
-0.0183979589,
-0.0597786568,
-0.0744546726,
-0.0024307966,
-0.0762656257,
-0.0118438555,
-0.026672028,
-0.0051589138,
-0.0453031659,
0.0193655342,
-0.0762183219,
0.0245611593,
0.0382657163,
0.0664041266,
0.0158770476,
-0.0037374187,
-0.0213497989,
0.0161453988,
0.0057764077,
0.0003606738,
-0.0370160341,
0.032895159,
-0.024813842,
-0.0077029336,
0.0429867357,
-0.0478305481,
-0.0025726433,
-0.0006016546,
0.0068062851,
-0.0635536313,
-0.0182046648,
0.0732362196,
0.0201321822,
0.0574889816,
0.0221554935,
-0.015644405,
0.0689454302,
0.0206789225,
-0.0090229046,
0.0378306471,
0.0188218243,
0.061324995,
-0.0014792731,
0.0469014458,
-0.0649926141,
-0.0069647864,
0.0013309817,
0.0609092042,
0.0043394426,
-0.0239797495,
0.0141738616,
-0.0339768268,
0.0491654798,
-0.0315116271,
-0.0174983609,
-0.046714928,
0.0189901814,
-0.0049221171,
-0.0009340949,
-0.0359814763,
0.0136360088,
0.0604683831,
0.0292771608,
0.0383743159,
0.0266289059,
0.0295406654,
0.0332447737,
-0.0078783091,
-0.0020610867,
0.0288614072,
0.0097265318,
0.0095461011,
0.0539246872,
0.0026666203,
0.0637167096,
0.0017680585,
0.0388556495,
0.0061667231,
0.0863134637,
0.0300002061,
0.0021323077,
-0.0521852225,
0.0011637025,
-0.0338920094,
-0.0641608313,
0.0077892095,
0.0318407007,
0.0331554972,
-0.0006461291,
0.0368642882,
-0.0355803482,
0.0025921573,
-0.0119237443,
0.0263043623,
-0.0341085643,
-0.0353800766,
0.0510671698,
0.022656912,
0.0034019155,
0.007124498,
0.0408755988,
-0.000037986,
0.0707471594,
-0.0335288905,
0.0309560616,
-0.0258730091,
0.0067490083,
0.0051729237,
0.0349756703,
-0.0367543884,
-0.0342282578,
-0.0025743898,
-0.0236666109,
-0.0580386445,
0.0478120223,
-0.0608458631,
-0.0416414104,
0.063622497,
-0.0277506951,
-0.0018774337,
-0.0177192166,
-0.0244312529,
-0.0257896725,
-0.0019905556,
-0.0101803448,
-0.0217375774,
-6.614807933e-33,
-0.0396199189,
-0.0102728,
0.0418684781,
-0.0212219395,
-0.0543220751,
-0.0269840378,
-0.0237014927,
0.0091327028,
0.0035609098,
-0.0119327754,
0.0060132886,
-0.0130482903,
0.0071817739,
-0.0034971486,
0.0117044365,
-0.0385335162,
-0.0034929542,
-0.037162181,
-0.0370361134,
-0.0182316471,
0.037351992,
0.0154454326,
0.0140015278,
-0.0155059602,
0.0024439723,
0.0136703029,
-0.0156090781,
0.0231823437,
-0.0426021405,
0.0506797805,
-0.0017007539,
0.0096412022,
-0.0060047172,
-0.0089477869,
-0.0357593931,
0.0228554793,
-0.0939832777,
-0.0350579508,
0.0047303019,
-0.0076319715,
-0.0652277321,
-0.005522518,
0.0474922583,
0.0072911577,
-0.0350518338,
0.0270866938,
0.007035248,
-0.0353421979,
-0.0281850789,
0.0357970744,
-0.0519379377,
-0.0138582503,
-0.0167745762,
-0.0257062949,
0.0372861028,
0.0058681546,
-0.0193882976,
-0.0198234171,
-0.0549921989,
0.0014319022,
0.007131842,
0.007427799,
0.0632753223,
0.0423503704,
0.035866186,
0.0517133698,
0.0089922575,
-0.010334623,
-0.0227606073,
-0.0026719891,
0.038564872,
0.050544899,
0.0111283185,
0.0461151525,
0.0097766295,
-0.0314901285,
-0.0353765152,
0.0359880477,
0.0214839112,
0.0470369756,
0.0431085825,
0.025867695,
-0.0109316381,
-0.0191343408,
-0.0645993948,
-0.0013676437,
-0.0118959183,
-0.0472286455,
0.0481695607,
0.0366800502,
0.0173568707,
-0.0945899561,
0.0342546254,
-0.0178993847,
0.015284108,
-0.0117382398,
-0.031351462,
0.0152901234,
0.0098754494,
-0.0156668834,
-0.0968484357,
-0.0789920241,
0.0214543603,
0.0001910823,
-0.0045644445,
0.0124254469,
-0.0181904603,
0.0633310825,
-0.0606578514,
0.0016850985,
-0.017191099,
0.00864405,
0.0297675431,
0.0217014123,
0.025964627,
-0.0453328118,
-0.0107307835,
0.0308062863,
-0.0162713472,
-0.0408340171,
-0.0127738668,
0.016278455,
-0.0382457636,
0.005849652,
-0.0146999769,
0.0299858637,
-0.036265064,
0.0101011591,
0.0367177539,
-0.0474471413,
0.0048472784,
0.0422253609,
3.186e-7,
-0.0024046206,
0.0373801552,
0.0032216799,
-0.0323564634,
0.032974638,
0.0412673987,
-0.0435852073,
0.0564068891,
-0.0428613611,
-0.0616240352,
0.0442243554,
0.0329297557,
0.0276065506,
-0.019333154,
-0.0809102729,
0.0035616653,
-0.0647599474,
0.0001417609,
-0.0543644205,
-0.0281078517,
0.0144876856,
0.0701290369,
0.0221680999,
0.0037575075,
0.0400512777,
-0.009024133,
0.0470941737,
-0.0576152503,
0.0389130116,
-0.0231463406,
0.0083160419,
-0.0009621363,
-0.0103029683,
-0.0037484814,
0.0137069589,
0.0572346486,
-0.034375526,
0.0842378736,
0.0083823195,
-0.0281329677,
-0.0008310555,
0.0121125719,
0.0186920669,
-0.0604100898,
0.0693737343,
-0.0915411413,
0.0002748497,
-0.0136425272,
0.0275997277,
-0.0251040347,
0.0169895999,
-0.0336870067,
-0.0210825242,
0.0079283603,
-0.0018161209,
0.019943703,
-0.0145035433,
-0.0748989359,
-0.0179777723,
0.0249528252,
-0.0393320359,
-0.044081904,
-0.0558173284,
0.0044739139,
-0.0040413728,
-0.0541586801,
-0.0216685049,
2.480826185e-34,
-0.0310271848,
-0.0117427334,
-0.0125861084,
0.0102553312,
0.0016536296,
0.008133552,
0.0151558071,
0.0071862983,
0.0672056228,
-0.0565224811,
-0.0120770289
] |
http://arxiv.org/abs/2401.14135v1 | http://arxiv.org/pdf/2401.14135v1 | Convolutional Neural Networks can achieve binary bail judgement classification | 2024-01-25T12:31:41 | 2024-01-25T12:31:41 | Amit Barman, Devangan Roy, Debapriya Paul, Indranil Dutta, Shouvik Kumar Guha, Samir Karmakar, Sudip Kumar Naskar | cs.CL | cs.CL, cs.CY, cs.LG | There is an evident lack of implementation of Machine Learning (ML) in the
legal domain in India, and any research that does take place in this domain is
usually based on data from the higher courts of law and works with English
data. The lower courts and data from the different regional languages of India
are often overlooked. In this paper, we deploy a Convolutional Neural Network
(CNN) architecture on a corpus of Hindi legal documents. We perform a bail
Prediction task with the help of a CNN model and achieve an overall accuracy of
93\% which is an improvement on the benchmark accuracy, set by Kapoor et al.
(2022), albeit in data from 20 districts of the Indian state of Uttar Pradesh. | Convolutional Neural Networks can achieve binary bail judgement classification: There is an evident lack of implementation of Machine Learning (ML) in the
legal domain in India, and any research that does take place in this domain is
usually based on data from the higher courts of law and works with English
data. The lower courts and data from the different regional languages of India
are often overlooked. In this paper, we deploy a Convolutional Neural Network
(CNN) architecture on a corpus of Hindi legal documents. We perform a bail
Prediction task with the help of a CNN model and achieve an overall accuracy of
93\% which is an improvement on the benchmark accuracy, set by Kapoor et al.
(2022), albeit in data from 20 districts of the Indian state of Uttar Pradesh. | [
0.0547549017,
0.0482369959,
-0.0197665412,
0.020279251,
-0.0902810544,
0.0240475629,
0.0286108498,
0.0362690203,
0.0264953971,
-0.0227416623,
0.005899902,
-0.0386885144,
-0.0151555827,
0.0631261542,
-0.002962111,
-0.02523751,
0.029399164,
0.0331910662,
0.0523597002,
0.006671187,
0.0251035467,
0.030799184,
-0.0167246498,
0.0286878292,
-0.0541609935,
0.0121156452,
-0.0334785692,
-0.0268257074,
-0.0137405153,
-0.0111256214,
0.0122966189,
0.0600750931,
0.0377966948,
0.0151432687,
0.0000016342,
-0.053270299,
0.0147548458,
-0.0323676392,
-0.0046268748,
-0.0036989851,
-0.0090004355,
-0.0178016331,
-0.0133812632,
-0.0097126001,
-0.0494140759,
-0.0048371935,
0.0231149066,
0.0432207063,
0.0180846266,
0.1047080904,
0.0153172622,
-0.0561744682,
0.0494753867,
-0.0287493244,
-0.0185339618,
-0.0408438742,
0.0060778982,
-0.0142027838,
0.0222672112,
-0.0481223986,
-0.0139463414,
-0.0197622553,
0.043404609,
-0.0427614525,
-0.0283240639,
-0.0244304426,
-0.0323182493,
-0.0205860436,
0.0227123462,
0.042926766,
-0.0149498619,
-0.0331404023,
-0.0478783548,
-0.0028431357,
0.0184976608,
-0.0058626607,
-0.0035998677,
0.0821625963,
-0.0401501656,
0.0686955974,
0.0295482427,
0.0284340959,
-0.0193527211,
0.0338878445,
0.0254902598,
0.0175053161,
-0.0233182032,
-0.0094994726,
0.000647113,
0.0424514897,
0.0255891811,
-0.0459612049,
0.0486268289,
0.0097105028,
-0.004704759,
0.00369523,
-0.0189736001,
-0.0638137683,
0.0373576097,
-0.0628599897,
0.0021454727,
-0.0213467814,
-0.0470997728,
0.0154074449,
-0.0548842363,
0.0345034488,
0.0116915684,
0.0372811556,
0.0274551436,
-0.0225419067,
0.0029866735,
0.01148762,
0.0007077364,
0.0437234417,
-0.0304984655,
0.0231513511,
-0.0510712489,
0.0080096228,
-0.034760695,
0.0168479923,
0.0202026423,
0.0272409879,
-0.0352562889,
-0.0256899372,
-0.0434790067,
-0.0776659921,
-0.0461687446,
0.0575323328,
0.0054131602,
-0.0256195869,
0.0212352872,
-0.0054023252,
-0.0087079117,
0.0030568566,
-0.0191980358,
0.0578301176,
-0.0303769559,
-0.0142180156,
-0.0796166062,
-0.0088437786,
-0.0081697023,
-0.0609930605,
-0.0569892228,
-0.0581171773,
-0.0143022444,
-0.0303539913,
-0.0213847291,
0.0022265327,
0.0384553336,
0.0450791642,
-0.0042804219,
0.0968462676,
0.0133011481,
-0.0296236929,
0.0028731343,
0.0185549501,
0.0641182214,
-0.0309814084,
-0.0104942312,
0.0482847989,
0.0098333377,
-0.0391969681,
0.0635575503,
0.0161595736,
-0.0033985043,
-0.0473206416,
-0.0242621601,
-0.0039004157,
-0.0483527444,
0.0117928553,
-0.0072558275,
0.0097895414,
0.014643644,
0.0075487047,
0.0021721241,
0.0491661839,
0.0244531073,
0.0787884146,
-0.0289213546,
-0.0041851499,
0.0201154202,
0.0104437927,
-0.047921557,
0.0194934681,
0.0022223068,
-0.0135643072,
-0.0298549924,
-0.020559039,
-0.0278583318,
-0.0166354422,
0.0193318576,
-0.0314910337,
0.0129898898,
-0.0402255766,
0.0474155955,
0.0063593108,
-0.0250289533,
-0.0310568269,
0.0582463779,
-0.0130585451,
-0.0169238076,
-0.0804217607,
0.042828586,
0.0195977073,
0.0035629214,
0.0027615649,
0.0229318924,
0.0002869142,
0.0106097115,
-0.0141057214,
0.0262322389,
-0.0050219148,
-0.0012878844,
-0.0296881292,
-0.0214106292,
0.0578916967,
0.0009140105,
0.0084109437,
-0.0085076308,
0.0368986018,
0.0173454341,
-0.0387259722,
0.0087100612,
0.0342466719,
0.0380725041,
-0.0130720809,
0.0370043553,
0.0357483216,
0.0523357131,
0.0000321436,
-0.0244929269,
-0.0599202588,
0.0348675773,
0.0250378083,
-0.003689551,
0.0025237638,
-0.0164842233,
-0.0108049177,
-0.0146596171,
-0.0237957835,
0.0488113649,
0.0421930701,
-0.0688057318,
-0.0248433761,
0.0221812874,
0.0058961431,
0.0089580743,
-0.0615394115,
0.0054810131,
0.0260412339,
-0.0316133909,
-0.0027560357,
0.0088216523,
-0.000317218,
-0.0326439664,
-0.1109214574,
0.0601648092,
0.0901537538,
-0.0142842559,
-0.006782732,
-0.0596327856,
-0.0327073261,
0.0220738109,
0.0133962212,
0.0106337462,
-0.0284676161,
-0.0300077591,
-0.0642184392,
-0.0182311721,
-0.0107239485,
-0.0372649543,
-0.0228997692,
-0.0063477512,
0.0181663185,
-0.0444869548,
-0.0353567228,
0.0377323031,
0.0459416807,
-0.0192707423,
0.0518037081,
-0.0050730132,
0.0030418132,
0.0029495754,
-0.0495034568,
-0.0299421176,
0.0260565933,
0.0062810825,
-0.0397315919,
0.0148469983,
0.017114995,
0.0091068586,
0.0025933697,
-0.0616776459,
-0.0197298899,
-0.0100092739,
-0.0104452604,
-0.0493767373,
0.0205341876,
0.0563840531,
0.0263856314,
-0.004553366,
0.0321882367,
-0.0309686847,
-0.0006526834,
0.0029172439,
0.0151112899,
0.0026400052,
-0.0042466093,
-0.0474937521,
-0.0231948849,
0.0028654952,
0.0418836474,
-0.0049197893,
-0.0238621105,
0.0248294063,
0.0156979356,
-0.0530630909,
0.0181139875,
-0.0220727921,
0.0503516756,
0.0194481146,
-0.061089702,
0.0436177067,
0.0344172753,
0.0808200538,
-0.040883787,
-0.0321110114,
-0.0079999035,
0.0728887096,
0.0201969407,
0.0158978123,
-0.013717887,
-0.0448866375,
0.0148312273,
0.0015449425,
-0.0009717555,
-0.0375246368,
0.0631483868,
-0.028720228,
0.0358276516,
-0.014320666,
-0.0603414699,
-0.0544579141,
-0.0688868091,
0.0122655537,
0.0014970405,
-0.0016192788,
0.0010800946,
0.0051260842,
0.0458086394,
-0.0725785643,
-0.0021835989,
-0.0289189219,
-0.0183491521,
-0.041946847,
0.0200269036,
0.0294247549,
-0.0029921969,
0.0174824186,
0.0054906793,
-0.0509444922,
-0.0028477174,
-0.0218737796,
0.0408897512,
-0.0047405427,
0.0221591797,
-0.00284258,
-0.0567863733,
-0.0138640134,
0.0102814902,
-0.0338748544,
-0.0281264447,
0.0148779573,
0.041364152,
0.0328131728,
-0.0444136187,
-0.0200789608,
-0.0079754479,
-0.0485932119,
0.0086760465,
0.0188416056,
-0.0666830838,
0.0044670878,
0.0116600124,
-0.0248142723,
-0.0281913932,
-0.0300640743,
-0.0000381571,
-0.0670499727,
0.0509552807,
0.0064259856,
0.0276442748,
-0.0621732436,
-0.0392777622,
-0.0352367982,
-0.0059980177,
-0.0024175141,
0.0027177441,
-0.0220415629,
-0.0300386902,
0.0370982997,
-0.0524588525,
0.0474442318,
0.0336681455,
-0.0540388599,
0.0286019947,
0.0392198004,
-0.0060704709,
-0.0764662474,
0.000612556,
-0.0074078469,
-0.0125553617,
-0.0124750538,
0.0132233007,
0.0608073696,
0.0323682912,
0.0076792403,
-0.036472477,
0.046790693,
0.009989243,
-0.0150783649,
0.0364611372,
0.0237666499,
-0.0326122306,
-0.0196922813,
-0.03034647,
-0.0290067401,
0.004016784,
0.009643401,
-0.0137246083,
-0.004152569,
-0.0570398197,
-0.0237815585,
-0.0179862939,
-0.0612949729,
0.050135497,
-0.0259428639,
-0.0191698596,
-0.0169699956,
0.0425718315,
-0.0476865359,
-0.0037952166,
0.0329427496,
0.0244577266,
0.0085409572,
0.0375567712,
-0.0303392727,
-0.0649822429,
0.0049618431,
-0.0056615728,
0.0529543385,
-0.0523025356,
0.0487479009,
-0.004824813,
0.0140497861,
-0.0289933793,
-0.0333216973,
-0.0227800142,
0.0144724892,
0.0348125882,
0.0209416617,
0.0090370812,
-0.0402672514,
-0.0779342577,
-0.0635619462,
0.0903633684,
-0.0044504204,
-0.0174268726,
-0.0352603644,
0.0243058596,
-0.0454302467,
-0.0236681439,
0.0072046421,
-0.007713181,
0.0743554831,
-0.0155002233,
-0.0414244756,
-0.038801793,
0.0133475391,
0.0059261802,
0.0207792725,
0.0826410577,
0.0059960969,
-0.0259592272,
-0.0409807488,
0.0230793711,
0.019987369,
0.0119461017,
0.002126043,
-0.0313079134,
0.0474395528,
0.0397654362,
-0.0355609134,
0.0121389255,
0.0212673619,
0.0107853841,
-0.0162829291,
0.029549472,
0.0696441904,
0.0623410866,
0.0326453783,
-0.0047029089,
-0.0140378894,
0.0114198355,
0.0146218417,
0.0412619822,
-0.0541480929,
0.0418290272,
-0.020752186,
0.0218987744,
-0.0049355719,
0.0306096897,
0.0107703973,
0.0308559965,
-0.0340207331,
0.0423086435,
0.0416754857,
-0.0343595259,
-0.0180362388,
-0.0320158452,
0.0188856982,
-0.0353325903,
0.0363875404,
-0.0425630398,
0.0201260746,
0.0405249484,
0.0099471156,
-0.0285190605,
0.0441045091,
-0.0125227347,
-0.0240167025,
-0.0098424479,
0.0039637061,
-0.0530699529,
-0.033933863,
0.0140396524,
-0.0138179632,
0.0421980768,
0.0340203866,
-0.0439212732,
-0.0513655059,
0.0031363121,
0.0164155196,
0.0085597392,
0.0245653912,
-0.0134948883,
-0.0003751268,
0.043564681,
0.0679873303,
0.0314976946,
-0.062683627,
0.0147692952,
0.0243677534,
-0.0185815711,
-0.0143657438,
-4.59540782e-33,
-0.0362005308,
-0.0336128399,
-0.0054976167,
0.0281605329,
0.0114817563,
-0.0036616356,
0.0101866955,
-0.0206775554,
-0.0562000722,
0.046391312,
-0.0395638421,
0.027422471,
0.0251260977,
-0.0231622271,
0.0359104425,
-0.0326671042,
0.0050431527,
0.0073284144,
0.0268392023,
-0.0101968003,
0.0462996401,
0.0276162643,
0.0221261792,
-0.0649868995,
-0.0073315152,
0.0143061606,
-0.041866418,
-0.0230748169,
-0.0733184963,
0.0379607156,
0.0182666015,
0.0352762192,
0.0489801876,
-0.008469047,
0.0301230773,
0.0456327759,
-0.0838733986,
0.0208489299,
-0.0548843369,
0.038414225,
0.0242009927,
0.0038958376,
0.008874773,
-0.0032897198,
-0.0521101691,
0.0245360602,
0.0325644389,
0.0097438265,
-0.0558649823,
-0.0541195124,
-0.0607889816,
0.0388875008,
0.0131041426,
0.0146717345,
0.0081445482,
0.0113643538,
0.0111055141,
-0.0131373014,
-0.0798225179,
-0.0141086141,
0.0012895735,
0.0086354539,
-0.0163054187,
-0.0285826214,
-0.0778992102,
0.0365517475,
-0.0049559642,
-0.0707848296,
-0.0582038909,
0.0289065596,
0.0507639088,
0.0700438991,
0.027905101,
0.0172503926,
0.0986642316,
-0.1166440398,
-0.0868564844,
0.0074787661,
0.0671813264,
0.0769628137,
0.0206512827,
-0.018585192,
-0.070728004,
0.0096869878,
0.0074219764,
-0.0414038301,
-0.0165979993,
-0.0057036909,
0.026821211,
-0.0291592069,
0.0940153822,
-0.0046601146,
-0.0069565028,
-0.0417507179,
0.0293480791,
0.0163305402,
-0.0155808013,
0.0227724798,
-0.0018405956,
0.0368388034,
-0.0630612448,
-0.070884198,
-0.0026568912,
0.0274664946,
0.0058827554,
0.0006571849,
0.0891487151,
0.0010010209,
-0.0573485941,
-0.0085645914,
0.0269226003,
0.0383921154,
0.0524911061,
0.1045545861,
-0.0035439904,
-0.0153344506,
0.0279808901,
-0.0400110371,
-0.0085142814,
0.0211621169,
-0.0059011965,
-0.0236464497,
0.0333061293,
-0.0258945711,
0.0355062932,
0.0010113723,
-0.0372662097,
0.0054902826,
0.0242577437,
-0.0384918451,
0.0229715779,
0.0824496746,
2.269e-7,
0.017566368,
0.0262700263,
0.0046761464,
-0.0172049347,
-0.040174678,
0.0246971194,
-0.014043319,
0.0627002791,
-0.0250113532,
0.0251649395,
0.0090636229,
-0.0107022859,
0.0677225515,
0.0637908205,
-0.0222315602,
-0.0180097781,
-0.054942105,
-0.0143757146,
-0.0481969565,
0.0542762615,
-0.0182982441,
0.1149733141,
0.0701674074,
-0.0309208818,
0.0692897141,
-0.0301043838,
-0.0185041595,
0.0493624881,
-0.0162291024,
0.0125507852,
0.0484719537,
-0.0262245052,
-0.0342049152,
0.0459430628,
-0.0039387248,
0.0492015556,
-0.0282186866,
0.0270248428,
-0.0269185379,
0.0382730439,
-0.0157451574,
-0.0367016569,
0.0544748418,
0.0063109603,
0.0630398318,
0.0476974137,
-0.0051948377,
-0.0130295325,
-0.0162031464,
0.0081684496,
0.0422566794,
0.0267045889,
-0.0029620104,
-0.0316868648,
0.0427102819,
-0.0604774356,
0.0084823072,
-0.0707039759,
0.0198950991,
0.0283383802,
0.0027194486,
-0.0014188539,
0.0038537653,
-0.0017371055,
0.1066892371,
-0.0575417392,
0.0193912908,
1.623012167e-34,
0.0427468084,
-0.0152642988,
0.0167500079,
-0.0215952415,
0.0194640048,
0.0533769466,
0.0008807234,
-0.017806476,
0.0009072224,
-0.0733025372,
0.0181116946
] |
http://arxiv.org/abs/2401.14113v2 | http://arxiv.org/pdf/2401.14113v2 | On the Affinity, Rationality, and Diversity of Hierarchical Topic Modeling | 2024-01-25T11:47:58 | 2024-02-01T03:47:28 | Xiaobao Wu, Fengjun Pan, Thong Nguyen, Yichao Feng, Chaoqun Liu, Cong-Duy Nguyen, Anh Tuan Luu | cs.CL | cs.CL | Hierarchical topic modeling aims to discover latent topics from a corpus and
organize them into a hierarchy to understand documents with desirable semantic
granularity. However, existing work struggles with producing topic hierarchies
of low affinity, rationality, and diversity, which hampers document
understanding. To overcome these challenges, we in this paper propose Transport
Plan and Context-aware Hierarchical Topic Model (TraCo). Instead of early
simple topic dependencies, we propose a transport plan dependency method. It
constrains dependencies to ensure their sparsity and balance, and also
regularizes topic hierarchy building with them. This improves affinity and
diversity of hierarchies. We further propose a context-aware disentangled
decoder. Rather than previously entangled decoding, it distributes different
semantic granularity to topics at different levels by disentangled decoding.
This facilitates the rationality of hierarchies. Experiments on benchmark
datasets demonstrate that our method surpasses state-of-the-art baselines,
effectively improving the affinity, rationality, and diversity of hierarchical
topic modeling with better performance on downstream tasks. | On the Affinity, Rationality, and Diversity of Hierarchical Topic Modeling: Hierarchical topic modeling aims to discover latent topics from a corpus and
organize them into a hierarchy to understand documents with desirable semantic
granularity. However, existing work struggles with producing topic hierarchies
of low affinity, rationality, and diversity, which hampers document
understanding. To overcome these challenges, we in this paper propose Transport
Plan and Context-aware Hierarchical Topic Model (TraCo). Instead of early
simple topic dependencies, we propose a transport plan dependency method. It
constrains dependencies to ensure their sparsity and balance, and also
regularizes topic hierarchy building with them. This improves affinity and
diversity of hierarchies. We further propose a context-aware disentangled
decoder. Rather than previously entangled decoding, it distributes different
semantic granularity to topics at different levels by disentangled decoding.
This facilitates the rationality of hierarchies. Experiments on benchmark
datasets demonstrate that our method surpasses state-of-the-art baselines,
effectively improving the affinity, rationality, and diversity of hierarchical
topic modeling with better performance on downstream tasks. | [
0.0077989376,
0.0625794455,
-0.0043157646,
0.0406717733,
-0.0313914753,
0.0246162191,
-0.0250659902,
-0.0125632863,
-0.0590025336,
0.0011200334,
0.0231743418,
0.0124224816,
0.0030150949,
0.0876069814,
-0.0391901247,
-0.1024410352,
0.0164362006,
-0.0333578438,
-0.010494465,
-0.0132689206,
-0.0129714049,
-0.0021978875,
-0.0163371246,
-0.0006469364,
0.0301979426,
-0.0293289106,
-0.0426209047,
0.0029895359,
0.0191159155,
-0.012085096,
-0.0093450574,
0.0765483081,
0.0756376982,
0.003084182,
0.0000022272,
0.0015413174,
-0.0264535248,
-0.0318832137,
0.0022556193,
0.0041451044,
0.0163662024,
-0.0856649801,
0.0402796865,
-0.0117451288,
0.0022731291,
0.0064161075,
-0.0306078419,
0.0595337935,
0.0539343432,
0.0348232202,
-0.0315218195,
-0.0376479253,
0.0570619628,
-0.0365372598,
0.010865178,
-0.0724024028,
-0.0019161579,
0.0454765223,
-0.0413492471,
0.0844088271,
-0.0660746098,
-0.0008895028,
0.0269823577,
0.0142901605,
0.0864677504,
0.0172070563,
0.0526497327,
-0.0164904576,
0.0005415284,
0.0839340538,
-0.0099713206,
0.0016276718,
0.0169979148,
-0.0150507856,
-0.0814385712,
-0.0160097573,
-0.0326665044,
0.0160194691,
-0.0349855945,
0.0333638787,
0.0911726356,
0.0397218503,
0.0081966957,
0.0206845812,
-0.0241985936,
0.0471511297,
0.0011011973,
-0.0269603189,
-0.0256804973,
0.0392753184,
0.0557025895,
-0.0197577458,
0.0336676389,
-0.0113674803,
-0.0083085867,
0.0030765822,
0.0170135032,
-0.033106897,
-0.0221365411,
-0.0120028267,
-0.0503459461,
0.0522468835,
0.0169791337,
0.061313767,
0.063605912,
-0.0136770625,
-0.0393715054,
-0.0323159285,
-0.0425009876,
-0.0260200053,
-0.0141620403,
0.0300136246,
-0.030115515,
0.0437702499,
-0.0622395128,
-0.0296092387,
-0.0041526314,
0.0167494845,
0.002419984,
-0.0413262248,
-0.0385664813,
-0.0012075162,
-0.0478604399,
-0.0359475166,
-0.0069445763,
0.0648134798,
-0.0634046793,
0.029854862,
-0.0029506211,
0.0489422642,
-0.003227351,
0.0359771699,
0.0171348732,
-0.0105827153,
-0.0526609607,
0.0050146659,
-0.0506002754,
-0.0439388119,
-0.0503690429,
-0.0015726675,
0.0348405764,
-0.0070391991,
0.0293302517,
-0.0608807877,
-0.0012207803,
0.0032602716,
0.0214391127,
-0.0150341811,
-0.0436892621,
-0.0279557649,
-0.0493265167,
0.0214070901,
-0.0236038156,
-0.0029171277,
0.0532400049,
0.0433222093,
-0.0053878911,
0.0883503631,
-0.0321424156,
0.0940854996,
0.0187004954,
0.008970893,
-0.029234888,
-0.048421707,
0.0121018188,
0.0380448997,
-0.0128839668,
0.0046704798,
-0.0504437461,
0.0067665689,
0.0819103718,
0.0332160443,
-0.0842723548,
-0.0088588297,
0.0239693839,
0.0582772829,
0.0096145701,
-0.0130591551,
-0.0355567299,
0.0378193781,
-0.0512183309,
0.0210613646,
0.0226627048,
0.0506892987,
0.0176696125,
-0.0402040631,
0.0579735115,
-0.0029457849,
-0.037367221,
-0.0305473991,
0.0146443099,
-0.0011532842,
0.0142682334,
-0.0339496061,
0.0724235177,
0.0309038181,
0.0010632124,
0.0076869377,
0.028179612,
-0.0013730186,
0.0322733447,
0.0079845432,
0.0464185774,
0.0713940188,
0.0611713901,
0.0114476597,
0.0120496228,
-0.0007738177,
0.0082966574,
0.0400978811,
-0.0287837107,
-0.013647995,
0.0291416906,
-0.056573268,
-0.0508694835,
0.0043838434,
0.0211986434,
0.0106869806,
-0.0226385538,
-0.0120930681,
-0.0284878667,
-0.0095548443,
0.0917357504,
0.0075867432,
-0.0088132368,
0.035678491,
0.041655127,
-0.0040632463,
0.0722570643,
0.0154984146,
0.0158308856,
-0.0164809804,
0.0272987317,
-0.0445053428,
-0.0394057408,
-0.0502370261,
-0.025640035,
-0.0552821569,
-0.0347878858,
-0.0964646265,
-0.0043546297,
0.0060473322,
-0.0284146965,
-0.0242900662,
0.0619785786,
-0.0264151394,
0.0076180845,
-0.0184665807,
-0.0380240418,
-0.0051178462,
0.0189536866,
-0.006677228,
0.0810777694,
-0.060501989,
-0.0075557465,
0.0056560379,
0.0062380787,
-0.0030849776,
-0.0159763247,
0.0394077823,
0.0406228863,
-0.0041493727,
-0.0018869991,
0.0322418287,
-0.0146633377,
0.0052151671,
-0.0215434972,
-0.0064550918,
-0.0114039481,
-0.0392342806,
-0.0007515497,
0.036017552,
0.0089800097,
0.0292521752,
-0.0060950029,
-0.0160471946,
0.0394877978,
-0.0436119027,
0.0149167003,
0.0124144563,
-0.0192034002,
0.0346673019,
-0.0100900689,
-0.0358727463,
0.0146527383,
0.0376750417,
-0.0361671448,
-0.00917398,
-0.0235873256,
0.0083798636,
0.0104963342,
-0.0219684187,
-0.0365620367,
0.0108114406,
0.0106701506,
-0.0335959792,
0.0567339659,
0.0175783709,
0.0029195324,
0.0783555433,
0.0596535914,
-0.0350793116,
-0.0156438127,
-0.0056200209,
-0.0341423675,
0.0650022179,
-0.0378897712,
-0.049662035,
0.0195725001,
0.1059467793,
0.0157851167,
0.0288819112,
-0.0579687953,
-0.0294238813,
-0.0314547084,
-0.0128122745,
-0.0551481172,
-0.0039692544,
0.0307732727,
0.0870917514,
-0.0315253213,
-0.0239227805,
0.0590007901,
0.0224638451,
0.026264606,
-0.059069261,
-0.0101630762,
-0.0567102954,
0.0741029754,
-0.0125775784,
-0.0014439384,
-0.0181027371,
-0.0410526097,
0.0087101758,
0.0250416063,
0.0267328694,
0.021008499,
0.0596797019,
-0.0222139508,
0.0098452521,
-0.0019046813,
-0.0249928553,
0.0034443524,
-0.0356204361,
0.0261741057,
0.002093269,
0.0619614422,
-0.0260942802,
-0.044941932,
-0.0096741281,
0.0091935862,
0.0316900648,
-0.0047225123,
-0.0163717326,
-0.0212779697,
0.0561530553,
0.0592636652,
-0.0203061942,
0.0070861671,
-0.0221853182,
0.0555480458,
0.0579171814,
-0.0178732723,
-0.0634687021,
-0.0256411228,
0.0527678318,
0.0266409311,
0.0090790316,
-0.0058210641,
-0.0148578975,
0.0103536462,
-0.0311453845,
-0.0097191688,
0.0680110082,
0.0201558881,
-0.0746804625,
-0.0019778628,
-0.0214927979,
0.0609961972,
-0.0134707214,
-0.0026122688,
-0.0715947673,
-0.004274121,
0.0140600996,
0.0458451025,
0.0352859199,
-0.032738924,
0.0302772522,
-0.0141275926,
0.0323857106,
0.0002944602,
-0.028209392,
-0.0317374468,
-0.094787471,
-0.0048524141,
-0.0328869857,
0.0111588826,
0.0186298452,
0.0183336958,
-0.0002640409,
0.0112465173,
-0.0131578473,
-0.0355204642,
0.0112465564,
0.0233381726,
0.0182779562,
-0.010129327,
-0.0026542125,
0.0110767912,
0.0089423005,
-0.07976567,
-0.0609229878,
-0.0285606459,
-0.0232573729,
0.0192935299,
-0.0183186084,
0.024642095,
-0.0476775654,
0.0396554396,
-0.0642523691,
-0.0427012593,
-0.0137765184,
0.0657385886,
-0.0577306636,
-0.0272271186,
-0.0645474568,
-0.0183077194,
-0.0085999174,
0.0609779842,
-0.0046705082,
-0.0250851214,
-0.0010663467,
-0.0372968987,
0.0384122618,
0.0019936438,
0.0442019403,
-0.0003609179,
-0.0219093319,
-0.0710479319,
-0.0770606995,
-0.0204609428,
-0.0137163103,
0.00001969,
-0.0490976535,
0.016923245,
0.0453395993,
-0.0520326421,
-0.0243553482,
0.0220488906,
0.010582055,
-0.0352504216,
-0.0171343535,
-0.0281033143,
0.0313565135,
-0.0106126396,
-0.0034939495,
0.0023574654,
-0.0433228128,
-0.0200700238,
-0.0106998431,
-0.0378706194,
0.0105430717,
0.0427948274,
-0.0471731536,
-0.054154288,
-0.0255223904,
0.0091983108,
-0.0722851157,
0.0156901143,
-0.0007949326,
0.0106797572,
-0.0129950549,
0.0181983355,
-0.0344916023,
0.0622082986,
-0.0305448547,
0.0461750887,
-0.0781646892,
-0.022746522,
-0.0018351554,
-0.0372535363,
0.065460138,
0.015505475,
-0.0277254097,
-0.0472656153,
0.0160381477,
-0.0285335518,
-0.0050333845,
-0.020516552,
0.0154259251,
0.0468799882,
-0.0206521247,
-0.0116962986,
0.0214237459,
0.0196830295,
0.0188943185,
0.0038641591,
0.0062092477,
-0.0379425846,
0.0313560516,
-0.0106960526,
0.0182948094,
-0.0067430367,
-0.0019128046,
-0.0271052029,
0.0246541034,
-0.0161134861,
0.0481375754,
0.0264980812,
-0.0125469537,
-0.0191341843,
-0.0355617106,
-0.0753491074,
0.0267836824,
0.0201278701,
-0.0018778879,
0.0229970235,
0.0392872803,
0.0290914513,
-0.0291005075,
0.0028323063,
0.0017604266,
0.0443786681,
0.0087954002,
-0.012214914,
0.0053782831,
0.0312392823,
-0.0022256509,
0.0165040903,
0.0669976398,
0.0157316644,
-0.0241358615,
0.0239307825,
0.0168263223,
-0.0413371734,
-0.0051068035,
0.016053848,
0.0846525803,
-0.0270364489,
0.0536134355,
-0.0609394833,
0.0081860097,
0.001058383,
-0.0351479016,
0.0033033278,
-0.0099280998,
0.0567888282,
0.0556443557,
-0.024421908,
-0.0510342754,
-0.0488663875,
0.032081157,
-0.0444757789,
-0.0359351486,
-0.007855555,
-7.319735732e-33,
-0.0124275954,
-0.0066372408,
-0.0034189492,
-0.0192994513,
0.0126526095,
0.0269540828,
-0.0453903265,
-0.0282671843,
0.004380466,
0.0314463489,
0.0013779316,
0.0461565033,
-0.0080039082,
0.0389293842,
-0.0090709664,
0.0248632748,
0.0543554612,
0.005168221,
0.0050877361,
0.0420261547,
0.059225522,
0.0247853007,
-0.0064292559,
-0.0398823135,
-0.0172959547,
-0.0351844691,
0.0132481996,
0.0272807144,
0.0517887883,
0.0193438362,
-0.0434167013,
0.0220032614,
0.0356508791,
-0.01764117,
-0.0117443353,
0.0283272769,
-0.1017135084,
-0.0542280711,
-0.0232112668,
-0.0335620083,
-0.0396610349,
-0.0437746681,
0.0020144321,
0.044097878,
0.029148411,
-0.0558477752,
0.0354827717,
-0.01656886,
-0.0425961651,
-0.0666670203,
-0.0423672833,
0.0448162258,
-0.0298270304,
0.0033347649,
0.0188674983,
-0.0057820734,
0.0133029483,
0.0407861955,
-0.0219715033,
0.0385479853,
0.0116653033,
0.0347220115,
0.013698102,
-0.0314456224,
-0.0104781426,
0.0520681478,
-0.0299952216,
0.0445370451,
-0.0279891379,
0.042804502,
0.0230312608,
0.0399592184,
0.0127998935,
0.0060597598,
0.0416168831,
0.024217017,
-0.0792636201,
0.0460039265,
-0.0015650131,
0.0557070188,
0.0157240815,
0.0734408423,
-0.0248829946,
-0.0256670825,
-0.0147356624,
0.0110292714,
0.0010951367,
-0.0430835299,
-0.0182600804,
0.0164386015,
-0.0204150993,
-0.0064447196,
0.0461127013,
-0.0131789641,
-0.0091478806,
0.0633864403,
-0.0086088721,
0.0369584262,
0.0119568631,
0.0457646959,
-0.0625260323,
-0.0417806879,
0.0457543768,
-0.0142117813,
-0.0291414727,
-0.0000975071,
-0.020049151,
-0.0438035652,
-0.0202337001,
-0.0287652668,
-0.021087775,
-0.0228758808,
-0.0236165971,
0.0476613604,
0.0014849158,
0.0038942725,
-0.0175517779,
0.0990486667,
-0.0036703208,
0.0405564792,
-0.0172709674,
-0.0172063299,
-0.0175656732,
0.0496495441,
-0.0028375282,
0.0001283856,
-0.0244848467,
-0.0261590723,
0.0875584185,
-0.0710852444,
-0.0289322156,
0.0462714024,
2.875e-7,
0.0202310383,
0.0808029175,
0.0114266044,
-0.0429421738,
0.0306380969,
0.0191182923,
-0.0189528335,
0.0242790617,
-0.0389239676,
-0.0524786748,
0.0105846636,
0.0599352904,
0.0186761096,
0.0258109197,
-0.0498163328,
0.0282498691,
-0.0326136202,
-0.0350435637,
-0.02977648,
-0.0134171862,
0.041402705,
0.0808001012,
0.0014403482,
0.0061676037,
0.0402201898,
0.0444128066,
0.0604190677,
0.0075162235,
-0.0346204229,
0.0143569764,
-0.0024826815,
-0.0108602205,
0.0016978966,
-0.002275141,
-0.0035598224,
0.0066356827,
0.0171622373,
0.066919893,
0.0030638727,
-0.0391001999,
0.0035328225,
-0.0230225064,
-0.0156149343,
-0.0396469869,
0.0378849693,
-0.0449952632,
-0.0545442551,
0.0070805773,
-0.0646116808,
0.0543259904,
0.0002309625,
0.0020698379,
-0.0084855575,
0.005177516,
-0.0207811277,
0.0024899517,
-0.0104827387,
-0.0609632321,
0.0328490585,
0.0277137961,
-0.0175304376,
-0.0319279619,
-0.0064437673,
0.0036856038,
0.0349076465,
0.0016099489,
-0.0034321561,
2.500110492e-34,
0.0473707318,
-0.0203362014,
-0.0634970739,
-0.0178963691,
0.0157771278,
-0.0350101776,
-0.0740439296,
0.0107410783,
0.01760488,
-0.071395345,
0.0279328134
] |
http://arxiv.org/abs/2402.01708v2 | http://arxiv.org/pdf/2402.01708v2 | Not My Voice! A Taxonomy of Ethical and Safety Harms of Speech Generators | 2024-01-25T11:47:06 | 2024-05-15T15:26:42 | Wiebke Hutiri, Oresiti Papakyriakopoulos, Alice Xiang | cs.CL | cs.CL, cs.AI, cs.CY, eess.AS | The rapid and wide-scale adoption of AI to generate human speech poses a
range of significant ethical and safety risks to society that need to be
addressed. For example, a growing number of speech generation incidents are
associated with swatting attacks in the United States, where anonymous
perpetrators create synthetic voices that call police officers to close down
schools and hospitals, or to violently gain access to innocent citizens' homes.
Incidents like this demonstrate that multimodal generative AI risks and harms
do not exist in isolation, but arise from the interactions of multiple
stakeholders and technical AI systems. In this paper we analyse speech
generation incidents to study how patterns of specific harms arise. We find
that specific harms can be categorised according to the exposure of affected
individuals, that is to say whether they are a subject of, interact with,
suffer due to, or are excluded from speech generation systems. Similarly,
specific harms are also a consequence of the motives of the creators and
deployers of the systems. Based on these insights we propose a conceptual
framework for modelling pathways to ethical and safety harms of AI, which we
use to develop a taxonomy of harms of speech generators. Our relational
approach captures the complexity of risks and harms in sociotechnical AI
systems, and yields a taxonomy that can support appropriate policy
interventions and decision making for the responsible development and release
of speech generation models. | Not My Voice! A Taxonomy of Ethical and Safety Harms of Speech Generators: The rapid and wide-scale adoption of AI to generate human speech poses a
range of significant ethical and safety risks to society that need to be
addressed. For example, a growing number of speech generation incidents are
associated with swatting attacks in the United States, where anonymous
perpetrators create synthetic voices that call police officers to close down
schools and hospitals, or to violently gain access to innocent citizens' homes.
Incidents like this demonstrate that multimodal generative AI risks and harms
do not exist in isolation, but arise from the interactions of multiple
stakeholders and technical AI systems. In this paper we analyse speech
generation incidents to study how patterns of specific harms arise. We find
that specific harms can be categorised according to the exposure of affected
individuals, that is to say whether they are a subject of, interact with,
suffer due to, or are excluded from speech generation systems. Similarly,
specific harms are also a consequence of the motives of the creators and
deployers of the systems. Based on these insights we propose a conceptual
framework for modelling pathways to ethical and safety harms of AI, which we
use to develop a taxonomy of harms of speech generators. Our relational
approach captures the complexity of risks and harms in sociotechnical AI
systems, and yields a taxonomy that can support appropriate policy
interventions and decision making for the responsible development and release
of speech generation models. | [
0.0169633795,
0.0709006637,
-0.0201662369,
-0.0419586226,
-0.0460037403,
-0.0176357161,
0.0819447488,
-0.0168216862,
0.0095134871,
-0.030186614,
-0.0081494385,
0.0029133789,
0.0167174488,
0.0194077231,
0.0640667602,
-0.0356860869,
0.0268206373,
-0.0213230904,
0.0058011198,
0.0170706213,
0.0201391876,
0.0270824935,
-0.0268801637,
0.0709001645,
-0.0641307458,
-0.0340544879,
0.0213474222,
0.0138859572,
0.0264725126,
0.0217183698,
-0.0209255945,
-0.0188705064,
0.0201046374,
0.0643548518,
0.0000015402,
-0.0544491597,
0.01781152,
0.020296229,
-0.0051061902,
-0.0273674354,
0.0257708263,
0.0373619013,
-0.0503763668,
0.0352472961,
-0.031188095,
0.0363601856,
0.0341881774,
-0.0115848342,
0.0206126422,
0.0570471622,
-0.0321524851,
-0.0049907826,
0.0586160496,
0.0215145946,
-0.0063292049,
0.0092412364,
0.0084987152,
-0.0291115232,
0.0007992798,
-0.0236701127,
0.0544809289,
0.0457007401,
-0.0130956108,
-0.0262421127,
-0.0007602215,
-0.0191850811,
-0.0120876282,
-0.0806846246,
0.0077071637,
0.0458624363,
0.0478753857,
0.0160877332,
0.0221958291,
0.0314847678,
0.0223856401,
-0.0157496389,
0.0253874455,
0.0507399961,
-0.0277548078,
-0.022671584,
0.0061746105,
0.0205326583,
0.0325083174,
0.0214275271,
0.0607960969,
-0.0208599251,
-0.0139538636,
-0.0004221366,
-0.0191470459,
-0.0540676191,
0.0115332389,
-0.0132956328,
0.02594533,
0.0157819502,
-0.0288422853,
0.023992382,
-0.0585299246,
-0.0509560406,
0.020064326,
-0.0517478287,
0.0028148626,
-0.0000951647,
-0.0026575041,
-0.0308845341,
0.0331802629,
-0.034712974,
-0.0230060294,
0.0291922744,
-0.0118232435,
0.032932654,
0.0194283593,
-0.014087623,
-0.0119425971,
-0.0085562989,
-0.0255088061,
0.0100403912,
-0.0009802646,
-0.0025527654,
-0.0920882598,
0.0248407107,
-0.0460258499,
0.0135895004,
-0.0387257449,
-0.0139919184,
0.0374213718,
-0.0474778824,
-0.0284859296,
0.0619355477,
-0.0015736994,
-0.0611049905,
0.0277512353,
-0.005501824,
-0.0042239754,
0.0250153989,
0.0054812725,
0.1256525069,
0.0173589252,
-0.0005974506,
-0.0360427536,
-0.0667158365,
-0.0062106042,
-0.0239120461,
-0.0128581747,
-0.0179063659,
-0.0375436805,
-0.0231921729,
-0.043649409,
-0.0485242866,
0.0137054622,
0.0062546758,
-0.013047358,
0.0214635096,
-0.017911993,
0.0502221063,
0.057368692,
0.0378911383,
-0.0168715771,
-0.0039348006,
0.0682447553,
-0.0065642265,
0.0094938995,
-0.0191598348,
0.0607473999,
0.0355346128,
-0.0084722983,
-0.0043728054,
-0.0126797697,
-0.0166606512,
-0.0744583905,
-0.0009289486,
0.0104839886,
0.0291530788,
-0.0199743547,
0.0027764244,
0.0677017346,
0.0707092732,
-0.0322321691,
0.0130023835,
0.0226805974,
-0.0134969074,
-0.042202808,
-0.0362434946,
0.0243411306,
-0.0363141149,
-0.0226935018,
-0.0395935401,
-0.0120859602,
0.017748002,
-0.0397247896,
-0.0049653924,
0.0317920074,
-0.0515124016,
0.0231460202,
0.0318502001,
-0.002564501,
0.0377254337,
0.0023975412,
-0.0352567732,
-0.0267794766,
-0.0630631745,
0.0176784024,
-0.0618620105,
-0.0037209282,
0.0814816058,
-0.002672967,
-0.0045491992,
0.0080497079,
-0.0739286393,
-0.0654016882,
0.0467765629,
-0.0536530949,
-0.0362867787,
-0.0096098408,
0.0034556384,
0.0140719134,
0.0298844017,
-0.0045473739,
0.0257173292,
-0.0329595469,
-0.012279355,
-0.0043567275,
0.0125129186,
-0.0252629332,
-0.0299657062,
-0.0151280686,
-0.0593568496,
-0.0301980153,
0.0070847487,
0.0370099247,
0.0450013094,
-0.0484846234,
0.0384687297,
0.0430531129,
0.0017906986,
-0.0055866041,
0.0275537018,
-0.0031304958,
0.0991000235,
-0.014376631,
-0.080160588,
0.0160674509,
0.0509584136,
0.0209323559,
0.0097212465,
0.0690133199,
-0.0639198869,
-0.0126464888,
-0.0214411896,
-0.0073868139,
-0.0244231932,
-0.0318480283,
0.0101515613,
-0.0081061162,
-0.0082609802,
-0.0571688414,
-0.0038010124,
0.0180807412,
0.022066975,
-0.0092735803,
-0.0465336703,
0.0103196502,
-0.0487991869,
0.0805692002,
0.0271165892,
-0.0466186889,
-0.0341919065,
0.0147159975,
-0.0514365323,
-0.0253067072,
-0.0008932632,
-0.0190955605,
0.0041150246,
0.0095205558,
-0.0166793372,
-0.0168064274,
-0.0030546763,
-0.0657930821,
0.0101489332,
0.0277547836,
-0.0323072411,
-0.0203235336,
-0.0521270633,
0.0017985177,
0.0044621048,
-0.0222726688,
0.0046589426,
0.01023502,
-0.0356055573,
-0.0364575796,
0.0022150334,
-0.0382265784,
0.0363966748,
0.0011591107,
-0.0101054059,
0.049922891,
0.0426645651,
0.0125574721,
0.0476757772,
0.0623789802,
0.0766918585,
0.0363560766,
0.0610334575,
-0.0088204443,
-0.0211230423,
0.0210567117,
-0.0020446982,
-0.0064340848,
-0.0374094658,
-0.0353473164,
-0.0119070858,
0.0500429198,
0.0051887874,
-0.0126063554,
0.0603922568,
-0.0518508367,
-0.0892549753,
0.0169152729,
-0.1223068982,
-0.0534848571,
0.0378693342,
-0.0082310541,
-0.0597179607,
0.0156095074,
-0.0020727471,
0.0526527353,
0.033634454,
-0.0444118418,
0.0244285613,
0.0461285301,
-0.0141332187,
0.040184699,
0.0620271303,
-0.0332339033,
-0.0156682543,
-0.014457006,
0.0335819237,
-0.064212285,
0.0004464238,
-0.0294007901,
0.0114197964,
0.0025351159,
-0.0376536325,
-0.0267802477,
-0.0324122421,
0.0008704953,
0.0303995516,
0.0186905395,
-0.0173239466,
0.0412519574,
-0.0185059123,
0.0759191737,
0.0040272409,
0.0455637574,
-0.0056417855,
-0.0039302995,
-0.0031058926,
0.0251111779,
-0.0091296667,
0.0197145957,
0.0567099117,
0.0047474317,
0.0085731894,
0.0395776108,
0.0177626107,
-0.0313619561,
-0.0091209002,
0.003318547,
0.0304339305,
0.0231913421,
0.0250128657,
0.0551021323,
0.0111165736,
0.0260835793,
0.0141886501,
-0.0301117748,
0.0052323062,
0.0427500531,
0.0272570141,
-0.0317465886,
-0.0053954469,
-0.0115952929,
-0.0652175546,
0.056363523,
0.015451692,
0.0632165372,
-0.1061955914,
-0.0147880372,
-0.0024633619,
-0.053333126,
0.074056156,
0.0239912625,
-0.0165617727,
-0.0228732806,
0.0164458118,
-0.0219739601,
0.0166924,
0.0498711243,
-0.0031921375,
0.0065518338,
-0.0379674956,
0.0502531528,
-0.0670318231,
-0.0009075545,
0.0043215128,
-0.0331669785,
0.1063204184,
0.0781403556,
0.027085783,
-0.0544005223,
0.0011962795,
-0.0281814504,
0.0362075828,
0.0000524828,
0.0396183468,
0.0209179465,
0.0325591862,
-0.0207149759,
-0.0406588614,
0.0607004128,
-0.0075439587,
-0.0335909463,
0.0200781953,
0.0441878177,
-0.0235804208,
-0.0338512212,
-0.0413054563,
0.0050948481,
0.0526062176,
0.0211712644,
0.0006069344,
-0.0356507264,
-0.0016033623,
-0.0296842866,
0.0198949724,
-0.0363082103,
-0.0536944941,
0.0055740289,
0.0205887388,
-0.0939824283,
-0.0588793941,
-0.0286404546,
-0.044113636,
0.0280372482,
0.0362371504,
-0.0305198301,
0.0376015939,
0.0183708891,
-0.0117518418,
-0.0034825699,
0.0444233865,
0.0296232998,
0.0259417798,
0.0098244827,
-0.0235163122,
0.0028219768,
0.0145068271,
0.0293484125,
0.0201039519,
-0.0383728929,
0.0619379766,
0.0463903211,
0.0315965824,
0.0228300691,
-0.0109215621,
-0.0247248616,
0.0549783036,
0.0226027984,
-0.0411798693,
0.0249817539,
0.0847108215,
-0.0103178564,
-0.0226205979,
-0.007855461,
-0.0005078124,
0.006635299,
0.0051545957,
-0.0088276174,
-0.0832226127,
0.0195441861,
0.0095431758,
0.0176781137,
-0.0138607817,
-0.0725943521,
0.046600379,
-0.0208904799,
-0.0402692556,
-0.000206208,
0.0091929128,
-0.0195155479,
0.0082728304,
0.0076815304,
0.0068924888,
-0.0002886622,
-0.0164695159,
-0.0437345915,
0.0102342525,
0.0072138035,
0.0438561179,
0.0275756866,
0.0579469949,
-0.0013207202,
0.0145268105,
-0.0059058554,
0.0876367092,
-0.0525365993,
0.0081091421,
0.0235419516,
-0.0172924791,
-0.0132705923,
0.0209405273,
0.0577608235,
0.0482961126,
-0.0281924661,
-0.0752912238,
-0.0134614399,
0.0166376606,
-0.0178727582,
0.041007679,
0.0235384628,
0.0237921998,
0.019549612,
-0.0120450705,
0.0183277205,
-0.0426442549,
-0.04084846,
-0.0556434095,
0.0183589421,
-0.0389300585,
0.0167916268,
0.0453892797,
-0.0173465125,
0.0856506526,
-0.0206834767,
0.0088937404,
0.0052119386,
-0.081230551,
-0.0020515686,
-0.0122227389,
0.0035236888,
-0.0298681688,
-0.0748568475,
-0.0634903163,
-0.0251114257,
-0.0150134312,
0.0058135092,
-0.0452737957,
0.0102964919,
0.0063630212,
-0.0099350214,
0.0232804082,
-0.0183491148,
0.0035393485,
-0.0498308577,
-0.0378851108,
0.0442009456,
-4.797928966e-33,
0.0091908565,
-0.0434241481,
-0.0002038827,
0.0200177655,
-0.0461888574,
-0.0000343005,
0.0210502613,
0.0312060714,
-0.0223868378,
-0.000384364,
0.010296002,
-0.0261015464,
-0.0074106674,
0.0000465824,
-0.0137128793,
-0.0295177326,
0.0380601361,
-0.0129681658,
0.028291652,
0.0440547839,
0.0518766046,
-0.0093019241,
-0.0138539327,
-0.1077971384,
0.0519167036,
0.0163776595,
-0.0735768005,
-0.0423641466,
0.0063805208,
0.0315022878,
0.0205481667,
0.055913616,
-0.0024623112,
-0.0364941321,
-0.0015219554,
0.095676437,
-0.0004172739,
0.0111504169,
0.000061448,
-0.0171294473,
-0.0043784669,
-0.0481608845,
0.0121858027,
-0.0225124601,
-0.0393507518,
0.0005800341,
0.0108491397,
-0.0273319539,
-0.0099371141,
-0.0525856689,
-0.0721390098,
-0.0178647488,
-0.0509813726,
-0.0220527966,
-0.0379471704,
0.0655199662,
0.003369937,
-0.0430092253,
-0.0360685065,
0.0389479809,
-0.0030645775,
-0.0453702807,
0.0305396579,
-0.0028814904,
0.0140723232,
0.055993855,
0.031215528,
-0.0116478764,
-0.0528174601,
-0.0061349538,
0.010459437,
0.0967066661,
0.0144129843,
0.0499387234,
-0.0293515809,
-0.0645218119,
-0.0380756222,
0.0125012221,
-0.0005067633,
0.0021647054,
-0.0031588359,
-0.0025499654,
-0.0450969264,
-0.0244253371,
-0.0382762477,
-0.0684113503,
-0.0300282594,
-0.0015122758,
0.0061595887,
-0.0246768128,
-0.0375024825,
-0.0008280746,
0.0184155572,
-0.0582674444,
0.058157742,
-0.0079606008,
0.0209370609,
0.0021636393,
-0.0005601158,
0.0097505944,
0.0192928873,
-0.0173562299,
0.056164667,
0.025941601,
0.0834303722,
0.0003883489,
0.0223379023,
0.0451625511,
-0.0026816248,
-0.0071225553,
0.0072994558,
0.0875168368,
0.0140412571,
0.0116940262,
0.0341295861,
0.0213108044,
-0.0156095503,
-0.0521313399,
0.0339180455,
0.0354508758,
0.0017192003,
0.0349923857,
-0.0978368893,
0.0119638024,
-0.0218021814,
0.0268737823,
-0.0285432972,
0.0399092659,
0.0778493881,
-0.0427682735,
0.001063776,
-0.0642451495,
2.49e-7,
-0.0436274111,
-0.0199259128,
0.0300697833,
0.0253701452,
0.0403300785,
0.0790505856,
0.0759939775,
0.0739719942,
0.0163524635,
-0.0349219143,
0.0261313189,
-0.0409677438,
0.05200902,
0.0509335548,
-0.0150449472,
-0.0418550223,
-0.0793051347,
0.0194140635,
-0.0485458821,
0.041902978,
0.0537565947,
0.0139857642,
0.0029896712,
-0.0769659206,
-0.014215949,
-0.0225118827,
-0.0265869591,
-0.0299444217,
0.0170292277,
-0.0443259031,
0.0006931213,
0.0318579301,
0.0132345483,
0.0151656661,
-0.0028782669,
-0.014836981,
0.0008590891,
0.0548037887,
-0.0171116367,
-0.052413784,
-0.0071160677,
-0.0414135754,
0.0099837612,
0.0180476904,
0.0193695705,
0.0354938768,
-0.0163701084,
-0.0242303461,
-0.0164300911,
-0.0142659703,
0.0093102064,
-0.0339796469,
0.022779813,
-0.0183018297,
0.0167778302,
-0.003523933,
0.0416788012,
-0.0292587876,
0.0114796059,
-0.0046963221,
-0.0130779417,
-0.0702003017,
0.0199683271,
0.0489857309,
-0.024677759,
0.0126055125,
0.0328541584,
1.575911693e-34,
-0.0329749212,
-0.0255414657,
-0.0011691095,
-0.0113599366,
0.0126720155,
0.0432515666,
-0.0158966482,
-0.016223181,
0.0372932777,
-0.0051990221,
-0.0466630161
] |
http://arxiv.org/abs/2401.14109v2 | http://arxiv.org/pdf/2401.14109v2 | CompactifAI: Extreme Compression of Large Language Models using Quantum-Inspired Tensor Networks | 2024-01-25T11:45:21 | 2024-05-13T10:48:36 | Andrei Tomut, Saeed S. Jahromi, Abhijoy Sarkar, Uygar Kurt, Sukhbinder Singh, Faysal Ishtiaq, Cesar Muñoz, Prabdeep Singh Bajaj, Ali Elborady, Gianni del Bimbo, Mehrazin Alizadeh, David Montero, Pablo Martin-Ramiro, Muhammad Ibrahim, Oussama Tahiri Alaoui, John Malcolm, Samuel Mugel, Roman Orus | cs.CL | cs.CL, cs.AI, cs.LG, quant-ph | Large Language Models (LLMs) such as ChatGPT and LlaMA are advancing rapidly
in generative Artificial Intelligence (AI), but their immense size poses
significant challenges, such as huge training and inference costs, substantial
energy demands, and limitations for on-site deployment. Traditional compression
methods such as pruning, distillation, and low-rank approximation focus on
reducing the effective number of neurons in the network, while quantization
focuses on reducing the numerical precision of individual weights to reduce the
model size while keeping the number of neurons fixed. While these compression
methods have been relatively successful in practice, there is no compelling
reason to believe that truncating the number of neurons is an optimal strategy.
In this context, this paper introduces CompactifAI, an innovative LLM
compression approach using quantum-inspired Tensor Networks that focuses on the
model's correlation space instead, allowing for a more controlled, refined and
interpretable model compression. Our method is versatile and can be implemented
with - or on top of - other compression techniques. As a benchmark, we
demonstrate that a combination of CompactifAI with quantization allows to
reduce a 93% the memory size of LlaMA 7B, reducing also 70% the number of
parameters, accelerating 50% the training and 25% the inference times of the
model, and just with a small accuracy drop of 2% - 3%, going much beyond of
what is achievable today by other compression techniques. Our methods also
allow to perform a refined layer sensitivity profiling, showing that deeper
layers tend to be more suitable for tensor network compression, which is
compatible with recent observations on the ineffectiveness of those layers for
LLM performance. Our results imply that standard LLMs are, in fact, heavily
overparametrized, and do not need to be large at all. | CompactifAI: Extreme Compression of Large Language Models using Quantum-Inspired Tensor Networks: Large Language Models (LLMs) such as ChatGPT and LlaMA are advancing rapidly
in generative Artificial Intelligence (AI), but their immense size poses
significant challenges, such as huge training and inference costs, substantial
energy demands, and limitations for on-site deployment. Traditional compression
methods such as pruning, distillation, and low-rank approximation focus on
reducing the effective number of neurons in the network, while quantization
focuses on reducing the numerical precision of individual weights to reduce the
model size while keeping the number of neurons fixed. While these compression
methods have been relatively successful in practice, there is no compelling
reason to believe that truncating the number of neurons is an optimal strategy.
In this context, this paper introduces CompactifAI, an innovative LLM
compression approach using quantum-inspired Tensor Networks that focuses on the
model's correlation space instead, allowing for a more controlled, refined and
interpretable model compression. Our method is versatile and can be implemented
with - or on top of - other compression techniques. As a benchmark, we
demonstrate that a combination of CompactifAI with quantization allows to
reduce a 93% the memory size of LlaMA 7B, reducing also 70% the number of
parameters, accelerating 50% the training and 25% the inference times of the
model, and just with a small accuracy drop of 2% - 3%, going much beyond of
what is achievable today by other compression techniques. Our methods also
allow to perform a refined layer sensitivity profiling, showing that deeper
layers tend to be more suitable for tensor network compression, which is
compatible with recent observations on the ineffectiveness of those layers for
LLM performance. Our results imply that standard LLMs are, in fact, heavily
overparametrized, and do not need to be large at all. | [
-0.0240083337,
0.0790277421,
-0.0349865034,
0.0347048938,
-0.0317755453,
0.007182857,
0.0025714068,
0.0458887964,
-0.0185937919,
0.053952381,
-0.0424150378,
-0.0529806539,
0.0072321529,
-0.0116169779,
0.0139816832,
0.0191181339,
0.0219447576,
-0.0027307754,
-0.0163278207,
-0.0596707314,
-0.0265766699,
-0.019600777,
0.0213605128,
0.0190958865,
0.0250853263,
-0.0143405097,
-0.0223346613,
-0.0124480585,
-0.0168857221,
0.0424169824,
-0.0354772396,
0.0500631109,
0.0333810747,
0.0197616387,
0.0000019119,
0.0198506154,
0.0166812744,
0.0127084246,
-0.0255007409,
0.0518246405,
-0.0154889748,
-0.0302650705,
-0.018846415,
0.014071933,
-0.00837164,
0.0028279298,
0.0721138865,
0.0162827391,
0.0491700433,
0.0411023162,
0.0101370318,
-0.0338771604,
0.0360000208,
-0.0272470023,
-0.0268830284,
-0.0386881121,
-0.0064304192,
-0.0011628831,
-0.0356880985,
0.0031878634,
-0.0282743014,
0.009579896,
0.0496052876,
0.0284900162,
-0.0236420967,
0.0278377216,
-0.0401234291,
-0.0038663533,
0.0067080487,
0.0736048371,
-0.0543471463,
-0.0537592284,
-0.0075088637,
0.0185741708,
-0.0409789048,
-0.0513527021,
-0.0198762398,
0.0286921915,
0.0285530165,
0.0444188602,
0.0253930986,
-0.0224649515,
0.0443335064,
-0.0073182802,
-0.0222028233,
-0.0008138271,
0.0555975698,
0.0031515462,
-0.0419149026,
-0.0260805003,
-0.0384606645,
0.0218809936,
-0.0095257517,
-0.003313421,
0.0201970953,
-0.0297367573,
-0.0421469435,
-0.0614398643,
0.0198986344,
-0.0574668534,
0.0527445003,
0.0996007919,
0.0238677878,
0.0209049024,
-0.041453559,
0.0663534701,
-0.0088490266,
-0.015704358,
-0.0409657396,
-0.0182094965,
-0.0153899742,
-0.0010470438,
-0.0301879849,
0.0186011437,
-0.0385899059,
-0.0552678593,
-0.0837058201,
-0.0098646898,
0.0036907161,
-0.0027386004,
0.025800081,
0.0276403408,
0.004133082,
0.0018123705,
0.0187958162,
0.0379929282,
-0.033051949,
0.0129505564,
-0.0580000766,
-0.0480619818,
-0.0426064357,
0.033248648,
0.0122523066,
-0.0042385976,
0.0341959409,
0.0602081381,
0.0009976708,
-0.0624305308,
0.0184106175,
0.04766123,
0.0171379801,
-0.0536368378,
0.051536493,
-0.0367986262,
0.0589639135,
-0.034998104,
0.001912011,
-0.0483620092,
0.0050689606,
0.0030554102,
-0.0622552633,
-0.0018754535,
0.0134110982,
-0.0050787078,
0.0144879809,
0.014313858,
0.0634167641,
0.0297818854,
0.0566461124,
0.0326858982,
-0.0080763036,
-0.0367990099,
-0.0011236522,
0.0039724107,
0.0715060979,
-0.0037698559,
-0.0808660164,
-0.0322595127,
-0.0503721647,
0.0326090753,
0.0589318834,
0.1119051054,
-0.0592365749,
-0.038517803,
0.0479134321,
0.0553146154,
0.0543925241,
0.084374249,
0.0130909383,
-0.0111645302,
0.0045282156,
-0.0144927865,
-0.0421436764,
0.0625780299,
-0.0204445478,
-0.0067188325,
0.0562101305,
0.0048213913,
-0.0560897663,
-0.0369195603,
-0.0060344753,
-0.0173118655,
0.0248078387,
-0.0029147589,
0.0377134494,
0.009822824,
-0.0171412285,
0.0578081198,
0.0161508191,
-0.0487863906,
0.0055869878,
-0.070990108,
0.018508533,
0.0523346961,
0.0253273007,
0.0116186775,
0.0290763304,
0.0547768101,
-0.0742389262,
0.0099987611,
0.0228739232,
0.0251294598,
0.0353057496,
-0.0562587343,
0.0156051647,
-0.0345783643,
0.0080966754,
-0.0237469375,
0.0066564879,
-0.0000862262,
0.0049305232,
-0.0205796082,
0.0096267676,
0.0306211133,
-0.0301328488,
-0.0484457761,
0.0397649445,
-0.0020097075,
0.0401523858,
-0.0196651053,
-0.0261088964,
0.0193809457,
0.0319605768,
-0.0460892767,
-0.0139514729,
-0.0926845595,
-0.0095224343,
-0.0254485551,
0.0188566279,
-0.0041043507,
0.0829326808,
0.0224373396,
-0.058618255,
-0.0452760607,
0.0919972882,
-0.0237837862,
0.0439096726,
-0.0452550389,
-0.0388436839,
0.0179682635,
0.0707687661,
0.0038239602,
0.0403152183,
-0.0461649075,
-0.0024354267,
0.0324576385,
-0.0352873132,
-0.0060419422,
-0.044384405,
-0.0316078067,
0.0074774544,
0.0283132885,
0.0282111298,
0.0090176221,
-0.004459579,
0.0578904226,
-0.0112280082,
-0.1048898324,
0.0090534389,
-0.0284422953,
0.0033075085,
0.0215693712,
0.0114333965,
0.0111630335,
0.0034266354,
-0.0280792005,
-0.0006842823,
-0.0044449926,
-0.0279186033,
-0.0072965389,
-0.0211616643,
-0.0256699361,
-0.0402985066,
-0.0538312308,
0.0096662482,
0.0543687977,
-0.0836361647,
-0.0191519,
-0.0133839864,
-0.009234868,
-0.0329912342,
0.0254571717,
-0.0169041883,
-0.0692764595,
-0.0077003902,
-0.0125042815,
0.0393263549,
0.0689622015,
0.0141010899,
0.0322509669,
0.0921404511,
0.0326934867,
-0.0184457023,
0.0239762738,
-0.0617619976,
-0.0045647961,
-0.0191529579,
-0.0233889818,
-0.04229559,
0.0705839172,
0.0315874331,
-0.001689609,
-0.0678719878,
-0.0172376595,
0.0105365962,
-0.007943566,
-0.0444420651,
0.0049524722,
-0.0028867275,
0.0218949541,
0.0097962627,
-0.0085618822,
0.042930752,
-0.0418430492,
0.0197707396,
-0.0022829312,
-0.0357260592,
-0.0371919647,
0.0571895987,
0.0243116617,
0.0553887002,
-0.053231962,
0.0621133298,
-0.044150766,
0.0003436576,
0.0055542067,
-0.0202972014,
-0.0471489877,
-0.0306948666,
-0.0068030166,
-0.0136160115,
-0.0064898697,
0.0252041463,
0.0248934682,
-0.0076145944,
0.0374104455,
0.0143663669,
-0.0358367898,
0.0386636704,
0.0211368501,
0.0010786222,
0.0524855964,
-0.0102972658,
0.020563161,
0.0103280041,
0.0304365158,
0.038525328,
-0.006483309,
0.0767994598,
0.0041687377,
0.0058217514,
0.0440881699,
0.0179608557,
0.0470195487,
-0.0548029356,
-0.0069839805,
0.0312535875,
0.0216019358,
-0.0470065139,
0.0135242017,
0.0209187437,
-0.0163883436,
-0.00922787,
0.0546655133,
0.00507027,
0.0080820341,
0.0322506838,
0.0613176599,
0.0047464096,
-0.0183987748,
0.0367796123,
-0.0806773305,
-0.0075562922,
0.0109847505,
0.0113670332,
0.0207780786,
-0.0010756988,
-0.0222824924,
-0.0328410119,
0.0012608685,
0.0317387842,
0.0539131276,
-0.0246693045,
-0.0319420993,
0.0017139482,
-0.0232987944,
0.0467490852,
0.0161723476,
0.0164977461,
-0.0213807058,
0.004671881,
0.025489904,
-0.0344093181,
0.0292503517,
0.0442057103,
0.0501899458,
0.0078801941,
0.0349726155,
-0.0019565253,
-0.0714091435,
0.0110554192,
-0.0283312667,
0.0025647443,
0.0108773308,
0.0051598847,
0.0364308506,
-0.011304562,
0.0091565326,
0.0357906446,
-0.0093739191,
-0.0192866474,
-0.0369976237,
0.0306244623,
-0.0177087709,
-0.0529061444,
-0.0034283977,
-0.016708063,
-0.0581685118,
0.0854677856,
0.0257474277,
-0.0397226512,
-0.0528554209,
0.0340428278,
-0.006553262,
-0.0178794339,
-0.0437171236,
-0.0313078612,
0.0210860111,
-0.0534374677,
-0.0570325367,
0.0379212499,
-0.043931026,
-0.0030935649,
-0.0046738838,
-0.0361686833,
0.0058111604,
0.0045441622,
-0.0334870256,
0.009174455,
-0.0144233871,
-0.0177938212,
-0.0388479568,
0.0251542237,
0.0173547734,
-0.0375552773,
-0.0181316808,
-0.02140202,
-0.0916818082,
-0.0299285725,
-0.0174514558,
0.0434646048,
0.0177021511,
-0.0089646988,
0.023967864,
-0.0452908687,
0.0512167327,
0.0381195359,
-0.0416546352,
0.051998686,
0.017402323,
-0.0421520248,
0.016486425,
0.0465252958,
-0.0123962946,
0.0446574986,
0.0020894713,
0.0439475067,
-0.0367326923,
-0.0403464027,
-0.0027722418,
0.0029859901,
0.0413928218,
0.0013139601,
0.0498525724,
-0.0090388311,
0.0185115766,
-0.006119194,
0.0321374089,
-0.0782121941,
-0.0056703142,
0.0299017541,
0.0527039692,
-0.0097657228,
-0.0083534308,
0.1019202024,
-0.0071506598,
0.0264379568,
-0.0276414026,
-0.0372069553,
0.0189067498,
0.011507527,
0.0007977217,
-0.0477784686,
-0.0094132172,
-0.012367052,
0.0347659364,
-0.0336158872,
0.0165624749,
0.0145886084,
0.0125671579,
-0.0636948869,
0.0478914231,
-0.0493720248,
-0.092214182,
-0.0212449022,
0.0606103204,
0.0342141353,
0.0507330783,
-0.0112335607,
-0.0581083633,
-0.0146493353,
0.006361139,
0.0490142219,
0.0358741879,
0.0009789569,
0.0208812915,
0.0170418471,
-0.0247162823,
-0.0302329119,
0.0649132878,
0.0144781033,
0.041508317,
-0.0333919041,
0.0134032201,
-0.0293617453,
-0.0553770587,
-0.0463911034,
0.0567996912,
-0.000130846,
-0.0003068124,
-0.0236407984,
-0.0349579491,
-0.0307202265,
0.0278420448,
-0.0081086103,
-0.0686199889,
0.0372520275,
0.0181015786,
0.0102944979,
-0.0208820794,
0.0041457438,
-0.0464669764,
0.0165485274,
-0.0217647962,
-0.0115751224,
-5.768767914e-33,
-0.0234407857,
0.0458370224,
0.0339662656,
-0.0154097108,
-0.007808825,
0.0205494333,
-0.0465984456,
-0.0311372094,
-0.0373118035,
-0.0332051404,
-0.0119086355,
0.0183129683,
-0.006800679,
0.0002043724,
-0.0232103765,
0.0196742862,
0.0631066784,
-0.002836263,
-0.0105983838,
0.0395373255,
0.0135086756,
-0.0157897472,
0.0377842151,
-0.0018697371,
-0.0684838519,
0.0134823304,
-0.0030480074,
0.0276542809,
-0.0161642227,
0.0149550987,
-0.0944653451,
0.0016948235,
-0.0277375896,
-0.0142896492,
0.033759173,
-0.0250482857,
-0.0658740476,
-0.0319385976,
0.0253812075,
-0.0300071817,
-0.0237293597,
-0.0230279155,
0.034781687,
0.0319241695,
-0.0349648595,
-0.0189881232,
-0.0211471766,
-0.0224694572,
-0.0396207646,
-0.0647626668,
-0.0293722842,
-0.0025607327,
0.0133806346,
-0.0009377099,
0.0182379689,
-0.0048906081,
-0.0447125882,
0.0015908983,
-0.0516982824,
-0.0017006785,
-0.0060325814,
0.0441683643,
0.0296448879,
-0.0076332255,
0.0486559905,
0.0234113019,
0.030348694,
-0.035245236,
-0.0476423874,
0.0066353097,
0.0380421877,
0.024026053,
0.0467633046,
0.0368186235,
0.067208074,
-0.0403964743,
-0.072028704,
-0.0576135702,
0.0763986558,
0.0266017579,
0.0433985069,
0.0155626833,
-0.0156091405,
-0.0401848517,
-0.0116873085,
-0.0064014294,
-0.0205769483,
-0.0194894429,
-0.0164402798,
0.0361558087,
-0.0638839155,
-0.0659195259,
0.0281429011,
-0.0587916896,
0.0416107215,
0.0645759255,
-0.0193712022,
0.0138649317,
0.0429874733,
-0.0083870236,
-0.0472996905,
-0.0535653643,
-0.0060624434,
-0.0429991484,
0.0205629207,
0.0199138783,
0.0412079059,
0.0665835217,
-0.0269461405,
-0.014788528,
-0.0355849676,
-0.0230636727,
0.0279410109,
0.0227385741,
0.0106227715,
0.0354711823,
0.0055661248,
0.0029629094,
0.0131475003,
0.0941318199,
-0.0101019666,
-0.0264144633,
0.0399000384,
0.0547433384,
-0.0345184095,
-0.0177578498,
-0.0722835511,
0.0331245176,
0.0396983847,
-0.098760739,
-0.0453158021,
0.0366706401,
2.652e-7,
0.0129245548,
-0.0017184785,
-0.0214898121,
-0.002349227,
0.0292405877,
0.0043722643,
0.0229406804,
0.0209910069,
0.0031831975,
0.0362157226,
0.0735782012,
0.0502054915,
0.0002379725,
-0.0037133005,
-0.0496660471,
0.0015630936,
-0.0747708902,
-0.0426089652,
-0.0302461609,
-0.0078817047,
0.0138167879,
0.1317220628,
0.0544799268,
0.0030995156,
0.0414532013,
-0.0064314501,
0.0642100498,
-0.0002256712,
0.0306215864,
0.0062979744,
0.0156821311,
0.0130309621,
-0.0376294032,
-0.0254259799,
0.0098051224,
0.0300397873,
-0.0419234671,
0.0558188148,
0.0034636136,
-0.0299080834,
0.0141931372,
-0.0205591395,
-0.0200220831,
-0.0409917198,
0.0600936897,
-0.0106352586,
-0.0241347868,
-0.0575815216,
-0.0087384433,
-0.002945621,
0.0353826359,
-0.0001528511,
0.0333698392,
0.0095664132,
0.0056105023,
-0.0175085254,
0.0209801178,
-0.0651000068,
-0.0066531105,
0.0102010854,
-0.0252526905,
-0.0294933803,
-0.0437852256,
-0.0488387607,
0.0332452543,
-0.0368044525,
-0.0029265061,
2.327147288e-34,
0.0225244593,
-0.0269811656,
0.0111075994,
0.0028053198,
0.0120095899,
-0.0221837834,
0.0013585353,
0.0048280447,
0.0081656529,
-0.0563456416,
-0.0498133861
] |
http://arxiv.org/abs/2401.14067v1 | http://arxiv.org/pdf/2401.14067v1 | Ta'keed: The First Generative Fact-Checking System for Arabic Claims | 2024-01-25T10:43:00 | 2024-01-25T10:43:00 | Saud Althabiti, Mohammad Ammar Alsalka, Eric Atwell | cs.CL | cs.CL, cs.AI | This paper introduces Ta'keed, an explainable Arabic automatic fact-checking
system. While existing research often focuses on classifying claims as "True"
or "False," there is a limited exploration of generating explanations for claim
credibility, particularly in Arabic. Ta'keed addresses this gap by assessing
claim truthfulness based on retrieved snippets, utilizing two main components:
information retrieval and LLM-based claim verification. We compiled the
ArFactEx, a testing gold-labelled dataset with manually justified references,
to evaluate the system. The initial model achieved a promising F1 score of 0.72
in the classification task. Meanwhile, the system's generated explanations are
compared with gold-standard explanations syntactically and semantically. The
study recommends evaluating using semantic similarities, resulting in an
average cosine similarity score of 0.76. Additionally, we explored the impact
of varying snippet quantities on claim classification accuracy, revealing a
potential correlation, with the model using the top seven hits outperforming
others with an F1 score of 0.77. | Ta'keed: The First Generative Fact-Checking System for Arabic Claims: This paper introduces Ta'keed, an explainable Arabic automatic fact-checking
system. While existing research often focuses on classifying claims as "True"
or "False," there is a limited exploration of generating explanations for claim
credibility, particularly in Arabic. Ta'keed addresses this gap by assessing
claim truthfulness based on retrieved snippets, utilizing two main components:
information retrieval and LLM-based claim verification. We compiled the
ArFactEx, a testing gold-labelled dataset with manually justified references,
to evaluate the system. The initial model achieved a promising F1 score of 0.72
in the classification task. Meanwhile, the system's generated explanations are
compared with gold-standard explanations syntactically and semantically. The
study recommends evaluating using semantic similarities, resulting in an
average cosine similarity score of 0.76. Additionally, we explored the impact
of varying snippet quantities on claim classification accuracy, revealing a
potential correlation, with the model using the top seven hits outperforming
others with an F1 score of 0.77. | [
0.0497523881,
0.0026847061,
-0.0121808518,
0.0345326066,
-0.0266070478,
0.0087798554,
0.0372232199,
0.0017844423,
-0.0165839605,
-0.0692257434,
0.0274750534,
-0.0199540518,
0.0465466604,
0.0212496109,
0.0571184717,
0.0129488222,
0.0366317742,
-0.0163045079,
0.0681045055,
-0.0046272692,
0.0428132117,
0.0604326986,
-0.0517393686,
0.0379615165,
-0.0409431458,
-0.0011137166,
-0.0020064504,
-0.0244950429,
0.0208256226,
0.0083478317,
-0.0202716589,
0.0289003588,
-0.0015959482,
0.0457770862,
0.0000019827,
-0.0851472542,
-0.0188558735,
0.0080011059,
-0.0773949176,
0.0097163487,
0.0056856079,
0.0068492577,
-0.0295482352,
0.0010241199,
-0.0027932797,
-0.0099674528,
0.0139714982,
0.0539980941,
0.0625259876,
0.0757444426,
0.0102577573,
-0.0614416264,
-0.0141232694,
0.0236325245,
-0.0005661311,
-0.0222456213,
-0.0010371719,
0.0453974269,
-0.0173747763,
0.0119484961,
0.0173722636,
0.0294953883,
-0.0135858785,
-0.0761059299,
-0.0194907002,
-0.0116259521,
-0.0660170391,
-0.0304687675,
0.0541902296,
-0.0015182956,
0.0288435575,
-0.0747904927,
0.0059614242,
-0.0011837563,
0.0040532509,
0.0579390451,
-0.0127136186,
0.093338646,
-0.0511435121,
0.0111006396,
0.0633944571,
0.0053420486,
0.0298212897,
0.0065502189,
-0.0150398985,
0.034560889,
-0.0268562995,
-0.0327078067,
0.0153497551,
-0.0084155239,
0.022341121,
-0.07163959,
-0.001073468,
0.0134785231,
0.0050139464,
0.0029566803,
-0.0108198198,
0.0076020602,
0.083461076,
-0.0865929499,
0.0060703545,
0.0318535864,
-0.0714731514,
0.0547376089,
-0.0924029499,
0.0202720724,
0.0307434853,
-0.0032762489,
-0.0211434215,
-0.0282357186,
-0.0408470258,
0.0019500325,
-0.0122526102,
0.0416175202,
-0.0509875119,
0.0440967716,
-0.0593171902,
-0.0138377426,
-0.0012218411,
0.0443792194,
-0.0008081439,
-0.010453037,
0.0139925657,
-0.0047309245,
-0.0798567384,
-0.0185145568,
-0.015614097,
0.0001689248,
-0.0323837027,
-0.082349591,
0.0070965365,
0.0244484078,
0.016210245,
0.0266137011,
-0.0002571617,
-0.0130836396,
-0.0020991052,
-0.0400165059,
-0.0703561753,
-0.0270005967,
-0.0104585411,
-0.0372330435,
-0.041721344,
-0.0451305695,
0.0218366086,
0.0090575665,
-0.0304700322,
0.0039217351,
0.0064200722,
0.053953249,
0.0173459463,
0.0353952982,
-0.0140678268,
0.0042367955,
0.04954274,
-0.0005084224,
0.0254603866,
0.0302229654,
0.039085716,
0.0246459283,
0.060252849,
0.0363216586,
0.031397935,
-0.0380702093,
0.0331841893,
0.0076637804,
0.0334683508,
-0.0100588892,
-0.0175878461,
0.0290680528,
-0.0180908963,
0.0054897116,
-0.0144151412,
-0.0311211888,
0.0422924124,
0.0143184885,
0.0015922295,
0.0214831159,
0.0408517905,
0.0100036273,
0.0521733128,
0.0094779059,
0.0048452271,
0.0207432304,
-0.0310425442,
0.0157563798,
0.0446353778,
0.0519570261,
0.00877757,
-0.0899432525,
-0.0358031131,
-0.0044617867,
-0.0329292528,
0.0108072506,
0.0394306965,
0.0003133282,
-0.0383591279,
0.0324179046,
0.0322220065,
-0.0389550813,
-0.0273484103,
-0.078386344,
0.0451117493,
0.0166758373,
0.0209379587,
0.0033476916,
0.0181729998,
-0.0245217197,
0.011083629,
0.0532795563,
0.0462138504,
-0.0030276293,
0.000546966,
-0.0142923733,
-0.0367734693,
-0.0240909439,
0.0392949246,
-0.0274395458,
0.0353974551,
0.01806904,
0.0248446576,
-0.048200842,
-0.0382297374,
0.0288621504,
0.0155867739,
0.0046978872,
-0.0100363092,
0.0109709995,
0.0064364178,
0.0378836505,
-0.0069893994,
0.0109614749,
0.0213936344,
-0.0286579113,
-0.0431753173,
0.0292706341,
0.0249450207,
0.0115104355,
0.0301203858,
-0.0348251238,
0.0052970429,
0.0206033159,
-0.0169674475,
-0.0399015099,
0.0079751927,
-0.0404846817,
0.0286253039,
-0.0398661681,
0.0295923613,
-0.0017058465,
0.0048329062,
-0.0130566824,
0.0843577906,
0.0021485633,
0.0204359163,
-0.0527874716,
-0.0277105495,
0.0831151903,
-0.0082709622,
-0.0013009699,
0.014583353,
-0.0068479278,
-0.0189615786,
0.0394179709,
-0.0484011546,
0.0012081853,
0.0130967349,
-0.0595991425,
-0.0200335197,
-0.009475478,
0.0145770926,
-0.0010994623,
-0.0360368825,
-0.0192632433,
0.0289183762,
-0.0107290205,
-0.0443769321,
0.0304350946,
0.0014070176,
0.074763,
-0.0250444915,
-0.0157081224,
-0.0639951751,
0.0293476433,
0.0082208356,
0.060628064,
0.0831769034,
-0.0628948882,
-0.019413054,
-0.033337526,
-0.0652140453,
-0.00720467,
-0.030128656,
-0.0111611197,
0.0236002449,
0.0217657611,
-0.0390952118,
0.010646956,
0.0367753766,
0.0196750052,
0.0016184417,
0.0540631153,
-0.0444686227,
-0.0126924301,
-0.0065937834,
-0.0172913969,
-0.0058647157,
-0.0761447102,
0.0153699946,
-0.0023125799,
0.0251783803,
0.0027218412,
-0.0218436569,
0.0172411613,
-0.0094663799,
0.0046339286,
0.0201303046,
-0.0254852604,
0.0069420272,
0.138678506,
0.0127809448,
-0.0481555201,
0.0425847992,
-0.0441742875,
0.0278150551,
0.0035118968,
-0.0303413179,
-0.0318231359,
0.0192824192,
-0.0419812389,
0.0320808962,
0.0250891857,
0.015915731,
0.0243042689,
-0.0422258824,
0.0508606583,
-0.0135447877,
0.0375034623,
-0.0304896478,
0.0332733579,
0.000221991,
-0.0040711719,
-0.0280625392,
-0.0407613851,
-0.0264267195,
-0.0152634811,
-0.0261262003,
-0.0005349916,
-0.0794227794,
0.0032879196,
0.0663104877,
0.0248162765,
0.0055539575,
-0.0024723946,
0.0179368369,
-0.0225866921,
0.0060553337,
0.0012861222,
0.0732793808,
0.0065572346,
-0.0234300718,
0.0454154275,
-0.0827360228,
0.0335760191,
-0.0342466347,
-0.0735767335,
-0.0126436399,
-0.0652227253,
-0.0371695459,
-0.0593515113,
-0.0161067322,
-0.0686181039,
0.0219029225,
0.074639298,
0.0172920823,
0.0414224491,
0.000802607,
-0.0186840091,
-0.0437145606,
0.0150961084,
-0.0067850305,
-0.1211846471,
0.0116262902,
-0.0143689765,
-0.0065255379,
-0.0317051038,
0.02046391,
0.0186246336,
-0.0918278247,
0.0636285469,
-0.0348923691,
0.0122713065,
-0.0675402954,
-0.0567476004,
-0.0071663284,
0.0261866022,
-0.0495241694,
0.014531563,
0.0080017215,
0.0228119213,
0.0297618341,
-0.0183158647,
-0.0346611179,
0.0383928493,
-0.0102381837,
0.0010313814,
0.035519626,
0.0288543701,
-0.0153016401,
-0.071780242,
-0.0195502006,
-0.0079038655,
-0.0378391296,
0.0001698315,
0.0354426056,
-0.0064246296,
0.0299812183,
-0.0284737069,
-0.0320952833,
0.0137288421,
-0.0029205487,
-0.0030704313,
0.032063216,
-0.0270603485,
0.0004325125,
0.0021395523,
0.0278860107,
-0.0330807045,
0.0802842602,
0.0073667215,
0.0572023615,
-0.0180145539,
-0.0250695664,
-0.0411475077,
-0.0600345284,
-0.0580364838,
0.0281357486,
0.0213564485,
0.0162163768,
-0.1002942249,
-0.0062816832,
-0.0284660161,
0.010186404,
-0.0207322612,
0.0193137806,
-0.0047461959,
0.0079222713,
-0.0432806872,
-0.0212008208,
-0.0114589762,
0.0055554491,
-0.0169214178,
-0.0214953478,
0.033060316,
-0.0961591229,
0.0223117657,
-0.0029814143,
0.0017094041,
-0.0141097642,
-0.0002506864,
-0.0146796471,
0.0230622906,
0.0190840475,
0.0093080802,
-0.014806916,
0.0368031673,
0.0094354944,
-0.0125745842,
0.0341340899,
0.0228981283,
0.0117675923,
-0.0069057909,
0.0241491757,
-0.0133509906,
0.011155556,
-0.0309136845,
-0.0388500541,
-0.0143536683,
-0.0619029105,
0.0303473659,
0.0120407417,
0.0237623006,
-0.0344377458,
-0.0290846191,
-0.03961033,
0.0165266208,
-0.0078975745,
-0.0008547515,
0.0027198948,
0.0454148985,
-0.0295668356,
0.0019852617,
0.0072859023,
-0.044196859,
-0.0748732239,
0.028825786,
-0.0251844991,
0.0207006428,
-0.0543154255,
0.0362972133,
-0.0443730466,
0.008577412,
-0.0082722576,
-0.0405793376,
-0.0303625725,
0.039164111,
-0.0092370175,
0.0209023654,
0.0016623721,
0.0082760043,
0.0460571907,
0.0036594479,
-0.0264003649,
-0.02530366,
-0.0266230386,
0.0066791526,
0.0382445082,
-0.0329586826,
0.0615421683,
0.0031914434,
-0.0007634122,
0.0312254988,
0.041409377,
0.0047111013,
-0.0326174721,
-0.00678002,
0.0060947989,
-0.0233512484,
0.0190841556,
0.0424037129,
0.0052242531,
0.0167267974,
0.0030463729,
-0.0559795238,
-0.0118142525,
-0.0775675103,
-0.0608965009,
0.018185949,
-0.0112458412,
-0.027508758,
-0.0880166143,
-0.0417215154,
0.0141940992,
0.0543504804,
0.0340091363,
0.0045118257,
0.0163317248,
0.1010095105,
0.0409436896,
-0.032356482,
-0.0816996992,
0.0232845377,
0.0011263447,
0.0041437759,
0.021428166,
-6.034507514e-33,
-0.0102024674,
-0.0720941648,
0.0124575756,
0.0131289307,
0.0327394083,
0.0155992722,
-0.0511826575,
-0.0193209257,
-0.0361974649,
0.0610029772,
0.0137478001,
0.0267411862,
0.031587556,
0.000022499,
-0.0055584549,
0.0022259208,
0.0291564204,
0.0328136869,
-0.0367425717,
0.0050023948,
0.0513843074,
-0.0076997448,
0.0945358574,
-0.0974901095,
0.0767428875,
-0.0146674402,
-0.0323808305,
0.009228006,
-0.0314884409,
0.0978626013,
0.015752431,
0.0370540023,
0.0137198288,
-0.0308792889,
0.0398628488,
-0.041222591,
-0.056015484,
0.0086968048,
0.0735231414,
-0.0099089788,
0.003675069,
0.0268681031,
-0.0130998353,
-0.0050160848,
-0.0405113101,
0.0384091586,
0.0465081707,
-0.0204517785,
0.0115968185,
-0.0435397737,
-0.0700301901,
0.0543178543,
0.0123174451,
0.0150402747,
0.0002310592,
0.049980443,
0.0191347022,
0.0218560696,
-0.0600758158,
0.0088184997,
-0.0241561439,
0.0093335109,
0.0449514128,
-0.049035944,
-0.0730732977,
0.0525810495,
-0.0041996543,
-0.0270940717,
-0.021358341,
0.0335165597,
-0.0372710451,
-0.0072352733,
0.0546840578,
0.1101324409,
0.0079515474,
-0.0171181317,
-0.1010154933,
0.0336833522,
0.0025360633,
0.0172928628,
0.0020108023,
-0.0333667062,
-0.068767108,
0.0003387069,
-0.0139176687,
0.008468058,
-0.0168610383,
0.028274361,
-0.0029682859,
0.0105809066,
-0.0268366188,
0.0615631342,
-0.027367318,
0.0064374059,
0.0598871373,
-0.0026150583,
-0.021414537,
-0.0197637249,
0.000723877,
-0.0181489754,
-0.0412593111,
0.0360858962,
-0.011735159,
0.0524200983,
0.0289880726,
0.0063586077,
0.0369969867,
0.0082704667,
-0.063007012,
0.0447790921,
-0.021548396,
0.0285779852,
-0.0326171555,
-0.0276652202,
-0.0053653,
0.0159921814,
-0.0201694127,
0.0168202221,
0.0103595071,
-0.0247913431,
-0.0393779837,
-0.0207126718,
-0.0255977511,
-0.0132198706,
-0.0074990699,
0.0135100577,
-0.0194933731,
0.0356551819,
0.0184195861,
-0.0776465461,
0.022295922,
0.0212323181,
2.74e-7,
0.0426346362,
0.0278471317,
0.0114775533,
0.0112985233,
-0.0155581813,
0.0229807887,
0.0546004251,
-0.0205424838,
0.0083665997,
-0.0431346297,
-0.0348973647,
0.0140174935,
0.0254815221,
0.0518961288,
-0.0843992159,
-0.014991059,
-0.0185924955,
-0.0513111055,
-0.0441791229,
-0.0122964848,
0.0394464061,
0.0466952212,
-0.0135027049,
-0.0328087769,
-0.0368365198,
-0.0350033715,
0.0452991389,
0.0231023021,
-0.0314518437,
0.0188430678,
0.10681995,
-0.0272281952,
0.0416470692,
0.0132025126,
-0.0480142683,
0.026449956,
-0.0064098584,
0.0613638312,
0.0311153997,
0.0056408751,
-0.0275517441,
-0.0483131185,
0.0278343149,
-0.0715277493,
0.0280369204,
0.0555894785,
0.0159830749,
0.0051717851,
-0.0257933382,
0.0358434282,
0.008806332,
-0.0179862697,
-0.0091189332,
0.0336570069,
-0.0023194826,
-0.0071706218,
0.0067006433,
-0.035227593,
-0.0112445271,
-0.018834129,
-0.007734865,
0.0201407894,
0.0302687269,
-0.0428870432,
0.0023649554,
-0.0422575884,
-0.0099585922,
2.29334402e-34,
0.0260992367,
0.0219753105,
0.0016018834,
-0.0340715684,
0.0913719535,
0.0504009612,
-0.0196742248,
0.000807375,
0.0106346626,
0.0357055776,
-0.0089823548
] |
http://arxiv.org/abs/2401.14043v2 | http://arxiv.org/pdf/2401.14043v2 | Towards Goal-oriented Prompt Engineering for Large Language Models: A Survey | 2024-01-25T09:47:55 | 2024-06-18T02:58:37 | Haochen Li, Jonathan Leung, Zhiqi Shen | cs.CL | cs.CL, cs.AI | Large Language Models (LLMs) have shown prominent performance in various
downstream tasks and prompt engineering plays a pivotal role in optimizing
LLMs' performance. This paper, not only as an overview of current prompt
engineering methods, but also aims to highlight the limitation of designing
prompts based on an anthropomorphic assumption that expects LLMs to think like
humans. From our review of 36 representative studies, we demonstrate that a
goal-oriented prompt formulation, which guides LLMs to follow established human
logical thinking, significantly improves the performance of LLMs. Furthermore,
We introduce a novel taxonomy that categorizes goal-oriented prompting methods
into five interconnected stages and we demonstrate the broad applicability of
our framework. With four future directions proposed, we hope to further
emphasize the power and potential of goal-oriented prompt engineering in all
fields. | Towards Goal-oriented Prompt Engineering for Large Language Models: A Survey: Large Language Models (LLMs) have shown prominent performance in various
downstream tasks and prompt engineering plays a pivotal role in optimizing
LLMs' performance. This paper, not only as an overview of current prompt
engineering methods, but also aims to highlight the limitation of designing
prompts based on an anthropomorphic assumption that expects LLMs to think like
humans. From our review of 36 representative studies, we demonstrate that a
goal-oriented prompt formulation, which guides LLMs to follow established human
logical thinking, significantly improves the performance of LLMs. Furthermore,
We introduce a novel taxonomy that categorizes goal-oriented prompting methods
into five interconnected stages and we demonstrate the broad applicability of
our framework. With four future directions proposed, we hope to further
emphasize the power and potential of goal-oriented prompt engineering in all
fields. | [
0.0546696335,
0.0491163209,
-0.0357363001,
0.0223780312,
-0.051054664,
-0.0112195211,
0.0268753637,
-0.0195182655,
0.0269433558,
-0.0035050528,
-0.0128228134,
-0.0438776463,
-0.008828586,
0.0521465838,
0.0214797389,
-0.0785307586,
0.0757093057,
-0.0694352388,
0.0125291869,
-0.0095230304,
-0.0173027907,
0.0025178702,
0.0048858547,
0.0403031148,
0.0086178184,
-0.0575837381,
-0.0440825745,
0.0102025522,
-0.0095494278,
-0.0407584161,
-0.0227175057,
0.0377137028,
-0.0084428554,
0.0362845883,
0.0000022408,
-0.0936666355,
0.0057529206,
-0.0178081617,
-0.0559907332,
-0.0094985412,
0.0361730121,
-0.0027853204,
-0.0458806418,
0.0115758907,
-0.0104008531,
-0.0130144404,
0.0367188007,
0.0212353636,
0.0549557172,
0.0462802686,
0.0030089617,
0.0016509409,
-0.0068578757,
-0.0080346018,
-0.0051045478,
-0.0356611051,
0.0025156604,
-0.0206155591,
0.0724964887,
0.0137287658,
0.0030402173,
0.0515338592,
0.0341176726,
-0.0361786522,
-0.0116966618,
0.0161320735,
0.0116662532,
-0.0563563481,
-0.0346585959,
0.0434477739,
0.0026476192,
0.0035874401,
-0.0061175185,
-0.0175243448,
0.0445776656,
0.0216703583,
-0.0321929231,
0.055867888,
-0.0213988274,
-0.0166020729,
0.0278947372,
0.0117363976,
0.0041473815,
0.0124315964,
-0.0335329734,
0.0908675566,
0.0313337147,
-0.0231305975,
0.0097813429,
-0.025031399,
-0.0028408456,
-0.0933129936,
-0.0067719547,
0.0442608036,
-0.1083334461,
0.0214878432,
-0.0353845544,
0.0021415823,
0.0178407598,
-0.0088196183,
0.0377898626,
0.0548595153,
-0.0075317197,
0.0357420854,
0.0211586114,
0.0470575392,
-0.0124292187,
0.0020974244,
-0.0400878564,
0.0120054903,
-0.0168442614,
-0.0111441454,
-0.0685290843,
-0.0003836982,
-0.0103531936,
0.0077044633,
-0.0124771213,
0.0889725164,
-0.0018469397,
0.0554636233,
-0.0105197215,
-0.0191147029,
-0.024283519,
-0.0413878672,
-0.0220391862,
0.0062147155,
-0.046568092,
-0.0145547744,
0.0323876478,
-0.0444108211,
0.0038148605,
0.0105622103,
-0.0236236081,
-0.0163708851,
-0.034182433,
0.0682297647,
0.009873827,
-0.0068828482,
-0.0160853341,
-0.0525435694,
0.0194916055,
-0.0565969571,
-0.0037403107,
-0.0530720316,
0.0039404281,
0.0149623351,
-0.0096830502,
-0.0294244178,
-0.0166881979,
0.0220602937,
0.0351041369,
0.063649632,
-0.0306709372,
0.0251876041,
0.0264629051,
0.008263058,
-0.0225851107,
0.0680098385,
0.0091239847,
0.0419921353,
0.0265538022,
0.0046728798,
-0.0136484215,
0.0690174922,
-0.0186610501,
-0.007340698,
-0.0163527541,
0.0284661334,
0.0376287811,
0.0433809571,
-0.0130151855,
0.0420658439,
0.0019691847,
0.0439413488,
0.0459980145,
0.0400309451,
-0.0049141222,
-0.0216045715,
0.059555389,
0.0189333335,
-0.0208954699,
0.0203713216,
-0.0633624196,
0.0146724191,
-0.0233579297,
0.0303143244,
-0.0187414289,
0.0417639911,
-0.0278549325,
-0.0228709038,
-0.0031027854,
0.0092359437,
-0.0241262317,
0.0332293771,
0.0321694314,
0.0238687079,
-0.0124807386,
0.0140869254,
-0.017961299,
-0.0230777245,
0.0049622068,
-0.0716335103,
0.0571064912,
0.0774235427,
0.0477521904,
-0.0356806405,
-0.0414381884,
-0.0003286828,
0.033695139,
0.0494607091,
0.0839100182,
-0.0418589674,
-0.0424827933,
-0.0171514284,
-0.0256378073,
0.0082266135,
-0.00921728,
0.0093777329,
-0.0043594618,
-0.0308570527,
-0.0102896094,
0.0172489267,
-0.0121966861,
0.0261438768,
-0.0223794337,
0.0093113231,
-0.0182515662,
0.0389194563,
0.0134179201,
0.0215917677,
0.0173425991,
0.0221395772,
0.0186163429,
-0.0086901831,
-0.0004604099,
0.0027733045,
-0.045172561,
-0.0171581432,
0.0183655694,
0.014865797,
-0.0496332571,
-0.004232523,
-0.0199171845,
-0.0413309485,
-0.0027720104,
-0.017292114,
0.0186536796,
-0.0245584417,
0.0501108468,
-0.0220499169,
-0.0489372797,
-0.0474907793,
0.0486746877,
0.0286895577,
-0.0008450935,
0.0820804685,
0.0303065591,
0.0167826228,
-0.0534548201,
-0.0459292196,
0.0397455506,
-0.0019836535,
-0.0036953173,
0.0463392586,
0.0389514826,
-0.0474828333,
-0.0355399288,
-0.0659323558,
-0.0016170016,
-0.0217842609,
0.011753194,
0.0071859695,
0.0353593305,
-0.0311949495,
-0.0249412972,
-0.024640495,
-0.0065507707,
0.0354941823,
0.031457033,
0.0337982625,
0.018749373,
-0.0528296381,
0.0016221065,
-0.0064245737,
-0.0241367649,
0.0062584835,
0.0067265523,
-0.0305899736,
-0.1001911834,
0.0398664884,
-0.004501903,
0.0294116084,
0.0087937545,
0.0218888205,
0.0459997579,
-0.0114380624,
0.0899590552,
0.0377833508,
-0.0075540505,
0.0043627541,
0.057013303,
-0.0224687308,
-0.0492465161,
0.0161631014,
-0.0162391812,
-0.017976705,
-0.0007669745,
-0.034592066,
0.0359860137,
0.0434292965,
-0.0134187108,
0.0066535254,
-0.0512332581,
0.0461686142,
-0.0338828787,
-0.0148084704,
-0.0560991988,
-0.0456682071,
-0.0276952423,
0.0963083059,
0.0002044738,
-0.0387212895,
0.1093770862,
-0.014497431,
0.0017867528,
-0.0240595546,
0.0244412832,
-0.0321271233,
0.0717450231,
-0.0234584622,
-0.0023006042,
0.068757005,
-0.0461006872,
-0.050161466,
0.0236991141,
-0.0113624688,
-0.000987157,
0.0532602035,
-0.062695384,
-0.0043026479,
-0.0176839307,
-0.0429011621,
-0.0692551732,
-0.0423832275,
-0.0306048989,
0.0056133084,
-0.0174997561,
0.013435591,
-0.0151311439,
0.0012498533,
0.0203331113,
0.0213473905,
-0.0175493248,
-0.0173601713,
-0.0378278457,
0.0003035858,
0.0451250561,
0.0354678668,
0.0087294532,
-0.0283297151,
0.0670420974,
0.0088025536,
0.0228462797,
-0.0101834405,
0.0011366308,
-0.0094699832,
0.0659352317,
-0.0387350768,
-0.0297525283,
0.0678530857,
0.0635529906,
-0.0489638783,
-0.030654218,
0.1305243075,
0.0430942848,
-0.0241624936,
0.0067967018,
-0.0535360165,
0.0537182279,
0.0359018631,
0.0525026061,
-0.064208582,
-0.0240415856,
0.0319355056,
-0.0063205143,
-0.025212327,
-0.0046378411,
-0.0398096703,
-0.1041732281,
0.0512124188,
-0.0128177963,
-0.0201002266,
0.0098663922,
-0.0133437142,
0.019239042,
0.03572632,
0.0150124896,
0.0421101376,
-0.0249498151,
-0.028775597,
-0.0072909808,
-0.0198101811,
-0.0426817313,
0.0090671768,
-0.0439922959,
0.0661021844,
0.0769214258,
-0.0122838235,
0.0114126904,
-0.0490246192,
0.0293656792,
0.0209065359,
0.0261163227,
-0.0050117136,
-0.0165356938,
0.0090534799,
-0.0106390826,
-0.0342122801,
0.0052695014,
-0.0219715405,
-0.0495291539,
0.0414084271,
0.1115143895,
-0.0512639321,
0.0105427569,
-0.0257143695,
-0.0037832621,
0.0201455839,
0.0727563724,
0.0292811822,
-0.0786231682,
-0.0224642158,
-0.0403473675,
-0.0325219855,
-0.0258230679,
-0.0244617872,
0.0179700721,
-0.0204354096,
-0.0259967465,
-0.0362005569,
-0.0602260269,
0.0429693237,
0.0289475992,
0.0417038985,
-0.0270023476,
0.0294171888,
-0.0050632753,
0.0031593675,
0.035370633,
-0.0446359962,
0.044520542,
-0.0777222216,
0.0118335476,
-0.0284936875,
-0.0524745695,
0.0036228506,
0.0168988351,
-0.0101911388,
-0.0542578958,
-0.0081203515,
0.0362723134,
0.0317536555,
0.0097357845,
-0.0237505324,
-0.0597580746,
-0.0012644298,
0.0147774825,
-0.0036474478,
-0.019572027,
0.0123539409,
0.0086379945,
-0.0045109298,
0.0456518568,
-0.0255051628,
0.0140741151,
-0.0037734965,
0.0309995003,
0.010729868,
-0.0632091612,
0.0415886231,
0.0523282886,
0.0269380137,
-0.0389380231,
-0.0171947628,
-0.0625065789,
-0.0104851881,
-0.0465411842,
-0.0527317636,
-0.0229316019,
0.0142554166,
-0.0163484868,
-0.0266693309,
0.0317907371,
-0.04224189,
-0.0554518215,
0.0318452492,
-0.0561940335,
0.0118509708,
-0.0249350853,
0.0447488688,
0.0450823642,
-0.0039848043,
-0.0249836575,
0.034109246,
-0.0092938906,
0.0295377281,
-0.0168889593,
0.0957827941,
0.0304319523,
-0.009109729,
-0.0225545503,
0.0319206193,
-0.0416483246,
-0.0825005844,
0.0399802402,
0.0473588593,
0.030252872,
0.0273953117,
-0.0000831695,
-0.0372529104,
-0.001146113,
0.0041632797,
0.0043754363,
0.0643510967,
-0.0742207393,
0.0214667339,
0.0313260779,
-0.0138961188,
-0.0610113516,
0.0140775507,
-0.0571489483,
0.004428783,
-0.0175740216,
-0.0023558883,
-0.0522164367,
-0.0203367434,
0.0047399099,
-0.0017594496,
-0.0575033501,
0.000561243,
-0.1068879068,
-0.0428816825,
-0.006175864,
0.058586549,
-0.0212297607,
-0.0175144766,
0.0527451448,
0.0447159931,
0.0042871679,
-0.0306665078,
-0.0287818313,
0.0098564262,
-0.0285542458,
-0.0335325301,
-0.0319386572,
-5.758497767e-33,
-0.0348186903,
-0.014361674,
0.0521365032,
-0.0169564933,
-0.0105265109,
-0.015084316,
-0.0588595532,
0.0235908311,
-0.0050942278,
-0.0070093856,
-0.037424732,
-0.0312033016,
0.0179160945,
-0.0103166234,
0.0246404577,
0.0204182267,
0.053481672,
0.007927089,
-0.0544601902,
0.0046518128,
0.0112654651,
-0.0095857428,
-0.0089478027,
-0.0079661878,
0.0592526868,
-0.0037378771,
-0.0227701273,
0.050379809,
-0.0112452796,
0.0377219617,
-0.0234469008,
0.044163622,
0.0083853807,
-0.0605544709,
-0.0421290435,
0.0952481031,
-0.0256238598,
-0.0474462211,
0.055166766,
0.0130124921,
-0.0425566472,
-0.0347699858,
-0.0259572808,
-0.0374392718,
-0.0231883004,
0.0378079489,
0.0070918757,
-0.0058292621,
0.0041820314,
0.0078976154,
-0.0897949561,
0.0063240603,
0.0113959908,
-0.0324443392,
0.0256489478,
0.0595770068,
-0.0004377453,
-0.0172367077,
0.0021198248,
0.0326449163,
0.0299184956,
0.0180120487,
0.0741198659,
0.0355209596,
-0.0195638742,
0.0405236222,
0.0296858754,
0.0430993438,
-0.0005847044,
0.011897401,
0.0160785634,
0.0108981775,
0.0295004062,
0.0115528563,
-0.0125245946,
-0.0083167208,
-0.0268057212,
0.0282417331,
0.0214566644,
0.008221251,
0.0257508866,
0.0034947293,
-0.044740092,
-0.0244833231,
-0.0244695079,
0.0197656062,
-0.035252627,
-0.0052578999,
0.0319742039,
-0.0316689201,
-0.0220646784,
-0.021481527,
0.0336947888,
-0.0134935984,
-0.0033166159,
0.0012507647,
-0.0112253986,
0.0395518988,
-0.0087200254,
0.0513312221,
-0.0102638286,
-0.0220990628,
0.017476568,
-0.0135263531,
0.0214022305,
0.0360493213,
-0.0048574619,
0.0340080447,
-0.0540034696,
0.0040352396,
-0.054446511,
0.0457852334,
-0.0054715029,
-0.0060039954,
0.0474147648,
-0.0040028119,
-0.0011894602,
0.0456400812,
0.0151075376,
-0.0959583819,
-0.0466966927,
0.0608190298,
0.0039519346,
0.005280843,
-0.0166554954,
-0.0096311746,
-0.0347574316,
0.0648188889,
0.0315782987,
-0.0206380989,
-0.0189141054,
0.0146651482,
2.869e-7,
0.0116837062,
0.0131239481,
0.0139416317,
0.0962942615,
-0.0005335859,
0.0310826451,
-0.0010216244,
0.0014229097,
0.0257892255,
-0.009871806,
0.0015064345,
0.0013098659,
0.0216343347,
0.0244531222,
-0.128455624,
-0.004752845,
-0.0148064755,
0.013842429,
-0.0777368769,
0.0340089202,
0.0863733515,
0.0362576358,
-0.0010180614,
-0.0165304858,
-0.018305419,
-0.0068012718,
0.017298013,
-0.0203360058,
-0.0034473448,
0.0362697057,
0.012396059,
-0.0302923266,
0.0065879901,
0.0416490659,
0.027302254,
-0.0058155372,
-0.0198956206,
0.0572776906,
0.0176328234,
0.0143770333,
-0.0075652315,
-0.0206177738,
-0.0049335063,
-0.0708141476,
0.0509614423,
-0.058428742,
-0.0169453882,
-0.0359543115,
0.0146432081,
-0.0341974236,
-0.0243692491,
-0.026627183,
-0.0098859193,
0.0233287849,
0.0199143868,
-0.0152917504,
-0.0063007642,
-0.0219010003,
-0.0055846581,
-0.0007920854,
-0.0319947787,
-0.0630161837,
-0.0461244062,
0.014324245,
0.0109570324,
-0.0803968459,
0.0005798053,
2.796845673e-34,
0.0272699073,
-0.0076900385,
0.0291148853,
0.0115140965,
0.0311567895,
0.0151956901,
-0.0177873168,
0.0330620669,
0.0497777723,
-0.0727957487,
-0.0644199997
] |
http://arxiv.org/abs/2401.14040v3 | http://arxiv.org/pdf/2401.14040v3 | (Chat)GPT v BERT: Dawn of Justice for Semantic Change Detection | 2024-01-25T09:36:58 | 2024-04-29T07:07:45 | Francesco Periti, Haim Dubossarsky, Nina Tahmasebi | cs.CL | cs.CL | In the universe of Natural Language Processing, Transformer-based language
models like BERT and (Chat)GPT have emerged as lexical superheroes with great
power to solve open research problems. In this paper, we specifically focus on
the temporal problem of semantic change, and evaluate their ability to solve
two diachronic extensions of the Word-in-Context (WiC) task: TempoWiC and
HistoWiC. In particular, we investigate the potential of a novel, off-the-shelf
technology like ChatGPT (and GPT) 3.5 compared to BERT, which represents a
family of models that currently stand as the state-of-the-art for modeling
semantic change. Our experiments represent the first attempt to assess the use
of (Chat)GPT for studying semantic change. Our results indicate that ChatGPT
performs significantly worse than the foundational GPT version. Furthermore,
our results demonstrate that (Chat)GPT achieves slightly lower performance than
BERT in detecting long-term changes but performs significantly worse in
detecting short-term changes. | (Chat)GPT v BERT: Dawn of Justice for Semantic Change Detection: In the universe of Natural Language Processing, Transformer-based language
models like BERT and (Chat)GPT have emerged as lexical superheroes with great
power to solve open research problems. In this paper, we specifically focus on
the temporal problem of semantic change, and evaluate their ability to solve
two diachronic extensions of the Word-in-Context (WiC) task: TempoWiC and
HistoWiC. In particular, we investigate the potential of a novel, off-the-shelf
technology like ChatGPT (and GPT) 3.5 compared to BERT, which represents a
family of models that currently stand as the state-of-the-art for modeling
semantic change. Our experiments represent the first attempt to assess the use
of (Chat)GPT for studying semantic change. Our results indicate that ChatGPT
performs significantly worse than the foundational GPT version. Furthermore,
our results demonstrate that (Chat)GPT achieves slightly lower performance than
BERT in detecting long-term changes but performs significantly worse in
detecting short-term changes. | [
0.0381157734,
0.0754017532,
-0.0292938817,
0.0104024028,
-0.0819797218,
-0.0301207844,
-0.0253682807,
0.0411034152,
-0.0342542864,
-0.0444410145,
0.0329699367,
-0.0246639922,
0.0197537486,
-0.0029930151,
-0.0312657319,
-0.0457978137,
0.1056135669,
0.0022693444,
-0.0288988221,
0.0523816012,
-0.0065080998,
-0.0192276891,
-0.0107807852,
0.0219458845,
-0.0165265016,
-0.0352703519,
-0.0409501791,
-0.0014643186,
0.0309050307,
-0.028624922,
0.0190915186,
0.0659803748,
0.0217541866,
0.0902409703,
0.0000019694,
-0.0624382161,
-0.0567181148,
-0.0038303379,
0.0122661646,
0.0584891476,
0.0377318822,
0.0089787254,
-0.0341031253,
0.0044633471,
-0.0433452092,
-0.0041450956,
0.0139871668,
0.0155227687,
0.0591752529,
0.0726530477,
-0.024774963,
0.0002631508,
0.0371060669,
-0.0019399973,
0.0226762798,
-0.0151322326,
0.0416613668,
-0.0218318384,
-0.0074358508,
-0.0172699038,
-0.0416436531,
0.012899261,
-0.0129589299,
-0.003854767,
-0.0173150748,
0.035854239,
-0.0638120696,
0.0409850702,
-0.033238817,
-0.0079264082,
-0.0110764485,
-0.0429976694,
0.0044824718,
0.0242284406,
-0.0001050512,
0.0431436487,
-0.0142223593,
-0.0302849934,
-0.0019530706,
-0.001235889,
0.0177215524,
0.0033552984,
-0.0115551706,
0.054411903,
-0.032029938,
0.0437319912,
0.066513449,
-0.0534860566,
-0.0675248876,
-0.0074885557,
-0.0065505514,
-0.0564921945,
0.0175168607,
0.0094414866,
-0.0009804241,
0.0180791896,
-0.0183757767,
-0.0191300791,
0.0101415599,
-0.0686871484,
-0.0497309789,
0.0339524895,
-0.0600653812,
0.0542096645,
-0.0265427716,
0.0660479069,
-0.0105460621,
-0.0291364361,
-0.0197234079,
0.0042287679,
-0.0875182822,
-0.0205367431,
-0.0866018087,
0.0184396561,
-0.0394069329,
-0.0411731713,
-0.0178755503,
0.05944556,
0.0305879321,
-0.0324428491,
0.0029949436,
-0.0234904978,
-0.0510197617,
0.030665949,
-0.0117928451,
0.0387366824,
-0.0038366811,
-0.0105233155,
-0.04239171,
-0.0661782473,
-0.0198380314,
0.0224207565,
-0.0027019118,
-0.0164156687,
-0.0257595088,
0.0616566278,
-0.0124199754,
-0.0573794059,
-0.0355883427,
0.0096394299,
-0.0747710764,
-0.0040044044,
-0.0172366947,
-0.0507683083,
-0.0147651089,
-0.0004601448,
0.0146986907,
0.015029477,
-0.0103153903,
-0.0000479327,
-0.0162181575,
0.0822103322,
-0.0077948179,
0.0173834413,
0.0294636376,
-0.0087556718,
0.0306332503,
0.0632947981,
0.0174543485,
-0.0357624702,
0.0668881461,
0.0083688386,
-0.0288454704,
0.0525863841,
-0.0436983407,
0.0567898527,
-0.0347311497,
-0.006921804,
0.0277023911,
0.048695527,
-0.0107712764,
0.0496609472,
-0.0141698988,
-0.0426982567,
0.0591346882,
0.0442884639,
0.0398962833,
0.0177618116,
0.0189354979,
0.0158880632,
-0.0110508837,
0.0248266235,
0.0003602961,
-0.0392782651,
-0.0262781736,
-0.0324596725,
-0.0509447232,
-0.0315790139,
-0.0122606549,
0.0123951873,
-0.0258198846,
0.0062464844,
0.0167000517,
-0.0840855688,
0.0422615893,
0.0262305662,
-0.028530458,
0.086855486,
0.0254848488,
-0.0146127585,
-0.0091651939,
-0.0643777698,
0.0099330852,
0.0755545497,
0.0935693085,
-0.0453363024,
-0.0207325183,
-0.0243237205,
-0.0419747643,
0.0508042201,
0.0227615964,
0.0053283358,
-0.01324898,
-0.0107420972,
-0.0113123246,
0.0006243534,
0.0211058091,
0.0090922154,
0.0562633052,
0.0410048701,
-0.0013342717,
-0.0400685593,
0.0377749875,
0.0498231798,
0.0085021807,
0.0107042901,
0.0024261617,
-0.0252939239,
0.0295064468,
-0.0120889405,
-0.0169719793,
-0.0071344618,
-0.0141777936,
0.0000265809,
-0.0495542288,
0.0755362362,
-0.0344754457,
-0.0410480686,
0.0437221415,
0.00002343,
0.0315208323,
0.0231110081,
-0.0304732379,
-0.0177672431,
0.0223219004,
-0.059811417,
0.0687907562,
-0.0427970178,
0.0359888673,
-0.0305966958,
0.0207501538,
-0.0499531031,
0.0551650226,
-0.0054979799,
-0.0335812271,
0.0018145296,
0.0561803989,
0.0353572927,
-0.0055468371,
-0.0268086866,
0.0517983846,
0.0126377679,
-0.0322597176,
0.0401686393,
-0.0016773036,
-0.0432792343,
-0.0330709331,
-0.1012889966,
0.0026428145,
-0.0387577116,
0.0225564223,
0.0481264517,
0.0260178074,
-0.0189260654,
0.0053791618,
0.0348429456,
0.0017878285,
-0.0155868186,
0.0477491505,
0.059869539,
-0.0064815101,
-0.0060079517,
-0.0602126643,
-0.0252933092,
-0.0275205113,
0.057196863,
0.0088786306,
-0.0511738174,
-0.050981909,
-0.04177792,
-0.0231186263,
0.0379580371,
0.0181086026,
-0.0277030561,
0.0074432893,
-0.0182751734,
0.1058155522,
-0.0099332361,
0.0824548602,
0.0656068698,
0.0180265903,
-0.059753634,
-0.0505217463,
0.0056060697,
-0.0211494621,
-0.0119412905,
-0.0172018502,
-0.0170198288,
0.0474212617,
0.0601184405,
0.0132513968,
-0.0254702065,
-0.0609166659,
-0.0261897165,
-0.0556283295,
-0.0041939248,
-0.007217689,
-0.0179008711,
-0.0599007867,
-0.0312476866,
-0.0185799804,
-0.0327999927,
0.0434760451,
-0.0126005867,
0.0269516706,
-0.0702637583,
-0.0134093948,
-0.0566639304,
0.0476028621,
-0.0101715848,
-0.0102051795,
-0.0026700888,
-0.0201610215,
-0.0285191443,
-0.0160770882,
0.0450485758,
-0.0245001502,
0.0371741019,
-0.0757108182,
0.0582913421,
0.0151608111,
0.008513662,
-0.0245972127,
-0.0426046103,
0.0080224881,
0.0176280197,
-0.0112082409,
-0.0412931517,
-0.0009670696,
0.0571641475,
-0.0323424861,
0.0817754567,
-0.0143606961,
-0.0350508168,
-0.0267959964,
0.0548285693,
0.0062008519,
-0.0427949987,
-0.0130546242,
0.0253595393,
0.0089015132,
0.0263914,
0.0286124907,
-0.0021734831,
-0.1006587446,
-0.0146125583,
-0.0188210011,
-0.0593974739,
-0.0156258103,
-0.0299418755,
0.0032220243,
-0.0334432051,
-0.0199955609,
0.0991211236,
0.0493401475,
-0.0200096294,
-0.0079217795,
-0.0103483377,
-0.0231946297,
-0.001995831,
-0.027514888,
-0.0793100819,
-0.0392251574,
0.0107495729,
-0.0151791116,
-0.0178307984,
-0.040442884,
0.0068241097,
-0.0484050326,
0.0058729905,
-0.0078664031,
-0.0156447068,
0.0638887212,
-0.0385650322,
0.0412210003,
-0.0154063478,
-0.0183459502,
-0.0232187528,
0.013818969,
-0.0034111256,
-0.0028152859,
-0.0568366535,
-0.0104926396,
0.036343731,
-0.0001804999,
0.0201921556,
0.0687582418,
0.0379980616,
-0.0240102839,
-0.0520112328,
0.0823906511,
-0.040783748,
-0.0058908341,
-0.0283998027,
0.0204692874,
0.010066174,
-0.0001675359,
0.0013596715,
-0.0473683514,
0.0090834117,
-0.0213326477,
0.0088014929,
0.0524793677,
-0.0298265889,
-0.0109148016,
-0.0439794697,
-0.0092972033,
0.0069212909,
0.0392899439,
-0.0169723425,
-0.0165005401,
-0.0183406994,
0.0204419196,
0.0160926357,
0.0177468602,
-0.0334145129,
-0.0589438453,
-0.0160239507,
-0.0267401487,
-0.0794301629,
0.0123680625,
-0.0208529718,
-0.0339551196,
0.010730342,
-0.0006108717,
0.0084016118,
0.0178486463,
-0.0749049485,
0.0027260417,
-0.0059806588,
0.0273449123,
-0.0452504531,
-0.0171119664,
0.0890251696,
0.0392989479,
-0.0016042093,
0.0538969077,
-0.0712225661,
-0.0265568811,
0.03195801,
-0.0418222845,
0.0177213829,
0.0243173838,
-0.0444314256,
0.0400143787,
0.0052748322,
0.0312566198,
0.0016810101,
0.043950703,
0.0026452504,
0.0290793367,
-0.0255672149,
0.0073660957,
-0.0144559667,
0.0017661994,
-0.0047589494,
0.040299233,
-0.0195209403,
-0.0274810866,
0.0444260724,
-0.0192050263,
0.0722142011,
0.0162220206,
-0.0191390514,
0.0165381618,
-0.008451757,
-0.003009852,
0.0173236169,
-0.0083129052,
-0.0205864664,
-0.0022780921,
0.00235581,
-0.0438490957,
0.0550402477,
-0.0664934814,
-0.0062400699,
-0.0371684432,
-0.012138376,
0.0216116868,
0.0271194335,
0.0100818761,
0.0196869932,
0.0058870511,
-0.0330674015,
-0.0106751006,
-0.0060628164,
-0.0169944614,
0.0612981357,
0.0266674105,
-0.0201215241,
0.0073315776,
0.0313965827,
-0.0691065416,
0.0191096421,
0.0079853935,
0.0555468611,
0.0166174099,
0.0059669134,
0.0625709742,
-0.0343989022,
-0.0199678335,
-0.0280554984,
-0.0063683847,
0.0296176877,
0.0176928099,
0.0076117977,
0.0161410067,
0.0035918404,
0.0349357128,
0.0219030492,
-0.0236890409,
0.0736995563,
0.0154093783,
-0.0401218683,
-0.0101296855,
-0.0115218656,
0.0149663975,
-0.0167944878,
-0.0101273097,
0.018636087,
-0.0471382402,
0.0116710775,
0.0129257971,
-0.0174466576,
0.0200433824,
0.019635221,
0.0408354625,
0.0765099302,
0.0045553008,
-0.02872766,
-0.0043719299,
0.0343338214,
0.0197789576,
-0.0106196469,
-0.026792502,
-5.999517455e-33,
-0.0239497852,
-0.0086055584,
-0.0084011741,
0.0130600678,
-0.0645365715,
0.0207952708,
-0.0148912584,
-0.0243770927,
0.0175479967,
0.0101318946,
0.0008274178,
0.0895573199,
0.0119076259,
-0.0393116251,
0.0231214017,
-0.0238051638,
0.1001111194,
0.0044158106,
-0.0274524055,
-0.0170393381,
0.0510153808,
0.0071006701,
0.085813731,
-0.0447130613,
0.0296975616,
0.0317022987,
0.0006740461,
0.0214938298,
0.015911648,
0.0119681042,
-0.0055465666,
0.0495693013,
0.0249918867,
-0.0379160717,
-0.0299586654,
0.0097943265,
-0.0803726688,
-0.0457280315,
0.0266905203,
0.0211726669,
-0.00726707,
-0.0119908387,
0.0172205959,
0.004396759,
-0.0218239445,
0.0282572843,
0.0162845198,
-0.0278929528,
-0.0218316577,
-0.0098337447,
-0.0710575655,
0.0614112429,
-0.025547687,
-0.0164768137,
-0.0142485583,
0.0058174925,
0.0284369793,
-0.0314789303,
-0.0550927445,
0.0022517971,
0.0351340733,
0.0134109771,
-0.0052142399,
-0.0063621853,
-0.0088480357,
0.0350523032,
-0.005478391,
-0.011042391,
0.0344485715,
-0.0329140611,
0.0248094071,
-0.0036238709,
0.0462785065,
-0.0258228742,
0.0120002022,
-0.0237124469,
-0.0033597846,
0.0107775107,
0.0497536659,
0.053312581,
0.0323067531,
0.0364599563,
-0.0149114328,
0.0092042442,
0.0151273245,
0.0261983462,
-0.0139914015,
-0.0043045282,
0.0296219327,
0.0229515266,
0.0170909055,
0.0339170285,
0.0156969912,
-0.0133700715,
0.0123346932,
-0.0025546004,
-0.0650040209,
0.0135064581,
-0.0008714558,
0.0549822114,
-0.0001032125,
-0.0099228676,
0.0399237126,
0.0507157221,
0.002353265,
-0.0335103571,
-0.0226958673,
0.0126209464,
-0.0628871843,
0.0033045695,
0.0101888925,
-0.0098468186,
-0.0701662749,
-0.0673125163,
-0.0190504845,
0.002203302,
-0.0060600201,
0.0272967853,
0.0653959587,
-0.0000455461,
-0.0037699556,
0.0244135763,
0.0080084652,
0.0282230619,
-0.0102720559,
0.0170749314,
0.0022272812,
0.0893389061,
0.0071853176,
-0.0288169906,
-0.002161178,
-0.0095428042,
2.726e-7,
0.0629286617,
0.0712947771,
0.0659690052,
-0.0149637135,
0.0352047309,
-0.0183702428,
-0.0208374839,
0.0280131865,
0.0008887849,
0.0445823371,
0.002037402,
0.0116287926,
0.0007541769,
0.0254623126,
0.0025786872,
-0.0276898686,
-0.0749012008,
-0.0227799118,
-0.1130593643,
0.0176785868,
0.0813617408,
0.0459742025,
0.0387843847,
0.0031458316,
-0.0425923206,
-0.0310721863,
0.0679229125,
-0.0229018144,
-0.0128142834,
0.0168179814,
0.0404237472,
0.0252773426,
0.0109315198,
0.0221142173,
-0.0142338816,
0.0227107462,
0.0027188475,
-0.0012005861,
-0.0183498021,
0.0169986188,
-0.0178485382,
-0.0223378297,
-0.0144684073,
-0.0505354628,
0.0967753902,
-0.053744074,
0.0155854141,
-0.0256710816,
-0.0231850892,
0.0136722177,
0.0496057756,
-0.0118361311,
-0.0723891333,
0.0224257074,
0.0059427461,
0.0639367923,
-0.0591977499,
-0.1101335064,
0.0208594091,
-0.0121275131,
-0.0377489217,
-0.0289504901,
-0.0560356677,
-0.0136004128,
0.0065726242,
-0.0343941525,
0.0298426449,
2.198583793e-34,
0.0282338299,
0.0596233197,
-0.019295333,
0.0116297649,
0.0077109528,
-0.0087746913,
0.0399590172,
0.0181127395,
-0.0136364326,
-0.053558629,
-0.0046793288
] |
http://arxiv.org/abs/2401.14021v1 | http://arxiv.org/pdf/2401.14021v1 | Accelerating Retrieval-Augmented Language Model Serving with Speculation | 2024-01-25T09:06:44 | 2024-01-25T09:06:44 | Zhihao Zhang, Alan Zhu, Lijie Yang, Yihua Xu, Lanting Li, Phitchaya Mangpo Phothilimthana, Zhihao Jia | cs.LG | cs.LG, cs.CL, cs.IR | Retrieval-augmented language models (RaLM) have demonstrated the potential to
solve knowledge-intensive natural language processing (NLP) tasks by combining
a non-parametric knowledge base with a parametric language model. Instead of
fine-tuning a fully parametric model, RaLM excels at its low-cost adaptation to
the latest data and better source attribution mechanisms. Among various RaLM
approaches, iterative RaLM delivers a better generation quality due to a more
frequent interaction between the retriever and the language model. Despite the
benefits, iterative RaLM usually encounters high overheads due to the frequent
retrieval step. To this end, we propose RaLMSpec, a speculation-inspired
framework that provides generic speed-up over iterative RaLM while preserving
the same model outputs through speculative retrieval and batched verification.
By further incorporating prefetching, optimal speculation stride scheduler, and
asynchronous verification, RaLMSpec can automatically exploit the acceleration
potential to the fullest. For naive iterative RaLM serving, extensive
evaluations over three language models on four downstream QA datasets
demonstrate that RaLMSpec can achieve a speed-up ratio of 1.75-2.39x,
1.04-1.39x, and 1.31-1.77x when the retriever is an exact dense retriever,
approximate dense retriever, and sparse retriever respectively compared with
the baseline. For KNN-LM serving, RaLMSpec can achieve a speed-up ratio up to
7.59x and 2.45x when the retriever is an exact dense retriever and approximate
dense retriever, respectively, compared with the baseline. | Accelerating Retrieval-Augmented Language Model Serving with Speculation: Retrieval-augmented language models (RaLM) have demonstrated the potential to
solve knowledge-intensive natural language processing (NLP) tasks by combining
a non-parametric knowledge base with a parametric language model. Instead of
fine-tuning a fully parametric model, RaLM excels at its low-cost adaptation to
the latest data and better source attribution mechanisms. Among various RaLM
approaches, iterative RaLM delivers a better generation quality due to a more
frequent interaction between the retriever and the language model. Despite the
benefits, iterative RaLM usually encounters high overheads due to the frequent
retrieval step. To this end, we propose RaLMSpec, a speculation-inspired
framework that provides generic speed-up over iterative RaLM while preserving
the same model outputs through speculative retrieval and batched verification.
By further incorporating prefetching, optimal speculation stride scheduler, and
asynchronous verification, RaLMSpec can automatically exploit the acceleration
potential to the fullest. For naive iterative RaLM serving, extensive
evaluations over three language models on four downstream QA datasets
demonstrate that RaLMSpec can achieve a speed-up ratio of 1.75-2.39x,
1.04-1.39x, and 1.31-1.77x when the retriever is an exact dense retriever,
approximate dense retriever, and sparse retriever respectively compared with
the baseline. For KNN-LM serving, RaLMSpec can achieve a speed-up ratio up to
7.59x and 2.45x when the retriever is an exact dense retriever and approximate
dense retriever, respectively, compared with the baseline. | [
0.0878517032,
0.063183479,
-0.013830387,
0.0520731658,
-0.0549825132,
-0.0253830347,
-0.0404901654,
-0.0082576051,
-0.0031464982,
-0.0442556962,
0.0071716383,
-0.0747266039,
-0.0295189712,
0.0365323424,
-0.008777923,
-0.0053050281,
0.0359239429,
-0.0155547578,
0.0403104872,
-0.0109686516,
-0.0024483663,
0.0284402631,
0.0304278191,
0.0097605698,
0.0089142295,
0.0062565152,
-0.0012914202,
0.0138757899,
-0.0487401113,
-0.0120606711,
0.0184594523,
0.0402459949,
0.0200099815,
0.0133274337,
0.0000023075,
-0.0264759231,
-0.0497040674,
0.0328301638,
-0.0254317317,
-0.0010334088,
0.0482970849,
0.0170684513,
0.0119376015,
0.037020497,
-0.0571055152,
0.0158260651,
0.0330930427,
-0.0078861238,
0.0111174006,
0.0695294365,
-0.0054375706,
-0.0254559107,
0.0058014044,
-0.0162980873,
0.0040614805,
-0.0426652208,
0.0187708903,
-0.0264256932,
0.0222991444,
0.0498728789,
0.0030043616,
0.053567376,
0.0128992135,
-0.0051574474,
0.0058618449,
0.0139624244,
-0.0258629359,
-0.0179088693,
-0.0253812708,
0.0438784249,
-0.009284866,
-0.0260854214,
-0.0190698523,
0.0233088378,
-0.0295540877,
0.0130298445,
-0.0267468337,
0.0314726941,
-0.0138867395,
0.008680216,
0.0252388585,
0.0066946107,
0.0247588679,
0.0129688308,
-0.0097455531,
0.0735645369,
0.0209018346,
-0.0371147394,
-0.0243700668,
-0.0087974193,
-0.0297113191,
-0.0179612748,
-0.0153048774,
0.0027886529,
-0.032294441,
0.0600856915,
-0.015645327,
-0.0194221828,
0.0317025557,
-0.0214866418,
0.0357347913,
0.0403802581,
-0.0067524398,
0.0326197967,
0.0004162372,
-0.0222718697,
-0.0102599319,
-0.0384876952,
-0.0421537608,
-0.0369119719,
-0.0158613734,
-0.0289848559,
-0.0671780929,
0.0467564724,
-0.0656602234,
-0.0248862579,
-0.0091190552,
0.030944407,
-0.0074466011,
-0.0163954366,
0.0326788388,
0.0432421304,
-0.036151018,
0.0017951371,
-0.0428904518,
0.0326999277,
-0.0184860416,
0.000881542,
0.0170357097,
-0.0997791141,
-0.0200432856,
-0.0140636247,
0.0092803435,
-0.0420803912,
0.0159677826,
0.0484203696,
0.0047246255,
-0.0288851466,
-0.0355254821,
-0.0610919483,
0.0084122373,
-0.0850182474,
-0.0204106439,
-0.0698889047,
0.0332904793,
0.0133162709,
-0.0271601789,
-0.0147965783,
-0.0049041808,
0.0504195131,
-0.0422524177,
0.0343486331,
-0.0304161813,
0.0242047161,
0.0058190646,
0.0095103942,
-0.0213994607,
0.1062194705,
-0.003978515,
0.014102662,
0.0595676079,
0.0106315613,
-0.0289747417,
0.0216674898,
0.0567011572,
0.0078167059,
-0.0206877701,
0.0459572263,
-0.0135030905,
-0.0123248659,
0.0168909431,
0.0498433597,
-0.0360128731,
-0.036267668,
-0.002807294,
0.0409427099,
-0.0058512609,
0.0038378055,
0.0495429784,
0.0374056138,
0.0073396619,
-0.0186682325,
-0.0024777176,
0.0157475658,
-0.0617859364,
0.0176188406,
0.0438668802,
0.0489911325,
-0.0187855586,
-0.0177972429,
-0.0297248233,
0.040997453,
-0.0033029278,
0.0274066124,
0.0517496914,
0.0078458078,
-0.0539285019,
0.0534035303,
0.0319968909,
-0.0114979409,
-0.0317058787,
-0.0411827378,
0.0063427705,
0.0560768582,
0.0565078966,
0.0042578746,
-0.0087971138,
0.0219167825,
-0.0051461952,
0.0349589586,
0.0378945917,
0.002300533,
-0.0115217287,
-0.028414363,
-0.037974678,
-0.0063584661,
0.0126178674,
0.0483079366,
0.0895780548,
0.0256489161,
0.0161128622,
-0.0113287205,
-0.0330786705,
-0.0153159425,
-0.0469183959,
-0.0128593063,
0.03108388,
-0.019412443,
0.0449448787,
0.0215364601,
-0.0153935673,
0.0341992304,
0.0422541425,
-0.0156507585,
-0.031758178,
-0.0008797867,
-0.0133760441,
-0.0298977513,
0.0136294141,
0.0618396699,
0.0128676733,
0.0070824865,
-0.0168942884,
-0.0345603824,
0.0408297107,
-0.0105928406,
0.0691003129,
-0.0742722452,
0.0110004358,
0.0004424309,
0.0276482329,
-0.0210100356,
0.0639723912,
-0.0115269488,
0.0419723801,
0.0239472296,
0.0005876712,
0.0061703343,
-0.0294888858,
-0.0373966433,
-0.034854047,
-0.0118891345,
0.0277773552,
0.0130080692,
-0.033606872,
-0.0209594443,
0.002074566,
-0.0678583086,
0.0164844468,
-0.0256055519,
-0.017932374,
-0.0067244517,
0.0477715172,
-0.0225882437,
-0.0191439558,
0.0009652627,
-0.0115741957,
0.0372675918,
0.0640166402,
-0.0283871256,
0.0585100949,
-0.0153776091,
-0.024441747,
-0.0262909569,
0.0255454388,
0.0499804579,
-0.0149433259,
0.0092617469,
-0.0242380667,
0.0057164631,
-0.0489369407,
0.0339347795,
-0.037079677,
0.0130273998,
-0.0174151454,
-0.0657386631,
-0.0074923648,
-0.0214327648,
0.0026763673,
-0.0257059764,
0.0094821025,
-0.0250271074,
-0.0651322678,
0.037358854,
-0.0442723148,
-0.0577591807,
0.0019373747,
-0.0214712508,
-0.0028695813,
0.0426959097,
-0.0037076822,
-0.0387402624,
-0.041880846,
0.0120378938,
0.018795114,
-0.0234467946,
-0.0621554181,
-0.0127420994,
0.0159367472,
0.0574139953,
-0.0424994864,
-0.0583409779,
0.0883316919,
-0.0363367684,
-0.0027183963,
-0.064408876,
-0.0011064034,
-0.0369615182,
0.0293136835,
-0.0043014688,
0.0091558732,
-0.0015038553,
-0.0001035658,
0.0044558048,
0.0137041444,
-0.0026369519,
-0.0569331497,
0.034540046,
0.0102013079,
0.0518198498,
0.0130301546,
-0.0409563109,
-0.059766043,
-0.0489378721,
0.0021712363,
-0.001191179,
0.027241338,
-0.035562288,
-0.0133113479,
0.0122182993,
-0.0104669994,
0.0148044713,
-0.0216495544,
-0.0060703107,
0.0109404484,
0.0000818316,
0.0743690953,
-0.0267565157,
0.0174118057,
0.027544504,
0.0282133836,
0.0596409217,
0.0577201769,
-0.0448936,
-0.0273524206,
-0.0012389109,
0.0324046426,
-0.004886575,
-0.019940963,
0.0361247733,
0.0081820637,
-0.0569786094,
-0.0328583606,
0.0807383582,
0.0260860808,
-0.0318967663,
-0.0273778494,
0.006384897,
0.0725640133,
0.01965525,
0.070244126,
-0.0948731974,
0.0007476709,
0.0054694391,
0.0448854268,
-0.0080078831,
-0.028663516,
-0.0446788967,
-0.1429096162,
0.0706863254,
-0.031131858,
-0.0070111812,
0.007928337,
-0.0900308937,
-0.0207083188,
0.0179150403,
0.0046950574,
0.0446423776,
-0.0143798022,
0.0114995455,
-0.0045542107,
0.0164411552,
-0.0293285623,
0.036272224,
-0.007309333,
0.0035570436,
0.0134233264,
-0.0121818576,
0.0290669501,
-0.0132131269,
-0.0167190786,
-0.0143531477,
0.0098992288,
-0.0248917714,
0.0229228064,
-0.0152168628,
0.0155360987,
0.0036701178,
-0.0128798801,
-0.0098142726,
0.0077239624,
-0.0147817209,
0.0743333772,
-0.0061914097,
0.0179476812,
-0.0577379428,
-0.0133210458,
0.0286044013,
0.0929890797,
0.0035487467,
-0.0032434808,
-0.0254491065,
0.0019710846,
-0.0163319651,
-0.045561038,
-0.015755672,
0.010144338,
-0.0061856029,
0.0189036354,
-0.055525668,
-0.0277222618,
-0.0408143289,
0.0252388939,
-0.0166532453,
0.0667233542,
-0.0053547388,
-0.0158146992,
-0.099021107,
0.0368341953,
-0.0157190617,
-0.0081255976,
-0.0091227256,
-0.0120641263,
0.0474357493,
-0.0356638022,
-0.0229370221,
0.003148647,
-0.0738965124,
-0.0711063892,
-0.0357423984,
0.0387530737,
0.0154019659,
0.0915489867,
-0.0273182262,
-0.030516712,
0.0512706712,
0.0385195985,
-0.0728813857,
0.018122334,
-0.0233780108,
-0.0128334425,
-0.0084773246,
0.035980586,
-0.0451161563,
0.0119897369,
-0.0327159949,
0.0396856703,
0.0049980022,
-0.0436383337,
-0.0063962117,
0.0432268754,
0.0617457405,
0.022442013,
-0.0145956744,
-0.0121946177,
0.0297197737,
-0.0446676947,
-0.038695123,
-0.0480391756,
0.0199632403,
0.0082192058,
-0.0464061871,
-0.0226527564,
-0.008146747,
0.0146538373,
0.0443455763,
-0.0233191978,
-0.033415053,
-0.0243141744,
-0.0274560321,
0.046667967,
0.0003974425,
0.0070625823,
0.0181808956,
-0.0052270675,
0.0636289343,
-0.0225437973,
0.0054567419,
0.0059607616,
-0.0052315718,
0.006188259,
-0.0126385698,
-0.0759448931,
-0.0897554457,
-0.0259573963,
0.0032698051,
0.0298143066,
-0.015319434,
0.0460554212,
-0.0307372566,
-0.0047967969,
-0.0138307465,
-0.0213627648,
0.0259881131,
-0.0610533692,
0.0512306057,
0.0201564487,
-0.0320610963,
0.0401675291,
0.0679617301,
0.0109182429,
0.0545580648,
-0.0249111969,
-0.0604094267,
-0.0150280008,
-0.0400647819,
-0.0161448829,
0.0469588004,
-0.0035528992,
0.0035843744,
-0.0191763509,
-0.0639960766,
0.0203278661,
-0.018257983,
-0.0412606262,
0.0276387539,
0.0372573584,
0.1005405188,
0.0278698467,
-0.0127640767,
-0.0650685802,
0.0160554685,
-0.0475689918,
-0.0328468569,
0.0158580337,
-7.449880589e-33,
0.008344695,
0.0083318539,
0.0523309037,
-0.0339281075,
0.0048311069,
0.0264423657,
-0.0093087703,
-0.0164499134,
-0.041420687,
-0.0165618043,
-0.0282240454,
0.0167803522,
0.0204234254,
0.0438770279,
-0.0076086177,
0.0239226799,
0.0151733337,
0.0064382367,
0.0040625175,
0.0126739265,
0.0453370661,
0.0593011901,
0.0365895964,
-0.0304642115,
0.0137628643,
-0.0218151435,
0.0301138479,
0.0402607061,
-0.0210281555,
0.0507060438,
-0.0277777035,
0.0556058511,
0.0027478153,
-0.0388969891,
-0.025188366,
-0.0199958459,
-0.073964119,
-0.0284770168,
0.0400840826,
-0.0071859695,
-0.0191409588,
0.0161955282,
0.0240967944,
0.0046624709,
0.0276161619,
-0.0029449605,
0.0127948625,
-0.0013530087,
-0.0068181981,
-0.0255397409,
0.0037221056,
0.0359097905,
0.0411951728,
0.0331394039,
0.0331065468,
0.052123908,
0.0252709128,
0.027932059,
-0.0681821853,
-0.0058181104,
-0.0143868262,
0.0683392286,
0.0433950089,
-0.0013315044,
0.0199043192,
0.0405799747,
-0.0004144643,
-0.0246396028,
-0.0131155588,
0.0173851326,
0.0163632985,
0.0543410443,
-0.0246637333,
0.0498359986,
0.0522997975,
0.0119771427,
-0.070856683,
0.0328661576,
0.0195419211,
0.0472781733,
0.0651771501,
0.011632422,
-0.0677521601,
-0.0464995466,
-0.0527938418,
0.0354118645,
-0.0567134209,
-0.054181464,
0.0400375985,
0.0570022017,
-0.0081131989,
0.0292408541,
0.0517972969,
-0.041844815,
-0.0245297067,
0.0646453947,
-0.0416968875,
-0.0180070288,
-0.0082140658,
0.0630277842,
-0.0470925681,
0.0014817885,
0.0228844527,
0.003406438,
-0.0607957542,
0.0307362769,
-0.0056632804,
0.0169079527,
-0.0522444621,
0.017688401,
-0.0307336245,
0.0672960281,
-0.0179048795,
0.0193527546,
0.0444756374,
0.0463626496,
-0.0004454437,
0.0525743961,
0.0344297402,
0.0361707471,
-0.0092550861,
-0.0179150775,
0.0224863403,
-0.0045650755,
-0.0105437702,
0.0239878986,
-0.0061347368,
0.0104275057,
-0.0099519594,
-0.0556807667,
-0.0107854251,
0.0452082343,
3.173e-7,
0.051174026,
0.0047116373,
0.0324140862,
0.0375949182,
-0.0112338038,
-0.0229175873,
0.0261154454,
0.0141021991,
0.0523529723,
-0.0334837511,
-0.0225129314,
0.0437172912,
0.0138630727,
0.0366883054,
-0.1067462265,
-0.0552981012,
-0.0788971931,
0.0026789391,
-0.057385236,
-0.039930515,
0.0398005247,
0.0951229334,
0.0039012667,
-0.0273038205,
-0.0222947411,
-0.020794753,
0.0691112056,
0.0001120069,
-0.0492105894,
0.0389905684,
0.0509612039,
0.0600110926,
0.0192637779,
0.0025469435,
-0.0361988433,
0.0371715799,
-0.0676105991,
0.0276881773,
0.0032632449,
-0.0665306449,
0.0316773094,
-0.0292085763,
-0.0245277118,
-0.0497535951,
0.0814259574,
0.0055467696,
-0.017737966,
0.0137391593,
-0.0274830516,
-0.0142464237,
-0.0110464683,
-0.0226076171,
-0.0068004807,
0.0251524411,
0.0040783929,
-0.0080645289,
-0.0283564255,
-0.1094329357,
-0.0430961289,
-0.0162779838,
-0.0477862172,
-0.0086370958,
-0.0334254429,
-0.0319443867,
-0.0052477242,
-0.0863371193,
0.0316697992,
3.380756421e-34,
0.0645681545,
0.0340279415,
-0.0043490184,
0.0339316018,
0.0227640495,
-0.0014145464,
0.000058368,
0.0324089676,
-0.0196207277,
-0.1093164831,
-0.0366604775
] |
http://arxiv.org/abs/2401.14019v1 | http://arxiv.org/pdf/2401.14019v1 | Unitxt: Flexible, Shareable and Reusable Data Preparation and Evaluation for Generative AI | 2024-01-25T08:57:33 | 2024-01-25T08:57:33 | Elron Bandel, Yotam Perlitz, Elad Venezian, Roni Friedman-Melamed, Ofir Arviv, Matan Orbach, Shachar Don-Yehyia, Dafna Sheinwald, Ariel Gera, Leshem Choshen, Michal Shmueli-Scheuer, Yoav Katz | cs.CL | cs.CL, cs.AI | In the dynamic landscape of generative NLP, traditional text processing
pipelines limit research flexibility and reproducibility, as they are tailored
to specific dataset, task, and model combinations. The escalating complexity,
involving system prompts, model-specific formats, instructions, and more, calls
for a shift to a structured, modular, and customizable solution. Addressing
this need, we present Unitxt, an innovative library for customizable textual
data preparation and evaluation tailored to generative language models. Unitxt
natively integrates with common libraries like HuggingFace and LM-eval-harness
and deconstructs processing flows into modular components, enabling easy
customization and sharing between practitioners. These components encompass
model-specific formats, task prompts, and many other comprehensive dataset
processing definitions. The Unitxt-Catalog centralizes these components,
fostering collaboration and exploration in modern textual data workflows.
Beyond being a tool, Unitxt is a community-driven platform, empowering users to
build, share, and advance their pipelines collaboratively. Join the Unitxt
community at https://github.com/IBM/unitxt! | Unitxt: Flexible, Shareable and Reusable Data Preparation and Evaluation for Generative AI: In the dynamic landscape of generative NLP, traditional text processing
pipelines limit research flexibility and reproducibility, as they are tailored
to specific dataset, task, and model combinations. The escalating complexity,
involving system prompts, model-specific formats, instructions, and more, calls
for a shift to a structured, modular, and customizable solution. Addressing
this need, we present Unitxt, an innovative library for customizable textual
data preparation and evaluation tailored to generative language models. Unitxt
natively integrates with common libraries like HuggingFace and LM-eval-harness
and deconstructs processing flows into modular components, enabling easy
customization and sharing between practitioners. These components encompass
model-specific formats, task prompts, and many other comprehensive dataset
processing definitions. The Unitxt-Catalog centralizes these components,
fostering collaboration and exploration in modern textual data workflows.
Beyond being a tool, Unitxt is a community-driven platform, empowering users to
build, share, and advance their pipelines collaboratively. Join the Unitxt
community at https://github.com/IBM/unitxt! | [
0.023790339,
0.0834632888,
-0.0245061778,
-0.0009766016,
-0.0245819464,
0.0206822548,
0.0623183288,
-0.0284435209,
-0.0296910591,
-0.0084497361,
-0.0215636976,
-0.0195183456,
-0.0025535673,
0.1005992666,
0.0409148075,
-0.0692183003,
0.0183358528,
0.012657349,
-0.033491753,
0.0101568848,
0.0237506442,
0.0493557975,
-0.0027127489,
0.0462079756,
-0.0285901465,
-0.0293689501,
-0.009908272,
0.0252044965,
0.0359591395,
-0.0041996278,
-0.0022972473,
0.087722607,
-0.0002928921,
0.071291998,
0.0000020535,
-0.0567281507,
-0.0169361345,
-0.0024773395,
-0.0052406592,
-0.0142231546,
0.0870799124,
-0.006025888,
-0.0142057994,
0.0151846223,
-0.0608701035,
-0.027735699,
0.0566704869,
-0.0097828796,
0.0407204367,
0.0592390932,
-0.0043739914,
0.0040384694,
0.0030799641,
-0.0335746966,
0.0045994851,
0.0226080604,
-0.0158132259,
0.0017391357,
0.035296414,
0.0400611684,
0.0083518689,
0.0240687169,
0.0592661425,
-0.0424377955,
0.0564349256,
0.0365937948,
-0.0216277745,
-0.0574634038,
0.0035558827,
0.0854083598,
0.0176590029,
-0.0379429683,
0.0170102641,
-0.0076808454,
-0.0199632123,
0.0026436939,
-0.0521912724,
0.0537127927,
0.0201945789,
0.0186696816,
0.0126362164,
-0.005136651,
0.0109062064,
-0.0205364507,
0.0162858348,
0.0389713757,
0.0410694554,
0.0176882036,
-0.0176901631,
0.0031976902,
-0.007259896,
-0.1024409384,
-0.0069651911,
0.0117567685,
-0.0133102415,
0.0019503961,
-0.042671565,
-0.063910991,
0.0412418917,
-0.0237881951,
0.0655698404,
0.0122730108,
-0.015301303,
0.0246704482,
0.0068779844,
0.0189788267,
0.045976799,
-0.0136099597,
-0.0375444368,
0.0094350027,
-0.0431361459,
-0.0028647345,
-0.0593490079,
-0.0150671797,
-0.0814631134,
-0.0167955402,
-0.0092151864,
0.0294538401,
-0.0140169682,
0.0117573868,
0.0740043372,
0.0500767417,
-0.0407962129,
0.014615613,
-0.0018381511,
0.007707146,
-0.0638602003,
-0.0018017745,
0.0211955607,
-0.0619713776,
-0.0328009129,
0.0175237227,
0.0238919593,
-0.0213184096,
-0.0061724912,
0.0178451482,
0.0036239012,
-0.004622174,
-0.0026935537,
-0.0486287437,
-0.0028996198,
-0.0429544263,
0.0008465165,
-0.0734995231,
0.0814782977,
0.009342893,
0.0113814464,
-0.0567385219,
-0.0161779821,
0.0096535143,
0.0321199931,
0.05781262,
-0.0215960033,
0.010348659,
-0.0124446964,
0.0148980366,
0.0271871984,
0.0483629256,
0.0291405078,
0.0530257262,
0.0432161875,
-0.0078988476,
0.0165433101,
0.0306284633,
0.019268034,
-0.0344189592,
-0.0289645921,
-0.0112067377,
-0.012566397,
0.0124365631,
0.0165715888,
0.0529958867,
-0.0163779277,
0.0123021435,
0.0449623019,
0.0258631408,
0.0437635928,
0.0285186004,
0.093266286,
0.0121064829,
-0.007320357,
-0.0512856394,
-0.0029250423,
0.0350963809,
0.0100661227,
0.0061105555,
-0.0365784168,
0.0750814751,
-0.0101784868,
-0.036883764,
-0.0363505743,
0.0049325558,
-0.0082399324,
-0.0240938719,
0.0165050831,
0.0678391755,
-0.0182535779,
-0.0137588978,
0.0236044899,
-0.0898756236,
0.0233880579,
-0.0109888399,
0.0520888194,
0.0926659629,
0.0773173273,
0.0152364252,
-0.0197898895,
0.009572628,
-0.0134146176,
0.0286400151,
0.0054214629,
-0.048657082,
0.0094178887,
-0.0362827331,
-0.0110356091,
-0.02229104,
0.018489426,
0.0263554882,
0.0183728691,
-0.0016373368,
0.0073962016,
-0.0535961352,
-0.0144460555,
0.0440894589,
-0.0608205348,
-0.0636747256,
0.010861693,
-0.0121121015,
0.0289749447,
0.0414812304,
-0.0299698636,
0.0023856598,
0.0318255648,
-0.0166814718,
-0.0237844326,
0.0624876395,
-0.0288398881,
-0.0156070702,
0.0252472926,
-0.0288765933,
-0.009324926,
0.0037600345,
-0.0402480401,
-0.039072223,
0.0405618772,
0.0030392902,
0.0025110699,
-0.0508021004,
-0.0062870663,
0.022017369,
0.0251943506,
-0.0318640135,
0.0552698486,
0.0110395551,
-0.0121634528,
-0.0030049558,
0.0296056475,
0.0232323855,
0.0176011715,
-0.0097279958,
0.0218810178,
0.0007173391,
-0.0134255504,
-0.0104788952,
0.0263732448,
-0.0508375652,
-0.0137546221,
-0.0938347131,
-0.0047872839,
-0.0252630003,
0.011491349,
0.0455984361,
0.0164774638,
-0.0255231541,
0.0123219239,
-0.0276367627,
-0.0136343306,
0.0475190692,
-0.0120238932,
0.0313049294,
-0.0339042619,
0.0181411896,
-0.037079379,
-0.011837271,
-0.0109478924,
0.041945003,
0.0125332037,
0.0006788076,
-0.0287567489,
0.0298919305,
-0.0241691731,
0.012925026,
0.0133630792,
-0.002325447,
0.0198670141,
-0.0219845194,
0.0971038416,
0.0329308882,
-0.0357569642,
0.049369771,
0.031958703,
-0.022138264,
-0.0437504351,
-0.0006420414,
-0.0279944092,
0.011298771,
-0.0020978444,
-0.0326679163,
0.0048958864,
0.0589408129,
0.006574675,
-0.0285096206,
-0.0304397512,
0.0459680222,
-0.0203155857,
-0.0414307714,
0.0085917767,
-0.0189006012,
0.0031202645,
0.016176369,
-0.0346258841,
-0.0634511113,
0.0587033704,
-0.0493275784,
-0.0402932875,
-0.0464416109,
-0.013292769,
-0.0522805415,
0.1191305667,
0.0180357397,
0.0079833549,
-0.026940152,
0.0142144728,
-0.0096906573,
0.0588439591,
0.0026520507,
0.0000793803,
-0.0121999402,
-0.0727390125,
-0.0305567496,
-0.0370417982,
-0.0080497498,
-0.0449773818,
0.0409909897,
-0.0107453987,
0.0273182485,
0.0547162592,
-0.0315302387,
0.0017382045,
-0.0273367558,
0.0009976861,
0.0328035168,
-0.0080413446,
-0.0223983023,
-0.0186467748,
0.0120328628,
0.0561364293,
-0.0091578243,
-0.0280194469,
0.0137970857,
0.0009482995,
0.0613441579,
-0.0005452872,
0.0075673345,
-0.0378882438,
0.0122968853,
0.0536999293,
-0.0107133621,
0.0297488533,
-0.0151169058,
0.0614394508,
-0.0362718403,
-0.0325600542,
0.1110441834,
0.0497271717,
-0.019624671,
0.0114981458,
-0.0166683886,
0.0670202076,
0.0261969566,
0.0304774269,
-0.0863629505,
0.0131275179,
0.0052134064,
0.0282567386,
0.0068935747,
-0.0109762643,
-0.008459148,
-0.0860033333,
0.0783355832,
-0.0052007758,
-0.0319649354,
-0.0045361407,
-0.0523372553,
-0.0019033971,
-0.0208804458,
-0.0044446932,
0.0041576582,
-0.0028679152,
0.0496231541,
0.013022474,
0.0023171881,
-0.0436671823,
-0.0250400491,
-0.0535499342,
0.0677299574,
0.0794448256,
0.03392433,
0.048594892,
-0.0336607061,
0.0011604031,
0.0005797564,
-0.0041297739,
-0.0403802246,
0.0273653436,
0.0120654674,
-0.0088872556,
-0.0291338023,
0.0764575675,
-0.0382689536,
-0.0476271361,
-0.025453262,
0.0458473712,
-0.0331074968,
-0.0161657128,
-0.0534128658,
-0.0112517225,
0.0179390423,
0.0754614919,
0.0405244008,
-0.0393956825,
-0.0316914357,
-0.0323449038,
0.0418120511,
-0.0970480293,
-0.033069253,
-0.0027113033,
-0.006610286,
-0.030307591,
-0.0806766823,
-0.0576112792,
0.0682734475,
0.058658585,
0.0227931198,
-0.0516724661,
-0.0411935151,
-0.0262397081,
-0.0515250824,
0.0386583731,
0.0004420878,
0.0036532443,
-0.0193417016,
-0.0952514857,
0.0088784918,
-0.0312010292,
-0.0186667368,
0.0721714869,
-0.0753478929,
-0.0468421876,
0.0341915525,
0.0491842292,
0.0239712913,
-0.0242299531,
-0.0079789842,
-0.045889359,
0.0088834781,
0.0476725064,
-0.0220899805,
0.0174191277,
-0.0143585159,
0.0158038791,
-0.0047416408,
0.0082831038,
0.0005640921,
0.0540508293,
0.0094464989,
0.0489696488,
0.0112951966,
-0.0209238958,
0.0054163663,
-0.0015200697,
0.0097687487,
-0.0291860979,
0.0130152116,
-0.0885304883,
-0.0261958577,
-0.0400232114,
-0.0128187947,
-0.0753106698,
0.0325154476,
0.0350656435,
0.0122538386,
-0.0175168607,
0.0139441304,
-0.0087719765,
0.0596256927,
-0.0629369766,
0.0140177859,
0.0440802835,
0.0121212136,
-0.0010270064,
0.0068630087,
-0.0017399419,
0.0273584872,
-0.0676382259,
0.0167660657,
-0.01869113,
0.0349769518,
0.0014333957,
-0.0213200524,
-0.0120317023,
0.0349995494,
-0.0205399152,
-0.0287423022,
-0.0522010699,
0.0141485529,
0.0053481972,
0.0150624029,
0.0711276755,
-0.047150027,
-0.0101546627,
-0.0022454206,
0.0354205295,
0.0032770308,
-0.0629799515,
-0.0113077415,
0.0025715407,
-0.0334580615,
-0.0785337314,
0.0592485406,
-0.0170537904,
0.0592965074,
-0.0135845691,
-0.0047764415,
-0.0295912344,
-0.0472413823,
0.0074586924,
-0.0257765576,
-0.0140163526,
-0.0094335377,
-0.0651699677,
-0.0572030246,
-0.0182346888,
0.0590724349,
0.0038984958,
0.0105723161,
0.0558023639,
0.0498370491,
-0.0226425659,
-0.0313131548,
-0.0105495378,
0.0275495574,
-0.0031291042,
-0.0529188775,
-0.0379565246,
-6.589226237e-33,
-0.0302134324,
-0.0234833676,
0.0089324471,
-0.0274238139,
-0.0108288527,
-0.0182234477,
-0.0388864167,
0.0196143184,
0.0350771956,
-0.0015707971,
0.0001364525,
-0.0144139761,
0.0110397413,
0.0328057259,
0.0079549784,
0.0203392878,
0.0598313287,
-0.013344408,
-0.0209787916,
0.0413364097,
0.0288347024,
0.0284005255,
0.0334755927,
-0.0888568759,
0.0176484007,
-0.0059743971,
-0.025095569,
0.0060476288,
-0.023497276,
0.0446359776,
-0.0219814721,
0.0029551587,
0.0083260918,
-0.0445491001,
-0.0252122786,
0.0083024474,
-0.0992040858,
-0.0376433097,
0.0226527471,
-0.015709063,
-0.0753305256,
-0.0198365599,
0.021690771,
-0.0201829839,
-0.0033463659,
0.0052720932,
0.0222099181,
0.0099760462,
0.0037037479,
-0.0420302711,
-0.0860513523,
0.0080982065,
0.0125991208,
0.0082510486,
0.0286303218,
0.0159798395,
0.0235042684,
-0.0041907192,
-0.0399643965,
0.02062946,
0.0021572371,
0.0339261256,
0.0870528817,
-0.0077694603,
0.0254112594,
0.0560852028,
0.0346215777,
-0.0084885061,
-0.0526328012,
0.0276375357,
0.0081391679,
0.068737641,
0.0267777052,
0.0690968037,
-0.0021542746,
-0.0344338901,
-0.0055988044,
0.0570030063,
0.0193435177,
-0.0183780361,
-0.013804866,
0.0164512414,
0.0047278162,
-0.0085230861,
-0.0296858568,
-0.0430365056,
-0.0197600573,
-0.0358240642,
0.0248729941,
0.038691856,
-0.0312193874,
0.0083935354,
0.0016435443,
-0.0010093782,
-0.0233639088,
0.0355577394,
-0.0221111495,
0.0286367908,
0.0075035868,
-0.0149653917,
-0.0184617732,
-0.0100530265,
0.0747382119,
0.0227330849,
0.0332034938,
-0.0046685571,
0.0466706119,
0.0507272109,
-0.0529156774,
0.0350478403,
-0.0688387379,
0.0641815215,
-0.01610508,
0.0268293712,
0.0127680935,
0.0012516137,
-0.026219869,
-0.0156960301,
0.0340707712,
0.0165698919,
-0.0745762885,
0.0234085228,
-0.0173895061,
-0.0151740601,
-0.0388221741,
0.0106713315,
-0.0302924924,
0.0083191255,
0.0752843693,
-0.0276897065,
-0.0082157366,
-0.0021165998,
2.734e-7,
0.0360642783,
0.0497522913,
0.0558261275,
0.0139647797,
0.0319653191,
-0.0123100299,
-0.0025133726,
0.0294148102,
-0.0011159106,
0.0120464806,
0.000677131,
0.0242911186,
0.0416284539,
0.0387983136,
-0.062002331,
-0.0118786022,
-0.0373049937,
-0.0000796174,
-0.0760007203,
-0.0277440157,
0.0274922866,
0.0641695932,
0.0311860275,
-0.0103500681,
0.0367256105,
0.0075681829,
0.0060341703,
-0.0299325269,
0.0295111798,
0.0354568772,
0.0290029626,
-0.0369483307,
-0.0037333113,
-0.045017384,
-0.0229347404,
-0.0125932954,
-0.0141861942,
0.0724717602,
-0.0180281289,
-0.0344779044,
0.0216159187,
-0.0213344581,
0.0088388883,
-0.1089346707,
0.0455033928,
0.0077035492,
-0.0097894883,
-0.0284744669,
0.008759412,
0.0455760248,
-0.0089141196,
-0.0223098807,
-0.0016611946,
0.0133127365,
0.0248457976,
-0.0145355444,
-0.0227949228,
-0.0745472163,
-0.0439756997,
-0.0290352497,
-0.0536390357,
-0.0493667871,
-0.052070383,
0.0055150017,
-0.0179332662,
-0.0273797046,
-0.0082716588,
2.944987119e-34,
0.0145218298,
0.0054844981,
-0.0061455858,
0.0018976814,
0.0366752036,
-0.0197425298,
0.0162707996,
0.0184893068,
-0.0135163497,
-0.0538085289,
-0.0617030114
] |
http://arxiv.org/abs/2401.14016v3 | http://arxiv.org/pdf/2401.14016v3 | Towards Uncertainty-Aware Language Agent | 2024-01-25T08:48:21 | 2024-05-30T13:26:38 | Jiuzhou Han, Wray Buntine, Ehsan Shareghi | cs.CL | cs.CL | While Language Agents have achieved promising success by placing Large
Language Models at the core of a more versatile design that dynamically
interacts with the external world, the existing approaches neglect the notion
of uncertainty during these interactions. We present the Uncertainty-Aware
Language Agent (UALA), a framework that orchestrates the interaction between
the agent and the external world using uncertainty quantification. Compared
with other well-known counterparts like ReAct, our extensive experiments across
3 representative tasks (HotpotQA, StrategyQA, MMLU) and various LLM sizes
demonstrate that UALA brings a significant improvement of performance, while
having a substantially lower reliance on the external world (i.e., reduced
number of tool calls and tokens). Our analyses provide various insights
including the great potential of UALA compared with agent fine-tuning, and
underscore the unreliability of verbalised confidence of LLMs as a proxy for
uncertainty. | Towards Uncertainty-Aware Language Agent: While Language Agents have achieved promising success by placing Large
Language Models at the core of a more versatile design that dynamically
interacts with the external world, the existing approaches neglect the notion
of uncertainty during these interactions. We present the Uncertainty-Aware
Language Agent (UALA), a framework that orchestrates the interaction between
the agent and the external world using uncertainty quantification. Compared
with other well-known counterparts like ReAct, our extensive experiments across
3 representative tasks (HotpotQA, StrategyQA, MMLU) and various LLM sizes
demonstrate that UALA brings a significant improvement of performance, while
having a substantially lower reliance on the external world (i.e., reduced
number of tool calls and tokens). Our analyses provide various insights
including the great potential of UALA compared with agent fine-tuning, and
underscore the unreliability of verbalised confidence of LLMs as a proxy for
uncertainty. | [
0.0268239547,
0.0694152266,
-0.0241132714,
0.0135607449,
-0.0313819647,
-0.0106783891,
-0.0144826202,
-0.0239528418,
0.0077040251,
-0.0432325862,
-0.0135337338,
-0.0530143268,
-0.0127695017,
0.0623455942,
0.0343593694,
-0.0354779512,
0.0428894646,
-0.0533962101,
0.0526593775,
-0.0179420821,
-0.0029986978,
-0.0254081152,
-0.0205610581,
0.0623711757,
-0.0141780684,
0.004404949,
-0.0547673218,
0.0084932707,
0.0110976733,
0.0006545293,
-0.0285952799,
0.0449063815,
0.0059667006,
0.0525056049,
0.0000018761,
-0.0600957833,
0.0242354255,
-0.0149717163,
-0.0316833965,
-0.0531194657,
0.0402385257,
0.0364656225,
-0.0683838874,
0.0322428681,
-0.0345368348,
0.0108969547,
0.0787015855,
-0.0119565642,
0.0046313931,
0.0404885188,
-0.0037415591,
0.0439579375,
0.0085262898,
0.0214728452,
-0.0023961731,
-0.0454405844,
-0.0173250157,
-0.0657089651,
0.046744369,
0.0046074735,
-0.002344917,
0.0433545671,
0.0375419855,
-0.0409524441,
-0.0567963421,
0.012064429,
-0.0285509806,
-0.0630651489,
-0.0362294614,
0.0557101704,
0.0223032013,
0.0104643758,
-0.0131514492,
-0.0115010478,
0.0199656561,
0.0287363045,
-0.0088857003,
0.0291229356,
0.0076310458,
-0.0030402564,
-0.0148865497,
0.0086277882,
0.0124515351,
0.008362432,
-0.0238291211,
0.118598409,
0.0046213944,
0.0333611555,
-0.0122748874,
-0.0009353111,
0.020756539,
-0.081675753,
-0.0114497645,
0.0741543919,
-0.0870351717,
0.0474625528,
-0.0279903915,
-0.0021580081,
0.0290896334,
-0.0533087514,
0.0077787274,
0.0665359274,
-0.0059883618,
0.0161031857,
0.0256246869,
0.0301315486,
0.017418148,
0.0026550738,
-0.0123468414,
-0.0322474726,
0.0086439447,
-0.0074849227,
-0.0232754461,
0.0085796742,
-0.0321844779,
-0.042151399,
-0.0316522792,
0.0744824857,
-0.0380935818,
0.0547947846,
-0.0501920618,
-0.015707938,
-0.0272006486,
-0.023832934,
-0.0609307475,
0.0135374339,
-0.0317184404,
0.0081086224,
0.0367288366,
-0.102022253,
0.0082386453,
-0.0033831936,
0.0033767771,
-0.0093952939,
-0.0347334482,
0.117429018,
-0.0110976826,
-0.0174224917,
0.0023538817,
0.0018247006,
0.0013765886,
-0.0778483152,
-0.0492379479,
-0.0576244295,
-0.0047501186,
0.0229083858,
-0.0314795747,
-0.0615939759,
0.0157314576,
0.011708756,
0.0066217161,
0.0779478699,
-0.020257147,
0.0039688782,
-0.0074867913,
0.0089169405,
-0.0210377909,
0.0788098574,
0.0578473844,
0.0414668247,
0.0418039002,
0.0279233456,
0.0493440181,
0.037121091,
-0.0428324938,
0.0116559397,
-0.0225763768,
-0.005393751,
-0.0207721628,
0.0198807251,
-0.0043765684,
0.048152253,
0.0012503252,
0.0263517145,
0.0906582922,
0.0274947975,
-0.0251906756,
-0.0094648311,
0.0097793182,
0.0292011872,
0.0208114199,
-0.0287597757,
-0.0542816184,
0.0498011857,
-0.0148337847,
0.0211211368,
-0.0115604438,
0.0257418249,
-0.0237754956,
-0.0525447913,
-0.0268052109,
-0.0352161564,
0.0282975975,
0.0294210259,
0.0339683183,
-0.0074884281,
-0.0368559845,
-0.0048400257,
-0.0025594956,
-0.0234483648,
-0.0162167437,
-0.0956802443,
0.0183592625,
0.0932280049,
0.037624415,
-0.0128102731,
-0.0580734834,
0.0171490181,
-0.0338058099,
0.0705276281,
0.0430635028,
-0.0116480002,
0.0156278182,
-0.0208597202,
-0.0255992804,
-0.0202443786,
0.011669579,
0.033382535,
0.0032573512,
0.038636107,
-0.0253428333,
0.0029941224,
0.0206534062,
-0.0075911595,
0.0322007909,
-0.0300808009,
-0.0070038917,
-0.0056671281,
0.0414951667,
-0.0151399225,
0.0014221923,
0.004267009,
0.002012419,
0.0278735515,
-0.0397959799,
0.0291303843,
0.0023327647,
0.0419466384,
0.0549509712,
0.0675368607,
-0.0196858067,
0.0056677316,
-0.035800416,
-0.0298471656,
0.055007685,
-0.0129680354,
0.0177836772,
0.0025429048,
0.0232221242,
-0.0322087035,
0.0201600827,
-0.0205160417,
0.0122656161,
0.0155302817,
-0.0274705254,
0.0139560951,
0.0793746412,
0.0043895789,
-0.0150923012,
-0.0141252074,
0.0161255188,
0.0331231952,
0.042459853,
0.0408564024,
0.0226919334,
-0.0208878703,
-0.0339135453,
-0.0550951548,
-0.022603279,
-0.0208595134,
-0.0118693728,
-0.0064293342,
0.0289053768,
-0.0066898442,
-0.004049113,
-0.0160751026,
0.032465443,
-0.0023218389,
0.0406855159,
0.0337690488,
0.0038581253,
-0.071162276,
-0.0552704185,
0.0067542703,
0.0176110156,
0.0080115329,
-0.0064102351,
-0.0280657373,
-0.0646660924,
0.040815521,
-0.0424776673,
0.0293672439,
-0.0130746067,
-0.0573724285,
0.0306087658,
-0.0249774214,
0.0054457188,
0.0070762406,
-0.0219734199,
0.0120682102,
0.0898148343,
-0.0520771444,
-0.057543911,
0.0241054576,
-0.0162031166,
-0.025004968,
-0.0602061041,
0.0124868508,
-0.0236948188,
0.0479202718,
-0.0038629936,
0.0027000627,
-0.0287857596,
0.0136437705,
0.0100106606,
-0.0554125682,
0.0072241817,
-0.013672187,
-0.0399788693,
0.0808346346,
0.0108946096,
-0.0593484528,
0.0877572075,
-0.032346122,
-0.0027715922,
-0.0034940583,
-0.0326315016,
-0.005618121,
0.060410358,
-0.0185030289,
-0.0117767118,
-0.0242817961,
-0.0222493708,
-0.0537093095,
-0.0365814567,
0.0146373082,
-0.0386599824,
0.0130812433,
-0.0675653145,
0.0113581829,
-0.0206435844,
-0.0352043808,
-0.0344428308,
0.0406276509,
-0.0507999472,
0.018806126,
0.0306998007,
-0.0429528728,
-0.0239448603,
0.00512427,
0.0239970889,
0.0464940965,
-0.0275695622,
-0.0214432422,
-0.0430290438,
-0.0382953435,
0.0177717041,
0.0583315492,
0.0640930533,
0.016072277,
0.0276981089,
0.0179094374,
0.0303341635,
-0.0593722686,
-0.0311361589,
-0.0253442191,
0.0515791252,
0.0025463447,
-0.0164354444,
0.0659687445,
0.083934702,
-0.0455132425,
-0.0227640178,
0.1119759455,
0.0103414543,
-0.0151140941,
0.0049829893,
-0.0784379244,
0.0490122698,
0.025751831,
0.0611315295,
-0.0489494801,
-0.0244353861,
0.0354425535,
0.0159132052,
-0.0249896925,
-0.0045830668,
-0.0138601679,
-0.0755993649,
0.0568695851,
-0.014196706,
-0.0292179715,
0.007103601,
-0.0182696506,
0.0314070322,
0.0281654894,
0.0763366148,
0.0097839786,
0.0122721158,
-0.0167478025,
-0.0206712373,
-0.0291866194,
-0.0364099294,
-0.0318761505,
-0.0399942063,
0.0694412664,
0.0413986668,
0.0137417857,
-0.0104100481,
0.0075412584,
0.0227555111,
-0.0056356583,
0.00318336,
-0.0114586474,
-0.0290624648,
0.0506176613,
-0.0173563231,
-0.0079443688,
-0.0269306172,
-0.0237812847,
-0.0095240362,
-0.0185835175,
0.0601036288,
-0.0422215797,
0.0237899907,
-0.0121603152,
-0.0440965071,
0.0418170281,
0.0236344412,
-0.004089396,
-0.0349507779,
-0.0668304116,
-0.0686413124,
-0.0378273502,
-0.0328230076,
-0.0132767446,
0.0118348021,
0.0135175698,
-0.0561180376,
-0.0401600152,
-0.0556074977,
0.0294772275,
0.0243191496,
0.0173823126,
0.0021805833,
0.0342770442,
0.0396985523,
-0.0117208557,
-0.008945886,
0.0008134882,
0.0462955199,
-0.0313756093,
-0.007487834,
0.0275914446,
-0.0347838178,
0.0176135562,
0.0344986171,
-0.0235810373,
-0.0792022273,
0.0191710815,
0.0730724335,
0.0372269936,
-0.0027138751,
-0.0350300632,
-0.0744673684,
0.0376202837,
0.0404385701,
-0.0634986013,
0.0044094394,
0.0450814664,
0.0384980254,
0.00884677,
0.0567697883,
-0.0776431561,
-0.0192009211,
-0.0322471447,
0.0381753966,
-0.0262779761,
-0.0599820241,
-0.0157860331,
0.0197336767,
-0.0146487914,
-0.0143237505,
0.0270621628,
-0.0200566854,
0.0297653973,
-0.040278852,
0.0101017263,
-0.0489060879,
-0.0066278945,
-0.0260639489,
-0.0228142105,
0.0324588493,
0.0125947678,
-0.0017562083,
0.0124108437,
0.0388806164,
0.009048773,
-0.0072339233,
0.0137635814,
0.049953863,
0.0252469555,
-0.0033866193,
0.0387584902,
0.0025226083,
0.0275261849,
0.0053080376,
0.0026762928,
0.0100908633,
-0.0536657125,
-0.0205400623,
0.031464681,
-0.070672363,
-0.1354409754,
0.0399780646,
0.0290914923,
0.0260342583,
0.0552156195,
0.0523137152,
-0.0708645359,
0.0168930534,
-0.0086927768,
0.0211301111,
0.0234504677,
-0.0159946773,
0.0380025022,
0.0221551396,
-0.0088192616,
-0.0186457708,
0.0175583418,
-0.0279822815,
0.0392691419,
-0.0637925044,
-0.0385615118,
-0.0147190448,
-0.0318231694,
-0.0316816643,
0.0321021155,
-0.0208929069,
-0.0178267397,
-0.0498246811,
-0.0643010437,
-0.014134381,
0.0280304831,
-0.0021818846,
-0.0644546151,
0.0167858526,
0.0393251739,
0.048207283,
0.0128155723,
-0.0428542271,
-0.0346154943,
0.0147485165,
0.0271555763,
-0.0038438577,
-5.778056156e-33,
-0.0354124568,
-0.0424006656,
0.0407460183,
-0.0185995698,
0.0113876378,
-0.0288462993,
-0.0600283407,
0.0244951919,
-0.008925071,
0.0115517573,
-0.0316516384,
0.0137629313,
0.0340891965,
-0.0008296271,
0.0448070467,
0.0015755137,
0.0350828394,
0.016044654,
-0.0195735209,
-0.0124325566,
0.0399738178,
-0.0117418393,
0.0241450183,
-0.047224436,
0.0264693312,
0.0270983167,
-0.0236289203,
0.0436200164,
-0.0031675671,
0.0362179354,
-0.0026597083,
0.0464285873,
-0.0230613686,
0.0103328777,
-0.0196811557,
-0.0216646045,
-0.0405590832,
-0.0461186394,
0.0072624092,
0.0640082955,
-0.0272144116,
-0.0311894584,
0.0151526639,
-0.0067710006,
-0.0652014837,
0.0566357858,
-0.0013800092,
-0.0224202443,
-0.0041748811,
-0.0195299964,
-0.0673788339,
0.0187538099,
0.0163433664,
-0.0175783634,
-0.0010090087,
0.0028501581,
0.0010537046,
-0.0400373787,
-0.0290109441,
0.0153284697,
-0.0262096357,
0.0032850665,
0.0873452872,
0.0414651148,
-0.0031533986,
0.0470295697,
0.0297790859,
0.0286961552,
-0.0111252964,
0.0044139642,
0.0092290388,
0.019321898,
0.0150630344,
0.0696939528,
-0.0250429697,
-0.0042595509,
-0.0275963675,
0.0241692532,
0.0954406708,
0.0327615254,
0.0274625979,
0.0358388759,
-0.0446837097,
-0.0157933533,
-0.0238180459,
0.0330923088,
-0.0258222464,
-0.0474634171,
0.0141821112,
0.0044917823,
0.0178356189,
-0.0366883576,
0.0143289929,
0.004258086,
-0.0348077863,
0.002108237,
-0.029057635,
0.0387554802,
0.0204234794,
0.0172603112,
0.0155870486,
-0.0097675044,
-0.0123950876,
0.0071858093,
0.0022455466,
0.0344399773,
-0.0402841903,
0.0708762929,
-0.0673647597,
0.0262573268,
-0.0475095846,
0.0477796048,
0.0206477176,
-0.0495109744,
0.0054934002,
-0.0590169318,
-0.0116348248,
0.061066065,
-0.0219717342,
-0.0181913003,
0.0073450161,
0.0537452362,
-0.0050793733,
0.0002981127,
-0.058541771,
-0.0356930271,
-0.0587821156,
0.0501936115,
0.00334197,
-0.0609975867,
-0.0091969194,
-0.0246793646,
2.743e-7,
-0.0007236922,
0.0477139875,
-0.0116095217,
0.0704627559,
0.0448235534,
0.0543549508,
-0.0360894687,
-0.0182789192,
-0.0174690615,
0.0087013096,
-0.0081035076,
0.036357861,
0.0298763085,
0.0721613839,
-0.0547770113,
-0.0162105281,
0.0079309102,
0.0156830456,
-0.0318102837,
-0.0005426951,
0.0215292834,
0.0703556985,
0.0222888123,
-0.001736342,
-0.0049855635,
-0.0489841551,
0.0160947647,
-0.0441450626,
0.0000738208,
0.0366432332,
-0.0101559414,
0.0125295874,
0.0020606271,
-0.0110920304,
0.0163289774,
0.0117519163,
-0.0694694445,
0.0452791527,
0.0128793055,
-0.0256931093,
0.0309477132,
-0.0489203371,
0.0066598235,
-0.0314664319,
0.0245702788,
-0.0285649151,
-0.0475098938,
-0.0144212898,
0.0120359436,
-0.0445990749,
-0.0104934406,
-0.0138205756,
-0.0233810358,
0.0140105849,
-0.0132406736,
-0.0206474792,
-0.0027923186,
-0.0389201045,
-0.0080823228,
-0.0099912658,
-0.0173572991,
-0.0621443354,
-0.044710204,
-0.0111879241,
0.00606988,
-0.0477940626,
0.0172474459,
2.404775145e-34,
0.0459091924,
0.0202955306,
-0.0107087288,
0.0719751716,
0.0504071265,
0.0247390158,
0.0017725446,
0.0204459205,
0.0835724771,
-0.0172818676,
-0.0508129448
] |
http://arxiv.org/abs/2401.14011v3 | http://arxiv.org/pdf/2401.14011v3 | CMMU: A Benchmark for Chinese Multi-modal Multi-type Question Understanding and Reasoning | 2024-01-25T08:22:10 | 2024-05-08T07:34:06 | Zheqi He, Xinya Wu, Pengfei Zhou, Richeng Xuan, Guang Liu, Xi Yang, Qiannan Zhu, Hua Huang | cs.CL | cs.CL, cs.AI, cs.MM | Multi-modal large language models(MLLMs) have achieved remarkable progress
and demonstrated powerful knowledge comprehension and reasoning abilities.
However, the mastery of domain-specific knowledge, which is essential for
evaluating the intelligence of MLLMs, continues to be a challenge. Current
multi-modal benchmarks for domain-specific knowledge concentrate on
multiple-choice questions and are predominantly available in English, which
imposes limitations on the comprehensiveness of the evaluation. To this end, we
introduce CMMU, a novel benchmark for multi-modal and multi-type question
understanding and reasoning in Chinese. CMMU consists of 3,603 questions in 7
subjects, covering knowledge from primary to high school. The questions can be
categorized into 3 types: multiple-choice, multiple-response, and
fill-in-the-blank, bringing greater challenges to MLLMs. In addition, we
propose an evaluation strategy called Positional Error Variance for assessing
multiple-choice questions. The strategy aims to perform a quantitative analysis
of position bias. We evaluate seven open-source MLLMs along with GPT4-V,
Gemini-Pro, and Qwen-VL-Plus. The results demonstrate that CMMU poses a
significant challenge to the recent MLLMs. The data and code are available at
https://github.com/FlagOpen/CMMU. | CMMU: A Benchmark for Chinese Multi-modal Multi-type Question Understanding and Reasoning: Multi-modal large language models(MLLMs) have achieved remarkable progress
and demonstrated powerful knowledge comprehension and reasoning abilities.
However, the mastery of domain-specific knowledge, which is essential for
evaluating the intelligence of MLLMs, continues to be a challenge. Current
multi-modal benchmarks for domain-specific knowledge concentrate on
multiple-choice questions and are predominantly available in English, which
imposes limitations on the comprehensiveness of the evaluation. To this end, we
introduce CMMU, a novel benchmark for multi-modal and multi-type question
understanding and reasoning in Chinese. CMMU consists of 3,603 questions in 7
subjects, covering knowledge from primary to high school. The questions can be
categorized into 3 types: multiple-choice, multiple-response, and
fill-in-the-blank, bringing greater challenges to MLLMs. In addition, we
propose an evaluation strategy called Positional Error Variance for assessing
multiple-choice questions. The strategy aims to perform a quantitative analysis
of position bias. We evaluate seven open-source MLLMs along with GPT4-V,
Gemini-Pro, and Qwen-VL-Plus. The results demonstrate that CMMU poses a
significant challenge to the recent MLLMs. The data and code are available at
https://github.com/FlagOpen/CMMU. | [
-0.0040180706,
0.0354228877,
-0.0006287484,
0.0581463017,
-0.0310807973,
0.0281334557,
0.0445059128,
-0.0371584781,
0.0131392898,
-0.0539355166,
-0.0286030676,
-0.0715098381,
-0.0243984014,
-0.0018728292,
0.0245359819,
-0.0200102907,
0.0479212441,
-0.0285131149,
0.0145519702,
-0.0144354329,
-0.0123641938,
-0.0194116477,
-0.0304525308,
0.0075107459,
0.0303389113,
-0.0670900792,
-0.0333471894,
0.00416897,
0.0392133743,
-0.0429783165,
-0.0009157992,
0.0617262274,
0.0194284469,
0.0383228101,
0.000002095,
-0.0703128278,
-0.0370787978,
-0.0068200319,
-0.0188172814,
0.0186630264,
0.0013904187,
0.0201814175,
0.0178989265,
0.0195395853,
-0.0247725435,
-0.0417248309,
0.0129051134,
0.0492571741,
0.0586974211,
0.0615061522,
0.0096566211,
-0.0153127536,
0.0057163308,
-0.0010631377,
0.0278827325,
-0.0622566789,
0.0257044155,
0.0525667034,
0.0019120184,
0.0327543914,
-0.0466139875,
0.035434451,
0.0296908896,
0.008762911,
-0.0796514824,
0.010906958,
0.0040274649,
0.0037321809,
-0.0008070053,
0.0962490067,
0.0251412131,
-0.0037963334,
-0.0488754883,
0.0174136423,
0.0249023512,
0.028025765,
-0.0535910949,
0.0245802626,
-0.0127767297,
-0.0381088406,
0.0059146136,
-0.043927528,
0.0132473418,
0.0125466734,
-0.0346051119,
0.056095235,
0.023896642,
-0.040701963,
0.0235550068,
-0.030255327,
-0.0436480455,
-0.0206756163,
-0.0096919406,
0.0384664051,
-0.0370572172,
0.0287588984,
0.0080573829,
0.0001558161,
0.0097076986,
-0.0184390675,
0.007871042,
0.0246112552,
-0.0249240287,
0.0693852454,
-0.0065840781,
0.0735187307,
-0.0092761582,
0.0266630519,
-0.0598642454,
-0.0329154879,
-0.0172849465,
-0.021571381,
0.0018371033,
0.0336659029,
-0.0310939495,
-0.0084799137,
-0.035268683,
0.0314676985,
-0.0106485179,
0.0259765554,
-0.0151343038,
-0.0050162254,
0.008128373,
-0.0252115224,
-0.03532039,
0.0880813003,
-0.0498653837,
0.0301591735,
0.006011094,
-0.0256820098,
-0.0004535784,
0.0289028473,
0.0052400013,
-0.0561204217,
0.0101108765,
0.0731057003,
-0.0223615095,
-0.0135807963,
0.0001646608,
-0.0133837666,
-0.008266815,
-0.0571002327,
0.0189026408,
-0.0545248985,
-0.0078424038,
0.017620191,
0.0136948004,
0.00938462,
0.014652024,
0.0132452194,
0.0282681566,
0.0246473085,
0.0450669862,
-0.0074977758,
-0.0033284833,
0.0212191809,
-0.0290830191,
0.0580381677,
0.002523568,
-0.0199336559,
0.0641793087,
-0.0428365991,
-0.0228267945,
0.0223816615,
-0.0207894519,
-0.0155620724,
-0.0172088929,
0.0594869591,
0.0224776343,
0.0333430432,
-0.0012211567,
0.0153388791,
-0.0087758824,
-0.0050932253,
0.0613342226,
0.01528739,
0.0080912756,
-0.0181402415,
0.0273226723,
-0.0018518867,
0.0304818768,
0.0103780795,
-0.0408204868,
0.0215495508,
-0.0769976005,
0.0138077587,
0.0015320013,
0.0412032418,
-0.0190513227,
-0.0519876331,
-0.0192047171,
0.0109878546,
-0.0223890208,
0.0482344441,
0.0462338068,
-0.050321307,
-0.0102343159,
0.0272837542,
0.000931078,
-0.0066836718,
-0.0010545525,
-0.0295271371,
0.0374764316,
0.002986707,
0.0155557888,
-0.0394714661,
-0.0764407739,
-0.000679409,
0.0085939514,
0.024909066,
0.0655554235,
-0.0310784914,
0.0055033104,
-0.0465308018,
-0.0088967523,
-0.0039692386,
0.0050661098,
0.0087255221,
-0.0197877735,
0.011727172,
-0.0091649238,
-0.0716930479,
0.0381086543,
0.0024179451,
0.0002561887,
-0.0357333198,
-0.0261926614,
0.0003464521,
0.0280315802,
0.0062900898,
-0.0152992737,
0.0189123619,
0.0310456958,
-0.0292817261,
-0.0261023361,
0.0127048325,
-0.072518535,
-0.0556649379,
0.0664777085,
-0.0436142348,
-0.0717207044,
0.0617724173,
-0.048758354,
-0.0375486314,
0.0078223245,
0.030332284,
-0.0093965167,
-0.0619328953,
0.0100376578,
-0.0201946069,
0.0074973446,
-0.0563781522,
0.1088347062,
0.0144111272,
0.0562845021,
0.0471250005,
0.0358468294,
0.0394926257,
0.0112682087,
-0.0412289202,
0.0236220676,
-0.0299901646,
0.0247358531,
-0.0062272949,
0.0159588046,
-0.0813444108,
-0.033033926,
-0.102735199,
0.0477678403,
-0.0237147082,
-0.0049301232,
0.0305453837,
0.0160419494,
-0.0211205054,
-0.0049931537,
-0.0525014326,
-0.0034720793,
-0.0197970308,
0.0087585347,
0.0172645189,
-0.0130931102,
0.020351056,
0.0027948734,
-0.0037832463,
-0.0579007305,
0.0371638574,
-0.000351652,
0.0602798909,
-0.0190380067,
-0.0063293963,
-0.0364970304,
-0.0064193024,
-0.0076845279,
-0.0066058766,
0.0236569401,
-0.0410923511,
0.0652070418,
0.0865464285,
0.0174311474,
-0.0143402088,
0.0521611348,
-0.0173936486,
-0.0536836982,
0.0560292378,
0.0136755621,
-0.0138916075,
-0.0648840368,
0.0044208704,
-0.0325833,
0.0784229338,
0.0109195272,
-0.0425207168,
0.0025852141,
0.0071392334,
-0.0254069418,
-0.0077154781,
-0.07049454,
0.0473541059,
-0.0217863452,
0.0237134844,
-0.0242098831,
-0.002940357,
0.0615961589,
-0.0304688141,
-0.0189965982,
-0.0734842718,
0.0115315337,
-0.0455237664,
0.0757207498,
-0.0150959967,
0.0159486122,
-0.0083545139,
-0.0190863628,
-0.0863223672,
0.0317688026,
0.0155097647,
-0.0302569941,
0.0576666966,
-0.0543576144,
-0.0279000178,
-0.0344244018,
-0.0112841669,
-0.0238575656,
-0.019054221,
-0.0156578757,
0.004415168,
0.0017092137,
-0.0051999772,
-0.0009290306,
-0.0016429268,
0.0321296789,
0.0013840733,
-0.0411189944,
0.0077747144,
-0.025501864,
0.0703329667,
0.0450364538,
0.0905944258,
0.0292166024,
-0.0372995771,
0.0580202825,
0.0578986779,
0.0386119373,
0.015025679,
0.018298395,
0.0288876705,
0.025690116,
-0.0023715508,
-0.003064343,
0.0282979086,
0.105042249,
-0.0488157049,
0.0127014518,
0.1345358044,
0.0249236375,
-0.0178224146,
-0.0202274378,
-0.0293447189,
0.0484040342,
0.0357441306,
0.0158592779,
-0.0586880259,
-0.0495029837,
0.0058824485,
-0.0093482239,
0.0160496701,
-0.0279734209,
0.0163846593,
-0.1273533702,
0.0214186106,
-0.0379723348,
-0.0728030503,
-0.0194511488,
-0.0290675927,
0.0072888872,
0.0218965802,
0.0274947472,
0.0232964139,
0.0114606284,
-0.0220667496,
0.032426253,
-0.0035334432,
-0.0602205321,
0.0636324659,
-0.0237627253,
0.0307696331,
-0.0031414372,
0.0374695212,
-0.0479269736,
-0.01653523,
0.0385549888,
-0.0227702614,
-0.036126852,
-0.0181258358,
-0.011163597,
-0.0029269885,
0.0009930782,
-0.0378238857,
0.0434825048,
-0.0163358897,
0.0123521024,
0.0636700764,
0.0881387442,
-0.0005945599,
-0.0473178141,
-0.016235115,
0.0007402922,
-0.0450175814,
0.0729656667,
0.0302269701,
-0.0574725382,
-0.0177476499,
-0.0347896256,
-0.0372736342,
-0.0239515249,
-0.0065118787,
0.0514662527,
0.0181430243,
0.0366988741,
-0.034105815,
-0.0587275773,
0.0477355421,
0.0354869477,
0.0009419587,
-0.0046297759,
-0.0017764762,
0.0080789141,
-0.0377063788,
0.021449605,
0.0101039456,
0.0155713223,
-0.0286912415,
0.0179491788,
0.0185592826,
-0.0182339773,
-0.0372044779,
0.0126900515,
-0.0934690163,
-0.0309490412,
-0.0150840953,
0.0070639346,
0.0122840144,
0.0027249916,
-0.0104454923,
-0.0058878767,
0.0098209549,
-0.0332954973,
0.0168801192,
-0.0053062751,
0.0045414516,
-0.0213588923,
-0.0082500409,
0.009031265,
-0.0207842719,
-0.0051394952,
0.0136840129,
0.0614544675,
-0.0157559607,
0.0056212293,
0.0403420329,
0.0288054254,
0.021986898,
0.0295016821,
0.0076055471,
-0.0389185622,
0.0450046957,
-0.028714966,
-0.0385516621,
-0.0109376693,
-0.0132440906,
0.0583571903,
-0.0588329174,
0.002103426,
0.0007681386,
0.0371778309,
-0.001170836,
-0.0000637684,
0.0291769765,
0.041923061,
0.0211070366,
0.0072450601,
-0.0047290288,
-0.0318330862,
0.0131770195,
0.0229650326,
0.0738361701,
-0.0055278693,
0.110329926,
0.0180884823,
-0.0143341683,
0.0273950025,
-0.0080711925,
-0.0669112802,
-0.0531934798,
0.0172138512,
0.040543776,
0.0440948755,
0.0091568381,
0.0187260192,
-0.0115225185,
-0.0039200489,
-0.0394621156,
0.0202254988,
0.0112025691,
-0.0227532294,
0.038982112,
0.0443973951,
0.0013455481,
0.0133427596,
0.0743947327,
-0.0008457439,
0.0307207573,
0.0128911817,
0.0090098865,
-0.049826324,
-0.0020458146,
0.0277538598,
0.0338604525,
-0.0181689542,
0.0247718263,
-0.0845735669,
-0.084043324,
0.012920497,
0.0191642158,
0.0168525055,
0.0106867049,
-0.0221878458,
0.0733638257,
-0.0264932439,
-0.0349598639,
-0.0336345881,
0.0172890704,
-0.0442006327,
0.0075121508,
-0.0046498664,
-6.70139852e-33,
-0.0368448347,
-0.0325910673,
0.0276813097,
-0.0749707818,
0.0118652107,
0.0004327841,
-0.0333328918,
0.0342390276,
-0.0551196001,
0.0036971893,
0.0119259804,
-0.0109850615,
-0.0029579273,
0.0387572348,
0.0024866394,
-0.0338627361,
-0.025249498,
0.0000722591,
-0.0133701283,
-0.0003883225,
0.0420042127,
0.0144999772,
0.0399330147,
-0.0091322856,
0.0516109802,
-0.0037535974,
0.0099783046,
0.048170004,
-0.0811285451,
0.0659023449,
-0.0121917976,
0.0173845701,
0.0490926728,
-0.0084694009,
0.0149763077,
-0.003399265,
-0.0628558174,
-0.0488751605,
0.0170940682,
0.0047173076,
-0.0452400744,
-0.0030912783,
0.0014327884,
0.0177406892,
0.0331462622,
-0.0433609374,
0.0493894145,
-0.0122121451,
-0.0018154396,
-0.0012793725,
-0.0133488663,
0.0109704062,
-0.063024126,
0.0063021374,
-0.0558201186,
0.0883116648,
0.0173235089,
0.0448981524,
-0.0349497385,
0.0003363265,
-0.0021357841,
0.0118242679,
0.0546715707,
0.0088427449,
0.0099648871,
0.0438397229,
0.0151126394,
-0.0267881677,
-0.0046734507,
0.0016361355,
0.0431251824,
0.0370772704,
0.0327441394,
0.0734540895,
0.0115168942,
0.0375645459,
-0.1096255407,
0.0216423664,
-0.0001548722,
0.0308359209,
0.0060079866,
-0.0310605913,
-0.0138031859,
0.0171015449,
-0.0756496266,
0.014417069,
-0.0525571667,
-0.0135271689,
0.0613326766,
-0.0580941103,
-0.0402160734,
-0.0175571013,
0.0611470714,
-0.0303807016,
-0.027439449,
0.0157433338,
0.011136964,
0.02912887,
-0.0136760492,
0.0850424394,
-0.038672857,
-0.0126167107,
-0.0144099323,
-0.0320002027,
0.0011416975,
0.0026639763,
0.020851586,
-0.0009312194,
-0.0398599096,
-0.0206975788,
-0.0455079786,
-0.0530078374,
-0.0015152049,
0.0454720855,
0.0287475605,
0.0099993525,
-0.0014399437,
0.007314919,
0.0290274341,
0.0080432342,
-0.0544093028,
-0.0344176255,
-0.0185062289,
-0.0107843066,
-0.0017047763,
0.0002231426,
-0.0298119131,
-0.0259260796,
0.0164635722,
-0.0431969352,
0.0136632146,
0.033335451,
2.703e-7,
0.0517810471,
0.0436057039,
0.0049479213,
0.0160050839,
0.013503979,
0.0462635718,
-0.0521062836,
0.004475385,
0.047419142,
0.0034882906,
0.0053154421,
0.0364742838,
0.033811897,
0.0477713794,
-0.0851490274,
-0.0023308708,
0.011790717,
-0.0094479034,
-0.0237242542,
0.0370296985,
0.0753236413,
0.0761895254,
-0.0473443791,
-0.005494542,
0.0156604443,
-0.0380214751,
0.0307204444,
-0.0128297992,
-0.0310548469,
0.0398205109,
0.0180131271,
0.0067559578,
0.0034403563,
0.0243212432,
0.0192686822,
-0.0057664909,
0.0389004312,
0.0616313927,
0.0279959925,
-0.0509550907,
-0.0144660156,
-0.0090265395,
-0.0011863222,
-0.0674617812,
0.0678728744,
-0.0151618384,
-0.0432326794,
-0.024887057,
-0.0270541292,
-0.0359248146,
-0.0017718052,
-0.0087090554,
-0.0124949608,
0.0053286469,
-0.0212036148,
-0.040394403,
0.0136600398,
-0.0424430408,
-0.0330454297,
-0.0009079242,
-0.041746188,
-0.0703627616,
-0.0043542087,
-0.0627259985,
0.0434921086,
-0.046204038,
-0.0206541084,
2.300417419e-34,
0.0251117703,
-0.0262942128,
0.0251767114,
0.0159568805,
0.0431904346,
-0.0326090343,
-0.0270489268,
-0.0298995934,
-0.0105230408,
-0.1156588346,
-0.0135807721
] |
http://arxiv.org/abs/2401.14003v1 | http://arxiv.org/pdf/2401.14003v1 | ConstraintChecker: A Plugin for Large Language Models to Reason on Commonsense Knowledge Bases | 2024-01-25T08:03:38 | 2024-01-25T08:03:38 | Quyet V. Do, Tianqing Fang, Shizhe Diao, Zhaowei Wang, Yangqiu Song | cs.CL | cs.CL, cs.AI | Reasoning over Commonsense Knowledge Bases (CSKB), i.e. CSKB reasoning, has
been explored as a way to acquire new commonsense knowledge based on reference
knowledge in the original CSKBs and external prior knowledge. Despite the
advancement of Large Language Models (LLM) and prompt engineering techniques in
various reasoning tasks, they still struggle to deal with CSKB reasoning. One
of the problems is that it is hard for them to acquire explicit relational
constraints in CSKBs from only in-context exemplars, due to a lack of symbolic
reasoning capabilities (Bengio et al., 2021). To this end, we proposed
**ConstraintChecker**, a plugin over prompting techniques to provide and check
explicit constraints. When considering a new knowledge instance,
ConstraintChecker employs a rule-based module to produce a list of constraints,
then it uses a zero-shot learning module to check whether this knowledge
instance satisfies all constraints. The acquired constraint-checking result is
then aggregated with the output of the main prompting technique to produce the
final output. Experimental results on CSKB Reasoning benchmarks demonstrate the
effectiveness of our method by bringing consistent improvements over all
prompting methods. Codes and data are available at
\url{https://github.com/HKUST-KnowComp/ConstraintChecker}. | ConstraintChecker: A Plugin for Large Language Models to Reason on Commonsense Knowledge Bases: Reasoning over Commonsense Knowledge Bases (CSKB), i.e. CSKB reasoning, has
been explored as a way to acquire new commonsense knowledge based on reference
knowledge in the original CSKBs and external prior knowledge. Despite the
advancement of Large Language Models (LLM) and prompt engineering techniques in
various reasoning tasks, they still struggle to deal with CSKB reasoning. One
of the problems is that it is hard for them to acquire explicit relational
constraints in CSKBs from only in-context exemplars, due to a lack of symbolic
reasoning capabilities (Bengio et al., 2021). To this end, we proposed
**ConstraintChecker**, a plugin over prompting techniques to provide and check
explicit constraints. When considering a new knowledge instance,
ConstraintChecker employs a rule-based module to produce a list of constraints,
then it uses a zero-shot learning module to check whether this knowledge
instance satisfies all constraints. The acquired constraint-checking result is
then aggregated with the output of the main prompting technique to produce the
final output. Experimental results on CSKB Reasoning benchmarks demonstrate the
effectiveness of our method by bringing consistent improvements over all
prompting methods. Codes and data are available at
\url{https://github.com/HKUST-KnowComp/ConstraintChecker}. | [
0.065240711,
0.056531921,
0.0061041792,
0.0413819663,
-0.0395142511,
-0.0027313454,
0.0009972327,
0.0242876001,
0.0622029006,
-0.0302694812,
0.0216695704,
0.0171247479,
-0.0161354784,
0.0025249785,
-0.0052156933,
-0.0057604024,
0.0228312928,
-0.0384346284,
0.0016546926,
-0.0127830151,
0.0062771905,
0.0422840975,
-0.0392391346,
0.0071888692,
-0.0491250977,
-0.0366981067,
0.0105681447,
-0.0417580567,
0.0207431614,
-0.0028441362,
-0.0193388853,
0.0597917549,
-0.0023451161,
0.007767065,
0.0000024292,
-0.065962933,
-0.009797764,
-0.0098893996,
-0.0301641989,
-0.0111217117,
0.0120887095,
-0.0330679268,
-0.0023025449,
-0.000853802,
-0.0490341,
0.0317142196,
0.0168659184,
0.0411443412,
0.0626259595,
0.0337918773,
-0.0097475406,
-0.0275389701,
-0.0004386588,
0.0176413339,
0.0421649851,
-0.0516017862,
0.0310681239,
0.0367131531,
0.0644433424,
-0.042446427,
0.0113811437,
0.033091411,
0.0418224744,
-0.0516281202,
-0.0411214605,
0.0289554689,
-0.0609740168,
-0.0126696695,
-0.0212174598,
0.0973390564,
0.0542081185,
-0.032939218,
-0.0056011272,
0.0008848256,
0.0206075106,
0.0653230846,
-0.0156033142,
0.0319390632,
-0.009847925,
-0.0204177368,
0.0206413697,
0.0008807058,
0.0170912035,
0.0378891602,
-0.0156963654,
0.0699628666,
-0.0091707325,
-0.0245366413,
0.016031472,
-0.0319206566,
-0.0248614009,
-0.0576511212,
0.0032385217,
0.0771013647,
-0.0696519539,
0.0179598741,
0.0008087501,
0.0308244806,
0.0201081224,
-0.0170125514,
-0.0191823747,
0.0528189801,
0.0040905233,
0.0577716716,
-0.021512527,
0.0701848492,
0.0098328842,
0.0064510973,
-0.0065444894,
0.0072936481,
-0.0752398744,
-0.0204110071,
-0.0217580833,
0.0116629899,
-0.013779304,
0.0026755217,
0.0112767778,
0.0163123589,
-0.035756398,
0.0676529258,
0.0086500682,
0.0000095536,
0.0216676313,
-0.0404493995,
-0.0600019135,
0.0252465513,
-0.0546555147,
0.001314526,
-0.0009262457,
0.0027047915,
-0.0043769265,
-0.0023114365,
0.0074623474,
-0.0243934039,
-0.0127295516,
0.0183596052,
0.0022163803,
0.0181129649,
-0.0236240681,
-0.0400671624,
-0.0012526128,
-0.0549094565,
-0.050576169,
-0.0697809607,
0.0082205823,
0.0155652845,
-0.0134779084,
-0.0153460335,
0.0167763047,
-0.0114090145,
-0.0071358341,
0.0751768127,
-0.0011290786,
-0.0064869956,
-0.0288280435,
0.0552391708,
-0.013053406,
0.0553216636,
-0.0193992294,
-0.0185023025,
0.0123736961,
0.023951143,
-0.002755787,
-0.0064519094,
-0.0029228108,
0.0018868994,
-0.0066864854,
0.0257629175,
0.0118196923,
0.0184066016,
-0.0055362904,
0.0262795631,
-0.0215154793,
0.006023326,
0.021260364,
0.0430525579,
-0.0151700675,
0.0006578005,
0.0283784922,
0.0170560721,
0.0129714934,
0.0037856787,
-0.067867063,
0.0158220436,
-0.0363746472,
-0.0054904288,
-0.031138828,
0.0243518911,
-0.0063783969,
-0.0123158311,
0.0040265354,
0.0075197602,
0.0024203337,
-0.0176963657,
0.0628278106,
0.0078066182,
0.0030059873,
0.0795998201,
0.0324807912,
-0.0469033346,
-0.0287363362,
-0.0068987976,
0.0579579659,
0.0443972237,
0.0715097338,
-0.0543220416,
-0.0449516214,
0.0069332276,
0.0011304145,
0.0580709204,
0.0192962848,
-0.0345421396,
0.0059567033,
-0.044691857,
-0.0229909308,
0.0245871097,
0.0122537911,
-0.0178559218,
0.0366411284,
0.0099017508,
0.0162410066,
-0.0162267871,
0.0114741987,
0.0071565784,
-0.0194391124,
-0.0144917024,
-0.0372475125,
0.0187526178,
0.0092232022,
0.0227779895,
0.0310194921,
0.0207266919,
0.0260377377,
-0.0201220084,
0.0016609526,
0.0321582071,
-0.0279743411,
-0.0603827648,
0.0656981394,
-0.0097830519,
-0.0425916426,
-0.0144882789,
-0.0165886786,
-0.0416995175,
0.0303741358,
0.0355181731,
0.0482145809,
-0.0316487662,
-0.030130107,
-0.0196231008,
-0.0732013211,
0.0221975856,
0.1118623093,
0.0207421593,
-0.0120444456,
0.0566121861,
-0.0274583511,
0.0047644642,
-0.0056097279,
-0.0161894765,
-0.0015445503,
-0.0178588647,
0.0184971653,
-0.0094842566,
0.0409471691,
-0.0550792105,
0.006556958,
-0.0369369201,
0.0140743041,
0.0121262148,
0.0087532056,
-0.0268787201,
0.0107036401,
0.0100856004,
0.0054069138,
-0.0305066984,
0.017871242,
0.007490614,
0.0195855368,
0.0034313221,
0.0483022071,
-0.0093217092,
0.0452475548,
0.0152394669,
-0.0374556333,
0.0316864997,
0.0245474856,
0.0663916618,
-0.0272153821,
0.0377110243,
-0.0468284748,
0.0117733143,
-0.0056049237,
-0.0026770984,
0.0189478174,
-0.0240755212,
0.0122073758,
0.0269900654,
-0.0365638249,
0.0025087057,
0.0556905232,
-0.0353669003,
-0.0593403019,
0.0280997138,
0.0091319177,
-0.0113458093,
-0.0055157309,
0.0102751702,
-0.0531952456,
0.0361054316,
-0.0278413836,
-0.0079450374,
-0.0787168071,
0.0435285605,
0.0167853702,
0.0149199804,
-0.0755082518,
-0.0074319579,
0.0101426756,
0.0661540553,
-0.0197395813,
-0.0127125531,
0.0859052166,
-0.0130411377,
0.0187408328,
-0.0149550997,
0.0235328749,
-0.0217750259,
0.0696958452,
-0.0307576265,
0.0047657797,
0.0355836302,
0.0035671673,
-0.0352574401,
-0.0071268622,
0.0528078601,
-0.0158516094,
0.0253706817,
-0.0444018655,
0.037675865,
-0.0066371611,
-0.0525392666,
-0.0589630306,
-0.0186527986,
-0.0456042588,
0.0148905031,
-0.0206740387,
-0.010899527,
-0.0057195881,
0.0130383885,
0.0030781261,
0.025540119,
0.0045955107,
-0.0042212289,
-0.0307130031,
0.0083278092,
0.0266521946,
0.0363146029,
0.0308414325,
0.0323718823,
0.0067807343,
0.0342620201,
0.0110490117,
-0.0386470482,
0.0107671758,
0.0458569154,
-0.0143277012,
-0.0250254627,
0.0259172656,
0.0228327662,
0.0105169034,
-0.0738654435,
-0.0258495566,
0.0863960758,
0.0600373968,
-0.0213846155,
-0.002205936,
-0.0530707873,
-0.004687326,
0.0101726595,
0.0589101836,
-0.1280035526,
-0.06050203,
0.008463379,
0.0043354873,
-0.0099359443,
-0.0632068813,
-0.0030181904,
-0.13861233,
-0.0022247683,
-0.0045542074,
-0.0793266818,
-0.0315163583,
-0.0373208672,
-0.0103780543,
-0.0114436587,
-0.0302402079,
0.0385856815,
0.0022784995,
-0.00953501,
-0.0201901626,
0.0172408838,
-0.071440883,
0.0336177349,
-0.0514048263,
0.0576526076,
0.0220418107,
0.022666093,
0.0369072035,
-0.077716887,
-0.0151296547,
-0.0570434853,
-0.0036557382,
-0.052964434,
-0.0451582298,
0.0125758667,
-0.0013148948,
-0.0365460999,
0.0482978113,
-0.0654888824,
-0.0565299168,
0.0239725243,
0.111249201,
0.0011781097,
-0.0019193252,
-0.0097926902,
0.0115292873,
0.0140365139,
0.0898366943,
0.0267202053,
0.032484144,
-0.0195735414,
-0.0602114946,
-0.0556973964,
-0.0223856457,
-0.0098317936,
0.0131594567,
0.0157924872,
0.0175166428,
-0.0591001101,
0.0111494586,
0.0590025634,
0.0484065786,
0.0125582004,
0.0023534112,
0.0024648451,
0.0122951949,
-0.0405731462,
0.0425406881,
-0.0510663167,
-0.0071867122,
-0.0221600328,
0.033546064,
-0.0016336854,
-0.0285927467,
-0.0224455427,
0.0101324897,
-0.0006849651,
-0.0634905025,
-0.0671901181,
0.0580163598,
0.0262493305,
0.0288443808,
-0.0645064488,
-0.0110713597,
0.022493124,
0.0673693269,
-0.0425027981,
-0.0531782657,
0.0089077139,
-0.0180494096,
0.0116886823,
0.0614457503,
-0.026239235,
-0.020615425,
-0.0090475157,
0.0651655942,
-0.0456196293,
0.0052676322,
0.0255381186,
0.0322388783,
-0.029768046,
-0.0001210194,
-0.0120506091,
-0.028145602,
-0.0031052616,
-0.0166215058,
-0.0723676756,
-0.0149260359,
0.0430653915,
-0.0100979414,
-0.0278022233,
0.0347634703,
-0.000244988,
-0.0201756228,
0.0323512517,
0.0022946591,
0.0105439108,
0.0815441012,
0.0067630508,
0.0562703498,
-0.0549775995,
0.0050828927,
0.0258922838,
-0.0054221191,
0.0388661325,
-0.0211064778,
0.0709285662,
0.0515896454,
-0.0187116712,
-0.0317279249,
-0.0353166945,
-0.0299624782,
-0.0112879649,
0.0464070477,
0.0388415195,
0.0409370475,
-0.0173459258,
0.0118499044,
-0.0223460421,
-0.0186521374,
0.0198917836,
0.0295027979,
0.0494934507,
-0.030071184,
0.0403786674,
0.0469904244,
0.0046765674,
0.0082041863,
0.0540828519,
-0.0540867746,
0.0474109389,
0.0285609998,
-0.0222502332,
-0.0373727493,
-0.0314675756,
0.0109180314,
0.0345546529,
-0.0193556715,
0.0430070795,
-0.0812603459,
-0.0523490161,
-0.0108856717,
0.0165907927,
-0.0368239395,
0.0248577781,
-0.0018551322,
0.0447839536,
0.0439657457,
-0.062665008,
-0.0096635129,
0.0057125208,
-0.0899241716,
-0.0216952469,
0.0140473424,
-6.737305464e-33,
-0.0175049342,
-0.0461612307,
0.0425615944,
-0.0245139208,
0.0191685688,
0.0073432093,
-0.0723840818,
-0.0252514165,
0.0034860093,
0.0321389809,
-0.0227323547,
0.0099437889,
0.0020492468,
-0.0217499863,
0.0097525408,
-0.016252527,
-0.0163250417,
0.0463058241,
-0.0438347161,
-0.003323378,
0.0654294193,
-0.0040327949,
0.0114161,
0.0051032999,
0.0394800976,
-0.0166685954,
-0.0103204837,
0.0537355579,
-0.0689144656,
0.0551253408,
-0.0240381341,
0.0431822985,
0.0202304143,
-0.0100323278,
0.0130276354,
0.0299436972,
-0.0980079472,
-0.0428732075,
0.0536847971,
-0.033215344,
-0.0152281057,
-0.0300468579,
0.001783998,
0.0147615001,
-0.0084607899,
0.0062018549,
-0.0002785149,
-0.0585813671,
0.0051805167,
-0.0243450366,
-0.0305842776,
0.0206383877,
-0.020989215,
-0.0334074832,
-0.0015489724,
0.0322454348,
-0.0140550788,
-0.0256229453,
-0.0444692485,
0.0477888845,
0.0446132198,
0.0054810988,
0.0642054752,
-0.023620639,
-0.0094940336,
0.0200164914,
-0.0898026451,
-0.0580674224,
-0.0464010648,
0.0115062827,
0.0084612221,
0.071975708,
0.0405229777,
0.0915652886,
0.0365574397,
0.0010414489,
-0.0630571917,
0.0415145792,
0.0044341846,
0.0759748295,
0.0241016373,
0.0023465999,
-0.0437535755,
-0.0371052399,
-0.0279434491,
-0.0032622276,
-0.0639530346,
-0.0000285771,
0.0213239118,
0.0122548537,
-0.0481416099,
0.0271625891,
-0.0054147937,
-0.0283742398,
-0.0158745926,
0.0626802519,
0.0229886286,
-0.0285708141,
-0.0119969016,
0.0325947851,
-0.044143796,
-0.0120045757,
0.047393579,
-0.0240904782,
0.0189458337,
0.0608860664,
0.0509608462,
0.0072583086,
-0.0048311031,
-0.0138566131,
-0.0394171588,
0.0207776539,
-0.0204665363,
-0.0121401176,
0.0272044744,
0.0240033716,
0.0258563496,
0.0283255447,
0.0027664555,
-0.0041682739,
-0.0054145944,
0.0481046289,
0.0211108737,
0.0064570615,
-0.0193870962,
0.0075814412,
-0.025197994,
-0.0098439828,
0.0121526551,
-0.0359399058,
-0.0076453355,
0.0227517243,
2.988e-7,
0.0037711866,
0.0345919318,
-0.0122516509,
0.0054062605,
0.0002020981,
0.0454117991,
-0.0110777868,
-0.0231897235,
-0.0191392936,
-0.0461409017,
0.0018834132,
0.0678485408,
0.0302172508,
0.0388291739,
-0.0654759631,
-0.0054326979,
-0.0439255834,
-0.0598605275,
-0.0767676607,
0.026662685,
0.0789336413,
0.0283667576,
0.0091718109,
-0.0121616609,
0.02585903,
-0.0815957636,
-0.0092123318,
-0.0203458313,
0.0099775419,
-0.0183380321,
0.0086644832,
-0.0117769223,
0.0165340491,
0.0725358054,
0.0094702309,
0.0090974402,
-0.0366262905,
0.0509446487,
0.0050862036,
0.0127558298,
-0.0124713825,
-0.047286287,
0.0174166393,
-0.0656273067,
0.1031749472,
-0.0330000408,
-0.0149524594,
-0.0105567817,
-0.0157163013,
-0.0638455674,
-0.0481746271,
-0.0109687988,
-0.0081102857,
-0.0231985003,
0.0042830245,
-0.0127523281,
-0.022464823,
-0.0588031262,
0.0085405791,
0.0301766638,
0.0103963893,
-0.0435647368,
-0.0162185412,
-0.0418794639,
0.0046096039,
-0.1023301035,
0.0073967017,
2.91202713e-34,
0.008565601,
0.0183033161,
0.0066350475,
0.0140064908,
0.0614901148,
0.0138982534,
-0.0959544033,
-0.004750465,
0.0396316908,
-0.0964338332,
-0.0002666844
] |
http://arxiv.org/abs/2401.13996v1 | http://arxiv.org/pdf/2401.13996v1 | Investigate-Consolidate-Exploit: A General Strategy for Inter-Task Agent Self-Evolution | 2024-01-25T07:47:49 | 2024-01-25T07:47:49 | Cheng Qian, Shihao Liang, Yujia Qin, Yining Ye, Xin Cong, Yankai Lin, Yesai Wu, Zhiyuan Liu, Maosong Sun | cs.CL | cs.CL, cs.AI | This paper introduces Investigate-Consolidate-Exploit (ICE), a novel strategy
for enhancing the adaptability and flexibility of AI agents through inter-task
self-evolution. Unlike existing methods focused on intra-task learning, ICE
promotes the transfer of knowledge between tasks for genuine self-evolution,
similar to human experience learning. The strategy dynamically investigates
planning and execution trajectories, consolidates them into simplified
workflows and pipelines, and exploits them for improved task execution. Our
experiments on the XAgent framework demonstrate ICE's effectiveness, reducing
API calls by as much as 80% and significantly decreasing the demand for the
model's capability. Specifically, when combined with GPT-3.5, ICE's performance
matches that of raw GPT-4 across various agent tasks. We argue that this
self-evolution approach represents a paradigm shift in agent design,
contributing to a more robust AI community and ecosystem, and moving a step
closer to full autonomy. | Investigate-Consolidate-Exploit: A General Strategy for Inter-Task Agent Self-Evolution: This paper introduces Investigate-Consolidate-Exploit (ICE), a novel strategy
for enhancing the adaptability and flexibility of AI agents through inter-task
self-evolution. Unlike existing methods focused on intra-task learning, ICE
promotes the transfer of knowledge between tasks for genuine self-evolution,
similar to human experience learning. The strategy dynamically investigates
planning and execution trajectories, consolidates them into simplified
workflows and pipelines, and exploits them for improved task execution. Our
experiments on the XAgent framework demonstrate ICE's effectiveness, reducing
API calls by as much as 80% and significantly decreasing the demand for the
model's capability. Specifically, when combined with GPT-3.5, ICE's performance
matches that of raw GPT-4 across various agent tasks. We argue that this
self-evolution approach represents a paradigm shift in agent design,
contributing to a more robust AI community and ecosystem, and moving a step
closer to full autonomy. | [
0.0012122929,
0.0321292914,
-0.0290940814,
-0.0392545946,
-0.0326824971,
-0.0116107464,
-0.0110124946,
-0.0286674201,
0.0608340353,
0.0139855612,
-0.0276965797,
0.058979217,
-0.0478572212,
0.0599277057,
-0.0130906282,
-0.0557597056,
0.0547468364,
-0.0390916094,
-0.0085186698,
-0.004768847,
-0.0335426368,
-0.0394113064,
-0.0029008333,
0.0220120307,
-0.0007241529,
-0.0240279343,
-0.0609026887,
0.0084087094,
0.0470762253,
-0.0008171209,
0.0126390588,
0.0022800069,
0.0192668494,
0.0415498279,
0.0000022086,
-0.0672862157,
-0.0267355591,
-0.0040522646,
-0.0084455609,
-0.0362736806,
0.0340878442,
0.0305369589,
-0.0037709156,
0.030445965,
-0.0536127016,
0.0216250438,
0.0349511243,
0.0132008269,
0.055812981,
0.0200498011,
-0.0088182222,
0.0571361333,
-0.0067849569,
0.0049734809,
-0.021606354,
-0.064535439,
0.011204496,
-0.0408613384,
0.0321908295,
-0.0325668454,
0.0108298203,
-0.028163977,
0.034094274,
0.0111042336,
0.0061582983,
0.0453868918,
-0.0011754688,
-0.0550883599,
-0.0601177365,
-0.006692756,
-0.0568426363,
-0.0163020249,
0.0397744514,
0.0083880452,
0.0335603282,
-0.0172998551,
-0.0024059489,
-0.030202603,
-0.00578041,
-0.0265078954,
0.019650599,
-0.0126707936,
-0.0005321734,
-0.0089479359,
-0.0480500087,
0.0782521144,
-0.0039131627,
0.0130343437,
-0.0120567065,
-0.0232991446,
0.0250258483,
-0.0296750218,
-0.0030813001,
0.0568432771,
-0.0681888238,
-0.0263882354,
-0.0467408784,
-0.0026089426,
-0.0294122659,
0.0247007813,
0.0476395003,
0.0108980564,
-0.0006003185,
0.0149729131,
-0.010707411,
0.0590273477,
-0.0183216259,
-0.0222185738,
-0.0196902324,
0.012492382,
-0.0926737785,
0.0240288284,
-0.05061239,
0.0534163192,
-0.0263700876,
-0.0002660716,
0.0354205333,
0.079890646,
-0.0319515951,
-0.0072766156,
-0.0186629947,
0.028681552,
-0.0410969369,
0.0020889416,
-0.0360396467,
-0.0181750339,
-0.041610878,
-0.0165649951,
0.0270066299,
0.0148261059,
-0.0226615835,
-0.026936654,
-0.0102496566,
-0.0054931371,
0.0074280063,
0.089965187,
-0.0215533562,
-0.0303269811,
-0.0343027227,
-0.0505143777,
0.0549569018,
-0.0248223729,
-0.0038982939,
-0.0104190689,
0.0502180122,
-0.0517409667,
-0.0221226402,
-0.0150753967,
-0.0312711112,
-0.0560236461,
-0.0085179461,
0.0468770899,
-0.0045099454,
0.0635652244,
0.0167124458,
0.0284550283,
-0.0230845399,
0.0698521957,
0.0067955884,
0.0381115116,
0.0100018289,
0.0262250267,
0.0538944453,
0.0219132975,
-0.0275701154,
0.0281083882,
0.1207037047,
-0.006069148,
0.0011394965,
0.0100390874,
0.0033149628,
-0.0247320812,
0.0035928576,
0.0105893323,
0.0342430845,
0.0338134579,
-0.0165605284,
0.017093759,
0.0215374827,
0.0502575971,
-0.0309936255,
-0.0161615126,
-0.0651925802,
0.0613361821,
-0.0138688162,
-0.0154707152,
-0.0550166182,
0.0573358648,
-0.0326745585,
-0.0330266058,
0.0342611149,
-0.0140647767,
0.011676305,
0.002084926,
0.0027447788,
0.0483488292,
-0.0480355285,
-0.0372267477,
-0.0066822884,
-0.0420531742,
-0.0339635238,
-0.1275586337,
0.0269774199,
0.0652429536,
0.0461658016,
-0.0548058636,
-0.0856950432,
0.0293504838,
-0.0283454526,
0.0052365963,
0.0063694934,
-0.0286647622,
0.0469253324,
0.0345002823,
-0.0505293831,
-0.0074924361,
0.0119135827,
0.0476805381,
0.0299425237,
-0.0168688558,
-0.0357699618,
-0.0016985198,
-0.0122756734,
0.0262289885,
-0.0339166149,
-0.0224268399,
0.0486213155,
0.0046579521,
0.056514997,
-0.0135105457,
-0.0148193296,
-0.0221001729,
0.0224306136,
0.0377265774,
-0.049030818,
0.0732546598,
0.0074085933,
0.0537380017,
-0.0035671433,
0.0126116434,
0.0280811992,
0.0144646885,
-0.0087732589,
-0.059477672,
0.0332855657,
0.0160167962,
0.0506727807,
0.0251024254,
0.0338523202,
-0.0256387088,
-0.0062579792,
0.0210219026,
0.0244681612,
0.031911213,
-0.1294970512,
-0.0070389886,
0.0421147645,
0.0168273468,
0.0175765101,
0.0378949903,
0.0101544112,
-0.0084673213,
-0.0034311963,
0.0476629809,
-0.0048502353,
-0.0083068348,
-0.0413974859,
-0.0665206909,
0.0391451307,
-0.006493323,
0.0063243825,
0.0322932154,
0.029174583,
-0.0190503933,
0.0063424343,
0.0195562001,
0.0636439845,
-0.006318429,
0.0053097159,
0.0222647544,
0.0090689575,
-0.0229867268,
-0.0482905544,
0.024650421,
-0.0361060165,
-0.0119639868,
-0.0043784664,
0.0252965223,
-0.0555442236,
0.0071401652,
-0.0097517343,
0.0253776908,
-0.0106987832,
-0.041973643,
0.0229015071,
-0.0382089987,
0.0065689781,
0.0844598413,
-0.0316782407,
0.0365075283,
0.100827992,
-0.0418925732,
-0.0164528042,
0.0477205068,
-0.0040152259,
-0.0112598613,
-0.018952759,
0.0285064802,
0.0407681353,
0.038533669,
-0.0488900505,
-0.0426771455,
-0.0396668725,
0.004558831,
0.0305845086,
-0.070514597,
0.0502966866,
-0.025905041,
-0.0162790045,
-0.010244431,
-0.0522453636,
-0.0303299688,
0.0556058809,
0.0290999506,
0.005585779,
-0.0564911477,
-0.0332393162,
-0.0138458516,
0.1032038555,
0.0093919327,
0.002092927,
0.0103123346,
0.001293314,
-0.0543877892,
-0.0172341559,
0.0158527493,
-0.0032633131,
0.0215522982,
-0.0756030232,
-0.0138154887,
0.0158173535,
-0.0550064631,
-0.0251688883,
0.0297121461,
0.0031283789,
0.0257584918,
-0.0272993222,
-0.0558663756,
-0.0495761335,
-0.0225965194,
0.018391598,
0.0550039448,
-0.0066012717,
-0.0393664762,
-0.0404228643,
-0.0806909129,
0.0312135462,
0.0514950529,
-0.02240468,
0.0300214533,
0.0732779801,
0.0195213836,
0.0693344474,
-0.0526526012,
-0.0429446362,
0.0077647301,
-0.0145360194,
-0.0649661943,
-0.002499274,
0.0176071022,
0.0428591967,
-0.0561312512,
-0.0155329788,
0.0429220572,
0.0326574668,
-0.0019481929,
0.0170876812,
-0.0451008826,
0.0285365488,
0.0215052795,
0.0751419887,
-0.0051344465,
-0.0130733792,
0.0137251653,
0.033543814,
-0.0254490785,
-0.0369123109,
0.027539555,
-0.0389377624,
0.0440691039,
-0.0162826795,
-0.020856047,
0.0433610156,
-0.065623872,
0.0308890603,
-0.0167058017,
-0.0495299548,
0.0312495008,
0.009010083,
-0.015490164,
-0.0154735995,
-0.02178929,
0.0109819733,
-0.0156219564,
-0.0381402746,
-0.0007247738,
0.0830573142,
0.0286676195,
-0.003762031,
0.0069326065,
-0.0218344778,
0.0135856541,
0.0449747257,
0.0044127638,
0.027139619,
-0.0011919655,
0.0083987266,
-0.0016732813,
0.0399273448,
-0.030129768,
-0.0279378053,
0.051634714,
0.0596577488,
-0.0218353532,
-0.0404538997,
-0.0556350127,
-0.0733727291,
0.0173587706,
0.0252494868,
-0.0614766702,
-0.0305224955,
-0.0610729009,
0.0248725358,
-0.0199649837,
-0.058029715,
-0.0243220944,
0.0111265928,
-0.0111485058,
-0.0539462008,
-0.0577823967,
-0.0516256057,
0.0437133089,
0.0412934087,
0.0325857922,
-0.0194040854,
-0.0106597235,
0.0624489784,
-0.0366395451,
-0.0609293804,
0.0220063608,
-0.0294758566,
0.0301189721,
-0.038940765,
0.0330779888,
-0.0149731133,
0.0236478709,
0.0147841014,
-0.0363065861,
-0.0297038071,
0.0093843509,
0.0945712775,
0.0228487123,
0.0411123373,
-0.0064898892,
0.02164183,
0.0695374832,
0.0427309051,
-0.0066758892,
0.0112873064,
0.029462263,
-0.0087745273,
0.0042204577,
0.0127526438,
-0.0684771165,
0.0049233241,
-0.0064747105,
0.0022618568,
-0.0292446148,
0.0094330525,
0.0126201035,
0.0046244031,
-0.0421430767,
-0.0196086299,
-0.0083484231,
-0.0037231389,
0.0011327137,
0.0177587233,
0.0144656431,
-0.0416915081,
0.0484702885,
-0.0521647856,
0.0170941371,
0.0721448138,
0.0527647622,
0.0438014306,
-0.0394745395,
-0.0077341916,
0.0000701557,
-0.012565895,
0.0326162018,
-0.0090892361,
0.0275358818,
0.0255005695,
-0.0035621861,
0.0270730015,
0.0177099928,
-0.001444813,
0.0297387801,
-0.0022876919,
0.0177202281,
-0.0073580192,
0.0633158162,
-0.1061799228,
-0.069991447,
0.0301808361,
0.0261330642,
0.0258660801,
0.0219187587,
0.0407759808,
-0.0588855296,
0.0108711133,
-0.0136526665,
-0.0208331384,
-0.0429185852,
0.0157900024,
-0.0539003834,
0.0272377115,
0.0203233957,
-0.0128463963,
0.0485174023,
-0.0481579378,
0.0860320851,
-0.0288931783,
-0.0303643961,
-0.0414059386,
-0.0542476065,
0.0164077096,
-0.0193552598,
0.0031817756,
-0.0392048247,
-0.0021409318,
0.0008107668,
-0.0075057521,
0.0528740659,
-0.0039476091,
-0.0294358637,
0.028987268,
0.0308210254,
-0.0049236566,
-0.0732630342,
0.0064511932,
-0.0733478814,
-0.0234053843,
0.0256090499,
0.029113153,
-5.843599884e-33,
0.0132379122,
0.0079086358,
0.0523226187,
0.0111611849,
-0.031599462,
-0.0304804146,
-0.0342797264,
0.0161100943,
0.0094904499,
0.0039923885,
-0.0298288334,
0.0282333437,
0.0095456066,
-0.041633632,
0.0520278774,
0.035739854,
0.0487801768,
-0.0234078057,
-0.0230373759,
0.0403910875,
-0.0222239178,
0.0122090261,
-0.0332270525,
-0.0406878144,
-0.0332991853,
0.0604943261,
-0.0053729285,
0.0197524484,
-0.0319073685,
0.0081688417,
0.0091117201,
-0.0068018199,
0.0007568399,
0.0014354815,
-0.0055474513,
-0.0115887122,
-0.0559034124,
-0.0576485135,
-0.0232591275,
-0.0387088098,
-0.0861820802,
-0.0390897505,
0.0190319326,
0.0269055311,
-0.0233413875,
0.0128935389,
0.0167821441,
0.0059288293,
0.0488283224,
0.0188925229,
-0.0677706748,
-0.0210551862,
0.0045576436,
-0.0276150871,
0.0384831205,
-0.0114715789,
0.0103539396,
-0.0742612332,
-0.0187180918,
0.0218588356,
-0.0338599905,
-0.0388027802,
0.0305743068,
0.0246008411,
0.0294713974,
-0.0093989493,
0.037433628,
0.0200946201,
0.0011749236,
-0.011469759,
0.0046863467,
0.0965703726,
0.0030362599,
-0.0050465865,
-0.05889735,
-0.0460900441,
0.009288162,
0.0660077929,
0.0627531409,
0.0103126904,
-0.0079880254,
0.003044469,
-0.0147372121,
0.0217235051,
0.0260745492,
-0.0108942445,
0.0001408291,
-0.0282522198,
0.0326981694,
0.0008420086,
-0.008397839,
-0.0098662488,
0.0016962434,
-0.0389022343,
0.0393842235,
-0.0376365855,
-0.0525789224,
0.0257054828,
0.0098205982,
0.0138750849,
0.0018150415,
-0.0702020004,
-0.0177285913,
0.0201921072,
0.0466548018,
0.0106288902,
-0.0173584558,
0.0760313272,
-0.0405553319,
0.0149078676,
-0.0346328057,
0.0333178453,
0.0346174426,
-0.0660514459,
-0.0003054737,
0.0074228127,
0.0312281586,
-0.028281441,
0.035416197,
-0.0227794629,
-0.065172717,
0.0644162968,
-0.0150320148,
0.0064823385,
-0.002000489,
0.0012155569,
-0.0457376875,
0.0163950268,
-0.0402599536,
-0.045298133,
-0.0337033123,
0.0232780445,
2.982e-7,
-0.0554679818,
0.0138751715,
-0.0110513568,
-0.0204476062,
0.0408272855,
0.0538091436,
-0.0300888456,
0.047888875,
-0.0335734263,
0.0106030721,
0.0357626975,
0.0286455527,
-0.0057554906,
0.0167025216,
-0.0404396653,
-0.0242667105,
0.064705506,
0.0036630055,
-0.0702205896,
-0.0264825597,
0.005232105,
0.0107841855,
0.0344011001,
0.0064111636,
0.0358299986,
-0.0194995124,
-0.016038144,
0.0179480202,
0.0490597188,
0.0434267484,
-0.0632691383,
0.0610399805,
-0.0364765115,
0.0152659304,
-0.015489094,
0.043098148,
-0.052842021,
0.0466658399,
-0.0248411559,
-0.0234808885,
0.0052787829,
0.0149430512,
0.0417329259,
-0.01229164,
0.0589888841,
-0.0332624465,
0.0052861748,
0.0110182902,
-0.0083104772,
0.0083373487,
-0.0657159835,
0.0275895782,
0.0059149894,
0.0111076599,
-0.002254776,
0.0151062906,
-0.0116860783,
-0.078869611,
0.0529929176,
0.026287064,
-0.0233398862,
-0.0317614861,
-0.0316138566,
0.0220395066,
0.0133827394,
-0.0231649633,
0.0099732354,
2.600102357e-34,
-0.0529123396,
-0.0085615972,
0.039841231,
0.0282826349,
0.0056991489,
0.0185611024,
-0.0196494404,
0.007063678,
0.0927420184,
0.0176901966,
-0.0381731205
] |
http://arxiv.org/abs/2401.13986v1 | http://arxiv.org/pdf/2401.13986v1 | Towards Consistent Natural-Language Explanations via Explanation-Consistency Finetuning | 2024-01-25T07:04:30 | 2024-01-25T07:04:30 | Yanda Chen, Chandan Singh, Xiaodong Liu, Simiao Zuo, Bin Yu, He He, Jianfeng Gao | cs.CL | cs.CL, cs.AI, cs.LG | Large language models (LLMs) often generate convincing, fluent explanations.
However, different from humans, they often generate inconsistent explanations
on different inputs. For example, an LLM may generate the explanation "all
birds can fly" when answering the question "Can sparrows fly?" but meanwhile
answer "no" to the related question "Can penguins fly?". Explanations should be
consistent across related examples so that they allow a human to simulate the
LLM's decision process on multiple examples. We propose explanation-consistency
finetuning (EC-finetuning), a method that adapts LLMs to generate more
consistent natural-language explanations on related examples. EC-finetuning
involves finetuning LLMs on synthetic data that is carefully constructed to
contain consistent explanations. Across a variety of question-answering
datasets in various domains, EC-finetuning yields a 10.0% relative explanation
consistency improvement on four finetuning datasets, and generalizes to seven
out-of-distribution datasets not seen during finetuning (+4.5% relative). Code
is available at https://github.com/yandachen/explanation-consistency-finetuning . | Towards Consistent Natural-Language Explanations via Explanation-Consistency Finetuning: Large language models (LLMs) often generate convincing, fluent explanations.
However, different from humans, they often generate inconsistent explanations
on different inputs. For example, an LLM may generate the explanation "all
birds can fly" when answering the question "Can sparrows fly?" but meanwhile
answer "no" to the related question "Can penguins fly?". Explanations should be
consistent across related examples so that they allow a human to simulate the
LLM's decision process on multiple examples. We propose explanation-consistency
finetuning (EC-finetuning), a method that adapts LLMs to generate more
consistent natural-language explanations on related examples. EC-finetuning
involves finetuning LLMs on synthetic data that is carefully constructed to
contain consistent explanations. Across a variety of question-answering
datasets in various domains, EC-finetuning yields a 10.0% relative explanation
consistency improvement on four finetuning datasets, and generalizes to seven
out-of-distribution datasets not seen during finetuning (+4.5% relative). Code
is available at https://github.com/yandachen/explanation-consistency-finetuning . | [
0.0394866019,
0.0376551263,
-0.0097970078,
0.0475478955,
-0.0108368928,
0.0132438121,
0.0102296192,
-0.0437905565,
-0.0611121766,
-0.0261325892,
0.0311151594,
0.0081799999,
0.027401397,
0.0011091181,
0.0438054204,
-0.0609883144,
0.0347060449,
-0.0425105952,
0.0114374012,
-0.0131115634,
0.0186415035,
0.0334635638,
-0.0179608054,
0.0524369814,
0.0522208959,
-0.0177050736,
-0.0447466262,
-0.0096462118,
0.0236704778,
0.0108129596,
0.0051499521,
0.0271757301,
0.0635917038,
0.0050636018,
0.0000021638,
-0.0275469534,
0.0308700614,
0.0109444577,
-0.0691834912,
0.0290054008,
-0.0396983027,
-0.015559203,
0.0281022061,
0.0144473696,
-0.0151531128,
0.031483151,
0.0541067719,
0.0388107374,
0.0765680522,
0.043091204,
0.0291639213,
-0.0398994349,
-0.0115718329,
-0.011674989,
0.0116499802,
-0.038496349,
0.0020749038,
0.0382851362,
0.0147023313,
0.0123791015,
-0.0279630274,
-0.0003249726,
0.0455840677,
-0.018078126,
-0.0230791867,
0.0364213623,
-0.0279554501,
-0.0307209194,
0.0196705163,
0.0657336861,
0.0089581413,
-0.0380317718,
-0.0076903291,
-0.0064955419,
-0.0070732762,
0.0241952538,
-0.0591327846,
0.0214250349,
-0.0068320134,
0.0167036206,
0.0178203657,
-0.0353329778,
0.0578563958,
0.0613100938,
-0.0121801374,
0.0714045167,
-0.018340053,
-0.0118125733,
-0.0239399895,
-0.010518779,
-0.0443771817,
-0.0637572259,
-0.0070119351,
0.0785537735,
-0.071008049,
0.0220524631,
-0.0098457616,
0.0108731054,
0.0274351165,
0.0097526647,
0.034840323,
0.0363173075,
-0.0161101818,
0.059956409,
0.0120660039,
0.0714468881,
-0.0114988983,
0.0268003754,
-0.0220186599,
-0.0320724249,
-0.055606436,
-0.019209452,
0.0065241386,
-0.0124535682,
-0.0341040082,
0.0008937321,
-0.0172059946,
0.0105792983,
-0.0423864499,
0.0222456306,
0.0027481907,
0.0281319451,
-0.028101949,
-0.0280315541,
-0.0580790676,
0.0260761958,
-0.0973084271,
0.0522642545,
0.0365971178,
-0.0597863346,
-0.0286331326,
0.0039907778,
0.008784052,
-0.013950143,
-0.0024392141,
0.0290823001,
0.0241448302,
0.0276722983,
-0.0159862675,
-0.0307284128,
-0.0199501291,
-0.0818612278,
-0.0491673499,
-0.0617662519,
0.0470108166,
0.0098226359,
-0.0193764493,
-0.0279020481,
0.0108183064,
0.0101814931,
-0.0406986102,
0.099192895,
-0.069134295,
0.0333247967,
0.0005488648,
0.0113139814,
-0.0060861348,
0.0442398116,
0.0179865658,
-0.0282790121,
0.023941895,
-0.0431960411,
0.0374582224,
-0.011516382,
0.0212322809,
-0.0199411828,
0.0125421369,
0.0356452838,
-0.0545292534,
0.0703892186,
0.0403860547,
0.0230462588,
-0.0185594093,
-0.0356051624,
0.034409672,
0.0150873205,
0.0039273761,
0.0102019478,
0.0722358674,
0.0192060135,
0.0058976659,
0.0474162921,
-0.0908485875,
0.0494135618,
-0.0033071111,
0.0163942017,
0.0241753422,
0.0117776738,
-0.0250723995,
-0.0678022504,
-0.01268626,
0.0132527305,
-0.0005122307,
0.017214939,
0.0320217572,
-0.0033667721,
-0.0149570284,
0.0368000418,
0.0459281355,
-0.0106254136,
-0.0042708553,
-0.0379044004,
0.0855191648,
0.0854365528,
0.0267181937,
-0.0196453296,
-0.0300069805,
-0.021368295,
0.0472096615,
0.0275928956,
0.0472025648,
-0.0179304872,
0.0356759131,
-0.0371291526,
-0.0418142043,
-0.0317786187,
0.0236229803,
-0.0041190861,
0.0348895602,
-0.0086974967,
0.025354594,
-0.0661603063,
-0.0163724925,
0.038107425,
-0.0169775337,
-0.0139848543,
-0.0128262844,
-0.0173107143,
0.0175171793,
0.0098744594,
0.0121397665,
0.0158943255,
0.0069789644,
0.0109111043,
-0.0082823541,
-0.003247895,
-0.0319284424,
-0.0373376794,
0.0392923914,
-0.0221223179,
-0.0515323952,
0.0398591124,
-0.0773313493,
-0.031611044,
-0.0089062816,
-0.0254867245,
-0.0032056023,
-0.0495522693,
-0.0105441073,
-0.0083742887,
0.0054980787,
-0.025165353,
0.0349266604,
0.0029817775,
0.0246845037,
0.0054348269,
0.0038312983,
0.0126856351,
-0.0527325422,
0.0495451279,
0.0089328,
-0.0249040369,
-0.0028733346,
0.0323675349,
-0.0124315461,
-0.0528821871,
0.0168096721,
-0.0405549817,
-0.0207357369,
-0.0284449104,
0.016471656,
0.0195662137,
-0.0194046013,
0.0160760395,
-0.0087625179,
-0.0163382981,
0.0195565112,
-0.006608448,
0.0126974629,
0.0452172086,
0.0011259178,
-0.0511578023,
-0.0524975099,
0.0073884251,
-0.0222327169,
0.0413288064,
-0.01272476,
0.0282650962,
-0.1075110286,
-0.0338022336,
-0.0644740388,
-0.0002126408,
-0.0161807146,
-0.0506933704,
0.0441736467,
0.0156447459,
-0.0215073079,
0.051277373,
-0.0060695214,
0.0296170153,
0.0538313687,
0.0295494106,
-0.0783510059,
0.0052122078,
-0.0288006179,
-0.0216961,
-0.0333888866,
-0.03729821,
-0.0090034232,
0.0746962056,
0.0017140176,
0.0051149288,
-0.0545709319,
0.0562072247,
0.0198776592,
-0.0156036466,
-0.0300360974,
-0.0251287073,
-0.0080094328,
0.0482276753,
-0.0442503802,
-0.0648392513,
0.1075769588,
-0.0328407437,
0.0018436098,
-0.0048206425,
-0.0056731757,
-0.0199677199,
0.0498306118,
-0.0250812359,
-0.0045535923,
0.0276453383,
0.0101197315,
-0.0309047103,
-0.058374241,
0.0490163863,
-0.0116311479,
0.0532831401,
-0.070787929,
-0.0064843674,
0.0047694836,
-0.0022281785,
-0.0462782234,
-0.0576405302,
-0.0644101053,
0.0159393493,
-0.0163308997,
-0.0256511793,
-0.016806297,
-0.020683594,
0.0431080274,
-0.0061687161,
-0.024923699,
0.0011083394,
-0.0245263688,
0.0034087256,
0.063041769,
0.0724865049,
0.0378620625,
-0.0207061954,
0.0350404158,
0.0205852464,
-0.0133063104,
0.0414640643,
-0.0376543887,
-0.0031143534,
-0.0162204485,
-0.0159096774,
0.0055464562,
0.024014065,
0.0719812214,
-0.08278808,
0.0082673272,
0.1288878769,
0.0112856934,
0.0007556315,
0.0212200079,
-0.026815068,
0.0287760142,
0.0379426591,
0.0233104285,
-0.0526334234,
-0.0571809486,
0.0028558145,
0.0353973359,
0.03217455,
0.0082326801,
-0.0120160598,
-0.079379499,
0.0248099696,
0.0123779392,
-0.0327174664,
-0.0469340459,
-0.0829327032,
0.0169164408,
-0.009633868,
0.0278870035,
0.0199386869,
0.0263911914,
0.0287710745,
0.0247542243,
-0.0215086751,
-0.0633773282,
0.0614784881,
-0.0568290353,
0.053676188,
0.0564327165,
0.0469045974,
-0.0034752251,
-0.0529665537,
0.0007199898,
-0.0274310336,
0.0147299459,
-0.0164187439,
0.0481687672,
0.0184561219,
-0.011782486,
-0.0486901328,
0.0403620228,
-0.0154799893,
-0.0315651558,
0.0004274789,
0.0813858956,
-0.0171707235,
-0.0524018221,
-0.0044842591,
0.0193596594,
-0.0095228944,
0.1114089265,
0.0553816259,
-0.009194525,
-0.0355924219,
-0.0233914834,
-0.0427364521,
-0.0387914404,
-0.0274586156,
0.0172588099,
0.0376336426,
0.0089090392,
-0.0347899385,
-0.0287353899,
0.0482564196,
-0.0031414446,
-0.0117585156,
0.0162716992,
-0.0134040173,
-0.0125468615,
-0.0113086412,
0.0355491452,
-0.031686604,
0.0464608893,
0.0078851916,
0.002088639,
0.0080199353,
-0.0173820574,
-0.0008582588,
-0.001466381,
-0.0007309528,
-0.0491715856,
-0.0165890921,
0.03092096,
0.027263429,
0.0630715489,
0.0036318931,
-0.0728938654,
-0.027822705,
0.0124017633,
-0.0228445325,
-0.01091123,
0.0350858457,
-0.0113312649,
0.0160424076,
0.0657164901,
-0.0406543724,
0.0060234219,
-0.0186634064,
0.0544942282,
-0.0315159746,
-0.0244663898,
0.0375373214,
-0.0139716789,
0.0253546592,
-0.0032250753,
-0.00923065,
-0.0051254984,
0.0102687739,
-0.0014384824,
-0.0209135134,
-0.0171976574,
0.0253711473,
-0.0154332574,
-0.043618504,
0.0103838863,
-0.0472549237,
-0.0270459298,
0.0006449698,
-0.0418582298,
-0.0025295322,
0.006314165,
-0.0079445839,
-0.0044040638,
0.0155773712,
-0.0349281207,
0.0114455065,
-0.0418226197,
0.0549367517,
0.0028853032,
0.0596492663,
0.0363483094,
0.0101214852,
-0.0141921332,
-0.0297221616,
-0.052948162,
-0.0482132398,
0.014024158,
-0.0032066545,
0.0257302485,
-0.0047350256,
0.048790805,
0.0159162078,
-0.0232471637,
-0.0124302488,
0.0533780567,
0.0101326909,
-0.0013080176,
0.0529180691,
0.0443891585,
-0.0400451384,
-0.0086816996,
0.0475141145,
0.0098507563,
0.0306658298,
-0.0080585908,
-0.0042470698,
0.0088258972,
-0.0194884762,
-0.03367633,
0.0301848501,
0.0064294031,
0.0122561771,
-0.0323789082,
-0.0651612431,
-0.036040213,
0.0443217233,
-0.037478812,
-0.0281383954,
-0.0004645081,
0.058534611,
-0.0179841388,
-0.0374550708,
-0.0114426157,
0.0293494407,
-0.0291676186,
-0.000819438,
-0.030816542,
-6.73913997e-33,
-0.0148389013,
-0.0436047539,
0.0130583262,
-0.0303997602,
-0.0133083472,
-0.0164294504,
-0.0720014349,
-0.0294599216,
-0.0463476516,
0.07991115,
0.0136977928,
0.025202062,
0.0069634751,
0.002939387,
0.0038900261,
-0.0595254265,
0.0290568564,
0.0530195311,
-0.0211325027,
0.0202153921,
0.0350808986,
-0.0167575032,
0.0586438477,
-0.0743629411,
0.0584267303,
0.0130714159,
-0.0309115052,
0.0382181667,
-0.0478025116,
0.0578752756,
-0.0030235851,
0.044912003,
0.0219494998,
-0.0496148132,
-0.0060987002,
0.0110088168,
-0.0904644206,
-0.0134949777,
0.0793401003,
-0.0019113385,
-0.0825984702,
-0.0398320556,
0.0500871874,
-0.0109254969,
0.0081608929,
-0.027019877,
0.0247510448,
-0.0229916722,
-0.0042203004,
-0.0002892739,
-0.035175126,
0.0177247934,
-0.0183576867,
-0.054839544,
-0.0146952784,
0.0472835936,
0.0030138737,
0.0132105853,
-0.0128913978,
-0.0157624502,
0.0508654527,
0.0166997015,
0.0723404661,
-0.0146679999,
-0.0108424854,
0.0684854165,
-0.0042948863,
-0.0287070405,
-0.0701893643,
-0.0178203788,
-0.0257081706,
0.0665422827,
0.0955769941,
0.0560820028,
0.0251591019,
0.0293206181,
-0.0577981435,
0.0328234546,
-0.0308171548,
0.0489184186,
0.0226358809,
0.0135774696,
-0.0404873975,
-0.0314966962,
-0.0851375088,
0.0157114994,
-0.0386148095,
-0.012559481,
0.0323791131,
-0.0085416734,
-0.0661575198,
0.0026342557,
-0.0121305091,
-0.0546183996,
-0.0159264021,
0.0230469257,
-0.0221785307,
-0.0436409637,
0.0280998852,
-0.0039538019,
-0.0110923201,
-0.0063818679,
0.0090858405,
-0.0053452626,
0.0430047438,
0.0235402174,
0.0433572754,
0.0007172433,
-0.0144148832,
-0.0072064605,
-0.0659578815,
-0.0133120622,
-0.0379019417,
-0.0362422764,
0.0328322202,
0.0319773592,
-0.0408527926,
0.016842939,
0.013609983,
-0.0241349917,
-0.0019476723,
0.0226028375,
-0.0392973423,
0.0281267129,
-0.018496247,
-0.0072193444,
-0.0378505476,
-0.0326984636,
0.0115887793,
-0.0383664705,
-0.0264768172,
0.0433306806,
2.602e-7,
0.0188840553,
0.0308175236,
0.0357372351,
0.006699455,
0.0288151335,
0.04259675,
-0.0080552753,
-0.0263326019,
-0.0058258069,
-0.0789028555,
-0.0143477218,
0.051730711,
0.0577136502,
0.0167546682,
-0.0149018308,
0.0034189683,
-0.0521153286,
-0.0044266605,
-0.0358791575,
0.0124291265,
0.0687172785,
0.0671023577,
0.029817054,
-0.000147133,
0.0141593162,
-0.0279830787,
0.0315038711,
-0.0325021818,
0.0277133156,
0.0091495914,
0.0129208378,
0.0015122794,
0.0506082028,
0.010415894,
-0.0004827404,
-0.0198637024,
-0.0260817725,
0.0742491856,
-0.0200383104,
0.0173820443,
-0.0705389902,
-0.083302334,
0.0382793248,
-0.0776850954,
0.0775931776,
-0.078836605,
-0.0278275795,
-0.0164975282,
-0.009209428,
-0.0239867941,
-0.0001358,
0.0014660603,
0.0101487217,
0.0247987043,
-0.0063869124,
-0.0005162231,
-0.0051787305,
-0.0629693642,
-0.0594126061,
0.0153289922,
-0.0268600732,
0.0049045244,
-0.0273458287,
-0.0249728803,
0.0518479235,
-0.0681447461,
-0.0046819667,
2.077574455e-34,
0.0174147971,
0.0234285537,
0.0303023607,
0.0241713971,
0.0479961038,
0.0208121575,
0.0005068739,
-0.0030148367,
0.0478567109,
-0.0547749102,
-0.0173264183
] |
http://arxiv.org/abs/2401.13979v1 | http://arxiv.org/pdf/2401.13979v1 | Leeroo Orchestrator: Elevating LLMs Performance Through Model Integration | 2024-01-25T06:45:32 | 2024-01-25T06:45:32 | Alireza Mohammadshahi, Ali Shaikh, Majid Yazdani | cs.CL | cs.CL, cs.AI, cs.LG | In this paper, we propose an architecture to harness the collective knowledge
of multiple trained LLMs to create a new state-of-the-art. At the core of this
framework is a LLM-based orchestrator that is adept at picking the right
underlying LLM experts for optimal task execution. Inspired by self-play in
reinforcement learning, we created a loop of query generation, orchestration,
and evaluation to generate training data for the orchestrator. Our evaluation
focused on the MMLU benchmark, employing models with 7B, 13B, and 34B
parameters available on Hugging Face. The results demonstrate new
state-of-the-art open-source models: Our Leeroo orchestrator achieves
performance on par with the Mixtral model while incurring only two-thirds of
its cost. Moreover, increasing the allowed cost surpasses Mixtral's accuracy by
over 5% at the same cost level, reaching an accuracy of 75.9%. Further
enhancements were observed when integrating GPT4 into the underlying model
pool. The Leeroo orchestrator nearly matches GPT4's performance at half the
cost and even exceeds GPT4's results with a 25% cost reduction. These findings
illustrate the potential of our architecture in creating state-of-the-art and
cost-effective LLMs by optimizing the synergy between multiple LLMs to achieve
superior performance outcomes. | Leeroo Orchestrator: Elevating LLMs Performance Through Model Integration: In this paper, we propose an architecture to harness the collective knowledge
of multiple trained LLMs to create a new state-of-the-art. At the core of this
framework is a LLM-based orchestrator that is adept at picking the right
underlying LLM experts for optimal task execution. Inspired by self-play in
reinforcement learning, we created a loop of query generation, orchestration,
and evaluation to generate training data for the orchestrator. Our evaluation
focused on the MMLU benchmark, employing models with 7B, 13B, and 34B
parameters available on Hugging Face. The results demonstrate new
state-of-the-art open-source models: Our Leeroo orchestrator achieves
performance on par with the Mixtral model while incurring only two-thirds of
its cost. Moreover, increasing the allowed cost surpasses Mixtral's accuracy by
over 5% at the same cost level, reaching an accuracy of 75.9%. Further
enhancements were observed when integrating GPT4 into the underlying model
pool. The Leeroo orchestrator nearly matches GPT4's performance at half the
cost and even exceeds GPT4's results with a 25% cost reduction. These findings
illustrate the potential of our architecture in creating state-of-the-art and
cost-effective LLMs by optimizing the synergy between multiple LLMs to achieve
superior performance outcomes. | [
-0.0120033678,
0.0313572437,
-0.0015145055,
-0.0080489479,
-0.0099031888,
0.0000106317,
0.0001737569,
-0.0134617779,
-0.0027910653,
-0.0368587524,
-0.0769180357,
0.052982077,
-0.0813159123,
0.0473327227,
0.0122087579,
-0.0124431513,
0.009266993,
-0.0362063609,
-0.0500843897,
-0.0161752719,
-0.0396508612,
-0.0080480771,
-0.0071769804,
0.0042915153,
0.0238366146,
0.0137493964,
-0.0635244697,
0.072262913,
0.0064879917,
-0.059684936,
0.0164242052,
0.017744787,
0.0261525828,
0.0500489511,
0.0000022409,
-0.0482049324,
-0.0636567101,
0.00593204,
-0.0217540879,
-0.012210371,
0.0243603624,
0.0214762874,
-0.0005134226,
0.0562946834,
-0.0451224372,
0.0540586524,
0.0244719144,
0.0486010127,
0.0547180213,
0.0644658804,
0.0105731422,
0.0215903092,
0.0150862839,
-0.0120241782,
-0.0140340514,
-0.0507284179,
-0.0063734823,
0.0258732252,
0.0365008377,
-0.041932568,
-0.0088618668,
0.0149875246,
0.0317317769,
0.0152564244,
-0.013556662,
0.0140224518,
-0.0184407365,
-0.0030767913,
-0.0545028746,
0.0020970921,
-0.0029708897,
0.0105734505,
0.0387537181,
0.0080686957,
-0.0141666131,
0.0157353021,
-0.0344571918,
0.011987363,
0.0121841934,
0.019238174,
-0.0621694699,
0.0076355343,
0.0416673794,
0.0226025414,
-0.041748561,
0.0732560158,
-0.0100622373,
-0.0014378787,
0.0264419578,
-0.0045091887,
0.0308443531,
-0.0389829949,
-0.0106698228,
0.0558529906,
-0.0824202895,
-0.0008990936,
-0.0244226661,
0.0332340039,
-0.012034283,
0.0053904988,
0.0640337765,
0.0024556902,
-0.0173231047,
0.0378846712,
-0.042973306,
0.0196963307,
-0.052376084,
0.0325337723,
-0.0226613618,
-0.0014805803,
-0.075388357,
-0.0043080631,
0.0082806321,
0.0636459887,
-0.0439039581,
-0.0282969642,
-0.0370756611,
0.063164942,
-0.0227566417,
0.0315763839,
0.0128415888,
0.0302452277,
0.0123826452,
-0.0009779797,
-0.090823777,
-0.0466186628,
0.0120998602,
0.0314865969,
0.0063795634,
-0.0635821521,
-0.0047207135,
0.0014742707,
-0.0513535887,
0.0189290196,
0.0296972226,
0.0677065477,
0.0077108466,
-0.0009189329,
-0.0509209894,
-0.0586041547,
0.043821115,
-0.0620140359,
-0.0339388177,
-0.0418164581,
0.01145228,
-0.0518019684,
-0.0201287419,
-0.0027433105,
0.0065622763,
-0.0219831783,
0.022028489,
0.0642261878,
0.0189364031,
0.0197449587,
0.0143621545,
0.0085502295,
-0.0022188281,
0.1089492217,
0.0296747759,
0.0666852295,
-0.0034065242,
0.0290533155,
0.015122938,
0.0309106279,
-0.0151300896,
0.005831569,
-0.0091278385,
0.0111734122,
-0.0217386466,
0.0743310973,
0.0261988025,
-0.0212102924,
-0.031917911,
0.0034018403,
0.044136852,
0.0126443226,
0.0497527383,
0.0066830493,
0.0094938371,
0.0149681941,
-0.0212032087,
-0.0660584569,
-0.0637601316,
0.0712350011,
-0.0144382045,
0.004190688,
-0.0010670301,
0.0440991633,
-0.0346484147,
-0.0543040782,
-0.008063836,
-0.0036539386,
-0.0887963176,
0.0504530743,
0.0395808816,
-0.0295460839,
-0.0745999739,
0.0364021771,
0.0096569778,
-0.012967214,
-0.0301560294,
-0.0693228915,
0.0037068159,
0.0636369362,
0.0479090847,
-0.0640344918,
-0.0687487498,
-0.0329563096,
-0.0215565804,
0.0466133058,
0.0503431074,
0.0027626525,
0.0074822172,
-0.0046250299,
-0.0907482877,
0.0040772054,
0.0109841339,
0.0665056929,
0.0304946713,
0.0102581512,
-0.0259559415,
-0.0198718142,
-0.0325124897,
0.0379533432,
-0.0271080825,
-0.0643710047,
0.0148416469,
-0.0034363642,
0.0537360646,
0.017572917,
-0.009052394,
-0.0141489431,
0.0382958502,
0.0069309585,
-0.06432271,
-0.0158920959,
-0.0218505803,
0.0270796046,
0.0050982204,
0.0241308492,
0.0352832265,
-0.0234234668,
-0.0660450608,
-0.0594242029,
-0.0135969073,
0.025428487,
0.0194874369,
-0.036894992,
-0.0179481823,
0.0158454739,
-0.028647963,
-0.0340474024,
0.0303646009,
-0.0060824943,
-0.0404776148,
-0.0317037813,
0.031010136,
0.0270997733,
-0.0675833672,
0.030101778,
-0.0019032407,
0.0218983982,
0.0107328137,
-0.0015344303,
0.0174444504,
-0.027096387,
-0.0648188815,
-0.0721950233,
0.0511481352,
-0.0141061815,
0.0213668831,
0.0135369757,
-0.0138572985,
-0.0280278362,
-0.0175008643,
-0.012365669,
-0.0117749348,
-0.0054283687,
-0.0131406579,
-0.0003974702,
0.0046055764,
-0.0250872318,
-0.0125728669,
0.0018475342,
0.0013943053,
0.0003296222,
-0.0019939081,
0.0034976753,
-0.0267233942,
0.0078023439,
-0.0473445803,
0.0228617266,
-0.0003664759,
-0.0196208749,
0.0199164189,
-0.0287708919,
-0.0620008335,
0.074401632,
-0.0301008001,
-0.0085377665,
0.0761864111,
-0.0143743949,
-0.0808044299,
0.0407021903,
-0.0346427262,
-0.0566608757,
-0.0394120365,
0.0286030062,
0.0329778492,
0.0106606344,
-0.0342006907,
0.0143897226,
-0.0254001636,
0.0024081655,
0.0695830211,
-0.090302445,
-0.0119934101,
0.0203704853,
-0.0493633561,
0.043284893,
-0.0055278419,
-0.0392650403,
0.0202954188,
0.0228214413,
0.0014166749,
-0.0032897354,
-0.0091642421,
-0.0616272129,
0.0432242639,
-0.0063969777,
0.0144689642,
-0.0396161191,
-0.0386240669,
-0.0425583906,
-0.0068088039,
0.0285040885,
-0.0517350957,
0.0206554756,
-0.0503147431,
0.0099928332,
0.0079287188,
-0.0568799414,
-0.030907914,
0.0779968873,
-0.0488246307,
0.0339604169,
-0.0578364916,
-0.0363285206,
-0.0481677577,
0.0022286316,
0.0065165223,
0.0392453969,
-0.0623984076,
-0.0231423154,
-0.0262676254,
-0.0289700963,
0.0170240253,
0.0266031604,
0.0231110323,
-0.0179930162,
0.0739712641,
0.0428513102,
0.0590367913,
-0.064761363,
-0.05767015,
0.0317313299,
0.0337808356,
-0.050820414,
0.0015433014,
0.0318016857,
0.0026928762,
-0.0532664731,
0.0318517759,
0.0836710706,
0.0100975148,
-0.0213696305,
-0.0280739907,
-0.0286510773,
0.0495680533,
0.0439090207,
0.0543745086,
-0.047747694,
-0.0211689211,
0.023018159,
0.0801457167,
0.0003395538,
-0.0511491373,
0.0018334599,
-0.079207845,
0.084206894,
-0.0156625398,
-0.0255445372,
0.0101729976,
-0.0367822796,
0.000047554,
-0.0651929528,
-0.0056162132,
0.0402256474,
-0.0151926214,
-0.0214262642,
0.0435322225,
0.0440655984,
-0.0086914767,
-0.012995719,
-0.0438307673,
0.0147386156,
0.0702376887,
-0.0019311982,
0.0028225244,
0.0126798786,
0.010458068,
0.034148436,
0.0248650871,
-0.0347093567,
0.0735984072,
-0.0178610552,
0.0194422901,
0.0057630427,
0.0248782933,
0.0084208185,
0.0211746711,
0.0214402247,
0.0723282397,
-0.0248871762,
0.002295305,
-0.0004993264,
-0.0454398356,
-0.0131527763,
0.0368707962,
-0.0487813577,
0.0176008586,
-0.0801943243,
-0.0021598306,
-0.0234563909,
-0.042136699,
-0.0213255379,
0.0000856202,
-0.0291811042,
-0.007159031,
-0.0207969528,
-0.0065900106,
0.0028126403,
0.0619184822,
0.0094579561,
0.0232319068,
-0.0235472731,
0.0271573067,
-0.0324923098,
-0.012288047,
-0.0189181864,
0.006743819,
-0.068778187,
-0.0020731869,
0.0497142486,
-0.0459484011,
0.0019752111,
-0.0326836482,
-0.0444904231,
-0.0398905911,
-0.0102402959,
0.0872659758,
0.0271672383,
0.0134961577,
0.0163416266,
-0.0100895204,
0.0871783346,
0.0748310387,
-0.0721968263,
0.013823214,
0.0396603867,
0.0165315643,
0.0018084714,
0.0198961832,
-0.0816630423,
0.0157410819,
0.0159243718,
-0.0194428377,
-0.0342875756,
-0.0204487201,
0.0030763552,
-0.009650615,
-0.0022317457,
-0.0019659635,
0.0170832127,
-0.0010097079,
0.0420405865,
0.0008331412,
0.005535481,
-0.0438878983,
0.0625525489,
0.0027709855,
-0.026552625,
0.0470757782,
0.0161591116,
0.0764334872,
-0.0169870108,
-0.0197374485,
0.0163941421,
0.0053942669,
-0.0147745265,
0.0257568862,
0.0168449804,
0.0129986946,
0.0218898412,
0.0458856486,
0.0541427322,
0.0002289452,
0.0144660901,
0.0159489345,
0.0360664986,
-0.0121310325,
0.0130403545,
-0.0880837664,
-0.1148895547,
0.0033426108,
0.0043979399,
0.0608925447,
0.020783674,
0.0438724719,
-0.0464136414,
0.0029900337,
0.0081032366,
-0.0261345636,
-0.0206909515,
0.0758319646,
0.0465831868,
0.0263395924,
0.0140614426,
0.0067833108,
0.0413618609,
0.0168415885,
0.068747364,
-0.0388113819,
-0.0040109809,
-0.0604925156,
-0.0402048416,
0.0611225031,
0.0200137813,
-0.0055574067,
-0.0258088615,
-0.0349964611,
0.02303911,
0.0142143304,
0.0212432183,
-0.0094549563,
-0.0539627969,
0.0484031476,
0.0373203456,
0.0080103287,
0.0063807317,
-0.0068716099,
-0.0621066391,
-0.0089370972,
-0.0058567659,
0.0346602052,
-6.17918442e-33,
-0.014682713,
-0.0187948272,
0.0692125559,
-0.0004007329,
-0.0106742205,
-0.0167205352,
-0.0220083911,
-0.0265058577,
0.0118060019,
-0.0175631102,
-0.0218128711,
0.0034812549,
0.0156566277,
-0.0268201269,
0.0806782469,
-0.0313067771,
0.0619026981,
-0.032145109,
-0.0201104358,
-0.0048276829,
-0.0128698964,
0.0439900495,
0.004138812,
-0.0315907449,
-0.036474321,
0.0427390561,
-0.0316943862,
0.0202353876,
-0.0588574708,
0.0305461884,
0.01462752,
-0.0047645979,
-0.0245779306,
-0.0346111022,
-0.0164086558,
-0.049378939,
-0.0366920307,
-0.0454535745,
-0.0318284705,
-0.0217125155,
-0.0252613556,
0.0099243186,
0.0588481352,
0.0306506101,
-0.0444464758,
-0.022080278,
0.0182507131,
-0.0105472356,
-0.0199437402,
-0.0499363206,
0.0000136707,
0.0291820485,
0.0160446372,
-0.0047362107,
0.1081980616,
-0.0152049595,
0.0009222528,
0.0164244436,
-0.0333013162,
0.0258119013,
-0.0220678281,
0.0204096492,
0.0534493774,
0.0821580663,
0.0144440001,
0.0393570103,
0.0062885312,
-0.0085519021,
-0.0135852601,
0.0031645889,
-0.0101893339,
0.1365389228,
0.0152854817,
-0.0017933596,
-0.0230969526,
0.0211330932,
-0.0203402955,
0.0235492624,
0.0178256482,
0.0197247211,
0.0072019082,
0.0013186185,
-0.0376856513,
0.0002720026,
-0.0030434262,
-0.0151496483,
-0.0056953235,
-0.0241240021,
0.0368041135,
0.0260591246,
0.0289820898,
-0.0345250592,
-0.0067082695,
-0.0111261597,
0.0256621782,
-0.0020612576,
0.0182718541,
0.0028538979,
0.0145330783,
0.0102091646,
-0.0760540813,
-0.0496362001,
-0.0444650352,
-0.0207052454,
-0.0066165207,
0.0050819088,
-0.0061726444,
0.0434887595,
-0.0646851361,
0.0306042787,
-0.0311520919,
0.0315220617,
0.027984906,
-0.0384190865,
0.0181068089,
-0.0318255462,
0.0213503521,
0.0425393693,
-0.0058894595,
-0.0127638206,
-0.0063368795,
-0.0031594634,
-0.0174360741,
0.0110846749,
0.0100701004,
0.0361583121,
-0.0435232259,
0.007133299,
-0.0192368235,
-0.0191800129,
-0.0216855798,
0.0309044421,
2.939e-7,
0.010825933,
0.0465448536,
-0.0045342743,
0.0334704407,
0.0108412178,
0.0461735688,
-0.0245211311,
0.0006596803,
0.0091641257,
-0.0080293827,
0.0662687346,
0.0481486544,
0.0506101064,
0.0004083978,
-0.070864059,
-0.0196755342,
0.0190769136,
0.0196113028,
-0.0700783134,
-0.0543078892,
0.0061985934,
0.0384464003,
0.054915458,
-0.0212580077,
0.0481391288,
0.0113810692,
0.0210542325,
-0.0332362019,
0.0197745748,
0.0347098783,
0.0309074223,
0.0174350813,
-0.0346451551,
-0.0060605351,
-0.0044032354,
0.0718522593,
-0.0367653705,
0.0414168835,
0.009256674,
-0.0477231443,
0.0237442441,
-0.035861589,
0.0205915514,
-0.0370895974,
0.0539856702,
0.0242169444,
0.0092968987,
0.0104760127,
-0.0182059556,
-0.0507095456,
0.0057027764,
0.0056072781,
0.002161344,
0.0074581401,
-0.0149889439,
-0.0255547836,
-0.0186679568,
-0.0489944369,
0.0157663748,
-0.0025349127,
-0.0348043665,
-0.0440075994,
-0.0247275922,
0.0726904646,
-0.0172700975,
-0.0091380812,
-0.0528238751,
2.846120009e-34,
0.0197937302,
0.006895252,
0.0091054998,
0.08414606,
-0.0091970526,
0.0181567818,
-0.0422349386,
-0.0019043855,
0.0307626631,
-0.0307303164,
-0.0162102468
] |
http://arxiv.org/abs/2401.13927v2 | http://arxiv.org/pdf/2401.13927v2 | Adaptive Text Watermark for Large Language Models | 2024-01-25T03:57:12 | 2024-06-09T03:52:21 | Yepeng Liu, Yuheng Bu | cs.CL | cs.CL | The advancement of Large Language Models (LLMs) has led to increasing
concerns about the misuse of AI-generated text, and watermarking for
LLM-generated text has emerged as a potential solution. However, it is
challenging to generate high-quality watermarked text while maintaining strong
security, robustness, and the ability to detect watermarks without prior
knowledge of the prompt or model. This paper proposes an adaptive watermarking
strategy to address this problem. To improve the text quality and maintain
robustness, we adaptively add watermarking to token distributions with high
entropy measured using an auxiliary model and keep the low entropy token
distributions untouched. For the sake of security and to further minimize the
watermark's impact on text quality, instead of using a fixed green/red list
generated from a random secret key, which can be vulnerable to decryption and
forgery, we adaptively scale up the output logits in proportion based on the
semantic embedding of previously generated text using a well designed semantic
mapping model. Our experiments involving various LLMs demonstrate that our
approach achieves comparable robustness performance to existing watermark
methods. Additionally, the text generated by our method has perplexity
comparable to that of \emph{un-watermarked} LLMs while maintaining security
even under various attacks. | Adaptive Text Watermark for Large Language Models: The advancement of Large Language Models (LLMs) has led to increasing
concerns about the misuse of AI-generated text, and watermarking for
LLM-generated text has emerged as a potential solution. However, it is
challenging to generate high-quality watermarked text while maintaining strong
security, robustness, and the ability to detect watermarks without prior
knowledge of the prompt or model. This paper proposes an adaptive watermarking
strategy to address this problem. To improve the text quality and maintain
robustness, we adaptively add watermarking to token distributions with high
entropy measured using an auxiliary model and keep the low entropy token
distributions untouched. For the sake of security and to further minimize the
watermark's impact on text quality, instead of using a fixed green/red list
generated from a random secret key, which can be vulnerable to decryption and
forgery, we adaptively scale up the output logits in proportion based on the
semantic embedding of previously generated text using a well designed semantic
mapping model. Our experiments involving various LLMs demonstrate that our
approach achieves comparable robustness performance to existing watermark
methods. Additionally, the text generated by our method has perplexity
comparable to that of \emph{un-watermarked} LLMs while maintaining security
even under various attacks. | [
0.0456940904,
0.0473013297,
0.0042025247,
0.0282997712,
-0.0370286629,
-0.0109780375,
0.018980924,
0.0642667264,
0.0124347052,
-0.0473973304,
0.0145576112,
0.0224815905,
-0.004722089,
0.0259192381,
0.0319112353,
-0.0416376702,
0.017205067,
-0.0449508131,
0.0032751812,
-0.0203729421,
-0.0224329382,
0.0109258285,
0.0698313564,
0.0209590457,
-0.0398166701,
-0.0057659773,
-0.0177445319,
0.0354508907,
0.0099844877,
-0.0118441284,
-0.091475673,
0.0556010641,
-0.0122392951,
0.0741019323,
0.0000022021,
-0.0477696806,
-0.0367813706,
0.0508488603,
-0.0295370258,
0.0383214578,
0.0649598911,
0.00603332,
-0.0427161939,
0.022493707,
-0.0088239647,
0.0460218377,
0.0650087446,
0.070786491,
0.0714308992,
0.0351816751,
-0.0260831341,
-0.0254469272,
0.0335492678,
-0.0010952166,
0.0319394991,
-0.0310041551,
0.0295910053,
0.0061606881,
-0.0282152239,
0.0295405425,
0.0114965783,
0.0580222718,
0.0146088805,
-0.0493545122,
-0.0069313226,
0.0064206803,
0.001269024,
0.0093011074,
0.0051505808,
-0.0047028488,
-0.0724127963,
-0.0165920127,
0.0063696611,
-0.0241178516,
0.0222194083,
0.027682703,
-0.0103341388,
0.0189574081,
0.0225963183,
0.0119899008,
0.0118427156,
0.006696471,
0.0303372759,
0.0234373845,
-0.0189786386,
0.0307457745,
0.0482425094,
-0.0201331377,
-0.0631786063,
-0.0484811999,
-0.0484319776,
-0.0728851482,
0.0402919464,
0.0201138314,
0.0467651486,
0.0189087223,
-0.118832238,
-0.0741644725,
0.0109452065,
-0.0346105546,
0.0416414514,
0.0162714366,
0.0091089942,
0.0441114306,
-0.0428362265,
0.0939727649,
-0.0271777548,
0.0617108233,
-0.0078065838,
-0.0485127568,
-0.0053412365,
-0.0550776199,
-0.0344480388,
0.043180272,
0.0006532123,
-0.0228173994,
-0.0427352525,
0.0528470464,
0.0145559758,
0.0336233042,
0.0164346136,
-0.0306587927,
-0.0050348216,
-0.0248101279,
0.0216645282,
-0.0555263683,
0.0154859629,
-0.0089167682,
-0.0083194869,
-0.0003893497,
0.020820979,
0.049876079,
0.0092352424,
0.0203097425,
-0.0196008794,
0.113639012,
0.0356001854,
-0.0350462496,
-0.0314136483,
-0.0378262326,
-0.0089893183,
0.0330920294,
0.0081829084,
-0.0505995825,
-0.0147937052,
-0.0276810937,
0.00961594,
-0.0634156466,
0.0170310158,
0.0234530848,
-0.0402829833,
0.0531846732,
-0.0444762371,
-0.0170676596,
0.0113610486,
-0.0052737533,
-0.0018371021,
0.0382398292,
0.0537879951,
0.0719314963,
0.0643610358,
-0.0199537687,
0.0042992546,
0.0449704267,
-0.0000833786,
0.0029780597,
-0.0209886227,
-0.0261024274,
0.0644366071,
-0.0152384518,
-0.0132503649,
0.0604638867,
0.0117958561,
-0.0372308791,
0.0665412322,
0.0778295547,
-0.0366610177,
0.0009370301,
-0.0307085942,
0.0090630865,
0.0035204866,
-0.0251885429,
-0.0246077478,
-0.0024705131,
-0.0066706808,
-0.0354260691,
0.0463498756,
-0.0195843447,
-0.0046753753,
-0.0097967442,
-0.0207304116,
-0.0262520872,
0.0090800943,
-0.013888048,
0.0132687287,
0.0532090366,
-0.0399525724,
0.0106345033,
-0.0051347367,
-0.0175913032,
0.0330371633,
-0.074667573,
0.0221848693,
0.1016419232,
0.079667002,
0.0145390546,
0.0941008776,
0.0033702636,
-0.0176579338,
0.0657703951,
-0.0362792052,
0.0159220174,
0.0030911262,
-0.0444226712,
-0.0205990616,
-0.0141190905,
-0.0092361476,
0.0349225439,
-0.0139016416,
0.0452957153,
0.0345470645,
-0.0386475064,
0.0061501255,
0.0213968605,
0.0272158906,
-0.0164838098,
0.0283179618,
-0.0316915177,
0.0226518363,
0.0244369432,
-0.009735574,
0.0555790849,
0.0138052963,
-0.0339308307,
-0.0340845212,
-0.0429737158,
0.0161543339,
0.0459490567,
0.0477649868,
-0.0318281353,
0.0046019778,
0.0297389552,
-0.0183398481,
-0.0250313114,
0.015509312,
-0.0445569083,
0.0494968444,
-0.0363168679,
0.031619966,
-0.0093769422,
0.0267726295,
-0.0490173101,
0.0029294111,
0.0007465383,
-0.0112696337,
-0.047721073,
0.0285635386,
-0.0106683783,
0.0039197467,
-0.0147877336,
0.0031696905,
-0.0444421954,
0.0098236995,
0.0025720007,
0.0094758291,
0.0040094107,
-0.0403129384,
-0.0203505233,
-0.0097814305,
0.0014340953,
0.0101781646,
0.0288219042,
-0.0007507825,
0.0135040255,
-0.007209009,
-0.0400179811,
-0.0301622692,
-0.0019616708,
0.0081699509,
0.0271750111,
-0.0179700442,
-0.0244078785,
0.0055410643,
-0.008842661,
0.0270638037,
0.0613613836,
0.010887498,
-0.0177073982,
-0.0862811357,
0.0253952071,
0.0020683019,
-0.0125292158,
0.0175568294,
-0.013155371,
-0.0110110007,
-0.0048917704,
0.025879452,
0.0164878238,
0.0308946595,
-0.0093912063,
0.026293911,
0.0094305035,
-0.0059250253,
-0.041136127,
-0.0141324373,
0.0113923643,
-0.0046799672,
-0.0297522675,
0.0595558509,
-0.0163215566,
-0.0037380012,
0.0164036453,
-0.0467573702,
0.0392015688,
-0.0063691745,
-0.024017347,
-0.0386197418,
-0.0312195625,
-0.0140953958,
0.006204545,
-0.0027167017,
-0.0484836474,
0.0041679377,
-0.0499713533,
0.0417777933,
-0.0225161593,
-0.0126105752,
0.0378346257,
0.0334453546,
-0.0107176881,
0.0412090346,
0.0071443291,
-0.0125677465,
-0.0272474159,
-0.0429808907,
0.0245325379,
-0.0354489721,
-0.0159235727,
-0.0055873818,
-0.0267682169,
0.0045491611,
0.0066751163,
-0.0195733905,
-0.0644145608,
-0.0125245517,
-0.0025153102,
0.0327901468,
-0.0353065617,
-0.0223851725,
0.0151051758,
0.0369788259,
-0.0000398211,
-0.0190951098,
-0.0099993851,
-0.0348302051,
0.0519233756,
0.0110430792,
-0.0658247545,
-0.0063853231,
-0.005237747,
0.0131804999,
0.0823507011,
0.0268454403,
-0.0144939702,
0.0267466884,
-0.0217256751,
0.0304036904,
0.0104990257,
-0.0266900826,
0.0255066901,
-0.0092602037,
-0.0147708142,
0.0279054008,
0.0780999959,
0.0424672365,
-0.0175344367,
0.0207143072,
-0.015882019,
0.0138863642,
-0.008657367,
-0.0070189596,
-0.0639078096,
-0.0156763382,
0.034888152,
0.0291499272,
-0.0278728474,
-0.0038999314,
-0.0349208452,
-0.0641412437,
0.0102489656,
0.0210410599,
-0.0241209883,
-0.0321453214,
-0.055589173,
-0.0164907537,
0.061999429,
-0.0162014309,
-0.0576638356,
0.0467752181,
-0.0273869429,
0.0280036312,
-0.0185409337,
-0.0429910272,
-0.0151655851,
-0.0241487008,
0.0219680276,
0.0663914457,
0.0344470292,
-0.0129250987,
-0.0416975506,
0.0341389067,
0.0161191709,
-0.0166599117,
0.0033177377,
0.0163335707,
-0.0047310391,
-0.0314442702,
-0.0242197458,
0.0010837025,
0.0352493897,
-0.051052466,
0.0227477551,
0.0187275484,
0.0037624822,
0.012322844,
-0.0269333851,
-0.0652434751,
0.017751202,
0.0505568236,
0.0124604972,
-0.0458543487,
-0.0271406807,
-0.0668981597,
0.0417497568,
-0.026384471,
-0.0296296049,
0.0073347422,
-0.0444391705,
-0.0693475157,
-0.042632062,
0.017182203,
-0.0498421863,
0.0219145138,
0.0375746526,
0.021281939,
0.034662772,
-0.0409704633,
-0.0236343723,
-0.0518254265,
-0.1199435294,
0.0400088727,
-0.0564509928,
0.0277329274,
-0.0037300158,
-0.0413782783,
-0.0190364756,
0.0440072939,
-0.05972315,
-0.0567711182,
0.0405211896,
0.0149145499,
0.0188255608,
-0.0299163461,
-0.0466330871,
-0.0819395557,
0.0782948211,
0.0295860097,
-0.0048581818,
0.0077823913,
0.0320809819,
0.0191680603,
0.0035875593,
0.0181232374,
-0.0568586774,
0.0025372761,
0.0527495109,
-0.0189611409,
-0.0144166686,
-0.0045339018,
0.0050194361,
0.0728143454,
0.0098939752,
-0.028073255,
0.0202313252,
0.0351992361,
0.0323494002,
-0.0642133802,
-0.0416621193,
0.0013791831,
0.0100330738,
0.0220815353,
-0.025456341,
0.0176511575,
-0.0261759497,
-0.0165994503,
-0.0014854815,
-0.0063776383,
-0.0104766088,
-0.0340958796,
0.0602572747,
0.0267835818,
-0.0052022664,
-0.0008374717,
0.0228170156,
0.020974854,
0.0022969549,
-0.0106144492,
-0.01569557,
0.023207847,
-0.0231874362,
-0.0292294752,
0.0646711066,
-0.0634987578,
-0.0319566205,
-0.0199252423,
0.0450229496,
0.031243166,
0.0466976054,
0.0021058044,
-0.0710630193,
0.0004001188,
-0.0320575796,
-0.0060217511,
0.0243235007,
-0.0502945855,
0.0064500016,
-0.023184495,
-0.0164351705,
0.0017660992,
0.0200055465,
0.014419144,
0.0299197808,
0.0040325904,
-0.0347280279,
-0.0366169959,
0.0098441457,
0.053433504,
0.0283333156,
-0.044258099,
-0.0312848464,
-0.0441809297,
-0.0576258712,
0.0102910995,
-0.0154318204,
0.0154960277,
0.0243479349,
-0.0133560421,
0.0212881211,
-0.0157785509,
0.0077673253,
-0.0333001241,
0.0238451697,
0.0073697474,
-0.0547541715,
0.0196888186,
-5.609583571e-33,
0.0041861949,
-0.049020201,
-0.0586275794,
-0.0260417089,
0.0579452626,
0.0400963798,
-0.0031419208,
-0.0153461648,
0.0129106622,
-0.0021169803,
0.000147385,
-0.0338809676,
0.0310941897,
0.0233717542,
-0.0586571582,
-0.009476658,
0.0255538356,
0.0129471896,
-0.0120940739,
0.0168835185,
0.051314123,
-0.0389507823,
0.0100462399,
-0.0843708813,
0.0144868912,
-0.0108916396,
-0.0766534209,
-0.0387925282,
0.0083289305,
0.0433172323,
-0.022780221,
0.0435380526,
-0.0268932898,
-0.012837681,
-0.0294839181,
0.0374862663,
-0.0021627205,
-0.0349676684,
-0.0120832976,
0.0323855057,
0.0236659069,
-0.0961027071,
-0.0113507761,
-0.0216978621,
-0.0280434284,
0.0438420326,
-0.0019508451,
0.0131824035,
-0.0324821733,
-0.0254264642,
-0.0004220496,
0.0269846544,
-0.0126910806,
0.0197671633,
0.021315679,
-0.0196342859,
0.0554537885,
0.0123599302,
-0.0718514025,
0.0569932684,
0.0059703765,
-0.0164235029,
0.0413816907,
0.0127715115,
-0.0071168933,
0.0252075717,
-0.0079700463,
0.0091081802,
0.0261332393,
0.0071279551,
0.0086740879,
-0.0025571317,
0.0997528285,
0.0551878363,
0.074688144,
-0.0435012132,
-0.011258726,
0.0348407589,
-0.0681063533,
-0.0173273459,
0.0818465203,
0.0296147019,
-0.0395209864,
-0.0110808229,
-0.0539306439,
-0.0092105297,
-0.0101377871,
-0.0132098272,
0.0306320768,
-0.0143473186,
-0.0553249978,
-0.0078633614,
0.018751774,
-0.0479300655,
-0.0460040793,
0.0237241071,
0.0044802204,
0.0368168578,
-0.0275412612,
0.0271818396,
0.0405646041,
-0.0144218458,
0.0731268227,
-0.0616152585,
0.008752225,
-0.0111801112,
0.0113643557,
0.0030431829,
-0.1057332233,
0.0021267077,
0.0227672979,
0.0106059778,
0.041920092,
-0.0381499007,
0.0126505727,
-0.0485005267,
0.0072062425,
-0.0408304743,
0.0358784012,
-0.0410714447,
0.030796146,
0.0109648686,
-0.0272569712,
0.0174304787,
0.005480458,
0.0340669528,
-0.0288689565,
0.0066538448,
0.0672876984,
0.0045291795,
-0.0212465338,
0.002107172,
2.827e-7,
0.0370702483,
0.0327176601,
0.0353811905,
-0.0020793132,
0.0196096655,
0.057342872,
0.0303390678,
0.0156876519,
-0.0370365009,
-0.0029868686,
-0.0465189926,
-0.0067713712,
0.0667109564,
0.0020233663,
-0.0545359142,
0.0123446602,
-0.0364617072,
-0.0007271867,
-0.0500520021,
0.0627952963,
0.062755838,
0.116718322,
0.0368404612,
-0.0270207264,
-0.0097982055,
-0.0775018558,
0.0096901013,
-0.0273065381,
0.0201099049,
-0.0259022862,
-0.0164103545,
0.0160822719,
0.0400738753,
0.0291206986,
-0.0400986932,
-0.0344883278,
0.0273110513,
0.0362852961,
-0.0098583903,
0.019553246,
-0.0412846617,
-0.0150751416,
0.0027688269,
-0.0715105012,
0.045729287,
0.0782832801,
-0.0484202355,
-0.0628470853,
-0.0354791172,
-0.0213009138,
0.0611768402,
-0.0359969772,
-0.0401470996,
0.015526128,
0.0309227146,
0.0388995409,
-0.0179242808,
-0.047405649,
-0.0484028198,
-0.1040537357,
-0.0234781466,
0.0162579753,
-0.0662859008,
0.0298384279,
0.083626546,
-0.0657615513,
0.045318868,
2.833592729e-34,
-0.0014112741,
-0.0765870214,
-0.0246533304,
0.0390678681,
0.0436165668,
0.040288724,
0.0159917902,
0.0381931476,
0.0138089526,
-0.0655107796,
-0.0076740743
] |
http://arxiv.org/abs/2401.13920v3 | http://arxiv.org/pdf/2401.13920v3 | LocMoE: A Low-Overhead MoE for Large Language Model Training | 2024-01-25T03:36:39 | 2024-05-23T10:03:35 | Jing Li, Zhijie Sun, Xuan He, Li Zeng, Yi Lin, Entong Li, Binfan Zheng, Rongqian Zhao, Xin Chen | cs.LG | cs.LG, cs.AI, cs.CL | The Mixtures-of-Experts (MoE) model is a widespread distributed and
integrated learning method for large language models (LLM), which is favored
due to its ability to sparsify and expand models efficiently. However, the
performance of MoE is limited by load imbalance and high latency of All-to-All
communication, along with relatively redundant computation owing to large
expert capacity. Load imbalance may result from existing routing policies that
consistently tend to select certain experts. The frequent inter-node
communication in the All-to-All procedure also significantly prolongs the
training time. To alleviate the above performance problems, we propose a novel
routing strategy that combines load balance and locality by converting partial
inter-node communication to that of intra-node. Notably, we elucidate that
there is a minimum threshold for expert capacity, calculated through the
maximal angular deviation between the gating weights of the experts and the
assigned tokens. We port these modifications on the PanGu-Sigma model based on
the MindSpore framework with multi-level routing and conduct experiments on
Ascend clusters. The experiment results demonstrate that the proposed LocMoE
reduces training time per epoch by 12.68% to 22.24% compared to classical
routers, such as hash router and switch router, without impacting the model
accuracy. | LocMoE: A Low-Overhead MoE for Large Language Model Training: The Mixtures-of-Experts (MoE) model is a widespread distributed and
integrated learning method for large language models (LLM), which is favored
due to its ability to sparsify and expand models efficiently. However, the
performance of MoE is limited by load imbalance and high latency of All-to-All
communication, along with relatively redundant computation owing to large
expert capacity. Load imbalance may result from existing routing policies that
consistently tend to select certain experts. The frequent inter-node
communication in the All-to-All procedure also significantly prolongs the
training time. To alleviate the above performance problems, we propose a novel
routing strategy that combines load balance and locality by converting partial
inter-node communication to that of intra-node. Notably, we elucidate that
there is a minimum threshold for expert capacity, calculated through the
maximal angular deviation between the gating weights of the experts and the
assigned tokens. We port these modifications on the PanGu-Sigma model based on
the MindSpore framework with multi-level routing and conduct experiments on
Ascend clusters. The experiment results demonstrate that the proposed LocMoE
reduces training time per epoch by 12.68% to 22.24% compared to classical
routers, such as hash router and switch router, without impacting the model
accuracy. | [
0.0505458266,
0.0275596827,
-0.0004400333,
0.0354616605,
-0.0271649528,
-0.0309357215,
0.0391308293,
-0.0088138934,
0.0320007354,
-0.0144987218,
-0.0582472645,
-0.0175809823,
0.0166620072,
0.0053776829,
0.0360306539,
-0.0279712286,
0.0304804016,
-0.0567690097,
-0.0067028091,
-0.0551494621,
-0.013003611,
0.0180820841,
0.0173156373,
0.0338624455,
0.0140494481,
0.0078799902,
-0.0122241601,
-0.0040485025,
0.005489965,
-0.0378326029,
-0.0094171502,
0.0428637303,
0.0228954032,
0.0344824754,
0.0000022897,
-0.0113569135,
-0.0345394723,
0.0073443609,
-0.0198286325,
0.0252793022,
0.0229643881,
0.0046495395,
-0.0104280729,
0.0514071919,
0.0112392111,
0.0069177998,
0.0691579506,
0.0405298099,
0.0191513691,
0.086415641,
-0.0051702545,
-0.0571012348,
0.0516653508,
-0.0327618644,
-0.0350496396,
-0.0351065993,
0.0426140651,
0.0365099199,
-0.0228654686,
0.0088498797,
-0.0502598062,
-0.0009961523,
0.0193091221,
0.0135809192,
0.0330965295,
0.0317946523,
0.0022300889,
-0.0059967926,
-0.0355273262,
0.0845789388,
-0.0427506454,
-0.005665896,
-0.0010256553,
0.0198283233,
-0.0574558564,
0.027961906,
-0.0252330601,
-0.0105224,
0.0165757872,
0.0136984941,
0.01777271,
-0.017337501,
0.0543087609,
0.0163619816,
-0.0594787039,
0.0270384029,
0.0368999168,
-0.0148310103,
-0.0149948234,
-0.0178162605,
-0.0416970477,
-0.0446911454,
0.0416914374,
0.0019227001,
-0.0324470103,
0.0018242749,
0.000499705,
-0.0009497316,
-0.0183268916,
-0.0145702567,
0.0355150513,
0.0489707254,
0.0227980185,
0.0544695444,
-0.0453838669,
-0.0225783624,
-0.0131880529,
0.0055835745,
-0.0383806303,
-0.0035718023,
-0.0426198691,
-0.010531377,
-0.0305639263,
0.0428212173,
-0.0028703131,
-0.0488755852,
-0.0364583582,
0.0014727658,
-0.0107211098,
-0.0273960978,
0.0304242298,
-0.0234845299,
0.0010236189,
-0.0166544821,
0.0013140757,
0.0724422038,
-0.0320354179,
0.0164935663,
-0.0071701952,
0.0028147376,
-0.0043493276,
0.0458814427,
-0.0219899602,
0.0142104449,
0.0228171125,
0.0751040429,
-0.0245781504,
-0.0121977059,
-0.0422639661,
0.0094685284,
0.0134362252,
-0.0584495366,
-0.0343353264,
-0.0769483298,
0.0507627353,
0.0130113298,
0.0285533722,
-0.0122023718,
-0.0029195617,
-0.005762375,
-0.0605031215,
0.0088674882,
-0.0152045358,
-0.0108199343,
0.0308190882,
0.0047973231,
-0.0339206234,
0.0547298416,
0.0048151296,
0.0085217915,
0.004019327,
-0.0154799083,
-0.0221916456,
0.0250053555,
0.0389587991,
0.0324212313,
-0.0139983986,
-0.0153614813,
0.0244193207,
0.0342274308,
0.0412350409,
0.0850277543,
-0.0761203095,
-0.0096910084,
0.0481774732,
0.0678660721,
0.0105945636,
0.0248248205,
0.0140800644,
0.0179269537,
0.0003697874,
-0.0017135285,
-0.0170622431,
0.0071528335,
-0.0470314622,
-0.0169358216,
0.1174617484,
0.056935139,
-0.0456111953,
-0.0002847519,
-0.0031427206,
-0.0119165732,
-0.0421745367,
0.0116001116,
0.0496732891,
-0.0302831884,
-0.0086544724,
0.0697247684,
0.0820919573,
-0.0407645218,
0.0040286849,
-0.0342484675,
0.0162722748,
0.0626442581,
0.0432911851,
0.0086034723,
0.0113727227,
-0.0273772534,
-0.0138839353,
0.0054939385,
-0.0026574789,
0.0045232098,
-0.0140652191,
-0.0809713528,
-0.0209023897,
0.0405318737,
-0.0184072107,
0.0124905789,
0.009862639,
0.0226116776,
0.0004074929,
-0.016115224,
-0.0612477437,
0.0059608892,
-0.0316335671,
0.0367663093,
0.0020259772,
0.0198772829,
0.0460105315,
0.0166495573,
-0.0229980517,
-0.0353347324,
0.0141089018,
-0.0164958853,
-0.0422822498,
-0.023655951,
-0.0737363547,
-0.0133063542,
0.0361155234,
-0.0077310717,
0.0301964544,
-0.0346659832,
-0.0182182528,
-0.0571433343,
0.0300253853,
-0.0366569497,
0.0441327132,
-0.0213855859,
0.0028366256,
0.0146394204,
0.0159051828,
-0.024561327,
0.056070812,
-0.0300624967,
0.0014773799,
0.0137467477,
0.0319241583,
0.0127386078,
-0.0183784273,
-0.0027775993,
0.0086040217,
0.0381625369,
0.0339388922,
0.0136502814,
-0.0126873869,
-0.010652028,
-0.0724079236,
-0.1200110614,
0.0068664728,
-0.0399374366,
0.0130779482,
0.0118036773,
0.0176094081,
0.0065861503,
-0.0114525706,
-0.0234987997,
0.0203016177,
-0.034985818,
0.005718126,
-0.0438421145,
-0.0037561506,
-0.0676676482,
-0.0103777489,
-0.0585125722,
-0.0050978488,
0.059181165,
-0.0328339823,
0.0013481876,
-0.0518229567,
-0.0039637927,
-0.0488246866,
0.0091783255,
-0.0120474296,
-0.0252163745,
-0.0254182983,
-0.0087431064,
0.0559385791,
0.0446171723,
0.0086988779,
-0.0140499584,
0.0587483943,
-0.0056939553,
-0.0323703103,
0.0047279666,
-0.0083577866,
-0.0035893724,
-0.086164467,
0.0082944063,
-0.0353310481,
0.0652193427,
0.0026961986,
-0.0385632403,
-0.0305857398,
-0.0612151213,
-0.0157494117,
-0.0227898043,
-0.0180213582,
0.0036068764,
0.0070195263,
-0.0099546835,
-0.0175271891,
-0.0181612261,
0.0533890612,
-0.0170174986,
0.0023628359,
-0.0448702462,
-0.0459584408,
-0.0395790078,
0.0078756129,
0.0195187535,
0.0734019503,
-0.0673366711,
-0.0343182273,
-0.0232080724,
0.0184544437,
0.0024375708,
0.0066071558,
-0.0236228481,
-0.0145700509,
0.0515824668,
-0.0268661138,
-0.05185166,
0.0277247373,
0.0131819267,
-0.0286732409,
0.0296178907,
0.0525262579,
-0.0660258457,
-0.0457575023,
0.0590603687,
0.0146507779,
0.0545391664,
-0.0215812344,
0.0001084727,
-0.0069798301,
0.0125714578,
0.0072545698,
-0.0056613944,
0.0146678928,
-0.0447754562,
0.0135210995,
0.0594104603,
0.0481632501,
-0.0513089821,
-0.015911283,
0.0052999635,
0.029449217,
-0.031498,
-0.0142374048,
0.0051946798,
0.0518826321,
-0.0334982648,
0.0301368814,
0.1044378132,
0.0044963341,
-0.0923881903,
-0.0101239542,
-0.0016337576,
0.043772053,
-0.0002175461,
0.0191206913,
-0.0756902397,
-0.0081156911,
0.0088065499,
0.0200570915,
0.0257625058,
-0.0151612423,
0.0101172449,
-0.098129645,
0.0188612361,
0.0363397636,
-0.0869959518,
-0.0169043448,
-0.0864620805,
-0.0214101821,
-0.0258097686,
0.0475614108,
0.0074309115,
0.0070973067,
-0.0282892063,
0.04266949,
-0.0078022489,
-0.0574359335,
0.0295411646,
-0.0357257724,
0.0753589347,
-0.0039493176,
-0.0116329249,
0.028146334,
-0.006063038,
0.0189517196,
0.0506312288,
0.0166809931,
-0.0311227962,
-0.0011566607,
-0.0381617658,
-0.0009946637,
0.002273846,
-0.0001263415,
0.0049640806,
0.0216927007,
0.0031774375,
0.0861990228,
0.0143257314,
-0.0008052837,
0.0100806141,
-0.021193549,
0.0035814901,
0.0638113916,
-0.010342177,
-0.0362564363,
-0.0109908907,
-0.0304055549,
-0.0296852309,
0.0450768843,
-0.0097038364,
0.007285018,
-0.0239516608,
-0.0141308876,
-0.0317745879,
0.0329321995,
0.0361469127,
0.0659817606,
0.0354748927,
0.0424195416,
0.0238613077,
0.0090990644,
-0.0166652165,
0.0165291652,
-0.0770915672,
0.0028587363,
-0.1019678488,
-0.0115258833,
0.0349274427,
0.0057881083,
-0.0177500937,
0.0111217638,
-0.0607711002,
-0.0951389894,
-0.016427204,
0.0344811231,
0.0156591106,
0.0540126972,
0.0311297905,
-0.0801348314,
0.1213037521,
0.0090642767,
-0.0917023942,
0.0337858684,
0.0078742998,
-0.0216130316,
0.0078989156,
0.0270000044,
-0.0201426502,
0.013651046,
0.0491172299,
0.0280212443,
-0.040955849,
-0.0049256147,
-0.0378528498,
0.036583975,
0.0336543471,
0.0020440889,
0.0326164924,
-0.0357920267,
0.0559048876,
0.0001912685,
-0.0070925048,
-0.014973443,
0.010797468,
0.0375576615,
0.0059273099,
0.0278177932,
0.0119724888,
0.081466347,
-0.0384821892,
0.028309796,
0.0078201685,
0.0030319307,
-0.0143523924,
0.0429315493,
0.0099188546,
-0.0125853885,
-0.0425620861,
0.032785695,
0.0091649778,
-0.025986176,
0.0303478036,
0.0087300008,
-0.019943377,
-0.0347277038,
0.0298453905,
-0.0608706288,
-0.0747123286,
-0.0160388984,
0.0533880405,
0.0899121165,
0.0417264774,
0.0501731001,
-0.023884112,
-0.0081439158,
-0.0302239303,
-0.0074731098,
-0.0314142257,
-0.0146704921,
0.0194147062,
0.0153611004,
-0.0001158847,
0.0646999106,
0.0615732521,
0.0095793363,
0.0178278852,
0.0004409197,
0.0085513322,
-0.0356213935,
0.010997233,
-0.0003692469,
0.0543870702,
-0.0381647013,
0.0343665369,
0.0079016807,
-0.035468325,
-0.0097906282,
-0.0399369709,
0.0183314979,
0.0013235414,
0.0483020283,
0.0535381176,
-0.0158145335,
0.0048079411,
-0.049343802,
-0.0294352975,
0.0378494784,
-0.0016030241,
0.0079272175,
-7.844617469e-33,
0.0096321199,
-0.0156506356,
0.0450183004,
-0.0250452962,
-0.031192245,
0.0131301833,
-0.0227266103,
0.0045528053,
-0.0149022397,
0.0068191867,
-0.0265037734,
0.0177331772,
0.0068991925,
0.047537718,
0.0104012638,
0.0284855962,
0.0716178045,
-0.001046779,
-0.013402475,
0.0399181768,
0.0392084271,
0.0169649627,
0.033572223,
-0.026344955,
-0.0285653714,
0.0108294636,
0.0080770357,
0.0258985888,
-0.0267669801,
0.0177201293,
-0.0480548665,
0.0354427025,
-0.012070043,
-0.0116006676,
-0.0161788948,
-0.0562119745,
-0.0365770571,
-0.0150140664,
0.0000958502,
-0.0199680664,
-0.0009258293,
-0.0282996688,
0.030824244,
0.0266546197,
0.0116161406,
-0.0329077952,
0.0236662142,
0.0167524647,
-0.0600944944,
0.035450317,
0.0023724223,
0.0373610593,
-0.0043392973,
0.0285950992,
0.0027759357,
-0.0511319377,
-0.0065371273,
-0.0204273295,
-0.0594593734,
0.007307183,
-0.0177146718,
0.058578629,
0.0280067362,
0.049888771,
0.0108815692,
0.0436109975,
0.0508880168,
0.0104906531,
-0.0024663575,
-0.003583309,
0.0343317427,
0.031896133,
0.0230856352,
-0.0406142883,
0.0026986876,
-0.0224519465,
-0.0840404183,
-0.0480038188,
0.0296107065,
0.0352183692,
0.0449627861,
0.0496901534,
-0.0469326526,
-0.0468603298,
-0.0183757022,
-0.0596881174,
-0.028460158,
-0.0423443057,
-0.023357911,
0.0471871272,
-0.0261196401,
-0.0308150984,
0.0303684548,
-0.0380669236,
-0.0253422093,
0.0210496988,
0.01955547,
0.0208612438,
0.0093147615,
0.0246443618,
-0.1358740181,
-0.0570439622,
-0.0245219599,
-0.042764239,
-0.0300108194,
-0.0092403302,
0.0199908856,
0.0273458157,
-0.0661421344,
-0.0077161347,
0.0021134939,
0.0011881896,
0.0158303864,
0.0246632546,
0.0319646522,
-0.0056742975,
0.0355725475,
0.0341086946,
0.0058793416,
0.0129832206,
0.020029135,
0.0159450248,
-0.0073426133,
0.0125405109,
-0.0246654004,
0.0329569131,
-0.0637121052,
-0.0105131678,
0.0437825397,
-0.0513387062,
-0.0352918319,
-0.006198646,
3.14e-7,
-0.0035190429,
0.0334162153,
-0.0171620529,
-0.0296421479,
-0.0047055725,
-0.011535353,
-0.0044793561,
0.0060030175,
0.0338878781,
0.0515579917,
0.0178952701,
0.0588989221,
0.0317533128,
-0.0091100074,
-0.1259810627,
0.0343014114,
-0.0935711712,
0.0265913606,
-0.04402164,
-0.011170526,
-0.0432530008,
0.0949374959,
0.0374966562,
0.0045029414,
-0.000893392,
-0.0451585874,
0.0296884552,
0.0053221532,
-0.0045864764,
0.0280001629,
0.0569641367,
0.0510188267,
-0.0128039885,
0.0283000208,
0.0047673816,
0.0501869656,
-0.0241834726,
0.0389234424,
-0.028633235,
0.0247775279,
0.0356756337,
0.0003438734,
-0.0178637952,
-0.0162525773,
0.0558153577,
-0.0134014096,
-0.035632845,
-0.042063497,
0.0274499319,
-0.0524278209,
0.0497459099,
-0.0124671012,
-0.0342442244,
0.0273571033,
0.0095879836,
-0.0315494984,
0.0236991085,
-0.0621767268,
0.0099841608,
0.0009077117,
-0.0277739875,
-0.1138079613,
-0.006364434,
0.0141391261,
0.0146433851,
0.0367175341,
0.0248714518,
2.927146926e-34,
0.0823356733,
-0.013580435,
0.0144144548,
0.0422676057,
0.0256145019,
-0.0585201047,
-0.0586650856,
-0.0268071629,
-0.0136839412,
-0.1144798622,
-0.041304443
] |
http://arxiv.org/abs/2401.13919v4 | http://arxiv.org/pdf/2401.13919v4 | WebVoyager: Building an End-to-End Web Agent with Large Multimodal Models | 2024-01-25T03:33:18 | 2024-06-06T18:37:34 | Hongliang He, Wenlin Yao, Kaixin Ma, Wenhao Yu, Yong Dai, Hongming Zhang, Zhenzhong Lan, Dong Yu | cs.CL | cs.CL, cs.AI | The rapid advancement of large language models (LLMs) has led to a new era
marked by the development of autonomous applications in real-world scenarios,
which drives innovation in creating advanced web agents. Existing web agents
typically only handle one input modality and are evaluated only in simplified
web simulators or static web snapshots, greatly limiting their applicability in
real-world scenarios. To bridge this gap, we introduce WebVoyager, an
innovative Large Multimodal Model (LMM) powered web agent that can complete
user instructions end-to-end by interacting with real-world websites. Moreover,
we establish a new benchmark by compiling real-world tasks from 15 popular
websites and introduce an automatic evaluation protocol leveraging multimodal
understanding abilities of GPT-4V to evaluate open-ended web agents. We show
that WebVoyager achieves a 59.1% task success rate on our benchmark,
significantly surpassing the performance of both GPT-4 (All Tools) and the
WebVoyager (text-only) setups, underscoring the exceptional capability of
WebVoyager. The proposed automatic evaluation metric achieves 85.3% agreement
with human judgment, indicating its effectiveness in providing reliable and
accurate assessments of web agents. | WebVoyager: Building an End-to-End Web Agent with Large Multimodal Models: The rapid advancement of large language models (LLMs) has led to a new era
marked by the development of autonomous applications in real-world scenarios,
which drives innovation in creating advanced web agents. Existing web agents
typically only handle one input modality and are evaluated only in simplified
web simulators or static web snapshots, greatly limiting their applicability in
real-world scenarios. To bridge this gap, we introduce WebVoyager, an
innovative Large Multimodal Model (LMM) powered web agent that can complete
user instructions end-to-end by interacting with real-world websites. Moreover,
we establish a new benchmark by compiling real-world tasks from 15 popular
websites and introduce an automatic evaluation protocol leveraging multimodal
understanding abilities of GPT-4V to evaluate open-ended web agents. We show
that WebVoyager achieves a 59.1% task success rate on our benchmark,
significantly surpassing the performance of both GPT-4 (All Tools) and the
WebVoyager (text-only) setups, underscoring the exceptional capability of
WebVoyager. The proposed automatic evaluation metric achieves 85.3% agreement
with human judgment, indicating its effectiveness in providing reliable and
accurate assessments of web agents. | [
0.0446800031,
0.026986232,
-0.0217384156,
0.0154630663,
-0.0228428915,
0.0213443991,
0.0406920277,
-0.0023976052,
0.0470145158,
-0.0727357343,
-0.0522248037,
-0.0555561706,
-0.0157572553,
0.032409817,
0.0335248411,
-0.0625677332,
0.0189816132,
-0.0460503548,
-0.0144617129,
0.010533154,
0.0128878048,
-0.0050940621,
-0.0406081825,
0.0684027895,
-0.0118971579,
-0.038442824,
-0.0591702238,
0.0175804477,
0.0046094768,
-0.0205837749,
-0.0139835086,
0.0409338959,
0.0346515328,
0.0296487827,
0.0000021906,
-0.0636997744,
0.020306617,
-0.0119162006,
-0.0174293835,
0.0037156849,
0.0440849289,
0.03506561,
-0.0251651909,
0.0260089114,
-0.0101879518,
0.0334097072,
0.0295854583,
0.0305561218,
0.0704643875,
0.0377525426,
-0.0122139333,
0.0289932154,
0.0255344491,
0.0249135848,
0.0011733866,
-0.0596691705,
-0.0129720764,
-0.0021931478,
0.0272443034,
-0.0327425115,
-0.0034946974,
0.0044695586,
0.0111753782,
-0.0028765309,
-0.0309877992,
0.0166660566,
-0.0313748121,
-0.0602313019,
-0.028339142,
0.0921828151,
0.0013237322,
0.0408386514,
0.0062424419,
0.0275033433,
-0.0062396689,
0.0411260761,
-0.0553304106,
0.0373620391,
-0.0233065449,
-0.0436183773,
-0.0597714148,
0.0096552772,
0.0087683946,
0.0135162063,
-0.032879997,
0.048468858,
0.0189087372,
-0.0216053985,
0.0108037451,
-0.0147565687,
0.0157194491,
-0.0510952286,
0.0041596051,
0.0452095866,
-0.0779288113,
0.0044969274,
-0.0200424325,
0.0148049761,
0.0342068411,
-0.0462087356,
0.0700628832,
0.0544613078,
-0.0527575314,
0.0641084164,
-0.0037436364,
0.0512590967,
-0.0220841542,
-0.0131237898,
-0.0166843776,
0.0007830964,
-0.074831225,
0.007419236,
0.0099877007,
0.0018396436,
-0.0273596756,
-0.0487303771,
-0.0043212776,
0.0694241375,
-0.012582914,
0.0353360996,
-0.0293436125,
-0.0118744439,
-0.0320339277,
0.0213121679,
-0.0551890135,
0.0284067616,
-0.0517454334,
-0.0020189586,
0.0053635002,
-0.0793845579,
0.025620563,
0.0419224761,
0.0029113269,
-0.0159606673,
-0.0041816966,
0.1147929206,
-0.0095985997,
-0.0314833485,
-0.0269606374,
-0.0620704405,
-0.0011660232,
-0.0787544996,
-0.0038047859,
-0.0563082993,
0.0412359387,
0.0229537059,
-0.0310640745,
0.0037890968,
0.0036263126,
0.0074131745,
0.0171961132,
0.0758654028,
0.0172262043,
0.0011904518,
0.0142079089,
0.0009299633,
-0.048482772,
0.0448261276,
-0.0085872076,
0.1116937548,
0.0416887477,
-0.005719698,
0.0641161427,
0.0478002615,
-0.0429998301,
0.0196149275,
0.0332164206,
-0.0019244192,
0.0264009088,
0.0203759558,
0.0133562787,
0.0432991013,
-0.0136719169,
0.0098549826,
0.0956582054,
0.0282885525,
-0.0165219251,
-0.0054027797,
0.0403431244,
0.0307145976,
0.0263479501,
-0.0041416325,
-0.0784758329,
0.0632091686,
-0.0511776283,
-0.014195337,
-0.0044550304,
0.060040731,
-0.021482138,
-0.0210682359,
0.0215098877,
-0.0172866471,
0.0588358641,
0.0022572002,
0.0533953123,
-0.031637948,
-0.0469334386,
-0.0322550759,
0.013201518,
-0.043182563,
-0.0119170649,
-0.0672101378,
-0.0194804408,
0.1060191691,
0.0617816523,
-0.0006559336,
-0.0483837947,
-0.0177780688,
0.0096855992,
0.0370657071,
0.0498549491,
-0.0038154637,
-0.0221418235,
-0.0319155641,
-0.0788562596,
-0.0199843924,
0.0013259678,
0.039848838,
-0.0100739896,
0.0063421,
0.0098826326,
-0.052364666,
0.0233562104,
0.0177415032,
0.0114939827,
-0.0228960905,
-0.0640771911,
0.0318393931,
0.039214991,
-0.0006492073,
-0.0004821004,
-0.0052012992,
0.011924753,
-0.0055185277,
-0.0267571993,
0.0015193032,
0.0282462928,
0.0314898342,
0.0595063455,
-0.0501409881,
0.0119625134,
0.079784736,
-0.0091559831,
-0.0318714194,
0.0154371774,
-0.0572610088,
0.0160103217,
-0.0256680064,
0.0188265983,
0.0003500526,
0.003993839,
-0.0070418604,
0.0463585109,
0.0183585212,
-0.0365579575,
0.0109037431,
0.0224230383,
0.0039994237,
-0.0134300236,
-0.034138795,
0.0257502906,
-0.0041604652,
0.0481255576,
0.0316439793,
-0.0207640249,
-0.057703644,
-0.0096997758,
-0.1611915231,
-0.0057697967,
-0.012333679,
0.0091143502,
0.0261722617,
0.006723199,
0.018166475,
-0.016869612,
-0.0146040637,
0.0420235209,
-0.0095756063,
0.0413316935,
0.0318344422,
0.0064357743,
-0.0600427575,
-0.0300042909,
-0.0093517862,
-0.0228466913,
0.0256333873,
0.0038712653,
0.0071381829,
-0.0304727983,
-0.0226437096,
-0.0333386622,
0.0223966278,
-0.0291143935,
-0.0767612234,
0.0387457721,
-0.0280449186,
-0.0022516146,
0.0624660291,
0.0554364733,
0.0439384542,
0.046028506,
-0.0226348937,
-0.0443808995,
0.0115587935,
0.0017366915,
-0.0191703439,
-0.0464065447,
0.004801658,
0.0064306483,
0.0676305443,
0.0081765233,
0.0395734496,
-0.0336457528,
0.0126940953,
0.0022818027,
-0.0330635607,
-0.0197407641,
0.0170157831,
-0.0303169247,
0.0423993208,
-0.0003357694,
-0.0417036265,
0.0565279424,
-0.039543055,
-0.0059307455,
-0.0090060951,
-0.0059812502,
0.0051724198,
0.0813118294,
-0.0113610541,
0.0348605551,
-0.0201196633,
-0.0214562975,
-0.0547896773,
0.0107478993,
0.049238205,
-0.0345591158,
0.02022985,
-0.0540253073,
0.0232618582,
-0.0107368054,
-0.0430378355,
-0.0410673097,
-0.0343082547,
-0.0362408124,
-0.0141806323,
-0.0005019205,
-0.0343964696,
-0.0544471256,
0.0110959439,
0.056352336,
0.0355117358,
-0.0096316431,
-0.028195085,
-0.0404894128,
-0.0231583454,
0.0355192609,
0.0355218872,
-0.0248294286,
-0.0211438257,
0.0156469587,
0.0635908023,
0.0253695305,
-0.0312815979,
-0.0502954498,
0.0075044865,
0.0103678126,
-0.0291927997,
-0.0245208759,
0.0369827785,
0.0091021499,
-0.0762398914,
0.0535207577,
0.0854477435,
0.0271084961,
-0.0638249218,
-0.0094955051,
-0.0377045609,
-0.0199468266,
0.0327914432,
0.018943565,
-0.0663852245,
-0.0562856272,
0.0248686261,
0.0246346388,
0.0195798222,
-0.0238829721,
0.0468276441,
-0.0733511299,
0.0345974155,
-0.0065071085,
0.0149866333,
0.0357762724,
-0.0172053929,
0.0330936201,
0.0159041286,
0.0088688042,
0.051395826,
-0.0133050131,
-0.0358250588,
0.0002724678,
-0.0168590453,
-0.0436376408,
0.0663770288,
-0.0361314155,
0.0392883159,
0.0764051452,
0.0326463915,
0.0061860252,
-0.0424693823,
0.0160496198,
0.0110271657,
0.0224785656,
0.0005795841,
0.0679076463,
-0.0158901159,
0.0004951658,
-0.0273827836,
0.0017554284,
-0.0457232706,
0.0137698427,
0.0014752158,
0.0705402642,
-0.0347021893,
0.0050760093,
-0.0368427411,
-0.0281881839,
0.0086101843,
0.0406820476,
0.0036713399,
-0.0656252876,
-0.0805686712,
-0.0364527963,
-0.0185913388,
0.0012551981,
-0.0227797031,
0.0145652527,
-0.0035839309,
-0.0225381255,
-0.0629196167,
-0.0675136,
0.0307298601,
0.0233916715,
0.0146532394,
-0.0405012146,
0.0219797995,
0.05473914,
-0.0328085087,
0.0097774947,
-0.0120769367,
0.0409729332,
-0.0482702777,
0.0033990629,
0.0402185731,
-0.0824059844,
-0.000508633,
0.0247878134,
-0.0072453115,
-0.0164695773,
-0.0154415416,
0.0402756967,
0.0285438411,
-0.0014406909,
-0.0186300017,
-0.0389548205,
0.0012706341,
-0.0201394223,
-0.0727337897,
-0.0254763588,
0.0521774553,
0.0277861841,
-0.0102271223,
0.0007261939,
-0.078281112,
0.004691557,
-0.0405185074,
0.0368112251,
0.003138426,
-0.0139300469,
0.0467519425,
0.0500361212,
0.0070627653,
-0.0166803617,
-0.0018848459,
0.0187987592,
0.0037917336,
0.0042148125,
-0.0170736481,
-0.0513793454,
0.0081375204,
-0.008634666,
0.0012489193,
0.0580194443,
0.0106890555,
-0.0064439885,
-0.0166127849,
0.0062368531,
0.0041519403,
0.0190139655,
-0.0119409356,
0.0191228501,
0.0048220297,
-0.0136078838,
0.0040808595,
0.0300811417,
0.0296961386,
-0.024471689,
0.0681528449,
0.0188805852,
-0.026001418,
-0.0335977823,
0.0526083745,
-0.0403229631,
-0.0536790825,
-0.0114739249,
0.0205756687,
0.037747331,
0.0429201983,
0.0538842343,
-0.0335605666,
0.0174554307,
-0.0335954763,
0.0361429527,
0.0174155068,
-0.0083090616,
-0.0066025378,
0.03923437,
0.0022119267,
-0.0198690854,
0.0406550877,
-0.0521261171,
0.0356045403,
-0.0232657194,
-0.0457201563,
-0.0661359802,
-0.0280706305,
0.0123894894,
0.0354158431,
-0.0214197505,
-0.0617148764,
-0.0460418537,
-0.0483960472,
-0.0348956622,
0.0294727553,
-0.0123454547,
-0.0265674107,
0.0280709974,
0.0813938975,
0.0284541864,
-0.0192135219,
-0.0500857942,
-0.0096900137,
0.0033733859,
0.0351105109,
-0.0014995497,
-6.438663043e-33,
-0.0009903273,
-0.0345938541,
0.0376877859,
-0.0133887324,
-0.0175916888,
-0.0317957662,
-0.0414611697,
0.0302669853,
0.0099993143,
0.0029614284,
0.0010954379,
-0.0221104864,
0.0233412888,
-0.0015250064,
0.0210007709,
-0.0017658095,
0.0410598107,
-0.0000660722,
-0.0507954583,
0.0372399166,
0.0423473418,
0.0024123176,
0.0008425874,
-0.056285169,
0.0235264637,
-0.0197792035,
-0.0424901322,
-0.005595047,
0.0071647041,
0.066508323,
-0.0392335244,
0.0519455634,
0.0356673412,
-0.0100069242,
0.0108334329,
0.0312512368,
-0.0448372513,
-0.004522684,
-0.0163864847,
0.0372298472,
-0.0506406575,
0.0018862537,
0.0086802738,
-0.0292975921,
-0.0236867275,
0.0094661266,
0.0336257182,
-0.019440921,
0.0120774768,
0.0084192846,
-0.0973155499,
0.0011994542,
-0.011195913,
-0.0518899411,
0.0212493502,
0.0741142035,
0.0128330011,
-0.0119065046,
-0.0340056568,
0.0234768651,
-0.0380914994,
-0.0287365634,
0.0923382565,
0.0162399821,
-0.0367308967,
0.0703272596,
0.0081408117,
-0.0153200421,
-0.033323098,
0.0113582816,
0.0370129459,
0.067285955,
-0.0104148416,
0.0558596216,
-0.019757539,
-0.0112900427,
0.0098379599,
0.0558618978,
0.0817374736,
0.0428663157,
0.0310987309,
0.0112039363,
-0.0379890129,
0.0159018822,
-0.0276013296,
0.0179243088,
-0.0661593601,
-0.0789960101,
0.0767804608,
0.0135765951,
-0.0233643074,
-0.0293118488,
0.0431672223,
-0.0331547447,
-0.0153753012,
0.0212733559,
-0.0442884564,
0.0465983301,
0.0066547417,
-0.0255225729,
0.0407119505,
-0.0335495546,
-0.0038267842,
-0.0119663877,
0.0316435061,
0.0042488961,
-0.0337114967,
0.0491542667,
-0.0206613485,
-0.0027065694,
-0.0522319898,
0.003888323,
0.0138475383,
0.0047536679,
0.0376459807,
-0.0372765362,
-0.0309069529,
0.0341196135,
-0.0205183495,
-0.0016661922,
-0.0559710301,
-0.041330263,
0.0037731894,
-0.0004019701,
-0.0465302952,
-0.0167999044,
-0.0608889349,
-0.0013465938,
0.011477774,
-0.0419580191,
-0.0042599034,
-0.0025322416,
2.919e-7,
-0.0161412396,
0.0585465655,
-0.0180728883,
0.0102202427,
0.0192039199,
0.0737120956,
-0.0270599183,
0.0246269256,
-0.017767733,
-0.0496328399,
-0.0131727075,
0.0341060348,
0.0517764017,
0.0524526015,
-0.0137068406,
-0.0322556309,
0.0269923881,
0.0001871054,
-0.0676822439,
0.0375758968,
0.06731157,
0.0437118262,
0.0041124998,
-0.0361456014,
0.0414490253,
-0.0266138669,
0.0276876651,
-0.0157234669,
-0.0243206229,
0.0020464533,
0.0092935776,
0.0040609636,
0.0029348163,
0.0356036983,
0.0076535316,
-0.0184422992,
-0.0019010026,
0.0612110384,
-0.0029247222,
-0.0202415399,
0.0008059766,
-0.0371946469,
0.0244987495,
-0.0569953062,
0.0532692075,
0.0057010953,
-0.0285501089,
-0.0148083149,
-0.0517174341,
-0.0538788363,
-0.0385420844,
-0.0292692296,
-0.0082570501,
0.0156211974,
-0.0151142385,
-0.013563592,
-0.0105383852,
-0.056229949,
0.0114891101,
-0.0042369426,
-0.0501633175,
-0.0667540506,
-0.0350167267,
-0.0042494969,
0.0528500117,
-0.0580528267,
-0.0092069386,
2.484229609e-34,
0.0229287446,
0.0175345745,
0.0354098231,
0.0102171199,
0.0359747745,
0.0421490818,
-0.0169182215,
0.0174425449,
0.0323888622,
-0.0602762103,
-0.0234850068
] |
http://arxiv.org/abs/2402.01706v1 | http://arxiv.org/pdf/2402.01706v1 | MULTIVERSE: Exposing Large Language Model Alignment Problems in Diverse Worlds | 2024-01-25T02:57:40 | 2024-01-25T02:57:40 | Xiaolong Jin, Zhuo Zhang, Xiangyu Zhang | cs.CL | cs.CL, cs.LG | Large Language Model (LLM) alignment aims to ensure that LLM outputs match
with human values. Researchers have demonstrated the severity of alignment
problems with a large spectrum of jailbreak techniques that can induce LLMs to
produce malicious content during conversations. Finding the corresponding
jailbreaking prompts usually requires substantial human intelligence or
computation resources. In this paper, we report that LLMs have different levels
of alignment in various contexts. As such, by systematically constructing many
contexts, called worlds, leveraging a Domain Specific Language describing
possible worlds (e.g., time, location, characters, actions and languages) and
the corresponding compiler, we can cost-effectively expose latent alignment
issues. Given the low cost of our method, we are able to conduct a large scale
study regarding LLM alignment issues in different worlds. Our results show that
our method outperforms the-state-of-the-art jailbreaking techniques on both
effectiveness and efficiency. In addition, our results indicate that existing
LLMs are extremely vulnerable to nesting worlds and programming language
worlds. They imply that existing alignment training focuses on the real-world
and is lacking in various (virtual) worlds where LLMs can be exploited. | MULTIVERSE: Exposing Large Language Model Alignment Problems in Diverse Worlds: Large Language Model (LLM) alignment aims to ensure that LLM outputs match
with human values. Researchers have demonstrated the severity of alignment
problems with a large spectrum of jailbreak techniques that can induce LLMs to
produce malicious content during conversations. Finding the corresponding
jailbreaking prompts usually requires substantial human intelligence or
computation resources. In this paper, we report that LLMs have different levels
of alignment in various contexts. As such, by systematically constructing many
contexts, called worlds, leveraging a Domain Specific Language describing
possible worlds (e.g., time, location, characters, actions and languages) and
the corresponding compiler, we can cost-effectively expose latent alignment
issues. Given the low cost of our method, we are able to conduct a large scale
study regarding LLM alignment issues in different worlds. Our results show that
our method outperforms the-state-of-the-art jailbreaking techniques on both
effectiveness and efficiency. In addition, our results indicate that existing
LLMs are extremely vulnerable to nesting worlds and programming language
worlds. They imply that existing alignment training focuses on the real-world
and is lacking in various (virtual) worlds where LLMs can be exploited. | [
0.0278028268,
0.0731602162,
-0.0370299369,
0.0465047136,
-0.0487016179,
-0.0115660569,
-0.0022761836,
-0.0040310565,
0.0274711624,
0.0124894092,
-0.0224600248,
-0.028215602,
0.0328784846,
0.0226836558,
0.0387668274,
-0.0005083502,
0.0701705888,
-0.0244529918,
-0.040143311,
0.0131952558,
0.0169013906,
0.0272596125,
0.0222652238,
0.052768331,
-0.0076739923,
-0.0620594062,
-0.0311766472,
0.035216812,
0.0594855621,
-0.0091987923,
0.0110442592,
0.0234546568,
0.0182200503,
0.0674605966,
0.000002027,
-0.0536830612,
-0.0100559024,
-0.0063345274,
-0.0203839298,
0.049607724,
-0.0118415942,
-0.0119085768,
-0.0327383056,
0.0084994892,
-0.0503998697,
0.0181168914,
0.0593365282,
0.0014208599,
0.0776967779,
0.0635638833,
-0.0010153538,
0.0425799266,
-0.0054388163,
0.0104504004,
0.013865822,
-0.0703697726,
0.0154519407,
0.0120363832,
0.0024044602,
-0.0011658309,
0.0045928094,
0.0055343774,
-0.0069738356,
-0.0126620978,
-0.0516884029,
0.002568285,
-0.0300042797,
-0.0210667551,
-0.0041576615,
0.075066261,
-0.029972747,
-0.0431848168,
-0.0038812209,
0.0114103546,
0.0006996804,
-0.0052777487,
-0.0011703451,
0.0265584011,
-0.0115199834,
0.0104265912,
0.0268777702,
-0.0130439941,
0.0495979041,
-0.0105360746,
0.0022586703,
0.084500581,
0.0191386621,
-0.0457015857,
-0.0500821173,
-0.0264319628,
0.0046819788,
-0.0610345192,
-0.0293330438,
-0.0152476244,
-0.043377094,
0.0258687027,
-0.0826971978,
0.0026684087,
0.0348491147,
-0.0710663646,
-0.0022878246,
0.0570639372,
0.0019486316,
0.0360675007,
0.0178133082,
0.0550620556,
-0.0388649963,
-0.0214094147,
-0.0069133281,
0.0297206175,
-0.0082159853,
-0.0270637292,
-0.0034295137,
-0.0118680876,
-0.0186515171,
0.0114953397,
-0.0297899004,
0.0648881122,
-0.032899335,
0.0619808771,
-0.0699412897,
-0.00878511,
-0.0261375625,
0.0074662156,
0.0418793894,
-0.002710816,
-0.0230037365,
-0.0041928235,
-0.017177714,
-0.0715415701,
-0.017401265,
0.0171967335,
-0.0604206212,
-0.0194079485,
-0.0028200701,
0.1199028417,
0.018966347,
0.0290323272,
-0.0353683829,
-0.062357571,
-0.0306894016,
-0.0215404481,
-0.0333736502,
-0.0859064683,
-0.0137669938,
0.0016326569,
-0.0086551011,
-0.0520125553,
0.0164965261,
-0.0096163414,
-0.0366427563,
0.0658013374,
-0.0299891513,
0.0080578271,
0.039770063,
0.027690921,
0.0437944531,
0.061962828,
0.0044506299,
0.0259852801,
0.0216813814,
-0.0522934794,
0.0203891192,
0.0735860541,
-0.0154592339,
0.0275906473,
-0.0249080453,
-0.0031921132,
0.0937729701,
0.0382051058,
0.0095476909,
0.0236724801,
0.0016740175,
-0.0016850926,
0.0506518707,
0.0333927907,
-0.0409826711,
0.0193840582,
0.030045405,
-0.0068594636,
-0.0393506661,
0.0187042803,
-0.0769481361,
-0.0029205673,
-0.0732864738,
0.0007360677,
-0.0233888458,
0.0749911666,
-0.0075119524,
0.02300613,
-0.0033407868,
0.0013359905,
0.0004163902,
-0.0130120944,
0.0035608306,
-0.0105523393,
-0.009116617,
0.0553410389,
0.0495775938,
-0.053804826,
-0.0141072702,
-0.035372179,
-0.0016569387,
0.0835651159,
0.0526339635,
-0.0200885255,
-0.0375381894,
-0.0143298376,
0.0015063775,
0.028652858,
0.0142316399,
0.0053198617,
0.0026835646,
-0.0233273059,
-0.0096042445,
-0.0493072681,
-0.0117047895,
0.0572620705,
0.0140538123,
-0.007261992,
-0.0091388412,
-0.0461012684,
-0.0254667103,
0.0235518999,
-0.0241491925,
-0.0219400935,
-0.0469791517,
0.046374511,
0.0253715198,
-0.0134742754,
-0.0425706394,
-0.0110851713,
0.0250547361,
-0.0283358805,
0.0155844809,
0.0498171411,
-0.0630450472,
0.0014526895,
0.0699919835,
-0.0757196024,
-0.0213832296,
0.0595002808,
-0.0158911105,
-0.0084366733,
0.0334010907,
-0.0175137892,
0.0431338549,
0.0197562091,
0.0622088425,
-0.0412822664,
0.0052323253,
-0.0470823124,
0.0411229469,
-0.0131055815,
-0.0054540033,
-0.0202785376,
-0.0048026857,
0.0204989053,
0.0070172618,
-0.030848423,
0.0096449507,
-0.0226128716,
0.0325823054,
0.0417768471,
0.018011488,
-0.024948841,
-0.0523321144,
-0.0778855458,
-0.0050615883,
0.0027594282,
0.0356259644,
0.0658957139,
0.0533106737,
-0.024799658,
-0.0091842357,
-0.0335587636,
-0.0069771046,
0.0129543655,
0.0153470682,
0.0411406159,
-0.0065381192,
-0.051167205,
-0.0339773484,
0.0126986876,
-0.0567102693,
0.055078011,
-0.0243649762,
-0.021283187,
-0.0756956413,
0.0277140401,
-0.0098139066,
0.0110194702,
-0.0166472085,
-0.0495743454,
0.0087845977,
0.0167206675,
0.0178066734,
0.0628670231,
0.0088994913,
0.0464440957,
0.0679350793,
0.0157854203,
-0.0774629712,
-0.0319733508,
0.0491439365,
-0.0473187976,
-0.0463125482,
-0.0701022074,
0.0311602801,
0.084743008,
0.018523043,
0.0241904948,
-0.0278085377,
0.0496022142,
-0.0017197077,
-0.0368237495,
0.0150434356,
-0.0615124255,
-0.0364906974,
0.0018713558,
-0.0275950246,
-0.0773923323,
0.0774767324,
-0.0645261109,
0.0113687217,
0.0038610676,
-0.0398841947,
-0.0042064395,
0.0273986198,
-0.0288234949,
0.058282271,
0.0092199547,
-0.069334209,
-0.0439832732,
-0.0073295105,
0.0491085611,
-0.0013014385,
0.0455066077,
-0.057589218,
-0.0054947007,
-0.0069399974,
-0.0535199419,
-0.0477518775,
-0.053610146,
-0.0050216247,
-0.0360117741,
0.0349958763,
-0.0258186143,
-0.0345833041,
-0.0076181157,
0.0242362916,
-0.0010625114,
0.0110101979,
-0.0086333361,
-0.0203419067,
-0.0407800041,
0.0493750945,
-0.0545836426,
-0.0066896831,
-0.0340232104,
0.0271261837,
0.0472388156,
0.0126381991,
0.0174075309,
-0.0279730428,
0.006963545,
0.0298438333,
-0.0173471309,
-0.0218528453,
0.0224182177,
0.0770249069,
-0.0361523367,
-0.0147199705,
0.0755668059,
0.0256591775,
-0.0159295686,
0.036701303,
-0.0350368507,
0.0103977062,
0.0274503771,
0.0218390282,
-0.0835413039,
-0.0224799085,
-0.0078668501,
0.0108489925,
-0.0253352337,
-0.0258002132,
-0.0042220517,
-0.0892529264,
-0.0262927208,
0.067683816,
-0.0376280248,
-0.0181076545,
0.0123918876,
-0.0025810599,
0.0107372981,
0.0619595461,
-0.0297201071,
0.0066893795,
-0.0163439624,
0.001224994,
-0.0314785466,
-0.0157347713,
0.0259264987,
-0.0304439366,
0.0633378997,
0.0847757161,
-0.007004994,
-0.0256437175,
-0.0252358578,
-0.0001990906,
0.0503419079,
0.0350511819,
0.0178163629,
0.0022120455,
-0.0182751138,
-0.015312939,
-0.0085737789,
0.0299508907,
-0.0148936464,
-0.0344303921,
0.0393558219,
0.072307758,
-0.0246025529,
-0.0003597097,
0.0005192219,
-0.0140817128,
0.034962222,
0.0422305055,
0.0012815008,
-0.0635020286,
-0.015762601,
-0.0405171141,
-0.0187404845,
0.0229625758,
-0.0209918693,
-0.0031026818,
0.0080730421,
-0.0020411524,
-0.0276920814,
-0.0151960524,
-0.0086240312,
0.0437428616,
0.059850771,
-0.0329681858,
-0.0120938169,
-0.0113519439,
-0.0357769653,
-0.0032843603,
-0.03330772,
0.0460118279,
-0.0244696476,
0.0440941527,
-0.0225898791,
-0.0353855602,
0.0255985428,
0.0145701226,
-0.068081513,
-0.1101652831,
-0.0068982509,
0.0334998742,
0.0219787452,
0.0642186627,
-0.0158005133,
-0.0171288792,
-0.0061768172,
0.0393949635,
-0.0494277664,
0.0379853547,
0.0377349481,
-0.0031825684,
0.0191829633,
0.0065440661,
-0.036449898,
-0.043470785,
0.0530574657,
-0.0041456558,
-0.0833594799,
-0.0408354327,
0.0491040535,
0.0436886512,
0.0122589068,
-0.0452996902,
0.001009661,
0.0020586741,
-0.0042641591,
-0.0021363695,
-0.0541517399,
0.0086305263,
-0.0113381352,
0.0403694399,
-0.0036367944,
0.0810043812,
0.0123892333,
-0.0155666973,
-0.0015824978,
-0.0158567335,
0.0225387365,
0.005319349,
0.0631250143,
0.0676878765,
0.0342193916,
-0.0243106876,
0.0364277177,
-0.0109591903,
0.0656187832,
-0.0040662591,
0.0643346012,
0.0158358812,
0.0195570234,
0.001904441,
0.0395236388,
-0.0336620398,
-0.0446233973,
-0.001913005,
0.0437716693,
0.0534592904,
-0.0006767396,
0.047960788,
-0.0110183926,
0.0065486715,
-0.0051812194,
0.0206438452,
0.0070005353,
-0.058324948,
-0.018670829,
0.0175477881,
-0.048971124,
-0.0001858646,
0.0428452268,
-0.0267527848,
0.0584865212,
0.014224791,
-0.0298601277,
-0.0412360206,
0.009351613,
0.0700270012,
0.0148759242,
-0.0371426903,
-0.0383204445,
-0.0777598247,
-0.0506711416,
-0.0245573781,
0.0317908712,
-0.0112147117,
-0.0014020221,
-0.0217951965,
0.0728838295,
0.0087111788,
-0.007780306,
-0.0321133174,
-0.0080710482,
0.0042313398,
-0.0302831195,
-0.02167801,
-5.6635597e-33,
-0.0232198574,
-0.0347077847,
-0.025942497,
0.0011070168,
-0.0299839638,
-0.0037105402,
-0.0294300653,
0.0207594801,
-0.0319999829,
-0.0163381565,
0.0154293496,
-0.0117409127,
0.0005649595,
-0.0313293971,
0.0098257232,
0.0136561105,
0.0298074279,
0.036119543,
-0.0640794337,
0.0133255394,
0.059521459,
0.0051093316,
0.0576436892,
-0.0873442069,
0.0097063845,
-0.042911917,
-0.0386161134,
0.0278850924,
0.0082708737,
0.0247594137,
0.0117475381,
0.0133539969,
0.0017177813,
-0.010608105,
-0.00030751,
0.0303379279,
0.0188661329,
-0.0212064721,
0.0565768518,
0.0235972758,
0.0310597494,
-0.0515173748,
0.0050597643,
-0.0227464586,
-0.0016527064,
0.0008335189,
0.0319840573,
0.0105903558,
-0.0076424382,
0.0224078353,
-0.0849038512,
-0.0141371395,
-0.0695855319,
0.0008555199,
-0.0367421247,
0.0176753979,
0.0005488491,
-0.0183559936,
-0.051170405,
0.0390818864,
0.0186042599,
-0.0151887098,
0.1024691314,
0.0363390446,
0.0059236824,
0.0817372501,
-0.0140029863,
0.0201693587,
-0.007547237,
0.0065952577,
0.0231430139,
0.0263481792,
-0.0101887733,
0.0816930234,
-0.0097355293,
-0.0191683304,
-0.0349587053,
0.0109055433,
0.0182408821,
0.0028683336,
0.0067555215,
0.007975216,
-0.0825613067,
-0.0110652158,
-0.0128564434,
0.0249155238,
-0.0017898733,
-0.0599829927,
0.0259893406,
-0.0524908826,
-0.0169369001,
-0.0351998806,
0.0060355556,
-0.0745555907,
0.0132479277,
0.0242915247,
-0.0234309342,
-0.0260370243,
-0.0182164647,
0.0474721156,
-0.0040767724,
-0.0273325667,
-0.0073900782,
-0.0105967689,
0.0778569728,
0.0332155973,
0.0339261182,
0.0033965297,
-0.0594439656,
0.015145231,
-0.0333605669,
0.0108360648,
-0.0245905519,
-0.0104663456,
0.0593897365,
-0.0513839275,
-0.0066667031,
-0.0122499205,
0.0489926599,
-0.000836619,
-0.0442127362,
-0.0159906279,
-0.0273742937,
0.0235989988,
-0.0466729067,
-0.0167538486,
-0.0359192453,
0.0102228289,
0.0749995038,
-0.0599100776,
-0.0209824704,
0.0053130523,
2.81e-7,
-0.0190085154,
0.0168788489,
0.0368266143,
0.0059881653,
-0.0055593904,
0.0889922902,
-0.0115538007,
0.0706516802,
0.0030873576,
-0.0285030063,
-0.0018719831,
-0.0066436562,
0.0613513738,
0.032185629,
-0.0481836461,
0.0530094653,
-0.0428355597,
0.0072682435,
-0.0764825791,
0.0198587291,
0.0360799246,
0.0604991205,
0.0169891436,
-0.0061814915,
0.0376011021,
-0.0393731967,
0.0105172638,
-0.0185449552,
-0.0201850291,
0.0184090789,
-0.0535443723,
0.0019068234,
-0.006539084,
0.0269341618,
-0.0073621478,
0.0040033339,
-0.0395325944,
0.0392120592,
-0.0366871804,
0.0298789106,
-0.047784891,
-0.0269767214,
-0.0212806631,
-0.0659326836,
0.0212124661,
0.0094517572,
-0.037956886,
-0.0241837129,
-0.0045334571,
-0.0349683389,
-0.0058253719,
-0.0113794794,
0.0326247588,
0.0078128725,
0.0068501555,
0.0139237735,
-0.0112722283,
-0.0344556943,
-0.034470059,
0.0286522228,
-0.0147851622,
-0.0496977977,
0.0104959998,
0.0190435182,
0.0610219277,
-0.0336123146,
0.0345652103,
2.744229903e-34,
-0.0288986992,
-0.0112806959,
0.055821564,
0.0046696565,
0.0265217219,
0.0153721264,
-0.0027805155,
0.0523235723,
0.0230731312,
-0.0341417044,
0.002719711
] |
http://arxiv.org/abs/2401.13907v1 | http://arxiv.org/pdf/2401.13907v1 | No More Distractions: an Adaptive Up-Sampling Algorithm to Reduce Data Artifacts | 2024-01-25T02:54:53 | 2024-01-25T02:54:53 | Han Chen | cs.CL | cs.CL | Researchers recently found out that sometimes language models achieve high
accuracy on benchmark data set, but they can not generalize very well with even
little changes to the original data set. This is sometimes due to data
artifacts, model is learning the spurious correlation between tokens and
labels, instead of the semantics and logic. In this work, we analyzed SNLI data
and visualized such spurious correlations. We proposed an adaptive up-sampling
algorithm to correct the data artifacts, which is simple and effective, and
does not need human edits or annotation. We did an experiment applying the
algorithm to fix the data artifacts in SNLI data and the model trained with
corrected data performed significantly better than the model trained with raw
SNLI data, overall, as well as on the subset we corrected. | No More Distractions: an Adaptive Up-Sampling Algorithm to Reduce Data Artifacts: Researchers recently found out that sometimes language models achieve high
accuracy on benchmark data set, but they can not generalize very well with even
little changes to the original data set. This is sometimes due to data
artifacts, model is learning the spurious correlation between tokens and
labels, instead of the semantics and logic. In this work, we analyzed SNLI data
and visualized such spurious correlations. We proposed an adaptive up-sampling
algorithm to correct the data artifacts, which is simple and effective, and
does not need human edits or annotation. We did an experiment applying the
algorithm to fix the data artifacts in SNLI data and the model trained with
corrected data performed significantly better than the model trained with raw
SNLI data, overall, as well as on the subset we corrected. | [
0.0165496394,
0.0753825605,
-0.0349455774,
0.0047889138,
-0.0288260337,
0.0180037096,
-0.0304373428,
0.0164215583,
-0.0310504492,
-0.0433863588,
-0.0321036056,
0.0115851061,
0.0432058573,
0.0386366174,
0.0237912703,
-0.0390285738,
0.0834975317,
-0.0426365212,
0.0388340391,
-0.001668398,
-0.0024181844,
0.0086950818,
0.0209870804,
0.0199336521,
-0.0617813878,
-0.0275401529,
0.0048321509,
0.0013438058,
-0.0391205214,
-0.0022414967,
0.0038821248,
0.0582439154,
-0.0105920685,
0.0494471416,
0.0000018688,
-0.0584616773,
0.009681711,
-0.0293072648,
-0.0175823029,
0.0415779948,
0.055981826,
-0.0062305117,
-0.0299133081,
0.0466880649,
-0.0228157118,
0.0032216585,
0.0337848626,
0.0491005518,
0.0212999675,
0.0265800972,
-0.0158543233,
0.0470280983,
-0.0152379423,
0.0002244298,
-0.0372602604,
0.0133507252,
0.040608447,
0.0085240323,
-0.0187010225,
-0.0394336507,
-0.0362544842,
0.0273288563,
-0.0364917889,
-0.0309010856,
-0.0173515044,
0.0353091918,
-0.0386959612,
-0.0308191869,
-0.0110804588,
0.0620909967,
-0.0605305135,
-0.0204944797,
-0.0263075549,
-0.0417053103,
0.0135993659,
0.0122467764,
-0.0427924991,
0.0036807505,
0.0145741124,
0.0582438074,
0.0340711921,
-0.0474088006,
0.0431493521,
0.0093808053,
-0.0333241262,
0.0564192571,
0.0179365817,
-0.0409332104,
0.006230603,
0.0020476915,
0.001955535,
-0.0549273714,
0.035710603,
0.0247169882,
-0.0541006476,
0.0388832428,
-0.0407643653,
-0.0900240764,
0.0218464341,
-0.0316873454,
0.0002657105,
0.0085046832,
-0.0204810668,
0.0336631201,
-0.0237965062,
0.004580061,
0.0324012078,
0.0591617934,
-0.0069913724,
-0.0210826583,
0.0239782147,
-0.0038017479,
-0.0911685005,
0.0191400163,
-0.0408558883,
0.0029510907,
-0.0239501856,
0.0078608552,
-0.0151892528,
-0.0244424585,
0.0067664059,
-0.0158515964,
-0.0321363434,
0.0242877547,
-0.0137089444,
0.0343620367,
-0.0324531868,
0.0323344991,
-0.0180245936,
-0.0907974467,
-0.0021186774,
0.0027819087,
0.0242717378,
-0.0250494275,
0.0043362393,
0.0356375687,
0.0072885919,
-0.0382205434,
-0.0487765819,
0.0156902652,
-0.0207564253,
-0.0042891316,
-0.0246539153,
-0.0656188205,
0.0096781254,
-0.0077099474,
-0.0050092041,
-0.043754749,
-0.013931673,
0.0311632194,
-0.0886891931,
0.0484643392,
-0.099208422,
0.0264565572,
-0.0061470941,
-0.0384062938,
0.0029459305,
0.0631761476,
0.0535448939,
-0.0075791599,
0.0523360893,
0.0168272704,
0.0171889011,
0.0597226582,
0.0204156879,
0.0068839951,
-0.0566831194,
-0.040720433,
0.0092535317,
0.0206558239,
0.0778535604,
0.06188979,
-0.001464963,
0.0005244655,
0.0392100662,
0.029463673,
0.0050868765,
0.0062232264,
-0.0382318869,
0.0167666879,
0.0107496073,
0.0000606902,
-0.0409302339,
0.015841797,
0.0077270106,
0.0110147744,
-0.0027119822,
0.0129639227,
-0.01894545,
-0.0447781719,
-0.0221268218,
-0.0248542577,
0.0274290107,
-0.0149408821,
0.0921027809,
0.0500908643,
-0.0330344252,
0.0437664203,
0.0527771525,
-0.0244820602,
0.0209905878,
-0.0405066758,
0.0236352906,
0.0623677075,
0.0436165631,
-0.0325042978,
-0.0106830783,
-0.0099475048,
0.0051109018,
0.0302169975,
0.0242526252,
-0.0036535461,
-0.0028290355,
-0.0353462473,
-0.0290718339,
0.0466668122,
0.0206502546,
-0.0043186382,
-0.0225685406,
0.015833877,
-0.0010024321,
-0.0653265417,
-0.0391137004,
0.0539970919,
-0.0243443307,
0.0390908308,
0.0357324779,
0.0031568063,
0.0325743221,
0.0100333076,
-0.0161694214,
0.0142895747,
0.0160998218,
-0.0071387244,
-0.0054607573,
0.003414202,
-0.0149546862,
-0.0242857225,
0.0529466569,
-0.0288293399,
-0.0259003639,
0.0356408097,
-0.0278694816,
-0.0471946672,
0.026977228,
-0.0339215249,
0.03323365,
-0.049669154,
0.0455042645,
-0.0341133587,
0.0425591655,
-0.0320922844,
0.0351177715,
0.0027791834,
0.015413682,
0.016455099,
0.0261565428,
0.0179114845,
-0.023213435,
0.0009151792,
0.0308165886,
-0.0462275557,
-0.0073311776,
0.0198664516,
-0.0040323455,
-0.0539554507,
0.0130136209,
-0.0734729692,
-0.0261549279,
-0.0350348018,
0.0055482937,
0.0091529563,
0.0272498876,
-0.0305297878,
-0.0415024869,
0.0040809605,
-0.0108652208,
0.0213630702,
0.0121004339,
0.0451420583,
-0.0056603863,
-0.0086547108,
-0.058402136,
-0.0121417027,
-0.036267288,
0.0192294065,
-0.0066950233,
-0.0972142071,
-0.0765930489,
0.0467747524,
-0.0192194227,
0.0197115205,
0.0125502357,
-0.0256971251,
0.0113809286,
-0.0084073311,
0.0524184667,
0.0108530605,
0.012225912,
0.0032514147,
0.0584943406,
-0.0214688238,
-0.0463448241,
-0.0052417521,
0.0097224014,
0.0301754531,
-0.0082769264,
-0.0112280548,
0.016635485,
0.058570426,
-0.0105705205,
-0.001361912,
-0.0034671724,
-0.000063155,
-0.0500595495,
-0.0122799492,
-0.0275674574,
0.0003188692,
-0.0561712943,
0.020758234,
-0.0394101813,
-0.0247538183,
0.0220258906,
0.0061586243,
0.0186875798,
-0.0708153844,
0.0296542142,
0.007590069,
0.0423950031,
-0.0081488714,
0.0109795332,
0.0057905265,
-0.0861562639,
-0.054683499,
-0.024368763,
-0.0334115066,
-0.035028033,
0.0295154415,
-0.0529765487,
0.0284724757,
0.0092717027,
-0.0181576721,
-0.0471716225,
-0.0493327044,
-0.0068637803,
-0.0064812778,
0.0002654943,
0.0016605164,
-0.0142097706,
-0.0399296172,
-0.034069553,
0.0537286587,
-0.0045988509,
0.0007142039,
-0.0206458438,
0.0194220375,
0.0775695071,
-0.002788651,
0.039902892,
-0.0074480372,
-0.0113395071,
0.0444825664,
0.0425322801,
0.0485809781,
-0.0282048509,
-0.0024916586,
0.0404895581,
-0.0394041874,
-0.0248325262,
0.0143599771,
0.0458202325,
-0.0000164007,
-0.0364630781,
0.1434372067,
0.0370434187,
-0.017163761,
0.0412579253,
-0.0152100483,
0.0413547046,
0.0305269975,
-0.0116351768,
-0.0880581811,
-0.0249079857,
0.0332331248,
-0.0091567272,
-0.0024512496,
-0.0169028044,
-0.0017058124,
-0.1237318814,
0.0345691368,
0.0068749422,
-0.0712655261,
-0.0182351451,
-0.0689278543,
0.0294035338,
-0.0034199397,
0.1033043936,
0.0335288793,
-0.017638443,
0.0027520435,
0.030941559,
-0.0358805098,
-0.0883886814,
-0.0186729953,
-0.0645088479,
0.0653093606,
0.0051943967,
0.0572378673,
-0.0274845157,
0.0007215185,
0.0676177144,
0.0048175785,
0.0191614833,
-0.0156537872,
0.0607875325,
0.0365556329,
-0.0260310285,
-0.0220911503,
0.0247467272,
-0.051755134,
-0.0155022806,
-0.0282022804,
0.0596464239,
-0.0190055706,
-0.0062935562,
-0.0314656794,
-0.0326020978,
0.0209282711,
0.045934204,
0.0628337041,
-0.0622596331,
-0.0167924445,
-0.0154763944,
0.0093451086,
-0.0094732568,
0.063993983,
0.0532483123,
-0.0275298692,
-0.0537927672,
-0.0696434155,
-0.0657073185,
0.0639491379,
0.0345757529,
0.0283118412,
0.0401523225,
-0.014295415,
-0.0068746852,
-0.0397800431,
0.0243147016,
-0.0451013073,
0.0375852734,
-0.0197142381,
-0.0003983657,
-0.0399860889,
0.0136173889,
-0.0229186248,
0.0074047782,
-0.0641209707,
-0.0457631201,
0.0777971447,
0.0372941084,
0.0241572,
0.040561676,
-0.0405558683,
-0.0547967665,
0.0107505033,
0.062783137,
-0.0219914354,
-0.0027280103,
-0.0122080185,
-0.0629052818,
0.0105241276,
0.0356038287,
-0.0641385391,
0.0142467888,
0.0103602149,
0.0382592939,
-0.0375835747,
-0.0006683741,
0.0241439175,
0.0287108142,
0.0781017989,
-0.0193815175,
-0.0037492593,
-0.0447560623,
0.0072956919,
-0.0314829275,
-0.0039655068,
-0.0148883862,
0.0114422441,
0.002579543,
-0.002102667,
-0.0317562073,
-0.0169686023,
-0.0628158152,
-0.0012656651,
-0.0051865038,
0.0010661972,
-0.0164872166,
0.0727784336,
0.033904627,
0.0217162948,
-0.0087047974,
-0.0013005417,
0.0329297781,
0.0164776389,
0.0070985002,
-0.0127949901,
0.021041669,
0.0192368384,
-0.0453508049,
0.0067588165,
-0.0194013938,
0.0101125501,
-0.0250825193,
0.0020698756,
0.0645402968,
-0.0321553126,
0.0695809796,
-0.0122352261,
0.0034906743,
-0.0208932329,
0.0370083265,
0.0009661327,
-0.0399752036,
0.0195734892,
-0.0070780045,
-0.0359465443,
-0.0126906456,
0.0635461286,
-0.0196717214,
0.0452814735,
-0.0306675583,
-0.0052076587,
-0.0110124089,
0.0160548761,
0.0062770462,
0.0235918798,
-0.0333880968,
-0.014213359,
-0.0257818755,
-0.0796209201,
-0.023059113,
0.0026810605,
0.0191964135,
0.009045057,
0.0378451161,
0.0483775064,
0.0008512784,
0.0018500921,
-0.0336377583,
-0.0127729839,
-0.0005270009,
-0.0718233138,
-0.0306998435,
-6.070816698e-33,
-0.0247316677,
-0.030292144,
-0.0213456266,
-0.0127928536,
0.0075072926,
-0.0009103104,
-0.0423212126,
0.0143471258,
-0.0333325304,
0.0320934653,
-0.0437278189,
0.0328188539,
0.0287165567,
0.0197332408,
0.0140770841,
0.0243962314,
0.0617018305,
-0.0174519122,
-0.0211929549,
-0.0038522868,
0.0485984273,
0.0680773333,
0.1021641791,
-0.0619487502,
0.0015118649,
0.0271485969,
-0.0096656345,
0.0146572804,
0.0079331081,
0.0576994643,
0.0038923756,
0.0766477957,
0.0210416559,
0.0154098235,
-0.0634832233,
0.0246942919,
-0.0951843709,
-0.0484356582,
0.0182137284,
0.0904552042,
-0.0071569053,
-0.011379526,
0.0935250372,
-0.0207974594,
-0.0463027731,
0.028008787,
0.0567107797,
-0.0358892344,
-0.0561478287,
0.035448283,
-0.057446342,
0.0200142916,
0.0037564596,
0.0229775533,
0.0112769827,
0.013544525,
-0.0137151917,
-0.0109953638,
-0.0440464951,
0.0297127869,
0.042778831,
0.0495243706,
0.0330829695,
0.0126490975,
0.0073909997,
0.0800514072,
0.021149341,
-0.0019694627,
0.0495163724,
-0.012546354,
0.0857659057,
0.0088371392,
0.0226818416,
-0.002489056,
0.059743464,
-0.0646894872,
-0.0436763205,
0.0125168199,
0.0320357345,
0.0175600424,
-0.006175858,
0.0174357127,
-0.0393292941,
-0.0424916223,
-0.0417533293,
-0.0045476598,
-0.0271557309,
-0.0349480286,
0.0262857266,
0.0042252652,
-0.003735794,
-0.0008715808,
0.0065990887,
-0.0280552879,
-0.0802779347,
-0.0142838173,
-0.0021200299,
0.0137388622,
-0.0071732299,
-0.002891989,
-0.0359080657,
-0.0372590125,
-0.0057103718,
0.0251041707,
0.0275649615,
0.0198508836,
0.0669842586,
0.0178236272,
-0.0834229738,
0.0212759823,
-0.004027904,
-0.0101946956,
0.0156510845,
-0.0112694874,
0.0436197333,
-0.0243778471,
0.0105709126,
0.0562116355,
0.0302802566,
-0.0165191479,
-0.0328384079,
-0.0412685685,
-0.006500727,
-0.004736776,
-0.0488431491,
-0.0392644294,
-0.0297598671,
0.0065675317,
0.0619669743,
-0.0653605387,
-0.0279806387,
0.0000781149,
2.54e-7,
0.0585495345,
0.0771295801,
0.0336113907,
0.0210439004,
0.0402146727,
0.0021723283,
-0.0332657769,
0.0223476421,
-0.0027316536,
0.043050468,
0.0095414734,
-0.034485817,
0.0164322015,
0.0051597133,
-0.1025864407,
0.0138359712,
-0.0574074835,
0.0081047183,
-0.0405834764,
-0.0036786802,
0.0123133268,
0.0944985822,
0.0195863955,
-0.0396101288,
-0.0349305831,
-0.0330399349,
0.0241968986,
-0.0247852225,
-0.009231382,
0.0285288617,
0.0072537027,
0.0199283715,
0.0346852802,
0.012489548,
-0.041552294,
-0.0103220185,
0.024944745,
0.0035210575,
-0.0536164157,
0.0334163979,
-0.001664394,
0.0086074499,
0.0138947172,
-0.0370486379,
0.0551421568,
-0.0305699427,
-0.0354967602,
-0.056163013,
-0.0238215979,
0.0193230007,
0.0032354123,
0.0043217824,
-0.020882776,
0.0274076313,
0.0446861871,
0.0272345152,
-0.038457077,
-0.0229700748,
-0.0098814704,
-0.0235567726,
-0.0076344335,
-0.0284633487,
-0.0127656078,
0.0108398283,
0.0375330336,
-0.0367262252,
0.0193551946,
2.239123654e-34,
0.0665525571,
-0.0092917979,
0.0251515321,
0.0337704942,
0.0156735182,
-0.0230558608,
0.0378375612,
0.0414697416,
-0.0023548349,
0.000850498,
-0.0162232835
] |
http://arxiv.org/abs/2401.13905v1 | http://arxiv.org/pdf/2401.13905v1 | Dynamic embedded topic models and change-point detection for exploring literary-historical hypotheses | 2024-01-25T02:50:03 | 2024-01-25T02:50:03 | Hale Sirin, Tom Lippincott | cs.CL | cs.CL | We present a novel combination of dynamic embedded topic models and
change-point detection to explore diachronic change of lexical semantic
modality in classical and early Christian Latin. We demonstrate several methods
for finding and characterizing patterns in the output, and relating them to
traditional scholarship in Comparative Literature and Classics. This simple
approach to unsupervised models of semantic change can be applied to any
suitable corpus, and we conclude with future directions and refinements aiming
to allow noisier, less-curated materials to meet that threshold. | Dynamic embedded topic models and change-point detection for exploring literary-historical hypotheses: We present a novel combination of dynamic embedded topic models and
change-point detection to explore diachronic change of lexical semantic
modality in classical and early Christian Latin. We demonstrate several methods
for finding and characterizing patterns in the output, and relating them to
traditional scholarship in Comparative Literature and Classics. This simple
approach to unsupervised models of semantic change can be applied to any
suitable corpus, and we conclude with future directions and refinements aiming
to allow noisier, less-curated materials to meet that threshold. | [
-0.0040063523,
0.0489909947,
-0.0550419055,
0.0345397182,
-0.0334726237,
-0.0044031516,
-0.0266738571,
0.0221925955,
-0.0603137724,
-0.0874983147,
0.0358292684,
0.0039339368,
0.0458293706,
-0.0158605613,
0.0166953858,
-0.0013921558,
0.0508005396,
-0.0006647616,
-0.0507739112,
0.0272121355,
0.001950119,
0.0025304649,
-0.0074416394,
-0.0104413833,
0.035927128,
-0.025968302,
-0.0564599782,
-0.0085434606,
-0.0020085278,
0.0109316744,
0.0035610173,
0.0514364615,
0.0424576402,
0.0417136662,
0.0000015836,
-0.044689294,
-0.0429622941,
-0.0474961773,
0.0300214235,
0.0433035232,
0.0589348078,
0.0286934376,
0.0061205276,
-0.0861018822,
-0.0268391892,
0.01892839,
-0.0201345049,
-0.0064281919,
-0.0402776003,
0.0436720699,
-0.0146178482,
0.0028030789,
0.0074527632,
-0.0126831969,
0.012915316,
0.0009359769,
0.0191462748,
-0.003819633,
0.0065929387,
0.0610453673,
-0.0047365734,
0.0078357663,
-0.0470688045,
-0.0100891488,
-0.0118370429,
-0.0213834103,
0.0150261261,
0.0582845658,
-0.0161745083,
-0.0524238758,
0.0342934392,
-0.0344382264,
0.0492420234,
0.0158819929,
-0.0114717688,
0.020682808,
-0.0199807305,
-0.0438732654,
0.0019814435,
0.0004212798,
0.1190944314,
-0.0122435875,
-0.0097444886,
0.0227879304,
-0.0249507688,
0.046846766,
0.0502735302,
-0.0496307574,
-0.0360867865,
-0.00029084,
-0.0019196997,
-0.0640431643,
-0.0487987362,
0.0248549096,
0.0027385391,
0.0102574946,
0.0114114638,
-0.0256769788,
-0.0275836308,
-0.0313501209,
0.0070947171,
-0.0290545579,
-0.0464795008,
0.0162477475,
0.0519199409,
0.0048450474,
0.0118826097,
-0.0763651282,
-0.0229725577,
0.0354593769,
-0.0191345457,
-0.0104262903,
-0.0298987422,
-0.0025038677,
-0.0083943633,
0.0291644856,
0.0190230571,
0.0392525718,
0.0395424329,
-0.0298421234,
-0.0770511925,
-0.048221793,
-0.035517592,
-0.0110608721,
-0.0098802811,
0.0238358155,
-0.0258527435,
0.0222915225,
-0.0714899227,
-0.0123115564,
-0.0017506597,
0.0103812087,
0.0103787221,
-0.0322783031,
-0.023890581,
0.0333915465,
-0.0139384288,
-0.0354047157,
-0.0288258307,
-0.0167297944,
-0.0462861694,
0.0081552323,
-0.014912514,
-0.0586898588,
-0.0258368067,
-0.047836341,
0.0097103203,
0.0049076583,
0.0024281163,
0.0190074425,
-0.0314052999,
0.010395458,
-0.0149128698,
0.0105968621,
-0.0026871231,
-0.0332925059,
0.0425108261,
0.0214820076,
0.0276175383,
-0.0216991808,
0.0836637393,
0.0339456759,
-0.0626220778,
0.0146341827,
-0.0165135879,
0.021348929,
-0.0453157239,
-0.0196423661,
0.0278485492,
0.0256751049,
-0.0218314417,
-0.0097233187,
-0.0160767976,
-0.0062361211,
0.0447197184,
0.0564326569,
-0.0071560084,
0.0081900153,
0.0134968497,
-0.0183463562,
-0.0012688882,
0.0389595367,
0.0049763466,
-0.0585980229,
0.0385521203,
-0.0278133228,
-0.0602195486,
-0.0068544876,
-0.0038833979,
-0.008583568,
-0.0117323976,
0.0179454107,
-0.0335036144,
-0.0752361417,
0.0559722558,
0.0149975503,
0.0203600042,
0.0333393477,
0.013773133,
-0.0244805645,
0.0263595078,
0.013182655,
0.0851164758,
0.0466654524,
0.0548201688,
-0.059986908,
0.0080424594,
0.0228956752,
0.0032417825,
0.0158287454,
-0.0609382987,
-0.0260970276,
0.0046503898,
0.0076257209,
0.0388137549,
-0.0216522068,
0.0345891975,
-0.0112490915,
-0.0563697591,
0.0163427796,
-0.0315998644,
-0.0628133714,
0.0546914116,
0.0573873632,
-0.0237477031,
0.0171915442,
-0.0237180442,
-0.0313388295,
0.0023381365,
-0.0209940188,
0.0019294817,
-0.0196911842,
0.0227854773,
-0.0440249294,
0.002859083,
0.075256139,
-0.0056811008,
0.0359953456,
0.0708677173,
-0.0408406481,
0.0331338681,
-0.0394686759,
0.0207249168,
-0.0135675324,
0.0509298965,
-0.0598386899,
0.0240457524,
0.0263533164,
0.0097246962,
-0.0332910493,
-0.0237811934,
-0.0495788716,
0.0517219417,
0.006421986,
0.0499278158,
0.0313478597,
-0.0370770507,
0.0167709831,
0.0299455747,
-0.0020664032,
-0.0257257279,
0.0175050534,
-0.0546875559,
0.0258418098,
-0.0376580544,
0.0162564423,
-0.0644462854,
-0.0400196686,
0.0274398625,
-0.0567023158,
0.000074978,
0.0701491162,
0.0150484648,
-0.0071807262,
0.034944728,
0.0303117223,
0.025902817,
-0.0408919267,
-0.0094174957,
0.0667641088,
-0.026931189,
0.0275375526,
-0.0591687113,
-0.0111679994,
-0.0161182694,
0.0684513152,
0.0122154336,
-0.0278061051,
-0.0515403524,
-0.0171944872,
-0.0174995307,
0.0133338869,
-0.046151273,
0.004942371,
-0.0195482187,
-0.0124380728,
0.0852954686,
0.0067828684,
0.0372268744,
0.0266096611,
0.0208419301,
-0.0723783076,
-0.0036853175,
-0.0497410074,
0.0139387203,
-0.0047477456,
-0.0724498704,
-0.0467751957,
0.0481765494,
0.0624631643,
0.0034092609,
-0.0069239885,
0.0101841502,
0.0273624808,
-0.0602938496,
0.0266480986,
-0.0174084753,
0.0049884571,
-0.0395965353,
0.0169730894,
-0.0031991322,
-0.0321367271,
-0.0027559518,
0.0254036672,
0.0433530621,
-0.0809958428,
-0.0270517878,
-0.0508701764,
0.0185542833,
0.0337328352,
-0.0214374233,
0.0266575906,
-0.0173751563,
-0.0425632596,
0.0401253887,
0.0631225482,
0.0137856258,
0.1200556606,
-0.0604566596,
0.0091885235,
0.0210442767,
0.0240116008,
0.0222153049,
-0.0270110182,
-0.0047820285,
0.0067693093,
0.0264728237,
-0.0249481499,
0.0219805073,
0.0088393278,
-0.0521776006,
0.0612687059,
0.0096237902,
-0.0171817541,
-0.0590388253,
0.0586155318,
0.0182469394,
0.0073145838,
-0.0602311753,
0.0253514517,
0.0251796208,
0.0253986847,
-0.0139260842,
-0.0403868072,
-0.0784828961,
0.0016660686,
0.0165056959,
-0.0460623652,
-0.0391508453,
-0.0101054758,
-0.0411529578,
-0.0243486501,
-0.0177127104,
0.0417024754,
0.0270393044,
-0.0222559944,
0.0188769158,
-0.0373302661,
0.0079356693,
0.0050228019,
-0.0100554256,
-0.0728600323,
0.0206205174,
-0.0082393689,
-0.0207274146,
0.0016896427,
-0.0530057475,
0.0034057254,
-0.0249285903,
0.020502083,
-0.019410478,
0.0204664376,
0.0560301021,
0.0166749358,
0.0346474461,
-0.0416084714,
-0.0833775103,
0.0110949185,
-0.0240797028,
0.0351949036,
0.0016092246,
-0.0224876646,
-0.0081437295,
-0.040517129,
-0.0162270367,
-0.0318383835,
-0.0104792221,
0.0547301061,
-0.0523307547,
0.0202044472,
0.0213764086,
-0.039550148,
0.0231519267,
0.013236776,
0.0257907752,
0.027788708,
0.0291779954,
0.0019457542,
-0.0128132142,
0.0140847042,
-0.0711560771,
0.080896683,
0.0696483478,
-0.015229024,
0.0164694972,
-0.0342578553,
0.0095097134,
-0.0091411369,
0.0251541827,
0.0062282551,
-0.0092831245,
0.0280548595,
0.0430023074,
0.0613459498,
0.0984268934,
-0.054625079,
-0.0565719008,
-0.0386035927,
-0.0177567694,
-0.0842734203,
0.015924681,
-0.0640889779,
0.0196496882,
0.0330069587,
0.026501473,
-0.0001684419,
-0.0402394645,
-0.0817552805,
0.0228338782,
-0.0054356968,
-0.0075078593,
-0.0111477626,
-0.0055516604,
0.0423121564,
0.0083654961,
-0.0073687816,
-0.0065553985,
-0.0885302424,
-0.0283131748,
-0.0011291459,
-0.0479391366,
0.0183779188,
0.0397137217,
-0.0524098203,
-0.0220499579,
-0.0034480819,
-0.0457874201,
0.0235141311,
0.0851539075,
-0.00846944,
-0.0023280801,
-0.0203449223,
0.0264352672,
-0.0124298735,
0.009646547,
-0.0013749943,
-0.0172470994,
-0.0358516164,
-0.066122286,
0.0163244493,
-0.0476300493,
0.0434146225,
-0.0170515981,
-0.045638103,
-0.038415473,
-0.0144435596,
-0.0164322108,
-0.0371181704,
0.0332335494,
0.0601392053,
-0.0266351383,
-0.0064799651,
-0.0316002071,
0.0115539199,
-0.0491365977,
0.0383776799,
-0.0367568396,
-0.022644043,
0.0508902892,
0.0238248166,
-0.0499221422,
0.0540795065,
0.038815055,
0.0319545344,
-0.0674477816,
0.0721975416,
-0.0345461294,
0.0159914196,
0.004232001,
-0.0171298552,
-0.008719964,
0.0081207165,
-0.0639018938,
0.0274816342,
0.0281678643,
0.0015250659,
0.0195907485,
-0.0029244847,
0.0774400756,
-0.0155141102,
-0.0273812655,
-0.014544908,
0.028830057,
0.0577825233,
0.0254640747,
-0.0146427425,
0.0241016243,
-0.0110955546,
0.0465358831,
0.0251374505,
-0.0786179602,
0.0220326241,
0.0436253063,
0.0020482079,
-0.0221276507,
0.0132955788,
0.0081841433,
-0.0157022998,
0.020043347,
0.0272922404,
-0.060922835,
0.0360082686,
0.0481004789,
-0.0098420447,
0.015244293,
-0.0354962163,
0.0054477514,
0.0588579066,
0.0113768447,
0.0197004788,
-0.05521027,
0.0957220793,
-0.0099578416,
-0.0258775987,
-0.0870030671,
-4.988122116e-33,
-0.0250107981,
0.0084519619,
-0.0159689523,
-0.0055480814,
-0.0277494099,
0.046321664,
0.0118130352,
-0.0005306129,
0.040579129,
0.0195923373,
0.0203086808,
0.0473121665,
0.0016226607,
-0.0638540909,
0.0203532521,
-0.00431633,
0.0424015,
0.0151872067,
-0.022541482,
-0.0170290042,
0.0717906952,
-0.0310596153,
0.039635729,
-0.0539675169,
0.0637451708,
-0.0259560887,
0.0117933834,
0.0207082182,
0.0375194326,
0.0129981274,
0.0115877241,
-0.0071859676,
0.0312611945,
0.0017211648,
0.0028792026,
0.0033680359,
0.0175224859,
-0.0729422197,
0.0433468856,
0.0173460282,
0.0451865755,
-0.0665956885,
-0.0446444973,
0.0147533305,
-0.0306582842,
0.0282169469,
0.0861964077,
-0.0221467428,
-0.0261323284,
-0.0500945412,
-0.0556453206,
0.0618673824,
-0.0278235562,
0.0159073398,
-0.01715317,
0.0118208397,
0.061305888,
0.024402976,
0.0021242124,
0.0482629873,
0.0033719805,
0.0480066724,
-0.0807953328,
-0.0294653252,
-0.0251234658,
0.0139324982,
-0.011355008,
0.0182072166,
0.0395130515,
0.0609023906,
0.0016491322,
0.0230679046,
0.0253216997,
0.0571356006,
0.0234245304,
0.0576854385,
0.009408134,
-0.0234074723,
0.0425485075,
-0.0137861529,
0.0136232805,
-0.0505009517,
0.0046504298,
0.0008926615,
-0.0248820949,
-0.0262053311,
0.0177504569,
0.0131868022,
-0.0315383002,
-0.0204220787,
0.0400835387,
-0.0075767916,
-0.0262015592,
-0.0129682189,
-0.0154329063,
0.0195311643,
-0.0561330356,
0.0540396832,
0.0028634898,
0.0237507951,
0.0120825218,
-0.0059943884,
0.0394526981,
0.0155739794,
0.0035447809,
0.0083688833,
-0.032990247,
-0.0222039241,
0.0015648092,
-0.0203669742,
0.0538016669,
-0.042188853,
-0.0754075199,
-0.0011257899,
0.0085272081,
0.0158165973,
0.0294441413,
0.051770661,
0.0838823542,
0.0304792002,
0.0217269268,
0.0005014031,
0.0085327066,
0.0509657413,
0.0097662238,
0.0032200397,
0.0191834737,
0.0526457988,
0.0487727486,
-0.0365035534,
0.0142978178,
-0.0253604092,
2.481e-7,
0.0512038618,
0.0192010216,
0.0143801235,
-0.0329309702,
0.0362222269,
0.0112234671,
-0.0127168102,
0.0112169068,
-0.0224283896,
-0.0155157074,
-0.0132679036,
-0.0416342951,
0.0044036089,
0.0632678047,
0.0029458308,
-0.0425716229,
-0.0204481836,
-0.047151234,
-0.0793235824,
0.0014947701,
0.0058113099,
0.0607229508,
0.017421471,
0.0042071948,
-0.0294590425,
0.0143320207,
0.0307674091,
-0.0268456452,
-0.0840025991,
0.038863726,
0.066920802,
-0.0230974406,
0.0049828058,
0.0264701266,
0.0049591078,
0.0627367049,
-0.0170876663,
-0.0269701406,
0.0212684963,
0.0769210681,
-0.019402599,
-0.0720594972,
-0.0137437768,
0.0182868075,
0.0277239103,
0.0001455926,
0.0148196258,
-0.0823929757,
-0.0263174754,
0.0485796258,
0.0288350508,
-0.0031621668,
-0.055816751,
0.0463734381,
0.031933751,
0.0188312847,
-0.0175718982,
-0.0900584161,
0.0012744546,
0.0331593417,
-0.0296002906,
-0.0366353653,
0.0279942416,
-0.0384077802,
0.0099797361,
-0.035454642,
0.0191693567,
1.731030915e-34,
0.035190165,
-0.0300019141,
-0.0415188931,
0.0311664585,
0.0029219661,
-0.0188602526,
-0.003658121,
0.0277670603,
0.0149475001,
-0.0184271261,
0.0130836572
] |
http://arxiv.org/abs/2401.13887v1 | http://arxiv.org/pdf/2401.13887v1 | A comparative study of zero-shot inference with large language models and supervised modeling in breast cancer pathology classification | 2024-01-25T02:05:31 | 2024-01-25T02:05:31 | Madhumita Sushil, Travis Zack, Divneet Mandair, Zhiwei Zheng, Ahmed Wali, Yan-Ning Yu, Yuwei Quan, Atul J. Butte | cs.CL | cs.CL, cs.LG | Although supervised machine learning is popular for information extraction
from clinical notes, creating large annotated datasets requires extensive
domain expertise and is time-consuming. Meanwhile, large language models (LLMs)
have demonstrated promising transfer learning capability. In this study, we
explored whether recent LLMs can reduce the need for large-scale data
annotations. We curated a manually-labeled dataset of 769 breast cancer
pathology reports, labeled with 13 categories, to compare zero-shot
classification capability of the GPT-4 model and the GPT-3.5 model with
supervised classification performance of three model architectures: random
forests classifier, long short-term memory networks with attention (LSTM-Att),
and the UCSF-BERT model. Across all 13 tasks, the GPT-4 model performed either
significantly better than or as well as the best supervised model, the LSTM-Att
model (average macro F1 score of 0.83 vs. 0.75). On tasks with high imbalance
between labels, the differences were more prominent. Frequent sources of GPT-4
errors included inferences from multiple samples and complex task design. On
complex tasks where large annotated datasets cannot be easily collected, LLMs
can reduce the burden of large-scale data labeling. However, if the use of LLMs
is prohibitive, the use of simpler supervised models with large annotated
datasets can provide comparable results. LLMs demonstrated the potential to
speed up the execution of clinical NLP studies by reducing the need for
curating large annotated datasets. This may result in an increase in the
utilization of NLP-based variables and outcomes in observational clinical
studies. | A comparative study of zero-shot inference with large language models and supervised modeling in breast cancer pathology classification: Although supervised machine learning is popular for information extraction
from clinical notes, creating large annotated datasets requires extensive
domain expertise and is time-consuming. Meanwhile, large language models (LLMs)
have demonstrated promising transfer learning capability. In this study, we
explored whether recent LLMs can reduce the need for large-scale data
annotations. We curated a manually-labeled dataset of 769 breast cancer
pathology reports, labeled with 13 categories, to compare zero-shot
classification capability of the GPT-4 model and the GPT-3.5 model with
supervised classification performance of three model architectures: random
forests classifier, long short-term memory networks with attention (LSTM-Att),
and the UCSF-BERT model. Across all 13 tasks, the GPT-4 model performed either
significantly better than or as well as the best supervised model, the LSTM-Att
model (average macro F1 score of 0.83 vs. 0.75). On tasks with high imbalance
between labels, the differences were more prominent. Frequent sources of GPT-4
errors included inferences from multiple samples and complex task design. On
complex tasks where large annotated datasets cannot be easily collected, LLMs
can reduce the burden of large-scale data labeling. However, if the use of LLMs
is prohibitive, the use of simpler supervised models with large annotated
datasets can provide comparable results. LLMs demonstrated the potential to
speed up the execution of clinical NLP studies by reducing the need for
curating large annotated datasets. This may result in an increase in the
utilization of NLP-based variables and outcomes in observational clinical
studies. | [
0.0249562468,
0.0514693148,
0.0076765437,
-0.026592046,
-0.0283281468,
0.0523088686,
0.0379602686,
0.0046628155,
-0.0177366603,
-0.0556635633,
0.0070457398,
-0.0444909818,
-0.0048240763,
-0.0186800081,
-0.0060498128,
-0.0356818065,
0.0094721736,
0.0061447006,
0.0315588042,
0.0213389117,
-0.0154478438,
0.0080826823,
0.0258639697,
0.0335041694,
-0.0146793397,
0.0239788294,
0.0531749725,
-0.0094650658,
-0.0214926656,
-0.0460223816,
-0.0047105807,
0.0269146692,
0.0019580547,
0.0401076972,
0.0000020806,
-0.090301007,
-0.0192564987,
0.0244281683,
-0.04792732,
-0.0122248521,
0.0263939761,
-0.0434823856,
0.0261059366,
-0.0009013522,
0.02976235,
-0.0653976128,
0.0512249619,
0.0584402569,
-0.0134059386,
0.0694171265,
-0.0173362698,
-0.0299424455,
0.0291052535,
-0.0312173851,
0.0264920536,
-0.0367206745,
0.0515819043,
-0.0301804505,
-0.017158594,
0.0061586513,
0.0046183909,
0.0602626652,
0.0055966638,
0.018265089,
0.0259441622,
0.0808005333,
-0.0531060733,
-0.016823316,
0.0186231732,
0.0440812819,
-0.0290178657,
0.0006602295,
0.0035955668,
-0.019336665,
0.0093708895,
0.043298196,
-0.0484885573,
0.0542539246,
-0.0288893506,
0.003231907,
0.0002101928,
-0.0163210072,
0.0362240411,
-0.003809442,
-0.0078997146,
0.0058747507,
0.0106626162,
0.0040622535,
0.0445195287,
0.0277880169,
-0.0261830464,
-0.029902203,
0.0070034373,
0.0250044428,
0.0124465339,
0.0134338299,
-0.0291549116,
-0.0272363797,
0.0584655106,
-0.0593236051,
0.022822937,
-0.0053734281,
-0.0599162728,
0.0506775342,
-0.0349610262,
0.0021398952,
-0.007464373,
0.0697980896,
-0.0318616889,
-0.0166447964,
-0.0598283596,
-0.0098316697,
0.0185995307,
0.0440102406,
0.0214278903,
0.001551758,
-0.0735524148,
-0.0254960395,
0.0039362996,
-0.0279516224,
0.0263694637,
0.0200422592,
-0.0197317749,
0.0200746246,
-0.0055517992,
0.0301575698,
0.0131936353,
0.028325187,
-0.0254428424,
-0.0533107184,
-0.0169414356,
0.0435346551,
0.0266989414,
-0.0330445841,
0.0155110182,
0.0739261657,
0.0226354413,
-0.049375996,
-0.0377893932,
0.0047041797,
-0.0571832024,
-0.0598417073,
0.015017814,
-0.0650880635,
-0.0070756106,
0.0314088985,
0.0125761013,
-0.0226342715,
-0.0011040613,
-0.0159796309,
-0.0694432035,
0.0432121791,
-0.0101370867,
-0.0010510815,
0.0535177998,
-0.0053848494,
-0.0032279405,
0.0440238044,
0.0245908871,
-0.0120144682,
-0.0272880681,
-0.0343013369,
0.0118413335,
0.0129561415,
0.0302659143,
-0.01388417,
-0.0224240534,
0.0029478236,
-0.0047647925,
0.0427022688,
0.0147991376,
0.0112887416,
-0.0013984127,
-0.0262575336,
0.0421790145,
0.0691084787,
0.0271714497,
0.0435046963,
0.0111632869,
-0.0323215574,
-0.0250878129,
-0.0550653711,
-0.0294007957,
0.0187315717,
-0.0423115157,
-0.0062922644,
0.0282001123,
-0.0436585248,
-0.0294315685,
-0.0334930383,
-0.059113726,
0.0040064487,
-0.0294619668,
0.0046291891,
0.0696939453,
-0.0215118229,
-0.0048329998,
0.0325427726,
0.0503962152,
0.0470839106,
-0.0607130937,
-0.0535224676,
-0.0093682865,
0.0316875242,
-0.020487776,
-0.0563770831,
-0.0351814367,
-0.026616862,
0.0356810614,
0.0126117384,
0.0039694351,
0.0358865447,
0.0273419414,
0.0142455082,
0.0183615685,
0.0023057801,
0.0399677306,
-0.0052312948,
-0.0080742389,
0.0581376962,
-0.020370217,
-0.0439052135,
-0.0407418013,
0.0605650023,
0.0024495244,
-0.0820952877,
0.003682693,
-0.0026369737,
0.0464917459,
-0.0184567496,
-0.040275,
0.0138376532,
-0.01300178,
-0.0362339169,
-0.0448546447,
-0.0346884951,
-0.0048274789,
-0.039591141,
0.0184021536,
-0.0064298925,
-0.0319619589,
-0.0046328111,
-0.0774086267,
-0.0012881589,
-0.0108624212,
-0.0366431512,
0.0075196512,
0.0162789151,
0.0073231189,
0.0134792225,
-0.0003677838,
-0.0366281569,
0.0257773362,
0.0058926768,
-0.0172394458,
0.0164719187,
-0.0740644112,
0.0048634377,
-0.0158692338,
-0.0220901016,
0.025263479,
0.0028319778,
-0.0154486904,
-0.0120001826,
-0.0000261267,
-0.0193528309,
-0.0008926392,
-0.0989332348,
0.019542383,
0.0159302298,
0.014888756,
-0.0103709716,
0.0124953305,
-0.0076351678,
-0.0278860815,
-0.0457387604,
-0.0232911427,
0.0144371726,
0.0605410337,
0.0488606803,
0.0196939427,
0.0029227091,
-0.0449067354,
-0.0150101231,
-0.0371198468,
0.061745543,
0.0481563434,
-0.0784166306,
0.0166681185,
0.0395609587,
-0.0153629566,
-0.0161882266,
0.0420574583,
0.0047713341,
-0.0273052398,
-0.0335007869,
-0.0470204987,
0.0532349646,
-0.0142811192,
0.0247126836,
0.041330602,
-0.0036539685,
-0.0347975604,
-0.0025236397,
-0.0291636381,
-0.0643489137,
0.0180929694,
-0.0289889742,
0.0063915821,
0.0580236427,
0.0034330506,
-0.0032919133,
-0.0187687762,
-0.0064151301,
0.000150663,
0.0394028462,
-0.0504525751,
0.0406282172,
-0.0151303131,
-0.001416233,
-0.0161945913,
-0.0417944789,
0.0915682912,
0.0206521358,
0.0263433792,
-0.0548551232,
0.0151963122,
-0.0290574804,
0.0243622214,
0.0223911759,
-0.0014963025,
0.0145825744,
0.0247736983,
0.0203855224,
-0.0171677489,
-0.0010338713,
-0.0236732941,
0.0399529114,
-0.0217045452,
0.0649637729,
-0.0146757318,
0.0074416446,
-0.0079954229,
-0.0328773186,
-0.0247920621,
0.0194958523,
-0.0212479737,
0.0215982422,
-0.084028855,
0.0261326134,
0.0113801789,
0.0798711032,
-0.0222243611,
-0.0304808933,
0.0240244996,
0.0207295083,
0.0263526309,
-0.0231294576,
-0.0340914875,
-0.0217007566,
0.0269509833,
0.0344801433,
-0.0211422481,
0.0688106194,
-0.00990777,
0.0073282234,
-0.0053630867,
-0.0246967711,
-0.0042579286,
0.0274670701,
0.0050421557,
-0.087535575,
-0.00830644,
0.1168929562,
-0.0009590012,
-0.0676750839,
-0.0225898568,
0.0045188637,
-0.0001627272,
-0.0306387208,
0.0285063721,
-0.0540418066,
-0.0601216108,
-0.015420517,
0.0110615054,
-0.011626686,
-0.0578377284,
0.0241360832,
-0.066982761,
0.0336395837,
-0.004133008,
-0.0990397558,
-0.0007526775,
-0.0650587529,
0.0088748187,
-0.0116805602,
-0.0284001455,
-0.0414113514,
0.0029153384,
-0.0279994942,
0.0327063501,
-0.0714455843,
-0.0739388466,
0.0148256859,
-0.0355892405,
0.0459514968,
0.0265267957,
0.0285126213,
-0.0530935042,
0.0271397084,
0.0304136667,
0.0467851199,
-0.016078379,
0.0130140604,
0.105409421,
-0.0095004039,
0.0215572603,
-0.0178553518,
0.0197163746,
-0.0078008054,
0.0103163356,
0.0074842395,
0.0876180977,
-0.0038516256,
-0.036701899,
-0.0235926546,
-0.0228942148,
0.0115187,
0.0490079336,
0.0101705408,
-0.0211232007,
-0.0578757562,
-0.0028697322,
0.00052895,
0.0031136898,
-0.0015007451,
0.075373657,
-0.0495594181,
-0.0786372572,
-0.0297306329,
-0.033027757,
0.0711092427,
-0.0551300384,
0.0278283563,
0.068804644,
-0.0043090642,
-0.0092344293,
-0.0531856902,
0.0824707076,
0.0025846364,
0.0056237625,
-0.0405796729,
-0.0386411101,
0.0524479225,
-0.0214379001,
0.0113169365,
0.0255038142,
-0.042284742,
-0.0055060377,
-0.0340553038,
0.0382738933,
0.0141934939,
-0.0076644537,
0.0006482264,
-0.0718299598,
0.0663158745,
0.0511709638,
-0.0243957788,
-0.0208791737,
-0.0117772929,
-0.0030693326,
0.0186788701,
0.0067405873,
-0.027840551,
0.0417046435,
0.0518358573,
-0.0185063742,
0.0300289765,
-0.0079934914,
0.0233391766,
0.0421599299,
0.0817991942,
0.0135027086,
-0.0260764956,
0.0040319632,
0.0010232474,
0.0085578095,
0.0028739176,
-0.0088619301,
-0.0274286456,
0.0419391058,
0.0274211317,
-0.0395117253,
-0.0011833088,
-0.0277742483,
-0.0064781066,
-0.0550794341,
0.0239080787,
0.030861957,
0.0043682829,
0.0087913042,
0.0014456274,
-0.0454404019,
-0.0535388738,
-0.0029372741,
0.0032097111,
0.0133914705,
0.0066347858,
0.0236002635,
0.0619540587,
-0.0259531327,
0.0062172958,
-0.0616708882,
-0.0358970277,
-0.0171974525,
0.0158279799,
0.0293984432,
-0.0618312024,
0.0453712791,
-0.0290170331,
-0.0087933075,
-0.0385550112,
0.0180940349,
-0.0359600075,
0.0281534456,
0.0346436724,
0.0289563648,
-0.0074525787,
-0.0180628132,
0.0780574232,
0.0844755098,
0.0382972546,
-0.0050724777,
-0.0694236755,
-0.0188998114,
0.0110507729,
0.045751784,
-0.0075344122,
-0.0096713873,
0.0200432129,
-0.0534658842,
-0.0230213627,
0.0568083413,
-0.0676152334,
-0.0041202479,
-0.0158221647,
0.0313184708,
0.0265644025,
0.0061739711,
-0.0475275815,
-0.040765442,
0.0253900606,
0.025309179,
-0.0289064981,
-0.0132295983,
-5.993315988e-33,
0.0313823521,
-0.0299744904,
0.0056409882,
-0.0223055147,
-0.0214290824,
0.0320588164,
-0.0593517721,
0.004824223,
0.0290297717,
0.0504340269,
-0.0140223037,
0.0097067496,
0.0052726786,
0.054029841,
0.0560243428,
-0.0403236263,
0.028033061,
0.0037188386,
0.0003762115,
0.012131989,
0.0107020717,
0.0357033163,
0.0361599922,
-0.0472672693,
0.0095688347,
0.0452529863,
0.0425037965,
-0.0231426638,
-0.0072951312,
0.0474595949,
-0.0295930635,
0.0740735978,
0.0453547165,
-0.0488526262,
0.012144519,
0.0161280986,
-0.0300795455,
-0.0287747104,
0.0054319287,
-0.0021658475,
-0.0113801481,
-0.0098197004,
0.0737312883,
0.0173619483,
0.0261002369,
-0.0261363443,
0.035567414,
-0.0397076905,
-0.0476835892,
-0.0668360591,
-0.0521920733,
0.051059287,
-0.0036680358,
0.0147336544,
0.0267363898,
-0.0119296918,
-0.0660705417,
0.0579324253,
-0.0355634652,
-0.0250487812,
0.0254145842,
0.0003273796,
0.0007846428,
0.0273832604,
-0.0287441071,
0.0550721511,
0.0078934496,
-0.0110551817,
-0.0532724895,
0.0039305626,
-0.0335788578,
0.0212269239,
0.0282998886,
0.0160270259,
0.0599735528,
0.0150045948,
-0.103965424,
-0.0048328927,
-0.0710758045,
0.0115549844,
0.0579362214,
-0.0246262681,
-0.0350620709,
-0.0184012298,
-0.0104547739,
0.0049289255,
-0.0204462167,
0.006984523,
0.0159417503,
-0.0285902098,
-0.0124562168,
0.0241783652,
0.0180871878,
0.0032401441,
-0.0424203314,
-0.0308849309,
0.0425031446,
0.0085023958,
-0.0278839413,
0.0008938799,
-0.056232363,
-0.0107809883,
0.005827629,
0.0058682626,
0.0442463309,
-0.0009998973,
0.0710513219,
-0.0164155085,
-0.0528084263,
-0.0594126545,
-0.0208022259,
0.0019853474,
-0.0237099566,
0.0670406371,
0.0227709487,
0.030374676,
0.0086021647,
0.0889280438,
0.0154219568,
-0.0163566377,
0.0088429693,
-0.0273165889,
-0.006843131,
-0.0397292413,
-0.0375303924,
-0.0037338096,
0.0004272515,
0.0058865468,
0.0194048416,
-0.0780645162,
-0.004618342,
0.0143400067,
2.867e-7,
0.1126882061,
0.0583949797,
-0.007459111,
-0.0575393736,
0.0426758081,
-0.0030503287,
0.0088059986,
0.0068459501,
0.0288091302,
-0.0079989824,
0.0364006795,
0.0272573698,
0.0397550128,
0.0153577672,
-0.0494437218,
-0.0360728763,
-0.061890766,
-0.026449101,
-0.0247542094,
-0.0004244303,
0.0665684119,
0.0785915107,
0.0072574043,
-0.0450181775,
0.0345997624,
0.0065913792,
0.0279907882,
-0.0599411279,
0.0481246337,
-0.0047744066,
0.0025421281,
0.0178842545,
0.0178594626,
0.0795308575,
0.034234263,
-0.0335812382,
0.0020153357,
0.0698901042,
-0.0642677769,
0.0159232896,
0.0203161594,
0.0408173613,
0.0293297116,
-0.0378076881,
0.062006209,
-0.0970463455,
-0.0611485392,
0.0129084038,
0.0217002537,
0.0004556628,
0.0887875855,
-0.0154038006,
-0.0233888775,
0.0214548931,
0.020077223,
-0.0022112653,
-0.0432088301,
-0.0558630563,
0.0013727162,
0.0301092118,
-0.0613447987,
-0.0230175853,
0.014840004,
-0.0116821658,
0.0456319451,
-0.0847676471,
0.0161301158,
2.495633511e-34,
0.01365068,
0.0079157213,
0.0106632831,
0.0281564835,
-0.0181994978,
0.0220993664,
0.0145921698,
0.0019464452,
0.0211304612,
-0.0840281546,
0.0017921509
] |
http://arxiv.org/abs/2402.01705v2 | http://arxiv.org/pdf/2402.01705v2 | Beyond Behaviorist Representational Harms: A Plan for Measurement and Mitigation | 2024-01-25T00:54:10 | 2024-05-06T21:00:00 | Jennifer Chien, David Danks | cs.CY | cs.CY, cs.AI, cs.CL, cs.LG | Algorithmic harms are commonly categorized as either allocative or
representational. This study specifically addresses the latter, focusing on an
examination of current definitions of representational harms to discern what is
included and what is not. This analysis motivates our expansion beyond
behavioral definitions to encompass harms to cognitive and affective states.
The paper outlines high-level requirements for measurement: identifying the
necessary expertise to implement this approach and illustrating it through a
case study. Our work highlights the unique vulnerabilities of large language
models to perpetrating representational harms, particularly when these harms go
unmeasured and unmitigated. The work concludes by presenting proposed
mitigations and delineating when to employ them. The overarching aim of this
research is to establish a framework for broadening the definition of
representational harms and to translate insights from fairness research into
practical measurement and mitigation praxis. | Beyond Behaviorist Representational Harms: A Plan for Measurement and Mitigation: Algorithmic harms are commonly categorized as either allocative or
representational. This study specifically addresses the latter, focusing on an
examination of current definitions of representational harms to discern what is
included and what is not. This analysis motivates our expansion beyond
behavioral definitions to encompass harms to cognitive and affective states.
The paper outlines high-level requirements for measurement: identifying the
necessary expertise to implement this approach and illustrating it through a
case study. Our work highlights the unique vulnerabilities of large language
models to perpetrating representational harms, particularly when these harms go
unmeasured and unmitigated. The work concludes by presenting proposed
mitigations and delineating when to employ them. The overarching aim of this
research is to establish a framework for broadening the definition of
representational harms and to translate insights from fairness research into
practical measurement and mitigation praxis. | [
0.0526121818,
0.0794655681,
-0.0658248067,
0.0357895158,
-0.0132500725,
0.0118115284,
0.0133272773,
0.0073719951,
-0.0094755897,
-0.0720365047,
0.0063730446,
-0.0100724753,
0.0399787016,
-0.0247363094,
0.0449633859,
-0.0575197972,
0.0233672727,
0.0265419781,
0.0400293916,
-0.0037807946,
0.0597800203,
0.0268194489,
0.0422811247,
0.0854185224,
-0.0205349755,
-0.0350075439,
0.0537459739,
-0.003956263,
-0.0202833358,
-0.0349778645,
-0.0274613407,
0.0376452878,
-0.0090507176,
0.0268132109,
0.0000019751,
-0.0711203218,
0.0177159589,
0.0196404904,
0.0115718832,
0.0291877445,
0.0472091176,
-0.0409588665,
0.0194964316,
0.0372339226,
0.0175627023,
0.0428056568,
0.0568300821,
-0.0064676018,
0.0264373012,
0.0165530555,
-0.0115559725,
0.013855109,
0.0127914064,
0.0292870607,
-0.0011837997,
0.0399751626,
0.0443499945,
-0.0054942607,
-0.0083719203,
-0.0623074211,
-0.0321388766,
0.0298399143,
0.0232524946,
0.0195823032,
-0.0046255798,
-0.0448682792,
-0.0426571742,
-0.0320951492,
0.0232768096,
0.0549025424,
0.0496269539,
-0.0004451548,
0.0069673769,
-0.0167548917,
0.0300546102,
-0.0038700246,
-0.0180106703,
0.0656541213,
0.0043783202,
-0.0193183478,
-0.0261974838,
0.0308345221,
0.0456712544,
0.0359470882,
0.0449903458,
0.0322586447,
-0.0099638598,
-0.0572669432,
-0.0143047534,
0.0014735772,
0.0106359785,
-0.0351245143,
0.0075193429,
-0.0026984336,
-0.0221508108,
0.0277197566,
-0.0197878387,
-0.0551080145,
0.046054896,
-0.0755931363,
0.029939441,
0.0479976945,
0.0027208664,
-0.0417829975,
0.0379353464,
0.0204377398,
-0.0146491267,
0.0341778323,
-0.0183633436,
0.0539061315,
0.0117630241,
-0.051553566,
0.0017716895,
0.0005132614,
-0.0097794337,
0.0271463469,
-0.0226905663,
0.0052122804,
-0.1035817862,
-0.0129458746,
-0.0555594228,
0.0197537374,
-0.0109314602,
0.0240155589,
-0.0085496083,
-0.0364037938,
-0.0504396707,
0.0334853157,
-0.0224644039,
-0.0482625067,
0.0214209184,
0.0050606858,
-0.0667493492,
-0.0148779675,
-0.0107703255,
0.0873963311,
0.0589207932,
-0.0237595942,
-0.011428385,
-0.0827960223,
-0.0499026291,
-0.0181299038,
-0.0322484113,
-0.0398241207,
-0.0343735665,
0.0176626146,
-0.00730502,
-0.0712020248,
-0.0174608007,
-0.0094808573,
-0.0617007576,
0.0508818887,
0.0070194681,
0.0001270692,
0.0065315505,
0.05185562,
-0.001068306,
0.0583803542,
0.0368676782,
0.0024963506,
0.0366330519,
-0.0074357716,
0.0010175678,
0.0018172063,
0.0019780362,
0.0053160465,
-0.031907957,
0.0142634558,
-0.0298508909,
0.0230673086,
0.0100496002,
0.0623018704,
-0.0172708388,
0.0026827704,
0.0705535188,
0.0472995006,
-0.0317535065,
0.0312090553,
-0.0259514935,
0.0075104581,
-0.0234665591,
-0.0049328348,
-0.0612149313,
-0.025794344,
-0.0060076285,
-0.0289461277,
0.0074306363,
-0.0067937435,
-0.0220196117,
-0.0283743236,
-0.014983709,
0.015720835,
0.0727733672,
0.0216477923,
-0.0222755279,
0.0099221244,
0.0186319388,
0.0282086954,
-0.001865994,
-0.0405306332,
0.0295949336,
-0.0560985543,
-0.0164376348,
0.0470431261,
0.0541717038,
-0.0328716822,
0.0778160915,
-0.0449062847,
-0.0209995732,
0.0222815108,
0.020955313,
0.0050877067,
-0.014529462,
0.0099322433,
-0.0189124588,
-0.021738546,
0.0060378164,
-0.0309599899,
-0.0280358549,
0.0106592849,
-0.0108225038,
-0.0277263243,
-0.0619684272,
0.0304044485,
0.0474300422,
-0.0437830202,
-0.0630147308,
0.0150894076,
-0.0131736444,
0.0155656161,
-0.0372991078,
0.008442061,
-0.0289747957,
0.032698676,
0.010000743,
0.0040381206,
-0.0372164212,
0.0693443418,
0.0332440473,
-0.0348196514,
-0.0110216923,
0.0443713441,
0.0286055207,
0.0257658251,
0.0607174411,
-0.0496453904,
0.0492270999,
0.0068789031,
0.011767772,
-0.0637116134,
0.0163145326,
-0.0326938741,
-0.0126517043,
-0.0185473431,
0.0036841538,
0.0112370336,
-0.0385209583,
0.0083432179,
-0.0305946022,
-0.0933486372,
0.0413748547,
-0.028674243,
0.0451815054,
0.0334466062,
-0.0509243608,
-0.0459118523,
0.026870938,
-0.0700976551,
-0.047626771,
-0.0080810646,
-0.0213391092,
-0.0030508677,
0.0212464631,
-0.0228159074,
-0.0060574729,
0.0194150452,
-0.0699905455,
0.0013969132,
0.0667922199,
-0.0052683717,
-0.0094274757,
-0.0554242991,
-0.0161200967,
-0.0282094404,
-0.0540797673,
0.0423136875,
-0.0034681952,
-0.0153035372,
-0.0496608913,
0.0060384488,
-0.0349412337,
-0.0067904764,
0.0106430287,
-0.0118631208,
0.0500839725,
0.0831075013,
0.0030845103,
0.0336806662,
0.0343769565,
0.0731662065,
0.0792373642,
-0.0200058762,
-0.0406380147,
-0.0281705037,
0.0302307624,
0.0293458719,
0.0059309108,
-0.0703116059,
-0.0138846971,
0.0340876281,
0.0506775752,
-0.0063189808,
-0.0348700248,
0.054686334,
-0.094140783,
-0.0829292983,
-0.050731726,
-0.1125357598,
-0.0470364019,
0.00486936,
-0.0321932212,
-0.0750902966,
0.0466819704,
0.0032225871,
0.0287592858,
-0.0065127886,
0.0063098571,
0.0242986828,
0.0497760437,
-0.0187333096,
0.0425951332,
0.0337503292,
-0.051324863,
-0.037390396,
-0.0588769093,
0.0431342833,
-0.0359419398,
0.049463775,
-0.0297322292,
0.0089651933,
0.0040138969,
-0.043218594,
-0.0228474922,
-0.0406126678,
0.0061462927,
0.0229265243,
0.0264776275,
0.0076094903,
-0.0046055904,
-0.0020523323,
0.0415332802,
-0.0268409997,
0.0026068583,
-0.0313325711,
0.0209871475,
-0.0727794543,
0.0180539843,
0.007789786,
0.0733561218,
0.0187939238,
0.0191834588,
0.0464473739,
0.0138461431,
0.0225649215,
0.0169716552,
0.0319960266,
0.0102940174,
-0.0335464962,
-0.0203703009,
0.0418565869,
0.0704768822,
-0.0193551108,
0.0101430016,
0.0498807393,
0.0134974271,
-0.050974898,
0.0632158965,
-0.0292516593,
-0.0108525259,
0.0161077678,
-0.0308553558,
-0.081001617,
0.0119141638,
0.0314239301,
0.0411654375,
-0.0887412801,
-0.0414214954,
0.0160201211,
-0.0389499702,
0.0100894375,
0.0142498184,
-0.0137984632,
-0.0335885622,
0.0020770512,
-0.001118989,
0.0259507336,
0.0575189404,
-0.0053033293,
-0.0129537228,
0.0006380808,
0.0308563635,
-0.0658736676,
0.0201390367,
-0.0067433156,
-0.0268846657,
0.0257522985,
0.047523737,
0.0180157404,
-0.0374868363,
-0.0195224062,
-0.0302550439,
0.0160454866,
-0.0210619662,
0.0139703834,
-0.0014450738,
0.0006348941,
0.0028451944,
-0.0460675433,
-0.0085694799,
-0.0080148941,
-0.0430080257,
0.0028950982,
0.0491586365,
-0.0172490496,
-0.008755195,
-0.0572856441,
-0.0136947539,
0.016403079,
0.0605322532,
0.0487031788,
-0.0189118516,
-0.0028593906,
-0.0676735789,
0.0108341174,
-0.0016257104,
-0.0815104246,
-0.023299776,
0.0507185087,
-0.0023039649,
-0.0275644567,
-0.0944106951,
-0.0379720889,
0.0102219805,
0.0492405258,
-0.0367051065,
-0.0191656947,
-0.01034611,
-0.0462261178,
0.021786496,
-0.0096837301,
0.0571090467,
0.0225294419,
0.0167799406,
-0.0493364446,
0.0204291716,
0.0078419121,
-0.00049166,
0.0015890464,
-0.0515292734,
0.0514640734,
0.0371308252,
0.0287666563,
0.0032501991,
0.0201072097,
-0.0533415079,
0.0343849547,
0.0180718284,
0.0085808514,
0.0415723287,
0.0440831371,
-0.0578525104,
0.0239391159,
0.0059444578,
-0.0155244349,
-0.0130008021,
0.0394907519,
0.0283613931,
-0.017399136,
0.0167963617,
0.0372999348,
0.0406798273,
-0.0028002076,
-0.0512686558,
-0.004176707,
-0.0021829284,
-0.0719895214,
-0.022189403,
-0.052053947,
-0.0044102604,
-0.0172927678,
0.0062424368,
0.0530057438,
-0.0429216512,
0.0205902886,
-0.0391311198,
-0.0244505983,
0.004675915,
0.0759967119,
0.0075462633,
0.0423395261,
0.0019941365,
0.0379577726,
-0.0202155616,
0.0317359157,
-0.0323045366,
0.0333680809,
0.0256082974,
0.0293670725,
0.017075276,
0.0361723714,
-0.031572599,
0.0222620722,
-0.0022940619,
-0.0406361483,
0.0201610364,
-0.0283237081,
-0.021829851,
-0.0347102806,
0.0074496828,
0.0270699039,
0.0039551714,
-0.0071740393,
-0.0276278276,
-0.0168178193,
-0.0370395258,
0.0229930822,
0.0074748578,
-0.035457734,
-0.0057711168,
0.0245978646,
-0.0571831539,
0.0784864277,
-0.035101749,
-0.0343254656,
0.0346302614,
-0.0552482903,
0.031132726,
0.0411789902,
-0.014999738,
-0.0487535857,
-0.1038269028,
-0.0780194625,
-0.0329923034,
-0.0288806614,
0.0458545573,
-0.0639242604,
0.0008303835,
0.0451570265,
0.0293425974,
-0.0154536022,
-0.0094652781,
0.0129956659,
-0.0483455807,
-0.0183533542,
-0.00862516,
-5.29970526e-33,
0.0178009197,
-0.0343779661,
-0.0419613495,
-0.0268202629,
-0.0424701311,
0.01358735,
-0.0025621203,
0.0413898565,
-0.0178862419,
0.0123060094,
0.0119768754,
0.0122700585,
-0.0015836171,
-0.0149395941,
0.0238165464,
-0.0193965603,
0.0103646098,
0.0079985913,
-0.0586026274,
-0.006779606,
0.0077721784,
-0.0261999257,
0.0225230902,
-0.067791827,
0.1119492203,
0.0219562035,
-0.0275652613,
-0.0397523753,
-0.0088230819,
0.0360113531,
-0.0003486265,
0.1408149898,
0.0351693295,
-0.0218346622,
0.0002484285,
0.0757381544,
0.0240124278,
-0.0189325586,
0.0718981698,
0.0219307188,
0.0176698398,
-0.0161072891,
0.0376658067,
0.015874438,
-0.0645475164,
0.0074127861,
0.0269050729,
-0.0132994503,
0.0152277881,
0.0279656947,
-0.0539322682,
-0.0078165298,
-0.050673712,
0.0272792522,
0.0142263062,
0.0135455243,
0.0074867462,
0.050286036,
-0.0696554706,
0.004153321,
-0.0213740449,
0.0020962681,
0.0491104424,
0.0089828055,
0.0247997027,
0.0805181041,
0.0391010791,
0.0421256274,
-0.0133308182,
0.0021446717,
0.0482254624,
-0.0223043598,
0.0282645263,
0.0564025044,
0.0225050244,
-0.0725340918,
0.0055434788,
-0.0116331019,
0.0072791073,
0.0019759634,
0.0268864557,
-0.000823123,
-0.0790379643,
-0.0698187798,
-0.0123662381,
-0.0498625375,
-0.026770616,
-0.0295026936,
0.0604546741,
-0.0443506278,
-0.0021733437,
0.030684948,
-0.0034962129,
-0.0590896048,
-0.0051854914,
-0.0001972949,
0.0062082913,
0.003962649,
0.0300008748,
-0.0067039318,
-0.035882391,
-0.0079160761,
-0.008952463,
0.0133921374,
0.0539761707,
-0.0022071649,
0.018924281,
0.0105897523,
-0.0117903482,
-0.0074697197,
0.0046845437,
0.0297891572,
0.0240253843,
0.0046108039,
0.0185231064,
0.034714669,
0.0062596397,
-0.0344225913,
0.0545919947,
-0.0024215709,
-0.056260433,
-0.0055156644,
-0.0551796779,
0.0220247712,
-0.0331906043,
-0.0410469063,
-0.0223050341,
0.0359266326,
0.0801593736,
-0.0622207783,
-0.0293132905,
-0.0721609965,
2.592e-7,
0.0253164526,
0.0039531197,
0.003658534,
0.0059605427,
0.0024799164,
0.0490272678,
0.0321915373,
0.0679990277,
0.0562599264,
-0.0103953965,
-0.0108303381,
0.0096830986,
0.0054443963,
0.0250690319,
-0.0206230599,
0.0109104933,
-0.0740320012,
0.0592457093,
-0.0185578186,
0.0489567891,
0.0275982171,
0.0388735421,
0.0079573989,
-0.002926433,
-0.0361662656,
-0.0281683952,
0.0023711151,
-0.0385792702,
0.03725712,
-0.0101348888,
0.0025089907,
0.0031523637,
0.0370358415,
0.0198753774,
0.027386304,
0.0187713541,
0.0158205163,
-0.0028717567,
-0.0252396744,
0.0028844974,
-0.0353325121,
-0.0259960126,
0.004628296,
-0.0331164338,
0.0154927345,
0.0157589931,
-0.0309723541,
-0.0097667864,
-0.0191961434,
-0.0316051356,
0.0684793517,
-0.0012111727,
-0.0482106172,
0.0092940526,
0.0230201781,
0.0126946121,
0.0375337675,
-0.0680021942,
-0.0107004205,
-0.0328975879,
0.0167990942,
-0.0528120659,
-0.010009137,
0.0201758333,
-0.0327263847,
-0.0419712737,
0.0309621189,
2.292393293e-34,
-0.0049044066,
0.0017325381,
0.0460251383,
-0.0189547203,
0.0077753752,
0.0350104943,
0.0066225869,
-0.0121542998,
0.0087630451,
-0.0073154173,
-0.0275641307
] |
http://arxiv.org/abs/2401.13867v1 | http://arxiv.org/pdf/2401.13867v1 | Unmasking and Quantifying Racial Bias of Large Language Models in Medical Report Generation | 2024-01-25T00:34:16 | 2024-01-25T00:34:16 | Yifan Yang, Xiaoyu Liu, Qiao Jin, Furong Huang, Zhiyong Lu | cs.CL | cs.CL | Large language models like GPT-3.5-turbo and GPT-4 hold promise for
healthcare professionals, but they may inadvertently inherit biases during
their training, potentially affecting their utility in medical applications.
Despite few attempts in the past, the precise impact and extent of these biases
remain uncertain. Through both qualitative and quantitative analyses, we find
that these models tend to project higher costs and longer hospitalizations for
White populations and exhibit optimistic views in challenging medical scenarios
with much higher survival rates. These biases, which mirror real-world
healthcare disparities, are evident in the generation of patient backgrounds,
the association of specific diseases with certain races, and disparities in
treatment recommendations, etc. Our findings underscore the critical need for
future research to address and mitigate biases in language models, especially
in critical healthcare applications, to ensure fair and accurate outcomes for
all patients. | Unmasking and Quantifying Racial Bias of Large Language Models in Medical Report Generation: Large language models like GPT-3.5-turbo and GPT-4 hold promise for
healthcare professionals, but they may inadvertently inherit biases during
their training, potentially affecting their utility in medical applications.
Despite few attempts in the past, the precise impact and extent of these biases
remain uncertain. Through both qualitative and quantitative analyses, we find
that these models tend to project higher costs and longer hospitalizations for
White populations and exhibit optimistic views in challenging medical scenarios
with much higher survival rates. These biases, which mirror real-world
healthcare disparities, are evident in the generation of patient backgrounds,
the association of specific diseases with certain races, and disparities in
treatment recommendations, etc. Our findings underscore the critical need for
future research to address and mitigate biases in language models, especially
in critical healthcare applications, to ensure fair and accurate outcomes for
all patients. | [
0.0343783908,
0.10362003,
-0.0143168122,
-0.0279852562,
0.0056580398,
0.0350156091,
0.0280323327,
0.0123988306,
-0.0133954212,
-0.0434308797,
0.031544093,
-0.0565995649,
0.0418606885,
0.0070780865,
0.0051408038,
-0.0541526042,
0.0288409889,
0.0656587258,
0.0302835442,
-0.0206302814,
-0.0068076882,
0.0469381362,
0.0063853227,
0.0717851445,
-0.0088673625,
-0.0046232808,
0.0670375824,
-0.0219004583,
-0.0124780629,
-0.0391379371,
-0.0013546156,
0.0576560423,
-0.0460868143,
0.0046607782,
0.0000019788,
-0.059205424,
-0.0091020484,
0.0067926045,
-0.0104213422,
-0.0278438032,
0.0646825358,
-0.0653573126,
-0.0176840294,
0.0310010016,
-0.0237303711,
-0.0302781109,
0.0547148176,
-0.0033699167,
-0.0175733529,
0.011796861,
-0.0166367497,
0.0246818643,
-0.0298366621,
-0.0250062682,
0.0036344666,
-0.021273924,
0.0263731759,
-0.0012023394,
-0.0029035299,
-0.1014519781,
-0.0156289265,
0.0267965831,
0.0099100396,
0.035262756,
-0.008027846,
0.0459022224,
-0.0273072459,
-0.0172431935,
0.0056139384,
0.0224517733,
0.002860558,
0.0060356278,
0.0139967604,
0.0011356244,
0.0140521396,
-0.0334370323,
-0.0163491014,
0.0417082943,
-0.0422852375,
-0.0001370775,
-0.0086912913,
0.0071095051,
0.0272106845,
0.0299394913,
0.0004844628,
0.0498919003,
0.0150695499,
-0.0203672685,
-0.0051660137,
0.0296640471,
0.0168425273,
-0.0397524834,
0.0182995815,
-0.0152338576,
-0.0454843044,
0.0279622674,
-0.0157469884,
0.0002847166,
0.0614292212,
-0.0819911584,
0.0074396436,
0.0330266207,
0.0216430798,
-0.0207744464,
0.0180118028,
-0.0126234423,
-0.0103051569,
-0.0171872694,
-0.0235532075,
0.0294219982,
0.0147711877,
-0.0190737508,
0.0088854004,
-0.0164278671,
0.0195834767,
0.0604710616,
-0.0925686881,
-0.0409777053,
-0.0564706996,
0.0127573265,
0.021353405,
0.0206393041,
-0.0738247856,
-0.0050987373,
-0.0284919254,
-0.0104181292,
0.0027564981,
0.0243344661,
-0.0238161702,
-0.0616189465,
0.03632522,
0.0036689257,
-0.0360303856,
-0.0110546062,
-0.0244183671,
0.0247616135,
0.0065699257,
0.0194502175,
-0.0033090191,
-0.049599003,
-0.0241412446,
-0.0303608757,
-0.0156456698,
-0.0745610595,
0.0132078361,
0.0267121624,
0.0283179488,
-0.0809741765,
0.0051596165,
0.0169631205,
-0.0846661925,
0.0329499729,
-0.0328123346,
0.0234303735,
0.0404717736,
0.0524173416,
-0.0019477428,
0.0314470083,
0.0417664126,
-0.0371881872,
-0.0084342388,
-0.045810964,
0.0002123023,
0.0186653417,
0.0173423234,
0.0151914414,
0.0063419263,
0.0167411473,
0.0237827376,
0.0376793891,
0.0212823693,
0.0455508754,
-0.0005756753,
-0.0147810588,
0.0225170609,
0.0660221651,
-0.0098199267,
0.0328327082,
0.0366846845,
-0.0186105557,
-0.0257136207,
-0.0009580491,
-0.0511325337,
-0.0397219397,
0.0053079035,
-0.036193911,
0.0190238673,
0.0394469351,
-0.0138358921,
-0.01555264,
-0.0276993029,
0.0113536334,
0.0251937211,
-0.0240600966,
-0.0308168102,
-0.0273629017,
-0.021539811,
0.0868894309,
0.0574455559,
-0.0049158195,
-0.0366093405,
-0.0676033273,
-0.0392175056,
0.0644861758,
0.011310027,
0.0086835716,
0.0100632701,
-0.0290443599,
-0.0125216534,
0.0280701965,
0.0620629489,
0.0202782396,
-0.0052321614,
-0.0334810689,
0.0019232542,
0.0557961054,
0.0403340869,
0.0151684862,
0.055489663,
0.0647887141,
-0.0029855333,
-0.0120101999,
-0.0704852045,
0.0284708198,
-0.0572284982,
0.007859665,
0.0081776539,
-0.0321310163,
0.049351234,
0.0054181535,
-0.0162218492,
0.0350548849,
-0.0657106787,
0.0254985727,
-0.001700776,
0.0424226336,
-0.0598482378,
0.0370450467,
-0.0041859727,
-0.055791717,
-0.0103167417,
-0.0103757689,
0.01671586,
0.0190262627,
0.0102640083,
-0.0152846025,
0.0258417539,
0.0169200003,
0.0163456555,
0.0314561203,
0.0120162591,
-0.0110228118,
0.0084649147,
-0.0312620029,
0.0053380304,
0.0109877307,
-0.0601103045,
-0.0076196063,
-0.0242894478,
-0.0003507185,
0.064892821,
0.0019837681,
0.0129848747,
-0.0023449243,
-0.0012156701,
-0.0093749631,
0.031287998,
-0.0838707685,
-0.0051328312,
0.0165056568,
-0.0308498871,
-0.0006484641,
-0.0063172416,
-0.031148307,
0.0081808046,
-0.0080212681,
-0.0622975901,
0.060087584,
0.0338818878,
-0.0228972081,
-0.0103329513,
0.0067500994,
-0.0278502051,
-0.0424521603,
-0.0603774823,
0.032105621,
0.0396702625,
-0.0012189972,
0.007315068,
-0.0128007485,
-0.0530707985,
-0.0295167323,
0.0102455998,
-0.0328397229,
0.0331165306,
0.0190386064,
0.0127900932,
0.0088548223,
-0.0075840228,
0.0377840251,
0.0548993312,
-0.0161877852,
-0.0155435279,
-0.0065136384,
-0.022940658,
-0.0092763565,
0.048122853,
-0.1373877227,
-0.017987188,
0.0739054158,
0.0648865104,
-0.0124318274,
-0.0585543551,
0.0361768976,
-0.082028307,
-0.0249126293,
-0.0054789516,
-0.0231744815,
-0.0133513045,
0.0530409664,
-0.0152516412,
-0.0358673595,
0.0369492881,
0.0246989187,
0.0672261342,
-0.0807375237,
-0.0116384234,
-0.035320431,
-0.0186691694,
-0.0128106354,
0.0171003193,
0.0285700578,
0.0086603789,
-0.0129983593,
-0.0139351245,
0.0109896874,
-0.053983476,
0.0758490786,
-0.0199270435,
0.0123238433,
-0.0047954344,
-0.0470848046,
-0.0225888379,
-0.0418834724,
-0.0289361551,
-0.010704861,
0.0306854304,
0.0107007837,
-0.0405624434,
0.039784234,
0.0244772341,
0.0709405243,
-0.0200370867,
-0.0088656144,
0.0256603379,
-0.0744330212,
0.0060861558,
-0.0480863787,
0.033344198,
-0.0504752733,
0.0457995795,
0.028183708,
-0.003946295,
0.0484682098,
0.0255491417,
0.013168769,
0.0074418909,
-0.0097541474,
-0.0387999304,
0.0368325934,
0.0134132477,
-0.0429474749,
0.0449897386,
0.0883246809,
-0.00688176,
-0.051981613,
0.033491347,
-0.0187732372,
-0.0087424805,
-0.0126821008,
0.0058004092,
-0.0927482843,
0.0236851871,
0.0038894427,
0.033379592,
-0.0605569296,
-0.0431925282,
-0.0181291327,
-0.0407074541,
0.0263398681,
0.0326426961,
-0.0156221408,
-0.018751286,
-0.0159784872,
-0.0029558907,
0.0170942694,
0.010258792,
-0.0303696524,
0.0151026472,
0.0192865189,
-0.0132783819,
-0.1253111213,
-0.0625368729,
0.0324635431,
0.0099571636,
-0.0147119882,
0.0435339771,
0.0293556936,
-0.0291343108,
0.0081011346,
0.0122031178,
0.0127719063,
-0.0182064883,
-0.0246785525,
0.0773859769,
0.0137758413,
0.0117176566,
-0.0839996859,
0.0152609013,
-0.002345751,
-0.0054359986,
-0.0033807976,
0.0348718502,
-0.015856903,
-0.0664824694,
-0.0335839391,
0.0555942617,
-0.0041023642,
0.075660035,
0.0145880505,
0.0085558398,
-0.0007340641,
-0.0592241362,
0.0533754863,
-0.0077134473,
-0.0587050915,
0.0338809006,
-0.0217620134,
-0.0387524888,
0.0062615066,
-0.0702321902,
0.0342191495,
-0.0269133449,
0.0076936996,
-0.0145892287,
0.0318310708,
-0.0396862999,
-0.0448928289,
0.0590280853,
0.011817229,
0.0488381758,
0.01153256,
-0.0363844298,
0.0140941702,
0.0248757135,
-0.0079354504,
0.0023323419,
0.0181900766,
-0.0148686348,
0.0481929146,
0.0140508376,
0.0300862528,
-0.0037925707,
0.0265048258,
-0.019059319,
-0.0070680087,
0.0447789505,
-0.013716233,
-0.0127078202,
-0.0473946929,
0.0007261344,
0.010768122,
-0.0013236477,
-0.0032517649,
0.0524742976,
0.0098389275,
0.0073243664,
-0.031903673,
0.0257895924,
0.0138161257,
0.0012216917,
0.0108301388,
-0.0496369153,
-0.0062949494,
-0.0304615367,
-0.0238629077,
-0.0543834269,
-0.0178492796,
-0.0123607721,
-0.0277462434,
0.0077500516,
0.0367044434,
-0.0539317168,
-0.0025596907,
-0.0685803071,
-0.0226756483,
-0.0431793034,
0.0709759668,
0.0001664884,
0.0071816309,
0.0716973767,
-0.0311556067,
-0.0114545124,
0.0166926831,
-0.0184133742,
0.0385352671,
0.0411236361,
0.0717657283,
0.0078390809,
0.0077565503,
0.0185420401,
-0.0027635091,
-0.0303112958,
-0.0677311048,
-0.0181922093,
0.0095126722,
0.0144094992,
-0.0228979886,
0.0261864383,
-0.0231294371,
-0.0023710404,
-0.0070139961,
-0.006375046,
0.0243185461,
0.0349580497,
0.0633083284,
-0.0069348956,
-0.01412935,
-0.041201584,
0.0840658471,
0.0160017461,
0.0614745319,
-0.0433379002,
-0.0511579886,
-0.0015689266,
-0.0681672692,
0.0325286165,
-0.0183300264,
-0.0329258032,
0.0075344546,
-0.1110832244,
-0.0372461565,
-0.0166616924,
-0.0220671035,
-0.0023432679,
-0.0352124237,
0.0005447666,
0.0591935143,
-0.0094552496,
-0.0467050746,
0.0049747908,
0.0624013171,
-0.0547546819,
-0.0220432095,
0.0523119308,
-5.810832612e-33,
-0.0078985598,
-0.0428738967,
-0.0053261933,
-0.0474690944,
-0.0014095502,
0.0265288968,
-0.0206158534,
-0.0005546287,
-0.007755748,
0.0369986296,
-0.0096995691,
0.0152580421,
0.0132665569,
0.0108603677,
0.0033023267,
0.0087375194,
0.0290596317,
0.0107569741,
-0.0271740668,
0.0012228686,
-0.0847103149,
0.0159366615,
0.0271090977,
-0.0390924811,
0.0561677553,
-0.0013903722,
0.0140971104,
-0.0200616699,
-0.0004630342,
0.0300206337,
-0.0057455972,
0.0479850806,
0.039789889,
-0.0423389003,
0.0095944572,
0.0125435935,
0.0190406796,
-0.0035299985,
0.035293553,
-0.0222478565,
-0.0198430978,
-0.0349703915,
0.041751463,
0.0301414374,
-0.0792633891,
-0.0304641966,
0.0281220954,
-0.0331552364,
0.0439182892,
0.0092404522,
-0.0082997987,
0.0673328042,
-0.0075664716,
0.0790327638,
-0.0416156836,
0.0171155073,
-0.0404077955,
0.0403950438,
-0.0172457229,
-0.021982396,
0.0349826254,
0.0070648408,
0.0359836109,
-0.0145079717,
-0.0273184273,
0.0133417826,
0.0477928035,
0.0474321246,
-0.0621937364,
-0.0159937404,
-0.0316987857,
-0.0350722186,
0.0261017941,
0.0672614351,
0.0236112643,
-0.0454781465,
-0.0272794534,
0.020648336,
-0.0418564454,
0.0273419674,
0.0404763669,
-0.0565217584,
-0.090648897,
-0.0581911691,
-0.038588699,
-0.022518713,
0.0002090883,
-0.0284330249,
0.0379617549,
-0.0126060871,
0.0283054803,
0.0419742055,
0.0152355796,
-0.007631795,
-0.0111965742,
-0.0298436955,
0.0018659184,
-0.0342979617,
-0.0239871591,
0.0217750166,
-0.0585621521,
0.0803052112,
-0.0264220983,
0.0298327357,
0.0477048345,
0.039388746,
0.0031493308,
-0.0421823002,
-0.0660481006,
0.0150831901,
0.0430007018,
0.0225073155,
0.0186038986,
0.0081797149,
0.0447534956,
0.0376596339,
0.0323998444,
0.0083683031,
0.0370059609,
-0.0298455544,
-0.0311357826,
-0.0592461564,
-0.0327800736,
0.0457385033,
-0.0623734519,
0.0074595981,
-0.0056494819,
0.0393497571,
0.0873044059,
-0.0530710891,
-0.0062352917,
-0.0635069609,
2.527e-7,
0.0498501696,
0.0499935448,
-0.0321000367,
-0.0003777256,
0.0328845307,
0.0111316787,
0.0288641937,
0.0103198616,
0.0406840891,
-0.0001397069,
0.0004600923,
0.0188735463,
0.0425870232,
0.0590804964,
-0.0282678809,
0.008380129,
-0.1168672815,
-0.0133367572,
-0.0531421229,
-0.0103020091,
0.0091263419,
0.0543109365,
-0.0319258496,
-0.0179927591,
-0.0534009896,
0.0211569443,
0.0243626609,
-0.0611103103,
-0.0008186549,
0.0503206924,
0.0229030661,
0.0069145299,
0.0164924171,
0.0599401444,
-0.0126773724,
-0.0013273811,
0.0196362883,
0.0727797374,
-0.0408122353,
0.0752734691,
0.02278441,
-0.025929939,
0.006702675,
-0.082983017,
0.0377066955,
0.0213006493,
-0.0417656377,
-0.0460626222,
-0.0260798093,
-0.0182803776,
0.0698613748,
-0.0032827354,
-0.0424440168,
0.0243144389,
0.0073940773,
0.040659111,
-0.004708563,
-0.0170440655,
-0.0291380174,
0.0025289799,
-0.0208134037,
-0.0531855673,
0.022053659,
0.0245471168,
-0.003792611,
-0.0267301276,
0.0547373146,
2.35965384e-34,
0.033164449,
0.0317871124,
0.0359764248,
-0.0579060987,
0.0522125512,
0.014027359,
0.0321338587,
0.0089521902,
0.0149797592,
-0.0029779391,
-0.0286063515
] |
http://arxiv.org/abs/2401.13849v1 | http://arxiv.org/pdf/2401.13849v1 | TPD: Enhancing Student Language Model Reasoning via Principle Discovery and Guidance | 2024-01-24T23:11:33 | 2024-01-24T23:11:33 | Haorui Wang, Rongzhi Zhang, Yinghao Li, Lingkai Kong, Yuchen Zhuang, Xiusi Chen, Chao Zhang | cs.CL | cs.CL, cs.AI | Large Language Models (LLMs) have recently showcased remarkable reasoning
abilities. However, larger models often surpass their smaller counterparts in
reasoning tasks, posing the challenge of effectively transferring these
capabilities from larger models. Existing approaches heavily rely on extensive
fine-tuning data or continuous interactions with a superior teacher LLM during
inference. We introduce a principle-based teacher-student framework called
``Teaching via Principle Discovery'' (TPD) to address these limitations.
Inspired by human learning mechanisms, TPD mimics the interaction between a
teacher and a student using a principle-based approach. The teacher LLM
generates problem-solving instructions and corrective principles based on the
student LLM's errors. These principles guide the refinement of instructions and
the selection of instructive examples from a validation set. This enables the
student model to learn from both the teacher's guidance and its own mistakes.
Once the student model begins making inferences, TPD requires no further
intervention from the teacher LLM or humans. Through extensive experiments
across eight reasoning tasks, we demonstrate the effectiveness of TPD. Compared
to standard chain-of-thought prompting, TPD significantly improves the student
model's performance, achieving $6.2\%$ improvement on average. | TPD: Enhancing Student Language Model Reasoning via Principle Discovery and Guidance: Large Language Models (LLMs) have recently showcased remarkable reasoning
abilities. However, larger models often surpass their smaller counterparts in
reasoning tasks, posing the challenge of effectively transferring these
capabilities from larger models. Existing approaches heavily rely on extensive
fine-tuning data or continuous interactions with a superior teacher LLM during
inference. We introduce a principle-based teacher-student framework called
``Teaching via Principle Discovery'' (TPD) to address these limitations.
Inspired by human learning mechanisms, TPD mimics the interaction between a
teacher and a student using a principle-based approach. The teacher LLM
generates problem-solving instructions and corrective principles based on the
student LLM's errors. These principles guide the refinement of instructions and
the selection of instructive examples from a validation set. This enables the
student model to learn from both the teacher's guidance and its own mistakes.
Once the student model begins making inferences, TPD requires no further
intervention from the teacher LLM or humans. Through extensive experiments
across eight reasoning tasks, we demonstrate the effectiveness of TPD. Compared
to standard chain-of-thought prompting, TPD significantly improves the student
model's performance, achieving $6.2\%$ improvement on average. | [
0.0547350645,
0.0449105725,
-0.034785904,
0.0646706,
-0.0539881326,
-0.0007606295,
-0.0011104462,
-0.0567040183,
-0.0640751272,
-0.0400662348,
0.0078189094,
-0.0108192721,
-0.0150891906,
-0.0212002378,
0.033945974,
-0.061795909,
0.0329556949,
-0.0425553732,
0.0371848457,
0.0121130534,
-0.0151426839,
-0.0111749759,
-0.036398109,
0.0400177948,
-0.0028542562,
-0.012719431,
-0.0393806919,
-0.0105775688,
0.0022187252,
-0.0355828963,
0.0013058411,
0.0523354299,
0.0408890508,
0.0155060906,
0.0000023544,
-0.0538077988,
-0.0051438557,
-0.0241282303,
-0.0219078958,
0.0391735248,
-0.005883568,
0.0197178219,
-0.0200877078,
0.0314437337,
-0.0471409932,
-0.0032509775,
0.023318477,
0.0567199662,
0.0301381685,
0.0371079147,
0.0168113261,
-0.0175407641,
0.0097898692,
0.0091151139,
0.0181717835,
-0.0454140902,
0.0222846195,
0.0205176286,
0.061434269,
0.0264289286,
-0.0452119969,
0.0236597992,
0.0504542217,
-0.0361426398,
-0.0286992751,
0.0356936008,
-0.0238007884,
-0.0143176289,
0.0249891691,
0.0533793718,
0.0308413245,
-0.0358256027,
-0.038372919,
0.0149810081,
0.0258055944,
0.0482527278,
-0.0587289967,
0.0317778029,
-0.0217593107,
0.0348960422,
0.0465855189,
-0.0607238822,
0.0246156659,
0.0493121482,
0.0032636558,
0.0692960024,
-0.0104629565,
-0.0370914191,
0.0039536795,
0.0190068725,
-0.0632846132,
-0.0003255703,
0.0185329728,
0.0554801226,
-0.017162459,
0.0454185456,
-0.0048887767,
-0.0078672506,
-0.0122734653,
-0.0063394541,
0.0489179604,
0.0734566525,
0.0100342818,
0.0338833518,
-0.0204714965,
0.0379359163,
-0.0050724894,
0.0458003357,
-0.0150681641,
0.0068392791,
-0.0537698828,
-0.0156766996,
-0.0272677448,
0.028514659,
0.0362874046,
0.0171360318,
-0.0214182138,
0.0239412691,
-0.0108174831,
0.0260843188,
-0.018832216,
-0.0124181956,
0.0171958972,
-0.0233880635,
-0.0563658811,
0.0631731749,
-0.0233792197,
0.0454354584,
0.004455362,
-0.0564789698,
-0.0251941606,
-0.0221832059,
-0.0220946465,
0.0137481084,
0.0084379241,
0.0113940407,
-0.0510517731,
-0.0460698195,
-0.0731081367,
-0.0166989416,
0.0198591482,
-0.0551343672,
-0.0020223577,
-0.0755536184,
-0.0112433406,
0.0033830199,
-0.0287963003,
-0.0071041202,
0.0400551893,
0.0120584937,
0.0148380753,
0.0709699988,
-0.0251077563,
0.0268899035,
0.0156070292,
0.055327937,
-0.0372999683,
0.0605029017,
0.0384442769,
-0.0205645394,
-0.0045360466,
-0.0167939607,
0.006829638,
0.0068516443,
0.0114634018,
-0.0015237374,
0.0072271302,
-0.033983469,
-0.0009800288,
0.0494741276,
-0.0179116037,
0.0420857742,
-0.0229450595,
0.0505658835,
0.0649405718,
0.0514102317,
-0.0366268307,
0.040454749,
0.0262559783,
-0.0120185614,
0.0364528634,
-0.0037062743,
-0.071327053,
0.0113948109,
-0.0369219817,
0.0039468072,
-0.0080131367,
0.0065707765,
-0.0104952864,
-0.0737435892,
-0.0114071677,
0.0245309994,
-0.0173436366,
0.0312072635,
0.0744357854,
0.0197421685,
-0.0033046613,
0.0491369553,
0.0318469554,
-0.0369167738,
0.0264585167,
-0.0713729933,
0.0702262074,
0.0721647143,
0.0762212649,
-0.0439773239,
-0.0135983564,
-0.0043360065,
0.0523153879,
0.0159725118,
0.0448547266,
-0.032039348,
-0.0425758027,
-0.0302820131,
-0.0129783684,
-0.0207878891,
0.0063940389,
-0.0220050886,
0.0308850016,
-0.0142832762,
0.0113959638,
-0.025482282,
-0.0058941608,
0.031386286,
-0.0621457323,
-0.0150991809,
-0.0303516015,
0.0053077312,
-0.0258384831,
0.0231788363,
0.0265758764,
0.016900437,
0.0287224781,
-0.0123541588,
0.0241055544,
0.0087246932,
-0.0368520096,
-0.0231342837,
0.0554087125,
-0.0525269955,
-0.0318388157,
0.0054227286,
-0.03165913,
-0.0223869868,
0.0468351655,
-0.0181186795,
0.0118402271,
-0.0134915262,
0.0423289239,
-0.0106958458,
-0.0274802074,
-0.0173201561,
0.0453739874,
0.0051392391,
0.0253986623,
0.038262289,
0.0293503422,
-0.01264684,
-0.0211533904,
-0.0038995263,
0.0013475096,
-0.0168130379,
-0.0048150718,
-0.005392266,
0.0057923398,
-0.0671859235,
-0.0050254823,
-0.0254422445,
-0.0324819125,
0.0023613803,
0.0140737072,
-0.001041075,
-0.0176397655,
-0.0147543959,
-0.0038334352,
0.0101572061,
-0.0251640622,
-0.0078232335,
-0.0039127083,
0.0202063005,
0.0091597782,
-0.0652937666,
-0.0315847322,
0.0032349764,
-0.0423982143,
-0.0015634015,
0.0292483345,
0.0138460184,
-0.0529859848,
0.0636417717,
-0.0269048419,
0.0339514837,
-0.0141852312,
-0.0301067606,
0.0392369963,
-0.0351106189,
0.0172943305,
0.0853181407,
-0.0223192368,
0.056851767,
0.0902609378,
-0.0083412761,
-0.05738803,
0.0303228311,
-0.0172781795,
0.003349971,
-0.0272935256,
-0.023135215,
-0.048562739,
0.0305435229,
-0.0019987451,
-0.0031039049,
-0.0360278897,
0.05171334,
-0.0132527398,
0.0033701449,
-0.0232817885,
-0.0278135948,
0.0069090729,
0.0088403327,
-0.0302693099,
-0.0352870636,
0.1072320193,
0.0193025954,
-0.0069371583,
-0.0217920877,
0.0075206906,
-0.0175085142,
0.0563583821,
-0.029196078,
-0.0180451833,
-0.0091775274,
0.0207544379,
-0.049361743,
-0.0001841402,
0.0627603456,
-0.0330005176,
0.0840800554,
-0.1110610291,
-0.0240566023,
0.0179355089,
-0.0452016927,
-0.082567744,
0.0417559035,
-0.0565821342,
0.021893952,
-0.0313033909,
-0.0268705487,
-0.0114249224,
-0.0309005491,
0.051658798,
-0.0079990625,
0.0015588781,
-0.0351996124,
-0.019460449,
-0.0370521247,
0.0008760331,
-0.0046759765,
0.0368230976,
-0.0471360348,
0.0690560862,
0.0437402725,
0.0137346005,
-0.0157854855,
-0.0085102953,
0.0227575563,
0.0165397488,
0.0171002522,
-0.0182565432,
0.056031011,
0.0190105978,
-0.0914410949,
-0.0468532592,
0.1256312877,
0.0576854944,
-0.0119960532,
0.0142282788,
-0.0657551885,
0.0260828231,
0.0220179651,
0.0345987342,
-0.0528710745,
-0.0375149734,
0.0114007276,
0.0056248577,
-0.0016861437,
-0.0383242965,
-0.0334434621,
-0.1046562493,
0.0425708182,
0.0332871974,
0.0105176084,
-0.0384726115,
-0.0748033449,
0.0092216413,
0.0080265459,
0.0152521711,
0.0256843194,
0.0122244889,
0.0218303371,
0.0567036942,
0.0092329057,
-0.0889703184,
0.0073225885,
-0.030498229,
0.0150840143,
0.0085066715,
-0.0058400375,
0.0024045308,
-0.124965407,
0.0249141306,
0.0181588959,
0.0180007685,
-0.0218322799,
-0.0031671817,
-0.0005604243,
0.0156581849,
-0.0213712342,
0.0823750198,
-0.0414515659,
-0.0713589638,
0.0164713897,
0.122367166,
-0.0166905429,
-0.0142457355,
-0.0156251118,
0.0124121355,
0.001288449,
0.0585654601,
0.0524744019,
-0.0131786522,
0.0008591411,
-0.0365193971,
-0.0628094822,
-0.013657134,
-0.0084281741,
0.0089358548,
0.0314159952,
0.0357688628,
-0.0062100827,
-0.0146880765,
0.0796868354,
0.0393395387,
0.0218694415,
0.0059970357,
-0.0045022769,
0.0104374737,
-0.0098059755,
0.0204221401,
-0.0146561014,
0.0233467426,
-0.0626192465,
-0.0022294456,
-0.0000931474,
-0.0398407057,
0.0094506238,
-0.0099473055,
-0.058963906,
-0.0892731175,
0.0184065904,
0.0323185697,
0.0194725785,
0.0396633483,
0.0105511956,
-0.0653514564,
-0.0116673466,
0.0401335508,
-0.0044739698,
-0.0294418819,
0.0261105467,
-0.033114776,
0.0398143232,
0.0196788162,
-0.0034192202,
0.0161652751,
0.0244039986,
0.0534270927,
-0.0276697073,
-0.0188755728,
0.0586437136,
0.0417718776,
0.053540159,
-0.0112088975,
-0.0075096702,
-0.0516745225,
0.0304295588,
-0.0166733749,
-0.0265191086,
-0.0292263366,
0.0262264088,
0.0244095642,
-0.0110821975,
0.014411971,
-0.0612994768,
-0.0042437734,
0.0251134578,
0.0248624291,
0.0139858024,
0.004784998,
0.0357689746,
0.045440916,
-0.0173715819,
0.0244307797,
0.0261322614,
-0.0091459118,
0.0712775737,
0.005885534,
0.0974357203,
0.0357995555,
-0.0031311205,
-0.0299141631,
-0.0285756774,
-0.0546299852,
-0.0534086786,
0.0027462698,
0.041365318,
0.0368469618,
0.014815405,
-0.0224764924,
-0.0918844268,
0.0075086141,
0.0241513085,
0.0443857424,
-0.0194486547,
-0.0506876297,
0.0656995624,
0.033829134,
0.0309200212,
-0.0184931979,
0.0131621631,
-0.0368127711,
0.039606709,
-0.0355153121,
0.0113492552,
-0.0103559662,
-0.0109704314,
0.0295584891,
0.0050502634,
-0.0215646997,
0.0599654131,
-0.0657117143,
-0.0662085563,
-0.0559872128,
0.0444894545,
0.0000207648,
-0.028421592,
0.0291164499,
0.0555938333,
-0.0341712385,
-0.0362926349,
-0.0068656383,
-0.0009877861,
0.0116924401,
-0.0250572376,
-0.0350249484,
-6.499347204e-33,
-0.0404382534,
-0.0316313878,
0.0519426875,
0.0486113504,
0.0186065193,
-0.0388358235,
-0.0618379451,
0.035898421,
-0.0014154286,
0.0504854284,
-0.0066524069,
0.0198941026,
-0.0205099843,
-0.0066033737,
0.0128661292,
-0.0075083729,
0.0023269365,
0.031780187,
-0.0378629491,
-0.0123628443,
0.029162664,
0.0008473821,
0.0552953258,
-0.0371205173,
0.0188051369,
-0.0107668526,
-0.0139564248,
0.0724883229,
-0.0936667472,
0.0385967903,
0.0173985492,
0.0143187316,
0.0085807424,
-0.0228720605,
-0.0142926648,
-0.005787875,
-0.0378619842,
-0.0609351359,
0.0636151284,
-0.0047065229,
-0.0464953035,
-0.0292558465,
0.0073079066,
0.0052267769,
-0.0149761867,
0.0195953473,
0.0360787064,
-0.0318729281,
0.0016602043,
0.0215338878,
-0.038501747,
-0.0141493147,
-0.0422781631,
-0.0457842238,
0.0188554917,
0.0264280867,
-0.0395989269,
-0.0492435247,
-0.0406565927,
0.0480250642,
0.10029421,
0.0182222109,
0.0715972334,
0.016715169,
-0.0151336547,
0.0432531349,
-0.0016652695,
0.0094720563,
-0.0309315529,
-0.0448462777,
0.0427814759,
0.0369004086,
0.0574379936,
0.0102423169,
-0.001719777,
0.0126549592,
-0.0581022054,
-0.0088226581,
0.0397093222,
0.0325303413,
0.0079718288,
-0.00105717,
-0.0485563166,
-0.0467674471,
-0.0539678261,
0.0077884272,
-0.0329092741,
-0.0173995029,
0.0306833852,
-0.0059457798,
-0.0412206016,
-0.064582631,
-0.0199321415,
-0.0425511114,
-0.035845235,
0.0039351862,
-0.0218250453,
-0.0030395268,
-0.0286705755,
-0.0078243101,
-0.0939373225,
-0.0487782769,
0.0126608955,
-0.0243811626,
-0.0006875696,
0.0146743599,
0.036518611,
0.0271240845,
-0.0127648814,
-0.0087578762,
-0.0599495955,
-0.0078473957,
0.0462084971,
-0.055215992,
-0.0054067345,
0.0343027301,
-0.0338356867,
0.0206121057,
0.0302015599,
-0.0352620184,
-0.0397713818,
0.06711144,
-0.0290017109,
-0.0026221559,
-0.0090880021,
-0.0226412192,
-0.0297176111,
0.032045044,
0.0307230949,
-0.0647572875,
-0.011437363,
0.0183995273,
3.048e-7,
0.0147208571,
0.0279057436,
-0.0088478802,
0.0078129852,
0.0026667935,
0.0364542343,
-0.0445918292,
-0.0486738868,
0.0569285266,
-0.0334890634,
0.0278879609,
0.070591718,
0.0249307808,
0.0277714282,
-0.0639737025,
0.0073012677,
-0.0026224805,
-0.0216068476,
-0.028352052,
0.0029452369,
0.050125856,
0.0330887847,
0.0464321449,
-0.0047008721,
-0.0086945705,
-0.0069577051,
0.0334762223,
-0.0218197908,
-0.009483221,
0.0203776024,
0.0151067702,
-0.0306456555,
0.0160120353,
0.0409153327,
-0.009171051,
-0.0031446619,
-0.0513171256,
0.0208254699,
0.0079595903,
0.0341306627,
-0.0381217934,
-0.0241083801,
0.0109029282,
-0.0518090688,
0.0291083399,
-0.065837577,
0.0016334042,
0.0603046268,
0.0246741958,
-0.0336296931,
-0.0153114721,
-0.0063378243,
-0.0132234255,
-0.0069288379,
-0.0168567244,
-0.0282592587,
-0.0021135989,
-0.0396686979,
-0.0208733007,
0.0007241647,
-0.0110311778,
-0.0292206462,
-0.0266973842,
-0.0592527911,
-0.0051535289,
-0.0246037319,
0.0223100968,
2.665955525e-34,
0.018443279,
-0.0212752316,
0.0440110974,
0.0415323451,
0.0342059545,
0.0035439509,
0.001846591,
0.0259669945,
0.0521061271,
-0.0330972336,
-0.0222112648
] |
http://arxiv.org/abs/2402.01704v2 | http://arxiv.org/pdf/2402.01704v2 | States as Strings as Strategies: Steering Language Models with Game-Theoretic Solvers | 2024-01-24T22:22:00 | 2024-02-06T08:53:11 | Ian Gemp, Yoram Bachrach, Marc Lanctot, Roma Patel, Vibhavari Dasagi, Luke Marris, Georgios Piliouras, Siqi Liu, Karl Tuyls | cs.CL | cs.CL, cs.AI, cs.GT | Game theory is the study of mathematical models of strategic interactions
among rational agents. Language is a key medium of interaction for humans,
though it has historically proven difficult to model dialogue and its strategic
motivations mathematically. A suitable model of the players, strategies, and
payoffs associated with linguistic interactions (i.e., a binding to the
conventional symbolic logic of game theory) would enable existing
game-theoretic algorithms to provide strategic solutions in the space of
language. In other words, a binding could provide a route to computing stable,
rational conversational strategies in dialogue. Large language models (LLMs)
have arguably reached a point where their generative capabilities can enable
realistic, human-like simulations of natural dialogue. By prompting them in
various ways, we can steer their responses towards different output utterances.
Leveraging the expressivity of natural language, LLMs can also help us quickly
generate new dialogue scenarios, which are grounded in real world applications.
In this work, we present one possible binding from dialogue to game theory as
well as generalizations of existing equilibrium finding algorithms to this
setting. In addition, by exploiting LLMs generation capabilities along with our
proposed binding, we can synthesize a large repository of formally-defined
games in which one can study and test game-theoretic solution concepts. We also
demonstrate how one can combine LLM-driven game generation, game-theoretic
solvers, and imitation learning to construct a process for improving the
strategic capabilities of LLMs. | States as Strings as Strategies: Steering Language Models with Game-Theoretic Solvers: Game theory is the study of mathematical models of strategic interactions
among rational agents. Language is a key medium of interaction for humans,
though it has historically proven difficult to model dialogue and its strategic
motivations mathematically. A suitable model of the players, strategies, and
payoffs associated with linguistic interactions (i.e., a binding to the
conventional symbolic logic of game theory) would enable existing
game-theoretic algorithms to provide strategic solutions in the space of
language. In other words, a binding could provide a route to computing stable,
rational conversational strategies in dialogue. Large language models (LLMs)
have arguably reached a point where their generative capabilities can enable
realistic, human-like simulations of natural dialogue. By prompting them in
various ways, we can steer their responses towards different output utterances.
Leveraging the expressivity of natural language, LLMs can also help us quickly
generate new dialogue scenarios, which are grounded in real world applications.
In this work, we present one possible binding from dialogue to game theory as
well as generalizations of existing equilibrium finding algorithms to this
setting. In addition, by exploiting LLMs generation capabilities along with our
proposed binding, we can synthesize a large repository of formally-defined
games in which one can study and test game-theoretic solution concepts. We also
demonstrate how one can combine LLM-driven game generation, game-theoretic
solvers, and imitation learning to construct a process for improving the
strategic capabilities of LLMs. | [
0.0917834491,
0.0850019455,
-0.0541198403,
0.0472345948,
-0.0748300925,
-0.03423585,
-0.0167488102,
0.0244486295,
0.0042328294,
-0.0276623517,
0.0127224,
-0.0432499386,
0.0132816238,
0.0587074794,
0.0140391449,
-0.0470540673,
0.0239780564,
-0.033202149,
0.0584370606,
-0.0089447238,
-0.0029735214,
-0.0198100042,
-0.0265701022,
0.0267806835,
-0.021345593,
0.0272907428,
-0.0641711727,
-0.0044563678,
0.0246834122,
-0.0084632384,
-0.0230135433,
-0.0046128826,
0.0350150838,
0.0028342099,
0.0000022089,
-0.0724076331,
-0.0149680106,
-0.0360603854,
-0.0096954387,
-0.0290611777,
0.0487578064,
0.0133493859,
0.0233470034,
0.0144755263,
-0.072583355,
0.0312219337,
0.0487695187,
0.0030549949,
0.075182803,
0.098200798,
0.0061298269,
-0.0302589592,
0.0035382204,
-0.0248475056,
-0.0429163091,
-0.1003833115,
0.0321840942,
0.031692069,
0.0415775031,
-0.0261348337,
-0.0180431828,
0.0163521878,
-0.0082801925,
-0.0213031601,
-0.0538994446,
0.0409671888,
0.0093728565,
0.0256219991,
-0.0383570977,
0.0525005385,
0.0381097384,
0.0020791846,
0.0555165894,
-0.0228316281,
0.0165387783,
0.0210391898,
0.01313686,
0.0241795406,
0.0089055169,
0.0394889377,
0.0208206251,
0.0020625119,
0.0145512456,
0.0327428505,
-0.0006205216,
0.0721240565,
-0.0230734367,
-0.0308517274,
-0.0007795165,
-0.0010371258,
-0.025291549,
-0.1037464589,
-0.0098404502,
0.0478342511,
-0.0804032534,
0.0444671102,
-0.0290950835,
0.0138886636,
-0.0140194288,
-0.0331531093,
0.0060025477,
0.0286027528,
0.0807215869,
0.0382087193,
0.0732459351,
0.0527089238,
-0.0091547286,
-0.0190479346,
-0.0231566597,
0.0204773545,
-0.0251459405,
-0.0258774571,
-0.0728095993,
0.021565767,
-0.0436808169,
-0.0206591841,
-0.0052960496,
0.0163660143,
-0.0607768744,
0.0464016311,
-0.0226628277,
0.0279674809,
-0.0225913525,
-0.0043639415,
-0.0606152415,
-0.0185852256,
-0.0260095559,
0.0191029124,
0.0103752445,
-0.05990218,
-0.001083968,
-0.0165038686,
0.0198445469,
-0.0357973203,
0.0025571301,
0.0773384869,
-0.0165093355,
-0.003748823,
0.0385269746,
-0.0370969884,
-0.0191523936,
-0.0787976608,
-0.0155254472,
-0.0412610434,
0.0544293076,
-0.0453309044,
-0.0089131575,
-0.0646327361,
0.0073887301,
-0.0157909412,
-0.0204275679,
0.0604288019,
0.030044673,
0.0206600688,
0.009839043,
0.0020861041,
-0.005256345,
0.0648903325,
0.0681829602,
-0.0165733714,
0.0346149281,
0.007748046,
0.0229998492,
0.0610990487,
0.0107463431,
0.010931178,
-0.0053383629,
0.0015613779,
0.0008281467,
0.0763039067,
-0.0361140892,
0.0227982718,
-0.0355282426,
-0.0416018553,
0.0286077932,
0.0616618805,
0.0043839575,
0.0164498426,
0.0256318059,
0.0451105125,
-0.0283758473,
-0.0065523549,
0.019041419,
0.0080264974,
-0.0451230519,
-0.0057281312,
0.0004226796,
-0.0122878524,
0.0000769499,
-0.0431651883,
0.0140841361,
0.0154156601,
0.0241758097,
0.0301423166,
0.00965147,
-0.0213960614,
-0.0467102118,
0.0217771232,
-0.0105619216,
-0.0232847743,
0.0117060477,
-0.0648586228,
0.0081176143,
0.0547852442,
0.0702227056,
-0.0323334299,
-0.0128251594,
-0.0080562169,
-0.0327449925,
0.0883847326,
0.0195274763,
-0.0142270951,
0.0039739134,
-0.029791398,
0.0040249913,
-0.0189173259,
0.012072443,
0.0262111165,
-0.0153198382,
0.0158342272,
-0.0203925427,
0.0285209268,
0.0344194733,
0.0197640005,
-0.0169116743,
-0.0369877219,
0.0148042776,
0.0012599544,
0.0850346163,
-0.0200221594,
-0.0138797788,
0.0280149747,
0.0105825458,
0.0162656568,
-0.0156072276,
0.0266909245,
-0.0022661923,
0.0015273204,
0.0396056771,
-0.0099314898,
-0.0053771939,
-0.0207402706,
-0.0437225811,
-0.0123035545,
0.0648175776,
-0.0413687788,
0.0405139066,
-0.0046586492,
-0.011539747,
0.004152243,
-0.013739015,
-0.0050912905,
0.0364607014,
0.0026675824,
-0.0107140504,
-0.0010106787,
-0.0330888107,
-0.0063696718,
-0.0731497854,
0.0424042568,
0.0053907083,
-0.0157768652,
0.0125735793,
0.0437251516,
0.0323010907,
-0.0166267455,
-0.0473930575,
-0.0750828087,
-0.0177757517,
0.0053613349,
-0.010393776,
-0.0011285675,
-0.0254990645,
-0.0098511335,
0.0241828244,
-0.0121962726,
-0.005665177,
0.054806862,
-0.0227731466,
0.0253822617,
0.0480315275,
-0.0642978996,
-0.0390206501,
0.0151301064,
-0.0323866084,
0.0112021323,
-0.0047351383,
0.00353882,
-0.057079073,
-0.0041933721,
-0.0420185998,
0.0673165619,
-0.0364730172,
-0.0403836034,
0.035354916,
-0.0187702533,
0.0085384883,
0.0388460346,
0.0413687862,
0.0515843667,
0.0283998884,
-0.0690989494,
-0.0719963014,
0.050898958,
-0.0333619677,
-0.0038203928,
-0.094191961,
-0.0654149726,
0.0515445732,
0.0816300362,
0.0019809441,
0.0115023451,
-0.063333258,
0.042071607,
0.0284472071,
-0.0479001738,
-0.0449683182,
-0.0413524918,
-0.0315232612,
0.0486724675,
0.0045732521,
0.0142382383,
0.0944078341,
-0.0072763059,
0.0068131946,
-0.0295251273,
0.019286776,
-0.0202059727,
0.0866993368,
-0.0140208462,
-0.0115784472,
0.0283191055,
0.0316540077,
-0.0243004225,
-0.0704510659,
0.049208235,
-0.0272389464,
0.018980708,
-0.063908048,
0.0272394996,
0.0343785658,
-0.0329384692,
-0.029003771,
-0.0257345811,
-0.0671878085,
-0.0006811436,
0.033847928,
-0.0079987552,
-0.0511060506,
0.0118226949,
0.0026264049,
0.0013647219,
0.0362915769,
-0.0074721682,
-0.029118998,
0.0471851453,
0.0430029072,
0.0427431911,
-0.0088017378,
-0.0021200604,
0.0120529924,
0.024357127,
-0.0267633423,
-0.009520581,
0.0146320052,
0.0086191827,
0.0285773445,
-0.011614983,
0.0180880949,
0.0535896383,
-0.0305174217,
-0.0261910781,
-0.0497771613,
0.0872394145,
0.0465435907,
0.0330354497,
0.0210363828,
-0.0369784348,
0.0381918252,
-0.0115146944,
0.0298962258,
-0.0902806669,
-0.0270558037,
-0.0054305382,
-0.0246159248,
-0.0530857295,
-0.0448145755,
-0.0123077864,
-0.126058355,
-0.0151938768,
-0.0037715232,
-0.0526797026,
0.0065717385,
0.0137584759,
-0.0065341666,
-0.0295368768,
-0.0253711548,
0.0076641557,
0.0472436771,
0.0140794041,
-0.0000197629,
0.0202940479,
0.0136106228,
-0.0011928781,
-0.0069156336,
0.0467065461,
0.0639451668,
0.0950706899,
0.003133598,
-0.0435666293,
0.0148551557,
0.0489317477,
-0.0218779594,
-0.0087496713,
-0.0230260789,
-0.0312252212,
-0.0057799434,
-0.0003909204,
0.0360258967,
-0.0045223902,
-0.0354000814,
0.0418597646,
0.0118005276,
-0.0196444578,
-0.0262385514,
0.0198252685,
0.0371983126,
0.0717117563,
0.0464065671,
-0.012460731,
-0.0560583994,
-0.0132196965,
-0.0303241517,
-0.007563543,
-0.0455388762,
-0.0611995682,
-0.0759794414,
0.0745250732,
0.0256598927,
-0.0562618487,
-0.0268688779,
-0.0179637279,
0.0107994322,
0.0000637094,
-0.0461332165,
0.0442673638,
0.0445563868,
0.0096455952,
-0.0091421492,
-0.0663214326,
0.0028455139,
-0.0102914404,
-0.0283536706,
0.0023435031,
-0.0606444106,
-0.0486057475,
0.0526766442,
-0.0216700695,
-0.0653660148,
-0.0202566851,
0.0700542107,
0.0225726143,
-0.0007297269,
-0.0040307641,
0.0016767246,
-0.0203857366,
0.0306008514,
-0.0111616226,
0.0201400686,
0.0307257101,
0.0145452265,
0.0142451786,
0.0366720445,
-0.007386446,
-0.0046829549,
0.0055931439,
0.0791079774,
-0.0505389534,
-0.049029123,
0.0238326825,
-0.0142717371,
-0.0014507459,
-0.006050406,
-0.0492962301,
0.006703069,
0.0015648665,
-0.0168369282,
-0.0479263179,
-0.0502852909,
-0.0119890086,
-0.0144038647,
0.0284014214,
0.0164338592,
-0.0416914038,
-0.0153940078,
0.0193541851,
0.0633997694,
-0.0159453452,
0.0690505952,
0.0051237731,
0.0280074161,
-0.0212297589,
0.0530877896,
0.0482570007,
-0.0182155836,
0.0492199063,
0.0210851338,
0.052827362,
0.0626764894,
-0.0468472727,
0.003169646,
-0.0057059419,
-0.0109143872,
-0.0686969161,
0.020700179,
0.0455658697,
0.0579968728,
0.0034863553,
-0.0149311563,
-0.0881649256,
0.0157847349,
0.0296828113,
0.0083884224,
-0.0003071704,
-0.0889539197,
0.0443458892,
0.0379588008,
-0.0116433511,
0.0223133992,
0.044011768,
-0.0116373841,
0.0421976037,
-0.018868668,
0.0046935752,
-0.033842355,
-0.0470781922,
0.0058594015,
0.0432272926,
-0.0136837997,
0.0213408899,
-0.0822060257,
-0.0327750072,
-0.0195006896,
0.0602353439,
0.0094596259,
0.0100887921,
0.0611747056,
0.0192889962,
-0.0005135306,
0.0258213077,
-0.0413396843,
0.0474056005,
-0.0493909158,
-0.0581455715,
-0.0423542149,
-6.040536698e-33,
-0.0213538744,
-0.0513928197,
0.0446213707,
0.0005266301,
-0.0180285014,
-0.0446376763,
-0.0434172526,
-0.0007473775,
-0.0071912049,
0.005311856,
-0.0137877762,
0.0067791296,
0.0113403192,
-0.0047557689,
-0.0195449851,
-0.022395486,
-0.0138217863,
-0.0109620867,
-0.0430803299,
-0.0205762293,
-0.0078826286,
-0.0477802046,
0.0281469095,
-0.0242891964,
0.0121073946,
0.0062754503,
-0.0544748679,
0.0605888516,
-0.0281897709,
-0.0016766915,
-0.0130121717,
-0.0010971712,
-0.010007062,
-0.0239111409,
0.0455317907,
0.0213801898,
0.0072504729,
-0.0085621187,
0.0687399805,
-0.0204711389,
-0.0279128682,
-0.0150726726,
-0.0043750252,
0.0085982634,
-0.0405687392,
0.0418663807,
0.0125543559,
-0.0109073892,
0.012154581,
0.0284159388,
-0.0251880307,
-0.0013509487,
-0.057041429,
-0.0354091898,
-0.0035636735,
0.033180479,
-0.0005676411,
-0.0401681736,
-0.0079904692,
0.0642620027,
0.0343216583,
0.0027643386,
0.0222976543,
0.0154922856,
0.058658801,
0.0661448166,
-0.0159760788,
0.004476367,
-0.0185599308,
-0.0290149339,
0.0043454859,
0.030886339,
0.0220404658,
0.0440778024,
-0.0098965336,
-0.0203912593,
-0.0314023346,
-0.0118937325,
0.033258792,
0.0368439928,
0.0190262571,
0.0248102099,
-0.0161476322,
-0.0071593323,
-0.0341735445,
-0.0209901053,
0.0045811371,
-0.013458712,
0.0002435012,
0.0184108596,
-0.0183174405,
-0.0066737523,
0.0258300249,
-0.0425816886,
-0.0448752232,
0.0036869869,
0.0143034253,
0.0114873601,
-0.00580783,
-0.0209738798,
-0.0360665061,
-0.0372039676,
0.0069164229,
-0.0180772115,
0.0148990527,
0.0259135384,
-0.0645382702,
0.0002409075,
-0.0121730734,
0.0158912688,
-0.0239596274,
0.0532125607,
-0.0294611957,
-0.023108365,
0.0153101757,
0.018353004,
-0.0189376213,
-0.0164248757,
-0.006405714,
-0.0735623166,
-0.0196012948,
0.0509499349,
-0.0311750192,
0.0172988661,
-0.0101046013,
0.0186092537,
-0.0159262549,
0.0149813537,
0.0063713249,
-0.0106789134,
-0.0112073524,
-0.034274023,
2.939e-7,
-0.0140787298,
0.0478545837,
0.0268481262,
0.0206264984,
0.0051927259,
0.0735146105,
-0.0301811509,
0.0103255492,
0.0068272385,
-0.064170137,
0.0472203717,
0.0503312238,
0.0396343134,
0.000307592,
-0.0959360227,
0.0131318104,
-0.0662989467,
-0.0505957492,
-0.0100117484,
-0.0345659964,
0.0410949178,
0.0279478207,
0.0472069643,
0.0115649486,
-0.0180594437,
-0.0264736805,
0.0326008722,
-0.0528798662,
0.0247918833,
0.0554607585,
-0.0347817354,
-0.0540076941,
-0.0089008464,
-0.0248128548,
0.0181370564,
0.0532453693,
-0.0364871323,
0.0679390132,
0.0320908055,
-0.0348522477,
-0.0708291605,
-0.0158062149,
-0.0090847937,
-0.0372602344,
0.0196565706,
-0.0559468083,
-0.0207118597,
-0.0035060046,
0.0165082999,
-0.0119437696,
-0.0386286415,
-0.0180992875,
-0.0396024138,
0.0297978465,
0.0027401643,
0.0276445858,
0.0561297722,
-0.023315968,
-0.0111748818,
-0.0090965573,
-0.0427222028,
-0.0086000813,
-0.0047262171,
0.0450222306,
-0.0155719416,
-0.0427425802,
-0.0219570436,
2.428433805e-34,
0.0299280733,
0.0002270411,
0.0084467921,
0.0433905162,
0.0781503171,
0.0348850414,
-0.0244116131,
0.0197975021,
0.0703999996,
-0.0452373438,
-0.0241170507
] |
http://arxiv.org/abs/2401.13835v1 | http://arxiv.org/pdf/2401.13835v1 | The Calibration Gap between Model and Human Confidence in Large Language Models | 2024-01-24T22:21:04 | 2024-01-24T22:21:04 | Mark Steyvers, Heliodoro Tejeda, Aakriti Kumar, Catarina Belem, Sheer Karny, Xinyue Hu, Lukas Mayer, Padhraic Smyth | cs.LG | cs.LG, cs.AI, cs.CL, cs.HC | For large language models (LLMs) to be trusted by humans they need to be
well-calibrated in the sense that they can accurately assess and communicate
how likely it is that their predictions are correct. Recent work has focused on
the quality of internal LLM confidence assessments, but the question remains of
how well LLMs can communicate this internal model confidence to human users.
This paper explores the disparity between external human confidence in an LLM's
responses and the internal confidence of the model. Through experiments
involving multiple-choice questions, we systematically examine human users'
ability to discern the reliability of LLM outputs. Our study focuses on two key
areas: (1) assessing users' perception of true LLM confidence and (2)
investigating the impact of tailored explanations on this perception. The
research highlights that default explanations from LLMs often lead to user
overestimation of both the model's confidence and its' accuracy. By modifying
the explanations to more accurately reflect the LLM's internal confidence, we
observe a significant shift in user perception, aligning it more closely with
the model's actual confidence levels. This adjustment in explanatory approach
demonstrates potential for enhancing user trust and accuracy in assessing LLM
outputs. The findings underscore the importance of transparent communication of
confidence levels in LLMs, particularly in high-stakes applications where
understanding the reliability of AI-generated information is essential. | The Calibration Gap between Model and Human Confidence in Large Language Models: For large language models (LLMs) to be trusted by humans they need to be
well-calibrated in the sense that they can accurately assess and communicate
how likely it is that their predictions are correct. Recent work has focused on
the quality of internal LLM confidence assessments, but the question remains of
how well LLMs can communicate this internal model confidence to human users.
This paper explores the disparity between external human confidence in an LLM's
responses and the internal confidence of the model. Through experiments
involving multiple-choice questions, we systematically examine human users'
ability to discern the reliability of LLM outputs. Our study focuses on two key
areas: (1) assessing users' perception of true LLM confidence and (2)
investigating the impact of tailored explanations on this perception. The
research highlights that default explanations from LLMs often lead to user
overestimation of both the model's confidence and its' accuracy. By modifying
the explanations to more accurately reflect the LLM's internal confidence, we
observe a significant shift in user perception, aligning it more closely with
the model's actual confidence levels. This adjustment in explanatory approach
demonstrates potential for enhancing user trust and accuracy in assessing LLM
outputs. The findings underscore the importance of transparent communication of
confidence levels in LLMs, particularly in high-stakes applications where
understanding the reliability of AI-generated information is essential. | [
0.0336391181,
0.058422491,
-0.0376564413,
0.0380565561,
-0.0232829079,
-0.0167062972,
-0.0062643248,
-0.0228178352,
0.0040417574,
-0.0245736018,
-0.0196414683,
-0.0442305543,
0.0576324016,
0.0072730342,
0.0327323042,
-0.0319681317,
0.0055015204,
-0.0412685797,
0.0212310515,
-0.0002939079,
0.0113372421,
0.0151922265,
0.0036132615,
0.0634657443,
0.009311744,
-0.011240595,
-0.0108019589,
0.0032445791,
-0.0098904753,
-0.0558440425,
-0.0030956618,
0.0629723966,
0.0096912822,
0.0290891547,
0.0000018763,
-0.0479218289,
0.0124631766,
0.0074473154,
-0.0785917416,
0.0537992008,
0.0167491082,
0.0313657895,
0.0170871075,
0.0274410192,
-0.025001945,
0.0670361966,
0.0474236459,
-0.0123522813,
0.0144508434,
0.0134898126,
0.0104999607,
0.0220480654,
-0.0029996722,
0.0094594685,
-0.0151491407,
-0.0472571366,
-0.0043715318,
0.011117992,
-0.0073259384,
0.0152781382,
-0.0017737634,
0.0328061134,
0.0225974862,
-0.0220976137,
-0.0063152639,
0.0213233344,
-0.0146312779,
-0.0654472858,
-0.0124380738,
0.0360102616,
0.0630241334,
-0.0276904497,
-0.0330947898,
-0.0296652671,
0.0171230156,
0.041090779,
-0.0370296873,
0.0261511765,
-0.0040079779,
0.0026543348,
-0.0373725779,
0.0131223891,
0.0397754498,
0.0163716022,
0.0121825961,
0.0830842629,
-0.0026737112,
-0.0077991909,
0.0023433862,
0.0412234701,
0.0195800923,
-0.0682485625,
0.0106049208,
0.0340578035,
-0.038728483,
0.0481597856,
-0.0070454017,
0.0137264505,
0.0400240459,
-0.0596683547,
0.045455277,
0.0386224985,
-0.0093837483,
0.0017982934,
0.0227215625,
0.0654565468,
0.0054638525,
0.0221083853,
0.0015634302,
0.023701001,
-0.0189680494,
-0.0083847381,
-0.0019371138,
-0.0428967662,
0.0149417501,
0.0084137861,
-0.0543395467,
-0.0025245331,
0.0021834029,
0.0435234755,
-0.008677775,
0.0053015724,
-0.0192676745,
-0.0359111391,
-0.0646749735,
0.0409635045,
-0.0321962945,
0.0292560253,
0.0240574609,
-0.0926591456,
0.0129734445,
0.015890304,
-0.0298339054,
0.0035434654,
-0.0179906152,
0.0495533459,
0.0068978705,
0.0064855022,
-0.0038092164,
-0.0545108058,
-0.0219894517,
-0.0682299361,
-0.0283137616,
-0.0458622053,
-0.0343351848,
0.0514497422,
0.0037070855,
-0.0414494388,
0.0129388915,
0.0138583779,
-0.0023220051,
0.0654342175,
-0.0481374487,
0.0350802019,
0.0179566834,
0.0241362564,
-0.0345640779,
0.0459954329,
0.0585030839,
0.0448763967,
0.0663295016,
-0.0147971306,
0.0678945705,
0.0271992665,
-0.0133549701,
0.0061080819,
-0.0329941958,
0.026170373,
-0.0365412384,
0.0374234654,
0.0146960691,
0.0729309916,
-0.0102987615,
0.0219535846,
0.1099419966,
0.04860029,
-0.0061665936,
-0.0108227935,
0.0133048026,
-0.0040106387,
0.0427258275,
0.0251001641,
-0.0434337966,
0.0175404623,
-0.0267301314,
0.0073623499,
0.0307921637,
0.0111721056,
0.0086963903,
-0.0707885623,
-0.0465529971,
-0.0018758528,
0.0184654221,
0.0652805045,
0.0370047763,
-0.049095612,
-0.0191309322,
0.0070547061,
0.0025312705,
-0.0296586268,
0.0109970998,
-0.0833357647,
0.0586662889,
0.0752429962,
0.0089857522,
0.0434400626,
0.0196410008,
-0.0072411094,
-0.0009937292,
0.0458200388,
0.0734240264,
-0.0076150093,
-0.0174790826,
-0.0272199288,
-0.0284924172,
-0.010128526,
0.013905257,
0.011042472,
0.0301084239,
-0.0060285614,
0.0075599723,
-0.0455265529,
-0.0726571605,
0.0287393313,
-0.045057673,
-0.0217723176,
-0.0485006794,
-0.0154893398,
-0.0010438465,
0.0211449228,
0.0053156605,
0.0340032727,
0.0074644391,
0.0515243076,
-0.0181585569,
-0.0015487435,
-0.0206980146,
0.0677060261,
0.0496124253,
-0.0039545591,
-0.0260064192,
-0.0099132787,
-0.0000273776,
-0.0027238559,
0.0034666255,
-0.0498308092,
0.0038829502,
-0.0234760977,
0.0255671125,
-0.0386656336,
0.0050565354,
-0.0340946354,
0.0039885226,
0.0061353259,
0.0080545805,
0.0101832217,
-0.0148478793,
0.0238162745,
-0.0227849465,
-0.0146330884,
0.041519247,
0.0059610731,
0.0273588691,
0.0537354648,
-0.0319955535,
-0.0715365335,
0.0130314548,
-0.0699934959,
-0.0218685027,
-0.011339142,
0.0217988789,
0.0129580367,
-0.0334638432,
-0.0199098997,
-0.0072450726,
-0.0330792181,
-0.025055524,
-0.0010653802,
0.0251969919,
0.0396676026,
-0.0092273122,
-0.0271221828,
-0.0506655239,
-0.0009998498,
-0.0197094679,
0.0427678451,
0.0795962662,
0.0137444651,
-0.0869627073,
0.0137805799,
-0.0309264287,
-0.0258738827,
0.0046963794,
-0.0386068784,
0.0532373451,
0.0313469507,
-0.0267593469,
0.0256643854,
-0.0030412867,
0.0133670503,
0.0705083013,
0.0089481222,
-0.0548006035,
-0.0394780822,
-0.0052384636,
-0.0704051033,
-0.027520664,
-0.065387696,
-0.0150564918,
0.0595879778,
0.0241051298,
0.0158829205,
-0.0078649949,
0.0415775739,
-0.0439974666,
-0.052202858,
-0.0240632519,
-0.051299721,
-0.0371377394,
0.1156911105,
0.0034118833,
-0.0928916708,
0.0710042268,
-0.0499371812,
0.0139681417,
-0.0008476556,
-0.0166129619,
-0.0046855151,
0.0485755093,
-0.0303221494,
0.0052347742,
0.0214559473,
0.0011705701,
-0.0479325429,
-0.0490327105,
0.0281899832,
-0.0608216785,
0.0255585797,
-0.0742304549,
-0.0208407361,
0.0204453003,
-0.0076419977,
-0.0414944105,
-0.0335090347,
-0.0848043337,
-0.0147517957,
0.0404706709,
-0.0537023321,
0.0104097007,
-0.0264102295,
0.054553587,
-0.0200010445,
-0.025467081,
-0.0276165977,
0.0116411261,
-0.0127042094,
0.030697003,
0.035360463,
0.0777210519,
-0.0757157877,
0.006486658,
0.0462907515,
0.0281654298,
0.0375443287,
-0.0390001647,
-0.0359904841,
0.0332411267,
0.0059074713,
-0.0114935599,
0.0721301958,
0.0745191574,
-0.0401567891,
0.0161348898,
0.1106325015,
0.0210652258,
-0.015952507,
0.0105887661,
-0.0530293286,
0.0024689934,
0.031050656,
0.0238145776,
-0.0599492677,
-0.0065038097,
0.0205837432,
0.032471139,
-0.034431126,
-0.0108978078,
-0.022004392,
-0.1034883261,
0.0390776582,
-0.0028954013,
-0.0058593,
0.0100791277,
-0.0743206888,
0.0076821698,
0.0013184135,
0.0701856315,
0.01018898,
0.0077600149,
0.0227666292,
0.0285101011,
-0.069898881,
-0.0547275208,
0.0266580824,
-0.0486676842,
0.0631821379,
0.0548526086,
0.017097244,
-0.0244150665,
-0.0560187809,
0.0191610456,
0.0245718695,
0.0096002808,
0.0146216815,
0.0440925136,
0.0059459168,
0.0162865855,
-0.0397875421,
-0.0168935545,
-0.0177284516,
-0.008964628,
0.0047604274,
0.0306839962,
-0.0004341215,
-0.0431730077,
-0.0077093784,
0.0017630552,
0.0024457418,
0.0701912716,
0.0461892374,
-0.0367895886,
0.0080463141,
-0.0458482616,
0.0007917691,
-0.0003516292,
-0.0500379726,
0.0427944697,
0.0177206323,
0.0063776835,
-0.0361245535,
-0.0411907211,
0.0287077986,
-0.0034226803,
0.0168383662,
-0.0027789366,
0.0148092061,
0.0137653062,
-0.0036886565,
0.0283506215,
-0.0308244955,
0.0585096814,
-0.0264489017,
0.0465287529,
0.0032132375,
-0.020953292,
0.0185450278,
0.0377959348,
-0.0329956897,
-0.0636272952,
-0.0057791574,
0.0429022536,
0.0359765999,
-0.0054997504,
0.008636266,
-0.0957815498,
0.0177766588,
-0.0066086729,
-0.0423249863,
-0.0142106777,
0.0112957107,
0.02633119,
-0.0084627895,
0.0321099684,
-0.0169526339,
-0.0265444666,
0.0059013106,
0.0118008573,
-0.0258681625,
-0.0816408992,
0.0309366584,
0.0521717109,
0.0589913018,
-0.0447632931,
-0.0003403465,
-0.0002383518,
-0.0279506892,
-0.0435942039,
-0.020825047,
-0.0024975659,
-0.0167529732,
0.0272331331,
-0.0268734824,
-0.0024441606,
-0.037886586,
-0.0426473767,
0.0092159174,
-0.023496937,
0.0177232325,
0.0156959817,
0.0330771357,
0.0237777922,
0.0091753518,
-0.0232087243,
0.0123884948,
-0.019734554,
0.0504500791,
0.0644507036,
0.0285236072,
-0.0062186751,
-0.0207024608,
0.0113282138,
-0.0110744881,
-0.0576677732,
-0.0734676868,
-0.0083861323,
0.014215203,
0.0057518817,
0.0160089508,
0.0326117836,
-0.0190082453,
0.0051264148,
-0.0168558229,
0.0013638401,
-0.0116998153,
-0.0027298869,
0.0351526961,
-0.0051441737,
-0.0597068705,
-0.0063482751,
0.058733806,
-0.0109156454,
0.0311620049,
-0.0621398278,
-0.0507415235,
-0.0054175975,
-0.0173836257,
0.0122154281,
0.0138060581,
-0.0317978896,
-0.0310247727,
-0.06959939,
-0.0843061358,
-0.0068793246,
0.0459863804,
0.0138469981,
-0.0506853014,
-0.0110179372,
0.0752685592,
-0.0135329198,
0.0063759587,
-0.0404254496,
-0.010902185,
-0.0318118595,
0.0329484269,
-0.0488789603,
-6.234638366e-33,
-0.0081993109,
-0.0756767318,
0.0017291431,
-0.0169341844,
0.0117330179,
0.0044621131,
-0.0437051058,
0.0658107027,
-0.0160490423,
0.0320873894,
-0.0207520071,
0.0013380779,
0.0228076372,
-0.0348105878,
0.0535767414,
-0.0191477723,
0.0314274132,
-0.0035464226,
-0.0480212681,
-0.020052826,
0.0053026215,
-0.0161796287,
0.0655180737,
-0.0934459791,
0.1012596488,
0.0066964314,
-0.0276878607,
0.0171938501,
-0.0139908129,
0.046361018,
-0.009147862,
0.042960763,
0.0077884095,
-0.041337084,
-0.0202668067,
0.0206850823,
0.0206926055,
-0.0227753762,
0.0382367782,
0.0258524437,
-0.0308022611,
-0.0155783761,
0.0371811204,
-0.0204183832,
-0.0661519989,
0.0260864049,
0.0263152886,
-0.025307605,
0.0005162535,
-0.0054382887,
-0.0119935395,
0.0188005567,
-0.0065838741,
0.0043396186,
-0.0144617911,
0.0614454262,
-0.0086078634,
-0.0192434937,
-0.0402422398,
0.0274645425,
0.0383302122,
0.0013156938,
0.1105511039,
0.016330108,
-0.0161376316,
0.0557109155,
0.0304299872,
0.0238133594,
-0.0504022837,
-0.0415394418,
0.0291943271,
0.0261675604,
0.0701066256,
0.0851116478,
0.0118367588,
0.0013344445,
-0.0282950923,
0.0390353762,
-0.0027189937,
0.0217620898,
0.0532849804,
0.0130094392,
-0.0936454013,
-0.0378799513,
-0.0569430031,
-0.0396306589,
-0.0351358689,
-0.0467325933,
0.0192031376,
-0.017842086,
-0.0023723794,
-0.003918122,
-0.0074925488,
-0.040530622,
-0.0365858786,
-0.0416299365,
-0.0132172825,
0.0159756318,
0.0046856459,
0.0264142361,
-0.0294566639,
0.0426549911,
-0.0318829976,
0.0323948227,
0.0365230665,
0.0235805623,
-0.0091054086,
0.0094006686,
-0.0478313342,
0.0105046984,
-0.0466802642,
-0.0248073023,
0.0428303964,
-0.0301003419,
0.0389135927,
-0.006447528,
-0.0443456732,
0.0232090708,
-0.0347302891,
0.016378466,
-0.0062530544,
-0.0183269493,
-0.0055031688,
-0.005614426,
-0.0270724427,
-0.0481173024,
-0.0316288285,
-0.0011787298,
0.0541589633,
-0.0049597146,
-0.027306553,
-0.028637439,
2.601e-7,
0.0431113876,
0.0618941858,
-0.0096789571,
0.016304357,
-0.0022185948,
0.0413474068,
-0.0256416965,
-0.0116937729,
0.0226772204,
-0.0721534789,
0.0099164965,
0.0062203961,
0.0659067035,
0.0540005416,
-0.0144157279,
-0.0391367972,
-0.0561546087,
0.0404959619,
-0.0182506777,
0.0422877036,
0.0487458706,
0.0348905735,
0.0131379347,
-0.0053744125,
-0.0379663892,
-0.0361322388,
0.0378818214,
-0.0516477376,
-0.0012383921,
0.0252997875,
0.0696009099,
0.034296453,
0.0260113906,
0.0124564199,
-0.0136075979,
-0.0265460517,
-0.0129225887,
0.0545562729,
0.0020199027,
0.0139282709,
-0.0249514394,
-0.0254433174,
0.0294425953,
-0.0735500306,
0.0252135806,
-0.0501167253,
-0.0360612981,
-0.0421976633,
0.0172808152,
-0.0760457441,
0.0553432815,
-0.0342667811,
0.0080379685,
0.0285162013,
0.0007311185,
0.0119267963,
-0.0218470115,
-0.0186466128,
-0.0399270281,
-0.0180424694,
-0.0182691384,
-0.0558389388,
-0.0172662456,
-0.0144491494,
-0.0128993466,
-0.0553504191,
0.0429157987,
2.026149103e-34,
0.0478085466,
0.0091995168,
0.0070899399,
0.0102871209,
0.0833744332,
0.0144000603,
0.0075745247,
0.0044742101,
0.0477761887,
-0.0588598996,
-0.0267376062
] |
http://arxiv.org/abs/2401.13810v1 | http://arxiv.org/pdf/2401.13810v1 | Automated Root Causing of Cloud Incidents using In-Context Learning with GPT-4 | 2024-01-24T21:02:07 | 2024-01-24T21:02:07 | Xuchao Zhang, Supriyo Ghosh, Chetan Bansal, Rujia Wang, Minghua Ma, Yu Kang, Saravan Rajmohan | cs.CL | cs.CL, cs.SE | Root Cause Analysis (RCA) plays a pivotal role in the incident diagnosis
process for cloud services, requiring on-call engineers to identify the primary
issues and implement corrective actions to prevent future recurrences.
Improving the incident RCA process is vital for minimizing service downtime,
customer impact and manual toil. Recent advances in artificial intelligence
have introduced state-of-the-art Large Language Models (LLMs) like GPT-4, which
have proven effective in tackling various AIOps problems, ranging from code
authoring to incident management. Nonetheless, the GPT-4 model's immense size
presents challenges when trying to fine-tune it on user data because of the
significant GPU resource demand and the necessity for continuous model
fine-tuning with the emergence of new data. To address the high cost of
fine-tuning LLM, we propose an in-context learning approach for automated root
causing, which eliminates the need for fine-tuning. We conduct extensive study
over 100,000 production incidents, comparing several large language models
using multiple metrics. The results reveal that our in-context learning
approach outperforms the previous fine-tuned large language models such as
GPT-3 by an average of 24.8\% across all metrics, with an impressive 49.7\%
improvement over the zero-shot model. Moreover, human evaluation involving
actual incident owners demonstrates its superiority over the fine-tuned model,
achieving a 43.5\% improvement in correctness and an 8.7\% enhancement in
readability. The impressive results demonstrate the viability of utilizing a
vanilla GPT model for the RCA task, thereby avoiding the high computational and
maintenance costs associated with a fine-tuned model. | Automated Root Causing of Cloud Incidents using In-Context Learning with GPT-4: Root Cause Analysis (RCA) plays a pivotal role in the incident diagnosis
process for cloud services, requiring on-call engineers to identify the primary
issues and implement corrective actions to prevent future recurrences.
Improving the incident RCA process is vital for minimizing service downtime,
customer impact and manual toil. Recent advances in artificial intelligence
have introduced state-of-the-art Large Language Models (LLMs) like GPT-4, which
have proven effective in tackling various AIOps problems, ranging from code
authoring to incident management. Nonetheless, the GPT-4 model's immense size
presents challenges when trying to fine-tune it on user data because of the
significant GPU resource demand and the necessity for continuous model
fine-tuning with the emergence of new data. To address the high cost of
fine-tuning LLM, we propose an in-context learning approach for automated root
causing, which eliminates the need for fine-tuning. We conduct extensive study
over 100,000 production incidents, comparing several large language models
using multiple metrics. The results reveal that our in-context learning
approach outperforms the previous fine-tuned large language models such as
GPT-3 by an average of 24.8\% across all metrics, with an impressive 49.7\%
improvement over the zero-shot model. Moreover, human evaluation involving
actual incident owners demonstrates its superiority over the fine-tuned model,
achieving a 43.5\% improvement in correctness and an 8.7\% enhancement in
readability. The impressive results demonstrate the viability of utilizing a
vanilla GPT model for the RCA task, thereby avoiding the high computational and
maintenance costs associated with a fine-tuned model. | [
0.019936176,
0.0267758258,
-0.0123051982,
0.0139215095,
-0.0413430333,
-0.0035959797,
-0.0174663309,
-0.0098689012,
-0.0025719786,
-0.0329521596,
0.0940284058,
-0.0179524776,
-0.0401640907,
0.0756444857,
0.0182290878,
0.0067582866,
0.030320717,
-0.0102634225,
0.0150570096,
0.0208126549,
0.0018847443,
0.0373754017,
0.0042531635,
0.0262086708,
-0.0452764146,
-0.0211131983,
-0.0074522072,
0.0280272719,
0.0238786452,
0.0177124329,
0.0089156432,
-0.0347341262,
0.0267691538,
0.0507801436,
0.000002361,
-0.0507762395,
-0.0410053506,
0.0157289691,
-0.0553567968,
0.0380218811,
-0.0278670117,
-0.0060537755,
-0.047133863,
0.0624187291,
-0.0205175057,
0.0267748367,
0.0445078313,
0.055060599,
0.0750138238,
0.0483440794,
-0.0280241109,
-0.0207381956,
0.01088499,
0.0161037594,
-0.0299001709,
-0.0503329076,
0.0128710857,
0.0151951117,
0.0027765087,
-0.0501103848,
-0.0100844819,
-0.0360142104,
0.0307258125,
0.002511553,
-0.0235562399,
0.057850413,
-0.0067180381,
0.023445921,
0.0014423018,
0.0099329334,
-0.030861292,
-0.0116887921,
0.0061649466,
-0.0227528997,
-0.0432469919,
-0.0355727598,
0.00474633,
0.0234906282,
-0.0118581476,
0.0110945599,
0.0417780355,
-0.0085189613,
0.023103727,
-0.0064702765,
-0.0284676626,
0.0683936924,
-0.020352779,
-0.0177905913,
-0.0379027314,
-0.0265538171,
-0.0009174596,
-0.0055703353,
-0.0274032485,
0.0261527561,
-0.0796686783,
0.0060512605,
0.0106167495,
0.0021015615,
0.023285415,
-0.0338471867,
0.0126434583,
0.0222513843,
0.0055868952,
-0.0016599583,
0.0095649855,
0.0623355731,
0.0015444966,
-0.0150264706,
0.0276674516,
-0.0138407387,
-0.0236722119,
-0.011458193,
-0.018234577,
-0.0069240327,
0.0187226981,
-0.0223993585,
-0.0129465191,
0.0534796603,
0.0042869695,
0.0480220541,
-0.0069719069,
-0.0419258513,
-0.0353489555,
0.0214109328,
0.016443396,
-0.0417512171,
-0.0386392996,
0.0214617439,
-0.0496907607,
-0.0420929044,
-0.0155688273,
0.0017506123,
-0.0125138229,
0.023006063,
0.0185943786,
0.0426973328,
0.0419918001,
0.0245097876,
-0.0493724495,
-0.0615986623,
-0.0014811348,
-0.027849175,
-0.1064323112,
-0.0367975272,
-0.0703385994,
0.0044288328,
0.0112711852,
-0.008816462,
-0.0098550972,
-0.0396743827,
-0.0355654731,
0.038003318,
-0.0859678239,
0.0331531093,
0.038022697,
0.0236337893,
-0.0208022222,
0.0391374901,
0.0287887305,
-0.0040468872,
-0.0006394412,
-0.0522974394,
0.1082309186,
0.0114888288,
0.0071733803,
0.0157180242,
0.009172095,
-0.0065102368,
-0.0205305666,
0.0293008145,
0.0351771377,
-0.0042482675,
0.0359960608,
0.0114209987,
0.0262338687,
0.0242663976,
-0.0747161061,
0.040757753,
0.0455863699,
0.0246769805,
-0.0644618571,
-0.0367966592,
-0.0174315739,
0.0024561672,
-0.0432610996,
-0.0007839534,
0.004196119,
0.0277657062,
-0.0063852784,
-0.0285737924,
-0.0066269445,
0.0068435264,
-0.078615129,
0.0036058095,
0.0365799814,
-0.0203097202,
-0.0182493441,
-0.0153703569,
-0.0218791571,
-0.0118224174,
-0.0416574515,
-0.0521776937,
0.0249091648,
0.0586346351,
0.0379979573,
-0.0581343994,
-0.1073726341,
-0.0549203642,
-0.0056276298,
0.0000858774,
-0.0287837535,
0.0085682729,
-0.0007898201,
-0.0370461568,
0.0092448015,
0.0199959632,
0.0253566559,
-0.0051479712,
-0.0255203974,
-0.0165865868,
-0.0103006111,
-0.0092657572,
-0.0784867927,
0.0199453644,
-0.0082748476,
-0.017741194,
0.0143345958,
0.0052077482,
0.0444385484,
-0.0334921256,
-0.0124931298,
-0.0104016615,
0.0068861833,
0.0099793132,
-0.0804900452,
0.0580403656,
-0.0025645767,
0.0337782167,
0.00427404,
-0.0164595731,
0.0148778865,
0.0000400191,
-0.0839205235,
-0.0479805171,
0.0653248355,
-0.0530415215,
-0.0311840083,
-0.0117785605,
-0.0003595454,
-0.007441395,
-0.0075624082,
-0.0203214232,
0.0454666317,
0.0531239882,
0.0021816955,
-0.0427375324,
0.0420049839,
0.0470049754,
0.000252982,
-0.0073509035,
-0.0048117083,
-0.0628878921,
0.0308471508,
0.0413556471,
-0.0085486565,
-0.013747924,
0.0150059005,
-0.0718170255,
0.0032454631,
-0.0075951912,
-0.0191204995,
0.0091025382,
0.0403319299,
-0.0415795594,
-0.0133846803,
0.0131918946,
-0.015178821,
0.0124586681,
0.0058140452,
0.0448082685,
-0.0143706398,
-0.0555061921,
-0.0084889885,
-0.0000600638,
-0.0617235415,
0.0329369679,
0.0364233628,
0.0092460969,
-0.0641091242,
0.0835098028,
-0.0064909295,
-0.0116961394,
-0.0267092846,
-0.0530127734,
-0.0183760598,
-0.0506099984,
0.0249346755,
0.0859055966,
-0.0313607343,
0.0357470103,
-0.0165841039,
0.0113319512,
-0.0455427282,
-0.0001326813,
-0.0152329244,
-0.0579257384,
0.0026069209,
-0.0076620029,
0.0118032955,
0.055722978,
-0.0014056707,
-0.0488567501,
-0.0545591265,
0.0112832785,
-0.0480137058,
0.0374725051,
-0.0407770723,
-0.0981799886,
-0.0200457312,
0.0588719808,
0.0106971962,
-0.0858909413,
0.0288045071,
0.0065976297,
0.019693803,
-0.0820818394,
-0.0600962825,
-0.0434818491,
0.0389624797,
-0.0017360433,
0.0316209421,
-0.0255696997,
-0.0554720759,
0.0142998891,
0.0202549081,
0.0309606493,
-0.0102306632,
-0.0115901157,
-0.0802617744,
0.0353564359,
0.0245929584,
-0.0381517224,
0.008810591,
-0.0232015625,
-0.0256995,
-0.005090998,
-0.0136707257,
-0.0114795351,
0.0190207511,
0.0085068755,
0.0349550173,
0.071246244,
0.0604772866,
-0.0420427099,
-0.0092719989,
-0.0629685894,
-0.0181930922,
-0.0384871364,
0.0184984729,
0.0184819233,
0.0421554781,
-0.0114726443,
0.1430786103,
-0.0339842066,
-0.0466170385,
0.0059847496,
-0.019787794,
-0.0272827279,
-0.0162287951,
0.0532125868,
-0.0039898199,
-0.065456003,
-0.0349739082,
0.0612182468,
0.0197326038,
-0.0320204273,
0.0092141572,
0.0005563559,
-0.0310313664,
0.0485960059,
0.0386815593,
-0.0562776476,
0.0016574998,
0.0120029179,
0.0294995178,
-0.0200449042,
-0.020686429,
0.0123008424,
-0.0812142864,
0.0460629612,
-0.0169982649,
-0.0080264686,
-0.042017065,
-0.0610173568,
0.0017395911,
-0.0381863378,
-0.0126292603,
-0.0258634575,
0.0340412334,
-0.001001966,
0.02745598,
-0.0624756962,
-0.0698374361,
0.0278020035,
-0.0563309118,
0.0221623331,
0.0617883317,
-0.0126240971,
-0.0117368177,
0.0557591803,
0.0483910292,
0.0397384167,
0.0223341826,
-0.015422835,
0.0410672501,
0.03007208,
-0.0004525768,
0.0177649241,
0.0603114255,
-0.0886865556,
-0.0046594939,
0.0239836387,
0.1271485686,
-0.043213401,
0.0171385761,
0.0084548434,
-0.0298659578,
0.0454413332,
0.0328704342,
0.0239598323,
-0.035944283,
-0.056048654,
-0.0091642216,
0.0297330339,
-0.0051431688,
-0.0076258034,
0.0056808358,
0.0340950154,
-0.037759956,
-0.0140361637,
-0.0092331842,
0.0266143065,
0.0419614539,
0.0307974275,
-0.002439464,
-0.0016853201,
0.0252316017,
-0.0122312242,
-0.0008534513,
-0.0129061267,
0.0583197773,
-0.0506658219,
0.0542905629,
0.0492686071,
0.0512977019,
-0.0325493515,
0.0036122198,
0.0190587137,
-0.0434486717,
0.0342450179,
0.0368910581,
0.0217414238,
0.0216652695,
-0.0358978808,
-0.0548144877,
0.0127041293,
0.0059837601,
-0.0650682971,
-0.0056154151,
0.0133829797,
-0.0409243852,
0.0003689392,
-0.0137650482,
-0.046393279,
0.0216632225,
-0.0157244671,
-0.0098738698,
-0.0743137226,
0.0172704309,
0.0367990211,
0.0185636114,
0.0080690607,
-0.0260989945,
-0.0256548822,
-0.0014105296,
-0.0010007159,
-0.0041120551,
-0.0111567583,
0.0114523126,
0.0768749043,
-0.0111958953,
-0.0290615745,
0.0050517982,
-0.019672811,
0.0048146448,
-0.0352049805,
-0.010059421,
0.00105612,
0.047852736,
0.0214623623,
0.0637836307,
-0.0039924621,
0.0214623045,
-0.0269965101,
0.0270820055,
0.0553446673,
-0.002188381,
0.0488852412,
0.0332699232,
0.0656547248,
0.0002055334,
0.0007406383,
-0.0256872624,
-0.0519957133,
0.0181339178,
-0.0537758805,
0.0716829747,
-0.0008765481,
0.0116814729,
-0.0314939916,
0.0193961617,
-0.0028238462,
0.0429712608,
-0.0043443535,
0.0124902027,
-0.017815819,
-0.0001996111,
-0.0306166783,
0.0046931743,
0.0984540805,
0.0317863747,
0.0607067198,
-0.0555958673,
-0.0158553179,
-0.022023309,
0.0128519945,
0.002335246,
0.020766126,
-0.0003165742,
-0.0039092568,
-0.0670689493,
-0.0047295373,
-0.0482774079,
-0.0159106608,
0.0059600389,
-0.0456138663,
0.0178960748,
0.1118854135,
0.0113798957,
0.0351521373,
-0.0118858125,
0.0049584215,
-0.037283767,
-0.0273480229,
-0.0061029028,
-6.255496778e-33,
-0.0105151581,
-0.0640672967,
0.0141576789,
0.0463698953,
0.0128562599,
0.0306447744,
-0.0385177322,
0.002476393,
0.0462481454,
0.0237513296,
0.0069692922,
0.0461879782,
0.01215111,
-0.0129289487,
0.0529913194,
0.0260832589,
0.0181948803,
0.0237294585,
-0.0103671635,
0.0938508287,
0.0569914319,
-0.0034908925,
-0.0104301134,
-0.1146812961,
-0.0106927827,
-0.0077053341,
0.0010193923,
0.0350714028,
-0.0118929176,
0.0212489199,
0.0096059022,
-0.0100065982,
0.0217157621,
-0.0556665808,
0.0254211482,
-0.0338735953,
0.0271358155,
0.0103393402,
-0.0187034197,
-0.042421896,
0.0036648621,
-0.0524717346,
-0.011252135,
0.008650545,
0.0032572171,
0.0066764285,
0.0429569483,
0.0044718445,
0.0024735464,
-0.0814194754,
-0.04151031,
0.0153171513,
-0.0740723386,
0.0312751979,
-0.0149068618,
-0.0203687139,
0.0094726253,
0.0498335026,
-0.0454159528,
0.0117490934,
-0.0123563558,
-0.0150152622,
0.027450731,
-0.0137152346,
-0.0616785437,
-0.0146066891,
-0.029493168,
-0.0162564982,
-0.0172386058,
0.0017701034,
-0.0026175214,
0.064752914,
0.0453431495,
0.0526425987,
0.0138781462,
-0.0377430804,
-0.0424483977,
0.0468520336,
0.0279743932,
0.0051233247,
0.0755758584,
-0.0178320482,
-0.0806375444,
-0.0235396605,
-0.0073451083,
0.0055155531,
-0.0386260711,
0.0222290903,
0.0297156777,
0.000802729,
-0.0198483299,
0.0178482179,
0.0470615588,
-0.0315277725,
0.0277982783,
-0.0072475406,
-0.0027105073,
0.000811197,
0.0006199095,
0.0287719388,
-0.0467133895,
-0.0095224818,
-0.0587496869,
0.0255294647,
0.0789291114,
-0.011131499,
0.042677924,
0.0480296426,
-0.038258642,
0.041832339,
-0.0223623067,
0.0222242214,
-0.0700023621,
0.020535918,
-0.0052465978,
0.0335142426,
0.0233196393,
0.0252949949,
0.0565667264,
0.0146659147,
-0.008033203,
-0.0473012626,
-0.0475752726,
-0.0196482521,
-0.0410043597,
-0.0139411259,
0.0168897007,
-0.0076727411,
0.0475376546,
-0.0266507827,
-0.0224822424,
0.0317634605,
3.143e-7,
0.0055078366,
0.0590566695,
0.0316175334,
0.0015861258,
0.024654638,
0.0722446665,
0.0379334204,
-0.0344151556,
0.0021858406,
0.0184355639,
0.0277320016,
-0.0466611087,
0.0438899808,
0.0176536553,
-0.0074133072,
-0.0116266143,
-0.0459115058,
0.0117681641,
-0.1062351763,
0.0086429082,
0.0501833446,
0.0671917722,
0.0243616179,
-0.0567348041,
0.0229501538,
-0.0167938638,
-0.0236601066,
0.0006854026,
0.0292825419,
0.029016519,
-0.0034324981,
0.0866643339,
0.0135914898,
0.0486099869,
-0.0144886598,
-0.0110769682,
-0.043943435,
-0.0048872419,
-0.0051579271,
0.0217779763,
-0.0053804424,
-0.030505551,
0.0309879854,
-0.0917593092,
0.0223063212,
-0.0006438876,
-0.0364666805,
0.0370557234,
0.0060879518,
-0.0300676171,
-0.0229650848,
-0.0124295931,
-0.0088616284,
0.0289930534,
-0.028392585,
-0.0275885761,
0.0308558606,
-0.0628721416,
0.0042177076,
0.0214261599,
-0.0187065769,
-0.0003619731,
0.007426254,
0.0036189356,
0.0696078092,
0.0259592198,
0.0342087969,
3.217564055e-34,
0.0086089857,
0.0031752856,
0.0231708977,
0.0358690582,
0.0216198489,
0.0367194526,
-0.0066311467,
0.0188161526,
0.0279318634,
-0.0024979503,
-0.0028544001
] |
http://arxiv.org/abs/2401.13802v3 | http://arxiv.org/pdf/2401.13802v3 | Investigating the Efficacy of Large Language Models for Code Clone Detection | 2024-01-24T20:43:36 | 2024-01-30T06:10:29 | Mohamad Khajezade, Jie JW Wu, Fatemeh Hendijani Fard, Gema Rodríguez-Pérez, Mohamed Sami Shehata | cs.SE | cs.SE, cs.AI, cs.CL, cs.LG | Large Language Models (LLMs) have demonstrated remarkable success in various
natural language processing and software engineering tasks, such as code
generation. The LLMs are mainly utilized in the prompt-based zero/few-shot
paradigm to guide the model in accomplishing the task. GPT-based models are one
of the popular ones studied for tasks such as code comment generation or test
generation. These tasks are `generative' tasks. However, there is limited
research on the usage of LLMs for `non-generative' tasks such as classification
using the prompt-based paradigm. In this preliminary exploratory study, we
investigated the applicability of LLMs for Code Clone Detection (CCD), a
non-generative task. By building a mono-lingual and cross-lingual CCD dataset
derived from CodeNet, we first investigated two different prompts using ChatGPT
to detect Type-4 code clones in Java-Java and Java-Ruby pairs in a zero-shot
setting. We then conducted an analysis to understand the strengths and
weaknesses of ChatGPT in CCD. ChatGPT surpasses the baselines in cross-language
CCD attaining an F1-score of 0.877 and achieves comparable performance to fully
fine-tuned models for mono-lingual CCD, with an F1-score of 0.878. Also, the
prompt and the difficulty level of the problems has an impact on the
performance of ChatGPT. Finally we provide insights and future directions based
on our initial analysis | Investigating the Efficacy of Large Language Models for Code Clone Detection: Large Language Models (LLMs) have demonstrated remarkable success in various
natural language processing and software engineering tasks, such as code
generation. The LLMs are mainly utilized in the prompt-based zero/few-shot
paradigm to guide the model in accomplishing the task. GPT-based models are one
of the popular ones studied for tasks such as code comment generation or test
generation. These tasks are `generative' tasks. However, there is limited
research on the usage of LLMs for `non-generative' tasks such as classification
using the prompt-based paradigm. In this preliminary exploratory study, we
investigated the applicability of LLMs for Code Clone Detection (CCD), a
non-generative task. By building a mono-lingual and cross-lingual CCD dataset
derived from CodeNet, we first investigated two different prompts using ChatGPT
to detect Type-4 code clones in Java-Java and Java-Ruby pairs in a zero-shot
setting. We then conducted an analysis to understand the strengths and
weaknesses of ChatGPT in CCD. ChatGPT surpasses the baselines in cross-language
CCD attaining an F1-score of 0.877 and achieves comparable performance to fully
fine-tuned models for mono-lingual CCD, with an F1-score of 0.878. Also, the
prompt and the difficulty level of the problems has an impact on the
performance of ChatGPT. Finally we provide insights and future directions based
on our initial analysis | [
0.015285735,
0.0379551984,
-0.0448374934,
0.0583221056,
-0.0546467006,
-0.0059940862,
0.0158489235,
0.0322516039,
0.0123296762,
-0.0297627617,
-0.0270804279,
-0.0057693687,
-0.014930198,
0.0084477151,
0.0403239019,
0.0411463454,
0.0492492504,
-0.0115332156,
-0.0356273837,
0.0167819448,
-0.0257882215,
0.0079697045,
0.0451903045,
0.0671273097,
0.0226955041,
-0.0267734472,
0.0014760444,
0.0169983413,
0.0011357018,
-0.02075045,
0.0104271928,
0.0560529381,
-0.0057228901,
0.087462984,
0.000002044,
-0.0435392149,
-0.0048758606,
0.0054067764,
-0.0029953243,
0.0546145886,
0.0071500731,
0.0174700953,
0.0036844623,
0.0010882644,
-0.008872686,
-0.0459707975,
0.0256159361,
-0.0211930703,
0.0645647943,
0.0421787351,
-0.0008704088,
-0.0063551776,
0.0344709083,
-0.0016300364,
0.0549840964,
-0.0186598189,
-0.0082253683,
0.0551602542,
0.0481565371,
0.0513804443,
0.0251635797,
0.0231239721,
0.0008159601,
0.0131410975,
-0.0211295355,
0.0202505104,
-0.0571448319,
0.0338200964,
0.0205915757,
0.0223311838,
-0.104419753,
-0.0806063712,
0.0387899354,
-0.0168869197,
-0.0070971535,
-0.0109961387,
-0.0190370493,
0.0351903848,
-0.0069391956,
0.0125193251,
-0.0011316601,
-0.0156834684,
0.0017493961,
-0.0271240976,
0.00626521,
-0.0127012767,
0.0039233565,
-0.0044174185,
-0.0292222705,
-0.0323629826,
-0.0204606745,
-0.0605070479,
-0.0446467064,
0.0159496982,
-0.0147492038,
-0.0204551518,
-0.0466881879,
-0.0164507702,
0.0472627915,
-0.0179590862,
0.0045795171,
0.0412254222,
-0.0482988395,
0.0052008717,
-0.0943005532,
0.10808824,
-0.0088339755,
-0.0023715026,
-0.011262035,
-0.0012784528,
-0.0928715765,
-0.0215363409,
-0.0488267355,
0.038065996,
0.0520370603,
0.0196039174,
-0.0184970126,
0.0280772857,
0.0071058744,
0.055566214,
-0.0365700983,
-0.069789432,
-0.0115063954,
0.019272266,
0.0544141307,
-0.0422238931,
0.000670305,
-0.0264332313,
-0.0350673832,
-0.1024220437,
-0.0038504838,
0.0343781225,
-0.0241453405,
-0.0146597531,
0.0026815005,
0.0337495767,
0.0293084942,
0.00940846,
-0.0562040173,
-0.0345172137,
0.0084415311,
-0.0045500197,
-0.0226181913,
-0.0882012323,
-0.0516742915,
-0.014054995,
0.0107293315,
-0.0450579152,
-0.0233537182,
-0.0211599246,
0.0094150119,
0.0344292782,
-0.0401544161,
0.0460335463,
-0.0017894021,
-0.0058862693,
-0.0253628809,
0.0179094989,
-0.0181445237,
-0.0043741455,
0.0080060549,
-0.0290130209,
0.0587616563,
0.0307532605,
-0.0368212871,
0.0178604126,
0.0017156847,
-0.0128417742,
0.0253371168,
-0.0226526912,
0.0270507894,
0.0082419543,
-0.0279391147,
-0.0061555919,
0.0083700875,
0.0458795577,
-0.0007110575,
0.0651631206,
0.0959516615,
0.0620033778,
-0.0272138454,
-0.0029933092,
-0.0145144016,
-0.0064821802,
-0.0390623882,
0.0113354111,
-0.0215225909,
0.0024695843,
-0.0267890915,
0.0003626607,
-0.0134859914,
-0.0123632075,
-0.0560691543,
-0.0402076282,
0.0131469863,
-0.0128223039,
-0.0135568809,
0.0082919514,
-0.0061911652,
-0.0419573896,
-0.0503062531,
-0.0958337039,
0.0506422073,
0.0132549927,
0.0555867702,
-0.0195941459,
0.0030472409,
0.0154672908,
-0.0268941447,
-0.0338890068,
0.0613900162,
-0.0041552717,
-0.0034930066,
-0.0384064279,
-0.0386600047,
-0.0421058498,
0.0003228398,
0.0452795215,
0.0153032709,
0.0058448096,
0.0629664287,
0.0001361038,
-0.050584875,
0.0264305063,
-0.0148507385,
-0.0115226312,
0.0051876688,
-0.0148109328,
0.0232885089,
-0.0292205587,
-0.0516378433,
0.0317319483,
0.0178417582,
-0.0163902435,
-0.0266206674,
-0.0096059004,
0.0043439791,
-0.0429502055,
0.0171364229,
-0.0390907265,
-0.0017693719,
0.0113293072,
-0.0316996984,
-0.0154228909,
0.0256614685,
-0.0297566969,
0.0274409112,
-0.0310864914,
0.0386892147,
-0.0640003979,
-0.0103400461,
-0.0230941139,
0.0537118949,
0.0411689132,
0.023254117,
0.0287362877,
0.0518047921,
0.0816129074,
-0.0094701685,
-0.0627400875,
-0.0262397118,
-0.0349161476,
0.0255035497,
0.007227499,
0.0762593299,
-0.0054099029,
-0.0195728764,
-0.0698153228,
0.0235116668,
0.006531518,
0.0344234034,
0.0346518457,
0.016402036,
-0.0340122916,
-0.0128367059,
-0.0244325753,
0.000125647,
0.0057556503,
0.0001129282,
0.0274067204,
-0.0318696797,
-0.0551791154,
-0.0207467265,
0.020004496,
-0.0198714081,
0.0518742688,
-0.0076891426,
-0.0345532335,
-0.1127670258,
0.0995183438,
-0.0199391376,
0.0240492355,
0.0016004669,
-0.0227250978,
0.0036594411,
-0.0369963497,
0.0600097924,
0.0157236699,
-0.0230878051,
-0.0614651591,
0.0249046367,
0.0348076448,
-0.0601150095,
-0.019902207,
0.02428771,
-0.0771671683,
0.0242964029,
-0.0315935798,
0.0516116992,
0.0471544415,
0.0253704246,
-0.0430670641,
-0.0302165896,
0.0185312573,
0.0109817386,
0.0345751978,
-0.002976062,
-0.0642910898,
-0.0051011397,
0.0364736468,
-0.0146120097,
-0.068654269,
0.0355498791,
-0.0559699349,
0.0138552943,
-0.0326927677,
0.0122310314,
-0.0357842483,
0.0758167058,
0.0071750549,
0.0252862498,
0.0551105626,
-0.0254575126,
-0.0138365189,
0.0284031518,
0.0120460084,
-0.0034925477,
0.0210530534,
-0.0501956083,
0.044978369,
-0.0035907235,
0.0027011908,
-0.0497939363,
-0.0100359954,
-0.0183374938,
0.0026915006,
0.0280549228,
0.0275481306,
0.0444465503,
0.0631095618,
-0.0106472373,
0.0914945677,
0.035062436,
-0.0259498209,
-0.0226796828,
0.0221969932,
-0.0336265638,
-0.0691807941,
0.0027998849,
0.0145687032,
0.0079666506,
-0.0102410149,
0.0138061522,
0.0722819045,
-0.0651920587,
-0.0074441768,
0.0012009302,
-0.0152593106,
-0.0630756915,
0.0448832102,
0.0390576459,
-0.0295589343,
-0.0169252176,
0.0803095922,
0.0200676359,
-0.0504880808,
0.0077336822,
-0.0216510948,
0.0346398391,
0.0533798561,
0.0225360282,
-0.0595621057,
-0.0421227477,
0.0043153232,
-0.0229892991,
-0.0205358006,
-0.0024770775,
-0.0123102516,
-0.1212324575,
0.0216536801,
0.002316975,
-0.0658737347,
-0.0360295065,
0.0037488113,
-0.0254651625,
0.0088639054,
-0.0390686207,
0.0078618973,
-0.0086735971,
0.0217824336,
-0.0043638078,
-0.0515236296,
-0.0220018234,
0.0423798226,
-0.0429009944,
0.0265911501,
0.0708233714,
0.0090390686,
0.0286188219,
-0.0099221831,
0.0888160914,
-0.0510815345,
0.0255165026,
-0.0136773326,
0.0234307013,
0.0346186049,
-0.0157751869,
0.0453800149,
0.012267218,
-0.0401035137,
0.0135823684,
0.033296559,
0.0459553562,
-0.0287898574,
-0.0379643738,
-0.0069159204,
-0.0106606279,
-0.0146139693,
0.0330110006,
0.0977873132,
-0.0058175311,
-0.0196209811,
-0.010870683,
-0.0026723742,
-0.0379152596,
-0.052605696,
0.0003755402,
-0.0181351788,
-0.0097497758,
-0.0601423122,
0.0217040833,
-0.0279692933,
-0.0154829463,
0.0447504781,
-0.0068091527,
-0.0065086922,
0.001899089,
-0.0123704523,
-0.0070934128,
-0.0493650623,
0.0120808352,
-0.0258276779,
0.0290597677,
0.0464797504,
0.0032006647,
-0.0372995846,
-0.016001204,
-0.1154106781,
-0.0318737216,
-0.0899587795,
0.038556803,
0.0174332224,
0.0083572976,
-0.0326138921,
-0.0248554144,
-0.0323638432,
0.0055603981,
0.0025509023,
-0.0154537018,
-0.0105232699,
-0.0283300094,
0.0103104701,
0.0156219192,
-0.0319915302,
0.0338179134,
0.0396026075,
-0.0275577065,
-0.0240515005,
0.0159140117,
0.0537647977,
0.0510872193,
0.038367141,
0.0148159973,
0.0179195069,
0.0041791517,
-0.025498841,
0.0219101328,
-0.0063731559,
-0.0384216234,
0.0312657207,
0.0086331237,
-0.0030876051,
0.0180555005,
-0.0069914106,
-0.0439405479,
0.0756004229,
-0.0090020811,
0.0248305947,
0.0060299938,
0.0877147838,
0.0258974005,
-0.021700222,
-0.0161624346,
0.0088929255,
0.0061760009,
0.007329891,
-0.0269975886,
0.0615075901,
-0.0002287588,
0.0290198121,
0.0064053233,
0.0281730145,
-0.0323739722,
-0.0702510476,
0.009421993,
0.0247376673,
0.0403973274,
0.0239353459,
0.0317563675,
-0.0218169298,
0.0242045838,
-0.0381971598,
0.0657186657,
-0.0198219959,
-0.0301472545,
0.0123306969,
0.0144109484,
-0.0415072031,
-0.0250350777,
0.107354261,
-0.0012782783,
0.0184520967,
-0.0285489187,
-0.020411307,
-0.030612262,
0.0222258493,
0.0760452524,
-0.0291773062,
-0.017071275,
0.0000437275,
-0.0413031168,
-0.0260108989,
-0.0248662252,
0.0476756468,
0.036991246,
0.0368264541,
0.0008784843,
0.0989219025,
-0.0189759228,
-0.0215219129,
-0.0201061126,
0.0130289784,
0.0154930651,
-0.0535698943,
-0.0241969693,
-5.028615692e-33,
0.0017311812,
-0.0056461086,
-0.0136262542,
0.0184841491,
-0.0153081724,
0.0169932414,
-0.0703336149,
0.0203665625,
-0.0185209289,
-0.0414246619,
0.0192056783,
-0.0128534343,
-0.0104322825,
-0.0708516911,
0.0521280169,
0.0257353261,
0.0400175266,
0.0420102887,
-0.0312172174,
0.0501309149,
-0.0200197306,
-0.01282313,
0.0230380222,
-0.0081649963,
-0.0143233892,
-0.0437703468,
0.0440293662,
-0.0186822955,
0.0126405321,
0.0512215793,
0.0019541234,
0.0127982814,
0.0251238625,
-0.0367944166,
-0.0436404459,
-0.044575315,
0.0238397364,
-0.0194194354,
-0.0065962235,
-0.029061541,
0.0514437072,
-0.0522880964,
0.0544446334,
-0.0089277346,
-0.041641511,
0.0341691673,
0.0426245704,
-0.0231658127,
-0.0365558118,
-0.0387055166,
-0.0747743919,
0.0350359231,
-0.0487716533,
-0.0048561608,
0.0043872856,
0.0723706782,
-0.0123751983,
0.0752872601,
0.005195458,
0.0329593159,
0.0182431787,
0.0379738323,
0.0576939359,
-0.0225691479,
0.0167536736,
0.0060629756,
0.0241943281,
-0.0257932618,
0.0283807907,
0.0098361308,
0.048013445,
0.0283453204,
0.0456892736,
0.0513854399,
0.0530730784,
-0.0087122936,
-0.0278373268,
0.0305506848,
0.0221055038,
0.0068155611,
0.023700133,
0.0229223333,
-0.0335470922,
0.011277115,
-0.0182150174,
-0.0541382916,
-0.0271898247,
-0.0305816419,
0.0574771315,
-0.0265693665,
-0.0105986325,
0.0015682849,
0.0512320213,
-0.050031174,
-0.0173483361,
-0.0024810908,
-0.0590388253,
0.0194129404,
-0.0258396156,
0.0114825154,
-0.0237076115,
-0.048396986,
0.0159196388,
0.0000065738,
0.0803241581,
0.0077659059,
0.0554703139,
0.0082737114,
-0.0753517002,
0.0117773125,
-0.0355369374,
-0.0286707878,
-0.0417274572,
0.018128721,
-0.0199876465,
-0.0040996652,
0.0403395742,
0.0024347084,
0.0348205976,
-0.0674017966,
-0.0410468392,
0.0211114697,
-0.0109946104,
0.0244358629,
-0.0208318215,
0.0078688096,
-0.0354747325,
0.0613380112,
0.0220302492,
-0.0201564543,
-0.0472522527,
0.0197805725,
2.725e-7,
-0.0141320359,
0.0065912502,
-0.0004548696,
-0.0425184891,
0.031804245,
0.0402142256,
-0.0328535251,
0.0031290704,
0.0086577516,
0.0152180539,
-0.002313839,
-0.0083298394,
0.0285186302,
0.0135470331,
-0.0618099533,
0.0556520261,
-0.0597184524,
0.0004206494,
-0.0842798501,
0.0288925041,
0.0832441598,
0.0725876689,
-0.0225527473,
-0.000479161,
0.0350712501,
-0.0492033362,
0.0055122548,
-0.0019948257,
0.0141478451,
0.0540139638,
-0.0160386432,
-0.008476791,
0.0078202747,
0.0111328755,
-0.0001297602,
-0.0247192699,
0.0167346057,
0.0150559945,
0.0088645751,
0.0706204399,
0.0363109931,
-0.0044235038,
0.0014544604,
-0.1004412994,
0.0593494512,
-0.0217372812,
0.0241633933,
-0.0379828326,
-0.0085017532,
0.0076872581,
-0.0087544397,
-0.0109865107,
-0.0087018274,
0.0332575403,
0.010697389,
-0.0051840162,
-0.0393412486,
-0.0259484742,
0.0071122344,
-0.0116249034,
-0.0326783918,
-0.037686687,
-0.0358591266,
0.0039516864,
0.0270381272,
0.0155408988,
0.0643644333,
2.392867065e-34,
0.0244713891,
-0.0440589748,
0.048892919,
-0.0171023943,
0.0187747423,
0.0112781767,
-0.0349403135,
0.0372843444,
-0.0384895504,
0.0176688042,
-0.0080126515
] |
http://arxiv.org/abs/2401.13789v3 | http://arxiv.org/pdf/2401.13789v3 | A Unified Approach to Emotion Detection and Task-Oriented Dialogue Modeling | 2024-01-24T20:17:11 | 2024-06-28T10:23:29 | Armand Stricker, Patrick Paroubek | cs.CL | cs.CL | In current text-based task-oriented dialogue (TOD) systems, user emotion
detection (ED) is often overlooked or is typically treated as a separate and
independent task, requiring additional training. In contrast, our work
demonstrates that seamlessly unifying ED and TOD modeling brings about mutual
benefits, and is therefore an alternative to be considered. Our method consists
in augmenting SimpleToD, an end-to-end TOD system, by extending belief state
tracking to include ED, relying on a single language model. We evaluate our
approach using GPT-2 and Llama-2 on the EmoWOZ benchmark, a version of MultiWOZ
annotated with emotions. Our results reveal a general increase in performance
for ED and task results. Our findings also indicate that user emotions provide
useful contextual conditioning for system responses, and can be leveraged to
further refine responses in terms of empathy. | A Unified Approach to Emotion Detection and Task-Oriented Dialogue Modeling: In current text-based task-oriented dialogue (TOD) systems, user emotion
detection (ED) is often overlooked or is typically treated as a separate and
independent task, requiring additional training. In contrast, our work
demonstrates that seamlessly unifying ED and TOD modeling brings about mutual
benefits, and is therefore an alternative to be considered. Our method consists
in augmenting SimpleToD, an end-to-end TOD system, by extending belief state
tracking to include ED, relying on a single language model. We evaluate our
approach using GPT-2 and Llama-2 on the EmoWOZ benchmark, a version of MultiWOZ
annotated with emotions. Our results reveal a general increase in performance
for ED and task results. Our findings also indicate that user emotions provide
useful contextual conditioning for system responses, and can be leveraged to
further refine responses in terms of empathy. | [
0.0320086405,
0.0296548903,
-0.0058293506,
0.063285403,
0.012095077,
0.0465071499,
-0.002016234,
-0.0182334706,
0.0681380183,
-0.0395923182,
-0.0588958859,
-0.0541065186,
-0.0541532561,
0.0447851233,
0.0054484867,
-0.0875714347,
0.0307822302,
0.0375992246,
0.0256864317,
0.0022604014,
0.0574086644,
0.040439371,
-0.0285487771,
0.0435915366,
-0.0202454608,
-0.0137045523,
-0.0503959432,
0.0088877063,
0.0294866338,
-0.0285547227,
0.0120257232,
0.0213506408,
0.0156760886,
0.0056294692,
0.0000020833,
-0.0452589057,
-0.0055631441,
-0.0413861983,
-0.0117105758,
-0.0117893675,
0.0583964624,
0.0084745446,
-0.0141784009,
0.0209736153,
-0.0382352024,
-0.0072894348,
0.0374822691,
0.0963739604,
0.0801720619,
0.0628599152,
-0.0439361557,
0.0235835947,
0.019248141,
0.0027377296,
0.0200064778,
-0.0251866393,
-0.0001165055,
0.0025014891,
0.0077245403,
-0.0538561046,
0.0035986777,
-0.0004665338,
-0.0040659402,
0.0036356654,
-0.0378411002,
-0.0033237247,
-0.000729645,
-0.0308734328,
-0.0706022009,
0.0609756522,
0.0303298049,
0.0293572731,
0.0206753667,
0.0126463221,
-0.0082666129,
-0.0299414881,
-0.0365505815,
-0.0495338291,
-0.0401142165,
0.0216633976,
0.0185686219,
0.0213499069,
-0.0039729369,
0.048026856,
-0.0020805949,
0.0754369721,
-0.023160968,
-0.0244776439,
0.0111427512,
0.0191619433,
-0.0446207933,
-0.0449902676,
0.0219443068,
0.0126496917,
-0.0171051528,
0.0255209859,
-0.0655075759,
-0.0127219129,
-0.0248846412,
-0.0653736442,
-0.0044280593,
0.0680124685,
-0.0389305614,
0.0320140421,
0.022784695,
0.0367060341,
-0.0057535861,
-0.0325580016,
0.0104823122,
-0.0204482153,
-0.0429014862,
0.0274840705,
-0.0287992582,
0.0253662802,
-0.0235567801,
-0.0515880175,
-0.0504287817,
0.0524757393,
-0.0175220706,
-0.0164932236,
-0.0152811781,
-0.0222239364,
-0.0233709738,
0.0168794934,
-0.0214363653,
0.047325179,
-0.0098323664,
0.0364624523,
0.0283494126,
-0.0564711839,
-0.0028518161,
0.0442626141,
0.02954318,
0.0257234313,
0.0185570437,
0.0046315575,
-0.0092272982,
-0.0383718759,
-0.0271276943,
-0.0209643487,
-0.0276262537,
-0.0488228798,
-0.0177074689,
-0.0358988009,
0.0459595211,
-0.0051144767,
0.0049963337,
-0.0013761272,
0.028238371,
-0.0054815505,
-0.0122802174,
0.0480780192,
-0.0470276289,
-0.0318897478,
-0.0086384593,
0.0130231958,
-0.0157130379,
0.0565070175,
0.0308796931,
0.0529755056,
0.0359640568,
0.0417099595,
0.0123869544,
0.033297237,
-0.0298452787,
0.0030752984,
0.0378506966,
-0.0085010193,
-0.0151076503,
0.0142150633,
0.041697681,
0.0313791856,
-0.046100378,
0.0260397326,
0.0462359898,
0.0507493541,
-0.0250970386,
0.0127081601,
-0.0114972787,
0.023807995,
0.0341508426,
-0.0034620545,
-0.0293138232,
0.0323163942,
-0.0148194442,
-0.0449592024,
0.0169315208,
0.0063786693,
-0.0359713435,
-0.0373993814,
-0.0209277067,
-0.0074605094,
-0.002743474,
0.0014705503,
0.0340091996,
-0.0425875597,
-0.0520355031,
-0.0114733288,
0.0101420023,
-0.0722465441,
-0.0117788687,
-0.048748102,
-0.016022427,
0.0900539234,
0.0244281646,
-0.0379792787,
-0.0462307371,
-0.0210422128,
0.0092602996,
0.0599040426,
0.0428085625,
0.0007951746,
0.0435322635,
-0.0543050319,
-0.0047344472,
0.0114612235,
0.0059217508,
0.0202903282,
-0.0248343889,
0.0111215236,
0.0000290056,
-0.0078953914,
0.0505033173,
0.0245047007,
0.0173701849,
0.0128431292,
-0.032425236,
0.0076843761,
0.050240986,
0.0392111726,
0.0140422089,
-0.056046281,
0.0097349016,
0.0026611332,
-0.0165703949,
0.0487521775,
-0.0260677617,
-0.0662117526,
0.0370481163,
0.0108462069,
-0.0162037127,
0.0497278646,
-0.0167709608,
-0.0396993719,
-0.0054830932,
-0.0341093205,
0.0373680405,
-0.0700812265,
-0.000411725,
-0.0069455956,
0.0187547505,
0.0105848024,
0.0151705109,
-0.0091535421,
-0.0123589626,
0.0111530041,
-0.0369380154,
0.0298038889,
-0.05171315,
-0.0288519021,
0.0029475656,
0.0184140671,
-0.00670199,
-0.0103713423,
-0.0067577078,
-0.0536956079,
-0.0325944051,
-0.1219655722,
0.0115506174,
-0.0262882188,
0.0440591425,
0.0161253549,
0.0013176341,
-0.0541497134,
0.0095007075,
-0.0347262919,
0.0234912857,
0.0642187372,
-0.0238395445,
0.0058342083,
-0.0052109626,
0.0003745865,
-0.0776344314,
0.0392971188,
-0.0429095812,
-0.0125798741,
0.0033186341,
-0.0807148069,
-0.0391022749,
0.0297167115,
-0.0518742651,
0.0606538504,
-0.0283579919,
-0.0709233209,
0.0256812647,
0.0093220389,
0.0511953495,
0.0387779549,
-0.0141283702,
0.0578044429,
-0.0027503276,
-0.0191224236,
-0.0630305633,
0.0557643175,
-0.0151423225,
0.0450650789,
-0.0361648537,
-0.0434668325,
0.0172380991,
0.0533553064,
0.0159201808,
0.0072035142,
-0.0593369715,
-0.0230214912,
0.0041564824,
-0.0256933812,
0.0239715166,
-0.0378470384,
-0.0330289714,
0.0105858976,
0.0001335265,
-0.0433131047,
0.058143314,
-0.0091207838,
-0.0429771282,
-0.0862436518,
0.0110239554,
-0.0202435236,
0.049711585,
-0.0465623401,
0.0179714356,
0.0195727814,
-0.0020714919,
-0.008935486,
0.0082589537,
-0.031674847,
-0.0532848127,
0.0075678839,
-0.014050955,
-0.0263866838,
0.0293073244,
-0.0380342565,
-0.0413255766,
-0.0098943207,
-0.0426140949,
-0.010355291,
0.0126547944,
-0.0309929531,
0.010665684,
-0.0162995029,
0.0336301327,
0.0074117403,
-0.0017504323,
-0.0385284647,
-0.0311223809,
0.0088804103,
0.0792436376,
-0.0297638923,
0.0470407084,
-0.0409091711,
0.082102567,
0.0716672391,
0.0902706236,
-0.0454633497,
0.023259338,
-0.0071924138,
0.0288620386,
0.018358374,
-0.0287928078,
-0.0092059551,
0.055737637,
-0.0255749896,
0.0383273847,
0.1440289766,
0.039616663,
-0.0633944944,
-0.0206264574,
-0.0408874713,
0.0050531072,
0.0267233159,
0.016933918,
-0.0639250651,
-0.0520664044,
0.0148920249,
0.0344001241,
-0.0232369173,
-0.0394290946,
0.0030862931,
-0.03704473,
0.0243127681,
-0.0257325117,
0.0336116031,
-0.0116165029,
-0.025689302,
0.0570686832,
0.0359739102,
-0.0173977744,
-0.0211182889,
0.0266971458,
-0.0054152664,
0.0276623964,
-0.0080738161,
0.0119493902,
0.0358063243,
-0.0329799205,
0.0662356839,
0.0436142273,
0.089230068,
0.0011302326,
-0.0107594859,
0.012685487,
0.0333176404,
-0.0216604341,
-0.0031692525,
0.0322780348,
-0.0393604897,
-0.0143646365,
-0.0624707565,
0.0504624657,
-0.0330377072,
0.0359857082,
-0.0074904575,
0.0632934123,
-0.0604532212,
-0.0018591151,
-0.0115849758,
0.0034294445,
0.0328415409,
0.0629496351,
0.0153774833,
-0.0759361088,
-0.0181856509,
0.0024040719,
0.0163891725,
0.0045265644,
-0.0081332661,
-0.0450074002,
-0.0102701774,
-0.0027917828,
-0.0722080097,
-0.0873727426,
0.0083383098,
0.051683832,
-0.0359945036,
-0.0372888669,
-0.0174616482,
0.0006314779,
-0.0111413635,
0.0117632393,
0.0061049922,
0.034298189,
-0.0405031703,
0.0184757076,
0.0023952525,
-0.0209222026,
0.006695725,
0.047488451,
-0.0494341366,
-0.041903764,
0.0023930832,
0.048817277,
0.0191640668,
-0.0117405597,
0.023468636,
-0.0164317526,
-0.0510751642,
-0.003165564,
-0.0299687181,
0.0266836565,
0.0617114939,
0.0668680668,
0.003475239,
0.0340535305,
-0.0060929894,
0.0376366638,
-0.0251018703,
0.0146995671,
-0.0134260282,
-0.0833353549,
-0.0170625281,
0.0004516443,
0.0798500627,
0.0187247954,
-0.0359822996,
-0.0237549338,
0.0020809078,
-0.0270717312,
0.0113049652,
-0.0706923306,
-0.0024341349,
-0.0230185557,
0.0198969655,
-0.0485384427,
-0.0090532415,
-0.0158966873,
-0.016894931,
-0.0300311595,
0.0097623067,
0.0055730031,
-0.0111913159,
-0.0079809716,
-0.003258198,
-0.0250186305,
-0.0202810429,
0.0278367866,
0.0053124786,
-0.0060119098,
0.1151555702,
0.0167422537,
-0.0771102682,
0.0062565794,
0.0263274238,
-0.0341038853,
0.0060484898,
-0.0019844035,
0.0298402812,
0.0172450989,
0.0365489796,
0.0704834312,
-0.0626155138,
0.0264550298,
0.0013307031,
0.0373105444,
-0.0072361501,
-0.0160033461,
-0.0035337685,
-0.0049347426,
-0.0190053377,
-0.0197999477,
0.0247264393,
-0.0395956039,
0.0631582811,
-0.0065324032,
-0.0284542944,
-0.0435708314,
0.0046602343,
0.0550466925,
0.0267261192,
-0.0061818128,
0.0244455189,
-0.0593952909,
-0.0370227844,
0.0229484793,
-0.0302214995,
-0.0252808221,
-0.0542397574,
0.0439573005,
0.0581212342,
-0.0352410078,
0.0240728837,
-0.0144780315,
-0.0093536098,
-0.0381433778,
-0.0102161467,
-0.0032579657,
-6.112049366e-33,
0.0031680344,
-0.0054564024,
-0.0250107758,
0.0463569164,
-0.0273274872,
0.0007502825,
-0.0346966945,
0.0298366267,
-0.0292662028,
0.0407492816,
-0.0044773095,
0.059861362,
0.0053637666,
-0.0061584837,
-0.0389690362,
0.01760621,
0.0280820001,
-0.0160459578,
0.0057703592,
0.0083835218,
0.0747228861,
0.0077046524,
0.0607344136,
-0.0707966089,
0.060772229,
0.0118200136,
-0.020319771,
0.0284900274,
0.0466942452,
0.0295811798,
-0.0379728153,
0.067821309,
0.0186299905,
-0.0221297722,
0.0007500195,
0.0471002124,
-0.0023693983,
-0.0072467229,
0.0148421032,
-0.0160689726,
-0.0338135473,
0.0008516633,
-0.0120295044,
0.0249642637,
-0.0259779524,
0.0133011593,
0.010157722,
-0.0026321216,
0.0487810671,
-0.0342801698,
-0.0444506556,
0.0134078301,
0.0008863765,
-0.0709113032,
0.0012602501,
0.0858961046,
0.0068898564,
-0.0381524526,
-0.0334534906,
-0.0299249943,
-0.0271232165,
-0.042435497,
0.0331933945,
-0.0119529264,
-0.0039223479,
0.1093608439,
0.0204022508,
0.0164592881,
-0.0459017679,
-0.0001443066,
0.0055680862,
0.0078812893,
0.0431836918,
0.0219867323,
-0.0384498723,
0.0203922018,
-0.0655890927,
0.0345318019,
0.0976123065,
0.0351799689,
0.0463848226,
0.0246761441,
-0.0677609518,
0.0140494835,
0.0035176838,
0.015266913,
-0.0178632792,
-0.0326080248,
0.0450829379,
0.004356924,
-0.0078873066,
-0.0170692392,
0.0094336951,
-0.0233694594,
0.0122870682,
0.0302074179,
-0.0487178788,
0.0430379286,
0.0036659911,
0.0202173144,
-0.0320535451,
0.0055756052,
0.0206568819,
0.0258009993,
0.0334617384,
-0.009264023,
-0.011172602,
0.0161334649,
-0.0599150322,
0.0034084409,
-0.053216517,
0.0590943322,
-0.0286205932,
0.0278981626,
0.0093767978,
-0.0318843052,
-0.017147528,
0.0410644524,
0.0056398469,
-0.022299068,
-0.0595573001,
-0.0315183252,
-0.0075171022,
0.0173558872,
-0.0336098112,
-0.0079436842,
0.0073649324,
-0.0430583209,
-0.0329492092,
-0.0179054663,
-0.0119241923,
-0.0041498346,
2.784e-7,
-0.0003643162,
0.0421776287,
0.0186411031,
-0.0274927672,
0.0171405766,
0.0641931742,
0.0100200372,
0.0530804954,
0.0557406694,
-0.025500739,
-0.0000044515,
0.003103154,
0.0640336275,
0.0126989037,
-0.1027540863,
-0.0420090966,
0.0309513193,
-0.0013226266,
-0.02701349,
0.027837893,
0.1087817475,
0.053084258,
0.0511147156,
0.0257846285,
0.0211774483,
0.0313167609,
0.0534614213,
-0.0406991541,
0.000859993,
0.0341042913,
0.0306961387,
-0.0388883464,
-0.0171328168,
-0.0129899578,
0.0131679866,
0.0440472253,
0.0177676026,
0.0775922164,
-0.0136522641,
-0.0477036536,
0.0364829563,
0.0090767043,
-0.0160982814,
-0.0578213185,
0.0332500562,
-0.0331877321,
-0.0295775924,
-0.0172836613,
-0.0062234439,
-0.024540633,
-0.0051540737,
0.0022226558,
-0.0447209775,
0.0240654107,
-0.0013589021,
-0.0457689278,
-0.0389910489,
-0.0704694912,
-0.0179656111,
-0.0092276931,
-0.0599656478,
-0.0538904034,
-0.0510088466,
0.0158785433,
0.0098352563,
-0.0434782133,
-0.0025496338,
2.314519678e-34,
0.0006112419,
-0.01207799,
0.0096711991,
0.0489878617,
0.0335748829,
0.013340109,
-0.0357323624,
-0.0098488545,
0.0736005455,
-0.0677990243,
0.0242539719
] |
http://arxiv.org/abs/2401.13782v2 | http://arxiv.org/pdf/2401.13782v2 | Tweets to Citations: Unveiling the Impact of Social Media Influencers on AI Research Visibility | 2024-01-24T20:05:49 | 2024-03-03T08:16:52 | Iain Xie Weissburg, Mehir Arora, Xinyi Wang, Liangming Pan, William Yang Wang | cs.DL | cs.DL, cs.AI, cs.CL, cs.CV, cs.LG, cs.SI | As the number of accepted papers at AI and ML conferences reaches into the
thousands, it has become unclear how researchers access and read research
publications. In this paper, we investigate the role of social media
influencers in enhancing the visibility of machine learning research,
particularly the citation counts of papers they share. We have compiled a
comprehensive dataset of over 8,000 papers, spanning tweets from December 2018
to October 2023, alongside controls precisely matched by 9 key covariates. Our
statistical and causal inference analysis reveals a significant increase in
citations for papers endorsed by these influencers, with median citation counts
2-3 times higher than those of the control group. Additionally, the study
delves into the geographic, gender, and institutional diversity of highlighted
authors. Given these findings, we advocate for a responsible approach to
curation, encouraging influencers to uphold the journalistic standard that
includes showcasing diverse research topics, authors, and institutions. | Tweets to Citations: Unveiling the Impact of Social Media Influencers on AI Research Visibility: As the number of accepted papers at AI and ML conferences reaches into the
thousands, it has become unclear how researchers access and read research
publications. In this paper, we investigate the role of social media
influencers in enhancing the visibility of machine learning research,
particularly the citation counts of papers they share. We have compiled a
comprehensive dataset of over 8,000 papers, spanning tweets from December 2018
to October 2023, alongside controls precisely matched by 9 key covariates. Our
statistical and causal inference analysis reveals a significant increase in
citations for papers endorsed by these influencers, with median citation counts
2-3 times higher than those of the control group. Additionally, the study
delves into the geographic, gender, and institutional diversity of highlighted
authors. Given these findings, we advocate for a responsible approach to
curation, encouraging influencers to uphold the journalistic standard that
includes showcasing diverse research topics, authors, and institutions. | [
0.0416081734,
0.0547795482,
-0.0335979648,
-0.0432018526,
-0.0134798028,
-0.0351301096,
0.0944628492,
-0.0057770321,
0.0318155177,
0.002337853,
0.0129011096,
0.0562133119,
0.010196195,
0.0651241392,
0.0260738879,
0.0128722107,
0.0045862463,
0.0360711887,
0.0548037104,
-0.0335911252,
-0.0072909258,
0.0050314595,
0.0041828598,
0.0249515194,
0.0460843518,
-0.0403121673,
0.0445102602,
-0.0150392745,
0.0240296628,
-0.0266089775,
0.0166499224,
0.023456974,
0.006932375,
0.0680662766,
0.0000016528,
-0.0194663759,
-0.0211768728,
0.0235708989,
0.0145368893,
0.0072593046,
0.0441034436,
-0.0329932868,
0.0140974298,
0.0019590787,
-0.0695848912,
0.0498824678,
0.0308145378,
0.0138965547,
-0.0760407895,
0.0212882292,
-0.0085786059,
0.0266054012,
0.053703405,
-0.0026808206,
0.0179861244,
-0.007333037,
-0.0120536713,
-0.0169151817,
-0.0229193363,
0.0354705006,
0.0346985906,
0.070335947,
-0.0137959989,
-0.0017127966,
0.0641263127,
0.0777865946,
0.001004423,
-0.0017562964,
0.0045965081,
0.016244607,
0.0426631868,
-0.0238952823,
0.0645719469,
-0.0200919285,
0.0137683274,
0.0625416189,
-0.0539459065,
0.0310976077,
0.0267055016,
-0.0225750674,
-0.0154639352,
0.018160034,
0.0316388868,
0.0684747472,
0.0171838049,
0.0326137133,
-0.019421177,
-0.0521884076,
0.0319784917,
0.0066629811,
0.0550242998,
-0.0570979901,
-0.0044812323,
-0.0139453616,
0.0276664011,
-0.0211447943,
0.0349664651,
-0.0354657099,
0.0058916663,
-0.0355337523,
-0.0067175035,
-0.0468536727,
0.0337618664,
0.0361743458,
0.0239266306,
-0.0476562679,
-0.0446880832,
0.0421012379,
0.0758152157,
0.04263255,
-0.0335177332,
-0.013067076,
0.0210679211,
-0.0109333079,
0.0263353996,
-0.0071340087,
-0.0246799327,
0.0172378328,
-0.0414286032,
-0.0275151823,
-0.0397300012,
0.0550977476,
-0.0149943512,
0.0204911977,
0.019666275,
0.0180267263,
-0.0237923451,
0.0411268137,
0.0042263269,
0.0087311706,
0.0056359465,
0.0069521675,
0.0545995645,
-0.032454364,
0.0283543356,
0.0005248424,
0.022636611,
-0.0482910424,
0.0088569988,
-0.0359710716,
0.034661267,
-0.0277065486,
0.0172045622,
-0.0380177163,
0.0141152376,
0.0454916582,
0.023034865,
-0.0523982756,
-0.0417849645,
-0.005495389,
-0.0265414827,
0.003278079,
-0.0267999861,
0.0101353144,
0.0399465524,
0.0236904379,
-0.0098695038,
0.0336548723,
0.0365064628,
0.0306379702,
0.05013014,
-0.0277127698,
0.0276387222,
0.0106844008,
0.0505934581,
0.0158844199,
0.0525706932,
-0.025370406,
-0.069893606,
-0.011647705,
-0.0179679915,
0.0211280044,
-0.0235104617,
0.0339275189,
0.0390904993,
0.0442391969,
-0.0300216824,
0.0447792374,
0.0007988991,
0.0131416777,
-0.0180005841,
-0.0458869115,
-0.0108361794,
0.0141644785,
-0.0153714819,
-0.0354554839,
0.042267859,
0.0199794695,
-0.0232566074,
-0.0068825618,
0.0376257561,
0.0076250881,
0.0349828079,
0.0058805263,
0.0300920084,
-0.0220114086,
0.0043382361,
-0.0131774964,
-0.0097076483,
-0.0090783369,
-0.0013257787,
0.0115522752,
0.0275302622,
-0.0276928209,
0.0292826593,
0.0040722336,
0.0484954715,
-0.0269889049,
0.0087764757,
-0.0011528063,
-0.0244622976,
0.0158735067,
-0.0239777341,
0.0235408675,
-0.0313548446,
0.0126557276,
0.0024868606,
-0.0210416932,
0.0601723045,
0.0645296872,
-0.0154823503,
-0.0331645496,
-0.0720365271,
-0.0050554196,
-0.0235169902,
-0.0077427807,
-0.0143891331,
-0.0284352992,
0.041275803,
0.0482669324,
-0.0317844786,
-0.0211938098,
0.0586321689,
-0.0167284925,
-0.0460499898,
0.031213291,
0.0380774811,
0.0933013707,
-0.0018579101,
-0.0796603262,
-0.0457638726,
0.0149020096,
-0.0346248038,
0.0133214863,
0.1079878137,
0.012278364,
0.0009123894,
0.0420790501,
-0.0645989627,
-0.0455652513,
0.0135453884,
0.0019171091,
0.0326936431,
-0.0415848419,
0.0034496298,
0.0169751067,
-0.0596411116,
-0.008670955,
0.0047387136,
0.0288111717,
-0.0240676105,
-0.0299323685,
0.0083728991,
0.0510353558,
-0.0731999204,
-0.0221737064,
0.0024232764,
-0.0310810413,
-0.0113907671,
-0.013743476,
-0.0386090837,
0.0007696765,
0.0103388401,
0.0091348048,
0.0101963738,
0.0173926726,
-0.0632653981,
-0.0027914641,
0.0147472033,
0.0217039157,
0.0260283332,
0.0087495055,
-0.041930709,
-0.0260933489,
-0.0267859846,
0.0304648839,
0.0797923431,
-0.0318262987,
0.0167252291,
-0.0245529432,
-0.0436277352,
-0.0168287735,
-0.0599675812,
0.0567037426,
-0.0029880102,
0.0250550564,
-0.0371580422,
0.0648154169,
-0.0064725149,
0.0561101437,
0.0437554531,
0.0464648865,
-0.0361739211,
-0.0087569552,
-0.0263663884,
0.0047645736,
-0.0168537796,
-0.0433014072,
0.0757387355,
-0.0321320817,
-0.0096192192,
-0.0091006458,
-0.0064409901,
0.0290700421,
-0.0021151868,
-0.039753899,
-0.0020091387,
-0.0054986086,
0.0043303217,
0.0746382028,
0.0034292308,
-0.0424223803,
-0.0010159027,
-0.0014452947,
0.032076519,
-0.0296850335,
-0.0510090366,
-0.014524268,
-0.003949218,
0.0289469976,
0.0104331262,
0.0181411207,
0.0341864936,
0.0031228028,
0.009370611,
-0.011381655,
-0.0095144119,
-0.047629416,
0.0435825102,
0.0212698877,
0.0053817714,
-0.0153481178,
0.0196795017,
-0.0906733871,
-0.0494492091,
-0.0070012435,
0.0604237951,
-0.0977939591,
-0.0551619679,
-0.0273728501,
0.0715979114,
-0.0044786264,
-0.0361324623,
-0.0444783568,
0.0005575022,
-0.0125149013,
0.0535999127,
-0.007481005,
0.0292419717,
-0.010902049,
-0.0527331121,
0.0156798493,
0.0400675759,
0.002367798,
-0.0601638183,
-0.040335767,
-0.0066620661,
-0.040667206,
0.0256940685,
-0.0086627025,
0.0732468739,
0.0117373513,
0.0586786941,
0.0437601209,
0.0226767827,
-0.0517427996,
-0.0017996345,
-0.0265840739,
0.0600372255,
-0.0095461905,
0.0150157148,
-0.1117786244,
0.0413163565,
-0.0045068422,
0.0034569108,
-0.0721293986,
-0.0398033485,
-0.0021183451,
-0.0665815249,
0.0459469371,
-0.0244408287,
0.0194645356,
-0.0224020649,
-0.0664619207,
0.0116590681,
-0.0477018617,
-0.0252421796,
-0.0181035008,
-0.003392075,
0.0148931621,
0.0036132175,
-0.0140786171,
-0.0190535206,
0.0026245553,
-0.054934293,
0.0056143678,
0.0332023539,
0.0559963472,
-0.0461292788,
-0.0119554857,
-0.0480354726,
0.0440219082,
0.0061399718,
-0.0127119841,
0.0745707899,
-0.0228084438,
0.0193621628,
-0.049880106,
0.0253895409,
-0.0208030287,
0.0254631136,
0.0031681301,
0.0423947684,
0.0160680301,
-0.0604558662,
-0.0405830182,
-0.0158078801,
-0.0104188491,
0.0212643184,
0.0187598672,
-0.0276508424,
-0.0215811264,
-0.0148027865,
0.0565527007,
0.0226169359,
-0.0377574302,
-0.06302578,
-0.0234968998,
-0.0247625411,
-0.0245454721,
-0.0335518792,
-0.0096685272,
-0.0123495907,
0.0195532944,
0.0072843768,
0.0416183509,
-0.008584829,
-0.0177934021,
0.0289051253,
-0.0129625015,
0.0159273073,
0.0198213533,
-0.0089644371,
0.0103811035,
0.0178528242,
-0.0372321121,
0.0204737596,
0.0234691892,
-0.0347070657,
-0.0215830598,
0.0349948481,
0.0244213101,
0.035700947,
-0.0225722659,
-0.0683201924,
0.0479520597,
0.0876787156,
-0.0090687731,
0.0118638296,
0.0204757247,
-0.0494070277,
0.0044511664,
-0.023664657,
-0.0273781307,
-0.0142063918,
0.0126763312,
-0.0675404593,
-0.0592690259,
0.0172801148,
-0.0236273706,
-0.0231164023,
0.0483349785,
-0.0309912767,
0.0152189182,
0.0254989415,
-0.0516402982,
0.0363967493,
0.0240960866,
-0.0158149004,
-0.0049644061,
0.0074966601,
0.0251743048,
-0.0190582871,
0.0470422432,
0.0029218732,
0.0212014355,
-0.0910315812,
0.0389860943,
-0.0180931743,
0.0671128109,
-0.0054144259,
-0.002232718,
-0.0010919054,
0.0087611452,
-0.0330770202,
0.0360955782,
0.0214434899,
-0.0202327706,
-0.022344647,
-0.0022544388,
0.0224821791,
0.0264948197,
-0.1182299107,
-0.0594103038,
0.0181414224,
0.0011753385,
0.0255210958,
0.009786983,
0.0558686107,
-0.0072899223,
0.0045461217,
-0.0199316908,
0.0042937188,
-0.0050335554,
0.0281639006,
0.0000535402,
0.0103017557,
-0.038072031,
0.0103875389,
0.0543191843,
0.0186086372,
-0.015181019,
0.0191674959,
-0.0327322707,
0.0379455835,
-0.0222539883,
-0.0089641884,
0.0696774498,
0.0311052892,
-0.0098604187,
-0.0047592637,
-0.035813041,
0.0527127534,
-0.0343414284,
0.0234182347,
-0.0482256636,
-0.0454855189,
0.0373390131,
-0.0143321371,
-0.0219542086,
-0.0029271364,
0.0159163214,
-0.0521247163,
-0.0893786997,
-0.0147993006,
-6.271552562e-33,
0.0542861968,
-0.0577850565,
-0.0395005457,
-0.0311032459,
-0.0228384696,
0.0283520576,
-0.01919863,
0.0158978384,
-0.034669742,
0.0182835497,
-0.0747737885,
0.023083698,
0.0042503634,
0.0020783197,
0.0670009702,
0.008101495,
0.0221705176,
-0.0083137061,
-0.0191048197,
0.0848531723,
-0.0309822895,
-0.0021921867,
-0.0254288018,
-0.1099570617,
0.0428293683,
-0.0169951785,
-0.0057746172,
-0.0338580608,
0.0233799331,
0.019796703,
-0.0267990436,
0.0291961692,
0.0212862212,
-0.0463627391,
-0.0035409927,
0.0346235968,
-0.0613088645,
-0.0088374466,
0.0213737227,
-0.0369281843,
-0.0162669569,
-0.0217899214,
0.0705017671,
-0.0080461744,
-0.0349796191,
-0.0042576646,
0.0616103746,
0.0347170942,
0.0077955304,
-0.0386582799,
-0.0105809178,
0.0186591186,
-0.0606473833,
0.0746352375,
0.003095164,
0.0379415527,
0.0163792241,
0.0936383381,
-0.0719566494,
0.0176384598,
-0.006980591,
-0.0204742327,
0.015690513,
-0.0511066504,
0.0104367081,
0.024669867,
0.0071902233,
0.0091198599,
-0.0121047897,
-0.0432373136,
0.0007607042,
0.0436118953,
0.0161676761,
0.0644678399,
0.0442348607,
-0.0392468087,
-0.0616997071,
-0.0056742569,
0.033178553,
-0.0221409146,
0.0436164886,
-0.0235202722,
-0.0309599228,
-0.0056899562,
0.0227760486,
-0.06441769,
0.0010663753,
-0.040462587,
-0.0506688617,
0.007596307,
-0.021307379,
0.0614552312,
0.0291212462,
-0.0396221504,
-0.0873985291,
0.0480161533,
0.0179296024,
0.0638656542,
0.0315561295,
-0.0261046682,
-0.0921240523,
0.004195157,
-0.0389168486,
0.0350458622,
0.028343508,
-0.0028682398,
0.0355512798,
0.0020260527,
-0.0118977968,
-0.0221207663,
0.0025315098,
0.0230603367,
-0.005890599,
-0.0113034081,
-0.0134673147,
0.0554449409,
0.0127655342,
-0.0398260951,
0.041739393,
0.075214088,
0.0034096742,
0.0291567799,
-0.0691634342,
0.0018758487,
-0.0188444965,
-0.0432904139,
-0.0465047434,
0.0418305658,
0.0334685966,
-0.0710207447,
-0.0163539238,
-0.0974145979,
2.571e-7,
0.021361731,
0.0354488939,
-0.0514466278,
-0.0112025086,
0.047076162,
0.0195809342,
0.040253859,
0.0486321263,
0.0292247869,
-0.0017635759,
-0.0135778813,
0.024640955,
0.0005792984,
-0.0129155414,
-0.0472770445,
-0.0443211608,
-0.0993492901,
-0.016381884,
-0.0406758785,
0.0350455195,
-0.0905052647,
0.0022492355,
0.0248772725,
0.0095429933,
-0.0185873229,
-0.0299319513,
0.0302026458,
-0.0393508151,
0.05723501,
0.0160151795,
0.0157625619,
-0.0131951179,
0.0137527632,
-0.0087202424,
-0.0182401408,
0.02927492,
-0.0411317572,
-0.0175971221,
0.0347211771,
0.0439310744,
0.0378244519,
-0.0314366445,
0.0052339816,
-0.0481177792,
0.0427622758,
-0.0257843044,
-0.089436546,
-0.0642129034,
-0.0076333056,
-0.016217744,
0.0538706817,
0.0094516538,
-0.0001980357,
0.0119676962,
-0.039599292,
-0.0407632589,
-0.0147440787,
-0.032508146,
0.0205752905,
-0.0219703149,
-0.000522516,
-0.0495146029,
0.0034071619,
-0.0406327359,
0.017067844,
0.0109344507,
0.0763433725,
1.8620649e-34,
0.0433702655,
-0.0104643842,
0.0213984028,
0.0278249402,
0.0323715359,
0.000363955,
-0.0256707817,
-0.0296692364,
0.0404507816,
-0.0308073722,
0.0236450005
] |
http://arxiv.org/abs/2401.13660v2 | http://arxiv.org/pdf/2401.13660v2 | MambaByte: Token-free Selective State Space Model | 2024-01-24T18:53:53 | 2024-04-03T02:36:27 | Junxiong Wang, Tushaar Gangavarapu, Jing Nathan Yan, Alexander M. Rush | cs.CL | cs.CL, cs.LG | Token-free language models learn directly from raw bytes and remove the
inductive bias of subword tokenization. Operating on bytes, however, results in
significantly longer sequences. In this setting, standard autoregressive
Transformers scale poorly as the effective memory required grows with sequence
length. The recent development of the Mamba state space model (SSM) offers an
appealing alternative approach with a fixed-sized memory state and efficient
decoding. We propose MambaByte, a token-free adaptation of the Mamba SSM
trained autoregressively on byte sequences. In terms of modeling, we show
MambaByte to be competitive with, and even to outperform, state-of-the-art
subword Transformers on language modeling tasks while maintaining the benefits
of token-free language models, such as robustness to noise. In terms of
efficiency, we develop an adaptation of speculative decoding with tokenized
drafting and byte-level verification. This results in a $2.6\times$ inference
speedup to the standard MambaByte implementation, showing similar decoding
efficiency as the subword Mamba. These findings establish the viability of SSMs
in enabling token-free language modeling. | MambaByte: Token-free Selective State Space Model: Token-free language models learn directly from raw bytes and remove the
inductive bias of subword tokenization. Operating on bytes, however, results in
significantly longer sequences. In this setting, standard autoregressive
Transformers scale poorly as the effective memory required grows with sequence
length. The recent development of the Mamba state space model (SSM) offers an
appealing alternative approach with a fixed-sized memory state and efficient
decoding. We propose MambaByte, a token-free adaptation of the Mamba SSM
trained autoregressively on byte sequences. In terms of modeling, we show
MambaByte to be competitive with, and even to outperform, state-of-the-art
subword Transformers on language modeling tasks while maintaining the benefits
of token-free language models, such as robustness to noise. In terms of
efficiency, we develop an adaptation of speculative decoding with tokenized
drafting and byte-level verification. This results in a $2.6\times$ inference
speedup to the standard MambaByte implementation, showing similar decoding
efficiency as the subword Mamba. These findings establish the viability of SSMs
in enabling token-free language modeling. | [
0.0587607957,
0.0805127397,
0.016935531,
0.0025149353,
-0.0361152925,
0.0238282736,
-0.0095887948,
0.0264590066,
-0.0419946723,
-0.0386741459,
0.0167991351,
-0.0689093992,
0.0033214027,
0.0509859547,
0.0320641138,
-0.0381171368,
0.0530142561,
-0.0533331074,
0.0430464037,
0.0053689913,
-0.0054883999,
0.0144984294,
-0.0024214357,
0.0568440519,
-0.0348063521,
-0.0008393566,
-0.0492167622,
0.0117023671,
-0.0348182991,
-0.0168358032,
-0.0091861533,
0.0275639463,
0.015049546,
0.0413287878,
0.0000021554,
-0.0146253286,
-0.0372831896,
-0.0050468207,
-0.0233905092,
0.0276470073,
0.0401713885,
-0.045939479,
-0.0396220312,
0.0416206606,
-0.0033700166,
-0.0462275222,
0.0626289397,
0.0085176053,
0.0393610038,
0.0938767791,
-0.0314470418,
-0.0334511138,
0.0203923341,
-0.0450092442,
-0.0437964536,
-0.010537588,
0.0148235522,
-0.0237900037,
0.0016765459,
-0.0174468849,
-0.0522722341,
0.0012079832,
-0.0046607424,
0.0177029613,
-0.0260337256,
0.0323386043,
-0.0219613463,
-0.0317871124,
-0.0295512676,
0.0280473717,
-0.0278532319,
-0.0727287903,
0.0074616894,
-0.0079840459,
-0.0275908355,
0.0053372453,
-0.009385203,
0.045933336,
-0.0049418923,
0.027986167,
0.0265794713,
0.0104561001,
0.0380595699,
-0.0039204531,
-0.0966393799,
0.0729888976,
0.0125812404,
-0.0397388861,
-0.0467268378,
-0.0515755154,
0.0207947474,
-0.0648626834,
0.015143468,
0.041037336,
-0.0149540547,
0.0000381853,
-0.0341847315,
-0.0465656072,
0.0153636737,
-0.0289197937,
0.0145388283,
0.0532384403,
-0.009694255,
0.008245023,
-0.0200160947,
0.0331839658,
-0.0322806425,
0.0302656889,
-0.0283298884,
-0.0019431981,
-0.0150852883,
-0.0152949318,
-0.0702406466,
0.0423459858,
-0.0435889699,
-0.0348475091,
0.001111304,
0.0092600659,
0.0568542667,
-0.0205070991,
0.0066002025,
-0.0136432508,
-0.0374949574,
0.0061496659,
0.0302942898,
0.0191200972,
0.0116907498,
-0.005213432,
-0.0096236337,
-0.0562881716,
-0.0254317783,
0.0285030771,
0.0088966712,
-0.0380046777,
0.0070009669,
0.097137481,
0.008333331,
-0.0627653152,
-0.0141141489,
0.0128869042,
0.0199820194,
-0.054485891,
-0.0127825076,
-0.0463303477,
0.0144004272,
0.0198592059,
0.017725477,
-0.0013819304,
-0.0017508283,
0.0183448065,
-0.0095495591,
0.0746427998,
-0.0170398541,
-0.0394932553,
-0.0066952892,
-0.0180967916,
0.0304844361,
0.0490125939,
0.0685471073,
0.0377833061,
0.001989,
-0.0058221933,
-0.0168544538,
0.0559230633,
0.0188341662,
-0.022089256,
-0.0340172946,
-0.007451552,
0.012942859,
-0.0200314224,
0.038025327,
0.0918901861,
0.0004283756,
-0.0046480489,
0.0124407355,
0.0463453382,
0.0004439798,
0.0277794711,
0.0421112627,
0.0381866619,
0.0134993484,
-0.0028130689,
-0.0100683272,
0.0219680946,
0.0259608235,
0.0258971713,
0.0450830385,
0.0466856137,
-0.0224485435,
-0.0177882854,
-0.0208726209,
-0.0428220741,
0.0291788615,
-0.0285279062,
0.0676721781,
0.0167267583,
-0.1080099642,
0.0838310272,
0.0169515833,
-0.0607038252,
-0.0093917875,
-0.059280742,
0.0331393108,
0.0426815934,
0.0595149584,
0.0056463955,
0.0245342422,
-0.0414050743,
0.0223410726,
0.0568294711,
-0.017896235,
0.024789853,
0.0216208976,
-0.0654686242,
-0.0058358316,
0.053874664,
0.0229531862,
0.0204252489,
0.0304086544,
0.0091151912,
0.0286134835,
-0.0422214493,
0.0069251475,
0.0303365868,
0.0159118939,
-0.0253936145,
-0.0092725884,
0.0151236551,
0.0690120235,
0.0123744532,
-0.0172872413,
0.0405265205,
0.0303394683,
-0.0502328351,
-0.0063726106,
0.0214933064,
-0.054877948,
-0.0306630228,
0.0199158452,
0.0316267572,
0.0475659519,
0.0580615252,
-0.0216555949,
-0.0670630112,
0.0032683737,
-0.0610638894,
0.1333592534,
-0.0517138094,
0.0111203082,
0.0317159072,
0.0509376563,
-0.004921935,
0.0246630535,
-0.0399174914,
-0.0164726526,
-0.0138050374,
-0.0010299974,
0.0283036288,
-0.0074260863,
-0.0259683281,
0.0292034317,
-0.0453960821,
0.005954871,
0.0019340154,
0.0010872424,
0.0182640981,
-0.014013581,
-0.0499855578,
-0.0022950876,
-0.0496031679,
0.039751932,
-0.0331585854,
0.0610289872,
-0.0046659713,
-0.017804822,
0.0012228151,
-0.0486739799,
-0.0207193457,
-0.0051010391,
0.010906144,
-0.0157721546,
-0.0218844879,
-0.0202502329,
-0.0275484174,
-0.0040395618,
0.0602787472,
-0.0324275605,
-0.0345647335,
-0.0272924956,
0.0014198127,
-0.0020979131,
0.0369397625,
-0.0208167806,
-0.0804599971,
0.0036169873,
-0.0410216078,
0.008806996,
0.0040066922,
0.0418521054,
0.010142006,
-0.0143051827,
0.0166886747,
-0.0187218823,
0.0140807293,
0.0089615285,
-0.0014228601,
-0.0112116635,
-0.0347355902,
0.0195225086,
0.0936841741,
-0.011701772,
0.012741819,
-0.0849022493,
-0.0302357618,
0.0073754182,
-0.011334612,
0.0385826565,
-0.0159153827,
-0.0344248191,
-0.0045521255,
-0.054664392,
-0.0509793125,
0.0530421138,
-0.07225205,
-0.0274409559,
-0.0109980954,
0.0155181224,
-0.0344994813,
0.0377564318,
0.0064905626,
0.00067907,
-0.0105197439,
-0.0544100553,
0.0157099087,
0.006694552,
-0.0116502671,
-0.0292233508,
0.0094804447,
-0.008902275,
0.0517017916,
-0.0160715058,
-0.004378601,
-0.0240198467,
-0.0005876144,
-0.0389394723,
0.0178944059,
0.0275139324,
-0.0325952284,
-0.011335765,
0.0737973675,
-0.0133116925,
0.0547743291,
0.0091177588,
0.0245866552,
-0.0250124,
0.0301626809,
0.0404494554,
-0.0467032827,
0.0000748574,
-0.0338870101,
-0.009943082,
0.0469049066,
0.0601990558,
0.0075949216,
-0.0263317116,
-0.0071608219,
0.0349087082,
-0.0185969621,
0.0060188184,
0.0480336137,
0.0498147942,
-0.0292287841,
-0.0401015468,
0.1234568283,
0.0303171743,
-0.0317852199,
0.0197786633,
0.0253722742,
0.0572193712,
-0.0194212813,
-0.0043788562,
-0.0626623929,
-0.051038716,
0.0342254974,
0.0126496693,
0.0052549355,
0.0013317313,
-0.0036059082,
-0.0814540088,
-0.0099826362,
0.0352377184,
-0.0731662363,
-0.020233348,
-0.0281633716,
-0.0482424684,
-0.0007396702,
0.0067986003,
0.0509221889,
-0.0190850236,
-0.0106884213,
0.0623003691,
-0.0256649572,
-0.0250867475,
0.0051669539,
-0.0229151323,
0.0656446889,
0.0804386958,
-0.0047690617,
0.002099755,
-0.0144889448,
0.0259059984,
-0.0010286076,
0.0008573362,
-0.0459749959,
-0.0172174647,
-0.0292414892,
-0.0070101377,
0.0301300492,
-0.0082200337,
-0.0354232155,
-0.0006875428,
-0.030765526,
0.0246566404,
0.0073303981,
-0.018319143,
-0.0223775003,
-0.0558257438,
0.019841373,
0.062923722,
0.0232707839,
-0.0144194653,
-0.0937356949,
-0.0379088558,
0.0105003547,
-0.051265873,
0.0379263014,
-0.039644368,
0.030355189,
-0.0633350089,
-0.0330319963,
-0.0074676191,
0.017855648,
0.0304298475,
-0.0212455168,
0.0299835708,
0.0002869883,
-0.0198318325,
-0.1019324064,
-0.0069026905,
-0.0298441537,
0.0381518193,
-0.0866878033,
-0.0396035388,
0.0153528908,
-0.0023344725,
-0.0261908155,
0.0000675666,
-0.0987259448,
-0.0637501031,
0.0027408651,
0.0746600553,
0.0253345761,
0.0042742868,
-0.0049401638,
-0.0257064477,
0.0586434864,
0.0330444463,
-0.0488095954,
-0.0061425678,
0.0554467328,
-0.0163010769,
0.0081639737,
0.014420175,
-0.0154187633,
-0.014783578,
0.0026948927,
0.0601878576,
0.0102928886,
0.020920489,
0.0094709843,
-0.0230617635,
0.0189151783,
-0.0250164792,
-0.0198344514,
0.0207102336,
0.0445160158,
-0.0133333625,
0.0446768962,
-0.0344428755,
-0.0039399965,
0.0033825648,
-0.009790834,
-0.018324567,
-0.0072249318,
-0.0302588772,
0.0125315292,
0.0789734274,
0.0386975259,
-0.0040735202,
0.0182294529,
0.0667474121,
-0.0343345962,
-0.0196005385,
-0.021956047,
0.0301842112,
-0.0157845411,
-0.0358188599,
0.0089550717,
0.0067813271,
0.0382757671,
-0.0144725991,
0.0398621745,
-0.0294688568,
0.0183617603,
-0.0360150114,
0.0402638093,
0.0204542968,
-0.0272032674,
0.0398320481,
0.0020779781,
0.0230045803,
-0.0233005621,
0.0450221933,
-0.0098520033,
-0.03965763,
0.043305777,
0.021775471,
-0.0499665476,
0.0247082375,
0.0700150877,
0.0670408979,
0.0466264188,
-0.0196801983,
0.0118988752,
-0.0397306047,
-0.0202796143,
-0.0114666494,
0.0103821615,
-0.0487186164,
0.0143524157,
-0.0620835237,
-0.0901077613,
0.0173243918,
0.0064465939,
0.0357248001,
0.0050963298,
0.0167000405,
0.0459955595,
0.0045438991,
-0.0323876664,
-0.033300031,
-0.0269586053,
0.0487632453,
-0.0455473997,
-0.0056951349,
-6.803069965e-33,
-0.0004652802,
-0.0017674372,
-0.0029650577,
-0.0380345657,
-0.0164132267,
-0.0109813055,
-0.042297367,
-0.0375038832,
0.006299797,
-0.0349401534,
-0.0357341431,
0.0115725119,
0.025051428,
0.0534324609,
-0.003891045,
0.0197817888,
0.0844171718,
0.0168743972,
-0.0209691022,
0.010904762,
0.0017837652,
0.0562999547,
0.0031981214,
0.0232058484,
0.0073838248,
-0.0401842967,
-0.0015880573,
-0.0129924435,
-0.0050303317,
0.0405071564,
-0.0325465761,
-0.0012787888,
0.0272880271,
0.0171173979,
-0.019162422,
0.0162846707,
-0.1116263717,
0.0298629273,
0.0350049995,
0.0113148866,
0.0182768367,
-0.040577393,
0.0080670994,
-0.000605333,
-0.0152954431,
0.0272161271,
-0.0047860998,
0.00241596,
-0.0525031537,
-0.0202549957,
-0.0479468964,
0.0207566861,
0.005872352,
0.0049939863,
0.0462268963,
0.0224406142,
-0.0175492894,
-0.0021771835,
-0.0231833104,
0.0316595621,
0.03016055,
0.0776447207,
0.0367849655,
0.0247110687,
-0.0090389615,
-0.0013539261,
0.0353451893,
-0.052096542,
-0.0283461642,
0.0131186247,
0.0430109352,
-0.0196484532,
-0.0035808235,
-0.0416910052,
0.0509895273,
-0.0694985539,
-0.0582508966,
-0.0261945035,
0.0322669931,
0.0648337677,
0.0087214774,
0.0318089463,
-0.0532324314,
-0.0101563865,
0.0009834408,
-0.0226358958,
-0.0293965563,
-0.0319650993,
0.045754008,
0.0452100746,
-0.022197444,
-0.0080662528,
0.0537369736,
-0.0185284317,
-0.0085743647,
-0.0005240742,
0.0242844597,
-0.0018651444,
-0.0253125522,
0.018937394,
-0.0313654579,
-0.051359985,
0.0065791421,
-0.0099025797,
-0.0028448687,
-0.00352773,
0.0113970526,
0.0511906333,
-0.0891630054,
-0.0042235251,
-0.0382400304,
0.0375794396,
-0.0037831627,
0.0272821523,
0.0358694084,
0.0077132317,
0.0141093647,
0.0089064231,
-0.0502010919,
-0.0150588611,
0.0037152932,
0.0127940709,
-0.0019577055,
0.0493156426,
-0.0331956968,
-0.0012214268,
-0.0530957356,
0.0492590219,
-0.0038902622,
-0.0542881787,
-0.0122611467,
0.059411291,
2.797e-7,
0.0124016162,
0.0551212803,
0.0878224373,
-0.025564665,
0.00691573,
0.0175809748,
-0.0270279292,
0.0148039237,
-0.0113922171,
0.0310194511,
-0.0232450422,
0.0025710773,
0.0694564059,
-0.0447819903,
-0.1152485088,
-0.00070249,
-0.0396511108,
0.0109391883,
-0.063023679,
-0.007041438,
0.01781011,
0.1137556285,
-0.0083950562,
-0.0135173481,
0.0003901955,
-0.0170134567,
0.0307716057,
0.0125743039,
0.0154519444,
0.000701639,
0.0119169932,
0.0414403602,
-0.0051047662,
-0.0358067304,
-0.0307819527,
-0.0050530373,
-0.0131043866,
0.0710258111,
-0.0179010686,
-0.0422643684,
0.0375220291,
0.0006077976,
-0.0281567462,
-0.0053297183,
0.069328621,
-0.0039491234,
-0.0037439975,
-0.0540216006,
-0.0321555212,
0.0565060861,
-0.041434586,
0.0128522906,
-0.0212878902,
0.0573647767,
0.0444315001,
0.006140973,
0.0157686919,
-0.0583562329,
-0.0337515697,
-0.0581278875,
-0.0439164452,
0.0112744393,
-0.0544641875,
0.0387923419,
0.0220799744,
-0.0310574602,
-0.0053836643,
2.82542488e-34,
0.0664878637,
-0.0351871327,
0.0112378234,
0.0597960018,
-0.0214778595,
-0.0092911897,
0.0451077484,
0.054514844,
-0.0292161163,
-0.0413282812,
-0.0435784832
] |
http://arxiv.org/abs/2401.13649v2 | http://arxiv.org/pdf/2401.13649v2 | VisualWebArena: Evaluating Multimodal Agents on Realistic Visual Web Tasks | 2024-01-24T18:35:21 | 2024-06-06T02:01:09 | Jing Yu Koh, Robert Lo, Lawrence Jang, Vikram Duvvur, Ming Chong Lim, Po-Yu Huang, Graham Neubig, Shuyan Zhou, Ruslan Salakhutdinov, Daniel Fried | cs.LG | cs.LG, cs.CL, cs.CV | Autonomous agents capable of planning, reasoning, and executing actions on
the web offer a promising avenue for automating computer tasks. However, the
majority of existing benchmarks primarily focus on text-based agents,
neglecting many natural tasks that require visual information to effectively
solve. Given that most computer interfaces cater to human perception, visual
information often augments textual data in ways that text-only models struggle
to harness effectively. To bridge this gap, we introduce VisualWebArena, a
benchmark designed to assess the performance of multimodal web agents on
realistic \textit{visually grounded tasks}. VisualWebArena comprises of a set
of diverse and complex web-based tasks that evaluate various capabilities of
autonomous multimodal agents. To perform on this benchmark, agents need to
accurately process image-text inputs, interpret natural language instructions,
and execute actions on websites to accomplish user-defined objectives. We
conduct an extensive evaluation of state-of-the-art LLM-based autonomous
agents, including several multimodal models. Through extensive quantitative and
qualitative analysis, we identify several limitations of text-only LLM agents,
and reveal gaps in the capabilities of state-of-the-art multimodal language
agents. VisualWebArena provides a framework for evaluating multimodal
autonomous language agents, and offers insights towards building stronger
autonomous agents for the web. Our code, baseline models, and data is publicly
available at https://jykoh.com/vwa. | VisualWebArena: Evaluating Multimodal Agents on Realistic Visual Web Tasks: Autonomous agents capable of planning, reasoning, and executing actions on
the web offer a promising avenue for automating computer tasks. However, the
majority of existing benchmarks primarily focus on text-based agents,
neglecting many natural tasks that require visual information to effectively
solve. Given that most computer interfaces cater to human perception, visual
information often augments textual data in ways that text-only models struggle
to harness effectively. To bridge this gap, we introduce VisualWebArena, a
benchmark designed to assess the performance of multimodal web agents on
realistic \textit{visually grounded tasks}. VisualWebArena comprises of a set
of diverse and complex web-based tasks that evaluate various capabilities of
autonomous multimodal agents. To perform on this benchmark, agents need to
accurately process image-text inputs, interpret natural language instructions,
and execute actions on websites to accomplish user-defined objectives. We
conduct an extensive evaluation of state-of-the-art LLM-based autonomous
agents, including several multimodal models. Through extensive quantitative and
qualitative analysis, we identify several limitations of text-only LLM agents,
and reveal gaps in the capabilities of state-of-the-art multimodal language
agents. VisualWebArena provides a framework for evaluating multimodal
autonomous language agents, and offers insights towards building stronger
autonomous agents for the web. Our code, baseline models, and data is publicly
available at https://jykoh.com/vwa. | [
0.0340478532,
0.032782007,
-0.0129471757,
0.0306002665,
-0.0351917222,
0.0203263257,
0.0340999588,
0.0190744083,
0.0485412776,
-0.0814056471,
-0.0451428108,
-0.0290208366,
-0.0276575498,
0.0551307201,
0.0182746314,
-0.0741575584,
-0.0262054205,
-0.0357528105,
0.0099324407,
0.0131865721,
0.0250117369,
-0.0253366437,
-0.0622094721,
0.0315657184,
-0.0340728387,
-0.0242674872,
-0.0563919991,
0.0211362094,
0.0123703619,
-0.0326072797,
-0.026464358,
0.0108737526,
0.0452808551,
-0.0116501711,
0.0000020875,
-0.0488592461,
0.0428214371,
-0.0200452134,
0.0182252657,
-0.016251415,
0.0304160025,
-0.0016504079,
-0.0245560948,
0.0008164975,
-0.0327951983,
0.0397459827,
0.0397996567,
0.0457239896,
0.0514350832,
0.0531013273,
-0.0194692723,
0.0417032838,
0.0111106979,
0.0381877162,
0.0306448229,
-0.0445441194,
-0.0205567684,
-0.0039282031,
0.0404041559,
-0.0128522255,
-0.0283952542,
0.025371084,
0.004978315,
-0.0092466734,
-0.0168562159,
0.0503075495,
-0.0241447818,
-0.0690812021,
-0.0487354398,
0.0886420086,
-0.0093125179,
0.0344960541,
0.0098180613,
0.0058981264,
-0.0095737539,
0.0176083334,
-0.0491673127,
0.0477933139,
-0.0161393117,
-0.0164769627,
-0.0681861788,
0.0094613787,
0.0032211898,
0.0161219984,
-0.0209305324,
0.0481878556,
0.0108948229,
0.0014400086,
0.0188122895,
-0.0008420701,
0.0087779975,
-0.0644095466,
0.0026426399,
0.0554118752,
-0.0310506355,
0.028380895,
-0.0244807843,
-0.0015614032,
0.0217269771,
-0.0567735434,
0.0752882436,
0.0439956449,
-0.0581530072,
0.0751910731,
-0.0295836534,
0.0904267132,
-0.0154337361,
0.0390047655,
0.0034368543,
0.0275932141,
-0.0498387925,
0.0133289602,
0.0151471794,
-0.0186822154,
-0.0455684289,
-0.0449092127,
0.010242803,
0.0674042329,
-0.0190154146,
0.0538173057,
-0.0345772877,
0.0132291699,
-0.0042945906,
0.0059954966,
-0.0891548619,
0.001721305,
-0.0450044125,
0.0085500246,
0.0072311242,
-0.0728428885,
0.0332525894,
0.0352481157,
0.0373748913,
-0.0243159812,
-0.0013354917,
0.1338303238,
-0.0280602183,
-0.0353198238,
-0.0418585613,
-0.034254659,
-0.0094288383,
-0.0460263342,
0.0277476534,
-0.0415545441,
0.0641428903,
0.0006601163,
-0.0217436366,
-0.0260176044,
0.0188015029,
0.0329941437,
0.0319318622,
0.1024236158,
-0.0027439333,
-0.0150777521,
0.0068718242,
0.0133259213,
-0.0472348556,
0.0511309355,
0.0297040697,
0.1064568833,
0.0508191958,
0.0280413888,
0.0419751927,
0.0463950746,
-0.0442812294,
0.0184196066,
0.0072308872,
0.0155471722,
0.0071283299,
0.0461125225,
0.0219450034,
0.0319636166,
-0.0316103585,
-0.0153774833,
0.1024321541,
0.0648106039,
-0.0221347716,
-0.0327377729,
0.0324150734,
-0.0022965018,
0.0230069086,
-0.0153986579,
-0.0688826218,
0.0826289281,
-0.0390856117,
-0.0080105672,
-0.0290528871,
0.0371523127,
-0.0221740827,
-0.0559897088,
-0.0039749318,
-0.0235807523,
0.0547091179,
-0.0185312498,
0.0451040864,
0.0027698742,
-0.0538678803,
-0.0501072146,
0.0006240298,
-0.0731605887,
-0.0071859295,
-0.0581924096,
-0.0146356653,
0.0967638195,
0.0482383519,
-0.0221127402,
-0.0477116257,
0.0121235456,
0.0396568738,
0.0576748326,
0.0008248705,
0.008442184,
0.0105942329,
-0.0242062733,
-0.0797565952,
-0.0386227854,
-0.0045412937,
0.0277603902,
-0.0095233899,
0.000903427,
-0.0197569411,
-0.0462005995,
0.0244605411,
0.0586184375,
0.0195117034,
-0.0064172866,
-0.0631757751,
0.0357526131,
0.017073581,
-0.0116813341,
-0.002760797,
0.001927817,
0.0026607576,
-0.0200916026,
-0.0266483501,
-0.0031918322,
0.0516731814,
0.0042000455,
0.0576865412,
-0.0757618994,
0.0069580334,
0.0562855452,
-0.0125230988,
-0.0047627264,
0.0216864236,
-0.0339485928,
-0.0127553232,
-0.0148117458,
0.0055879196,
0.017589055,
0.0162123553,
0.002556087,
0.0572779514,
-0.010482653,
-0.0426069759,
0.0070760683,
0.0324878395,
0.0122214416,
-0.0001045254,
-0.0153270876,
0.009350366,
0.0020955005,
0.0402624346,
0.0205059163,
0.0177755449,
-0.0529503785,
-0.0233458411,
-0.0926374048,
-0.0236134585,
-0.00700858,
-0.01343353,
0.0011768332,
-0.0191362798,
0.012570329,
-0.0180205051,
-0.0225601085,
0.0640752986,
-0.0258011203,
0.0409286842,
0.0445056595,
-0.0042304038,
-0.0425641313,
-0.0631388798,
-0.0018973991,
-0.0073950063,
0.0150896152,
0.0259862859,
-0.0104743233,
-0.0228190646,
0.0104270903,
-0.0664140955,
0.0415956341,
-0.0319493711,
-0.0714583546,
0.0586227477,
-0.0331873558,
-0.0029245145,
0.0868351683,
0.0139594944,
0.051104296,
0.0437805206,
-0.0194231961,
-0.0452429913,
0.0221771169,
0.0046736659,
0.0148680527,
-0.0497678183,
-0.013258514,
0.0074235634,
0.0581155345,
-0.0145643866,
0.0407519229,
-0.0449795574,
0.0194015596,
0.0204505473,
-0.0424069799,
-0.0462058373,
0.0243478902,
-0.0262149312,
0.0642465875,
-0.0365021005,
-0.0322219357,
0.0292715225,
-0.0366011299,
-0.011035447,
0.0473759286,
0.0113037908,
-0.0186415762,
0.0953561068,
0.0367838629,
0.023067819,
-0.0302558709,
0.0056674741,
-0.055760406,
0.0272640567,
0.0392585509,
-0.0169495698,
0.0328863375,
-0.0601630025,
0.0084877731,
-0.0206084847,
-0.0359588005,
-0.0394792184,
0.0109133711,
-0.0194591712,
-0.0229097791,
-0.0233645551,
-0.0255106501,
-0.0478003398,
0.0102862399,
0.0501612127,
0.0228109807,
0.005339405,
-0.0381681919,
-0.045310989,
-0.0212402306,
0.0298871584,
0.0421921909,
-0.0376554765,
-0.018210765,
0.0094389003,
0.0635999739,
0.0301347151,
-0.0570230857,
-0.0605662689,
0.048187051,
0.0150839379,
-0.0068775746,
0.0037471009,
0.0435203835,
0.0278270636,
-0.034231931,
0.0264821779,
0.0671713278,
0.0273702573,
-0.0150633119,
-0.00830054,
-0.0242830031,
0.0030172882,
0.0490146093,
0.0133238211,
-0.0467401147,
-0.0452872589,
-0.0058375392,
-0.006990951,
0.008490446,
-0.029747786,
0.0095622838,
-0.0714980811,
0.0337091126,
-0.0268162824,
0.0275886767,
0.0232690591,
-0.0329905897,
0.0435765311,
0.0010269126,
0.0050524343,
0.0446246564,
0.0029587985,
-0.032487452,
-0.0016420755,
-0.0120101674,
-0.0325735211,
0.0146179758,
-0.0521696992,
-0.0029306903,
0.0671785846,
0.0394262932,
-0.0048717498,
-0.0680748373,
0.0114218825,
0.0270162653,
0.0223538782,
-0.0124625294,
0.0674210563,
-0.0165709462,
-0.0285977162,
-0.0067087165,
0.0098421834,
-0.0230253506,
-0.0109842587,
-0.0112218,
0.0414524265,
-0.0341623239,
-0.00388262,
-0.0470413305,
-0.0137434145,
-0.0102673611,
0.0297643021,
0.0160871223,
-0.0508289039,
-0.0930516645,
-0.0394922607,
-0.0315375812,
-0.0438872129,
-0.040505901,
0.0016237342,
-0.0026769033,
-0.0315612368,
-0.0964034796,
-0.0596202239,
0.0383037813,
0.0363719352,
-0.0009275353,
-0.049922578,
0.0208300464,
0.034817405,
-0.0401374772,
0.0027277323,
-0.0057952986,
0.0391541496,
-0.022101609,
0.0106716463,
0.0143940942,
-0.0856048539,
0.0030388094,
-0.0038289148,
-0.0482302308,
-0.0117033338,
-0.0246535204,
0.0656566843,
0.0171347875,
-0.0112157101,
-0.0148180984,
-0.0607642829,
0.0055841678,
-0.002515919,
-0.0177645274,
-0.0103365472,
0.0234333817,
0.0336012915,
0.0162614044,
0.0009732489,
-0.0809207857,
0.009435717,
-0.0223205965,
0.037122298,
-0.0086352602,
-0.0030597348,
0.0212706849,
0.0210439023,
0.0003051179,
-0.0037522176,
-0.0058083041,
0.0090658637,
-0.0088378061,
-0.0079510938,
-0.0167486239,
-0.0650086254,
0.0018771944,
-0.0068354006,
0.0259478297,
0.0305184405,
0.0204995945,
-0.0157730412,
-0.0187136233,
-0.0088460734,
0.0170652233,
0.0143554239,
-0.0024003324,
-0.0211033877,
0.0124429353,
-0.0078311954,
0.0248924159,
-0.0245916527,
0.0297141615,
0.0113198152,
0.0602000654,
0.015451733,
-0.0303675104,
-0.0415734239,
0.0555090979,
-0.0410809815,
-0.0095547521,
0.015711911,
0.0362491086,
0.0454797409,
0.0430433378,
0.0567077547,
-0.0451240465,
0.0123765329,
-0.0027803876,
0.0129307369,
0.0159555636,
-0.0116816442,
0.0377427675,
0.0328200608,
0.0261309147,
-0.0295896195,
0.0423591696,
-0.013284022,
0.0338158831,
-0.0217211694,
-0.0197496675,
-0.0490519889,
-0.0424584933,
0.0144334491,
0.0193441343,
0.0010419404,
-0.0401183888,
-0.030033242,
-0.051987648,
0.0160896312,
0.0623962581,
-0.0087377289,
-0.0096705733,
0.0379841626,
0.039608065,
0.0236483943,
0.0081759328,
-0.0343932845,
-0.0448743477,
0.0018577152,
0.0225846302,
-0.0012082737,
-5.969945691e-33,
-0.0066486672,
-0.0432742052,
0.0576173998,
0.0029946191,
0.0092474418,
-0.0144465258,
-0.0316894427,
0.0299779791,
-0.0036227233,
0.0055210805,
-0.0039198943,
-0.0227744486,
0.0079543004,
0.0069444487,
0.0008964688,
-0.0159328524,
0.0011647916,
-0.0046356576,
-0.0537669621,
-0.0285815038,
0.0100329453,
-0.0026668538,
0.002842708,
-0.0480207093,
0.0155062322,
0.0097039398,
-0.0611429401,
0.0011307106,
0.0125875473,
0.076592952,
-0.0474332944,
0.0322319604,
0.0465255268,
0.032421574,
-0.0046991734,
0.0264175609,
-0.05757121,
-0.03479141,
-0.026923364,
0.0427813753,
-0.0936880931,
0.0166069809,
0.0330002941,
-0.0169614814,
-0.0407971554,
0.0024153497,
0.0585680678,
-0.030699499,
0.0151442708,
0.0195151698,
-0.0759823173,
0.0168870743,
0.0007986056,
-0.0488783605,
0.029522296,
0.0427570716,
0.0295323208,
0.014428664,
-0.0385264568,
0.0345154218,
-0.0264689606,
-0.0219667852,
0.0807415694,
0.0231154449,
-0.0054619876,
0.1030230299,
0.0478738099,
0.0104947146,
-0.033844538,
0.0140229277,
0.0377691649,
0.0779195577,
0.0165443625,
0.0672448277,
-0.026709782,
0.0352380723,
-0.0171345528,
0.0664546415,
0.0013084424,
0.0464336202,
0.0089633474,
0.028028829,
-0.0376595184,
0.0271061342,
-0.0190511551,
0.0094860448,
-0.06672436,
-0.027524462,
0.0630066991,
0.0161454976,
0.0098147783,
-0.0328870453,
0.0147340028,
-0.0133536691,
-0.019976683,
-0.005999051,
-0.0334282815,
0.0784175396,
0.0089867981,
-0.0385077,
0.01587498,
-0.0289366823,
0.0307552572,
-0.0002652841,
0.0301529001,
0.0004633525,
-0.0224482082,
0.0287548322,
-0.0303269923,
-0.0234607868,
-0.0567145869,
-0.0230915006,
0.0395581052,
-0.0045012333,
0.0229056943,
-0.0554888137,
-0.0529862717,
0.03196894,
-0.0064878538,
-0.0460548811,
-0.0248945467,
-0.0471234396,
-0.006213767,
-0.0075166253,
-0.0601760596,
-0.0051892744,
-0.0631304085,
0.0294389166,
-0.009247628,
-0.0715422034,
0.0040118513,
0.0121333972,
2.734e-7,
-0.0014049902,
0.0445202962,
-0.0086523984,
0.0404489338,
0.0293175671,
0.0659815893,
-0.0466921031,
0.0194994379,
-0.0382567719,
-0.0679662079,
0.0106553482,
0.0512710102,
0.0224403441,
0.0066939895,
0.0045895209,
-0.003246577,
0.0230632052,
-0.022776315,
-0.0613621324,
0.0125840753,
0.0836190656,
0.0195276607,
-0.0102795567,
-0.0360256322,
0.0350240953,
-0.0481572077,
0.0459021963,
-0.0273911767,
0.000095506,
-0.0195878856,
0.0057159155,
0.0194551498,
-0.0074802539,
0.0046143141,
-0.0058327573,
-0.0272904895,
0.0059368904,
0.0526544489,
-0.0312268306,
-0.0277815573,
0.0075576957,
-0.0160649959,
0.0027551798,
0.0014580337,
0.0781275406,
0.0214911867,
-0.0130663915,
0.0257021971,
-0.0285518989,
-0.0261511561,
-0.018013645,
-0.0440501571,
-0.0438061282,
0.0084609417,
-0.0062173428,
-0.0196082499,
-0.0326572657,
-0.0566320978,
0.0016594641,
-0.0292511582,
-0.0253997538,
-0.0301131103,
-0.0597712211,
0.026800992,
0.0348426178,
-0.0368999168,
-0.048287183,
2.160355201e-34,
0.0053004264,
0.0043327291,
0.0146305356,
0.0191310439,
0.0484201983,
0.0479203351,
-0.0520681553,
-0.0055542719,
0.0263633765,
-0.0289345868,
-0.0214152578
] |
http://arxiv.org/abs/2402.01702v1 | http://arxiv.org/pdf/2402.01702v1 | Fluent dreaming for language models | 2024-01-24T17:57:12 | 2024-01-24T17:57:12 | T. Ben Thompson, Zygimantas Straznickas, Michael Sklar | cs.CL | cs.CL, cs.AI | Feature visualization, also known as "dreaming", offers insights into vision
models by optimizing the inputs to maximize a neuron's activation or other
internal component. However, dreaming has not been successfully applied to
language models because the input space is discrete. We extend Greedy
Coordinate Gradient, a method from the language model adversarial attack
literature, to design the Evolutionary Prompt Optimization (EPO) algorithm. EPO
optimizes the input prompt to simultaneously maximize the Pareto frontier
between a chosen internal feature and prompt fluency, enabling fluent dreaming
for language models. We demonstrate dreaming with neurons, output logits and
arbitrary directions in activation space. We measure the fluency of the
resulting prompts and compare language model dreaming with max-activating
dataset examples. Critically, fluent dreaming allows automatically exploring
the behavior of model internals in reaction to mildly out-of-distribution
prompts. Code for running EPO is available at
https://github.com/Confirm-Solutions/dreamy. A companion page demonstrating
code usage is at https://confirmlabs.org/posts/dreamy.html | Fluent dreaming for language models: Feature visualization, also known as "dreaming", offers insights into vision
models by optimizing the inputs to maximize a neuron's activation or other
internal component. However, dreaming has not been successfully applied to
language models because the input space is discrete. We extend Greedy
Coordinate Gradient, a method from the language model adversarial attack
literature, to design the Evolutionary Prompt Optimization (EPO) algorithm. EPO
optimizes the input prompt to simultaneously maximize the Pareto frontier
between a chosen internal feature and prompt fluency, enabling fluent dreaming
for language models. We demonstrate dreaming with neurons, output logits and
arbitrary directions in activation space. We measure the fluency of the
resulting prompts and compare language model dreaming with max-activating
dataset examples. Critically, fluent dreaming allows automatically exploring
the behavior of model internals in reaction to mildly out-of-distribution
prompts. Code for running EPO is available at
https://github.com/Confirm-Solutions/dreamy. A companion page demonstrating
code usage is at https://confirmlabs.org/posts/dreamy.html | [
0.0215790272,
0.0812172145,
-0.028847592,
0.065169394,
-0.0325542577,
0.0135631198,
0.0083268676,
-0.0178172495,
-0.0005965728,
-0.0294030961,
0.0010751728,
-0.0569403581,
0.0096796406,
0.0655526817,
0.0582688637,
-0.016322542,
0.0328750536,
-0.0450925566,
-0.0329093188,
-0.0268754195,
-0.023726508,
0.0015757956,
-0.0150065478,
0.0722594112,
-0.0111400485,
-0.0302966684,
-0.0208231956,
0.0063585299,
0.0222296808,
-0.0063063605,
-0.0010944769,
0.0563772842,
0.011180399,
0.0451816469,
0.0000025122,
-0.0331288874,
0.0103598526,
-0.0219988078,
0.000114693,
0.0351633057,
-0.0054899845,
-0.0429988541,
-0.0059205093,
-0.0408288091,
-0.031833183,
-0.0093705291,
0.0684716329,
0.0414753146,
0.0407283381,
0.0853853002,
0.0129738897,
-0.0264418349,
0.0521836765,
-0.0326388255,
-0.0045320978,
0.0084462157,
0.0282103438,
-0.058538463,
-0.0159951858,
-0.0058648158,
-0.0210238099,
0.0215424262,
0.0284774825,
-0.0291621685,
-0.0046207341,
0.0351905748,
0.0320760161,
-0.0422649607,
-0.0091652023,
0.0359010287,
-0.0524751395,
0.0020660469,
-0.0007830385,
-0.0108211143,
0.0135465926,
-0.0566777177,
-0.0361626819,
-0.0048816642,
0.008652431,
0.0422533676,
0.0282455496,
-0.0385633931,
0.0236798692,
-0.0189242009,
-0.0230722427,
0.0064985869,
0.0202286486,
0.0066309394,
-0.0033302486,
-0.0173257794,
-0.0533194616,
-0.034260612,
0.0220824517,
0.0005503263,
-0.027823044,
0.0077903979,
-0.0536131375,
-0.0630051345,
-0.0013999657,
0.0237213019,
0.1181767136,
0.029940702,
-0.010471168,
0.0121037103,
-0.0048113889,
0.0744191781,
-0.032067325,
0.0687335059,
-0.0254332852,
0.0261541102,
-0.0062094945,
-0.0113991546,
-0.0026215606,
-0.0255490504,
0.0243931785,
-0.0067325383,
-0.0127861612,
0.0204990245,
-0.0272416789,
0.0023778302,
-0.0220966581,
0.0177702326,
-0.0364602394,
-0.0236435849,
-0.0269279666,
-0.0263191368,
-0.0074346978,
-0.0117179183,
0.0130596794,
-0.0548072904,
-0.0463039018,
0.0140588451,
-0.0201495513,
-0.0149831679,
-0.051462099,
0.0934102908,
0.0118486062,
-0.0269296318,
-0.0038785464,
-0.0084532546,
-0.0230646785,
0.0050625247,
0.001149505,
-0.0751892403,
-0.0081391297,
-0.03452814,
-0.0150632672,
-0.0524223819,
0.0174557306,
-0.0020355152,
-0.0046324017,
0.0537617058,
-0.0252461527,
-0.0220915824,
0.0281584971,
0.0065839761,
-0.0017941205,
0.0679639131,
-0.0006122716,
0.0503038242,
-0.0044167265,
0.0053905277,
0.0373757407,
0.0779954121,
0.0191976763,
0.0268041771,
-0.0248504076,
-0.0231291447,
0.0762209371,
0.0427440368,
0.0306490436,
0.0196045302,
-0.0386690907,
-0.0885944888,
0.0624887608,
0.0426062979,
0.0115663633,
-0.0013572264,
0.0151994694,
0.0145383934,
-0.0108254729,
-0.0053698309,
-0.0257823579,
0.0345853679,
-0.0219952036,
-0.0111604137,
-0.0160101019,
0.0328348763,
-0.0077752601,
-0.0432914309,
-0.0076082945,
0.0023576613,
-0.0388450362,
0.0051025259,
0.0364428684,
0.0311785173,
-0.0328540467,
-0.0032074593,
0.0298692249,
0.0062959916,
-0.0112333447,
-0.0676160753,
0.014655591,
0.0275903456,
0.0186193474,
-0.028574232,
0.0105765332,
-0.0120644886,
0.0087228911,
0.0500673316,
0.0316783935,
0.0037497282,
0.0308826827,
-0.012263407,
-0.0306280591,
-0.0412774533,
0.0420508347,
0.0399085544,
-0.0274079796,
0.0117674842,
-0.001464272,
-0.0034697109,
-0.0488195196,
0.0448080227,
-0.0152687673,
0.0015183642,
0.0261984859,
0.0010245997,
-0.0053624269,
-0.0012510307,
-0.0202409942,
0.0298962928,
0.025768606,
-0.0525060557,
0.012161863,
-0.000668795,
-0.0366185345,
-0.0175236519,
0.0512136519,
-0.0170664974,
-0.0033078447,
0.0266243033,
-0.0384144112,
-0.0091677513,
0.0169950947,
-0.0497376136,
0.0252966806,
-0.0385045782,
-0.0076569039,
-0.0401236787,
0.0371330604,
-0.0443942547,
0.0678861961,
-0.0083508082,
0.003844551,
0.0273906402,
0.0191774685,
0.0252726376,
-0.0500442535,
-0.0060774125,
0.0272075199,
-0.000977813,
0.0086743757,
0.0139011536,
-0.0012659308,
-0.0356519446,
-0.0187019594,
-0.1010829508,
-0.0105926702,
-0.0238098428,
-0.0143127376,
0.0550430641,
-0.0233700685,
-0.0271569174,
-0.0064963312,
-0.0230272338,
0.0584552661,
-0.0558919348,
0.0210782979,
0.0135978768,
0.0003980748,
-0.0348736346,
-0.0246750806,
-0.0268470179,
0.0021093613,
0.0569591224,
-0.0014765937,
-0.010031594,
-0.0818858072,
-0.0065005957,
0.0155484844,
0.0192538258,
-0.0164268874,
-0.0215919409,
0.035269253,
-0.038591776,
0.0192752946,
0.0772829652,
0.0254925415,
0.0617144816,
0.0765089095,
0.015206567,
-0.0192828905,
-0.0228355173,
-0.0206864402,
0.01701889,
-0.0164070558,
-0.0355471559,
0.040451657,
0.0841367841,
0.0467854738,
-0.029038053,
-0.0254174825,
0.0583347604,
-0.0206364542,
-0.0387666598,
-0.0679376721,
0.0159188416,
-0.0434789099,
0.0830004439,
-0.013464774,
-0.0515449718,
0.0948260278,
-0.0245784707,
0.0194911081,
-0.0232533216,
-0.0006220679,
0.000387087,
0.0671337098,
0.0230280068,
0.0178497396,
0.0300512034,
-0.0292453561,
-0.0556959547,
0.0177475549,
-0.0117376111,
-0.0250661168,
0.0217224006,
-0.034186542,
-0.0278085936,
-0.0050793169,
-0.0233182516,
-0.0574206412,
-0.0415086187,
0.0040631234,
-0.0042728879,
-0.0051449304,
-0.0152442893,
-0.0268643107,
-0.0358822159,
0.0046102819,
0.0395304039,
-0.0419420823,
-0.0018399234,
-0.0079301922,
0.0355436951,
0.0805786178,
0.0389755256,
0.0033015572,
-0.0390233658,
0.0221195202,
0.0488528982,
-0.0345103256,
0.063828446,
-0.0533095971,
0.0248508714,
0.0321966186,
0.0338986404,
0.005912303,
-0.0071121664,
0.0527172647,
-0.0447001606,
-0.0127259642,
0.080514729,
0.0220199395,
-0.0465510674,
0.0337720625,
-0.0382622667,
0.0693868697,
0.0374295078,
0.0439403281,
-0.0669967383,
-0.0036633536,
-0.0132357096,
0.031424582,
0.0011549733,
0.0030312655,
-0.008134014,
-0.0815310255,
0.0414529294,
0.0069883945,
-0.0113657918,
-0.0138209816,
-0.006238468,
0.0201767143,
0.0039882455,
0.0151230618,
-0.0001833597,
0.0202248693,
0.0165921822,
-0.0237230938,
0.07802549,
-0.0358387008,
0.0152286617,
-0.0295878444,
0.0424076878,
0.0981703028,
0.0169033725,
-0.0090751275,
-0.0521531925,
0.0680660829,
0.0014695232,
0.0091227796,
0.0203298405,
0.0166464858,
0.0411074422,
-0.039420303,
0.0184348337,
0.0550149381,
-0.0025261599,
-0.0207436662,
0.0009480041,
0.0133522153,
-0.0232982542,
-0.0256119668,
-0.0355593972,
-0.0001269738,
0.0119386446,
0.0349224582,
0.0780816004,
-0.0949679464,
-0.007018615,
-0.0693099573,
-0.0005429058,
-0.0299758799,
-0.0736344531,
0.0303075947,
0.0851057544,
-0.0600927807,
-0.0653468892,
-0.0121383062,
0.0191420633,
0.0299020037,
0.0159503762,
-0.0267120413,
-0.0752523616,
-0.0365705267,
-0.0273611359,
0.029741222,
0.0062068626,
0.0093986718,
-0.03983064,
-0.0266813803,
0.0206269398,
-0.0630858094,
-0.0200552996,
0.0234893765,
-0.1066427976,
-0.0901044607,
0.0078332536,
0.0423860662,
0.0093528861,
0.043281626,
-0.0415820107,
-0.0677025318,
0.0794587284,
0.0830977485,
0.0423299968,
0.0228170455,
0.0234495234,
-0.0425416231,
-0.0094124386,
0.0323615409,
-0.0097279334,
-0.0111558838,
0.0449381024,
-0.0021853587,
0.0258023534,
-0.0152593441,
0.0217642039,
-0.0213194396,
0.0793062076,
-0.0310505014,
-0.0395298116,
-0.0575594902,
-0.0255341474,
-0.0035684479,
-0.0049819858,
-0.0305940118,
-0.015904326,
0.0585615896,
-0.0206578113,
0.0468032658,
-0.0093219215,
0.0218990482,
-0.0376303531,
-0.0489133894,
-0.0049599516,
0.0212650057,
0.0001476363,
0.030898191,
0.0184891112,
-0.0327952281,
0.0738936067,
-0.0525570475,
0.0216159262,
-0.016190147,
0.0368866287,
0.0177771579,
-0.0186490286,
-0.0114105092,
0.036698442,
-0.0519525371,
-0.0884633362,
-0.0180595629,
0.0398913212,
0.0519586243,
-0.0169577636,
0.0490839034,
-0.0316017419,
-0.0086193262,
-0.0137978438,
0.0023047538,
0.0100051826,
-0.0529678576,
0.0522969477,
0.0189558752,
0.0142666912,
-0.0272159763,
0.0639736205,
-0.0074841725,
0.0372736901,
0.0064367936,
0.0387009457,
-0.019717183,
0.0152834086,
0.0230212435,
0.0016662098,
-0.0368777104,
0.0068453578,
-0.0113112358,
-0.0552005842,
-0.0022938338,
0.028781306,
-0.0257154219,
0.0480383672,
0.0345042907,
0.0371334367,
-0.0224021822,
0.0187845659,
-0.0365833417,
0.0160100851,
0.0269871242,
-0.07794898,
-0.0012157862,
-7.189783362e-33,
-0.0031976476,
-0.036556825,
-0.0099373236,
0.010703926,
-0.0021105856,
-0.0208954532,
-0.0378336832,
0.0234821104,
-0.0195473228,
0.0057172501,
-0.0065213204,
-0.0048205154,
-0.0095740501,
0.0263069291,
-0.0134539464,
-0.0122466637,
-0.0038447054,
-0.0179612245,
-0.0234863963,
0.0096190562,
0.0125161996,
0.0052523711,
0.0458025001,
-0.035627488,
0.030092679,
0.0140876193,
-0.041910965,
0.0299725141,
-0.0349871665,
0.0302140843,
-0.0430858582,
0.0490291417,
0.0048419861,
-0.0255428683,
-0.0381841436,
-0.0294862445,
0.0075728116,
-0.0376130827,
-0.0086459769,
0.018228244,
-0.0616293959,
-0.0499349423,
0.0270651393,
-0.0237052571,
-0.0156290364,
-0.0296660792,
0.0430952758,
-0.0323275775,
-0.0224761143,
0.0137666576,
-0.0562594905,
-0.012921405,
-0.0168742705,
-0.0135660851,
0.0168056488,
0.0029427453,
0.0355580822,
-0.0125180101,
-0.0482106656,
0.0268358607,
0.0251142569,
0.0429071896,
0.0577595532,
0.0454394482,
0.0293955468,
0.088313587,
0.0544201098,
-0.0179454461,
-0.0061560078,
0.0376544856,
0.0669378564,
-0.0017648613,
0.0468898676,
0.0167467967,
0.0420387462,
-0.0310908873,
-0.027113298,
0.0008899733,
0.006423445,
-0.0287464168,
0.0735865831,
0.0213475469,
0.0160257295,
-0.0311830249,
-0.0472802222,
-0.0650603771,
-0.0626990199,
0.0045916331,
0.0139409928,
-0.0032043352,
-0.0311900899,
-0.0249379221,
-0.0116476091,
-0.0179125257,
0.0064491741,
0.0290413518,
-0.0501338877,
0.0126287509,
0.0090462854,
-0.0036877033,
0.0141107924,
0.0042446963,
0.0108613325,
-0.0100804511,
0.0299889743,
0.0352242962,
0.0062757935,
0.006883326,
-0.0207363311,
0.0034484763,
-0.0612347908,
0.009638669,
0.0489922501,
0.0432057492,
0.0269518998,
-0.0319511071,
-0.0446614698,
0.0221008714,
0.0242152438,
-0.00167356,
-0.0418287814,
0.0583057702,
-0.0331126601,
0.0465247892,
-0.011955115,
-0.0113627492,
-0.0247461032,
0.0217208397,
0.0428380817,
-0.0595182516,
-0.0399574339,
-0.0073753004,
3.072e-7,
0.034160614,
0.0318870842,
-0.0128897205,
0.0222043656,
0.040223293,
0.0450158454,
-0.0282607041,
0.0551075973,
0.0150983613,
-0.0770499259,
-0.0379954502,
0.0076027117,
0.0351377837,
0.0242813006,
-0.1131091192,
0.0596765019,
-0.0362182744,
0.0396406315,
-0.0687805638,
-0.008257553,
0.0472598933,
0.1005999222,
0.0113385879,
-0.0242919475,
0.0318431295,
-0.0349960662,
0.0593887679,
-0.0335314162,
0.0390644446,
-0.0034199357,
-0.0003051567,
0.0174691975,
0.0186614376,
-0.0218681712,
-0.0193865243,
-0.0074053784,
-0.0142652206,
0.025499424,
0.0051326011,
-0.0317176469,
-0.0076995683,
0.0334573574,
-0.0058273263,
-0.0866733044,
0.0776172131,
-0.0248249751,
-0.0167287532,
-0.05052137,
0.0647321418,
-0.0384589694,
0.0564457551,
-0.0927715674,
-0.0145632466,
0.0230210572,
0.0101054329,
-0.0166107137,
0.0023618534,
-0.0147801638,
-0.0352275856,
-0.0354369432,
-0.0495845936,
-0.0000768845,
-0.0463147797,
0.0409273468,
0.0027392155,
-0.0271293242,
-0.0056915404,
3.026521428e-34,
0.0268026721,
-0.0147418473,
0.0248308107,
-0.0000142907,
0.0331297442,
-0.0028458943,
0.0711647123,
-0.0115761589,
0.0011981877,
-0.0526763424,
-0.072169967
] |
http://arxiv.org/abs/2401.13621v1 | http://arxiv.org/pdf/2401.13621v1 | DenoSent: A Denoising Objective for Self-Supervised Sentence Representation Learning | 2024-01-24T17:48:45 | 2024-01-24T17:48:45 | Xinghao Wang, Junliang He, Pengyu Wang, Yunhua Zhou, Tianxiang Sun, Xipeng Qiu | cs.CL | cs.CL | Contrastive-learning-based methods have dominated sentence representation
learning. These methods regularize the representation space by pulling similar
sentence representations closer and pushing away the dissimilar ones and have
been proven effective in various NLP tasks, e.g., semantic textual similarity
(STS) tasks. However, it is challenging for these methods to learn fine-grained
semantics as they only learn from the inter-sentence perspective, i.e., their
supervision signal comes from the relationship between data samples. In this
work, we propose a novel denoising objective that inherits from another
perspective, i.e., the intra-sentence perspective. By introducing both discrete
and continuous noise, we generate noisy sentences and then train our model to
restore them to their original form. Our empirical evaluations demonstrate that
this approach delivers competitive results on both semantic textual similarity
(STS) and a wide range of transfer tasks, standing up well in comparison to
contrastive-learning-based methods. Notably, the proposed intra-sentence
denoising objective complements existing inter-sentence contrastive
methodologies and can be integrated with them to further enhance performance.
Our code is available at https://github.com/xinghaow99/DenoSent. | DenoSent: A Denoising Objective for Self-Supervised Sentence Representation Learning: Contrastive-learning-based methods have dominated sentence representation
learning. These methods regularize the representation space by pulling similar
sentence representations closer and pushing away the dissimilar ones and have
been proven effective in various NLP tasks, e.g., semantic textual similarity
(STS) tasks. However, it is challenging for these methods to learn fine-grained
semantics as they only learn from the inter-sentence perspective, i.e., their
supervision signal comes from the relationship between data samples. In this
work, we propose a novel denoising objective that inherits from another
perspective, i.e., the intra-sentence perspective. By introducing both discrete
and continuous noise, we generate noisy sentences and then train our model to
restore them to their original form. Our empirical evaluations demonstrate that
this approach delivers competitive results on both semantic textual similarity
(STS) and a wide range of transfer tasks, standing up well in comparison to
contrastive-learning-based methods. Notably, the proposed intra-sentence
denoising objective complements existing inter-sentence contrastive
methodologies and can be integrated with them to further enhance performance.
Our code is available at https://github.com/xinghaow99/DenoSent. | [
0.0041694674,
0.0573034063,
0.0011516338,
0.0591537431,
-0.0438130014,
0.0051157749,
0.0063374406,
-0.0145176277,
0.0132168876,
-0.032159958,
-0.0321584828,
0.0300599691,
-0.0356359705,
-0.0181573518,
0.0169935226,
-0.0588810965,
0.0478449464,
0.0094818762,
0.0391341634,
0.0106307892,
0.0508807525,
0.009945889,
0.0118413195,
0.0105017349,
-0.0352407098,
-0.0274562221,
0.0329243615,
0.0151058063,
0.0386202075,
-0.0303185228,
0.0755229592,
0.0260098875,
0.027022155,
-0.0209500529,
0.0000018187,
-0.0377440192,
-0.0355838016,
-0.0266481806,
0.0302478243,
0.00040671,
0.069155857,
-0.0292517729,
0.0427493788,
-0.017471144,
-0.0443629175,
-0.0083302697,
0.0332543515,
0.0854481533,
0.0631175935,
0.0685006231,
-0.0408399068,
-0.049442865,
-0.0192745123,
-0.0159955733,
0.0072137769,
-0.0094551994,
0.0088269114,
0.0388152786,
-0.0339083672,
0.0284010861,
0.0030278838,
0.0285588522,
0.0095109297,
0.0002476108,
0.051083073,
0.0136461584,
-0.0044750934,
0.0075300201,
-0.0200886317,
0.0835044682,
-0.0259858388,
-0.0286110267,
0.0002070256,
-0.0033967774,
0.0072694155,
-0.0174715482,
-0.0457624793,
-0.0019129403,
0.0253604725,
0.0734724626,
0.0126254195,
0.0568060651,
0.0048358836,
0.0027370397,
-0.0087955808,
-0.0157796983,
0.0029284039,
-0.0182692148,
0.0001759009,
0.0150922919,
-0.0347206518,
0.0163709112,
0.0313376188,
0.0109680407,
-0.0170636196,
-0.0327432565,
-0.0459346659,
-0.056326855,
0.0060553132,
-0.0118201412,
0.0253207535,
0.0188008212,
-0.0030487832,
0.0225894209,
-0.0414078534,
0.0205973983,
0.0102857184,
0.0188597515,
-0.0609746948,
-0.0246816184,
-0.0696505532,
0.0058358395,
-0.0392823406,
0.0731823966,
-0.0412883759,
-0.0346197188,
-0.0409027301,
-0.0057126656,
-0.0189562067,
-0.0224236567,
0.0087558413,
0.0298670884,
-0.0539135151,
-0.0071528852,
0.0022810197,
0.021156827,
-0.0278666932,
0.031415686,
0.0222254004,
0.0015898438,
-0.053079024,
0.0285279732,
0.0494757667,
-0.0317083783,
-0.0100215757,
0.0479743034,
-0.0368491299,
-0.0176146347,
-0.0331870168,
-0.0026505946,
-0.0348165445,
-0.0211561453,
-0.0329835266,
-0.067357026,
0.0068567824,
-0.0381502919,
0.0109927757,
-0.0854530111,
0.0066378405,
0.0227172524,
-0.026310958,
0.0641144216,
-0.0167438071,
-0.0474068224,
-0.0451451875,
0.014164363,
0.0685500205,
0.0310814939,
0.0068877162,
0.0077009797,
0.0034681745,
0.0342969,
-0.0264737699,
-0.0103330445,
0.0155748911,
-0.0033402713,
-0.0703291893,
0.0118955225,
0.0332940519,
0.0204519816,
0.0674528331,
0.0671008527,
-0.0258090235,
-0.0314214043,
0.0510941818,
0.0110207098,
0.022266211,
0.0082632881,
0.0508493111,
0.0470841378,
-0.0421773344,
0.0294707548,
-0.0310918,
0.0126345651,
-0.0081628636,
-0.0044473526,
0.0254227389,
0.0017063379,
-0.0410235971,
-0.020749772,
0.0380697101,
-0.0432719849,
-0.0054245517,
-0.0178186223,
0.0152877765,
0.0832287073,
-0.0259229783,
0.045676887,
0.0515260026,
-0.0444119424,
-0.0353280082,
0.0051024086,
0.0135802813,
0.0293366145,
-0.0000920133,
-0.043747481,
0.010315231,
-0.0103791952,
-0.0331478156,
0.0408252552,
0.0246954579,
0.0007618122,
0.0543409027,
-0.0548976474,
-0.0233213492,
-0.0714505464,
0.0191905871,
-0.0133108823,
-0.0144000333,
0.0212346241,
0.0503746346,
-0.0665602684,
-0.0245060939,
0.0479814783,
0.0164769422,
-0.0471060239,
0.0444139913,
-0.0150050549,
-0.0006173762,
0.0142250406,
-0.0096740015,
0.0147693595,
0.0189373624,
-0.0522309914,
-0.0073040463,
-0.0169587154,
-0.1058133766,
-0.1262182742,
0.0469910167,
-0.0604106113,
-0.0170976911,
0.0685711727,
-0.055143293,
-0.0163777638,
0.0297607593,
-0.0183442812,
-0.011896315,
-0.0256919023,
-0.0284944735,
0.048960045,
0.0576997921,
-0.0323031507,
0.0611826591,
0.0196436029,
-0.0057224459,
-0.0199861154,
0.001028448,
0.0194683578,
-0.0155898081,
0.0173747092,
-0.0695319101,
-0.0019185352,
0.0190206263,
0.0348410644,
0.0181753039,
-0.0426167883,
-0.0115154739,
-0.055676017,
0.0130306762,
0.0063921874,
-0.0025657692,
0.0642458797,
0.0405417532,
0.0140801445,
-0.0524844043,
-0.0329219513,
0.0177650042,
-0.0071720947,
-0.0063969572,
0.0897860229,
-0.0373675935,
0.0110826967,
-0.0610389523,
-0.0036727015,
-0.012463457,
0.0508860126,
-0.0335623547,
-0.0166160781,
-0.0245302301,
0.0576826818,
-0.0125866234,
0.0094210114,
0.0224919319,
-0.0354828127,
-0.0168793947,
-0.0403536931,
0.0411640964,
0.0399323702,
0.0170856044,
0.02025567,
0.0342575051,
-0.0181001369,
-0.0427425504,
0.0170895327,
-0.0277094208,
0.0646675825,
-0.0217915494,
-0.0155791584,
0.0176725425,
0.0680376142,
0.0138492454,
-0.0011818744,
-0.006719803,
0.0059348647,
0.0187990572,
-0.0220314804,
-0.0153261768,
0.0293496158,
-0.0072887293,
-0.0080183465,
-0.0512989238,
-0.0111603811,
0.0556711853,
-0.0170591865,
-0.0112014813,
-0.0548432916,
-0.0502995923,
-0.0231627822,
0.0663489774,
-0.0064392937,
0.0114422161,
0.0381473191,
0.0212814044,
-0.0418701135,
-0.0386672132,
0.0266545452,
0.0260714274,
0.0091716582,
-0.0089608319,
0.0075282021,
-0.0124415895,
-0.0364038236,
-0.0248514749,
-0.0627836511,
-0.0160757974,
0.0115742227,
-0.0102844415,
0.0253694933,
-0.0332460999,
-0.04504117,
-0.0030397056,
0.0389467962,
-0.0134101389,
-0.0200784728,
-0.0329079218,
0.1129794195,
0.0582301244,
0.0239367671,
0.017774161,
-0.0037911711,
0.0551433302,
0.0596477017,
-0.0051489491,
0.0049392181,
-0.0613542534,
0.0363635644,
-0.0156165464,
-0.005607788,
0.0105536999,
0.0061683664,
0.105398275,
-0.0240625236,
-0.0262717456,
0.0992266312,
-0.001528451,
-0.0351502337,
0.0079462742,
0.0205703489,
0.0537484437,
0.0321211666,
-0.0091457218,
-0.0389892012,
-0.0654819012,
-0.022094842,
0.0041048252,
0.017690191,
-0.0239783302,
-0.0094505856,
-0.069639422,
0.0229720604,
0.001820961,
-0.0404631197,
-0.0004098319,
-0.0288339667,
0.0460008532,
-0.0033132264,
0.03294212,
-0.039918419,
-0.0344871692,
0.0028762382,
0.0024424139,
-0.0058059553,
-0.023305485,
-0.0121333757,
-0.0336298309,
0.0495578423,
0.0096059227,
0.0458969586,
0.0184089057,
0.0005127168,
0.0337257087,
-0.0615981594,
-0.0203478616,
-0.012630458,
0.0877172723,
0.0046734489,
-0.0109538864,
-0.0228266958,
0.0263611041,
-0.0748413354,
-0.0321475714,
0.0118414694,
0.0294636879,
0.0120186219,
-0.0306290202,
-0.036930088,
0.0280091651,
0.0209421106,
0.070694536,
0.0575975589,
-0.0089803161,
-0.0680456087,
0.0041255308,
-0.0083016902,
-0.0444211215,
-0.0387696847,
0.004375746,
0.0106746703,
0.0001219052,
-0.0594342537,
-0.0783392563,
-0.0270139873,
0.0071683619,
-0.0239796136,
0.0143153071,
-0.023881495,
-0.041433204,
-0.030212516,
0.0128322011,
-0.0489345044,
-0.0198260713,
0.0424389355,
-0.0120577989,
0.0249386169,
-0.0197066478,
-0.0281794388,
-0.0369279124,
-0.0849250257,
-0.0417864658,
-0.0190901384,
0.0417560786,
-0.0002928373,
0.050926961,
-0.0074267234,
0.0187981762,
0.0566939376,
0.071490325,
-0.0785641447,
0.0003813875,
0.0703077167,
-0.0010668517,
0.0237646755,
0.057904169,
-0.0123061053,
0.0639071986,
0.0480033159,
0.027518874,
0.009099761,
-0.0378265679,
0.0549736731,
-0.0002083304,
0.0400793739,
0.0268643145,
0.0161780156,
-0.0238583628,
-0.0480062477,
-0.0272313897,
-0.0383404717,
-0.0586337037,
0.0312447753,
0.0286391266,
0.0310767256,
-0.0055405949,
0.0249524936,
-0.001877627,
-0.0018476601,
-0.0517594442,
-0.0066444995,
0.030652646,
0.0072119068,
-0.0374242589,
0.0089985728,
-0.0225740708,
0.0055936859,
0.0001027157,
0.0072478433,
-0.0110754706,
-0.0171806868,
0.0393171534,
0.0089359256,
-0.0429257974,
0.0176962353,
-0.0699985251,
0.0422808416,
-0.0226148739,
-0.0439000092,
0.0400093906,
-0.0386792868,
0.0665417612,
0.005293509,
0.013032577,
-0.0075097396,
0.0499215871,
-0.0090144137,
0.0287665501,
-0.0112186968,
-0.0120778754,
-0.0101933926,
0.0230869148,
0.0896406993,
-0.0489851832,
0.0449679755,
0.0277911574,
0.0267278142,
-0.0736389756,
-0.0170063321,
0.0055766227,
0.0355330594,
0.0270536151,
0.0184066854,
0.0511482172,
-0.0571779646,
0.0295613017,
-0.0374207199,
-0.0010771392,
0.0153392544,
0.0630367175,
0.016244296,
-0.000152073,
-0.0185770802,
-0.0087443748,
0.0023843991,
0.043569576,
-0.06257388,
-0.0572863482,
-6.397719105e-33,
-0.007596937,
-0.0023803597,
-0.0443786159,
-0.0067767189,
-0.0108859446,
-0.0157909151,
-0.0695589259,
-0.0168820489,
-0.0041259653,
0.0836421251,
0.0292747691,
0.0298975781,
0.0072057131,
0.0386103801,
-0.0016133778,
0.0115154572,
0.0548903793,
-0.0151352352,
0.016590463,
0.0207826011,
-0.0074886903,
0.0480149761,
0.0484139211,
-0.0913026035,
-0.0116245784,
-0.0071960669,
-0.0145172095,
0.0175135043,
0.0712678954,
0.033246655,
-0.0270597041,
0.0520079061,
0.032613188,
-0.0089179929,
-0.0337868929,
0.0149069028,
-0.0894628987,
-0.080658488,
-0.0295265764,
-0.0621490963,
-0.0513592698,
-0.0412709154,
0.048741769,
-0.0331678577,
0.0182430241,
-0.0233947448,
0.0688240007,
-0.0209423192,
0.0182336122,
-0.0160533078,
-0.0590569228,
0.0058953026,
-0.0455588661,
-0.0252420455,
0.0200972278,
0.0265457518,
0.0022666748,
0.0250474755,
0.000429464,
-0.0027975496,
-0.0577908717,
0.0601248257,
-0.0256190263,
-0.023754254,
-0.0279824287,
0.0554976426,
0.0139824441,
-0.0608491935,
-0.0144985849,
-0.0083023813,
0.0128860287,
0.0507292449,
0.0240163039,
0.063063778,
0.0158823431,
0.0030819483,
-0.0839747861,
0.0426998027,
-0.0037001655,
0.0215264671,
0.0098852906,
0.0228588451,
0.0058305603,
-0.0090843923,
0.0174046773,
-0.0432995632,
-0.0170896892,
0.0092699472,
0.0287178643,
0.0171712339,
-0.0094131958,
-0.0043325261,
-0.0108708888,
-0.0517595336,
-0.0080474354,
-0.0053507965,
-0.0315040573,
-0.019983951,
0.0438576899,
-0.0070669772,
-0.0497976691,
-0.0434276238,
0.0241578128,
0.0465194695,
0.0280587673,
-0.0121365599,
0.0670579448,
-0.0239709113,
-0.0503491759,
-0.0112905465,
-0.0242035259,
-0.0031100179,
-0.0494136363,
0.0176895261,
-0.01745262,
0.0188324433,
-0.0243782829,
0.0432511419,
0.0259925164,
-0.0238190982,
-0.0186308902,
-0.0016713304,
0.0211848263,
0.045915857,
-0.0073432936,
-0.0277546402,
-0.0578955784,
0.003495144,
-0.0049561784,
-0.0207610894,
-0.0136627126,
-0.0074489145,
2.464e-7,
0.0283916648,
0.011655475,
0.0205938388,
-0.0166411437,
0.0154001135,
0.0401843749,
0.0014707559,
0.0580057427,
-0.0216197893,
-0.0175728891,
0.0173429754,
0.0616498105,
0.0449177176,
0.006470535,
-0.0751383454,
0.0614842176,
-0.0808859095,
-0.0356509574,
-0.0300296955,
0.0183823612,
0.0895516798,
0.0753130317,
0.0292771421,
-0.0089427354,
0.0236936044,
-0.0215772241,
0.0500867367,
-0.0107802805,
0.0139502501,
-0.0003255387,
0.0019989677,
-0.0101850312,
-0.0141656809,
-0.0009654852,
0.0033353146,
0.0288509391,
-0.0036648614,
0.0399188176,
-0.0188257657,
-0.0414544158,
0.014037516,
0.0407821126,
0.0416397788,
-0.0392694473,
0.0507583767,
-0.054367099,
-0.025069965,
-0.0195762031,
-0.0318468362,
0.0167595837,
0.040551208,
0.0241682157,
-0.0583044216,
-0.016903175,
0.0535856932,
0.0427351072,
-0.0947893336,
-0.0697004423,
-0.0255389512,
0.02129733,
-0.0321726054,
0.0182934795,
-0.0338037312,
-0.0014082224,
0.0243178327,
-0.0319912955,
0.0263868738,
2.018650505e-34,
0.0093596531,
0.0034617297,
-0.0094200457,
0.0100157559,
0.0078823101,
-0.0161421355,
-0.0260273237,
-0.01493295,
0.0153035382,
-0.051348798,
0.015755292
] |
http://arxiv.org/abs/2401.13601v5 | http://arxiv.org/pdf/2401.13601v5 | MM-LLMs: Recent Advances in MultiModal Large Language Models | 2024-01-24T17:10:45 | 2024-05-28T05:36:23 | Duzhen Zhang, Yahan Yu, Jiahua Dong, Chenxing Li, Dan Su, Chenhui Chu, Dong Yu | cs.CL | cs.CL | In the past year, MultiModal Large Language Models (MM-LLMs) have undergone
substantial advancements, augmenting off-the-shelf LLMs to support MM inputs or
outputs via cost-effective training strategies. The resulting models not only
preserve the inherent reasoning and decision-making capabilities of LLMs but
also empower a diverse range of MM tasks. In this paper, we provide a
comprehensive survey aimed at facilitating further research of MM-LLMs.
Initially, we outline general design formulations for model architecture and
training pipeline. Subsequently, we introduce a taxonomy encompassing 126
MM-LLMs, each characterized by its specific formulations. Furthermore, we
review the performance of selected MM-LLMs on mainstream benchmarks and
summarize key training recipes to enhance the potency of MM-LLMs. Finally, we
explore promising directions for MM-LLMs while concurrently maintaining a
real-time tracking website for the latest developments in the field. We hope
that this survey contributes to the ongoing advancement of the MM-LLMs domain. | MM-LLMs: Recent Advances in MultiModal Large Language Models: In the past year, MultiModal Large Language Models (MM-LLMs) have undergone
substantial advancements, augmenting off-the-shelf LLMs to support MM inputs or
outputs via cost-effective training strategies. The resulting models not only
preserve the inherent reasoning and decision-making capabilities of LLMs but
also empower a diverse range of MM tasks. In this paper, we provide a
comprehensive survey aimed at facilitating further research of MM-LLMs.
Initially, we outline general design formulations for model architecture and
training pipeline. Subsequently, we introduce a taxonomy encompassing 126
MM-LLMs, each characterized by its specific formulations. Furthermore, we
review the performance of selected MM-LLMs on mainstream benchmarks and
summarize key training recipes to enhance the potency of MM-LLMs. Finally, we
explore promising directions for MM-LLMs while concurrently maintaining a
real-time tracking website for the latest developments in the field. We hope
that this survey contributes to the ongoing advancement of the MM-LLMs domain. | [
0.0278459713,
0.0687081963,
-0.005919301,
0.0482222661,
-0.0385746397,
0.0045547877,
0.0245177373,
-0.0137180453,
-0.0189982485,
-0.098835513,
-0.0667172819,
-0.0530503131,
-0.0233018901,
0.0223531295,
0.063852042,
-0.054253947,
0.0338299163,
-0.0886108503,
-0.0535958633,
-0.0003286035,
0.0264246408,
0.0123442989,
-0.0401068106,
0.0421240889,
-0.0361650698,
-0.0139067499,
-0.0270337313,
0.026121147,
-0.0192600545,
-0.0496265776,
-0.0152398404,
0.0257926583,
0.0839173347,
0.0287125688,
0.0000018558,
-0.0317328088,
0.0003705796,
-0.0273827575,
0.0102625089,
-0.024704719,
0.0652606562,
-0.0373327918,
0.0176721271,
0.0002780975,
-0.007433217,
-0.025757201,
0.061400868,
0.048357442,
0.0651618689,
0.0700834021,
-0.0069343587,
-0.0344405323,
0.0257482603,
-0.0160340536,
-0.0146491276,
-0.0231204573,
0.0254722442,
0.0153570333,
-0.0018198795,
0.0109244296,
-0.0426292233,
0.0376781225,
0.0070934426,
-0.0005403171,
0.0035719473,
0.0239414349,
0.0410099812,
-0.0312583074,
-0.0378116779,
0.0735863969,
0.0094431732,
0.0248323921,
0.0000153829,
-0.0066724131,
-0.027631294,
-0.0016322643,
-0.0706804097,
0.0445696339,
0.0221071225,
0.0152858272,
-0.0466087312,
-0.0090185795,
0.0335843936,
0.0387882628,
-0.0124581549,
0.0058883848,
0.029425364,
-0.0207367651,
-0.0043416414,
-0.0011626751,
-0.0131337661,
-0.0457662456,
-0.0176592302,
0.0514124893,
-0.0169199873,
0.0077198097,
-0.0226980075,
-0.0234592166,
-0.0005363927,
-0.0354589932,
0.0633826554,
0.0119151734,
-0.033384867,
0.0507079102,
-0.0084097339,
0.0395787917,
-0.0502491072,
0.0199706685,
-0.0178227667,
-0.0114133339,
-0.0519927479,
-0.0000119277,
0.0038405126,
-0.0039708065,
-0.0053795693,
-0.0558864735,
-0.0304723289,
0.0322853401,
0.0113100624,
0.0160547569,
0.0079908669,
0.0126746893,
-0.0314184576,
-0.0315240212,
-0.0233354867,
0.0771411955,
-0.003417911,
0.0325287953,
0.0148114907,
-0.0252927653,
0.0242571682,
0.0722544938,
-0.0084642172,
-0.0408680774,
0.0036194271,
0.126261875,
-0.0335400738,
-0.051007513,
-0.036328733,
-0.0084974458,
-0.031468749,
-0.0741805509,
0.0538364016,
-0.0566240028,
0.052028574,
-0.019269105,
0.0302591883,
0.0036873452,
0.0037665823,
-0.0169993266,
0.000370505,
0.0854257941,
0.028796846,
0.0101745399,
0.0381466001,
-0.0016621533,
-0.0061439769,
0.1077998728,
0.0422155149,
0.0359342806,
0.0435284302,
-0.0264410581,
-0.0110278912,
0.0727493763,
-0.033156354,
0.0169070717,
-0.0497994684,
-0.019049922,
0.0283998959,
0.0759329274,
0.0489365272,
0.0469833575,
-0.0305589065,
-0.0061109141,
0.117092289,
0.0345049426,
0.0233039409,
-0.0142750647,
0.0480210967,
-0.0197996888,
-0.0014534874,
-0.0022750949,
-0.0539126135,
0.0498521328,
-0.0197269171,
0.0009653299,
0.0185172129,
0.050607916,
-0.0390461646,
-0.0048876866,
-0.0144032938,
-0.0325949565,
-0.0028619869,
0.0173997451,
0.072018899,
-0.032623224,
-0.0441637635,
0.0256004427,
0.0118340151,
-0.0298954397,
0.0247329231,
-0.049429927,
0.0046847258,
0.0601593629,
0.0377898477,
-0.0072323312,
-0.01637429,
-0.0235833582,
0.0275169108,
0.0757159814,
0.0143640079,
-0.0168399885,
0.016591087,
-0.0800650716,
-0.036696326,
-0.0066038598,
-0.0012453778,
0.0015740595,
0.0193225443,
-0.0151240258,
-0.0085406872,
-0.0127805872,
0.0273002908,
0.0354403444,
-0.0011542944,
-0.009712806,
-0.0300686043,
0.0016456397,
0.0319374353,
-0.010033722,
-0.0078036967,
-0.0066831377,
0.0337450057,
-0.0391988233,
0.0033860318,
0.0074531739,
-0.0411340296,
-0.0286790263,
0.0477417931,
0.0008911069,
-0.0094634648,
0.0279093646,
-0.0505416133,
-0.0236714929,
-0.0021381448,
-0.0337199904,
0.0171554741,
-0.070063971,
-0.003026682,
0.0317135043,
-0.0086646089,
-0.0462626256,
0.0356138386,
-0.0271742456,
-0.0110734971,
0.0050845989,
-0.0012640457,
-0.016270414,
-0.0423185192,
-0.0354617052,
0.0385900475,
0.0199575853,
0.0310251415,
-0.0099702021,
-0.0141754104,
-0.0238990467,
-0.0519541651,
-0.0874154419,
-0.0029508544,
-0.006386125,
-0.0028936127,
0.0016534352,
-0.0161934402,
0.0015201718,
-0.00604629,
-0.0106232967,
0.0268272199,
0.0291868951,
0.0318635367,
0.0080269398,
-0.0009160212,
-0.0002390676,
-0.0191030502,
-0.0287906118,
0.0079200119,
0.0506261401,
-0.0410760753,
-0.0135711608,
-0.0476366542,
-0.0350147858,
-0.0380286463,
0.0448265262,
-0.0340336226,
-0.0353487171,
0.0242993589,
-0.0494022109,
0.0474454015,
0.0328326561,
0.026210729,
0.0275154244,
0.0557412766,
-0.0207152944,
-0.030762108,
0.0110634742,
-0.0101396795,
-0.0205063447,
-0.0408729538,
0.0023580501,
-0.030868521,
0.063948676,
-0.0139597682,
0.0297163595,
-0.0261717848,
-0.009405151,
-0.0023582377,
-0.0560885556,
-0.041465465,
0.0125407213,
-0.0252213851,
0.0009417196,
-0.0223513972,
-0.0478775613,
0.0411785021,
-0.0560163595,
-0.012530514,
0.0196170285,
-0.0040748487,
-0.0350267105,
0.0761367753,
0.017143447,
0.018622255,
-0.0214123446,
-0.008108804,
-0.055285722,
-0.0162048973,
0.023750484,
-0.0087390356,
0.0629668236,
-0.0509219356,
0.0129387323,
-0.0202710163,
-0.0201926529,
-0.0051381872,
-0.0183563456,
-0.0460507497,
-0.0055348058,
-0.0364979468,
0.003901378,
-0.019554209,
0.0306194164,
0.0258107837,
0.0033846051,
-0.0172983408,
0.0028338009,
-0.0446581654,
0.0469234474,
0.0619142465,
0.0625182092,
-0.0234497767,
-0.0802435502,
-0.0172896851,
0.0442898907,
0.0236931946,
-0.0122232344,
-0.0530112088,
0.0293757264,
0.0502766296,
0.0169642605,
-0.0081795491,
0.0143615333,
0.0428678356,
-0.0698804781,
0.0336342193,
0.0783379674,
-0.0056222021,
-0.0575079806,
-0.0176036265,
0.0213471074,
0.0326128788,
0.0084248325,
0.005476234,
-0.0596642308,
-0.0576410778,
0.0013916986,
-0.0005370979,
0.0205276515,
-0.006373378,
-0.0029301404,
-0.0907784402,
0.0280084424,
0.0080623906,
-0.0141648268,
-0.0102783497,
-0.0160413254,
-0.0034651714,
0.002224823,
0.0134762246,
0.0145644555,
-0.0279780291,
-0.0354286246,
0.0227636266,
-0.0132127851,
-0.0653802454,
0.0090577612,
-0.0515114963,
0.0394230001,
0.0481880195,
0.0035363217,
-0.0134740304,
-0.0506524816,
0.0463287011,
0.0111009246,
0.0140732648,
-0.0143989632,
0.023225842,
-0.0546524115,
-0.0188301951,
-0.0253680386,
0.007373007,
0.0132776415,
-0.0172262639,
-0.0145253548,
0.1030574217,
-0.0335807353,
0.009781043,
-0.0173700377,
-0.001232582,
-0.003162266,
0.0687075257,
0.0074929963,
-0.0501160137,
-0.0393198095,
-0.0021853142,
-0.0159500446,
-0.0296895355,
-0.0102688977,
-0.0033217014,
-0.0228250027,
0.0114355357,
-0.0565788448,
-0.0027707119,
0.025900593,
0.0376197137,
0.00110467,
-0.0107800178,
-0.0029393653,
0.0024927072,
-0.0218646955,
0.0355277508,
-0.0414677933,
0.0219691843,
-0.0668112859,
0.0112519609,
0.0374035425,
-0.0244842377,
0.0269250348,
0.0373327173,
-0.1067417637,
-0.0585950874,
-0.0557028577,
0.0595725738,
0.0146321021,
0.0384518951,
0.0307405833,
-0.0905348361,
-0.0014481737,
0.0009455599,
-0.0057188505,
0.0189838745,
0.0295398422,
0.0270419288,
-0.0182981659,
0.0292479321,
-0.0378191546,
-0.004334155,
0.0060450956,
0.0523620062,
-0.0131720621,
-0.0089670047,
-0.0133530097,
-0.0250808764,
0.0462507159,
-0.0371193774,
-0.0122351749,
-0.0572527722,
0.0121871429,
-0.026217049,
-0.0187201537,
-0.0609702021,
0.0133665819,
0.0381784402,
0.0102998838,
0.0100736069,
-0.0062198942,
0.0382227823,
0.0418355092,
-0.0313992091,
0.05030616,
0.0034758081,
-0.0368698537,
0.0082791578,
0.0384045355,
-0.0146585684,
0.0266597439,
-0.029560009,
0.0016922785,
-0.0161461346,
0.0939957723,
0.0133473156,
-0.0318845063,
-0.0344093554,
0.0146599375,
-0.0034204624,
-0.0328366458,
-0.0095767882,
0.0595759228,
0.0778202042,
0.0174063966,
0.0411791317,
-0.0136094512,
-0.0271293614,
-0.0246361308,
0.0724926814,
-0.0125890672,
-0.0174252763,
0.0627329573,
0.037939772,
-0.0012802088,
-0.0032171493,
0.0640594065,
0.0143502587,
0.033321172,
-0.0042803474,
0.0184722468,
-0.0346780196,
-0.0281163808,
0.0463342257,
0.0274050031,
-0.0158713721,
0.015168041,
-0.0203253888,
-0.0637389198,
-0.0136963055,
0.0205103215,
-0.0258809924,
-0.0188701097,
0.0501300991,
0.0662953109,
-0.0001268179,
-0.0198880695,
-0.0432270281,
-0.0013803407,
0.0191232264,
0.0063192868,
0.0052733207,
-6.173415681e-33,
-0.0154706687,
-0.006270099,
0.0660981238,
-0.0421213172,
-0.0494551882,
-0.0017984555,
-0.011095765,
0.0201198552,
0.0044743894,
-0.0234809332,
0.036607068,
-0.0278489217,
0.0006601588,
0.0300111454,
0.005147648,
0.0165709238,
0.0360916257,
-0.0365362316,
0.0010190661,
-0.0060558212,
0.0315344594,
0.0171481129,
0.0487123877,
-0.0121108666,
-0.0098376395,
0.010334204,
-0.0329619125,
0.0524330661,
-0.0160635691,
0.0833627284,
-0.0671585873,
0.0244659558,
0.0325899161,
-0.014870652,
-0.0346052982,
0.0404770225,
-0.0443318523,
0.0206627715,
0.0134253716,
0.0154250804,
-0.0567924827,
-0.0098107541,
0.0414989553,
-0.0166391898,
-0.0149582978,
-0.0110569559,
0.0251359306,
-0.0148655018,
-0.0407559499,
0.0190098509,
-0.029573217,
0.0237755794,
0.0068996539,
-0.0139377899,
0.0353232175,
0.0417242572,
0.0141519764,
0.0197015181,
-0.0431396216,
0.0334813222,
-0.0121144634,
0.0278793965,
0.076031819,
0.0682756826,
0.0180486925,
0.0747281164,
0.0156366751,
-0.022466328,
-0.0220595486,
0.0100511927,
0.0472683236,
0.0593076274,
0.0485489145,
0.0669311136,
-0.0260662902,
0.0165685732,
-0.0276021361,
-0.0035215125,
0.0148713924,
0.0688369721,
0.0358227827,
0.0450237617,
0.0125631308,
-0.011778431,
-0.0344866887,
0.0072433976,
-0.0259471573,
-0.0142681776,
0.0458215028,
0.0418787859,
-0.0011191774,
0.0103438292,
0.1038996056,
-0.0280018933,
-0.0387655608,
-0.012009671,
-0.0037959516,
0.0698770732,
0.0082756002,
0.0205266569,
-0.0563456453,
-0.0402637273,
0.0229992382,
-0.0443144478,
0.0089942655,
-0.0078688348,
0.0001384277,
0.0446350686,
-0.039387621,
-0.0001354703,
-0.0445750095,
-0.0061590024,
0.0130469175,
0.0205837768,
0.0422011167,
-0.0212185662,
-0.0403197296,
0.0528429821,
-0.0176162459,
-0.0276643839,
-0.0099403225,
-0.0605816282,
-0.0194344576,
0.0389764085,
-0.0082948795,
0.0369436778,
-0.0531900823,
-0.050294321,
0.0292437766,
-0.0529076681,
-0.0004654747,
0.0294513684,
2.484e-7,
0.0418084934,
0.03862001,
0.00196295,
0.0001662688,
0.0329987034,
0.0406011455,
-0.0525326654,
0.0283360407,
0.0120041622,
-0.0444825403,
-0.0199230686,
0.0549814664,
0.0507495403,
0.0159321595,
-0.0403225757,
-0.0061355145,
-0.032683745,
0.0255912673,
-0.0619123504,
0.005401656,
0.0489350446,
0.0700263605,
-0.0192704555,
-0.0314151123,
-0.0019825506,
-0.0423582569,
0.0994031057,
-0.0290346202,
-0.0464669243,
-0.0112185022,
0.0276464373,
-0.0016841323,
-0.0129232211,
0.0010767921,
0.020755101,
-0.0233021304,
0.0265523996,
0.0778959692,
0.0085023418,
-0.0550322942,
0.0171921067,
0.0232511256,
-0.0482820868,
-0.0491372421,
0.0731685013,
-0.0339282453,
-0.0272546541,
-0.0485396385,
-0.0052406429,
-0.0219107922,
-0.0006150696,
-0.0295001566,
-0.0584304668,
0.0169581957,
0.0232151058,
-0.0041612522,
-0.0279510058,
-0.0971431732,
-0.0058561722,
-0.0014543604,
-0.0801558048,
-0.0463660359,
-0.0497610569,
-0.0131709501,
0.0253702197,
-0.0670996085,
-0.0604808964,
2.200795192e-34,
0.031788297,
-0.0107240733,
0.0131660774,
0.0182152893,
0.0059329923,
-0.0178494044,
-0.0339980461,
0.0077787004,
-0.0078395484,
-0.1160965562,
-0.0115617467
] |
http://arxiv.org/abs/2401.13598v1 | http://arxiv.org/pdf/2401.13598v1 | Consistency Guided Knowledge Retrieval and Denoising in LLMs for Zero-shot Document-level Relation Triplet Extraction | 2024-01-24T17:04:28 | 2024-01-24T17:04:28 | Qi Sun, Kun Huang, Xiaocui Yang, Rong Tong, Kun Zhang, Soujanya Poria | cs.CL | cs.CL | Document-level Relation Triplet Extraction (DocRTE) is a fundamental task in
information systems that aims to simultaneously extract entities with semantic
relations from a document. Existing methods heavily rely on a substantial
amount of fully labeled data. However, collecting and annotating data for newly
emerging relations is time-consuming and labor-intensive. Recent advanced Large
Language Models (LLMs), such as ChatGPT and LLaMA, exhibit impressive long-text
generation capabilities, inspiring us to explore an alternative approach for
obtaining auto-labeled documents with new relations. In this paper, we propose
a Zero-shot Document-level Relation Triplet Extraction (ZeroDocRTE) framework,
which generates labeled data by retrieval and denoising knowledge from LLMs,
called GenRDK. Specifically, we propose a chain-of-retrieval prompt to guide
ChatGPT to generate labeled long-text data step by step. To improve the quality
of synthetic data, we propose a denoising strategy based on the consistency of
cross-document knowledge. Leveraging our denoised synthetic data, we proceed to
fine-tune the LLaMA2-13B-Chat for extracting document-level relation triplets.
We perform experiments for both zero-shot document-level relation and triplet
extraction on two public datasets. The experimental results illustrate that our
GenRDK framework outperforms strong baselines. | Consistency Guided Knowledge Retrieval and Denoising in LLMs for Zero-shot Document-level Relation Triplet Extraction: Document-level Relation Triplet Extraction (DocRTE) is a fundamental task in
information systems that aims to simultaneously extract entities with semantic
relations from a document. Existing methods heavily rely on a substantial
amount of fully labeled data. However, collecting and annotating data for newly
emerging relations is time-consuming and labor-intensive. Recent advanced Large
Language Models (LLMs), such as ChatGPT and LLaMA, exhibit impressive long-text
generation capabilities, inspiring us to explore an alternative approach for
obtaining auto-labeled documents with new relations. In this paper, we propose
a Zero-shot Document-level Relation Triplet Extraction (ZeroDocRTE) framework,
which generates labeled data by retrieval and denoising knowledge from LLMs,
called GenRDK. Specifically, we propose a chain-of-retrieval prompt to guide
ChatGPT to generate labeled long-text data step by step. To improve the quality
of synthetic data, we propose a denoising strategy based on the consistency of
cross-document knowledge. Leveraging our denoised synthetic data, we proceed to
fine-tune the LLaMA2-13B-Chat for extracting document-level relation triplets.
We perform experiments for both zero-shot document-level relation and triplet
extraction on two public datasets. The experimental results illustrate that our
GenRDK framework outperforms strong baselines. | [
0.0546410829,
0.0430749618,
0.0266023539,
0.0754924342,
-0.0178124998,
-0.0017877523,
-0.0102197388,
0.0486136824,
0.02958248,
-0.0560425222,
-0.0227397606,
-0.0163954925,
-0.0473009422,
-0.0040007862,
-0.005471861,
-0.0308692493,
0.0423495471,
0.0070451768,
0.0577313378,
-0.0037396192,
0.0243874565,
0.0185665563,
0.0259145256,
-0.036334984,
-0.0545422584,
-0.0200153925,
0.0175457522,
-0.0050515896,
0.0024395974,
-0.0482624844,
0.0224310737,
0.0390723273,
0.0266943686,
0.017839605,
0.0000019485,
-0.0711341128,
-0.0592371561,
0.0283965264,
-0.0192195233,
-0.0117801847,
0.0064453371,
-0.0378018282,
0.0110974424,
-0.0073638922,
-0.0367877074,
-0.0476534441,
-0.0038454018,
0.0582781881,
0.0400706343,
0.0869877562,
-0.0303653255,
-0.0440796465,
0.0420589,
-0.0465352088,
0.0690665245,
-0.0221891683,
0.0351142064,
0.0314933248,
0.0194917191,
0.0405196659,
0.0287663806,
0.0192647949,
0.0020463737,
-0.0257259104,
0.0254230518,
0.0418638773,
-0.0493486524,
-0.0332984477,
0.0222740434,
0.1031529456,
0.0283971932,
0.0165450796,
0.0437524207,
0.0313381068,
-0.0203173831,
-0.0257082544,
-0.0483684577,
0.0010084838,
-0.0348553322,
0.0258661136,
0.0085383588,
-0.0070320815,
0.0385892875,
0.0101784477,
-0.0145434849,
0.0403532609,
0.0130317016,
-0.0557017028,
0.0072250739,
0.0021421802,
0.003586031,
-0.0467353873,
0.0075718397,
0.04624613,
0.0103354435,
0.0353472792,
-0.0383624174,
-0.0477530994,
0.0178910028,
-0.0193712395,
-0.0595770031,
0.0110751018,
-0.030574657,
0.0842526779,
-0.0573538989,
0.011278566,
0.0011179205,
-0.0157782342,
-0.0170172229,
-0.0203585289,
-0.0275944676,
-0.0126789436,
-0.044744052,
0.0845397711,
-0.0453563184,
-0.0107744141,
-0.025589563,
-0.0090505974,
0.0098439706,
0.0439807139,
0.0296602007,
-0.0116206761,
-0.0443788059,
-0.0301481504,
-0.0067128646,
0.0189488623,
-0.0209060311,
-0.0113716703,
-0.0066258884,
-0.0337795354,
-0.0050247773,
0.0078673707,
0.0409121476,
-0.022966085,
0.013536144,
0.0820879415,
-0.0035938667,
-0.0365602225,
-0.019092219,
-0.0595576949,
0.0222884621,
-0.0899653584,
0.0266497806,
-0.0471752398,
0.0650621802,
0.020559188,
-0.0157726798,
-0.042090632,
0.0574509874,
0.0049005207,
-0.0606324039,
0.0630510971,
-0.0530920662,
0.0032608509,
-0.0050690081,
0.0345242806,
0.0399768353,
0.0684105083,
0.0315807648,
0.0033187883,
0.0817323625,
0.0299156476,
-0.0135794673,
0.0339721404,
0.0505759045,
-0.0065778857,
-0.0000593808,
0.0176462475,
0.0151716443,
-0.0006692628,
0.0473156981,
0.0611740686,
-0.0299039241,
-0.0305823907,
0.0388853438,
-0.0090448102,
0.0351831093,
0.0310161076,
0.047917787,
0.016437456,
0.0012153917,
0.0233275797,
-0.0205208492,
0.0004300734,
-0.0005957603,
0.0282059237,
0.0067432425,
0.0258796494,
-0.0207270253,
0.0003432926,
-0.0103713609,
-0.0108722895,
-0.0508562848,
-0.0209646914,
0.0903972238,
0.0270349886,
-0.0417615362,
0.0660180449,
0.0651301071,
-0.0406223424,
0.0059227231,
-0.0015521041,
0.0498015247,
0.0351402573,
0.0595468543,
-0.0656467006,
-0.0281544216,
0.0110683301,
-0.016784912,
0.0400212444,
0.0095202094,
-0.0236484874,
0.0043727001,
-0.0456342809,
-0.0343395509,
0.0173600297,
0.0021935985,
0.0084716491,
0.0620653108,
0.0329862013,
0.0497319624,
-0.0469365753,
0.034554068,
0.0326130763,
-0.043449115,
-0.0563693829,
0.0179388709,
-0.0130372653,
0.0007604067,
0.0361752585,
-0.0225671381,
-0.0001914616,
0.0547935329,
-0.0399160348,
0.0011818482,
-0.0268778875,
-0.0320565663,
-0.0547534302,
0.0364911295,
0.0136991078,
-0.0287404917,
-0.0303361323,
-0.034341868,
-0.0366152227,
-0.0279964581,
0.0299277678,
0.0562007464,
-0.0443331152,
-0.0282812845,
0.0634742305,
0.009883157,
-0.0357303917,
0.0883015543,
0.0061903149,
-0.0189263821,
0.0300711989,
-0.0495373309,
-0.0077771991,
0.0042915186,
-0.0215728395,
0.0323337428,
0.0095043546,
0.0073475768,
0.0181264095,
0.0292934794,
-0.0414614603,
-0.0381430909,
-0.0584757701,
0.0454130098,
0.0172371343,
0.0404588543,
0.0015043634,
0.0286819804,
-0.0069061834,
-0.0145464297,
-0.0498381481,
0.0454859436,
0.0101960255,
-0.002375718,
0.0383611284,
0.0287306085,
-0.0279835425,
-0.0236263014,
-0.0017395649,
-0.0013953532,
0.025845794,
0.0002594939,
-0.0288467165,
0.0022721272,
0.0701629966,
-0.0595434308,
0.0369032323,
-0.0152694881,
0.0187161434,
-0.0142684467,
-0.0403190255,
0.0287789684,
-0.0082857097,
-0.0172172971,
0.0488977768,
-0.0150374286,
-0.0257154275,
-0.0425485484,
0.0248894561,
-0.044752188,
0.0159477592,
0.0253229421,
0.0158856548,
-0.0166661348,
0.0398232192,
-0.0188411884,
-0.0093495799,
-0.0152319977,
-0.0000528846,
0.0125544639,
0.0063251602,
-0.0334566124,
0.0437745117,
0.0219289493,
0.0101232557,
-0.0591390133,
-0.0194117371,
0.0462618023,
-0.0502441935,
-0.0144196404,
0.0153355002,
0.02110474,
-0.0515072159,
0.0684472844,
-0.006692783,
0.0057173315,
-0.0108486218,
0.0048739524,
-0.0255715866,
-0.0404501781,
0.034759447,
-0.034492556,
0.0001603507,
-0.0302477665,
0.0526688658,
-0.033014141,
-0.0395523347,
-0.0422151163,
-0.0227378029,
0.0039937817,
0.0218260195,
-0.0122971488,
-0.0006494405,
0.003580593,
0.009328519,
0.0266414247,
0.048876334,
-0.0221232101,
-0.0266023744,
-0.0149481082,
0.0953691155,
0.0540806726,
-0.0454322137,
-0.0186161734,
0.0036415479,
0.0320147164,
0.0659361482,
-0.0035727902,
0.0102022309,
0.0033495342,
-0.0017600983,
-0.0072771437,
-0.0138908746,
-0.0007794784,
-0.0312203635,
-0.0033046501,
-0.0739070103,
-0.0167335253,
0.0937642008,
0.0140424594,
-0.0587598756,
-0.0113037713,
0.0208953675,
0.0273240954,
0.0201850403,
0.0146937575,
-0.1264240146,
-0.0193671864,
-0.0085800393,
0.0000227432,
0.014874151,
-0.0591103993,
-0.0163481627,
-0.1153602079,
0.0318083502,
-0.0080583924,
-0.0422393568,
0.020997297,
-0.0565738007,
0.0079108896,
0.0233488716,
-0.0195049383,
0.0212301761,
-0.024697708,
0.0019512506,
-0.0225143135,
0.0349308625,
-0.0642999783,
-0.0032750352,
-0.0065292963,
0.0359674394,
0.013899155,
0.0342447087,
0.0079918168,
0.0078320559,
0.03455659,
-0.0015460147,
-0.0476293787,
-0.048065424,
0.0472275615,
-0.040813975,
-0.0117406277,
-0.0355232805,
0.0059319539,
-0.0415011793,
-0.0080679562,
0.0003039202,
0.0592599884,
0.0126277059,
-0.00793867,
-0.0184283573,
-0.0128136054,
0.0063734204,
0.0745121762,
0.0029736825,
0.0033384245,
-0.0080263419,
0.0205440354,
0.0174661353,
-0.0340451747,
0.0018745863,
0.007556404,
-0.0873251632,
-0.0208571274,
-0.0623062626,
-0.0568080284,
0.036015071,
0.0238930993,
-0.0111148683,
0.0107082725,
0.0068563032,
0.0040867762,
-0.0397546291,
-0.0111172479,
-0.0455290973,
-0.0300077256,
-0.0122945923,
0.0137386443,
0.0384410433,
-0.0334215462,
-0.0118079549,
-0.016954653,
-0.0410303995,
-0.044880487,
-0.0285222344,
0.0278014559,
0.0259670801,
0.0584791787,
0.0011580553,
0.0009167573,
0.042686224,
0.1077364236,
-0.0615976527,
-0.00006691,
-0.0041866647,
0.0059431205,
0.0035233896,
0.048817113,
-0.0526442677,
0.0464400612,
-0.0022609797,
0.0447770879,
-0.0383603275,
-0.036453627,
0.0154445535,
0.0557845384,
0.037584316,
0.0066491631,
0.0526569821,
-0.0521212742,
0.0191269983,
-0.0687804073,
-0.0239698924,
-0.0678384677,
0.0298901219,
-0.0013481895,
-0.0213085525,
-0.0054163043,
0.0121468399,
-0.0072196401,
0.0521151423,
-0.0186408646,
-0.0384652689,
-0.0126548363,
-0.0043444433,
0.0294185989,
-0.0483849905,
-0.008031412,
0.0452954136,
-0.0027217597,
0.0126732411,
-0.0548996665,
0.0211175289,
0.024347486,
-0.0080299703,
0.0011416129,
-0.0104952035,
0.0069051851,
0.0147120655,
-0.044228632,
0.0341911763,
0.0594233871,
-0.0017511608,
0.0750745907,
-0.0360782258,
-0.0265055299,
-0.0103822853,
0.0348562859,
0.0087188529,
0.0085435957,
0.0056907046,
0.0209873393,
-0.0339758433,
0.0397511274,
0.0983778387,
0.0088436911,
0.0519524179,
0.0185154155,
-0.0235615391,
-0.0590351149,
-0.0138663258,
0.0220765918,
-0.0066834409,
0.0039024458,
-0.0033224162,
-0.0647746623,
-0.0514271669,
-0.0123826787,
-0.014522925,
-0.062935777,
0.0088505326,
0.0186224356,
0.0531918295,
0.0235078298,
-0.0326927379,
-0.0408299044,
0.0140215643,
-0.04703594,
-0.0409515165,
0.0244956519,
-6.904281498e-33,
-0.0478921086,
-0.0460764803,
0.0193728916,
0.0185470209,
-0.0616951138,
0.0477805249,
-0.0477028675,
-0.0198896211,
0.0106694559,
0.0471103564,
-0.0331119299,
-0.0050729765,
0.0054761185,
0.016290037,
-0.045192793,
-0.000002095,
0.0036026575,
-0.0268723555,
0.0119120227,
0.051614821,
0.0491405763,
0.0401992649,
0.0499443412,
-0.0962844416,
0.0195762403,
-0.0272699147,
-0.0133514246,
0.0249717087,
0.0327897221,
0.0370690823,
-0.03539031,
0.0418462045,
0.0375862122,
-0.0334708765,
0.0133512765,
-0.004086222,
-0.1420062929,
-0.0135753322,
-0.0095223673,
-0.045174025,
-0.0332609043,
-0.0257737804,
0.0299537983,
-0.0038700658,
0.0508229025,
-0.0339732319,
0.023607567,
-0.0307078697,
-0.0406892225,
-0.0961365849,
-0.0527687892,
0.0502756499,
0.0363412201,
-0.0122183291,
0.033943858,
-0.0038211332,
0.0232451055,
0.0241672192,
-0.0350856371,
-0.021038102,
-0.016596349,
0.0324098095,
0.0136783803,
-0.0014062244,
0.0069622784,
0.0484602116,
-0.0144958133,
0.0058908011,
-0.0249601342,
-0.0243731607,
0.011643271,
0.0657345876,
-0.035274215,
0.0647297576,
0.0194873456,
-0.0139126917,
-0.1064908952,
0.0337692797,
0.0089830393,
0.0377057157,
-0.020684965,
-0.0107901758,
-0.0150200827,
-0.0048110001,
0.007538612,
-0.0215049982,
-0.0359966233,
-0.0215921849,
0.0145443529,
0.0334266201,
-0.0371434428,
0.0010372942,
0.0126741771,
-0.0130003048,
-0.0380403288,
0.032113418,
-0.0220329203,
0.0028740887,
-0.0118012056,
-0.0025084917,
-0.0612331443,
-0.0676650926,
0.0150912944,
0.0088497186,
-0.0132970819,
0.0283281747,
0.0409771726,
-0.0273135435,
-0.0592432134,
-0.0089076879,
-0.047079619,
0.0329566859,
-0.0134432288,
-0.0041411878,
0.0030211406,
0.0097963968,
0.0166232605,
0.0466180593,
0.0514038168,
0.0317097865,
-0.0089617204,
0.0227356553,
0.0148186451,
0.0169781372,
0.0147886593,
0.0100964922,
-0.0193685908,
-0.0213990584,
0.0374688245,
-0.041527614,
0.0070738387,
0.0136536509,
2.772e-7,
0.0069397041,
0.0275731552,
0.0574764423,
0.0255418476,
0.0228799414,
0.0326051526,
0.0111028738,
0.0429087244,
-0.013323321,
-0.061625082,
0.0149546163,
0.0090813199,
0.0245140716,
0.0365746357,
-0.0854034349,
-0.011664819,
-0.0667803064,
-0.0697671622,
-0.0721368045,
-0.0258737374,
0.0662194714,
0.0555134043,
0.0210699234,
-0.0052724583,
-0.0324493386,
-0.0426640213,
0.0739669129,
-0.0109463558,
-0.0253358651,
0.0442388803,
0.0792621896,
0.0044614696,
-0.0177066997,
0.0032385604,
-0.0043517961,
0.0277960971,
0.0121962186,
0.0923155025,
0.0007520284,
0.0187004376,
0.049254179,
-0.0203918405,
-0.0029094091,
-0.0267543811,
0.0831406042,
-0.0280193817,
0.0046184049,
-0.0068781483,
-0.0182521716,
-0.0201892126,
-0.0159749743,
-0.0109751252,
-0.0153249828,
0.0450465158,
0.0176455136,
0.0093503082,
-0.0407656468,
-0.044126153,
0.0041958019,
0.0164353605,
-0.0229462888,
-0.0206827316,
-0.0072910539,
-0.0228656996,
-0.0251229946,
-0.0399624258,
0.0230219979,
2.832292109e-34,
0.0161554497,
-0.0042535765,
-0.0310600773,
0.0149852242,
0.0324830934,
-0.0424221605,
-0.0334520712,
0.0259580798,
-0.0252387431,
-0.0911366045,
0.022181401
] |
http://arxiv.org/abs/2401.13594v1 | http://arxiv.org/pdf/2401.13594v1 | Graph Guided Question Answer Generation for Procedural Question-Answering | 2024-01-24T17:01:42 | 2024-01-24T17:01:42 | Hai X. Pham, Isma Hadji, Xinnuo Xu, Ziedune Degutyte, Jay Rainey, Evangelos Kazakos, Afsaneh Fazly, Georgios Tzimiropoulos, Brais Martinez | cs.CL | cs.CL, cs.AI, I.2.7 | In this paper, we focus on task-specific question answering (QA). To this
end, we introduce a method for generating exhaustive and high-quality training
data, which allows us to train compact (e.g., run on a mobile device),
task-specific QA models that are competitive against GPT variants. The key
technological enabler is a novel mechanism for automatic question-answer
generation from procedural text which can ingest large amounts of textual
instructions and produce exhaustive in-domain QA training data. While current
QA data generation methods can produce well-formed and varied data, their
non-exhaustive nature is sub-optimal for training a QA model. In contrast, we
leverage the highly structured aspect of procedural text and represent each
step and the overall flow of the procedure as graphs. We then condition on
graph nodes to automatically generate QA pairs in an exhaustive and
controllable manner. Comprehensive evaluations of our method show that: 1)
small models trained with our data achieve excellent performance on the target
QA task, even exceeding that of GPT3 and ChatGPT despite being several orders
of magnitude smaller. 2) semantic coverage is the key indicator for downstream
QA performance. Crucially, while large language models excel at syntactic
diversity, this does not necessarily result in improvements on the end QA
model. In contrast, the higher semantic coverage provided by our method is
critical for QA performance. | Graph Guided Question Answer Generation for Procedural Question-Answering: In this paper, we focus on task-specific question answering (QA). To this
end, we introduce a method for generating exhaustive and high-quality training
data, which allows us to train compact (e.g., run on a mobile device),
task-specific QA models that are competitive against GPT variants. The key
technological enabler is a novel mechanism for automatic question-answer
generation from procedural text which can ingest large amounts of textual
instructions and produce exhaustive in-domain QA training data. While current
QA data generation methods can produce well-formed and varied data, their
non-exhaustive nature is sub-optimal for training a QA model. In contrast, we
leverage the highly structured aspect of procedural text and represent each
step and the overall flow of the procedure as graphs. We then condition on
graph nodes to automatically generate QA pairs in an exhaustive and
controllable manner. Comprehensive evaluations of our method show that: 1)
small models trained with our data achieve excellent performance on the target
QA task, even exceeding that of GPT3 and ChatGPT despite being several orders
of magnitude smaller. 2) semantic coverage is the key indicator for downstream
QA performance. Crucially, while large language models excel at syntactic
diversity, this does not necessarily result in improvements on the end QA
model. In contrast, the higher semantic coverage provided by our method is
critical for QA performance. | [
0.0618365817,
0.011931885,
-0.0125882048,
0.0111391749,
-0.0151357455,
-0.0076008374,
-0.026719097,
-0.0220020991,
-0.010076507,
-0.0073400317,
0.0000250209,
-0.031814646,
-0.0201502386,
0.035532061,
0.0076464331,
-0.0378577411,
0.0092577869,
-0.0096913502,
-0.022657536,
-0.0129606845,
0.003408656,
0.0294893216,
0.0357021727,
0.0137113323,
-0.0033698666,
-0.033054933,
-0.0363973267,
0.0182635244,
-0.0150874108,
-0.0530737415,
0.0233053807,
0.0373771228,
0.0191359222,
-0.0073233065,
0.0000020254,
0.0067091468,
-0.0111497445,
0.0021853538,
-0.0064039351,
-0.032169763,
0.0202486869,
0.0037047053,
0.0248829369,
0.0221084394,
-0.0572502427,
-0.0204621684,
0.0386473946,
0.0259385668,
0.0879472047,
0.0797014385,
-0.0070814341,
-0.0110397013,
0.0039003768,
0.0114855049,
0.0233472586,
-0.0393622555,
-0.0229973271,
0.0410728231,
0.0482095778,
0.0176333208,
-0.0142463688,
0.014317709,
0.0581560358,
0.02451149,
0.0253573637,
0.0409201421,
-0.0021312619,
0.0226319078,
-0.0093463669,
0.1173590571,
-0.010228877,
-0.0418799967,
0.006781124,
0.008421679,
-0.0052574538,
0.0084546525,
-0.0630890802,
0.0213433821,
-0.053944096,
-0.0065077785,
-0.0276404489,
-0.014393256,
0.0107262051,
-0.0313512087,
0.0040503405,
0.0274925511,
0.007040076,
-0.0261381194,
-0.0018778286,
0.0128956139,
-0.0506112985,
-0.0597592853,
-0.0169711038,
0.0685026273,
-0.0395115055,
0.0156095279,
-0.0597432666,
-0.011155664,
0.0102531202,
0.0159553885,
0.0690780357,
0.0255027302,
0.0223619957,
0.0027369775,
-0.0100549152,
0.0130228372,
-0.0163019393,
0.0284533445,
-0.0914510787,
-0.0016652433,
-0.0461803228,
-0.0323196352,
-0.0817486867,
0.0554333776,
-0.0477170944,
-0.032604374,
0.0039703702,
0.0610019118,
0.0201012641,
-0.009174549,
0.0349485762,
0.0601597093,
-0.0550586954,
-0.0113856457,
-0.0086223222,
-0.0148873916,
-0.0852208138,
0.0167385135,
0.0122518865,
-0.0309459046,
-0.0296988226,
-0.0020891321,
0.0544794947,
-0.0055805058,
-0.002019211,
0.0754421502,
-0.0101512391,
-0.0061187316,
-0.0112116467,
-0.0357007459,
-0.0141357929,
-0.0871817246,
0.0211713482,
-0.0771281272,
0.0183232073,
0.0097697768,
-0.0149079598,
0.01212468,
0.028259635,
0.0650411248,
-0.0153269172,
0.0474621207,
-0.0363579355,
0.0262998696,
0.0073200618,
-0.002426893,
-0.0079021538,
0.0531585254,
0.0496745817,
0.0166276749,
0.030190086,
0.0070560165,
-0.0098439353,
0.0134274103,
0.0246480163,
-0.0030858146,
0.05389015,
0.038610708,
0.0046367613,
0.0020057652,
0.0134896971,
0.0496334992,
-0.0358927101,
-0.0281536449,
0.0291217044,
0.0261041299,
-0.0314789824,
-0.0058581042,
0.0727872327,
0.0330245011,
0.0111362645,
0.0408274829,
-0.054884553,
0.0204662159,
-0.0179096851,
0.0393571667,
-0.0165905356,
-0.0134659754,
-0.0397499651,
-0.0229410175,
-0.0060978173,
0.0082925959,
0.0435432605,
0.0126993395,
0.0202936493,
0.0604434051,
-0.0355503112,
0.000268195,
0.0125736063,
-0.0323215649,
-0.0200136732,
-0.0091046514,
0.0618761592,
0.0646498278,
0.0579210781,
0.0067992406,
-0.0551139303,
0.0087210843,
0.0370011702,
0.0263528675,
-0.0042120339,
-0.0332072824,
0.0219817385,
-0.0447411314,
-0.0111180237,
-0.0744003057,
-0.0065474571,
0.0449551083,
0.0430605635,
-0.0019512033,
0.0122053958,
-0.1044266894,
-0.0349109098,
0.0142060407,
0.0157931037,
-0.0292641502,
0.027592903,
-0.0106376223,
-0.0075797602,
0.0128581785,
-0.0183075536,
0.0258274674,
0.0228131544,
-0.0124164475,
0.0101750987,
-0.0027657256,
-0.0412796028,
-0.0710700899,
-0.0041250214,
-0.0264125522,
-0.032575421,
0.0169102456,
-0.0550682023,
-0.0572609715,
0.0011126862,
-0.0745938122,
-0.0042644003,
-0.0293651391,
-0.0235136896,
0.0659780651,
-0.0094344793,
-0.0234931018,
0.0605550706,
0.029164765,
0.0224564541,
0.0312963761,
0.0339111984,
0.0293071158,
-0.0681499243,
0.0324519426,
0.0192937236,
-0.006132747,
0.0107861031,
0.0164898671,
0.0666426793,
-0.0566119216,
-0.0229782313,
-0.1140200049,
0.0261656996,
-0.0416612215,
0.0295309722,
0.059202347,
0.0336383022,
-0.0062404992,
-0.0446734242,
-0.0761383995,
0.0458469205,
0.0509438589,
0.0201210845,
0.0281311758,
-0.0036625264,
-0.0007630765,
-0.0326065011,
-0.0180802513,
-0.008374908,
0.0213910192,
0.0415640436,
0.0237147864,
-0.0464322045,
0.0031549323,
-0.0323422253,
0.0323784873,
0.0134850722,
-0.0202610735,
0.0385960378,
-0.0239598323,
0.0457931645,
0.0481549092,
-0.0060840189,
0.0082393941,
0.0283548441,
-0.0343309939,
-0.0376668535,
0.0390020832,
-0.0340861268,
-0.0243230946,
-0.0153458361,
-0.0475707948,
0.011098925,
0.0508281477,
0.0246396847,
0.0005594081,
-0.0603453107,
0.0280788988,
0.0159396417,
0.0144160101,
-0.0312433206,
0.018463891,
-0.0130954571,
0.0011023202,
-0.0394347161,
-0.0220125746,
0.0580856614,
-0.0026734127,
-0.0060768728,
-0.0649221465,
-0.0127996812,
-0.0547650792,
0.0963819921,
0.0498387404,
0.0215806607,
0.0437856503,
-0.0288658738,
-0.0731702745,
0.0216813795,
0.0286184885,
-0.0115337241,
0.0112246517,
-0.0545610785,
0.0228987336,
-0.0052573751,
-0.0514821857,
-0.0515248701,
-0.0235981327,
0.012947022,
0.0129860668,
0.025960682,
-0.0234785881,
0.0242177397,
0.0047930647,
-0.019134609,
0.0335905962,
-0.0410757735,
-0.0099370042,
-0.0385727026,
0.1016977355,
0.0840744972,
0.0271147508,
-0.0216875337,
-0.0393311009,
0.0112727294,
0.0591382086,
0.0502130874,
-0.0245487206,
-0.0025712287,
0.0166321266,
-0.007433956,
-0.0312173106,
-0.0148286223,
0.0447730385,
0.0997394696,
-0.0498788245,
-0.0378592014,
0.086245358,
0.0304349288,
0.0088454206,
-0.0058910511,
0.0086256312,
0.0693131015,
0.0067877606,
0.042349264,
-0.0999055654,
-0.0331256613,
-0.02247441,
0.0419539325,
0.0204837322,
-0.0110419681,
-0.0241450053,
-0.1008124426,
0.0320024043,
0.0002931088,
-0.0586119443,
-0.0256097391,
-0.0120251263,
-0.0262961704,
0.0303552765,
-0.0200316459,
-0.0007592894,
-0.0329198353,
0.0380632505,
0.0077644773,
-0.0371278003,
0.004565083,
0.0381860621,
-0.0124064945,
-0.0026652303,
0.0402384847,
0.0110094743,
0.0217057057,
0.0308295172,
0.005186778,
-0.0187435783,
0.0022136229,
0.0074861809,
0.0408668779,
0.0082286336,
0.0002904486,
-0.0113367038,
0.0409702882,
-0.0549312383,
-0.0018956864,
0.0403879173,
0.0529279001,
-0.0034975694,
-0.0356163979,
-0.0376550928,
0.0280956328,
0.0243491046,
0.0378528014,
0.0452236086,
-0.0202010814,
-0.093393974,
-0.0426371992,
0.0155826379,
-0.11793001,
0.0086539937,
0.0009253611,
0.0009619498,
0.0426711701,
-0.0944536924,
-0.0780106708,
0.0497408882,
0.0396706387,
-0.0215867944,
-0.0089276331,
-0.0025332295,
-0.0041056327,
-0.0648406744,
0.0108563267,
-0.036869254,
0.0237478688,
-0.0242709629,
-0.0552986637,
-0.0212103575,
-0.0391701125,
0.0008169278,
0.0260221418,
-0.1082452461,
-0.0449885391,
-0.0517398305,
0.0471501276,
0.0135647515,
0.0592000782,
-0.0289734341,
0.0354956202,
0.0320106447,
0.0086944401,
-0.0719072446,
-0.0120173227,
0.0113405818,
-0.0206249636,
-0.0084905978,
0.0331253372,
-0.059301205,
0.0192709919,
0.0030381787,
0.0883393288,
-0.010291188,
-0.0151159354,
0.0669553578,
0.012440254,
-0.0236440636,
0.0181068536,
0.0221786965,
-0.01377156,
0.0093955528,
-0.029865345,
-0.0279037952,
-0.0505462214,
0.0128468741,
-0.0133708622,
-0.0324363224,
0.0056674117,
-0.0276986863,
-0.0171199888,
0.0663827583,
-0.0241363216,
-0.0369912907,
0.0145859197,
-0.0391881391,
0.0334462784,
-0.0222372636,
-0.0239901263,
-0.0099567678,
0.0193039812,
0.0618361309,
0.0156345144,
0.0520732589,
0.0124967163,
-0.0237149633,
-0.0288986955,
0.0216757096,
-0.0722902715,
-0.03913863,
-0.0447715931,
0.0003420017,
0.052011203,
0.0272724312,
0.0174458399,
-0.0249847211,
-0.0075752665,
-0.0189438313,
0.0078437803,
0.0139984963,
-0.0278709847,
0.0502251014,
0.0160964131,
-0.010592049,
-0.0341507718,
0.0984251276,
0.0466589443,
0.0289559662,
-0.0055981874,
-0.0022092254,
-0.0414653756,
-0.045960471,
-0.0198748261,
0.0409972742,
0.0194899496,
0.0070428089,
-0.0174774099,
-0.0677515939,
-0.0068579693,
0.0257618614,
0.0050352761,
-0.0186284259,
0.0576267764,
0.0408743322,
0.0188525245,
-0.0480739959,
-0.0399851017,
0.0374680385,
-0.0193392653,
-0.0084190955,
-0.0300286934,
-5.91122938e-33,
-0.0403326526,
-0.0499935262,
0.0433775149,
0.0002993911,
-0.01304664,
-0.004061786,
-0.0562614389,
-0.0144975632,
0.0033845762,
-0.0028511554,
-0.0186474286,
0.0097867614,
0.0070881587,
0.0467795655,
-0.0062273913,
-0.0203178767,
0.0385302231,
-0.0000423358,
-0.0312443562,
0.0488921367,
0.0192214102,
0.004473384,
0.030795183,
0.0040709833,
0.0069613392,
-0.0234754607,
0.0148330657,
-0.0013689531,
-0.0137442024,
0.0541574359,
-0.0088960566,
0.0363114811,
0.0310465451,
-0.0010036541,
-0.0203448981,
0.0336689241,
-0.0824356899,
-0.0728569403,
0.0113148019,
-0.0183326229,
-0.0951326042,
-0.0348613225,
0.0091396384,
-0.0383327864,
0.0335228816,
-0.0370124169,
0.0344428793,
-0.0002347755,
-0.0171701349,
-0.044904083,
-0.031623859,
0.008036199,
-0.0031950767,
-0.0287616681,
-0.0577894188,
0.0440664962,
-0.0053353314,
-0.0476933271,
-0.010673346,
0.0039646397,
0.0163133554,
0.0059708529,
-0.0018776625,
0.0159878563,
0.0153426714,
0.0222330894,
0.0129635651,
0.0072182817,
-0.04758472,
0.0467590243,
-0.0612774193,
0.0491404422,
0.0190215092,
-0.0030259367,
0.0179488454,
0.035953559,
-0.0654211789,
0.0485794209,
-0.0163520332,
0.0587746985,
-0.009783186,
-0.0202149656,
-0.0170159973,
-0.0260564517,
-0.0246614311,
0.0276938602,
-0.001431542,
-0.0526885092,
0.0481317118,
-0.0058797458,
-0.0188225526,
-0.0149543863,
0.0065993294,
-0.0510023423,
0.0089527797,
0.0423864312,
-0.0444288366,
0.036586795,
-0.0060915761,
0.0540181622,
-0.0232840907,
0.0020571288,
0.0373350978,
-0.0169833861,
-0.0165080465,
0.0416336097,
0.017899435,
0.0289936345,
-0.0616785474,
0.0194523074,
-0.0426339209,
0.0318117999,
-0.0442701727,
-0.0023841402,
0.0452184491,
0.028217053,
-0.022281982,
0.0613134466,
-0.0230934657,
0.0189548191,
-0.0487296954,
0.0023795916,
0.0122264652,
0.0293087587,
-0.010196723,
0.0420417972,
-0.0637093037,
0.0274338033,
-0.0332704261,
-0.020443175,
0.004978932,
0.0417942069,
2.602e-7,
0.0268648621,
-0.0037929635,
0.0291689057,
0.0814295933,
0.0042158486,
-0.0023123161,
0.0209129378,
0.0235979389,
0.0148066934,
-0.0108216405,
0.0079563037,
0.0363400243,
0.0234049764,
0.0519860461,
-0.0866807476,
0.0410481766,
-0.0489511564,
-0.024699118,
-0.0350333005,
-0.0155510735,
0.0560107343,
0.0621651113,
-0.0233333576,
0.0012302558,
-0.0110663576,
-0.0289927758,
0.060211774,
0.0204794928,
-0.0130677186,
0.0298905317,
0.006127832,
-0.0149154607,
-0.0181070548,
0.0139470147,
-0.0189595036,
0.0068788221,
0.0258820225,
0.0454742387,
0.0027479986,
-0.0105083715,
0.0012527644,
-0.0666012913,
0.0294679161,
-0.0242965464,
0.0806222111,
-0.0439630486,
-0.026077183,
0.0309597924,
-0.0094404276,
-0.0311250631,
-0.0157761723,
-0.0241646096,
-0.0243009776,
0.0140287513,
0.0289188176,
-0.0274107959,
0.0181995239,
-0.0367634892,
-0.0014306267,
0.0355953574,
-0.0211835075,
-0.0276553892,
-0.0358077399,
0.0139526259,
0.0073798387,
-0.0251462031,
-0.0131333778,
2.385666474e-34,
0.0400908962,
-0.0223454293,
0.0218316838,
0.0068681459,
0.0119834226,
-0.0162324458,
0.0136166289,
-0.0112696011,
0.0505439453,
-0.1037720814,
-0.0357220806
] |
http://arxiv.org/abs/2401.13588v1 | http://arxiv.org/pdf/2401.13588v1 | Evaluation of General Large Language Models in Contextually Assessing Semantic Concepts Extracted from Adult Critical Care Electronic Health Record Notes | 2024-01-24T16:52:37 | 2024-01-24T16:52:37 | Darren Liu, Cheng Ding, Delgersuren Bold, Monique Bouvier, Jiaying Lu, Benjamin Shickel, Craig S. Jabaley, Wenhui Zhang, Soojin Park, Michael J. Young, Mark S. Wainwright, Gilles Clermont, Parisa Rashidi, Eric S. Rosenthal, Laurie Dimisko, Ran Xiao, Joo Heung Yoon, Carl Yang, Xiao Hu | cs.CL | cs.CL, cs.AI, cs.SE | The field of healthcare has increasingly turned its focus towards Large
Language Models (LLMs) due to their remarkable performance. However, their
performance in actual clinical applications has been underexplored. Traditional
evaluations based on question-answering tasks don't fully capture the nuanced
contexts. This gap highlights the need for more in-depth and practical
assessments of LLMs in real-world healthcare settings. Objective: We sought to
evaluate the performance of LLMs in the complex clinical context of adult
critical care medicine using systematic and comprehensible analytic methods,
including clinician annotation and adjudication. Methods: We investigated the
performance of three general LLMs in understanding and processing real-world
clinical notes. Concepts from 150 clinical notes were identified by MetaMap and
then labeled by 9 clinicians. Each LLM's proficiency was evaluated by
identifying the temporality and negation of these concepts using different
prompts for an in-depth analysis. Results: GPT-4 showed overall superior
performance compared to other LLMs. In contrast, both GPT-3.5 and
text-davinci-003 exhibit enhanced performance when the appropriate prompting
strategies are employed. The GPT family models have demonstrated considerable
efficiency, evidenced by their cost-effectiveness and time-saving capabilities.
Conclusion: A comprehensive qualitative performance evaluation framework for
LLMs is developed and operationalized. This framework goes beyond singular
performance aspects. With expert annotations, this methodology not only
validates LLMs' capabilities in processing complex medical data but also
establishes a benchmark for future LLM evaluations across specialized domains. | Evaluation of General Large Language Models in Contextually Assessing Semantic Concepts Extracted from Adult Critical Care Electronic Health Record Notes: The field of healthcare has increasingly turned its focus towards Large
Language Models (LLMs) due to their remarkable performance. However, their
performance in actual clinical applications has been underexplored. Traditional
evaluations based on question-answering tasks don't fully capture the nuanced
contexts. This gap highlights the need for more in-depth and practical
assessments of LLMs in real-world healthcare settings. Objective: We sought to
evaluate the performance of LLMs in the complex clinical context of adult
critical care medicine using systematic and comprehensible analytic methods,
including clinician annotation and adjudication. Methods: We investigated the
performance of three general LLMs in understanding and processing real-world
clinical notes. Concepts from 150 clinical notes were identified by MetaMap and
then labeled by 9 clinicians. Each LLM's proficiency was evaluated by
identifying the temporality and negation of these concepts using different
prompts for an in-depth analysis. Results: GPT-4 showed overall superior
performance compared to other LLMs. In contrast, both GPT-3.5 and
text-davinci-003 exhibit enhanced performance when the appropriate prompting
strategies are employed. The GPT family models have demonstrated considerable
efficiency, evidenced by their cost-effectiveness and time-saving capabilities.
Conclusion: A comprehensive qualitative performance evaluation framework for
LLMs is developed and operationalized. This framework goes beyond singular
performance aspects. With expert annotations, this methodology not only
validates LLMs' capabilities in processing complex medical data but also
establishes a benchmark for future LLM evaluations across specialized domains. | [
0.0623674691,
-0.0129561163,
0.0014560011,
-0.0020174813,
-0.0554104485,
0.0427125543,
-0.0157712419,
0.0213058572,
-0.0072758058,
-0.0391373225,
-0.0003352613,
-0.0868403912,
0.0023633973,
0.0292624645,
-0.0114246765,
-0.0442814119,
0.0220548976,
0.0030832468,
0.0376831181,
0.023276167,
-0.0036734513,
0.0062944787,
-0.0355588607,
0.0503356904,
0.0165513325,
-0.0067708055,
0.0169711001,
-0.0091089075,
0.0124752354,
-0.0892287418,
-0.0484109484,
0.0150549747,
-0.0147702675,
0.0066185738,
0.0000021872,
-0.075425446,
-0.0359623618,
0.0329494923,
-0.0483343452,
-0.0306807291,
0.0462708771,
-0.0377094038,
-0.0212022439,
-0.0011726279,
-0.0363411866,
-0.0591724738,
0.0334742069,
0.0472036675,
0.0444581434,
0.05932004,
-0.0227668211,
-0.0115725854,
0.0202909559,
-0.0169042889,
0.0206921771,
-0.055680871,
0.0456925035,
0.0539875478,
0.0430302545,
-0.0259969793,
0.0193420351,
0.0511724986,
-0.0012642508,
0.0403595231,
0.0014014046,
0.0505101234,
-0.0137797082,
-0.0100889681,
0.0136122899,
0.0380516797,
0.0516409688,
0.0454848893,
0.024158081,
-0.008028633,
-0.0228922069,
0.0346058384,
-0.0802802891,
0.0219304729,
-0.0287380219,
0.0012276667,
0.0154539458,
0.0181903671,
0.0167425275,
-0.0142519539,
-0.0436342247,
0.0216359254,
0.00809382,
-0.0490540452,
0.0354389213,
-0.0075598876,
-0.0635774136,
-0.019106986,
0.0404722691,
0.024678953,
-0.0166378785,
0.0328215435,
0.0401393287,
-0.0220491998,
0.0856782496,
-0.085828349,
0.0235565286,
0.0024499667,
0.0314316563,
0.0411175862,
-0.0009713673,
-0.0021210203,
-0.0226907693,
-0.0065260082,
-0.0483649112,
0.0051337257,
-0.0782745928,
-0.0292562284,
-0.0301119182,
0.0285890959,
-0.0036457193,
-0.0113868508,
-0.0826932564,
-0.0123024872,
0.0050383131,
0.0303639174,
0.0220721252,
-0.0023835276,
-0.0063807014,
0.0091556143,
-0.0446031801,
-0.0030340443,
-0.0042982381,
0.0108828954,
-0.0323785096,
-0.0413996801,
0.0506181233,
0.0209366363,
0.0084800245,
-0.0210671257,
-0.0023185436,
0.0377777629,
0.0302108452,
-0.0149865989,
-0.0061742146,
-0.0359154791,
-0.0461903811,
-0.0398051739,
-0.0032330675,
-0.0474889614,
-0.0075861705,
0.0288296994,
0.0119103566,
-0.0222358909,
0.0149254985,
0.0144755756,
-0.0352963619,
0.0291446075,
-0.0407640114,
0.019422859,
0.0145661132,
0.0381776802,
0.0229204409,
0.009529056,
0.0294815917,
-0.0367183909,
0.0066772732,
-0.0263050795,
-0.0004045216,
0.011510836,
-0.0242139921,
-0.0474786945,
-0.0279172324,
0.0519904345,
0.0317257941,
0.0520041175,
-0.058172863,
0.0921205357,
0.009499683,
0.0075056432,
0.0371198878,
0.0750599205,
0.0190541204,
0.0080259088,
0.0249246527,
-0.0248144921,
0.0204245765,
-0.0297106504,
0.0080861682,
-0.0044787279,
0.0330475532,
0.0026416548,
-0.0132322256,
-0.0281446762,
-0.0244373176,
-0.0363007858,
-0.0716865808,
-0.0010288742,
-0.0088118929,
-0.0027623817,
0.0513839237,
-0.0419820249,
-0.0310139488,
0.0266155954,
0.0073528951,
-0.0384884551,
0.0043997983,
-0.0362332687,
-0.0225148872,
0.1197304055,
0.0486331023,
-0.0196440034,
-0.0248082019,
-0.0293356162,
0.0240176897,
0.0134670511,
0.0055299983,
0.0021809454,
-0.0290588364,
-0.0346292481,
-0.0088759465,
0.0293047391,
-0.0027252098,
-0.0277732685,
-0.0322687104,
0.007053352,
-0.0016786344,
-0.0274635516,
-0.0319777913,
0.0325763412,
-0.047456041,
-0.0115330499,
-0.0349240601,
-0.0076186163,
0.0338308066,
0.0429661274,
-0.0012690038,
0.0458024554,
0.0510730334,
-0.0428822935,
-0.0231612287,
0.0107459165,
-0.051722493,
0.0211214628,
0.0038611898,
-0.0171316117,
-0.005532572,
-0.0279981997,
-0.0020934418,
-0.0511176251,
-0.0516289063,
-0.0432328321,
0.0413953066,
-0.0137262978,
0.0041506644,
0.0294949021,
-0.021355994,
-0.07702519,
0.0339415856,
-0.019829575,
0.0284339003,
-0.0041802693,
-0.0218492057,
0.0177083276,
-0.0234387387,
-0.0466920622,
0.0452682786,
-0.0066530365,
0.0215353873,
0.0004855282,
0.0111751473,
-0.0379012339,
-0.0342384875,
-0.1394989341,
0.0011545313,
0.0054171197,
0.0453078412,
0.058449138,
0.0127479387,
-0.0112986742,
0.0145752896,
-0.0358808823,
-0.0098314332,
0.0154336737,
0.0268142633,
0.0127797751,
0.0679085627,
-0.0006650981,
-0.0133734895,
-0.0043615648,
-0.0391541757,
0.0318127647,
0.0161649045,
-0.0255824085,
0.0064542484,
-0.0064554843,
-0.0267587025,
-0.0317909643,
0.0078931404,
-0.0003528554,
0.0245551951,
-0.0391414277,
0.0197973046,
-0.00832457,
-0.0068704509,
0.0336037874,
0.0164398067,
-0.0250656307,
-0.0056139408,
-0.0141530968,
-0.0347970314,
-0.0165129229,
-0.0194185451,
-0.040005073,
-0.0197807644,
0.0605533645,
0.0286695622,
0.0206107348,
-0.0340886712,
-0.0195283648,
-0.0773146972,
0.0365059003,
-0.0450212881,
0.0081953015,
-0.0516074076,
0.0340028442,
0.0143543156,
-0.0158996377,
0.0632844865,
-0.0255750436,
0.0171057209,
-0.0784556121,
-0.0464336351,
-0.0537429042,
0.0382330716,
-0.0513424948,
0.0206946526,
0.0004276685,
-0.0222539157,
0.0182956792,
0.0013924949,
-0.0110557945,
-0.0617010668,
0.0940347984,
-0.0338688307,
0.0518849343,
0.0067192605,
0.0030021877,
-0.0276308898,
-0.007432533,
-0.0409578346,
0.0100799398,
0.0022645122,
0.0339936353,
-0.0358252786,
0.0666468441,
0.0086929714,
0.0592235401,
0.0083138375,
-0.0206285901,
0.0014519666,
-0.0018088655,
0.0044808621,
-0.0163943134,
0.0122952862,
-0.0010220208,
0.0541598536,
0.0815012902,
0.0113326609,
0.0096542099,
0.006360162,
-0.0128103727,
-0.0096506737,
-0.0535721593,
-0.021698175,
0.0271067061,
0.0114570493,
-0.0626269802,
-0.0162511617,
0.1308549643,
0.0520818718,
-0.0506570637,
0.013286924,
-0.0165920351,
-0.0366149582,
-0.0165671073,
-0.0274677426,
-0.0329218507,
-0.0207946226,
-0.0350812897,
0.0398931727,
-0.0211226568,
-0.0391582176,
0.0192435067,
-0.0243744031,
0.0146130035,
-0.0101390183,
-0.0659263283,
0.0049129729,
-0.0433573462,
0.0199837107,
-0.0147241335,
0.0082532354,
0.0033507394,
-0.0383870602,
0.0032963061,
0.0229623709,
-0.0636360273,
-0.0837621242,
0.0620112494,
-0.0063936743,
0.0724055022,
0.0622669794,
0.0218362901,
-0.0382213481,
-0.0084948847,
0.0530018955,
-0.0547150262,
0.0099553056,
-0.0126344683,
0.0159919094,
0.0122388359,
0.0220870953,
-0.0446727797,
-0.0073907734,
-0.039457947,
-0.0276390743,
0.0153245861,
0.0929582268,
-0.061271593,
-0.0166950673,
-0.036488086,
0.0337572508,
0.0103121027,
0.0650467128,
0.0088382717,
-0.0072146324,
-0.0246954504,
0.0398097746,
0.031639576,
0.0386828519,
-0.0223782528,
0.0286770035,
-0.0071507315,
-0.0074586705,
-0.0316545293,
-0.0496245921,
0.0473265015,
-0.0053518475,
0.0055876207,
0.0665545091,
0.0468812026,
0.0110055087,
-0.0565101728,
0.0505454689,
0.0216265209,
0.0256842673,
-0.0629836619,
-0.0245133843,
0.0415081382,
0.0272361841,
-0.0118365223,
-0.0250761062,
-0.0084873075,
0.0199291985,
0.0370567963,
0.0113108028,
0.0255848374,
-0.0484160222,
0.0105429245,
-0.0376841947,
0.0090625742,
-0.0084422454,
-0.0340816453,
-0.0246887002,
-0.0143433223,
0.0293490849,
-0.0046350369,
0.0037568379,
-0.0126603078,
0.0517749488,
-0.0036261752,
0.0486452542,
-0.0081365276,
-0.0069337497,
0.0105281528,
0.036542967,
0.0185868219,
-0.0010754104,
-0.0209778957,
-0.0026716087,
0.0045740022,
-0.0378191434,
0.0243127327,
-0.0176336803,
0.0019758637,
0.0327916928,
-0.0124788787,
-0.0173077565,
0.0054139928,
-0.0789242983,
-0.0596041344,
-0.0300373342,
0.0570417494,
0.0079507688,
0.0043990556,
0.0224557072,
-0.019663794,
-0.0167478248,
-0.0030127715,
-0.015289464,
0.0017515486,
0.0080470303,
0.0710276142,
-0.0237089563,
-0.0047932947,
0.0135509679,
0.0547878072,
-0.0502471775,
-0.0185958706,
0.0012826181,
0.0197724402,
0.0517672375,
0.018829288,
0.0336358063,
-0.0335151367,
0.0286477972,
-0.0506362505,
0.002577598,
-0.0019320786,
0.0334722959,
0.0545872822,
0.0237361249,
-0.0150221754,
-0.0228683408,
0.0641129836,
-0.0056836866,
0.0274114031,
-0.0035692581,
-0.0468893722,
-0.0225081164,
-0.0274090562,
-0.0108407913,
-0.0035942749,
-0.0817502886,
0.0000860844,
-0.1135452688,
-0.0020379839,
0.0235653259,
0.0001578463,
-0.0193011649,
-0.0249529872,
0.0527431741,
0.0374149531,
0.0055745845,
-0.0177866891,
-0.0359941088,
0.0298147462,
-0.0370240882,
0.0135737024,
0.0131200058,
-6.129587007e-33,
0.0100609493,
-0.0572484918,
0.0307363216,
-0.0132952239,
-0.0053460845,
0.0345547535,
-0.0207856204,
0.0064001358,
0.0048293425,
0.0121276919,
-0.0074936063,
-0.0029987826,
0.0158154387,
0.0184089895,
0.0168799534,
-0.0441420451,
0.0010116196,
0.0292680152,
-0.0062667714,
0.0303187892,
-0.0036759831,
0.0299876668,
0.0494278483,
-0.101115793,
0.0562160276,
-0.0223183278,
0.017709488,
0.0021682743,
0.0091809658,
0.0735753849,
-0.0399574228,
0.0318254456,
0.0651856959,
-0.0355521888,
0.0032271475,
0.0113958884,
-0.0506455526,
-0.014272131,
0.0616950244,
-0.0394681096,
-0.008384726,
-0.029723864,
0.0408521183,
0.0140584288,
-0.0321801119,
0.0283744838,
0.0103472667,
-0.0262312107,
-0.0170247406,
-0.0160217751,
-0.0366285592,
0.0664638504,
0.0078922054,
0.0392512418,
0.0047607743,
-0.0086299116,
-0.062067043,
0.0190238934,
-0.0383220762,
-0.0155226672,
0.0169931855,
0.0070421617,
0.0094814645,
-0.0208553448,
-0.0296003632,
0.0518654399,
0.0036151544,
0.0309310928,
-0.0239485092,
0.0050053881,
-0.0430756882,
0.0286047179,
0.0476237424,
0.0424932092,
0.0023415484,
-0.0167033933,
-0.0973461121,
0.0330784246,
-0.0786913335,
0.0672115833,
0.0745424852,
-0.0366030373,
-0.0507761352,
-0.0140149593,
-0.0555917844,
0.0549046174,
-0.0233282726,
-0.0278491396,
0.0488314293,
-0.0210730508,
0.0131887291,
0.0414989702,
0.022642741,
-0.0063147163,
0.0148125337,
-0.0467340797,
0.0264890306,
0.0394920371,
-0.0285401344,
0.0096860463,
-0.0611780025,
-0.0072470093,
0.0366000719,
0.0000990084,
0.0315485522,
0.004335023,
0.0032692531,
-0.0190284904,
-0.0243432336,
-0.0157507751,
0.0242224485,
-0.0261799432,
-0.0363025777,
-0.0125824278,
0.0358928405,
0.0208985973,
0.0112537546,
0.0783317685,
0.0017221151,
-0.0266660862,
-0.0301674772,
-0.0671571344,
0.014637528,
-0.005923023,
-0.0357412808,
0.0263034571,
-0.0150306448,
0.0517835245,
0.063856706,
-0.0120910704,
0.0216777697,
0.0147348214,
2.914e-7,
0.121418409,
0.0671939701,
-0.0027730039,
-0.044925604,
0.0269833338,
-0.0026790197,
-0.0199938305,
-0.0129399672,
0.0146781411,
-0.0008371779,
0.0563983358,
0.006406263,
0.0362959206,
0.0539359152,
-0.0250077173,
-0.0435892045,
-0.0835891515,
-0.028796263,
-0.0538337417,
0.0294727683,
0.1004189029,
0.0372541547,
-0.0599081069,
-0.0171022899,
-0.0283651408,
-0.0295213312,
0.0392542742,
-0.0742974803,
-0.0125462702,
0.0304858424,
0.0158894919,
0.0133239487,
0.0369719416,
0.0542036444,
0.0193397161,
-0.017968785,
0.0338037349,
0.1230023578,
-0.0628481507,
0.03835411,
0.023767544,
-0.0312626772,
0.0368742049,
-0.0514912046,
0.0563437045,
-0.0688623562,
-0.0446400456,
-0.0037021053,
-0.0253175721,
-0.0113985492,
0.0711217076,
0.0011529777,
-0.0347260945,
0.0248186346,
0.0110989967,
0.0071742274,
-0.0489237495,
-0.0355482586,
0.0046496401,
0.047460027,
-0.016317619,
-0.0533361062,
0.023272641,
-0.0409340449,
-0.0157918632,
-0.0694258884,
0.0000627254,
2.680593874e-34,
0.0288467091,
0.0137565713,
-0.0028790005,
0.0137357879,
0.0249436349,
0.020766465,
0.0005348198,
0.0383262299,
-0.0189441554,
-0.0915737599,
0.0120136784
] |
http://arxiv.org/abs/2401.13586v3 | http://arxiv.org/pdf/2401.13586v3 | Instruction Fine-Tuning: Does Prompt Loss Matter? | 2024-01-24T16:51:23 | 2024-06-18T10:37:08 | Mathew Huerta-Enochian, Seung Yong Ko | cs.LG | cs.LG, cs.AI, cs.CL | We present a novel study analyzing the effects of various prompt loss token
weights (PLW) for supervised instruction fine-tuning (SIFT). While
prompt-masking (PLW = 0) is common for SIFT, some fine-tuning APIs support
fractional PLWs and suggest that using a small non-zero PLW can help stabilize
learning when fine-tuning on short-completion data. However, there has never
been a study confirming this claim, and OpenAI, a major cloud-based SIFT
provider, recently removed this parameter from their fine-tuning API. We found
that performance of models fine-tuned on short-completion data had a
statistically-significant negative quadratic relationship with PLW. Using small
values (0.01 - 0.5) of PLW produced better results on multiple-choice and
short-generation benchmarks (outperforming models fine-tuned on long-completion
data) while large values (~ 1.0) of PLW produced better results on
long-generation benchmarks. We explained this effect and verified its
importance through additional experiments. This research serves as a warning to
API providers about the importance of providing a PLW parameter for SIFT. | Instruction Fine-Tuning: Does Prompt Loss Matter?: We present a novel study analyzing the effects of various prompt loss token
weights (PLW) for supervised instruction fine-tuning (SIFT). While
prompt-masking (PLW = 0) is common for SIFT, some fine-tuning APIs support
fractional PLWs and suggest that using a small non-zero PLW can help stabilize
learning when fine-tuning on short-completion data. However, there has never
been a study confirming this claim, and OpenAI, a major cloud-based SIFT
provider, recently removed this parameter from their fine-tuning API. We found
that performance of models fine-tuned on short-completion data had a
statistically-significant negative quadratic relationship with PLW. Using small
values (0.01 - 0.5) of PLW produced better results on multiple-choice and
short-generation benchmarks (outperforming models fine-tuned on long-completion
data) while large values (~ 1.0) of PLW produced better results on
long-generation benchmarks. We explained this effect and verified its
importance through additional experiments. This research serves as a warning to
API providers about the importance of providing a PLW parameter for SIFT. | [
0.0005175491,
0.0293879025,
-0.0078615537,
0.0042439299,
-0.0085153384,
0.0063670706,
0.0350440256,
0.0208989568,
-0.015734978,
-0.0258945227,
-0.0127978846,
-0.0224956442,
0.0201526973,
0.0414040275,
0.0164469294,
-0.0664658546,
0.0527131818,
-0.0439286083,
0.0024575708,
-0.0249506328,
-0.0475644767,
-0.0197650585,
-0.0051786969,
0.0645281971,
0.0229598973,
-0.0026344827,
-0.0003174141,
0.039772626,
0.0347772948,
0.0013073934,
0.0359804519,
0.05862673,
-0.0172299668,
0.0099059474,
0.0000022009,
-0.0666636601,
0.0095877629,
-0.000186227,
-0.0424680859,
0.0282859523,
0.0198637042,
0.0441669673,
0.0086524617,
0.0359867774,
0.0336271636,
0.0506976396,
0.0413768925,
0.0271111317,
0.0253868289,
0.0560548678,
0.0155432792,
-0.0038085917,
0.0112486435,
-0.0247633755,
0.003413233,
-0.0135635985,
-0.006910041,
-0.0534076877,
-0.011993777,
0.0045038722,
-0.0568250231,
0.0196170229,
0.0407369472,
0.0295320097,
0.0204199683,
0.0557874925,
-0.0209162217,
-0.04243147,
-0.0219392776,
0.0532405227,
-0.0550420694,
0.0190047137,
-0.0104149068,
-0.0365252607,
0.0273735393,
-0.0147322984,
-0.0287298355,
0.0326034129,
-0.016593257,
0.0069453255,
-0.0213209372,
-0.0119402306,
0.0012497726,
0.0146073522,
-0.0193851963,
0.0524677634,
0.0036966496,
-0.0123833446,
-0.0004760782,
-0.0331488065,
-0.0491161719,
-0.0430366173,
-0.0299385041,
0.0309923589,
-0.0431938544,
-0.00232439,
-0.0572185256,
-0.0127704814,
0.0321606994,
0.0368311964,
0.0565517955,
0.0458424799,
-0.061130628,
0.0055122511,
-0.0210604612,
0.0474623181,
0.0261213891,
0.0763198957,
-0.0200369842,
0.0308155678,
-0.0015896811,
0.0236963928,
-0.0264424719,
0.0766913518,
-0.0543446578,
-0.0273036771,
-0.063749373,
-0.0047900914,
-0.0377876945,
0.0200259425,
-0.005711515,
0.0166642088,
-0.0692632645,
0.0318314843,
-0.0306169968,
0.044613298,
-0.0692565143,
0.0138046667,
0.0365656205,
-0.0553357303,
-0.0185148176,
0.0011429735,
-0.0182498507,
-0.0059376671,
0.0633569732,
0.0166603047,
-0.0183009058,
-0.0346460082,
-0.0103520676,
-0.0626005903,
-0.0161030013,
-0.0458937846,
-0.0173348151,
-0.0668930411,
0.005176723,
0.0288195536,
0.0177209545,
-0.0133826006,
-0.0082869856,
0.0371082313,
0.0010267107,
0.0282253493,
-0.0000809368,
-0.0156145981,
0.0368162505,
0.0149701675,
-0.0267356392,
0.0569084585,
-0.0041825385,
0.0429447293,
0.0380599387,
-0.0478219539,
0.0024761453,
-0.0108337598,
0.0408734493,
0.0016278673,
-0.0127361314,
-0.0077751968,
0.0402671322,
0.0232536905,
0.0391951539,
0.0332378484,
-0.0152793517,
-0.0280419346,
0.0445352979,
0.0711481199,
0.0219150148,
-0.0557230935,
0.0250742696,
0.0106169749,
0.0042607635,
0.0267950092,
-0.0587858148,
0.0040818793,
-0.033123903,
-0.0016330164,
0.0337340012,
0.0320853032,
-0.0294160731,
-0.038900461,
0.0135433786,
-0.0096539445,
-0.0359305404,
0.0525934063,
0.0300847162,
-0.0250304956,
-0.0301360413,
0.0066697286,
0.0268505309,
-0.0259971116,
-0.0032446538,
-0.1000313386,
0.0466347374,
0.0231974181,
0.0266310479,
-0.006337279,
-0.0215495955,
-0.0068608378,
0.0659284517,
-0.0288778823,
0.0560117587,
-0.0046660234,
0.0376030579,
-0.04596534,
-0.0484103411,
-0.021296734,
-0.0287033077,
-0.0016728941,
0.0295008402,
0.019284552,
0.0039333864,
-0.0188504793,
-0.0586731657,
0.0199749935,
0.0080938153,
-0.0774544403,
0.0067333425,
0.033641912,
0.0097645819,
0.0038956129,
0.0117265293,
0.0083611216,
0.0343167,
0.0316050239,
0.0216251817,
-0.0697218999,
-0.0593162179,
0.0578954257,
0.0250594448,
-0.0202853084,
0.034477178,
0.0463765822,
-0.0670810193,
-0.061768841,
-0.0194829218,
-0.0171776041,
-0.0117914546,
0.0031608739,
0.0337946005,
-0.0449100286,
0.0334401503,
-0.0500272475,
0.0088299699,
0.0035456528,
0.0295138005,
0.0301197413,
0.0391033068,
0.0079110814,
-0.0304020364,
-0.0240617096,
0.0255323611,
-0.0316613242,
-0.0132242115,
0.0131033286,
-0.0082307337,
-0.0813703388,
0.0246752948,
-0.0830351263,
0.0312149506,
0.0023882966,
0.0036187277,
0.0008812873,
0.0124648595,
-0.0130211655,
-0.0338595957,
-0.021396881,
0.0010721819,
-0.0057007005,
0.0312181953,
0.0065497463,
0.0272954106,
-0.0253685899,
-0.0062507023,
-0.0420081802,
-0.0084122149,
-0.0054608625,
0.0065326579,
0.0044989674,
-0.0443598665,
0.0193005968,
-0.0048750895,
0.017589936,
0.0029175382,
-0.0317428,
0.017200958,
-0.003632447,
-0.0088561326,
0.1142778844,
0.0176491905,
0.0372678302,
-0.0006790554,
0.0263855737,
-0.052924186,
0.0214413311,
-0.0379253626,
-0.0478450879,
-0.0257360917,
0.0012052261,
0.0042737522,
0.067197375,
-0.0018983217,
-0.0047688838,
-0.0625370964,
0.0109755071,
-0.0324576609,
0.0014364257,
-0.0114902128,
0.016558798,
-0.0262727905,
0.0149656357,
-0.0303886402,
-0.0966548473,
0.081212908,
0.002444247,
0.014011384,
-0.024315197,
-0.0038741806,
-0.0025833987,
0.0286878869,
0.0003486667,
0.0045716777,
0.0138176959,
-0.0343938433,
-0.0754379109,
-0.0262302104,
0.0063891574,
-0.018033836,
0.0376368016,
-0.028726289,
-0.0589308403,
-0.0556017272,
-0.0368277431,
-0.0229206793,
-0.0553472005,
0.0307726525,
0.0183949955,
0.0607646368,
-0.0535663776,
0.0061774184,
-0.053526476,
0.0073764715,
0.047281906,
-0.0442769751,
-0.0399361216,
-0.0393672287,
-0.010504568,
0.0645770952,
-0.0151333977,
0.0364787877,
-0.0486261509,
0.0517432429,
0.0417881459,
0.080024533,
-0.0242001731,
-0.0010635264,
0.0198505428,
0.0376794972,
-0.0549861118,
-0.007826251,
0.091745995,
0.117025286,
-0.0691619664,
0.0047623967,
0.0621756017,
0.0288370531,
-0.0343817435,
-0.007391829,
0.0254766252,
0.049198959,
0.0283897649,
0.008584273,
-0.0479834154,
-0.0509732813,
0.014926468,
0.1022310629,
-0.0309765879,
-0.0115365162,
0.0298121665,
-0.0977689475,
0.0404341444,
0.0104860039,
-0.0646063611,
-0.0239779986,
-0.0383021571,
-0.0158815905,
-0.0012274913,
0.0124064405,
0.0048293876,
-0.05144272,
0.014495004,
0.062755987,
0.0013185899,
-0.0357692987,
0.0440564267,
-0.057266701,
0.0234506186,
0.0225636158,
-0.0164120067,
0.0103857433,
-0.0535330214,
0.0746635646,
-0.0298169255,
0.0283946637,
-0.0024586581,
0.0831553787,
-0.0218066573,
0.0132615501,
-0.0042941864,
0.0632273629,
0.0004830463,
-0.0100867916,
0.0036597156,
0.0447653309,
-0.0299430881,
-0.0169781912,
-0.0349722914,
-0.0288205221,
-0.0283916257,
0.0408582725,
0.0277938228,
-0.05154882,
-0.0224901978,
-0.0432299003,
0.0188417323,
-0.046689596,
0.0396059006,
0.0266796704,
-0.0131481998,
-0.0475882627,
-0.0173262488,
-0.0565138906,
0.0492726937,
0.0421142541,
0.0021098787,
-0.0213806015,
-0.0521957688,
0.018343376,
-0.0346383639,
0.0554470308,
-0.0471209325,
0.0337382108,
-0.1036371216,
0.0078702988,
0.0026460683,
-0.0001227242,
-0.0232380498,
0.0255475659,
-0.0985204801,
-0.056180656,
0.0208518822,
0.1044699699,
0.0244303588,
0.018531315,
0.0222157724,
-0.0513215885,
0.0522426367,
0.05817746,
-0.0164033119,
0.0190373622,
0.0313655026,
-0.0548678301,
-0.0121096149,
-0.0091379909,
-0.0308392085,
0.0125822732,
0.0426709652,
0.0152365938,
0.0279754996,
-0.0205449332,
0.0036228616,
0.0381415002,
0.0713160858,
-0.0440460779,
-0.0372994021,
-0.0118238982,
0.0480973944,
0.0408339202,
0.0128087299,
-0.0101296129,
-0.052146852,
0.0145994304,
0.0007709638,
0.0071798586,
0.0051609315,
-0.0003848858,
0.0214878786,
-0.038070824,
0.0444317758,
-0.0151561694,
0.0484840125,
0.0454570688,
0.0043252381,
-0.0222615004,
-0.0000645837,
0.0163613632,
0.0410816967,
0.0155108441,
0.071756117,
0.0144140003,
0.0515098199,
-0.0081346333,
0.0074528758,
-0.0653965697,
-0.0970499963,
-0.0584233291,
-0.0171671249,
-0.0036677448,
-0.018031396,
-0.022512842,
-0.0129131777,
-0.0126187857,
-0.0247135479,
0.0137888389,
0.0009479318,
-0.0182564966,
0.0084778899,
0.0231200047,
-0.0016628866,
-0.0164795201,
0.0554925054,
-0.0474768281,
0.0665497929,
-0.0406091996,
-0.0016985844,
-0.0229517203,
-0.0063428404,
0.023167694,
0.0478731766,
0.0034301151,
0.0190491099,
-0.0531984381,
-0.0239489619,
-0.0096872449,
-0.0122180441,
0.0370969325,
-0.0191509165,
0.050820902,
0.009427622,
-0.0275070686,
-0.0513498634,
-0.0258908737,
-0.0028063483,
0.0143700726,
-0.00031907,
0.0119996071,
-6.584401567e-33,
-0.032704398,
-0.0316746123,
0.004036908,
0.0176973995,
0.0068434635,
0.0012595,
-0.0076589608,
0.009156934,
-0.0221115928,
-0.0127775157,
-0.0117586013,
0.0001739977,
0.0280751418,
0.0398692228,
0.0284851808,
-0.017858116,
0.0475846343,
-0.0158282779,
-0.025801979,
0.0231632162,
0.0421295837,
0.0139908185,
0.0148095554,
0.0100416662,
-0.0066848644,
0.00240923,
-0.0016256184,
-0.0315041952,
-0.0395387411,
0.0481796488,
0.0151283415,
0.0392402709,
-0.002261302,
-0.0611306094,
-0.0090113552,
0.0159258861,
-0.0575623102,
-0.0427651852,
-0.0365595296,
0.0056459899,
-0.0407277644,
-0.0563182943,
0.0340443626,
-0.0069264527,
-0.0469852798,
-0.0383842178,
0.0357121751,
-0.0092818672,
-0.0302810837,
-0.0312819034,
-0.0455211177,
0.0038937982,
-0.0042668618,
-0.0431590863,
0.0038424779,
0.0810163021,
0.0080951937,
0.0118313748,
-0.0300803632,
0.0168801434,
0.0606932007,
0.0642438829,
0.0634174868,
0.0472377166,
-0.0213571992,
0.0044910177,
0.0877767652,
0.0132050533,
0.0295857694,
-0.0228283498,
0.0589198582,
0.0314555839,
0.0395664051,
-0.0212712903,
0.078069441,
-0.0481171757,
-0.0360177495,
0.0368527286,
-0.0221548323,
0.0330856778,
-0.0052633137,
-0.0185425226,
-0.0403349735,
-0.0132290162,
-0.0518273674,
-0.0145643242,
-0.0344117582,
0.00768713,
0.0691014081,
0.0065022674,
0.0032868348,
0.0233389288,
0.0478213839,
-0.0312855542,
0.004358632,
-0.0459719971,
-0.0162174851,
0.0378449932,
-0.0020719836,
-0.0283355787,
-0.0292215347,
-0.0693500638,
0.0381411165,
-0.0410733074,
0.0640734881,
0.0003343503,
0.0511867628,
0.0375838391,
-0.0621430054,
-0.0046239146,
-0.0272422452,
0.0254434533,
0.0492916666,
0.0024819272,
0.0477716252,
0.0025425188,
-0.0056038951,
0.0530767702,
-0.0153272506,
-0.022219222,
-0.0421067663,
-0.0252802763,
-0.0268073957,
0.0188869312,
-0.02637472,
0.0220530517,
-0.0152846565,
0.0116057834,
-0.0243511908,
-0.027790606,
-0.052108068,
0.0136557966,
2.905e-7,
0.0074948161,
0.0345876403,
0.0335882567,
0.0194909386,
-0.0369316004,
0.0316371806,
0.0033072627,
0.0027741345,
0.0146551933,
0.0239115302,
-0.0071007628,
0.0002682961,
0.0377209857,
0.0117706992,
-0.0635413229,
0.0048754518,
0.0048174174,
0.0603414141,
-0.0665432587,
0.06011305,
0.019391641,
0.0655389503,
0.0140490234,
-0.0295162722,
0.0605417974,
0.0368132032,
0.0252520368,
-0.0251237396,
0.042261254,
0.0154159022,
0.0036848411,
0.0055896705,
0.0112042977,
0.0463406481,
-0.0388714075,
-0.0220436621,
-0.0030566033,
0.0442805067,
-0.009005242,
-0.0354676358,
0.0051795919,
-0.0006886658,
0.0118892323,
-0.0667861626,
0.0601075664,
-0.0445533693,
-0.0173173957,
-0.0125059243,
-0.019550465,
-0.040439222,
0.0659654289,
-0.0367894955,
0.0296237096,
-0.0063274242,
0.0102992179,
-0.0030507906,
-0.0130260652,
-0.0766173303,
-0.0041119438,
0.044594001,
-0.0242452603,
-0.0785931945,
-0.0788916349,
-0.0206843466,
0.0126645062,
-0.0400463566,
-0.0500411838,
3.014485927e-34,
0.0269406158,
-0.0650735497,
0.0596867688,
0.0018174481,
-0.0029244106,
0.0005281165,
0.000032555,
0.0334861092,
0.0278601926,
-0.0163106136,
-0.0440839864
] |
http://arxiv.org/abs/2401.13565v3 | http://arxiv.org/pdf/2401.13565v3 | Large Malaysian Language Model Based on Mistral for Enhanced Local Language Understanding | 2024-01-24T16:21:28 | 2024-02-04T06:52:28 | Husein Zolkepli, Aisyah Razak, Kamarul Adha, Ariff Nazhan | cs.CL | cs.CL | In this paper, we present significant advancements in the pretraining of
Mistral 7B, a large-scale language model, using a dataset of 32.6 GB,
equivalent to 1.1 billion tokens. We explore the impact of extending the
context length, releasing models with context lengths of 4096 and 32768 tokens,
and further refining performance with a specialized 16384 context length
instruction-tuned model, we called it Malaysian Mistral.
Our experiments demonstrate the efficacy of continue pretraining and the
influence of extended context lengths on Mistral 7B's language understanding
capabilities. Additionally, we release a model specifically tuned with a 16384
context length instruction, showcasing its potential for capturing nuanced
language intricacies.
Furthermore, our research contributes to the benchmarking of Malaysian
Mistral against prominent language models, including ChatGPT3.5 and Claude 2.
We present compelling results indicating Malaysian Mistral's superior
performance on Tatabahasa (Malay grammar) test set, particularly when
fine-tuned with instructions.
All models released at
https://huggingface.co/collections/mesolitica/malaysian-mistral-7b-6528f2ec825f4bba46c1700c | Large Malaysian Language Model Based on Mistral for Enhanced Local Language Understanding: In this paper, we present significant advancements in the pretraining of
Mistral 7B, a large-scale language model, using a dataset of 32.6 GB,
equivalent to 1.1 billion tokens. We explore the impact of extending the
context length, releasing models with context lengths of 4096 and 32768 tokens,
and further refining performance with a specialized 16384 context length
instruction-tuned model, we called it Malaysian Mistral.
Our experiments demonstrate the efficacy of continue pretraining and the
influence of extended context lengths on Mistral 7B's language understanding
capabilities. Additionally, we release a model specifically tuned with a 16384
context length instruction, showcasing its potential for capturing nuanced
language intricacies.
Furthermore, our research contributes to the benchmarking of Malaysian
Mistral against prominent language models, including ChatGPT3.5 and Claude 2.
We present compelling results indicating Malaysian Mistral's superior
performance on Tatabahasa (Malay grammar) test set, particularly when
fine-tuned with instructions.
All models released at
https://huggingface.co/collections/mesolitica/malaysian-mistral-7b-6528f2ec825f4bba46c1700c | [
0.0334309936,
0.0169019047,
0.004383787,
0.0476903133,
-0.0042756475,
0.0282894187,
0.0248397589,
-0.0172090046,
0.0071285544,
-0.0405711196,
-0.0869425312,
-0.0652692765,
-0.0107949227,
0.0364243723,
0.0241621435,
-0.0887724459,
0.0934623554,
-0.0581818223,
-0.0034146321,
0.0240606181,
0.0117522078,
0.0222580004,
0.008499166,
0.0242068302,
-0.0076753846,
-0.0432110503,
-0.0349348113,
0.0167213716,
0.0180198122,
-0.0492754392,
0.003016154,
0.0366095155,
0.0545136109,
0.0645825863,
0.000002217,
-0.0465421304,
0.0037094809,
-0.006388051,
-0.0378121063,
0.040294122,
0.0482944325,
0.0127256103,
-0.0099879457,
0.0321647599,
-0.0333260596,
-0.0088712024,
0.0076763164,
0.0348243192,
0.0534623191,
0.0987161696,
-0.008942347,
-0.043136552,
-0.0074753477,
-0.0347877741,
-0.0064293463,
-0.0404958986,
-0.0123817269,
-0.0055893403,
-0.0053007244,
-0.0235082954,
-0.046564769,
0.0347844847,
0.0344348699,
-0.0012444046,
0.0111763049,
0.0072137872,
-0.0325186774,
-0.0421652757,
-0.0054047932,
0.0520736389,
-0.0160954539,
-0.0328947045,
-0.0187970363,
0.0291351248,
-0.0112320064,
-0.0329580307,
-0.0370004922,
-0.0177139007,
-0.0323745944,
0.0303667951,
0.0549080819,
-0.0156467538,
0.0342897326,
0.0036990505,
-0.0354482867,
0.0068027661,
-0.0129975034,
-0.0521143675,
-0.0192558672,
-0.0198641364,
-0.0242606904,
-0.0615434386,
0.0346564651,
0.0492122062,
-0.0452050492,
0.0261743907,
-0.02160386,
0.0107504744,
0.0189993437,
-0.0636706725,
0.0140757021,
0.076815024,
-0.0254009794,
0.0450338051,
-0.0312178303,
0.0412780643,
-0.0181284063,
0.0246321168,
-0.0691332743,
-0.0279102474,
-0.0466104448,
-0.0116651962,
-0.067654483,
0.0086638201,
-0.0861620381,
-0.0227299239,
-0.021985447,
0.0319493674,
0.0330062583,
0.0092624296,
0.0673664585,
-0.0215220954,
-0.0565221794,
-0.0058758454,
0.0016046765,
0.0062214783,
-0.0258717854,
0.0132502755,
0.0244851951,
-0.0763477683,
-0.0247350037,
0.0265850239,
0.0011311756,
-0.0438580364,
-0.0093088821,
0.0438571982,
-0.0301443785,
-0.0530625097,
-0.0232114941,
0.0126201566,
0.0075890748,
-0.0710724145,
0.0040180283,
-0.0212796349,
0.0228299666,
0.0192350373,
0.0178845394,
0.0551880449,
-0.0156726409,
-0.0051738382,
0.0163548924,
0.0404787846,
-0.0427662507,
0.0156079913,
0.0183478519,
0.0255539045,
0.0085308421,
0.0628929436,
-0.0098189572,
0.0236802511,
0.0116571765,
-0.0335676894,
-0.0402591527,
0.0152190914,
0.0271296166,
0.0043389746,
0.0011628658,
0.0338406079,
-0.0353171937,
0.0234844759,
0.0523174442,
0.0385084823,
-0.0321048051,
0.008287481,
0.0298815984,
0.0359322876,
0.0377112329,
0.0263630953,
0.0691654533,
0.0282118879,
0.0114331273,
0.0078943511,
-0.0314266086,
0.0034923148,
-0.0168094579,
0.0316447914,
-0.0022626452,
0.0423411913,
-0.0464027151,
-0.0214743037,
-0.0328465402,
-0.0023047831,
-0.0201275144,
0.0378297009,
0.0428205654,
0.022743715,
-0.0385019928,
0.0882092044,
0.0513247438,
-0.0202640221,
0.0070769908,
-0.0607971735,
0.0249070302,
0.0625646561,
0.0006983154,
-0.0243539438,
-0.0264646132,
-0.0072876913,
0.011196062,
0.0381812863,
0.0409570932,
-0.0090731718,
-0.0014095046,
-0.0582312644,
0.0122135784,
-0.0090916604,
0.0246020146,
0.042425517,
0.0183104649,
0.0209153965,
0.0016931196,
-0.0467803255,
0.0255078319,
0.0429524705,
-0.0062498604,
-0.020209264,
-0.0667227954,
0.000037281,
0.0444994345,
0.0123999715,
-0.016173536,
-0.0261680149,
0.0223184619,
-0.0317975059,
0.0303317122,
0.0377041772,
-0.1207515225,
-0.0622548722,
0.0043051965,
0.0224027578,
-0.0183457881,
0.0240083989,
-0.0480649956,
-0.0592835397,
0.0128613906,
0.0074929716,
0.0586078614,
-0.0577946752,
0.011066149,
-0.0173242874,
0.0185298827,
-0.013748426,
0.0548603944,
-0.0252937507,
-0.0164508559,
0.005036124,
0.0155376447,
0.0289796293,
-0.041569721,
-0.022773331,
-0.0126319844,
-0.0059064324,
0.0001363087,
-0.0386207439,
0.0185631663,
-0.0354141779,
-0.0424497686,
-0.1371343881,
-0.0106880786,
-0.0490795523,
0.0277027301,
0.0166600887,
0.0221257657,
-0.0285084099,
-0.0395923294,
0.0212946311,
-0.0259627402,
0.0038520931,
0.0142723201,
-0.0021450627,
0.0044261469,
-0.0249689166,
-0.0403896421,
-0.0345631838,
-0.0062793149,
0.0396728739,
-0.0207947697,
-0.0460829847,
-0.0460314751,
-0.0075114947,
-0.0121019464,
-0.0010741291,
-0.0226936415,
-0.0563281104,
0.01516324,
-0.0163588114,
0.0210603643,
0.0586138219,
0.0060473778,
0.0324723907,
0.0044575343,
-0.0119115356,
-0.0459423475,
0.0202605873,
0.0114108566,
-0.0273218714,
-0.0483249277,
0.0150727751,
-0.0177250523,
0.1280705035,
-0.00926953,
-0.0035719646,
-0.0456720479,
-0.014098566,
-0.021581592,
-0.0152467247,
-0.0479761995,
-0.007168008,
-0.0332078934,
0.0308246091,
-0.0305717718,
-0.0353057161,
0.0777546987,
-0.0262814835,
-0.0060165529,
-0.070822373,
-0.0183852725,
-0.029963864,
0.0091218846,
-0.0237616748,
0.0208257157,
-0.0356189162,
-0.0405721404,
-0.0242784917,
-0.0056182942,
0.004970639,
-0.0006542691,
0.0090933908,
-0.034306705,
-0.0022452837,
-0.0245118197,
-0.0432160757,
-0.0396570377,
-0.0295120664,
-0.0268744882,
0.0283810142,
0.0375022367,
0.0171799455,
0.0021560634,
0.0273607615,
-0.0087290592,
0.0428230055,
-0.0148266442,
0.0069761849,
-0.0280813985,
0.0050316136,
0.0140838278,
0.0006312245,
0.0282596126,
-0.0436060093,
0.0544046089,
0.0181028042,
0.0308997557,
-0.0191440824,
-0.0167506356,
0.0313045755,
-0.0096119838,
-0.0244240202,
-0.0351298563,
-0.0091756172,
0.1140410006,
-0.0623169467,
-0.0333049819,
0.1054002419,
0.0241912846,
-0.0047031199,
-0.0101050129,
-0.0631863326,
0.0290148575,
0.0205902494,
0.0064387186,
-0.0725964159,
-0.0249730628,
0.0144437132,
-0.0249563586,
0.0316778123,
0.0113951173,
0.0038004764,
-0.0564352497,
0.0041498933,
-0.0118205464,
-0.0628327802,
-0.0055229552,
-0.0218715109,
-0.0449421555,
0.0017675327,
0.1100196987,
0.0497313775,
-0.0222024471,
-0.0198324192,
0.0644607767,
0.0078618322,
-0.0701392516,
0.066086933,
-0.0255905334,
0.0108673777,
0.0334726088,
-0.0010638173,
0.0093216365,
-0.0315392129,
0.0587932989,
-0.0080089886,
0.036580231,
-0.013367964,
0.0037096867,
0.0170457792,
-0.0330628976,
-0.0166495536,
-0.000133139,
-0.0228133854,
0.0140581094,
0.005991458,
0.0541005768,
0.0141607905,
0.0167381428,
0.01317601,
-0.0068557509,
-0.0190179683,
0.0509269424,
0.0516475365,
-0.0337811708,
-0.0278946906,
-0.00722691,
-0.0295521133,
-0.0322959311,
0.0305784289,
0.0292720422,
-0.0307533741,
-0.0222493559,
-0.0144506376,
-0.0421803556,
0.0121692,
0.0410057642,
-0.006079434,
0.0043845726,
0.0164415687,
0.0064246622,
-0.0372326002,
0.0872835815,
-0.019215554,
-0.0103270421,
-0.073346585,
-0.0122593725,
-0.0125642484,
-0.0052556233,
-0.0343453176,
0.008763901,
-0.0678645447,
-0.0358834155,
0.0394471511,
0.02588569,
0.0238913503,
0.0442211479,
-0.0054920106,
-0.0252222437,
-0.0182690956,
-0.0009152736,
-0.0025817316,
0.0187071022,
0.0395547338,
0.0006222928,
0.0016518894,
0.0280603841,
-0.0107684759,
0.0008125122,
0.0240724236,
0.014332979,
-0.0348380692,
0.0886452943,
0.021880256,
0.0101957181,
0.0501069315,
-0.0349557176,
0.0194513611,
-0.0506924167,
0.0469668619,
0.0009559388,
0.0291740149,
-0.026246272,
0.0027942965,
0.0372535065,
-0.0386407785,
0.0039250976,
-0.0075721778,
-0.0150935445,
0.0037885832,
-0.0418742374,
0.039443709,
-0.0000815426,
0.0198682379,
0.0504193231,
0.0039198669,
-0.0343735963,
-0.0043588318,
-0.0034260296,
0.0294167511,
-0.0276893098,
0.0899231806,
-0.0036415679,
0.0166214947,
0.0024083862,
0.0396675654,
-0.0058182701,
-0.0290395617,
-0.0194617435,
0.0513313077,
0.0868593305,
-0.0289881136,
0.012358794,
0.0112138549,
0.0026379761,
-0.0190501325,
0.0244405102,
0.0035352618,
-0.0370435379,
0.0302580371,
0.0106752245,
-0.030202806,
0.0325801522,
0.0578920357,
-0.013180281,
0.0492286533,
0.039574977,
0.0390056558,
-0.029034324,
0.0206765141,
0.0271107554,
0.0554891676,
-0.010936508,
0.0181792881,
-0.0499808639,
-0.0826331675,
-0.0069629084,
0.0313492976,
-0.0027144377,
-0.0020047568,
0.0281999074,
0.0535351932,
-0.0045559378,
-0.0573019758,
-0.0206653718,
-0.0076819276,
0.0396811813,
0.0129652563,
0.0017082078,
-6.843584847e-33,
-0.0548046827,
0.0577124022,
0.0169797335,
-0.0509668663,
-0.0093088727,
-0.0126201138,
-0.0513866358,
-0.0276220385,
-0.0435610674,
0.0023176223,
-0.0333296023,
0.0320139192,
0.031154735,
0.0335981026,
0.0040608579,
-0.0425447747,
0.0498601012,
0.0184538309,
0.0330548026,
0.0210156273,
0.0406585597,
0.0417051837,
0.0938576981,
-0.0112941042,
-0.0080389539,
-0.0090489443,
0.0018445302,
0.037891537,
-0.050069645,
0.0672619566,
-0.0372548588,
0.0519950986,
0.0145059219,
-0.0062359893,
-0.0639185235,
-0.0316827409,
-0.1061599776,
-0.0141611574,
0.0376254208,
0.0444252901,
-0.006730495,
-0.0011873208,
0.0316229016,
-0.0011138666,
-0.0105080418,
0.0035716642,
-0.0060901144,
0.0063959872,
-0.0370584242,
-0.0209084898,
-0.0462503098,
0.0300739147,
-0.0077525894,
-0.0199315567,
0.0464876816,
0.0191990454,
-0.0226326548,
0.0055085961,
-0.0852630213,
-0.0473150313,
0.0560297668,
0.0475798063,
0.0511398725,
0.0286184773,
-0.0002091038,
0.0119577805,
0.0789443552,
-0.020172881,
0.0104454244,
0.0485862643,
0.0573239252,
-0.0050859177,
-0.0090941684,
-0.001937451,
0.0500111729,
-0.0318819098,
-0.068767339,
-0.0412755832,
0.0104577076,
0.0452858731,
0.0288643446,
0.0299385618,
-0.0515796952,
-0.0255809706,
-0.0142976604,
-0.0233577602,
-0.0357362516,
-0.0005002026,
0.0316447914,
0.0160813648,
0.0214386117,
-0.0200324096,
0.048056107,
-0.0377699286,
0.0073476178,
-0.0172103271,
0.0155146504,
-0.004695273,
-0.0271663517,
0.0659463182,
-0.0513392128,
-0.0585547835,
0.0153166261,
0.01973176,
-0.0076691438,
-0.0098456377,
0.0292242039,
-0.0014932528,
-0.0498430692,
0.0196949299,
-0.0425829925,
0.0007772522,
-0.0137489606,
0.0337699428,
0.0480134226,
0.0139358323,
-0.0165413879,
0.0685052052,
-0.0023142744,
-0.0147568528,
-0.0253946241,
0.0026023029,
-0.0308271442,
0.0395342484,
-0.024237249,
0.0058801281,
-0.0357305259,
-0.0125356652,
0.0202682223,
-0.0494651906,
-0.0013920394,
0.0632745177,
2.797e-7,
0.0229059663,
0.0608699061,
0.0347945951,
-0.0102273347,
0.0146167707,
0.0112884054,
-0.0189853646,
-0.0121643785,
0.0230577309,
-0.0171200652,
-0.0043518855,
0.023436958,
0.0310293641,
0.0144787533,
-0.1219018772,
0.0080895461,
-0.1082001403,
0.0012919753,
-0.0417586826,
0.0225749072,
0.0051168869,
0.1397717595,
-0.0116653461,
0.0024741376,
0.0320118479,
0.0205199812,
0.0320578329,
0.005510977,
-0.0350730047,
0.034122292,
0.0543905757,
-0.0025328903,
0.0110924104,
-0.0112712365,
-0.0089223972,
0.0381511748,
-0.0462092534,
0.0742102265,
-0.0130320927,
0.0010718934,
0.0209957939,
-0.0193978101,
-0.0375091918,
-0.0880481973,
0.0662487522,
-0.0330576152,
-0.0182281248,
-0.0046652765,
-0.0182290599,
-0.0092745135,
-0.0006708641,
0.0060508307,
-0.0080886707,
0.0650248006,
0.0129568614,
-0.0358220041,
-0.0228738692,
-0.0761910453,
-0.0198102482,
0.0539846011,
-0.0428384319,
-0.0375207588,
-0.0119965794,
0.0153713441,
0.04906746,
-0.0375053585,
-0.0034609891,
2.414746413e-34,
0.0550825298,
0.0073240078,
0.0164557621,
0.0291380286,
0.0470697768,
-0.0274019782,
-0.0135037741,
-0.0010705304,
0.0020400751,
-0.0718351603,
-0.0086140111
] |
http://arxiv.org/abs/2401.13527v2 | http://arxiv.org/pdf/2401.13527v2 | SpeechGPT-Gen: Scaling Chain-of-Information Speech Generation | 2024-01-24T15:25:01 | 2024-01-25T17:24:52 | Dong Zhang, Xin Zhang, Jun Zhan, Shimin Li, Yaqian Zhou, Xipeng Qiu | cs.CL | cs.CL, cs.SD, eess.AS | Benefiting from effective speech modeling, current Speech Large Language
Models (SLLMs) have demonstrated exceptional capabilities in in-context speech
generation and efficient generalization to unseen speakers. However, the
prevailing information modeling process is encumbered by certain redundancies,
leading to inefficiencies in speech generation. We propose Chain-of-Information
Generation (CoIG), a method for decoupling semantic and perceptual information
in large-scale speech generation. Building on this, we develop SpeechGPT-Gen,
an 8-billion-parameter SLLM efficient in semantic and perceptual information
modeling. It comprises an autoregressive model based on LLM for semantic
information modeling and a non-autoregressive model employing flow matching for
perceptual information modeling. Additionally, we introduce the novel approach
of infusing semantic information into the prior distribution to enhance the
efficiency of flow matching. Extensive experimental results demonstrate that
SpeechGPT-Gen markedly excels in zero-shot text-to-speech, zero-shot voice
conversion, and speech-to-speech dialogue, underscoring CoIG's remarkable
proficiency in capturing and modeling speech's semantic and perceptual
dimensions. Code and models are available at
https://github.com/0nutation/SpeechGPT. | SpeechGPT-Gen: Scaling Chain-of-Information Speech Generation: Benefiting from effective speech modeling, current Speech Large Language
Models (SLLMs) have demonstrated exceptional capabilities in in-context speech
generation and efficient generalization to unseen speakers. However, the
prevailing information modeling process is encumbered by certain redundancies,
leading to inefficiencies in speech generation. We propose Chain-of-Information
Generation (CoIG), a method for decoupling semantic and perceptual information
in large-scale speech generation. Building on this, we develop SpeechGPT-Gen,
an 8-billion-parameter SLLM efficient in semantic and perceptual information
modeling. It comprises an autoregressive model based on LLM for semantic
information modeling and a non-autoregressive model employing flow matching for
perceptual information modeling. Additionally, we introduce the novel approach
of infusing semantic information into the prior distribution to enhance the
efficiency of flow matching. Extensive experimental results demonstrate that
SpeechGPT-Gen markedly excels in zero-shot text-to-speech, zero-shot voice
conversion, and speech-to-speech dialogue, underscoring CoIG's remarkable
proficiency in capturing and modeling speech's semantic and perceptual
dimensions. Code and models are available at
https://github.com/0nutation/SpeechGPT. | [
-0.0316239037,
0.0604908168,
0.0243134834,
0.0243209079,
-0.0236852393,
-0.0056169145,
0.0041450863,
0.0075112083,
-0.0588303991,
-0.0484619513,
-0.0419954285,
-0.0481020473,
0.00787473,
0.0310826357,
0.0329599604,
-0.0546428822,
-0.0074847341,
-0.0539154857,
-0.0275727343,
-0.0120512182,
0.0068579572,
0.0521538705,
-0.0250722058,
0.0238197129,
-0.0469087027,
-0.0180242974,
0.0004358342,
0.0415181294,
0.0192112699,
-0.0297104884,
-0.0360557698,
0.0057306006,
0.0591896586,
0.0586807095,
0.0000021481,
-0.0443793498,
0.0188625641,
0.0077972119,
-0.0341658928,
0.033684358,
0.0072740554,
-0.0014248886,
-0.0410339981,
0.0354293287,
0.0004986576,
-0.0257418565,
0.050903827,
0.022992773,
0.0212978963,
0.0726560652,
-0.0149883395,
-0.0700604096,
0.0248629581,
-0.0355886221,
0.0351622067,
-0.0202472825,
-0.011585624,
0.0000803998,
0.0003541808,
-0.0026008876,
-0.0026919048,
0.033295732,
0.0127698863,
-0.0196259879,
0.0103201913,
0.0469644926,
0.0259804074,
-0.0200385265,
-0.018312281,
0.0893066302,
-0.0176875349,
-0.0245968252,
0.029265834,
-0.0065562781,
-0.0353348106,
-0.0419128835,
-0.0308042932,
0.0437773801,
0.0251053497,
0.0187764876,
0.0066020149,
0.0277120005,
0.0214711986,
-0.0210806988,
-0.0244428683,
0.020150438,
0.0339078084,
0.0030344075,
-0.043132145,
-0.0344461091,
-0.0371070616,
-0.0398453735,
-0.0072022043,
0.034040872,
-0.0174724776,
-0.0036909576,
-0.0732501298,
-0.0031595542,
-0.0180938412,
0.010939735,
0.0307145491,
0.038046401,
0.0043141418,
0.0034810582,
-0.0104607064,
0.0604032092,
-0.0339375138,
-0.0416624583,
-0.0303964,
0.0088421823,
-0.0058578625,
-0.0027777872,
-0.0426795632,
0.0432158001,
-0.0754210874,
-0.0463771261,
-0.0202047601,
0.0275816191,
-0.0031205534,
-0.0236555692,
0.0097180232,
0.0095962677,
-0.0417550653,
-0.0281353295,
0.0288684312,
0.0172972456,
0.0019892,
0.0118719544,
0.024226645,
-0.0296522491,
-0.0097619249,
0.0510652587,
0.0320065618,
0.0070822146,
0.0103373025,
0.0773865059,
-0.0191880129,
0.021961987,
-0.0481762551,
-0.0475230105,
-0.0038302864,
-0.0813277438,
-0.0223794207,
-0.0506324954,
0.0200192891,
-0.0151017895,
-0.0207517538,
-0.0150296586,
-0.0004516744,
-0.0012327159,
-0.0219286513,
0.0100929793,
-0.0224942882,
0.0442105979,
0.0422791913,
0.0113035981,
0.0580216944,
0.040886879,
0.0677538291,
-0.0029719546,
-0.0292749181,
-0.0295040105,
-0.0142883165,
0.0510238633,
0.0268745348,
0.0404426157,
0.0028373869,
-0.024689246,
0.008090023,
0.0485260859,
0.0962874964,
0.0768494159,
-0.049905546,
0.0145536894,
0.0853785798,
0.0282433331,
0.051902622,
0.0130375167,
0.0874453411,
0.055753544,
-0.0276313461,
-0.0129425135,
-0.0144711435,
0.0255775135,
-0.0011441456,
0.0216289163,
0.0505463667,
0.0700773597,
-0.0587459058,
0.0239547305,
0.0418345481,
-0.0224852636,
-0.0343560167,
-0.0006503934,
0.0034769657,
0.0318840817,
-0.0607710071,
0.049770195,
0.016716361,
-0.0740777627,
0.0328223743,
-0.0271347202,
0.024416035,
0.0550052412,
0.0346187465,
0.0297574718,
0.0036567945,
-0.0352960601,
-0.0122228116,
0.0815498605,
-0.0286223814,
0.0036129227,
0.0275078136,
-0.0740009174,
-0.0015950555,
0.024343932,
0.0063412301,
0.0497898273,
-0.0406126045,
0.0243999753,
0.066637449,
0.0210565627,
0.0218949299,
-0.0087702964,
-0.055535879,
-0.0238769986,
0.0381867923,
-0.0254982337,
0.0442964137,
0.0140070682,
-0.0432797521,
0.0237929039,
-0.0093318848,
-0.0349339806,
0.0008305896,
0.0243367869,
-0.0755172148,
-0.009905057,
0.0088625671,
-0.0127022499,
0.0575894639,
-0.0172516033,
0.0057830396,
-0.1000189707,
0.042423524,
-0.0092879264,
0.0501048341,
-0.037801601,
0.0430014059,
0.0710536912,
0.0178522058,
-0.0109962048,
0.0125907091,
0.0058478634,
-0.0420589745,
0.0501963608,
0.0941944718,
-0.0138073191,
-0.0645028055,
0.0181369986,
0.0281857569,
-0.0031209046,
0.0162051357,
0.0038077417,
0.0292057618,
-0.0116159515,
-0.0241985545,
-0.0620193817,
0.0176476613,
0.0106042093,
0.0132926144,
0.034228269,
0.036030788,
0.0012054207,
0.0382521339,
-0.0654794276,
0.0257431231,
0.0398272164,
-0.0253500193,
-0.0329119042,
-0.0101995533,
-0.0257254764,
-0.0097730672,
0.0293494891,
0.0294735283,
0.0195794404,
-0.0005912716,
-0.0086785415,
-0.042549789,
0.0580164902,
-0.022323858,
0.0251241084,
-0.004562878,
-0.0375999361,
-0.0084501002,
-0.000527276,
0.0665912703,
-0.0309280027,
0.0095930649,
-0.0424986929,
0.007626316,
-0.0484169684,
-0.0574367605,
0.0190354101,
0.0061217868,
0.0016069699,
-0.0173449628,
0.006069168,
0.0068348809,
0.0441002995,
0.0554386973,
0.0228311867,
-0.0604679473,
-0.0183656495,
0.0021253156,
-0.040361017,
0.0062752068,
-0.0234007295,
-0.020010652,
0.0152258417,
-0.020397475,
-0.0327372439,
-0.0168779716,
-0.0862751231,
-0.0251596533,
-0.0023641463,
-0.007890149,
-0.0384656191,
0.0323640443,
0.0089854486,
0.0358250439,
-0.0117117567,
-0.0064790412,
-0.0241493396,
0.0141175827,
-0.0056299185,
-0.0467036553,
0.0169167183,
-0.0408261269,
0.0183162261,
-0.0085361497,
-0.0335377827,
-0.0320289768,
-0.0223280583,
-0.0017943201,
0.0084800208,
0.0138167636,
-0.0668476,
-0.0081923353,
0.0392151698,
0.0212077312,
0.0374594741,
0.0094603952,
0.0044812635,
-0.0365617573,
0.0629920661,
0.0699348897,
-0.0064395354,
-0.0019876051,
0.0245532636,
0.0442490801,
0.035059642,
0.0465558879,
0.0676039308,
-0.0354010798,
0.0066966913,
0.0252325069,
0.01069461,
-0.0332300402,
0.0449076854,
0.0419207253,
-0.0317921788,
0.0209141169,
0.1113984436,
-0.0227133688,
-0.0483059026,
0.0166430809,
0.0373956636,
0.0866840854,
-0.017986808,
-0.0079550557,
-0.0945106298,
-0.0359847397,
0.0200406536,
0.0010571418,
-0.0212689564,
-0.0060120723,
-0.048338037,
-0.0985877514,
0.0598650463,
0.0727573857,
-0.0142379487,
-0.0244535971,
-0.0066916198,
-0.0158969164,
0.0101827886,
0.02665966,
0.0326692797,
-0.0243470781,
0.0082896631,
0.0266062692,
-0.0261370298,
-0.0355301201,
0.0001147848,
-0.0266593602,
0.0873589218,
0.0691399947,
0.0210063178,
0.0205348674,
-0.000038593,
0.0328196101,
-0.0131756887,
-0.01466709,
-0.0111626275,
0.0155603308,
-0.0079259351,
0.0007658078,
-0.0068474193,
0.080072619,
-0.0283242073,
0.0107691996,
0.0140770227,
0.0288489778,
-0.0161923729,
-0.0162539072,
-0.0044707227,
0.0389505401,
0.0160046611,
0.0775445327,
0.0296935812,
-0.0329253413,
-0.0238957759,
-0.0127045754,
0.0029114909,
-0.0300266314,
-0.0328937471,
0.0139523773,
0.0205501746,
-0.08460778,
-0.0600571521,
0.0076242927,
-0.0065176664,
0.0643104464,
-0.0038650238,
-0.0102628497,
-0.0020838741,
-0.0205383878,
-0.0304727033,
-0.0436965823,
-0.0519448593,
0.0150372246,
-0.0830880105,
-0.0175676644,
0.0452124178,
-0.0537266284,
0.0111934273,
0.0558248535,
-0.0958854854,
-0.0789859369,
-0.029435534,
-0.0136577049,
0.0242263395,
0.0657979026,
-0.0101906713,
-0.0268262923,
0.0841662437,
0.0214562658,
-0.0368452854,
0.0528267734,
-0.0031103941,
0.0227062274,
0.0022730369,
0.0316461548,
-0.0319583528,
0.0380839445,
-0.025660364,
0.0368596204,
-0.0397652797,
-0.0446376465,
0.0148605378,
0.0514533333,
-0.0323526077,
-0.0371212326,
0.0453608781,
-0.0179196056,
-0.0237521213,
-0.0331683345,
0.018081665,
-0.0760370865,
-0.0052727326,
0.0584439561,
0.0087229209,
0.0066917744,
-0.0356980972,
0.0075220102,
0.0072570378,
-0.0022934822,
-0.0451922491,
-0.0326672383,
0.015696194,
0.0169325545,
-0.0170257017,
-0.0003395208,
0.0182329807,
-0.0176896136,
-0.0059298035,
-0.014364697,
0.0536029413,
-0.0059033777,
-0.0031512568,
0.0198000092,
0.0063219341,
-0.0329540819,
-0.0215091556,
-0.0374353006,
0.0274500847,
0.073691465,
0.0452533327,
0.1129471436,
-0.0421780981,
0.0278316829,
-0.0487401336,
0.0662268102,
-0.0394898504,
-0.017215265,
0.0248040296,
-0.0131265,
-0.0380814262,
-0.0062481919,
0.0684233755,
0.0655292049,
0.0230761804,
-0.0128855472,
0.0494289324,
-0.0186381675,
-0.0422699563,
-0.0360300392,
-0.0049054031,
0.0040840986,
0.0097855832,
-0.0301519986,
-0.0482701957,
-0.0112794954,
-0.0270691179,
0.0162086692,
-0.0353167579,
0.061708346,
-0.005777244,
-0.0536589473,
0.006293661,
-0.0241127834,
-0.042701818,
-0.016902918,
-0.0343957804,
-0.01280268,
-6.571592352e-33,
-0.0412019715,
0.0005203396,
0.0515426844,
0.0269064344,
-0.063893728,
-0.0078511229,
-0.0241473205,
-0.0067274286,
0.008244467,
-0.002658129,
0.0021267352,
-0.0015441002,
0.0011570905,
0.0523552224,
-0.0315700062,
0.0123413363,
0.082030125,
-0.0272084288,
0.009096534,
0.0895249322,
0.0393575355,
0.0148882633,
-0.0045960159,
-0.0587030649,
-0.0751879066,
-0.0408793241,
-0.032363154,
0.000718052,
0.0294419192,
0.0008289642,
-0.031003736,
0.0172288325,
-0.0094221532,
-0.0190588012,
-0.0340225399,
-0.0060088267,
-0.0393545441,
0.0013624652,
0.0121399499,
-0.0000659899,
-0.0184596945,
-0.0413145237,
0.0040150066,
-0.0040410487,
-0.0289826188,
-0.0403650105,
-0.0108127743,
-0.005803667,
-0.0159861967,
-0.0564353131,
-0.0369888619,
0.0360690467,
-0.0228719376,
-0.0339441448,
0.0181818698,
0.03789635,
-0.0346603543,
-0.0284646954,
-0.0273834448,
0.0545061082,
0.0001934302,
0.0473846123,
0.0558597855,
0.0531988367,
0.012698493,
0.0354419313,
0.0288426317,
-0.048105929,
-0.0160585362,
0.0355930962,
-0.0041578831,
0.0457062088,
-0.009007046,
0.0476413593,
-0.0048466451,
-0.012558721,
-0.045019798,
0.0079167718,
-0.0029801396,
0.0040164883,
0.0174485818,
0.0279286616,
-0.0050694379,
-0.0213352442,
-0.0431415774,
-0.0475009866,
-0.0036568693,
-0.0135874273,
0.0470652618,
0.0138125531,
0.0132793859,
-0.0344717056,
0.0781033859,
-0.0009468748,
0.0106858332,
0.0278819278,
-0.0655261353,
0.0090427175,
-0.0293179266,
0.0249435343,
-0.0505403168,
-0.0556326248,
0.0682315826,
0.0360768102,
-0.0050411485,
0.0233836416,
-0.0444078706,
0.0423964262,
-0.0278740283,
-0.0087304618,
-0.0305054206,
0.0494014435,
0.0039577004,
-0.0304360166,
0.0089084003,
-0.0463984273,
-0.0053621195,
0.0437605828,
-0.0179683864,
-0.0024143076,
0.024809191,
-0.0158288684,
0.0010779933,
0.0307690334,
-0.0586632565,
0.0449209996,
-0.0293297693,
-0.0322128683,
0.0814799741,
-0.0387706943,
-0.0416327752,
0.0095729632,
2.938e-7,
-0.0562965646,
-0.0248554926,
0.0079774559,
0.0097469622,
0.0464455932,
0.0269635525,
0.0244276989,
0.035845425,
0.0162091795,
-0.0019339633,
0.0170717351,
0.022415733,
0.0855628327,
-0.0271945558,
-0.0860779062,
-0.02254791,
-0.0721059367,
-0.0099193696,
-0.0328044295,
-0.0229331702,
-0.0279322024,
0.049590338,
-0.0315132439,
-0.0329557396,
0.0688979477,
-0.0298047531,
0.0631426871,
-0.0153363626,
-0.0183188673,
0.0110892365,
-0.0015709213,
-0.0110566569,
-0.0170024354,
-0.0149284182,
-0.0241548158,
0.0081788348,
0.0023643684,
0.0732575506,
0.0082271602,
-0.0534117259,
0.0084205149,
0.0191502813,
-0.0153032355,
-0.02292585,
0.0236213338,
0.000839901,
-0.013172905,
-0.0033842523,
-0.0543662012,
0.0029578644,
-0.026778141,
-0.0279095061,
-0.0108353496,
0.0022249517,
0.0255960748,
-0.0102441339,
-0.0151552465,
-0.0237042215,
-0.0167323239,
-0.005044424,
-0.0857503265,
-0.035638161,
-0.0562291369,
-0.0106408112,
0.039708823,
0.0023831534,
-0.0363436192,
2.59515449e-34,
0.0146485334,
-0.0258466881,
0.046836216,
-0.0402866565,
0.028818896,
-0.0029935879,
0.0090011824,
0.0224217381,
0.0100640198,
-0.067393586,
-0.0439451262
] |
http://arxiv.org/abs/2401.13512v1 | http://arxiv.org/pdf/2401.13512v1 | Can GPT-3.5 Generate and Code Discharge Summaries? | 2024-01-24T15:10:13 | 2024-01-24T15:10:13 | Matúš Falis, Aryo Pradipta Gema, Hang Dong, Luke Daines, Siddharth Basetti, Michael Holder, Rose S Penfold, Alexandra Birch, Beatrice Alex | cs.CL | cs.CL | Objective: To investigate GPT-3.5 in generating and coding medical documents
with ICD-10 codes for data augmentation on low-resources labels.
Materials and Methods: Employing GPT-3.5 we generated and coded 9,606
discharge summaries based on lists of ICD-10 code descriptions of patients with
infrequent (generation) codes within the MIMIC-IV dataset. Combined with the
baseline training set, this formed an augmented training set. Neural coding
models were trained on baseline and augmented data and evaluated on a MIMIC-IV
test set. We report micro- and macro-F1 scores on the full codeset, generation
codes, and their families. Weak Hierarchical Confusion Matrices were employed
to determine within-family and outside-of-family coding errors in the latter
codesets. The coding performance of GPT-3.5 was evaluated both on prompt-guided
self-generated data and real MIMIC-IV data. Clinical professionals evaluated
the clinical acceptability of the generated documents.
Results: Augmentation slightly hinders the overall performance of the models
but improves performance for the generation candidate codes and their families,
including one unseen in the baseline training data. Augmented models display
lower out-of-family error rates. GPT-3.5 can identify ICD-10 codes by the
prompted descriptions, but performs poorly on real data. Evaluators note the
correctness of generated concepts while suffering in variety, supporting
information, and narrative.
Discussion and Conclusion: GPT-3.5 alone is unsuitable for ICD-10 coding.
Augmentation positively affects generation code families but mainly benefits
codes with existing examples. Augmentation reduces out-of-family errors.
Discharge summaries generated by GPT-3.5 state prompted concepts correctly but
lack variety, and authenticity in narratives. They are unsuitable for clinical
practice. | Can GPT-3.5 Generate and Code Discharge Summaries?: Objective: To investigate GPT-3.5 in generating and coding medical documents
with ICD-10 codes for data augmentation on low-resources labels.
Materials and Methods: Employing GPT-3.5 we generated and coded 9,606
discharge summaries based on lists of ICD-10 code descriptions of patients with
infrequent (generation) codes within the MIMIC-IV dataset. Combined with the
baseline training set, this formed an augmented training set. Neural coding
models were trained on baseline and augmented data and evaluated on a MIMIC-IV
test set. We report micro- and macro-F1 scores on the full codeset, generation
codes, and their families. Weak Hierarchical Confusion Matrices were employed
to determine within-family and outside-of-family coding errors in the latter
codesets. The coding performance of GPT-3.5 was evaluated both on prompt-guided
self-generated data and real MIMIC-IV data. Clinical professionals evaluated
the clinical acceptability of the generated documents.
Results: Augmentation slightly hinders the overall performance of the models
but improves performance for the generation candidate codes and their families,
including one unseen in the baseline training data. Augmented models display
lower out-of-family error rates. GPT-3.5 can identify ICD-10 codes by the
prompted descriptions, but performs poorly on real data. Evaluators note the
correctness of generated concepts while suffering in variety, supporting
information, and narrative.
Discussion and Conclusion: GPT-3.5 alone is unsuitable for ICD-10 coding.
Augmentation positively affects generation code families but mainly benefits
codes with existing examples. Augmentation reduces out-of-family errors.
Discharge summaries generated by GPT-3.5 state prompted concepts correctly but
lack variety, and authenticity in narratives. They are unsuitable for clinical
practice. | [
0.0286790263,
0.0412928201,
0.0106951715,
-0.0446967669,
0.000630669,
0.0594922118,
0.0077954577,
0.0397536829,
-0.0254404414,
0.0123893274,
-0.0014848255,
-0.0636490062,
0.0449512489,
0.0650292262,
0.0134520708,
-0.0346622989,
0.048071228,
0.0482024066,
0.0289865639,
0.0208151471,
-0.0267320387,
0.0149023654,
-0.020299565,
0.048659496,
-0.0350266024,
-0.0362284705,
0.0068874019,
-0.0015020174,
0.0236020945,
-0.0971695036,
0.0042480351,
0.051692836,
0.0090890778,
0.0067167338,
0.0000022518,
-0.0454230346,
0.0010462662,
0.0224160124,
-0.0616303049,
-0.0268215295,
0.0082740886,
-0.0788532719,
0.0225382559,
0.0089938557,
-0.0082219457,
-0.0809327438,
0.0615554601,
0.0952697173,
0.0315200426,
0.0593322366,
-0.0201641843,
0.0062842323,
0.0327535532,
-0.0231640618,
0.0443362892,
0.0063616522,
0.0082734274,
0.051982332,
0.0180947389,
-0.0317928009,
0.0165253505,
0.0176660623,
0.0501133911,
0.0471925065,
0.0562152751,
0.1000888869,
-0.0038084271,
0.0110545792,
-0.0030778868,
0.0257093217,
0.0504011363,
0.0046825274,
-0.0153621938,
-0.0291565321,
0.015868498,
-0.0780149102,
-0.0539811514,
-0.0134495972,
-0.0428616777,
0.0257606953,
0.0027051773,
0.0305008031,
0.0091673443,
-0.0448219217,
0.0057363748,
-0.0406764522,
-0.0098638656,
-0.0384405702,
-0.0131918192,
-0.0069008428,
-0.0015708801,
-0.0388264544,
0.0129158068,
0.0122124599,
-0.0002737901,
0.0052081039,
-0.0553288721,
-0.036538925,
0.065178521,
-0.0064400532,
0.01690384,
0.0427072123,
-0.0147563545,
-0.018390445,
-0.0329419598,
0.0568037257,
-0.003431632,
0.0235038139,
-0.055105377,
0.0300078373,
-0.0631699637,
-0.0179316457,
-0.0516558252,
0.0527750663,
0.024219662,
0.0045245807,
-0.1022442505,
0.0035679403,
0.0294171609,
0.0141694229,
0.0234870221,
0.0420867167,
-0.05525003,
0.0461728051,
-0.008384319,
-0.0073948912,
-0.0132568209,
-0.0130642373,
0.0186762791,
-0.0622617416,
0.0071644587,
-0.0045436011,
-0.0157891978,
0.0083891591,
-0.0104487194,
0.032575205,
-0.0243165325,
-0.0353214554,
-0.04882209,
-0.0345106162,
-0.0389398448,
-0.0064416192,
0.028869275,
-0.0419437848,
0.0257965345,
0.0164336357,
0.0277107004,
-0.0365773328,
0.003719897,
-0.0070893737,
-0.0156510118,
0.0215905383,
-0.1058554128,
0.0090375608,
0.0090832608,
0.0198363885,
0.0100512113,
0.0183719248,
0.0387972333,
-0.0330965482,
-0.0243560243,
-0.0386665799,
0.0205011237,
-0.0288873464,
0.0156961754,
-0.0223165154,
0.0076698144,
-0.0105188889,
-0.010845622,
0.0459684469,
0.0151504874,
0.0308757685,
-0.025257973,
-0.0112962872,
0.0517896414,
0.0219799541,
0.031620048,
0.0135290734,
0.0151573708,
-0.0453024134,
0.0049221111,
0.0025498911,
0.0086930618,
0.0132189216,
0.0486549772,
0.025282355,
0.0309924707,
-0.0321167409,
-0.0315791182,
-0.0652588308,
-0.0324792564,
-0.006916835,
-0.0113409152,
-0.0127495881,
-0.003567751,
-0.0130668879,
-0.0320051685,
0.0200337339,
0.0034458258,
-0.0422499664,
-0.0151245575,
-0.0484440327,
-0.022185633,
0.0968883187,
0.0491397269,
0.0208798703,
-0.0505295843,
-0.046686098,
-0.0166881979,
0.024090847,
0.0282817222,
0.0075112637,
0.0179128777,
-0.0584928691,
-0.0257949401,
0.010438947,
0.0075544971,
-0.0017754073,
-0.0116923833,
-0.0213698521,
-0.0106064677,
-0.0060807969,
-0.0657580867,
0.0352243111,
-0.045777902,
-0.0548942462,
0.0125338165,
-0.0492665656,
0.0470519215,
0.0204220973,
-0.0062127346,
0.0321123041,
0.0266759861,
0.0105284182,
-0.022736514,
-0.0204434227,
-0.0797981843,
-0.0179192908,
0.0021144662,
-0.0054172524,
-0.0010929429,
-0.0492929146,
-0.014054588,
-0.0273510832,
-0.0535099581,
-0.0929423943,
0.0410202593,
-0.0391775966,
0.0124075366,
0.0542860739,
-0.0041460707,
-0.0090674357,
0.0447424091,
-0.0363682061,
0.0026193662,
0.0066858502,
0.0337068476,
0.0426097736,
-0.0240955222,
0.0183825195,
0.0093408842,
-0.0493136384,
-0.0030052939,
-0.0138246352,
0.0284951385,
-0.0236300901,
0.0120397955,
-0.1330569834,
-0.0121020023,
0.0221403018,
-0.0216496028,
0.0354153104,
-0.0066095102,
-0.0169843622,
-0.0044004265,
-0.0348920822,
-0.0166073404,
0.0352630243,
0.0350341052,
0.0124045638,
0.0556403622,
-0.0003030612,
0.0107232239,
0.0041731456,
-0.0281649288,
0.0062534022,
0.0296811368,
-0.0370970033,
-0.0385405272,
0.0224514026,
-0.0268463176,
-0.0007331325,
-0.0028234925,
-0.0244169608,
0.0530568138,
0.0194287188,
0.0724828392,
-0.0356222019,
0.0133555233,
0.0256144814,
0.0298402291,
0.0028440799,
-0.0222769994,
0.0096632689,
-0.0229202174,
-0.0461274981,
0.0237684846,
-0.0692829341,
0.0106053678,
0.0631789044,
0.0657378063,
0.0120797418,
-0.0736792535,
-0.0297127534,
0.0171187464,
0.0203860886,
-0.0272876061,
-0.044987306,
-0.0360090211,
0.0516231246,
-0.0049729529,
-0.0664249286,
0.0766704306,
-0.0212322213,
0.013793475,
-0.0706659332,
-0.0293755308,
-0.0168679748,
0.0201989338,
-0.0140011003,
0.0125944838,
0.0838601738,
-0.0330915302,
0.0189112555,
0.0247479603,
-0.0335836336,
-0.0397902504,
0.0611265078,
0.0031606131,
0.0025616765,
-0.0007127125,
-0.0388829932,
-0.0472216569,
0.0114826607,
0.0438977778,
0.0081070475,
-0.0149852773,
-0.0024074288,
-0.0280904993,
0.0235587209,
-0.0064877621,
0.0880818665,
-0.0271486975,
-0.0327485837,
-0.0017049628,
0.0068234932,
-0.0073396135,
-0.019620277,
0.0022445018,
-0.0242725518,
-0.0181677081,
0.0294857975,
-0.0067353374,
0.0495564789,
-0.0145007987,
0.021203775,
-0.0202450212,
-0.0000039458,
-0.0119269621,
0.0308867432,
0.0210945308,
-0.0381912477,
-0.007914573,
0.1031788141,
0.021752039,
-0.0925917178,
0.0167206377,
-0.007180864,
-0.0045041316,
-0.0018159245,
-0.0150102768,
-0.049924098,
0.0226396602,
-0.0082996711,
0.0699380934,
-0.0537270233,
-0.0152758574,
0.0073034731,
-0.0547000729,
0.0543390512,
0.0298928246,
-0.0574508347,
-0.0990939513,
-0.0131342327,
-0.016384745,
0.0515684374,
-0.0249845982,
0.0046734279,
0.0012456521,
-0.0004334971,
0.0332826674,
-0.0494089574,
-0.0774128884,
-0.0009531391,
0.0104437387,
0.062198896,
0.0812913328,
0.0217175316,
0.0031663668,
-0.0507873222,
0.0010159125,
-0.0417278223,
0.0029398873,
0.0042489595,
0.0801912695,
0.0067961682,
0.0160997566,
-0.029385915,
0.0398906395,
0.0257855859,
-0.0256151222,
-0.0298847184,
0.0606451854,
-0.0650608987,
-0.0215478223,
-0.0364043042,
0.0247542784,
0.0088427765,
0.0484040454,
0.0375548638,
0.0196643304,
0.0280684736,
0.0110316221,
0.0959377959,
-0.0302266404,
0.0688442141,
0.046722956,
-0.0457721278,
-0.052100528,
-0.0620185137,
-0.0694503114,
0.02354878,
-0.0041557048,
0.0163961984,
0.0023536566,
-0.0381985717,
-0.0087238429,
-0.0586125478,
0.0462527201,
0.0264088586,
-0.044358328,
-0.0417247266,
-0.0613865219,
0.0041400641,
-0.0103464872,
-0.0133065544,
0.0267557781,
0.0208208896,
0.0129128629,
0.006528398,
0.0285622459,
0.0243700985,
0.0120996181,
0.0134620508,
0.0097823553,
0.0097855125,
0.0444220006,
0.0061808974,
-0.0172471907,
-0.0407842435,
-0.013521025,
-0.0087703448,
-0.0233703498,
-0.0283451341,
0.0509122014,
-0.014041137,
0.0130449422,
0.0182675403,
0.017040953,
0.0616585314,
0.0289427228,
0.0075695496,
-0.0396539308,
-0.0269968528,
-0.0169840865,
0.0095171267,
-0.0297310594,
0.0200643297,
-0.0118833929,
-0.0351859331,
0.0388700701,
0.0053203409,
-0.0340973102,
-0.0104495063,
-0.0511232056,
-0.0100425249,
-0.0471082292,
-0.0250899382,
-0.0318278223,
0.0324818045,
-0.0096983956,
-0.020268051,
-0.0474439859,
-0.0085419649,
-0.0248229057,
-0.0018435683,
-0.0160295852,
0.0802287832,
0.0127614327,
0.0290933382,
0.0234094802,
0.0415630005,
-0.0275826827,
-0.0374708809,
-0.0711324662,
-0.0038828973,
-0.0010453016,
0.0288002528,
-0.0055847499,
-0.0738815069,
0.0370096192,
-0.0322362706,
0.00165406,
-0.0209744107,
0.024734769,
0.0075443094,
0.0186522547,
-0.0059160125,
-0.0607038811,
0.0590837076,
0.0094393324,
0.0573498085,
-0.0476737916,
-0.0504870526,
-0.0428059883,
0.015064639,
0.0383159146,
0.017451359,
-0.049094446,
-0.0096075265,
-0.0682398453,
-0.012694695,
0.0614588223,
0.0116870096,
0.009776853,
-0.0186106656,
0.07747785,
-0.0017681906,
-0.0483840518,
-0.0524797849,
0.0147158271,
0.0321995206,
0.012250334,
-0.0106016099,
-0.0300314762,
-6.182497844e-33,
-0.0397880338,
-0.0235085618,
0.0457289964,
-0.0179271698,
-0.0630962476,
0.0121220313,
-0.0530689172,
-0.0210524425,
0.0342769995,
0.0075171925,
-0.0044899122,
0.003388203,
0.0131902015,
0.0405436344,
0.0142254801,
-0.0226174444,
0.0508153811,
-0.0101559218,
-0.0266508181,
0.0474000201,
0.0017901219,
0.0212613605,
0.0679280311,
-0.0991471112,
0.0305882189,
0.0070424587,
0.0155189587,
-0.0217691883,
0.0005346993,
0.0381659716,
-0.0290077645,
0.0152426353,
0.0609826781,
-0.0010086381,
-0.0070817694,
-0.0152291665,
-0.063204281,
-0.0303471927,
0.0264240392,
-0.062468566,
-0.0307509005,
-0.0484857708,
0.0591107272,
-0.0065193637,
-0.0458812676,
0.0059975428,
0.0033172229,
-0.0364195518,
0.0110536106,
-0.0344810337,
-0.0120700896,
0.0226338133,
0.0217710342,
0.0168557316,
-0.05181152,
0.0471365042,
-0.0438279472,
0.0636355728,
-0.04169048,
-0.0027097233,
0.0531038456,
0.0316163115,
-0.0139256762,
-0.0300036781,
0.0118910233,
0.0488286763,
0.0464227572,
0.0144356582,
-0.0464010984,
0.0200993735,
-0.0177973285,
-0.0117145833,
0.0845303684,
0.027189197,
0.0758632869,
-0.0434970036,
-0.024166204,
0.0444263965,
-0.0852123275,
0.0200650506,
0.0516117513,
-0.0204087347,
-0.0517548174,
-0.0222678948,
-0.0002211563,
-0.0119286915,
0.0044906237,
-0.054988984,
0.0567273907,
0.008984575,
0.0168702267,
0.017995147,
0.0188068748,
0.0178785659,
0.0349102356,
-0.0304107033,
0.0531990938,
0.0581015572,
-0.0229556859,
-0.0277225208,
-0.0146958949,
0.0142156864,
0.0806781352,
-0.0000203683,
0.0501683056,
0.0186796617,
0.0511218682,
-0.0057862597,
-0.0477410965,
-0.0101224547,
0.0237991791,
0.0271822661,
-0.0107823228,
-0.0067130895,
0.0131104887,
0.0219607484,
-0.0035112861,
0.0111896293,
-0.0058080498,
-0.0305626541,
-0.041311346,
-0.052917961,
0.0171318669,
0.0150388321,
0.0149344848,
0.0187096093,
-0.0437884219,
0.0507531911,
0.0211629607,
-0.0129302889,
-0.0002204774,
0.0533564575,
2.94e-7,
0.0506976694,
0.0224245284,
0.052012179,
-0.0392274074,
0.0319716483,
-0.0148967132,
-0.005912994,
0.0015773944,
0.0234237667,
-0.0227845814,
0.0597981438,
0.0043022595,
0.0366704017,
0.0447779074,
-0.0445756875,
-0.034220092,
-0.0536293946,
-0.0169728342,
-0.0329104923,
0.0174518488,
0.0789463744,
0.0514052883,
-0.0307760909,
-0.0220635962,
0.0527864061,
0.0045137983,
0.0274083335,
-0.0082972944,
0.0002697614,
-0.0017686401,
-0.000559967,
0.0250151288,
0.0235494524,
0.0710668489,
-0.0102474261,
-0.0379671976,
-0.0116516035,
0.0528703928,
-0.0487824604,
0.0159861315,
0.0219087079,
0.0339640565,
0.0156790819,
-0.0485457629,
0.0244893525,
-0.0352595747,
-0.0046046833,
0.0309635662,
-0.040964853,
0.0097974446,
0.0613991395,
-0.047501009,
-0.0473115295,
0.0344201103,
0.0029489782,
0.0281949379,
-0.0378508456,
-0.033127781,
0.0164746977,
0.0472857431,
-0.0516497493,
-0.0214076061,
-0.0072767399,
0.0133976582,
0.0191165004,
-0.0588289611,
-0.0125071369,
2.983732515e-34,
0.0291730724,
-0.0348768532,
0.0181507207,
0.0045390329,
0.0272091329,
-0.0236226618,
0.0300881937,
0.0384798869,
0.0026769761,
-0.0499413684,
-0.0184830055
] |
http://arxiv.org/abs/2401.13481v1 | http://arxiv.org/pdf/2401.13481v1 | How AI Ideas Affect the Creativity, Diversity, and Evolution of Human Ideas: Evidence From a Large, Dynamic Experiment | 2024-01-24T14:29:39 | 2024-01-24T14:29:39 | Joshua Ashkinaze, Julia Mendelsohn, Li Qiwei, Ceren Budak, Eric Gilbert | cs.CY | cs.CY, cs.AI, cs.CL, cs.HC | Exposure to large language model output is rapidly increasing. How will
seeing AI-generated ideas affect human ideas? We conducted an experiment (800+
participants, 40+ countries) where participants viewed creative ideas that were
from ChatGPT or prior experimental participants and then brainstormed their own
idea. We varied the number of AI-generated examples (none, low, or high
exposure) and if the examples were labeled as 'AI' (disclosure). Our dynamic
experiment design -- ideas from prior participants in an experimental condition
are used as stimuli for future participants in the same experimental condition
-- mimics the interdependent process of cultural creation: creative ideas are
built upon prior ideas. Hence, we capture the compounding effects of having
LLMs 'in the culture loop'. We find that high AI exposure (but not low AI
exposure) did not affect the creativity of individual ideas but did increase
the average amount and rate of change of collective idea diversity. AI made
ideas different, not better. There were no main effects of disclosure. We also
found that self-reported creative people were less influenced by knowing an
idea was from AI, and that participants were more likely to knowingly adopt AI
ideas when the task was difficult. Our findings suggest that introducing AI
ideas into society may increase collective diversity but not individual
creativity. | How AI Ideas Affect the Creativity, Diversity, and Evolution of Human Ideas: Evidence From a Large, Dynamic Experiment: Exposure to large language model output is rapidly increasing. How will
seeing AI-generated ideas affect human ideas? We conducted an experiment (800+
participants, 40+ countries) where participants viewed creative ideas that were
from ChatGPT or prior experimental participants and then brainstormed their own
idea. We varied the number of AI-generated examples (none, low, or high
exposure) and if the examples were labeled as 'AI' (disclosure). Our dynamic
experiment design -- ideas from prior participants in an experimental condition
are used as stimuli for future participants in the same experimental condition
-- mimics the interdependent process of cultural creation: creative ideas are
built upon prior ideas. Hence, we capture the compounding effects of having
LLMs 'in the culture loop'. We find that high AI exposure (but not low AI
exposure) did not affect the creativity of individual ideas but did increase
the average amount and rate of change of collective idea diversity. AI made
ideas different, not better. There were no main effects of disclosure. We also
found that self-reported creative people were less influenced by knowing an
idea was from AI, and that participants were more likely to knowingly adopt AI
ideas when the task was difficult. Our findings suggest that introducing AI
ideas into society may increase collective diversity but not individual
creativity. | [
0.0248035211,
0.0250714682,
-0.0702726766,
-0.0134736886,
-0.0558927059,
-0.0155637963,
0.0429345369,
-0.0132730827,
0.0358786099,
0.0143409213,
-0.0247239899,
-0.0073066009,
0.0060780286,
0.0698391423,
0.035885442,
0.0186542291,
0.00571007,
-0.0331307352,
0.0437469371,
-0.0471082777,
-0.0347926691,
0.0233894493,
0.0271258559,
0.0576222241,
0.0166394766,
-0.0150093362,
0.0084890798,
-0.0320357196,
0.0126289017,
-0.0383622944,
-0.0353894569,
0.0517890863,
0.0197144262,
0.0114278197,
0.0000019485,
-0.0471069403,
0.0193277169,
-0.0321049429,
0.005972696,
-0.0156646241,
0.0290246624,
0.0802274048,
-0.0229701456,
-0.0026712166,
-0.0383635759,
0.0409402326,
-0.0011428533,
0.0057291156,
-0.0204097815,
0.0093515478,
0.0132470215,
-0.0200519282,
0.0253283512,
-0.0156399924,
0.0359658636,
0.0203002468,
0.0150289517,
-0.0104475031,
0.0008329431,
-0.0090547567,
-0.009781966,
0.007898109,
-0.0011418299,
-0.0055340971,
0.0137739433,
0.0358594991,
-0.0288590994,
-0.047343377,
0.0352191329,
0.0417763181,
0.0654073134,
-0.0073596472,
0.0186035577,
-0.0144456746,
0.0561660007,
-0.0049917689,
-0.0416527912,
0.054573115,
-0.0021881282,
-0.0100338589,
0.0474999212,
-0.0193510205,
-0.0029858937,
0.0235179309,
0.0170575418,
0.0603012852,
-0.0180941671,
-0.0032781796,
-0.0401183143,
-0.0141937183,
0.016474016,
-0.0545721762,
0.0008540298,
-0.017594019,
0.0237929337,
0.0166121759,
0.0048008165,
-0.012153008,
-0.001460841,
-0.0449779406,
0.0379050449,
0.0236090124,
0.0096851699,
0.0208626781,
0.0295314193,
0.0072130337,
-0.0387194678,
0.0453114845,
-0.0144830896,
0.0165401213,
-0.0845233053,
-0.0244788025,
-0.0097321291,
0.0315923393,
0.0428177416,
-0.009383739,
-0.0319835395,
0.0601926632,
-0.043764241,
-0.0229218416,
-0.0727796182,
-0.0061167991,
0.0073733926,
-0.0020739206,
-0.038510289,
0.023100445,
-0.0200436693,
0.0091561824,
0.0268503819,
-0.0204701833,
0.0069820047,
-0.0147235822,
0.0210631937,
0.0005697199,
0.0359883942,
0.0216256902,
0.0322580971,
-0.0553055033,
-0.010789765,
-0.0234749801,
0.0381167866,
-0.0155422799,
0.0420235731,
-0.0452795662,
-0.0138947945,
0.0362078398,
0.0014631764,
-0.0468125045,
-0.0205142647,
0.0094341394,
-0.0528436676,
-0.0081472304,
-0.0255575702,
0.0569615029,
0.0565189682,
0.0206509773,
-0.0407114774,
0.0357852466,
0.0737989396,
0.0062901401,
0.0135878846,
0.0149014201,
0.0544305854,
0.0305522885,
-0.0192053113,
0.0148939956,
0.0322381072,
-0.0130793555,
-0.0405286327,
-0.0011659159,
-0.0008014427,
0.0328216292,
-0.0072052339,
0.0972912908,
0.0666268095,
0.0806877613,
0.0018476513,
0.0529905967,
-0.0108028399,
-0.0000640449,
-0.0179834161,
-0.0057438076,
0.0070895832,
0.014905409,
-0.0107809259,
-0.0052309451,
0.0289917588,
0.0440292247,
-0.051514551,
0.0187950023,
0.0042292797,
0.0047000898,
0.0534571074,
0.0059610498,
0.0009066208,
-0.0035497642,
0.0379894264,
-0.0156136183,
-0.0116337156,
-0.0066436017,
-0.0306068584,
-0.0744891316,
0.0895131454,
0.0836009532,
0.0371092893,
0.0246164259,
0.0484282225,
-0.0304948147,
0.004670199,
0.0377744623,
0.0428316258,
0.021917887,
-0.0325106271,
0.0531098172,
0.0166715495,
-0.0278227385,
-0.0009438207,
0.0189723559,
-0.0170608982,
0.0099707376,
-0.0380224362,
0.0431603268,
-0.0085691558,
-0.0146301705,
-0.0679829419,
0.035600882,
-0.0244114753,
-0.0278996304,
0.0211999491,
0.0326243751,
-0.0708098039,
0.034356039,
0.0029973267,
-0.0420785956,
-0.0840052664,
0.0573846735,
-0.0075432048,
0.0198513921,
-0.0277411621,
-0.0675827265,
-0.012145156,
-0.0569407493,
0.0469434932,
0.0178814474,
0.0851987004,
0.0001062566,
0.0371320434,
0.020792393,
-0.0061466629,
-0.034863241,
-0.0001179593,
0.0023605626,
0.0297575518,
-0.0436862931,
-0.0246610418,
-0.0289734043,
-0.0237290934,
0.016426893,
0.000983604,
0.0084262332,
0.0296451617,
0.0133094471,
0.0101075675,
0.0221241619,
-0.0113640167,
-0.0380940661,
-0.0095037026,
-0.0786760896,
-0.0285733975,
0.0275710989,
-0.0265298542,
0.0089152418,
0.0095315576,
-0.0295121875,
0.044782877,
-0.0542928763,
0.001053011,
0.0481469631,
0.0670413002,
-0.0114556709,
-0.0031032301,
-0.0323876664,
0.0188641045,
0.0249199942,
-0.010870236,
0.0119366925,
0.0499236844,
-0.0436358973,
-0.0256409794,
-0.0077738836,
-0.0749981627,
-0.0240927134,
-0.0572558455,
-0.0067021027,
0.0047404924,
-0.0213401429,
0.0436367281,
0.0225834865,
0.0015612587,
-0.0010315911,
0.0812257975,
0.0386698321,
-0.0214427561,
-0.0122817671,
0.0178440697,
0.0700877234,
0.0230615251,
-0.0394103825,
0.0040376186,
0.012041539,
0.0287338067,
0.0146823404,
-0.0010048588,
0.1144333929,
-0.0251303706,
-0.0741951913,
-0.0035144181,
-0.1040171534,
-0.0354619697,
0.0777570307,
-0.0184760336,
-0.07150089,
0.0325114392,
0.0404164977,
0.0196719225,
0.0126439789,
-0.0100427978,
-0.015024771,
0.0404486209,
-0.0258351564,
0.0192165095,
0.0426779464,
0.025154151,
-0.0617677942,
0.0121417213,
0.0351499319,
-0.0357774273,
0.0150372889,
-0.0051902635,
-0.0231199618,
-0.0131589752,
-0.0293637495,
-0.0454489365,
-0.0092209652,
0.0045295027,
0.0052406089,
0.1013621241,
-0.0046678428,
0.0317615941,
-0.0545980185,
0.0892345235,
-0.0072465437,
0.0314625315,
0.0133540025,
0.0387673303,
-0.0206075367,
0.0173032284,
-0.0022115957,
0.0127663771,
-0.0518459305,
0.0311794207,
-0.0139902318,
-0.0095400084,
0.0048090499,
-0.0155850044,
-0.0003843925,
0.0040300456,
0.039689254,
-0.0178675409,
0.0130529152,
0.0622177608,
0.0071939244,
0.0194098987,
0.0313167796,
-0.0079811839,
-0.0333404206,
0.01193015,
-0.0055959262,
-0.0133517943,
0.0316152573,
0.027594734,
-0.0715293288,
0.0519345887,
-0.0121548465,
-0.0427224748,
-0.0605845861,
-0.0233440958,
-0.0328795277,
-0.0882285312,
0.0426167734,
0.0091640931,
-0.0255543739,
-0.0165566802,
0.0117152454,
-0.0286950879,
0.0355604999,
-0.035670612,
-0.0070019374,
0.0091456184,
0.0291580576,
-0.0242587738,
-0.0540909842,
-0.0208536964,
0.0045183203,
-0.0345311314,
0.0457693227,
0.1068076342,
0.0197830461,
-0.0249007307,
-0.0592360646,
-0.0112747131,
0.0452113077,
0.0189843904,
0.0170311481,
0.0240441002,
0.0272706766,
0.0168444589,
-0.0255193468,
0.0023067086,
-0.0279033277,
-0.0192648023,
0.0197751559,
0.0409764275,
0.0059578465,
-0.0529082268,
-0.0024062174,
-0.0280228835,
-0.0166849028,
0.0368166529,
0.0198071171,
-0.0657936111,
0.0581543967,
-0.0163436625,
-0.0343093239,
-0.050847657,
-0.0728366077,
-0.0402879119,
0.0011194898,
0.0286601912,
-0.0858745575,
-0.0289164092,
-0.0583744012,
-0.0007570395,
0.0333805382,
0.0136536695,
-0.0379999205,
-0.0070946598,
0.0280422829,
-0.0179732665,
0.0099078929,
-0.0258824565,
0.0143066915,
-0.0519095361,
-0.0486210324,
-0.0617455319,
0.0045751231,
0.0280556194,
-0.1098475605,
-0.0584413335,
0.066501677,
0.0154800666,
0.0235123429,
0.0093508102,
0.0376037806,
-0.0224350896,
0.1023543328,
-0.0005269433,
0.0228568465,
0.0334444456,
-0.0088987425,
0.069773443,
0.0163442492,
-0.0021731243,
-0.0115923099,
0.0009809897,
0.0514912568,
-0.0025616789,
-0.0054307999,
0.0029496313,
0.0303535275,
0.0308356322,
-0.0279210359,
-0.0696808398,
0.0043905061,
-0.02884271,
-0.0374096073,
0.0159200579,
-0.0242004301,
-0.0132223116,
-0.0446708202,
0.0186146572,
-0.0620475523,
0.0350284651,
0.0257588159,
0.0070177461,
0.0215631332,
-0.051638484,
0.0404648036,
-0.0039179767,
0.0509206206,
-0.0112311617,
0.0211762786,
0.0093136895,
0.1061291546,
-0.0606831834,
0.0290500578,
0.0011709984,
0.0877764151,
-0.0286692698,
-0.0375468619,
0.0293387678,
0.0661461353,
-0.0834658518,
-0.1169457957,
0.0462711155,
0.0630245879,
0.0417751856,
0.0459526666,
0.0181644671,
-0.0115097342,
0.0053922418,
-0.0167679153,
0.0352045745,
-0.0469195209,
-0.037601836,
-0.0365548655,
0.00109148,
-0.0350083001,
-0.0140147684,
0.0706259906,
-0.0733185783,
0.0062094601,
-0.0236498341,
0.0044232099,
-0.0075869886,
0.0368757732,
0.0207034182,
0.0123671554,
0.0061079958,
0.0582624674,
-0.0477058142,
-0.0200926457,
-0.007197402,
-0.0102882748,
0.0537844598,
-0.0165516008,
-0.0140557596,
0.0476296246,
-0.0104106385,
-0.0312468875,
-0.0199107733,
-0.0046105837,
-0.0749084055,
-0.0053784018,
-0.0267053228,
-6.017453295e-33,
0.0070609646,
-0.0349736921,
0.0197733399,
-0.0126205748,
-0.0160783306,
-0.0307022091,
-0.0412093513,
0.0766595826,
-0.0292793531,
-0.0035181846,
-0.0186625328,
-0.0102831628,
-0.011585813,
-0.0601466857,
0.0255290829,
-0.0007981348,
0.0132851256,
-0.0311399736,
-0.0277262125,
-0.0033344838,
0.0134677971,
-0.0037810116,
0.0148221916,
-0.0509000272,
0.0032174063,
0.0057062432,
0.0002362978,
-0.0700189173,
-0.0127554014,
0.010087884,
-0.0117301745,
0.0268861596,
0.0631365478,
-0.0418462045,
-0.023141386,
-0.0156414602,
0.0409205444,
-0.0093791457,
-0.0213951133,
-0.0423635244,
0.0260429215,
-0.0539683737,
0.0067838863,
-0.014455297,
-0.0484788381,
0.030388359,
0.0344698653,
-0.0057361852,
0.0067677014,
-0.0077862265,
-0.0495149121,
0.0023095137,
0.0033445237,
0.0069379266,
-0.0664668083,
0.0136673795,
0.0180465896,
-0.0099319536,
-0.046531152,
0.0588146187,
0.0109734107,
-0.000624766,
0.0434467867,
-0.0254501011,
0.0331805199,
0.0395908877,
0.0697016641,
0.0917365104,
-0.036145553,
-0.044411663,
0.0054579871,
0.026987426,
0.0013763122,
0.0506027825,
0.0092591057,
-0.0411670096,
0.046553947,
-0.00809565,
0.0046911533,
-0.0557903685,
0.0284350999,
0.0279036909,
-0.0381581038,
-0.0439697206,
0.0183601137,
-0.0734626502,
-0.0132470159,
-0.0293839108,
0.0092803147,
0.0014509431,
0.0143704573,
-0.0457498841,
0.0470280685,
-0.0548151098,
0.0052356948,
-0.0414387882,
-0.0666987598,
-0.0115063796,
0.0035636974,
-0.0462870039,
-0.0583165362,
-0.0154760126,
-0.0237416215,
0.0789105073,
0.0243826751,
0.0498506203,
0.0074301427,
0.0062576118,
-0.0343811288,
-0.0119947363,
-0.0207708403,
-0.0001337022,
0.0643474758,
-0.0655661225,
0.0351342931,
0.0018818735,
0.0010110186,
-0.0703118145,
0.0675532445,
0.0004221959,
-0.0546349473,
-0.0437316149,
-0.0605393313,
0.0218453947,
-0.0395405143,
-0.0075131059,
-0.0351098515,
0.0184532981,
0.0600173064,
-0.0296768863,
-0.0461494289,
-0.0479171388,
2.787e-7,
-0.0300777406,
0.0255838204,
-0.0208876245,
0.0576746501,
0.0480933711,
0.0229852367,
-0.0426022597,
0.026132958,
0.0165745132,
0.0321269333,
0.0005033812,
0.0164345913,
0.0026431477,
0.0179503225,
-0.0464461222,
-0.0008529918,
-0.0667827502,
0.0494068004,
-0.0340625532,
-0.0072554187,
0.0180991311,
0.0015817413,
0.0335306376,
-0.0167096537,
-0.0326895528,
-0.043386735,
0.0450866818,
-0.0132479332,
0.0224019438,
-0.0124449385,
-0.000694591,
0.013322616,
-0.0090666478,
0.0043165912,
-0.0251495503,
0.0055015418,
-0.0442150123,
-0.0248495787,
0.0343248621,
0.0407318212,
0.0380182005,
0.0209869035,
0.0129374703,
-0.0403641872,
0.0495249368,
0.0025926882,
-0.0158547089,
-0.0401888117,
0.0206546113,
-0.0220309999,
0.0184324868,
-0.0097501446,
0.0467888936,
-0.0188916493,
-0.0136205805,
0.0147633404,
0.0246047154,
-0.0269007925,
-0.0241863877,
-0.0471961983,
-0.0456583947,
-0.0552890375,
-0.0297594965,
0.0184526518,
-0.0390799344,
0.004575653,
0.0261736885,
2.398604258e-34,
0.0170837641,
-0.0340462513,
-0.00231621,
0.0485298187,
0.0612823516,
0.0027464903,
0.0253031421,
-0.0341683626,
0.002959965,
-0.015779417,
-0.0271902364
] |
http://arxiv.org/abs/2401.13478v2 | http://arxiv.org/pdf/2401.13478v2 | SciMMIR: Benchmarking Scientific Multi-modal Information Retrieval | 2024-01-24T14:23:12 | 2024-06-11T10:18:08 | Siwei Wu, Yizhi Li, Kang Zhu, Ge Zhang, Yiming Liang, Kaijing Ma, Chenghao Xiao, Haoran Zhang, Bohao Yang, Wenhu Chen, Wenhao Huang, Noura Al Moubayed, Jie Fu, Chenghua Lin | cs.IR | cs.IR, cs.CL, cs.CV, cs.MM | Multi-modal information retrieval (MMIR) is a rapidly evolving field, where
significant progress, particularly in image-text pairing, has been made through
advanced representation learning and cross-modality alignment research.
However, current benchmarks for evaluating MMIR performance in image-text
pairing within the scientific domain show a notable gap, where chart and table
images described in scholarly language usually do not play a significant role.
To bridge this gap, we develop a specialised scientific MMIR (SciMMIR)
benchmark by leveraging open-access paper collections to extract data relevant
to the scientific domain. This benchmark comprises 530K meticulously curated
image-text pairs, extracted from figures and tables with detailed captions in
scientific documents. We further annotate the image-text pairs with two-level
subset-subcategory hierarchy annotations to facilitate a more comprehensive
evaluation of the baselines. We conducted zero-shot and fine-tuning evaluations
on prominent multi-modal image-captioning and visual language models, such as
CLIP and BLIP. Our analysis offers critical insights for MMIR in the scientific
domain, including the impact of pre-training and fine-tuning settings and the
influence of the visual and textual encoders. All our data and checkpoints are
publicly available at https://github.com/Wusiwei0410/SciMMIR. | SciMMIR: Benchmarking Scientific Multi-modal Information Retrieval: Multi-modal information retrieval (MMIR) is a rapidly evolving field, where
significant progress, particularly in image-text pairing, has been made through
advanced representation learning and cross-modality alignment research.
However, current benchmarks for evaluating MMIR performance in image-text
pairing within the scientific domain show a notable gap, where chart and table
images described in scholarly language usually do not play a significant role.
To bridge this gap, we develop a specialised scientific MMIR (SciMMIR)
benchmark by leveraging open-access paper collections to extract data relevant
to the scientific domain. This benchmark comprises 530K meticulously curated
image-text pairs, extracted from figures and tables with detailed captions in
scientific documents. We further annotate the image-text pairs with two-level
subset-subcategory hierarchy annotations to facilitate a more comprehensive
evaluation of the baselines. We conducted zero-shot and fine-tuning evaluations
on prominent multi-modal image-captioning and visual language models, such as
CLIP and BLIP. Our analysis offers critical insights for MMIR in the scientific
domain, including the impact of pre-training and fine-tuning settings and the
influence of the visual and textual encoders. All our data and checkpoints are
publicly available at https://github.com/Wusiwei0410/SciMMIR. | [
0.0044440203,
0.0631747916,
0.0212451071,
0.0249614995,
-0.0225390885,
0.015590772,
-0.0168566909,
0.0432758853,
-0.0280159283,
-0.0411191173,
-0.0835572183,
0.0136196287,
0.0142378202,
0.044790931,
0.0321161188,
-0.0247591008,
0.0478302799,
0.0101347473,
-0.0094182128,
0.0262731146,
0.017504815,
-0.0067693866,
-0.0018733745,
0.0387706645,
-0.0068033533,
-0.0149532128,
0.0196978394,
0.029360462,
0.0194005258,
-0.0505905412,
-0.0204890799,
0.0478642061,
0.0755876228,
0.0286005586,
0.0000017951,
-0.0164944027,
-0.0324950889,
0.0458388217,
0.0500078499,
0.0101914303,
0.0833644867,
-0.0696541071,
0.0495193042,
-0.0510842279,
-0.0064064381,
0.0099092256,
0.0708371177,
0.0527608655,
0.0434475392,
0.0270589702,
-0.0198337678,
-0.0251284614,
0.0256722011,
-0.0185577627,
0.0134947747,
-0.0152951134,
0.0140305655,
0.0385810137,
0.0091969464,
0.0575960428,
-0.0212455392,
0.0504872985,
-0.0451315604,
-0.0147515461,
-0.018064009,
0.0506140776,
-0.0370187052,
-0.0207969733,
-0.0048125153,
0.0994529426,
0.0276162252,
0.0118957926,
-0.0113956649,
0.0117318667,
-0.0470824391,
0.0462281071,
-0.0572608784,
-0.0122757219,
0.0013105015,
-0.0127052497,
-0.0482782461,
0.0125751551,
-0.0010111483,
0.0030128409,
-0.0006391578,
-0.0019370263,
0.0207564887,
-0.0266694725,
-0.0016221754,
0.0062590824,
-0.0271099024,
0.0024102558,
0.0008794902,
-0.0038587821,
0.0093434574,
-0.0069192676,
-0.0303924847,
-0.0512229204,
0.0531609394,
0.006022587,
-0.0105722742,
0.0343421809,
-0.0568577759,
0.0784308612,
-0.0049112756,
0.0255568698,
-0.0282290317,
0.0582779422,
0.0070923781,
-0.0130021321,
-0.0508972555,
0.0073169861,
0.0071895439,
-0.0232567061,
-0.0491859056,
-0.0529919304,
-0.0174047332,
-0.0301810745,
-0.0063631414,
0.0067613893,
-0.0048683123,
0.0018924781,
-0.0559951439,
0.0033997723,
-0.0499224551,
0.0659363493,
0.0109821018,
-0.0057445103,
-0.0179495923,
-0.0213757437,
-0.0097356578,
0.0476293303,
0.0119672474,
-0.0449827649,
0.0176602881,
0.0636809692,
-0.053716965,
-0.036372304,
0.0423375741,
-0.0533305556,
-0.040912874,
-0.0534497686,
0.0629676655,
-0.0168764405,
0.0625898913,
-0.0012461941,
0.0031826512,
-0.0223319419,
0.0119877104,
0.0269998536,
-0.0500785783,
0.1188450158,
-0.0058906423,
-0.0463499948,
-0.0060216808,
0.0165678319,
-0.0205373447,
0.080400303,
0.0246542636,
-0.007507531,
0.0116653601,
-0.0139961205,
-0.0102272155,
-0.0142170107,
0.0289909765,
0.0005720927,
-0.0084904339,
0.0177007578,
-0.0006195204,
0.0577141941,
0.0309189558,
0.027399417,
-0.051992625,
-0.004760961,
0.0192870647,
0.0327231251,
0.0149766346,
-0.0008977567,
-0.0036514772,
-0.0024491611,
0.0011656204,
0.002403874,
0.0005468502,
0.0573690571,
0.0210853498,
0.0233169943,
0.0589585379,
0.0426870622,
-0.0115275858,
-0.0083785914,
0.0092363385,
0.0209216997,
-0.0182502884,
-0.0435203388,
0.0663213357,
-0.027664233,
-0.0454601906,
0.027104795,
0.0284658149,
-0.0355507061,
-0.0125017716,
-0.0043729902,
-0.0289309956,
0.0176325105,
0.0480321012,
-0.060614679,
-0.0127140097,
0.0074019348,
0.0157306306,
0.0418861583,
-0.035532143,
0.0024398428,
0.0121320048,
-0.0424519777,
-0.0519800037,
-0.0452988185,
0.0037993451,
0.0077083008,
-0.0329721645,
0.0348530784,
0.0170043018,
-0.0497219265,
0.008370786,
0.0360564366,
0.0324306637,
0.0208148994,
-0.0294674113,
-0.0014022066,
0.0129868248,
-0.0257641412,
-0.052504655,
0.0420812145,
0.0373144187,
-0.0138161853,
0.0213858634,
0.0221120585,
0.0012518812,
-0.0509793945,
0.0385073908,
-0.1083111614,
-0.0138756838,
0.0041167708,
-0.0203486886,
-0.0437579416,
0.0033933739,
-0.0211858172,
-0.0299833771,
-0.0102375885,
-0.0039161057,
-0.0096799508,
0.0278583635,
-0.0176969524,
0.0376824886,
-0.0127430493,
0.0343722552,
0.0440639183,
-0.0421738364,
0.046977032,
-0.0097566331,
-0.0311854184,
0.0238635261,
0.0062667108,
0.0243559573,
0.0113952374,
0.0034895937,
-0.0510791503,
-0.0104552349,
-0.1030631065,
-0.0002157663,
-0.0058118189,
-0.019083783,
-0.003381972,
-0.0471914038,
0.0314551555,
-0.0333396457,
-0.0154687734,
0.0005486281,
-0.0224613659,
0.0086651118,
0.0652520582,
0.0113660945,
0.0161599983,
-0.0402896032,
-0.0225057434,
-0.0317696519,
0.0632108524,
0.0044128601,
-0.0312962309,
0.0143951392,
-0.0022755074,
-0.0330752544,
-0.0119328136,
-0.0421602055,
0.0199218076,
0.0359384976,
-0.0173993297,
0.0114644282,
0.0387687646,
0.0111823846,
0.0088738455,
0.0288159475,
-0.0039219176,
-0.0619253553,
0.0229125321,
0.0083899386,
-0.0194447581,
0.0322369859,
-0.0223512556,
0.0074438048,
0.0000983191,
-0.0401371121,
0.0403926112,
-0.0290617552,
0.0013682852,
0.020685371,
-0.0300047919,
-0.0813466385,
0.0427786112,
0.0131736519,
0.0391333438,
-0.0325979665,
-0.0364046954,
0.0441587754,
-0.051073879,
-0.0215110034,
-0.0182193425,
0.0182661675,
-0.0496497154,
0.0484156795,
0.0044447407,
0.0208810288,
0.0222028326,
-0.0050838259,
-0.0237088855,
0.020747738,
0.0157640483,
-0.0372572318,
0.040339224,
-0.0390830897,
0.0354687683,
-0.046870932,
-0.0371162221,
0.0371043533,
-0.040218398,
-0.0162662249,
0.0156898461,
0.0024312346,
-0.0225943401,
-0.0756814256,
0.0177898929,
0.0364982374,
0.0015015232,
-0.0227285679,
-0.023968637,
-0.0266048014,
0.0634888262,
0.0568895191,
0.0229242314,
-0.0583963804,
-0.0333033167,
-0.0460142456,
0.0291397385,
-0.0246078894,
0.0070958463,
-0.0306931455,
0.0119723454,
0.0167797487,
0.0463933237,
0.0028577447,
-0.0120855616,
0.0494667962,
-0.0164425895,
0.0063837571,
0.0379029848,
0.001050203,
-0.0477607846,
-0.0696285442,
0.0067353523,
0.0572303683,
0.0152934529,
0.0068880673,
-0.0920578018,
-0.0515272915,
-0.0324993469,
-0.0028137891,
0.0207564775,
0.0052889506,
-0.0216603056,
-0.0538320392,
0.0465312675,
0.0043776724,
-0.0577222481,
-0.047619082,
-0.0571751371,
-0.0004562327,
-0.0180625655,
-0.0074238479,
-0.0246299729,
-0.0211826898,
-0.0173782296,
0.0233663991,
0.018591607,
-0.0932298675,
-0.0025103942,
-0.0473805033,
-0.0278350618,
0.0029048966,
0.0640064105,
-0.0329804756,
-0.0426009744,
-0.0148568517,
0.0054176738,
0.0070009362,
0.0004705393,
0.0755732656,
-0.0430844426,
-0.0269305315,
-0.0030327605,
0.0109268324,
-0.0084415739,
0.0032635839,
-0.0322827697,
0.0453911498,
0.0010546829,
0.0204140246,
-0.0980937928,
0.0129096676,
-0.0065451548,
0.0410854593,
0.0020133732,
0.008869566,
-0.0212443285,
-0.0115281753,
0.0478285365,
-0.0170367993,
-0.0454691462,
-0.0054138084,
-0.0261554029,
0.0164855793,
-0.0364603847,
0.0235935617,
0.0186124258,
0.0359930024,
-0.0171253569,
-0.0070014773,
-0.0617116317,
0.0151196206,
-0.0606316812,
0.0731312335,
-0.0411081389,
0.017309567,
-0.0244752131,
-0.0017027379,
0.0653957427,
0.0024904872,
-0.0394098014,
0.0239178203,
-0.0482011028,
-0.0279430319,
-0.0747294873,
0.0325135924,
0.0091991173,
0.0213854033,
0.0328426734,
-0.0298801828,
0.0331166722,
0.0666672066,
0.006574878,
-0.0204673465,
-0.0114657069,
-0.0261988044,
-0.0502761453,
-0.0284221545,
-0.042471163,
0.0025783433,
-0.021883117,
0.0302592274,
0.0117187593,
-0.0334636532,
0.0155029055,
0.0195183549,
0.0522494018,
-0.0350594483,
-0.0161816739,
-0.0122405607,
-0.0300916806,
-0.020166168,
-0.0264904052,
-0.0220470689,
0.0018120827,
0.0204673242,
-0.0200413875,
-0.0033382827,
0.0482987054,
0.0129116075,
0.0276995227,
-0.0384899266,
0.0272654723,
0.0473275669,
-0.0180605855,
-0.0088179102,
0.000335118,
-0.020696586,
-0.027306471,
-0.0174765904,
0.0290979017,
-0.0243654698,
0.0446750559,
0.0358808376,
0.0067305313,
-0.0290877447,
0.0604697727,
-0.0697883144,
0.0200323723,
-0.0270279422,
0.0027064818,
0.0609690472,
0.0058937287,
0.0544722751,
-0.0119027607,
-0.0464113764,
-0.0143656144,
0.0480245091,
0.0113673946,
0.0384963863,
0.0880524814,
0.0116191246,
0.0015329691,
0.0056637824,
0.0858862475,
0.0136776417,
0.017868951,
-0.0050776522,
-0.0392898694,
-0.0199770126,
0.0081951087,
-0.0146820694,
0.0089626042,
0.0118805487,
-0.0313519016,
-0.0310775619,
-0.0182637256,
0.0575394742,
0.0185655542,
-0.0411501117,
0.0091277361,
0.031690754,
0.0569810793,
0.0392554775,
-0.00045899,
-0.0219618883,
-0.0191276185,
-0.0022797212,
-0.0014400636,
-0.0428308733,
-5.916332863e-33,
-0.0309983529,
-0.0020334546,
0.0175997987,
-0.0324832723,
-0.0120181823,
0.04499593,
-0.0453197062,
-0.0034630098,
-0.0229398087,
0.0295183212,
0.0212567709,
-0.0251987725,
0.0008223479,
-0.0035202254,
0.0132978847,
0.0137014706,
0.0281994902,
-0.0169493072,
-0.033305265,
0.0044328012,
0.0537672266,
0.0072736135,
0.0187611897,
-0.088065505,
-0.0039377129,
0.0460475199,
-0.0051472955,
-0.0141806323,
0.0406267941,
0.0718692467,
-0.0445846915,
0.0161909405,
0.0218303576,
-0.0189853124,
-0.0264105443,
0.0043494133,
-0.0607101321,
0.0020053333,
-0.0051723225,
-0.0090051461,
-0.066471234,
-0.0143282861,
0.0347448662,
-0.0291713011,
0.0044737486,
-0.0705448762,
0.0775464177,
0.0103268903,
-0.0321806557,
-0.0231122505,
-0.0449722819,
0.019068744,
-0.0520190522,
-0.0001671373,
0.0712277517,
0.0527606234,
0.0523432605,
0.1059859172,
-0.039756421,
0.0547408983,
-0.025157202,
-0.0083491756,
0.0386957228,
0.0123931039,
0.0213699564,
0.081232667,
0.0144341113,
-0.0047787013,
-0.0111412238,
-0.0357226729,
0.0359308273,
0.0801211298,
0.0464888737,
0.0688750967,
0.0511712655,
0.024930533,
-0.0351824574,
0.0514010899,
-0.0360306427,
0.0260823034,
-0.0122171193,
0.0436525829,
0.0095292572,
0.009438036,
-0.0369552486,
-0.0137691041,
-0.0529924072,
-0.0024455988,
0.0205815136,
0.0066112191,
-0.0091678044,
0.0701026171,
0.0633796901,
0.0053830012,
-0.0187750608,
0.0227729809,
-0.0038697924,
0.0598492958,
-0.0098797046,
-0.0295447111,
-0.0390436873,
-0.0580030642,
0.083425194,
-0.0036566062,
0.0362582356,
-0.0237927269,
0.0596395396,
-0.0146961613,
-0.024972653,
-0.0177324899,
-0.0177936722,
-0.0848303586,
-0.0310666189,
0.0132241836,
0.0361514874,
0.0244562551,
-0.0317262448,
0.002178374,
0.0211586952,
0.0568032488,
-0.0093604112,
-0.0533269607,
0.012905404,
0.0152453622,
0.034815304,
0.0113601089,
-0.0392848738,
-0.0697963983,
0.0054341205,
-0.0593074374,
-0.0088461507,
0.0399129838,
2.626e-7,
0.0844222456,
0.0544691905,
0.0488046594,
-0.0187824667,
0.0202457514,
0.0236509163,
-0.0646704659,
0.022728188,
-0.0087391278,
-0.0413184464,
0.0067780437,
-0.0176559146,
0.044724308,
0.0353469066,
-0.0055463593,
-0.0368097648,
-0.054047104,
-0.0291248057,
-0.0431824066,
-0.0054784738,
0.070019342,
0.0712957755,
-0.0765870139,
-0.0080700098,
0.0196268335,
-0.0549642779,
0.1281258911,
-0.0415019169,
0.0177404452,
0.016496595,
-0.0299556702,
0.0440629572,
-0.0202590097,
0.0352538191,
-0.0139035322,
-0.0013839756,
0.0734382719,
0.0284795109,
-0.0130553115,
-0.0075821457,
0.009664705,
0.0716376901,
0.0001814233,
-0.0090747718,
0.0712239668,
-0.0572612993,
-0.0301288534,
-0.0284638032,
-0.0370609313,
0.0353727043,
0.0670104325,
0.0012676194,
-0.0593212061,
0.0456100143,
0.0031448726,
0.0093783196,
-0.0306480713,
-0.0620001443,
0.0117820352,
0.0292341672,
-0.0420880206,
-0.0092334347,
-0.0239214618,
-0.0465061292,
0.0280421171,
-0.0397421159,
-0.05318214,
2.315187093e-34,
0.0550512038,
-0.0446265489,
0.0204730984,
0.0022975549,
0.029158853,
-0.0320947915,
-0.0167121124,
-0.0246302206,
-0.0332089886,
-0.087222822,
0.0137246577
] |
http://arxiv.org/abs/2401.13463v2 | http://arxiv.org/pdf/2401.13463v2 | SpeechDPR: End-to-End Spoken Passage Retrieval for Open-Domain Spoken Question Answering | 2024-01-24T14:08:38 | 2024-03-18T06:08:31 | Chyi-Jiunn Lin, Guan-Ting Lin, Yung-Sung Chuang, Wei-Lun Wu, Shang-Wen Li, Abdelrahman Mohamed, Hung-yi Lee, Lin-shan Lee | cs.CL | cs.CL, cs.IR, cs.SD, eess.AS | Spoken Question Answering (SQA) is essential for machines to reply to user's
question by finding the answer span within a given spoken passage. SQA has been
previously achieved without ASR to avoid recognition errors and
Out-of-Vocabulary (OOV) problems. However, the real-world problem of
Open-domain SQA (openSQA), in which the machine needs to first retrieve
passages that possibly contain the answer from a spoken archive in addition,
was never considered. This paper proposes the first known end-to-end framework,
Speech Dense Passage Retriever (SpeechDPR), for the retrieval component of the
openSQA problem. SpeechDPR learns a sentence-level semantic representation by
distilling knowledge from the cascading model of unsupervised ASR (UASR) and
text dense retriever (TDR). No manually transcribed speech data is needed.
Initial experiments showed performance comparable to the cascading model of
UASR and TDR, and significantly better when UASR was poor, verifying this
approach is more robust to speech recognition errors. | SpeechDPR: End-to-End Spoken Passage Retrieval for Open-Domain Spoken Question Answering: Spoken Question Answering (SQA) is essential for machines to reply to user's
question by finding the answer span within a given spoken passage. SQA has been
previously achieved without ASR to avoid recognition errors and
Out-of-Vocabulary (OOV) problems. However, the real-world problem of
Open-domain SQA (openSQA), in which the machine needs to first retrieve
passages that possibly contain the answer from a spoken archive in addition,
was never considered. This paper proposes the first known end-to-end framework,
Speech Dense Passage Retriever (SpeechDPR), for the retrieval component of the
openSQA problem. SpeechDPR learns a sentence-level semantic representation by
distilling knowledge from the cascading model of unsupervised ASR (UASR) and
text dense retriever (TDR). No manually transcribed speech data is needed.
Initial experiments showed performance comparable to the cascading model of
UASR and TDR, and significantly better when UASR was poor, verifying this
approach is more robust to speech recognition errors. | [
-0.0047784648,
0.043345809,
0.0000195741,
0.0298982635,
-0.0189496465,
-0.0085571529,
0.0071147126,
-0.0073050805,
0.0087872762,
-0.0282838792,
0.0176657047,
-0.0393529981,
-0.0125072189,
0.0666050687,
-0.0242869277,
-0.0291999076,
0.0456884615,
-0.0426470935,
0.0195838138,
-0.0095367385,
-0.0024419799,
0.0166939795,
-0.0188162457,
-0.0017642371,
0.0105339345,
-0.005033887,
-0.0321874022,
-0.0092576332,
0.043441575,
-0.0032326856,
0.0563494302,
-0.033252351,
0.0758520886,
-0.009888418,
0.0000018394,
-0.0318632424,
-0.0734310299,
0.0193428062,
-0.0414220281,
-0.0347512625,
-0.0024800689,
0.0479358546,
0.0493896566,
0.043985609,
-0.0228695832,
-0.0112280222,
-0.0024419602,
0.0537654348,
0.0403679647,
0.0481751859,
-0.009385827,
-0.0276762489,
0.0182390343,
0.0089155696,
0.0024552825,
-0.0733163431,
0.0049347579,
-0.0389945842,
0.044045683,
0.0271726195,
0.0170410834,
0.0464859307,
-0.0065050479,
0.030803103,
-0.0019326293,
-0.010270441,
-0.0263888333,
0.0373624302,
-0.0122069176,
0.0775467679,
0.0854640156,
0.0003850007,
0.0220412873,
0.0023889574,
-0.0335093103,
0.010717947,
-0.0171514302,
0.0268712062,
-0.0301020853,
0.004633503,
-0.0578586087,
0.0457828604,
-0.0297283549,
0.0088058934,
-0.0451609492,
0.0151255298,
0.0078326697,
-0.0203676336,
0.0053208764,
-0.0451278165,
-0.0224504266,
-0.008615328,
0.0328982137,
0.0385135263,
0.002280141,
0.0234713759,
-0.0017063963,
-0.0118449423,
0.0053734966,
-0.0231871046,
0.0251917783,
0.0348332338,
-0.0708580986,
0.0314721316,
-0.0066433614,
0.06255126,
-0.0224669762,
-0.015228603,
-0.0424728543,
-0.0513383672,
-0.0443945676,
-0.0048262281,
-0.0679321364,
0.0772749409,
-0.082545273,
-0.0100935204,
0.0132981027,
0.0075811427,
0.0270117652,
-0.0209872574,
0.0016813384,
0.0515417121,
-0.0259315018,
-0.0001712772,
-0.0293554366,
0.0561813712,
-0.0062557547,
0.0249719936,
0.0122575676,
-0.0188743565,
-0.0320804082,
-0.0061871861,
0.0472137481,
-0.0051333751,
0.0032177106,
0.0294661541,
-0.0263616443,
-0.0521505959,
-0.0394394286,
-0.0094919875,
-0.0226672404,
-0.0440475307,
-0.0226142015,
-0.053129565,
0.0292011201,
0.0295813028,
-0.0423691198,
-0.0238476973,
0.0273748413,
0.0476402491,
-0.0263435822,
0.0247509256,
-0.0043337867,
-0.0165883135,
0.0366731659,
0.0206007753,
0.0521759987,
0.0383617952,
0.0162290093,
0.0085206721,
0.0366222747,
0.027812453,
-0.0006359566,
0.0014082999,
0.0023764384,
0.001682349,
-0.0209868662,
0.0661395192,
0.023406595,
-0.0105330208,
-0.0057574152,
0.0667750835,
-0.0259708613,
-0.0242246352,
-0.0073006721,
0.0677912235,
0.0087597556,
-0.0057745674,
-0.0023074178,
0.0289822165,
0.0161032528,
0.0141629949,
0.028820416,
0.0168595128,
-0.0071004042,
0.0011660182,
0.0511262231,
0.0028149069,
-0.0250839125,
-0.0135220969,
-0.0043748571,
-0.0019331947,
-0.0094141727,
0.0737752169,
0.0604532585,
0.0082615977,
-0.0672605187,
0.0560763776,
0.0055364273,
-0.0705359429,
-0.0054230285,
-0.0251493528,
0.0321390592,
0.0141604953,
0.0484338105,
-0.072808139,
0.0038852871,
-0.037507683,
-0.003156543,
0.038551759,
-0.0165091623,
-0.0056957076,
0.0394574404,
-0.0494264811,
0.0123980017,
0.0037700154,
-0.0146577647,
0.055242274,
-0.0040279282,
0.0273707695,
0.0225585103,
-0.0347657092,
0.015840143,
0.0079256836,
-0.0206129681,
-0.0184865762,
0.0306149311,
0.002064409,
0.0492602997,
0.0572193339,
-0.0270564463,
-0.0086644962,
0.0427358635,
-0.0565502271,
-0.052975364,
-0.0198259186,
-0.0784543753,
-0.0151483603,
0.0164984576,
-0.0056350036,
-0.0019524765,
0.0625512376,
0.0028396093,
-0.0600902252,
0.0102790268,
-0.0368238464,
0.014059335,
-0.0766325369,
-0.0142741948,
0.0273323115,
-0.0123734102,
-0.0544228293,
0.0465301648,
0.0282409787,
0.0365026966,
0.0088308929,
0.0211772099,
0.0362032615,
-0.044676099,
-0.0135145262,
-0.0072689517,
-0.0483636521,
0.0425506234,
0.0521817245,
-0.0648914576,
-0.0162723847,
-0.0175749399,
-0.045225542,
0.0326943472,
-0.0231011715,
0.0250078775,
0.0039870124,
0.0516779944,
-0.0029910861,
-0.0357000642,
-0.0239464752,
-0.0112582156,
0.0440466702,
0.0028403576,
-0.0137179252,
0.0143375108,
-0.0048212279,
-0.0216762163,
0.0350325331,
-0.0009136428,
0.0084320297,
-0.0420211703,
-0.0344529711,
-0.0592218786,
0.042081587,
-0.0032544138,
0.0746900514,
-0.0429341979,
-0.0042730165,
-0.0141716283,
-0.0365343057,
-0.0065874541,
0.0486174636,
0.0225372966,
-0.0308780372,
-0.0088448264,
-0.0212353058,
-0.0530563556,
0.0609386526,
-0.0083439322,
-0.0235541202,
0.0014419127,
0.0431729816,
0.0170591772,
0.0072991247,
0.0474511907,
0.0094597246,
0.0124344844,
-0.053518679,
0.0030446246,
0.0150511833,
-0.049631808,
-0.0284247473,
-0.0463326313,
0.0610416718,
0.0104056187,
0.0045326464,
0.0144050354,
-0.0432354994,
-0.0357472114,
-0.0664650425,
-0.0563978702,
-0.0756219551,
0.0714751855,
-0.0359084681,
0.0275501348,
0.0202561207,
-0.0645068437,
-0.0406750739,
-0.0111895073,
-0.0149565861,
-0.0673826933,
0.0527893901,
0.0043696612,
0.0322370529,
-0.0378130265,
-0.0377991274,
-0.0720886365,
-0.0099527286,
-0.0179895572,
0.048862271,
-0.0460114107,
-0.0022482581,
-0.0303148571,
0.0275765434,
0.0289188121,
0.0324645564,
-0.0094889635,
0.0192056205,
-0.0275305975,
0.0823832229,
0.0575150326,
0.0372642577,
-0.016688522,
0.0019147564,
0.0281215906,
0.0423359498,
0.0235206056,
-0.0285834298,
0.0081420606,
-0.0151660098,
0.039781969,
-0.0345584415,
0.0269782227,
0.0223272536,
0.0547713973,
-0.0280161966,
-0.0375500731,
0.0644951984,
0.0278879292,
0.0001775539,
-0.0718442574,
-0.0285273343,
-0.005169176,
0.0283593461,
0.0416652411,
-0.0892789066,
-0.0271484479,
0.0128462957,
0.0666661263,
0.0371695831,
-0.0115985796,
-0.0090939729,
-0.075603649,
0.0371361822,
-0.0025973029,
-0.0720014125,
-0.0496150814,
-0.0519716889,
0.0050480687,
0.0175529979,
0.0158058889,
0.0320741236,
-0.0314385109,
-0.0136865424,
0.0001686266,
-0.0029872693,
0.0450898632,
0.0383780189,
-0.0173571277,
0.0058338465,
0.0345028117,
0.0203673467,
0.0197119284,
-0.0059655188,
-0.033716619,
-0.0943193585,
0.0048730639,
0.0280891992,
0.0566066504,
0.0022142471,
0.0003081664,
-0.0331426635,
0.0147280144,
-0.007102666,
-0.0060376246,
0.0211996362,
0.0513703674,
-0.0062190653,
0.0141837513,
-0.0192724299,
0.0309746135,
0.005960355,
0.0512554012,
-0.0142174177,
0.0279596969,
-0.0593751296,
0.0133210598,
0.0084265461,
-0.037052352,
-0.0171549302,
0.0115706893,
0.0136876889,
0.0359084271,
-0.0401084349,
-0.0209149588,
0.0018834106,
0.0440461747,
-0.018581368,
0.0246649981,
-0.002159345,
0.0218467005,
-0.0633238554,
-0.0043995818,
-0.0504959561,
0.0229271185,
-0.0553036444,
-0.0144663863,
0.0394711718,
-0.0658104494,
-0.0157542024,
0.020543253,
-0.0216014329,
-0.0508372709,
0.004454568,
-0.0149230827,
0.0111285402,
0.0122344214,
-0.0051501808,
0.0028972989,
0.1247212738,
0.0138844782,
-0.0963004678,
0.0068044434,
0.0218714792,
0.0424123369,
-0.0545274653,
0.0169543568,
-0.0707902983,
0.025163155,
-0.0412565544,
0.0486833826,
0.0213055611,
-0.0960071236,
0.0160339084,
0.0203499179,
0.057807371,
0.0799098313,
-0.0176392421,
-0.0295913909,
0.0510687865,
0.0021915289,
-0.0040324996,
-0.0391494669,
0.0154353147,
-0.0221797563,
-0.0553402379,
0.0068849237,
0.0247617755,
-0.0319166444,
0.0432651676,
-0.0025363038,
0.0283535644,
0.0666912794,
0.0128075201,
0.0169504471,
0.0156706795,
-0.0128508406,
-0.0158485789,
0.0077600181,
0.0809397623,
0.0117991064,
0.0790684745,
-0.0211112779,
-0.0050057606,
0.0007349492,
-0.0111803804,
-0.1085023955,
-0.0317533538,
0.0012819159,
0.0220250115,
0.0093986159,
-0.00373951,
0.0174252875,
0.031426549,
0.0123477736,
-0.0446844064,
0.0154560842,
-0.0275396109,
-0.0028124326,
0.0298057087,
0.0164482053,
-0.0163865425,
0.0457442664,
0.0362454019,
-0.0058213999,
0.0373766869,
0.0028520278,
0.0038165012,
-0.0212988779,
-0.055404935,
0.0021563759,
0.0274577718,
-0.0128769241,
-0.0178181343,
0.0021226755,
-0.0409359671,
0.0520129837,
-0.0304146539,
-0.0005813212,
-0.0254893843,
-0.0033836057,
0.0119489236,
0.0191619489,
-0.0234988499,
-0.0190909132,
-0.0031716244,
-0.0405681804,
-0.0049281316,
0.0029970845,
-6.106476788e-33,
-0.0391636267,
0.0142457411,
0.0053686709,
0.0120987417,
-0.0368581042,
0.0028560972,
-0.0172978882,
0.0128452405,
-0.0040780501,
-0.0367689244,
-0.012237695,
-0.0142223258,
0.0043947967,
0.0494559482,
-0.0323496424,
-0.014562659,
0.0433487482,
-0.0374183655,
-0.012676782,
-0.0021556392,
0.0946929306,
0.065055266,
-0.0025888006,
-0.0185912568,
-0.0291258935,
-0.0319329426,
0.0398566946,
0.0080815768,
0.0633057654,
0.0345485695,
-0.0022292042,
0.0237966627,
-0.0039134948,
0.0037505687,
-0.0079953056,
0.0586100295,
-0.1209294274,
-0.0168074202,
-0.0074695176,
-0.0031905682,
-0.0120687131,
-0.0050295983,
-0.0166807715,
-0.0395677462,
0.039379511,
-0.0291657634,
0.0165047254,
-0.002430938,
-0.0040316363,
-0.0290769115,
-0.001708306,
0.0236009378,
-0.0056938487,
-0.0268801227,
0.0138884084,
0.082373932,
0.0405761078,
-0.0325107239,
-0.0187171195,
0.0003902516,
-0.0149350772,
0.0170931555,
-0.0247169547,
0.0162168574,
-0.0371121988,
0.0484220423,
0.0265924968,
-0.0907851905,
-0.0370379165,
-0.0599073581,
-0.0170205086,
0.0111907236,
-0.0086129773,
0.0429761373,
0.012203089,
0.0011848626,
-0.0577776395,
0.0155729065,
0.0096342806,
0.0272486769,
0.0285003539,
0.0152647076,
-0.0362235308,
-0.0061741034,
-0.03019787,
0.0483266376,
-0.0385513492,
-0.0217099693,
0.0781094283,
0.0230945405,
0.0172730666,
-0.020080667,
0.0694818944,
-0.0438303947,
0.0601461232,
0.0225247703,
0.0175536815,
-0.0121615417,
-0.0031943256,
0.0551397465,
-0.0542546995,
0.0139113124,
0.0460353084,
-0.0065646027,
-0.0437221415,
-0.0185019858,
-0.0193229038,
0.0093222056,
-0.0865649804,
0.0311497934,
-0.0015744705,
0.0401420034,
-0.0323434062,
0.0428064764,
-0.0039919489,
0.0190153215,
0.0014449084,
0.0849719718,
-0.0295411088,
0.0184956454,
-0.0578563586,
-0.0362541601,
0.0076804971,
-0.0079792496,
-0.0389183275,
0.0112228263,
0.004130146,
0.0463473164,
-0.0251695532,
-0.0074967504,
0.0277627744,
0.0388457477,
2.62e-7,
-0.0068250312,
-0.0316374302,
0.0567473099,
0.0185586493,
-0.0207456276,
-0.0055472245,
0.0131244464,
0.0146710388,
0.0302104205,
0.0008542947,
-0.0199847966,
-0.0023041237,
0.060646873,
0.0422328599,
-0.1451721489,
0.0075486335,
-0.0200878382,
-0.034603633,
-0.0060771615,
-0.042480804,
0.0742637515,
0.0500958934,
-0.0338324793,
-0.0066289948,
-0.0423846729,
-0.0202688314,
0.04969218,
0.0346129797,
0.0105948802,
-0.0212497227,
-0.0446783416,
0.0115371104,
-0.0266838986,
-0.0154621555,
0.0129452571,
0.0946571752,
0.0259174388,
0.04346589,
0.0184127633,
-0.1287138313,
-0.0249274243,
0.0350342207,
0.0164445993,
0.0023595488,
0.0717333481,
-0.0296709109,
-0.0159235597,
-0.0201406274,
-0.0696390048,
0.0024442959,
-0.022637669,
0.0549541935,
0.0066021555,
0.0245460924,
-0.0080619669,
-0.0104874307,
0.0095757833,
-0.1028984115,
0.0380212776,
-0.0042534624,
-0.0343757309,
-0.0086809676,
-0.0017986648,
0.0221096836,
-0.0172983762,
-0.0580006577,
0.0174760912,
1.959861324e-34,
0.00984638,
0.0109836031,
-0.0309172627,
-0.0119915837,
0.0308293458,
-0.0465416126,
-0.0097570401,
0.0145737035,
0.0400006399,
-0.1078505293,
-0.0157411862
] |
http://arxiv.org/abs/2402.01700v1 | http://arxiv.org/pdf/2402.01700v1 | Question answering systems for health professionals at the point of care -- a systematic review | 2024-01-24T13:47:39 | 2024-01-24T13:47:39 | Gregory Kell, Angus Roberts, Serge Umansky, Linglong Qian, Davide Ferrari, Frank Soboczenski, Byron Wallace, Nikhil Patel, Iain J Marshall | cs.CL | cs.CL, cs.AI | Objective: Question answering (QA) systems have the potential to improve the
quality of clinical care by providing health professionals with the latest and
most relevant evidence. However, QA systems have not been widely adopted. This
systematic review aims to characterize current medical QA systems, assess their
suitability for healthcare, and identify areas of improvement.
Materials and methods: We searched PubMed, IEEE Xplore, ACM Digital Library,
ACL Anthology and forward and backward citations on 7th February 2023. We
included peer-reviewed journal and conference papers describing the design and
evaluation of biomedical QA systems. Two reviewers screened titles, abstracts,
and full-text articles. We conducted a narrative synthesis and risk of bias
assessment for each study. We assessed the utility of biomedical QA systems.
Results: We included 79 studies and identified themes, including question
realism, answer reliability, answer utility, clinical specialism, systems,
usability, and evaluation methods. Clinicians' questions used to train and
evaluate QA systems were restricted to certain sources, types and complexity
levels. No system communicated confidence levels in the answers or sources.
Many studies suffered from high risks of bias and applicability concerns. Only
8 studies completely satisfied any criterion for clinical utility, and only 7
reported user evaluations. Most systems were built with limited input from
clinicians.
Discussion: While machine learning methods have led to increased accuracy,
most studies imperfectly reflected real-world healthcare information needs. Key
research priorities include developing more realistic healthcare QA datasets
and considering the reliability of answer sources, rather than merely focusing
on accuracy. | Question answering systems for health professionals at the point of care -- a systematic review: Objective: Question answering (QA) systems have the potential to improve the
quality of clinical care by providing health professionals with the latest and
most relevant evidence. However, QA systems have not been widely adopted. This
systematic review aims to characterize current medical QA systems, assess their
suitability for healthcare, and identify areas of improvement.
Materials and methods: We searched PubMed, IEEE Xplore, ACM Digital Library,
ACL Anthology and forward and backward citations on 7th February 2023. We
included peer-reviewed journal and conference papers describing the design and
evaluation of biomedical QA systems. Two reviewers screened titles, abstracts,
and full-text articles. We conducted a narrative synthesis and risk of bias
assessment for each study. We assessed the utility of biomedical QA systems.
Results: We included 79 studies and identified themes, including question
realism, answer reliability, answer utility, clinical specialism, systems,
usability, and evaluation methods. Clinicians' questions used to train and
evaluate QA systems were restricted to certain sources, types and complexity
levels. No system communicated confidence levels in the answers or sources.
Many studies suffered from high risks of bias and applicability concerns. Only
8 studies completely satisfied any criterion for clinical utility, and only 7
reported user evaluations. Most systems were built with limited input from
clinicians.
Discussion: While machine learning methods have led to increased accuracy,
most studies imperfectly reflected real-world healthcare information needs. Key
research priorities include developing more realistic healthcare QA datasets
and considering the reliability of answer sources, rather than merely focusing
on accuracy. | [
0.0940623358,
-0.00689389,
-0.0105384672,
-0.0382180177,
-0.0106868232,
0.0159486439,
0.0266683046,
0.014852725,
-0.0137542291,
0.026075609,
0.0206368547,
-0.0187919829,
0.001172725,
-0.0023392874,
-0.0057289405,
-0.0258745532,
-0.012038379,
0.0199109614,
0.0469339862,
0.0147934351,
0.0166936815,
-0.0004228193,
-0.0490904488,
0.0617881678,
0.0178438928,
-0.0010477421,
0.0028788296,
-0.0360046178,
0.0217408147,
-0.0832020938,
-0.0150196552,
-0.0090798596,
-0.0420787446,
-0.0495181233,
0.0000023455,
-0.0663930178,
-0.0064888559,
0.0450346768,
-0.0754444599,
-0.0558092631,
0.0123833641,
0.0057452102,
0.0221813787,
0.0377171487,
-0.0159283988,
-0.0134870866,
0.0500116982,
0.0025868509,
0.0383414328,
-0.0042928602,
-0.0135634942,
0.0090888729,
0.0071120681,
0.0147003196,
0.0374088213,
-0.0538214594,
0.031467732,
0.0363904573,
0.0600365214,
-0.0134563055,
0.0494764186,
0.036470063,
-0.0006062554,
0.0261663832,
0.025499206,
0.0289168004,
-0.0049152058,
-0.0380477458,
-0.0012857252,
0.0558969341,
0.0586745813,
0.0141700292,
-0.0135165853,
-0.0223658886,
0.0104178861,
0.0212734491,
-0.0366101898,
-0.0248543657,
-0.0451687649,
-0.0378094427,
-0.030592002,
0.0149917891,
0.0177643616,
-0.0259474237,
0.0097211264,
-0.0439664461,
0.0044572228,
-0.0070463414,
0.0325448811,
-0.0121138282,
0.0397878587,
-0.0285400227,
0.0447863191,
0.0461277887,
-0.0368960984,
-0.0099738296,
0.0636830702,
-0.0054137306,
0.0909921452,
-0.0302197095,
0.0129473079,
-0.0086132856,
-0.0143205738,
0.001675826,
0.0416586697,
-0.0127848787,
0.0032952847,
0.0371421576,
-0.068683587,
0.0157267414,
-0.0496424399,
-0.0343151949,
-0.0127609577,
0.0190148111,
0.0027907591,
0.032451041,
-0.042574279,
-0.0393682234,
0.0154155465,
0.0165677704,
-0.0207272694,
0.0643605366,
0.0104385726,
0.0217900909,
-0.0622983873,
0.0041529438,
-0.0544360727,
0.032794416,
-0.0087301685,
-0.0353871025,
0.0263622832,
-0.0252587218,
0.0366601273,
0.0329603776,
0.0336340107,
0.0324903168,
-0.0071844994,
-0.0282660276,
0.0102715939,
-0.0371760987,
-0.0339761823,
-0.0206049159,
-0.0068933209,
-0.0281216707,
0.0396978445,
0.0385080166,
-0.0115004098,
0.0093011465,
0.0273352005,
0.0602804013,
-0.0421805792,
0.0420128778,
-0.0631646737,
0.0424229763,
0.0533825643,
0.0225968696,
0.008357862,
-0.0193006899,
-0.0048495461,
-0.0248210616,
0.0161768273,
-0.033715006,
0.0184142645,
-0.0476590618,
0.0172683094,
-0.0687389895,
0.0277669672,
0.0949526355,
-0.0328866057,
0.0727778077,
-0.085147053,
0.0953818783,
0.0096488129,
-0.0024168473,
-0.0041243122,
0.0741774663,
-0.0049436642,
0.0197760127,
-0.0159115903,
-0.0455311015,
0.0211574975,
-0.0292721614,
0.0237034522,
0.0229670368,
0.0219798349,
-0.0043950295,
-0.0125598973,
-0.0311029963,
-0.0316460878,
-0.0511831865,
-0.0705299005,
-0.0069812681,
-0.028980203,
0.0528183021,
0.0011924156,
-0.0261521935,
0.0031589218,
0.0118449694,
0.0159965623,
-0.0168771911,
-0.0329761617,
-0.0603354275,
0.0484228805,
0.0741110966,
0.0289457161,
-0.0114491964,
-0.0081792958,
-0.0408914015,
0.0257773995,
-0.0243230499,
0.0003369759,
-0.0369619802,
0.008614732,
-0.0299938861,
-0.0058382591,
0.0009171162,
-0.0190201178,
-0.0246458761,
-0.0098446766,
0.0367014147,
-0.0324178934,
-0.0404680409,
-0.0487493351,
0.0341146365,
-0.0256357361,
-0.0368383601,
-0.0007073985,
-0.0072286944,
-0.0031379717,
0.0586512759,
0.0103516215,
0.0455605909,
0.0383452736,
0.0003986428,
-0.0115342261,
-0.0052196402,
0.0025106112,
0.0192424823,
0.0138687929,
-0.0158161614,
-0.0133044831,
-0.0144884875,
-0.0284649264,
-0.0296395235,
-0.07412,
-0.0904599577,
-0.0277704205,
0.0103241792,
-0.0394811668,
0.0273893643,
-0.0064441836,
-0.0671004355,
-0.0330924504,
0.0079688895,
0.0248434022,
0.0250398815,
-0.0418064483,
0.0092100846,
-0.0225778688,
-0.02850754,
0.0262958165,
-0.018924851,
0.0105693089,
-0.0175878219,
0.0138682881,
-0.0468680933,
0.0334684774,
-0.0706275627,
-0.0212678351,
0.0125277443,
0.007556037,
0.0309244767,
0.0071092942,
0.0047490378,
-0.042838756,
-0.0478552729,
-0.0053573013,
0.0113012241,
0.0240753181,
0.0474205539,
0.0828228369,
0.0465050265,
-0.06461972,
-0.0148772104,
-0.0043114386,
0.0247478206,
0.0622735731,
0.0373723768,
0.0079129012,
-0.0771459341,
-0.0493681841,
-0.0273732767,
0.0353530832,
0.0036081597,
0.0380947217,
0.0046538543,
-0.0378914513,
0.0433705077,
0.0195566099,
-0.0289351624,
0.0173389539,
0.0230807867,
-0.0144726625,
0.0190666895,
-0.0491577648,
-0.0496527478,
0.0439705513,
-0.0626455992,
0.004314844,
0.0216892473,
0.0377127267,
0.0123554766,
-0.0181546323,
0.0278597493,
-0.0118022114,
0.0267693643,
-0.090001747,
0.0016139678,
-0.0360601693,
0.0654984117,
0.0105890175,
-0.0034745538,
0.0867881924,
0.014860535,
0.005611991,
-0.0800606087,
-0.0192272943,
-0.0566991009,
-0.0224054847,
-0.0522732548,
0.0289784409,
0.0229984634,
0.020070916,
0.0004149717,
-0.0358148143,
-0.0190845262,
-0.0456563495,
0.0049384395,
-0.0049509495,
0.0327502862,
-0.0045324024,
-0.0374164954,
-0.0602107272,
0.0128217712,
-0.0509478599,
-0.0209548268,
0.0048365425,
0.0266542472,
-0.0098675601,
0.032928355,
0.0364974327,
0.0328408591,
0.0228008032,
-0.0379135869,
0.0215024576,
0.0174726676,
0.0244686045,
0.0444400944,
0.0569943823,
0.0392035432,
0.0377662219,
0.0786187127,
0.0186745357,
-0.0066517354,
0.0129329124,
-0.0225085691,
-0.0138222333,
-0.0263343155,
0.0080970433,
0.0217570383,
0.0208257306,
-0.0364424698,
0.0055002649,
0.0343028307,
0.0609445833,
-0.0412479751,
0.0084378468,
-0.0368706658,
-0.076897949,
-0.0246436074,
0.013717629,
-0.0539964512,
-0.0338836536,
-0.0369809531,
0.0831846595,
-0.0385074355,
-0.0051952228,
0.0059694615,
-0.010988418,
0.0207390394,
-0.0468809083,
-0.0365148969,
-0.0205127504,
-0.0304256175,
0.0204719733,
0.0087043019,
0.001443687,
-0.0016643836,
-0.0098518608,
0.0019044884,
0.0142192338,
-0.0900532901,
-0.0213343035,
0.0597465932,
-0.0070484946,
0.0210902356,
0.0785609782,
0.0291018803,
0.0021887408,
-0.0471138656,
-0.0276590232,
-0.0608068034,
0.0274075773,
0.002809816,
0.0736971796,
0.033588931,
0.0209131837,
-0.0593933165,
0.0490340702,
-0.0249987096,
0.0146027124,
0.0325885378,
0.0642353743,
-0.0228511095,
-0.0032399897,
-0.0777668655,
0.039860189,
0.0124601545,
0.0170279797,
-0.0022599043,
-0.0074613211,
-0.0081212753,
0.0213821959,
0.0515851267,
0.008649271,
-0.0023882273,
0.0187823232,
0.0154674146,
0.0117127839,
-0.0505523831,
-0.0149297472,
0.0549816489,
-0.0430196412,
-0.0037678131,
0.0214720126,
0.0190748554,
0.0298889447,
-0.0453357659,
0.0210139658,
0.0557602793,
0.003704983,
-0.0136666493,
-0.0226942804,
0.0697739348,
-0.01169906,
-0.0144479033,
-0.0027379892,
0.0296370909,
0.0345452912,
-0.0195943732,
0.0728252158,
0.0080552204,
-0.0663328469,
-0.0215751175,
-0.0227397885,
0.0063185105,
0.0217773765,
-0.035988722,
-0.0364395939,
-0.0353107899,
0.0286557283,
0.0018701351,
-0.0382518284,
-0.0350461006,
0.0112827718,
-0.0255580321,
-0.0172555409,
-0.010377625,
-0.0201807935,
0.0355339088,
0.0510450937,
0.010664287,
0.0169376601,
-0.034970738,
0.028003566,
0.0459843352,
-0.0162509233,
0.0219703317,
-0.0069295205,
-0.0126646897,
-0.0039165928,
-0.0109644402,
0.0120772012,
-0.0014909047,
-0.0833520591,
-0.0202282853,
-0.0161177199,
0.0367869921,
0.0273842756,
0.0303223841,
0.0093340157,
-0.0061509828,
-0.0151485745,
-0.0289302748,
0.0212438274,
0.033845108,
0.0405131355,
0.0298395567,
-0.0445482209,
-0.006384172,
0.0160579197,
0.0360889435,
-0.0567627922,
-0.0995030776,
0.0018627378,
0.0158046875,
-0.0077571813,
0.0506577455,
-0.0239998177,
0.0377366096,
-0.012379365,
0.0010681222,
-0.0288622119,
-0.0033029919,
0.0183056723,
0.0868586972,
0.0065175076,
-0.0034044639,
-0.0345240943,
0.0315920599,
-0.0138315819,
0.0100948652,
-0.0462627448,
-0.0505976863,
-0.0253841784,
-0.0106530003,
-0.0340780057,
0.0275898036,
-0.0498435795,
-0.0174255185,
-0.0598607175,
-0.0120752109,
-0.0066649639,
-0.0069382377,
-0.0110516744,
-0.0334052853,
0.0546389706,
0.0317204744,
0.0333309695,
-0.0559714884,
-0.0268124528,
0.0775597692,
-0.0853302553,
0.0124261566,
-0.0132461581,
-6.249209353e-33,
0.0201067291,
-0.0393891595,
0.0295326859,
-0.0417574719,
0.0497548617,
0.0188532323,
-0.0129997479,
0.0014491515,
-0.0262167193,
0.0254905093,
-0.0165170208,
-0.0233140122,
-0.0053235609,
0.0401188731,
0.0106341094,
-0.0825280324,
0.0136180855,
-0.0041678199,
0.0155181019,
0.0493684486,
-0.0276797321,
0.0085198469,
-0.0023468158,
-0.0530972667,
0.0129020531,
-0.0152108707,
0.0281554647,
0.0069021052,
-0.0269855112,
0.0446577966,
-0.0100415973,
0.0447257534,
0.0363064148,
0.0227737669,
0.0058992286,
0.0482665114,
-0.0528367162,
-0.0211310275,
-0.0017408453,
-0.0453193262,
-0.0774848163,
0.0134996157,
0.0511252545,
0.0030730276,
0.0158434324,
0.0166542064,
0.0315342434,
-0.014180718,
0.0163674764,
-0.0423186645,
0.0251143649,
0.0209305249,
-0.025797382,
0.0012341157,
-0.0553861111,
0.0920197368,
-0.0480585881,
0.0031545917,
-0.0330992118,
-0.0138700567,
0.0359547101,
-0.034674298,
-0.0023422542,
-0.009504457,
-0.028862603,
-0.0150898518,
0.0180380866,
-0.0368588381,
-0.0432922356,
0.023524249,
-0.0943249464,
0.0089034867,
0.1025288105,
0.057647571,
-0.0024777027,
0.0083477478,
-0.0739636123,
0.0597635955,
-0.0828887895,
0.0594862886,
0.0299641006,
-0.0268762689,
-0.0536349416,
0.0107571604,
-0.0415673703,
0.0316175595,
-0.0399022326,
-0.0448361747,
0.0502603874,
-0.0294271447,
-0.0240340494,
-0.0054858052,
-0.0309966076,
0.0159010515,
0.0393026173,
-0.031712614,
0.0179131683,
0.0093853185,
0.0071390597,
-0.0304944906,
-0.1114260331,
0.0826929063,
0.0088092266,
-0.0193559453,
0.0317205526,
0.0101728039,
-0.0204988103,
-0.0498928502,
-0.0255982187,
-0.0011868388,
0.0373581685,
0.0693636984,
-0.04198201,
-0.0068647177,
-0.0031047983,
0.0529954694,
-0.0054415753,
0.075127326,
-0.014377228,
0.0221162532,
-0.0456732996,
-0.0507257357,
-0.0170369986,
0.0203393772,
-0.0285468008,
0.0189964157,
-0.0422280766,
0.060635604,
0.0012224889,
-0.0368915126,
0.019409148,
0.0023297388,
2.955e-7,
0.1058550403,
0.0008790758,
0.0240515824,
-0.0149911484,
0.0315714888,
0.0050891894,
0.0475331321,
-0.0292264093,
0.0429519452,
-0.0082366429,
0.0395785011,
-0.0151498057,
-0.0096836146,
0.0633455887,
-0.0069614458,
-0.0841203481,
-0.0754336491,
-0.0028913375,
-0.0272398461,
-0.0271638408,
0.0539505444,
0.0078333039,
-0.0606823936,
-0.029058082,
-0.0127317263,
0.0083858082,
0.0232577473,
-0.0188507158,
0.0269135851,
0.0281990767,
0.0359481536,
-0.0118527636,
0.0331356488,
0.0455809571,
-0.0043853656,
-0.0326167904,
-0.0255709197,
0.0215831809,
-0.0294555966,
-0.0424311534,
0.0016883609,
0.0044851275,
0.0713031068,
-0.0676994249,
0.0625534132,
-0.0566315763,
-0.0461579487,
-0.0002187149,
0.0064011863,
-0.0138294231,
0.0760177746,
-0.0368595682,
-0.0545207001,
0.0100244591,
0.0040519023,
0.0278833155,
0.0085269026,
-0.0128272455,
0.0296592657,
0.0413161255,
0.0098464703,
-0.0316911601,
0.0364583507,
-0.0755996257,
-0.0612360165,
-0.0408670232,
0.0003255478,
2.090315023e-34,
-0.0084140208,
0.0204985291,
0.0017466644,
0.024795251,
0.0236578193,
0.0143825486,
0.0199698731,
-0.0153175415,
0.0293770619,
-0.0132676298,
0.0099174529
] |
http://arxiv.org/abs/2401.13444v1 | http://arxiv.org/pdf/2401.13444v1 | Clue-Guided Path Exploration: An Efficient Knowledge Base Question-Answering Framework with Low Computational Resource Consumption | 2024-01-24T13:36:50 | 2024-01-24T13:36:50 | Dehao Tao, Feng Huang, Yongfeng Huang, Minghu Jiang | cs.CL | cs.CL, cs.AI | In recent times, large language models (LLMs) have showcased remarkable
capabilities. However, updating their knowledge poses challenges, potentially
leading to inaccuracies when confronted with unfamiliar queries. While
integrating knowledge graphs with LLMs has been explored, existing approaches
treat LLMs as primary decision-makers, imposing high demands on their
capabilities. This is particularly unsuitable for LLMs with lower computational
costs and relatively poorer performance. In this paper, we introduce a
Clue-Guided Path Exploration framework (CGPE) that efficiently merges a
knowledge base with an LLM, placing less stringent requirements on the model's
capabilities. Inspired by the method humans use to manually retrieve knowledge,
CGPE employs information from the question as clues to systematically explore
the required knowledge path within the knowledge base. Experiments on
open-source datasets reveal that CGPE outperforms previous methods and is
highly applicable to LLMs with fewer parameters. In some instances, even
ChatGLM3, with its 6 billion parameters, can rival the performance of GPT-4.
Furthermore, the results indicate a minimal invocation frequency of CGPE on
LLMs, suggesting reduced computational overhead. For organizations and
individuals facing constraints in computational resources, our research offers
significant practical value. | Clue-Guided Path Exploration: An Efficient Knowledge Base Question-Answering Framework with Low Computational Resource Consumption: In recent times, large language models (LLMs) have showcased remarkable
capabilities. However, updating their knowledge poses challenges, potentially
leading to inaccuracies when confronted with unfamiliar queries. While
integrating knowledge graphs with LLMs has been explored, existing approaches
treat LLMs as primary decision-makers, imposing high demands on their
capabilities. This is particularly unsuitable for LLMs with lower computational
costs and relatively poorer performance. In this paper, we introduce a
Clue-Guided Path Exploration framework (CGPE) that efficiently merges a
knowledge base with an LLM, placing less stringent requirements on the model's
capabilities. Inspired by the method humans use to manually retrieve knowledge,
CGPE employs information from the question as clues to systematically explore
the required knowledge path within the knowledge base. Experiments on
open-source datasets reveal that CGPE outperforms previous methods and is
highly applicable to LLMs with fewer parameters. In some instances, even
ChatGLM3, with its 6 billion parameters, can rival the performance of GPT-4.
Furthermore, the results indicate a minimal invocation frequency of CGPE on
LLMs, suggesting reduced computational overhead. For organizations and
individuals facing constraints in computational resources, our research offers
significant practical value. | [
0.090243876,
0.0104446551,
0.0065394416,
0.0598783605,
-0.0551373698,
0.0054908344,
-0.0442558676,
-0.0108624808,
0.0513786264,
-0.0299603511,
0.015882289,
0.0234621651,
-0.041442968,
0.0095962575,
-0.0119631188,
-0.012586345,
0.0240030419,
-0.0570130199,
0.0080494126,
-0.0373832248,
-0.0051305881,
0.0402317978,
0.0012034653,
0.0108916182,
0.0375158861,
-0.0031517774,
-0.0397696532,
-0.001853906,
-0.0239415802,
-0.011737166,
0.0086754197,
0.0523230694,
0.0283267517,
0.0578408167,
0.0000023632,
-0.0540782101,
-0.0205620248,
-0.0037618098,
-0.0758406669,
0.013229508,
-0.0158944242,
0.0439083949,
0.0300240666,
0.0402723961,
-0.0458448492,
-0.0084481779,
0.0245767273,
0.0129977968,
0.0536797941,
0.0263031386,
0.0145995002,
0.0110568041,
-0.0167650972,
-0.0084393565,
0.0393264368,
-0.0814683661,
0.053736113,
0.014699595,
0.0200160481,
0.0250706207,
-0.0035522252,
0.0460401289,
-0.0128432596,
0.0139386104,
0.0072739897,
0.0383105725,
-0.019391166,
0.0127117336,
-0.0174337681,
0.0879984349,
0.0202573016,
-0.0058624074,
-0.0037381749,
0.0131489243,
0.0084537631,
0.0494594648,
-0.047709208,
-0.0105313091,
-0.0419789143,
-0.0192633588,
0.0158610903,
-0.0482605435,
0.0261980947,
0.0349104889,
-0.0162030291,
0.092527397,
0.0224367026,
-0.0228627436,
0.0207154453,
-0.0141342804,
-0.0348606855,
-0.0588580668,
0.0005976395,
0.0569549538,
-0.0175843202,
0.0607104972,
-0.0225796085,
0.057212159,
0.0242055245,
-0.0618959703,
0.0059259082,
0.0324354619,
0.0102175511,
0.1151703596,
-0.0018888218,
0.02060063,
-0.0112818461,
-0.0128969457,
-0.0595810451,
-0.0644334629,
-0.0535576902,
-0.0297344066,
-0.0568983853,
0.0402053483,
-0.0448174365,
-0.0362041406,
0.0107391393,
0.0570427738,
0.0185558386,
-0.0015952783,
0.0403139889,
0.0274268016,
0.0456278548,
-0.0089744832,
-0.0851129293,
0.0153981242,
-0.0487263538,
-0.0089370254,
-0.0118082706,
-0.0358943008,
0.0012873762,
-0.0116880704,
0.0264287256,
-0.0191404838,
0.0405471846,
0.0394593664,
-0.0017042601,
-0.0206456222,
-0.048923593,
-0.0373077691,
0.0213572253,
-0.0411401428,
-0.0141682019,
-0.0888036713,
0.0737181529,
0.03515907,
-0.0343837105,
-0.0278929695,
0.0259253476,
0.0426967144,
-0.0726010874,
0.044600673,
-0.0447697788,
0.0309616774,
0.0556757934,
0.0194411054,
-0.0163198169,
0.0646925345,
-0.0100856032,
0.0135646798,
0.0488070212,
0.0112241497,
-0.0171533041,
0.0118019311,
0.0342203714,
0.0298865754,
0.0564814322,
0.0666823387,
-0.0101508461,
0.0261521265,
-0.0056102015,
0.0478448048,
-0.0234711226,
-0.0181552395,
0.028912371,
0.0549952984,
-0.0024333044,
0.0144089432,
0.0155735789,
0.0424796976,
0.0067233075,
-0.0452637747,
-0.0279115736,
0.0269525126,
-0.0803387463,
0.0053864196,
0.0321787484,
0.0164993722,
-0.0263316687,
-0.0021875452,
-0.0092830891,
0.034621425,
0.0099016447,
0.053407155,
0.0415926538,
-0.016126018,
-0.0414159596,
0.0723684058,
0.0242238231,
-0.0126207545,
-0.0327130444,
-0.0231998544,
0.0263864063,
0.0981615558,
0.0891181156,
-0.0217460636,
-0.0429354347,
0.0324620828,
0.0281675719,
0.0098924898,
0.0309443288,
-0.0221458469,
-0.0014083562,
-0.0437160209,
-0.0392939746,
0.0012427712,
-0.0215274151,
0.0147956526,
0.0656494573,
-0.0058412929,
-0.0117843142,
-0.0598079748,
0.0101314569,
-0.0032371127,
0.0009803864,
-0.0164188668,
-0.0007680962,
-0.0018290534,
-0.0171353146,
0.0303921513,
0.0260639004,
0.0245609302,
0.0382697843,
-0.012159925,
-0.0653844625,
-0.0241220053,
-0.0039414368,
-0.0488141626,
0.0409905091,
0.0261375327,
-0.0041725831,
-0.0219680872,
-0.0295913499,
-0.0375033766,
-0.0023377643,
-0.0291373916,
0.0878081396,
-0.0299601145,
0.0042533875,
0.0107384874,
-0.0435674451,
-0.0481472909,
0.0775254369,
-0.0259606354,
0.035143245,
0.0054996856,
0.0144319618,
0.020562904,
-0.0450282916,
-0.0274983793,
-0.0073791058,
0.0077961935,
0.011854548,
0.0045273229,
-0.0042312429,
-0.0461296029,
-0.0073699076,
-0.1057922244,
0.0233005863,
0.01336639,
0.0273992512,
0.0157022532,
0.0420632958,
-0.0064322478,
-0.0095346952,
-0.0386119224,
0.0173083823,
-0.0017517691,
0.026691379,
0.0094427895,
0.03693562,
-0.0065178946,
-0.0411013886,
-0.012356611,
0.0029955148,
0.0234544799,
-0.0098098796,
-0.0100692157,
-0.022740189,
0.0354165882,
-0.0639261156,
0.0289158653,
-0.0195410457,
-0.0045648701,
-0.0175968334,
-0.0514788851,
0.0262958426,
0.0052892892,
-0.010216576,
-0.0200921204,
0.0259893723,
-0.0438764878,
-0.0620384254,
0.0174728781,
-0.015740376,
-0.0212234873,
-0.0166426506,
-0.0141965589,
0.0019585309,
0.0161659792,
0.0205115937,
0.0005905045,
-0.0307174176,
0.0096566891,
0.002682203,
0.0167451948,
-0.0943433568,
0.0113248713,
0.0272227358,
0.0425955392,
-0.0298709031,
-0.0100395568,
0.078376554,
-0.0098052397,
-0.0005756496,
-0.0255424995,
-0.0100660305,
-0.0448422544,
0.0732402578,
-0.025550751,
0.035221085,
-0.008532905,
-0.0113078393,
-0.0209133253,
0.0063410373,
0.031746801,
-0.0579976253,
0.0446127281,
-0.0455125645,
0.0364863053,
0.0284606609,
-0.031987071,
-0.0291488022,
-0.0265314467,
-0.0538639948,
-0.0114091495,
0.0546444133,
-0.0252480097,
0.0098938718,
0.0108063268,
0.0235913303,
0.0149342716,
0.0167118628,
-0.0203624442,
0.0065609841,
0.0243089143,
0.0623444095,
0.0154930241,
0.0196606163,
0.023140585,
0.0432015993,
0.0743776336,
0.0263557713,
-0.0503514372,
0.0100759352,
-0.0238998961,
0.0209401846,
-0.0459112264,
-0.0075274706,
-0.0011054657,
-0.0072128377,
-0.0334679149,
-0.0056786537,
0.110829398,
0.0480478965,
-0.0269387178,
-0.032009054,
-0.0277817324,
0.0220391266,
0.0281026885,
0.0582782589,
-0.0895790756,
-0.0484415926,
-0.0031442663,
0.0281372517,
0.0014584975,
-0.0601464994,
-0.0094710495,
-0.1237037182,
0.0138156628,
-0.0343884118,
0.010151566,
-0.0184711795,
-0.0505405925,
-0.0102345245,
0.0003169035,
-0.0574344397,
0.0477158427,
-0.0223943628,
-0.0033175913,
-0.0230663624,
-0.0249523744,
-0.0232289638,
0.05631008,
-0.0252668411,
0.0207117088,
0.0170368161,
-0.0000665701,
0.0330475718,
0.0105381738,
-0.0145344874,
-0.0124670025,
0.009094974,
-0.0461288802,
0.0318986923,
-0.0092289019,
0.0063500176,
-0.0398819186,
0.0004183798,
-0.0623843856,
0.0122087924,
0.0523343943,
0.1018826663,
-0.0090442039,
0.0277097318,
0.0072898092,
-0.0109405182,
0.0043269526,
0.1106089428,
-0.0029603634,
-0.0052092071,
-0.0149788326,
-0.0285396762,
-0.0293809883,
-0.0046500298,
-0.003476189,
0.023973288,
-0.0063081575,
0.0392852053,
-0.0639709756,
-0.0083786538,
0.0854732841,
0.0073491484,
-0.0232495274,
0.0499920063,
-0.0147704538,
0.0161567722,
-0.0704284087,
0.0038387256,
-0.0545138866,
0.0088143721,
-0.0557330772,
-0.0272438265,
0.0278712027,
-0.0250107665,
-0.0173072871,
0.0384467915,
-0.069908306,
-0.0563947,
-0.0250044391,
0.0164020639,
0.0214476474,
0.0247860458,
-0.045822829,
-0.020938633,
0.0480648503,
0.0162166432,
-0.0156521723,
0.0173426177,
-0.0171437487,
0.0343378186,
-0.0013279777,
0.0332517698,
-0.086176306,
-0.0387292616,
-0.0250346493,
0.0637129843,
-0.0649051666,
-0.0511541218,
0.0315574408,
0.0295302495,
-0.017785484,
0.0220699254,
0.0149725741,
-0.0362247191,
0.0215911716,
-0.0089548985,
-0.0417330489,
-0.0378040895,
-0.0008385053,
0.0013081534,
-0.0604412034,
0.0311257169,
-0.0033058668,
-0.0032365802,
0.0105116433,
-0.0498522185,
-0.0270127282,
0.0342848189,
-0.0109780626,
0.0529125407,
-0.0161843449,
-0.001897661,
0.0139328726,
-0.012641794,
0.0208988898,
-0.0265113655,
0.0573151223,
-0.0093896938,
-0.047716137,
-0.0238924045,
0.0080233747,
-0.0551063865,
-0.1210764796,
-0.0001217822,
0.048629161,
0.0728787407,
0.0352791809,
0.0053440956,
-0.0442321412,
-0.0356679335,
-0.0108960951,
-0.0313987695,
0.0057769627,
-0.029252369,
0.0841620937,
0.0610444769,
0.00322689,
0.0227056164,
0.0731721073,
0.0218553282,
0.0722006336,
-0.0011160523,
-0.059103217,
-0.0391344577,
-0.0356326252,
0.0170443822,
0.062079031,
-0.0326460712,
0.0337340757,
-0.0530176051,
-0.045294933,
-0.0092871496,
-0.0105796289,
-0.0314593799,
-0.0210957397,
0.0769199505,
0.0567899942,
0.035704758,
-0.0385836326,
-0.0430126376,
0.0364212878,
-0.0847231895,
-0.0332077369,
0.0081196791,
-7.039380329e-33,
-0.0132465009,
-0.0186474714,
0.0573260859,
-0.0053421995,
0.0060122609,
0.0176585931,
-0.0335160606,
0.0012806025,
-0.0413719043,
-0.0007180605,
-0.0130975591,
0.00668588,
-0.0050088149,
-0.0002963372,
0.0001535571,
-0.0248903316,
-0.0030748711,
0.0258550197,
0.00940431,
0.0098383334,
0.0716298446,
0.0371838249,
0.0182911139,
-0.0227099322,
-0.015528148,
-0.0220464785,
-0.0024062709,
0.0611692108,
-0.002706825,
0.0599903613,
-0.0244790558,
0.021553304,
0.0227801856,
-0.0192791671,
0.0099873217,
-0.0219643954,
-0.0739945322,
-0.02138911,
0.0431824364,
-0.0214865841,
-0.0279489923,
0.0016504278,
-0.0000229308,
-0.000355762,
0.0455205888,
0.0232084636,
-0.0067119845,
-0.0091811735,
-0.0405544266,
-0.0182198044,
-0.011400505,
0.0211953241,
0.0163858384,
-0.0452399924,
-0.0206620786,
-0.0179811232,
-0.0040834006,
-0.0126397293,
-0.0276582111,
0.0196445137,
0.030003529,
0.0317346305,
0.0375394635,
0.0238508061,
0.0292378459,
0.0188668054,
-0.0192264058,
-0.0355025791,
-0.0174033716,
0.0277808886,
-0.0213773511,
0.0396932103,
-0.0033872866,
0.0278655551,
-0.0008256455,
0.016388448,
-0.0401139259,
0.0327064209,
0.03510933,
0.0143521195,
0.0058174445,
0.02884911,
-0.0453952849,
-0.0585907884,
-0.0377137735,
0.0314172283,
-0.0502892323,
-0.0225676838,
0.0333215743,
0.0035495204,
-0.0351195037,
-0.036810331,
0.0101401675,
-0.0192950759,
0.0093068499,
0.0597843006,
0.003732288,
0.0030213865,
-0.0005647399,
0.0273188297,
-0.0776260495,
0.0011329136,
-0.0088384114,
-0.044203274,
-0.0363139771,
0.0370972864,
-0.0464402623,
0.006536766,
-0.0455373116,
-0.0041768523,
-0.0372014344,
0.0164469611,
-0.0391604006,
-0.0402231216,
0.0124512855,
0.010535487,
0.0116143357,
0.0337586738,
0.0409473144,
0.0314043239,
-0.0312543064,
-0.0082714753,
0.0016052302,
-0.0316569358,
0.0028476168,
0.0192912184,
-0.009056258,
-0.0175710227,
-0.0016267952,
-0.0381948426,
-0.0045566289,
0.0400095657,
3.11e-7,
0.0315822437,
0.0315953605,
-0.0015120178,
0.0608201846,
0.0106153684,
0.0059134853,
0.0286427159,
-0.0037040734,
0.0254724212,
-0.0406585857,
0.0121527556,
0.058182966,
0.0035216503,
0.0365244225,
-0.0872420222,
-0.0161865037,
-0.0071770111,
-0.0481675789,
-0.0636509508,
-0.0149346339,
0.0147981467,
0.0671610162,
0.0235546045,
-0.0029342123,
-0.0339117013,
-0.0594190732,
0.0549735613,
-0.0141656324,
-0.0146617983,
-0.0027977442,
0.0313348547,
0.0440970324,
-0.0122151962,
0.0283422843,
-0.0051017446,
0.0147468951,
-0.0078920182,
0.0549615361,
0.0144970687,
-0.0124308253,
-0.0161673129,
-0.0698693842,
0.0174511876,
-0.0534694903,
0.0883805677,
-0.0345988907,
-0.0017610897,
0.0182043035,
0.0173209645,
-0.0333571732,
-0.0207956135,
-0.0363760479,
-0.035622824,
-0.0222256947,
-0.0005556376,
-0.0193902515,
0.01028171,
-0.0656548217,
0.0090352707,
-0.0071768099,
-0.0394895896,
-0.0597229823,
-0.0353216492,
-0.0219021384,
-0.0232788436,
-0.0499728471,
0.0351720601,
2.915911772e-34,
0.0288491808,
0.0002089021,
-0.0061704963,
0.0286435913,
0.0493494123,
0.0054370309,
0.0031080495,
-0.0019992199,
0.0317763723,
-0.1175749302,
-0.0131695066
] |
http://arxiv.org/abs/2401.13398v1 | http://arxiv.org/pdf/2401.13398v1 | Text Categorization Can Enhance Domain-Agnostic Stopword Extraction | 2024-01-24T11:52:05 | 2024-01-24T11:52:05 | Houcemeddine Turki, Naome A. Etori, Mohamed Ali Hadj Taieb, Abdul-Hakeem Omotayo, Chris Chinenye Emezue, Mohamed Ben Aouicha, Ayodele Awokoya, Falalu Ibrahim Lawan, Doreen Nixdorf | cs.CL | cs.CL, cs.LG | This paper investigates the role of text categorization in streamlining
stopword extraction in natural language processing (NLP), specifically focusing
on nine African languages alongside French. By leveraging the MasakhaNEWS,
African Stopwords Project, and MasakhaPOS datasets, our findings emphasize that
text categorization effectively identifies domain-agnostic stopwords with over
80% detection success rate for most examined languages. Nevertheless,
linguistic variances result in lower detection rates for certain languages.
Interestingly, we find that while over 40% of stopwords are common across news
categories, less than 15% are unique to a single category. Uncommon stopwords
add depth to text but their classification as stopwords depends on context.
Therefore combining statistical and linguistic approaches creates comprehensive
stopword lists, highlighting the value of our hybrid method. This research
enhances NLP for African languages and underscores the importance of text
categorization in stopword extraction. | Text Categorization Can Enhance Domain-Agnostic Stopword Extraction: This paper investigates the role of text categorization in streamlining
stopword extraction in natural language processing (NLP), specifically focusing
on nine African languages alongside French. By leveraging the MasakhaNEWS,
African Stopwords Project, and MasakhaPOS datasets, our findings emphasize that
text categorization effectively identifies domain-agnostic stopwords with over
80% detection success rate for most examined languages. Nevertheless,
linguistic variances result in lower detection rates for certain languages.
Interestingly, we find that while over 40% of stopwords are common across news
categories, less than 15% are unique to a single category. Uncommon stopwords
add depth to text but their classification as stopwords depends on context.
Therefore combining statistical and linguistic approaches creates comprehensive
stopword lists, highlighting the value of our hybrid method. This research
enhances NLP for African languages and underscores the importance of text
categorization in stopword extraction. | [
0.0567824394,
0.0692339689,
-0.0204444584,
0.0479575917,
-0.0064157867,
0.0364440568,
0.0590578727,
0.0197343472,
-0.0099771675,
-0.0592654645,
0.0160260461,
-0.0398149863,
0.0352493711,
0.0216908921,
-0.0047625238,
-0.0532252155,
0.03616824,
0.0016091734,
0.0181842279,
0.0259316564,
0.052309569,
0.0335706472,
-0.0178824905,
0.0220958367,
0.0123479012,
0.0329029858,
0.0054207873,
-0.0279118232,
-0.0149192698,
0.021468997,
-0.0208688509,
0.0338146165,
-0.0475814268,
0.0677079782,
0.0000016679,
-0.0805730447,
-0.044639077,
-0.0206490494,
-0.0031474244,
-0.0280160271,
0.0610638335,
-0.0618055761,
-0.0475930125,
0.0111805266,
0.0105794072,
0.0337347835,
0.0747667849,
0.0107902074,
-0.0305355731,
0.0694464669,
-0.010099479,
-0.0312257577,
-0.0031699697,
-0.0118555101,
0.0005415041,
-0.01472425,
-0.0092569562,
-0.0552416109,
0.00841795,
-0.0427364334,
-0.0039874818,
0.0274883136,
-0.0354396068,
0.0113506289,
-0.0496624187,
0.0184410904,
-0.0641687363,
-0.0615257919,
0.037144348,
-0.0030267462,
-0.0510029234,
-0.0123864692,
0.0271554161,
-0.0057819984,
0.0223934129,
0.0323715359,
-0.0376060829,
0.0539318249,
-0.0525434539,
-0.0002434881,
0.0337078683,
0.038078852,
0.0264066104,
0.0145011479,
-0.0680764169,
0.0408783592,
0.0059974571,
-0.0183133651,
0.0574910007,
0.0114219924,
-0.0158995613,
-0.0307668764,
0.0384864211,
-0.043015562,
-0.0034551509,
-0.0151456892,
-0.0089149494,
0.0111624505,
0.0238514729,
-0.0587389171,
-0.0206026845,
-0.0042409552,
-0.0133638093,
-0.0131737143,
-0.0270762611,
-0.0514794812,
-0.016299786,
-0.0455424227,
-0.0027395112,
-0.0189109724,
-0.0395045429,
-0.0372088924,
-0.0155136576,
-0.0015656808,
-0.0224829949,
0.0048233559,
-0.0431397632,
-0.0290443394,
0.0075541665,
-0.0726724491,
-0.0746339858,
-0.0341413766,
-0.0003693165,
-0.0328172371,
0.0427031107,
0.0130400127,
-0.0461108275,
0.0445662215,
-0.0191810802,
-0.0672662705,
0.0418108515,
-0.013453397,
0.0317037255,
0.0145392232,
0.0197434593,
0.0209805556,
-0.044585783,
-0.0513740852,
-0.0831541121,
-0.0358467437,
-0.0217178855,
-0.0088105816,
-0.056276653,
-0.0684269667,
-0.0162860043,
0.021425819,
0.0531708188,
-0.0321258157,
0.0092024328,
0.0023278876,
-0.02667276,
0.0865582526,
0.0040341532,
-0.0404818952,
0.0572059304,
0.0258850455,
0.0502637886,
0.0896728411,
0.0202825759,
0.0192183387,
0.0779685676,
-0.0097324429,
-0.0121538267,
0.0367584154,
-0.0290811323,
-0.022328902,
-0.0352376588,
-0.0538757145,
-0.020223897,
0.0057954788,
0.0330744274,
0.0097751757,
0.0076090363,
0.0014514057,
-0.009373568,
0.0514589362,
0.0543679707,
0.0039633899,
0.0083213914,
0.0237857793,
-0.0311259944,
0.0055924663,
0.0384401791,
-0.0761023238,
-0.0366359986,
-0.0121909119,
0.0311133675,
0.0249712951,
-0.0130804339,
-0.0022631041,
0.0078020957,
0.0043791728,
0.0240436513,
-0.0276858062,
0.0529448986,
0.0051463936,
-0.0237109102,
-0.0026336042,
0.0281893723,
-0.0587818809,
0.0108466009,
-0.0219589137,
0.0190989077,
0.0196482278,
0.0087239482,
0.0074625202,
0.042194996,
-0.0021050829,
0.0411280021,
0.0777784958,
-0.026758546,
-0.0128246574,
0.0230276436,
-0.0536545664,
0.0210644323,
0.0716718659,
0.0274599269,
0.0032107723,
0.0252079573,
0.1112877801,
0.0197194014,
-0.0189640597,
-0.0161718149,
0.0248068888,
-0.0104319146,
0.0086036464,
0.0211569853,
-0.0003579211,
-0.013828449,
0.0097797578,
-0.0236224085,
0.011762483,
-0.0526644103,
-0.0457658134,
-0.0002224493,
0.0185715333,
0.0321355462,
0.0373637006,
-0.0024981063,
-0.0388812684,
0.0131904194,
0.0492786951,
0.0673302189,
-0.0176614411,
0.0340974033,
0.0061782571,
0.0438037701,
-0.0414930582,
0.0354175195,
-0.0360463299,
0.0603024624,
-0.0146397604,
0.0346326083,
0.0306106806,
0.0123471301,
-0.0121232653,
-0.0581507273,
0.0374369994,
0.0022076841,
0.0042942469,
0.0760757551,
-0.0187797304,
-0.0044267615,
0.0326619893,
0.009808803,
-0.0204090551,
-0.0218084585,
-0.1083354875,
0.012008342,
-0.0188636836,
0.0198592283,
-0.0327825025,
0.0610219017,
-0.0367609672,
0.0145448707,
0.0705209374,
-0.018130593,
0.0196710229,
0.0216521006,
-0.0006023615,
-0.0275941771,
-0.0360386334,
-0.0284577385,
-0.0051662349,
0.0257548206,
0.043774616,
0.0085081859,
-0.0887842774,
-0.0084239012,
0.0060925256,
0.0270251837,
0.0151087539,
0.0101214573,
-0.012227417,
0.0130221322,
-0.0005107427,
-0.0541367084,
0.0074341516,
0.0072521856,
0.0342333391,
-0.0091934344,
-0.0221630111,
-0.0034096658,
-0.0349956192,
-0.0118940426,
0.0396482274,
-0.0017479287,
-0.0534062944,
0.0457032882,
0.0770903602,
-0.0066237124,
0.0061099911,
-0.0196456686,
-0.0241291076,
-0.0278927367,
-0.0366272107,
-0.0363584943,
0.0193973202,
-0.0143563114,
0.0544730015,
0.0035773327,
-0.0320235752,
0.0069324942,
-0.0024212203,
-0.0078769233,
-0.0502613299,
-0.0131533211,
-0.0510911979,
0.0036149144,
-0.0624629483,
-0.0044633662,
0.0217054933,
-0.0832600668,
0.0246582255,
0.0589130707,
0.0505960025,
0.0171659011,
0.0175662581,
-0.0452976115,
-0.0169305392,
-0.0119280526,
-0.0141832894,
0.0004339265,
-0.071292758,
-0.0459373295,
-0.0125937993,
0.0368669182,
-0.0078919055,
-0.040565066,
0.0339303799,
0.0177131891,
0.0313824154,
-0.0092050694,
0.0074690063,
0.013759573,
0.05842641,
-0.0107033579,
-0.031180976,
-0.0060614152,
-0.0199980196,
0.0248754285,
0.0230478793,
-0.0317565501,
0.0093672965,
0.0424310975,
-0.0270964876,
0.0454404391,
-0.0510058217,
-0.05443592,
-0.0126353689,
-0.0114042135,
-0.0026172639,
-0.0177675448,
0.0994140282,
0.046183195,
0.0139663937,
-0.000684293,
-0.0110488441,
0.0169206038,
-0.0304605588,
-0.0232063681,
-0.0344260447,
-0.0435906574,
0.0328111611,
-0.0160081089,
-0.0157160312,
-0.0029257343,
0.0167410355,
-0.0712444708,
0.0018735706,
0.0095361313,
-0.0786734819,
-0.0162609443,
-0.0142841749,
-0.0166808814,
0.0192605089,
-0.0378104672,
0.0056629707,
-0.0333783887,
-0.0131074609,
0.0468253717,
-0.0398007557,
-0.0728985667,
-0.0338468961,
-0.0181067605,
-0.0053946376,
0.0382087678,
-0.0054865656,
-0.0774903372,
0.0319069289,
0.0616403334,
-0.0810025111,
-0.0426599607,
-0.0349680409,
0.0097597251,
0.005424046,
-0.016880149,
0.0277799144,
0.0003462187,
-0.0119737247,
0.0081268018,
0.0206029247,
0.0071121189,
0.0136204725,
-0.0467578731,
-0.0366041139,
0.0161553267,
-0.0092540216,
0.0126331281,
0.0169303101,
-0.0617234446,
-0.0194593612,
0.0011341355,
-0.0336575918,
-0.0499149114,
-0.0487749949,
0.0029206653,
-0.0065603461,
-0.0890399888,
-0.0440275185,
-0.0843458027,
0.0347987674,
0.0042328439,
0.0142227989,
0.0204547513,
-0.0129375681,
0.0188119281,
-0.0503082387,
-0.0270485729,
0.0225036014,
0.0018135909,
-0.0778327957,
0.0086065633,
-0.0179312974,
-0.0165962018,
-0.0130761685,
0.0025794522,
-0.0313262492,
-0.0358604491,
0.0361039191,
-0.0738981441,
0.0222069938,
-0.0235472862,
0.0150204012,
-0.0961975753,
0.0207022037,
0.0169494152,
-0.0004633034,
-0.0193270296,
0.0316702873,
-0.001762549,
0.0051982636,
0.018576622,
0.0216365457,
0.0413693376,
-0.0081492821,
-0.0105996057,
-0.0316521302,
0.0104999384,
-0.039571628,
-0.077530928,
0.015838271,
-0.0407188423,
-0.0270596743,
-0.0731053501,
-0.015368443,
0.0265419949,
0.0449462794,
0.0055213943,
-0.0069565233,
0.0381259099,
0.0061804429,
-0.036088068,
0.0335074067,
-0.0498763733,
-0.0217141062,
-0.0231793597,
0.0395329893,
0.0584088676,
0.0710160658,
-0.0028411339,
-0.000738751,
0.0202198066,
0.0301383175,
0.0019010168,
0.0433948524,
-0.0358746238,
-0.006325569,
0.0162099544,
0.0225317292,
-0.0189428478,
0.0344333537,
0.0092476327,
0.0513988473,
-0.0467841513,
0.037843205,
0.0470442921,
-0.0379417688,
0.0317724869,
0.0028891703,
-0.0011023827,
-0.0125123858,
0.0182612147,
-0.0073928591,
-0.0134859029,
-0.0242326073,
0.0090869563,
-0.0519455783,
0.017248055,
0.0452993326,
0.0603935942,
-0.0226642117,
0.0227362812,
-0.0366564617,
-0.0040260977,
0.0101699205,
0.0062901829,
0.0370876379,
-0.0049929605,
0.0459900498,
-0.0904080868,
-0.0258397982,
0.0522856638,
-0.0306032915,
0.033991769,
0.0409610346,
-0.0196105726,
0.053520672,
0.0218610913,
0.0471137278,
-0.0397280641,
0.0012953579,
0.0110573433,
-0.0414632112,
-0.0008659043,
-5.219330467e-33,
0.0219747219,
-0.0071414215,
-0.0238815118,
0.0066163898,
0.0115753412,
0.0183286909,
-0.0102738766,
0.0447691828,
-0.0024056607,
0.0305681657,
-0.0401657633,
0.0148669761,
0.0556679592,
0.0002430285,
0.0779247507,
0.0316369161,
-0.0037438308,
0.0435267352,
-0.0411899835,
-0.0083983373,
0.0582484826,
0.0753472298,
0.0803256184,
-0.043625053,
0.0591856167,
-0.0162234027,
-0.0113975136,
-0.0404862389,
0.0808773339,
0.0585463196,
-0.0296494458,
0.0351924188,
0.0361064002,
-0.0035139846,
0.0070920559,
0.0522290468,
-0.0494967811,
-0.0032843943,
0.0300383046,
0.0403222255,
-0.0004777101,
-0.0308820475,
0.0183735061,
0.0072353515,
-0.0057204785,
0.0494800657,
0.0150905559,
-0.026256429,
-0.020740984,
-0.0226961952,
-0.1115525067,
0.064498432,
0.0250815954,
0.0122773396,
0.0707714707,
0.0108086159,
-0.0321616493,
-0.0075144377,
0.0356130712,
0.0025125043,
0.053693261,
0.0710430741,
-0.0026943157,
-0.0011324866,
-0.0952889174,
0.0286495425,
0.016774185,
-0.0517690629,
0.0404679105,
0.0001170585,
0.0685056299,
-0.0377943553,
-0.0114574386,
0.0586818345,
0.0665126592,
-0.064951919,
-0.0295092687,
0.0211897381,
0.02190906,
-0.0066069355,
-0.0104772272,
-0.0222140774,
-0.0669685453,
-0.0388268977,
0.0199459326,
-0.0314144045,
-0.0234936997,
0.0332677886,
0.060949102,
0.0049755326,
-0.0149885807,
0.0528447144,
0.0417748615,
0.0330251232,
-0.0312103983,
-0.024467526,
0.0063362634,
0.037856847,
0.0120964069,
-0.0185220372,
-0.0420617573,
-0.0330161192,
-0.0147218946,
0.0222328156,
-0.0122843422,
-0.0488972962,
0.0177651271,
-0.0124617061,
-0.0670798421,
0.0235137921,
-0.0110120056,
0.0331494212,
-0.0112695564,
0.0700254515,
-0.0248924922,
-0.0289002415,
0.0020753457,
0.0032988582,
0.0319380425,
-0.0523346849,
-0.0146906376,
0.0546488389,
-0.0280771013,
0.0141046047,
-0.0535582267,
-0.009769422,
0.013169135,
0.0323446207,
0.0134003703,
-0.0222064927,
0.0022125903,
0.0043258667,
2.3e-7,
0.0120517313,
0.0146238618,
0.0386270247,
-0.0027075552,
0.0355458483,
0.018947361,
0.0352820829,
0.0116003798,
-0.0154162385,
0.0694918036,
-0.0160273593,
-0.0599640831,
0.0847500339,
0.0191608034,
-0.0845768899,
-0.0609059967,
0.0420699604,
0.0229190961,
-0.0574286953,
-0.0585619658,
-0.0216586646,
0.0954509452,
-0.0298788622,
-0.020775279,
-0.0504330397,
-0.0503391735,
0.0230159275,
-0.0215093754,
0.0350600556,
-0.0032583403,
0.0217141975,
0.0340488404,
-0.0030685531,
0.0032350782,
-0.0304807574,
0.0258076731,
0.0518797114,
0.0415112376,
-0.0092032515,
0.0456861183,
0.0220802035,
-0.0187016148,
-0.0492851213,
-0.0260142051,
0.0487640761,
0.0447662435,
-0.0106718941,
-0.0055844318,
0.0067244554,
-0.0270102099,
-0.0022487587,
-0.0338513069,
-0.0563473403,
0.0437912233,
0.0347935595,
-0.0113713471,
-0.0287695676,
-0.0430929027,
0.0068411217,
-0.0415043905,
-0.0120969713,
0.0451470315,
0.0234742444,
-0.001141021,
0.0117167188,
-0.008657855,
0.0174261425,
2.160295279e-34,
0.0271252971,
0.0331568234,
0.0138512449,
0.0585114658,
0.0056349034,
-0.0065242262,
0.0199374799,
0.0168649387,
-0.0166739542,
0.0030048208,
0.017689595
] |
http://arxiv.org/abs/2402.01698v1 | http://arxiv.org/pdf/2402.01698v1 | Large language model empowered participatory urban planning | 2024-01-24T10:50:01 | 2024-01-24T10:50:01 | Zhilun Zhou, Yuming Lin, Yong Li | cs.CL | cs.CL, cs.AI | Participatory urban planning is the mainstream of modern urban planning and
involves the active engagement of different stakeholders. However, the
traditional participatory paradigm encounters challenges in time and manpower,
while the generative planning tools fail to provide adjustable and inclusive
solutions. This research introduces an innovative urban planning approach
integrating Large Language Models (LLMs) within the participatory process. The
framework, based on the crafted LLM agent, consists of role-play, collaborative
generation, and feedback iteration, solving a community-level land-use task
catering to 1000 distinct interests. Empirical experiments in diverse urban
communities exhibit LLM's adaptability and effectiveness across varied planning
scenarios. The results were evaluated on four metrics, surpassing human experts
in satisfaction and inclusion, and rivaling state-of-the-art reinforcement
learning methods in service and ecology. Further analysis shows the advantage
of LLM agents in providing adjustable and inclusive solutions with natural
language reasoning and strong scalability. While implementing the recent
advancements in emulating human behavior for planning, this work envisions both
planners and citizens benefiting from low-cost, efficient LLM agents, which is
crucial for enhancing participation and realizing participatory urban planning. | Large language model empowered participatory urban planning: Participatory urban planning is the mainstream of modern urban planning and
involves the active engagement of different stakeholders. However, the
traditional participatory paradigm encounters challenges in time and manpower,
while the generative planning tools fail to provide adjustable and inclusive
solutions. This research introduces an innovative urban planning approach
integrating Large Language Models (LLMs) within the participatory process. The
framework, based on the crafted LLM agent, consists of role-play, collaborative
generation, and feedback iteration, solving a community-level land-use task
catering to 1000 distinct interests. Empirical experiments in diverse urban
communities exhibit LLM's adaptability and effectiveness across varied planning
scenarios. The results were evaluated on four metrics, surpassing human experts
in satisfaction and inclusion, and rivaling state-of-the-art reinforcement
learning methods in service and ecology. Further analysis shows the advantage
of LLM agents in providing adjustable and inclusive solutions with natural
language reasoning and strong scalability. While implementing the recent
advancements in emulating human behavior for planning, this work envisions both
planners and citizens benefiting from low-cost, efficient LLM agents, which is
crucial for enhancing participation and realizing participatory urban planning. | [
0.0510990731,
0.0724954307,
-0.0191494115,
0.054351043,
-0.0399391763,
-0.006244251,
-0.0138592739,
-0.0335037075,
-0.0008397237,
-0.0006246269,
0.0333333239,
0.0336428173,
-0.0431227349,
0.0604573265,
-0.0256913658,
-0.0490044057,
0.0012490592,
-0.0175889675,
0.0065065883,
-0.0113493111,
0.007062498,
0.0016981439,
-0.0122218896,
0.0599433295,
-0.019172946,
0.0588942282,
-0.05446155,
-0.0131039778,
0.0181459505,
-0.0338991024,
0.0045975652,
0.0142403953,
0.0548655577,
0.0355942808,
0.0000022145,
-0.0663975999,
-0.0350877084,
-0.0017456409,
0.0295460094,
-0.0617678948,
0.0494315252,
0.0098539433,
-0.0120187225,
-0.0059398538,
-0.0584756173,
0.0154632209,
0.0191634316,
0.0319817588,
0.0516086556,
0.0166083872,
0.0128664915,
-0.0681337044,
-0.0035518517,
0.0150162354,
0.0119234053,
0.017352799,
0.0151636973,
-0.0192367975,
0.0900769755,
-0.005519853,
-0.0107059265,
-0.0363372415,
0.0157341585,
-0.0440708101,
-0.0344521292,
0.023603579,
-0.0080505908,
-0.0318206325,
0.0052694669,
0.0439428911,
0.0384475067,
0.0314758718,
0.0485262834,
-0.0097721666,
-0.0003021529,
0.0587092824,
-0.0406847186,
0.01533785,
0.0151821151,
0.0349838957,
-0.0421512499,
-0.0093081212,
0.0141722495,
0.0399654843,
-0.001652822,
0.1296842992,
-0.0345539823,
0.0052676885,
-0.0436401591,
0.0232150052,
0.0200347919,
-0.0327302963,
0.0004462567,
0.0700063631,
-0.0825959221,
0.0288567245,
0.0046038995,
0.0175277051,
0.0411233455,
-0.0443024859,
0.0478948131,
0.020548543,
0.0170047656,
0.0471559912,
0.0261844508,
0.0193977617,
0.0130675677,
0.0222390126,
-0.0164121091,
-0.0136534059,
-0.0640590191,
-0.0179355666,
-0.0581075288,
0.0494823903,
-0.0667653084,
-0.0113911713,
0.0085747922,
0.0404264703,
-0.023925174,
0.0575053282,
-0.032538116,
0.0493221916,
0.015582799,
-0.0036418075,
-0.0776937827,
-0.0002133791,
-0.0567366816,
0.0106624728,
-0.0111838505,
0.002133687,
0.0324891657,
-0.0240508914,
0.0007012113,
-0.0036827021,
-0.0237208605,
0.0656747445,
0.0238981526,
0.0217599254,
-0.0379277058,
-0.026710609,
0.0169347674,
-0.0658793151,
0.0202357136,
-0.0840818286,
-0.0196271669,
-0.0239206087,
-0.0188713875,
-0.0297694914,
0.0074346773,
0.0109264217,
0.0044612638,
0.0576412603,
-0.0467644222,
0.0271632932,
0.0666107386,
-0.0000742096,
0.0344145782,
0.0902744904,
0.0633690432,
0.0496341325,
0.0266102497,
0.0249975938,
0.0053081727,
-0.012828758,
-0.004012756,
0.013163642,
0.0321836621,
0.0253432095,
-0.036388725,
0.0385034494,
-0.0462863557,
0.0415794626,
0.0009242613,
0.0143497977,
0.0609010383,
0.0531044751,
-0.0040146746,
-0.0066414368,
0.0470946394,
0.1100672409,
-0.0370426476,
-0.0623775274,
-0.038541235,
0.0228089578,
-0.0075951344,
0.0125387656,
-0.0339504555,
0.033129178,
-0.0065457192,
-0.0395601504,
0.0265450273,
-0.0191700459,
0.0747274682,
-0.0045685037,
0.033549495,
-0.002700953,
-0.0161867496,
-0.0091164215,
0.0068629226,
-0.0006797121,
0.0264879297,
-0.071085684,
0.0463322364,
0.0381719917,
0.0012649789,
-0.0503093787,
0.0176085513,
-0.0190694854,
-0.0140068075,
0.0384219438,
0.0268941261,
-0.0086806305,
-0.0608465821,
-0.0463151075,
-0.0257977676,
0.0099464208,
0.0377100259,
0.0302887838,
0.0020563786,
0.0056156442,
0.0204215087,
-0.0001126693,
-0.0161143467,
0.0182399359,
-0.0203229692,
-0.0138231041,
-0.0414496548,
0.0062003839,
0.0252070408,
-0.0061102402,
-0.0301469769,
0.029689122,
0.0171020571,
0.0085712783,
-0.086064294,
0.0171490386,
-0.000504824,
0.0108060483,
0.002284307,
-0.0076936404,
-0.0090027601,
0.0358712226,
-0.04066718,
-0.008174886,
0.0088037932,
-0.0256147664,
0.0029378764,
-0.0464850776,
0.0069708917,
-0.0159503315,
-0.0382610559,
-0.0358762294,
0.0274554994,
-0.0074751331,
-0.0103371469,
0.016960986,
0.0280618034,
0.0053560436,
-0.0114524169,
-0.0277413856,
0.0209679436,
0.0023207213,
0.0126036843,
0.0100383535,
0.0487525761,
-0.0156184658,
-0.0531037934,
-0.0614130646,
-0.0437492877,
-0.0141289989,
0.0014961327,
-0.0270152353,
-0.0282770284,
-0.0128138838,
0.0161607321,
-0.021066308,
0.0352917612,
0.0121302325,
0.0082705496,
0.0056607141,
-0.0111476453,
-0.0005852701,
-0.0058308067,
0.0165700261,
-0.0197345261,
0.0197613724,
0.0553309806,
-0.0167987701,
0.0011779274,
-0.0400193892,
-0.0219948553,
-0.0036429365,
0.0219987798,
-0.0268255342,
0.038199041,
0.0203835461,
0.0047662156,
-0.027692033,
0.0273795687,
0.0408968739,
0.0241178572,
-0.0250438694,
-0.0879518539,
0.0320895053,
-0.0471269749,
-0.0175081193,
-0.0511314124,
-0.0229976792,
0.033884149,
0.0706061497,
-0.0384361893,
0.0100387,
-0.0381538942,
0.0659145638,
-0.0271437988,
-0.0420857407,
-0.014455976,
-0.0091098044,
0.0002880757,
0.0472640432,
-0.0478463545,
-0.0411198065,
0.0809181705,
0.0110950712,
-0.0139037808,
-0.0108535709,
0.0271840617,
-0.0781991333,
0.0972410589,
0.0279130023,
0.02267115,
-0.0069704442,
0.0037936587,
-0.0176731832,
-0.0530678444,
-0.016507661,
-0.0097718975,
0.0303057823,
-0.076009281,
-0.0088923024,
-0.0059951115,
-0.0700488836,
-0.0164048169,
-0.0217117015,
-0.0338849276,
0.0399627648,
0.0100875972,
-0.0072065531,
-0.0160306338,
0.0333698094,
0.0188906416,
0.017887488,
-0.0034163978,
-0.0363492891,
-0.0432464331,
-0.0305181909,
0.0465174913,
0.0884978324,
-0.0144025832,
0.0339012593,
0.0041984096,
0.0262505859,
0.0124765458,
-0.0182517357,
0.0005529165,
0.0275828,
-0.0047981879,
-0.0275586881,
0.0231440943,
0.0279836133,
0.00561067,
-0.0410586707,
-0.0143741965,
0.0799679086,
0.0132827042,
-0.0105286501,
0.0135713415,
-0.0922359303,
0.0256182607,
0.0344951339,
0.0270670168,
-0.0545487441,
0.0113948891,
0.0133977765,
-0.0187338106,
-0.0110208737,
-0.0373215079,
0.0299251042,
-0.129086107,
0.0815402567,
-0.0140725411,
-0.034658704,
0.0250151698,
-0.0523740537,
-0.0161310527,
0.0064480053,
-0.0555240251,
-0.0319036655,
0.028639935,
0.0047546434,
0.0365732014,
-0.0435406268,
0.0033953595,
-0.0262748245,
-0.0538292676,
0.0863522664,
0.1013547108,
0.0669470131,
0.0006150995,
0.0103334403,
-0.004050035,
0.0672436208,
0.0282578859,
-0.0193014871,
-0.0662041306,
0.0101526752,
0.0029373534,
0.0171924904,
0.0199849233,
-0.0272781998,
-0.039088577,
0.0784417838,
0.0785063431,
-0.0110414997,
0.0104958722,
-0.0524613895,
0.0346110165,
0.0552608334,
0.0024650889,
0.0062901052,
-0.0178291257,
-0.0380921699,
-0.0421388447,
-0.067973502,
-0.0298234057,
-0.0871082768,
0.0064757601,
0.0043461188,
0.0456743799,
0.001794374,
-0.0507877395,
0.0193735547,
0.0273222774,
0.0162860528,
-0.030420417,
0.0570263043,
-0.0115413675,
-0.0197236892,
0.0049836989,
-0.0018967108,
0.0386811681,
-0.0209503844,
-0.0447438881,
0.0147406617,
-0.0253591426,
-0.0208449773,
-0.0226203222,
-0.0159312356,
-0.0586514249,
0.0312117375,
0.0510501638,
0.0218775738,
-0.0132329371,
0.0157813579,
-0.0250417441,
0.0305338167,
-0.0243045352,
0.014311377,
-0.0201225504,
-0.0042410335,
0.0647666752,
0.0334777199,
0.0164295584,
-0.0098396139,
0.052819781,
0.0337749869,
0.0548257157,
-0.0795333982,
0.0346560255,
-0.0009290086,
0.0087945936,
0.0261725988,
-0.0237734672,
-0.0132782059,
-0.0147508159,
-0.0159519706,
-0.0262728371,
-0.0113856271,
-0.037027549,
-0.0099263405,
-0.0040755193,
0.0258511882,
0.0182562768,
0.0168940704,
0.0075579002,
0.0037901723,
0.0233082175,
0.0241361093,
0.0272419285,
-0.0126634687,
0.0289689247,
0.0303632095,
-0.0056811459,
0.0067516495,
-0.0021269009,
0.0071822032,
0.0338540524,
0.0140236141,
0.0170850214,
-0.006980367,
-0.0159690659,
-0.0059260847,
0.0295291096,
-0.0978665352,
0.0087608323,
0.02409973,
0.0187088531,
0.025527278,
-0.0195557978,
-0.0612640716,
0.0193169564,
-0.024518799,
-0.0143717621,
0.0230221879,
-0.0654000416,
0.0342871919,
0.0022680943,
-0.0537981652,
-0.0465232395,
0.0299274214,
-0.0023642571,
0.0318599194,
-0.0332880914,
-0.0317728482,
-0.0153879616,
-0.0460301153,
-0.0106657455,
0.0234081186,
-0.0192395244,
-0.0065887501,
-0.0710469633,
-0.0092495764,
-0.0204432644,
0.0565666854,
-0.011376447,
0.0048340415,
0.0346706249,
-0.0265425742,
-0.0102235125,
0.007269348,
-0.0635402948,
-0.0063715014,
-0.0661684573,
-0.0110365227,
0.0183815416,
-5.848096518e-33,
-0.0185873676,
-0.0743836835,
0.0911746547,
-0.0299170725,
-0.0155925285,
-0.0496638641,
-0.0417919867,
0.0193416327,
0.0416060574,
0.0292460583,
0.0186635945,
-0.0385418683,
0.0274908971,
0.0178641547,
0.0293802209,
0.0051387521,
0.0354868174,
0.0016950705,
-0.0069837668,
-0.0331523344,
0.0045081992,
-0.0087713217,
-0.0456381403,
-0.0524201989,
-0.0071907882,
0.0166166369,
-0.0209158268,
0.0288760848,
-0.0326921418,
0.0380865522,
0.0150111411,
0.0399647504,
0.0295230504,
-0.0253408663,
-0.0599475875,
0.0524480939,
-0.0440922193,
-0.0315552764,
0.0075912909,
-0.0163966659,
-0.0531773455,
-0.0227814373,
-0.0023384995,
0.0226821955,
-0.0099712452,
0.0407098532,
0.041435305,
-0.0207681563,
-0.0088019986,
-0.0118263615,
-0.1127857938,
0.0087039592,
-0.0574560687,
0.0118045695,
0.0009381629,
-0.0274021588,
0.0152360322,
-0.0128222629,
-0.0279812813,
0.0338084698,
0.0143373478,
-0.0350221246,
0.0070286291,
0.047012832,
0.0176291298,
0.0720419511,
0.0607110672,
0.0262156036,
-0.0504080206,
-0.0170538332,
0.0439277776,
0.0741125196,
-0.0069811246,
0.074903883,
-0.0276631359,
-0.0010079469,
-0.0086514521,
0.0260858648,
0.0090112248,
0.0188541785,
0.0548640862,
0.0165973064,
-0.0159304496,
-0.0241341665,
-0.0578655079,
-0.0121119339,
-0.0187565181,
-0.0405012071,
0.0219753291,
0.0177805889,
-0.0198030639,
0.0640174821,
0.0044010873,
-0.0616193041,
-0.0274332315,
0.001560615,
-0.0041204956,
-0.0032973927,
-0.0226527471,
0.0614358746,
-0.0853621736,
-0.0405516662,
-0.0474579558,
0.0127406847,
0.0094795506,
0.0392933488,
-0.0607704744,
0.0311690383,
-0.0292657446,
-0.0055876439,
-0.0221838839,
-0.0080042519,
0.0258033667,
-0.0164917894,
0.0065806713,
-0.0655911192,
-0.0095271124,
-0.0152696269,
0.0017497403,
-0.0923701748,
0.0243427828,
0.0222257357,
-0.0413594693,
0.0151668852,
-0.0028031997,
0.0159550607,
-0.0334260352,
0.0177523475,
0.0151774753,
0.0001577563,
0.025034301,
0.0016336102,
2.855e-7,
-0.0281307716,
0.0664031506,
-0.0258264784,
-0.044896882,
0.0421846136,
0.0382543765,
0.0014823413,
0.0016753057,
-0.0431620143,
0.0513329916,
0.0631543621,
0.0458366126,
0.0452114977,
0.0390422679,
-0.0911585614,
0.047347147,
-0.0295879394,
0.0029936393,
-0.1094778925,
0.0047404389,
0.0208117217,
0.0302436985,
0.0473701842,
0.006571522,
0.0032057534,
-0.0112698423,
-0.0055816104,
-0.0774769261,
-0.01413853,
0.0388225801,
0.005088361,
-0.040555913,
0.0289499331,
-0.0346650556,
-0.0055850851,
-0.0181629453,
-0.0266581923,
0.0411370732,
0.0089965379,
-0.0201983806,
0.0012053153,
-0.0747832954,
0.0124592157,
-0.0256166197,
0.0317138173,
-0.0564818457,
-0.0307714455,
0.0197125655,
0.022268977,
-0.044918146,
0.0141043644,
-0.0199791063,
-0.0169649441,
-0.0034232391,
0.003956818,
-0.0214529205,
0.0045832223,
-0.0601537675,
0.0529693179,
-0.0098850671,
-0.060452532,
-0.0476337261,
-0.0478432328,
0.0327425301,
-0.0228159074,
0.0001768888,
0.0003307717,
2.0052051e-34,
-0.0234144535,
0.0505488962,
-0.0195515156,
0.0057135066,
0.0543276034,
0.0498928651,
0.0201289728,
0.0048562777,
-0.0067680464,
-0.0404044092,
-0.0130202835
] |
http://arxiv.org/abs/2402.01697v4 | http://arxiv.org/pdf/2402.01697v4 | APT-Pipe: A Prompt-Tuning Tool for Social Data Annotation using ChatGPT | 2024-01-24T10:09:11 | 2024-02-20T07:54:12 | Yiming Zhu, Zhizhuo Yin, Gareth Tyson, Ehsan-Ul Haq, Lik-Hang Lee, Pan Hui | cs.CL | cs.CL | Recent research has highlighted the potential of LLM applications, like
ChatGPT, for performing label annotation on social computing text. However, it
is already well known that performance hinges on the quality of the input
prompts. To address this, there has been a flurry of research into prompt
tuning -- techniques and guidelines that attempt to improve the quality of
prompts. Yet these largely rely on manual effort and prior knowledge of the
dataset being annotated. To address this limitation, we propose APT-Pipe, an
automated prompt-tuning pipeline. APT-Pipe aims to automatically tune prompts
to enhance ChatGPT's text classification performance on any given dataset. We
implement APT-Pipe and test it across twelve distinct text classification
datasets. We find that prompts tuned by APT-Pipe help ChatGPT achieve higher
weighted F1-score on nine out of twelve experimented datasets, with an
improvement of 7.01% on average. We further highlight APT-Pipe's flexibility as
a framework by showing how it can be extended to support additional tuning
mechanisms. | APT-Pipe: A Prompt-Tuning Tool for Social Data Annotation using ChatGPT: Recent research has highlighted the potential of LLM applications, like
ChatGPT, for performing label annotation on social computing text. However, it
is already well known that performance hinges on the quality of the input
prompts. To address this, there has been a flurry of research into prompt
tuning -- techniques and guidelines that attempt to improve the quality of
prompts. Yet these largely rely on manual effort and prior knowledge of the
dataset being annotated. To address this limitation, we propose APT-Pipe, an
automated prompt-tuning pipeline. APT-Pipe aims to automatically tune prompts
to enhance ChatGPT's text classification performance on any given dataset. We
implement APT-Pipe and test it across twelve distinct text classification
datasets. We find that prompts tuned by APT-Pipe help ChatGPT achieve higher
weighted F1-score on nine out of twelve experimented datasets, with an
improvement of 7.01% on average. We further highlight APT-Pipe's flexibility as
a framework by showing how it can be extended to support additional tuning
mechanisms. | [
0.049415037,
-0.0104149533,
-0.0174303837,
0.0056821746,
-0.0111691905,
0.0395301729,
0.0137082003,
0.0124368807,
0.0722384006,
-0.04788417,
-0.0139550669,
0.0165796801,
-0.0129792504,
0.0339094326,
-0.0091822781,
-0.0380262397,
0.0282697175,
0.0135386996,
0.025602378,
0.0183382425,
0.0259237476,
0.0146776484,
-0.0228437576,
0.0416554548,
-0.0210606549,
-0.0219262615,
0.0151781114,
0.032397639,
0.0069523561,
-0.0421431661,
0.0485742874,
0.0868496001,
-0.0420231186,
0.0638239235,
0.0000019652,
-0.0653103963,
-0.005378833,
-0.0155687146,
-0.0373155661,
-0.0036910251,
0.0361429639,
0.0405914187,
0.015973337,
0.0276040975,
-0.0327441506,
0.0136837065,
0.0534076355,
0.0272986088,
0.0526683293,
0.0701275989,
-0.0108575439,
0.0235499498,
-0.0122373756,
-0.013962578,
0.0465811044,
-0.0149958134,
0.0058491863,
-0.0162926931,
0.0473885164,
0.0779475272,
0.0221398808,
0.0544876494,
-0.0053109541,
-0.0258521345,
0.0544742569,
0.0119301761,
-0.0355811119,
-0.0347324535,
-0.0261109043,
0.0720162466,
0.048229713,
-0.0138429748,
0.0091094133,
-0.0333732069,
0.0556438416,
0.0077554709,
-0.0398038626,
-0.0161559079,
-0.0214301869,
-0.0119347041,
-0.0083654374,
0.0229535196,
0.0016264206,
-0.0066833696,
0.062882334,
0.0629647598,
-0.0236910582,
-0.0133550428,
0.023420725,
0.0347723588,
-0.0169974081,
-0.056618426,
0.034811981,
0.0104534952,
-0.0303642675,
0.0026182274,
-0.0370770507,
0.008847082,
0.0451600887,
-0.0389105193,
-0.0161987767,
0.0300253239,
-0.0306105968,
0.03911056,
0.0198647529,
-0.0017905049,
0.0388642959,
0.0406405032,
0.0181674939,
0.0075314115,
-0.0749991089,
0.0125941578,
-0.0287013873,
-0.0045406609,
0.0034474789,
0.017280329,
-0.0552872941,
0.0411929227,
-0.0088648889,
0.0238466486,
0.0150621515,
-0.0203688871,
0.0054478026,
0.0483031236,
-0.0224348158,
0.0025488711,
-0.0261944979,
-0.0048421579,
0.0095919296,
-0.0486836396,
0.0083610732,
0.007537839,
0.0480132177,
-0.0212742724,
0.0009434734,
0.0005296486,
0.0119800335,
-0.0520193726,
-0.0124691455,
-0.0432288907,
-0.0226207562,
-0.0129947839,
-0.0072516585,
-0.0416495763,
0.0379362553,
0.0502050668,
0.0044539543,
-0.0128558576,
-0.0137468958,
0.0558840334,
-0.0225765929,
0.0702317506,
-0.0644646063,
-0.0058247419,
0.0579052009,
0.0462974161,
-0.0492370613,
0.0573973842,
0.0010545627,
0.0669461265,
0.076909177,
-0.0267161801,
0.0261604674,
0.0556116253,
0.0013466836,
0.0004755768,
0.0500183739,
0.0071313055,
0.0009231942,
-0.0032147204,
0.0017864187,
0.0273483787,
0.00544815,
0.002854943,
0.0537890941,
0.0744122639,
0.0445524938,
-0.007512087,
0.0134395333,
0.06025365,
0.0186499842,
-0.0135739353,
-0.037995398,
0.0382520109,
-0.0583288483,
-0.0202049203,
0.0211416055,
0.00337016,
-0.0143679976,
-0.036532443,
-0.0232725013,
-0.0033052475,
0.0196279027,
0.0056853048,
0.0004955118,
0.0006428849,
-0.0160445627,
0.0269351229,
0.0392614156,
-0.0328959376,
-0.010766997,
-0.0367233828,
0.0363141596,
0.07212428,
0.0216852091,
-0.0545108952,
-0.0144887026,
-0.0540630147,
0.0658211336,
0.027530143,
0.06582167,
-0.0312544517,
-0.0522885583,
-0.0141532356,
-0.0521888845,
0.015268377,
-0.0183939375,
-0.00701688,
0.0435024388,
-0.0036958023,
-0.0010544475,
0.0178280734,
-0.0447250269,
0.0304010119,
0.0295312218,
-0.0177465845,
0.0003519245,
0.0051027359,
0.0193311535,
0.0562696159,
0.0314241834,
-0.0279644877,
0.0020408237,
0.0219725966,
-0.0298192296,
-0.0200333055,
-0.0063378466,
-0.0214056764,
0.0338754058,
0.0157730356,
-0.0445726998,
-0.013942901,
-0.0596273057,
-0.0215773806,
-0.0382208042,
-0.0349357873,
0.0292437784,
-0.0140057812,
-0.0192387216,
-0.0228268169,
0.0060741524,
-0.0374319889,
0.0418752693,
0.0137186646,
0.0114713544,
-0.0124808326,
0.0121964905,
-0.0018881913,
-0.0304884613,
-0.0259558875,
-0.0208510216,
-0.0011325151,
0.0050112749,
0.037466675,
0.0301806182,
-0.0724199191,
-0.0164628867,
-0.1101964489,
-0.0015540931,
-0.0199113302,
0.0114619657,
-0.0129625453,
0.0235003456,
-0.0301542804,
-0.0344980173,
-0.0360403769,
-0.0036568535,
-0.0128937932,
0.0435897335,
0.0550547801,
-0.0102789458,
0.0051670102,
-0.0144225284,
0.0222962238,
-0.0252599567,
0.0231559332,
0.0553869084,
-0.0838343948,
-0.0419901237,
-0.0081546586,
-0.0287891626,
-0.0358242281,
0.0098942397,
0.0185703896,
0.0097836582,
0.038271293,
0.0129798679,
-0.0028482105,
0.004138181,
0.0166292302,
-0.0045389403,
0.0442289822,
-0.0503550395,
-0.0006992653,
-0.0179027356,
-0.0020107578,
0.0104692318,
-0.0404995419,
0.0150121665,
0.0452846661,
-0.0179498307,
0.0144179435,
-0.0715242475,
0.0264240559,
-0.0181770492,
0.011150009,
-0.0702019855,
0.0043074898,
0.0014122268,
0.0298894811,
-0.0196153168,
-0.0510487258,
0.0606838278,
-0.0260553453,
-0.0162900053,
-0.0146977184,
-0.0527664386,
-0.0448967926,
0.060911987,
-0.0023162111,
-0.0070021907,
0.0251793973,
-0.0301218275,
0.0044892975,
0.0284930486,
0.0235603973,
-0.0456811562,
0.0195317119,
0.0184939094,
0.0073049786,
-0.0179285184,
-0.0391327962,
-0.0223182868,
-0.0740394071,
-0.0676762909,
-0.0126102911,
-0.0051773461,
-0.0218592416,
0.0295236073,
0.012095008,
-0.0141438646,
0.0486037619,
-0.0202458091,
-0.0343731493,
0.0029822628,
0.0075680339,
0.0202360693,
-0.0097036101,
-0.0344058648,
0.0036588761,
-0.0079597877,
0.0271050874,
-0.0028717201,
-0.0371461473,
-0.0078976927,
-0.0280964039,
0.0318309218,
-0.0559357665,
-0.021524461,
0.0011724321,
0.0702351853,
-0.0280142985,
-0.0001163477,
0.1164027303,
0.0393022187,
-0.0438839011,
-0.0008297717,
0.0058038831,
0.0593645647,
0.0307499021,
0.0185639895,
-0.1188983768,
-0.0522054471,
-0.0078675253,
0.045538716,
-0.0088020572,
-0.0314238369,
-0.0318559259,
-0.0886570364,
0.0409720391,
-0.0172812156,
-0.0325633995,
0.0133864973,
-0.0739949271,
0.0022353751,
-0.0345089398,
0.0000864648,
-0.0488170683,
-0.0234387182,
0.025692571,
0.0363719724,
-0.05616742,
-0.0197901111,
0.0269630104,
-0.0239967275,
0.0846202895,
0.0192227177,
0.0245090686,
0.0160283279,
-0.0362376384,
0.0651454479,
0.0177645348,
0.0048242155,
-0.0378353223,
0.0826727301,
-0.0193731003,
0.0187670123,
-0.0111931944,
0.0130342571,
-0.0122366184,
-0.0390164666,
0.027407838,
0.0504492857,
-0.0152491247,
0.0119891372,
-0.0135430628,
0.0253717676,
0.0170240086,
0.0916751474,
0.0594297983,
-0.0633470789,
0.004430138,
-0.0678913593,
-0.0020816238,
-0.0413430184,
-0.0344437957,
-0.0021632407,
-0.0432386436,
-0.0632770732,
-0.0356218219,
-0.0336370394,
0.0432825908,
0.0002411868,
0.0020004632,
-0.0114344936,
-0.0236965027,
0.0458425656,
-0.0543556213,
0.0559320301,
-0.0403533205,
0.0291427132,
-0.018306898,
-0.0398055241,
0.0231135096,
-0.0230293106,
-0.0176070761,
0.0112565383,
-0.070247151,
-0.0396023877,
0.0149875209,
0.0175221581,
0.0209301282,
0.0246344432,
-0.0332872048,
-0.0940784737,
0.0392310657,
0.1017801538,
-0.0256922599,
-0.0277475491,
0.0131497569,
0.0571382828,
-0.0187203307,
0.015859019,
-0.0400056951,
0.0008745226,
0.0116558624,
-0.0021812096,
-0.0316472687,
-0.0411655419,
-0.0200509727,
0.0388316959,
0.0641810372,
-0.0178802349,
-0.0693740547,
-0.0386750065,
-0.020425301,
0.0224134829,
-0.0322829336,
-0.0082602268,
-0.0003638272,
-0.0260836035,
-0.0607638843,
-0.0488726646,
-0.0051463875,
-0.0335977934,
0.0723656192,
-0.1040296182,
0.0318057239,
0.0149542633,
0.0293983091,
0.0029543405,
-0.0134464512,
-0.0288885012,
0.0226597488,
0.0495143086,
0.0317777693,
-0.0014758032,
0.0469502732,
0.0030126544,
0.0090748891,
0.0292647965,
0.0510535426,
-0.0648478642,
-0.005971089,
-0.0105624991,
0.0799054429,
0.0255223457,
0.0050838511,
0.0369012989,
-0.0030895302,
-0.0364382043,
0.0023238994,
0.0182665102,
-0.003600718,
-0.0362009555,
0.0115712266,
-0.0179041922,
-0.0429315977,
-0.000250125,
0.0374937803,
-0.059333425,
0.0122555019,
0.0356262401,
-0.0545847863,
-0.0512646809,
-0.0125540616,
-0.0144434385,
0.0557210855,
-0.0031067447,
-0.0130209802,
-0.0503794812,
-0.0550634786,
0.0233270191,
-0.0247158222,
-0.0174066834,
0.0262169391,
0.0328593776,
0.0397246331,
-0.015265502,
-0.0180738792,
-0.0361980274,
-0.019965848,
-0.0022809561,
-0.0218973476,
-0.0055153416,
-6.32236257e-33,
-0.0218199901,
-0.0537225045,
-0.0081807887,
0.0887144804,
-0.0004466051,
0.0131027075,
-0.0167461373,
0.0226095747,
0.024745455,
-0.0065975087,
-0.044133421,
0.0031799683,
0.0172184017,
0.0087819137,
0.0151937967,
-0.0635766909,
0.0366729572,
-0.0012776505,
-0.0479413308,
0.0660820156,
0.034754537,
0.0261084456,
0.0177574344,
-0.090252392,
0.0806880817,
0.0073758294,
-0.0186059326,
-0.0010920576,
0.0390663445,
0.0507050343,
-0.0112045072,
0.0673028156,
0.0352670923,
-0.0229611415,
0.0124929342,
0.0737077892,
-0.0698351264,
-0.0451456606,
-0.0133871064,
0.0116370972,
-0.0164178163,
-0.0249533504,
0.0094136391,
-0.0242021848,
-0.0373561233,
-0.0025409067,
0.0679330677,
-0.0060565281,
-0.004128858,
-0.0334990025,
-0.0234644786,
0.0469688512,
-0.0183690079,
-0.0059064142,
0.0114355646,
0.0578739569,
0.050942827,
0.0403323956,
-0.0559140928,
0.0195748638,
-0.0232788958,
-0.0098845186,
0.0380689129,
0.0162724834,
-0.0168477353,
0.0540983789,
0.0180006549,
0.0455758795,
-0.0000702612,
0.0034767897,
-0.0090591991,
0.0055579133,
0.0437027887,
0.0383400433,
0.0136129847,
-0.046480611,
-0.0283240471,
0.0603419803,
0.069673799,
-0.081381239,
-0.0162464064,
0.0025709725,
-0.0494208224,
-0.0291296784,
-0.0201285947,
-0.0259678736,
-0.0307842493,
-0.0148432702,
0.0026003758,
-0.0104770223,
-0.0500513576,
0.041359324,
-0.0131542394,
-0.0237437207,
-0.0446603037,
-0.0309609249,
-0.0066098669,
0.0438017808,
-0.0035816401,
-0.0260885321,
-0.0231856871,
-0.0173729584,
0.0247134119,
0.0107340952,
0.0422920436,
-0.0073417081,
0.0872360021,
0.0172832198,
-0.0825943574,
-0.0116391769,
-0.0036433314,
0.0485572405,
-0.0397116654,
-0.0190523136,
0.0107470294,
0.0164883863,
-0.015380498,
0.0640462562,
0.0542324185,
0.0506767929,
-0.0222867075,
-0.0282702167,
-0.024507951,
-0.0345004462,
-0.0079780463,
-0.0081788171,
0.0242082756,
0.0180862378,
0.000039885,
-0.0539133698,
0.0010988888,
-0.0270992424,
2.606e-7,
0.0592414103,
0.0661845654,
-0.0068070758,
-0.0065787835,
0.0144817941,
0.0293892734,
0.0254757889,
0.0291742105,
0.0158448052,
-0.0359468348,
0.004711464,
0.0169515777,
0.0429663323,
0.0376924947,
-0.1029258147,
-0.0647705793,
-0.0800203755,
0.0061371014,
-0.0704995617,
0.0490358025,
0.0302835368,
0.0421314314,
0.0397702716,
-0.0221374836,
-0.0175753236,
-0.0239073541,
0.0163116697,
-0.0527550913,
-0.0081376201,
0.0399055555,
0.0618458465,
-0.0046021626,
0.0223687217,
0.0192566533,
-0.0435670353,
-0.0252558179,
0.0126778688,
0.0450181514,
0.017743038,
-0.0267754234,
-0.0144687332,
-0.0021338672,
0.0303103682,
-0.0399040356,
0.058105398,
-0.0410455875,
-0.0245555099,
-0.0528658703,
0.0167133231,
-0.0341951661,
0.0521723256,
0.0083338823,
-0.0254959222,
0.0088242562,
0.0120723136,
-0.0199139193,
-0.0631378815,
-0.084716484,
-0.0111533301,
-0.0200272147,
-0.0156103661,
-0.069452621,
-0.0254562963,
0.017650092,
0.0050747655,
-0.0621561259,
0.0267159119,
2.513504469e-34,
0.0205089487,
-0.0157481637,
0.0158423372,
-0.0282594878,
0.0413071029,
0.0197589956,
0.0250360966,
0.0082647987,
0.0445464849,
-0.0468306541,
0.0080984235
] |
http://arxiv.org/abs/2401.13313v1 | http://arxiv.org/pdf/2401.13313v1 | InstructDoc: A Dataset for Zero-Shot Generalization of Visual Document Understanding with Instructions | 2024-01-24T09:09:37 | 2024-01-24T09:09:37 | Ryota Tanaka, Taichi Iki, Kyosuke Nishida, Kuniko Saito, Jun Suzuki | cs.CV | cs.CV, cs.CL | We study the problem of completing various visual document understanding
(VDU) tasks, e.g., question answering and information extraction, on real-world
documents through human-written instructions. To this end, we propose
InstructDoc, the first large-scale collection of 30 publicly available VDU
datasets, each with diverse instructions in a unified format, which covers a
wide range of 12 tasks and includes open document types/formats. Furthermore,
to enhance the generalization performance on VDU tasks, we design a new
instruction-based document reading and understanding model, InstructDr, that
connects document images, image encoders, and large language models (LLMs)
through a trainable bridging module. Experiments demonstrate that InstructDr
can effectively adapt to new VDU datasets, tasks, and domains via given
instructions and outperforms existing multimodal LLMs and ChatGPT without
specific training. | InstructDoc: A Dataset for Zero-Shot Generalization of Visual Document Understanding with Instructions: We study the problem of completing various visual document understanding
(VDU) tasks, e.g., question answering and information extraction, on real-world
documents through human-written instructions. To this end, we propose
InstructDoc, the first large-scale collection of 30 publicly available VDU
datasets, each with diverse instructions in a unified format, which covers a
wide range of 12 tasks and includes open document types/formats. Furthermore,
to enhance the generalization performance on VDU tasks, we design a new
instruction-based document reading and understanding model, InstructDr, that
connects document images, image encoders, and large language models (LLMs)
through a trainable bridging module. Experiments demonstrate that InstructDr
can effectively adapt to new VDU datasets, tasks, and domains via given
instructions and outperforms existing multimodal LLMs and ChatGPT without
specific training. | [
0.0059522456,
0.0369478688,
0.0277084894,
0.0611814149,
0.0029401735,
0.0504098088,
0.0112535004,
-0.0022687609,
0.0148715712,
-0.028210381,
-0.0404196009,
0.00584203,
-0.0206949078,
0.0119309267,
0.0204721559,
-0.0512014553,
0.0351687521,
-0.0253711417,
-0.0152350729,
0.0158128198,
0.0404774062,
0.0234177597,
-0.0242525339,
-0.008201574,
0.0029771843,
-0.0577068888,
-0.0435369425,
0.0224418566,
0.020229483,
-0.0228373036,
-0.0086815748,
0.0324401036,
0.0877395421,
0.0193438046,
0.0000015517,
-0.0006017561,
-0.0009577062,
-0.0058813347,
-0.0322009325,
-0.005815316,
0.0275914278,
-0.0879305452,
0.0251909178,
-0.0207622144,
-0.0136346752,
-0.0604988374,
0.030957289,
0.0657934025,
0.0703145117,
0.075479269,
-0.0211925395,
-0.014779483,
0.0204263292,
-0.030386623,
0.0914103016,
-0.0673988536,
0.0076005487,
0.0499610379,
0.0116681634,
0.0240830258,
-0.0533841625,
0.0320668295,
0.0388924107,
-0.0206578728,
0.0018574564,
0.0548212938,
0.0058886674,
-0.0313148536,
-0.0197680462,
0.1051858589,
-0.0025258646,
0.0324282534,
-0.0185326394,
0.0207164623,
-0.0027256866,
0.0094212778,
-0.0902850926,
0.0129953548,
-0.0312049463,
-0.0158811957,
-0.021358842,
-0.0310518537,
-0.0114344927,
0.0041071521,
0.0064099403,
0.0039414125,
0.0028622008,
0.0028478985,
0.0088605937,
0.0183907263,
-0.0357685983,
-0.0180747788,
-0.0090539437,
0.0521572419,
0.0031645792,
0.0115957689,
-0.043669194,
-0.0243050028,
0.0014931583,
-0.0350481011,
0.0606320687,
0.0211509801,
-0.0494907089,
0.1046546623,
-0.0561109111,
0.0515309647,
-0.0343291387,
0.0588001981,
-0.0404354967,
-0.0012453799,
-0.032999564,
0.0102100922,
-0.0394628607,
-0.008032864,
-0.0463259332,
-0.0436148234,
0.0228688028,
0.0304060653,
0.0054952055,
0.0344192088,
0.0041615791,
0.0175237954,
-0.0490917116,
-0.0489031821,
-0.016376812,
0.0282966886,
-0.0444407687,
0.0176109783,
-0.0258227848,
-0.0221537296,
-0.0300208014,
0.0383889116,
0.0322193801,
-0.0208006352,
0.0054995138,
0.0561509281,
-0.0749093369,
-0.0419738404,
-0.0255532153,
0.0101207262,
0.0363895148,
-0.0481452271,
0.0372677967,
-0.0237123724,
0.0533931032,
-0.0038113508,
-0.0153661976,
-0.0022721954,
-0.010074784,
0.0104455939,
-0.0139102684,
0.1084926352,
-0.0132553931,
-0.0176823325,
0.0458592102,
0.029773077,
-0.0179946441,
0.0758252293,
0.004258337,
0.0436778627,
0.027867211,
0.0084432922,
0.0254344139,
0.044153288,
-0.005986949,
0.0179096982,
0.017149495,
0.0431011952,
0.0207348354,
0.0104181925,
0.0743164048,
0.0046177735,
-0.0763230771,
-0.0412943251,
0.0340867043,
0.0180684291,
-0.0061336295,
-0.0155396964,
0.024264181,
0.0278410204,
0.0153960641,
0.0061199251,
-0.0624934174,
0.0450584888,
0.0197952185,
0.0137301972,
-0.0076552709,
-0.0292782709,
-0.0395165011,
-0.012077014,
0.0204378068,
0.0037849848,
0.0187228285,
0.004153321,
0.0519628078,
0.0231879652,
-0.066050455,
-0.0048361463,
-0.0062501053,
-0.0163268894,
-0.0182241332,
0.0152199734,
0.0658587664,
0.0739775896,
0.0578547008,
-0.0182993189,
-0.02728278,
-0.00752402,
0.0725817755,
0.0333654955,
-0.0109524531,
-0.0148977181,
0.0613810234,
-0.104577221,
-0.0309432577,
-0.0266702771,
-0.0164091215,
0.0371881723,
0.0212705322,
-0.0340579599,
-0.0036442264,
-0.05772838,
-0.0100450311,
0.0498956256,
0.0141801713,
-0.0663257763,
-0.0319314599,
0.0052383612,
0.0005381879,
-0.0086011281,
-0.02169732,
-0.0360549055,
0.0076042274,
-0.0315939225,
0.0261643287,
-0.0237772726,
-0.0306766704,
-0.0395803899,
0.0029430399,
-0.0636348501,
-0.0473977663,
0.0097591393,
-0.0684273317,
-0.027418308,
-0.0157963168,
-0.0071102227,
-0.0093948189,
-0.0596981794,
-0.0305566899,
0.0511438474,
0.0185135268,
-0.0117728477,
0.0014108764,
-0.0079401778,
-0.008613525,
0.0083060013,
0.0103562716,
0.0087149134,
-0.0285308417,
-0.001289443,
0.0039415169,
0.0051285499,
-0.0120960483,
-0.0159507468,
0.0662897229,
-0.0609108917,
-0.0031692865,
-0.1147193536,
0.0086845038,
-0.0190196596,
0.0037129973,
0.0220047031,
0.0013992853,
-0.0093028946,
-0.0614305474,
-0.0116984313,
0.019691214,
0.0016875661,
-0.0016848389,
0.0511871129,
0.0097279353,
0.0105798859,
-0.0182016678,
-0.0207653418,
0.0280784946,
0.0207998659,
0.0305579249,
0.0272109509,
-0.0425198153,
0.0144576058,
-0.0011184176,
0.042008426,
-0.0437452905,
-0.0077292868,
0.0316274501,
-0.04007392,
0.0286367554,
0.1145151183,
-0.0221931692,
0.0438698679,
-0.0229563899,
-0.0015136177,
-0.0592712834,
0.0054367026,
0.0033548733,
-0.0255511701,
-0.0216070507,
-0.0296102837,
-0.0582353212,
0.0506247431,
-0.0208705217,
0.0243656281,
-0.0360369608,
-0.0026989537,
0.0185668562,
-0.0188531075,
-0.0797470212,
0.0198961999,
0.0310048144,
0.0157692675,
-0.0702179521,
-0.0576326214,
0.0436272807,
-0.0408966951,
-0.0652121231,
-0.0141873537,
0.0102306381,
-0.0472127944,
0.0454524904,
0.0246240851,
0.0022236211,
-0.0181379393,
-0.0445608199,
-0.0359466821,
0.0004285942,
0.0265023392,
0.0115504134,
0.032858856,
-0.0449700318,
0.0283093583,
-0.0631999746,
-0.0027748765,
-0.0227563158,
-0.0083788037,
0.0020370472,
-0.0076794303,
-0.0173694547,
-0.007069381,
-0.0461230129,
0.0089010494,
0.0337309092,
0.0829207972,
-0.0355992205,
-0.0256567355,
-0.0488004498,
0.0512172766,
0.0610842183,
0.0160465278,
-0.0442253351,
-0.061047595,
0.0203844626,
0.0470776111,
0.0028743078,
-0.0230998453,
-0.0348007157,
0.0313799903,
-0.0043532201,
0.0008899748,
0.0471584611,
-0.0011872988,
0.0332816616,
-0.0839303285,
-0.0167228468,
0.0683612302,
0.0132890474,
-0.0462533385,
-0.0612548217,
-0.0156591367,
0.0198287666,
0.0215299707,
0.0323063657,
-0.1083238348,
-0.0510491021,
-0.01143556,
0.0066475049,
0.0500917956,
-0.000047435,
-0.0167337805,
-0.0588561706,
0.0466078408,
0.0054326477,
-0.0021547442,
-0.0208245423,
-0.0114053013,
-0.013828041,
-0.0020057501,
0.0092631821,
0.0173129886,
-0.0137311788,
-0.0064173969,
0.0100780167,
0.0258716252,
-0.0126512181,
-0.0012170003,
-0.0176167376,
-0.0137238428,
0.0402578488,
-0.0023104816,
0.0251494553,
-0.012730279,
0.0352832228,
0.0307419114,
-0.0113341175,
-0.0104355672,
0.1063146815,
-0.0384021364,
-0.0189660154,
-0.0113900043,
0.0572569817,
-0.0423336476,
0.0024750845,
0.0105657438,
0.0109178005,
-0.0217292923,
-0.0249265525,
-0.02983406,
-0.0242327489,
-0.0085186167,
0.0514822416,
0.0067423638,
-0.0111347502,
-0.0519303121,
-0.0316058844,
0.005879302,
-0.0884914696,
0.038840916,
-0.0030646957,
-0.03204583,
-0.0160361864,
-0.0412285179,
-0.0455835238,
0.0833442062,
0.0224287007,
-0.0068953978,
-0.0585213192,
0.0141713331,
-0.0051314444,
-0.0604128055,
0.0641948655,
-0.0249311291,
0.0110345529,
-0.0731717497,
-0.025852453,
0.002280314,
-0.0171416998,
-0.0291892607,
0.0239005089,
-0.0622378439,
-0.0167350546,
-0.0162752997,
0.0609478951,
0.0148472823,
0.0754293799,
0.0067163939,
-0.031258516,
0.0258344151,
0.005807769,
-0.0186436828,
-0.0126448302,
0.024453748,
-0.0147902993,
0.011917321,
0.0082005309,
-0.0285618138,
0.0290483311,
-0.0035371524,
-0.0105687231,
-0.0360216722,
0.0309939384,
-0.0099353315,
0.0005311267,
-0.000767162,
0.0105795264,
0.0485253595,
-0.0452811755,
0.0259856097,
-0.0071640499,
0.0122324349,
-0.0640672222,
-0.0036486394,
0.0280543398,
-0.0303334147,
0.0031661736,
0.0231296849,
0.0033991865,
0.0368474871,
-0.0206601508,
0.0100204153,
0.0295395181,
0.020430794,
0.0101403249,
0.0005319388,
-0.0180049371,
0.0420548059,
-0.0200907458,
0.053973753,
0.0040352992,
0.087710835,
0.0400218479,
0.0053618285,
-0.014750246,
0.0091781449,
-0.0191774778,
-0.0368842781,
-0.023073867,
0.0783142,
0.0275159795,
-0.000104476,
0.0153477741,
-0.0047111139,
-0.0257313289,
-0.0043311096,
0.0217839237,
-0.0205575544,
-0.028247254,
0.0507862121,
0.0434192829,
0.0363427736,
-0.0077743973,
0.0944156423,
0.0262680147,
0.0315833539,
-0.0317602791,
0.000698878,
-0.0609869733,
-0.0201898981,
0.059598539,
-0.0090720383,
0.0119020622,
0.0491124317,
0.0198123101,
-0.0337341987,
-0.0166868661,
0.0397446528,
-0.0369573832,
0.0257550925,
0.0315211378,
0.0286514163,
0.0103578176,
-0.0437721089,
-0.0122912088,
0.0130831655,
-0.0285344869,
-0.0313930288,
-0.0082201371,
-5.294970957e-33,
-0.040019501,
-0.0294748992,
0.0501862839,
-0.0169303212,
0.0082803583,
0.0111371735,
-0.0207517911,
-0.0195881333,
-0.0012940433,
0.0196013413,
-0.0005570273,
0.0092288489,
0.0221347362,
0.0530658364,
-0.0076987417,
-0.0098978085,
0.0308414493,
-0.0302578453,
-0.0078551536,
0.0243218988,
0.0230059717,
0.0397874154,
0.0443101227,
-0.0288307685,
0.0463831015,
-0.0149251614,
-0.0360234268,
0.0335958935,
0.0389671251,
0.0370407142,
-0.022002887,
-0.0166652892,
0.0249231625,
-0.013438453,
-0.0328956544,
0.0587636009,
-0.1437104195,
-0.0002098713,
0.015648745,
0.0543480217,
-0.0896704793,
-0.009080614,
0.0205531139,
-0.035296578,
0.0500144809,
-0.0562927239,
0.0364732556,
-0.0102565475,
-0.0032593089,
0.0177377257,
-0.0348831937,
-0.0025086796,
-0.026401978,
-0.0441496372,
0.0354530066,
0.0279436558,
0.0110959914,
0.0113220587,
-0.0584125109,
-0.0207739435,
-0.0115032168,
0.0169232525,
0.0448527634,
0.0388838798,
-0.0000289186,
0.0105481744,
0.0109421303,
-0.0371904075,
-0.0250086021,
0.0016932588,
0.0257059634,
0.0473808795,
-0.0009025033,
0.0293748248,
0.0463659614,
0.0397387557,
-0.0256070904,
0.0590275079,
-0.0051540695,
0.0689471811,
-0.0087111462,
0.0115678292,
-0.0017533306,
0.0174883381,
-0.0082139131,
-0.0029068633,
-0.03426392,
-0.0160240661,
0.0458798707,
0.0205948912,
0.0010964185,
0.0015864962,
0.0886817873,
-0.0406527668,
-0.0022402767,
0.0427171029,
-0.0079612043,
0.041234903,
0.0046544652,
0.0180372559,
-0.038775444,
-0.0152944578,
0.0573046766,
-0.0146527737,
-0.0224815495,
0.0089441119,
0.0698907673,
-0.0142708281,
-0.0467825606,
-0.0199039225,
-0.0364202894,
0.0232202392,
-0.0154042318,
0.0535847209,
0.0015281101,
0.0046433802,
-0.0568205491,
0.0471737422,
-0.0419744328,
-0.0044091493,
-0.0300958939,
-0.0475905128,
0.0053821448,
0.0319480412,
-0.0266456828,
0.0053746104,
-0.0289584529,
-0.0548072271,
0.0201283786,
-0.0652663261,
0.0238118004,
0.0810793117,
2.211e-7,
0.034390714,
0.024143694,
0.0601678416,
0.0546170287,
0.0065274523,
0.0400941893,
-0.0208786633,
0.0126541117,
0.0092355357,
-0.0498342477,
0.0245106891,
0.0648201108,
0.037359301,
-0.0026751338,
-0.0641056374,
-0.0086565353,
-0.0156612825,
-0.004960103,
-0.0636047572,
-0.0193339232,
0.0376899131,
0.0657609329,
-0.0506794304,
-0.0063974219,
0.0453943051,
-0.0438720584,
0.0698087215,
-0.0061420472,
0.0026926375,
0.0010250905,
0.0140385376,
0.0130218053,
-0.0163222011,
0.0134129608,
0.0201066565,
-0.0178026985,
0.0418736823,
0.0620935522,
-0.0141344983,
-0.0081491126,
-0.0195938256,
-0.0116900569,
-0.0123048509,
-0.0414017476,
0.0728663951,
-0.0549930036,
-0.0148603627,
-0.0016118512,
-0.0646000877,
-0.0127603794,
0.0050178398,
0.0029522006,
-0.0066641914,
0.0429041348,
0.0574495606,
-0.0672877133,
-0.0012039697,
-0.0680166334,
0.0179642793,
0.006072911,
-0.0292858314,
0.0076551093,
-0.0420167334,
0.0119258892,
0.0238649379,
-0.0244878046,
-0.0699627697,
1.36142817e-34,
0.0301983375,
-0.0057284199,
0.0212452486,
0.0354252011,
0.0336207747,
-0.0218414925,
-0.0674640536,
-0.0069945622,
0.0053235744,
-0.0986209512,
-0.0191123579
] |
http://arxiv.org/abs/2401.13303v2 | http://arxiv.org/pdf/2401.13303v2 | MaLA-500: Massive Language Adaptation of Large Language Models | 2024-01-24T08:57:39 | 2024-04-03T08:23:06 | Peiqin Lin, Shaoxiong Ji, Jörg Tiedemann, André F. T. Martins, Hinrich Schütze | cs.CL | cs.CL | Large language models (LLMs) have advanced the state of the art in natural
language processing. However, their predominant design for English or a limited
set of languages creates a substantial gap in their effectiveness for
low-resource languages. To bridge this gap, we introduce MaLA-500, a novel
large language model designed to cover an extensive range of 534 languages. To
train MaLA-500, we employ vocabulary extension and continued pretraining on
LLaMA 2 with Glot500-c. Our intrinsic evaluation demonstrates that MaLA-500 is
better at predicting the given texts of low-resource languages than existing
multilingual LLMs. Moreover, the extrinsic evaluation of in-context learning
shows that MaLA-500 outperforms previous LLMs on SIB200 and Taxi1500 by a
significant margin, i.e., 11.68% and 4.82% marco-average accuracy across
languages. We release MaLA-500 at https://huggingface.co/MaLA-LM | MaLA-500: Massive Language Adaptation of Large Language Models: Large language models (LLMs) have advanced the state of the art in natural
language processing. However, their predominant design for English or a limited
set of languages creates a substantial gap in their effectiveness for
low-resource languages. To bridge this gap, we introduce MaLA-500, a novel
large language model designed to cover an extensive range of 534 languages. To
train MaLA-500, we employ vocabulary extension and continued pretraining on
LLaMA 2 with Glot500-c. Our intrinsic evaluation demonstrates that MaLA-500 is
better at predicting the given texts of low-resource languages than existing
multilingual LLMs. Moreover, the extrinsic evaluation of in-context learning
shows that MaLA-500 outperforms previous LLMs on SIB200 and Taxi1500 by a
significant margin, i.e., 11.68% and 4.82% marco-average accuracy across
languages. We release MaLA-500 at https://huggingface.co/MaLA-LM | [
0.0285364054,
0.0826409906,
0.0003486082,
0.0511963889,
0.0102820648,
0.016266454,
0.0062653157,
-0.003895669,
-0.0342986137,
-0.0394146182,
-0.0855987892,
-0.0272758175,
-0.0024040269,
0.0076287729,
0.0348108001,
-0.0496688075,
0.0505048372,
-0.0621668287,
-0.0115121044,
-0.0128514571,
-0.0086525027,
0.0258588661,
0.013720057,
0.0037373451,
0.0101878354,
0.0149119748,
0.0259332489,
0.0013935607,
0.0227715466,
-0.0544202998,
0.0152451852,
0.0298347883,
0.0471468344,
0.0592599697,
0.0000020055,
-0.0539723113,
0.010404964,
-0.0134070246,
-0.024145795,
0.0563618541,
0.0597982146,
-0.0258472282,
-0.0056492933,
0.0244768206,
-0.0263860617,
-0.0121572316,
0.0476174913,
0.0098713264,
0.0162051748,
0.0972611755,
-0.0271804277,
-0.0291604437,
0.0227690432,
-0.0780857876,
-0.0157012995,
-0.055424571,
0.059723258,
-0.0088857152,
0.0112006767,
0.0135091608,
-0.0401522629,
0.0119965263,
0.0013005957,
0.0076792715,
-0.0018246282,
0.0220739208,
-0.0110185845,
-0.023004787,
-0.0163179189,
0.0757387727,
-0.0215840116,
-0.0193038024,
-0.022697024,
0.0116687911,
-0.0268741399,
-0.0091055054,
-0.0327312388,
-0.0133311786,
0.0396253243,
0.0226283129,
0.0734379664,
-0.0225190651,
0.0603404976,
0.0198492948,
-0.0592447519,
0.0213968698,
0.033884231,
-0.012362659,
0.0369476192,
0.0161093753,
-0.032413058,
-0.0118723307,
-0.0029272954,
0.0176737793,
-0.034957841,
0.0206463858,
-0.0234561432,
-0.0327823311,
0.0314851105,
-0.0436458923,
0.0603658482,
0.0238772184,
-0.0168357901,
0.0720137209,
-0.0419660956,
0.0229572114,
-0.0147462962,
0.0031433385,
-0.0448452309,
-0.029023381,
-0.0155644557,
0.0231571738,
-0.0145180384,
-0.0155868363,
-0.0541467965,
-0.0114287967,
-0.034679722,
0.0075640935,
0.0100646922,
-0.0157047249,
0.0149472496,
-0.0414501652,
-0.0515767783,
-0.0406843014,
-0.0051978203,
0.0721908137,
0.0007888785,
0.0510486029,
-0.031483911,
-0.0328578614,
-0.0096334266,
0.0135399103,
0.0043224548,
-0.0347250216,
0.0469730012,
0.0468229763,
-0.0172532294,
-0.0645508543,
-0.0297629088,
0.0591507591,
-0.0285135042,
-0.0460780822,
-0.0174056776,
-0.0764358938,
0.0588791296,
-0.0116243046,
0.0318802819,
-0.0534185581,
0.012421946,
-0.0047833989,
-0.0403334834,
0.064985849,
-0.0262259282,
0.0051574823,
0.0322632156,
-0.0054414603,
0.0431107394,
0.0799533427,
0.029723797,
-0.0035054456,
0.0391852893,
-0.0115874801,
-0.0326972641,
0.1008513048,
0.0145843308,
0.0476599224,
-0.0308262259,
-0.0129323564,
0.0047208653,
0.0337645262,
0.0710409805,
0.0184535682,
0.0125635248,
0.0111951809,
0.0619887561,
0.041462142,
0.0063171983,
0.0384843163,
0.0346581005,
-0.0087348791,
-0.0084738769,
-0.0777940229,
-0.0550974011,
-0.0113083972,
-0.0266666282,
-0.0075446004,
0.0169490315,
0.0324940905,
-0.0070688212,
-0.0219084397,
-0.013865849,
-0.0062203663,
0.0257308893,
0.0175514314,
0.0636913776,
0.0006078402,
-0.0075559169,
0.014389066,
0.0765769705,
-0.0015513272,
0.025205696,
-0.0268310532,
0.0651563406,
0.0486780442,
0.0204631798,
-0.0299041122,
-0.009806443,
-0.0001001789,
0.0088301133,
0.0653381571,
0.0450733639,
-0.0147633152,
0.0223149564,
-0.0780011714,
0.0191788618,
0.016493082,
0.0234871954,
0.0391864888,
0.0297365077,
0.0579620376,
0.0002969503,
-0.0289242454,
-0.0394825488,
0.0301473159,
-0.0707618594,
-0.0505871996,
0.0286549293,
-0.0161856897,
0.0483182222,
-0.0066956477,
-0.0187007245,
-0.0227890629,
0.025322469,
-0.0508503579,
0.0137873236,
-0.0026756369,
-0.0519070253,
-0.0437621251,
0.0555796884,
0.0343399905,
0.0033374422,
0.0100591583,
-0.0398793109,
-0.0275345687,
0.0074368785,
-0.0174675062,
0.0304244552,
-0.0183190797,
0.0324241444,
-0.0017371791,
0.0740810409,
-0.0236683227,
0.0321782753,
0.0132100815,
-0.0273089092,
-0.0058379225,
0.0100124804,
0.0093848333,
-0.0089924298,
-0.029664984,
0.0549671091,
0.0165414847,
0.0046256627,
-0.0267965682,
0.0404447429,
-0.0256338567,
0.0033933767,
-0.1150618494,
0.0265713185,
-0.016756041,
0.0251068417,
0.0080738794,
0.0212259758,
-0.0308512077,
-0.0267654117,
-0.0129915904,
0.0183456205,
-0.0253123101,
0.0080841463,
-0.0060447524,
-0.0022200409,
0.0002050812,
-0.0383650288,
-0.0034792896,
0.0229074042,
0.0421027541,
-0.0309061315,
-0.0412497148,
-0.0295767039,
0.0293628573,
-0.0189406313,
0.0066895811,
-0.0071115531,
-0.0266372543,
-0.0560675189,
-0.0276399143,
0.0308609158,
0.0491516665,
-0.0094442023,
0.0059116706,
0.0460438617,
0.0093138982,
-0.0339949578,
0.0118153663,
0.0154922726,
-0.0544352904,
-0.0630902871,
-0.0373783298,
0.0059622582,
0.0644121468,
-0.0145664653,
0.0082168449,
-0.0492474213,
-0.0165289529,
-0.0334792025,
-0.0101951035,
-0.0235050898,
0.0333363488,
-0.0280951671,
0.0281199459,
-0.0265747868,
-0.0358315557,
0.0317523777,
-0.0274451263,
-0.0102416156,
-0.0664619505,
-0.0322145596,
-0.0391237587,
0.0295662172,
0.0098324409,
0.0122729819,
-0.0532506369,
0.0053850692,
-0.0453466065,
-0.0304940753,
0.0096960608,
-0.014731979,
0.0474051908,
-0.0545978583,
0.0178772043,
-0.0180478524,
-0.033809226,
-0.0103252884,
-0.0276059564,
-0.0551529713,
-0.0018668408,
-0.0321001634,
-0.0273383316,
-0.0392998159,
-0.0267825723,
0.0235912111,
0.0613542572,
0.0103875864,
0.0059550605,
0.002710348,
0.0121508474,
0.0378905199,
0.0517603867,
0.0230205934,
-0.0481430627,
0.0179749932,
-0.0016494895,
-0.0144025534,
-0.0021457414,
-0.0157092884,
0.015793452,
0.038793236,
0.016261898,
-0.0136021627,
0.013228992,
0.0572890192,
-0.0374829546,
-0.0093773799,
0.0899880305,
0.0079606296,
-0.0821306556,
0.0207356308,
-0.0448864289,
0.025674399,
0.0124759087,
0.0361657105,
-0.0624004528,
-0.0402195156,
0.0153500596,
0.0138761941,
0.0301745608,
-0.0006754544,
0.0003521165,
-0.131013602,
0.0331656076,
0.0280398708,
-0.0739758685,
-0.0313163251,
-0.0576574691,
0.0303500425,
-0.027145287,
0.0240080487,
-0.0129338829,
-0.0245740563,
-0.0364745595,
0.0009190663,
-0.0236306954,
-0.0669272393,
0.0296643805,
-0.0216570683,
0.0477110185,
0.0318345018,
0.027128458,
-0.0399086922,
0.0065256967,
0.0415206887,
-0.0237814877,
-0.0274122506,
-0.0491185859,
0.0271645933,
0.0057592154,
0.000202753,
0.0256588552,
0.0447574183,
-0.0367325321,
-0.0198234729,
0.0514184013,
0.0672200546,
0.0417450294,
-0.0284888353,
0.0152529776,
-0.0289193951,
0.0077407565,
0.0512342863,
0.0190607775,
-0.0695349947,
0.0074874763,
-0.0156766661,
-0.0545174666,
-0.0303002987,
-0.0177463051,
0.0103985397,
-0.0752917603,
-0.0173963513,
-0.0600848459,
-0.0086340355,
0.012952893,
0.0023654229,
0.0471783467,
0.0536324196,
-0.0137074552,
-0.0242758207,
-0.0233528968,
0.0425610058,
-0.0476322956,
0.0230735857,
-0.0689259171,
0.0009689641,
0.0332975797,
-0.0090639973,
-0.0031995219,
0.0213130116,
-0.0741972923,
-0.0605831072,
0.0077153412,
0.0261822492,
0.0101257451,
0.0446614884,
0.0108709419,
-0.08263468,
0.0779151395,
0.0182980094,
-0.0178893264,
0.0542092584,
0.0209386982,
-0.0040342375,
-0.0135240117,
0.0405620001,
-0.0347838588,
0.0343492478,
0.0391517766,
0.0404455587,
-0.0035879419,
-0.0143098906,
-0.0325816646,
0.0150534399,
0.0153396465,
-0.0423290171,
-0.0381750353,
-0.0644038096,
0.0157969054,
0.0094443318,
0.0018387694,
-0.0203867443,
-0.0256863646,
0.0702243969,
0.0140588125,
-0.0228659175,
-0.0196845084,
0.0328641459,
-0.0041259821,
-0.0206711665,
0.0544562973,
0.0117856888,
-0.0048201168,
0.0480894484,
0.0303998794,
-0.0018289672,
0.0002150754,
-0.0085082063,
0.0444718488,
-0.0287601836,
0.0313555598,
-0.00956129,
-0.0348948278,
-0.0172455236,
0.0201419592,
-0.0195726901,
-0.0249322932,
-0.0328189023,
0.0419821329,
0.0831564516,
-0.0174415391,
0.0887762234,
-0.0112723336,
-0.0232637245,
-0.0386318788,
0.0183945391,
-0.0525893234,
0.00490078,
0.0474012755,
-0.0022031271,
-0.0293820761,
0.0388604142,
0.0546932518,
0.0379836261,
0.0281025749,
0.0186563488,
0.0348962285,
-0.0275016129,
-0.0054979408,
0.0271880478,
-0.0155416988,
-0.0311254188,
0.0173647124,
0.014492915,
-0.0829257369,
0.0126754651,
-0.0470800512,
-0.0375776216,
-0.0041328389,
-0.0023779646,
0.0330624692,
-0.0090364721,
0.0335083902,
-0.0448142402,
0.0138352895,
0.0049545048,
-0.0416318998,
-0.0006116669,
-6.524785636e-33,
-0.0451940894,
0.0062945969,
0.0130253024,
0.0026690285,
-0.0251898803,
-0.0051472061,
-0.0379388407,
-0.0204978399,
-0.0262612477,
0.0086770123,
-0.0145076429,
0.0006330547,
0.0170238409,
0.0272360519,
0.0053353664,
0.0154682007,
0.0412848778,
0.0119149508,
-0.0203945693,
0.0060611875,
0.0093460567,
0.0431482755,
0.0744323209,
-0.0578503758,
-0.0004982949,
0.0152273346,
0.0033691821,
0.0040797931,
-0.0073289061,
0.0687550604,
-0.0366787389,
0.0583376959,
0.0134855788,
0.0150740827,
-0.0057613878,
0.0203855913,
-0.0803145394,
-0.011965475,
0.0174367651,
0.0327319428,
-0.0070216078,
-0.0272430256,
0.0192794707,
0.0079404833,
0.002291417,
-0.0063060173,
0.0168861058,
-0.0648010299,
-0.0383814909,
0.0138089089,
-0.0624754168,
0.0342767388,
0.0284521934,
0.0127029102,
0.0008257482,
0.0415898375,
0.0032878418,
0.0278751254,
-0.0499235392,
0.0144768916,
-0.0174433179,
0.036517594,
0.0357106701,
0.0259815957,
-0.0024888027,
0.0414185934,
-0.0160637163,
-0.0222067237,
0.0232912675,
-0.0292933173,
0.03724939,
-0.0032474587,
0.0301366001,
0.0609952509,
0.0257866401,
-0.0419007838,
-0.0952431336,
0.0189670585,
0.0195045713,
0.0231933985,
0.0319314934,
0.0331880301,
-0.0189900659,
0.0044899983,
-0.0387000367,
-0.0128587009,
-0.0406908952,
-0.0104309814,
0.046795208,
0.0183397271,
0.0087705748,
0.0068571325,
0.0512132384,
-0.0619405583,
-0.0340183824,
0.0183983911,
0.0000265651,
-0.0132445134,
-0.0152734099,
0.0295756143,
-0.0481646247,
-0.05369002,
-0.0003580567,
-0.0163149945,
0.0229381323,
-0.04292465,
0.0547184162,
-0.0042850031,
-0.0876062363,
0.0085395854,
-0.0051336242,
-0.0021675613,
0.034946505,
0.0882912949,
0.0025316332,
-0.0145710791,
0.008576123,
0.0552860834,
0.0440427177,
-0.0183122624,
0.0082348995,
-0.0414079167,
-0.0129999556,
0.010001637,
-0.0494673848,
-0.0014982111,
-0.0165168568,
-0.0015326422,
0.014333725,
-0.0476479195,
-0.0213274509,
-0.0211575776,
2.755e-7,
0.0527725555,
0.0640011728,
-0.0604485497,
-0.01903788,
0.0093036825,
0.0069179698,
-0.0379145965,
0.0274285432,
0.0200919881,
0.0675872043,
0.0175457802,
0.0385777056,
0.0743832141,
0.0036575145,
-0.0548067018,
-0.0183650348,
-0.0546464957,
0.0211907811,
-0.0637870207,
-0.0059229042,
0.0024734733,
0.136665225,
0.0145116486,
-0.0418329947,
-0.0120039526,
-0.0812863633,
0.0762692764,
-0.0597024746,
-0.0393490866,
0.0148435216,
-0.0129716434,
0.0664105043,
-0.0410371348,
-0.0155893853,
-0.0383299068,
0.0740391016,
-0.0144691542,
0.0859264508,
-0.0075850496,
0.0007334484,
0.0447107218,
0.0316418558,
-0.0119894184,
-0.0721095651,
0.0450225361,
-0.0144901993,
-0.0315877907,
-0.0224252865,
-0.0210194886,
-0.0064452491,
0.0323157832,
-0.0553910583,
0.0031542701,
-0.0063884249,
0.0557284057,
-0.0105861844,
-0.0698701814,
-0.0494982563,
-0.0167682879,
0.0263869837,
-0.0362249315,
-0.0257415846,
-0.043337442,
0.000984899,
-0.0102380542,
-0.0486236215,
0.0485822223,
2.278876485e-34,
0.029034581,
-0.024755884,
-0.0025907194,
0.0266374312,
0.0233120415,
-0.0263772216,
0.0105675478,
0.0235449728,
-0.0042303423,
-0.1065285131,
0.010934649
] |
http://arxiv.org/abs/2401.13298v1 | http://arxiv.org/pdf/2401.13298v1 | Towards Explainable Harmful Meme Detection through Multimodal Debate between Large Language Models | 2024-01-24T08:37:16 | 2024-01-24T08:37:16 | Hongzhan Lin, Ziyang Luo, Wei Gao, Jing Ma, Bo Wang, Ruichao Yang | cs.CL | cs.CL, cs.AI | The age of social media is flooded with Internet memes, necessitating a clear
grasp and effective identification of harmful ones. This task presents a
significant challenge due to the implicit meaning embedded in memes, which is
not explicitly conveyed through the surface text and image. However, existing
harmful meme detection methods do not present readable explanations that unveil
such implicit meaning to support their detection decisions. In this paper, we
propose an explainable approach to detect harmful memes, achieved through
reasoning over conflicting rationales from both harmless and harmful positions.
Specifically, inspired by the powerful capacity of Large Language Models (LLMs)
on text generation and reasoning, we first elicit multimodal debate between
LLMs to generate the explanations derived from the contradictory arguments.
Then we propose to fine-tune a small language model as the debate judge for
harmfulness inference, to facilitate multimodal fusion between the harmfulness
rationales and the intrinsic multimodal information within memes. In this way,
our model is empowered to perform dialectical reasoning over intricate and
implicit harm-indicative patterns, utilizing multimodal explanations
originating from both harmless and harmful arguments. Extensive experiments on
three public meme datasets demonstrate that our harmful meme detection approach
achieves much better performance than state-of-the-art methods and exhibits a
superior capacity for explaining the meme harmfulness of the model predictions. | Towards Explainable Harmful Meme Detection through Multimodal Debate between Large Language Models: The age of social media is flooded with Internet memes, necessitating a clear
grasp and effective identification of harmful ones. This task presents a
significant challenge due to the implicit meaning embedded in memes, which is
not explicitly conveyed through the surface text and image. However, existing
harmful meme detection methods do not present readable explanations that unveil
such implicit meaning to support their detection decisions. In this paper, we
propose an explainable approach to detect harmful memes, achieved through
reasoning over conflicting rationales from both harmless and harmful positions.
Specifically, inspired by the powerful capacity of Large Language Models (LLMs)
on text generation and reasoning, we first elicit multimodal debate between
LLMs to generate the explanations derived from the contradictory arguments.
Then we propose to fine-tune a small language model as the debate judge for
harmfulness inference, to facilitate multimodal fusion between the harmfulness
rationales and the intrinsic multimodal information within memes. In this way,
our model is empowered to perform dialectical reasoning over intricate and
implicit harm-indicative patterns, utilizing multimodal explanations
originating from both harmless and harmful arguments. Extensive experiments on
three public meme datasets demonstrate that our harmful meme detection approach
achieves much better performance than state-of-the-art methods and exhibits a
superior capacity for explaining the meme harmfulness of the model predictions. | [
0.0439150929,
0.0705846399,
-0.011325093,
0.0480668843,
-0.0225830954,
0.0299506392,
0.0485316552,
0.0057636537,
0.037489526,
-0.0683006048,
-0.0061184685,
-0.036808379,
-0.0272549782,
0.0381880142,
0.0359773487,
-0.0596898273,
0.0039791381,
-0.0127523895,
0.0440386087,
-0.0184092913,
0.0624311753,
0.0312781781,
-0.005225786,
0.0636635944,
-0.035356883,
-0.0383073576,
0.0134472633,
0.0266130362,
0.0296057705,
-0.0419483408,
0.0025516269,
0.0025785479,
0.0541598238,
0.0570901856,
0.0000017952,
-0.0314287543,
0.0102902418,
-0.0211965591,
-0.0019141483,
-0.0368491858,
0.0461174399,
-0.0121427961,
-0.0019940299,
0.017082924,
-0.0062460159,
0.0701541901,
0.079181008,
0.070191212,
0.1003682613,
0.071202904,
-0.0268876348,
-0.0495213233,
-0.0136339767,
0.0106786648,
0.0020061596,
-0.0019113346,
0.0110456096,
-0.0004851475,
-0.0365050063,
0.0419978984,
-0.0105067315,
0.0016527666,
0.0092020892,
-0.0405991189,
-0.0081524272,
-0.0250216611,
0.0062735067,
-0.0551524423,
0.0207534041,
0.0726802349,
-0.0223348178,
-0.0262986049,
0.0463144518,
-0.0055613942,
0.0039147954,
0.0196211729,
-0.0671333149,
0.0546804219,
-0.0013681625,
0.0226307251,
0.02887295,
0.0244659968,
0.058880806,
0.0806120634,
-0.014412852,
0.0137748308,
-0.0204580408,
0.0010985591,
0.0079765758,
-0.012681256,
-0.0555563048,
-0.0119714076,
0.0187087525,
0.0365785025,
-0.0373481363,
0.0402078666,
-0.0463715754,
-0.0441280976,
0.0245365594,
-0.0789712444,
0.0006611343,
0.0000495832,
-0.0145830736,
0.0495494492,
0.0074070022,
-0.0177784301,
-0.0450222082,
0.0248052087,
0.0216334984,
-0.0368108489,
0.0015752011,
-0.0146871461,
-0.0057178163,
-0.0076580453,
-0.0145839006,
-0.022525236,
-0.0188274365,
0.0113133099,
-0.0741273314,
-0.021168625,
0.0148950322,
0.0302876383,
0.0033608279,
-0.017749818,
-0.0303262025,
0.0073893103,
-0.0373187661,
0.0239228085,
0.0102551868,
-0.0305316094,
0.0071991803,
0.0503916331,
-0.002021485,
-0.038472373,
0.0096291639,
0.1058048308,
0.0127900634,
0.0326833241,
-0.0338038728,
-0.0869315341,
-0.0621624254,
-0.0385997482,
0.0240063742,
-0.0497958176,
-0.0062646251,
-0.0144625325,
0.0068741492,
-0.054294046,
0.0256494787,
0.0158724282,
-0.0414456725,
0.0783840641,
0.0004458506,
-0.0032433015,
0.0371550098,
0.0511792079,
-0.0500916764,
0.0719594806,
0.016703926,
0.0581647307,
0.0473604761,
-0.0058670891,
-0.0352804773,
0.0130706914,
0.0426813476,
0.0114663579,
-0.0280896593,
-0.0161503609,
-0.0581851266,
0.0092053916,
-0.0048237992,
0.0480111837,
-0.0158921015,
-0.0249384679,
0.0776306316,
0.0186372753,
-0.0131229106,
0.0089939702,
0.0290458519,
-0.0057110703,
-0.0254324153,
-0.0331699364,
-0.0167117044,
0.0453783944,
-0.0330570042,
-0.0448023044,
0.0388005935,
0.0552174672,
0.0061851349,
-0.0123680774,
-0.0026889152,
-0.0269345473,
0.0382081643,
-0.026113186,
0.0328581408,
0.0196094736,
-0.0455102623,
0.0383612365,
0.0179915261,
-0.0225502979,
-0.0269523393,
-0.0252162032,
0.0501111671,
0.0310780816,
0.0291700792,
-0.0181175079,
-0.000283592,
-0.041414503,
0.0395422652,
0.0467482805,
-0.0068684514,
-0.0069727302,
-0.0065453756,
-0.0492670238,
-0.0278678853,
-0.0304913968,
0.0010385665,
-0.0145409433,
0.0078062937,
-0.0004996352,
0.0120910257,
0.0183643177,
-0.0255889557,
0.0071616401,
0.0206237566,
0.0135916555,
-0.0168962199,
0.0384249203,
0.0489392765,
0.001909883,
-0.0035812196,
-0.0155670457,
0.0095207999,
-0.04124026,
-0.0483453609,
0.0253983922,
-0.0651143417,
-0.0019624343,
0.0414587855,
-0.0727212951,
-0.000719681,
0.0170802716,
-0.0308932513,
0.015819598,
0.0724486336,
-0.0430403389,
-0.0103639578,
-0.0419045277,
-0.0332533866,
0.0017049329,
0.0037989418,
0.0020494903,
0.0681887642,
-0.0155369174,
-0.0287127681,
0.0131063676,
-0.0022109116,
0.0142710209,
0.0134636201,
-0.0238341857,
0.0115592498,
-0.0078327851,
0.0154097341,
0.0122695426,
-0.0114549398,
-0.0521292984,
0.0003615479,
-0.0692716315,
-0.0143110342,
-0.0113585154,
-0.0020455031,
0.0314464942,
-0.0236611534,
0.0325850956,
-0.0160297211,
-0.0414653942,
0.052709341,
0.0330288373,
0.0376571342,
0.0278980061,
-0.0260981694,
0.0176837239,
-0.0340345204,
0.0135143939,
-0.012110631,
0.1001640186,
0.0027134798,
0.004563055,
-0.0118257338,
-0.0059189987,
-0.048204191,
0.0005781557,
0.0049351356,
-0.0356989466,
0.0503313243,
0.0235465914,
0.0304832514,
0.0446313396,
0.0422815457,
0.0875213072,
0.0204415265,
-0.0046727289,
-0.0461419709,
-0.0313105546,
0.0078162774,
0.0218961667,
-0.0115221934,
-0.0599602051,
-0.0284704193,
0.0587079786,
-0.0147634279,
0.0138564855,
-0.0590558872,
0.1185984984,
-0.0182112306,
-0.0364973247,
-0.0255902354,
-0.079529956,
0.0227122791,
0.0324960351,
-0.0553272516,
-0.0631895959,
0.0683032125,
-0.0118526956,
0.0157608613,
0.0475866199,
-0.0407989398,
-0.015804207,
0.0833575204,
0.000078305,
0.0503357723,
0.0129874162,
0.0238227695,
-0.0033773987,
-0.0419008955,
0.0617634989,
-0.0053356956,
0.0341271833,
-0.0426060706,
-0.0038122246,
0.0064228205,
-0.0308929291,
-0.046566464,
-0.0466688126,
-0.0340541974,
-0.0379862785,
0.0024954507,
-0.0010144294,
-0.0154527463,
0.0254522432,
0.0249705296,
-0.0627125874,
0.0040895399,
0.0044294153,
0.0120143658,
0.015820628,
0.047042843,
0.0183685813,
0.0272774324,
-0.0360387042,
0.0303670391,
0.056208238,
-0.0427129753,
-0.0314671323,
-0.0653481558,
0.005250046,
-0.0321337059,
0.0315924585,
-0.0020476612,
-0.0248470977,
0.0761150494,
-0.0457264036,
0.0115206568,
0.070547238,
-0.0105241118,
-0.0030578743,
0.0180591512,
0.0221930742,
-0.0073833941,
0.0099479584,
-0.0310340412,
-0.0408050381,
-0.0413326509,
-0.0296159703,
0.0096483836,
-0.0371348821,
-0.0353923887,
-0.001799484,
-0.0911265984,
0.0132414177,
0.0293353572,
0.0028322446,
0.0074411226,
-0.0079218559,
-0.019367801,
-0.026782155,
-0.0051597599,
-0.0197516307,
0.0241161026,
-0.0206014421,
-0.006005473,
0.0329993851,
-0.0168576967,
-0.0061288006,
-0.016084943,
0.0306539796,
-0.0100090597,
0.0439968109,
0.0045948974,
-0.0592161901,
0.0350861736,
-0.0011835344,
-0.0160419792,
-0.0243713874,
0.0415813141,
0.0164049398,
-0.0335135981,
-0.0658811107,
0.0330556445,
-0.0023007083,
-0.0818727091,
-0.0278830603,
0.0679078773,
-0.017035028,
-0.0084122363,
-0.0499647968,
0.0436951369,
0.0128918253,
0.0654594153,
0.0377735235,
-0.0297518354,
-0.0203885436,
-0.0458737575,
-0.0202021841,
-0.0504783243,
-0.0294403527,
-0.0277645197,
0.0089472653,
0.0217728298,
-0.0582975782,
0.0031997729,
-0.0280438531,
0.0389973447,
-0.0021852348,
-0.0215615164,
-0.0212947205,
0.0003264227,
-0.010389545,
0.036638476,
0.0074570961,
0.0367882438,
-0.0049861064,
0.0104275513,
-0.0103987791,
-0.0266460851,
-0.0122508658,
0.0496306792,
-0.0623295829,
-0.0654500946,
-0.0376851931,
0.0617414638,
0.0257499572,
0.0316727571,
0.0170317013,
-0.0563713871,
0.0190343075,
0.0373436511,
-0.004916226,
-0.026027834,
0.0365314335,
-0.0029221969,
0.0269025806,
0.0357507989,
-0.0129486611,
-0.0019874922,
0.0017292129,
-0.0097291199,
-0.06871292,
0.0034065589,
0.0022839115,
-0.0176951475,
0.0120449858,
-0.0122640403,
-0.012160385,
-0.0257275365,
-0.0110110277,
-0.0156756211,
-0.0293262303,
-0.0176684894,
0.0391124375,
0.0022661902,
-0.0370844305,
-0.0090674423,
-0.0143257715,
0.0150569566,
0.0357803181,
0.0023917784,
0.0288894791,
0.033317592,
0.023454586,
-0.0001738823,
0.0021641452,
-0.0080478853,
0.0538555011,
-0.0120561738,
0.0116862524,
-0.0131621929,
0.0825422481,
0.0075114979,
-0.0344554894,
0.0251763221,
-0.0012140685,
0.0106175616,
-0.0509506166,
0.0174104292,
-0.0274402108,
0.0398409702,
-0.0230216794,
0.0524153523,
0.0066372729,
-0.0266492851,
0.0013057496,
0.0558320358,
-0.0011080019,
-0.0575050712,
0.013370852,
0.019030109,
-0.0157477204,
0.012068131,
0.0607455149,
-0.0363026187,
0.007924852,
0.0510811359,
-0.0440718196,
-0.0112296306,
-0.022935411,
0.0185433459,
0.0567130819,
0.013573681,
0.0281923395,
-0.0411272198,
-0.0703641772,
-0.0295981299,
-0.0117924474,
0.0042348858,
0.0294554643,
-0.0107592465,
0.0546369515,
0.0151902102,
0.0347841717,
-0.03692393,
0.0259024315,
-0.0199099388,
-0.0003629737,
-0.0076388475,
-6.10594304e-33,
0.0016450315,
-0.0887163654,
-0.0270955581,
-0.0395761281,
-0.0210573617,
-0.0101843178,
-0.0322935805,
-0.0079446631,
0.0188355222,
0.0396887809,
0.0502821393,
0.0075370115,
-0.0210427362,
0.0120047107,
-0.0093794418,
-0.0118349418,
0.0244888198,
0.0404286459,
-0.0219184496,
0.0162411127,
0.0790581852,
0.004662069,
0.0045900387,
-0.0812552646,
0.0267428961,
0.0056091719,
-0.0304128192,
0.0303702317,
-0.0193091743,
0.080663681,
-0.014444517,
0.0427084267,
0.0383665115,
-0.0239711907,
-0.0024085194,
0.0536632016,
-0.0296365712,
0.0051331315,
0.0376531668,
-0.0154392039,
-0.0465454385,
-0.0648160204,
0.0514983572,
-0.0108460216,
0.0078576962,
-0.0013552562,
0.0247415919,
-0.0274287518,
0.0493329093,
-0.020946037,
-0.0555025227,
0.0215285439,
-0.0449417084,
-0.0241169464,
-0.0048006452,
0.0228306223,
0.0338916332,
0.0623062998,
-0.0478384532,
0.0353236273,
-0.0204186793,
-0.0531436503,
0.0593069568,
-0.0262100939,
-0.011983512,
0.0841120631,
-0.0354175977,
0.0177606698,
-0.041479703,
-0.0084892623,
-0.0312542692,
0.0543359108,
0.081428133,
0.0713905618,
0.0444435515,
-0.0051055131,
-0.1115724146,
-0.0192409176,
-0.0233784411,
0.0196847636,
0.0289071724,
0.0128130391,
0.0097183371,
0.0106382044,
-0.0242045727,
-0.047366932,
-0.0548777431,
-0.0455557369,
-0.0046140468,
-0.0227908269,
-0.049225159,
0.0002481095,
0.0284031406,
-0.0474665686,
-0.0022076792,
0.013462944,
0.0234201346,
-0.0317043997,
0.0129704867,
0.0474329218,
-0.0005222155,
0.0080718268,
0.0559106991,
-0.0280216355,
0.0259266794,
0.00964907,
0.0624772534,
0.0113373725,
-0.03000772,
0.0036795919,
-0.0759053752,
0.0235414114,
-0.0609734207,
0.0501236841,
0.0277791098,
0.0438763723,
-0.0395487174,
0.0038765031,
0.0332331434,
0.0116316574,
-0.0071860906,
-0.0326813199,
-0.0323054157,
0.0234536938,
-0.018779587,
-0.0364469625,
-0.0434919298,
-0.0330607966,
0.0246508215,
-0.1004574671,
-0.0008652372,
-0.0075364304,
2.492e-7,
0.0135994293,
0.0430014767,
-0.0091776028,
0.0294327065,
0.0560202561,
0.0591358654,
0.0110629853,
0.0095929503,
-0.0310352817,
-0.0671032295,
-0.0220835991,
0.0674297735,
0.0369462632,
0.0457698815,
0.0227281936,
-0.0329431519,
-0.0670332313,
-0.0229082461,
-0.0784068406,
0.0194980539,
0.0948014483,
0.0601982847,
0.0409690328,
-0.0586316995,
0.0063272393,
-0.0293080751,
0.080334641,
-0.041307576,
-0.011548562,
-0.0507008731,
0.0541435704,
-0.0171433426,
0.0058956076,
-0.0128644407,
0.0200857911,
-0.0585165136,
-0.0078442283,
0.0311437584,
0.0068952721,
-0.030753538,
-0.008383695,
-0.0103818011,
0.0082630217,
-0.0659954101,
0.0506302528,
-0.0318626836,
-0.0603522435,
-0.0488948636,
-0.0240846761,
-0.0457630455,
0.029462425,
-0.0493299067,
-0.0323008262,
-0.0072686695,
0.0029839468,
0.0033092205,
0.0128958849,
-0.0854455233,
-0.0030882186,
0.0185873639,
-0.0425791927,
-0.0048103905,
-0.045010455,
-0.0101299761,
0.0687479377,
-0.0125954403,
0.0080672596,
1.940019576e-34,
0.0209989492,
-0.0329883359,
0.0001854791,
-0.0169311278,
0.0447227284,
0.0317263454,
-0.0267849397,
-0.0162746888,
0.0234453008,
-0.0278050918,
-0.0051325643
] |
http://arxiv.org/abs/2401.13275v2 | http://arxiv.org/pdf/2401.13275v2 | Can AI Assistants Know What They Don't Know? | 2024-01-24T07:34:55 | 2024-01-28T09:07:13 | Qinyuan Cheng, Tianxiang Sun, Xiangyang Liu, Wenwei Zhang, Zhangyue Yin, Shimin Li, Linyang Li, Zhengfu He, Kai Chen, Xipeng Qiu | cs.CL | cs.CL, cs.AI | Recently, AI assistants based on large language models (LLMs) show surprising
performance in many tasks, such as dialogue, solving math problems, writing
code, and using tools. Although LLMs possess intensive world knowledge, they
still make factual errors when facing some knowledge intensive tasks, like
open-domain question answering. These untruthful responses from the AI
assistant may cause significant risks in practical applications. We believe
that an AI assistant's refusal to answer questions it does not know is a
crucial method for reducing hallucinations and making the assistant truthful.
Therefore, in this paper, we ask the question "Can AI assistants know what they
don't know and express them through natural language?" To answer this question,
we construct a model-specific "I don't know" (Idk) dataset for an assistant,
which contains its known and unknown questions, based on existing open-domain
question answering datasets. Then we align the assistant with its corresponding
Idk dataset and observe whether it can refuse to answer its unknown questions
after alignment. Experimental results show that after alignment with Idk
datasets, the assistant can refuse to answer most its unknown questions. For
questions they attempt to answer, the accuracy is significantly higher than
before the alignment. | Can AI Assistants Know What They Don't Know?: Recently, AI assistants based on large language models (LLMs) show surprising
performance in many tasks, such as dialogue, solving math problems, writing
code, and using tools. Although LLMs possess intensive world knowledge, they
still make factual errors when facing some knowledge intensive tasks, like
open-domain question answering. These untruthful responses from the AI
assistant may cause significant risks in practical applications. We believe
that an AI assistant's refusal to answer questions it does not know is a
crucial method for reducing hallucinations and making the assistant truthful.
Therefore, in this paper, we ask the question "Can AI assistants know what they
don't know and express them through natural language?" To answer this question,
we construct a model-specific "I don't know" (Idk) dataset for an assistant,
which contains its known and unknown questions, based on existing open-domain
question answering datasets. Then we align the assistant with its corresponding
Idk dataset and observe whether it can refuse to answer its unknown questions
after alignment. Experimental results show that after alignment with Idk
datasets, the assistant can refuse to answer most its unknown questions. For
questions they attempt to answer, the accuracy is significantly higher than
before the alignment. | [
0.0253038779,
0.0537132211,
-0.0367056392,
0.0231786314,
-0.0527958348,
0.0139388125,
0.016127931,
-0.0108886668,
0.0484834351,
-0.0273468047,
-0.0146177169,
-0.0450255722,
-0.0067535364,
0.0206131823,
0.0162348319,
-0.0125629976,
0.0397430733,
-0.0407727249,
0.0377198532,
-0.0286162347,
-0.0246070027,
0.0057718945,
-0.0436405763,
0.0365900099,
-0.0387634635,
-0.0482009314,
-0.0409507751,
-0.0371054821,
0.000520904,
0.011711305,
-0.0021257293,
0.0339367241,
-0.0066019399,
0.0464401469,
0.0000018103,
-0.0513739549,
0.0061305887,
-0.0110652326,
-0.0624040216,
-0.0291526262,
-0.0426532663,
0.0129594915,
0.0272451527,
0.0189129766,
-0.0841999725,
-0.0117921121,
0.0466129854,
0.0588771179,
0.06982366,
0.0811889172,
0.0040109348,
-0.00555549,
0.0501578152,
-0.0010908911,
0.0554337613,
-0.061413072,
0.0443039909,
0.0202009305,
-0.0015707493,
-0.0414840914,
0.0133187249,
0.0339623168,
-0.0147171756,
-0.0070561245,
-0.0253652148,
0.0277059153,
-0.0615348183,
-0.0513553098,
0.0158861391,
0.0658850297,
-0.0147995874,
-0.0330170244,
0.0153621258,
0.0198915508,
0.0141853541,
-0.0070726797,
-0.0335621797,
0.029043518,
-0.0424856953,
0.0064413529,
0.0231650136,
-0.008696191,
0.0259708799,
-0.0086906077,
-0.0403488316,
0.0893112868,
-0.0136973644,
-0.0631612763,
0.0195187926,
-0.0210079346,
-0.0179553814,
-0.086485751,
-0.0184365809,
0.0505020656,
-0.0561457388,
0.0295494348,
-0.0561072715,
0.0006531478,
0.0202727206,
-0.0645478517,
0.0150095886,
0.0292576216,
-0.0262945537,
0.0439759567,
-0.0140186567,
0.0570944399,
-0.0232162364,
0.0683993101,
-0.0154614002,
-0.0456650294,
-0.0239189174,
-0.0347570889,
-0.0494671315,
0.0439555049,
0.0012807712,
-0.0034121466,
0.0141847115,
0.0367823169,
-0.0416064933,
0.1077305451,
-0.031953536,
0.031522762,
-0.019301312,
-0.0522206239,
-0.0518644266,
0.0007538772,
-0.0309670195,
0.0200686846,
0.0167677943,
-0.0723188296,
-0.0136048887,
0.0101610655,
-0.0014164798,
0.0181598701,
0.0224341638,
0.0479872338,
0.0358506404,
0.0159180108,
-0.0231401995,
-0.0180903673,
0.0058788336,
-0.0820903108,
0.011738616,
-0.0399640165,
0.0265356302,
0.0132016968,
-0.0308724772,
-0.063634403,
0.0242568515,
0.0400765389,
-0.0296570808,
0.0813290179,
-0.0306354221,
0.0527728274,
0.0501222089,
0.0170944016,
-0.0327275917,
0.0495306067,
0.0111547839,
-0.0021323084,
0.051065091,
-0.0249773804,
0.0565781109,
-0.0101151383,
0.030470157,
-0.0095077744,
-0.0057564918,
0.0499982312,
-0.0126914764,
0.0165041499,
-0.033035472,
0.0433793664,
0.0508781075,
-0.0528949425,
0.0621822998,
0.031367369,
-0.0548481084,
0.0304028559,
0.0154941743,
0.0005692012,
0.0393750332,
0.02563335,
-0.0641135499,
0.0265623499,
-0.063002646,
0.0190803427,
0.0491057709,
0.0013595258,
-0.04061573,
-0.0481817648,
-0.003747981,
-0.0111360531,
-0.0118517037,
0.0569006428,
-0.0063574859,
-0.0085452907,
-0.031907782,
0.0267891437,
0.0053234398,
-0.0258333571,
-0.0387452096,
-0.0520408526,
0.054848399,
0.0257112868,
0.0388279259,
-0.0837363899,
-0.0421184786,
-0.0262431614,
0.0332101658,
0.0406871773,
0.0558715202,
0.0057362011,
0.0198491756,
-0.0227506626,
0.0082281614,
-0.0116368299,
0.0229911041,
-0.0226114821,
-0.025301801,
0.0053818249,
-0.0056469315,
-0.0462006293,
-0.0326768905,
0.0292997528,
-0.0282127932,
-0.0422591604,
-0.0073358756,
0.0079009514,
0.0277730338,
-0.0020841507,
-0.0122177275,
0.0250764713,
0.0357748084,
-0.0117606325,
-0.0539082102,
-0.0099118603,
-0.0339739621,
0.0105796689,
0.0300459322,
-0.0106572639,
-0.0259946138,
0.0381777659,
-0.0140390145,
-0.0452293865,
0.0310682617,
-0.0351196527,
0.0244569425,
0.0084361974,
0.0506839454,
0.0091338735,
0.0309039913,
0.0075387079,
0.0760560781,
-0.0141653121,
0.0150385918,
-0.0066842018,
-0.0044660876,
0.0270278268,
-0.0277812164,
-0.015297519,
-0.0017517918,
-0.00737258,
0.0200183038,
0.0069904854,
-0.0426552445,
-0.0332965814,
0.0027333526,
-0.0850700065,
0.0090537462,
-0.0045129899,
-0.0026689305,
0.0178939216,
-0.0036035869,
-0.0140515789,
-0.0314318575,
-0.0808719769,
0.0341178812,
-0.0041615455,
0.0004206217,
0.0211239085,
0.009464737,
-0.0072901202,
-0.0269554835,
0.0279642325,
-0.0340296626,
-0.0200543404,
-0.0259644594,
-0.0262152608,
-0.0614355616,
0.0525737219,
-0.0829085112,
0.0826192647,
-0.0237660464,
-0.0076191342,
-0.0050556348,
-0.0038585165,
0.0040394347,
0.1029936373,
0.0178174153,
-0.0101817092,
0.0902503952,
0.0135312537,
-0.0044522714,
0.0438723564,
0.0159508735,
-0.0169734135,
-0.0100614745,
-0.0483262874,
0.0056541702,
0.0440823622,
0.0444723219,
0.0082647558,
-0.0401703864,
0.031052256,
0.0047937292,
-0.0334466919,
0.0049180146,
-0.0142823262,
-0.0059510875,
0.0713374391,
-0.0025472208,
-0.0518254302,
0.0989696458,
-0.0086594215,
-0.0159762129,
-0.0254131556,
-0.0205860324,
0.0014833519,
0.0408800878,
-0.0066815852,
0.0117722852,
0.0000846057,
0.0191219971,
-0.0464350879,
-0.0120582413,
-0.0235618036,
-0.0680293441,
0.0424100049,
-0.0520290174,
0.0039392659,
0.0022303825,
-0.0394764617,
-0.0959907919,
-0.0390788876,
-0.0368293189,
-0.0199423768,
-0.0239359699,
-0.0089513985,
-0.0672176257,
-0.0221552309,
0.0572073609,
0.0258383099,
-0.0056170151,
-0.019252697,
-0.0110151945,
-0.0188868772,
0.0927811414,
0.0060788379,
0.0148672843,
-0.0077121365,
0.056383267,
0.0308756568,
0.0246506408,
0.0203740615,
-0.0002569744,
-0.0425981209,
-0.0018057653,
0.0169126391,
-0.0216685999,
0.0396031886,
0.0239966642,
-0.0090732016,
-0.0430616513,
0.1035444587,
0.0576771572,
-0.022446014,
0.0024348951,
-0.0672171637,
-0.0012453519,
0.0063498076,
0.0179832596,
-0.0698834956,
-0.0189479478,
-0.0135260914,
0.0125260111,
-0.0371745676,
-0.0381542556,
-0.0165548138,
-0.1078311503,
0.0048239101,
0.0015168682,
0.0150702745,
-0.0518978275,
0.0112309372,
0.0105153602,
0.016238017,
0.0224494636,
-0.0312596932,
0.0395414755,
-0.0475519225,
0.0004258246,
-0.0338882133,
0.0035411136,
0.0350547284,
0.0049061924,
0.0805496871,
0.0527781174,
0.0392060392,
-0.0091767879,
-0.0477586202,
-0.0532563142,
0.0285876431,
-0.0189689044,
0.0328945927,
0.0546624698,
0.0159958508,
-0.02145941,
0.0224920809,
0.0295495912,
-0.0568698309,
0.0139941154,
0.0558518618,
0.0791264325,
-0.0588189848,
-0.0206290819,
-0.0120835397,
0.0277077612,
0.0122560784,
0.082686156,
-0.0029039739,
-0.0141533362,
-0.0372756235,
-0.0183840282,
0.0315325372,
-0.0052405004,
-0.015528488,
0.0382505693,
0.0519921444,
0.0204736479,
-0.0761781558,
-0.0343351066,
0.0207362305,
0.0145155713,
0.0234601293,
0.015615765,
0.0334380567,
0.0012324603,
-0.0145344092,
-0.0250872727,
-0.0274963211,
0.0013487156,
-0.0135548534,
-0.0070392485,
0.0304406099,
-0.0193194486,
-0.0011539642,
0.0209028516,
-0.0509376042,
-0.0685205087,
0.0333744884,
0.0420428254,
0.0325918682,
0.0417572707,
-0.0294122938,
-0.0249010567,
0.0134926336,
0.0157461762,
-0.0164466202,
0.0065905619,
0.0256024878,
0.0212222952,
0.0064785611,
0.0586505532,
-0.0619249232,
-0.0149148051,
-0.0033344736,
-0.0007616068,
-0.060383305,
-0.0523923598,
0.0460760035,
0.0234317482,
-0.0161163043,
0.0419523381,
0.0531344935,
0.0077618468,
0.0338986777,
-0.0208719168,
-0.0209218897,
0.0072429711,
-0.0044946363,
0.0188116953,
-0.0424318574,
0.0545797199,
-0.0167134851,
-0.0473031364,
-0.0219956581,
-0.0340039879,
-0.0492113829,
-0.0079799565,
0.0362470634,
0.0042568948,
-0.0175392143,
-0.0232443307,
0.0089566791,
0.0448231734,
0.0654920042,
0.0341801196,
0.0807407424,
0.0120488405,
-0.0202030167,
-0.0228777267,
0.0082612503,
-0.0649153441,
-0.1136400178,
0.0302217808,
0.0467423499,
0.0330916047,
0.0057160649,
0.0014503858,
-0.0399289243,
0.02055154,
-0.0118763158,
0.0198082551,
0.069253318,
-0.0393273681,
0.0348706134,
0.0261702165,
-0.0231634229,
-0.001960349,
0.0849793702,
-0.0021020488,
0.0650118589,
-0.0144606913,
-0.0042133247,
-0.0489917621,
-0.0344058983,
-0.0019258731,
0.0111535126,
-0.0211331435,
0.0291453823,
-0.0612270199,
-0.0790282115,
-0.0084779952,
-0.0073441355,
0.0109023107,
-0.0541094504,
0.0032033366,
0.0495220758,
0.0247680862,
-0.0237799603,
-0.0330018848,
0.0337482616,
-0.0286946669,
0.0124612888,
-0.0437042341,
-6.02702586e-33,
-0.0150207244,
-0.0156496223,
0.0072492454,
-0.0224046726,
0.0237461068,
-0.0046586511,
-0.0157413278,
0.012619867,
-0.0661254004,
0.0022288603,
-0.0293663349,
0.0463089906,
-0.010378249,
0.0170203727,
0.0118793333,
-0.0354473367,
-0.0042051412,
-0.0100204675,
-0.0408034213,
0.0234506484,
0.0520866923,
-0.0322130807,
0.0192251392,
-0.0887468383,
0.0209757071,
-0.0517314374,
-0.0135794142,
0.0402220003,
-0.0542622656,
0.0237315129,
-0.04730241,
0.065959923,
-0.0023586741,
0.016906932,
-0.0030138055,
0.0051661045,
0.0153035056,
-0.0386771709,
0.0284205861,
0.0032406859,
-0.0208804458,
-0.0125591336,
0.0022591597,
0.0114674531,
0.015750641,
0.0154382717,
0.0185845457,
-0.0390822962,
0.0131596094,
0.039852798,
-0.0241949484,
-0.0196716581,
-0.0262477882,
-0.0772079453,
-0.0104576144,
0.0526494682,
0.0153666418,
-0.0585587472,
-0.0538281053,
0.0121849775,
0.004850354,
-0.0127050364,
0.0707582384,
-0.0441480651,
-0.0034319854,
0.0387239382,
0.0302517377,
0.013220503,
-0.053503193,
0.0057956218,
-0.0252851769,
0.0557818413,
0.0437492095,
0.0189006794,
0.0088237869,
-0.0003485398,
-0.0452267751,
-0.0021342905,
-0.0306507144,
0.0423032939,
0.0570086278,
-0.0323472805,
-0.091106832,
0.0014708312,
-0.0248674359,
0.0155530879,
-0.0428577401,
0.0066094291,
0.0452992059,
-0.0003933702,
0.0072527924,
-0.0027662637,
0.0071285898,
-0.0497719757,
0.051904045,
0.0193826351,
-0.0302208513,
-0.0019455604,
0.0067231185,
0.0147082647,
-0.0252027959,
-0.0075064427,
0.0291449353,
0.0190871991,
0.0430255458,
0.0427124016,
0.0363470279,
0.0016008648,
-0.0380261764,
0.0066150688,
-0.0384607203,
0.0390168019,
0.0063985023,
-0.0412265919,
0.0454282463,
0.0012002637,
-0.0263620503,
0.0094882688,
-0.0095175887,
-0.0180905703,
-0.0387607813,
0.0011727479,
-0.0245927796,
0.0057692295,
-0.0467952974,
-0.0217053872,
-0.0399550982,
-0.0149449836,
0.0073558143,
-0.0396211557,
-0.0011156037,
-0.001019472,
2.674e-7,
-0.032168448,
0.0231395066,
0.0218916256,
0.0472438931,
0.0006430433,
0.0797907114,
-0.0225931164,
0.0497654155,
0.0091468776,
-0.0514373928,
-0.0165773947,
0.0082729217,
0.0571558066,
0.0421666056,
-0.0497093834,
-0.0125393681,
-0.0301178712,
-0.001061825,
-0.0755165219,
0.0065315105,
0.0509569012,
0.0911042839,
0.0361314081,
-0.0027234268,
0.0229146685,
-0.0774143264,
0.0373172089,
0.0228482392,
0.0122425817,
0.0174832363,
0.0586000346,
0.0385597013,
-0.0055508688,
0.0728056878,
-0.0011263803,
0.0100147035,
0.0036967413,
0.0265480727,
-0.0206667073,
-0.0382028371,
-0.0430148728,
-0.0049365107,
0.0293257963,
-0.0604978502,
0.0744839534,
-0.0173604097,
-0.0549655445,
0.0042964211,
0.0190796182,
-0.0010506072,
0.0160625614,
0.0132184606,
0.0061700898,
-0.0054997997,
-0.0082977479,
0.0054316721,
0.0218885858,
-0.0416501425,
-0.0045858496,
0.0015230634,
-0.0045393175,
-0.0044520115,
0.0030187154,
0.0322655104,
0.0195866916,
-0.0686728433,
0.063930802,
2.156566528e-34,
-0.0145379892,
0.0114313066,
0.0123168286,
0.0414986797,
0.0208061021,
-0.0184437279,
0.0289366003,
0.0062503177,
0.0139192129,
-0.0419128574,
-0.0095846681
] |