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/2401.16055v1 | http://arxiv.org/pdf/2401.16055v1 | Stolen Subwords: Importance of Vocabularies for Machine Translation Model Stealing | 2024-01-29T11:04:01 | 2024-01-29T11:04:01 | Vilém Zouhar | cs.CL | cs.CL | In learning-based functionality stealing, the attacker is trying to build a
local model based on the victim's outputs. The attacker has to make choices
regarding the local model's architecture, optimization method and, specifically
for NLP models, subword vocabulary, such as BPE. On the machine translation
task, we explore (1) whether the choice of the vocabulary plays a role in model
stealing scenarios and (2) if it is possible to extract the victim's
vocabulary. We find that the vocabulary itself does not have a large effect on
the local model's performance. Given gray-box model access, it is possible to
collect the victim's vocabulary by collecting the outputs (detokenized subwords
on the output). The results of the minimum effect of vocabulary choice are
important more broadly for black-box knowledge distillation. | Stolen Subwords: Importance of Vocabularies for Machine Translation Model Stealing: In learning-based functionality stealing, the attacker is trying to build a
local model based on the victim's outputs. The attacker has to make choices
regarding the local model's architecture, optimization method and, specifically
for NLP models, subword vocabulary, such as BPE. On the machine translation
task, we explore (1) whether the choice of the vocabulary plays a role in model
stealing scenarios and (2) if it is possible to extract the victim's
vocabulary. We find that the vocabulary itself does not have a large effect on
the local model's performance. Given gray-box model access, it is possible to
collect the victim's vocabulary by collecting the outputs (detokenized subwords
on the output). The results of the minimum effect of vocabulary choice are
important more broadly for black-box knowledge distillation. | [
0.04083848,
0.0502883345,
-0.0459527895,
0.0803632066,
-0.0990082845,
0.0080314418,
0.0230070911,
0.0211449303,
-0.0370104313,
-0.0365859382,
-0.0123888925,
-0.0019964627,
0.0086586447,
0.0352770612,
0.0573842302,
-0.0203253385,
0.0258652978,
-0.0253715571,
-0.0468812883,
-0.0152056636,
-0.0356016904,
0.0162372943,
0.0346505605,
0.019092273,
0.0100327963,
0.0145475529,
-0.0066095614,
0.0252176709,
-0.0052922689,
0.05988428,
-0.0181771498,
0.019912973,
-0.0147628784,
0.0639223754,
0.000002075,
-0.0098432116,
-0.09583731,
0.05060561,
-0.0352201946,
-0.0016509072,
0.0357870236,
-0.0261445828,
0.0273725111,
-0.0129479617,
-0.0330869071,
0.0244814344,
0.0370474644,
0.0288407914,
0.0566808879,
0.0727455914,
-0.015477878,
-0.0079687275,
-0.0011002414,
-0.0232462212,
0.0017660992,
-0.0767264217,
-0.013668457,
0.0236544982,
0.0141039826,
0.038703911,
0.0058067953,
0.0348641127,
-0.0125355422,
0.0461951159,
-0.0310862567,
-0.0167183764,
0.0196500234,
-0.0343372338,
0.0126477191,
0.0549300089,
-0.0638989657,
0.0178475343,
0.0106964195,
0.0124070356,
-0.0030142211,
0.0243356545,
-0.0018973115,
0.0359497815,
-0.004906151,
-0.0325278826,
0.0310087688,
-0.0117288716,
0.0249745473,
-0.0231609512,
-0.0367855504,
0.0225419924,
0.0237771291,
-0.0376503915,
-0.0420688875,
-0.0511203222,
-0.0313787535,
-0.061086949,
-0.0016070933,
-0.0153643563,
0.0282054096,
0.0608835109,
-0.0677954555,
-0.0275010541,
-0.0125077004,
-0.0472573936,
0.0384725183,
0.0419941805,
0.0134705268,
0.0203149375,
0.009741961,
0.0233781468,
-0.0569612309,
0.0536192544,
-0.0663859174,
-0.0194211397,
-0.0711724982,
-0.0075799837,
-0.0147149349,
0.0637667328,
-0.0227030236,
-0.0237302464,
-0.008764443,
0.014262897,
-0.0160245374,
0.0241443329,
-0.0407208838,
-0.0146518229,
-0.0507460646,
-0.0232431889,
0.023906311,
-0.0251124315,
0.0009149484,
-0.0205694512,
-0.0059285504,
-0.0876729563,
-0.0412059017,
-0.0386588499,
-0.0305446647,
-0.0330213718,
-0.0122279413,
0.104275614,
0.017750008,
-0.0483065471,
-0.0385834984,
-0.1042873189,
-0.0032344526,
-0.0152552053,
0.0063346084,
-0.0759505928,
0.0116830897,
-0.0082157152,
-0.0040070144,
-0.0443947092,
-0.0001375733,
-0.0064324546,
-0.065826349,
0.0912428126,
0.0142950024,
-0.0245936476,
0.008118066,
0.0202418063,
-0.0020483085,
0.0809933841,
-0.0131501388,
0.0637202114,
0.04850775,
-0.0152860396,
-0.0291819014,
0.0521895885,
0.0110914223,
0.0242308117,
-0.0453530699,
-0.0227331277,
0.0617610812,
-0.0368566848,
0.0007699287,
0.0165652819,
-0.0383490436,
0.0040676328,
0.0591576025,
0.0804443657,
-0.0407259688,
0.0706698671,
0.0234082919,
0.0092174523,
-0.040494401,
0.0385050885,
-0.0231997799,
-0.0564734451,
-0.0499086976,
-0.034728609,
-0.0115664415,
0.017081853,
0.0182567351,
-0.0117604751,
-0.0020930048,
-0.0065998915,
0.0043787248,
-0.0094308918,
0.0356398299,
0.0781307146,
-0.0003270225,
0.1110812277,
-0.000799421,
-0.0317212939,
-0.0168426055,
-0.0417318717,
-0.0641019493,
0.0085530877,
0.0738685131,
-0.0223660562,
0.0183188748,
0.0017240485,
-0.0406455509,
0.0320976377,
0.0438362174,
0.0092615327,
0.0126973158,
0.0164426025,
-0.0612216294,
-0.0101604601,
0.023478236,
0.0204641614,
0.0098718544,
-0.0226208158,
0.0271585006,
-0.0498549193,
-0.080293417,
0.0086766025,
-0.0257508401,
-0.0737428665,
-0.0129995598,
0.0132172816,
0.0435937904,
0.0093058804,
-0.0451957025,
0.0365771763,
0.0216936786,
-0.0279701594,
-0.0366224684,
0.000616698,
0.0071836207,
0.0229668096,
0.0055592754,
-0.0015637807,
0.0661924407,
0.0445254967,
-0.0260490142,
-0.0179845504,
0.0402512513,
-0.056520015,
0.0429105237,
0.0297586042,
0.0623139553,
-0.0208175294,
0.0116313389,
-0.0564949438,
0.0272791237,
-0.0460856967,
0.01922727,
-0.0288596675,
-0.0201219581,
0.0150178606,
0.0162386689,
-0.0357976109,
0.0612111129,
0.0015682895,
0.0088579739,
0.0403569713,
-0.0011500638,
-0.001049774,
-0.0108884182,
-0.0294068661,
0.0328899324,
-0.0065165344,
0.0405089222,
0.0519630499,
0.0639512986,
0.0160086956,
-0.0010852349,
0.0079798913,
-0.0097528473,
-0.0434154496,
0.0268077478,
-0.0185705591,
0.0127857979,
-0.0045720832,
0.0033789731,
-0.0452433228,
0.0300275479,
0.0905544311,
-0.0321402922,
0.0106604882,
-0.0248273741,
0.0369760394,
-0.0124210054,
0.0351003967,
0.0031183229,
-0.0361895189,
-0.001401622,
-0.0024454049,
-0.0335236825,
0.0553800948,
-0.0185729824,
0.0080059087,
0.0347109102,
0.0091666691,
-0.0625256598,
-0.0312423222,
-0.0145701328,
-0.0712627918,
-0.0313849263,
-0.058206711,
0.0319967121,
0.051824864,
0.0499262214,
0.0003893848,
-0.0020381026,
0.014541544,
-0.0103444653,
-0.0428377055,
-0.0151833622,
-0.0188793577,
-0.0266163703,
0.0289804339,
-0.038505394,
-0.0468827635,
0.0420861244,
-0.0522711985,
0.0444879793,
-0.0512485839,
0.0021544769,
0.0219884384,
0.0492713787,
0.0021686775,
0.0467808023,
0.020066062,
-0.0099297604,
-0.0292267352,
-0.0137158865,
0.0617816634,
-0.0551286601,
0.0061442559,
-0.0627658367,
-0.0100481873,
-0.006805487,
-0.022352213,
-0.0500292778,
0.0138919195,
0.0159428474,
0.0270763934,
0.0800123885,
-0.0300439503,
-0.0078828828,
-0.0300159864,
0.0404833667,
0.0370007828,
0.0127681661,
-0.0044610258,
-0.0392492972,
0.0308402106,
0.0132085057,
-0.0481698923,
0.0028837109,
-0.0050224015,
-0.0071031069,
0.0556218997,
0.0287284665,
-0.0218323618,
0.0019588894,
-0.0102608725,
0.0520545505,
-0.0322847739,
-0.0501789972,
-0.0070905844,
0.0202644505,
-0.0386580639,
-0.0312244128,
0.0636455342,
0.0250109099,
-0.0055927872,
0.0020203178,
-0.0159269087,
-0.0025078438,
0.0229609758,
0.012332676,
-0.0665626377,
-0.0043870681,
0.0219390523,
0.0668918863,
-0.0234644022,
-0.0328850187,
-0.0371754728,
-0.1042488962,
-0.0198451262,
0.0356708579,
-0.0190828983,
-0.0170650017,
0.0007544855,
-0.0129412049,
-0.0337284654,
-0.0264583379,
-0.0278650485,
0.0124979056,
-0.0116918283,
0.0153757725,
0.0026693565,
-0.0277158394,
0.0049103484,
-0.0115707461,
-0.0232258849,
0.0978524163,
0.008317274,
-0.0369145498,
0.0230413266,
0.0222709849,
-0.0171414427,
0.0082637742,
-0.0121328272,
0.053438399,
-0.0053827264,
-0.0257489812,
0.0713623837,
0.0053436416,
-0.0154068107,
0.0075210407,
0.0009883041,
0.0278362595,
-0.0028334595,
0.0214411262,
0.0005288463,
-0.0510308929,
0.0543384105,
0.0638140291,
0.0294473693,
-0.003525093,
-0.0164069664,
-0.0170402862,
-0.0431230552,
-0.0273593012,
-0.0594180375,
-0.0147822592,
-0.062847957,
-0.0873346105,
-0.073209852,
-0.0007863736,
-0.0256167725,
0.0646305457,
0.0332104228,
0.0203010216,
0.031804692,
-0.0175727867,
-0.0520615876,
-0.0368140861,
-0.036788661,
0.0218311846,
0.0059968866,
-0.0159846935,
0.0016866976,
-0.0184344929,
-0.0264399573,
0.0074057668,
-0.0729832575,
-0.0830103904,
0.0191069394,
0.0731448531,
0.0295594484,
0.0451493263,
-0.0218398105,
-0.0564487763,
0.0630245134,
-0.0046442505,
0.0142724048,
-0.0056283632,
0.0689562187,
-0.0492880382,
0.0091606248,
0.0561027266,
-0.0316365473,
-0.007059359,
0.0736661628,
-0.0130232563,
-0.0252597947,
0.0035355312,
0.0362455845,
0.0496494249,
0.0443688557,
-0.0140640168,
0.010457987,
0.0128980391,
0.0037349712,
-0.0080942959,
-0.0271255188,
-0.0517677702,
0.0133132469,
0.0201215241,
0.0134497937,
0.0436115041,
0.0094996393,
-0.0213739481,
-0.0286547858,
0.0162044019,
-0.00573648,
0.0237256363,
0.0047103777,
0.041737996,
0.0223194882,
-0.0135033298,
0.036244221,
-0.0143406428,
0.0338517725,
-0.0334492847,
0.0435074158,
0.0078431731,
0.012564715,
-0.0386533476,
0.0520172492,
-0.0418225937,
-0.0906931013,
0.011349,
0.0223260634,
0.0141577404,
-0.0620259531,
0.0241838228,
-0.0326293632,
0.0012853827,
-0.0241907071,
-0.0076522413,
-0.0102653829,
-0.0128024602,
0.0055144327,
0.0033976468,
-0.0023374616,
0.0267562959,
0.1002638564,
-0.0074493675,
0.0661059469,
-0.0124524813,
-0.0353422463,
-0.0356321745,
-0.0276771933,
0.0443317592,
0.0049738307,
-0.0158347432,
-0.010098407,
-0.0713339373,
-0.080902487,
-0.0033305218,
-0.0025578358,
0.0268915296,
0.0138272932,
0.0197918508,
0.1135761812,
0.0063682208,
0.0220586043,
-0.0943809226,
0.0455743223,
0.0062415255,
-0.035165567,
0.0105532482,
-6.022098702e-33,
-0.0072964304,
-0.0350168385,
-0.049748756,
0.0063378704,
0.0211978666,
0.0219365265,
0.008280525,
0.0189566519,
-0.0148838377,
-0.0060306657,
-0.0221638121,
0.0011559785,
0.0159967449,
-0.0102928681,
-0.0149779627,
0.0070438408,
0.0311828833,
0.0164869502,
-0.0373869389,
0.0171504878,
0.0889725611,
0.026336262,
0.068175517,
-0.0495890528,
0.073146008,
-0.0345784947,
-0.0022575115,
-0.0044580055,
0.0065092989,
0.0206567664,
0.0087736044,
0.0501306057,
0.0047562984,
-0.0474051461,
-0.0355581678,
-0.0289711207,
-0.0027520717,
-0.0085061351,
0.0589561015,
0.005118913,
-0.0053541712,
-0.0473011136,
0.0540163033,
-0.0191123411,
-0.0003587644,
-0.0100031644,
-0.0200222339,
-0.0363591127,
-0.076630041,
0.0299989078,
-0.0649567842,
-0.0009803569,
-0.0064575812,
-0.0104779461,
0.0554193482,
-0.0154492743,
0.001714292,
-0.0708637312,
-0.0078510251,
0.0161355399,
0.0509146787,
0.0543108508,
0.0821167827,
0.0305252094,
0.0088218814,
0.0373686776,
0.0094568264,
-0.0009888354,
0.0589985661,
-0.0118257962,
-0.0055009234,
0.0786068887,
0.037443947,
0.0173848979,
0.0613836013,
-0.0242627393,
0.0177888311,
0.019491259,
-0.0205002073,
0.0378552079,
0.0324072726,
0.0028715599,
-0.069206886,
-0.0035532191,
-0.0293742772,
-0.0541859232,
-0.0350039341,
-0.026533708,
-0.0045509245,
0.0223473329,
-0.0479017422,
0.0267630797,
0.0243977532,
-0.0371250622,
0.0456051491,
0.0107037546,
-0.0208412781,
-0.0393734723,
-0.0188798625,
-0.0083208475,
0.0032681823,
-0.0169188008,
-0.0145983864,
-0.0269403066,
0.0201689098,
0.0111197708,
0.017749263,
0.0237973444,
-0.0246506725,
-0.0273689721,
-0.045940578,
0.0037122054,
0.0072398633,
0.035188403,
0.0477903672,
-0.012507529,
-0.0083505278,
-0.0207504258,
0.0557806827,
-0.0198285636,
-0.0268991403,
0.0503619425,
-0.0048883655,
0.0113418642,
-0.0369546562,
0.0019680287,
0.0020225765,
0.0408733562,
0.0483637974,
0.0057098335,
-0.0306978561,
-0.0098771323,
2.937e-7,
0.0426653624,
0.0002299835,
-0.0007006694,
0.02744334,
0.0153999459,
0.0696073547,
0.0080997897,
0.0681488961,
-0.0261841938,
-0.0413102135,
-0.0128837861,
-0.0058818199,
0.0629028007,
0.0102766408,
-0.0504590273,
-0.0368354395,
0.0118923252,
0.0307370648,
-0.0518762097,
-0.0280246586,
0.0659055412,
0.0739579052,
0.0188950952,
-0.0193636157,
-0.0475396924,
-0.0279155374,
0.0180201139,
-0.0165308956,
-0.0284095239,
0.0058602896,
0.0630531088,
0.0574377328,
0.0000648558,
0.0255705826,
-0.0387187786,
0.0017168806,
0.0227353629,
0.0552523471,
0.0120257838,
-0.0153744873,
-0.031769298,
-0.0688128471,
-0.0086600753,
-0.0205256306,
0.0608466826,
-0.0190000888,
-0.0394564122,
-0.0315842666,
-0.0125279035,
0.0311432015,
0.0069407369,
-0.0104335146,
-0.0069630551,
0.0341332033,
0.0098218555,
0.0212576818,
-0.0153741594,
-0.040293429,
-0.0044681551,
-0.0149606541,
-0.0140889883,
-0.0507760346,
-0.0058684996,
0.0059701866,
-0.0096144127,
0.0007797243,
-0.0031760784,
3.01201119e-34,
-0.0182099082,
-0.0124386474,
0.0152995214,
0.0201003253,
0.0278047752,
-0.0250701457,
-0.0146952933,
0.0065830671,
-0.0304770321,
-0.0321309306,
-0.0145055698
] |
http://arxiv.org/abs/2402.07909v1 | http://arxiv.org/pdf/2402.07909v1 | Prompt4Vis: Prompting Large Language Models with Example Mining and Schema Filtering for Tabular Data Visualization | 2024-01-29T10:23:47 | 2024-01-29T10:23:47 | Shuaimin Li, Xuanang Chen, Yuanfeng Song, Yunze Song, Chen Zhang | cs.HC | cs.HC, cs.AI, cs.CL, cs.DB | Data visualization (DV) systems are increasingly recognized for their
profound capability to uncover insights from vast datasets, gaining attention
across both industry and academia. Crafting data queries is an essential
process within certain declarative visualization languages (DVLs, e.g.,
Vega-Lite, EChart.). The evolution of natural language processing (NLP)
technologies has streamlined the use of natural language interfaces to
visualize tabular data, offering a more accessible and intuitive user
experience. However, current methods for converting natural language questions
into data visualization queries, such as Seq2Vis, ncNet, and RGVisNet, despite
utilizing complex neural network architectures, still fall short of
expectations and have great room for improvement.
Large language models (LLMs) such as ChatGPT and GPT-4, have established new
benchmarks in a variety of NLP tasks, fundamentally altering the landscape of
the field. Inspired by these advancements, we introduce a novel framework,
Prompt4Vis, leveraging LLMs and in-context learning to enhance the performance
of generating data visualization from natural language. Prompt4Vis comprises
two key components: (1) a multi-objective example mining module, designed to
find out the truly effective examples that strengthen the LLM's in-context
learning capabilities for text-to-vis; (2) a schema filtering module, which is
proposed to simplify the schema of the database. Extensive experiments through
5-fold cross-validation on the NVBench dataset demonstrate the superiority of
Prompt4Vis, which notably surpasses the state-of-the-art (SOTA) RGVisNet by
approximately 35.9% and 71.3% on dev and test sets, respectively. To the best
of our knowledge, Prompt4Vis is the first work that introduces in-context
learning into the text-to-vis for generating data visualization queries. | Prompt4Vis: Prompting Large Language Models with Example Mining and Schema Filtering for Tabular Data Visualization: Data visualization (DV) systems are increasingly recognized for their
profound capability to uncover insights from vast datasets, gaining attention
across both industry and academia. Crafting data queries is an essential
process within certain declarative visualization languages (DVLs, e.g.,
Vega-Lite, EChart.). The evolution of natural language processing (NLP)
technologies has streamlined the use of natural language interfaces to
visualize tabular data, offering a more accessible and intuitive user
experience. However, current methods for converting natural language questions
into data visualization queries, such as Seq2Vis, ncNet, and RGVisNet, despite
utilizing complex neural network architectures, still fall short of
expectations and have great room for improvement.
Large language models (LLMs) such as ChatGPT and GPT-4, have established new
benchmarks in a variety of NLP tasks, fundamentally altering the landscape of
the field. Inspired by these advancements, we introduce a novel framework,
Prompt4Vis, leveraging LLMs and in-context learning to enhance the performance
of generating data visualization from natural language. Prompt4Vis comprises
two key components: (1) a multi-objective example mining module, designed to
find out the truly effective examples that strengthen the LLM's in-context
learning capabilities for text-to-vis; (2) a schema filtering module, which is
proposed to simplify the schema of the database. Extensive experiments through
5-fold cross-validation on the NVBench dataset demonstrate the superiority of
Prompt4Vis, which notably surpasses the state-of-the-art (SOTA) RGVisNet by
approximately 35.9% and 71.3% on dev and test sets, respectively. To the best
of our knowledge, Prompt4Vis is the first work that introduces in-context
learning into the text-to-vis for generating data visualization queries. | [
0.015406603,
0.0303108227,
-0.0373356901,
0.0469280407,
-0.0193101205,
0.0174735468,
-0.0219189804,
0.0042331209,
0.0276107546,
0.0103708776,
-0.0083368681,
0.0090455972,
-0.0355602391,
0.0704930052,
0.0193459112,
-0.0172594003,
0.0506016985,
0.0155926887,
-0.0138219614,
0.0619902797,
0.0151757039,
0.0166381188,
-0.0090952553,
0.055054605,
-0.0033969004,
-0.041446127,
-0.0318541713,
0.0190390982,
0.0016933135,
-0.0572428815,
-0.0117636034,
0.0490360782,
0.0408844016,
0.0251506511,
0.0000022999,
-0.0481275842,
-0.0246096756,
-0.00892343,
-0.0274900422,
-0.0129535655,
0.0404864773,
-0.0074170884,
0.0189509876,
0.0197772179,
-0.0681565776,
-0.0443735607,
0.0278147124,
0.0434640385,
0.0955594108,
0.024110036,
-0.0194341969,
-0.0037877588,
0.0079860045,
-0.0151511971,
0.0292180106,
0.0188188441,
-0.0346681252,
0.005701235,
0.0122216549,
0.0494314134,
-0.0095470371,
0.0408441015,
0.0249377433,
-0.021315651,
0.0044576386,
0.0409718677,
-0.0400155224,
-0.0040556402,
0.0082588214,
0.1090747789,
-0.0336780809,
0.0078777783,
-0.0093782982,
-0.0047660568,
0.0180293862,
0.0142728528,
-0.058100801,
-0.0136978403,
-0.019792648,
0.0004204436,
0.0055139051,
0.0199711286,
-0.0011867261,
-0.030546017,
-0.0047306428,
0.0050557386,
-0.0178175867,
0.001250239,
0.0560855493,
-0.0244386308,
-0.011061714,
-0.0329682194,
0.0232561976,
0.0224619657,
-0.0739784539,
0.0236472767,
-0.0011416458,
-0.0640035719,
0.0168378521,
-0.0099660521,
0.0665680543,
0.0316147618,
-0.0313499346,
0.0213147923,
-0.0102841007,
0.0165230129,
0.0260700472,
0.0281312857,
-0.0034554573,
-0.028056683,
-0.0472532772,
0.0019281763,
-0.0790365264,
-0.0264191013,
-0.0669874027,
-0.0061960421,
-0.0146331564,
0.027573958,
0.0088502057,
0.0283671599,
0.0114399763,
0.0272790473,
-0.0086834729,
0.0437433682,
-0.0792432427,
0.0028968991,
-0.0346958824,
-0.0295152236,
0.0020997741,
-0.0807300881,
-0.0043322509,
-0.0107930088,
0.0157230943,
-0.0391676128,
-0.0019886736,
0.0549959801,
0.0164261889,
-0.0051856898,
-0.0550443269,
-0.0424500443,
0.002687678,
-0.0127796978,
0.0079012392,
-0.0597182438,
0.0408799425,
-0.0055634542,
-0.0024993126,
-0.0109019335,
0.0004192159,
0.0668928996,
-0.0082297642,
0.0890985131,
-0.0326731876,
-0.013700583,
-0.0006946435,
-0.0195709243,
-0.016830612,
0.0559774078,
0.0356440954,
-0.0029834288,
0.0103883576,
-0.0335950963,
0.0220713299,
0.0339730382,
0.0248559583,
-0.0006601431,
-0.0472987667,
0.0325539857,
-0.0010085782,
0.0421367958,
0.01370537,
0.0210918225,
-0.0157657508,
-0.0090684285,
0.0347020961,
0.0660468116,
0.0372219905,
0.0128653459,
0.0351528153,
0.0407154523,
0.0109349666,
-0.0340462886,
-0.003553167,
0.0237656068,
-0.0006143986,
-0.0133780958,
-0.024571253,
0.0495588668,
-0.0002587764,
-0.0687211603,
0.0166475419,
0.0420874842,
0.017348025,
0.0045189634,
0.0178318322,
0.0322486758,
-0.0395677276,
-0.0161106233,
0.0363830104,
-0.0635441691,
-0.023711862,
-0.0233208071,
0.0576910637,
0.0366348997,
0.0586742423,
-0.0606849976,
-0.0177479312,
-0.0154145658,
0.0217099618,
0.046506118,
0.0594241768,
-0.0248817205,
0.002077745,
-0.0342109092,
-0.0368838795,
0.0052010366,
-0.0039541251,
-0.0195251219,
0.0148076564,
0.01412246,
-0.003107711,
-0.0553395562,
-0.0799372271,
0.0386177748,
0.0177183952,
-0.0041505732,
0.0370188169,
0.015726475,
-0.0098976372,
0.0445152186,
0.0063091717,
0.0358806401,
0.0447799377,
-0.0346413553,
0.0295289345,
0.0016172766,
-0.0201995969,
-0.1183192134,
-0.0168731175,
-0.0429447368,
-0.009730638,
-0.0098082963,
-0.0719617829,
-0.0227973331,
-0.0358120836,
-0.035679277,
-0.0035966095,
-0.0188533626,
-0.0401057377,
0.0069621219,
0.0264742523,
-0.0596717969,
0.0718049258,
-0.000011458,
0.0250488855,
0.0103702573,
-0.0011177633,
0.0042753164,
-0.0090816319,
0.0144158332,
0.0037356261,
0.0021783845,
0.0112403985,
0.013614784,
0.0110612549,
-0.0768449455,
0.0341589563,
-0.0829150602,
0.0312100556,
-0.0067882296,
0.0131525267,
0.0070159915,
0.0109083895,
-0.0431645066,
-0.0604750961,
-0.0684571639,
0.0275161471,
0.022055313,
-0.0094133168,
0.0792939514,
-0.0081695681,
-0.009200355,
-0.0157627352,
-0.0086139832,
-0.0400606059,
0.0508413874,
0.0070037763,
-0.035694018,
-0.0039953296,
0.0322236903,
0.0038518892,
0.0074006035,
0.0115558812,
0.0482206568,
0.0347928144,
-0.016139783,
0.041964706,
0.0101245632,
-0.0102262702,
0.0620125383,
0.0040787682,
-0.0230292231,
-0.0591931418,
0.0462852232,
-0.0374069251,
0.026597403,
0.0088024922,
-0.0289767608,
0.0644260421,
0.0083105154,
0.0065637249,
-0.0145457806,
-0.0202549342,
0.0413036346,
0.0078766244,
0.0159780644,
-0.0864050165,
0.0195849352,
0.0238557756,
0.0236392468,
-0.0203522313,
-0.0301632322,
0.025972886,
-0.0347600468,
-0.0113857808,
-0.0577101782,
-0.025448665,
-0.0521015897,
0.0974807739,
0.0316076651,
-0.003381138,
0.0221078098,
-0.0262542218,
-0.0393514447,
0.0061060875,
-0.0293026082,
-0.0371367894,
0.071571812,
-0.0324660055,
-0.0186949819,
-0.0107950848,
0.0009446624,
-0.0535311699,
-0.0234741662,
-0.0025408689,
0.0126979845,
0.0202315822,
-0.0026776462,
-0.0230007749,
-0.0110354843,
0.0059446804,
0.0534157008,
0.000725863,
-0.0555194095,
-0.025867749,
0.0450051092,
0.0303951278,
0.0307427552,
0.004405316,
-0.010129354,
0.0492547452,
0.0464187488,
-0.0414793231,
-0.0031979659,
-0.0236446001,
0.0306546651,
-0.0156172682,
-0.0027697913,
0.0092045078,
0.0231566615,
0.0233871043,
-0.0645478964,
-0.0096249115,
0.0861108974,
0.0555019565,
-0.0373196527,
0.0261867475,
0.007068282,
0.0486970469,
0.0495358929,
0.0636120066,
-0.104360655,
-0.0403555594,
-0.0061718817,
0.0002588645,
0.0063044019,
-0.0709534287,
-0.0521302223,
-0.1204986274,
0.0858810097,
-0.0237758849,
-0.0822147653,
-0.0029368515,
-0.0515713133,
0.0027244024,
-0.029150892,
-0.0012846773,
-0.0081875762,
-0.003514335,
0.0303791761,
0.003708777,
-0.0157396663,
-0.0352081209,
0.0034736507,
-0.0590536036,
0.0020050691,
0.0214539189,
0.0604814105,
-0.007020371,
-0.0171635095,
0.0064900955,
0.0202041455,
0.0265867412,
-0.0315105841,
0.1075676978,
0.0278350841,
-0.0255222581,
-0.045884002,
0.0140009047,
-0.0354789197,
-0.0177418552,
0.0225592237,
0.0771047473,
-0.0447839238,
0.0213440377,
-0.036783509,
-0.012518202,
0.0366813429,
0.0668215454,
0.0613199323,
-0.04471609,
-0.0152641395,
-0.0290782619,
0.0062879259,
-0.0425314866,
-0.0474833027,
0.0140223736,
0.0123368008,
-0.0201933216,
-0.052460175,
-0.023384532,
0.073487334,
0.0381116495,
0.00537208,
0.0370292068,
0.0020019319,
0.0008735157,
-0.0395520777,
0.0423353463,
-0.0250691921,
0.0253311209,
-0.0110184317,
-0.0165847037,
-0.0100606317,
-0.0244915318,
-0.039313525,
0.0103129549,
-0.0527913682,
-0.011348322,
-0.026066022,
-0.0182012226,
0.0135583812,
0.0173782464,
-0.0055973786,
-0.0010195344,
0.0331791379,
0.113683708,
-0.0416102037,
-0.0881615058,
-0.0365973301,
0.0160793122,
0.0026162672,
0.0075072828,
0.0165400282,
0.0263587106,
-0.029912455,
0.040800631,
-0.0103648137,
0.0026006238,
0.0493867137,
0.0083801374,
0.0275526624,
0.0262271874,
-0.0468187258,
-0.073793456,
-0.0199796669,
-0.0098138126,
-0.0425481498,
-0.0485172197,
0.0268945917,
-0.0072426824,
-0.0453982651,
0.010537616,
-0.0139934123,
0.0033259159,
0.027978139,
-0.0682865158,
0.0123460153,
0.0449579023,
0.0111955507,
0.022516612,
0.0201291014,
0.0001706108,
0.0072427997,
-0.0118877431,
0.0356488898,
-0.0182402693,
0.05366138,
0.0167825073,
-0.0226355866,
-0.0089585632,
0.0341241919,
-0.0593620837,
-0.0695535243,
-0.0016229389,
0.0324930884,
-0.0048335623,
-0.0243200194,
0.0082126576,
-0.0214830004,
0.0015658552,
0.0039389576,
0.004699457,
0.02649335,
0.005155148,
0.0643059388,
-0.0091520473,
-0.0305327307,
-0.0805712864,
0.0434133485,
-0.0317855813,
0.0231040809,
0.0103449831,
-0.0452254601,
-0.0492209308,
0.0242200475,
0.0460031405,
0.005569824,
-0.011478941,
0.0237672664,
-0.0530008264,
-0.0471649505,
-0.0105614169,
0.0059439233,
-0.0482805967,
0.0301951468,
0.0423902795,
0.0447488092,
0.0393081717,
-0.032413993,
-0.0167601705,
0.0329390578,
-0.0458748527,
-0.0906432718,
-0.0059279497,
-6.481939602e-33,
-0.0297457632,
-0.065665856,
0.0408224054,
-0.0264949743,
0.0080346018,
-0.0073527992,
-0.0488170274,
-0.0154041676,
-0.0272444878,
0.0160906557,
-0.0398349799,
0.0186263062,
0.0123307221,
0.0263006445,
-0.019742392,
-0.031600371,
0.0350334644,
0.0064876373,
0.0063611586,
0.0377849564,
0.0126371542,
0.0468062796,
0.0380522646,
-0.0762695596,
0.0508150682,
0.036366377,
0.0074998979,
0.0277785324,
-0.0181878917,
0.0395584367,
-0.0414694473,
0.0551390573,
0.0231836941,
-0.0445817672,
-0.0465863012,
0.0469051413,
-0.0882342085,
-0.0639923364,
-0.0117151337,
0.0086534768,
-0.0811393932,
0.0056507187,
0.0347755849,
-0.0122585474,
0.0020063673,
0.0101201497,
0.0500101373,
-0.0205321517,
-0.016981082,
0.0002873713,
-0.0145297442,
0.0353321359,
-0.0127379633,
-0.0025764208,
-0.0000170102,
0.0794629455,
0.0114216749,
0.0534630083,
-0.0696549341,
0.0334736742,
0.0207889006,
-0.0091538616,
0.0333170183,
0.0054274318,
0.0205553398,
0.0744164065,
0.0512248762,
-0.0072549647,
-0.0006335273,
0.0256852731,
0.0044851946,
0.0721409619,
0.0500490479,
-0.0230316427,
0.0388743505,
-0.0006279378,
-0.0647716224,
0.0523160286,
-0.0024573116,
-0.0136871571,
0.0002952596,
-0.0194752067,
-0.0339692421,
-0.0280091316,
-0.0191332176,
-0.0365619399,
-0.0387684405,
-0.0055923588,
0.0495167263,
-0.0305087566,
0.0129396357,
0.0057961927,
0.0020669156,
-0.0150885088,
-0.0514030978,
0.0051200828,
-0.0193259418,
-0.010032462,
0.0025668608,
-0.0080405883,
-0.0102658784,
-0.010019619,
-0.0008664538,
-0.0279539768,
0.019980954,
0.021047879,
0.0621113926,
0.0089510968,
-0.0787490755,
0.0087622227,
-0.0255825892,
0.024822196,
0.0072419746,
0.030215539,
0.0086242184,
-0.0143499253,
-0.0307802614,
0.0644879565,
0.0407781042,
-0.0126598589,
-0.0246375408,
-0.0255699437,
-0.02040378,
0.0042615631,
0.0170061514,
-0.0122360084,
-0.0055462294,
-0.0012662694,
0.0862773284,
-0.0552119948,
-0.0076157609,
-0.0125159007,
2.811e-7,
0.0720097199,
0.0512780771,
0.0017608365,
0.0358603559,
-0.0080676237,
0.0295092054,
0.0049105575,
0.0147290817,
-0.0170056988,
-0.0044516819,
-0.0032527,
0.0153375156,
0.0250087846,
0.0730594993,
-0.0773672238,
-0.0240908563,
-0.0461412929,
-0.0421967022,
-0.0781460926,
0.0491007902,
0.1063534617,
0.0784410462,
-0.0073456191,
-0.0355547331,
0.0174211189,
-0.0459824763,
0.015998695,
-0.0582017265,
0.0131683713,
0.0230313502,
0.0206346978,
0.0157814231,
0.0170318279,
0.0321457461,
-0.0016950795,
-0.00241002,
0.0073195416,
0.0220000539,
0.0162813701,
0.030532077,
0.0210943129,
-0.0145228542,
0.0092724608,
-0.0534557961,
0.073613137,
-0.0600685924,
-0.0286189262,
0.0145888906,
0.0373419784,
-0.0143061737,
0.0434697308,
-0.0164306946,
-0.0352575704,
0.0020631067,
0.014507127,
-0.0240024813,
-0.0187912825,
-0.0340649225,
-0.0487762354,
0.0244702566,
-0.0047630742,
-0.0254689567,
-0.0437389351,
0.033930473,
-0.0188666042,
-0.0111408653,
0.0114039164,
3.408769921e-34,
0.0489947423,
0.0229169577,
0.0311063435,
0.0204606522,
0.0622347631,
-0.0095826965,
0.0384255536,
-0.0028247072,
-0.0041250181,
-0.0985429883,
-0.0398207381
] |
http://arxiv.org/abs/2401.16458v1 | http://arxiv.org/pdf/2401.16458v1 | Credit Risk Meets Large Language Models: Building a Risk Indicator from Loan Descriptions in P2P Lending | 2024-01-29T10:11:05 | 2024-01-29T10:11:05 | Mario Sanz-Guerrero, Javier Arroyo | q-fin.RM | q-fin.RM, cs.AI, cs.CL, cs.LG | Peer-to-peer (P2P) lending has emerged as a distinctive financing mechanism,
linking borrowers with lenders through online platforms. However, P2P lending
faces the challenge of information asymmetry, as lenders often lack sufficient
data to assess the creditworthiness of borrowers. This paper proposes a novel
approach to address this issue by leveraging the textual descriptions provided
by borrowers during the loan application process. Our methodology involves
processing these textual descriptions using a Large Language Model (LLM), a
powerful tool capable of discerning patterns and semantics within the text.
Transfer learning is applied to adapt the LLM to the specific task at hand.
Our results derived from the analysis of the Lending Club dataset show that
the risk score generated by BERT, a widely used LLM, significantly improves the
performance of credit risk classifiers. However, the inherent opacity of
LLM-based systems, coupled with uncertainties about potential biases,
underscores critical considerations for regulatory frameworks and engenders
trust-related concerns among end-users, opening new avenues for future research
in the dynamic landscape of P2P lending and artificial intelligence. | Credit Risk Meets Large Language Models: Building a Risk Indicator from Loan Descriptions in P2P Lending: Peer-to-peer (P2P) lending has emerged as a distinctive financing mechanism,
linking borrowers with lenders through online platforms. However, P2P lending
faces the challenge of information asymmetry, as lenders often lack sufficient
data to assess the creditworthiness of borrowers. This paper proposes a novel
approach to address this issue by leveraging the textual descriptions provided
by borrowers during the loan application process. Our methodology involves
processing these textual descriptions using a Large Language Model (LLM), a
powerful tool capable of discerning patterns and semantics within the text.
Transfer learning is applied to adapt the LLM to the specific task at hand.
Our results derived from the analysis of the Lending Club dataset show that
the risk score generated by BERT, a widely used LLM, significantly improves the
performance of credit risk classifiers. However, the inherent opacity of
LLM-based systems, coupled with uncertainties about potential biases,
underscores critical considerations for regulatory frameworks and engenders
trust-related concerns among end-users, opening new avenues for future research
in the dynamic landscape of P2P lending and artificial intelligence. | [
0.0329361334,
0.0778879002,
-0.0216098856,
0.0392001793,
-0.0200797264,
0.0326284952,
0.0014320272,
0.0509288423,
0.0304018911,
-0.0295235086,
-0.0350294448,
-0.0422290564,
0.0089420946,
0.0274193455,
0.0157082938,
-0.0100464849,
0.0195310321,
-0.0022111372,
0.0522635207,
-0.0104344878,
0.0048544891,
0.0165773407,
0.0243520103,
0.0193978623,
-0.0330402777,
-0.0170526467,
0.0114679486,
-0.0252757613,
0.012631963,
-0.0368188806,
-0.0032531496,
0.0325706936,
0.0006015155,
0.0198917314,
0.0000018307,
-0.0674466118,
-0.0432826281,
-0.0017706987,
-0.0122624822,
0.0136239128,
0.0236631315,
0.0058952663,
0.038383171,
0.0039880029,
-0.0796076134,
-0.0111086657,
0.0591924079,
0.0432083793,
0.0378474668,
0.0243947953,
-0.0316073932,
-0.0508272313,
-0.006147786,
-0.0144833177,
0.0370386392,
-0.0259857252,
0.0861307308,
-0.0014107616,
-0.0160852354,
-0.0415625088,
-0.0373340696,
0.0049123801,
-0.0145493057,
-0.0356754325,
0.0573294759,
0.0435191169,
-0.0167951435,
-0.0025888782,
-0.0691501051,
0.0474085547,
0.0219043326,
0.0321093276,
-0.009917588,
-0.0142381536,
0.0178712234,
0.0883979723,
-0.0541072115,
0.0597146861,
-0.0066699483,
0.0086599011,
-0.0374294296,
-0.0028425881,
0.0544482693,
0.0285449773,
-0.0421663336,
0.0843385011,
-0.0028773309,
-0.0130208172,
0.0236734953,
0.0225777123,
0.0140282856,
-0.0680297166,
0.0644001737,
0.0210632216,
0.0337070115,
0.0101459501,
0.0223112106,
0.0185926575,
0.0275187306,
-0.0520263836,
0.0443844646,
-0.030025145,
-0.0100893443,
0.0514570773,
-0.0144369304,
0.0011914965,
0.0389795192,
-0.0059547876,
-0.0199590195,
0.0185479149,
-0.0674485192,
0.0196906812,
-0.0205490179,
0.0788900927,
-0.0009216436,
0.0005269888,
-0.0672281533,
-0.0019752244,
-0.0016790061,
-0.0171460453,
0.007149139,
-0.0289240945,
0.0039943964,
0.0245484505,
-0.0353847407,
-0.0332346782,
-0.0032730668,
0.0321431533,
-0.0006862124,
0.0054826951,
0.0262115858,
0.0033802665,
0.011140733,
-0.0764309466,
0.0405348092,
0.061067123,
0.0663266033,
-0.0409940593,
-0.027908016,
0.0130710555,
-0.0423395857,
0.0095750615,
-0.0379376486,
-0.0136192152,
0.0061201826,
0.0162722021,
-0.0073657976,
-0.0200487524,
-0.0113919964,
0.006649076,
-0.0799789429,
0.0725430548,
-0.0205740314,
0.0393402949,
0.00074752,
-0.0185686927,
-0.0317611732,
-0.0598129891,
0.0526529029,
0.0441132747,
0.0513432883,
-0.0404493846,
0.0308797378,
0.0549620949,
-0.0120521123,
0.0044936677,
0.0358891077,
0.0129931178,
0.0201547612,
-0.0630720854,
-0.0071208435,
0.0568070374,
0.0033613492,
-0.0130294459,
0.0953012928,
0.0705323741,
0.0187339913,
0.0342113599,
0.0066417572,
0.0665626526,
-0.0260417648,
-0.0351306163,
0.0001584673,
-0.0118006961,
-0.0397858806,
-0.0438581184,
-0.009219381,
0.0138243027,
-0.0176035576,
0.0249169227,
-0.0043630498,
-0.0207532533,
0.0172573347,
0.0241880361,
0.0191105027,
0.0043642521,
-0.024789989,
0.0620934814,
0.0646735579,
0.0203253571,
0.0190007705,
-0.0602873601,
0.0744526535,
0.0568016917,
0.0310094934,
-0.0092522157,
0.0863956288,
-0.0460506491,
0.0091467537,
-0.0026687719,
0.0141715687,
-0.001444153,
0.0110887969,
-0.0110125374,
-0.0284499303,
0.0496431589,
-0.0009614849,
-0.0050690481,
-0.0153175611,
0.0073404955,
0.015712671,
-0.00710967,
-0.0411291942,
0.0377242267,
0.0585267507,
-0.0132423723,
0.0408364795,
-0.0096285241,
0.0009145496,
-0.0383019671,
0.0149025936,
-0.01123848,
0.0332095735,
0.004123467,
-0.0553558357,
-0.0134677133,
-0.0216785911,
0.0454289913,
0.0272337329,
-0.0403416455,
0.0094862292,
0.0053231856,
-0.0298548397,
0.0526819564,
-0.0159422029,
-0.0109696006,
0.0069659483,
-0.0351348855,
0.0234897286,
-0.0001620464,
0.0429887287,
-0.0117548695,
0.0456095748,
0.0073597492,
0.0090411399,
-0.0612585396,
-0.0207478348,
-0.0073703271,
0.0255642887,
0.0478149727,
-0.0125797652,
-0.058516793,
0.0172482524,
0.0026941325,
0.0323622301,
0.0244585164,
0.0032644849,
0.0056996732,
-0.0335653126,
0.0118470164,
0.0382502116,
-0.0099069355,
-0.0164280273,
-0.0155163705,
0.022103738,
-0.0434838161,
0.0198304113,
-0.0252688471,
0.0175759234,
0.0769521147,
0.0253550392,
-0.042667985,
-0.0125670163,
-0.0199595224,
-0.0121142147,
0.0524366871,
0.0587154664,
-0.0287165362,
-0.0481351689,
-0.0198054127,
0.0069549354,
-0.0519269779,
0.0241351333,
0.0424826778,
-0.0432806462,
0.0175821036,
-0.0331937149,
-0.0135364756,
-0.0062392242,
0.0590648353,
0.0221472438,
-0.0155037977,
0.0031639722,
-0.000789047,
-0.0139825027,
-0.0220491141,
-0.0333112702,
-0.0483731963,
0.0244789384,
0.0427249223,
0.0073136603,
-0.0106731271,
-0.0238684453,
-0.0179581959,
-0.0268979054,
-0.0035108458,
-0.0120603228,
-0.0232332274,
-0.0215579867,
0.0237779617,
-0.0040205754,
-0.0219425168,
0.0494253747,
-0.0121318614,
0.0216428265,
0.019782668,
-0.0772184134,
-0.0316896178,
0.0351774208,
-0.0273195505,
0.0202951264,
-0.0092108427,
-0.0249949731,
-0.0000147226,
-0.0292975046,
0.0438365601,
-0.0055549424,
0.0215959623,
-0.0226237588,
-0.0100321062,
0.0177501328,
-0.055435732,
0.0012404326,
-0.0952472463,
0.0177291557,
0.0316208042,
0.0272761453,
-0.0225885529,
-0.0468503498,
0.0248974338,
-0.0399184972,
-0.0096413093,
-0.024881633,
-0.0242841523,
0.0248355959,
0.0242091641,
0.0109981075,
-0.0753122121,
0.0462616049,
-0.0693457574,
0.0114102699,
0.0073882942,
-0.0533232801,
0.021156542,
0.0133716743,
0.0241753515,
0.0056710308,
-0.0643270388,
-0.0071284547,
0.0200951491,
-0.0068421317,
-0.0615705028,
-0.0261519663,
0.0575217195,
0.0325409435,
-0.015680261,
0.0409056135,
-0.0205418896,
-0.0051723826,
-0.0134583069,
0.0447188988,
-0.0772980228,
-0.0165980309,
-0.031017324,
0.0300169829,
0.0148579143,
-0.0573756844,
-0.0061163688,
-0.0557934716,
0.0507437028,
-0.0333317891,
-0.0147665897,
-0.0365006179,
-0.0470442884,
-0.0045827734,
-0.0512453765,
-0.0053508924,
-0.0205231663,
-0.053218659,
-0.0025060205,
0.002535698,
-0.0417414866,
0.0120409736,
0.0459739715,
-0.0583353341,
0.0007962025,
0.0252590496,
0.0033931981,
-0.0379611365,
0.0074953102,
-0.0219413936,
-0.0147235543,
-0.0410829894,
0.0084182769,
0.0436223559,
-0.0055077295,
0.0029048894,
0.0121908849,
0.0240759607,
-0.0470564291,
0.0072470726,
0.0395557322,
0.0445894562,
0.0174719356,
-0.0653904304,
0.0120898616,
-0.0221949741,
0.046240814,
0.1052533016,
0.0005706302,
-0.0720269904,
0.0033934091,
-0.0346686095,
-0.0890278369,
0.0410763212,
0.0006515658,
0.0354505777,
-0.0234701317,
-0.0050845398,
-0.0429279804,
-0.0445523225,
0.0150777055,
-0.0281430036,
0.0524008609,
0.0314539112,
0.065127939,
-0.0383214615,
-0.0457931347,
0.0124744959,
-0.0618469007,
0.0427125916,
-0.0081116622,
0.0611653067,
-0.0105781592,
-0.0263360888,
-0.0393032096,
-0.0021949324,
-0.0448820181,
-0.0202914532,
-0.0183359571,
0.0333361253,
0.0186769292,
-0.0390675478,
-0.026480481,
-0.0519037619,
0.0289000832,
0.0571384132,
-0.062148165,
-0.0351178907,
0.0114073576,
-0.0123640783,
-0.0114620505,
0.0429645255,
-0.0239655133,
0.0174459778,
0.0374831669,
-0.0381977931,
-0.0485925451,
-0.0029693418,
0.0081551541,
0.0466205887,
0.0334436819,
0.0037356783,
-0.0423690788,
-0.0250941515,
-0.0386014655,
0.0236887895,
0.00718879,
-0.0337338336,
-0.0268235337,
0.0827736109,
-0.0079200212,
-0.0064409981,
0.0127351414,
-0.0149421822,
-0.0157011393,
-0.0315984264,
0.066864796,
0.0335401185,
0.0472382866,
0.0381187648,
-0.030404726,
0.0687754974,
0.0040594977,
0.0444125794,
-0.0012190728,
0.0195731539,
-0.0168008152,
0.0259406622,
0.0144168604,
0.0195123833,
-0.0244424026,
-0.0168080255,
-0.005874678,
-0.0075538531,
0.0142735681,
0.0065577403,
-0.0258272942,
0.0034613556,
-0.0345486403,
-0.0123274513,
-0.0458665155,
0.0303104147,
-0.0259271543,
-0.0751194358,
0.0275939107,
-0.0115279844,
-0.0581876189,
0.0190197006,
0.0691827014,
0.0087251188,
-0.0015889416,
-0.0330837183,
-0.0363399126,
-0.0318891257,
-0.0218984801,
0.0603634268,
-0.0243943669,
-0.0473513901,
0.0101637179,
-0.0616893955,
-0.0467885248,
-0.0138121974,
-0.0795802027,
0.060000062,
-0.0030294836,
0.0037747447,
0.0207278859,
-0.0156115796,
0.0134667847,
-0.0433544032,
-0.0081694862,
-0.0253560375,
-0.0235058926,
-0.0662825406,
-6.113256819e-33,
-0.0133047728,
-0.0600314811,
-0.0050643818,
0.0310817771,
0.0219153129,
0.0367671475,
0.0210556518,
-0.009336221,
0.0189385619,
0.027023688,
-0.0114779109,
0.0193956885,
-0.0023949537,
0.0225944482,
0.0251627341,
0.0241912622,
0.0028423099,
-0.0094432021,
-0.0146127474,
-0.0023528165,
-0.0136521496,
0.0460329279,
0.0635770261,
-0.08771649,
0.0204542652,
0.0056522377,
-0.0255992748,
0.00707972,
0.0194533747,
0.0369107649,
-0.0176077373,
-0.0014026384,
0.0230837855,
-0.0554905832,
0.024306437,
0.02121781,
-0.0221956391,
0.0327347554,
-0.0065482538,
0.0199906547,
0.0386442915,
-0.0347288772,
0.0401599929,
-0.0199532751,
-0.0106111169,
-0.0110699898,
0.0376988128,
-0.0702443197,
-0.0148726096,
-0.0389808454,
-0.028912954,
0.0498153083,
-0.0329895541,
0.0222743601,
0.0429331325,
0.0649191961,
0.0227497257,
0.0240443051,
-0.0112739792,
0.0021822688,
-0.0257875845,
-0.0171610452,
0.0434872694,
-0.0237332974,
-0.0250633713,
0.0480479598,
-0.0264917426,
-0.0703054145,
0.0175925605,
-0.0369878411,
-0.0167560875,
0.0870754868,
0.0298958942,
0.056003999,
0.0208838545,
-0.0316035263,
-0.0928073749,
0.0230958164,
0.0387709029,
0.00064501,
0.0496915355,
-0.0031612357,
-0.038105607,
-0.0314692333,
-0.0260831919,
-0.075092487,
-0.0148132937,
-0.0015257032,
0.0268460885,
-0.042403467,
-0.003719046,
0.0488555916,
0.0032422,
-0.0834164321,
-0.1266786903,
-0.0005703062,
-0.0118099609,
-0.039021723,
-0.0092202742,
0.0205299538,
-0.0466793999,
-0.04204759,
0.0048215636,
-0.0124073094,
0.0409519337,
0.016389342,
0.0886844546,
0.002774928,
-0.0417298228,
-0.0205375161,
0.0180846807,
-0.0080951042,
0.0085957274,
0.1062073931,
0.0299780071,
0.0183834396,
0.0151644265,
0.0420675203,
0.0281731989,
0.046793852,
0.0123275314,
-0.0479965694,
-0.0447093695,
0.0186163224,
-0.0016868621,
-0.0039036612,
-0.0118423058,
0.0178289115,
0.0682140887,
-0.0530360676,
-0.0177914165,
0.0050635356,
2.657e-7,
0.0471776016,
0.0832655579,
-0.094811976,
-0.0200629774,
-0.0251274854,
0.0388733521,
-0.0204320047,
0.0640698969,
0.0128064305,
-0.0063773948,
0.0003595107,
0.0339063853,
0.0703095943,
0.0110201752,
-0.0831130445,
-0.0264175497,
-0.0825604275,
-0.0080610719,
-0.0768529177,
0.0046810932,
0.0347121134,
0.0851082578,
0.0366228968,
-0.0672359318,
0.0184241198,
-0.0783943608,
0.0186438281,
-0.0588879921,
0.0010036827,
0.0332780071,
0.0266149472,
0.028727524,
0.011794501,
0.0670165792,
-0.0268759895,
0.021778699,
0.0138206454,
0.0424137861,
0.0114067327,
0.0303883199,
0.0010112462,
-0.0049011256,
0.0014295386,
-0.0282337181,
0.0243134927,
-0.0520513877,
-0.0550519526,
-0.0346012563,
-0.0220474862,
-0.0195043851,
0.0316370316,
-0.0041955961,
0.0291890688,
0.0192401111,
0.020456763,
-0.0566646419,
-0.0031626425,
-0.0243139155,
0.0242477879,
-0.006507121,
0.0086369654,
-0.0049317349,
-0.0057459101,
0.0298785809,
0.0056160921,
-0.070587717,
0.0555614941,
2.340471241e-34,
-0.0079234671,
-0.0238814652,
-0.0167871527,
0.0135986349,
0.0461929478,
0.0110950274,
0.0108431531,
0.0348469503,
-0.0258311983,
-0.065600805,
0.0375205986
] |
http://arxiv.org/abs/2401.16012v1 | http://arxiv.org/pdf/2401.16012v1 | Finding Challenging Metaphors that Confuse Pretrained Language Models | 2024-01-29T10:00:54 | 2024-01-29T10:00:54 | Yucheng Li, Frank Guerin, Chenghua Lin | cs.CL | cs.CL | Metaphors are considered to pose challenges for a wide spectrum of NLP tasks.
This gives rise to the area of computational metaphor processing. However, it
remains unclear what types of metaphors challenge current state-of-the-art
models. In this paper, we test various NLP models on the VUA metaphor dataset
and quantify to what extent metaphors affect models' performance on various
downstream tasks. Analysis reveals that VUA includes a large number of
metaphors that pose little difficulty to downstream tasks. We would like to
shift the attention of researchers away from these metaphors to instead focus
on challenging metaphors. To identify hard metaphors, we propose an automatic
pipeline that identifies metaphors that challenge a particular model. Our
analysis demonstrates that our detected hard metaphors contrast significantly
with VUA and reduce the accuracy of machine translation by 16\%, QA performance
by 4\%, NLI by 7\%, and metaphor identification recall by over 14\% for various
popular NLP systems. | Finding Challenging Metaphors that Confuse Pretrained Language Models: Metaphors are considered to pose challenges for a wide spectrum of NLP tasks.
This gives rise to the area of computational metaphor processing. However, it
remains unclear what types of metaphors challenge current state-of-the-art
models. In this paper, we test various NLP models on the VUA metaphor dataset
and quantify to what extent metaphors affect models' performance on various
downstream tasks. Analysis reveals that VUA includes a large number of
metaphors that pose little difficulty to downstream tasks. We would like to
shift the attention of researchers away from these metaphors to instead focus
on challenging metaphors. To identify hard metaphors, we propose an automatic
pipeline that identifies metaphors that challenge a particular model. Our
analysis demonstrates that our detected hard metaphors contrast significantly
with VUA and reduce the accuracy of machine translation by 16\%, QA performance
by 4\%, NLI by 7\%, and metaphor identification recall by over 14\% for various
popular NLP systems. | [
0.063898325,
0.030603433,
-0.0607014634,
0.013756129,
-0.0233576521,
0.0336124152,
-0.0178152267,
-0.0066775773,
0.0104883844,
-0.0670887306,
-0.0088868905,
-0.0117438231,
0.0150298774,
-0.0091583068,
0.0104816034,
-0.0683626905,
0.0356421396,
0.0142356483,
-0.019573601,
0.0014016501,
-0.0543842651,
0.0396503694,
-0.001580399,
0.0813232064,
-0.0104271928,
-0.0354158916,
0.0325840265,
0.0395447724,
-0.0117289731,
-0.0379023217,
0.0062176869,
0.0233149529,
0.0162881892,
0.026659755,
0.0000020506,
-0.0357664973,
-0.0003211058,
-0.0254374575,
0.0107193161,
-0.0096237501,
0.0859713852,
-0.014023561,
0.027723508,
0.0031206983,
0.0037102716,
-0.0176102873,
0.0072732475,
0.0682244003,
0.0163582955,
0.025246866,
0.0004134373,
-0.0141207548,
-0.0480409563,
0.0045386492,
-0.0297553185,
-0.0307067707,
0.0055532781,
-0.0086681936,
0.0099538499,
-0.0223281942,
-0.0252702609,
0.0213228948,
0.0216912217,
0.0726671889,
-0.0204962995,
0.0176793467,
-0.0317127407,
0.001240003,
-0.0316003226,
0.0621999353,
-0.0100954715,
-0.0367161557,
-0.0289861243,
-0.0209208429,
-0.0159743689,
0.0390237346,
-0.0068284483,
-0.0049819849,
0.0122554153,
0.0185970571,
0.0413176157,
-0.0004448501,
-0.0024626036,
0.0088049583,
-0.023911627,
-0.0280744769,
0.0149728125,
0.0035010967,
0.0067627463,
-0.0190419275,
-0.0174983032,
-0.0543628894,
0.0086867688,
-0.0093025714,
-0.0309678726,
0.0245176479,
-0.0531461127,
-0.0395891964,
0.0231053699,
-0.0715352967,
0.0927519277,
0.0529704429,
-0.0168187898,
0.037522018,
-0.0470481738,
0.0210168008,
0.0014659382,
0.0888152421,
-0.0370843969,
-0.032677453,
-0.0494103692,
-0.0222990904,
-0.0732167736,
0.0291627366,
-0.0213433933,
-0.0391251855,
-0.0182207543,
-0.0157573614,
-0.0445136689,
-0.0244841408,
-0.0521820672,
0.0257489942,
-0.0369075276,
0.0086833863,
-0.0842180401,
-0.0203004833,
-0.0691354051,
0.0237828996,
-0.0354477353,
-0.0994481295,
0.0020984334,
-0.0174998511,
-0.0192408524,
-0.0449984409,
0.0329022445,
0.0860908702,
0.0106676323,
-0.032215625,
-0.0214754734,
-0.0285555478,
0.014425219,
-0.0186734423,
-0.0151459118,
-0.0224425606,
0.0308178961,
0.0100848936,
0.0571068041,
-0.0553144924,
0.0087271752,
0.0088705169,
-0.0566016696,
0.0964184403,
0.0184194725,
0.0033230088,
0.0126701063,
0.012575306,
-0.0721403658,
0.0472223051,
0.0072288471,
0.0045202421,
0.048751656,
-0.0108436132,
-0.0251157861,
-0.0133977067,
-0.0201677326,
-0.0223325379,
-0.0617615469,
0.0082059298,
-0.0012605406,
0.0568438657,
-0.0098081008,
0.0606054254,
-0.0070335069,
0.000546627,
0.0365477204,
0.0602254979,
0.0136393132,
0.0247224402,
0.0582194217,
-0.0080383224,
0.0214156434,
0.0179931801,
-0.0196868386,
-0.0192623362,
0.0251486972,
-0.0350058675,
-0.0075677168,
0.0612715706,
-0.0119617619,
0.003936409,
-0.0483278781,
0.0120607195,
-0.0158976894,
-0.0419287607,
0.0147502925,
0.039945282,
-0.0197113659,
0.0351019129,
0.0199786797,
0.0085450746,
-0.0106326621,
-0.0080411807,
-0.0471183471,
0.062816903,
0.0652676448,
-0.0322887711,
-0.0080511831,
-0.0164505541,
0.0081692683,
0.0207664967,
0.0334739126,
0.0300045125,
-0.0194842536,
-0.0078726299,
0.0360893048,
-0.0403865054,
0.0310670584,
0.0131242694,
0.0186394528,
0.040529605,
-0.0008165556,
-0.0173975956,
-0.0646417141,
0.0573906936,
-0.0252599828,
-0.0038008548,
0.0006897087,
0.0274916869,
0.0262271743,
-0.006596942,
-0.0095443176,
0.008429369,
0.0163574852,
-0.0660274625,
-0.0424415208,
0.0849436224,
-0.0381301753,
0.0091154957,
0.0371882804,
-0.0592418127,
-0.0107280053,
0.028411055,
0.0219798088,
-0.0009134993,
0.0625785664,
-0.0369495787,
0.0501759201,
-0.0186983757,
0.0257197823,
-0.0210090522,
0.0333043374,
0.0095136343,
0.0299017914,
0.0121400626,
0.0637178048,
0.0099060908,
-0.0156273767,
-0.0048903837,
0.0677266717,
-0.0229421891,
0.0019304529,
0.0111639602,
0.0020018003,
-0.0259751659,
-0.0523388349,
-0.0458262376,
-0.0187770836,
-0.0605521537,
0.0091753788,
-0.0090573905,
-0.0279073585,
0.0376366824,
-0.0333175361,
-0.048492454,
-0.0072960085,
0.0312797837,
-0.0061112526,
-0.0123755988,
0.0171108041,
0.0270370971,
-0.0478079766,
0.0065221223,
-0.0058358773,
0.03056412,
-0.0438110568,
0.0916293636,
-0.0300737526,
0.0340588465,
0.0339230113,
0.0095241135,
-0.0226943213,
-0.0207490381,
-0.0074062925,
-0.0362467468,
0.0209501348,
0.0636771098,
0.0857741088,
0.014811201,
0.0153910397,
0.0843663439,
0.0489943363,
-0.0413087979,
-0.0213648304,
0.0026063803,
0.0118155302,
-0.0586133674,
0.0115230205,
-0.0555190928,
0.0102989245,
0.0544225462,
0.0370193683,
-0.0006243488,
-0.0284773689,
0.0720022768,
-0.0382948816,
-0.0137353549,
0.0118960617,
0.0461027995,
-0.0197615307,
0.0153350299,
-0.0148413936,
0.0312901065,
0.0572251529,
-0.0165672451,
0.0189571306,
-0.025536865,
-0.0108790966,
-0.0396810435,
0.071391888,
-0.0236395355,
0.0389516801,
0.0373982005,
-0.0022867629,
-0.0098678833,
-0.0356374793,
-0.0139849,
-0.0313471183,
0.045080591,
-0.0377005488,
-0.0234904289,
0.0163285322,
-0.0040959157,
-0.0266103148,
-0.0260875132,
-0.0127642155,
0.0185241029,
0.0146514596,
0.0141878724,
-0.0414608195,
-0.0208473038,
-0.0169401262,
0.0066662282,
-0.0064557372,
-0.0233261399,
-0.0463294685,
0.0263676737,
0.0364159569,
0.0111339865,
0.0267824177,
0.0115938187,
0.0402596258,
0.0120454645,
-0.0269422382,
-0.0284338295,
-0.04041088,
0.0208230969,
0.0205643065,
-0.0599348471,
0.011149738,
-0.0251339898,
0.0586014874,
-0.0342556089,
-0.0410582386,
0.0608797334,
0.0703034177,
0.0266281739,
0.0479465462,
-0.0010143726,
0.0150991306,
0.0501147062,
-0.009622301,
-0.0933408216,
-0.0601020455,
0.025193058,
-0.0128954565,
-0.0372004844,
-0.0315571241,
-0.0231998079,
-0.0881194919,
0.0259468909,
-0.0071660536,
-0.0124889016,
-0.0090401089,
-0.0053083501,
-0.0144087523,
0.0041949181,
0.0518127456,
-0.0443771519,
-0.0179962069,
0.0203624982,
0.0062627327,
0.050966043,
-0.049814187,
0.0081392592,
-0.0308845192,
0.053176526,
-0.0651851147,
0.0434858054,
-0.0219082758,
-0.0017813731,
-0.0130799487,
-0.0546010248,
0.0155774876,
0.0006173532,
-0.0109786065,
-0.011765861,
-0.012824513,
-0.0029490092,
-0.0027757748,
-0.0345730521,
-0.0352839231,
0.0127737392,
0.0496760979,
-0.0207723398,
-0.0104382858,
-0.0274103545,
-0.0037215981,
-0.0152216628,
0.1041708514,
0.0564955659,
-0.0476650521,
-0.0019021149,
-0.0360612869,
0.0134978322,
-0.0064544156,
-0.0122032231,
-0.0263437461,
-0.047490418,
0.0285915397,
-0.0510836244,
-0.0345235728,
0.0404585339,
0.0299229287,
0.0434748642,
0.0024984146,
-0.0750666484,
-0.0008087297,
-0.0883892253,
0.0649218187,
-0.014151643,
0.0319670588,
-0.0150349857,
-0.0052402588,
0.0430323333,
-0.0174477678,
-0.023670882,
0.0834628716,
-0.0545289554,
-0.0663041845,
-0.009076206,
-0.0012128322,
0.0168557186,
0.0058741318,
-0.0042550578,
-0.0218484029,
0.039578326,
0.0537636206,
-0.0425723046,
0.0282547642,
-0.0111581562,
-0.0053058672,
0.0071241814,
0.0455608144,
0.0152644198,
0.0026009919,
0.0161565859,
0.0436762311,
0.009962217,
0.0507422574,
0.0323253274,
0.0002899839,
0.0138818566,
-0.0005076869,
-0.0976421386,
-0.0173264928,
-0.0260794219,
-0.0296689905,
-0.0448584482,
-0.0033739326,
-0.0082145203,
0.0494057387,
0.052130267,
-0.0575725436,
-0.0525215268,
-0.0079265218,
0.050871145,
0.0529461242,
-0.0157334693,
0.0386797152,
0.045291435,
-0.0151691092,
-0.0005517153,
-0.0277485475,
-0.0198390298,
-0.045086883,
0.0269268453,
-0.0190287959,
0.0115968548,
0.0225055162,
-0.0015807936,
-0.0376651064,
0.0348737612,
0.0155287068,
-0.0851995796,
0.0301852357,
0.0540048666,
0.035576947,
-0.0571549721,
-0.0020779551,
-0.0360045955,
-0.0017158061,
0.020038275,
-0.0042296792,
0.0129600447,
0.0010591001,
0.0064926432,
-0.0007896872,
-0.0217078663,
-0.0305313915,
0.0088329837,
-0.0457794704,
0.0517562144,
0.0400879197,
-0.0297607835,
-0.0034470451,
0.0153607009,
0.0202234369,
0.0326084159,
0.0416869372,
0.0378929563,
-0.097811766,
-0.0540003143,
0.0585880466,
-0.0114309397,
-0.0069601135,
0.015778102,
0.0000161309,
0.0605631806,
0.0398707166,
-0.0032628016,
-0.0536340736,
-0.005767799,
-0.043706134,
0.0114126308,
-0.0194548108,
-6.267246579e-33,
-0.0340658352,
-0.0332047418,
-0.0313053615,
-0.0723902062,
-0.0150439302,
0.0370644182,
-0.0474056192,
-0.0192014202,
0.0082209026,
0.0053367079,
-0.0141176116,
-0.0152691593,
-0.006310611,
-0.0244674049,
0.0162774399,
0.0085050231,
0.0335997641,
0.0086296033,
-0.0494146645,
-0.0483559631,
0.0842826813,
0.0153531423,
0.0410367139,
0.0053679063,
-0.0088395486,
0.0478339493,
0.0056567849,
-0.0268222187,
0.0386634059,
0.034789037,
-0.0236827843,
0.0600521378,
0.022384027,
-0.005841142,
0.0134715978,
-0.0192802716,
-0.0670631155,
-0.0428390279,
0.0590557978,
0.0300860461,
0.0057037869,
-0.0351906158,
0.0792188272,
-0.0042397692,
-0.015596454,
0.0022745633,
0.0318429545,
-0.0127021065,
0.0275601093,
0.0083434684,
-0.063026838,
-0.020524377,
-0.0210923217,
-0.0146923661,
-0.0080988491,
0.0126970327,
0.0076150252,
0.0758201703,
-0.0897540599,
0.0089951102,
-0.0185002424,
-0.0108033,
0.0232867952,
-0.0054434678,
-0.001618184,
0.0844222233,
-0.0015297333,
-0.0018315542,
-0.0305687711,
0.0330914743,
0.0390554145,
0.0216974672,
0.0496358089,
0.0191393215,
0.0857787356,
-0.0508417375,
0.009482448,
-0.0031996865,
0.0088450219,
-0.0010385667,
0.0609202906,
-0.014812286,
-0.0503957197,
0.0133900307,
-0.0610409938,
-0.0332693383,
-0.0241739731,
-0.033496853,
0.0339611992,
-0.023685677,
-0.0189633854,
0.0341306143,
-0.008931404,
-0.0080490401,
-0.0297011435,
-0.0312947854,
-0.0053734891,
0.0109589156,
-0.0150361387,
0.0264937598,
-0.021630358,
-0.0381776616,
0.0363650732,
0.0012109157,
0.0437440984,
-0.007556702,
0.0080041969,
-0.0057646423,
-0.0875512287,
-0.0110023189,
-0.0291663185,
0.0081980005,
-0.0394396,
-0.0336300805,
0.0241016839,
0.0288633592,
-0.0134868212,
-0.0025682929,
0.0296936464,
-0.0266088471,
-0.0485744439,
0.0718036443,
-0.0337296873,
0.057641536,
-0.0247441586,
-0.0181158632,
-0.0375307165,
-0.0213767104,
0.0326296277,
-0.0773316249,
-0.0140044028,
-0.0348433219,
2.711e-7,
0.0153688444,
0.0539737642,
0.0078046592,
0.0303273536,
-0.0040102494,
0.0404067151,
0.0070061707,
0.022516232,
-0.024558045,
0.0026664329,
-0.0121976323,
0.0568252653,
0.0169996023,
0.0341496468,
-0.0867646039,
0.0521115065,
-0.0288641956,
0.0260494146,
-0.0632279664,
-0.0074791727,
0.0707708374,
0.0545017011,
0.0449430011,
-0.0061978595,
-0.0123016909,
-0.0668038353,
0.0492909364,
-0.0725460723,
0.0229044408,
0.0264400132,
0.047227148,
0.0388472006,
0.0112901544,
0.0327328555,
0.0042534578,
0.0164670423,
0.040383894,
-0.0203003082,
-0.0140138203,
-0.037747778,
-0.0403961577,
-0.0308262762,
-0.0269159805,
-0.0882974491,
0.0448066108,
-0.0745923147,
-0.032928627,
-0.0085843848,
-0.0058561116,
0.0150901312,
0.0766574368,
-0.0006538626,
-0.0433403254,
0.0349240787,
0.0240614545,
0.0193580035,
0.0106022023,
-0.0830623582,
0.026270099,
0.0225202683,
-0.0050205206,
0.0172379781,
-0.0026475112,
-0.0089875888,
0.008720953,
-0.1026800498,
-0.0071720197,
2.323007344e-34,
0.023208335,
0.0050471094,
0.0190977771,
0.0379060209,
0.0725143179,
-0.0202329699,
0.0377783924,
-0.0043372074,
-0.0255298223,
-0.0193031114,
0.004355832
] |
http://arxiv.org/abs/2402.01738v1 | http://arxiv.org/pdf/2402.01738v1 | C4Q: A Chatbot for Quantum | 2024-01-29T09:44:45 | 2024-01-29T09:44:45 | Yaiza Aragonés-Soria, Manuel Oriol | cs.CL | cs.CL, quant-ph | Quantum computing is a growing field that promises many real-world
applications such as quantum cryptography or quantum finance. The number of
people able to use quantum computing is however still very small. This
limitation comes from the difficulty to understand the concepts and to know how
to start coding. Therefore, there is a need for tools that can assist
non-expert in overcoming this complexity. One possibility would be to use
existing conversational agents. Unfortunately ChatGPT and other Large-Language
Models produce inaccurate results. This article presents C4Q, a chatbot that
answers accurately basic questions and guides users when trying to code quantum
programs. Contrary to other approaches C4Q uses a pre-trained large language
model only to discover and classify user requests. It then generates an
accurate answer using an own engine. Thanks to this architectural design, C4Q's
answers are always correct, and thus C4Q can become a support tool that makes
quantum computing more available to non-experts. | C4Q: A Chatbot for Quantum: Quantum computing is a growing field that promises many real-world
applications such as quantum cryptography or quantum finance. The number of
people able to use quantum computing is however still very small. This
limitation comes from the difficulty to understand the concepts and to know how
to start coding. Therefore, there is a need for tools that can assist
non-expert in overcoming this complexity. One possibility would be to use
existing conversational agents. Unfortunately ChatGPT and other Large-Language
Models produce inaccurate results. This article presents C4Q, a chatbot that
answers accurately basic questions and guides users when trying to code quantum
programs. Contrary to other approaches C4Q uses a pre-trained large language
model only to discover and classify user requests. It then generates an
accurate answer using an own engine. Thanks to this architectural design, C4Q's
answers are always correct, and thus C4Q can become a support tool that makes
quantum computing more available to non-experts. | [
0.0437225066,
0.0600708574,
-0.02731096,
0.0410160199,
-0.0221691206,
-0.004250478,
-0.0142981727,
0.0281712152,
0.0514595397,
0.0038974844,
0.0099959811,
-0.0254276879,
-0.0307023358,
0.0421707518,
-0.0109297009,
-0.0275009666,
0.0223888122,
-0.0005341901,
-0.0397956073,
-0.0468045734,
-0.0046943496,
-0.0186978225,
-0.0118169403,
0.0311884657,
0.0369440094,
-0.0462799706,
-0.0321493857,
-0.0006358449,
-0.0003515965,
0.0088870022,
0.0023584452,
0.0662122145,
-0.0131984139,
0.0555638596,
0.0000018659,
-0.0301010348,
0.0017607141,
0.0348733924,
-0.056471169,
-0.0060736118,
-0.0386827812,
-0.0061001224,
-0.0084108431,
0.0395800881,
-0.0447430834,
0.038664151,
0.0887026638,
-0.00026699,
0.081684567,
0.0747455209,
0.0197899733,
-0.0115555795,
0.0154902069,
0.0334438682,
0.0319116302,
-0.0855373144,
0.0167222712,
0.0478513502,
0.0340714231,
0.014374421,
-0.0268217027,
-0.0008277981,
0.0639791116,
-0.0290711187,
0.0145505974,
-0.0080501102,
-0.0265205465,
0.0089081619,
-0.0452028476,
0.0712537691,
-0.0020256659,
-0.0309567619,
-0.0014238204,
0.0329042859,
-0.0253176205,
-0.0182253122,
-0.0317757428,
0.0072228983,
-0.0092502646,
-0.0312625058,
-0.0017943305,
-0.0208871011,
-0.0266680717,
0.0008760372,
0.0181865636,
0.0300006643,
0.0431024097,
0.0051520187,
-0.0439999439,
-0.0001944882,
0.0101237819,
-0.0471682698,
0.0137869557,
0.0150608551,
-0.032028079,
0.0063707167,
-0.0034089349,
0.0528155491,
-0.015531485,
-0.0570117608,
-0.0495568402,
0.0871863738,
0.0404732339,
0.0327891521,
0.0158905908,
0.0389773436,
0.0117835412,
-0.0052757002,
-0.0120795202,
-0.0042125834,
-0.0867914036,
-0.0022740152,
-0.0213168152,
0.0407163948,
0.0033045057,
-0.0603763312,
-0.0961371213,
0.0509553775,
0.039613232,
0.0160165317,
-0.0163119752,
0.0216840841,
-0.0096415104,
0.0097801043,
-0.0323801637,
0.0341673605,
-0.0514837429,
-0.032759048,
-0.0411785357,
-0.0921053216,
-0.0199491996,
0.0258583091,
-0.0269197077,
-0.0158191044,
-0.0056233159,
0.077694349,
0.0050841388,
-0.0258921646,
0.0811377689,
0.0092896139,
0.0131228548,
-0.0451584086,
0.0076317727,
-0.0274345074,
0.0128321918,
0.0294224937,
0.0037266582,
-0.0130144563,
0.0211169701,
-0.0158003345,
0.0016290372,
0.056458097,
-0.0302843135,
-0.0448290296,
0.0332206562,
0.0299497284,
0.0215055626,
0.0131636355,
-0.0219003707,
0.024571484,
0.0136718443,
-0.0421814807,
0.0865590051,
-0.0287822913,
0.000361143,
-0.0426124558,
-0.0321362168,
0.0245244056,
-0.0243363082,
0.0355352238,
-0.038586542,
0.0612269081,
-0.0378240943,
0.0007003914,
0.0069340584,
0.0510342047,
0.0218505617,
0.0647492781,
-0.01584718,
0.0291976947,
0.02172767,
-0.0297449287,
-0.0365316309,
0.1034518778,
-0.0542947687,
-0.0244262274,
0.0196182784,
-0.0292013418,
-0.0212632529,
-0.0401844457,
0.0126013402,
0.0053686653,
0.0488373488,
-0.0126574319,
0.0526206419,
0.0167179983,
-0.0317249931,
-0.0341595076,
0.0285583846,
-0.0330064595,
-0.03768038,
-0.0858574584,
0.0457834713,
0.0726570487,
0.0373208933,
-0.0151423709,
0.0106935501,
0.0370616168,
-0.0489204489,
0.0570718497,
0.0777937397,
0.0147853056,
-0.0114547908,
-0.0214350708,
-0.0158248059,
-0.0343776643,
-0.0251785088,
0.0084524537,
0.0404316522,
0.0038250545,
-0.029766025,
0.0412487201,
0.0155965779,
0.0090922285,
0.0236050133,
-0.0658276752,
-0.0179962739,
-0.0007283645,
0.0261052381,
0.0196488332,
0.0028016258,
0.0485215411,
0.0133869499,
0.0093671074,
-0.0573841594,
-0.0459237322,
0.0154933389,
-0.0244758986,
0.0444504209,
-0.087750189,
-0.0012994922,
0.0198466554,
-0.0078383926,
-0.0181373749,
0.1046152413,
-0.0232883003,
0.0601742044,
-0.0148965875,
-0.0390440747,
-0.0128595671,
0.0100633129,
-0.0096375057,
0.0601782277,
0.0617239065,
0.0172103811,
-0.0012503755,
-0.005125043,
-0.0218152665,
-0.0651587918,
-0.0126862051,
-0.0126447342,
0.0022046587,
0.0463585109,
0.0363975726,
0.0031249435,
-0.0194231607,
0.0076383725,
-0.1009176672,
-0.0111630019,
-0.0133114578,
-0.0108117042,
0.0068218224,
0.0161407813,
0.0104644662,
-0.0648481995,
-0.0344711728,
-0.0007905561,
-0.0081763938,
-0.0193312149,
0.0172772836,
0.0267309714,
-0.0429134332,
-0.0475002192,
0.00212355,
-0.0297682788,
0.0030983158,
0.0320469737,
0.0084426934,
-0.0410768576,
-0.0037194232,
-0.0224873275,
-0.0081685819,
-0.0016716958,
-0.0507936962,
0.0351243019,
-0.0271826629,
-0.0212159883,
0.0445284583,
0.0369464532,
0.0550441295,
0.037624117,
0.0305613782,
-0.0430478044,
0.0350812413,
-0.036081519,
-0.0132285478,
0.0004926104,
-0.0024733124,
0.0100027239,
0.0304262992,
0.0261075962,
0.0269922409,
-0.0568856336,
0.072923243,
0.0221150871,
0.0037409076,
-0.0332599282,
0.0119520267,
0.0008556009,
0.0304466337,
0.0205545779,
-0.0098535931,
0.0970458165,
-0.0060029868,
-0.0193018522,
-0.0344903916,
-0.0628985614,
-0.002264905,
0.0691551566,
-0.0334503315,
0.0502822287,
-0.0629333481,
-0.0056646653,
-0.0087351101,
-0.0323836505,
0.0203634296,
-0.0282070842,
-0.049882587,
-0.0146629969,
0.0152340429,
0.0235393252,
-0.0052430481,
-0.0345143341,
-0.0032863086,
-0.0766475126,
-0.0108001139,
0.0037493221,
-0.0358499996,
0.0061803204,
0.0222519673,
0.0313921943,
0.0002926954,
-0.0154699683,
-0.0132167134,
-0.0121192839,
-0.0493646711,
0.0141518917,
0.0012590908,
0.0444154926,
-0.0273316428,
0.058469262,
0.0534472391,
0.0066613196,
0.0390728563,
-0.0238417182,
-0.0478288792,
0.0292696375,
-0.0052919588,
0.0117558325,
0.0231033005,
0.0269617587,
-0.0235476922,
-0.0228827596,
0.0532498807,
0.0136908339,
0.0011716062,
0.00803997,
-0.0086057447,
-0.0725098029,
-0.0043705408,
0.0723253787,
-0.1176054925,
-0.0354486406,
-0.0372047536,
0.0243006386,
0.0111770425,
0.0000812828,
-0.0229700785,
-0.0223629139,
0.0038799273,
0.0331218876,
-0.0000914318,
-0.0698898137,
-0.0404849611,
-0.000364796,
-0.0143524492,
0.0349298753,
0.0398855247,
0.0133779151,
-0.0053519919,
-0.0244558658,
-0.0025429584,
0.017985737,
0.046465043,
0.0120945061,
0.0559475534,
0.0758353323,
0.0242105909,
-0.0233225059,
-0.0203502942,
-0.0283612851,
-0.0131349713,
-0.009141529,
-0.0005893299,
-0.0311450157,
0.0134692965,
-0.0423549525,
-0.0590676665,
0.0270592961,
-0.0090446742,
0.0027567786,
0.0327915959,
-0.0239089578,
-0.0532934256,
-0.0521273874,
0.0227601174,
0.0394518413,
-0.0031621051,
0.0617222376,
-0.0181153435,
-0.0309845768,
-0.0891534537,
-0.0642952472,
-0.0065660896,
-0.0057065892,
-0.0462830402,
-0.060063947,
-0.0033452541,
-0.0065056044,
-0.0776688159,
0.0070413188,
0.0235460158,
0.0099585364,
-0.0065522371,
-0.018898448,
0.0593686514,
0.0293195751,
0.0178103503,
0.0029436254,
-0.0423638746,
0.0062452927,
-0.0807988867,
-0.045103319,
-0.0023918666,
-0.0916458219,
-0.0177952554,
-0.0315257311,
0.0259855725,
-0.0476457179,
-0.0103869746,
0.0299735889,
0.0265482068,
-0.0124734845,
-0.0113307806,
-0.0285660122,
-0.0269710701,
0.0168586466,
-0.0226450618,
-0.0126082739,
0.0070384108,
0.0554081649,
-0.0227132346,
-0.0037350173,
-0.0238740165,
-0.0099072251,
-0.0176060703,
0.03905534,
-0.1152030528,
-0.0868600234,
0.0228862483,
0.0015200527,
0.0126357162,
0.032619223,
0.0233994145,
0.031746801,
0.014897814,
-0.0304915458,
0.0255196709,
-0.0570315123,
-0.0228955336,
0.000407433,
-0.0007008836,
0.0766839683,
-0.0243923552,
-0.0219842102,
-0.0091327084,
-0.0122771803,
0.0357448198,
-0.020563649,
0.0367811508,
0.0423591696,
-0.0082736611,
-0.0203268286,
0.0455649868,
0.027560696,
0.0826815665,
-0.0053835623,
0.0359381102,
0.0286637321,
-0.0045682914,
0.0009145607,
0.0051034223,
-0.1009478122,
-0.118395552,
0.0296497475,
0.0857247859,
0.0311042797,
0.1003413498,
-0.0192910675,
-0.0501879789,
-0.0065487828,
-0.0065431767,
0.0691177472,
0.0391969457,
-0.0167916268,
0.0318219885,
0.0300686359,
-0.023916468,
-0.059722472,
0.0141007109,
0.0010346539,
0.0146102197,
-0.0137676243,
-0.0190653652,
-0.0490053445,
-0.0141591467,
0.016814746,
0.0381555371,
-0.0012209419,
0.0160611439,
-0.0126851676,
-0.0038410865,
-0.0453329049,
0.0307826344,
-0.0497496314,
-0.0400314257,
0.0881692618,
0.0333262049,
0.0025169309,
-0.0651620105,
-0.0241041332,
0.000704695,
0.0068411967,
-0.0014345045,
0.0235476512,
-5.887841451e-33,
-0.0403516218,
-0.0352051929,
0.0228078999,
0.0136703635,
0.0397538878,
-0.0235639997,
-0.0392181166,
-0.0160037633,
-0.0286001172,
-0.0415736809,
0.0186617766,
0.0354509279,
0.0247150902,
-0.025247639,
-0.0026869811,
0.008234866,
0.0849851444,
0.0580574647,
-0.0101525364,
0.0500443727,
-0.0022172895,
-0.0134357642,
0.0232879017,
-0.0095095206,
0.007001522,
-0.0140449004,
0.006151272,
0.0328846052,
-0.0260803252,
0.0119283283,
-0.0347335115,
-0.0012224265,
-0.0525285676,
0.021830203,
0.0372569337,
0.0122343171,
-0.0088983122,
-0.0179708041,
0.0133033805,
0.0047467821,
-0.010526997,
-0.0141130928,
-0.0110140294,
-0.017583413,
-0.0107570346,
0.0076694107,
0.0120522277,
0.0286426097,
0.0246140286,
-0.0056581507,
-0.0365073718,
0.0076551079,
-0.0236846246,
-0.025076434,
-0.0464885533,
-0.005776573,
-0.0287538096,
-0.0519652404,
-0.0250961687,
0.0380398557,
0.0251425765,
-0.0154029699,
0.0072901538,
0.0180095956,
0.0207844209,
0.0117058372,
0.0777780339,
-0.0206472483,
-0.0002626934,
0.0091021759,
-0.0167643856,
0.0034078695,
0.0400675684,
0.0449261889,
0.0514933467,
-0.0163766798,
-0.0449309945,
-0.0191629361,
0.1376378238,
0.0214558858,
0.0249451865,
-0.0053016045,
-0.0605973229,
-0.0387867689,
0.0394533165,
0.0625387058,
-0.0130297318,
-0.0436239094,
-0.0172001012,
0.014996781,
-0.0621307716,
-0.0433341451,
0.0384188481,
-0.058913596,
0.0424176604,
0.0029747502,
0.0196305178,
-0.0181904342,
-0.0070880642,
0.0140018743,
0.0012261471,
-0.0153791904,
0.0177043546,
-0.0652694553,
0.0079229306,
0.031169815,
-0.0504866578,
0.0257650129,
-0.0212378651,
0.0118088955,
-0.0265239608,
-0.0044404306,
0.0062552681,
0.0148542458,
-0.0281293765,
0.036799565,
-0.0087108715,
0.0405167751,
-0.027755212,
0.0331109986,
-0.0371603034,
-0.0285678413,
0.0086108269,
0.0335093178,
-0.042735558,
-0.0269269291,
-0.068995215,
0.0596020296,
0.0053427201,
-0.0817734301,
-0.0268198345,
0.0060483799,
2.732e-7,
0.0098496545,
0.0444857329,
-0.0013309696,
0.0176756587,
0.0039966423,
0.0149822654,
0.004509646,
0.0229943544,
0.0042379163,
0.0245106705,
-0.0055457135,
0.0483581014,
0.0057154517,
0.0114810951,
-0.0678331032,
-0.0537125953,
-0.0497339144,
-0.0476535968,
-0.047400225,
0.0343046077,
0.0675350875,
0.0599356294,
0.0145676723,
0.0330075957,
0.0267417934,
-0.0165346023,
0.033517357,
0.0441469923,
-0.0042298562,
0.0349207446,
-0.0517926477,
0.0348531976,
0.0314383134,
0.0126396092,
0.0444832258,
0.0018747535,
-0.0128483949,
0.0420595668,
0.0060137953,
-0.0502884276,
-0.0274389163,
0.0024209002,
0.0084743099,
-0.0059366454,
0.0285146944,
-0.0211819112,
-0.0169565771,
-0.0180413853,
-0.0153241968,
-0.0352066159,
0.0174173675,
-0.015585701,
-0.0095315073,
-0.019877227,
-0.0215067118,
-0.0143192923,
0.0323748961,
-0.0113894641,
0.0060919328,
-0.0170601662,
-0.0299955625,
-0.0722977147,
-0.013946739,
-0.0600717217,
0.0053010923,
-0.0069264509,
0.0217801481,
2.113943608e-34,
0.0225436166,
0.0401720144,
0.0112789404,
0.0164253842,
0.0449964926,
0.0483628176,
0.0235401597,
-0.0108475788,
0.0690488964,
-0.018481249,
-0.03314697
] |
http://arxiv.org/abs/2402.01737v2 | http://arxiv.org/pdf/2402.01737v2 | Assistive Large Language Model Agents for Socially-Aware Negotiation Dialogues | 2024-01-29T09:07:40 | 2024-06-18T13:10:16 | Yuncheng Hua, Lizhen Qu, Gholamreza Haffari | cs.CL | cs.CL, cs.AI, I.2.7 | We develop assistive agents based on Large Language Models (LLMs) that aid
interlocutors in business negotiations. Specifically, we simulate business
negotiations by letting two LLM-based agents engage in role play. A third LLM
acts as a remediator agent to rewrite utterances violating norms for improving
negotiation outcomes. We introduce a simple tuning-free and label-free
In-Context Learning (ICL) method to identify high-quality ICL exemplars for the
remediator, where we propose a novel select criteria, called value impact, to
measure the quality of the negotiation outcomes. We provide rich empirical
evidence to demonstrate its effectiveness in negotiations across three
different negotiation topics. The source code and the generated dataset will be
publicly available upon acceptance. | Assistive Large Language Model Agents for Socially-Aware Negotiation Dialogues: We develop assistive agents based on Large Language Models (LLMs) that aid
interlocutors in business negotiations. Specifically, we simulate business
negotiations by letting two LLM-based agents engage in role play. A third LLM
acts as a remediator agent to rewrite utterances violating norms for improving
negotiation outcomes. We introduce a simple tuning-free and label-free
In-Context Learning (ICL) method to identify high-quality ICL exemplars for the
remediator, where we propose a novel select criteria, called value impact, to
measure the quality of the negotiation outcomes. We provide rich empirical
evidence to demonstrate its effectiveness in negotiations across three
different negotiation topics. The source code and the generated dataset will be
publicly available upon acceptance. | [
0.0576420352,
0.0540252253,
-0.0227536298,
0.0455494858,
-0.0385568179,
-0.0186062604,
0.0362957679,
0.0071332534,
0.0090305051,
-0.0602539964,
-0.0420585871,
-0.0249450319,
-0.0226336271,
0.0425389633,
0.0080228802,
-0.0578405745,
0.0253191981,
-0.0022816111,
0.0763843283,
-0.0240525715,
0.0023441263,
0.0005141573,
-0.0074655893,
0.0586804263,
-0.0501850024,
0.0037374529,
-0.0076299859,
0.010413032,
0.0264177788,
0.0011596791,
0.0257119723,
0.0617849827,
0.0281639509,
0.0267926864,
0.000002024,
-0.082075119,
-0.0032214061,
-0.0256554876,
-0.0136359725,
0.0024056928,
0.0708879605,
-0.0254225843,
-0.0297441464,
0.0482291915,
-0.0768374279,
0.005511397,
0.0201838873,
0.0714726448,
0.0692398325,
0.083049275,
-0.0084383124,
0.0063927416,
-0.035019692,
-0.0362334885,
-0.015507333,
-0.0920779407,
0.0206052754,
0.0028915789,
0.0034376641,
0.0176668204,
-0.0029254404,
0.0317449532,
0.0079337405,
-0.0776833221,
-0.0421849154,
0.026381636,
-0.0725217685,
0.0011634266,
-0.004351147,
0.0821248591,
0.0315545648,
0.010451966,
0.0039943638,
-0.0270697288,
0.0365916267,
0.0263518039,
-0.0254398528,
0.0135047566,
0.0333001763,
0.0072283531,
-0.0198292546,
0.0206260141,
0.0263035521,
0.0174894594,
0.0112727117,
0.0508144051,
-0.0254106428,
-0.0157254506,
-0.0049519027,
-0.0125250109,
-0.024015747,
-0.0686114207,
-0.0060632951,
0.0238597654,
-0.0928486213,
0.0323941261,
-0.0266013965,
0.0409597792,
-0.0082587954,
-0.046869617,
0.0146279214,
-0.0007375144,
0.0169781744,
0.055925969,
0.0401095971,
-0.0013831438,
0.0021239454,
0.0047390345,
0.0120086242,
0.0051344899,
-0.0225176346,
-0.0088038025,
-0.0704128742,
0.0150979972,
-0.0653338358,
-0.0348888934,
-0.018408142,
0.0269991457,
-0.0707428232,
0.0506592095,
-0.0167770144,
0.0218751114,
-0.0048254495,
-0.0090171415,
-0.0529044904,
0.0078900959,
-0.0182077158,
0.010086515,
0.0020716041,
-0.0483760163,
0.01786156,
0.0073934589,
0.0079315938,
-0.0126785012,
-0.0050542713,
0.0456368402,
-0.0088696778,
-0.0109717809,
0.0061716069,
-0.0145188812,
0.0066017793,
-0.0625643283,
-0.0433923192,
-0.046212066,
0.0198641289,
-0.0378423706,
-0.0308163632,
-0.092556566,
0.0239853263,
0.0002410657,
-0.0130320955,
0.0646277145,
0.0316071697,
0.0087875705,
0.0519952588,
0.0338184275,
-0.0233286358,
0.0670947284,
0.0550443456,
-0.0262239166,
0.0648228377,
-0.0256449152,
0.038490694,
0.0554349422,
-0.0000870395,
0.0196886323,
0.0180949699,
0.0118118897,
0.0393194519,
0.0329371542,
-0.0282818731,
0.0416117273,
-0.048624523,
-0.0487203188,
0.097153306,
0.0427895673,
0.0083079301,
0.0423347354,
-0.0125614544,
0.0814828649,
-0.0156340376,
-0.0403236374,
-0.0455480255,
0.020705739,
-0.0115602612,
-0.0290641952,
-0.0012713715,
0.0057218722,
-0.0361263454,
-0.0535703115,
0.0495055728,
-0.0097507015,
0.0660909116,
0.0469031855,
0.0075394064,
0.0019663605,
-0.0807886496,
0.0095988698,
0.0092673963,
-0.0347566009,
-0.0200729202,
-0.0212492608,
-0.0083382474,
0.0370848738,
0.0645080879,
-0.0684342682,
-0.0141115412,
-0.0563492849,
-0.0153787704,
0.0678994358,
0.0423083827,
-0.0243326072,
-0.0213815495,
-0.0527577624,
-0.0172269382,
0.004564533,
-0.0044914265,
0.0004501661,
0.0185618252,
-0.0054029427,
0.0019845937,
0.0204550847,
0.0227623396,
-0.001476194,
0.01512207,
-0.0292779151,
0.013125821,
0.0111683086,
0.0403215177,
0.012560633,
-0.001918765,
-0.0255906899,
0.0113007277,
-0.0010917125,
-0.0052357377,
0.0519184507,
-0.0565358475,
-0.0303493459,
0.0380997248,
0.0062200166,
-0.0256931074,
0.0078440672,
-0.0154845463,
-0.028114479,
0.04211393,
-0.0084495805,
0.0420483612,
-0.0273647066,
-0.0058571943,
-0.0240851007,
0.0116789462,
-0.0287754647,
0.0154369073,
0.030122485,
-0.0327348374,
0.0153440209,
0.0132370209,
-0.0235562827,
-0.0408180319,
-0.0033590277,
0.0367677771,
0.0091237128,
0.0230023097,
0.03058777,
-0.0016567216,
-0.0189142078,
-0.0537989698,
-0.0730872825,
-0.0331506506,
-0.0194203313,
-0.0294424649,
0.0066977418,
-0.0248994119,
0.0046089152,
0.0098076547,
-0.041993238,
-0.0006806413,
0.0368626304,
-0.0080637503,
0.0255491436,
-0.0102014029,
-0.0567689277,
-0.0246280096,
0.0135017699,
0.0027198114,
0.0385912433,
0.0030542442,
0.0151024424,
-0.0557503887,
0.014858813,
-0.0319619887,
0.0288821124,
-0.0236953106,
-0.0492104888,
0.0198367182,
0.0341322683,
-0.0052801403,
0.0703295767,
0.0285720676,
0.0859950408,
0.0100476025,
-0.0366383977,
-0.0850864053,
0.0291425269,
-0.0158286449,
0.0300785881,
-0.1148293391,
-0.0188240763,
0.0027779771,
0.0364024118,
-0.0152800586,
0.000740157,
-0.0434187725,
0.049436681,
0.0385247022,
-0.0587856211,
0.0101155108,
-0.0346335955,
-0.0357008278,
0.027884962,
0.0010610365,
-0.0034729741,
0.0497827977,
0.0073829787,
-0.0284017343,
-0.0134225711,
-0.0040082685,
-0.0447676182,
0.0602804013,
-0.0537860394,
-0.0104216505,
0.0075970218,
-0.0239961389,
-0.038110245,
-0.0427392237,
0.0674850121,
-0.0419410504,
0.0355898961,
-0.0821066797,
0.0052240533,
0.025519602,
-0.0189249758,
-0.0440971889,
-0.034807723,
-0.0120420745,
0.0548353121,
0.0325160399,
-0.010544572,
-0.0645476282,
0.0003211774,
0.0085582463,
0.0257754978,
0.051143866,
-0.0018008547,
-0.0153173972,
-0.0303247403,
0.0314373821,
0.052648589,
0.0077281035,
0.0010351765,
0.0717103183,
0.0279327519,
-0.0189260356,
-0.0499397554,
-0.0030718204,
-0.0046046963,
0.0162843075,
-0.0077116173,
0.0204592999,
0.014187539,
0.025028605,
-0.0480535291,
0.0038557528,
0.0826660916,
0.0250238311,
-0.0165086798,
0.0289215911,
-0.0113750016,
0.0277122147,
0.0179295279,
0.0144118248,
-0.0871858001,
0.0061785728,
-0.012281497,
-0.0086921528,
-0.0228834003,
-0.0640648827,
0.0052548531,
-0.118189171,
0.0775017515,
0.001430659,
-0.0166457221,
0.02284218,
0.0134853479,
0.0118995858,
-0.0218349937,
0.0377553403,
-0.0176281594,
0.0250691976,
-0.0507277325,
-0.0012840625,
0.0111167748,
-0.0256763231,
0.0269858036,
-0.0182095133,
0.0643034354,
0.0280007832,
0.0547993556,
0.0130562037,
0.0368011519,
0.0286926199,
0.0547949187,
-0.0377157182,
-0.003808989,
0.0213826485,
-0.0020747979,
-0.0162288714,
0.0080623692,
0.0478746891,
-0.014166683,
-0.0331948437,
0.0150534846,
0.0406750962,
-0.0273875687,
-0.0002295041,
0.0431837924,
0.0308513101,
0.0565487854,
0.0548372306,
0.0141577562,
-0.0589521378,
-0.0194499008,
-0.0343446881,
-0.0336574614,
0.0123542957,
-0.0632273182,
-0.0102443108,
0.0178425126,
-0.006122937,
-0.0372632481,
-0.0750543326,
0.0573064163,
0.0431784019,
-0.0036543624,
-0.0031185579,
0.0517451614,
0.0054810229,
0.0167065952,
-0.0018444898,
-0.0481580496,
0.0270331688,
-0.0248945113,
-0.0228479914,
0.0022074082,
-0.0380929708,
-0.0201275814,
-0.0277559627,
-0.0076141027,
-0.0706555769,
-0.0124834813,
0.0453435034,
0.026886303,
0.0432098284,
0.0061332374,
-0.0535242669,
-0.0095865848,
0.0282035656,
-0.0091599086,
0.0166926775,
0.0614185743,
0.0351286121,
0.0030068266,
0.0370242633,
-0.0143512553,
0.033373259,
0.0167944413,
0.0433417968,
-0.0634314269,
-0.0685092211,
0.0087067271,
-0.0290690977,
-0.0408358313,
0.0308347568,
-0.0205800701,
-0.0386010185,
0.0301045477,
-0.0538771115,
-0.0198905747,
-0.0535404459,
-0.0245840773,
0.0024865202,
-0.0228435844,
0.0234494992,
0.0041205506,
-0.0545822755,
0.0127804046,
-0.0085294098,
0.0504452549,
-0.0232686419,
-0.0209989734,
0.0291974582,
0.0020739303,
0.0193041265,
0.0305805746,
0.0207052417,
0.0296453238,
0.0411096774,
0.0726176128,
0.0464205258,
-0.0320900157,
-0.0073723556,
0.0009032157,
-0.0185103957,
-0.021076519,
0.0281132944,
0.0616674423,
0.0170217287,
-0.0100090429,
0.0368098207,
-0.0532967858,
0.0174765624,
-0.017499363,
0.0312164295,
0.0109792138,
-0.0841768384,
0.060935095,
0.0204451419,
-0.0156062515,
0.0533496179,
0.0580805577,
-0.0255963095,
0.0179038178,
-0.0138548315,
0.0232529044,
-0.0470373072,
-0.0484889895,
0.0289559774,
0.0031761266,
-0.0315362625,
0.0158672873,
-0.065538466,
-0.0354044884,
-0.0219258144,
-0.0076398561,
-0.0273130238,
-0.0296003856,
-0.0009984334,
0.0221442152,
0.0013193395,
0.01123532,
-0.045499932,
-0.0008348167,
-0.0299882777,
-0.0294376817,
-0.0419191085,
-6.170316784e-33,
0.0075673,
-0.0678923428,
0.0317767151,
0.0144078014,
-0.0198984668,
-0.01583069,
-0.0278000366,
-0.0069800974,
-0.0043820278,
0.0380386189,
-0.0239206515,
0.0154930456,
0.0158707723,
0.0014995424,
-0.0268762186,
-0.0263773557,
0.0364786722,
0.0130482642,
-0.035447184,
0.0180062745,
-0.0165816918,
-0.0030833916,
0.0207604002,
-0.1227015853,
0.0309118796,
-0.0129365074,
-0.0323431864,
0.0335343108,
-0.0067702704,
0.0426549353,
-0.0014532824,
0.0292536821,
-0.0166668929,
-0.0386292376,
0.0104300715,
0.0364588052,
-0.0295257568,
0.0183608159,
0.0275095757,
0.0001578209,
-0.0063350522,
-0.0262525529,
0.0021427074,
-0.0039597624,
-0.0364993662,
0.0015546019,
0.0054632756,
-0.0227507651,
-0.0308162086,
0.0315440968,
-0.0787040293,
0.0180081055,
-0.0566164665,
-0.03854651,
0.0208227355,
0.0038184673,
0.0086166421,
-0.0224146508,
-0.0395244919,
0.0326918066,
-0.0141022075,
-0.0169500709,
0.0253103822,
0.0039959918,
0.0326022096,
0.0571684875,
-0.0075223721,
0.08820232,
0.0246082917,
-0.098692894,
0.0150643373,
0.0628933758,
-0.0048493845,
0.0827544406,
-0.0590228178,
-0.0238254499,
-0.0244715121,
-0.0090491725,
0.0567975231,
0.0099375807,
0.0132328542,
0.0240048114,
0.0077055269,
-0.0424915999,
-0.046029374,
0.0041184067,
-0.0126120532,
-0.0437900275,
0.0262364224,
0.0294393413,
-0.0034684662,
0.0272979327,
0.0239852872,
-0.0530012958,
-0.0411398634,
0.014394667,
0.0223763157,
0.0176374838,
-0.0201068837,
-0.03802412,
-0.0518145189,
-0.0206595212,
0.0217111018,
-0.0212828759,
0.030165147,
0.0061552874,
-0.0176266208,
0.0077106967,
-0.021518128,
-0.0012106778,
-0.0200786535,
0.0687605739,
-0.0294720624,
-0.0054786243,
-0.0089967083,
0.0020200498,
-0.0126467673,
0.0249837674,
0.0006450558,
-0.0321030989,
-0.0275515988,
0.0103377569,
-0.021556979,
0.0209952649,
-0.0501221418,
-0.0113266734,
-0.0025629273,
0.0112837264,
0.0528681502,
0.000914763,
-0.0068211881,
-0.0579500012,
2.742e-7,
-0.0007823007,
0.0904232115,
-0.0084204422,
0.0442540087,
0.0214286111,
0.0845208988,
-0.0109875845,
0.0304510109,
0.0369613878,
-0.0142194945,
0.0614956915,
0.0248835888,
0.0661397055,
0.0321458355,
-0.0783068314,
-0.004607812,
-0.0568715222,
-0.0581275746,
-0.0357809253,
-0.0369645879,
0.0465215147,
0.0440009013,
0.0298158024,
0.0123866182,
-0.034472663,
-0.013223093,
0.0518344864,
-0.1117488667,
-0.0193800293,
0.0512626432,
0.0048189638,
-0.0707665458,
0.0309031811,
0.0164330248,
-0.0094388546,
0.0608037189,
-0.0259498749,
0.0643745512,
-0.0041564396,
-0.0339649245,
-0.0098392591,
-0.004123834,
-0.0015494473,
-0.0258699618,
0.0119772796,
-0.0674186572,
-0.0442650877,
-0.0010991603,
-0.0234992225,
-0.0484683588,
0.0219033714,
-0.0128651438,
-0.0066682026,
0.0100474916,
-0.0057669394,
-0.0183585621,
0.053650666,
-0.0436944738,
-0.0071424036,
0.0029829089,
-0.0377779342,
-0.0148884216,
0.017020775,
0.0034173131,
-0.0172570478,
0.0152846752,
0.0461749658,
2.101103169e-34,
0.0536869243,
0.0152347237,
0.0187339839,
0.0931190252,
0.0760399923,
0.0297568254,
-0.0001561963,
0.041202426,
0.0379391089,
-0.0482784808,
-0.0013415047
] |
http://arxiv.org/abs/2402.01736v1 | http://arxiv.org/pdf/2402.01736v1 | SADAS: A Dialogue Assistant System Towards Remediating Norm Violations in Bilingual Socio-Cultural Conversations | 2024-01-29T08:54:21 | 2024-01-29T08:54:21 | Yuncheng Hua, Zhuang Li, Linhao Luo, Kadek Ananta Satriadi, Tao Feng, Haolan Zhan, Lizhen Qu, Suraj Sharma, Ingrid Zukerman, Zhaleh Semnani-Azad, Gholamreza Haffari | cs.CL | cs.CL, cs.AI, I.2.7 | In today's globalized world, bridging the cultural divide is more critical
than ever for forging meaningful connections. The Socially-Aware Dialogue
Assistant System (SADAS) is our answer to this global challenge, and it's
designed to ensure that conversations between individuals from diverse cultural
backgrounds unfold with respect and understanding. Our system's novel
architecture includes: (1) identifying the categories of norms present in the
dialogue, (2) detecting potential norm violations, (3) evaluating the severity
of these violations, (4) implementing targeted remedies to rectify the
breaches, and (5) articulates the rationale behind these corrective actions. We
employ a series of State-Of-The-Art (SOTA) techniques to build different
modules, and conduct numerous experiments to select the most suitable backbone
model for each of the modules. We also design a human preference experiment to
validate the overall performance of the system. We will open-source our system
(including source code, tools and applications), hoping to advance future
research. A demo video of our system can be found
at:~\url{https://youtu.be/JqetWkfsejk}. We have released our code and software
at:~\url{https://github.com/AnonymousEACLDemo/SADAS}. | SADAS: A Dialogue Assistant System Towards Remediating Norm Violations in Bilingual Socio-Cultural Conversations: In today's globalized world, bridging the cultural divide is more critical
than ever for forging meaningful connections. The Socially-Aware Dialogue
Assistant System (SADAS) is our answer to this global challenge, and it's
designed to ensure that conversations between individuals from diverse cultural
backgrounds unfold with respect and understanding. Our system's novel
architecture includes: (1) identifying the categories of norms present in the
dialogue, (2) detecting potential norm violations, (3) evaluating the severity
of these violations, (4) implementing targeted remedies to rectify the
breaches, and (5) articulates the rationale behind these corrective actions. We
employ a series of State-Of-The-Art (SOTA) techniques to build different
modules, and conduct numerous experiments to select the most suitable backbone
model for each of the modules. We also design a human preference experiment to
validate the overall performance of the system. We will open-source our system
(including source code, tools and applications), hoping to advance future
research. A demo video of our system can be found
at:~\url{https://youtu.be/JqetWkfsejk}. We have released our code and software
at:~\url{https://github.com/AnonymousEACLDemo/SADAS}. | [
0.0762681514,
0.0263327658,
-0.0474499166,
0.0544944219,
0.0000829771,
0.014397759,
0.0283451341,
-0.0273743421,
0.0612345636,
-0.0154052051,
-0.037007831,
-0.0271211974,
0.0078031444,
0.0307689551,
-0.0166838635,
-0.0797743425,
0.0550372861,
-0.008101739,
0.0634076297,
-0.0351562388,
0.0639473051,
0.0201242361,
0.0017552517,
0.0272238608,
-0.0203923229,
0.0092185456,
0.0189436115,
-0.0323302671,
0.0487294942,
0.0028376651,
0.0342227891,
0.0655305982,
-0.0115600247,
0.041095145,
0.0000021333,
-0.0541431829,
0.0188055933,
-0.0229075924,
0.0555117615,
0.0039815656,
0.0193341691,
0.005492955,
-0.0277020279,
0.0202658512,
-0.062926352,
0.0126519762,
0.0175412204,
0.0444884114,
0.0613056011,
0.0130474698,
-0.0029834898,
-0.0117533915,
-0.027913969,
-0.0215461887,
-0.0622231588,
-0.0587456971,
0.0129602673,
0.0039980928,
-0.0013457695,
0.0203503799,
0.0050687734,
0.0189712867,
0.0034882568,
-0.0495031774,
-0.0070645348,
-0.04315418,
-0.05664308,
-0.0385028943,
-0.0119233225,
0.0919145346,
0.0083444891,
0.0114477044,
0.0213760696,
0.0242726468,
-0.0220358036,
-0.0259540323,
-0.0522499904,
-0.0048917709,
-0.0097987959,
0.0156870168,
0.0016598406,
0.0013655828,
-0.0243232828,
0.034385547,
0.0260810982,
0.04043575,
-0.0147159677,
0.0030861525,
-0.0496283397,
-0.0071341908,
0.0201323703,
-0.0501036383,
0.0646547526,
0.0256466884,
-0.0966712981,
0.0155901322,
0.0225420371,
-0.0012375459,
0.0412703678,
-0.0484498478,
-0.0152610634,
0.0714834109,
-0.0302113537,
0.0267430693,
0.0224956982,
0.043730583,
0.0046448018,
0.0042502303,
0.0087389257,
-0.0158232227,
0.001231205,
0.0002083228,
-0.000720413,
-0.0058092684,
-0.0745843053,
0.0349715427,
-0.0086823879,
0.0043438543,
-0.0505747944,
0.0468739122,
-0.030021904,
0.0131012639,
0.0033383744,
0.0207668375,
-0.0162175056,
0.0488065071,
-0.0627445579,
0.0015527369,
-0.0189989209,
-0.0739192739,
0.0263954718,
0.0082570044,
-0.0183607396,
0.0208126195,
-0.0203293338,
0.0286006592,
-0.0388499983,
0.0023122632,
-0.0118768131,
-0.0089465678,
-0.0346217379,
-0.0495136939,
-0.01772324,
-0.0638544858,
0.0682016611,
0.0112109687,
-0.0127336122,
-0.0228799488,
0.0497440733,
0.0237676278,
-0.0239860639,
0.036001645,
0.0269862823,
0.0072750715,
0.0864099711,
0.0598381758,
0.0047975318,
0.0517606363,
0.0229449235,
-0.0197065044,
0.0868995264,
-0.0132350959,
0.0014051241,
0.0372140184,
-0.0249890983,
-0.0032917853,
-0.0108664287,
0.0325895809,
-0.0161700919,
-0.0104620494,
-0.0010382731,
0.064608641,
-0.0178943109,
-0.0454503074,
0.0215389095,
0.0931420997,
-0.0723991618,
0.0427153297,
0.0182954893,
0.0185989402,
0.0124251693,
0.0077070026,
-0.0330898389,
0.004000193,
-0.0173900351,
-0.0511868894,
-0.0184797421,
0.0900299102,
-0.0399426855,
-0.0333086215,
0.0195416734,
0.010961269,
0.081029743,
-0.0063916184,
0.0093159983,
0.004370498,
0.0241845697,
-0.0099012107,
0.0202239975,
-0.0625905767,
-0.003070212,
-0.0191542543,
0.0377891399,
0.06609492,
0.0418917462,
-0.0222136527,
0.0089724939,
-0.0440984927,
-0.0007227818,
0.058702182,
0.0386622921,
-0.0358404294,
-0.0030692557,
-0.0585379675,
-0.0191604421,
0.0598281212,
-0.0188207068,
0.0249045715,
0.0152851548,
0.0199979786,
0.0044488027,
0.036311388,
-0.0193257239,
-0.023462005,
-0.0043514394,
0.0234488342,
0.0029168385,
0.0081493491,
0.0257177763,
0.0598503239,
-0.006929887,
-0.0474761948,
-0.0080522383,
0.0183631871,
0.0135059804,
0.0689158067,
-0.0897656158,
-0.0126564177,
0.0300937332,
-0.0386014543,
-0.0389260426,
0.0201687627,
-0.0016278564,
-0.0177364368,
0.0613670573,
0.0140241636,
0.0158439856,
-0.0276130512,
-0.0157193765,
-0.0175777972,
0.008322374,
-0.0275465753,
-0.0241557322,
-0.0220045336,
-0.008343826,
0.0214516055,
-0.035313271,
0.0631917119,
-0.0319776163,
-0.0140558891,
0.0342833214,
-0.0101568196,
0.0119886054,
0.0097624762,
0.0249096118,
-0.021466976,
0.0089596072,
-0.116483584,
-0.0392566696,
-0.0262805708,
-0.0214227475,
-0.0070760455,
0.025569424,
-0.022632556,
-0.0600293316,
-0.0567067154,
-0.0052875597,
0.0693147555,
0.0037627066,
0.0345093124,
-0.0111912945,
-0.0390285105,
-0.0338480137,
-0.0021294877,
-0.018619094,
0.0233972147,
-0.0278765429,
-0.036216177,
-0.0271886401,
-0.0612772368,
-0.0764514059,
0.0487247184,
-0.0184353311,
-0.0964778513,
0.0393409505,
0.0822661594,
0.0889410898,
0.0667741671,
0.0728282407,
0.0329306461,
0.0243181624,
-0.0224072542,
-0.0380611755,
-0.0223556291,
0.0152113428,
0.0720207915,
-0.0128170196,
-0.0479275659,
-0.0187270436,
0.0084655667,
0.0276149586,
0.0183880795,
-0.0297806263,
-0.0051729358,
-0.0124303596,
-0.0473218299,
-0.0238964893,
-0.0429601409,
-0.0283934101,
0.0321660973,
-0.0149392225,
0.0001638396,
0.0602839105,
-0.0028160976,
-0.0313239247,
-0.0338045098,
0.0054827407,
-0.0531048216,
0.0303436425,
-0.0270517617,
0.0422146916,
-0.0216237307,
-0.0567534901,
-0.0005329858,
-0.038577389,
0.0495523252,
-0.0482492335,
0.0260444842,
-0.0716115013,
-0.0382889472,
0.007503551,
-0.043266844,
-0.0524257869,
0.0353644006,
-0.0295205452,
0.0269964989,
-0.0064538536,
-0.0119093759,
0.0222099107,
0.0065143132,
0.042586524,
0.0114607168,
0.0054207002,
0.0413880534,
-0.0064091678,
0.0048161582,
0.0027626907,
0.0062729381,
0.018496763,
0.0160782877,
0.0373109467,
0.0380424485,
-0.0155894002,
-0.0402150005,
0.0199562423,
-0.0028283538,
0.0550576821,
0.053197585,
-0.0362615734,
-0.0160434805,
0.085725762,
-0.0253573135,
0.0150310518,
0.062480405,
0.0282109994,
-0.0396851711,
-0.0232386421,
-0.0642885193,
0.0081376843,
0.0180288423,
0.0261348505,
-0.1132162213,
0.0131805902,
0.0035345533,
0.0330794938,
-0.0475025177,
-0.0467910692,
0.0066784206,
-0.0469414368,
0.0214186125,
0.014283387,
-0.0367449969,
-0.0276117604,
-0.0033753652,
0.0181074422,
-0.0080942623,
0.048406627,
-0.0132412538,
0.0475219935,
-0.0376984067,
-0.0139935445,
-0.0393864252,
-0.001301872,
0.0124519607,
-0.0314982869,
0.0629220158,
0.0445494577,
0.0618705004,
0.0375153013,
-0.0582305044,
0.0026255518,
0.0118135521,
-0.0028718619,
-0.0397914685,
-0.0220332928,
0.036424391,
-0.0362120233,
-0.0637132153,
0.0484028161,
0.0081985183,
-0.038351398,
0.0157302208,
0.0485933386,
0.0101315007,
0.0140531445,
-0.0164023936,
0.0364818685,
0.0347092971,
0.0492162704,
0.0384104662,
-0.0365149975,
-0.0041442025,
-0.0637263507,
-0.0465955697,
-0.0261568148,
-0.0473380871,
-0.0270521045,
-0.0052556493,
-0.0334528014,
-0.0542542413,
-0.0501235574,
-0.0071234331,
0.0014084281,
0.0174964648,
-0.0695936903,
0.0230862759,
0.0177898966,
0.0219422393,
-0.0146479737,
-0.0345910005,
0.0358493216,
-0.0384782255,
0.0209338311,
-0.0189345274,
-0.0253215693,
-0.0137435431,
-0.0124016758,
0.0032439299,
-0.0749444962,
0.0268551763,
0.0532287322,
0.0106743826,
0.0301301442,
-0.0086271167,
-0.0417995192,
-0.0089042988,
0.0364658982,
-0.0090363072,
-0.0061458512,
0.0507463105,
0.015654197,
-0.0020873994,
0.0387253202,
0.0330267362,
0.0285064783,
-0.0269593652,
-0.0088331923,
-0.0803066716,
-0.0578104034,
-0.0435297303,
-0.0096582929,
-0.0088232122,
0.0098679336,
-0.0402882658,
-0.052605141,
-0.0195153933,
-0.0371426679,
-0.0430170149,
-0.0119220009,
0.032624118,
-0.0393620171,
-0.0477643386,
0.0069255577,
0.0219286084,
-0.0583988056,
0.025409285,
-0.0206937566,
0.0444124788,
0.0266310815,
0.0471638367,
-0.0163777824,
0.0080270218,
-0.0053149005,
0.0364240892,
0.0108487708,
0.0396677516,
-0.0141000971,
0.1070146114,
0.0178586692,
-0.0427188203,
0.0503532924,
0.0145493504,
0.0040668244,
-0.0190433972,
0.0214419719,
0.0402773209,
0.0308714602,
0.0027929258,
0.0400862955,
0.0236144047,
-0.001519519,
0.0129809435,
0.0414400585,
-0.0067249779,
-0.0483091623,
0.0146774128,
-0.0028247966,
-0.0095586479,
0.0240727905,
0.0266754143,
-0.121327199,
-0.024700636,
0.039731428,
-0.0393816084,
-0.0307173878,
-0.0080771223,
0.0477125645,
0.0025880872,
-0.0177603476,
-0.0088510429,
-0.0441176742,
-0.0069284788,
-0.0244207121,
-0.0028122363,
0.0186790023,
0.0018333689,
0.0023565558,
0.0662882552,
0.0003294851,
-0.0157494582,
-0.0323297195,
-0.009588006,
-0.0599662326,
-0.0388969034,
-0.0160173867,
-5.785105816e-33,
-0.0094874194,
-0.0636577681,
-0.0213068388,
0.012693245,
-0.0161437206,
-0.0597753003,
-0.0344529413,
0.023880912,
-0.0225903317,
0.0279103871,
-0.0184381362,
0.0056457636,
0.0066299872,
0.0006516733,
0.0155537548,
0.0067057377,
0.0263159368,
0.0047554895,
-0.0225589629,
0.0310702883,
0.0693329349,
-0.0169418305,
0.03063103,
-0.1282448471,
0.0312321875,
0.04148807,
-0.0104999086,
0.0490180179,
0.003735594,
0.0195754785,
-0.0093023973,
0.0630381182,
0.0152028576,
0.0178751629,
0.0019178651,
0.0905241892,
-0.0443892069,
-0.035599608,
0.0530354232,
0.0446426272,
0.0002770357,
-0.0191349518,
-0.0184463263,
-0.0532402135,
-0.0028597843,
0.0476874374,
0.0127827916,
-0.0224126559,
0.0345366187,
0.0515116453,
-0.0592931062,
0.0246313829,
-0.0351563022,
-0.022702381,
-0.0207057297,
0.0682192445,
0.0135476571,
-0.0536147431,
-0.0514395423,
0.0581362657,
-0.0087477695,
-0.0418595448,
0.0431425236,
-0.0100013884,
0.0049007144,
0.0233279448,
0.0090146204,
0.0230153073,
0.0238666739,
-0.0433145948,
0.002389651,
0.0208776873,
0.0271871239,
0.0789367259,
-0.0478428341,
-0.0506040193,
-0.0352014862,
-0.0038629728,
0.0979142562,
-0.0038602329,
-0.0317555554,
0.0227351487,
-0.0487531498,
-0.0278774127,
-0.0101902289,
-0.0508837476,
-0.0030919369,
-0.0565284677,
0.024613807,
0.0111372015,
-0.0445170924,
-0.0063937609,
0.0335854739,
-0.0185207687,
-0.0208882671,
0.0165685508,
-0.0227588266,
-0.0192733593,
0.0499253087,
0.0060178204,
-0.0668893307,
-0.0208634399,
0.0018990298,
-0.0238323137,
-0.0026914482,
0.005242778,
0.0043018381,
0.0060098474,
-0.0253330693,
0.0153244538,
0.0082852254,
0.0773637146,
-0.0391792431,
0.011512192,
-0.0077098082,
-0.0012600725,
-0.0330527835,
0.0713182241,
0.0430024005,
0.0134820947,
-0.0574220903,
-0.0357544348,
-0.0193934869,
0.0320244767,
-0.0508056022,
-0.0017284639,
-0.0380579643,
0.0102246441,
0.0097128144,
0.0262744315,
-0.0035821486,
-0.0588324405,
2.627e-7,
-0.0465078764,
0.0265434217,
0.0022603341,
0.0397239961,
0.039115876,
0.022970641,
0.0256726854,
0.0369554199,
0.0300109219,
0.0211335514,
-0.0190064237,
0.0207634252,
0.0431020036,
0.061482463,
-0.0563003458,
-0.0651509836,
-0.0326113664,
-0.0531790555,
-0.0150473053,
0.0075830859,
0.0015714121,
0.0059071025,
0.017572999,
0.0236772317,
-0.0273037851,
0.0092611779,
0.0326035358,
-0.0213400591,
-0.0097778924,
-0.0023632015,
0.0391315632,
-0.0514604449,
0.0348931774,
0.0004348721,
0.0234174635,
0.0374307074,
-0.0216570292,
0.0488845296,
-0.0039548925,
-0.0266262107,
0.0199306831,
0.0065753022,
0.0053847497,
-0.0926216319,
0.0109128179,
0.0132356016,
-0.0043139467,
0.0032894865,
-0.0042812675,
-0.0221816618,
0.0342217721,
-0.0397999771,
-0.0547012724,
0.0009437468,
0.0145292012,
-0.014959381,
0.0368481949,
-0.0294771753,
0.0109133329,
-0.0170538276,
-0.0098914327,
-0.0350757167,
0.0017001504,
-0.017304685,
-0.0725815371,
-0.0003722665,
0.0184492581,
2.032292439e-34,
0.0217701923,
-0.0013494029,
0.020486787,
0.0097210184,
0.0393632725,
0.0164171588,
-0.0133336354,
0.013243963,
0.0677969083,
0.0218156669,
-0.0148147959
] |
http://arxiv.org/abs/2401.15966v1 | http://arxiv.org/pdf/2401.15966v1 | Response Generation for Cognitive Behavioral Therapy with Large Language Models: Comparative Study with Socratic Questioning | 2024-01-29T08:53:41 | 2024-01-29T08:53:41 | Kenta Izumi, Hiroki Tanaka, Kazuhiro Shidara, Hiroyoshi Adachi, Daisuke Kanayama, Takashi Kudo, Satoshi Nakamura | cs.CL | cs.CL, cs.AI | Dialogue systems controlled by predefined or rule-based scenarios derived
from counseling techniques, such as cognitive behavioral therapy (CBT), play an
important role in mental health apps. Despite the need for responsible
responses, it is conceivable that using the newly emerging LLMs to generate
contextually relevant utterances will enhance these apps. In this study, we
construct dialogue modules based on a CBT scenario focused on conventional
Socratic questioning using two kinds of LLMs: a Transformer-based dialogue
model further trained with a social media empathetic counseling dataset,
provided by Osaka Prefecture (OsakaED), and GPT-4, a state-of-the art LLM
created by OpenAI. By comparing systems that use LLM-generated responses with
those that do not, we investigate the impact of generated responses on
subjective evaluations such as mood change, cognitive change, and dialogue
quality (e.g., empathy). As a result, no notable improvements are observed when
using the OsakaED model. When using GPT-4, the amount of mood change, empathy,
and other dialogue qualities improve significantly. Results suggest that GPT-4
possesses a high counseling ability. However, they also indicate that even when
using a dialogue model trained with a human counseling dataset, it does not
necessarily yield better outcomes compared to scenario-based dialogues. While
presenting LLM-generated responses, including GPT-4, and having them interact
directly with users in real-life mental health care services may raise ethical
issues, it is still possible for human professionals to produce example
responses or response templates using LLMs in advance in systems that use
rules, scenarios, or example responses. | Response Generation for Cognitive Behavioral Therapy with Large Language Models: Comparative Study with Socratic Questioning: Dialogue systems controlled by predefined or rule-based scenarios derived
from counseling techniques, such as cognitive behavioral therapy (CBT), play an
important role in mental health apps. Despite the need for responsible
responses, it is conceivable that using the newly emerging LLMs to generate
contextually relevant utterances will enhance these apps. In this study, we
construct dialogue modules based on a CBT scenario focused on conventional
Socratic questioning using two kinds of LLMs: a Transformer-based dialogue
model further trained with a social media empathetic counseling dataset,
provided by Osaka Prefecture (OsakaED), and GPT-4, a state-of-the art LLM
created by OpenAI. By comparing systems that use LLM-generated responses with
those that do not, we investigate the impact of generated responses on
subjective evaluations such as mood change, cognitive change, and dialogue
quality (e.g., empathy). As a result, no notable improvements are observed when
using the OsakaED model. When using GPT-4, the amount of mood change, empathy,
and other dialogue qualities improve significantly. Results suggest that GPT-4
possesses a high counseling ability. However, they also indicate that even when
using a dialogue model trained with a human counseling dataset, it does not
necessarily yield better outcomes compared to scenario-based dialogues. While
presenting LLM-generated responses, including GPT-4, and having them interact
directly with users in real-life mental health care services may raise ethical
issues, it is still possible for human professionals to produce example
responses or response templates using LLMs in advance in systems that use
rules, scenarios, or example responses. | [
0.1142656505,
0.0066364156,
-0.04584039,
-0.0034099715,
-0.0237115379,
0.0065235575,
0.0336603187,
-0.0031065105,
0.0717096925,
-0.0299916659,
-0.0064187003,
-0.020051064,
-0.0241799168,
-0.0089951595,
-0.01690007,
-0.0673935264,
-0.0076697702,
0.0345011353,
0.050431367,
0.018169418,
0.0110791754,
0.0180968065,
-0.0235337224,
0.0559892729,
-0.0217647124,
-0.0230510309,
-0.0028053815,
0.0342252143,
0.0335896984,
-0.0850062817,
0.0315468125,
0.035528373,
-0.014243573,
-0.027116064,
0.0000021783,
-0.0398125201,
-0.0048803417,
-0.0072505036,
0.018295655,
-0.031565316,
0.0499447584,
-0.0298937466,
0.0100564612,
0.048885297,
-0.0506459922,
0.0068978993,
0.0284579657,
0.0672268793,
0.0976284146,
0.067587778,
-0.0329830498,
-0.0261387005,
0.0140419044,
0.0293110069,
-0.037431471,
0.0030160367,
-0.004914274,
0.0043526902,
0.0377986208,
-0.0258922242,
0.0014475568,
0.0100522796,
-0.0097754961,
0.0082771853,
-0.0268494282,
-0.0146682346,
-0.0016934192,
-0.0512330383,
-0.0254251882,
0.0926390216,
0.0734255165,
0.0435767099,
-0.0039920527,
-0.0125003215,
0.0399326757,
-0.0100317486,
-0.0487480722,
0.0268681962,
-0.0373595729,
0.0125630554,
0.0087579563,
0.0081009911,
-0.0172740649,
-0.0048611872,
0.023217421,
0.0163164381,
-0.0228794217,
-0.0308702644,
-0.0216038749,
-0.0113846883,
-0.028157277,
-0.0445064157,
0.0191105921,
0.0710579902,
-0.0723463073,
0.0246047284,
-0.0186628997,
-0.023700038,
0.0304011814,
-0.0301705636,
0.0596428066,
0.0760014281,
-0.002934265,
-0.0479921885,
0.0529709645,
0.0423953012,
0.0281477589,
0.0263674036,
-0.0460921675,
0.042412661,
-0.0687533617,
-0.0063153459,
-0.0352044851,
0.0309815425,
-0.0474446453,
-0.0122992629,
-0.0654524565,
0.0336316638,
-0.0160354562,
-0.0160991978,
-0.0034272242,
-0.0075621097,
-0.0023554605,
0.0271019824,
-0.0476503856,
-0.0063422229,
-0.025388062,
0.0216229707,
0.030779738,
-0.06993521,
0.0347153433,
-0.0026197552,
-0.0095576402,
0.0072293752,
0.0069727474,
0.0557922423,
-0.0169937331,
0.0025191975,
-0.0427653231,
-0.0256837513,
-0.0569591746,
-0.0426092707,
0.0178916324,
-0.029936064,
0.0129877888,
0.0193019807,
0.0017101866,
-0.0172553696,
0.0563004203,
0.0395266153,
-0.0741586238,
0.0509493239,
-0.0161555242,
0.0412181765,
0.0263175163,
0.0420306884,
0.0005605905,
0.0504053198,
0.034185946,
-0.0012417166,
0.0469626822,
-0.0287143476,
0.0401608236,
0.0046413476,
-0.0105069177,
-0.0252462458,
0.0144068273,
0.0208836198,
-0.044873748,
0.0455266051,
-0.0329656377,
0.0410974957,
-0.0212954544,
0.0267387126,
0.0679854304,
0.0737824291,
0.0161725581,
0.0516719334,
0.0071927295,
0.0084299557,
0.006890669,
0.0005721588,
-0.0289673246,
0.0148791708,
-0.0416582972,
-0.0170250535,
-0.0180636719,
0.0012842693,
-0.0321227871,
-0.059228614,
-0.0302486848,
-0.0003521975,
0.0027450488,
0.029112434,
-0.0554396287,
0.0241172537,
-0.0192903858,
0.050354965,
0.0269004814,
-0.0440481864,
-0.0605266169,
-0.0723916739,
0.0403519422,
0.0911157057,
0.0450656004,
-0.026990436,
-0.0176058337,
-0.0113569712,
0.0054387706,
0.0515006036,
0.0359715596,
0.0007212453,
0.0075910515,
-0.0231529623,
0.0014796646,
-0.0088469703,
-0.0041050632,
-0.006108623,
-0.0120002199,
0.0179920793,
-0.021120239,
0.0382856131,
-0.017895272,
0.0402807556,
-0.0064945337,
-0.0270637516,
0.029220812,
-0.0027844303,
0.0683524013,
0.0343982801,
0.0064594606,
0.0212129317,
0.0315397196,
0.0129779177,
-0.0159619767,
0.0454922505,
-0.0698840395,
-0.0676492453,
0.0170597136,
0.0213761125,
0.005528349,
0.0145935565,
-0.0050570234,
0.002392597,
0.0424837247,
-0.042785868,
0.0324018635,
-0.0361610427,
0.0241857376,
0.0102596078,
0.0249952469,
-0.0505224504,
-0.000139515,
-0.0058350712,
-0.004882297,
0.014345862,
-0.0193352848,
-0.0165220108,
-0.0107289944,
0.0124785714,
0.0386443287,
-0.0300451275,
0.0021913929,
-0.0369064361,
0.0333588608,
-0.0590828732,
-0.0416218042,
-0.1364212781,
0.0218689889,
-0.0301503558,
-0.0306579899,
0.0296235066,
-0.0017094288,
-0.0242629256,
-0.0000419308,
-0.0527880937,
0.0352933444,
0.0575890131,
0.0281840097,
0.0143058645,
0.0180636216,
-0.0072439834,
-0.0286735427,
0.0162538923,
-0.0147753414,
0.0081880046,
-0.0178489611,
-0.0101995291,
-0.0351446569,
0.0036860334,
-0.0365981422,
0.0006355035,
-0.0105023375,
-0.0822745264,
0.0482154787,
0.0208296739,
0.0829594806,
0.0005307541,
0.0405906737,
0.0474476069,
-0.009603261,
-0.0133445682,
-0.053839419,
0.0278277695,
-0.0135220429,
-0.0565437898,
-0.0462929942,
-0.0722113475,
0.0116109792,
0.0377258398,
0.0499947965,
0.0227311198,
-0.0447359234,
0.0041228663,
0.024184294,
-0.013067876,
0.0085618692,
-0.0138108479,
-0.0233175438,
0.0115739135,
0.0056902394,
-0.0160657912,
0.0464276783,
0.0181448068,
-0.0196944587,
-0.0581038222,
-0.0301609598,
-0.0317627974,
0.0284112263,
-0.0173407644,
0.0267695244,
0.0197667256,
-0.0086120972,
0.0089777764,
-0.0666192472,
-0.0156404171,
-0.0573478565,
0.031468302,
-0.0680496693,
-0.016265396,
-0.0032718622,
-0.0221504103,
-0.0623506196,
-0.0356065705,
-0.0217334479,
-0.0282157529,
0.0052882009,
0.0405927263,
-0.0046373652,
-0.0192918349,
0.0266106892,
0.0356947742,
-0.0211669356,
-0.0380942598,
0.0107363816,
0.0112102544,
0.0290602334,
0.0282302089,
0.041847039,
0.0068202936,
0.0846998543,
0.059829846,
0.0149153229,
0.0023872645,
0.0111961318,
0.0080435025,
0.0003355155,
0.0328135155,
0.0292253196,
0.0422927998,
0.0781652331,
0.0054364693,
-0.0308807828,
0.0935906842,
0.0298448652,
-0.0413372964,
0.0231766067,
-0.0449386761,
-0.0434572808,
-0.0073796948,
-0.0143292844,
-0.0457255989,
-0.0156223327,
-0.0220073368,
0.0459437594,
-0.0667390972,
-0.0240860768,
0.00178215,
-0.0706927553,
0.0452161692,
-0.0045254435,
0.0243006144,
0.0014284088,
0.0243016463,
0.0075836373,
0.0383761562,
0.0069445218,
-0.018628573,
0.0073341355,
-0.0017221154,
-0.0088616675,
-0.050150916,
0.0053210375,
0.0299590472,
0.0076022688,
0.0795022026,
0.0360916145,
0.0564951561,
0.0165940449,
-0.0224494636,
-0.0041379337,
0.0349306948,
0.028306121,
-0.0019456617,
0.0241106041,
0.025269663,
0.0330361314,
-0.0295621306,
0.0616291091,
-0.0516198277,
-0.0522865988,
0.0614248961,
0.046811495,
-0.0293929242,
-0.0180781353,
-0.0459880456,
0.0690484568,
0.0483278818,
0.0396549515,
0.0301459301,
-0.0493711643,
-0.0031098092,
0.005882856,
0.00367276,
-0.0110956384,
-0.01562123,
-0.0065179598,
-0.0186726786,
0.0427340604,
-0.075301446,
-0.0574094839,
0.0096551226,
-0.00399645,
-0.0232186113,
-0.0559814386,
0.0232200511,
0.0418166332,
-0.0081979614,
0.0027579372,
0.0271939673,
0.0216343924,
-0.0228785593,
0.0111875078,
0.0472031794,
-0.0425708666,
0.0193486679,
0.0099488096,
0.00260321,
-0.0265477672,
0.0179456994,
0.0634140074,
0.0144816153,
0.0105876755,
0.0489266552,
0.0086856196,
-0.024106184,
0.0626926869,
-0.0397620946,
0.0037590223,
-0.0002083817,
0.109183237,
-0.008423388,
0.0093499664,
0.0067288494,
0.0552351512,
0.0566472635,
0.0342823491,
-0.0297417194,
-0.0449844785,
0.022781672,
0.0316696875,
0.0217317678,
-0.0150827402,
-0.0554235466,
0.0126942275,
-0.0126706883,
-0.0487963483,
-0.0187184978,
-0.0440074131,
-0.0254494529,
-0.0023282636,
-0.0233473033,
-0.0155750103,
-0.0517590903,
-0.0907632187,
0.0045532291,
-0.0386803485,
-0.0144484714,
0.0222941581,
-0.0142779304,
-0.0239965674,
-0.0094409762,
-0.0273485482,
0.0167061351,
0.0057732714,
0.0249261279,
-0.0074559604,
0.1435074955,
-0.0169751719,
-0.0153117636,
0.0250394307,
0.033763323,
-0.0510994829,
-0.0594025329,
-0.0280682668,
0.0333284363,
0.0227529071,
0.0536351688,
0.0106972978,
-0.0257387385,
0.0086318282,
0.0011567198,
0.0269497484,
-0.0143106291,
-0.0281508826,
0.0216458812,
0.0013200574,
-0.0232013669,
-0.0316134468,
0.005034911,
-0.0544664711,
-0.0005017441,
0.0108549511,
-0.0373383388,
-0.0545509756,
-0.02682372,
0.019619612,
-0.0022828036,
-0.0070414566,
-0.0088260025,
-0.0654115602,
-0.025603259,
0.0054539554,
-0.0331704579,
-0.0002088337,
-0.0633372664,
-0.002253911,
0.0353079922,
0.0113180429,
-0.023798706,
-0.0433890149,
0.0496350229,
-0.0549132787,
-0.0163972117,
-0.0579006262,
-5.735280284e-33,
-0.0239716582,
-0.0225932263,
0.0219956897,
-0.0074274335,
-0.0273341816,
-0.0491146259,
-0.0242288373,
0.0323468111,
-0.007159356,
0.0132124424,
-0.0157886706,
0.0293368399,
-0.0043650097,
0.0374344848,
-0.0223483238,
-0.0189935677,
0.0077920295,
-0.0212261416,
-0.0276998281,
0.0265839342,
-0.0207940321,
0.0308864899,
0.018514391,
-0.1050980091,
0.0043023699,
0.0063312296,
-0.0127978465,
-0.0037786409,
-0.0081237564,
0.0275197905,
-0.0252840426,
0.0714871585,
0.0462855101,
0.0246444736,
0.020940084,
0.035314288,
-0.0205570795,
-0.0351047628,
0.0259931982,
-0.0320121758,
-0.06926734,
-0.0347891338,
0.0186905228,
-0.0154840061,
-0.0282707792,
-0.0310966633,
0.0127104595,
-0.0440072343,
0.0183173958,
-0.0086318022,
-0.0504960306,
-0.0355172567,
-0.0108802654,
-0.0308823753,
-0.0869212076,
0.0834821686,
-0.0530542918,
0.02319053,
-0.0368382595,
0.0179409031,
0.0321530998,
-0.0483259223,
0.00266579,
-0.0281944573,
-0.016538281,
0.0213954337,
-0.0023396949,
0.0202857628,
-0.0496809892,
-0.0133498386,
-0.0298802275,
0.0367538258,
0.0577978641,
0.0867364481,
-0.0318987705,
-0.0637407228,
-0.0293823667,
0.0148577411,
0.0319167301,
-0.0077221883,
0.0444481634,
0.0046103136,
-0.0181035213,
-0.0202092957,
-0.0228757747,
-0.0131038977,
-0.0182438027,
-0.0432824045,
0.0425674245,
-0.0045646518,
-0.0253432654,
-0.0445726961,
0.0015330651,
-0.0396334343,
0.026351247,
-0.0112274401,
-0.0253522322,
0.0252267756,
-0.0092021879,
-0.0068330923,
-0.039896097,
0.0049239341,
0.0144506879,
0.0188041311,
0.0428610295,
0.0104133431,
0.0131502077,
-0.0096554756,
-0.0228908826,
-0.0111173522,
-0.0091506243,
0.1119457632,
-0.0027179325,
0.0156816915,
-0.0684502125,
0.020104846,
-0.0326203406,
0.0422153585,
-0.0108393887,
-0.0331578068,
-0.0786747709,
-0.0368529074,
0.0086412299,
0.0204214305,
-0.0417174287,
-0.0020663661,
-0.0084504532,
0.0223224722,
0.0407134704,
-0.0107972957,
-0.015058008,
0.0069236192,
2.839e-7,
-0.0015539179,
0.0365320891,
-0.0233390164,
0.0080353543,
0.0294245034,
0.0376626365,
0.0226076767,
0.0613091029,
0.0873352587,
-0.0342374071,
0.0023301085,
0.0101317354,
0.0571649894,
0.0468433946,
-0.0236718841,
-0.037583679,
-0.0332446918,
-0.0386854932,
-0.0416289493,
0.0345690735,
0.1100856811,
0.0247391146,
-0.0133151403,
-0.0326376185,
0.0002111751,
0.0336297639,
0.0305993911,
-0.0674490929,
-0.0176858101,
0.0382237136,
-0.0046782722,
-0.0145772733,
0.0380302928,
-0.0021905445,
0.011547762,
0.0009697492,
-0.0249842759,
0.0827449635,
-0.0165340863,
-0.0637510493,
0.0081934892,
0.0116903968,
0.0360773988,
-0.05408125,
0.0290975012,
-0.1056576967,
-0.0410829186,
-0.0052234107,
-0.0082158716,
-0.0543569885,
0.013417826,
-0.0385431126,
-0.0243819822,
0.0558002517,
0.0254392102,
0.0112325605,
0.0367635116,
-0.0310707185,
0.0106941918,
-0.0013674032,
-0.0449836254,
-0.0327874869,
-0.0177912079,
0.0109327352,
-0.0228708275,
-0.0578798987,
0.0291065313,
2.35035894e-34,
-0.0029032207,
0.020862408,
0.0407051221,
0.0133932512,
0.028230723,
0.0399563983,
0.0181392059,
0.0468276218,
0.0194957051,
-0.0720075071,
-0.0242053885
] |
http://arxiv.org/abs/2401.15963v2 | http://arxiv.org/pdf/2401.15963v2 | NoFunEval: Funny How Code LMs Falter on Requirements Beyond Functional Correctness | 2024-01-29T08:47:31 | 2024-02-02T18:11:27 | Manav Singhal, Tushar Aggarwal, Abhijeet Awasthi, Nagarajan Natarajan, Aditya Kanade | cs.SE | cs.SE, cs.AI, cs.CL, cs.LG | Existing evaluation benchmarks of language models of code (code LMs) focus
almost exclusively on whether the LMs can generate functionally-correct code.
In real-world software engineering, developers think beyond functional
correctness. They have requirements on "how" a functionality should be
implemented to meet overall system design objectives like efficiency, security,
and maintainability. They would also trust the code LMs more if the LMs
demonstrate robust understanding of requirements and code semantics.
We propose a new benchmark NoFunEval to evaluate code LMs on non-functional
requirements and simple classification instances for both functional and
non-functional requirements. We propose a prompting method, Coding Concepts
(CoCo), as a way for a developer to communicate the domain knowledge to the
LMs. We conduct an extensive evaluation of twenty-two code LMs. Our finding is
that they generally falter when tested on our benchmark, hinting at fundamental
blindspots in their training setups. Surprisingly, even the classification
accuracy on functional-correctness instances derived from the popular HumanEval
benchmark is low, calling in question the depth of their comprehension and the
source of their success in generating functionally-correct code in the first
place. We will release our benchmark and evaluation scripts publicly at
https://aka.ms/NoFunEval. | NoFunEval: Funny How Code LMs Falter on Requirements Beyond Functional Correctness: Existing evaluation benchmarks of language models of code (code LMs) focus
almost exclusively on whether the LMs can generate functionally-correct code.
In real-world software engineering, developers think beyond functional
correctness. They have requirements on "how" a functionality should be
implemented to meet overall system design objectives like efficiency, security,
and maintainability. They would also trust the code LMs more if the LMs
demonstrate robust understanding of requirements and code semantics.
We propose a new benchmark NoFunEval to evaluate code LMs on non-functional
requirements and simple classification instances for both functional and
non-functional requirements. We propose a prompting method, Coding Concepts
(CoCo), as a way for a developer to communicate the domain knowledge to the
LMs. We conduct an extensive evaluation of twenty-two code LMs. Our finding is
that they generally falter when tested on our benchmark, hinting at fundamental
blindspots in their training setups. Surprisingly, even the classification
accuracy on functional-correctness instances derived from the popular HumanEval
benchmark is low, calling in question the depth of their comprehension and the
source of their success in generating functionally-correct code in the first
place. We will release our benchmark and evaluation scripts publicly at
https://aka.ms/NoFunEval. | [
0.0307104513,
0.0659166351,
-0.0065734214,
0.0373259671,
-0.0233118162,
-0.0113313925,
-0.0426023528,
-0.0114243673,
-0.0129173063,
-0.0050174487,
-0.0165526792,
-0.0068173516,
0.0443621278,
0.042763304,
0.0078869555,
-0.0081655243,
0.0525634326,
-0.0244711135,
-0.0538855568,
0.0281792022,
-0.0144710438,
-0.000154245,
-0.0154327629,
0.0559545346,
0.0028462303,
-0.0635285825,
-0.0132191433,
0.0490085408,
-0.0189800821,
-0.0289048571,
0.0293363109,
0.0485632382,
-0.0398189574,
-0.011817581,
0.000002463,
-0.0502786972,
-0.0189489722,
0.0122886719,
0.0057983021,
0.023075968,
-0.004148711,
-0.0383905657,
0.0008369767,
0.0118744392,
-0.0284316186,
0.0041385693,
0.0561422072,
-0.0119527569,
0.0213802122,
0.0415214486,
0.0059044333,
0.0198755227,
-0.0039114272,
0.0095067611,
0.0752078146,
-0.0260976944,
-0.0184917934,
0.0032453353,
0.0258827917,
-0.0061198226,
0.001666693,
0.0152668608,
0.0451714471,
0.0101351989,
0.0439263247,
0.0352866501,
-0.0309361257,
-0.0198748447,
0.0355867781,
0.0100019239,
-0.0088557033,
0.0012122828,
-0.0187576879,
0.0136288861,
-0.063023597,
-0.0396881811,
-0.010799394,
-0.0135118132,
0.0077294847,
0.0008282976,
0.0042436365,
0.0109851109,
0.0331575461,
0.0001984709,
-0.0099837286,
0.067275323,
-0.0400158688,
-0.0173011031,
0.0244003404,
-0.0435155928,
-0.028461121,
-0.012129643,
-0.030394759,
0.0480581559,
0.0118874302,
0.0001045502,
0.0101542668,
0.0395229012,
0.0809107497,
-0.0577469505,
0.0038394073,
0.0007097037,
-0.0316003896,
0.0168136936,
-0.0641938075,
0.0715917572,
0.0162710398,
-0.0276235584,
-0.0267760064,
0.0305690896,
-0.0539076999,
-0.0132132117,
0.0154485926,
0.0561268218,
0.0193533711,
0.0513732322,
-0.0374339372,
0.0473890789,
-0.0351769812,
0.1184037328,
-0.0652687997,
-0.0188026521,
-0.0116082523,
0.0033971756,
-0.0167540684,
-0.0267317351,
-0.0079196561,
-0.0395938158,
-0.0264686383,
-0.0854924545,
0.0190751478,
0.0023776721,
-0.0999211967,
-0.0009993716,
0.0271387119,
0.0532614253,
0.0312682092,
0.0176815707,
-0.0325913131,
-0.0522189587,
0.0105338227,
-0.0057459697,
-0.0190134756,
-0.1013759151,
-0.0827323124,
0.0221223161,
0.0207314435,
-0.064377442,
-0.0099131158,
-0.0145630091,
0.0070327003,
0.0565903559,
-0.0308165215,
0.0195268597,
0.0177772678,
0.0658672974,
-0.0152713824,
0.0829526782,
0.0118753882,
-0.0061660986,
0.0431055427,
-0.013656971,
0.0096040601,
0.0114009799,
-0.0499891788,
0.0020540522,
0.0042702979,
0.0480418205,
0.0248686783,
-0.030423779,
0.0118526565,
0.0283756629,
0.0049562217,
0.0155374464,
0.0028146789,
0.0327590443,
0.0271294247,
-0.0041096965,
0.0290136915,
0.0397422351,
0.0202058144,
-0.0135118011,
-0.0475208685,
0.0264087077,
-0.0462731235,
0.0102892276,
-0.0280181505,
-0.0172158349,
-0.0286449548,
-0.0848189518,
-0.0135041969,
-0.0357491672,
0.0046489914,
-0.0211561248,
0.0433601439,
0.0128599508,
-0.0222805552,
0.0086675212,
-0.0063840342,
-0.0377311036,
-0.0379322022,
-0.0428782627,
0.009926999,
0.0363014117,
0.0459145978,
-0.0772423223,
-0.0593193397,
0.0135508375,
-0.0291953739,
-0.0055049541,
0.0573366322,
-0.0219178889,
0.0077969613,
-0.0456740409,
-0.0128330886,
-0.0394161381,
0.0113513423,
0.0123220515,
0.0012062674,
-0.0172056518,
0.0384962074,
-0.0248025749,
-0.0885987654,
0.0585967489,
-0.003651011,
0.0073094498,
0.011663476,
-0.0031194107,
0.0182423666,
-0.0146402754,
-0.007317449,
0.0203435458,
0.0285040252,
-0.0077443887,
0.011417211,
-0.0439205803,
-0.0412911549,
0.0346799642,
0.0012280216,
-0.0368529744,
-0.0225594081,
0.0889400989,
-0.032691095,
0.0161268283,
0.0386107974,
0.0051622777,
-0.0150359869,
-0.0035666062,
0.0349868461,
-0.0274291821,
-0.0554443076,
-0.0221532378,
0.0097376732,
0.0178526472,
0.0470955558,
-0.0644091517,
0.0536646247,
0.0218151845,
0.0361661613,
-0.0672408789,
0.0194978658,
-0.0587181821,
0.0572386459,
0.0028548276,
0.0598065481,
-0.0176650994,
-0.0222079791,
-0.0786959603,
0.0120291449,
-0.0061675874,
0.0162773877,
0.0202854667,
0.0108696911,
0.0198812224,
0.0101271169,
0.0018869671,
-0.0264570732,
-0.0427993275,
0.0166470818,
0.0510921851,
-0.0178561844,
-0.0533914417,
0.0072068344,
0.0216416959,
-0.0411211513,
0.0516320914,
-0.0384440683,
0.0476517491,
-0.0215674657,
0.0710016862,
-0.0221343916,
-0.0111971572,
-0.0288503524,
-0.0021961834,
0.0375394076,
-0.0172032155,
0.0359117277,
-0.0138674416,
-0.0101169711,
0.0136624873,
0.0589975715,
0.0624387339,
-0.0476020463,
-0.0097013526,
0.037913911,
-0.043460343,
-0.0249858499,
-0.0305918418,
0.0711454675,
0.0732461512,
-0.0523514487,
-0.0308218766,
-0.0357406065,
-0.0000582853,
-0.0238303319,
0.0136018684,
-0.0442748889,
-0.0159827825,
0.0138144242,
0.0474756546,
-0.0360064432,
-0.0701303259,
0.0407428741,
-0.0454465896,
-0.0370584913,
-0.0365017541,
-0.0314758234,
-0.0434300341,
0.0943180397,
-0.0180999395,
0.0054741832,
0.0780861601,
-0.0169752445,
-0.0242648944,
-0.0004992132,
-0.0071744104,
-0.0505041443,
0.0927680805,
-0.0204962529,
0.0111804307,
-0.0146443546,
-0.0049168468,
-0.0717480183,
0.013177949,
-0.0007592401,
0.022146441,
-0.0201239176,
0.0242510065,
0.0183088109,
0.0142677026,
-0.031691175,
0.0546688847,
0.0542327464,
-0.0597980358,
-0.0078977505,
-0.0948126838,
-0.0230085757,
-0.0555701256,
-0.0255333167,
-0.0239867792,
0.0082216142,
0.0430569574,
0.0111518819,
0.0278353356,
-0.0618244261,
-0.0173609983,
0.0035605179,
-0.0303100217,
-0.0287049096,
0.0525505841,
0.0281517394,
-0.0279212315,
-0.044925984,
0.0648673028,
0.0498591885,
-0.047108721,
0.0299951956,
-0.0242064465,
0.0484530851,
0.089480713,
0.0316069089,
-0.0707594752,
-0.0156107619,
-0.0000587736,
0.0693967566,
-0.0126138218,
-0.0277902968,
-0.0235677175,
-0.060178116,
0.0085544717,
0.020867411,
-0.035419412,
-0.0420525111,
-0.0597075373,
-0.0043413439,
0.0027999952,
-0.0662308782,
0.0473451689,
0.0187089834,
0.0070347111,
0.0095415758,
-0.0960455984,
-0.0516613498,
0.0634310022,
-0.0217352584,
0.0015705401,
0.0906818882,
-0.0331049152,
0.0087365871,
-0.0170213562,
0.0370884836,
-0.0198937934,
0.0245968848,
-0.0263783708,
0.0285138115,
0.0045887511,
0.0046933214,
0.0448557585,
-0.0120825833,
-0.068180725,
-0.0223839898,
0.0778670982,
0.0243820921,
-0.0275727473,
-0.016882373,
-0.0244803615,
0.0090372479,
-0.0031756081,
0.0406297296,
0.0561608896,
-0.0060611987,
-0.0322047099,
-0.0646178648,
-0.017193839,
-0.0360665359,
-0.0133541264,
0.0005758754,
0.0108632548,
-0.0226870645,
-0.0340079404,
0.0117446138,
-0.0095753865,
0.0483725853,
0.0114376796,
-0.0427407287,
0.0185420644,
-0.0020469325,
-0.0662817881,
0.0136333639,
-0.0078366958,
0.0205518212,
-0.0183005836,
-0.0109758787,
-0.0072543439,
-0.0022284994,
-0.0121523179,
-0.0297956746,
-0.0270299017,
-0.0128721809,
-0.0191560369,
0.0630819947,
0.032423798,
-0.0252855811,
-0.0575789735,
-0.0297711696,
-0.0434280783,
0.0518873148,
-0.0412977152,
-0.0133272251,
0.0590784475,
-0.0488763228,
0.0310114436,
-0.020805113,
-0.0435896888,
0.0440977328,
-0.0015304749,
0.0108681014,
-0.0071591404,
0.0595471822,
0.0226661526,
0.0204265062,
0.0291657839,
0.0047599035,
-0.0433097295,
0.0269824062,
-0.024523437,
0.0343182608,
-0.0316235609,
-0.0147574609,
0.0034967868,
0.0101043871,
0.0236733146,
0.0364943221,
0.0231465865,
-0.0349923745,
0.0681064129,
0.0428048596,
0.0328555778,
0.0395717584,
0.0644710138,
0.0460948832,
-0.0041821171,
-0.0415210463,
-0.007183631,
-0.0327988751,
0.0323930532,
0.0024260315,
0.0287875924,
0.0160766318,
0.1074639559,
0.0074647143,
0.0197290778,
-0.0241834763,
-0.069211334,
0.0177097302,
-0.0258612446,
0.0136516979,
0.0618559793,
0.0076337596,
-0.0385597087,
0.0415753946,
-0.04646172,
0.0264658444,
-0.0303650778,
-0.0476236418,
-0.0168799143,
0.0503395721,
-0.0005058806,
-0.0532299839,
0.0930081755,
-0.058719486,
-0.0192952622,
-0.0465119444,
-0.0209202208,
-0.053785149,
0.0281240828,
0.0772243515,
-0.0073988503,
-0.0431077741,
-0.057562951,
-0.0523125529,
-0.072916694,
0.0535516292,
-0.0070570903,
0.0490884148,
0.0013858415,
0.016400056,
0.0943164602,
0.0134993847,
-0.0279328395,
-0.0282090008,
-0.0413141325,
0.0034465462,
-0.0058034072,
-0.0291200895,
-6.196140191e-33,
0.0014802561,
-0.0155951371,
-0.0071847765,
0.041759707,
0.0182173215,
0.0381681994,
-0.0316457339,
0.0065342612,
-0.0060188556,
-0.0407651179,
0.0359316841,
0.0081568332,
-0.0223154891,
-0.0393519141,
0.0567816384,
0.0146713518,
0.0364604481,
-0.0036190092,
-0.0596565157,
0.0033263508,
-0.0152960019,
0.0057421466,
-0.0093313055,
-0.0430198982,
0.0430260003,
-0.0177281238,
-0.0040861559,
0.0092397369,
0.0375588052,
0.0327654816,
0.0109023489,
0.0314455368,
0.025697032,
-0.0337699279,
-0.0187465865,
-0.0266886353,
-0.008137567,
-0.043270763,
0.0244279671,
-0.0163206626,
-0.0105309533,
-0.036720112,
0.0618260428,
0.0338920839,
-0.0124837486,
-0.0192750338,
0.0066565429,
-0.0038920844,
0.0057312851,
-0.006973967,
-0.0357123129,
0.051366318,
-0.0520930514,
0.0062383213,
-0.0172228292,
0.0661534891,
-0.010255117,
-0.0132002318,
-0.0098067243,
0.0056723282,
0.0095188301,
0.0428413525,
0.0577256233,
0.0141102131,
-0.0481130891,
0.0345756747,
0.0205792282,
0.0665978864,
-0.0188172478,
0.0554355346,
0.0658382103,
0.0498453565,
0.0092639262,
-0.0006064592,
0.0522318296,
-0.021119101,
-0.0321989432,
0.0628582612,
-0.0490167737,
0.0344035849,
0.0281921569,
0.0084346542,
-0.0775838718,
-0.0198714733,
-0.0270756446,
0.0247237757,
-0.0168937203,
-0.0391269177,
0.0291882046,
-0.010779568,
-0.0376456305,
0.0283869486,
-0.0405301042,
-0.0362802073,
0.0092145,
0.0065422524,
-0.0258035902,
0.0194742884,
-0.0085309278,
0.0366611369,
-0.018277755,
-0.030322643,
-0.0059390236,
0.020283008,
0.0884976983,
-0.0011991233,
0.0500042737,
0.0259119999,
-0.0566874556,
-0.0178605076,
-0.0209512375,
0.0064486773,
-0.0140284188,
0.0470097885,
0.0055999248,
-0.0354345068,
0.0196883064,
0.0131535809,
-0.0346916988,
-0.0475557446,
-0.0603195764,
0.0493670851,
0.0047375648,
0.0360775292,
0.0019115765,
-0.0656219944,
-0.0477698259,
0.0199292172,
0.0545185395,
0.0024046164,
-0.0501345135,
0.0280518867,
3.221e-7,
0.002784142,
-0.024335999,
-0.0168785118,
0.0367208086,
-0.0024869349,
0.0590217859,
-0.0464169197,
-0.0027099005,
-0.0075360881,
-0.0006203013,
-0.0308143571,
0.0292341597,
0.0461369865,
0.0480474606,
-0.020399943,
-0.0449997224,
0.0254194047,
-0.0001269467,
-0.0376981273,
0.0267521739,
0.029316131,
-0.0127940318,
0.0017675095,
-0.0157974511,
0.0196125414,
-0.0376336016,
-0.0207971111,
-0.004623001,
0.0465965271,
0.0480005741,
-0.0277551394,
0.02411668,
0.0175193157,
0.0106471209,
0.0161612816,
-0.0367352553,
-0.0146581307,
0.0868152529,
-0.00370698,
0.0291539449,
0.0129050575,
0.0095702084,
-0.0026841462,
-0.0589002743,
0.0364414491,
0.0137953889,
-0.0288902856,
0.0031408495,
-0.0093899854,
-0.0095281061,
0.0139653049,
0.0300994366,
-0.0271856301,
-0.0169194303,
-0.0027440982,
-0.0433471203,
-0.0397675298,
-0.007636501,
-0.0311068818,
0.0332844555,
-0.0026961155,
-0.0456354953,
-0.0389796831,
0.05663082,
0.010540436,
-0.0257789642,
0.0262118671,
3.182183512e-34,
0.0108426241,
-0.0352811702,
0.0249268226,
0.0088257343,
0.0180465095,
-0.0010265521,
-0.0361846387,
0.0169025529,
-0.0145652788,
0.0385130458,
-0.0062292758
] |
http://arxiv.org/abs/2402.01735v2 | http://arxiv.org/pdf/2402.01735v2 | VIALM: A Survey and Benchmark of Visually Impaired Assistance with Large Models | 2024-01-29T08:28:32 | 2024-02-10T14:08:55 | Yi Zhao, Yilin Zhang, Rong Xiang, Jing Li, Hillming Li | cs.CL | cs.CL, cs.AI, cs.CV | Visually Impaired Assistance (VIA) aims to automatically help the visually
impaired (VI) handle daily activities. The advancement of VIA primarily depends
on developments in Computer Vision (CV) and Natural Language Processing (NLP),
both of which exhibit cutting-edge paradigms with large models (LMs).
Furthermore, LMs have shown exceptional multimodal abilities to tackle
challenging physically-grounded tasks such as embodied robots. To investigate
the potential and limitations of state-of-the-art (SOTA) LMs' capabilities in
VIA applications, we present an extensive study for the task of VIA with LMs
(VIALM). In this task, given an image illustrating the physical environments
and a linguistic request from a VI user, VIALM aims to output step-by-step
guidance to assist the VI user in fulfilling the request grounded in the
environment. The study consists of a survey reviewing recent LM research and
benchmark experiments examining selected LMs' capabilities in VIA. The results
indicate that while LMs can potentially benefit VIA, their output cannot be
well environment-grounded (i.e., 25.7% GPT-4's responses) and lacks
fine-grained guidance (i.e., 32.1% GPT-4's responses). | VIALM: A Survey and Benchmark of Visually Impaired Assistance with Large Models: Visually Impaired Assistance (VIA) aims to automatically help the visually
impaired (VI) handle daily activities. The advancement of VIA primarily depends
on developments in Computer Vision (CV) and Natural Language Processing (NLP),
both of which exhibit cutting-edge paradigms with large models (LMs).
Furthermore, LMs have shown exceptional multimodal abilities to tackle
challenging physically-grounded tasks such as embodied robots. To investigate
the potential and limitations of state-of-the-art (SOTA) LMs' capabilities in
VIA applications, we present an extensive study for the task of VIA with LMs
(VIALM). In this task, given an image illustrating the physical environments
and a linguistic request from a VI user, VIALM aims to output step-by-step
guidance to assist the VI user in fulfilling the request grounded in the
environment. The study consists of a survey reviewing recent LM research and
benchmark experiments examining selected LMs' capabilities in VIA. The results
indicate that while LMs can potentially benefit VIA, their output cannot be
well environment-grounded (i.e., 25.7% GPT-4's responses) and lacks
fine-grained guidance (i.e., 32.1% GPT-4's responses). | [
0.0224137157,
0.0036822534,
-0.0015379412,
0.0276868455,
-0.0205095075,
-0.0026004778,
0.0079166014,
-0.0149288261,
0.0574316755,
-0.0595703311,
-0.0248216148,
0.0126422439,
0.0126834949,
0.0354701132,
0.0068709003,
-0.1182847992,
-0.018095579,
-0.0540095605,
-0.0504316501,
-0.0017007466,
-0.0124032954,
0.0101399459,
-0.0521477386,
0.0606069639,
0.0084932614,
-0.0069979131,
-0.012089313,
0.0266905911,
-0.0059317662,
-0.0563406944,
-0.0628987923,
-0.0028337699,
0.0718508139,
0.0020541891,
0.0000022191,
-0.0447931327,
0.024845846,
0.0132780988,
0.0243476015,
-0.0382072777,
0.0081302458,
-0.073910214,
0.0124886753,
0.0100992657,
0.0043762685,
-0.0098672304,
0.0463373512,
0.0655794889,
0.025295062,
0.0059563736,
-0.0130801685,
0.0076777725,
0.0553175993,
0.022353949,
-0.0183822587,
-0.0174184144,
0.0026342329,
0.0020174254,
0.0573358871,
-0.0080976924,
-0.0208887346,
0.0500810593,
0.0210704766,
0.0119888699,
0.0024299547,
0.0495350882,
0.0235516243,
-0.035645701,
-0.0406791531,
0.0736446977,
0.0088103721,
0.0423466414,
-0.0125401393,
-0.0006243059,
-0.0130000599,
0.0458723381,
-0.0365996361,
0.03984892,
-0.015206649,
0.0163437519,
-0.0757954493,
0.010653791,
-0.0034502132,
0.026972821,
-0.016595263,
-0.0020925424,
0.0229317918,
-0.0118546383,
0.0238750465,
-0.0102459081,
-0.0383467376,
-0.0427280441,
0.0532027595,
0.0316794515,
-0.0340279639,
-0.0048414767,
-0.0343804993,
0.0236210339,
-0.0252983179,
-0.0479167812,
0.0769631788,
0.0640838072,
-0.081362091,
0.0590007976,
-0.021562716,
0.0611900091,
-0.0481748953,
0.0114146229,
0.0277371239,
0.0333652571,
-0.036049556,
-0.0086963717,
-0.0423086621,
-0.0158849079,
-0.1123268828,
-0.047204826,
0.0101880785,
0.0216852631,
0.0206433982,
-0.0113698989,
-0.0230691228,
0.0463936552,
0.0091735376,
-0.0194359012,
-0.0405460894,
0.0006892795,
-0.0295653436,
0.0463121943,
-0.0396631174,
-0.0679635629,
0.0304503944,
0.0246672314,
0.0137769375,
-0.0360966846,
-0.0151149677,
0.1392664611,
0.0118339052,
-0.063143231,
-0.0256020855,
0.0030597684,
0.0112582063,
-0.0496645272,
0.0355427079,
-0.0201896541,
-0.0189172309,
-0.0033402811,
0.0355318338,
-0.0576490089,
0.0262711346,
0.0154983336,
-0.003770174,
0.1020994559,
0.0040022563,
0.0180933177,
0.0430922434,
0.0027199574,
-0.0004209678,
0.0266881641,
0.006736022,
0.0200032759,
0.0191208497,
-0.0253258534,
0.0268430822,
0.0108073726,
-0.0739166886,
-0.0059581012,
-0.0766412765,
-0.0057740342,
0.0025108911,
0.0480631478,
0.0131497188,
0.057004828,
-0.0067445748,
0.0174043868,
0.0800593048,
0.0374524184,
0.0299119521,
-0.0290382598,
0.0391172171,
-0.010357229,
0.0035499295,
-0.0117866015,
-0.0441511087,
0.0109971641,
-0.0154901547,
-0.0196923278,
-0.0074126422,
0.022955846,
-0.008505227,
-0.0333282277,
-0.0200153124,
-0.0042783036,
0.0279565584,
-0.0121677089,
0.101955004,
-0.0083653443,
-0.0261249579,
-0.0447195023,
-0.0127974693,
-0.0058255317,
-0.0418839343,
-0.0173497107,
0.0034151569,
0.0682478249,
0.0149031337,
-0.0165317152,
0.0343931615,
0.0198531654,
0.023826465,
0.0779567361,
-0.0420048274,
0.03648559,
0.0358101502,
-0.0287858956,
-0.037238881,
0.0141714988,
-0.0087617543,
0.027414443,
-0.0061242892,
0.0232160743,
0.0045841937,
-0.0164939668,
0.0017841702,
0.0621261075,
0.0221672226,
-0.0074104331,
-0.0288353655,
0.0202572402,
-0.0082753906,
0.0548549034,
-0.0213693678,
0.0079566408,
0.0109181935,
-0.0475522839,
0.0072596539,
0.0212483946,
-0.0391823873,
-0.0024548317,
0.0233972967,
-0.0577323548,
-0.0012611859,
0.0533037335,
0.021780258,
-0.0150326807,
0.0060668765,
-0.025078794,
-0.0004540675,
-0.0446276963,
0.0010076707,
0.0370765515,
-0.0088265352,
-0.0317285843,
0.0260206517,
-0.0286593661,
-0.011826382,
0.0140426029,
-0.0147280945,
-0.0117606036,
-0.0609484352,
0.0038753643,
0.0207619742,
0.0107240044,
0.0133290682,
0.0000626001,
0.0328373648,
-0.0289230607,
-0.006029611,
-0.0892562941,
-0.0784088969,
-0.0135950577,
-0.0262935571,
-0.0229929909,
-0.041434519,
-0.016292518,
-0.033336781,
0.0160539076,
0.0484477244,
0.0101982458,
0.0002475975,
0.0105790347,
0.0002317842,
-0.039786011,
-0.0246858448,
0.0146946544,
0.0112598222,
0.0239144526,
0.040771801,
-0.093691662,
-0.0796009749,
-0.0293335821,
-0.0099399015,
0.0475013703,
-0.0197879709,
0.0238953233,
0.0497542359,
-0.0135284448,
0.0599709712,
0.0284489729,
-0.0066865687,
0.0488252454,
0.0489835776,
-0.0006463308,
-0.0302470215,
-0.0017030685,
-0.0048547047,
0.0263959691,
-0.016096944,
0.0269161761,
0.0079414099,
0.0194769502,
-0.0154640041,
0.046605017,
-0.0308069419,
0.0445727967,
-0.0289809164,
-0.0390752628,
-0.0516418144,
0.0342024155,
-0.0438627228,
0.001835116,
-0.0035094395,
-0.0812145099,
-0.0071673482,
-0.0459469818,
0.0002726631,
0.0458209403,
0.009695027,
-0.0149751203,
0.1040292531,
0.0090114716,
0.0024755651,
-0.0197157711,
-0.0387859605,
-0.0082869092,
-0.0037195969,
-0.0361292437,
0.0110743353,
0.0304766018,
-0.1162222698,
0.0391897708,
-0.015062687,
-0.0123260822,
-0.0081582256,
0.0179467928,
-0.0424700268,
0.0257683769,
-0.0549466573,
-0.0007263566,
-0.0232101977,
0.0207575783,
0.0267666355,
0.0683820248,
0.0295730177,
-0.0252824705,
-0.0417829007,
0.0213724803,
0.0047198171,
0.0568904653,
-0.026332289,
-0.0179353822,
0.0237749163,
0.0508314073,
-0.0264022462,
-0.029513279,
-0.0425217859,
0.012371581,
0.0438606106,
0.0413844474,
0.0020825176,
0.0376644805,
0.069743365,
-0.04395945,
0.0051086689,
0.0482756197,
0.0221585594,
-0.0402562842,
0.0176649094,
-0.0263091121,
-0.0317268595,
0.0283722784,
-0.0058053951,
-0.0408306159,
-0.0463893674,
0.0142805194,
0.0380854607,
0.021533899,
-0.0213393942,
-0.003487367,
-0.0863915384,
0.0509859957,
0.0036978396,
-0.0233610012,
-0.0497077778,
-0.0101775108,
0.0201919489,
0.0267464016,
-0.0078243408,
0.0229341388,
-0.0357343294,
-0.0230573118,
-0.00121866,
0.0040684137,
-0.0471552275,
-0.0081142355,
-0.0372746512,
0.027260473,
0.0450872034,
0.0019729538,
-0.0065975222,
-0.0194590501,
0.0497453175,
0.0247762855,
0.0721615106,
-0.0488903597,
0.0054032798,
-0.0253901258,
-0.0326597169,
-0.0017042789,
0.0486500785,
-0.0512998439,
-0.0318658836,
0.016452346,
0.0269625597,
-0.063121818,
0.0291579291,
-0.0553181246,
-0.0115763927,
-0.001302115,
0.0305118151,
0.0100451633,
-0.0355485305,
-0.0516000427,
-0.0198271349,
-0.0500691608,
0.0157864038,
-0.044559218,
-0.0160783846,
-0.0339756906,
-0.0163969025,
-0.0687746555,
-0.00661994,
0.064932026,
0.0874189436,
0.0113891037,
-0.0274100583,
0.0188557152,
0.0656801686,
-0.0047085793,
0.0773866251,
-0.0205363892,
0.0313125141,
-0.0666563734,
0.0193487499,
-0.0053824056,
-0.0829615295,
-0.0067804256,
0.0047148764,
0.0218173061,
-0.0237378757,
0.0203538612,
0.0426940545,
0.0243672598,
0.0015494381,
0.0227571242,
-0.05101531,
-0.0014616032,
0.0234559942,
0.0058103767,
0.0478287041,
-0.0005173071,
0.010390562,
-0.0102727935,
0.0050272667,
-0.0140540898,
0.0384783447,
-0.0192241576,
0.0395261459,
0.0132934349,
-0.0003624748,
-0.0102928123,
0.0132661192,
0.0829185322,
-0.0077148885,
0.0062290546,
-0.0462660976,
0.0149207013,
-0.0281387754,
0.0074140863,
-0.0968100354,
0.0458852239,
0.0089616375,
0.033241339,
0.0110570155,
-0.0359226093,
-0.0151696503,
0.0249696486,
-0.0451224037,
0.0337141342,
-0.0126696434,
0.0079262359,
-0.0141827762,
-0.0018100083,
-0.0129735265,
0.0423730947,
-0.0629854798,
0.034827888,
0.0293621123,
0.0862393379,
0.0232110098,
-0.0172393192,
-0.0348521769,
0.0409351625,
0.0405576527,
-0.0695678741,
-0.009023021,
0.0592832938,
0.0391144827,
0.0334812179,
-0.0030587462,
-0.0492166579,
-0.0160770379,
-0.0214999672,
0.0156452581,
0.0634976923,
-0.0136191761,
0.0676617101,
0.0162259471,
-0.0263817366,
-0.018102644,
0.0168361124,
0.0030025535,
0.0107606063,
0.0009905987,
-0.0188856516,
-0.0434917286,
-0.0300278813,
0.007097275,
0.0058004465,
-0.0077187004,
0.0549525619,
-0.0002790462,
-0.0299867932,
-0.0130106537,
0.0156610906,
-0.0395233408,
-0.0329355784,
0.0469899364,
-0.0104009723,
0.0312563963,
-0.0281139854,
-0.0041739014,
-0.0437335633,
-0.041088488,
0.0020813656,
0.0151190767,
-6.204176164e-33,
-0.0396347158,
-0.0722210333,
0.0224385578,
-0.0281960424,
-0.0230709445,
0.0277897231,
-0.0305674598,
0.0419005454,
-0.0057552382,
0.0182090886,
0.0411793254,
-0.0513683073,
0.0203113593,
0.0121358605,
0.0159446634,
0.013153756,
0.0129849231,
0.0101197511,
-0.0129294628,
-0.0453615859,
0.0141069712,
0.0192004591,
0.00657578,
-0.0755953118,
0.0293817315,
0.0360444747,
-0.0169509389,
0.0258762483,
0.0015412134,
0.0256460104,
-0.0735223144,
0.0311289597,
0.0259802844,
0.0085146837,
-0.0176624432,
0.0753518865,
-0.0443369299,
0.0022574866,
0.000235625,
0.042144794,
-0.1121713743,
-0.0034157,
0.0226245075,
0.0144491224,
-0.0267581008,
-0.0081743533,
0.0129888486,
-0.0480217002,
-0.012877469,
0.0481439419,
-0.0442659147,
0.0357873589,
0.017924482,
-0.0520770214,
0.0079362933,
0.0309870448,
-0.0007524277,
0.0380010009,
3.961e-7,
-0.0087941233,
0.0066068084,
-0.0216294806,
0.0393478014,
0.0117635643,
0.0275185574,
0.061365135,
0.0146628935,
-0.0766355172,
-0.0597770624,
-0.0597964637,
0.008382733,
0.0447434001,
0.0398645103,
0.0027639018,
0.0028235272,
0.0522798002,
-0.0108045582,
-0.0065524033,
-0.0417345054,
0.0842362642,
0.03283564,
0.037612699,
-0.0467564352,
0.0326963663,
-0.018447347,
-0.0108457142,
-0.0310428068,
-0.0103506884,
0.0723296031,
0.0277621634,
-0.000813465,
0.006692356,
0.0449078754,
-0.0042518717,
-0.0122382985,
-0.0507264435,
0.0123960078,
0.0423732512,
-0.0086576864,
-0.003545284,
-0.0674220547,
-0.029964136,
0.0471551865,
-0.0224889349,
0.0150630949,
-0.0200169217,
-0.0651526973,
0.0353151262,
-0.0412759036,
-0.0152631085,
-0.0166430846,
-0.0150146764,
0.0235270187,
0.0594305061,
0.0612820834,
-0.0407072082,
-0.0495791361,
0.08646366,
-0.0510424152,
-0.0533040464,
-0.0313313752,
-0.0238084998,
-0.0121851685,
0.0268838666,
-0.0329831801,
0.000806232,
-0.0146999136,
-0.0175998732,
0.0089465734,
-0.023369886,
-0.0013774114,
0.0294158664,
2.982e-7,
0.0225508288,
0.0335961953,
0.0310734883,
0.0144400345,
-0.0043083099,
0.0116304085,
-0.0065381434,
0.0450505316,
-0.0501192808,
0.0039042449,
0.0105172712,
0.0472158529,
0.0403858237,
-0.0218523629,
-0.0362234488,
0.0258395933,
0.0187760014,
-0.0090950448,
-0.0718898326,
0.0148151275,
0.0607918464,
0.0193909518,
-0.0557678752,
-0.0116885742,
0.0264373012,
-0.0820022747,
0.0622124784,
-0.0555056036,
0.0333365053,
-0.0774580836,
0.040186815,
-0.0017751093,
-0.0198236294,
-0.0011188326,
0.016578231,
0.0191326886,
0.0259052515,
0.0560371093,
-0.0528405644,
0.0130773708,
0.0304016657,
0.0498779304,
-0.0049076779,
-0.0224655773,
0.0675658584,
-0.0684155226,
-0.0174878929,
0.0084058177,
0.0026856565,
-0.0507775582,
0.0167406444,
0.0001090104,
-0.0209173728,
-0.0182505902,
0.0080971075,
-0.0280348994,
-0.0336262509,
-0.0675772205,
-0.0012802392,
-0.0149624953,
-0.0513236448,
-0.0230849665,
-0.0477543138,
0.0043315529,
0.0101456931,
-0.0406394675,
-0.0608210303,
2.246731995e-34,
0.021126356,
0.0207161084,
0.0202504322,
0.0211464278,
-0.0007975856,
0.0118804183,
0.006433358,
0.0157007948,
0.0043148417,
-0.0731995925,
-0.000599248
] |
http://arxiv.org/abs/2402.01734v2 | http://arxiv.org/pdf/2402.01734v2 | CFTM: Continuous time fractional topic model | 2024-01-29T08:07:41 | 2024-02-07T03:51:42 | Kei Nakagawa, Kohei Hayashi, Yugo Fujimoto | cs.CL | cs.CL, cs.LG, q-fin.CP, stat.AP | In this paper, we propose the Continuous Time Fractional Topic Model (cFTM),
a new method for dynamic topic modeling. This approach incorporates fractional
Brownian motion~(fBm) to effectively identify positive or negative correlations
in topic and word distribution over time, revealing long-term dependency or
roughness. Our theoretical analysis shows that the cFTM can capture these
long-term dependency or roughness in both topic and word distributions,
mirroring the main characteristics of fBm. Moreover, we prove that the
parameter estimation process for the cFTM is on par with that of LDA,
traditional topic models. To demonstrate the cFTM's property, we conduct
empirical study using economic news articles. The results from these tests
support the model's ability to identify and track long-term dependency or
roughness in topics over time. | CFTM: Continuous time fractional topic model: In this paper, we propose the Continuous Time Fractional Topic Model (cFTM),
a new method for dynamic topic modeling. This approach incorporates fractional
Brownian motion~(fBm) to effectively identify positive or negative correlations
in topic and word distribution over time, revealing long-term dependency or
roughness. Our theoretical analysis shows that the cFTM can capture these
long-term dependency or roughness in both topic and word distributions,
mirroring the main characteristics of fBm. Moreover, we prove that the
parameter estimation process for the cFTM is on par with that of LDA,
traditional topic models. To demonstrate the cFTM's property, we conduct
empirical study using economic news articles. The results from these tests
support the model's ability to identify and track long-term dependency or
roughness in topics over time. | [
0.0020689457,
0.0237770155,
-0.0410055704,
0.0155298188,
-0.0286696628,
0.0136551382,
-0.0399004146,
0.02903172,
-0.0731704012,
-0.0021382333,
0.0240796562,
-0.0607269667,
-0.0225596521,
0.0737624392,
-0.0386858806,
-0.0386250168,
0.0104840929,
0.003695752,
0.0599870607,
0.0012978303,
0.0207857117,
-0.0498472564,
0.0155795095,
-0.0095188292,
0.0055313166,
-0.0074799936,
-0.0030335926,
-0.0496946499,
-0.0103476522,
-0.0208822675,
0.0021558139,
0.0526932701,
0.0439821929,
-0.0324768126,
0.0000017729,
-0.030883655,
-0.0220234487,
-0.0334803723,
0.0573932864,
0.0171267372,
0.0584665239,
0.0056434181,
0.0259308796,
0.0078813005,
-0.0718087703,
0.0596549287,
0.000712491,
0.0987562761,
0.0210706759,
0.0353229195,
-0.022316644,
-0.0025248353,
0.0307188444,
-0.0361865424,
0.0313058719,
-0.0226132646,
0.0166607425,
0.0478711724,
-0.0205493495,
0.0228165425,
-0.0224523004,
0.0341950506,
-0.0302104317,
0.0271452926,
0.0434863679,
0.0240750276,
0.0245600715,
0.0119689628,
0.0022571096,
0.0497276224,
0.0374597721,
-0.0100312568,
0.0547662452,
0.0211447161,
-0.0575304553,
0.0339051746,
-0.0861523598,
-0.0065832837,
0.021477934,
0.0768894404,
0.0290237535,
0.0131929535,
-0.0013495524,
0.0193363093,
0.0224562343,
0.0614419878,
-0.0187212098,
-0.0243966691,
-0.0736492053,
-0.0156908501,
0.0348374918,
-0.0406264141,
-0.0028546976,
0.0126286419,
0.0555651411,
0.0163618382,
0.0322303921,
0.0123570487,
-0.0006975349,
-0.0812957659,
-0.0816977695,
-0.003358132,
-0.0160669331,
0.0459551401,
0.0401827209,
-0.0555315167,
-0.0190999433,
-0.0361939743,
0.0280401856,
0.0174487159,
-0.0313227251,
0.0327141806,
0.0426981002,
0.0697006285,
0.0068348916,
-0.0331841558,
-0.0201808941,
0.0025607736,
-0.0245041139,
-0.0241295639,
-0.0074003446,
0.0051547932,
-0.0155546786,
-0.003131846,
-0.0382807814,
0.0035713254,
-0.0198167562,
0.0437064618,
-0.0181378443,
0.086680688,
0.0084671704,
0.0075787143,
0.0213133562,
0.0112780025,
-0.0184043981,
0.043928273,
-0.0186301339,
-0.0265230518,
-0.0512410179,
0.0042332932,
0.0036577315,
0.0070245373,
-0.0213170014,
-0.0595000423,
-0.0329494625,
0.0318326429,
0.0005399621,
0.0383028686,
-0.0155133903,
0.0033489405,
-0.058937557,
0.0243182201,
0.0002634834,
-0.0147312153,
0.0334577709,
-0.0132755078,
0.0309821963,
0.0788670629,
0.0176788326,
-0.0098166578,
0.0455956683,
0.0197590943,
0.0008178607,
-0.0028080544,
-0.0053804331,
0.020349035,
-0.0173548851,
-0.012474807,
-0.0166452993,
-0.0296499934,
0.0403269492,
-0.0033879241,
-0.0058600274,
-0.0067479648,
0.0394451991,
0.026935054,
0.0005652829,
0.022630943,
-0.0179800298,
0.0233148728,
-0.0046029692,
0.029974822,
0.0532489493,
0.013019016,
-0.003042425,
-0.0135153094,
0.0071054436,
0.0112112304,
-0.0399506912,
-0.0259706303,
0.0044105174,
-0.0263080783,
0.0503364094,
-0.0718597546,
0.0886171311,
-0.0724912137,
0.0181638841,
0.0440054014,
0.0409276411,
0.0616307296,
0.0042512552,
-0.0145954108,
0.0820329338,
0.0326876678,
0.030650096,
-0.0337489173,
0.0255242661,
-0.031093359,
0.0380708762,
-0.0163340252,
-0.0406001545,
0.0281106587,
-0.0523323268,
-0.0257415809,
0.0135383569,
0.0158331785,
0.0539667904,
0.0251302905,
-0.0146084195,
0.0463563502,
-0.0076558921,
-0.0293774065,
0.0461016782,
-0.0156799927,
0.0321394168,
0.0353169367,
0.0142030623,
-0.0639702231,
0.0588645935,
0.0260474477,
-0.0006506733,
-0.0111027444,
0.0184022337,
-0.0230491422,
-0.0937612578,
0.0155226467,
-0.0033993367,
0.0458178036,
-0.0191670898,
-0.0181503408,
0.0324292816,
-0.026576452,
0.0129401358,
-0.0284397546,
0.0705493093,
0.0016718139,
0.0141921286,
-0.0434023589,
-0.0269887112,
-0.0028337487,
0.0284054987,
0.0099441353,
0.0351792015,
-0.0246936176,
0.0349766165,
-0.0029421563,
-0.0072077094,
-0.0016577872,
-0.0414561778,
0.0379455239,
0.0374406688,
-0.0302211698,
0.0127038201,
0.0128819598,
-0.0107438592,
0.0376167186,
-0.0143074887,
-0.0487098917,
-0.0196006279,
-0.0177296158,
-0.022974534,
0.0233347118,
-0.0160511341,
-0.0149535043,
0.028420696,
-0.0001862865,
0.0156007223,
0.002135114,
-0.0150885899,
0.0284164865,
-0.0253274925,
0.0071904268,
-0.0506306253,
-0.0035502689,
-0.0449105576,
0.0101491883,
0.0070925858,
-0.0027517832,
0.0068870657,
-0.0062835836,
-0.0191227179,
0.001527169,
-0.0246733427,
-0.00102456,
-0.0011302949,
-0.0221688282,
0.0985623524,
0.0096349819,
0.0353246294,
0.0396768525,
-0.0004380142,
-0.0700746775,
-0.0200063605,
-0.0119328229,
-0.0489687659,
0.1118599251,
0.0019637924,
-0.043131575,
0.0214795265,
0.0802048743,
0.0292450115,
0.017624069,
-0.021499373,
0.0365692601,
-0.0467752554,
0.0572613254,
-0.0270053279,
0.0219385363,
-0.0289250053,
0.0232968088,
0.0159950852,
-0.0276418198,
0.0139282718,
0.003112115,
0.0556800663,
-0.0154325254,
-0.0593093932,
-0.0544241443,
-0.014070834,
-0.030437341,
-0.0173576195,
-0.0385444388,
-0.0157178305,
0.0135963662,
-0.0189467333,
0.0167573541,
0.0029334773,
0.0338503048,
-0.0223765187,
0.0510652997,
0.041589465,
0.0314078443,
0.013288442,
-0.0908721313,
0.0174570475,
0.0164045226,
0.028271487,
-0.0716814324,
0.0198324621,
0.0475467816,
0.0349173546,
-0.0038726514,
-0.0213576332,
-0.0001573111,
0.0003091172,
0.086891301,
0.0681257546,
-0.0416407809,
-0.0191647038,
0.0126299541,
0.0206389148,
-0.0051997127,
0.0696094036,
-0.0121159349,
-0.0261872634,
-0.0126333917,
-0.0455783755,
-0.008315959,
-0.0192635767,
0.0017103999,
0.0169578381,
0.0127291214,
0.0076913713,
0.0887850448,
0.0142351352,
0.0032640896,
-0.0104257502,
0.0110175386,
-0.0163449533,
-0.0092406943,
-0.0295251925,
-0.030699145,
0.0049840375,
-0.0073494036,
-0.0002228587,
-0.0024872543,
-0.0424986519,
-0.0104445592,
0.0088471072,
0.0075351833,
-0.0230672769,
0.0156959388,
0.0127834901,
-0.0955863893,
0.020124441,
-0.0351415314,
-0.048200883,
-0.0108386884,
0.0019501974,
0.0171882249,
0.0274917856,
-0.0991460085,
-0.0423301794,
0.0277437083,
-0.0103110643,
0.0280482527,
-0.033844512,
0.0213275962,
-0.0522387363,
0.0196322165,
-0.0423301868,
-0.0211483948,
-0.0368083268,
-0.0252371412,
0.0101133194,
0.0017659991,
-0.0070357905,
-0.006995908,
0.0243898127,
-0.0095204702,
0.0154092787,
-0.020879522,
0.0433232002,
-0.0412542708,
-0.0424655527,
-0.051509764,
-0.0062430613,
0.002403646,
0.0593066253,
-0.0125380959,
-0.0379915647,
0.0002697797,
0.0090240017,
-0.0044168006,
0.0274296459,
0.0156906471,
-0.0822852924,
-0.0544004589,
-0.0609907843,
-0.0838363916,
-0.0291756578,
-0.0572466031,
0.0458493456,
-0.0128058661,
0.0427275375,
0.0802399293,
-0.0365577973,
-0.0100288512,
-0.0138894999,
-0.0150901861,
-0.0300605092,
-0.0183289219,
-0.0224299934,
0.0845486373,
0.0319672599,
-0.0346826091,
0.0465052463,
-0.1246329769,
-0.0461500324,
0.0215026755,
-0.0370877758,
0.0210019797,
-0.0266213864,
-0.0283102021,
0.0190126207,
-0.0165419672,
0.0180535764,
-0.0332052,
0.0173329376,
-0.0187012684,
0.023903396,
0.0199941043,
-0.0020842601,
0.0144969299,
0.041405078,
-0.0310653262,
0.0055094822,
-0.0941164047,
0.0150048835,
-0.0465415046,
-0.0415701456,
0.0401964001,
0.0411730893,
-0.0336056575,
-0.0286427047,
-0.0358585268,
-0.0248744916,
0.0413638093,
-0.0095335795,
-0.0349735022,
0.0063758492,
-0.0051366119,
-0.0683171973,
0.0017917226,
-0.0227935109,
-0.011873438,
0.0108823488,
-0.0202367064,
-0.0554557592,
-0.007098645,
-0.0002225338,
-0.0126858708,
0.0323027596,
-0.001933257,
-0.0310522709,
0.0118097411,
0.0265863221,
0.0591223463,
-0.0119620096,
-0.0332910083,
0.0102833286,
-0.0433411412,
-0.0600913353,
0.0213074256,
-0.0108860498,
0.0540398099,
0.0074303071,
-0.0384078585,
0.0279757697,
-0.0317545235,
0.0313500948,
0.0096126562,
0.0231409669,
-0.0213253815,
0.0233944152,
0.0372343101,
-0.0023032862,
-0.0645155013,
0.0710817352,
0.0553503074,
0.0247958824,
-0.0562739447,
-0.0188638791,
-0.0390396006,
-0.0181856193,
0.0579091795,
0.097215414,
0.0010520763,
0.0111295469,
0.0343024954,
-0.0560345687,
0.0213530529,
0.0776242465,
-0.0653967485,
0.006600888,
-0.0362029448,
0.0337775946,
0.1028949991,
-0.0138543556,
0.0552318655,
-0.0391406082,
0.0300785694,
-0.004063216,
-0.0343178213,
-0.0480718911,
-6.497428209e-33,
0.0059872442,
-0.0254912656,
-0.0269578695,
0.0255565587,
0.0174072292,
0.0242416598,
-0.0152133266,
0.0179571696,
0.0192833468,
-0.0202859193,
0.0113905137,
0.0426133946,
0.0070863292,
-0.0127492649,
0.0336712301,
-0.0152370892,
0.0046095164,
-0.0332747996,
0.050888706,
0.0223141648,
0.0473934263,
0.0316023007,
0.0154962149,
-0.044724267,
0.0230613071,
-0.0096758408,
-0.0129228728,
-0.0228066035,
0.106007047,
-0.0055675916,
-0.040164154,
0.0095160054,
0.0433834381,
-0.0711043924,
-0.0088835685,
0.0006714602,
-0.0032816031,
-0.0506517738,
0.0046979981,
-0.0246228129,
0.0821060017,
0.0239039864,
0.0425687395,
0.0343203954,
-0.0332712084,
0.0356870629,
0.0733359978,
0.0085836714,
-0.012129508,
-0.0601042397,
-0.026623331,
0.0565671511,
-0.0203504097,
0.0334973335,
-0.038957838,
-0.0323028192,
0.0107155992,
0.0314857736,
-0.0737097561,
0.0123893265,
0.0576954186,
0.0543741286,
0.0003286781,
-0.0597565509,
0.0127842119,
0.025270056,
-0.0439223982,
0.0146064013,
0.0006383656,
0.0536476858,
0.0505262092,
0.013862838,
0.0131977629,
0.0217906535,
0.023673851,
-0.013374608,
-0.0372406431,
0.0533809178,
0.0499591529,
-0.0010190343,
0.0444348641,
0.018659655,
-0.0320679396,
-0.0166671965,
0.0155922743,
0.0489150919,
0.0067616659,
-0.0586088039,
0.0028383248,
0.0037071821,
0.0173914321,
0.0100649055,
0.0247562621,
-0.0218267273,
-0.0697284266,
-0.0085800607,
-0.0513684861,
0.0452182516,
0.0087490352,
0.0018849042,
-0.0724933669,
-0.0671502203,
-0.0050612832,
0.0380264334,
-0.0398012176,
-0.0387813374,
-0.0297312941,
-0.0131460838,
-0.0489156954,
-0.044943288,
0.0399267823,
-0.0272119064,
-0.01312418,
0.0101597346,
-0.009368686,
-0.0186666567,
-0.0009156563,
0.0577600673,
0.044181861,
0.0352846421,
-0.0038079498,
-0.0470703766,
-0.0275730398,
0.0382538512,
-0.0143695856,
0.0007042178,
-0.0015448824,
0.0123621728,
-0.0280089267,
-0.0369548835,
-0.001672301,
-0.0102482243,
2.618e-7,
0.0010936366,
0.0716957897,
-0.030030828,
-0.0548446737,
0.0372881778,
0.0154941771,
-0.0366166309,
-0.0007622411,
-0.0147846015,
-0.0800087973,
0.0131220976,
0.0296131112,
-0.0290943105,
0.031420581,
-0.069301419,
0.0092062475,
-0.0534144267,
-0.033609245,
-0.0411699899,
-0.0156278946,
-0.0164269805,
0.0416093469,
0.0416548438,
0.0255314019,
-0.0181737728,
-0.0193397403,
0.0604232475,
-0.0192074832,
-0.0924798101,
0.0232696738,
0.0479340367,
0.0206702631,
-0.0025945751,
-0.0614381693,
-0.0216754954,
0.0347954147,
0.0058279787,
-0.0025858637,
0.000849285,
0.0000106101,
0.0206016004,
-0.0219543763,
-0.0212196521,
-0.0227591749,
0.0543773063,
-0.0345875695,
-0.0385082476,
-0.0423760079,
-0.0058732112,
0.0587016754,
0.0109255426,
0.0122959102,
-0.0269241408,
0.0033638724,
-0.0381074883,
-0.0056329235,
0.0007476773,
-0.0179478619,
-0.0005944888,
0.0425667614,
-0.0538973808,
-0.0300739016,
-0.04153613,
-0.0112428693,
0.0155298142,
-0.0195086822,
0.0299495999,
1.879296109e-34,
0.0830605179,
0.0104927002,
-0.0692053512,
0.0506965704,
0.0080614621,
0.0266356412,
-0.0940135643,
0.0257026199,
-0.0042176922,
-0.0483111218,
0.0312372055
] |
http://arxiv.org/abs/2401.15927v1 | http://arxiv.org/pdf/2401.15927v1 | E-EVAL: A Comprehensive Chinese K-12 Education Evaluation Benchmark for Large Language Models | 2024-01-29T07:34:37 | 2024-01-29T07:34:37 | Jinchang Hou, Chang Ao, Haihong Wu, Xiangtao Kong, Zhigang Zheng, Daijia Tang, Chengming Li, Xiping Hu, Ruifeng Xu, Shiwen Ni, Min Yang | cs.CL | cs.CL | With the accelerating development of Large Language Models (LLMs), many LLMs
are beginning to be used in the Chinese K-12 education domain. The integration
of LLMs and education is getting closer and closer, however, there is currently
no benchmark for evaluating LLMs that focuses on the Chinese K-12 education
domain. Therefore, there is an urgent need for a comprehensive natural language
processing benchmark to accurately assess the capabilities of various LLMs in
the Chinese K-12 education domain. To address this, we introduce the E-EVAL,
the first comprehensive evaluation benchmark specifically designed for the
Chinese K-12 education field. The E-EVAL consists of 4,351 multiple-choice
questions at the primary, middle, and high school levels across a wide range of
subjects, including Chinese, English, Politics, History, Ethics, Physics,
Chemistry, Mathematics, and Geography. We conducted a comprehensive evaluation
of E-EVAL on advanced LLMs, including both English-dominant and
Chinese-dominant models. Findings show that Chinese-dominant models perform
well compared to English-dominant models, with many scoring even above the GPT
4.0. However, almost all models perform poorly in complex subjects such as
mathematics. We also found that most Chinese-dominant LLMs did not achieve
higher scores at the primary school level compared to the middle school level.
We observe that the mastery of higher-order knowledge by the model does not
necessarily imply the mastery of lower-order knowledge as well. Additionally,
the experimental results indicate that the Chain of Thought (CoT) technique is
effective only for the challenging science subjects, while Few-shot prompting
is more beneficial for liberal arts subjects. With E-EVAL, we aim to analyze
the strengths and limitations of LLMs in educational applications, and to
contribute to the progress and development of Chinese K-12 education and LLMs. | E-EVAL: A Comprehensive Chinese K-12 Education Evaluation Benchmark for Large Language Models: With the accelerating development of Large Language Models (LLMs), many LLMs
are beginning to be used in the Chinese K-12 education domain. The integration
of LLMs and education is getting closer and closer, however, there is currently
no benchmark for evaluating LLMs that focuses on the Chinese K-12 education
domain. Therefore, there is an urgent need for a comprehensive natural language
processing benchmark to accurately assess the capabilities of various LLMs in
the Chinese K-12 education domain. To address this, we introduce the E-EVAL,
the first comprehensive evaluation benchmark specifically designed for the
Chinese K-12 education field. The E-EVAL consists of 4,351 multiple-choice
questions at the primary, middle, and high school levels across a wide range of
subjects, including Chinese, English, Politics, History, Ethics, Physics,
Chemistry, Mathematics, and Geography. We conducted a comprehensive evaluation
of E-EVAL on advanced LLMs, including both English-dominant and
Chinese-dominant models. Findings show that Chinese-dominant models perform
well compared to English-dominant models, with many scoring even above the GPT
4.0. However, almost all models perform poorly in complex subjects such as
mathematics. We also found that most Chinese-dominant LLMs did not achieve
higher scores at the primary school level compared to the middle school level.
We observe that the mastery of higher-order knowledge by the model does not
necessarily imply the mastery of lower-order knowledge as well. Additionally,
the experimental results indicate that the Chain of Thought (CoT) technique is
effective only for the challenging science subjects, while Few-shot prompting
is more beneficial for liberal arts subjects. With E-EVAL, we aim to analyze
the strengths and limitations of LLMs in educational applications, and to
contribute to the progress and development of Chinese K-12 education and LLMs. | [
0.0322390832,
0.0227674805,
-0.0182652865,
0.0277346503,
-0.0381275676,
0.0271531418,
0.0264373478,
-0.0295459647,
0.028797403,
-0.0724353045,
-0.0056629223,
-0.032637354,
-0.0089261485,
0.002129768,
-0.0098107969,
-0.0615170524,
0.0240515824,
0.0123518026,
0.0262736585,
0.0226502605,
-0.0172722917,
-0.0215529483,
-0.012214696,
0.0276102051,
0.0466438383,
-0.0506056547,
0.0118511626,
-0.0197474305,
0.0324027613,
-0.0536761023,
0.0083415043,
0.0726758614,
-0.0106357485,
0.0211055707,
0.0000023554,
-0.0747588128,
-0.0237676743,
-0.0120110437,
-0.0147608044,
0.0582562722,
0.0401753709,
0.0396481305,
-0.0201759581,
0.0064534936,
-0.0373362862,
-0.0070031537,
0.0052078525,
0.0267279483,
0.0545472838,
0.0390052833,
0.0211167652,
-0.039999608,
0.0506202728,
0.0194857102,
0.0046890443,
-0.0345217697,
0.0394133441,
0.0290485099,
0.0262985639,
0.0289531983,
-0.0629048496,
0.0182612259,
0.0245006271,
-0.0061037908,
-0.026096426,
0.0217635911,
-0.0311409645,
0.017855864,
-0.0013945561,
0.0694015846,
0.0137795005,
-0.0010808632,
-0.0022212791,
0.0090525411,
-0.0097218975,
0.0422140323,
-0.0616399683,
0.0330516361,
-0.0133997155,
-0.0109987929,
0.0408253223,
-0.0267295856,
-0.0166349653,
0.0313727632,
-0.0211807452,
0.0360518433,
0.0124774873,
-0.0431341268,
0.0036471048,
0.0348848142,
-0.0174547099,
-0.0315772519,
0.005715745,
0.0034726751,
-0.0048794248,
0.0370441899,
0.0192751028,
-0.0268298369,
-0.0285398178,
-0.0204561669,
0.021341918,
0.0554804429,
0.0227919687,
0.027671326,
0.002229552,
0.086604394,
0.0207014624,
0.0242431276,
-0.0242131334,
0.0262240525,
-0.0682053715,
-0.0093218172,
-0.0144253513,
0.054832764,
-0.0148258144,
-0.0232396927,
-0.0550418906,
-0.0308173895,
-0.0221905485,
-0.005930746,
-0.0117684081,
-0.0262501463,
0.0304971244,
-0.0050313929,
-0.0649971664,
0.0768420547,
-0.0205209777,
0.0024866923,
0.0026540556,
-0.0194099117,
0.0168641414,
-0.0016783174,
-0.0131527595,
-0.0155846132,
0.0036227936,
0.0125869168,
-0.005387329,
-0.0202659778,
-0.0192263797,
-0.0083336867,
-0.0113844713,
-0.0615870878,
0.0247477815,
-0.0852844715,
-0.0238948967,
0.0135974754,
0.0395532586,
-0.0038266578,
0.0208942164,
0.0164345987,
0.0152224367,
0.0048907776,
0.0401589088,
-0.0075705801,
-0.0073079774,
0.0181395542,
0.0166366212,
0.0548737794,
0.0340883099,
-0.0330401175,
0.044983577,
-0.0358767696,
-0.0243001506,
0.0261033811,
-0.0357754789,
-0.0218378026,
-0.0191894025,
0.0114433784,
-0.0118932966,
0.0201664623,
-0.0405288748,
0.0402291976,
0.0012934933,
0.0378452539,
0.0378984585,
0.0723477304,
-0.0021196974,
0.0181300882,
0.0274463072,
0.0347785205,
0.0671489686,
0.0355102718,
-0.040524438,
0.0107942987,
-0.0555168428,
0.019847,
0.0147446338,
0.0265478473,
-0.0040748795,
-0.0890716612,
-0.0232483875,
0.0459788144,
0.0198458042,
0.02450376,
0.0571484417,
-0.0377905816,
-0.0257685687,
-0.0003776343,
-0.0015228613,
-0.0333425254,
0.0270894151,
-0.0255458541,
0.0238108784,
0.0750460476,
0.0459018685,
0.0021233046,
-0.0263183415,
0.0115817897,
0.0202108007,
0.0763666332,
0.0473628044,
-0.017885074,
-0.0407845862,
-0.0175942592,
-0.0243874956,
0.0054633948,
0.0254280381,
-0.029824594,
0.0153303565,
0.017826179,
0.0017596582,
-0.0128850574,
-0.0289274585,
0.0147138815,
-0.0670140758,
-0.0496939979,
-0.054017961,
0.0054212105,
0.025239436,
0.0428866819,
-0.0276811607,
0.0280740038,
0.008927769,
-0.0248412564,
-0.0453010276,
-0.0218598936,
-0.0488175601,
-0.0404902473,
0.0534313805,
-0.0659470409,
-0.0287380759,
0.0562077314,
-0.0411280952,
-0.0417745896,
0.0290473867,
0.0010804115,
0.0106237847,
-0.0185318347,
0.0506247804,
-0.0652448088,
-0.0252018068,
-0.0240849238,
0.0934399813,
0.0033563019,
0.0757591948,
0.0202105623,
0.0288509522,
0.0155089563,
0.0228381548,
-0.065933235,
-0.0024209109,
0.0052757594,
0.0274949074,
0.0307160113,
0.0230123047,
-0.0821500346,
-0.0250356328,
-0.050674919,
0.0122695388,
-0.0000624844,
0.0284801796,
0.0301988311,
-0.0554937944,
-0.0134556368,
0.0506124496,
-0.0198995341,
0.0002705314,
-0.0180585869,
0.0071314098,
-0.0022744099,
-0.0304572228,
-0.0179598015,
-0.0143910218,
-0.0170446113,
-0.0722876042,
0.0223600995,
0.0217209645,
-0.0055593583,
0.0176042337,
0.0374089368,
-0.0283251721,
-0.0095959641,
0.0105943484,
-0.034671586,
0.0371760167,
0.0000960699,
0.0541889817,
0.0676260591,
0.012245262,
-0.0020260506,
0.0384226702,
-0.0450196229,
-0.045829121,
0.0073635783,
0.0183042455,
0.0018421477,
-0.026688097,
-0.0251845792,
-0.0361654609,
0.1136382967,
0.031780377,
-0.0338733867,
-0.0123898378,
0.0322150923,
-0.0709294677,
0.014605375,
-0.0692261234,
-0.00023871,
-0.0362594426,
0.0154445805,
-0.0301280376,
0.0282791927,
0.0910750851,
0.0079644546,
0.0088058496,
-0.0869574696,
0.0469422415,
-0.0065510222,
0.0485510007,
-0.0092304479,
-0.0289596841,
0.0130042555,
-0.0012545087,
-0.0594594292,
0.0338220447,
0.025011465,
-0.0552687235,
0.0525588952,
-0.045864094,
-0.0278148539,
-0.0104141161,
0.011555613,
-0.0292239282,
0.0363153405,
-0.0141582564,
0.0290569682,
0.0130717391,
-0.003759142,
-0.0008490899,
-0.0026039206,
0.0336315893,
0.0085426401,
-0.0416788533,
-0.0217744242,
-0.0192860477,
0.0455760956,
0.0220474675,
-0.0240877513,
0.0269421134,
-0.0360420421,
0.1256621927,
0.0770862624,
0.0593569614,
0.0173049178,
0.0359120183,
0.0195875559,
0.0310772154,
-0.0290697236,
-0.0147939799,
0.0450729206,
0.0309381112,
-0.1065493822,
0.0161234066,
0.1301492155,
0.0350798965,
0.0084463097,
-0.0157178361,
-0.0299181733,
-0.01537746,
0.0393388607,
-0.0109051019,
-0.0823752359,
-0.0118492888,
-0.0236196183,
-0.0328593962,
-0.0172265712,
-0.0537850894,
0.0265477709,
-0.1548830122,
-0.0028292395,
0.0041708085,
-0.0460570529,
0.0194624513,
-0.0327006653,
-0.0078487387,
0.0225121174,
0.0082424795,
0.0075680539,
-0.0208145771,
0.0074625663,
0.0295793954,
-0.0167225618,
-0.0390091203,
0.0809210166,
-0.0255929977,
-0.0006867552,
-0.0062148031,
0.0415331312,
-0.0490039513,
-0.0171716418,
0.0283149388,
-0.0329446904,
-0.0452339351,
0.0227213744,
-0.0057274974,
0.0136624686,
0.0157769751,
-0.0496406592,
0.0066258796,
-0.0072015207,
0.0236961357,
0.0657507554,
0.1069202498,
0.0185968224,
-0.0239587296,
-0.0192340948,
0.0220273286,
-0.0510418788,
0.0935587287,
0.0426949523,
-0.0734517425,
0.0016255643,
-0.0589396879,
-0.0440864451,
-0.0355629772,
-0.0693638772,
0.0637610033,
-0.0035113299,
0.0087454449,
-0.0660422817,
-0.0526918545,
0.039051678,
0.0033232591,
0.0200756546,
0.0407146104,
0.0600730553,
0.0062338738,
-0.0080404477,
0.0261042323,
0.007272651,
-0.0007990359,
-0.0457994416,
0.0137383984,
0.0103313783,
0.0114923064,
-0.020245919,
0.029154636,
-0.0638742074,
-0.0538904779,
-0.0175298732,
-0.0003774685,
0.0101096081,
-0.0374266505,
0.0209414288,
0.0021134466,
0.0023698607,
-0.041826997,
0.0040292414,
-0.0176427923,
-0.0204223841,
0.0005727796,
0.02005261,
-0.022206787,
0.0013291548,
0.0089893294,
0.0235960763,
0.0770051926,
-0.0310237948,
-0.0200309604,
0.0345817618,
0.047484424,
0.0434306487,
-0.0054147514,
-0.0301709492,
-0.0408145897,
0.0007303959,
-0.0723558441,
-0.0175506789,
-0.0249272864,
-0.0247717798,
0.0538746566,
-0.005832823,
-0.0085433098,
0.0258461144,
0.0008200825,
0.0296747033,
-0.0386250801,
0.0504673459,
-0.0321380384,
0.0439140089,
0.0001225594,
-0.0079499455,
-0.0254540667,
-0.0300560687,
0.0078816405,
0.0642537549,
0.0011870174,
0.0890919268,
-0.0066615599,
0.0085485727,
0.0516872853,
-0.0117913745,
-0.0324599817,
-0.0444811285,
-0.0103571992,
0.0198454782,
0.0309935343,
0.0120816575,
0.0303139519,
-0.0414760448,
-0.0054683192,
-0.0488820747,
-0.009887429,
0.0138671612,
-0.0519328453,
0.0420728959,
0.0288792998,
0.0254416447,
-0.0196110085,
0.0393122174,
0.0021238909,
0.0081511559,
0.033363577,
-0.0175444093,
-0.0240404177,
-0.0253486149,
-0.0178413,
0.0290709715,
-0.003547177,
0.0319569074,
-0.1073050722,
-0.0600218847,
-0.0244559273,
0.0180324949,
-0.0008437168,
-0.025220735,
0.0216735825,
0.0636954159,
-0.0410762541,
-0.0231027715,
-0.0284331013,
0.0122028077,
-0.0163107477,
-0.0335379988,
0.0069269859,
-6.66341242e-33,
-0.0198919475,
-0.0381569825,
0.045811642,
-0.0019628359,
0.0210559145,
-0.0049510719,
-0.0444993824,
0.0535471253,
-0.0533044003,
0.0091926185,
0.0078158956,
-0.0118123367,
-0.0053809769,
0.0296812225,
0.0009827671,
0.0119851464,
-0.0141568407,
-0.007844572,
-0.0140645988,
-0.0251621176,
0.0510980524,
0.0062208939,
0.0823523179,
-0.0182164982,
0.0423233695,
0.0032922786,
0.036540702,
0.0030539297,
-0.0479582138,
0.0399471968,
-0.0066316412,
0.0389140286,
0.0365397595,
-0.0821638703,
-0.0105893724,
-0.0020715625,
-0.0637396201,
-0.0666574538,
0.0370204002,
0.0042584958,
-0.0126172053,
-0.0149181336,
0.0141384983,
0.0225599483,
-0.0307374988,
-0.039166037,
0.0644631907,
-0.0523364358,
-0.0036361376,
-0.010996257,
-0.065793395,
0.0168819353,
-0.0303122196,
0.0134111494,
-0.0219537262,
0.0722873136,
-0.0242093112,
0.0211391728,
-0.0415809788,
0.0271866303,
0.0519729331,
0.0486607924,
0.0550519079,
-0.0274837743,
0.0348275416,
0.0616710074,
0.0102852453,
-0.0138464496,
-0.0228696913,
-0.0023224242,
0.0654868633,
0.0102240648,
0.0698541254,
0.047939878,
0.0181769896,
0.0185500868,
-0.0991580337,
0.0263066515,
0.0249184873,
0.0233068112,
-0.0083196368,
-0.0497037992,
-0.057786122,
-0.0173375458,
-0.0573398136,
0.0162446909,
-0.0301552024,
-0.0175755881,
0.0485796072,
-0.0495889448,
-0.0248924494,
-0.0326638073,
0.0313378349,
-0.047277078,
-0.0273471344,
0.0499539077,
-0.0069297766,
0.0333583318,
-0.0194323752,
0.0368883535,
-0.0828611106,
-0.0011495883,
-0.0241456907,
-0.0168806538,
0.0088803554,
0.0015648914,
-0.0004197157,
-0.012927575,
-0.0349820629,
-0.0224275142,
-0.0445945896,
-0.039851293,
0.0490571186,
-0.0043283813,
0.02124287,
0.0092113521,
-0.0032614181,
0.0422166325,
0.0276953746,
-0.033408694,
-0.0220573712,
-0.0087401718,
0.0006398703,
-0.0071741426,
0.0074355016,
-0.0088205552,
-0.0340075418,
0.0131134512,
0.028695913,
-0.0507169776,
-0.0082422951,
0.0131019615,
2.902e-7,
0.0342384391,
0.0412416086,
0.0028861964,
0.0110160913,
-0.0104751354,
0.0230229348,
-0.0367429405,
-0.0196486358,
0.0389920734,
-0.0059074867,
-0.0105955685,
0.0200059991,
0.0748294592,
0.0098942099,
-0.0684176609,
-0.0315164439,
-0.0134255504,
0.0033927909,
-0.0458991192,
0.0298434459,
0.0541327335,
0.0071372874,
0.0056672678,
0.0024933477,
-0.059643086,
-0.017055545,
0.0269425195,
-0.0073035858,
-0.0243695211,
0.0317723565,
-0.016613096,
0.0008547241,
0.0048659672,
0.0314804986,
-0.0154527863,
0.033628922,
0.0276784748,
0.0730206594,
0.0356612504,
0.0191368368,
0.0081180921,
-0.0385744572,
-0.0137935961,
-0.0626578629,
0.0575536415,
-0.0398923047,
-0.0123689361,
0.0628729165,
-0.0320541374,
-0.0157780368,
-0.0094205467,
-0.0059554977,
-0.0009421134,
0.011682949,
-0.0072542611,
-0.0213581566,
0.0265331771,
-0.0075849285,
-0.0330568589,
0.0028882998,
-0.0138756959,
-0.0758004338,
-0.035247758,
-0.0381652899,
0.0126763331,
-0.0273740888,
-0.0262635313,
2.533815496e-34,
0.0223542415,
-0.0439835005,
0.0465195887,
-0.011805105,
0.0379681066,
-0.0037971241,
-0.0339672603,
-0.0184997134,
-0.00615391,
-0.0849319771,
-0.0100113275
] |
http://arxiv.org/abs/2401.16454v1 | http://arxiv.org/pdf/2401.16454v1 | KAUCUS: Knowledge Augmented User Simulators for Training Language Model Assistants | 2024-01-29T06:57:02 | 2024-01-29T06:57:02 | Kaustubh D. Dhole | cs.HC | cs.HC, cs.AI, cs.CL, cs.IR, I.2.7; H.3.3 | An effective multi-turn instruction-following assistant can be developed by
creating a simulator that can generate useful interaction data. Apart from
relying on its intrinsic weights, an ideal user simulator should also be able
to bootstrap external knowledge rapidly in its raw form to simulate the
multifarious diversity of text available over the internet. Previous user
simulators generally lacked diversity, were mostly closed domain, and
necessitated rigid schema making them inefficient to rapidly scale to
incorporate external knowledge. In this regard, we introduce, Kaucus, a
Knowledge-Augmented User Simulator framework, to outline a process of creating
diverse user simulators, that can seamlessly exploit external knowledge as well
as benefit downstream assistant model training. Through two GPT-J based
simulators viz., a Retrieval Augmented Simulator and a Summary Controlled
Simulator we generate diverse simulator-assistant interactions. Through reward
and preference model-based evaluations, we find that these interactions serve
as useful training data and create more helpful downstream assistants. We also
find that incorporating knowledge through retrieval augmentation or summary
control helps create better assistants. | KAUCUS: Knowledge Augmented User Simulators for Training Language Model Assistants: An effective multi-turn instruction-following assistant can be developed by
creating a simulator that can generate useful interaction data. Apart from
relying on its intrinsic weights, an ideal user simulator should also be able
to bootstrap external knowledge rapidly in its raw form to simulate the
multifarious diversity of text available over the internet. Previous user
simulators generally lacked diversity, were mostly closed domain, and
necessitated rigid schema making them inefficient to rapidly scale to
incorporate external knowledge. In this regard, we introduce, Kaucus, a
Knowledge-Augmented User Simulator framework, to outline a process of creating
diverse user simulators, that can seamlessly exploit external knowledge as well
as benefit downstream assistant model training. Through two GPT-J based
simulators viz., a Retrieval Augmented Simulator and a Summary Controlled
Simulator we generate diverse simulator-assistant interactions. Through reward
and preference model-based evaluations, we find that these interactions serve
as useful training data and create more helpful downstream assistants. We also
find that incorporating knowledge through retrieval augmentation or summary
control helps create better assistants. | [
0.0733994022,
0.0136380866,
-0.0449123308,
0.0452195555,
-0.0166548155,
-0.0073214262,
-0.0243451539,
-0.0203913804,
0.0243828446,
-0.0215268061,
-0.0775065348,
-0.0456619523,
-0.0388728157,
-0.0097826915,
0.0090500154,
-0.0967761949,
0.061111521,
-0.0462621376,
-0.0273805838,
0.0109878574,
0.0188483763,
0.0266420525,
-0.0263348632,
0.06459409,
0.0025395351,
0.0067148446,
-0.0365462601,
-0.0033701963,
-0.0076668039,
-0.0246744603,
0.0322195739,
0.0389731489,
0.0462986566,
0.0290566981,
0.0000022058,
-0.0185143873,
0.0077022486,
0.002214656,
-0.0386815108,
0.0056219199,
0.0874298289,
0.0493986383,
-0.0083506936,
0.001155919,
-0.0418279544,
0.009244265,
0.0111803664,
0.0175275691,
0.0756805688,
0.0383054204,
-0.0057099597,
-0.0033623737,
-0.0170637164,
0.0012973759,
0.013141443,
-0.0248730294,
-0.0036166159,
0.0092255743,
0.0867142528,
0.0178957935,
0.0235199444,
0.026550658,
0.0450253561,
-0.0028632751,
-0.0001639133,
0.0455023125,
-0.0177190565,
-0.0145160072,
-0.0346848704,
0.0474763885,
0.0095809465,
-0.0028165779,
-0.0052992958,
0.0111809354,
0.0237929914,
0.0174010396,
-0.0176728461,
-0.0262212027,
-0.032274656,
-0.0243007373,
0.0136744119,
0.0113922348,
0.0052331318,
0.017461,
-0.025170248,
0.0337329842,
0.035934858,
-0.0725229755,
0.0053903912,
0.0045412001,
0.001130297,
-0.0266309995,
0.0138507048,
0.0633884743,
-0.05717678,
0.0540293232,
-0.0302888732,
-0.0165027827,
-0.0060452321,
-0.0441648327,
0.0742503405,
0.0600511357,
-0.0358668938,
0.0349178463,
0.0148313809,
0.0374712572,
0.0065300968,
0.0193782933,
-0.0238107387,
0.0122866416,
-0.0551762879,
-0.0227398109,
-0.0419557765,
0.0315808617,
-0.0385561921,
-0.0144563038,
-0.0218601786,
0.0667160079,
-0.0059332382,
-0.0123327319,
0.0223749969,
0.0386920795,
-0.051645305,
0.0103790704,
-0.0744769499,
0.0152639477,
-0.0435454845,
0.0142630609,
-0.001990946,
-0.0620510504,
0.0106103579,
-0.0055119544,
-0.0260006376,
-0.0323814489,
-0.0083836308,
0.0725699216,
-0.0134001048,
-0.0401743315,
-0.0160888657,
-0.0023660031,
0.0187905282,
-0.0434392504,
-0.0064583379,
-0.0441676714,
0.04675382,
-0.0058497661,
-0.0155414175,
-0.0434340611,
-0.0095655769,
0.0452102385,
-0.0111205494,
0.0559513941,
-0.011154783,
0.004874018,
0.0484851226,
0.004097776,
-0.057119634,
0.0591415018,
-0.0065830606,
0.0242707934,
0.0148520963,
0.0071561616,
0.0084360633,
0.0531251132,
0.0225410964,
0.0294484459,
0.0007712609,
0.0257488862,
-0.0239628386,
0.0231165923,
0.0012896,
0.0221471693,
-0.0404645503,
0.0360272937,
0.0676626787,
0.0460062325,
0.0388080627,
0.0123220431,
0.0817327723,
0.005237788,
0.0378238671,
0.0264208578,
-0.0704062283,
0.0371743217,
-0.0196712706,
0.0325331502,
0.0097420234,
0.0589794964,
-0.0250789858,
-0.0272779111,
0.0281954631,
-0.0176694784,
0.0100406343,
0.013810575,
0.0427059717,
0.0196247175,
-0.0328606442,
0.0129055204,
0.0352026895,
-0.040391352,
0.0031512727,
-0.0546874069,
0.0008052706,
0.1124722213,
0.0489625186,
-0.0224918425,
-0.0303976387,
0.0092496332,
-0.0288775954,
0.0408519022,
0.038486246,
-0.0363759324,
-0.0291683525,
-0.0392634161,
-0.0371190533,
-0.0010388425,
0.0155697977,
0.0554968566,
0.0541975275,
0.0003920465,
0.0162799265,
0.0043315012,
0.0009671412,
0.0192578156,
-0.0557985753,
-0.0256508812,
-0.0530593656,
0.0109547451,
0.017379541,
0.0189158581,
0.0046552522,
0.0026520162,
0.0225159694,
0.0254089255,
-0.0033521336,
0.0329055265,
-0.0461377352,
0.0000884436,
0.0393215865,
0.0217682384,
0.0074249827,
0.0403714031,
0.0186439641,
-0.0346494056,
0.034417171,
-0.0070166057,
0.0759379417,
-0.0560785457,
0.0096854325,
0.0129455542,
-0.0272007026,
-0.0193297826,
0.0732692555,
0.0076956125,
0.0099039264,
0.0710000768,
-0.0091308998,
-0.0069814799,
-0.0285260938,
-0.0213236921,
-0.032388892,
-0.0014144296,
0.0438857675,
0.0186606236,
-0.007207376,
-0.0431556776,
-0.006366482,
-0.1051778272,
0.0091084801,
-0.0382071473,
0.0115634929,
0.0536653548,
0.053586144,
-0.0301894695,
-0.0533117093,
-0.0032682554,
-0.0263281595,
-0.004338732,
0.0216599591,
-0.0074105239,
0.0328462832,
-0.0040131956,
-0.0555212349,
0.0241297595,
-0.0061014853,
-0.0169452578,
-0.0211539548,
0.0538025685,
-0.0444710813,
-0.0262496807,
-0.0176269468,
0.0575283803,
-0.0307537764,
-0.0143666826,
0.0456256978,
-0.0229001474,
0.0364691205,
0.01437084,
0.0197711345,
0.0629159361,
0.044640936,
-0.006796699,
-0.0808906332,
0.0541713946,
-0.007988519,
-0.0506781042,
-0.0513846055,
-0.0150825875,
0.026215218,
0.0846001133,
-0.0160421897,
0.015459761,
-0.0652288571,
-0.0111626452,
-0.0313887186,
-0.0370892063,
-0.036976669,
0.012702344,
-0.0475068651,
-0.0055262381,
-0.0278683975,
-0.0233563576,
0.0980752632,
-0.0137707284,
-0.0233607981,
-0.0750861242,
-0.0236048829,
-0.0093640294,
0.0359071977,
-0.0082476595,
-0.0022176476,
-0.0133669134,
0.026115505,
-0.047554452,
0.0283388495,
0.0032201339,
-0.0400728099,
0.0399372429,
-0.079742603,
-0.0003104867,
0.005528647,
-0.0147267245,
-0.0518075116,
-0.004801624,
-0.0092641059,
-0.0221365374,
0.0576204993,
-0.010030455,
-0.0192367602,
-0.012598672,
0.0620716736,
0.0611673668,
-0.025434196,
-0.0275301822,
-0.0148413889,
0.0066633723,
0.0837037042,
-0.0030032194,
-0.0444787145,
-0.0446409583,
0.0007107759,
0.0164528824,
0.0339300036,
-0.0760739297,
0.0042249914,
-0.0067782202,
0.0412802584,
-0.0491151325,
-0.0168349184,
0.0610656589,
0.0171312504,
-0.0331903435,
-0.0180989951,
0.1204177588,
0.0487097502,
-0.0724099129,
-0.0043027648,
-0.051189959,
0.0301072728,
0.0434160903,
0.0886329785,
-0.0677100345,
0.004989943,
-0.0343222059,
0.0370789282,
0.0136674717,
-0.0290443022,
-0.0186779629,
-0.1124615818,
0.0075000385,
-0.0098998286,
-0.0724417344,
0.0181585867,
-0.0443175659,
0.0238997564,
0.0025891645,
0.0198613759,
-0.0003908619,
0.0023406558,
0.0043118182,
-0.0053404979,
0.0184948929,
-0.0618519783,
0.0165053867,
-0.0053431406,
0.0307465326,
0.0372576863,
0.0117370468,
0.0175507758,
-0.0200534966,
-0.0417486914,
0.0113214981,
0.0359655768,
-0.0004422879,
0.0463039763,
-0.0269253254,
-0.002473712,
-0.0243187509,
0.0258717705,
-0.0022465503,
-0.0023552012,
0.0153712882,
0.0633527115,
-0.0253862161,
0.0316641033,
-0.0416036919,
-0.0469074249,
0.0766658932,
0.0321861878,
0.0104388008,
-0.0629368275,
0.0023025563,
-0.0377260596,
-0.0402937457,
-0.0406176783,
-0.031258598,
0.0086211488,
-0.0079037622,
-0.0010712465,
-0.0473774634,
-0.0885664225,
0.014038966,
0.0121789146,
0.0083841737,
0.0009015423,
0.0128516089,
-0.0226229168,
-0.0000018882,
-0.005727971,
-0.0364646763,
0.0322225951,
-0.0232412592,
-0.0390870869,
0.0383743346,
-0.071024172,
0.0206881426,
-0.0128340647,
-0.0137997447,
-0.059198834,
0.0099928789,
0.0910242721,
0.0206686296,
0.0815341547,
-0.0194609724,
-0.0162057765,
0.0223414022,
0.0125876348,
-0.111459583,
-0.0023262312,
0.0073643089,
-0.0320034996,
-0.0001983742,
0.0013449383,
-0.0596606247,
0.0258707032,
-0.0510294698,
0.0064393673,
-0.0201909039,
-0.0184325892,
0.0452866629,
0.0070569925,
-0.0022791929,
-0.0179283824,
0.0034933151,
-0.0167372357,
0.0328338705,
-0.0416321568,
-0.0323924683,
-0.0142956153,
-0.0250308681,
0.0150165241,
-0.0101670232,
0.0320251137,
-0.0300417822,
-0.0056673242,
0.0253505372,
-0.033856146,
-0.0187000539,
-0.0000565219,
0.0447728001,
0.0230288245,
0.0110795069,
-0.0105720423,
-0.003430353,
-0.0127298646,
0.0565573089,
-0.0236871224,
0.081888184,
0.0011568717,
-0.0305678416,
0.0118028196,
0.0416160375,
-0.0485252291,
-0.0398536697,
-0.0400553755,
0.0147691211,
0.0333277993,
-0.0225254074,
0.0409772322,
-0.0698515922,
0.0197771359,
-0.0019163741,
-0.0163957048,
-0.0006531333,
-0.0445551649,
0.0332528837,
0.0364290103,
0.0306279249,
-0.0114464127,
0.0731421858,
-0.020443067,
0.0641050488,
-0.0207041763,
0.0027523178,
-0.059522599,
-0.0375056751,
0.0423018262,
0.058766894,
-0.0512347706,
-0.0205501448,
-0.0552229881,
-0.0533410646,
-0.0227734931,
0.0145217022,
-0.0329405554,
-0.0485135466,
0.0627574921,
0.0442726575,
-0.0260107554,
0.0057166717,
-0.036853306,
-0.0306726936,
-0.0424832553,
-0.0048454669,
-0.0221459214,
-6.237915791e-33,
-0.0423248969,
0.0141476514,
0.0790717229,
-0.0374156646,
-0.0024054819,
-0.0225867648,
-0.0281714313,
-0.0240940787,
-0.0275312215,
-0.0240013208,
-0.0107069742,
-0.0002977007,
0.0038135834,
0.0059584719,
0.0173538309,
0.0260493103,
0.0434574783,
-0.0232342575,
-0.0299692005,
0.0230000131,
0.022412207,
0.0184123889,
0.0395078212,
-0.0812502205,
0.0602431633,
-0.0154428212,
0.0152879935,
-0.0039738398,
-0.0465282984,
0.0822582245,
-0.0519564338,
0.0506717972,
-0.0011291227,
-0.0449676141,
-0.0110291839,
-0.0254088137,
-0.0305450205,
-0.0415170491,
0.045642402,
0.0168055911,
-0.029156195,
-0.0148829194,
-0.0139840664,
0.0062360563,
0.0115691805,
0.0196860526,
0.036015451,
0.004553183,
-0.0092230076,
0.0425571389,
-0.0675997213,
0.0015870854,
0.0160053056,
-0.0509510525,
-0.0036541815,
0.0364181586,
0.002547595,
-0.0372844674,
-0.028729422,
0.019244872,
0.0162021015,
0.0100365197,
0.105124034,
0.0157913268,
-0.0059710522,
0.0084046572,
-0.0334494412,
-0.0005058481,
-0.0520052314,
-0.0266694222,
0.0264975075,
0.0407123491,
-0.0051208371,
0.0416427962,
0.0126087544,
-0.0193903428,
0.0449378565,
0.042827148,
0.0109008811,
0.0635930002,
0.0102552921,
0.037636213,
-0.0516933613,
0.0071597206,
-0.0313676856,
0.0215110332,
-0.0202760883,
-0.0964580029,
0.0419444405,
0.0140985511,
-0.0096714431,
-0.0235143527,
-0.0053783399,
-0.0181268658,
-0.0167318508,
-0.000413299,
-0.014974203,
0.0361651443,
-0.0083252536,
-0.0053489618,
-0.06751775,
-0.0290308688,
0.0276877973,
0.0169298835,
-0.0204270817,
0.0494024381,
-0.0020589498,
0.0412732475,
-0.0308712479,
0.0111850677,
-0.0409943983,
0.0331157446,
0.0466319956,
-0.0103046969,
0.0204816647,
0.0040277126,
-0.0415256768,
0.047743082,
0.0098596895,
-0.071524024,
-0.0340329073,
-0.0501079634,
0.0095982272,
-0.0053305863,
-0.0049374634,
-0.0010834232,
-0.0575930513,
0.0136114452,
0.0060381303,
-0.0422578231,
-0.0099969991,
-0.0069604004,
2.926e-7,
0.0134611679,
0.0044578454,
0.0306007396,
0.0236951374,
-0.0056237914,
0.0011564809,
-0.0139223449,
0.0411283709,
0.0137307271,
-0.0178682655,
0.0107202064,
0.0587389953,
0.0239604376,
0.0388307124,
-0.0956530571,
-0.0132061755,
-0.0273047555,
-0.0051207654,
-0.0713924989,
-0.0072853058,
0.0543400757,
0.0350951217,
0.0067045507,
0.0242137369,
-0.0053868052,
-0.0085307537,
0.0422837622,
-0.0147452662,
-0.0066377167,
-0.0038818712,
0.0476722643,
0.0542606562,
-0.0238973014,
0.00471181,
-0.0277908575,
0.0442864262,
-0.0170201417,
0.0426519364,
-0.0314492248,
0.0058908802,
0.0155138662,
-0.0231500231,
-0.0083579803,
-0.0718167052,
0.0335375406,
-0.0418505222,
0.0407342277,
-0.0173544362,
-0.0167076811,
0.0106603606,
-0.0300922804,
-0.0363965966,
-0.0024100591,
0.0166377332,
-0.0216490272,
0.0095948046,
-0.0301078744,
-0.0646524355,
-0.016393818,
0.0350821316,
-0.0129012195,
-0.0254679788,
-0.0935091451,
-0.0090483073,
-0.0240903459,
-0.0595814623,
0.0318804309,
2.391681928e-34,
0.0030679435,
0.0037129705,
0.0445847251,
0.0612845197,
0.0348091982,
0.0013766433,
0.0450800285,
0.0564452484,
0.0408985578,
-0.0382457189,
-0.0223195702
] |
http://arxiv.org/abs/2402.01733v1 | http://arxiv.org/pdf/2402.01733v1 | Development and Testing of Retrieval Augmented Generation in Large Language Models -- A Case Study Report | 2024-01-29T06:49:53 | 2024-01-29T06:49:53 | YuHe Ke, Liyuan Jin, Kabilan Elangovan, Hairil Rizal Abdullah, Nan Liu, Alex Tiong Heng Sia, Chai Rick Soh, Joshua Yi Min Tung, Jasmine Chiat Ling Ong, Daniel Shu Wei Ting | cs.CL | cs.CL, cs.AI | Purpose: Large Language Models (LLMs) hold significant promise for medical
applications. Retrieval Augmented Generation (RAG) emerges as a promising
approach for customizing domain knowledge in LLMs. This case study presents the
development and evaluation of an LLM-RAG pipeline tailored for healthcare,
focusing specifically on preoperative medicine.
Methods: We developed an LLM-RAG model using 35 preoperative guidelines and
tested it against human-generated responses, with a total of 1260 responses
evaluated. The RAG process involved converting clinical documents into text
using Python-based frameworks like LangChain and Llamaindex, and processing
these texts into chunks for embedding and retrieval. Vector storage techniques
and selected embedding models to optimize data retrieval, using Pinecone for
vector storage with a dimensionality of 1536 and cosine similarity for loss
metrics. Human-generated answers, provided by junior doctors, were used as a
comparison.
Results: The LLM-RAG model generated answers within an average of 15-20
seconds, significantly faster than the 10 minutes typically required by humans.
Among the basic LLMs, GPT4.0 exhibited the best accuracy of 80.1%. This
accuracy was further increased to 91.4% when the model was enhanced with RAG.
Compared to the human-generated instructions, which had an accuracy of 86.3%,
the performance of the GPT4.0 RAG model demonstrated non-inferiority (p=0.610).
Conclusions: In this case study, we demonstrated a LLM-RAG model for
healthcare implementation. The pipeline shows the advantages of grounded
knowledge, upgradability, and scalability as important aspects of healthcare
LLM deployment. | Development and Testing of Retrieval Augmented Generation in Large Language Models -- A Case Study Report: Purpose: Large Language Models (LLMs) hold significant promise for medical
applications. Retrieval Augmented Generation (RAG) emerges as a promising
approach for customizing domain knowledge in LLMs. This case study presents the
development and evaluation of an LLM-RAG pipeline tailored for healthcare,
focusing specifically on preoperative medicine.
Methods: We developed an LLM-RAG model using 35 preoperative guidelines and
tested it against human-generated responses, with a total of 1260 responses
evaluated. The RAG process involved converting clinical documents into text
using Python-based frameworks like LangChain and Llamaindex, and processing
these texts into chunks for embedding and retrieval. Vector storage techniques
and selected embedding models to optimize data retrieval, using Pinecone for
vector storage with a dimensionality of 1536 and cosine similarity for loss
metrics. Human-generated answers, provided by junior doctors, were used as a
comparison.
Results: The LLM-RAG model generated answers within an average of 15-20
seconds, significantly faster than the 10 minutes typically required by humans.
Among the basic LLMs, GPT4.0 exhibited the best accuracy of 80.1%. This
accuracy was further increased to 91.4% when the model was enhanced with RAG.
Compared to the human-generated instructions, which had an accuracy of 86.3%,
the performance of the GPT4.0 RAG model demonstrated non-inferiority (p=0.610).
Conclusions: In this case study, we demonstrated a LLM-RAG model for
healthcare implementation. The pipeline shows the advantages of grounded
knowledge, upgradability, and scalability as important aspects of healthcare
LLM deployment. | [
0.0657579228,
0.0277580209,
-0.0071665663,
-0.0293517075,
-0.0127643244,
0.0033094527,
-0.0170179848,
0.0387445427,
0.0048502078,
-0.0437813811,
-0.0019463931,
-0.0444842465,
-0.0110448729,
0.029182842,
0.007898136,
-0.040030051,
0.0423263945,
0.0341065228,
-0.0108136553,
0.0182790272,
-0.0242662281,
0.0167025179,
-0.0192973036,
0.0438661985,
0.0255227834,
0.0081375651,
0.0086823972,
0.0098045999,
-0.011207697,
-0.0712912232,
-0.0044973679,
0.0155258942,
0.0241931863,
0.0224019252,
0.0000023677,
-0.0567335561,
-0.0413256101,
0.0283893123,
-0.0608304255,
-0.0324262828,
0.0623129942,
-0.0017112865,
-0.0021887347,
0.0203226525,
-0.0329062156,
-0.0381770618,
0.0240527671,
0.0222721063,
0.0635088459,
0.0500860065,
-0.0054920991,
-0.0210823361,
0.0354817137,
-0.0025115421,
0.0256939884,
-0.0275966823,
0.0165993869,
0.0393555723,
0.0777459219,
0.0173761826,
0.0398375951,
0.0629533231,
-0.0071784975,
0.0211672615,
0.0342438258,
0.04862931,
-0.0287569743,
-0.0037866293,
-0.0041290168,
0.0195285566,
0.0342837088,
-0.0213143975,
0.02492157,
0.0059143743,
-0.018573733,
0.0072639473,
-0.0785406604,
-0.0016248298,
-0.0170349665,
-0.0034767336,
-0.0202446338,
-0.0295216255,
0.014940084,
-0.0021732505,
-0.0242121425,
0.0255142767,
0.0249470528,
-0.0285608862,
0.0062265755,
-0.0281144101,
0.0177713986,
-0.0062592858,
0.0403235257,
0.0621853806,
-0.0407685861,
0.0256532077,
-0.0046680458,
-0.0355655774,
0.0974702388,
-0.06123982,
0.0392802246,
0.0128505742,
0.0075335735,
0.0159534533,
-0.0470893681,
0.0376383476,
-0.0546957441,
0.006532534,
-0.0685464516,
0.0102847181,
-0.0693674982,
-0.0436628312,
-0.0245194696,
0.057348717,
-0.0137721002,
0.0084857177,
-0.0693687722,
0.0176186636,
0.0136051876,
0.0275185965,
0.0134289358,
0.0430535637,
-0.0298857912,
0.0097479979,
-0.0326085985,
0.0337030813,
0.0155089954,
0.003928632,
0.0115357116,
-0.0369766541,
0.0307722967,
0.0270914976,
0.0163945165,
-0.0034342236,
0.0189402644,
0.0480104275,
-0.01438314,
-0.0131574562,
-0.0086842794,
-0.082925722,
-0.0077078496,
-0.0269489009,
0.0242355149,
-0.080534175,
0.0203141887,
0.0360189341,
-0.0072616339,
-0.057185296,
0.0200735163,
0.0634245723,
-0.0521755666,
0.0161570422,
-0.0411775596,
0.0272407178,
0.06180894,
0.0436283164,
0.0356465168,
0.0115389479,
-0.0072320634,
0.0010175828,
-0.005569614,
-0.0304328986,
0.001075841,
-0.0208969433,
0.0110191377,
-0.0392922051,
-0.0157201793,
0.0489510596,
0.0163417514,
0.0744802803,
-0.0491803847,
0.0792799145,
-0.0066969385,
0.0139813339,
0.0046288026,
0.0802944973,
0.0616966262,
-0.0192404427,
0.0400157385,
-0.0059648887,
0.0047856071,
-0.0484004617,
-0.0065992465,
0.014714594,
0.0050725024,
0.0205766018,
0.0111566382,
0.0151199074,
-0.0324884541,
-0.0169920176,
-0.0572682545,
-0.0110683823,
-0.0208779667,
-0.0114365108,
0.0349299982,
-0.0087861773,
-0.049265895,
0.0791055411,
0.0142571991,
-0.0778623149,
-0.0044843596,
-0.0669884458,
-0.0228842739,
0.0953627303,
0.0949435234,
-0.0424923375,
-0.0512300506,
-0.017065376,
-0.0088281473,
0.0155228227,
0.0163383391,
-0.013194927,
-0.0368357189,
-0.063671872,
-0.0203490984,
0.04596534,
-0.0074918764,
0.0010337434,
0.0207725838,
0.0215284564,
0.0253907926,
-0.0388383046,
-0.0149587831,
0.0271242131,
-0.0928725302,
-0.0347781144,
-0.0249719694,
-0.0237199087,
0.0293297973,
0.06500002,
-0.0030930492,
0.047444202,
0.0755543336,
-0.0172937382,
-0.0408976786,
-0.0130964052,
-0.0298295878,
0.0423779227,
0.0130949337,
-0.0020261444,
0.0415884964,
-0.0251702406,
0.0038122134,
-0.0376227163,
-0.0423448235,
-0.047293283,
0.0601928718,
-0.0202939529,
0.0177912284,
0.0446606837,
-0.0216026045,
-0.0653115734,
0.015406101,
-0.0102946656,
0.0176289808,
0.0344549119,
-0.0158490743,
0.0283938386,
-0.0394952558,
-0.0116188964,
0.0403194614,
-0.020543782,
0.043067459,
0.0139901116,
0.0248088352,
-0.0296744108,
-0.0200821124,
-0.1032473594,
0.0018861431,
0.0143418796,
0.0209529623,
0.0234009419,
0.0326278955,
-0.0081227217,
-0.0188051555,
-0.0395465009,
-0.0156188319,
0.0192340743,
0.0287614409,
0.019562548,
0.0575430579,
0.0116440691,
-0.0428047366,
-0.0133601129,
-0.0148847364,
0.0143109141,
-0.0057396702,
0.0093425279,
-0.0348272435,
0.0237023234,
-0.0457703695,
-0.0103761358,
0.0062357984,
-0.0228870343,
0.0268234015,
-0.0479044095,
0.0366170928,
-0.0103223715,
-0.0152852777,
-0.0011914221,
0.0360907204,
-0.0229978655,
-0.0529057309,
0.0015144709,
-0.0570249856,
-0.0484570749,
0.033659216,
-0.0446587987,
0.0044106175,
0.0358791053,
0.016096551,
0.0067265504,
-0.0392041169,
-0.0269091874,
-0.0259256717,
0.0116419243,
-0.0615348145,
-0.0010973501,
-0.0366696455,
0.0478224866,
0.035153605,
-0.0393082201,
0.0792975575,
-0.0103732273,
0.0367078967,
-0.0984910652,
-0.0289282482,
-0.0429218076,
0.0271238871,
-0.0178482272,
0.0225252174,
-0.0266904533,
-0.0241518579,
-0.0262062643,
0.0283669662,
-0.0167636573,
-0.0894539803,
0.0470136926,
-0.0096953241,
0.0217654351,
0.0027800233,
0.0082203485,
-0.0559258088,
0.0048753587,
0.001308102,
-0.0020422363,
-0.0069792769,
0.0235276353,
-0.034538582,
0.0332320184,
0.0241063107,
0.0671066865,
0.0280199386,
-0.0175874531,
-0.0050161048,
0.0204799641,
0.002434971,
-0.0020047589,
-0.0135302935,
0.0230490528,
0.0255025756,
0.0642153695,
0.0555082001,
-0.0229434241,
-0.0040671099,
-0.0162640233,
0.0139755262,
-0.0482769683,
-0.0145652015,
0.0502769016,
-0.0117781516,
-0.0584273264,
-0.0006448722,
0.0836249441,
0.0673348755,
-0.0433163792,
-0.004597418,
-0.0369289815,
-0.0272655562,
0.0138244461,
0.0347879268,
-0.0588251427,
-0.0018700893,
-0.0077505489,
0.0578848757,
0.0036038258,
-0.0111708427,
0.0265148114,
-0.0649012029,
0.0353672393,
0.0091836685,
-0.0032645417,
-0.0444795527,
-0.0393772498,
0.0014588911,
0.03451588,
-0.0174033791,
0.0545984656,
-0.0433127806,
-0.0008922769,
0.0014254885,
-0.0429755636,
-0.0736180767,
0.0097643388,
-0.0017703244,
0.0362909399,
0.0983211473,
0.0297735594,
-0.0153819183,
-0.0479684472,
0.0075782561,
-0.0109556923,
0.0271294601,
-0.051970806,
0.0513997078,
0.0456832908,
0.0344853066,
-0.0262806229,
0.0048798299,
-0.041745577,
-0.0386200659,
-0.0102907233,
0.0856889933,
-0.0592085756,
0.0071053896,
-0.0705194846,
0.0090991659,
-0.0030990725,
0.0517507456,
0.0002780737,
0.023304157,
-0.0226285141,
0.0113756508,
0.0470305383,
0.0279662274,
-0.0279200543,
0.0243302733,
-0.0309853647,
0.0184723176,
-0.0762857497,
-0.0070079756,
0.0205711816,
-0.002154866,
0.0069232392,
0.0399331041,
-0.0095038051,
-0.0114985863,
-0.0906035751,
0.0174331833,
0.0104481289,
0.0025565331,
-0.0714171454,
-0.0222692359,
0.0694220364,
-0.0306917708,
-0.0193743147,
0.0144618982,
0.0144506423,
-0.0209585093,
0.0202555973,
0.0570849292,
0.032264933,
-0.0115900785,
0.012356258,
-0.0040930645,
0.050830517,
0.0512175485,
-0.0501221456,
-0.0050716093,
-0.0433506481,
-0.0095083257,
-0.0147399148,
-0.0183920525,
-0.0453441031,
0.011747295,
-0.0253147446,
0.0400116369,
0.0253537577,
-0.0440644622,
0.0155947814,
0.0669350699,
0.0194943622,
-0.0130719896,
-0.0238744803,
-0.0065778005,
0.0402452312,
-0.0619308874,
0.0091882842,
-0.0096760439,
-0.0017846959,
-0.0087311054,
-0.0319258906,
0.0330556668,
0.0023535662,
-0.0504361391,
0.0170267206,
-0.0313064605,
0.043599356,
-0.0113666393,
0.0249722991,
0.0509394333,
-0.0183188282,
-0.0267442819,
-0.018964937,
-0.0167902913,
0.059602607,
-0.0317321494,
0.0660423934,
-0.0344723836,
0.0013840435,
-0.0073198085,
0.0463656038,
-0.054725498,
-0.0639102682,
-0.0351765007,
0.000339513,
0.021425372,
0.0273356084,
0.0313379541,
-0.0612885393,
-0.002215001,
-0.0350876525,
0.0084508704,
0.0063464483,
0.0317312442,
0.0202940144,
0.0247241575,
-0.0033036878,
-0.0340659805,
0.0736063719,
-0.0039666346,
0.0406325422,
-0.0440528095,
-0.0634975135,
-0.0590444393,
-0.0287138801,
0.0278891344,
0.0408711173,
-0.0646842271,
-0.0232113842,
-0.0764380917,
-0.0294260215,
0.0172771215,
0.0144675551,
-0.0498685278,
-0.0473264158,
0.0657467917,
0.0457610711,
0.0104020191,
-0.0157228988,
-0.0266431291,
0.0701381341,
-0.0748698786,
-0.0264342315,
0.0299525633,
-6.606440617e-33,
-0.0205117483,
-0.0359968171,
0.0655503571,
-0.0352496877,
-0.0118094767,
0.0183022637,
-0.0148159275,
0.0048609381,
0.0049698381,
-0.028230615,
-0.022731768,
-0.0163970813,
0.0121926861,
0.0333263017,
0.0030890224,
-0.0007929963,
0.0078565385,
0.0024093571,
-0.0012274989,
0.0513665341,
0.0229526795,
0.0450232662,
0.0297098141,
-0.0791495442,
-0.0000092938,
0.0106791127,
0.0355353728,
-0.0148459403,
-0.0135173062,
0.0670315251,
-0.0371287353,
0.0360732898,
0.0159356855,
-0.0359167121,
0.0032962197,
0.0101571176,
-0.0452126078,
-0.0163883585,
0.0028490333,
-0.0447056592,
-0.0395451002,
-0.0346943103,
0.0554097742,
-0.03437813,
-0.0378983617,
0.0291172545,
0.0106675187,
-0.021868363,
-0.038886372,
-0.0230500847,
-0.0423361212,
0.0338327065,
0.038947165,
0.0086862966,
-0.0051234034,
0.0349725671,
-0.0257514715,
0.0401414745,
-0.0630800501,
-0.0091540664,
0.0144115947,
0.0460298657,
0.046327997,
-0.0335478969,
-0.0205528159,
0.0347839408,
0.0130169019,
-0.0193573777,
-0.0109411133,
0.0089737969,
-0.0245443583,
0.0366122015,
0.0579961836,
0.0170741193,
0.0051734862,
-0.0417131223,
-0.0482415557,
0.0813981444,
-0.080379732,
0.0674715489,
0.0521263778,
-0.0090893516,
-0.0497242659,
-0.0054580988,
-0.0174833555,
0.042335283,
-0.0230324604,
-0.0627151951,
0.0428647213,
-0.000636015,
-0.0135560483,
0.0253487173,
0.0291322004,
-0.0004142954,
0.0138936462,
-0.0136973765,
0.0106174583,
0.0187105983,
-0.0523799732,
0.0074806707,
-0.0391169153,
0.0049804272,
0.0272506401,
-0.0176162906,
0.0209616739,
0.0128024304,
-0.0220579915,
0.0010271142,
-0.074446179,
0.0098292073,
0.0143768191,
-0.0043417108,
-0.0259196274,
0.0070805699,
0.0289542302,
0.0301308632,
0.0225462392,
0.0571731664,
0.0173225924,
-0.038421981,
-0.0467916466,
-0.0838610455,
0.0270785112,
-0.0204728823,
-0.0391564965,
0.0569409095,
-0.0456603691,
0.0608139597,
0.0301539674,
-0.0058900765,
0.0142508326,
0.0372031964,
3.17e-7,
0.0991036668,
0.0699270442,
0.0271241833,
-0.0377907082,
0.0174519122,
-0.0151737649,
0.0186449885,
-0.0233757403,
0.0100538507,
-0.0199557338,
0.0329782777,
-0.0212509874,
0.0336629562,
0.0565885678,
-0.0722718239,
-0.0207867026,
-0.0453069694,
-0.0374924578,
-0.0377259851,
-0.025325181,
0.0774772242,
0.0310023017,
-0.049642168,
-0.0447511263,
-0.0011356273,
-0.0034457103,
0.0271333847,
-0.0273526851,
0.0138029801,
0.0195807777,
0.0202098936,
0.0337175801,
0.0285412408,
0.0029427558,
-0.0201881435,
-0.0289771426,
0.0020932308,
0.0648373216,
-0.0469899364,
0.0150749506,
0.0357533395,
-0.031860061,
0.0313520394,
-0.051500842,
0.0405862071,
-0.0263358261,
-0.021054415,
-0.0118766101,
-0.0296662804,
-0.0098986374,
0.033173997,
-0.0155096222,
-0.0337811336,
0.0204705298,
-0.00502413,
0.0257442724,
-0.0396073088,
-0.0688188151,
-0.0070286323,
0.0373728164,
-0.0418905988,
-0.0398594029,
0.0006675399,
-0.0118769705,
-0.0263471287,
-0.0866672471,
0.0111767696,
3.370497248e-34,
0.009679893,
0.0013388227,
0.0101104015,
0.0416431762,
0.0271728132,
0.0077365339,
-0.0087105641,
0.0301896818,
0.016230667,
-0.0771970376,
-0.0088877222
] |
http://arxiv.org/abs/2401.15884v2 | http://arxiv.org/pdf/2401.15884v2 | Corrective Retrieval Augmented Generation | 2024-01-29T04:36:39 | 2024-02-16T19:10:36 | Shi-Qi Yan, Jia-Chen Gu, Yun Zhu, Zhen-Hua Ling | cs.CL | cs.CL | Large language models (LLMs) inevitably exhibit hallucinations since the
accuracy of generated texts cannot be secured solely by the parametric
knowledge they encapsulate. Although retrieval-augmented generation (RAG) is a
practicable complement to LLMs, it relies heavily on the relevance of retrieved
documents, raising concerns about how the model behaves if retrieval goes
wrong. To this end, we propose the Corrective Retrieval Augmented Generation
(CRAG) to improve the robustness of generation. Specifically, a lightweight
retrieval evaluator is designed to assess the overall quality of retrieved
documents for a query, returning a confidence degree based on which different
knowledge retrieval actions can be triggered. Since retrieval from static and
limited corpora can only return sub-optimal documents, large-scale web searches
are utilized as an extension for augmenting the retrieval results. Besides, a
decompose-then-recompose algorithm is designed for retrieved documents to
selectively focus on key information and filter out irrelevant information in
them. CRAG is plug-and-play and can be seamlessly coupled with various
RAG-based approaches. Experiments on four datasets covering short- and
long-form generation tasks show that CRAG can significantly improve the
performance of RAG-based approaches. | Corrective Retrieval Augmented Generation: Large language models (LLMs) inevitably exhibit hallucinations since the
accuracy of generated texts cannot be secured solely by the parametric
knowledge they encapsulate. Although retrieval-augmented generation (RAG) is a
practicable complement to LLMs, it relies heavily on the relevance of retrieved
documents, raising concerns about how the model behaves if retrieval goes
wrong. To this end, we propose the Corrective Retrieval Augmented Generation
(CRAG) to improve the robustness of generation. Specifically, a lightweight
retrieval evaluator is designed to assess the overall quality of retrieved
documents for a query, returning a confidence degree based on which different
knowledge retrieval actions can be triggered. Since retrieval from static and
limited corpora can only return sub-optimal documents, large-scale web searches
are utilized as an extension for augmenting the retrieval results. Besides, a
decompose-then-recompose algorithm is designed for retrieved documents to
selectively focus on key information and filter out irrelevant information in
them. CRAG is plug-and-play and can be seamlessly coupled with various
RAG-based approaches. Experiments on four datasets covering short- and
long-form generation tasks show that CRAG can significantly improve the
performance of RAG-based approaches. | [
0.0550376102,
0.0439218506,
-0.0003829938,
0.0086374832,
-0.0409051217,
-0.0161537621,
-0.0480870008,
0.0211655889,
0.0076451628,
-0.0776118711,
0.0065935729,
-0.0483417511,
-0.0163810924,
-0.0134210298,
0.0246152114,
-0.0116357468,
0.0422717966,
0.0175096039,
-0.0198565777,
0.0127596334,
0.0083306162,
0.0168255195,
0.0183227863,
0.0269069243,
-0.0009244304,
0.0199883636,
-0.0114223938,
0.0259964019,
-0.0090780854,
0.0033238167,
0.0391599797,
0.0418519303,
0.0070990711,
0.0511358753,
0.000002065,
-0.045476567,
-0.0367858857,
-0.0002222831,
0.0070194895,
0.023186544,
0.0572960936,
0.0236191191,
-0.0027160007,
0.0438540317,
-0.0323757492,
0.0119112888,
0.0110042328,
0.0334231667,
0.0512239747,
0.079409644,
0.0075732772,
-0.0087248432,
0.0316620059,
-0.0122251622,
0.042511452,
-0.0229195952,
0.0044297427,
0.0141005535,
0.0280561559,
0.0303286109,
0.0381283984,
0.0454318225,
-0.0109059047,
-0.0291606188,
0.0356718153,
0.0071627451,
-0.0841539204,
-0.0230362732,
0.020075243,
0.0398194157,
0.0243797265,
-0.0386096239,
0.0287931021,
0.0087720975,
0.0087726936,
0.0513919666,
-0.0514390953,
0.0274117943,
-0.0191250108,
0.0043196888,
0.0031483793,
-0.0308858249,
0.0323026218,
0.0369807184,
-0.0200377703,
0.0495290272,
0.0285795759,
-0.0379134864,
-0.0177551452,
0.0215886272,
0.0086937277,
-0.0022843769,
0.0005525851,
0.0412238166,
-0.0260569789,
0.0329545848,
-0.0303031616,
-0.0598587207,
0.0668868423,
-0.0482229963,
0.0188238528,
0.0169003047,
-0.0042657824,
0.009840888,
-0.0521698743,
0.0446868725,
-0.008793042,
-0.0269124303,
-0.0479774177,
-0.0067682811,
-0.0489466749,
-0.0340112634,
-0.0312452186,
0.0351013951,
-0.047122132,
-0.0170862824,
-0.0314732343,
0.052841682,
-0.0058492618,
0.0182613283,
-0.0009520029,
0.0283826087,
-0.0388384201,
-0.0155069306,
-0.032518141,
0.0094062863,
0.0156700276,
0.01643968,
0.0375414155,
-0.0537947789,
0.0022635069,
0.0166252218,
0.0173958912,
-0.0119629297,
-0.004638907,
0.0104513159,
-0.0091623031,
-0.0250045452,
-0.0691139027,
-0.0790779367,
0.0006758815,
-0.0181762446,
0.0020979948,
-0.0798924565,
0.0312977433,
0.016041562,
-0.0290684979,
-0.0324683636,
0.0143905515,
0.0498872399,
-0.0371545851,
0.0134015093,
-0.0130062578,
0.0294211302,
0.0345669538,
0.0149008334,
0.0316810645,
0.0586682074,
0.0133575657,
0.0553440079,
0.0518287383,
0.0264376868,
0.0329224057,
0.0029304582,
0.0360336676,
0.0222501475,
-0.0261056609,
0.0101214228,
-0.0198906679,
-0.0020043871,
0.013097804,
0.0594679713,
-0.0355427563,
0.0297438372,
0.0290067159,
-0.0382446349,
0.0318280347,
0.0101481341,
0.0498646088,
0.0290533025,
-0.0018008889,
0.0117284711,
-0.0124629941,
0.0017853156,
-0.0763343498,
0.0310919527,
0.0422527343,
0.0171773322,
-0.0213279761,
-0.0420854241,
-0.0051134108,
0.0126443105,
0.0002052655,
0.0007725302,
0.0444425642,
0.0378746875,
-0.0601848029,
0.0368072279,
0.00972318,
-0.0925159156,
0.0416689627,
-0.0422714353,
0.0223885477,
0.0878073052,
0.1011783853,
-0.0185829643,
-0.0031965449,
0.0364936143,
-0.0236517936,
0.0224295687,
0.0355131328,
-0.0096540963,
-0.0313355736,
-0.0755165592,
-0.063134633,
-0.0080118757,
0.01571602,
0.0104822339,
0.0479497351,
-0.0096313162,
0.0842646956,
-0.0499298833,
-0.0495127216,
0.0200000796,
-0.0384064242,
-0.0372066982,
0.0122657456,
-0.0062027443,
0.0030563981,
0.0185579956,
-0.0258873515,
0.0442270637,
0.0239379648,
-0.0042579868,
-0.0540547296,
-0.0346325189,
0.0194267631,
-0.0059805359,
0.0055179,
0.0069033136,
0.0414371677,
0.0212901477,
-0.0304498784,
-0.0516673625,
-0.0022733163,
-0.0551792607,
0.0336920433,
-0.0902972594,
0.0134207616,
0.0054238481,
-0.0026446669,
-0.057740245,
0.0837848336,
0.0076203072,
0.0115719652,
0.0341324955,
0.0138309989,
-0.0040203673,
-0.0334179178,
-0.0434314683,
-0.0206617396,
-0.0132241556,
0.0532930568,
0.0790633336,
0.011612216,
-0.0366906486,
-0.0468162224,
-0.0796402544,
0.0085538784,
-0.016013287,
0.0187411886,
0.0205471851,
0.0233500432,
0.0051176506,
0.0104614291,
-0.034821175,
-0.0049224012,
0.0387525894,
0.0354504921,
0.0059355646,
0.0279509053,
-0.0010762424,
-0.0241792016,
-0.0110501787,
0.0304364152,
0.0115205981,
0.0223402791,
0.0155557198,
-0.0568156093,
0.0866999403,
-0.068594873,
-0.0014699211,
-0.0005219438,
-0.0188185964,
-0.0087383296,
-0.0524032712,
0.0270850919,
0.0006121455,
0.0429603718,
-0.006695685,
0.0450649709,
-0.0170554612,
-0.0763451532,
-0.0072624921,
-0.0625861362,
-0.041002214,
0.0118845468,
-0.0672953874,
0.0392309353,
0.0105477348,
0.0115219522,
-0.0100407712,
-0.0463783853,
-0.0104205525,
-0.0074382075,
-0.0113895461,
-0.0524387769,
-0.0274580587,
0.000579599,
0.0431441814,
-0.0108737135,
-0.0525801592,
0.0507621653,
-0.03886218,
0.0316438898,
-0.0492095314,
0.0234995447,
-0.0089601362,
0.0414454937,
-0.0018921428,
0.0294412021,
-0.021322526,
-0.0015198096,
-0.046040453,
0.0151494099,
0.0291963592,
-0.0859401077,
0.0206418857,
-0.0428924933,
0.036561016,
0.013138121,
-0.0218335688,
-0.0694442987,
-0.013905365,
0.040236149,
-0.0062479577,
0.0178767759,
-0.0155733675,
-0.0308702104,
-0.0017779813,
0.0537230112,
0.019395018,
-0.0296802223,
-0.0272848085,
-0.0079587428,
0.0320611447,
0.044925943,
-0.0171683896,
0.0173719209,
-0.0114321066,
-0.0343669988,
0.0539093129,
0.0467088111,
-0.0175400302,
0.0060119932,
-0.0323477127,
0.0387190804,
-0.0064665982,
-0.0278659835,
0.0107813207,
-0.0112188831,
-0.0580723621,
-0.0426602587,
0.0782416388,
0.0491552092,
-0.0288386419,
-0.0358785242,
-0.0215889718,
0.0348939523,
0.0092889071,
0.0591652729,
-0.0476575121,
0.0119359922,
0.0237811301,
0.0891500339,
-0.0310336482,
-0.0116986185,
0.0021375332,
-0.1406259537,
0.0652690828,
0.0211352743,
0.0206977576,
-0.0141883595,
-0.0487592444,
-0.0109996004,
0.0216863379,
-0.006826866,
0.0804215968,
-0.0167864263,
0.0116599565,
0.0276000183,
0.0040091015,
-0.06097598,
-0.0456401706,
0.0021668812,
-0.0118983528,
0.0617356487,
0.0139749721,
0.028408451,
-0.0298424102,
0.0256365724,
-0.0452329293,
0.0071232277,
-0.0211097952,
0.0294854119,
0.0275661368,
0.0148203569,
-0.0267504565,
-0.0141243674,
-0.015768813,
-0.0131105091,
-0.0077814274,
0.0453698821,
-0.0525305122,
0.0276164431,
-0.0507098213,
-0.0239309426,
0.0160541963,
0.0727549791,
0.0133012049,
0.0057362616,
-0.0398279093,
-0.0059433905,
0.0202561896,
-0.0658113584,
0.000601186,
0.0274516176,
-0.038467899,
0.0142538408,
-0.1167379543,
-0.0422060117,
-0.0268870909,
0.0130173946,
-0.0010605246,
-0.0027367831,
-0.0108990623,
-0.0256645195,
-0.0865950137,
0.0128890527,
-0.0329950713,
-0.0165771786,
-0.0386121087,
-0.0174393617,
0.033308506,
-0.0638859123,
-0.0128998505,
0.0203048699,
-0.0352366306,
-0.0587009229,
-0.0204026494,
0.025426846,
0.0227459334,
0.0454269722,
-0.0166535564,
-0.0374369808,
0.05827903,
0.070294477,
-0.0483645312,
0.0070390287,
-0.0154995667,
-0.0162876491,
-0.0223049652,
0.0141038187,
-0.0883494616,
0.0088693835,
-0.0347151645,
0.0756543726,
0.027945336,
-0.0192232784,
0.0438965522,
0.0631670579,
0.0119250892,
-0.0388778634,
0.0298902784,
-0.0315034352,
0.0024295687,
-0.0431540459,
-0.0346120745,
-0.0343157575,
0.0422880538,
0.0133357998,
-0.0117512774,
0.0298509225,
0.0502506234,
-0.0192306526,
0.0841010138,
-0.0342972316,
-0.0448527038,
-0.0290327538,
0.0099375546,
0.040744435,
-0.0175088719,
-0.0164328683,
0.0053986036,
0.0161576569,
0.0401612297,
-0.0403152294,
0.0278532188,
0.0178836957,
-0.0071597807,
-0.0230768118,
0.011206002,
-0.0689923987,
-0.0375013947,
-0.0619381219,
-0.0128121059,
0.0043132193,
-0.0044663046,
0.0678571984,
-0.0511698052,
-0.0183180179,
-0.0345286801,
0.0066607581,
0.0448926985,
-0.0188138559,
-0.0142167192,
0.0184580721,
-0.013929314,
0.0167354923,
0.0881176814,
-0.0264491066,
0.0235616229,
-0.0651978478,
-0.0493256785,
-0.0400058031,
-0.0694780275,
0.0028989036,
0.0104470635,
-0.0335461609,
-0.0314458348,
-0.0445319749,
-0.03738866,
0.0057777753,
0.0190672595,
-0.0068723741,
-0.0082281251,
0.0516906977,
0.1071673706,
-0.0149999261,
0.0186796598,
-0.0639062524,
0.0241341442,
-0.049504973,
-0.0124524068,
0.0017003311,
-6.816208318e-33,
-0.0159632973,
-0.0217331946,
0.0261681657,
-0.0515418351,
-0.016768463,
-0.002585375,
-0.0356184095,
-0.0165663846,
0.02060977,
-0.0070290598,
0.0050288732,
0.0047957557,
0.0146576678,
0.0398160964,
0.0050939536,
0.0291011073,
0.0358376391,
0.0135725252,
-0.050797347,
0.0585104264,
0.1053457782,
0.015750749,
0.0104171485,
-0.063821286,
0.0142090619,
-0.0275761038,
-0.018690953,
0.0100765396,
0.0290584508,
0.0579488762,
0.0111523084,
0.0668309852,
0.0307378564,
-0.0025037681,
0.0008979064,
0.0036543875,
-0.0536817461,
-0.0043629725,
-0.0145901376,
-0.0264582466,
-0.0404466093,
-0.0214937422,
0.0473792367,
-0.0231740922,
-0.0145001272,
0.0323638842,
0.0298030451,
-0.0416287035,
-0.0268840641,
-0.0608127676,
-0.0442395844,
0.019098131,
0.0379782878,
-0.0159460101,
-0.0190666299,
0.0800376236,
0.0005050598,
0.0304366946,
-0.0367618613,
0.0423181802,
-0.0108830705,
0.0679322854,
0.0533729978,
0.015682986,
-0.0055021723,
0.0426440015,
-0.0008692695,
0.0097668087,
-0.0044655022,
-0.0115150521,
0.0184943154,
0.0322247036,
-0.0006079467,
0.0748775452,
-0.0024004681,
-0.0248302966,
-0.0303570051,
0.0782225132,
-0.0206345394,
0.0464790724,
0.0439607538,
-0.0191788115,
-0.0437423326,
-0.0252525378,
-0.0295440573,
0.0166062694,
-0.0485566594,
-0.0257078893,
0.0469956473,
0.07098189,
-0.0456733406,
-0.041939687,
0.0250214525,
-0.0128646037,
0.0002472961,
0.0483702458,
-0.0604445711,
-0.0166263673,
-0.0295165982,
0.0034116805,
0.0159861818,
-0.0408089757,
0.0361168422,
0.0035226736,
-0.0546067245,
0.0219392236,
0.0004011314,
0.0311575271,
-0.0393807814,
0.0321592726,
-0.0506096408,
0.0427994281,
-0.0155737074,
-0.0348795243,
0.0203739218,
0.0272317659,
-0.0007367985,
0.0223676506,
0.0305421147,
0.0388817117,
-0.0095426608,
-0.0480659455,
0.0262017939,
0.0164952166,
-0.0063293134,
0.0340777524,
-0.0190905519,
0.0378023535,
0.0248450711,
-0.030578861,
-0.0049664774,
0.012554164,
2.961e-7,
0.0310512483,
0.0396612585,
0.0137660149,
0.0127544301,
-0.0025072854,
0.0020625428,
0.0106802518,
0.0116059491,
0.0153087359,
-0.0087597147,
0.0097938366,
-0.0221405514,
0.0496029556,
0.0348625481,
-0.0685304478,
0.0014734579,
-0.0639598519,
-0.0414510518,
-0.0359737985,
-0.0230795406,
0.0622662976,
0.0461182818,
-0.0129421735,
-0.0359549634,
-0.0177711565,
-0.0692288652,
0.0771165714,
0.0202285033,
-0.0466901734,
0.0330466032,
0.0747528151,
0.0383995175,
0.0263861716,
-0.0279036351,
-0.0447655208,
0.00035318,
0.00461365,
0.0291784983,
0.0183217414,
-0.0227908399,
0.0170678552,
-0.0512756296,
0.0242604818,
-0.054454308,
0.0430939719,
0.0025958393,
0.03651537,
0.0198179074,
-0.0640509054,
0.0027951889,
-0.0055656363,
-0.0376034789,
-0.0176795889,
0.0241193194,
-0.0134517523,
0.0432050265,
-0.0165008157,
-0.0564298965,
-0.0230228044,
0.0078698033,
-0.0421699807,
-0.0462130383,
-0.0259493534,
-0.0027415804,
-0.0229975749,
-0.0577289611,
0.0189396311,
2.850216561e-34,
0.0357474871,
0.0117786014,
-0.0182264447,
-0.0152273858,
0.0255538765,
0.01832412,
-0.0257679522,
0.0364962779,
-0.0061189164,
-0.0688812956,
-0.0224697106
] |
http://arxiv.org/abs/2401.15861v3 | http://arxiv.org/pdf/2401.15861v3 | BPDec: Unveiling the Potential of Masked Language Modeling Decoder in BERT pretraining | 2024-01-29T03:25:11 | 2024-06-08T10:37:52 | Wen Liang, Youzhi Liang | cs.CL | cs.CL, cs.AI | BERT (Bidirectional Encoder Representations from Transformers) has
revolutionized the field of natural language processing through its exceptional
performance on numerous tasks. Yet, the majority of researchers have mainly
concentrated on enhancements related to the model structure, such as relative
position embedding and more efficient attention mechanisms. Others have delved
into pretraining tricks associated with Masked Language Modeling, including
whole word masking. DeBERTa introduced an enhanced decoder adapted for BERT's
encoder model for pretraining, proving to be highly effective. We argue that
the design and research around enhanced masked language modeling decoders have
been underappreciated. In this paper, we propose several designs of enhanced
decoders and introduce BPDec (BERT Pretraining Decoder), a novel method for
modeling training. Typically, a pretrained BERT model is fine-tuned for
specific Natural Language Understanding (NLU) tasks. In our approach, we
utilize the original BERT model as the encoder, making only changes to the
decoder without altering the encoder. This approach does not necessitate
extensive modifications to the encoder architecture and can be seamlessly
integrated into existing fine-tuning pipelines and services, offering an
efficient and effective enhancement strategy. Compared to other methods, while
we also incur a moderate training cost for the decoder during the pretraining
process, our approach does not introduce additional training costs during the
fine-tuning phase. We test multiple enhanced decoder structures after
pretraining and evaluate their performance on the GLUE tasks and SQuAD tasks.
Our results demonstrate that BPDec, having only undergone subtle refinements to
the model structure during pretraining, significantly enhances model
performance without escalating the finetuning cost, inference time and serving
budget. | BPDec: Unveiling the Potential of Masked Language Modeling Decoder in BERT pretraining: BERT (Bidirectional Encoder Representations from Transformers) has
revolutionized the field of natural language processing through its exceptional
performance on numerous tasks. Yet, the majority of researchers have mainly
concentrated on enhancements related to the model structure, such as relative
position embedding and more efficient attention mechanisms. Others have delved
into pretraining tricks associated with Masked Language Modeling, including
whole word masking. DeBERTa introduced an enhanced decoder adapted for BERT's
encoder model for pretraining, proving to be highly effective. We argue that
the design and research around enhanced masked language modeling decoders have
been underappreciated. In this paper, we propose several designs of enhanced
decoders and introduce BPDec (BERT Pretraining Decoder), a novel method for
modeling training. Typically, a pretrained BERT model is fine-tuned for
specific Natural Language Understanding (NLU) tasks. In our approach, we
utilize the original BERT model as the encoder, making only changes to the
decoder without altering the encoder. This approach does not necessitate
extensive modifications to the encoder architecture and can be seamlessly
integrated into existing fine-tuning pipelines and services, offering an
efficient and effective enhancement strategy. Compared to other methods, while
we also incur a moderate training cost for the decoder during the pretraining
process, our approach does not introduce additional training costs during the
fine-tuning phase. We test multiple enhanced decoder structures after
pretraining and evaluate their performance on the GLUE tasks and SQuAD tasks.
Our results demonstrate that BPDec, having only undergone subtle refinements to
the model structure during pretraining, significantly enhances model
performance without escalating the finetuning cost, inference time and serving
budget. | [
0.0423335023,
0.0174091719,
0.0018868918,
0.0541252457,
0.0104086017,
0.0193818919,
0.0243780687,
-0.0150754163,
-0.030440107,
-0.0636028051,
-0.0237499624,
-0.0013374764,
-0.0106339538,
0.0192808509,
0.0288621597,
-0.0814399496,
0.0395431779,
-0.0052523715,
0.0167069454,
0.0022091407,
0.0056160707,
0.0105635505,
-0.0208793022,
0.039309781,
0.0140034771,
-0.0109365229,
-0.0237115622,
0.0272126403,
0.019642489,
-0.0322487988,
0.0410907939,
0.0407542437,
0.0427740067,
0.0267995484,
0.000002174,
-0.0408120304,
-0.0271389354,
0.00531842,
-0.0484985299,
0.0315027907,
-0.0264762286,
-0.0277413744,
-0.0081182392,
0.0378516354,
-0.018775804,
-0.0277297832,
0.0697301403,
0.0420933552,
0.0764311627,
0.0696788728,
-0.0121383611,
-0.0464576706,
-0.0033605252,
-0.0185697805,
-0.0635403991,
-0.0482090339,
-0.0014523307,
-0.0210315455,
-0.0403739139,
-0.0228833631,
-0.0431349166,
0.0265965965,
0.0301653221,
0.0462611131,
0.037991561,
0.05488047,
-0.0352403969,
-0.0275595821,
-0.0377910435,
0.0469483435,
-0.0713557005,
-0.0224327426,
-0.0280195288,
0.0088749537,
-0.0059105791,
-0.0070308684,
-0.0035684977,
-0.0138405329,
0.0133135086,
0.0164284948,
-0.0044520632,
0.0211156812,
0.0029080447,
0.0022355537,
-0.0519449413,
0.02722154,
0.0271873474,
-0.0426980779,
-0.0482786447,
-0.0019899944,
-0.0177060459,
-0.0513152182,
0.066606909,
0.0332876332,
0.0050905934,
0.0145112174,
-0.0621700659,
-0.0866473094,
-0.0175472461,
-0.0345376395,
0.0249220226,
0.0673543885,
-0.0258877575,
0.0298203602,
-0.0553196818,
0.0580138639,
-0.0175282583,
0.0398698039,
-0.0554409735,
-0.0110930121,
-0.0133313257,
-0.0292691384,
-0.102471903,
0.027725229,
-0.0585241355,
-0.0495491512,
0.0105659589,
0.0286971219,
0.032721322,
-0.00980921,
0.037301179,
0.0112892706,
-0.059617687,
-0.0215286408,
0.0195336547,
0.0192805454,
-0.0309361052,
-0.0077223354,
-0.020413639,
-0.0390723608,
-0.0415385067,
-0.0033925117,
0.0175422803,
-0.036355339,
0.0022927977,
0.0409850925,
-0.0290907044,
-0.0487104952,
-0.0175892543,
0.0298445653,
0.0002425764,
0.0009219974,
0.0081018917,
-0.019910153,
0.016067706,
0.0102920961,
0.0175860245,
-0.025768213,
-0.0024188517,
-0.0052052056,
0.0058422522,
0.0945224538,
-0.0432806611,
-0.0154632479,
0.0211385246,
-0.0029029332,
0.0009142686,
0.0557726398,
0.0318042971,
0.0122382585,
0.0190516822,
-0.0268884469,
-0.0038154772,
0.0410113074,
0.0083783427,
-0.0340443477,
0.0020057547,
0.0007090626,
-0.0103616575,
-0.0130185271,
0.0526359938,
0.0117554832,
-0.04036038,
-0.0082095787,
0.0505539365,
0.0386364199,
0.0165432505,
0.0440423675,
0.0236097854,
0.0504461639,
0.0215133987,
-0.0023269851,
-0.0294183958,
0.0146605968,
-0.0147980908,
0.0010794885,
-0.0223307237,
0.0353210904,
-0.0320226774,
-0.0126045179,
-0.003075137,
-0.0114448592,
0.0225388892,
-0.0028986293,
0.0386874452,
0.019084353,
-0.0918949917,
0.1072880179,
-0.0058228164,
-0.0390586965,
0.009200586,
-0.0526978262,
-0.0292096864,
0.0618294701,
0.0483307987,
-0.0297831036,
-0.0689010173,
0.0130276419,
-0.0319304392,
0.0423057675,
0.0278869849,
0.0196444932,
0.0184322484,
-0.0785294473,
-0.0323636495,
-0.0304274298,
0.0108365268,
0.0440152064,
0.0048099183,
-0.0100173736,
0.000167521,
-0.0417140834,
0.0312934816,
0.0558953248,
0.0106255831,
-0.0224907473,
-0.0210712682,
0.0123759974,
0.0224153511,
0.0127401706,
-0.0274497587,
0.0050973743,
0.0165662337,
-0.0462052822,
0.0120320376,
0.0071955482,
-0.0492331013,
-0.0717433915,
0.0240804777,
-0.0501011759,
-0.0187245067,
0.1063974798,
-0.024284903,
-0.0291806832,
-0.0314401537,
-0.073023878,
0.0756510496,
-0.0426255837,
0.0355815068,
-0.0112574268,
0.0263657272,
-0.0035266802,
0.0488159731,
-0.0216836594,
-0.0123747522,
0.0304347631,
0.0159472618,
0.0334373154,
-0.0228125062,
-0.0444731824,
0.0117505584,
-0.0164253265,
-0.0076635503,
0.0053292168,
-0.007052612,
-0.0195221398,
-0.0163220987,
-0.0673291832,
-0.0027705831,
-0.0475910977,
0.042036023,
0.0324661806,
0.0017824466,
-0.0208374392,
-0.0324865356,
-0.0106096091,
-0.0510772392,
-0.04452236,
-0.0100032697,
0.0054900488,
-0.0304041356,
-0.0342085809,
-0.0001358599,
-0.0115607101,
0.0156666972,
0.0381747633,
-0.0111847902,
-0.0593892224,
-0.0723494962,
-0.0054218466,
0.0181029961,
0.0380746759,
0.0082258135,
-0.088744618,
-0.0067842207,
0.0187728442,
0.0533378907,
0.0541392528,
0.028802814,
0.0294469781,
0.0229339004,
0.0125320731,
-0.0445066318,
0.0429446436,
-0.0138842594,
-0.0067275977,
-0.0359476618,
-0.0096983146,
0.0267599467,
0.0755305067,
-0.013861062,
0.0083063319,
-0.0941778719,
-0.0281561706,
0.0044119861,
-0.0440400504,
-0.0272341166,
-0.028645236,
-0.0284409802,
0.0003763978,
-0.0466337167,
-0.0188727491,
0.0836046711,
-0.054723464,
-0.0316561982,
-0.068226397,
0.0122022899,
-0.0105666202,
0.0187535565,
0.0083051091,
0.0391577445,
-0.0097141042,
-0.0728003755,
-0.0018168984,
0.0197646599,
0.0044608503,
-0.0495547093,
-0.0230166614,
-0.0046829749,
-0.010533846,
-0.016645249,
-0.0145596741,
-0.0590021685,
-0.0311907865,
-0.037172161,
0.0323129147,
0.0454406403,
-0.0283466782,
-0.0197624583,
0.0305071045,
0.0035088127,
0.0596175715,
0.0206778757,
0.0243480597,
-0.0569552854,
0.0263549127,
0.0114624882,
0.0175059456,
0.0112186894,
-0.0420214161,
0.0153788859,
0.0329516828,
0.0357384309,
-0.0053176139,
-0.04547197,
0.0241425075,
0.0467028134,
-0.0405944102,
-0.0008759073,
0.0122136772,
0.0460168831,
-0.0740620792,
-0.0093474416,
0.1212580875,
0.0467395075,
-0.00582251,
-0.0111692548,
-0.0106964558,
0.0244199354,
0.0150929112,
0.0176007282,
-0.0468495488,
-0.0516483895,
0.031393487,
0.0461459309,
0.0197279006,
0.0142886182,
-0.0102099665,
-0.1167603731,
0.0160138234,
0.0027136677,
-0.002754899,
-0.0457869843,
-0.0584208518,
-0.0312102884,
0.0170872714,
0.0330609418,
0.0078431321,
0.0347676016,
-0.015395903,
0.039643947,
-0.0457558855,
-0.0654792339,
0.009014708,
-0.0478341505,
0.0356760062,
0.0765616968,
0.0191837735,
-0.008445085,
-0.0420233831,
0.0360000357,
-0.0309501104,
-0.0125564151,
0.0018415464,
0.0656448156,
-0.0595926605,
-0.0312711298,
-0.0121111376,
0.0056841183,
-0.020081697,
-0.013867476,
-0.0370415077,
0.0530838966,
-0.0004054867,
0.013929286,
-0.0095853135,
-0.0142397778,
-0.0139289713,
0.0811351314,
0.0364837535,
-0.0345478728,
-0.0192201436,
-0.0227597542,
0.033448156,
-0.0755900592,
-0.0032421574,
0.0400652997,
-0.0390512571,
-0.0490351766,
-0.0388492905,
-0.0356557518,
0.0448493883,
0.0120981103,
0.0071531655,
-0.0300109107,
0.0424844362,
-0.00559896,
-0.0665251985,
0.0365955941,
0.0001642747,
0.036175821,
-0.0490147248,
-0.04359008,
0.0442814194,
-0.0116854841,
-0.0178825669,
0.0045190402,
-0.0084364656,
-0.0361292101,
0.0224392693,
0.0625908002,
0.0173722319,
0.0493004546,
-0.0462503582,
-0.0743456706,
0.0316103622,
0.0224309042,
-0.0279204454,
0.004029206,
0.0540577956,
-0.0385622606,
-0.0017271077,
0.0190971307,
-0.0291871447,
0.0203239545,
-0.0052990192,
0.0380406901,
-0.0223168787,
-0.0052197613,
0.0355816521,
0.0299713425,
0.0578495488,
-0.0248535965,
0.0277062282,
-0.0144534055,
0.0538113676,
-0.0178578738,
-0.0070630377,
-0.0491769165,
-0.0138317123,
0.0353505984,
-0.019095134,
-0.007526638,
0.0708673149,
-0.011675478,
-0.0088124471,
-0.0355068706,
0.0524872914,
-0.0124898627,
-0.002670194,
0.0176935215,
-0.005680732,
-0.0383569337,
-0.0120133916,
0.0094766421,
0.0085462518,
-0.0047521391,
0.0441441312,
0.0227239467,
0.0018943519,
-0.0293826181,
0.0556754656,
-0.0526262149,
-0.0159797985,
-0.0312920958,
0.0179736558,
0.0299317874,
-0.0188433155,
0.0486013405,
0.0012215019,
-0.0073724636,
0.0217673928,
0.0090164486,
-0.0100338142,
-0.0244593751,
0.0261602383,
0.0142517993,
0.0006985904,
-0.0136035429,
0.052912578,
0.0044748234,
0.0627999082,
-0.025058914,
-0.0193043817,
-0.048930537,
-0.0017702599,
0.0532115139,
0.0335042067,
-0.026655836,
0.0196264163,
-0.0233978499,
-0.0574380122,
0.0043259328,
-0.0117815994,
0.0380590148,
-0.0190658048,
0.0370097794,
0.0291358437,
-0.037451122,
-0.051995907,
0.0109024113,
0.0131896259,
0.0374987833,
0.0192287266,
-0.0283719338,
-6.648995716e-33,
-0.0363541655,
0.0149446288,
0.0009796806,
-0.0299975742,
-0.0345344283,
0.0205868036,
-0.0481678136,
-0.0273254048,
-0.0193205066,
0.0112046162,
-0.0053642988,
0.014200096,
0.0200426318,
0.0360851251,
-0.0550467186,
-0.0288810823,
0.0807948336,
-0.0057547702,
-0.0234082881,
-0.0328910388,
0.0084224818,
0.059772633,
0.0416583456,
-0.0315546319,
-0.0392236374,
-0.0132671027,
-0.0280089285,
0.0279866662,
0.0304560531,
0.0387156196,
-0.0216496252,
0.0228654817,
0.0124851381,
-0.0354839154,
-0.0347970836,
0.0427392498,
-0.1134961098,
0.0140722971,
0.0052983188,
0.0585579798,
-0.0189699586,
-0.0278804619,
0.0514414832,
0.0232547987,
0.0089138839,
0.013356735,
0.0175350644,
-0.031299483,
0.0050749057,
-0.0305618029,
-0.0327990949,
0.0224268716,
-0.034092553,
-0.0576567091,
0.0639818013,
0.0363699123,
0.0105334027,
-0.0010511791,
-0.0773808584,
-0.0328865685,
0.0446406268,
0.0211811215,
0.0645802468,
0.0531421229,
-0.009657613,
0.0290427878,
0.0291854292,
-0.0200154409,
-0.0138305239,
0.0065945387,
0.043240644,
0.0350490697,
0.0510144867,
-0.0250200815,
0.0504852459,
-0.0008765929,
-0.0280248392,
0.0259362906,
-0.0335441157,
0.0203403328,
0.0404679589,
0.0443617925,
-0.0253239051,
0.0032913624,
-0.0249289982,
-0.0047424776,
-0.0203750767,
0.0114046559,
0.047791332,
0.0134931849,
-0.0521776527,
0.0209262539,
0.0466597006,
-0.0253487956,
0.0763555467,
0.0047127488,
0.0176676065,
0.037775632,
-0.0132269571,
0.0692526996,
-0.0252626408,
-0.0322522707,
0.0436336249,
-0.0023743585,
-0.0165429134,
-0.0051133493,
0.0242061373,
0.0247457232,
-0.053634081,
0.0129345981,
-0.036219731,
-0.0058048498,
-0.001627534,
0.058529567,
0.0278133787,
0.0143961115,
-0.0218402315,
0.007533296,
0.0248722285,
0.0006848361,
-0.0344335288,
0.0271618217,
0.0072102067,
0.0347154737,
0.0058920924,
0.012693814,
-0.0413686149,
0.035649512,
0.0281324293,
-0.063595295,
-0.0287781395,
0.0099254716,
2.89e-7,
0.0730113909,
0.0483465493,
0.0741029307,
0.0141372634,
0.0031067498,
0.0261977632,
0.0089044292,
0.0620898306,
0.0388464667,
-0.0096047223,
-0.0020484328,
0.0427782722,
0.0241252668,
-0.0214382596,
-0.0538950823,
-0.0570655055,
-0.0365060791,
0.0115381163,
-0.0389043912,
0.0056830333,
0.0657307431,
0.095309563,
-0.0039100023,
0.0180155858,
0.0234298576,
-0.0097510926,
0.0449905954,
-0.0082710069,
0.0148551967,
0.0192693006,
-0.0006457769,
0.0396163762,
-0.0232699811,
0.027302919,
-0.0262934305,
0.0557604581,
0.0261530075,
0.0627940372,
-0.0421643108,
-0.0194412414,
-0.0045184791,
-0.011623973,
-0.0372189917,
-0.0251560491,
0.0686440691,
-0.0674966723,
-0.0136212576,
0.0095046675,
-0.0997602642,
0.0210456047,
-0.0149084376,
-0.029545106,
-0.0340993069,
0.0401190966,
0.034816362,
-0.008126637,
-0.0259135999,
-0.0953612924,
-0.0354817994,
-0.0152984615,
-0.0570220575,
-0.0087591801,
-0.0355840325,
-0.0206399933,
0.0598654114,
-0.071064949,
0.0058966945,
2.65697585e-34,
0.0343989953,
-0.0106538478,
0.0129026575,
0.065287374,
0.0388520397,
-0.0303891003,
0.0252370536,
0.0018546769,
0.0294425152,
-0.0866185352,
-0.0263748039
] |
http://arxiv.org/abs/2401.15854v2 | http://arxiv.org/pdf/2401.15854v2 | LSTM-based Deep Neural Network With A Focus on Sentence Representation for Sequential Sentence Classification in Medical Scientific Abstracts | 2024-01-29T03:05:35 | 2024-05-31T08:37:04 | Phat Lam, Lam Pham, Tin Nguyen, Hieu Tang, Michael Seidl, Medina Andresel, Alexander Schindler | cs.CL | cs.CL | The Sequential Sentence Classification task within the domain of medical
abstracts, termed as SSC, involves the categorization of sentences into
pre-defined headings based on their roles in conveying critical information in
the abstract. In the SSC task, sentences are sequentially related to each
other. For this reason, the role of sentence embeddings is crucial for
capturing both the semantic information between words in the sentence and the
contextual relationship of sentences within the abstract, which then enhances
the SSC system performance. In this paper, we propose a LSTM-based deep
learning network with a focus on creating comprehensive sentence representation
at the sentence level. To demonstrate the efficacy of the created sentence
representation, a system utilizing these sentence embeddings is also developed,
which consists of a Convolutional-Recurrent neural network (C-RNN) at the
abstract level and a multi-layer perception network (MLP) at the segment level.
Our proposed system yields highly competitive results compared to
state-of-the-art systems and further enhances the F1 scores of the baseline by
1.0%, 2.8%, and 2.6% on the benchmark datasets PudMed 200K RCT, PudMed 20K RCT
and NICTA-PIBOSO, respectively. This indicates the significant impact of
improving sentence representation on boosting model performance. | LSTM-based Deep Neural Network With A Focus on Sentence Representation for Sequential Sentence Classification in Medical Scientific Abstracts: The Sequential Sentence Classification task within the domain of medical
abstracts, termed as SSC, involves the categorization of sentences into
pre-defined headings based on their roles in conveying critical information in
the abstract. In the SSC task, sentences are sequentially related to each
other. For this reason, the role of sentence embeddings is crucial for
capturing both the semantic information between words in the sentence and the
contextual relationship of sentences within the abstract, which then enhances
the SSC system performance. In this paper, we propose a LSTM-based deep
learning network with a focus on creating comprehensive sentence representation
at the sentence level. To demonstrate the efficacy of the created sentence
representation, a system utilizing these sentence embeddings is also developed,
which consists of a Convolutional-Recurrent neural network (C-RNN) at the
abstract level and a multi-layer perception network (MLP) at the segment level.
Our proposed system yields highly competitive results compared to
state-of-the-art systems and further enhances the F1 scores of the baseline by
1.0%, 2.8%, and 2.6% on the benchmark datasets PudMed 200K RCT, PudMed 20K RCT
and NICTA-PIBOSO, respectively. This indicates the significant impact of
improving sentence representation on boosting model performance. | [
0.0072789984,
0.0364927463,
0.0249577258,
0.0022211405,
-0.0474889167,
0.0029418436,
0.0089623788,
-0.001326442,
-0.0018996143,
-0.0340815261,
-0.0120280534,
-0.0797907561,
0.0150817074,
0.0302518904,
-0.0012545022,
-0.077872023,
0.0420388542,
0.058886122,
0.0409672186,
0.0378690027,
0.0101215709,
-0.000787405,
-0.0132715171,
0.0423597805,
0.0138318082,
-0.0330157615,
0.0413756333,
-0.003392563,
0.0432041883,
-0.0954030901,
0.0515071712,
0.0269713961,
0.0307343192,
0.0189325176,
0.0000017977,
-0.0334264599,
-0.0083423052,
-0.0081480546,
-0.0108694686,
0.0255151503,
0.0399500467,
-0.0649660081,
0.0197385643,
0.015375725,
-0.0241761971,
-0.0157007296,
-0.0019239396,
0.0944514498,
0.0200119969,
0.0764713958,
-0.0146340337,
-0.026688572,
0.046119649,
-0.0377728343,
-0.0228446331,
-0.0133004813,
0.0145229213,
-0.0253080614,
0.0082327146,
0.027399037,
-0.0033095826,
0.0439349078,
0.0362888612,
0.0049523883,
0.0337913893,
0.0275257733,
-0.0403933637,
0.0202492345,
0.0102931643,
0.0467682518,
-0.0503395759,
-0.0309846923,
0.0156868789,
-0.0082992027,
-0.0122606456,
-0.0469736978,
-0.0478333756,
0.0139289303,
-0.0044523124,
0.0338050462,
0.0359977819,
0.0209270343,
0.0367216691,
0.0347110219,
-0.0398489498,
0.0021448408,
-0.0053555667,
-0.0669110864,
-0.0023990462,
-0.0030720995,
-0.0201311968,
-0.0138061503,
0.0620992072,
0.0362503938,
-0.0423617959,
-0.0089446222,
-0.0209026225,
-0.0720899627,
0.056610696,
-0.0595639236,
-0.0271960348,
-0.0113314204,
-0.021124227,
0.0146172429,
-0.0275722072,
0.0226444043,
-0.0360630043,
0.0364586003,
-0.0347079895,
0.0241546761,
-0.0208363589,
-0.0458054468,
-0.0052818432,
0.047827363,
-0.0153727569,
-0.0276295077,
-0.0821845531,
-0.0210413933,
0.0113788582,
-0.0088163018,
-0.0125730494,
0.0357454009,
-0.0607575364,
0.0058923452,
0.0116597293,
0.0258987136,
0.0092740217,
-0.0120192952,
-0.0043286206,
-0.0095255198,
-0.0222117044,
0.0372917056,
0.031792134,
0.0029962263,
0.0206454564,
-0.0015642657,
-0.0032168317,
-0.0380864665,
0.0321385302,
-0.054778263,
-0.0218874179,
-0.0208497345,
0.0362989269,
-0.0337912813,
0.0178406276,
-0.0022606959,
-0.015947992,
-0.0077251117,
0.007848938,
0.0110816564,
-0.0081716878,
0.058217667,
-0.0260868128,
-0.062046703,
-0.0088945832,
0.0386853069,
0.0745087862,
-0.002008881,
0.0496963859,
0.0190067254,
-0.0185047109,
0.010363684,
0.0062435581,
-0.043883387,
-0.0036305273,
-0.0830460638,
0.0071186395,
0.0021906765,
-0.0420962572,
0.0935288444,
-0.0090964651,
0.0379893631,
0.0021178424,
0.0042303219,
0.0048008906,
0.0657457709,
0.0188558996,
0.0389902443,
0.0123194773,
0.0287223496,
0.0091909636,
0.0008223712,
0.0207279716,
0.0058044684,
0.0617729127,
0.0288767554,
-0.0313970782,
-0.0114074768,
-0.0214819796,
-0.0228759442,
-0.0055550677,
-0.0273170341,
-0.0435514301,
-0.0248850435,
0.0353483781,
0.0013817059,
-0.0460425727,
0.0480295457,
0.0096938489,
-0.0117603131,
-0.0355024375,
-0.0102361813,
-0.0108483657,
0.0230297539,
0.0484755896,
-0.0458509363,
-0.0664672181,
-0.0213899091,
0.0258266497,
0.066754438,
0.0388763323,
0.0136294505,
0.0127163185,
-0.0570594296,
-0.0072313854,
0.0038533956,
-0.0033683709,
-0.0068710488,
-0.0463457555,
0.0063033011,
0.0263457112,
-0.0086050071,
-0.0096847368,
0.047221452,
-0.0029522849,
-0.0155265881,
0.0032934546,
0.0054337084,
0.0527054071,
0.0092317984,
0.0044995691,
0.0312144812,
0.0369126238,
-0.0638091415,
-0.0385912508,
0.0115266815,
-0.0835449919,
-0.0615950599,
0.0037012345,
-0.0363353156,
-0.0075798188,
0.0534726828,
-0.0666440204,
-0.0392818265,
0.0086743617,
-0.0303861164,
-0.0252465755,
-0.0147454524,
0.0050137029,
0.0419767946,
0.026613703,
-0.0278753769,
0.0126390588,
-0.0282335039,
0.0172849521,
0.0075011104,
-0.0236539394,
0.0513768792,
-0.0267649554,
-0.0091403285,
-0.0046296488,
-0.0125860693,
-0.017687548,
0.0064876429,
0.0222339183,
-0.0137130208,
-0.0036496618,
-0.0882912427,
0.015364402,
0.0131649151,
0.0084123304,
0.0089642387,
0.0092365658,
0.0260067005,
-0.0216468349,
-0.0497487597,
-0.0364933722,
0.023770133,
-0.0090859309,
0.0828474239,
0.0301834065,
-0.0162674133,
-0.0305040926,
-0.0124690402,
-0.0539743714,
0.0350905545,
-0.0049756346,
-0.0445223972,
-0.0169661436,
-0.0286380909,
0.0146650616,
-0.0006317989,
0.0015859453,
-0.0416409597,
-0.0039671296,
-0.0088404678,
-0.0021429483,
0.0588188507,
0.0300950948,
0.0321806297,
0.025229862,
0.0635554567,
0.0153781036,
0.0285143591,
-0.0362674929,
0.0906740725,
0.0076847724,
-0.0123946555,
0.0290679038,
0.0779205412,
0.0188604705,
0.0174947493,
0.0079413746,
-0.0513053201,
0.0084695453,
0.0176456999,
-0.043452017,
-0.0137159834,
-0.0438378192,
-0.000907617,
0.005853455,
-0.0361397192,
0.0605960749,
-0.000419142,
0.0108321579,
-0.1006496847,
-0.027664803,
-0.0517214313,
0.0141118979,
0.0209510494,
0.0068515614,
0.0032420824,
-0.0093427859,
-0.0003666638,
0.0149876364,
-0.0453815721,
-0.0342711061,
0.0204003844,
-0.0179990437,
-0.0019621667,
-0.0251289103,
-0.0587041825,
-0.0182788633,
-0.0364362374,
-0.0221304614,
0.0283285696,
0.0049566319,
-0.0017191317,
-0.0369488373,
0.0143998144,
-0.0075298641,
0.0608077943,
-0.0376496315,
0.0221687127,
-0.0476232059,
0.059497945,
0.0675004199,
0.0100128213,
0.0181154422,
-0.0277154483,
0.064289622,
0.0461724102,
0.0082749957,
0.029173987,
-0.0429522321,
-0.0321837142,
-0.0148214232,
-0.0587406829,
0.0052079572,
0.0144968154,
0.0231699012,
-0.0954689905,
-0.0119520966,
0.1135956571,
0.021589214,
-0.0685831755,
-0.0037970606,
0.0224738251,
-0.0129096163,
-0.0298060291,
-0.001661217,
-0.0578507297,
-0.0236309115,
-0.0411151946,
-0.0200623367,
-0.0000116589,
-0.0049949903,
-0.0012107386,
0.0092272982,
0.0130673554,
0.0138055086,
-0.0778853223,
-0.0243979935,
-0.0112625305,
-0.0480417721,
0.0269542504,
0.0258544516,
-0.0045357719,
-0.0269355308,
-0.0233858693,
0.0213625226,
-0.1301975399,
-0.0408399291,
-0.003092709,
-0.0298941229,
-0.0081971316,
0.0410949439,
0.0560516864,
-0.0411114767,
0.0101644816,
0.0562645718,
-0.0518945605,
-0.0291640591,
0.0150974514,
0.0455572866,
-0.0212160069,
-0.0148122692,
-0.0212777723,
0.0076465905,
-0.0529762805,
-0.0432789996,
-0.0271686036,
0.0719503611,
-0.0606960729,
-0.0398166366,
-0.0869478434,
0.014185112,
-0.0191257205,
0.0445986614,
0.0052333819,
-0.0234109312,
-0.0864131749,
0.0217783432,
0.0387942567,
-0.0536934063,
0.0154403718,
0.0035219835,
-0.0215483941,
0.0315878317,
-0.0607584715,
-0.0579417795,
0.012737059,
-0.0356048159,
-0.0078426115,
0.0427855738,
-0.0053028744,
-0.022236865,
-0.0711244717,
0.037627846,
0.0316811278,
0.0037211953,
-0.0394777879,
-0.0094393659,
0.0676445886,
0.0088058002,
-0.0358265527,
-0.0109697012,
-0.0299939569,
0.0137164481,
-0.0051212008,
-0.0081142038,
0.0038339032,
-0.0023180805,
0.002736921,
-0.028508842,
0.0886561796,
0.0078847129,
0.0249279346,
-0.0385768414,
0.0301592629,
0.0130411163,
-0.0133497883,
0.02264864,
-0.0095318872,
0.0526351146,
-0.0082371272,
0.0406714082,
0.0037202139,
-0.0009544035,
0.0068417192,
0.0022690254,
0.0647128001,
-0.0165552385,
-0.0085321153,
-0.0140303271,
0.0208422057,
-0.007455789,
0.0316760689,
-0.0249363016,
0.0355707482,
0.036200393,
0.0172975622,
-0.0284181312,
0.0228690654,
-0.0145340478,
-0.0145114446,
-0.0232822485,
0.0321483836,
-0.0155234495,
0.0096103884,
-0.0062143924,
-0.0008421107,
-0.0258259326,
-0.0172180943,
-0.0143762026,
0.0138686048,
0.009315067,
0.0077519594,
0.0303067863,
0.0539530329,
-0.0415415801,
0.0734553859,
-0.0800616145,
0.0417201258,
-0.0426009819,
0.0001488239,
0.0542367585,
-0.0379624851,
0.0346833505,
-0.0360209495,
0.0113698542,
-0.0291620102,
0.0616471879,
-0.0188300405,
0.0043354384,
0.0481990613,
0.0136720445,
-0.0220939182,
-0.0294278339,
0.0668602586,
0.0251988787,
0.0199450012,
0.0413476527,
-0.0461168066,
-0.0404191241,
0.0064322506,
0.001738817,
0.0176205095,
-0.0153486347,
0.0095075164,
-0.0070481701,
0.0121293375,
0.1005175859,
-0.0220063906,
0.0204601511,
-0.03683291,
0.0532401577,
0.0081932126,
0.0144854868,
-0.0343435109,
-0.0036787633,
0.0329917297,
0.0088996012,
-0.0424087308,
-0.026683541,
-5.526754933e-33,
0.0207105912,
-0.0119469222,
-0.0167530552,
0.0139114289,
-0.0204640683,
0.0452972576,
-0.0762359276,
-0.0186048076,
-0.0049968697,
0.0547644235,
-0.0195063967,
0.0211279821,
0.0083490517,
0.0702183917,
0.0024677513,
0.0003224681,
0.0509440862,
-0.0160614885,
0.0051356996,
0.0189461745,
0.009039632,
0.0763992742,
0.0769819021,
-0.0581106097,
-0.0038847437,
-0.0161471441,
0.0297748186,
-0.0084343962,
0.00254159,
0.040526621,
-0.0365123935,
0.0298136566,
0.0609530695,
0.0030277239,
-0.0066418997,
0.0378855802,
-0.0476439521,
-0.030358186,
-0.0330464579,
-0.0081742937,
-0.0824000388,
-0.0449520685,
0.0470611453,
-0.0025436699,
0.0206627771,
-0.0025089951,
0.0458085239,
-0.025592126,
-0.0165745169,
-0.0811054036,
-0.044371672,
0.0317948796,
-0.0046640141,
0.0175757818,
0.022980826,
-0.0042606606,
-0.0321134329,
0.0603337102,
0.003805469,
-0.0378966257,
0.0172381122,
0.0653206855,
-0.0135611193,
-0.0124198562,
-0.0293478854,
0.0458723344,
-0.0022393174,
-0.0079571512,
-0.0081222057,
0.0476191081,
-0.0342605375,
0.0213477183,
0.0759034529,
0.0080139283,
0.0700225607,
-0.0023935798,
-0.113162905,
0.0104524586,
-0.0158407539,
0.0512328483,
0.0355835594,
-0.0084452368,
-0.0165944248,
0.0053724903,
0.0342085212,
0.0119136041,
0.0085982354,
-0.0169691201,
0.0539609008,
0.0102376807,
0.0176400598,
0.0355098248,
-0.0106752515,
-0.0345066711,
0.0155708119,
-0.0240943879,
0.0336712189,
0.0031125962,
0.0076078712,
-0.0087792668,
-0.0581990741,
-0.0355692245,
0.0103318123,
-0.0042760372,
0.0227732603,
-0.000632384,
0.0597585775,
-0.0297089089,
-0.0158806089,
-0.0114653818,
0.0368114747,
-0.0424937233,
-0.0231312793,
0.0676734596,
-0.0065598134,
0.0107205166,
0.0090009058,
0.0709961355,
-0.0072942949,
-0.0127086155,
-0.0137459394,
-0.0157019217,
0.0244679786,
0.0415136851,
-0.0275460817,
-0.0257426947,
-0.0205389652,
0.0299355723,
0.0005118836,
-0.0873260796,
-0.0056281504,
-0.0119752381,
2.529e-7,
0.0720282719,
0.0309713092,
0.0317530595,
-0.01561186,
0.0402322747,
-0.0274150074,
0.0097708534,
0.0516168587,
-0.0477260463,
0.0085918382,
0.0670359358,
-0.0125178937,
0.0357832424,
-0.0026166537,
-0.1339653432,
0.0186963081,
-0.0840959474,
-0.0330097638,
-0.0318277478,
-0.0013381515,
0.0519286394,
0.0997964889,
-0.057676401,
-0.0210978352,
0.0271481313,
-0.0145306997,
0.0333181769,
-0.0072880709,
0.0057558897,
0.0416647978,
0.0010317183,
0.0122629739,
-0.0089780195,
0.012813312,
-0.0269612763,
0.032699462,
0.0022576391,
0.0474203452,
-0.0077375062,
-0.0313507058,
0.0390697047,
0.0233960338,
0.0520646796,
0.007752846,
0.0297073014,
-0.0702894032,
-0.0296724699,
-0.0238833465,
-0.0289033595,
-0.0109913405,
0.0931201503,
0.0082375351,
-0.0493925139,
0.0136690522,
0.0257934667,
-0.0350295119,
-0.0331340879,
-0.1142050847,
-0.0276732538,
0.0529941469,
-0.0326889008,
0.0284598377,
-0.0048860889,
-0.0131511698,
0.0530036278,
-0.0657377616,
-0.0257948618,
2.534703775e-34,
0.0617614202,
-0.0179251432,
-0.003631151,
0.0632821023,
0.0012062708,
-0.0228672754,
-0.0075047715,
-0.0063090506,
0.0286031347,
-0.0650369003,
0.0135713359
] |
http://arxiv.org/abs/2401.15847v3 | http://arxiv.org/pdf/2401.15847v3 | Muffin or Chihuahua? Challenging Multimodal Large Language Models with Multipanel VQA | 2024-01-29T02:43:40 | 2024-06-27T15:38:17 | Yue Fan, Jing Gu, Kaiwen Zhou, Qianqi Yan, Shan Jiang, Ching-Chen Kuo, Xinze Guan, Xin Eric Wang | cs.CV | cs.CV, cs.AI, cs.CL | Multipanel images, commonly seen as web screenshots, posters, etc., pervade
our daily lives. These images, characterized by their composition of multiple
subfigures in distinct layouts, effectively convey information to people.
Toward building advanced multimodal AI applications, such as agents that
understand complex scenes and navigate through webpages, the skill of
multipanel visual reasoning is essential, and a comprehensive evaluation of
models in this regard is important. Therefore, we introduce Multipanel Visual
Question Answering (MultipanelVQA), a novel benchmark comprising 6,600 triplets
of questions, answers, and multipanel images that specifically challenge models
in comprehending multipanel images. Our evaluation shows that questions in the
MultipanelVQA benchmark pose significant challenges to the state-of-the-art
Multimodal Large Language Models (MLLMs) tested, even though humans can attain
approximately 99% accuracy on these questions. Distinctively, the MultipanelVQA
benchmark features synthetically generated multipanel images specifically
crafted to isolate and assess the impact of various factors, such as the
layout, on MLLMs' multipanel image comprehension abilities. As a result, in
addition to benchmarking the capabilities of MLLMs in understanding multipanel
images, we analyze various factors of the multipanel image that affect MLLMs'
performance with synthetic data and offer insights for enhancement. Code and
data are released at https://sites.google.com/view/multipanelvqa/home. | Muffin or Chihuahua? Challenging Multimodal Large Language Models with Multipanel VQA: Multipanel images, commonly seen as web screenshots, posters, etc., pervade
our daily lives. These images, characterized by their composition of multiple
subfigures in distinct layouts, effectively convey information to people.
Toward building advanced multimodal AI applications, such as agents that
understand complex scenes and navigate through webpages, the skill of
multipanel visual reasoning is essential, and a comprehensive evaluation of
models in this regard is important. Therefore, we introduce Multipanel Visual
Question Answering (MultipanelVQA), a novel benchmark comprising 6,600 triplets
of questions, answers, and multipanel images that specifically challenge models
in comprehending multipanel images. Our evaluation shows that questions in the
MultipanelVQA benchmark pose significant challenges to the state-of-the-art
Multimodal Large Language Models (MLLMs) tested, even though humans can attain
approximately 99% accuracy on these questions. Distinctively, the MultipanelVQA
benchmark features synthetically generated multipanel images specifically
crafted to isolate and assess the impact of various factors, such as the
layout, on MLLMs' multipanel image comprehension abilities. As a result, in
addition to benchmarking the capabilities of MLLMs in understanding multipanel
images, we analyze various factors of the multipanel image that affect MLLMs'
performance with synthetic data and offer insights for enhancement. Code and
data are released at https://sites.google.com/view/multipanelvqa/home. | [
0.0252544004,
0.0514223203,
-0.0173415318,
0.0675017461,
-0.0263960138,
0.035240721,
0.0579025745,
0.0074153049,
0.0140858246,
-0.0580260679,
-0.0394420475,
-0.0290866774,
-0.0179712623,
0.0248840768,
0.049010545,
-0.0754125565,
0.0257407501,
-0.0358139537,
-0.0145759918,
-0.0137147326,
0.0129433032,
0.0083246939,
-0.0260500219,
0.0098002367,
-0.0449208468,
-0.0604333319,
-0.0315308645,
0.0202687513,
0.0120786754,
-0.0265028793,
-0.014063389,
0.0255496893,
0.0469647385,
-0.0001592256,
0.0000019076,
-0.0086724907,
0.0219585858,
-0.0193932261,
0.0089121396,
0.0274062641,
0.0149739943,
-0.0160285961,
0.0173629802,
-0.0247801356,
-0.0442333296,
-0.0000321834,
0.0453104861,
0.0610229634,
0.1035669371,
0.0433042906,
-0.0131734889,
-0.0358059332,
-0.0083679864,
0.0205591917,
0.0376173072,
-0.0652815551,
-0.0021713118,
0.0494002178,
0.015469851,
-0.0263074432,
-0.057158161,
0.0403696671,
0.0430178978,
0.0138496878,
-0.0353428982,
0.0414194912,
-0.0322445892,
-0.0630494505,
-0.0338658057,
0.1065048873,
-0.0193565339,
0.0125897098,
-0.0326061994,
-0.0035460421,
0.0085031362,
0.0201898273,
-0.0739220306,
0.0196083318,
-0.0286632031,
-0.0055922568,
-0.0620875396,
-0.0227178205,
-0.0085162595,
0.0080618113,
-0.0250226948,
0.016759634,
0.023503188,
0.0015915858,
-0.0108341444,
0.0109519614,
-0.0384765156,
-0.0241622962,
-0.0125363758,
0.046205271,
-0.0078365495,
0.0118790912,
-0.0106106447,
-0.0436855815,
0.0025294009,
-0.0261612274,
0.006453773,
-0.0054108575,
-0.0513683483,
0.086342454,
-0.0273292568,
0.0720517933,
-0.0348507501,
0.0542556681,
-0.033998888,
-0.0173486937,
-0.0130523276,
-0.0134901479,
-0.0504947305,
-0.0030924852,
-0.0521541685,
-0.0466915108,
0.0139032332,
0.0245117452,
-0.0102381585,
0.0130613195,
-0.0034403263,
0.0425044894,
-0.0052142907,
-0.0173696335,
-0.0279463939,
0.0438729599,
-0.0646448284,
-0.0016848306,
-0.0113219805,
-0.0717614666,
0.0045611262,
0.0411599986,
0.0137206968,
-0.0450266376,
0.0036685579,
0.0788573399,
-0.0307356175,
-0.0257152822,
-0.0035809956,
-0.0253944341,
-0.0195533633,
-0.0693989322,
0.0532582104,
-0.0441409089,
0.0562836118,
-0.0069801742,
0.0064304476,
0.0046071694,
-0.0166043509,
0.0184936039,
-0.0090462789,
0.0782900974,
0.0188225675,
0.0016059935,
0.0149303088,
0.0092220809,
-0.0224521216,
0.0742993951,
0.0136341918,
0.0572723895,
0.0568823554,
-0.0246505439,
0.0216929223,
0.0570295602,
-0.0011351695,
-0.0142271668,
-0.0138446921,
0.045273345,
0.0170527287,
0.0278468356,
0.0275565684,
0.0358438678,
-0.0278312974,
-0.0585738681,
0.0829212964,
0.0404240042,
-0.0199333802,
-0.0229269601,
0.0163876396,
0.0029637923,
0.0414891429,
0.0341079719,
-0.0442392193,
0.0729888529,
0.0048596272,
0.0053363401,
0.0355487056,
0.0694061518,
-0.0539329089,
-0.0172125604,
-0.0069040051,
0.0008118656,
-0.0063480651,
0.0337375365,
0.0486583598,
0.0133303059,
-0.0433262773,
0.0261865705,
0.0061374852,
0.0017825786,
-0.0249677617,
-0.0172769316,
0.0491155982,
0.0491836965,
0.0380847864,
0.0016297355,
-0.0075661275,
0.0048692259,
0.0530529246,
0.0397717468,
0.0231812224,
-0.000815239,
0.0377077945,
-0.0375148356,
-0.0100800088,
-0.0445648469,
-0.0061585819,
0.0027333805,
0.0295880791,
-0.0260968972,
-0.0235001352,
-0.0459534489,
0.0526115261,
0.0313695185,
0.0306881908,
-0.00104308,
-0.050914336,
0.0290830117,
-0.0063948887,
-0.0125508886,
-0.0198560786,
-0.0150180934,
0.0113896774,
-0.0361238495,
-0.0049607051,
0.0011436401,
-0.0499615669,
-0.0593376458,
0.0196773633,
-0.0731328353,
-0.0462492257,
0.0591920763,
-0.0476468727,
0.0144381737,
0.0022844134,
-0.0307371709,
-0.0245950855,
-0.0551984422,
-0.0335037746,
0.0357893594,
0.0385962427,
-0.0328302085,
0.0595991202,
-0.0299214628,
0.000539956,
-0.0031455222,
0.0021858132,
0.0073805982,
-0.0195787679,
-0.0029029527,
0.0349165723,
-0.0176553056,
0.011641169,
-0.0249400213,
0.0248948485,
-0.0766467154,
-0.0191931576,
-0.0647653788,
-0.0222789105,
-0.0335123986,
-0.0381769016,
0.0231497269,
0.0017067007,
-0.0012203635,
-0.0543685183,
-0.0417662524,
0.0381588116,
0.0203262102,
0.031752415,
0.0241144858,
-0.003718233,
0.0245310478,
-0.0257280674,
-0.0089597609,
0.0079436135,
0.0995316878,
0.0102976374,
0.0062560472,
0.0009457456,
-0.0455624834,
-0.0476854667,
0.0314301513,
-0.0085917404,
-0.0131978942,
0.0355949774,
-0.0447223037,
0.0411658809,
0.0966690183,
0.0054674395,
0.0185494144,
-0.0026511909,
-0.002651104,
-0.0484722704,
0.0177110229,
0.0308862161,
-0.0284960233,
-0.0317418724,
-0.0404783674,
-0.035812486,
0.0354220942,
-0.0069961431,
0.0132406168,
-0.0383833349,
0.0674665794,
0.0240316018,
-0.0612490736,
-0.0729487315,
0.0368093252,
0.0252133571,
0.051176101,
-0.0549187586,
-0.0516488366,
0.0344151892,
-0.026158154,
-0.0351208709,
0.0046354658,
0.033970464,
-0.0177129451,
0.0797428936,
0.0314692184,
0.0500560962,
-0.0100476332,
0.0105499476,
-0.0480047986,
-0.0051757307,
-0.0047723912,
-0.014474608,
0.0382663459,
-0.0430416949,
-0.0021731281,
-0.0234552193,
-0.0403544977,
-0.0391400196,
-0.0244952124,
-0.0285325423,
-0.0202447604,
-0.0707188174,
0.0102325464,
-0.0223612674,
0.0312733687,
0.0374289304,
-0.0147360554,
-0.0025851771,
-0.0081703244,
-0.004297954,
0.0865106955,
0.0805734321,
0.0605143718,
-0.0004408924,
-0.0699181557,
0.0289342366,
0.042314589,
0.0147172119,
0.0166019276,
-0.0377979353,
0.0409990661,
0.0077960473,
0.050048463,
0.005326374,
0.0314372443,
0.0639810786,
-0.0364107005,
0.0068051796,
0.0712930262,
0.0169911664,
-0.0393813886,
-0.026965702,
-0.0121581294,
0.019353088,
0.0407929868,
0.0229758509,
-0.0815138146,
-0.0522730798,
-0.0279906671,
-0.0272107385,
0.0402096361,
0.0073646302,
-0.0017725674,
-0.0731816366,
0.0705627576,
-0.0214994922,
-0.0504934415,
-0.0146574285,
-0.0053641801,
-0.0203543883,
0.0042039505,
0.0265300479,
0.000293974,
-0.001762433,
-0.0014801695,
-0.0079088947,
0.0299604293,
-0.0122565841,
0.0301809404,
-0.0439367853,
0.002781813,
0.0108589446,
0.0162368659,
-0.0155625725,
-0.0716046989,
0.0117089823,
0.0023307463,
0.0076946672,
-0.0045038229,
0.0586582422,
-0.0055793794,
-0.0479143895,
-0.0075738612,
0.0076339189,
-0.0233020969,
0.0024666535,
0.0011757386,
0.0645551011,
-0.0061584543,
-0.0332192704,
-0.0468942784,
0.0080456994,
-0.0149076674,
0.0479120016,
0.055956997,
-0.0416427925,
-0.0760816187,
-0.0036597054,
-0.0115047954,
-0.0564225614,
-0.0163236652,
0.0272228364,
0.0337412581,
0.0042887428,
-0.0751824975,
-0.0387525149,
0.0332315713,
0.0486982614,
-0.0429026559,
-0.0369080976,
-0.0341176651,
0.037159171,
-0.0429786779,
0.0467999429,
-0.0288798772,
0.03915466,
-0.009320979,
0.0081287473,
0.0058322181,
-0.05174594,
-0.0234755576,
0.0238373335,
-0.1075465083,
-0.0084210709,
-0.0326745994,
0.0537053719,
0.0155929541,
0.0178262945,
0.0137481876,
-0.0468582697,
-0.0108031016,
0.0006226447,
0.0114357406,
0.0076437877,
-0.0073864483,
-0.0060808961,
0.0090943174,
0.013423821,
-0.0351219587,
0.0134825679,
-0.0109033724,
0.0277449749,
0.0009180926,
-0.0126390411,
0.0168232825,
0.0248804316,
-0.0166187417,
-0.0012634373,
-0.0049687494,
0.0076736049,
0.0100476211,
0.0131393587,
-0.0095647825,
-0.0461535789,
-0.0064912452,
0.0212731361,
0.005265682,
0.0082140742,
0.0005204743,
0.0165932998,
0.0477294661,
-0.0339282602,
-0.0056624645,
0.0259487424,
-0.0217133723,
-0.0214264076,
0.0107789915,
-0.0027027105,
0.0187874585,
-0.0213228352,
0.0546623059,
0.0205096193,
0.0836261883,
0.0300560836,
-0.0466032885,
-0.0337756313,
0.0316299349,
-0.0277081188,
-0.0604528226,
0.005144407,
0.0524732023,
0.0623135865,
0.0107488083,
0.0072763343,
0.0235744733,
-0.025471108,
-0.0066676084,
0.0013594634,
0.0356204621,
-0.0301859528,
0.0742233172,
0.0440085828,
0.0382293537,
-0.0165294502,
0.0531067066,
0.0068429466,
0.0297736488,
-0.0053276964,
0.0163192581,
-0.0398606807,
-0.0290065575,
0.0092282929,
0.0139655545,
0.0133489305,
0.0264139082,
-0.0364343077,
-0.0732566938,
-0.0229427014,
0.052641362,
-0.0089566326,
0.0181917027,
0.0179169104,
0.0679899529,
0.027217377,
-0.0324922651,
-0.0133236889,
0.0079634115,
-0.023409117,
0.0004982171,
-0.0071442653,
-6.076913105e-33,
-0.0121224076,
-0.0299930181,
0.0427121632,
-0.0394982547,
-0.0269183647,
0.0054730577,
-0.0359198451,
0.0033345586,
-0.0492203571,
-0.0022495512,
0.0060420851,
-0.0199693702,
0.0085379016,
0.0428405739,
-0.0211500376,
-0.0211274959,
0.0044415141,
-0.0224476662,
-0.0392676517,
-0.0005926064,
-0.0087174783,
0.0054707578,
0.013590944,
-0.0193574745,
-0.0306921881,
-0.012507665,
-0.0272671189,
0.0289141629,
-0.0242962036,
0.0861272141,
-0.0543280914,
0.0105435094,
0.0446935147,
-0.0110272048,
-0.0281743892,
0.0336168185,
-0.0728117377,
0.0131726013,
0.019369131,
0.0439989604,
-0.0956566781,
0.0092857378,
0.0345024727,
-0.0301555227,
0.0329262018,
-0.0401456319,
0.0606363863,
-0.0087865945,
0.0035141255,
0.0242178682,
-0.0352144279,
-0.0025345257,
-0.0446561873,
-0.0265310314,
0.0092861522,
0.0720605403,
0.0258519799,
0.0636608377,
-0.0219039805,
0.0111535685,
-0.0073003252,
-0.0165333003,
0.093226485,
0.0348232761,
0.0182039756,
0.0579101928,
0.0511566363,
-0.0101924008,
-0.0342327319,
-0.0202796403,
0.0305209421,
0.0790227875,
0.0441919267,
0.0626907796,
0.0268966034,
0.0380792283,
-0.0391595289,
0.0221999101,
-0.0299998447,
0.0364393406,
0.0089551555,
-0.0002401089,
0.0023356592,
0.0329944566,
-0.0446859747,
0.0354741663,
-0.0658492669,
-0.0060183699,
0.0600236878,
-0.0026419081,
-0.01258393,
0.007956299,
0.0746500567,
-0.0514875762,
0.0124970218,
0.0144457594,
-0.0218059346,
0.0581081063,
0.0163511168,
0.0358998775,
0.0123945745,
-0.0039779153,
0.0416804962,
0.0132258432,
-0.0012221509,
0.0062130513,
0.0454550274,
-0.0036660233,
-0.0373640507,
-0.0221174657,
-0.0536053777,
0.0004506343,
-0.021508202,
0.0133032138,
0.0292219836,
-0.0027184431,
-0.0706760287,
-0.0017034799,
-0.0139119318,
-0.0184995271,
-0.0210327897,
-0.049488429,
-0.0207354929,
0.0038034478,
-0.0403692536,
-0.0038820242,
-0.0681338459,
-0.0585162938,
-0.0051546725,
-0.085309647,
0.0006151358,
0.0627488866,
2.471e-7,
0.026862422,
0.0239557065,
0.0130428169,
0.0373764001,
0.0499893725,
0.0296299718,
-0.0295073185,
0.0272938069,
-0.0144905727,
-0.0678603724,
-0.0342998691,
0.0820418373,
0.0325696468,
0.0095034745,
-0.0098280618,
0.0511631966,
-0.0278063621,
-0.0069641653,
-0.0715894103,
0.0075159036,
0.0858302638,
0.0543988496,
-0.0355912931,
-0.0430321395,
0.0400502272,
-0.0253181607,
0.0704936087,
-0.0066448939,
-0.0044474611,
-0.0262395758,
0.0105831372,
0.0196711663,
0.0198694766,
0.0184003897,
0.0191180389,
-0.0349242948,
0.0519380718,
0.0388766937,
0.0073859757,
-0.0518571138,
-0.0156580415,
-0.0056269998,
-0.0162818041,
-0.0751060694,
0.0920330584,
-0.0270768069,
-0.0333871879,
-0.0255561005,
-0.0445254296,
-0.0164834838,
0.0299540423,
-0.0276339408,
-0.0226246994,
0.0294761248,
0.016022427,
-0.0397150777,
-0.007932283,
-0.0668750256,
-0.0212620106,
-0.0088336617,
-0.0546794981,
-0.015772732,
-0.0650028288,
-0.0204217155,
0.0457148701,
-0.0242824163,
-0.0571009852,
1.575627462e-34,
0.0353039205,
-0.0170783121,
0.0253354833,
0.0175600499,
0.0202166177,
-0.0263268948,
-0.0484054275,
-0.0350881144,
-0.0126441335,
-0.0818553716,
0.0144479396
] |
http://arxiv.org/abs/2401.15840v1 | http://arxiv.org/pdf/2401.15840v1 | Emergent Explainability: Adding a causal chain to neural network inference | 2024-01-29T02:28:39 | 2024-01-29T02:28:39 | Adam Perrett | cs.CL | cs.CL | This position paper presents a theoretical framework for enhancing
explainable artificial intelligence (xAI) through emergent communication
(EmCom), focusing on creating a causal understanding of AI model outputs. We
explore the novel integration of EmCom into AI systems, offering a paradigm
shift from conventional associative relationships between inputs and outputs to
a more nuanced, causal interpretation. The framework aims to revolutionize how
AI processes are understood, making them more transparent and interpretable.
While the initial application of this model is demonstrated on synthetic data,
the implications of this research extend beyond these simple applications. This
general approach has the potential to redefine interactions with AI across
multiple domains, fostering trust and informed decision-making in healthcare
and in various sectors where AI's decision-making processes are critical. The
paper discusses the theoretical underpinnings of this approach, its potential
broad applications, and its alignment with the growing need for responsible and
transparent AI systems in an increasingly digital world. | Emergent Explainability: Adding a causal chain to neural network inference: This position paper presents a theoretical framework for enhancing
explainable artificial intelligence (xAI) through emergent communication
(EmCom), focusing on creating a causal understanding of AI model outputs. We
explore the novel integration of EmCom into AI systems, offering a paradigm
shift from conventional associative relationships between inputs and outputs to
a more nuanced, causal interpretation. The framework aims to revolutionize how
AI processes are understood, making them more transparent and interpretable.
While the initial application of this model is demonstrated on synthetic data,
the implications of this research extend beyond these simple applications. This
general approach has the potential to redefine interactions with AI across
multiple domains, fostering trust and informed decision-making in healthcare
and in various sectors where AI's decision-making processes are critical. The
paper discusses the theoretical underpinnings of this approach, its potential
broad applications, and its alignment with the growing need for responsible and
transparent AI systems in an increasingly digital world. | [
0.0106808329,
0.0773203373,
-0.0436966866,
-0.018665323,
0.0134560056,
-0.0019992162,
0.0200345758,
0.012194315,
0.0067162928,
0.0564487986,
0.0499344952,
0.0507493988,
-0.0034945332,
0.0814532638,
0.0202326328,
-0.0227686409,
-0.0077979355,
0.0072209793,
0.0050871898,
0.0301636197,
-0.0306941941,
-0.002744324,
-0.0011326297,
0.0645255074,
-0.0203913394,
-0.0161531046,
-0.0262663532,
-0.0087393383,
0.021852972,
-0.0087482007,
-0.0506569594,
-0.0491186008,
0.03952365,
-0.0000162473,
0.0000022698,
-0.0221140664,
-0.004509083,
0.0256305374,
-0.1093745679,
-0.0409187116,
-0.0461523719,
0.0088725751,
0.0197720453,
0.0013648465,
-0.0078688646,
0.0420347527,
0.0666880384,
0.0219143257,
0.010497449,
0.0236569121,
-0.018238524,
0.0066033481,
0.0371470973,
0.0053434307,
-0.0196929742,
0.0005257138,
-0.0020757543,
-0.0226132143,
-0.0086030215,
-0.0459783711,
-0.0051196548,
0.0139094852,
0.0476307608,
-0.0235584639,
0.0631273463,
0.0092875306,
0.0776090994,
-0.0507338718,
-0.0201912709,
0.0450041108,
0.0187063571,
-0.0254663955,
0.0250366349,
-0.0336316228,
0.0234319586,
0.0216202214,
-0.0442936532,
-0.0229854602,
-0.002023902,
0.0209389012,
0.0151548618,
0.031174425,
0.0839174986,
0.0165697839,
-0.0261207279,
0.0549454018,
-0.0386567041,
0.0034291437,
-0.024816785,
0.0024579216,
0.0507196561,
-0.0327218734,
-0.0028661059,
0.0661659762,
-0.0622704774,
0.0053787897,
0.0253934879,
-0.0017080598,
-0.0315354243,
-0.0118345171,
0.0372800976,
0.0047234283,
0.015243426,
0.0031527863,
0.0173762888,
0.0264690518,
-0.0438519232,
0.0202998668,
-0.019681083,
0.0130661139,
-0.0605831519,
-0.0346918926,
0.0301118549,
-0.0029579932,
-0.0121979453,
0.018806912,
-0.0207028799,
0.0312821381,
-0.0749765858,
0.0160054248,
-0.0042177881,
0.0293352399,
-0.011410349,
-0.0151320752,
-0.0414911844,
-0.0257975478,
-0.0422878563,
0.0473386608,
0.01104178,
-0.044582963,
-0.0280418079,
0.0225228481,
-0.005118493,
0.018369101,
-0.0231982134,
0.0777098462,
0.0167019535,
-0.0149231935,
-0.0012652919,
-0.0679826811,
0.018276656,
-0.0302941911,
0.0372500829,
0.019292051,
-0.0112672225,
0.0008015178,
0.0178756453,
-0.0330706201,
-0.0109378649,
-0.05380008,
-0.0340077281,
0.0687842667,
-0.0472899154,
0.0211700145,
-0.0268600155,
0.0298945047,
0.0079992423,
-0.0065312381,
0.0593827665,
-0.0094654327,
0.0246772077,
-0.0305563789,
0.073782824,
-0.0355856493,
0.0542980395,
-0.0549089685,
0.008438169,
-0.0335842334,
-0.0763572976,
0.0216206051,
-0.0113644982,
-0.0306878574,
-0.0009850173,
0.0508941412,
0.0477817208,
0.0126004051,
0.0242883079,
0.0243510567,
0.0474947169,
-0.0443362407,
-0.0330506526,
0.0108797979,
-0.0448964834,
0.0382828526,
-0.0061450913,
-0.0056210263,
-0.032752417,
-0.0230010245,
-0.0306768063,
-0.0841580778,
0.0268785283,
-0.0463935323,
0.0264304187,
0.0262599029,
0.0359417535,
-0.0000329076,
0.007856953,
-0.0356595181,
0.0098703131,
-0.0325260423,
-0.0265646689,
-0.135373801,
0.0368445218,
0.0781195238,
0.0418154038,
-0.0253132228,
0.0129332384,
-0.0172728766,
-0.0350461453,
0.0151004242,
-0.0301036648,
-0.0113541922,
0.0320099629,
-0.0167837124,
-0.0255996697,
-0.0086325807,
0.0253833961,
0.0014950217,
-0.0231137481,
-0.023141779,
-0.0333348066,
0.0133945877,
-0.0464646891,
0.0226789806,
-0.0315295756,
-0.0267085675,
0.0002666544,
0.0103797792,
0.0580402724,
0.0496049263,
-0.0058269114,
0.0270634387,
0.0308366828,
-0.0105974125,
-0.0476536006,
0.0712242126,
0.0038347968,
0.0723375604,
-0.0473039038,
-0.053643696,
0.0261900686,
0.0206542201,
-0.0240800995,
-0.0502346754,
0.0421325937,
-0.029071698,
0.0258429572,
0.0061119609,
-0.0031727357,
-0.0016543313,
0.0017480211,
0.0003782391,
-0.0019975479,
-0.0021264029,
-0.0584906265,
-0.0112023167,
0.0092963185,
0.0111776646,
0.0050281947,
0.0640216097,
-0.0109321503,
-0.0145547697,
-0.0215955116,
0.0156859662,
-0.0343520567,
0.0051523014,
0.0208454076,
-0.0227977782,
-0.0338623784,
0.0041348571,
-0.030133741,
-0.0100159338,
-0.0253230315,
0.021529749,
-0.0182544179,
0.0091047538,
0.0610914417,
0.0407191105,
-0.0038314145,
0.0462749079,
0.0030386327,
-0.0458539128,
-0.0163545329,
0.0143038956,
-0.0282402635,
0.0526622161,
0.0648601353,
-0.003069001,
-0.0486800782,
-0.0207382347,
0.0062972633,
-0.0180388913,
-0.0394774787,
-0.0183256846,
0.030976899,
0.0315905027,
-0.0544650033,
0.0426332839,
-0.0346866287,
0.1069800332,
0.0184944645,
0.0151681211,
0.0005519365,
-0.0133097367,
-0.0593265258,
-0.0117574474,
-0.019888822,
-0.0520580336,
-0.0077940952,
0.0601098351,
0.0100907683,
-0.0090651736,
-0.0346471667,
0.04587432,
0.014473943,
-0.0366902053,
-0.0129284924,
-0.0916716009,
-0.0072747972,
0.0788008198,
-0.0098365555,
-0.0795812234,
0.0712282807,
0.0178952869,
0.0347120762,
0.0182655621,
-0.0480994508,
-0.0093792612,
0.0222197846,
-0.0082684401,
-0.0091974996,
0.0560122579,
0.0069291368,
-0.0240200628,
-0.0388866253,
-0.0078626443,
0.0004610138,
-0.0094044022,
-0.0568518713,
-0.0145867923,
0.0735240355,
-0.0253368709,
-0.0437032282,
0.0012228268,
-0.0174352471,
0.0371778682,
0.0115613155,
-0.0330443755,
-0.0142617794,
-0.0271264892,
0.0501408018,
0.0093087545,
0.001360688,
-0.0394488312,
-0.0225814264,
-0.0565394759,
-0.001257472,
0.0356979072,
0.0593888387,
-0.0266086794,
0.030132575,
-0.0186495949,
0.0131058367,
-0.0127501274,
-0.0382140055,
0.0030652904,
-0.0654467419,
0.0038925288,
0.0201010667,
0.0162032805,
0.0271713324,
-0.0343837366,
-0.0208266322,
0.0282186605,
0.0062524769,
-0.0229408219,
0.038855616,
-0.0121968314,
-0.0424594469,
-0.01803004,
0.017145317,
-0.0248420555,
0.0065846825,
0.0066791051,
0.0449747704,
-0.0662598684,
-0.0221890938,
0.0303030219,
-0.0086749829,
0.0655973479,
-0.0159088224,
-0.0394519642,
-0.0595446751,
-0.0612729453,
-0.0480004363,
-0.0223672241,
-0.0174019728,
-0.0220759623,
0.0541569404,
0.0036571815,
0.022625763,
-0.0132791754,
-0.0204340573,
-0.0137598049,
0.0022129549,
0.0537356772,
0.0801887512,
0.0493022576,
-0.0194074474,
-0.0079928497,
-0.0545108356,
0.0048203948,
0.0329619758,
0.013904647,
0.0647509545,
0.0540553182,
-0.018742295,
-0.0584322661,
0.0423721708,
-0.0171780754,
-0.0047515146,
-0.0078469096,
0.0611849874,
-0.0568250641,
-0.0177333839,
-0.0208945069,
-0.0122782039,
0.0443135314,
0.0509540737,
0.04403124,
-0.0295776017,
-0.0393029004,
-0.0037146083,
0.0316071771,
0.0075970744,
-0.0429144949,
0.009385339,
0.0345333703,
-0.072005108,
-0.0244211797,
-0.0450034961,
-0.0470456406,
0.0225793831,
0.0264098588,
-0.0163889769,
0.0315660164,
-0.004298693,
-0.0312443897,
0.04838356,
0.0533134192,
-0.0228603855,
0.0235862918,
-0.0231207032,
-0.0097480211,
-0.0786141455,
-0.0138857188,
0.0142335119,
0.0315887295,
0.0141452067,
0.0064376728,
0.0675299838,
0.0360706635,
0.0227043387,
-0.0124500049,
-0.0556999035,
0.0389775969,
0.0494107939,
-0.0149492398,
-0.0226878691,
0.0470934622,
0.0104281427,
-0.0063420436,
0.0263648834,
-0.0090689724,
0.0118320528,
0.0084896442,
-0.0361005254,
-0.0378586091,
0.0075104134,
0.0333961509,
-0.0065563531,
0.0270573888,
-0.0076787299,
-0.0226729717,
-0.0282656569,
0.0194503739,
-0.010924642,
0.030818833,
-0.0418558046,
0.0026156572,
-0.0027635219,
0.0110480748,
0.0188962556,
-0.0713276714,
-0.0258091744,
-0.0576736927,
-0.0249709412,
0.0227891598,
-0.0045236233,
0.0592634864,
-0.0290073659,
0.0280269142,
0.0373474173,
0.0505183823,
-0.0359077863,
0.0361678526,
0.0263828728,
0.0561488904,
0.0125836469,
0.012818614,
-0.0092640603,
0.0324659683,
-0.0974903256,
-0.0877324939,
0.0446212478,
0.0226339716,
0.0076268478,
0.0225957558,
0.0054290281,
-0.02177394,
0.0251311697,
0.0343778506,
0.0598009303,
0.0348843411,
0.0098208338,
-0.0197550338,
-0.002536285,
-0.0328973681,
-0.0174328964,
0.0594596043,
-0.0246084854,
0.0419597141,
-0.0718998015,
-0.0112091042,
-0.0122157112,
-0.0526599884,
0.0094592394,
0.0293720979,
-0.0251865853,
0.0164143462,
-0.0374837182,
-0.0485758409,
-0.0160622299,
0.0269832276,
-0.0143468603,
-0.047965657,
0.0387213379,
0.0490860045,
-0.0181089994,
-0.062278077,
-0.0015094889,
0.0169180557,
-0.0408857726,
0.0200779196,
-0.0057853539,
-6.246588001e-33,
0.0127602881,
-0.0646318346,
0.0143845016,
0.0003283406,
0.0203613732,
0.0264720041,
-0.0666723102,
-0.0101673556,
-0.0085036205,
0.054032933,
-0.0096899271,
0.0631143227,
0.0030049889,
0.023406256,
0.0069837808,
-0.038572289,
0.0321996324,
0.0095311329,
0.0285858214,
0.0614603013,
0.0270787105,
0.0101138232,
0.0021078885,
-0.1269730926,
0.0590332747,
0.0251721889,
-0.0223441236,
0.0130306361,
-0.0482542403,
0.0380008705,
-0.0092299618,
0.0205454025,
0.0235051196,
-0.0319129564,
0.0198412929,
0.0049538654,
-0.0116057228,
0.0020087715,
0.0187262371,
-0.023680741,
-0.0800519213,
-0.0659077689,
0.0520655438,
-0.0104328701,
-0.0186973251,
0.019634515,
0.0226022098,
-0.0438684449,
0.0174944028,
-0.0215133931,
-0.0233333185,
-0.0077073486,
-0.0676767007,
-0.0177126247,
0.0160804577,
0.028176127,
-0.0319196545,
0.007957181,
-0.0469601192,
-0.0073815333,
0.0423537754,
-0.0580261461,
0.0498958938,
0.0086169476,
0.0153856259,
0.0276062526,
0.0612484105,
-0.013164849,
-0.095741719,
-0.0209122021,
-0.0417198576,
0.0633949637,
0.0990843549,
0.0004667114,
-0.0304461624,
0.0003642374,
-0.1155880615,
0.0350462534,
-0.0103700114,
-0.0076879463,
0.023621304,
-0.0082717687,
-0.0733100474,
-0.0060505834,
-0.0006369826,
-0.089987956,
-0.0223672185,
0.0044194572,
-0.0154385725,
-0.0130627602,
-0.0297275204,
0.0177955944,
-0.002402443,
-0.0608798228,
0.0621173978,
-0.0159025881,
0.0651661381,
-0.0193062201,
0.0159071274,
-0.0479385257,
-0.0043504052,
-0.0157540999,
-0.0189646352,
0.035842929,
0.0896478817,
0.0245255418,
0.0010632476,
0.0103796851,
0.0310463142,
-0.0121475318,
-0.0449204072,
-0.0106016612,
-0.003595097,
-0.0697920769,
0.033894591,
0.0510240346,
-0.0169670824,
0.0340184234,
0.0253416002,
0.062169835,
-0.0021937641,
0.0031924543,
-0.0712370351,
0.0188810285,
-0.0316793807,
-0.0408721827,
-0.0427316204,
0.0362872742,
0.0343598537,
-0.0753145814,
-0.0243998803,
-0.0114865676,
2.952e-7,
0.0042105806,
-0.0037499706,
-0.0025977897,
0.0048425389,
0.0664214939,
0.0399236456,
0.0174621213,
-0.0012923068,
-0.024896469,
-0.0324431621,
0.0528617315,
0.0181987528,
0.0289756842,
-0.0106332814,
0.0257406887,
-0.0571893491,
-0.0542007871,
-0.0245592706,
-0.021520881,
0.0247988272,
0.0538587794,
0.0185972564,
0.0118167875,
-0.0144059351,
0.0563481636,
0.0378986225,
-0.0173970014,
-0.0407717116,
0.0682861134,
-0.0248420257,
0.0054641715,
-0.0040089837,
0.0025490159,
0.0396160185,
-0.014693968,
-0.0004430485,
-0.0377413407,
0.0483076684,
-0.0456853732,
-0.0011306838,
-0.0286222808,
-0.0283937417,
0.0494374894,
-0.0473821945,
0.0577877536,
-0.0531033613,
-0.0684201941,
0.0421224684,
0.023863392,
-0.0108290156,
0.0167083479,
-0.0001227437,
0.0236213319,
0.0028379308,
-0.036488153,
-0.0116826547,
0.0513913408,
-0.041998554,
-0.0019139127,
0.0122103347,
-0.0126010962,
-0.0180095918,
-0.0149225146,
-0.0118180541,
0.0721123591,
-0.0175653771,
0.0226266943,
2.481151742e-34,
-0.0095997909,
0.0237102527,
0.0153772701,
0.0054433555,
0.0485946834,
0.0104785729,
-0.0166128147,
0.0126929879,
0.0789272338,
-0.0436135717,
-0.0224130563
] |
http://arxiv.org/abs/2401.15798v1 | http://arxiv.org/pdf/2401.15798v1 | UnMASKed: Quantifying Gender Biases in Masked Language Models through Linguistically Informed Job Market Prompts | 2024-01-28T23:00:40 | 2024-01-28T23:00:40 | Iñigo Parra | cs.CL | cs.CL | Language models (LMs) have become pivotal in the realm of technological
advancements. While their capabilities are vast and transformative, they often
include societal biases encoded in the human-produced datasets used for their
training. This research delves into the inherent biases present in masked
language models (MLMs), with a specific focus on gender biases. This study
evaluated six prominent models: BERT, RoBERTa, DistilBERT, BERT-multilingual,
XLM-RoBERTa, and DistilBERT-multilingual. The methodology employed a novel
dataset, bifurcated into two subsets: one containing prompts that encouraged
models to generate subject pronouns in English, and the other requiring models
to return the probabilities of verbs, adverbs, and adjectives linked to the
prompts' gender pronouns. The analysis reveals stereotypical gender alignment
of all models, with multilingual variants showing comparatively reduced biases. | UnMASKed: Quantifying Gender Biases in Masked Language Models through Linguistically Informed Job Market Prompts: Language models (LMs) have become pivotal in the realm of technological
advancements. While their capabilities are vast and transformative, they often
include societal biases encoded in the human-produced datasets used for their
training. This research delves into the inherent biases present in masked
language models (MLMs), with a specific focus on gender biases. This study
evaluated six prominent models: BERT, RoBERTa, DistilBERT, BERT-multilingual,
XLM-RoBERTa, and DistilBERT-multilingual. The methodology employed a novel
dataset, bifurcated into two subsets: one containing prompts that encouraged
models to generate subject pronouns in English, and the other requiring models
to return the probabilities of verbs, adverbs, and adjectives linked to the
prompts' gender pronouns. The analysis reveals stereotypical gender alignment
of all models, with multilingual variants showing comparatively reduced biases. | [
0.0509679429,
0.0876419917,
-0.0517827347,
-0.003849684,
-0.0231459867,
0.0031333789,
0.0717148632,
-0.0262962133,
0.0136209857,
-0.0695586726,
0.0146051524,
-0.0305961557,
0.0162780713,
0.0141870659,
0.031885542,
-0.016956877,
0.0311044678,
0.0208663922,
-0.0251571685,
-0.0461704098,
-0.0185831971,
0.0624379069,
0.0090955431,
0.0873071998,
0.0234821402,
-0.0326513648,
0.049279213,
-0.0006063876,
-0.004947158,
0.0124358302,
0.0771303922,
0.0476637743,
-0.0056452164,
0.0227039009,
0.0000020183,
-0.0810665339,
0.0079081142,
-0.0031285209,
-0.0234291553,
0.0080075879,
0.0397926569,
-0.0066152262,
-0.0277903527,
0.0455772243,
-0.0408672243,
0.000454155,
0.047648564,
0.0054390687,
0.0184362493,
0.0128741758,
-0.0094459299,
0.0057681971,
-0.0734556541,
-0.0261059925,
-0.0463832915,
-0.0018839819,
0.0465401784,
-0.0675071329,
-0.0320699513,
-0.0193474106,
-0.0010297799,
0.0215232261,
-0.0433915928,
-0.0186422821,
-0.0258036181,
0.0243402719,
-0.0290309936,
-0.0561860353,
-0.0019520072,
0.0124371247,
0.0193043277,
-0.021098325,
0.0022168115,
-0.0131913936,
0.0377527773,
0.0389247127,
-0.010177779,
0.0319282003,
-0.0006522659,
-0.0046954835,
-0.0042265533,
-0.0400509872,
0.0036266698,
0.0368985087,
0.0006752815,
0.1374488026,
0.0003020102,
-0.0203514155,
0.0073981141,
-0.0305922292,
0.0462401509,
-0.0605053119,
0.0333241671,
-0.0226658154,
0.0086662285,
0.0246217176,
-0.0049656476,
-0.0239473879,
0.0176038463,
-0.0874145031,
-0.0092204297,
0.0307152662,
0.0064051389,
-0.0028053832,
-0.0194846336,
0.0335029736,
0.0132597852,
-0.0056414148,
-0.039338652,
-0.0181186777,
-0.0343298428,
-0.0398564525,
-0.0350956395,
-0.0186967142,
-0.0125953127,
0.0398590863,
-0.0598156787,
-0.0048578014,
-0.0507012345,
0.022224104,
-0.0132712806,
0.0098266592,
-0.0183113739,
0.0108020622,
-0.0307482313,
0.0097312778,
-0.022800127,
0.0324718393,
-0.0229629595,
-0.0767876357,
0.0437271781,
0.0023443911,
-0.0024139422,
-0.0596791208,
0.006718134,
0.023785647,
0.0137224691,
-0.0108035468,
-0.0297897924,
-0.0115136383,
-0.0232679117,
-0.0327813812,
0.0004622221,
-0.0547728613,
-0.0008943151,
0.0360906944,
0.0077164154,
-0.0559479743,
0.0148730017,
0.0270842556,
-0.037501052,
0.0381186642,
0.0230930634,
0.0147013878,
0.0820070729,
-0.0139414854,
-0.0190049149,
0.0903941095,
0.0322920382,
0.0279409774,
0.0367494375,
-0.0472599491,
0.0057338788,
0.0705514997,
0.0056735808,
0.0239755567,
-0.0382270738,
0.0105581339,
-0.0010873318,
-0.0030670173,
0.094753623,
-0.02833887,
0.0248590577,
0.0063082874,
0.0451025553,
0.0732506067,
0.0329784453,
0.0222192518,
0.0451268554,
0.0320498198,
-0.0179779772,
0.0069390964,
-0.0341024362,
-0.0311688911,
-0.0332309604,
-0.0184741914,
-0.0613358133,
0.0379366316,
-0.0143338861,
-0.0282753278,
-0.0187578369,
0.0160602406,
0.0422746725,
0.0314032845,
0.0285728257,
0.0099059995,
-0.0138799818,
0.055595234,
0.0368787348,
-0.0266293809,
-0.0339927338,
-0.0400372371,
-0.0397585742,
0.0620981306,
-0.0120031359,
0.0032667704,
-0.0053724442,
-0.0440785959,
-0.0199589804,
0.0669390336,
0.0589505471,
0.0229374357,
-0.0040398091,
-0.0285642631,
-0.0040250886,
0.0209872276,
0.0219941549,
0.0404358767,
0.0311130527,
0.0250498615,
0.0171211176,
-0.0180796236,
-0.047974553,
0.009997597,
-0.0340491198,
-0.0428849682,
0.0312180202,
-0.0081573315,
0.010650849,
0.0162497833,
-0.0205960721,
0.0275780763,
0.0022527403,
-0.0097036995,
-0.0016690898,
0.0465257056,
0.0164586026,
0.0389374532,
-0.0054647406,
-0.0323070176,
-0.0578443855,
0.0277414061,
0.0033281355,
0.0088685676,
0.0446691476,
-0.0038465022,
0.0260140896,
-0.0277237445,
0.003303139,
-0.0188955199,
0.0356940441,
-0.0127435764,
0.0065764487,
-0.0178194307,
0.0085768765,
0.0107093956,
-0.0236680135,
-0.0008753713,
-0.083067432,
-0.0059195915,
0.0308095012,
-0.0067511052,
0.0280689579,
0.0301621798,
-0.0113721984,
-0.0059415437,
-0.0315881744,
-0.0557383038,
-0.0295000169,
0.0152313765,
-0.0026054974,
0.000186544,
-0.0181810688,
-0.0539943203,
0.0055808313,
-0.0287913177,
-0.0458016209,
0.007900035,
0.0370923206,
0.0199791025,
-0.0320699811,
-0.0335359797,
-0.029927142,
-0.025021581,
-0.0383287854,
0.0340749994,
0.0306265578,
-0.018513022,
-0.0598812625,
-0.0136360349,
-0.0108009409,
-0.0001968408,
-0.0203294959,
-0.0223901793,
-0.0153355999,
0.0384716578,
0.0354415439,
0.0512110889,
0.0008604267,
0.0211908743,
0.0368165337,
0.008797219,
-0.0478949845,
-0.0202473607,
0.0154961105,
0.0166947562,
-0.0014940316,
-0.120033212,
0.0037284635,
0.0240718164,
0.0086289216,
-0.0269441381,
-0.0186608192,
0.0505071133,
-0.0580811836,
-0.0520419143,
-0.0481756292,
-0.0192904733,
-0.0241534933,
0.0835281312,
-0.012830507,
-0.0552745759,
0.0040736441,
-0.0171694066,
-0.0024593547,
-0.0307135768,
-0.052770853,
-0.0257740878,
0.0174833033,
-0.0322258547,
0.0316181667,
0.0304293446,
-0.0409240536,
-0.0562304519,
-0.0278173462,
0.0189942718,
-0.0348366611,
0.0056932955,
-0.0273854714,
0.0181991588,
0.0282980911,
-0.0177428219,
-0.0406268798,
-0.0564210936,
-0.0103833759,
0.0169145092,
0.0530643761,
0.0170269776,
-0.0305727925,
0.0152318096,
0.0007149789,
0.0229148511,
-0.0035782987,
0.0140362289,
-0.0185229667,
-0.0154969534,
0.0555795394,
0.0233693458,
0.0138644911,
-0.0317232907,
0.0015784262,
-0.0197640713,
-0.0038127995,
0.0358583443,
0.0079203621,
0.0154183302,
0.0541419871,
-0.0524266399,
-0.0226021204,
-0.0060681519,
-0.0064082919,
-0.0521613918,
0.0362373702,
0.0724456981,
0.0139988782,
-0.0070804101,
0.0100976946,
-0.0179273561,
0.0278534554,
0.0415105,
0.0452289432,
-0.0793667883,
0.0406917334,
0.0152600361,
0.0305493847,
-0.0459440276,
-0.0154094798,
-0.0449448191,
-0.1413555741,
0.0426565036,
-0.0096195042,
-0.0195233393,
-0.0079722321,
-0.043520987,
-0.023298718,
0.0007587816,
0.0302452501,
-0.0150062041,
-0.0171419699,
-0.00066388,
0.0244818479,
-0.061023239,
-0.0410990305,
0.0363245644,
-0.0603513904,
0.0340627618,
0.0498975143,
0.0384571105,
-0.0455310084,
-0.0237398203,
0.028234994,
0.006230528,
0.0013666513,
-0.0087021841,
0.0068445238,
-0.0433976538,
-0.015228332,
-0.0301353745,
0.0178208072,
0.0154455611,
0.0070180297,
0.0242269151,
0.0050106938,
0.0119560501,
-0.0362176746,
-0.0283981487,
0.0064989822,
-0.0118242512,
0.0866919309,
0.0019515882,
-0.0795807764,
-0.0074646208,
-0.0552216433,
0.0215105303,
0.0434306897,
-0.0519466661,
0.0247516967,
-0.0116103152,
-0.0380480997,
-0.0125097185,
-0.0841294676,
0.0531206615,
0.0075620152,
0.0285905041,
-0.0276632216,
0.0352092795,
-0.0042072991,
-0.0573102385,
0.0340002403,
-0.0105734998,
0.0571752861,
-0.0093851285,
-0.0328578725,
-0.0034535113,
-0.0420297533,
-0.0204113554,
0.0035136368,
-0.0181803871,
-0.0484938137,
0.0475873798,
0.019780349,
0.0305445809,
0.0315590352,
0.0068196617,
-0.0684455931,
0.0262480397,
0.0478974469,
-0.0220606122,
0.0487640314,
0.0315948986,
0.0476408079,
-0.0080432938,
0.0169648994,
-0.0426176637,
0.0298615023,
0.0021222318,
0.0018831188,
-0.0227412898,
-0.0536711626,
0.0291949175,
0.0141375773,
0.0021754529,
-0.0555776842,
0.0394636542,
-0.0587170981,
-0.0214857552,
-0.0190615207,
-0.0447651744,
-0.0132839931,
-0.0334056318,
0.048894994,
0.0069002691,
0.0149322208,
0.0116219781,
-0.0525693223,
0.0020247248,
-0.0275039021,
0.0400100388,
-0.0058018845,
0.0722221434,
-0.0066800364,
0.0021072121,
0.0090953056,
-0.0003856569,
-0.0204295237,
0.0327056907,
0.0514899753,
0.034390673,
-0.0036151018,
0.006544942,
-0.0042286874,
0.0159270372,
-0.0216095205,
-0.0451539494,
-0.0024047089,
0.021125868,
0.015132443,
-0.0602144562,
0.062573202,
-0.0050498196,
-0.0128169032,
0.0075059081,
-0.0039630048,
0.0200919751,
-0.0048471247,
0.0491460562,
-0.0058360095,
-0.0231391639,
-0.020835042,
0.0527176671,
0.0823082328,
0.0383654945,
-0.0264086928,
-0.0351830386,
-0.0281303655,
-0.0246022027,
0.0628028736,
-0.0025148096,
-0.0126471417,
0.0028119148,
-0.1118863598,
-0.0610653609,
-0.0121531859,
-0.0181082208,
0.0112499828,
-0.0343954004,
-0.0246249232,
0.071534887,
-0.0140883513,
-0.0560423508,
-0.0598535277,
0.0085096257,
0.0073363772,
-0.0049621873,
-0.0467369258,
-6.116987912e-33,
-0.0277620163,
-0.0633494183,
0.0045096353,
-0.0013792901,
0.0088390904,
-0.0008275571,
0.0203805361,
0.0116296997,
-0.0137603683,
0.0017312284,
-0.011674122,
0.0132167153,
0.0181307793,
-0.0064284508,
0.0212915726,
0.0421453342,
0.0440598838,
0.0140245957,
-0.0039692787,
-0.0583808646,
-0.0188239235,
0.0266106576,
0.0626325905,
-0.0546539947,
0.114645727,
-0.0075000017,
-0.0393080786,
-0.0016063037,
0.002770056,
0.0154102081,
-0.0121642603,
0.0476317294,
0.0112528838,
-0.0432698354,
0.0021057643,
0.0632644892,
-0.0386784375,
-0.0240793042,
0.0437476635,
0.0311172083,
0.0700873062,
-0.0361716822,
0.0385934822,
0.0045341142,
-0.0474498048,
0.0219167806,
0.0783376172,
-0.0031456007,
0.0465100296,
0.0423146896,
-0.0315239318,
0.0185281187,
-0.0474592149,
0.0289812684,
0.0162952151,
0.0831959546,
-0.010662565,
0.0029771924,
-0.1135815978,
0.0295116622,
0.0053047407,
-0.0043472699,
0.0765953213,
0.0225211959,
0.0492185801,
0.0288525019,
0.1020065919,
0.0388257578,
-0.0011697756,
-0.0392955765,
0.0524687208,
0.0110109244,
-0.0031266136,
0.037145786,
0.0104512824,
-0.0271650162,
-0.0180440024,
-0.0383875072,
0.0186445508,
-0.0227458365,
0.0279424582,
-0.0116463946,
-0.060169667,
-0.0780408382,
-0.0428679548,
-0.0550995022,
-0.0041083242,
0.0037546095,
0.0461384691,
-0.0163669717,
-0.0206674486,
0.018773606,
-0.0079334173,
-0.0574327782,
-0.0643146709,
-0.043646872,
-0.0313440822,
-0.0327109732,
0.0373396613,
0.0415042192,
-0.0672392547,
0.0306287948,
-0.0258820038,
0.0399412736,
0.0400390327,
0.0508457348,
-0.0380292162,
-0.0056375926,
-0.0842470452,
-0.009359099,
-0.0306054093,
0.0106673166,
-0.0102730477,
0.0153887998,
0.0371980667,
0.0091597661,
0.0174383353,
-0.0104662282,
0.0628163293,
0.0188170597,
-0.0208195914,
-0.0213841926,
-0.0197103582,
0.0566569604,
-0.0100416355,
0.0041821599,
-0.0026849757,
0.0450841151,
0.1044210047,
-0.0275621358,
-0.0033627311,
-0.0307171624,
2.599e-7,
0.0384087637,
0.0600753315,
-0.0166662224,
0.0459078439,
0.0180645175,
0.0409434922,
0.0084838867,
0.0263636261,
0.0941230878,
0.0079613477,
-0.0547210574,
-0.0123216845,
0.003903517,
0.0138287814,
-0.0732833296,
0.0135386484,
-0.0636260584,
0.0142406188,
-0.0402050838,
0.0328167677,
0.003750399,
0.0628244951,
0.0485417396,
-0.0047506643,
-0.0558744036,
-0.0312637463,
0.0253420118,
-0.0402973071,
0.0064740898,
0.0326426178,
0.0259035658,
-0.0003357607,
0.0150347063,
0.03661846,
-0.004482287,
0.0386499316,
0.0254499968,
0.0282167755,
-0.0182815101,
0.034297429,
0.0070927441,
-0.0289685912,
-0.051248271,
-0.0770933554,
0.0512347817,
0.0002009994,
-0.0382237583,
-0.0526622459,
-0.0440882854,
0.0115265436,
0.0649056733,
-0.0420573093,
-0.0335271508,
0.0090098772,
0.033028733,
-0.009155537,
-0.0113698989,
0.0009225188,
-0.0111088622,
-0.0205766689,
-0.0314838812,
-0.0616769381,
-0.00397676,
0.0438912585,
0.0071877767,
-0.0434544571,
0.0418280475,
2.438466374e-34,
0.0190557465,
0.047044497,
-0.0037808693,
0.0160883199,
0.0522447117,
0.0096368315,
0.0348867923,
0.0288659185,
0.0202744156,
-0.0132069821,
0.0027012683
] |
http://arxiv.org/abs/2401.15780v1 | http://arxiv.org/pdf/2401.15780v1 | Fine-Tuned Large Language Models for Symptom Recognition from Spanish Clinical Text | 2024-01-28T22:11:25 | 2024-01-28T22:11:25 | Mai A. Shaaban, Abbas Akkasi, Adnan Khan, Majid Komeili, Mohammad Yaqub | cs.CL | cs.CL, cs.LG | The accurate recognition of symptoms in clinical reports is significantly
important in the fields of healthcare and biomedical natural language
processing. These entities serve as essential building blocks for clinical
information extraction, enabling retrieval of critical medical insights from
vast amounts of textual data. Furthermore, the ability to identify and
categorize these entities is fundamental for developing advanced clinical
decision support systems, aiding healthcare professionals in diagnosis and
treatment planning. In this study, we participated in SympTEMIST, a shared task
on the detection of symptoms, signs and findings in Spanish medical documents.
We combine a set of large language models fine-tuned with the data released by
the organizers. | Fine-Tuned Large Language Models for Symptom Recognition from Spanish Clinical Text: The accurate recognition of symptoms in clinical reports is significantly
important in the fields of healthcare and biomedical natural language
processing. These entities serve as essential building blocks for clinical
information extraction, enabling retrieval of critical medical insights from
vast amounts of textual data. Furthermore, the ability to identify and
categorize these entities is fundamental for developing advanced clinical
decision support systems, aiding healthcare professionals in diagnosis and
treatment planning. In this study, we participated in SympTEMIST, a shared task
on the detection of symptoms, signs and findings in Spanish medical documents.
We combine a set of large language models fine-tuned with the data released by
the organizers. | [
0.028227929,
0.0395402312,
0.0028552152,
0.0159677193,
-0.0009087313,
0.0470722802,
-0.0126089631,
0.039546594,
0.0616799183,
-0.0483907945,
-0.0105706379,
0.0079133473,
-0.0269007795,
-0.0116013885,
-0.0125723844,
-0.0458215326,
0.0256515574,
-0.0030284508,
0.0397012308,
0.0006144659,
-0.0025925008,
0.0086844256,
-0.0044365823,
0.0303877555,
0.0294927265,
-0.0127438158,
0.071310088,
-0.0200711582,
0.038999036,
-0.0715900958,
0.0119141825,
0.0282585248,
-0.0365637243,
-0.0110749658,
0.0000018142,
-0.038547609,
-0.0502447709,
0.0127343172,
0.0189340059,
-0.0445200764,
0.061109148,
-0.0689858571,
-0.0139532005,
-0.0002055468,
-0.0583586358,
-0.0715378746,
0.0386831425,
0.0548357368,
0.0409966186,
0.0579618029,
-0.033019349,
-0.0361497886,
0.0087235831,
-0.0003684673,
0.0150071895,
-0.0106936675,
0.0674299151,
-0.0213806797,
0.0886657164,
-0.0433021002,
0.0013167551,
-0.0003277141,
-0.0223659426,
0.0312289391,
-0.0187681578,
0.0322882049,
-0.0341029428,
-0.002462216,
0.0267413761,
0.0397071093,
-0.0413397737,
-0.0026623548,
0.0342328884,
0.0070995684,
-0.0245677233,
0.0040958561,
-0.0124766715,
-0.0131439744,
-0.048121877,
0.0312049408,
0.0899538547,
0.025246704,
0.0294745155,
0.0240583364,
-0.0262412,
0.0456913188,
0.0022520109,
-0.0435034856,
0.0354721509,
-0.0003966167,
-0.0610314496,
-0.038701579,
0.0471613817,
0.021243114,
-0.0420772359,
-0.0003297371,
0.0093184793,
-0.0761782899,
0.0504363365,
-0.0899101198,
-0.0158342086,
0.005223278,
0.002598125,
0.0275937524,
-0.0070103677,
-0.0029749051,
-0.0093358522,
-0.0007816731,
-0.0445553958,
-0.0266793389,
-0.0321276076,
0.0288403295,
-0.089998126,
-0.0022917981,
0.0057554538,
-0.0076338681,
-0.0465283133,
-0.0107906675,
-0.006281734,
0.006090736,
0.0278489124,
0.010773601,
0.0124001941,
0.0235464573,
0.0235128775,
0.0355769433,
-0.0515252464,
-0.0029525282,
-0.0160574894,
-0.0409124158,
0.0460463427,
0.0010053143,
0.0055643162,
-0.0480728708,
0.0597857349,
0.0175787751,
0.0281720608,
-0.0188806579,
-0.0126692038,
-0.0368516296,
-0.0300804935,
-0.0331974849,
-0.0481598116,
-0.0887118652,
0.0223276298,
0.0245377198,
0.034568686,
-0.0271264538,
0.0099809216,
-0.0047695539,
-0.039953839,
0.0438974053,
-0.0810334459,
0.00464718,
0.0033657127,
0.0145856719,
0.0498094149,
0.0315423831,
0.0198565442,
0.0036616367,
0.0229084007,
-0.0434896648,
0.0150875691,
0.0174198169,
-0.0233943034,
-0.0270257946,
0.0099635795,
-0.0026122807,
0.0056838556,
0.0654487759,
-0.0379444659,
0.0509131849,
0.0141839786,
0.0138046984,
0.0236951839,
0.0782907754,
0.0401413925,
0.0091182068,
0.0087247742,
-0.0386151634,
0.020621784,
-0.0273342822,
-0.011899625,
-0.0369438641,
-0.0019151701,
0.0002427916,
-0.0163975656,
0.0466287248,
-0.0318155587,
-0.0546841212,
-0.0676097125,
0.0099909613,
-0.0580658987,
0.0018564898,
0.0589530356,
-0.0161297731,
-0.0118039194,
-0.0072993897,
0.0418750234,
-0.0434864722,
0.0054099755,
-0.0211538952,
0.0152943376,
0.0716261417,
0.0256186668,
-0.0332714766,
-0.0087731173,
0.0045160549,
0.0173740331,
-0.0262033902,
0.0308036488,
-0.0055009783,
-0.0104163866,
-0.0424871631,
0.0230392721,
0.0678745881,
0.0291035138,
-0.0523192734,
-0.0675669238,
0.0173105597,
-0.0388515592,
-0.0330952033,
-0.0518208779,
0.0431096628,
-0.0059724506,
-0.0545219406,
0.0033531203,
-0.0021534439,
0.0276877806,
0.0122076012,
-0.0070378883,
-0.004260608,
0.0473130122,
-0.0819887668,
-0.0189586207,
0.0421339087,
0.0017527036,
-0.0440899357,
-0.0073944153,
-0.028946938,
0.0072055943,
-0.0137097649,
-0.0423892215,
-0.0065958411,
0.0011813564,
-0.0482553691,
0.0210392457,
0.0050928565,
0.0265608244,
0.0156697612,
0.035374999,
-0.0387765132,
0.029923331,
-0.0206830204,
0.0664952844,
0.0013909788,
-0.0670578703,
-0.0165948384,
-0.0417171456,
-0.0582762174,
0.0414927006,
0.0158530604,
0.0129298363,
-0.0172162689,
-0.002711908,
0.0088234963,
-0.0361330882,
-0.1369079202,
0.0099306181,
0.0052266787,
0.0038334946,
0.0308670029,
0.0154294753,
-0.0273397472,
-0.0051522884,
-0.0395704061,
-0.0202501584,
0.0437028371,
0.0267062243,
0.0225520767,
0.0098282266,
0.0041982634,
-0.0583556332,
0.00059351,
-0.0678051487,
0.0217344984,
0.0311166476,
-0.0289280117,
-0.0357191041,
-0.0052393139,
-0.0301133282,
0.0263360776,
0.0514106639,
-0.0436627269,
-0.0286472421,
-0.0232120976,
-0.0070988256,
0.0805962682,
-0.0073854257,
0.0751313642,
0.0334278531,
0.0054023573,
0.0016885536,
0.000641892,
-0.0419497825,
-0.0000730125,
-0.0359976403,
-0.0648888946,
0.0275340639,
0.1210203469,
0.0321902037,
-0.01105446,
-0.0237500388,
0.0089138001,
-0.0424711928,
0.0518064201,
-0.0146642914,
0.0199761856,
0.0041459934,
0.0158522259,
-0.0104535585,
-0.0399274528,
0.0592369661,
-0.0375549681,
0.0228220671,
-0.0968719274,
-0.0531155393,
-0.0292538162,
-0.0251124911,
0.0075650043,
0.0371549539,
0.0215709545,
-0.0014885919,
0.0300370138,
0.0057017067,
0.0111794258,
-0.067632705,
0.0454646051,
-0.0097029889,
-0.0126472414,
0.0204539504,
-0.039153181,
-0.0337726362,
-0.0280037932,
-0.0407221541,
0.0224624351,
0.0311377198,
0.0165072419,
-0.0715460703,
0.0409209616,
-0.0505844057,
0.0512300506,
0.0296721496,
-0.0301007181,
0.0031050597,
-0.0015130346,
0.0224546827,
-0.008899427,
0.0184449293,
-0.048710987,
-0.0016469274,
0.0639608279,
-0.0084383637,
0.044775255,
-0.0065711709,
0.0308478475,
0.0154098859,
-0.0520276688,
-0.0174332801,
-0.0260398164,
-0.0147298705,
-0.0424784124,
-0.0072683725,
0.1259512603,
0.0247980021,
-0.066400364,
0.0169795156,
-0.0428689308,
0.0010600273,
-0.0177756064,
0.007027756,
-0.0324808657,
-0.0211117864,
0.0071106227,
0.0252719373,
-0.02059705,
-0.0356998183,
0.0247535277,
-0.0195587557,
0.0189765021,
-0.0065732971,
-0.0385069661,
-0.0345763676,
-0.0332433432,
0.0048078476,
-0.0262115728,
-0.028293971,
-0.031229632,
-0.0044656801,
0.0005994159,
0.0306627806,
-0.0481982864,
-0.0584484749,
0.0356853604,
-0.027139198,
0.0085760914,
0.0298911538,
0.0548384674,
0.0154467216,
-0.021271782,
-0.0183896665,
-0.0445705093,
-0.0042136442,
-0.0087865163,
0.0404233113,
0.0094399713,
-0.0143587356,
-0.0448570065,
0.0268697254,
-0.0899570957,
0.020393217,
0.0024246646,
0.1062173918,
-0.0159710385,
-0.0179472603,
-0.023874959,
0.0195165295,
0.0438812375,
0.0313025638,
-0.0113007864,
-0.0180743989,
-0.0166848227,
0.0577616021,
0.0092477724,
0.0237771478,
-0.0230533984,
0.0258945227,
-0.0299511272,
-0.0364238434,
-0.0702258199,
-0.022875119,
0.0271115024,
-0.0058240122,
0.0231602341,
0.0057784435,
-0.0062926658,
-0.0102906926,
-0.0664377064,
0.0643771589,
0.058838848,
0.028013926,
-0.0754553676,
-0.0158907846,
0.0153761506,
0.0494199209,
-0.0078852167,
0.0186188985,
-0.0768424943,
-0.01702342,
0.0141320582,
-0.0119470218,
0.0210355669,
-0.0219059847,
-0.0304239951,
-0.0465250574,
0.014864713,
0.0302389357,
-0.0272309408,
-0.0077509866,
0.0143897729,
-0.0364015028,
0.0083743976,
0.0288969893,
0.006550659,
0.0477694161,
0.0152066369,
0.0190219432,
0.0082656741,
-0.0286460724,
0.0168523993,
0.0228388347,
0.0547299609,
-0.0120425159,
-0.0657277927,
-0.0380628556,
-0.0281816553,
0.0068087364,
-0.0005464794,
0.0043484685,
-0.0044689723,
0.0162699129,
0.017532099,
-0.0287183821,
0.0404339582,
-0.0326366909,
-0.0440986231,
-0.0411963426,
0.0791465119,
0.029224677,
-0.0190467983,
0.0680826828,
0.0150966458,
0.0264695082,
0.0123334201,
-0.0014209809,
0.0139858238,
-0.0154150361,
0.0360808671,
-0.0074497615,
0.0445860028,
0.0017528216,
0.0690545589,
0.0154815763,
-0.0458527431,
-0.0406386405,
-0.0173104685,
0.0597570315,
-0.0429910272,
0.0547725074,
-0.0008724972,
0.0000296733,
-0.0345389992,
0.0060954671,
-0.021302728,
0.0221301224,
0.0558610521,
-0.0063751889,
-0.0104973568,
-0.0399816558,
0.0707388595,
-0.0358044766,
0.02350826,
0.0377174057,
-0.0180944242,
0.0063818996,
0.005276293,
0.0229061302,
-0.0140110534,
-0.0864823982,
-0.0164765175,
-0.0503069758,
-0.0424579978,
0.0424638428,
-0.0403608903,
-0.0322715193,
0.0157447532,
0.0257841777,
0.0632292181,
-0.0063855923,
-0.0043842997,
-0.0297024604,
0.0452276021,
-0.0454541408,
-0.037439581,
0.0044904756,
-5.210736134e-33,
-0.007355737,
-0.0390280969,
0.0046977727,
-0.0109768165,
0.0074478067,
0.0228402391,
-0.0229004733,
0.0036540786,
0.0118929567,
0.0371357501,
-0.0063563925,
-0.0012818979,
0.0188908521,
0.006677208,
0.0077443048,
-0.0048462804,
0.0375716053,
0.0163505487,
-0.0431016944,
0.0044542658,
0.0124147488,
0.0272666812,
0.0700976849,
-0.1046912894,
0.0183282495,
-0.0131324707,
0.0275563318,
-0.0118490178,
0.0340453424,
0.0691574365,
-0.0367469527,
0.0914550871,
0.0535184294,
0.0122512672,
0.0010237825,
0.0238935314,
-0.0344162881,
-0.0391119681,
0.0311907474,
-0.0024060721,
0.0081839813,
-0.0344827995,
0.031868238,
-0.019192921,
-0.0043973625,
0.0244041402,
0.0563961305,
-0.0718211383,
-0.001666186,
-0.0333530419,
-0.0494808033,
0.0509551167,
0.030057082,
0.0120058022,
0.0391326398,
0.0310928728,
-0.0418121032,
0.0374485627,
0.0107459109,
-0.0069076442,
0.019560378,
0.0458130203,
0.0162353497,
-0.0603004955,
-0.0010040591,
0.0521913618,
0.0377161764,
-0.0252013151,
-0.0027246426,
0.0606884435,
0.0033979442,
0.0313426703,
0.0431584939,
0.0514816754,
0.0348766074,
-0.0166206378,
-0.0371079817,
0.0165976044,
-0.0417165756,
0.0153312143,
0.0488387272,
-0.0109123299,
-0.0229995288,
-0.0092213014,
-0.0282245334,
0.0001106651,
-0.0320821144,
-0.0127543015,
0.0250054915,
-0.0043686833,
0.027657602,
-0.0002343534,
0.0045691188,
-0.0201507695,
-0.0306230523,
0.0125066498,
0.0161864664,
0.0298347119,
-0.0211991295,
0.0276047401,
-0.0700320452,
-0.0425658673,
-0.0056041637,
0.00361614,
0.0457644388,
0.0206495728,
0.0283765774,
-0.0438382514,
-0.04507,
0.0167712532,
0.0201257449,
0.0138689168,
0.0025203449,
0.0930863321,
-0.0054824678,
-0.0131561365,
0.0128605776,
0.0841415748,
0.0055514844,
-0.0056849299,
-0.0241846759,
-0.0177125204,
0.0156618804,
-0.0029260968,
-0.0170268677,
-0.006616286,
-0.0412403196,
0.0187079962,
0.0515292622,
-0.0299570113,
0.0125533463,
-0.0487830043,
2.448e-7,
0.0908498764,
0.0459005423,
-0.0351081081,
-0.0321840458,
0.0256400295,
-0.0040839212,
0.0210974719,
0.0460311398,
0.0033016,
0.0312741846,
0.001162649,
-0.0172187705,
0.0493599027,
0.0712911561,
-0.0463253707,
-0.0071454891,
-0.0474297702,
-0.0254670065,
-0.0410047062,
-0.0200248193,
0.0710647702,
0.1096340716,
-0.0288936105,
-0.021929726,
0.0247456022,
-0.0616134144,
0.038487535,
-0.0599423461,
0.0005458206,
0.0627292246,
0.0032641522,
0.0110219866,
0.0143270958,
0.0283934195,
-0.0059196264,
-0.0183444507,
0.0155171985,
0.0529957302,
-0.0310692508,
0.0438339449,
0.0329382271,
-0.0512066148,
-0.0032068808,
-0.0597783141,
0.050356824,
-0.0729505345,
-0.0297197383,
0.0136149153,
0.0285149757,
-0.0412686691,
0.0396873876,
-0.0506355837,
-0.0410244577,
0.0428711623,
0.0451755114,
0.002568457,
-0.0440702401,
-0.0736055896,
-0.0060165944,
0.043327868,
-0.0181692448,
-0.0364548638,
0.0373836234,
-0.0293774661,
-0.0058859158,
-0.0649610013,
0.0194135513,
2.135533444e-34,
0.0090224845,
0.0314388797,
-0.0292888544,
0.0192831047,
0.0268375017,
-0.017167706,
-0.0094651505,
0.0053211576,
-0.0031329805,
-0.0811149403,
0.0175599344
] |
http://arxiv.org/abs/2401.15777v1 | http://arxiv.org/pdf/2401.15777v1 | cantnlp@LT-EDI-2024: Automatic Detection of Anti-LGBTQ+ Hate Speech in Under-resourced Languages | 2024-01-28T21:58:04 | 2024-01-28T21:58:04 | Sidney G. -J. Wong, Matthew Durward | cs.CL | cs.CL | This paper describes our homophobia/transphobia in social media comments
detection system developed as part of the shared task at LT-EDI-2024. We took a
transformer-based approach to develop our multiclass classification model for
ten language conditions (English, Spanish, Gujarati, Hindi, Kannada, Malayalam,
Marathi, Tamil, Tulu, and Telugu). We introduced synthetic and organic
instances of script-switched language data during domain adaptation to mirror
the linguistic realities of social media language as seen in the labelled
training data. Our system ranked second for Gujarati and Telugu with varying
levels of performance for other language conditions. The results suggest
incorporating elements of paralinguistic behaviour such as script-switching may
improve the performance of language detection systems especially in the cases
of under-resourced languages conditions. | cantnlp@LT-EDI-2024: Automatic Detection of Anti-LGBTQ+ Hate Speech in Under-resourced Languages: This paper describes our homophobia/transphobia in social media comments
detection system developed as part of the shared task at LT-EDI-2024. We took a
transformer-based approach to develop our multiclass classification model for
ten language conditions (English, Spanish, Gujarati, Hindi, Kannada, Malayalam,
Marathi, Tamil, Tulu, and Telugu). We introduced synthetic and organic
instances of script-switched language data during domain adaptation to mirror
the linguistic realities of social media language as seen in the labelled
training data. Our system ranked second for Gujarati and Telugu with varying
levels of performance for other language conditions. The results suggest
incorporating elements of paralinguistic behaviour such as script-switching may
improve the performance of language detection systems especially in the cases
of under-resourced languages conditions. | [
0.071261175,
0.0301441997,
-0.0206746776,
0.0404066332,
-0.0020308299,
0.0432458036,
0.0036712911,
0.0458369665,
0.0301454794,
-0.0832768455,
-0.0154422577,
-0.0560562462,
-0.0108159259,
0.0206487551,
-0.0277136713,
-0.0357111841,
0.0116494652,
0.0558853336,
-0.0162968673,
-0.0153245097,
0.0247970279,
0.0307568908,
-0.0083844708,
0.011208537,
0.0004899551,
0.0386805348,
0.035293512,
0.0149590671,
0.0145580666,
0.0127182053,
0.0346054398,
0.0492511541,
-0.009990369,
0.0552445129,
0.0000018828,
-0.0609108061,
-0.0005715576,
0.0113359056,
0.0161382314,
-0.0489029698,
0.0656869784,
-0.067392841,
-0.0092965299,
0.0484746285,
-0.0156894028,
0.0014318142,
0.0912697464,
0.01044326,
0.0533438958,
0.101659067,
-0.024190817,
-0.052159369,
-0.024425732,
0.0354310349,
-0.027187176,
0.0136418678,
0.0339954533,
-0.0683897287,
-0.0048537436,
-0.0266841762,
-0.0474055298,
-0.0069980109,
-0.0049992129,
0.0246853568,
-0.0087137083,
-0.0500276089,
-0.0185784567,
-0.0537961125,
0.0488036498,
0.0171268508,
-0.0530856401,
-0.0091461865,
0.0229602102,
0.0073387539,
-0.0267421901,
-0.0019536824,
-0.0072172275,
0.0119930143,
-0.042663727,
0.0274400506,
0.015326797,
0.0017790432,
0.0343396626,
0.0619563274,
0.0142518859,
0.0266350601,
0.0014283336,
-0.0189807247,
0.0045772716,
-0.0082752667,
0.0161746591,
0.0064131208,
0.0528422892,
0.0160985589,
0.0204239935,
0.0487674698,
-0.0150159756,
-0.0722749233,
0.0047092517,
-0.0777663216,
0.0299186613,
0.0743003637,
-0.0288730916,
-0.0270374734,
-0.0241732039,
0.0467439666,
-0.0047397763,
-0.0088300668,
0.0279564541,
-0.0150712663,
-0.0033655015,
0.0232569724,
-0.004489495,
-0.0279691461,
0.0011152081,
0.0821021497,
-0.0575298667,
0.0432476327,
-0.0270021837,
-0.0254967269,
0.0152769154,
-0.0283917431,
-0.0151752587,
-0.014700627,
0.0214416627,
0.0398381203,
0.0145265516,
0.0076996312,
-0.0533458479,
-0.0677633509,
0.0403550752,
0.044998318,
0.001717388,
-0.0253933165,
0.011601218,
0.053262379,
0.0116824768,
-0.0455915183,
-0.0367583521,
0.0295293909,
-0.0586561896,
-0.0327243581,
0.0127506536,
-0.0579836145,
0.0294110924,
0.0332328565,
-0.0152001828,
-0.0397055559,
0.001429155,
0.004729141,
-0.036680419,
0.0861389711,
0.0315748341,
-0.0170141142,
0.0703115612,
0.0247189179,
0.0169325024,
0.0899196938,
0.0080723194,
0.0188440159,
0.0401880518,
-0.0311898049,
0.011287882,
0.0596337281,
0.006023048,
0.0326090306,
-0.0372406915,
0.0238498524,
-0.0316878036,
-0.0180818774,
-0.0175626408,
0.0107020168,
0.0245978236,
-0.0352457017,
-0.0067432337,
0.0254473295,
-0.0029101509,
0.0325977951,
0.0371220745,
0.0351485163,
-0.03434496,
-0.0050806375,
-0.0372062922,
-0.0317385346,
-0.0214827769,
-0.0391108468,
-0.0142333498,
0.0426914878,
-0.0212125499,
-0.0212696362,
-0.0350399204,
-0.0304553527,
0.0528638028,
-0.0600648783,
0.0500622131,
0.0206211992,
-0.0344157144,
0.031883318,
0.0668269768,
-0.0369252972,
0.0394597501,
-0.0605741926,
0.0012518796,
0.034461055,
-0.0114928456,
-0.0373439156,
0.0218326282,
-0.0520380512,
-0.0102446182,
0.0526076406,
0.0154785579,
0.0075771511,
-0.0076050479,
-0.0354911424,
-0.0041673426,
0.0799091235,
-0.0051851189,
0.0112925656,
0.0330733173,
0.0590632334,
0.0213694684,
0.0184696894,
-0.0205226373,
0.0079912189,
0.0028420358,
-0.0093145268,
0.0143209863,
0.0098044416,
0.052662842,
-0.0023083698,
-0.015851697,
-0.0810769796,
-0.0115338126,
-0.006934491,
-0.0484685749,
0.0407300442,
-0.0112912804,
-0.0198166668,
0.0000724325,
-0.068509303,
-0.0023438488,
0.1105096936,
0.0008329598,
-0.0005459922,
0.0885802433,
-0.0276910253,
0.0408772528,
-0.031929601,
0.0125083858,
-0.0720841885,
0.0364090502,
-0.0150047326,
0.0237871502,
0.0031110519,
0.0067981542,
-0.0237392671,
-0.0547109544,
0.0616267435,
0.0335683413,
-0.0090835989,
0.0038881891,
-0.028956838,
0.0329479612,
0.0019341403,
0.0086688446,
0.0015150529,
-0.0428841077,
-0.088107571,
-0.038676016,
-0.022136569,
-0.025631953,
0.0020806915,
0.0282602478,
-0.0229869764,
0.0133029437,
-0.0013512515,
-0.0023119636,
0.0540756509,
0.0313500613,
-0.0046923319,
-0.060320206,
-0.0041157366,
-0.0038926664,
-0.0083482731,
-0.0300148912,
0.0135231055,
0.0207007676,
-0.0626390129,
-0.0213623215,
-0.0035949936,
0.0051070456,
0.0193681363,
0.0150721734,
-0.0695277601,
0.0147022018,
0.0227706991,
0.0180730764,
0.0106659196,
0.0620402358,
0.043778453,
0.028071288,
0.0267185606,
-0.0075109601,
-0.0175272934,
-0.0021469051,
0.0122789145,
-0.0448274501,
-0.0375344455,
0.0088557824,
0.040517658,
0.0350666903,
-0.005850628,
-0.0874676257,
-0.0129082892,
-0.0859866068,
-0.0177368708,
0.0353721492,
-0.0319864601,
-0.0195523631,
0.0073879762,
-0.0016280666,
-0.0272302441,
0.0486047827,
0.0053003039,
0.0053588171,
-0.0809389949,
-0.0177438967,
0.0012953526,
0.0133288326,
0.0003182884,
0.0582560115,
-0.0412510261,
-0.0257275235,
-0.0023978443,
-0.0179769211,
-0.0048941509,
-0.0129982494,
0.0280075707,
-0.0226769038,
0.0235397629,
-0.0182367489,
-0.0665897354,
-0.06453906,
-0.058770936,
-0.0364044644,
-0.0537822731,
-0.0130069414,
-0.035540536,
0.0202504657,
0.0592764542,
-0.0441473685,
0.0084969504,
0.0001388651,
-0.0050200089,
0.0551312305,
-0.0184987243,
-0.0092556728,
-0.0188127421,
-0.0080169039,
0.0422684103,
-0.0285812486,
0.027891174,
0.0006369967,
0.0086375866,
-0.007593493,
-0.0039587985,
0.0348852649,
-0.0114336619,
0.0062417048,
-0.0568362661,
-0.0011237767,
0.005276408,
-0.0017193735,
0.0697521493,
0.0141210593,
-0.0392436534,
0.0002940474,
-0.0364337191,
-0.0111200325,
0.005934827,
0.0209176261,
-0.0920844227,
0.0018261502,
-0.0205581654,
0.0142856603,
-0.0436348803,
-0.0155310063,
0.0006725696,
-0.0494553298,
0.0078706704,
-0.0257637966,
-0.0382209122,
-0.07858859,
-0.0131063834,
0.0360909998,
-0.0194024127,
0.0086897025,
-0.0219626855,
-0.0112312455,
0.0007197064,
0.0366214961,
-0.0764026418,
-0.0164667554,
-0.0093890755,
-0.0081883147,
0.0047813496,
-0.0129955756,
0.0506340899,
-0.0805832073,
-0.0280086845,
0.042204652,
-0.0329438485,
-0.0224463362,
-0.0612671264,
0.0302652158,
0.0048997309,
-0.0527901575,
0.0103564328,
0.0043275501,
-0.0293934103,
0.0020510592,
0.0111596426,
0.0622109175,
0.008331486,
-0.039744813,
-0.0263347011,
0.0340304673,
-0.0075461227,
0.0540101305,
0.0055175778,
-0.0947281048,
0.0365783311,
-0.0660061315,
-0.0035745881,
-0.0257221498,
-0.0607510395,
-0.0258955918,
-0.0104126707,
-0.059193667,
-0.0337540098,
-0.0113398116,
0.0099487584,
0.0119779343,
0.0269748941,
-0.045566842,
-0.0222305655,
0.0056172637,
-0.0593802221,
-0.0391236693,
0.0208254363,
0.0393728949,
-0.026701238,
-0.0018951879,
0.0274682287,
-0.0139551507,
-0.0210918505,
-0.017845042,
-0.0060974532,
-0.0482220873,
0.0874021947,
0.0281842593,
0.019680718,
-0.0046747085,
-0.0250554495,
-0.0489717312,
0.0474578105,
0.0332044289,
-0.0078440905,
0.0114398692,
0.0733002722,
-0.0038360036,
-0.0389928147,
0.0437802859,
0.0024591694,
0.0163637027,
0.0220329706,
-0.0726955086,
-0.0685904175,
-0.0451183952,
-0.0277826376,
-0.0178181622,
0.0019880035,
-0.0276958384,
-0.0557030812,
-0.0223255549,
-0.0253981762,
0.0415343605,
-0.0122354645,
-0.0038955908,
-0.047983259,
0.0201015677,
0.0015171518,
-0.038647268,
0.0007562264,
-0.0513534211,
-0.019463107,
-0.0063039139,
0.0632964969,
0.0739570707,
0.0813414454,
0.0204002205,
0.0189375971,
-0.0282323565,
0.0013113085,
0.0092262123,
-0.0097114779,
-0.0110591715,
0.0327421129,
0.019251937,
0.0141370697,
0.0442026779,
-0.0106179602,
0.0536373295,
0.0149337454,
-0.0222704671,
-0.0503401794,
0.0001282562,
-0.0333072543,
0.054113593,
0.0015173688,
-0.0119993575,
-0.0146442801,
0.0229533575,
-0.0555392839,
-0.0164631885,
0.0316911601,
0.0033123065,
-0.0135229444,
0.0034162691,
0.0206352994,
0.0321916007,
0.0310224667,
0.0609061867,
-0.0182697158,
-0.0300150961,
-0.0007015676,
0.0025781344,
-0.0006331219,
0.0374634005,
0.0275492761,
-0.038096372,
-0.0420877524,
0.0048100716,
-0.0859475136,
0.0499316193,
-0.0239265598,
0.014949603,
0.0726235807,
0.0184761584,
0.0138659524,
-0.0223602671,
0.0004539887,
0.0348579586,
-0.0235794429,
0.0036023101,
-5.813816164e-33,
-0.0206577927,
-0.0012984567,
-0.0353145748,
0.033339303,
0.0120357107,
-0.0498460568,
0.0373662002,
0.0073551987,
-0.0108984839,
0.0028088102,
0.0012078019,
0.042145025,
0.0182558205,
0.0101497294,
0.0060687559,
-0.0250903256,
0.0404407941,
0.014351245,
-0.0221993867,
-0.0060781292,
0.0606445782,
0.0221969951,
0.0584168993,
-0.0816749185,
0.0097980099,
0.0057473844,
-0.0099691516,
-0.0295402408,
0.0270592012,
0.0279747322,
-0.0182364751,
0.0873817131,
0.0593196042,
-0.0068482771,
0.0364453085,
0.0772892982,
-0.0386798307,
0.0123747187,
0.0227932204,
0.086195685,
0.041923821,
-0.0072826669,
0.0358368233,
-0.0237169098,
-0.0406682566,
-0.0398351625,
0.029986186,
-0.0393156111,
0.0398311317,
-0.0308886673,
-0.0926133543,
0.0462969691,
-0.0083009619,
0.0196276214,
0.0036127747,
0.0920320824,
0.0354100242,
0.0011396097,
-0.0648445785,
0.013485794,
-0.0351161249,
-0.0099174622,
-0.0040652729,
0.0366862565,
-0.0269059222,
0.0083394777,
0.0088736508,
-0.0138928164,
-0.0014153688,
0.01506689,
0.039386794,
0.0039998833,
0.0047728238,
0.03187095,
0.0091143828,
-0.106403105,
-0.0307820067,
0.0116150612,
0.0762319788,
-0.0208646003,
0.0168850869,
-0.0172125623,
-0.0355944894,
-0.0630540177,
0.0150827821,
-0.0499958135,
0.0004804048,
-0.0355848894,
0.0102245174,
-0.0332180597,
-0.0411837511,
0.0322969966,
0.0146730496,
-0.0330620483,
-0.0211010743,
0.0282505807,
-0.0356306955,
-0.005882842,
0.0208651293,
0.0460956953,
-0.0151794367,
-0.0250799973,
0.0037142443,
0.0179308131,
0.0391611122,
-0.0046771136,
0.042646572,
0.0462494828,
-0.0916379318,
0.0120940562,
0.0068499148,
0.0528242141,
-0.0156017141,
0.1173996925,
-0.0241673756,
0.0471758917,
-0.0112131555,
0.0044899024,
0.0399383828,
-0.0004070915,
-0.0364398994,
0.0142422877,
-0.0324289128,
0.0413642749,
-0.0434159897,
-0.0491298288,
-0.0066901632,
0.0250811186,
0.0461044237,
-0.0963220075,
0.0189932249,
-0.0419547111,
2.746e-7,
0.0247354954,
0.0222823955,
-0.0131432535,
-0.0744168535,
0.0011443028,
0.0286049657,
0.0085468236,
0.0189571679,
0.0256293006,
0.0383720212,
-0.032044664,
0.0268424861,
0.0150277037,
0.0403964333,
-0.0299578886,
-0.063808836,
-0.0437993221,
0.0171621982,
-0.0455442443,
0.0361689553,
-0.0335788988,
0.0934457034,
0.0932597965,
-0.0550759174,
-0.0226281676,
-0.0026850302,
0.0128643103,
-0.0198125131,
0.0066688205,
-0.0293783955,
0.0142984428,
0.0073271021,
0.0132368309,
-0.0085801408,
-0.0108513338,
0.0218220763,
0.0147373183,
0.0363687538,
-0.0112715205,
0.0556831695,
0.019495111,
-0.012285593,
-0.04852698,
-0.0462309606,
0.0422632843,
-0.0010369448,
-0.0169739481,
-0.0326917134,
-0.0217877626,
-0.0023348855,
0.0143204182,
-0.0210848842,
-0.0327144489,
-0.0039199507,
0.0339413807,
-0.0383335799,
-0.0018411146,
-0.069889389,
0.0244582426,
0.0091822278,
-0.0206116233,
-0.0016653838,
-0.0010969213,
-0.004991414,
0.0510393679,
0.0130047901,
0.0529100411,
2.150969613e-34,
0.0618406273,
0.0143434592,
-0.0116898781,
-0.0466377437,
0.0158717707,
0.0165377092,
0.0189866535,
0.0291249938,
0.0462433733,
-0.0060142865,
0.0043472424
] |
http://arxiv.org/abs/2401.15770v3 | http://arxiv.org/pdf/2401.15770v3 | PILOT: Legal Case Outcome Prediction with Case Law | 2024-01-28T21:18:05 | 2024-04-13T01:59:17 | Lang Cao, Zifeng Wang, Cao Xiao, Jimeng Sun | cs.CL | cs.CL | Machine learning shows promise in predicting the outcome of legal cases, but
most research has concentrated on civil law cases rather than case law systems.
We identified two unique challenges in making legal case outcome predictions
with case law. First, it is crucial to identify relevant precedent cases that
serve as fundamental evidence for judges during decision-making. Second, it is
necessary to consider the evolution of legal principles over time, as early
cases may adhere to different legal contexts. In this paper, we proposed a new
framework named PILOT (PredictIng Legal case OuTcome) for case outcome
prediction. It comprises two modules for relevant case retrieval and temporal
pattern handling, respectively. To benchmark the performance of existing legal
case outcome prediction models, we curated a dataset from a large-scale case
law database. We demonstrate the importance of accurately identifying precedent
cases and mitigating the temporal shift when making predictions for case law,
as our method shows a significant improvement over the prior methods that focus
on civil law case outcome predictions. | PILOT: Legal Case Outcome Prediction with Case Law: Machine learning shows promise in predicting the outcome of legal cases, but
most research has concentrated on civil law cases rather than case law systems.
We identified two unique challenges in making legal case outcome predictions
with case law. First, it is crucial to identify relevant precedent cases that
serve as fundamental evidence for judges during decision-making. Second, it is
necessary to consider the evolution of legal principles over time, as early
cases may adhere to different legal contexts. In this paper, we proposed a new
framework named PILOT (PredictIng Legal case OuTcome) for case outcome
prediction. It comprises two modules for relevant case retrieval and temporal
pattern handling, respectively. To benchmark the performance of existing legal
case outcome prediction models, we curated a dataset from a large-scale case
law database. We demonstrate the importance of accurately identifying precedent
cases and mitigating the temporal shift when making predictions for case law,
as our method shows a significant improvement over the prior methods that focus
on civil law case outcome predictions. | [
0.0376414508,
0.0837889984,
-0.0360568017,
-0.0310761072,
-0.0633966401,
-0.002819798,
0.011829759,
0.0344984382,
-0.0203293525,
-0.041565612,
0.036358647,
-0.0009114998,
-0.0476076826,
0.0477175564,
-0.0420405231,
-0.0491496958,
0.0318702422,
0.0376507342,
0.0342426337,
0.0302510168,
0.0150584988,
-0.0071041547,
0.0252426006,
0.0380038656,
-0.0463105142,
-0.0188519619,
-0.0163215492,
-0.0077969497,
-0.0358456112,
-0.0316337496,
0.0462841913,
0.0466205105,
0.0309733823,
-0.00093575,
0.0000016091,
-0.067031078,
0.0017892376,
-0.041293256,
-0.00141693,
-0.0051647965,
-0.0206364486,
-0.0186022669,
-0.0115294764,
0.0128898742,
-0.0677269697,
0.0124080116,
-0.0144877955,
0.0127690397,
-0.0054515591,
0.0513180643,
-0.0043411837,
-0.0464030392,
0.0580394752,
-0.0302451178,
0.0236440599,
-0.0158120878,
0.0242133364,
0.0136003336,
0.0174713861,
-0.0568132363,
0.0337844789,
-0.0193161201,
0.0342182405,
0.0130998883,
-0.0411054604,
0.02366478,
-0.0206572805,
0.0023635877,
-0.0106871249,
0.0284318551,
0.0536251068,
-0.0711764023,
0.0074242763,
-0.0029907236,
0.0533741117,
0.0022463144,
-0.0219413675,
0.0561185889,
-0.0254646819,
0.037644431,
0.0331865847,
0.0539455302,
-0.0097160069,
0.0364162996,
-0.0122148357,
0.0788702667,
-0.0113726715,
-0.0268333759,
-0.015326811,
0.0055272384,
0.0824496448,
-0.0353949964,
0.0509982221,
0.0216682293,
-0.0066948226,
0.0139500946,
-0.0071326368,
0.0184909664,
0.0677569434,
-0.0751394928,
-0.0049572396,
0.0071854978,
-0.0173413064,
0.0688628107,
-0.0068860734,
-0.0403901301,
0.0073082624,
-0.006678496,
0.0129934801,
-0.0095835431,
-0.0493574627,
0.0324764065,
0.0076003,
0.035440091,
-0.0743066818,
0.033613164,
0.0000078739,
0.0199529566,
-0.0219727866,
0.0055549238,
0.0196311325,
0.0287287794,
-0.0050613955,
-0.0271249637,
-0.0600107759,
-0.0636864677,
-0.023036018,
0.0572403185,
-0.0022235794,
0.0148790143,
0.0268743485,
-0.027308153,
0.0029550316,
-0.0017025019,
0.0344972722,
0.0939726532,
0.0225617383,
0.0031287172,
-0.071645461,
-0.0636030808,
0.0038327854,
-0.040868111,
-0.0365193039,
-0.0493348762,
-0.0117096603,
-0.0238002911,
-0.0292858761,
-0.0549164191,
0.0133028599,
0.0363746621,
-0.0043849126,
0.0113618588,
0.0346751586,
-0.0101858675,
-0.0116754435,
0.0264688525,
0.013474456,
-0.0254498832,
-0.0339527763,
0.0352653861,
0.0108131506,
-0.0163080394,
0.0225409959,
0.0246981103,
0.0372459851,
0.0196131133,
-0.025967015,
0.0279353075,
-0.0120450659,
0.0157553181,
-0.0093536228,
0.0118894065,
0.0338194631,
-0.0699565262,
0.0053480854,
0.0104653277,
0.019269336,
0.0748052523,
-0.0013112666,
-0.0282374732,
0.0180823579,
-0.0294290092,
-0.037442781,
0.0031362367,
-0.0476091467,
0.0104498994,
-0.0930806771,
0.000074704,
-0.0424789265,
-0.0474606045,
0.0314902961,
-0.0073326663,
0.0166797843,
0.0156602431,
0.0405264609,
-0.0126724066,
-0.0407298394,
0.0049252831,
0.0455611981,
0.0167507753,
-0.0428208821,
-0.0496479794,
0.0418179892,
-0.0089778388,
0.0352457613,
-0.0470885411,
-0.0111962054,
-0.0114094894,
-0.0196194332,
-0.0283817649,
0.0183252487,
0.0034280499,
-0.0048565893,
0.013753009,
-0.0280480273,
0.0606905483,
0.0293915328,
0.0082772383,
0.0283886157,
0.0480890386,
0.0177706853,
-0.0344647132,
0.0061839358,
0.0376136042,
-0.037911579,
-0.0177122802,
0.0370597728,
-0.0159769524,
0.0385519378,
0.0106528318,
0.0117428945,
-0.0210698787,
0.0365330465,
0.0252921246,
-0.0136738839,
0.0287569687,
0.0009160468,
-0.0014669105,
0.040013805,
0.0042629046,
-0.013620792,
-0.0102900257,
-0.0446364433,
0.0176982526,
0.0055577122,
0.0161423683,
0.0197482239,
-0.0471344143,
-0.0455215946,
-0.0076117874,
-0.0246834457,
-0.0094624329,
0.0331356004,
-0.0014372261,
-0.059135031,
-0.0378908105,
-0.0246027634,
0.031429626,
-0.0044247163,
-0.0024219872,
-0.0496376418,
-0.0594462007,
0.0305196233,
0.0179631468,
-0.0418372825,
-0.0759033933,
0.0076157739,
-0.0476522595,
0.0020688772,
-0.034299884,
-0.014402993,
-0.0278373249,
-0.0282407627,
-0.0208188891,
-0.0641378388,
-0.0170767251,
-0.0368886553,
0.0005858312,
-0.0120054521,
0.0145203164,
0.0115242815,
0.0099077327,
-0.0234457459,
-0.0231843013,
-0.0282387491,
0.070118539,
-0.0065313908,
0.0166903213,
0.0069491859,
0.0886615515,
-0.0366153903,
0.0021850944,
-0.0493886843,
0.0332797691,
-0.0389042608,
-0.0040708375,
-0.0513704531,
0.0136199165,
0.0504231453,
0.0081074974,
0.0039342609,
0.0084686177,
-0.0667938888,
0.017122874,
-0.0241749436,
0.0266809613,
-0.0166982859,
0.0154062351,
0.0146506084,
-0.0181152094,
0.026483912,
0.0028353818,
-0.0094361193,
0.0205950607,
0.0193083957,
-0.0203158669,
-0.0310490653,
-0.0025509226,
-0.0006245003,
0.0525514446,
0.0048156739,
-0.0967748687,
0.0688541159,
0.0652934313,
0.0625034794,
-0.0111428909,
-0.0211322065,
-0.0566913299,
0.102387324,
0.0194320343,
-0.0294069089,
-0.039100308,
-0.0334595293,
0.0166546106,
-0.0292772483,
0.0360055938,
-0.0515147932,
0.0793904811,
-0.0028037457,
0.0243021436,
0.0044389553,
-0.0581426322,
-0.0420521684,
-0.0543842912,
0.0108620916,
0.0322067104,
-0.0555472784,
0.0425767116,
-0.021698419,
0.0032049238,
-0.0496669933,
0.0067698718,
0.0289806128,
-0.026544163,
0.0010627992,
0.0216618143,
0.0052494179,
-0.0448420569,
-0.0239303205,
-0.0115634575,
-0.0130732013,
-0.0107663525,
0.0171693899,
-0.0002224015,
-0.029246686,
0.0204621833,
-0.000093645,
-0.0670725331,
-0.0017226592,
0.0673942417,
-0.0376068763,
-0.0689757913,
0.0484461486,
0.0742874593,
0.016593013,
-0.0369464159,
-0.0075077801,
0.0090899486,
-0.0062269201,
0.0367921479,
0.0392093472,
-0.0813289136,
0.0116896983,
-0.0066666752,
-0.0321335532,
-0.0255059581,
-0.0486135855,
0.0017357727,
-0.0509622321,
0.0859770328,
0.0361130945,
0.009450445,
-0.0449436232,
-0.0486990698,
-0.0062939851,
-0.0420057364,
-0.0383663587,
0.0115244184,
-0.0520122573,
0.0020375685,
0.014737187,
-0.0890445262,
0.0141719636,
0.0727909505,
-0.0456868336,
0.0358914509,
0.0238063913,
0.0030534002,
-0.0451848134,
0.0051019648,
-0.0338315219,
-0.0049282326,
0.0337728113,
0.0073881689,
0.0528680496,
0.0150976041,
0.0280383751,
-0.0573037416,
0.0389103144,
0.0101240724,
-0.0467147678,
0.0661948472,
0.0679442659,
-0.0749895722,
-0.0463811271,
-0.0394999348,
-0.0311744809,
-0.0132475467,
0.0335485302,
-0.0289198197,
-0.000792573,
-0.0803475454,
0.0168556888,
-0.0080716237,
-0.0068407324,
-0.0041616037,
-0.043592684,
-0.02065829,
0.0430434942,
-0.0047947848,
-0.0200603642,
-0.0132541908,
0.0315976776,
-0.007488803,
0.0380710848,
0.0066539948,
0.0064011216,
-0.0855539143,
0.0415293202,
0.0088244602,
0.0257298015,
-0.0155477691,
0.0460573174,
0.0292504672,
-0.0044556279,
-0.0197191872,
0.0154353315,
0.0122563681,
0.0066109579,
-0.0112468032,
0.011714736,
0.0171771795,
0.0237815026,
-0.0127165075,
-0.0118859839,
0.058866825,
0.0282061864,
0.033411622,
-0.0405385196,
0.062977232,
-0.0450123064,
-0.0334368646,
-0.0165779218,
0.0124966512,
0.0507104173,
-0.0061197402,
-0.015213958,
-0.0408205837,
0.0241180696,
0.0090044821,
0.0067121093,
0.0747089311,
0.0143239861,
-0.0651906058,
-0.0139649445,
0.0071061384,
0.0037597902,
0.0195834227,
-0.0111843757,
-0.0191615596,
0.0229925495,
0.0318366475,
-0.0643965974,
0.0541633703,
-0.007763519,
0.0084457342,
-0.0721924827,
0.0366112404,
0.029773308,
0.0281054545,
0.0294202678,
-0.0072245691,
0.0194890201,
0.0032496976,
0.0064588822,
0.0557724908,
-0.0124435658,
0.0234770216,
-0.0139212226,
0.0637483299,
-0.0167775881,
-0.0264397431,
-0.0337340198,
0.0008596828,
-0.0067096138,
0.0144008491,
0.0498197116,
-0.0312646627,
-0.0312979817,
-0.0197359901,
0.0060648387,
-0.0222679339,
0.040575549,
-0.023362698,
-0.0014723714,
0.0219656732,
0.0111879902,
-0.0395218618,
0.0277565848,
0.0080024945,
0.0126958601,
0.0353439413,
-0.0220629945,
-0.0286784042,
0.01269772,
-0.0058575599,
0.0330423042,
0.0111309625,
-0.0240062885,
-0.0835679621,
-0.0822095871,
-0.0026853662,
0.004990824,
0.00069432,
0.0202174485,
0.0235426985,
-0.0190909971,
0.0661866739,
0.0383513235,
0.0281599816,
-0.0383994952,
0.0168172326,
-0.0195297077,
-0.0266216155,
-0.0262763128,
-4.95916969e-33,
0.0044556279,
-0.0319457389,
0.0155916447,
0.063347742,
-0.0208403468,
-0.0048746546,
-0.0032806182,
-0.0149311749,
-0.0271949861,
0.0197654031,
-0.0334641188,
0.0336385854,
0.0134484684,
-0.0205858592,
0.0653623492,
-0.0241826624,
0.0101675075,
-0.0093359044,
0.019537244,
0.0244022459,
-0.0383635089,
0.0357422717,
0.0072557605,
-0.091560185,
-0.0137423147,
0.0624392182,
0.0044420832,
-0.0385171771,
-0.070060432,
0.0342571326,
0.014211759,
0.0345290452,
0.054663118,
-0.0492849238,
-0.0094365673,
0.0053685275,
-0.0529608466,
0.0162312295,
-0.0322632566,
-0.0502352715,
0.017607905,
0.0262460317,
0.0109637417,
0.0173863899,
-0.037482772,
0.0040021134,
0.0345130824,
-0.0425234735,
-0.050349094,
-0.0910342112,
-0.0372590683,
0.0652545914,
-0.0265927035,
0.0352302603,
0.0103926351,
0.0153099885,
0.0280969478,
0.0448759645,
-0.0628769398,
0.0252708308,
-0.0565150492,
-0.0386377163,
-0.0141536482,
-0.0756504983,
-0.0278841313,
0.0870281011,
0.0192303956,
0.0017080298,
-0.0411662832,
-0.0115753021,
0.014472574,
0.1109368727,
0.0245201681,
0.0241567697,
0.0414874367,
-0.0791600123,
-0.0867251828,
0.0058157803,
0.0572563522,
0.0118988883,
0.0109168151,
-0.0451965407,
-0.056427978,
0.0166267902,
-0.0213396475,
-0.0488179438,
0.0106652239,
0.0112485234,
0.0114423232,
0.0125625255,
0.0806035325,
-0.005946904,
0.0043151034,
-0.0334284455,
0.033061374,
0.0665471777,
-0.0027050183,
0.0100530321,
0.0035648977,
0.0278579164,
-0.0409172587,
-0.0617043488,
-0.007244071,
-0.0342749357,
-0.0054050898,
-0.0510831848,
0.0473424457,
-0.0228617378,
-0.0337069109,
-0.0037534032,
0.043876376,
0.0271995552,
0.0333431438,
-0.0117549747,
-0.0026027358,
0.0336463675,
0.04026049,
0.0145908268,
0.0069404426,
0.0171550252,
-0.0428892896,
-0.0505247973,
0.0120914048,
-0.017896045,
0.022044085,
-0.0354929492,
0.0019003236,
0.0423333533,
0.0215976126,
0.0164766908,
0.0187707581,
0.0646530539,
2.394e-7,
0.0774168447,
0.0520773456,
-0.0246539433,
0.0134103475,
-0.0037994755,
0.0273602996,
-0.0106932176,
0.0424090512,
0.0669839233,
0.0094280001,
0.052789636,
-0.0087694358,
0.0456275493,
0.0322555341,
-0.0538400821,
-0.0627916008,
-0.0585470647,
-0.0432267524,
-0.0838714242,
0.0038310268,
0.0451016761,
0.0931529254,
0.0536105186,
-0.0236790106,
0.0229753759,
0.0063505876,
0.0546804257,
-0.0238510184,
-0.0196139328,
-0.019215649,
0.0435446352,
-0.032271374,
-0.0107174525,
0.0230412465,
-0.0370621197,
0.0718875229,
-0.0634711385,
0.0148580251,
0.0177139435,
0.0543797798,
-0.0024606371,
-0.0140666375,
0.0218863599,
-0.0182788875,
0.0323163345,
-0.0278076436,
0.0120940944,
-0.0249724779,
-0.0549627282,
-0.0051379222,
0.0385187604,
0.0485815667,
-0.0186273921,
-0.0238302797,
0.0080247559,
-0.002017864,
0.0066310521,
-0.0729734898,
-0.002458751,
0.058374092,
-0.0199220292,
0.0217139181,
0.0269986149,
-0.0340286382,
0.0677331015,
-0.0768581182,
0.048214782,
1.681914881e-34,
0.0627265051,
0.020709252,
0.0011372465,
0.0212410856,
0.016125815,
0.019543238,
0.0090310592,
0.0046786312,
-0.0118677085,
-0.0296671968,
0.0422288738
] |
http://arxiv.org/abs/2401.15724v3 | http://arxiv.org/pdf/2401.15724v3 | RE-GAINS & EnChAnT: Intelligent Tool Manipulation Systems For Enhanced Query Responses | 2024-01-28T18:26:31 | 2024-06-20T10:47:47 | Sahil Girhepuje, Siva Sankar Sajeev, Purvam Jain, Arya Sikder, Adithya Rama Varma, Ryan George, Akshay Govind Srinivasan, Mahendra Kurup, Ashmit Sinha, Sudip Mondal | cs.CL | cs.CL | Large Language Models (LLMs) currently struggle with tool invocation and
chaining, as they often hallucinate or miss essential steps in a sequence. We
propose RE-GAINS and EnChAnT, two novel frameworks that empower LLMs to tackle
complex user queries by making API calls to external tools based on tool
descriptions and argument lists. Tools are chained based on the expected
output, without receiving the actual results from each individual call.
EnChAnT, an open-source solution, leverages an LLM format enforcer, OpenChat
3.5 (an LLM), and ToolBench's API Retriever. RE-GAINS utilizes OpenAI models
and embeddings with a specialized prompt based on the $\underline{R}$easoning
vi$\underline{a}$ $\underline{P}$lanning $(RAP)$ framework. Both frameworks are
low cost (0.01\$ per query). Our key contribution is enabling LLMs for tool
invocation and chaining using modifiable, externally described tools. | RE-GAINS & EnChAnT: Intelligent Tool Manipulation Systems For Enhanced Query Responses: Large Language Models (LLMs) currently struggle with tool invocation and
chaining, as they often hallucinate or miss essential steps in a sequence. We
propose RE-GAINS and EnChAnT, two novel frameworks that empower LLMs to tackle
complex user queries by making API calls to external tools based on tool
descriptions and argument lists. Tools are chained based on the expected
output, without receiving the actual results from each individual call.
EnChAnT, an open-source solution, leverages an LLM format enforcer, OpenChat
3.5 (an LLM), and ToolBench's API Retriever. RE-GAINS utilizes OpenAI models
and embeddings with a specialized prompt based on the $\underline{R}$easoning
vi$\underline{a}$ $\underline{P}$lanning $(RAP)$ framework. Both frameworks are
low cost (0.01\$ per query). Our key contribution is enabling LLMs for tool
invocation and chaining using modifiable, externally described tools. | [
0.0327506103,
0.054027319,
-0.0293166079,
0.0290567111,
-0.0484043323,
-0.0126572326,
-0.0410408676,
0.0350588933,
0.0565880798,
-0.0460257679,
-0.0411719196,
0.0280470699,
-0.0103895655,
0.0328257009,
0.0039824098,
-0.059019234,
0.0363076888,
-0.0202301014,
-0.0087374076,
0.0260497667,
-0.0113406777,
0.0160591863,
-0.030762529,
0.0865113735,
0.0348157249,
-0.0152024357,
-0.0489628315,
0.0009134643,
-0.0317175649,
0.0071142265,
-0.0009953215,
0.0710400194,
-0.0219657402,
0.0529561453,
0.0000023483,
-0.0610840432,
-0.0044734324,
0.0172206331,
-0.0308555569,
-0.0156931728,
0.0168028716,
0.0135843717,
0.0077750706,
0.0517708436,
-0.0467594825,
-0.0335684679,
0.055598367,
-0.0010622797,
0.0722959638,
0.0593873635,
0.0019461046,
0.0117019061,
0.0138383172,
-0.0214396399,
-0.0141458828,
-0.0180619247,
0.0185660254,
-0.0466404259,
0.0740113929,
0.0084243389,
0.0298367683,
0.0201953463,
0.0334218964,
-0.0263841748,
0.0037513992,
0.0032313829,
-0.0551373363,
-0.0309987292,
0.0114474213,
0.0416570567,
-0.0187984332,
-0.012823551,
-0.0090224482,
0.0005387891,
0.0170973539,
0.0286793187,
-0.0724293739,
0.0004410119,
-0.0003715481,
-0.0241005886,
-0.0545320958,
-0.0200339947,
0.0087258331,
0.0189889297,
-0.030706143,
0.062655367,
0.0205059294,
0.0109194946,
-0.0119634522,
-0.0172314923,
0.007846348,
-0.0715859681,
0.0059247874,
0.0673094839,
-0.0563977882,
0.0272786878,
-0.0214965679,
-0.0331590213,
0.0281439368,
-0.0400401354,
0.0017606197,
0.051172629,
-0.053438507,
0.0288411416,
-0.0171296243,
0.011790324,
-0.0048467624,
0.0384087265,
-0.0353694037,
-0.02811254,
-0.0404850356,
-0.032371074,
-0.0457035638,
0.0201905891,
-0.0531208515,
-0.0345991328,
-0.0182808153,
0.0764069781,
0.0061934437,
0.0178523697,
0.001019153,
-0.0169825125,
-0.0205920991,
-0.0203967225,
-0.045944836,
0.0244197063,
0.0037711959,
-0.0104615949,
0.0155811673,
-0.0791663975,
-0.0280514657,
0.0130598107,
-0.0073802532,
-0.0173367541,
-0.025236953,
0.0955437794,
-0.0189529713,
-0.0014961966,
-0.0510519594,
-0.0502584837,
0.0368238278,
0.0000900896,
-0.0145159625,
-0.0588304959,
0.045876231,
0.019362554,
-0.0516966581,
-0.0438409708,
0.0017338352,
0.0656933039,
-0.045639202,
0.0438130312,
-0.0111554628,
0.0247933418,
0.0292805638,
0.0153432144,
-0.0058537852,
0.0722352266,
-0.0055368673,
0.0540985279,
0.0400993377,
-0.0054273414,
0.0394664779,
0.0569211207,
-0.0035603389,
-0.0373756886,
-0.0518077128,
0.0284151565,
0.0371791385,
0.0238268003,
-0.0187290087,
0.0331794135,
-0.025967788,
-0.0078753857,
0.0376149006,
0.0512030311,
0.0433960706,
-0.0125393746,
0.0181730222,
0.0020914585,
-0.0027277577,
-0.0155522181,
-0.0618802905,
0.0290024523,
-0.0348865464,
0.0061050653,
0.0074358638,
0.0168557875,
-0.0151369432,
-0.042512238,
-0.0005505488,
-0.0101947095,
-0.0395931154,
0.0113860536,
0.0363586769,
0.0159288123,
-0.0190984383,
-0.0092620002,
0.0282194074,
-0.035551019,
-0.0244746823,
-0.0890775546,
-0.0038665163,
0.0880897492,
0.0850437507,
-0.0299422145,
-0.0182867404,
0.0159317441,
-0.0400931127,
0.0251293778,
0.0265449136,
-0.0224295985,
-0.0049886121,
-0.0289885774,
-0.017787369,
0.0305735674,
0.0066974168,
0.006410636,
0.0498231389,
-0.0101520289,
0.0024481651,
-0.043202173,
0.0389360003,
0.0205102004,
0.0241491273,
-0.0463297628,
0.0078913011,
0.0301858652,
0.0080530252,
0.0044695386,
0.01504457,
0.0399392508,
0.0184256192,
-0.0103560509,
-0.0381317958,
-0.0079977075,
-0.0047171195,
0.0261547118,
0.0349510871,
0.0584766157,
0.0110552208,
0.0032478662,
-0.029733317,
-0.0316871777,
0.0117669366,
0.0045767846,
0.0779418349,
-0.0206760392,
0.0107988222,
0.0154268304,
-0.0023236116,
-0.0484499298,
0.0389936417,
0.0042564841,
-0.021515809,
0.0060500745,
-0.0155510306,
-0.0160560682,
-0.0310738012,
-0.010599955,
0.053325478,
-0.0001073708,
0.0279760715,
-0.0069623343,
0.0331895612,
-0.0195313245,
-0.0216126088,
-0.1147735864,
0.0012889771,
-0.0266059898,
0.0018046764,
0.008302303,
0.067891404,
-0.0320453048,
-0.0399542451,
-0.0116620911,
-0.0053636292,
-0.0016452817,
0.0590060093,
-0.0042812545,
0.0121803144,
-0.0455703139,
-0.0311156549,
-0.0379205346,
-0.0094054304,
0.0469071716,
-0.0190639459,
-0.0440389812,
-0.0268906727,
0.0045834593,
-0.0409657843,
0.0293461997,
0.0032279787,
-0.0208638143,
0.0279819444,
-0.0546281971,
0.003553875,
0.0700611249,
-0.0032684987,
0.0251777954,
0.0514662527,
0.0049895509,
-0.0885159373,
0.0014295195,
-0.0118528102,
-0.0582274646,
0.028871445,
0.0029063905,
0.050099466,
0.0151711069,
0.0047111916,
0.0179527458,
-0.0208529327,
-0.0048526502,
0.0263302829,
-0.0402204469,
-0.0321504287,
-0.0335432068,
-0.0085713221,
0.0235176682,
0.0038868175,
-0.0619659163,
0.0698989183,
-0.0233372133,
-0.0146870455,
-0.0166193862,
-0.0350423865,
-0.0249429196,
0.1113464162,
0.0185202695,
0.012964502,
-0.0008496881,
-0.049339056,
-0.0302773882,
-0.0108511299,
0.0114650596,
-0.0626317486,
0.0263230167,
-0.0616039746,
0.0208551995,
-0.0084588546,
-0.0203038212,
-0.0489129871,
-0.0013617855,
-0.0313589014,
0.0136922039,
-0.004001698,
-0.0301025957,
0.029533308,
0.0194102693,
0.0368266553,
0.049291674,
-0.0171829108,
-0.0132555543,
-0.0197297037,
0.0011114334,
0.0532466769,
0.0084370589,
-0.0540237203,
0.0056542349,
0.0512023084,
0.0507949777,
0.0485803746,
-0.0877994224,
-0.0285465177,
-0.0173680596,
0.0530217253,
-0.0468935482,
0.0089286184,
0.0422608852,
-0.0052667721,
-0.0640397593,
0.0100861443,
0.072598286,
0.0517950729,
-0.0724548027,
0.0025888877,
-0.0188794527,
0.0380879603,
0.0276606884,
0.0822355822,
-0.1143335104,
-0.0235630237,
0.0170407537,
0.0588860735,
0.0136943609,
-0.0236672964,
-0.0142928706,
-0.1557094008,
0.064829953,
-0.0167276561,
-0.055562254,
0.0068435329,
-0.0317405686,
0.0314495601,
0.0077064661,
0.0098489951,
0.0431250036,
0.0042245016,
0.0028690761,
-0.0151585704,
0.0101075983,
-0.0307242051,
-0.0242049862,
-0.0232203491,
0.0156080835,
0.0565410964,
-0.01137379,
0.0617339313,
0.0159053952,
0.0741846636,
-0.0102603361,
0.0095037119,
-0.0559114777,
0.0304493587,
0.0201737005,
-0.0160499271,
-0.0338424295,
0.0245807003,
-0.0165383089,
-0.0229763947,
0.0003338641,
0.0628719032,
-0.0495022163,
0.0526424162,
-0.0075876005,
-0.0470433198,
0.011133315,
0.0397927985,
0.0373933464,
-0.0192165356,
-0.0187123846,
-0.0498916954,
-0.0397100411,
-0.0085832197,
-0.0417003408,
-0.0002343874,
0.004806926,
-0.0369352624,
-0.0606729537,
-0.0202481914,
0.0682508498,
0.0198847689,
0.0192781389,
0.0168202668,
0.0224184543,
0.0226360466,
-0.0235584974,
0.0062983488,
-0.0458366834,
0.0159375109,
-0.0550861657,
-0.0058715618,
0.0220576134,
-0.0322741382,
0.0034390416,
0.0086214785,
-0.0035264743,
-0.0789585114,
-0.0167926215,
0.0755201876,
0.0336167328,
0.0039871559,
-0.0505419448,
-0.0588401072,
-0.0640898123,
0.0914690942,
-0.0632760152,
-0.0118619902,
0.0015431858,
0.0270195995,
0.0055166301,
0.0347117856,
-0.0762963966,
-0.0105598681,
-0.0435845964,
0.0281172153,
-0.0264143012,
-0.0918844342,
0.0477179959,
0.0501995347,
-0.0065545095,
0.0357408635,
0.0051068836,
-0.0294474587,
0.0403532051,
-0.0288734678,
-0.0163734294,
-0.0393238552,
0.0225842651,
-0.0399447493,
-0.0067166206,
0.0301301368,
0.0036537552,
-0.0006053919,
0.0289632138,
-0.0410455465,
0.0036021473,
-0.0325553492,
0.0134816822,
0.0298436712,
0.0296786744,
-0.0261358302,
0.0299698915,
0.015381542,
0.0570057929,
-0.0548599288,
0.0601025932,
0.0042873775,
-0.0047541913,
-0.0154019436,
0.0281428192,
-0.0684835091,
-0.130682677,
0.0262413882,
0.025149487,
0.0291091055,
-0.0270217992,
0.0251861792,
-0.022912506,
0.0026397198,
-0.0261820182,
0.0177488383,
0.016828198,
-0.0077344235,
-0.0038334564,
0.02638123,
-0.0062885014,
-0.0098914439,
0.0701651052,
-0.0713154227,
0.0227586199,
0.0101049235,
-0.040052969,
-0.0418494046,
-0.0793312192,
0.0519664548,
0.0284592435,
-0.0391123705,
-0.0386921912,
-0.0264858734,
-0.0486233719,
-0.0710342228,
0.00313247,
-0.0174557101,
-0.0077140005,
0.0364571773,
0.081703119,
0.0451336354,
-0.0208748635,
-0.0430143848,
0.0287201069,
-0.0456513651,
-0.0443472899,
0.0166650973,
-6.392784968e-33,
-0.0303933546,
-0.0283461642,
0.0345531777,
-0.0051815594,
-0.0061627487,
-0.035673447,
-0.0355015248,
0.0020134854,
-0.0307775605,
-0.0323180631,
-0.0199575014,
0.0055008898,
0.0253956616,
0.0094571095,
0.02435489,
-0.0036536248,
0.0233718064,
0.004714868,
-0.00608279,
0.0033869434,
0.0565155223,
0.0037332023,
-0.0027828708,
-0.0759398639,
-0.0109913703,
-0.0078475233,
0.0092152804,
0.038872119,
-0.0224587042,
0.0498461165,
0.0008580025,
0.0389756225,
0.0102957357,
-0.0209767818,
-0.0027507783,
0.0110652195,
-0.0591877811,
-0.0306587908,
0.0241117273,
0.0200404264,
-0.0024515863,
-0.0090764761,
0.0533784255,
-0.008469726,
0.0055136415,
0.0203401521,
0.0125478311,
-0.0131656956,
-0.040588975,
0.0412872732,
-0.0177993607,
0.0175644159,
0.0266147088,
-0.0492753573,
0.0016400719,
0.0423979647,
0.0072274655,
-0.0025541391,
-0.034518797,
-0.0003883808,
-0.0003063925,
0.0537947193,
0.0541606657,
0.0436825901,
-0.0008014624,
0.0244617667,
0.0059632766,
-0.0181673672,
-0.0093068788,
0.0157999005,
0.0032487935,
0.0833647996,
0.0240649842,
-0.0198102817,
-0.0298618861,
0.0078569753,
-0.0219556782,
0.0374094695,
0.042450428,
0.0484816357,
0.0057677431,
0.0484353378,
-0.0370201916,
-0.0136491517,
0.0059857015,
0.0001745679,
-0.0655852407,
-0.0319404118,
0.0447660945,
0.0196070597,
0.0287410226,
-0.0339209363,
0.0215277709,
0.0011643026,
0.0073554735,
0.0095441146,
-0.0099274777,
0.0294604152,
-0.0128587028,
0.0011335444,
0.0027903004,
0.0117967483,
0.0082274182,
-0.0413154773,
0.0060348464,
0.0327277593,
-0.0259568524,
0.0302399173,
-0.066674456,
0.0008644381,
-0.0329621248,
0.0304659903,
-0.0010278194,
-0.0066211442,
0.0385283232,
0.0121768974,
0.0005068499,
0.0682121739,
-0.0120953852,
-0.0186891146,
-0.0555218719,
-0.0537362918,
0.025347935,
-0.0251823291,
-0.021382805,
0.019472573,
0.002179353,
0.0249833409,
0.0228034016,
-0.0281867888,
-0.0060042851,
0.0191093069,
3.117e-7,
0.0312959552,
0.0098338043,
0.0371280015,
0.0866000503,
-0.0095217181,
0.0132994168,
-0.0037578582,
0.0069805919,
0.0169991814,
0.0038021232,
0.0049608252,
0.0285173655,
0.0408956148,
0.0287981275,
-0.0835543945,
-0.0143366605,
-0.0397208594,
-0.0244528633,
-0.0751535818,
-0.0060496326,
0.0629524961,
0.066576764,
0.0103093451,
0.0011084207,
-0.0173064098,
-0.0591389276,
0.0177189279,
-0.0206718873,
-0.0201883931,
0.0093680369,
0.0372578017,
0.0268796925,
0.0194094591,
0.0212215744,
-0.0175923966,
0.0051873485,
0.0201940238,
0.0557471067,
-0.0147440331,
-0.0092163021,
0.0036423851,
-0.0317917317,
0.0109294225,
-0.0560820773,
0.0570005924,
-0.0124754142,
-0.0095422463,
0.0100886282,
0.0366203077,
-0.0202768017,
-0.0227484126,
-0.0184500702,
-0.0062602488,
0.0222752653,
0.0227322746,
0.0093737971,
0.0004498732,
-0.0605454519,
-0.0494804606,
0.009201481,
-0.0228931457,
-0.0442423448,
-0.0632870197,
-0.0169410799,
-0.0124993017,
-0.0050877081,
-0.0268022548,
3.30680451e-34,
0.0267044362,
0.0119539043,
0.0194588751,
0.0974417031,
0.0478872284,
0.0324598886,
0.0361251906,
0.0533413589,
0.0378229544,
-0.1085675359,
-0.0569560267
] |
http://arxiv.org/abs/2401.15717v1 | http://arxiv.org/pdf/2401.15717v1 | Check News in One Click: NLP-Empowered Pro-Kremlin Propaganda Detection | 2024-01-28T17:51:47 | 2024-01-28T17:51:47 | Veronika Solopova, Viktoriia Herman, Christoph Benzmüller, Tim Landgraf | cs.CL | cs.CL | Many European citizens become targets of the Kremlin propaganda campaigns,
aiming to minimise public support for Ukraine, foster a climate of mistrust and
disunity, and shape elections (Meister, 2022). To address this challenge, we
developed ''Check News in 1 Click'', the first NLP-empowered pro-Kremlin
propaganda detection application available in 7 languages, which provides the
lay user with feedback on their news, and explains manipulative linguistic
features and keywords. We conducted a user study, analysed user entries and
models' behaviour paired with questionnaire answers, and investigated the
advantages and disadvantages of the proposed interpretative solution. | Check News in One Click: NLP-Empowered Pro-Kremlin Propaganda Detection: Many European citizens become targets of the Kremlin propaganda campaigns,
aiming to minimise public support for Ukraine, foster a climate of mistrust and
disunity, and shape elections (Meister, 2022). To address this challenge, we
developed ''Check News in 1 Click'', the first NLP-empowered pro-Kremlin
propaganda detection application available in 7 languages, which provides the
lay user with feedback on their news, and explains manipulative linguistic
features and keywords. We conducted a user study, analysed user entries and
models' behaviour paired with questionnaire answers, and investigated the
advantages and disadvantages of the proposed interpretative solution. | [
0.0667529851,
-0.0015932949,
-0.0265623592,
0.0018618893,
-0.0145840552,
0.0137559632,
-0.0014194674,
-0.0092472881,
0.0577066801,
-0.0171686336,
0.0111355148,
0.0171523895,
-0.0319257602,
0.0382099673,
0.0178692378,
-0.0634225681,
0.0667708144,
0.0003175914,
0.0435559452,
-0.0105731199,
0.0443786532,
0.0420978442,
-0.020322863,
0.025663998,
0.018503787,
-0.0047316337,
0.0045413971,
-0.0040257047,
0.0043301713,
-0.0300686806,
-0.041332487,
0.0333864205,
-0.0345783494,
0.0725913942,
0.0000018357,
0.0020659075,
-0.0267497804,
-0.0210672617,
-0.0183515362,
-0.0152463568,
0.0877987146,
-0.0632403493,
-0.0083859591,
-0.0091006765,
-0.0543299653,
0.0223488808,
0.0263783652,
0.0419467874,
0.0861203671,
0.0779220164,
-0.0113393599,
0.0127416411,
-0.0388374701,
0.0242345296,
0.0016707361,
0.0045812591,
0.0180525128,
-0.0424459204,
-0.0268135797,
0.0254211724,
0.0522472411,
0.0513034649,
0.0027175341,
0.0155342044,
-0.0631082281,
-0.0049332799,
0.0079858145,
-0.0683964714,
-0.0245437715,
0.0092132241,
-0.010774943,
-0.0252406932,
0.0010043968,
0.0053177909,
0.0069028316,
0.0470525697,
-0.0337282531,
-0.0128431171,
-0.0366761126,
0.0106243677,
0.0115897087,
0.0005290975,
-0.0038637328,
0.0297657102,
0.0203683749,
0.0606138259,
-0.0357319228,
0.0132171707,
0.0385174155,
0.0062192315,
-0.0014222969,
-0.0462521575,
0.0541090742,
0.0093626948,
-0.0220711548,
0.0025253589,
-0.024464028,
-0.0275650695,
0.0607349724,
-0.1104076505,
-0.0228057299,
0.0353710987,
0.0100181038,
0.0534303151,
-0.0173647106,
-0.0243052784,
-0.0004286485,
0.0281475596,
-0.009147469,
-0.0112148914,
-0.022750847,
-0.0112358686,
-0.0082989987,
-0.0141482959,
-0.0011265324,
0.0076311803,
-0.0179719049,
-0.0095766317,
0.0090481471,
0.0607728511,
-0.0505700782,
0.0194661003,
0.0008757327,
0.0219666511,
-0.0373486876,
0.0598578006,
-0.091576308,
-0.0031823774,
-0.0385682024,
-0.0290159695,
0.012375813,
0.025284458,
-0.0116934702,
-0.0227758139,
-0.0049410788,
0.0044689556,
0.0245277304,
0.009380972,
-0.0256106444,
-0.0738682225,
-0.007824719,
-0.0355542973,
-0.0016383704,
-0.041497454,
-0.0124327336,
0.0326496139,
-0.0115516791,
-0.0436537638,
0.0007404039,
0.0370282978,
-0.0354744978,
0.0834650695,
0.0005725729,
0.0184541121,
0.0058727707,
0.0112821395,
0.0345400833,
0.0959315598,
-0.0495666973,
0.1070366353,
0.0524254069,
-0.0071400749,
0.0528965853,
0.0099792229,
0.000204304,
0.0240139738,
-0.0109812804,
0.0148185175,
-0.0615976974,
-0.0358980596,
-0.0022127726,
0.022603631,
0.001149244,
-0.0256463792,
0.0769459084,
0.0097994795,
0.0046548257,
-0.0078703016,
0.0146923782,
0.0180586074,
-0.0280324053,
0.0209261794,
-0.0282423478,
0.0219619088,
-0.0514720567,
-0.0348774977,
-0.0413831547,
0.0244684182,
-0.001081593,
-0.0295539536,
-0.017801011,
0.0133628631,
0.0659450814,
-0.0031037331,
0.0138095915,
-0.0288451724,
-0.0062757218,
0.0211004931,
0.0366460904,
-0.0221457481,
0.0332721844,
-0.0492474474,
0.0183664765,
0.0444711223,
0.0072629536,
-0.0126293739,
0.0377896763,
-0.029233912,
0.0025272497,
0.0476418845,
0.0506940596,
-0.0248084869,
-0.038927678,
0.0101823946,
-0.0090287179,
-0.0282627176,
0.0135445213,
0.0376945026,
-0.0054753544,
0.0452481285,
-0.0091251703,
-0.0379063897,
-0.0230807215,
0.0259798598,
0.0615237094,
-0.0433510095,
-0.0613493137,
0.0597814284,
-0.0248702019,
0.0595810711,
0.027316967,
0.0298011433,
0.068627663,
0.0136990715,
-0.0016910497,
-0.0175506547,
0.0327817164,
0.0294456929,
-0.0037708371,
-0.0296462271,
-0.0325632282,
0.0183100477,
0.0089086751,
0.0092491256,
0.0561253428,
-0.0366948098,
0.0030842931,
-0.0300540719,
-0.0221742895,
-0.0304290149,
-0.0089128893,
-0.0218365099,
0.0488867648,
0.0023833355,
0.0174881238,
0.007932445,
-0.0426096506,
0.0182290971,
0.0167775527,
-0.0212659687,
0.0237386208,
-0.0047169961,
-0.0149703724,
0.0463094041,
-0.0403106958,
-0.0354750119,
0.0127843078,
-0.068567507,
-0.0139390547,
-0.028003009,
0.051424522,
0.0636513084,
0.0182073321,
-0.0096585127,
-0.0008849589,
-0.0581635237,
-0.0610230044,
0.0049462044,
0.0369119793,
0.0016004734,
0.0009108152,
0.0167028513,
-0.0018886697,
0.0220088605,
-0.0414976627,
-0.0058305068,
0.0435853601,
-0.0462209061,
-0.0238068812,
-0.0430360213,
-0.0277126282,
-0.0132708708,
-0.0271153748,
-0.0496472158,
0.0417224579,
0.0009608514,
-0.0096247345,
0.0495859981,
0.0022136208,
0.0831753463,
0.0575031936,
0.0499158651,
-0.0462851524,
-0.0321038067,
-0.0774202123,
0.01474432,
-0.0298335254,
-0.0394245796,
0.026031984,
0.0278899781,
0.0317873061,
0.0307742283,
-0.0291151106,
0.0436870679,
-0.0412409157,
-0.0088247107,
-0.0417127907,
-0.0146414563,
-0.0016241712,
0.0623162352,
0.0146344304,
0.0080600847,
0.079141885,
-0.0004758722,
0.0258153696,
-0.0299030673,
-0.0728488639,
-0.0092080543,
0.0734697506,
-0.0383041315,
0.0237845574,
-0.009173464,
-0.0247160289,
-0.0122107016,
-0.0997192562,
0.0444057882,
-0.0258633625,
0.0267602839,
-0.0298684668,
-0.0064353147,
0.0214061085,
0.016206298,
-0.024649037,
-0.1221485734,
-0.0313817114,
-0.0438367762,
0.0195274409,
-0.0289106723,
-0.0404804535,
0.0314618833,
0.0656685531,
0.0053348169,
-0.0267292727,
-0.0025117025,
-0.0079646921,
0.0192026235,
0.0325311385,
-0.0228710398,
0.0512760058,
-0.003180481,
-0.0208627302,
0.0687105283,
-0.0116526894,
0.0271892659,
0.0161450803,
-0.0311501306,
0.0329924338,
-0.048938442,
-0.0033077789,
-0.0347349681,
0.0062913653,
-0.0458875783,
0.0487236865,
0.051127445,
0.0408885181,
-0.0279017705,
0.0091588041,
-0.01409961,
-0.003976882,
0.0243092272,
0.0419603139,
-0.0751759335,
0.0110441763,
0.0084629217,
0.0226949323,
-0.0541353263,
-0.0313471742,
-0.0299571976,
-0.0490246303,
-0.0369295329,
-0.0024034933,
-0.0088303946,
-0.0267059486,
-0.0932400674,
0.0023465587,
-0.0160857141,
-0.0219979417,
0.0282915402,
0.0009238775,
0.0321016796,
-0.0010869972,
-0.0634535253,
-0.0458283015,
-0.0096274111,
-0.0111062173,
0.0159403533,
0.0474824086,
0.0491637997,
-0.0212436169,
-0.0081747444,
-0.0314472802,
-0.0001977527,
-0.0210200325,
-0.0245324727,
0.0648954287,
-0.002397134,
-0.0056334799,
-0.0318552852,
0.0274545643,
0.0029048992,
0.0152537767,
0.0320358835,
0.0195511784,
0.0036463155,
-0.0373876393,
-0.0041830218,
0.0165403094,
0.0824899599,
0.0524594001,
0.0068716272,
0.0131770186,
0.0067357887,
-0.0307742022,
-0.0184260197,
-0.0289654434,
-0.082574673,
0.0533613488,
-0.0210541822,
-0.0811186656,
-0.0462952293,
-0.076566346,
0.0456903949,
0.0324668847,
-0.0047898004,
0.0288454369,
-0.0120294383,
-0.0075282734,
-0.0078912647,
0.0271859933,
0.0420189947,
0.0529662222,
-0.0295126978,
-0.0355541185,
-0.0345635824,
0.0087119881,
-0.0112410132,
0.0359553918,
-0.0179399513,
-0.0261513703,
0.0139503963,
-0.0067755044,
0.0295797978,
0.0120917903,
-0.0047566844,
-0.0144626945,
0.0219480116,
0.0230043624,
-0.0557844415,
-0.004150663,
0.0085768038,
0.0247140452,
0.0390149169,
0.0212910622,
-0.0051042032,
0.0118439458,
0.0087261302,
-0.0614283867,
-0.0655983984,
0.0021399416,
0.0202690978,
0.0043263775,
0.0725132897,
-0.0218099747,
-0.0098585971,
-0.0276234411,
-0.0440127142,
-0.0018556215,
-0.0282925181,
0.0140682291,
0.0038603237,
0.0107690161,
-0.0296001472,
-0.0093161343,
0.0503824092,
-0.0680664033,
-0.0333289765,
-0.0481252484,
0.0149023747,
-0.0010293096,
0.0187465046,
0.0224605817,
0.0075775604,
-0.0334396958,
0.0578943081,
-0.0067697475,
0.0155398212,
-0.015973907,
-0.0259491745,
0.0072120042,
-0.0387197547,
0.056614168,
0.0099997129,
-0.043062266,
-0.0457275584,
0.017343238,
-0.0249399226,
0.0072783371,
-0.0197062641,
0.0639187768,
-0.0225706976,
-0.0049178679,
0.0316712782,
0.0032206406,
0.0060979673,
-0.0306050405,
0.0136816092,
-0.0026974052,
-0.0455269031,
-0.0238208976,
0.0126103004,
-0.0452138372,
0.0242850985,
0.0382388346,
-0.0750412047,
-0.0169992913,
-0.0182389338,
0.010462705,
0.0048963791,
-0.0401092693,
0.0038885695,
-0.0712832138,
-0.0254018363,
-0.0155037669,
0.0130332988,
-0.0341590978,
-0.0090536429,
0.0318399966,
0.064977698,
0.0259266105,
0.0342328474,
-0.0920332745,
0.0121359769,
-0.0572784543,
-0.0235843714,
0.008850202,
-6.533152217e-33,
-0.0149981137,
-0.0507839024,
-0.0109707816,
-0.0475630537,
0.0519170873,
-0.0166039187,
-0.0126481,
-0.0006491618,
-0.0373737663,
-0.0126936249,
0.0408760756,
-0.0084886607,
0.0240920726,
-0.0508124828,
-0.0364574604,
0.00376876,
0.05830485,
0.0198570695,
-0.0209620427,
-0.0029621397,
0.099427484,
-0.011158024,
0.0543156452,
-0.0743454024,
0.0924125686,
0.020946566,
-0.0456456132,
0.0030086886,
0.0367086381,
0.0221831314,
0.0018977049,
0.0355327837,
0.0197847057,
-0.0172230192,
0.0045838729,
0.0564291328,
-0.028484514,
-0.0469753742,
0.0633661151,
0.0259613693,
-0.0479548089,
0.0296259858,
0.0552254803,
-0.0441557989,
0.0137257678,
0.0824375823,
0.0369643718,
-0.0086803446,
0.0004819995,
0.0173542909,
-0.0726959035,
0.0033220409,
-0.0169234816,
0.0214890018,
-0.0290165618,
0.097307846,
0.0379181206,
0.0027578641,
-0.1111459062,
0.0331302136,
-0.0002503435,
0.0232307818,
0.0636982992,
-0.0045214351,
0.0089030424,
0.0489199087,
-0.0136796609,
-0.0196840614,
-0.0572060756,
0.0504353419,
0.030952543,
0.0010725826,
0.063587442,
0.0506350584,
0.0100611821,
-0.0782120973,
-0.0480488203,
0.0227116384,
0.0940769315,
0.0430465415,
0.0017175588,
-0.0104978541,
-0.0323393494,
-0.0128717758,
-0.0363568589,
-0.0360716283,
-0.0499891341,
-0.0811580271,
-0.0084616942,
0.0023821048,
-0.0304665901,
0.0193046406,
0.0067695505,
-0.0045241513,
-0.0100024138,
-0.0027263986,
-0.0133358715,
0.0005011397,
-0.0032814653,
0.042889066,
0.0033397498,
0.0198703054,
-0.0165670682,
-0.0122277234,
0.0216969289,
0.0027842964,
0.0266509838,
0.0264622867,
-0.0408337377,
-0.0025447756,
-0.0118532712,
0.0462419465,
0.0145522878,
0.0121584674,
0.0466794483,
-0.0132964952,
-0.0312612541,
0.0431933068,
0.0093053756,
-0.011890498,
0.022177374,
-0.0146819074,
-0.0087603377,
0.0444972441,
-0.0244147982,
-0.0334117189,
-0.0566421822,
0.0383673012,
0.0398282744,
-0.0461470373,
-0.0007534015,
-0.0389726087,
2.657e-7,
0.0416861326,
0.0389541909,
-0.0489579886,
0.0106493458,
0.0154489009,
0.0379060134,
0.0102587948,
0.0135211227,
-0.0790436789,
-0.0586317368,
-0.0115306322,
-0.0049093957,
0.0111005083,
0.0714218989,
-0.0649677739,
0.0162299741,
-0.0234919358,
0.0186519995,
-0.0783037767,
0.0620277077,
0.0715836063,
0.0354678519,
0.0465409234,
0.0220763944,
-0.0175093729,
-0.0324808247,
0.0124886287,
0.0116341971,
-0.015173764,
-0.0098062521,
0.0537286215,
-0.0052333833,
0.0369109958,
0.0105737057,
0.0091246301,
-0.0151984887,
-0.0147829885,
0.0263192933,
0.0058457069,
-0.0333342031,
0.0018534822,
-0.0644003525,
0.0003345459,
-0.0919148624,
0.10105858,
-0.0583183654,
-0.0421551503,
-0.0309295189,
-0.0021721411,
-0.0340958498,
0.0207828972,
-0.0103928596,
-0.0583749153,
0.0271665361,
-0.0025308428,
-0.016553048,
0.0153573034,
-0.0244439486,
0.0092198756,
-0.0250199549,
-0.0251049828,
-0.0327295251,
-0.023788482,
0.0148245879,
-0.023042636,
0.0040308558,
0.0015779479,
2.139835019e-34,
0.0220926516,
0.0452737138,
-0.0109247994,
-0.0408684686,
0.070267491,
0.0405958444,
-0.0231624786,
-0.0012799975,
0.0305157341,
-0.0351029672,
-0.0330493189
] |
http://arxiv.org/abs/2401.15713v2 | http://arxiv.org/pdf/2401.15713v2 | Contrastive Learning and Mixture of Experts Enables Precise Vector Embeddings | 2024-01-28T17:34:42 | 2024-05-30T19:55:58 | Logan Hallee, Rohan Kapur, Arjun Patel, Jason P. Gleghorn, Bohdan Khomtchouk | cs.LG | cs.LG, cs.AI, cs.CL | The advancement of transformer neural networks has significantly elevated the
capabilities of sentence similarity models, but they struggle with highly
discriminative tasks and produce sub-optimal representations of important
documents like scientific literature. With the increased reliance on retrieval
augmentation and search, representing diverse documents as concise and
descriptive vectors is crucial. This paper improves upon the vectors embeddings
of scientific literature by assembling niche datasets using co-citations as a
similarity metric, focusing on biomedical domains. We apply a novel Mixture of
Experts (MoE) extension pipeline to pretrained BERT models, where every
multi-layer perceptron section is enlarged and copied into multiple distinct
experts. Our MoE variants perform well over $N$ scientific domains with $N$
dedicated experts, whereas standard BERT models excel in only one domain.
Notably, extending just a single transformer block to MoE captures 85% of the
benefit seen from full MoE extension at every layer. This holds promise for
versatile and efficient One-Size-Fits-All transformer networks for numerically
representing diverse inputs. Our methodology marks significant advancements in
representing scientific text and holds promise for enhancing vector database
search and compilation. | Contrastive Learning and Mixture of Experts Enables Precise Vector Embeddings: The advancement of transformer neural networks has significantly elevated the
capabilities of sentence similarity models, but they struggle with highly
discriminative tasks and produce sub-optimal representations of important
documents like scientific literature. With the increased reliance on retrieval
augmentation and search, representing diverse documents as concise and
descriptive vectors is crucial. This paper improves upon the vectors embeddings
of scientific literature by assembling niche datasets using co-citations as a
similarity metric, focusing on biomedical domains. We apply a novel Mixture of
Experts (MoE) extension pipeline to pretrained BERT models, where every
multi-layer perceptron section is enlarged and copied into multiple distinct
experts. Our MoE variants perform well over $N$ scientific domains with $N$
dedicated experts, whereas standard BERT models excel in only one domain.
Notably, extending just a single transformer block to MoE captures 85% of the
benefit seen from full MoE extension at every layer. This holds promise for
versatile and efficient One-Size-Fits-All transformer networks for numerically
representing diverse inputs. Our methodology marks significant advancements in
representing scientific text and holds promise for enhancing vector database
search and compilation. | [
0.0326586813,
0.0346534327,
0.0124969529,
-0.0284622237,
-0.013704597,
-0.0034039561,
0.0449582674,
-0.0014518682,
0.0016036681,
-0.0198471118,
-0.038030412,
-0.0173059478,
-0.0197994653,
-0.0326323546,
0.0370573997,
-0.0405284204,
0.0170800369,
0.0467061251,
0.0301242545,
-0.0122278612,
-0.0114997625,
-0.0027040003,
0.0133167133,
0.0553209707,
0.0276464056,
-0.016124066,
0.0177066922,
-0.0012538498,
0.067455478,
-0.024347445,
0.0611039586,
0.0476705506,
0.05696477,
0.0137135638,
0.0000018211,
-0.0022567643,
-0.0308064949,
0.0326575376,
0.0127462568,
0.0010182092,
0.0193089452,
-0.0324693434,
0.0580412373,
-0.0047357981,
-0.0028586925,
-0.0316008143,
0.0550762229,
0.0493673384,
0.0365198776,
0.0616958477,
-0.0168758817,
-0.0309788045,
0.0613853037,
-0.0015584235,
-0.0093957521,
0.0182201359,
0.0354081579,
-0.0155919883,
0.0496854819,
0.0591301695,
0.0147889918,
0.0386404805,
0.0147589454,
0.0163047332,
0.065319322,
0.0422604047,
0.0080978144,
0.0197136402,
0.0119553935,
0.0527837686,
-0.0426800698,
-0.0301022138,
0.0171189774,
-0.0161334965,
-0.0206526406,
0.0267276522,
-0.0246151537,
-0.0117912441,
-0.0137428138,
0.0162694287,
0.0080688745,
0.0285679344,
0.0338272005,
-0.0041849152,
-0.0510096997,
0.0016700599,
0.0377446488,
-0.0406484194,
-0.0157481655,
0.0104832696,
-0.0091822576,
0.0265182294,
0.0541659519,
0.0056003705,
-0.0203153975,
-0.0170052219,
-0.0242438987,
-0.0940130129,
0.0526663698,
-0.0326318182,
0.0148885315,
0.0182988718,
-0.0553764068,
0.0501197651,
-0.0206981748,
-0.006422997,
-0.0745226145,
0.0048168795,
-0.0403567739,
-0.0490005352,
-0.0848995447,
-0.0351103432,
-0.0046688044,
0.0231095329,
-0.0207684077,
-0.0359851792,
-0.0610306226,
0.0041172183,
0.0028478971,
-0.0148537401,
0.0312858038,
0.0601730905,
-0.0387718305,
0.0047436971,
0.0103130797,
0.0614786074,
-0.0105381245,
0.0141110197,
-0.021972416,
-0.0327559859,
-0.0153417047,
0.034483742,
0.012046922,
-0.0403241254,
0.0311508682,
0.024572432,
-0.006095177,
-0.0646398738,
0.0214338098,
0.0098107802,
-0.0227000881,
-0.0283225253,
0.0382439382,
-0.0676746145,
0.0869749561,
-0.0029808823,
-0.0091031706,
-0.0591195226,
-0.0081374841,
0.0116674826,
-0.0435302928,
0.0456273295,
-0.0484150201,
-0.0386288725,
0.0227726307,
0.0076826685,
-0.0069460627,
0.0093170498,
0.0129363574,
-0.0429681689,
0.0001673536,
0.0130950836,
0.0128763951,
-0.0227117371,
0.0110917706,
-0.0363913402,
0.0269561615,
-0.0076410347,
-0.0573930703,
0.0774754435,
0.0279825684,
0.057388097,
-0.0491759107,
-0.0235486254,
0.0211621504,
0.0602814928,
0.0422610492,
0.073430039,
0.042579975,
0.0112855714,
0.0049764193,
-0.0053358926,
0.0188655015,
0.0210228432,
0.0268775914,
0.0013603441,
0.0313475132,
0.0310056917,
-0.0519280322,
0.0105363457,
0.0247340314,
0.0064434502,
-0.0527215339,
-0.0285829678,
0.0553310663,
-0.035634771,
-0.0269173477,
0.0674493685,
0.0389640816,
-0.0422462486,
0.0013727841,
0.0144515112,
-0.0238387305,
0.047488898,
0.03819732,
-0.0693084747,
-0.0445376709,
-0.0049298666,
-0.0303356033,
0.0192616638,
0.005527636,
0.0170296151,
0.0121674938,
-0.0624705479,
-0.0267498251,
-0.0446380936,
0.0250606425,
0.0118680634,
-0.0111897215,
0.0089446427,
0.0164848361,
-0.0571006909,
0.009614178,
0.032675676,
-0.0706409961,
0.0066700908,
0.0067073819,
-0.0184847713,
0.0081156269,
0.0269016549,
-0.036218252,
0.0357395336,
0.0320675522,
-0.0707689226,
-0.0646982267,
0.0042946846,
-0.0391431488,
-0.037982326,
0.0538847148,
-0.0730328411,
0.0045682467,
0.0043407213,
-0.0249944497,
-0.0166006424,
-0.0040710564,
-0.0464401543,
0.0359206609,
-0.0220516846,
-0.0068337815,
-0.0001122708,
0.0118682273,
-0.0363595933,
0.0661674291,
-0.0180042107,
0.0360497795,
0.0269595347,
-0.0414762981,
0.0393795557,
-0.0203590151,
-0.0506460406,
-0.0568550304,
0.0497541651,
0.0253638402,
0.0184767693,
-0.045191355,
-0.0098338323,
-0.0415079519,
-0.1014020517,
0.0146382283,
0.0096435156,
-0.0086207027,
-0.0044173924,
0.0157648381,
0.0254597422,
0.0242331419,
-0.0364310816,
-0.0629324019,
-0.0221356601,
0.0030498786,
0.0510666184,
-0.0013364258,
-0.0039065788,
-0.0476141907,
-0.0415705517,
-0.0540224202,
0.0854650289,
-0.0351616181,
-0.0032486515,
-0.0161687657,
-0.0340918489,
-0.009486313,
0.0082239211,
-0.0260816701,
0.0254496839,
0.0045198142,
-0.0055140629,
0.0464998335,
0.0539404936,
0.0233755223,
0.0181017481,
0.0214790907,
0.0225935299,
-0.0503265336,
0.0220919065,
-0.0411283709,
0.0347290002,
-0.0080835419,
0.0030467615,
0.0099185994,
0.0080180997,
-0.0054984884,
-0.0390433483,
-0.0398968607,
-0.0084671173,
0.0452879332,
0.0062117218,
0.0042276508,
-0.0094675058,
0.0016040617,
0.0006937144,
-0.0026115999,
-0.0075648818,
0.1055404097,
-0.002287623,
0.0101024313,
-0.0556915998,
-0.0301741902,
-0.0277648252,
-0.0187247228,
0.0128373979,
0.0135416538,
-0.0193537232,
0.0286902469,
-0.0257088207,
0.023180129,
-0.0156852454,
-0.0215932168,
0.0161902662,
0.010798445,
0.0140760569,
0.0073112343,
-0.0559338704,
-0.0172777716,
-0.0162329413,
-0.0284208041,
0.0712657124,
0.0743796974,
-0.0550262704,
-0.0895106271,
-0.008476737,
0.0544225462,
0.0461540446,
-0.016166117,
-0.0142063638,
-0.0290180836,
0.0672906265,
0.026551187,
-0.0076767318,
0.0156487375,
-0.0269237757,
-0.0177097842,
0.0371204652,
0.0116645088,
-0.0261543877,
-0.0677058101,
-0.0164957289,
0.0094473958,
-0.0478863455,
-0.0293122083,
-0.017633941,
0.0193592291,
-0.0143590048,
0.001650458,
0.084348768,
0.02005364,
-0.0617177859,
-0.0205951538,
-0.0190224536,
0.0017631777,
-0.0077506863,
0.0434475653,
-0.0661267638,
-0.0094276099,
-0.0365228467,
-0.0050177802,
0.0335366279,
-0.0379476175,
0.0158219989,
-0.0540153123,
0.0328112394,
-0.0101340162,
0.0092224441,
-0.0425413065,
-0.0469611883,
-0.0056410464,
-0.0135498066,
-0.0126678571,
-0.0130303996,
-0.0035415108,
-0.0282864608,
0.0242646988,
0.0157465171,
-0.0792049691,
-0.0134723028,
-0.072111547,
0.015535377,
0.0115821166,
0.0281785578,
-0.0006932163,
-0.0666681007,
-0.0380202085,
-0.0518780127,
-0.0159828402,
0.0231674481,
0.0462691672,
-0.0085666561,
-0.0113872224,
-0.0333716385,
0.0216982383,
-0.0138574839,
-0.0350984521,
-0.0190053694,
0.0921360925,
-0.0094963396,
0.0009291599,
-0.0802803189,
-0.0106823696,
-0.0258878376,
0.0548161827,
0.0260938033,
-0.0383038297,
-0.026423417,
0.0181173347,
0.0429624133,
-0.0244877152,
-0.005548961,
0.0078061293,
-0.0168492738,
-0.0231729839,
-0.0244851466,
-0.0087429015,
-0.0006106306,
0.0137629658,
-0.0005046595,
0.0072141592,
0.0156885646,
-0.0288851932,
-0.0494356751,
0.0549417846,
0.0189923234,
-0.0200333856,
-0.0236707274,
-0.0015906494,
0.0890583545,
0.0050330013,
-0.0439758077,
-0.0303074438,
0.0002198848,
-0.0510935746,
-0.0242608655,
0.0586725026,
-0.0000995994,
0.0027108539,
0.0140773337,
-0.0647408441,
0.1001290679,
0.0776218101,
-0.0303379633,
-0.0119243432,
0.0300080515,
-0.0494195297,
-0.0258975588,
0.001866471,
-0.034313757,
0.0233500618,
-0.0124585889,
0.0253422484,
-0.022580659,
-0.0412191823,
0.0238959193,
-0.0237927549,
0.0701755956,
-0.0284220185,
0.0404256023,
-0.0341877341,
-0.0154851442,
-0.0033727111,
0.0246405993,
-0.030980166,
0.0004961761,
0.0152982445,
0.0117894253,
-0.0137643488,
0.0588501617,
-0.015299011,
0.0333588086,
-0.0513966344,
0.0442335345,
-0.0040177964,
-0.0030177999,
-0.0091897584,
-0.0127823791,
-0.0366022848,
-0.0289251991,
-0.0274415743,
0.0010760688,
-0.0282295961,
-0.0119228056,
0.0550879613,
0.0305835698,
-0.0328202397,
0.0498672649,
-0.1066620648,
-0.0260078534,
-0.0274196956,
-0.0218830481,
0.0590845048,
-0.0229225177,
0.025806414,
-0.0283744894,
-0.044400502,
-0.0195070971,
0.0198163502,
-0.0169786867,
0.0393232815,
0.0285090804,
0.0351675265,
-0.0071045067,
-0.0304533765,
0.0745155141,
-0.0005806555,
0.050963264,
0.0065470948,
-0.0507067479,
-0.0364744961,
-0.0236021169,
0.0353158899,
0.0500766523,
-0.0131805139,
-0.0142782731,
-0.0217360668,
-0.0009351716,
0.0880669355,
-0.0008957579,
-0.048052758,
-0.0008239901,
0.0502711236,
0.0112627726,
-0.0139928441,
-0.0296547376,
0.0068646828,
0.0297762621,
0.0306639876,
-0.0615693852,
-0.0094834259,
-6.394902327e-33,
0.0312647186,
-0.0013630358,
-0.0045842268,
-0.0537901446,
0.0104521429,
0.030059915,
-0.0255670659,
-0.0233375542,
-0.0299278274,
0.0325821228,
-0.0355552062,
0.0200891923,
-0.0134496149,
0.02874933,
0.0060686772,
0.014161488,
0.0580939651,
0.0208266322,
0.0216962546,
0.0052082469,
0.005734927,
0.0716388673,
0.0718734264,
-0.1150679514,
-0.0123347333,
0.0116457716,
0.0291303247,
0.0175910946,
0.018777119,
0.0663014352,
-0.024344245,
0.043236319,
0.0011528863,
-0.059085533,
-0.0055207647,
-0.0140968859,
-0.0559752509,
-0.012196525,
-0.0030312813,
-0.0536426418,
-0.0349837169,
-0.0340740234,
0.0539984666,
0.0040421602,
0.0208211876,
-0.0482096039,
0.082827732,
-0.0213200934,
-0.0252457485,
-0.0771313235,
-0.0160013232,
0.0210251231,
-0.0001252532,
0.0352090299,
0.0018051341,
0.034337163,
-0.0113068232,
0.0924049765,
-0.0449218825,
0.0065199197,
0.0100988308,
0.0316377729,
-0.0162372198,
-0.0215173699,
0.0240956657,
0.0487787947,
0.0195877086,
-0.0474619046,
-0.0065352162,
-0.0340887494,
0.0114722168,
0.0293432437,
0.0573966391,
0.0382622443,
0.0118000461,
0.0316279046,
-0.0707174614,
0.0284518749,
-0.0210432634,
0.0065703965,
0.0473664403,
0.02757282,
-0.0048416266,
-0.0182991568,
0.0208863076,
-0.0218069758,
-0.0117823267,
-0.0247280877,
-0.0092742434,
-0.0055010486,
-0.0083789611,
0.0440186337,
0.01998467,
-0.0337872244,
-0.0138515644,
0.0630653203,
0.0132553782,
0.0045280294,
0.0140121309,
0.0100077651,
-0.0582489409,
-0.053549774,
0.0255312249,
-0.0108110923,
0.0151196392,
0.0033994757,
0.0512352549,
0.0004716169,
-0.0488525033,
-0.0120287351,
0.017135974,
-0.0562798679,
-0.0185305681,
0.0547966361,
-0.012265563,
0.0396932289,
0.0156241283,
0.0447234064,
0.0496533252,
0.0253453422,
-0.0107577397,
-0.0038198144,
0.0365406051,
0.0280717965,
0.0201391969,
-0.0021436822,
-0.0434674323,
0.0602389239,
-0.0048349695,
-0.0701868013,
-0.0159911774,
-0.015935855,
2.76e-7,
0.0765539855,
0.0447533838,
0.0220630504,
-0.0606430732,
0.0093463687,
0.0252833683,
0.0169537421,
0.0483880602,
-0.023917336,
0.0168388821,
0.063680321,
0.0129504791,
-0.0020094393,
0.009758153,
-0.080546692,
-0.028391052,
-0.0902294517,
-0.0474326611,
-0.0415390991,
-0.0200270768,
0.0289751478,
0.0886360258,
0.0201823972,
-0.0075237104,
0.0379970856,
-0.0200094692,
0.0720333979,
-0.036527019,
0.0237499755,
0.0396220535,
0.0673806593,
0.0300477352,
0.0081470832,
0.0095611811,
0.0024389494,
0.0560948439,
0.0107065551,
0.0317423977,
-0.0109901605,
-0.0309378952,
0.0668973178,
0.007575748,
-0.0106542436,
-0.0243646111,
0.0083680386,
-0.0885538757,
-0.0251546353,
-0.0149768721,
-0.0420773551,
0.022832768,
0.0576663166,
-0.0230803955,
-0.0188533105,
0.0196778364,
0.0090835942,
0.0122596668,
-0.0024386311,
-0.0754731819,
-0.0049861083,
0.0702165142,
-0.0761644021,
-0.0139679974,
-0.016735103,
-0.0318187326,
0.0684915856,
-0.0578019992,
0.0408038795,
2.587788595e-34,
0.0500093922,
0.0223577023,
-0.0042454577,
0.0351135544,
0.0032868145,
-0.0204987414,
-0.0183951873,
-0.0175470859,
0.0343168005,
-0.0707458332,
-0.0105325896
] |
http://arxiv.org/abs/2401.15670v1 | http://arxiv.org/pdf/2401.15670v1 | YODA: Teacher-Student Progressive Learning for Language Models | 2024-01-28T14:32:15 | 2024-01-28T14:32:15 | Jianqiao Lu, Wanjun Zhong, Yufei Wang, Zhijiang Guo, Qi Zhu, Wenyong Huang, Yanlin Wang, Fei Mi, Baojun Wang, Yasheng Wang, Lifeng Shang, Xin Jiang, Qun Liu | cs.CL | cs.CL, cs.AI, cs.LG | Although large language models (LLMs) have demonstrated adeptness in a range
of tasks, they still lag behind human learning efficiency. This disparity is
often linked to the inherent human capacity to learn from basic examples,
gradually generalize and handle more complex problems, and refine their skills
with continuous feedback. Inspired by this, this paper introduces YODA, a novel
teacher-student progressive learning framework that emulates the
teacher-student education process to improve the efficacy of model fine-tuning.
The framework operates on an interactive \textit{basic-generalized-harder}
loop. The teacher agent provides tailored feedback on the student's answers,
and systematically organizes the education process. This process unfolds by
teaching the student basic examples, reinforcing understanding through
generalized questions, and then enhancing learning by posing questions with
progressively enhanced complexity. With the teacher's guidance, the student
learns to iteratively refine its answer with feedback, and forms a robust and
comprehensive understanding of the posed questions. The systematic procedural
data, which reflects the progressive learning process of humans, is then
utilized for model training. Taking math reasoning as a testbed, experiments
show that training LLaMA2 with data from YODA improves SFT with significant
performance gain (+17.01\% on GSM8K and +9.98\% on MATH). In addition, we find
that training with curriculum learning further improves learning robustness. | YODA: Teacher-Student Progressive Learning for Language Models: Although large language models (LLMs) have demonstrated adeptness in a range
of tasks, they still lag behind human learning efficiency. This disparity is
often linked to the inherent human capacity to learn from basic examples,
gradually generalize and handle more complex problems, and refine their skills
with continuous feedback. Inspired by this, this paper introduces YODA, a novel
teacher-student progressive learning framework that emulates the
teacher-student education process to improve the efficacy of model fine-tuning.
The framework operates on an interactive \textit{basic-generalized-harder}
loop. The teacher agent provides tailored feedback on the student's answers,
and systematically organizes the education process. This process unfolds by
teaching the student basic examples, reinforcing understanding through
generalized questions, and then enhancing learning by posing questions with
progressively enhanced complexity. With the teacher's guidance, the student
learns to iteratively refine its answer with feedback, and forms a robust and
comprehensive understanding of the posed questions. The systematic procedural
data, which reflects the progressive learning process of humans, is then
utilized for model training. Taking math reasoning as a testbed, experiments
show that training LLaMA2 with data from YODA improves SFT with significant
performance gain (+17.01\% on GSM8K and +9.98\% on MATH). In addition, we find
that training with curriculum learning further improves learning robustness. | [
0.0308090858,
0.0515870675,
-0.0451224968,
0.0168139897,
-0.0254121907,
0.0106551712,
0.0024226638,
-0.0404153988,
-0.0204144027,
-0.0242932756,
-0.0137722744,
-0.0178492703,
-0.0234399941,
-0.0039582788,
0.0373685397,
-0.095304139,
0.0410196111,
-0.0582090504,
0.0136251869,
-0.0053858543,
-0.0047418415,
-0.0284127556,
-0.0536297821,
0.012369561,
-0.0064827618,
-0.0282040611,
-0.0262439735,
-0.0464689955,
-0.0061715837,
-0.018655695,
0.0186707135,
0.0612641759,
0.0499711446,
0.0205422658,
0.0000022816,
-0.0488624908,
0.0075539076,
-0.0119161298,
-0.031703949,
0.0172591228,
0.0215330999,
0.0359969847,
-0.0278120786,
0.0115735512,
-0.0549468137,
0.0132282758,
0.0322922058,
0.0033880661,
0.0789737701,
0.0346141979,
0.0194253363,
-0.0312868617,
0.0423866771,
-0.0045158975,
0.0249747746,
-0.0423849374,
0.0119158793,
0.0554929301,
0.0354915261,
-0.0204707868,
-0.0425786264,
0.0278613288,
0.0749137402,
-0.0102820396,
0.0013495692,
0.0404670946,
-0.0246850234,
-0.0158045255,
0.004935916,
0.0506033711,
-0.0133447908,
-0.0307699796,
-0.0350676738,
-0.002439925,
0.0101334071,
0.0002164209,
-0.0835511759,
0.0169872511,
-0.0082939146,
0.0400365368,
-0.0076915361,
-0.0610546097,
0.0060594296,
0.0290880706,
-0.0242266972,
0.0555826686,
-0.0071315081,
-0.0377466604,
0.0182118528,
0.0135706002,
-0.0617255196,
-0.0129602514,
0.012904196,
0.0902086794,
-0.0521859862,
0.0397258997,
-0.0309716053,
-0.0405295417,
0.006655639,
-0.0221090689,
0.0657432675,
0.0710170716,
0.006833144,
0.0390885398,
-0.0127999811,
0.0577048995,
0.0285458136,
0.0446435362,
-0.0289792996,
0.0374692157,
-0.0760846511,
-0.0254771449,
-0.0486026295,
0.0385464765,
0.0230084639,
0.0146466587,
-0.0239238758,
0.0386384018,
-0.0119685847,
0.0191266071,
-0.0248611122,
-0.0158719793,
-0.032560885,
0.0006315325,
-0.0497111641,
0.0465308055,
-0.0652724653,
0.0292599816,
0.0213100296,
-0.0430844016,
-0.0158578623,
-0.0148374904,
-0.0318776928,
0.0055335276,
0.0127256084,
0.0297736786,
-0.0590982474,
-0.0203421749,
-0.0396306105,
-0.0149322972,
0.0516473949,
-0.080260843,
0.0143392738,
-0.0702170357,
0.0146731064,
0.0012475188,
-0.0193553325,
-0.0307164881,
0.017178677,
0.0151546607,
0.0140231047,
0.0572940968,
-0.0288849641,
0.0065129418,
0.0406243391,
0.022662051,
-0.0279508457,
0.1013792604,
0.0484210476,
-0.0097290901,
0.0100349765,
-0.0285149831,
0.0161309093,
0.0432328135,
0.0242477506,
0.0184283685,
-0.0012360683,
0.0054218699,
0.0369282961,
0.0345502421,
-0.0229517724,
0.0258083269,
-0.0336090364,
0.0491736792,
0.0493841805,
0.0386832617,
-0.0352953337,
0.0313075148,
0.0166589599,
0.0072349319,
0.0715006068,
0.0211562943,
-0.0814611539,
0.011860556,
0.0004779329,
0.0024347398,
-0.0195245035,
0.003607634,
-0.0118802255,
-0.0746988058,
0.0021727888,
0.0184121057,
0.0090361657,
0.0258974079,
0.0756880715,
0.0143141663,
-0.0077358759,
0.0517936535,
0.0407121889,
-0.0265973564,
0.0125453211,
-0.0703408495,
0.0481499434,
0.0358766615,
0.0414428748,
-0.0669093579,
-0.0638109446,
-0.0064836764,
0.0399246551,
0.0044210525,
0.0546105541,
-0.0272103269,
-0.0234787818,
-0.0229982026,
-0.0121432533,
-0.0107856151,
0.0202964544,
-0.0035384765,
0.051123552,
-0.0044354871,
-0.0059808521,
-0.0202744715,
-0.0155063989,
0.0428483486,
-0.0660894737,
-0.0389892124,
-0.0494695231,
-0.0160384178,
-0.0127289612,
-0.0120656267,
0.0073190136,
-0.0012959889,
0.0281616356,
-0.014512579,
0.0054422012,
-0.008368074,
-0.0581983738,
0.0000387274,
0.0709387958,
0.0161384717,
-0.0218337886,
0.0020549761,
-0.0462427139,
-0.0137626631,
0.0508169308,
-0.0049972651,
0.0177619141,
-0.023142783,
0.0291857366,
-0.0239325408,
-0.0337544717,
-0.0337198153,
0.0407977998,
0.0170517489,
0.007705336,
0.0281623788,
0.0203226171,
0.019041162,
-0.0359222442,
0.0240031872,
0.0142329177,
-0.0086016487,
0.0121215396,
-0.0238028374,
0.014222539,
-0.0886814669,
-0.0171896759,
-0.0400755666,
-0.0059936671,
-0.011837339,
0.0358830504,
0.0048660068,
-0.0227973163,
-0.0214286335,
-0.0319521613,
-0.0215082988,
0.0111004291,
-0.0072914972,
0.0083081825,
0.0015152309,
0.0353387706,
-0.0301121548,
-0.0727205947,
-0.0157460775,
-0.0360821597,
-0.0100401482,
0.026206173,
0.0499829687,
-0.0453366973,
-0.0034574126,
-0.014497472,
0.0411470421,
-0.0174658354,
-0.0388592891,
0.0487327129,
-0.0364869088,
0.028997412,
0.078739427,
-0.0225528218,
0.0572467893,
0.0503032319,
-0.0105838152,
-0.0492145754,
0.0519515462,
-0.0096829748,
-0.0516543165,
-0.0456186533,
0.0145563846,
-0.0496096201,
0.0456831232,
-0.017570883,
-0.0107196532,
-0.0425154753,
0.0399698913,
-0.0303925052,
-0.0120528536,
-0.0466333739,
-0.002861188,
-0.0108113382,
0.0010440188,
-0.0231055934,
-0.038585335,
0.089211151,
0.018073393,
-0.0163298231,
-0.0393394679,
0.0142276427,
-0.0116985617,
0.0526299626,
0.0089620687,
-0.0258511677,
-0.0041084429,
0.0076146624,
-0.0833666772,
-0.0064623347,
0.0447103791,
-0.0358757302,
0.0384715609,
-0.0805063918,
-0.0328503251,
-0.0086672409,
-0.0329761878,
-0.0618661419,
0.0168541912,
-0.0473952889,
0.0510754623,
-0.039030917,
-0.0134171369,
0.0000850033,
-0.0386280455,
0.0100685013,
-0.0009896555,
-0.0476434529,
-0.0117745493,
-0.020776229,
-0.0155157847,
0.0381796397,
0.038638439,
-0.0061336802,
-0.0724479482,
0.0538696907,
0.0412089452,
0.0677214414,
-0.0169107746,
-0.0100201191,
0.0221508648,
0.0116914054,
-0.0296298526,
0.0081547974,
0.0325074233,
0.0571478568,
-0.1099184006,
-0.0213848706,
0.1188595966,
0.0402555428,
-0.0159754772,
0.0057071405,
-0.0618263744,
0.071482271,
0.0475895368,
0.0676281825,
-0.0533792675,
-0.0177022032,
0.0106047289,
0.0448239967,
0.0256830361,
-0.0283603929,
-0.0064226007,
-0.0976447538,
0.0004105166,
0.0323682986,
-0.0520726144,
0.0309284143,
-0.0545360558,
-0.0041851359,
0.0043653394,
0.009380497,
0.026411267,
-0.0032646877,
0.0396943055,
0.0478970259,
0.0309345014,
-0.0333727039,
0.0081888288,
-0.0223503988,
0.0179288182,
0.0583920665,
-0.0218287166,
-0.0283223465,
-0.0641344488,
0.0430303961,
0.0153877288,
0.0004993425,
-0.0069086659,
0.0140162101,
0.0054663806,
-0.0187989622,
-0.0111046126,
0.0643743351,
-0.0084356209,
-0.0173261855,
0.0442442447,
0.1275981665,
0.0080565093,
-0.0017534223,
-0.0150521854,
-0.0247368477,
-0.0040298183,
0.0318126604,
0.0482627042,
-0.0458326191,
-0.0175886434,
-0.0585568398,
-0.0430119485,
-0.0156397317,
-0.0273705684,
0.0152875148,
0.0146758296,
0.0238461178,
-0.0265648514,
-0.0452725515,
0.0774930492,
0.0485563017,
0.014518884,
-0.0041752374,
0.0254084431,
0.0087690223,
-0.0264682919,
-0.0004179214,
-0.019440379,
0.023173511,
-0.0771056786,
-0.0294072628,
-0.005834525,
-0.0238249041,
-0.0107967136,
-0.0390541591,
-0.0982041359,
-0.0798873976,
0.0092511317,
0.0712126419,
0.0229443386,
0.0301640369,
0.007495075,
-0.0587754585,
0.0044514779,
0.0143763945,
-0.0325462446,
-0.0028388568,
0.0291341208,
-0.0530179963,
0.0331080928,
0.0219590627,
-0.0239508227,
0.0034643465,
0.0189750977,
0.0328313597,
0.007852098,
0.0220389608,
0.0458567068,
0.0108964108,
0.0180502851,
0.0029031739,
-0.0101615693,
-0.0314223543,
0.0451452844,
-0.0102788275,
-0.0032250518,
-0.0221309457,
0.0027105045,
-0.0023030548,
-0.0090444349,
0.0275726598,
-0.0246088002,
0.0145668946,
0.0676111877,
0.0016720173,
0.0422540642,
-0.002014152,
0.0072435569,
0.0357254259,
0.0037434511,
-0.0094498396,
-0.0025225398,
-0.0028457188,
0.0405677781,
-0.0115682846,
0.0702333674,
0.0297556929,
0.0021137709,
0.0022008871,
-0.0103127938,
-0.0331352055,
-0.0991086438,
-0.0060900524,
0.0256211963,
0.0308621489,
0.0205508266,
-0.0173588935,
-0.0726107806,
-0.0036167286,
0.0101739476,
-0.0057160002,
-0.0174068306,
-0.0260802098,
0.0421465598,
0.0346786827,
0.0516032204,
-0.0102667194,
0.0409749113,
-0.0144106746,
0.0594236739,
-0.0252192896,
0.0449144766,
-0.0226937514,
-0.0123768495,
0.0238901451,
0.0192039311,
0.0082755657,
0.0292810723,
-0.0098005775,
-0.0618702918,
-0.0421900824,
0.0346338004,
-0.0270666629,
-0.05165378,
0.04249762,
0.0575825721,
-0.0484162793,
-0.0092112273,
0.0001640534,
0.0394040495,
0.0134854,
-0.0185980834,
0.0056227962,
-6.40592479e-33,
-0.0474141352,
0.0032413553,
0.0616913214,
0.0301472023,
-0.034343034,
-0.0626280531,
-0.048499696,
0.0433692895,
-0.0234647226,
0.0226337481,
-0.0199011583,
0.0191274099,
0.0048506074,
0.0107923085,
0.0259269346,
-0.0012141772,
0.0213638917,
0.0009014499,
-0.0200310443,
0.0178904682,
-0.0218637809,
0.030966932,
0.0641344935,
-0.0518918112,
0.0045382674,
0.0081738876,
0.0025594214,
0.0480235182,
-0.0636296421,
0.0550528504,
0.0058040842,
0.0440995768,
0.0042728474,
-0.0363501795,
-0.0187535957,
-0.0105977617,
-0.0624335147,
-0.0582478233,
0.0065006875,
0.0306653231,
-0.059910655,
-0.0307471361,
0.0338107497,
-0.0187164359,
-0.0243163928,
-0.0132431658,
0.0251376033,
-0.0460826568,
-0.0150514608,
0.042089887,
-0.0577346049,
0.0056731924,
-0.0217853487,
-0.0204107556,
-0.0128364926,
0.0203704797,
-0.0207528695,
-0.0335254483,
-0.0070954747,
0.0198091473,
0.0937895551,
0.0504112989,
0.0507269241,
0.025208069,
0.0134966411,
0.0337504186,
0.0152544826,
-0.0133902095,
-0.0187135469,
-0.0182898752,
0.0618308708,
0.0246918667,
0.0647558644,
0.0017920228,
0.0321103819,
0.0021753788,
-0.0653864294,
-0.0033174532,
0.0279294234,
0.0511501245,
-0.0085406797,
0.0029523347,
-0.0213328861,
-0.0247853119,
-0.0692186132,
0.0066820076,
-0.0152059952,
-0.044931192,
0.0409598276,
-0.0038737115,
-0.0173461866,
-0.036553394,
-0.0105960937,
-0.0348296314,
-0.0249844473,
-0.003334956,
-0.0316913053,
0.03498872,
-0.0096729584,
0.0420659147,
-0.0667728931,
-0.0434016697,
-0.0141474009,
-0.0107850358,
0.0226712357,
0.0337796472,
0.0229628719,
0.0460662805,
-0.038900286,
-0.0054598413,
-0.0496037342,
0.0020233791,
0.0781513751,
0.0084359339,
0.0112718074,
0.0051782201,
-0.0104351789,
0.0716790259,
-0.0075634299,
-0.0553730614,
-0.0623728409,
0.0051863645,
-0.0127289323,
0.0141745824,
-0.0080285035,
-0.027225716,
-0.0433956124,
0.0329880007,
-0.0083499672,
-0.0908772349,
-0.0079955058,
0.0146896467,
2.876e-7,
0.0183915831,
0.0463712439,
-0.006397557,
0.028583128,
0.0224187374,
0.035069447,
-0.0599820949,
-0.0223168749,
0.0446333848,
-0.0010753772,
-0.0130542684,
0.0609646961,
0.0158690847,
0.0392636396,
-0.0705406293,
-0.0187926516,
-0.0300345439,
0.0168100111,
-0.0577174947,
0.011250658,
0.0589401498,
0.072564736,
0.0251347311,
-0.0046588699,
0.0021207582,
-0.0360511541,
0.0503613465,
-0.0098674782,
0.0246945471,
0.0251910165,
-0.023878444,
0.0107900659,
0.0112703741,
0.0011732108,
-0.0106765795,
-0.0010897283,
-0.0472830012,
0.0515308157,
0.0113341706,
0.0145935137,
0.005187165,
-0.0205279756,
-0.0000788634,
-0.0653145984,
0.0650602505,
-0.0708752722,
-0.0082427179,
0.0397919044,
0.0111379148,
-0.0401385836,
-0.0127182603,
-0.0239460785,
0.0019332846,
-0.0037525825,
0.002000754,
-0.0016308597,
-0.0140232332,
-0.0295173656,
-0.0272021014,
-0.0022276018,
-0.0285699386,
-0.0391933694,
-0.0628392249,
-0.0299899876,
-0.0253160652,
-0.0002612816,
0.0304976348,
2.350725823e-34,
0.0450539142,
-0.0120045748,
0.0596158244,
0.0573947169,
0.0100717302,
-0.0168415233,
0.0185505711,
0.0231723394,
0.0552924015,
-0.0551300198,
-0.0289589278
] |
http://arxiv.org/abs/2401.15656v3 | http://arxiv.org/pdf/2401.15656v3 | LLsM: Generative Linguistic Steganography with Large Language Model | 2024-01-28T13:21:44 | 2024-04-08T03:50:39 | Yihao Wang, Ruiqi Song, Ru Zhang, Jianyi Liu, Lingxiao Li | cs.CL | cs.CL | Linguistic Steganography (LS) tasks aim to generate steganographic text
(stego) based on secret information. Only authorized recipients can perceive
the existence of the stegos and extract secrets, thereby preserving privacy.
However, existing LS methods do not consider the controllable generation of
stegos containing specific discourses such as style, genre, and theme. And they
are difficult to simulate high-quality natural texts. As a result, the stegos
are easily perceived and detectable, compromising covert communication. This
paper proposes the LLsM, the first LS work with the Large Language Model (LLM).
Regarding open-source LLMs, we reconstruct the token generator of LLM to the
"stego generator" so that it can control the generation of stego based on the
secret. In this "stego generator", the candidate pool is encoded by range
coding, and the adjustment factor for the interval length is also given. The
secret determines the interval, thereby determining the next token. This better
simulates the distribution of natural texts and controls the adjustment of the
embedding rate. In addition, we preliminarily built an LLsM-c architecture for
closed-source LLMs. It encodes discourse to obtain high-quality prompts
containing discourse based on secrets, and generates pure natural texts
containing discourse. Experiments show that LLsM performs superior to prevalent
LS and related-task baselines regarding various kinds of concealment and
anti-steganalysis. LLsM's MAUVE surpasses baselines by 60%-80% and
anti-steganalysis exceeds baselines by 20%-30%. Notably, LLsM can also generate
longer stegos with high quality, showing its advantages in understanding and
coherence. | LLsM: Generative Linguistic Steganography with Large Language Model: Linguistic Steganography (LS) tasks aim to generate steganographic text
(stego) based on secret information. Only authorized recipients can perceive
the existence of the stegos and extract secrets, thereby preserving privacy.
However, existing LS methods do not consider the controllable generation of
stegos containing specific discourses such as style, genre, and theme. And they
are difficult to simulate high-quality natural texts. As a result, the stegos
are easily perceived and detectable, compromising covert communication. This
paper proposes the LLsM, the first LS work with the Large Language Model (LLM).
Regarding open-source LLMs, we reconstruct the token generator of LLM to the
"stego generator" so that it can control the generation of stego based on the
secret. In this "stego generator", the candidate pool is encoded by range
coding, and the adjustment factor for the interval length is also given. The
secret determines the interval, thereby determining the next token. This better
simulates the distribution of natural texts and controls the adjustment of the
embedding rate. In addition, we preliminarily built an LLsM-c architecture for
closed-source LLMs. It encodes discourse to obtain high-quality prompts
containing discourse based on secrets, and generates pure natural texts
containing discourse. Experiments show that LLsM performs superior to prevalent
LS and related-task baselines regarding various kinds of concealment and
anti-steganalysis. LLsM's MAUVE surpasses baselines by 60%-80% and
anti-steganalysis exceeds baselines by 20%-30%. Notably, LLsM can also generate
longer stegos with high quality, showing its advantages in understanding and
coherence. | [
0.03054424,
0.0270652846,
-0.01045903,
0.0566937253,
-0.0736057833,
-0.0259770565,
0.070410803,
0.0556880645,
0.0000134606,
-0.0030386557,
0.0006031485,
-0.0091191549,
-0.0007556769,
0.0665558204,
0.0262804404,
0.0133327749,
0.0105582038,
-0.0446301401,
-0.0116646355,
-0.039401859,
0.0001666021,
-0.009381705,
0.0403495505,
-0.0180326849,
0.0260720681,
-0.035034962,
-0.0276571307,
0.0112175476,
0.0059400704,
-0.0019785843,
-0.0270371959,
0.0370066501,
0.0453685001,
0.0703221187,
0.0000025376,
-0.0364556946,
-0.0341162458,
0.0083395988,
-0.0590287,
0.0446881317,
0.0497195497,
-0.023484638,
-0.0621185526,
-0.021345865,
-0.061601188,
0.0290204324,
0.0266996492,
0.0546009019,
0.0624295101,
0.087760292,
-0.0010280751,
-0.0166773628,
0.0008575171,
-0.0337610915,
0.0051470324,
-0.0231153388,
-0.0057877232,
-0.0167638548,
-0.002177001,
0.0928489864,
-0.0056010652,
0.0103778271,
0.000970758,
-0.0562154204,
0.0049704816,
0.0208793953,
-0.0161352996,
-0.0093203308,
-0.0077502159,
0.0028957601,
-0.0390930735,
-0.0185435805,
0.028139703,
0.0208056234,
0.009231139,
0.0232253019,
-0.00667017,
0.0570891201,
-0.0056544412,
-0.0026745282,
0.0278176814,
0.0161857922,
-0.0019821886,
-0.0223060865,
0.0293559004,
-0.0080389157,
-0.0008452187,
-0.009904244,
-0.0644468069,
-0.0817730427,
-0.0279084872,
-0.0626859665,
0.0476140715,
0.0090670642,
0.0370769612,
0.015002016,
-0.0561561994,
-0.0819124803,
0.0318278782,
-0.0569322184,
0.0171021689,
0.0141862882,
-0.0214387197,
0.0376278646,
-0.0469614603,
0.0745287165,
-0.0385179408,
0.007808588,
-0.0334021971,
0.0002940909,
0.0546517521,
-0.0419071279,
-0.0479135923,
0.0648876131,
-0.096038729,
-0.0527421348,
-0.0431900695,
-0.0048782779,
-0.0002791912,
0.0376696624,
0.0018371871,
-0.0077370969,
-0.0330580398,
-0.0215890147,
0.0308938064,
-0.0441932045,
-0.0023673901,
0.0065614628,
0.0037492116,
-0.0060275998,
-0.0060105501,
0.0431168266,
-0.005153711,
0.0185056329,
-0.0215637423,
0.1030820012,
-0.0032332656,
0.0016910068,
0.001289644,
-0.0448295847,
-0.0363554843,
-0.0174926035,
0.010951899,
-0.0657843724,
-0.0240570772,
-0.0222692844,
0.0010303496,
-0.0619155392,
0.0409183912,
0.0134672178,
-0.0072477669,
0.0683532804,
-0.0134130893,
-0.0385046117,
0.0246028211,
0.0389993861,
-0.0268702172,
0.0624819212,
0.0183307901,
0.0992935002,
0.0606826246,
-0.015201767,
0.0175152645,
0.0162342526,
0.021856403,
0.0259380043,
-0.0100653963,
-0.0178626794,
0.0936666802,
-0.0156110581,
0.0411572531,
0.0616176203,
-0.0188308377,
-0.1099880636,
0.0585028417,
0.0538763031,
-0.0007380781,
0.0304381363,
0.0232358184,
0.0483752638,
-0.0005697473,
0.0215661712,
0.0242162403,
0.0108134607,
-0.0422241129,
-0.0437067896,
-0.0165142678,
0.0113709336,
-0.0035506068,
-0.0001791854,
0.0160463899,
-0.0618643984,
0.012475946,
-0.0300899204,
0.039764151,
0.0414972045,
-0.0637569651,
0.0204009786,
0.0196641646,
-0.0419912338,
-0.0054550623,
-0.0439996086,
0.0534592904,
0.0892085209,
0.0384087488,
0.0355785489,
0.0941225663,
-0.0278699324,
0.0249136519,
0.0408287607,
-0.034665361,
0.020616686,
0.0028563456,
-0.0532763787,
-0.0099163176,
-0.0561681092,
-0.01446026,
0.0628893077,
-0.0048767203,
0.025657611,
0.0270989444,
-0.0082747098,
0.0215577073,
-0.0142101031,
-0.0334020443,
-0.0287101623,
0.0076032709,
-0.0383439213,
0.0499124154,
0.0138803096,
-0.0324595682,
0.0310045574,
0.0411185883,
-0.0646211728,
-0.0155392978,
-0.0415815674,
-0.0184100661,
-0.0075262785,
0.0038561032,
-0.0609708428,
0.037508484,
-0.0160716679,
-0.0136149628,
0.0013363506,
0.0123726837,
-0.0291274432,
-0.0024160212,
-0.0475377701,
0.0228258651,
-0.0092609832,
0.0514161363,
-0.0326268747,
0.0322802812,
0.0003623968,
-0.0176400598,
0.0054486678,
0.0122739803,
0.0086116111,
-0.0328687355,
0.0349040665,
-0.0087497942,
-0.0178087633,
-0.0031258003,
0.0120139979,
0.0533419624,
0.0041125044,
-0.0959569141,
-0.0671875328,
-0.0122120865,
-0.0055258218,
-0.0131164184,
0.0246236548,
-0.0153047368,
-0.0028134696,
0.0037085447,
-0.0425789207,
-0.0451045781,
-0.0258330181,
-0.0102534331,
0.0032591049,
-0.0122544914,
-0.0292032957,
0.0373065211,
0.0091872541,
0.0261820536,
0.0435550734,
-0.0267227907,
-0.0149920266,
-0.0563807115,
0.0308250152,
0.0148057994,
0.0099457698,
-0.0407738946,
-0.0544062406,
-0.0055044387,
-0.011246264,
0.047545366,
0.0733564049,
0.0645209476,
0.0347620249,
0.0154644763,
-0.022095371,
-0.016382074,
-0.0258007329,
-0.0193140879,
0.0147293974,
-0.0172457732,
-0.0341852754,
0.0090319915,
0.0156653598,
0.03798379,
0.0122724082,
-0.0314845964,
0.0492288321,
-0.0078271171,
-0.0055657346,
0.0169024169,
-0.0457495786,
-0.0131419031,
-0.0025576428,
-0.0314069539,
-0.0522379912,
0.0279493183,
-0.0571405627,
0.0133004524,
0.0069925394,
-0.015719153,
-0.0130961044,
0.0295914076,
-0.0069196522,
0.0558451042,
0.0221150722,
-0.024228489,
-0.0054212362,
-0.0089965807,
0.0533463359,
-0.062987268,
-0.0366000012,
-0.0461591929,
-0.0250435974,
-0.0048620924,
-0.01875039,
-0.0391457453,
-0.0566313118,
0.0021775167,
-0.0185540486,
0.0276358575,
-0.0340631455,
-0.0087306807,
-0.0101017095,
0.0315744393,
0.0281077996,
-0.0118591087,
0.0234784726,
-0.0200161226,
0.0111396136,
-0.0093692234,
-0.0675087646,
-0.0021471581,
-0.0089054517,
0.0609516725,
0.0731444061,
-0.0209692121,
0.0387847722,
0.0206511188,
0.0054927818,
0.0085058901,
0.038460277,
-0.0594466403,
0.0187316984,
0.0674885437,
-0.0017092852,
0.0193455033,
0.0814479962,
0.0129073244,
0.0096918484,
-0.0087602288,
-0.0053233546,
-0.0063375579,
-0.0051400959,
-0.011466681,
-0.111665383,
0.0196468644,
0.0169157907,
0.0297005735,
-0.0491149426,
-0.0036718836,
-0.0416281894,
-0.0907658488,
0.0504019223,
0.0600064024,
0.014908012,
-0.0015349921,
-0.0543433875,
-0.0455864407,
0.0635921583,
-0.0164410919,
-0.0056940131,
0.0278559253,
-0.0456748158,
-0.0188409463,
-0.0340816043,
0.012676537,
-0.0087613277,
0.0115194516,
0.0533616096,
0.0132089918,
0.0190658215,
0.0168990809,
-0.0653167889,
-0.0105769355,
0.0207271185,
-0.0013167744,
-0.0276308991,
0.0438011475,
0.0160456821,
-0.0287131332,
-0.0130743412,
0.0553285107,
0.015599967,
-0.0285759643,
0.031198943,
-0.0027455918,
0.0013109613,
-0.0551529564,
-0.0015851599,
0.027131876,
0.0205185078,
0.050353989,
0.0188673977,
-0.0273234677,
-0.0295791868,
-0.05347858,
0.0185243916,
-0.0330400951,
-0.0141432183,
-0.0085632587,
0.0221162979,
-0.0405213907,
-0.028943887,
0.0399222188,
-0.0388809294,
0.0250861701,
0.0278531369,
0.0629725382,
0.0161093958,
-0.0242351536,
-0.0408364385,
-0.0697400719,
-0.0698489249,
0.0467091091,
-0.0317626111,
-0.0211893097,
0.0110551137,
-0.0369315185,
-0.001584209,
0.0198466741,
-0.0346828923,
-0.0272948407,
0.0251124892,
0.0217912439,
0.0178884957,
0.0167753492,
-0.0561553128,
-0.0453664735,
0.0810804665,
0.0524857752,
-0.0544233471,
0.0285640117,
0.0593318604,
-0.0055692685,
0.0199556556,
0.0063546929,
-0.0190237574,
0.0133077214,
0.059562739,
-0.0489473827,
-0.0454902463,
0.011262103,
0.0073343948,
0.0056874375,
-0.0026997831,
-0.0223225374,
0.0569082908,
0.0081296396,
0.025502881,
-0.0909531787,
-0.0144057097,
-0.0227063596,
0.0051759118,
0.0301128216,
-0.0501861013,
0.0674369559,
0.0157963242,
-0.0346906334,
0.0060367668,
0.0017980436,
0.0321037285,
-0.0117670642,
0.0308184568,
0.0447228663,
0.0096478425,
-0.0147959609,
0.0247191805,
-0.023546271,
0.0384606123,
-0.0143905384,
0.0319416746,
0.0228370894,
-0.024074778,
0.0078328289,
0.0421617702,
-0.0302427486,
0.0129919779,
-0.0147202406,
0.0782159194,
0.0428198949,
0.0227252003,
0.0213337354,
-0.0147956256,
0.0133644482,
-0.0497099385,
0.0312568471,
-0.0174064171,
-0.0308997836,
0.0105140842,
0.0122306142,
-0.0232225806,
0.0436692461,
0.0330754519,
0.0229558107,
-0.0194038004,
0.0302531272,
-0.0111511583,
-0.0319360122,
-0.0036477305,
0.0060617127,
0.0447438732,
-0.0175022576,
0.0092712929,
-0.0419022478,
-0.0169979353,
0.0256723631,
-0.0120272068,
0.0216613952,
0.007137897,
-0.0037588563,
0.036714755,
-0.0471853055,
-0.0211756248,
-0.0714473799,
0.0348513164,
0.0333500057,
-0.0449304804,
0.0378849395,
-6.638353084e-33,
-0.0126903001,
-0.0805076882,
-0.0063762828,
0.0212766193,
0.0321680568,
0.0353820585,
-0.0091247214,
-0.0268902909,
0.006772995,
-0.0125943571,
-0.0206443332,
-0.010468469,
0.0167026129,
0.0177656617,
-0.0748892352,
0.040468093,
0.0377494991,
-0.0112343859,
-0.0102749309,
0.0350974984,
0.0625996143,
-0.0046679927,
-0.000156148,
-0.0928070992,
0.0140311439,
-0.0073145144,
-0.054753419,
-0.0142440191,
0.022865044,
0.0353062749,
0.0034497781,
0.0048229923,
-0.0014474895,
-0.0168171097,
-0.0189174842,
0.0826527253,
-0.0366241075,
-0.0290458985,
0.0300463699,
0.0056812689,
0.0254373513,
-0.1041190848,
-0.0178927109,
-0.0167224482,
-0.0060012857,
0.0091182608,
-0.0128579568,
-0.0337282829,
-0.0503003597,
-0.0660062209,
0.0055847815,
0.0341197252,
-0.0005996958,
-0.0128533132,
0.0042604813,
-0.008029026,
0.026982313,
-0.0141891828,
-0.0636349991,
0.0558907799,
0.0011516393,
0.0080288043,
0.0213920698,
-0.0031529306,
0.0057834918,
0.0399976298,
0.0230882298,
0.0031578662,
0.0003690728,
0.0311757438,
0.0170458909,
0.0081515973,
0.0534426495,
0.0153269814,
0.0494965464,
-0.056794364,
-0.0592111126,
-0.0235126633,
-0.0214563906,
-0.012599065,
0.0833473951,
0.0300602671,
-0.0484154075,
-0.0269560479,
-0.0444222763,
0.0071138181,
-0.00878671,
-0.0331303626,
0.0120777888,
0.0317671709,
-0.0439343117,
-0.041461423,
0.0340417549,
-0.0634139478,
-0.0081711803,
0.0032757095,
0.0093458286,
-0.0096894763,
-0.0125602819,
0.0305874404,
-0.0045169168,
-0.0308420118,
0.0602018759,
-0.0114627732,
0.0086798053,
0.0010947293,
0.029832555,
0.0076525365,
-0.0708544776,
-0.0186302271,
0.0098081678,
0.0140733887,
0.0526345335,
-0.0484286398,
-0.0079855686,
-0.0205180161,
-0.0079886699,
0.023309527,
0.0003199702,
-0.0275666788,
-0.0173037983,
0.009210119,
-0.0418356322,
0.0576267987,
-0.0251268838,
0.0425299592,
-0.0137661742,
0.0424098969,
0.0788554698,
0.0197106004,
-0.0124473702,
0.0055288319,
3.325e-7,
-0.0193306413,
0.0152224824,
0.0274796113,
0.0302480645,
0.0420279391,
0.0288325362,
0.0466348045,
0.0135097839,
-0.0291465931,
-0.044797156,
-0.0559442453,
0.0439898893,
0.0671873093,
0.0342832804,
-0.0827136412,
0.0683681294,
-0.0066940319,
-0.035608314,
-0.0628519505,
0.0272599831,
0.0661460981,
0.0802256241,
0.0153662385,
-0.0075366464,
-0.0108865388,
-0.0581809543,
0.0370964147,
-0.0212173238,
-0.0237536337,
-0.0297055487,
0.0332096145,
0.0164782722,
-0.0106477132,
0.0262725204,
-0.0079620108,
-0.0132765258,
-0.0102042565,
0.0785528496,
0.009608889,
-0.007740018,
0.0155185666,
-0.0788742676,
-0.0530088209,
-0.0179757047,
0.0353626199,
0.0225265827,
-0.040434584,
-0.053184025,
-0.0323070958,
-0.0639028773,
0.0075180619,
-0.0717560947,
-0.0205529798,
0.0124730282,
0.0275757499,
-0.0011984158,
0.0190696865,
-0.0352863483,
-0.0243201926,
-0.0415148623,
-0.023422692,
-0.0410743952,
-0.052208595,
0.0093758712,
0.0550626665,
0.0056391363,
0.0531332977,
3.210059947e-34,
0.0178689193,
-0.0618159026,
-0.0401528627,
0.0265353471,
0.0477047935,
-0.0001423311,
0.0190761592,
0.0248592999,
0.038644515,
-0.050938122,
-0.0167724062
] |
http://arxiv.org/abs/2401.15641v2 | http://arxiv.org/pdf/2401.15641v2 | PRE: A Peer Review Based Large Language Model Evaluator | 2024-01-28T12:33:14 | 2024-06-03T11:11:13 | Zhumin Chu, Qingyao Ai, Yiteng Tu, Haitao Li, Yiqun Liu | cs.IR | cs.IR, cs.CL | The impressive performance of large language models (LLMs) has attracted
considerable attention from the academic and industrial communities. Besides
how to construct and train LLMs, how to effectively evaluate and compare the
capacity of LLMs has also been well recognized as an important yet difficult
problem. Existing paradigms rely on either human annotators or model-based
evaluators to evaluate the performance of LLMs on different tasks. However,
these paradigms often suffer from high cost, low generalizability, and
inherited biases in practice, which make them incapable of supporting the
sustainable development of LLMs in long term. In order to address these issues,
inspired by the peer review systems widely used in academic publication
process, we propose a novel framework that can automatically evaluate LLMs
through a peer-review process. Specifically, for the evaluation of a specific
task, we first construct a small qualification exam to select "reviewers" from
a couple of powerful LLMs. Then, to actually evaluate the "submissions" written
by different candidate LLMs, i.e., the evaluatees, we use the reviewer LLMs to
rate or compare the submissions. The final ranking of evaluatee LLMs is
generated based on the results provided by all reviewers. We conducted
extensive experiments on text summarization tasks with eleven LLMs including
GPT-4. The results demonstrate the existence of biasness when evaluating using
a single LLM. Also, our PRE model outperforms all the baselines, illustrating
the effectiveness of the peer review mechanism. | PRE: A Peer Review Based Large Language Model Evaluator: The impressive performance of large language models (LLMs) has attracted
considerable attention from the academic and industrial communities. Besides
how to construct and train LLMs, how to effectively evaluate and compare the
capacity of LLMs has also been well recognized as an important yet difficult
problem. Existing paradigms rely on either human annotators or model-based
evaluators to evaluate the performance of LLMs on different tasks. However,
these paradigms often suffer from high cost, low generalizability, and
inherited biases in practice, which make them incapable of supporting the
sustainable development of LLMs in long term. In order to address these issues,
inspired by the peer review systems widely used in academic publication
process, we propose a novel framework that can automatically evaluate LLMs
through a peer-review process. Specifically, for the evaluation of a specific
task, we first construct a small qualification exam to select "reviewers" from
a couple of powerful LLMs. Then, to actually evaluate the "submissions" written
by different candidate LLMs, i.e., the evaluatees, we use the reviewer LLMs to
rate or compare the submissions. The final ranking of evaluatee LLMs is
generated based on the results provided by all reviewers. We conducted
extensive experiments on text summarization tasks with eleven LLMs including
GPT-4. The results demonstrate the existence of biasness when evaluating using
a single LLM. Also, our PRE model outperforms all the baselines, illustrating
the effectiveness of the peer review mechanism. | [
0.0760097802,
0.0200616103,
0.020341903,
-0.0010533181,
-0.0377135426,
0.0024919345,
0.0196188744,
-0.0260914247,
-0.0131686917,
-0.0653696284,
-0.0602677949,
-0.0150557365,
0.0078697307,
0.0103975972,
0.0203771833,
-0.0467252731,
0.0141933551,
-0.0323051624,
0.0513726026,
-0.0017858557,
0.0192259103,
-0.0053910133,
0.0179323629,
0.051028464,
0.0442205556,
-0.0076664351,
0.0301214717,
-0.0363514908,
0.0218081046,
-0.0002396906,
0.0012501326,
0.0611858144,
0.0156564657,
0.0466503985,
0.000001952,
-0.0525139533,
-0.012877889,
0.0266373586,
0.0174076501,
0.0683981478,
0.0746858194,
0.017558476,
0.0019243632,
0.0143261077,
-0.0141595034,
-0.0157702286,
0.003281977,
0.0165360309,
0.0978964344,
0.0428851508,
0.0122107966,
-0.0296403635,
0.0172820557,
-0.041648075,
-0.0415895507,
-0.0332098342,
0.0307429172,
0.0556590594,
0.0298189651,
0.034478303,
0.0134928972,
0.0206694789,
0.0195964444,
-0.0403886512,
0.0380091965,
0.0277717933,
-0.0452594124,
-0.0048352238,
0.0032012961,
0.0532101095,
0.0122694429,
-0.0116349431,
0.0116649782,
0.0037935663,
0.0117851775,
0.0240498725,
-0.0791004002,
0.0410345718,
0.0107152332,
0.0139789302,
-0.0135213193,
0.0417060852,
0.0117283873,
0.0295636375,
0.0174614862,
0.049086798,
0.0131770214,
-0.0211880188,
0.0009200614,
0.046877183,
-0.0154355308,
-0.0639374107,
0.0314561762,
0.0238555968,
-0.0257543549,
-0.0161328632,
-0.0075553521,
-0.0296642873,
0.0403978415,
-0.0399436839,
0.0336487107,
-0.0431997925,
0.0337662436,
0.0617871806,
-0.0236616619,
0.0452129394,
0.0219793059,
0.0208307877,
-0.0525275618,
-0.0017445453,
-0.0609994344,
-0.0021126668,
-0.0073592444,
0.0297387801,
-0.0076117078,
-0.0145088686,
-0.0618795939,
0.0139025245,
-0.011139621,
-0.0065766475,
0.0457542427,
0.0145061594,
0.0055707721,
0.0139983865,
-0.02316246,
0.0347208567,
-0.007895248,
-0.0031480126,
-0.0139030498,
-0.0021017662,
0.0070076804,
0.0187903065,
0.0034476593,
-0.0064397934,
0.0169158913,
0.032695841,
-0.0037103845,
-0.0128909489,
0.0173791517,
-0.0032224262,
-0.0108632501,
-0.0867046788,
-0.0316751264,
-0.0925314799,
-0.0065299068,
0.0090010436,
0.0049528079,
-0.0222898517,
-0.0025065385,
0.0033214039,
-0.0289095193,
0.0214894135,
-0.0372320339,
0.0200551227,
0.0302873608,
0.0124595258,
-0.0093357731,
0.0295378175,
0.0500778891,
0.0366792679,
0.0195959676,
-0.0058795549,
0.0439980701,
0.0404666364,
-0.0263378099,
0.0125669306,
0.0114211123,
0.0054983431,
-0.0416460373,
0.0020555772,
0.0221131798,
0.072322607,
-0.0308590587,
0.0346081145,
0.0725390017,
0.0514233373,
-0.001246511,
0.0241367761,
0.0373146608,
0.0157334954,
0.0343006812,
0.021815069,
-0.013391139,
0.0310466886,
-0.0143414084,
0.0229719672,
0.0738146007,
0.0158853158,
-0.0124931103,
-0.0537950695,
-0.0100559313,
0.0155187445,
-0.0198481176,
0.0202375744,
0.0625738353,
-0.0164667647,
-0.0248980708,
0.0142546827,
0.0028939713,
-0.0599384941,
0.0489730276,
-0.038876038,
0.0350470692,
0.0537860282,
0.0789035782,
-0.0222125258,
0.0006623286,
-0.0146984681,
-0.0300203674,
0.0318839401,
0.0676893443,
-0.0305643231,
-0.0584408119,
-0.0299056228,
-0.0468130149,
0.0046065878,
-0.0044742669,
0.0010586834,
0.0087328898,
-0.022001313,
0.0280940495,
-0.0420846939,
-0.1090967208,
0.0283983778,
-0.0135937603,
0.0034604962,
-0.0380478613,
-0.0377253518,
0.0391559377,
0.03228122,
-0.0363100544,
0.0397410244,
0.0199841466,
-0.0241224952,
-0.0314512812,
-0.009439745,
-0.0004623807,
0.0025623075,
0.0452854894,
-0.0600919835,
0.002548337,
0.0191096868,
-0.0276970584,
-0.0630278289,
0.0397150926,
-0.022878997,
-0.0030882533,
-0.031631615,
0.0106148962,
-0.0330894291,
-0.0366419256,
-0.0407258011,
0.052847553,
0.0232102629,
0.0226131603,
0.0326754376,
-0.0192001257,
0.0072742086,
-0.0098033752,
-0.06516359,
-0.0038153941,
0.0223338827,
0.0327842943,
0.0608533844,
0.0243068896,
-0.0307571199,
-0.0599875599,
-0.0982110202,
0.0058707912,
-0.0116084144,
0.0351656415,
0.0369877778,
-0.0217486676,
0.0054178368,
0.0151830725,
-0.0447378121,
-0.0347381718,
0.0289610457,
0.0049188817,
0.0112152454,
0.0153095517,
0.021809604,
-0.076078251,
-0.0240240656,
-0.0084853265,
0.0567467809,
0.0495410562,
-0.003326444,
-0.0383428149,
0.0612033382,
-0.0341206677,
-0.0217678677,
0.0210481584,
-0.0184675641,
0.0163772553,
0.0046800808,
0.0183429271,
0.0306952391,
-0.0017319256,
-0.0164499711,
0.0722182095,
-0.0232175626,
-0.0427441783,
-0.0215535108,
0.0083939629,
-0.0074466108,
-0.0097109824,
-0.0403003059,
-0.0016151562,
0.0473492444,
0.0161270313,
-0.0239382088,
-0.0367900245,
0.0469430089,
-0.0610958636,
-0.0117347464,
-0.046419695,
-0.047686711,
-0.0063087405,
0.023754131,
-0.0338137038,
-0.0191042852,
0.0640717745,
-0.0600439534,
0.0032759944,
-0.0532124639,
0.0045880764,
-0.0619879477,
0.0324411169,
0.020778602,
-0.0042203851,
-0.0310687423,
0.0183438361,
-0.0335068963,
0.0042142989,
0.0288774669,
-0.0361218527,
0.0386629365,
-0.0611817017,
0.0287899747,
-0.0313218497,
0.0010444465,
0.0078704506,
-0.0310536027,
-0.051773116,
0.0568350591,
0.0285551883,
0.0185252689,
-0.0409108438,
0.073828198,
0.0414847732,
0.0692460984,
-0.0482149161,
-0.0114881732,
-0.0169079676,
0.032519117,
0.0442495383,
-0.0239056628,
0.027159825,
-0.0511228852,
0.0075795078,
0.0400580205,
0.0247810874,
0.0108233849,
-0.0554862916,
-0.0255111977,
0.0035106395,
-0.0739794746,
-0.0039777602,
-0.0175803285,
0.0470936783,
-0.0580800362,
0.0104161119,
0.1297579557,
0.0218973383,
-0.0907329023,
-0.0303811487,
0.0248438045,
0.0006976992,
0.0238008145,
0.0069023748,
-0.0439395867,
-0.0039281053,
0.0008586722,
0.032637354,
-0.0015105944,
-0.0177426655,
0.0017288157,
-0.1168031767,
0.0592051521,
-0.000548022,
-0.0369142406,
0.0058433553,
-0.0343039557,
0.0215648301,
-0.0524351858,
-0.0040434594,
0.0328613147,
-0.0195390191,
0.0378348418,
0.0160467047,
-0.0643981844,
-0.0453471206,
0.0436759666,
-0.0544980764,
0.053704109,
0.0709148869,
0.0017583492,
-0.0246655289,
0.0194002539,
0.0394926891,
-0.0396092273,
-0.0061422759,
0.0120338844,
0.0455503985,
0.0086513208,
0.0157525204,
-0.0330228955,
0.002760299,
-0.0374794602,
0.0212469455,
0.0304780696,
0.0728317872,
0.022446977,
-0.0123366229,
-0.0239867307,
0.020928463,
-0.0518482849,
0.0674761161,
0.0234514009,
-0.0280018318,
-0.056121435,
-0.0111334473,
-0.0288296826,
-0.0359515734,
-0.0655993447,
0.0037155941,
-0.016702963,
-0.0252275597,
-0.0739742741,
-0.0447479375,
0.0334687717,
-0.0151890079,
0.0228343531,
0.0234328061,
0.0110456087,
0.0077677025,
-0.0029418478,
0.0303773768,
-0.0297187679,
0.0088112969,
-0.0530671105,
0.0018241273,
0.0765113235,
-0.0025991814,
-0.0268797874,
0.0214837175,
-0.0665804297,
-0.0400411785,
-0.0500765964,
0.0729828924,
0.020805873,
0.0159426313,
-0.0314354524,
-0.0680338293,
0.0418089628,
-0.0230674017,
-0.0103455232,
0.0089868773,
-0.0083460882,
-0.0031083019,
-0.033576142,
-0.0080160545,
-0.0626572669,
0.0425995253,
0.0016318462,
0.0206632931,
-0.0106160883,
-0.0663225353,
0.0107622175,
0.0415026993,
0.0114536248,
-0.0178190954,
0.0314663611,
-0.033326827,
-0.0309734698,
-0.0337975621,
-0.037245784,
-0.0451728888,
0.0288979951,
0.0101545714,
-0.0042452742,
0.0020671694,
0.0567159094,
-0.0123871481,
0.0379028544,
-0.0695804879,
0.0500208065,
0.0110958908,
-0.0154930642,
-0.0338275135,
-0.022002548,
-0.0438349843,
-0.0523672923,
-0.0418012589,
0.004324764,
-0.0080324858,
0.025178289,
-0.000777641,
-0.0086109126,
0.0194615945,
0.0508180484,
-0.1048978642,
0.0108943833,
-0.0451747328,
0.0358294323,
0.0526009873,
0.0279068146,
0.0843293741,
-0.0026554763,
-0.0010165444,
-0.0487708263,
0.0161117967,
0.0150035452,
0.0089241564,
-0.0013356401,
0.0226348191,
0.006506342,
0.0021101618,
0.0918457881,
0.0116639389,
-0.0081059169,
-0.002081207,
-0.0464432202,
-0.0394448303,
-0.0590614974,
-0.0463711545,
-0.0056060478,
-0.018479703,
-0.0302820057,
-0.0494830273,
-0.0195174254,
0.0012669268,
0.0375772044,
0.0088825012,
-0.0504273549,
0.0137786875,
0.0561711602,
-0.0181249399,
-0.0086010322,
-0.0632073581,
-0.0199040957,
0.0028939438,
0.0153797651,
-0.0300912485,
-6.010841139e-33,
-0.0288890861,
-0.0209620278,
0.0126775829,
0.002208886,
-0.0445103571,
0.0252187196,
-0.0474067219,
-0.0089564212,
0.0299039278,
-0.0025569657,
-0.0284373593,
-0.0011614949,
0.0030101878,
-0.00681121,
0.0626288801,
0.0299426373,
0.0704962984,
-0.0186077375,
-0.0470696315,
0.0414667502,
-0.0172388032,
-0.0110116694,
0.0537388846,
-0.0297654197,
0.0517596714,
-0.0014970112,
-0.0378132612,
-0.0186164696,
0.0074674217,
0.0492502265,
-0.0222611092,
0.0156196402,
0.0409871042,
-0.0600918457,
-0.0365798138,
-0.024952529,
-0.0482663773,
0.0160164479,
0.0445642993,
-0.0198249519,
-0.0247717258,
-0.0231732763,
0.059722349,
-0.0170013476,
-0.0406865068,
-0.0021141407,
0.0676869452,
-0.0499104373,
-0.013030014,
-0.0424338877,
-0.0240948889,
0.0415829495,
-0.0209393688,
-0.0041440362,
0.0147613818,
0.0325370021,
-0.0263944417,
0.0379055291,
-0.0343887284,
0.0326516293,
0.0519237593,
0.0403812155,
0.068776086,
0.0113813477,
-0.0031828957,
0.0592092164,
-0.0376510993,
-0.0084187435,
-0.0103377579,
-0.0485419221,
0.0069635329,
0.015838027,
0.0643750355,
0.0677986145,
-0.0036688049,
0.0472629368,
-0.0380872861,
0.0753009543,
0.008192216,
0.0363344513,
0.0672517046,
-0.008832451,
0.0035441578,
-0.0265864301,
-0.0532510541,
-0.0382190384,
-0.0183001887,
-0.0089799929,
0.0152912075,
0.0070554772,
-0.0325933993,
-0.0176498964,
0.0342710614,
-0.0496517718,
-0.0033619795,
0.0273750834,
-0.0232959948,
0.074136585,
0.001075899,
-0.0014848959,
-0.0626582876,
0.0301353391,
-0.0043109125,
0.0247849897,
-0.0077566914,
-0.0179649293,
0.0190796517,
-0.0027450235,
-0.0279492792,
0.0157938041,
-0.0119125769,
0.0035483753,
-0.0164441224,
0.0431703776,
-0.007186871,
-0.0083830673,
0.0156337526,
0.0246342458,
-0.0133010624,
0.0206485894,
-0.0070884852,
0.0381118692,
0.018798219,
0.0025325038,
-0.0243158862,
0.0279089995,
-0.007721974,
0.0222341307,
0.0107983286,
-0.0130206011,
-0.001985658,
-0.0022379251,
2.565e-7,
0.0648744255,
0.0518167429,
0.0141014671,
-0.0196510442,
0.0247935858,
0.0160391405,
-0.0396442339,
0.0075337375,
0.0485680364,
0.0019682313,
0.0449045673,
0.0178992022,
0.0472967103,
0.0310139749,
-0.080924876,
-0.0113614583,
-0.0959313512,
-0.018881999,
-0.0392133854,
-0.0397063307,
0.0153645342,
0.0593265556,
-0.0061828187,
0.0057214946,
-0.0560709611,
-0.0647252873,
0.0554780327,
-0.0351613313,
-0.0431140512,
0.0567980558,
0.0427458547,
0.0091864374,
0.0058838245,
0.0157126263,
0.0190743357,
0.0066061746,
0.0512955524,
0.0760261714,
-0.0014442906,
0.0172134265,
0.0324590057,
-0.0115709202,
0.0391237997,
-0.0525575951,
0.0594731532,
-0.0536569171,
-0.0181744825,
-0.056167908,
-0.0373966321,
-0.0273242816,
0.0174613725,
-0.0142167946,
-0.0197202545,
0.0472930409,
-0.0213249102,
-0.0211989786,
-0.0210129675,
-0.050269518,
-0.0491520129,
0.0062103625,
-0.0276638381,
-0.0790436864,
0.0011225477,
-0.0413904674,
-0.030764617,
-0.0556123145,
0.0246062633,
2.336490861e-34,
0.1002155468,
-0.0123731308,
-0.0101523055,
-0.018415045,
0.0392364115,
0.0174038839,
-0.060410589,
-0.0048783445,
-0.0029355134,
-0.0708863214,
-0.0155093363
] |
http://arxiv.org/abs/2401.15635v2 | http://arxiv.org/pdf/2401.15635v2 | RecDCL: Dual Contrastive Learning for Recommendation | 2024-01-28T11:51:09 | 2024-02-19T03:09:40 | Dan Zhang, Yangliao Geng, Wenwen Gong, Zhongang Qi, Zhiyu Chen, Xing Tang, Ying Shan, Yuxiao Dong, Jie Tang | cs.IR | cs.IR, cs.CL | Self-supervised learning (SSL) has recently achieved great success in mining
the user-item interactions for collaborative filtering. As a major paradigm,
contrastive learning (CL) based SSL helps address data sparsity in Web
platforms by contrasting the embeddings between raw and augmented data.
However, existing CL-based methods mostly focus on contrasting in a batch-wise
way, failing to exploit potential regularity in the feature dimension. This
leads to redundant solutions during the representation learning of users and
items. In this work, we investigate how to employ both batch-wise CL (BCL) and
feature-wise CL (FCL) for recommendation. We theoretically analyze the relation
between BCL and FCL, and find that combining BCL and FCL helps eliminate
redundant solutions but never misses an optimal solution. We propose a dual
contrastive learning recommendation framework -- RecDCL. In RecDCL, the FCL
objective is designed to eliminate redundant solutions on user-item positive
pairs and to optimize the uniform distributions within users and items using a
polynomial kernel for driving the representations to be orthogonal; The BCL
objective is utilized to generate contrastive embeddings on output vectors for
enhancing the robustness of the representations. Extensive experiments on four
widely-used benchmarks and one industry dataset demonstrate that RecDCL can
consistently outperform the state-of-the-art GNNs-based and SSL-based models
(with an improvement of up to 5.65\% in terms of Recall@20). The source code is
publicly available (https://github.com/THUDM/RecDCL). | RecDCL: Dual Contrastive Learning for Recommendation: Self-supervised learning (SSL) has recently achieved great success in mining
the user-item interactions for collaborative filtering. As a major paradigm,
contrastive learning (CL) based SSL helps address data sparsity in Web
platforms by contrasting the embeddings between raw and augmented data.
However, existing CL-based methods mostly focus on contrasting in a batch-wise
way, failing to exploit potential regularity in the feature dimension. This
leads to redundant solutions during the representation learning of users and
items. In this work, we investigate how to employ both batch-wise CL (BCL) and
feature-wise CL (FCL) for recommendation. We theoretically analyze the relation
between BCL and FCL, and find that combining BCL and FCL helps eliminate
redundant solutions but never misses an optimal solution. We propose a dual
contrastive learning recommendation framework -- RecDCL. In RecDCL, the FCL
objective is designed to eliminate redundant solutions on user-item positive
pairs and to optimize the uniform distributions within users and items using a
polynomial kernel for driving the representations to be orthogonal; The BCL
objective is utilized to generate contrastive embeddings on output vectors for
enhancing the robustness of the representations. Extensive experiments on four
widely-used benchmarks and one industry dataset demonstrate that RecDCL can
consistently outperform the state-of-the-art GNNs-based and SSL-based models
(with an improvement of up to 5.65\% in terms of Recall@20). The source code is
publicly available (https://github.com/THUDM/RecDCL). | [
-0.0073519768,
0.0126680275,
0.0123187294,
0.0514319576,
-0.0131946569,
-0.0167707149,
0.0170298368,
-0.0338976346,
0.0629746988,
-0.001379848,
-0.0297929868,
-0.0138865812,
-0.0169707369,
0.0579725578,
-0.0065214983,
-0.041854687,
-0.0081546912,
0.0331825241,
-0.0591510199,
-0.0120701874,
0.0175443739,
0.0075590392,
0.0051605464,
-0.0210138485,
0.0305955242,
-0.0010840473,
-0.002658193,
-0.0010325228,
0.0078338906,
-0.0205469765,
0.0001550169,
0.0306791496,
0.0225826614,
-0.0251073986,
0.0000024301,
-0.0250078607,
-0.0179741867,
-0.024991896,
0.0241795052,
0.0179923661,
0.0288809799,
0.0735512897,
0.019182561,
0.0200375672,
-0.0666687861,
-0.019765256,
0.0221010502,
0.1093500257,
0.012787791,
0.0139805768,
-0.0331028998,
-0.008974026,
0.0108350683,
-0.033647228,
0.0086751701,
0.0100635178,
0.0232271235,
-0.0518956445,
0.002700821,
0.0242298916,
-0.024462143,
-0.0116605926,
0.0359184928,
0.0011736886,
0.0684549659,
0.037339475,
0.0205508526,
0.020115478,
0.0189783499,
0.115868926,
-0.0125219021,
0.0139653813,
-0.0091368537,
-0.0189397875,
0.0324596651,
-0.0061832233,
-0.0096606044,
-0.0823566467,
-0.007500147,
0.056759011,
0.000547741,
0.0149366418,
0.0307974722,
-0.0101862745,
-0.0155265704,
0.0082257669,
0.0102042668,
-0.0228936467,
0.0003467373,
0.0129325846,
-0.0020597761,
-0.0111350622,
0.0400402248,
-0.0169922505,
0.0081661055,
-0.0339060239,
0.0044148988,
0.0255481098,
-0.0034465801,
-0.0229144804,
0.0390201807,
-0.0028950556,
-0.0583538972,
0.0327780917,
0.0026911476,
0.0146191753,
-0.0123038748,
-0.0080917925,
-0.0252261478,
0.0427734926,
-0.0118595418,
0.026334459,
0.0271572527,
0.0639512092,
-0.0205862783,
-0.0171282068,
-0.0320111141,
0.049571272,
-0.0183134768,
0.0405114815,
0.0082267346,
0.0720087811,
0.0310204718,
0.0489242859,
-0.050468009,
0.0299392808,
-0.0585487746,
0.010733665,
-0.007165493,
0.0085782409,
0.0358372442,
0.0068743816,
0.025780715,
-0.0555287823,
0.0071928985,
0.0275466181,
0.02506762,
-0.0326780118,
-0.014574144,
0.0067761908,
-0.0076570162,
-0.033832185,
-0.0546808988,
-0.079779245,
0.0351474173,
0.0017220383,
-0.0061458265,
-0.0607593097,
0.0125769936,
-0.0091774631,
-0.0120272236,
0.0934405699,
-0.0092846034,
-0.0677914843,
-0.0045063705,
0.0078215562,
-0.0402308851,
-0.0149201779,
-0.0208405014,
0.0564748235,
0.0448678434,
-0.007232246,
-0.0027678781,
-0.0398753509,
0.0218780898,
-0.0064744591,
-0.0271159355,
0.008689885,
-0.0281580649,
0.0233965311,
0.0538817011,
0.0729822144,
-0.055021774,
-0.0811801329,
0.0505530052,
0.0200683754,
0.0076908115,
0.0856869072,
0.0191777367,
0.033747986,
-0.023861954,
0.0286535788,
-0.0112802135,
0.0638025329,
-0.0055161892,
-0.0491724312,
0.0569361635,
0.0332127251,
-0.0801908895,
-0.0425864495,
0.0458795764,
-0.0445347838,
0.0553336367,
0.0253642481,
0.0560472757,
-0.0213791355,
-0.0351872593,
-0.0686848089,
0.0445618257,
-0.0298609901,
0.0132544544,
-0.0107891867,
0.0570177995,
0.0937922895,
0.0470686257,
-0.0104157757,
0.0095364843,
-0.0210493449,
-0.0451242924,
-0.0649406165,
0.0396198109,
-0.0267010462,
0.0224408433,
-0.0204748176,
-0.0708527714,
-0.0257721599,
-0.0460336246,
0.0519508198,
-0.0350671932,
0.0233441144,
0.0086532878,
0.0188821033,
0.0290384535,
-0.0268027335,
-0.0215866156,
-0.008928461,
0.0821658224,
0.0144389411,
0.0242379848,
0.0108505776,
0.0402342938,
-0.0085013965,
0.0184320863,
-0.0391668044,
-0.0528018363,
-0.0357955471,
-0.0408398956,
-0.0584917516,
0.062400043,
-0.0335384198,
0.0485367961,
0.0761731341,
-0.0025467174,
-0.0176191758,
0.0739181563,
0.0199001897,
0.0155270407,
-0.0172283743,
-0.0304673612,
0.0501596816,
0.0406468399,
-0.0164221879,
0.0362851955,
-0.0234928206,
-0.0039965478,
-0.0415735021,
0.0097435461,
0.0142411515,
-0.0325110592,
-0.0527709983,
-0.0348416492,
-0.0069350344,
0.0077161496,
0.0063028187,
-0.0051972149,
-0.0366367139,
0.0022951113,
-0.0417192169,
-0.0039354884,
0.0184451807,
-0.0124807367,
-0.0165491458,
0.0146735441,
0.0363709182,
-0.0329662748,
-0.0615876019,
-0.0025022042,
-0.0479375422,
0.0078029488,
0.0185171086,
0.0449523814,
0.0092027774,
-0.0372570753,
-0.0297387447,
0.0092163607,
0.0128628192,
-0.0374740772,
0.0216852054,
0.0018573854,
-0.011968147,
-0.0296245478,
0.0081041232,
-0.0148232114,
0.0264306143,
-0.0524840839,
-0.0251734592,
0.0223223902,
0.0217273403,
0.0042499448,
0.0414911769,
0.0321652964,
-0.0153921703,
-0.0472572781,
0.0180094577,
-0.033059895,
0.0431684703,
-0.0022038347,
0.0104591986,
-0.0461853109,
0.0072715962,
-0.0135546513,
-0.0318483599,
-0.04556958,
-0.0162172392,
0.0391892456,
-0.0079696048,
-0.0257799998,
0.0093008699,
0.0000841199,
0.0129096564,
0.0079878131,
-0.012490591,
0.0754776374,
0.0119340746,
0.0014197744,
-0.0094249211,
-0.0497312211,
-0.0387480296,
-0.0118931821,
0.0008075864,
-0.0017352871,
-0.0516213253,
0.0127291838,
-0.0261959545,
-0.0171280671,
0.0100750886,
-0.0330334492,
0.0341238566,
0.0329400934,
-0.0111136585,
-0.0006314438,
-0.0781106353,
-0.0363308489,
-0.0721357092,
0.0376889296,
-0.012762323,
-0.0125843883,
-0.0318515263,
0.0054170075,
-0.0415926538,
-0.0535856746,
0.0405507833,
-0.0094486251,
-0.0389787294,
0.0340143032,
0.029963756,
0.0369824842,
-0.0572369955,
0.0125983432,
0.0093648518,
0.0353957303,
0.051798638,
0.0405816585,
-0.0618898049,
-0.0519105978,
0.0188939255,
0.0270290058,
0.0089588193,
-0.0081381071,
0.0438520275,
-0.0130259944,
0.0346979648,
0.0008924997,
0.0782956108,
0.0128859915,
-0.1167380959,
0.0128476396,
-0.0019287069,
-0.0100193024,
0.041631382,
0.0254835617,
-0.0428084359,
-0.0037913511,
-0.0270767771,
0.0477217808,
0.0460485891,
-0.0425282344,
0.0057779225,
-0.0775011331,
0.0258730594,
0.0091651343,
0.01120965,
-0.0292103551,
-0.0795965195,
0.0009212705,
-0.0731321201,
-0.008584328,
-0.0086103566,
-0.0141621586,
0.0019378444,
-0.0174346399,
-0.0071352245,
-0.0018768592,
0.0303642992,
-0.0481152087,
0.0066240379,
0.0073932549,
0.010925171,
0.0036785717,
-0.0341179594,
-0.0412131138,
-0.0351649672,
-0.0253292136,
-0.0364883319,
0.0684337839,
-0.0265565421,
0.0117152724,
-0.0275335461,
0.0277362056,
-0.0550783612,
-0.0394651368,
-0.0150693273,
0.0837059245,
-0.0225513838,
-0.0034359999,
-0.0262915026,
0.0158106424,
0.0346059836,
0.0474970527,
0.0063263527,
-0.04765369,
0.0166838057,
-0.0194427762,
-0.0173543785,
0.0180784501,
-0.0155600673,
-0.0538210645,
0.0379552394,
0.02758082,
-0.0455973297,
-0.0412989147,
0.0180713553,
0.0303915087,
-0.0461605452,
0.0396613218,
-0.0661973432,
-0.0429894626,
-0.0338517725,
0.0167471375,
0.0140185058,
0.0267718919,
-0.0250351429,
-0.0101888683,
0.0163289309,
-0.0620270669,
-0.0087685091,
-0.0390927792,
-0.0353601277,
-0.0249054991,
-0.0392126329,
0.0718000755,
0.0029345567,
0.0431480408,
-0.0304815192,
-0.0460866913,
0.0312992148,
0.0506477244,
-0.0414093994,
-0.027373787,
0.0103771556,
0.0105545772,
0.0077564828,
0.0367470533,
-0.0310826395,
0.0080585219,
-0.0222195182,
-0.0029044009,
-0.0148919718,
-0.0532191917,
0.011844513,
-0.0209847353,
0.078730911,
0.020156838,
0.0664737672,
-0.0373612866,
-0.0114439242,
0.0096480371,
-0.0300378669,
-0.0383799523,
0.0102158831,
0.038670402,
0.0020897177,
-0.0467351638,
-0.0164741371,
-0.0127885705,
0.0299970191,
-0.0313950144,
-0.0037144322,
-0.0053829988,
-0.031642843,
-0.0062335199,
0.0613899119,
0.0131112179,
-0.0183714256,
0.0560439266,
0.029437976,
0.0102292495,
0.0423845984,
-0.0211176015,
0.0216583237,
-0.0450186841,
0.0462862477,
-0.0217614416,
0.0008040373,
-0.0088859266,
-0.0237722117,
0.0205464736,
-0.0428014584,
0.0204900187,
-0.0100888088,
-0.0101996837,
-0.0812728107,
0.0136929648,
-0.0732744709,
0.0053934297,
0.0326940417,
0.0267093237,
0.0053441715,
0.0294677839,
0.0851117522,
-0.0675338283,
-0.0008874819,
0.0088323383,
-0.0137442574,
-0.0374709815,
0.0059193089,
0.042735476,
0.0896316767,
-0.0341501683,
-0.0001198936,
0.0042737862,
-0.056457065,
0.0266391113,
-0.048325032,
0.0045545478,
0.0130656222,
0.0403343439,
0.0349209644,
0.022842098,
0.054523509,
-0.0410090089,
0.0052518318,
0.0079415943,
-0.0339914709,
-0.0195545051,
-7.106279182e-33,
0.0381327197,
-0.0222491026,
-0.0032131984,
0.0411043763,
0.027055189,
-0.005744494,
0.0208580419,
0.0106899142,
0.0173734445,
0.010720388,
-0.0086178528,
0.0038783858,
-0.0100716492,
0.0274162702,
0.0010202026,
-0.0120776258,
0.0507428944,
0.0547925793,
0.0178745035,
0.0443898812,
0.0378623419,
0.0511516854,
0.0222804733,
-0.111323297,
-0.0097894585,
-0.0109676933,
0.0199202541,
-0.0065490007,
0.0241535176,
0.065002054,
-0.022353027,
0.0515838079,
0.0093756244,
0.0043266667,
-0.0126483608,
-0.0253352337,
-0.1043857336,
-0.0382039174,
-0.0167365242,
-0.0263979938,
-0.0746538714,
-0.0170564428,
0.0116357487,
0.0089884391,
0.0007529906,
-0.0545654967,
0.0444830842,
0.0137326196,
-0.0214238558,
0.0151072601,
0.0396077223,
0.0354028903,
-0.0247208569,
-0.0125192069,
0.0137220733,
0.0325462036,
-0.0080901589,
0.0126194423,
-0.0548634008,
0.0017675937,
-0.0662372336,
0.0094586713,
-0.0064595053,
-0.0442325361,
-0.0271978099,
0.0505838692,
0.0286119431,
-0.0086989785,
0.0481629707,
-0.0199882928,
-0.0270010848,
0.0467838868,
0.0150413513,
0.0764842704,
0.0543157198,
-0.0079411278,
-0.0524187014,
0.0694084242,
0.0225847252,
-0.007891017,
0.0083422316,
0.0515919477,
-0.0249601323,
-0.0406978428,
-0.0014804917,
-0.0329575539,
-0.0533419847,
-0.0663078502,
0.003956378,
0.0391125716,
-0.0074115214,
0.0367653146,
0.0166063737,
-0.0599933304,
-0.0204999447,
0.0272998828,
-0.0362778306,
-0.0225390755,
0.0558236726,
-0.0272926241,
-0.065662168,
0.010898035,
0.0032749854,
-0.0683709458,
-0.0431497358,
0.0296791885,
0.0444663763,
0.0172713455,
-0.0360735841,
-0.01192299,
0.0191721879,
-0.0538633429,
-0.0066569131,
0.0556278192,
0.02258428,
0.0536671765,
-0.0418590009,
0.1028637215,
0.0260642022,
0.0440402664,
-0.0457212441,
-0.0627254769,
0.0227013715,
-0.0187154505,
0.0088976622,
-0.0363422856,
-0.0468629189,
0.0158002637,
0.0580274537,
-0.0867834315,
-0.0235768948,
-0.0510311984,
3.122e-7,
-0.0118931467,
-0.0096582323,
0.0117907086,
-0.0418336578,
0.01573346,
0.0390633717,
-0.0027201851,
0.0741695389,
0.0103299469,
0.0581704788,
0.0595365763,
0.0340479463,
-0.0113236746,
0.0246492196,
-0.0000902114,
0.042005565,
-0.0610322133,
0.0072551137,
-0.0242915154,
-0.022799287,
0.0355036296,
0.0723531321,
0.082192257,
-0.0292313341,
0.0526829809,
0.0307780877,
0.0710930452,
-0.0489672311,
-0.0003388877,
-0.0655488297,
0.0069425134,
0.0062327641,
-0.0079649976,
-0.000251873,
-0.0346584767,
0.0267483294,
0.0079133231,
0.0100715123,
-0.0182732325,
-0.0655952543,
0.0307487175,
0.0326827206,
0.0046322737,
-0.0171264168,
0.083523117,
-0.0087794811,
-0.032923732,
-0.0070569902,
-0.0379957296,
0.0083115166,
0.0574394874,
-0.0114711327,
-0.0048715346,
-0.0014146139,
0.0021493924,
-0.0020057624,
-0.0379137285,
-0.0630706251,
0.0248132497,
0.0133285234,
-0.0231399462,
-0.0087900329,
-0.0460432395,
0.031347543,
0.0215833373,
0.0069336421,
0.0272600576,
2.637586393e-34,
0.0174324047,
0.0296407919,
0.0050976407,
-0.0306645706,
-0.0022348517,
-0.0268629473,
-0.0373839065,
0.003747351,
0.0840637833,
-0.0451829657,
-0.0239440631
] |
http://arxiv.org/abs/2403.12973v1 | http://arxiv.org/pdf/2403.12973v1 | C Analyzer : A Static Program Analysis Tool for C Programs | 2024-01-28T11:43:16 | 2024-01-28T11:43:16 | Rajendra Kumar Solanki | cs.PL | cs.PL, cs.CL | In our times, when the world is increasingly becoming more dependent on
software programs, writing bug-free, correct programs is crucial. Program
verification based on formal methods can guarantee this by detecting run-time
errors in safety-critical systems to avoid possible adverse impacts on human
life and save time and money.
This project work tries to leverage Abstract Interpretation techniques for
static analysis of C programs. C Analyzer is a tool developed for static
analysis of C programs. This implementation of C Analyzer provides a
plug-and-play domain architecture for multiple abstract domains to be used. C
Analyzer supports four abstract domains - Interval, Octagon, Polyhedra, and Bit
Vector. We use these different domains for required precision in program
verification. C Analyzer tool uses LLVM C/C++ compiler frontend Clang API to
generate and traverse the Control Flow Graph (CFG) of a given C program. This
tool generates invariants in different abstract domains for statements in basic
blocks of CFG during CFG traversal. Using these invariants, some properties of
a program, such as dividing by zero, modulus zero, arithmetic overflow, etc.,
can be analyzed. We also use a source-to-source transformation tool, CIL
(Common Intermediate language), to transform some C constructs into simpler
constructs, such as transforming logical operators, switch statements, and
conditional operators into if-else ladders and transforming do-while and for
loops into while loops.
Using C Analyzer, C program constructs such as declarations, assignments,
binary operations (arithmetic, relational, bitwise shift, etc.), conditions
(if-else), loops (while, do while, for loop), nested conditions, and nested
loops can be analyzed. Currently, this tool does not support arrays,
structures, unions, pointers, or function calls. | C Analyzer : A Static Program Analysis Tool for C Programs: In our times, when the world is increasingly becoming more dependent on
software programs, writing bug-free, correct programs is crucial. Program
verification based on formal methods can guarantee this by detecting run-time
errors in safety-critical systems to avoid possible adverse impacts on human
life and save time and money.
This project work tries to leverage Abstract Interpretation techniques for
static analysis of C programs. C Analyzer is a tool developed for static
analysis of C programs. This implementation of C Analyzer provides a
plug-and-play domain architecture for multiple abstract domains to be used. C
Analyzer supports four abstract domains - Interval, Octagon, Polyhedra, and Bit
Vector. We use these different domains for required precision in program
verification. C Analyzer tool uses LLVM C/C++ compiler frontend Clang API to
generate and traverse the Control Flow Graph (CFG) of a given C program. This
tool generates invariants in different abstract domains for statements in basic
blocks of CFG during CFG traversal. Using these invariants, some properties of
a program, such as dividing by zero, modulus zero, arithmetic overflow, etc.,
can be analyzed. We also use a source-to-source transformation tool, CIL
(Common Intermediate language), to transform some C constructs into simpler
constructs, such as transforming logical operators, switch statements, and
conditional operators into if-else ladders and transforming do-while and for
loops into while loops.
Using C Analyzer, C program constructs such as declarations, assignments,
binary operations (arithmetic, relational, bitwise shift, etc.), conditions
(if-else), loops (while, do while, for loop), nested conditions, and nested
loops can be analyzed. Currently, this tool does not support arrays,
structures, unions, pointers, or function calls. | [
-0.0367868617,
0.0289715342,
-0.0286377575,
0.0117303003,
-0.0204389654,
0.0011409491,
-0.0126491366,
-0.0204183459,
-0.0368030742,
-0.0300493799,
0.063344337,
0.0169224292,
0.0202935915,
0.0496930517,
-0.0223164242,
0.0026073672,
0.055008363,
-0.0256924611,
-0.0336847715,
0.0600950532,
-0.0037353311,
0.0259935576,
-0.026350379,
-0.0095556583,
0.0132766105,
-0.0419252254,
0.0148584256,
0.0753553212,
0.0163454935,
-0.0481445342,
-0.0214274656,
-0.0129802041,
-0.023110548,
-0.03943583,
0.0000024331,
0.003661959,
-0.0118579818,
0.084856391,
-0.013739734,
0.0827629045,
0.0125056719,
0.0735358447,
0.0238910969,
-0.010202636,
0.0082896249,
-0.0514597148,
0.0043349196,
-0.021699179,
0.0504547656,
0.0297236294,
0.0187048949,
-0.0147571685,
0.0218446534,
0.029184062,
-0.0292179827,
-0.0262848549,
0.0402836725,
0.0451715998,
-0.0159469359,
0.0250175539,
0.0231358986,
-0.0488821119,
-0.0007979879,
-0.0306373835,
-0.0187262706,
-0.0112488773,
-0.0137559352,
-0.0215537492,
0.0018715538,
-0.021880582,
0.0029906624,
-0.0355804078,
0.0277622547,
0.01312673,
-0.0446684882,
-0.0176669657,
-0.0052281069,
-0.0011810616,
-0.0201307703,
-0.0575326234,
0.013974933,
0.0117167896,
-0.0269771367,
-0.0502186306,
-0.0043937615,
0.0412628017,
-0.0243183542,
-0.0307523441,
-0.0152271967,
0.0123224687,
0.0143854935,
-0.0662995204,
0.0224030297,
0.0129297515,
0.0285013672,
-0.0062097437,
-0.0061869449,
-0.0101214778,
0.0436602756,
-0.0489693955,
-0.0614248365,
0.0257227961,
-0.0026835946,
-0.0365198068,
-0.013405636,
0.019134121,
-0.047241129,
0.0309032928,
0.0024761276,
0.061354883,
0.0273602512,
-0.0166270807,
-0.0693549216,
0.0531058498,
0.0032159882,
0.076211445,
-0.0787944943,
-0.0069154375,
-0.0572873726,
0.0834920183,
-0.0422390252,
-0.0223041419,
0.0814301819,
0.0304328185,
0.046760954,
-0.070003055,
-0.01217815,
-0.0550462231,
-0.0486317426,
-0.004832807,
-0.0036105029,
-0.008186792,
-0.0289595984,
-0.0157993417,
-0.0076327967,
0.0539086908,
0.0147200339,
0.0387000553,
-0.0025940191,
-0.0314396322,
0.0483920425,
-0.0527160093,
-0.0336343274,
-0.0600847229,
-0.058570303,
-0.0056499755,
0.0189480875,
0.036711514,
-0.0204477347,
-0.0147100417,
0.0332155488,
-0.0057395753,
0.0157942139,
-0.0044606244,
-0.0206491239,
0.0608665906,
0.0189386811,
-0.0286683012,
-0.0153042935,
-0.0435838364,
0.0158633105,
-0.0389647521,
0.0348588191,
-0.0217496194,
-0.0127952239,
-0.0048181387,
0.0154069662,
0.0468991958,
-0.03583299,
-0.0341219455,
-0.0629050434,
0.0544634499,
-0.0045807469,
-0.0102560241,
-0.0950908512,
-0.0036877152,
-0.0170587562,
0.0257749744,
0.0524044149,
0.0290581137,
0.0316181332,
0.0021969939,
-0.0445798598,
0.0079149269,
0.0009862378,
-0.0313209817,
-0.0526330434,
0.0059086061,
0.0313165821,
-0.034638904,
-0.0232313201,
0.0402322374,
0.005395188,
-0.0274317134,
0.0019018797,
0.00335899,
0.0135330586,
0.0456886068,
0.0306156557,
-0.0712834895,
-0.0432962328,
-0.0222383905,
0.0510560311,
-0.0103448033,
0.0380207673,
-0.0667756796,
0.0130474251,
-0.0065997802,
0.0188005678,
-0.0432487279,
0.0694372207,
-0.0077538146,
-0.0137759522,
-0.0229545012,
0.0188363697,
0.0114228502,
0.0031992313,
0.0722858086,
-0.008359273,
0.0728667751,
0.001039312,
-0.033580441,
-0.0412032567,
-0.0167467091,
0.0553758554,
-0.0275538657,
0.0039648991,
0.0373922959,
-0.0606118925,
0.0229729563,
-0.0759171247,
0.0119876489,
0.0459285937,
0.0257501267,
-0.002394336,
-0.0355898701,
-0.0530066118,
0.069259055,
0.0231647491,
-0.0923151001,
-0.0394713692,
-0.0223371871,
-0.0410262048,
0.0018880345,
0.1245919317,
-0.0036986421,
-0.0204010643,
0.0321395509,
0.063825056,
-0.0699483156,
-0.0078481659,
0.0307651982,
0.0007107209,
0.0349037647,
0.0004524153,
0.0031532082,
0.0573689565,
0.0164241213,
0.0537707619,
-0.0131454486,
-0.0382879451,
-0.020731885,
-0.0084017832,
-0.0186150279,
0.0432453342,
0.0371687748,
0.0601521991,
-0.0340962932,
0.0234759152,
0.0170834027,
0.0115870219,
0.0318954177,
0.0252392292,
0.0252930615,
0.0366896726,
0.0302535184,
0.0103977965,
-0.0136959888,
-0.0226712525,
0.0164982099,
-0.0450753011,
-0.0662291571,
0.0029588607,
-0.0328160003,
-0.0760872066,
0.0646667853,
-0.0285221562,
-0.0282961763,
-0.0131325703,
0.0359278098,
0.0363700464,
-0.0002628764,
-0.0361452587,
-0.003485366,
0.000520897,
-0.0404957794,
-0.0559301525,
0.0190045908,
-0.033150021,
0.0664879903,
0.0049739936,
-0.0258648302,
-0.0370618366,
-0.0183786694,
0.077462025,
-0.0049327975,
0.0736697167,
-0.0265881401,
-0.002008799,
0.0585502125,
-0.0199853927,
-0.0011523745,
0.0205549784,
-0.0273767579,
-0.0133303013,
0.050181061,
0.0182663258,
-0.0577610135,
-0.0242576245,
-0.0099232765,
-0.0404254161,
-0.0060988739,
0.0318788514,
-0.0401280895,
-0.0258549396,
0.0051750508,
-0.0433410332,
-0.0350205451,
0.0100469803,
-0.0287450012,
0.0023933651,
0.0583592355,
-0.0344721191,
0.0530743189,
-0.0226187687,
0.0167225935,
-0.0613996796,
0.0471831635,
-0.0186969824,
-0.0313246138,
0.0266127698,
0.0244552922,
-0.011196509,
0.0324854441,
0.0214362256,
0.0112653468,
0.0164421983,
-0.0231286604,
0.0514479131,
0.0198583174,
0.0126478309,
-0.0273666438,
0.0121123139,
-0.0189350825,
0.0361864306,
-0.028530471,
-0.0324045978,
-0.0645073876,
-0.0126659591,
-0.042334713,
-0.0156029882,
0.0174087025,
-0.0004487434,
0.0158757381,
-0.0628107488,
-0.0636361018,
0.0064137829,
-0.0259357635,
-0.0106987944,
-0.0056825853,
0.0286625158,
-0.0415460058,
-0.0655667335,
-0.0362414382,
-0.0117725134,
-0.0320544094,
0.0210450031,
-0.0235159416,
-0.010164951,
0.0282732174,
0.0544128157,
-0.0907609686,
0.0192646738,
-0.0226265769,
0.0212871321,
-0.0093087554,
-0.0017144172,
-0.0027094029,
-0.0175291225,
0.0027550135,
0.0040841335,
0.0007667327,
-0.0154942479,
-0.0020316024,
0.0036547633,
0.0233677756,
-0.0567914099,
0.0194483511,
0.0041115535,
-0.0165888164,
0.0270812269,
-0.0209266748,
0.0462946743,
0.042984806,
-0.0090848263,
-0.0160794239,
-0.0034588077,
-0.0118862409,
-0.0226001944,
-0.0129939057,
0.0786201954,
-0.0204979535,
0.0278962478,
-0.040010035,
0.0198044218,
0.0178279709,
0.0461662859,
0.0115242908,
-0.0542072095,
0.0202457458,
-0.0103409719,
0.0386747308,
0.0491425171,
-0.0189086553,
-0.0592823327,
-0.0191478208,
0.0015232919,
0.0311302897,
-0.0247209482,
0.0053967414,
0.0682531446,
-0.0644766614,
-0.0557260588,
-0.0472743362,
-0.085384585,
0.0054523731,
-0.0132340062,
-0.0171496216,
-0.0197886806,
0.0733115822,
0.1047602817,
-0.0157890301,
0.0076499004,
0.0231202189,
-0.0977364033,
0.0495050624,
-0.0160024073,
-0.0929255262,
0.0268350672,
-0.0358960554,
0.0518679023,
0.0314059891,
0.0294338539,
-0.0107877329,
0.0449040085,
-0.0209916886,
0.0590353571,
-0.0173209421,
-0.0381603278,
0.0052551902,
0.0810454637,
0.004442242,
-0.0141578717,
-0.0512655787,
0.0204120036,
-0.034072116,
-0.0148371719,
-0.0137749445,
0.032225199,
-0.0224781614,
-0.0494017079,
0.0035142219,
-0.078871645,
-0.0382126048,
-0.0016986674,
0.0530675389,
-0.0608951636,
-0.0203157458,
0.0304852277,
-0.0499276333,
0.0667493492,
-0.0250115171,
0.0662947297,
-0.0173178799,
0.0411704779,
0.0343917459,
-0.0106150834,
0.0057691229,
-0.0225418638,
0.016457025,
-0.0163102727,
-0.0490461774,
-0.013216869,
-0.0155444667,
-0.0599204563,
0.0130183222,
0.0254066698,
0.07717219,
-0.0480978191,
0.0285964198,
0.0717711449,
0.0009922179,
0.0100347139,
-0.0209314004,
0.0032070414,
0.0653006285,
0.0183037017,
-0.0306401178,
0.026331136,
0.0551449656,
-0.0017114845,
0.0189100616,
0.0410041809,
0.0082833851,
0.0419550575,
0.0104102455,
0.0114623727,
0.0462833866,
0.0389107987,
0.0312598683,
0.0467583761,
-0.0314669199,
0.0209531728,
-0.0616289452,
-0.0723334476,
0.0021245098,
0.0551085062,
0.0312646702,
-0.0695796162,
-0.0266762674,
-0.0198832173,
-0.0409686528,
-0.0359481722,
-0.0283515956,
-0.013832503,
0.0405548215,
-0.0014043623,
-0.0028211903,
-0.0323603153,
-0.001605724,
-0.050708916,
-0.0586934723,
-0.0283642579,
0.0286611356,
0.0572095253,
0.0441708267,
-0.0016258742,
0.002499447,
0.0645168647,
-0.0664310232,
0.0132086975,
0.0065234778,
-0.0318695791,
-0.0500845276,
0.0019921337,
-5.715169045e-33,
-0.0231562834,
-0.0390811153,
-0.0148951076,
0.0053218575,
-0.0129810441,
-0.0139857344,
-0.0580037385,
0.0179046076,
0.0325233154,
-0.0207525231,
0.0529424176,
0.0104412436,
-0.0002250267,
-0.0666142628,
-0.0231349319,
0.0152740618,
0.0069475723,
0.0541550219,
-0.0524813011,
-0.006028635,
0.0089472365,
-0.003764793,
0.0560492203,
-0.0644116923,
-0.0108752586,
-0.0434749201,
0.0192749053,
0.0125778988,
-0.0224332344,
0.044271756,
0.0690866113,
-0.0539835952,
-0.0101765655,
0.0371386893,
-0.0127468696,
-0.0147870546,
-0.0687552318,
-0.0584138222,
0.0510223322,
0.0561557375,
0.040505182,
-0.0472358726,
-0.0022069677,
0.0443543196,
-0.0010808067,
0.0007037881,
-0.0194913466,
0.0631077439,
0.0689179674,
-0.0561877675,
0.0031039964,
0.0168492179,
-0.0550545603,
0.0279795025,
-0.0297813844,
0.0088961646,
-0.033359848,
-0.0433481522,
0.0396872312,
-0.0065947431,
0.0360173248,
-0.0283175595,
0.0050985073,
-0.0149265854,
-0.0156315211,
-0.0138175692,
-0.0055643022,
-0.00167166,
-0.0636956543,
0.0236115493,
0.018528128,
0.0243691113,
-0.039788086,
-0.0213442948,
0.061305318,
0.0015513264,
0.0060361265,
0.0296751577,
-0.0164246429,
0.0482719317,
-0.0365260355,
0.0127055198,
-0.0219640825,
-0.0154039497,
0.0322412588,
0.0394455716,
-0.0459452346,
-0.0130206216,
0.0305424891,
-0.0205668043,
-0.0375727601,
0.0326985754,
-0.0265256874,
-0.0344144106,
0.0315008648,
-0.01928268,
0.0196392126,
-0.0040644314,
-0.021868689,
0.0297182128,
0.0439335182,
-0.0421217047,
-0.0289459731,
-0.0302729774,
0.0926249623,
0.0085472232,
-0.0346145034,
0.0155816413,
-0.003374368,
0.0220146328,
0.0360457376,
-0.0184568856,
-0.0051665064,
0.0287999958,
-0.0354891494,
0.0419964306,
0.0019455971,
0.0081043895,
0.02398392,
-0.0037041537,
-0.0426372588,
0.0484249443,
0.0104289604,
0.0560038909,
-0.0192447659,
-0.0154787321,
0.0264095161,
-0.0084925406,
0.032452222,
-0.0253145024,
-0.0111974766,
0.0340814292,
3.309e-7,
0.0061806398,
0.0071504731,
0.060549356,
0.0963734388,
-0.0133249182,
-0.0073433472,
0.0312410537,
0.00436704,
0.017380923,
0.0307395663,
-0.0083917165,
0.0034940639,
0.0142019,
-0.006117072,
0.0755632967,
0.0426544435,
0.0475323312,
0.020843368,
0.0054052016,
0.0331464335,
0.0426641218,
-0.0635527745,
0.0258759018,
0.0077843582,
-0.0162953194,
-0.0277467817,
-0.0348195471,
0.0460310169,
0.026113309,
0.01373355,
-0.0529331677,
0.0100093465,
0.0226474088,
0.0192092881,
-0.0185812246,
0.0039397748,
-0.0125332251,
0.0590655915,
0.0103886407,
0.0584752262,
0.0032602511,
-0.0735437647,
0.009509651,
-0.0159268547,
0.0224711578,
0.0057124766,
0.0084818862,
-0.0271756984,
-0.0498507023,
-0.0380456261,
-0.0385156199,
0.0280086659,
0.0000763138,
0.0142139951,
0.0167133193,
-0.035871882,
-0.0143094277,
0.0429776981,
-0.0462948643,
0.0376615599,
0.0008043158,
-0.0175836347,
0.0311039556,
0.0331214517,
-0.0272764359,
0.013024929,
0.0027309465,
3.113888438e-34,
0.0512574352,
-0.037300922,
0.0051193587,
-0.0050341692,
-0.0221999846,
0.0000525061,
-0.0101029798,
-0.0070977616,
0.0127667813,
0.040778622,
-0.0231908206
] |
http://arxiv.org/abs/2401.15626v1 | http://arxiv.org/pdf/2401.15626v1 | TA&AT: Enhancing Task-Oriented Dialog with Turn-Level Auxiliary Tasks and Action-Tree Based Scheduled Sampling | 2024-01-28T11:02:23 | 2024-01-28T11:02:23 | Longxiang Liu, Xiuxing Li, Yang Feng | cs.CL | cs.CL, cs.AI | Task-oriented dialog systems have witnessed substantial progress due to
conversational pre-training techniques. Yet, two significant challenges
persist. First, most systems primarily utilize the latest turn's state label
for the generator. This practice overlooks the comprehensive value of state
labels in boosting the model's understanding for future generations. Second, an
overreliance on generated policy often leads to error accumulation, resulting
in suboptimal responses when adhering to incorrect actions. To combat these
challenges, we propose turn-level multi-task objectives for the encoder. With
the guidance of essential information from labeled intermediate states, we
establish a more robust representation for both understanding and generation.
For the decoder, we introduce an action tree-based scheduled sampling
technique. Specifically, we model the hierarchical policy as trees and utilize
the similarity between trees to sample negative policy based on scheduled
sampling, hoping the model to generate invariant responses under perturbations.
This method simulates potential pitfalls by sampling similar negative policy,
bridging the gap between task-oriented dialog training and inference. Among
methods without continual pre-training, our approach achieved state-of-the-art
(SOTA) performance on the MultiWOZ dataset series and was also competitive with
pre-trained SOTA methods. | TA&AT: Enhancing Task-Oriented Dialog with Turn-Level Auxiliary Tasks and Action-Tree Based Scheduled Sampling: Task-oriented dialog systems have witnessed substantial progress due to
conversational pre-training techniques. Yet, two significant challenges
persist. First, most systems primarily utilize the latest turn's state label
for the generator. This practice overlooks the comprehensive value of state
labels in boosting the model's understanding for future generations. Second, an
overreliance on generated policy often leads to error accumulation, resulting
in suboptimal responses when adhering to incorrect actions. To combat these
challenges, we propose turn-level multi-task objectives for the encoder. With
the guidance of essential information from labeled intermediate states, we
establish a more robust representation for both understanding and generation.
For the decoder, we introduce an action tree-based scheduled sampling
technique. Specifically, we model the hierarchical policy as trees and utilize
the similarity between trees to sample negative policy based on scheduled
sampling, hoping the model to generate invariant responses under perturbations.
This method simulates potential pitfalls by sampling similar negative policy,
bridging the gap between task-oriented dialog training and inference. Among
methods without continual pre-training, our approach achieved state-of-the-art
(SOTA) performance on the MultiWOZ dataset series and was also competitive with
pre-trained SOTA methods. | [
-0.0042785057,
0.0424997844,
0.0024541037,
0.0006570636,
0.0003801783,
0.025488941,
0.000042882,
-0.0291334111,
-0.0456023067,
-0.0308231972,
-0.0468089879,
-0.0330383331,
-0.0210707523,
0.0619028509,
0.0131705338,
-0.0760222822,
0.0201707631,
-0.0356428176,
0.0206150375,
-0.0071813385,
0.0017734792,
0.0109012527,
-0.0478707589,
0.0711121336,
0.0056100613,
-0.0441460349,
-0.0716999322,
0.0202730354,
0.0240831804,
-0.0327125862,
-0.0133005437,
0.0092145726,
0.0585062914,
0.0033151235,
0.0000020169,
-0.0091043441,
-0.0116742691,
-0.0247524362,
-0.0494550504,
-0.0319198221,
0.0167108867,
-0.0160101261,
-0.0173130482,
0.0133751426,
-0.022761602,
-0.0136554139,
0.0367445536,
0.0625522509,
0.0767737478,
0.0790551156,
-0.0054514939,
-0.0134504242,
-0.0043410109,
-0.0201424211,
0.000363871,
-0.0452004373,
-0.0176476575,
0.0202621147,
0.0057707555,
-0.0087325741,
-0.0335066281,
-0.0104668252,
0.0522551239,
-0.0419609472,
-0.0248962343,
0.0160949789,
0.0245870594,
-0.012231335,
-0.0384656824,
0.0726996064,
-0.0066934316,
0.0180698633,
-0.0021591131,
-0.0201808903,
-0.026694579,
-0.038405221,
-0.0326034166,
0.023272818,
-0.0240408368,
0.0342529602,
0.0099947918,
0.0469565652,
-0.0251994282,
-0.0239557084,
-0.0148560256,
0.0592619106,
-0.0119226715,
-0.0447703674,
0.0115418676,
0.0078879064,
0.0336727165,
-0.0494599938,
0.0472488031,
0.0117655685,
-0.0697176754,
0.0098076696,
-0.028488053,
0.0178152751,
-0.0462188199,
-0.0044491543,
0.0427772179,
0.0576910116,
-0.0208048187,
-0.0043945927,
0.0484010689,
0.058857765,
0.0084691457,
0.0301497635,
-0.0171690267,
0.0074799145,
-0.0507715568,
-0.0040426655,
-0.0669906363,
0.0222514439,
-0.0654438362,
-0.0053571975,
-0.014418372,
0.052607812,
-0.0575330332,
-0.0081639038,
0.0421080329,
0.0094459979,
-0.0745778829,
0.0062485728,
-0.0171862785,
0.0621595271,
-0.0250768065,
0.03933632,
0.0403347947,
-0.0089395968,
-0.0565937385,
0.0283837412,
0.0310218036,
-0.0019837124,
-0.0188401323,
0.0523726903,
-0.0474870838,
-0.0115313325,
-0.0123306112,
-0.0188402012,
-0.0297218449,
-0.0321116671,
0.0088258404,
0.0039168191,
0.034758985,
-0.0319705717,
0.0084941983,
0.0029121879,
-0.0123766735,
-0.0025326351,
0.0283033177,
0.0393052027,
0.0293733366,
-0.0157719217,
-0.008157406,
-0.0028510445,
-0.0273319222,
0.0587217361,
0.0454403087,
0.0326738209,
-0.0012350293,
0.028086612,
0.0205527134,
0.0329955854,
0.0206691511,
-0.0084812371,
0.0677855164,
0.0025142699,
-0.04231067,
0.0088088578,
0.0885541737,
0.0077561606,
-0.078228198,
0.0110446122,
0.0490712747,
0.0060541024,
-0.0312884487,
-0.0108891791,
0.0441052802,
0.0454736874,
-0.0104901772,
0.0057485262,
-0.0133132515,
0.0284481104,
0.0020600869,
0.0061122179,
0.0131559204,
0.0131301619,
-0.0531743281,
-0.0417438783,
0.0351155959,
-0.0094470289,
0.0114313578,
-0.0016920706,
0.0543273464,
0.0419730507,
-0.0765731782,
0.0270131733,
0.0045567732,
-0.0351011567,
0.0036854707,
-0.0559105389,
-0.0072282301,
0.0664184391,
0.0212673992,
-0.0148465643,
-0.0449028201,
-0.0272538271,
0.0067756954,
0.049068775,
-0.0039487849,
-0.045913782,
0.0836162716,
-0.0491826944,
-0.0589238964,
-0.0233206097,
0.007481995,
0.0429555438,
-0.0456526466,
-0.0069281175,
0.0030651053,
0.0042184712,
-0.0005898401,
-0.0134722581,
0.0185880084,
-0.0113277575,
0.0414682366,
0.0030415945,
0.0645222515,
0.0392749235,
-0.0243248008,
-0.0056052026,
0.0082896473,
-0.0321616679,
0.0136425458,
0.0363488346,
-0.0832166746,
-0.1091035604,
-0.0114495335,
-0.0234706067,
-0.0126468986,
0.0407993048,
-0.0471551977,
-0.0756928399,
0.0137702553,
-0.0469640344,
0.0282308273,
-0.0304393843,
0.0084583852,
0.0482880846,
-0.0119149191,
-0.0011724384,
0.0196332056,
-0.016424492,
-0.0472736657,
0.0383899845,
0.0043452182,
-0.0129330643,
-0.1001649871,
0.0130203972,
0.0269094519,
-0.0309092402,
-0.0276628062,
0.0119365873,
0.0434133112,
-0.0171190519,
0.0029225012,
-0.067047976,
-0.0327514447,
-0.0347555764,
0.0267891083,
0.0076467171,
-0.0101823565,
-0.0044869906,
-0.0156714264,
-0.0256848186,
-0.0082274992,
0.0302097313,
-0.0216852389,
-0.0008943945,
0.0238606129,
-0.0334092639,
-0.0111037325,
0.0237287916,
0.0301635694,
-0.0080912132,
-0.022148706,
-0.0339320712,
-0.0438773669,
0.0102010341,
0.0180143863,
0.0153933987,
-0.0297059808,
-0.0143097928,
0.0517822243,
0.0104630813,
0.0385160185,
0.0400518216,
-0.0153913824,
0.0534801893,
0.0033372613,
-0.0194036495,
-0.0479554199,
0.0517768115,
-0.023231823,
0.0139834341,
-0.0515061207,
-0.0170801915,
0.0371019058,
0.0794091001,
0.0128399432,
0.0521218106,
-0.0752937943,
-0.0370928235,
0.0045225653,
-0.057617873,
0.0152778085,
-0.0263280645,
-0.0473098643,
-0.0019370671,
-0.0392780527,
-0.0179479495,
0.0603421032,
-0.0090016155,
-0.0692410469,
-0.0604554377,
0.0067015071,
0.0216338728,
0.089194268,
-0.0029093826,
0.0078275697,
0.0517436117,
-0.0547020361,
-0.04032306,
0.0135840653,
0.0390218236,
-0.0109116258,
0.0012235146,
-0.0397897251,
0.0052854992,
-0.0171831306,
-0.0542780161,
-0.0375666805,
0.0132205319,
0.0033659292,
0.0110461051,
0.0260469653,
-0.0235758349,
-0.0020088181,
0.015789235,
-0.0009350871,
0.0440044105,
-0.0072465255,
-0.0102713211,
-0.0422402397,
0.0268953741,
0.0655196905,
0.0317899324,
-0.0228957664,
-0.0189240966,
0.0612251014,
0.0303434264,
0.0890407115,
-0.0278457813,
-0.0114234639,
-0.0023435082,
0.0229255874,
0.0070520113,
0.0139991026,
0.0316390507,
0.0584177598,
-0.0294876117,
-0.0396072641,
0.0955378786,
0.0059291553,
-0.0382141955,
0.013351555,
-0.0165133122,
0.0301444456,
-0.0182779618,
0.0029871729,
-0.0764988661,
-0.0373338088,
0.0072121629,
0.0448358096,
-0.0115181208,
0.0000055602,
0.0109716039,
-0.097211197,
0.0389394388,
0.0166431442,
-0.0260233283,
-0.0473068953,
0.003942492,
-0.0069593415,
-0.0202595964,
0.0282133408,
-0.0034605269,
0.0334927589,
0.0010510033,
0.0502266251,
-0.0003046216,
-0.0379142985,
0.0219490603,
-0.002291278,
0.0538068265,
0.0812307149,
0.050656151,
0.0516439974,
0.0245735645,
0.0089372667,
-0.0066392305,
-0.0127795618,
-0.0064740307,
-0.0090082353,
-0.0188232902,
0.007120579,
-0.0012854132,
0.04838074,
-0.006971621,
-0.0166294239,
-0.0026288254,
0.0562289059,
-0.0997561142,
-0.0009585608,
0.0111201983,
-0.0012487433,
0.0828609169,
0.0537305102,
0.003471571,
-0.0212466046,
-0.0228697751,
-0.0187273733,
0.0098090982,
-0.1111175939,
-0.0018806617,
-0.015391768,
0.0197668262,
0.0005555245,
-0.0536242872,
-0.0578454249,
0.0069766971,
0.0549932607,
-0.0320769846,
-0.0441628657,
0.0404739156,
0.0203614887,
-0.0132673923,
-0.0304681342,
-0.0306827389,
0.0028237873,
-0.004293615,
-0.0862619206,
0.0124606611,
0.0083194515,
-0.0160551742,
0.029386336,
-0.0559828915,
-0.0431508422,
-0.0236776806,
0.0613919199,
0.0214014091,
0.0703048185,
0.0029755607,
0.0049370807,
0.0342455916,
-0.0015732171,
-0.0322038643,
-0.0013597405,
0.0906397253,
0.0414477065,
-0.0116555998,
0.0466462038,
-0.0653473735,
0.0382963903,
-0.0345819853,
0.0534455031,
-0.0469358005,
-0.0607674606,
0.0089947935,
-0.0415503792,
0.0472322181,
0.0121712834,
-0.0098832417,
-0.0184006263,
0.0459621102,
-0.0513880625,
0.0222281348,
-0.0869659036,
0.0062123882,
-0.0113238851,
-0.0176755991,
0.0243375115,
0.0395676009,
-0.013091052,
0.0175527316,
-0.0149422493,
-0.0055864109,
0.0168565493,
-0.0010756934,
-0.0277677979,
-0.0040014461,
-0.0081392592,
-0.0107815191,
0.0656271651,
0.00060864,
0.0188463628,
0.0833119899,
0.0180872176,
-0.00018968,
0.0190810189,
0.0207165219,
-0.0770122409,
-0.0170761198,
-0.0370599069,
-0.0200409777,
0.0163108669,
0.0369222313,
0.0583135597,
-0.0309291258,
0.0395589918,
0.0066930116,
0.0525665693,
0.016104117,
-0.043274004,
0.0482258648,
0.0056189159,
0.0081772041,
-0.0226194188,
0.0793460682,
0.0162563417,
0.0351337455,
-0.0213869959,
0.0278048534,
-0.0530088916,
-0.0777202398,
-0.0362421125,
0.0205026716,
-0.0153576992,
-0.0201832242,
-0.0228069089,
-0.0124492403,
-0.0229058508,
-0.003274228,
-0.0039253007,
-0.016134914,
0.0755309239,
0.0033294589,
-0.0494643301,
-0.0411882587,
-0.0032709979,
-0.0165617876,
-0.0057363189,
-0.0005589226,
-0.0379556902,
-5.950649584e-33,
-0.0612537116,
0.0005776079,
0.0326711498,
-0.0117338123,
-0.0377321243,
-0.0257966518,
-0.0618930645,
-0.0258416943,
0.0034886294,
0.0413583331,
0.0157323666,
0.0083913039,
0.0332881883,
0.0546483099,
-0.0011231193,
-0.0233709812,
0.0760812759,
-0.0307081211,
-0.0286499821,
0.0560827404,
0.0449480377,
-0.001917633,
0.0456155613,
-0.0552571379,
0.0188196898,
0.0188704375,
-0.0254751127,
0.0080619818,
0.023398906,
0.0409132093,
0.0181740038,
0.0340506323,
-0.0101527702,
-0.0862684473,
-0.0304862484,
-0.0037293155,
-0.0710422844,
-0.0399730727,
-0.0044419058,
0.0145329228,
-0.0325241797,
-0.010799895,
-0.0249995179,
0.0328180119,
-0.0017445054,
0.0016952935,
-0.0153488237,
-0.0101951268,
-0.0121759512,
-0.045203764,
-0.052419927,
0.0036423705,
-0.0129439961,
-0.0669336095,
-0.0037118157,
0.1038603634,
-0.0007383592,
-0.0550941452,
-0.0145508219,
0.0117379837,
0.0037509012,
-0.0194130484,
0.0509549864,
0.0498713069,
0.0277304593,
0.0450430699,
0.0107696559,
-0.0023168977,
-0.0298527554,
-0.0103586735,
-0.0345774181,
0.0380527973,
-0.0244503058,
-0.0032230895,
0.000290046,
0.0281167887,
-0.0264958851,
0.036531534,
0.0574567392,
0.0726295114,
-0.0024726475,
0.0275840461,
-0.0199529417,
-0.0277002007,
0.00171254,
-0.0085068019,
0.0300704073,
-0.0272165854,
0.0345520601,
0.0618869327,
0.0112334248,
0.0414579809,
0.0693098158,
-0.0115262903,
0.0173381567,
0.0294970777,
-0.0192361698,
0.002014026,
-0.0220546164,
-0.0358539708,
-0.0523057543,
-0.0497038998,
0.0817364454,
0.0174749121,
0.016914241,
-0.0071967593,
-0.0087839644,
0.0371082239,
-0.0333939418,
0.0015803583,
-0.034475375,
0.0570527986,
-0.0046025263,
0.0344918631,
0.0411332175,
-0.0175140165,
-0.0325001217,
0.0792367011,
-0.0213524196,
-0.0319066532,
-0.026177967,
-0.0075656893,
0.0284304582,
0.0171491597,
-0.0290287193,
0.0136286439,
-0.0083841411,
-0.0429876857,
-0.0294677764,
-0.0456304662,
-0.0125275711,
-0.0040642535,
2.755e-7,
-0.0303806867,
0.0056629232,
0.0598728731,
0.0147470394,
0.0643881708,
0.0265231654,
-0.0095033217,
0.0088448916,
-0.0001249099,
-0.0227588098,
0.0345215462,
0.0087246764,
0.0629988909,
0.0253639314,
-0.0681920499,
-0.0361437015,
-0.0416100398,
-0.0241954755,
-0.0337985866,
-0.0059717726,
0.0960219502,
0.081934087,
-0.0121006155,
0.0024076954,
0.0441135131,
0.0138475718,
0.0517839789,
0.0086245956,
0.013146176,
-0.0071334247,
0.0041436763,
-0.0397636034,
-0.0070915339,
0.019312432,
-0.0247328281,
0.0160657354,
-0.0361862965,
0.0666990876,
0.0009871918,
-0.0928754285,
-0.0170139652,
0.0203101244,
0.0116727641,
0.0105234161,
0.0613660179,
-0.0709475279,
-0.0133145759,
0.0230923351,
-0.0187535007,
0.0083033219,
-0.054872103,
0.0043774284,
-0.0041106725,
0.0161337443,
0.003324345,
0.0247838255,
0.0226172656,
-0.0748093501,
0.02541475,
-0.0214535594,
-0.0410574302,
-0.05618174,
-0.0347598605,
-0.0057362933,
0.0226390455,
-0.0233613271,
-0.0321253873,
2.378055837e-34,
0.0292545762,
-0.0180932116,
0.0066047069,
0.051410988,
-0.0066996831,
-0.0018888873,
0.0033824418,
-0.0118307536,
0.051136516,
-0.0276267827,
-0.0418458544
] |
http://arxiv.org/abs/2402.01730v1 | http://arxiv.org/pdf/2402.01730v1 | Evaluating LLM -- Generated Multimodal Diagnosis from Medical Images and Symptom Analysis | 2024-01-28T09:25:12 | 2024-01-28T09:25:12 | Dimitrios P. Panagoulias, Maria Virvou, George A. Tsihrintzis | cs.CL | cs.CL, cs.AI, cs.CV | Large language models (LLMs) constitute a breakthrough state-of-the-art
Artificial Intelligence technology which is rapidly evolving and promises to
aid in medical diagnosis. However, the correctness and the accuracy of their
returns has not yet been properly evaluated. In this work, we propose an LLM
evaluation paradigm that incorporates two independent steps of a novel
methodology, namely (1) multimodal LLM evaluation via structured interactions
and (2) follow-up, domain-specific analysis based on data extracted via the
previous interactions. Using this paradigm, (1) we evaluate the correctness and
accuracy of LLM-generated medical diagnosis with publicly available multimodal
multiple-choice questions(MCQs) in the domain of Pathology and (2) proceed to a
systemic and comprehensive analysis of extracted results. We used
GPT-4-Vision-Preview as the LLM to respond to complex, medical questions
consisting of both images and text, and we explored a wide range of diseases,
conditions, chemical compounds, and related entity types that are included in
the vast knowledge domain of Pathology. GPT-4-Vision-Preview performed quite
well, scoring approximately 84\% of correct diagnoses. Next, we further
analyzed the findings of our work, following an analytical approach which
included Image Metadata Analysis, Named Entity Recognition and Knowledge
Graphs. Weaknesses of GPT-4-Vision-Preview were revealed on specific knowledge
paths, leading to a further understanding of its shortcomings in specific
areas. Our methodology and findings are not limited to the use of
GPT-4-Vision-Preview, but a similar approach can be followed to evaluate the
usefulness and accuracy of other LLMs and, thus, improve their use with further
optimization. | Evaluating LLM -- Generated Multimodal Diagnosis from Medical Images and Symptom Analysis: Large language models (LLMs) constitute a breakthrough state-of-the-art
Artificial Intelligence technology which is rapidly evolving and promises to
aid in medical diagnosis. However, the correctness and the accuracy of their
returns has not yet been properly evaluated. In this work, we propose an LLM
evaluation paradigm that incorporates two independent steps of a novel
methodology, namely (1) multimodal LLM evaluation via structured interactions
and (2) follow-up, domain-specific analysis based on data extracted via the
previous interactions. Using this paradigm, (1) we evaluate the correctness and
accuracy of LLM-generated medical diagnosis with publicly available multimodal
multiple-choice questions(MCQs) in the domain of Pathology and (2) proceed to a
systemic and comprehensive analysis of extracted results. We used
GPT-4-Vision-Preview as the LLM to respond to complex, medical questions
consisting of both images and text, and we explored a wide range of diseases,
conditions, chemical compounds, and related entity types that are included in
the vast knowledge domain of Pathology. GPT-4-Vision-Preview performed quite
well, scoring approximately 84\% of correct diagnoses. Next, we further
analyzed the findings of our work, following an analytical approach which
included Image Metadata Analysis, Named Entity Recognition and Knowledge
Graphs. Weaknesses of GPT-4-Vision-Preview were revealed on specific knowledge
paths, leading to a further understanding of its shortcomings in specific
areas. Our methodology and findings are not limited to the use of
GPT-4-Vision-Preview, but a similar approach can be followed to evaluate the
usefulness and accuracy of other LLMs and, thus, improve their use with further
optimization. | [
0.0256880876,
0.0461782664,
-0.0035296076,
0.026513895,
-0.0227641705,
0.033911638,
0.0219903868,
0.0662710369,
0.0585094281,
-0.0382181853,
-0.0146135818,
-0.0484268591,
-0.019177543,
0.0256988667,
0.0272013266,
-0.0583004542,
0.0055455994,
-0.0077444324,
-0.014118447,
-0.0030192165,
0.007006899,
-0.001009512,
-0.0195518248,
0.0322021395,
-0.0359974578,
-0.010428261,
0.0152602773,
-0.0109465579,
0.0206031706,
-0.0782883987,
-0.0441858061,
0.0424239524,
-0.0004652193,
0.0004881175,
0.0000022479,
-0.0651728213,
-0.0220430512,
0.0144480336,
0.0064879279,
-0.0265191905,
0.0386093743,
-0.0234573111,
0.0249864664,
-0.00391021,
-0.0337240286,
-0.0627622753,
0.0306168478,
0.0598436259,
0.0573015213,
0.0497130975,
-0.0275855549,
-0.0198881943,
0.0579216182,
0.0068208976,
0.0197621882,
-0.0307956878,
0.0551136322,
0.025019614,
0.0338740461,
-0.0172657315,
-0.0225499645,
0.0581468083,
-0.0148877278,
0.0231442433,
-0.0086899623,
0.0634810477,
0.0208763033,
-0.0465235226,
-0.0079732509,
0.0861546323,
0.0165317021,
0.0291257799,
0.0028068433,
-0.0098556131,
-0.0069368551,
0.0602594391,
-0.0349445604,
-0.011994564,
-0.0221800525,
-0.0007576096,
-0.0101178382,
-0.0198575836,
0.0250994842,
-0.004407783,
-0.0085696224,
0.059680216,
0.0164747071,
-0.0180099513,
0.0273084417,
-0.0043746191,
-0.0573019087,
-0.0324641578,
0.0068576513,
0.041603487,
-0.0266221594,
0.0152074294,
-0.007866716,
-0.0489939339,
0.0648611337,
-0.0450931638,
0.0249476582,
-0.0136447297,
-0.0532208197,
0.0420475639,
-0.0309051219,
0.0225551333,
-0.0513864756,
0.0362542421,
-0.0435648188,
-0.0010910305,
-0.0593482256,
-0.0075068842,
-0.0147503931,
0.03708205,
-0.0039070933,
-0.0347200036,
-0.0601559095,
0.0148962326,
-0.0014562289,
-0.0010241455,
0.0215833448,
0.0315238051,
0.0230306704,
0.0345021784,
-0.0629417822,
0.0228548441,
0.0111917555,
0.0043103495,
-0.0333500057,
-0.0085683567,
0.0618968382,
0.0354649685,
-0.0183551591,
-0.0274651032,
0.0408798419,
0.1239594966,
0.015948046,
-0.0303543154,
0.0003821346,
-0.0568311512,
-0.0606091209,
-0.0380625539,
0.0317791402,
-0.0636737123,
0.0300411042,
0.0468777493,
0.0158025287,
-0.0120171653,
0.0141696855,
-0.0009305274,
-0.0484323427,
0.0870096236,
-0.0510631204,
0.0107232081,
0.0456509367,
0.0233667735,
0.0064341645,
0.0107329609,
0.0318912677,
-0.0102450429,
0.0273370165,
-0.0570964701,
0.0207689535,
0.0328647532,
-0.0051625092,
-0.0115258181,
-0.0059173894,
-0.0064209867,
0.0047934856,
0.063069135,
-0.0150607266,
0.0798260421,
0.028239768,
-0.0214437172,
0.0404035896,
0.0750995204,
0.0352529809,
-0.0106595987,
0.0163204335,
-0.0440639332,
-0.005746346,
0.020110134,
-0.0094327321,
0.0047430061,
-0.0381287225,
-0.0092556057,
0.0013145501,
0.0223256797,
0.0092660934,
-0.0360587351,
-0.0700076073,
0.0057797474,
0.0022495412,
0.0138842901,
0.0432060994,
-0.0602010675,
-0.0420812555,
0.01670832,
0.0419242233,
-0.005398802,
-0.0276182722,
-0.0559435114,
0.0210581273,
0.0891747177,
0.0525204465,
-0.0314161293,
-0.037318211,
0.0198280588,
-0.011257981,
0.0084476518,
-0.0060957465,
-0.0034219669,
0.0012700638,
-0.0321017168,
-0.0079727396,
0.0327032246,
-0.0062300321,
-0.030809639,
-0.0249839686,
0.0171578899,
-0.0310005136,
-0.0359844156,
0.0181254987,
0.0346679017,
0.0239534993,
-0.0163679235,
-0.0094037829,
0.0288041625,
0.0233581476,
0.0163262859,
-0.0090713073,
0.0056667509,
0.0136715816,
-0.0329741761,
-0.0433032513,
0.0006986589,
-0.0184545666,
-0.0214200467,
0.013515519,
-0.0294148978,
-0.029297201,
-0.0166096315,
-0.0572251156,
-0.0285538658,
-0.0307839103,
-0.0455758348,
-0.0199214648,
-0.0304566976,
-0.0039573917,
0.0485672802,
0.0063038976,
-0.0468107574,
0.066300109,
-0.0317334048,
0.0029497885,
0.0050572157,
-0.0293867569,
0.0220225547,
-0.0146694891,
-0.0529037677,
0.0154828867,
0.0272388272,
0.0155298309,
-0.0242959615,
0.0198527705,
-0.0279580113,
-0.0097279334,
-0.1364435554,
-0.0147614162,
0.0322041847,
-0.019308174,
0.0173479822,
-0.0003066681,
-0.0119090881,
-0.0204949807,
-0.0397837348,
0.0432367288,
0.0577308275,
0.0314232446,
0.0716056377,
0.0402490832,
-0.000755723,
-0.0267554037,
-0.0017533115,
-0.0632628724,
0.0384060964,
0.0250025466,
-0.024897445,
-0.0055957278,
0.0088185342,
-0.0589029565,
0.0292157494,
0.0252753235,
-0.0086510815,
0.0067056306,
-0.0506620929,
-0.003260585,
0.0374389775,
-0.0156646091,
0.0407088324,
0.0296389293,
0.0016292458,
-0.0153186023,
0.0013269798,
-0.0363832563,
-0.024449734,
0.0044749239,
-0.0573040172,
-0.0192940403,
0.0605407022,
0.0415130667,
0.0096162921,
-0.0333777778,
0.0199067499,
-0.0219995733,
0.0066631348,
-0.0714877322,
0.0169870891,
-0.0058935368,
0.0301593374,
0.0066299043,
-0.0647974461,
0.099882029,
-0.0146826599,
0.0411862023,
-0.0120601701,
-0.023385942,
-0.0307547394,
0.0567854308,
0.0257988796,
0.0394841656,
0.0152206123,
-0.0144451037,
-0.0219587944,
-0.0116771935,
0.0048605115,
-0.0639726743,
0.0716352984,
-0.050109949,
0.0264925472,
0.0251462664,
-0.0332610607,
-0.0187708773,
0.0016147228,
-0.0231926031,
-0.0002642375,
-0.0136584099,
0.0181476623,
0.0213381536,
0.0760529637,
-0.006980476,
0.0096615134,
0.0239535496,
-0.0105723375,
-0.0120583512,
0.0775171667,
0.0293308198,
0.0092218937,
-0.0106884371,
-0.0426265709,
0.0153229339,
0.038992878,
-0.0182399023,
0.0285782441,
-0.0247496367,
0.0279091783,
-0.0154067948,
0.0170722697,
-0.0090553015,
-0.0029868239,
0.0039135711,
-0.0429471433,
0.0307051372,
0.116540648,
0.0153440926,
-0.0598122701,
-0.00273966,
0.0123890629,
-0.0636892989,
-0.0215648841,
0.0015895197,
-0.0829312056,
-0.0370692834,
-0.0412559584,
0.0034125061,
-0.0159744415,
-0.0421433039,
0.0329523645,
-0.0443937778,
0.0468900427,
-0.0308272559,
-0.0266718399,
-0.0203553271,
-0.0203690883,
-0.003357148,
-0.001483742,
-0.0160204638,
-0.0272331461,
0.0114425588,
-0.0173907634,
0.0082425755,
-0.0308165699,
-0.0581903756,
0.0591193214,
-0.0375395082,
-0.0003869164,
0.0508537851,
0.0219419636,
-0.0259891432,
-0.0324010476,
0.0113979038,
0.0113849621,
0.035729751,
-0.0215505846,
0.062676169,
-0.0018303711,
-0.0151459984,
-0.0704116523,
0.0222816747,
-0.0056447592,
0.0059522321,
-0.0131514538,
0.1241640225,
-0.050563626,
-0.0319010243,
-0.071806863,
0.000775997,
0.0178339817,
0.0409735627,
-0.0197238494,
-0.0405448675,
-0.0361138731,
0.0352113955,
0.049679555,
0.0060766707,
-0.0087980805,
0.037173856,
-0.0138127254,
0.0095249685,
-0.0719926581,
0.0011926224,
0.0392700247,
-0.0261533018,
0.0147479558,
0.0247774143,
0.001368925,
0.0204591639,
-0.0874898285,
0.0638080686,
0.0532577112,
0.0240254086,
-0.0563771538,
-0.0207217913,
0.0714044198,
-0.002488228,
0.0005403861,
0.0329338722,
-0.0527870059,
-0.0058480557,
-0.0324250087,
0.0326456539,
0.0142068509,
-0.0361789726,
-0.0067171683,
-0.0515004583,
0.0454414003,
0.0345002525,
-0.0155861173,
0.0063941139,
-0.0269947387,
-0.0042877551,
-0.0092776325,
-0.0043989364,
-0.0234237872,
0.0191429108,
-0.0184869207,
0.0047654593,
-0.0214824285,
-0.0290638562,
0.0272270422,
0.0542212687,
0.0595896691,
-0.0311180186,
-0.0253130887,
-0.0072163343,
-0.0103512229,
0.0047023967,
-0.0005096261,
-0.0030124248,
-0.0095770499,
0.0157069191,
-0.0010803201,
-0.0189552028,
-0.0064517106,
-0.0382163189,
-0.0220683962,
-0.0496146232,
0.0385679826,
0.0253425874,
-0.0502034426,
0.0513618253,
-0.0081009325,
0.0217355173,
-0.0059012487,
-0.001918993,
-0.0049415962,
0.005724276,
0.0940443426,
-0.0057048392,
0.0156721417,
-0.0015454422,
0.0627269894,
0.0114087854,
-0.0384338982,
-0.0338403545,
0.0414593145,
0.0590298958,
0.0297478009,
0.0136788813,
-0.0261469465,
-0.0178297665,
-0.0266504027,
-0.0038087771,
-0.0230342783,
-0.0011490831,
0.0973909944,
0.0124392863,
0.009798863,
-0.0108616678,
0.0773926824,
-0.0032792813,
0.015296774,
0.0013073828,
-0.0576815009,
-0.0491215289,
-0.0356291719,
0.0463135578,
-0.0025012109,
-0.0682655722,
0.0180528257,
-0.0659431741,
-0.0818082094,
0.0082108444,
-0.0016204182,
-0.0281436872,
0.0057295831,
0.0476937741,
0.0448442958,
0.0317883268,
-0.0460377634,
-0.0368994586,
0.0332322456,
-0.0618360937,
-0.0155103486,
0.0530547351,
-6.359309093e-33,
-0.0203662086,
-0.0642149374,
0.0580077358,
-0.0101743955,
0.0073380508,
0.0352066718,
-0.014860685,
0.0050613852,
0.005811648,
0.0321210362,
0.0126605211,
-0.0466728434,
-0.0012704588,
0.0480741002,
0.0061977617,
-0.0271504186,
-0.0107039511,
-0.0149873085,
-0.0275724381,
0.00994654,
0.0097955521,
0.037187621,
0.0253301393,
-0.0800890252,
-0.0118697137,
0.0055484483,
-0.0057937726,
0.0054315045,
-0.0277236588,
0.0677019432,
-0.0595448576,
0.0611636713,
0.048903361,
0.005857707,
0.0068426342,
0.0102567915,
-0.0576916449,
-0.0190789998,
0.0199395977,
-0.0110248942,
-0.0790938661,
-0.0004361964,
0.0538607538,
-0.0215662774,
0.0019294831,
-0.0031596189,
0.047257416,
-0.031649705,
0.0103935078,
-0.0400643833,
-0.018874798,
0.0348146111,
0.0319779031,
0.0042012325,
0.0106266383,
0.0016924398,
-0.009802781,
0.0467382222,
-0.0330251418,
0.0217308272,
0.0310545601,
-0.0150584392,
0.059401121,
0.007627286,
-0.0148878079,
0.0476781763,
0.0205092523,
-0.0628969744,
-0.0338196829,
0.0079977699,
-0.031208355,
0.0719110891,
0.0534843728,
0.0230346266,
-0.0159861092,
0.00982669,
-0.0597532056,
0.0210813414,
-0.0655516163,
0.028959142,
0.0246550962,
-0.0404874831,
-0.0276148729,
-0.0080081569,
-0.0290263481,
0.0237472877,
-0.0499229208,
-0.0425953977,
0.0485263169,
-0.0228857901,
0.0046626423,
0.0228168909,
0.0270060841,
-0.0322186388,
-0.0302919075,
-0.0503844358,
0.0283350293,
0.0557382107,
-0.0159450453,
0.0243431889,
-0.028461853,
-0.0273519866,
0.0262103379,
-0.0344835445,
0.0452610403,
0.0133580742,
0.0045591528,
-0.0217814036,
-0.0349605829,
-0.0045547877,
0.0037537154,
-0.0088381227,
-0.0182098206,
0.0073529943,
0.0126109552,
0.0236753542,
-0.0267662723,
0.0411781669,
0.0261862613,
0.0307811294,
-0.0335895754,
-0.0650886148,
-0.0111972876,
-0.0107726557,
-0.0256495103,
0.0035845861,
-0.0119600836,
-0.0228034891,
0.0448297523,
-0.054132048,
-0.0081717027,
0.0319760367,
3.031e-7,
0.1116565615,
0.0636248589,
-0.0207559206,
-0.0044180141,
0.0334877074,
0.0211016517,
-0.0188890286,
0.0271898899,
-0.0098158885,
-0.0371856801,
0.0314270817,
0.0192089006,
0.0103296973,
0.0528174117,
-0.0144198202,
0.0106772073,
-0.0058140471,
-0.045578897,
-0.0632012933,
0.0415052511,
0.0889555067,
0.0578546785,
-0.0390869975,
-0.0691937059,
0.0476455055,
-0.0236704014,
0.0481563061,
-0.0502552614,
-0.0097587472,
0.0099232038,
-0.0035535553,
0.0125402128,
0.0231132992,
0.0875124484,
0.0061204582,
-0.0426447764,
0.031042669,
0.0332725793,
-0.0351147018,
-0.0220793765,
-0.0005676674,
-0.0132989353,
0.0216483921,
-0.0690435991,
0.083463937,
-0.0506782085,
-0.0317778289,
0.0099358428,
0.0151370727,
-0.0549735054,
0.0448420122,
-0.0348525681,
-0.0421482921,
0.035129562,
0.0216508973,
0.0076020644,
-0.0232873708,
-0.0723045692,
0.0160851106,
0.0221610852,
-0.0442204773,
-0.0549180992,
0.0246834513,
-0.059900742,
0.0067350701,
-0.0896354094,
-0.0263551977,
2.63729183e-34,
0.0132361166,
0.019934848,
0.0157318749,
-0.0037550335,
0.0148305716,
0.0078810174,
-0.0111994622,
-0.0035423059,
-0.0293792933,
-0.1187281534,
-0.0024073529
] |
http://arxiv.org/abs/2402.01729v3 | http://arxiv.org/pdf/2402.01729v3 | Contextualization Distillation from Large Language Model for Knowledge Graph Completion | 2024-01-28T08:56:49 | 2024-02-24T07:01:22 | Dawei Li, Zhen Tan, Tianlong Chen, Huan Liu | cs.CL | cs.CL, cs.AI | While textual information significantly enhances the performance of
pre-trained language models (PLMs) in knowledge graph completion (KGC), the
static and noisy nature of existing corpora collected from Wikipedia articles
or synsets definitions often limits the potential of PLM-based KGC models. To
surmount these challenges, we introduce the Contextualization Distillation
strategy, a versatile plug-in-and-play approach compatible with both
discriminative and generative KGC frameworks. Our method begins by instructing
large language models (LLMs) to transform compact, structural triplets into
context-rich segments. Subsequently, we introduce two tailored auxiliary tasks,
reconstruction and contextualization, allowing smaller KGC models to assimilate
insights from these enriched triplets. Comprehensive evaluations across diverse
datasets and KGC techniques highlight the efficacy and adaptability of our
approach, revealing consistent performance enhancements irrespective of
underlying pipelines or architectures. Moreover, our analysis makes our method
more explainable and provides insight into generating path selection, as well
as the choosing of suitable distillation tasks. All the code and data in this
work will be released at
https://github.com/David-Li0406/Contextulization-Distillation | Contextualization Distillation from Large Language Model for Knowledge Graph Completion: While textual information significantly enhances the performance of
pre-trained language models (PLMs) in knowledge graph completion (KGC), the
static and noisy nature of existing corpora collected from Wikipedia articles
or synsets definitions often limits the potential of PLM-based KGC models. To
surmount these challenges, we introduce the Contextualization Distillation
strategy, a versatile plug-in-and-play approach compatible with both
discriminative and generative KGC frameworks. Our method begins by instructing
large language models (LLMs) to transform compact, structural triplets into
context-rich segments. Subsequently, we introduce two tailored auxiliary tasks,
reconstruction and contextualization, allowing smaller KGC models to assimilate
insights from these enriched triplets. Comprehensive evaluations across diverse
datasets and KGC techniques highlight the efficacy and adaptability of our
approach, revealing consistent performance enhancements irrespective of
underlying pipelines or architectures. Moreover, our analysis makes our method
more explainable and provides insight into generating path selection, as well
as the choosing of suitable distillation tasks. All the code and data in this
work will be released at
https://github.com/David-Li0406/Contextulization-Distillation | [
0.0144073209,
0.053658627,
0.0105070462,
0.0619417652,
0.0153768146,
0.0217572507,
-0.0491826795,
-0.0036788585,
0.0305092055,
-0.035886731,
0.0082296701,
0.0165394302,
-0.0088275513,
0.0131702758,
0.0269188099,
-0.0251361262,
0.023143081,
-0.0156252757,
-0.0201198719,
-0.020469144,
-0.0021308463,
0.0422899053,
0.0205085985,
-0.0142240524,
0.0130036809,
-0.0014706866,
0.0120595004,
-0.0020349841,
-0.0197585803,
-0.0141293108,
0.008926712,
0.0305793229,
0.0399137251,
-0.0049741836,
0.000002025,
-0.0450355262,
-0.0038334406,
-0.0222999845,
-0.028559573,
0.0182722136,
0.030131299,
-0.0379639305,
0.0484210476,
-0.0356944054,
-0.022954911,
-0.0064235688,
0.0156058948,
0.0250818226,
0.0071862247,
0.0357384309,
-0.012903926,
-0.0594469532,
0.0247548223,
-0.0379116908,
0.0102450401,
-0.0323001072,
0.0354716964,
0.0404207259,
-0.0264032185,
0.0136086177,
-0.0403274931,
0.033675123,
-0.013054451,
0.009108793,
0.0510304719,
0.0317243114,
-0.0332878716,
-0.0118845506,
0.0317781903,
0.1202175021,
-0.0081131523,
0.0129302777,
0.0052810074,
0.0048396522,
-0.0225605424,
-0.0002518577,
-0.045535095,
-0.0185407624,
0.0032457972,
0.0333234817,
-0.0014329209,
-0.0257033017,
0.0682692751,
-0.009663051,
-0.042184297,
0.0745150149,
0.0169635098,
0.0154307224,
0.0017553157,
0.0402146764,
-0.0373807773,
-0.0406469442,
0.0021079432,
0.0091903498,
-0.008369443,
0.0420457684,
-0.0185073446,
0.0136932777,
0.0101150181,
-0.0012817093,
-0.0581528619,
0.0060073463,
-0.0046303212,
0.0406928435,
-0.0563847162,
-0.0078556705,
-0.0214783661,
0.0076688249,
-0.0698889121,
-0.0349484757,
-0.029450424,
0.0206413064,
-0.0116253672,
0.0318194218,
-0.0393371619,
-0.0261143148,
-0.0284470581,
0.036868792,
-0.0169008132,
0.0235255267,
0.0329076722,
0.0136387451,
-0.0090632671,
-0.0198754594,
0.0027007463,
0.0353236981,
-0.04531179,
0.0251964498,
-0.0338118188,
0.0199680943,
-0.0135942055,
0.0034875821,
0.0397894681,
-0.0326627083,
0.0019860559,
0.0042129359,
-0.013300105,
-0.046529714,
-0.0553380959,
-0.0282467026,
-0.0134048844,
-0.0427737236,
-0.0013756139,
-0.0743672997,
0.0549670681,
0.0245569274,
0.0034532929,
-0.0605842881,
0.0243817493,
-0.0243975855,
-0.0507177003,
0.0766019151,
-0.0466895513,
-0.0111056361,
-0.0121495901,
0.0054993662,
-0.0023899311,
0.0627424046,
0.0243814606,
0.0197751988,
0.0534826666,
-0.008648674,
-0.0509966426,
0.0143784303,
0.0503779277,
0.0318872184,
-0.0084617315,
0.0103115402,
-0.0203443412,
0.0221713483,
0.0782310367,
0.0419127904,
-0.0618314855,
0.017023908,
0.0160730891,
0.0235647392,
0.0394846834,
0.0209291242,
0.0459558293,
-0.0030697403,
-0.0249842554,
0.0266002435,
-0.0318458676,
-0.0394277424,
-0.0406855196,
0.0292397477,
0.0371278897,
0.0474330746,
-0.0191623662,
-0.008291252,
0.0412570052,
-0.0350298025,
0.0415639728,
0.0309895799,
0.0758052245,
0.0253885314,
-0.0151203759,
0.0518856794,
0.0784823447,
-0.0555982254,
0.0474143848,
0.0199410319,
0.0243491754,
0.0645256713,
0.037206769,
-0.0227998365,
-0.0124757271,
0.0125913583,
-0.0030560787,
-0.0244959947,
-0.0060264799,
-0.0788275823,
0.028645549,
-0.0449311845,
-0.073941648,
-0.03473581,
-0.0065694861,
0.0029843459,
0.0625253543,
0.0193571523,
0.048567459,
-0.1018048376,
0.0172781274,
0.0154967019,
-0.0625121891,
-0.0349500328,
0.0030565411,
0.0121371243,
0.0072171367,
0.0425406508,
0.008957576,
-0.0221480001,
0.0407393016,
-0.0324559025,
0.0132536059,
-0.0120626977,
-0.019070752,
-0.0902677551,
0.0537490472,
0.0122259231,
-0.0005725679,
-0.0092966147,
-0.0537726432,
-0.0485799126,
0.0257866159,
0.0277426355,
0.0639350638,
-0.0132948281,
0.0078202346,
0.0090017878,
0.008897718,
-0.0133151077,
0.0821702257,
-0.0084365588,
-0.0066028312,
0.0261581279,
-0.0041087852,
-0.0109321345,
0.027670797,
-0.0427597016,
0.006176685,
0.0135037052,
-0.0051631546,
0.039915096,
-0.0331208594,
-0.0675455928,
0.0174059197,
-0.0969563499,
0.015887944,
-0.0048197447,
-0.001953363,
-0.0074938675,
0.031501092,
0.0293832831,
-0.000477012,
-0.0090575162,
0.026175471,
-0.0236409791,
0.0165914334,
-0.0047785887,
0.0028208618,
-0.0191245954,
-0.0386306047,
-0.0478456616,
0.0073344652,
0.0715103745,
-0.0158100221,
-0.0278914403,
-0.0064368257,
0.04991965,
-0.0114134541,
0.0170923639,
-0.0484537855,
0.0369287468,
-0.0208265707,
-0.0475207642,
0.0371414237,
0.0119131692,
-0.0545933954,
0.01813058,
0.0457638763,
-0.0136330221,
-0.0460428484,
-0.0017133285,
-0.0377590358,
0.0030010522,
-0.0720456764,
-0.0028456647,
-0.0071618455,
0.0270115118,
-0.0106018484,
-0.0316543765,
-0.0320101902,
-0.045080591,
0.0062717921,
-0.0096940119,
-0.0285478681,
-0.0193159897,
0.0168174915,
0.0138970707,
-0.0239812993,
-0.0038683277,
0.0580126084,
-0.038685333,
0.0144423954,
0.0016608997,
0.0079377415,
-0.0227395575,
0.0737313777,
0.0103272125,
0.0054442566,
-0.0043590106,
-0.0168747362,
-0.0354401767,
-0.007195713,
0.0129470704,
-0.0152622489,
0.0521657579,
-0.0186574068,
0.0284068696,
0.0139506357,
-0.0056814724,
-0.0212463047,
-0.0162547212,
-0.0262296293,
0.0305693112,
0.0620220304,
0.0031461208,
0.0244371817,
-0.006915825,
0.0070792129,
0.0307254456,
0.032427907,
-0.0067424425,
-0.0444875881,
0.0578551181,
0.0503166877,
-0.024278665,
0.02853713,
0.0098818429,
0.0156516116,
0.05340866,
0.0483119413,
-0.0186433829,
-0.0111045009,
0.0225265659,
0.0301960539,
-0.0430009514,
-0.0337015204,
-0.0186426546,
0.0460577421,
-0.0000105892,
-0.0010425808,
0.1029496491,
0.0027581232,
-0.0763387829,
0.0113479011,
0.0521770082,
0.0064049778,
0.0354423076,
-0.0240668617,
-0.1121672094,
-0.0148603199,
-0.0167725254,
0.0120111359,
0.0624862947,
-0.0513633639,
-0.002660468,
-0.0839656889,
0.003299176,
0.0403733738,
-0.0058523095,
-0.0390255377,
-0.0617806502,
0.0201712344,
-0.0092619257,
0.0120446086,
0.0075077992,
-0.0090289246,
0.0148984883,
-0.0064900825,
-0.0099563142,
-0.0911754519,
0.025058724,
-0.0314792208,
0.0614230484,
0.0017730899,
0.0254707001,
0.0116197607,
0.0298414156,
0.004624729,
-0.0400055945,
-0.0334997252,
-0.0011128022,
0.0431033932,
-0.0076406109,
0.0184200481,
0.0196502693,
0.0150572816,
-0.0618079454,
-0.0301449597,
-0.0065317969,
0.0824354291,
0.0199067723,
-0.0180840064,
0.0046199271,
-0.031118853,
0.0180481877,
0.0494914353,
0.0280440822,
0.0113239503,
0.0156576205,
-0.006678381,
0.0136550283,
-0.0491116419,
-0.000830249,
0.0108103454,
-0.0426175892,
0.0130187618,
-0.053130053,
-0.0468599051,
0.0678570494,
0.0032901308,
-0.0118974103,
0.0520247966,
0.0145734418,
-0.0173429847,
-0.0675887465,
0.031711325,
-0.0460354351,
-0.0043887864,
0.0024513206,
0.0002702546,
-0.0118787792,
-0.0437677726,
-0.0095528662,
0.0146796918,
-0.0890703946,
-0.0543657653,
-0.0268136337,
0.0201824959,
0.0025354906,
0.0944752023,
-0.0360552147,
-0.0592294335,
0.0604992695,
0.0239152443,
-0.0214601438,
0.0189538244,
-0.0040914174,
0.0220367275,
-0.0042801895,
0.0441983193,
-0.0684235916,
-0.0026855418,
-0.0057965154,
0.0622444302,
-0.0393854938,
-0.0387344211,
-0.001362726,
0.035117723,
0.0295302458,
-0.0272903889,
0.0550248921,
-0.0765922442,
-0.0116019156,
-0.0210701507,
-0.0138725685,
-0.047912173,
-0.0260491651,
0.0600464977,
-0.0323843956,
0.013420403,
-0.0153376702,
-0.0090017347,
0.0350557901,
-0.0732401609,
-0.035457883,
0.0310341343,
-0.0233047511,
0.0751373693,
0.00734554,
-0.0134000676,
0.0187021792,
-0.0106588295,
-0.0056406995,
-0.0581034087,
-0.0009089563,
-0.016378304,
-0.028882347,
-0.0504193306,
0.002882546,
-0.0452968366,
0.006834418,
-0.0408624634,
0.0242907237,
0.0697908774,
-0.0207097009,
0.0778706521,
-0.0041507841,
-0.0371311121,
-0.0220184326,
0.0132919261,
-0.011727605,
-0.0179250501,
0.0085679451,
0.0180420745,
-0.026950784,
0.030647127,
0.1017352194,
0.0125813037,
0.0569428131,
0.0032619652,
-0.0256864876,
-0.0264808424,
-0.0496571399,
-0.0192528013,
0.0759541765,
-0.0394904539,
0.0186890755,
-0.0050822096,
-0.0873583704,
0.0067055016,
-0.0521957204,
-0.0791666955,
0.0074649635,
0.0080835037,
0.0588426068,
0.0315991901,
-0.0260287896,
-0.0325216316,
0.0112199988,
-0.0103339842,
-0.0213745609,
0.007177671,
-6.201043472e-33,
-0.0087623438,
-0.0252137166,
0.0140173035,
-0.0253219344,
-0.0369712152,
0.019637771,
-0.0310579222,
-0.0112498645,
-0.043499358,
0.0233199913,
-0.0184399113,
0.0267116968,
-0.0029217354,
0.0574650988,
-0.0254281983,
0.0137902824,
0.040792454,
0.0280786753,
0.0119725028,
0.0684494227,
0.0278827008,
0.0342775099,
0.0662687793,
-0.090077728,
0.0400200002,
-0.0129687795,
0.0163263511,
0.0071725748,
0.0559274778,
0.0341369621,
-0.0316089801,
0.0889673308,
0.0136516877,
-0.0010831077,
-0.0096194232,
-0.0135508785,
-0.1270546317,
-0.0390347578,
0.0326778777,
-0.0088922614,
-0.0446019843,
-0.0297568347,
0.0536210351,
-0.0199389234,
0.0604815185,
-0.0393213443,
0.0200336389,
-0.0078043686,
-0.0666543394,
-0.0699236244,
-0.0417976901,
0.0290053692,
-0.0058992789,
-0.0003216683,
0.0509294942,
-0.0298689026,
0.0321864523,
0.0135337692,
-0.0843245611,
0.0163141061,
-0.0097348178,
0.0231391415,
0.0301684532,
0.0232032668,
0.0234549306,
0.0330777504,
-0.0213203412,
0.0129637877,
-0.0276943818,
0.0059432597,
0.0051021376,
0.0680024698,
-0.0408560932,
0.0440129526,
-0.0268775523,
-0.0258763917,
-0.0576792695,
0.0082758423,
0.0155223478,
-0.0053094556,
-0.0190417767,
0.0204247851,
0.0072691371,
-0.0431827381,
-0.0577517562,
-0.0300341677,
-0.0065469248,
-0.0218257923,
-0.0216463767,
0.0132376878,
-0.0275881682,
0.0078491364,
-0.0072912439,
-0.0685111955,
-0.0201415792,
0.0705277771,
0.0068355524,
0.0076841582,
0.0127183981,
0.0217343736,
-0.0284378007,
-0.0231946502,
0.0254518725,
-0.0601677895,
-0.0217639022,
0.0075438335,
0.0110902255,
0.020635929,
-0.0010838583,
-0.0104133748,
-0.0415479727,
-0.029464772,
-0.0109272087,
0.0176885389,
0.0528646968,
0.0543017797,
-0.0000779568,
0.016589867,
0.0790506676,
0.0426420234,
0.0200589169,
-0.0033464073,
0.0280896407,
0.0202862602,
-0.0020009589,
-0.011380964,
-0.028318543,
-0.0004234136,
0.0605044104,
-0.0448884442,
-0.0199452806,
-0.0154837258,
2.673e-7,
0.0690137073,
0.0060121352,
0.0192422401,
0.0011154077,
0.0019189174,
-0.000145695,
0.0338232405,
0.0149855427,
-0.0071169292,
-0.0545814782,
0.0542273521,
0.0215948261,
-0.012640967,
0.0392164215,
-0.0635421127,
0.0067773573,
-0.0175345112,
-0.0141320443,
-0.0382948071,
-0.0189889316,
0.0278344303,
0.059927851,
0.0379479937,
-0.0198181588,
-0.0204057805,
-0.0638763085,
0.0448598601,
-0.0390966609,
-0.0013266311,
0.0599476472,
-0.0013222034,
0.0254909117,
-0.0069943219,
0.0291606374,
-0.01867982,
0.0485698022,
0.0120558701,
0.0651267543,
0.0098044211,
-0.0076707555,
-0.0133487545,
-0.0358494185,
0.047965683,
-0.0651646852,
0.1086867526,
-0.0461420678,
0.0132005652,
0.0086582052,
-0.0585587882,
-0.0135856904,
-0.0030933153,
0.0111872656,
0.0087814918,
0.0168182626,
0.0223999452,
0.0313710459,
-0.0352025889,
-0.0960109234,
-0.001720148,
0.0361639895,
-0.0413142592,
-0.0338893682,
0.0245561767,
-0.0614395328,
0.0231919549,
-0.0346209668,
0.0603516325,
2.585540921e-34,
0.0263305493,
0.0089396965,
-0.0026651917,
0.0106115732,
0.0414267518,
-0.0429256223,
-0.0550860688,
-0.0067242985,
0.0038105499,
-0.1054766029,
0.008904106
] |
http://arxiv.org/abs/2401.15585v1 | http://arxiv.org/pdf/2401.15585v1 | Evaluating Gender Bias in Large Language Models via Chain-of-Thought Prompting | 2024-01-28T06:50:10 | 2024-01-28T06:50:10 | Masahiro Kaneko, Danushka Bollegala, Naoaki Okazaki, Timothy Baldwin | cs.CL | cs.CL | There exist both scalable tasks, like reading comprehension and
fact-checking, where model performance improves with model size, and unscalable
tasks, like arithmetic reasoning and symbolic reasoning, where model
performance does not necessarily improve with model size. Large language models
(LLMs) equipped with Chain-of-Thought (CoT) prompting are able to make accurate
incremental predictions even on unscalable tasks. Unfortunately, despite their
exceptional reasoning abilities, LLMs tend to internalize and reproduce
discriminatory societal biases. Whether CoT can provide discriminatory or
egalitarian rationalizations for the implicit information in unscalable tasks
remains an open question.
In this study, we examine the impact of LLMs' step-by-step predictions on
gender bias in unscalable tasks. For this purpose, we construct a benchmark for
an unscalable task where the LLM is given a list of words comprising feminine,
masculine, and gendered occupational words, and is required to count the number
of feminine and masculine words. In our CoT prompts, we require the LLM to
explicitly indicate whether each word in the word list is a feminine or
masculine before making the final predictions. With counting and handling the
meaning of words, this benchmark has characteristics of both arithmetic
reasoning and symbolic reasoning. Experimental results in English show that
without step-by-step prediction, most LLMs make socially biased predictions,
despite the task being as simple as counting words. Interestingly, CoT
prompting reduces this unconscious social bias in LLMs and encourages fair
predictions. | Evaluating Gender Bias in Large Language Models via Chain-of-Thought Prompting: There exist both scalable tasks, like reading comprehension and
fact-checking, where model performance improves with model size, and unscalable
tasks, like arithmetic reasoning and symbolic reasoning, where model
performance does not necessarily improve with model size. Large language models
(LLMs) equipped with Chain-of-Thought (CoT) prompting are able to make accurate
incremental predictions even on unscalable tasks. Unfortunately, despite their
exceptional reasoning abilities, LLMs tend to internalize and reproduce
discriminatory societal biases. Whether CoT can provide discriminatory or
egalitarian rationalizations for the implicit information in unscalable tasks
remains an open question.
In this study, we examine the impact of LLMs' step-by-step predictions on
gender bias in unscalable tasks. For this purpose, we construct a benchmark for
an unscalable task where the LLM is given a list of words comprising feminine,
masculine, and gendered occupational words, and is required to count the number
of feminine and masculine words. In our CoT prompts, we require the LLM to
explicitly indicate whether each word in the word list is a feminine or
masculine before making the final predictions. With counting and handling the
meaning of words, this benchmark has characteristics of both arithmetic
reasoning and symbolic reasoning. Experimental results in English show that
without step-by-step prediction, most LLMs make socially biased predictions,
despite the task being as simple as counting words. Interestingly, CoT
prompting reduces this unconscious social bias in LLMs and encourages fair
predictions. | [
0.0299929362,
0.073929958,
-0.0329006873,
0.0229837243,
-0.0274173878,
0.028176561,
0.0613180958,
-0.0306930542,
0.0197610073,
-0.036622718,
-0.004256112,
-0.0496151075,
0.0376941971,
-0.0181314703,
0.0540400408,
0.0011388052,
0.0299385954,
-0.0310867839,
0.0348194689,
-0.0654215887,
-0.0049929121,
0.011905428,
-0.0065371641,
0.0241749976,
0.0156546179,
-0.0368844531,
0.0103271361,
-0.0127572175,
-0.0207424052,
-0.0136415036,
0.0028848839,
0.0716622993,
0.014178629,
0.0178529862,
0.0000020374,
-0.0900885537,
0.0284972228,
0.006590426,
-0.04869169,
-0.0078009972,
-0.0224103797,
-0.0071262973,
0.0101601109,
0.026101796,
-0.008642883,
0.023290636,
-0.0035446249,
0.0571074672,
0.0017062171,
-0.0016712105,
0.0097965281,
0.0368004143,
-0.0516749732,
-0.0241677444,
-0.000043259,
-0.0201452915,
0.0265462864,
-0.015828982,
-0.047926303,
0.0180027429,
-0.0300758183,
0.0021142797,
0.014478609,
-0.0098863281,
-0.0035093443,
0.0176561121,
-0.0080898413,
-0.0437462516,
-0.0125154639,
0.0214067847,
0.0309349112,
-0.0063114506,
0.0189071503,
-0.0124160433,
0.0251609385,
0.0293070078,
-0.0576559417,
0.0494733267,
-0.0109031163,
-0.0284552649,
0.0059758746,
-0.0092451805,
0.0306590423,
0.0517900363,
0.0000707884,
0.1288417578,
-0.0077514946,
-0.0402626209,
0.037788894,
0.0223451331,
-0.004117291,
-0.0781673864,
-0.0010507472,
-0.0117036495,
-0.0222022831,
0.0118040675,
0.0231855065,
-0.0016484297,
0.0190436263,
-0.0542767681,
0.036758177,
0.0801146477,
0.0169510338,
0.0214654934,
-0.0250653811,
0.0314398706,
-0.0113462377,
0.0202061608,
0.0016168852,
-0.0177610237,
0.0056664799,
-0.0441788025,
-0.0062666233,
0.0171789583,
-0.0032761346,
0.0416958444,
-0.0254815575,
-0.0253179334,
-0.0450385399,
-0.0037597073,
0.0227441043,
0.0195433274,
0.0172394328,
-0.0090406016,
-0.059078861,
-0.0056116637,
-0.0839941129,
0.0459382311,
-0.0036090733,
-0.0829923153,
-0.0101728598,
-0.0290137213,
-0.0084749423,
-0.0217605438,
0.0158092249,
0.031185206,
-0.0223488882,
-0.0296250861,
-0.019412091,
-0.0215628408,
-0.0113334944,
-0.0613720939,
0.0175217465,
-0.0609756857,
0.0123612722,
0.0509184971,
0.0115445107,
-0.0192159098,
0.0156371798,
0.0258883778,
-0.0380002186,
0.0577709079,
0.0737838522,
0.0244312752,
0.0443511941,
0.0189152397,
-0.0587658733,
0.0988430604,
0.054538887,
-0.0029914426,
0.0122969439,
0.0085892305,
-0.0078860363,
0.0265192352,
0.0305198599,
0.0115510989,
-0.0325438939,
0.0300562195,
-0.0236320496,
0.045883514,
0.0716965422,
0.0241030939,
0.011321838,
0.0083692549,
0.061831452,
0.0854831561,
0.0244749021,
0.0507370345,
0.032284826,
-0.0230711065,
-0.0010575894,
0.0165670756,
-0.0899640247,
0.0327229016,
-0.0206991043,
-0.0450402535,
0.0110301459,
-0.0062131532,
-0.0279453546,
-0.0459063016,
-0.0069793114,
0.0544705354,
0.0204016734,
0.0331579037,
0.0392720401,
0.0051357313,
-0.0068404423,
0.0439872518,
0.0315564051,
-0.0265359953,
-0.0060127387,
-0.0661487058,
0.0341278911,
0.0285487678,
0.0100399107,
-0.00375267,
0.0167871229,
-0.0255726371,
-0.0075298073,
0.0403824225,
0.0644796416,
0.0158673953,
-0.0160114244,
-0.0222485419,
-0.0120852711,
0.0477555022,
0.0256070457,
0.0069754305,
0.0402666666,
0.0397759452,
0.0010593645,
-0.0208589677,
-0.0264837034,
0.0229759291,
-0.0262395684,
-0.0379817449,
0.0001220102,
0.0059160548,
-0.0028063033,
-0.0066945506,
0.0023406469,
0.0130484132,
-0.0070344745,
-0.0105652837,
-0.0481836684,
0.0068287342,
-0.0873314813,
0.0102594327,
0.0147322705,
-0.0111232856,
-0.0681281388,
0.0291924272,
0.0029364508,
0.0090263588,
0.055866342,
-0.0298430808,
-0.0178713407,
-0.0115956152,
0.0213362575,
-0.016955439,
0.0035403636,
-0.0120302336,
0.0531597547,
-0.0352204815,
-0.0122839743,
0.0179053154,
-0.029497819,
-0.0220381301,
-0.0465885997,
-0.0190253947,
0.01420963,
0.0113010546,
-0.0293786842,
0.021470217,
-0.0415791348,
-0.0503199026,
-0.0317956172,
-0.0752574503,
-0.0090332665,
0.0162034668,
-0.0132331792,
-0.0328188874,
-0.0050891433,
-0.0221818145,
0.0080270544,
-0.0138287125,
0.0113474652,
0.0207373966,
0.0601031184,
0.044289764,
-0.0261005666,
0.0003162997,
-0.0199551862,
-0.0271249712,
-0.0659513101,
0.0538613871,
0.0375629552,
0.0006659577,
-0.0405309312,
0.0195019804,
-0.0570595339,
0.0236015711,
-0.0226667039,
0.0006144004,
0.0247590654,
0.0453576148,
0.0684862584,
0.1029598489,
0.0241075289,
-0.000880308,
0.0775915533,
0.0101708025,
-0.0584427156,
-0.0012160098,
0.0188862216,
0.0110480506,
0.0333358683,
-0.0535150282,
-0.0478772633,
0.0540682264,
0.0267679114,
-0.0158150438,
-0.0380709209,
0.0734448507,
-0.0587323569,
-0.0698108897,
-0.0597390905,
-0.0452248007,
0.0024946211,
0.0757029802,
-0.0415703766,
-0.0412833169,
0.0359392576,
0.0105041051,
-0.0056884331,
-0.0038014699,
0.0181125253,
-0.0358305201,
0.0777996257,
-0.0244165603,
0.0007214826,
-0.0055024014,
0.0027203094,
-0.1075353324,
-0.0493140295,
0.017781226,
-0.0317769386,
0.0637345836,
-0.0523341931,
0.0130350236,
-0.010550525,
-0.0506963544,
-0.057271488,
-0.0192061346,
-0.0373816639,
0.0010694194,
-0.0040454548,
-0.0099778064,
-0.002675948,
-0.013013646,
0.0406108163,
0.0158051755,
-0.0163238216,
-0.028705854,
0.015168108,
-0.0349691436,
0.0251641106,
0.0216496866,
0.0762720108,
-0.0326235443,
0.0449894443,
0.0395707563,
0.0219811685,
0.020239301,
-0.0034629849,
0.0223812256,
0.0284077059,
-0.0167676639,
-0.0117502566,
0.0107431207,
0.0377590917,
-0.0562475659,
0.0063552717,
0.1126959175,
0.0226006974,
-0.0571760945,
0.0171343312,
-0.0406071506,
0.0138999289,
0.0466219559,
0.0026547934,
-0.0703426078,
0.0020556508,
0.0624740683,
-0.0228534192,
-0.05413232,
-0.0815351233,
-0.0301388539,
-0.1284003109,
-0.0121917352,
-0.0054051657,
0.0085656298,
0.0033200544,
-0.0338387527,
0.005508753,
0.0180493295,
0.0020127918,
0.0293871313,
0.0104831122,
0.0132525591,
-0.001737305,
-0.0281812437,
0.0137160663,
0.070728533,
-0.0430410244,
0.0317100883,
0.0020902839,
0.0243367162,
-0.0275509041,
-0.0132776694,
0.0723592117,
-0.0084544988,
-0.0016619476,
-0.0275093634,
-0.0085685011,
-0.0245860852,
-0.0058970759,
-0.0658179373,
0.0323719829,
0.0280226953,
0.0095998123,
0.0198421199,
0.0509715229,
0.0279815104,
-0.0501028597,
-0.0351455435,
-0.0049381214,
-0.0273880213,
0.0804790333,
0.0148921022,
-0.0430021808,
-0.012161158,
-0.0433171764,
-0.0393040441,
0.0591289103,
-0.0739965513,
-0.0004336488,
0.0703437626,
-0.0064158533,
-0.0362878852,
-0.0484611876,
0.0419851094,
0.01774396,
0.006233952,
-0.0372070447,
0.0360540748,
0.0251884069,
-0.0579333268,
0.0131724384,
-0.0309418701,
0.0252680443,
0.0183347836,
0.0068280608,
-0.038299609,
-0.0341727994,
-0.0218307767,
0.0345905945,
-0.0470791198,
-0.0421029776,
0.0113402689,
0.0244059414,
0.0332980826,
0.0287528839,
0.0169199836,
-0.0335856751,
0.0159701016,
0.0071290289,
-0.0314372368,
0.0649677813,
0.0159513541,
0.0462639332,
0.0348192006,
0.0477917269,
-0.0306818429,
0.0296310168,
0.0118993483,
0.0053116763,
-0.0058796224,
-0.014168486,
0.0172316749,
0.0191145018,
0.0137628671,
-0.0302954856,
0.0289999023,
-0.0329158045,
-0.007234476,
-0.0145890806,
-0.0590980314,
0.0117190108,
-0.0380608886,
0.0183942933,
0.0168855283,
-0.0297908373,
-0.0615853556,
-0.00666617,
0.0122409416,
-0.0046001282,
0.0056366026,
0.0009339169,
0.0563420877,
0.0065097879,
0.0081107905,
-0.0015688075,
-0.0058613084,
-0.0357783884,
0.051262442,
0.0320392027,
0.0604916848,
0.0172843318,
0.0144406967,
-0.0499561094,
-0.0369531326,
-0.0513110682,
-0.0509330668,
0.0371229611,
0.0130697861,
-0.0152998352,
-0.0445293151,
0.0143020814,
-0.0193767697,
0.003444355,
0.0084507521,
-0.0205342695,
-0.0012547408,
0.0019032765,
0.0359060355,
0.0158013608,
0.0206406545,
-0.0335822329,
0.0417284779,
0.0171463545,
0.0374709703,
0.0038861539,
-0.0324376374,
-0.0148085905,
-0.0430426858,
-0.0180244427,
-0.033496961,
0.0070732348,
0.0134897856,
-0.119588986,
-0.0567731149,
-0.0326209925,
-0.00409716,
0.0431889221,
-0.0438734591,
0.0076942486,
0.0619497336,
-0.0276972372,
-0.0460754931,
-0.0089389207,
0.0245390907,
0.0059501752,
0.0071339305,
-0.034498781,
-6.291355968e-33,
-0.034477789,
-0.0595162697,
0.0070697367,
-0.0051113716,
-0.0055343574,
0.0042639934,
-0.0323468819,
0.0197862834,
0.0012432657,
0.0387937874,
-0.0336105488,
0.0502256304,
0.0101651875,
-0.0279016327,
0.0521610081,
-0.0050150407,
0.033658307,
-0.0204781648,
-0.0446340889,
-0.0464240201,
-0.0020496519,
0.0112063885,
0.0125127444,
-0.0458676293,
0.085756585,
0.0000016121,
0.0187439676,
0.0217595808,
-0.0462534353,
-0.0120774675,
-0.0115550645,
0.0445339531,
0.0299644321,
-0.0358195491,
0.0049090898,
0.0912340358,
-0.0542212836,
-0.0510979779,
0.0452205464,
0.0050582201,
0.0011241158,
0.0089206314,
0.0122415274,
0.0345970653,
-0.0656489283,
-0.0362684354,
0.0216786489,
-0.0048716059,
0.0500670262,
0.0312596671,
-0.0295899045,
0.0032856539,
-0.0186013002,
-0.0041587106,
0.0004008564,
0.0140912281,
-0.013097466,
-0.028016489,
-0.0594669133,
0.022736365,
-0.0031477162,
-0.0093416953,
0.0487202033,
0.0041321185,
0.0195766743,
0.0435802937,
0.0907369107,
0.0521757528,
-0.0686204359,
-0.0575110577,
0.0235602241,
-0.0435043909,
0.0645289049,
0.0092462637,
0.0727937371,
0.021889532,
-0.0593842939,
-0.0672525242,
0.047447253,
0.028161129,
0.0402806364,
-0.0181720946,
-0.0660037994,
-0.0575120524,
-0.0611547306,
-0.0345294438,
-0.0397223793,
-0.0113495989,
0.0282457843,
0.0113270422,
-0.0265372097,
-0.0200645011,
-0.015599885,
-0.0221939851,
-0.0342097878,
0.0057550007,
0.0006892096,
-0.0258207638,
0.0363717452,
0.0085527459,
-0.0759521648,
0.0273573808,
-0.0362056829,
0.0593790784,
0.0270553622,
0.0296328999,
-0.0269959867,
0.007569347,
-0.0716354027,
-0.0203889459,
-0.0614338294,
0.0097787948,
0.0169736166,
-0.0057054139,
0.0379895456,
0.0298936758,
-0.0144320242,
-0.0128140841,
0.0235607009,
-0.020168148,
-0.0395691395,
0.0357092917,
-0.0276382379,
0.0271437988,
-0.0247248709,
-0.0331037529,
-0.0244829059,
0.0136079201,
0.0548342392,
-0.0548043549,
-0.0010946955,
-0.0150055131,
2.714e-7,
-0.0273670908,
0.0781338662,
-0.0045590908,
0.0305740573,
0.0225644615,
0.0284992941,
0.0037958547,
0.0225912128,
0.0673411191,
-0.0449276604,
-0.05601741,
0.0747176185,
0.0138397543,
0.0304175094,
-0.0408301502,
-0.0315692164,
-0.074169457,
0.0282763187,
0.0044599902,
0.014977118,
-0.0164867565,
0.026288243,
0.0530095883,
0.0038692988,
-0.0340749472,
-0.019933667,
0.0453949459,
-0.0031402875,
0.0315375365,
-0.0257633086,
-0.0018729229,
-0.0119397799,
0.0317263566,
0.059492711,
0.0511106849,
0.0025805337,
-0.0061235586,
0.0225786455,
-0.0011516196,
0.0684367344,
-0.0284311473,
-0.0158843417,
-0.0219890755,
-0.049245514,
0.0645862445,
-0.0067954799,
-0.0410790965,
-0.0157523975,
0.023578763,
-0.01589581,
0.0515328608,
0.0187732875,
-0.0059082587,
0.0030375582,
0.031322442,
0.0008562645,
0.0166835729,
-0.0574087352,
-0.0368601978,
-0.0260511115,
-0.0206014272,
-0.0819206312,
-0.0347830653,
0.032170821,
0.0152682755,
-0.0325895213,
0.0086823711,
2.285570375e-34,
-0.0015998262,
0.0223577265,
0.0300392341,
-0.0132651245,
0.0268707499,
-0.0163623262,
-0.0054554739,
-0.0028389485,
0.0908176899,
-0.0299728513,
-0.010410931
] |
http://arxiv.org/abs/2401.15579v1 | http://arxiv.org/pdf/2401.15579v1 | MunTTS: A Text-to-Speech System for Mundari | 2024-01-28T06:27:17 | 2024-01-28T06:27:17 | Varun Gumma, Rishav Hada, Aditya Yadavalli, Pamir Gogoi, Ishani Mondal, Vivek Seshadri, Kalika Bali | cs.CL | cs.CL, cs.SD, eess.AS | We present MunTTS, an end-to-end text-to-speech (TTS) system specifically for
Mundari, a low-resource Indian language of the Austo-Asiatic family. Our work
addresses the gap in linguistic technology for underrepresented languages by
collecting and processing data to build a speech synthesis system. We begin our
study by gathering a substantial dataset of Mundari text and speech and train
end-to-end speech models. We also delve into the methods used for training our
models, ensuring they are efficient and effective despite the data constraints.
We evaluate our system with native speakers and objective metrics,
demonstrating its potential as a tool for preserving and promoting the Mundari
language in the digital age. | MunTTS: A Text-to-Speech System for Mundari: We present MunTTS, an end-to-end text-to-speech (TTS) system specifically for
Mundari, a low-resource Indian language of the Austo-Asiatic family. Our work
addresses the gap in linguistic technology for underrepresented languages by
collecting and processing data to build a speech synthesis system. We begin our
study by gathering a substantial dataset of Mundari text and speech and train
end-to-end speech models. We also delve into the methods used for training our
models, ensuring they are efficient and effective despite the data constraints.
We evaluate our system with native speakers and objective metrics,
demonstrating its potential as a tool for preserving and promoting the Mundari
language in the digital age. | [
-0.0029655965,
0.0869837105,
-0.0011175955,
-0.0270347726,
-0.0043435032,
-0.0373424403,
-0.0062812059,
-0.0019572712,
-0.0316349529,
-0.0112124486,
-0.0462097563,
-0.0457117707,
0.0675910339,
0.0234428905,
0.0343106575,
-0.1077160537,
0.0054051303,
-0.0451247618,
-0.0074689314,
0.0073558106,
0.0302256644,
0.0705465972,
-0.0456023142,
0.0456560776,
-0.049758032,
-0.0097429901,
-0.0241194032,
-0.0132135032,
-0.0041038729,
0.0048146248,
-0.019513661,
-0.0068680877,
0.023456743,
-0.0003926894,
0.0000018239,
-0.0634151697,
0.0021072049,
-0.0005757038,
-0.0151938871,
0.0299743507,
0.0702720731,
0.0172001086,
0.0033774877,
0.0637769476,
0.0253634714,
0.0409539044,
0.0542209335,
0.0157349557,
0.0422708206,
0.0778452605,
-0.0290214345,
-0.0325619876,
0.0425306447,
-0.0046412209,
-0.0224078223,
0.0263015535,
-0.0506614856,
-0.0127053484,
0.0306829158,
-0.0655730367,
-0.0090903398,
0.0279416405,
-0.0132237095,
-0.0042130044,
-0.0286392495,
-0.008259262,
-0.0119860619,
-0.0672861114,
0.0041528279,
0.043468859,
-0.0190540254,
-0.0340192132,
0.0258706193,
0.017052047,
-0.0450629257,
-0.0452857092,
-0.0101544736,
-0.0046851444,
-0.0184437595,
0.0221767481,
0.0220974144,
0.0193067901,
0.0048757843,
-0.0078786872,
0.0217342842,
-0.0299910102,
0.031251125,
-0.0338252559,
-0.0162397549,
-0.036893636,
0.0148260491,
-0.0204675421,
0.0546897314,
0.0249092802,
-0.0237578247,
0.0283365548,
-0.031686727,
-0.0801024064,
-0.0300475135,
-0.004979664,
-0.0158032943,
0.0439968333,
-0.0101736477,
-0.0387950428,
-0.0030278952,
0.0049258517,
-0.0167480111,
-0.0032141374,
-0.0176943243,
-0.0234995987,
0.0042445189,
-0.0438618027,
-0.0576420203,
-0.0067253029,
-0.0432033911,
-0.0043158284,
-0.0532205664,
-0.0076432535,
0.0028010069,
-0.0227228329,
0.0404160768,
-0.0127320234,
-0.060791865,
-0.0572703965,
0.0455582924,
0.0231459383,
0.0145160556,
0.047321707,
0.0315685049,
-0.0971577093,
0.0247728322,
0.0307000838,
0.0015511016,
0.011947386,
0.00069573,
0.0387880094,
-0.0326271616,
-0.0265395474,
-0.0786258727,
-0.006767618,
-0.0330140665,
-0.0551237836,
-0.0167478379,
-0.0415077172,
0.0408125632,
0.016138263,
0.0031637636,
-0.0211659316,
-0.0174777284,
-0.0023738951,
0.0249274261,
0.0336229578,
-0.0116722295,
0.0425743386,
0.0942579582,
0.0274298675,
0.0353316665,
0.0235678107,
0.0902551711,
0.0128134284,
0.0161872208,
-0.021726951,
0.0088992193,
0.072530508,
0.0042926045,
0.0247704126,
-0.016075965,
0.0117707327,
-0.0455139577,
0.0357076265,
0.0262586325,
0.0569803864,
-0.0022332855,
0.0747225136,
0.059547551,
0.0117991995,
0.1000714675,
0.0584817231,
0.0578377098,
0.0059569236,
0.0034968317,
-0.0180826969,
0.0203318931,
0.0156556331,
0.020291768,
0.0123554301,
0.0815421715,
0.0363727137,
-0.0276080109,
0.00650902,
-0.0015828462,
-0.0604089722,
-0.0560135543,
-0.0034953591,
0.0384703763,
0.0530265942,
-0.0564583838,
0.0387575105,
0.042550344,
-0.0927892849,
0.0580009185,
-0.0274415035,
0.0266602989,
0.0368382595,
0.0205914434,
-0.0205434971,
0.0278649721,
-0.04444547,
0.0102166897,
0.0493960194,
0.0037141123,
-0.0204689708,
0.0254110582,
-0.0875032991,
0.0363632962,
0.0617358238,
0.0030657179,
0.0386761576,
-0.0340923704,
0.0664583519,
0.0265192296,
-0.0265315995,
0.0379023924,
0.0178493913,
-0.0532790162,
-0.0117798951,
-0.0160722621,
-0.0004737852,
0.0406051837,
0.0348625518,
-0.0138910208,
0.0145654213,
0.0014186193,
-0.0126259346,
-0.0066354522,
0.050808318,
-0.0468148552,
0.0058548842,
0.0042975568,
-0.055182334,
0.038084995,
0.0642862394,
0.0449442491,
-0.1055876166,
0.0682733506,
-0.0191544406,
0.035995584,
-0.0237482954,
0.0157326665,
0.0433352552,
-0.037415836,
0.0549760796,
-0.0216398686,
0.0066615515,
-0.023861751,
0.0443190634,
-0.0264014099,
0.0478159562,
-0.0995313972,
0.0112727638,
0.0365631282,
-0.0186725371,
0.0498022959,
-0.0168454628,
0.0226565041,
-0.0140641741,
-0.0326770768,
-0.0688320473,
-0.0133792683,
-0.0044349316,
0.018237086,
-0.01175409,
0.0843946263,
-0.0047256271,
0.0009428195,
-0.022235062,
-0.0392464213,
0.0359433144,
-0.0328755341,
-0.0051440732,
-0.0127076181,
-0.0200124513,
-0.0239393897,
-0.0130824586,
0.0057245367,
0.0274167061,
-0.0271156877,
-0.057920564,
-0.0374549255,
-0.0346497074,
0.0092648193,
-0.0018151145,
-0.010693267,
-0.0182512868,
0.0037913928,
-0.020690823,
0.0346115492,
-0.0283646118,
-0.0121076433,
-0.0294802692,
-0.0211893413,
0.0124282837,
-0.0220434237,
-0.0060869697,
0.0282613169,
0.016146671,
0.0039781472,
-0.0213424731,
0.0080096005,
0.0278775189,
0.0230240058,
0.0171589702,
-0.0577582084,
-0.005016908,
-0.0250281729,
-0.004059792,
0.0188573394,
0.0076267011,
-0.0488738082,
0.0168568753,
-0.0107137142,
-0.0545044467,
0.009184015,
-0.047486525,
-0.0478078686,
-0.0265243035,
0.0071331249,
-0.0117283305,
-0.0144236768,
-0.0384390987,
0.0226880815,
-0.008761134,
-0.0386015177,
-0.0489589535,
0.036294844,
0.0015481989,
-0.0295266137,
-0.0044190283,
-0.0419409201,
0.0170951746,
-0.0300049949,
-0.0177170672,
0.0040067905,
0.0010270685,
0.0066936975,
0.0424463227,
0.0062655658,
0.0051359846,
0.0308831036,
0.0578895845,
-0.0200166404,
0.0189802963,
0.0115412269,
0.0067023993,
-0.0024352588,
0.0545880347,
0.0360325128,
0.0207415354,
-0.0194232725,
-0.0084379297,
0.0349478647,
-0.0563553534,
0.0660587624,
0.0027532475,
-0.0049821269,
-0.0290123969,
0.0108451145,
0.0180187263,
0.0214813016,
0.0073392736,
0.0582174547,
0.0180806331,
0.0052021234,
0.0426126122,
0.0024558201,
0.015007237,
-0.0098753385,
-0.0137397731,
0.0537269637,
-0.0223616455,
0.004743855,
-0.0760505199,
-0.0302481689,
0.0672818646,
0.0754735321,
0.0220241118,
0.0404012464,
-0.0452718064,
-0.0937881768,
-0.0135839544,
0.0148439314,
-0.0366613269,
-0.0271059088,
-0.0024610211,
-0.0689716488,
0.0076330341,
0.029671412,
0.0583694614,
-0.0008615931,
-0.0319183134,
0.0457327776,
-0.0584090687,
-0.0267980602,
-0.0194951296,
-0.0144527853,
0.0744738355,
0.0225750934,
0.0121787917,
-0.0146758324,
-0.0090149632,
0.0032655697,
-0.0338699184,
-0.0141209988,
0.007655059,
0.0111830831,
0.0249983799,
-0.0068924846,
-0.0118582407,
0.0724507123,
-0.0025440517,
0.0123303989,
0.0065304921,
0.0305412319,
0.0064081191,
-0.036036104,
0.0099798758,
0.0009675894,
0.036151398,
0.0016102191,
-0.002548465,
-0.0702992603,
-0.0048071919,
-0.0230284166,
0.0290911216,
-0.0438095219,
-0.0235544853,
0.0296347756,
-0.0204540268,
-0.0254218467,
-0.0429168753,
-0.0312378462,
0.0125529123,
0.0358643085,
0.0362883322,
-0.0151914842,
0.0002686297,
0.014417734,
-0.0211746022,
-0.0102374423,
-0.0412187316,
0.0458990894,
-0.0641222522,
-0.0154429786,
0.0055156448,
-0.036495965,
-0.0003450135,
0.0205556303,
-0.0592870079,
-0.0530646928,
0.0038336806,
0.055378627,
0.0333246142,
0.0010793369,
0.0002164677,
-0.0259448681,
0.057614509,
-0.0188645646,
-0.0206698067,
0.0121494634,
0.0626587719,
0.003534673,
-0.0054566055,
0.0090797469,
-0.0363351591,
0.017549539,
-0.0188840274,
0.0211575255,
0.0105441771,
-0.0073740357,
0.0363186486,
0.0088441186,
-0.0193700865,
-0.0869909599,
-0.0030657488,
-0.0537771657,
-0.0084876204,
0.0269176867,
0.0451715775,
-0.0445277616,
0.0112247374,
0.0788541809,
0.0837066472,
-0.0115529718,
-0.0618738122,
-0.010609393,
-0.0153891193,
0.0156301185,
0.0651295781,
0.0409397781,
0.0541316755,
0.0449839681,
-0.0196785647,
-0.0219492894,
0.0507563949,
-0.044245556,
-0.0060961982,
-0.0207983293,
0.0274705384,
-0.005585962,
0.0105242608,
0.0498367697,
0.0289184935,
0.0702909008,
0.0189875346,
-0.0258876681,
0.0453900024,
0.0599406622,
0.035613209,
0.103313759,
-0.0188190117,
0.0454673618,
-0.0308307149,
0.0391312875,
-0.0331444703,
0.0205060355,
-0.0111404294,
-0.0119190859,
-0.0283358488,
0.0222952571,
0.0317547023,
-0.0279292427,
0.0244522914,
0.0028730428,
0.0265231039,
-0.0107666049,
-0.0483928546,
-0.043553818,
-0.0078136306,
-0.011770701,
-0.0302462168,
-0.0388987176,
-0.0523071475,
0.0255516581,
-0.0121339429,
0.0109735662,
-0.0482768938,
0.0233644005,
-0.055274982,
-0.0155248335,
0.00934703,
-0.0226152278,
-0.025693953,
-0.0554695502,
-0.031211881,
0.0218247604,
-5.669587415e-33,
-0.0248251576,
0.0235963706,
0.0032460557,
0.0217710324,
-0.030298112,
-0.044597134,
-0.0235406626,
-0.0024686123,
-0.0027809888,
-0.0090845777,
-0.0374018699,
-0.0072150389,
0.0057696644,
0.0262012295,
-0.0038382488,
0.0317931287,
0.0407278426,
-0.0054922057,
-0.008402436,
0.0163489934,
0.0431240201,
0.023387894,
0.0466935299,
-0.0406526811,
0.0052593574,
-0.0664739981,
-0.0556550659,
-0.0165425241,
-0.0324802548,
0.0115849013,
-0.0275346451,
0.0638999343,
-0.0071365777,
-0.0165147614,
-0.0230720136,
0.0233672168,
-0.0575995333,
0.0612835139,
-0.0206366312,
0.054563459,
-0.0050624497,
-0.0134059209,
0.0306701027,
-0.011247362,
-0.0421636924,
0.0178064462,
-0.0102167269,
-0.0070407349,
-0.0002837943,
-0.0334904753,
-0.061976742,
0.0352097079,
-0.0173028968,
-0.0354760885,
-0.0002068486,
0.0930632204,
-0.0260929838,
-0.0753274336,
-0.0254447665,
0.0298381243,
0.0308197457,
0.0396527834,
0.0387227237,
0.0256844461,
-0.0292176269,
-0.0274568163,
0.049113702,
-0.0906383768,
0.0480861664,
0.0370754451,
0.0286247805,
0.0300174039,
0.0409264639,
0.033712931,
0.0250375848,
-0.0861492008,
-0.018623434,
-0.0181094017,
0.0015539745,
0.0191311724,
-0.0152393281,
0.0453398973,
-0.0189716928,
-0.0134701142,
0.0070225936,
-0.0539659113,
0.003228632,
-0.0172806233,
0.1000811011,
0.013056878,
0.0145674711,
0.0168845914,
0.053628698,
0.0168366618,
0.0031233709,
-0.0289148893,
-0.0472456552,
0.0198226739,
-0.0198595636,
0.0182002094,
-0.0401940905,
-0.047390338,
0.0521465242,
0.0273644309,
0.0184391104,
-0.0241287779,
-0.0105559258,
0.0482502356,
-0.0632440299,
0.0131535586,
-0.0137486886,
0.071666345,
-0.0091735367,
0.0179662686,
0.0264056437,
-0.0461826399,
0.0023920103,
-0.0164722446,
0.0033602263,
-0.0135077881,
-0.0007608227,
0.0083347289,
0.0274723042,
-0.0091378624,
-0.058628425,
0.0220847744,
-0.0197541136,
-0.0270509887,
0.0400539078,
-0.0135029582,
-0.0028227763,
0.0149936015,
2.583e-7,
-0.0177379381,
0.0199311785,
0.0484671965,
-0.042849604,
0.0029162241,
0.0105036171,
-0.0484017469,
0.0100213839,
-0.026982449,
0.0725189671,
-0.0022953125,
-0.0317613333,
0.0931615829,
0.0269618575,
-0.1249743551,
-0.0276888665,
-0.0614497364,
0.0199992787,
-0.0378366709,
-0.0161862541,
-0.0324611515,
0.0415528566,
-0.0269856323,
-0.0510857254,
-0.0077600656,
-0.0396122038,
0.0426992998,
-0.0094852746,
0.0063546542,
-0.0176046304,
0.0472804941,
-0.0569815189,
-0.0432351083,
-0.0778747723,
-0.0475947522,
0.0043102526,
0.0522945076,
0.0978967026,
-0.0286923721,
-0.0132824751,
0.0408916883,
0.0431784615,
-0.0617500618,
0.012730049,
0.005002704,
0.0514427647,
0.0077974149,
-0.0348921679,
0.0081892461,
0.0310315508,
-0.0008174434,
-0.0173845049,
0.0195017736,
-0.0074058152,
0.0018107404,
-0.0232859775,
0.0152142346,
-0.0113732507,
-0.0315153711,
0.0239644125,
-0.0388785042,
-0.0688593313,
-0.0204772428,
0.0381249599,
0.03828834,
-0.0278469473,
-0.0170953441,
1.792778232e-34,
0.036447607,
-0.0207538698,
0.0353225023,
-0.0158674661,
0.013518394,
-0.048898261,
0.0482614338,
0.0108569246,
0.0301504899,
-0.0259210113,
-0.0471032523
] |
http://arxiv.org/abs/2401.15569v1 | http://arxiv.org/pdf/2401.15569v1 | Efficient Tuning and Inference for Large Language Models on Textual Graphs | 2024-01-28T05:12:09 | 2024-01-28T05:12:09 | Yun Zhu, Yaoke Wang, Haizhou Shi, Siliang Tang | cs.CL | cs.CL | Rich textual and topological information of textual graphs need to be modeled
in real-world applications such as webpages, e-commerce, and academic articles.
Practitioners have been long following the path of adopting a shallow text
encoder and a subsequent graph neural network (GNN) to solve this problem. In
light of recent advancements in large language models (LLMs), it is apparent
that integrating LLMs for enhanced textual encoding can substantially improve
the performance of textual graphs. Nevertheless, the efficiency of these
methods poses a significant challenge. In this paper, we propose ENGINE, a
parameter- and memory-efficient fine-tuning method for textual graphs with an
LLM encoder. The key insight is to combine the LLMs and GNNs through a tunable
side structure, which significantly reduces the training complexity without
impairing the joint model's capacity. Extensive experiments on textual graphs
demonstrate our method's effectiveness by achieving the best model performance,
meanwhile having the lowest training cost compared to previous methods.
Moreover, we introduce two variants with caching and dynamic early exit to
further enhance training and inference speed. Specifically, caching accelerates
ENGINE's training by 12x, and dynamic early exit achieves up to 5x faster
inference with a negligible performance drop (at maximum 1.17% relevant drop
across 7 datasets). | Efficient Tuning and Inference for Large Language Models on Textual Graphs: Rich textual and topological information of textual graphs need to be modeled
in real-world applications such as webpages, e-commerce, and academic articles.
Practitioners have been long following the path of adopting a shallow text
encoder and a subsequent graph neural network (GNN) to solve this problem. In
light of recent advancements in large language models (LLMs), it is apparent
that integrating LLMs for enhanced textual encoding can substantially improve
the performance of textual graphs. Nevertheless, the efficiency of these
methods poses a significant challenge. In this paper, we propose ENGINE, a
parameter- and memory-efficient fine-tuning method for textual graphs with an
LLM encoder. The key insight is to combine the LLMs and GNNs through a tunable
side structure, which significantly reduces the training complexity without
impairing the joint model's capacity. Extensive experiments on textual graphs
demonstrate our method's effectiveness by achieving the best model performance,
meanwhile having the lowest training cost compared to previous methods.
Moreover, we introduce two variants with caching and dynamic early exit to
further enhance training and inference speed. Specifically, caching accelerates
ENGINE's training by 12x, and dynamic early exit achieves up to 5x faster
inference with a negligible performance drop (at maximum 1.17% relevant drop
across 7 datasets). | [
0.0490996242,
0.0499781035,
-0.0063027525,
0.0777387545,
-0.021675434,
0.0093795927,
0.0037546568,
0.0157238245,
0.0401488468,
-0.0459271669,
-0.0158588458,
-0.017932212,
0.0058946055,
0.0507457368,
0.0279231109,
-0.0014052748,
0.0036772178,
-0.0216571018,
-0.0047707562,
-0.019412538,
0.0096027851,
0.058565788,
0.0577271245,
-0.0003690024,
0.0295127612,
-0.0051281611,
-0.0159481149,
0.0661780313,
-0.0182675309,
-0.0317627117,
0.0004410589,
0.0062620258,
0.0430951528,
0.0161867458,
0.0000020554,
-0.0032387367,
-0.0111530824,
-0.0021867394,
-0.0212247372,
0.0374426618,
0.0547250956,
-0.0355466157,
0.0340864845,
0.0171780773,
-0.0037769605,
-0.0295961816,
0.0466898382,
0.0220564939,
0.0173842777,
0.1108748987,
-0.0278053433,
-0.0550075658,
0.0186175592,
-0.046499528,
-0.0265710074,
-0.0740165934,
0.0491205193,
-0.0065138992,
-0.029270472,
0.0089197559,
-0.0680088922,
0.0099643162,
0.0270262305,
0.009023224,
0.0214600228,
0.0356651917,
0.004119433,
0.0215933714,
0.0020143108,
0.1007665321,
-0.0265928321,
-0.0228411891,
0.0109062921,
-0.0287042372,
-0.0197776109,
0.0332500264,
-0.0325739048,
-0.0195403304,
-0.0074191722,
0.0299371835,
-0.0202691518,
-0.0183047652,
0.0563747063,
0.0166673996,
-0.0259267911,
0.0237950403,
-0.0038765862,
0.0056983246,
0.0046043438,
0.006327861,
-0.0384034701,
-0.0312236696,
0.0342375711,
0.0246947892,
0.0105649009,
0.0071874107,
-0.0427827872,
-0.0007611534,
-0.0183889586,
-0.0440469682,
0.0277788825,
-0.0026246605,
0.0276540052,
0.0515407361,
-0.029711986,
0.0331905261,
-0.0467329212,
0.0446704887,
-0.0477438942,
-0.0126662161,
-0.0069877952,
0.0167583022,
-0.0400186479,
0.0110576674,
-0.0013076797,
-0.0640659183,
-0.041642908,
0.0333962962,
0.0228713024,
0.0071933074,
0.026625637,
0.0113918176,
-0.0070283697,
-0.0433021523,
0.0357552804,
0.0221273601,
-0.0033621951,
0.0181445796,
-0.0254411399,
-0.0033456967,
-0.0095956335,
0.037230961,
0.0895214379,
-0.0311102513,
0.024775397,
0.0644074827,
0.0241927877,
-0.0164455064,
-0.0582899041,
-0.0227175578,
-0.0061118961,
-0.0465948991,
0.0270076189,
-0.0819278657,
0.0511813313,
0.0284270011,
0.0012053365,
-0.0623685382,
0.011444089,
0.0030660627,
-0.0869982094,
0.0232227221,
-0.0205520485,
0.0016830617,
-0.0011443993,
-0.0337382481,
0.0133599406,
0.0510655306,
0.051446218,
0.0744903386,
0.0403022319,
-0.0002123998,
-0.0427755192,
0.0514165647,
0.0548463501,
0.0152759319,
-0.0089160465,
-0.0119073354,
-0.0280169174,
0.0108513935,
0.0131334234,
0.0567712858,
-0.0539239496,
-0.0585884154,
0.0511935502,
0.0407076068,
0.0209938288,
0.0291456487,
0.016039364,
0.03293024,
-0.006083637,
-0.0067848265,
-0.0177778173,
0.0103803314,
-0.001192313,
0.0088498462,
0.0107697677,
0.0066149542,
-0.0224992074,
-0.0023249725,
-0.0295871478,
-0.0270309299,
0.0709470287,
0.018090304,
0.0768849626,
0.0027702404,
-0.0603218637,
0.0196678434,
0.0283893757,
-0.0017691032,
0.0038066017,
0.0000547774,
-0.0172099769,
0.0785603672,
0.0467552319,
0.0214104597,
-0.0431719646,
-0.0052755503,
0.0075271074,
0.019348206,
0.0239184834,
-0.0159423575,
-0.0040458897,
-0.0841835663,
-0.0490884818,
-0.0253495146,
-0.011891434,
0.0234468281,
0.045939371,
0.0237507932,
0.0342892781,
-0.0519643649,
-0.0349546559,
-0.0033471943,
-0.0026050289,
0.0237688515,
0.0056829643,
-0.0050097341,
-0.0077821901,
0.0029672249,
0.0011769817,
-0.00775945,
0.0176696796,
-0.016220957,
-0.0152291488,
-0.0505243503,
-0.0409775004,
-0.0645757169,
0.023691332,
-0.0149287228,
-0.0026697977,
0.0216378216,
-0.071654439,
-0.019800907,
0.0266307462,
-0.0732259005,
0.0297756214,
0.0188525394,
-0.0145611707,
-0.0305814445,
0.0272310153,
0.0077044005,
0.0284272488,
-0.0179633126,
0.0003655614,
-0.0055098939,
0.0224002879,
-0.023211496,
-0.0370276906,
-0.0150077958,
0.0190196019,
0.0234747529,
-0.0008030549,
0.0253820773,
0.0090563679,
-0.024662761,
-0.0193236023,
-0.0998060033,
0.0100479396,
-0.013491014,
0.0219637584,
0.0085410429,
0.0150694419,
0.005058995,
-0.0234635547,
-0.057478942,
0.0616757944,
-0.0174524039,
0.0188430473,
0.0104451533,
-0.019467257,
-0.0224916469,
-0.0194712877,
-0.0659021735,
-0.0043174345,
0.0965960696,
0.0063948594,
-0.0782961771,
-0.0281069707,
0.0429023243,
-0.018910544,
0.0076576215,
-0.019842634,
-0.0220513605,
-0.0431137346,
-0.0038674627,
0.0225788467,
0.072952047,
-0.0017007217,
0.0298529491,
0.0252912752,
-0.0062999623,
-0.0446641967,
-0.0141057177,
-0.0518890992,
-0.0078171268,
-0.0495888554,
-0.0259141065,
-0.0397579074,
0.0332810059,
0.0001677575,
0.009941441,
-0.0343052335,
-0.0022612205,
0.0017088186,
0.0253352337,
-0.0168119036,
-0.0267158654,
0.0285703707,
-0.0072234985,
-0.0452726595,
-0.0055739949,
0.0623236299,
-0.0216505826,
0.0044842903,
0.0283518862,
-0.0252656639,
-0.0798499733,
0.0513631217,
0.0525007024,
0.0354782976,
-0.0088361986,
-0.0232004505,
-0.0426433161,
0.0378922075,
0.0344563089,
-0.0347005725,
0.010278509,
0.020863235,
0.0397643559,
-0.0254042968,
-0.0274505168,
0.0075943447,
-0.011473109,
-0.0167814437,
0.0016769752,
0.0410295129,
-0.0426522978,
-0.0077670389,
0.0206516925,
-0.0454017594,
0.0413097404,
-0.0477000065,
0.0173425823,
-0.0287992042,
0.0565871112,
0.0720572844,
-0.0298793595,
-0.0111505408,
-0.0614894517,
-0.0001557663,
0.0609594658,
-0.0279139969,
0.0032197516,
-0.0387905762,
0.0008141313,
0.0023342073,
0.016291894,
-0.0026604601,
-0.0153030055,
0.0176197514,
-0.0727913529,
0.0025370831,
0.1073754728,
0.0034037768,
-0.0743035674,
-0.0251957644,
0.0283398479,
0.0354915932,
0.0048590656,
0.0161650013,
-0.0926713273,
-0.0275458004,
-0.004729942,
0.0275535323,
0.0273946915,
-0.0298721585,
-0.0033644335,
-0.099450998,
-0.0030241893,
0.0495989136,
0.013304376,
-0.0340828747,
-0.0471898653,
-0.0236335639,
-0.0046075047,
0.0024614271,
-0.0056891036,
0.0064284033,
-0.0267425533,
0.002533426,
-0.0133595336,
-0.0383147225,
0.0112006972,
-0.0107924519,
0.0051208236,
-0.0263588801,
0.0249152668,
0.018914463,
0.0158585459,
0.0854168236,
0.0203014594,
-0.027903812,
-0.0275148395,
0.0762452185,
-0.020372251,
0.0154526345,
-0.007809178,
0.0145936739,
-0.0410178676,
0.0020662702,
-0.0279134158,
0.0659978986,
0.0251517314,
-0.0441974476,
0.0074668438,
-0.0334585719,
-0.0531133711,
0.0211506207,
0.0522606187,
-0.007303725,
-0.0515458547,
-0.0534294806,
-0.0222008526,
-0.0556845702,
-0.0500503145,
0.0164142791,
-0.0051550744,
0.0016551948,
-0.0423521772,
-0.0200401582,
0.0079842964,
-0.0136973849,
-0.0151535384,
-0.0094463592,
0.0680298284,
-0.0350140557,
-0.0905389264,
-0.0043989397,
-0.0579604432,
0.013522258,
-0.0027943745,
0.0053683328,
0.0565382242,
0.0034004136,
-0.003627849,
0.02510833,
-0.1040973812,
-0.0873537585,
-0.0446129963,
-0.008167997,
0.008084612,
0.0888704881,
-0.0186506025,
-0.039601706,
0.0264696479,
0.0400356874,
-0.0184551664,
0.0103401104,
0.0050406992,
-0.0103493752,
-0.0157041419,
0.0291368738,
-0.0688171759,
0.0136175156,
0.0186364297,
0.0248487499,
-0.0323077701,
-0.0428660586,
-0.009972007,
0.0372987948,
0.013436066,
0.0271046087,
0.0031263747,
-0.0575267561,
0.0241933521,
-0.0165125001,
-0.0216565896,
-0.0225666184,
-0.0146136023,
0.0318371616,
-0.0090135923,
-0.0291541964,
-0.0330329314,
0.0873015895,
0.0162964631,
-0.0615707487,
-0.009953307,
-0.0093159899,
-0.0219092257,
0.0452788807,
0.0151363574,
-0.0429909937,
0.0247280542,
0.0206261259,
0.0126257008,
-0.0334000103,
0.0189719088,
0.023223348,
-0.0270334668,
-0.0714235157,
0.0066167666,
-0.0581014939,
-0.0080763903,
-0.0611386783,
0.0240716971,
0.0591047294,
-0.0009059245,
0.0387868732,
0.0087813111,
-0.0175752379,
-0.0300384406,
0.054101944,
-0.0071242894,
-0.0437167026,
0.0273677837,
0.0184528567,
-0.0074676271,
0.0253435709,
0.0572708286,
0.077546753,
0.0354379825,
0.0190037135,
-0.0583577044,
-0.0179553702,
-0.0299989227,
-0.024710767,
0.0174259972,
0.0050666579,
0.0024402849,
-0.0032042451,
-0.0772669464,
0.018750377,
-0.0134870019,
0.0302570835,
0.016824387,
0.0161089599,
0.0362183154,
0.0177509449,
0.000555328,
-0.0361707918,
0.0509211048,
-0.0063167969,
-0.0241677221,
-0.0106815714,
-6.583123952e-33,
0.0016962665,
-0.0619970858,
-0.0009167981,
0.0272759572,
-0.072764121,
0.0236982219,
-0.0508672744,
-0.0457475819,
-0.0228196289,
-0.011820823,
0.0064104134,
0.0176680814,
0.002186869,
0.0380597077,
-0.0361289755,
-0.030566128,
0.0399535596,
-0.0014169981,
0.0079623573,
0.0559419729,
0.0296385996,
0.0428399295,
0.0442343242,
-0.0531372279,
-0.031078767,
-0.025407631,
0.0237951167,
-0.0102380253,
0.0361138918,
0.0425257906,
-0.0033044436,
0.0309653524,
0.042061016,
0.0057569807,
0.0084622083,
0.0204860829,
-0.0763854533,
-0.0174012538,
-0.0230070557,
0.0313207246,
-0.0160673391,
-0.0360227525,
0.0622608662,
0.0217242334,
0.0385750383,
-0.0454005487,
0.0397457182,
-0.0011538533,
-0.0531043559,
-0.0501961634,
-0.0309858397,
0.0652787089,
-0.0030230863,
0.0184826814,
0.0348879173,
0.0383129828,
0.0225009229,
0.0390435904,
-0.0377944335,
-0.0002317194,
-0.0355554521,
0.0353776328,
0.013339486,
0.029100256,
0.0117508508,
0.0364498571,
-0.0154587934,
0.0346629694,
-0.0432277396,
0.0477160215,
0.0440632254,
0.0457774624,
0.0345120132,
0.014675294,
0.0039969175,
-0.0131553793,
-0.1166038662,
-0.0208907928,
0.0045704544,
0.0284802038,
0.0302873515,
0.0475266241,
-0.0139220413,
-0.0568797328,
-0.0258150473,
-0.0268338062,
-0.0382778607,
-0.008552012,
-0.0011719773,
0.0209391285,
-0.0418965779,
0.0079057943,
0.0260793716,
-0.0834552348,
-0.0226919483,
0.0860117301,
-0.0040705306,
0.0493253879,
-0.0212768354,
0.0260251891,
-0.0427127667,
-0.0559040196,
-0.0033014023,
-0.0264592227,
0.011708539,
-0.0020609256,
0.0748692006,
0.0189140644,
-0.0446547605,
-0.0154271415,
-0.0456556678,
-0.0307702813,
-0.0216856673,
0.0253296178,
0.0464798026,
0.0321902595,
-0.0083538704,
0.0688395202,
0.0579595231,
0.0463601388,
-0.0152080497,
-0.0154134305,
0.0240154732,
0.0278257411,
-0.0321740992,
0.0005929112,
-0.0195031781,
0.0174156483,
0.0205402859,
-0.0736015961,
-0.0259764679,
0.0217747521,
2.779e-7,
0.0477281213,
0.0688165873,
0.0574506931,
0.0264557377,
0.028279176,
0.0037646808,
0.0086146025,
0.0284479149,
0.0230177082,
-0.0030967325,
0.0522231869,
0.0497297719,
0.0189107079,
0.0068744728,
-0.0841163546,
0.0520647168,
-0.0874546543,
0.0004781917,
-0.0218172874,
-0.0061178668,
0.0337736383,
0.0723959878,
0.0255472101,
-0.0001159354,
-0.0187669732,
-0.0486361645,
0.0714553744,
-0.0174672473,
-0.0136326337,
0.02268962,
-0.0158614032,
0.0283380579,
-0.0442679152,
-0.0104320087,
-0.0036143726,
0.0538733266,
0.0038792845,
0.0562442765,
-0.0135993995,
0.019132955,
0.0351083912,
-0.0362970307,
-0.0125225978,
-0.0372969471,
0.0794054344,
-0.0627273917,
-0.0286336113,
-0.0315321572,
-0.0063928259,
-0.0296365917,
0.0040588463,
-0.0216879565,
-0.0065509882,
0.0471926332,
0.0290945582,
-0.0424577966,
0.0076675387,
-0.0841785595,
0.0076254746,
0.0098767886,
-0.0381860249,
-0.0393871889,
-0.0236506294,
0.0371551961,
0.0239253864,
-0.0168228839,
0.0121355588,
2.676107939e-34,
0.0565413423,
-0.010992418,
-0.0201477297,
0.0227251146,
0.0382343084,
-0.0262543652,
-0.0055216132,
0.0250732228,
-0.0001681992,
-0.1240560636,
-0.0276363716
] |
http://arxiv.org/abs/2401.15560v1 | http://arxiv.org/pdf/2401.15560v1 | An Analysis of Letter Dynamics in the English Alphabet | 2024-01-28T03:54:41 | 2024-01-28T03:54:41 | Neil Zhao, Diana Zheng | cs.IT | cs.IT, cs.CL, math.IT, 94A15 | The frequency with which the letters of the English alphabet appear in
writings has been applied to the field of cryptography, the development of
keyboard mechanics, and the study of linguistics. We expanded on the
statistical analysis of the English alphabet by examining the average frequency
which each letter appears in different categories of writings. We evaluated
news articles, novels, plays, scientific publications and calculated the
frequency of each letter of the alphabet, the information density of each
letter, and the overall letter distribution. Furthermore, we developed a metric
known as distance, d that can be used to algorithmically recognize different
categories of writings. The results of our study can be applied to information
transmission, large data curation, and linguistics. | An Analysis of Letter Dynamics in the English Alphabet: The frequency with which the letters of the English alphabet appear in
writings has been applied to the field of cryptography, the development of
keyboard mechanics, and the study of linguistics. We expanded on the
statistical analysis of the English alphabet by examining the average frequency
which each letter appears in different categories of writings. We evaluated
news articles, novels, plays, scientific publications and calculated the
frequency of each letter of the alphabet, the information density of each
letter, and the overall letter distribution. Furthermore, we developed a metric
known as distance, d that can be used to algorithmically recognize different
categories of writings. The results of our study can be applied to information
transmission, large data curation, and linguistics. | [
0.0498653911,
0.0308025032,
-0.0370440334,
0.0345485434,
-0.0674793869,
-0.0031514824,
0.0176444463,
0.0564653054,
0.0006543039,
0.0089380043,
0.0567444675,
-0.0663221106,
0.0426627845,
-0.0005894275,
0.0163588375,
0.0007776029,
0.0197117086,
-0.0311570875,
0.0116795246,
-0.0075621782,
0.0128037818,
-0.0032974274,
0.0039028183,
-0.0100050727,
0.0485059246,
-0.0094723701,
-0.043626532,
-0.0281690955,
0.0241127014,
0.0077342191,
-0.0624426566,
0.0723977908,
-0.0252684467,
0.068333149,
0.0000017873,
-0.0265742093,
-0.0107140522,
0.0014197591,
-0.0043699201,
0.0256768186,
0.0910498202,
0.0114732571,
-0.0505085103,
0.0029344133,
0.0191225074,
0.003240014,
0.0793674588,
-0.0036126673,
-0.0215329323,
0.0379054062,
-0.0169252902,
-0.0485175848,
0.0401264057,
0.0114134056,
0.0187822618,
0.0070561576,
0.0274122935,
-0.0210689418,
0.0116195446,
0.0453188345,
-0.047844559,
0.0178032219,
-0.0140899327,
-0.0115450872,
0.0193537753,
0.050111834,
0.0063845483,
0.0028320414,
-0.0143490378,
0.0136479456,
0.005721671,
-0.0349938683,
0.0563569739,
0.001728487,
-0.0113225337,
0.041355215,
-0.0087020341,
0.0050026807,
-0.0230794102,
-0.0422632247,
0.031982217,
0.0293754842,
0.0027641377,
0.0025002076,
-0.0180402491,
0.0247206148,
0.0206655376,
-0.0307932869,
-0.0109490668,
-0.0476918742,
0.0800479501,
-0.0647795498,
0.0196122527,
0.0044904309,
0.044607196,
0.0188655369,
-0.0343489163,
-0.0386350155,
0.0240067411,
-0.0974016115,
0.0292138997,
-0.0277445484,
0.0639013276,
0.0207230076,
0.0204092059,
0.0119713647,
-0.0005880884,
0.0001154585,
0.0107023111,
0.0461574756,
-0.0790986344,
-0.0389502458,
0.0214055777,
0.0395010225,
0.0688949078,
-0.0306638405,
-0.0213474222,
0.0120259179,
0.036676459,
-0.0110207172,
-0.0318526365,
-0.0430698916,
0.0042156498,
-0.0189709961,
0.0226507038,
-0.0400265865,
-0.0462841727,
0.0034010671,
-0.0325829275,
0.0204490777,
0.0291471388,
0.0068441187,
0.046252653,
-0.0112438099,
-0.0139484582,
0.0649357587,
-0.0198894832,
0.0019864193,
-0.0022273245,
-0.0257056653,
-0.0354226492,
0.0078185359,
-0.0371188223,
-0.0691038296,
-0.0043935468,
0.00756475,
-0.0036725793,
0.0036646416,
-0.0126151936,
0.0166444853,
-0.047803659,
0.0520038977,
-0.0101673547,
0.0271619745,
0.0548341759,
-0.0098367622,
0.063623026,
0.0129412794,
0.0043856869,
0.0217548665,
0.0415392108,
-0.0385915302,
0.0399270952,
0.0376046598,
-0.0389409102,
-0.0077002845,
-0.006765719,
-0.038966693,
-0.0036759172,
-0.0042090802,
-0.0352730565,
0.0603874214,
-0.0189250577,
-0.0309829246,
0.0586614385,
0.1168914586,
0.0550835282,
0.0341329128,
-0.0669709444,
0.00485483,
0.0254431702,
0.0061754868,
0.0318340398,
0.0235539358,
-0.0183518752,
-0.0245762486,
0.0435820781,
-0.0055088704,
-0.000995259,
-0.0099961553,
0.0018818234,
-0.0039589279,
0.0355609171,
-0.048089426,
0.0314499587,
-0.0004006453,
0.0019851176,
0.0048503359,
-0.0133240186,
-0.033333946,
-0.0171988364,
-0.0088194069,
0.0227223914,
0.0919369534,
0.0720029771,
0.0086144991,
0.0558590144,
0.021497326,
0.0344466195,
0.0076892092,
-0.0252672769,
-0.0022227347,
0.0099656992,
-0.0280815531,
0.0078817727,
0.0667234063,
-0.0072615314,
0.0180005115,
-0.0456795767,
0.0910576135,
-0.0078773508,
-0.0263632797,
-0.0299494304,
0.0025739837,
0.0057826275,
0.0091692572,
-0.0111442572,
-0.0036021075,
0.007951797,
0.0216953233,
-0.0395223498,
0.0283317268,
0.0021638495,
-0.0578196272,
-0.0005093642,
-0.0256732348,
0.0578760765,
0.1211324558,
0.0522896573,
-0.0672777519,
0.0451152138,
0.0060177739,
0.0334943384,
-0.0162625741,
0.0410784259,
-0.0431340076,
0.0498806164,
-0.0120993257,
0.013223703,
-0.0560602471,
-0.0097843697,
-0.039520964,
0.0450317077,
0.0259288847,
0.0728485435,
-0.0710482076,
-0.0626384169,
0.0664014444,
-0.018171655,
0.0093207052,
0.0479866564,
-0.0386487134,
0.0221168455,
0.075810045,
-0.0332950577,
-0.0001597125,
-0.0875282362,
-0.0332691744,
-0.0374807566,
0.0188406445,
0.04575013,
0.0084299557,
0.0051629688,
-0.007574833,
0.006940966,
-0.0284653734,
-0.0485855304,
-0.0289128553,
-0.021845825,
0.0599457026,
0.0029781654,
-0.058182586,
-0.0137255797,
-0.0315638892,
-0.0067220451,
0.0288500767,
0.0361278094,
-0.0429326408,
-0.0527936406,
0.0260625575,
-0.0092468215,
-0.0374397077,
-0.0023736567,
-0.0596971177,
0.0279078744,
-0.0129719665,
0.0243803151,
0.0327871591,
0.0389862731,
-0.029307982,
0.0024609552,
-0.0360285603,
0.0208444055,
-0.0526092052,
-0.007778212,
0.0595969632,
-0.02822919,
-0.0698238015,
0.0272729266,
0.0095214136,
-0.0092290938,
-0.0042417091,
-0.0244648047,
0.0062146224,
-0.0070243361,
-0.0100055495,
-0.0453641191,
-0.0207790993,
-0.0348098204,
0.012807495,
-0.0203377884,
-0.0167693105,
-0.0085160192,
-0.0378366932,
0.0396837406,
-0.0237190984,
-0.013581072,
0.0022838202,
0.0514003932,
-0.0452203788,
0.0280419476,
0.0289319903,
-0.0844772458,
-0.0146323973,
-0.0054159998,
0.0552616641,
-0.044485379,
0.002995644,
-0.0058735204,
0.0206068736,
-0.0013901358,
0.0099014323,
0.0079791378,
-0.0483163856,
0.0044817906,
-0.0760914236,
0.0330593735,
-0.0968938395,
0.0085950484,
0.0133777615,
0.0536331683,
0.0184493251,
0.0102093797,
0.0007576183,
-0.007231195,
0.0258619003,
0.0072243866,
-0.0375664346,
-0.0370846912,
-0.0422193632,
-0.0186944567,
0.0334092639,
-0.0106775342,
0.0458425619,
-0.0058995299,
-0.0366668813,
-0.0012078645,
-0.0364351049,
-0.0399663635,
0.0449418761,
-0.0164495818,
0.0196618699,
0.0134540275,
0.0820218176,
0.0396168195,
-0.0368565172,
0.0000819552,
0.0364588723,
-0.0303478409,
-0.0122175682,
0.0143397348,
-0.0631343499,
0.0400500931,
0.0100311963,
-0.0026462532,
-0.0252828617,
-0.035745278,
-0.0160956271,
-0.0686216578,
0.0090421923,
0.0183227006,
-0.003941901,
-0.0501383729,
-0.0223836303,
-0.0140100187,
0.0274367668,
-0.0646626204,
0.008258664,
0.0203427393,
0.0186136253,
0.024413297,
-0.0854724795,
-0.0135505367,
-0.0209425222,
-0.003601128,
-0.0023105522,
-0.0100265862,
0.0578556396,
-0.0790857002,
-0.0342354849,
-0.0505126566,
0.0329407454,
-0.0240158029,
0.0017681401,
0.0357684568,
0.0648804009,
-0.027336739,
-0.0570680276,
0.0002084023,
-0.018579416,
-0.0203128643,
0.0160657894,
0.0208074637,
-0.0039666272,
-0.033584699,
-0.0220814794,
-0.0279359445,
0.0202936083,
0.0045453315,
-0.0006050956,
-0.032977473,
-0.0171935614,
-0.0420510471,
0.0747709423,
-0.0350509249,
-0.0585284233,
-0.0483284146,
0.0150361843,
-0.0874219984,
-0.0795399621,
0.0395891741,
-0.0594038889,
0.0529726073,
0.0406036675,
0.0558735132,
0.0456295088,
-0.035696283,
-0.053663712,
-0.1158842444,
-0.0187054891,
0.042422194,
-0.0609754026,
0.0021688871,
-0.0305898376,
-0.0037924394,
-0.0306876134,
0.0316002592,
-0.0328137055,
-0.0574253574,
0.0391138047,
-0.0148423044,
0.0339573286,
-0.055844605,
-0.0340827405,
-0.065999575,
0.0376796946,
0.0490764491,
0.030760102,
-0.0028910011,
-0.0217352863,
-0.0774300918,
-0.0046536401,
-0.0247306302,
-0.0035680295,
-0.0167861208,
0.0014396461,
-0.046986945,
-0.0622280315,
-0.0077343769,
0.0009618709,
0.02513122,
-0.0380929112,
0.0025765754,
0.0411515757,
-0.0050651706,
0.0007608751,
-0.0437593125,
0.0057882383,
0.0267143566,
-0.0165431537,
0.0202944018,
-0.0124933384,
-0.0175633319,
0.0224111993,
0.0026598279,
-0.0266670417,
0.0154752275,
0.0336728878,
0.0141719505,
0.1214110926,
0.0146208899,
-0.0010035986,
0.0357197374,
0.0225023013,
-0.0016991502,
0.0548484325,
-0.0110281538,
-0.063570939,
0.021392148,
-0.0050439644,
-0.0181582049,
0.0309621133,
-0.0774391815,
-0.0340052284,
0.0335655361,
-0.0032195905,
0.0415290371,
0.0156533346,
-0.0300457291,
-0.0135444459,
0.0093392441,
-0.0191883445,
-0.0045920629,
0.001208798,
0.0483808666,
-0.0649267733,
0.0327181183,
-0.0001461694,
0.0349513702,
0.063372016,
-0.0039974838,
-0.0315266624,
0.01680406,
-0.0788435414,
-0.0155083407,
0.0351217464,
0.0163390916,
0.0184498243,
0.0001650699,
0.010974125,
-0.0407822356,
-0.03012578,
0.089125663,
0.0001006479,
0.0539355725,
0.0117372992,
0.0293533895,
0.0551220849,
0.000944846,
0.0335039459,
-0.0236626416,
0.0100898389,
-0.0438737944,
-0.0359991863,
0.0068733129,
-5.859561261e-33,
0.0574085079,
-0.0135807367,
-0.0623755343,
0.0352362841,
-0.014341115,
0.0001242774,
-0.0243446734,
0.0076162824,
-0.010693891,
0.0377851725,
-0.0200385507,
0.0205479991,
0.022579724,
-0.0354631282,
0.0507823452,
0.0114620272,
-0.0294073801,
-0.0062260563,
0.0074033411,
-0.0152556607,
0.0769542009,
0.026589388,
0.0610786788,
-0.0485132709,
0.0527048595,
-0.0267271325,
-0.0283153299,
-0.0444646515,
0.0479505658,
0.0185208712,
-0.0044935336,
0.0202108305,
0.0023656487,
-0.0313088186,
0.0061713434,
0.0550040118,
0.0085763158,
-0.0504753254,
0.0128431786,
0.0154786212,
0.0642297566,
-0.0643629134,
0.0144348014,
-0.0271837078,
-0.0545652732,
0.0162944887,
0.0393229127,
-0.0189626515,
-0.0422383361,
-0.0367213823,
0.0207838099,
0.0115089966,
-0.0393884815,
0.0361250453,
0.0219576862,
-0.0717682987,
-0.0030402078,
-0.0170905236,
-0.0177728627,
0.0743438676,
0.0714882463,
0.0577649809,
-0.0200320389,
0.0072289128,
0.0163063686,
0.0444404408,
0.0519538745,
-0.0015367633,
0.0096457386,
0.0636646897,
0.0228022095,
-0.0058930446,
0.0196288712,
0.0943659171,
0.0857031792,
-0.0414591916,
-0.055726923,
0.007567565,
-0.0009425615,
-0.0224031862,
0.0219654832,
0.0003964638,
-0.0769258291,
-0.0377970301,
0.0147595992,
0.0329384767,
0.0045852172,
-0.0276042763,
-0.0271805935,
-0.0322758965,
0.0077251107,
-0.024308607,
-0.0013696554,
-0.0486629456,
-0.0762111768,
0.002601872,
0.0237377528,
0.065585576,
-0.0056780102,
-0.0097505348,
-0.023242224,
-0.0128022991,
-0.0160685871,
0.0030865292,
0.005266529,
-0.0299174935,
-0.0332125165,
0.0313485414,
-0.0759825334,
-0.0169814248,
0.0483495332,
-0.0280911326,
0.0391856506,
0.0160276014,
-0.01147993,
-0.0259703789,
0.0198724456,
-0.0180322565,
0.0222071689,
-0.005446739,
0.0192596577,
0.010773791,
-0.0342112854,
0.0175616667,
0.0119034844,
-0.0028033443,
-0.0461935513,
0.0064468752,
0.0457686186,
-0.0160578024,
-0.0193512645,
-0.0016161753,
2.689e-7,
0.0544958264,
0.052993536,
0.0700955912,
0.0429070257,
0.0387758687,
-0.0142467795,
-0.0273751002,
0.0323293395,
-0.0438093841,
-0.0283815283,
0.0093568992,
-0.0036751712,
0.0509084798,
0.016356986,
-0.0787319988,
0.0044006188,
-0.0528434068,
0.0025821663,
-0.0483613722,
0.0262280554,
0.0019938922,
0.0336055532,
0.0140661243,
0.0038058953,
-0.0114240991,
-0.0464760587,
-0.0021614297,
-0.01743632,
0.0360643677,
-0.0011495416,
0.0106244832,
-0.0362731256,
-0.0030271567,
0.0389002152,
-0.0271823313,
0.0049468745,
0.0341514125,
-0.0265984982,
-0.0212003458,
0.0484111011,
0.0064753699,
-0.041942317,
-0.0190258864,
-0.0142720984,
0.038343329,
0.0723227262,
-0.0267395861,
-0.000564315,
-0.0359137356,
0.0021424263,
0.012989142,
0.0260286424,
-0.0321862623,
0.0123738367,
0.010955805,
0.0031568478,
0.0501371399,
-0.0359801762,
-0.0064198035,
-0.0036632747,
-0.0182537325,
-0.003100815,
-0.0375486165,
-0.0202077273,
0.0322358683,
-0.0226132926,
0.0362119898,
2.301325902e-34,
0.01984521,
-0.0233654045,
-0.0399876945,
0.0252913684,
0.002195311,
0.0201322734,
0.0181827806,
0.0241099726,
0.0210576504,
-0.0478585958,
-0.0123879677
] |
http://arxiv.org/abs/2401.15555v1 | http://arxiv.org/pdf/2401.15555v1 | Augment before You Try: Knowledge-Enhanced Table Question Answering via Table Expansion | 2024-01-28T03:37:11 | 2024-01-28T03:37:11 | Yujian Liu, Jiabao Ji, Tong Yu, Ryan Rossi, Sungchul Kim, Handong Zhao, Ritwik Sinha, Yang Zhang, Shiyu Chang | cs.CL | cs.CL | Table question answering is a popular task that assesses a model's ability to
understand and interact with structured data. However, the given table often
does not contain sufficient information for answering the question,
necessitating the integration of external knowledge. Existing methods either
convert both the table and external knowledge into text, which neglects the
structured nature of the table; or they embed queries for external sources in
the interaction with the table, which complicates the process. In this paper,
we propose a simple yet effective method to integrate external information in a
given table. Our method first constructs an augmenting table containing the
missing information and then generates a SQL query over the two tables to
answer the question. Experiments show that our method outperforms strong
baselines on three table QA benchmarks. Our code is publicly available at
https://github.com/UCSB-NLP-Chang/Augment_tableQA. | Augment before You Try: Knowledge-Enhanced Table Question Answering via Table Expansion: Table question answering is a popular task that assesses a model's ability to
understand and interact with structured data. However, the given table often
does not contain sufficient information for answering the question,
necessitating the integration of external knowledge. Existing methods either
convert both the table and external knowledge into text, which neglects the
structured nature of the table; or they embed queries for external sources in
the interaction with the table, which complicates the process. In this paper,
we propose a simple yet effective method to integrate external information in a
given table. Our method first constructs an augmenting table containing the
missing information and then generates a SQL query over the two tables to
answer the question. Experiments show that our method outperforms strong
baselines on three table QA benchmarks. Our code is publicly available at
https://github.com/UCSB-NLP-Chang/Augment_tableQA. | [
0.0151091916,
0.0718025342,
-0.0211607926,
0.00237149,
-0.0003205862,
0.048934456,
-0.0499399938,
0.0180732161,
-0.0006196514,
-0.0138899311,
-0.0076714819,
0.0405824147,
-0.0597260259,
0.047932744,
0.047840558,
0.0219739862,
0.0451542214,
0.0176150575,
-0.0560566522,
0.02476166,
0.0173115581,
-0.0157177206,
-0.0173495375,
-0.0130713871,
-0.004794687,
0.0221207645,
-0.0724509507,
-0.010695423,
-0.0230979808,
0.0262227096,
0.0018655439,
0.0727925748,
0.0170753598,
0.0128344316,
0.0000018964,
0.0002576782,
-0.0489506423,
0.00797961,
-0.0429400839,
-0.0281941015,
-0.0107605495,
-0.0348732322,
0.0389145166,
-0.004655831,
-0.0416963995,
-0.0123631107,
0.0236455444,
0.0639226288,
0.0267996732,
0.0169870295,
0.0152291497,
-0.0031874205,
-0.0011477186,
0.0135435574,
0.0639945641,
-0.0423524491,
0.0174789168,
0.0284733288,
0.037816558,
-0.0121683814,
0.0179771632,
0.0653781295,
0.0328406766,
-0.021195149,
0.0466154255,
0.0292869061,
-0.0076817684,
0.0616975911,
-0.010161614,
0.0916441828,
0.0594699085,
-0.0309798829,
-0.0047557349,
-0.0016272016,
0.0376501046,
-0.0034340394,
-0.0563451871,
-0.0526343435,
0.0093943886,
-0.0382010713,
-0.0167576168,
0.0026266263,
-0.0076003475,
-0.0181509294,
-0.0204073824,
0.0657668784,
-0.0044942498,
-0.0408075638,
-0.0066798907,
-0.0114510898,
0.0116277952,
-0.0341298543,
0.0112138493,
0.0848430246,
-0.0181414708,
0.0338940248,
0.0420837067,
-0.0429084785,
0.0315743163,
-0.0276756026,
-0.0016901478,
0.0473278351,
-0.0120215528,
0.0399296843,
-0.0510812812,
0.0020157075,
-0.0149616981,
0.0312960148,
-0.0581964254,
-0.0366490074,
-0.0533083268,
-0.0251713619,
-0.0526002683,
0.0313474387,
-0.0731902644,
-0.0119932592,
0.0319376998,
-0.0001926785,
-0.0093105128,
0.0267113801,
0.008882557,
0.061473202,
-0.0159468371,
0.026627399,
-0.0686158538,
0.0020820319,
-0.0744079128,
-0.0032582253,
-0.0150511581,
-0.0713503584,
-0.0277588218,
-0.0448580235,
0.0287660751,
0.0030835234,
0.0242190938,
0.0801614225,
-0.0435568579,
0.0060809199,
-0.0334373601,
-0.0193088334,
-0.0023729308,
-0.0461325273,
-0.0030381449,
-0.0062304591,
0.0714929402,
0.0130477995,
-0.0229282677,
-0.0237053856,
0.011217854,
0.0441819616,
-0.012755638,
0.0659275725,
-0.0245277788,
0.020861866,
0.0011848562,
-0.0220839847,
-0.0201221667,
0.0971521735,
-0.0084557002,
-0.0297497511,
-0.0293917358,
-0.016580781,
-0.0222805124,
0.0068535088,
0.0794368535,
-0.0036283496,
-0.0009415636,
0.0834898427,
0.0017049606,
0.0247889385,
-0.0335840248,
0.0081844041,
-0.0184876546,
0.0324715152,
0.0022508341,
-0.0087296683,
0.0056086136,
0.0522628762,
0.0005191769,
-0.0017098479,
0.0188591946,
0.0219581164,
-0.0179786626,
0.060793709,
-0.0099755181,
0.0098697664,
0.0413231514,
-0.0027520661,
-0.0248357747,
-0.0495968275,
0.0203942005,
0.0033852539,
0.0224046465,
0.0446149074,
0.02019733,
0.0442536995,
-0.0242854171,
-0.0271845516,
-0.0034969039,
-0.0303868167,
-0.0131661939,
-0.0089878375,
0.0475797467,
0.0562355965,
0.0431048572,
-0.0747463554,
-0.0278720576,
-0.0034285043,
-0.0079792421,
-0.0280620307,
0.0111848637,
-0.0505716056,
0.0043051178,
0.0242376495,
-0.0215988755,
0.0224146843,
0.0410119854,
-0.0123190396,
0.0379826389,
-0.0082082618,
0.0125352386,
-0.0641783401,
-0.0013923373,
0.0191097409,
0.0196823422,
-0.0193103999,
-0.0137437088,
-0.0064675729,
-0.0076265126,
0.0408001617,
0.0185277537,
0.0446552485,
0.0453139767,
-0.0212330017,
-0.0339169204,
-0.0167065263,
0.0063852472,
-0.0026477817,
0.0335886031,
0.0265938863,
-0.0584740601,
-0.0308553614,
-0.0342286155,
-0.0566261336,
0.0250043105,
-0.0211760253,
-0.0189578254,
-0.0346828364,
-0.0201589037,
0.032892067,
0.0051522227,
-0.0125023406,
0.079589732,
0.0211855266,
0.0338116661,
0.0021498734,
-0.1015955582,
0.0174066294,
-0.0337828174,
0.0379459485,
-0.0562464967,
-0.0123800198,
0.0166030955,
-0.0008451074,
-0.0089852707,
-0.1023198292,
0.0177645255,
-0.0681720451,
0.031627439,
-0.0274316035,
0.0067762025,
0.0438946858,
0.0381346308,
-0.016288856,
-0.0378730707,
-0.0637910366,
-0.0173441321,
-0.004299535,
0.0100170719,
0.0662919879,
0.0386315063,
-0.0206930824,
-0.0352019258,
-0.0016551353,
-0.0191284306,
0.0147786029,
-0.0147467125,
-0.03051126,
0.0297626872,
-0.0482790619,
-0.0076170512,
0.0397277176,
0.0254235826,
-0.0081124362,
-0.0312135257,
-0.0331230573,
-0.0165790319,
0.0196402278,
-0.0497649275,
-0.0413424931,
0.0275578573,
-0.0128041795,
-0.0362456813,
0.0383787788,
0.0047652344,
0.024054097,
0.0330454111,
-0.0294315107,
0.0234136712,
0.0071372283,
-0.0218653977,
-0.0212057494,
-0.0097526666,
0.0147869177,
0.0275814123,
-0.0228565242,
-0.0790080205,
0.0364881828,
0.0252702404,
0.0296798348,
-0.013279696,
0.0298944879,
0.0364770629,
-0.0189077202,
-0.0036525801,
-0.0181523468,
-0.0057773837,
-0.0744137838,
0.057976909,
0.0055988859,
0.0038181299,
-0.0228688326,
-0.0151371155,
-0.0213121809,
-0.0212004501,
0.0137939863,
-0.0215874519,
0.0967438817,
-0.0311207641,
0.0355138779,
0.0080457618,
-0.0185146406,
-0.0301641505,
-0.0261230115,
-0.0332547054,
0.0305344034,
0.0316612422,
-0.0157858972,
0.0101311598,
-0.0386825018,
0.0404904671,
0.0460917503,
-0.0041781026,
-0.0124096954,
-0.0183254462,
0.0412231907,
0.0600906052,
0.0536391623,
-0.0274834987,
0.0127396183,
0.0399638787,
0.0392128155,
-0.0088151544,
-0.064136073,
-0.0160936993,
0.0191221442,
-0.0103165675,
-0.0272244271,
0.0412697792,
0.0278321635,
-0.0260336343,
-0.050774008,
-0.015297222,
0.0550005324,
0.0435216427,
0.03976373,
-0.0092705237,
-0.0144327236,
0.0297378674,
0.0226962809,
0.0760543272,
-0.098733671,
-0.0274524242,
-0.0087724933,
-0.0270063169,
-0.0005409503,
-0.0561087169,
-0.0151112005,
-0.0944781527,
0.0146263605,
-0.0624998771,
-0.0466006361,
-0.0785539299,
-0.0121544935,
0.0229921434,
-0.0073229424,
-0.0015202495,
0.046646174,
-0.0062512858,
0.0272220131,
-0.0024421129,
-0.0190960262,
-0.0154976118,
-0.0113117676,
-0.0375221744,
0.0089485878,
0.0491181239,
0.0088493777,
-0.0075679696,
0.0080700135,
-0.0183233321,
-0.0139520122,
0.0082296124,
0.00541178,
0.0754534751,
0.0269676652,
0.0084080845,
0.0044823429,
0.0062399241,
-0.0174520835,
0.0350226648,
0.0347259641,
0.0537462831,
-0.0108522037,
0.0427118354,
0.0117215309,
-0.0135501921,
-0.0006735293,
0.0591467768,
0.0544635057,
0.0385939218,
-0.0399574526,
-0.0159723554,
-0.0763403922,
-0.0638587326,
0.049651742,
0.053060662,
0.0008651576,
0.0067586303,
-0.0652830601,
-0.0378745534,
0.0872553885,
0.0246965121,
-0.0008216382,
0.0711982027,
-0.0643359497,
0.0008285973,
-0.032380417,
-0.0304624736,
-0.0557330064,
0.0038130889,
0.0382997096,
-0.0522067063,
-0.0197876897,
-0.0429281518,
-0.0203257482,
0.0098271593,
-0.0347597897,
-0.0080608483,
-0.0251955856,
-0.0058466839,
0.0153595489,
0.0139554832,
-0.0322833769,
0.0096765226,
-0.0274841543,
-0.0042701671,
-0.1055790856,
-0.0438586995,
0.0000085398,
-0.0331461616,
0.0115497708,
0.0461879596,
-0.0420599394,
-0.0240834858,
-0.0296021532,
0.0327452868,
-0.0138472775,
0.040938776,
0.0681955963,
0.0271305852,
-0.0257186703,
0.0581908002,
-0.0575910769,
-0.0133100571,
0.0650728196,
-0.0749084875,
-0.0570712946,
-0.0134514226,
0.0137124211,
0.0014146111,
-0.0014643434,
0.0003532135,
-0.0186921209,
0.0021868653,
0.0330106653,
-0.0270842873,
-0.0185569879,
0.0695103183,
-0.0408571325,
0.007569348,
-0.0264464542,
0.003023074,
-0.0212807283,
0.0290334783,
0.0346389972,
-0.0068263756,
-0.016761383,
0.0349305421,
-0.0349805728,
-0.0472954661,
0.0251104012,
-0.085870631,
-0.0409290195,
-0.0040490828,
0.0194846299,
0.0031439413,
-0.0091478741,
-0.0058113476,
0.0128679331,
-0.0383429602,
0.0417910367,
-0.0112013444,
0.0195342973,
-0.0144648412,
0.0593943633,
0.010211518,
0.0157379508,
-0.0065874299,
0.0559835993,
0.0131025687,
0.039871715,
-0.0168049783,
-0.0093901418,
-0.0489178412,
0.0144424774,
-0.0056706844,
0.0148251057,
-0.0433292203,
0.00969557,
-0.0021002195,
-0.0476990603,
-0.0041583031,
0.0367865264,
-0.055478394,
-0.0017716871,
0.0398041122,
0.0343361124,
0.0696746856,
-0.0315026306,
-0.043114543,
0.0187109038,
-0.0115956627,
-0.029078899,
-0.0006665147,
-6.177002408e-33,
-0.0484935157,
-0.001649039,
0.0303774513,
-0.0776526481,
0.0146554895,
-0.0028010225,
-0.0454892218,
-0.0384539217,
-0.0295339935,
-0.0007905521,
-0.0381434597,
0.0466665067,
0.0075403601,
-0.0208577327,
0.0085168034,
-0.096150659,
0.0011509343,
0.0093319416,
0.0227633547,
0.0051362435,
0.0084376987,
0.0434744917,
-0.0006534943,
0.0002858797,
0.0493698902,
0.0032372456,
-0.0192835778,
0.0000463331,
-0.0131696174,
0.0423117988,
-0.0472494438,
0.0244888794,
0.0361475497,
-0.0455542803,
-0.0301749371,
-0.0153477341,
-0.1681563407,
-0.0252314266,
-0.00022329,
0.0027602436,
-0.0408791304,
-0.002110943,
0.0140806623,
-0.0141875641,
0.0386204794,
-0.0064128302,
0.0114216888,
-0.0371173248,
0.0136748943,
0.0090642916,
-0.0131870573,
-0.0062267515,
-0.0300390702,
0.0129896449,
0.0341739878,
0.0507360622,
0.0441570319,
-0.0356752463,
-0.0173612572,
0.0122805201,
0.0472041965,
0.0190026965,
0.0327809341,
-0.0364383385,
0.0026891604,
-0.0085010231,
-0.0108140213,
-0.0375245996,
-0.0546385199,
0.0263226405,
-0.0703963041,
0.0496797375,
0.0701074079,
0.0231772121,
0.0328103974,
0.0362170041,
-0.04979866,
0.0699543059,
0.0017814566,
0.0219928455,
-0.0527108461,
-0.0152261127,
-0.0324152447,
-0.0089426655,
-0.0392719433,
0.0460559912,
-0.0458383262,
-0.0243668575,
0.0247386545,
0.0059277397,
0.0032795204,
-0.0154617913,
-0.0131405387,
-0.00486323,
-0.0054677892,
0.0490696803,
0.0082158437,
-0.0068371869,
0.0265654046,
0.0152695896,
-0.0082088504,
0.0066392738,
0.0486931279,
-0.0388080776,
0.0074896491,
0.0226569548,
0.0332073197,
-0.0005065814,
-0.0575802401,
0.0000129184,
-0.033068154,
0.0608238317,
-0.0131707098,
-0.0103914477,
-0.0061661289,
0.0421485454,
-0.0025538544,
0.0571915507,
0.0337963477,
0.0352386087,
-0.0038148283,
-0.0039192024,
-0.0001634489,
-0.0067067491,
0.0175769143,
0.0230914988,
-0.01562458,
-0.0196830053,
-0.0049223574,
-0.0332363062,
0.0188505743,
0.0388675928,
2.438e-7,
0.0834604576,
-0.0236129761,
0.0583836809,
0.039908953,
-0.0097292569,
-0.0126457708,
-0.0170321111,
-0.0124825593,
-0.0223632641,
-0.0531110689,
0.0500502028,
0.0060729743,
0.034507744,
0.0403150246,
-0.0326878279,
-0.0555327535,
-0.0407165736,
-0.0304459333,
-0.0396086946,
0.0011083552,
0.0425411537,
0.0717012808,
-0.0515522994,
0.0225305129,
-0.0397953279,
-0.0334507711,
0.0090023829,
-0.0314387418,
0.0284229796,
0.0255827587,
0.0671471879,
-0.0235893335,
0.0152959805,
0.0804995522,
-0.0347102955,
0.0250443369,
-0.0027511772,
0.0068429229,
0.017020585,
0.0265746787,
-0.0297452677,
-0.0707242116,
0.0570570044,
-0.0584151298,
0.1223085523,
-0.0322095044,
0.0088578723,
0.0837366283,
-0.0303834938,
0.0117302584,
0.0031425955,
0.0340110809,
-0.0186454654,
-0.0096950568,
0.0286539886,
-0.0031135979,
0.0237268992,
-0.0426581129,
0.005283169,
0.0257935859,
-0.0149827162,
-0.0071415268,
-0.0054819118,
-0.0466397926,
-0.0334137306,
-0.0455872789,
-0.0252495483,
2.097256409e-34,
0.0517228432,
0.0118725691,
0.0323439538,
0.0615945011,
0.0762768462,
-0.0497355871,
0.000550297,
-0.031251628,
0.0258614421,
-0.0860342309,
-0.0153760938
] |
http://arxiv.org/abs/2401.15545v1 | http://arxiv.org/pdf/2401.15545v1 | PPM: Automated Generation of Diverse Programming Problems for Benchmarking Code Generation Models | 2024-01-28T02:27:38 | 2024-01-28T02:27:38 | Simin Chen, Xiaoning Feng, Xiaohong Han, Cong Liu, Wei Yang | cs.SE | cs.SE, cs.AI, cs.CL, cs.PL | In recent times, a plethora of Large Code Generation Models (LCGMs) have been
proposed, showcasing significant potential in assisting developers with complex
programming tasks. Benchmarking LCGMs necessitates the creation of a set of
diverse programming problems, and each problem comprises the prompt (including
the task description), canonical solution, and test inputs. The existing
methods for constructing such a problem set can be categorized into two main
types: manual methods and perturbation-based methods. However, manual methods
demand high effort and lack scalability, while also risking data integrity due
to LCGMs' potentially contaminated data collection, and perturbation-based
approaches mainly generate semantically homogeneous problems with the same
canonical solutions and introduce typos that can be easily auto-corrected by
IDE, making them ineffective and unrealistic. In this work, we propose the idea
of programming problem merging (PPM) and provide two implementation of this
idea, we utilize our tool on two widely-used datasets and compare it against
nine baseline methods using eight code generation models. The results
demonstrate the effectiveness of our tool in generating more challenging,
diverse, and natural programming problems, comparing to the baselines. | PPM: Automated Generation of Diverse Programming Problems for Benchmarking Code Generation Models: In recent times, a plethora of Large Code Generation Models (LCGMs) have been
proposed, showcasing significant potential in assisting developers with complex
programming tasks. Benchmarking LCGMs necessitates the creation of a set of
diverse programming problems, and each problem comprises the prompt (including
the task description), canonical solution, and test inputs. The existing
methods for constructing such a problem set can be categorized into two main
types: manual methods and perturbation-based methods. However, manual methods
demand high effort and lack scalability, while also risking data integrity due
to LCGMs' potentially contaminated data collection, and perturbation-based
approaches mainly generate semantically homogeneous problems with the same
canonical solutions and introduce typos that can be easily auto-corrected by
IDE, making them ineffective and unrealistic. In this work, we propose the idea
of programming problem merging (PPM) and provide two implementation of this
idea, we utilize our tool on two widely-used datasets and compare it against
nine baseline methods using eight code generation models. The results
demonstrate the effectiveness of our tool in generating more challenging,
diverse, and natural programming problems, comparing to the baselines. | [
-0.0349135362,
0.0670507476,
-0.036242716,
0.0014706869,
-0.0491488278,
-0.0028758573,
-0.0080265664,
-0.0117971804,
-0.0155541981,
-0.0090727322,
0.0597537421,
0.0555031337,
0.0160406828,
0.0360011607,
0.020630395,
-0.0218370706,
0.0352280438,
-0.0147915222,
-0.0561188497,
0.0570766218,
-0.0273872465,
-0.0045318659,
0.0037258877,
0.0297077969,
-0.0009796242,
-0.0177181978,
-0.0184151139,
0.0483592227,
0.0224602353,
-0.0277294088,
0.0182358548,
0.0479719639,
-0.0086098416,
0.01333695,
0.0000025314,
-0.0497695059,
-0.0157722738,
0.0198462456,
0.0140098324,
0.0196465421,
0.0070405705,
0.0507757217,
0.0074366252,
0.0331891552,
-0.0415463634,
-0.0124964919,
0.0187259763,
-0.0007383996,
0.0917308852,
0.0106933834,
0.0097602746,
0.0128833754,
-0.0028558199,
0.0180761032,
0.0480812602,
-0.0104124844,
0.015693767,
0.059399426,
0.0337693505,
-0.0452807993,
0.0243728086,
-0.0147449896,
0.0395451821,
0.0180492327,
0.0144083016,
0.0136033157,
-0.0547233634,
-0.0283579752,
-0.0006921573,
0.0405129232,
-0.0701197907,
-0.0394913368,
0.0395498648,
-0.0124544129,
-0.0103959739,
-0.0185839906,
-0.0160888936,
-0.0197498575,
-0.0214044005,
-0.0022351721,
-0.0103427311,
-0.0011542999,
-0.0025894593,
-0.0156174544,
-0.0369378515,
-0.000423318,
0.0125849489,
-0.0070152706,
-0.0258897636,
-0.0493460298,
-0.0166258849,
-0.0419126749,
0.0120443227,
0.0308547188,
-0.0436964259,
-0.0161959901,
-0.046735622,
-0.0710196942,
0.034509901,
-0.0078129321,
0.0535320677,
0.0433644392,
0.0157808643,
-0.0058558369,
-0.0306754597,
0.0786743164,
-0.048770126,
-0.0407416001,
-0.0294296537,
0.0550239235,
-0.0696499571,
0.0132719567,
-0.0304317791,
0.0236707516,
0.0010810617,
0.0539578497,
-0.0129473647,
0.0501983985,
-0.0256188009,
0.055250369,
-0.0112745827,
0.0149035128,
0.0035041252,
0.0113823581,
0.0461780615,
-0.0623049773,
-0.0327135697,
-0.0428908505,
-0.0049249148,
-0.0277778208,
0.0102756787,
0.000953096,
-0.0351392068,
-0.0049421331,
0.027494451,
0.0687386692,
-0.0120313605,
0.0608960651,
-0.0066895396,
-0.080001533,
0.0505378731,
-0.0140445866,
-0.0499706417,
-0.109790042,
0.0309679415,
-0.0312599204,
-0.0390379317,
-0.050411433,
0.030723555,
0.0228869673,
0.0326006822,
0.038275227,
-0.0417497717,
0.0513823219,
0.0617153384,
0.0386120379,
-0.0491665937,
0.0228427835,
0.0118354699,
-0.03183515,
-0.0110096317,
-0.0288070049,
0.0259911623,
0.0037123787,
0.0117034512,
-0.0023723026,
0.0193855111,
0.0161115881,
0.0584719591,
-0.0035602904,
-0.0394862555,
0.0416183844,
-0.0207336023,
0.0212882068,
-0.0333410576,
0.0077012638,
-0.0254054386,
0.0396643877,
0.0090970462,
0.051774092,
0.0044423258,
0.0103551792,
-0.0101828258,
0.0467846319,
0.0059710527,
0.0013218645,
-0.0055833026,
0.0874935761,
0.0019938089,
-0.0517907403,
-0.0148820942,
0.0169055983,
-0.0354171656,
0.0012949198,
0.0042799269,
0.0001800608,
-0.0650698096,
0.0212239791,
-0.0184911862,
-0.0817760825,
-0.0367187746,
-0.0206565596,
0.0264582206,
-0.0049279593,
0.0294687338,
-0.0600223839,
-0.0481498353,
0.0253989529,
0.0103789549,
-0.0759491101,
0.0598332509,
-0.0077520395,
-0.0013502878,
-0.0554625727,
-0.033878006,
-0.0295776892,
-0.0265264176,
0.0519597679,
0.0451354347,
-0.0603383407,
0.0610091239,
-0.0001817235,
-0.0464457795,
0.0436639935,
-0.0200972874,
-0.0009691357,
0.0793380365,
-0.0387981012,
0.0134528717,
-0.0281963404,
-0.0291115474,
0.0219601858,
0.0454674475,
-0.0290507544,
-0.0306731984,
0.0076137865,
-0.008498962,
-0.0097596971,
0.0237895828,
-0.0801910684,
-0.0456862375,
0.0321326517,
-0.070491232,
-0.0275765881,
0.0331212133,
0.013614025,
-0.0439723358,
0.0351652056,
0.0467672758,
-0.0239343699,
-0.0212821756,
-0.0157604367,
0.0618150756,
0.029764045,
0.0076255477,
-0.0062572001,
0.0580892898,
0.037847016,
0.0044544023,
-0.0087353932,
-0.0115101393,
-0.0605515614,
-0.0038038369,
0.0081351623,
0.0450902991,
-0.0142426249,
-0.0489080958,
-0.0709802061,
0.0558972396,
0.0360291898,
0.028422527,
0.0515421145,
0.0342550427,
-0.0021628025,
0.0285250247,
-0.0282003321,
-0.0495482087,
0.003539179,
0.003719541,
0.030414531,
-0.0188175216,
-0.0545898229,
-0.0003084225,
-0.0074394788,
-0.0550832488,
0.0430596583,
-0.0228688773,
-0.0019336971,
-0.0336510837,
0.0891059786,
-0.0018660439,
0.0010882033,
-0.0079961056,
-0.0170620773,
0.0081504416,
-0.0128631927,
0.1226967722,
0.0065884674,
-0.0127156833,
-0.0191258099,
0.0435084328,
-0.0556003638,
-0.0523122177,
0.0236652754,
0.0584173761,
-0.0123502072,
0.0902348235,
-0.0813016668,
0.0467825681,
0.0363369249,
-0.0080936728,
-0.0377659723,
-0.0309869517,
0.0419971831,
-0.0078885974,
0.0265973825,
-0.0033062997,
-0.0585229956,
0.0146488855,
0.0505343527,
-0.0351280645,
-0.0635861456,
0.0670610592,
-0.0156373791,
0.0245961137,
-0.026861364,
-0.0161889084,
-0.0643784702,
0.0293431114,
-0.0104212426,
0.027116023,
-0.029094426,
-0.0085214367,
-0.0369780362,
0.0154361753,
0.0242409129,
-0.0383944474,
0.0329285003,
-0.0571382232,
-0.0052001271,
-0.0071255825,
-0.0238207094,
-0.052709695,
0.0187459141,
0.0500403382,
-0.0434011221,
0.0342437625,
0.0139936265,
0.0292622708,
-0.0053201122,
0.0112518892,
0.0197502282,
0.0282767434,
-0.0192000661,
0.019263301,
-0.0317170769,
-0.0344520323,
-0.0426239111,
-0.0294363331,
0.0172517393,
0.0351031497,
0.0110869007,
0.0757654831,
0.0488513522,
-0.0603019483,
-0.0146103492,
0.0005018391,
-0.0136046233,
-0.0317660607,
0.0408432931,
0.0046825013,
-0.0540504456,
0.0061012735,
0.0225771591,
0.0605093241,
-0.0129429055,
0.0045079356,
-0.002518432,
0.0392490551,
0.0784506053,
0.0573627204,
-0.0879010558,
-0.0020199111,
0.0031984979,
0.0077293273,
-0.0586070344,
-0.0115297018,
-0.0612465739,
-0.091685988,
-0.0121759614,
0.0218002051,
-0.0100121927,
-0.0195634533,
0.0014092933,
-0.0085006049,
-0.0022513156,
-0.0647012219,
0.038492728,
0.0327776968,
0.0093843834,
0.002178211,
-0.0337437093,
-0.0522153191,
0.0015474702,
0.0026388979,
-0.0176384468,
0.0747043192,
-0.0075629624,
0.0304197408,
0.0051122666,
0.0543118343,
0.0414032936,
0.0147571824,
-0.0237933937,
0.0103595788,
0.0525038168,
-0.0184088405,
0.030059386,
0.0068447762,
-0.0753468275,
-0.0726947635,
0.0634539798,
0.0315975174,
0.0258710813,
0.0124124801,
-0.0283799469,
-0.015037749,
0.0105855586,
0.0151469177,
0.0299076606,
0.020927012,
0.000955548,
-0.0420486592,
-0.0185911767,
-0.049400039,
0.0582189746,
-0.026365649,
0.011916968,
-0.0115574626,
-0.0211057682,
0.0486541018,
0.0329326019,
0.0485836416,
0.0220098626,
-0.0649641231,
-0.0174011793,
-0.0163679756,
-0.0877662227,
0.0186058488,
0.0364373662,
0.0184707902,
-0.0164361615,
-0.0107439412,
-0.0185391251,
0.0126889041,
0.0007907044,
0.0394254215,
-0.1202344373,
-0.084011972,
-0.037951421,
0.0401711129,
0.0116029307,
-0.0243717283,
-0.0226146281,
0.0016945443,
-0.0337807313,
0.058896102,
0.0055826362,
-0.0108585758,
-0.0234669819,
-0.0357019454,
0.0372860618,
-0.0623378158,
-0.0588370971,
-0.0070054075,
-0.0221694577,
-0.0275447946,
-0.0325865373,
0.0400444157,
0.0610392801,
0.052115228,
-0.0295491666,
0.0144694084,
-0.0544762835,
0.0021011017,
-0.0244543403,
-0.0206018351,
-0.063255854,
-0.0321681276,
0.0494582243,
-0.0288207624,
0.0079469047,
0.0903868973,
0.0251736715,
-0.0091931596,
0.0280841775,
0.017518552,
0.0361052938,
-0.0229879469,
0.0246286374,
0.0628484413,
-0.0003357982,
-0.025848005,
0.0035685061,
-0.0174604375,
0.014174928,
-0.018112829,
0.0164719671,
0.0346787721,
0.0874503031,
0.0077349045,
-0.0133518679,
-0.0404870212,
-0.0461234078,
-0.0108709456,
-0.0212083794,
0.0022511391,
0.0620237328,
-0.029501576,
-0.0162044503,
0.0022295003,
-0.0041411184,
0.0034848773,
-0.0295001715,
0.0011452304,
-0.0001434873,
0.0453072712,
0.0214443896,
-0.0763493851,
0.0885691568,
-0.0341418013,
-0.0340090469,
-0.0176471863,
0.0224986132,
-0.038313102,
0.0465312414,
0.0807562768,
-0.0086907875,
0.0171933901,
-0.0719143301,
-0.0362546183,
-0.0313777737,
0.0212424975,
0.0241360571,
0.027823912,
0.047486905,
0.0502000675,
0.0324340276,
-0.0034912601,
0.0129194101,
0.0435997583,
-0.0238952152,
-0.0281494148,
-0.0529456995,
-0.0102953259,
-6.269106799e-33,
-0.0043253736,
-0.0408127569,
-0.0220425539,
-0.0236124005,
0.0341593772,
-0.0193389598,
-0.0304796435,
0.0043946933,
0.0445438698,
-0.0591040812,
0.0252173655,
0.018752357,
-0.0203860551,
-0.0519728139,
0.0172830392,
0.010557645,
0.0434696153,
0.0267859846,
-0.0675517991,
-0.0056826817,
-0.0463100597,
0.005853469,
-0.042424649,
-0.0291665234,
-0.0312560685,
-0.0295007266,
0.0095932735,
-0.0120077915,
0.0335146748,
0.0207931362,
0.0294047687,
-0.0032005222,
-0.0351834409,
0.0072265961,
-0.0342156626,
-0.0091030775,
0.0035124754,
-0.0716545805,
0.0252449084,
-0.0186768714,
0.0071988506,
-0.051496204,
0.020414209,
0.0001859747,
-0.0031836091,
0.0047371727,
0.0057040649,
0.0266334321,
0.0385979041,
-0.0401575007,
-0.0160775892,
0.0193620268,
-0.0531977043,
-0.012387516,
-0.0263774954,
0.0310445894,
-0.006373303,
0.0394525975,
0.053251069,
0.0197225995,
0.0413087644,
0.0221909918,
0.0800436735,
-0.0296320412,
0.0193300135,
-0.0262440629,
0.0016698228,
-0.0086830165,
-0.037731193,
0.0292531382,
-0.0079667894,
0.0468851775,
0.0137443356,
0.0562780872,
0.0756542534,
-0.0096725179,
-0.0125978431,
0.0525070652,
0.0012067098,
-0.0474327616,
-0.0078003258,
0.019828273,
-0.0185616463,
-0.0275091305,
-0.0395239331,
-0.0049958779,
-0.0370525829,
-0.020834066,
0.0500019416,
-0.0131336004,
-0.0434372798,
0.0288439821,
-0.0083387215,
-0.0138242561,
0.0155110843,
0.0137940729,
-0.0495919101,
-0.0005823133,
-0.0171009786,
0.0118648689,
0.0202251878,
-0.0481589213,
0.0514419526,
0.0064601665,
0.1033438817,
0.0234585498,
-0.0137649532,
-0.0093613416,
-0.0571056493,
0.0227194428,
-0.0336282998,
-0.0141809061,
-0.0067658504,
0.007798986,
-0.003580275,
0.0156755615,
0.0060907458,
-0.0016367844,
0.0286182314,
-0.0416716412,
-0.0567716695,
0.0288712028,
0.0087059857,
0.0321136937,
0.0386042744,
-0.030208502,
-0.0102896821,
0.0002061135,
0.0625641197,
0.0317062624,
-0.080115661,
0.0402904153,
3.241e-7,
-0.0237046834,
0.0358713977,
0.0055818981,
0.0205287244,
-0.0454742126,
0.0117554674,
-0.0193240866,
-0.0165072531,
-0.0092805671,
-0.0081596021,
-0.0268499833,
0.0181039665,
0.0504424386,
0.0643360317,
0.001453355,
0.0077739395,
0.0326472744,
-0.0174417533,
-0.0802385509,
0.0323110148,
0.0969805494,
0.0033496099,
0.0066985912,
0.0198794007,
0.0088923164,
0.0022589855,
-0.0063724061,
-0.0282430369,
0.0290609691,
0.0542252548,
-0.08405523,
-0.0297559891,
0.0045556929,
0.0504910871,
-0.0085247122,
-0.0325445868,
0.0295466166,
0.0220158156,
0.0139167374,
0.0264373012,
0.0495483652,
0.015844686,
-0.0250232648,
-0.0363831334,
0.0087311035,
-0.0391625315,
-0.0257066097,
0.0013183354,
-0.0119271949,
-0.0267188679,
-0.0098543372,
-0.0108407745,
0.0318234563,
0.0028425879,
-0.0219971295,
0.0111325588,
-0.0375456698,
-0.0001653899,
-0.0450002924,
0.0174967535,
-0.0285082571,
-0.0336397365,
-0.0278562885,
0.0275603272,
0.0297177527,
0.0215546694,
0.0328354388,
3.596013773e-34,
0.0056909369,
-0.0364077687,
0.0337706804,
-0.0056591984,
0.0275879484,
-0.0048911967,
-0.0448169596,
0.0207858384,
-0.0156932212,
0.0252951849,
-0.0152609255
] |
http://arxiv.org/abs/2401.15535v1 | http://arxiv.org/pdf/2401.15535v1 | Quantifying Stereotypes in Language | 2024-01-28T01:07:21 | 2024-01-28T01:07:21 | Yang Liu | cs.CL | cs.CL | A stereotype is a generalized perception of a specific group of humans. It is
often potentially encoded in human language, which is more common in texts on
social issues. Previous works simply define a sentence as stereotypical and
anti-stereotypical. However, the stereotype of a sentence may require
fine-grained quantification. In this paper, to fill this gap, we quantify
stereotypes in language by annotating a dataset. We use the pre-trained
language models (PLMs) to learn this dataset to predict stereotypes of
sentences. Then, we discuss stereotypes about common social issues such as hate
speech, sexism, sentiments, and disadvantaged and advantaged groups. We
demonstrate the connections and differences between stereotypes and common
social issues, and all four studies validate the general findings of the
current studies. In addition, our work suggests that fine-grained stereotype
scores are a highly relevant and competitive dimension for research on social
issues. | Quantifying Stereotypes in Language: A stereotype is a generalized perception of a specific group of humans. It is
often potentially encoded in human language, which is more common in texts on
social issues. Previous works simply define a sentence as stereotypical and
anti-stereotypical. However, the stereotype of a sentence may require
fine-grained quantification. In this paper, to fill this gap, we quantify
stereotypes in language by annotating a dataset. We use the pre-trained
language models (PLMs) to learn this dataset to predict stereotypes of
sentences. Then, we discuss stereotypes about common social issues such as hate
speech, sexism, sentiments, and disadvantaged and advantaged groups. We
demonstrate the connections and differences between stereotypes and common
social issues, and all four studies validate the general findings of the
current studies. In addition, our work suggests that fine-grained stereotype
scores are a highly relevant and competitive dimension for research on social
issues. | [
0.0268540047,
0.0879393741,
-0.0140405158,
0.0273152534,
0.017510565,
0.0592981055,
0.0275285803,
-0.0364827737,
0.0089303106,
-0.1068701148,
-0.0129224015,
-0.0164348874,
0.0153546166,
0.0205874462,
0.03330056,
-0.048597347,
0.0342445672,
0.0698517859,
0.045283746,
-0.0435044952,
0.0390114821,
0.029495718,
0.032369595,
0.0367088094,
-0.0319912508,
-0.0265525393,
0.0765210241,
-0.044083152,
0.0059477063,
-0.0115148127,
0.0895165876,
0.0644231439,
0.0285019353,
0.0044285655,
0.0000018088,
-0.0695143789,
0.0493658669,
-0.0037021206,
0.0168658998,
0.0014782564,
0.0321211517,
-0.0250136293,
0.0206719656,
0.0361242518,
-0.0347478315,
-0.037222065,
0.0286024269,
0.0386627316,
0.0323583148,
0.0167258475,
-0.008794005,
-0.0626021773,
-0.0759559721,
-0.0183190536,
0.001654099,
0.0267720353,
0.0638285503,
-0.0474775247,
-0.104002811,
-0.045349922,
-0.0105509153,
-0.0328724086,
-0.0081320209,
0.0736699328,
0.0303339567,
0.0043902039,
-0.0797834247,
-0.0391494818,
0.0117807444,
0.0697600022,
-0.0016083339,
-0.0350404531,
-0.0185600631,
-0.0079207029,
0.0003481932,
0.0021623317,
0.0163060464,
0.0409854278,
-0.0138837704,
0.017777238,
-0.0130582303,
0.0146173974,
0.0231833905,
0.0403524935,
-0.0177883301,
0.0355983116,
-0.0017973452,
-0.0151641527,
-0.0012006906,
0.0058049732,
-0.0330625214,
-0.0162669085,
-0.0162836276,
-0.0174435135,
0.0250390675,
0.0104403282,
-0.0195852127,
-0.0131446607,
0.0329567268,
-0.0742832273,
0.0270750038,
0.0219833683,
-0.0194981135,
0.0142186945,
0.0163978022,
-0.0127748381,
0.0247459076,
0.0048450492,
0.025460178,
-0.0027316865,
-0.0292809475,
0.0033987376,
0.0111203697,
0.0196621362,
0.0727945715,
0.0513842776,
-0.0508086905,
-0.0459491089,
-0.0496115908,
0.0123926159,
-0.0146864569,
0.0270721428,
-0.0161809158,
-0.012681447,
-0.0729142576,
0.0124577023,
-0.0598940477,
0.0380562842,
-0.0898539722,
-0.0788198933,
0.0329586007,
0.0164637603,
-0.0012263882,
-0.0271272715,
0.0220675208,
0.0513857156,
0.0302920677,
-0.0288654566,
-0.010296938,
-0.0234538782,
-0.0409947746,
-0.0373230055,
-0.011215644,
-0.0409838296,
0.0224813279,
0.0169594083,
0.0063938666,
-0.0684340298,
0.0016622938,
0.024881769,
-0.0484141447,
0.0883152932,
0.0374559797,
-0.0198472496,
0.0315785035,
0.0275129024,
-0.0166735388,
0.0639314204,
0.0123509895,
-0.0031669256,
0.0427983813,
-0.0180740338,
0.0094952434,
0.027400529,
0.0329959318,
-0.0116308993,
-0.012275978,
0.0441043973,
0.0219298713,
-0.0292403325,
0.0210405458,
0.0283422396,
0.0185878221,
-0.0470460989,
0.0350648165,
0.0762029067,
-0.0204910226,
0.0335231163,
0.0310893748,
0.0128449239,
-0.0033849217,
-0.0014778455,
-0.0594538301,
-0.0371089168,
0.0080617517,
-0.0412403233,
-0.0053275307,
0.0694501549,
-0.0147320395,
-0.0387145206,
-0.0352838039,
0.0005594633,
0.0572457463,
-0.0220547691,
0.024124058,
0.0246246886,
-0.0099165142,
0.046128381,
0.0368583351,
-0.0322858617,
-0.0158361681,
-0.0220467076,
0.0270757787,
0.0212602336,
0.0126430942,
-0.0222495105,
-0.0025868807,
-0.0505932868,
-0.0179900825,
0.0036902865,
0.0510108024,
-0.0058112522,
0.0177865103,
-0.0124534694,
-0.0024532019,
0.0273821969,
0.040653225,
0.0168668944,
0.0591359027,
0.0327710994,
0.0278984401,
-0.0113230655,
-0.0388624221,
0.0240199305,
0.0146706104,
-0.0327125862,
-0.020423511,
0.0440543853,
0.0456319824,
0.0011899766,
-0.0255276654,
-0.0239010043,
-0.0256779063,
-0.0024501602,
0.0121750701,
0.053768374,
-0.0503285713,
-0.0125933662,
0.0082939574,
-0.0756519288,
-0.0601724163,
0.0540390834,
-0.0305265561,
0.0232701898,
0.0562195145,
0.0111884186,
0.0339895822,
-0.0259108432,
-0.0352108851,
-0.0749179572,
0.0421836488,
-0.0147288535,
0.0589969195,
0.0186253544,
-0.0013421064,
-0.0224497542,
-0.0735246614,
0.0580527,
0.0211289953,
-0.0504718721,
-0.0359771289,
-0.003802551,
0.006222792,
0.0264572632,
-0.0240939185,
-0.0217118394,
0.0311227236,
-0.0771051869,
-0.0110582467,
-0.0086692069,
-0.0111983214,
-0.00413715,
-0.0124127688,
-0.0134357046,
-0.0058006365,
-0.0450398214,
-0.0259292033,
-0.0151936756,
0.0228512604,
0.0407006182,
-0.0376338847,
0.0195575561,
-0.0335648507,
-0.0019845616,
-0.041494932,
0.0633420199,
-0.0054077161,
0.0024842883,
-0.000090317,
0.0203268714,
-0.0573872179,
0.0300208945,
0.0237443577,
-0.0413675271,
0.0107990867,
0.0513581298,
0.0132756084,
0.0480143987,
0.0304786433,
0.0317528024,
0.0452015512,
0.0219095815,
-0.0176580511,
-0.0421152748,
-0.0378594324,
0.0439415835,
0.0092733949,
-0.05791834,
-0.0172016229,
0.0879405215,
0.0540292189,
-0.0531451181,
-0.055489786,
0.0660511553,
-0.0262348354,
-0.0228128526,
-0.0189932995,
-0.0305161774,
-0.0213588197,
0.0136629194,
-0.0347527526,
-0.0462502465,
0.044133801,
0.0068759066,
0.0252708066,
0.0122134425,
0.0157693308,
-0.0207491014,
0.0527453013,
0.0349984504,
0.0026273557,
0.0012814184,
0.0187587161,
-0.0539774597,
-0.0497827344,
0.0326244719,
-0.0251980629,
0.0059758658,
-0.0269617364,
0.0018850283,
-0.0103418753,
-0.0389903709,
-0.043940708,
-0.0730940551,
-0.0591408275,
-0.0239756703,
-0.0045352583,
-0.012702439,
0.0044451896,
-0.0208105836,
-0.0299496595,
-0.0134754041,
-0.0043721632,
-0.0045551681,
0.0206197519,
-0.0197779275,
0.0024904171,
-0.0137494812,
-0.0018628279,
0.0276233237,
-0.0176247507,
0.0150883766,
-0.0608244948,
0.0671562478,
-0.0421175919,
0.0016622992,
0.0093886983,
-0.0069764247,
0.0016323712,
-0.008133308,
0.0342938006,
0.0004450897,
0.0100728003,
0.1004160792,
0.0226262268,
-0.0135216406,
0.017687453,
0.0463233367,
0.0199729726,
0.049123425,
-0.0182057954,
-0.1075753719,
-0.0147008086,
-0.0184643716,
0.01170664,
-0.0468670465,
-0.0226899926,
-0.0206899568,
-0.0833823383,
0.0010408222,
0.0031155969,
-0.0540606715,
-0.0231375303,
0.0000919198,
0.0216086395,
-0.0244781673,
0.0169895291,
-0.0437794775,
0.0037796635,
0.0389573425,
0.0167319998,
-0.0811914504,
-0.0414331257,
0.0638164431,
-0.0154131362,
0.0568916723,
0.0145801017,
0.0434073284,
-0.0086486638,
0.0179372877,
0.0329757668,
-0.0228390098,
-0.0219021868,
-0.0220460948,
0.0398614183,
-0.0575266406,
-0.0445244424,
-0.0204541013,
0.0255310256,
-0.0124703106,
0.0095129013,
0.0013730311,
0.0444911197,
0.0287704114,
-0.0427117571,
-0.0347328633,
0.0127542699,
-0.0026903558,
0.1111242622,
0.0340843536,
-0.0062587718,
0.0135778021,
-0.0693595633,
-0.0002313565,
-0.0147689693,
-0.0666248649,
0.03448027,
-0.0178815648,
-0.0187692046,
-0.0100322254,
-0.0724873245,
0.0585936457,
-0.0363794118,
0.0174970441,
-0.0430185162,
0.0270089135,
-0.0003084299,
-0.1096339151,
0.0350235999,
-0.0153869316,
0.0149549525,
0.0449464396,
-0.0481407568,
0.0325652994,
-0.0233744942,
-0.0562747791,
0.0044954615,
-0.0362198949,
-0.0393363908,
0.0378116295,
0.0121721756,
0.0142499758,
0.0279150028,
0.0148481848,
-0.0530031547,
0.0649565309,
0.0513641238,
0.0031677999,
0.0001888796,
0.0152892834,
0.0243062023,
0.0109074097,
0.018988505,
-0.0156501681,
0.0668466166,
0.0099511659,
-0.0715196058,
-0.0444050804,
-0.0364837348,
-0.0127401678,
-0.0164902117,
0.0470623523,
-0.0129446946,
-0.0113814296,
-0.0475125387,
-0.0289685484,
-0.0241492633,
-0.075485602,
0.0182251986,
-0.0421437323,
0.0047964687,
0.0144007783,
-0.054244224,
-0.0186174568,
-0.067521736,
0.053199701,
-0.0374824144,
0.0215410404,
0.0606147647,
0.0238662418,
0.0137017965,
0.0009246632,
-0.0007770712,
0.0182925723,
-0.0596702509,
-0.0026123514,
0.0198667347,
0.0158831067,
0.0136239063,
0.0626533926,
-0.0035850417,
-0.0181447268,
0.0223426893,
-0.0390871838,
-0.0236051604,
-0.015252091,
-0.0111106699,
-0.0522063002,
0.0185623914,
-0.0020605065,
-0.0214803554,
0.0102818673,
0.0366209336,
-0.0124037899,
-0.0006481428,
0.0578899607,
0.0018354034,
-0.0343136974,
0.0058864807,
0.0574369319,
0.0068161101,
0.0326203331,
0.0237179883,
-0.0655874908,
-0.0305441692,
0.0096872263,
0.041405011,
0.0029108142,
0.0393622294,
0.0039300411,
-0.0758884996,
-0.0660435408,
-0.0112826265,
-0.0109817097,
0.0213906392,
0.0279026702,
0.012733981,
0.0508841053,
-0.0075374548,
-0.0075820675,
-0.0050225896,
0.0037487766,
-0.0296208337,
-0.0147483395,
-0.0113318898,
-6.078763774e-33,
-0.0085042417,
-0.0234334152,
-0.0370741636,
-0.0082452344,
0.0062177647,
-0.0203357935,
0.0074696671,
-0.0103044892,
-0.0430210195,
0.0388622321,
0.0141113549,
0.0294014141,
0.0025826537,
0.0003695188,
0.0273999218,
0.0024159008,
0.045453582,
-0.0351626575,
-0.0266345907,
-0.0331978649,
0.0166502148,
0.0059798961,
0.0621254966,
-0.0361951701,
0.0614823215,
-0.0216336381,
-0.0170532912,
-0.0028792999,
0.029957436,
0.0242158864,
-0.0379264466,
0.0926712155,
0.0476002172,
0.0018129415,
0.014943338,
0.1194996461,
-0.0592467412,
-0.0299335793,
0.0067846412,
0.0348781422,
0.0290638339,
-0.0111139091,
-0.0008852295,
0.0032200534,
-0.0445470922,
0.0043092403,
0.0839466527,
-0.0201133713,
0.0580340624,
-0.0169604141,
-0.0587325655,
0.0104953339,
-0.0343106501,
0.0076349322,
-0.0165398289,
0.0182988681,
0.0153542459,
0.0419138446,
-0.0309122652,
0.0532547086,
-0.0321623199,
-0.0149754733,
0.0059007914,
0.0126685305,
0.0353340767,
0.0732107013,
0.0262305196,
0.0534413867,
-0.0128198033,
0.0092569683,
0.0477796569,
-0.035269998,
0.0035183337,
0.0702412724,
0.0538658872,
-0.0338861458,
-0.010982207,
-0.0063102646,
0.0704002455,
-0.0248469058,
0.0173890926,
-0.0295538884,
-0.0646282658,
-0.0590659678,
-0.0341855697,
-0.052544985,
-0.0164976958,
-0.01457824,
0.0144545771,
-0.037266165,
-0.0245989058,
0.0226930324,
0.0140902502,
-0.0290788468,
-0.0258504208,
0.0262925699,
-0.0379631668,
-0.0341217816,
0.0220152475,
-0.0034027337,
-0.0792173222,
-0.0017988952,
0.006728767,
0.035033647,
0.0560778938,
0.0102435322,
0.0446787737,
0.0049812831,
-0.0759558305,
0.0114952195,
0.0015800502,
-0.0240487177,
0.0012483282,
0.0422510244,
0.0171201024,
0.0528705306,
0.0009806708,
-0.0113293296,
0.0166239627,
-0.0303312968,
-0.0263847895,
0.00940366,
-0.0103335837,
0.0349499285,
-0.0256682634,
-0.0334729068,
-0.0212121289,
-0.0086182449,
0.0439841449,
-0.037505094,
-0.0031888802,
-0.0729284063,
2.303e-7,
0.0448176041,
0.0564112552,
-0.0262988228,
0.0494426824,
0.0341417342,
0.0454436205,
0.047538653,
0.0383177847,
-0.0049332101,
-0.0080171051,
-0.010311937,
0.0287740529,
-0.0120506939,
0.0310191773,
-0.0330350511,
0.0049820696,
-0.0601107776,
0.0283235479,
-0.0199436359,
-0.0165523198,
0.0540784001,
0.0881377235,
0.0350587517,
-0.0097340029,
-0.0420175642,
-0.0230232198,
0.0275217481,
-0.0454223491,
0.0103577645,
0.002007927,
-0.0090714684,
0.0577665381,
-0.0073754163,
0.0188318379,
-0.0050722384,
-0.003250913,
0.0031896469,
-0.0086310217,
-0.0130060539,
0.0124632232,
-0.0158331562,
0.0442108996,
-0.0362963416,
-0.0566713624,
0.0563057028,
-0.0556448996,
-0.049122896,
-0.0469911397,
-0.0164024457,
-0.0030413282,
0.04829932,
0.009519781,
-0.0071048471,
-0.0158981979,
0.0294713117,
-0.0163193345,
0.0008110727,
-0.0753986463,
-0.0220712386,
0.0312051009,
-0.0180365443,
-0.0565453097,
0.0132663734,
-0.0218038633,
0.0342942178,
-0.0127943195,
0.0289279241,
1.620778613e-34,
0.0396033265,
0.0104001965,
0.0033739018,
-0.033071015,
0.0405718535,
-0.0264556278,
-0.0155434748,
-0.0172091294,
0.0626646876,
-0.021984525,
-0.0018811174
] |
http://arxiv.org/abs/2401.15532v1 | http://arxiv.org/pdf/2401.15532v1 | Byte Pair Encoding Is All You Need For Automatic Bengali Speech Recognition | 2024-01-28T00:41:21 | 2024-01-28T00:41:21 | Ahnaf Mozib Samin | cs.CL | cs.CL, cs.SD, eess.AS | Byte pair encoding (BPE) emerges as an effective tokenization method for
tackling the out-of-vocabulary (OOV) challenge in various natural language and
speech processing tasks. Recent research highlights the dependency of BPE
subword tokenization's efficacy on the morphological nature of the language,
particularly in languages rich in inflectional morphology, where fewer BPE
merges suffice for generating highly productive tokens. Motivated by this, our
study empirically identifies the optimal number of BPE tokens for Bengali, a
language known for its morphological complexity, thus enhancing
out-of-distribution automatic speech recognition (ASR) performance.
Experimental evaluation reveals that an excessively high number of BPE tokens
can lead to overfitting, while approximately 500-1000 tokens result in superior
OOV performance. Furthermore, we conduct a comparative analysis of BPE with
character-based and unigram-based tokenization methods. By introducing BPE
tokenization to Bengali ASR, we achieve a substantial reduction in the word
error rate (WER) from 66.44% in our character-based baseline system to 63.80%
on the LB-ASRTD eval set and from 46.34% to 42.80% on the SHRUTI eval set, both
of which include out-of-distribution data. | Byte Pair Encoding Is All You Need For Automatic Bengali Speech Recognition: Byte pair encoding (BPE) emerges as an effective tokenization method for
tackling the out-of-vocabulary (OOV) challenge in various natural language and
speech processing tasks. Recent research highlights the dependency of BPE
subword tokenization's efficacy on the morphological nature of the language,
particularly in languages rich in inflectional morphology, where fewer BPE
merges suffice for generating highly productive tokens. Motivated by this, our
study empirically identifies the optimal number of BPE tokens for Bengali, a
language known for its morphological complexity, thus enhancing
out-of-distribution automatic speech recognition (ASR) performance.
Experimental evaluation reveals that an excessively high number of BPE tokens
can lead to overfitting, while approximately 500-1000 tokens result in superior
OOV performance. Furthermore, we conduct a comparative analysis of BPE with
character-based and unigram-based tokenization methods. By introducing BPE
tokenization to Bengali ASR, we achieve a substantial reduction in the word
error rate (WER) from 66.44% in our character-based baseline system to 63.80%
on the LB-ASRTD eval set and from 46.34% to 42.80% on the SHRUTI eval set, both
of which include out-of-distribution data. | [
-0.0030251141,
0.0006670992,
0.0138820447,
0.0121583231,
-0.0482440479,
0.0188647714,
-0.0080653075,
0.049011033,
-0.040840406,
0.0007565275,
-0.0093417596,
-0.0628117993,
0.0765538067,
0.0344902016,
0.0091426251,
-0.0324805416,
0.0377285406,
-0.0490530618,
0.0311815348,
0.0017234285,
0.0192402303,
0.0220334046,
-0.0236115474,
0.0048609069,
-0.0326907746,
0.033000011,
0.013219784,
0.0420725495,
-0.0298567303,
-0.0089906249,
-0.086456418,
-0.0087751849,
-0.0099087898,
0.0282988008,
0.0000017088,
-0.0675236434,
-0.0470317192,
0.0237464793,
-0.0593634509,
0.0408036485,
0.0678239986,
0.0112963635,
-0.0339577794,
0.0552645437,
0.0594402552,
-0.0010093979,
0.0875020176,
0.0191687178,
-0.027322175,
0.0497546792,
-0.0169131141,
-0.0392940491,
0.031991221,
0.023962304,
-0.0490384884,
0.0520334579,
-0.0246146936,
-0.0554343499,
0.0198015906,
-0.0341038965,
-0.0211169794,
-0.0125511903,
-0.0484460331,
-0.0002257916,
0.0014248624,
0.0019614703,
-0.0236893594,
-0.0364052951,
0.0051013194,
0.0385515131,
-0.0720901936,
-0.0000886711,
0.01239138,
-0.0099709285,
-0.014526546,
-0.0121012311,
-0.0015786717,
0.0139983855,
-0.0573614836,
0.0366271175,
0.039455682,
0.023970956,
-0.0176557172,
-0.0320078135,
-0.012136355,
-0.002715257,
-0.0021045709,
-0.0311854053,
-0.0297641531,
-0.0698814765,
-0.0426938571,
0.016415067,
0.0586701147,
-0.0077323769,
-0.00804675,
0.0262163114,
0.010857299,
-0.0086545246,
0.0213978551,
-0.0084505687,
0.0416265912,
0.0533100553,
-0.0240690727,
-0.005577839,
-0.0392455347,
0.0414454602,
0.0052020052,
0.0167156551,
-0.0376060158,
-0.000494907,
0.0039102696,
0.0018385443,
-0.0414429195,
0.026264118,
-0.0246050023,
-0.0202043504,
0.0096731829,
-0.0097178519,
0.0303403791,
-0.0187157299,
0.0417264625,
-0.0221141484,
-0.0132478457,
-0.0368355475,
0.0151539203,
-0.0032016176,
-0.0256900508,
-0.0046356572,
-0.0047950251,
-0.101567395,
-0.0135570494,
0.0163020771,
0.0278704502,
0.0176309012,
0.013285446,
0.0781630725,
0.0059318379,
-0.0433049463,
0.0047671325,
0.0189423114,
-0.0159988925,
-0.044628907,
-0.0256658513,
-0.0274124555,
0.0345876925,
0.0349331498,
-0.0187659804,
0.0240033548,
0.0152366674,
0.0352140144,
0.018855121,
0.0475640558,
-0.0257126316,
0.0128431898,
0.0652069449,
-0.0037521964,
0.0590558127,
0.0362005197,
0.0099905897,
-0.0406699665,
-0.0096651921,
-0.028040709,
-0.000141056,
0.0576563813,
-0.0334133357,
-0.0109429723,
-0.023316104,
-0.0110420883,
0.0724614039,
0.0293791667,
0.0550909489,
0.082492657,
-0.0515847467,
0.0331004523,
-0.0182258617,
0.06136005,
0.035085313,
0.0295743365,
0.000703515,
0.008200407,
0.027008649,
0.0168901011,
0.0343290269,
-0.0051419539,
0.0484848469,
-0.0007640639,
0.0512063056,
0.0294775125,
-0.0076673864,
-0.0013454542,
-0.0197615027,
-0.047390908,
-0.001634879,
-0.0346409865,
0.0566684753,
0.0003251162,
-0.0703079551,
0.0276502948,
0.0613006428,
-0.0476730131,
0.019637078,
-0.0776940733,
-0.0186663363,
0.0085157072,
0.0678796247,
-0.0060425629,
0.0302869398,
0.0218983963,
-0.0705557615,
0.0278783552,
-0.0195015259,
0.0083879987,
0.0271693524,
-0.0905372947,
0.0319581963,
0.0643859878,
-0.0191389825,
0.0542788617,
-0.0396413505,
0.0207570344,
0.0014452938,
-0.0298042987,
-0.0166025236,
0.0106657455,
0.0444854833,
-0.0028728293,
0.0099697346,
-0.0163891427,
0.0686588734,
0.0208687428,
-0.0179953873,
0.0029213098,
-0.0089034792,
-0.031308677,
0.0438095294,
0.0108007137,
-0.0494432226,
-0.0220446493,
0.0338439494,
0.0127142807,
0.0274255909,
-0.0459680483,
0.0807607993,
-0.080312334,
0.0401577204,
-0.063027367,
0.0747704059,
-0.0349041224,
0.0599286184,
-0.0233267881,
0.0132363774,
0.0136254849,
0.0217870269,
0.0155226104,
-0.0027431166,
-0.0051588994,
-0.0124046775,
0.0507672913,
-0.0409233943,
-0.000105203,
0.0682240129,
0.0045353849,
-0.0161753539,
-0.001285564,
-0.0103784846,
-0.0215663705,
-0.0193178542,
-0.1170203462,
-0.0441713668,
-0.0037813261,
0.0360688716,
-0.0249886643,
0.0290075447,
-0.0223907772,
-0.0300983805,
-0.0012736566,
0.0149450898,
-0.0063572046,
-0.0329688005,
-0.0193760227,
-0.014241376,
-0.0432405286,
0.0365193449,
-0.0261339769,
0.0143455025,
0.0351827107,
-0.0031738766,
-0.0593086816,
-0.0670975521,
-0.0090061305,
0.0218397025,
-0.0031380649,
-0.0078411847,
-0.0459877588,
0.00108814,
-0.0314297676,
0.0601780117,
-0.0154675357,
0.0019497044,
-0.0679187551,
0.015845323,
-0.0122625642,
-0.0190912224,
-0.0348380543,
0.0407487564,
0.022169264,
0.0360023938,
-0.016867632,
-0.017514633,
-0.0141532645,
0.0041967984,
0.0566486642,
0.0143695492,
-0.0717222691,
-0.0125055118,
-0.0049811928,
-0.0029766134,
-0.0194772296,
-0.0713615641,
0.0338444114,
-0.0097467229,
-0.0376232527,
0.0170787256,
-0.072461158,
-0.0358518474,
-0.0704386085,
0.0479944423,
-0.0042246641,
-0.0099493787,
-0.007301766,
0.0214345437,
-0.0348602459,
-0.1081533656,
-0.0121431863,
0.0029582004,
0.0146104125,
-0.0569949113,
-0.0203012265,
-0.037283279,
0.0064692497,
-0.0401645787,
0.017136177,
0.0176055599,
0.007200229,
0.0152339526,
0.0426136293,
0.0181485824,
-0.0190621149,
0.0223241951,
0.0921895057,
0.0045661782,
-0.0004508125,
-0.0194242056,
0.0219894815,
0.0121239219,
0.0621395148,
-0.0034774784,
0.0202300604,
-0.003794289,
0.0051723807,
-0.0182647184,
-0.050374642,
0.017350547,
-0.01442827,
0.0241495129,
-0.0121763805,
0.0227112584,
-0.0055079646,
-0.0108426185,
0.0237058308,
0.0597049706,
0.048628062,
-0.0302457027,
0.0529244766,
0.0276516583,
-0.0050877994,
-0.045536384,
-0.0462593138,
0.0082800547,
-0.0195740294,
-0.0310171843,
-0.0005713518,
-0.0099140899,
0.0511664599,
0.0176744238,
0.02263809,
0.0132457744,
-0.0073663942,
-0.1056581587,
-0.0024552832,
0.0309912134,
-0.0118947569,
-0.0596163422,
-0.0323858783,
-0.0163325034,
0.0122735044,
0.04447034,
0.0618256032,
-0.0151912626,
-0.0373240039,
0.0806224048,
-0.0481835306,
-0.0012464458,
-0.0197485685,
-0.0435004756,
0.037624009,
0.027322337,
-0.0158244893,
-0.0373284444,
0.0557414331,
0.0180333164,
-0.0628887191,
0.0168053936,
-0.0170010701,
0.0097704055,
0.0277264286,
-0.0126740178,
0.0041222041,
-0.0014108574,
0.0080232471,
0.0065371813,
-0.0160672348,
0.0315094329,
0.0218754355,
0.0162538197,
0.0031647882,
-0.0116395028,
-0.0186632201,
-0.0015740353,
0.024600653,
-0.0347338803,
-0.034520667,
-0.0125767328,
0.0045065302,
0.0040412107,
0.0410669222,
0.0000893647,
-0.0236432217,
-0.0595807061,
-0.0359521769,
-0.0019303623,
-0.0127997799,
0.0740743056,
0.0093660057,
0.0532065444,
-0.0158932339,
0.0693460405,
-0.0327492766,
-0.0499903299,
-0.0007706436,
0.0102203432,
-0.0871808082,
0.0096130837,
-0.0088919858,
-0.0087687811,
-0.0081287716,
0.0054017697,
0.0372131616,
-0.0221980009,
0.0798286796,
-0.0391746201,
0.020133879,
-0.0019618731,
-0.0160331652,
-0.0350868665,
0.1031018496,
0.0474229194,
-0.0310882218,
-0.0133638652,
0.0379395597,
0.0221965704,
-0.0101419277,
-0.0015860517,
0.0255282614,
-0.0281965937,
-0.0348314941,
-0.0186034516,
-0.0121167414,
0.0029261489,
0.0262500048,
0.0295035634,
0.0081473319,
-0.0663360208,
-0.0044553471,
-0.0107071856,
-0.0162347369,
0.0065411134,
0.0318105817,
0.0096598146,
-0.0192259718,
0.0414718203,
0.0270659756,
-0.011224662,
0.0078412658,
-0.0542556532,
-0.0257933903,
0.1127372757,
0.0948405489,
0.0136982733,
0.0407630652,
0.0099600833,
-0.0279433522,
-0.0118739633,
-0.0326596908,
0.0000905583,
-0.019105712,
-0.0562644154,
0.038636867,
-0.0388522297,
0.0131977312,
0.0075136982,
0.0524321161,
0.0646648034,
0.0776062012,
-0.0207690783,
0.0504731946,
-0.0016824604,
0.0122872358,
0.0215188731,
-0.0259800758,
0.0298844967,
-0.0285945665,
0.0059212595,
-0.0851475075,
0.0505988225,
0.0142728388,
0.0010720768,
-0.0318204612,
-0.0088056419,
0.0393217392,
-0.0103119444,
-0.0343070105,
-0.0018296978,
0.0352583118,
-0.0020644458,
-0.0296414997,
-0.0423317067,
0.0318699367,
-0.0333881378,
-0.0054786238,
-0.0200278722,
-0.0360805206,
0.0081548477,
0.0063061365,
0.0470249653,
0.0005590302,
0.0336305611,
0.0182868335,
0.0050732554,
0.0503241606,
0.0021703194,
-0.013293446,
0.0043742359,
0.0016450338,
0.0224144179,
-5.211995749e-33,
0.0049154339,
0.0129214674,
-0.0332468413,
-0.0510312989,
0.0133684939,
0.0079913549,
0.0097620254,
0.0040779025,
0.0010847729,
-0.0593701825,
-0.046658691,
-0.0165438689,
0.0313269496,
-0.0346248746,
0.0349590704,
0.0119014718,
0.024822345,
0.0018713698,
0.0016657008,
-0.0211473666,
0.0535713322,
0.0251337588,
0.0148605648,
0.0153962197,
-0.0296851397,
-0.028238995,
-0.0202058088,
-0.0395603441,
0.0144586386,
0.0334573835,
-0.0176995173,
0.0518490598,
0.0003247476,
0.0158526599,
-0.0075661736,
0.0356782824,
-0.0986015946,
0.0149816032,
0.018693924,
0.0690894872,
0.0245659947,
0.0000415608,
0.0088490434,
-0.0033543373,
-0.0737006739,
0.0173249524,
-0.0463064313,
-0.0153385373,
-0.0532523245,
0.020355694,
-0.0290129706,
0.0538388565,
0.0205679238,
0.0325824134,
0.0721718147,
-0.0187043566,
-0.0076555349,
-0.0953007936,
-0.0029974913,
0.0188633297,
0.0336079337,
0.0542640425,
-0.0099223601,
0.0322122052,
-0.0437743887,
-0.0312067531,
0.0529343635,
-0.1210765764,
0.0474863648,
-0.0178871229,
0.0487332307,
-0.0413790643,
-0.0014951862,
0.0381229259,
0.1133311614,
-0.0858574659,
-0.0428337343,
-0.0075654634,
0.0674305111,
0.0213056728,
0.0088353259,
0.0230760574,
-0.0608776249,
-0.0330394581,
0.0262690932,
-0.0261084531,
-0.0517701767,
-0.0365023129,
0.0400007553,
-0.0255130306,
0.0048912643,
-0.0001636755,
0.0570292473,
0.0231827833,
0.0256500766,
-0.0217369143,
-0.0024336982,
0.0491280034,
-0.0078531485,
0.0320193805,
-0.0514370501,
-0.0982809737,
0.0330773965,
0.0499698073,
0.0001996279,
-0.0155721111,
0.013992372,
0.0128611336,
-0.0747674853,
0.0008763206,
0.0037287921,
0.0055044154,
-0.0077820527,
0.0166616626,
0.0095090242,
-0.0419012308,
0.0342144705,
-0.0315981582,
0.004464352,
0.0320704468,
0.0022034578,
0.0371107981,
-0.0484138131,
0.0060623176,
-0.0297412258,
0.0497575067,
-0.0131309312,
-0.0133444592,
-0.002747786,
0.0277073812,
-0.0123613132,
0.0581685901,
2.531e-7,
-0.0262782369,
0.0652421266,
0.0643733889,
-0.05645293,
0.0157471094,
-0.0241468288,
-0.0244816057,
-0.0195902102,
-0.0275940597,
0.0182252787,
0.0432607271,
-0.0356100984,
0.0236695204,
0.0156956334,
-0.0974991247,
-0.0166541208,
-0.0157781839,
0.0081611779,
-0.0490138978,
0.027363386,
0.0003102917,
0.0536650084,
-0.0428605042,
-0.0359360091,
-0.0183955096,
-0.0304303579,
0.0152858272,
-0.0198470205,
0.0175546743,
-0.0271229222,
-0.0275402069,
0.0121194785,
-0.0063435617,
0.0198135003,
-0.0106552737,
0.025291305,
0.0697884113,
-0.0122917462,
-0.0261746831,
-0.0397010744,
0.0575417317,
0.0439784192,
-0.0343760066,
-0.0212253705,
0.0518617481,
0.0519937649,
-0.0436671861,
-0.0298591144,
-0.0395498835,
0.0138072753,
-0.0137259783,
0.002155262,
-0.0143105658,
0.0372745804,
0.0294720624,
-0.0384466499,
-0.0178658534,
-0.0050959983,
0.0203725696,
0.0326857418,
-0.0173150878,
-0.0585686415,
-0.0175800007,
0.0664985105,
0.0534906499,
-0.0449241363,
0.0008429708,
2.151175325e-34,
0.0385491848,
-0.0266363621,
0.0473006479,
-0.021219993,
-0.0009694097,
-0.023426393,
-0.0073190047,
0.0435332842,
-0.0169130452,
-0.0276045892,
-0.0548879951
] |
http://arxiv.org/abs/2402.01728v1 | http://arxiv.org/pdf/2402.01728v1 | Hardware Phi-1.5B: A Large Language Model Encodes Hardware Domain Specific Knowledge | 2024-01-27T22:49:43 | 2024-01-27T22:49:43 | Weimin Fu, Shijie Li, Yifang Zhao, Haocheng Ma, Raj Dutta, Xuan Zhang, Kaichen Yang, Yier Jin, Xiaolong Guo | cs.CL | cs.CL, cs.AI, cs.AR | In the rapidly evolving semiconductor industry, where research, design,
verification, and manufacturing are intricately linked, the potential of Large
Language Models to revolutionize hardware design and security verification is
immense. The primary challenge, however, lies in the complexity of hardware
specific issues that are not adequately addressed by the natural language or
software code knowledge typically acquired during the pretraining stage.
Additionally, the scarcity of datasets specific to the hardware domain poses a
significant hurdle in developing a foundational model. Addressing these
challenges, this paper introduces Hardware Phi 1.5B, an innovative large
language model specifically tailored for the hardware domain of the
semiconductor industry. We have developed a specialized, tiered dataset
comprising small, medium, and large subsets and focused our efforts on
pretraining using the medium dataset. This approach harnesses the compact yet
efficient architecture of the Phi 1.5B model. The creation of this first
pretrained, hardware domain specific large language model marks a significant
advancement, offering improved performance in hardware design and verification
tasks and illustrating a promising path forward for AI applications in the
semiconductor sector. | Hardware Phi-1.5B: A Large Language Model Encodes Hardware Domain Specific Knowledge: In the rapidly evolving semiconductor industry, where research, design,
verification, and manufacturing are intricately linked, the potential of Large
Language Models to revolutionize hardware design and security verification is
immense. The primary challenge, however, lies in the complexity of hardware
specific issues that are not adequately addressed by the natural language or
software code knowledge typically acquired during the pretraining stage.
Additionally, the scarcity of datasets specific to the hardware domain poses a
significant hurdle in developing a foundational model. Addressing these
challenges, this paper introduces Hardware Phi 1.5B, an innovative large
language model specifically tailored for the hardware domain of the
semiconductor industry. We have developed a specialized, tiered dataset
comprising small, medium, and large subsets and focused our efforts on
pretraining using the medium dataset. This approach harnesses the compact yet
efficient architecture of the Phi 1.5B model. The creation of this first
pretrained, hardware domain specific large language model marks a significant
advancement, offering improved performance in hardware design and verification
tasks and illustrating a promising path forward for AI applications in the
semiconductor sector. | [
-0.0058612726,
0.0540624671,
-0.014462458,
0.0466583259,
-0.0356305242,
-0.0129980929,
0.0749057382,
-0.0305107366,
0.0299813189,
-0.0229768995,
-0.0129530635,
0.0240419861,
0.005737796,
0.0608809702,
0.0516820811,
0.0402980447,
0.069408156,
-0.045224607,
-0.0081949839,
0.0059360624,
-0.016508935,
-0.006354094,
0.0252646264,
0.0360605046,
-0.0094216522,
-0.0290095489,
-0.0376502499,
0.0087374765,
0.0227182806,
0.0029664387,
0.0054868972,
0.0478660576,
0.0385807194,
0.0041035796,
0.000001979,
-0.00394578,
-0.0248130299,
0.0197980162,
-0.0411334857,
0.0306891389,
-0.0162120778,
-0.0315144062,
-0.0305903312,
0.0289602298,
-0.000645142,
-0.0608055815,
0.0618212521,
0.0658390895,
0.0513609201,
0.0749894753,
0.0108443024,
-0.0338687114,
0.05007324,
-0.0308328383,
-0.0395428389,
-0.0342318863,
0.0348255932,
0.0027561767,
0.0441420712,
0.0294728186,
-0.0065120081,
-0.0131801348,
0.0448821187,
0.0231051706,
0.0195003822,
0.0130719943,
-0.0201460905,
0.0544840917,
-0.0242494829,
0.0604721308,
-0.0568478182,
-0.0261874422,
-0.0120174428,
-0.0160898976,
-0.0417736843,
-0.028101908,
0.038969364,
0.0010610801,
-0.0318934843,
-0.0166233871,
0.011330449,
-0.006452166,
0.0530168228,
-0.0183853656,
-0.0212893989,
0.0698624924,
0.0120898588,
-0.0176546704,
-0.0205880105,
-0.0609815493,
0.0020815213,
-0.0589112528,
-0.0092494851,
0.0347459912,
-0.0168978088,
0.0065732682,
-0.0556960553,
0.0395890921,
-0.0022557182,
-0.0679095015,
-0.0074031348,
0.007905907,
-0.0060116644,
0.0386784449,
-0.0417459123,
0.0347465537,
-0.018075319,
0.0066246488,
-0.0528324321,
0.0183711685,
-0.0274329688,
-0.0527032986,
-0.0337593183,
0.0475295968,
0.0293960832,
-0.0383657031,
-0.040141128,
-0.0016107137,
-0.0359207839,
0.0488509685,
-0.004922688,
-0.0349853672,
0.0116334474,
-0.0409168676,
0.015291987,
-0.0091469754,
-0.0240037087,
-0.0039614527,
-0.022106111,
-0.0544751436,
-0.0115838302,
-0.0028898495,
-0.0104757585,
0.0022654836,
-0.0015422056,
0.0742881224,
-0.0024451793,
-0.0568583012,
-0.0763146579,
0.0025814758,
-0.0031730549,
-0.0627355874,
0.0126026031,
-0.0407508872,
0.029456811,
-0.0243026856,
0.0349168889,
-0.0387093835,
-0.0069988156,
0.0391110368,
0.0050280867,
0.0808799267,
0.0146649173,
-0.0251611359,
-0.0056856363,
-0.0016109303,
0.0347787961,
0.0166014079,
0.040252097,
0.0315538272,
0.0194489472,
-0.0234409403,
-0.0315416269,
0.036518205,
0.0036911829,
-0.0372985415,
-0.0316253975,
0.0287242215,
0.0216709506,
-0.0467785522,
-0.0505138151,
0.0303847212,
0.0240520537,
-0.0208513513,
0.0107197203,
0.0310619548,
0.0261087939,
0.0270741731,
0.0315199383,
0.0445882529,
0.0034913081,
-0.0541429929,
-0.019966213,
0.039704565,
-0.0433680117,
-0.0012833747,
0.0210980605,
-0.0009456325,
-0.038337741,
-0.0213695411,
-0.0290333759,
-0.0550096743,
-0.0243264809,
0.0166110061,
0.066077657,
-0.019126825,
-0.033656586,
0.0020551903,
0.0323561095,
-0.0291801039,
-0.0425867364,
-0.0634965673,
0.0434247553,
0.0332107507,
0.0370223708,
-0.0527611896,
-0.0247983169,
0.023057064,
0.0309704021,
0.0098502506,
0.0072680563,
-0.0080925757,
-0.0070647569,
-0.0310075562,
-0.0207163524,
-0.0166310407,
0.0245687049,
0.0151001336,
-0.0072962465,
-0.0361489169,
0.0118416529,
0.0368508138,
-0.0303075984,
-0.0097688716,
-0.0370048732,
-0.0108724926,
0.0246046782,
0.0457495376,
0.0046052588,
-0.0199975427,
-0.0460799597,
0.0224595107,
0.0224013533,
-0.0317770094,
0.0295418743,
-0.0434462577,
-0.025221061,
0.0183722954,
0.0338740684,
-0.0405089632,
0.0138382623,
0.0091917478,
-0.0538506508,
-0.0440496542,
0.0534531772,
-0.044734098,
0.0017068445,
-0.0189207848,
0.0026350659,
-0.0031457825,
0.0094784033,
-0.0154940449,
0.0621172301,
0.029027462,
-0.0201089047,
-0.0253069066,
0.0114532746,
0.0302317198,
-0.0220054761,
-0.0367726125,
-0.0161291659,
-0.0196766686,
0.0346522406,
0.0366463847,
-0.0171839222,
0.004945274,
0.0145557411,
-0.0412484743,
-0.0041014263,
0.0002666599,
0.0123800952,
-0.0171289984,
-0.0096017281,
0.0252153203,
-0.0238470864,
-0.0049874182,
0.0063430909,
-0.0140486797,
-0.034871567,
-0.0122118909,
0.0256770011,
-0.0651583672,
0.0025930451,
-0.0284973457,
-0.012224338,
0.0436447263,
0.002100975,
0.0390800647,
-0.0700023398,
0.0619561709,
0.0122158593,
0.0173374712,
-0.0434145443,
-0.0320153795,
0.0076153828,
-0.0076439744,
-0.0111384429,
0.0773417205,
0.0063681854,
0.0101877544,
0.0609125197,
-0.0259912573,
-0.0148411319,
0.0188508853,
0.0399771743,
-0.0102314036,
-0.0238346756,
0.0245325603,
-0.0141300242,
0.0614611246,
0.0029145817,
-0.0067741093,
-0.0442596115,
0.0214220807,
-0.0101586003,
-0.0133392187,
-0.0049189944,
-0.015726665,
0.0285083279,
0.0411504395,
-0.0466332436,
-0.0446968861,
0.0736892521,
-0.0302227233,
0.0154864239,
-0.0142721711,
-0.0360945426,
-0.0228470005,
0.0488011688,
0.0041846717,
0.0328768641,
0.055003453,
0.0032006002,
-0.0297535267,
-0.010090814,
-0.003238932,
-0.0128873512,
0.0273267254,
-0.0406670235,
0.0555044971,
-0.0326364934,
-0.036243245,
-0.0310626086,
0.0117624719,
-0.0032680752,
0.0160849802,
-0.0029354896,
-0.0511790849,
-0.0295755714,
0.0514054298,
0.0164011568,
0.0037689256,
0.0375695601,
-0.041660428,
-0.0283340551,
-0.0220480897,
0.0372984149,
-0.0120250154,
0.0090668425,
-0.0091125332,
0.0068623368,
0.0099048112,
0.032481961,
0.0249828435,
-0.1343643218,
0.0109576583,
0.0106208678,
-0.0506058261,
0.0077610221,
0.0351232886,
-0.0198696684,
-0.0243381187,
-0.0369305201,
0.0858204141,
-0.0225781109,
-0.0024166068,
0.0174745601,
0.0741340145,
0.0117057376,
0.0173384324,
0.0671544373,
-0.1037547365,
-0.0107542733,
0.0011157922,
0.0069838795,
-0.0215305611,
-0.0067810663,
-0.0371468067,
-0.1223357692,
0.0570549071,
0.0266719759,
-0.0300631113,
-0.0873780251,
-0.0717903078,
-0.0166903436,
0.0019696122,
0.0292255897,
-0.0109844021,
-0.0137022743,
-0.0518547557,
0.0002789366,
-0.0597618297,
-0.0899744704,
0.0655137524,
-0.0016454866,
0.0251433291,
0.0531581938,
-0.0204784591,
-0.009683921,
0.0139220282,
0.0216190834,
0.0231431555,
0.0229451023,
-0.0086111091,
-0.0021321292,
0.0046637179,
-0.0428990833,
-0.0046134903,
0.0185368098,
-0.0513344817,
-0.0128791826,
0.0060351193,
0.0636432692,
-0.0330883525,
0.005080028,
-0.0001050509,
-0.039161101,
0.0246964954,
0.0044454746,
-0.0084738806,
-0.0083091538,
-0.0468636751,
-0.0235371441,
-0.0222784244,
-0.0342821851,
-0.0057687461,
0.0114284875,
-0.0000773133,
-0.0126137119,
0.0016851255,
0.0062907883,
-0.0114390738,
0.0250588488,
0.045486901,
-0.0628188923,
0.0908061415,
0.0062862351,
-0.0874131098,
0.0216634069,
-0.0443758741,
0.0337801054,
-0.0452593006,
-0.0027997594,
0.0203377828,
-0.0147415847,
-0.0602760538,
0.0326770134,
-0.0208638981,
-0.0757246614,
0.0049337992,
0.095239915,
0.0198437106,
-0.0179016385,
-0.0269575324,
-0.047505375,
0.0343213975,
0.0281784143,
-0.0476068147,
0.0128519917,
0.0246859677,
-0.0578420945,
-0.0252502766,
-0.0095439721,
0.0034393566,
0.0502177551,
0.0993260071,
-0.0086928224,
0.0176341869,
-0.0212450717,
0.0281558298,
0.0143378517,
0.0123445159,
0.0406594984,
0.0461683124,
-0.040437676,
0.0287554339,
0.0076231607,
-0.0037488162,
-0.0385819189,
-0.0290517733,
0.0252255872,
-0.0693569183,
0.0088228164,
-0.0625020936,
0.0387336835,
0.0245490633,
-0.0256344192,
0.0429760441,
0.0278974976,
0.0752609223,
0.0097549949,
-0.0210310202,
-0.0003315563,
0.0228616334,
-0.0020281421,
0.0545883365,
-0.0314862132,
0.0203436241,
0.0318391323,
0.0295960736,
-0.062338151,
0.0366466455,
-0.0240297876,
-0.100372076,
-0.0099368077,
-0.0052491026,
0.0549454577,
0.0126945525,
-0.0279639568,
-0.0472912528,
0.0098153027,
-0.0740634874,
0.0699629039,
0.0049808165,
-0.0339031592,
0.0155509077,
0.0431657769,
-0.0111520374,
-0.0117098466,
0.0459930934,
0.0196824558,
0.0475830473,
0.0004068534,
0.0004652046,
-0.0749701932,
-0.029589491,
0.0313254669,
0.0023540168,
-0.0342487432,
0.0009353959,
-0.0287924334,
-0.0660472885,
-0.0089714779,
-0.0008883419,
0.0655488968,
0.0421252511,
0.0325905494,
0.0591119714,
0.0095612546,
-0.0265560858,
-0.0010390083,
-0.0390795358,
-0.0225283038,
-0.0147898383,
0.0064759878,
-5.320883995e-33,
-0.0343152918,
-0.0045199771,
0.0426727049,
-0.0087089175,
0.0348139368,
0.0466823727,
-0.0163992923,
-0.0044609578,
-0.0166594423,
-0.0303331614,
-0.0033095633,
0.0060521397,
0.0241760463,
0.0427237339,
0.0107876593,
-0.0121749649,
0.0638600066,
0.0073560853,
-0.0196092799,
0.0537808724,
0.0216523856,
-0.0053066928,
0.0349868089,
-0.0473677516,
0.012824419,
-0.0008426362,
-0.0288373157,
-0.0203161631,
-0.0307966061,
0.0641968027,
-0.0123139564,
0.0509325042,
0.0062440326,
-0.0157094207,
0.0108712176,
0.0604665242,
-0.0673346967,
-0.0248992518,
-0.0196543466,
0.0162000433,
0.047224991,
-0.0547922887,
0.0225743297,
0.0061139632,
-0.0423598364,
-0.0094566038,
0.0403204113,
-0.0108682141,
0.0008567215,
-0.0061907629,
-0.0460664444,
0.0299685411,
-0.0066382592,
0.0006831212,
0.0154613648,
0.006158642,
-0.0392763764,
-0.020789329,
-0.0459564179,
0.0410074368,
0.0328666717,
0.05842584,
0.0827509537,
0.0105893454,
0.0150542231,
0.0486439243,
0.0436892807,
0.0247444846,
0.0276229586,
0.0219072532,
0.0270641465,
0.0726962835,
-0.0228215978,
0.0300254077,
0.0296914093,
-0.0464881584,
-0.0960446373,
0.0160779115,
-0.014132713,
0.0684974045,
0.0334059298,
0.0128772538,
-0.05917136,
-0.0019616396,
0.0123486724,
0.0553562306,
-0.0056759114,
-0.0224973466,
0.0107610896,
0.0517833829,
-0.0614113733,
0.0195113663,
0.0503043197,
-0.0794524774,
-0.0012610717,
0.0274639763,
0.0656828061,
-0.0075906082,
-0.0228058826,
0.0376493521,
-0.0445459113,
-0.0897976458,
-0.0169237684,
-0.0387753658,
0.0322082601,
-0.0136246094,
-0.0157640371,
0.0896070302,
-0.0573094971,
-0.0125416471,
-0.0220960826,
-0.0138549283,
-0.0178389978,
0.075938344,
0.0117577557,
0.0411419533,
0.0040463922,
0.0404022597,
-0.0568098314,
-0.082914196,
0.0083759343,
0.0478634387,
-0.0056626513,
0.0189716388,
-0.0164329838,
-0.0120694377,
-0.0605897233,
0.039514754,
0.0221547652,
-0.0372263193,
-0.005896531,
0.0169091653,
2.808e-7,
-0.017002007,
0.0316755734,
0.0322712958,
0.0814914852,
-0.019082047,
0.0245210454,
-0.001160557,
0.026974909,
-0.024683293,
-0.012527857,
0.005772593,
0.042946279,
0.0371407419,
-0.0342599563,
-0.0209274273,
0.0070042945,
0.0084894486,
0.0333890431,
-0.0606439076,
0.0043260441,
0.0523802191,
0.1042005494,
-0.0013099381,
0.0009853999,
0.0241764951,
-0.0561225414,
0.0148829482,
-0.0110297166,
0.0057279929,
0.0386430323,
-0.00866991,
0.0220463797,
-0.0509664491,
0.1119242385,
-0.0019148221,
0.0230487883,
-0.0450933874,
0.065237686,
0.0127546135,
-0.0062664011,
-0.0149489176,
-0.0070442692,
-0.0322540365,
-0.0587483793,
0.0372694805,
-0.0322911665,
-0.0466157943,
-0.0675888136,
0.0033698713,
-0.0295366533,
0.0000746325,
-0.0140225561,
0.0493675508,
-0.014944681,
0.0150854997,
-0.0348643325,
0.033762522,
-0.0552278459,
-0.0079824207,
0.0348766036,
-0.0390200205,
-0.0177576505,
0.0035424158,
0.0047355923,
0.0524114855,
-0.0120143276,
0.0170496367,
2.399672994e-34,
0.0035976502,
-0.0332294591,
-0.0092523452,
0.0587562732,
0.0627628639,
-0.0192935895,
-0.0457307436,
-0.0108758649,
-0.0158208609,
-0.0586655103,
-0.0259710141
] |
http://arxiv.org/abs/2401.15509v1 | http://arxiv.org/pdf/2401.15509v1 | Style-News: Incorporating Stylized News Generation and Adversarial Verification for Neural Fake News Detection | 2024-01-27T21:35:29 | 2024-01-27T21:35:29 | Wei-Yao Wang, Yu-Chieh Chang, Wen-Chih Peng | cs.CL | cs.CL, cs.AI, cs.SI | With the improvements in generative models, the issues of producing
hallucinations in various domains (e.g., law, writing) have been brought to
people's attention due to concerns about misinformation. In this paper, we
focus on neural fake news, which refers to content generated by neural networks
aiming to mimic the style of real news to deceive people. To prevent harmful
disinformation spreading fallaciously from malicious social media (e.g.,
content farms), we propose a novel verification framework, Style-News, using
publisher metadata to imply a publisher's template with the corresponding text
types, political stance, and credibility. Based on threat modeling aspects, a
style-aware neural news generator is introduced as an adversary for generating
news content conditioning for a specific publisher, and style and source
discriminators are trained to defend against this attack by identifying which
publisher the style corresponds with, and discriminating whether the source of
the given news is human-written or machine-generated. To evaluate the quality
of the generated content, we integrate various dimensional metrics (language
fluency, content preservation, and style adherence) and demonstrate that
Style-News significantly outperforms the previous approaches by a margin of
0.35 for fluency, 15.24 for content, and 0.38 for style at most. Moreover, our
discriminative model outperforms state-of-the-art baselines in terms of
publisher prediction (up to 4.64%) and neural fake news detection (+6.94%
$\sim$ 31.72%). | Style-News: Incorporating Stylized News Generation and Adversarial Verification for Neural Fake News Detection: With the improvements in generative models, the issues of producing
hallucinations in various domains (e.g., law, writing) have been brought to
people's attention due to concerns about misinformation. In this paper, we
focus on neural fake news, which refers to content generated by neural networks
aiming to mimic the style of real news to deceive people. To prevent harmful
disinformation spreading fallaciously from malicious social media (e.g.,
content farms), we propose a novel verification framework, Style-News, using
publisher metadata to imply a publisher's template with the corresponding text
types, political stance, and credibility. Based on threat modeling aspects, a
style-aware neural news generator is introduced as an adversary for generating
news content conditioning for a specific publisher, and style and source
discriminators are trained to defend against this attack by identifying which
publisher the style corresponds with, and discriminating whether the source of
the given news is human-written or machine-generated. To evaluate the quality
of the generated content, we integrate various dimensional metrics (language
fluency, content preservation, and style adherence) and demonstrate that
Style-News significantly outperforms the previous approaches by a margin of
0.35 for fluency, 15.24 for content, and 0.38 for style at most. Moreover, our
discriminative model outperforms state-of-the-art baselines in terms of
publisher prediction (up to 4.64%) and neural fake news detection (+6.94%
$\sim$ 31.72%). | [
0.0454157591,
0.0435275324,
0.0006084389,
0.0349351205,
-0.027322663,
0.0110532297,
0.0032025073,
0.0168405101,
0.0019518759,
0.0100490227,
-0.0051775356,
-0.027485501,
0.0140847238,
0.0312404688,
0.045016665,
-0.002826428,
0.0318654813,
0.0511175096,
0.0600667372,
-0.0255339071,
0.0555910133,
0.0449905172,
0.007529363,
0.0037925718,
-0.0288278721,
-0.0247103088,
0.0193293504,
0.0489503145,
0.0337896124,
-0.006332383,
0.0296669882,
0.0546363033,
-0.0034284408,
0.0957141742,
0.0000016954,
-0.0021818411,
0.0010723844,
-0.0122818686,
0.011303314,
0.0272692163,
0.0681592971,
-0.06700591,
-0.0349355601,
-0.0061914218,
-0.0302659404,
-0.0343313254,
0.0120381592,
0.0916824713,
0.0953700244,
0.1205894724,
-0.0279014669,
-0.0296682063,
0.0460036956,
-0.0124635911,
0.0065507451,
-0.0045729047,
0.0240887105,
-0.018552117,
-0.0175874922,
0.0065813172,
0.0389831439,
0.0177519843,
0.0323295593,
-0.0107125593,
-0.0574282631,
0.0127274822,
-0.0159535371,
-0.000853051,
0.0065378491,
0.0371550322,
-0.0313095637,
-0.0685504749,
0.0370417237,
-0.0039749076,
0.0265632346,
0.0414345227,
0.0020312297,
0.0171019696,
-0.0336787328,
-0.0003647358,
-0.0059796269,
-0.0191857889,
0.0462349057,
0.0315823331,
0.0369332358,
0.001444511,
0.0172818005,
-0.0319421515,
-0.0145649184,
-0.0290754214,
0.0202708915,
-0.0498357266,
0.028964486,
0.0041153715,
-0.0033081698,
-0.0302864816,
-0.0854599997,
-0.063654162,
0.007167696,
-0.1008739024,
-0.0138484351,
-0.0169506632,
0.0099139316,
0.0340354927,
-0.0332237408,
0.0054933955,
-0.0374667607,
0.0539410077,
0.0138341328,
0.0063178078,
0.0140748508,
0.0013446368,
-0.0111417305,
0.0043358253,
0.0239432082,
-0.0209758878,
-0.0252639819,
-0.0043605925,
-0.0134945111,
0.0599454381,
-0.019459784,
0.039393384,
-0.0109515563,
-0.032726083,
0.0206753034,
-0.0132564008,
-0.0099569596,
-0.0021715891,
0.0106846532,
-0.0250310227,
-0.0033096655,
0.0304023996,
-0.0255748704,
0.0125010284,
0.0066293115,
0.0289549176,
-0.0133407367,
0.0260284916,
-0.0713002309,
-0.0270877928,
-0.0035598781,
-0.05316028,
0.0463275611,
-0.0573070571,
-0.0375275463,
-0.0511971228,
-0.0103830695,
-0.079222478,
-0.024237765,
0.0320179313,
-0.0061753802,
0.028008515,
0.0065386277,
-0.0274869464,
0.0660707206,
0.0404208526,
0.0147543466,
0.0809119269,
0.0332880542,
0.1098653078,
0.0375208519,
0.0173335895,
0.034000054,
0.0354298875,
0.0485418178,
0.0192317739,
-0.0232738685,
-0.0455977172,
0.0006560204,
0.0103182578,
0.0144481258,
0.0340446979,
-0.0050884676,
-0.0867924541,
0.0857351944,
-0.0611698218,
0.0221661255,
0.0124061005,
0.0683944449,
0.0149197364,
0.0176772326,
0.0263527464,
-0.0240716711,
0.0230162907,
-0.0712846592,
-0.0146239744,
-0.0173923057,
0.033942692,
-0.0338060297,
0.0129697872,
0.0153161325,
-0.003369512,
0.0283128526,
-0.032755211,
-0.0321039557,
0.0373779684,
-0.0389078744,
-0.0004841974,
0.0066777798,
0.0021251112,
0.012484408,
-0.039313063,
0.0527732745,
0.0350872613,
0.0590039007,
0.0218722001,
0.0436736234,
-0.030680785,
0.0291949827,
0.0648098141,
0.0424062461,
-0.0292195547,
-0.0326747075,
-0.0349143185,
0.000491549,
-0.0409477688,
0.018288672,
0.0229446162,
0.034294419,
0.0450669974,
0.0181749631,
0.0120761851,
-0.0552065633,
0.0141713247,
-0.0075130654,
-0.0351364538,
-0.0063988375,
0.0252656769,
0.0224779975,
0.0415064208,
-0.0209522769,
-0.0029296803,
0.0222989172,
-0.0229410939,
-0.0127959931,
-0.023956079,
-0.0223350134,
-0.0271115843,
-0.0068307966,
-0.0264967736,
-0.0334800966,
0.0573540963,
-0.0245329514,
0.0047772611,
0.0472539142,
-0.0310940836,
-0.0082984474,
-0.0723637491,
-0.0448366255,
-0.0322748497,
0.0399838053,
0.0090558846,
0.0792498887,
-0.0273395497,
-0.0049261968,
0.0089441864,
-0.0030814023,
-0.0149784479,
0.0019748546,
-0.0113529908,
-0.0080670882,
-0.0291099809,
0.0337194055,
0.0188143477,
-0.0197717436,
0.0009152784,
-0.0352924056,
-0.0031944136,
-0.0150314709,
-0.040633712,
-0.0138121489,
0.0440950654,
-0.0235978216,
-0.0116892112,
0.0126203997,
-0.0588253625,
-0.0060193408,
-0.0179665834,
0.023373967,
-0.0095706796,
-0.0209187828,
-0.0127062192,
0.0102819558,
0.0090777492,
-0.0102886772,
0.0563522875,
0.0379205607,
-0.0264682639,
-0.0224719476,
-0.0188850276,
-0.0434320532,
-0.0025781209,
-0.0334979743,
-0.0230971258,
0.0039326176,
0.0019866787,
-0.0119782221,
0.0603336543,
0.0685869753,
0.0131259104,
0.0062043495,
0.0612228364,
-0.0271344297,
-0.0343088061,
-0.0309163816,
0.0772719681,
-0.0103865899,
-0.0554225557,
0.0382192507,
-0.0090038264,
0.0415850244,
0.0292831678,
-0.0005613393,
0.0687290728,
0.0258907415,
-0.0261590518,
0.0593748912,
-0.0255858693,
0.0020497453,
-0.0121085597,
-0.0407279879,
-0.0159533434,
0.0799228027,
-0.0388107561,
0.0378595702,
0.0226537995,
-0.0237481333,
0.0150134014,
-0.0383419394,
-0.0187147018,
0.0294913892,
0.0094889654,
-0.0100329192,
-0.0195908286,
-0.0148187857,
0.0841479301,
-0.0323365442,
-0.0049464591,
0.0148230698,
0.030595839,
0.021903038,
0.0046169334,
-0.0672303587,
-0.0757207721,
0.015885666,
0.000183677,
-0.003715182,
-0.0300846919,
-0.0398865379,
-0.0245181657,
0.0303100478,
0.0570434406,
-0.0461017042,
-0.0082259122,
0.0049791806,
0.0711090639,
0.0168241635,
-0.0252139568,
0.0277839582,
-0.0278155468,
-0.0345066078,
0.0301411767,
-0.0219926909,
0.060437113,
-0.0489002168,
-0.030696122,
-0.008085696,
-0.0077733397,
0.0064212205,
0.0063881138,
-0.0040580714,
-0.0285451766,
0.0226443056,
0.0756703615,
-0.0067272391,
-0.0202492774,
-0.0148040662,
0.0405297168,
0.0154203102,
-0.000735816,
0.0149019463,
-0.0704632849,
0.0214626621,
-0.0079355994,
0.0285246633,
-0.0198223088,
-0.0149002075,
-0.0310906135,
-0.0381614827,
0.0081129204,
0.0100923013,
-0.0105013009,
-0.0202124622,
-0.0311580189,
-0.0169945154,
0.0152355032,
-0.0205084961,
-0.0295092743,
0.0429294966,
-0.0406612493,
-0.0104631083,
-0.0066670445,
-0.0729140118,
-0.0088845547,
-0.0252824426,
0.044205144,
0.059990909,
0.0369974189,
-0.0199865159,
-0.0731918365,
0.0000560179,
0.0161387287,
-0.0475292504,
0.0036949909,
0.0561531894,
-0.0355534293,
-0.0025553061,
-0.0162916277,
0.057357803,
0.0357803814,
-0.0145201245,
-0.0014169965,
0.0285797138,
0.0212307591,
-0.0967774242,
-0.0394507237,
0.0334008783,
0.0417204797,
0.0476849154,
0.0346774496,
-0.0286578257,
-0.00934783,
-0.0461238548,
0.0214236677,
-0.0689749718,
-0.0397233851,
0.0040452024,
0.0033656373,
-0.0334089957,
-0.0785392225,
-0.017604351,
-0.0632272586,
0.0301624946,
-0.0114035718,
-0.0323988944,
0.0053148791,
-0.0506216548,
-0.0382015891,
-0.0498356372,
-0.0059977034,
0.0374726728,
0.0029445931,
-0.0127971042,
-0.0149239935,
-0.0275338851,
-0.0113324113,
0.0367440768,
-0.0817448497,
-0.0458357856,
0.013673583,
0.0361709371,
0.0073806089,
0.0539100058,
-0.0520892851,
-0.0525001548,
0.0911199525,
0.0966242701,
-0.0068246033,
0.0092459973,
0.0338985324,
-0.0369244926,
0.0243745297,
0.0055196211,
-0.0252412055,
0.048343163,
0.0335266925,
-0.0572921,
0.001600586,
0.0338944383,
0.0156964064,
0.0060659386,
-0.0164497066,
-0.0266671963,
0.010721813,
-0.0303057916,
-0.055389706,
-0.020337183,
-0.0292545073,
0.0177803542,
0.0102232555,
0.0142634735,
-0.0286962651,
-0.0139228329,
0.0291616712,
-0.0033877569,
0.0358849168,
-0.0643977076,
-0.0542398132,
-0.0098427758,
0.0636230931,
-0.0188859459,
0.0042454093,
-0.0360163786,
0.059716884,
-0.0056425594,
0.0283030774,
-0.0214076303,
0.0285639856,
0.0317472331,
-0.0219278913,
0.0104189683,
-0.0065366975,
-0.0679738969,
0.0238477308,
-0.053163588,
0.0512362011,
0.0356541611,
-0.0083516371,
-0.0002412698,
0.0074583283,
0.0066023422,
-0.0113934521,
-0.01540705,
0.0176557992,
-0.0769428089,
-0.0267083179,
0.0311069507,
-0.0101686995,
0.0157285053,
0.0553139262,
0.007168937,
0.0186124444,
0.0151981805,
-0.0384397022,
-0.0355290696,
-0.0253559984,
-0.0222032536,
0.0440729409,
-0.0182244163,
-0.0366193019,
-0.0535096601,
-0.06105946,
0.0478488393,
0.0324976593,
0.0337883644,
0.0411075167,
-0.0374827869,
0.0678390935,
0.0031774226,
0.0270605516,
-0.0248902235,
0.0306104347,
0.0089493087,
-0.0524597168,
0.0121526131,
-5.542778023e-33,
0.0177667178,
-0.0381388366,
-0.0316447578,
-0.0029394713,
0.0013850323,
0.0100855008,
-0.0382449552,
-0.0669058561,
-0.0182630029,
0.033438541,
0.0034135715,
0.03914937,
-0.0220927913,
0.0180391334,
-0.0685468167,
-0.0213636979,
0.0547724329,
-0.0064494517,
-0.0564301051,
0.0317574181,
0.0548750199,
-0.0220984984,
0.0236623287,
-0.0367636345,
0.0278125424,
-0.0548210889,
-0.0480673388,
-0.0457971208,
-0.0062881047,
0.0314878412,
0.0122186858,
0.0161386319,
0.031441059,
-0.0339609385,
-0.0000949216,
0.0268945117,
-0.0214945097,
-0.0369714499,
0.0117344065,
-0.0130826095,
-0.0325533785,
-0.0062816236,
0.0477021448,
-0.0392890312,
-0.0077479845,
0.0100680739,
0.0271728244,
-0.0147554167,
-0.0311278012,
-0.0539275408,
-0.0298555121,
0.0183307007,
-0.007225947,
0.0073670228,
-0.0288148317,
0.0361643136,
0.032548584,
-0.0035736263,
-0.1067308336,
0.0252068713,
-0.0451831594,
-0.0070954626,
0.0233292207,
-0.0374422558,
-0.034649048,
0.0618859529,
-0.0106536131,
0.0502613708,
-0.0680284798,
0.0215643924,
0.0374064036,
-0.0061261267,
0.0757350847,
0.1024445221,
0.0187871698,
-0.0662559047,
-0.0494259894,
0.0510060377,
0.0371171497,
-0.0244332645,
0.016918283,
-0.0357576311,
-0.0229933355,
-0.0007883889,
-0.0053577237,
-0.0534465127,
-0.0292942394,
-0.0208010506,
-0.008572977,
0.0370587744,
-0.027989585,
0.0254188776,
-0.0077497987,
-0.0554382466,
0.0073939459,
0.0636280999,
-0.0421395898,
-0.0036250935,
0.0091061145,
0.0274027567,
-0.0238472689,
0.0290098805,
0.0464363098,
0.0053035035,
0.0053736256,
-0.01131461,
0.091505751,
0.0445255376,
-0.0114721116,
-0.0055253385,
-0.003923323,
0.0490886904,
0.0059947544,
-0.0351383053,
0.0238290392,
-0.0198539272,
-0.0223060083,
-0.0367963687,
-0.0089667635,
0.0033359476,
0.0160257109,
0.0298347306,
-0.0057758926,
0.0340791456,
-0.036927484,
-0.0133441444,
-0.0354925282,
0.0002536043,
0.0064097033,
-0.0940823853,
0.0275088213,
-0.0154436901,
2.543e-7,
-0.0307723153,
0.0205693338,
0.0100334994,
0.0340151116,
0.0349954702,
0.0246423297,
0.0316158608,
0.0593507253,
-0.0645492449,
-0.0824365094,
-0.040153604,
0.0062710373,
0.0400869772,
0.0553235747,
-0.0752343982,
0.0408691801,
-0.0990426168,
0.0377486311,
-0.0565568022,
-0.0202482808,
0.0029563825,
0.0332805775,
0.0376837365,
-0.0385960601,
0.0297369659,
0.0007050092,
0.0525514334,
0.0608362406,
0.0118581019,
-0.0022808851,
0.0853442177,
-0.024070438,
0.0190838855,
-0.0138577474,
-0.0453503728,
-0.0040527969,
-0.008193424,
0.0501458198,
0.0358928591,
-0.0004606712,
0.0154037196,
-0.0454175361,
0.0329484381,
-0.0434121825,
0.0786358565,
0.0004325252,
-0.0202595256,
-0.0656367689,
-0.0775936395,
-0.031379059,
0.0519307069,
-0.0151103474,
-0.0129392585,
0.0225766394,
-0.0199341979,
0.0036729516,
0.031478595,
-0.0353798755,
0.00680889,
-0.0139014497,
-0.0318358466,
0.046212215,
-0.0091306735,
-0.0146737136,
0.0596900024,
-0.0308645368,
0.0445339717,
1.934009402e-34,
0.0410700142,
-0.0073485081,
-0.0204200074,
-0.0264600758,
-0.0160133988,
0.0204616878,
0.0107819391,
-0.019138677,
0.0135543197,
-0.0223274585,
-0.0081666466
] |
http://arxiv.org/abs/2401.15499v1 | http://arxiv.org/pdf/2401.15499v1 | Semantic Properties of cosine based bias scores for word embeddings | 2024-01-27T20:31:10 | 2024-01-27T20:31:10 | Sarah Schröder, Alexander Schulz, Fabian Hinder, Barbara Hammer | cs.CL | cs.CL | Plenty of works have brought social biases in language models to attention
and proposed methods to detect such biases. As a result, the literature
contains a great deal of different bias tests and scores, each introduced with
the premise to uncover yet more biases that other scores fail to detect. What
severely lacks in the literature, however, are comparative studies that analyse
such bias scores and help researchers to understand the benefits or limitations
of the existing methods. In this work, we aim to close this gap for cosine
based bias scores. By building on a geometric definition of bias, we propose
requirements for bias scores to be considered meaningful for quantifying
biases. Furthermore, we formally analyze cosine based scores from the
literature with regard to these requirements. We underline these findings with
experiments to show that the bias scores' limitations have an impact in the
application case. | Semantic Properties of cosine based bias scores for word embeddings: Plenty of works have brought social biases in language models to attention
and proposed methods to detect such biases. As a result, the literature
contains a great deal of different bias tests and scores, each introduced with
the premise to uncover yet more biases that other scores fail to detect. What
severely lacks in the literature, however, are comparative studies that analyse
such bias scores and help researchers to understand the benefits or limitations
of the existing methods. In this work, we aim to close this gap for cosine
based bias scores. By building on a geometric definition of bias, we propose
requirements for bias scores to be considered meaningful for quantifying
biases. Furthermore, we formally analyze cosine based scores from the
literature with regard to these requirements. We underline these findings with
experiments to show that the bias scores' limitations have an impact in the
application case. | [
0.0039958432,
0.0428671502,
-0.0249146949,
0.0638369843,
-0.0433491133,
0.0283189062,
0.0098930979,
-0.0280372426,
0.0395980403,
-0.0411571562,
-0.0375595167,
-0.0384285487,
0.0580983087,
0.0006529931,
0.0185508579,
-0.0387172662,
0.0664367676,
0.0495799184,
0.0586219132,
-0.0324406214,
0.0191641375,
0.0026658918,
0.013080758,
0.0402294472,
-0.0286772195,
0.0085350387,
0.061666958,
0.0033211848,
-0.0069018202,
-0.0431369692,
0.0169719607,
0.0393403396,
0.0124435658,
0.038382221,
0.000001779,
-0.0686311349,
-0.0251565147,
-0.0120174671,
0.0176252015,
0.0029675097,
0.0896192789,
-0.0467043258,
0.0069681201,
0.0104110856,
-0.041783046,
-0.0162482988,
0.0366177633,
-0.0096884156,
0.0170431659,
0.0520594269,
-0.0278792623,
-0.050795991,
-0.0703737736,
-0.0300416183,
-0.0025400058,
-0.0256852433,
0.0228212774,
0.0106089879,
-0.0537386835,
-0.0155271878,
-0.0569911376,
0.0302243847,
-0.0274074357,
0.0505797639,
0.0278711263,
0.0128857093,
-0.0238257311,
-0.0330086909,
0.0170633737,
0.0427995771,
0.0156794209,
-0.0149894357,
-0.0082678273,
-0.0016388467,
0.0139998207,
0.065838106,
-0.0206208192,
-0.0071476791,
0.0381854698,
0.0215156395,
0.0385690033,
0.0473457873,
0.0200169925,
0.0293832403,
-0.010750737,
0.0414669886,
0.0377538241,
-0.0488565862,
0.0113748256,
0.0259729084,
0.0171734318,
-0.0114414264,
-0.0022861296,
-0.0257267877,
0.0042914362,
0.0349118002,
-0.0260120519,
0.0156787485,
0.0178986602,
-0.0813857168,
0.0199346896,
-0.012964366,
-0.0007167088,
0.0442810394,
-0.0227074064,
-0.0131171932,
0.0254325774,
-0.0092880279,
-0.0076831784,
0.0080878176,
0.0102306632,
0.0069287671,
-0.0324509107,
-0.0296753384,
0.0228621289,
-0.0116402302,
-0.0273185093,
0.0028168124,
-0.0546625033,
0.0214382056,
0.0092049427,
0.0289035738,
-0.0034966774,
0.0089708129,
-0.0092443116,
0.0207507387,
-0.0686979368,
0.0301661696,
-0.0775004104,
-0.0642054304,
-0.0063821208,
0.0140346903,
0.0349772833,
-0.0521429554,
-0.0273479559,
0.0366470814,
0.0557721145,
-0.0567589775,
0.0182553791,
-0.0264991615,
-0.0308667161,
-0.0062754094,
-0.0218490511,
-0.0324486122,
0.0382005796,
0.0373206437,
0.0167271383,
-0.0373988673,
-0.0471256673,
0.0028949117,
-0.0811206102,
0.0442678481,
0.029628031,
0.0329923034,
0.0062563727,
0.0131120142,
-0.006193947,
0.0878131911,
0.0123314084,
-0.0070215715,
0.0787481889,
-0.0180152655,
-0.0162442867,
0.0238680784,
0.0148972282,
-0.02389195,
-0.0275324397,
0.014987886,
-0.0030899059,
0.0467196293,
0.0441055112,
0.05254611,
-0.0227209143,
-0.0301605277,
0.0455855057,
0.0632831007,
0.0272895247,
-0.0081530698,
0.0165759046,
0.0005835617,
0.0095126517,
0.0261680391,
-0.0517507382,
-0.0169875734,
-0.0188972931,
-0.0183885824,
0.0178161524,
0.007299317,
-0.0006099924,
0.007877849,
-0.0281407349,
0.0028159251,
0.0208904594,
-0.0403416231,
0.0327203535,
0.008565139,
-0.0187599473,
0.0200434942,
0.0325719044,
-0.0431147553,
0.0602054186,
0.0086683929,
0.0196068026,
0.0224266611,
0.025616819,
-0.0311033223,
0.0295272935,
0.0413624011,
-0.030010825,
0.031117538,
0.0787989795,
-0.03300675,
-0.0034595004,
-0.0294829514,
-0.0141803082,
0.0136626335,
0.0660698637,
0.0028125809,
0.0417738184,
0.0405648611,
0.0199984685,
-0.0676757246,
-0.0198463462,
0.0381695516,
0.0184553005,
-0.0013412981,
-0.0627288967,
0.0198953059,
0.0040555685,
0.036795482,
-0.0515729375,
0.0212281533,
0.0021038465,
-0.0067290729,
0.0144502148,
0.0209555291,
-0.0525341704,
-0.0267738719,
0.0014201627,
-0.0897097066,
-0.0095706517,
0.0324869528,
-0.0194723681,
0.0018991991,
0.0579484478,
-0.0237492546,
0.0891505703,
-0.0024349014,
0.0019619914,
-0.0489805974,
0.0332340561,
-0.0295754299,
0.0487621762,
0.005505871,
0.0194133054,
0.0227757618,
-0.0555192828,
0.0338259265,
-0.0042974609,
-0.0519720241,
-0.0060634846,
0.0113304853,
-0.0398522429,
0.0408363268,
-0.0499539226,
-0.0308517776,
-0.0434157401,
-0.0789154544,
-0.0396224968,
-0.0411366262,
0.0166987479,
0.0383216962,
0.0107747633,
-0.01278337,
-0.0215356331,
-0.0294111576,
-0.0117175225,
-0.0186290666,
0.0503042787,
0.0358706005,
-0.0520718023,
-0.0148717957,
-0.0846157223,
-0.0190208741,
-0.0318998396,
0.0583953373,
0.039425943,
-0.0198654998,
-0.0448132083,
-0.0040417891,
-0.0433180965,
-0.0327222981,
0.0161510911,
0.0153016215,
0.0157030933,
0.0431738645,
0.0220429674,
0.0560841784,
0.0163681973,
0.0252869483,
0.0568182133,
-0.0097993668,
-0.0573644005,
-0.074659124,
-0.037539877,
0.0343812704,
-0.0389724188,
-0.0617498867,
-0.0039151218,
0.044509612,
0.0445780382,
0.0023811734,
-0.0495064706,
0.0541724116,
-0.0612035543,
-0.0191997122,
0.0112792663,
-0.0123563986,
-0.0014877178,
-0.0004944982,
-0.0551535003,
-0.0599327572,
0.0233343653,
-0.027499184,
0.0095193787,
-0.0326869674,
0.0333315022,
-0.0490870439,
0.0056786491,
0.0202510301,
0.02990539,
-0.0111573003,
-0.0079726465,
-0.0455787331,
0.0103122061,
0.0398719981,
-0.022617463,
-0.0146210622,
-0.0106117856,
-0.0188749135,
-0.0024069042,
0.0083400952,
-0.0120270038,
-0.0598441847,
-0.0374363512,
0.0204305947,
0.0269451812,
-0.0076026451,
0.0306036323,
-0.0115434574,
-0.0247687381,
-0.0323276669,
0.0070878607,
-0.0031444652,
-0.0011509059,
0.0175797828,
0.0572124943,
0.0015255248,
0.0377370827,
0.0386794731,
-0.0195191912,
0.0025268048,
0.0199764762,
0.0252579544,
-0.037560638,
0.0091688586,
0.0395077914,
-0.0317458585,
-0.0057589407,
-0.01935919,
0.0431403108,
0.0168920681,
0.0179215409,
0.1066053435,
0.0158974361,
-0.0508091561,
0.0461011976,
-0.0012922176,
0.0303222314,
0.0190094896,
-0.0502865538,
-0.0640468821,
-0.0058001792,
0.0119447904,
-0.0067092716,
0.0099148378,
-0.0335432366,
-0.0168010965,
-0.1163469627,
-0.0028488347,
-0.0039270995,
0.0375961438,
-0.0195172112,
-0.0716058165,
0.0381199718,
-0.0238044336,
0.0126701035,
-0.0281048249,
-0.008691038,
0.0268448722,
0.0066428888,
-0.0575608686,
-0.0769692138,
-0.0020718027,
-0.0677675009,
0.062007457,
0.0379142165,
0.0498458296,
-0.0283304472,
-0.0279241707,
0.0071778712,
-0.0139145171,
-0.0274517871,
-0.029194802,
0.0407640785,
-0.0693311468,
-0.0049349163,
-0.0043906597,
0.0234485324,
0.0185830202,
0.0242115762,
-0.0102068754,
0.0685194656,
0.0235723592,
-0.0604651757,
-0.0346224569,
-0.0054898802,
0.0075226971,
0.110213019,
0.040039707,
-0.01340215,
-0.0014420402,
-0.0061284369,
0.022811193,
0.0636913106,
-0.041442994,
-0.0129751721,
0.0233166721,
-0.0188085418,
-0.0719311386,
-0.0660151392,
0.0050052758,
0.0252358578,
0.0290095899,
-0.0168800224,
-0.0037745258,
-0.0322344191,
-0.004727134,
0.018467823,
-0.0105929486,
0.029744966,
0.0218717083,
0.0332248509,
-0.0096243396,
-0.0107590379,
-0.033999946,
-0.0433877222,
-0.029611744,
-0.059612859,
0.0563425049,
-0.0275724847,
0.0165428706,
0.0357155576,
-0.0336678438,
-0.0722657815,
0.0784836188,
-0.0042399876,
-0.0326495245,
0.0428478867,
0.0330219492,
-0.0106538627,
0.0069805803,
0.0512039289,
-0.0405531786,
0.0032996046,
0.011071167,
0.0286804438,
0.0053740581,
-0.03327531,
-0.0201822743,
-0.0137079787,
0.0858260691,
-0.0426004454,
0.0623069257,
-0.0475096554,
-0.0585514009,
-0.0379333347,
-0.0223495644,
0.003056172,
-0.0432183407,
0.0316563174,
0.0194873549,
-0.062915206,
-0.0039037666,
-0.0580163151,
0.0206710789,
-0.0144851431,
0.0597747527,
0.040604718,
0.0144664077,
-0.021905249,
0.0064818752,
-0.0162495319,
0.0197482239,
-0.0615703464,
0.0067337574,
-0.0137450099,
-0.0103210052,
0.0048635281,
-0.0089449659,
-0.01726171,
0.0240955893,
-0.0390148349,
0.0135605121,
0.0119320853,
0.0073992917,
0.0145873986,
-0.0911343172,
0.0658264607,
-0.0205626469,
0.0215376019,
0.0040193121,
-0.0127258822,
-0.0032092326,
-0.0248798858,
0.0299961083,
-0.0092413444,
-0.023728082,
0.0322862193,
0.0839093924,
-0.0054210396,
0.0411855727,
-0.034409903,
-0.0535526052,
0.0237082038,
-0.0313373059,
0.0613048933,
0.0146999741,
-0.0564434566,
-0.0329802446,
-0.1180859804,
-0.0564370342,
-0.0149614913,
0.009738123,
0.0049941051,
0.0090037165,
0.011366874,
0.0748234987,
-0.0300113093,
0.0051739411,
-0.0076818326,
-0.0092728334,
-0.0462565869,
-0.0145135345,
-0.0345292427,
-5.954259821e-33,
0.005334009,
0.001487921,
-0.0410542451,
-0.0183379054,
-0.0282152779,
0.0084623387,
-0.0313198753,
0.0286081005,
-0.0559813641,
0.0267944857,
0.004372566,
0.0109757287,
0.0179980844,
-0.0183264855,
0.0068941633,
-0.0268012378,
0.0254752245,
0.0130736288,
-0.0332721509,
-0.0223362222,
-0.0181667842,
-0.0042059924,
0.0488400906,
-0.0576051325,
0.0870438367,
0.0171179194,
0.0352788456,
-0.0003596475,
0.0524397902,
-0.0044185375,
-0.0380564369,
0.0723432451,
0.0399777591,
0.0142578948,
-0.0119670145,
0.045802664,
-0.0398740359,
-0.0505889729,
0.0644132942,
0.0237059724,
0.0053539062,
0.0231347103,
0.0260802358,
0.0002830517,
-0.0740639716,
0.0029296591,
0.0449817926,
0.0224795975,
-0.0005710124,
0.0675470755,
-0.0712513179,
0.0447136052,
-0.0078088474,
0.0440003946,
0.0346056744,
0.0047954172,
0.0511143282,
0.029254755,
0.002369517,
0.019339785,
-0.0613559224,
0.0044505084,
0.0325011387,
0.0030395675,
0.0206936914,
0.1025451422,
0.0506527387,
0.0058139437,
-0.0090387911,
-0.0572223365,
0.0726492628,
-0.0281346925,
0.0233594757,
0.0602333806,
0.0323577188,
0.0183938835,
-0.0213703513,
-0.0176550075,
0.0156568997,
0.0352142304,
0.0255937539,
-0.0138828298,
-0.1062485427,
-0.0392317064,
-0.0542115234,
-0.0781014413,
-0.0165109951,
-0.0351952016,
0.0120449755,
0.001849412,
-0.0186743364,
0.0274693556,
0.0078202989,
-0.0397355147,
-0.0242720656,
0.0230659712,
-0.0042572678,
-0.0142035773,
0.0113549922,
0.0326555856,
-0.0414874703,
-0.0104468837,
-0.0083730025,
0.0659564212,
0.0387382656,
-0.0311699696,
0.0123281395,
0.0116873886,
-0.0340096317,
-0.0209056474,
-0.0212057475,
-0.0695454106,
-0.0227750875,
-0.0290442109,
0.0590205044,
0.0236467719,
-0.0217080507,
-0.0137401046,
0.0378156975,
0.0462215543,
-0.0213056635,
-0.0164534077,
-0.0174056105,
0.0159253925,
-0.0377737433,
-0.0211419761,
0.0072240504,
0.0147313736,
0.0344801135,
-0.029229667,
-0.0136310449,
-0.0424246602,
2.458e-7,
0.0407341123,
0.0477133356,
-0.001443533,
0.007175805,
0.0211339574,
0.0082996991,
0.0144179305,
0.0734983906,
-0.0070534465,
-0.0005792834,
-0.0012610555,
0.0025734468,
-0.0016819201,
0.0241769962,
-0.0562994704,
0.0224521924,
-0.0668230429,
0.0419503041,
-0.0100594852,
0.0039370302,
0.0149312206,
0.0508159511,
0.0204365123,
0.0276137963,
-0.0407742597,
-0.0225791074,
0.0056124469,
-0.0586209744,
0.0220485348,
0.0048602833,
0.0080991024,
0.0077810073,
0.0710332245,
0.0067175659,
0.0058159647,
0.0385361984,
0.0260087531,
0.0033126969,
0.0005728789,
0.0295596011,
-0.0429056585,
-0.0354808085,
-0.0228907745,
-0.0883056968,
0.0721068084,
-0.0111273481,
-0.0305720791,
-0.0752750486,
-0.0025942763,
0.0116043705,
0.0540020876,
0.0046099331,
-0.0253037084,
-0.0204553921,
0.0565194562,
-0.0017448268,
-0.012227132,
-0.0687262416,
-0.0279927291,
-0.0134987766,
0.0011343801,
-0.0498680584,
0.0121417837,
0.0440298505,
0.0435738452,
-0.0088269301,
0.028167028,
1.995006309e-34,
0.0315068029,
-0.0091108643,
0.0394793637,
-0.0468306914,
0.0446486361,
-0.0130693223,
-0.0200488381,
-0.0188997127,
0.0441248491,
-0.075974226,
0.0070799552
] |
http://arxiv.org/abs/2401.15498v2 | http://arxiv.org/pdf/2401.15498v2 | Do We Need Language-Specific Fact-Checking Models? The Case of Chinese | 2024-01-27T20:26:03 | 2024-02-17T13:10:25 | Caiqi Zhang, Zhijiang Guo, Andreas Vlachos | cs.CL | cs.CL | This paper investigates the potential benefits of language-specific
fact-checking models, focusing on the case of Chinese. We first demonstrate the
limitations of translation-based methods and multilingual large language models
(e.g., GPT-4), highlighting the need for language-specific systems. We further
propose a Chinese fact-checking system that can better retrieve evidence from a
document by incorporating context information. To better analyze token-level
biases in different systems, we construct an adversarial dataset based on the
CHEF dataset, where each instance has large word overlap with the original one
but holds the opposite veracity label. Experimental results on the CHEF dataset
and our adversarial dataset show that our proposed method outperforms
translation-based methods and multilingual LLMs and is more robust toward
biases, while there is still large room for improvement, emphasizing the
importance of language-specific fact-checking systems. | Do We Need Language-Specific Fact-Checking Models? The Case of Chinese: This paper investigates the potential benefits of language-specific
fact-checking models, focusing on the case of Chinese. We first demonstrate the
limitations of translation-based methods and multilingual large language models
(e.g., GPT-4), highlighting the need for language-specific systems. We further
propose a Chinese fact-checking system that can better retrieve evidence from a
document by incorporating context information. To better analyze token-level
biases in different systems, we construct an adversarial dataset based on the
CHEF dataset, where each instance has large word overlap with the original one
but holds the opposite veracity label. Experimental results on the CHEF dataset
and our adversarial dataset show that our proposed method outperforms
translation-based methods and multilingual LLMs and is more robust toward
biases, while there is still large room for improvement, emphasizing the
importance of language-specific fact-checking systems. | [
0.0524824411,
0.0518139042,
0.0056778723,
0.0817715153,
-0.0651323795,
-0.0027036429,
0.0296269581,
-0.0133047327,
0.0216760114,
-0.0306663085,
-0.0020742458,
-0.036715202,
0.0174816903,
0.0183698386,
0.0188592654,
0.0051832302,
0.0641006753,
0.0185469352,
0.0797113478,
-0.0255651213,
0.0207811967,
0.0151684349,
0.0269488674,
0.0159396306,
-0.0533698536,
-0.0181946009,
0.0199049599,
-0.000032344,
0.0343438908,
-0.0098751122,
0.0063328044,
0.0210209098,
-0.0112915253,
0.1197750047,
0.000001627,
-0.0239488389,
-0.0243785158,
0.0300135668,
-0.0150765609,
0.0276307892,
0.0269894246,
-0.033479955,
-0.0610825494,
0.0120879514,
-0.0232556295,
-0.0503044091,
0.0387663357,
0.0859294683,
0.0697095394,
0.0930469483,
-0.009168175,
-0.0350293405,
0.0066149421,
-0.0121617522,
-0.0020250268,
-0.0359614715,
0.0521977358,
0.0286914557,
-0.0580333918,
0.0209356491,
-0.0030574023,
0.0230035968,
0.0469966009,
-0.0047448552,
-0.0788556784,
0.0100927399,
-0.0426480547,
0.0174006112,
0.0155343153,
0.0352594592,
-0.0094382996,
-0.0636969954,
-0.0168546792,
-0.0052638329,
0.012985399,
0.008263614,
-0.008295808,
0.0234835315,
-0.0456846021,
0.0122308293,
0.0310505927,
-0.0344089903,
0.045140516,
0.0394926518,
-0.0128909443,
0.052029103,
0.005665964,
-0.0609960817,
-0.0088002384,
-0.0009312012,
0.0161255486,
-0.0744485557,
0.0097545991,
-0.007547433,
-0.0148779657,
0.0175139364,
-0.0202810057,
0.0006734009,
0.0445612669,
-0.0464760922,
-0.0394351371,
-0.0236413572,
-0.0778450593,
0.0572260953,
-0.0319840237,
0.0026949584,
-0.0304849502,
0.0181125961,
-0.0088094976,
-0.0690544546,
0.020121498,
0.0142353717,
-0.0199346002,
0.0097791115,
-0.0078865336,
0.0508970618,
-0.0359051339,
-0.0028589654,
-0.0280207824,
0.0540718064,
-0.0049586678,
-0.0207372028,
-0.0221655294,
-0.0211151075,
-0.0496431254,
0.0315579548,
-0.0231957566,
0.0423427112,
-0.0137482025,
-0.0542246252,
-0.0110058011,
0.019132074,
-0.0220064055,
-0.0123291314,
0.0204484109,
0.0433059782,
0.0243239962,
0.015845757,
-0.0283016078,
-0.0017166714,
-0.0301477276,
-0.0795745105,
-0.0211854018,
-0.0787284747,
-0.0083803097,
-0.0191270318,
0.002242822,
-0.0118401702,
-0.0134845898,
0.0298371073,
-0.0144893415,
0.0672973171,
0.0108256824,
-0.0242313705,
0.0434721783,
-0.0049679712,
-0.0088703707,
0.0477736443,
0.0113692526,
0.0517119728,
0.079502888,
0.0417137481,
0.0040292991,
0.0007825606,
0.0322116539,
0.0232723635,
0.0080844229,
0.0114041539,
-0.0020924499,
-0.0140212951,
-0.0067082983,
0.0573873781,
-0.0159300491,
-0.0291216709,
0.0701599643,
-0.0420297347,
-0.0312717147,
-0.0180205982,
0.0147113819,
-0.0193306059,
0.0204437822,
0.0348270983,
-0.0474498384,
0.0247043986,
-0.0315644257,
0.0025177526,
0.0019527727,
0.048073031,
-0.0196585506,
-0.0687604398,
-0.0332031697,
0.0072402195,
-0.0058414792,
0.002868386,
0.0304719005,
0.012104922,
-0.024497645,
0.0005423063,
0.0170936901,
-0.0389193408,
-0.0197749827,
-0.0464004613,
0.0234596115,
0.0197842512,
0.0608765185,
-0.0403697863,
0.0415858515,
-0.0092274603,
0.0054993983,
0.0520365238,
0.0621335693,
-0.0232443884,
-0.0063128467,
-0.0188993514,
-0.0136417663,
-0.0263445042,
0.0317513943,
0.0249413494,
0.0511455871,
0.0554164611,
-0.0026803997,
-0.0515153669,
-0.0533242263,
0.030922953,
0.0058189328,
-0.0115203448,
0.0104397936,
0.0284013823,
0.0268877577,
0.0350633413,
-0.0414458103,
0.0006144809,
0.0355040282,
-0.013680798,
-0.0358518846,
0.0043779537,
0.0083254557,
-0.0607578568,
0.0458750203,
-0.0371906757,
-0.0526831672,
0.0259122979,
-0.0730945021,
-0.0157852899,
-0.0262757428,
-0.0002116207,
0.0361692682,
-0.0619055666,
-0.0126871774,
-0.0074802474,
0.0286324825,
-0.0080329292,
0.0883845836,
-0.0160919987,
0.0140329013,
0.0017919278,
0.0257811733,
0.0371711813,
-0.0222123135,
-0.0563848168,
0.038606409,
0.0145816375,
0.0351743996,
0.0407734178,
-0.001419411,
0.004426348,
-0.013572026,
-0.065556854,
0.0077569876,
-0.0230347384,
0.0082135061,
0.0242362749,
-0.0287965387,
-0.0094929608,
-0.0060341586,
-0.037309397,
-0.0243578274,
-0.0079576122,
0.0279685017,
0.0394243449,
-0.0195068829,
0.0159879662,
-0.0260645077,
0.0062265145,
-0.0016055108,
0.0093086055,
0.0244582314,
-0.0240535326,
-0.0381976776,
-0.0122286277,
-0.0515151918,
0.032710433,
-0.0251632798,
-0.038455613,
0.0177706853,
0.0138548529,
0.0052990159,
0.0609783344,
0.0078256167,
0.0223553628,
0.0586212389,
0.001721564,
-0.002111536,
0.0030550314,
0.0126287891,
0.0070644733,
-0.0013350103,
-0.0582945012,
0.0173557196,
0.0092325434,
0.0612364784,
-0.023084864,
0.0098388614,
0.0282385945,
-0.0184021909,
-0.020552611,
0.0262695812,
-0.0023723771,
-0.031532716,
0.0208923519,
-0.0386584252,
-0.033752732,
0.0853747949,
-0.0403746404,
0.038490355,
-0.0532945022,
0.0056856438,
-0.0077738701,
0.0440337099,
-0.0039260848,
0.033095777,
-0.0102041196,
0.0247147754,
-0.0340579599,
0.0003248881,
0.0597645417,
-0.033522781,
0.0068431706,
-0.0135939009,
-0.0001007052,
-0.0057098879,
-0.0264029745,
-0.0563749596,
-0.033989843,
-0.0049576736,
0.005394374,
-0.0108669708,
-0.0060236328,
-0.0592398345,
-0.0079978732,
0.0380901508,
0.0568702333,
-0.0098379953,
0.0116221309,
0.0130951554,
0.0184683166,
0.0170265269,
-0.0219863094,
0.0514244959,
0.0045652874,
0.0191040542,
0.080437988,
-0.0094220098,
0.0169373937,
0.0128718857,
-0.0644957721,
-0.0141713945,
-0.0256394334,
0.0150219239,
-0.0201364439,
0.0139369546,
-0.0623641424,
0.0159513168,
0.0777270272,
0.0050586653,
-0.003868317,
-0.0221639778,
0.0064156419,
-0.0348012894,
0.0210509747,
-0.0172977801,
-0.1196850911,
-0.0110449735,
0.0099464087,
-0.0236353725,
-0.0390797183,
0.006345442,
-0.025568353,
-0.107728079,
0.0469692647,
-0.0020493567,
-0.0437031165,
-0.0650060102,
-0.0495146811,
0.003008313,
-0.00512368,
0.0394874625,
-0.0037674017,
0.0499392226,
-0.0238043256,
-0.0344405025,
0.0157519747,
-0.047554668,
0.0436171405,
-0.0485467054,
0.0171515476,
0.0375287645,
0.0625213385,
-0.0268350802,
-0.0083664209,
0.0251675025,
0.0117345182,
-0.046213299,
-0.0032191835,
0.0441660658,
-0.0392304286,
-0.0031207425,
-0.023977451,
0.0176792871,
0.0444365703,
-0.0026937942,
0.037452165,
0.0454153232,
0.0232416783,
-0.0253481418,
-0.025394123,
-0.0205140132,
0.0161871836,
0.0952339321,
-0.017142605,
0.0032947254,
-0.0242072027,
-0.0455675572,
-0.0421951152,
-0.0228511132,
-0.0146775376,
0.0557096601,
-0.0409906805,
-0.0347046889,
-0.0830603763,
-0.0193801578,
-0.0088996151,
0.0238758046,
-0.0226588156,
0.0244526062,
0.0124158887,
-0.0070696701,
-0.0082205525,
-0.0252785999,
-0.039878007,
0.0348867439,
-0.0201981086,
-0.0146037601,
0.013717222,
-0.0270903222,
0.0145816496,
0.0488656051,
-0.0533895083,
-0.063047789,
0.0024478217,
-0.0083135786,
0.0122513874,
0.0821594149,
-0.0550780259,
0.0177347697,
0.0452679433,
0.0453069583,
-0.0319056585,
-0.0046137357,
0.0368506387,
-0.0431362838,
0.0094428621,
0.0204227231,
-0.0567954592,
0.0286190659,
0.036002174,
0.0054891896,
-0.0432733409,
-0.0089727081,
0.0537034981,
0.0504210107,
0.047592856,
-0.0145252179,
0.0012160892,
-0.0577336363,
0.0352694839,
-0.0423126966,
-0.0533179715,
0.024421338,
0.014910005,
-0.0186969675,
-0.0446123667,
-0.020791227,
0.0238809083,
-0.0203026496,
0.0185528062,
-0.0134113422,
-0.0033930894,
0.0076720305,
0.0731053203,
-0.0068767397,
-0.0254485402,
-0.0358418897,
-0.0064480179,
-0.011124948,
0.0085769473,
-0.0042126263,
0.0183262117,
0.0416922234,
0.0014200548,
0.0306682251,
-0.0259151142,
-0.0950022861,
0.0215474032,
-0.0362376012,
0.0058959876,
0.0651653484,
-0.0200455952,
0.0167923104,
-0.0472722165,
-0.031606365,
-0.0190744866,
-0.0032017503,
0.0495425463,
-0.0160362888,
0.0124531314,
0.0032971303,
-0.0484752692,
0.0267605446,
0.0490043983,
0.0189599283,
0.0623520613,
0.0138390558,
-0.0678503215,
-0.0272649657,
-0.0435679592,
-0.0070509966,
-0.0132683981,
-0.0169661809,
-0.0283227451,
-0.0644425228,
-0.0724187121,
-0.0245992821,
0.0580387376,
0.0113149034,
0.0227173083,
-0.0369190387,
0.086376749,
0.0321105495,
-0.0068541048,
-0.0426220223,
0.0588043854,
0.0239587929,
-0.0271744411,
0.0191739928,
-5.509599695e-33,
-0.0187348537,
-0.060837891,
-0.0245465543,
-0.0082588578,
0.0350599959,
0.0273237955,
-0.0201583095,
-0.0539121367,
-0.0853888318,
0.0557020828,
0.0035512024,
0.0360356495,
0.0253843945,
-0.0150142536,
-0.0436265543,
0.0061468245,
0.0728430375,
0.0161767937,
-0.0127932737,
-0.0151461242,
0.0599866137,
-0.0278025288,
0.085157983,
-0.0629488826,
0.0851779804,
-0.0284069106,
-0.0325191803,
0.0155387521,
-0.0514125079,
0.0394217409,
0.0260292813,
0.0565723516,
0.0288586952,
-0.0265247244,
-0.0045574689,
-0.0386281759,
-0.0859859362,
-0.0169099569,
0.0576949604,
0.0275670234,
0.0170425419,
0.0087718898,
0.0249667,
-0.0148415333,
0.0067321039,
0.0171576384,
0.0310470685,
-0.0401074998,
0.0039464612,
-0.0105579039,
-0.0617069826,
0.044255197,
0.051405631,
0.0485723279,
-0.0394267328,
-0.0001847916,
0.0426365994,
0.0070242332,
-0.095641546,
0.0017491491,
-0.0006136561,
0.0155962231,
0.0268301386,
-0.0249154922,
-0.0029634556,
0.0381100774,
0.0009915916,
0.022690611,
-0.0402834825,
-0.0145769622,
0.0016617067,
0.023223279,
0.0442191698,
0.1410265118,
0.0213074517,
-0.0511852652,
-0.1047045365,
0.0282660574,
0.0504906736,
0.0311325304,
0.0208839588,
-0.010851982,
-0.0432986952,
-0.0009960666,
-0.0128673734,
-0.0155751538,
-0.0335977338,
0.0010820915,
0.0031711734,
0.0295920353,
-0.0101986034,
-0.0236676726,
0.0028678384,
-0.0496920869,
0.0073453407,
0.0600133836,
-0.0384715274,
-0.036072094,
-0.0124694286,
0.0323884077,
-0.0685097352,
0.0172453355,
-0.0114549007,
0.026465876,
-0.0103201754,
0.00865536,
0.0881416127,
-0.0105806775,
-0.0582374372,
0.0427375436,
-0.0272782128,
0.0274541192,
0.0116660288,
-0.0273568723,
0.0243844967,
-0.0152118355,
0.0024799567,
-0.0059740827,
0.0186682027,
0.0079796417,
-0.0532617643,
-0.0056582531,
-0.0177375823,
0.0243070088,
-0.0387793332,
0.0086358031,
-0.0066108326,
0.0008696605,
0.0208140854,
-0.064794749,
0.0114500811,
0.0215260591,
2.324e-7,
-0.0068933112,
0.0122850593,
0.0285556745,
0.0290330648,
0.0154959746,
0.0439435281,
-0.0077640223,
0.0122152949,
-0.0434095562,
-0.0441526845,
-0.010739712,
0.0324170813,
0.0305266399,
0.0738951936,
-0.0587879904,
0.0528555512,
-0.0627400428,
-0.0069483006,
-0.0311498884,
0.0069896299,
-0.0145017942,
0.0909668654,
0.0246240236,
-0.0013063343,
-0.0179419592,
-0.0393529348,
0.036295753,
0.0476540811,
-0.0357549079,
0.0207551531,
0.0823391825,
0.0093572326,
0.0261835009,
0.0076782056,
-0.0142444884,
-0.0016230423,
0.018463634,
0.0299788266,
0.0069769393,
0.001363523,
-0.0341687985,
-0.0488900952,
-0.0083744917,
-0.0793411955,
0.0707379356,
0.0641064867,
-0.0064861476,
-0.0556031466,
0.0085860137,
0.0079399263,
0.0331605449,
-0.0119166533,
-0.0265079252,
0.023220554,
-0.002083384,
0.0191166904,
0.0093225548,
-0.021680031,
-0.0434838608,
-0.0164383277,
-0.0292382035,
-0.0563844852,
0.0500300452,
-0.004664233,
0.0240081046,
-0.0559414364,
0.0204498675,
1.831359932e-34,
0.0394322686,
-0.021744661,
-0.0258944836,
-0.0422902666,
0.0594682992,
0.0070528071,
-0.0303866789,
-0.0149962958,
-0.0241798926,
0.0128992992,
-0.0364171192
] |
http://arxiv.org/abs/2401.15496v3 | http://arxiv.org/pdf/2401.15496v3 | Baichuan2-Sum: Instruction Finetune Baichuan2-7B Model for Dialogue Summarization | 2024-01-27T20:20:39 | 2024-04-04T03:15:15 | Jianfei Xiao, Yancan Chen, Yimin Ou, Hanyi Yu, Kai Shu, Yiyong Xiao | cs.CL | cs.CL, cs.AI, cs.LG | Large language models (LLMs) like Llama, Baichuan and Bloom models show
remarkable ability with instruction fine-tuning in many natural language tasks.
Nevertheless, for the dialogue summarization task, which aims to generate
summaries for different roles in dialogue, most of the state-of-the-art methods
conduct on small models (e.g Bart and Bert). Existing methods try to add task
specified optimization on small models like adding global-local centrality
score to models. In this paper, we propose an instruction fine-tuning model:
Baichuan2-Sum, for role-oriented diaglouge summarization. By setting different
instructions for different roles, the model can learn from the dialogue
interactions and output the expected summaries. Furthermore, we applied NEFTune
technique to add suitable noise during training to improve the results. The
experiments demonstrate that the proposed model achieves the new
state-of-the-art results on two public dialogue summarization datasets: CSDS
and SAMSUM. We release our model and related codes to facilitate future studies
on dialogue summarization task. | Baichuan2-Sum: Instruction Finetune Baichuan2-7B Model for Dialogue Summarization: Large language models (LLMs) like Llama, Baichuan and Bloom models show
remarkable ability with instruction fine-tuning in many natural language tasks.
Nevertheless, for the dialogue summarization task, which aims to generate
summaries for different roles in dialogue, most of the state-of-the-art methods
conduct on small models (e.g Bart and Bert). Existing methods try to add task
specified optimization on small models like adding global-local centrality
score to models. In this paper, we propose an instruction fine-tuning model:
Baichuan2-Sum, for role-oriented diaglouge summarization. By setting different
instructions for different roles, the model can learn from the dialogue
interactions and output the expected summaries. Furthermore, we applied NEFTune
technique to add suitable noise during training to improve the results. The
experiments demonstrate that the proposed model achieves the new
state-of-the-art results on two public dialogue summarization datasets: CSDS
and SAMSUM. We release our model and related codes to facilitate future studies
on dialogue summarization task. | [
0.0614281259,
0.0029341818,
0.028141886,
0.0501148663,
-0.0112620424,
-0.0006723399,
-0.0201033633,
-0.0669632405,
-0.0116993999,
-0.0319391899,
-0.0344616883,
-0.0799228996,
0.0127731645,
0.0037563385,
0.0088170841,
-0.089620471,
0.0525285825,
-0.0135536781,
0.010787718,
-0.0297202263,
0.0217848867,
0.002971391,
0.0157170203,
0.0349441431,
0.030777134,
0.0128181605,
0.0059060263,
0.0153587377,
0.0440743417,
0.0098440349,
0.0262241308,
0.0342800096,
0.0120335016,
-0.0064826282,
0.0000019392,
-0.0487205535,
-0.0181228034,
-0.0176282432,
0.0135836583,
0.0464740992,
0.0253541078,
-0.0060340744,
0.0182044543,
0.0450513475,
-0.0238198731,
0.014902439,
0.0387948453,
0.0920013711,
0.0855826065,
0.0766783282,
0.0145775825,
-0.0671985,
-0.0124310348,
-0.0535904132,
-0.0129496772,
-0.0637038276,
0.0111437654,
0.019674629,
-0.011417903,
0.0009462386,
-0.0054974989,
0.0165275224,
-0.0078288177,
-0.0373816602,
0.0014056691,
0.0092228679,
-0.0220402926,
0.0038724011,
0.0268655699,
0.0770364106,
0.0297737345,
0.0404652096,
0.0156437829,
0.0303599741,
0.0201695282,
0.000259613,
-0.0692871436,
0.0452197753,
0.022680074,
0.0115668746,
0.0048091235,
0.0338491574,
-0.014182345,
0.0298554506,
-0.0375970341,
0.0077637406,
-0.0110834278,
-0.0223137308,
-0.0554561503,
-0.0338251814,
-0.0557380542,
-0.028123185,
0.0003510978,
0.0337797925,
-0.0598915219,
-0.0101837143,
-0.0312991403,
0.0421429463,
0.0013585426,
-0.0143290805,
0.029500436,
0.025686305,
0.0111718941,
0.0328391939,
-0.0092811929,
0.0492003821,
0.0428388491,
0.0180391427,
-0.0636657923,
-0.025108343,
-0.026408663,
-0.0057347715,
-0.0579547733,
0.0313830413,
-0.0659157559,
-0.0469742119,
-0.0497827604,
0.011402851,
0.004920369,
-0.011519881,
0.0433301479,
0.0342869647,
-0.0122176548,
0.0115477163,
0.0009420908,
0.0101427846,
-0.0361498296,
-0.0109834364,
0.0372729562,
0.0154639985,
-0.0360320434,
0.0178585295,
0.0151128126,
0.0031754149,
-0.0137776006,
0.0428921208,
-0.0366438292,
-0.0137309423,
0.0045650718,
-0.0205052234,
-0.046071209,
-0.0511604957,
-0.0195711106,
-0.0454870053,
-0.0066565629,
-0.0375473015,
-0.0001768592,
-0.0046927407,
0.0193696879,
-0.0066037597,
-0.0332261734,
0.0209873728,
-0.0005049192,
0.0139928311,
0.0203786436,
0.0175591204,
0.0500966124,
0.046453014,
0.0501192287,
0.0210415423,
0.0538446419,
0.0172886085,
-0.0422660671,
0.0202650707,
0.0061997073,
0.0104928855,
0.0428648293,
-0.0140329916,
-0.0410781056,
-0.0082714427,
0.0315309837,
0.0664466396,
-0.0621366501,
0.0231199842,
0.0630817637,
0.0669740513,
0.0148760714,
0.0007595879,
0.0162708405,
0.0493805297,
0.018716773,
0.0682450458,
-0.0322662704,
0.0379074067,
-0.0242784601,
0.0564650781,
0.1145053059,
0.0201979931,
-0.0448474549,
-0.0504448153,
0.0311647672,
-0.01889617,
0.0198560003,
0.0069423281,
-0.0060886317,
0.0212651454,
-0.0457372032,
0.0471352227,
0.0061811204,
-0.0773507059,
-0.0001389133,
-0.0374392271,
0.0008729987,
0.0466821417,
0.0331218764,
-0.0573875271,
-0.0333729014,
-0.0219767708,
-0.0081788702,
0.0304691829,
0.0355611853,
-0.0214298815,
0.0028986596,
-0.0689087063,
-0.0288472865,
-0.0182725377,
0.0281894542,
-0.0145232901,
-0.0034821071,
-0.049396541,
0.0332682878,
0.0262897611,
-0.0679203495,
-0.000387466,
-0.0189167019,
-0.0167073105,
0.0137472693,
0.0034572433,
0.0226947796,
0.0457595438,
-0.0293786302,
0.0098616704,
0.024453491,
-0.0297717825,
-0.0249329302,
0.0010851756,
-0.0713041201,
-0.0564994588,
0.0530872494,
-0.03699404,
0.0148638943,
0.0110302158,
-0.048726622,
-0.085266009,
0.0440414064,
-0.0652207732,
-0.0336237103,
-0.052976042,
-0.0024445786,
0.0076196278,
0.006136545,
-0.0432483926,
0.0531053357,
-0.022241855,
-0.0114607895,
0.0622493066,
0.0233067311,
-0.0010746387,
-0.0584739968,
-0.0176463835,
-0.0040239133,
0.0102564963,
0.0176184028,
0.059286207,
0.0224893056,
0.0025375029,
-0.0608233921,
-0.1076356843,
0.0268634334,
-0.027067285,
-0.0014645428,
0.0639474392,
0.0339196809,
-0.0029721512,
0.0141097242,
-0.0488845557,
-0.0262060408,
0.0391588844,
-0.0329942107,
0.0179855172,
0.0164434928,
-0.0022659602,
-0.0603625067,
0.0006584665,
0.0223054551,
0.0080249095,
-0.0233973302,
-0.01408746,
-0.0072367392,
0.0259806253,
-0.0138277365,
-0.0019146528,
-0.0060467212,
-0.041609358,
-0.028148029,
0.0214064382,
0.0410861038,
0.0528639853,
0.0538454428,
0.0140464846,
-0.0026845129,
-0.0158810187,
-0.0310728811,
0.0096278787,
-0.029942682,
-0.001809617,
-0.007432159,
-0.0504569001,
0.0083448235,
0.0620333552,
0.0216811541,
-0.001936878,
-0.0232694224,
-0.0003529384,
0.009441142,
-0.0092820683,
0.0257709026,
-0.017534893,
-0.0370936915,
-0.0064183483,
-0.0579850897,
0.0074895518,
0.0558443964,
-0.0320522562,
-0.0006800378,
-0.0741769746,
-0.0118801333,
-0.0739351138,
0.0352630429,
-0.0122290784,
0.0482786596,
-0.0065375674,
-0.0378628783,
-0.0499835722,
-0.0295190122,
0.0708056837,
-0.0238722712,
0.0673359111,
-0.0166244637,
0.0025827347,
-0.0352982208,
0.0007272425,
-0.0182454623,
-0.0346189588,
-0.0226364341,
0.082442373,
0.0634518862,
-0.0217740946,
-0.0413898788,
0.0303497054,
0.0270288493,
0.0401544236,
-0.0375778936,
0.0181546658,
-0.0328923464,
0.0274733212,
0.0579099096,
0.0307999607,
0.0086975489,
-0.0496024415,
0.0581905581,
0.0158192292,
-0.0252554994,
-0.0104572643,
0.0176455583,
-0.0220862199,
-0.004742987,
-0.0435194559,
0.0020619405,
-0.0015040416,
0.0885991752,
-0.0125207584,
0.0055238493,
0.1369555593,
0.0164905638,
-0.0020191644,
0.001301488,
-0.0065977531,
0.0124358675,
0.0265713595,
-0.0029597813,
-0.0631434023,
-0.0354618989,
-0.0061945859,
0.0355436057,
0.0175452102,
-0.0259629525,
-0.0165129714,
-0.1044958904,
0.0181300361,
0.0218095575,
-0.0354801938,
-0.0356040895,
-0.0220608208,
0.0118014328,
-0.0345002152,
0.0435430557,
0.0098109478,
-0.024760941,
-0.0138643179,
0.0273221787,
-0.0266664047,
-0.0238133762,
-0.00238676,
-0.0294792205,
0.0446082801,
0.0784550086,
0.0322795399,
0.0207864009,
0.0934684947,
0.0382490754,
-0.0170721281,
-0.00070148,
0.0142266983,
0.0591095425,
0.0007943023,
0.0111578368,
-0.0162722617,
0.0528344177,
0.027627565,
0.0047702398,
0.0136428932,
-0.0309848953,
-0.0060333717,
-0.0194750763,
-0.0163808987,
0.0805402398,
-0.0096604479,
0.0786276236,
0.0457110517,
-0.0080971345,
-0.019895155,
-0.0292664804,
-0.0250714999,
-0.0732349157,
-0.0407693088,
-0.0224794857,
-0.0230539385,
0.0000829081,
-0.0828360245,
-0.0414005183,
0.021703925,
-0.0034985221,
-0.0159112625,
-0.0031999275,
0.0296417493,
0.0181764271,
-0.0074214879,
-0.0187512971,
-0.0211409759,
0.0129911099,
-0.0580206551,
-0.0250552148,
0.0798488781,
-0.0253239665,
-0.0152162276,
0.0045321039,
-0.0429018848,
-0.0764040872,
-0.0418219864,
0.0633178055,
0.0206497405,
0.0547892191,
0.0088075306,
-0.0431558266,
0.0180756487,
0.0504250452,
0.009518547,
0.0244105961,
0.0336502902,
0.0154643841,
-0.0366538577,
0.0358421803,
-0.0277116187,
0.0609064847,
-0.0199769698,
0.0568034723,
-0.0710540488,
-0.0781821162,
-0.0083320886,
0.0330410264,
0.0005856495,
0.0019895132,
-0.0064797876,
-0.0445680767,
0.0136224367,
-0.0292714033,
-0.0299694799,
-0.0460592769,
0.0283325352,
-0.0122476919,
-0.0333395191,
0.0333511755,
0.0288976617,
-0.0013538314,
-0.0105892401,
-0.0449575223,
0.0120968716,
-0.0143149216,
-0.0188743509,
-0.0670730919,
-0.0018524149,
-0.050887119,
-0.0015089586,
0.0079285214,
0.0174901374,
0.0140680866,
0.0527646653,
0.0093388464,
0.0089164311,
-0.0097255027,
0.0451967344,
-0.085007295,
-0.0115838684,
-0.0513438731,
0.0140075656,
0.0595263168,
-0.0269925464,
0.0487599447,
0.0166555587,
0.0132820122,
-0.0433218256,
0.039437212,
0.0054375664,
0.0083568944,
-0.0193143804,
0.0267862827,
0.0126587003,
0.017696809,
0.1181701049,
-0.0244223382,
0.0163444765,
-0.0176390149,
0.0143632879,
-0.0502957553,
-0.0662073866,
0.020576939,
0.060612563,
-0.0220738277,
0.0291614532,
-0.0268328171,
-0.0055236416,
0.0156323891,
-0.0074690836,
0.021538863,
-0.0223285016,
0.0528387986,
-0.0106513351,
-0.0143496478,
0.0190616976,
-0.0402914137,
-0.0425534211,
-0.0156789459,
0.0209920313,
-0.013109725,
-6.318267442e-33,
-0.069471404,
-0.0193584487,
-0.0135614937,
0.0153339943,
-0.0412897021,
0.004688106,
-0.0585092641,
-0.001287501,
-0.0076145809,
-0.0126212304,
-0.0178993102,
0.0078119896,
0.0112315835,
0.0514417887,
-0.0219527893,
-0.0034705552,
0.0582320206,
0.0013643735,
-0.050914567,
0.0579617471,
0.0185743608,
0.0054850993,
0.0000984394,
-0.0430683978,
0.009271767,
-0.0363872088,
-0.0436404087,
0.0233609695,
0.0191382989,
0.0357915051,
0.000204849,
0.0149658676,
0.0177217014,
-0.0092135193,
-0.0569783412,
0.0084190881,
-0.0763817951,
-0.0153602762,
0.0518547557,
-0.0351514034,
-0.0316935405,
-0.0315049253,
0.0124551458,
0.0040099346,
0.0179365892,
-0.0493676476,
0.0294918418,
-0.0475093052,
-0.0290262513,
-0.033514075,
-0.036468666,
0.0232883543,
-0.0410124063,
-0.0750146508,
-0.0008278489,
0.0409252644,
0.0058171935,
-0.0493197739,
-0.0004908902,
0.0456458963,
0.0116599239,
0.0568898953,
0.0211306214,
0.0616190135,
-0.0045139375,
0.0488315672,
0.0172614586,
0.0168958437,
-0.010116606,
-0.0387723632,
-0.026836304,
0.0367471389,
0.0003311634,
0.0452894047,
-0.0200811159,
0.0296921283,
-0.0629188046,
0.0623122975,
0.0454086363,
0.005403568,
0.0307115484,
-0.0021252776,
0.025732765,
-0.0382050388,
-0.0385159515,
0.0244668443,
-0.0057231616,
0.0152613586,
-0.0264516249,
0.0492731109,
-0.0592426658,
-0.0344660468,
0.0307748541,
-0.0526593477,
0.047444772,
0.0145701002,
0.0056923418,
-0.004197761,
-0.0150334928,
-0.0215682518,
-0.0541892536,
-0.0135971569,
0.0439176075,
-0.0147811286,
-0.0231712647,
0.0271726474,
0.0400282852,
-0.0079220282,
-0.0343655571,
0.0150082717,
0.0144564649,
0.0466470234,
-0.0416467264,
0.0699369237,
-0.0343822576,
-0.0060766609,
0.0123526454,
0.0449321643,
-0.0221819431,
0.0194568243,
-0.0194820967,
0.0067740595,
0.0178263877,
0.0306501221,
-0.04391459,
0.0471946932,
-0.0010157645,
0.0082635386,
-0.0032540665,
-0.0233773291,
-0.0363208018,
-0.0182707906,
2.621e-7,
0.02889877,
0.0311126467,
0.0670511499,
0.0046872487,
-0.0048470148,
0.039594423,
-0.0019150295,
0.0310692135,
0.0469299555,
0.002597414,
0.0435206741,
0.0516425222,
0.0378766768,
-0.0244174153,
-0.0988249704,
-0.018637618,
-0.0685494766,
-0.0318496674,
-0.0006150466,
-0.0223713163,
0.0226281732,
0.0777476653,
-0.0323144831,
0.01909055,
-0.0225698669,
-0.0426983573,
0.0315431952,
-0.0004533509,
-0.0312073901,
0.0716242418,
0.0392117426,
0.0232411399,
-0.0248101335,
-0.0024390277,
0.005598668,
0.030629497,
0.0190623589,
0.0670434684,
0.0140282651,
-0.0346818492,
0.0011088373,
-0.0057642153,
0.0239870772,
0.0140021807,
0.0504746661,
-0.0723687559,
0.0020048514,
-0.0080229668,
-0.0299154948,
0.0058381017,
-0.0022261473,
-0.0184266102,
-0.0295549221,
0.0523919724,
-0.0085394746,
-0.0054112757,
0.0121793002,
-0.0825621784,
-0.0230198894,
-0.013238891,
-0.0225858409,
-0.0474948958,
0.0188945085,
-0.0031768491,
-0.0173991732,
-0.0222410485,
0.0058450233,
2.390144373e-34,
0.090159148,
0.0034395203,
-0.0217739362,
0.023046935,
0.0362028629,
-0.0393350609,
-0.0434204936,
-0.0163001511,
0.0280155316,
-0.0776630938,
-0.0376661532
] |
http://arxiv.org/abs/2402.04268v1 | http://arxiv.org/pdf/2402.04268v1 | ProtAgents: Protein discovery via large language model multi-agent collaborations combining physics and machine learning | 2024-01-27T20:19:49 | 2024-01-27T20:19:49 | A. Ghafarollahi, M. J. Buehler | cond-mat.soft | cond-mat.soft, cs.AI, cs.CL, q-bio.BM | Designing de novo proteins beyond those found in nature holds significant
promise for advancements in both scientific and engineering applications.
Current methodologies for protein design often rely on AI-based models, such as
surrogate models that address end-to-end problems by linking protein structure
to material properties or vice versa. However, these models frequently focus on
specific material objectives or structural properties, limiting their
flexibility when incorporating out-of-domain knowledge into the design process
or comprehensive data analysis is required. In this study, we introduce
ProtAgents, a platform for de novo protein design based on Large Language
Models (LLMs), where multiple AI agents with distinct capabilities
collaboratively address complex tasks within a dynamic environment. The
versatility in agent development allows for expertise in diverse domains,
including knowledge retrieval, protein structure analysis, physics-based
simulations, and results analysis. The dynamic collaboration between agents,
empowered by LLMs, provides a versatile approach to tackling protein design and
analysis problems, as demonstrated through diverse examples in this study. The
problems of interest encompass designing new proteins, analyzing protein
structures and obtaining new first-principles data -- natural vibrational
frequencies -- via physics simulations. The concerted effort of the system
allows for powerful automated and synergistic design of de novo proteins with
targeted mechanical properties. The flexibility in designing the agents, on one
hand, and their capacity in autonomous collaboration through the dynamic
LLM-based multi-agent environment on the other hand, unleashes great potentials
of LLMs in addressing multi-objective materials problems and opens up new
avenues for autonomous materials discovery and design. | ProtAgents: Protein discovery via large language model multi-agent collaborations combining physics and machine learning: Designing de novo proteins beyond those found in nature holds significant
promise for advancements in both scientific and engineering applications.
Current methodologies for protein design often rely on AI-based models, such as
surrogate models that address end-to-end problems by linking protein structure
to material properties or vice versa. However, these models frequently focus on
specific material objectives or structural properties, limiting their
flexibility when incorporating out-of-domain knowledge into the design process
or comprehensive data analysis is required. In this study, we introduce
ProtAgents, a platform for de novo protein design based on Large Language
Models (LLMs), where multiple AI agents with distinct capabilities
collaboratively address complex tasks within a dynamic environment. The
versatility in agent development allows for expertise in diverse domains,
including knowledge retrieval, protein structure analysis, physics-based
simulations, and results analysis. The dynamic collaboration between agents,
empowered by LLMs, provides a versatile approach to tackling protein design and
analysis problems, as demonstrated through diverse examples in this study. The
problems of interest encompass designing new proteins, analyzing protein
structures and obtaining new first-principles data -- natural vibrational
frequencies -- via physics simulations. The concerted effort of the system
allows for powerful automated and synergistic design of de novo proteins with
targeted mechanical properties. The flexibility in designing the agents, on one
hand, and their capacity in autonomous collaboration through the dynamic
LLM-based multi-agent environment on the other hand, unleashes great potentials
of LLMs in addressing multi-objective materials problems and opens up new
avenues for autonomous materials discovery and design. | [
0.0127578229,
0.057824403,
-0.0282458812,
0.0384994037,
-0.0422641113,
-0.0321128555,
0.0272584669,
-0.0107474476,
0.0064576073,
0.0211335849,
-0.0211965255,
-0.0049873386,
-0.0476922169,
0.0615448914,
-0.0192836504,
0.020621784,
-0.0036257282,
-0.0462262183,
0.0258327294,
-0.0471001752,
-0.0257921349,
-0.031840805,
0.0044498304,
0.0226315279,
0.0136747016,
-0.0037390243,
-0.0705440417,
-0.0275481213,
0.0156940706,
-0.0072649377,
-0.031411279,
0.0495848991,
0.0199894626,
0.0408507138,
0.0000020945,
-0.0445777737,
0.0092631439,
0.0594958626,
-0.0068711638,
0.0120156072,
0.0191156287,
-0.0143080121,
-0.0456260517,
-0.0259956103,
-0.0136693073,
0.0256361812,
0.0475413539,
-0.0579569526,
-0.0221347343,
-0.0169541836,
0.001239722,
-0.0093016876,
0.0461706258,
0.0143617755,
0.0487543084,
-0.0250506233,
0.0247153156,
0.0208474137,
0.0876443461,
0.0102059506,
0.0160774738,
0.0090308534,
0.0143073751,
-0.0491037928,
0.0549112037,
0.0004938074,
-0.0359188654,
-0.045349028,
-0.0234312285,
0.0922936574,
0.0190427806,
-0.031237701,
-0.0690845326,
-0.0066044512,
-0.021226326,
0.0422124751,
-0.0198730323,
0.0047502974,
0.0323228873,
0.0141735673,
0.0070636352,
-0.0051231286,
0.0321268551,
0.0020118342,
-0.0013399405,
0.0181451682,
0.0118743489,
-0.0093417335,
-0.0263392739,
-0.0236554816,
-0.0389620028,
0.0057379929,
0.0166137554,
0.016297143,
0.0049350662,
-0.0216632225,
0.0422855094,
-0.0507572964,
0.0410244502,
-0.0322879516,
0.0325582996,
0.0144686196,
0.0415233634,
0.0374639407,
-0.0184534732,
-0.0196542293,
-0.0353806354,
-0.0264764782,
-0.0443656594,
-0.0062387045,
-0.092712149,
0.0077219326,
-0.0222128462,
0.0089339092,
-0.0281499773,
-0.0385885723,
0.0080296295,
0.0709939599,
-0.0189110618,
-0.0083843749,
-0.025547009,
0.0239544809,
-0.0060890955,
-0.0110728061,
-0.0357264653,
-0.0018451476,
-0.0275816396,
0.0218196977,
0.0126665561,
-0.0580225252,
-0.0173029769,
-0.004264941,
-0.0291544795,
0.0208289996,
0.0103169475,
0.1461051553,
-0.0177028757,
0.0115702236,
0.0470289364,
-0.0591337942,
0.0145635558,
-0.0235819705,
-0.020312462,
-0.0142325005,
0.0580786057,
0.0144813638,
-0.0214349572,
-0.0513696447,
0.0211249888,
-0.0530556776,
-0.0010569381,
0.0764988512,
-0.0127571514,
0.0034110383,
0.0330285095,
0.0018965462,
-0.0237381347,
0.0218347274,
0.0065018805,
0.0160771329,
-0.0295530967,
0.0263474416,
0.0003001646,
0.0056393384,
-0.019300798,
-0.0303151812,
-0.029032873,
-0.0193543639,
0.0524380207,
0.059603747,
-0.0528856628,
-0.0264879912,
0.000907511,
0.0215341095,
0.0521651953,
0.0531500541,
-0.0042040106,
0.0758230314,
0.0072884285,
0.0458505601,
0.0227003787,
-0.0445287116,
0.0097548347,
0.032999862,
-0.0291279964,
0.0046093618,
0.0243178047,
0.0287161935,
0.0056248941,
-0.0362448655,
0.0106781954,
-0.0277681891,
0.028698124,
-0.0413471237,
0.0612470433,
-0.0102792354,
-0.0470552929,
-0.0392901264,
0.0097439466,
-0.0445932336,
-0.016722871,
-0.0245770179,
0.045051191,
0.0613432787,
0.060142301,
-0.0879916847,
-0.0418433174,
0.0604158044,
-0.0275738481,
0.0554541536,
0.0029755498,
-0.0130749568,
-0.0260334536,
0.0037893464,
-0.0283668302,
-0.0099361986,
0.0350832418,
0.0072372882,
-0.0327266902,
-0.0333312266,
-0.0168487485,
-0.0057491935,
-0.0610897467,
0.003501392,
-0.0241273157,
-0.0252033044,
0.0435048677,
0.0092544146,
0.064586103,
0.0057778927,
0.0209664013,
0.0174260065,
0.1049364507,
-0.0734020397,
-0.0210560001,
-0.0407902747,
0.0612693876,
0.0301559344,
0.0255956855,
-0.0101440158,
0.0366600454,
-0.0498247072,
-0.0090944739,
-0.0405218117,
0.0980667919,
0.0400061086,
-0.012554286,
-0.0126163978,
-0.0132632256,
-0.0326320678,
0.041706048,
-0.0359092057,
0.0481157936,
0.0553404912,
-0.0487160794,
0.0296053234,
0.0242008138,
0.0173296947,
-0.0294732843,
0.018879991,
0.0408358052,
0.0574664548,
0.0320481956,
0.0737507641,
-0.0200470164,
0.0246567298,
-0.061888922,
0.0188250989,
-0.0028488573,
0.021994384,
-0.0055396007,
0.0104635721,
-0.0286414847,
-0.0182052739,
0.0357802063,
0.0172791686,
0.0376944207,
-0.0634887591,
-0.050479155,
0.0440180227,
0.009101307,
-0.0845535472,
-0.0311232321,
-0.0207346212,
-0.054988347,
0.0534139983,
0.0583912618,
0.0194726493,
-0.0150866155,
0.0040550497,
0.0015692223,
0.040457014,
0.0108636729,
-0.0542132333,
0.0260399636,
-0.0324596502,
0.0135693224,
0.0251841582,
-0.0446336977,
0.0331342705,
0.092341952,
0.0055783829,
-0.009725905,
0.0139141753,
-0.0342599414,
0.0207720716,
0.0262376629,
0.0086507061,
0.0184299983,
0.0531144962,
-0.0129516292,
-0.0043216231,
-0.0144311618,
0.0509920083,
0.0343043469,
-0.074921988,
0.0235105287,
-0.0075638955,
0.0303178746,
0.021145666,
0.0415081233,
0.0101624271,
0.0875683278,
-0.0112573113,
0.0295751411,
-0.054290574,
-0.0383049287,
-0.0072855619,
0.0483581163,
-0.0076395618,
0.0081959767,
0.0304690339,
0.0684014633,
0.0243504699,
-0.0014004686,
0.0574145205,
0.0032085264,
-0.0179538913,
-0.031508036,
-0.0028870627,
0.0304277446,
-0.0477293059,
0.0022399051,
0.0614415854,
-0.0345960297,
0.0505191684,
-0.0191701762,
-0.0318441279,
-0.046640601,
0.0179508254,
0.0376079828,
-0.0136106219,
-0.020057356,
0.023139542,
-0.0778470486,
-0.0055887867,
-0.0174643137,
-0.026543105,
-0.0273500141,
-0.0063890247,
0.0772949532,
0.0158244651,
0.0105434572,
-0.0024767441,
-0.0659584254,
0.0027033079,
0.0121176234,
0.0180333219,
0.010188763,
0.0078877816,
-0.0265098531,
-0.0405521654,
-0.0173527729,
0.0284259561,
0.0303275418,
-0.0667317286,
0.0163163822,
0.0214200225,
0.0125900414,
-0.0284346007,
0.0267828126,
-0.0477275215,
-0.0368378162,
-0.0059361164,
-0.0093787322,
-0.0192617383,
-0.0158081073,
-0.0067461478,
-0.0273305383,
0.0761984959,
0.0005065392,
-0.0129595101,
-0.00755976,
-0.0576137826,
-0.0254657306,
-0.0416017361,
0.0172021929,
0.0529577099,
0.0236583278,
0.0107111707,
-0.0174547844,
0.0146723986,
-0.0706176385,
0.0260152146,
0.0009886109,
0.0010605251,
0.0806203112,
0.0423561446,
-0.0336591341,
-0.0309592616,
0.0001896616,
0.0399182178,
0.0463908017,
-0.0402678251,
-0.0181815438,
0.0722894818,
-0.0181186981,
-0.0249963384,
0.0476996712,
-0.044474747,
-0.0456866473,
-0.006726942,
0.0957292169,
-0.0048481212,
0.0226624478,
-0.0618603155,
-0.0583698601,
-0.0078226645,
0.0733385682,
0.0000607733,
-0.078656815,
-0.0792774111,
-0.0109018162,
-0.0188570246,
-0.0186315402,
0.0084940344,
-0.0411477573,
-0.0033857347,
-0.0084338682,
-0.0117523735,
0.0096539529,
0.0363292024,
-0.0126776332,
0.0280643255,
0.0066501931,
0.0434861071,
-0.0175267477,
-0.0521622561,
-0.0286855102,
0.010339573,
0.0053569921,
0.0056540873,
-0.0400543921,
0.0420884788,
-0.0544158705,
-0.0027749692,
0.0362279788,
0.0174101293,
-0.0496390313,
-0.0206832513,
0.0511919744,
0.0248458013,
0.0049419147,
-0.041815795,
-0.044967562,
0.0350864679,
0.0469673723,
-0.029563278,
-0.028692279,
-0.0105240215,
-0.0178299565,
-0.0280717146,
-0.0044840882,
-0.0057650814,
0.0168334283,
-0.008671714,
0.0060198982,
0.0119467704,
-0.0691710934,
0.0318068527,
0.0386207402,
0.0128813805,
0.0140631665,
-0.0191837698,
-0.0262913797,
-0.006099002,
0.0120586567,
0.0137565369,
-0.0462995954,
-0.0064409873,
0.0410637259,
-0.0369883664,
0.0447021089,
-0.0145169143,
0.0243063103,
0.0101226699,
0.0199986249,
0.0862382948,
0.0212524962,
0.0444890447,
0.0626674667,
-0.0104466313,
-0.0032816462,
0.0461546741,
-0.0721772835,
0.0472938716,
-0.0073084589,
0.0010969695,
0.0014619981,
0.0437607095,
-0.0446868464,
0.0181282144,
-0.0326169021,
-0.0584163442,
-0.0284836031,
0.0160998348,
0.0449692458,
0.0573971048,
0.0188960545,
-0.0285631251,
-0.0320171453,
-0.0198887475,
0.0337768942,
0.0119745824,
-0.0169880372,
0.0226313435,
0.0201777127,
-0.0268158689,
-0.0651988834,
0.0620714277,
0.0169039071,
0.0130372401,
-0.019190412,
-0.0383585989,
-0.0240597781,
0.0116358669,
-0.0306032132,
-0.0054769116,
-0.038829565,
-0.0050496035,
-0.046708487,
-0.0017386066,
0.0088383593,
0.0574322902,
-0.0247435011,
0.0041513867,
-0.0032557843,
0.0563517213,
-0.0179046206,
-0.0799241066,
0.0144166676,
-0.0184953958,
-0.0226236433,
-0.0387406312,
0.0063238735,
-5.60097234e-33,
0.0202586669,
-0.0035511691,
-0.0044717202,
0.0431881994,
0.0087640285,
-0.0199219324,
-0.0590572618,
-0.045576226,
0.0233084615,
0.052900821,
-0.0016013609,
-0.0417437032,
0.0018781868,
-0.0620554797,
0.0648544356,
-0.0044762846,
0.0164360292,
0.0071331542,
0.0287989639,
-0.009085576,
-0.0028927943,
0.0197597705,
-0.0110393325,
-0.0090595484,
0.0154396808,
0.0174896866,
-0.0300374217,
-0.0195448846,
-0.0525660813,
0.0456843078,
-0.0326910578,
0.0404944122,
-0.0714853182,
-0.0157295391,
-0.0356342606,
-0.0234969165,
-0.0313395411,
0.0043592853,
-0.0131293237,
-0.0183175057,
0.0115883742,
-0.0135451667,
-0.0138410712,
0.0268950351,
-0.092039898,
0.0077358885,
0.033430554,
-0.0802420154,
0.0004877481,
0.0098657347,
-0.0783215612,
-0.0146038625,
0.027559327,
0.0120198326,
0.0641549751,
-0.0336616635,
-0.0149837276,
-0.0206470396,
-0.0034841835,
0.0685057342,
0.0304105505,
-0.0434796028,
0.0357224941,
0.0066263461,
0.038644392,
-0.0005463854,
0.0642082468,
0.0173020214,
-0.006537341,
-0.0087831113,
0.0062284726,
0.0786212161,
0.0543569326,
0.0390718617,
0.0048749861,
0.0280680917,
-0.1225351244,
0.0024124319,
0.0300614238,
-0.0244254973,
0.0165501703,
0.0605028421,
0.0159589406,
-0.0310826544,
-0.0368332379,
0.0250702649,
-0.0005053791,
-0.0417921096,
-0.0445492119,
-0.0309798848,
-0.0319840237,
-0.0349581651,
-0.0022704261,
-0.0458236597,
-0.0074953591,
0.0469742268,
-0.0135268662,
0.0096181398,
-0.0327530056,
-0.0169829037,
-0.0229581054,
-0.0463611372,
-0.0171698406,
-0.005812719,
0.0074188891,
-0.0074028471,
-0.05473179,
0.0481960438,
0.0364237018,
0.0290843118,
-0.0551174544,
-0.009566389,
0.0567276217,
0.0192980077,
0.0270215534,
-0.03246589,
-0.0066671753,
0.0085644908,
-0.0094881915,
0.056015633,
0.0301857572,
0.086711511,
0.0072377352,
0.0211440865,
-0.068482399,
-0.0530598536,
-0.0473127253,
-0.0023641125,
0.0332859121,
-0.056839671,
-0.0288522504,
0.009110529,
2.854e-7,
0.0380697064,
0.0698504001,
0.0440124273,
0.0262987409,
-0.0174357891,
0.0597954951,
-0.0297104847,
-0.0324474052,
-0.010618885,
-0.0227014776,
0.0576998852,
0.0366463959,
0.0110178674,
-0.0096569955,
-0.0548135862,
0.0046131704,
-0.0025806737,
-0.0033979174,
-0.0739543661,
0.0007653396,
-0.0563930348,
0.0211315081,
0.0259648841,
-0.0098056989,
0.0214171018,
-0.0510613061,
-0.0056892489,
-0.0324931107,
0.0399065353,
0.0529474616,
-0.0568922721,
-0.0512577333,
-0.0254457891,
0.0476324558,
-0.0139012793,
-0.017456362,
-0.0061470745,
0.0275133681,
0.0003555609,
0.0256824344,
0.0454127938,
-0.0058272257,
0.0081272637,
0.0140779214,
-0.0280195046,
-0.0410910062,
-0.0547553897,
0.0296836514,
0.008433382,
-0.0092167007,
0.0128877219,
-0.0259610172,
0.0796766803,
-0.0115710674,
-0.0365346558,
0.0088321548,
0.0185770951,
0.0175236221,
-0.0099440692,
-0.0044053881,
-0.0672025383,
0.0327332541,
-0.0414707623,
-0.020327609,
0.0510335788,
-0.0155928405,
0.0377023853,
2.600719492e-34,
0.084628731,
-0.0143099427,
0.0009456757,
0.0500465482,
0.000425852,
0.020883685,
-0.0156274624,
0.0016630862,
-0.0069461851,
-0.055074919,
-0.0334003083
] |
http://arxiv.org/abs/2401.15479v3 | http://arxiv.org/pdf/2401.15479v3 | Navigating the Post-API Dilemma | Search Engine Results Pages Present a Biased View of Social Media Data | 2024-01-27T19:04:30 | 2024-04-02T15:28:07 | Amrit Poudel, Tim Weninger | cs.IR | cs.IR, cs.CL, cs.SI | Recent decisions to discontinue access to social media APIs are having
detrimental effects on Internet research and the field of computational social
science as a whole. This lack of access to data has been dubbed the Post-API
era of Internet research. Fortunately, popular search engines have the means to
crawl, capture, and surface social media data on their Search Engine Results
Pages (SERP) if provided the proper search query, and may provide a solution to
this dilemma. In the present work we ask: does SERP provide a complete and
unbiased sample of social media data? Is SERP a viable alternative to direct
API-access? To answer these questions, we perform a comparative analysis
between (Google) SERP results and nonsampled data from Reddit and Twitter/X. We
find that SERP results are highly biased in favor of popular posts; against
political, pornographic, and vulgar posts; are more positive in their
sentiment; and have large topical gaps. Overall, we conclude that SERP is not a
viable alternative to social media API access. | Navigating the Post-API Dilemma | Search Engine Results Pages Present a Biased View of Social Media Data: Recent decisions to discontinue access to social media APIs are having
detrimental effects on Internet research and the field of computational social
science as a whole. This lack of access to data has been dubbed the Post-API
era of Internet research. Fortunately, popular search engines have the means to
crawl, capture, and surface social media data on their Search Engine Results
Pages (SERP) if provided the proper search query, and may provide a solution to
this dilemma. In the present work we ask: does SERP provide a complete and
unbiased sample of social media data? Is SERP a viable alternative to direct
API-access? To answer these questions, we perform a comparative analysis
between (Google) SERP results and nonsampled data from Reddit and Twitter/X. We
find that SERP results are highly biased in favor of popular posts; against
political, pornographic, and vulgar posts; are more positive in their
sentiment; and have large topical gaps. Overall, we conclude that SERP is not a
viable alternative to social media API access. | [
0.0923582539,
0.1007148102,
-0.0453788415,
-0.0190785211,
-0.0252385903,
-0.0281520635,
0.0035310083,
0.0161353741,
0.0505726188,
-0.0723005757,
0.0340525284,
0.0168480407,
-0.0085632326,
0.0442667045,
-0.0166452657,
-0.0171885565,
-0.0050426736,
0.090007104,
0.0302408636,
-0.014607273,
0.0263056215,
0.0251815002,
0.0096209086,
0.0105883591,
0.0222927276,
-0.0172928125,
0.0495476648,
0.0045682997,
-0.0275919233,
-0.0401622467,
0.0204038303,
0.0384142511,
-0.0069886097,
0.0277283471,
0.0000017103,
-0.0261399802,
0.0311233383,
-0.0005086042,
0.0256540161,
-0.0320149809,
0.0892507285,
0.0100444406,
0.0494656526,
0.0478494316,
-0.0512869768,
-0.0013309985,
0.0626377016,
-0.031907782,
0.026011074,
0.020918034,
-0.0129634496,
-0.0289381687,
-0.0624425299,
0.0151542723,
0.0100732753,
0.0268137436,
0.0060192491,
-0.0557444133,
0.0149767539,
-0.000556979,
0.0071019847,
0.0486843586,
-0.0267692003,
-0.0052967449,
0.0034976001,
0.0094288662,
-0.0932370871,
-0.0661799461,
0.0265661925,
0.0329388268,
0.0601898171,
-0.0190743618,
0.0065953936,
-0.0007470954,
0.008513039,
0.0343270712,
-0.0728069395,
0.0354603492,
-0.0140871573,
0.0353318416,
-0.0232888684,
-0.0053605363,
0.0275378022,
0.0277213641,
0.0350404717,
0.0191353392,
-0.0038345132,
-0.0432481132,
0.0284771044,
0.0283028986,
0.0571112484,
0.0012427884,
0.0379324742,
-0.0201308709,
0.0129343644,
0.0222495291,
-0.0011576907,
-0.0537512377,
0.0512220263,
-0.0586492829,
-0.0353360474,
-0.0018578316,
-0.0331672318,
0.0336522721,
0.0707962811,
-0.0564303361,
-0.0293742865,
0.0228377469,
0.0471364222,
0.0013759551,
0.0073522488,
0.0243562534,
0.0588614382,
-0.0075388039,
-0.0113627156,
0.0193582736,
-0.0125985434,
0.0308031701,
-0.0384484567,
-0.025717495,
-0.0394037366,
0.0333478972,
-0.0169254933,
0.0218236446,
-0.0242653228,
-0.0158471894,
-0.007482673,
0.0394415408,
-0.0020363231,
-0.0358346663,
0.0295252148,
-0.0348473117,
0.0883652046,
-0.00242595,
0.0100415684,
0.0319972672,
0.044344414,
-0.0159791335,
-0.0485926904,
-0.0390977338,
0.0299883578,
0.0130084706,
0.0290542021,
-0.0289698727,
0.0692482814,
0.0565480813,
-0.0081087239,
0.0189022124,
0.0014177052,
0.0342977531,
-0.0687195584,
0.0021214695,
-0.0188798234,
0.0322761051,
0.0192385968,
0.0216735099,
-0.0163943395,
0.1044668853,
-0.0536239035,
0.0542390235,
0.0977261439,
0.0127680693,
-0.0123979691,
0.0146840019,
0.0391655117,
0.0037266479,
0.0084393313,
0.0086916611,
-0.0340346061,
-0.0437791273,
-0.0431918092,
0.0225197542,
0.0102593563,
0.0103268744,
0.0359827913,
0.0355790444,
-0.0278990027,
0.0341035649,
-0.0163788032,
0.0143349506,
0.0120096086,
-0.0205853097,
-0.0545055568,
0.0030827497,
0.0079538524,
-0.0238967873,
0.0805073977,
0.0302414522,
0.0043187891,
-0.0009896304,
-0.015993135,
0.0141691156,
0.0114576221,
-0.0036678375,
0.0185251124,
-0.0414434038,
-0.005099583,
-0.0206776205,
0.0286526792,
-0.0114015453,
0.017179681,
-0.0005327562,
0.0152203599,
0.0170410778,
0.0031392814,
-0.0355365388,
0.0455509834,
-0.0217280127,
0.0231274404,
0.039654132,
0.010260338,
0.0116039533,
-0.0069687977,
0.0322264545,
-0.0286139697,
0.0320542566,
0.0372922234,
-0.0440355018,
0.0492605455,
0.0467255339,
-0.0010101997,
-0.022117326,
-0.0331595205,
-0.0087362295,
0.0427225083,
0.0280469395,
-0.0537390038,
0.0410551839,
0.0288569015,
0.0648405179,
-0.0035635745,
-0.0160395578,
0.0060246023,
0.0108522531,
-0.029249737,
-0.0431502759,
0.0418638252,
0.0772108063,
0.0004900053,
-0.0720745325,
-0.0371693112,
0.1247288063,
-0.0207312014,
0.0494260415,
0.0667738393,
-0.0240463745,
0.0577241108,
0.0444260277,
-0.0509462841,
-0.0502075255,
0.0089184362,
0.0114801275,
0.0071778037,
-0.027804574,
0.0181622691,
-0.016731469,
-0.0757458061,
-0.016724566,
0.0180439614,
-0.0198459532,
-0.0149577623,
-0.0235115848,
0.0350603051,
0.0455495454,
-0.0642354861,
-0.0855279863,
0.0245809704,
-0.0768626556,
-0.0211054906,
-0.0503555201,
0.00540864,
0.0235861596,
0.0108831478,
0.0021475086,
0.0070479498,
-0.0111777987,
-0.0360927097,
0.0002191562,
0.0567425825,
0.007676702,
-0.0036621785,
0.0176131688,
-0.0477375723,
-0.0466919467,
-0.0162298158,
0.009902752,
0.0395018198,
-0.0050254795,
0.059179049,
0.0052718935,
-0.0648367554,
-0.0051095565,
-0.0141048245,
-0.0165719055,
0.0036460138,
-0.018043125,
-0.0625774562,
0.0124963047,
0.0518926866,
0.0350532942,
-0.0018314001,
0.02804262,
-0.0185139589,
-0.0333172046,
-0.0574101396,
-0.0195058491,
0.0119498968,
-0.0847819299,
0.0537591241,
-0.0538054816,
0.0205726847,
0.0321477875,
-0.0301297233,
0.0358573683,
-0.0166971032,
-0.0261947177,
-0.0139074381,
0.0324166007,
-0.0056556631,
0.037252944,
-0.0162964836,
-0.0442266129,
-0.0011221336,
0.0516062193,
0.0545961373,
0.0000643166,
-0.0255567096,
0.0083312346,
0.0499184653,
0.0038065193,
0.0063537732,
-0.0354803242,
0.0103164725,
0.0298624616,
-0.0062398673,
0.0347846784,
-0.0679902509,
0.0423188321,
0.0302585028,
0.0017792692,
-0.0133775566,
-0.0288598705,
0.0017665315,
-0.1162785888,
-0.0559209362,
-0.0356344059,
0.0066216546,
-0.0352099389,
-0.0096111884,
-0.0228882022,
0.0067860219,
-0.0235289093,
-0.004952583,
-0.0445717424,
0.0489899032,
-0.0202718191,
-0.004158563,
-0.0197243039,
-0.002355499,
0.0730569512,
-0.0351227745,
0.0227865484,
0.0549275801,
-0.0210071709,
0.0008159298,
-0.0281667467,
0.0084585808,
-0.0322823375,
-0.0267123617,
-0.0259094592,
-0.0073485281,
0.0030553162,
0.0533236675,
0.0203826092,
0.0095000966,
-0.042787537,
-0.0182362348,
0.00254813,
-0.0807858929,
0.0305865984,
-0.0001033459,
-0.0556480885,
0.0195169412,
-0.0137442863,
0.0370184295,
-0.0627624467,
-0.051479958,
-0.0212934799,
-0.1164995059,
0.0312475283,
-0.0306210481,
0.0094100051,
-0.0580247641,
-0.1145570502,
0.0216829926,
-0.0393889286,
-0.0250507239,
-0.0180174783,
-0.0319860913,
0.0027502717,
-0.0076192264,
0.0173116196,
-0.0395590998,
-0.0051744999,
-0.0640283301,
0.0009587821,
0.005039081,
0.0478524715,
-0.0351988748,
0.0094701909,
0.0108805345,
0.0332516581,
-0.0150496596,
-0.0384621099,
0.076340735,
-0.0360444449,
0.0062306197,
-0.0510713123,
0.0085154362,
-0.0224446207,
0.0232535396,
-0.0292321611,
0.0105822785,
0.0054622954,
-0.0252467878,
-0.0723020658,
-0.0009350572,
0.0175981596,
0.0212933347,
0.0295505263,
-0.0142894899,
0.0182089116,
-0.0488636456,
0.0335212126,
0.0000724234,
-0.0224972274,
-0.0080087716,
-0.0503179245,
-0.0073742042,
-0.0047220001,
-0.0153051261,
-0.017253425,
-0.0101082372,
0.0035784701,
-0.0187274162,
0.0343806557,
-0.0226011295,
-0.0532838516,
-0.0138567109,
-0.0061793467,
-0.0184249803,
0.0311982576,
-0.0464112386,
0.0074760839,
0.0167638175,
-0.0070313052,
-0.0066344654,
0.0040133246,
-0.0296641532,
0.0638411343,
-0.0044106701,
0.0205316767,
0.0084048398,
-0.0058655799,
-0.0339854173,
0.0290263947,
0.1360768229,
-0.0060814163,
-0.0191085916,
0.0118175121,
0.0012957919,
0.0294034835,
0.0046271388,
-0.0821580961,
-0.0593751706,
-0.0023721159,
-0.059270326,
0.0038429333,
0.0176616386,
-0.0074965875,
0.0584491454,
0.0703478456,
0.0161693897,
0.0152616743,
-0.0295913387,
-0.0541479923,
-0.0077988156,
0.0115476558,
0.0057115164,
-0.0057036746,
-0.0412305817,
0.0225527119,
-0.0524108149,
0.0657672063,
-0.0365895368,
0.060049504,
-0.126654014,
0.0184763242,
-0.0307584796,
0.0043973275,
-0.0011458544,
0.0128338421,
-0.0336783007,
-0.0161745399,
0.0325431451,
0.0211253967,
-0.0114823049,
-0.0079042884,
-0.0022367497,
-0.0004510362,
0.026957171,
0.0111371651,
-0.0533667915,
-0.0647617355,
-0.0018519155,
-0.0261489339,
-0.010093268,
-0.0475228801,
0.0031338769,
-0.0166027918,
-0.0190018658,
0.0257541649,
-0.0071396651,
0.0082940208,
-0.0031105953,
0.0069283014,
-0.0099604717,
-0.0265263394,
0.0414672606,
0.0267626531,
-0.0563138463,
0.0068575325,
0.0119251544,
-0.0510075279,
0.0263974033,
-0.0114863086,
0.0353761502,
0.0417869128,
0.0017535915,
-0.013642366,
-0.0320018791,
-0.0427935086,
-0.000582005,
-0.0421721376,
0.0398417227,
-0.0087093422,
-0.0120329428,
0.0670498088,
0.026554333,
-0.000160545,
-0.0492325388,
0.0116795879,
-0.1038533226,
-0.0255419146,
-0.0039380984,
-6.826829645e-33,
0.0622513965,
-0.0078403791,
-0.0483069718,
-0.0113561582,
-0.0315573774,
0.0089082578,
-0.0066608773,
0.0135503085,
-0.0299204513,
-0.0078912294,
-0.0265829898,
0.0180913247,
0.0131584695,
-0.0123990765,
0.0438070036,
-0.0434294157,
0.018633794,
-0.0083393669,
-0.0420832746,
0.008925382,
0.0260200612,
-0.0097524496,
-0.0135573409,
-0.0746274292,
0.0324187949,
-0.0203543603,
0.0136380736,
0.0391269512,
0.0769845024,
0.005523066,
0.0116468798,
0.0814856663,
0.0480069444,
-0.0064281728,
0.0042734719,
0.0472066104,
-0.0418496244,
-0.0186452102,
0.0340606198,
-0.0193170905,
0.0158006586,
0.0409155339,
0.0456958786,
0.0306537915,
-0.0208577104,
-0.0076529826,
0.0304731168,
0.0243242867,
0.035017252,
-0.0135983406,
0.0204563234,
-0.0066466141,
-0.015480645,
0.043271035,
-0.0131333638,
0.049276717,
0.0466757081,
0.1043300629,
-0.0256661419,
0.0118032163,
-0.0312703289,
-0.0211779419,
0.0288274996,
-0.0117601324,
0.0014967442,
0.0651353523,
0.0192251019,
0.0477771685,
-0.003655989,
-0.0347631201,
-0.0056280131,
-0.0093111694,
-0.0422619395,
0.0525031313,
0.0140204541,
-0.0870729312,
-0.0329410806,
-0.0214589816,
0.0175340082,
-0.0227355976,
0.0417066589,
-0.0371837355,
-0.0237892196,
0.011439831,
-0.004361317,
-0.0555411391,
-0.0435961112,
0.0024131888,
-0.0415275693,
0.011415313,
-0.0360065326,
0.0205202494,
0.0010123488,
-0.0288068112,
-0.0275411289,
0.0286247469,
-0.005425835,
0.0056785895,
0.0078945616,
-0.0369344279,
-0.0424467847,
0.024094779,
-0.0489698723,
0.0383089185,
0.0144907767,
-0.0309226792,
-0.0414714962,
0.0319236256,
-0.027936466,
-0.0091443537,
-0.0282683913,
0.0244988967,
-0.0052648862,
-0.0633422583,
0.0109772123,
0.0716469735,
0.0166106094,
0.0346725062,
0.0834536105,
0.1054292172,
0.0122096399,
-0.0066939411,
-0.0193796642,
-0.024964394,
0.0091303028,
-0.0577823818,
-0.0066150632,
0.0247266274,
0.0020463916,
-0.0363851078,
-0.0196370613,
-0.0509777553,
2.673e-7,
0.0516501814,
0.0076889112,
-0.0095570795,
0.0732912496,
0.0329801477,
0.0304697379,
0.0671781898,
0.0493855551,
0.0121671762,
-0.0426297449,
-0.0331843309,
-0.0039188741,
0.0072326981,
0.0451255552,
-0.0243425667,
-0.0199377015,
-0.0402342752,
-0.0254576318,
-0.0279301535,
0.0291038044,
0.0000670755,
0.0213276353,
0.0399226956,
-0.0034379326,
-0.0414197333,
-0.0204778109,
0.0025582016,
-0.0313761234,
-0.0156592205,
0.009257054,
0.020348439,
0.0348781347,
0.017406147,
-0.0178526249,
-0.0330272466,
-0.0090609398,
-0.007220096,
-0.0205992851,
0.0320464075,
0.0078150118,
0.0060805306,
-0.0422355309,
0.0333418958,
-0.0404337682,
0.0452044494,
-0.0082100676,
-0.0297731925,
-0.0529376641,
-0.0295419823,
-0.0009253406,
0.0629336536,
-0.0352041908,
-0.0375791751,
0.017645536,
-0.0450883135,
0.01693286,
-0.0272596907,
-0.0655032247,
0.0503539965,
-0.014288445,
-0.0017346386,
-0.0623717122,
0.0203952566,
0.005110295,
0.0371023752,
0.0468172804,
0.060954906,
1.702546873e-34,
0.0299282223,
-0.0171168018,
0.0135046178,
-0.0471232049,
-0.0036059485,
0.0314619169,
0.0004994816,
-0.0488718078,
0.0100696338,
-0.0110864602,
0.0241231155
] |
http://arxiv.org/abs/2401.15476v1 | http://arxiv.org/pdf/2401.15476v1 | To Burst or Not to Burst: Generating and Quantifying Improbable Text | 2024-01-27T18:34:29 | 2024-01-27T18:34:29 | Kuleen Sasse, Samuel Barham, Efsun Sarioglu Kayi, Edward W. Staley | cs.CL | cs.CL | While large language models (LLMs) are extremely capable at text generation,
their outputs are still distinguishable from human-authored text. We explore
this separation across many metrics over text, many sampling techniques, many
types of text data, and across two popular LLMs, LLaMA and Vicuna. Along the
way, we introduce a new metric, recoverability, to highlight differences
between human and machine text; and we propose a new sampling technique, burst
sampling, designed to close this gap. We find that LLaMA and Vicuna have
distinct distributions under many of the metrics, and that this influences our
results: Recoverability separates real from fake text better than any other
metric when using LLaMA. When using Vicuna, burst sampling produces text which
is distributionally closer to real text compared to other sampling techniques. | To Burst or Not to Burst: Generating and Quantifying Improbable Text: While large language models (LLMs) are extremely capable at text generation,
their outputs are still distinguishable from human-authored text. We explore
this separation across many metrics over text, many sampling techniques, many
types of text data, and across two popular LLMs, LLaMA and Vicuna. Along the
way, we introduce a new metric, recoverability, to highlight differences
between human and machine text; and we propose a new sampling technique, burst
sampling, designed to close this gap. We find that LLaMA and Vicuna have
distinct distributions under many of the metrics, and that this influences our
results: Recoverability separates real from fake text better than any other
metric when using LLaMA. When using Vicuna, burst sampling produces text which
is distributionally closer to real text compared to other sampling techniques. | [
0.0301650073,
0.0843670517,
-0.0203526374,
0.026278751,
-0.0257023871,
-0.0062635629,
-0.0146395983,
0.0455927178,
-0.0122421961,
-0.0331464559,
-0.0180369746,
-0.0747728199,
0.0226141382,
0.0597463809,
0.0460385345,
-0.0615051873,
-0.0081593078,
-0.0213094186,
0.0165237393,
-0.0044264449,
0.0428654402,
0.0198945049,
0.0188407544,
0.0226889346,
-0.0041869497,
-0.008267764,
0.0044304794,
0.0018824663,
0.024965385,
0.0172054023,
0.0105369082,
0.0456141233,
-0.0280758142,
0.0609622076,
0.0000019022,
-0.0154883135,
-0.0086185038,
-0.0245412942,
-0.0023640378,
0.0434912294,
0.0556136034,
-0.0343305729,
-0.0323514678,
0.0343861692,
-0.0310228467,
0.0084657995,
0.0165454987,
0.0222024806,
0.0190831032,
0.0641998053,
-0.0102415048,
0.0009206532,
0.006847111,
-0.0286808889,
0.0128922965,
-0.0010767843,
0.0220149606,
-0.0011324388,
0.0296637826,
0.0224575195,
0.0179326143,
0.0291893426,
0.0285972822,
-0.0420510173,
0.0129659176,
0.0488364808,
-0.027546728,
-0.0212785471,
-0.0045897798,
0.045018781,
-0.0358528271,
-0.0391125083,
0.0338743106,
-0.036662925,
-0.0114036109,
0.0095715122,
-0.0324626938,
0.0537599549,
-0.0176617205,
0.0528286025,
-0.0073075057,
0.0133327153,
0.0445220172,
-0.0017522664,
0.0147561803,
0.0149951726,
0.0202146862,
-0.0020974947,
-0.0239895396,
0.0320163332,
-0.0197955407,
-0.0648665875,
0.0387636721,
0.0203076173,
-0.0407666862,
0.0094220433,
-0.1072945371,
-0.0377154686,
0.049882032,
-0.005319783,
0.0770446509,
-0.0096586337,
-0.0078499028,
0.0095430296,
-0.0168988872,
0.0217875633,
-0.0162294768,
0.0042200284,
-0.0338764824,
0.0141448081,
-0.0048646885,
-0.0163761638,
0.0005108824,
-0.0026452674,
-0.0480986908,
-0.0125543326,
-0.0508567728,
0.008071946,
-0.0055571785,
-0.0161046106,
-0.0113729276,
-0.0047261207,
-0.0293407273,
-0.0112138642,
-0.000992479,
-0.0493443348,
-0.0260629561,
0.0341500826,
0.033435408,
-0.0562818348,
-0.0179455634,
0.0330008306,
0.0125983916,
-0.0006575905,
-0.0108492542,
0.0261121839,
-0.035767179,
-0.0230836384,
-0.0496692657,
-0.0484801047,
-0.0268019196,
-0.0585830398,
-0.0163166318,
-0.0880349278,
0.0061980602,
-0.0176076666,
-0.0162469298,
0.0010204593,
0.0182815529,
0.0452136211,
-0.0136258574,
0.0376943573,
-0.0108256415,
0.0265312027,
0.0307943095,
-0.0041566934,
0.0120714149,
0.0408838168,
0.0405243747,
0.0700200871,
0.0457925424,
0.0592376888,
0.0457512476,
0.0782946572,
0.0375661552,
0.0175406616,
-0.0428386368,
-0.059634652,
-0.018074071,
0.0169983786,
0.0747831613,
0.0872436985,
-0.0412357189,
0.0017214812,
0.0692936555,
0.0266000591,
-0.032550402,
0.0035695639,
0.0517466255,
0.0141209243,
0.0359147303,
0.0384040736,
-0.0207366701,
0.0097948248,
-0.0023133766,
0.0316715091,
0.0533471778,
0.0177007318,
0.0004863705,
-0.0205172971,
-0.0068784403,
-0.019799456,
-0.0362610258,
-0.0400903784,
0.0288414564,
0.0317820199,
-0.0261699874,
0.0229323693,
0.0040442585,
-0.040072646,
0.0323457383,
-0.0420987681,
0.0539102592,
0.0605571568,
0.0656024739,
-0.02100165,
0.069685556,
0.0117695602,
-0.0343847945,
0.0754136667,
0.0089702113,
-0.0156505909,
-0.0034459089,
-0.0240558106,
-0.000488872,
0.0003010578,
0.0286716521,
0.0184941329,
0.0292194579,
0.0248586182,
0.025286328,
-0.0501204804,
-0.0425399169,
0.0389233083,
-0.0678425506,
0.0093929488,
0.0212441236,
-0.0380904861,
0.0010817983,
0.009732306,
-0.0398637764,
0.0351169109,
0.0097448546,
-0.0434905998,
-0.0409153998,
-0.0094979471,
-0.0114748497,
-0.0375675224,
0.0159184691,
-0.0392431021,
0.0057586147,
0.0209779795,
-0.0062351571,
-0.0414538793,
-0.007289581,
-0.0496922433,
-0.0139608812,
-0.0766642168,
0.0409961492,
0.0153394705,
0.0466636345,
-0.0306349769,
0.0660473853,
-0.0282539446,
0.017921092,
0.0220272522,
0.0015649837,
0.0108058425,
-0.0358060226,
0.001886835,
-0.0008203586,
0.0063483929,
0.0346289575,
0.0088101467,
0.0023366134,
-0.0437348299,
-0.0575989299,
-0.0727702305,
0.0103230895,
-0.026372252,
0.0236847252,
0.0352764651,
-0.0255497005,
-0.0306790918,
0.0043187216,
-0.045073811,
-0.0263562296,
0.052719675,
-0.0035289344,
0.0511861853,
-0.0213408172,
0.0207038391,
-0.0261933655,
0.0039000949,
0.0276243929,
0.0571040511,
0.005969679,
-0.0451497324,
-0.071222119,
0.009922212,
-0.0169102438,
-0.028566964,
0.0071192347,
-0.057464689,
-0.003527442,
-0.0035901843,
0.0711132735,
0.0276413113,
0.0359960422,
0.0233971514,
0.020964399,
0.023841355,
0.0063804057,
-0.0502626374,
-0.0200802889,
0.0123383403,
0.0104483832,
-0.075969018,
0.0370974541,
0.035534665,
0.0432741754,
0.0050310059,
-0.0091549922,
0.0505274497,
-0.0022457079,
-0.0410588197,
0.0327824689,
0.0083663305,
-0.0549229942,
0.0366179012,
-0.040441338,
-0.0606914461,
0.0183368642,
-0.0469788238,
0.0299921092,
0.0030144132,
-0.0058848532,
-0.0058500604,
0.0249286834,
0.0116365664,
0.0075538382,
0.0304644015,
-0.00377904,
-0.0186530538,
0.0346798226,
0.0070086801,
-0.0632568076,
0.0222832728,
-0.053856574,
0.0110377586,
-0.0114023061,
-0.0128418682,
-0.060248293,
-0.0644034445,
0.0000034276,
0.0112081766,
0.0290282592,
-0.0302050877,
-0.0399738029,
-0.0347902253,
0.0236648247,
0.0376853384,
-0.045685932,
-0.0135772424,
-0.0207554027,
0.116593577,
0.0782843903,
-0.0296731517,
0.018784279,
-0.0772812292,
-0.0169944372,
0.0723425969,
0.0097279213,
0.0510988496,
0.0080209868,
-0.0108416975,
0.0023433829,
0.0126764206,
-0.0166793577,
0.0234057978,
0.0489920713,
-0.0038817348,
-0.0277981758,
0.0732992217,
0.0170659944,
-0.0319108367,
0.0221285019,
0.0071248021,
0.0582740232,
-0.0102645401,
0.0214495063,
-0.0530748256,
-0.0021806802,
0.0119866338,
0.0405813754,
-0.0557302423,
-0.0291498397,
-0.031773258,
-0.111243099,
0.0525110774,
0.0217347164,
-0.0395629853,
-0.0346122943,
-0.0309622157,
0.0001485979,
0.0413988195,
0.0299306381,
0.0168537945,
-0.0036177866,
0.0311058629,
0.0307952389,
0.0269082375,
-0.0552753955,
-0.022522673,
-0.0261676833,
0.0802937895,
0.0505750142,
0.0423476286,
0.0079135345,
-0.0012175707,
-0.0210820325,
-0.010919814,
-0.0124887228,
0.0177198853,
0.0677410364,
0.0408091657,
-0.0214776453,
-0.0191800594,
0.0563556626,
-0.0157949906,
-0.0210239962,
0.0023360311,
0.0300826281,
-0.0084523847,
-0.0612565912,
-0.035919074,
-0.0116270399,
0.0280312877,
0.0588122047,
0.0489818193,
-0.0371126793,
-0.0433260314,
-0.0383086465,
0.0195119195,
-0.0762901008,
-0.0450451672,
0.036379531,
-0.0089581609,
-0.0311854854,
-0.1148412675,
-0.0389744043,
-0.0307906512,
0.0799145475,
-0.0086025167,
0.0158303995,
0.026736265,
-0.0589966364,
-0.0296305213,
-0.0278729443,
-0.0862789154,
0.0205315836,
-0.0272178669,
-0.0051655043,
-0.0146433879,
-0.0565820336,
-0.0094718765,
0.027734682,
-0.1012600139,
-0.0580672622,
-0.0077294027,
0.0122498572,
0.0293382518,
0.0136294058,
0.0124577247,
-0.0895603746,
0.1018961519,
0.0745373741,
-0.0239734165,
0.0574177317,
-0.0113623645,
-0.0050046761,
-0.0093347616,
0.038451653,
-0.0127596129,
0.0159580503,
0.0123787886,
0.0311561916,
-0.0097369067,
-0.0266852081,
-0.0086147925,
0.0167114995,
-0.0106226346,
-0.061273437,
0.0209347531,
-0.0378227681,
-0.0273885541,
-0.0352356546,
0.0021800711,
-0.0584446974,
-0.0013443492,
0.0451419428,
-0.0268315449,
-0.0243918095,
-0.0174133275,
-0.0405421406,
0.0441087037,
-0.0341814682,
-0.0564159378,
-0.0177730657,
0.0376178548,
0.0125036631,
-0.0254785512,
-0.0038053244,
-0.0108716944,
-0.0391872004,
0.0601287074,
0.0038604827,
0.0331476405,
0.0098736733,
-0.0207259823,
-0.0000472085,
0.0178937986,
-0.0618502945,
0.0107407151,
-0.0525577329,
0.0582362153,
0.0370742232,
-0.0040747491,
0.0668159798,
-0.0295206849,
0.0077458904,
-0.0455467254,
0.0286586136,
0.021571923,
-0.0297686569,
-0.0316593871,
-0.0131571693,
-0.0305958316,
0.0360568129,
0.106267646,
0.0247362461,
-0.0045677722,
-0.0036640679,
-0.019914845,
-0.0221587326,
-0.0261795912,
0.0029644477,
-0.0219354741,
0.0004149466,
-0.0488995351,
-0.0665045306,
-0.0547064394,
0.0726400092,
0.0330244005,
-0.0023629048,
-0.0144567564,
-0.0099322153,
0.0517176166,
-0.0152804535,
0.0652236417,
-0.0736731067,
0.0507098883,
0.0007286803,
-0.0031420446,
0.005957271,
-6.309371888e-33,
-0.0007822792,
-0.0802148059,
-0.0222245865,
-0.0453419574,
-0.0311336387,
0.0467723124,
-0.0447458588,
0.0044205585,
0.0411837399,
0.0306191333,
-0.0233782642,
0.0005853939,
0.0213468987,
0.0455514975,
-0.0173028186,
0.0170086902,
0.0211311541,
-0.0322493352,
-0.0615902655,
0.0339222141,
0.0444100462,
0.0213965904,
0.0061862422,
-0.0447215773,
0.034753155,
-0.0390321538,
-0.0188008081,
-0.0120416889,
-0.0029464532,
0.0396447852,
-0.0500078127,
0.0534601919,
0.0134367067,
-0.0131448535,
-0.0449355096,
0.0265330821,
-0.0168287009,
-0.0101063401,
0.0036613219,
-0.0107237343,
-0.0211652964,
-0.0334943235,
0.0438328944,
-0.0294698887,
-0.0095641492,
0.0104478821,
0.0239722691,
-0.0461919792,
-0.0362743698,
-0.039736487,
-0.0232638698,
0.0256484151,
0.0124800503,
-0.0056039705,
0.0003365795,
0.0090439664,
-0.0133636352,
0.0330162682,
-0.0285879634,
0.0103671411,
0.0167633314,
0.039748013,
0.0180634931,
0.0076622046,
0.0036191242,
0.0935395062,
0.0274928622,
-0.0029676815,
-0.0259108748,
0.00698255,
0.0126154497,
-0.0019968795,
0.0198549032,
0.1079676151,
0.0150355529,
-0.0272194687,
-0.0458660163,
0.0513387248,
0.0188108198,
-0.0642693266,
0.0597312711,
-0.008343826,
-0.0291228425,
-0.0055919774,
-0.0199300107,
-0.0501953699,
-0.0031876133,
0.0209762007,
0.0101108123,
0.0390289985,
-0.0373413004,
-0.0391359851,
0.0054260562,
-0.01651939,
-0.0268679373,
0.004348034,
-0.0491921455,
0.0219294913,
-0.0109934919,
0.0104171243,
-0.0431755334,
0.0013765681,
0.0307210088,
0.0311796349,
-0.0109835938,
0.0102536986,
0.0151747381,
0.0244507845,
-0.0493937545,
0.0320622511,
-0.0301147588,
0.055354666,
0.0180358998,
0.0018599286,
0.0265412387,
-0.0399810672,
-0.0067886128,
0.0061440417,
0.015268296,
-0.0089312522,
-0.009579259,
0.0243227184,
-0.0194394067,
0.0191655587,
-0.0613606162,
0.0138646737,
-0.0517220162,
0.0334296189,
0.0174349062,
-0.0568106472,
0.018652169,
-0.0079817325,
2.476e-7,
0.011190841,
0.0224527474,
0.0452558845,
0.0378148369,
0.0397685431,
0.0417215712,
0.0024467441,
0.0311272349,
-0.0089806626,
-0.0270768199,
-0.0225225501,
0.0038665058,
0.0639608279,
0.0152689945,
-0.0721581057,
0.0256869178,
-0.0612336323,
0.0134241311,
-0.0397314206,
-0.070146434,
0.0337038413,
0.0648202449,
0.0103960549,
-0.0443836898,
0.0311779305,
-0.0556649566,
0.0793937594,
0.0208594985,
-0.0403816216,
0.0236067083,
0.0657003,
0.006993053,
0.0113157453,
-0.0121778138,
-0.0207767915,
-0.0037890207,
0.030265443,
0.0665326864,
0.0366097018,
-0.020378558,
0.0112281246,
-0.0447777212,
0.0199420471,
-0.0548167229,
0.0468872003,
0.0120725567,
-0.0010452766,
-0.029337002,
-0.0130564515,
-0.0192839317,
0.0107332505,
-0.0348189101,
-0.0237341337,
0.0518295132,
0.0068791574,
0.0373115242,
-0.0145278685,
-0.0136352032,
-0.0536269844,
-0.0674312487,
-0.0555633232,
-0.0280625708,
-0.0727725774,
0.0313939936,
0.0117987618,
-0.0748753399,
0.0288989805,
2.154937596e-34,
0.0292574428,
-0.0231845547,
-0.0154858846,
0.0151252868,
0.0102154361,
-0.0093471268,
0.0589730404,
0.0286299903,
0.0025383187,
-0.0393739194,
-0.0209088568
] |
http://arxiv.org/abs/2401.15471v1 | http://arxiv.org/pdf/2401.15471v1 | ConvoSense: Overcoming Monotonous Commonsense Inferences for Conversational AI | 2024-01-27T17:51:05 | 2024-01-27T17:51:05 | Sarah E. Finch, Jinho D. Choi | cs.CL | cs.CL | Mastering commonsense understanding and reasoning is a pivotal skill
essential for conducting engaging conversations. While there have been several
attempts to create datasets that facilitate commonsense inferences in dialogue
contexts, existing datasets tend to lack in-depth details, restate information
already present in the conversation, and often fail to capture the multifaceted
nature of commonsense reasoning. In response to these limitations, we compile a
new synthetic dataset for commonsense reasoning in dialogue contexts using GPT,
ConvoSense, that boasts greater contextual novelty, offers a higher volume of
inferences per example, and substantially enriches the detail conveyed by the
inferences. Our dataset contains over 500,000 inferences across 12,000
dialogues with 10 popular inference types, which empowers the training of
generative commonsense models for dialogue that are superior in producing
plausible inferences with high novelty when compared to models trained on the
previous datasets. To the best of our knowledge, ConvoSense is the first of its
kind to provide such a multitude of novel inferences at such a large scale. | ConvoSense: Overcoming Monotonous Commonsense Inferences for Conversational AI: Mastering commonsense understanding and reasoning is a pivotal skill
essential for conducting engaging conversations. While there have been several
attempts to create datasets that facilitate commonsense inferences in dialogue
contexts, existing datasets tend to lack in-depth details, restate information
already present in the conversation, and often fail to capture the multifaceted
nature of commonsense reasoning. In response to these limitations, we compile a
new synthetic dataset for commonsense reasoning in dialogue contexts using GPT,
ConvoSense, that boasts greater contextual novelty, offers a higher volume of
inferences per example, and substantially enriches the detail conveyed by the
inferences. Our dataset contains over 500,000 inferences across 12,000
dialogues with 10 popular inference types, which empowers the training of
generative commonsense models for dialogue that are superior in producing
plausible inferences with high novelty when compared to models trained on the
previous datasets. To the best of our knowledge, ConvoSense is the first of its
kind to provide such a multitude of novel inferences at such a large scale. | [
0.0432296097,
0.0930728316,
-0.0052730646,
0.0390972421,
-0.0142127248,
0.00880164,
-0.0185253341,
-0.0141666187,
0.039700307,
-0.0184253268,
-0.0032527228,
0.0346347205,
0.0051279799,
0.0229925644,
0.0078122891,
-0.0393473133,
0.0293609649,
-0.0108434362,
-0.0017879637,
0.0016744154,
0.0212357063,
0.0409387313,
-0.0194042493,
0.0633707345,
-0.0041251336,
-0.0305213463,
-0.0001210268,
0.0060264119,
0.0426289625,
0.0185586214,
-0.0140406564,
0.043873135,
0.0491817221,
0.0149979694,
0.0000015495,
-0.0217848048,
0.0309270304,
-0.0299874898,
-0.0052334368,
-0.0183206722,
0.0111746024,
0.0064592916,
0.0044669788,
-0.0007234985,
-0.0592035279,
0.0289215613,
0.0279188324,
0.0428678729,
0.0331741348,
0.0409115814,
-0.0211512428,
-0.0362847745,
-0.0317143947,
-0.020036215,
0.0206965897,
-0.08322341,
0.0044083865,
0.0310308821,
-0.0100097731,
-0.0031178098,
-0.0095372722,
0.0121429889,
0.045243755,
-0.0411776565,
-0.043705482,
0.0525576361,
-0.0633030608,
0.0044480152,
-0.0184909161,
0.1240735352,
0.0439735427,
-0.0162121393,
-0.0266402382,
0.0063210661,
0.04417523,
-0.0019759969,
-0.0381401703,
0.0129688196,
-0.0073435218,
0.0432446636,
0.0218020249,
-0.0016300294,
-0.0050335154,
0.0367857665,
-0.003753406,
0.0096707148,
-0.0078481929,
0.0115481773,
0.0105012143,
-0.0009429942,
-0.0387055874,
-0.0172126051,
-0.0294335689,
0.0144824516,
-0.0339345783,
0.0281646065,
-0.0379594117,
0.018899288,
0.0211010873,
-0.0334697217,
-0.0035687576,
0.0199427679,
-0.0299058761,
0.067835249,
-0.0017290289,
0.0106292255,
-0.0276753884,
0.0083416523,
-0.0157378465,
-0.048347082,
-0.0403304622,
0.0001867026,
-0.0306586623,
0.0070243157,
-0.0653094873,
-0.0139851486,
-0.0197675843,
0.0083197057,
-0.0064837309,
0.0391316004,
0.0108189313,
0.0607638806,
-0.0689604282,
-0.036044009,
-0.0247654989,
0.0385879576,
-0.0667246059,
0.0550491549,
0.0463472679,
-0.0599405468,
-0.0505650938,
0.0159213897,
0.0308708753,
-0.0297358185,
0.0063357884,
0.0304349475,
-0.0312873609,
-0.0017076932,
-0.0511882082,
-0.0104632573,
-0.0054912674,
-0.0414675474,
-0.0183192268,
-0.0352949239,
0.0443064049,
-0.0092091868,
-0.0097579053,
0.0082638543,
0.0237571727,
0.0070368927,
-0.0172293112,
0.0745024234,
-0.0121295741,
-0.0033961746,
0.0045596655,
0.0316908769,
-0.0042527942,
0.0775189251,
0.004004762,
-0.013226171,
0.0388799682,
-0.0012268372,
0.0122533599,
-0.0058798003,
0.0138660828,
-0.0105940867,
0.0076049767,
0.0271482784,
-0.0401713736,
0.0323919654,
0.0477364324,
0.0153783178,
-0.045858968,
-0.0406784303,
0.0552946776,
0.0772291422,
-0.0492948219,
0.0472986847,
0.0235217866,
0.026920516,
0.0135859782,
-0.001877018,
-0.0436141044,
0.0206422564,
0.0006594434,
-0.0269357506,
-0.0037323132,
0.0174072552,
-0.0274638068,
-0.0156277977,
0.0199862514,
-0.0076975529,
0.0243890323,
-0.0088598412,
0.0035055054,
0.0342067629,
-0.0310350619,
0.0251788646,
0.0254815537,
-0.0702066422,
-0.0690617114,
-0.0264966898,
0.0186718144,
0.0257839523,
0.0449060462,
0.00626744,
-0.0694363639,
-0.0248866994,
0.0146376854,
0.064667061,
0.0315550193,
-0.0308279041,
0.0741512403,
-0.0595162995,
0.0073975273,
0.0042806934,
0.0189268347,
0.0293401796,
0.0300208684,
-0.0021553433,
-0.0094463425,
-0.0284320656,
0.0488008112,
-0.0137572028,
-0.0133391637,
-0.0186982267,
0.0226558466,
-0.0275347847,
0.0401494615,
0.0059206164,
0.0034781462,
-0.034975078,
0.0187408924,
-0.0015372572,
-0.011596459,
0.0929185599,
-0.0465597883,
-0.0857864395,
0.0214807447,
-0.0574562512,
-0.0826389194,
0.0200157128,
-0.0655693039,
-0.0346675888,
0.0232145917,
-0.0002439894,
0.0559509434,
-0.0640058219,
-0.0598226078,
0.0152746774,
0.0268049054,
0.0415959172,
0.0835343003,
-0.0261809882,
-0.0160606913,
0.0417490005,
0.0158167072,
0.0104381572,
-0.0457464457,
0.0330824815,
-0.0191963259,
-0.0078481901,
-0.0039565377,
-0.0099193966,
0.0360170938,
-0.0457969755,
-0.0021203237,
-0.1146204397,
-0.0102608968,
0.0019460032,
-0.0075985887,
0.0022246435,
0.0053714248,
-0.0103523694,
-0.0287704431,
-0.0349623896,
0.0051439898,
0.0122649912,
-0.007991815,
0.0460832678,
-0.0217430238,
0.0254035424,
-0.0209379736,
0.0292733274,
-0.0045090672,
0.0210265033,
0.0046042111,
0.0860882252,
-0.0013130221,
-0.004893105,
-0.0393074714,
0.0461698994,
-0.0490368046,
-0.069305256,
0.0214434173,
-0.009641449,
0.0103601683,
0.0731513798,
0.0211522244,
0.0422322415,
-0.0180116426,
-0.037475992,
-0.0382122211,
0.0436005965,
-0.0170719139,
0.0060389172,
-0.0468582995,
0.012386132,
-0.0253972113,
0.0263496283,
0.0177175906,
0.025071837,
-0.0647348762,
0.002428076,
0.0487684757,
-0.0173509866,
-0.0007608596,
0.028581528,
-0.0186446607,
0.0081221238,
-0.0237782821,
-0.0180098265,
0.0669670105,
-0.001272193,
-0.0253946092,
-0.0438127555,
-0.0103864418,
-0.0251914617,
0.0556031689,
-0.0453749783,
-0.0206154939,
-0.0192862898,
-0.0092363479,
-0.0321721621,
-0.0137780113,
0.0330349728,
-0.019231081,
0.0286194999,
-0.0562394671,
-0.0008524295,
-0.0137156714,
-0.0528643057,
-0.0523611084,
-0.0372127816,
-0.0085486686,
-0.001072776,
0.0320525803,
-0.0537032187,
-0.0341670923,
-0.0463642217,
-0.0160374399,
0.0470098034,
0.0047584097,
-0.0055062338,
-0.019968491,
0.0441217795,
0.0468058027,
0.069975771,
-0.0043527386,
0.0238440111,
0.0212546904,
0.074557513,
0.0036836932,
-0.0207510516,
-0.0241187718,
0.0277935211,
-0.0125587368,
0.0266681425,
0.0178010613,
0.0016688186,
0.0698435456,
-0.028524559,
-0.0162941422,
0.0771050528,
0.0121895354,
-0.0200927965,
-0.0017289303,
-0.0061616553,
0.0453918613,
-0.0105054155,
-0.0015187921,
-0.104331024,
-0.0640462339,
0.0115684662,
-0.0049478952,
-0.0295826606,
-0.0632344112,
-0.0098139513,
-0.1024458557,
0.038705375,
-0.0199611858,
-0.0672853664,
-0.0459258184,
-0.0217848644,
0.0223857556,
-0.0325797275,
-0.0127432048,
-0.0131438393,
0.035829179,
0.0008497204,
-0.0031765501,
0.0221279226,
-0.0771009624,
0.0346005112,
0.0073525654,
0.0679120049,
0.0255484879,
0.0723448023,
0.0303209331,
-0.0485942662,
-0.0047799442,
-0.0446109585,
-0.027854627,
-0.0147993807,
0.0126644131,
-0.0216661803,
-0.0095756752,
-0.0179732312,
0.0994232893,
-0.064181827,
-0.0385773294,
0.0253170002,
0.0336748622,
-0.0294660907,
-0.0163523797,
-0.0263486989,
0.0355283655,
0.0509682298,
0.1015500203,
0.038343817,
0.0195155982,
-0.0151304891,
-0.0469014309,
-0.0249824151,
-0.0528497249,
0.0075338115,
-0.0080607301,
0.0675302371,
0.0030875879,
-0.0581734851,
-0.0259739496,
0.0135032348,
0.0306036808,
-0.0360577293,
-0.0168684013,
-0.0323750079,
0.0310144164,
-0.022676358,
0.0555586666,
-0.0191240218,
-0.0484516695,
-0.0124379126,
-0.0114809135,
0.0355222709,
-0.0177545473,
-0.0513654649,
0.0773294047,
-0.0603275783,
-0.0527639538,
-0.0573748387,
0.0473951288,
0.0231363513,
0.046157863,
-0.0222227741,
0.0194675736,
0.0131302718,
0.0529525504,
-0.0120983329,
-0.019675497,
0.0637358129,
0.0264552459,
0.0333668292,
0.083476536,
0.0030183794,
0.0111945784,
-0.0078020855,
0.039140489,
-0.0575531125,
-0.0240986161,
0.0003201265,
-0.0376471728,
0.0051248795,
0.0211938899,
0.0264912397,
-0.0445190966,
0.0174031239,
-0.0190661922,
-0.0241466817,
-0.0613656081,
0.0267632753,
0.0068123667,
-0.0212717708,
-0.0297086388,
-0.0108106788,
-0.0414524414,
0.0498978719,
-0.0282326005,
-0.0294888802,
0.0616944917,
0.0005511697,
-0.0157460775,
-0.0437458977,
-0.0195640624,
-0.0067586177,
-0.0339035913,
0.0378369056,
0.01110176,
0.0849568248,
0.060779009,
-0.0564951748,
0.0307503492,
0.0027231537,
-0.0514726229,
-0.0112471776,
-0.0075619239,
0.0781348646,
0.0253898855,
-0.0050769057,
0.039239984,
-0.0373150781,
0.0170487612,
0.0355539396,
0.0838687867,
-0.0068464214,
-0.0226834137,
0.0582846329,
0.0085443035,
-0.0196363255,
0.0107143614,
0.0645398647,
-0.0657569021,
0.0404655933,
0.000717562,
-0.0111083221,
-0.03843778,
-0.0549566634,
0.0167983454,
0.0402495228,
0.0564021841,
0.0536656715,
-0.0478720963,
-0.0447807238,
-0.0247498956,
0.0310365558,
0.0145063745,
0.0180178974,
0.0259326212,
0.0714855492,
-0.0046612578,
-0.053813234,
-0.003860574,
0.0139020877,
-0.0628892332,
-0.0124531379,
-0.0738153681,
-5.543337485e-33,
-0.00970688,
-0.007521986,
0.0033631702,
-0.0198604483,
-0.0214276686,
-0.0275222342,
-0.0635514036,
-0.0314044654,
-0.042217955,
0.0495687015,
-0.0199643224,
0.0672369674,
0.0001702038,
0.0457148924,
-0.0099695371,
-0.0324591845,
0.0090498896,
0.0319034904,
-0.0245358441,
0.0450939834,
0.0529602282,
-0.024633579,
0.0764668807,
-0.0324628577,
-0.0057305461,
-0.000745006,
-0.0160729066,
0.0401747711,
-0.0163133163,
0.0376477167,
-0.007890637,
0.0449952595,
0.0064088241,
-0.0171652585,
0.0245669223,
0.043519102,
-0.0588418432,
-0.0588227659,
0.0452917218,
-0.0124649825,
-0.0330217257,
-0.0048677158,
0.0011330852,
0.0016664396,
0.0222861059,
-0.0016435108,
0.022063639,
-0.051635731,
0.0298702978,
-0.0099571645,
-0.0257908329,
0.0243599638,
-0.045015756,
-0.0503202081,
-0.062920697,
0.0303893834,
-0.0006767522,
-0.0261395182,
-0.016806744,
0.0259507801,
0.0138662215,
-0.0182367116,
0.0343218371,
-0.0393958874,
0.0301988851,
0.0251827016,
-0.0320968702,
-0.026995359,
-0.0579670742,
-0.0017704953,
-0.0127961691,
0.0603676476,
0.0113039315,
0.0842152983,
0.0233127698,
0.0176775735,
-0.0460487232,
0.0373965167,
0.0122449687,
0.0507023148,
0.0049789734,
-0.0060001533,
-0.0205250029,
0.0051563699,
-0.0194845349,
-0.0306273568,
0.0105101913,
0.0062559154,
-0.000431017,
-0.0155266263,
-0.019780077,
0.0301267952,
0.0387221277,
-0.0265249778,
-0.0013858132,
0.0592091307,
-0.0266286656,
-0.0324534439,
0.0235404167,
0.0177059527,
-0.0359723978,
-0.0039433343,
0.0779041648,
0.0045450637,
0.0105631733,
0.0292582829,
0.0260377638,
-0.010550675,
-0.0018979418,
-0.0127391908,
-0.0277993381,
0.036375206,
-0.0858840942,
0.0171625037,
0.0221966375,
0.0209926032,
-0.0149659803,
0.0285868794,
-0.0159706715,
0.0285282843,
-0.0291403793,
0.0255908407,
0.0037060319,
-0.0011116358,
-0.0321741514,
0.0014535765,
-0.0219959412,
-0.0579101145,
0.0109056346,
-0.0386295952,
-0.0012275703,
0.0286492035,
2.157e-7,
-0.008966431,
-0.000255944,
0.0346473418,
0.0071007642,
0.0086426167,
0.0542399846,
0.0077113174,
0.032913167,
-0.0038160149,
-0.0399846584,
0.0275124237,
0.063591592,
0.0593849607,
0.0238224436,
-0.0770028979,
-0.00105489,
-0.0758964717,
-0.0355383903,
-0.0282212514,
-0.0392163619,
0.0630774647,
0.0879583135,
0.0296218041,
-0.0259868484,
0.0677384734,
-0.0306252576,
0.0265139472,
-0.0300143845,
0.0081434362,
-0.009135535,
0.0259667467,
0.0168565996,
-0.0084702559,
0.0320137925,
0.0146012278,
0.0107526984,
-0.0336584039,
0.0429868922,
-0.0058508366,
-0.0612879395,
-0.0217730943,
-0.0349891745,
0.0205842592,
-0.0450341366,
0.0660963356,
-0.0702978298,
-0.0313285068,
-0.0194872003,
-0.0115721738,
-0.0156864524,
0.0236288924,
-0.0197339654,
-0.0248462167,
-0.0016500707,
0.0255618133,
-0.0243600067,
0.0016805424,
-0.0738678724,
0.0179226119,
0.0022330538,
-0.066843152,
-0.0372581333,
-0.0039781663,
-0.0578091778,
0.0235715862,
-0.0529908016,
0.0084582055,
1.75932038e-34,
0.0042342972,
-0.0003559912,
-0.000123164,
0.0476465896,
0.0512693152,
0.0013941117,
0.0353618711,
-0.0186984073,
0.0330329239,
-0.0313848145,
-0.0097814472
] |
http://arxiv.org/abs/2401.15463v1 | http://arxiv.org/pdf/2401.15463v1 | DataFrame QA: A Universal LLM Framework on DataFrame Question Answering Without Data Exposure | 2024-01-27T17:06:53 | 2024-01-27T17:06:53 | Junyi Ye, Mengnan Du, Guiling Wang | cs.CL | cs.CL, cs.AI | This paper introduces DataFrame question answering (QA), a novel task that
utilizes large language models (LLMs) to generate Pandas queries for
information retrieval and data analysis on dataframes, emphasizing safe and
non-revealing data handling. Our method, which solely relies on dataframe
column names, not only ensures data privacy but also significantly reduces the
context window in the prompt, streamlining information processing and
addressing major challenges in LLM-based data analysis. We propose DataFrame QA
as a comprehensive framework that includes safe Pandas query generation and
code execution. Various LLMs, notably GPT-4, are evaluated using the pass@1
metric on the renowned WikiSQL and our newly developed 'UCI-DataFrameQA',
tailored for complex data analysis queries. Our findings indicate that GPT-4
achieves pass@1 rates of 86% on WikiSQL and 97% on UCI-DataFrameQA,
underscoring its capability in securely retrieving and aggregating dataframe
values and conducting sophisticated data analyses. This approach, deployable in
a zero-shot manner without prior training or adjustments, proves to be highly
adaptable and secure for diverse applications. | DataFrame QA: A Universal LLM Framework on DataFrame Question Answering Without Data Exposure: This paper introduces DataFrame question answering (QA), a novel task that
utilizes large language models (LLMs) to generate Pandas queries for
information retrieval and data analysis on dataframes, emphasizing safe and
non-revealing data handling. Our method, which solely relies on dataframe
column names, not only ensures data privacy but also significantly reduces the
context window in the prompt, streamlining information processing and
addressing major challenges in LLM-based data analysis. We propose DataFrame QA
as a comprehensive framework that includes safe Pandas query generation and
code execution. Various LLMs, notably GPT-4, are evaluated using the pass@1
metric on the renowned WikiSQL and our newly developed 'UCI-DataFrameQA',
tailored for complex data analysis queries. Our findings indicate that GPT-4
achieves pass@1 rates of 86% on WikiSQL and 97% on UCI-DataFrameQA,
underscoring its capability in securely retrieving and aggregating dataframe
values and conducting sophisticated data analyses. This approach, deployable in
a zero-shot manner without prior training or adjustments, proves to be highly
adaptable and secure for diverse applications. | [
0.025007192,
0.1196043044,
-0.0032760676,
0.0504409112,
-0.0066419248,
0.0385976695,
0.0610228069,
0.0009082214,
0.0416156054,
-0.0067620799,
-0.0058850306,
-0.0329643413,
-0.0284564514,
0.0347360708,
0.0235291701,
-0.0063899048,
0.0291115828,
-0.0585715137,
-0.0185054746,
-0.0175882187,
-0.0271365345,
-0.0075738784,
0.0082987305,
-0.0030074641,
0.007844694,
-0.0294317473,
-0.0542206354,
-0.0237722993,
0.0119708385,
-0.0225702673,
0.0332479589,
0.0559115373,
0.0345354937,
0.0438481197,
0.0000021349,
-0.0426489748,
-0.0366827771,
0.0347812399,
-0.0469889939,
-0.0086412793,
0.0169296656,
-0.0136282826,
0.0516921617,
0.0149420165,
-0.0382786803,
-0.0289112031,
0.0413765386,
0.0380914547,
0.0966421068,
0.0507181287,
0.0013790891,
-0.0463844091,
0.0094470587,
-0.0007087668,
0.0364696719,
-0.0778588429,
0.0532752723,
-0.0133779291,
-0.0136658987,
-0.0191182159,
0.0192012675,
0.0172099099,
0.0159889068,
0.0143009108,
-0.0348462239,
0.0118397744,
-0.0487084351,
-0.0196939781,
-0.0195338689,
0.0519563369,
0.0246335063,
-0.0205014646,
0.006057377,
0.0237158667,
0.0145406611,
0.0244057663,
-0.0625200495,
0.0366961583,
-0.0170984399,
-0.0258331392,
-0.0090369238,
0.0216218811,
0.012864342,
0.0384507217,
-0.032389041,
0.0639150217,
0.0079859318,
-0.0247852374,
-0.0226791445,
-0.06712953,
0.0129328053,
-0.057901375,
0.0262507387,
0.0600079335,
-0.0325677879,
0.0119062075,
-0.0118570048,
-0.0565708838,
0.0493460968,
-0.0348754153,
-0.0259924605,
0.0023546913,
0.0150780678,
0.0466863811,
0.0172744729,
0.0461894013,
-0.0106721902,
0.013586035,
0.0010510805,
-0.0379630364,
-0.0117973713,
-0.0328577273,
-0.0430994593,
0.0526262857,
-0.0562517531,
0.0241057295,
-0.0047860332,
0.0643579438,
-0.0049561607,
0.0523783565,
-0.0070902267,
0.0155953784,
0.0010446712,
0.0323181525,
-0.0381430164,
-0.0187960416,
-0.0381370373,
0.0018305668,
0.0062886747,
-0.0455064215,
-0.0149707282,
0.0016480451,
0.0222141538,
-0.0374004953,
0.0342915989,
0.0984198079,
0.0412122495,
-0.0043758862,
-0.0448266752,
-0.0668811724,
0.0097328704,
-0.0366785303,
-0.0010764765,
-0.061965175,
0.0184500366,
0.029401809,
-0.00425538,
-0.0278880745,
0.0059136637,
0.0686873272,
-0.0244003441,
0.0627726316,
-0.0357226245,
0.012007351,
-0.0192030817,
0.0317082964,
0.0147059048,
0.0402740054,
-0.0152371116,
0.0564509369,
0.0482702851,
-0.020759711,
0.071602501,
-0.0196791906,
0.0368445031,
0.0334456936,
-0.0164124779,
0.1164059341,
0.0320271514,
-0.0345739424,
-0.024125861,
0.0415577143,
0.0122541264,
-0.064580217,
0.0637339577,
0.0744951814,
-0.0141874067,
-0.0019545904,
-0.0464840457,
0.0363248475,
0.0100695267,
-0.0237689838,
-0.0129348952,
0.0292513035,
-0.0361652672,
-0.0067997929,
-0.0233670771,
0.0249249861,
-0.0227973778,
-0.0168878548,
-0.0087016476,
-0.0202694219,
0.018191563,
0.0646809712,
0.0063909632,
0.0086316401,
-0.0059907027,
0.048769094,
0.0043093381,
-0.075550586,
-0.0547059923,
-0.0392872617,
0.0374221392,
0.0790208429,
0.0314672738,
-0.0276288651,
-0.0402982011,
-0.0496902391,
0.0564425364,
0.0062326337,
-0.0153875994,
-0.0057218694,
0.0181937199,
-0.029564796,
-0.0294920132,
-0.006288223,
0.0174568836,
-0.033905074,
0.0150885368,
0.0056190277,
0.0036163232,
-0.0055890889,
-0.0265339836,
0.0376617536,
-0.0011318957,
-0.0651563555,
0.0203377623,
0.014220302,
0.0291973278,
0.0183803756,
-0.0080656931,
0.0298957303,
0.0406858623,
0.0190306008,
-0.0088930903,
-0.0412907824,
0.0383665189,
0.0262993351,
-0.0016491915,
-0.0069708265,
-0.0275149904,
0.0149675859,
-0.0435741805,
-0.0037468863,
-0.0325625725,
-0.0110915545,
-0.030370418,
-0.0322109982,
-0.0302263983,
0.0112413811,
0.0109671056,
-0.0438792109,
0.0140350787,
0.0056691207,
0.0241347682,
-0.00315838,
-0.0144838998,
0.0437855572,
-0.0508737192,
-0.0089305975,
0.0141855646,
-0.0353803262,
0.0463232584,
0.0106678335,
0.0210678503,
-0.0402751155,
0.0159493666,
-0.0357364528,
0.0591143295,
-0.0033393395,
0.0350540765,
0.0147875361,
0.021966977,
-0.0539752357,
-0.0390861966,
-0.0675737262,
-0.037442077,
-0.0262608752,
0.0145144006,
0.0627140179,
0.0185349938,
-0.0265367366,
0.0141841955,
0.0023359524,
0.0054650651,
0.0519032367,
-0.0201085079,
-0.0145642338,
0.0307396986,
0.0155735863,
-0.0367920622,
-0.0103035802,
-0.0019600408,
-0.0086388616,
0.0004174398,
-0.0271358602,
-0.0393838212,
0.0399632864,
0.0303253271,
-0.0239949245,
0.0108963754,
-0.0397708938,
-0.0310580954,
0.0269386452,
-0.0139046079,
-0.0179902148,
0.0161038134,
0.0141348392,
0.0285988208,
-0.0049819774,
0.000676776,
-0.0349165834,
-0.0254023373,
0.0667166337,
0.0297224987,
-0.0204045214,
-0.0932108983,
0.0081117516,
0.0137484958,
0.0343377925,
0.0134306429,
-0.077336356,
0.0518761426,
-0.0138062751,
-0.0063370592,
-0.0160673987,
-0.034152586,
-0.0214232802,
0.1031216681,
-0.0187010858,
0.0180926565,
-0.0373387933,
-0.0364725478,
-0.0298395045,
-0.0222482011,
0.0052860235,
-0.0859793425,
0.0433999263,
0.000283613,
-0.0144693796,
-0.0268261638,
-0.0272571407,
-0.0660932139,
-0.0241774283,
-0.0355460569,
-0.0091111436,
-0.0411937721,
-0.0474766716,
0.0156073188,
0.0049182642,
0.0638684258,
-0.0030634783,
-0.022149913,
-0.0008125146,
0.0032007855,
-0.0035121143,
0.0248898137,
-0.0300111417,
-0.010501463,
-0.0047653145,
0.0557335652,
0.0675592199,
0.0401704162,
0.0129520446,
0.0209388547,
-0.0152975079,
-0.017372245,
-0.0296274163,
0.0118304957,
0.0216603,
-0.0068391985,
-0.0750929341,
0.0104863849,
0.0481106378,
0.0356179923,
-0.0008446209,
-0.0224595182,
-0.0003110724,
-0.0255679712,
0.0117506348,
0.0525313318,
-0.1646828353,
-0.011979904,
0.0124803092,
0.0707027987,
-0.0204846952,
-0.0756788254,
-0.0138499206,
-0.0903039724,
0.0589110889,
-0.0214950033,
0.0155974338,
-0.0344184898,
-0.1044392735,
-0.0136511195,
0.000421152,
0.0267581008,
0.0313526168,
-0.0011582745,
0.0093391519,
0.0119158439,
0.0156550314,
0.0080049029,
0.0180622917,
-0.0082683749,
-0.0191318635,
0.0267556366,
0.0294228718,
-0.02111784,
-0.0444893353,
-0.0319708474,
0.0055279862,
-0.0712742209,
0.0001796224,
0.0546624325,
-0.0043387325,
-0.0243908558,
-0.0468594916,
0.0393479168,
-0.0360507667,
-0.0043516094,
0.0455932952,
0.0527713113,
-0.0364618488,
-0.0236521736,
-0.0129504297,
-0.0462027676,
-0.0009673373,
0.0968588665,
0.0174509753,
-0.0082834754,
-0.0128797852,
-0.0242648553,
-0.0376022421,
-0.0410755835,
-0.0110452939,
0.0062816585,
0.0076975599,
0.005200923,
-0.0440992825,
0.014493295,
0.0588894077,
0.0155574186,
-0.0135975163,
0.0397858582,
-0.0177447516,
0.0166840684,
-0.0628989339,
-0.0203472115,
-0.0469508134,
0.02768833,
-0.0178133696,
0.0080072526,
0.0401864611,
-0.0036790341,
-0.0435063429,
0.0444344394,
-0.053675551,
-0.0241651367,
-0.0012941703,
0.0326373652,
0.0221523643,
-0.0578163788,
-0.0082883183,
-0.0196170155,
0.0625646114,
0.1396046281,
-0.0350547768,
-0.0251400191,
-0.0057162261,
-0.0304986089,
0.0311141126,
0.0040325955,
-0.0578310341,
-0.0395381562,
0.0190424062,
-0.0116808088,
-0.0204875711,
-0.0123821823,
0.0155989723,
0.0863589272,
-0.0268001407,
0.0371722765,
-0.0280917436,
-0.0035325859,
0.0732403249,
-0.0218985807,
-0.0455347337,
-0.0097059822,
0.0193591099,
0.0118310833,
-0.0137686683,
0.030914681,
0.0322845429,
0.0311372206,
0.0219184197,
-0.0416288935,
0.0192024615,
0.0325235389,
0.043840427,
0.0421076417,
-0.0260288753,
-0.017298514,
-0.0134244869,
0.0281952675,
0.0566853359,
0.0094571123,
0.0149743203,
0.0104781408,
-0.0224498175,
-0.0165646952,
0.0378983244,
-0.0828313231,
-0.0916757882,
0.0118485652,
-0.0039030686,
-0.0316207595,
0.0273129251,
-0.0308382399,
-0.0359821469,
-0.0214544814,
-0.0236604623,
-0.0101349661,
0.0126996664,
-0.0034959377,
0.0319823287,
0.0327874087,
-0.0410375595,
-0.0044411873,
0.0746520907,
0.0098196464,
0.0489795245,
0.0026351113,
-0.0674439222,
-0.0313460454,
-0.0352214873,
0.0279942583,
0.0263971053,
-0.0229345001,
-0.0285730064,
-0.0305213779,
-0.0636045933,
-0.019208597,
-0.0062273727,
0.0026076378,
0.0139884241,
0.0073178145,
0.0167327113,
0.0436048247,
-0.0168072339,
-0.0612808503,
0.0462976173,
-0.0046526291,
-0.0369307473,
0.0000815911,
-5.911667252e-33,
0.0048750225,
-0.0709211528,
0.0035112791,
0.0176786799,
0.0283194538,
0.0106057031,
-0.0106607554,
-0.0098019456,
-0.0135492012,
-0.0014040471,
-0.0281132236,
0.0033461456,
0.0241741147,
0.0276624579,
-0.034700416,
-0.0608857311,
-0.0205454864,
-0.0216062553,
0.0175606702,
0.0322206654,
0.0805079341,
0.0314112566,
0.0029804027,
-0.051999893,
0.0396748558,
-0.0350257829,
-0.0456708074,
0.0242108982,
-0.0199568924,
0.0462807417,
-0.0122640701,
0.0081786942,
0.0236672498,
-0.0091372347,
-0.0032840418,
0.0386427343,
-0.060214296,
-0.0144495741,
-0.0109013207,
-0.031023249,
0.0224982686,
-0.0362025164,
0.0648061335,
-0.0091197155,
0.0208614096,
0.0162361655,
-0.0186489969,
-0.0378965847,
0.0096686827,
-0.039275609,
0.0229896959,
0.0142836114,
-0.0170093048,
0.0193027649,
0.0172670465,
0.0195448492,
0.0063516754,
0.0214702114,
-0.0221439153,
0.0435482971,
0.0789737627,
-0.0249495003,
0.0199582092,
0.0365912579,
0.0600988753,
-0.0026664382,
0.042353794,
-0.036639113,
-0.0188736189,
-0.0026559816,
-0.0382665731,
0.0474217087,
0.0470379442,
0.046378497,
0.0090942103,
-0.0069189556,
-0.1064884588,
0.0235157721,
-0.0090217721,
0.0117337136,
0.0361050218,
-0.0204299428,
-0.0560791418,
-0.0056505068,
-0.0553864464,
0.0640351102,
-0.0356227309,
0.0003246459,
0.0181660838,
-0.0067925584,
-0.0539169237,
0.007429813,
0.016015457,
-0.0513507389,
-0.0216564797,
0.037054345,
0.0190102365,
-0.0296503399,
0.0166899841,
-0.0029442227,
-0.0085615795,
0.0449364781,
0.0048350897,
-0.0619738437,
0.0071731945,
0.0149264438,
0.0177455097,
0.0132578649,
-0.101309292,
-0.0167161897,
-0.0205049049,
0.0593871325,
0.0217960235,
0.0316075012,
0.0073610567,
0.0116081508,
0.0125856306,
0.0391895995,
-0.0044852803,
0.0035040057,
-0.0532586388,
0.0183654409,
-0.0334010236,
-0.0112154977,
0.0206878241,
-0.0281552579,
0.0059669833,
0.0022494607,
0.0422193632,
-0.0141847543,
0.020235341,
0.0109894052,
2.846e-7,
0.0386882238,
0.0072148261,
-0.0039841332,
0.0781483278,
0.0019144466,
0.023138199,
-0.0121332994,
0.0445885025,
-0.0035930793,
-0.0237419568,
0.0081708161,
0.0077989018,
0.0263641309,
0.0598420799,
-0.0841842443,
-0.0112310685,
-0.0460972637,
-0.0343892016,
-0.0651040971,
0.0081613837,
-0.0004375591,
0.064684473,
0.0061919,
-0.0075585237,
0.0250240881,
-0.0561496504,
0.0091807274,
-0.0212848987,
0.0126931174,
0.017601708,
0.0219338611,
0.0410816185,
0.0340754054,
0.0275726039,
-0.0100701973,
0.0050345832,
-0.0393843874,
0.0895770192,
0.0467971787,
-0.0192800481,
-0.003867575,
-0.0895374641,
0.0267737024,
-0.0702238455,
0.0932014138,
0.0073522227,
-0.0260236897,
0.0125506409,
0.0121716326,
-0.0549102053,
0.0217102561,
-0.0400228612,
0.0167336557,
0.0226323176,
-0.0166552532,
0.0100836568,
0.0541133955,
-0.0253000706,
-0.015125474,
-0.0566453226,
-0.0176266544,
-0.0512941293,
-0.0284173507,
0.0047317762,
-0.0080373837,
-0.0141384043,
0.0608791076,
2.935301918e-34,
0.0102000339,
-0.0083241332,
-0.0514549986,
0.0095183877,
0.0569718815,
-0.0206785463,
-0.0092338743,
-0.0037584559,
0.0450824127,
-0.0586103983,
-0.011437227
] |
http://arxiv.org/abs/2401.15449v1 | http://arxiv.org/pdf/2401.15449v1 | Learning to Trust Your Feelings: Leveraging Self-awareness in LLMs for Hallucination Mitigation | 2024-01-27T16:19:30 | 2024-01-27T16:19:30 | Yuxin Liang, Zhuoyang Song, Hao Wang, Jiaxing Zhang | cs.CL | cs.CL | We evaluate the ability of Large Language Models (LLMs) to discern and
express their internal knowledge state, a key factor in countering factual
hallucination and ensuring reliable application of LLMs. We observe a robust
self-awareness of internal knowledge state in LLMs, evidenced by over 85%
accuracy in knowledge probing. However, LLMs often fail to express their
internal knowledge during generation, leading to factual hallucinations. We
develop an automated hallucination annotation tool, Dreamcatcher, which merges
knowledge probing and consistency checking methods to rank factual preference
data. Using knowledge preference as reward, We propose a Reinforcement Learning
from Knowledge Feedback (RLKF) training framework, leveraging reinforcement
learning to enhance the factuality and honesty of LLMs. Our experiments across
multiple models show that RLKF training effectively enhances the ability of
models to utilize their internal knowledge state, boosting performance in a
variety of knowledge-based and honesty-related tasks. | Learning to Trust Your Feelings: Leveraging Self-awareness in LLMs for Hallucination Mitigation: We evaluate the ability of Large Language Models (LLMs) to discern and
express their internal knowledge state, a key factor in countering factual
hallucination and ensuring reliable application of LLMs. We observe a robust
self-awareness of internal knowledge state in LLMs, evidenced by over 85%
accuracy in knowledge probing. However, LLMs often fail to express their
internal knowledge during generation, leading to factual hallucinations. We
develop an automated hallucination annotation tool, Dreamcatcher, which merges
knowledge probing and consistency checking methods to rank factual preference
data. Using knowledge preference as reward, We propose a Reinforcement Learning
from Knowledge Feedback (RLKF) training framework, leveraging reinforcement
learning to enhance the factuality and honesty of LLMs. Our experiments across
multiple models show that RLKF training effectively enhances the ability of
models to utilize their internal knowledge state, boosting performance in a
variety of knowledge-based and honesty-related tasks. | [
0.0308433529,
0.0393337794,
-0.0197185818,
0.0278623495,
-0.0548120029,
-0.002218745,
-0.0356896855,
-0.0028487458,
0.0530072115,
-0.029620951,
-0.042516809,
-0.0176259354,
-0.0031608713,
-0.0214644447,
0.0332240276,
-0.0285537206,
0.0727639124,
-0.0436725765,
0.0302597079,
-0.0265858769,
-0.0145010334,
0.0232669655,
-0.0349921621,
0.0255814102,
-0.0630084351,
-0.022272883,
-0.0043451255,
-0.0045053246,
0.034141589,
-0.0218099002,
0.0208986998,
0.0731498599,
-0.0015610359,
0.051711984,
0.0000022224,
-0.0507147014,
-0.0116624823,
-0.0330303423,
0.005429768,
-0.0107770404,
0.0140087279,
0.0079873772,
0.0073233191,
-0.0031694628,
-0.0410791822,
0.0417440683,
0.0206020121,
0.0957554057,
0.079528369,
0.0751324594,
0.0076973583,
0.0041555665,
0.0029101733,
-0.0031854273,
0.0635115802,
0.0224075764,
0.0248301961,
0.0356766172,
-0.0165377371,
-0.0586217307,
0.0096318936,
0.0249013267,
0.0016019986,
-0.012402867,
0.0326885432,
0.0218912717,
-0.0497706383,
-0.0365343653,
-0.0097807888,
0.0371004008,
0.0038125892,
-0.0305184647,
0.0118350396,
-0.0058090952,
0.0129156746,
-0.0003813994,
0.002904041,
-0.0030181303,
-0.0390012562,
0.0207923483,
0.018319726,
0.0262682103,
0.0314051025,
0.01216821,
-0.0187253878,
0.0826067552,
-0.0208103824,
-0.0571186803,
0.0681922883,
0.0205492638,
0.0111881495,
-0.0767720491,
0.0162568148,
-0.0012475377,
-0.0376571715,
-0.0001764295,
-0.0391916037,
0.0096067935,
0.040098317,
-0.064807646,
-0.0238840804,
0.0090297498,
-0.0533335507,
0.03047527,
-0.0083904257,
0.0453902148,
0.0113939876,
0.1090762913,
-0.0032360721,
0.0279210303,
-0.0139135886,
-0.0108939968,
-0.0134925395,
0.0321265087,
0.0165309589,
-0.0084393378,
-0.0485588387,
-0.0162247103,
-0.0652685687,
0.0548814237,
-0.0323807597,
0.0274782833,
0.0050189318,
-0.0211095717,
-0.0656729266,
0.0155255264,
-0.009067554,
-0.0083919521,
0.0304657891,
-0.0292376895,
-0.0116535136,
0.0080756377,
-0.0134236151,
0.0027406756,
0.0071957028,
0.0086831842,
0.0042426493,
-0.0139990225,
-0.0316285789,
-0.0342406929,
-0.0183104426,
-0.0934567228,
-0.0183519423,
-0.0505723804,
0.0025501419,
-0.0437283628,
-0.0053599817,
-0.0375105292,
0.0455973782,
0.0177415013,
-0.0165737607,
0.0509559587,
-0.0404876582,
-0.0254881699,
0.0143181253,
0.0387484096,
-0.0158135835,
0.0501073822,
-0.0043836366,
0.0701064318,
0.0135062169,
-0.0155373747,
0.0581650697,
-0.0023258515,
-0.0004085541,
-0.019497199,
-0.0045217853,
-0.0162967611,
0.005223271,
0.0568061471,
0.0403776467,
0.0398856029,
-0.0323898382,
-0.0298599973,
0.1059165895,
0.0152808428,
-0.0066719986,
-0.0055166767,
-0.0306020249,
0.0451532081,
0.0187537856,
0.0145254452,
-0.0548101813,
0.0602551363,
-0.018883761,
-0.0108318133,
0.0492516495,
-0.0055211852,
0.0063700085,
-0.0377791375,
-0.0113891875,
-0.0153276036,
0.0090142442,
0.0464170054,
0.0011959212,
0.016866399,
-0.0409037843,
0.0376101583,
0.0254795421,
-0.0206481908,
0.0003668533,
-0.0593271479,
0.0517628677,
0.1105659828,
0.0308517683,
-0.037831299,
-0.0115799811,
-0.0470922403,
-0.0084700212,
0.0693977401,
0.038695462,
-0.0183378644,
0.008294913,
-0.0379255898,
0.0116945263,
-0.0131178545,
0.0320739485,
0.0068986802,
-0.0207300875,
0.0243392903,
-0.0191058386,
-0.0301907547,
-0.0264888834,
0.0224680062,
0.0095715774,
-0.0214942265,
0.001934798,
-0.0080003189,
-0.0034550636,
0.0171451773,
0.001704841,
0.0185241997,
0.0040210248,
-0.0038737229,
-0.0447063334,
0.0274341535,
-0.0588272959,
-0.0328078009,
0.0279703103,
0.0251703113,
-0.0083213672,
0.0010731143,
0.0052989204,
-0.0027407338,
0.0024596395,
-0.0394448936,
0.0758001208,
-0.0375378616,
0.0243841317,
-0.009485703,
0.0195775181,
-0.0185121726,
0.0546478704,
-0.0506681614,
0.0118166925,
-0.010939341,
-0.0403699763,
-0.0298286341,
-0.0195814911,
0.0072282841,
0.0000318989,
0.0099484501,
0.0417424366,
0.0038003346,
-0.0120466352,
-0.0593816377,
-0.0166545864,
-0.0587745532,
-0.0242445916,
-0.007065773,
-0.0209232196,
0.0165549982,
-0.0474559888,
-0.0112181948,
-0.0228632875,
-0.0395015739,
0.0177981853,
-0.0025853768,
0.0031571279,
0.0018784476,
0.0041807843,
-0.0018170518,
-0.0251403674,
0.0429867804,
-0.0220398195,
-0.0047706026,
-0.0154073229,
-0.0258936118,
-0.0868024379,
0.0196372364,
-0.0669745207,
0.0150169618,
-0.0086346259,
-0.0231536701,
0.0229248591,
0.0170809049,
0.0332098193,
0.0714900047,
0.0458091646,
0.0222983919,
0.1105498597,
-0.0078224037,
-0.0598048493,
0.0037871974,
0.0149895502,
0.0223032851,
-0.0299565364,
-0.0265704393,
0.0156227387,
0.0484180227,
0.042095542,
0.0204784516,
-0.0477204584,
0.0293537341,
-0.0327635221,
-0.0868638977,
0.0145508125,
0.0001613835,
-0.012109654,
0.0735370517,
-0.0089086965,
-0.0667226985,
0.1172669232,
-0.0029446075,
-0.0025985441,
-0.0027342471,
0.002697736,
0.0346425064,
0.0284257699,
-0.0265087523,
0.0124585647,
0.0014861709,
0.0077743158,
-0.0333281942,
-0.0349081308,
0.0012963807,
-0.0635845214,
0.0206293669,
-0.0430571139,
-0.0174850021,
0.0425509624,
-0.0337431915,
-0.0874958038,
-0.0526785888,
-0.0315902941,
-0.0514178202,
-0.060045816,
0.007846795,
-0.0082362657,
-0.0268059298,
0.0445982143,
-0.0008714225,
-0.0185684655,
-0.0549151115,
0.0312471334,
-0.0193708912,
0.0720808432,
0.0083821248,
0.0706234947,
-0.0153939622,
0.0173401572,
0.0682240799,
-0.001633413,
0.0546881817,
0.0009041678,
-0.0165748391,
0.014420433,
-0.0115754604,
-0.0020945107,
0.0193394106,
0.0381076038,
-0.0436414555,
-0.0211247224,
0.1104258075,
0.0342485346,
-0.0340847597,
0.0126214894,
-0.0655985475,
0.0473363586,
0.021755036,
0.0128798001,
-0.0924797803,
0.0045567011,
-0.0118688028,
0.0153354341,
-0.0332977176,
-0.0476702489,
-0.0095547307,
-0.093624562,
0.0281239059,
0.0186826494,
0.0018992041,
-0.0248858463,
-0.0120305223,
0.0055156853,
0.0125755686,
0.0124883316,
-0.010376689,
0.0167417806,
-0.0387856811,
0.0036923992,
-0.0240831058,
-0.0194379501,
0.0456943065,
-0.0277866442,
0.0791735798,
0.0859551653,
0.0538089685,
0.010411391,
-0.0898434296,
-0.0471633524,
0.0395527519,
-0.0038349801,
-0.0010982748,
0.0469755344,
-0.015810268,
0.0103183938,
-0.0021702086,
0.0251406841,
-0.009549737,
-0.0267489552,
0.0141533278,
0.0565752685,
-0.0484080501,
-0.0754446164,
0.0141665181,
0.0544195175,
0.0419608876,
0.0820749477,
-0.0000076176,
-0.0305832606,
-0.0200483985,
-0.0298896916,
0.0420658737,
-0.0085092168,
-0.0509096719,
0.0319540873,
0.0677311867,
0.0026404257,
-0.0941739753,
-0.0448482744,
0.0007826654,
0.0299981683,
-0.0006639113,
0.0285551101,
-0.0183247793,
-0.0090561006,
-0.0379867069,
-0.0123065151,
-0.0536687672,
0.0160248708,
-0.0001367596,
-0.0071179126,
0.0051107937,
-0.0500449538,
0.0216339678,
0.0094529213,
-0.0666246414,
-0.0579264089,
0.0196799655,
0.0807575285,
0.0220044814,
0.0862813219,
-0.0271410216,
-0.0342417471,
0.0600944906,
0.0385149904,
-0.0488734841,
0.0156581663,
0.0389204323,
0.0395773612,
0.0080128266,
0.0250223465,
-0.0193488449,
-0.0246287473,
0.001045104,
0.0058567589,
0.0097107952,
-0.0103354594,
0.0380356088,
-0.000410612,
-0.0056717047,
-0.0467801541,
0.0415331014,
-0.0421011001,
-0.0094902767,
-0.0376208387,
-0.013717927,
-0.0028766105,
-0.0136512462,
0.0523598529,
-0.0099454727,
-0.0055932696,
-0.0515540987,
-0.0164024327,
-0.0567385517,
-0.0091618998,
-0.044608254,
0.0282356422,
0.0206188783,
0.0198852941,
-0.0102102375,
-0.0389354117,
0.0137414774,
0.008683241,
0.0638934746,
0.0092463205,
0.081843704,
0.0084387809,
-0.0679183304,
0.0044710333,
0.0210179668,
-0.0035042521,
-0.0562171005,
-0.0106748417,
0.0231215935,
0.0410635136,
0.0150463535,
0.0529613309,
-0.0547029451,
0.025174493,
0.0074530682,
0.0112502677,
0.0060447743,
-0.0522356927,
-0.0315575004,
0.0138718067,
-0.0420676433,
-0.0056111217,
0.0483509973,
-0.0338386558,
0.0292707663,
-0.0229531154,
-0.0159610007,
-0.0592393465,
-0.0239282157,
-0.0385711342,
0.0175599568,
-0.0432208106,
-0.0006170063,
-0.0705793425,
-0.070073396,
0.0117189065,
0.0107970163,
0.0119428951,
-0.0250304472,
0.0201855097,
0.058411926,
-0.0168321356,
-0.0161263999,
-0.0387898833,
-0.0398640633,
0.0360755622,
-0.0099784359,
-0.0206318572,
-6.701842269e-33,
0.0076135434,
-0.0302027632,
0.021381883,
-0.0358973555,
-0.0149448812,
-0.0175385755,
-0.0096341893,
0.0038155888,
0.029687224,
0.0128564546,
0.0040588793,
0.0374069363,
0.0116299745,
0.0018074728,
0.0402761474,
-0.0048381453,
0.0403407589,
0.0150337322,
-0.0296441019,
0.0065311659,
0.0128180059,
0.0009295875,
0.087149635,
-0.0674003586,
0.0403217673,
0.0278098341,
-0.050001096,
0.0207300242,
-0.0533503816,
0.0174881741,
-0.0542272814,
0.0131532243,
-0.0010871443,
-0.03566017,
-0.0078830933,
-0.0302516147,
-0.0029200329,
-0.0393634029,
0.0049019717,
0.0153523032,
-0.0429182984,
-0.0116396975,
0.0152921993,
-0.0328138284,
0.0062737078,
-0.0025467796,
0.0142592359,
-0.0122094499,
-0.0048781005,
0.0213278122,
-0.0532268807,
-0.0297158957,
0.0150957815,
-0.0435694829,
0.0049029817,
0.0290184338,
-0.0088286307,
-0.0559795648,
-0.1058590859,
0.0442938581,
-0.0287290495,
-0.0036836248,
0.0570894293,
-0.0101077743,
-0.0033085789,
0.0989784971,
0.0536896773,
0.0245504323,
-0.0207097307,
0.0269989315,
0.0509783961,
0.0330027491,
0.0332762226,
0.0460798293,
0.0401994698,
-0.0260008872,
0.0015144936,
-0.0015368696,
0.0016718578,
-0.0091494462,
0.068824403,
0.0023926976,
-0.0384891964,
-0.0175549053,
-0.0065629371,
0.0001544457,
-0.0215000156,
-0.0249898452,
0.0349368304,
0.03994973,
-0.0092813158,
-0.0158769432,
-0.0116883647,
-0.0290914066,
0.0024685652,
0.0008095665,
-0.0424825661,
0.0090730954,
-0.0138530172,
-0.0073561575,
-0.0286078919,
0.0304355305,
0.0147293136,
0.0074974946,
0.0528198704,
0.0285925083,
0.0275050811,
0.0233389419,
-0.0621864647,
0.0150370626,
-0.0344960503,
0.0622656047,
0.0331713445,
-0.069862619,
0.0344459377,
-0.0100463377,
-0.0411340073,
-0.002114004,
-0.0029186893,
0.0120142018,
-0.0187007319,
0.0301662963,
-0.0219375994,
0.0155486967,
-0.000704858,
-0.0122797554,
-0.0272893924,
-0.0454861932,
-0.0086336471,
-0.0483582951,
-0.0065081264,
-0.0105808433,
3.076e-7,
-0.0619928911,
0.0485827066,
-0.0081905648,
0.0329940803,
-0.0058348309,
0.0388876498,
0.0084553109,
0.0226296131,
0.0392914303,
-0.092316255,
-0.0005176383,
0.0128988437,
0.0711333454,
0.0537724569,
-0.0975495577,
-0.0286579765,
-0.0110455491,
0.0047843852,
-0.0496549383,
-0.0756833255,
0.0380309671,
0.0276940539,
0.0687495992,
-0.0242074262,
0.0389215462,
-0.0138247376,
0.057010036,
-0.0099551678,
0.0194659792,
0.0298781041,
0.0655104294,
0.0015527749,
0.0168059058,
0.0317368731,
-0.0078551685,
-0.043567244,
-0.0321215168,
0.0111207291,
-0.0108514279,
-0.0486479029,
-0.0285493433,
0.0366934799,
0.0157129001,
-0.0643409267,
0.0202926211,
-0.0306736808,
-0.0100193927,
-0.0213866681,
0.0117885927,
-0.0451239012,
0.057011839,
-0.0122465184,
-0.0470366552,
0.0207257885,
0.0127505651,
0.0565013103,
0.0078438269,
-0.0481375568,
-0.0205065701,
0.0209963061,
-0.0122546777,
-0.002244815,
-0.0141344462,
0.0236033499,
-0.0498807728,
-0.089591451,
0.0260350518,
2.632979924e-34,
0.0101995794,
-0.0022122117,
0.0335797332,
0.0205296278,
0.0425670147,
0.034219861,
0.0522704981,
0.000973961,
0.0227551423,
0.0223669764,
-0.0095275734
] |
http://arxiv.org/abs/2401.15439v1 | http://arxiv.org/pdf/2401.15439v1 | Pre-training and Diagnosing Knowledge Base Completion Models | 2024-01-27T15:20:43 | 2024-01-27T15:20:43 | Vid Kocijan, Myeongjun Erik Jang, Thomas Lukasiewicz | cs.CL | cs.CL | In this work, we introduce and analyze an approach to knowledge transfer from
one collection of facts to another without the need for entity or relation
matching. The method works for both canonicalized knowledge bases and
uncanonicalized or open knowledge bases, i.e., knowledge bases where more than
one copy of a real-world entity or relation may exist. The main contribution is
a method that can make use of large-scale pre-training on facts, which were
collected from unstructured text, to improve predictions on structured data
from a specific domain. The introduced method is most impactful on small
datasets such as ReVerb20k, where a 6% absolute increase of mean reciprocal
rank and 65% relative decrease of mean rank over the previously best method was
achieved, despite not relying on large pre-trained models like Bert. To
understand the obtained pre-trained models better, we then introduce a novel
dataset for the analysis of pre-trained models for Open Knowledge Base
Completion, called Doge (Diagnostics of Open knowledge Graph Embeddings). It
consists of 6 subsets and is designed to measure multiple properties of a
pre-trained model: robustness against synonyms, ability to perform deductive
reasoning, presence of gender stereotypes, consistency with reverse relations,
and coverage of different areas of general knowledge. Using the introduced
dataset, we show that the existing OKBC models lack consistency in the presence
of synonyms and inverse relations and are unable to perform deductive
reasoning. Moreover, their predictions often align with gender stereotypes,
which persist even when presented with counterevidence. We additionally
investigate the role of pre-trained word embeddings and demonstrate that
avoiding biased word embeddings is not a sufficient measure to prevent biased
behavior of OKBC models. | Pre-training and Diagnosing Knowledge Base Completion Models: In this work, we introduce and analyze an approach to knowledge transfer from
one collection of facts to another without the need for entity or relation
matching. The method works for both canonicalized knowledge bases and
uncanonicalized or open knowledge bases, i.e., knowledge bases where more than
one copy of a real-world entity or relation may exist. The main contribution is
a method that can make use of large-scale pre-training on facts, which were
collected from unstructured text, to improve predictions on structured data
from a specific domain. The introduced method is most impactful on small
datasets such as ReVerb20k, where a 6% absolute increase of mean reciprocal
rank and 65% relative decrease of mean rank over the previously best method was
achieved, despite not relying on large pre-trained models like Bert. To
understand the obtained pre-trained models better, we then introduce a novel
dataset for the analysis of pre-trained models for Open Knowledge Base
Completion, called Doge (Diagnostics of Open knowledge Graph Embeddings). It
consists of 6 subsets and is designed to measure multiple properties of a
pre-trained model: robustness against synonyms, ability to perform deductive
reasoning, presence of gender stereotypes, consistency with reverse relations,
and coverage of different areas of general knowledge. Using the introduced
dataset, we show that the existing OKBC models lack consistency in the presence
of synonyms and inverse relations and are unable to perform deductive
reasoning. Moreover, their predictions often align with gender stereotypes,
which persist even when presented with counterevidence. We additionally
investigate the role of pre-trained word embeddings and demonstrate that
avoiding biased word embeddings is not a sufficient measure to prevent biased
behavior of OKBC models. | [
0.038657371,
0.0510108247,
0.0095271906,
0.0249569714,
-0.0051654093,
0.0028130319,
-0.0360983163,
0.0040517072,
0.0501837023,
-0.0387693457,
0.0087135658,
0.0077326046,
-0.0096531035,
-0.0056551187,
0.0054036998,
-0.0007304727,
0.0375969037,
0.0451132245,
-0.0269584991,
-0.0343359448,
-0.0307951737,
0.0138804931,
-0.0008934595,
-0.0111648692,
0.0076385909,
-0.0186334699,
0.0238229949,
-0.0388724543,
-0.0023551274,
0.027349541,
-0.0052409475,
0.0581820756,
0.0246957168,
-0.039215941,
0.0000020321,
-0.0720951408,
0.0185796153,
-0.0003795539,
-0.0388676301,
-0.0183807444,
-0.014439879,
-0.0241071191,
0.0278318431,
0.0061748149,
-0.0230772141,
-0.0609812737,
0.0215437803,
0.0727160051,
0.0195448529,
0.0288766585,
-0.0037388119,
-0.0369667374,
0.00840382,
-0.0152526498,
0.0116261262,
-0.0152850524,
0.0577112138,
0.0096454024,
-0.0235504266,
-0.0112457639,
-0.0357669145,
0.0000590386,
-0.016247699,
0.0289713256,
0.0562768839,
0.0426830538,
-0.0411470234,
-0.0081618084,
0.0605022237,
0.1086221263,
-0.0064918022,
-0.0112951975,
0.0096700722,
0.0291574728,
0.002093757,
0.0564287342,
-0.0031661924,
-0.0040919515,
-0.0243910048,
-0.0074010259,
-0.0274165738,
-0.0186257791,
0.0506584384,
0.0130070774,
-0.0606175587,
0.0807851627,
0.0097406842,
-0.0568254478,
0.0388592631,
0.0143171418,
-0.0350045152,
-0.067848444,
-0.0174126942,
-0.0099832285,
0.0201403815,
0.0501467697,
-0.0522001646,
-0.006549886,
0.0038518806,
-0.0378744863,
-0.0265949331,
0.0295086093,
-0.0204664357,
0.0704321191,
-0.0843499079,
0.0118980305,
-0.0155962044,
0.0264488198,
-0.0251959413,
-0.0370127335,
-0.0540487431,
0.022750644,
-0.0211630687,
0.0448755622,
0.0143756596,
-0.0075322003,
-0.015286102,
0.0136750164,
-0.0332095698,
0.0455321893,
0.0316599794,
0.0102638751,
-0.0076755104,
-0.0212223995,
-0.0575111024,
0.0033612638,
-0.0353053659,
0.0112486435,
-0.0607426427,
-0.006768337,
-0.0134040294,
-0.0395167917,
0.017152736,
-0.0338398814,
0.0111254249,
0.016132839,
0.032501597,
0.0062412564,
-0.0499871448,
-0.0341459997,
-0.0073916079,
-0.0506995395,
0.0165189318,
-0.0446157753,
0.0407935418,
0.0460427962,
-0.0066582412,
-0.0235180072,
0.0067520658,
0.0009530833,
-0.0993352085,
0.088538982,
-0.0226156935,
-0.0031005649,
0.0038581267,
0.0105477925,
-0.0447143391,
0.0615454838,
0.0032150184,
-0.0079078544,
0.0485311821,
-0.0022645192,
-0.0632040575,
-0.001152434,
0.0871504471,
0.0055601159,
0.0244544782,
0.0462415516,
0.0067800023,
0.0531183407,
0.0450857691,
0.0092667807,
-0.0271146167,
-0.0224756151,
0.0167498309,
0.0056909863,
0.0289364979,
0.0598637387,
0.0237759687,
-0.0077563054,
0.0159542765,
0.0410311185,
-0.0523423478,
-0.0221155174,
-0.0281724185,
0.0112186866,
0.0027478042,
0.0260150861,
-0.0203019157,
0.0066394582,
-0.0033024286,
-0.031119911,
-0.0009955842,
-0.0188064966,
0.0805563256,
0.0521034263,
-0.0175311491,
0.0700039342,
0.0420714095,
-0.0254289303,
0.0060611507,
0.0065638851,
0.0374139957,
0.0276517477,
0.06345395,
-0.039873682,
-0.0326968767,
-0.0249592084,
-0.0175987128,
-0.0123900678,
0.0183622688,
-0.0248411726,
-0.0120137352,
-0.0438082777,
-0.0417053811,
0.0126344217,
0.0306132436,
0.0148349013,
0.0272419639,
0.0502371341,
0.0295213163,
-0.0794532523,
0.0232194252,
0.0330293179,
-0.0404853523,
-0.0069380202,
0.0017754543,
0.0433744565,
0.031506855,
0.0261678621,
-0.0140638025,
0.0003035351,
0.0298094265,
-0.0150910402,
-0.0596556924,
0.0323356166,
-0.0178429782,
-0.0357089862,
0.0202321131,
-0.0385677814,
-0.0463697389,
0.0288733002,
-0.0325151347,
-0.0269286577,
0.040423885,
0.0214828979,
0.0844512805,
-0.0331834704,
0.0084649231,
0.0092907446,
-0.0236858446,
0.023216879,
0.1004579365,
-0.0172132794,
0.033012107,
0.0455286168,
-0.0447785556,
0.0544119962,
0.0113489199,
-0.0552995391,
0.0167848431,
0.0128532732,
-0.0132893184,
0.0081267534,
-0.0689550564,
-0.0652404875,
0.0199333839,
-0.1021594778,
0.0206703171,
0.0125732571,
0.0049142274,
0.0053411354,
-0.0023628201,
0.0145935724,
-0.0091746291,
-0.0001466865,
-0.0064851861,
-0.0248710085,
0.030896401,
0.0154605117,
-0.0017801889,
0.0179579202,
0.006261724,
-0.0213411767,
-0.0199299939,
0.070616506,
-0.0230520461,
0.0231499132,
-0.0027957254,
0.0597826466,
-0.0546993315,
0.0310073271,
-0.0139228068,
-0.0056410208,
-0.023392342,
0.0031447178,
0.0358715169,
0.0471161008,
0.0012579112,
-0.0255342368,
0.0406489596,
0.0042801606,
-0.0480629914,
-0.011194909,
-0.0190749709,
-0.0023471413,
0.0053087855,
-0.0038427832,
-0.0562441573,
0.0123873064,
0.0114255175,
-0.0538637526,
-0.0695448369,
-0.0073152795,
0.0027616401,
-0.0115393605,
-0.0710251927,
-0.0058297557,
0.0161850229,
0.0112769092,
-0.0391507261,
0.0081024328,
0.064548865,
0.0071140365,
0.0320930816,
0.005104884,
0.0272700358,
-0.0053729028,
0.0397969335,
-0.0055813449,
0.0145820184,
0.0068449881,
0.0161592886,
-0.0728937164,
-0.0043059071,
0.0357197523,
-0.0186202955,
0.0039922572,
-0.0037340771,
0.0138035901,
0.0390980877,
-0.0106878579,
-0.058674708,
-0.0314996094,
0.0053326455,
0.0031061263,
0.0177857634,
-0.029586086,
0.0011558948,
-0.0128724743,
-0.0035323224,
0.0111247944,
0.0213555843,
-0.0006494217,
-0.0196477398,
0.0112130791,
0.0366584808,
-0.0250825267,
0.0162681006,
0.0115727559,
-0.0075065158,
0.0214806721,
0.0241406076,
-0.0339680538,
-0.0104404902,
0.0117375832,
-0.0045057382,
-0.0730187744,
-0.0413166471,
-0.0269789714,
-0.0292021055,
-0.0012397817,
0.0047831642,
0.10273204,
0.0095622586,
0.0162836313,
-0.0051823431,
-0.0000989807,
-0.0154050719,
0.0157741904,
-0.0133859627,
-0.0994971991,
0.0028396305,
0.0078821378,
-0.0009523124,
0.0433060154,
-0.080092147,
-0.0096189417,
-0.12907888,
-0.0348979123,
-0.0101810033,
-0.0165392384,
-0.0602880828,
-0.0168825332,
0.0172873326,
-0.0337941013,
-0.0298656598,
-0.0160986762,
0.0150552001,
0.0146624213,
-0.0238740016,
-0.026188083,
-0.0447079018,
0.0471267179,
-0.0952764824,
0.0709413812,
0.0081935506,
0.042608358,
-0.0073902234,
-0.0109928837,
-0.0403349176,
-0.0301832147,
-0.0337829478,
0.0136218341,
0.0794862658,
-0.0555634275,
-0.0084463228,
-0.0282120965,
-0.0056179888,
-0.0025087202,
0.0253311079,
-0.0139446948,
0.0877305791,
0.0235498659,
-0.0423111133,
-0.0212505385,
-0.0144213205,
-0.0007843501,
0.1010036319,
-0.0316396654,
0.0118006496,
-0.0056601991,
0.016782118,
0.0165428538,
0.0104737291,
0.0188793801,
0.0367863588,
-0.0601953603,
-0.0125204427,
-0.069024533,
-0.0440764576,
0.0583664291,
0.0083402758,
0.0064148302,
-0.001450656,
0.0337935872,
-0.0116119366,
-0.0676214322,
0.0465882942,
-0.0092585227,
-0.0049696886,
-0.0025442978,
0.0007391688,
0.0194519032,
-0.0152591374,
-0.0463138148,
-0.0283024795,
0.0137229441,
-0.0382590033,
0.0134178195,
0.0155218877,
0.0158440713,
0.0747739077,
-0.0605542287,
-0.0295248535,
0.0543948002,
0.0256872606,
-0.0447204001,
-0.0027858419,
0.0583421588,
-0.0184991863,
0.0320782438,
0.0630193278,
-0.0418953821,
-0.005223576,
-0.0034235434,
0.0193560477,
-0.0118449386,
-0.0608134605,
0.0157611892,
0.021857867,
0.0324619897,
0.0101651503,
0.0173949804,
-0.0645238832,
0.0172596779,
-0.020573698,
-0.0191560034,
-0.0333523937,
-0.0087058702,
0.0494945087,
0.0107619539,
0.0043261466,
0.0013699288,
-0.0256547425,
0.0492150187,
-0.0184367243,
-0.0445675142,
0.0105713764,
0.0110668577,
0.0495848954,
-0.0134098129,
-0.0376854278,
-0.0168953966,
0.0108901067,
-0.0037412948,
-0.0257433224,
0.0029861224,
0.020787375,
0.0123664886,
0.0129321739,
-0.0029863217,
-0.0294948295,
-0.0021529796,
-0.0422371812,
0.0447221659,
0.0335191116,
-0.0314885229,
0.0142221358,
-0.0309676528,
-0.0317300074,
0.002241482,
0.0049693254,
-0.0273213089,
0.0173087977,
0.0400948897,
0.0167275481,
-0.019894544,
0.0266501438,
0.0701448247,
0.018483324,
0.0898025557,
-0.0482246354,
-0.0431210846,
-0.0195599198,
-0.0195494797,
0.0475632809,
0.0369321667,
-0.0188005026,
0.0382423066,
-0.0669165552,
-0.0992037728,
0.0113721695,
-0.0236333907,
-0.0628840178,
0.0275052916,
0.033070907,
0.0379059203,
0.0279337484,
-0.0447846279,
-0.0147758201,
0.0073096533,
-0.0226848647,
-0.0093462495,
0.0149038406,
-6.067486743e-33,
-0.0408356711,
-0.0092196837,
0.002168458,
-0.0383241214,
-0.0351420678,
-0.0231039319,
0.0200408529,
-0.017663531,
-0.0483728349,
0.0528830066,
-0.0191315245,
0.0455728024,
-0.0139538972,
0.0007428889,
-0.0079424987,
-0.0025913594,
0.0485499948,
0.0255889352,
-0.0051891836,
-0.0276078694,
0.0561510213,
0.0646844432,
0.0596060529,
-0.0871555731,
0.0459066592,
-0.0274248812,
0.002851974,
-0.000998587,
0.0069431062,
0.0270217545,
-0.012954847,
0.0547644272,
0.0193918403,
-0.0175588559,
-0.0055178669,
0.0272075795,
-0.138311699,
-0.0700606853,
0.0465598628,
0.010881329,
-0.0094952788,
0.0225194246,
0.0325404182,
0.0074643954,
0.0390913635,
-0.0373435579,
0.0359482877,
-0.0256941896,
-0.0100311805,
0.0093310094,
-0.0235094782,
0.0158434007,
0.0372601636,
-0.0027293682,
0.0037668718,
-0.0001992059,
0.0186070669,
-0.0266285408,
-0.0919347852,
0.017726941,
-0.0015172673,
-0.0158579294,
0.0266394671,
-0.0326482058,
0.0412272103,
0.0224885717,
-0.0070422343,
-0.0068243132,
-0.0378216431,
-0.028152097,
-0.0054508639,
0.0570542924,
-0.0051956833,
0.0511833988,
0.0372411609,
-0.0005430817,
-0.0638703033,
-0.003149627,
0.0033287818,
0.0273796748,
-0.0191440824,
-0.0280258674,
-0.0601504147,
-0.0190547854,
-0.0437402129,
0.0042092069,
-0.0153481327,
-0.0536945947,
0.0139732137,
0.0042649894,
-0.0256062429,
-0.0076236427,
-0.0252739917,
-0.0714701936,
-0.0118816923,
0.0360293537,
-0.0014449869,
-0.0378763452,
0.0149219614,
0.033636108,
-0.0700074285,
-0.0217203908,
0.0431865603,
-0.0178308655,
-0.0004710219,
0.0533118397,
0.0211346969,
0.0013323094,
-0.0441916212,
-0.0179732386,
-0.0117021408,
-0.0492366031,
-0.0157190021,
0.0035661287,
0.0394712761,
0.0487912744,
0.0028373764,
0.0315206088,
0.0816019475,
0.065824464,
-0.017741954,
-0.0220728535,
0.0496886745,
-0.0043219971,
0.0121829836,
-0.0263871178,
-0.0065231957,
-0.002640272,
0.015206947,
-0.0418983884,
0.0046369694,
-0.0046099788,
2.881e-7,
0.0277556237,
0.0127667962,
0.0057722186,
0.0425873026,
-0.0035383385,
0.0290452484,
0.0047396198,
0.0425650962,
0.0264733154,
-0.051162716,
0.0288082156,
-0.0065704058,
-0.0234784633,
0.0246219877,
-0.0919905156,
-0.0424682125,
-0.0325019844,
-0.0399950556,
-0.0217272062,
-0.0461270064,
0.0485840142,
0.0588123649,
0.0771670565,
0.0168535784,
-0.0071846549,
-0.0064827329,
0.0580969341,
-0.0108090611,
-0.006845864,
-0.0172213279,
0.0614548214,
0.0475879572,
-0.0128935454,
0.0912780538,
-0.0158635844,
0.0473776162,
0.0425395928,
0.0023428013,
-0.0016914777,
0.037416108,
-0.0344568379,
-0.0597372316,
0.0022373323,
-0.0494244732,
0.0964665413,
-0.0577124283,
-0.0094845463,
0.0419907756,
-0.0649447814,
-0.0081440834,
0.0442366451,
0.0127548221,
0.0169412512,
-0.0347968824,
0.0315063894,
0.0445448346,
-0.0047137621,
-0.0846991614,
0.0287015587,
0.0448235497,
-0.0362050682,
-0.0794777945,
0.0311417449,
-0.0261213109,
0.0600469485,
-0.0845892876,
0.0383804962,
2.544820144e-34,
0.0119045172,
0.0452443138,
0.0205911081,
-0.0347337462,
0.0265295599,
-0.0621434487,
-0.0105518755,
-0.0271478072,
0.0419641994,
-0.0591973625,
0.0063494048
] |
http://arxiv.org/abs/2402.01726v2 | http://arxiv.org/pdf/2402.01726v2 | AI Does Not Alter Perceptions of Text Messages | 2024-01-27T14:32:12 | 2024-02-07T17:04:31 | N'yoma Diamond | cs.CL | cs.CL, cs.AI, cs.HC | For many people, anxiety, depression, and other social and mental factors can
make composing text messages an active challenge. To remedy this problem, large
language models (LLMs) may yet prove to be the perfect tool to assist users
that would otherwise find texting difficult or stressful. However, despite
rapid uptake in LLM usage, considerations for their assistive usage in text
message composition have not been explored. A primary concern regarding LLM
usage is that poor public sentiment regarding AI introduces the possibility
that its usage may harm perceptions of AI-assisted text messages, making usage
counter-productive. To (in)validate this possibility, we explore how the belief
that a text message did or did not receive AI assistance in composition alters
its perceived tone, clarity, and ability to convey intent. In this study, we
survey the perceptions of 26 participants on 18 randomly labeled pre-composed
text messages. In analyzing the participants' ratings of message tone, clarity,
and ability to convey intent, we find that there is no statistically
significant evidence that the belief that AI is utilized alters recipient
perceptions. This provides hopeful evidence that LLM-based text message
composition assistance can be implemented without the risk of
counter-productive outcomes. | AI Does Not Alter Perceptions of Text Messages: For many people, anxiety, depression, and other social and mental factors can
make composing text messages an active challenge. To remedy this problem, large
language models (LLMs) may yet prove to be the perfect tool to assist users
that would otherwise find texting difficult or stressful. However, despite
rapid uptake in LLM usage, considerations for their assistive usage in text
message composition have not been explored. A primary concern regarding LLM
usage is that poor public sentiment regarding AI introduces the possibility
that its usage may harm perceptions of AI-assisted text messages, making usage
counter-productive. To (in)validate this possibility, we explore how the belief
that a text message did or did not receive AI assistance in composition alters
its perceived tone, clarity, and ability to convey intent. In this study, we
survey the perceptions of 26 participants on 18 randomly labeled pre-composed
text messages. In analyzing the participants' ratings of message tone, clarity,
and ability to convey intent, we find that there is no statistically
significant evidence that the belief that AI is utilized alters recipient
perceptions. This provides hopeful evidence that LLM-based text message
composition assistance can be implemented without the risk of
counter-productive outcomes. | [
0.071211949,
0.0049754311,
-0.0567795858,
-0.0417910516,
-0.0549199022,
-0.0005885535,
0.064634189,
-0.000536046,
0.0409238935,
-0.0610246696,
-0.0030751023,
-0.0443646088,
0.0202825759,
0.0127269253,
0.0112062553,
-0.0256111473,
-0.0295233261,
-0.0000988836,
0.0626945049,
0.0303393938,
0.0405183248,
0.048040472,
0.018018771,
0.0517071411,
-0.0476653799,
-0.0309387855,
0.0037385717,
0.0486303382,
-0.0222329758,
-0.0457089581,
-0.0151872281,
0.0616260991,
0.011281007,
-0.0016266328,
0.0000019571,
-0.0403981246,
0.0282383598,
-0.0156527366,
-0.019822713,
0.0308162794,
0.0915243849,
0.0191930681,
-0.0137689393,
0.040240746,
-0.0395806842,
0.0620609298,
0.0437038429,
0.0233403016,
0.0388734601,
0.0413948894,
-0.0343634337,
0.0186874513,
0.0160198342,
-0.0252573937,
0.0170940217,
-0.0003562769,
0.0065648975,
-0.0446557961,
0.0182972625,
0.0060161087,
0.0421860665,
0.0652705133,
0.0109203346,
-0.0166692026,
0.0171908475,
0.0190293565,
-0.0034389428,
-0.0757780895,
-0.0157643706,
0.0406997278,
0.0747727528,
-0.022843359,
-0.0028926516,
-0.03843024,
0.0342107005,
0.006285253,
-0.0265172683,
0.002037355,
-0.0029712876,
0.0035990286,
-0.00867505,
0.0167175662,
0.0205777604,
-0.0032720736,
-0.0156962518,
-0.0075340327,
-0.0154214287,
-0.0634939447,
-0.0098069366,
0.0169984642,
0.0018443601,
-0.072005406,
0.0589316115,
0.010205104,
0.0143616889,
0.0239315964,
-0.0272975005,
-0.0824239776,
0.0039799358,
-0.0513940565,
0.0342074446,
0.0333329625,
-0.0175457262,
-0.0176843852,
0.0224222802,
0.0022660589,
0.0195944495,
-0.0055113914,
-0.0029636924,
0.0304011814,
-0.0273204204,
-0.0412868783,
-0.0477012843,
-0.0000463703,
0.006437086,
-0.0288741738,
-0.0496576615,
0.0079072025,
-0.01559602,
-0.0328933969,
-0.0087452233,
-0.0196015667,
-0.006603641,
0.014874557,
0.0173795819,
0.0098469481,
-0.0304282811,
0.064513281,
0.0151633685,
-0.0495910831,
0.0433114059,
-0.0020060358,
0.0067668008,
0.0035639859,
-0.0102128536,
0.0270368773,
-0.0022834695,
-0.0541856997,
-0.022309456,
-0.0221140124,
-0.0406015292,
-0.0468439497,
-0.034306787,
-0.0368171744,
-0.0064331843,
0.0384929553,
-0.0116625046,
-0.0246896502,
0.0179179888,
0.021064708,
-0.0635859892,
-0.0051171524,
-0.0367889777,
0.0597564429,
0.0247048587,
0.0372226499,
0.0078702988,
0.0220850129,
0.0486794934,
0.0550107881,
0.0261374041,
0.0036590071,
0.047054965,
0.0441389047,
-0.0164532941,
0.0166214798,
0.0020638187,
-0.0496046506,
-0.081281662,
0.0137367733,
-0.0311911441,
0.0735744536,
0.0136830034,
0.0701842308,
0.1244358048,
0.0712446943,
0.0143466042,
0.042295482,
-0.0019974906,
0.0443159342,
0.0171997137,
0.0010965791,
-0.031604033,
0.018660659,
-0.069916755,
0.0051606232,
0.0456985794,
-0.0052016601,
-0.043715965,
-0.0666271746,
-0.0351723246,
-0.0385303274,
0.0223048907,
0.0196219217,
0.0259452593,
-0.0073691625,
-0.0175107848,
-0.0158898681,
-0.0309644006,
-0.0522086732,
-0.0308898259,
-0.0315807424,
0.0263004042,
0.083897382,
0.0072585708,
-0.0116724083,
0.047943633,
-0.0292105153,
0.0262594838,
0.0483309329,
0.0075800503,
0.025500847,
-0.0204496514,
-0.0068802764,
0.0313281566,
-0.0021988496,
-0.0007230691,
-0.0548488423,
-0.021274602,
0.0477050245,
0.0183968935,
0.053905122,
-0.0363457799,
-0.0055576828,
0.0195139814,
-0.0421234556,
-0.0028280194,
0.0150611736,
-0.007624683,
0.0618554242,
-0.0242779329,
0.0259131882,
-0.0093577001,
-0.00431976,
-0.0316152088,
0.0227525644,
-0.0441490859,
0.0504851826,
-0.0046224813,
-0.0038647966,
-0.0343362428,
0.0070072743,
0.0095488019,
-0.0096800383,
0.0411466099,
-0.0453395173,
0.050556574,
0.003838103,
0.0515692942,
-0.0251927152,
-0.0116398418,
0.0235238951,
-0.0183378328,
0.0001981217,
-0.0456476212,
-0.0184970405,
-0.0080664791,
-0.0265994556,
0.0140035907,
-0.0362564065,
0.0149890985,
-0.0048339786,
0.0423677899,
0.0207926091,
0.0068156989,
-0.0285582822,
-0.0333664007,
-0.0974100009,
-0.0028382358,
-0.0123399291,
-0.0283295084,
0.0570720024,
0.0265509468,
0.003684802,
0.0203449484,
-0.0457530841,
-0.0144226654,
0.0135685066,
0.0413455479,
-0.0005533685,
-0.0213862341,
-0.0030024471,
-0.0467827246,
-0.0109438514,
-0.0451278575,
0.0098020965,
0.0535143241,
-0.0698162019,
-0.0334551744,
0.0556576587,
-0.0316970721,
-0.008669205,
0.0074272836,
-0.0654820874,
0.0447611883,
-0.002121852,
0.0475344062,
0.0635016412,
0.0084233638,
0.0062641161,
0.0409570858,
0.0296471864,
-0.0189171489,
-0.0145612517,
0.0149438558,
-0.0143886367,
-0.0063812868,
-0.0934301391,
0.0166135821,
0.0483414121,
0.03437379,
0.0427976437,
-0.0247487389,
0.0133637898,
-0.0455600247,
-0.0550602861,
-0.0014503998,
-0.027767621,
-0.0330590233,
0.0259065907,
0.0022468239,
-0.0826998726,
0.0291323196,
0.0100449491,
-0.0193696674,
-0.0028169586,
-0.0341612622,
0.0060076402,
0.0217351839,
-0.0477270149,
0.0469156802,
0.0421898142,
0.0025136631,
0.0018110528,
-0.0276930053,
-0.0224068798,
-0.0279943161,
0.0177316852,
-0.0405665673,
-0.0230542887,
0.0509651378,
-0.0527656041,
0.0185699854,
-0.0625284761,
-0.0591316074,
-0.0097991405,
0.0511767268,
-0.0210189,
0.0000433682,
-0.0020143532,
0.0629409105,
-0.0072190631,
-0.0118921334,
-0.0130048916,
0.007843269,
-0.0224132109,
0.0310118645,
-0.0552460328,
0.0302046314,
-0.0379247665,
0.0468643196,
0.0212177169,
-0.0075713643,
0.0020592005,
-0.0028870883,
-0.0289072823,
0.0036324919,
0.0279870555,
-0.0375712514,
0.0570575558,
0.0683240592,
-0.0020163702,
0.0273169149,
0.0855235457,
0.0263464916,
-0.0654802099,
0.0140635939,
-0.0285226461,
-0.0092718471,
0.0088025499,
-0.0124461958,
-0.0531801023,
-0.004998663,
0.0184901897,
-0.0026085153,
-0.0535524637,
0.0220006984,
-0.0003569359,
-0.0593108721,
0.0137382485,
0.036621239,
0.0118724247,
0.0301501118,
0.0021418235,
-0.0036302872,
0.0466129221,
0.0040716915,
-0.0359062403,
-0.0178957116,
-0.012549527,
-0.0123837115,
-0.1093572751,
-0.0417462289,
-0.000687167,
-0.0299036764,
0.0625444502,
0.0644633174,
0.0601268187,
-0.0146978907,
-0.027412042,
0.0225384012,
0.0629899129,
0.0440415069,
0.0035465949,
0.0431574956,
-0.0029236129,
0.0261356644,
-0.0871385038,
0.0260010026,
-0.0341062583,
-0.0099048382,
0.0224950984,
0.0328625515,
0.0051646028,
-0.0293595269,
-0.0228946917,
-0.0004060922,
0.0543529205,
0.0460781418,
0.0486041792,
-0.0926792398,
-0.0092348708,
-0.0164630916,
0.0024956965,
0.0097369067,
-0.0308484286,
0.0114143007,
0.0007857536,
-0.0155001273,
-0.0511230677,
-0.0591027588,
0.0204122216,
-0.0073987725,
0.0193647053,
0.0167064741,
0.0185245126,
-0.0134373233,
0.0044262223,
-0.0301300008,
0.0147445137,
0.046993833,
-0.0225973111,
-0.0016473631,
-0.0300552417,
-0.025423022,
0.0643733144,
-0.0067667672,
-0.0477554463,
-0.055029206,
0.0196948014,
0.0441769548,
0.025213575,
-0.0028620488,
-0.0193293132,
-0.0972163305,
0.0112849027,
0.0219841339,
-0.0178020727,
0.0527336523,
0.0204634368,
0.0311818123,
0.0310665164,
0.0144365523,
-0.0089040156,
0.0209586211,
0.0396869369,
0.0045143133,
-0.0303284787,
-0.0304610319,
0.0191184022,
0.0545779467,
-0.0019272565,
-0.0028729057,
0.0149539197,
0.0039061974,
-0.027282998,
-0.0503398292,
-0.0204903074,
-0.0212429706,
-0.0113218548,
-0.0113616213,
0.0295208991,
-0.0203393195,
-0.0348124355,
-0.0400810093,
-0.0101916967,
-0.0019281525,
0.0460470915,
-0.0235355925,
0.030970905,
0.0137649113,
0.0174128506,
-0.0330652632,
0.0550487265,
-0.0368794762,
-0.0136208478,
0.0376765467,
0.0548162982,
-0.0249758773,
-0.017663924,
0.0109949028,
0.0518407412,
-0.0362644866,
-0.07286302,
0.0072365645,
0.0252803061,
0.0404919758,
-0.0170979965,
0.051106941,
-0.0151985502,
0.0140305972,
-0.0480546579,
0.03939417,
-0.0167903267,
-0.024280116,
-0.033740472,
-0.0296669491,
-0.0434610993,
-0.0034219716,
0.0343146399,
-0.0482204296,
0.0291437563,
-0.0078527918,
-0.032245826,
-0.0222126991,
0.0099850567,
0.0399171896,
0.0227981135,
-0.0109111983,
-0.009801235,
-0.0793524683,
-0.0546112992,
0.0148586212,
-0.0606572367,
0.0271395985,
-0.0733884722,
0.0107518928,
0.0356713906,
-0.0063677896,
0.001962082,
-0.028176425,
-0.0430775657,
-0.054421071,
-0.0068231905,
-0.041687686,
-5.789743876e-33,
0.0389625132,
-0.0176375285,
-0.0375309438,
0.0070238733,
-0.011227875,
0.0243257023,
-0.0323700756,
0.0435925871,
0.0372230262,
0.0265952758,
-0.0333740786,
0.0479980074,
-0.0112567171,
-0.034734793,
0.029877644,
0.042763073,
0.0510381162,
-0.0390882231,
-0.0281065628,
0.0172083192,
0.0154179372,
0.0121541088,
0.0127892438,
-0.1178334951,
0.0473034382,
-0.032559026,
-0.0372349098,
-0.0556800403,
0.0273089595,
0.0229590051,
-0.0126682064,
0.0349017456,
0.0326049663,
0.021790389,
-0.0155881271,
0.0816645995,
0.0518788174,
-0.0408361591,
0.0229842793,
-0.0161582846,
-0.0231189653,
-0.0528018214,
0.0188332479,
-0.0044446089,
-0.0577207915,
0.0335019566,
0.0404043756,
-0.0026593111,
0.028856026,
-0.0376486555,
-0.0428623334,
-0.0333058685,
-0.009883659,
-0.0221769996,
-0.0131756635,
0.0624747463,
-0.0451787785,
0.0223967712,
-0.027968565,
0.0361331515,
-0.0163157322,
-0.0396197736,
0.0214400981,
-0.0225160904,
0.0306470599,
0.0396244042,
0.0035058991,
0.0418764837,
-0.0780120715,
-0.0339897163,
-0.0018573666,
0.0177615732,
0.0516962782,
0.0458634347,
-0.0011402012,
-0.0548625588,
0.0015686085,
-0.0062624668,
-0.0193137322,
-0.043968305,
0.0705624074,
0.0194944292,
-0.0864606425,
-0.0312264729,
-0.0157148857,
-0.0105306571,
-0.0495613031,
-0.0599851236,
0.0266112983,
-0.0236818474,
0.0138949174,
-0.0318269692,
-0.0000291133,
-0.0289801992,
-0.0118801519,
-0.0565740988,
-0.0328761935,
0.0700911209,
-0.0011504808,
-0.0031553202,
-0.0733501688,
0.0356265157,
0.0119411927,
0.073735036,
0.0534596704,
0.051353991,
0.0285266191,
-0.0029320782,
-0.0469571799,
-0.0227943864,
-0.0008247857,
0.0029650924,
0.0479603335,
-0.0516898558,
0.033835534,
0.0221375003,
-0.0273506306,
0.0261451751,
-0.0066415914,
-0.0132774478,
-0.0479719453,
-0.0321335793,
-0.0284141172,
0.002595454,
-0.0509675741,
-0.0149490377,
-0.036019884,
-0.0022779072,
0.0695284456,
-0.0239321701,
-0.0369370393,
-0.0299520344,
2.595e-7,
-0.0168369282,
0.0585465617,
-0.0290987436,
0.0390918739,
0.0253914855,
0.0652483702,
-0.0178760067,
0.0085315565,
0.0335299335,
0.0455284268,
0.0300351921,
0.0189154129,
0.0457654744,
0.0009724267,
-0.0292359106,
0.0116744787,
-0.0755383149,
0.0194135252,
-0.0451222733,
0.034046635,
0.1108372882,
0.0301356558,
0.0355500393,
-0.0115069421,
-0.0277892463,
-0.0370817296,
0.068922624,
-0.0554979332,
0.0336595289,
-0.0092741009,
0.0384392105,
0.0019085265,
0.0023708188,
0.0010831796,
-0.001181132,
-0.0522682033,
0.0032970104,
0.0715685189,
-0.018612111,
0.0264369864,
0.0329823531,
0.024817653,
-0.0039952477,
-0.0400633253,
0.0210459027,
-0.06807331,
-0.0435834602,
-0.0304844994,
0.0058312244,
-0.0438199304,
0.0331376754,
-0.0144672561,
-0.0036830872,
0.0327800959,
-0.006930647,
0.0109060546,
0.0177514851,
-0.0515608229,
0.0027109066,
0.0116805229,
0.0436331667,
-0.0449151881,
-0.0715522021,
0.0698864311,
-0.007170768,
-0.05928123,
0.0367122777,
1.813288543e-34,
0.0313531123,
-0.0301396083,
0.03211952,
0.0426977202,
0.0680523142,
0.0265078526,
0.0042688888,
0.0253113639,
0.0228953492,
-0.0653359517,
0.0142369848
] |
http://arxiv.org/abs/2401.15422v2 | http://arxiv.org/pdf/2401.15422v2 | A Survey on Data Augmentation in Large Model Era | 2024-01-27T14:19:33 | 2024-03-04T16:55:15 | Yue Zhou, Chenlu Guo, Xu Wang, Yi Chang, Yuan Wu | cs.LG | cs.LG, cs.CL, cs.CV | Large models, encompassing large language and diffusion models, have shown
exceptional promise in approximating human-level intelligence, garnering
significant interest from both academic and industrial spheres. However, the
training of these large models necessitates vast quantities of high-quality
data, and with continuous updates to these models, the existing reservoir of
high-quality data may soon be depleted. This challenge has catalyzed a surge in
research focused on data augmentation methods. Leveraging large models, these
data augmentation techniques have outperformed traditional approaches. This
paper offers an exhaustive review of large model-driven data augmentation
methods, adopting a comprehensive perspective. We begin by establishing a
classification of relevant studies into three main categories: image
augmentation, text augmentation, and paired data augmentation. Following this,
we delve into various data post-processing techniques pertinent to large
model-based data augmentation. Our discussion then expands to encompass the
array of applications for these data augmentation methods within natural
language processing, computer vision, and audio signal processing. We proceed
to evaluate the successes and limitations of large model-based data
augmentation across different scenarios. Concluding our review, we highlight
prospective challenges and avenues for future exploration in the field of data
augmentation. Our objective is to furnish researchers with critical insights,
ultimately contributing to the advancement of more sophisticated large models.
We consistently maintain the related open-source materials at:
https://github.com/MLGroup-JLU/LLM-data-aug-survey. | A Survey on Data Augmentation in Large Model Era: Large models, encompassing large language and diffusion models, have shown
exceptional promise in approximating human-level intelligence, garnering
significant interest from both academic and industrial spheres. However, the
training of these large models necessitates vast quantities of high-quality
data, and with continuous updates to these models, the existing reservoir of
high-quality data may soon be depleted. This challenge has catalyzed a surge in
research focused on data augmentation methods. Leveraging large models, these
data augmentation techniques have outperformed traditional approaches. This
paper offers an exhaustive review of large model-driven data augmentation
methods, adopting a comprehensive perspective. We begin by establishing a
classification of relevant studies into three main categories: image
augmentation, text augmentation, and paired data augmentation. Following this,
we delve into various data post-processing techniques pertinent to large
model-based data augmentation. Our discussion then expands to encompass the
array of applications for these data augmentation methods within natural
language processing, computer vision, and audio signal processing. We proceed
to evaluate the successes and limitations of large model-based data
augmentation across different scenarios. Concluding our review, we highlight
prospective challenges and avenues for future exploration in the field of data
augmentation. Our objective is to furnish researchers with critical insights,
ultimately contributing to the advancement of more sophisticated large models.
We consistently maintain the related open-source materials at:
https://github.com/MLGroup-JLU/LLM-data-aug-survey. | [
0.0293631908,
0.112118125,
-0.0088280644,
0.0721561164,
0.0029315506,
-0.0005684081,
0.0028945685,
0.0040504527,
-0.0095481165,
-0.0250763781,
-0.0568992347,
-0.0094183385,
0.004951077,
0.0295370072,
0.0652995035,
-0.086027734,
0.0330709629,
-0.0576287806,
-0.0127974935,
-0.0125670182,
-0.0071253954,
0.0035026257,
0.0076634726,
0.0101539576,
-0.0311267078,
0.0175751243,
-0.0060617854,
0.0418945625,
0.0512289964,
-0.009167213,
-0.0493826158,
0.031528268,
0.0642312765,
0.0527659953,
0.0000020746,
-0.0308700502,
0.0328959562,
0.0315949805,
-0.047845941,
0.0196499135,
0.0477463007,
-0.0649309009,
-0.021600537,
-0.0251953173,
-0.0169648957,
-0.0334470943,
0.0745718628,
0.0582957193,
0.0511287153,
0.0745899156,
-0.0176642705,
-0.0530374423,
0.1156601533,
-0.0374451466,
0.0246400554,
0.013540782,
0.0675529763,
0.0504231043,
-0.000691105,
0.0385986939,
-0.0240350664,
0.0474421643,
0.0201977938,
0.001223582,
0.0796357915,
0.0546934009,
0.0224212371,
-0.0012109408,
-0.0123262629,
0.057499554,
-0.0028722628,
-0.0032986619,
-0.0324810334,
0.0015691043,
0.0043662447,
-0.0101110879,
-0.0429393239,
-0.0403707437,
0.0558206439,
0.0212768745,
-0.0311223269,
-0.0530649573,
0.0715050772,
0.0360836498,
-0.0166326333,
0.0363021493,
0.0166297853,
-0.0389917977,
-0.0149797574,
0.0153311538,
-0.0402458087,
-0.0588099919,
0.0077246046,
0.0138062872,
0.0216780547,
0.00086627,
-0.0561968423,
-0.0661518797,
-0.0091095287,
-0.0142912669,
0.041287601,
0.0098874187,
-0.0347093455,
0.0541740619,
-0.0253607854,
0.0169623513,
-0.0623293594,
0.0170209017,
-0.0420795605,
-0.0008029848,
0.008103027,
-0.0165196452,
-0.0223996919,
0.0074431216,
-0.0493048429,
-0.0343178958,
-0.0644210204,
0.0067336624,
-0.0218135137,
0.0184518192,
-0.0097216349,
0.0124788079,
-0.039894063,
-0.0342790298,
-0.0018803476,
0.0336926319,
-0.0010425269,
0.0792806819,
0.0121490527,
-0.0222107824,
-0.0082065528,
-0.001540013,
0.0486440957,
-0.0167251527,
0.015458243,
0.0561513342,
-0.0176422279,
-0.0438115001,
-0.0375544913,
-0.0070749414,
-0.0061397315,
-0.0306680966,
-0.020936884,
-0.0752843842,
0.0252710115,
-0.0441892259,
-0.0005612897,
-0.0398222022,
0.0004322722,
-0.0147472294,
-0.0519784354,
0.0819406658,
-0.0815689638,
0.0159673914,
0.0330604464,
-0.0097308429,
0.0009713349,
0.0682893321,
0.0741174072,
0.0075767753,
0.0229168441,
-0.0314996168,
0.0013781349,
0.0892573297,
0.0122379148,
0.0423131101,
-0.0415313169,
-0.0226199944,
-0.0068741292,
0.0289318096,
0.0489441641,
0.0341345593,
-0.0668847561,
-0.0009976969,
0.0929862037,
0.0303137582,
0.0090762628,
-0.001492339,
0.0210169591,
-0.0167851523,
0.0169251058,
-0.0097540226,
-0.0405053422,
0.0008082724,
-0.0317514837,
0.0208013672,
0.0305711515,
0.0464450307,
-0.0316208676,
-0.0542728379,
0.0419828147,
-0.0264683329,
0.063187547,
0.0373729654,
0.0631895512,
0.0045183417,
-0.0008873583,
-0.0167722572,
0.0318017341,
-0.0432734862,
0.0180938505,
-0.0476778932,
0.0064511909,
0.0749396756,
0.0465338752,
-0.0288990643,
-0.0138888098,
-0.016851984,
0.0141544091,
0.0402741246,
-0.0189795531,
0.01426508,
0.0334824808,
-0.0529938042,
-0.032885801,
0.0189443212,
0.0107572218,
0.0155100739,
0.0248358194,
0.0002710521,
0.0206054319,
-0.0100722602,
-0.038852565,
0.0071782856,
-0.0572923496,
-0.0013309092,
0.0002602758,
-0.0058699711,
0.0164028071,
0.0420007035,
-0.0226048436,
-0.0182766449,
0.0274634026,
-0.0195678808,
-0.0160296764,
-0.0063263616,
-0.0225630794,
-0.0103000896,
0.0443834253,
-0.0169725101,
-0.000784579,
0.0295898691,
-0.0420359336,
-0.0474923998,
0.0029350198,
-0.0474515669,
0.017854508,
-0.0478759259,
0.0168960411,
-0.0058091311,
0.0467534289,
-0.0046610055,
0.0409808159,
-0.0139409583,
-0.0287489817,
-0.0020965307,
0.0231577884,
-0.0236665718,
-0.0433873609,
-0.0179821961,
0.0056773364,
0.0161864627,
0.0239195749,
0.0055665127,
-0.0444650874,
-0.0584766157,
0.001694433,
-0.0983669907,
-0.0173780583,
-0.0072179982,
0.0051536104,
-0.0109870564,
0.0301638115,
-0.0133790113,
-0.0192334242,
-0.0121695213,
-0.0201154314,
0.0236382075,
0.0063237445,
0.0009310303,
0.0134562459,
-0.0286325086,
-0.0338105671,
-0.0305995345,
0.0257431827,
0.0277251918,
-0.0002488822,
-0.0498582534,
-0.0508157536,
-0.0134906266,
-0.0228229761,
0.0563911386,
-0.0178216174,
-0.023195399,
-0.0322306789,
0.0011588201,
0.0550157353,
0.0495978296,
0.0002126697,
0.0022982091,
0.0191446003,
0.0136650736,
-0.0267326403,
0.0094816368,
-0.0127617298,
0.0260347594,
-0.0596946627,
-0.0374074467,
-0.0314765386,
0.0336535126,
-0.0433004908,
-0.0014092914,
-0.005981707,
0.0058480334,
-0.0015129499,
-0.0668457299,
-0.0191038884,
-0.0420493148,
-0.0259115845,
0.0189810544,
-0.0507482477,
-0.07259579,
0.0239612814,
-0.0464849547,
-0.0224184301,
0.0019045499,
-0.0433410592,
-0.0175336152,
0.042721469,
0.0340451077,
0.0145967128,
-0.0427718498,
0.002808695,
-0.0257270206,
-0.0229163487,
0.0081818029,
-0.0390145257,
0.0359854437,
-0.0361202247,
0.0047204983,
-0.0022574796,
-0.0282501746,
-0.0138615994,
-0.0374359787,
-0.0266256724,
-0.0026673286,
-0.0208061431,
-0.044726029,
-0.0161781907,
-0.0356851891,
0.0252417699,
0.0238021947,
-0.0110136513,
0.0175381545,
-0.0417964607,
0.0563144274,
0.0710670277,
0.0454258993,
-0.0122430837,
-0.0318274647,
0.0232185367,
0.0193017721,
0.0217439011,
-0.0205099285,
-0.0794134811,
0.0053663366,
0.0326248966,
0.0289138928,
-0.0178229362,
-0.0071561378,
0.0756214559,
-0.0645143166,
0.0307291187,
0.078089729,
0.0055301157,
-0.0306539778,
0.00363384,
0.0232598893,
0.0198711418,
0.0091824606,
-0.0197015665,
-0.0785024688,
-0.0297511015,
0.0062137502,
-0.0058029625,
-0.0145441126,
0.0026463587,
-0.0007452515,
-0.1000876799,
0.0655667931,
0.0490995571,
0.0022950799,
-0.0384829566,
-0.0246745553,
0.0399396308,
-0.0115860896,
0.0247884057,
-0.0329783261,
0.0064213681,
-0.0123891281,
0.0194964614,
0.0081970217,
-0.0920558721,
-0.0312562361,
-0.0296398681,
0.0324081481,
0.0650399104,
0.0217432026,
-0.0531942993,
-0.0519589148,
0.0363777205,
0.0443851575,
0.0160546489,
0.020493824,
0.0782552958,
-0.0058368193,
-0.0101963291,
-0.0010156451,
0.0650613531,
0.0380429216,
-0.0147602856,
0.0014136699,
0.0591470562,
-0.0075678751,
0.0205693115,
-0.0180360805,
-0.0106312539,
-0.0043609766,
0.0440834686,
0.0187418442,
-0.041362334,
0.0254978891,
-0.0064217802,
-0.0175707527,
-0.0169896707,
0.0029288097,
-0.0012167572,
-0.0342158265,
-0.0376059636,
-0.0476020612,
-0.0391609743,
0.0140710315,
-0.0065663569,
0.0046589649,
0.0204357505,
0.0018835685,
-0.0384055749,
-0.0345814563,
0.0631339848,
-0.0522102155,
0.0375768095,
-0.03176805,
-0.0258213114,
-0.014825061,
-0.0680566505,
-0.0010099743,
0.0250733234,
-0.1208457202,
-0.0631333068,
-0.0031140316,
0.01571621,
0.0183890704,
0.0987162441,
-0.0192396492,
-0.0810834169,
0.0535107404,
0.0508780293,
-0.0368544422,
0.0367509797,
0.0289844256,
-0.0375522226,
-0.0000431595,
0.0104686944,
-0.0375378802,
0.0624637604,
0.0196533352,
-0.0045829145,
-0.0388601311,
-0.0310067832,
0.0008359049,
0.0205660686,
0.0555688776,
-0.0013919401,
0.0311419517,
-0.0265366118,
0.0265072081,
-0.0397748984,
-0.0100149391,
-0.0588330291,
-0.0257791355,
0.0644872338,
0.0150580797,
-0.0063165505,
-0.0216881558,
0.025055198,
-0.0023373859,
-0.0911921263,
-0.013150271,
-0.0178661235,
-0.0162048191,
0.0331923217,
0.0341338031,
0.0072098253,
0.0333494321,
-0.0024267931,
0.0218554176,
0.0170810167,
0.0108758537,
0.0081417607,
-0.0297212489,
-0.0673063174,
0.0465143584,
-0.0841581076,
-0.015934512,
-0.0394513942,
0.0691166371,
0.0500015207,
-0.0029355169,
0.0701272339,
-0.0262655094,
-0.0257514287,
-0.0342270546,
0.0136832325,
-0.0036753044,
-0.0271909032,
0.0244594775,
0.0090854494,
-0.020032648,
0.0223803613,
0.0323727019,
-0.0164050236,
0.0454545431,
-0.0375723653,
-0.008531115,
-0.00785268,
-0.0258548632,
0.0132268593,
-0.0004188209,
-0.0213378984,
0.0427048728,
-0.0000474816,
-0.0837464556,
-0.0214580651,
-0.0117144808,
-0.0151220886,
-0.0162889827,
0.0141521757,
-0.0124159819,
-0.0291654654,
0.0162307266,
-0.0048180339,
-0.0184890423,
-0.0094865179,
-0.0224496517,
0.051577922,
-6.152002582e-33,
-0.0189333912,
0.0246300008,
0.0108364057,
-0.0416568965,
-0.0565587617,
0.0197990369,
0.01597156,
0.0064896913,
-0.0228369609,
0.0372742079,
0.0016899032,
0.0247373078,
0.0218069721,
0.0425375178,
-0.0232149716,
0.0217295941,
0.0892974213,
-0.0036630274,
-0.0038027635,
0.0066559594,
-0.0018666222,
0.0155374222,
0.048278816,
-0.0772674456,
0.024266487,
0.0209598951,
-0.0253439806,
0.0176449306,
0.0155294957,
0.0487806164,
-0.0275774524,
0.0461384542,
0.0033298312,
-0.0294420719,
-0.0389179923,
0.0278289709,
-0.0865466595,
0.0409284122,
0.0028333708,
0.0305584557,
-0.0777967125,
-0.0288205687,
0.0667830259,
-0.0058312588,
-0.0045202114,
-0.0080009783,
0.025174953,
-0.0895695686,
-0.0052312622,
0.0150536969,
-0.0382715501,
0.0286235083,
-0.0257255081,
-0.0027195427,
0.0503734015,
-0.0066479328,
0.0077392999,
0.0219581667,
-0.057670027,
0.0127901873,
0.083560437,
0.0141273336,
0.0600268394,
0.0540599041,
0.0135995913,
0.0359716713,
0.0292584952,
-0.0221177805,
-0.000020143,
0.0198064093,
0.0072632204,
0.0485269018,
0.031767562,
0.0351899602,
0.0451042652,
0.0001508742,
-0.0046925736,
0.0127427336,
-0.042551171,
0.0016359043,
0.0027545281,
0.0204220843,
-0.0190065801,
-0.0019629754,
-0.0337983631,
-0.002063364,
-0.0112999203,
-0.0294220205,
0.0105792014,
0.023153808,
-0.0121436184,
0.0632381961,
0.0217292123,
-0.0519498214,
-0.0183629692,
-0.0137785403,
-0.0082905227,
0.0423070081,
0.0123584922,
-0.0129776523,
-0.0181397106,
-0.0244637411,
0.071678713,
0.0237235241,
-0.018570181,
-0.0101164402,
0.047728017,
0.0248866454,
-0.0484441705,
-0.0026638675,
-0.0095112715,
0.0092661316,
0.0284872595,
0.0132024074,
0.0298453085,
0.0115041789,
-0.0173061937,
0.0559111461,
0.0215199888,
0.0367996655,
-0.0068459488,
-0.0353892334,
-0.0335827991,
0.0042580524,
-0.0095051127,
0.0197442882,
-0.0174431335,
-0.0494050607,
0.0443313792,
-0.047344964,
-0.024954468,
0.004768393,
2.698e-7,
0.006006157,
0.0178597644,
0.0294710733,
-0.0090153869,
0.0375344791,
0.0202205051,
-0.0189773235,
0.0068499753,
0.040956594,
-0.0286333039,
0.026311351,
0.0722530186,
0.0472475179,
-0.0327230506,
-0.0206415877,
-0.0143717267,
-0.0544680282,
-0.0041088751,
-0.0388052054,
-0.0020319484,
0.0256806947,
0.0957124382,
0.0171653684,
-0.0337964669,
0.0433390178,
-0.0194672327,
0.0634834766,
-0.0497934632,
-0.0239023715,
-0.0003866678,
0.0261544567,
0.0475834385,
0.0043572774,
0.0418998003,
-0.0227599367,
0.0461018011,
0.0250362549,
0.0262768809,
-0.0423743576,
0.0631419122,
-0.0448876694,
0.0386061929,
-0.0028784978,
-0.0273028575,
0.077020444,
-0.0231563617,
-0.0345038846,
-0.0635226592,
-0.0794335976,
-0.0010438475,
0.0120351026,
-0.0301059075,
-0.0017636467,
-0.0119569004,
0.0460259616,
0.0141793927,
-0.0349834748,
-0.0513883531,
-0.0213517509,
0.0108900182,
-0.0490832292,
-0.0662202239,
-0.0446558557,
0.0017821272,
0.0219669398,
-0.0551034696,
-0.0185566153,
2.184842218e-34,
0.0395052209,
-0.0080931839,
0.0321469568,
-0.0085069099,
0.0039412146,
-0.0173090212,
-0.0037132767,
-0.0035431131,
-0.0064473553,
-0.1087174416,
0.004968306
] |
http://arxiv.org/abs/2401.16441v1 | http://arxiv.org/pdf/2401.16441v1 | FaKnow: A Unified Library for Fake News Detection | 2024-01-27T13:29:17 | 2024-01-27T13:29:17 | Yiyuan Zhu, Yongjun Li, Jialiang Wang, Ming Gao, Jiali Wei | cs.LG | cs.LG, cs.AI, cs.CL | Over the past years, a large number of fake news detection algorithms based
on deep learning have emerged. However, they are often developed under
different frameworks, each mandating distinct utilization methodologies,
consequently hindering reproducibility. Additionally, a substantial amount of
redundancy characterizes the code development of such fake news detection
models. To address these concerns, we propose FaKnow, a unified and
comprehensive fake news detection algorithm library. It encompasses a variety
of widely used fake news detection models, categorized as content-based and
social context-based approaches. This library covers the full spectrum of the
model training and evaluation process, effectively organizing the data, models,
and training procedures within a unified framework. Furthermore, it furnishes a
series of auxiliary functionalities and tools, including visualization, and
logging. Our work contributes to the standardization and unification of fake
news detection research, concurrently facilitating the endeavors of researchers
in this field. The open-source code and documentation can be accessed at
https://github.com/NPURG/FaKnow and https://faknow.readthedocs.io,
respectively. | FaKnow: A Unified Library for Fake News Detection: Over the past years, a large number of fake news detection algorithms based
on deep learning have emerged. However, they are often developed under
different frameworks, each mandating distinct utilization methodologies,
consequently hindering reproducibility. Additionally, a substantial amount of
redundancy characterizes the code development of such fake news detection
models. To address these concerns, we propose FaKnow, a unified and
comprehensive fake news detection algorithm library. It encompasses a variety
of widely used fake news detection models, categorized as content-based and
social context-based approaches. This library covers the full spectrum of the
model training and evaluation process, effectively organizing the data, models,
and training procedures within a unified framework. Furthermore, it furnishes a
series of auxiliary functionalities and tools, including visualization, and
logging. Our work contributes to the standardization and unification of fake
news detection research, concurrently facilitating the endeavors of researchers
in this field. The open-source code and documentation can be accessed at
https://github.com/NPURG/FaKnow and https://faknow.readthedocs.io,
respectively. | [
0.0732836276,
0.0245711878,
0.0039382624,
0.0821510926,
-0.0292536933,
0.0394807979,
0.0055749975,
0.0306762978,
0.009163673,
-0.0092831245,
-0.0109694274,
-0.0195363928,
0.0224258006,
0.0697507784,
0.00255664,
-0.0223824959,
0.0546074472,
0.0596148856,
0.0668710545,
-0.0059255864,
0.0544779114,
0.0513668768,
-0.0092276465,
0.0291817561,
-0.0375058055,
-0.02039974,
0.0308131948,
0.0196958818,
0.0557779744,
-0.0150077445,
0.0312157124,
0.0214061383,
0.0019876517,
0.0948203877,
0.0000019184,
-0.0015919848,
0.0063424087,
-0.0047484906,
-0.0301464461,
-0.0107580032,
0.066970706,
-0.0770926923,
-0.02213897,
0.0193299092,
-0.0515959822,
-0.0407461151,
0.0252801124,
0.0663473755,
0.0627176762,
0.1019484997,
-0.0147311995,
-0.0037023916,
0.0164351668,
-0.0004624907,
-0.0107083172,
-0.0132240783,
0.0510467552,
0.0047299606,
-0.029819712,
0.0343967713,
0.0227882639,
0.0158168077,
0.0302778836,
-0.006424861,
-0.0363641828,
-0.0027285013,
-0.0105706695,
-0.0439402238,
0.0009702353,
0.0272718184,
-0.0695256218,
-0.0705498382,
0.0141170202,
-0.0163931977,
0.0143783959,
0.0027431888,
0.0180964228,
0.0343009531,
-0.0072379196,
0.0168171003,
0.0076361741,
-0.0086318851,
0.0479291826,
0.0656769276,
0.0463734865,
-0.0012778351,
-0.0273622163,
0.008672609,
-0.0108529972,
-0.0029594325,
0.0158254579,
-0.0501997508,
0.0452884249,
0.006018851,
0.0035951431,
-0.0187645685,
-0.0821501538,
-0.0434124917,
0.0143792862,
-0.0700908452,
-0.0323035084,
-0.0169879124,
0.0019667225,
0.023497209,
-0.0231611673,
-0.0235791355,
-0.0250289403,
0.015377257,
-0.0079201935,
-0.0062493286,
0.0187327843,
0.0176301561,
0.012892331,
-0.0169053618,
-0.0382018238,
-0.007970904,
-0.0579625815,
0.0199077837,
-0.0176615138,
0.0147605762,
-0.0290178712,
0.0206573009,
-0.0337153822,
-0.0102946414,
0.0064772377,
0.0175228883,
0.0186450444,
0.0154724196,
-0.0103110401,
-0.0353007838,
-0.0001469572,
0.0355872735,
-0.015235493,
0.0095917974,
0.0147841973,
-0.0381047949,
0.0233620964,
0.0007379117,
-0.0436722711,
-0.0165065639,
-0.007134662,
-0.0245058052,
-0.0107884696,
-0.0889508501,
-0.0176766552,
-0.0564535484,
-0.0171885509,
-0.0158048328,
-0.0251890048,
0.0296730231,
-0.0153482081,
0.0318977758,
-0.0280206837,
-0.0185944028,
0.0678773299,
0.038149707,
0.0154180303,
0.085552454,
-0.0159581266,
0.117533192,
0.0414743535,
0.007569931,
0.0049983175,
0.0116177388,
0.0538549908,
0.0034074925,
0.0333685875,
-0.0107795177,
-0.0409456901,
0.0063208146,
0.0046221218,
-0.0069503891,
-0.0029139426,
-0.0671085119,
0.0415353067,
-0.0149391573,
0.0162680615,
-0.0028240865,
0.0770674795,
0.0399297886,
-0.0035866213,
-0.0271367822,
-0.0441349633,
0.0089407191,
-0.073235549,
-0.0109977797,
-0.0562721975,
0.0712351725,
0.0116892159,
-0.006801872,
0.0134036904,
-0.0265625231,
0.0083915507,
-0.030085085,
0.0140078058,
0.0345644467,
-0.0419462845,
-0.0102591207,
0.0192359947,
0.0246551689,
0.0023430113,
-0.0401604995,
0.0340317711,
0.0240028501,
0.0291547477,
0.0258615501,
0.0302320272,
-0.0349169187,
0.0084924549,
0.0203962363,
0.0141961267,
-0.0177270193,
-0.0267883744,
-0.0008310291,
-0.0281783696,
-0.020429913,
-0.0049922927,
-0.002735019,
0.0589220598,
0.0546642467,
0.0263210312,
0.0073567717,
-0.0360352471,
0.0086039538,
-0.0211563595,
-0.0028482575,
-0.0194948353,
0.034929432,
0.0128728747,
0.0290098879,
0.0186712798,
-0.036844559,
0.0532021597,
0.0300378986,
-0.0253872443,
-0.0362204909,
-0.0266075861,
-0.0442801379,
-0.0079159867,
-0.0620843358,
-0.040054474,
0.1057951227,
-0.0185149312,
0.0359461643,
0.0362704545,
-0.0375212766,
0.0067616897,
-0.0512608737,
-0.0555796549,
-0.0347623713,
0.0350704789,
0.0566486977,
0.0741008967,
-0.0401959307,
-0.0217904896,
0.0297014061,
0.0374253802,
-0.0296579525,
-0.0110944109,
0.00667773,
-0.0260779802,
-0.0438273922,
-0.015515225,
0.0104918387,
0.0159052126,
-0.0081033008,
0.0190226212,
-0.0741671175,
-0.0255070645,
-0.038341511,
-0.0339848846,
0.0551657043,
0.011056981,
-0.0070532202,
0.007328412,
-0.0237785783,
-0.0184826795,
0.0152418511,
0.0024703073,
-0.0236118305,
-0.0441960394,
0.0111922491,
-0.0100587076,
0.0200441759,
-0.0127040921,
0.0222401954,
0.0074151717,
-0.0830551237,
-0.0256391335,
-0.0305572841,
-0.0393892676,
0.0049043237,
-0.02601799,
-0.0026885958,
-0.0044922568,
0.0105105108,
-0.0190152135,
0.0534586236,
0.0605704933,
0.0876928195,
0.0034839886,
0.0788535476,
-0.0178232305,
-0.0399887376,
-0.0432611108,
0.080631122,
-0.0313435905,
-0.080218859,
0.0126053318,
0.010461011,
0.0280113574,
0.0239825919,
0.0041735214,
0.027879823,
-0.0017370386,
-0.0262244102,
0.0401570909,
-0.0309368316,
-0.0098731359,
-0.0203260258,
-0.0182366446,
-0.0265751444,
0.1019025743,
0.0134515883,
0.0240318738,
-0.0308541954,
-0.0584158376,
0.0026929658,
0.0015919008,
-0.0252801944,
0.0130600249,
0.0048314175,
0.0105615985,
0.0367921852,
0.0051947595,
0.0370087437,
-0.0259837899,
-0.0371689275,
-0.0172955897,
0.0258127581,
0.0253717862,
0.0048978934,
-0.0669795424,
-0.0965506509,
-0.0286983512,
-0.0481482521,
-0.0177151505,
-0.024415439,
-0.0475980453,
-0.033618547,
0.0136413788,
0.0458019562,
-0.0083589377,
0.0071026245,
0.0104670404,
0.0458961949,
0.0331985913,
-0.0188446436,
0.0573090278,
-0.012568607,
-0.0354132243,
0.0701110289,
-0.0095718913,
0.0300804414,
-0.0258530993,
-0.0343118124,
0.0112078339,
0.0188254099,
0.0238316413,
-0.0189733598,
-0.0041422034,
-0.0450737774,
0.0045990171,
0.0760383531,
0.0014941532,
-0.0293664243,
0.0144754844,
0.0375422835,
-0.0209675096,
0.0013575635,
-0.0320348665,
-0.0578447543,
-0.0266232975,
-0.0244307145,
-0.0117800338,
-0.0626418069,
0.0001020047,
-0.0107941786,
-0.0437091142,
0.0042401901,
-0.0077665779,
-0.0433484353,
-0.0504649132,
-0.0750914067,
-0.0137363737,
-0.0016708875,
0.0213599578,
-0.0208764784,
0.0261359848,
-0.0404505618,
-0.0253958646,
-0.0061986358,
-0.0709287748,
-0.010512922,
-0.0065826266,
0.0802692771,
0.036952354,
0.0350323245,
-0.0078388592,
-0.0430939868,
-0.0131920781,
0.0077655194,
-0.0366956182,
-0.0307859592,
0.0532496497,
-0.0581377856,
-0.0142297931,
-0.0389794931,
0.0522972941,
-0.0012267337,
0.0050744177,
-0.0291453488,
0.0342902094,
0.0131559568,
-0.1177587211,
-0.0537387989,
0.021057982,
0.043255195,
0.0588077791,
0.0153507069,
-0.0130421435,
0.0070187403,
-0.0080218744,
-0.0092207128,
-0.0084072761,
-0.0578245446,
0.0134871649,
0.0050335238,
-0.0281958785,
-0.0302778669,
-0.0298865978,
-0.000506583,
0.001981386,
-0.0370802581,
-0.0290211122,
0.004507849,
-0.0256533958,
-0.0757799596,
0.0125766713,
-0.0070714806,
0.0676643774,
0.0004832125,
0.003287304,
-0.0067174518,
-0.0116864881,
-0.015647985,
0.0596538074,
-0.1009732708,
-0.0211681984,
0.0069072,
0.031179728,
0.0077996179,
0.0392786674,
-0.0253157094,
-0.0293296594,
0.0618121997,
0.0217729565,
-0.0239190497,
0.0204853285,
0.0263642911,
-0.0182533916,
0.0548402928,
0.0144321807,
-0.0043497696,
0.0544077046,
0.0129303662,
-0.048476316,
-0.059802521,
0.0262003224,
-0.0098094139,
-0.0056555057,
0.0196829513,
-0.0047589922,
-0.0149447033,
-0.0146793928,
-0.0714671165,
-0.0135837132,
-0.0141793517,
0.0105575854,
0.0443893336,
0.0466762148,
0.0081506912,
0.0261752922,
0.0158001687,
-0.0108605316,
0.007703077,
-0.0316960774,
-0.0310598128,
0.0164173245,
0.0570443533,
0.0198392589,
0.0102116894,
-0.0397919193,
0.0381540135,
-0.0091546969,
0.0208020862,
0.014093969,
-0.0099326102,
0.0130481254,
-0.0652260482,
0.0437648669,
-0.0220229272,
-0.0235604253,
0.0438810475,
-0.0684702173,
0.0577444844,
0.0114502702,
-0.0120757353,
0.0393897258,
-0.0268306136,
-0.0007269201,
-0.004645356,
-0.0277797077,
-0.0294961557,
-0.0509152152,
-0.0265701078,
0.0073511354,
-0.0254872255,
0.0588968918,
0.0460717157,
0.0173646789,
0.0178268775,
0.0311830249,
-0.0561759695,
-0.0326585993,
-0.0052352981,
-0.0000753678,
0.0713209137,
-0.0350296088,
-0.0017549357,
-0.043891646,
-0.036458265,
0.0103628905,
-0.0075262804,
0.0248640217,
0.0399076119,
-0.0116872955,
0.0853414237,
0.0005944209,
0.0167724341,
-0.0048814747,
0.017096987,
0.0011765963,
-0.081811972,
0.0088074012,
-6.094991841e-33,
0.052120138,
-0.0669125542,
-0.0507128835,
-0.0113412291,
0.0123262051,
0.0761792213,
0.0136742508,
-0.0308577716,
-0.0287121776,
0.0420036539,
0.0549606346,
0.074404411,
-0.0138453431,
0.0201502815,
-0.1057294011,
0.0013585715,
0.0271263607,
0.0134278927,
-0.0251992885,
0.030648157,
0.0379861034,
-0.0005959768,
0.0343468189,
-0.0623005107,
0.0484528057,
-0.0508457609,
-0.0392596684,
0.0048699011,
-0.0191363767,
0.0346086212,
-0.0448038802,
0.0401429683,
0.0329374447,
-0.0122127747,
-0.0206129849,
0.0399717651,
-0.0172799472,
-0.0208353493,
0.0281077269,
0.0018820106,
-0.0560283624,
0.018677745,
0.0580934547,
-0.0010321971,
-0.0432316735,
0.0057208552,
0.0356732272,
-0.023817217,
0.0127358995,
0.0020463204,
-0.0294693783,
0.0021131812,
0.000881208,
0.0107675251,
-0.0007542241,
0.0606229678,
0.0283609554,
0.0181630477,
-0.0948935822,
0.0005581022,
-0.0612249151,
-0.0131010013,
0.0156849679,
-0.0618236773,
-0.0309127681,
0.0505683087,
0.0199299715,
0.0442818776,
-0.0534896106,
0.00874874,
0.0186145194,
0.0163034201,
0.0891937613,
0.0981263667,
0.0003527697,
-0.0678353012,
-0.0598637573,
0.0352489054,
0.0581590235,
-0.0334112309,
0.0038296075,
-0.0400814228,
-0.0324340947,
0.0139054768,
0.0050018118,
-0.0121482145,
-0.0234581698,
-0.0089614755,
-0.0053038211,
0.0292078126,
-0.0583754964,
0.0724349543,
-0.0111861126,
-0.0813359916,
0.0146570336,
0.0364193469,
-0.0132509414,
0.0067036473,
0.0391340852,
0.0517666824,
-0.026205238,
0.027956387,
0.0482084341,
-0.0008241934,
-0.0035958558,
-0.0290508009,
0.0480463356,
0.0225663017,
-0.0036960668,
0.0147702275,
-0.0170774981,
0.0601840839,
-0.0189867485,
-0.0178261064,
0.004724883,
-0.0128222136,
-0.0113561414,
-0.0174264368,
0.000298394,
0.021513816,
-0.0107110646,
0.0202192459,
-0.0317152031,
0.0408117287,
-0.0156810302,
-0.0252876822,
-0.0509150214,
0.0000750232,
0.0187686365,
-0.0964669362,
0.0158108175,
0.0110699795,
2.596e-7,
-0.0436277688,
0.0327162035,
0.0166240558,
-0.0032059175,
0.0084150638,
0.0526198409,
0.0681587756,
0.0313722417,
-0.0050489069,
-0.0794459879,
-0.0388165526,
0.0386107564,
0.0159746669,
0.0105778575,
-0.0623206086,
0.0227542501,
-0.0574625917,
-0.0072279009,
-0.055243548,
-0.0061844373,
-0.0475237928,
0.0630120486,
0.0710127875,
-0.0611728951,
0.0330035612,
0.015517923,
0.0227598138,
0.0348162279,
-0.0186589081,
-0.0150999893,
0.048900485,
-0.0311667863,
0.0335524864,
-0.0386897512,
-0.0156998783,
-0.0215977058,
-0.0070208972,
0.0293242242,
0.0253698975,
0.0003607754,
-0.0163464192,
-0.0548245572,
0.0380006768,
-0.0634102672,
0.0545080416,
-0.0060725762,
-0.0169057976,
-0.0718175322,
-0.0273732711,
0.0167719889,
0.0377677456,
-0.0042226235,
-0.0030995808,
0.0282204933,
0.0258336868,
-0.0191269349,
-0.0039049149,
-0.0287743174,
-0.0019735617,
-0.0227269474,
-0.0440943688,
0.0016882443,
-0.0133532202,
-0.0483817533,
0.0582687594,
-0.0080022002,
0.0314475559,
1.905312187e-34,
0.033054404,
-0.0097412271,
-0.0024815146,
-0.0236068349,
0.0010245505,
0.026335327,
-0.0284780506,
-0.0278318133,
0.0311034918,
0.0191248674,
0.0084661189
] |
http://arxiv.org/abs/2401.15400v1 | http://arxiv.org/pdf/2401.15400v1 | Indexing Portuguese NLP Resources with PT-Pump-Up | 2024-01-27T12:33:07 | 2024-01-27T12:33:07 | Rúben Almeida, Ricardo Campos, Alípio Jorge, Sérgio Nunes | cs.CL | cs.CL, cs.IR, 68P20, I.7.1 | The recent advances in natural language processing (NLP) are linked to
training processes that require vast amounts of corpora. Access to this data is
commonly not a trivial process due to resource dispersion and the need to
maintain these infrastructures online and up-to-date. New developments in NLP
are often compromised due to the scarcity of data or lack of a shared
repository that works as an entry point to the community. This is especially
true in low and mid-resource languages, such as Portuguese, which lack data and
proper resource management infrastructures. In this work, we propose
PT-Pump-Up, a set of tools that aim to reduce resource dispersion and improve
the accessibility to Portuguese NLP resources. Our proposal is divided into
four software components: a) a web platform to list the available resources; b)
a client-side Python package to simplify the loading of Portuguese NLP
resources; c) an administrative Python package to manage the platform and d) a
public GitHub repository to foster future collaboration and contributions. All
four components are accessible using: https://linktr.ee/pt_pump_up | Indexing Portuguese NLP Resources with PT-Pump-Up: The recent advances in natural language processing (NLP) are linked to
training processes that require vast amounts of corpora. Access to this data is
commonly not a trivial process due to resource dispersion and the need to
maintain these infrastructures online and up-to-date. New developments in NLP
are often compromised due to the scarcity of data or lack of a shared
repository that works as an entry point to the community. This is especially
true in low and mid-resource languages, such as Portuguese, which lack data and
proper resource management infrastructures. In this work, we propose
PT-Pump-Up, a set of tools that aim to reduce resource dispersion and improve
the accessibility to Portuguese NLP resources. Our proposal is divided into
four software components: a) a web platform to list the available resources; b)
a client-side Python package to simplify the loading of Portuguese NLP
resources; c) an administrative Python package to manage the platform and d) a
public GitHub repository to foster future collaboration and contributions. All
four components are accessible using: https://linktr.ee/pt_pump_up | [
0.0419287346,
0.0223725531,
-0.048461657,
0.0278590899,
0.0127570648,
0.0122612631,
-0.0489642359,
0.0279186927,
-0.0380320661,
-0.0463408493,
-0.0248237029,
-0.0153352106,
0.0087233568,
0.0068807118,
0.0153998425,
-0.0765057877,
0.0410426557,
0.0156816691,
-0.0352535248,
-0.0206307471,
0.0031291763,
0.0336942412,
-0.0191799812,
0.0293777529,
0.0693140179,
0.0148717714,
-0.0137856705,
-0.0556110553,
0.0014637951,
-0.0141635882,
-0.0052023781,
0.0287183654,
-0.0463989079,
0.0414529331,
0.0000021143,
-0.0619825944,
-0.0586190745,
0.0109644439,
-0.024336176,
-0.0133721409,
0.0593107194,
-0.0119526871,
-0.0230655838,
0.0163257886,
-0.0694632828,
-0.0363365635,
0.0551894046,
-0.0181646794,
0.0631125271,
0.0460321382,
0.0059993295,
-0.0221895482,
-0.023488462,
0.010414294,
0.0059572342,
0.0104579069,
-0.0109529588,
-0.0007726513,
0.0686371997,
0.0279929228,
0.0288603175,
0.0259309895,
-0.0110483747,
-0.0323272943,
0.0024111322,
0.0128094694,
-0.0439323597,
-0.0775134191,
-0.0087882821,
0.0086028297,
0.0165220834,
-0.0082821632,
-0.0007277212,
0.0062030815,
-0.0034085233,
0.0244188756,
-0.054805588,
0.0392965563,
0.0100182062,
0.0162192751,
0.0600508861,
-0.0057878797,
0.0235935766,
0.0245173983,
0.0388045125,
0.0203612987,
0.02778963,
-0.0185509883,
0.0692548826,
0.0137130348,
0.0539785959,
-0.0327221006,
0.0184955616,
0.0426310301,
0.0094752321,
0.0159987714,
0.0294974335,
-0.0340870582,
0.0809588954,
-0.065121077,
-0.0251532588,
-0.0089896414,
-0.0169510599,
0.0282290708,
0.0285591744,
-0.0297635533,
-0.0280080475,
-0.0167650264,
-0.0062530637,
-0.0424441844,
-0.0841366649,
-0.0356316045,
-0.067620866,
0.0408603735,
-0.0687778443,
0.0340105444,
-0.0444796979,
0.0250528213,
-0.0119172391,
-0.0071365186,
0.0399411395,
0.0085311532,
0.002627698,
0.0296899229,
-0.0454190671,
0.0447521359,
-0.0363691151,
-0.0238748938,
-0.0494989231,
-0.0724842846,
0.0368438214,
-0.0086946702,
-0.015644744,
-0.0285944231,
0.0250520241,
0.0239242837,
-0.0074372618,
-0.0577337593,
-0.0688440651,
-0.0402018316,
0.0119255967,
-0.0618949309,
-0.01928721,
-0.0626439303,
-0.0075070178,
0.0114860516,
0.045485612,
-0.0158601385,
0.0042508943,
0.0221316535,
0.0104153901,
0.0055717067,
-0.0755086094,
-0.0095326584,
0.0097377496,
0.0028731155,
-0.0029890833,
0.0849408358,
-0.0043719537,
-0.0571934581,
0.0711485818,
-0.0109398086,
-0.0277322214,
0.016713893,
0.0012919977,
-0.0254475083,
0.0506996736,
0.0218812097,
0.008163563,
0.0198900159,
0.0020141255,
0.0196009856,
-0.0147337392,
0.0301126093,
-0.0201928969,
0.045871783,
0.0350819416,
0.0131642846,
0.0365437828,
-0.006846928,
0.0068048271,
-0.0676718578,
0.042420011,
-0.0047386005,
0.0366974995,
-0.0021080875,
-0.0064108032,
0.098342523,
0.0017706605,
-0.0162375402,
0.0007777085,
0.0098928912,
-0.02500844,
-0.0251013488,
0.0310215633,
-0.0147312945,
0.0158534143,
0.0282353573,
0.0355024002,
-0.0365726277,
0.019120777,
-0.006957151,
-0.0139447553,
0.0559434108,
0.023358345,
-0.0144837191,
0.0175339635,
-0.0120868729,
0.0010757304,
0.0280570537,
-0.0195088126,
-0.0332564414,
-0.0098185902,
-0.0338709876,
-0.0004007453,
0.0073409481,
0.022117028,
-0.002421407,
0.0114869047,
0.0830682367,
0.0081485314,
-0.03439923,
-0.0032350842,
0.037951652,
-0.0198974758,
0.0081130862,
-0.0283119548,
0.0104208002,
0.0224273093,
0.0905585513,
0.0155338822,
-0.0220999755,
0.0319531225,
-0.0031033636,
0.0185148455,
0.0070895264,
0.0788954124,
0.0536885932,
0.0508545972,
-0.0407864563,
-0.0455657318,
0.0174353588,
0.0412073694,
-0.0144945914,
-0.0102353664,
-0.0353878662,
0.0955831707,
-0.0151775125,
0.0188980345,
-0.0259228274,
0.0522882007,
0.0087147932,
0.0507256612,
0.0298635997,
0.0548726693,
0.0444632545,
-0.0069922428,
0.0367193446,
-0.069548279,
-0.0183124486,
0.0494610146,
-0.0097580226,
-0.0302096438,
0.0239820145,
0.0370433331,
-0.0466561764,
0.0318717882,
-0.0891850367,
-0.0147391213,
-0.0178043079,
0.0255774055,
0.0147118615,
0.0373078547,
-0.0140874339,
-0.0154991169,
0.0349025689,
-0.0457469448,
-0.0022438799,
0.0095044374,
-0.0151195144,
0.0088050393,
-0.0004957726,
-0.0337001346,
0.0299326628,
0.0566854849,
0.0210320689,
0.05653698,
-0.0919109136,
-0.0305910241,
0.0267690346,
-0.0308813713,
-0.01520254,
0.0326753519,
-0.0403422825,
-0.0086812116,
0.0390838049,
0.0252788384,
0.0108736856,
-0.008555444,
-0.0015661457,
0.0788056031,
-0.0165891517,
-0.0610185862,
-0.0312219225,
-0.029073501,
-0.0487887152,
-0.0513942875,
-0.0307858009,
0.061171744,
0.0475614555,
-0.0177080333,
0.0149122598,
-0.0196210202,
-0.0344917774,
-0.0836176351,
0.0055772015,
0.0083741797,
0.0176214743,
-0.0409008563,
0.0145855332,
0.0406911671,
0.0051488108,
0.0577756874,
-0.0312499776,
-0.0158607885,
-0.1075082794,
-0.0600835495,
-0.0353170037,
0.044019036,
-0.0210033376,
0.0042471415,
0.0037481906,
-0.0066607757,
0.0138386348,
0.0638039634,
0.0432809331,
-0.0271002874,
-0.0278225206,
-0.0429990478,
0.022035215,
-0.0137579758,
-0.0444857031,
-0.0008698392,
-0.0089340443,
-0.0295243431,
-0.0106228059,
0.0245624185,
0.0144059826,
0.0051126024,
0.0096722757,
0.0375627652,
0.0217019785,
0.0339084193,
-0.0390016809,
0.0168614332,
0.0210636128,
-0.0045825103,
-0.0013033205,
0.0045486023,
0.031890966,
-0.0214431919,
0.0212348998,
0.0191504285,
-0.0504233129,
0.0126104932,
0.004219431,
0.0171665102,
-0.04902393,
-0.0679027364,
-0.0306796227,
-0.0304439086,
-0.0058211638,
-0.0378631428,
0.04592067,
0.0532893129,
-0.0808758065,
0.0056387712,
-0.0170527156,
0.0139865531,
0.0283158123,
0.0491422117,
-0.0783792883,
0.0066918647,
-0.0096090874,
0.0446143858,
-0.0328881815,
0.0034434185,
-0.0258155875,
-0.1165460423,
0.044913739,
-0.0210441295,
0.011052656,
-0.0021809589,
-0.0902436078,
-0.0157035906,
-0.0337208398,
-0.0333395787,
0.0116395801,
-0.0119541846,
0.0533466116,
0.0637948662,
0.0047531915,
-0.0787378699,
-0.041870378,
-0.0165181067,
0.0336888433,
0.0155070703,
0.0275143571,
-0.0002039806,
0.0188361052,
0.0379244499,
-0.014896472,
-0.0179763231,
-0.0420242921,
0.0026222304,
-0.0025424215,
-0.0143767018,
-0.0401563607,
0.0027630292,
-0.0223858207,
-0.0045301346,
0.0155998031,
0.1138556078,
0.0044930973,
0.0554365553,
-0.0339768417,
-0.0130929481,
0.0070272419,
0.0162389558,
0.00081476,
-0.0085327374,
0.0328508988,
0.0005192765,
-0.0146099357,
0.0172000807,
-0.05682512,
0.056213513,
-0.0730103701,
-0.0288685281,
-0.0872413144,
-0.0651746392,
0.0057527572,
0.0117240883,
0.0226423386,
0.0330129825,
-0.0232479665,
0.0234931652,
-0.0516400822,
0.0007521312,
0.0354702473,
0.0204576869,
-0.0612930954,
-0.0897831917,
0.0245370623,
0.0319352187,
-0.0199548043,
-0.0328238793,
-0.0118254879,
-0.0369339772,
0.0470307879,
0.0021427975,
0.0324540883,
-0.0278712194,
0.0138315652,
-0.063862443,
0.0079495348,
0.0585878864,
-0.1022726744,
0.0314261131,
-0.017375974,
0.0080070039,
0.0080007482,
0.0053023789,
-0.0428470932,
-0.0060966825,
-0.0605086163,
0.0389067754,
-0.0322737396,
-0.0152717875,
0.0104353288,
0.0357300267,
0.0419589318,
-0.0340621844,
-0.0469738916,
-0.0314120278,
0.0238923747,
0.0310685877,
-0.0007996833,
-0.0041806907,
0.0399269722,
0.0203293581,
-0.0389088877,
0.0010497395,
0.0692614242,
0.017466113,
-0.0043383031,
0.0098157078,
0.0380005389,
0.0543220975,
0.0018289457,
-0.009256037,
0.0432483591,
0.0081463419,
-0.0115923546,
-0.0284535699,
0.0424052104,
-0.0418852456,
0.0208958127,
0.0102731893,
-0.0534586199,
0.0201574583,
0.008523779,
0.0173114613,
-0.0189154577,
-0.0019166515,
-0.0172830839,
0.0608828254,
-0.0069075287,
0.0717525631,
-0.0576138124,
0.0105278436,
0.0185492299,
-0.0051304423,
-0.0173350815,
-0.0517198183,
0.0196106676,
-0.0116598736,
-0.024105886,
-0.0019195945,
0.0686421841,
-0.0301306024,
0.0176887941,
-0.0001336774,
-0.0013036348,
-0.0115913358,
-0.0092468811,
0.0159184653,
0.0137670618,
-0.0506455898,
-0.0137314713,
-0.0665684119,
-0.0388141386,
-0.0553830229,
0.0218192991,
-0.0081345849,
0.0029535936,
0.0508565158,
0.0227054507,
0.0218226872,
0.0011773743,
-0.0628455803,
0.0516195782,
-0.0681190938,
0.0229189377,
-0.0272285659,
-5.74650001e-33,
-0.0233492162,
-0.0504555963,
0.0165920388,
0.0072117415,
0.0215460807,
0.0165823363,
-0.0100487955,
0.0170819052,
-0.0188930575,
-0.0076042223,
0.0098067075,
-0.0007715896,
0.0303002335,
-0.0303829536,
0.0337850489,
0.0285221133,
-0.0166854206,
0.0113767525,
-0.0112994509,
-0.0195535421,
0.0614236146,
0.0198842008,
0.0866069198,
-0.0375537165,
0.0327339582,
0.011184264,
-0.0020872029,
0.0349901095,
0.0780050084,
0.0891358107,
-0.0109113762,
0.033717595,
0.0133475438,
-0.0441693664,
0.0107790306,
-0.0159729887,
-0.0572348908,
-0.057705313,
0.0454207398,
0.0370791554,
0.0207301956,
0.0028163667,
0.0225070864,
-0.0110291447,
0.0009680389,
0.0723519698,
0.0303119607,
-0.017624937,
-0.011426664,
0.0104500866,
-0.0407786556,
0.038134262,
-0.0003243311,
0.0587826297,
0.0614106059,
0.0240335986,
-0.0235167574,
-0.0513234735,
-0.0404130891,
0.0084491856,
-0.0282970145,
0.0575219579,
0.0496211685,
-0.0167761538,
-0.0088636586,
0.0281609204,
0.0173889641,
-0.0846921057,
0.0432023108,
0.0701339021,
0.0300443061,
-0.0007260414,
0.0159415305,
0.0831087679,
0.0129448604,
0.0040330281,
-0.0283627938,
0.0161740389,
0.0464486927,
0.0069101602,
-0.0198165737,
0.0013117281,
-0.0011463324,
0.0052630529,
-0.0229022,
0.0399114713,
-0.0028227735,
-0.0405710973,
-0.0075889048,
-0.0036342908,
-0.0314166732,
0.0101975333,
-0.0079654288,
-0.0088754343,
-0.0568709373,
0.0044793473,
-0.001223804,
0.0334305093,
0.003063794,
-0.0035468654,
-0.0590427406,
0.0116658788,
-0.0296185985,
0.0062975693,
0.0221733097,
-0.0253644548,
-0.0174042899,
-0.0186462104,
-0.0499667637,
0.0346483067,
-0.0301507767,
0.0447083712,
-0.0187003259,
0.0103601245,
0.0087094661,
0.0118476758,
-0.0175716728,
0.0360968672,
0.0449833646,
0.0137616079,
-0.0277123526,
-0.0330996551,
0.0430197306,
0.0025554823,
-0.0577974506,
0.0200257152,
-0.036839705,
0.0916502327,
0.0303879511,
-0.0139437513,
0.0100515177,
-0.0058621047,
2.827e-7,
0.0524990596,
0.0400118232,
0.0070819831,
-0.0196590684,
-0.0059059085,
-0.020224411,
0.0055057337,
0.0044356459,
-0.0681058019,
-0.0010971639,
-0.0272368174,
-0.0347967334,
0.0380174145,
0.0929061025,
-0.0970478803,
-0.0350761525,
-0.0260943882,
-0.0517812967,
-0.0846336707,
-0.0108041726,
-0.0183486175,
0.0746888071,
0.0244111754,
0.0045503993,
-0.0619839244,
-0.0184323322,
0.0384266339,
-0.0018559501,
0.0287530813,
0.0127473036,
0.031759128,
0.0410846695,
0.0025687192,
-0.0485874452,
-0.0328052454,
0.0088630645,
-0.0540106073,
-0.0137037756,
-0.0190014523,
0.0209927298,
0.0217988286,
-0.0801011175,
-0.0478269495,
-0.0696071088,
0.0063926983,
-0.0099395318,
-0.0279032197,
0.0035775455,
-0.0175246894,
-0.0213384498,
-0.003534941,
-0.0069571189,
-0.0728917196,
0.0440708697,
-0.0084109642,
-0.054948505,
-0.0476488248,
-0.0697788298,
-0.0152375558,
0.0154258609,
0.0062619625,
-0.0145450998,
-0.0269994568,
0.0060586524,
0.0081373546,
-0.015588061,
0.0440294929,
2.245070462e-34,
-0.0353559628,
-0.0156944301,
-0.0057601659,
0.0365411937,
0.039478533,
0.0100846412,
-0.0020616313,
0.0233513135,
-0.0134025756,
0.0132699944,
-0.0014096549
] |
http://arxiv.org/abs/2401.15393v1 | http://arxiv.org/pdf/2401.15393v1 | Semantics of Multiword Expressions in Transformer-Based Models: A Survey | 2024-01-27T11:51:11 | 2024-01-27T11:51:11 | Filip Miletić, Sabine Schulte im Walde | cs.CL | cs.CL | Multiword expressions (MWEs) are composed of multiple words and exhibit
variable degrees of compositionality. As such, their meanings are notoriously
difficult to model, and it is unclear to what extent this issue affects
transformer architectures. Addressing this gap, we provide the first in-depth
survey of MWE processing with transformer models. We overall find that they
capture MWE semantics inconsistently, as shown by reliance on surface patterns
and memorized information. MWE meaning is also strongly localized,
predominantly in early layers of the architecture. Representations benefit from
specific linguistic properties, such as lower semantic idiosyncrasy and
ambiguity of target expressions. Our findings overall question the ability of
transformer models to robustly capture fine-grained semantics. Furthermore, we
highlight the need for more directly comparable evaluation setups. | Semantics of Multiword Expressions in Transformer-Based Models: A Survey: Multiword expressions (MWEs) are composed of multiple words and exhibit
variable degrees of compositionality. As such, their meanings are notoriously
difficult to model, and it is unclear to what extent this issue affects
transformer architectures. Addressing this gap, we provide the first in-depth
survey of MWE processing with transformer models. We overall find that they
capture MWE semantics inconsistently, as shown by reliance on surface patterns
and memorized information. MWE meaning is also strongly localized,
predominantly in early layers of the architecture. Representations benefit from
specific linguistic properties, such as lower semantic idiosyncrasy and
ambiguity of target expressions. Our findings overall question the ability of
transformer models to robustly capture fine-grained semantics. Furthermore, we
highlight the need for more directly comparable evaluation setups. | [
0.0326538868,
0.0633944944,
-0.018306423,
0.0457100011,
-0.0339845754,
-0.0149950143,
0.0113347601,
-0.0076793474,
-0.0798150152,
-0.0512554012,
-0.0522724018,
-0.0432616509,
0.0118480371,
0.0220028721,
0.0059784255,
-0.0321392715,
0.0643309727,
-0.00041649,
-0.0307223331,
0.0048130141,
-0.0086708097,
0.0237028152,
0.0019950757,
0.0596712194,
-0.044241067,
-0.0127112959,
0.0018337653,
-0.017002983,
0.0055279322,
-0.0009462842,
-0.02828376,
0.0174101908,
0.0171664059,
0.0238581095,
0.0000019656,
-0.0338011608,
-0.0477793738,
0.0096465498,
-0.0067238356,
-0.0175670218,
0.0833406523,
-0.0525204018,
-0.0147193437,
-0.0215833783,
-0.0171631947,
0.0159196723,
0.0800448358,
0.0656653047,
0.038913317,
0.0168743692,
-0.0347084664,
-0.0434694588,
0.0413494632,
-0.007909012,
0.010149451,
-0.0420662686,
0.0146089718,
-0.0352054574,
-0.0299312416,
-0.0301239137,
-0.0347708836,
0.0014245111,
0.0203900058,
0.0484688766,
-0.0335949659,
0.0219482351,
-0.0090886988,
-0.0418559052,
-0.0315256789,
0.039306853,
-0.0450929888,
0.013879844,
-0.0187363811,
-0.0330919735,
0.0027676888,
0.0323502906,
-0.024660958,
0.0386196859,
0.0001973872,
0.002969916,
-0.0420678444,
0.0194209814,
0.0152996127,
0.0363599323,
-0.0700813681,
0.0498781838,
0.0289451554,
-0.0287347063,
0.0032353001,
-0.0361582711,
-0.0049886541,
-0.0038317661,
0.0349028893,
0.0345369466,
-0.0256124046,
0.0443220995,
-0.0434185565,
-0.0261289794,
0.0163060781,
-0.0399712324,
0.0520867594,
0.0552063845,
-0.0092477947,
0.0349319167,
-0.0506227911,
0.0493065789,
-0.0331509374,
0.0377673954,
-0.0407215767,
-0.0352557078,
-0.0344545357,
-0.0519959852,
-0.0559311658,
0.0565771312,
-0.0302832015,
-0.0698193088,
-0.0239110067,
0.0239914246,
-0.0117813973,
-0.0502836965,
0.0007874446,
0.0337558091,
-0.0508215837,
0.0030056187,
-0.0236850847,
0.0503595509,
-0.0049564671,
0.0127879437,
0.0165408812,
-0.0816577822,
-0.0331730172,
-0.0045602061,
0.0059704706,
-0.0538597256,
0.010091058,
0.067585133,
-0.0205516238,
-0.0454080477,
0.0208220724,
-0.0364355557,
-0.0518203378,
-0.0236652847,
0.0578701273,
-0.0418230556,
0.0554504506,
0.0276861656,
0.0165366698,
-0.0473108254,
0.0346691087,
0.0310260449,
-0.0322196148,
0.0779028982,
0.0392120928,
-0.0106021715,
-0.0289778244,
0.0184947681,
0.0034208067,
0.0631609783,
0.0748348683,
0.0350564457,
0.0059272139,
-0.0171375349,
-0.0061194152,
0.037859533,
-0.0000348791,
-0.0318202488,
-0.0548024923,
-0.0268705878,
-0.0050550746,
0.0208035465,
0.0331368111,
0.0661827773,
-0.0203125514,
-0.0023262752,
0.0245035533,
0.0446122736,
0.015136702,
0.0436855145,
0.0628070161,
0.0227314252,
-0.0337058343,
0.0080150589,
-0.0356009677,
0.0330891088,
-0.0014190894,
0.0129760737,
-0.0334580615,
0.0559283756,
0.0024340076,
-0.0034254414,
-0.0101697659,
-0.0178289227,
0.0339799598,
-0.03493369,
0.0356368758,
0.0618576184,
-0.0478994958,
0.1500570923,
0.019890435,
-0.0466866195,
-0.0250164531,
-0.0646058917,
-0.0421283357,
0.0295447633,
0.0253438167,
0.0045291637,
-0.0304448064,
0.0023378811,
-0.0461045019,
0.0464522615,
0.0217655282,
0.0173943248,
0.0156755205,
-0.0520634465,
-0.0299859438,
-0.0159025155,
0.0259120632,
0.0256934557,
0.0087968614,
-0.0460656248,
0.0190243907,
-0.0270592701,
0.0196502246,
0.0283367336,
0.0281214062,
0.0015847965,
-0.0095286919,
0.0105146905,
0.0073723006,
-0.0232951064,
-0.007430933,
0.0392375514,
0.0413321406,
-0.0117839137,
-0.0503317155,
0.0046207062,
-0.0849644467,
0.0356876589,
0.0208249427,
-0.0083166976,
-0.0099656619,
0.0178213548,
0.0000917522,
-0.033777874,
0.0105281575,
-0.0645659342,
0.0797406361,
-0.0028982796,
0.0281249788,
-0.0090772817,
0.0525268465,
-0.0195083655,
0.0732839629,
-0.0192875862,
0.0056545669,
-0.0319356881,
-0.0082646143,
0.0011464317,
0.004276304,
-0.0153511995,
-0.0010531328,
-0.0458876416,
0.0075472724,
-0.0071917996,
-0.0272006057,
-0.0298064519,
-0.0641087815,
-0.0423408039,
-0.0008216111,
0.0025398154,
-0.0057094535,
0.0079612276,
0.0454052463,
0.0286058392,
0.0189104173,
0.0279675052,
0.0107318508,
-0.024120653,
0.0473577529,
0.0083525842,
-0.0120686647,
-0.0189726017,
-0.0119071342,
-0.0223950874,
-0.0607025176,
0.0864321217,
-0.0207595062,
-0.0188365337,
-0.0026427873,
-0.0095854094,
-0.0243705548,
-0.0178296715,
-0.0142761609,
-0.0226665176,
0.0386461094,
0.0020677329,
0.0816226676,
0.0211895909,
0.0188473053,
0.0761686638,
0.021866953,
-0.0474697538,
-0.0685874894,
0.0270735379,
-0.0318041816,
-0.0108595276,
-0.0524804816,
0.0144129079,
-0.0013768587,
0.0889581516,
0.0363466889,
0.0248305239,
-0.0398432054,
0.0241473503,
-0.0313049555,
-0.038786754,
-0.0070518539,
-0.0136400219,
-0.0412742347,
-0.0008480826,
-0.0151050761,
-0.0053479173,
0.0489039384,
-0.0270949975,
-0.0148771405,
-0.0318229832,
0.0250239428,
-0.022005491,
0.0800957009,
-0.0404794924,
0.0406288244,
0.0156282224,
-0.0273862667,
-0.0480679162,
-0.0069305026,
-0.0335965045,
-0.0376201309,
0.0646896288,
-0.0322642364,
-0.0224983692,
0.0106372759,
-0.0207626205,
-0.0177063402,
-0.0062641734,
-0.0340570286,
0.0425655283,
0.0414680466,
0.0046083881,
-0.0719450712,
0.0225057881,
0.0176257975,
-0.0002886985,
-0.014987031,
-0.0228201561,
-0.0515433215,
0.0051992559,
0.060674835,
0.0501847379,
0.0359224267,
0.0028009533,
-0.017575888,
0.0285814591,
-0.0028000427,
-0.025243286,
-0.0315508358,
0.0163703319,
0.0279474948,
-0.0299654063,
-0.0043674158,
0.009621012,
0.0209414605,
-0.0714340881,
-0.0191637203,
0.0932384878,
0.0484803766,
-0.015927555,
0.0305666514,
-0.037365593,
0.0053390162,
0.01783875,
-0.0078609698,
-0.0728563145,
-0.0374775454,
-0.0012757799,
-0.0463228785,
0.0118508143,
0.0076170159,
0.0027952394,
-0.0837735161,
0.0197678357,
-0.0275818724,
-0.0394529477,
0.0077802045,
-0.0244205035,
-0.0072470112,
-0.0086058397,
0.0261008274,
-0.0229913313,
-0.028483443,
-0.0347334817,
-0.0185277965,
0.0002367886,
-0.0326013155,
0.0502839386,
-0.0092662778,
-0.0226034913,
0.0724735036,
0.0290168002,
-0.0502123795,
-0.0983757898,
0.0301110093,
0.0155989071,
0.0427449495,
-0.0125298211,
-0.0056997868,
0.0070821797,
-0.0140595306,
-0.0179095939,
-0.0294773746,
-0.0232636239,
-0.0629458874,
-0.0681430623,
0.0594334751,
-0.0217862744,
-0.0107712485,
-0.0235739015,
-0.0039236895,
-0.0247620847,
0.0728020221,
0.0089015113,
-0.0524102934,
-0.0326325074,
-0.0112749161,
0.0075065461,
-0.0176855065,
-0.0665500909,
0.0230478365,
0.0262289532,
-0.0273310766,
-0.0242312085,
-0.0248309895,
0.0223774798,
0.02573389,
-0.014694497,
-0.0386070721,
0.0094096186,
-0.0138279805,
-0.0581222698,
0.0217907727,
-0.016692888,
0.0417683423,
-0.015597959,
0.0214442275,
0.0240699947,
-0.0235982966,
-0.0526130907,
0.0239002779,
-0.0622350797,
-0.0089854617,
0.0503788963,
0.0429308973,
0.0219762065,
0.0011585641,
-0.019414328,
0.0070739002,
0.0199391935,
0.020553641,
-0.0447533987,
0.0107570011,
0.0538204983,
-0.0173611548,
-0.0153706539,
0.0237030406,
-0.0405215621,
0.0495974384,
0.0156051293,
0.0695620403,
0.0015703309,
-0.0440975279,
0.052167289,
-0.0413851701,
0.0740081295,
0.0159686767,
0.0164285395,
0.0163159315,
0.0092853112,
0.003265959,
-0.025379261,
-0.0351607017,
-0.0179886017,
0.0332962759,
0.0064200261,
0.0076543577,
-0.0064605377,
-0.0058526588,
0.0251804013,
0.0103385439,
-0.0142617933,
0.0294965766,
-0.0445419177,
-0.0095457733,
0.0335392989,
-0.0239203479,
0.012708365,
-0.0133831734,
0.0003005097,
-0.0219262764,
0.0537394099,
-0.0140972892,
-0.0080530085,
-0.0801483095,
0.022859633,
-0.0435191765,
-0.0113049028,
-0.0242219623,
-0.0434447043,
-0.0136266053,
-0.0109547731,
0.0398056433,
-0.0019645039,
-0.0049123806,
-0.000031088,
0.002080742,
-0.0107481107,
-0.0229061544,
0.0723721609,
0.0249699559,
0.0052598156,
-0.0493039899,
0.0380585901,
-0.0721418262,
0.0774910077,
-0.02166491,
-0.0168781392,
-0.0375057645,
-0.0446628034,
0.031291917,
0.022322638,
-0.0020945482,
-0.0180452671,
-0.098374106,
-0.0230264757,
0.057181865,
0.0074111298,
0.04046572,
-0.0484341346,
0.0572339855,
0.0931701213,
-0.007094807,
-0.0243316069,
-0.020778846,
-0.0127040008,
0.014539293,
0.0129533131,
-0.0266515967,
-6.13091605e-33,
-0.0468104519,
0.0085407877,
0.0282259323,
-0.0010124254,
-0.0245136805,
0.0123818833,
-0.0197489448,
0.0016954662,
-0.0180379432,
0.0195613895,
-0.0083114319,
0.0115131773,
0.0295833517,
0.0080366684,
0.0339984745,
-0.0195236411,
0.0332304053,
0.0240637176,
-0.0240233764,
-0.0229854751,
0.0267706681,
0.0185773876,
0.0815327018,
-0.0394323692,
0.0305972584,
0.0569554493,
-0.0216078274,
-0.0286502019,
-0.0128842276,
0.0955764949,
-0.0360091962,
0.0346313901,
0.0318471566,
-0.0438932814,
-0.0232812911,
0.0355145372,
-0.0534225442,
-0.003291152,
0.0454711355,
0.0185365565,
-0.0256096125,
-0.0429406278,
0.0679403171,
-0.0189000573,
-0.012011827,
-0.032205753,
0.0122796642,
-0.0031274159,
-0.0196218826,
-0.0475364029,
-0.103483066,
0.0036004183,
0.0370437689,
-0.0322902091,
-0.010737828,
0.0028117038,
-0.0126840807,
0.0490881056,
-0.0591725446,
0.0156319812,
-0.0266861394,
0.0055791908,
0.0421135537,
0.0480003804,
0.0023261858,
0.063817732,
0.0126762241,
-0.0275144447,
0.0206297282,
0.0034779566,
0.0483179837,
0.0715188459,
0.0249391422,
0.0081798304,
0.0428095907,
0.0233791042,
0.0152311036,
-0.0141280368,
-0.0390858576,
0.0767584294,
0.0493770204,
0.0140336622,
0.0000285811,
0.0157875735,
-0.0543587655,
0.0261991117,
-0.0505903475,
-0.0158420987,
0.0174873043,
0.021683054,
-0.0234560519,
0.0849192739,
0.0380782895,
-0.0254202392,
-0.0064970804,
-0.0253061038,
0.0028772398,
0.0299905371,
0.0073560243,
0.0232760683,
0.0029885888,
-0.0339081511,
0.0416163169,
0.0094660278,
0.0282938983,
0.0309947189,
0.0138377389,
0.0426459312,
-0.0599757433,
-0.019409541,
-0.0303671826,
-0.0474947654,
-0.0220965315,
-0.0127676399,
0.0445435755,
0.0372937098,
-0.0400057808,
0.0088485973,
0.0065496461,
-0.0012111133,
-0.0252224822,
0.0109863104,
-0.0133786909,
0.0497897416,
-0.0152300224,
-0.0118294507,
-0.0486467183,
-0.0245578922,
0.0549547598,
-0.0429819301,
-0.0222755838,
0.0343448222,
2.775e-7,
0.095371291,
0.0567153133,
0.0297826193,
0.0352495573,
0.0009358256,
0.0191875603,
0.0098056123,
0.0354209878,
-0.0291951913,
0.0118530141,
-0.017512178,
0.0346080437,
0.0050526378,
0.0087105902,
-0.0329500809,
0.0248016436,
-0.0537610054,
0.0141367074,
-0.0463958234,
0.033212997,
0.0777370781,
0.1050091535,
-0.0044521275,
-0.0142972982,
-0.0367081724,
-0.000066637,
0.0613053702,
-0.0202570613,
0.0142265568,
0.0267551634,
-0.0121181887,
-0.0070365435,
0.0117450869,
0.0200697761,
0.0025896335,
0.0367995091,
0.0417153463,
0.0490198359,
-0.033704225,
-0.001332238,
0.0049273064,
-0.0513755195,
-0.0184615534,
-0.0395814478,
0.0628320128,
-0.1065744236,
-0.0197236445,
-0.0452003479,
-0.0123306839,
0.0606675744,
0.0134655731,
-0.0100845173,
-0.0118750092,
-0.0358694196,
0.0203391984,
0.0131564429,
-0.0293348301,
-0.07908573,
-0.0150397541,
0.0162176266,
-0.0388607606,
-0.0097176535,
-0.0656377673,
-0.061578203,
0.0661114827,
-0.0399849378,
-0.0436260663,
2.430830941e-34,
0.0171298943,
0.0008024449,
0.0035060099,
0.0322659984,
0.0310087539,
-0.0120029096,
0.0136442175,
0.0397060253,
-0.0258323029,
-0.0716866255,
-0.0127576562
] |
http://arxiv.org/abs/2401.15391v1 | http://arxiv.org/pdf/2401.15391v1 | MultiHop-RAG: Benchmarking Retrieval-Augmented Generation for Multi-Hop Queries | 2024-01-27T11:41:48 | 2024-01-27T11:41:48 | Yixuan Tang, Yi Yang | cs.CL | cs.CL | Retrieval-augmented generation (RAG) augments large language models (LLM) by
retrieving relevant knowledge, showing promising potential in mitigating LLM
hallucinations and enhancing response quality, thereby facilitating the great
adoption of LLMs in practice. However, we find that existing RAG systems are
inadequate in answering multi-hop queries, which require retrieving and
reasoning over multiple pieces of supporting evidence. Furthermore, to our
knowledge, no existing RAG benchmarking dataset focuses on multi-hop queries.
In this paper, we develop a novel dataset, MultiHop-RAG, which consists of a
knowledge base, a large collection of multi-hop queries, their ground-truth
answers, and the associated supporting evidence. We detail the procedure of
building the dataset, utilizing an English news article dataset as the
underlying RAG knowledge base. We demonstrate the benchmarking utility of
MultiHop-RAG in two experiments. The first experiment compares different
embedding models for retrieving evidence for multi-hop queries. In the second
experiment, we examine the capabilities of various state-of-the-art LLMs,
including GPT-4, PaLM, and Llama2-70B, in reasoning and answering multi-hop
queries given the evidence. Both experiments reveal that existing RAG methods
perform unsatisfactorily in retrieving and answering multi-hop queries. We hope
MultiHop-RAG will be a valuable resource for the community in developing
effective RAG systems, thereby facilitating greater adoption of LLMs in
practice. The MultiHop-RAG and implemented RAG system is publicly available at
https://github.com/yixuantt/MultiHop-RAG/. | MultiHop-RAG: Benchmarking Retrieval-Augmented Generation for Multi-Hop Queries: Retrieval-augmented generation (RAG) augments large language models (LLM) by
retrieving relevant knowledge, showing promising potential in mitigating LLM
hallucinations and enhancing response quality, thereby facilitating the great
adoption of LLMs in practice. However, we find that existing RAG systems are
inadequate in answering multi-hop queries, which require retrieving and
reasoning over multiple pieces of supporting evidence. Furthermore, to our
knowledge, no existing RAG benchmarking dataset focuses on multi-hop queries.
In this paper, we develop a novel dataset, MultiHop-RAG, which consists of a
knowledge base, a large collection of multi-hop queries, their ground-truth
answers, and the associated supporting evidence. We detail the procedure of
building the dataset, utilizing an English news article dataset as the
underlying RAG knowledge base. We demonstrate the benchmarking utility of
MultiHop-RAG in two experiments. The first experiment compares different
embedding models for retrieving evidence for multi-hop queries. In the second
experiment, we examine the capabilities of various state-of-the-art LLMs,
including GPT-4, PaLM, and Llama2-70B, in reasoning and answering multi-hop
queries given the evidence. Both experiments reveal that existing RAG methods
perform unsatisfactorily in retrieving and answering multi-hop queries. We hope
MultiHop-RAG will be a valuable resource for the community in developing
effective RAG systems, thereby facilitating greater adoption of LLMs in
practice. The MultiHop-RAG and implemented RAG system is publicly available at
https://github.com/yixuantt/MultiHop-RAG/. | [
0.047234904,
0.0557873622,
-0.009452194,
0.0315068997,
-0.0657770932,
-0.0130146276,
-0.0119048674,
-0.0077342968,
0.0275741257,
-0.0499545895,
-0.0125535782,
-0.0112585844,
-0.0433591604,
-0.0023469988,
0.0274854507,
0.0187110063,
0.0664773583,
0.0118968086,
-0.0131698633,
-0.0072272052,
0.0080577135,
0.047022488,
0.0016202402,
0.0442593917,
-0.0238298066,
-0.0100862375,
-0.0159736592,
0.0031907808,
0.0087758768,
-0.0154874548,
-0.0136815207,
0.0503410846,
0.0384523943,
0.0467689,
0.0000018353,
-0.0477038845,
-0.0112603828,
0.013071469,
-0.0248113014,
0.0010191391,
0.0024597053,
0.0400491059,
0.0258608852,
0.0370592512,
-0.0461258627,
0.0029458832,
0.0153699834,
0.0569665469,
0.0500420891,
0.0743169263,
0.0049831686,
-0.0415567458,
0.0239378512,
-0.0209880266,
0.0606069714,
-0.0310981683,
0.0209079478,
0.0292682163,
0.0577721,
0.0361099094,
0.0193368029,
0.0323975459,
0.0193335395,
-0.0110125914,
-0.045043651,
0.016865164,
-0.0692451224,
0.0001905786,
0.0115708234,
0.0756167322,
0.0331176519,
-0.0344002694,
0.0029088487,
0.0222313013,
0.0117144035,
0.0751368254,
-0.0510757379,
0.0002803302,
-0.0363795199,
0.0200472418,
0.0117508508,
-0.0140311113,
0.0465074144,
0.0279076789,
-0.0188533682,
0.0991105959,
0.0149322636,
-0.0360803343,
0.0079256203,
0.0084313201,
-0.0071002645,
-0.0267655738,
0.0065090731,
0.0319497697,
-0.0392852686,
0.0281804688,
-0.0079505872,
-0.0157739297,
0.0659862533,
-0.0313147865,
0.0039440808,
0.0351600423,
-0.0358120315,
0.0689390153,
-0.0378341973,
-0.0092581129,
-0.0497990586,
-0.0055640861,
-0.0408580527,
-0.0371829718,
-0.0534870811,
-0.0341900736,
-0.0227238182,
0.0470204093,
-0.0607364178,
-0.0156313293,
-0.0000390246,
0.0193739757,
0.0045728027,
-0.018519558,
-0.0067672296,
0.021997869,
-0.0199384466,
-0.0054123257,
-0.0395769142,
0.0484628826,
-0.0409107246,
0.0360761993,
0.0422305129,
-0.0759272352,
-0.0187358279,
0.0255570151,
0.0093457513,
-0.0160018485,
0.015251643,
0.0385906994,
-0.009151849,
-0.0190681946,
-0.0358205326,
-0.0834825858,
-0.0150793316,
-0.0720341355,
-0.0016009625,
-0.0686219484,
0.0519316159,
0.0079769343,
-0.0293799527,
-0.0078486158,
0.0090733441,
0.0667802468,
-0.0551723801,
0.0180632807,
0.0024055617,
0.0046411809,
0.035053011,
0.0205495525,
-0.0060420819,
0.1010007709,
-0.0091874925,
0.0441111177,
0.0405934602,
0.0161193162,
-0.0065031243,
0.0015663857,
0.0393646471,
0.0363644063,
-0.0188980233,
0.0364561193,
-0.0146234687,
0.0227337461,
-0.0072145327,
0.0533573143,
-0.0626737028,
-0.0130945444,
0.0199132748,
0.0111855632,
0.0407070853,
-0.029333245,
0.0629963428,
0.0049813553,
0.0083602602,
0.0134146903,
-0.0178590789,
0.0271582026,
-0.0523330048,
0.0153124779,
0.0703974739,
0.0397914499,
-0.0309034698,
-0.0205017198,
-0.0153832408,
0.0200208314,
-0.0491804853,
0.0243707541,
0.0258735046,
-0.0010518248,
-0.0275747161,
0.0450122468,
0.0169092063,
-0.0478050075,
0.0209091008,
-0.0447333008,
0.0442330427,
0.0464605317,
0.0847343728,
-0.0282148365,
-0.0063892924,
-0.0024950772,
0.0031997848,
0.0208176896,
0.0424778946,
-0.0140430359,
-0.0088028396,
-0.0441451557,
-0.0252872985,
-0.0223818589,
0.0189680494,
0.0035193195,
0.0275321584,
0.0076505593,
0.0429974794,
-0.0433705635,
0.0139338793,
0.0129863005,
-0.0338171609,
-0.0004812926,
0.0061494112,
-0.0193394776,
0.0182314999,
0.0602518618,
-0.003884953,
0.0499573164,
0.0492104664,
-0.0206881296,
-0.0360121801,
-0.0074432618,
-0.0008810952,
-0.0455612205,
0.0175055023,
-0.0279187653,
0.0185403787,
0.0031892264,
-0.0251934677,
-0.0681116357,
-0.0114013469,
-0.0318709947,
0.0528045185,
-0.0759263262,
-0.0238837488,
0.0154369017,
0.005908092,
-0.0416359417,
0.0808521062,
-0.0133724948,
0.0122538581,
-0.0050044893,
0.0229782872,
0.0226612184,
-0.0277269818,
-0.0307913814,
0.0146956276,
0.0085675344,
0.0336884111,
0.0385506116,
-0.0114437416,
-0.062703535,
-0.0079930555,
-0.087796025,
0.0212170798,
-0.0267243907,
0.0104329037,
0.0095836837,
0.0190829914,
-0.0128573803,
0.0018961567,
-0.0656283274,
-0.0004956042,
0.0319867991,
0.0584628135,
0.0031333831,
0.0326533653,
0.0075917752,
-0.0220755562,
-0.0154085523,
-0.0145068523,
0.0406550616,
-0.022655338,
0.0035176938,
-0.0274460297,
-0.0171959046,
-0.0722071081,
0.0064452812,
-0.0196235292,
0.0010862804,
0.036323823,
-0.0549546219,
0.0293690506,
-0.0108037768,
0.0035629412,
-0.0194689706,
0.0394541621,
-0.0151011376,
-0.0790425017,
0.0175902788,
-0.0594302453,
0.0075003183,
0.0073560895,
-0.0246887896,
0.0142490044,
0.0291295182,
0.0191819966,
-0.016515553,
-0.0502852499,
0.0068922713,
0.0218997877,
-0.0286372993,
-0.0495613329,
-0.0262536723,
0.01891228,
0.0526079834,
-0.0063907895,
-0.0449964106,
0.0810853541,
-0.0276197642,
0.0371146575,
-0.0503261983,
0.0138870617,
-0.0403421037,
0.0580511019,
-0.033679422,
0.0379886962,
-0.0476652645,
-0.0105295973,
-0.0125870053,
0.0021718137,
0.0199535508,
-0.0664238334,
0.036338713,
-0.0298939999,
0.0314291492,
0.0102132075,
-0.0192986466,
-0.0312068202,
-0.0377360694,
-0.022473786,
-0.0194428097,
-0.0234327912,
-0.0030041728,
-0.0362592153,
0.0185062382,
0.0470290296,
0.0076764319,
-0.0208407287,
-0.0022813599,
0.0074473447,
0.0200022254,
0.078983061,
-0.0008261912,
0.0317494646,
0.0166062135,
0.0144426003,
0.0687190518,
0.0435609184,
-0.0242509861,
0.0302322302,
-0.0118071185,
0.0185832363,
-0.0081403553,
-0.0229871068,
0.0014252436,
0.0156177348,
-0.035155151,
-0.0078156106,
0.0622640736,
0.0554354154,
-0.0445609689,
-0.0153307104,
-0.0284858476,
0.0320068039,
0.0137206782,
0.0461333916,
-0.1027522162,
-0.037793681,
-0.0021444494,
0.0126175387,
-0.011654309,
-0.0237252954,
-0.0150223188,
-0.1420318335,
0.0393079259,
-0.003569586,
-0.0106001571,
-0.0170039032,
-0.0614480004,
0.0042295908,
0.020207515,
0.0175220612,
0.0492664315,
0.0161634181,
0.0042354628,
0.0033622072,
0.0321649984,
-0.0434609801,
0.0208846871,
-0.0081078,
0.0020750172,
0.0564770438,
0.0126075251,
0.0278621726,
-0.0561774559,
0.0154836336,
-0.0196546111,
0.0111480448,
-0.0214074608,
0.0351628251,
-0.0121678999,
0.0134318573,
-0.0410633795,
0.0227568168,
-0.0353858098,
-0.0149077699,
-0.0058652661,
0.088423714,
-0.0378668606,
0.0282544158,
-0.0700589418,
0.0159657821,
0.0044777966,
0.0962352157,
0.0152521757,
0.0050952588,
-0.0265283044,
-0.0006446691,
0.0061938344,
-0.029485371,
-0.0115281353,
0.0458079875,
-0.0030528612,
0.0135269454,
-0.0686578676,
-0.0427789874,
0.0050588665,
0.0501746386,
-0.0155846206,
0.0431423672,
-0.013408876,
-0.0171681028,
-0.0390571728,
0.0144336466,
-0.0430054553,
0.0034007749,
-0.0132748783,
0.0306542013,
0.0553324632,
-0.0201074947,
-0.017495729,
0.015976239,
-0.0533950143,
-0.0713633969,
-0.0156255011,
0.0069697909,
0.0296995882,
0.0214453712,
-0.0297382139,
-0.0372891538,
0.0125462553,
0.0394493565,
-0.0418779477,
-0.0180992819,
-0.0005215193,
-0.0168879963,
-0.0201539323,
0.0383669063,
-0.0685507134,
0.0122954957,
-0.0369317308,
0.070447959,
0.0007012107,
-0.0114238914,
0.0468743853,
0.0302570593,
0.035862945,
0.0007363627,
0.0060527786,
-0.0208149906,
0.0188282058,
-0.0691472068,
-0.0298918262,
-0.0123355445,
0.0182999875,
0.0024352265,
-0.0695712045,
0.011116026,
-0.0218781922,
-0.0255755484,
0.0748863295,
-0.0362606198,
-0.0147529794,
-0.0067543471,
-0.0132650258,
0.0253237374,
-0.0229572114,
-0.0363590829,
0.0052218246,
-0.0051928139,
0.0553755984,
-0.0301847029,
0.0525688119,
0.0188791193,
-0.0320365168,
0.0118722757,
0.0012199128,
-0.0927468985,
-0.0363486148,
-0.0181427486,
0.0002313787,
0.0321877412,
0.0030018839,
0.0599571057,
0.0005612763,
-0.024067292,
-0.0098539433,
0.0108098285,
0.0312218051,
-0.0198774636,
0.0319165587,
0.0390316211,
-0.0335191712,
0.0438989885,
0.0829464272,
-0.0284353122,
0.0460518524,
-0.0195179507,
-0.050159514,
-0.0285707694,
-0.0701468587,
-0.0027142959,
0.0403644778,
-0.0322801098,
-0.0147584556,
-0.0897383168,
-0.0406717621,
-0.0452362821,
0.0126876067,
-0.0049410076,
0.002276378,
0.0124339368,
0.0994502902,
0.0333200507,
-0.0286717284,
-0.0582760088,
0.0280366577,
-0.0708587244,
-0.015559284,
0.0254869219,
-6.593776135e-33,
-0.0104764523,
-0.0194467548,
0.0553523079,
-0.0738992319,
-0.0311036129,
-0.007811219,
-0.018488178,
-0.0170731712,
-0.0634032935,
0.0025596167,
-0.0087722251,
0.0168076288,
0.0137946187,
0.015948575,
0.001632869,
-0.0017911647,
0.0136707257,
0.0222264342,
-0.0215117875,
0.0501234159,
0.0932592452,
0.0041932091,
0.0267846677,
-0.0086975535,
0.0158030149,
-0.0304804053,
-0.0069352034,
0.026828954,
-0.0488380902,
0.0689246878,
-0.0133393873,
0.0396599621,
0.0243990254,
-0.0389983468,
0.0046524322,
0.0132018859,
-0.0603260808,
-0.0328144468,
0.0349223576,
-0.0496224873,
-0.03016131,
-0.009731086,
0.0202798005,
-0.0065946537,
0.0217731912,
0.0137428502,
0.0005302728,
-0.0168633275,
-0.0423897281,
-0.0346285105,
-0.0339195319,
0.0346811041,
0.0063026673,
-0.002556819,
-0.0337192602,
0.1001182273,
0.0192532893,
0.0314929448,
-0.0679956973,
0.0415816754,
-0.0169828702,
0.0226757228,
0.0597660132,
-0.0211036019,
-0.0091352342,
0.0471106209,
-0.0098576974,
-0.0199216064,
-0.0123101622,
0.0109786922,
0.007192737,
0.0428945124,
0.0138802677,
0.0672916621,
0.033925429,
-0.0191680789,
-0.0878716558,
0.0385724828,
0.0262135956,
0.0436093174,
0.0435037874,
-0.0001074546,
-0.0547442958,
-0.0184777081,
-0.0370233804,
0.0176261552,
-0.0409367383,
-0.0457250625,
0.0289148688,
0.0384879336,
-0.0265270472,
0.0150401751,
0.0471563376,
-0.0321003683,
-0.0092520509,
0.0458259247,
-0.0602464005,
-0.0191555098,
-0.0161764007,
0.0565611646,
-0.0375446342,
-0.0067541264,
0.0297442041,
-0.0089007048,
-0.0228031948,
0.0189639404,
0.0419596322,
-0.0059910859,
-0.0422955565,
0.0023523881,
-0.0520983674,
0.0296034571,
-0.0366142914,
-0.0632303283,
0.0424917899,
0.0273815393,
0.0036275329,
0.0406901911,
0.0445780233,
0.0198594313,
-0.0157818682,
-0.0327486508,
-0.0010055116,
0.0026807336,
-0.0070760036,
0.0403136536,
-0.0133911045,
0.0121483477,
0.0259875003,
-0.0415626504,
-0.0171276238,
0.0568848662,
2.681e-7,
0.0488834828,
0.0153279258,
0.0414272361,
0.0169086438,
-0.0065800459,
0.0062095546,
0.0040895334,
0.0258141644,
0.0211762581,
-0.0422560126,
-0.0216205698,
0.0148235122,
0.0337458812,
0.0580228008,
-0.0687135383,
-0.0242378954,
-0.0717385784,
-0.0415389352,
-0.0668392703,
-0.0263007972,
0.0802338943,
0.0628264323,
-0.000716462,
-0.0184729341,
-0.019643791,
-0.0171735734,
0.0686919168,
-0.0000447939,
-0.071081616,
0.0183567051,
0.0572694056,
0.0120642809,
0.0301570687,
-0.0024426533,
-0.0257283654,
0.0232227873,
-0.0124648605,
0.0286686737,
0.0197275486,
-0.0649650097,
-0.0052854591,
-0.0506861471,
0.0281045437,
-0.0596075095,
0.1100472659,
-0.0078130467,
-0.0260205995,
-0.0183322225,
-0.0475162603,
-0.0127233481,
0.0169487875,
-0.0243569165,
-0.0248389114,
0.0052763461,
-0.0060348362,
0.018827403,
0.0056704213,
-0.0888092741,
-0.0134679722,
0.0298030302,
-0.0521564037,
-0.04287415,
-0.0169042237,
-0.031351395,
-0.0090373792,
-0.0627059788,
-0.006488956,
2.634007793e-34,
0.0442203693,
0.0091772219,
0.0169259366,
0.0081911338,
0.0578978285,
0.0008428833,
-0.0316737145,
0.0007793991,
0.0142063471,
-0.1068250537,
-0.0238945242
] |
http://arxiv.org/abs/2401.15385v1 | http://arxiv.org/pdf/2401.15385v1 | Towards Event Extraction from Speech with Contextual Clues | 2024-01-27T11:07:19 | 2024-01-27T11:07:19 | Jingqi Kang, Tongtong Wu, Jinming Zhao, Guitao Wang, Guilin Qi, Yuan-Fang Li, Gholamreza Haffari | cs.CL | cs.CL, cs.MM | While text-based event extraction has been an active research area and has
seen successful application in many domains, extracting semantic events from
speech directly is an under-explored problem. In this paper, we introduce the
Speech Event Extraction (SpeechEE) task and construct three synthetic training
sets and one human-spoken test set. Compared to event extraction from text,
SpeechEE poses greater challenges mainly due to complex speech signals that are
continuous and have no word boundaries. Additionally, unlike perceptible sound
events, semantic events are more subtle and require a deeper understanding. To
tackle these challenges, we introduce a sequence-to-structure generation
paradigm that can produce events from speech signals in an end-to-end manner,
together with a conditioned generation method that utilizes speech recognition
transcripts as the contextual clue. We further propose to represent events with
a flat format to make outputs more natural language-like. Our experimental
results show that our method brings significant improvements on all datasets,
achieving a maximum F1 gain of 10.7%. The code and datasets are released on
https://github.com/jodie-kang/SpeechEE. | Towards Event Extraction from Speech with Contextual Clues: While text-based event extraction has been an active research area and has
seen successful application in many domains, extracting semantic events from
speech directly is an under-explored problem. In this paper, we introduce the
Speech Event Extraction (SpeechEE) task and construct three synthetic training
sets and one human-spoken test set. Compared to event extraction from text,
SpeechEE poses greater challenges mainly due to complex speech signals that are
continuous and have no word boundaries. Additionally, unlike perceptible sound
events, semantic events are more subtle and require a deeper understanding. To
tackle these challenges, we introduce a sequence-to-structure generation
paradigm that can produce events from speech signals in an end-to-end manner,
together with a conditioned generation method that utilizes speech recognition
transcripts as the contextual clue. We further propose to represent events with
a flat format to make outputs more natural language-like. Our experimental
results show that our method brings significant improvements on all datasets,
achieving a maximum F1 gain of 10.7%. The code and datasets are released on
https://github.com/jodie-kang/SpeechEE. | [
-0.0274351686,
0.0788979158,
0.0288546067,
0.0596804507,
-0.0175610054,
-0.0098362165,
-0.015264974,
-0.0013266684,
0.0067225238,
-0.0651567578,
0.0214083884,
0.0208670069,
-0.0214081407,
0.0786133334,
0.0022976869,
-0.0875076577,
0.049914442,
-0.0025298218,
-0.0074104113,
0.0446632504,
0.0207217764,
0.0355011001,
-0.0518594459,
0.0301147159,
-0.0488403328,
0.033626426,
-0.0364243351,
-0.0430678688,
0.0291263722,
-0.0465263054,
-0.0106084906,
-0.001693884,
0.0622185282,
-0.0097312275,
0.0000018701,
-0.0540805161,
-0.025562048,
0.0061539346,
-0.0557263605,
-0.0266724117,
-0.0253884867,
-0.0068302546,
-0.0233385935,
0.012709368,
-0.0216547567,
-0.0266604666,
0.0013993059,
0.0724258199,
-0.0144357663,
0.0958532542,
-0.0250306278,
-0.0510569662,
0.0394602679,
-0.0036184657,
0.015482395,
-0.0494204313,
-0.0077740313,
-0.0187412016,
0.0228465665,
-0.0158700403,
0.0082798284,
-0.0017383446,
0.0169378407,
0.0155651858,
-0.0108943572,
0.0588685833,
-0.0423116498,
-0.0203587525,
-0.0477840044,
0.0597017407,
-0.0086494396,
0.0170223098,
0.0603250675,
-0.0028453337,
-0.0132976389,
-0.0271451883,
-0.0601629987,
0.0378145203,
-0.0124811865,
0.0250383541,
0.0403377414,
0.0522023253,
0.0095800273,
-0.0159305669,
0.0131526375,
0.0265901722,
0.0155470781,
0.0362868458,
-0.0533021986,
-0.0114456145,
0.0021490715,
-0.0460504629,
0.0439888351,
0.0082893185,
-0.0572206788,
-0.0000191759,
-0.0696559623,
0.0197734702,
-0.0247500315,
-0.0782735869,
-0.0156253055,
0.0456502587,
-0.0382703468,
0.0311333071,
-0.0052533913,
0.0379114784,
-0.0057294238,
0.0159354769,
-0.0361036994,
0.0009311872,
0.0022997514,
0.0673277006,
0.0011268248,
0.0195983201,
-0.0869137421,
-0.0308531635,
-0.0200431086,
-0.0033726697,
0.040000312,
-0.001416077,
-0.0257299878,
0.0258013196,
-0.0570025891,
0.0080165314,
0.0353140496,
-0.0510639139,
-0.0407097861,
0.0510934293,
0.0198165234,
-0.0527996272,
-0.0245749671,
0.0359583274,
0.0838037357,
0.0029142965,
-0.0024897738,
0.0488347821,
-0.0365146175,
-0.0237653758,
-0.0724179819,
-0.0552540049,
-0.0139846643,
-0.0305292252,
-0.0386180617,
-0.0322369561,
0.0491517372,
-0.0239405781,
-0.0427066498,
-0.0113703012,
0.045034457,
-0.0370573178,
0.0082594845,
0.0484853387,
-0.0265942216,
0.0049681277,
0.0084904199,
0.0278970283,
0.0123380013,
0.0435048342,
0.0981201828,
0.031688761,
0.072510615,
-0.007449185,
-0.0192488786,
0.0312188379,
0.0109270373,
-0.0330151245,
0.0067352266,
-0.0070240861,
0.0212297812,
0.0149369789,
0.0784991458,
-0.0131496787,
-0.0181273278,
-0.0110395486,
0.0195923802,
0.028431423,
-0.0033832616,
-0.0039343061,
0.0252353009,
0.0120291747,
-0.0341332704,
0.0252786744,
-0.0150978034,
-0.0311433952,
-0.0103315115,
0.0121503547,
0.0427377559,
-0.0149782067,
-0.0115226721,
0.004741054,
0.0113361925,
-0.0285070688,
0.002804782,
0.0224767067,
0.0260787457,
0.0025941643,
-0.053957805,
0.0290411208,
0.0059119761,
-0.0872335508,
0.040547844,
0.0010288123,
0.0138880704,
0.0260882732,
0.0242219158,
-0.0373789221,
-0.0411368981,
0.0041448055,
-0.0118979029,
0.047243014,
0.0045898999,
-0.034525279,
0.1014864221,
-0.0567878038,
-0.0298551265,
0.0219492484,
-0.002450923,
-0.0030921802,
0.0149538945,
0.0193842687,
0.019612439,
0.0353328958,
0.0143520143,
0.0432940125,
-0.0286533404,
-0.067251794,
-0.0015700067,
-0.0210784972,
0.0566790998,
0.0233871173,
-0.0070559075,
-0.0134053351,
0.0789055154,
-0.0737774521,
-0.0091537256,
0.0721118674,
-0.0554307029,
-0.0627414286,
0.0117089814,
0.0187407974,
0.0086334329,
0.0265097413,
-0.0410225466,
-0.0664032549,
0.0049386765,
0.0129641537,
0.0515724532,
-0.0185246635,
0.007888318,
0.035097193,
-0.002356685,
-0.0052341362,
0.0197032578,
0.0182139836,
-0.0323436372,
-0.0177102294,
0.0168782342,
-0.0148212574,
-0.0591745339,
0.0149423582,
0.0275692604,
-0.0131405434,
-0.0091991881,
0.0455029272,
-0.0035853409,
-0.0510725044,
0.0243882835,
-0.0901450366,
0.0292996205,
-0.0174059272,
0.0561046079,
0.0203581732,
0.0229808409,
-0.0124629531,
-0.0314876027,
-0.0375857353,
-0.078596577,
0.0076461956,
-0.0348275751,
-0.0056152721,
0.0235602241,
-0.0621278323,
-0.0697551295,
0.030703878,
-0.0207492299,
0.0204912648,
0.0299720839,
-0.0180920083,
-0.0341363214,
0.0432344638,
0.0461909436,
0.0586030744,
-0.0453995503,
0.0089826025,
-0.0034497113,
-0.0064978921,
0.0379444696,
0.0377342924,
0.0296065006,
0.0546289608,
-0.0268477499,
-0.0633983985,
-0.0372699238,
0.0198272038,
-0.0468620993,
0.0485773198,
-0.0369156934,
0.0078055058,
0.0181429032,
0.0276217908,
0.0542227924,
0.0545053147,
-0.0100921597,
-0.0578184463,
0.0047869212,
-0.039436534,
-0.0192596484,
-0.0363055058,
-0.0643413141,
0.0364703275,
0.0066331667,
0.0031209157,
0.0227101464,
-0.0598490611,
-0.047768008,
-0.0400683731,
-0.0244462099,
-0.0468879677,
0.0807128176,
-0.0300370771,
0.0024175458,
0.0093829064,
-0.0698281005,
0.0061679678,
-0.0302396547,
-0.0054886793,
-0.0403167419,
0.0021314789,
-0.0270080883,
-0.0054190587,
-0.0267340466,
-0.0203272272,
-0.0807528123,
-0.0245073754,
0.0030773988,
0.0089568021,
-0.0031279868,
-0.0111214407,
-0.0324363671,
0.0033475384,
-0.0164454561,
0.054821372,
0.0130864773,
0.0277713556,
-0.0116727734,
0.1119976044,
0.0382335968,
0.0473854877,
-0.0455089845,
0.0015980264,
-0.0291782115,
0.055260621,
0.0722828507,
0.0361897759,
-0.0326319486,
-0.0084426487,
0.0171850622,
-0.0248059668,
0.0044293958,
0.0352695659,
0.0410290323,
-0.0769253522,
0.0125318849,
0.0944525972,
0.0088578621,
0.0298317932,
0.0195779279,
0.0559175387,
0.0180933475,
0.0076221707,
-0.0247513093,
-0.0599217527,
-0.0651074648,
0.0299077015,
0.0249023605,
-0.0081563517,
0.0196256172,
-0.0051124478,
-0.0403233804,
0.0326600112,
0.018875476,
-0.0372931063,
-0.0622946694,
-0.0410367362,
0.0094726672,
-0.0118994769,
-0.0081673265,
0.0203075856,
0.0036141931,
-0.0193130933,
0.0123417638,
-0.0722032338,
-0.0819125324,
0.0353313349,
-0.0324296243,
0.0350077078,
0.0437914506,
0.0073924209,
0.0039870134,
0.0195104182,
0.0454492941,
-0.0063110068,
-0.0224039555,
0.0234388523,
0.0195776187,
-0.0127847968,
-0.0019700036,
0.0045714295,
0.0796974674,
-0.0214253999,
-0.0309164356,
0.0316484869,
0.0591831729,
-0.0205669627,
0.0018451686,
-0.0147213889,
0.0236074924,
0.0320753567,
0.0362452306,
0.0063851243,
0.0084229233,
0.0147800222,
0.016006548,
0.0251712743,
-0.0644641817,
-0.0238764919,
-0.0010477263,
-0.0049303938,
-0.0331922434,
-0.0800938606,
-0.0493397675,
-0.0167642497,
-0.0005505741,
-0.0288360119,
-0.0222385433,
0.0216131229,
0.0110342447,
-0.0612417385,
0.0671448484,
0.0182594601,
0.0063167866,
-0.0349290036,
-0.0106527144,
0.0123496931,
-0.011534756,
-0.0288436599,
0.0971114188,
-0.0426811539,
-0.0308020338,
0.0046936916,
-0.0272399262,
0.0034065316,
0.0727591887,
0.0207755957,
0.0346839465,
0.0146496091,
0.0110941865,
-0.0444159396,
-0.0060539735,
0.0308681894,
-0.0028653434,
-0.0370249599,
0.0021778834,
-0.0338831842,
0.0292003546,
-0.0361933,
0.0794993713,
-0.0483224802,
-0.0193409231,
0.0057159327,
-0.0046917438,
0.0703281388,
-0.0336072519,
-0.0024658798,
-0.0513882637,
0.0036725665,
-0.0352026932,
0.0280307848,
-0.0603928976,
0.0267481487,
-0.025471244,
0.0321746916,
-0.0060756514,
0.0023674553,
-0.0582423955,
-0.0282005724,
-0.0504514314,
0.0549721643,
0.0387290753,
0.0080695571,
-0.0045368713,
-0.020572789,
-0.0187848695,
0.0291464031,
-0.0101282885,
0.0185889304,
-0.03467324,
0.031167116,
0.019230146,
-0.0184951592,
-0.0001817083,
-0.018074479,
-0.0350310653,
0.0739965886,
-0.052943971,
0.0107106147,
0.0574572943,
-0.0214418154,
0.0318375789,
0.0039756387,
0.0174501482,
-0.0093379505,
0.0760820359,
-0.0242050495,
-0.0274670869,
0.027149342,
0.0129287085,
-0.0291864872,
0.0086387377,
0.0388645902,
-0.0122743044,
0.0431000888,
0.0049606152,
0.0204577055,
-0.0419462733,
-0.052791018,
0.0408591069,
0.0482760035,
-0.0219309963,
-0.0137519231,
-0.0230897255,
0.0176436249,
0.0538679585,
0.0064591244,
0.0172198322,
-0.0075214221,
0.0335318856,
0.002983368,
-0.0256323256,
-0.0166387334,
0.0044612624,
-0.0019589278,
-0.0126372622,
-0.0383904874,
-0.027935775,
-5.713679473e-33,
-0.0398025364,
-0.0463462174,
0.0194876138,
0.0889502391,
-0.0487859212,
0.0025811961,
-0.0534351282,
0.0043782154,
0.0289242584,
0.000436753,
-0.002432198,
-0.0171272755,
0.0245409124,
0.027954001,
0.0161866043,
-0.0249717906,
0.0516529232,
0.0023405091,
0.0066487989,
0.0367909931,
0.0657396913,
0.031812124,
0.0465232506,
-0.1026397124,
-0.0120282788,
-0.0064790333,
0.0059339264,
-0.0351940691,
0.0754646137,
0.0230925586,
-0.0038042832,
0.0045079263,
0.0002947921,
-0.0435465202,
-0.0184571836,
0.020524554,
-0.0815494061,
0.0057194829,
-0.0095727062,
0.0303329248,
-0.0158776063,
-0.0162853915,
0.000817798,
-0.0254420694,
0.0001675432,
-0.00147569,
0.0327113532,
-0.0155834928,
-0.0194466077,
-0.0559119768,
-0.0669915751,
0.0634629726,
-0.0217440277,
-0.0272765066,
0.0355125666,
0.0118265003,
0.0370207131,
-0.0346248299,
0.002044399,
0.013312649,
-0.0106857801,
0.0316620804,
0.0022911229,
0.0213840194,
-0.0535631366,
0.051434081,
0.0423128679,
-0.087332651,
0.0032594046,
0.037937358,
-0.0076208133,
0.0740970597,
-0.0015635673,
0.0405991524,
0.0281150751,
0.0077054636,
-0.0702152848,
0.0445683002,
0.039845258,
-0.0346999876,
-0.0297826454,
0.0294654761,
-0.0409089848,
0.0154011529,
-0.0299540032,
0.0091840252,
0.0028019617,
0.0468020961,
0.0347395688,
0.0419175178,
0.0229831226,
0.0368121862,
0.0532049127,
0.0254601408,
0.0520204306,
0.0002660949,
-0.007583871,
-0.0180806257,
-0.0269394219,
0.0021942034,
-0.0398399979,
-0.0450323224,
0.0640145689,
0.0188513175,
0.0278331805,
-0.0221078321,
-0.0245401766,
0.0425733328,
-0.0231080074,
0.0191343259,
0.0003234163,
0.0213103592,
-0.0294456948,
-0.0092983441,
0.0390641615,
-0.0267092232,
-0.0242461246,
0.0405453704,
-0.0135570075,
0.0556185916,
-0.00373044,
-0.0084893946,
0.0091666738,
0.041916389,
-0.012724014,
0.0224522464,
-0.0266146716,
0.0278659854,
0.0255491361,
0.0092432341,
-0.0043242839,
0.0082558459,
2.488e-7,
-0.0318256989,
-0.013830672,
0.0505955257,
0.0391359068,
0.0018671006,
0.0355234295,
0.0222893674,
0.0243480559,
-0.0099500213,
0.0584638119,
-0.0020605791,
-0.0481942594,
0.0619748682,
0.045561742,
-0.0612144433,
-0.0678158998,
-0.054239396,
-0.0160237774,
-0.0521235988,
-0.0156196561,
0.0499560721,
0.0517647825,
0.0228408091,
-0.0004425303,
0.0264013223,
-0.0227165688,
0.0171194747,
0.0154501442,
-0.026882818,
-0.0098174633,
-0.0755163878,
0.0040274789,
-0.0476818159,
0.008119368,
-0.0294375382,
0.036969997,
-0.0096835969,
0.0770518929,
-0.0063134516,
-0.0451925285,
-0.0163664967,
0.0225106012,
-0.0122789135,
0.0134696942,
0.054809615,
-0.0185404979,
-0.0224732477,
-0.0113790026,
-0.0572867356,
0.0512838773,
-0.0736975297,
0.0022844325,
-0.0267860312,
0.0231775343,
0.0321309678,
-0.0103602866,
-0.0452749021,
-0.0789633244,
-0.0085212374,
0.0184372179,
-0.0505447499,
-0.0020818443,
-0.0166852735,
-0.029975459,
-0.0260731261,
-0.0371475741,
-0.0259603281,
2.10786272e-34,
-0.0198758952,
0.0307095461,
-0.0319984518,
0.0042207409,
0.0201922692,
-0.0393184312,
0.0142441215,
-0.0077480236,
0.0304183941,
-0.0087584304,
-0.0177693795
] |
http://arxiv.org/abs/2401.15378v4 | http://arxiv.org/pdf/2401.15378v4 | A RAG-based Question Answering System Proposal for Understanding Islam: MufassirQAS LLM | 2024-01-27T10:50:11 | 2024-02-01T20:28:11 | Ahmet Yusuf Alan, Enis Karaarslan, Ömer Aydin | cs.CL | cs.CL, cs.AI | Challenges exist in learning and understanding religions, such as the
complexity and depth of religious doctrines and teachings. Chatbots as
question-answering systems can help in solving these challenges. LLM chatbots
use NLP techniques to establish connections between topics and accurately
respond to complex questions. These capabilities make it perfect for
enlightenment on religion as a question-answering chatbot. However, LLMs also
tend to generate false information, known as hallucination. Also, the chatbots'
responses can include content that insults personal religious beliefs,
interfaith conflicts, and controversial or sensitive topics. It must avoid such
cases without promoting hate speech or offending certain groups of people or
their beliefs. This study uses a vector database-based Retrieval Augmented
Generation (RAG) approach to enhance the accuracy and transparency of LLMs. Our
question-answering system is called "MufassirQAS". We created a database
consisting of several open-access books that include Turkish context. These
books contain Turkish translations and interpretations of Islam. This database
is utilized to answer religion-related questions and ensure our answers are
trustworthy. The relevant part of the dataset, which LLM also uses, is
presented along with the answer. We have put careful effort into creating
system prompts that give instructions to prevent harmful, offensive, or
disrespectful responses to respect people's values and provide reliable
results. The system answers and shares additional information, such as the page
number from the respective book and the articles referenced for obtaining the
information. MufassirQAS and ChatGPT are also tested with sensitive questions.
We got better performance with our system. Study and enhancements are still in
progress. Results and future works are given. | A RAG-based Question Answering System Proposal for Understanding Islam: MufassirQAS LLM: Challenges exist in learning and understanding religions, such as the
complexity and depth of religious doctrines and teachings. Chatbots as
question-answering systems can help in solving these challenges. LLM chatbots
use NLP techniques to establish connections between topics and accurately
respond to complex questions. These capabilities make it perfect for
enlightenment on religion as a question-answering chatbot. However, LLMs also
tend to generate false information, known as hallucination. Also, the chatbots'
responses can include content that insults personal religious beliefs,
interfaith conflicts, and controversial or sensitive topics. It must avoid such
cases without promoting hate speech or offending certain groups of people or
their beliefs. This study uses a vector database-based Retrieval Augmented
Generation (RAG) approach to enhance the accuracy and transparency of LLMs. Our
question-answering system is called "MufassirQAS". We created a database
consisting of several open-access books that include Turkish context. These
books contain Turkish translations and interpretations of Islam. This database
is utilized to answer religion-related questions and ensure our answers are
trustworthy. The relevant part of the dataset, which LLM also uses, is
presented along with the answer. We have put careful effort into creating
system prompts that give instructions to prevent harmful, offensive, or
disrespectful responses to respect people's values and provide reliable
results. The system answers and shares additional information, such as the page
number from the respective book and the articles referenced for obtaining the
information. MufassirQAS and ChatGPT are also tested with sensitive questions.
We got better performance with our system. Study and enhancements are still in
progress. Results and future works are given. | [
0.1016471162,
0.0061431783,
-0.045239836,
0.0208766479,
-0.010465689,
0.0094397971,
0.0195490681,
0.0107761091,
0.0264716744,
-0.036069084,
0.007602124,
-0.0163646154,
-0.0042087403,
-0.0378273688,
-0.0041652564,
-0.0351768024,
0.0542386919,
0.0153177259,
0.0043349722,
-0.0049410327,
0.0605112575,
0.0456246287,
-0.0308482181,
0.0313391723,
-0.027292585,
-0.0227886606,
-0.037089929,
-0.0075388588,
0.0293250438,
0.0263120867,
-0.0145586431,
0.0133044664,
0.0006679551,
0.0391961485,
0.000002046,
-0.0565356389,
-0.0384630896,
0.0008882587,
-0.0333958827,
-0.0588861518,
0.0439293049,
0.0360086411,
0.0006470656,
0.0074321269,
-0.044027172,
0.0178975817,
0.0271175895,
0.081685245,
0.0793393925,
0.0482487641,
0.0125716897,
-0.0750504956,
0.012380084,
0.0458309464,
0.0677529499,
-0.0087679634,
0.0072559104,
0.0359814055,
0.0211931393,
0.0416918173,
-0.0140416957,
0.0322277248,
0.0180699341,
-0.0262154453,
-0.0114982035,
-0.0273898728,
-0.0502963662,
-0.0492907017,
-0.0006592852,
0.0266835522,
0.046247445,
-0.0270137582,
0.0388790928,
0.0242044199,
0.0166748632,
0.0447344705,
-0.0300943088,
0.0129181296,
-0.084823288,
-0.0154709704,
-0.0036124855,
-0.012300605,
-0.0239957701,
0.0253687147,
0.0039715404,
-0.0005521123,
0.0171931479,
-0.02748904,
-0.0250887573,
-0.0174051039,
-0.0053275409,
-0.027408056,
0.0196251124,
0.057816457,
-0.0159958638,
0.0141550321,
-0.024790816,
-0.0367072262,
0.0257473923,
-0.0874761045,
-0.0158748422,
0.0704716444,
-0.0069454121,
0.0128534408,
-0.0216730405,
0.0392082445,
0.0202463809,
0.0171183571,
-0.0392542332,
0.0013041294,
-0.0440281592,
0.0039553428,
-0.0559752285,
0.0607963651,
-0.0216999687,
0.0503071584,
-0.0573365614,
0.0234369077,
0.0139259677,
-0.0100066466,
-0.0422082022,
0.0201708339,
-0.0138599928,
0.0046473001,
-0.0483703762,
-0.0217614975,
-0.0710570812,
-0.0293750223,
-0.0256628115,
-0.1180091202,
0.0334980488,
0.0259907637,
-0.0176235605,
-0.0039908285,
0.018602239,
0.0771748796,
-0.0024424656,
0.0063781254,
-0.0552462861,
-0.038533695,
-0.0172780808,
-0.0364682823,
0.0146323154,
-0.0533104204,
0.0374446064,
0.0312882774,
-0.0539076068,
0.0306118224,
0.029023312,
0.0778665915,
-0.0176317096,
0.0266908724,
-0.0148003157,
0.0137102632,
0.1263110936,
0.009344331,
0.0258014817,
0.008283929,
-0.0179311968,
0.034741044,
0.0590266734,
0.0191941485,
0.0192116741,
-0.0306564067,
0.0411005281,
-0.0429635085,
0.0002091627,
0.0475790389,
-0.0222081821,
0.0150206918,
-0.0456777252,
0.0482267924,
-0.0411818922,
-0.0265185703,
-0.0263754558,
0.031546317,
0.0011359659,
0.07429263,
0.005193613,
0.0399253145,
0.0680120587,
0.0129775163,
0.0044912724,
0.0457238741,
-0.0401887596,
-0.0237914082,
0.0756869167,
0.0359989889,
-0.0148845697,
-0.0590691119,
-0.0017373794,
0.0098562371,
0.0343803614,
0.0151933152,
0.0312572233,
0.0255389363,
0.0071028038,
0.008185775,
0.0434053987,
-0.0349739194,
-0.0223018602,
-0.0544174686,
0.0431908704,
0.0438194126,
0.0668338388,
-0.0218434501,
-0.0130767692,
-0.0175959896,
0.0362959094,
0.0497233234,
0.0226456728,
0.0053634024,
-0.0411608629,
-0.033186309,
-0.0066517028,
0.0260141063,
-0.0284951627,
-0.0055972063,
0.0247193333,
0.0168967173,
0.027547406,
0.02151029,
0.00703634,
0.0444318764,
0.0286828335,
-0.0250349566,
-0.0462813973,
0.0069912351,
0.0281917714,
0.0383006297,
-0.0188843943,
0.0048916689,
0.0265225284,
-0.0033777968,
-0.0582940541,
0.0046997005,
-0.0165784545,
0.0242407285,
0.0428112634,
-0.0233990047,
-0.0061012502,
0.0422320664,
0.0038486943,
-0.0075816251,
0.0314529799,
-0.070115149,
-0.0184839871,
-0.0447196811,
-0.0154297175,
-0.0162869003,
0.0037719971,
-0.0503963977,
0.0764037892,
0.0126576442,
0.0469368435,
0.0268153213,
-0.0476839989,
0.0528562553,
0.0001970493,
-0.0032833365,
-0.009961497,
-0.022768775,
0.0279500745,
0.061034888,
-0.0334543586,
-0.0148205608,
0.0183598716,
-0.0677201375,
0.0027858904,
0.0198716521,
0.0119898291,
-0.0090354113,
0.0133580239,
-0.0144415647,
-0.0293187164,
-0.0622093268,
0.0130253993,
0.0459111035,
0.0182561651,
0.0420397073,
-0.0125374133,
-0.0088413116,
-0.0334862173,
-0.0054173996,
0.0044364617,
0.0247911755,
0.050464876,
-0.014239003,
-0.0098281335,
-0.0562032051,
-0.0848408118,
0.0109365406,
0.0090984181,
-0.0418538526,
0.013756969,
-0.0221393034,
-0.051740665,
0.0143400459,
0.1034898087,
0.0188350901,
-0.026692789,
0.0117990617,
-0.064806737,
0.0102162631,
0.0037734762,
-0.0362212285,
0.0245635416,
-0.0344970413,
-0.0353123695,
-0.0439892188,
0.0296363179,
0.0190417171,
-0.0553764738,
0.0676757172,
-0.0117501365,
-0.0333577618,
-0.0401807353,
-0.0300779007,
0.0073119048,
0.0587908737,
-0.0075334888,
-0.0383302346,
0.0454211645,
0.0174796339,
0.0090139359,
-0.0557119362,
0.0220467076,
-0.0323798619,
-0.0266877655,
-0.0349862128,
0.0662426427,
-0.0410932004,
-0.0044763018,
-0.0397044048,
-0.0228890497,
0.0090278452,
-0.0720251054,
-0.0074090753,
0.0432588793,
0.0210939441,
-0.048798386,
0.0236591473,
-0.0955862477,
-0.0241638552,
-0.0025644277,
-0.060638763,
0.0082693622,
0.043660257,
-0.0215352401,
-0.0056606675,
0.0552094094,
-0.0747554302,
-0.0133356191,
-0.0374050736,
0.0268175099,
-0.0170010813,
0.0168567952,
0.0244756211,
0.0322485194,
0.0192522388,
-0.0070521482,
0.0776697248,
-0.0207610168,
-0.015340873,
-0.0120176943,
-0.0615244284,
0.0061077848,
-0.0230959654,
-0.0134324748,
-0.0390114263,
-0.0097243758,
-0.0479975343,
0.020976821,
0.0801707134,
0.0421978682,
0.0326674059,
-0.0182024147,
-0.0675467029,
-0.0439522155,
-0.0238117073,
0.0718592405,
-0.1439391077,
0.0153554697,
-0.0213178564,
0.0126894386,
-0.046050746,
-0.0170840193,
0.0314150974,
-0.0562738478,
-0.0214124378,
-0.0184657332,
-0.0303600673,
-0.0632792935,
-0.0160416625,
-0.0215494558,
0.0270693507,
0.0036724787,
0.0424737856,
0.0003960882,
0.0400713049,
0.0088920286,
0.0159595907,
0.0391356796,
-0.0076046572,
-0.004782855,
0.0042293132,
0.0673300251,
0.0622273907,
-0.0182950422,
-0.0316483192,
-0.0142682744,
-0.0061149704,
0.0095765227,
-0.0211049486,
0.0325241908,
-0.0030247916,
-0.0115285926,
-0.0844475478,
-0.0331991836,
0.0156151606,
-0.0064200358,
0.0358576141,
0.0084707644,
-0.0313293003,
0.0373436809,
-0.054359071,
0.0623455048,
-0.0181826688,
0.0529212877,
-0.0363726653,
-0.0188243724,
-0.0080287764,
-0.0402908549,
0.0047787568,
-0.0000849585,
-0.0634608418,
-0.0052826633,
0.0116212657,
-0.0059780832,
-0.1052416712,
-0.0215464588,
-0.046058666,
0.0292661004,
0.0142412093,
-0.0053726383,
-0.0196485464,
0.0546446145,
-0.0360269509,
-0.0541730374,
-0.0209706537,
-0.0163666401,
0.0121772857,
0.0036957378,
0.0149171837,
-0.0680450648,
-0.0097521041,
0.0243534409,
0.0010329436,
-0.0295231584,
0.0362669528,
0.0363798738,
0.0189431161,
-0.0113674486,
0.0071775382,
-0.010185089,
0.0186454114,
0.0371245667,
-0.0320964865,
0.0158659369,
0.0313173309,
0.0821954384,
-0.0167229008,
0.0089749498,
-0.0339678302,
0.0146946255,
0.0005183039,
-0.0375475213,
-0.008684678,
-0.041855026,
0.0192959569,
0.0078173708,
-0.0726027563,
0.0334998295,
-0.040915262,
0.0005633419,
0.0285491608,
-0.0525489226,
-0.0079410849,
-0.0029521214,
0.0282184631,
-0.0352373458,
-0.0059572924,
0.0332380421,
-0.0449505225,
-0.0363936387,
0.0173756685,
-0.0217543188,
0.0016891781,
0.0229120478,
0.0371166617,
-0.0643299371,
0.0160565842,
-0.0204937402,
0.0146306977,
0.0153154489,
0.0201520603,
-0.0225065425,
0.070559904,
-0.0013633128,
-0.017851783,
0.0525039956,
-0.0010497018,
-0.0366046764,
-0.0869589075,
0.0042966977,
0.0470022224,
-0.0031529048,
0.0295486748,
-0.0031542205,
0.0346120894,
0.0019386031,
-0.0055938051,
0.0215181019,
0.0515393093,
-0.0425893776,
-0.0020706784,
0.0118259015,
-0.0235144142,
-0.026060177,
0.0232059285,
-0.0617317632,
0.0116835022,
0.0164485518,
-0.0438392349,
-0.0409587994,
-0.0058769044,
-0.0227052942,
0.0449489132,
-0.0021406941,
0.0418811254,
-0.0426568575,
0.0021977508,
0.002936959,
0.0188381672,
-0.0120794866,
-0.0235084072,
0.0566648804,
0.0777211115,
0.0437943302,
-0.0401883349,
-0.0718951002,
0.0804450214,
-0.0915490165,
-0.03045444,
-0.0034485806,
-5.87451061e-33,
-0.0409497768,
-0.0250302181,
0.0097866906,
-0.0034834633,
0.0078282505,
-0.0601206981,
0.0092124641,
0.0177038498,
-0.0742550641,
0.0039145104,
-0.0288440157,
0.0165690519,
0.0141244708,
-0.0170796812,
-0.0040524742,
-0.0617978461,
0.0349045806,
-0.0041191042,
-0.0147915157,
0.0392836221,
0.0634035543,
0.0229481645,
0.0359238721,
-0.0886932909,
0.0159264002,
-0.0633394271,
-0.0072383541,
-0.0003009855,
0.0155841438,
0.0636169314,
-0.0125342729,
0.0669678524,
0.0180667546,
-0.0201805737,
0.0100459727,
0.0631212592,
-0.042022679,
-0.0315483063,
0.0372272059,
0.0216894597,
-0.041407153,
-0.0154066039,
0.0178774372,
-0.0373108163,
-0.0002303225,
0.0542492419,
0.0505968034,
-0.0229551401,
0.0072545023,
-0.0170280505,
-0.0824933052,
0.0116384495,
0.002487292,
-0.0122465575,
-0.0188242532,
0.0648172051,
0.0265492443,
-0.0009808518,
-0.0389756039,
-0.0092337606,
0.0004300608,
0.0032346258,
-0.0114573101,
-0.0125125125,
0.0063158842,
0.0171088967,
0.0196401384,
0.0030642997,
-0.0002121813,
0.0067064911,
-0.0284551382,
-0.0072952863,
0.0609756261,
0.0847649872,
0.0116686327,
-0.014378069,
-0.0718427896,
0.0308220275,
0.0416627601,
0.0473777764,
0.0140541159,
-0.0136125321,
-0.0581199601,
0.0554012023,
0.0132888909,
0.0183351878,
-0.0270167887,
-0.0630673468,
-0.0064792484,
0.0085494583,
-0.0338210277,
-0.032868091,
-0.0029372803,
-0.0376834273,
0.0370567329,
0.0301796179,
-0.0015586483,
-0.0512079895,
-0.0091374898,
0.00839788,
-0.0131898168,
0.0121480683,
0.0408907607,
-0.0452741086,
0.0075606206,
0.0036435379,
-0.0200395584,
-0.0235520098,
-0.0465625562,
0.0101471189,
-0.0079189893,
0.0620316565,
-0.0232199542,
0.0137293003,
0.0130922981,
0.0309306551,
-0.0130350292,
0.006276486,
0.0317146592,
-0.0209625829,
-0.0407352895,
-0.0330294035,
0.0032019757,
-0.0407115072,
-0.0153140826,
0.033916153,
-0.0109671727,
0.0494635738,
-0.0172503442,
-0.0835306346,
-0.004679109,
0.0021179479,
2.858e-7,
0.0362274535,
-0.0264381841,
0.0420579948,
0.0362698585,
-0.0180772897,
0.0048224335,
0.0821841434,
0.0186857507,
0.0132192699,
-0.0263833348,
-0.028673837,
0.0180469435,
0.0489023738,
0.0319117308,
-0.0672143176,
-0.0707040206,
-0.0448594168,
-0.0081247529,
-0.0407508723,
0.0090953382,
0.0782391801,
0.062286403,
-0.0139845666,
-0.009568559,
-0.0456470326,
0.0005630486,
0.0152101712,
0.0291115455,
-0.0190814622,
0.0349042453,
0.0751293302,
0.0189862233,
0.0452533625,
0.0035240171,
0.0066534188,
0.0105619123,
0.0348917618,
0.0124560166,
0.0274174642,
-0.037608698,
0.0141081754,
0.0017740136,
0.0164266322,
-0.0519580282,
0.0696136355,
-0.0026546428,
0.0093831541,
-0.0305018723,
-0.0307662748,
0.0090406341,
0.0164584462,
0.0004962757,
-0.0456717759,
0.0213118456,
0.0216775872,
0.0021185831,
0.0197330453,
-0.0355597213,
0.0255615469,
-0.0302568339,
-0.0258976799,
-0.0309136882,
-0.0135012288,
-0.0219194442,
-0.0073947562,
0.0141067971,
-0.0184446219,
2.185608815e-34,
0.0264280923,
-0.0051111244,
0.0047101132,
0.0309973098,
0.035706386,
0.0303149782,
-0.0096040154,
0.0158238802,
0.0366230607,
-0.0003048175,
-0.0336432345
] |
http://arxiv.org/abs/2401.15371v2 | http://arxiv.org/pdf/2401.15371v2 | LegalDuet: Learning Effective Representations for Legal Judgment Prediction through a Dual-View Legal Clue Reasoning | 2024-01-27T10:28:27 | 2024-02-28T04:38:25 | Pengjie Liu, Zhenghao Liu, Xiaoyuan Yi, Liner Yang, Shuo Wang, Yu Gu, Ge Yu, Xing Xie, Shuang-hua Yang | cs.CL | cs.CL | Most existing Legal Judgment Prediction (LJP) models focus on discovering the
legal triggers in the criminal fact description. However, in real-world
scenarios, a professional judge not only needs to assimilate the law case
experience that thrives on past sentenced legal judgments but also depends on
the professional legal grounded reasoning that learned from professional legal
knowledge. In this paper, we propose a LegalDuet model, which pretrains
language models to learn a tailored embedding space for making legal judgments.
It proposes a dual-view legal clue reasoning mechanism, which derives from two
reasoning chains of judges: 1) Law Case Reasoning, which makes legal judgments
according to the judgment experiences learned from analogy/confusing legal
cases; 2) Legal Ground Reasoning, which lies in matching the legal clues
between criminal cases and legal decisions. Our experiments show that LegalDuet
achieves state-of-the-art performance on the CAIL2018 dataset and outperforms
baselines with about 4% improvements on average. Our dual-view reasoning based
pretraining can capture critical legal clues to learn a tailored embedding
space to distinguish criminal cases. It reduces LegalDuet's uncertainty during
prediction and brings pretraining advances to the confusing/low frequent
charges. All codes are available at https://github.com/NEUIR/LegalDuet. | LegalDuet: Learning Effective Representations for Legal Judgment Prediction through a Dual-View Legal Clue Reasoning: Most existing Legal Judgment Prediction (LJP) models focus on discovering the
legal triggers in the criminal fact description. However, in real-world
scenarios, a professional judge not only needs to assimilate the law case
experience that thrives on past sentenced legal judgments but also depends on
the professional legal grounded reasoning that learned from professional legal
knowledge. In this paper, we propose a LegalDuet model, which pretrains
language models to learn a tailored embedding space for making legal judgments.
It proposes a dual-view legal clue reasoning mechanism, which derives from two
reasoning chains of judges: 1) Law Case Reasoning, which makes legal judgments
according to the judgment experiences learned from analogy/confusing legal
cases; 2) Legal Ground Reasoning, which lies in matching the legal clues
between criminal cases and legal decisions. Our experiments show that LegalDuet
achieves state-of-the-art performance on the CAIL2018 dataset and outperforms
baselines with about 4% improvements on average. Our dual-view reasoning based
pretraining can capture critical legal clues to learn a tailored embedding
space to distinguish criminal cases. It reduces LegalDuet's uncertainty during
prediction and brings pretraining advances to the confusing/low frequent
charges. All codes are available at https://github.com/NEUIR/LegalDuet. | [
0.0429026745,
0.0397636853,
-0.0029775866,
0.0162568316,
-0.0468748063,
0.0441050678,
0.019722281,
0.0137072941,
0.0205734018,
-0.0680969059,
-0.0318885446,
-0.0168816391,
-0.0575846434,
0.0019199626,
-0.0135114072,
-0.0503799841,
0.0570024215,
0.0204312578,
0.0781006739,
-0.0013966039,
0.0407112055,
0.0272301435,
0.0123414993,
0.0113790017,
-0.0226675887,
-0.0203185845,
-0.0240301266,
-0.0213948414,
0.0078982636,
-0.0309148971,
0.0092591336,
0.0936298072,
0.0570729785,
0.0139066912,
0.0000018737,
-0.044347588,
0.0157489497,
-0.0567852482,
-0.0241118371,
0.0035340216,
-0.0067430022,
-0.01457344,
0.0029071763,
0.0064527774,
-0.054340478,
-0.0127007626,
-0.0148007656,
0.0616347641,
0.0303371698,
0.0666399226,
0.0130040068,
-0.068125926,
0.0185333136,
-0.06186134,
0.0075978842,
-0.0539654009,
0.0264329053,
0.0292371009,
-0.0282307975,
-0.002555263,
0.0135067692,
-0.0265472047,
0.0559400953,
-0.0132673541,
-0.0161810182,
0.0381567553,
0.0019481491,
0.0008736448,
-0.0025969073,
0.0668056682,
0.0264766924,
-0.053258419,
-0.0377850123,
-0.0112647545,
0.0577302612,
0.0073187011,
-0.0096804537,
0.0700165108,
-0.0259971917,
0.0358162671,
0.0412730835,
0.0332165286,
0.0063587148,
0.0300147608,
-0.0278039109,
0.0639248118,
-0.0026582801,
-0.0411232226,
0.0262402594,
0.016967874,
0.0258403216,
-0.0377588049,
0.0231863707,
0.032484822,
-0.0045593432,
0.0290391464,
-0.0095768711,
0.0155628268,
0.0369093306,
-0.0635869578,
0.0122465873,
0.0222491659,
-0.0367918983,
0.0803749189,
-0.0766611174,
-0.0100358482,
0.0376683585,
0.0383539721,
-0.0208790991,
-0.030290328,
-0.0300841611,
-0.0117860595,
-0.0370489955,
0.0387679935,
-0.0625820383,
0.0039607231,
-0.0053232717,
-0.0099913534,
-0.0613816939,
0.0203865767,
0.0221949145,
0.0091498652,
-0.0097818682,
-0.0228361282,
-0.0594715811,
-0.0457637981,
-0.0641322583,
0.045289617,
0.0007385024,
0.013386921,
-0.0261882767,
-0.0239507612,
-0.0213482454,
-0.0123750633,
-0.0002921504,
0.0820852965,
-0.0159211978,
0.0051620463,
-0.0590817295,
-0.0599398278,
-0.0021430287,
-0.0583468862,
-0.0416333973,
-0.059306629,
0.0056570806,
-0.0301189218,
0.0044577061,
-0.0644430667,
0.0232367646,
0.0337323993,
-0.0286072716,
0.1132107601,
-0.0018544955,
-0.040929269,
-0.0555904396,
0.0551731363,
0.0234080087,
0.0287471619,
-0.0369142182,
0.0328244939,
0.0002823537,
-0.0137059418,
0.0107530486,
-0.0186598711,
0.0285105314,
0.0112652062,
-0.0174966361,
0.0217394587,
0.0101399347,
0.0318206064,
0.0360130705,
-0.003627541,
-0.0113464491,
-0.0141738774,
-0.0028006679,
0.0271544084,
0.0387125388,
0.0578828789,
0.0386168323,
-0.0399948508,
0.0196774993,
-0.002932393,
-0.0889365673,
0.0405556783,
-0.0282899365,
0.0003043346,
-0.0815894529,
0.0131390234,
-0.0499760844,
-0.00262328,
0.0136192925,
-0.0009024526,
0.0045931111,
-0.0128362868,
-0.0016604649,
0.0170427896,
-0.0327987038,
0.0405576676,
0.0624767318,
-0.0487484932,
-0.0331516899,
-0.0353087708,
0.0202040561,
0.0500480197,
0.026150506,
-0.0117931347,
-0.0464193746,
-0.0138124432,
0.0046411888,
0.0178203154,
0.0388845839,
-0.0177866574,
-0.0140025206,
-0.0510898642,
-0.0561172292,
-0.0027059615,
0.0351388343,
0.0329717025,
0.0048772376,
0.025233658,
0.0580825806,
-0.057374604,
0.0495267063,
0.0457540266,
0.0001634015,
-0.0380908586,
-0.0246861577,
0.0131205767,
0.0254785623,
0.0079902727,
0.0283536073,
-0.0204611532,
0.017688483,
0.0184346661,
0.0202813093,
0.041638989,
-0.0445165336,
-0.0348634347,
0.0590143017,
-0.0273827091,
0.005115808,
-0.0158593021,
-0.0501222983,
-0.001834065,
-0.0179190151,
0.0382629335,
0.035535302,
-0.0694317892,
-0.0284535121,
0.0127215991,
-0.0317557529,
-0.0147645539,
0.0682988837,
-0.009853133,
-0.0379809886,
-0.014366609,
0.0162066929,
0.046797093,
0.0287903585,
-0.0045987866,
-0.0424912572,
-0.0019502514,
0.0031355938,
-0.0025930975,
0.0244321395,
-0.0597328767,
-0.0194375534,
-0.058227431,
-0.000733135,
-0.0401632637,
-0.025021838,
0.0126521802,
-0.041351229,
0.0068271342,
-0.0402933285,
-0.0268940758,
-0.0050171232,
-0.0256907921,
0.0017559808,
0.0127194487,
-0.0325395539,
0.0155071039,
-0.0040710527,
-0.0096728746,
-0.0256631766,
0.0305606611,
-0.0044150744,
0.0086173732,
-0.0390034057,
0.0782991201,
-0.0148106618,
-0.0172463246,
-0.0307605695,
-0.0037281755,
-0.0029673215,
0.0196421966,
-0.0150751751,
0.0308362655,
0.0410359837,
0.0258215666,
0.0374225639,
-0.0031835409,
-0.0705815777,
0.0058189067,
0.0005281026,
0.0439732336,
-0.0186039526,
0.0067447685,
-0.0478720888,
0.033121679,
-0.0062864935,
-0.011082733,
-0.0486099534,
0.0179052334,
0.0185978953,
-0.0140171414,
-0.046113275,
-0.0153878229,
0.0039549679,
0.0254731849,
-0.0094095375,
-0.0730894357,
0.0813765973,
0.0364123397,
0.0577442683,
-0.0458690412,
0.009324139,
-0.0303869061,
0.0790774375,
0.0142978858,
-0.0097772051,
-0.0449279919,
-0.0517403781,
0.0069209551,
-0.0084116217,
0.0572677329,
-0.0272982977,
0.0806141719,
-0.01827766,
0.0027091145,
-0.0110715944,
-0.0426094942,
-0.0617705435,
-0.0433612391,
-0.0355363861,
0.0144545622,
-0.0181435682,
0.0219854787,
-0.0034089419,
-0.0127579784,
-0.0445348173,
-0.0030605174,
0.0166682657,
-0.0163482726,
-0.0388613306,
-0.004699091,
-0.0029744811,
-0.02023964,
0.0111532798,
0.0028099981,
0.022490982,
0.0353454128,
0.0018744955,
-0.0332388729,
-0.0053838617,
0.0485797375,
0.0145329023,
-0.0454076082,
-0.0116710719,
0.0353059992,
0.0030367626,
-0.0366042815,
0.0188767146,
0.1014119089,
0.0356843658,
-0.0413772129,
0.0000607237,
0.0134678716,
0.0186953787,
0.0494441316,
0.0004935169,
-0.0840719342,
-0.0271917917,
0.0149755822,
-0.0189747028,
0.003618465,
-0.032729581,
-0.0105500836,
-0.0711070746,
0.0674815103,
0.0291217454,
0.023881048,
-0.0337303579,
-0.0326024964,
0.0005370335,
-0.0190595686,
-0.0162118729,
0.0163457599,
-0.0206015408,
0.0118707633,
0.0033571492,
-0.0397050343,
0.0073404126,
0.0585950315,
-0.0541534163,
0.0621821359,
0.007984655,
0.0305673424,
-0.0092406105,
-0.0348703973,
0.0013447006,
-0.0131201027,
0.025067715,
-0.0173777267,
0.0305710174,
-0.0082188183,
0.0194501989,
-0.0460226648,
0.0616948791,
-0.0150280138,
-0.0537198037,
0.0386374816,
0.0689170063,
-0.0449443907,
-0.0261656232,
-0.0102310292,
0.014802618,
-0.0084869685,
0.0822726041,
0.0165317487,
0.0378149971,
-0.0363119654,
-0.0153212287,
-0.0087582581,
-0.0284798797,
0.004144371,
-0.0125687709,
0.0032984461,
0.0180383604,
0.0107085081,
-0.0403545573,
0.0266041663,
0.0397600494,
0.0204890464,
0.0095920227,
-0.0235381462,
-0.0220178831,
-0.0626891851,
0.052444078,
-0.0405441821,
0.0069331266,
0.0176234469,
0.0146051217,
0.0068368865,
-0.0157120861,
-0.0135357482,
-0.0211429074,
0.0110056289,
0.0101848319,
0.0105720554,
0.0157560036,
0.0181883108,
0.0488649309,
-0.0499006063,
-0.0151389819,
0.0419580676,
0.0109827453,
-0.0033980589,
-0.0121662728,
0.0819825456,
-0.0400870666,
0.0152599998,
0.0198400989,
0.0066544474,
0.0236855987,
0.0426015519,
-0.0061459234,
-0.049800355,
-0.0151170446,
0.0493412875,
0.0243229512,
0.0595047809,
0.0134841697,
-0.0056464286,
-0.0558126234,
-0.0030992394,
-0.0227142815,
-0.0118504157,
-0.0112183085,
0.0015665122,
0.0372660421,
-0.0147951264,
-0.0366607308,
0.0287626255,
-0.0108069936,
-0.0035674609,
-0.0442506075,
0.0131085562,
0.0753604621,
0.0132545885,
0.0017787375,
-0.0365948118,
-0.0112301409,
-0.006081244,
-0.0280208848,
0.0424009003,
-0.0193018392,
0.068998307,
0.0134310499,
0.0019738192,
-0.0446326211,
0.0062920363,
-0.05647346,
0.0132110333,
-0.0110187298,
0.0483475998,
0.0619120002,
-0.0429512039,
-0.014536093,
-0.0255873669,
-0.0039364886,
0.0328901522,
0.0443784408,
-0.0069016181,
-0.0516838841,
0.0560721904,
0.0227565523,
-0.0187516585,
-0.0035802678,
0.0257012378,
-0.0434872173,
0.0415667556,
-0.0070225899,
-0.0173972491,
-0.0283007659,
-0.0045458805,
0.0145130036,
0.0644076616,
-0.0175206549,
-0.0138785169,
-0.0962340012,
-0.0415941849,
-0.0344626978,
0.0219713803,
-0.0260438286,
0.03652592,
-0.0067720795,
0.0778678507,
0.0202746671,
-0.0232368633,
-0.0210098512,
0.0208669007,
0.027757531,
-0.0173220951,
-0.0361344665,
-5.524132479e-33,
-0.0364823453,
-0.0080382721,
0.0125523079,
0.0068658856,
-0.0297735278,
-0.0144651942,
0.0113528958,
-0.0027801122,
-0.0481355712,
0.0495558679,
-0.0106423181,
0.0284736827,
0.0101259062,
-0.0375293344,
0.0160853993,
-0.0336749069,
0.0023838924,
0.0142291402,
0.0094323857,
0.0226490255,
-0.0106264995,
0.0276254099,
0.0431907997,
-0.0953002647,
0.0338722654,
0.0506779067,
-0.0207809061,
-0.0074381889,
-0.0691207871,
0.0309390165,
-0.0001642332,
0.0088235047,
0.0258651823,
-0.057069052,
-0.0096083498,
0.0326078422,
-0.1091069877,
-0.0317500904,
0.028013628,
-0.0188750792,
0.0260104053,
-0.0384701937,
-0.0166674722,
-0.0074977898,
0.0267420057,
0.0037615455,
0.0207125638,
-0.0451355949,
-0.0322479084,
-0.0266265478,
-0.0585854724,
0.0336816609,
-0.0169056859,
-0.0030554277,
0.0082556419,
-0.0273980871,
0.0246809367,
0.0300406255,
-0.086685881,
0.0138136428,
-0.0233813673,
-0.0344274081,
0.0200729668,
-0.0557229072,
-0.0507275686,
0.0860404,
-0.0108236521,
-0.0267213825,
-0.0725069121,
-0.0123210279,
0.0261805765,
0.0949735343,
0.0427465066,
0.0438355953,
0.0639938414,
-0.0633007437,
-0.0789633393,
-0.0208569709,
0.0756841004,
0.0331400447,
0.0252436772,
-0.0104687,
-0.0627645999,
-0.0186994933,
-0.0396718644,
-0.0300952401,
-0.0208086614,
-0.0198471285,
0.0133488718,
0.0145059694,
0.0405152887,
-0.0233836975,
-0.0301303957,
-0.0389000177,
0.037693359,
0.0609650351,
-0.0133683514,
-0.0336961597,
-0.0164155215,
0.0563247316,
-0.0478006527,
-0.0465441681,
0.0342285745,
-0.0322181433,
-0.0077907764,
-0.010264839,
0.1034307927,
-0.0175528917,
-0.0425978638,
-0.0145302759,
-0.0135827074,
-0.0023103154,
0.0252980273,
0.0233860146,
0.017340675,
0.0524312742,
0.0191795807,
0.0036243149,
0.0282279439,
-0.0109550152,
-0.0137760714,
-0.0181899685,
0.0140366247,
0.0342961214,
0.0263701472,
-0.0144228842,
-0.011235767,
-0.0145399403,
0.0264446642,
-0.0143095795,
-0.0041751186,
0.0784536973,
2.574e-7,
0.0595499501,
0.0113196457,
-0.0270002093,
0.0057732258,
-0.019269323,
0.0515803471,
0.0006389599,
0.0440665111,
0.056119781,
-0.0351215415,
0.0083797509,
0.0547182746,
0.0313789248,
0.0559226386,
-0.0320454054,
-0.0480857305,
-0.0401505791,
-0.0616724007,
-0.0459193364,
0.017486332,
0.0529066138,
0.1018365994,
0.0546085127,
-0.0095694605,
0.0393308215,
-0.0113792811,
0.0230937805,
-0.0185550191,
-0.0042438535,
-0.0233545061,
0.0591880716,
-0.0347881429,
0.0045266221,
0.0767759308,
0.0124437353,
0.0490408167,
-0.0285793096,
0.0156989712,
-0.0118589764,
0.0271342192,
-0.0457327701,
-0.041915793,
0.0084284069,
-0.0400558598,
0.0822921693,
-0.0145128397,
-0.0211675391,
-0.0119621698,
-0.0363026708,
-0.0191928297,
0.0197025035,
0.0320334695,
-0.0259246379,
-0.0310894474,
0.0322621092,
-0.0105697736,
0.0029134338,
-0.1168233529,
-0.0178402793,
0.0432988368,
-0.0101200035,
0.0032943985,
0.0203923229,
-0.0452029556,
0.0939533114,
-0.1057978049,
0.034090519,
2.111069386e-34,
0.0364307873,
0.0034853283,
0.0160248782,
0.0198921841,
0.029152412,
0.0173124429,
-0.0036741763,
-0.0181519147,
0.0145048257,
-0.0565729812,
0.0216121264
] |
http://arxiv.org/abs/2401.15360v1 | http://arxiv.org/pdf/2401.15360v1 | Importance-Aware Data Augmentation for Document-Level Neural Machine Translation | 2024-01-27T09:27:47 | 2024-01-27T09:27:47 | Minghao Wu, Yufei Wang, George Foster, Lizhen Qu, Gholamreza Haffari | cs.CL | cs.CL | Document-level neural machine translation (DocNMT) aims to generate
translations that are both coherent and cohesive, in contrast to its
sentence-level counterpart. However, due to its longer input length and limited
availability of training data, DocNMT often faces the challenge of data
sparsity. To overcome this issue, we propose a novel Importance-Aware Data
Augmentation (IADA) algorithm for DocNMT that augments the training data based
on token importance information estimated by the norm of hidden states and
training gradients. We conduct comprehensive experiments on three widely-used
DocNMT benchmarks. Our empirical results show that our proposed IADA
outperforms strong DocNMT baselines as well as several data augmentation
approaches, with statistical significance on both sentence-level and
document-level BLEU. | Importance-Aware Data Augmentation for Document-Level Neural Machine Translation: Document-level neural machine translation (DocNMT) aims to generate
translations that are both coherent and cohesive, in contrast to its
sentence-level counterpart. However, due to its longer input length and limited
availability of training data, DocNMT often faces the challenge of data
sparsity. To overcome this issue, we propose a novel Importance-Aware Data
Augmentation (IADA) algorithm for DocNMT that augments the training data based
on token importance information estimated by the norm of hidden states and
training gradients. We conduct comprehensive experiments on three widely-used
DocNMT benchmarks. Our empirical results show that our proposed IADA
outperforms strong DocNMT baselines as well as several data augmentation
approaches, with statistical significance on both sentence-level and
document-level BLEU. | [
0.0137103247,
0.074249737,
0.0238448307,
0.0227963235,
-0.0119366907,
0.0042841909,
0.023632573,
0.0037402755,
-0.0221371949,
-0.0444855466,
-0.0243565924,
-0.0127570583,
0.0312466286,
0.0047990568,
0.0473475344,
-0.0579784848,
0.0470390469,
0.0036923625,
-0.0062472187,
-0.0073999525,
-0.0024502554,
-0.0142751308,
0.0345770828,
0.0303883255,
-0.0379529819,
0.0070476891,
0.0156262647,
0.0521752201,
0.0666893125,
-0.0204966683,
-0.0037782574,
0.003815138,
0.0357270129,
0.0415475592,
0.000002121,
-0.0338915177,
-0.0378638208,
0.0270188581,
0.0231082309,
0.0192728341,
0.0201083757,
-0.0616718009,
-0.0127808256,
0.0001338719,
-0.0082402686,
0.009073941,
0.0385536887,
0.0536802523,
0.0428169742,
0.0905465335,
-0.0158068798,
0.0029582488,
0.0468503647,
-0.0749538317,
0.0387375057,
-0.0561742075,
0.0021087481,
0.0025872316,
-0.0516991094,
0.0401606299,
-0.0243133772,
0.0305864438,
0.0312678441,
-0.0402276777,
0.0293831863,
0.017038282,
0.0061023976,
0.0427676104,
0.0111730061,
0.0979542881,
-0.0130425412,
0.0078209369,
-0.0276832711,
-0.0244440883,
0.0100199906,
-0.070432663,
-0.0556959435,
-0.0254771039,
0.0394451879,
0.0194082912,
0.0263484791,
0.0484953076,
0.0633962899,
0.0202874877,
-0.0359531753,
0.0323758423,
0.0148987854,
-0.0265196972,
-0.0458678715,
-0.0074290568,
-0.0584445074,
-0.0158575624,
0.060237132,
0.0116378982,
0.0207621418,
-0.0027095689,
-0.056151513,
-0.0958878323,
0.0192885436,
-0.0363391116,
0.0117642395,
0.0612983406,
-0.0177199952,
0.0594266243,
-0.0170197915,
0.0517160781,
-0.005315173,
-0.0012209303,
-0.0406614728,
0.0008122955,
0.0032678444,
0.0344718136,
-0.0048991041,
0.0364047326,
-0.0269894823,
-0.0313476101,
-0.0171842556,
-0.0146371508,
-0.0063750609,
0.0089755924,
0.0172721669,
-0.0041615386,
-0.0862183049,
-0.0351222195,
0.0082472721,
0.0299462676,
-0.0137091009,
0.0002817342,
0.0218613204,
0.0025233696,
-0.0259366874,
-0.0002228454,
0.0286617987,
-0.0200009812,
-0.0046261298,
0.0472585782,
-0.0223277565,
-0.0491387211,
0.0012309388,
0.0149106048,
-0.0013425363,
-0.0477404371,
-0.015085713,
-0.0371037908,
0.0517249517,
-0.024411235,
0.014696327,
-0.0360925756,
0.0055138171,
0.0089551359,
-0.0144777456,
0.0413564742,
-0.0251285546,
-0.0577981286,
0.0133276917,
-0.0012886155,
0.0027750656,
0.0224153008,
0.0422192812,
0.0669556856,
-0.0050134831,
0.0236388762,
-0.0293835588,
-0.0023623984,
0.0182601195,
0.0253784712,
-0.0047735292,
-0.0140989628,
-0.0251534693,
0.0343901291,
0.0425928161,
0.0157599077,
-0.056285996,
-0.0080233635,
0.0800329298,
0.0143796904,
-0.0067463266,
0.0060560065,
-0.00204399,
0.0278099868,
-0.0099593224,
-0.0064861211,
-0.0264206342,
0.0332810022,
0.0322941802,
0.004269707,
0.0714324787,
0.03017205,
0.0025915462,
-0.0545030162,
0.0102002313,
-0.0350911617,
0.0089289909,
-0.0131962523,
0.0372232497,
0.0220835283,
-0.0417041071,
0.0115311071,
-0.011719659,
-0.0478615761,
0.0159196779,
-0.017040085,
-0.0258325581,
0.056848038,
0.0759192556,
-0.0664914772,
-0.0553962849,
-0.0202561487,
0.028073851,
0.0370876193,
0.0281943008,
0.0333958827,
0.0540557094,
-0.0650731549,
-0.0210332014,
-0.0462400727,
0.0240392815,
0.0324721336,
-0.0285411775,
0.030566847,
0.0501472354,
-0.0548139624,
-0.0822862089,
0.0280699637,
0.0217828676,
-0.0328097939,
0.0678479373,
-0.0003047152,
0.0108252214,
0.0091422377,
-0.0018271187,
0.0061403178,
0.0102950893,
-0.0370356739,
-0.0282197855,
-0.0221820455,
-0.0565642975,
-0.0481969751,
0.0205009393,
-0.0486580208,
-0.0175048877,
0.0377792567,
-0.0837670043,
-0.0649742261,
0.0170444921,
-0.067789793,
-0.0132499272,
-0.0429674312,
0.0160673242,
0.0323695913,
0.0445909277,
-0.0301415846,
0.0335439555,
-0.0494711325,
-0.0290853977,
0.013006526,
0.0107969185,
0.0104117543,
-0.0098993285,
0.0020011927,
-0.007578786,
-0.0018394743,
0.0219411794,
0.0339985117,
0.0238276292,
0.0197197646,
-0.0310742799,
-0.0714262128,
-0.0120720239,
-0.028865004,
0.0583292916,
0.0448139831,
0.0841066539,
0.0079820184,
-0.0287877619,
-0.0204496346,
-0.0034961901,
-0.0388878882,
-0.0113696856,
0.0320366733,
0.0163478572,
0.0301999077,
-0.040017128,
-0.0138926841,
0.0203864332,
0.0176992938,
0.0046375971,
-0.0215964541,
-0.0273582954,
-0.0271125883,
-0.0181597304,
0.0215668119,
-0.0243982002,
-0.0310025103,
-0.004021313,
-0.0166565813,
0.0234436356,
0.0571389794,
0.008309843,
0.0147260241,
0.037281137,
0.0375826433,
-0.0676324144,
0.0152545786,
-0.0229253117,
0.0847021863,
-0.012434544,
-0.0507846661,
0.0343969427,
0.0367332362,
-0.0226526633,
-0.0161830448,
-0.0370258316,
0.0132713607,
0.0139007755,
-0.0353548229,
-0.0111267818,
-0.0031852177,
-0.0124700535,
-0.0128468769,
-0.0218030363,
-0.0558027737,
0.0570871904,
-0.0629063845,
-0.0014398203,
-0.0691267103,
0.0092643648,
-0.0627620369,
0.0246298965,
-0.0024293771,
0.0000249378,
-0.0719024241,
-0.0409425236,
0.027239915,
0.0107119437,
-0.0178908389,
-0.0010461183,
0.0087986048,
-0.0377748497,
0.0153473839,
-0.0305910502,
-0.0268783886,
0.006217631,
-0.0206177179,
0.0058906809,
0.0070678107,
0.0516867787,
-0.0452889428,
-0.0036478762,
-0.0290000476,
0.025096247,
0.1284353584,
-0.0497357547,
-0.0178616047,
-0.0381362103,
0.0154753644,
0.0092703365,
0.0188117027,
0.0655093193,
-0.0253576376,
-0.0096897082,
0.0276395027,
0.0150509933,
-0.0430906415,
-0.0399506763,
-0.0294953622,
0.040522363,
0.0050200722,
0.0112212077,
-0.0347984694,
0.0383787453,
-0.0812298581,
-0.0137800947,
0.0739749074,
0.0138134724,
-0.0788405016,
0.0052988161,
-0.0249933843,
0.0440734439,
0.0035432163,
0.0329960883,
-0.0592834689,
-0.0477907993,
0.0004585069,
0.0547614098,
0.0249738153,
-0.0000277936,
-0.0349789821,
-0.0083415518,
0.0216442458,
0.0212605,
-0.0091683939,
-0.0290303435,
-0.0204047617,
0.0163079984,
0.0044472008,
0.0359929875,
-0.024638908,
0.0139450282,
-0.0334287509,
0.0005026847,
-0.0021837621,
-0.0835353881,
-0.0417819917,
0.0069588772,
-0.0228057429,
0.0903632492,
0.043339394,
-0.0038358942,
-0.0089209834,
0.0818704665,
0.0322339796,
0.0150087168,
-0.0161287487,
0.024933368,
-0.0255827252,
0.0033987958,
-0.0027430085,
0.0570551865,
0.0040794951,
0.0294383764,
-0.0019347343,
0.0193257704,
-0.0476300158,
-0.0066263624,
-0.0100730192,
-0.0602389574,
0.0167782325,
0.0188351814,
0.0382957868,
-0.0469731018,
0.0089838505,
-0.0145482048,
0.0052451245,
-0.0593284592,
0.0690147355,
-0.0480109341,
-0.022786757,
-0.053565193,
-0.1131679714,
-0.0629757941,
-0.0117111541,
0.0229414161,
-0.0099810986,
0.001019513,
-0.0020012148,
-0.0628546253,
0.01044033,
0.0492729396,
-0.0155119225,
-0.0105095953,
-0.0293975677,
-0.0001936044,
0.0111908829,
-0.0389458314,
-0.0123959072,
-0.020165313,
-0.0662088469,
-0.0427043512,
0.0180876125,
0.0720806271,
0.0099538099,
0.0653649643,
-0.0404383726,
-0.0600565933,
0.0529269576,
0.0286493022,
-0.0209945105,
0.0111884689,
0.0412490107,
-0.0510695167,
0.0017407226,
0.0394673161,
-0.0223258119,
0.0102292122,
0.0282532815,
0.0371208601,
0.0048794127,
0.0116911959,
-0.0144888116,
-0.0050933962,
0.0128207011,
-0.0193150509,
0.010677767,
-0.0345554911,
0.0625477806,
-0.0293428805,
0.0087960362,
-0.0698701814,
-0.0007032967,
0.0298291873,
-0.0101135112,
-0.0057229176,
0.0309494808,
0.0328367352,
-0.0420480147,
-0.0167245753,
-0.0220642947,
-0.0352503657,
0.017164873,
0.0012337859,
0.0018854393,
-0.0315100625,
0.0371243022,
-0.0187823419,
0.0161388069,
-0.0377710648,
0.0093355449,
0.0248800088,
-0.0093959169,
-0.022665469,
0.0723253116,
-0.0715510473,
0.0204820354,
-0.0402610078,
0.0090957861,
0.0496460944,
0.0140757849,
0.1267002523,
-0.0285224915,
-0.003838551,
-0.0129132727,
0.022848662,
0.0320756212,
-0.0194451381,
-0.0378707089,
-0.0061748768,
0.0253692865,
0.0953333899,
0.0643386096,
-0.0081539536,
-0.0199828278,
-0.0174643658,
-0.0107330019,
-0.0153952511,
0.0133735249,
0.0386661962,
0.0697350353,
-0.0225811396,
-0.0014310417,
-0.0068675182,
-0.0145911165,
-0.0073270649,
0.0089447964,
-0.0152360033,
-0.005294146,
0.0210236963,
0.0234425794,
-0.0302143153,
-0.0071239304,
-0.0473716669,
0.0024842452,
0.002558504,
-0.0423045047,
-0.0007805417,
-6.783754389e-33,
-0.0095619503,
0.0077055194,
-0.0437083989,
-0.0310310069,
-0.0166529492,
0.0378675163,
-0.0592903569,
-0.0379710123,
0.0116424952,
0.0424028747,
-0.0280578826,
0.0246160999,
0.0171422847,
0.0520131104,
-0.0566903576,
0.0193836577,
0.0673016831,
0.0549177155,
-0.0444179289,
0.040909376,
0.0431012847,
0.0127733713,
0.1010645479,
-0.0691343918,
0.0407979488,
0.0208645202,
0.0004318712,
-0.0263220314,
0.0006528192,
-0.0238655899,
0.0051792962,
0.0090225032,
0.0233714432,
-0.0456776097,
-0.0621049963,
-0.0372281596,
-0.0689003766,
-0.0295033883,
-0.0151906414,
0.0149559071,
-0.0694665313,
-0.0418680497,
0.0676160902,
-0.0067615043,
0.0097396262,
-0.0283396151,
0.0110491253,
-0.0217704978,
-0.0520315878,
-0.0364000127,
-0.0569704957,
0.0451040901,
-0.0196873304,
0.025541693,
-0.0044577774,
0.0155153517,
-0.0030317162,
-0.004265334,
-0.0869348049,
-0.0262761321,
0.0212003589,
0.0548397042,
0.0602227561,
-0.0183795877,
-0.0212027542,
0.0216935109,
0.0273185819,
0.0276854839,
0.0096228467,
0.0037012352,
0.0461578704,
-0.0080167614,
0.0444417149,
-0.0051108645,
0.0565053895,
0.0080824699,
-0.0608713813,
0.0893929973,
0.0217025876,
0.0115621593,
0.0212632082,
0.0176396053,
-0.0267710499,
-0.0082528023,
-0.0209219363,
0.0174251832,
-0.0171390157,
-0.0119634643,
-0.0111918785,
0.0367104374,
-0.0648591518,
-0.0072808219,
0.0447183289,
-0.0245409012,
-0.0011293548,
0.0140500078,
0.0012951962,
0.0253647454,
0.0051301885,
-0.0229674634,
0.0001653718,
-0.0046167895,
0.0471343137,
-0.0002667766,
0.0286338851,
-0.0378570333,
0.0711843446,
0.0075994809,
-0.0244419053,
-0.0116725657,
-0.0278612822,
-0.0057529733,
-0.0074584363,
0.0199107341,
0.0205599256,
-0.0158587005,
-0.0040005539,
0.0796117112,
0.021794511,
0.0636807308,
-0.0076633645,
-0.0319863744,
0.0062930612,
0.0605782606,
-0.0524928011,
-0.00718723,
-0.0402009934,
0.0142357862,
0.028720988,
-0.0379274189,
-0.0323742777,
0.0454964451,
2.828e-7,
0.0395337455,
0.0495965704,
0.0221269503,
-0.0176935811,
0.0197643097,
0.040275868,
-0.001080495,
0.0491194054,
0.0078711864,
0.034446124,
0.0441375338,
0.0146367466,
0.0611906871,
-0.0199355055,
-0.0913265422,
-0.047996249,
-0.0452033803,
0.0233594757,
-0.0538114309,
-0.041681692,
-0.0083907861,
0.0928064883,
0.0030191541,
-0.0142031247,
0.0376529805,
0.0072441432,
0.0535477959,
0.0137041593,
0.0029171186,
0.0137637286,
0.087397866,
0.0027830761,
-0.0488925874,
-0.0350238979,
-0.0337742791,
0.0370215997,
0.0366140306,
0.0689267442,
-0.0128500424,
0.0308174882,
0.0058023515,
0.0169880092,
0.0127256112,
-0.0348487347,
0.0755523071,
-0.0410363749,
-0.038021069,
-0.0194252729,
-0.0419015586,
0.0105525423,
0.0306907948,
-0.0193294529,
-0.0546867065,
0.0281258225,
0.0292516574,
-0.0164239835,
-0.0253538843,
-0.0678498521,
-0.0311076231,
-0.0408347771,
-0.0033946498,
-0.0122486399,
-0.047822129,
0.0357193686,
0.0047238548,
-0.0327946916,
-0.0326398499,
2.830375961e-34,
0.0390733182,
-0.022789957,
0.0135629503,
-0.0343151055,
0.0255823769,
-0.0064848294,
0.010918634,
-0.0063607413,
0.0071181161,
-0.0568911433,
-0.0226490758
] |
http://arxiv.org/abs/2401.15351v2 | http://arxiv.org/pdf/2401.15351v2 | A Survey on Neural Topic Models: Methods, Applications, and Challenges | 2024-01-27T08:52:19 | 2024-06-24T12:08:06 | Xiaobao Wu, Thong Nguyen, Anh Tuan Luu | cs.CL | cs.CL, cs.AI, cs.IR | Topic models have been prevalent for decades to discover latent topics and
infer topic proportions of documents in an unsupervised fashion. They have been
widely used in various applications like text analysis and context
recommendation. Recently, the rise of neural networks has facilitated the
emergence of a new research field -- Neural Topic Models (NTMs). Different from
conventional topic models, NTMs directly optimize parameters without requiring
model-specific derivations. This endows NTMs with better scalability and
flexibility, resulting in significant research attention and plentiful new
methods and applications. In this paper, we present a comprehensive survey on
neural topic models concerning methods, applications, and challenges.
Specifically, we systematically organize current NTM methods according to their
network structures and introduce the NTMs for various scenarios like short
texts and bilingual documents. We also discuss a wide range of popular
applications built on NTMs. Finally, we highlight the challenges confronted by
NTMs to inspire future research. We accompany this survey with a repository for
easier access to the mentioned paper resources:
https://github.com/bobxwu/Paper-Neural-Topic-Models. | A Survey on Neural Topic Models: Methods, Applications, and Challenges: Topic models have been prevalent for decades to discover latent topics and
infer topic proportions of documents in an unsupervised fashion. They have been
widely used in various applications like text analysis and context
recommendation. Recently, the rise of neural networks has facilitated the
emergence of a new research field -- Neural Topic Models (NTMs). Different from
conventional topic models, NTMs directly optimize parameters without requiring
model-specific derivations. This endows NTMs with better scalability and
flexibility, resulting in significant research attention and plentiful new
methods and applications. In this paper, we present a comprehensive survey on
neural topic models concerning methods, applications, and challenges.
Specifically, we systematically organize current NTM methods according to their
network structures and introduce the NTMs for various scenarios like short
texts and bilingual documents. We also discuss a wide range of popular
applications built on NTMs. Finally, we highlight the challenges confronted by
NTMs to inspire future research. We accompany this survey with a repository for
easier access to the mentioned paper resources:
https://github.com/bobxwu/Paper-Neural-Topic-Models. | [
0.0285791624,
0.024540564,
-0.0258785821,
0.0579364598,
-0.0266174637,
-0.0030913702,
-0.005848527,
-0.051283367,
-0.0500332154,
0.0157408807,
-0.0368339419,
-0.0252538137,
-0.0035196093,
0.0076620034,
-0.0062505817,
-0.0045991978,
0.0084895678,
-0.0268533565,
0.009112847,
0.0018497413,
-0.0004868217,
0.0022463412,
-0.0050787083,
0.0202132035,
-0.0022372399,
-0.0089917658,
-0.0017140519,
-0.0282808524,
-0.0003012602,
-0.0035395082,
0.0312384758,
0.0294974241,
0.0098465793,
-0.0356239192,
0.000001762,
0.0059653413,
-0.0209721569,
-0.0569141209,
0.0694607124,
-0.0086222095,
0.0642489642,
-0.0312763788,
0.0306356791,
0.0063412092,
-0.0317012072,
0.0237503313,
-0.0178592224,
0.081165269,
0.0289973952,
0.0309335161,
-0.0246278513,
0.0112356683,
0.0454348214,
-0.0428069681,
0.0427675433,
-0.0392838083,
0.0155160883,
-0.0317998044,
-0.0125019057,
0.0618516803,
-0.0053489762,
0.0207619518,
0.0235666968,
-0.0267490447,
0.0051617119,
0.027809361,
0.002054197,
-0.0093001183,
0.0048188758,
0.0674830154,
-0.016579425,
0.0013592595,
0.0234824698,
0.0009801293,
-0.0423255675,
-0.0057013524,
-0.0476970375,
0.0203796346,
-0.0013480035,
0.0378232375,
0.0869734064,
0.0237122346,
0.039180126,
0.0291578081,
0.0034589306,
0.0505939126,
0.0127184847,
0.0094270473,
-0.0438623689,
-0.0139526464,
0.0138885593,
-0.021789616,
0.0688475892,
0.0304153822,
0.0295988172,
0.01355456,
0.0195876267,
-0.0640312955,
0.0182175469,
-0.0773517042,
-0.0515016504,
-0.0030152421,
-0.0406242907,
0.0571923144,
0.0286849774,
-0.0118491352,
-0.013088759,
-0.0596907325,
-0.0414615721,
-0.0514317863,
0.0059452346,
0.0131412083,
-0.0228531603,
0.0254569706,
0.0051699295,
-0.021928804,
0.0024723066,
0.0491928943,
0.0211851988,
-0.0181635525,
-0.0115757044,
0.0283993371,
-0.0125028649,
-0.0043535782,
0.0096014412,
0.0345637389,
-0.057664644,
0.0303839017,
-0.0449731201,
0.0279013216,
0.0187778231,
0.0398231149,
0.0076093897,
-0.0074141626,
0.0036999709,
0.0188363884,
-0.0037194961,
-0.0589034781,
-0.0488990508,
-0.004483439,
-0.0088755721,
-0.0144585436,
0.0277262311,
-0.1215793416,
0.0084216539,
0.0007156637,
-0.0050982642,
0.0086233029,
0.0065139187,
0.0415640809,
-0.0495654009,
0.0469248332,
-0.0112411743,
-0.0216786899,
0.0759081468,
0.0367480591,
0.0400709547,
0.0626056939,
-0.0022482527,
0.0506658852,
0.0691093057,
0.048731301,
-0.030845305,
-0.0063430932,
0.0105685536,
0.0312333703,
0.0034465916,
-0.0069648209,
-0.0055205887,
-0.0082739955,
0.0245016403,
0.0420800149,
-0.0413173661,
-0.0549157709,
0.0442090742,
0.0998809636,
0.0158944745,
0.0404084697,
0.0120421108,
-0.0061849495,
-0.0050332593,
-0.0018247234,
0.0341113321,
0.0180494115,
-0.0124977268,
-0.0338059664,
0.0644210055,
0.027367603,
-0.0799165294,
-0.0534795299,
0.0336791016,
-0.0386759825,
0.0118530365,
-0.043850407,
0.0893545672,
-0.0092449915,
0.0626114085,
0.0145116607,
0.0397826731,
-0.0007654896,
0.0438056514,
0.0234454013,
0.035092026,
0.020998098,
0.0335516594,
-0.0479111522,
-0.04718991,
-0.0145212151,
-0.0062933918,
-0.0144472569,
-0.0164272729,
-0.0226830635,
0.0203299243,
-0.0310143977,
0.0165566597,
-0.0387968086,
0.0217383597,
0.0201847218,
-0.0346955918,
0.0259386394,
0.0235388428,
-0.0406947732,
0.0316311754,
0.0081005963,
-0.0227528363,
0.0036583042,
0.0755224153,
-0.0191379748,
0.0429178104,
0.0503144041,
0.0247761514,
-0.016387688,
0.0482030436,
-0.0559906475,
-0.0317874365,
-0.0355413333,
-0.0234039649,
-0.0332940035,
0.0433429442,
-0.0684407428,
0.0545746423,
-0.0090871956,
0.004660476,
-0.0286143832,
0.0489824824,
-0.0217370056,
-0.0347257741,
-0.0517533235,
-0.0189505089,
-0.0238872226,
-0.0051705721,
-0.0162067581,
0.0315245278,
-0.0244933292,
0.051775977,
0.0226797853,
-0.0091479244,
0.0218521096,
-0.0348645672,
-0.0214553736,
0.0080122985,
0.0241062455,
0.0100581441,
0.0536602624,
0.0164139681,
0.0284190886,
-0.0262739416,
-0.0801285207,
0.0260303412,
-0.0482349321,
0.0137851564,
0.039022062,
0.000027004,
-0.0040364545,
-0.027215939,
-0.062313363,
0.0112138428,
-0.0230886266,
-0.0034385764,
0.0331792943,
0.0005018591,
0.0437465422,
-0.041301474,
-0.0296465363,
-0.016206393,
0.0487794727,
-0.005947521,
-0.0459278971,
-0.0441308878,
-0.003126445,
-0.0221788026,
0.0038865912,
-0.0336646028,
0.011401209,
0.0029541885,
-0.0376140922,
0.0696035251,
-0.013128167,
0.0073528606,
0.0190919656,
0.0489378013,
-0.0226031244,
-0.0151662249,
-0.0081842178,
-0.0426455028,
0.0696000829,
-0.0082917362,
-0.0743112862,
0.0389287286,
0.062054988,
-0.0097518312,
-0.01096451,
-0.0093881562,
0.0133416057,
-0.0243876092,
0.0535762236,
-0.0480973758,
0.0287142787,
0.0107778702,
0.0116752116,
-0.0010431184,
-0.0381868333,
0.0615835711,
-0.0274448246,
0.0404460356,
-0.0953413919,
-0.0496816263,
-0.1073006168,
0.0179080665,
-0.0050419299,
0.0011262798,
-0.0048971437,
0.0034465108,
0.0051525892,
0.0346138924,
-0.0084395325,
0.004872961,
0.074033387,
0.013275126,
0.0020831681,
-0.0001454923,
-0.0091378773,
-0.0088074366,
-0.0653196573,
0.0274063088,
0.0221205279,
0.0101675075,
-0.0037545292,
-0.0016104778,
0.0042361026,
0.0254545882,
0.023181349,
-0.0282860324,
-0.0119834356,
-0.0096749226,
0.0510168858,
0.0572608598,
-0.0894100219,
0.0052065803,
-0.0111675663,
-0.0011881499,
0.027522536,
0.0208042525,
0.0099970531,
-0.048623167,
0.0100385975,
0.0307364482,
-0.0087255565,
-0.0242188908,
-0.0227154661,
-0.0038026646,
0.0252099764,
-0.005023574,
0.0655594245,
0.0266169533,
-0.1132971495,
-0.0064625782,
-0.0207679681,
-0.0238510072,
-0.0022945956,
0.0245862007,
-0.0310270395,
0.0201166607,
-0.0313422419,
0.0310269706,
0.0318141989,
-0.0155205755,
-0.0184540078,
-0.0201954227,
0.0055861585,
-0.0089516668,
-0.0139167663,
-0.0061387578,
-0.0790719017,
0.044222381,
-0.0189383682,
-0.0205180049,
-0.0203097872,
-0.0048674596,
-0.0245740768,
-0.0013673686,
-0.0107190162,
-0.0384487472,
-0.0174736883,
-0.0646469593,
0.0102906721,
-0.0166274942,
0.0151546756,
-0.0029485128,
0.0135865612,
-0.0058946628,
-0.0433887318,
-0.0437147431,
-0.0112575889,
0.0232573282,
-0.0129051059,
0.0038007905,
-0.0166450534,
0.0285153184,
-0.0651155114,
-0.0156035004,
0.0140389726,
0.0830118656,
-0.0558988303,
0.0054624747,
-0.0355306864,
0.0000515754,
-0.0010286095,
0.0420119315,
0.0237324536,
-0.0827412754,
0.047927469,
-0.0066916938,
0.009356915,
0.0289563127,
0.0080001922,
-0.0471849367,
-0.0054755616,
-0.0595241785,
-0.0584356487,
-0.0596482381,
-0.0149667393,
-0.004123278,
-0.0135697648,
0.0476120226,
0.0338431858,
-0.0651058406,
0.007688242,
0.0411107913,
0.0009073116,
-0.0165823549,
-0.0294165425,
0.0065060179,
0.0648309663,
-0.0617278926,
0.0072258203,
-0.029634878,
-0.1191153228,
-0.0340916589,
-0.0280292723,
0.0010693504,
-0.0017556145,
0.0405788794,
-0.0745942444,
-0.1049456894,
0.007336034,
0.0424542241,
-0.043075081,
-0.0127486605,
-0.0148785254,
0.0206798278,
-0.0095824534,
0.0251846965,
-0.0463906229,
0.0489559807,
0.0130057549,
0.0261344872,
-0.0444320664,
-0.0594194196,
-0.0183358286,
-0.0569686778,
0.0753811225,
0.0101410775,
-0.0189807769,
-0.0939458087,
-0.0142052192,
-0.0121333953,
-0.0025236111,
-0.0176707841,
-0.0098273335,
0.0124842264,
-0.0009615906,
-0.0520342477,
-0.0171626136,
-0.0099775773,
0.0357691757,
-0.0036890015,
0.0147180855,
0.0135804703,
0.0291063506,
-0.0381840728,
0.0189874694,
0.0062705977,
0.0208008159,
-0.0343892239,
0.0232859291,
-0.0296197925,
0.0446275659,
0.0103451526,
-0.0405313373,
-0.023584364,
0.0138833337,
-0.049799934,
-0.0002025349,
-0.0026810183,
0.0429339483,
0.0608837493,
-0.0153982695,
-0.0057190028,
-0.0386073142,
0.0105881747,
-0.0291004684,
0.0172635224,
0.0246321298,
0.0010881705,
0.0047093844,
0.0266689323,
0.0038647901,
0.0457374342,
0.072072506,
-0.0017362322,
-0.0520361662,
0.0331050716,
-0.0272111464,
-0.0045721605,
0.0536478236,
0.0101200482,
0.0509606004,
-0.0777303204,
0.0204453226,
-0.0152307013,
0.001189142,
0.0460062549,
-0.0436941162,
-0.0280122925,
0.0059998473,
0.0491657071,
0.0995600671,
-0.0182175841,
0.0252915006,
-0.0590624325,
0.0066765733,
-0.0414135903,
-0.0510464683,
-0.0511980876,
-6.470741549e-33,
0.0213778745,
-0.0288981833,
0.0085983537,
0.0310585741,
0.0454071015,
0.0287405085,
-0.03964727,
0.0020170186,
0.0199191663,
-0.0017812239,
-0.0010219101,
0.0408080965,
-0.006948526,
0.0082917009,
-0.0110834371,
0.0101513444,
0.0240405165,
0.0038536352,
0.0062722121,
0.0490353629,
0.0344683751,
0.0326142274,
0.0406733975,
-0.0869882703,
0.0435836278,
-0.0338379666,
0.0402601659,
0.0357825607,
0.1065963805,
0.070701398,
-0.0317844898,
0.0493482687,
0.0637524724,
-0.0454303734,
0.0188275632,
0.0323538259,
-0.0656348169,
-0.1064908281,
-0.0309325419,
-0.0365451947,
-0.0046077031,
-0.0604443662,
0.0174188092,
0.0021707325,
0.0198037252,
0.0287043639,
0.0816479772,
0.0105986632,
-0.0073142764,
-0.0066752844,
-0.008386286,
0.0448858365,
0.0032885559,
0.0110949473,
0.0055198465,
-0.0011718695,
0.0124711664,
0.0101747643,
-0.064476721,
0.0420201048,
0.0092988322,
0.0265170801,
-0.0092863003,
-0.0252001509,
-0.0051856427,
0.0289346911,
-0.012833667,
0.04211054,
-0.0070587778,
-0.0115001751,
0.0339487493,
0.0168520845,
0.0662926361,
-0.005452049,
0.0021245212,
0.0311445538,
-0.0586543083,
0.0503699034,
0.0467886403,
0.0061043482,
0.0389419496,
0.0022527368,
-0.021354774,
-0.0358859412,
-0.0263632238,
-0.0314872786,
-0.0328151397,
-0.030560689,
-0.0383765772,
0.0447468087,
0.0028514995,
-0.0126062967,
0.0428095199,
-0.024743719,
0.0213446897,
0.0735858902,
-0.0087042814,
0.0270146765,
0.020192286,
0.0406431295,
-0.0976029485,
0.0002465729,
-0.0120258173,
-0.0421142653,
-0.006965918,
0.0178829506,
0.005929593,
-0.0233266503,
0.0089817848,
-0.0468349904,
0.0070123044,
-0.0469401143,
-0.0206421465,
0.0608123392,
0.000189464,
0.0252012499,
-0.0118431821,
0.068212457,
0.0616128221,
0.0440220051,
0.0071853995,
-0.0635390207,
-0.0185575634,
0.0187659748,
-0.0153163681,
0.0035284816,
-0.0192650184,
0.0454717614,
0.0235068165,
-0.0542070232,
0.0001702324,
0.0080112945,
2.547e-7,
0.0623922087,
0.0144548491,
-0.024620112,
-0.0248039402,
0.0300316215,
0.0179066788,
-0.0281720608,
0.0149106458,
-0.0029532565,
0.0221585836,
0.0211141389,
0.0406348258,
-0.0025904875,
0.0610021539,
-0.0863415673,
-0.0015984472,
-0.0407712311,
0.0020731981,
-0.0375333279,
0.0000517884,
-0.0189042054,
0.09006273,
0.0356031768,
-0.0059014168,
-0.0268428307,
-0.033104524,
0.088198334,
-0.0405624695,
-0.0539462268,
0.0305964034,
0.0447047912,
0.0208816491,
-0.0050085597,
-0.0566085316,
-0.0111606373,
0.0430486947,
0.0127307288,
0.0468206145,
0.0137552954,
-0.0175395533,
0.0419898182,
-0.019409515,
-0.0010439891,
-0.0359217115,
0.0625313744,
-0.0167893842,
-0.0116950674,
-0.0088361157,
0.0380915292,
0.029121628,
0.036302004,
-0.0247515067,
-0.0544178411,
0.000555396,
0.0021073478,
-0.0344622321,
0.0039809821,
-0.056382902,
0.0117040323,
0.0322453268,
-0.0300144199,
0.0040914044,
-0.0417068116,
-0.0052371807,
0.020561479,
-0.0322314128,
0.0664179176,
2.073073368e-34,
0.0127246678,
-0.0375683904,
-0.0817386881,
-0.0291267019,
-0.0145388283,
-0.0121971918,
-0.0781913623,
0.0054917512,
0.0162161607,
-0.0521534905,
-0.006025475
] |
http://arxiv.org/abs/2401.15347v1 | http://arxiv.org/pdf/2401.15347v1 | A Comprehensive Survey of Compression Algorithms for Language Models | 2024-01-27T08:38:56 | 2024-01-27T08:38:56 | Seungcheol Park, Jaehyeon Choi, Sojin Lee, U Kang | cs.CL | cs.CL, cs.AI, 68T50, I.2.7 | How can we compress language models without sacrificing accuracy? The number
of compression algorithms for language models is rapidly growing to benefit
from remarkable advances of recent language models without side effects due to
the gigantic size of language models, such as increased carbon emissions and
expensive maintenance fees. While numerous compression algorithms have shown
remarkable progress in compressing language models, it ironically becomes
challenging to capture emerging trends and identify the fundamental concepts
underlying them due to the excessive number of algorithms. In this paper, we
survey and summarize diverse compression algorithms including pruning,
quantization, knowledge distillation, low-rank approximation, parameter
sharing, and efficient architecture design. We not only summarize the overall
trend of diverse compression algorithms but also select representative
algorithms and provide in-depth analyses of them. We discuss the value of each
category of compression algorithms, and the desired properties of low-cost
compression algorithms which have a significant impact due to the emergence of
large language models. Finally, we introduce promising future research topics
based on our survey results. | A Comprehensive Survey of Compression Algorithms for Language Models: How can we compress language models without sacrificing accuracy? The number
of compression algorithms for language models is rapidly growing to benefit
from remarkable advances of recent language models without side effects due to
the gigantic size of language models, such as increased carbon emissions and
expensive maintenance fees. While numerous compression algorithms have shown
remarkable progress in compressing language models, it ironically becomes
challenging to capture emerging trends and identify the fundamental concepts
underlying them due to the excessive number of algorithms. In this paper, we
survey and summarize diverse compression algorithms including pruning,
quantization, knowledge distillation, low-rank approximation, parameter
sharing, and efficient architecture design. We not only summarize the overall
trend of diverse compression algorithms but also select representative
algorithms and provide in-depth analyses of them. We discuss the value of each
category of compression algorithms, and the desired properties of low-cost
compression algorithms which have a significant impact due to the emergence of
large language models. Finally, we introduce promising future research topics
based on our survey results. | [
-0.0200356208,
0.0892039388,
-0.0204217825,
0.03390963,
-0.0466716588,
-0.0014869255,
-0.0075121606,
0.0854766667,
-0.046205502,
0.0188799053,
0.0172947515,
-0.0426243655,
0.0387001745,
-0.0075328592,
0.0142234741,
-0.0355470255,
0.0271794125,
-0.0319069065,
0.0105942721,
-0.0431831852,
-0.0102190543,
-0.0085926475,
0.0619828925,
0.0371899679,
0.0119395591,
-0.0103208022,
-0.018045919,
-0.0134679172,
-0.0497884192,
0.0540298633,
-0.0232204366,
0.0625490919,
0.0336842872,
0.024984628,
0.0000020106,
-0.0031161224,
-0.0190007687,
-0.0023165043,
-0.0453962572,
0.0450050011,
0.0581254959,
-0.0251523741,
-0.0090291481,
0.0085769966,
-0.0044279532,
-0.0560551062,
0.0763477013,
-0.011903299,
0.0556648821,
0.023750471,
0.0033454441,
-0.0242329948,
-0.022561444,
-0.022199234,
-0.0516815148,
-0.0200688429,
0.0101846168,
0.014656364,
-0.0132784955,
0.0101114111,
-0.0536425747,
0.0179063827,
-0.0117794843,
0.0195059627,
-0.0365119278,
0.0161283091,
-0.0537511706,
-0.0233017094,
0.0248668976,
0.0343919396,
-0.0768341348,
-0.0655646846,
0.0038723128,
0.0468189679,
-0.006285761,
-0.0492159016,
-0.0639048368,
0.0460914522,
0.0364589654,
0.0439485013,
0.002039358,
-0.0068162205,
0.0289403405,
-0.0172128342,
-0.0126351416,
0.0404455662,
0.0274596699,
-0.0088650361,
-0.0448048674,
-0.0334949717,
-0.0285709482,
-0.0710676834,
-0.0114176376,
0.0037513999,
0.0388308875,
-0.0055775545,
-0.0821462199,
-0.0351157077,
0.0078697,
-0.0321142934,
0.018583646,
0.0434964933,
0.0481121205,
0.0145652881,
-0.0151863853,
0.0714677423,
0.018330561,
-0.017758863,
-0.0393613838,
-0.0186937824,
-0.0061640316,
-0.0057789073,
-0.0591904372,
0.0539412014,
-0.0237001795,
-0.0042592855,
-0.0327276401,
-0.0038097859,
0.0179943834,
0.0196556933,
0.0266234521,
-0.0090009235,
-0.0073720929,
-0.008282044,
0.034825176,
0.0244555604,
-0.0259195399,
-0.0091010164,
-0.0237665419,
-0.0354555026,
-0.0221188273,
0.0144442106,
0.0225920342,
0.0001638752,
-0.0010400463,
0.0042881817,
0.0248198584,
-0.0243843254,
0.0133025423,
0.0027903016,
-0.0047844071,
-0.0492317528,
-0.0147883557,
-0.0826158896,
0.0436366461,
-0.0377603583,
0.0195345785,
-0.0718845576,
-0.0009323625,
0.0036071267,
-0.063864626,
0.0171449333,
-0.0200448185,
0.033006791,
0.005491768,
0.014851179,
0.0511628948,
0.0736136958,
0.0264832545,
0.0092572179,
0.0200670995,
-0.0539086312,
-0.044003658,
0.0276844446,
0.0722922832,
0.0147340735,
-0.0529857837,
-0.019019112,
-0.0167509317,
-0.0071423245,
0.0353966914,
0.1194075942,
-0.0636024699,
-0.067681089,
0.0327583924,
0.0320251361,
0.0540490001,
0.0778407007,
0.0115248384,
0.0197645016,
0.0035863619,
0.007870256,
0.0091041736,
0.0106741143,
-0.0156326778,
0.0021683222,
0.0390608683,
0.045359578,
-0.0451066159,
-0.0100731542,
-0.0304767694,
-0.041800037,
0.0367779918,
0.0035582262,
0.0251998827,
0.0010386884,
-0.0041395565,
0.0904632807,
0.0192799512,
-0.0718989,
-0.01852246,
-0.050533127,
-0.0311183836,
0.0075544068,
0.0285443142,
0.0051460126,
0.0183948912,
0.0057245712,
-0.0187787078,
-0.0311697777,
0.0113359625,
-0.0101595623,
0.0216953717,
-0.0396913476,
-0.0094539449,
0.008505404,
0.0124102095,
0.003995222,
0.0106896842,
0.0156527646,
0.0078118364,
-0.0765217245,
-0.0143943019,
0.0159335807,
-0.0102640232,
-0.0097380215,
0.0176872052,
-0.0084494837,
0.0300744716,
-0.0132422894,
-0.0119747585,
0.0014822612,
0.0477338545,
-0.0465598181,
0.0115398187,
-0.0279782955,
0.0002963368,
-0.0571578853,
0.0545821637,
-0.0030571767,
0.0814697072,
0.0825900137,
-0.0261283182,
-0.0455203168,
0.0247135516,
-0.044651784,
0.052596055,
-0.0495604351,
0.0273498073,
-0.0519473217,
0.0543747917,
0.0139026036,
0.0481380112,
-0.0543201528,
0.0203033034,
0.0482220203,
-0.0281708594,
-0.0052529802,
-0.0171841756,
-0.0642928556,
0.0411394387,
-0.0251551978,
0.0371183716,
0.0376095958,
-0.0152262505,
0.0427973792,
-0.0133518176,
-0.0834813938,
-0.0036548087,
-0.0073133805,
0.0131060351,
0.0241459031,
0.046615053,
0.0225388445,
0.0063927704,
-0.0313497484,
0.0093168886,
0.0029346871,
0.0175329,
-0.0181614906,
-0.0047240243,
-0.0337997228,
-0.0246081855,
-0.071799241,
0.0286953747,
0.1009582281,
-0.097130388,
-0.0364727974,
-0.0607827306,
0.0413108505,
-0.0585579127,
0.0482867658,
-0.0022101568,
-0.0502067544,
-0.0325877443,
-0.0288813151,
0.0126651917,
0.0229305662,
0.002944805,
0.0301112235,
0.0909274518,
0.0013925133,
-0.0405378193,
0.0062141772,
-0.0417412743,
-0.0209182892,
-0.0140015641,
-0.0668178424,
-0.0217245314,
0.0953258201,
-0.0086213294,
-0.0341531076,
-0.0681788176,
-0.0173833445,
-0.0093615875,
-0.0216776729,
-0.0367423669,
-0.0098422533,
-0.0017887485,
0.0506942198,
-0.0006335111,
-0.0391367786,
0.0280969907,
-0.0681384355,
0.0008753847,
0.0176010299,
-0.0209266357,
-0.0128466086,
0.0454122201,
0.0300019253,
0.0737767965,
-0.0114714215,
0.0252475329,
-0.0469481163,
0.0315345079,
0.0098012099,
-0.0508086123,
-0.013389538,
0.0035163136,
0.0361472629,
-0.0024760286,
-0.0222565569,
0.0196601413,
-0.0113723939,
0.0037717889,
0.0494443141,
0.0445395596,
-0.0238382202,
0.006785064,
0.0929258391,
-0.0130355814,
0.0287860464,
0.0355175696,
0.0209865551,
0.016044898,
0.0148206567,
0.0347559378,
-0.0024543153,
0.0397800021,
-0.0005931091,
0.0307928976,
0.0531846434,
0.0639755353,
0.0495174788,
-0.0328922085,
-0.0123550463,
0.046398595,
-0.0097529292,
-0.0483742133,
0.0245425273,
-0.0001583575,
-0.0175366327,
0.0025971876,
0.0948063433,
0.022175286,
-0.0249792971,
0.0374079458,
-0.0035687978,
0.0829300508,
0.0008700183,
0.0234571751,
-0.0762162879,
-0.0021068533,
0.022140542,
-0.0039221812,
0.0138495257,
-0.023537321,
-0.0146646453,
-0.068776615,
0.0087152841,
0.0489485078,
-0.0004613726,
-0.0283396505,
-0.0306060612,
0.008535713,
-0.0047036097,
-0.0021202564,
0.0261338037,
-0.0010021153,
-0.0006705421,
0.0048316987,
-0.0560979843,
-0.0450248308,
0.0074290582,
0.0309318528,
0.0615325682,
0.0273171887,
0.0664170235,
0.0074368864,
-0.0091638295,
-0.0235168431,
0.0171164051,
-0.02773677,
-0.0328203365,
0.0050214767,
0.0364311151,
-0.0032174531,
0.0601518005,
0.0167003591,
-0.0140674002,
0.0061295088,
-0.024993008,
0.0720526427,
-0.0005059328,
-0.0541907884,
0.0097536556,
-0.0255590379,
-0.0220513083,
0.0644855201,
0.0394889116,
0.0104705356,
-0.0239109956,
-0.0113423308,
0.0096814446,
0.0175850242,
-0.0504291952,
0.0035493881,
-0.0007759925,
-0.0494930111,
-0.043518275,
0.0185052715,
0.0101523697,
-0.0029650133,
-0.0379582718,
0.011713881,
0.0126626799,
-0.0047109025,
-0.0798206329,
-0.0479579531,
-0.0409552678,
-0.0159543678,
-0.0095966998,
0.0141737359,
-0.0276825521,
0.0144740008,
-0.0360667557,
-0.0099010281,
-0.1187587008,
-0.0488280207,
-0.0142317675,
0.0552762747,
0.0249632057,
0.0376643129,
0.000360332,
-0.0547051802,
0.034345068,
0.0135580525,
-0.0202465057,
0.0732905716,
0.0243676733,
-0.0771036148,
0.0238294285,
0.0754353255,
-0.0237653293,
0.0270763002,
-0.0312056337,
0.0396157056,
-0.0366943665,
-0.0499891303,
0.0346060656,
0.0428735167,
0.0009421846,
-0.0102572432,
-0.0247044321,
0.030543128,
0.0209090877,
-0.0207551643,
-0.0039257384,
-0.0558898449,
0.0244642384,
0.0152133964,
0.0439660437,
0.0219803769,
-0.0005478932,
0.0474251099,
0.0044141342,
0.013227649,
-0.0127844736,
-0.0218661726,
-0.0104228631,
0.0670105889,
0.0141723705,
-0.0045098932,
-0.0193233229,
-0.0031466067,
0.0192003902,
-0.0593679622,
-0.029160032,
0.0124653978,
0.0432463363,
-0.0378747545,
0.0322624072,
-0.0105499644,
-0.0838776827,
-0.038035728,
0.0235110763,
0.0389524475,
0.0169829335,
0.0271791611,
-0.0257342216,
-0.0040798755,
-0.0312974043,
0.0055464595,
0.0273637976,
-0.0230330359,
0.00254752,
0.0146609722,
-0.0229567029,
0.0077835359,
0.1030054241,
0.0261106156,
0.0422001816,
-0.0142440023,
-0.0082550412,
-0.0492817573,
-0.028715726,
-0.0721506104,
0.0502744988,
0.0011033114,
-0.0132189216,
-0.0421159267,
-0.0534349717,
-0.0002624929,
-0.0096633416,
-0.0074462146,
-0.0166727547,
0.0293109454,
0.0327208452,
0.0221918616,
0.0184546784,
-0.0335349664,
-0.0431312583,
-0.0321169347,
-0.030994093,
0.0311934073,
-5.859686892e-33,
-0.0334593318,
0.0324145518,
-0.0108431503,
-0.033600457,
-0.0127767613,
0.028360121,
-0.0141920159,
-0.0051008496,
-0.0356879793,
-0.0039499644,
-0.0134262163,
0.0129312715,
0.0018172524,
0.010457891,
-0.0512586907,
0.036416661,
0.0588684268,
-0.0202462804,
-0.027092291,
0.0431079157,
0.0166408643,
0.0149451345,
0.040877603,
0.0043052551,
-0.0455099419,
-0.0084675262,
-0.0016049726,
0.0018265079,
0.0354148075,
0.0311813541,
-0.0460835807,
0.0399885215,
0.0018294284,
-0.0291375015,
-0.0340458117,
0.0314945206,
-0.0584772229,
-0.0225252919,
0.0350143537,
-0.017987242,
-0.0047190282,
-0.0245857742,
0.0304925609,
0.0224372856,
0.0153905377,
-0.0290219057,
-0.0252682026,
-0.0442212932,
-0.0461410582,
-0.0354085788,
-0.0119010303,
0.0060203383,
-0.0082672369,
-0.0018118636,
0.0913816988,
0.0260980241,
-0.0405441262,
0.0215111058,
-0.0359738544,
0.0248879399,
0.0117945438,
0.1321819425,
0.0472737104,
-0.0325099789,
0.031134028,
0.0214901194,
-0.0089660259,
-0.016630413,
-0.0318097658,
0.0626959205,
0.0658786073,
0.0337623507,
0.0295527577,
0.0436082855,
0.042718906,
-0.0251321234,
-0.0538753159,
-0.000046271,
0.0332107693,
0.0280228071,
0.0062731425,
0.0218087099,
-0.002797442,
-0.0558121763,
-0.0595524162,
-0.0230758879,
-0.0190136284,
-0.0264737364,
-0.000193132,
0.0148318242,
-0.0619712248,
-0.0230280645,
0.0391158462,
-0.0768466443,
-0.0347385854,
0.0399349146,
-0.0206269063,
0.0236931667,
0.0275271591,
0.0459050536,
-0.0441914611,
-0.0238798857,
0.0022419172,
-0.0352841653,
0.0041879825,
-0.0150727257,
0.0006428565,
0.0567681715,
-0.0518824197,
-0.0146357436,
-0.0176730435,
-0.0092890281,
-0.0143156173,
0.0192335304,
0.020453589,
0.0193265211,
0.0277439859,
-0.0056302762,
0.0421851352,
0.0143202301,
-0.0043621282,
-0.0031858885,
0.0294835139,
0.0365241952,
-0.0241543092,
-0.0245538931,
-0.0798150077,
0.0057378062,
0.0777006745,
-0.029310815,
-0.0426650457,
0.022333283,
2.555e-7,
0.0309149083,
0.0218351651,
0.0173681062,
-0.0275579747,
0.0224242583,
0.0047415514,
0.0250998344,
0.0183280166,
-0.018720964,
-0.0178703386,
0.0275029521,
0.024804214,
0.0303132292,
-0.0035967631,
-0.0698769912,
0.0155980429,
-0.0598507077,
-0.0431903303,
-0.024957763,
-0.0338851474,
-0.0134299025,
0.1138889343,
0.0176516008,
-0.0158836916,
-0.0199457556,
0.0048174653,
0.0377417244,
-0.0275267288,
0.0226299837,
0.0150743974,
0.0281858016,
-0.0050888779,
-0.0241038576,
-0.0248161182,
-0.0055040172,
0.0280481707,
-0.013114268,
0.0808021054,
-0.0210714377,
0.031248711,
0.0278456081,
-0.034222994,
-0.0613804124,
-0.0432850383,
0.0609380826,
0.0220617373,
-0.0043197107,
-0.0761815608,
0.0039700749,
0.0033355176,
-0.0343516469,
-0.0140829328,
0.0210010428,
0.0180896074,
0.0178443007,
0.0032763334,
0.021783337,
-0.04997335,
-0.0519160591,
0.0096416678,
-0.0337253809,
-0.0376857221,
-0.0121022202,
-0.0188726895,
0.0188785326,
-0.034558028,
0.0049111242,
2.547070573e-34,
0.0255007651,
0.015380743,
0.0092985099,
0.0345780365,
-0.0212963205,
-0.0679145306,
-0.0152009521,
0.0238808002,
-0.0391567051,
-0.067092061,
-0.0411937721
] |
http://arxiv.org/abs/2402.01725v1 | http://arxiv.org/pdf/2402.01725v1 | Fortifying Ethical Boundaries in AI: Advanced Strategies for Enhancing Security in Large Language Models | 2024-01-27T08:09:33 | 2024-01-27T08:09:33 | Yunhong He, Jianling Qiu, Wei Zhang, Zhengqing Yuan | cs.CL | cs.CL, cs.AI | Recent advancements in large language models (LLMs) have significantly
enhanced capabilities in natural language processing and artificial
intelligence. These models, including GPT-3.5 and LLaMA-2, have revolutionized
text generation, translation, and question-answering tasks due to the
transformative Transformer model. Despite their widespread use, LLMs present
challenges such as ethical dilemmas when models are compelled to respond
inappropriately, susceptibility to phishing attacks, and privacy violations.
This paper addresses these challenges by introducing a multi-pronged approach
that includes: 1) filtering sensitive vocabulary from user input to prevent
unethical responses; 2) detecting role-playing to halt interactions that could
lead to 'prison break' scenarios; 3) implementing custom rule engines to
restrict the generation of prohibited content; and 4) extending these
methodologies to various LLM derivatives like Multi-Model Large Language Models
(MLLMs). Our approach not only fortifies models against unethical manipulations
and privacy breaches but also maintains their high performance across tasks. We
demonstrate state-of-the-art performance under various attack prompts, without
compromising the model's core functionalities. Furthermore, the introduction of
differentiated security levels empowers users to control their personal data
disclosure. Our methods contribute to reducing social risks and conflicts
arising from technological abuse, enhance data protection, and promote social
equity. Collectively, this research provides a framework for balancing the
efficiency of question-answering systems with user privacy and ethical
standards, ensuring a safer user experience and fostering trust in AI
technology. | Fortifying Ethical Boundaries in AI: Advanced Strategies for Enhancing Security in Large Language Models: Recent advancements in large language models (LLMs) have significantly
enhanced capabilities in natural language processing and artificial
intelligence. These models, including GPT-3.5 and LLaMA-2, have revolutionized
text generation, translation, and question-answering tasks due to the
transformative Transformer model. Despite their widespread use, LLMs present
challenges such as ethical dilemmas when models are compelled to respond
inappropriately, susceptibility to phishing attacks, and privacy violations.
This paper addresses these challenges by introducing a multi-pronged approach
that includes: 1) filtering sensitive vocabulary from user input to prevent
unethical responses; 2) detecting role-playing to halt interactions that could
lead to 'prison break' scenarios; 3) implementing custom rule engines to
restrict the generation of prohibited content; and 4) extending these
methodologies to various LLM derivatives like Multi-Model Large Language Models
(MLLMs). Our approach not only fortifies models against unethical manipulations
and privacy breaches but also maintains their high performance across tasks. We
demonstrate state-of-the-art performance under various attack prompts, without
compromising the model's core functionalities. Furthermore, the introduction of
differentiated security levels empowers users to control their personal data
disclosure. Our methods contribute to reducing social risks and conflicts
arising from technological abuse, enhance data protection, and promote social
equity. Collectively, this research provides a framework for balancing the
efficiency of question-answering systems with user privacy and ethical
standards, ensuring a safer user experience and fostering trust in AI
technology. | [
0.0936897025,
0.0835391432,
-0.0242417622,
0.0719108582,
-0.0558515266,
-0.0212424714,
0.0574131757,
-0.0046407729,
0.0287755821,
-0.040897835,
-0.025362812,
-0.050032977,
-0.0307559911,
0.0419753455,
0.011659747,
-0.0079718325,
0.0261475127,
-0.0318408348,
0.0364000909,
-0.0299340356,
0.0055260928,
0.0162313282,
0.0315783024,
0.0425115265,
0.0084266439,
-0.0013571644,
0.0249224901,
-0.0214616321,
0.0292684343,
0.0072264224,
0.0191706661,
0.0403749496,
0.0051222057,
0.0798306465,
0.0000021054,
-0.091153197,
-0.0266875699,
0.0376334041,
-0.0239119884,
-0.0042220349,
0.0315159969,
-0.0132397851,
0.0062230355,
0.0338294953,
-0.0729459152,
0.0403210446,
0.0506934486,
0.0559410937,
0.1012870595,
0.0807065144,
-0.0157116428,
-0.0409711823,
0.0438242704,
-0.0130474679,
-0.0175782442,
-0.0526546054,
0.0531896427,
-0.0407022797,
-0.0021687089,
0.0366549864,
0.01333813,
0.0385951139,
0.0283721313,
-0.006298664,
-0.0170077402,
0.0136289056,
-0.0155887175,
-0.051822722,
-0.0013182127,
0.0580613352,
-0.0212832093,
-0.0199008174,
0.015897816,
0.0243890081,
0.0404253826,
0.0358285941,
-0.0180478785,
0.0952828899,
-0.0269877743,
-0.0221706219,
0.0264133327,
0.0037605555,
0.0274827387,
0.0663052276,
0.0104340771,
0.0364118889,
0.0132948626,
-0.0364274718,
-0.0402585752,
-0.0340987667,
-0.0129542826,
-0.0422432534,
0.0567154363,
0.0414302647,
-0.0210801139,
0.0174230579,
-0.0714279413,
-0.0081694331,
0.0417995527,
-0.0355411246,
0.0022596894,
0.0153906215,
0.0478105918,
0.0230074115,
0.022385072,
0.0305138063,
-0.0374947079,
0.0382144488,
-0.0094071859,
-0.0150882211,
0.0252820607,
-0.0409669615,
-0.0398619249,
0.0648444444,
-0.0410948284,
0.0204277579,
-0.0108702658,
0.03980821,
-0.0367901027,
0.0342137292,
0.0101503432,
0.0147313522,
-0.0213218536,
-0.0197889749,
-0.0025307024,
-0.0309508108,
-0.0410681404,
0.0086266482,
0.0045392509,
-0.0690859333,
0.0001324004,
-0.0019924399,
-0.0292914286,
-0.0269644503,
-0.0012159087,
0.1000537723,
0.0195602793,
-0.0153743597,
-0.0217565931,
-0.0962482095,
-0.0122855688,
-0.0458388887,
0.0092694806,
-0.0359618738,
0.0368654504,
-0.0174773838,
-0.0235301591,
-0.0601478554,
0.0121470038,
0.0539104789,
-0.0301958751,
0.0624364763,
0.0033763442,
-0.003473805,
0.0329433009,
0.0581321009,
-0.0096425926,
0.0038718558,
0.0369710959,
0.0414577201,
0.0800175071,
0.0009414266,
0.0442058854,
0.0301469024,
0.0006339055,
-0.0125013599,
0.0046233693,
0.0351911746,
-0.0234091152,
-0.0053971866,
-0.0361130126,
0.0841603354,
-0.0194059946,
-0.0624657273,
0.105559431,
0.0769675598,
-0.0415299796,
0.0338526219,
0.0035777362,
0.0162421167,
-0.0006766504,
-0.0264259335,
-0.053013511,
0.0121801887,
-0.0532961935,
-0.045944877,
-0.0587126426,
0.055638358,
0.0178184472,
-0.0196760539,
0.0139305014,
-0.0396725498,
0.0383232199,
0.038229391,
-0.0246780366,
0.0155117176,
-0.0012191575,
0.0062091961,
-0.0004920966,
-0.060176868,
-0.0041991621,
-0.0907276124,
0.0308070146,
0.0568860583,
0.0559051447,
-0.0117469216,
0.0042928653,
-0.0615464747,
-0.008191105,
0.0473952964,
0.0012266667,
0.0133397058,
-0.0036314682,
-0.0365092233,
-0.0350806937,
0.0116869807,
-0.022067165,
-0.0145500191,
0.0310666785,
-0.0084365923,
0.019688433,
0.0013169284,
-0.0008726111,
0.0173414052,
-0.000778597,
-0.0313332491,
0.015820086,
0.0123404022,
0.0338426456,
0.0082666539,
-0.0070864684,
0.0169062037,
0.0308902133,
-0.0043944689,
-0.0228388906,
-0.0405803844,
-0.0243576802,
0.0500698648,
0.0035696307,
-0.0333722197,
0.0168680903,
0.0840654373,
-0.011071288,
0.0303752385,
-0.0216272157,
-0.0271230433,
-0.0159780178,
-0.0445549563,
-0.0077046137,
-0.0221288931,
0.0194044933,
-0.0619420484,
0.0535506271,
-0.0322665051,
-0.0184645839,
-0.0241554417,
-0.022815723,
-0.0024625515,
-0.0166327842,
-0.0215637349,
0.0318828076,
-0.0188544486,
0.0454552025,
0.0429216102,
0.0152416648,
-0.0153009528,
-0.0404103622,
-0.0320415646,
0.0144599117,
-0.0060723131,
0.0295332093,
0.0128665175,
0.0131560368,
-0.0139020681,
-0.0012724062,
-0.0574388057,
-0.0256053302,
0.0034820011,
0.0254155342,
0.027511334,
-0.0122773107,
-0.0456459746,
0.0055080764,
-0.0030009996,
-0.0077385185,
0.0287876949,
0.0034529814,
-0.0162553117,
-0.03353405,
-0.0060151638,
-0.0367714874,
0.0232142974,
0.0120413369,
-0.0100301625,
-0.0020500077,
0.0027251025,
-0.0045570638,
0.0399162546,
0.0677800328,
0.0355678834,
0.044856634,
0.0292767491,
-0.0284181163,
-0.02257232,
0.0161512941,
-0.0095132738,
-0.0358547345,
-0.0084196543,
0.0283352733,
0.0056187664,
0.0317140631,
-0.0252219215,
-0.0378093347,
0.0735251382,
-0.0287222713,
-0.0649244487,
0.0200345069,
-0.0501086935,
-0.0238257181,
0.0008657179,
-0.0100327795,
-0.0396894366,
0.0773515999,
-0.0053765639,
0.0303651392,
-0.0217075665,
-0.0253771506,
-0.0094804484,
0.0530009009,
0.0141303055,
0.0733859241,
0.0338553861,
-0.0190557986,
-0.0281943474,
-0.0065958435,
0.0390395857,
-0.0558153391,
0.0233521238,
-0.0214617308,
-0.0083939806,
-0.0039545051,
-0.0593511164,
-0.060931161,
-0.0861256346,
-0.0037470954,
-0.0041806572,
-0.0038332129,
-0.0385632589,
-0.0089741452,
-0.0377135016,
0.0457669385,
-0.0076231291,
-0.0269760862,
-0.0094222734,
-0.0329127163,
0.0403973833,
0.0397393182,
-0.0033973837,
0.0087205358,
0.008727761,
0.0438012555,
0.094389759,
0.0069523584,
-0.0120810606,
0.0349039398,
-0.0098089892,
0.0334545895,
-0.0377102457,
-0.0036301389,
0.0207621455,
0.011077988,
-0.0471224152,
-0.002772664,
0.0770935118,
0.0307660848,
-0.0157307815,
0.0033496618,
-0.0027871218,
-0.0052667665,
-0.0072571477,
0.0046758251,
-0.0925581157,
-0.0004714105,
-0.002756814,
0.0367447026,
-0.0520760678,
-0.0471630506,
-0.0188340582,
-0.0856210887,
0.0243474618,
-0.0127418293,
-0.0178990588,
-0.0269728936,
-0.0529154092,
-0.0433078334,
0.0167159867,
0.0214254335,
-0.0370901451,
-0.0218506251,
-0.0332375132,
0.0007864495,
0.0136391781,
0.0025830241,
0.0207859762,
-0.0586721972,
0.0397518314,
0.0771679059,
0.0387597457,
-0.0415905118,
-0.0383303687,
-0.0163104553,
0.0397860557,
-0.0291468687,
0.0079776058,
0.0255486201,
-0.0214947816,
-0.0101504149,
-0.0254602246,
0.042886436,
0.0095919594,
-0.0278878883,
0.0284128599,
0.0758917108,
-0.0228155442,
-0.0374606252,
-0.0240912512,
-0.0058316924,
0.0303135496,
0.0700215325,
-0.0189521816,
-0.0558378063,
-0.0022879655,
-0.0387957208,
-0.0093574645,
-0.0105014937,
-0.0419616215,
0.0080327354,
0.0285729542,
-0.0180150624,
-0.0700087473,
-0.016171068,
-0.0212078951,
-0.0128277885,
0.0277444236,
-0.0246679336,
0.0380650721,
0.0149831511,
-0.0178464558,
-0.0167781468,
-0.0923131928,
0.0470337868,
-0.0220494159,
0.0418054461,
0.0265515223,
-0.0415081531,
-0.0188383032,
0.0232137777,
-0.0199706983,
-0.0366724283,
0.0301558264,
0.069064796,
0.0222702362,
-0.0021410943,
-0.0324811228,
-0.0496133193,
0.0408563539,
0.0472077318,
-0.0494826622,
0.0117735285,
0.0447752774,
-0.0242731515,
-0.0026580375,
0.0186524354,
-0.0152461641,
-0.0069301142,
0.0123123219,
-0.0123319104,
-0.0637965053,
-0.049627766,
0.0128871035,
0.0560827404,
-0.0215803068,
0.0204959214,
0.0130456947,
-0.0134033104,
-0.006231261,
-0.0410903879,
-0.0346686095,
-0.0088805538,
0.015048746,
0.0463759713,
-0.0066337506,
0.0296278894,
0.0027290268,
-0.0109503288,
-0.0169981569,
-0.043534074,
0.0100848647,
0.0114473067,
0.0340447798,
0.0035478943,
0.0079948464,
-0.0349515006,
0.0444591194,
0.024741713,
0.035875313,
-0.0029655478,
0.0433122069,
0.0000012288,
-0.0235383753,
-0.0125857247,
0.0673085377,
-0.0703658164,
-0.0675010905,
0.0099954195,
0.027958354,
-0.0138160577,
0.0291458908,
0.0020464503,
-0.0172811504,
-0.0066333101,
-0.00875175,
-0.0285891779,
0.0187901147,
-0.0637649819,
0.0214914419,
0.0078117987,
-0.0103829149,
0.0339830555,
0.0668990538,
0.0154754445,
0.0389060676,
-0.0024517099,
-0.0366218574,
-0.0234734621,
-0.0495662801,
0.013597602,
-0.0022576845,
-0.0178663973,
-0.0268057119,
-0.0683726072,
-0.0515314266,
-0.036021743,
-0.0067791231,
0.0019211652,
-0.0170985647,
0.0176150668,
0.0495334305,
0.0190866087,
-0.0287359394,
-0.0808224604,
0.0094470233,
0.0006304496,
-0.0254237987,
-0.0069841389,
-5.866564269e-33,
0.0192301702,
-0.052001521,
-0.0156115806,
0.0213914942,
-0.0087987566,
-0.0009678408,
0.0321989693,
-0.012078994,
-0.0546159148,
-0.0043830541,
-0.0367083065,
0.0096842907,
-0.00807912,
0.0283594728,
-0.0445310622,
-0.0500336103,
0.037125621,
-0.0262070168,
-0.0211058948,
0.0461976156,
0.0408918299,
-0.0065229633,
0.0047538136,
-0.0685526654,
0.0513641723,
0.0059608924,
-0.0529155172,
0.0352645926,
-0.0719419718,
0.0469112471,
-0.0039055571,
0.0594360232,
0.0264865141,
-0.0368616283,
0.0213726852,
0.0921958312,
-0.0326405652,
-0.0073251612,
0.0148500511,
-0.0165966637,
-0.0021533305,
-0.0993006453,
0.0653547272,
-0.0371832475,
-0.0081686685,
-0.0046132579,
0.0001738597,
-0.0009595847,
0.014192773,
-0.0565669984,
-0.0502428748,
-0.0244915318,
-0.0102570569,
-0.0204791725,
-0.0094582727,
0.0528321043,
0.0182823725,
0.0034151501,
-0.0239559095,
0.0279686097,
0.0201362409,
-0.0396784917,
0.0431767777,
-0.0093052471,
-0.0194122382,
0.0322921276,
-0.0024771173,
0.0102896374,
-0.0238543153,
-0.0270676296,
-0.043931108,
0.0469905585,
0.0603769533,
0.0597481951,
-0.0186792258,
-0.0208335444,
-0.0697305202,
0.025456531,
-0.0057529383,
0.0502639152,
0.0449586324,
-0.0412687138,
-0.0588492602,
0.0127744097,
-0.0286350567,
0.0039593517,
-0.0515323058,
-0.0000490395,
0.0317476876,
-0.0015021437,
-0.0723943636,
0.0065473504,
0.0172259789,
-0.0856268108,
0.0028970097,
0.0440030098,
0.0514892749,
-0.035804294,
0.0029504551,
0.0376571752,
0.0095892735,
0.0117195221,
0.0183195397,
-0.0676152185,
0.0199428014,
0.0042612697,
0.0118965674,
0.0329600759,
-0.0485925674,
-0.003952547,
-0.0133621283,
0.0561634004,
0.0099416897,
0.0336292721,
0.0173006244,
0.0187321659,
0.0034793282,
0.0044344408,
0.0363577418,
0.0146102877,
-0.0526363477,
0.0300251655,
-0.0107984897,
0.011692185,
0.00582308,
0.0110113947,
-0.0250808876,
0.0488379598,
0.0400031656,
-0.0428490676,
-0.0130511047,
-0.0319713652,
2.904e-7,
-0.0013190135,
0.0059562465,
-0.0136126298,
0.0279684346,
-0.001923927,
0.066379562,
0.0675111115,
0.0870133042,
-0.0020581791,
-0.0029380377,
-0.0016480717,
0.0238110404,
0.0206748359,
0.0442711785,
-0.0485093258,
-0.0487226062,
-0.0525492691,
-0.0055313413,
-0.0602346882,
-0.0033187431,
0.080211103,
0.0741087645,
0.0185101219,
-0.0413093418,
-0.0371433236,
-0.067741625,
0.023473436,
-0.0281992182,
-0.0231623221,
-0.0012915591,
0.0595472753,
0.0282742344,
0.0391959473,
0.0122177033,
-0.0127125466,
0.0049176821,
-0.0124107813,
0.0967982709,
0.0049039046,
-0.0052295756,
-0.0110675422,
-0.0754667744,
0.0117503079,
-0.0766126961,
0.0770946592,
-0.0028601752,
-0.021758873,
-0.0159364659,
0.0162967164,
-0.0672294945,
0.0279085264,
-0.0701822713,
-0.0337019749,
-0.0050196652,
0.0193616953,
0.0206910782,
0.0413127951,
-0.0352217555,
-0.0235280748,
-0.0409468897,
-0.0124409245,
-0.037122611,
-0.0022427642,
0.0070840642,
-0.0088703725,
-0.034720026,
0.0423488319,
2.802945617e-34,
-0.0291513875,
-0.0292458944,
0.0175982136,
0.0359458439,
0.0399595313,
0.0482450724,
-0.0074246074,
0.0020642038,
0.0161795542,
-0.0461029559,
-0.0153400945
] |
http://arxiv.org/abs/2401.15328v2 | http://arxiv.org/pdf/2401.15328v2 | Equipping Language Models with Tool Use Capability for Tabular Data Analysis in Finance | 2024-01-27T07:08:37 | 2024-01-30T06:08:33 | Adrian Theuma, Ehsan Shareghi | cs.CL | cs.CL | Large language models (LLMs) have exhibited an array of reasoning
capabilities but face challenges like error propagation and hallucination,
particularly in specialised areas like finance, where data is heterogeneous,
and precision is paramount. We explore the potential of language model
augmentation with external tools to mitigate these limitations and offload
certain reasoning steps to external tools that are more suited for the task,
instead of solely depending on the LLM's inherent abilities. More concretely,
using financial domain question-answering datasets, we apply supervised
fine-tuning on a LLaMA-2 13B Chat model to act both as a 'task router' and
'task solver'. The 'task router' dynamically directs a question to either be
answered internally by the LLM or externally via the right tool from the tool
set. Our tool-equipped SFT model, Raven, demonstrates an improvement of 35.2%
and 5.06% over the base model and SFT-only baselines, respectively, and is
highly competitive with strong GPT-3.5 results. To the best of our knowledge,
our work is the first that investigates tool augmentation of language models
for the finance domain. | Equipping Language Models with Tool Use Capability for Tabular Data Analysis in Finance: Large language models (LLMs) have exhibited an array of reasoning
capabilities but face challenges like error propagation and hallucination,
particularly in specialised areas like finance, where data is heterogeneous,
and precision is paramount. We explore the potential of language model
augmentation with external tools to mitigate these limitations and offload
certain reasoning steps to external tools that are more suited for the task,
instead of solely depending on the LLM's inherent abilities. More concretely,
using financial domain question-answering datasets, we apply supervised
fine-tuning on a LLaMA-2 13B Chat model to act both as a 'task router' and
'task solver'. The 'task router' dynamically directs a question to either be
answered internally by the LLM or externally via the right tool from the tool
set. Our tool-equipped SFT model, Raven, demonstrates an improvement of 35.2%
and 5.06% over the base model and SFT-only baselines, respectively, and is
highly competitive with strong GPT-3.5 results. To the best of our knowledge,
our work is the first that investigates tool augmentation of language models
for the finance domain. | [
0.0345167518,
0.0530431494,
-0.0205272809,
-0.0006009253,
0.0075265351,
0.0340001434,
0.020441018,
0.0123242158,
0.0203784369,
-0.0277251191,
-0.0464777723,
-0.0205027554,
-0.0377956368,
0.0480573066,
0.05914535,
-0.0277503468,
0.0282782167,
-0.0191203263,
0.0209351126,
-0.0252821688,
0.0191481542,
-0.0071511329,
0.0146680856,
0.0195084251,
0.0243973397,
-0.0225694794,
-0.0303395633,
-0.01355597,
-0.0217658263,
0.0470422432,
0.0387514681,
0.0765965134,
0.0081041809,
0.0701081306,
0.0000022025,
-0.045903489,
-0.0341932625,
-0.0096980119,
-0.0459901243,
-0.0178537555,
0.0136472844,
-0.0101180393,
0.0447921678,
0.030272942,
-0.0687765926,
0.0179465972,
0.0511682518,
0.0250794869,
0.0982967839,
0.0743901655,
-0.0041980599,
-0.0160591751,
0.0262036454,
-0.0520063415,
0.0390968323,
-0.0609146021,
0.014067756,
0.0646290928,
0.0037075488,
0.009865487,
-0.0044199056,
0.0438140668,
0.0274886638,
0.0045245388,
0.0253014751,
-0.0010119017,
-0.0507207364,
0.0222160667,
-0.0268933475,
0.0559487715,
0.0005784708,
0.0060166083,
0.0012730159,
-0.0025975166,
0.0236236174,
0.0494539477,
-0.0746297762,
0.0089834556,
-0.0149478894,
-0.0314453021,
0.0357172452,
-0.0567848571,
0.0082197962,
-0.0031301556,
-0.0342569575,
0.0592268221,
-0.0318915509,
-0.0112995701,
0.0149834063,
0.0212946441,
-0.0138498191,
-0.0585892238,
0.0004704826,
0.0574699715,
-0.0387933627,
0.0033380128,
-0.0012813587,
-0.0311599243,
-0.001770623,
-0.0709885582,
0.019167006,
0.0571778193,
-0.0123923765,
0.0380341485,
-0.0257894453,
0.0092794793,
-0.0036458392,
0.0256355219,
-0.0323428549,
-0.0093672751,
-0.0622838959,
-0.0250827745,
-0.0546665899,
0.0401791446,
-0.0551065691,
-0.0206255559,
-0.0406963974,
0.0232494585,
-0.0105335787,
0.0074825827,
0.0193341319,
-0.0240107197,
-0.0436771996,
-0.0256315377,
-0.039306175,
-0.0079889139,
-0.0472144596,
0.0219761524,
0.0089302603,
-0.0602771193,
-0.0118075721,
0.0038520899,
-0.0351285152,
-0.032769531,
0.0170074087,
0.109177649,
0.0023026674,
0.0009831631,
-0.0500376411,
-0.050471846,
0.0139532955,
-0.0544793978,
-0.0057262816,
-0.0536952876,
0.0366949067,
0.0207419004,
-0.0238085836,
-0.0120485425,
-0.0025361313,
0.0070089465,
-0.0254949685,
0.0747640654,
-0.011204931,
-0.0080433795,
0.0165985711,
-0.0198017713,
-0.0237289518,
0.090598233,
-0.0071258158,
0.0281572249,
0.0179465432,
-0.0081683509,
0.0099945134,
0.0608817376,
0.0214108676,
-0.034608271,
-0.016751891,
0.0582449771,
0.0188996047,
-0.0184324346,
0.0380747914,
0.0045248638,
-0.0391091779,
-0.0046685841,
0.0591985285,
0.0191079266,
0.0625392497,
0.0498405173,
-0.0065918285,
0.0075225877,
0.0333717428,
0.0146635463,
-0.0281760097,
0.0354371369,
-0.0229020473,
0.0238933396,
0.034898784,
0.0100701805,
-0.0391697474,
-0.0457383096,
0.002661661,
0.0064800307,
-0.0273515414,
0.0617458075,
0.019535033,
-0.0195698403,
-0.0311872195,
0.0793943852,
0.0289799236,
-0.0023046969,
-0.0489684679,
-0.0658605993,
0.0350785702,
0.0776120275,
0.0471940972,
-0.0422339328,
-0.0305350181,
-0.0337709896,
0.0039119371,
0.0357433297,
0.0599566884,
-0.0207466409,
-0.0437855758,
-0.0468204618,
-0.0052384911,
0.0097875586,
0.0249539353,
0.0166852754,
0.0328227915,
-0.0171430539,
-0.0148951774,
-0.0309444238,
-0.0265055262,
0.0076887202,
0.017758524,
-0.0317634121,
-0.0083230799,
-0.0213631559,
-0.004779805,
-0.003669082,
-0.0021972968,
0.0092670992,
-0.018562153,
-0.0118506765,
-0.0417580232,
0.0001106778,
-0.0259526893,
-0.0025411984,
0.025486825,
0.0303173698,
-0.0312083289,
0.0083841933,
-0.0587183982,
-0.0216658022,
-0.0008542722,
-0.0352887921,
0.0113233989,
-0.0604385138,
-0.0166157354,
0.0065351459,
0.0289654341,
-0.0264481511,
0.0820714459,
-0.0093857106,
0.0287569705,
-0.0313888751,
-0.024717465,
-0.0158478133,
-0.0418698639,
0.0481595248,
0.0144565795,
-0.0046298001,
0.0535542406,
0.0126898317,
0.0084534111,
-0.0361193754,
-0.0354536138,
-0.078293696,
-0.0054004714,
-0.031549111,
0.0436783917,
0.0314436406,
0.0133096771,
-0.0166969784,
-0.0215162225,
-0.0617078394,
0.0181836691,
0.0115479818,
0.0416665189,
0.0626801848,
0.0079172561,
-0.0061287447,
-0.0463403799,
-0.0221626647,
-0.0104232002,
0.0082841292,
0.0178305004,
-0.0192398056,
-0.035172645,
0.0038968937,
-0.0338908248,
0.0184638891,
-0.0060586198,
0.0183496699,
-0.0030062851,
-0.036026869,
0.0106490804,
0.0843667537,
-0.0312879905,
0.0605695397,
-0.0160158444,
0.001217912,
-0.044245448,
0.0304706302,
0.0043328917,
-0.0284056142,
-0.0261547528,
-0.0121641001,
0.0137803964,
0.0642514974,
-0.0430143848,
-0.0086979326,
-0.0541662984,
0.0670136064,
0.0309588052,
0.0061365068,
-0.0403005965,
0.0333278738,
0.005959833,
0.0729943737,
-0.0163723193,
-0.0573019609,
0.0519943498,
-0.0027277353,
-0.0440051965,
-0.0140448781,
-0.0322293229,
-0.0824845582,
0.0409497321,
-0.0067526577,
-0.0051068934,
-0.0308690928,
-0.0372498482,
-0.069872275,
-0.0534361042,
0.0512060858,
-0.0391021371,
0.0723247379,
-0.0367052034,
0.0216939934,
-0.0076066437,
-0.0097232955,
-0.0137595143,
-0.0600651093,
-0.0346230939,
0.0215431154,
0.0237444229,
-0.0499102362,
-0.0364845395,
0.0159293823,
0.0027336921,
0.0106667299,
-0.0421009734,
-0.0306577068,
-0.0049592308,
0.041900754,
0.0752595589,
0.0320867412,
0.0221715644,
-0.0430954918,
0.0610519014,
0.031905137,
0.0367774516,
-0.0358892754,
0.0102928681,
0.0094044879,
0.0117279943,
-0.0457295217,
-0.0022122858,
0.019348437,
0.0227159802,
-0.0614811182,
0.0025456883,
0.1058182865,
0.030050341,
0.0120650372,
-0.0187440738,
-0.0388401262,
0.0472962335,
0.0119140241,
0.0948465392,
-0.1117305085,
-0.030854702,
0.0148637285,
0.011671124,
0.0175835695,
-0.0582018569,
-0.0228728186,
-0.1151345819,
0.0587806813,
-0.0165597927,
-0.015687041,
-0.016807735,
-0.0272957981,
-0.0177487321,
-0.0254742727,
0.0098038148,
0.0217315927,
-0.0096840393,
0.019317938,
0.0110578109,
-0.0227505192,
0.0039523123,
0.0017658863,
0.001913041,
0.0154604539,
0.0225824919,
-0.0103576481,
-0.0153590841,
0.0124872439,
0.0093440218,
0.0284304246,
-0.0020451052,
-0.0068484866,
0.0547074117,
-0.0057115592,
-0.0193004105,
0.0081390366,
0.0236063115,
-0.0178153161,
0.0559328087,
0.0047456427,
0.0359582007,
-0.0309586097,
-0.0020478137,
0.0163591281,
0.0191330574,
-0.0197193325,
0.0703885779,
0.0313567519,
-0.0191065874,
-0.0271942969,
-0.0797773749,
-0.0589133948,
-0.0394038558,
-0.001637428,
0.0104178861,
0.0409406126,
-0.0108486963,
-0.0737252012,
-0.0907591581,
0.0751743987,
0.030912688,
-0.0230305959,
0.0357569233,
-0.0079570841,
-0.0053243679,
-0.0321844667,
0.0295290612,
-0.0772069767,
0.0180773214,
-0.0125094773,
0.0004969081,
-0.0063469424,
-0.0275364891,
-0.041056931,
0.038389381,
-0.0739948675,
-0.0525999032,
-0.0118948808,
0.0229782816,
0.0375819542,
0.0071759382,
-0.0310950987,
-0.0063757231,
-0.0149948439,
0.0528078601,
-0.0704244375,
-0.0153670255,
0.0425273404,
-0.0078338245,
0.0013100142,
0.0525300018,
-0.0539445095,
0.0179781523,
-0.0175188612,
0.0306942184,
-0.0263625551,
-0.0165828131,
0.0255806223,
0.0001057363,
-0.0173026863,
0.061404001,
-0.0338435732,
-0.0181045569,
0.0181071572,
-0.0111098895,
-0.0109160459,
-0.0122003779,
-0.0310659874,
0.0058733635,
-0.028193187,
0.0082125962,
0.0174991153,
-0.0207273569,
-0.0100999847,
-0.0408023223,
0.013403154,
-0.0147897471,
-0.0060888561,
0.0208605956,
-0.0171654169,
-0.0008296993,
-0.0145726977,
0.0673450902,
0.0466474965,
0.000352086,
0.040059872,
0.0127466694,
-0.0057230429,
-0.0102825407,
-0.0006227204,
-0.0836171433,
-0.1290493906,
0.004364606,
0.0586676858,
0.0150244124,
-0.0207338892,
-0.0005072,
-0.0216589775,
-0.0207509454,
0.0158203505,
-0.0063015129,
0.0347513705,
-0.0170045141,
0.0363395028,
0.0245670881,
-0.0377277918,
-0.0663141087,
0.1005896851,
0.0132561997,
0.026281517,
-0.0295914523,
-0.0246871896,
-0.0496803895,
-0.0484756902,
0.0241230857,
0.0057532704,
0.0063766236,
0.0214974433,
-0.0606663227,
-0.0923768729,
-0.0126367621,
0.0063036093,
-0.0281740855,
-0.0239586178,
0.0146707408,
0.0693615675,
0.0218293834,
0.0172706824,
-0.0306892637,
0.0296776555,
-0.0123446882,
-0.018547846,
-0.0298171304,
-6.635344553e-33,
-0.0411863104,
-0.0183964968,
0.0599144287,
0.0062277773,
0.0288641043,
-0.0150445644,
-0.0321899168,
0.0035422663,
-0.0302602723,
0.0137185762,
-0.0300391708,
0.0483463146,
0.0210698228,
-0.0035698221,
0.0296591632,
-0.0155951818,
0.0261492766,
-0.0277666599,
-0.0073565696,
-0.0062134741,
0.0020800419,
0.052993875,
0.0577377416,
-0.0167335812,
0.0434963964,
0.0030351412,
-0.0204953831,
0.028771529,
0.0264968742,
0.0500961505,
0.0074393586,
0.0450437218,
0.023159273,
-0.0963830128,
-0.0238670446,
-0.0010910499,
-0.0772999525,
-0.0520385429,
0.0450728945,
0.0284904633,
0.0011400155,
-0.005267961,
0.0396727808,
-0.0279306136,
0.0162987709,
-0.0333371423,
0.0095849363,
-0.0442671478,
0.0258430634,
0.0094000492,
-0.0457977057,
0.0235138461,
-0.0019295809,
-0.0142459823,
0.0169767309,
0.081666939,
0.0035768733,
0.0034714027,
-0.008516456,
0.0173694156,
0.0308156833,
0.0293956902,
0.0382700488,
0.0109972246,
-0.0123067545,
0.0263273213,
-0.0118135223,
0.0108236996,
-0.0067107412,
-0.0021652898,
-0.0032213661,
0.0478311218,
0.0516685694,
-0.015173709,
0.006086322,
0.0000781527,
-0.0430608429,
0.0478596203,
0.055649668,
0.032773871,
0.0121788401,
0.0157589316,
-0.0452670529,
-0.0419143327,
-0.0098103015,
0.0169349369,
-0.039414838,
-0.0278048795,
0.0143934777,
0.004144385,
0.0139321852,
0.0216973517,
0.0204291102,
-0.0323039964,
-0.0410780609,
0.0142491953,
0.0006124027,
0.0106854057,
0.0150102051,
0.0180608127,
-0.0202882886,
-0.0073940093,
-0.0146681722,
0.0048802923,
-0.0041318256,
0.0107272025,
0.0434696376,
0.0005978563,
-0.0793818086,
-0.0062959478,
-0.0442649797,
0.0604872704,
0.0106784189,
0.0389387943,
0.0135695394,
0.0215729121,
-0.0001513594,
0.0357052572,
0.0045086928,
0.0021411274,
-0.0393496342,
0.0180125199,
-0.0234572198,
0.0235902648,
0.0094814235,
0.0056551178,
0.00781989,
0.0172597934,
-0.0000016465,
-0.0336904004,
-0.0101955198,
0.0115585849,
2.915e-7,
0.0499943532,
0.0591021851,
0.0029633958,
0.0617161132,
0.0296817143,
0.032238178,
-0.0224746633,
-0.0055035856,
0.0208241325,
-0.0253891349,
-0.0000101708,
0.0455518216,
0.076916106,
0.0138918022,
-0.0618195273,
-0.0545994528,
-0.044425115,
-0.0037646634,
-0.0550307408,
0.006013948,
0.0241211448,
0.1068233475,
-0.0249747783,
-0.0251515396,
-0.0109527661,
-0.0882938132,
0.05458083,
-0.0296315141,
-0.0408159792,
0.0446864069,
0.0182520859,
0.0487165488,
-0.0059439568,
0.0470499061,
0.0022205089,
0.0048803193,
0.0038654364,
0.049419485,
0.0080729006,
0.0124664959,
-0.0129261361,
-0.0351075083,
0.0050880895,
-0.0650420561,
0.0695420206,
-0.0788718238,
-0.0548272952,
-0.0118119717,
0.0082632238,
-0.0150288483,
0.0280621964,
-0.0199602302,
-0.0073615052,
0.0154461004,
0.0190462619,
-0.020929629,
0.0304527786,
-0.0371215716,
-0.0448062308,
0.0558475517,
-0.0426848941,
-0.0263672303,
-0.0445116311,
-0.0233832877,
-0.0002921108,
-0.0432311259,
0.0082036937,
2.883401776e-34,
0.0824083686,
0.0183696281,
0.0125189377,
0.0825515985,
0.0724213049,
-0.0107456939,
0.0117513807,
0.0408511721,
0.0325071663,
-0.1129340827,
-0.0185493845
] |
http://arxiv.org/abs/2402.01724v1 | http://arxiv.org/pdf/2402.01724v1 | CERM: Context-aware Literature-based Discovery via Sentiment Analysis | 2024-01-27T06:40:08 | 2024-01-27T06:40:08 | Julio Christian Young, Uchenna Akujuobi | cs.CL | cs.CL, cs.AI, cs.LG | Driven by the abundance of biomedical publications, we introduce a sentiment
analysis task to understand food-health relationship. Prior attempts to
incorporate health into recipe recommendation and analysis systems have
primarily focused on ingredient nutritional components or utilized basic
computational models trained on curated labeled data. Enhanced models that
capture the inherent relationship between food ingredients and biomedical
concepts can be more beneficial for food-related research, given the wealth of
information in biomedical texts. Considering the costly data labeling process,
these models should effectively utilize both labeled and unlabeled data. This
paper introduces Entity Relationship Sentiment Analysis (ERSA), a new task that
captures the sentiment of a text based on an entity pair. ERSA extends the
widely studied Aspect Based Sentiment Analysis (ABSA) task. Specifically, our
study concentrates on the ERSA task applied to biomedical texts, focusing on
(entity-entity) pairs of biomedical and food concepts. ERSA poses a significant
challenge compared to traditional sentiment analysis tasks, as sentence
sentiment may not align with entity relationship sentiment. Additionally, we
propose CERM, a semi-supervised architecture that combines different word
embeddings to enhance the encoding of the ERSA task. Experimental results
showcase the model's efficiency across diverse learning scenarios. | CERM: Context-aware Literature-based Discovery via Sentiment Analysis: Driven by the abundance of biomedical publications, we introduce a sentiment
analysis task to understand food-health relationship. Prior attempts to
incorporate health into recipe recommendation and analysis systems have
primarily focused on ingredient nutritional components or utilized basic
computational models trained on curated labeled data. Enhanced models that
capture the inherent relationship between food ingredients and biomedical
concepts can be more beneficial for food-related research, given the wealth of
information in biomedical texts. Considering the costly data labeling process,
these models should effectively utilize both labeled and unlabeled data. This
paper introduces Entity Relationship Sentiment Analysis (ERSA), a new task that
captures the sentiment of a text based on an entity pair. ERSA extends the
widely studied Aspect Based Sentiment Analysis (ABSA) task. Specifically, our
study concentrates on the ERSA task applied to biomedical texts, focusing on
(entity-entity) pairs of biomedical and food concepts. ERSA poses a significant
challenge compared to traditional sentiment analysis tasks, as sentence
sentiment may not align with entity relationship sentiment. Additionally, we
propose CERM, a semi-supervised architecture that combines different word
embeddings to enhance the encoding of the ERSA task. Experimental results
showcase the model's efficiency across diverse learning scenarios. | [
0.0499514751,
0.0881256461,
0.0176644884,
0.0006695863,
-0.0068700374,
0.0222217515,
-0.0346008688,
0.0216843281,
0.0077211484,
-0.0133413412,
-0.0489886329,
0.0000835075,
-0.0264966283,
0.0547954515,
-0.0113652889,
0.0139583517,
0.0464199409,
0.0781383663,
0.0321629085,
0.0224540811,
0.018182775,
0.0197486691,
0.0085534388,
0.0178033393,
-0.0159650128,
-0.0221483391,
0.0385564938,
-0.0080231726,
0.048414804,
-0.0626865923,
0.0358886905,
0.0074397298,
-0.0164930131,
-0.0485330448,
0.000002062,
0.0068668239,
-0.0345522948,
0.0131299915,
0.0124168219,
0.0343573466,
0.0760847628,
-0.0335764848,
0.0128876064,
0.0103251813,
-0.0272395946,
-0.0290333647,
0.03504356,
0.0723737106,
-0.0120288413,
0.0621535406,
-0.0242216308,
-0.0421556905,
-0.0210803021,
-0.0169479251,
0.0247464161,
-0.0112936059,
0.0325448439,
-0.0561703295,
0.0075857481,
0.069963187,
0.004632703,
0.0154894572,
0.0047791051,
0.0038640136,
0.0349876881,
0.0553519093,
0.0444517247,
0.0020036262,
0.0091556953,
0.069823876,
0.0307754427,
0.0059305872,
0.009989325,
-0.0091615869,
-0.0538273081,
-0.0017026935,
0.0343010761,
-0.0602096356,
-0.042643059,
0.0635441765,
0.0406013317,
0.0227924772,
0.042670045,
0.0165762883,
-0.0000285054,
0.0543703735,
-0.0052794782,
-0.0433463864,
-0.0050889426,
0.0175280329,
-0.0163191296,
-0.0086552827,
0.0395328552,
0.0007657006,
-0.0136152552,
0.0306453947,
0.0217727628,
0.0116889235,
0.0418776423,
-0.0632477701,
-0.035075292,
-0.012434463,
-0.0229671784,
0.0228988808,
0.0119615272,
-0.0551568195,
0.0160277784,
-0.0281647909,
-0.0593124181,
0.0316924825,
-0.04527371,
0.0012277993,
0.0056034192,
0.0315947793,
-0.019499151,
-0.0225027092,
-0.0565190502,
-0.0149629805,
-0.0099942014,
0.0088678068,
0.0281721037,
0.0588512793,
0.0179520808,
0.0236931033,
-0.0386609547,
0.0234184731,
-0.0446630381,
0.0269830693,
-0.0533731952,
0.039982494,
-0.0102902362,
-0.0307486337,
0.0009525147,
-0.0038895069,
0.0131971957,
0.044232551,
0.0393947884,
-0.0937436968,
-0.0069898902,
-0.0190724414,
-0.0642151833,
-0.0493013822,
-0.0144694978,
-0.0336309075,
0.0408423953,
0.0174345616,
0.00842804,
-0.0102208648,
0.0297862273,
0.0377407111,
-0.0584022216,
0.0718835294,
-0.0098177027,
-0.038096454,
-0.0116073946,
0.0446775407,
0.0676849782,
0.0248785168,
-0.0021496511,
-0.0085228737,
0.021420626,
0.0162279513,
-0.0447159931,
-0.0525507927,
0.0260652434,
0.0007678816,
0.0006215727,
0.0216130223,
-0.0121738454,
0.0499398075,
-0.0477933697,
0.0504624397,
-0.0101502864,
-0.0345780179,
0.032693591,
0.0314126089,
-0.0153297372,
0.0469771512,
-0.0124295456,
0.0123598427,
-0.0066107395,
-0.0263903197,
0.0159472357,
0.0341406912,
-0.0224098265,
-0.0366877392,
0.0205523521,
0.0082839783,
-0.0156623535,
-0.0519398488,
-0.0179779381,
-0.0184976719,
0.0197310094,
-0.0198598318,
0.0462410711,
-0.0083952639,
-0.0479093343,
-0.0183163676,
0.0597340502,
-0.0525243133,
-0.000664826,
0.0182986967,
0.073745124,
0.0546977334,
0.0391976945,
-0.0489254817,
0.0012112102,
-0.0206860993,
-0.0375755243,
0.0072683524,
0.0058030584,
-0.0399696082,
0.0081634857,
-0.018613141,
-0.0134232352,
-0.0528084859,
0.0220989734,
-0.0492973849,
-0.0061620511,
0.0713349655,
-0.0166708175,
-0.0382235125,
0.0267253332,
0.0107702138,
-0.0080689685,
-0.0165711679,
0.0566974692,
0.010368797,
-0.0015392695,
0.0670531392,
0.0181718897,
0.0500654317,
0.0428144559,
-0.0663033426,
-0.0632203594,
0.0201729145,
0.0060941326,
-0.0404319279,
0.0094076637,
-0.0153629957,
-0.0003569008,
-0.050879702,
-0.0340642035,
-0.0185663309,
-0.0102453781,
-0.0254185479,
0.0141665377,
-0.0104355067,
-0.0218959153,
0.0312876739,
0.0088714585,
-0.0332205258,
0.0480614975,
-0.0442607626,
0.0245611481,
-0.0412671976,
-0.0488031581,
0.0065892963,
0.0422150716,
-0.0404815599,
-0.0411164761,
0.0513603613,
-0.0242364556,
0.0397447236,
0.0055664205,
-0.0296625942,
-0.0165600087,
-0.0945260227,
0.0115203215,
0.0117783183,
0.0188836642,
0.0013992958,
-0.0067431647,
0.0251211673,
0.0115236714,
-0.0278186798,
-0.0067898305,
-0.0582004003,
0.0362867378,
0.0206508171,
0.0313597731,
0.0303623397,
-0.0691947192,
0.0020723972,
-0.0257536937,
0.0224497821,
0.0043201451,
-0.0084658293,
-0.0366891772,
0.0105817821,
-0.0074665318,
-0.0112441899,
0.0096558826,
0.0121274842,
-0.0216024332,
0.0335199796,
0.0007655697,
0.021601025,
-0.0009833972,
0.0632434934,
0.0574017614,
0.0335934609,
-0.0353613533,
0.0327860229,
-0.0893849507,
0.0665282086,
0.0076894197,
-0.046404358,
0.0202061459,
0.043529924,
-0.0050879945,
0.0041314694,
-0.0475110672,
0.0138013232,
-0.0057410859,
-0.0047653117,
-0.0421417505,
0.0172182117,
0.0447985642,
-0.0060915751,
-0.003725159,
-0.0128852436,
0.0775457025,
0.0135838501,
0.0255451519,
-0.0854155719,
-0.0205403306,
-0.0321673155,
-0.0174869131,
-0.0045696744,
0.0279433429,
0.0319090821,
0.0544367805,
0.0145055763,
0.0156961922,
-0.0347681455,
-0.019563172,
0.0591322668,
-0.0150855249,
-0.0159871709,
0.0305775963,
-0.0599273592,
-0.0031050283,
0.0014026368,
-0.0398239642,
-0.0030147396,
0.0438527949,
-0.0073777619,
-0.0221042819,
-0.0199048463,
-0.0250066668,
-0.0457558781,
0.0305660032,
-0.038056694,
0.0188454874,
0.0128728673,
0.0860255435,
-0.0610386431,
0.0086241299,
-0.0028867975,
0.0281200707,
0.0782479793,
-0.0588107817,
-0.0377218835,
-0.0089486316,
-0.01109083,
0.0034872023,
-0.0082655093,
0.0091840839,
-0.0178218521,
0.0088058393,
0.0109112263,
-0.0505316369,
0.0966173261,
0.0565464348,
-0.0547915138,
0.0337237008,
0.0011601379,
-0.0496106185,
0.0196617786,
0.0282631032,
-0.0689589307,
-0.0220165644,
-0.0768614933,
-0.0150819384,
-0.0206799228,
-0.0325832702,
-0.0227633957,
-0.0488471128,
-0.010964117,
-0.0265339892,
-0.0328952856,
-0.0858909339,
-0.102555193,
0.0109198755,
-0.0422461964,
-0.0072129034,
-0.0564986616,
0.0037957823,
0.0224299356,
-0.0134067973,
-0.0215893798,
-0.0367707014,
0.03722528,
-0.0220217071,
0.0319119021,
0.0463688262,
0.0607466772,
0.0126279742,
-0.0285173263,
-0.0212798547,
-0.0421869271,
-0.0082057873,
-0.0326377749,
0.1030256376,
-0.0698559806,
-0.0065322062,
-0.0229350198,
0.0199853051,
-0.0617367327,
-0.0467914678,
-0.005733348,
0.0700159296,
0.0026611809,
-0.0218398925,
-0.0753800198,
-0.0055725733,
0.0733940676,
0.0477881245,
-0.0176594406,
-0.0422360897,
-0.0137215173,
0.0141714653,
-0.0027543688,
-0.003594469,
0.006302895,
-0.0235057436,
-0.0371291637,
-0.0142462337,
-0.0640339628,
-0.0198636111,
0.0166877229,
-0.0087453658,
-0.0148252398,
0.038376417,
0.0324671492,
-0.0234828256,
-0.0583964474,
0.0166650582,
0.0476680771,
0.0403798446,
0.0002056191,
-0.0229562316,
0.0984190926,
0.006147447,
-0.0259313527,
-0.0279418267,
-0.0443747677,
-0.0296450797,
-0.0658914149,
0.030612072,
-0.0068630707,
0.0337782502,
-0.0215821471,
-0.0318125486,
0.035738267,
0.0741673931,
-0.0519064926,
-0.0297935475,
0.0176013671,
0.0424380451,
0.0259808544,
0.0291636381,
-0.0116626872,
0.0400059,
0.0156837404,
0.0222099498,
-0.0369074643,
-0.0711392015,
0.0224440731,
-0.0551121458,
0.0589184277,
0.0030463459,
0.0061955606,
-0.0492590889,
0.0153172612,
-0.0067585516,
-0.0094979526,
-0.0291376542,
-0.0230804104,
0.0076040807,
-0.003194476,
-0.075520508,
0.0349988751,
-0.0299890433,
0.0525659211,
-0.0384046994,
0.0460508466,
0.0215939227,
0.0293211062,
0.0064440463,
0.0124169327,
0.013268494,
-0.0187863875,
-0.0153056709,
-0.0047575547,
-0.0257887803,
0.0440631583,
0.0029587701,
0.0195361711,
-0.032654684,
0.0307814498,
-0.0148312999,
0.0498873815,
-0.0236267727,
-0.0067774248,
-0.0035584799,
-0.0619106367,
0.0305129346,
-0.0328643546,
-0.0062457072,
-0.0165800061,
0.0482802503,
-0.039533399,
-0.0046670348,
0.0931443498,
0.0055093085,
-0.0170074403,
-0.0152391456,
0.0675930902,
-0.0552786961,
0.0856992751,
0.0265768543,
-0.0290164296,
-0.0124036977,
-0.0162816066,
0.0252765007,
0.0265118033,
-0.0279985722,
0.0387331285,
-0.0546558723,
-0.0566031374,
0.0505391844,
-0.0180531237,
-0.0196969453,
-0.0108504202,
0.0266603194,
0.0204429533,
0.0213881247,
0.0207894091,
-0.0341056511,
0.0296436884,
-0.0405773856,
-0.0555849448,
-0.0064352951,
-6.206620458e-33,
0.0158283412,
-0.0356636085,
-0.0062942239,
0.0115327099,
0.0352171361,
0.029884642,
-0.0340561531,
0.0012820944,
0.0202031173,
0.0592241473,
0.0242710393,
0.0300363414,
-0.0191792306,
0.0036384419,
0.0064489185,
-0.0381291881,
0.0333110057,
0.0019166619,
0.022900762,
0.0172259416,
0.015134804,
0.0533272289,
0.0377192385,
-0.1129279882,
0.0181775261,
0.0273113828,
-0.0065016113,
0.0039323359,
-0.0081995921,
0.0484006591,
-0.069098264,
0.0792092457,
0.0187162179,
-0.0021936807,
0.0071747974,
0.0174723696,
-0.0627076402,
-0.0394217446,
-0.0150407627,
-0.0520367101,
-0.0656480715,
-0.0371772647,
-0.0057273032,
0.0406457372,
0.007119474,
0.0295392256,
0.0525900871,
-0.0134953512,
-0.0244186409,
-0.0730486363,
-0.0514351986,
0.0200149547,
0.0064144339,
0.0228788499,
0.0361243933,
0.018849358,
0.0317329541,
0.0304617919,
-0.0088271881,
-0.0136292577,
-0.0449697673,
0.0117819943,
-0.0436880626,
-0.0334280394,
-0.023735052,
0.0647355616,
-0.0228409283,
-0.026666319,
-0.0373206176,
0.0410512052,
-0.0396994837,
0.024479175,
0.0136518404,
0.035239771,
0.0230817664,
-0.0223115031,
-0.0773445889,
0.0384355374,
-0.0013481097,
-0.0203076396,
0.0381688997,
0.025716275,
-0.0016499452,
-0.01353411,
0.0062976214,
-0.0094727445,
-0.0529453792,
-0.0115758684,
-0.0101089366,
0.0254183039,
-0.020992469,
-0.0235937387,
-0.0542621613,
-0.034569554,
-0.0213969946,
0.0464366935,
-0.0203709472,
-0.0192053504,
0.010265938,
-0.0033805941,
-0.1113898382,
-0.001758049,
-0.0024099261,
0.0055073104,
0.0137788607,
-0.0080755642,
0.0504244007,
-0.0098528014,
-0.039604824,
-0.0286378711,
-0.0151839964,
-0.0055919029,
-0.0133668259,
0.0156077016,
-0.0267035756,
0.0396751091,
0.0000614596,
0.100942485,
0.0340660699,
0.0517206565,
-0.0230782181,
-0.0018975292,
0.0424788743,
0.0076397816,
-0.0034837353,
-0.0360197797,
0.0272229724,
0.011997845,
-0.0025293694,
-0.0812122896,
0.003321612,
-0.0230253972,
2.819e-7,
0.0928649157,
0.032332968,
0.0072432705,
-0.0263932142,
0.0209874231,
0.0463019349,
0.0333336666,
0.0418290049,
-0.0206322316,
-0.0143291773,
0.0237653814,
0.0103864148,
-0.0078826481,
0.0362287499,
-0.0341297053,
0.014766315,
-0.0252308547,
-0.0289690867,
-0.0052661835,
-0.0000407096,
0.0728653818,
0.055592332,
0.0373369567,
0.0036041092,
-0.0213840958,
0.0703828931,
0.0565315634,
-0.0716794655,
0.0045145629,
0.0256493725,
-0.0389882587,
0.0265535321,
0.0028684649,
-0.0288126674,
-0.0406419709,
-0.0094275195,
0.0246580448,
0.0518496148,
0.0122634005,
-0.0478369743,
0.0122435987,
0.0252647027,
0.0186544918,
0.0049264841,
0.0662579313,
-0.0783920065,
-0.0472722538,
-0.0229941122,
0.0328231342,
0.0277615096,
0.050465554,
-0.0003222754,
-0.0112571968,
0.0109427953,
0.0082639167,
-0.0029607366,
-0.0032856085,
-0.1020618156,
-0.0093606142,
0.0571585186,
-0.0120770019,
0.007729644,
0.004975047,
0.0211407151,
-0.0004892992,
-0.0545999371,
0.0083228601,
2.609310014e-34,
0.0318571702,
-0.0119111286,
-0.0095163966,
0.0234852023,
0.0026640124,
0.0016761578,
-0.0586677864,
-0.020912325,
0.0251950007,
-0.0485874973,
0.0042992514
] |
http://arxiv.org/abs/2401.15316v3 | http://arxiv.org/pdf/2401.15316v3 | UNSEE: Unsupervised Non-contrastive Sentence Embeddings | 2024-01-27T06:29:07 | 2024-02-02T14:06:31 | Ömer Veysel Çağatan | cs.CL | cs.CL | We present UNSEE: Unsupervised Non-Contrastive Sentence Embeddings, a novel
approach that outperforms SimCSE in the Massive Text Embedding benchmark. Our
exploration begins by addressing the challenge of representation collapse, a
phenomenon observed when contrastive objectives in SimCSE are replaced with
non-contrastive objectives. To counter this issue, we propose a straightforward
solution known as the target network, effectively mitigating representation
collapse. The introduction of the target network allows us to leverage
non-contrastive objectives, maintaining training stability while achieving
performance improvements comparable to contrastive objectives. Our method has
achieved peak performance in non-contrastive sentence embeddings through
meticulous fine-tuning and optimization. This comprehensive effort has yielded
superior sentence representation models, showcasing the effectiveness of our
approach. | UNSEE: Unsupervised Non-contrastive Sentence Embeddings: We present UNSEE: Unsupervised Non-Contrastive Sentence Embeddings, a novel
approach that outperforms SimCSE in the Massive Text Embedding benchmark. Our
exploration begins by addressing the challenge of representation collapse, a
phenomenon observed when contrastive objectives in SimCSE are replaced with
non-contrastive objectives. To counter this issue, we propose a straightforward
solution known as the target network, effectively mitigating representation
collapse. The introduction of the target network allows us to leverage
non-contrastive objectives, maintaining training stability while achieving
performance improvements comparable to contrastive objectives. Our method has
achieved peak performance in non-contrastive sentence embeddings through
meticulous fine-tuning and optimization. This comprehensive effort has yielded
superior sentence representation models, showcasing the effectiveness of our
approach. | [
0.0069260029,
0.0729620978,
0.0185113158,
0.0428103171,
-0.033316832,
0.0125018293,
0.0152347004,
-0.034062624,
0.0173305105,
-0.0321566164,
-0.0175070763,
-0.0115394909,
-0.0102801267,
0.005462368,
0.038701009,
-0.0592149422,
0.0334203281,
0.0240834132,
-0.00729369,
0.0163203347,
0.0268638078,
0.0221302919,
0.018171303,
0.0178804304,
-0.0004806595,
-0.0571427904,
0.0027268475,
0.0236248095,
0.0343422778,
-0.0209980421,
0.05495359,
0.0322735235,
0.0456010178,
0.0055073495,
0.0000018488,
0.0013986777,
0.0031413094,
-0.0265784282,
0.0041061877,
0.0569030568,
0.0716094747,
-0.0735374913,
-0.0001248583,
0.0189476516,
-0.0193059817,
-0.0154777123,
0.0186558794,
0.0735624284,
0.051068943,
0.0894750953,
-0.0194128137,
-0.0622457564,
0.0073391963,
-0.0529367365,
-0.0506457649,
-0.0042429436,
0.0183916967,
-0.0264630914,
-0.0526766926,
0.0446360074,
-0.0224335808,
0.0265713409,
0.0379470512,
0.0147309992,
0.0242735073,
0.0299840756,
-0.0277915727,
0.0356689915,
-0.0086361645,
0.0838824809,
-0.0157263242,
0.003623951,
0.0036365513,
-0.0137645956,
-0.0139041487,
-0.0412582457,
-0.0162361823,
0.0146516245,
0.0400705449,
0.0597619824,
0.0413673557,
0.0635343716,
0.0431032404,
-0.0050684069,
-0.007491665,
-0.0515953526,
0.0051657013,
-0.0256887265,
-0.0071908059,
0.0000801576,
-0.0607206896,
0.0192360524,
0.0337732993,
-0.0369386524,
-0.0145450663,
-0.0199220777,
-0.0110653434,
-0.0381794162,
0.0135019934,
-0.043915499,
0.0568909347,
0.0200002678,
0.0091427993,
0.0526359603,
-0.0221031867,
0.0230094623,
-0.0259279944,
0.0060459557,
-0.0557906777,
-0.0072927307,
-0.0279026907,
-0.0089059463,
-0.0187652186,
0.0375896655,
-0.0618317835,
-0.0559751391,
-0.0147255147,
0.0081416247,
-0.0179569963,
0.0079634301,
0.0203409176,
0.041984845,
-0.0308689382,
-0.0063488176,
0.0200649407,
0.0336008891,
-0.0439870991,
-0.0088465344,
-0.0009801273,
-0.0212486796,
-0.045554366,
0.051296182,
0.0148579739,
-0.0205939598,
0.0066481926,
0.0551766306,
-0.015951883,
-0.0394554101,
-0.0284318496,
-0.0020742882,
-0.0173094738,
-0.054161191,
-0.0136019289,
-0.040129371,
0.0273637883,
-0.0282378532,
-0.0039971499,
-0.0303317569,
-0.0012839503,
0.0185184404,
-0.0197725836,
0.0500578247,
0.0375358164,
-0.0581477508,
-0.0540778227,
0.017586723,
0.0435113646,
0.0743974522,
-0.0132992137,
0.0014398097,
0.0129333464,
0.0296547357,
-0.0112376278,
-0.0063954182,
0.0084327348,
0.0059165112,
-0.0110764373,
0.0354880244,
0.0276161116,
0.0440628827,
0.0364927761,
0.0427331664,
-0.0288480688,
-0.0577810332,
0.0294988789,
0.0014650737,
0.0280185658,
0.044694569,
0.0422274061,
0.0423051156,
-0.0561829545,
0.0283791088,
-0.0139714805,
0.0100594871,
0.0090251165,
0.00712563,
0.0433775112,
0.022789903,
-0.0587378591,
-0.0111101605,
-0.0089819478,
-0.0384322219,
0.0007779279,
-0.0205543283,
0.0125802048,
0.0258326642,
-0.0625175908,
0.0254500378,
0.0019450904,
-0.037865106,
-0.0423271433,
0.036391221,
-0.0086602466,
0.0600234978,
0.0452178158,
-0.0226190481,
-0.0077289669,
-0.0258875266,
0.016421577,
0.0718870386,
0.033019539,
-0.0080223829,
0.0633251444,
-0.0744613111,
-0.0183418989,
-0.0502483547,
0.0348081701,
0.0007699828,
-0.02019611,
-0.0100406287,
0.0434232578,
-0.0218570493,
0.010997124,
0.0142241577,
-0.0183626637,
-0.0309062842,
0.0315598957,
0.0123071503,
-0.0064207558,
0.0136856372,
-0.0283827353,
0.0041287476,
0.0173062179,
-0.0741257891,
-0.0193897523,
-0.0312293917,
-0.1091513708,
-0.1133868322,
0.0088185491,
-0.0918286145,
-0.0286439899,
0.1162883118,
-0.0643918142,
-0.0226996113,
0.0247922391,
-0.0486763641,
0.0099497437,
-0.0453937091,
-0.0389883183,
0.0187095627,
0.0473463908,
-0.0063623767,
0.0751752257,
-0.0341439433,
-0.0002321215,
0.0054332661,
0.0122379521,
0.0267847236,
-0.0501552373,
-0.0293753669,
-0.0256684385,
0.0256822389,
0.0304492265,
-0.0167077202,
-0.019495409,
-0.010823465,
-0.0056045069,
-0.0810213089,
0.0138289761,
-0.0371869206,
-0.0014902676,
0.0351389721,
0.00759415,
0.0155753437,
-0.0084478548,
-0.0433335975,
-0.026874464,
-0.0341618694,
0.0214869268,
0.051308699,
-0.0384986736,
0.0137294456,
-0.0291027706,
-0.0246102996,
-0.0267605633,
0.0770317093,
-0.0614124797,
0.0158073772,
-0.0251514819,
0.0527056232,
0.0023074625,
-0.0080378475,
0.0017045288,
-0.0210807361,
-0.0312924273,
-0.0189144351,
0.0523310751,
0.0703179464,
0.0156596899,
0.0266022291,
0.0139876092,
0.0235816427,
-0.0197516121,
-0.0006544027,
-0.0128392121,
0.0855713561,
-0.0026259162,
-0.0202817693,
0.0091963746,
0.0928867757,
0.0131426062,
-0.010349826,
-0.0491209999,
0.0103533054,
0.0102454349,
0.0172129869,
-0.037219435,
0.0181292221,
-0.0053602252,
-0.0425714329,
-0.0446026176,
-0.0017008833,
0.0450982377,
-0.0009082783,
-0.0109710479,
-0.067703709,
-0.0437740721,
-0.0363478139,
0.019610092,
0.0113575682,
0.0210633315,
0.0269422587,
-0.011711251,
-0.0162949171,
-0.0412361696,
0.0038557597,
0.0085958512,
0.0025248409,
0.0360436291,
-0.0168723837,
-0.0301043261,
-0.0451548472,
-0.0060500167,
-0.1106669977,
-0.0275593773,
0.0298364721,
0.0661794469,
0.0049660285,
-0.0282207057,
-0.0258022361,
-0.0155022908,
0.0401753448,
-0.0153242266,
0.0014467065,
-0.0446994565,
0.0709920079,
0.0590657629,
0.010862262,
0.0022335378,
-0.0483965911,
0.0479771942,
0.0595036,
0.0118760858,
0.008463054,
-0.0579385385,
0.0101883188,
-0.0131510329,
-0.0072261589,
-0.0050244955,
-0.0070728189,
0.0214591511,
-0.0051520625,
-0.0100245429,
0.1058033332,
0.0116841281,
-0.0317162015,
0.0095634786,
0.0191653948,
0.046314422,
0.0158717837,
-0.0066478057,
-0.0636608824,
-0.0469365045,
-0.0216398537,
-0.0098727727,
0.0372786783,
-0.0246476494,
-0.0120986756,
-0.060577888,
0.0103428187,
0.0271165017,
-0.0041707521,
-0.0328777544,
-0.0050020004,
-0.0213814415,
-0.0265200082,
0.0220272187,
-0.0125246579,
-0.0094956644,
-0.0182224158,
0.0022544803,
-0.0392395556,
-0.05769163,
-0.0019574231,
-0.0441177227,
0.0159262419,
0.0265870709,
0.0205704831,
0.0117459763,
-0.0175038949,
0.039000608,
-0.0608865619,
-0.0169878285,
-0.0046251048,
0.066709362,
-0.0391420051,
-0.0122368932,
-0.0381937735,
0.0228076931,
-0.0568874888,
-0.0226130746,
-0.0038486556,
0.0725509003,
0.0284470599,
-0.0225621555,
-0.0301044229,
0.0235432237,
0.0038552885,
0.0649130866,
0.0348474644,
-0.0609814338,
-0.0420700088,
-0.0210079961,
0.0196482744,
-0.0645727813,
-0.0343527906,
0.00608443,
0.0581207015,
0.0223595332,
-0.0414389707,
-0.0897030979,
-0.0318987034,
0.0190105382,
-0.0365384482,
-0.0024860164,
-0.0528356694,
-0.0335504971,
-0.0234760884,
0.0277496558,
-0.03648182,
0.0019131278,
0.0137525685,
0.026656175,
0.016991267,
0.0098629678,
-0.0294009931,
-0.0356865674,
-0.0729417801,
-0.0447769426,
-0.0423311554,
0.0038459741,
-0.0037313537,
0.0160017144,
-0.0004730814,
-0.0071229758,
0.0947401449,
0.0376357958,
-0.0225646738,
0.018441923,
0.0649300441,
-0.0200705864,
0.0208773129,
0.0565921552,
0.0130495895,
0.012791873,
0.0134644788,
0.0553066246,
-0.0082950518,
0.0269467179,
-0.0001403255,
-0.0523954742,
0.0484338254,
0.0193079822,
0.0008399196,
-0.0293937661,
-0.0078490647,
-0.0267414004,
-0.0204329137,
-0.0479951724,
0.0353902951,
0.0363686271,
0.0269985348,
-0.0355170295,
0.0528912358,
-0.0088890977,
-0.001584777,
-0.0065524345,
0.0113890171,
0.0321420915,
-0.0208845083,
-0.0184320435,
0.0196444988,
-0.0471861921,
-0.0087573007,
-0.0319736153,
0.0040674163,
-0.0048584072,
0.0222985279,
0.010122926,
0.0037375833,
-0.0338694565,
0.0279550403,
-0.0447239503,
0.0604428723,
-0.0378183797,
-0.0296895336,
0.0724948347,
-0.0485093556,
0.0171235725,
0.0054095285,
0.0004444473,
-0.0329519846,
0.0346239917,
-0.0193139669,
-0.0005367471,
-0.0080748191,
0.0165630225,
-0.0449582525,
0.0196229257,
0.0619651228,
-0.0116570182,
0.014990584,
0.0314566419,
0.00564946,
-0.057608787,
-0.0100382371,
0.0223646015,
0.0581890196,
-0.0035472466,
-0.0025640973,
-0.001625888,
-0.0487748981,
0.0342190452,
0.0246599391,
-0.0256846994,
0.0155385286,
0.0341245346,
-0.0127899591,
-0.0049297526,
-0.0030295823,
-0.025794724,
0.0176100153,
0.0017898467,
-0.0214136075,
-0.0254160166,
-6.597652327e-33,
0.0224648472,
0.0267875716,
-0.0486060642,
-0.0542121865,
-0.0106704505,
0.0368542336,
-0.0431372933,
-0.0152943516,
0.0031404831,
0.0276395772,
0.0104408385,
0.0290602706,
-0.0046097408,
0.0661329478,
-0.0313747339,
-0.0123457424,
0.0550616495,
-0.0080666831,
-0.0152488584,
0.00868819,
-0.0053219432,
0.0412173234,
0.039241977,
-0.0881632641,
-0.0120230028,
-0.017357843,
0.0075965622,
0.0197115373,
0.0428217277,
0.0634943098,
-0.039209608,
0.0606195629,
0.0260239616,
-0.0308818035,
-0.0290240888,
0.0276162904,
-0.0515918806,
-0.0229762942,
-0.0231846981,
-0.0093875257,
-0.0816377625,
-0.0594587959,
0.0138143469,
-0.0016933816,
0.03901539,
-0.056406375,
0.0309600048,
0.0027898434,
-0.0053562513,
0.025246229,
-0.0336844027,
0.0405799858,
-0.004156535,
0.008036959,
0.0069468063,
0.0442430004,
0.0342601761,
0.0335975587,
-0.0419539623,
-0.0245188382,
-0.0129034184,
0.0294108093,
-0.0078389803,
0.0010721348,
-0.0085938359,
0.0675262213,
0.0136526851,
-0.0039743003,
-0.0192180872,
-0.0396743752,
-0.0187090263,
0.0240493696,
0.0605882071,
0.0632207766,
0.0243496373,
0.0142712723,
-0.0867526457,
0.0027243325,
0.0009141897,
-0.0324863531,
0.0455354862,
0.022618372,
-0.0245278068,
-0.0133441957,
0.0169520155,
-0.0361146405,
-0.018656509,
-0.0146965571,
0.0373450369,
0.0325320065,
-0.0118186651,
0.0376456343,
0.0144254453,
-0.060363356,
0.0072756088,
0.0530500598,
0.0034835816,
0.0006981755,
0.0422698893,
0.0140434382,
-0.0762907714,
0.0023636306,
0.0043683401,
0.0334948227,
0.0019945607,
-0.0144833419,
0.0475138277,
-0.0000749607,
-0.0687883049,
-0.0266313236,
-0.001404301,
-0.0385711268,
-0.050021898,
0.0131251859,
0.0356545299,
0.0069255023,
-0.0222191103,
0.0757827684,
0.0178464036,
0.0171403512,
-0.024602687,
0.021464048,
0.0065461132,
0.0532424673,
0.0215179585,
-0.0018314524,
-0.0314135589,
0.0318969078,
0.0382358134,
-0.0615825616,
-0.0111224642,
0.0005942044,
2.319e-7,
0.015224006,
0.0024700484,
0.0450180024,
-0.0110168718,
0.0446578749,
0.0670681745,
0.0118945036,
0.069531709,
-0.0611148737,
0.0198903773,
0.0373729803,
0.0333757438,
0.0490468256,
-0.0005846284,
-0.0978882015,
0.0824321806,
-0.0842330977,
-0.0046069063,
-0.0227024574,
-0.0030163904,
0.057798624,
0.1241447181,
0.0027230999,
0.0130686034,
0.0152099002,
-0.0106474357,
0.0815744773,
-0.0189534836,
-0.0146046979,
-0.0274503082,
-0.0383091718,
0.0289327502,
-0.0267972983,
-0.0293196756,
0.0037437684,
0.069383502,
0.0176523142,
0.0691397265,
0.023699129,
-0.046365682,
0.0119371088,
0.0169417281,
-0.0127295218,
0.01888844,
0.0697545037,
-0.0557872355,
-0.0262928996,
-0.0294401124,
-0.0348650105,
0.0215020329,
0.0411325432,
0.0302793067,
-0.0447329432,
0.0165869147,
0.0194290206,
-0.0214489829,
-0.0249390062,
-0.0865331218,
-0.0203331001,
0.0254639685,
-0.0517648421,
0.0060967882,
-0.0264094863,
0.0475648157,
0.0605836771,
-0.0425011776,
0.0122434609,
2.298616755e-34,
0.029814465,
-0.023347998,
0.0254937597,
-0.0102306018,
0.0014543824,
-0.0352263749,
-0.0260743462,
-0.0659179688,
0.0609112792,
-0.0653364807,
-0.0166398417
] |
http://arxiv.org/abs/2401.15312v1 | http://arxiv.org/pdf/2401.15312v1 | How We Refute Claims: Automatic Fact-Checking through Flaw Identification and Explanation | 2024-01-27T06:06:16 | 2024-01-27T06:06:16 | Wei-Yu Kao, An-Zi Yen | cs.CL | cs.CL | Automated fact-checking is a crucial task in the governance of internet
content. Although various studies utilize advanced models to tackle this issue,
a significant gap persists in addressing complex real-world rumors and
deceptive claims. To address this challenge, this paper explores the novel task
of flaw-oriented fact-checking, including aspect generation and flaw
identification. We also introduce RefuteClaim, a new framework designed
specifically for this task. Given the absence of an existing dataset, we
present FlawCheck, a dataset created by extracting and transforming insights
from expert reviews into relevant aspects and identified flaws. The
experimental results underscore the efficacy of RefuteClaim, particularly in
classifying and elucidating false claims. | How We Refute Claims: Automatic Fact-Checking through Flaw Identification and Explanation: Automated fact-checking is a crucial task in the governance of internet
content. Although various studies utilize advanced models to tackle this issue,
a significant gap persists in addressing complex real-world rumors and
deceptive claims. To address this challenge, this paper explores the novel task
of flaw-oriented fact-checking, including aspect generation and flaw
identification. We also introduce RefuteClaim, a new framework designed
specifically for this task. Given the absence of an existing dataset, we
present FlawCheck, a dataset created by extracting and transforming insights
from expert reviews into relevant aspects and identified flaws. The
experimental results underscore the efficacy of RefuteClaim, particularly in
classifying and elucidating false claims. | [
0.0574361533,
0.0384829417,
-0.0009991501,
0.0211813841,
-0.0529642999,
0.0123090381,
0.0519164428,
-0.0145512223,
-0.0293029901,
-0.0480902828,
0.005654823,
0.0098551549,
0.0399605371,
0.0989505649,
0.0115829529,
-0.0029034165,
0.0270423554,
0.0177580714,
0.0637750775,
0.0041076718,
0.0163309332,
0.0301577616,
-0.0530820005,
0.0127496058,
-0.0442162119,
-0.0117259296,
0.0418709517,
-0.0046114554,
0.0403648093,
-0.0466858968,
-0.0049198642,
0.040422298,
-0.0050666607,
0.053566467,
0.0000018877,
-0.0273486897,
-0.0004475432,
0.0342735276,
-0.0183099508,
0.0301988721,
0.0109572541,
0.0022432276,
0.0009148901,
0.0362344272,
0.0212592613,
-0.0277673118,
0.0106310854,
0.0979911312,
0.0836726204,
0.1131036282,
-0.0177610591,
0.0177356992,
0.0074042743,
-0.0005792077,
-0.0274920072,
-0.0443774611,
0.0171612632,
0.0276169591,
-0.0608839989,
0.0486142747,
0.0449430607,
-0.0232402161,
0.0057956185,
-0.0422178805,
0.0061548394,
0.0117991362,
-0.012302299,
-0.0258811302,
0.0074241273,
0.0323454849,
0.0455203652,
-0.0712205917,
0.0224854629,
-0.0007785609,
0.0217837915,
0.0716427118,
0.0152988825,
0.0403320417,
-0.0453716964,
-0.0011347109,
0.0193046145,
-0.0091017755,
0.0374665558,
0.0246065184,
0.0584900081,
0.0316006877,
-0.010723778,
-0.0485887602,
-0.0340726338,
0.0104590133,
0.0357373841,
-0.052521646,
-0.0113167074,
-0.0174169708,
-0.0012474,
0.0136999693,
-0.0097382283,
0.0149926655,
0.0553341545,
-0.046824839,
-0.0039174147,
-0.030473372,
-0.0495203733,
0.0560009815,
-0.0106225321,
-0.0114084212,
0.0185393039,
-0.0162470452,
0.0143177174,
0.008286831,
-0.0181393679,
0.0265527088,
0.020412771,
0.0215876494,
-0.0120710945,
0.0262668729,
-0.0124282753,
-0.0104943523,
-0.0297331307,
0.069965072,
-0.0060831597,
0.0124151176,
-0.0024792019,
-0.0572032891,
-0.0709507689,
0.0415243767,
-0.0013522155,
0.0457976162,
-0.0200522039,
-0.0043721898,
-0.0245232545,
0.0116704404,
-0.0178191476,
0.0183905326,
-0.0056911684,
-0.0021313371,
0.0464257486,
0.0153520945,
-0.0546799526,
-0.0558401421,
-0.0304562785,
-0.0429724231,
-0.0208199061,
-0.0686535835,
-0.0037301008,
-0.024595391,
-0.0185888987,
-0.0034640005,
-0.0030440171,
0.0336697847,
-0.0270751733,
0.037500374,
-0.0921880901,
-0.0051077991,
0.0344077982,
0.0122934766,
-0.0029107542,
0.0303159337,
-0.0344984606,
0.0298145339,
0.0453490391,
0.044599928,
0.0246545039,
-0.0380617455,
0.0485664085,
0.0142417308,
0.06815584,
-0.0115297372,
-0.0281465184,
-0.0264413804,
-0.0283838678,
0.0080318842,
-0.0250797383,
-0.0354742818,
0.0708103031,
-0.0507278107,
-0.032659661,
0.020780623,
0.0370217562,
-0.0157151725,
0.0053705689,
-0.0004467205,
-0.0487859026,
0.0269825291,
-0.0462730676,
-0.0409197547,
-0.0122253243,
0.0577758811,
-0.019940529,
-0.0593066327,
-0.0400045589,
-0.0046673412,
-0.0530439951,
-0.0031138773,
0.0293822736,
0.04058595,
-0.0277981758,
-0.0009964502,
0.0453863181,
-0.0235647876,
-0.0157518536,
-0.0311432984,
0.0382683873,
0.0378588401,
0.0357347466,
-0.0059463368,
-0.0070101935,
-0.0294704605,
0.0139136426,
0.0152655076,
0.0687494129,
-0.0485384725,
-0.0418960042,
0.0064376346,
-0.0151749561,
-0.0535981506,
0.0282074362,
-0.0210913178,
0.0278128814,
0.0057164542,
0.0147116184,
-0.0378880762,
-0.0186887234,
0.0113153476,
0.0071416209,
0.0207997039,
0.0140479384,
-0.0029652463,
0.0022222667,
0.0307759587,
-0.0047798818,
-0.0089886291,
0.031167008,
0.0123701356,
-0.0396108404,
0.009186415,
0.0554247573,
-0.020186035,
0.0302043799,
-0.0790695176,
-0.0206186567,
0.0218525827,
-0.0259755999,
0.0010146069,
0.0035727331,
0.0022889401,
-0.0199340954,
-0.0278336704,
-0.0158408601,
0.0001728698,
-0.0133740297,
0.0162926782,
0.0698163882,
-0.0189384222,
0.0001888872,
-0.0283316299,
-0.005263946,
0.0082450537,
0.0063858838,
-0.0136290099,
-0.0191479102,
-0.0142961433,
-0.0011508137,
0.0337101258,
-0.0196983423,
-0.0142684393,
0.017169062,
-0.0871907473,
-0.0238310024,
0.0068132179,
0.0149411093,
0.0618745126,
-0.036144305,
0.0207862426,
0.0311288759,
-0.0236385409,
-0.030756047,
0.0003096978,
0.0091496343,
0.0528495722,
0.0059580528,
-0.0138428826,
-0.0160583314,
-0.0095924903,
-0.0298330281,
0.023347877,
0.0590458214,
-0.0342554338,
-0.054097876,
0.026047308,
-0.0519703142,
0.0047353762,
-0.0326956511,
0.0037031563,
0.0072157639,
0.0342026018,
-0.0641023815,
-0.0093375649,
0.0207218658,
0.0533423536,
0.0217909906,
0.0351667851,
0.0133346189,
-0.0218831748,
-0.0268712416,
0.06544815,
0.0013766114,
-0.0589914508,
0.0532730892,
0.0302339308,
0.0131707611,
0.0039734929,
-0.0050845067,
0.0688220188,
-0.0294091906,
-0.0030721771,
0.0635748506,
-0.0434015542,
0.0158617944,
0.0477987267,
-0.0291332789,
-0.0114363329,
0.0864025652,
-0.0301675443,
0.0686263815,
-0.0346737839,
-0.0030615013,
-0.0267441496,
0.0081986748,
-0.030252872,
0.0476162992,
0.0390158333,
0.0237688608,
0.0235001165,
0.0065992433,
0.1013336256,
0.0030291486,
-0.011870618,
-0.0615699328,
0.0166402329,
0.0370021649,
-0.0488778502,
-0.021913996,
-0.0933615118,
-0.0091121085,
-0.0358835235,
-0.0501829721,
-0.0043843878,
-0.063386187,
-0.0245111492,
0.0159241129,
0.0099618733,
0.0031790023,
-0.0122524695,
0.038422782,
0.0046947356,
0.0169358458,
0.0049923155,
0.0732394233,
0.0113999639,
0.0009615449,
0.076868929,
-0.03819561,
0.0178530682,
0.0008000553,
-0.0511331484,
-0.0453434736,
-0.028616704,
0.0045598135,
-0.0359855965,
0.032691285,
-0.0603510253,
-0.0092221033,
0.0509474874,
0.0214873757,
0.0053944248,
-0.004432254,
0.0433704481,
-0.0171041004,
0.026006626,
-0.0191338621,
-0.1122302711,
0.0159221254,
-0.0205449872,
0.012805324,
-0.0231832303,
-0.0104239536,
-0.0056526144,
-0.098099269,
0.0401577763,
-0.0170393009,
0.0240036007,
-0.0779418498,
-0.0624028891,
0.0032468904,
0.0041620638,
-0.0062890835,
0.0262466036,
0.0443100855,
0.0047824588,
-0.0061959061,
-0.0416576378,
-0.0395049565,
0.0439224169,
0.0062410594,
0.0642978773,
0.0095128659,
0.0136924414,
-0.0055808434,
-0.0145895416,
0.0061227586,
-0.008694822,
-0.0444207899,
-0.0118322084,
0.1129412949,
-0.0136470608,
0.0003946983,
-0.0021216583,
0.0344533697,
0.0021364775,
-0.0284207519,
-0.0099287201,
0.0581724606,
0.0079136835,
-0.0386416055,
-0.0254986882,
-0.0001216496,
0.0082495855,
0.0643144324,
-0.0024203546,
0.015825469,
-0.031512551,
-0.0143129686,
-0.0406514741,
-0.047016196,
-0.0297975615,
0.0442591868,
0.0527630299,
-0.0476511046,
-0.090159677,
-0.0061586997,
-0.0286892485,
0.0293735508,
-0.0007172997,
-0.0675471202,
0.0160372034,
0.0129879238,
-0.028550107,
0.0098427469,
0.0007263145,
0.0256093126,
0.008774234,
-0.0172421094,
0.041156169,
-0.0191814471,
0.0145905502,
0.0420472473,
-0.0471947752,
-0.0366935544,
-0.0259267148,
0.0249455329,
0.0122147286,
0.064610675,
-0.0414948016,
0.0010086666,
0.0214117374,
0.0049608666,
-0.0532727689,
-0.0168496966,
0.0356111191,
-0.0246322379,
0.0332827866,
0.0052998941,
-0.0230593029,
0.0226459671,
0.0040777735,
-0.006053675,
-0.0403885879,
0.001665268,
0.0485924482,
0.0211775973,
0.0599300526,
0.0105525311,
0.0481485873,
-0.0465408601,
0.0130819557,
-0.0455747433,
-0.0434063002,
0.0054459283,
0.057815887,
-0.0312575214,
-0.031915009,
-0.0056154318,
0.0270716492,
-0.0323880389,
0.0334740952,
-0.0247077812,
-0.004565198,
0.0383702591,
0.0082176477,
-0.0015144044,
-0.006034466,
-0.0388398394,
-0.013036808,
0.0137611208,
0.0127551015,
0.0155153023,
-0.0068206047,
0.0119791152,
-0.0157232229,
-0.0043800161,
-0.0448626839,
-0.0516945459,
0.0086097559,
-0.0253720023,
-0.0046538403,
-0.0128400838,
0.011114927,
0.0380095169,
-0.0042521497,
-0.0426729731,
0.0158431455,
0.0005899682,
-0.0246006288,
0.007543915,
0.0197775923,
0.0312649645,
-0.0530694313,
0.0175328199,
0.036535278,
-0.0290268548,
0.0394007675,
0.0124469604,
-0.0606389344,
-0.0257576406,
-0.03913128,
-0.0864700526,
0.0398161858,
0.0154774887,
-0.0376924314,
-0.1060272083,
-0.0453411117,
0.0078031067,
0.0363729075,
0.0828884244,
0.0547392108,
-0.0287959781,
0.0785523504,
0.0587900542,
-0.000471661,
-0.0354188085,
-0.018795168,
-0.0004188126,
-0.0279615521,
-0.0041741431,
-6.270569555e-33,
-0.0079139033,
-0.079425931,
-0.0097963726,
0.0089138653,
0.0221210625,
0.0110385064,
-0.0425989144,
-0.0156872366,
-0.0492558926,
0.080219835,
0.0325046964,
0.0417506061,
0.0004344323,
-0.0085604731,
-0.0179594215,
-0.0010220207,
0.0675071925,
0.0013034726,
-0.0228706915,
0.0531539358,
0.048018489,
-0.0264197011,
0.0477991253,
-0.0873413384,
0.0514311306,
0.0144591304,
-0.052011922,
0.0341099203,
-0.0351978391,
0.0558876507,
0.0176619831,
0.0361039229,
0.0385425575,
-0.0146234185,
0.0081034722,
-0.057765644,
-0.0749036595,
-0.0047704442,
0.0220442545,
-0.0326254033,
0.0343182273,
0.0259049367,
0.0202010907,
0.0079027126,
-0.001705124,
-0.0467020683,
0.0529093444,
-0.0371420048,
0.0202446524,
-0.0700622723,
-0.0515443943,
0.0064184736,
-0.0237150826,
-0.0012608927,
-0.0308924802,
0.0321530364,
0.0235046353,
-0.0469236225,
-0.0493457839,
0.0444725566,
0.0068247193,
-0.0025409218,
0.0027133364,
-0.0594839267,
-0.0397064909,
0.059662696,
0.0027656013,
0.0435528383,
-0.0539354309,
0.0584598891,
-0.0230310075,
0.0477798991,
0.0295221861,
0.0807232633,
0.0012339577,
-0.0630259439,
-0.092205368,
0.0088615175,
-0.0148753719,
-0.0046079601,
0.020512592,
-0.0674897358,
-0.033899013,
0.0115621891,
-0.0036485966,
-0.0361997597,
-0.0416682735,
0.0162776709,
-0.0182049628,
0.0436542891,
-0.0357612558,
0.0296191126,
-0.0288900379,
-0.0677408427,
0.0226587411,
0.0165074207,
-0.0368714295,
0.0018159301,
0.0117716622,
-0.0227463357,
-0.0621630996,
0.0250420216,
-0.0327555798,
0.0539665893,
0.0079111224,
-0.0243735276,
0.0738683715,
0.005735449,
-0.0494813919,
0.0285344645,
-0.0424549282,
0.0368780978,
0.0108343074,
0.011940795,
-0.0181891657,
-0.0086063994,
-0.0320715532,
0.0159753207,
-0.0207558107,
0.0594895259,
-0.061428085,
0.0657418072,
-0.017706899,
0.0147929508,
-0.0065264255,
-0.0242517591,
-0.0007868367,
0.0106122484,
0.0099414345,
-0.0571992472,
-0.0031237123,
0.0090160836,
2.711e-7,
0.0108125154,
0.0429733433,
0.0242384654,
-0.0371568352,
0.0548227876,
0.0477263369,
0.0358514562,
0.0124711189,
-0.0135795763,
-0.0937396362,
0.0162425824,
0.0129384063,
-0.0090547716,
0.0540331043,
-0.0494275913,
0.0149550596,
-0.0265393015,
-0.0313460529,
-0.0512182042,
-0.012131664,
0.0096148578,
0.0169957429,
0.0375128388,
-0.0336400121,
0.0141251842,
0.0230982658,
0.0506014563,
0.0451854616,
-0.0605598837,
0.0233037528,
0.0387993827,
-0.0088993497,
0.00119989,
-0.013326291,
-0.0436439365,
0.0110854758,
-0.0395316891,
0.0124131273,
0.0155408084,
-0.0377189554,
-0.0322686359,
-0.0070927958,
0.0546376146,
-0.0131001659,
0.0494879447,
0.0256572738,
-0.0198122393,
-0.0935115591,
-0.0379675739,
0.0199979972,
0.0167209078,
-0.0044290721,
-0.014710458,
0.0387806073,
-0.0414386317,
0.0001263751,
0.0071168127,
-0.0452546366,
0.0046394784,
0.0598090366,
-0.0444072001,
-0.0457169227,
0.0951278508,
-0.0515858382,
0.0219047088,
-0.0175715778,
0.0345712081,
2.388592123e-34,
0.0332489014,
-0.0150913354,
-0.010895133,
-0.0235215891,
0.0493853651,
0.0239674244,
-0.0578255057,
-0.0314668603,
0.0138533451,
0.0250331368,
-0.0108286915
] |
http://arxiv.org/abs/2402.01723v1 | http://arxiv.org/pdf/2402.01723v1 | An Empirical Study on Large Language Models in Accuracy and Robustness under Chinese Industrial Scenarios | 2024-01-27T03:37:55 | 2024-01-27T03:37:55 | Zongjie Li, Wenying Qiu, Pingchuan Ma, Yichen Li, You Li, Sijia He, Baozheng Jiang, Shuai Wang, Weixi Gu | cs.CL | cs.CL, cs.AI | Recent years have witnessed the rapid development of large language models
(LLMs) in various domains. To better serve the large number of Chinese users,
many commercial vendors in China have adopted localization strategies, training
and providing local LLMs specifically customized for Chinese users.
Furthermore, looking ahead, one of the key future applications of LLMs will be
practical deployment in industrial production by enterprises and users in those
sectors. However, the accuracy and robustness of LLMs in industrial scenarios
have not been well studied. In this paper, we present a comprehensive empirical
study on the accuracy and robustness of LLMs in the context of the Chinese
industrial production area. We manually collected 1,200 domain-specific
problems from 8 different industrial sectors to evaluate LLM accuracy.
Furthermore, we designed a metamorphic testing framework containing four
industrial-specific stability categories with eight abilities, totaling 13,631
questions with variants to evaluate LLM robustness. In total, we evaluated 9
different LLMs developed by Chinese vendors, as well as four different LLMs
developed by global vendors. Our major findings include: (1) Current LLMs
exhibit low accuracy in Chinese industrial contexts, with all LLMs scoring less
than 0.6. (2) The robustness scores vary across industrial sectors, and local
LLMs overall perform worse than global ones. (3) LLM robustness differs
significantly across abilities. Global LLMs are more robust under
logical-related variants, while advanced local LLMs perform better on problems
related to understanding Chinese industrial terminology. Our study results
provide valuable guidance for understanding and promoting the industrial domain
capabilities of LLMs from both development and industrial enterprise
perspectives. The results further motivate possible research directions and
tooling support. | An Empirical Study on Large Language Models in Accuracy and Robustness under Chinese Industrial Scenarios: Recent years have witnessed the rapid development of large language models
(LLMs) in various domains. To better serve the large number of Chinese users,
many commercial vendors in China have adopted localization strategies, training
and providing local LLMs specifically customized for Chinese users.
Furthermore, looking ahead, one of the key future applications of LLMs will be
practical deployment in industrial production by enterprises and users in those
sectors. However, the accuracy and robustness of LLMs in industrial scenarios
have not been well studied. In this paper, we present a comprehensive empirical
study on the accuracy and robustness of LLMs in the context of the Chinese
industrial production area. We manually collected 1,200 domain-specific
problems from 8 different industrial sectors to evaluate LLM accuracy.
Furthermore, we designed a metamorphic testing framework containing four
industrial-specific stability categories with eight abilities, totaling 13,631
questions with variants to evaluate LLM robustness. In total, we evaluated 9
different LLMs developed by Chinese vendors, as well as four different LLMs
developed by global vendors. Our major findings include: (1) Current LLMs
exhibit low accuracy in Chinese industrial contexts, with all LLMs scoring less
than 0.6. (2) The robustness scores vary across industrial sectors, and local
LLMs overall perform worse than global ones. (3) LLM robustness differs
significantly across abilities. Global LLMs are more robust under
logical-related variants, while advanced local LLMs perform better on problems
related to understanding Chinese industrial terminology. Our study results
provide valuable guidance for understanding and promoting the industrial domain
capabilities of LLMs from both development and industrial enterprise
perspectives. The results further motivate possible research directions and
tooling support. | [
0.0080935685,
0.030347988,
-0.046644289,
0.0349576697,
-0.0148739247,
-0.0327869207,
0.0495209582,
0.0169729106,
0.0146502489,
-0.0434984416,
-0.0179386958,
-0.0462047607,
-0.0179651678,
0.0077456036,
0.0234938823,
-0.0272880569,
0.0409643129,
-0.0378737673,
-0.0519408658,
-0.0030092364,
-0.0053608539,
-0.0380489863,
0.0285035539,
0.0420189649,
0.0066758697,
-0.0089557134,
0.0246464852,
-0.0118189668,
0.0232817829,
-0.0195648093,
0.0217087828,
0.0256121513,
0.0265020207,
0.0390275195,
0.0000022875,
-0.085558109,
-0.0278615281,
0.0331805944,
0.0042123008,
0.0462640971,
0.0260711014,
-0.0054286565,
-0.0436605252,
0.0479026511,
-0.0398481376,
-0.0167201664,
0.0149970315,
0.0380107164,
0.0389777832,
0.0484411195,
-0.00513088,
-0.023929853,
-0.0103130275,
0.0065076388,
-0.044164937,
-0.0535366721,
0.0355788283,
0.0673898831,
0.0357856415,
-0.0160738956,
-0.0335822403,
-0.0034489045,
0.0264444519,
0.0097652376,
-0.0810716152,
0.0123484302,
-0.0205031354,
0.0169486981,
-0.000910821,
0.0538231917,
0.0046384158,
0.0104228547,
-0.0269052219,
-0.0154166464,
-0.0008323213,
0.0322960056,
-0.019891018,
0.0252376795,
-0.0151592679,
0.0081408704,
-0.0059763594,
-0.0293151159,
-0.0037042538,
0.0159051493,
-0.053934779,
0.1055399179,
-0.0016697717,
-0.0492604673,
-0.0041904096,
-0.024546193,
-0.0441520549,
-0.0167806074,
0.0191139337,
0.0148037886,
-0.0401904695,
0.0384142101,
-0.0019403162,
0.0144023057,
0.022486927,
-0.0888799652,
-0.0251925085,
0.0415004678,
-0.01784743,
0.0152386436,
-0.0202783346,
0.0829368606,
0.0051607182,
-0.0127499085,
-0.0491250865,
-0.0037269983,
-0.0551969036,
-0.0036778587,
-0.0448598973,
0.0252529494,
-0.0236978028,
-0.0086118635,
-0.061711967,
0.0286301989,
-0.0223478489,
0.0536261722,
-0.0104404762,
-0.0312634185,
0.0065678214,
-0.0242830571,
-0.0335031152,
0.0388382971,
-0.0240203701,
0.0294828918,
-0.0418555848,
-0.0353967622,
0.0468959101,
0.0325782001,
-0.0376461707,
-0.0223816298,
0.0073041925,
0.0851150826,
0.0168206505,
0.0069210394,
-0.053538166,
-0.0029296654,
-0.021589078,
-0.0481453352,
-0.0173402056,
-0.0591509007,
-0.0311637316,
0.0106577817,
0.012754756,
-0.0426267758,
0.0063645304,
0.0039954917,
0.025538072,
0.0373295918,
-0.0177875087,
0.0246111453,
0.0117133046,
0.0079915551,
0.0164940003,
0.0672815368,
0.0318079926,
0.0132429134,
0.0574976206,
-0.0515538007,
-0.0067458255,
0.0369861051,
-0.0627382472,
0.0040037776,
-0.050935559,
0.0350364186,
0.0084395008,
-0.0241081435,
-0.0072400034,
0.0476208776,
0.0161160249,
0.0348390825,
0.07191246,
0.0094139501,
0.0088811181,
0.0170763042,
0.0346425511,
0.026546916,
-0.0176552422,
-0.034313295,
-0.0299345385,
-0.0347803384,
-0.0774272457,
0.0366649032,
-0.0127662253,
0.0118712699,
-0.0181490183,
-0.0621163808,
-0.0471877195,
0.0205130763,
-0.0236333162,
0.0422373191,
0.0723406002,
-0.0587132312,
-0.0155648887,
0.0456423536,
0.0020882695,
-0.0309035983,
-0.0114508495,
-0.0307489373,
-0.0028732915,
0.0803968683,
0.0271725431,
-0.0479577184,
-0.0347989202,
0.0116728591,
-0.0040257075,
0.0670531094,
0.0624798946,
0.0093774647,
-0.0447574034,
-0.0586703159,
-0.0095090773,
0.035141807,
-0.0053713731,
-0.00003214,
0.0025984726,
0.0288204364,
-0.0090492452,
-0.0604950823,
-0.0409154519,
-0.0038510174,
-0.0481062345,
-0.0189332832,
-0.0257795583,
0.0142100155,
0.0030747959,
0.0205973443,
-0.0361474641,
0.0144917667,
0.0204757042,
-0.0251229815,
-0.0292639621,
0.0254511535,
-0.0224835649,
0.0468569621,
0.0737309605,
0.0162119698,
-0.0165622644,
-0.0350853205,
-0.0268215705,
-0.0357026011,
0.0207123794,
0.0106306951,
0.0113277007,
-0.0391751043,
0.0494143292,
-0.0667883009,
0.0286960565,
-0.0456769988,
0.0463862419,
0.0107515752,
0.0419435054,
-0.0156559423,
0.025238717,
0.0339248292,
-0.0249487739,
-0.0752780065,
0.0476584695,
-0.0049920683,
0.0590115376,
0.000530577,
0.0313818567,
-0.0192742199,
-0.0140820546,
-0.0655990094,
0.0192141309,
-0.0113817509,
0.0392487459,
0.0140529759,
-0.0393041633,
-0.0353245884,
0.0255631544,
-0.0263027344,
0.0125141647,
-0.0087493574,
0.0121207247,
0.0158970784,
0.0054311524,
-0.0524688587,
0.004936886,
0.0036604474,
-0.0141821103,
0.0180691443,
-0.0021786203,
0.0069397041,
-0.0971018821,
0.0790243521,
-0.0291107632,
0.0258804373,
-0.0177057162,
-0.0431743599,
-0.0235841759,
0.010648964,
0.0411981754,
0.0177974086,
-0.0282594115,
-0.0055780923,
0.0457296781,
-0.0170456488,
-0.0430803411,
-0.005896287,
0.026052434,
-0.0402214713,
-0.040911153,
-0.051304698,
-0.0161894821,
0.0692603886,
0.0063037411,
-0.0170112178,
-0.0123957423,
-0.0049510081,
-0.0694134682,
0.0094944984,
-0.0287973247,
0.0281880051,
-0.0233886056,
0.0352141224,
0.0095481509,
-0.0546784177,
0.0613788739,
-0.0197145753,
0.0145118209,
-0.045643311,
-0.0314206779,
-0.0541103594,
0.044329349,
-0.0414848141,
0.0102154519,
0.0237578992,
-0.0478801392,
-0.0503887124,
-0.0088174446,
-0.0074004619,
-0.0237962566,
0.0581587926,
-0.047617387,
0.0226306692,
-0.0224900264,
-0.0253498312,
-0.0077932188,
0.0094642593,
-0.019801626,
0.0733712316,
0.0203966293,
0.0169807263,
-0.0056345048,
0.0664767027,
0.0223403629,
-0.0056460388,
0.002825055,
0.0021648994,
-0.0049007819,
-0.0394878909,
0.0259201024,
0.0439716838,
0.0288869813,
-0.0353717878,
0.0555113629,
0.0314419195,
0.0462162048,
-0.0085525475,
-0.0464586429,
-0.0050164852,
0.020054929,
-0.0452431254,
-0.0177701395,
0.023529971,
0.0308913607,
-0.0647390261,
0.0040476043,
0.1004563421,
0.0159133971,
0.0061046532,
0.0271299537,
-0.0270756353,
-0.0012506096,
0.0577195548,
0.0432880037,
-0.0916529596,
-0.0100877546,
-0.0056324648,
0.0176810306,
0.0131316464,
-0.000765621,
0.0122272056,
-0.1525644064,
0.0507876612,
-0.007889526,
-0.0525366552,
-0.061557278,
-0.0045541627,
0.0134261465,
0.0117233787,
0.1115030125,
0.0256529972,
-0.0135153029,
-0.0372190326,
0.0394579805,
-0.0787295401,
-0.0937020406,
0.0232760757,
-0.0431390144,
0.0246818494,
0.0533977076,
-0.0197607521,
-0.0189783685,
0.0116348788,
0.0627099499,
0.0119843232,
-0.0060099559,
-0.0182938892,
-0.0065687094,
0.0329026133,
-0.0091388151,
0.01472505,
-0.0166499577,
-0.0224703588,
0.0021771432,
0.0715591609,
0.0640236884,
-0.0165191069,
-0.0203632209,
0.002867853,
0.0132145165,
-0.0202627238,
0.0514146723,
-0.0049309796,
-0.0687017739,
-0.0395280346,
-0.0448344573,
-0.0584372766,
-0.0100184949,
-0.0203297995,
0.0661405474,
-0.0457612127,
0.0092272088,
-0.0423410684,
-0.0242765117,
0.0466570258,
0.004661649,
0.0143523831,
-0.0050791944,
0.0220874902,
0.0138675971,
-0.0120968632,
-0.0050743902,
-0.0340563431,
0.0619477443,
-0.0857740343,
0.0314818323,
0.0477270298,
-0.0344449058,
-0.0039626253,
-0.0108224703,
-0.0353728198,
-0.0572961532,
-0.0043986207,
0.0123731894,
0.0233416632,
-0.0326370187,
0.0017089688,
-0.0360729545,
0.0566476099,
0.0043510366,
-0.0094972663,
0.0062609972,
0.0081619313,
0.0000706254,
-0.0103981486,
0.0107741123,
-0.0332507975,
0.0229459386,
0.0174247883,
0.0736189336,
-0.0041257767,
-0.0153277442,
0.0102539258,
0.0748542398,
-0.0181828011,
0.0066928212,
-0.0158752631,
-0.021080967,
0.0469387956,
-0.0350616314,
-0.0465541743,
-0.0106976656,
0.0026388874,
0.0086303437,
-0.0212801211,
0.0166859925,
0.0199323259,
-0.0102491248,
0.0149591584,
-0.0067888624,
0.06560193,
-0.0496559478,
0.0494417958,
0.0451718606,
0.0140202949,
-0.0164738037,
-0.018000612,
0.0135396449,
0.0520434231,
0.0118649229,
0.092206575,
-0.0076951915,
-0.001873582,
-0.0422441699,
0.0062072668,
-0.0141561367,
-0.0732215568,
0.0085445074,
0.027370844,
0.0640827417,
0.0133894961,
0.0066907909,
-0.0510444976,
0.0178050809,
-0.021464061,
-0.0045007453,
0.0237625521,
-0.0147606255,
0.0331216902,
0.0165279936,
-0.0354627185,
-0.020105971,
0.0831127539,
0.0306524839,
0.0001736981,
-0.0086892825,
-0.0114966678,
-0.0539547093,
-0.0126782348,
0.0732958093,
-0.0201101024,
-0.0514710583,
-0.0065055783,
-0.1165754721,
-0.067207776,
-0.0128287263,
0.0255749673,
-0.0119311223,
-0.0175294615,
-0.0137680648,
0.0743690953,
0.0354022197,
-0.0029801007,
-0.0224153604,
-0.0016912016,
-0.0319509245,
0.0196607187,
0.0053935815,
-6.342301893e-33,
-0.0149120111,
-0.0402587727,
0.0461999811,
-0.0373951904,
-0.0023675859,
-0.0000688832,
0.0058906833,
0.0142421322,
-0.0270600989,
-0.0318603218,
-0.0007547427,
-0.0036503368,
0.0229852665,
0.0079804538,
0.0027703263,
-0.0131783551,
0.0092193158,
0.0232377406,
0.0123278759,
-0.0347814299,
0.0350396149,
-0.0032127895,
0.0737980008,
-0.0530573986,
0.0460717268,
0.0127852233,
-0.0051088082,
0.016221514,
-0.016239062,
0.0867452398,
0.0101286601,
0.0901279673,
0.0187269002,
-0.0322140269,
-0.0349534452,
-0.0178693738,
-0.0468723737,
-0.0275765453,
0.0094443075,
0.0344764069,
0.0704902038,
-0.0082068816,
0.0524349585,
-0.0093574002,
-0.0278539248,
0.0316309072,
0.0357691981,
-0.0622021034,
-0.0229480974,
0.0007869682,
-0.0394498855,
0.0255117994,
0.0204599239,
0.0626525134,
-0.0216139257,
0.0298625864,
-0.0214210488,
0.0040912433,
-0.037297409,
0.0030300503,
0.0457129814,
0.0454331562,
0.0849593952,
0.0103715165,
0.0291337315,
0.0231798217,
0.0046644462,
0.0372751504,
0.0407105722,
-0.0517959185,
0.0454725511,
0.0518649369,
0.0163099468,
0.0552156828,
0.0221452005,
-0.0289062299,
-0.0462544896,
0.0257711019,
0.0227145702,
0.0260192007,
0.0483452454,
0.0151464427,
-0.0714761317,
-0.0459963381,
-0.0192860886,
0.0751856789,
-0.0306108873,
-0.0463801324,
0.0347617008,
-0.015254599,
-0.023789322,
0.0095598046,
0.0416046642,
-0.045250535,
-0.0413036309,
-0.0142529681,
-0.0106445104,
0.0163278133,
-0.0219544005,
0.036198698,
-0.0497191027,
-0.0527496189,
-0.0527284779,
-0.0159543883,
0.0019226093,
0.0118763298,
-0.0207813531,
-0.001949131,
-0.0735901967,
0.0175171513,
-0.0050771493,
-0.0219367519,
0.0145171573,
-0.0004815631,
0.0247597694,
-0.0032655676,
0.0288261045,
0.0129435165,
0.0414343812,
-0.0556039326,
-0.0328128114,
0.0006964994,
-0.0141773513,
0.0297942217,
-0.0126088532,
0.0158120915,
-0.0341231264,
0.037758667,
0.0442856774,
-0.033174403,
-0.0161104184,
-0.0167311952,
2.993e-7,
0.0494846851,
0.0519000031,
-0.0168103445,
0.0451771542,
-0.0180661716,
0.0157071333,
-0.0284908321,
0.0002967217,
0.0128109138,
0.0348188728,
0.0204054732,
0.0142629156,
0.040494103,
0.0575817488,
-0.0636270866,
-0.0025212162,
-0.0021691672,
0.0389217846,
-0.0570039451,
0.0595243946,
0.0412764028,
0.0547266677,
-0.0328303464,
0.0067921435,
-0.0143698463,
-0.0960517004,
0.0214338489,
-0.0451905541,
-0.0103599625,
0.0537132993,
-0.0155264623,
0.0159260463,
0.0063537951,
0.0311728027,
0.0172575526,
0.0345585421,
0.0137872268,
0.0573385693,
0.0173001457,
0.0064587472,
-0.0081774276,
-0.0588421226,
0.0090475548,
-0.0390185937,
0.0295338854,
-0.0391066708,
-0.0413897522,
0.0062209452,
-0.0226521548,
-0.0570127778,
-0.023396872,
-0.0415306762,
-0.0038051775,
0.0168164335,
0.0160866585,
-0.0084303059,
-0.0008469302,
-0.0169213321,
-0.0222406015,
0.0348375291,
-0.0398100056,
-0.0915321112,
0.0141742155,
-0.0320789255,
-0.0148273576,
-0.0072255111,
0.0238323677,
3.158208708e-34,
0.037422549,
0.0045036315,
0.006871284,
0.0451222062,
0.0481625721,
0.0069559859,
-0.0179417115,
0.0176721849,
-0.011095711,
-0.0686971918,
-0.0288168974
] |
http://arxiv.org/abs/2401.15269v3 | http://arxiv.org/pdf/2401.15269v3 | Improving Medical Reasoning through Retrieval and Self-Reflection with Retrieval-Augmented Large Language Models | 2024-01-27T02:29:42 | 2024-06-18T02:10:15 | Minbyul Jeong, Jiwoong Sohn, Mujeen Sung, Jaewoo Kang | cs.CL | cs.CL, cs.AI, cs.IR | Recent proprietary large language models (LLMs), such as GPT-4, have achieved
a milestone in tackling diverse challenges in the biomedical domain, ranging
from multiple-choice questions to long-form generations. To address challenges
that still cannot be handled with the encoded knowledge of LLMs, various
retrieval-augmented generation (RAG) methods have been developed by searching
documents from the knowledge corpus and appending them unconditionally or
selectively to the input of LLMs for generation. However, when applying
existing methods to different domain-specific problems, poor generalization
becomes apparent, leading to fetching incorrect documents or making inaccurate
judgments. In this paper, we introduce Self-BioRAG, a framework reliable for
biomedical text that specializes in generating explanations, retrieving
domain-specific documents, and self-reflecting generated responses. We utilize
84k filtered biomedical instruction sets to train Self-BioRAG that can assess
its generated explanations with customized reflective tokens. Our work proves
that domain-specific components, such as a retriever, domain-related document
corpus, and instruction sets are necessary for adhering to domain-related
instructions. Using three major medical question-answering benchmark datasets,
experimental results of Self-BioRAG demonstrate significant performance gains
by achieving a 7.2% absolute improvement on average over the state-of-the-art
open-foundation model with a parameter size of 7B or less. Overall, we analyze
that Self-BioRAG finds the clues in the question, retrieves relevant documents
if needed, and understands how to answer with information from retrieved
documents and encoded knowledge as a medical expert does. We release our data
and code for training our framework components and model weights (7B and 13B)
to enhance capabilities in biomedical and clinical domains. | Improving Medical Reasoning through Retrieval and Self-Reflection with Retrieval-Augmented Large Language Models: Recent proprietary large language models (LLMs), such as GPT-4, have achieved
a milestone in tackling diverse challenges in the biomedical domain, ranging
from multiple-choice questions to long-form generations. To address challenges
that still cannot be handled with the encoded knowledge of LLMs, various
retrieval-augmented generation (RAG) methods have been developed by searching
documents from the knowledge corpus and appending them unconditionally or
selectively to the input of LLMs for generation. However, when applying
existing methods to different domain-specific problems, poor generalization
becomes apparent, leading to fetching incorrect documents or making inaccurate
judgments. In this paper, we introduce Self-BioRAG, a framework reliable for
biomedical text that specializes in generating explanations, retrieving
domain-specific documents, and self-reflecting generated responses. We utilize
84k filtered biomedical instruction sets to train Self-BioRAG that can assess
its generated explanations with customized reflective tokens. Our work proves
that domain-specific components, such as a retriever, domain-related document
corpus, and instruction sets are necessary for adhering to domain-related
instructions. Using three major medical question-answering benchmark datasets,
experimental results of Self-BioRAG demonstrate significant performance gains
by achieving a 7.2% absolute improvement on average over the state-of-the-art
open-foundation model with a parameter size of 7B or less. Overall, we analyze
that Self-BioRAG finds the clues in the question, retrieves relevant documents
if needed, and understands how to answer with information from retrieved
documents and encoded knowledge as a medical expert does. We release our data
and code for training our framework components and model weights (7B and 13B)
to enhance capabilities in biomedical and clinical domains. | [
0.0656627715,
0.0262076724,
0.0154166063,
-0.0243530534,
-0.0291504804,
-0.0061163045,
0.0186892785,
0.025396673,
0.0218418334,
-0.0378826186,
0.0008251295,
-0.0283016842,
-0.0510416701,
0.0166003276,
0.0247235224,
-0.0040912693,
0.0480335392,
0.0053047328,
0.0272010118,
-0.0104353903,
-0.0123594208,
0.0222544614,
-0.0179985519,
0.0305752624,
0.024038773,
-0.0314574279,
0.0128716333,
-0.0008378496,
-0.0068698754,
-0.062573418,
0.0268311426,
0.0307600144,
0.0351829566,
0.0019022542,
0.0000019719,
-0.0555131324,
-0.041373454,
0.0278361887,
-0.037608467,
-0.0140841575,
0.0241740029,
-0.0015821655,
0.023338845,
0.03506165,
-0.0477787554,
-0.036029987,
0.0003840464,
0.053074766,
0.0731342211,
0.0639919266,
0.0036589159,
-0.0230013616,
0.0290325303,
-0.0221510697,
0.0569755994,
-0.0121201975,
0.0300593153,
0.0189351048,
0.0388777293,
0.0422813594,
0.0216897111,
0.0312063955,
0.0079311952,
0.0044918032,
0.0372513533,
0.043923486,
-0.020225266,
0.0014557866,
0.0109774657,
0.0661093742,
0.0588851422,
-0.0294802804,
0.0177535377,
0.0282870363,
-0.0033907227,
0.0318258815,
-0.053433992,
0.0214208346,
-0.0227930453,
-0.0022417565,
0.0066193035,
-0.0225325841,
0.0256491024,
0.0235485528,
-0.0146803465,
0.0476996042,
0.0283398628,
-0.0304087251,
-0.023087386,
-0.0283678882,
-0.0065541924,
-0.0339610241,
0.0312624387,
0.0843242556,
-0.0243423339,
0.0198438019,
-0.0053794654,
-0.0233403854,
0.0881620571,
-0.0018721162,
0.0169072188,
0.0120392526,
-0.0271425508,
0.0303663686,
-0.0446709208,
0.0224269461,
-0.0524569303,
0.0112257255,
-0.0719053224,
-0.0207717009,
-0.0683081746,
-0.0406406596,
-0.0425159484,
0.064443104,
-0.0332541578,
0.006174718,
-0.0705300644,
0.0078773871,
0.009292502,
-0.0148206474,
-0.002123018,
0.045861572,
-0.038866777,
-0.0181280989,
-0.0500831231,
0.031456057,
-0.0178882927,
-0.0127495397,
0.0301205851,
-0.0483728051,
-0.0169400387,
-0.0016478554,
0.0184570272,
-0.0168437976,
0.0383960083,
0.0598462038,
-0.0059003844,
-0.007596666,
0.0044320133,
-0.0732599422,
-0.038220752,
-0.06065109,
0.0064651687,
-0.0591176823,
0.0110650314,
0.0345469788,
-0.0405209884,
-0.0296748467,
0.0360140987,
0.0469011217,
-0.0237027239,
0.0392067395,
-0.0253009759,
0.0199717768,
0.0514398068,
0.0283902586,
0.0038457308,
0.011188278,
-0.009820777,
-0.0325650983,
0.0108275767,
-0.0313671045,
0.0169113446,
-0.045118589,
0.0463984571,
-0.0357247815,
-0.0013271944,
0.0626486465,
0.0051558823,
0.0850736201,
-0.0096907234,
0.064598456,
-0.0235887766,
-0.0043004146,
0.0073604756,
0.0647355691,
0.0102430517,
0.0235550627,
0.0326518491,
-0.0073253461,
0.0163809359,
-0.0094095068,
-0.025569899,
0.0264355093,
0.0082120197,
0.0226202291,
0.006243513,
0.0110293999,
-0.0405613892,
-0.0332201868,
-0.0303210858,
0.0113444775,
-0.0315253995,
0.0265207868,
0.0564063229,
0.0047249636,
-0.0438681059,
0.0762383491,
0.0322590396,
-0.0681441948,
0.0149068562,
-0.0822474286,
0.0563217252,
0.0621936247,
0.083361432,
-0.0435055532,
-0.0517980568,
-0.004246437,
0.0090239551,
0.0171656944,
0.0388497114,
-0.0183314383,
-0.0179123916,
-0.0625322834,
-0.008321424,
-0.0153648071,
-0.0134739513,
-0.00212686,
0.025465196,
0.0084563233,
0.0250042286,
-0.0333476327,
-0.0078419549,
0.044890251,
-0.054565981,
-0.0435183048,
0.0297030248,
-0.0323491655,
0.0164109897,
0.0370479226,
-0.0029571126,
0.0679441616,
0.0482877903,
-0.0362518579,
-0.0537402071,
-0.0199296121,
-0.0593289994,
-0.0359328017,
0.0271750875,
-0.0122131165,
-0.0135952905,
-0.005675904,
-0.0327142626,
-0.0610575899,
-0.0104096113,
-0.0474664941,
0.0221078247,
-0.0595161319,
-0.0193776004,
0.0418510623,
-0.014532513,
-0.0698775724,
0.0382284075,
0.023047056,
0.023185011,
0.0046421378,
-0.0028632861,
0.0249580238,
-0.057633888,
-0.0041335016,
-0.0049613295,
-0.0298236851,
0.0064973356,
0.0223207735,
-0.003630436,
-0.0617810413,
-0.0256748311,
-0.0976227969,
0.0191724524,
0.0094320299,
0.0084625259,
0.0242871214,
0.012348244,
0.0170567222,
-0.0255177077,
-0.067063272,
-0.0153446048,
0.0431501828,
0.0260770563,
0.0377916694,
0.0277330745,
0.0345660783,
-0.0217220578,
-0.0147514148,
-0.0413559638,
0.0045156125,
0.0032424948,
0.0606986955,
-0.0468429215,
-0.009639536,
-0.0526221953,
0.0030425477,
0.0048742974,
-0.0295164585,
0.0349222235,
-0.0413669161,
0.0137958918,
0.0586437099,
0.021553902,
-0.0024696977,
0.0333431661,
0.0136692319,
-0.0518013611,
0.0360742323,
-0.0657964945,
-0.0250965916,
0.0202242713,
-0.0227215961,
-0.0155595951,
0.0519704781,
0.0229311269,
-0.0336919501,
-0.0429396369,
0.0041528377,
-0.0051099546,
0.0068875374,
-0.0479194224,
-0.0154971518,
-0.0031113701,
0.0193053856,
0.0036307569,
-0.0224467944,
0.1053272635,
-0.0097375065,
0.01594688,
-0.0830846652,
0.005138916,
-0.0504030921,
0.0370265357,
-0.018790435,
0.0219527744,
0.0235098545,
-0.0029062636,
-0.0199388508,
0.0218271334,
0.0164938029,
-0.0716279373,
0.04539188,
-0.0398448668,
0.0249680988,
0.0191123653,
-0.0598778017,
-0.0598498359,
-0.0166406073,
0.0081933951,
-0.0113873268,
-0.0131370202,
0.0077808029,
-0.0476013348,
0.0087684495,
0.0355062447,
0.0384583436,
-0.0227051172,
-0.0120482929,
-0.0077416277,
0.0334425494,
0.0681716502,
0.0202027261,
0.0257056579,
-0.0074138818,
0.0447702706,
0.070185788,
0.0494729802,
-0.0092527661,
0.0143460538,
-0.0016046856,
-0.0008776349,
-0.0423330404,
0.0224403758,
0.0194711592,
0.0460585393,
-0.0837084875,
-0.0378115252,
0.1085145399,
0.0436886884,
-0.0358018205,
-0.0207747519,
-0.0551291965,
-0.0046348218,
0.0075792368,
0.0464255288,
-0.0375990905,
-0.0193562862,
-0.012999896,
0.0392707884,
-0.0032353951,
-0.0185388457,
0.025509648,
-0.0919075161,
0.0446363054,
-0.0014060775,
-0.0178205203,
-0.0231410414,
-0.05151283,
-0.0293931104,
0.0219149385,
-0.0058951378,
0.0378449969,
-0.0253639407,
0.0148833664,
0.0219597984,
-0.0245540962,
-0.0350433104,
0.0417108424,
0.0063200295,
0.0231980085,
0.0678701326,
0.0237544477,
0.0101762554,
-0.0226386376,
0.0127093103,
-0.0420443863,
0.0142447697,
-0.0093498612,
0.0821989849,
0.0121317804,
0.0095226523,
-0.0716206804,
0.0668128431,
-0.0492545404,
-0.0231075101,
0.0094206352,
0.0677031949,
-0.0280164443,
-0.0086232815,
-0.0665259808,
-0.0009222635,
-0.0043170429,
0.0832434744,
0.00089359,
0.0040277853,
-0.0350026786,
0.0192532409,
0.0524686575,
-0.0206327625,
-0.0064858119,
0.0313810259,
-0.014229957,
0.0394422784,
-0.0689898208,
-0.0333308429,
0.0275298133,
0.0170585662,
-0.004599269,
0.0337371379,
-0.0044113961,
-0.0104308482,
-0.0727366433,
0.0573727675,
-0.0014487018,
-0.0299109649,
-0.0436059274,
-0.0388757475,
0.0684059933,
-0.0301738419,
-0.0121256141,
0.0054424223,
0.0091743451,
-0.0317017063,
-0.0221587345,
0.0568920299,
0.0204852987,
0.0251890421,
-0.0083055254,
-0.0216317233,
0.0586181059,
0.033065699,
-0.0736564547,
-0.0389674418,
-0.0133328224,
-0.0223001447,
-0.0191700924,
0.0323309153,
-0.0359913893,
0.0334201194,
-0.0099035222,
0.0606165268,
0.0253102742,
-0.0215812232,
0.0615410358,
0.0409965888,
0.0289002582,
-0.013533379,
0.0231450014,
-0.0094752349,
0.0276668649,
-0.0491886549,
0.0052120904,
-0.0162378401,
0.0269012321,
0.0255226083,
-0.0506864265,
0.013278285,
-0.0143022379,
-0.0557518601,
0.0274076611,
-0.0209103469,
0.0163777974,
-0.022809675,
0.0128877787,
0.0401619449,
-0.0221043006,
-0.0204062946,
-0.0147892945,
0.0016945442,
0.029548062,
-0.0257075801,
0.0959055945,
0.0099775698,
0.0238352008,
0.0151824905,
0.006019555,
-0.0734418482,
-0.0345691368,
-0.0168856457,
0.0083585698,
0.0002506121,
0.0183432978,
0.0207938664,
-0.0127461944,
-0.0243282374,
-0.0245499089,
0.035341721,
-0.0125808716,
0.0144644799,
0.049222447,
0.0355212614,
-0.0166979451,
-0.031275399,
0.0692759082,
0.0117840031,
0.0437288322,
0.0147751207,
-0.0316391736,
-0.0411218069,
-0.0416712835,
-0.0012330632,
0.0184491053,
-0.0425786823,
0.0122025162,
-0.0565409511,
-0.0575281344,
0.0093270093,
0.0215558112,
-0.0357373543,
-0.0245652571,
0.056525901,
0.0667746887,
-0.0005125416,
-0.0718567967,
-0.0383415893,
0.1107427627,
-0.0866335928,
-0.0183320399,
0.0140530942,
-6.349208657e-33,
0.0012207712,
-0.0382567607,
0.0531996116,
-0.0348915234,
-0.0002513224,
0.0076406226,
-0.0461286753,
-0.0182166304,
-0.0154037839,
0.0268227831,
-0.0144560197,
0.0059652058,
-0.0021125211,
0.0507496446,
0.009287593,
-0.0193221234,
0.0201849956,
-0.0047303857,
-0.0256656595,
0.0548277497,
0.0227893759,
0.0561847687,
0.0176384486,
-0.0409259014,
-0.0184382834,
-0.0245398134,
0.0401144326,
0.0180839468,
-0.0361030474,
0.0764245763,
-0.033326298,
0.0367176756,
0.0294579323,
-0.0234565698,
0.020845186,
0.0200897902,
-0.1074617058,
-0.0250179581,
-0.0055430634,
-0.0611111447,
-0.0755254552,
-0.0532513335,
0.0387659073,
-0.0293295123,
-0.0085933637,
-0.0110176429,
0.0161816161,
-0.0044305073,
-0.011645915,
-0.017885834,
-0.0224112179,
0.010210637,
0.0153812682,
-0.0186027884,
-0.0358259156,
0.0315104835,
-0.0234450512,
0.0595537052,
-0.0051178206,
-0.0087625403,
0.0035662283,
0.049316287,
0.0112376893,
-0.0268365946,
0.0055415123,
0.0346660726,
0.0186005812,
-0.0140754404,
-0.0346543752,
-0.0132658323,
-0.0418319292,
0.0331698805,
0.0542594232,
0.0425585546,
0.0337781049,
0.000831957,
-0.1125807539,
0.0463332161,
-0.0571970753,
0.0751868412,
0.0265539438,
-0.0308969449,
-0.0136392843,
0.0166684724,
-0.0338495076,
0.0450972766,
-0.0314743519,
-0.0474471599,
0.0492007509,
0.002966332,
-0.0263117347,
-0.0103985509,
0.00816641,
-0.0243215207,
0.0016503811,
0.0011084008,
-0.0102258632,
-0.0368424654,
-0.0299879741,
0.0386467688,
-0.0574811511,
-0.0017872757,
0.0254862588,
-0.0178998616,
-0.0025229028,
0.0264985636,
0.0114849079,
-0.020276444,
-0.0469308645,
-0.0033142515,
-0.0342025086,
0.030244831,
-0.0325458832,
0.002296244,
0.0039334572,
0.0680277124,
0.0077051297,
0.0302052274,
0.0044793291,
0.0037772695,
-0.058267694,
-0.0376511477,
0.0187051613,
-0.0002930215,
-0.0151134841,
0.0315706022,
-0.026017243,
0.0587606691,
0.0074027847,
-0.0407188833,
-0.0052491077,
0.0400724597,
2.759e-7,
0.0807577893,
0.0476997495,
0.0362760499,
-0.0061820564,
0.0278505385,
0.0216414686,
0.0022938638,
-0.0061257556,
0.0407627523,
-0.0179958288,
0.0064207986,
0.0268004369,
0.0232122336,
0.0167298596,
-0.0680362061,
0.0012505075,
-0.0512067787,
-0.0365560241,
-0.0131242964,
-0.0295627154,
0.0765895694,
0.0580465421,
-0.0479026325,
-0.0412725359,
0.0013670953,
-0.0191454981,
0.0701685995,
0.0043287892,
-0.0175578184,
0.0504240803,
0.0049173892,
0.0140228942,
0.0103590116,
0.0293770544,
-0.0210216008,
-0.0194590464,
-0.0098197898,
0.0565706044,
-0.0298221335,
-0.0398565419,
0.0096314903,
0.0037929916,
0.0292625092,
-0.0559305586,
0.0524669737,
-0.0747467428,
-0.0348593593,
-0.0138616031,
-0.0317282416,
-0.0193217341,
0.0181154609,
-0.0349434763,
-0.0222300328,
0.0438501537,
-0.0024838252,
0.0141948862,
0.0139007363,
-0.0878352374,
-0.0175553113,
0.019840451,
-0.0551916473,
-0.0209441651,
-0.0234016273,
0.0036232411,
-0.0057032662,
-0.0832705051,
-0.0003056078,
2.559987005e-34,
0.0157043934,
0.0087747956,
-0.0123778563,
0.0127325095,
0.037532948,
0.0008180059,
-0.0269777384,
0.0306100957,
0.0206965338,
-0.0913187116,
-0.0220812578
] |
http://arxiv.org/abs/2402.01722v1 | http://arxiv.org/pdf/2402.01722v1 | Enhancing Large Language Model Performance To Answer Questions and Extract Information More Accurately | 2024-01-27T00:18:07 | 2024-01-27T00:18:07 | Liang Zhang, Katherine Jijo, Spurthi Setty, Eden Chung, Fatima Javid, Natan Vidra, Tommy Clifford | cs.CL | cs.CL, cs.AI | Large Language Models (LLMs) generate responses to questions; however, their
effectiveness is often hindered by sub-optimal quality of answers and
occasional failures to provide accurate responses to questions. To address
these challenges, a fine-tuning process is employed, involving feedback and
examples to refine models. The objective is to enhance AI models through
continuous feedback loops, utilizing metrics such as cosine similarity, LLM
evaluation and Rouge-L scores to evaluate the models. Leveraging LLMs like
GPT-3.5, GPT4ALL, and LLaMA2, and Claude, this approach is benchmarked on
financial datasets, including the FinanceBench and RAG Instruct Benchmark
Tester Dataset, illustrating the necessity of fine-tuning. The results showcase
the capability of fine-tuned models to surpass the accuracy of zero-shot LLMs,
providing superior question and answering capabilities. Notably, the
combination of fine-tuning the LLM with a process known as Retrieval Augmented
Generation (RAG) proves to generate responses with improved accuracy. | Enhancing Large Language Model Performance To Answer Questions and Extract Information More Accurately: Large Language Models (LLMs) generate responses to questions; however, their
effectiveness is often hindered by sub-optimal quality of answers and
occasional failures to provide accurate responses to questions. To address
these challenges, a fine-tuning process is employed, involving feedback and
examples to refine models. The objective is to enhance AI models through
continuous feedback loops, utilizing metrics such as cosine similarity, LLM
evaluation and Rouge-L scores to evaluate the models. Leveraging LLMs like
GPT-3.5, GPT4ALL, and LLaMA2, and Claude, this approach is benchmarked on
financial datasets, including the FinanceBench and RAG Instruct Benchmark
Tester Dataset, illustrating the necessity of fine-tuning. The results showcase
the capability of fine-tuned models to surpass the accuracy of zero-shot LLMs,
providing superior question and answering capabilities. Notably, the
combination of fine-tuning the LLM with a process known as Retrieval Augmented
Generation (RAG) proves to generate responses with improved accuracy. | [
0.0388002582,
0.0710135698,
-0.0196250211,
0.0436030999,
-0.0219897237,
-0.0001855869,
-0.0091145607,
0.0288572852,
0.0038491047,
-0.0513652191,
-0.033987768,
-0.0473108143,
-0.0515768193,
0.0164244641,
0.0304866377,
-0.0121185072,
0.0450345166,
-0.0125744324,
0.0139075825,
-0.0302480478,
0.000664948,
0.010133571,
0.0154404016,
0.0136794662,
0.0205865558,
-0.0127501208,
-0.0014249299,
-0.0167743005,
-0.0054889554,
-0.007772997,
0.022889344,
0.0255411174,
0.0174474157,
0.060343802,
0.0000021195,
-0.076202862,
-0.0110324966,
0.008812665,
-0.0430248193,
-0.0016666083,
0.0104281595,
0.022666214,
0.0554907285,
0.0482495762,
-0.0772700608,
0.0214341097,
0.0224243347,
0.0317350402,
0.0878587589,
0.0770050958,
0.0080288956,
-0.0585308336,
0.0435028672,
-0.0348926932,
0.046435196,
-0.0713171288,
0.0225731581,
0.0790335909,
0.0408945717,
0.0092867743,
-0.0060108984,
0.0424861349,
0.0153125413,
-0.0055982573,
0.0204528756,
0.0161792058,
-0.0682761595,
0.0106164636,
-0.0111043612,
0.0656728745,
0.0185551215,
-0.0047167437,
-0.0196640361,
-0.020205766,
0.0021330952,
0.0309599489,
-0.0616177097,
0.0085350284,
-0.0181875043,
-0.0117541049,
0.0097842384,
-0.01728951,
0.0009030086,
0.0065369667,
-0.0460712835,
0.0414982662,
0.0006053066,
-0.0498105362,
0.0095011238,
0.0027364511,
-0.0354007743,
-0.0266711079,
-0.0087264087,
0.0664501935,
-0.009622422,
0.0122391945,
0.0016693284,
-0.0326484889,
0.0280062631,
-0.0339929014,
0.0336039364,
0.0433045998,
-0.0104884095,
0.0431830063,
-0.0286944415,
0.0855665505,
-0.0115852514,
-0.0216809735,
-0.0607394129,
-0.0265572369,
-0.0791999102,
-0.0335847996,
-0.0344432294,
0.0950772539,
-0.0334098898,
-0.0191493817,
-0.0307202712,
0.0140264053,
-0.0030844018,
0.0099390885,
0.0007266592,
0.0046480512,
-0.033508271,
-0.0145783471,
-0.0770612583,
0.0105129816,
-0.059013892,
0.0133861369,
0.0050680633,
-0.0637139976,
0.0021320281,
0.0200342275,
-0.0059598214,
-0.0279100612,
0.04591823,
0.0708902255,
0.0000921353,
-0.0396605395,
-0.0160127487,
-0.0409875922,
0.0133529389,
-0.050969597,
-0.014364874,
-0.04777373,
0.0398779809,
0.016178567,
-0.045423843,
-0.0018457923,
0.0051255287,
0.0217005722,
-0.0224495977,
0.0548754968,
0.0096653327,
0.0200156942,
0.0479755141,
-0.0205152947,
0.0219632462,
0.0418378823,
0.012495569,
0.0325768776,
0.0403042361,
-0.0075739538,
0.0160042476,
0.032524135,
0.0331071354,
-0.0093323244,
0.0076993736,
0.094562687,
-0.000234244,
0.0225308407,
0.0292296689,
0.053129714,
-0.0350261405,
-0.001511458,
0.0424893498,
0.0288027637,
0.046205353,
0.0246999376,
0.0148575138,
0.0233965609,
0.047138188,
-0.0140351662,
-0.0141469091,
0.0119277593,
-0.0438462496,
0.0458911695,
0.0309983641,
0.0062353564,
-0.0410189666,
-0.0392454416,
-0.0053671375,
0.0059231687,
-0.0518025123,
0.0715878084,
0.0266964231,
-0.0199164432,
-0.0526921265,
0.0381610058,
0.0047741728,
-0.0076864213,
-0.0171331856,
-0.0800582692,
0.0733697712,
0.0457590856,
0.0477775224,
-0.0291038956,
-0.0402401276,
-0.0103661669,
-0.0271148141,
0.0315778404,
0.045787733,
-0.0106418207,
-0.0305261239,
-0.04557053,
-0.0176209603,
-0.0076129842,
-0.0066452143,
0.0436101221,
0.0180199817,
0.0054827388,
0.0307567175,
-0.0359022319,
-0.0126156416,
0.0174427908,
-0.0243727025,
-0.0658735931,
0.0199900288,
-0.0337267444,
0.0074492078,
0.0349884145,
-0.0117995944,
0.0384424813,
0.0101892538,
-0.0251804814,
-0.0365716107,
-0.0327509195,
-0.0239002351,
0.0066490411,
0.0221221317,
0.006494516,
0.003019488,
0.0332975946,
-0.0297023747,
-0.040415179,
-0.025180975,
-0.0316038504,
0.0225264337,
-0.0848465338,
-0.015001216,
0.0003721555,
0.017746428,
-0.0604814067,
0.0864667445,
0.0069517544,
0.0299023632,
0.0061434563,
0.0476239845,
0.0287158396,
-0.0711433068,
0.0180217661,
-0.0091268755,
-0.0154190063,
0.0506189838,
0.0529851019,
0.0293166414,
-0.0308019612,
-0.0438417196,
-0.0640818924,
0.0144191757,
-0.0152626773,
0.0487150513,
0.0032079287,
0.0059434953,
-0.0135640698,
-0.0153606404,
-0.0607867017,
0.005663773,
0.0416876636,
0.0243211593,
0.051069323,
0.0189856868,
-0.0142779993,
-0.0671388656,
0.001103061,
0.0084526716,
0.0014141586,
0.0207373034,
0.0213810485,
-0.0247454643,
0.0053243558,
-0.0350991748,
0.0148133701,
-0.010141247,
0.0087174289,
0.0118854223,
-0.0383654684,
-0.0094357328,
0.0470053181,
0.0052168192,
0.0149359629,
-0.0352269411,
0.0206259135,
-0.0797226653,
0.040078517,
-0.0192771293,
-0.0608918257,
-0.0123229725,
0.0018061762,
0.0334843919,
0.0289880466,
0.0023398995,
-0.007296199,
-0.0670343563,
0.0667203143,
0.011997777,
-0.0132062482,
-0.0594224446,
0.0037001544,
-0.0033353739,
0.0444767773,
-0.0133701181,
-0.0496153831,
0.0519215614,
-0.017482182,
-0.0080441376,
-0.0432599671,
-0.0174176134,
-0.073972486,
0.0535242967,
-0.0057874923,
0.0099026188,
-0.0226976909,
-0.0207909402,
-0.0563463345,
-0.0265556537,
0.0256974678,
-0.0581534281,
0.0461204611,
-0.0149246017,
0.0279794559,
-0.0006321169,
-0.0364563987,
-0.0589296073,
-0.0563978888,
-0.0167509373,
0.0277766213,
0.0094659468,
-0.022176167,
-0.0353439152,
0.0112564228,
0.0218800623,
-0.015070471,
-0.0360338092,
-0.0227872711,
-0.0040778252,
0.0556778274,
0.0959859788,
0.0407939292,
0.0302130021,
-0.032014288,
0.0360276364,
0.0172891915,
0.0287025999,
0.0109877056,
-0.0001296103,
-0.014715408,
0.0054299245,
-0.0444130525,
0.0009753709,
0.0254521184,
-0.0065798848,
-0.0500179641,
-0.0393788964,
0.1147242412,
0.0281330012,
0.0124476142,
-0.0205572788,
-0.0367383622,
0.0261948071,
0.014252604,
0.0926972702,
-0.0788586065,
-0.0161067769,
-0.0194607787,
0.0360653028,
0.0101805236,
-0.0124134384,
0.0061857929,
-0.1792817265,
0.0407938473,
-0.0288294423,
-0.0401058309,
-0.0362400934,
-0.0385411829,
-0.0239103865,
-0.0047885855,
0.0244326424,
0.0419015028,
-0.0233534463,
0.0217270125,
0.0157867782,
-0.0049233567,
-0.0085283592,
0.0328322686,
-0.0060180696,
0.0094908262,
0.0785142854,
0.0108420691,
-0.0166783407,
-0.0224701334,
-0.0065601142,
-0.0423948206,
0.0109119015,
0.0047450564,
0.0199316628,
-0.0038529185,
-0.0042420714,
0.0023859788,
0.0287352651,
-0.0179979522,
0.0495298207,
0.043005798,
0.0327181704,
-0.0372653753,
-0.002437453,
-0.0226987433,
-0.0016599944,
-0.0176939368,
0.0800595433,
0.0328378938,
-0.0130020864,
-0.0257941373,
-0.0170023441,
-0.0036331476,
-0.0677969158,
-0.0118957451,
0.0117614139,
0.010627903,
0.0504999161,
-0.077765286,
-0.0706218854,
0.0126550356,
0.0189089719,
-0.0189677253,
0.0053880308,
-0.0089491615,
-0.0052236929,
-0.0175600462,
0.0239127018,
-0.0641207024,
0.004576541,
-0.0410905145,
0.0237540305,
0.0234061461,
-0.0488565154,
-0.0569167547,
0.0077134892,
-0.0773168653,
-0.0377710797,
-0.0051840772,
0.0606869869,
0.025373036,
0.0022875969,
-0.0196342412,
-0.0211974457,
0.0313680284,
0.0405157506,
-0.0511156656,
-0.0042014369,
-0.0033997272,
-0.0001031682,
-0.0333969928,
0.0359265581,
-0.0726760626,
0.0325909443,
0.008161555,
0.0618583895,
-0.0235833339,
-0.0458118916,
0.0290265679,
0.0449112915,
0.0052818013,
0.0442900285,
-0.0260986648,
0.0056885057,
0.015003318,
-0.0236245058,
0.0017492919,
-0.0255719498,
-0.0266120862,
0.0160911735,
-0.0451974496,
0.0000828862,
-0.0289124995,
-0.0364058949,
0.0418770872,
-0.059038721,
0.0030563141,
-0.0140801687,
0.009542251,
0.011067275,
-0.0288868956,
-0.0035125448,
0.0034879267,
0.0595617518,
0.0528821461,
0.0033793806,
0.0497508906,
0.0011031072,
0.0006607593,
0.0020128775,
0.0006441628,
-0.0728819743,
-0.0743210316,
-0.0349292606,
0.0404418558,
0.0270323977,
0.0026375137,
0.001890794,
-0.0128119048,
-0.0080161905,
-0.0102155376,
-0.0069344779,
0.0331300087,
-0.0226628911,
0.0592576675,
0.0264580231,
-0.0192502029,
-0.0238145534,
0.0922123715,
0.0073321024,
0.0238792971,
-0.0177517217,
0.0087335929,
-0.036238499,
-0.0366843008,
0.0448114872,
0.0144357281,
-0.004748587,
0.0199461281,
-0.0517745167,
-0.0515742041,
0.0083658965,
0.0275181588,
-0.0285792481,
-0.0085020922,
0.0457975008,
0.068758592,
0.0104958564,
-0.0144352447,
-0.0500262715,
0.043332532,
-0.0829794705,
-0.0066266637,
-0.0380001552,
-6.81427463e-33,
-0.0339950249,
-0.0213801153,
0.0702244788,
-0.0130328201,
-0.0161135551,
-0.0074500679,
-0.0227987058,
-0.0119531443,
-0.0583770536,
-0.0194980446,
-0.0189227555,
0.0147357266,
-0.0006569112,
0.0190907642,
0.013558561,
-0.0341919512,
0.0070320121,
0.0035533246,
-0.0138050178,
0.0140116364,
0.0082821818,
0.047709845,
0.0327756554,
-0.0313238986,
0.0085616447,
-0.0353573523,
-0.0148891173,
0.0380988084,
-0.0030270591,
0.0793544576,
-0.0172889717,
0.0699604377,
0.0199910682,
-0.0660637319,
0.0136135621,
-0.0071811825,
-0.0346731953,
-0.0613454171,
0.0156703107,
-0.0183199495,
-0.0052468246,
-0.014053789,
0.0339306444,
-0.0171283484,
-0.017487146,
-0.0013555437,
0.0258915592,
-0.0109492559,
-0.0089444444,
-0.0154518085,
-0.0588556528,
0.0289509948,
0.0122530004,
0.0022240116,
-0.0236682687,
0.0846673772,
-0.0032066063,
0.0453897491,
-0.0057360367,
-0.0125111528,
0.0383060277,
0.0358235687,
0.0330830328,
0.0113638276,
-0.0079121543,
0.025689343,
0.0036017911,
0.0002596972,
0.0044743246,
-0.0190042965,
0.0141527066,
0.0392852053,
0.0512157045,
0.0723018572,
0.0189668573,
-0.0036731239,
-0.0629077479,
0.0480842106,
-0.0034415938,
0.0793019086,
0.0104513643,
0.0088980151,
-0.0533597879,
-0.0015143589,
-0.0538513996,
0.0419628061,
-0.0582470968,
-0.0218299609,
0.0455232114,
0.0019166066,
-0.0253815223,
0.0074499045,
0.0298085269,
-0.0488985777,
-0.011725314,
-0.0146665629,
-0.0246154238,
0.0037911034,
-0.0104008131,
0.0466663353,
-0.0376916528,
-0.032719329,
-0.0197495203,
-0.0043058768,
-0.0090822503,
0.0035734717,
-0.0008507626,
0.0002071118,
-0.0621177368,
-0.009337374,
-0.0185692944,
0.0277557299,
-0.024350496,
-0.0113964695,
-0.0155602489,
0.0569663979,
0.0163985882,
0.0346319191,
0.0074777803,
-0.0106065115,
-0.0226041712,
-0.0389451794,
0.0276643932,
0.0167122036,
-0.0143681653,
0.0516231805,
-0.0053499131,
0.0523045436,
-0.0042702104,
-0.0430791825,
-0.028228391,
0.0566884577,
2.817e-7,
0.0659079105,
0.0531189181,
0.0117970454,
0.062667571,
0.0020110593,
0.0276174471,
-0.0014724121,
0.0081800418,
0.0524543077,
0.0138648245,
0.0140264016,
0.0255041234,
0.0527996905,
0.0024236443,
-0.0716636628,
-0.0038912843,
-0.0758204386,
-0.0154325282,
-0.061844226,
0.0219842084,
0.0555754453,
0.0781823844,
-0.0365376063,
-0.0396558456,
-0.023719212,
-0.0811850801,
0.066604279,
-0.0076229312,
-0.0654994175,
0.0378657989,
0.0314026102,
0.0417140499,
0.0293434709,
0.0167308096,
-0.0049782884,
0.0275209118,
0.0199401788,
0.0297385342,
0.0048602666,
-0.0328237824,
-0.0175690427,
-0.015736999,
0.0203249734,
-0.0730459988,
0.0880105644,
-0.0435031503,
-0.0199500099,
-0.0438100696,
-0.044139199,
-0.0219531823,
0.0167460926,
-0.0425885879,
-0.0207063965,
0.0172517747,
0.0234730858,
-0.0114533231,
0.0186124034,
-0.0264746435,
-0.0109226117,
0.0184903722,
-0.0437833965,
-0.0203886125,
-0.0349545553,
0.0137288859,
-0.010525493,
-0.0363307483,
0.0193742327,
2.615001748e-34,
0.0559007041,
-0.0082876394,
0.0080673425,
0.0440581478,
0.0602627844,
0.0203314591,
-0.0329545103,
0.0383931212,
0.0173664372,
-0.1123601273,
-0.0181071218
] |
http://arxiv.org/abs/2401.15241v2 | http://arxiv.org/pdf/2401.15241v2 | Unlearning Traces the Influential Training Data of Language Models | 2024-01-26T23:17:31 | 2024-06-13T16:28:47 | Masaru Isonuma, Ivan Titov | cs.CL | cs.CL, cs.AI | Identifying the training datasets that influence a language model's outputs
is essential for minimizing the generation of harmful content and enhancing its
performance. Ideally, we can measure the influence of each dataset by removing
it from training; however, it is prohibitively expensive to retrain a model
multiple times. This paper presents UnTrac: unlearning traces the influence of
a training dataset on the model's performance. UnTrac is extremely simple; each
training dataset is unlearned by gradient ascent, and we evaluate how much the
model's predictions change after unlearning. Furthermore, we propose a more
scalable approach, UnTrac-Inv, which unlearns a test dataset and evaluates the
unlearned model on training datasets. UnTrac-Inv resembles UnTrac, while being
efficient for massive training datasets. In the experiments, we examine if our
methods can assess the influence of pretraining datasets on generating toxic,
biased, and untruthful content. Our methods estimate their influence much more
accurately than existing methods while requiring neither excessive memory space
nor multiple checkpoints. | Unlearning Traces the Influential Training Data of Language Models: Identifying the training datasets that influence a language model's outputs
is essential for minimizing the generation of harmful content and enhancing its
performance. Ideally, we can measure the influence of each dataset by removing
it from training; however, it is prohibitively expensive to retrain a model
multiple times. This paper presents UnTrac: unlearning traces the influence of
a training dataset on the model's performance. UnTrac is extremely simple; each
training dataset is unlearned by gradient ascent, and we evaluate how much the
model's predictions change after unlearning. Furthermore, we propose a more
scalable approach, UnTrac-Inv, which unlearns a test dataset and evaluates the
unlearned model on training datasets. UnTrac-Inv resembles UnTrac, while being
efficient for massive training datasets. In the experiments, we examine if our
methods can assess the influence of pretraining datasets on generating toxic,
biased, and untruthful content. Our methods estimate their influence much more
accurately than existing methods while requiring neither excessive memory space
nor multiple checkpoints. | [
0.0160633419,
0.1098901108,
-0.0257008485,
0.0549961887,
-0.0049774311,
0.0274822544,
0.0110487808,
0.0201921239,
0.0022175026,
-0.025470946,
-0.0236672759,
0.0104039786,
0.0018139674,
0.0317978784,
0.0400305875,
-0.0068985187,
0.0523644574,
-0.0044782502,
-0.0144913848,
-0.0091007184,
0.0290968437,
0.015618152,
0.0279253814,
0.0214164462,
-0.0271614902,
-0.0336101651,
0.01316074,
0.0117058791,
0.0075118253,
-0.0061400938,
0.044486057,
0.0328722708,
0.0173224565,
0.0738690272,
0.0000020118,
-0.0045965263,
-0.0350001305,
0.0043214774,
-0.0365230925,
0.0453361087,
0.0442490093,
0.0165564083,
-0.0236765295,
0.0452201031,
-0.0299556982,
-0.0218040682,
0.0399380736,
0.0137445899,
0.0543763191,
0.0898879319,
0.0012427261,
0.0024737003,
-0.0141074806,
-0.0039671902,
-0.0089541776,
0.0229460616,
0.0511935689,
-0.0282367207,
-0.0912682042,
-0.0095485495,
-0.026900975,
0.0027894718,
0.0165384728,
-0.017557634,
0.0019009594,
-0.0103503466,
-0.0380799286,
-0.0320315436,
0.0175720658,
0.0568807051,
-0.0600732267,
-0.0651211739,
0.0072059808,
-0.0167155359,
0.0085788174,
0.0046122791,
-0.0055223959,
0.0133323753,
0.0035688123,
0.0313388482,
-0.0024974397,
-0.026715003,
0.0675560832,
0.0364672728,
0.0236299541,
0.0316354446,
0.0174237378,
-0.016625613,
0.012863433,
0.0144970985,
0.0311940815,
-0.049138356,
-0.0004174084,
-0.0257403739,
-0.0110463919,
0.0395839512,
-0.0336789265,
-0.0348504372,
0.0140943434,
-0.0599426292,
0.0442439355,
0.0007506193,
-0.0295776911,
0.0180381332,
-0.0229803585,
0.0165447295,
-0.0012843922,
0.0267747976,
-0.0161127783,
0.0252293479,
0.0182214063,
0.0057743532,
-0.003285228,
0.0067967731,
0.0200554784,
0.0431606025,
-0.0390331782,
0.0003772425,
-0.0384915136,
-0.0025615764,
-0.0064883716,
0.0260532741,
-0.0339076184,
0.0363979414,
-0.0152382525,
-0.0057383804,
-0.037799295,
0.0197409168,
-0.0140431924,
-0.0831732973,
-0.0279592592,
0.0246784333,
0.0027542121,
-0.0111111561,
0.0299238879,
0.0643366873,
0.0101248706,
-0.0341276489,
-0.0519795902,
-0.0550036877,
-0.0017045717,
-0.0630837902,
-0.0044732974,
-0.0584637485,
-0.0094678458,
-0.0204981696,
-0.0175336096,
-0.0807854906,
-0.0034843103,
0.0183735862,
-0.0546560735,
0.0402942859,
-0.0182472542,
-0.0243783426,
0.0189501997,
0.0241849869,
-0.0285100192,
0.077279076,
-0.0227951054,
0.0638275594,
0.0424506813,
-0.0110293115,
0.017640112,
0.0431454256,
0.0842704996,
0.0392169468,
-0.0406163186,
0.0049865693,
0.0173635837,
-0.0195075572,
0.0124315536,
0.0274854861,
-0.0227076504,
-0.0257750936,
0.0965846777,
-0.0205574054,
-0.0173898786,
0.0289599374,
0.029323481,
-0.02102229,
0.0151651939,
0.0230896771,
-0.0690451115,
-0.0415061004,
-0.0388860963,
-0.028346641,
0.0259828791,
0.0406294502,
-0.0135392724,
-0.0429639556,
-0.0015659394,
-0.0197314844,
0.0361834019,
-0.0048979861,
0.0329879522,
0.0553713888,
-0.0313131623,
0.0420933031,
0.076487653,
-0.0702016354,
0.0108266659,
-0.0308144614,
0.0539790206,
0.0201797932,
0.0722521544,
-0.0147411199,
0.0084435605,
-0.0342199542,
0.0164084192,
0.0003781976,
0.0170939751,
-0.0280045439,
0.0347646326,
-0.0198370293,
-0.0117714917,
-0.000291497,
0.0286611486,
0.0207624082,
0.0725745484,
0.0123622743,
0.0149634071,
-0.0290974602,
-0.0780065358,
0.0457380079,
-0.0176271535,
-0.0212355144,
0.0046147532,
-0.0004431407,
0.0393057019,
0.0029694622,
-0.0460686795,
-0.0317956656,
0.0408831425,
0.0146478666,
-0.0183038991,
-0.0130224638,
-0.0114241662,
-0.0302513242,
0.0375289917,
-0.0558296777,
-0.0104833525,
0.0690542459,
-0.0323601216,
0.0011240022,
0.0454997048,
-0.0267816577,
0.0480305441,
-0.0272926539,
0.0126470625,
-0.0699474961,
0.033492554,
0.0138629638,
0.0627842546,
-0.0029758764,
0.025360588,
0.0225550495,
-0.0043223943,
-0.013508724,
0.0360764787,
-0.0284055434,
-0.0057083624,
-0.0155605469,
0.0094854329,
0.0454459302,
-0.0109035494,
-0.0264242291,
0.0198838226,
-0.0515598208,
-0.019509038,
-0.0247558951,
0.0134715708,
0.0354041085,
-0.0042966343,
-0.0133759584,
-0.0184214972,
-0.0242225025,
-0.0255435463,
-0.0001412571,
0.0309744552,
-0.0045244494,
-0.0299166702,
0.004907561,
-0.0437647924,
-0.0025378964,
0.0025574053,
0.0781393871,
0.0055258148,
-0.0041802349,
-0.0203222074,
0.0331937782,
-0.0457204133,
0.0311921518,
0.0024600597,
-0.0669049993,
0.0082295043,
0.0190711692,
-0.0131598925,
0.0578501523,
0.0162591394,
0.0628818274,
0.0588574968,
0.030297501,
-0.0546577647,
-0.0413318686,
-0.0314866081,
-0.0139021315,
-0.0159440339,
-0.0265754834,
-0.0227258597,
0.0693288594,
0.0306865107,
-0.0400485918,
-0.0107149314,
0.0432891771,
-0.0456138887,
-0.04222643,
-0.0039728028,
-0.051431708,
-0.0318866335,
0.0289564207,
-0.0109536154,
-0.0960309729,
0.0603990816,
-0.0112988614,
0.0256049689,
-0.0234847236,
-0.0186263174,
0.0088733034,
0.0299396347,
0.0269198976,
0.0334830545,
0.0228105914,
0.0005120863,
-0.0231200978,
-0.0116534205,
0.0751972347,
-0.0293684956,
-0.0026403761,
-0.0450432673,
0.0165425614,
0.0216138661,
-0.0446979813,
-0.063862443,
-0.0683686882,
0.030790085,
-0.0378829129,
0.0209187604,
-0.0364319496,
-0.0291283168,
-0.0631078333,
0.0126131149,
0.0179898739,
-0.0222923998,
0.0089206882,
0.0135836713,
0.0153186321,
0.0448895246,
-0.059167508,
0.0335554741,
-0.04462431,
-0.0174976178,
0.0828658044,
-0.0067361738,
0.0280856546,
-0.0266316906,
-0.0044039967,
-0.0175407287,
0.0127686895,
0.0218584388,
0.0167326108,
0.0252246615,
-0.038518928,
-0.0385818295,
0.0825631395,
0.0179526582,
-0.0311342329,
0.023278309,
0.0470329598,
0.0298711061,
0.0291572679,
-0.0078676865,
-0.0918757319,
-0.0125785181,
-0.0009445351,
0.0193835888,
-0.04757699,
-0.0476810485,
0.0006568631,
-0.1045760959,
0.0069697169,
0.0094377063,
-0.0222740527,
-0.0158322807,
-0.0678357184,
-0.0221741479,
-0.0324024372,
-0.0025960305,
-0.0219179112,
0.0231940262,
0.0192671977,
0.0038272776,
-0.022187097,
-0.062292859,
0.0139935417,
-0.0245234314,
0.0364292823,
0.0453496762,
0.0271982551,
-0.0286582578,
0.0014578672,
0.0562736765,
0.0211033393,
-0.0279363673,
-0.0206446145,
0.0844819024,
-0.002692061,
-0.0269117374,
0.0010500128,
0.049391333,
-0.0043689287,
0.012439942,
-0.0074019386,
0.0464691482,
0.0262822751,
-0.0569791198,
-0.0101945298,
-0.0319109298,
0.0420610569,
0.0783273131,
0.0294140317,
-0.0491342247,
0.0091705993,
-0.0479363985,
0.0064825439,
0.0233684853,
-0.035129942,
0.0166359432,
0.005416818,
-0.024185434,
-0.0754156932,
-0.0185203161,
-0.0069309701,
0.0014582745,
0.0182904527,
0.0020321519,
-0.0095283221,
-0.0041145002,
-0.048464708,
0.0245019495,
-0.0170002021,
0.0127878077,
-0.0087142838,
-0.0358957127,
-0.0173819996,
0.0136749018,
-0.0146672353,
0.0312019624,
-0.1180184186,
-0.0928777382,
0.0303878672,
0.0641621649,
0.0160266925,
0.0644863024,
0.014188461,
-0.0445791371,
0.0890579,
0.01487255,
-0.0493717305,
0.0143982703,
0.037401069,
-0.0625461563,
0.0318273716,
0.0403611362,
-0.0227369852,
0.0181215573,
0.0665758401,
-0.0567446351,
-0.0298033915,
0.003783647,
0.0180194471,
0.0056247362,
0.0635144338,
-0.0262036901,
-0.0210752841,
-0.0451100282,
-0.000576126,
-0.0263356641,
-0.0412289351,
-0.0224121138,
0.035213545,
0.0402857699,
0.0089143142,
-0.0225914419,
0.0022414629,
-0.0288007222,
0.0251379982,
-0.0299581066,
0.0010345683,
0.022333825,
0.0361187905,
0.067098923,
0.0199771579,
-0.0175661799,
0.0216823071,
-0.0286440682,
0.0237102043,
-0.0070841028,
0.0013233509,
0.0287853647,
0.0243662354,
0.0359758958,
-0.0156071046,
-0.030218048,
0.0045880959,
-0.0664454699,
0.0024914704,
0.0229027793,
-0.0339050107,
0.050201159,
0.0017001369,
-0.0072268969,
-0.0175596643,
-0.0028164035,
-0.0188781284,
-0.0425758623,
0.0040346431,
-0.0115751233,
-0.055354055,
0.0478944555,
0.0690822601,
0.0122006051,
0.0738445967,
0.0158681162,
-0.033462774,
-0.0245559718,
-0.0124268094,
0.0321082957,
0.0115501117,
-0.0120421574,
-0.0019114837,
0.008234431,
-0.0805481151,
-0.037191622,
0.0096134441,
0.0271091294,
0.0197873209,
0.0429170132,
0.0475726612,
-0.0099825151,
0.0248668715,
-0.0572710522,
-0.0037934936,
-0.002840112,
-0.0405417718,
0.0012025585,
-6.543477466e-33,
0.0340299867,
-0.028390659,
-0.0412069522,
0.0220251922,
0.0174030326,
0.0171904545,
-0.0191861968,
-0.0125215119,
-0.0556482375,
0.0223165918,
-0.0005214299,
0.0512375347,
-0.0024431876,
0.0284959693,
-0.0234444924,
-0.0118803782,
0.0592398793,
-0.028131973,
-0.053073179,
0.0436134934,
0.0463879108,
0.0392573476,
0.0752201602,
-0.0878157392,
0.0187403541,
-0.0251901578,
-0.0435050651,
0.0088879773,
-0.0363021418,
0.0348625295,
-0.0021273685,
0.0439588688,
0.0007784514,
-0.0580687039,
-0.0163267553,
-0.0117286518,
-0.0102258204,
-0.0273821801,
0.0112821003,
0.0226816088,
-0.053298302,
-0.004597866,
0.0403096713,
-0.0095451623,
-0.001931067,
-0.0370408595,
0.0178949311,
0.0031273181,
-0.0142895682,
-0.0068163052,
-0.0454102345,
0.0304648113,
-0.0207435656,
-0.0115343556,
0.0104139885,
0.0250693299,
0.0322832316,
-0.0243928768,
-0.063791424,
0.0011321475,
0.0241293516,
0.0341151208,
0.0391336158,
-0.0117582837,
0.0090908362,
0.0929192677,
0.0457936674,
0.0259531997,
0.0160501823,
-0.006590073,
0.0270932689,
0.0159439612,
0.0393447913,
0.0646303669,
0.050828483,
-0.0545357652,
-0.0886305794,
0.0022250824,
0.020688694,
-0.0002350573,
0.0010800044,
-0.0116045019,
-0.0096111316,
-0.0332901031,
-0.0559761114,
-0.0370984115,
-0.0236496553,
-0.0305433106,
0.0132290069,
0.0104892291,
-0.0489363298,
-0.0115836496,
-0.0088373553,
-0.070339255,
-0.0126952743,
0.0520924926,
-0.0423944592,
-0.0298305154,
0.0142981075,
0.0127346311,
-0.0510039106,
0.0126004191,
0.0061220122,
-0.023718778,
0.0416263528,
0.005162599,
0.0845475942,
0.0439940467,
-0.0541728586,
0.0253493246,
-0.0569542758,
0.0285358336,
0.0359985121,
0.0358867161,
0.036473874,
0.004420938,
-0.0097573539,
-0.0003069635,
0.0610404015,
0.0411223173,
-0.0588386506,
0.000006775,
0.0065659904,
0.0147725549,
-0.0420738719,
-0.0643633679,
-0.0329051353,
0.0456982031,
0.0759975314,
-0.0877446085,
0.0115517471,
0.0000497255,
2.685e-7,
0.0032095888,
0.0561511554,
0.0002321537,
0.0223911274,
0.0403002873,
0.0382672995,
0.0101762079,
0.0367898606,
-0.0063140681,
-0.0265997648,
-0.0223562121,
0.0052113519,
0.0557776093,
0.0505559109,
-0.0706326216,
0.006226195,
-0.1124476418,
-0.0173814856,
-0.0550889,
-0.0379450917,
-0.0071917702,
0.1196109429,
0.0487064905,
-0.0396253392,
0.0025598535,
-0.0033306675,
0.0387824066,
-0.0208780076,
-0.010812331,
-0.0081218509,
0.0132950665,
0.0271413662,
-0.0196696818,
-0.0102276038,
-0.0620205104,
-0.0008679453,
-0.0249695927,
0.0060547749,
-0.0005384542,
0.0054533561,
-0.0109862052,
-0.0098638888,
-0.0125536509,
-0.0470253266,
0.0774754211,
0.0338581838,
-0.015711654,
-0.0837537423,
0.0059245611,
-0.0338228755,
0.070110172,
-0.0234133787,
-0.0002288026,
0.0485025011,
0.0049673207,
0.0119201345,
-0.0220495276,
-0.0531827994,
-0.0508962572,
-0.0268499162,
-0.0339146145,
-0.0592996106,
-0.0120862918,
0.002868636,
-0.0060081938,
-0.062425971,
0.0447833985,
2.324893645e-34,
0.011075967,
-0.043824099,
0.0254854839,
-0.0428536646,
0.0466775969,
-0.0162144098,
0.0262091514,
-0.0156126283,
-0.007954102,
-0.0240245517,
-0.0105287125
] |
http://arxiv.org/abs/2401.15222v2 | http://arxiv.org/pdf/2401.15222v2 | Transfer Learning for the Prediction of Entity Modifiers in Clinical Text: Application to Opioid Use Disorder Case Detection | 2024-01-26T22:19:31 | 2024-02-05T17:13:41 | Abdullateef I. Almudaifer, Whitney Covington, JaMor Hairston, Zachary Deitch, Ankit Anand, Caleb M. Carroll, Estera Crisan, William Bradford, Lauren Walter, Eaton Ellen, Sue S. Feldman, John D. Osborne | cs.CL | cs.CL, cs.AI, cs.LG | Background: The semantics of entities extracted from a clinical text can be
dramatically altered by modifiers, including entity negation, uncertainty,
conditionality, severity, and subject. Existing models for determining
modifiers of clinical entities involve regular expression or features weights
that are trained independently for each modifier.
Methods: We develop and evaluate a multi-task transformer architecture design
where modifiers are learned and predicted jointly using the publicly available
SemEval 2015 Task 14 corpus and a new Opioid Use Disorder (OUD) data set that
contains modifiers shared with SemEval as well as novel modifiers specific for
OUD. We evaluate the effectiveness of our multi-task learning approach versus
previously published systems and assess the feasibility of transfer learning
for clinical entity modifiers when only a portion of clinical modifiers are
shared.
Results: Our approach achieved state-of-the-art results on the ShARe corpus
from SemEval 2015 Task 14, showing an increase of 1.1% on weighted accuracy,
1.7% on unweighted accuracy, and 10% on micro F1 scores.
Conclusions: We show that learned weights from our shared model can be
effectively transferred to a new partially matched data set, validating the use
of transfer learning for clinical text modifiers | Transfer Learning for the Prediction of Entity Modifiers in Clinical Text: Application to Opioid Use Disorder Case Detection: Background: The semantics of entities extracted from a clinical text can be
dramatically altered by modifiers, including entity negation, uncertainty,
conditionality, severity, and subject. Existing models for determining
modifiers of clinical entities involve regular expression or features weights
that are trained independently for each modifier.
Methods: We develop and evaluate a multi-task transformer architecture design
where modifiers are learned and predicted jointly using the publicly available
SemEval 2015 Task 14 corpus and a new Opioid Use Disorder (OUD) data set that
contains modifiers shared with SemEval as well as novel modifiers specific for
OUD. We evaluate the effectiveness of our multi-task learning approach versus
previously published systems and assess the feasibility of transfer learning
for clinical entity modifiers when only a portion of clinical modifiers are
shared.
Results: Our approach achieved state-of-the-art results on the ShARe corpus
from SemEval 2015 Task 14, showing an increase of 1.1% on weighted accuracy,
1.7% on unweighted accuracy, and 10% on micro F1 scores.
Conclusions: We show that learned weights from our shared model can be
effectively transferred to a new partially matched data set, validating the use
of transfer learning for clinical text modifiers | [
0.0137647232,
0.0763119981,
-0.0078378962,
-0.0353179611,
-0.0165049173,
0.0306803714,
0.0040667476,
0.0028925349,
-0.0025219249,
-0.0215731468,
-0.0329481401,
-0.0296008307,
-0.0483238548,
0.044146359,
-0.026382383,
0.0030532714,
0.0357257091,
0.0464021936,
-0.0227289777,
0.0220878795,
0.0012371551,
0.0121786706,
-0.0056316326,
0.0293518528,
-0.0021409597,
0.0080781886,
0.021682661,
-0.003390085,
0.0271623787,
-0.0582661629,
0.0270781852,
0.0662548766,
-0.0070436639,
-0.027397627,
0.0000018635,
-0.0501972027,
-0.0660182163,
0.010909575,
-0.0157557726,
-0.0774271786,
0.0591058433,
-0.053162232,
0.0023793287,
0.0099226218,
-0.045074515,
-0.035747949,
0.0722408667,
0.0865760222,
0.0388441943,
0.0637328923,
-0.0611225367,
0.0341070108,
0.0063124467,
-0.0182657056,
-0.0054653659,
-0.0373654813,
0.0862553045,
-0.0024746917,
-0.0208927281,
-0.0252507012,
-0.0026009483,
0.0201205313,
-0.036700502,
0.0127491467,
0.02455681,
0.0286336411,
0.0073109982,
0.0009228482,
-0.0037805457,
0.079536207,
0.0037058871,
0.0075677652,
0.0414858125,
0.0032286507,
-0.0178949162,
0.0397715606,
-0.0168042369,
-0.0463098697,
-0.0027847805,
0.0476145595,
0.0088602835,
-0.0189992525,
0.0023746695,
-0.0148550114,
-0.0209652446,
-0.0172809977,
0.0223273039,
0.0122219166,
0.0264469627,
0.0072611999,
-0.0477096513,
-0.0390475504,
0.0528477281,
0.0116071776,
-0.0619746149,
0.0287191365,
-0.0694597587,
-0.0991617963,
0.0186661258,
-0.0172175169,
0.0098009892,
0.0088902619,
-0.0407922342,
0.0175553802,
-0.0415311791,
-0.0081438571,
-0.0050724288,
0.0394491889,
-0.0470936522,
-0.0384247601,
-0.0873993188,
0.0121166399,
-0.0193978846,
0.061108727,
0.0076825055,
0.0031630902,
-0.050510373,
0.0094467737,
-0.0241043754,
-0.0393501446,
0.0194160398,
0.0398564339,
-0.037463408,
0.0117633156,
0.0349980406,
0.0552688576,
0.0338829309,
0.0414488986,
-0.0189407784,
-0.005131749,
-0.0087025501,
0.0030608685,
0.0334798843,
-0.0476987064,
0.0319210999,
0.0318558328,
-0.0073485724,
-0.0144272437,
-0.0204090104,
-0.0233401749,
-0.0383443125,
0.0348212421,
0.0118983537,
-0.0265314281,
0.0134942532,
0.0090493988,
0.0177570805,
-0.0675892606,
0.0200898796,
-0.0228015129,
-0.0541403256,
0.0729398653,
-0.0243071914,
0.0063768867,
-0.00924954,
0.0107907653,
-0.0263580736,
0.0029404552,
0.0351885743,
0.0209092274,
0.0178046897,
-0.0287240688,
-0.00537267,
0.0249043629,
0.0077369036,
-0.0489682294,
0.0213126149,
-0.0131866001,
0.0189348049,
0.0032085022,
-0.0343525633,
0.065604575,
0.0246840473,
-0.0070328359,
0.0388649218,
0.0148579786,
0.0317326821,
-0.0012309876,
0.0063142078,
-0.0028729166,
-0.0386579745,
-0.0212897658,
-0.037558198,
-0.0091151819,
0.0076885461,
-0.0493751653,
-0.029161673,
-0.0215246566,
-0.0064100409,
0.0038364751,
0.0119621456,
-0.0379584916,
-0.0362596437,
-0.0191886704,
0.0653162524,
-0.003432354,
-0.0500970297,
0.0764847621,
0.0466242656,
-0.0137464395,
-0.0219781585,
-0.0331083052,
-0.0353366472,
0.0253454708,
0.0198952761,
-0.0533926487,
-0.0701253489,
-0.0363747589,
0.0101601472,
-0.0061667021,
-0.005154164,
-0.016447816,
0.0097842868,
-0.0112363147,
0.0095289126,
0.0620354377,
-0.0017385718,
-0.0315838084,
-0.0210372731,
0.0386576504,
-0.006865276,
-0.0194830112,
-0.0052992445,
0.0736162215,
-0.0342305489,
-0.0547262989,
0.0315497704,
-0.0044425619,
0.0517877042,
-0.0385719985,
-0.0119082993,
0.0148846423,
0.0331036933,
-0.038498871,
-0.0415985286,
-0.0013962479,
-0.0204816107,
-0.0019147213,
0.0184716564,
-0.0220738836,
0.0512524657,
0.034685079,
-0.0242511313,
-0.0288315639,
0.0155905243,
-0.0286488477,
0.0092493454,
-0.0075134207,
0.0089539764,
0.0296781976,
0.0629536882,
-0.0372405052,
0.0235638563,
0.0193467848,
-0.020657843,
-0.0272269044,
-0.073507458,
-0.0004865301,
0.0134359719,
0.0028562834,
-0.0460404754,
-0.0522560216,
-0.0347538926,
-0.0533371642,
-0.0089625912,
0.0045743547,
-0.0291213337,
-0.111808151,
0.0073943259,
-0.0021454466,
0.0436425395,
0.0316620655,
0.0123452265,
-0.0051246914,
-0.0243171919,
-0.0213163514,
-0.0427601226,
0.0435272716,
0.008174805,
0.0099419923,
0.0256530214,
0.0180482417,
-0.0137079842,
0.0037267401,
-0.035231255,
0.0184081383,
-0.0496537015,
-0.0707206577,
-0.0238168128,
0.0166938752,
0.0082306769,
0.0040879212,
0.0060977112,
0.0319599062,
-0.0022329702,
0.0306858737,
0.0621336475,
0.0213155057,
-0.0056729498,
0.0413339734,
0.0429105237,
-0.0302627422,
0.0058757961,
0.0162308291,
-0.0027339528,
-0.0177580807,
-0.0282155778,
-0.0618438832,
0.0108972434,
0.0947839096,
0.0053354627,
-0.0279163644,
-0.0248617884,
-0.0254132561,
0.0039206292,
-0.0110133383,
0.0079509197,
-0.0193106234,
-0.020799147,
-0.0370526984,
-0.0227746218,
-0.02146722,
0.0478225537,
0.0100249248,
0.0236418527,
-0.0668916479,
-0.0290187784,
-0.0260257553,
-0.0129312156,
0.0127417287,
0.0135179525,
0.061852172,
0.004035376,
0.0292202029,
0.0241337679,
0.008062386,
-0.0596088208,
0.0502693281,
-0.0257528517,
0.0316700414,
-0.0133075165,
-0.0437146649,
-0.0372266322,
-0.0347464941,
-0.0093146628,
0.0320520625,
0.0462899096,
-0.0020929275,
-0.0542664751,
0.0231148824,
-0.0473380424,
0.0341148488,
0.0180122573,
-0.013400591,
0.0051607634,
0.0173167456,
0.0646467283,
0.0310900658,
-0.0112400912,
-0.0120663317,
-0.0051091481,
0.0601180531,
0.0097916014,
0.006755841,
0.0055443696,
0.014247505,
0.0323686078,
-0.0551151149,
-0.0151665267,
0.010014493,
0.0376626216,
-0.003350206,
-0.0031053135,
0.0670736581,
0.0359677039,
-0.0482069142,
0.0309077185,
-0.0220054183,
-0.0390655771,
-0.0139593845,
0.0137794502,
-0.0666569099,
-0.0566328205,
-0.0043087942,
0.0195787381,
-0.0391312912,
-0.0581694543,
0.0082996916,
-0.0190951172,
0.0183901712,
-0.0434205569,
-0.0586910658,
0.0141991582,
-0.0666241571,
0.0362396985,
-0.0498831831,
-0.0474828668,
-0.044569321,
-0.0417248905,
-0.0201962702,
0.033353705,
-0.0198274199,
-0.0685122311,
0.0180825833,
-0.0122050401,
0.0543886386,
0.0130937295,
0.0082717659,
-0.024261035,
-0.0376514457,
0.0037230037,
0.0179214682,
0.0042061964,
0.0063840779,
0.0931979865,
-0.0162820853,
-0.0082124537,
-0.0654555485,
0.0243830308,
-0.0478441417,
-0.0667353347,
-0.0131349275,
0.1116063371,
-0.0109212091,
-0.0089826724,
-0.0305675659,
-0.0119541949,
0.0489252768,
0.0480184108,
-0.0388517268,
0.0062261974,
-0.018682044,
0.0437372364,
-0.0057709534,
-0.0303123109,
-0.006260186,
0.0152404699,
-0.0523115098,
0.0035519828,
-0.0623887628,
0.0164594501,
0.0545342974,
-0.0192451179,
0.0414202809,
-0.0059967795,
0.0129469568,
-0.0164363515,
-0.0812625587,
0.0546991415,
0.0010163329,
0.022383444,
-0.0183542967,
-0.0172941331,
0.0350065157,
0.0202395525,
-0.0158124324,
0.0027792403,
-0.0553534664,
0.0045724465,
0.0308754258,
0.0058962889,
0.0085192556,
-0.0380221829,
-0.0208228137,
0.0043542041,
0.0115657914,
0.0337080285,
-0.054927282,
-0.0447692834,
0.0202795248,
-0.0634698272,
-0.0223543532,
0.0319437012,
0.0157226231,
0.0447614901,
0.0505322292,
-0.0092591243,
-0.0118441591,
-0.012699333,
0.0219743177,
0.0368603133,
0.0379514769,
0.0252148975,
0.0030901446,
-0.0438906327,
0.0127917817,
0.0078399768,
0.0272796154,
-0.0262498185,
-0.0023115366,
0.039415095,
-0.0072823497,
-0.043309506,
0.0899899751,
-0.0567711964,
-0.0150265386,
-0.0641389787,
0.0421687849,
0.016777074,
0.0066180932,
0.0213189479,
-0.0027398532,
0.0252006166,
-0.0441691093,
0.0588875972,
0.053290356,
-0.0306430589,
0.0464401357,
-0.0096429987,
0.0506704859,
-0.0168040879,
0.0529705174,
-0.0377280228,
0.0016875714,
-0.0367712714,
-0.0167290084,
0.0330847092,
-0.0244549643,
0.0531541482,
-0.0211888608,
0.0133166583,
-0.0358435325,
0.0257087424,
-0.045982331,
0.0465622544,
0.0381475426,
0.0013870558,
-0.0014582281,
0.0013978551,
0.086838074,
-0.009729608,
0.0313230716,
-0.0013044957,
-0.049849242,
-0.0428618789,
0.0040897783,
0.0365781896,
-0.0207963847,
-0.0622233003,
0.0077187447,
-0.0187897906,
-0.0256787017,
0.0646706149,
-0.1033352986,
-0.0088475402,
0.0002069197,
0.0139284981,
0.035019137,
0.0104567446,
-0.0645488277,
-0.0146723213,
0.0153480396,
0.0206910744,
-0.0126619451,
-0.0143975355,
-5.786657469e-33,
-0.0078821555,
-0.0403209589,
-0.0045001325,
-0.0266124681,
-0.0131443106,
0.0110749668,
-0.0009793263,
-0.0029911157,
0.0082009099,
0.0174748898,
-0.018149592,
0.0065880613,
0.0336019062,
0.0349319652,
0.0150229605,
-0.0272571705,
0.0398663878,
0.0314003192,
-0.0103189116,
0.0112956697,
0.0236133654,
0.1054606289,
0.0949980095,
-0.1317469925,
0.0413589887,
0.0575871468,
0.0140340058,
-0.0047217668,
-0.0019764134,
0.0360621847,
-0.0304478277,
0.0447420031,
0.0551685989,
-0.0549103618,
0.0005504126,
-0.0025344826,
-0.0672472268,
-0.0059186891,
0.0176127218,
-0.0424474329,
-0.012852856,
-0.0348773077,
0.047407452,
-0.0029861112,
0.020678509,
0.0024458377,
0.0420110114,
-0.0508972406,
0.0245834757,
-0.0114192823,
-0.064895466,
0.0491700694,
0.0135893915,
-0.0033079372,
-0.0159230549,
0.0186608322,
-0.0115595162,
0.0186429359,
-0.0240308624,
-0.0128141437,
-0.0137001909,
0.0298482105,
0.0022072906,
-0.0167069547,
-0.0327792875,
0.0581352077,
-0.0101241544,
-0.018113615,
-0.0306540877,
0.0100371204,
-0.0346313715,
0.1058659106,
0.0375428833,
0.0345672369,
0.0422755927,
-0.0174821932,
-0.0846540406,
0.0237631164,
-0.0681278482,
0.0576506965,
0.0131319547,
-0.0087350421,
-0.0011853297,
-0.0008877741,
-0.0144062312,
0.022016516,
-0.008676419,
-0.0248284582,
0.0007509905,
-0.0052205608,
-0.0357702337,
0.0639454126,
0.0130359251,
-0.0425570756,
-0.0126960948,
-0.0653248429,
0.0558894835,
0.0004902065,
-0.0194428694,
-0.0190717615,
-0.0615847632,
-0.0156564135,
0.0731997415,
-0.0308147874,
0.0473572761,
-0.0116669014,
0.0552941039,
-0.0052863867,
-0.0375470966,
-0.0216092207,
0.0363456756,
0.0241784081,
-0.0087615699,
0.1153292879,
0.0213617403,
0.0576106422,
-0.0067119785,
-0.008111666,
0.0095993187,
0.0000444519,
-0.0515865274,
-0.0291752387,
0.0302018579,
0.0057953629,
-0.0171449967,
0.0039634812,
-0.0341946222,
0.0060535339,
0.0159982964,
-0.0536042154,
0.0043046903,
-0.0127234953,
2.837e-7,
0.1084393188,
0.0444400981,
-0.0169266704,
-0.0412860848,
0.0095415134,
0.0090225749,
-0.0099479249,
0.0541005395,
0.0178237278,
-0.0014721414,
0.0337729715,
0.0394947566,
0.0378743485,
0.0720616281,
-0.0234006476,
-0.0609547198,
-0.0113328453,
-0.0210410375,
-0.0072520315,
-0.004130309,
0.0546327308,
0.0690996498,
-0.0105622383,
-0.0331376716,
0.0360509269,
-0.0106763178,
0.010030698,
-0.0409878902,
-0.003586757,
0.0504515134,
0.0053856983,
-0.0357856937,
0.0082883453,
0.103420876,
-0.0228718854,
0.0073165311,
-0.0158735476,
0.0816984549,
-0.0663946718,
-0.0162033103,
-0.0136640426,
-0.0154955965,
0.0336343199,
0.0073138843,
0.0917009115,
-0.0459193513,
-0.0369144604,
-0.0266026743,
-0.0333881304,
0.0290019885,
0.0320706405,
0.0013124055,
-0.0116000548,
0.0322589502,
0.043788638,
0.0504203923,
-0.0519895777,
-0.054967206,
0.003710917,
0.060351748,
-0.0375884958,
-0.0116934506,
0.01227092,
-0.024984872,
0.0231824704,
-0.0858533457,
0.0312211774,
2.408873304e-34,
-0.0185365882,
0.0269046631,
0.0071903761,
0.0351196788,
0.0130848996,
-0.019855788,
-0.0300317369,
-0.0028933231,
-0.0260411128,
-0.0694078803,
0.0108960532
] |
http://arxiv.org/abs/2401.15207v3 | http://arxiv.org/pdf/2401.15207v3 | HiFT: A Hierarchical Full Parameter Fine-Tuning Strategy | 2024-01-26T21:14:32 | 2024-06-17T10:35:06 | Yongkang Liu, Yiqun Zhang, Qian Li, Tong Liu, Shi Feng, Daling Wang, Yifei Zhang, Hinrich Schütze | cs.LG | cs.LG, cs.CL | Full-parameter fine-tuning has become the go-to choice for adapting language
models (LMs) to downstream tasks due to its excellent performance. As LMs grow
in size, fine-tuning the full parameters of LMs requires a prohibitively large
amount of GPU memory. Existing approaches utilize zeroth-order optimizer to
conserve GPU memory, which can potentially compromise the performance of LMs as
non-zero order optimizers tend to converge more readily on most downstream
tasks. In this paper, we propose a novel optimizer-independent end-to-end
hierarchical fine-tuning strategy, HiFT, which only updates a subset of
parameters at each training step. HiFT can significantly reduce the amount of
gradients and optimizer state parameters residing in GPU memory at the same
time, thereby reducing GPU memory usage. Our results demonstrate that: (1) HiFT
achieves comparable performance to parameter-efficient fine-tuning and standard
full parameter fine-tuning. (2) HiFT supports various optimizers including
AdamW, AdaGrad, SGD, etc. (3) HiFT can save more than 60\% GPU memory compared
with standard full-parameter fine-tuning for 7B model. (4) HiFT enables
full-parameter fine-tuning of a 7B model on single 48G A6000 with a precision
of 32 using the AdamW optimizer, without using any memory saving techniques. | HiFT: A Hierarchical Full Parameter Fine-Tuning Strategy: Full-parameter fine-tuning has become the go-to choice for adapting language
models (LMs) to downstream tasks due to its excellent performance. As LMs grow
in size, fine-tuning the full parameters of LMs requires a prohibitively large
amount of GPU memory. Existing approaches utilize zeroth-order optimizer to
conserve GPU memory, which can potentially compromise the performance of LMs as
non-zero order optimizers tend to converge more readily on most downstream
tasks. In this paper, we propose a novel optimizer-independent end-to-end
hierarchical fine-tuning strategy, HiFT, which only updates a subset of
parameters at each training step. HiFT can significantly reduce the amount of
gradients and optimizer state parameters residing in GPU memory at the same
time, thereby reducing GPU memory usage. Our results demonstrate that: (1) HiFT
achieves comparable performance to parameter-efficient fine-tuning and standard
full parameter fine-tuning. (2) HiFT supports various optimizers including
AdamW, AdaGrad, SGD, etc. (3) HiFT can save more than 60\% GPU memory compared
with standard full-parameter fine-tuning for 7B model. (4) HiFT enables
full-parameter fine-tuning of a 7B model on single 48G A6000 with a precision
of 32 using the AdamW optimizer, without using any memory saving techniques. | [
0.046456445,
0.0718050897,
0.0076305927,
0.0159548819,
-0.0001500534,
0.0221924242,
0.0121416198,
0.005610466,
-0.0349520259,
-0.071845755,
-0.0355234258,
-0.0359374024,
0.0393544249,
-0.0044011069,
0.0320939198,
-0.06855838,
0.0511197038,
-0.0616842136,
0.0171458144,
-0.0178381596,
-0.0172779392,
0.0091743991,
0.0028828548,
0.0202145614,
0.0286399294,
-0.0098693706,
-0.0201761369,
0.0284546688,
0.0290072858,
0.042516131,
0.0215236526,
0.0487755723,
0.0277806111,
0.0311093032,
0.0000023308,
-0.0158239137,
-0.0071526757,
-0.0192979556,
0.0161165334,
0.0379200988,
0.0552946255,
-0.0124859093,
-0.0481309816,
0.0262025259,
-0.0230190419,
-0.0124422684,
0.0307313893,
-0.0259646028,
0.0562342964,
0.0701774955,
-0.0023168521,
-0.0615361556,
0.0320317224,
-0.0496026278,
-0.0501861125,
-0.0469573811,
0.0076263938,
-0.0392331704,
-0.0291598048,
-0.0095523922,
-0.0435737483,
0.04813198,
-0.0059297802,
-0.0046975096,
-0.0274922326,
-0.0147243068,
-0.0091641732,
-0.0195585359,
0.015820045,
0.0343589447,
-0.100402616,
-0.0296823923,
0.0134092458,
-0.0267989244,
-0.0311115347,
-0.0610154234,
-0.0073749535,
0.0449067429,
0.0060108537,
0.0047675553,
0.0237714294,
-0.0564106852,
0.0325179137,
0.0162648596,
-0.0127244228,
0.0418806188,
0.0205469858,
0.0167355351,
-0.026859263,
-0.021433834,
-0.024061963,
-0.038265001,
0.0026678771,
0.0276075061,
-0.0254376177,
0.0220287703,
-0.0827833936,
-0.0538339131,
0.0017753891,
-0.0278713889,
0.0815872252,
0.0743121505,
0.016621355,
0.0169332363,
-0.0241545998,
0.041876208,
-0.0012499153,
-0.0204182081,
-0.0471748412,
0.0147379721,
0.0071223783,
0.0191354137,
-0.0573213212,
-0.0084385369,
-0.0788109303,
-0.0327227935,
-0.0174831599,
-0.0052358382,
0.0031661426,
-0.0134856375,
0.0179591179,
0.0238799863,
-0.0402383991,
0.002204506,
0.0393030494,
0.0484087095,
-0.0494785458,
-0.003684513,
-0.0032397956,
-0.0388731621,
0.0186502989,
0.0233565383,
0.0280581266,
0.0066713127,
0.0179268941,
0.0468699187,
-0.0222972464,
-0.024192892,
-0.0688902065,
0.0082277916,
0.0191417392,
-0.0738270655,
0.008496426,
-0.0279676821,
0.0284348838,
-0.0050240331,
0.014148159,
-0.0175939482,
-0.0010977685,
0.0221701395,
-0.0293625519,
0.0278493576,
-0.0377446637,
-0.0110372081,
0.0339883976,
0.0044173873,
0.0378074646,
0.0949953571,
0.0593533441,
0.0562303513,
0.0164155122,
-0.0173389465,
-0.0489556976,
0.0185099766,
0.0219536256,
0.0424138494,
-0.0342833437,
-0.0109792054,
-0.0267782621,
0.0549829602,
0.0781895071,
0.0399255194,
-0.0282989144,
-0.0465602875,
0.0625788197,
0.024012154,
0.0234413072,
0.0520968325,
0.0144679286,
-0.0006654472,
0.0099219037,
-0.028786175,
-0.0362355635,
0.0150862709,
-0.0183050707,
0.026842637,
0.0470646173,
0.0511093177,
0.0018551134,
-0.0053290906,
-0.0138927028,
-0.009739154,
-0.0184743237,
0.0219438113,
0.0188586228,
-0.0072793975,
-0.0524551906,
0.019978663,
0.041308172,
-0.0438074619,
-0.0288500898,
-0.0556529947,
-0.0005842697,
0.0731301159,
0.0101289283,
0.0114559159,
-0.0496411957,
-0.0089441156,
0.0341109037,
0.0213337652,
0.0021240781,
0.0090747178,
0.046111688,
-0.0703307912,
0.0150473081,
-0.0165289193,
-0.0026096758,
0.0253829788,
0.0214428548,
0.0075889528,
0.0131644187,
0.0377054773,
-0.0432441793,
0.0047379206,
0.003653426,
-0.0055204881,
0.019514557,
0.0080218371,
0.0033031001,
-0.0124379201,
0.0012344682,
-0.0211473126,
0.0136923986,
-0.0291706882,
0.0323891714,
-0.0617649145,
-0.0569229163,
0.0094616339,
0.0317520648,
0.0468955934,
0.070805721,
0.0259466749,
-0.044347439,
-0.056170363,
0.0379089601,
-0.0138580361,
0.0641536936,
-0.0273310114,
0.0269357171,
-0.0292146616,
0.035857752,
-0.0400900804,
0.0288366117,
-0.0112289544,
-0.0224829558,
-0.0123270797,
0.0335069522,
-0.0061325626,
-0.0259758085,
-0.0100023141,
0.0273465272,
0.0123604974,
0.033529751,
0.0087078223,
0.0086162994,
-0.014381161,
-0.0059273383,
-0.1005708128,
0.0215709563,
-0.0425053984,
0.014679498,
0.0350051224,
0.0245627332,
-0.0053789727,
-0.0125942761,
0.0224021878,
0.0611774884,
-0.0311644021,
0.016895989,
-0.0181084666,
0.0196526181,
-0.0001712877,
-0.0276563168,
-0.0334236547,
0.008640687,
0.0523102432,
-0.0372103341,
-0.0252085309,
-0.0521898381,
0.0075155604,
-0.0178968217,
0.0647493824,
-0.0003329473,
-0.0507499501,
-0.0218511093,
0.0050699166,
0.0060962155,
0.1086415127,
0.0178788379,
0.0016328325,
0.0308871958,
-0.0051672077,
-0.0242516827,
0.0219058394,
0.009283592,
-0.0385671705,
-0.025678752,
0.023501493,
-0.0441984236,
0.106470935,
-0.0293553341,
0.0004092618,
-0.0495585576,
0.0174304489,
-0.0212415494,
-0.0043906062,
-0.001795525,
0.0066629858,
-0.0331219621,
0.0068798745,
-0.0155616431,
-0.0726058036,
0.0190237463,
-0.0656660274,
-0.0087003261,
-0.0325044654,
-0.0033440909,
-0.0249787197,
0.0171165466,
0.0321947522,
0.0419888012,
-0.0255544595,
-0.0040489691,
0.0052642608,
0.0064463909,
0.0061866306,
-0.0103284651,
0.0585234687,
-0.0334123187,
0.0263114255,
-0.0549577884,
-0.0427247845,
0.0009549683,
0.0295334328,
-0.0242088325,
0.025390666,
0.0557583682,
-0.0758725628,
-0.0381565765,
0.0262886584,
0.0006036023,
0.0949014127,
-0.0412128903,
-0.0111573217,
-0.0284517948,
0.0040830593,
0.0832372755,
0.0114575112,
0.0074270177,
-0.012339443,
-0.0228692386,
0.0162965283,
0.0369151346,
-0.0184786338,
-0.0726844743,
-0.0031096367,
0.040759556,
-0.0121661769,
-0.0274817683,
0.0178591702,
0.0070503508,
-0.0757404715,
0.0139089348,
0.0739511028,
0.0126991551,
-0.0589846298,
0.0122173727,
0.0151222227,
0.0712021813,
0.0036569089,
0.0495485105,
-0.0644287318,
-0.0386309996,
0.0215324871,
0.0457144231,
0.0151043776,
-0.0053988062,
-0.0159441829,
-0.0694502071,
0.0098026861,
0.0280207712,
-0.0417890958,
-0.0038990588,
-0.0130950706,
0.0044034394,
-0.0054354426,
0.0165849961,
0.0333105214,
-0.0331048854,
-0.0170393754,
0.0108144665,
-0.0089815315,
-0.0414974988,
0.0034076818,
-0.0157116242,
0.0265265163,
0.031420406,
-0.0012819796,
0.0455701761,
0.020268904,
0.0212963969,
-0.0091775656,
0.0193008855,
-0.0661384538,
0.0110907732,
-0.0427019633,
-0.000616772,
0.0287529733,
0.0168869998,
-0.0470391177,
0.0000319824,
-0.0235615689,
0.0371016301,
-0.0125330295,
-0.0204993524,
0.0195592735,
-0.028233001,
-0.0066295369,
0.0067268689,
0.0112123927,
-0.0471590012,
-0.0154586425,
-0.0320844799,
-0.0345754325,
-0.0487136841,
0.018826507,
-0.0358433537,
0.0225444436,
-0.0166515764,
-0.0164216589,
-0.0026888219,
-0.0036075634,
0.0537077263,
-0.0120183369,
0.0095068384,
0.0197084844,
0.0181057658,
-0.059009593,
0.0193757229,
-0.0430662297,
0.0088585867,
-0.0835311562,
-0.0407369845,
0.0467704087,
0.0180833023,
-0.0209714957,
0.0441424027,
-0.071767047,
-0.075366728,
0.0249466449,
0.0470990688,
0.0170748141,
0.0482458696,
-0.0002857673,
-0.0976892784,
-0.0031158251,
0.0175534692,
-0.0174409952,
0.0658741966,
0.0205064025,
-0.0438135155,
-0.0207953788,
0.0493045785,
-0.0145425871,
-0.0089864293,
0.0291122291,
0.0452773795,
0.0296607707,
-0.0638211742,
-0.0349142253,
-0.0110348044,
0.0116127916,
-0.0318020098,
-0.0213118605,
-0.0218969844,
0.0567027666,
0.0210522581,
0.0305208787,
-0.0212344266,
0.0030119545,
0.0029180439,
0.0045610885,
-0.023290243,
-0.0051338174,
0.0608715974,
-0.0316468477,
-0.0268663149,
0.0458232649,
-0.0345310941,
0.0055809063,
0.0398718454,
0.0106651606,
-0.0026485855,
0.0478397235,
0.0238196235,
0.0307836048,
-0.062959373,
0.0402510986,
0.0211714637,
-0.0050147534,
-0.0849815458,
0.056334883,
0.0013237786,
-0.0773728117,
-0.0868872628,
-0.002714549,
0.0729354694,
0.0079412255,
0.0602531396,
0.0176650006,
-0.0127033461,
-0.0124968169,
-0.0079607712,
-0.0103000319,
-0.009631237,
0.0095865848,
0.0135199828,
-0.0299274214,
0.0514290556,
0.0268726218,
0.0326899178,
0.0318384655,
0.0183717087,
0.0223077592,
-0.0195001196,
-0.0363114141,
-0.0145897884,
0.0194707904,
-0.0258193687,
-0.0266429223,
0.0148550039,
-0.0445360281,
-0.0204455219,
-0.0131434929,
-0.0232997555,
-0.0064322981,
0.0400014296,
0.0221359562,
-0.0195437539,
-0.0439054519,
-0.0477655418,
-0.0052611181,
-0.0031524955,
0.0005024482,
0.0117108412,
-6.912302777e-33,
-0.0357833058,
0.0399067886,
0.0226545967,
-0.0094567649,
-0.0089587634,
-0.016303869,
-0.0237843879,
0.0041217529,
-0.0211928133,
-0.012055032,
-0.0311150011,
-0.0102758631,
0.0238560829,
0.0176912863,
-0.0165098384,
0.0253366195,
0.0557162277,
-0.0225376207,
-0.0296796169,
0.0518850572,
0.027435245,
0.0448774099,
0.0595312417,
-0.0024669205,
-0.0799228176,
0.0080026872,
-0.0047245594,
0.0184014942,
-0.0084481817,
0.0248210356,
-0.0161699671,
0.0415595248,
0.0022575716,
-0.0065932013,
-0.0241243746,
-0.0168623738,
-0.090776667,
-0.0100888554,
0.0454723537,
0.0323867947,
-0.0458073243,
-0.0497743487,
0.0222273096,
0.0416930132,
-0.0191436466,
-0.0196532179,
-0.0268931035,
-0.0086429985,
-0.0481548645,
-0.053117957,
-0.0794279054,
0.0234724097,
0.0359867811,
0.0235414561,
0.0541781113,
-0.024159234,
-0.0113211302,
-0.0028277903,
-0.0542350151,
-0.0067862174,
0.0593185462,
0.0903410017,
0.0816519409,
0.0468630008,
0.0130733317,
0.0234279484,
0.0446090102,
0.0009974593,
0.0196890701,
0.0328912251,
0.086422123,
0.0006245344,
-0.0077064801,
-0.0333578773,
0.0376944505,
-0.0577406287,
-0.0446962677,
-0.0573803708,
0.023785973,
0.0512254797,
0.01905757,
0.0593447387,
0.0014924344,
-0.0268071536,
-0.0425109603,
-0.0218693689,
-0.0303721186,
-0.0298356302,
0.0422528014,
0.0420629121,
-0.0452403575,
0.0017390266,
0.0420204662,
-0.0191064831,
0.0460339971,
0.0071471622,
-0.0027284997,
0.0224707406,
0.008501783,
0.0436183847,
-0.0451589897,
-0.0604532473,
0.0150082177,
-0.0374995396,
0.023311073,
-0.0408108421,
0.0078410376,
0.0563921519,
-0.0797669962,
0.0105569148,
-0.0194256045,
0.034828525,
0.0207488071,
0.0780658573,
0.0022673139,
-0.0085906619,
0.0180118792,
0.087247923,
-0.0214777831,
0.0019566284,
-0.0215588715,
0.026318308,
0.0145456865,
0.0148868896,
-0.0190633666,
0.0152640082,
-0.0180133823,
0.0112386951,
0.0209403057,
-0.0760448128,
-0.0318504944,
0.0575003214,
2.948e-7,
0.0062028053,
0.0730473176,
-0.0252183508,
-0.039854642,
0.0248208065,
-0.002271519,
0.0323254466,
-0.0266481601,
0.0243646521,
0.0345850959,
-0.0101445317,
0.0051772692,
0.0327386633,
0.0021960617,
-0.0616629384,
0.0188171268,
-0.0433606878,
0.0361667536,
-0.0218518581,
-0.012727445,
-0.055255983,
0.1333459467,
-0.0016584989,
-0.0198141113,
0.049719654,
-0.0246561207,
0.0186015647,
0.0036231424,
0.0045404476,
0.0476849303,
-0.0062307799,
0.0237651579,
-0.0483812764,
-0.0026670999,
-0.0151190776,
0.021921996,
-0.0513657518,
0.0956290364,
-0.0535773709,
-0.0005695245,
0.0449021123,
0.0082190055,
-0.0123074828,
-0.0792858452,
0.0425643176,
-0.022800507,
0.0018651842,
-0.0612864159,
0.0221582744,
-0.0121845566,
0.0030421377,
-0.0443857498,
0.0171526279,
0.0273311734,
0.0419684462,
-0.022909591,
-0.037281841,
-0.0812376887,
-0.0168664511,
0.0241367668,
-0.0197872072,
-0.0408994742,
-0.0836935639,
0.0052351225,
0.0222310852,
-0.0492474101,
0.0018292367,
2.966655016e-34,
0.0586096309,
-0.0391998142,
0.0113499817,
-0.0061904876,
0.0272578374,
0.0036356559,
-0.0166230202,
0.0271344427,
-0.0103336889,
-0.09610302,
-0.0520836934
] |
http://arxiv.org/abs/2401.15170v2 | http://arxiv.org/pdf/2401.15170v2 | Scalable Qualitative Coding with LLMs: Chain-of-Thought Reasoning Matches Human Performance in Some Hermeneutic Tasks | 2024-01-26T19:25:43 | 2024-02-12T23:04:10 | Zackary Okun Dunivin | cs.CL | cs.CL, cs.AI | Qualitative coding, or content analysis, extracts meaning from text to
discern quantitative patterns across a corpus of texts. Recently, advances in
the interpretive abilities of large language models (LLMs) offer potential for
automating the coding process (applying category labels to texts), thereby
enabling human researchers to concentrate on more creative research aspects,
while delegating these interpretive tasks to AI. Our case study comprises a set
of socio-historical codes on dense, paragraph-long passages representative of a
humanistic study. We show that GPT-4 is capable of human-equivalent
interpretations, whereas GPT-3.5 is not. Compared to our human-derived gold
standard, GPT-4 delivers excellent intercoder reliability (Cohen's $\kappa \geq
0.79$) for 3 of 9 codes, and substantial reliability ($\kappa \geq 0.6$) for 8
of 9 codes. In contrast, GPT-3.5 greatly underperforms for all codes
($mean(\kappa) = 0.34$; $max(\kappa) = 0.55$). Importantly, we find that coding
fidelity improves considerably when the LLM is prompted to give rationale
justifying its coding decisions (chain-of-thought reasoning). We present these
and other findings along with a set of best practices for adapting traditional
codebooks for LLMs. Our results indicate that for certain codebooks,
state-of-the-art LLMs are already adept at large-scale content analysis.
Furthermore, they suggest the next generation of models will likely render AI
coding a viable option for a majority of codebooks. | Scalable Qualitative Coding with LLMs: Chain-of-Thought Reasoning Matches Human Performance in Some Hermeneutic Tasks: Qualitative coding, or content analysis, extracts meaning from text to
discern quantitative patterns across a corpus of texts. Recently, advances in
the interpretive abilities of large language models (LLMs) offer potential for
automating the coding process (applying category labels to texts), thereby
enabling human researchers to concentrate on more creative research aspects,
while delegating these interpretive tasks to AI. Our case study comprises a set
of socio-historical codes on dense, paragraph-long passages representative of a
humanistic study. We show that GPT-4 is capable of human-equivalent
interpretations, whereas GPT-3.5 is not. Compared to our human-derived gold
standard, GPT-4 delivers excellent intercoder reliability (Cohen's $\kappa \geq
0.79$) for 3 of 9 codes, and substantial reliability ($\kappa \geq 0.6$) for 8
of 9 codes. In contrast, GPT-3.5 greatly underperforms for all codes
($mean(\kappa) = 0.34$; $max(\kappa) = 0.55$). Importantly, we find that coding
fidelity improves considerably when the LLM is prompted to give rationale
justifying its coding decisions (chain-of-thought reasoning). We present these
and other findings along with a set of best practices for adapting traditional
codebooks for LLMs. Our results indicate that for certain codebooks,
state-of-the-art LLMs are already adept at large-scale content analysis.
Furthermore, they suggest the next generation of models will likely render AI
coding a viable option for a majority of codebooks. | [
0.0377661586,
0.0779989958,
-0.0200954471,
0.0130773503,
-0.0257338993,
0.0586739257,
-0.0029722045,
0.0136470255,
0.0079219434,
0.0081264554,
0.0307844467,
-0.0099597415,
0.0287460778,
0.0315626115,
-0.0179554112,
0.0233348664,
-0.0149284862,
0.0036831053,
0.0500633866,
0.0069873449,
-0.0350857042,
0.0521148778,
0.01852444,
0.0270800255,
0.0138229262,
-0.0303792749,
-0.0284849629,
-0.0119383959,
0.027097499,
-0.0423639789,
-0.0225063022,
0.0318866335,
0.0266146511,
0.0336161107,
0.0000024591,
-0.0451123454,
-0.0207990222,
-0.007502635,
-0.0281016398,
0.0162684824,
0.0026592156,
-0.0412144884,
-0.0074301558,
-0.0200721063,
-0.053469032,
0.0272119753,
0.0161300004,
0.0180625618,
-0.0002742425,
0.0434145071,
-0.0062202211,
-0.0554274991,
0.0409438312,
0.0097380476,
0.00409202,
-0.0355497897,
0.0250158459,
0.0550355315,
0.014672162,
0.0687048584,
-0.0437374525,
0.0679583102,
0.0768075883,
-0.0083904527,
-0.0101149622,
0.0218520127,
-0.0260625891,
-0.0100428313,
0.0270490125,
0.0123793138,
0.0358547345,
-0.0215072054,
0.0255943406,
0.0046824757,
-0.0194914974,
-0.0082959542,
-0.0287644546,
0.0444461703,
-0.0322733112,
0.0212090984,
0.0463077873,
0.0216359962,
0.0071320757,
0.0194879081,
-0.0412374176,
0.0528691113,
-0.0235074032,
-0.0215261504,
0.0352931432,
0.0197232123,
-0.008524918,
-0.0382702574,
-0.0124517102,
0.0168026239,
0.0319880359,
0.014982136,
0.0147710759,
0.055300083,
-0.0079938127,
-0.0413861685,
-0.0003930078,
0.0572450981,
-0.0000710029,
0.0131404158,
0.0112384781,
0.0688793436,
-0.005126846,
0.0084531698,
-0.0197515301,
0.0061130077,
-0.0900207311,
-0.0150870234,
-0.0125434566,
0.025498325,
-0.0158015434,
0.0232084114,
-0.027958829,
0.0146387639,
0.0314519331,
0.0100626741,
-0.0316111706,
-0.0075666835,
-0.00041457,
-0.0030685195,
-0.0147813829,
0.0362231247,
-0.0415494032,
0.0105249127,
-0.0297651105,
-0.0450019948,
0.0155566623,
0.0084283715,
-0.0076203551,
-0.0488344692,
0.0130592128,
0.0007536822,
0.0032614765,
-0.0032324183,
-0.0549422763,
-0.0309313834,
-0.0437217243,
-0.0335413739,
0.042539537,
-0.0401789062,
0.0307403412,
0.0360301957,
0.0236843005,
-0.0796504617,
-0.0118383346,
0.0035328821,
-0.0281994529,
0.0440687612,
-0.0134386402,
0.0189058464,
0.0426055193,
0.0320794545,
0.0000763035,
0.0511380248,
0.0713718235,
0.0333259031,
0.0431095921,
-0.0197644308,
0.0063629281,
-0.0144185089,
-0.0263573993,
0.0192210581,
-0.0393217392,
0.0117191738,
-0.0454653278,
0.0247549154,
0.0189122353,
0.0497688688,
-0.0239439383,
-0.0139129264,
0.0637137294,
0.1059472039,
0.0386863053,
0.0576448552,
0.0509623103,
0.01796332,
0.0493603125,
-0.0009771065,
-0.0247480813,
0.0275733825,
0.0023255895,
-0.0395181179,
0.0087852338,
-0.0091940649,
-0.0118013453,
-0.0604249723,
-0.0110928575,
0.0177790783,
0.0126109533,
-0.0113551775,
0.0248170551,
-0.0463056117,
0.0024121108,
-0.044114694,
-0.0114776604,
-0.0735403374,
0.004230239,
-0.0530851111,
0.0584352091,
0.0767736733,
0.0600177534,
0.0092892442,
0.0520715788,
-0.0231377706,
0.0114289252,
0.0256774388,
0.0056029879,
-0.0113606863,
-0.0173282698,
-0.0625847802,
-0.0247610305,
-0.027819464,
-0.0154973641,
0.007001251,
-0.0146762477,
0.029510716,
0.0033063598,
-0.0011191721,
-0.027172029,
0.020411497,
-0.036511384,
-0.0444580726,
-0.0133371856,
-0.0093350708,
0.0309995506,
0.0421482734,
0.0005038111,
0.0042778593,
-0.006643808,
-0.0474439189,
-0.0448224582,
0.0034930902,
-0.027113732,
0.0109861167,
-0.0067272671,
-0.0608383082,
-0.0413857959,
0.0267944541,
-0.0022358224,
-0.0120163905,
0.0136397332,
-0.0452790968,
0.0127398185,
-0.0229538791,
-0.0083413161,
-0.0236852281,
-0.0078143543,
-0.0167600717,
0.0922307298,
-0.0113655739,
0.0028208701,
-0.009843397,
-0.0564552173,
0.0455490872,
0.0135228056,
-0.0074821017,
0.0313495249,
0.0203185119,
0.0053022415,
0.0294314995,
0.0113507649,
-0.0199842807,
-0.0352022089,
-0.0900764689,
0.0066305762,
-0.0136537161,
-0.008436243,
0.016738141,
-0.0241371319,
-0.0141352294,
0.0365576446,
-0.0195318125,
-0.0391132832,
-0.0383815393,
0.0188488979,
0.0131942583,
-0.0101231011,
0.0120977173,
0.0142784622,
0.0068753036,
-0.0420994982,
0.0648272634,
0.0561778061,
-0.0031406761,
-0.039309673,
0.0280015469,
-0.0190693233,
-0.0458885171,
-0.0248809494,
-0.0610321127,
0.0281719193,
-0.0010269927,
0.0167251732,
0.0572092161,
0.0095919902,
0.0445922837,
0.0323924869,
-0.0006010616,
-0.0263558608,
-0.0048583155,
-0.0081979437,
-0.0322612971,
-0.0445334725,
-0.0983902141,
0.0061499537,
0.0520435013,
0.0391751267,
0.0182473715,
-0.0998594016,
0.0362124629,
-0.0440718979,
-0.025801193,
-0.0415265337,
-0.0600555353,
0.0066287201,
0.1322900504,
0.0092685474,
-0.0105382996,
0.04200368,
0.0108908238,
0.0179092083,
-0.0359198786,
-0.0225317292,
-0.0629576221,
0.106592074,
-0.0362370424,
-0.0020267258,
-0.0144069036,
-0.0079022795,
-0.0253941342,
0.0048777107,
-0.0255314559,
-0.0438474156,
0.0513982885,
-0.0743403956,
0.0029355625,
-0.0164500736,
-0.0318205319,
-0.0143553615,
-0.011098288,
-0.0261910334,
-0.0233208425,
0.0169577729,
0.0021154729,
0.0359401442,
0.021848496,
0.0333784856,
0.0305169392,
-0.0144652948,
-0.0175756942,
-0.0122430054,
-0.0031806349,
-0.0121475821,
-0.0229967926,
0.0408214256,
0.0056174365,
0.0484105535,
0.0524695106,
-0.0328921638,
0.0307093803,
-0.033545766,
0.0366061218,
0.0028309962,
-0.0126017854,
-0.0476724654,
0.0074427794,
0.0642054901,
-0.0673072264,
-0.0044225524,
0.1358189285,
0.0387782156,
-0.1020309106,
0.0052995011,
-0.0698338076,
-0.0461010039,
-0.0013468129,
0.0105978753,
-0.0858131871,
0.0001265875,
-0.0007982294,
0.017192876,
-0.0494214147,
-0.0709707215,
0.0166537706,
-0.0841025487,
0.0276078656,
0.0194076523,
0.00734512,
-0.061087478,
-0.0242088102,
-0.0257075746,
-0.0028266811,
-0.0571063198,
0.0348523743,
0.022003632,
0.0196569208,
-0.0173103176,
-0.0001821829,
0.0039258571,
0.0479728766,
0.0178810228,
0.0386576913,
0.0569896474,
0.077229239,
-0.0233343039,
-0.0398727693,
-0.0082018096,
-0.0516892932,
-0.0025946738,
0.005671083,
0.0261824224,
0.0037761899,
0.0129127484,
0.0024228033,
0.043587666,
-0.0427838527,
-0.0279211272,
0.05185844,
0.0731993169,
-0.0406450033,
-0.0326991081,
-0.0284098107,
0.0085601006,
-0.0310990252,
0.0729563981,
0.0452041663,
-0.0305220503,
-0.0202454664,
-0.0607584566,
0.0196757913,
-0.0213910099,
-0.0722575113,
0.0229380578,
0.0149106998,
-0.0292862132,
-0.0498861559,
-0.0486166067,
0.0045335288,
-0.0068479944,
0.0243619513,
-0.039675191,
0.034543477,
-0.0141720818,
-0.0765294135,
0.0160168838,
-0.0365078971,
-0.0131776612,
0.0048806607,
0.0015248542,
0.0199606922,
-0.0481447019,
0.0037732583,
0.0136848381,
0.0121509759,
-0.0205547344,
0.0014493568,
-0.0225640982,
0.0265736654,
0.0683571249,
-0.0050398796,
-0.0344695933,
-0.0080147823,
0.0153770763,
-0.0544365197,
0.0613418333,
-0.0068806363,
0.0772045478,
0.0004528592,
0.0174725745,
-0.0257813279,
0.0390140638,
-0.043855764,
0.0174629409,
-0.0557376519,
-0.0586107746,
0.0253355131,
-0.026353335,
-0.0214313585,
-0.0211400222,
-0.0327298045,
0.0128940847,
-0.0099268965,
-0.0489814803,
-0.0200216714,
-0.0352407917,
-0.007330697,
0.0578844622,
-0.0275052339,
-0.0316605419,
-0.0107698785,
-0.0185730215,
0.0084796445,
-0.0363040939,
0.0306211896,
-0.0261892565,
0.0771165192,
-0.0466555171,
-0.0072311186,
-0.0367183015,
0.0136506883,
-0.0121652083,
0.0602221787,
-0.005997723,
0.0713407472,
0.0053118747,
-0.0051825806,
0.0357006267,
0.0219239872,
-0.1055802256,
-0.0550329946,
0.0257097445,
-0.0631793141,
-0.00085424,
0.0317956693,
0.0316622816,
-0.028223021,
-0.0042012618,
0.0148657905,
0.0161637105,
0.0233381111,
0.0027089058,
-0.0256935526,
0.0414775871,
-0.0247021075,
-0.0312275719,
0.0312698632,
-0.01182763,
0.0160137061,
0.0157592148,
-0.0599174313,
-0.0207001194,
-0.0394966528,
0.0072879689,
0.0100457063,
0.0245583039,
-0.0089867478,
-0.0844361633,
-0.0376747623,
0.0451127142,
0.005271181,
0.0219970439,
-0.0266947951,
0.0285386778,
0.0500855744,
-0.0266600884,
-0.0192751642,
-0.0504180565,
0.0176490285,
-0.0241463277,
-0.0191860609,
-0.022947615,
-6.989985322e-33,
-0.0324881487,
-0.0484428406,
0.0370808728,
-0.0121204229,
-0.0334221348,
0.0230884627,
-0.0420000963,
-0.0364769585,
-0.0076652318,
0.0457312241,
-0.0090066968,
0.0150984386,
0.000967384,
-0.0105046537,
0.052215755,
0.0110347336,
0.0278225094,
0.0075926827,
-0.0225062072,
0.0207059644,
0.026742626,
-0.0280539393,
0.0339273401,
-0.0659465343,
0.0441383906,
0.0267317668,
0.0124123273,
-0.0181209873,
0.0235524364,
0.028089745,
-0.0217113569,
0.0124805961,
0.0480589122,
-0.0105201062,
0.0312277153,
0.0286439136,
-0.0557394475,
-0.0617908388,
0.0357546285,
-0.0136250434,
-0.0137795368,
-0.0849941969,
0.0343852006,
-0.0083123557,
-0.0478963703,
-0.0036214248,
0.0453118198,
-0.0213812757,
0.0160247721,
-0.0737188458,
-0.0333072767,
0.037733227,
-0.0470057465,
0.0082963305,
-0.0101928171,
0.0469085276,
-0.0265249889,
0.0850189999,
-0.0420218632,
0.0098241223,
-0.0158479642,
-0.0088949995,
0.0014633014,
-0.0624560118,
0.0114242239,
0.0458731428,
0.0035084556,
-0.0194454659,
-0.0792838112,
0.0204292461,
0.0066879052,
-0.0211615581,
0.0567069091,
0.0626743659,
0.0117126219,
0.003246051,
-0.0664578751,
-0.0000064825,
-0.0259475708,
0.0182395056,
0.042333439,
0.0019727794,
-0.0776029229,
-0.0151059357,
-0.0284457412,
-0.033839833,
-0.001364975,
-0.0206677634,
0.0152767226,
-0.0467460006,
-0.0376849435,
0.0186723098,
0.0438384786,
-0.060656894,
0.0520446524,
0.0190341119,
-0.0326784961,
0.0642873198,
0.0322267152,
0.0318846665,
-0.0366405211,
0.0299559161,
0.024372397,
0.0087999292,
0.0745413676,
0.0070206849,
0.0086192386,
0.0129788229,
-0.01924471,
-0.0460245982,
-0.0369469151,
-0.0136345131,
-0.0004711547,
-0.0202629399,
-0.0138146207,
0.0548261255,
0.0186845344,
0.0106856385,
0.0280364696,
0.001780382,
-0.0088947741,
-0.0205725152,
0.0112810573,
-0.0087666623,
-0.0320164636,
-0.0165960081,
-0.0091618877,
0.056838166,
0.0457107686,
-0.068683885,
-0.0219182894,
0.0416985974,
3.333e-7,
0.0539169163,
0.0434231758,
0.0329801403,
-0.0074109547,
0.0264714863,
0.0110411085,
0.0276458208,
-0.0073073474,
-0.0153027009,
-0.0387113839,
0.037411198,
0.063541986,
0.0501223989,
0.0561593845,
-0.0434856303,
-0.0177407246,
-0.0151749281,
-0.0379485078,
-0.079400152,
0.0711937398,
0.0964477211,
0.0421027169,
0.0359727554,
-0.007674566,
0.0134457722,
0.0146591151,
0.0398976766,
-0.0117943501,
0.0013652628,
0.0238198396,
-0.0550054982,
0.0143038947,
0.0096863788,
0.0221114513,
0.0151027562,
0.0087079369,
-0.0038191774,
0.0840593502,
-0.0003694254,
-0.0246626567,
0.0235053822,
-0.0125633068,
0.0140562709,
-0.0638096407,
0.0986423269,
-0.0014430274,
-0.0599482358,
0.0120273242,
-0.0100785242,
-0.0241042152,
0.01811908,
-0.0008401151,
0.0088289017,
0.0137469703,
0.0034127431,
-0.0042403,
-0.0013191153,
-0.073620148,
0.0219933596,
-0.0186053962,
-0.035297893,
-0.0356119461,
-0.0176371746,
0.020449359,
0.0320837013,
-0.0890510529,
0.0132369855,
3.126558063e-34,
0.0553264581,
-0.0348006487,
0.0094365906,
0.0086757615,
0.0448886044,
0.0387327038,
-0.0437549539,
0.0121924477,
0.0336149372,
-0.0497252196,
-0.0182885155
] |
http://arxiv.org/abs/2401.15077v2 | http://arxiv.org/pdf/2401.15077v2 | EAGLE: Speculative Sampling Requires Rethinking Feature Uncertainty | 2024-01-26T18:59:01 | 2024-02-04T17:18:34 | Yuhui Li, Fangyun Wei, Chao Zhang, Hongyang Zhang | cs.LG | cs.LG, cs.CL | Autoregressive decoding makes the inference of Large Language Models (LLMs)
time-consuming. In this paper, we reconsider speculative sampling and derive
two key observations. Firstly, autoregression at the feature
(second-to-top-layer) level is more straightforward than at the token level.
Secondly, the inherent uncertainty in feature (second-to-top-layer) level
autoregression constrains its performance. Based on these insights, we
introduce EAGLE (Extrapolation Algorithm for Greater Language-model
Efficiency), a simple yet highly efficient speculative sampling framework. By
incorporating a token sequence advanced by one time step, EAGLE effectively
resolves the uncertainty, enabling precise second-to-top-layer feature
prediction with minimal overhead. We conducted comprehensive evaluations of
EAGLE, including all models from the Vicuna and LLaMA2-Chat series, the MoE
model Mixtral 8x7B Instruct, and tasks in dialogue, code generation,
mathematical reasoning, and instruction following. For LLaMA2-Chat 70B, EAGLE
achieved a latency speedup ratio of 2.7x-3.5x, doubled throughput, while
maintaining the distribution of the generated text. | EAGLE: Speculative Sampling Requires Rethinking Feature Uncertainty: Autoregressive decoding makes the inference of Large Language Models (LLMs)
time-consuming. In this paper, we reconsider speculative sampling and derive
two key observations. Firstly, autoregression at the feature
(second-to-top-layer) level is more straightforward than at the token level.
Secondly, the inherent uncertainty in feature (second-to-top-layer) level
autoregression constrains its performance. Based on these insights, we
introduce EAGLE (Extrapolation Algorithm for Greater Language-model
Efficiency), a simple yet highly efficient speculative sampling framework. By
incorporating a token sequence advanced by one time step, EAGLE effectively
resolves the uncertainty, enabling precise second-to-top-layer feature
prediction with minimal overhead. We conducted comprehensive evaluations of
EAGLE, including all models from the Vicuna and LLaMA2-Chat series, the MoE
model Mixtral 8x7B Instruct, and tasks in dialogue, code generation,
mathematical reasoning, and instruction following. For LLaMA2-Chat 70B, EAGLE
achieved a latency speedup ratio of 2.7x-3.5x, doubled throughput, while
maintaining the distribution of the generated text. | [
0.0329258852,
0.0594493486,
-0.0229117647,
0.0010341912,
-0.0373568088,
-0.0155812846,
-0.0329165533,
-0.0027608597,
-0.0337266773,
-0.0296400674,
0.00585949,
-0.1038037762,
0.0290223714,
0.0529231839,
0.0015366405,
-0.0286800135,
0.0171930585,
-0.1048298329,
0.0416613482,
-0.0346578434,
-0.0261447076,
0.0338966921,
0.0172067005,
0.0695706159,
0.0032714813,
-0.0090446444,
-0.0457574315,
-0.0121039795,
-0.0461674966,
-0.0031053489,
-0.0100910598,
0.0472431444,
0.0061990567,
0.003918075,
0.0000023435,
-0.0144801158,
-0.0087606199,
-0.0118385311,
-0.0257689003,
0.0489429347,
0.0510585345,
-0.0075706076,
-0.0423117653,
0.0314379409,
-0.010768909,
-0.0100225266,
0.0326571912,
-0.0130712716,
0.0118362894,
0.0154958684,
-0.0021633147,
-0.0100787152,
0.0225822162,
0.0025025054,
-0.0181411486,
-0.0566000715,
0.0044664941,
-0.0117368074,
0.0181410108,
0.0177676547,
-0.008196529,
0.0443244278,
0.0335207805,
-0.0107721146,
-0.0241575278,
0.0429112203,
0.0035750556,
-0.0222831052,
-0.0350415297,
0.0507464744,
-0.0189405065,
-0.0168097857,
-0.0158780366,
-0.0044308198,
-0.042803809,
-0.0239965953,
-0.0360416919,
0.0136106322,
-0.0241821986,
0.0140234213,
0.0715523809,
0.0073148687,
0.0128901759,
-0.0178283937,
-0.0370649919,
0.0516445562,
0.0098043615,
0.0295380931,
-0.0394591987,
0.0085312594,
-0.0017528209,
-0.0550808832,
0.0350159928,
0.0312896781,
-0.0377597436,
0.0147434166,
-0.0639095157,
-0.0324630104,
-0.0140025914,
-0.0491138846,
0.0606322549,
0.0697654113,
-0.0338714905,
0.0020797711,
-0.0002781931,
0.0184233915,
-0.0168678463,
-0.0243081842,
-0.0292136874,
0.0140036894,
0.0209658667,
-0.0245780237,
-0.0772777647,
0.0468843915,
-0.037554767,
-0.0390165448,
-0.0283410195,
-0.0056046774,
0.0133123705,
-0.0284805018,
0.0310080796,
0.0065962025,
-0.0476716422,
-0.0040263841,
0.0335880034,
-0.0049728407,
-0.0018671065,
0.012859419,
-0.0052019069,
-0.106164813,
-0.0111263376,
-0.0079010753,
-0.0146627724,
-0.0064612529,
-0.0223913342,
0.1152866632,
-0.0312535092,
-0.0295925345,
-0.0271314085,
-0.025034165,
0.0017905784,
-0.0435844064,
0.0094849793,
-0.0809498504,
-0.0065924572,
0.0246051047,
0.0057457001,
0.0231789295,
-0.0028430684,
0.0220813025,
-0.0175574776,
0.0199422259,
-0.001870565,
0.0227458496,
0.0143127255,
-0.0018618389,
-0.0018621766,
0.1163309291,
0.0583867356,
0.0450569279,
0.0044535287,
0.0432677306,
-0.0083028143,
0.0866744295,
0.0285806544,
-0.0192704946,
-0.031883914,
-0.0254059341,
-0.0113913929,
-0.0099603236,
0.0457199961,
0.0622992627,
-0.0388738066,
-0.0182135403,
0.0381676257,
0.0410327539,
-0.023804808,
0.0033692163,
0.0542077385,
0.0320059806,
0.0332132876,
0.0036995993,
-0.02346275,
0.0519622415,
-0.0019828202,
0.0285880323,
0.0238071978,
0.0615106411,
-0.0219423026,
-0.0180722009,
-0.017911803,
0.0034640315,
0.0033254544,
0.0132142277,
0.0635255426,
0.0076107159,
-0.0493993536,
0.0746323019,
0.0109481867,
-0.0084533403,
-0.0346593857,
-0.0535830557,
0.0010954813,
0.0772383809,
0.0380517617,
0.0161023214,
-0.0089814095,
0.0003341296,
0.0069630658,
-0.016224362,
0.0230946634,
-0.0102212392,
0.0403228961,
-0.039699275,
-0.030834116,
0.0496078581,
0.0024835898,
0.0632789955,
0.01936833,
0.0164352488,
0.0165434834,
-0.0073604384,
-0.0311343987,
0.0032186869,
-0.0578738824,
0.0178910438,
0.020914929,
-0.0051324032,
0.0305787306,
0.0086226696,
-0.0077778865,
0.0124862203,
0.0053661736,
-0.0495103113,
-0.0042783427,
0.0173369907,
-0.0972701684,
-0.0382990576,
0.0269366074,
-0.00454885,
0.0106037874,
0.0003188973,
-0.0188425519,
-0.0538020581,
0.0370581225,
-0.0496031456,
0.0615865849,
-0.0422669947,
0.0455775,
-0.0033652766,
0.0359747373,
0.0053122742,
0.0170961507,
-0.0571162999,
-0.0310828742,
0.0238001812,
0.0079719387,
0.0068161562,
-0.0249041282,
-0.0161378775,
0.0268720668,
-0.0001644649,
-0.0060188388,
-0.0138475094,
0.0266527552,
-0.0074641407,
-0.0038466579,
-0.093559809,
-0.0090798372,
-0.049931936,
0.0058897207,
-0.0157025028,
0.0386804976,
-0.0130350115,
-0.0120882699,
-0.0282072965,
-0.0153759513,
-0.027385965,
0.0085587911,
-0.0243712496,
0.0033973942,
-0.0357176848,
-0.0159967784,
-0.0291583203,
-0.0033777819,
0.0470929258,
-0.0208082553,
-0.0431261063,
-0.0597357117,
0.0204088408,
0.0094942367,
0.0189943388,
-0.0250446741,
-0.0416854583,
-0.0086739045,
-0.056564752,
0.0155190937,
-0.0162776634,
0.0026258451,
0.0165641885,
0.0057756971,
-0.002336307,
-0.0460213087,
0.0022462958,
-0.0236862805,
-0.003896971,
-0.0215248875,
-0.0363786668,
-0.0393808745,
0.0962601528,
0.0126427179,
-0.0100388788,
-0.0505858995,
-0.0224503484,
-0.005655448,
-0.0331912227,
0.0068925247,
-0.0170245674,
-0.0271186661,
0.0389364325,
-0.03021577,
-0.0608094223,
0.069661811,
-0.0670131743,
-0.0094367452,
-0.0134742362,
0.0139538217,
-0.010033235,
0.0640563071,
-0.0053310231,
0.0012481228,
0.0004105831,
-0.0701574758,
-0.0108089428,
0.0485810079,
0.0153063033,
-0.0622617975,
0.0258387476,
-0.0330098346,
0.0384952091,
0.0052787601,
-0.0241529215,
-0.028620068,
-0.0221654493,
-0.048162654,
0.0152197424,
0.0387927033,
-0.0256558545,
0.0399904288,
0.0274511166,
-0.064340435,
0.0497266948,
-0.0241566449,
0.0052474136,
-0.0107106529,
0.0029393453,
0.0571101829,
-0.014041339,
0.0072103045,
-0.0443128832,
0.0377284661,
0.0238091182,
0.0546165295,
0.0243047699,
-0.033145044,
0.0150750559,
0.060444463,
0.0310013164,
-0.028117666,
0.0577181801,
0.0790845007,
-0.0379916579,
-0.0169465691,
0.1004420891,
0.0190205649,
-0.0397809893,
0.0330512896,
-0.0000210555,
0.0815129131,
0.0071642976,
0.020208789,
-0.0627692565,
-0.0271864273,
0.0519921407,
0.0272561964,
-0.0079071429,
-0.0189097188,
0.0011481514,
-0.1269331425,
0.0520411283,
-0.0020110789,
-0.0472415537,
0.0073254029,
-0.0280652288,
-0.0082906177,
0.0344828926,
0.037151631,
0.0164612588,
-0.0035961361,
0.0061189192,
0.0456428155,
-0.0278234649,
-0.0050587892,
0.0318887196,
0.0048334198,
0.0629537106,
0.0455199368,
0.0063046655,
0.0242331177,
0.0416294709,
0.0093157468,
0.0169322602,
0.0204035994,
-0.0194384959,
0.0330059417,
0.0266233962,
-0.0060905726,
0.0262779165,
0.0352801643,
-0.0006388198,
0.0286662281,
-0.0380934291,
0.060244631,
-0.0159806907,
-0.0311659332,
-0.0303766001,
-0.0006155462,
0.029012911,
0.0361914523,
0.030278476,
-0.0326083861,
-0.0449076891,
-0.0530782379,
-0.0195493605,
-0.0659899339,
-0.0138209909,
0.010053291,
0.0192088429,
-0.0515584983,
-0.0445877761,
-0.010142168,
0.0436282232,
0.059296947,
-0.0066324747,
0.045305226,
0.0162899867,
-0.020946091,
-0.052114021,
-0.0495853052,
-0.0048964806,
0.058182057,
-0.007417751,
0.0059485966,
0.0264888834,
-0.0235918779,
-0.0213602688,
0.0372103341,
-0.1211550757,
-0.0931555703,
0.0348801576,
0.0391858295,
0.0245001223,
0.0543230399,
-0.0303002112,
-0.0520951077,
0.0764665008,
0.0302418079,
-0.0664026961,
0.0303041544,
0.0099447221,
-0.0267548952,
0.0101121003,
0.0433574803,
-0.0352016836,
0.0125774266,
-0.0207521617,
0.0437399782,
0.0068944707,
-0.0534814037,
-0.0104541918,
0.0243868362,
0.0294573642,
-0.022653006,
0.0173742063,
0.0212921742,
0.0549350865,
-0.0200750604,
0.0109534403,
-0.0577754341,
0.0568976812,
-0.0015130488,
-0.0346963033,
-0.0056598517,
-0.0233752988,
0.0373976156,
0.0324482881,
0.0433169603,
-0.0048911585,
-0.0484945811,
0.0335386768,
0.0591982566,
-0.0156580433,
-0.0098111946,
0.012153673,
0.0248215012,
0.0319369212,
-0.0064746705,
0.0165434573,
-0.0020206925,
0.0165812876,
-0.0179235954,
-0.0202258509,
-0.0626482144,
-0.0810583234,
-0.0701958239,
0.0457707159,
0.0234574545,
-0.0092274603,
0.0612124838,
-0.0213874411,
0.0228098966,
-0.0208192598,
0.0446224362,
-0.0264061298,
-0.0462298207,
0.0464401096,
-0.0015080872,
-0.0194859281,
0.0043277577,
0.0628006831,
0.0520578586,
0.0387237929,
-0.0023402229,
-0.024388276,
0.0013909774,
-0.0229568221,
0.0206729192,
0.0229607075,
-0.0234334338,
-0.0098127816,
-0.029734876,
-0.0682412311,
0.021069726,
0.0068191555,
-0.0014108815,
-0.0403450429,
0.0437335297,
0.07806205,
0.0105326306,
0.0061331298,
-0.0498579852,
-0.0095925452,
0.0135499975,
-0.0180752929,
0.0066251387,
-7.174415977e-33,
-0.034902636,
-0.0043681064,
0.0004061548,
-0.008254909,
-0.0230957884,
-0.0129943844,
-0.0272375941,
0.0037894771,
0.0022695709,
0.013430248,
-0.0207502842,
0.0254092105,
0.0386208445,
0.017129207,
0.0184380636,
0.0269075949,
0.0612346642,
-0.0295489971,
-0.0151048871,
0.0335315056,
0.0306272078,
0.0719408765,
0.0315974429,
-0.0275995489,
-0.0318897292,
-0.0244099461,
0.0356913544,
-0.0111271795,
0.0206108522,
0.0400472134,
-0.0200468395,
0.0335414149,
-0.0110953087,
-0.0354014263,
-0.0575386658,
-0.0449062474,
-0.086881645,
-0.0392972752,
0.0422486812,
0.0493001752,
0.0144064063,
-0.0013100215,
-0.0166406464,
0.0224137958,
-0.0294090789,
-0.0263973679,
-0.0049818614,
-0.0192768611,
-0.0603833012,
-0.0223036539,
-0.0281247254,
0.023840839,
0.0176391546,
0.0073580714,
0.0560660809,
0.0546061322,
-0.0284393933,
0.0001154814,
-0.0287144054,
-0.0063423072,
0.0239507779,
0.0563617721,
0.0700128376,
0.0491119064,
0.0333748572,
0.0502228737,
0.0348002128,
-0.0210907757,
-0.0561047457,
0.0329183489,
0.032768853,
0.0261617173,
-0.0641646013,
-0.0093790032,
0.0451550372,
-0.0235642437,
-0.0344730951,
0.0060514826,
0.0607092418,
0.0056222933,
0.0651762784,
0.0017157916,
-0.03820014,
-0.0642231405,
-0.0610056669,
-0.010891702,
-0.0170729179,
-0.063873373,
0.0141577916,
0.0334843211,
-0.0194303337,
-0.0122361248,
0.0654861107,
-0.0335995667,
-0.0367954597,
0.017832255,
-0.053458035,
0.0369478911,
-0.02073415,
0.0473055951,
-0.0277447831,
-0.0218966808,
0.0040314244,
-0.0114045888,
0.0180557333,
0.0087101599,
-0.0110339643,
0.0330568478,
-0.0505257547,
-0.0147305978,
-0.0344563648,
0.0213972908,
0.0112225553,
0.0356453843,
0.0526135154,
-0.0268803947,
-0.0041506933,
0.0532929525,
-0.022359686,
0.0312347636,
-0.0339468978,
-0.0134332506,
0.0146602001,
0.0252424777,
-0.0536118522,
-0.0059189172,
-0.0632255897,
-0.0221521072,
0.0068623186,
-0.0813286975,
-0.0572472848,
0.0541731268,
3.067e-7,
0.0120264953,
0.0331470259,
0.0650520548,
0.025833169,
0.0245747641,
0.0132224029,
0.0012485121,
0.0176520757,
0.0418076031,
0.003051836,
-0.0094405413,
0.0417284928,
0.0430165194,
0.0237318445,
-0.0841781944,
-0.0058179474,
-0.0491836444,
-0.0108077601,
-0.0314907022,
-0.0644231811,
0.0168953389,
0.0991505831,
-0.0294422638,
-0.0277521499,
0.0310554393,
-0.0047491281,
0.0230372213,
0.0109255752,
-0.0349566452,
0.0237513278,
0.0077661988,
0.033600498,
-0.0281512849,
-0.0321394391,
0.0055072601,
0.0083950078,
-0.0128608551,
0.0387902185,
-0.0047149193,
-0.0085993689,
0.0389734097,
-0.0111351861,
-0.0178591628,
-0.0154061299,
0.0507691354,
-0.0374516398,
-0.0049277968,
-0.0708058849,
0.0039816569,
0.0000605059,
-0.0400846303,
-0.0085577127,
0.0510811917,
0.0501259305,
0.0216608588,
-0.0272920281,
0.0039935373,
-0.0605759993,
-0.0413749367,
-0.000997439,
-0.0537125766,
-0.0606503896,
-0.0565885603,
0.0088213626,
0.0202368293,
-0.0358104222,
0.025364453,
2.943813691e-34,
0.0622715764,
-0.0071159359,
-0.024871001,
0.0872700363,
-0.0150580592,
-0.0314959288,
0.0558717325,
0.0448592976,
0.0036522588,
-0.0536405817,
-0.0555988699
] |
http://arxiv.org/abs/2401.15068v1 | http://arxiv.org/pdf/2401.15068v1 | Pairing Orthographically Variant Literary Words to Standard Equivalents Using Neural Edit Distance Models | 2024-01-26T18:49:34 | 2024-01-26T18:49:34 | Craig Messner, Tom Lippincott | cs.CL | cs.CL | We present a novel corpus consisting of orthographically variant words found
in works of 19th century U.S. literature annotated with their corresponding
"standard" word pair. We train a set of neural edit distance models to pair
these variants with their standard forms, and compare the performance of these
models to the performance of a set of neural edit distance models trained on a
corpus of orthographic errors made by L2 English learners. Finally, we analyze
the relative performance of these models in the light of different negative
training sample generation strategies, and offer concluding remarks on the
unique challenge literary orthographic variation poses to string pairing
methodologies. | Pairing Orthographically Variant Literary Words to Standard Equivalents Using Neural Edit Distance Models: We present a novel corpus consisting of orthographically variant words found
in works of 19th century U.S. literature annotated with their corresponding
"standard" word pair. We train a set of neural edit distance models to pair
these variants with their standard forms, and compare the performance of these
models to the performance of a set of neural edit distance models trained on a
corpus of orthographic errors made by L2 English learners. Finally, we analyze
the relative performance of these models in the light of different negative
training sample generation strategies, and offer concluding remarks on the
unique challenge literary orthographic variation poses to string pairing
methodologies. | [
0.0181299262,
0.0667535365,
-0.014374557,
0.0366967693,
-0.0663884655,
-0.0079588238,
0.0029861112,
0.0442090482,
-0.0170063972,
-0.02775166,
-0.0159084629,
-0.0154824471,
0.0768287107,
-0.027853746,
0.0291768331,
-0.0050255582,
0.051867798,
-0.0038980017,
-0.0494116545,
-0.0114776809,
0.0116064018,
0.0095674666,
0.0025783998,
0.0433621332,
0.0270711072,
-0.0347589143,
-0.0020444999,
0.0259134155,
0.0264709741,
0.0433669202,
-0.005573466,
0.0425905958,
-0.0460456461,
0.1063726172,
0.0000014806,
-0.0672290921,
-0.0179775655,
-0.0196678899,
0.0074100872,
0.0059173414,
0.0346699916,
0.0211570691,
-0.0515896007,
0.0011179524,
-0.0069327666,
0.0208573695,
0.0776877925,
-0.0019933642,
0.0068319528,
0.070644103,
0.0036155505,
-0.0323543288,
0.0363270342,
-0.0434288792,
0.003608377,
0.0159212165,
-0.0006691372,
-0.0429949053,
-0.0281731449,
0.0257316995,
-0.0256620739,
-0.007502296,
-0.0630235896,
-0.0041997852,
-0.0119990371,
0.0068200305,
-0.0230008196,
0.0439878292,
0.0179712474,
0.0168471336,
-0.0195948817,
-0.0109749539,
-0.0025815808,
0.0061704698,
-0.0227080695,
0.0449221954,
0.032446757,
-0.0341181234,
-0.0417093597,
-0.0069646453,
0.0346318968,
-0.0767602548,
0.0279677752,
0.0124528892,
-0.0233539287,
0.0389867984,
0.0628844947,
-0.0633037761,
-0.0286953542,
-0.0255607944,
-0.0535271391,
-0.0445754789,
0.0451035909,
-0.0197179709,
0.0265749674,
-0.0000496823,
-0.0426722802,
-0.1111039743,
-0.0159376003,
-0.0290578138,
0.0364926159,
0.0185778756,
0.0087229926,
-0.0015850918,
-0.0430397987,
0.0402212031,
0.0290303361,
-0.0057236031,
-0.0263484549,
0.0173070561,
-0.0603243262,
0.0058481297,
-0.0481614172,
0.0012024657,
-0.0649399161,
-0.0034004699,
-0.035445597,
-0.0487650931,
0.0115342801,
0.0117817586,
0.0224116128,
-0.0037459701,
-0.0016394637,
-0.0268435068,
-0.0525423251,
0.0040274849,
-0.0103674978,
0.0082099931,
-0.0497383587,
0.0043988284,
-0.0154919149,
0.0057157478,
-0.004822439,
-0.0727841184,
-0.0065572723,
0.0239865594,
-0.0339082517,
-0.0112300282,
-0.0045494009,
0.0067007076,
-0.008520755,
0.0241377633,
-0.0437876768,
-0.0836601779,
0.0385215767,
-0.0562225506,
0.0061567011,
-0.1013174504,
0.0021427036,
0.0148247834,
-0.0797581971,
0.0912441537,
-0.0289062466,
0.0112437196,
0.0508395955,
-0.0049256356,
-0.0247810092,
-0.0130662983,
0.0833364651,
-0.0451949537,
0.0724226981,
0.0067061195,
-0.0181961879,
0.069406569,
-0.0224139839,
-0.0527554341,
-0.0143331457,
-0.0468043089,
0.0692744553,
0.0880611688,
0.0506295823,
0.0705927908,
-0.0456939787,
-0.0500376448,
0.0506666303,
0.091659762,
0.0916279703,
0.0219136775,
0.0391546339,
-0.0132728266,
0.0133889373,
0.0147764292,
-0.0056365584,
-0.0151203889,
-0.0089451345,
-0.0146030625,
0.0474148616,
0.0052570188,
0.0079761576,
-0.0133843105,
-0.0169740431,
-0.0231368877,
0.0115750907,
-0.0730175748,
0.0041622934,
-0.0144170579,
-0.0509665608,
0.0116836783,
0.0008457177,
-0.0463689528,
0.0272042472,
-0.0159060471,
-0.0195590612,
0.0997287929,
0.0781367645,
0.0017598895,
0.0005065594,
0.0186265372,
-0.0636129901,
0.0463556014,
0.0271156747,
-0.0370191969,
0.0492697544,
-0.0204910897,
0.0214411113,
0.0242181905,
0.0120738586,
-0.0000894233,
-0.0261868648,
0.0178074259,
-0.0238179248,
-0.0134774353,
-0.054878965,
0.0470781997,
-0.0544362217,
-0.0553940088,
-0.007181434,
0.0239418782,
0.00034432,
-0.0245455783,
-0.0318831578,
0.0487925708,
-0.0204425696,
-0.0339350216,
0.0125404764,
0.0303995889,
-0.0043924875,
-0.0292636361,
0.0740945339,
-0.0379706547,
0.0295043718,
0.0124187553,
-0.0021346931,
-0.0536921769,
0.0268303417,
-0.034606494,
0.0285246428,
-0.0144460807,
0.0544335656,
0.0163705461,
-0.0341330431,
-0.0694311112,
0.0201978032,
0.0179590564,
0.0506484285,
-0.0306506436,
0.0379228778,
0.0243753791,
-0.0126024643,
-0.048066631,
-0.005530193,
0.024961317,
-0.0072812373,
-0.0112623852,
-0.0201780107,
-0.0276335124,
-0.0997533649,
-0.0404494926,
0.0088120718,
-0.0034366907,
0.03457487,
0.0534762703,
0.005793714,
-0.008302358,
-0.0111206053,
-0.0015465678,
-0.0116400756,
-0.019089248,
0.0241143424,
0.0490268655,
0.0080562951,
-0.0310517866,
-0.0377900638,
-0.010083938,
-0.0228462834,
0.0188789442,
0.0037212141,
-0.0432323515,
-0.0736645684,
0.0110850837,
-0.0240579993,
0.007415127,
0.0046561672,
-0.0194024332,
0.015514872,
-0.0144171519,
0.0731063932,
0.0479920246,
-0.0058042533,
-0.0452547595,
0.0513811931,
-0.0349419713,
-0.0355826281,
-0.0361746326,
0.0258053858,
-0.0125608239,
-0.0136778587,
-0.0556354336,
-0.0080149574,
0.0428357013,
-0.0369236618,
-0.0194697436,
-0.0236304738,
0.0481376685,
0.0297038499,
-0.005714796,
-0.0036037192,
-0.0300758202,
-0.0464842021,
0.0233900547,
-0.0503906645,
-0.0152197955,
0.0327765346,
-0.0286125764,
0.0535409935,
-0.1118167564,
0.0136661595,
-0.0158670489,
0.0359955914,
-0.0115954159,
0.0371270217,
-0.0221482292,
-0.0553429015,
-0.0642083138,
0.0082531339,
0.0463022441,
-0.0289265439,
0.0779251382,
-0.0600424856,
0.0275172982,
-0.0010757315,
0.0174942613,
-0.0025320493,
-0.0514267944,
0.0024002928,
0.0131364614,
0.0346095301,
-0.0440347642,
-0.0005509391,
0.0121508185,
-0.0467512049,
0.0963553116,
0.0080704307,
0.0030159622,
-0.0286009032,
0.0389252044,
0.006100751,
0.0004959317,
-0.0105396733,
0.0011558956,
0.0377037451,
0.0174093768,
-0.0088871755,
0.0008013367,
-0.0252717715,
-0.0184895284,
0.0063356971,
-0.08805006,
-0.026897233,
-0.0095415246,
0.0015202335,
0.0019079263,
-0.0180971976,
0.1074681655,
0.0337040722,
-0.0513477102,
-0.0004303466,
-0.0735542476,
0.0226673186,
0.0162511505,
0.0350613408,
-0.0406488739,
0.0103466259,
0.0143254604,
0.0317519791,
0.0176594127,
-0.0811528638,
-0.0003939486,
-0.1249298006,
-0.0308045633,
0.0261415746,
-0.0228979047,
0.0077991802,
0.0056340192,
0.0239397772,
-0.0296728332,
-0.0387561843,
0.0302633308,
-0.014266612,
0.0115432572,
-0.0049595409,
0.0237287562,
0.0040097251,
-0.0377245843,
-0.1252692491,
0.027870914,
0.0361864977,
0.0502628349,
-0.0155089665,
-0.0302610472,
-0.0169063117,
0.036139641,
0.022740921,
0.0005158288,
0.0391903333,
0.0186617058,
0.0075456593,
0.0297491718,
-0.0086832978,
0.0387891494,
-0.0446554907,
0.0523260608,
0.0161380731,
-0.0072053843,
0.0453032218,
0.0196711794,
0.0284654275,
-0.0046289042,
0.0404388607,
0.031675186,
-0.0330786705,
0.0167016573,
-0.0035437485,
0.0572468825,
-0.0456978418,
-0.0253593307,
-0.0143849878,
-0.0203562994,
0.0140082706,
-0.1140228957,
-0.0381101482,
-0.0259866305,
0.0729026496,
0.0711455196,
0.023200497,
0.0106576253,
-0.0164393522,
-0.0260030888,
0.0162118301,
0.0041948641,
0.0021819277,
-0.0775341541,
0.0082711084,
0.0154448543,
-0.0014536155,
-0.0071232421,
-0.0389685445,
-0.054502707,
-0.0667251945,
0.0123603074,
0.0448755473,
0.0290632397,
0.054122705,
-0.0347148143,
-0.0665584058,
-0.0430007167,
0.0368713513,
0.067423597,
0.051336024,
-0.0042120893,
0.032032989,
-0.0272516087,
0.0332697146,
0.0311581381,
-0.0362887681,
0.0182417464,
0.0452074409,
-0.0058359168,
-0.0298472159,
0.0344131552,
0.0098339673,
0.0485545062,
-0.0394538939,
0.0692945197,
-0.0539926589,
0.0204583928,
-0.0204913821,
0.0180545989,
0.021479683,
-0.0060685235,
0.0325994976,
-0.0501302332,
-0.0321832746,
-0.0023492409,
-0.0143593978,
0.0242604557,
-0.0540003441,
0.0033609702,
0.0329924412,
0.0655505359,
-0.0001212604,
0.0315818675,
0.0137680611,
0.0387534089,
-0.0022531308,
0.0629155338,
-0.0140840663,
0.0024854145,
0.0083425809,
0.0435755812,
-0.0341692902,
0.0386141986,
-0.0426559933,
-0.0247716792,
-0.019926019,
-0.0159750599,
0.0304597281,
-0.0301777739,
0.0553977117,
-0.0073834369,
-0.0413809121,
-0.015026723,
-0.0218594391,
0.0107541895,
0.0370130874,
-0.0240173806,
0.0161350034,
0.0348106697,
0.0068008704,
0.0571174771,
-0.0859877095,
0.0489333607,
0.0333844945,
-0.0465360656,
-0.0057334225,
-0.0185612626,
0.0050437422,
-0.0002261837,
-0.0200547874,
-0.0353429243,
-0.018074939,
-0.0370182618,
0.0451590121,
0.0363549702,
0.0076354821,
0.0318744928,
0.0049580005,
0.0339602418,
-0.0237636808,
0.0537873618,
-0.0361325853,
0.0182189457,
0.0039988919,
-0.0773608983,
-0.0608152226,
-4.833802115e-33,
-0.0200475156,
0.0196848717,
0.0082372408,
-0.0254636258,
-0.0090180375,
-0.0179887135,
-0.0153189022,
0.0118471393,
0.0042744866,
-0.0156363975,
-0.0285885204,
0.0200103857,
-0.0018297768,
-0.0328426249,
0.0156754237,
0.0290743206,
0.0338221714,
-0.0087477071,
-0.0099365823,
-0.0556037687,
0.0853322595,
0.0258458238,
0.0465106331,
-0.032002978,
0.0395461954,
-0.0284062084,
-0.0047993874,
0.0104070753,
-0.0012529582,
0.047716137,
0.0147507871,
-0.0300956126,
-0.0038820147,
0.0169027261,
-0.0281325318,
0.007331349,
-0.0568421893,
-0.0437809303,
0.0224096384,
-0.0082758376,
0.0235694908,
-0.0392929651,
0.0346623585,
-0.000611526,
-0.0738003775,
-0.0112316906,
0.0269591194,
-0.011748625,
-0.0088353809,
0.0317115821,
-0.0557185002,
0.0214681216,
0.0208327491,
-0.0167064779,
0.0106539307,
-0.0242727678,
0.025437789,
-0.0324943364,
-0.0785204247,
0.0071174977,
0.0158108957,
0.0699696988,
-0.0085739316,
-0.0055424212,
-0.0194226429,
0.0357912034,
0.0475856625,
0.0052620494,
0.0323587656,
0.0446066111,
0.0413740762,
0.0005590617,
0.0542719774,
0.0713941008,
0.0790041611,
-0.0037236507,
-0.0123834405,
0.0184622146,
0.0162458848,
0.0258218274,
-0.0468198992,
0.0121906167,
-0.0035458815,
-0.0138536319,
-0.0414656214,
-0.0409986302,
-0.0349339768,
-0.0177288931,
0.0123221474,
-0.0136849824,
0.023715673,
-0.001539765,
-0.0027074588,
-0.0289005283,
-0.012308307,
-0.0154409399,
-0.0203414205,
0.0249704905,
-0.0225033518,
0.0075675524,
0.0116808433,
-0.0111278621,
0.0185203459,
0.0308765378,
0.0467532836,
-0.0089490218,
0.0422303937,
-0.0230814014,
-0.0528949089,
-0.0081063416,
-0.0251903031,
-0.0198926441,
-0.0367164798,
0.0015166602,
0.0285756718,
0.0041875462,
0.0393622443,
0.0111241508,
0.0630892739,
-0.0279588029,
0.0003444006,
0.0116143767,
0.0241434015,
0.0326602049,
0.0048090466,
0.0380267911,
0.0315082371,
-0.0101839406,
0.0411018543,
-0.0349679887,
0.0021931352,
0.0116138179,
2.325e-7,
0.0557344034,
0.0299512148,
0.0629893988,
-0.0299174506,
-0.0027291295,
-0.0040443516,
-0.0265408847,
0.0321776606,
-0.0165225416,
0.0438548252,
0.0350754559,
-0.0272117928,
0.022326706,
0.012161823,
-0.0261845943,
0.0078342957,
-0.0803439543,
0.0274876263,
-0.0520091802,
0.0060297479,
0.0507943369,
0.0318176448,
0.0249180105,
-0.0306368619,
-0.0321135521,
-0.0170226227,
0.0242286902,
-0.0126245888,
-0.0047814352,
0.0063877366,
0.033152476,
-0.0253785253,
0.0269345567,
0.0656788871,
-0.0296320897,
0.0151491836,
0.0485944115,
-0.0173193626,
-0.0256463308,
0.0357836261,
-0.0044173501,
-0.0250930227,
0.005492745,
0.0005027781,
0.0406622775,
0.0407734737,
0.0106459139,
0.0247937813,
-0.0563918576,
0.0644561499,
0.0318575837,
-0.0360411927,
-0.0204746276,
0.0170299504,
0.0455290079,
0.0337893367,
0.0110803265,
-0.0350293294,
-0.000726714,
0.0122923134,
0.0027499474,
-0.0162750222,
-0.0049102711,
0.0262425374,
0.0176814795,
-0.0618777648,
-0.0399245881,
1.511331595e-34,
-0.0104228081,
-0.0043816995,
0.0163971093,
0.016295407,
-0.0000394477,
-0.0127152251,
0.0070946156,
0.0009322966,
-0.0161101017,
-0.0447745621,
-0.0367890187
] |
http://arxiv.org/abs/2402.01720v1 | http://arxiv.org/pdf/2402.01720v1 | Deep Learning Based Amharic Chatbot for FAQs in Universities | 2024-01-26T18:37:21 | 2024-01-26T18:37:21 | Goitom Ybrah Hailu, Shishay Welay | cs.CY | cs.CY, cs.AI, cs.CL, cs.LG | University students often spend a considerable amount of time seeking answers
to common questions from administrators or teachers. This can become tedious
for both parties, leading to a need for a solution. In response, this paper
proposes a chatbot model that utilizes natural language processing and deep
learning techniques to answer frequently asked questions (FAQs) in the Amharic
language. Chatbots are computer programs that simulate human conversation
through the use of artificial intelligence (AI), acting as a virtual assistant
to handle questions and other tasks. The proposed chatbot program employs
tokenization, normalization, stop word removal, and stemming to analyze and
categorize Amharic input sentences. Three machine learning model algorithms
were used to classify tokens and retrieve appropriate responses: Support Vector
Machine (SVM), Multinomial Na\"ive Bayes, and deep neural networks implemented
through TensorFlow, Keras, and NLTK. The deep learning model achieved the best
results with 91.55% accuracy and a validation loss of 0.3548 using an Adam
optimizer and SoftMax activation function. The chatbot model was integrated
with Facebook Messenger and deployed on a Heroku server for 24-hour
accessibility. The experimental results demonstrate that the chatbot framework
achieved its objectives and effectively addressed challenges such as Amharic
Fidel variation, morphological variation, and lexical gaps. Future research
could explore the integration of Amharic WordNet to narrow the lexical gap and
support more complex questions. | Deep Learning Based Amharic Chatbot for FAQs in Universities: University students often spend a considerable amount of time seeking answers
to common questions from administrators or teachers. This can become tedious
for both parties, leading to a need for a solution. In response, this paper
proposes a chatbot model that utilizes natural language processing and deep
learning techniques to answer frequently asked questions (FAQs) in the Amharic
language. Chatbots are computer programs that simulate human conversation
through the use of artificial intelligence (AI), acting as a virtual assistant
to handle questions and other tasks. The proposed chatbot program employs
tokenization, normalization, stop word removal, and stemming to analyze and
categorize Amharic input sentences. Three machine learning model algorithms
were used to classify tokens and retrieve appropriate responses: Support Vector
Machine (SVM), Multinomial Na\"ive Bayes, and deep neural networks implemented
through TensorFlow, Keras, and NLTK. The deep learning model achieved the best
results with 91.55% accuracy and a validation loss of 0.3548 using an Adam
optimizer and SoftMax activation function. The chatbot model was integrated
with Facebook Messenger and deployed on a Heroku server for 24-hour
accessibility. The experimental results demonstrate that the chatbot framework
achieved its objectives and effectively addressed challenges such as Amharic
Fidel variation, morphological variation, and lexical gaps. Future research
could explore the integration of Amharic WordNet to narrow the lexical gap and
support more complex questions. | [
0.0892998949,
0.0138828252,
-0.0218964275,
0.0185057446,
-0.0240466148,
0.0282220803,
0.0372725576,
-0.027203653,
0.0397182927,
-0.0228921566,
0.0230293926,
-0.0795590729,
0.0015273044,
0.0464878939,
0.0016702102,
-0.0296240691,
0.0366486572,
-0.0076956833,
0.0305726714,
-0.0130394008,
0.0297228359,
0.0475394614,
-0.0524305776,
0.0460539982,
-0.0493234918,
-0.0136208711,
-0.0426730551,
-0.0372549072,
-0.0225334726,
0.0364953615,
0.049488537,
0.0445586294,
-0.0276427716,
0.0468851812,
0.0000020355,
-0.0700786039,
-0.0135019328,
-0.039426621,
-0.0405997261,
-0.0332029872,
0.0740091428,
0.0503487922,
-0.0344993845,
0.0436175279,
-0.067410633,
0.0250045769,
0.0619122349,
0.0506225601,
0.0334227383,
0.0725835338,
-0.0048974822,
-0.0826988295,
0.0594023988,
0.0190763324,
-0.002935796,
-0.0316620059,
0.0077173188,
0.0071160682,
-0.0107745323,
-0.005182703,
-0.010606667,
0.0284091737,
0.0286055114,
-0.0619880818,
0.0391395949,
-0.0076038945,
-0.0387977026,
-0.0866355523,
0.0065043443,
0.0568920746,
-0.0193638764,
-0.0201348346,
0.0055282786,
0.0262080785,
0.0332171358,
0.0316634625,
-0.0410151891,
0.0278759282,
-0.0149048204,
-0.0067889029,
0.0335677713,
-0.0282984171,
-0.0082104653,
0.0393720306,
-0.0239636209,
-0.0019526005,
0.0134276655,
-0.0233746246,
-0.0033806921,
-0.0049016965,
0.0121150222,
-0.0484913848,
0.0662231892,
0.0571021885,
-0.0397958048,
-0.0008540989,
-0.0031855141,
-0.0658388287,
0.0269186981,
-0.0668788329,
-0.049398724,
0.0382274576,
-0.0340046398,
0.0255370364,
-0.0382313356,
0.0452583097,
0.0168221649,
0.0015800401,
-0.0048437924,
-0.0118698031,
-0.0597901009,
-0.0310712978,
-0.0083642155,
0.0737299174,
-0.0263532009,
-0.0088211382,
-0.0593007132,
0.0188192707,
0.0393012762,
0.0106810788,
-0.0129526919,
0.0356285833,
-0.0193129368,
0.0031778952,
-0.0366833545,
0.0258064549,
-0.0662215725,
0.0184868313,
-0.0189846661,
-0.0870999545,
0.0529876761,
0.0133997854,
0.0304984059,
0.0000485979,
0.0236811601,
0.0565473512,
-0.031565804,
-0.0186798237,
-0.0353359953,
-0.0325268693,
-0.0007039705,
-0.0309551042,
-0.0262936242,
-0.0848848447,
0.0292551108,
0.0044033653,
-0.0225967113,
-0.0022024373,
0.0102057774,
0.0302391425,
0.0164332017,
0.0617496483,
-0.0118531566,
-0.0142977983,
0.1407737285,
0.0022947141,
0.0621521436,
-0.0122531867,
0.0073411069,
0.026836291,
0.0486259945,
0.0063051633,
0.03992071,
0.0003404169,
-0.0218145568,
-0.0592163764,
0.0190531965,
0.0039790091,
-0.0171239749,
0.0619630218,
-0.0302790999,
0.0817375332,
0.0089063691,
-0.0038641307,
0.0073506627,
0.0779612511,
-0.0369876511,
0.0567364842,
0.0108133424,
0.0506645329,
0.0340131298,
-0.0036751467,
0.0048519173,
0.0222951081,
-0.0111297704,
-0.0162503328,
0.0411252081,
0.0329755098,
-0.0226943772,
-0.0805794746,
0.0181541126,
-0.0439476483,
0.012733967,
0.0104625756,
0.0669291392,
-0.0219785031,
-0.014799634,
0.008152944,
0.0383759961,
-0.0434616543,
-0.0185225625,
-0.054499492,
0.0309638698,
0.0599561743,
0.0317026898,
-0.0234583151,
-0.0161045473,
-0.0483213849,
-0.018696066,
0.0704329535,
0.0372856036,
-0.0102619985,
-0.0148558654,
-0.0523755476,
0.0161929559,
0.0273999814,
-0.0075885188,
-0.0167678818,
-0.0025998845,
0.0724112019,
0.0134624997,
0.0011917109,
-0.0404136628,
0.0412580222,
-0.0381939672,
-0.0197139308,
-0.0134294303,
0.0238809716,
0.0058786315,
0.062815398,
-0.0058041494,
-0.0013245764,
0.0116266394,
-0.0204839073,
-0.0369928442,
-0.0103128664,
-0.0579748601,
-0.0204402041,
0.0659949332,
-0.0556620471,
0.0018881508,
0.0744407251,
-0.0057132528,
-0.0482301377,
0.053838186,
-0.0585432202,
0.04028989,
-0.0279890504,
0.0496198833,
-0.0170296095,
0.0304277092,
-0.0316126011,
0.0241947155,
0.0244560149,
0.0117043676,
0.0148495352,
-0.0134007027,
0.039851293,
0.0202373806,
-0.009964725,
0.0355385244,
-0.0375738516,
-0.0165528394,
0.0357105397,
0.0179977324,
-0.0488600023,
0.0096270479,
-0.115525499,
-0.0051062065,
-0.0107948193,
-0.0061283908,
-0.0206904598,
0.0506981723,
-0.0010514085,
-0.0364297666,
-0.0643483326,
0.0187041853,
-0.0050812541,
-0.0228566546,
0.0739504918,
-0.0315913558,
-0.0140703432,
-0.0472335778,
-0.0216951258,
0.0055634235,
0.029632289,
0.0619355291,
-0.0627872273,
0.0318773389,
-0.0124882692,
-0.0362863019,
0.0236156471,
0.0243342444,
-0.0631660074,
0.0052138078,
-0.0233982448,
-0.0003503878,
0.0692302808,
0.0489328951,
0.0150554888,
0.0014472414,
0.046963267,
-0.0474907607,
0.0157110412,
-0.0040930063,
-0.0140220486,
-0.0135063548,
-0.0338246562,
-0.0260903183,
0.0017892373,
0.0101221185,
0.0079069445,
-0.0221425593,
0.0533226393,
-0.0262075234,
-0.0244917274,
-0.0328221619,
0.0226021539,
-0.0392947011,
0.0306862853,
0.0115202283,
-0.0108382376,
0.0748517737,
0.0192350112,
-0.023090858,
-0.0638629869,
-0.0279028751,
-0.0318063535,
0.0197095852,
-0.0522085689,
0.0377361625,
-0.0122780874,
-0.0583754815,
-0.041212365,
0.0066687874,
-0.000322384,
-0.0460794121,
-0.0197879169,
-0.0087798825,
0.0005212714,
-0.0169347823,
-0.003071537,
-0.0372930504,
-0.0077170646,
-0.0414629728,
0.0061401092,
0.0329247452,
-0.015729839,
-0.0258922148,
0.0151504055,
-0.0017296982,
0.0224493928,
-0.0265386906,
-0.0059104166,
-0.0064986101,
0.0601983964,
0.0445887446,
0.0130295716,
-0.0011556002,
0.0353890173,
0.0585185066,
0.0219563246,
-0.0144237196,
0.013377049,
-0.0006474484,
-0.0276952051,
0.0142310234,
-0.0315256193,
-0.0145955868,
-0.0171006527,
0.0174193047,
-0.0524501689,
-0.0102363164,
0.1314854175,
0.0523865633,
-0.0422360078,
-0.0073360591,
-0.0692031011,
-0.0308843628,
-0.0241119973,
0.0418114029,
-0.1132397652,
-0.0210347362,
0.0077447579,
0.007483752,
-0.0093195625,
-0.0072313165,
0.038591627,
-0.0936716422,
-0.0301764049,
-0.0218741428,
-0.0469275638,
-0.0489130206,
-0.0295834262,
-0.0251109619,
0.0091325846,
0.0365719125,
-0.0139933666,
0.0282726064,
-0.0045711412,
0.0380583778,
-0.0208455268,
-0.0312695131,
0.00077442,
-0.0346722864,
0.0573507957,
0.0502405502,
0.0516612753,
-0.0365449823,
-0.0024192629,
-0.0012049159,
0.0279209558,
-0.0267039482,
-0.0174348764,
-0.0028678353,
0.0011452343,
-0.0104063191,
-0.0360787585,
0.0086771911,
-0.0314335898,
0.0373234525,
0.0245645344,
0.045867227,
-0.0295373052,
-0.0109645836,
-0.0456737466,
0.0033870505,
-0.0100609763,
0.0445184447,
0.0093196295,
-0.0561563149,
-0.0399517417,
-0.0305858459,
-0.0124993473,
-0.0373897329,
-0.0347105786,
-0.0248231329,
0.0139303682,
-0.0105607649,
-0.0943444446,
-0.0746959299,
0.0297067091,
-0.0310529787,
0.0128587019,
0.0111180814,
0.0348423123,
0.036888916,
-0.042502176,
-0.0188188199,
0.008784648,
0.0268425196,
-0.0522807129,
-0.0059465156,
0.0034243057,
-0.0578383915,
-0.0117270481,
-0.0213892069,
-0.0114986505,
-0.0423930772,
0.0095703471,
0.0032134026,
0.0125265904,
-0.0119391717,
-0.0358345732,
-0.0832770318,
0.0135953827,
0.0279026236,
-0.0466929041,
-0.0200316068,
0.0130384164,
0.0429058336,
-0.0064660828,
0.0323610194,
-0.0270438362,
0.0066305581,
-0.0237676781,
-0.0122947525,
-0.0103781996,
-0.0525210164,
0.0032330193,
0.0146394903,
-0.0119875753,
0.0031746763,
-0.0606972612,
-0.0385265686,
0.0190967117,
0.0257405173,
-0.0045775459,
-0.0034094506,
0.0336114876,
0.0140739866,
0.00741367,
0.0050433804,
-0.0174263902,
-0.0183272194,
0.0017018074,
0.0364283957,
0.0234059878,
0.0307228118,
0.0284570735,
-0.0129457908,
0.0042168158,
-0.0164189115,
0.0080741979,
0.0256107803,
0.0416403413,
-0.0307762176,
0.0109856678,
-0.0335498191,
-0.0046959347,
0.0264147837,
0.0327147506,
0.008308243,
-0.0597726777,
-0.0254893694,
0.0717714354,
0.0434541851,
-0.0141780823,
0.0043632053,
-0.0066503622,
0.0140104452,
-0.0314277075,
0.0229432099,
0.0031615726,
-0.0119968802,
0.0120581593,
0.0014624239,
-0.0109358784,
0.0030010729,
0.016047487,
-0.0130200544,
-0.0080854204,
0.001586213,
-0.000851865,
-0.0391313434,
0.0108946003,
0.0330027193,
0.0347752646,
-0.0089640841,
0.0245663449,
-0.001140081,
-0.0530408956,
-0.0271072611,
-0.019955704,
0.0013333278,
-0.003339364,
0.0405055061,
0.07162828,
0.0095169796,
-0.0113727255,
-0.0602048747,
0.049981229,
-0.0218534563,
-0.0475306585,
-0.0222193711,
-6.263900828e-33,
-0.0090821823,
-0.0400584452,
-0.0021417278,
-0.0412496254,
-0.0141389174,
-0.0300445184,
-0.0325680599,
0.0206543952,
-0.0645398125,
0.0006231673,
-0.0357479192,
0.0302442554,
0.0284000095,
0.0469755568,
0.0413672738,
0.0138436528,
0.0367281847,
-0.0051706312,
0.023123432,
0.0497741066,
0.0195713285,
0.0679292828,
0.0526481159,
-0.0372422002,
0.0452350192,
-0.0479329452,
0.022563044,
0.0139848543,
0.01630692,
0.0483227707,
-0.0434001051,
0.0503921323,
0.0180004705,
-0.008806048,
0.0238981396,
0.079865776,
-0.065594241,
-0.0326048695,
-0.0000793643,
0.0546156317,
-0.035400074,
-0.0615223944,
0.0545649827,
-0.0268040206,
0.0024416568,
0.0549304634,
0.0733857229,
-0.008795822,
0.0221814327,
-0.0147693641,
-0.0760797486,
0.0245542787,
-0.0210502185,
-0.020288216,
0.0024633368,
0.092711702,
-0.0067625665,
-0.0435886495,
-0.0228712633,
0.0204403829,
0.0147057949,
0.0272808336,
0.0480727106,
0.0267820414,
-0.0567048937,
0.019805016,
0.0461946912,
-0.0370648839,
0.0208985861,
0.0172845349,
0.0326146558,
-0.0180602763,
0.0592792295,
0.0382879339,
0.0199293811,
-0.0365931503,
-0.0558214486,
0.015411322,
0.058417175,
0.051513657,
-0.0074589797,
-0.0132721914,
-0.0513010845,
0.0075700744,
0.0001829783,
0.0350071155,
-0.0314064771,
-0.0261219945,
0.0184926316,
-0.0384529121,
-0.0326586179,
-0.0137949605,
0.0021353937,
-0.0236956459,
0.008105685,
0.005901359,
0.0177483242,
0.0103868162,
0.0019743596,
0.0268235672,
-0.0339375399,
0.0108361151,
-0.0201080367,
-0.0257815085,
0.0106688272,
0.0251765009,
-0.0087628886,
0.0014485121,
-0.037372224,
-0.0036145083,
-0.009756349,
-0.0003338371,
0.0237917844,
0.0888592005,
-0.031270802,
0.0023073428,
-0.0186789595,
0.0363186412,
0.0165720433,
0.0193094034,
-0.0489025079,
-0.0565439202,
-0.0035308069,
-0.0331193954,
-0.071730569,
0.0118321441,
-0.0656680167,
0.0583823472,
-0.0239134431,
-0.0957196206,
0.0034930317,
-0.0266496278,
2.809e-7,
0.0332109816,
0.0640741587,
0.0162971746,
0.0024158123,
0.0012949193,
0.0168211348,
0.0711660683,
0.0194427185,
-0.0296284929,
0.0222041979,
0.0068390989,
0.014307471,
0.048320692,
0.0156275015,
-0.1264270991,
-0.0442752205,
-0.0419307984,
-0.0071709235,
-0.0594707653,
0.0156121422,
0.0400732905,
0.1143730506,
0.0055490476,
-0.0427978337,
-0.0325717404,
-0.0560039282,
-0.019145906,
0.034524627,
0.0096012373,
0.0147633366,
0.0370767415,
0.0289720464,
0.0401284248,
0.0037519298,
-0.0060800775,
0.003550668,
0.0344482064,
0.0395843908,
-0.0128264045,
-0.0165005066,
0.0415558852,
-0.0030316105,
0.0029896409,
-0.0582408644,
0.0604998432,
0.006674699,
-0.0077629876,
-0.0151161831,
0.0059161442,
-0.0120339524,
0.0445704088,
-0.0213341787,
-0.0237053763,
0.0184144787,
0.0162309185,
-0.0432025939,
0.015012756,
-0.0409863703,
0.0072526517,
-0.0247647278,
-0.0242328309,
-0.0200072229,
-0.006890282,
-0.0256475229,
-0.0024563784,
0.0264139958,
0.0315151438,
2.134491111e-34,
0.0384948477,
0.0205037426,
0.0166081749,
0.0724645257,
0.0178682804,
0.0102799423,
0.0006746973,
0.0177839287,
0.0294335019,
-0.0369057804,
-0.020325195
] |
http://arxiv.org/abs/2401.15055v1 | http://arxiv.org/pdf/2401.15055v1 | Deep learning-based approach for tomato classification in complex scenes | 2024-01-26T18:33:57 | 2024-01-26T18:33:57 | Mikael A. Mousse, Bethel C. A. R. K. Atohoun, Cina Motamed | cs.CV | cs.CV, cs.CL | Tracking ripening tomatoes is time consuming and labor intensive. Artificial
intelligence technologies combined with those of computer vision can help users
optimize the process of monitoring the ripening status of plants. To this end,
we have proposed a tomato ripening monitoring approach based on deep learning
in complex scenes. The objective is to detect mature tomatoes and harvest them
in a timely manner. The proposed approach is declined in two parts. Firstly,
the images of the scene are transmitted to the pre-processing layer. This
process allows the detection of areas of interest (area of the image containing
tomatoes). Then, these images are used as input to the maturity detection
layer. This layer, based on a deep neural network learning algorithm,
classifies the tomato thumbnails provided to it in one of the following five
categories: green, brittle, pink, pale red, mature red. The experiments are
based on images collected from the internet gathered through searches using
tomato state across diverse languages including English, German, French, and
Spanish. The experimental results of the maturity detection layer on a dataset
composed of images of tomatoes taken under the extreme conditions, gave a good
classification rate. | Deep learning-based approach for tomato classification in complex scenes: Tracking ripening tomatoes is time consuming and labor intensive. Artificial
intelligence technologies combined with those of computer vision can help users
optimize the process of monitoring the ripening status of plants. To this end,
we have proposed a tomato ripening monitoring approach based on deep learning
in complex scenes. The objective is to detect mature tomatoes and harvest them
in a timely manner. The proposed approach is declined in two parts. Firstly,
the images of the scene are transmitted to the pre-processing layer. This
process allows the detection of areas of interest (area of the image containing
tomatoes). Then, these images are used as input to the maturity detection
layer. This layer, based on a deep neural network learning algorithm,
classifies the tomato thumbnails provided to it in one of the following five
categories: green, brittle, pink, pale red, mature red. The experiments are
based on images collected from the internet gathered through searches using
tomato state across diverse languages including English, German, French, and
Spanish. The experimental results of the maturity detection layer on a dataset
composed of images of tomatoes taken under the extreme conditions, gave a good
classification rate. | [
-0.0328428969,
0.0657167658,
0.0114006819,
0.0703925267,
-0.026611004,
-0.0138217285,
0.0472642668,
0.0028251312,
0.0263674557,
0.0438753776,
0.0373117812,
-0.0326025411,
-0.0308312364,
0.0070754532,
-0.0487248264,
-0.0076688579,
-0.0087144421,
0.0350629278,
-0.0044036773,
-0.0092751132,
-0.0062929993,
-0.0004541133,
-0.0043882178,
-0.0019925917,
-0.1276363432,
0.0137323663,
-0.0194447432,
0.013348409,
0.0241337046,
-0.0196058117,
0.0011537805,
-0.0086437101,
-0.0092096645,
0.0503631234,
0.000001882,
-0.0233159345,
0.0257881954,
-0.0359524637,
-0.0528659895,
0.0220396034,
0.0234130323,
-0.0180997886,
-0.0360010937,
-0.0295195561,
-0.0348891653,
0.0133683281,
0.0752226487,
0.0187452268,
-0.018259611,
0.0137117654,
-0.003160527,
-0.0063894428,
0.034599416,
0.0358691216,
-0.0191457663,
-0.0370506532,
0.0464069359,
-0.0420524739,
0.0208313782,
-0.0066475896,
-0.0284707919,
0.0427385047,
0.0677236989,
-0.0082701426,
-0.0426758043,
0.0206110142,
0.0075343829,
-0.0446106456,
-0.0166450702,
0.0218517985,
-0.0937672704,
0.0197544359,
-0.0510679111,
-0.0334074274,
-0.0790993199,
-0.0511503257,
-0.0210345574,
0.024858905,
-0.0042927358,
0.0136138052,
-0.0436716452,
0.0357271843,
0.0128495358,
0.0061161774,
-0.0043626535,
0.0500628725,
-0.0220974591,
0.0001021857,
-0.000954513,
0.005638306,
0.0342464969,
-0.0181772709,
0.0354811884,
0.0281709731,
0.0186816119,
-0.0574439205,
-0.0005567557,
-0.0313899517,
0.0264140014,
-0.0886845365,
-0.0440807305,
-0.0395564809,
-0.1084505841,
0.0073745307,
-0.012944825,
0.0407153405,
-0.020486135,
0.0532885492,
0.0008775852,
-0.0235580485,
-0.0645361021,
0.0264552459,
-0.0010524322,
0.025357483,
0.0401290767,
0.0005904352,
-0.0324049108,
-0.0345085189,
-0.00877818,
-0.0515825078,
0.016565986,
0.0684978142,
-0.028980637,
-0.0178324338,
-0.047680635,
0.0147538576,
-0.0238184314,
0.027508948,
-0.0102093602,
-0.0593636036,
0.0119958362,
-0.0061939182,
0.0383922122,
0.0271640066,
0.0729050264,
0.0291909855,
-0.0072834869,
-0.0062973867,
-0.0135403452,
-0.0202743616,
0.0035789798,
0.0205043498,
0.0130306585,
-0.0268177669,
0.0292279217,
-0.065623045,
0.0365920588,
-0.0255566072,
0.0070888391,
0.0208409354,
0.0171099175,
0.033239603,
0.0416193455,
-0.0265119933,
0.0331005082,
-0.0147651751,
-0.0012699568,
0.0292926636,
0.0023638625,
0.083088696,
-0.0236356296,
-0.0134742372,
0.0633275434,
0.009445874,
-0.0692767873,
-0.0295243282,
-0.0470158681,
-0.0721208453,
-0.0577078201,
0.0256734341,
0.0156234773,
0.0078323558,
0.0298822578,
-0.002281128,
0.0536086783,
0.0226153992,
0.0065054921,
0.0261190962,
0.0433745086,
0.0536853112,
0.000794115,
-0.0523962453,
-0.0132052433,
0.0747907385,
-0.0423679501,
-0.0422592796,
-0.1042818427,
-0.0131897386,
-0.0505859032,
-0.0783380494,
0.011602303,
-0.0259722192,
-0.010217227,
0.0279856063,
0.094286561,
-0.0123287374,
0.0155383125,
-0.0476353988,
-0.0104210936,
0.0273832809,
0.0166510828,
-0.0646656901,
0.0229496248,
0.0051328107,
-0.0187126081,
-0.0548251309,
0.007161072,
0.0414351858,
0.0395782515,
-0.0177312754,
-0.0109418407,
-0.0209543426,
0.0449185185,
-0.0040855478,
-0.0153568042,
0.010455288,
-0.0223689619,
0.0007000156,
-0.0224456303,
0.0259761047,
0.0065006916,
-0.0075340434,
0.0134031,
0.0243591368,
0.0903881416,
-0.0381492563,
0.025394205,
0.038334962,
-0.0399437584,
-0.008004427,
0.0195935369,
-0.0292800069,
0.0098193418,
-0.0335462578,
-0.0004071203,
0.0085156457,
0.02409181,
0.0176878199,
0.0270506777,
-0.0230045859,
-0.01135555,
0.0449384116,
-0.0343071893,
-0.0166226346,
-0.0114821261,
0.044333484,
-0.0407447629,
-0.040404465,
0.029156724,
-0.0217013229,
0.0410136543,
-0.0076424126,
-0.0066486546,
0.0085192705,
-0.0235059019,
-0.0210820977,
-0.034379445,
0.0245233122,
0.0063037183,
-0.0040918617,
-0.0293929931,
-0.075319685,
0.0347320922,
0.0213055313,
0.0872731581,
-0.024966199,
-0.0145489611,
-0.02854692,
0.0058374335,
-0.0237645041,
-0.028551165,
0.0157666691,
-0.0140397567,
0.0269477013,
-0.0028016192,
-0.0138132293,
0.1063777506,
-0.0594119728,
-0.0697915778,
0.0101479776,
0.0121807214,
-0.0115646441,
-0.026554998,
-0.0189134944,
0.0027914287,
0.0456601754,
0.0585605241,
-0.0296349823,
0.0576557033,
-0.0804306641,
0.0173400454,
0.0452647433,
0.0209062714,
-0.0470921397,
-0.020943312,
-0.01906709,
-0.0289927013,
0.0859346539,
-0.0093569737,
0.0089686355,
-0.0089252293,
0.0422574729,
-0.0170231108,
-0.0143669294,
-0.0295451917,
-0.0034536291,
0.0043428722,
-0.0209686831,
-0.042955149,
-0.0115851164,
0.0254412033,
0.0173884127,
0.0070535759,
0.0382442325,
-0.0030146448,
0.0011137612,
-0.0297240559,
0.0835134238,
0.0395317636,
0.0010879388,
-0.022270849,
-0.0529459603,
-0.0110664554,
-0.0000040445,
0.0070820856,
-0.0041381316,
-0.0729696378,
-0.0102088936,
0.026791472,
0.0178593863,
0.0310392156,
-0.0317320861,
-0.0055050785,
-0.0239935461,
0.0389503874,
-0.0446392894,
-0.0205275714,
-0.0042912643,
-0.0441897139,
-0.017172426,
-0.0168645307,
-0.0202346537,
-0.0512136929,
0.0301633365,
-0.04551645,
0.0718979239,
0.0384062566,
0.0252040029,
-0.0073106177,
0.0142108807,
-0.0028027373,
0.0745993257,
0.0133269839,
-0.0152849313,
-0.0129035162,
0.0423244908,
-0.0140526919,
-0.0061788564,
-0.0333007947,
0.0228067823,
0.046767965,
-0.0324198566,
-0.0207872968,
0.0305400975,
-0.0887831226,
0.0435739756,
-0.0164345521,
0.0137461573,
0.023328213,
-0.0379823633,
-0.0240180213,
-0.0707546324,
-0.0221423153,
0.0158371162,
0.0037888817,
-0.0545348339,
-0.0023494682,
0.0165926628,
-0.0917609259,
0.0181851517,
0.0088353055,
0.0291846208,
0.0015210749,
-0.0020191208,
-0.0603956655,
-0.0334011167,
0.0186960045,
-0.0017115024,
-0.0104485126,
0.0289256386,
-0.0233263932,
0.0082608229,
-0.0286539067,
0.0506734513,
-0.0170095265,
-0.0061204876,
-0.015639957,
-0.0371578373,
-0.0013243555,
-0.0423250571,
0.0734506771,
-0.0152014969,
0.010093526,
0.0320565738,
-0.0477797873,
-0.0571168549,
0.0815574825,
0.02763585,
-0.0474505238,
-0.0236746669,
0.0475732349,
-0.0184330698,
0.0238426812,
-0.0218143985,
0.0132411886,
0.0410418101,
-0.0126368431,
0.0123029742,
0.0048315884,
0.0263655446,
0.035691347,
0.017141277,
0.0830155835,
-0.0069044079,
-0.0166972838,
-0.0213175695,
-0.0336510539,
0.0218242258,
-0.023247445,
0.0319365561,
-0.0457306951,
-0.0365837775,
0.0005359381,
-0.0492692441,
-0.0797856823,
0.0090606809,
0.013494879,
-0.0025751893,
-0.0384484977,
-0.0143877063,
-0.0308293253,
0.0410498753,
0.0232471339,
0.009464466,
-0.0178863294,
0.0064961989,
-0.0205096062,
-0.039477393,
0.0148241734,
0.0097121689,
0.0187318269,
-0.0371780358,
0.0029092394,
0.0686890036,
-0.0630767494,
0.0062410524,
-0.0195858739,
-0.0441949889,
0.0293593016,
-0.0049795839,
0.0456030555,
0.0027657454,
0.0079962974,
-0.018423466,
-0.0784743875,
0.0900745317,
0.0294162016,
0.1028155908,
0.0255596433,
0.0130666653,
-0.0570080951,
0.0055367541,
0.0029998713,
-0.0175787713,
0.0966590494,
-0.0020896595,
-0.0335131772,
-0.0453628637,
-0.021561252,
0.005225969,
-0.0348635577,
0.077475138,
0.0455978364,
-0.0142432712,
-0.026227532,
0.0272567626,
0.0582370311,
-0.0018963846,
0.0107020484,
0.0134758288,
0.0181429181,
0.0232914537,
0.0029296337,
0.0300806761,
0.0802604333,
-0.0204653647,
0.0407435633,
0.0115194796,
0.0190357976,
0.064233385,
-0.0501317233,
0.0446966887,
0.0188313648,
0.0045190551,
0.0115612671,
0.012641835,
-0.0046603438,
0.0083713122,
-0.0058461321,
0.0151276123,
-0.0787596405,
0.0389011577,
0.0310890079,
-0.0233727247,
-0.0309180617,
0.0593967028,
0.0494041666,
-0.0674484447,
-0.0084363138,
0.0040320521,
0.00007274,
-0.0411018021,
0.0444638617,
0.019932542,
0.0008246743,
0.0516290553,
-0.0123159243,
0.0195355359,
-0.006679133,
-0.0086424556,
0.0248235557,
-0.0507730916,
0.0284954626,
0.0157736652,
-0.0036713197,
0.0765696019,
-0.0037355833,
-0.0403830782,
0.0047177952,
0.0009936758,
0.0420679003,
0.01119297,
0.0517187864,
0.0151148029,
0.0510164835,
0.0031408444,
0.0111932568,
0.0082093393,
0.0042731883,
-0.0567374974,
0.010308695,
-0.0096808784,
-0.0357529633,
-0.0183321238,
0.0022385486,
-5.780977994e-33,
0.0054602586,
-0.035435196,
0.0087829204,
0.0058130058,
-0.0027514191,
0.0654510632,
0.0285499897,
-0.0639319271,
-0.018830698,
0.0286691058,
-0.0101407124,
-0.0003690607,
0.0227005947,
0.0113996621,
0.0662951246,
-0.0545649752,
0.0618809089,
0.0152093209,
0.0338578373,
0.0115047386,
-0.0133289993,
0.038464658,
0.0571121648,
-0.010995931,
-0.0591564588,
0.0109551875,
-0.0039564553,
-0.0259584952,
0.009814471,
-0.0113678379,
0.002642388,
0.0252826903,
0.0140473358,
-0.0346179828,
-0.0233639386,
0.0398452245,
-0.0343052559,
-0.0163447633,
-0.0918320492,
0.0552430637,
-0.0621256642,
-0.0004733248,
0.020089807,
0.0121012479,
0.0207200032,
-0.0028517607,
0.0707934499,
-0.0184915271,
-0.0187952034,
-0.0978077129,
0.0151558537,
0.0079434356,
-0.0038312997,
-0.0532477908,
0.1241994947,
0.001698588,
0.0257738382,
0.0467647351,
-0.0483417325,
-0.0293368753,
-0.0441959836,
0.0016688792,
0.0352550894,
0.0865454376,
-0.003633512,
0.0180678125,
0.0056623165,
-0.0525436588,
0.0087920763,
-0.0405508764,
0.021206053,
0.0187774096,
0.0898900107,
-0.0269598514,
0.0853494033,
-0.0596192554,
-0.0614058934,
0.0354157872,
-0.027349174,
0.0392628387,
-0.0479682907,
-0.0226174593,
0.0262676664,
0.0077991663,
-0.0122741265,
-0.0059909308,
-0.0301478785,
0.0138806468,
0.0136441449,
-0.0237000342,
-0.016569363,
-0.0019434843,
-0.034429457,
0.0038226743,
0.0457123592,
-0.0076534385,
-0.0018738792,
0.0199019723,
0.0112556843,
-0.0156855956,
-0.0636590198,
-0.0231822301,
-0.0556612574,
0.0571909733,
0.0106582688,
0.0296883434,
0.0068352884,
0.0740201548,
0.0046425471,
-0.0124382852,
-0.0135016181,
-0.0238896031,
0.0202794075,
0.0374535844,
-0.0136171393,
-0.0031340043,
0.02376955,
0.0661266074,
0.0007839889,
-0.0039359271,
0.0060834675,
-0.0096647823,
0.0388149731,
0.0057021007,
-0.0220408998,
-0.0315778553,
-0.0654252023,
-0.0603513867,
-0.0189975854,
-0.0830287337,
0.0067152069,
-0.0073098997,
2.646e-7,
0.0088704685,
0.0299523287,
0.0133058289,
-0.0014229023,
-0.0502914824,
-0.0058415001,
-0.0189807005,
0.0025327408,
-0.0536869504,
-0.0088913376,
0.035283491,
-0.0057576094,
-0.0196136218,
0.0455564559,
-0.0259389915,
-0.0311590564,
-0.0298456419,
0.0327323079,
-0.047864683,
0.0068533141,
0.0422397219,
0.0820995569,
0.0177348703,
-0.0291383881,
0.0597586408,
-0.0262654629,
0.0196548663,
0.0032937431,
0.0670508519,
-0.0156794749,
0.0604798496,
0.0219413023,
0.0264270846,
0.021223573,
0.000591217,
0.0177520066,
0.0089264866,
-0.0148910722,
-0.0212881062,
-0.0213234574,
0.0061035822,
0.0081979102,
0.0311431549,
0.0142855952,
0.0381965376,
-0.0445428938,
0.0202425569,
-0.0091487179,
0.0640665293,
0.0229650792,
0.0426369794,
-0.0099277711,
0.0225827713,
-0.0216923077,
-0.0140808402,
-0.0505810231,
-0.0242092349,
-0.0212827008,
0.0414120518,
-0.0080255717,
-0.0049432977,
-0.0088441782,
0.0017387937,
-0.0036899799,
-0.0250983443,
0.0807906985,
-0.0435292125,
1.624776557e-34,
-0.0042572473,
0.0096956892,
-0.0149868624,
-0.0346625187,
-0.0578606389,
0.0434897691,
0.0109541481,
-0.0077361502,
-0.0509039797,
0.0063054538,
-0.0405449606
] |
http://arxiv.org/abs/2401.15050v1 | http://arxiv.org/pdf/2401.15050v1 | LongFin: A Multimodal Document Understanding Model for Long Financial Domain Documents | 2024-01-26T18:23:45 | 2024-01-26T18:23:45 | Ahmed Masry, Amir Hajian | cs.CL | cs.CL | Document AI is a growing research field that focuses on the comprehension and
extraction of information from scanned and digital documents to make everyday
business operations more efficient. Numerous downstream tasks and datasets have
been introduced to facilitate the training of AI models capable of parsing and
extracting information from various document types such as receipts and scanned
forms. Despite these advancements, both existing datasets and models fail to
address critical challenges that arise in industrial contexts. Existing
datasets primarily comprise short documents consisting of a single page, while
existing models are constrained by a limited maximum length, often set at 512
tokens. Consequently, the practical application of these methods in financial
services, where documents can span multiple pages, is severely impeded. To
overcome these challenges, we introduce LongFin, a multimodal document AI model
capable of encoding up to 4K tokens. We also propose the LongForms dataset, a
comprehensive financial dataset that encapsulates several industrial challenges
in financial documents. Through an extensive evaluation, we demonstrate the
effectiveness of the LongFin model on the LongForms dataset, surpassing the
performance of existing public models while maintaining comparable results on
existing single-page benchmarks. | LongFin: A Multimodal Document Understanding Model for Long Financial Domain Documents: Document AI is a growing research field that focuses on the comprehension and
extraction of information from scanned and digital documents to make everyday
business operations more efficient. Numerous downstream tasks and datasets have
been introduced to facilitate the training of AI models capable of parsing and
extracting information from various document types such as receipts and scanned
forms. Despite these advancements, both existing datasets and models fail to
address critical challenges that arise in industrial contexts. Existing
datasets primarily comprise short documents consisting of a single page, while
existing models are constrained by a limited maximum length, often set at 512
tokens. Consequently, the practical application of these methods in financial
services, where documents can span multiple pages, is severely impeded. To
overcome these challenges, we introduce LongFin, a multimodal document AI model
capable of encoding up to 4K tokens. We also propose the LongForms dataset, a
comprehensive financial dataset that encapsulates several industrial challenges
in financial documents. Through an extensive evaluation, we demonstrate the
effectiveness of the LongFin model on the LongForms dataset, surpassing the
performance of existing public models while maintaining comparable results on
existing single-page benchmarks. | [
-0.0006183521,
0.0598281287,
0.0219230186,
0.0452674814,
-0.0116830599,
0.0287229288,
0.0466211215,
0.0328433774,
-0.0310746934,
-0.0389016755,
-0.0445539169,
-0.0330234878,
-0.0006502689,
0.0759070292,
0.0206091143,
-0.0678668991,
0.0200038459,
-0.0138967754,
0.0480733141,
0.0353397392,
0.0255522169,
-0.0027519325,
-0.0209112708,
-0.0015379474,
-0.0628721714,
-0.0513783656,
-0.0685507655,
-0.0030829364,
-0.017372271,
-0.0276158396,
-0.0141150635,
0.0093158567,
0.046024248,
0.0340848714,
0.0000019971,
-0.0316207521,
-0.0316725001,
0.0214214325,
-0.0204313565,
0.0038549739,
0.082153745,
-0.083039023,
0.023351552,
-0.0425139442,
-0.0159563497,
-0.004047886,
0.0181550533,
0.1122904643,
0.0498609282,
0.0705519542,
-0.0367784873,
-0.0654259622,
0.0489513576,
-0.0433482379,
0.0098522324,
-0.0948957577,
0.002549344,
-0.0018967396,
0.005311769,
-0.0233344026,
-0.0488298386,
0.0214788485,
0.019000249,
-0.024989536,
0.0162562113,
0.0177773815,
-0.0134443659,
-0.0089849066,
-0.0354353711,
0.0771347135,
0.0079197846,
0.0643370673,
0.017415097,
-0.0304599963,
0.0078569241,
-0.0005956291,
-0.0613294281,
0.0280513261,
-0.0376544967,
-0.0293840971,
-0.0317220017,
0.0103939474,
0.0030519285,
0.0008374519,
-0.0429061279,
-0.018972924,
-0.0300811362,
-0.0419168957,
0.0196655151,
-0.0020778349,
0.011262672,
0.0066610258,
0.0242799222,
0.0259884484,
0.0577919669,
0.0090625193,
-0.0237104446,
-0.0843829513,
0.0243931375,
-0.0361199826,
0.0293170493,
-0.0177608132,
-0.0796092525,
0.1026189327,
-0.0410792977,
0.0171223301,
-0.0420288444,
0.0152483014,
0.0088834278,
0.0184277482,
-0.0263010859,
-0.0053904555,
-0.0531577952,
0.038636446,
-0.0321177393,
-0.0758573636,
-0.0016836267,
-0.0009720968,
0.0077076592,
0.042823635,
0.0062132473,
-0.0096788844,
-0.0146672623,
-0.0360402055,
-0.0253317989,
0.0176310875,
-0.0429570191,
0.0232564732,
-0.0048512723,
0.0001863218,
0.0275303256,
0.0394375958,
0.0224036053,
-0.0304165017,
0.0080135865,
0.0667501688,
0.0108492766,
-0.0508791171,
-0.040654242,
-0.0461774208,
0.0073518204,
-0.0435180701,
0.0311324373,
-0.0054745348,
0.0711801201,
-0.0237060748,
-0.0191655308,
0.0261030681,
-0.0286472179,
-0.0036596099,
-0.0042785886,
0.094916217,
0.0130533092,
-0.0226943791,
0.0488386378,
-0.0260008629,
-0.0161899664,
0.0376664624,
0.0200064853,
0.0712112561,
0.0399356224,
-0.0173280891,
0.0121015059,
0.0564473495,
0.0066802856,
-0.0649802163,
0.0299136788,
0.0302361883,
0.010525763,
-0.0064047063,
0.0640393719,
0.004545697,
-0.0395230725,
-0.0367071889,
0.0445703678,
0.0416650474,
0.0439834967,
-0.0152890729,
-0.0017214462,
0.0375241712,
0.0386100225,
-0.0058489568,
-0.0193584412,
0.0275830925,
0.0264505241,
0.0304733533,
-0.0232171044,
-0.0325982943,
-0.0197736286,
-0.0205560736,
-0.0136427246,
-0.0342910923,
-0.0134248445,
-0.0215885136,
0.0779708326,
0.0018355678,
-0.0576213486,
0.003175633,
-0.0170038864,
-0.0364230871,
-0.0217029881,
-0.0135287261,
0.0386072136,
0.0767943263,
0.0584539324,
-0.0485235639,
0.0137577169,
-0.0219350457,
0.0004660907,
0.0448663868,
0.0702728853,
-0.0224796366,
0.0323397331,
-0.0694122761,
-0.0287838839,
-0.0198191442,
-0.0144997332,
0.008858243,
-0.0291420016,
-0.0131891351,
-0.015578513,
-0.0499147102,
0.0253785104,
0.0396273471,
0.0328125283,
-0.060831666,
-0.0083619477,
0.0266169161,
0.0287076589,
0.0098779993,
-0.0054021785,
-0.0159035958,
0.021722028,
-0.0198520739,
0.0106373588,
-0.0034244319,
-0.0373334773,
0.040069405,
-0.0040096818,
-0.0267264806,
-0.0291377157,
0.011030334,
-0.0692669228,
-0.0192377027,
-0.0246275961,
-0.0138850007,
0.0090431832,
-0.0435974374,
-0.0156309027,
0.0458367467,
0.038979996,
-0.0000702147,
0.061110083,
0.0031178216,
-0.0180783141,
-0.0491284505,
0.0020930909,
0.0144938286,
-0.0537831001,
0.0210948642,
0.0074892691,
-0.0418971553,
0.0071424767,
0.0252677426,
0.0077397721,
-0.0234191455,
-0.0199428387,
-0.0893251821,
-0.0374975353,
-0.005437443,
0.0212714486,
-0.0156015158,
-0.0093768677,
-0.010065306,
-0.0553034618,
-0.0342359506,
0.0000303544,
-0.0203703605,
-0.0055974596,
0.0709003657,
0.0332608372,
0.0011403584,
-0.0082953367,
-0.0235147681,
-0.0140148699,
0.0305333305,
0.0405880958,
-0.0476598553,
-0.0174286161,
-0.0317506753,
0.0257707871,
0.019120397,
-0.0215019397,
0.0355539843,
0.0138723645,
-0.0575533994,
-0.0060025235,
0.092326127,
0.016672859,
0.0729468018,
-0.0443923436,
-0.0150991837,
-0.0359036513,
0.0189343039,
-0.0041422267,
0.082621865,
-0.0202327128,
-0.0002290584,
0.016775256,
0.0409598611,
-0.0357744396,
0.0493565463,
-0.024274556,
0.0052614044,
0.0225810669,
-0.0132497251,
-0.0432730541,
0.0114349863,
-0.0070718718,
0.0472703464,
-0.016986303,
-0.0285144094,
0.0350546874,
-0.0229790974,
-0.0305644087,
0.0094380612,
0.0233657416,
-0.0551059209,
0.094961971,
0.0022078715,
0.0194352046,
-0.0393849015,
-0.0415034592,
-0.0090407049,
0.002061052,
0.036356084,
0.0037307804,
0.0233475864,
-0.0119163189,
0.0367189944,
-0.0387248583,
-0.0264538601,
-0.0069063618,
-0.0580410175,
0.0125207491,
-0.0094712945,
-0.0059731458,
-0.0259395167,
-0.079300262,
0.037069492,
0.0027125839,
0.0549917594,
-0.0155073814,
-0.0241887886,
-0.0004116986,
0.0593443103,
0.055979412,
-0.004020853,
0.0125231342,
-0.0710109323,
0.0241815075,
0.0298100505,
0.008509553,
-0.004950942,
-0.0283485968,
-0.0075540249,
-0.0354315639,
-0.0311649218,
0.016082814,
-0.016188316,
-0.0207300056,
-0.0796652138,
0.0209266338,
0.0529468693,
0.0061146128,
-0.0297423732,
-0.0203747991,
-0.0089683346,
-0.0405249372,
-0.0365477651,
0.0244803503,
-0.096277006,
-0.0283288881,
-0.001118072,
0.0295581985,
0.0635865778,
-0.0423712768,
-0.0031195518,
-0.0712369606,
0.0788667575,
-0.0189230442,
-0.017531164,
-0.0100746593,
-0.035439536,
-0.0181187894,
-0.0252082366,
0.0070913988,
0.0194481891,
-0.0267569013,
-0.0457376204,
0.0121353576,
-0.0026077409,
0.0113989785,
-0.0150074204,
0.0077230539,
-0.0756589994,
0.0228585266,
0.018741237,
0.0018870519,
-0.0268320628,
0.017078409,
0.0126205673,
-0.006692091,
0.0009627844,
0.0792373791,
-0.0280029383,
0.0074139037,
0.0164173525,
0.0479816832,
-0.0376667939,
0.01189938,
0.005747736,
0.0115624219,
-0.0152196493,
-0.0282540992,
-0.0475226454,
-0.0315993689,
-0.027040109,
0.0607371703,
0.0174312927,
-0.0326945931,
-0.081690751,
-0.0029063786,
0.0062419442,
-0.0690704808,
0.070686698,
-0.0103769507,
-0.0098899975,
-0.0147716356,
-0.0735856965,
-0.0749870464,
-0.0191832408,
0.0482739843,
-0.0116915386,
-0.036006812,
0.0350137949,
-0.0036199635,
-0.052043166,
0.0451088324,
-0.0347232707,
0.0160338525,
-0.0525418073,
0.0244230796,
-0.0096004345,
-0.019448746,
-0.0657816008,
0.0357996747,
-0.0671019703,
0.014915593,
0.009449535,
0.0199481901,
0.0185815115,
0.0051482585,
-0.0068629785,
-0.0008826389,
-0.0019072065,
0.0184340272,
-0.0827850252,
-0.0454137549,
0.043078728,
-0.008001592,
-0.0163831506,
0.0071955416,
-0.0552102923,
0.0463792048,
-0.0188589711,
-0.0194855295,
-0.0137646412,
0.0179735124,
-0.0184633024,
0.0256472547,
-0.013481061,
0.0205003973,
0.0051284307,
-0.0449304022,
0.0070042722,
0.002983219,
0.008882843,
-0.0137818307,
-0.0040691681,
0.0148120308,
0.0035978451,
-0.0072650583,
0.0269954242,
0.0084683904,
0.0082107596,
-0.0212090574,
0.0292866379,
0.0104706837,
0.0159990005,
-0.0172187928,
-0.0350691862,
0.016619917,
0.0330542177,
-0.0007171119,
0.0277070589,
-0.002576092,
0.0425474308,
0.0263378173,
0.0173634514,
-0.0158822816,
0.0085045435,
-0.0150633929,
-0.0070357462,
-0.0118533671,
0.0325744264,
0.0091345776,
-0.0445636138,
-0.0138667356,
-0.05445683,
0.0025496415,
-0.0038208733,
0.0183131583,
0.0538126417,
-0.0372986943,
0.0220363569,
0.0308719929,
0.0003236154,
-0.0188746788,
0.068084985,
-0.0183128268,
0.0225499533,
-0.0327035077,
0.007028156,
-0.0872323811,
-0.0253433008,
0.0600413196,
0.0276609287,
0.0048943888,
0.0062344633,
-0.0526968949,
-0.0048129624,
0.0282745995,
0.0025659841,
0.0039097685,
0.0284692943,
0.0252730474,
0.0849483013,
0.0293412264,
0.0083480142,
-0.0070386538,
0.0307454821,
-0.018977752,
-0.0112742009,
-0.0249141399,
-5.569709333e-33,
0.0103175519,
-0.0105071003,
0.0237875003,
0.0202287845,
0.0116552776,
0.0290964562,
-0.0168802906,
-0.0511808656,
0.0433465205,
0.0054846685,
-0.0133214388,
-0.0077804057,
0.036818549,
0.0282865483,
0.0047177928,
0.0178744346,
0.0294978172,
-0.0504424684,
-0.025232954,
-0.0006806767,
0.0339764506,
0.0577751845,
0.0292060394,
-0.0561049283,
0.0446927473,
0.0052057616,
-0.0586345978,
-0.0082936659,
0.0489521585,
0.0574010387,
-0.032144621,
-0.0264524035,
0.0411103703,
-0.0450017303,
-0.0103036882,
0.0766171142,
-0.1063925922,
0.0202816632,
-0.0357328467,
0.0713214278,
-0.021662645,
-0.0204100404,
0.0407859981,
-0.0344107375,
0.0036308584,
0.001977274,
0.0434083603,
-0.0174736381,
0.0126156285,
-0.0357711241,
-0.0892339647,
0.0461528115,
0.0014555029,
0.0142010413,
0.0860469341,
0.0432241857,
0.0088139763,
0.022745654,
-0.0521577299,
-0.0212969705,
-0.0346363261,
-0.0302201807,
0.0511301085,
-0.0193136502,
-0.0117174778,
0.0317804776,
0.0243145414,
-0.0020710535,
0.0013235395,
-0.0130843464,
0.0360333025,
0.0564380102,
0.0282553639,
0.0059640533,
0.053675361,
0.0050672023,
-0.0179955065,
0.0461908393,
0.0145880803,
0.062863037,
0.014255221,
0.0016135027,
-0.0122511396,
0.0366377383,
0.0115127275,
-0.0280361846,
-0.0169261899,
0.0195280444,
0.035166055,
0.0427033976,
0.0072917896,
0.0139357829,
0.0591793731,
-0.0267237481,
-0.0587215796,
0.0241327323,
0.0180472955,
0.0546167605,
0.0127958385,
-0.0498639159,
-0.0174191799,
-0.0440851562,
0.0169709362,
0.0094956569,
0.0152569087,
-0.0223418362,
0.0788982809,
-0.006184143,
-0.046186585,
-0.0049827662,
0.0060132658,
0.0169227496,
0.0144456448,
0.0551615059,
0.0313082561,
-0.0032835379,
-0.024668023,
0.0261181239,
-0.0210516043,
0.0234001391,
-0.0084065134,
-0.0445652679,
-0.0013201387,
0.0565427355,
0.0123540703,
-0.0142775495,
-0.018759273,
-0.0248507634,
0.0481488109,
-0.0313843191,
0.007119297,
0.0857546106,
2.525e-7,
0.0372346155,
0.0537972786,
0.068297334,
-0.0055254619,
0.0529720783,
0.0358322002,
-0.0154634463,
0.022998482,
-0.0820843726,
-0.0117823621,
0.0403436013,
0.0230638012,
0.0609214492,
0.0011364026,
-0.0715990886,
-0.0306215957,
-0.0508570373,
0.0007075911,
-0.0740890652,
0.0419311598,
0.0192472413,
0.0654136017,
-0.0400409959,
-0.0605803281,
0.0145624392,
-0.0684098229,
0.0550744049,
-0.0101189157,
0.0109846601,
-0.0199559573,
0.0450626612,
-0.0135026155,
-0.0189483613,
-0.0100646298,
-0.000080379,
0.014060732,
0.0301810987,
0.0631179288,
0.0015242591,
0.0224093422,
0.0228571538,
0.0131358691,
-0.01942968,
-0.0292561762,
0.0360801406,
-0.0265083443,
-0.0362032317,
-0.04800459,
-0.0567765683,
0.0032346791,
-0.0045690141,
0.0118831107,
0.024466021,
0.0431388915,
0.0433137864,
-0.0770300105,
0.0168967154,
-0.0535621867,
0.0169360097,
0.0059050536,
-0.018643178,
0.0539086871,
-0.0412912667,
0.0034802174,
0.0499680787,
-0.0585871674,
-0.0500644334,
2.048150592e-34,
0.0659611002,
-0.0380812064,
-0.0254065581,
0.0331091471,
0.0391165204,
-0.0244299956,
-0.0134511851,
0.0450906605,
-0.0145629216,
-0.0917989612,
0.0128566679
] |
http://arxiv.org/abs/2401.15043v2 | http://arxiv.org/pdf/2401.15043v2 | Health Text Simplification: An Annotated Corpus for Digestive Cancer Education and Novel Strategies for Reinforcement Learning | 2024-01-26T18:13:57 | 2024-03-29T19:01:27 | Md Mushfiqur Rahman, Mohammad Sabik Irbaz, Kai North, Michelle S. Williams, Marcos Zampieri, Kevin Lybarger | cs.CL | cs.CL, cs.AI, cs.LG | Objective: The reading level of health educational materials significantly
influences the understandability and accessibility of the information,
particularly for minoritized populations. Many patient educational resources
surpass the reading level and complexity of widely accepted standards. There is
a critical need for high-performing text simplification models in health
information to enhance dissemination and literacy. This need is particularly
acute in cancer education, where effective prevention and screening education
can substantially reduce morbidity and mortality.
Methods: We introduce Simplified Digestive Cancer (SimpleDC), a parallel
corpus of cancer education materials tailored for health text simplification
research, comprising educational content from the American Cancer Society,
Centers for Disease Control and Prevention, and National Cancer Institute.
Utilizing SimpleDC alongside the existing Med-EASi corpus, we explore Large
Language Model (LLM)-based simplification methods, including fine-tuning,
reinforcement learning (RL), reinforcement learning with human feedback (RLHF),
domain adaptation, and prompt-based approaches. Our experimentation encompasses
Llama 2 and GPT-4. A novel RLHF reward function is introduced, featuring a
lightweight model adept at distinguishing between original and simplified
texts, thereby enhancing the model's effectiveness with unlabeled data.
Results: Fine-tuned Llama 2 models demonstrated high performance across
various metrics. Our innovative RLHF reward function surpassed existing RL text
simplification reward functions in effectiveness. The results underscore that
RL/RLHF can augment fine-tuning, facilitating model training on unlabeled text
and improving performance. | Health Text Simplification: An Annotated Corpus for Digestive Cancer Education and Novel Strategies for Reinforcement Learning: Objective: The reading level of health educational materials significantly
influences the understandability and accessibility of the information,
particularly for minoritized populations. Many patient educational resources
surpass the reading level and complexity of widely accepted standards. There is
a critical need for high-performing text simplification models in health
information to enhance dissemination and literacy. This need is particularly
acute in cancer education, where effective prevention and screening education
can substantially reduce morbidity and mortality.
Methods: We introduce Simplified Digestive Cancer (SimpleDC), a parallel
corpus of cancer education materials tailored for health text simplification
research, comprising educational content from the American Cancer Society,
Centers for Disease Control and Prevention, and National Cancer Institute.
Utilizing SimpleDC alongside the existing Med-EASi corpus, we explore Large
Language Model (LLM)-based simplification methods, including fine-tuning,
reinforcement learning (RL), reinforcement learning with human feedback (RLHF),
domain adaptation, and prompt-based approaches. Our experimentation encompasses
Llama 2 and GPT-4. A novel RLHF reward function is introduced, featuring a
lightweight model adept at distinguishing between original and simplified
texts, thereby enhancing the model's effectiveness with unlabeled data.
Results: Fine-tuned Llama 2 models demonstrated high performance across
various metrics. Our innovative RLHF reward function surpassed existing RL text
simplification reward functions in effectiveness. The results underscore that
RL/RLHF can augment fine-tuning, facilitating model training on unlabeled text
and improving performance. | [
0.064312838,
0.0122889988,
-0.0099588884,
-0.0532120392,
-0.0031615004,
0.0629464239,
-0.0034095203,
0.0345498994,
-0.0012099381,
-0.0354354568,
-0.0087965187,
-0.0135378661,
-0.0056778798,
-0.0128860204,
-0.0029608258,
-0.0938217789,
0.0263950117,
0.010005841,
0.0403899178,
0.0198988076,
0.0010647239,
-0.0065595564,
-0.0281095393,
0.0332052335,
0.0372558497,
0.015991345,
0.0476453453,
0.0004563064,
0.0182037111,
-0.0879789665,
0.0386740416,
0.0330029652,
-0.0336122923,
0.0356448032,
0.0000020476,
-0.0351177603,
-0.0225304868,
0.024224164,
-0.0241138805,
0.0260107312,
0.1002911925,
-0.0464121625,
-0.0272889696,
0.0010763438,
0.0009478889,
0.0018439498,
0.0381416455,
0.0828541964,
0.0469937995,
0.0861409158,
-0.0084320698,
-0.0306094289,
-0.0143020339,
-0.0320169479,
0.026736537,
-0.0012224611,
0.0463771336,
-0.0002828309,
0.0303657707,
-0.0018480761,
0.0088551408,
0.0713581815,
-0.005045238,
-0.0079217739,
-0.0059706331,
0.0511836633,
0.0248440299,
-0.0068513067,
0.0064159464,
0.0684433207,
-0.010559272,
-0.0226459894,
0.0474786311,
-0.0304707829,
0.0026964596,
0.0365797393,
-0.0446448289,
-0.0226527657,
-0.0143203121,
0.0345489793,
-0.0181445852,
-0.0007608185,
0.0339503437,
0.007924444,
0.0267464146,
0.0196656846,
0.0063215173,
-0.0340598412,
0.0184406806,
0.0070676557,
0.0112416362,
-0.058829993,
0.0380721539,
0.0377214737,
-0.0075325179,
0.0295791496,
-0.0054848497,
-0.0688383281,
0.0798265263,
-0.0538336672,
-0.0015409187,
0.0038308532,
-0.0076993993,
0.0302428473,
-0.0125475544,
0.0013681934,
-0.0151325278,
0.0669869632,
-0.0329393074,
0.0731530935,
-0.026674645,
-0.005783522,
-0.0358195752,
0.0495375395,
-0.0093134753,
0.0147595964,
-0.0903360918,
-0.0194378588,
0.0015346031,
0.0088276723,
0.0127137052,
-0.0058837016,
-0.0575177781,
-0.0002145972,
-0.0192121863,
0.0616137162,
-0.0082132611,
0.0222236123,
0.010231561,
-0.0188175738,
0.0003426121,
-0.0070663881,
0.0054690936,
-0.007373489,
-0.0169074573,
0.0299065299,
-0.0388548225,
-0.0367345214,
-0.0185285695,
-0.0280355215,
-0.011615579,
-0.0238563269,
-0.0096230675,
-0.0638669357,
0.0163184814,
-0.0172150806,
0.0484117046,
-0.0401216485,
0.0202520825,
0.0058309794,
-0.0396227129,
0.059950158,
0.0030683498,
0.0036298591,
0.0740440264,
0.0400831178,
0.0260871723,
0.0528734103,
0.0281544738,
0.0064782556,
-0.0057648611,
-0.0046474552,
-0.0105433706,
0.0038207185,
0.0400134549,
0.0241428204,
0.0394146331,
-0.0083429618,
-0.0304925982,
0.0786626935,
-0.0142362779,
0.0592684411,
-0.0399441719,
0.0257096346,
0.0500687584,
0.0485255495,
0.0119755473,
0.0100964308,
0.0219959877,
0.0361391082,
0.0294443909,
0.0206723791,
-0.0250423141,
-0.003848884,
-0.0402774066,
0.0124791134,
0.0171720311,
-0.0135040702,
-0.0380179025,
-0.054636728,
-0.0006685359,
-0.0118841575,
0.0213624705,
-0.0373596251,
0.0532081835,
0.0120211262,
-0.0790234283,
-0.0122026671,
0.0682013184,
-0.0266757794,
0.0150283482,
-0.0454485603,
0.0365330838,
0.0546984561,
0.0619348027,
-0.0666395053,
-0.0361214317,
-0.038955465,
0.0319357365,
0.0202312991,
0.0196833834,
-0.0312805474,
-0.0411737114,
-0.0291803163,
-0.0361855738,
0.017135296,
0.0223008022,
-0.0306018889,
0.0199338831,
0.0224348512,
-0.0172135793,
0.0048946035,
-0.0507661588,
0.0544113033,
-0.0209507011,
-0.0592787936,
-0.0217332486,
0.0220808536,
0.0309901685,
0.0021540564,
0.016808534,
0.0120512573,
-0.0000254756,
-0.0124335112,
-0.0308772922,
-0.0201897547,
-0.0219722316,
-0.048989825,
0.0020168228,
-0.0546074659,
-0.004510026,
0.012777186,
-0.0501923561,
0.0015638443,
-0.0458232798,
-0.0534193963,
-0.0013020268,
0.0110604586,
0.0240018237,
0.056118574,
-0.0134730162,
-0.0407735929,
0.0125676561,
-0.0458132587,
-0.0029635571,
0.0260791387,
-0.0655638576,
-0.012916537,
-0.0455451868,
0.0256969035,
-0.038808085,
0.0054000085,
-0.0143365283,
0.0067911507,
0.0442612395,
-0.0158443041,
-0.0338732488,
-0.0607194938,
-0.0203484241,
0.0009414747,
0.0004650395,
0.0501975305,
0.0316773988,
-0.0143384188,
-0.0315109566,
-0.0427161306,
-0.0178064723,
0.0164640807,
0.0226806961,
0.0387169123,
0.0371138118,
0.0031888513,
-0.0593763255,
-0.0201186333,
-0.0156660248,
0.0005868254,
0.0511144474,
-0.0637383461,
-0.0136428345,
-0.0060064215,
0.0150437774,
-0.00639239,
0.0210503042,
-0.0281042028,
-0.0259149186,
0.0134288352,
-0.0102197742,
0.0485770702,
0.0140380356,
0.0416262113,
0.0306982864,
0.0241500847,
-0.0578219779,
0.0265489798,
-0.0713596418,
-0.0341060162,
-0.0354732461,
-0.0698459819,
0.0477004983,
0.0822788179,
0.0107158097,
0.0159182586,
-0.0650189221,
-0.0298651326,
-0.0331645012,
-0.0172888171,
-0.0881279707,
0.0195392221,
-0.0239022169,
-0.0060350434,
-0.0012222922,
-0.0356894918,
0.1153033525,
0.012487472,
0.0588953085,
-0.0959923938,
0.0042413031,
-0.0106071308,
-0.0559536852,
-0.0231145713,
0.025251301,
0.0608294383,
-0.0190443341,
-0.0239406247,
-0.0074945586,
0.0095973872,
-0.0428334773,
0.0490268953,
-0.0490691438,
-0.0011261941,
-0.0039755404,
-0.0513223372,
-0.0184559282,
-0.0332339108,
0.0030066541,
-0.0221297126,
0.0420036763,
0.0014897333,
-0.0940812454,
0.0131575046,
0.0062868944,
0.0712235272,
0.0128751062,
-0.0213050451,
-0.0101305181,
0.0088133402,
0.0680553839,
-0.0115750665,
-0.0159711018,
-0.0460221916,
0.0349898785,
0.0347492248,
-0.0164699871,
-0.0102595622,
0.062612541,
0.0093934005,
0.0146266865,
-0.0591813214,
0.0154871093,
0.0425984859,
-0.0032624153,
-0.0239904542,
-0.0113845626,
0.0743751079,
0.0247135833,
-0.0455639139,
0.0191516262,
-0.0039641694,
-0.0071193674,
-0.0510043986,
0.0377951972,
-0.0806776509,
-0.0124272453,
-0.0039965236,
0.0390730239,
-0.0205416102,
0.0006876475,
0.0177272167,
-0.0227564797,
0.0196597446,
0.020886261,
-0.0522772707,
-0.022373125,
-0.0301957186,
0.005005769,
-0.0033006102,
-0.0146925347,
0.001765357,
-0.030183455,
-0.0050792005,
0.0175563376,
-0.0483022891,
-0.0973441899,
-0.0358752571,
0.0106257703,
0.0062660864,
0.0536724068,
0.021396542,
0.0025860323,
-0.0215563346,
0.0049607563,
0.0214991048,
-0.0165596735,
-0.0184858274,
0.0828334317,
0.0146883866,
0.0279249307,
-0.0451873057,
0.0618791208,
0.0068751262,
-0.0098614469,
-0.0084964018,
0.0764666498,
0.0254934505,
-0.0418871306,
-0.0398892947,
0.0252885204,
0.0385035165,
0.0600003265,
0.0198439416,
0.0074831028,
-0.0283330977,
-0.0120576527,
0.0122153498,
-0.0019750686,
-0.0039219414,
0.0297704879,
-0.0629578233,
-0.0286126994,
-0.0714668036,
0.0045906808,
0.0281804241,
-0.0317835659,
0.0318653956,
0.0318350382,
0.014649827,
-0.0090821423,
-0.0371444076,
0.0262600239,
-0.020734394,
0.0063087186,
-0.0904208571,
-0.0687275603,
0.0033758834,
-0.0262947679,
-0.0077036615,
0.0388329998,
-0.0629864186,
-0.0207691174,
-0.0248242207,
0.04593179,
0.0182798486,
0.0279673114,
-0.0019548079,
-0.0642272532,
0.0945289135,
0.0391084664,
-0.0433677994,
0.0221524145,
-0.0183256119,
0.0122711957,
0.0041445727,
0.0032384335,
-0.0347156487,
0.0134958029,
-0.0127931312,
0.0301840995,
-0.0203702915,
-0.0008761191,
0.0306912437,
0.0306317601,
0.0277391113,
-0.0098715564,
-0.0383483022,
-0.0389609225,
0.0237303246,
-0.0108798528,
0.0037275534,
-0.0007402524,
-0.0014566014,
0.0588385798,
0.0139057245,
-0.0405482426,
0.0348154344,
-0.0305806119,
0.003104269,
-0.0495058112,
0.0257268269,
0.0033886353,
0.0280487165,
0.0452833958,
-0.0173545685,
0.0055627343,
-0.0078931991,
-0.0121555412,
-0.008874936,
-0.0003788765,
0.0484392941,
0.03604801,
0.0302872248,
-0.0062651308,
0.0286945291,
-0.0268378071,
-0.0047752736,
-0.053661745,
0.001436324,
0.0113156382,
0.0052742246,
0.0557640977,
-0.0364071093,
-0.0292809289,
-0.0267105233,
0.022105623,
-0.0242324341,
0.0224289522,
0.0728111342,
-0.0086023966,
0.0221363157,
-0.0222753454,
0.0773862377,
0.0014818358,
0.0587558933,
-0.0071082348,
-0.0234495439,
-0.0471050739,
-0.0318691693,
0.027163282,
0.0563715436,
-0.0518563874,
-0.0337705351,
-0.0657332763,
-0.0167470556,
0.0284487195,
0.0069481065,
-0.0092104906,
-0.0353660509,
0.064235054,
0.0022043737,
-0.012287817,
-0.0672694147,
-0.0467384346,
0.0298471171,
-0.0017308205,
-0.0456215367,
-0.0036186101,
-6.192009063e-33,
-0.0291148648,
0.000465168,
0.0196164418,
-0.0097300038,
-0.0098771881,
0.0064047505,
-0.0617981739,
-0.0232943706,
0.0509006679,
0.0695414469,
-0.0013106419,
-0.0050433148,
0.0057665352,
0.0150202131,
-0.0379768424,
-0.0192596726,
0.0774174854,
-0.0253410693,
-0.0400520451,
0.0315497518,
-0.0211671107,
0.0434987806,
-0.0022620417,
-0.0832366943,
0.0163294114,
0.0324860103,
-0.0280484986,
-0.0233513918,
-0.0137523096,
0.0488293394,
-0.0508570522,
0.0282637924,
0.0198232979,
-0.0312155318,
-0.0112280613,
0.0452656783,
-0.0930353627,
-0.0346839726,
0.0245301295,
-0.0393915512,
-0.0607391819,
-0.0539294891,
0.0742217004,
-0.0035119397,
-0.0363618694,
0.0031258839,
0.0478567183,
-0.0344096869,
-0.0181790721,
-0.0191760082,
-0.0567074567,
0.029716827,
-0.0030024927,
0.006347937,
0.0076438938,
0.0137539627,
-0.0365791954,
0.0555870458,
-0.0001157697,
0.0313781351,
0.0912069678,
0.0465969294,
0.034243159,
-0.018977765,
-0.0031922697,
0.0320730396,
-0.036978215,
-0.005828484,
-0.0325262211,
0.0042559984,
-0.0135881025,
-0.0174881089,
0.0340160467,
0.0355123915,
0.0154566811,
-0.0181407519,
-0.0822091028,
0.0663800687,
-0.0628815964,
0.0238723457,
0.006970102,
-0.0190260466,
-0.0338364914,
-0.010933416,
-0.0404391699,
0.0351561792,
0.0154022947,
-0.0334142298,
0.0186969861,
0.023236258,
-0.0014580168,
0.0094037578,
-0.0477616787,
0.0095078321,
-0.0213352051,
0.0072566024,
0.0476098172,
0.019461384,
-0.0344199426,
-0.0556003638,
-0.066875115,
0.0072722398,
0.0244379677,
-0.0161827393,
0.0268746242,
0.0109024802,
0.0569905452,
-0.0081653809,
-0.0640546829,
-0.0010955768,
0.0185952652,
0.0180418305,
0.0362308659,
-0.0210488606,
0.0121122152,
0.0107203098,
0.0117571522,
0.05951811,
0.006799391,
-0.0527243726,
0.0056778621,
-0.0045551788,
0.0172170345,
0.0192081481,
-0.0422862321,
0.0079305572,
-0.0149543704,
0.0542529039,
0.0521402173,
-0.0847695917,
0.0097895656,
0.0103410967,
2.836e-7,
0.0854920074,
0.0639350787,
0.0421766043,
-0.064744398,
0.0161697846,
0.0037506132,
0.0148868384,
-0.0160660949,
-0.0201078001,
0.0402579196,
0.0662249252,
-0.0006167707,
0.0061324001,
-0.0037143291,
-0.066591993,
-0.0184961911,
-0.0734914467,
-0.0082028378,
-0.0285283532,
-0.0370084755,
0.0512928329,
0.024000125,
0.002575303,
-0.0215809494,
0.0075393347,
0.0207185,
0.0374769829,
-0.0713015273,
0.0268974509,
0.0489349812,
0.0055339765,
-0.0166705139,
0.0149394134,
-0.0213637445,
-0.0248349588,
-0.0275049899,
0.0129499417,
0.1019870117,
0.0015629568,
0.0459643565,
0.0244610123,
-0.0099134753,
0.0207713377,
-0.032789059,
0.0134374704,
-0.0907120556,
-0.0071027847,
0.0460048094,
0.0033276596,
-0.0217717644,
0.0287574138,
-0.0286904629,
-0.0575230233,
0.0338290744,
0.0112733999,
0.0492468551,
-0.0027800191,
-0.0717919245,
-0.0021180355,
0.0169889219,
-0.0203040391,
-0.0103299255,
-0.0254736468,
0.0163804945,
-0.0455031693,
-0.094422698,
0.0138799809,
2.582423336e-34,
0.0377273709,
-0.0255798083,
-0.0061758514,
-0.0198233705,
0.0103009716,
0.0097093927,
-0.0106499977,
0.0392107181,
0.040450722,
-0.0636762977,
0.0112062134
] |
http://arxiv.org/abs/2401.15042v4 | http://arxiv.org/pdf/2401.15042v4 | PROXYQA: An Alternative Framework for Evaluating Long-Form Text Generation with Large Language Models | 2024-01-26T18:12:25 | 2024-06-04T12:46:47 | Haochen Tan, Zhijiang Guo, Zhan Shi, Lu Xu, Zhili Liu, Yunlong Feng, Xiaoguang Li, Yasheng Wang, Lifeng Shang, Qun Liu, Linqi Song | cs.CL | cs.CL, cs.AI | Large Language Models (LLMs) have succeeded remarkably in understanding
long-form contents. However, exploring their capability for generating
long-form contents, such as reports and articles, has been relatively
unexplored and inadequately assessed by existing benchmarks. The prevalent
evaluation methods, which predominantly rely on crowdsourcing, are recognized
for their labor-intensive nature and lack of efficiency, whereas automated
metrics, such as the ROUGE score, demonstrate discordance with human judgment
criteria. In this paper, we propose ProxyQA, an innovative framework dedicated
to assessing long-text generation. ProxyQA comprises in-depth human-curated
meta-questions spanning various domains, each accompanied by specific
proxy-questions with pre-annotated answers. LLMs are tasked to generate
extensive content in response to these meta-questions, by engaging an evaluator
and incorporating the generated texts as contextual background, ProxyQA
assesses the generated content's quality through the evaluator's accuracy in
addressing the proxy-questions. We examine multiple LLMs, emphasizing ProxyQA's
demanding nature as a high-quality assessment tool. Human evaluation
demonstrates that the proxy-question method is notably self-consistent and
aligns closely with human evaluative standards. The dataset and leaderboard is
available at \url{https://proxy-qa.com}. | PROXYQA: An Alternative Framework for Evaluating Long-Form Text Generation with Large Language Models: Large Language Models (LLMs) have succeeded remarkably in understanding
long-form contents. However, exploring their capability for generating
long-form contents, such as reports and articles, has been relatively
unexplored and inadequately assessed by existing benchmarks. The prevalent
evaluation methods, which predominantly rely on crowdsourcing, are recognized
for their labor-intensive nature and lack of efficiency, whereas automated
metrics, such as the ROUGE score, demonstrate discordance with human judgment
criteria. In this paper, we propose ProxyQA, an innovative framework dedicated
to assessing long-text generation. ProxyQA comprises in-depth human-curated
meta-questions spanning various domains, each accompanied by specific
proxy-questions with pre-annotated answers. LLMs are tasked to generate
extensive content in response to these meta-questions, by engaging an evaluator
and incorporating the generated texts as contextual background, ProxyQA
assesses the generated content's quality through the evaluator's accuracy in
addressing the proxy-questions. We examine multiple LLMs, emphasizing ProxyQA's
demanding nature as a high-quality assessment tool. Human evaluation
demonstrates that the proxy-question method is notably self-consistent and
aligns closely with human evaluative standards. The dataset and leaderboard is
available at \url{https://proxy-qa.com}. | [
0.0363538973,
0.0429144837,
0.0048075635,
0.0076990495,
-0.0462793894,
-0.0094009601,
0.0300348811,
-0.0237920154,
-0.0178865306,
-0.0637242794,
-0.0276896823,
-0.055595167,
0.0206738133,
0.0292558894,
0.039664682,
-0.042865146,
-0.0183472764,
-0.0207255632,
-0.0136289597,
0.0015087725,
0.0276398826,
-0.0010882856,
0.0152076399,
0.0351708941,
0.0208104439,
-0.0108325165,
-0.024104761,
-0.0038828393,
0.0200876761,
-0.033521045,
0.0077123474,
0.0815954506,
-0.0131983627,
0.0263580978,
0.0000018878,
-0.0586436428,
0.0121867722,
0.0292465519,
0.0186790545,
0.0296192821,
0.0423888043,
0.0076483679,
0.0069132652,
0.0175475739,
-0.0227297489,
-0.0093825627,
0.009587219,
-0.0047136326,
0.0886125341,
0.0505901761,
0.0086974548,
-0.0147905722,
0.0104161697,
-0.0360445492,
0.0136125442,
-0.0198688004,
-0.0086703096,
0.0161327403,
0.0474661961,
0.0198724382,
0.0174858533,
0.0045066318,
0.0428649373,
-0.0700005144,
0.0215415135,
-0.0020015952,
-0.0312336329,
-0.0259057973,
0.0212519113,
0.0754596293,
0.050260365,
-0.0153344842,
0.0153796896,
0.0007989754,
0.0134937139,
0.0175635386,
-0.0426163785,
0.0588201806,
0.0143573396,
-0.0233861953,
0.0131144552,
0.0347410738,
0.0192899536,
0.0247926675,
0.0158340987,
0.0178563092,
0.0166535228,
-0.013426126,
-0.0062344093,
0.0175367668,
-0.0165781993,
-0.0999336168,
0.0049982984,
0.0444555953,
-0.0592053346,
0.0001432757,
-0.0408446155,
-0.0616079532,
0.0538838133,
-0.021591628,
0.0255682301,
0.0084449854,
0.0083918292,
0.0183739718,
-0.0165911149,
0.0526720099,
-0.0020992476,
-0.0310542528,
-0.0499600545,
0.0138743976,
-0.0416378342,
-0.0180614442,
-0.028529536,
0.0167278778,
-0.0326969624,
-0.0047494732,
-0.0441853255,
-0.0052283062,
-0.0205098335,
0.0153970597,
0.0190813169,
0.0158295669,
-0.0131340344,
-0.0197935458,
-0.027720267,
0.0202348027,
-0.059232004,
0.0020133688,
0.0607739538,
-0.0051759691,
0.0249042753,
0.0346371345,
-0.0071166987,
-0.0231928639,
-0.0190970264,
0.0511465929,
-0.0040765791,
0.0070901755,
-0.0187959895,
-0.0581751727,
-0.0051131858,
-0.0676988512,
0.0066665392,
-0.0883312002,
-0.0005121013,
0.009673995,
-0.0234200135,
-0.0122452416,
0.0214576013,
0.0516848825,
0.0201519486,
-0.0095348479,
-0.0224952977,
0.010529072,
0.0543001145,
0.0104746325,
0.012735053,
0.0616024397,
0.0564342253,
0.0497451052,
0.0194451474,
0.0173074137,
0.0324433818,
0.0169531424,
0.0270066075,
0.020840453,
-0.0067869755,
0.027869748,
0.0024395448,
0.0534971431,
0.0148742627,
0.0947603732,
-0.0064664413,
0.0311390385,
0.053265281,
-0.0008721041,
0.001527347,
0.0059734094,
0.0860123485,
0.0130467545,
0.0270997509,
0.0336521342,
0.0073377071,
0.0421296097,
-0.0142219681,
0.0115181981,
0.0580132641,
0.0292265452,
0.0012421882,
-0.0475628898,
-0.0124176443,
0.0136658819,
-0.0016492428,
-0.0049338154,
0.0082328534,
0.0245140791,
-0.0321476273,
0.0048789033,
-0.0002607362,
-0.064916797,
0.0591717139,
-0.0379959084,
0.0670925155,
0.0568303578,
0.0523308627,
0.0103833051,
-0.0011779757,
-0.0282515325,
0.0331356041,
0.0684546232,
0.0558985807,
-0.0337626711,
-0.0434394851,
-0.0114586428,
-0.0116105089,
-0.0303141456,
0.0073807617,
-0.0208623614,
0.016427286,
0.0094919996,
0.065498665,
-0.0579020083,
-0.0775074363,
0.0363881178,
-0.0597098134,
-0.0075842598,
0.0075287144,
0.0026880149,
-0.007689618,
0.0271920338,
-0.0185711831,
0.0568129458,
-0.0092085525,
-0.0431577228,
-0.0224448722,
-0.0093543306,
-0.0256486665,
-0.00092761,
0.0213725977,
-0.072519578,
0.0064242324,
0.0243567787,
-0.013412443,
-0.0422633961,
-0.0213191211,
-0.0465209372,
-0.0109849451,
-0.0357505381,
-0.0013482898,
0.0185436066,
-0.0075875395,
-0.0350453444,
0.0652914196,
0.0065897973,
0.0374431238,
0.0548261367,
0.0079274764,
0.0268017445,
-0.0254672058,
-0.020417463,
0.0155352112,
0.0002081507,
0.0274735503,
0.0412598588,
0.0310912002,
-0.0637548491,
-0.0435995646,
-0.0886437967,
0.0230178703,
-0.0379754603,
0.0158974417,
0.0542370826,
-0.0142478906,
0.0015338274,
0.0536615923,
-0.0960439593,
-0.0342904776,
0.0254086573,
0.0205628499,
0.0427368879,
0.0087064551,
0.0484026,
-0.036584042,
-0.0090652779,
0.0143974414,
0.0517167039,
0.087412551,
0.0466568023,
-0.0513314866,
0.0355190597,
-0.0359060131,
-0.0695275217,
0.0150998943,
-0.0601558723,
0.0064786565,
0.0081624854,
0.0459078997,
0.0350205638,
0.0626663491,
0.0033463023,
0.0363850854,
0.0209853239,
0.0127288476,
-0.0212488715,
-0.0088935932,
-0.0029970552,
0.0037336026,
-0.0535169989,
0.0271691307,
0.0350328945,
-0.0042581554,
-0.009058373,
-0.0393979028,
0.0314355418,
-0.0694112927,
-0.0430884473,
-0.0421878397,
-0.0153715862,
-0.0108257802,
0.0565126091,
-0.0039832233,
-0.0216721203,
0.04978697,
-0.036779765,
0.0199631471,
-0.0344846174,
0.0228841603,
-0.0624067895,
0.0589844175,
-0.00890618,
0.0256881062,
-0.0152358944,
0.0285354555,
-0.0627148896,
0.018190261,
0.032174997,
-0.0468458682,
-0.0047051227,
-0.0440998673,
-0.0058953329,
-0.0210116226,
-0.0230274647,
-0.022542648,
-0.0423047356,
-0.0049396092,
0.0144329537,
0.0452849716,
-0.0147491908,
0.0078531206,
0.0162532832,
0.047957506,
0.026120672,
-0.0517389923,
0.0003322451,
0.0066486835,
0.0761380196,
0.0705887303,
0.0270632412,
0.0412518904,
-0.0771159083,
0.0350067131,
0.0344591737,
0.0027677875,
0.0424963199,
0.01858286,
-0.0148110222,
-0.0067479438,
-0.0271976609,
-0.0261130929,
0.019935444,
0.0612393692,
-0.0690112039,
-0.0245618392,
0.1002302095,
0.0390265398,
-0.0195734538,
-0.0004466788,
-0.0238522757,
0.0768165067,
0.0044768704,
0.0294644311,
-0.0692079663,
0.0038693054,
0.0070724827,
0.0428901538,
-0.0181692243,
-0.0355486237,
-0.0111351032,
-0.1202863976,
0.0440937094,
0.0031858287,
-0.0451271795,
0.003731889,
-0.0220981073,
-0.0110023897,
0.0117348749,
-0.0159198344,
0.0344798751,
-0.0237825084,
0.0271806773,
0.0226957612,
-0.0333755873,
-0.0473850332,
0.0212134961,
-0.0315543711,
0.0309645236,
0.0704283193,
0.0253132693,
-0.0052027907,
-0.0150719425,
0.0311168917,
-0.0576758347,
-0.0135098463,
0.0400241017,
0.0456682034,
0.005908431,
-0.0039204531,
-0.0372567661,
0.0403408706,
-0.018427521,
-0.0054575512,
0.0294348635,
0.0537509583,
0.0159596559,
-0.0092777107,
-0.0145923607,
0.0101311831,
0.0125234183,
0.0886725485,
0.0407595113,
-0.0879752934,
-0.0375901125,
-0.0184890721,
-0.0101531632,
-0.080005914,
-0.0381667092,
0.0421580262,
0.0102408994,
-0.0196326729,
-0.1261018366,
-0.0898278654,
0.0365738124,
0.0226680003,
0.0046579437,
-0.0057418328,
-0.0073401928,
-0.0135625638,
-0.0346104279,
0.0140079604,
-0.0155958822,
0.0270109884,
-0.0773431808,
-0.0086595323,
-0.0013006973,
-0.0350516811,
-0.0164594185,
0.0302534644,
-0.1011191905,
-0.0347504616,
-0.0025739965,
0.0159438793,
0.0174043924,
-0.0179117341,
-0.0064938362,
-0.0442408957,
0.0472412482,
0.0313414671,
-0.0293246824,
0.0209654439,
-0.0304515101,
0.0141397901,
-0.0104725361,
0.0231159348,
-0.0483711511,
-0.019921165,
-0.0238184147,
0.0512063801,
0.0391127355,
-0.0347148851,
0.0368082076,
0.0429598652,
-0.045046594,
-0.0480900742,
0.0185667537,
-0.0434810258,
-0.0249982551,
-0.0556045398,
-0.046803318,
-0.0523829646,
0.0138266105,
0.0033037893,
-0.0053740945,
0.0156974029,
0.0365852676,
-0.0134143624,
0.0300893541,
-0.0338715389,
0.0245691501,
-0.0206538066,
0.0242656991,
0.0223497227,
-0.0154842995,
-0.0415781811,
-0.04455138,
-0.0487826318,
0.0447510183,
-0.0111123947,
0.0429101959,
0.0299156923,
0.0031263912,
0.036913652,
0.0394056179,
-0.0526480712,
-0.0208885223,
-0.0416420512,
0.0166441258,
0.0260820016,
0.0161764175,
0.0405209363,
-0.0196921621,
-0.0107701402,
-0.0287718158,
0.045101326,
0.0147124929,
-0.0232424978,
-0.0070081614,
0.013902883,
-0.0320195891,
-0.0031478042,
0.0911520645,
0.0194669217,
-0.01033457,
0.0171821807,
0.0173174944,
-0.0395092145,
-0.0480867624,
-0.0332196951,
-0.0094228294,
0.0013848752,
-0.0203818269,
-0.0863935277,
-0.0557021163,
-0.0031754433,
0.0430645123,
-0.0016463255,
0.0010456454,
0.0327828638,
0.0442260429,
-0.0288036186,
0.0177188292,
-0.0980728492,
0.0332372859,
-0.0398564488,
0.001784864,
0.0177534763,
-5.974157634e-33,
-0.0254723132,
-0.0769103765,
0.0329820439,
-0.0453242138,
-0.0540002957,
0.0114518683,
-0.0603036992,
0.0017672209,
0.0399125963,
0.0411546938,
0.0017773641,
-0.0313416459,
-0.007073014,
0.0391902998,
-0.0075751408,
-0.0276944973,
0.0355925374,
-0.007778042,
-0.0646583885,
0.0438658856,
0.0173035059,
-0.0165529307,
-0.0263266545,
-0.0118347276,
0.0525906384,
-0.04664151,
-0.0189964026,
-0.0135066267,
0.0079427715,
0.0577182323,
-0.0113945697,
0.0005057821,
0.0371284671,
-0.0235919356,
-0.0092467209,
0.0316571929,
-0.0344606042,
0.0030230137,
-0.0005721422,
-0.043300949,
-0.0768791288,
-0.0440590121,
0.0504348874,
-0.0086300354,
0.0101969372,
-0.0157526881,
0.0393390022,
-0.0364219248,
0.0212967284,
-0.044141911,
-0.053203363,
0.0370472781,
-0.0093454951,
-0.0032617797,
-0.0543563403,
0.0765324309,
-0.0114618409,
0.0454213843,
-0.022571899,
0.0147667071,
-0.0221610256,
0.043414481,
0.060631521,
0.006547132,
0.0055662678,
0.0495781265,
-0.0188561715,
0.0356143937,
-0.0090602674,
0.021500295,
-0.0098775132,
0.008130555,
0.0542383045,
0.1029056311,
0.0016598486,
-0.0118959416,
-0.0609284267,
0.0619445816,
-0.0396431163,
0.0001608029,
0.0111988932,
-0.0656814948,
0.0074497634,
-0.0068726344,
-0.0701781437,
-0.0145994639,
-0.0185549799,
-0.0224704724,
0.039407637,
-0.0106934914,
-0.0337236635,
-0.0498741828,
0.0231712237,
-0.0108433953,
-0.0034564924,
0.0153127452,
-0.0570857823,
0.0454091989,
0.000539835,
0.037175186,
-0.0534672625,
0.0397812724,
0.0146576008,
0.0414220579,
0.018810017,
0.012817869,
0.0205999129,
0.0139109055,
-0.0464868285,
0.0145187834,
-0.0536154434,
0.0442676917,
-0.0137301218,
-0.0330157913,
0.0048010801,
-0.0076598627,
-0.0122590568,
0.0363456234,
0.0022746697,
-0.0056781764,
-0.0120098954,
-0.0270824078,
0.0238745306,
0.023002468,
-0.0466292873,
0.0227459166,
-0.0244098101,
0.038398236,
0.0447609536,
-0.0357371159,
0.0038688665,
-0.0095163565,
2.471e-7,
0.0345838033,
0.0316786356,
-0.0023638315,
0.0289911944,
0.0073040524,
0.0104094939,
-0.0145034576,
0.0222001113,
0.002625121,
-0.0167527832,
0.0029882186,
0.0065342649,
0.0576757975,
0.0467674211,
-0.0841357633,
0.0287051462,
-0.0881470516,
-0.0022239592,
-0.0356454775,
-0.0287960079,
0.0583268553,
0.0203199927,
-0.0272298101,
-0.014922319,
-0.0039232438,
-0.0396481156,
0.0803516582,
-0.0020971799,
-0.0525308959,
0.0374522358,
0.0278276429,
-0.0029838169,
0.0108310245,
-0.0278088991,
-0.0233521834,
-0.0129543431,
0.0334984027,
0.0470265187,
0.0527813621,
-0.0201367028,
0.0132494923,
-0.036320325,
0.0182314869,
-0.0618988201,
0.066546306,
-0.0201181229,
-0.0201618932,
-0.0351546705,
-0.0308612622,
-0.0216644183,
0.0146075245,
-0.0233007409,
-0.0124512864,
0.0431173369,
-0.0105636874,
0.0044759,
0.0151088824,
-0.0039127241,
-0.0540913492,
-0.0148502188,
-0.0328578241,
-0.0646671429,
-0.0488991067,
-0.0097786421,
-0.0268878005,
-0.0549078286,
0.0166445319,
2.212740234e-34,
0.0498776063,
-0.0333514735,
-0.0087411338,
-0.0203304254,
0.0535680093,
0.0039528529,
-0.02852102,
-0.0013398714,
0.011249762,
-0.0551869981,
-0.0073328465
] |
http://arxiv.org/abs/2402.01719v3 | http://arxiv.org/pdf/2402.01719v3 | Measuring Moral Inconsistencies in Large Language Models | 2024-01-26T18:05:47 | 2024-03-01T06:35:29 | Vamshi Krishna Bonagiri, Sreeram Vennam, Manas Gaur, Ponnurangam Kumaraguru | cs.CL | cs.CL, cs.LG | A Large Language Model (LLM) is considered consistent if semantically
equivalent prompts produce semantically equivalent responses. Despite recent
advancements showcasing the impressive capabilities of LLMs in conversational
systems, we show that even state-of-the-art LLMs are highly inconsistent in
their generations, questioning their reliability. Prior research has tried to
measure this with task-specific accuracy. However, this approach is unsuitable
for moral scenarios, such as the trolley problem, with no "correct" answer. To
address this issue, we propose a novel information-theoretic measure called
Semantic Graph Entropy (SGE) to measure the consistency of an LLM in moral
scenarios. We leverage "Rules of Thumb" (RoTs) to explain a model's
decision-making strategies and further enhance our metric. Compared to existing
consistency metrics, SGE correlates better with human judgments across five
LLMs. In the future, we aim to investigate the root causes of LLM
inconsistencies and propose improvements. | Measuring Moral Inconsistencies in Large Language Models: A Large Language Model (LLM) is considered consistent if semantically
equivalent prompts produce semantically equivalent responses. Despite recent
advancements showcasing the impressive capabilities of LLMs in conversational
systems, we show that even state-of-the-art LLMs are highly inconsistent in
their generations, questioning their reliability. Prior research has tried to
measure this with task-specific accuracy. However, this approach is unsuitable
for moral scenarios, such as the trolley problem, with no "correct" answer. To
address this issue, we propose a novel information-theoretic measure called
Semantic Graph Entropy (SGE) to measure the consistency of an LLM in moral
scenarios. We leverage "Rules of Thumb" (RoTs) to explain a model's
decision-making strategies and further enhance our metric. Compared to existing
consistency metrics, SGE correlates better with human judgments across five
LLMs. In the future, we aim to investigate the root causes of LLM
inconsistencies and propose improvements. | [
0.0635798872,
0.0573870428,
-0.030418044,
0.0582219511,
-0.0970877782,
0.0010990325,
-0.0147810262,
0.0171606559,
-0.0024964861,
-0.0484731831,
-0.011888979,
-0.0387784317,
0.0248203352,
-0.0288334973,
-0.0131515842,
-0.0062965881,
0.0365270041,
-0.0244101305,
0.0373190865,
-0.0251249187,
0.0107698971,
0.0133945011,
0.0455887318,
0.0189137384,
0.0009447823,
0.0085479338,
-0.0100620473,
0.0182824638,
0.0176902283,
-0.0281801149,
-0.0238696057,
0.0393341146,
-0.0021542145,
0.0596204586,
0.0000018912,
-0.0725650713,
0.0217480529,
-0.0102076288,
-0.0127675217,
0.0328061357,
0.0664865598,
0.0152644422,
-0.0096096033,
-0.0027895554,
-0.0283778068,
0.0256404206,
0.0445849411,
0.0515685901,
0.0142699545,
0.0501644649,
0.0032850332,
0.0143336905,
0.0377025045,
-0.0047531286,
0.0089388276,
-0.0218070671,
0.024491407,
0.0528198853,
-0.023688484,
-0.030461194,
-0.0254243538,
0.026407335,
0.0408200994,
-0.0127564957,
-0.0425693654,
-0.007353141,
0.0353454202,
-0.0176321138,
-0.0071376683,
0.0642545149,
0.0112278396,
-0.022228837,
0.0065542571,
-0.009599559,
0.0097679645,
0.0026953158,
0.0005538684,
0.0303986296,
-0.0150586748,
0.0102331163,
0.0031956721,
0.0077377423,
0.0131605417,
0.0354296789,
0.042663496,
0.0613030232,
-0.0065575656,
-0.0488635562,
0.0007272611,
-0.0167069342,
-0.0603991002,
-0.034973491,
-0.0292093828,
0.0163103752,
-0.0608492196,
0.0366935767,
-0.0054469057,
0.0289162118,
0.0731933862,
-0.0507952012,
0.0236233938,
0.0189507939,
0.0270332005,
0.0288367439,
0.05709498,
0.0641777739,
0.0004589295,
0.0605556965,
-0.0573760681,
0.0195770878,
0.0015029002,
0.0043409113,
0.0108897798,
0.0117341736,
0.0165435467,
0.0047974517,
-0.0316250175,
0.0312018357,
-0.056434717,
-0.0062145512,
-0.0102795716,
0.007016662,
0.01185383,
-0.0006548554,
-0.0622106008,
0.0134670399,
-0.0235947575,
0.0385119691,
-0.0059747184,
-0.0320464373,
0.0495268106,
-0.0287212096,
-0.0268320311,
0.0252634194,
0.0114315655,
0.0898955092,
0.0059010787,
-0.0150325792,
-0.0136919301,
-0.0276833996,
-0.0715642646,
-0.0775845051,
-0.0571505018,
-0.0621533841,
0.031585712,
-0.0055774245,
-0.0272797346,
-0.0505118929,
0.0214580912,
0.0114038223,
-0.0508691669,
0.032561481,
0.0037688883,
0.0305266343,
-0.0110465949,
0.0163156707,
0.0019376971,
0.044709947,
0.0271590669,
0.007991625,
0.0852594599,
-0.0001946057,
-0.0201209616,
-0.0060426681,
-0.0227854177,
0.0267073344,
-0.0087148426,
0.0101238871,
-0.0025275622,
0.0198284425,
0.0115141235,
0.0786632895,
-0.0258036442,
-0.0729738325,
0.0629761368,
0.0827295557,
-0.0140731391,
-0.0310317129,
0.007551888,
0.00677153,
0.0138880853,
0.0407331735,
-0.0649782717,
-0.0076388065,
-0.0064856801,
0.0010226095,
0.0117273359,
0.0247626174,
-0.0018349178,
-0.0378939137,
-0.0013558005,
0.0046437276,
0.0430194251,
0.0038045833,
-0.0078155184,
-0.0202774182,
-0.0111965397,
-0.0409877561,
0.0656181127,
-0.0499183759,
-0.0181129389,
-0.027712319,
0.0705931708,
0.0875360519,
0.0626725703,
0.0273260456,
-0.0092441756,
-0.0164879821,
-0.0220862515,
0.0250511169,
0.0275842045,
-0.0067399386,
0.0131212743,
-0.0400669239,
-0.0131449364,
0.0204932578,
0.0019138866,
-0.0016195229,
-0.0216168687,
0.0043631005,
0.0214540586,
-0.0364452749,
-0.0105020031,
0.0250820518,
0.0130884741,
0.0524016768,
-0.0398681946,
-0.0131881041,
0.0504587218,
-0.0184144825,
-0.0297161955,
-0.0206252448,
0.0067166118,
0.0271746535,
0.0077426671,
0.0049866531,
-0.0595077127,
0.012567441,
0.0477252193,
-0.0492828004,
-0.0247866753,
0.0283099525,
-0.0084753996,
0.0126931975,
-0.025103569,
-0.0674868152,
0.04288432,
-0.058528997,
0.035413973,
-0.0679748356,
-0.027667582,
-0.0402523987,
0.0269270614,
-0.0263088681,
0.0582903326,
-0.024396034,
-0.0183404107,
0.0325068682,
-0.0403822251,
-0.0109808277,
0.0431095921,
-0.0235003531,
0.0253837984,
0.0412029698,
0.0408600345,
-0.0298369657,
-0.0483673662,
-0.0930188373,
-0.025127003,
0.0020308716,
0.0036392605,
-0.0013588035,
-0.0061407844,
-0.0062168315,
0.0085415971,
-0.0384655483,
0.0609722845,
0.0310529489,
0.0077081462,
0.0123824291,
-0.0338732004,
-0.0528207608,
-0.0258108135,
-0.0096550481,
-0.0221180581,
0.0272705108,
0.0088181244,
0.0283974819,
-0.0451223627,
0.0610111691,
-0.0886350796,
0.0094084358,
-0.0025545901,
-0.055229228,
0.0108621838,
0.0577864945,
0.0189708564,
0.0498217978,
0.0744301975,
0.0007602617,
0.0384402387,
-0.0126591986,
-0.0744398087,
-0.027521126,
-0.0366203785,
-0.010199286,
-0.0489187911,
-0.1013384387,
-0.0107032517,
0.040908061,
0.0266052764,
0.0160928685,
-0.0332058519,
-0.0280922614,
-0.042056635,
-0.0660062879,
-0.030545719,
-0.0365434997,
-0.0072174263,
0.0274455231,
-0.0162660927,
-0.0633934736,
0.0947424471,
-0.0135600213,
0.0443983153,
-0.0012733936,
-0.0072068609,
-0.0166404638,
0.0676830783,
0.0268438514,
0.0569753237,
0.0279462915,
-0.011923627,
-0.0369002372,
-0.0321101174,
0.0475058407,
-0.0488878116,
0.0647275448,
-0.0262147877,
0.0123505406,
0.0226806644,
-0.0700794384,
-0.0223359298,
-0.0832797289,
-0.0224409476,
-0.0411517248,
0.0204943456,
-0.0292716715,
0.0551968701,
-0.0020215458,
0.0154496431,
-0.0282710977,
0.0038845255,
-0.0062514809,
-0.0216028485,
-0.0296165515,
0.0657661259,
0.0196980312,
0.0849559307,
0.0078983409,
0.0444081835,
0.0671197996,
0.0257022809,
0.038974423,
0.0139756436,
0.0190810151,
-0.0057307538,
0.0206288733,
0.0147988452,
0.0482377708,
0.0448242053,
-0.0089956867,
0.0092577683,
0.0925604701,
0.0155963665,
-0.0432176292,
0.0343512222,
-0.0275377706,
0.0133810928,
-0.0015357877,
-0.025215026,
-0.0804966092,
-0.0371166728,
0.0055084978,
0.006105124,
-0.0848994479,
-0.0615308024,
0.0140103195,
-0.0872524828,
-0.0146227283,
0.0148218721,
-0.012944269,
-0.0312586986,
-0.0308485888,
0.0088880565,
0.0060586599,
0.0748969167,
0.0098364437,
-0.015566946,
-0.0329740196,
-0.0187523644,
0.0096884668,
0.0303371046,
0.056495849,
-0.0144720497,
0.0865227804,
0.0544857606,
0.0360231474,
-0.0189449526,
-0.0531082191,
0.0216567703,
-0.0011668745,
-0.0393494628,
0.0121342642,
0.0110900942,
-0.0090490272,
0.0050697033,
-0.0719954818,
-0.0128738927,
0.0176239908,
-0.014682062,
-0.0115753105,
0.034352079,
-0.0041071032,
-0.05134397,
0.0173351504,
0.0182924736,
0.0030271176,
0.0873081684,
0.0176093467,
-0.0203935802,
-0.0169633012,
-0.0440053083,
-0.0132384142,
0.0014264757,
-0.0356818251,
-0.0069862027,
0.0200042538,
-0.0310353749,
-0.0457654297,
-0.0441629216,
-0.062894851,
-0.017802462,
0.0243764818,
-0.0458448753,
0.045838695,
-0.0126279462,
-0.0470522866,
-0.0299059767,
-0.0455227196,
0.0238846838,
-0.0356972776,
0.0754291192,
-0.010893777,
-0.0003618277,
-0.0057845861,
-0.0063389069,
-0.0221517999,
-0.0346718691,
0.0323730186,
0.0163027029,
0.0180304125,
0.0110474741,
-0.0373347141,
-0.0050632861,
0.0136938477,
0.0151473749,
-0.0707300156,
0.065253213,
0.0392066985,
0.0053174137,
0.0157074295,
0.0538076311,
-0.032253813,
0.0050879545,
0.0310991462,
0.0193936434,
-0.0232449286,
-0.0492905416,
0.02150736,
0.0560381971,
0.0016011981,
0.0270114318,
-0.0160889756,
0.0027870394,
-0.0140283713,
0.0038057363,
-0.0289354902,
-0.0349156298,
-0.0146868089,
0.0056927861,
-0.0011403671,
-0.0198857728,
0.0245338734,
-0.0278916191,
0.0232511535,
0.0154323867,
-0.0484663323,
0.0172458421,
0.0317587033,
-0.0046281493,
-0.0089898948,
-0.0423657633,
0.0307310764,
0.0000558153,
0.0377399959,
0.0256974716,
0.0836210996,
0.0307305101,
-0.0511365607,
-0.011618129,
0.0268440228,
-0.0743248612,
-0.0309738349,
0.0003013154,
0.0276949033,
0.0074012838,
0.0669475943,
0.0439853445,
-0.0261187945,
0.0144440373,
-0.0351690426,
0.005600384,
0.0136721348,
-0.0338277258,
0.0244514588,
0.0122383926,
-0.0263244305,
0.0417878479,
0.0358249582,
-0.0360384285,
0.0267843027,
0.003626155,
-0.0317562297,
-0.0006404454,
-0.0328598954,
0.0082367007,
-0.0099781314,
-0.0043464489,
-0.0308621079,
-0.1034792885,
-0.050244879,
-0.0493860282,
0.0359560028,
0.0482038632,
-0.0138466079,
-0.006914414,
0.0729267299,
-0.0024270788,
0.0033117377,
-0.0175564718,
-0.0030573888,
-0.0468765236,
0.0183138102,
-0.0565442517,
-5.664965518e-33,
-0.0019563066,
-0.0475046858,
-0.0463780276,
0.0090554757,
-0.0243338104,
-0.0260230303,
-0.0085897204,
0.0229186434,
-0.0435602963,
0.0480112657,
0.0232756734,
0.0245197825,
0.0018281909,
-0.0163700208,
-0.0103503391,
-0.0664876401,
-0.0025116077,
0.0307929069,
-0.0439708121,
0.0133885415,
0.037712723,
-0.0617408343,
0.0195505116,
-0.0673630387,
0.0650173873,
0.0039179591,
-0.0224591792,
0.0244505443,
-0.0208221152,
0.0268722326,
0.00611676,
0.1150317714,
0.0244925749,
-0.0333827063,
0.0122116674,
0.0396504812,
0.0166160017,
-0.0296679474,
0.0786714628,
-0.0052504344,
-0.044543121,
-0.0347179137,
0.011277901,
-0.0209853295,
-0.0755041167,
0.0051668989,
0.0167525057,
0.0242752079,
0.0249498151,
-0.01075067,
-0.0391739644,
-0.0158187989,
-0.0337816253,
-0.0482425652,
-0.0442013405,
0.0351909325,
-0.0140905166,
-0.0096772304,
0.0062944824,
0.0313895084,
-0.0254253652,
-0.0195168909,
0.0214264877,
-0.0343772359,
0.0116437487,
0.0937654898,
-0.0037417377,
0.0078561604,
-0.0224727932,
-0.014173869,
-0.0114561766,
0.0361277461,
0.0574291795,
0.0542317145,
0.0041092271,
-0.0463951789,
-0.055737488,
-0.0054712733,
0.0439187773,
0.0290389918,
0.0076777032,
-0.0121835498,
-0.0711664557,
-0.0248359796,
-0.0502938405,
-0.0488174893,
-0.0299849138,
-0.0385090597,
0.032394778,
-0.0314716585,
-0.007371841,
-0.0288391169,
0.0142259253,
-0.051278986,
-0.0033747582,
0.0377657786,
0.0049728206,
0.0203416552,
0.0074905674,
0.0441949964,
-0.0222878195,
-0.0106329415,
0.0226087905,
0.0001985081,
0.0169712435,
-0.0107637979,
0.0328574739,
-0.021047838,
-0.0164107569,
0.0158284027,
0.003032903,
-0.0032873198,
-0.019169189,
-0.0707786307,
0.035729222,
0.0074973074,
-0.0277659688,
-0.007954496,
0.0270870309,
-0.0005120084,
-0.0549949519,
0.0025411986,
-0.026510654,
0.04218724,
-0.0109007526,
-0.009326553,
-0.0451662652,
-0.0122402795,
0.0140248286,
0.006364855,
-0.0103467964,
0.006377826,
2.598e-7,
-0.0110034822,
0.0165475961,
-0.0072252196,
-0.0048033902,
-0.001125631,
0.078268826,
0.0318915844,
0.0390452109,
0.0451936871,
-0.0889396667,
-0.0208473671,
0.0206999201,
0.0501107872,
0.0286702942,
-0.0077409185,
0.0037005814,
-0.0590587333,
-0.0299863126,
-0.0094969729,
0.0531910993,
0.0359032713,
0.0590642877,
0.0457934029,
-0.0037740725,
0.0165264755,
-0.0210693125,
0.0086151799,
-0.0924976096,
-0.0267004836,
0.0120762717,
-0.0069325529,
-0.0642684251,
0.0458536968,
0.024251882,
0.050713893,
-0.0108825956,
-0.0162942614,
0.0321726017,
-0.0060918038,
-0.0276213847,
-0.0678001344,
-0.0552029349,
0.0019521386,
-0.0386273563,
0.0493392646,
-0.0058414727,
-0.0195168164,
-0.0260804556,
0.0426143259,
-0.0256803688,
0.037829563,
0.0029606186,
-0.0422090627,
0.0093866484,
0.0242971312,
0.0016426007,
0.0025856551,
-0.026014071,
-0.0100020235,
-0.0039097425,
0.0162098855,
-0.1155053228,
0.0030083673,
0.0427324474,
0.0455544256,
-0.0961551145,
0.0277632978,
2.237967445e-34,
0.0342557952,
-0.0196642149,
0.0502592959,
-0.005005809,
0.0446902141,
0.0398557819,
0.0038997359,
-0.0011434174,
0.0286996644,
-0.0486799367,
0.0138411885
] |
http://arxiv.org/abs/2401.15024v2 | http://arxiv.org/pdf/2401.15024v2 | SliceGPT: Compress Large Language Models by Deleting Rows and Columns | 2024-01-26T17:35:45 | 2024-02-09T17:59:40 | Saleh Ashkboos, Maximilian L. Croci, Marcelo Gennari do Nascimento, Torsten Hoefler, James Hensman | cs.LG | cs.LG, cs.CL | Large language models have become the cornerstone of natural language
processing, but their use comes with substantial costs in terms of compute and
memory resources. Sparsification provides a solution to alleviate these
resource constraints, and recent works have shown that trained models can be
sparsified post-hoc. Existing sparsification techniques face challenges as they
need additional data structures and offer constrained speedup with current
hardware. In this paper we present SliceGPT, a new post-training sparsification
scheme which replaces each weight matrix with a smaller (dense) matrix,
reducing the embedding dimension of the network. Through extensive
experimentation, we show that SliceGPT can remove up to 25% of the model
parameters (including embeddings) for LLAMA2-70B, OPT 66B and Phi-2 models
while maintaining 99%, 99% and 90% zero-shot task performance of the dense
model respectively. Our sliced models run on fewer GPUs and run faster without
any additional code optimization: on 24GB consumer GPUs we reduce the total
compute for inference on LLAMA2-70B to 64% of that of the dense model; on 40GB
A100 GPUs we reduce it to 66%. We offer a new insight, computational invariance
in transformer networks, which enables SliceGPT and we hope it will inspire and
enable future avenues to reduce memory and computation demands for pre-trained
models. Code is available at:
https://github.com/microsoft/TransformerCompression | SliceGPT: Compress Large Language Models by Deleting Rows and Columns: Large language models have become the cornerstone of natural language
processing, but their use comes with substantial costs in terms of compute and
memory resources. Sparsification provides a solution to alleviate these
resource constraints, and recent works have shown that trained models can be
sparsified post-hoc. Existing sparsification techniques face challenges as they
need additional data structures and offer constrained speedup with current
hardware. In this paper we present SliceGPT, a new post-training sparsification
scheme which replaces each weight matrix with a smaller (dense) matrix,
reducing the embedding dimension of the network. Through extensive
experimentation, we show that SliceGPT can remove up to 25% of the model
parameters (including embeddings) for LLAMA2-70B, OPT 66B and Phi-2 models
while maintaining 99%, 99% and 90% zero-shot task performance of the dense
model respectively. Our sliced models run on fewer GPUs and run faster without
any additional code optimization: on 24GB consumer GPUs we reduce the total
compute for inference on LLAMA2-70B to 64% of that of the dense model; on 40GB
A100 GPUs we reduce it to 66%. We offer a new insight, computational invariance
in transformer networks, which enables SliceGPT and we hope it will inspire and
enable future avenues to reduce memory and computation demands for pre-trained
models. Code is available at:
https://github.com/microsoft/TransformerCompression | [
0.0051798825,
0.0947202519,
-0.0066811405,
0.0700120255,
-0.011526485,
0.004983685,
0.0041050282,
0.0190677196,
-0.0244577583,
0.0027248675,
-0.0630505979,
0.0060674641,
0.0205036476,
0.0326196216,
0.0059928959,
-0.0324559994,
0.0368576348,
-0.0104498677,
-0.0287358798,
-0.0013056891,
0.0082935989,
0.0001771079,
0.026987182,
0.0195455439,
0.0139296139,
-0.0161917079,
-0.0215784237,
-0.0198238455,
-0.0174953509,
0.0058934991,
0.0083106626,
0.0407467894,
0.0299060885,
0.0471166521,
0.0000021955,
0.0203964598,
0.0074644042,
-0.0105947116,
-0.0441103131,
0.0550676212,
0.0437912419,
-0.0161051955,
0.0217443891,
0.0010564666,
-0.0188709293,
-0.0752043501,
0.0327012241,
0.0436353385,
0.0454518124,
0.0570103414,
-0.014884105,
-0.050238885,
0.0255006123,
-0.0674020723,
-0.0181400459,
-0.02543783,
-0.0067563057,
-0.0549222492,
-0.0457147993,
0.0166343395,
-0.0185143631,
0.0379804261,
0.0435292386,
0.0349941589,
-0.037185546,
0.0472114533,
-0.049351722,
-0.0430010594,
0.0236032698,
0.0436818749,
-0.0685077235,
-0.0405438393,
-0.0048003229,
0.004489576,
0.0067873173,
-0.0201240368,
-0.0421538167,
0.0265915077,
0.0468550734,
0.0421625488,
-0.0078094276,
0.0002209628,
0.0511010848,
0.0046993736,
-0.03271956,
0.0367857963,
0.0259715915,
0.0254444499,
0.0000295759,
-0.0026521489,
-0.0291827619,
-0.0402081013,
-0.0064193928,
0.004227093,
-0.018174598,
-0.0084311236,
-0.0494455844,
-0.031281516,
-0.0035673082,
-0.0216639843,
0.0594137087,
0.0568604879,
0.0190093573,
0.0122804977,
-0.0351926312,
0.027783392,
-0.0333721302,
0.016599955,
-0.0646310523,
-0.0165129006,
0.0153246038,
-0.0319949314,
-0.0526031591,
0.0373565257,
-0.0495614447,
-0.0331465751,
-0.0338588469,
-0.0235537849,
-0.0236704834,
-0.0256494749,
0.0225838553,
0.0539251715,
-0.0289703924,
0.0005404543,
0.0188487321,
0.0620465018,
-0.0219167694,
0.0256558899,
-0.0029964815,
-0.0593728684,
-0.0469194911,
0.0335340388,
0.014129837,
-0.0273089651,
0.0510611311,
0.0246413928,
-0.0236406662,
-0.0746259838,
-0.0300086662,
0.0204261672,
0.0090342453,
-0.0697123483,
0.0301986597,
-0.0644103065,
0.0596877635,
-0.0387293771,
0.0201136451,
-0.0512835979,
0.0083984938,
0.0465875342,
-0.0594405606,
0.0686647296,
-0.0096910251,
-0.0154775409,
0.0083039533,
0.0200310778,
0.0537905879,
0.0465767644,
0.0389207527,
0.0508370958,
0.0049407342,
-0.0346651599,
-0.0287845135,
0.0516027771,
0.0404330865,
-0.0175403375,
-0.0472173169,
-0.0347780734,
-0.0236526504,
0.0058797817,
0.0696330518,
0.0618205406,
-0.0484380238,
-0.0503233895,
0.0478972606,
0.0797826648,
0.000026379,
0.0559684634,
0.025071064,
-0.015035606,
0.0100915153,
-0.0029420874,
-0.0359387361,
0.0263190754,
0.0212111138,
-0.0187382214,
0.0151881464,
0.0655420646,
-0.0513663925,
-0.0194980856,
0.0058658049,
-0.0435611121,
0.0239969846,
0.001712987,
0.0231465474,
0.0281398222,
-0.0306746885,
0.1131852567,
0.0348427631,
-0.0473543815,
-0.050292898,
-0.0513432175,
0.007283261,
-0.0119985659,
0.0176878273,
0.0218295697,
-0.0398071222,
0.0157547388,
-0.003554241,
0.007108253,
-0.000176735,
-0.0042200373,
0.0322889686,
-0.075684078,
0.023242183,
-0.0009718936,
0.0367073938,
-0.0003761506,
0.0722627491,
-0.0037926408,
0.0187563878,
-0.0115840519,
0.005263716,
0.0244875513,
-0.0456787124,
-0.0180777106,
0.0267267749,
-0.0251137502,
0.0329237804,
-0.0236159991,
-0.004068749,
0.0002916118,
0.0690796226,
-0.0504169017,
0.0100540677,
-0.0158846676,
-0.0661085472,
-0.0189218,
0.0342218578,
-0.0045152074,
0.0350864157,
0.0823455751,
-0.0506404117,
-0.0220084172,
-0.0008641754,
-0.0378373563,
0.0167258512,
-0.0407235026,
0.003161886,
0.0028305799,
0.0762725025,
-0.002930525,
0.027185034,
-0.0520509332,
0.0309788529,
0.006025027,
-0.0228631794,
0.0151330428,
-0.0293638818,
-0.0445069186,
0.0093605081,
-0.0237957966,
-0.0046148957,
-0.0241609607,
-0.0008316856,
0.0137738353,
-0.0143753095,
-0.103335917,
0.0393974036,
-0.0211938843,
0.0017373041,
0.0297822468,
0.0328739956,
0.01952371,
-0.0057196203,
-0.0053120186,
0.0268042237,
0.0184794106,
0.0249282066,
-0.002278944,
-0.0246059485,
-0.0006433824,
-0.0353686213,
-0.0544207841,
0.0334261917,
0.0764411241,
-0.0444323979,
-0.0195576064,
-0.0164005309,
0.0364809409,
-0.0041892487,
0.0377781764,
-0.0261873491,
-0.0535134561,
-0.0127294017,
0.0012315351,
0.0394282155,
0.0681925863,
-0.0021908435,
0.0516590327,
0.0606567971,
0.0156385135,
-0.0151755512,
0.0064142514,
-0.0238786619,
0.0069358828,
-0.0208847038,
-0.0247293767,
-0.0126000093,
0.098474957,
-0.0176366512,
-0.0067997989,
-0.0376239419,
-0.0127101103,
0.0003320435,
-0.0270417891,
0.0007772368,
0.0023307672,
-0.0305258092,
-0.0116405478,
-0.0258770324,
-0.0531003624,
0.0028906967,
-0.0572000593,
0.0062794015,
-0.0070140953,
-0.0339435786,
-0.0224297289,
0.0423856936,
0.057728257,
0.0102132214,
-0.0383296087,
0.0131700318,
-0.0367463566,
0.0066308179,
-0.0177277438,
-0.0218563806,
0.044840958,
-0.0242311321,
0.0048857536,
-0.0431030206,
-0.0330436304,
-0.0008260475,
-0.0033848723,
-0.0210970845,
0.0385030173,
0.0432437435,
-0.0416889861,
-0.0430577807,
0.0055530909,
-0.0032040561,
0.089806512,
-0.0042239847,
0.0297856983,
-0.0115734544,
0.0416481644,
0.0369496308,
0.034204565,
0.0176965632,
-0.0019116356,
-0.0070438404,
0.069008328,
-0.0041853087,
0.0089043183,
-0.0676756501,
0.0011645528,
0.0322898515,
0.0004442312,
-0.0328010097,
-0.011243728,
0.0261860825,
-0.0333252996,
-0.0135244429,
0.0790633038,
0.0217885245,
-0.0618377961,
0.0078827888,
0.0219509024,
0.0541447476,
0.0387526117,
0.0215764381,
-0.0633606687,
-0.0331487246,
0.0281161573,
-0.0317086466,
0.0286891349,
-0.0317327157,
-0.0111050671,
-0.0998062342,
0.0474662073,
-0.0186523534,
0.0011511168,
-0.0232359245,
-0.0395686254,
-0.0177048743,
-0.023327902,
-0.0075856764,
-0.0106191495,
-0.0258401223,
-0.0037600473,
0.0254219864,
-0.0068553057,
-0.0167522132,
0.0148471128,
0.0036196643,
0.019246174,
-0.0063419216,
0.0330360122,
-0.0073178071,
-0.0571120158,
0.0402041934,
-0.0010995627,
-0.023448173,
-0.0178271197,
0.0722520351,
0.0149231991,
-0.0100885862,
-0.0097303456,
0.0225165021,
-0.0330363587,
-0.0533785969,
-0.0450981744,
0.0901055038,
0.0036729043,
-0.0419448242,
0.0120208021,
-0.0201865863,
-0.0622252971,
0.0689343587,
0.0328555778,
-0.0108255269,
-0.0294287875,
0.005908024,
-0.0200654417,
-0.0159863755,
-0.0565558448,
0.0180730335,
0.0245052166,
-0.056566678,
-0.0167025365,
0.000497723,
0.0283455048,
0.0197705571,
-0.0284081064,
-0.0287264902,
-0.0181663521,
-0.0022805827,
-0.065778479,
0.0435141698,
-0.0493481867,
0.0030538894,
-0.0518287346,
0.0325084627,
0.0396851674,
0.0138903121,
-0.0329868309,
0.0178297553,
-0.1296559125,
-0.0366839804,
0.0176249668,
0.0396444276,
0.0271199346,
0.0168985333,
0.0329244435,
-0.0372205451,
0.0455679446,
0.0549160019,
-0.0508834906,
0.0725191683,
0.0181110986,
-0.0718811825,
0.0061030267,
0.0420495532,
0.0177641977,
0.0193673987,
0.0126792761,
0.0101768821,
-0.0135136368,
0.00531084,
-0.0087707099,
0.0044195098,
0.0476831868,
0.0201815292,
0.0154204955,
-0.0399407633,
0.0359125845,
-0.0097142551,
0.0150361666,
-0.0539530925,
-0.0044117332,
0.0383266173,
0.0302086603,
-0.0233787857,
0.0203958675,
0.0465268791,
0.0023794889,
-0.0051640407,
-0.0141002042,
0.0341951884,
-0.0433960855,
0.0241314452,
0.0247528795,
-0.0203843694,
-0.0190873705,
-0.0208789315,
0.0165415891,
-0.0504138693,
0.0032390316,
-0.000653861,
0.0017669664,
-0.0759786069,
0.056088239,
-0.0267869402,
-0.0698439404,
-0.0244361032,
0.0090413233,
0.0255162772,
-0.0358696952,
0.0060405,
0.0000669736,
-0.0222042091,
-0.0086254943,
0.0102917273,
-0.0037687519,
0.0009968367,
0.0638227537,
0.010313103,
-0.0202343166,
-0.0091772024,
0.086261712,
0.0427715629,
0.0527762994,
0.0086232694,
-0.007030102,
-0.0402000695,
-0.0550032221,
-0.0005073113,
0.0084640514,
-0.0142551167,
-0.0204206854,
-0.0039847698,
-0.0888504237,
0.0084768562,
-0.0091092093,
0.0010774174,
-0.0214537568,
0.0281666424,
0.0342323743,
0.0173034985,
-0.0018870918,
0.0013209961,
0.0007232627,
0.0181855522,
-0.0380618945,
-0.0040664678,
-6.917198711e-33,
-0.0034376825,
-0.0065351506,
0.0214832518,
-0.0178329889,
-0.0207321886,
0.0068833553,
-0.0273733027,
-0.0087612662,
-0.0346202813,
0.0278074723,
-0.042315647,
0.0294379797,
0.0119865611,
0.0215698108,
-0.0148661807,
-0.0077135428,
0.0549562238,
-0.018717844,
-0.0097193168,
0.0172508508,
0.0015839098,
0.0291760638,
0.0440980718,
0.0017075281,
-0.0565246902,
0.0043201991,
0.0001720747,
0.0151625527,
-0.0117508713,
0.0160557479,
-0.067439124,
0.0217188708,
0.0404145718,
-0.0178718455,
0.0052416823,
0.0151161058,
-0.0951650888,
-0.0213638823,
0.0407807082,
0.0024121949,
-0.0081231464,
-0.0366809145,
0.0493564792,
0.0307105575,
0.0060485308,
-0.0325612128,
-0.0216898005,
-0.0016341095,
-0.0278627947,
0.0107911481,
-0.0332448781,
0.0047599724,
0.0178983398,
-0.0057789036,
0.0496353842,
0.0060077896,
-0.0210070536,
0.0515656844,
-0.0646637902,
-0.0356030576,
0.0103174774,
0.0790414438,
0.0477655008,
0.012769741,
0.0040361509,
0.044842869,
0.000615543,
-0.0367039591,
-0.0401434116,
0.0158179328,
0.0539813377,
0.035059426,
0.0578801408,
-0.0111118853,
0.0815806314,
-0.0261560809,
-0.089078024,
-0.0145695349,
-0.0077946191,
0.0740212202,
0.0194871835,
0.0240052249,
0.0221731346,
-0.0250488464,
-0.0165284891,
-0.0379423983,
-0.0391259938,
-0.0125318263,
0.0146685969,
0.0390706062,
-0.0259633046,
0.0176492836,
0.0302864034,
-0.0382460169,
0.0217334088,
0.039642375,
-0.0016269492,
0.0208720621,
0.0241506267,
0.0168913957,
-0.0393742025,
-0.0184018053,
-0.0070411665,
-0.0114439847,
0.0320680812,
-0.0058734664,
0.0546212941,
0.0434945934,
-0.0869508013,
-0.0135713378,
-0.0435107499,
0.0201600175,
0.0131446365,
0.119656153,
0.0230626594,
0.0358517244,
-0.0021964272,
0.0274892077,
0.0042908303,
0.0121147232,
-0.0323390998,
0.0002576186,
0.036022421,
0.0232007559,
-0.0243481919,
-0.0076836264,
-0.0654336363,
0.0186226442,
0.0326532498,
-0.0882019103,
-0.0190845691,
0.025035955,
2.874e-7,
0.0312318411,
0.0060960455,
-0.0036025918,
-0.0076237507,
0.0155489631,
0.0111217843,
0.0196593422,
0.0161789488,
-0.0123745827,
0.0013222048,
0.0009016133,
0.0242550261,
0.0315476693,
0.0178850573,
-0.0898907557,
0.0016062539,
-0.1050224528,
-0.0291170105,
-0.0210332349,
-0.0536409505,
-0.0187436324,
0.1568336487,
0.0103591299,
-0.0179683473,
0.0077619427,
0.02232554,
0.0441644043,
-0.0283082891,
0.0075358525,
0.007449856,
0.0312540978,
0.0099685956,
-0.0351061374,
-0.0322392322,
-0.0206629019,
0.0290488154,
0.0007309392,
0.0980414301,
-0.005565126,
0.0194935873,
0.018942032,
-0.0231621172,
-0.0345929712,
-0.0540043712,
0.0703249723,
-0.0504176207,
-0.0398692638,
-0.0533305816,
-0.0049562911,
0.0144211352,
0.0210035592,
-0.0096641285,
0.0459726974,
0.0310917459,
0.020351354,
-0.0300844051,
-0.0248712394,
-0.0858059525,
-0.020957591,
-0.0124592539,
-0.0705417171,
-0.0179551654,
-0.0717647895,
-0.0412358269,
0.0531534553,
-0.0260628685,
0.0028785584,
2.987750548e-34,
0.0082676681,
0.0111636007,
0.0197343659,
0.0438524634,
0.0121990694,
-0.0422309563,
0.0334767662,
0.0101314904,
-0.0347770043,
-0.1002238095,
-0.0227644797
] |
http://arxiv.org/abs/2401.15006v2 | http://arxiv.org/pdf/2401.15006v2 | Airavata: Introducing Hindi Instruction-tuned LLM | 2024-01-26T17:07:08 | 2024-02-26T12:17:25 | Jay Gala, Thanmay Jayakumar, Jaavid Aktar Husain, Aswanth Kumar M, Mohammed Safi Ur Rahman Khan, Diptesh Kanojia, Ratish Puduppully, Mitesh M. Khapra, Raj Dabre, Rudra Murthy, Anoop Kunchukuttan | cs.CL | cs.CL, cs.AI | We announce the initial release of "Airavata," an instruction-tuned LLM for
Hindi. Airavata was created by fine-tuning OpenHathi with diverse,
instruction-tuning Hindi datasets to make it better suited for assistive tasks.
Along with the model, we also share the IndicInstruct dataset, which is a
collection of diverse instruction-tuning datasets to enable further research
for Indic LLMs. Additionally, we present evaluation benchmarks and a framework
for assessing LLM performance across tasks in Hindi. Currently, Airavata
supports Hindi, but we plan to expand this to all 22 scheduled Indic languages.
You can access all artifacts at https://ai4bharat.github.io/airavata. | Airavata: Introducing Hindi Instruction-tuned LLM: We announce the initial release of "Airavata," an instruction-tuned LLM for
Hindi. Airavata was created by fine-tuning OpenHathi with diverse,
instruction-tuning Hindi datasets to make it better suited for assistive tasks.
Along with the model, we also share the IndicInstruct dataset, which is a
collection of diverse instruction-tuning datasets to enable further research
for Indic LLMs. Additionally, we present evaluation benchmarks and a framework
for assessing LLM performance across tasks in Hindi. Currently, Airavata
supports Hindi, but we plan to expand this to all 22 scheduled Indic languages.
You can access all artifacts at https://ai4bharat.github.io/airavata. | [
-0.0124408072,
-0.0126128998,
-0.0159755349,
0.0450622141,
-0.0254419222,
0.0201237146,
-0.0247200225,
0.0198118333,
0.0241194777,
0.0003079397,
-0.0923324451,
-0.0157764237,
0.0039016758,
-0.0004007757,
0.0115757762,
-0.0545999333,
0.0506214797,
-0.0479412749,
-0.0199118126,
0.0145205138,
0.0149555812,
0.0283522531,
-0.0847905427,
0.040930558,
-0.0035049412,
-0.0114752324,
-0.0044428236,
0.0441112444,
0.0144604025,
-0.0575490594,
0.0027967726,
0.0309482031,
0.0279324949,
-0.0429932177,
0.0000018035,
-0.0491529554,
0.0118483854,
-0.0363180451,
-0.0233828593,
0.0014993559,
0.0470074005,
-0.0133911539,
-0.0218066331,
0.0397154391,
0.0067328196,
0.0430787653,
0.0293518975,
0.0145015139,
0.0293100346,
0.1379288435,
0.0059375865,
0.0222097747,
0.0220978968,
-0.0077930582,
0.0132354451,
0.0054661189,
-0.0265205037,
-0.0242040362,
0.0262474697,
-0.0141943684,
-0.0267001558,
0.0207914412,
0.0262318347,
-0.0059264814,
0.0048857615,
0.0039384095,
-0.0507333577,
-0.0168168638,
-0.0161245782,
0.0467880182,
0.0074741822,
0.0088083707,
0.0049157161,
-0.0112282801,
-0.020158343,
-0.0006794048,
-0.0410023406,
0.0448766015,
-0.0004080718,
0.028741831,
0.05624808,
-0.0043655606,
0.0121524027,
-0.0058426308,
0.0789263099,
0.0430181101,
0.0098769153,
-0.0339053348,
0.0082069421,
0.004569219,
0.0033124615,
-0.0200117659,
0.0342132077,
0.068160899,
-0.0343561359,
0.0227388199,
-0.0677971616,
0.0069888663,
-0.0003902312,
-0.0309843402,
0.1055013612,
0.0834914222,
-0.0421322212,
0.0019064856,
-0.0135751599,
0.0176906809,
-0.0097197173,
0.0315577984,
-0.0174035113,
0.03447292,
0.0253607985,
-0.0055903965,
-0.0205630604,
0.0297007132,
-0.0726206303,
0.0361817181,
-0.0679879561,
0.0469176993,
-0.0091772396,
-0.0028306027,
0.0567080714,
-0.0420817547,
-0.0481830575,
-0.0128791248,
-0.0009773496,
-0.0150859952,
-0.0314745903,
0.043816179,
0.0057944045,
-0.0789457485,
0.0128555018,
0.036291942,
-0.033648517,
-0.0061907042,
0.0129640428,
0.0605575554,
-0.0195117835,
-0.0275850557,
-0.0769340619,
0.0256974902,
0.0165404137,
-0.0624671616,
-0.0030117268,
-0.0616804175,
0.001190431,
-0.0125563638,
0.0013291874,
-0.0404633023,
-0.0074073928,
0.0161438771,
0.0335012339,
0.0610188805,
0.0013249015,
0.0017657331,
0.0290561747,
0.0483812578,
0.0035703073,
0.0701724812,
0.0231034439,
0.0098055173,
0.0477253683,
-0.0099445358,
0.0250080749,
0.0799123272,
-0.0364663526,
0.009901627,
-0.0258128103,
-0.0079865223,
-0.0230462868,
0.0236907732,
0.015504933,
0.0086209122,
0.0002862811,
0.0503401048,
0.0327306651,
0.0712935925,
0.0426540747,
0.000730234,
0.025665978,
0.0009899765,
0.0283666756,
0.0098883221,
-0.0543910973,
0.044688642,
-0.0215789154,
-0.0063810959,
0.0038903458,
0.0262996089,
-0.0255995989,
-0.0264012255,
-0.0148122869,
-0.0332784541,
-0.0577883944,
-0.0117389131,
0.0252751634,
-0.0322017521,
-0.0412551612,
0.0384254828,
0.0547235794,
-0.0629601255,
0.0052973181,
-0.0430897139,
0.0057891938,
0.0480247512,
-0.0012989915,
-0.0569249317,
-0.0471730344,
0.0284649096,
0.017962845,
0.0360053778,
0.0200240649,
-0.0034541076,
-0.0070539718,
-0.0631973967,
0.0038332306,
0.0230179466,
-0.0042120847,
0.017752381,
0.0069387932,
0.0352791734,
-0.01049776,
0.0190792456,
-0.0011008243,
0.0318982489,
-0.048303578,
-0.0214309376,
-0.003552702,
0.0186589677,
0.0484735183,
0.0356551073,
-0.0118678017,
-0.0339303799,
0.0517620519,
-0.011162281,
0.0370087661,
0.016986357,
-0.0256557316,
-0.0344699882,
0.0094265817,
-0.0218194891,
-0.0013185783,
0.0147472862,
-0.0045791287,
-0.0749578476,
-0.0171024501,
-0.0111421859,
0.0264174119,
-0.0121191107,
0.0482898206,
0.0164739657,
-0.028107211,
-0.0109250657,
0.0140442569,
-0.0349056646,
-0.0109503828,
0.0446475148,
0.0635145307,
0.0126469145,
-0.0463339053,
-0.0142155485,
0.0374455638,
0.0029744094,
0.0232557822,
-0.0203123596,
0.0459486879,
-0.0570641644,
-0.0515687615,
-0.0958413109,
0.0299475938,
-0.0283566192,
0.0253853202,
-0.00213086,
0.0084740845,
-0.036847733,
-0.0165199656,
-0.0185540374,
0.003298918,
0.0366789736,
-0.0299320221,
-0.0079307407,
-0.0360531136,
-0.0389671922,
-0.0477228463,
-0.0215740297,
0.0138792088,
-0.0498802587,
-0.0157425273,
-0.0644115955,
-0.0773678124,
0.0120542832,
-0.0082062026,
0.0440249071,
-0.0238734484,
-0.0506081991,
0.0086870901,
-0.0071683107,
0.019466985,
0.0664394721,
-0.0297364183,
0.0158248339,
0.0164494999,
-0.0191824585,
-0.0485076569,
0.0131998369,
0.0196494367,
-0.0175965,
-0.0248232242,
-0.0140967332,
-0.0174954608,
0.0417154394,
-0.0078910412,
0.0306743849,
-0.0427024029,
-0.0657605901,
-0.0381838605,
-0.0122608105,
0.0286481436,
0.0528987534,
-0.0701315999,
0.0184430927,
-0.0163259096,
-0.0818778053,
0.0275467504,
-0.0418571085,
-0.0034103321,
-0.0488389879,
-0.0317801237,
-0.0320957527,
-0.002059612,
-0.0513319336,
-0.0102754943,
-0.0290369969,
-0.0388552211,
-0.0442916043,
0.053327132,
-0.0153312832,
-0.029232638,
0.0556172356,
-0.0941161215,
0.0119524784,
-0.0236659572,
-0.0210291948,
-0.0169310793,
0.0295780543,
0.0127513586,
-0.0073664756,
0.0034836996,
-0.0415763296,
-0.0087356567,
-0.0124047259,
0.0057052099,
0.0187580548,
-0.019348124,
0.0318640843,
-0.0235403739,
-0.0352948941,
0.0421222225,
0.0401889905,
-0.0236668177,
-0.0099489382,
0.0055164001,
0.0003600066,
0.0612808987,
-0.0441856533,
0.0046581342,
0.0616340004,
0.0820392147,
-0.0435126759,
-0.0318241939,
0.055041898,
0.0949759111,
-0.0214544777,
0.0016373323,
0.0723613724,
0.0254809782,
-0.0628057495,
-0.012703198,
-0.0422657169,
0.0046147979,
0.0124909077,
0.0205833875,
-0.0881402865,
-0.0335782766,
0.0102466485,
0.0461571328,
0.0278296154,
0.0031160312,
-0.012886567,
-0.0976409987,
0.0237902459,
0.0144618237,
-0.000969963,
-0.0238874685,
0.0054446966,
-0.0243830532,
0.0087633012,
0.0425341502,
0.0265512262,
-0.0634260997,
-0.0193010904,
0.0521728806,
-0.0072797923,
-0.0459752232,
-0.0064284164,
-0.0232012589,
0.0927391723,
0.0415450484,
-0.037259426,
-0.0086898161,
-0.0324181914,
0.0252902713,
-0.0082258023,
0.0072887763,
-0.0294162836,
0.0104482779,
-0.0196574479,
-0.0168461483,
0.0050952476,
0.0507640801,
-0.0358997062,
0.005275717,
0.0428735614,
0.0356202014,
-0.054904364,
0.0467583723,
-0.0044993176,
-0.0352446437,
-0.001479546,
0.0186782181,
0.0240509734,
-0.029361371,
-0.0115574161,
-0.0477147289,
-0.0306875557,
-0.0501824394,
-0.0338282622,
0.0324259698,
-0.0791832879,
-0.0444500148,
-0.0285929963,
-0.0483977273,
0.0441684574,
0.0256458018,
0.0348129869,
0.0303979851,
0.0096546803,
0.0135649908,
-0.0511420555,
0.0188424177,
-0.0212547667,
0.0697600469,
-0.123496756,
-0.0510279387,
0.0141134094,
0.0067523918,
0.028474113,
-0.009485065,
-0.0431250073,
-0.0504142158,
0.0371108465,
0.0732358098,
0.022860622,
0.0532187037,
-0.0103123784,
-0.067061305,
0.0135288304,
-0.0046257302,
-0.0278778318,
0.0565306358,
0.0324925184,
0.0042561884,
0.0150361955,
0.0287357327,
-0.0486527048,
0.0112160817,
0.0057604453,
0.0234371237,
-0.0084651178,
-0.0081020473,
-0.0168353505,
0.0298008416,
0.0015353203,
-0.0640719384,
-0.0111531038,
-0.0382333957,
0.0456105657,
0.0118388403,
0.0160676576,
-0.0589646399,
-0.0028885808,
0.0312806666,
-0.0129187582,
0.0058970195,
-0.0150449788,
-0.0285974257,
0.0089041302,
-0.0435181707,
0.1073886678,
0.0048354329,
0.0647395998,
0.0215907842,
0.0312344618,
-0.0181505103,
0.002528301,
-0.0082142726,
0.0216452349,
0.0391070247,
0.0827174187,
-0.0320861302,
0.0264638904,
0.0256590843,
0.0612754673,
0.0496213473,
0.0139141176,
-0.01620817,
0.0580686666,
0.0763930008,
0.0024613012,
0.039930962,
-0.0169128031,
0.0173403881,
-0.0192076564,
0.0244377349,
-0.0310906749,
0.0205645747,
0.0447422713,
0.0055712541,
0.0056462991,
-0.0090609724,
0.0620552525,
-0.038698379,
0.023874335,
0.0118455496,
0.0354507752,
-0.0455605537,
-0.0336340591,
0.0262418091,
-0.0028063278,
0.0225736424,
0.0222513136,
-0.0405246578,
-0.0418277383,
-0.0217205901,
0.0067194919,
0.0116543015,
-0.0452682227,
0.0259942245,
-0.0187599268,
0.0048347097,
-0.0063542323,
-0.0242564511,
-0.0361129157,
-0.0762812346,
-0.0199245699,
0.0441029482,
-5.632887012e-33,
-0.0793079585,
-0.003140792,
0.0238227639,
0.0245477166,
-0.0301770866,
-0.0003550215,
-0.0258463398,
-0.0007763984,
0.0040985285,
-0.0287513565,
-0.0231228434,
0.0068947473,
0.0297598653,
0.0007429881,
0.0472568385,
-0.0096911043,
0.0145451166,
-0.004758955,
-0.0235094931,
0.0131810009,
0.0116803581,
0.0249257628,
0.0906913504,
-0.0415451489,
0.0236802269,
-0.0321983211,
-0.007169175,
0.0313402452,
-0.0231982525,
0.072648488,
-0.0034350753,
0.0210372731,
0.00451208,
-0.0204808954,
-0.0138036227,
0.009547676,
-0.0416592471,
-0.0407854356,
0.0237199683,
0.0695594177,
-0.0293066353,
-0.0482430011,
0.0123916594,
0.0083310856,
-0.0555307418,
0.0312831029,
-0.0011671709,
-0.0156956967,
-0.0049714255,
0.0148307858,
-0.0289635975,
0.0401276909,
-0.0080366228,
-0.0269698165,
0.0049649021,
-0.0153562827,
-0.0170592647,
-0.0595787801,
-0.0204473976,
-0.0123209553,
0.0292304438,
0.0836288705,
0.0776834264,
0.0119364131,
-0.0505919382,
0.0192749333,
0.0258841347,
-0.0431815274,
0.0230878238,
0.0141930794,
0.006180957,
0.0811188146,
0.0083573619,
-0.0084619448,
0.0043615866,
-0.014647048,
-0.0065868832,
0.0557822958,
0.0454758182,
0.0636546835,
-0.0274462067,
0.0047895322,
-0.0340956636,
-0.0046634055,
0.0307546481,
0.0122860521,
-0.0124540748,
-0.0463158637,
0.0718163997,
-0.0168830976,
0.0136785191,
0.0029394601,
0.0220405944,
0.0007181739,
0.0592228845,
-0.0180020854,
-0.0278716236,
0.0399471968,
0.0089952517,
0.0082324296,
-0.0776900575,
-0.0643169954,
-0.0254268982,
-0.0007675855,
0.034754198,
-0.001238044,
0.0124089606,
0.0283896029,
-0.0537664071,
0.0151191698,
-0.0179103967,
0.0323472247,
0.0380362459,
0.0339448228,
0.0215123389,
-0.0354500674,
-0.0339456499,
0.0668825209,
-0.0320519209,
-0.0402309559,
-0.029015528,
-0.0442179888,
0.0174142849,
0.0153423678,
-0.0413039103,
0.0511746332,
-0.0513192117,
-0.0117095523,
-0.0299715959,
-0.0448071025,
0.0144539112,
0.0217945613,
2.486e-7,
0.0003868699,
0.0403536074,
0.0492758639,
0.0140047437,
-0.0238328837,
-0.01207685,
-0.0671186596,
-0.0101495031,
-0.0089629237,
0.0499655344,
0.020925669,
0.0265912656,
0.0588349588,
-0.004351717,
-0.0753858387,
-0.0057997997,
-0.0259425007,
0.045125287,
-0.0484811589,
0.0269579701,
-0.0066188155,
0.0509365499,
0.0009298837,
-0.0066970279,
-0.0200816188,
0.0087658027,
0.0133590745,
-0.0312581621,
0.0006293345,
0.0110495612,
0.023142565,
-0.0007015941,
-0.0555445179,
-0.0421958454,
-0.0536287688,
-0.0020741129,
0.0033084401,
0.1220138967,
-0.0144532397,
-0.0233075507,
0.0653427392,
0.0281003695,
-0.0216812287,
0.0105999494,
0.0461706929,
0.0504761264,
0.0029230334,
0.0242357999,
-0.0007442586,
-0.0431215428,
-0.0152850104,
-0.0167937465,
0.020352697,
0.0036604153,
0.0009497055,
-0.0690877512,
-0.0420561284,
-0.0302574784,
-0.0045006881,
0.0398717448,
-0.0105619011,
-0.0684911832,
-0.0353139304,
0.0169008747,
-0.0366979353,
-0.0578363799,
-0.0320952199,
2.197250112e-34,
0.0214851219,
-0.0048315856,
0.0492590964,
0.0646603927,
0.0005417631,
0.0112554757,
-0.0232109949,
0.042854555,
0.0234673023,
-0.0802604035,
-0.0372797102
] |
http://arxiv.org/abs/2401.14975v1 | http://arxiv.org/pdf/2401.14975v1 | Embedding-based search in JetBrains IDEs | 2024-01-26T16:07:42 | 2024-01-26T16:07:42 | Evgeny Abramov, Nikolai Palchikov | cs.SE | cs.SE, cs.CL, cs.IR, cs.LG | Most modern Integrated Development Environments (IDEs) and code editors have
a feature to search across available functionality and items in an open
project. In JetBrains IDEs, this feature is called Search Everywhere: it allows
users to search for files, actions, classes, symbols, settings, and anything
from VCS history from a single entry point. However, it works with the
candidates obtained by algorithms that don't account for semantics, e.g.,
synonyms, complex word permutations, part of the speech modifications, and
typos. In this work, we describe the machine learning approach we implemented
to improve the discoverability of search items. We also share the obstacles
encountered during this process and how we overcame them. | Embedding-based search in JetBrains IDEs: Most modern Integrated Development Environments (IDEs) and code editors have
a feature to search across available functionality and items in an open
project. In JetBrains IDEs, this feature is called Search Everywhere: it allows
users to search for files, actions, classes, symbols, settings, and anything
from VCS history from a single entry point. However, it works with the
candidates obtained by algorithms that don't account for semantics, e.g.,
synonyms, complex word permutations, part of the speech modifications, and
typos. In this work, we describe the machine learning approach we implemented
to improve the discoverability of search items. We also share the obstacles
encountered during this process and how we overcame them. | [
0.0309256539,
0.0250210315,
-0.0465713404,
0.0240217056,
0.0011700024,
-0.0337597579,
-0.0368178673,
0.0334324352,
0.0112099927,
0.0012918287,
0.0120121865,
0.0439391248,
-0.0286972895,
0.0101517905,
-0.0030397004,
-0.0124915559,
0.0799019113,
0.0237524956,
-0.0402536094,
0.0655844584,
-0.0270148814,
-0.005975021,
0.0017926337,
0.0995388478,
0.0032593769,
-0.0036570374,
-0.0188030135,
0.0036988237,
0.0098528564,
0.0079109445,
-0.0161731876,
0.0108771352,
-0.024961574,
0.0427996106,
0.0000018737,
-0.0148906708,
-0.0305084866,
-0.0273599718,
-0.0018526609,
0.0359026194,
-0.0031432195,
0.0407412425,
0.0059309117,
0.0211452376,
-0.0298099313,
-0.0055254153,
0.0431445129,
-0.0197313484,
0.040064469,
0.0119449217,
0.0155773377,
0.0362435207,
-0.0294105094,
0.0015190222,
0.0831742734,
-0.0451073833,
-0.0307466816,
-0.0250822566,
0.0583899319,
0.0265708007,
0.0184022319,
0.0585609227,
-0.0364322476,
-0.0056101172,
0.0669998527,
0.0284659099,
-0.0325763188,
-0.0313947834,
0.0445110612,
0.0114816306,
-0.0394609086,
-0.0270553175,
0.0210876036,
-0.0212387573,
-0.048197262,
0.0033843417,
-0.0089309495,
-0.0291695446,
-0.0187261906,
-0.0276598986,
0.0240364596,
-0.0517854989,
0.0093161101,
-0.032078959,
-0.0008800247,
0.0315986685,
0.0106945978,
0.0022857771,
-0.0132972859,
-0.0575585589,
-0.0200295616,
-0.0189667307,
0.0131998006,
0.0458928905,
-0.0184175614,
0.0214539338,
-0.0219463427,
-0.0546197705,
0.0506897308,
-0.0051138387,
-0.0278621875,
0.0482879616,
-0.0184879191,
0.0344808213,
-0.0201276708,
0.0491182953,
-0.0056487843,
-0.0158422701,
0.0132000335,
-0.0121034747,
-0.0893346742,
-0.0402913019,
-0.0533694886,
0.0186376348,
-0.0063866125,
-0.0138340993,
0.0141385533,
0.0480934717,
0.0283843148,
0.0067325849,
-0.0101136854,
-0.0084801894,
-0.0298911706,
0.0514196754,
0.0064959195,
-0.0489378348,
-0.0228157695,
-0.055461973,
-0.0420489497,
0.0143301403,
-0.0022297117,
-0.0138062909,
-0.0136109805,
0.0079255048,
-0.0077432403,
0.0234778672,
0.0608606599,
-0.0284858476,
-0.0239947569,
-0.0302120186,
0.0288961809,
0.0759745687,
-0.0384110212,
-0.0449547954,
-0.0558902547,
0.0098597249,
-0.0330154859,
-0.0542579368,
-0.0229109507,
0.0298106018,
-0.0063457955,
0.0308069959,
-0.0630471781,
0.0464948826,
0.0570302755,
0.0336049311,
-0.010529615,
0.0223666523,
-0.0180860274,
0.0130549837,
0.0397243127,
-0.0468110181,
0.0658458099,
0.0358166583,
-0.0103865881,
-0.0447078086,
-0.0006200974,
0.0027760854,
0.0151266595,
0.0041852347,
-0.0360835046,
-0.0012207122,
-0.0514289699,
-0.0022933194,
-0.0378564,
0.0729745552,
0.0745010749,
0.0174333695,
0.0304541402,
0.0405376516,
-0.0154960835,
-0.0593248196,
-0.0166617259,
0.0299873985,
-0.0169337504,
-0.017303573,
-0.0381253697,
-0.0021898879,
0.0418992229,
-0.0268248916,
-0.0176102538,
0.0225253906,
-0.027921807,
-0.0567609482,
0.0029626132,
0.0084037194,
-0.0094253933,
0.0282785222,
0.0082111191,
-0.0174717531,
-0.0390946977,
0.0060411496,
0.021259265,
0.0894207656,
0.1183638051,
-0.0859642923,
-0.0286881644,
0.0257674046,
-0.0578464456,
-0.0376649089,
0.0368060805,
-0.040768642,
-0.0095736561,
-0.0391341224,
-0.0305652469,
-0.0611801967,
-0.0041684569,
0.0666109249,
-0.0086568873,
-0.0083587011,
0.0233725645,
-0.039891541,
-0.0058804159,
0.0519934222,
0.0087695252,
-0.049683556,
0.0384956859,
0.013173189,
-0.0417818166,
-0.0068352511,
-0.0121758869,
0.0061814887,
0.0397577658,
0.0108131692,
-0.0373259597,
-0.0093585569,
0.0610479377,
0.0344932117,
0.0119591635,
-0.0281519759,
0.0057722032,
-0.0041295006,
0.0030464914,
0.0187369995,
0.0198943764,
-0.0469839759,
0.0489319488,
0.0387895927,
0.0409313887,
-0.0599414259,
-0.0436012708,
-0.0478035212,
0.0429405458,
0.0411855057,
0.0023148009,
0.011614833,
0.0430864692,
0.0621597581,
-0.0256336797,
-0.032911256,
0.0164269451,
-0.0125552099,
0.0148287946,
-0.0046982495,
0.0079264082,
0.0299276561,
-0.0263408739,
-0.0910554752,
-0.0034361116,
0.0046145837,
0.0328676254,
-0.0228006095,
0.0289254356,
0.0016088906,
-0.0401598364,
0.0087428456,
0.0087468997,
-0.0424794704,
-0.0194638614,
0.0380330011,
0.0102008069,
-0.0312463511,
-0.0655083805,
-0.0146390237,
-0.0630930588,
0.0436471924,
-0.0201318916,
-0.0357064009,
-0.0504275076,
0.0519576557,
-0.0162821151,
0.0115994774,
0.0197803564,
-0.018955335,
0.0049233697,
0.0038269726,
0.0322397836,
0.077832818,
-0.0179502685,
0.0419164598,
0.0503073297,
0.0165964793,
-0.0413029306,
-0.0042989473,
0.0443567894,
0.0236276165,
0.0313089676,
0.0036348859,
0.07343591,
0.0050940253,
-0.0319449604,
-0.0013558456,
0.0209319554,
0.0178891737,
-0.0084573049,
0.0000353065,
0.0064964117,
-0.0381354541,
0.0129575832,
0.0393703468,
0.0126503287,
-0.0849508718,
0.052638229,
-0.0063812304,
0.0010982058,
-0.0789091736,
-0.0486814715,
-0.0463442318,
0.0683857203,
-0.0426001251,
0.0065668547,
-0.0178941507,
-0.0518059023,
-0.0201198813,
0.0347069576,
0.0327897593,
-0.012266228,
0.0243488606,
-0.0031569118,
-0.0178587344,
0.0258047115,
-0.0149775026,
-0.0203548819,
-0.047490783,
-0.0127332378,
-0.0095563726,
0.0710105747,
-0.0174660515,
-0.0104894228,
0.060111735,
0.0126668867,
0.0475076474,
0.0625364259,
-0.0481719077,
-0.0239466149,
-0.0549461544,
0.0049113678,
-0.0855566263,
0.0048262095,
0.0456858166,
0.0002555408,
0.0077165412,
0.0580823906,
-0.013656579,
-0.0249558855,
-0.0500373244,
0.0125474185,
-0.058480259,
-0.0674336851,
0.0152228726,
-0.0306090787,
-0.0154121844,
-0.0342700854,
0.0503621027,
0.0914690569,
-0.05474158,
0.0253529195,
-0.0263287611,
-0.0013194322,
0.0805235878,
0.0663591474,
-0.0507148057,
-0.0262341481,
0.0038789867,
0.0065327496,
-0.0056019421,
0.023904996,
0.0197802819,
-0.0506341755,
-0.0006518787,
-0.019367924,
0.0116938064,
-0.015561237,
-0.0114570651,
-0.0168843381,
-0.0244497675,
-0.0699846596,
0.0125604719,
-0.0103105763,
0.0333744586,
-0.012395164,
-0.0507379398,
-0.0318794474,
0.0111994036,
-0.0482949018,
-0.0340511017,
0.0930346772,
0.0472423993,
0.04511052,
0.0022664426,
0.0062377858,
0.0022626463,
0.0237305798,
-0.0405672677,
0.0569693334,
-0.0074210302,
-0.0198501013,
0.0232523363,
-0.0167045314,
-0.0097528361,
-0.0254969914,
0.0520335771,
0.0650198832,
-0.0546301082,
0.0417028069,
-0.0718289539,
-0.0264076497,
-0.0012711097,
0.0064834608,
0.0611483641,
-0.0041909367,
-0.0096667008,
-0.0685572475,
0.039009288,
0.0439699367,
-0.0338036865,
-0.0664197206,
-0.0118110972,
0.0353500843,
-0.0582642965,
0.0013800982,
0.0434869789,
0.0791406706,
0.0311004315,
0.0111348992,
-0.0389242247,
0.0137269739,
-0.0751034543,
0.0075143683,
0.0258537028,
0.0312353875,
0.0087185884,
0.0074644964,
0.0265139975,
-0.0020231181,
-0.0298096184,
0.0116250254,
0.0041071484,
-0.0224836208,
-0.082479246,
-0.0009265491,
0.016087018,
0.0182502978,
-0.0823428929,
-0.0898392498,
0.0042097131,
-0.0005322815,
0.0078340909,
-0.02183998,
-0.0055643213,
-0.0357924998,
0.0106671341,
-0.0192275699,
-0.0378647372,
-0.0389437154,
-0.0301707145,
-0.0177938137,
-0.0314623043,
-0.0406185724,
0.025856996,
0.0126106506,
0.0196226742,
0.0681286678,
-0.0034391251,
-0.028316889,
0.0197219364,
0.0154940924,
-0.0137762818,
-0.0072173453,
0.0622493289,
-0.0030440572,
-0.0080517465,
0.0192089267,
0.0234518144,
-0.0379485935,
0.0417366661,
0.0143477693,
0.0498509705,
0.0144714247,
0.045508001,
0.0126356781,
-0.0031583544,
0.0029754657,
-0.0434079953,
-0.0062118308,
0.0572972968,
-0.0203823727,
0.0237383004,
0.0162463132,
-0.0073931725,
-0.0086216694,
-0.0252475161,
-0.0501431897,
-0.0459916145,
-0.0098922038,
0.0134720793,
0.0379330218,
-0.0130120879,
-0.0004436198,
-0.0282547586,
-0.0200453084,
0.0193279516,
0.0006460117,
-0.0168810859,
-0.0281863231,
-0.0265744552,
0.0237322208,
-0.0224614628,
-0.0554276556,
0.060781952,
-0.0330633894,
0.0232763663,
-0.0175820105,
-0.0533372648,
-0.0127090923,
0.0560140014,
0.1176025644,
0.0470880084,
-0.0366605259,
-0.0167224016,
-0.073271893,
-0.0039793262,
0.0412863754,
0.0431758314,
-0.0281596053,
-0.0041014203,
0.0083558904,
0.1182009131,
0.055382356,
-0.0386294834,
0.004490748,
-0.0339019187,
-0.0948496461,
-0.0170735456,
-0.0446738712,
-4.007292568e-33,
-0.0292246398,
-0.0664677024,
-0.0115787098,
0.0361098833,
0.0342708007,
0.0101340944,
-0.017888695,
0.0001395092,
-0.0056778579,
-0.0325640701,
-0.0073793828,
0.038740702,
-0.0073980652,
-0.0935331583,
0.0687932372,
0.0503946692,
0.0304339919,
0.0630233064,
-0.069519788,
0.0236158036,
-0.0097309221,
0.0325608477,
0.0214555208,
-0.046238333,
-0.0858795866,
0.0335116088,
0.0604901463,
0.0298758056,
0.0017917084,
0.0637970269,
-0.0115361251,
-0.0192339756,
-0.0293676276,
0.003200701,
-0.0031762812,
-0.0685839429,
0.0018683933,
-0.0391683355,
-0.0111393351,
-0.0044648848,
0.0442248508,
-0.0257646441,
0.0365882628,
-0.0044256966,
0.0101104202,
0.0459051356,
0.03721058,
0.0227894187,
-0.0559387691,
-0.0244536381,
-0.0466377251,
0.0220892504,
-0.0296651404,
0.0132977916,
0.0353792906,
0.0223128535,
0.0356687494,
0.0066351369,
-0.0337073542,
-0.0120563172,
-0.0123272324,
0.0488450155,
0.0621426366,
-0.0025437083,
-0.0002825953,
0.0319483466,
0.0363766514,
-0.0044592991,
0.0213602986,
0.013772497,
0.0354222283,
0.059384089,
0.0550448634,
0.019954741,
0.0427610204,
-0.0076015689,
0.0529645532,
0.0216073487,
-0.0015167459,
0.0271104779,
0.0150085026,
0.0772852078,
-0.0586389452,
0.0018372653,
-0.0293967221,
0.0050386349,
-0.0402893871,
-0.0483947583,
0.0362274721,
-0.0231393818,
-0.0190742034,
0.0021626863,
-0.0055466476,
-0.0075234007,
0.0153429275,
0.0526234247,
-0.0348809212,
0.0354267508,
-0.0258908719,
-0.0001760959,
0.0101382369,
-0.0044491962,
0.0060476479,
-0.0224452876,
0.0516796969,
0.0023447943,
-0.0143892532,
-0.0044551389,
-0.0549762249,
-0.0214416701,
0.0134607786,
-0.0703039989,
-0.0858235508,
-0.0396431088,
0.005389343,
-0.0319657661,
0.0086131841,
-0.0060074343,
0.0556480885,
0.0067302301,
-0.0541953705,
-0.0390498117,
0.0033035628,
-0.0146548133,
0.0010447158,
-0.0056197406,
-0.0314384848,
0.0153595479,
0.0257737841,
-0.0223267786,
-0.0221169293,
0.0267063379,
2.46e-7,
0.0003733426,
0.0837414041,
0.0120539526,
-0.0018218874,
0.0037531534,
0.0389368348,
-0.0109120794,
-0.0417080224,
0.0086160582,
0.0417206772,
0.0002645626,
0.0044854153,
0.0118345479,
0.0299701728,
-0.0440736935,
-0.0106126778,
0.0204048567,
-0.0345304906,
-0.0945527703,
-0.0094038341,
0.0486451909,
0.0757055953,
0.0271457452,
0.0332232937,
0.0065082433,
-0.0364755243,
-0.0133114159,
0.0266093612,
0.0455163456,
0.0436298922,
0.0120514529,
0.0146123469,
0.004670802,
-0.0280374959,
-0.0389203914,
0.0102595007,
0.0467699096,
-0.0241067931,
-0.0393736809,
0.0320008695,
-0.0087223416,
-0.0252623297,
-0.0137342047,
-0.0274823308,
0.0054194229,
-0.0258202255,
0.0304486956,
0.0000420709,
-0.0209537186,
-0.0299661737,
0.000991711,
-0.0236063469,
0.0233774707,
0.0063501252,
-0.0270701982,
-0.0221083779,
-0.030868331,
-0.0613921657,
0.0048960936,
0.0658345968,
-0.0187582523,
-0.0039945268,
-0.051630348,
0.0298609566,
0.0511782914,
0.0115800416,
0.0196258035,
2.563942131e-34,
0.0032583668,
-0.029243568,
0.0084750988,
0.0344934836,
0.0622907132,
0.0018595073,
0.0019353819,
0.0221663993,
0.0195527319,
-0.0355539843,
-0.0174424388
] |
http://arxiv.org/abs/2401.14931v1 | http://arxiv.org/pdf/2401.14931v1 | Do LLMs Dream of Ontologies? | 2024-01-26T15:10:23 | 2024-01-26T15:10:23 | Marco Bombieri, Paolo Fiorini, Simone Paolo Ponzetto, Marco Rospocher | cs.CL | cs.CL, cs.AI | Large language models (LLMs) have recently revolutionized automated text
understanding and generation. The performance of these models relies on the
high number of parameters of the underlying neural architectures, which allows
LLMs to memorize part of the vast quantity of data seen during the training.
This paper investigates whether and to what extent general-purpose pre-trained
LLMs have memorized information from known ontologies. Our results show that
LLMs partially know ontologies: they can, and do indeed, memorize concepts from
ontologies mentioned in the text, but the level of memorization of their
concepts seems to vary proportionally to their popularity on the Web, the
primary source of their training material. We additionally propose new metrics
to estimate the degree of memorization of ontological information in LLMs by
measuring the consistency of the output produced across different prompt
repetitions, query languages, and degrees of determinism. | Do LLMs Dream of Ontologies?: Large language models (LLMs) have recently revolutionized automated text
understanding and generation. The performance of these models relies on the
high number of parameters of the underlying neural architectures, which allows
LLMs to memorize part of the vast quantity of data seen during the training.
This paper investigates whether and to what extent general-purpose pre-trained
LLMs have memorized information from known ontologies. Our results show that
LLMs partially know ontologies: they can, and do indeed, memorize concepts from
ontologies mentioned in the text, but the level of memorization of their
concepts seems to vary proportionally to their popularity on the Web, the
primary source of their training material. We additionally propose new metrics
to estimate the degree of memorization of ontological information in LLMs by
measuring the consistency of the output produced across different prompt
repetitions, query languages, and degrees of determinism. | [
0.0223691277,
0.0391430333,
-0.0085608577,
0.0227250364,
-0.0139145516,
0.0220553875,
-0.015088344,
0.0044429637,
-0.009129121,
-0.065863274,
-0.0554910712,
-0.0638410598,
0.0007543281,
0.0208060294,
-0.0138606448,
-0.0129618607,
0.0117533356,
-0.0319974683,
-0.0446046367,
-0.0112849856,
-0.0315217264,
0.0330883563,
-0.0035265347,
0.0114815608,
-0.0176504515,
-0.0220900066,
-0.0222767871,
-0.0160071328,
-0.0184309874,
-0.0171415843,
0.0108934995,
0.0360715017,
0.030123299,
0.0463193394,
0.0000018057,
-0.0534268804,
-0.0178165864,
-0.0296274275,
-0.0280144177,
0.0120130032,
0.034056671,
-0.0057873446,
0.0207182709,
-0.0035734375,
-0.0365007408,
-0.0110532297,
0.0269383192,
0.037342228,
0.0358069465,
0.0404569842,
0.0141360341,
-0.0082909465,
0.0550515465,
0.0054164235,
0.0563756749,
-0.017236894,
0.0272395592,
0.0263123214,
-0.0242034402,
0.024055779,
0.0243978966,
0.024265008,
-0.0020155476,
0.0271156412,
0.0030466819,
0.0363946185,
0.0191207658,
0.001835533,
0.0044066287,
0.0526349433,
0.0320805609,
-0.0046993452,
0.0110552264,
-0.0052314741,
-0.004011774,
0.0323662497,
-0.0316274501,
0.0772008449,
-0.0156364776,
0.0016460511,
0.0151284439,
0.0027297307,
0.0288839657,
0.0334081799,
-0.0455221348,
0.0920365229,
-0.0134267267,
-0.0108755445,
0.0056255022,
0.0083416244,
0.003471798,
-0.0712209046,
-0.0008665664,
0.054054819,
0.0006917118,
0.0570856482,
-0.0643865243,
0.0048699002,
0.0217216536,
-0.0615467988,
0.0386476517,
0.0291481782,
-0.0190544967,
0.0115576191,
-0.0524494573,
0.0494581684,
-0.0037788108,
0.0042966451,
-0.029368734,
0.0099590812,
-0.0875577554,
-0.0389837511,
-0.0394506156,
0.0738821328,
0.0487124622,
-0.0295949709,
-0.0526605695,
0.0674615726,
0.0059911986,
0.0363009982,
-0.0206080973,
0.0205054041,
-0.0103105679,
-0.0231926814,
-0.073315829,
-0.0279105268,
0.0098714167,
0.00435889,
0.0388037413,
-0.0480863005,
0.0371648818,
-0.0082254289,
0.0190539099,
-0.0059235902,
-0.0070267613,
0.0867713168,
0.0004603959,
-0.0326355882,
-0.1067723185,
-0.0992450416,
-0.0058185975,
-0.0450439975,
0.0236036945,
-0.0497706421,
0.0007764925,
0.0335980579,
0.0048223794,
-0.0496443622,
0.0062940335,
0.0278616566,
-0.0478292517,
0.0421416163,
-0.0085799098,
0.0209896434,
0.0505710989,
0.021534957,
-0.0016568915,
0.0588012375,
0.0065006833,
0.0527327694,
0.0433681048,
0.0143385762,
-0.0247608498,
-0.0069520036,
0.0252408665,
-0.0130121838,
-0.0443307459,
0.014463882,
-0.0015273133,
0.0550757758,
0.045852799,
0.0479008444,
-0.0238850862,
0.0199404769,
0.0811814591,
0.0440165661,
0.0273544062,
0.0261898674,
0.0920875221,
-0.0195006113,
0.0298262965,
0.0587028824,
-0.0428522862,
0.0418528356,
-0.0131136924,
0.0415161513,
0.0181688871,
0.0037705265,
-0.0408679582,
-0.039100457,
-0.0071100122,
-0.0002284728,
-0.0971572995,
-0.0086661791,
0.0393492728,
0.0159885157,
-0.0250539687,
0.0816575363,
0.0123168891,
-0.0246349443,
0.0562571622,
-0.0379398018,
0.0442219526,
0.0623396859,
0.0334308706,
-0.0268570781,
-0.0504599437,
0.0117680468,
-0.0101208249,
0.0268337782,
0.0502114482,
-0.033477705,
-0.0130364532,
-0.0252571236,
0.0204542819,
-0.0445960909,
0.0367640033,
0.0373371951,
0.0302138776,
0.0097879432,
0.0531351157,
-0.0561633222,
-0.0523329824,
0.0383627601,
-0.0549450032,
-0.0436902419,
-0.0088781146,
0.0175033715,
0.0156694166,
-0.0007681478,
-0.0262144655,
0.0159887355,
0.0145113235,
-0.0085673342,
-0.0227843989,
-0.0224918369,
-0.0685375333,
0.0166077092,
0.018525809,
0.0112376837,
0.0218879543,
0.0194913968,
-0.0130236764,
-0.0336704887,
-0.0171560831,
-0.007488417,
0.0562154129,
-0.0485208035,
0.0224983599,
0.0132882437,
-0.0274441801,
-0.0137561141,
0.0657091141,
-0.0055177412,
0.0340551957,
0.0119894613,
-0.0153088607,
0.0161896143,
-0.0324670859,
-0.0041986201,
-0.0279755015,
-0.0190977901,
0.0348724648,
0.0123687796,
-0.0114772441,
-0.0504803546,
-0.0139937904,
-0.0591447279,
0.0014920828,
0.0023977649,
0.0371368192,
0.0400989465,
-0.0203692056,
-0.0011449271,
0.0160368625,
-0.0844256207,
0.0268710013,
0.0093613584,
0.024771627,
0.0585367717,
0.0209208634,
0.0107893888,
-0.0136055807,
0.0096661542,
-0.0003403524,
0.0354755037,
0.0092444001,
0.032920707,
-0.0480734296,
0.0368928947,
-0.0721032396,
-0.0232159346,
-0.0009794292,
-0.0087135434,
0.0101621607,
-0.0182509627,
0.0046357131,
0.0753102824,
0.0605098642,
-0.0046500321,
0.0703049675,
0.0071921665,
-0.0322177708,
-0.0334150791,
-0.0376897641,
-0.0398890488,
-0.0105891367,
-0.0404479019,
0.0400681943,
0.0873423219,
-0.0089949379,
-0.0157691762,
-0.0134576345,
0.0364474431,
-0.0434885994,
-0.0135230897,
-0.0717831552,
0.0232848395,
-0.0124151064,
0.067449823,
0.0243140385,
-0.0651211664,
0.0442225821,
-0.0265887938,
0.0260301158,
0.0278285295,
-0.0016709408,
-0.0198297556,
0.0681961104,
0.0100392085,
0.0759968311,
0.0456237793,
-0.0366409197,
-0.064569287,
-0.0091702174,
0.0269860346,
-0.0241092723,
0.0320438035,
-0.033023268,
0.0248850193,
0.0032266877,
0.0095416103,
-0.0786770433,
0.0379524715,
-0.0187083501,
0.0208650697,
0.0212856233,
-0.0366683379,
-0.0065860068,
-0.00814474,
0.0304445922,
0.0136533277,
0.0056851134,
-0.0553499758,
-0.0235235784,
0.0672336668,
0.0597591214,
-0.0274038073,
0.0119419117,
-0.0452022366,
-0.0162750948,
0.0195995681,
0.0050688004,
-0.0059678294,
-0.0217084121,
-0.0174733326,
0.0365351997,
-0.0910256356,
-0.0150416624,
0.0566423945,
-0.0169864763,
-0.0343110897,
-0.0137004601,
0.1199068278,
0.0608082078,
-0.0078182342,
0.0010617085,
0.0205669515,
0.0112037575,
0.004308064,
-0.0009059314,
-0.0842069313,
0.0193056185,
-0.0097514428,
0.0083720554,
-0.0269045588,
-0.0385534465,
-0.0493028164,
-0.088681452,
-0.0026002815,
-0.0097648567,
-0.0587035418,
0.0347488336,
0.0123299705,
-0.0263519604,
-0.0214881506,
0.0061913803,
0.0333728641,
-0.0625804141,
-0.0111255106,
0.0364255682,
0.0056429054,
-0.0305410028,
0.0182331484,
-0.0191733446,
0.0360585563,
0.0085908575,
0.0334844328,
-0.0090613058,
-0.0265915599,
0.0275402572,
-0.0172045827,
0.0365241393,
-0.0022312365,
0.0725113004,
-0.0330190063,
0.0119936308,
-0.0282539856,
-0.0089996336,
-0.0338556208,
0.0104573732,
0.0013661773,
0.0222373493,
-0.0269958843,
-0.0370003805,
0.0146911973,
0.0050408873,
-0.0374231935,
0.070934087,
0.0075917495,
-0.0060993875,
-0.0077441307,
-0.0208762903,
0.0140950475,
-0.0510106049,
-0.0189077444,
0.0225515533,
-0.0642438754,
-0.0229242295,
-0.1124402881,
-0.0195289552,
-0.0027203893,
0.0459508002,
0.0195744485,
-0.0209839512,
0.029675087,
-0.0097280359,
-0.0430932529,
0.0117543321,
-0.0566598698,
-0.0393114276,
-0.0662265569,
-0.0510310344,
0.0672617555,
-0.046776548,
-0.0092761097,
-0.0194005184,
-0.052885104,
-0.0633601472,
-0.0130237192,
0.0278825015,
0.0330002755,
0.0129311047,
0.0417500362,
-0.0155552188,
0.0787466317,
0.0350272954,
-0.0606211759,
0.01249664,
0.0113692172,
0.0486320518,
-0.016511986,
0.0445206054,
-0.0399243385,
-0.0031072935,
0.004726788,
-0.0017011277,
-0.0042991373,
-0.0213862211,
0.0608802959,
0.0106190555,
-0.022796452,
-0.0059372033,
0.0064442609,
-0.0543672554,
-0.0054900991,
0.0042538056,
-0.0056293579,
-0.0176911671,
0.0189866386,
0.0297660735,
0.0147039276,
-0.0107521676,
0.0101608951,
-0.0034073577,
0.0638951659,
0.0110089947,
-0.0839217976,
-0.0251921676,
0.0331522264,
0.0471699387,
-0.0064444211,
-0.0577848442,
0.013308702,
-0.0285561029,
0.0349412784,
0.000077114,
0.0809884146,
0.0126794977,
0.0273691211,
-0.043037191,
0.0107136099,
-0.0569276474,
-0.1165873632,
-0.0175453033,
0.0076939659,
0.0710858107,
-0.0403719582,
0.01985755,
-0.0846923739,
0.0113276318,
-0.0294413622,
0.0138705363,
0.017398404,
-0.0157503579,
0.0008009437,
0.000074033,
-0.006331014,
-0.0116265398,
0.1286561489,
0.0347770937,
0.0404777676,
0.0016724769,
-0.039591115,
-0.0243065413,
-0.0260315239,
0.0540452339,
0.0253650118,
-0.0529973544,
-0.015788503,
-0.0297480319,
-0.0842998624,
0.0252897348,
-0.0005446463,
0.0140304696,
-0.0088568712,
0.0522571392,
0.0709694251,
0.0135393562,
0.0171941947,
-0.0556056723,
0.0358409248,
-0.0596731342,
0.0069969124,
-0.0045178193,
-6.287774384e-33,
-0.0289478302,
-0.0336861312,
0.0378040411,
0.0128768198,
-0.017029658,
0.0120128617,
-0.0145857232,
0.0128291156,
-0.0003524611,
-0.0270046815,
0.0037281122,
-0.0165748559,
0.0004999707,
0.0044350908,
0.0180285219,
-0.035198912,
0.0209421348,
-0.0137513736,
-0.0365941897,
0.0283848736,
0.0906020328,
0.0203500502,
0.0097193439,
-0.0092924871,
0.0512302145,
-0.0045945086,
0.0232919604,
-0.0134580061,
0.0030506421,
0.045200441,
-0.0417729169,
0.0336153954,
0.0481788032,
-0.0142111536,
-0.0068316311,
0.0032711704,
-0.0434509329,
-0.0369440466,
0.019204827,
-0.0034342846,
0.0007088465,
-0.0384842455,
0.0328939818,
-0.0181616843,
0.0116754109,
0.0022989537,
0.0339675024,
-0.0177833941,
-0.0414896756,
-0.0661174282,
-0.0535469316,
-0.0006026852,
0.0265758224,
-0.0168774314,
-0.0065310411,
-0.008220043,
0.0212583803,
-0.0121335108,
-0.0334072597,
0.0375549048,
0.0178000871,
0.0444389246,
0.050576292,
0.0493656546,
0.0171888955,
0.0719928741,
0.0049136579,
0.0123582771,
0.0196192898,
0.0206762794,
0.0237796865,
0.0261792839,
-0.0417679697,
0.0112532042,
0.027421128,
-0.0144198174,
-0.0292342696,
0.0349166505,
-0.0116528962,
-0.0067888754,
0.0500444658,
-0.0329253115,
0.0059392536,
-0.0081777144,
-0.0499140918,
0.0584801808,
-0.0440377183,
-0.0472366773,
0.026125066,
0.0214364883,
0.0178753305,
-0.0426957607,
0.0037772586,
-0.015644528,
-0.0246284287,
-0.0105976239,
-0.012992396,
-0.004893647,
-0.0344428942,
0.0104340259,
-0.0760750771,
-0.0444541946,
-0.0276489723,
-0.028739769,
-0.0171073899,
0.0449247062,
0.0245098993,
0.0192860048,
-0.0457941256,
0.0010408729,
-0.0403595641,
-0.0238301214,
0.0158380456,
-0.0127877034,
0.0408380292,
-0.0047556465,
-0.0129850581,
0.0286515243,
0.0076351566,
-0.0356284976,
-0.0399602875,
0.0040523494,
0.0024481935,
0.0477499738,
-0.0537169762,
-0.0095676752,
-0.0428533331,
-0.0239344593,
0.0189688858,
-0.0149799138,
0.0080830846,
-0.0153895589,
2.59e-7,
0.0182825178,
0.0248928461,
-0.0285599679,
0.0473541915,
0.023288196,
0.0123363128,
0.0009171686,
0.0323004574,
-0.0117358873,
-0.0460102744,
0.0128340935,
0.0294784121,
0.0143735455,
0.0370159186,
-0.0881601274,
-0.0284852944,
-0.0000411048,
-0.0187651943,
-0.0590606369,
-0.0034768539,
0.0318879485,
0.0172531717,
0.0019921202,
-0.0147657515,
0.0180311296,
-0.0524170808,
0.0990429297,
-0.0203800984,
-0.0026206938,
0.0485732034,
0.0528163761,
0.0519539937,
-0.0289335493,
0.0317380391,
-0.0049315481,
-0.0201900806,
-0.0066961739,
0.0283645019,
0.0377266817,
-0.0001716502,
-0.0052082292,
-0.0579024553,
0.003075968,
-0.0824788362,
0.0390398428,
-0.0711634606,
-0.0073784813,
-0.015436342,
0.0140319774,
-0.0193858091,
-0.0140613429,
-0.0280723814,
-0.0081572635,
0.0148070753,
-0.0228825603,
0.030617157,
0.0146329338,
-0.0437555127,
0.018776346,
0.045569811,
-0.0106748249,
-0.1006352156,
-0.0587051846,
-0.0091828359,
0.044723846,
-0.0775581971,
0.0099425046,
1.847957131e-34,
0.012275042,
0.0198356081,
-0.0207571611,
0.0221942924,
0.0178570766,
-0.0125766676,
0.0243225023,
-0.0009066259,
0.0101715969,
-0.0806732923,
0.002123585
] |
http://arxiv.org/abs/2401.14890v1 | http://arxiv.org/pdf/2401.14890v1 | Comparison of parameters of vowel sounds of russian and english languages | 2024-01-26T14:15:20 | 2024-01-26T14:15:20 | V. I. Fedoseev, A. A. Konev, A. Yu. Yakimuk | cs.SD | cs.SD, cs.CL, eess.AS, 68T10, H.2.8 | In multilingual speech recognition systems, a situation can often arise when
the language is not known in advance, but the signal has already been received
and is being processed. For such cases, some generalized model is needed that
will be able to respond to phonetic differences and, depending on them,
correctly recog-nize speech in the desired language. To build such a model, it
is necessary to set the values of phonetic parameters, and then compare similar
sounds, establishing significant differences. | Comparison of parameters of vowel sounds of russian and english languages: In multilingual speech recognition systems, a situation can often arise when
the language is not known in advance, but the signal has already been received
and is being processed. For such cases, some generalized model is needed that
will be able to respond to phonetic differences and, depending on them,
correctly recog-nize speech in the desired language. To build such a model, it
is necessary to set the values of phonetic parameters, and then compare similar
sounds, establishing significant differences. | [
-0.0039618756,
-0.0245628655,
0.0241433252,
0.0215868019,
-0.0578424446,
-0.0056264643,
0.0115403356,
-0.0140805896,
-0.0159041081,
0.0170314461,
-0.0295332074,
-0.0798900649,
0.065069668,
-0.0388547331,
0.02658589,
-0.0512919128,
0.0051656216,
-0.068921566,
0.0211353097,
-0.0327742323,
0.0208179895,
0.0008490222,
-0.0735372752,
0.0310015325,
-0.0089849588,
0.0558279492,
0.0133082755,
0.0093816267,
-0.0047404421,
0.0123424912,
-0.0321303532,
0.0084847659,
-0.0248571318,
0.013733766,
0.0000016258,
-0.0462240763,
-0.0287493113,
-0.0180302374,
-0.0183772836,
-0.018581124,
0.0658940896,
-0.0028060784,
-0.0539389923,
-0.0004432611,
0.0185967237,
-0.0194222294,
0.0902728215,
-0.018130092,
-0.0250261575,
0.0359322764,
-0.0040551191,
-0.0842429996,
-0.0188704841,
0.0226679668,
-0.0712147132,
-0.0253477339,
0.026363112,
-0.0132446075,
-0.0044931909,
-0.1070849076,
-0.0128318677,
0.0372684039,
-0.0420694016,
0.0046711029,
-0.0718789548,
-0.0131421778,
0.0132732205,
-0.0490472503,
-0.0378526263,
0.0176329203,
-0.0568828881,
-0.0337597691,
0.063655518,
0.0180540774,
-0.0327466466,
0.0031882504,
-0.0026463866,
-0.000815661,
-0.0327703618,
-0.0348424315,
0.0108329868,
0.0303934906,
-0.0354777575,
-0.0371454805,
0.0112930527,
0.0348428264,
0.0123428619,
-0.0118970834,
-0.0032651594,
-0.0441297516,
0.0439873263,
0.0464110598,
0.0810162351,
0.0349534824,
-0.0377296582,
-0.0134792356,
0.0192625746,
-0.0736388117,
0.0016146987,
-0.0227474142,
0.0076693981,
-0.0046474538,
0.0030152285,
-0.0195512343,
-0.0429239571,
0.0598073751,
-0.0394196473,
-0.0060465713,
-0.0754421726,
-0.0340151079,
-0.0211572573,
0.0161085539,
-0.024260385,
-0.03118301,
0.0512824319,
0.0311489329,
-0.0194522534,
-0.013133131,
0.0025042102,
0.0318263769,
0.010765682,
-0.0357714705,
-0.0159865599,
-0.0212726835,
0.0259152688,
0.0254797824,
-0.0191101264,
0.0035811483,
-0.0776420087,
-0.0718733147,
0.0579397865,
0.0174997579,
0.0181543697,
0.0552133769,
0.0100976462,
0.0775959119,
0.0207398031,
0.0320833363,
-0.0238515958,
0.0187475439,
0.0129631609,
-0.0650624633,
-0.0166277252,
-0.037671905,
0.0279425792,
-0.0144920843,
-0.0068110265,
0.0004107992,
-0.0339175314,
-0.0429805294,
0.0297108702,
0.0192064065,
0.0816823617,
0.0495252647,
0.0595803,
-0.0344679281,
0.0536143929,
0.075252682,
-0.0221512225,
-0.0342009589,
0.0315199196,
-0.0590764768,
0.0097870911,
0.0610192083,
-0.0039113495,
-0.0178681146,
-0.0147854062,
-0.0176997762,
0.0155477906,
0.0111179948,
0.045854196,
0.0325839967,
-0.0335194729,
-0.0378283188,
0.046715077,
0.0580996834,
0.0645478666,
0.0310109034,
0.0539037809,
-0.0208932739,
0.0150927231,
0.0536737069,
0.0189385992,
-0.0097634355,
-0.0101400269,
0.0458492264,
-0.0268578883,
0.0479553305,
-0.043482136,
0.0036784706,
-0.0171956141,
-0.0407433696,
-0.0036339697,
-0.0162178222,
0.0780995637,
-0.0348980166,
-0.0478538349,
0.0083653042,
-0.0077016074,
-0.0212250464,
0.0204517096,
-0.0195201021,
0.0405821614,
0.0229756273,
-0.0040925308,
-0.0462957695,
0.0341013968,
0.0121749844,
-0.0722048208,
0.0141351921,
0.0315906964,
-0.0184116475,
0.034379337,
-0.0529706106,
0.0046460084,
0.0774216577,
-0.019101629,
0.083899565,
-0.0458739027,
0.0334540494,
0.0200915411,
-0.0343546048,
0.0418927595,
-0.0523323119,
0.0160646848,
0.0109092202,
-0.0274075288,
0.0049357866,
0.05037985,
0.0572609417,
-0.0016472093,
0.0495125353,
0.0215782858,
-0.0107469959,
0.0485332608,
0.028492799,
0.0135635445,
0.0104302382,
0.0463714451,
-0.0052972068,
0.0187958032,
0.0065513658,
0.0399841852,
-0.0520173423,
0.0903075412,
-0.0030894149,
0.0434073918,
-0.026825998,
0.0807894021,
-0.0019754646,
-0.0044105477,
0.0110592609,
-0.035371542,
0.0168670751,
0.018453205,
0.0126104178,
-0.0250835326,
-0.0133647704,
-0.1062406972,
-0.0302846916,
0.1019941717,
-0.0194573682,
0.012290434,
-0.0197993908,
-0.0496232994,
0.0026351558,
0.0089403875,
-0.0367119648,
-0.0294916425,
-0.0054481765,
0.0576585867,
-0.0233307928,
0.0013702952,
0.001034284,
-0.0272770841,
-0.0594581626,
-0.0042041028,
0.0036463977,
-0.0090747578,
-0.0389216654,
-0.008986718,
-0.0110759912,
-0.0208465792,
-0.0310585108,
0.0249283575,
0.0243949629,
-0.0426635668,
-0.045688916,
-0.1014119461,
-0.0200566351,
-0.0035861847,
0.0490489751,
0.023899477,
-0.0623984337,
0.006121194,
-0.0134691112,
0.0641836822,
0.0157788675,
-0.0224068556,
-0.0297960751,
0.0531944782,
-0.0206906758,
0.0058259419,
-0.0203286074,
-0.0150650367,
0.0589930639,
0.004231995,
0.0086342357,
-0.0307922997,
0.0164901055,
0.0163512286,
0.0128489994,
-0.0290542152,
-0.013876982,
-0.0308411103,
-0.0130750807,
-0.0124201365,
-0.0220949482,
-0.0181458872,
-0.0014027902,
0.0043784305,
-0.0150721315,
0.0388830118,
-0.1191017553,
-0.0249000117,
-0.0612041317,
-0.0291299839,
-0.0029480255,
-0.0293832049,
-0.0098537123,
-0.0232363921,
-0.0300736334,
-0.049323447,
-0.0043825964,
-0.0017788822,
0.0746934786,
-0.051260028,
0.0078959409,
-0.0147988312,
-0.0221759137,
-0.0351316966,
0.0532640554,
0.041263856,
0.0068037254,
-0.0370716378,
0.0368293673,
-0.0092209028,
-0.0093772933,
0.0483078621,
0.0632926747,
-0.0093523022,
0.0262546334,
-0.0275519844,
0.0769697577,
-0.024274895,
0.0379730649,
0.0297464728,
0.0271211844,
-0.0577073283,
0.0151262628,
0.0383053944,
-0.0564383604,
0.0666139647,
-0.0092377169,
-0.0306226909,
-0.0231248904,
0.0343169346,
-0.0126161426,
-0.0007292548,
-0.004173615,
-0.0284047853,
-0.0033846027,
0.0567875318,
0.0465365499,
-0.0241489597,
0.0096642394,
-0.0021517605,
-0.0151150245,
0.0011836645,
-0.0233613122,
-0.015931638,
0.0275684334,
0.0307662264,
0.0150568699,
0.0060942448,
-0.0093723014,
0.0353298821,
-0.0132547468,
-0.0782011449,
-0.0245197453,
0.0266709011,
-0.0195812937,
-0.0420928597,
-0.0258607827,
-0.0042964001,
-0.0194655694,
-0.0082301833,
0.0503283888,
-0.0135907624,
-0.0268370099,
0.0450235978,
-0.0436864011,
-0.0456864126,
-0.0367235281,
-0.0519433543,
0.0434536077,
0.0132347448,
-0.0067880135,
-0.0325619467,
-0.0051137465,
-0.031718608,
-0.0274210144,
-0.005802758,
-0.0185178462,
-0.0208114106,
0.0625156984,
-0.027667474,
-0.0036340342,
0.0049750376,
0.0222148336,
0.0247809496,
-0.0047651152,
-0.0129820127,
0.0014805221,
0.0043050754,
0.0236124098,
0.0546124801,
0.0056497385,
0.0215401165,
0.0297764074,
0.0093872109,
-0.0277700201,
0.0511840433,
0.0210609231,
0.0038606003,
0.0161314029,
0.0247821659,
-0.0126767168,
-0.1084756479,
-0.0701335594,
0.0129481032,
-0.0266649406,
0.056473542,
0.0164006017,
0.0641479641,
-0.0050240145,
0.0295384955,
-0.002094527,
-0.0066357013,
0.0589891337,
0.0451279916,
-0.0467344224,
-0.000546261,
0.007413472,
-0.0742916167,
-0.0133548621,
-0.0082414318,
-0.005566468,
-0.0471644811,
0.0309765711,
-0.0261177085,
0.023139311,
0.054336302,
-0.0475226641,
-0.0607036389,
0.0174822938,
-0.081069015,
0.0245411303,
0.0132502634,
0.0555119216,
0.0107568959,
-0.0106991092,
0.0070666247,
-0.028901834,
0.0213280115,
-0.0067809382,
-0.0376683883,
-0.0186412912,
-0.0444571786,
-0.032646887,
0.0236559212,
0.0403571837,
0.0153129082,
0.0154891778,
-0.0184955485,
0.0267448388,
0.0247722771,
0.072320722,
0.029330207,
-0.0132868811,
-0.0070350128,
0.0240521431,
-0.0221602563,
-0.0593811981,
-0.0328756608,
-0.0544632971,
0.0376202427,
0.0599054731,
0.0704581961,
0.0432055704,
0.0109473513,
-0.0124591431,
-0.0466815531,
0.0621475615,
-0.0173629578,
-0.0034577274,
-0.0255977903,
-0.0241904743,
-0.0424729064,
-0.0348263457,
0.0369724073,
0.0505628102,
0.0025662158,
0.006384823,
0.015443394,
-0.0005048352,
0.0338503383,
-0.0265596975,
0.0519302897,
-0.0151208006,
0.038447693,
-0.0250764154,
-0.0064616357,
-0.0061249025,
0.0299085509,
0.0010605947,
-0.0043711974,
0.0108056674,
0.0033901406,
0.0189462304,
0.009546021,
0.0308692306,
-0.0379071832,
0.0315846168,
-0.0071898713,
-0.0249917079,
-0.0691689029,
-0.0507598557,
-0.0672816485,
-0.0270788334,
-0.0024080875,
-0.0272018854,
0.0190052036,
0.0296590626,
0.0360299125,
0.0070056254,
0.0058147581,
-0.003152733,
-0.0313961729,
0.0217105206,
0.0076413816,
-0.0247944295,
-0.0206750799,
0.0048074522,
0.0266090631,
-5.239696641e-33,
0.0142603721,
0.0345835164,
-0.0106825689,
-0.0064557623,
0.0030125624,
-0.0225733612,
0.0013052773,
0.0085552363,
-0.0422746651,
-0.0429348461,
0.0175070968,
-0.0208633896,
0.0433688723,
-0.0719318241,
-0.0293728337,
-0.0076096202,
0.0477734506,
0.0000092853,
0.0263764374,
-0.0212289467,
0.0586197563,
0.0215919781,
0.0668640733,
-0.0140039129,
0.0008247594,
-0.0567946248,
-0.0174178779,
-0.0026422699,
-0.0012344562,
0.0198566131,
0.0169921014,
0.0718609467,
-0.0080691893,
-0.0201787222,
-0.0237555038,
0.0010260058,
-0.0422717854,
0.0094403299,
0.0611369275,
0.0969904289,
0.0265876502,
-0.0201869514,
0.0124728,
-0.0171116889,
-0.0255642477,
0.0182338059,
-0.0221075006,
-0.0473609865,
0.0175033547,
0.0013108082,
-0.0333807468,
0.035383001,
0.0113214813,
-0.0164426453,
0.1254193187,
0.0697519183,
-0.0389405116,
-0.0248155817,
0.0009071197,
0.0650788397,
0.0188556481,
0.0605787002,
0.0555654503,
-0.0089662131,
0.002332123,
-0.0133356294,
0.0644702539,
-0.0932198986,
0.0428493954,
-0.0199377853,
0.0087129138,
0.0183069501,
0.0490108021,
0.0484973602,
0.0068764458,
-0.0664338097,
-0.0021569503,
-0.0427058898,
0.077988863,
0.029906841,
-0.0093440767,
0.0019746164,
-0.0100455312,
0.003083982,
-0.0114475023,
-0.0651326105,
-0.0429544412,
-0.0134871164,
0.0293485429,
-0.0522428118,
0.0062216832,
-0.0049455143,
0.041543182,
0.012155558,
0.0409097895,
0.0216318462,
0.037985675,
0.0003227583,
-0.0212320648,
0.0243218802,
0.0072447951,
-0.0508497469,
0.0519444384,
0.0279128123,
0.014312665,
-0.0047628549,
-0.0302753914,
0.0467513762,
-0.0073050912,
0.0077585983,
0.0232267454,
-0.0023672369,
0.0008069593,
0.0218527056,
0.0293805543,
-0.0392749645,
-0.005691634,
0.0319169834,
0.0300915968,
-0.0479648039,
0.0302057546,
0.009619127,
0.0394720025,
0.0764328465,
-0.0122191999,
0.0146720344,
0.034499187,
-0.0374950171,
0.0065350942,
0.0559154041,
0.0083616842,
0.0206511039,
2.495e-7,
-0.0303614121,
-0.0111313239,
0.0499083735,
-0.0424280241,
0.0390064158,
-0.0149547867,
-0.0652721822,
-0.0340685397,
0.0253292155,
0.0390669703,
0.010404597,
-0.0781543031,
0.0860219449,
0.0185480583,
-0.0677414462,
-0.0322262608,
-0.0502647273,
-0.021679664,
-0.0288025066,
0.0593654476,
0.0085246852,
0.0092057027,
-0.0295347646,
-0.0192997269,
-0.0156193916,
-0.0175652355,
-0.03641247,
-0.0271008462,
0.0127831567,
-0.013095296,
0.040975932,
-0.0515578426,
0.0137668587,
0.0751537755,
-0.0430397578,
0.0647124648,
0.0555175766,
-0.0046459828,
-0.0094966237,
-0.0217275303,
0.0481855758,
0.0023459899,
-0.0486074612,
-0.0391575061,
0.0492712371,
0.0392411985,
-0.0011429035,
-0.0336343087,
-0.0095176259,
0.0004784337,
0.0096309138,
-0.0165925305,
0.0099503696,
0.0092983851,
0.0229766592,
-0.0336663388,
-0.0104977824,
0.0054440247,
-0.004853772,
0.0197957531,
-0.0444717221,
-0.0468016639,
0.0125794662,
-0.0251678079,
0.0575493313,
-0.0174885541,
-0.0143992184,
1.100413897e-34,
0.0159149077,
-0.0382779092,
0.0054738428,
0.0039631096,
-0.0289249402,
-0.0172493812,
0.0099550113,
0.0543973073,
-0.0035107168,
0.0092162788,
-0.0539399907
] |
http://arxiv.org/abs/2401.14887v4 | http://arxiv.org/pdf/2401.14887v4 | The Power of Noise: Redefining Retrieval for RAG Systems | 2024-01-26T14:14:59 | 2024-05-01T08:15:07 | Florin Cuconasu, Giovanni Trappolini, Federico Siciliano, Simone Filice, Cesare Campagnano, Yoelle Maarek, Nicola Tonellotto, Fabrizio Silvestri | cs.IR | cs.IR, cs.CL | Retrieval-Augmented Generation (RAG) has recently emerged as a method to
extend beyond the pre-trained knowledge of Large Language Models by augmenting
the original prompt with relevant passages or documents retrieved by an
Information Retrieval (IR) system. RAG has become increasingly important for
Generative AI solutions, especially in enterprise settings or in any domain in
which knowledge is constantly refreshed and cannot be memorized in the LLM. We
argue here that the retrieval component of RAG systems, be it dense or sparse,
deserves increased attention from the research community, and accordingly, we
conduct the first comprehensive and systematic examination of the retrieval
strategy of RAG systems. We focus, in particular, on the type of passages IR
systems within a RAG solution should retrieve. Our analysis considers multiple
factors, such as the relevance of the passages included in the prompt context,
their position, and their number. One counter-intuitive finding of this work is
that the retriever's highest-scoring documents that are not directly relevant
to the query (e.g., do not contain the answer) negatively impact the
effectiveness of the LLM. Even more surprising, we discovered that adding
random documents in the prompt improves the LLM accuracy by up to 35%. These
results highlight the need to investigate the appropriate strategies when
integrating retrieval with LLMs, thereby laying the groundwork for future
research in this area. | The Power of Noise: Redefining Retrieval for RAG Systems: Retrieval-Augmented Generation (RAG) has recently emerged as a method to
extend beyond the pre-trained knowledge of Large Language Models by augmenting
the original prompt with relevant passages or documents retrieved by an
Information Retrieval (IR) system. RAG has become increasingly important for
Generative AI solutions, especially in enterprise settings or in any domain in
which knowledge is constantly refreshed and cannot be memorized in the LLM. We
argue here that the retrieval component of RAG systems, be it dense or sparse,
deserves increased attention from the research community, and accordingly, we
conduct the first comprehensive and systematic examination of the retrieval
strategy of RAG systems. We focus, in particular, on the type of passages IR
systems within a RAG solution should retrieve. Our analysis considers multiple
factors, such as the relevance of the passages included in the prompt context,
their position, and their number. One counter-intuitive finding of this work is
that the retriever's highest-scoring documents that are not directly relevant
to the query (e.g., do not contain the answer) negatively impact the
effectiveness of the LLM. Even more surprising, we discovered that adding
random documents in the prompt improves the LLM accuracy by up to 35%. These
results highlight the need to investigate the appropriate strategies when
integrating retrieval with LLMs, thereby laying the groundwork for future
research in this area. | [
0.0651410818,
-0.0000159664,
-0.0076270439,
-0.0085861664,
-0.0220074076,
-0.0221548192,
-0.0239053052,
0.0245113522,
0.0190668423,
-0.0359438695,
-0.031323351,
-0.0296261348,
-0.0224490333,
-0.0133592151,
0.0101402588,
-0.0090507539,
0.0544778816,
-0.012224596,
0.0095342174,
0.002025729,
-0.0172302127,
0.0112297628,
0.0264946949,
0.0101741571,
-0.01830923,
0.0265863612,
-0.0204877406,
0.0388220474,
0.0049276045,
-0.0360945873,
0.0003022148,
0.0040416848,
0.0359914005,
0.0386371091,
0.000001922,
-0.0706487074,
-0.0280186106,
0.0089310398,
-0.0608255975,
0.0034086874,
0.0648579299,
0.0669614822,
-0.0041123829,
0.0008915027,
-0.0443112552,
-0.0130697154,
0.0172311962,
0.0436319709,
0.0045234193,
0.0839779899,
-0.0095623573,
0.0101615861,
0.0332125612,
-0.0303661376,
0.1099242344,
-0.0637654737,
0.0071793902,
0.0029786164,
0.0459836312,
0.0546394885,
0.0244550183,
0.0477702059,
0.007739068,
-0.0101986546,
0.0213992968,
0.030802574,
-0.0483026542,
0.0060224589,
-0.0219014771,
0.0378254764,
0.0479590483,
-0.022582341,
-0.002919453,
0.0087144207,
0.0282489154,
0.0641504154,
-0.0561462604,
0.0324950963,
-0.0260522906,
0.0095482999,
0.000939926,
0.0061750417,
0.0248910505,
0.0026727412,
-0.0134030106,
0.0413309895,
0.0168796666,
-0.0524524041,
0.0243740976,
0.0104008038,
0.0031940888,
-0.0367645845,
-0.0153383184,
0.0523012355,
-0.0185009632,
0.0240813158,
-0.0079799425,
-0.009981663,
0.0526400581,
0.010459925,
0.0338765383,
0.0133899329,
-0.0330896862,
0.0594909936,
-0.0154220406,
0.0113818832,
-0.0096246712,
-0.0242577121,
-0.0535101667,
-0.0208856612,
-0.0327774994,
-0.0358265825,
-0.0369436592,
0.0679626763,
-0.0597717054,
-0.023260491,
-0.0235273354,
0.0279324017,
0.0124768298,
0.001712146,
-0.011805784,
0.0437421501,
-0.0689939409,
-0.0062660351,
-0.0622810908,
-0.0106724389,
-0.0123656029,
0.0057192892,
0.0326005481,
-0.0554991849,
-0.0083622979,
0.0022502707,
0.0098164054,
-0.0165351965,
0.012745291,
0.0348581038,
-0.0157341063,
-0.0559716448,
-0.069388561,
-0.0592355393,
-0.0066601885,
-0.041753158,
0.0058412366,
-0.0506667383,
0.0163497906,
0.0025812399,
-0.0176934842,
-0.0262133405,
0.0145068001,
0.0543823577,
-0.0201739613,
0.0116688861,
0.0097999135,
0.0184632875,
0.0649332255,
0.017368054,
0.0425463244,
0.0248953439,
-0.0012789047,
0.0664532334,
0.0610396378,
0.0203723889,
0.0468807146,
0.020093238,
0.0159575045,
0.0144518567,
-0.019651996,
0.0181805268,
-0.0027602674,
-0.0141539928,
-0.0111475028,
0.0585401542,
-0.0570166223,
0.0120209567,
0.0240679476,
0.035604652,
0.0188140888,
0.0111970361,
0.0398823358,
0.014346783,
-0.0089833839,
0.0397385806,
0.004253522,
0.0148230325,
-0.0605971105,
0.0196672585,
0.070445545,
-0.0171204247,
-0.0355560854,
-0.0572519861,
-0.0119573055,
0.0087837819,
-0.0311486293,
0.0345424041,
0.0448370501,
0.0444995165,
-0.0410284176,
0.0037527168,
0.0112249963,
-0.0733615533,
0.0058063883,
-0.0441533811,
0.0474169664,
0.0869876593,
0.109748736,
-0.0569269955,
0.0072159758,
0.0146105634,
-0.0185405724,
0.0213829856,
0.0540707931,
-0.0026833178,
-0.0194058679,
-0.0519553274,
-0.0433965586,
-0.0132891331,
-0.0042766961,
0.048888918,
0.0449235514,
0.0012275787,
0.0314729139,
-0.031623736,
-0.0519066751,
0.0150914509,
-0.0407942161,
-0.0680381432,
0.0298948865,
-0.0080675436,
0.0072299577,
0.0294711329,
-0.0172534101,
0.0183396842,
0.0577723123,
-0.0129178008,
-0.0304861683,
-0.035657838,
-0.0209015664,
0.0272743534,
-0.0060510179,
0.0160483308,
0.0115965689,
-0.0085014403,
-0.0018019467,
-0.0357709303,
-0.0257954113,
-0.0412596315,
0.0099531952,
-0.0836987719,
-0.0076071243,
0.0332371145,
0.023036737,
-0.0748622492,
0.0735876858,
0.0161325894,
0.0166413244,
0.0503282174,
0.0342184156,
0.0361422785,
-0.0755850151,
-0.0181660671,
-0.0017607585,
-0.0051615136,
0.0223986097,
0.071042262,
-0.01871299,
-0.0173071939,
-0.0308595039,
-0.0955056101,
-0.003552357,
-0.0110445945,
-0.0301340632,
-0.0117649632,
0.0316594839,
-0.0232508592,
-0.023984259,
-0.0331458785,
-0.0153383948,
0.0631387085,
0.0303123761,
0.0193346236,
0.0297674239,
0.0280067623,
-0.0439108945,
-0.0106342174,
0.0246852096,
0.0232008789,
0.0203579478,
0.0126985162,
-0.070853211,
0.0742886961,
-0.0356350355,
0.017717801,
-0.0117242672,
0.0155831408,
0.0194076393,
-0.0693616271,
0.0517332032,
0.0399069078,
0.0239447057,
0.0190528836,
0.0306642875,
-0.0376631096,
-0.097037077,
0.0308843609,
-0.0435961746,
-0.0407324843,
-0.0055375462,
-0.0302188676,
0.0338014551,
-0.0236424413,
0.0130989337,
0.0041232063,
-0.0113372523,
0.0116510354,
0.0154414335,
-0.0533424094,
-0.0482170433,
-0.0201950334,
-0.0097122695,
0.0699740723,
-0.0036468669,
-0.0523597933,
0.0441481806,
-0.0165267531,
0.0021763705,
-0.0401328988,
0.0039437222,
-0.0363510549,
0.0503847636,
-0.0282013658,
0.0136127472,
-0.0383863226,
-0.0215677451,
-0.0431410931,
-0.0029946016,
0.0016572248,
-0.0620562993,
0.0247500073,
-0.005763601,
0.0130438656,
-0.0022977849,
-0.0332525335,
-0.0790392905,
-0.0379569121,
0.0491569862,
-0.0046317317,
-0.0162346084,
-0.0111560933,
-0.0575103685,
-0.0217992831,
0.0603692755,
0.0093678581,
-0.0196548514,
-0.0441024825,
0.0006657821,
0.0370716974,
0.0711802319,
0.008478567,
0.0123152956,
0.0028790256,
-0.0308042821,
0.0310975742,
0.0421154536,
-0.0297397692,
0.0156741012,
-0.0149816955,
0.0504738837,
-0.0219034776,
-0.0107336128,
0.0289929248,
0.0103708003,
-0.0261884257,
-0.0462141968,
0.0737447664,
0.0338901617,
-0.0223344378,
-0.0430475883,
-0.039164912,
0.0303058829,
0.0236352533,
0.0816004649,
-0.0557148084,
0.0105384849,
0.0081031537,
0.0711239055,
-0.0220152587,
-0.0000137478,
-0.0172658358,
-0.1312138885,
0.0755086988,
0.0050187157,
0.0070772977,
-0.0324799456,
-0.0345131829,
-0.0192931648,
0.0484664664,
0.0038475785,
0.0398658887,
-0.0274582952,
-0.0027183515,
0.0130266985,
0.0293960385,
-0.0225460511,
-0.0340471305,
0.0016912244,
0.0116599835,
0.0782109573,
-0.0041402522,
0.0020333673,
-0.0390449464,
0.0110146543,
-0.0052072834,
0.0405527428,
0.0069246809,
0.0510555021,
0.0066366917,
0.0225387067,
-0.0199360419,
-0.013282218,
0.011618129,
-0.0038301379,
0.0052607548,
0.0315291658,
-0.0512056425,
0.0134095792,
-0.0478614345,
-0.0072485562,
0.0251145549,
0.0879715458,
-0.0106820501,
0.0058994796,
-0.0162153207,
-0.040433418,
0.0558748208,
-0.0125729637,
-0.0188494809,
0.0338150486,
-0.0399459526,
-0.0371078067,
-0.0816093981,
-0.0679786205,
-0.0303591471,
0.0382069387,
0.0210796911,
0.026829103,
-0.0444588475,
0.0008040592,
-0.0288623329,
-0.014263385,
-0.060499955,
-0.0328881964,
-0.0296677127,
-0.0219835006,
0.0185540766,
-0.1034271196,
-0.0042812359,
0.0088139819,
-0.0366845839,
-0.0452745333,
0.0176718943,
0.0257102009,
0.0253561717,
0.0328482836,
0.0095632831,
-0.0437255204,
0.0976955071,
0.0596138947,
-0.0549516343,
0.0340774879,
-0.0024853074,
0.0188405663,
-0.0292931329,
0.0028959014,
-0.0970260426,
0.0301727559,
0.0148856286,
0.0351378694,
-0.0002826937,
-0.0457412191,
0.0154715469,
0.0373343602,
0.0175701864,
0.0122111961,
-0.0031196049,
-0.0381039828,
0.0295582097,
-0.0060093859,
-0.0178155079,
-0.0116836186,
-0.0002430336,
0.0169449765,
-0.0114498641,
0.0063932394,
0.0405886658,
-0.0293687601,
0.1113884747,
-0.041557841,
-0.054463543,
0.0264897794,
0.0249751173,
0.0085703433,
-0.004264419,
-0.0286737662,
0.0435876399,
-0.0013425255,
0.0703441575,
-0.0095474618,
0.0562921911,
0.0112802461,
-0.0291066635,
-0.0081979204,
0.0089367023,
-0.0905698985,
-0.0368794799,
-0.0001516745,
0.0329531804,
0.0232587475,
-0.0081516979,
0.0162803009,
-0.0405618213,
0.0096653532,
-0.040468175,
-0.0362317897,
0.0498659462,
-0.0228516553,
-0.0101387696,
0.0334560275,
-0.0134940092,
0.052979134,
0.0465573184,
-0.0243772287,
0.0316702686,
-0.0561115593,
-0.0375955328,
-0.042270422,
-0.0526773296,
0.0204810016,
0.0367469825,
-0.0075660348,
-0.0196152106,
-0.0584714226,
-0.0269123968,
0.0297738891,
0.0260473881,
-0.0120858978,
-0.0271706861,
0.0385235474,
0.0954217985,
-0.0108692963,
0.0083405916,
-0.0451171026,
0.0235298593,
-0.0698774531,
-0.0139223207,
-0.0329927839,
-6.105584147e-33,
-0.0238065664,
-0.0307747573,
0.0509763323,
-0.0246671457,
-0.0081071826,
-0.0033385854,
-0.0230099186,
-0.0197559353,
0.0052453801,
-0.0156674702,
-0.0474290363,
-0.001535655,
0.0196804944,
0.0178389158,
-0.0132395383,
0.0057006851,
0.0558140427,
0.0084524229,
-0.0380946882,
0.0386550762,
0.087187618,
0.0332887471,
0.0102229193,
-0.0570946038,
0.0289760064,
-0.0437458679,
0.0241250843,
-0.00269815,
0.0091269361,
0.0561698452,
-0.0199525636,
0.0545411855,
0.0071097468,
-0.0218386836,
0.0034629742,
0.035282433,
-0.0666646436,
-0.0425878353,
-0.0049713142,
-0.0302037597,
-0.0107385945,
-0.0383202583,
0.0265587624,
-0.0343565084,
-0.0258162227,
0.0485106483,
0.0348546244,
-0.0178920776,
-0.0337311029,
-0.063156642,
-0.0299058147,
0.0058723977,
0.0299665108,
0.0004570977,
-0.0115564158,
0.0453622751,
0.0319590233,
0.0137964962,
-0.0510541052,
0.0374738425,
-0.0243801605,
0.0608725622,
0.0241423976,
0.0030947335,
-0.0188316088,
0.0389650837,
0.0350814387,
0.0061084465,
-0.0130527159,
-0.0342997834,
0.0022481165,
0.0230191313,
-0.014016483,
0.0604970157,
0.0421614461,
-0.0086719058,
-0.0409968831,
0.0865491107,
0.0024012367,
0.0338393301,
0.0456984378,
-0.0018766546,
-0.0686031953,
0.0007173793,
-0.0006892339,
0.0259678792,
-0.0362458825,
-0.0336299799,
0.0591765232,
0.0462819412,
-0.0149028171,
-0.0295409076,
0.0469662435,
0.0007048915,
-0.0110174213,
0.0342561938,
-0.01584333,
-0.0098433187,
-0.0220818315,
-0.0157063063,
-0.0148474015,
-0.0217401683,
0.0677841976,
-0.0082961852,
-0.0498244241,
0.0171773657,
0.008088802,
0.0336166434,
-0.0629209876,
0.0146875838,
-0.016527053,
0.0618773326,
0.0027722812,
-0.0354286879,
0.0352621712,
0.0268900748,
0.00629625,
0.0246559624,
0.0056900256,
0.0099195922,
-0.0444142967,
-0.0394200273,
-0.0002475432,
-0.0218699705,
-0.0131787071,
0.0541901924,
-0.0047725532,
0.0437106825,
0.0018379391,
-0.0127429599,
0.0137783382,
0.0447946191,
2.7e-7,
0.0153911673,
0.0007738253,
0.0496137701,
0.01417882,
-0.014895346,
-0.0018296429,
0.0021336551,
-0.000210891,
0.0334608667,
-0.0001868268,
0.0054160645,
0.0229206011,
0.0300638322,
0.0160321612,
-0.1184404939,
-0.0380556472,
-0.0239886139,
-0.0127642099,
-0.0460112877,
-0.0092587136,
0.0470632054,
0.048073601,
0.0012796524,
-0.0394776203,
-0.015701687,
-0.0549924895,
0.0598482862,
-0.0144952321,
-0.0238564946,
0.0245436579,
0.0724140555,
0.0367878117,
-0.0110532809,
-0.0297757424,
-0.020323094,
0.0300522074,
-0.0023291882,
0.0133127579,
0.0148270838,
-0.0992733762,
-0.016969176,
0.0106410906,
0.0068657571,
-0.019730689,
0.053908769,
-0.0001837059,
0.0279406868,
0.0126858633,
-0.0279729161,
-0.0110187046,
-0.0113957906,
-0.0167543162,
-0.0245534703,
0.0009230083,
-0.0118331807,
0.0182394739,
0.0002621533,
-0.0710065365,
0.0185178965,
-0.0217217002,
-0.0398977399,
-0.0505107939,
-0.0244209524,
0.0176666081,
-0.0413998142,
-0.0586779676,
-0.0090424875,
2.62690271e-34,
0.037647713,
0.0080904998,
-0.0095141483,
0.0039975471,
0.0260104425,
0.0085412413,
-0.0374567099,
0.0385894403,
0.0042345906,
-0.0975639224,
-0.0439449288
] |
http://arxiv.org/abs/2401.14869v1 | http://arxiv.org/pdf/2401.14869v1 | F-Eval: Asssessing Fundamental Abilities with Refined Evaluation Methods | 2024-01-26T13:55:32 | 2024-01-26T13:55:32 | Yu Sun, Keyu Chen, Shujie Wang, Qipeng Guo, Hang Yan, Xipeng Qiu, Xuanjing Huang, Dahua Lin | cs.CL | cs.CL | Large language models (LLMs) garner significant attention for their
unprecedented performance, leading to an increasing number of researches
evaluating LLMs. However, these evaluation benchmarks are limited to assessing
the instruction-following capabilities, overlooking the fundamental abilities
that emerge during the pre-training stage. Previous subjective evaluation
methods mainly reply on scoring by API models. However, in the absence of
references, large models have shown limited ability to discern subtle
differences. To bridge the gap, we propose F-Eval, a bilingual evaluation
benchmark to evaluate the fundamental abilities, including expression,
commonsense and logic. The tasks in F-Eval include multi-choice objective
tasks, open-ended objective tasks, reference-based subjective tasks and
reference-free subjective tasks. For reference-free subjective tasks, we devise
new evaluation methods, serving as alternatives to scoring by API models. We
conduct evaluations on 13 advanced LLMs. Results show that our evaluation
methods show higher correlation coefficients and larger distinction than other
evaluators. Additionally, we discuss the influence of different model sizes,
dimensions, and normalization methods. We anticipate that F-Eval will
facilitate the study of LLMs' fundamental abilities. | F-Eval: Asssessing Fundamental Abilities with Refined Evaluation Methods: Large language models (LLMs) garner significant attention for their
unprecedented performance, leading to an increasing number of researches
evaluating LLMs. However, these evaluation benchmarks are limited to assessing
the instruction-following capabilities, overlooking the fundamental abilities
that emerge during the pre-training stage. Previous subjective evaluation
methods mainly reply on scoring by API models. However, in the absence of
references, large models have shown limited ability to discern subtle
differences. To bridge the gap, we propose F-Eval, a bilingual evaluation
benchmark to evaluate the fundamental abilities, including expression,
commonsense and logic. The tasks in F-Eval include multi-choice objective
tasks, open-ended objective tasks, reference-based subjective tasks and
reference-free subjective tasks. For reference-free subjective tasks, we devise
new evaluation methods, serving as alternatives to scoring by API models. We
conduct evaluations on 13 advanced LLMs. Results show that our evaluation
methods show higher correlation coefficients and larger distinction than other
evaluators. Additionally, we discuss the influence of different model sizes,
dimensions, and normalization methods. We anticipate that F-Eval will
facilitate the study of LLMs' fundamental abilities. | [
0.0024476363,
0.0310013909,
-0.0124951657,
0.0208098441,
-0.0446028896,
0.0360120684,
0.0231395252,
-0.0334324427,
0.0219252575,
-0.043784488,
-0.0816948265,
-0.0304770991,
0.0016139031,
-0.0288806763,
0.0221004356,
-0.0225410685,
0.0650644675,
-0.0379015915,
0.0124411238,
-0.0106261466,
-0.0341171809,
-0.003642193,
-0.0111760283,
0.026274588,
0.0519386157,
-0.004717025,
0.0276702866,
-0.0406037159,
0.0168580879,
0.0053903935,
-0.007904591,
0.0525256582,
0.002732865,
-0.0264504515,
0.0000021629,
-0.0608421043,
0.0426128507,
-0.0004088461,
0.0082758963,
0.0235681161,
0.0570216253,
0.0256221388,
-0.0207071286,
0.0067095403,
-0.0341887735,
-0.0529477969,
0.0122027667,
0.0234735254,
0.0406732634,
0.0453663915,
0.0175147522,
-0.0237611327,
0.0314596929,
-0.0130954739,
-0.0108344862,
-0.0514247268,
-0.0038464735,
0.0144800814,
-0.0005237825,
0.000076746,
-0.0063245208,
0.0385158397,
0.0577074289,
-0.0028780927,
-0.0191771127,
-0.011154023,
-0.0606336482,
-0.0092179216,
-0.0021057725,
0.0371835977,
-0.0053637717,
0.0076342011,
-0.0357129723,
0.0234274864,
0.0040018591,
-0.0085174767,
-0.0415752716,
0.0249136239,
0.0316917934,
0.0124731306,
0.0645054355,
0.0118007036,
0.0104400059,
0.0168400314,
-0.0050299913,
0.0833178833,
0.0112519888,
-0.0426419638,
0.0499480926,
0.0154112112,
-0.0022410892,
-0.0292356163,
-0.0033528672,
0.0194246266,
-0.0369080529,
0.0106614083,
-0.0184925757,
-0.0009264133,
0.015772447,
-0.0748932362,
0.0633855239,
0.0326455422,
-0.0248015001,
0.0377949402,
-0.0354533754,
0.1188978478,
0.0263795685,
0.0538903698,
-0.0515233725,
0.01057084,
-0.0662621036,
-0.0020537018,
0.0005570072,
0.0321055017,
-0.0303628538,
-0.0111021893,
-0.0567177385,
-0.0145317819,
-0.0246975552,
0.0038730165,
0.0232617185,
-0.022974059,
0.0084135151,
0.0146446051,
-0.0548340753,
0.0359458774,
-0.0238812212,
0.0056473063,
-0.0076452219,
-0.0398087278,
0.0318010934,
0.0328449085,
-0.0182191748,
-0.0287148636,
0.029574709,
0.069897972,
-0.0279397406,
-0.0204173513,
-0.0354190245,
0.0026248912,
-0.0040342682,
-0.1247311831,
-0.0032815281,
-0.0826266333,
0.0307280105,
0.0141305383,
0.0004309722,
-0.0138008585,
-0.0006161986,
0.0161168184,
0.0069394466,
-0.0092545627,
0.0705421939,
0.0337559953,
0.0033475144,
0.0053393147,
-0.0001814345,
0.0695353299,
-0.0106577622,
-0.0306038279,
0.0221488271,
-0.0278256163,
-0.022482818,
0.045720268,
-0.004151511,
0.0055370578,
-0.0084803049,
0.0177111793,
0.0180556588,
0.0169703253,
0.0529504046,
0.0533176661,
-0.0273111444,
-0.0071131717,
0.050139945,
0.0338120386,
-0.0001066659,
-0.0048190923,
0.0948776081,
0.0304274298,
0.0663211867,
0.0245876014,
-0.0627999976,
0.0152440872,
-0.0251112692,
0.0010056241,
0.0009383213,
0.0459361002,
-0.0407577232,
-0.047973305,
0.0103633245,
0.0022017062,
0.0360635482,
0.0058543989,
0.0282021575,
-0.0089876968,
-0.0132151218,
0.0391485132,
0.010779365,
-0.0572879128,
0.0303904545,
-0.0403774902,
-0.0261387806,
0.0404364057,
-0.001369738,
0.0420704186,
-0.0305799898,
0.0154692493,
0.0015580474,
0.046440132,
0.0638668165,
-0.0068531539,
0.0401786193,
-0.0407707244,
-0.0043130461,
-0.0246181842,
0.0048797629,
0.0000237068,
-0.0242935475,
0.0053368462,
0.0049174605,
-0.0612992048,
-0.0690592527,
0.0146362716,
-0.0206575096,
-0.0262866504,
-0.0155336978,
0.0049634622,
0.0322503261,
0.0338653587,
-0.00442388,
0.0144626563,
0.0232910365,
-0.0145522365,
0.0147154313,
0.0331057049,
-0.0494983457,
-0.0229612961,
0.0450886376,
-0.0081252232,
-0.0467738807,
0.032271307,
-0.0065857908,
-0.0418248512,
0.0267699603,
-0.0214530695,
-0.0086974036,
-0.029729018,
0.053853102,
-0.0437222831,
-0.0327115767,
-0.01097578,
0.0798547417,
-0.037433207,
0.0701684877,
0.0046271854,
-0.0047199572,
0.0333505198,
-0.0506684817,
-0.0671420097,
-0.0079099797,
-0.024988031,
0.0303827226,
-0.0028378249,
0.0554675758,
-0.0347557142,
-0.0493707433,
-0.1025909558,
-0.001089007,
-0.0222779121,
0.0465574302,
0.0322397351,
-0.0557865761,
0.0118040824,
0.0393521823,
-0.0587284975,
-0.0285919011,
0.0269196443,
0.0211998709,
-0.0035684605,
0.0072267591,
-0.0324377008,
-0.0076864604,
-0.0182948131,
-0.0152728939,
0.0105396556,
-0.0020751688,
0.0655373409,
0.0009175175,
0.0275925938,
-0.0292801633,
0.0144172292,
0.0083550671,
-0.0652640238,
0.0294989683,
0.0432663634,
0.0416552536,
0.0944900736,
0.0011719742,
-0.0137294428,
0.0294447895,
-0.029606102,
-0.035483826,
0.0177485906,
-0.0216593891,
-0.0190894231,
-0.0796219409,
-0.0601988286,
-0.0220726449,
0.0612914898,
0.0177372396,
0.0183859207,
-0.0650678203,
-0.0064640013,
-0.0557065606,
-0.0556203276,
-0.0392157137,
0.0257770251,
-0.0203299262,
0.0306901708,
-0.0402326807,
-0.0240945928,
0.0576763712,
-0.0375304185,
-0.0197622385,
-0.0828827396,
0.0028811453,
-0.0411961935,
0.0467366315,
-0.021323422,
0.0079483287,
-0.0455906317,
-0.0023091969,
-0.0481160954,
0.0176971257,
0.0023429161,
-0.062144652,
0.057352934,
-0.0426268317,
-0.0308916848,
0.0265311021,
-0.0305200946,
-0.0172740538,
0.0239108447,
0.0097375093,
0.0586599223,
0.0096286312,
-0.0110802585,
-0.0623139478,
-0.026027292,
0.0498531796,
0.0041601905,
-0.0434472114,
0.0261944905,
-0.006632111,
-0.0229234025,
0.018064402,
0.0052519008,
0.0025612321,
-0.04610461,
0.0109005533,
0.007410964,
0.0284586046,
0.0079523325,
0.000180939,
0.0264556296,
0.0274272282,
0.0036479738,
-0.0431215912,
0.0246849153,
0.0731313154,
-0.0417547822,
-0.0030865797,
0.1072403714,
0.0294025447,
0.0027033531,
0.0018492313,
-0.0292544775,
-0.0152200609,
0.0364012718,
0.0249078758,
-0.0382555276,
-0.0079733413,
-0.0191123765,
0.0165845398,
-0.0197489932,
-0.0088603366,
0.0132247005,
-0.1347533017,
0.0161226746,
0.0196585562,
-0.0682412684,
-0.0241623931,
-0.0172245614,
0.0073877051,
-0.0268127061,
0.0348398425,
0.0202470943,
0.0043343506,
-0.0163197629,
0.0188761298,
0.0000160403,
-0.0887370855,
0.0761337951,
-0.0419351384,
0.0489804745,
0.0172233675,
0.0168642532,
-0.0205253232,
-0.0752559677,
0.0204330552,
-0.0468948931,
0.0005048162,
0.0009176795,
0.010900625,
-0.0065466622,
-0.0142165264,
-0.0444702767,
-0.0117168436,
0.0030267083,
0.0200092457,
-0.0026791659,
0.0852078274,
-0.0086046066,
0.0076821251,
0.0141867986,
0.0125549762,
-0.0035103727,
0.0759990886,
0.0501215048,
-0.0678232461,
-0.0158206671,
-0.0104182865,
-0.0557758212,
-0.0652810782,
-0.0519492999,
0.0498238131,
0.0058298348,
-0.0179847702,
-0.0537416413,
-0.0464950614,
0.0164081324,
0.0116652614,
-0.0080145309,
0.0079787085,
-0.0096493335,
-0.0259582438,
-0.0002068201,
0.0570618026,
0.0075831166,
0.0216588285,
-0.0948527604,
-0.0123086432,
0.012915046,
-0.0172300674,
-0.0455031581,
0.0291253179,
-0.0269282609,
-0.0508122705,
0.0030613316,
0.0672213659,
0.01715458,
-0.0217179451,
-0.0262559094,
-0.0627959147,
-0.016240539,
0.0214233827,
-0.0355303697,
0.0476094484,
0.0358037949,
-0.0128396917,
0.0114821671,
0.0236569233,
-0.0067551928,
0.0375798419,
0.0165746398,
0.0343560688,
0.010166619,
-0.0260511953,
0.0054943929,
0.0272333827,
0.0124201914,
-0.0022758306,
0.0156874359,
-0.0184405353,
-0.0164930671,
-0.0311484318,
-0.0314575881,
-0.0410561115,
-0.0193244293,
0.0320591964,
0.0173028987,
0.0203496795,
0.0031026904,
-0.0031340932,
0.0209632479,
-0.0109899705,
0.0452910513,
-0.012842875,
0.019445112,
-0.0104459478,
0.0389655717,
-0.0412065759,
-0.0292297509,
-0.0098098256,
0.0430055223,
-0.00228349,
0.0766262338,
-0.0257451124,
0.0219308082,
0.020839287,
0.0092941094,
-0.0126373591,
-0.023205474,
0.0192975197,
0.0277970191,
0.0518026203,
-0.0081694657,
0.0948737636,
-0.0094297687,
0.0030975412,
-0.0579452217,
-0.0042109145,
0.000972299,
-0.022859538,
0.0418373533,
0.021172395,
0.0187726356,
-0.027207287,
0.0384799615,
-0.0314077698,
-0.0014379696,
-0.0410772711,
0.0315079764,
-0.0436182171,
-0.0423157588,
0.0044998005,
-0.0077195358,
-0.032031782,
0.0117514525,
-0.0683015212,
-0.0699098334,
-0.0219137259,
0.056860026,
0.0037018494,
-0.0236003883,
0.0116398465,
0.078805387,
-0.0449931063,
-0.0455596857,
-0.0462706201,
-0.0207206495,
-0.0212589726,
-0.0032500918,
-0.0363471694,
-6.559533249e-33,
-0.0399612859,
-0.0005561846,
0.0354046114,
-0.0351164192,
0.0063014482,
0.0165368319,
-0.0236717872,
0.0408660807,
-0.0382345207,
0.0015226733,
0.022146238,
-0.0338138305,
0.0107836528,
0.0035577829,
-0.0012325961,
0.005804948,
-0.0038427862,
-0.0054597878,
-0.0276757516,
0.0142786345,
0.0301475767,
-0.0463581942,
0.0732636228,
-0.0744505525,
0.036755003,
-0.013408497,
0.0005101805,
0.0205238517,
-0.0322900899,
0.0966791809,
0.0191569552,
0.0435568132,
0.0268856864,
-0.0442554206,
-0.0144157307,
-0.014636891,
-0.0410196297,
-0.0429266691,
0.0942316949,
0.021712359,
-0.0446292721,
-0.0431320183,
0.0294930469,
0.0000774076,
-0.0661223531,
-0.0167499185,
0.0164802726,
-0.0178531241,
0.0231056754,
0.0205734335,
-0.0494349338,
0.0267403666,
0.0088634519,
-0.0073308968,
-0.0021538758,
0.0709528252,
-0.0532734245,
-0.0319492929,
-0.0608410463,
0.0216817334,
-0.0010787699,
0.0433047228,
0.0873776972,
-0.0139733944,
0.0155752702,
0.0376131833,
0.0222345274,
0.0174245462,
0.0304598007,
-0.0112640206,
0.0815356597,
0.0468298756,
0.0584777705,
0.0354395732,
-0.0118028279,
0.014255479,
-0.0418073684,
0.0111053316,
0.0331100412,
0.059318468,
0.0288451314,
-0.0142028984,
-0.0376382098,
-0.0114521757,
-0.0731636658,
-0.0041464451,
-0.0268710554,
-0.0327098779,
0.0597787723,
-0.0058043371,
-0.0011662563,
-0.0412735306,
0.0686038509,
-0.014313261,
0.005687573,
0.0371399447,
-0.0339217037,
0.0666205287,
0.0024758566,
0.0340001695,
-0.1010617539,
-0.0352843665,
-0.0010008219,
0.0078627346,
0.032080885,
0.0007818436,
0.0112227276,
0.019176336,
-0.0389688797,
-0.0039465488,
-0.0360264443,
0.0065891454,
-0.0014364073,
0.0193691272,
0.037422765,
-0.0168083236,
-0.0208171308,
0.0529399961,
-0.0365577489,
-0.026189439,
-0.0406247005,
-0.0208683014,
0.016223738,
0.0438026339,
-0.0261360053,
0.0451492667,
-0.0279588681,
-0.0208148975,
-0.0166920647,
0.0003434688,
0.0113217654,
0.0586573109,
2.842e-7,
0.0408637859,
0.0283092652,
-0.0228738766,
0.0501632653,
0.0091849007,
0.0229475964,
-0.043344412,
-0.0183851048,
0.0468599722,
0.0121726766,
-0.0298612993,
0.0607134923,
0.061164625,
0.0030834228,
-0.0615033358,
-0.0072518536,
-0.0243363678,
-0.0015091008,
-0.0190791767,
0.0192608889,
0.0461496748,
0.045927044,
-0.0033859734,
0.0063779308,
-0.0317486674,
-0.0553888083,
0.047445856,
-0.0023953028,
-0.0353597477,
0.008137376,
0.0024726808,
0.0323966816,
0.0179875698,
0.059684474,
-0.0022648363,
0.0401881561,
0.0020728335,
0.1206591725,
-0.002072013,
0.0151105123,
0.0417726487,
-0.0197141413,
-0.0311612561,
-0.0749679878,
0.0690867081,
-0.0543817058,
-0.0513642356,
-0.0075184228,
-0.0212451126,
0.0139398668,
0.011226113,
-0.0182989556,
-0.0383879915,
0.0308513418,
0.0362447388,
-0.0438380986,
-0.011504069,
0.0017333481,
-0.0285157803,
0.0471351929,
-0.0102817845,
-0.0825521201,
-0.0428471081,
0.004031688,
-0.0012408776,
-0.0418016016,
-0.0053718612,
2.518071219e-34,
0.0197356828,
-0.0128277624,
0.0551755428,
-0.0330199711,
0.0232259929,
0.0185286496,
-0.0567909889,
0.0021325725,
-0.0172539819,
-0.0469589271,
0.0143253775
] |
http://arxiv.org/abs/2401.15127v3 | http://arxiv.org/pdf/2401.15127v3 | Evaluation of LLM Chatbots for OSINT-based Cyber Threat Awareness | 2024-01-26T13:15:24 | 2024-04-19T09:40:04 | Samaneh Shafee, Alysson Bessani, Pedro M. Ferreira | cs.CR | cs.CR, cs.CL, cs.LG | Knowledge sharing about emerging threats is crucial in the rapidly advancing
field of cybersecurity and forms the foundation of Cyber Threat Intelligence
(CTI). In this context, Large Language Models are becoming increasingly
significant in the field of cybersecurity, presenting a wide range of
opportunities. This study surveys the performance of ChatGPT, GPT4all, Dolly,
Stanford Alpaca, Alpaca-LoRA, Falcon, and Vicuna chatbots in binary
classification and Named Entity Recognition (NER) tasks performed using Open
Source INTelligence (OSINT). We utilize well-established data collected in
previous research from Twitter to assess the competitiveness of these chatbots
when compared to specialized models trained for those tasks. In binary
classification experiments, Chatbot GPT-4 as a commercial model achieved an
acceptable F1 score of 0.94, and the open-source GPT4all model achieved an F1
score of 0.90. However, concerning cybersecurity entity recognition, all
evaluated chatbots have limitations and are less effective. This study
demonstrates the capability of chatbots for OSINT binary classification and
shows that they require further improvement in NER to effectively replace
specially trained models. Our results shed light on the limitations of the LLM
chatbots when compared to specialized models, and can help researchers improve
chatbots technology with the objective to reduce the required effort to
integrate machine learning in OSINT-based CTI tools. | Evaluation of LLM Chatbots for OSINT-based Cyber Threat Awareness: Knowledge sharing about emerging threats is crucial in the rapidly advancing
field of cybersecurity and forms the foundation of Cyber Threat Intelligence
(CTI). In this context, Large Language Models are becoming increasingly
significant in the field of cybersecurity, presenting a wide range of
opportunities. This study surveys the performance of ChatGPT, GPT4all, Dolly,
Stanford Alpaca, Alpaca-LoRA, Falcon, and Vicuna chatbots in binary
classification and Named Entity Recognition (NER) tasks performed using Open
Source INTelligence (OSINT). We utilize well-established data collected in
previous research from Twitter to assess the competitiveness of these chatbots
when compared to specialized models trained for those tasks. In binary
classification experiments, Chatbot GPT-4 as a commercial model achieved an
acceptable F1 score of 0.94, and the open-source GPT4all model achieved an F1
score of 0.90. However, concerning cybersecurity entity recognition, all
evaluated chatbots have limitations and are less effective. This study
demonstrates the capability of chatbots for OSINT binary classification and
shows that they require further improvement in NER to effectively replace
specially trained models. Our results shed light on the limitations of the LLM
chatbots when compared to specialized models, and can help researchers improve
chatbots technology with the objective to reduce the required effort to
integrate machine learning in OSINT-based CTI tools. | [
0.0498879664,
0.0213805493,
0.0077329767,
0.0073459181,
-0.055094935,
0.0154945934,
0.0373833738,
0.0240737479,
0.074294433,
-0.0505901948,
-0.0279272385,
0.0032267871,
-0.0420506448,
0.0403352194,
0.0088245636,
-0.0062784711,
0.0461712852,
-0.0126589676,
0.0220750179,
0.016136786,
-0.001505551,
0.0338466316,
-0.0636229739,
0.0626322702,
-0.0113039371,
0.01092193,
-0.0525000915,
-0.0203021523,
0.0641942993,
-0.027949227,
0.0091160638,
0.0323078334,
0.0141936867,
0.0623309836,
0.0000019547,
-0.0499954,
-0.0457011797,
-0.0050496235,
-0.0744283721,
-0.0580854043,
0.0559720024,
-0.019415969,
-0.0182681642,
0.0480255634,
-0.021257665,
0.0171387009,
0.0424559489,
0.0801512748,
0.0550356731,
0.102154173,
0.0065104393,
-0.0024741171,
0.0439188555,
0.0015002855,
-0.0149302687,
-0.0215304811,
0.0091242222,
-0.011213542,
-0.0040312777,
0.0697032735,
0.0328968391,
-0.0049782437,
-0.0016630279,
0.0168411974,
-0.0199604686,
0.0613940582,
-0.0660934225,
0.0102731464,
-0.0172105078,
0.0489966087,
0.0125142233,
0.0078427792,
0.0031056369,
0.008001606,
0.0258111339,
-0.0045649209,
-0.0230445452,
-0.0072930898,
-0.0215044823,
-0.0258389097,
0.0031700192,
0.0130593218,
0.0057151373,
0.0196219478,
0.0011233584,
0.0256707445,
-0.0172057487,
-0.0173124466,
0.0092665721,
-0.0584159121,
-0.0002521905,
-0.0501277223,
-0.0081274323,
0.0419509374,
-0.0141461426,
0.0067341602,
-0.0283568427,
-0.0202880148,
-0.0008396329,
-0.0603933446,
-0.0655554757,
0.0304062925,
0.0079298001,
0.0192654021,
-0.0283276699,
0.0639915764,
0.0230813827,
0.050739985,
0.0257743634,
-0.0313379094,
-0.1145929173,
-0.0287454687,
-0.0247677844,
0.073530972,
0.0473254099,
-0.0250732526,
-0.059577927,
0.0139721753,
-0.0285798423,
0.0183679536,
0.004676396,
-0.013837616,
0.0090277996,
0.0094468221,
-0.0379605182,
-0.0001222963,
-0.020903904,
0.0121415295,
-0.0154668568,
-0.0805319324,
-0.0128874071,
0.0249266829,
-0.0224170405,
-0.0473370589,
0.0026117202,
0.1008075103,
0.019426411,
-0.0545201041,
-0.0100672524,
-0.0020702023,
0.0109249679,
-0.0198253095,
0.0046649626,
-0.0246302634,
0.0132385716,
0.0161744468,
0.0095081395,
-0.0294483025,
0.0062814509,
-0.0179640278,
-0.0397490263,
0.0759486035,
-0.0511601828,
-0.0174339842,
0.0299435258,
0.0254176371,
-0.0031012951,
0.0261000525,
0.0325384438,
0.0642863289,
0.0060659489,
-0.0109309191,
0.0260598641,
0.0138825523,
-0.0154426983,
-0.01340036,
0.0256346911,
0.0102802729,
-0.0056078099,
0.0106008714,
0.0052187638,
0.0254711471,
0.0114697646,
-0.0256115813,
0.049810078,
0.0386783779,
-0.0135309743,
0.0247589387,
0.0050211297,
0.0554838553,
0.0369069539,
-0.0287745055,
-0.0406005643,
0.0295919143,
-0.0749072507,
-0.0429307818,
-0.0019840619,
0.0251128785,
-0.0307739917,
-0.0068888925,
-0.0074160546,
-0.0479588509,
0.0234920681,
-0.0192401502,
0.049417749,
-0.01389948,
-0.0404853784,
0.0200196598,
0.0317531899,
-0.0219485685,
-0.0447311364,
-0.0951164961,
-0.0155994967,
0.0749162287,
0.0484438092,
-0.0345336422,
-0.0719010755,
-0.0437760763,
-0.0456176437,
0.0597522557,
0.0079624457,
0.0156254284,
-0.0271037556,
-0.0239504427,
0.0036760801,
-0.0236780941,
0.0282824151,
0.0175884217,
0.025573248,
0.0306173488,
0.0222005565,
0.0121421106,
-0.0427538827,
0.0207742006,
0.0212223306,
-0.0145230442,
-0.0511877984,
0.0122717815,
0.0792399719,
0.0415142924,
0.0074178367,
0.0090168286,
0.0408130847,
-0.0129810283,
-0.0640794709,
0.0107336016,
0.0490655862,
0.0098962933,
0.0356551521,
-0.0101365177,
0.0026942454,
0.0739067197,
-0.0471893698,
-0.0094847782,
0.0487685911,
-0.0556818433,
0.0145130986,
-0.0201027859,
0.0007029551,
-0.0244705416,
0.0197551679,
0.0060095838,
0.0618801378,
0.0069625131,
-0.0336742736,
-0.0030422171,
0.0062522981,
0.0073471516,
-0.0226170421,
-0.0010456573,
-0.0006031885,
-0.0130472314,
-0.0141404383,
0.0113547221,
0.0141482512,
-0.0398333408,
-0.0300602019,
-0.1256976873,
-0.0190850925,
0.0060576932,
0.0294859372,
0.016601393,
-0.0239786692,
-0.032709036,
-0.0048315576,
-0.0267297942,
0.0093881749,
-0.006734171,
0.0184017308,
0.0509546101,
-0.0017117965,
-0.0606139712,
-0.0018095977,
0.0330222398,
-0.0172149129,
0.003722508,
0.05364573,
-0.0676974505,
0.0037730746,
0.0450115353,
-0.0743432269,
-0.0055987709,
-0.0343983136,
-0.062314488,
0.0549229756,
-0.0199155137,
0.0000747194,
0.0157023501,
0.0632342547,
0.0510660373,
0.0448219515,
0.0224770159,
-0.0639229938,
0.0318045542,
-0.0087075615,
-0.060000401,
-0.0287542362,
-0.037762437,
0.0061125779,
0.0504703932,
0.0005857273,
0.0351805128,
-0.0456102341,
0.0242305044,
0.0415655859,
-0.0211908724,
-0.0380438156,
-0.0266639665,
0.0033212299,
0.0132769495,
0.0131048821,
-0.0615023002,
0.0750806481,
-0.025066549,
0.0202630796,
-0.027202934,
-0.0683422685,
-0.0447826236,
0.0525432006,
-0.0148387952,
0.0246854108,
0.0327033959,
-0.0343167186,
-0.0173179209,
0.0041661812,
-0.0157054085,
-0.0399542078,
-0.0441652462,
0.0180595964,
0.0127827842,
0.0003814739,
0.0017326769,
-0.066933699,
-0.0235470179,
-0.083243385,
0.0023665382,
0.0119794998,
-0.0481851771,
-0.0268804133,
0.0541619658,
0.0578348339,
-0.0006706938,
0.028076468,
-0.0140413325,
0.0056566442,
-0.0283576045,
0.0057388581,
0.0075381785,
0.011770892,
0.0299566444,
0.0234023277,
0.0368903503,
0.0497149937,
-0.0256347507,
-0.0381547101,
-0.0376875065,
0.0042467052,
-0.0342843644,
-0.015952006,
-0.0016057206,
0.0311374776,
-0.0201394483,
0.0219714232,
0.1257338226,
0.0240966044,
0.003265528,
-0.0194340162,
-0.0283178091,
-0.0227280017,
-0.014801315,
-0.0035467867,
-0.0672640651,
-0.038920518,
0.022819452,
0.0224155728,
-0.052888643,
-0.0651182011,
-0.0124984607,
-0.0960948989,
0.0206374936,
-0.0133562796,
-0.0365153812,
-0.0517656058,
-0.0725851581,
0.0045809145,
-0.0100722238,
-0.0071365479,
-0.0367226563,
-0.0006662706,
-0.005533454,
0.0019009875,
-0.0567848124,
-0.00076292,
0.0345207714,
-0.0284531135,
0.0582082681,
0.0394902043,
0.034557797,
-0.0131017184,
0.0220530592,
0.0493939593,
0.001459631,
0.0154082617,
-0.0041525862,
0.0776323751,
-0.0306642186,
0.0163540971,
-0.0472260863,
0.0115813222,
0.010805645,
-0.023413891,
0.0333763696,
0.0510079786,
-0.0636659637,
0.0110438615,
-0.0179735627,
0.0306083951,
0.0012899135,
0.0896270648,
0.0109049287,
-0.0142005468,
-0.0054796552,
-0.0212265924,
0.0103041045,
0.0411627665,
-0.0390603878,
0.002119432,
-0.01366108,
-0.0667426288,
-0.0516797304,
-0.044058051,
0.0478187203,
-0.0208445415,
0.0027786829,
-0.0298445188,
-0.0106325531,
0.0465850048,
-0.0146459546,
0.0046437494,
-0.0350321047,
0.0322225131,
-0.0472676195,
-0.014104953,
0.0381641909,
-0.0174660068,
-0.0326612256,
0.0488563702,
-0.0469052717,
-0.042040918,
-0.0194451381,
0.0274465065,
0.0282866023,
-0.0181374978,
-0.0478177816,
-0.0073126107,
0.0844163746,
0.0619733781,
-0.0292349402,
-0.0289078504,
0.0473177247,
0.0483131558,
0.0187883135,
0.0257955231,
-0.0263797697,
-0.007295778,
0.0129757384,
-0.0276937913,
-0.0665052161,
-0.0518564656,
0.0310647786,
0.0579759628,
0.0370970629,
0.0033797757,
-0.0217184927,
-0.016106341,
0.0271271802,
0.0069071245,
0.0147924703,
0.0151500348,
-0.0017223137,
0.0115035549,
-0.0003365837,
0.034264572,
0.0591857359,
-0.0518837683,
0.0102015436,
-0.0212718938,
-0.0067662047,
0.042374637,
0.0577367656,
0.0223449524,
-0.0218331311,
-0.0088994214,
0.0184455123,
0.0065832296,
-0.0017564894,
0.0092170741,
0.0378847197,
-0.0355459303,
0.0034530989,
0.012807467,
0.0483700559,
-0.0727857724,
-0.0729401633,
0.0140503366,
0.0798740238,
0.0405138545,
0.0752019808,
0.0331990868,
-0.0546602309,
0.0030055596,
-0.0091932211,
0.0274670888,
-0.0289417766,
-0.0260426737,
0.0310893096,
0.008392863,
-0.0444485284,
0.008306724,
0.0631943271,
-0.0280052703,
0.0554499254,
0.0138064958,
-0.0418321341,
-0.0236864928,
-0.0419395939,
0.0047258786,
0.021802986,
-0.0090314532,
0.0238149147,
-0.041117616,
-0.0448942147,
0.012419763,
-0.0563589893,
0.0017840817,
-0.0096311998,
0.034928754,
0.0330458954,
-0.0212442838,
-0.0021572004,
-0.0734330937,
-0.0209007058,
-0.0170270763,
-0.0327927209,
0.0099019781,
-5.972379331e-33,
-0.0057235146,
-0.0323718302,
0.0115114748,
0.0382645652,
-0.009931609,
0.0121367984,
-0.0113556264,
0.0156693049,
-0.0328825265,
-0.0239841435,
-0.0019639013,
0.0075481897,
0.0041688541,
-0.0235859044,
-0.0243664216,
-0.0301656164,
0.0146242371,
-0.0178354904,
-0.0131593999,
0.0358002372,
0.0680633187,
0.0110467067,
0.0459939055,
-0.0959689319,
0.0502776168,
0.0072844941,
-0.0305833668,
-0.0147155933,
0.0071796733,
0.0389260538,
-0.0129075246,
0.0360360332,
0.0365339182,
-0.0141692823,
0.0255205017,
0.0143330954,
-0.0083790533,
-0.0094229225,
-0.0334084257,
-0.0193599593,
0.0356802382,
-0.0509648956,
0.0108166598,
-0.0172657222,
-0.0045270501,
0.0053507378,
0.0171200596,
-0.0207064264,
0.0350564532,
-0.0169332009,
-0.0134529714,
0.0254803039,
-0.0362477489,
-0.01167579,
-0.0169736389,
0.0558934584,
0.0385512747,
0.0241809934,
-0.0457114317,
0.0672569275,
0.026513597,
-0.0317724459,
0.0689341873,
-0.0079289619,
-0.0332592875,
0.0438284874,
0.0217150263,
-0.0540851653,
-0.0354046449,
0.0000626692,
-0.0011811318,
0.0424558073,
0.0534316599,
0.035080146,
-0.0243134666,
-0.035331618,
-0.0902904496,
0.0385570973,
0.0852593705,
0.0421934202,
0.0043573743,
0.0001064951,
-0.072911486,
0.0194794834,
0.0082172686,
-0.0192215666,
-0.0417860895,
-0.0558467284,
-0.0101697668,
-0.0233215503,
-0.0177749675,
-0.0142082861,
-0.009628457,
-0.0161016826,
0.0047094524,
0.0347527675,
-0.0056440192,
0.0256450102,
0.0011039028,
0.0224112906,
-0.0496988147,
-0.0370217152,
0.0537109114,
-0.0223685764,
0.0364956968,
-0.0312506557,
0.0301022232,
0.0335803218,
-0.0517263599,
-0.0079975594,
-0.0193926971,
0.0685075447,
-0.0210386533,
0.013258853,
0.0007514449,
0.0039557619,
0.015059053,
-0.0000861252,
0.0203351937,
-0.0027386739,
-0.0188592263,
0.0314303078,
-0.0445843972,
-0.022369625,
-0.0190734118,
0.0277160034,
-0.0049474221,
0.0294090305,
-0.0331751741,
-0.0467726961,
-0.0284210704,
-0.0334151387,
2.885e-7,
0.0080601936,
0.0092495084,
0.0056350757,
0.0387154073,
-0.0188039504,
0.0559216328,
0.0042350073,
0.069077976,
-0.0304348432,
-0.0166502409,
-0.0012830222,
0.045030769,
0.0358355381,
0.034860637,
-0.0560977682,
-0.0695346892,
-0.0323889852,
-0.0578120723,
-0.089412868,
0.0443303138,
0.0691170692,
0.0650967956,
0.0163102951,
-0.0149004478,
-0.0247513819,
-0.0523669496,
0.0089721857,
-0.0302171037,
-0.0286305211,
0.0343242511,
0.0425526984,
0.0713701844,
-0.0014575991,
0.0700002983,
0.0238127224,
-0.0074064764,
-0.0091626719,
0.0299409479,
0.0471246243,
-0.0219200123,
-0.0027956802,
-0.0161159169,
-0.0065618167,
-0.0539688654,
0.0785600469,
-0.0390266217,
-0.0544777885,
0.0174162127,
-0.0117685469,
-0.004467038,
0.0381064191,
-0.0449610166,
-0.0250254069,
0.0092057195,
-0.0294872951,
-0.0535990149,
0.0151260411,
-0.0528932996,
0.0484679714,
-0.0344015136,
-0.0544513054,
-0.063264668,
-0.012668726,
-0.0182081424,
0.0614254177,
0.0193292592,
0.0424602516,
2.2582911e-34,
0.0034356145,
0.0038051831,
-0.046525348,
0.0406896509,
0.0299014915,
0.0353237651,
0.0048398017,
0.0107053919,
-0.0035753609,
-0.015222176,
0.0108784605
] |
http://arxiv.org/abs/2401.14818v1 | http://arxiv.org/pdf/2401.14818v1 | ChemDFM: Dialogue Foundation Model for Chemistry | 2024-01-26T12:45:55 | 2024-01-26T12:45:55 | Zihan Zhao, Da Ma, Lu Chen, Liangtai Sun, Zihao Li, Hongshen Xu, Zichen Zhu, Su Zhu, Shuai Fan, Guodong Shen, Xin Chen, Kai Yu | cs.CL | cs.CL, cs.DL | Large language models (LLMs) have established great success in the general
domain of natural language processing. Their emerging task generalization and
free-form dialogue capabilities can greatly help to design Chemical General
Intelligence (CGI) to assist real-world research in chemistry. However, the
existence of specialized language and knowledge in the field of chemistry, such
as the highly informative SMILES notation, hinders the performance of
general-domain LLMs in chemistry. To this end, we develop ChemDFM, the first
LLM towards CGI. ChemDFM-13B is trained on 34B tokens from chemical literature,
textbooks, and instructions as well as various data from the general domain.
Therefore, it can store, understand, and reason over chemical knowledge and
languages while still possessing advanced free-form language comprehension
capabilities. Extensive quantitative evaluation shows that ChemDFM can
significantly outperform the representative open-sourced LLMs. Moreover,
ChemDFM can also surpass GPT-4 on a great portion of chemical tasks, despite
the significant size difference. Further qualitative evaluations demonstrate
the efficiency and effectiveness of ChemDFM in real-world research scenarios.
We will open-source the ChemDFM model soon. | ChemDFM: Dialogue Foundation Model for Chemistry: Large language models (LLMs) have established great success in the general
domain of natural language processing. Their emerging task generalization and
free-form dialogue capabilities can greatly help to design Chemical General
Intelligence (CGI) to assist real-world research in chemistry. However, the
existence of specialized language and knowledge in the field of chemistry, such
as the highly informative SMILES notation, hinders the performance of
general-domain LLMs in chemistry. To this end, we develop ChemDFM, the first
LLM towards CGI. ChemDFM-13B is trained on 34B tokens from chemical literature,
textbooks, and instructions as well as various data from the general domain.
Therefore, it can store, understand, and reason over chemical knowledge and
languages while still possessing advanced free-form language comprehension
capabilities. Extensive quantitative evaluation shows that ChemDFM can
significantly outperform the representative open-sourced LLMs. Moreover,
ChemDFM can also surpass GPT-4 on a great portion of chemical tasks, despite
the significant size difference. Further qualitative evaluations demonstrate
the efficiency and effectiveness of ChemDFM in real-world research scenarios.
We will open-source the ChemDFM model soon. | [
0.0557620786,
0.0715224817,
0.0181610584,
0.0128025543,
-0.0276445393,
0.0003019932,
0.0396620147,
-0.000565415,
0.0854220837,
-0.0206381977,
-0.0378739536,
-0.0395570248,
-0.0482695177,
0.0638424009,
0.0371789485,
-0.0027151559,
0.047030855,
-0.0508206524,
0.0434565842,
-0.009662631,
0.0296888594,
-0.0146598127,
-0.0159333814,
0.0200771913,
0.0192939471,
-0.0082133533,
-0.0124152135,
-0.0275965612,
0.0250267219,
-0.0289702341,
-0.0018635707,
0.0583792105,
0.0328342654,
0.0465415902,
0.0000020697,
-0.0812723339,
-0.0019807457,
0.0320279785,
-0.0471180677,
-0.0205447003,
0.0349046327,
-0.0227542538,
-0.0121339774,
0.0129040824,
-0.0367890224,
0.003935019,
0.0549491979,
0.0520337708,
0.037636932,
0.0492516607,
0.0053100022,
-0.0276684333,
0.0536839515,
-0.0167187769,
0.039903719,
-0.0291075706,
0.045426134,
0.0310104024,
0.0208409447,
0.0080420794,
0.0117459549,
0.0493095741,
-0.0328028947,
-0.0446748212,
0.0289012566,
0.0045308331,
-0.0652156174,
-0.0098011531,
-0.0178531427,
0.0953102112,
0.0235376637,
-0.0119260354,
-0.0424503088,
0.0194966849,
-0.0263615828,
0.0070759095,
-0.0825307518,
0.0154772708,
0.0020764002,
0.003619717,
-0.0051876674,
-0.0009687361,
0.0240032636,
0.0117236255,
-0.0245384686,
0.0818225071,
-0.0177047383,
-0.0290625431,
0.0391153209,
-0.0304612443,
-0.0878510028,
-0.0054045427,
0.0037779473,
0.0318190865,
-0.0139966942,
0.0271180514,
0.0213825796,
-0.0034698967,
0.0651605502,
-0.0412568636,
-0.0032675795,
0.0219108146,
-0.0089536924,
0.0572545379,
-0.0169217456,
0.0587692149,
-0.0257341191,
0.0528010726,
-0.0506636836,
-0.02173624,
-0.0768789724,
-0.0248542856,
-0.0387499183,
0.0375822894,
-0.0298452638,
-0.060357444,
-0.0290238913,
0.0300492756,
0.0064304303,
0.0227015503,
0.0393427126,
0.0448976383,
-0.0186246987,
0.0114289466,
-0.0567384697,
0.0366854444,
-0.004456826,
-0.0076003731,
0.0081503792,
-0.0469223522,
-0.0400317572,
-0.013080732,
-0.0222472772,
-0.0249806233,
0.016257083,
0.1372709423,
0.0013834055,
-0.0175218247,
0.0092555378,
-0.0360658951,
0.0350545421,
-0.0696430951,
0.0337819457,
-0.0134284683,
0.0637070909,
0.0498616584,
0.0148276882,
-0.051824525,
0.0179921146,
-0.009314266,
-0.0231102314,
0.099071607,
-0.0035748871,
-0.0606956035,
0.0293808151,
-0.0168942343,
-0.0167140961,
0.002880309,
0.0091670929,
-0.0322324634,
0.0138286231,
-0.0152460653,
0.0003160098,
-0.0302502122,
0.0037868181,
-0.0505474396,
0.024658069,
-0.0022471636,
-0.0000785044,
0.0752832964,
-0.0098220427,
0.0388706848,
0.008617213,
-0.027567232,
0.0270732008,
0.0441646278,
0.0129847759,
0.0120659461,
0.0135194827,
0.0446327068,
0.0398140065,
-0.0335912555,
-0.0456753969,
0.0513476506,
-0.050977435,
-0.0048735915,
0.035608653,
0.0151526602,
-0.0088715153,
-0.0223786663,
-0.0008336678,
0.0063689691,
0.0418739505,
-0.0122005651,
0.0581698343,
-0.0543834716,
-0.0773416981,
0.0351871066,
0.0337727889,
-0.0297369622,
-0.0350635201,
-0.0170548093,
0.0646713749,
0.0783484504,
0.0403841734,
-0.0703976378,
-0.0459845029,
-0.0108188698,
-0.029066259,
0.0118265096,
0.0010840124,
-0.0103697069,
0.0200215597,
-0.0352079235,
-0.0183700882,
-0.0506338775,
-0.0069925482,
-0.0081672976,
-0.0262506548,
-0.0061647389,
-0.002788356,
-0.0214987732,
-0.0149317561,
0.0134376725,
0.0161049459,
-0.0408391356,
-0.0215349179,
-0.0282778759,
0.051876545,
0.0242783688,
0.0059166532,
0.0464598984,
0.06104609,
-0.0628120154,
-0.0145231392,
-0.0186417904,
-0.0091515006,
-0.0500286631,
0.0380058251,
-0.0220305156,
0.0423070528,
-0.0010769833,
-0.056941051,
-0.0432174169,
0.0445252769,
0.0031184088,
0.0016638873,
-0.0233363397,
-0.0032053513,
0.0061039971,
0.0554760732,
-0.0437275097,
0.0577071607,
0.0170151312,
-0.0186960753,
0.0010641885,
0.037144646,
0.0442658663,
-0.0391720682,
0.0136843622,
0.0206092391,
0.0296389572,
0.0056346217,
0.0153274722,
-0.0173124336,
-0.0375099331,
0.0061846692,
-0.0725508332,
0.0112979952,
0.0095002344,
-0.0044247876,
0.0056749922,
-0.0211639814,
0.0209422242,
-0.0020665473,
-0.0150739066,
0.0008730727,
-0.0072154445,
-0.040851064,
0.0499515608,
0.0325892419,
-0.0287568662,
-0.0515323989,
-0.0291335844,
-0.034969613,
0.0524374545,
0.0011755096,
-0.0379121117,
-0.009636377,
-0.0122614624,
-0.0036081513,
0.0403325446,
-0.0151017373,
-0.0055948482,
0.0093577364,
-0.022884177,
-0.0058335825,
0.0649033189,
0.0067122462,
0.0113160936,
0.0173269901,
0.0071653537,
-0.0591912195,
0.0280600339,
-0.008139289,
0.0211971123,
0.0138432439,
0.0122097051,
0.0150952525,
0.0578171089,
-0.0323463082,
0.0202837102,
-0.033697851,
0.048937425,
0.0274527706,
-0.0317843854,
-0.048549857,
0.0037294251,
0.0072053224,
-0.0107885823,
0.0100159114,
-0.0458388329,
0.0935344547,
-0.0395841822,
0.00415581,
-0.0767189264,
-0.0016791051,
-0.0103560891,
0.0257871617,
-0.0090892809,
0.0208578929,
0.0410721265,
-0.0079730703,
0.002081567,
-0.0146751758,
-0.0043455795,
0.0014724226,
0.0293963142,
-0.0781217888,
0.0211006161,
0.011307626,
-0.0203886721,
-0.0136761842,
0.0090521108,
-0.0411398374,
0.0575934388,
0.0379453488,
-0.0470753908,
-0.0243573207,
0.0380206034,
0.0448388085,
0.0088580027,
-0.0278749205,
0.0136522129,
-0.0426052622,
0.0122828139,
0.0416485891,
-0.0422724709,
0.0421678275,
-0.0240287688,
0.0597896166,
0.0541857891,
0.0120106973,
-0.0186921358,
-0.0085430602,
0.0009386978,
0.0029473177,
0.0038471397,
-0.0143146059,
-0.0155095961,
-0.003293917,
-0.042982582,
-0.0167607293,
0.1011213958,
0.0168178473,
-0.031835068,
0.0148600331,
-0.0014711407,
-0.0352568179,
-0.0053179315,
0.0488917455,
-0.0920959562,
-0.0316094272,
-0.0340017527,
-0.0065268944,
-0.0202391278,
-0.0222551376,
-0.0243592672,
-0.0723570958,
0.0338240117,
0.0278990846,
-0.0552416891,
-0.014977647,
-0.0412977748,
-0.0294228252,
-0.0118608596,
0.0138412528,
0.0344171636,
0.0023050059,
-0.0390317775,
0.0260324664,
-0.0140198721,
-0.0781313553,
0.0169497747,
0.0004666701,
0.0237807371,
0.0520965308,
0.000296573,
-0.0030773235,
-0.008109821,
0.0355115831,
0.0178412069,
0.031592086,
-0.0473838337,
0.0348244868,
0.0266510714,
-0.0271063764,
-0.0292043537,
0.0427220203,
-0.0458876751,
-0.0397043712,
-0.0177997723,
0.1126027852,
-0.0618851557,
0.0024297168,
-0.0535595566,
0.0061777113,
-0.0056255679,
0.0715513974,
-0.0260128509,
-0.0185710825,
-0.0887509882,
-0.023516139,
-0.0201407168,
-0.0430052914,
0.0150786703,
-0.0183009282,
0.0143966852,
0.0011209482,
-0.0288241114,
0.0057610762,
0.0507357828,
0.0066043609,
-0.0214284621,
0.0044037728,
0.0240010247,
-0.0171459969,
-0.0298139267,
0.0599814169,
-0.0005969525,
0.0213993844,
-0.0823105574,
-0.0659689754,
0.0439257436,
-0.0189494155,
-0.0376232453,
0.0458131172,
-0.0515163578,
-0.0644232109,
-0.010583194,
0.0617701523,
0.0222254936,
0.0241259523,
-0.0174280945,
-0.0502379797,
0.0454660468,
0.0034001151,
-0.0258329026,
-0.0151176453,
-0.0045812442,
0.0042444253,
-0.0071323127,
-0.0066450476,
-0.0197291169,
0.0219292138,
-0.0135670919,
0.0393088795,
-0.0218678806,
-0.0321790017,
0.0557020493,
0.0354383402,
0.0473933928,
0.0407486185,
0.0296955481,
-0.0329247713,
0.0072403313,
-0.0003508518,
0.0065806657,
-0.055072505,
0.0038475457,
0.0460051,
-0.048604846,
-0.0048719286,
-0.0316672213,
-0.0431434438,
0.032508295,
0.0344808139,
0.0200308915,
0.0092838453,
0.0182949137,
0.0802996159,
-0.0207632445,
-0.0158837717,
-0.0172447767,
-0.03991073,
0.0388958752,
-0.0274096783,
0.070515275,
0.0056226742,
0.0136868097,
-0.0326122008,
0.0558268428,
-0.0539228395,
-0.0852657706,
-0.0376126319,
0.0322095081,
0.0177545473,
0.0110192057,
0.0273741912,
-0.038563665,
-0.0091323033,
-0.0356766693,
0.0362229794,
0.0259326082,
-0.0101537276,
0.0782100409,
0.0357600264,
-0.0049444293,
-0.0197754167,
0.0662462339,
-0.0141563676,
0.0182624776,
-0.0255910866,
-0.0458162986,
-0.0525132604,
-0.0366614424,
-0.0246402975,
0.0279792044,
-0.0132721635,
0.0164825302,
-0.0338953249,
-0.0177397542,
0.0362637714,
0.0183787402,
-0.0179476291,
-0.01907284,
0.0507537834,
0.0543824024,
0.0075475075,
-0.0702008829,
-0.0231091604,
0.0308217444,
-0.0169088673,
-0.0416935571,
-0.0205838252,
-6.202368842e-33,
-0.0138985906,
0.0097409505,
-0.009681684,
-0.0042950311,
-0.0255035032,
-0.0401570238,
-0.0904841945,
-0.0159714893,
-0.0058474718,
0.0540323853,
-0.0162146166,
0.006024594,
0.0213992149,
-0.0082588783,
0.0088089313,
-0.0072789937,
0.0065259179,
0.021448642,
0.0089653647,
0.0475197993,
0.0900771096,
0.0629254505,
0.0492977314,
-0.0748673603,
-0.0021598639,
0.0272959918,
0.02049697,
0.0063064829,
-0.0591086298,
0.0635190159,
-0.0558662564,
0.0641882792,
-0.0026181741,
-0.0199088734,
-0.0165495574,
0.0137952128,
-0.0572033338,
0.005057028,
0.0279510394,
-0.0124744885,
-0.0238925721,
-0.0380916446,
0.0042456058,
-0.0115303136,
-0.0278063305,
0.0038792326,
0.0242748056,
-0.0278950911,
0.0074616447,
-0.0102985967,
-0.0831075758,
-0.0055957008,
-0.0034293262,
0.027553495,
-0.0121790273,
0.0408331305,
-0.015399307,
0.0110431015,
0.0171061251,
0.0358369872,
0.0022254835,
0.0110365674,
0.0319732912,
0.0247243345,
0.0171414111,
0.034179844,
-0.008902167,
0.0015151849,
0.0167032089,
0.0095439972,
0.0035877759,
0.0486120917,
0.0278685912,
0.0378236584,
0.006497419,
0.0203870647,
-0.1053272858,
-0.0219230093,
-0.0008333718,
0.0546600036,
0.0040817889,
0.0328005627,
0.0035761378,
-0.0417751446,
-0.0193790626,
0.0795827508,
-0.0151660182,
-0.032589823,
0.0016574082,
-0.0447034501,
-0.0420497768,
-0.0114532197,
0.0295770075,
-0.0532044768,
-0.0010655938,
0.0642813668,
0.0080148755,
0.032838095,
-0.0091144005,
0.0012310072,
-0.1040882319,
-0.0427960232,
0.0069771698,
-0.0293965694,
0.020428896,
0.0057078148,
-0.0106591573,
0.0305893738,
0.0020180554,
-0.0211293083,
-0.0893944502,
0.0128885601,
0.0034384741,
0.0273851007,
-0.0306200385,
0.0318062454,
-0.0083077112,
0.0603107251,
-0.0372487046,
-0.030890815,
0.0051919399,
0.0023784756,
0.0257085469,
0.033764232,
-0.0441139489,
0.0006338456,
-0.0448528603,
-0.0091029555,
0.0125249885,
-0.0612771995,
0.0110609904,
0.0129677495,
2.856e-7,
0.0762799531,
0.033547163,
0.0549679622,
0.0053950548,
-0.0134597588,
0.0107279308,
-0.018321082,
0.0235501118,
-0.017350411,
0.002980151,
0.0277257841,
0.0120304096,
0.0113409273,
0.0078765173,
-0.0580581203,
0.0040628165,
-0.0402623303,
-0.0655104294,
-0.0624377728,
0.0038639759,
0.021448981,
0.0931166038,
-0.0220840555,
-0.0173963699,
0.0271022357,
-0.0396217294,
0.0489844419,
-0.0270110145,
0.0097723622,
0.0760290325,
0.0047834804,
-0.0341631025,
0.0016805676,
0.0835774839,
-0.0006791866,
-0.0097953929,
-0.0136790648,
0.0538243204,
-0.0219007209,
-0.046605885,
0.0197205748,
-0.0312232189,
0.0429346673,
-0.0297515783,
0.0337348133,
-0.0225106925,
-0.0115095107,
0.0039338409,
-0.012917296,
-0.0299553648,
0.0085321777,
-0.0187276416,
0.0059456769,
0.0166911837,
-0.0092976959,
-0.0048992122,
0.0255643334,
-0.0729786083,
-0.0357007347,
-0.0091412049,
-0.0599937439,
-0.005131104,
-0.0311412681,
-0.0204100516,
0.01363518,
-0.0312232953,
0.003821776,
2.497407543e-34,
0.0710990056,
-0.010159418,
0.0315629207,
0.0122364275,
0.0599771366,
0.0129181882,
-0.0190953221,
0.016801212,
0.018774366,
-0.1199882329,
-0.0117445681
] |
http://arxiv.org/abs/2401.14777v1 | http://arxiv.org/pdf/2401.14777v1 | Large Language Model Adaptation for Financial Sentiment Analysis | 2024-01-26T11:04:01 | 2024-01-26T11:04:01 | Pau Rodriguez Inserte, Mariam Nakhlé, Raheel Qader, Gaetan Caillaut, Jingshu Liu | cs.CL | cs.CL, cs.AI | Natural language processing (NLP) has recently gained relevance within
financial institutions by providing highly valuable insights into companies and
markets' financial documents. However, the landscape of the financial domain
presents extra challenges for NLP, due to the complexity of the texts and the
use of specific terminology. Generalist language models tend to fall short in
tasks specifically tailored for finance, even when using large language models
(LLMs) with great natural language understanding and generative capabilities.
This paper presents a study on LLM adaptation methods targeted at the financial
domain and with high emphasis on financial sentiment analysis. To this purpose,
two foundation models with less than 1.5B parameters have been adapted using a
wide range of strategies. We show that through careful fine-tuning on both
financial documents and instructions, these foundation models can be adapted to
the target domain. Moreover, we observe that small LLMs have comparable
performance to larger scale models, while being more efficient in terms of
parameters and data. In addition to the models, we show how to generate
artificial instructions through LLMs to augment the number of samples of the
instruction dataset. | Large Language Model Adaptation for Financial Sentiment Analysis: Natural language processing (NLP) has recently gained relevance within
financial institutions by providing highly valuable insights into companies and
markets' financial documents. However, the landscape of the financial domain
presents extra challenges for NLP, due to the complexity of the texts and the
use of specific terminology. Generalist language models tend to fall short in
tasks specifically tailored for finance, even when using large language models
(LLMs) with great natural language understanding and generative capabilities.
This paper presents a study on LLM adaptation methods targeted at the financial
domain and with high emphasis on financial sentiment analysis. To this purpose,
two foundation models with less than 1.5B parameters have been adapted using a
wide range of strategies. We show that through careful fine-tuning on both
financial documents and instructions, these foundation models can be adapted to
the target domain. Moreover, we observe that small LLMs have comparable
performance to larger scale models, while being more efficient in terms of
parameters and data. In addition to the models, we show how to generate
artificial instructions through LLMs to augment the number of samples of the
instruction dataset. | [
0.0414456688,
0.0997632518,
-0.0010130604,
0.0090955058,
-0.0115537485,
0.0201358646,
0.007299941,
0.0414847322,
0.0071668481,
-0.0760947913,
-0.027949553,
-0.0459657237,
-0.0514191464,
0.0216541234,
0.013725738,
-0.0303522572,
0.0184882712,
-0.0152055575,
0.0425448604,
-0.0161319561,
0.0158068296,
0.0189881362,
0.0336370021,
-0.0049214265,
0.0137891993,
0.002042033,
0.0025368966,
-0.0003599655,
-0.0157575756,
-0.0050211558,
-0.0020146167,
0.0153999478,
0.0258742664,
-0.0004002602,
0.0000019028,
-0.0506262109,
-0.0330311172,
-0.0016594006,
0.0160421059,
0.0260371957,
0.0886081532,
-0.0659383535,
0.0268092453,
0.0369426385,
-0.0655350611,
-0.0120507246,
0.0548648685,
0.0932993814,
0.0703328848,
0.1148085594,
-0.0310086124,
-0.0680013224,
-0.014448585,
-0.0442377888,
0.0004048443,
-0.0615634471,
0.0484732687,
0.0261951648,
-0.0065271449,
0.0248706155,
-0.0271234848,
0.0115406243,
-0.0453135669,
0.0038186971,
0.0129078804,
-0.02671309,
-0.0254813135,
0.0272809733,
-0.0631510615,
0.019320434,
-0.0109090162,
0.0364729315,
0.0062346389,
-0.0384632312,
-0.0103493584,
0.0156727377,
-0.0263047293,
0.0436866358,
-0.0029384366,
0.0152101833,
0.0347718038,
-0.0105996598,
0.0074548335,
0.0006925206,
-0.0363224633,
0.0276049413,
-0.0116195502,
-0.0280504804,
0.0416578874,
0.0217765179,
-0.0093685277,
-0.0153119434,
-0.0033148027,
0.0027319835,
0.0421357378,
0.0329051763,
-0.028633574,
-0.0135337356,
0.0003807493,
-0.1163087934,
0.0721981004,
0.0178773999,
0.0129833641,
0.0695425123,
-0.0603740998,
-0.0100991623,
-0.004216834,
-0.0328049064,
-0.0221167468,
0.0167134795,
-0.0721240789,
0.0266999193,
-0.0379721336,
0.0727769285,
-0.0582387447,
-0.0289209653,
-0.0579050556,
0.0108821932,
0.0134342955,
0.0010256948,
0.0329955593,
-0.0498507805,
0.0138457371,
-0.0275818594,
-0.0232632011,
0.0049623391,
-0.0272489656,
0.0258878302,
-0.064573288,
-0.0240484979,
0.009792367,
0.0375980623,
-0.0114029208,
-0.0036579736,
0.0347095877,
0.0752798989,
0.049484551,
-0.0369261205,
-0.0744087473,
-0.0045859483,
0.0033440264,
-0.0193862729,
-0.0110183237,
-0.0176088568,
0.012088188,
-0.0011775339,
-0.0083779283,
0.0064609661,
0.0012723804,
-0.0138350679,
-0.0601350032,
0.075713031,
-0.0020223965,
-0.0226709545,
-0.0025817759,
-0.0591265894,
0.0031769974,
0.0561739653,
-0.0053327521,
0.0105125234,
0.0402044132,
-0.023791898,
-0.024283018,
0.0355107561,
-0.0049788617,
-0.0266812593,
-0.0012511046,
0.0358364843,
-0.0267455876,
-0.0306070652,
0.0569995381,
0.0305745434,
-0.002692153,
-0.0120172827,
0.0444929413,
0.0127551211,
0.0513280444,
0.0325811654,
-0.0090164812,
0.0375273563,
0.016201105,
-0.0424451642,
0.0087215407,
-0.0157764722,
-0.0471951067,
0.0218419619,
-0.0208401494,
0.0208120476,
-0.0288079847,
-0.0382201113,
-0.0332717188,
-0.018740207,
0.0037166378,
-0.015895335,
0.0451554134,
-0.0318965465,
-0.063809216,
0.0249221735,
0.0538181923,
-0.022577364,
-0.0081467591,
-0.0202638004,
0.0593998544,
0.0958622172,
0.0036785083,
-0.0271898117,
0.0274696313,
-0.0097297458,
-0.0094354376,
0.0614345744,
0.0499904491,
0.0257751476,
-0.039559871,
-0.0603198484,
0.0045770626,
0.0048881029,
0.0436840467,
0.0503135324,
-0.0244538523,
-0.0022521892,
0.0012336108,
0.0147970151,
-0.03811156,
0.0223917551,
0.0086926315,
-0.0554288104,
0.0274172891,
-0.0523532927,
0.014193546,
0.022807464,
-0.0197406691,
-0.0093973204,
0.0044570602,
-0.0275965538,
-0.0531771481,
-0.0262919385,
-0.0244538039,
-0.0101075936,
-0.0234504733,
-0.0095262453,
0.0358121395,
0.0072095725,
-0.0469036885,
-0.007314533,
-0.0002878838,
-0.0377677083,
0.0401448198,
-0.0337701552,
-0.0018398451,
0.0059312517,
0.0402540751,
0.0018627411,
0.05087905,
0.003370198,
-0.0051682699,
-0.0400005355,
0.0222944729,
-0.0065005221,
-0.0161515754,
0.0183149278,
0.0034705184,
-0.005774071,
0.0364073664,
0.034798827,
0.0278564673,
0.0053116437,
-0.0421010405,
-0.097289972,
-0.0098482845,
0.0021196383,
0.0476449952,
0.0117873484,
0.0228085872,
-0.0193809457,
-0.0212719571,
-0.0338617451,
0.0010973809,
-0.0299379099,
0.0032855091,
0.0532807745,
-0.000134661,
-0.0214975197,
-0.0474947356,
0.0061286218,
0.0060786498,
0.023529157,
0.0324331224,
-0.0480498597,
-0.0580232032,
0.0296137966,
0.0057087834,
0.0265167318,
0.010503496,
0.0082439389,
-0.0290255137,
-0.0257267822,
0.0081518488,
0.0621175952,
0.0093335938,
0.0568793938,
-0.0252851322,
0.0004299761,
-0.031606257,
0.0268804617,
-0.0206489712,
-0.0031972432,
-0.030606227,
-0.0207892749,
0.0248903539,
0.053686332,
-0.0223357845,
-0.0126741342,
-0.0894544423,
0.0342724882,
-0.0151681853,
0.0246324446,
0.0014609814,
-0.0008357691,
0.0055399816,
0.0105359666,
0.0344799682,
-0.0369195491,
0.0237855222,
-0.0104703754,
0.0016308116,
-0.0642736852,
-0.0376644321,
-0.0875087902,
0.0418738686,
-0.0247852802,
0.0048259217,
-0.0060306387,
-0.0036676424,
0.0120629231,
-0.032399822,
0.021458976,
-0.0023562862,
0.0354458466,
-0.0309150163,
0.040457584,
-0.006237661,
-0.0047114925,
0.006889808,
-0.0779078007,
-0.0346579775,
0.0107865185,
0.0806264579,
-0.0475730188,
-0.0441482216,
0.0477416813,
-0.0294266623,
-0.0207359195,
0.0063612345,
-0.0202359613,
0.0100935362,
-0.0149521148,
0.0797003135,
0.0022469778,
0.0393622145,
-0.018806519,
0.0413658135,
-0.006037436,
0.0039499751,
-0.0065666484,
-0.0319610499,
-0.0014146469,
-0.0045153722,
-0.0368396677,
-0.0025280418,
-0.0128382342,
0.0125814471,
-0.0385604613,
0.0076285116,
0.1148899943,
0.0047591818,
0.000357956,
0.0076627783,
0.020169327,
-0.0131844813,
-0.0332199074,
0.0521850772,
-0.0477334335,
-0.0346005112,
-0.0212882049,
0.0219904594,
0.015298672,
-0.0337722115,
-0.0191348195,
-0.1523334831,
0.0345987193,
0.0106411427,
-0.0081236176,
-0.0394533463,
-0.059561871,
-0.0248452462,
-0.0504692383,
0.0201429334,
-0.0108771781,
-0.0399602056,
0.0059780302,
0.0026436204,
-0.0720197856,
-0.0356157757,
-0.0104312738,
0.008849835,
-0.0057410691,
0.0163884312,
-0.0079290168,
-0.0041113989,
0.0105831306,
-0.0116217993,
-0.0243729725,
0.0098220743,
-0.041881036,
0.056446109,
-0.0117275743,
-0.0362421572,
0.0333048292,
0.0169215277,
-0.0716995895,
0.0756469071,
0.0087672761,
-0.0176151972,
-0.0117341522,
-0.0480098985,
-0.0175045431,
0.0380594321,
-0.0021753237,
0.0876451805,
-0.0007546194,
-0.0826232955,
-0.0068909503,
-0.0226950757,
-0.0424291454,
-0.0631337017,
-0.0000726418,
-0.0132797053,
-0.034460187,
-0.0039168526,
-0.1048545018,
-0.0509209447,
0.0195398629,
0.0275892764,
-0.0047545712,
0.0252794884,
0.0604590327,
-0.0065885116,
-0.0162176508,
0.0370952599,
-0.030748507,
0.0005849275,
-0.0374546424,
-0.0130740497,
0.0471897013,
0.0112294024,
-0.0493825674,
0.0263962112,
-0.0501354858,
-0.0498559177,
0.0125559429,
0.0340294056,
0.0135510694,
-0.0007759435,
-0.0088343127,
0.0182973221,
0.0293753073,
0.0445189402,
-0.0578596778,
0.0111483596,
0.056776464,
0.0429258309,
-0.0126369363,
0.0326255821,
-0.0439936519,
0.0514298752,
0.0111510288,
0.0346112922,
-0.03258292,
-0.011005138,
-0.0138067128,
0.0284172744,
0.0208491012,
0.0437137745,
-0.027942,
-0.0075682448,
0.0011555307,
-0.0184344687,
0.0106105916,
-0.0449641235,
-0.0349737555,
0.03743672,
-0.0145553695,
-0.0521227345,
0.0136840399,
-0.0206661206,
-0.0009944886,
-0.0063844505,
0.0609514005,
0.0070954105,
0.0249652639,
0.0107004456,
-0.0220262986,
-0.0048494604,
-0.0130225718,
0.0482491516,
-0.0252109226,
0.0034010811,
0.00413793,
0.0183543675,
0.028135132,
-0.0191020928,
-0.0188396424,
-0.0221748147,
-0.0124861188,
-0.0284119938,
0.0455976315,
-0.0033950985,
-0.0381188914,
0.0388463363,
-0.0504725464,
0.0146985184,
0.0005086785,
0.0212563686,
0.0180692244,
-0.0117282774,
0.0461738482,
0.0082056094,
-0.0567706488,
-0.0183652584,
0.0777900517,
0.0189753678,
0.0522521697,
-0.0119249728,
0.0026576156,
-0.0578469746,
-0.0165902693,
0.0729442909,
-0.0099575613,
-0.0073712911,
0.0384880006,
-0.0678603202,
-0.0363069624,
0.0365588106,
-0.0354267135,
0.0166399218,
-0.0104003064,
0.0467416346,
0.0416108072,
0.0119246878,
0.0621210858,
-0.0227396004,
0.017729301,
-0.0310999248,
-0.027974287,
-0.0577749796,
-6.312915269e-33,
-0.0361916088,
0.0005309306,
0.0168706365,
0.0285892822,
0.0247462094,
-0.0013938433,
-0.0028131464,
-0.0072468002,
-0.0081744194,
0.0105697243,
0.0103148576,
0.0309220087,
0.0065987138,
0.0313624814,
0.0085158814,
0.0185090788,
0.042944245,
-0.0264484063,
0.0031631081,
-0.0010542668,
0.023162324,
0.0507065803,
0.0541013815,
-0.0900008678,
0.0229291059,
0.0038741531,
-0.0502535067,
-0.0121790813,
0.0634448305,
0.0478694327,
-0.0133434022,
0.0701088384,
0.0142063247,
-0.0628579408,
-0.0158035327,
0.0099531217,
0.0020793073,
-0.0453443229,
0.0307747312,
0.0377039649,
0.0352239609,
-0.0050663711,
0.0277586281,
0.0095121637,
-0.0261415597,
0.0002759362,
0.0105977487,
-0.0307405628,
0.0092834709,
-0.0534243137,
-0.0769548714,
0.0573696829,
0.0343716256,
0.0450370088,
0.0218131281,
0.0510357246,
-0.0122221736,
0.0238516908,
0.0154383658,
0.004091145,
0.0293308645,
0.0314409919,
-0.0068080719,
-0.0147551438,
-0.0162537433,
0.0428734906,
-0.0315507054,
0.01500504,
0.0142771509,
0.0232122559,
0.0184061937,
0.0339856483,
0.0529375821,
0.0525840446,
0.0315029137,
-0.0252478458,
-0.0451497175,
0.0121608125,
0.0314522944,
0.0189731196,
0.0494595878,
0.0282901898,
-0.0241177101,
-0.0237133764,
0.0012810737,
0.0070057642,
-0.0175829902,
-0.0120903486,
0.0155258784,
0.0248248503,
-0.0332667455,
-0.0033928256,
0.00527036,
-0.0642527491,
-0.0725255832,
-0.0039038234,
0.0010638317,
0.0045806901,
0.0037035805,
-0.0290400703,
-0.0292146597,
-0.0799735412,
0.0023041337,
0.0008329228,
-0.0289252736,
-0.0333949514,
0.0218706857,
0.0172431301,
-0.0876695812,
-0.019254161,
0.0259736236,
0.0254223868,
0.0047645979,
0.0770706758,
-0.0086137317,
0.0218942985,
0.0004756459,
0.0395987034,
-0.0038343207,
0.0118093565,
0.025773555,
-0.0367664099,
0.0018827214,
0.0450522937,
-0.0073582171,
0.001985654,
0.031096898,
0.0315581746,
0.021965092,
-0.040922679,
-0.0429216921,
0.0377005488,
2.71e-7,
0.0485574603,
0.1169086769,
-0.0366648696,
0.0189436786,
0.0158184655,
0.0313104577,
-0.0203440022,
0.0323494673,
-0.0162343662,
-0.010663284,
0.0169169027,
0.0191893522,
0.0487171635,
-0.0274718367,
-0.077497974,
-0.0247438345,
-0.030057257,
0.0487790816,
-0.0514827482,
0.0547219478,
0.0413857102,
0.1191983446,
-0.0090917302,
-0.0280438624,
-0.0177391768,
-0.0853685588,
0.0462180711,
-0.0701457486,
-0.0327308215,
0.0142245097,
0.0188991688,
0.0180234704,
-0.0203898288,
0.0029557005,
-0.0020518587,
0.0441825725,
-0.0010504408,
0.0594552308,
0.0092437286,
0.0300668459,
0.0334529877,
0.0107067572,
-0.008734812,
-0.0139177591,
0.0325084887,
-0.0746927038,
-0.0258127172,
-0.0695521757,
-0.0395627283,
0.0101552028,
0.0174407326,
-0.001031411,
-0.0170003213,
0.0040739919,
0.0402138568,
-0.0267292075,
-0.019128317,
-0.0351336971,
0.0208439175,
0.0302849878,
-0.0304808915,
0.001940487,
-0.0299071521,
0.0342289656,
0.0261902213,
-0.0562491678,
0.0378139876,
2.390437098e-34,
0.0905196443,
-0.0506936274,
-0.0385918655,
0.0389501154,
0.0734047443,
0.0234171152,
-0.0544319227,
0.0605941303,
0.0091572087,
-0.1357814968,
0.0169470627
] |
http://arxiv.org/abs/2401.14717v1 | http://arxiv.org/pdf/2401.14717v1 | Turn-taking and Backchannel Prediction with Acoustic and Large Language Model Fusion | 2024-01-26T08:59:07 | 2024-01-26T08:59:07 | Jinhan Wang, Long Chen, Aparna Khare, Anirudh Raju, Pranav Dheram, Di He, Minhua Wu, Andreas Stolcke, Venkatesh Ravichandran | cs.CL | cs.CL, cs.AI, cs.LG, cs.SD, eess.AS | We propose an approach for continuous prediction of turn-taking and
backchanneling locations in spoken dialogue by fusing a neural acoustic model
with a large language model (LLM). Experiments on the Switchboard human-human
conversation dataset demonstrate that our approach consistently outperforms the
baseline models with single modality. We also develop a novel multi-task
instruction fine-tuning strategy to further benefit from LLM-encoded knowledge
for understanding the tasks and conversational contexts, leading to additional
improvements. Our approach demonstrates the potential of combined LLMs and
acoustic models for a more natural and conversational interaction between
humans and speech-enabled AI agents. | Turn-taking and Backchannel Prediction with Acoustic and Large Language Model Fusion: We propose an approach for continuous prediction of turn-taking and
backchanneling locations in spoken dialogue by fusing a neural acoustic model
with a large language model (LLM). Experiments on the Switchboard human-human
conversation dataset demonstrate that our approach consistently outperforms the
baseline models with single modality. We also develop a novel multi-task
instruction fine-tuning strategy to further benefit from LLM-encoded knowledge
for understanding the tasks and conversational contexts, leading to additional
improvements. Our approach demonstrates the potential of combined LLMs and
acoustic models for a more natural and conversational interaction between
humans and speech-enabled AI agents. | [
0.0520609319,
0.0149844745,
-0.0077942433,
0.0227303877,
-0.0121852318,
-0.010943261,
-0.0125580924,
-0.0392848291,
0.0111356732,
-0.0592673086,
-0.0565913655,
-0.0440856293,
0.0304926001,
0.0261181388,
0.0302971043,
-0.019378297,
0.016164599,
-0.0679662451,
0.0106655359,
-0.0068852687,
-0.0088749621,
0.0111420946,
-0.0447381884,
0.0696720928,
-0.0185218342,
-0.0250505693,
-0.0459881686,
0.0115755107,
-0.0003574357,
0.0113518424,
0.0162673648,
0.021619631,
0.0487856902,
0.0517702997,
0.0000015715,
-0.0135262804,
0.0538211763,
-0.0441722125,
-0.0840862244,
-0.0202033967,
-0.0157625563,
0.021473268,
-0.04194602,
0.0507524051,
-0.0313656181,
0.0097665368,
-0.0013192048,
0.0684433132,
0.0564838722,
0.0496670492,
0.0115910554,
-0.0253406875,
0.0670036823,
-0.0312496722,
-0.002746237,
-0.0588743798,
-0.003392814,
-0.0091144461,
-0.0064772787,
0.0256623402,
-0.0065283524,
0.0315582305,
0.0080692749,
-0.0534764528,
-0.0238179266,
0.0556653887,
-0.032990694,
-0.0145858768,
-0.0445975922,
0.1208066046,
-0.0201865733,
-0.0088313492,
-0.0170522649,
0.0081603881,
0.0025015376,
-0.0398546159,
-0.028518552,
0.0066499244,
0.0167532098,
-0.0074416799,
-0.0292892475,
-0.0068449965,
0.0065184799,
0.0113192629,
-0.0142823746,
0.0711804703,
-0.0286423415,
-0.0268702265,
0.0214683302,
-0.0125727104,
0.0223529097,
-0.0257880334,
0.0134101175,
0.0178487524,
-0.0651382357,
0.0236469191,
-0.0098011736,
0.0388761945,
0.0036907457,
-0.0345125496,
-0.011188685,
0.0138927093,
-0.0440497212,
0.0343605876,
0.0119989337,
0.0436143428,
-0.0268587787,
-0.0252218451,
-0.0113720419,
-0.0285341013,
-0.030015232,
0.0050697597,
-0.0356789641,
0.0037775168,
-0.0629427508,
-0.061368499,
-0.0054229139,
0.0248086918,
-0.0166625325,
0.0265777819,
0.0659692287,
0.0077086352,
-0.0103708608,
-0.0082658082,
0.0047832849,
0.0399645753,
0.0186007246,
0.0572648831,
-0.0083215665,
-0.0623453073,
-0.0000260195,
0.0311972369,
-0.0142970337,
-0.0292865057,
-0.0007754687,
0.0818766281,
-0.0128338551,
0.0090945326,
-0.0253520627,
0.0150086051,
-0.0294116717,
-0.0665302426,
-0.0078426562,
0.0030654413,
0.0873351023,
-0.0172865205,
-0.0205434356,
0.0219393391,
0.014450253,
0.0042323386,
-0.0418577492,
0.0717472807,
0.0745773092,
0.0411996096,
0.081403248,
-0.0114804795,
-0.0290526915,
0.0328356475,
0.0498736724,
0.0175726898,
0.0235537328,
-0.0152572403,
0.0127936928,
0.0795874521,
-0.003243305,
0.0020861926,
0.0181003865,
0.018937014,
0.0089370944,
0.0374694169,
0.0394443795,
0.0507810339,
-0.0359842628,
-0.0346526019,
0.0668362603,
0.0417700596,
0.0096435333,
-0.0022078401,
0.0339673124,
0.0600734018,
-0.0274371076,
-0.0082347132,
-0.0627650395,
0.0216449741,
-0.039539855,
-0.0140700499,
0.0293919537,
0.0191128235,
-0.0372223109,
-0.0179467816,
0.021008417,
-0.0236148518,
0.0047151935,
0.0306116119,
0.0205546618,
-0.0189616065,
-0.0869679078,
0.0400201567,
0.0070919953,
-0.0238754656,
0.0048570195,
-0.011216538,
0.0046388269,
0.0525927506,
0.0034200973,
0.0128352661,
-0.0513398871,
-0.0660652295,
-0.0268488135,
0.0475051478,
0.0097922441,
-0.0313706659,
0.0537338816,
-0.060858041,
-0.0114892833,
0.0483633354,
-0.0227984674,
0.0275954474,
-0.0231359564,
0.0008860141,
0.0147737991,
0.0046626502,
0.0203514025,
-0.0052093663,
-0.0033013062,
-0.0171456616,
-0.0129208993,
0.0197793208,
0.0300719533,
0.0082052508,
-0.0114144357,
-0.014672731,
0.0485864431,
0.0070885662,
-0.0123131555,
0.037730515,
-0.0624155924,
-0.0384280905,
0.0383343473,
0.0113888262,
-0.0408565663,
0.0153016998,
-0.0092051672,
-0.0641714633,
0.0200676676,
-0.0220909901,
0.0569557622,
-0.0207929593,
0.045231197,
0.0474211685,
0.0121807447,
-0.0104196761,
0.0076721725,
0.0064153667,
-0.0123250121,
0.0272210874,
-0.0012987464,
-0.0072020432,
-0.0458480716,
-0.0039799339,
0.0237355046,
-0.0068718283,
-0.0068210089,
-0.0044765272,
0.0059378105,
-0.0368816108,
-0.0457501858,
-0.0819003657,
-0.0091249505,
-0.0118315052,
-0.02542096,
0.0174236111,
-0.0058811302,
0.0086913425,
-0.0068841022,
-0.0240501091,
-0.0319710188,
0.0062993662,
-0.0135068912,
-0.0387989916,
0.0308618844,
-0.0282547269,
-0.0566198006,
0.0321009979,
0.0116099725,
0.0218252428,
-0.024092162,
0.0032453251,
-0.0577165857,
-0.0066439048,
0.0139961634,
0.0807658955,
-0.048078876,
-0.0676488504,
-0.0028774659,
-0.0210424177,
0.0204665884,
0.0752844363,
0.0296267886,
-0.0216814727,
0.008851367,
-0.0710184723,
-0.0720203444,
0.0358490683,
-0.0126761235,
-0.0156994704,
-0.0871068612,
0.0147082107,
-0.0280228797,
0.0438096672,
0.003129611,
0.0402377583,
-0.0392997153,
-0.0807585195,
0.0196265131,
-0.0549711287,
-0.016277751,
-0.0438830554,
-0.0646355897,
0.0261408165,
-0.0184439737,
-0.0046249786,
0.0380005278,
-0.0558921173,
-0.0487950146,
-0.0160888266,
-0.018037511,
-0.0218187813,
0.0173625201,
-0.0147643359,
-0.0020980183,
-0.0100126313,
-0.0473562554,
-0.0345809907,
0.0078513259,
-0.0063566738,
-0.0190495644,
0.0621419363,
-0.0461998954,
-0.0061613317,
0.0285546444,
-0.0313960239,
-0.0381744541,
-0.0206762888,
-0.0251069479,
0.0141350152,
0.0334311575,
-0.0553853884,
-0.0366603769,
0.0411172174,
0.0056501264,
0.0451540574,
0.0052032871,
0.0255101044,
-0.0116066923,
0.0083723925,
0.0586662777,
0.0287930723,
-0.0380829684,
0.0312900618,
0.0857319534,
0.0039535514,
0.0586490519,
0.013925435,
-0.01422955,
-0.0078415824,
0.060640838,
0.0202248115,
-0.0607903898,
0.0174255464,
0.0179029349,
-0.0197702795,
-0.017722955,
0.1393093765,
0.0189596396,
-0.0707219765,
-0.021468617,
-0.0289394855,
-0.0046080309,
0.0104408646,
0.0143450517,
0.0036303292,
-0.0314974226,
0.0076910821,
-0.0122326883,
0.0140425023,
0.0074550025,
-0.0045846356,
-0.13224186,
0.0352050103,
0.0104984371,
-0.0542053133,
-0.0249128714,
-0.0262744781,
0.0277493913,
-0.0263710935,
0.0510879941,
-0.0041522235,
0.0058871442,
-0.0673172697,
0.0203855485,
-0.0254440103,
-0.0011518908,
0.0513387322,
-0.0353652649,
0.0505178086,
0.0573939681,
0.0744818822,
0.00582015,
-0.0001390574,
0.0446507297,
0.0126049398,
0.0193414018,
0.0011679938,
0.0365096666,
-0.0380379334,
0.017848378,
0.0007535153,
0.0522408709,
0.0020113331,
0.0102919098,
0.0112298261,
0.0603740662,
-0.0700545907,
0.0314252824,
0.0273027103,
0.0446863882,
0.0478096083,
0.0702135339,
0.0007952111,
-0.0711565465,
0.0023186146,
-0.0059793303,
-0.0158308744,
0.0118986061,
-0.0285943449,
0.0161749758,
0.0100695463,
-0.0168368835,
-0.0323591009,
-0.0546370819,
0.029408114,
0.0049883723,
0.0096854102,
0.0224662442,
0.0433626957,
-0.0137261981,
0.0205944814,
-0.0128581878,
-0.0119290035,
0.0762435049,
-0.0626056865,
-0.0018056523,
0.1057318002,
-0.0446042828,
-0.005964708,
0.0484426953,
-0.0482280403,
-0.0826061964,
-0.0108817965,
0.0127542997,
0.0210939534,
0.0962188169,
0.0004778013,
-0.0211470257,
-0.0007176392,
-0.0383320265,
-0.0249388702,
0.006014735,
0.0462016165,
0.0574072711,
-0.0198833514,
0.0341969542,
-0.0536138453,
-0.0122957006,
-0.0021849077,
0.0059480369,
-0.086588487,
-0.0840125978,
-0.0086885467,
-0.0338363796,
0.0918273553,
0.0019829662,
0.0213483628,
-0.024410734,
0.0587887838,
0.0014533582,
0.0007063656,
-0.0510007329,
-0.0012804013,
0.0139240343,
-0.0187948234,
-0.0026738816,
0.0002343117,
-0.0350358486,
0.0166843943,
-0.0428291373,
0.0307580233,
0.0168516673,
0.0083946129,
0.0221559536,
0.0094202254,
0.0040531224,
0.010943057,
0.0431218296,
0.0171385389,
0.0468102507,
0.0628455579,
0.0198322926,
-0.0337116681,
0.0032766394,
0.0288208816,
-0.0531547703,
-0.0546150431,
-0.0094919764,
0.0249960255,
0.0412458479,
-0.015815407,
0.0347311571,
-0.0038293996,
0.0364264473,
-0.0014465972,
0.0511535481,
-0.0313521922,
-0.0362410769,
0.0654763132,
0.0169771258,
-0.003684554,
0.0433209091,
0.0631530285,
0.0063695121,
0.0457421467,
-0.0134121906,
0.0368814133,
-0.0539245717,
-0.0337738506,
0.0253072865,
0.0127824824,
-0.0120786019,
0.0425017588,
-0.022752136,
-0.0659201518,
-0.0026523997,
0.0031226294,
0.0226192493,
-0.0370735563,
0.0406568982,
0.0097582452,
-0.0068327137,
0.0103208162,
-0.0427520573,
-0.033959683,
0.0072978353,
-0.0207901392,
-0.0712163001,
-5.577608655e-33,
-0.0223119855,
-0.0351665542,
0.0291926973,
0.0269899517,
-0.0649116561,
-0.0234365035,
-0.0051419823,
0.0053645507,
-0.0334799923,
-0.0038859562,
-0.0177337658,
0.0539364852,
0.0216720067,
0.0088084806,
-0.0011331125,
0.0064161471,
0.040170785,
-0.0158485714,
0.0145026911,
0.0194588117,
0.0028224466,
0.0138944844,
0.0286351424,
-0.0961552858,
0.000115564,
-0.017224323,
-0.0385461375,
0.0394572206,
-0.0341929533,
0.0540013611,
0.0163720604,
0.0248565692,
-0.0123303849,
-0.0724524111,
0.009391861,
0.0604579933,
-0.0310061332,
0.014186237,
0.0250990521,
0.0368797742,
0.0219312664,
0.0088660661,
-0.0023925491,
0.0006450602,
-0.0037400525,
0.0060456474,
0.0212381836,
-0.007804004,
-0.0289437473,
0.0155388918,
-0.0224973001,
0.0016869799,
-0.0429176353,
-0.0768456385,
0.0071798488,
0.0523323715,
0.0147480043,
-0.0419335552,
-0.0822107494,
0.0390574299,
-0.0168009624,
-0.0198745523,
0.0423819572,
0.0445273332,
0.0140045714,
0.0348943621,
0.0391021147,
-0.0042260289,
0.0040200334,
-0.0542676598,
-0.0018825486,
0.0648277104,
-0.0255738292,
-0.019219676,
-0.0447656363,
-0.0061539942,
-0.0103231017,
-0.0050088023,
0.0586995557,
0.0566909015,
-0.0080256211,
0.0239449795,
-0.0263397899,
-0.0246838052,
-0.0747971535,
-0.0313403346,
-0.0074208137,
0.0076817716,
0.0205855668,
0.0359007679,
0.0415496863,
-0.0283741374,
0.0619791374,
-0.0210025106,
-0.0073035625,
0.0431363024,
-0.0126423249,
0.0132424124,
-0.0189279895,
0.0216113999,
-0.0568536147,
-0.0421971492,
0.0403982922,
-0.0209834855,
0.0146454778,
0.0117644779,
-0.0034034958,
0.0385991335,
-0.0229047704,
-0.0119406814,
-0.0257466808,
0.022365056,
0.0035372502,
0.0334458686,
0.0324513093,
-0.0107405754,
-0.041916471,
0.0501237102,
-0.0440970734,
0.0232924875,
-0.0062366123,
-0.0193567276,
0.0271709189,
0.0160906371,
-0.0742292032,
-0.0276156254,
0.006752911,
-0.0198579952,
0.0487663671,
-0.0243631452,
-0.0233496279,
0.0097900331,
2.454e-7,
-0.0341923796,
0.0275439247,
-0.001489328,
0.0396826603,
0.0279171709,
0.0476562083,
-0.0195660442,
-0.0121090086,
0.0393142514,
-0.0168485828,
0.0398608148,
0.0176684279,
0.0665587485,
-0.0154290432,
-0.1247417033,
-0.0640754029,
-0.0724111348,
-0.0304385815,
-0.0047694282,
-0.0082069431,
0.0062490231,
0.0939501747,
-0.0076513528,
-0.0173765253,
-0.0110457391,
0.0139091499,
0.0237846132,
-0.042044349,
0.0075594527,
-0.0021267626,
0.0008462418,
-0.0080197612,
-0.0436396375,
0.0479269959,
-0.0139205316,
0.0284702163,
-0.011574639,
0.0572560653,
-0.0297806952,
-0.082972832,
0.0058993995,
0.0367503092,
-0.0194895789,
-0.0034043726,
0.0315227732,
-0.0334842615,
-0.0257134158,
-0.0238144528,
-0.0069679152,
0.021729216,
-0.0145340487,
-0.0199889392,
0.0251467042,
0.007883139,
-0.0085795391,
-0.0186925083,
0.0226016361,
-0.0787338689,
-0.0033586991,
0.0101227509,
-0.0726507083,
-0.0418605581,
0.0174314342,
-0.0403132737,
0.0043981387,
-0.0109314444,
0.0159457047,
1.472393459e-34,
0.0322725289,
-0.0121904109,
0.0296878051,
0.0643879697,
0.0448137596,
-0.0042761737,
0.0033940868,
0.0200997218,
0.0343522914,
-0.0731510073,
-0.0341441929
] |
http://arxiv.org/abs/2402.01717v1 | http://arxiv.org/pdf/2402.01717v1 | From RAG to QA-RAG: Integrating Generative AI for Pharmaceutical Regulatory Compliance Process | 2024-01-26T08:23:29 | 2024-01-26T08:23:29 | Jaewoong Kim, Moohong Min | cs.CL | cs.CL, cs.AI, cs.IR, I.2.7; I.2.1; J.3 | Regulatory compliance in the pharmaceutical industry entails navigating
through complex and voluminous guidelines, often requiring significant human
resources. To address these challenges, our study introduces a chatbot model
that utilizes generative AI and the Retrieval Augmented Generation (RAG)
method. This chatbot is designed to search for guideline documents relevant to
the user inquiries and provide answers based on the retrieved guidelines.
Recognizing the inherent need for high reliability in this domain, we propose
the Question and Answer Retrieval Augmented Generation (QA-RAG) model. In
comparative experiments, the QA-RAG model demonstrated a significant
improvement in accuracy, outperforming all other baselines including
conventional RAG methods. This paper details QA-RAG's structure and performance
evaluation, emphasizing its potential for the regulatory compliance domain in
the pharmaceutical industry and beyond. We have made our work publicly
available for further research and development. | From RAG to QA-RAG: Integrating Generative AI for Pharmaceutical Regulatory Compliance Process: Regulatory compliance in the pharmaceutical industry entails navigating
through complex and voluminous guidelines, often requiring significant human
resources. To address these challenges, our study introduces a chatbot model
that utilizes generative AI and the Retrieval Augmented Generation (RAG)
method. This chatbot is designed to search for guideline documents relevant to
the user inquiries and provide answers based on the retrieved guidelines.
Recognizing the inherent need for high reliability in this domain, we propose
the Question and Answer Retrieval Augmented Generation (QA-RAG) model. In
comparative experiments, the QA-RAG model demonstrated a significant
improvement in accuracy, outperforming all other baselines including
conventional RAG methods. This paper details QA-RAG's structure and performance
evaluation, emphasizing its potential for the regulatory compliance domain in
the pharmaceutical industry and beyond. We have made our work publicly
available for further research and development. | [
0.1087095067,
0.0247116722,
-0.0215221606,
-0.0473492444,
0.0082084052,
-0.0152614359,
-0.0040169014,
0.0287446529,
0.0680960938,
-0.0147065008,
0.0144074736,
-0.0275126025,
-0.0490828082,
0.0553770475,
-0.0163209923,
-0.0272495877,
0.0112077296,
0.0469151214,
0.0420852453,
0.0346916355,
0.0419724658,
0.0316319875,
-0.0265363026,
0.0704272538,
-0.0242963843,
-0.0280630011,
-0.0055495743,
-0.0034359875,
0.0291014537,
-0.0392164253,
0.0257903598,
0.0280107763,
0.0114024589,
0.0211206898,
0.0000018312,
-0.05399042,
-0.0303475633,
0.0238859747,
-0.0361684598,
-0.0508458763,
0.0409081504,
0.0150775313,
-0.0126402928,
0.069864504,
-0.0508683771,
0.0010992943,
0.0501718856,
0.0295472927,
0.053558588,
0.0741233975,
-0.0161675829,
-0.0166541319,
0.0426437855,
-0.0290394109,
0.0404088087,
-0.0767813623,
0.0200938284,
0.0564639196,
0.0533465371,
0.0093754698,
0.0355042852,
0.03049214,
-0.0020147476,
-0.0479612797,
0.0172490962,
0.0171837676,
-0.0089305146,
-0.0381213203,
-0.025186643,
0.0756445304,
0.0496059507,
-0.0018389653,
0.0073606428,
0.0070835664,
0.0508759767,
0.044818785,
-0.0425345451,
0.0088422867,
-0.0533078164,
-0.0142916366,
-0.0032154785,
0.004109086,
-0.0265814327,
-0.0365944318,
0.0240308456,
0.00414446,
0.0135893449,
-0.0420637913,
-0.0136949718,
-0.0119692795,
0.0370597541,
-0.0340070203,
0.0220023841,
0.0686687157,
-0.039265994,
0.0151523529,
0.0037302938,
-0.0190499537,
0.0771221891,
-0.0228642058,
0.0082820365,
0.0076766144,
-0.0165265556,
0.0400293618,
-0.0141602065,
0.0512770675,
0.0173381604,
0.0416531824,
-0.0550031699,
-0.000275902,
-0.0793009251,
-0.0413498245,
-0.0529945269,
0.0476522148,
-0.0463712029,
-0.0260778237,
-0.0708474889,
0.0159970913,
-0.01453989,
-0.0095556546,
0.0304126833,
0.0561429374,
-0.0403693914,
0.0173371434,
-0.054754436,
-0.0183645505,
-0.0580146983,
0.0141475135,
0.0367772542,
-0.0543790348,
0.018765837,
0.0118332077,
0.0217039399,
0.0378071591,
-0.0095784217,
0.0673607737,
0.0013667415,
-0.0013232736,
-0.0204937402,
-0.0581967048,
0.0028255461,
-0.0360653549,
-0.0029793237,
-0.0165478177,
0.0632572696,
-0.0343636908,
-0.0224522036,
-0.0354461484,
0.0064626713,
0.0600986518,
0.0086278701,
0.0229584798,
-0.0031845332,
0.0096367775,
0.0355767496,
0.0176503863,
0.0265952386,
-0.0106450124,
-0.0180999935,
0.0232895333,
0.008505349,
0.0191546641,
0.0740164891,
-0.0233090017,
-0.0003384874,
-0.0456072055,
0.0508078672,
0.0315991268,
-0.0152182933,
0.0354082659,
-0.0365108214,
0.0758834258,
-0.0490130745,
0.0170196332,
0.0093385652,
0.0413835682,
0.0482862294,
0.0044376552,
0.0512415022,
0.0259547755,
0.019982731,
-0.0201112647,
0.0180850755,
0.0227366257,
-0.0272361543,
0.0042297547,
-0.0069651376,
-0.0040264842,
-0.0355996341,
-0.0478769243,
-0.0056287623,
0.0129754255,
-0.0200652368,
0.0019496063,
0.0083596492,
0.0442742072,
-0.0632257387,
-0.048841536,
-0.0003493227,
-0.0307523403,
-0.0442060605,
-0.0723174438,
0.0455943532,
0.0982724279,
0.0932153314,
-0.0230268873,
-0.0198932514,
-0.0232296549,
-0.012328743,
0.029507013,
0.0309185646,
-0.0317587592,
-0.0343236439,
-0.0398935415,
-0.0353421979,
-0.0082116509,
0.0044363821,
0.0327247232,
-0.0071981265,
0.0266537759,
0.0050107939,
0.0098077245,
-0.0323997699,
0.0490931049,
-0.0171967112,
-0.0685664788,
-0.0103447409,
-0.0050165309,
0.0224370863,
0.0529245175,
0.0032377192,
0.0542808324,
0.0519903786,
-0.037424989,
-0.0268173367,
-0.0103669902,
0.0138789192,
0.0050291573,
0.0134251947,
-0.0235596709,
0.0214624126,
-0.0143842762,
-0.0197477043,
-0.0328566059,
-0.0490765423,
-0.0349079967,
0.0124425068,
-0.0598144531,
-0.0233316571,
0.0227573682,
0.0083024809,
-0.0512577519,
0.0325097144,
0.0255136508,
-0.0318245888,
0.0434006974,
-0.0277412031,
0.0192494579,
-0.0278367121,
0.0363113657,
0.0102272788,
-0.0351512358,
0.0280654375,
0.0584556758,
0.0172649212,
-0.0308385603,
-0.0134996241,
-0.0578233749,
-0.0045852978,
0.0053200554,
-0.0042410456,
0.0088084312,
-0.0077938805,
-0.0018724239,
-0.0395230353,
-0.0822285786,
-0.0155767193,
0.0207981113,
-0.0157789141,
0.0383479483,
0.041590061,
-0.0245925207,
-0.045470532,
0.0024857505,
-0.0116569214,
0.004056422,
0.0675237998,
0.0067212326,
-0.0595893338,
-0.0023053773,
-0.057514172,
0.0172581207,
0.016999444,
-0.013983476,
0.0428507626,
-0.0192922931,
0.038328778,
0.0267896149,
0.0179761481,
0.0214416347,
0.0280457456,
0.022606954,
-0.073759608,
0.0307025425,
-0.0166948345,
-0.041737888,
0.0466077626,
-0.0437406451,
0.0416747928,
-0.0115324212,
0.0276530888,
0.0092995903,
-0.0721733272,
0.0387865901,
0.0085523967,
-0.011955604,
-0.0386153758,
-0.0061047571,
-0.013966308,
0.0594041385,
0.0006276564,
-0.0383047163,
0.0952591747,
0.0133622214,
0.0116005559,
-0.0679531619,
-0.011941405,
-0.0327822752,
0.0274101458,
-0.0192296635,
0.0377484486,
0.0037660568,
-0.0003286556,
-0.0284723975,
-0.0035952858,
0.0435044244,
-0.0699797571,
-0.0061596106,
-0.0315129086,
0.0346185938,
0.0014526416,
-0.0348800533,
-0.0697887391,
-0.0141281178,
0.0525034703,
0.0132148303,
0.0301782563,
0.0096500479,
-0.0509587824,
0.0210281275,
-0.0031667894,
0.0404616073,
0.0277309734,
-0.0492684618,
0.0136109833,
0.0163418446,
0.0779460818,
-0.0282951631,
0.0030054317,
0.0109437201,
0.0435217656,
0.0271525979,
0.0187569819,
-0.0126758805,
0.0069674421,
-0.0210019294,
-0.0128030013,
-0.0126183331,
0.012999448,
0.0376113728,
0.0136583736,
-0.0384441502,
0.0018000039,
0.0572616346,
0.0529113375,
0.006735797,
-0.0025259964,
-0.0283838008,
-0.0183771383,
-0.0069224942,
0.0804573074,
-0.0766269192,
0.0182578657,
-0.0180772562,
0.0605846792,
-0.0539209805,
-0.0211206619,
-0.0189941656,
-0.0673125833,
0.0575069115,
-0.0130534088,
-0.011442095,
-0.0496082194,
-0.0100488169,
-0.0297337007,
0.0070197275,
0.0003498888,
0.0288769603,
-0.0171468649,
-0.0266614035,
0.0113432147,
-0.0331150666,
-0.0120890243,
0.0065401057,
-0.0156005733,
0.0518152229,
0.1060787663,
0.0168346204,
0.0485073254,
-0.0244581457,
-0.0008243139,
-0.0092399372,
0.0133084478,
0.0205546711,
0.0397383273,
0.0124141993,
0.0173449181,
-0.055889152,
0.0316341743,
-0.0494552776,
0.0075891265,
0.0183023307,
0.0325962156,
-0.0270364545,
-0.0208463892,
-0.0749730021,
0.0455618426,
0.010659582,
0.0529761985,
0.016132459,
0.0031227726,
-0.0452402048,
-0.0367507935,
0.0465217754,
-0.060642682,
-0.0033125505,
-0.0190195311,
-0.0386791155,
0.0009734537,
-0.1315616369,
-0.0417700298,
0.0043015797,
0.0104661537,
0.0113391094,
-0.0061997473,
-0.0142433457,
0.0047157104,
0.0371999964,
-0.010762468,
-0.0039232327,
-0.0102413567,
-0.0342627913,
-0.0403252095,
0.0288685132,
-0.0174089689,
-0.0102848681,
0.0603551194,
0.0055527836,
-0.0371539332,
-0.014118515,
0.0458478034,
0.0262839328,
-0.0015094319,
-0.0322157852,
0.0172957517,
0.0162038412,
0.0408065356,
-0.0073101143,
-0.0426482335,
-0.0222506355,
0.0165627915,
-0.0335238762,
-0.0092298882,
-0.0799613297,
0.0299879145,
-0.0381114148,
0.0601275712,
-0.0004068346,
-0.0477344356,
0.0885521695,
0.0403925106,
-0.0524506271,
0.022964824,
0.0197711289,
-0.0503884926,
0.017085636,
-0.0364646986,
-0.0011151502,
-0.0291184913,
0.0069267652,
0.0143908206,
-0.057424631,
0.0242580976,
0.0295889154,
-0.0801139399,
0.0713226199,
-0.0367936604,
-0.0338288248,
-0.0135979913,
0.0507395528,
-0.005693797,
-0.0448412187,
-0.0046020043,
0.0446376987,
-0.0314466171,
0.0563596077,
-0.0050998488,
0.0680682138,
-0.0228123851,
-0.0097462311,
-0.0048524025,
0.0178823695,
-0.0662719682,
-0.0847684667,
-0.0415920429,
0.048445154,
0.0141429938,
0.0480826236,
0.005385438,
-0.0392869711,
0.0072249924,
-0.0141198216,
0.0273670293,
0.0220100842,
-0.0154827237,
0.0175157413,
0.0323372893,
-0.0062123626,
-0.0537746064,
0.056434866,
-0.0047605475,
0.0078273257,
-0.0906277373,
-0.0373412594,
-0.0545229614,
-0.0390302576,
0.0030811029,
0.0202733502,
-0.0118350107,
-0.0166534223,
-0.0475764088,
-0.0209277626,
-0.0377462804,
0.0314535089,
0.0046991892,
0.0035939075,
0.0298304204,
0.0729096681,
0.0012308081,
-0.0662894845,
-0.0612887368,
-0.0163604654,
-0.0884187371,
0.0021515037,
-0.0338479094,
-5.842066599e-33,
-0.0144281657,
-0.0241260082,
0.046773728,
-0.0117566399,
0.0086951163,
-0.025560515,
-0.0486914292,
-0.0057477779,
-0.002932373,
0.0149694672,
-0.0020304839,
0.0043285294,
0.0106893014,
-0.0054095639,
0.0149564007,
-0.0251139645,
0.0505549498,
0.0027985701,
-0.0183486473,
0.0382502191,
0.0168320518,
0.0325997286,
-0.0445287041,
-0.0512552857,
0.044429101,
-0.0187943298,
0.0015029508,
0.009697618,
-0.0270366129,
0.0946451426,
-0.0036067422,
0.0299899895,
0.0402435325,
-0.0095378347,
0.0101298941,
0.021293737,
-0.0758387074,
-0.0325077586,
-0.0200370457,
-0.0676870197,
-0.0545483567,
-0.0404609777,
0.0271806978,
-0.0452966988,
-0.0020728984,
0.0379278325,
0.0404257178,
-0.0090251509,
-0.0069905967,
-0.0590771921,
-0.072551243,
-0.0096941227,
-0.0190809388,
-0.0160778351,
-0.048410207,
0.0965808257,
0.0097717643,
0.0017128221,
-0.0142317154,
-0.0032585263,
-0.0198114868,
-0.0099243922,
0.0280100014,
0.0024139041,
-0.0250140764,
-0.00003045,
-0.0164156165,
0.0001133284,
-0.0210912433,
-0.0354597792,
-0.0405811854,
0.038537249,
0.0035932234,
0.0273744408,
0.0216746721,
-0.0087087294,
-0.056544058,
0.1056933254,
-0.0031818037,
0.0660834014,
0.012987718,
-0.0147293545,
-0.0521863252,
0.0183830746,
0.0128751015,
0.0528676324,
-0.0380275883,
-0.0445187949,
0.0431739502,
0.0209072977,
-0.028252501,
-0.0500945002,
-0.0147466483,
-0.0218059868,
0.0469242856,
-0.004626527,
-0.0062389956,
-0.0190760028,
-0.0237452406,
-0.0076919463,
-0.0614878871,
0.0124375587,
0.0715324879,
-0.0176436491,
-0.0086353133,
0.0063845576,
0.0073637022,
-0.0224658027,
-0.0460965484,
0.0284073316,
0.0018515022,
0.0702646747,
-0.0024284634,
-0.0017143262,
-0.0136575252,
0.0290691108,
0.003439181,
0.0246644281,
0.0042424821,
0.0161374323,
-0.0219123699,
-0.0134603437,
0.0330461189,
-0.0207688678,
-0.0388020203,
0.0527827442,
-0.0320728421,
0.0708842352,
0.0266407412,
-0.0378482901,
-0.0250160433,
0.0189983901,
2.601e-7,
0.0636882558,
0.0025764299,
0.0423672423,
0.0412426814,
0.0084810657,
-0.0252875071,
0.0263068825,
0.0258134101,
0.0036105048,
-0.0108761741,
0.0379676335,
0.0030970383,
0.0168789588,
0.0184979998,
-0.065028429,
-0.0619187579,
-0.0396073647,
-0.0491066389,
-0.0551508665,
0.0020039184,
0.0800947696,
0.0089093056,
-0.029964827,
-0.0260079335,
0.0171281062,
-0.0203303415,
0.040233817,
0.0313027278,
0.0122922445,
0.0464544222,
0.0832681209,
0.0107648484,
0.0075920625,
0.0092791673,
0.0035509556,
0.0138910823,
-0.0127006574,
0.0421805643,
-0.0175616089,
-0.0385214537,
0.0420033373,
-0.0146498149,
0.037756335,
-0.0319866836,
0.0282821767,
-0.0146982139,
-0.000255646,
-0.0161759686,
-0.0295507405,
-0.0008234866,
0.0343445167,
-0.0430029221,
-0.0396158509,
-0.0029097118,
-0.0119374683,
0.0015789637,
0.0336122699,
-0.0242522191,
0.0473966785,
-0.0031180838,
-0.0291894283,
-0.0392314978,
-0.0161347874,
-0.0276477039,
-0.0361246951,
-0.058803428,
0.0316025838,
2.108141211e-34,
0.0392010659,
-0.0149999065,
0.0018040878,
0.0242982116,
0.0605974719,
0.0194877647,
0.0108566852,
0.0237431265,
0.0091682402,
-0.0644088164,
-0.0239531547
] |
http://arxiv.org/abs/2402.01716v1 | http://arxiv.org/pdf/2402.01716v1 | Bloom-epistemic and sentiment analysis hierarchical classification in course discussion forums | 2024-01-26T08:20:13 | 2024-01-26T08:20:13 | H. Toba, Y. T. Hernita, M. Ayub, M. C. Wijanto | cs.CY | cs.CY, cs.CL, cs.LG, I.2.7 | Online discussion forums are widely used for active textual interaction
between lecturers and students, and to see how the students have progressed in
a learning process. The objective of this study is to compare appropriate
machine-learning models to assess sentiments and Bloom\'s epistemic taxonomy
based on textual comments in educational discussion forums. Our proposed method
is called the hierarchical approach of Bloom-Epistemic and Sentiment Analysis
(BE-Sent). The research methodology consists of three main steps. The first
step is the data collection from the internal discussion forum and YouTube
comments of a Web Programming channel. The next step is text preprocessing to
annotate the text and clear unimportant words. Furthermore, with the text
dataset that has been successfully cleaned, sentiment analysis and epistemic
categorization will be done in each sentence of the text. Sentiment analysis is
divided into three categories: positive, negative, and neutral. Bloom\'s
epistemic is divided into six categories: remembering, understanding, applying,
analyzing, evaluating, and creating. This research has succeeded in producing a
course learning subsystem that assesses opinions based on text reviews of
discussion forums according to the category of sentiment and epistemic
analysis. | Bloom-epistemic and sentiment analysis hierarchical classification in course discussion forums: Online discussion forums are widely used for active textual interaction
between lecturers and students, and to see how the students have progressed in
a learning process. The objective of this study is to compare appropriate
machine-learning models to assess sentiments and Bloom\'s epistemic taxonomy
based on textual comments in educational discussion forums. Our proposed method
is called the hierarchical approach of Bloom-Epistemic and Sentiment Analysis
(BE-Sent). The research methodology consists of three main steps. The first
step is the data collection from the internal discussion forum and YouTube
comments of a Web Programming channel. The next step is text preprocessing to
annotate the text and clear unimportant words. Furthermore, with the text
dataset that has been successfully cleaned, sentiment analysis and epistemic
categorization will be done in each sentence of the text. Sentiment analysis is
divided into three categories: positive, negative, and neutral. Bloom\'s
epistemic is divided into six categories: remembering, understanding, applying,
analyzing, evaluating, and creating. This research has succeeded in producing a
course learning subsystem that assesses opinions based on text reviews of
discussion forums according to the category of sentiment and epistemic
analysis. | [
0.0554985292,
-0.0135684526,
-0.0137658771,
0.0443710946,
-0.0564372204,
0.0242948458,
0.0048155016,
-0.008922603,
0.022658322,
-0.1108082458,
0.0166069604,
-0.0292609371,
-0.007379055,
0.0227085892,
-0.0289610643,
-0.0342134945,
0.0242898017,
0.0334357433,
0.0150153991,
0.0240312479,
0.0462030098,
0.0008592107,
-0.0376027152,
0.0204436444,
0.0053111934,
-0.0082781129,
0.0157117099,
-0.0441561788,
0.0296270885,
-0.0005164716,
0.0191712826,
0.0356848426,
-0.0342438854,
0.0273023732,
0.0000018852,
-0.043577645,
-0.0372541398,
-0.0045108371,
0.0259676315,
0.0173197016,
0.0545547158,
0.032253921,
-0.006435845,
0.0454350002,
-0.0532196686,
0.0125505393,
0.0419104993,
0.0923392624,
0.0583134145,
0.0328679681,
-0.0054100975,
-0.0835345834,
-0.0032628558,
0.0493576862,
0.0153828487,
0.027984513,
0.04192679,
0.0122707505,
-0.0383312181,
0.0386266597,
-0.0029888647,
-0.0070382026,
-0.0436051972,
-0.0555973053,
0.0333395191,
-0.0265477449,
-0.039039176,
-0.0319534726,
0.0370815061,
0.0074804383,
0.0737135857,
-0.0050444105,
0.0060038744,
0.0182832815,
0.0081332512,
0.0550982431,
-0.0247933865,
0.0150964763,
-0.0278454982,
0.0192639772,
-0.0009388519,
0.0249665733,
0.0002866333,
0.0111139733,
0.0119814584,
0.0259862542,
-0.0367473997,
-0.0339918286,
0.0134117603,
0.0516352355,
0.0276694521,
0.0078620669,
0.0153032402,
0.0045762565,
0.0284557249,
0.0254200548,
0.0222429596,
0.0081942445,
0.0223445836,
-0.093257606,
-0.0118081663,
0.0161163881,
0.022484716,
-0.0088369176,
-0.015049194,
-0.045925878,
0.0415238328,
-0.0221289676,
0.0169020537,
0.0189212207,
-0.1100419611,
0.0146156754,
-0.0044430881,
0.082054846,
0.0232035015,
0.0133843664,
-0.0848893225,
0.0166479349,
-0.0319343768,
-0.0161888581,
-0.0004240871,
-0.0271808263,
0.0702306181,
0.061360769,
-0.0462293476,
0.0238079764,
-0.0315931104,
0.0402660593,
-0.0589776933,
0.0106898136,
0.056646537,
0.0380956046,
0.0146815637,
0.0310996994,
-0.0133011891,
0.0330569036,
0.0285157803,
0.0115035186,
-0.0893581957,
-0.0026615376,
-0.0065322276,
0.0153739136,
0.0042482088,
-0.0441551283,
-0.0276862178,
-0.0022434543,
-0.0181700811,
0.0847665891,
0.0244072322,
0.0271475688,
-0.0051438226,
0.0217702854,
-0.0581892096,
-0.0094066123,
0.0804644898,
0.0255204905,
0.0261676293,
0.0080537293,
-0.0031370001,
-0.0034989619,
0.0509493612,
-0.0105573218,
0.0236520134,
-0.0251825284,
-0.0078344457,
-0.0693522245,
-0.0070273085,
0.0136899706,
-0.0488457046,
-0.0209187455,
-0.0741270483,
0.0476690941,
-0.0063186255,
-0.0027210703,
-0.0227410849,
0.0187898744,
-0.0430573933,
0.0349949934,
-0.0220853798,
0.0364050679,
0.0425036587,
-0.0097103557,
0.0350575894,
-0.0021033157,
0.0022217473,
-0.024391463,
0.0371422656,
-0.0408586971,
0.0123619977,
-0.0995378047,
-0.0349317864,
-0.0006122936,
0.0134743936,
-0.0361705981,
0.0621897802,
-0.0265264455,
-0.0221515503,
0.004023972,
0.0421183147,
-0.0454613306,
-0.0007508449,
-0.0591426119,
0.0375957638,
0.0286893584,
-0.0360151641,
-0.0554411113,
-0.004990709,
-0.0258969404,
0.0120576899,
0.0338805728,
0.0207194518,
-0.0347914174,
-0.0601255894,
-0.04049889,
0.0083783641,
-0.0106562506,
0.0022933178,
-0.0152266482,
-0.0417419784,
0.0154562984,
-0.0131934136,
0.029658528,
-0.0561617278,
0.0195465926,
0.0492961928,
0.0205083676,
-0.0185271855,
0.0015557928,
0.0467777476,
0.0342928097,
-0.0236091129,
0.0033770422,
0.0316688456,
0.0115904799,
-0.0683448911,
-0.020795228,
0.0252863709,
-0.0301136728,
0.0397107117,
-0.0963848755,
-0.0113813737,
-0.0254393294,
-0.0249258615,
-0.0117149428,
0.0412798226,
-0.050450746,
0.0491692275,
-0.0163906831,
0.0101604927,
-0.0698636547,
-0.0455439426,
-0.0347483419,
0.0270369351,
-0.0047442075,
0.0343114026,
-0.0336114615,
-0.0600842126,
0.0639951676,
0.0166185722,
-0.0337785482,
0.0005038861,
-0.0048080059,
0.0142939342,
0.0344571657,
-0.0228895433,
-0.0521030165,
-0.0026441475,
-0.1298299581,
-0.0423358455,
0.0195053499,
0.0450417772,
-0.0035296311,
-0.0720003098,
0.0127726914,
-0.0004696735,
-0.0142057715,
0.0181819052,
-0.0454444401,
-0.0000613973,
0.0458428897,
-0.001186042,
0.0501443222,
-0.0397067182,
-0.0176033992,
-0.0345226713,
0.0232078563,
0.0970414877,
-0.033252798,
-0.0031096123,
0.0083876299,
-0.0710718334,
-0.0347876996,
-0.0139859049,
-0.0408748426,
0.0022693588,
-0.0062492755,
-0.0258073751,
0.0092658829,
0.069324635,
0.0518501773,
0.0209791809,
0.0246993639,
0.0054795705,
-0.0038703117,
-0.0365625173,
0.0458144285,
-0.0180371515,
-0.0094140396,
0.0046808021,
0.0008714699,
0.0185930952,
-0.0236017015,
-0.0301363971,
0.0507849492,
-0.0368528292,
0.0254622214,
-0.046925921,
-0.0246880073,
0.0247910116,
0.0011469433,
0.0310224686,
-0.022408925,
0.0506069027,
0.014142788,
0.0362283476,
-0.0898552984,
-0.0360252894,
-0.0347756967,
-0.0212405715,
-0.0159766544,
0.0216304921,
-0.0299467128,
0.0129211815,
-0.0012178057,
0.0148890046,
0.0278768092,
-0.0383961387,
-0.002484312,
-0.0103027485,
-0.0018680705,
0.0045568477,
-0.0006580449,
-0.013715066,
0.0096113477,
-0.014560055,
-0.023425715,
0.0152605278,
-0.0054850667,
-0.0165720303,
0.0156839844,
-0.0082544154,
-0.0461410768,
0.0072385152,
-0.0318028964,
0.0448000692,
0.0188138504,
-0.0160536282,
-0.0461031087,
0.0256366674,
0.0050805258,
0.0070093288,
0.0730648488,
-0.0187308807,
-0.0230868831,
-0.009111166,
-0.0658256933,
-0.0124523509,
-0.0428507254,
-0.0244510621,
-0.0362796187,
-0.0182389356,
-0.0365842469,
0.0037497971,
0.0984726548,
0.0474532023,
-0.0384475701,
-0.0235919245,
-0.04021753,
-0.0215634685,
0.0078856777,
-0.0119217699,
-0.0491308309,
0.0093247751,
-0.0643329546,
0.031022463,
-0.0397607833,
-0.0115406718,
0.0494054928,
-0.0944433585,
0.0044895206,
-0.0134171108,
-0.0287588965,
-0.0046780761,
-0.0538893268,
-0.006911749,
0.0073304898,
-0.0226997025,
-0.0347485729,
-0.0008710012,
0.0351374038,
0.0712460205,
-0.0521316268,
-0.021720238,
0.008365538,
0.0003567714,
0.0328923278,
0.0070937863,
0.0745875314,
-0.0552078076,
-0.0328371041,
0.0449856557,
-0.0409087762,
-0.0096620452,
-0.014924706,
0.0559676886,
-0.0251901112,
0.0155770024,
-0.0366644412,
-0.0194552913,
-0.0202432778,
-0.0095092105,
0.0425110497,
0.0795276165,
-0.0226076003,
-0.0363568738,
-0.0798765719,
0.0628951788,
-0.0188619513,
0.0534838438,
-0.0200740024,
-0.0419282839,
-0.0013448726,
-0.0231215078,
0.0062240683,
0.0056804842,
-0.021317631,
-0.0256168861,
-0.0062883049,
0.0021843936,
-0.0884241387,
-0.0262702182,
-0.0154661117,
0.0250935536,
-0.0168988109,
0.0321826562,
0.0518459864,
0.041826684,
-0.0595892593,
-0.0050060088,
0.0425414741,
-0.0037080741,
-0.0462683439,
-0.046741806,
0.0787165537,
0.0054399711,
-0.0319260955,
-0.0405597761,
0.0400208198,
-0.0551512986,
-0.0158959664,
0.0422718488,
0.0191695839,
-0.0079877051,
0.0105614327,
-0.0176356863,
0.0411307961,
0.0548953786,
0.0574122816,
-0.0112165976,
0.0275075324,
0.0610871762,
0.0033248961,
-0.0161569268,
0.0228458475,
0.03451306,
-0.0116435541,
-0.0319407694,
-0.0563374273,
-0.0050169057,
-0.0107537089,
0.0029858344,
0.0361317173,
0.0350792445,
-0.0347975604,
0.0015630389,
0.0021323797,
-0.0067950394,
-0.0271044914,
-0.0045385309,
-0.0014205212,
0.0092531005,
0.0178733207,
-0.0338498056,
-0.0009490442,
-0.0260320269,
-0.0246233009,
0.0251569469,
0.0484746471,
0.0424762107,
0.0289387349,
-0.0105709312,
-0.0050679795,
-0.0247997213,
-0.0424513444,
0.0401005261,
-0.015812451,
-0.0025617715,
0.0890200958,
0.0158505607,
-0.0075967438,
0.0179731436,
0.0108511345,
-0.030800961,
0.0338803455,
-0.026784163,
0.0216712654,
-0.0172626674,
0.011512408,
0.0058628628,
0.0114219384,
-0.0104799289,
-0.0051689148,
0.0705974177,
-0.0217117593,
-0.0175725222,
0.0267287958,
-0.015885355,
-0.0356811471,
0.0257162601,
0.0048392825,
-0.0632733479,
-0.0079558603,
0.0107504679,
-0.0564044639,
-0.0240970477,
-0.0041929749,
0.0124244252,
0.0143493516,
0.0186914913,
0.0200898163,
-0.0503270105,
-0.0246549752,
0.0346169807,
0.0118346941,
0.0661714002,
-0.0224794298,
0.0374372713,
0.0529518761,
0.031579867,
0.053312853,
-0.0630227253,
0.0075006587,
-0.0592278019,
0.0029511831,
-0.0452779233,
-5.519284299e-33,
-0.0188341998,
-0.0391311496,
-0.0510070808,
0.0691115782,
-0.0062265736,
-0.00739562,
-0.0113904448,
0.0366572104,
-0.0159965102,
0.0432179831,
0.0434319004,
0.0487733334,
-0.005735775,
0.0264122784,
0.0402970053,
-0.0104880277,
0.0245896112,
-0.019478865,
0.0181613434,
0.024532102,
0.0556608923,
-0.0130063659,
0.0278679226,
-0.1144656688,
0.0559046827,
-0.0134923933,
-0.0000453476,
0.0247173309,
-0.0045614378,
0.0453232639,
-0.0235508196,
0.0567886941,
0.0368541293,
0.0226980876,
0.0315636881,
0.0315165594,
-0.0061377021,
-0.0316159129,
-0.0135926949,
-0.0156012783,
0.0063764146,
0.0040825587,
0.0287722126,
0.0215505715,
-0.04999616,
0.0113088703,
0.0827243552,
0.0051819272,
0.0335063301,
-0.0715174377,
-0.0686654821,
0.023711985,
-0.0346441008,
0.042640198,
-0.0226340387,
0.0354437903,
0.0247488022,
0.0074790074,
-0.0558635704,
0.0391289815,
-0.0110611226,
-0.0234649051,
0.0024742633,
-0.0093381386,
-0.0131604215,
0.0654169768,
0.0416872874,
0.0174110737,
-0.0318751819,
0.0214679092,
0.0282321963,
0.008288132,
0.0717555955,
0.0726554692,
0.0117573543,
-0.0292732157,
-0.0738611221,
0.0097954394,
0.0591463596,
-0.0163241532,
0.0445402451,
-0.0220722202,
-0.0086104851,
0.0039921463,
0.0041855383,
-0.0481293425,
-0.0070668138,
-0.0060405494,
-0.030362919,
-0.0526334271,
-0.0375785157,
-0.019868562,
-0.0160160884,
-0.0400526449,
-0.0316642374,
0.0300836619,
-0.022853924,
0.0547311045,
-0.0071674939,
0.0096006449,
-0.0464312993,
0.0015083631,
-0.0280266982,
-0.0110897012,
0.0194905009,
-0.0046975762,
0.0291867647,
-0.0214969143,
-0.0321167856,
-0.0185543355,
0.0104040923,
-0.0118546756,
-0.0050679194,
0.0527505539,
-0.0188093912,
0.0315249301,
-0.0002746069,
0.0329298601,
0.0073796757,
0.0638425648,
-0.0049991426,
-0.0191660952,
0.0043408643,
-0.0401409715,
-0.010020039,
-0.0347239748,
-0.0141346827,
0.0149864899,
-0.0064350558,
-0.0764904767,
0.0154733015,
-0.024705464,
2.64e-7,
0.0659576952,
0.0808577761,
-0.0324455537,
-0.0797560215,
0.0310174115,
0.0430435129,
0.0252309758,
0.0032626835,
0.0477556027,
0.0206651092,
0.0048154355,
-0.0151097635,
0.0084920516,
0.0252834279,
-0.0814911947,
-0.1096437573,
-0.0303220414,
-0.0540841669,
0.0072668637,
0.0859217122,
0.1046473235,
0.0457119718,
0.0326522328,
0.0184967313,
-0.0483365469,
-0.0145850992,
0.0199166667,
-0.0159782469,
0.0189091899,
0.0512554385,
-0.0124954069,
0.0027192181,
0.0159701966,
0.0306910053,
-0.0085517894,
0.0049005072,
0.0348693766,
0.0094598392,
0.0042012949,
-0.0012981803,
0.0349251665,
0.0308702793,
0.0444474071,
-0.0153324408,
0.0083660567,
-0.0439155176,
-0.0295920931,
-0.0289384797,
-0.0364015475,
-0.0427691936,
0.0581419468,
-0.0106562264,
-0.02236332,
0.0234677903,
-0.0323111713,
-0.0512030795,
-0.0175821278,
-0.063792482,
0.0331773646,
0.0028954274,
-0.0288570542,
-0.0488587134,
0.0457745343,
0.0277304221,
0.0478665866,
-0.004294008,
0.0344914645,
1.904139447e-34,
0.0414489135,
-0.0297657419,
-0.0081904,
0.0039650644,
0.0352174118,
0.0466618687,
-0.011807275,
-0.0026656559,
0.0032232923,
-0.0189094394,
0.0270434413
] |
http://arxiv.org/abs/2401.14698v2 | http://arxiv.org/pdf/2401.14698v2 | Under the Surface: Tracking the Artifactuality of LLM-Generated Data | 2024-01-26T07:53:27 | 2024-01-30T05:36:06 | Debarati Das, Karin De Langis, Anna Martin-Boyle, Jaehyung Kim, Minhwa Lee, Zae Myung Kim, Shirley Anugrah Hayati, Risako Owan, Bin Hu, Ritik Parkar, Ryan Koo, Jonginn Park, Aahan Tyagi, Libby Ferland, Sanjali Roy, Vincent Liu, Dongyeop Kang | cs.CL | cs.CL, cs.AI | This work delves into the expanding role of large language models (LLMs) in
generating artificial data. LLMs are increasingly employed to create a variety
of outputs, including annotations, preferences, instruction prompts, simulated
dialogues, and free text. As these forms of LLM-generated data often intersect
in their application, they exert mutual influence on each other and raise
significant concerns about the quality and diversity of the artificial data
incorporated into training cycles, leading to an artificial data ecosystem. To
the best of our knowledge, this is the first study to aggregate various types
of LLM-generated text data, from more tightly constrained data like "task
labels" to more lightly constrained "free-form text". We then stress test the
quality and implications of LLM-generated artificial data, comparing it with
human data across various existing benchmarks. Despite artificial data's
capability to match human performance, this paper reveals significant hidden
disparities, especially in complex tasks where LLMs often miss the nuanced
understanding of intrinsic human-generated content. This study critically
examines diverse LLM-generated data and emphasizes the need for ethical
practices in data creation and when using LLMs. It highlights the LLMs'
shortcomings in replicating human traits and behaviors, underscoring the
importance of addressing biases and artifacts produced in LLM-generated content
for future research and development. All data and code are available on our
project page. | Under the Surface: Tracking the Artifactuality of LLM-Generated Data: This work delves into the expanding role of large language models (LLMs) in
generating artificial data. LLMs are increasingly employed to create a variety
of outputs, including annotations, preferences, instruction prompts, simulated
dialogues, and free text. As these forms of LLM-generated data often intersect
in their application, they exert mutual influence on each other and raise
significant concerns about the quality and diversity of the artificial data
incorporated into training cycles, leading to an artificial data ecosystem. To
the best of our knowledge, this is the first study to aggregate various types
of LLM-generated text data, from more tightly constrained data like "task
labels" to more lightly constrained "free-form text". We then stress test the
quality and implications of LLM-generated artificial data, comparing it with
human data across various existing benchmarks. Despite artificial data's
capability to match human performance, this paper reveals significant hidden
disparities, especially in complex tasks where LLMs often miss the nuanced
understanding of intrinsic human-generated content. This study critically
examines diverse LLM-generated data and emphasizes the need for ethical
practices in data creation and when using LLMs. It highlights the LLMs'
shortcomings in replicating human traits and behaviors, underscoring the
importance of addressing biases and artifacts produced in LLM-generated content
for future research and development. All data and code are available on our
project page. | [
0.056055177,
0.0435563624,
-0.0436539911,
0.0586528964,
-0.0326042026,
0.0170438103,
-0.0187253337,
-0.0077066626,
0.0103754122,
-0.0549404435,
-0.0423575155,
-0.0006804212,
-0.0040321741,
0.0100430772,
0.0290565565,
-0.0432342924,
0.0369342119,
-0.0343504734,
-0.0048679309,
0.0081813131,
0.0049272869,
0.0198016856,
0.0033920305,
0.0497724414,
-0.0314060673,
-0.0195892863,
-0.0083261514,
0.0057749576,
-0.0089719072,
-0.0285940655,
0.0074702292,
0.0413925126,
0.0062602819,
0.0339735746,
0.0000021303,
-0.0538579747,
0.0367069393,
-0.0121793346,
-0.0512101948,
0.0313687399,
0.0687782019,
0.0259189717,
-0.0300110355,
-0.0005841398,
-0.0532205626,
0.0052878093,
0.0260362085,
0.0141268875,
0.0521889105,
0.072596103,
-0.0016105471,
0.0423519872,
0.0006889449,
0.0121151442,
0.0334483795,
-0.0055497815,
0.0135153206,
0.0125624044,
0.0258580726,
0.0460525341,
0.0096614519,
0.0251658782,
0.0202429872,
0.0007449607,
0.0058773216,
0.0322789289,
-0.0163415447,
-0.0488071926,
-0.0201891791,
0.0717543513,
0.0270390585,
-0.0118603399,
0.0206960123,
-0.0286044646,
0.0218957886,
-0.0032280739,
-0.0283974912,
0.0688428208,
-0.0099105919,
0.0351678841,
0.0103029637,
-0.0067905667,
0.0230502691,
0.0193935335,
0.0136961117,
0.013596545,
-0.0248246361,
-0.0317983888,
0.0085710455,
-0.0011767523,
0.0004837884,
-0.0777404159,
-0.0068584839,
0.0361025855,
-0.0204823911,
0.0383212194,
-0.0744688213,
-0.0367965586,
0.0536568128,
-0.0541011281,
0.0629486293,
0.0236168317,
-0.0292432271,
0.0062545142,
0.0027353363,
0.054077737,
-0.0056522931,
0.0397993475,
-0.0428144075,
0.0247376561,
-0.0360123999,
-0.0088359024,
0.0036582386,
0.0248143412,
-0.0313866101,
-0.0000095942,
-0.0084466022,
0.0420833193,
-0.0201755762,
0.0311366674,
0.0077922186,
0.0190337095,
-0.0214738119,
0.0263823867,
-0.0668169186,
-0.0465133488,
-0.0353939943,
0.0255782474,
0.0195978452,
-0.059748169,
0.0354921184,
0.0323568173,
-0.0272760745,
-0.0228425022,
0.0237436723,
0.0945023075,
-0.0165185593,
-0.0099583762,
-0.054441046,
-0.0129822027,
-0.0031416079,
-0.0494094416,
0.0147157293,
-0.0565357618,
0.02033378,
0.0119947847,
-0.017295504,
-0.0375793427,
0.0124873398,
0.055026345,
-0.0277939048,
0.0585893989,
-0.0379508622,
0.0381935947,
0.0397898257,
-0.0018020145,
-0.0083855446,
0.0696130916,
0.0736116767,
0.0326860398,
0.0335268453,
-0.0226285085,
0.0274348687,
0.0532960892,
-0.0101905083,
-0.0003484148,
-0.0356402844,
0.0090081198,
-0.0020494445,
0.0363489576,
0.0648314804,
0.0673034266,
-0.0220667012,
0.0084139481,
0.0905390903,
0.0693492293,
-0.0235822778,
0.0223095678,
0.0366500877,
0.0281859729,
0.0302476883,
0.021777954,
-0.0607480295,
0.0077616232,
0.0179617237,
0.0080541437,
-0.0287115127,
0.0213854332,
-0.0171549171,
-0.0664257258,
-0.026454322,
-0.0197196212,
0.0199227426,
-0.0037868726,
0.0202599149,
0.0101643205,
-0.0400757715,
0.0108048301,
0.027920017,
-0.0627370998,
-0.0004779911,
-0.0565906093,
0.0436915718,
0.0940835774,
0.0406573117,
-0.035011135,
-0.0122489035,
-0.0042300667,
0.0134956073,
0.0797568485,
0.0684577227,
-0.026825849,
-0.0267348401,
-0.0484549738,
-0.0266493559,
-0.0369449295,
-0.0060508926,
-0.0038527884,
0.0073316274,
-0.0136192404,
0.0084497109,
-0.030041378,
-0.1084795147,
0.0730985999,
-0.0687223375,
-0.0458521061,
-0.0072555528,
-0.0118691064,
0.0062861186,
0.0202703234,
-0.0091338214,
0.0183588285,
0.0408439003,
-0.0059053586,
0.0080383839,
0.0068110269,
-0.0531331152,
0.0293681454,
0.0285876691,
-0.0485215336,
-0.0210424587,
0.0257720463,
-0.0254983511,
-0.0249665715,
-0.0394167528,
-0.0448860228,
0.0401876755,
-0.028684685,
0.0126851453,
-0.0181694832,
-0.0027364532,
-0.0452832952,
0.0431199335,
-0.0136048608,
0.0027988402,
0.0253856238,
0.0133894058,
0.0034846033,
-0.0191588681,
-0.0090762749,
0.0095058549,
-0.0122470446,
0.037648309,
0.0146093331,
0.0260354895,
-0.0664061159,
-0.0257147625,
-0.1004826501,
-0.0041892151,
0.0056997752,
0.0039843223,
0.0216212962,
-0.0356711112,
-0.0374475047,
0.0075157438,
-0.0478521436,
-0.0419908054,
0.0028569596,
0.0162042622,
0.0538930111,
-0.0016606237,
0.0169518776,
-0.0496804491,
0.0194423907,
-0.0247808564,
0.0144142322,
0.003011578,
-0.0198681299,
-0.0656379834,
0.0661634728,
-0.0063063176,
0.0129467612,
-0.0143780177,
-0.039054174,
0.0469227843,
0.0235922746,
0.0611749515,
0.0848224685,
0.0180291347,
0.0398957469,
0.0536468253,
0.0120783355,
-0.0765362829,
0.0010617481,
-0.010020663,
-0.0176956616,
-0.0415690541,
-0.0656219721,
0.0105503052,
0.0532031059,
0.0044754734,
0.0380547568,
-0.0333730094,
0.0458151624,
-0.0364628844,
-0.0516702905,
-0.0049218619,
-0.0181425642,
-0.0618063398,
0.0332405232,
-0.0237309188,
-0.068224214,
0.0540036298,
-0.0317652151,
0.0074962503,
-0.0101085296,
-0.009352643,
-0.021394765,
0.0494117662,
-0.0096299164,
0.0078087361,
0.0261539891,
-0.0052579478,
-0.0508758724,
0.0195910074,
0.0129084513,
-0.0615625493,
0.0617456958,
-0.0497908853,
-0.011756463,
0.0021624761,
-0.0323001035,
-0.0564715303,
-0.0291691665,
-0.0152447708,
-0.021019455,
0.029680457,
-0.0144220842,
0.0021070631,
-0.0358510464,
0.0278592091,
0.0364442319,
-0.0094321752,
-0.0132310884,
-0.0258932821,
0.0020867535,
0.0453697219,
0.0061275139,
0.0161752589,
-0.0381141901,
0.0048066103,
0.0258747526,
0.0283578262,
0.0194242056,
-0.0029432431,
-0.0086527355,
0.0213045459,
-0.0214361418,
-0.0145259416,
0.0550322011,
0.0639593676,
-0.0545857698,
-0.0095722564,
0.1082174033,
0.0259556472,
-0.0692068934,
0.0253257435,
-0.0107032685,
0.0233804509,
0.0379693806,
0.0193173513,
-0.1092233807,
-0.0127198678,
0.004426932,
0.049478773,
-0.054549709,
-0.0360735618,
-0.0081487242,
-0.139815405,
0.0187420528,
-0.0006362173,
-0.0403253138,
0.0107793771,
-0.0377957076,
-0.012886947,
-0.0084335888,
0.0431717746,
-0.0061230282,
-0.0299890209,
-0.0086313235,
0.0544526279,
-0.0283840466,
-0.0573811792,
0.0246898551,
-0.0363787934,
0.0940610766,
0.0701478347,
0.043882627,
-0.0223011039,
-0.0317148305,
0.0237680543,
0.0140378149,
0.0277363677,
0.0053935787,
0.0535170101,
-0.0200631525,
-0.005906974,
-0.0522608347,
0.0262871217,
-0.0302016977,
-0.0112488987,
0.0197464898,
0.0773684531,
-0.0171806738,
-0.0114724096,
-0.0410081409,
-0.0267888736,
0.0332903154,
0.0877619982,
0.0506523512,
-0.0782551169,
-0.0186982155,
-0.0331839882,
-0.0030380855,
-0.0495317727,
-0.0375581011,
0.0438064821,
-0.0359011553,
0.0035408831,
-0.0818411708,
-0.1008394659,
0.0189935509,
0.0021600428,
0.0277013257,
-0.0120448964,
0.0048368592,
-0.018499488,
-0.04332361,
0.0450999029,
-0.0370051265,
0.02332028,
-0.0555357113,
-0.0613114461,
-0.0058091599,
-0.0447883606,
0.0196864568,
0.0211966652,
-0.0583328493,
-0.0574990176,
0.0215935204,
0.0765917376,
0.0257565435,
0.0506069213,
0.0426237993,
-0.0498868972,
0.0095530665,
0.0533868149,
-0.0479219593,
0.02907639,
0.0206527077,
0.0064336387,
0.0020240678,
0.0074526905,
-0.0284437034,
0.0204840079,
0.0153826782,
0.0067190286,
-0.024330914,
-0.0439750254,
0.0329504088,
0.0095426207,
0.0096154204,
-0.0353015512,
-0.0163617413,
-0.0354454853,
-0.0202896222,
-0.0148954671,
-0.0363290161,
-0.0405741632,
-0.0145031987,
0.0000576189,
0.0121397171,
0.0155554656,
0.0097571481,
-0.0402576774,
0.0584078841,
-0.0399192944,
0.0092726089,
-0.0006916829,
0.0378253236,
0.0197563581,
0.0234677419,
-0.0286117718,
0.0435728133,
-0.0187081378,
0.0509594232,
0.0280383024,
0.0698225722,
0.0111478958,
-0.0099026933,
0.0368806496,
0.0430907458,
-0.0437342376,
-0.0427092537,
-0.0336855315,
-0.0123180896,
0.0363488793,
-0.020664297,
0.0600546114,
-0.0258198492,
0.0151517428,
-0.0028364749,
0.0149871968,
-0.0093374066,
-0.0262123849,
0.0043220134,
0.02272176,
-0.0288869701,
-0.0365319997,
0.0883301347,
0.0220003482,
0.0365612619,
-0.0195466485,
-0.0048082024,
-0.0371007398,
-0.0229013879,
0.0434983782,
-0.0183742438,
-0.0227986034,
-0.0429647043,
-0.0674785748,
-0.0786199719,
-0.0021760003,
0.0327011943,
0.0122470446,
-0.049378816,
0.0407135189,
0.0353592746,
-0.0282641556,
-0.0001399812,
-0.0523197353,
0.0032863873,
-0.0231321473,
-0.0038674721,
-0.0287220906,
-5.862582281e-33,
-0.0293898229,
-0.0547726862,
0.0351939425,
0.0157222189,
-0.0278902743,
0.0172067992,
-0.0281633064,
-0.0001585494,
-0.0157597233,
-0.008301965,
-0.0128648691,
-0.0168737136,
0.0067652781,
0.0061596381,
0.0284839235,
-0.0206423067,
0.0363350697,
-0.0265503731,
-0.0316018239,
0.024362538,
0.0133920545,
-0.009330078,
0.0488837138,
-0.0729027689,
0.0543725342,
-0.0324922279,
0.0018690461,
-0.0039331415,
-0.0027332881,
0.0751264542,
-0.0348285437,
0.0602832846,
0.0227321889,
-0.0196862761,
-0.0387442783,
0.0409884416,
-0.0165601391,
-0.0347665623,
0.0232947115,
0.0084122559,
-0.0108468756,
-0.0448025092,
0.0435864106,
-0.026740443,
-0.0467246436,
0.0287516266,
0.0561354607,
-0.0275282059,
-0.0038882771,
-0.0112309866,
-0.0707865953,
0.0131493807,
-0.0039124279,
-0.0232118014,
0.0217075367,
0.0430887379,
-0.0061431574,
0.0025292742,
-0.0535350777,
0.0426374003,
0.0109954514,
-0.0064799637,
0.0808273256,
0.0469473079,
0.0357883871,
0.0773938596,
0.0524639152,
0.0889845118,
-0.0326370932,
-0.0018567761,
0.0428756885,
0.0485316515,
0.0257466268,
0.044547502,
-0.0046605412,
-0.029981453,
-0.0078228321,
0.0233158227,
0.0108918827,
-0.0144750001,
-0.0085091209,
0.0072609945,
-0.0531609915,
-0.0313270725,
-0.0322543122,
0.0062151453,
-0.0024286557,
-0.0141213564,
0.0435375124,
-0.0065813046,
0.0003190236,
0.0046468498,
-0.001514283,
-0.026938919,
-0.014942443,
-0.0198717602,
-0.0181904063,
0.0368919969,
0.0049066297,
-0.0233168975,
-0.0823260844,
-0.0341166928,
0.0024863284,
-0.0020249875,
0.0587787926,
0.0469533205,
0.0277203843,
0.0306180026,
-0.0387933664,
-0.0043193558,
-0.0375903361,
0.0145160314,
0.015382302,
-0.025848588,
0.0424596928,
-0.005726221,
-0.0182377249,
0.0558130182,
0.0234361216,
-0.0096162772,
-0.052862756,
-0.0232844427,
-0.0044038598,
0.0032019301,
-0.0262955874,
-0.019927321,
-0.0668727085,
0.0439455844,
0.0455357544,
-0.0526472293,
0.0006980597,
-0.0040892605,
2.679e-7,
0.0187067911,
0.0552579314,
0.0230612233,
0.0228827335,
0.0016876714,
0.0498937853,
-0.0143157262,
0.0212718695,
0.0047506657,
-0.0432059765,
0.0035746116,
0.0271375366,
0.0468976982,
0.0425160006,
-0.0807813779,
-0.0265518818,
-0.0631148517,
-0.0053694942,
-0.0812627226,
0.0295764972,
0.0355190374,
0.0292675998,
0.0007468704,
-0.023749359,
-0.0043230518,
-0.005935112,
0.038048394,
-0.0715270042,
0.0056541199,
0.0149123669,
0.0012010725,
0.007263138,
-0.0123850377,
0.0275373608,
-0.0174272191,
-0.0163295306,
0.0040188129,
0.0758367404,
-0.0150912274,
-0.0298531745,
0.012731771,
-0.0375870578,
-0.0082219802,
-0.050497096,
0.0422832482,
-0.024426626,
-0.0073214048,
-0.0471791141,
-0.0018039662,
-0.0200291388,
-0.004853182,
-0.0402501486,
0.0112648308,
0.0102946423,
0.003178223,
-0.0048179985,
-0.0055111954,
-0.0419616997,
-0.02635823,
0.004846157,
-0.038942825,
-0.0644637272,
-0.061146725,
0.0225114785,
-0.0081746168,
-0.0556757003,
0.0222859792,
2.368197526e-34,
-0.0006719379,
-0.0078994855,
0.0443360247,
0.0292673539,
0.0143325627,
0.0354764573,
0.0224099476,
0.0328202061,
0.0206128266,
-0.0455361232,
-0.0100927101
] |