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/2401.17396v1 | http://arxiv.org/pdf/2401.17396v1 | Fine-tuning Transformer-based Encoder for Turkish Language Understanding Tasks | 2024-01-30T19:27:04 | 2024-01-30T19:27:04 | Savas Yildirim | cs.CL | cs.CL, cs.AI | Deep learning-based and lately Transformer-based language models have been
dominating the studies of natural language processing in the last years. Thanks
to their accurate and fast fine-tuning characteristics, they have outperformed
traditional machine learning-based approaches and achieved state-of-the-art
results for many challenging natural language understanding (NLU) problems.
Recent studies showed that the Transformer-based models such as BERT, which is
Bidirectional Encoder Representations from Transformers, have reached
impressive achievements on many tasks. Moreover, thanks to their transfer
learning capacity, these architectures allow us to transfer pre-built models
and fine-tune them to specific NLU tasks such as question answering. In this
study, we provide a Transformer-based model and a baseline benchmark for the
Turkish Language. We successfully fine-tuned a Turkish BERT model, namely
BERTurk that is trained with base settings, to many downstream tasks and
evaluated with a the Turkish Benchmark dataset. We showed that our studies
significantly outperformed other existing baseline approaches for Named-Entity
Recognition, Sentiment Analysis, Question Answering and Text Classification in
Turkish Language. We publicly released these four fine-tuned models and
resources in reproducibility and with the view of supporting other Turkish
researchers and applications. | Fine-tuning Transformer-based Encoder for Turkish Language Understanding Tasks: Deep learning-based and lately Transformer-based language models have been
dominating the studies of natural language processing in the last years. Thanks
to their accurate and fast fine-tuning characteristics, they have outperformed
traditional machine learning-based approaches and achieved state-of-the-art
results for many challenging natural language understanding (NLU) problems.
Recent studies showed that the Transformer-based models such as BERT, which is
Bidirectional Encoder Representations from Transformers, have reached
impressive achievements on many tasks. Moreover, thanks to their transfer
learning capacity, these architectures allow us to transfer pre-built models
and fine-tune them to specific NLU tasks such as question answering. In this
study, we provide a Transformer-based model and a baseline benchmark for the
Turkish Language. We successfully fine-tuned a Turkish BERT model, namely
BERTurk that is trained with base settings, to many downstream tasks and
evaluated with a the Turkish Benchmark dataset. We showed that our studies
significantly outperformed other existing baseline approaches for Named-Entity
Recognition, Sentiment Analysis, Question Answering and Text Classification in
Turkish Language. We publicly released these four fine-tuned models and
resources in reproducibility and with the view of supporting other Turkish
researchers and applications. | [
0.0554596782,
0.0285358969,
0.0027818007,
0.0420274399,
-0.0060238424,
0.0671150684,
0.0236122552,
-0.0337014347,
-0.0369876735,
-0.0332248919,
-0.0308394339,
-0.0689579099,
0.0029607874,
0.0201934949,
-0.0011405637,
-0.0289884023,
0.0542882122,
-0.027202649,
0.0047147046,
0.0151770962,
-0.0047283466,
0.034022782,
-0.050631348,
0.0147318542,
0.006122306,
0.0019504221,
-0.0124141313,
-0.0200380851,
0.0185970962,
-0.0019870156,
0.0317857675,
0.0232751034,
-0.0000802079,
0.0497591347,
0.000001924,
-0.0220702048,
-0.0201686434,
-0.0276124589,
-0.0097331526,
-0.039551571,
-0.0054824227,
-0.069959797,
-0.019314114,
-0.0162442476,
-0.040801324,
-0.0390146077,
0.0000316742,
0.0665101781,
0.0736473799,
0.0678547099,
0.0087529514,
-0.0632108152,
0.0437655337,
-0.0155143533,
-0.0308838598,
-0.0539394915,
0.0106820185,
-0.0703052133,
-0.0635471344,
0.0022944794,
-0.0150118982,
0.0399539508,
0.0354229473,
0.0125154518,
-0.0231685434,
-0.0012450127,
-0.0720626265,
-0.0447908454,
0.0220514424,
0.0724347457,
-0.0191723183,
-0.0210986398,
-0.0031892483,
0.0120430999,
-0.0148081323,
0.0061822496,
-0.0330683105,
0.0066834996,
-0.0183491558,
0.0066777957,
0.0191217102,
0.0116955321,
0.0076783504,
0.0203210376,
-0.0246250425,
0.042471312,
0.0238075163,
-0.016450936,
-0.0102579845,
0.0235666092,
-0.0042335191,
-0.0183029994,
0.0593651719,
0.0531552583,
-0.0022689614,
0.0031765825,
-0.0035884858,
-0.0405852795,
-0.00688614,
-0.0628146082,
-0.0028697397,
0.0381409004,
-0.0142097957,
0.0366862193,
-0.1021122932,
0.0447015874,
-0.0163160954,
0.0105841681,
-0.0237419624,
-0.0761531666,
-0.0292275958,
-0.0225851145,
-0.0496492684,
0.0361266732,
-0.0583361648,
0.0198016707,
-0.0327987224,
0.0340777636,
0.0144726504,
-0.0001216078,
0.0201761909,
0.029243676,
-0.0524989851,
-0.0089315046,
-0.0372697562,
0.0589993931,
-0.0478952006,
-0.0036162012,
-0.0446853675,
-0.0412650481,
0.0033091048,
0.0004351796,
0.0224746093,
-0.0237515271,
0.0127033023,
0.0316728503,
-0.0184945501,
-0.0484799333,
-0.0275414512,
0.0554830283,
-0.0092204185,
-0.0389834382,
0.0259123649,
-0.0367042832,
0.0436216332,
0.0152134802,
-0.0179517251,
-0.0200025979,
0.0046027717,
0.0279238522,
-0.0125567494,
0.0796567649,
0.0050631249,
-0.0079749646,
0.0363021828,
-0.0005738413,
0.0414568894,
0.0585510135,
0.0474527553,
0.0246338435,
0.0004403093,
-0.0309062395,
0.0077761472,
0.0153468028,
0.0114794094,
-0.0363348126,
0.0190142673,
0.0323191844,
-0.0429577194,
0.0263280105,
0.019666655,
0.0260831397,
0.0191827081,
-0.0062096035,
0.0428773314,
0.0278543178,
-0.0153184952,
0.0835575834,
0.0125318859,
0.0117961504,
-0.0155572258,
-0.0149408681,
-0.0012834747,
0.0363224372,
0.0214344673,
0.0012981433,
-0.0263861474,
0.0379931331,
-0.046632871,
-0.0281147324,
0.0263598002,
-0.0115160327,
0.0151479701,
-0.021022277,
0.0853362232,
0.0368798599,
-0.0619694628,
0.12469282,
0.0201182757,
-0.01683129,
0.0262751859,
-0.0534978323,
0.0265467167,
0.0458942056,
0.0188388992,
-0.0083026057,
-0.0150737129,
-0.0452067181,
0.0127882203,
0.0467998125,
0.0174399372,
-0.0204935037,
0.0179748312,
-0.0413966663,
-0.0007001711,
-0.0070126019,
0.0043702284,
0.0260391105,
-0.0397950001,
0.0311881322,
0.0217013918,
-0.050118763,
0.0312569737,
0.0301642232,
-0.0014690252,
-0.0110067977,
-0.0015479857,
0.0269924346,
-0.0194358323,
0.0250579733,
0.0160260033,
-0.0243251789,
0.0341032334,
-0.063180469,
0.0015474176,
-0.0142442361,
-0.0324914865,
-0.0432207696,
0.0055935946,
-0.0447902828,
-0.0478282981,
0.0879408494,
-0.049219884,
-0.0497299358,
0.0187908076,
-0.0418841019,
0.049522303,
-0.0278657358,
0.0255171768,
-0.0032825596,
0.0348695777,
-0.0143259699,
0.0470122173,
-0.0131076155,
0.0033018377,
-0.0034049172,
-0.0203743093,
0.0236090906,
0.0293064564,
-0.0412845053,
0.017814951,
-0.0210432298,
-0.0109565435,
0.0056354604,
-0.0179661289,
-0.0241960213,
-0.0282014813,
-0.0979642123,
0.0129999863,
-0.0444433056,
0.0610274561,
0.0266279094,
0.0263880156,
-0.005430501,
-0.0188364573,
-0.0269646551,
-0.0278689209,
0.0101112062,
-0.003794499,
0.0706861019,
-0.0296469796,
-0.0077563673,
-0.0106217405,
-0.0270517394,
-0.0188467242,
0.0386943147,
0.013060716,
-0.0242830738,
-0.0285930168,
-0.0431511737,
0.002297774,
0.0071603418,
-0.0063981344,
-0.0604877695,
-0.0010386182,
0.0454014614,
0.0279389434,
0.0508539081,
0.0317594633,
0.0296442676,
-0.0220718533,
0.0083625223,
-0.0366675258,
0.0225517973,
-0.04179129,
-0.0190858282,
-0.0318569504,
0.0085703963,
-0.0015579212,
0.0846239626,
-0.0367310457,
0.00309281,
-0.0732997134,
-0.0301497299,
-0.0015708959,
0.0047653089,
0.0067002848,
0.0246171989,
-0.0154412594,
0.0486696921,
-0.0216052886,
0.0222971272,
0.058666978,
-0.0379939303,
-0.0022111065,
-0.0618045144,
0.0176238101,
-0.0493711568,
0.0084357057,
0.0058605005,
0.0041307504,
-0.0335749686,
-0.0233297814,
-0.0276313685,
0.0160851981,
-0.0170158111,
-0.0387720093,
0.0546078943,
0.001944188,
0.0117992572,
-0.0257757269,
-0.0325102732,
-0.0607271977,
-0.0323006734,
-0.0291648749,
0.0438278988,
0.0390036888,
0.0002960782,
-0.0670772791,
0.0354330689,
0.0062590409,
0.0247654133,
-0.0051381122,
0.0007003444,
-0.0459975116,
0.052220948,
0.0596624017,
0.0092502628,
0.0430878624,
0.0156052746,
0.0273425486,
0.0521649569,
-0.0131607922,
0.0024629289,
-0.0429999195,
-0.0063923732,
0.0218191929,
-0.0618851557,
0.0073769251,
-0.0481038652,
0.0331944823,
-0.0539516583,
-0.039277263,
0.126642704,
0.0007713191,
-0.0158451926,
-0.012490958,
-0.0048422329,
-0.0139172701,
0.0079096602,
0.0363373198,
-0.0494944677,
-0.0484975241,
-0.0075350385,
0.0222108923,
0.037443433,
-0.0120363608,
0.062460687,
-0.1216868311,
-0.0260561947,
-0.0576331355,
-0.0700422674,
-0.0528570898,
-0.0694791749,
0.0012138167,
-0.0335106775,
0.0281251464,
-0.0306775887,
0.0028078244,
0.0103691658,
0.0405642726,
-0.03707923,
-0.0651406571,
0.0059263092,
-0.0377774648,
-0.0033911825,
0.0403193831,
0.0363581665,
-0.029745046,
-0.0130877579,
0.0062218546,
-0.0484117344,
-0.0089442227,
0.0264859013,
0.0380538963,
-0.0421961099,
-0.0419935286,
-0.0324609056,
-0.008546683,
-0.099487409,
0.0105507327,
-0.037341699,
0.0529701672,
0.0015669395,
-0.0049046855,
0.0107424343,
-0.0078330152,
-0.0073604328,
0.0640123114,
0.0044994978,
-0.0388006903,
-0.0374041423,
-0.0105936723,
0.0002268178,
-0.0727977976,
-0.0081077237,
0.024976844,
0.043430686,
0.0012260909,
-0.0553465299,
-0.0406311378,
0.0418674871,
-0.0087157572,
-0.030375503,
-0.0135084549,
0.0201435871,
0.00143868,
-0.0542450175,
0.0448633954,
0.0114408378,
-0.0017176381,
-0.0617095754,
-0.0230588689,
0.0728865787,
-0.0046919817,
-0.037833672,
0.0042477134,
-0.0895090774,
-0.0095130578,
0.019330712,
0.0329734609,
0.009375507,
0.0189401954,
-0.0200351365,
-0.019349549,
0.0220500045,
0.0059081567,
-0.0488428473,
-0.0187451132,
0.0078513902,
0.0089836773,
-0.028008651,
0.0448436514,
-0.0227825232,
0.0096762944,
0.0245972127,
0.0079817595,
-0.0359015279,
-0.029267149,
0.0076094284,
-0.0157461613,
0.0422789827,
0.0193222109,
0.0024192352,
-0.0339087211,
0.0540752187,
0.063848421,
0.0222087167,
-0.0168182682,
-0.007988126,
0.0271259118,
-0.032867562,
-0.0121016456,
0.0455804802,
-0.0121711157,
-0.0163318273,
-0.0623871945,
0.028834492,
0.0222665668,
-0.0035011575,
-0.0173202138,
0.0118606035,
-0.0279237572,
0.0357621945,
0.025807796,
0.0577843152,
-0.0223082993,
0.0396017283,
-0.0084972065,
0.0184360389,
0.0102735944,
0.0487685613,
-0.0589322634,
-0.0227092728,
-0.019466674,
0.0522289649,
0.0286681242,
-0.0395409763,
0.0234628357,
0.0257939063,
-0.0018741809,
0.0122169163,
0.010530916,
-0.0340933912,
0.0057489066,
0.0375238918,
-0.0031718542,
0.0006819469,
-0.0075635202,
0.0420715399,
-0.0138485301,
0.0534917526,
0.007662287,
0.02010509,
-0.0258501749,
-0.0068036662,
0.0168152098,
0.0017501628,
-0.0135013573,
0.0347961821,
-0.0130477082,
-0.0672827214,
0.0176325757,
-0.0383669883,
0.0125363348,
-0.0514628291,
0.0534811169,
0.0498543158,
0.038376607,
-0.0128056649,
-0.0348397195,
0.0198850613,
0.0055555948,
0.0093404306,
-0.0476707928,
-6.116186739e-33,
-0.0294915829,
0.031683173,
-0.0198502671,
-0.0219326895,
0.011306609,
-0.0245430171,
-0.0233955104,
-0.0157002173,
-0.0464854911,
0.0088047301,
-0.0176884811,
0.0377429724,
0.0067545935,
0.0624673888,
-0.0314238928,
-0.0973870531,
0.0465019494,
0.0023877213,
-0.0196262151,
0.025128467,
0.0607037023,
0.070687227,
0.0925641507,
-0.0130019458,
-0.0306928456,
-0.0319589451,
0.0270356834,
0.0258436538,
0.0259476621,
0.0798201784,
-0.0169554185,
0.0343493223,
0.0405988768,
-0.0743646771,
0.0150913782,
0.0356965214,
-0.1289597303,
-0.0003311207,
-0.0288843829,
0.0728533864,
-0.0315103345,
-0.0175820738,
0.0467596278,
0.0243614111,
0.0238688756,
0.0097798118,
0.0311557259,
-0.0235190634,
0.0025972191,
-0.0421423502,
-0.0649464279,
0.0112840319,
-0.0390989557,
-0.0287414212,
0.0366837829,
0.0669858754,
0.0034838056,
0.0212583393,
-0.0691839233,
-0.0362971723,
-0.0249732789,
0.0227084439,
0.0650506243,
-0.0146201337,
-0.0436446369,
0.0270222612,
-0.0394211262,
0.0084437337,
-0.0238415375,
-0.0055264491,
0.0056108083,
0.0125016253,
0.0364466235,
0.0341407657,
0.0186995529,
-0.0185371991,
-0.0776200369,
0.0374861024,
0.0111174788,
0.0611410253,
0.0146828378,
-0.0318186767,
-0.0233434159,
-0.0089458358,
-0.0035874282,
-0.0002267343,
-0.0078500556,
0.0041955905,
0.0162877124,
-0.0289222933,
-0.0582891703,
0.0196766201,
0.0143122133,
-0.0533102378,
0.0407173447,
0.020324355,
0.0226118807,
-0.0100808777,
0.0042630993,
0.0560737737,
-0.0279648509,
-0.0383589678,
0.0147854323,
0.007300945,
0.0169100799,
0.0148313977,
0.0228973627,
-0.012730781,
-0.0555657782,
0.0093359118,
-0.0351748504,
0.033199925,
-0.0302952081,
0.1036185771,
0.0150355883,
0.0233631711,
-0.0239912812,
0.0518360063,
-0.0000990753,
0.0093805566,
-0.011868434,
0.0202208087,
0.0099023953,
0.0020330125,
-0.0180584304,
-0.0102081224,
-0.0365591161,
0.0412095264,
0.0022990445,
-0.0908985734,
-0.0061032656,
0.0294075217,
2.667e-7,
0.0701583475,
0.0313174613,
0.0471643284,
-0.022024218,
0.0111044561,
0.0241597202,
0.0389556065,
0.0367598273,
0.0233639181,
-0.0207513217,
0.0146664511,
0.0341348946,
0.0453191884,
0.0148756048,
-0.0645514727,
-0.0335004516,
-0.0714143887,
0.0363259465,
-0.0262692664,
0.0445832089,
0.0580056757,
0.1123157516,
-0.0327652358,
-0.0245047212,
-0.006993603,
0.0000812464,
0.0637047961,
-0.0068029645,
0.0204540268,
0.0318421051,
-0.0087001855,
0.0358813405,
-0.0011541898,
0.0275091752,
-0.0211363975,
0.0562402047,
0.0318267792,
0.0618899018,
0.0007613045,
-0.0217497014,
0.0313484222,
-0.0219918936,
0.0147737795,
-0.0737620518,
0.1114478484,
-0.092352502,
-0.0444907248,
-0.0061988616,
-0.0255497601,
0.0176911671,
0.0094517432,
-0.0038631372,
-0.006703252,
0.0428303741,
0.0418170355,
-0.0258275717,
-0.022452686,
-0.0761312842,
0.0442828909,
-0.0150030823,
-0.0252384748,
0.020502843,
-0.0200899076,
-0.029579863,
0.0438031554,
0.0217939802,
-0.005899908,
1.710479853e-34,
-0.0073163207,
0.0301971007,
-0.0212019999,
0.0270318035,
0.0150649128,
-0.0146861523,
0.0158612095,
-0.0306945685,
0.0267566536,
-0.0650537312,
-0.0064118449
] |
http://arxiv.org/abs/2401.17390v2 | http://arxiv.org/pdf/2401.17390v2 | Customizing Language Model Responses with Contrastive In-Context Learning | 2024-01-30T19:13:12 | 2024-04-08T05:22:51 | Xiang Gao, Kamalika Das | cs.CL | cs.CL, cs.AI | Large language models (LLMs) are becoming increasingly important for machine
learning applications. However, it can be challenging to align LLMs with our
intent, particularly when we want to generate content that is preferable over
others or when we want the LLM to respond in a certain style or tone that is
hard to describe. To address this challenge, we propose an approach that uses
contrastive examples to better describe our intent. This involves providing
positive examples that illustrate the true intent, along with negative examples
that show what characteristics we want LLMs to avoid. The negative examples can
be retrieved from labeled data, written by a human, or generated by the LLM
itself. Before generating an answer, we ask the model to analyze the examples
to teach itself what to avoid. This reasoning step provides the model with the
appropriate articulation of the user's need and guides it towards generting a
better answer. We tested our approach on both synthesized and real-world
datasets, including StackExchange and Reddit, and found that it significantly
improves performance compared to standard few-shot prompting | Customizing Language Model Responses with Contrastive In-Context Learning: Large language models (LLMs) are becoming increasingly important for machine
learning applications. However, it can be challenging to align LLMs with our
intent, particularly when we want to generate content that is preferable over
others or when we want the LLM to respond in a certain style or tone that is
hard to describe. To address this challenge, we propose an approach that uses
contrastive examples to better describe our intent. This involves providing
positive examples that illustrate the true intent, along with negative examples
that show what characteristics we want LLMs to avoid. The negative examples can
be retrieved from labeled data, written by a human, or generated by the LLM
itself. Before generating an answer, we ask the model to analyze the examples
to teach itself what to avoid. This reasoning step provides the model with the
appropriate articulation of the user's need and guides it towards generting a
better answer. We tested our approach on both synthesized and real-world
datasets, including StackExchange and Reddit, and found that it significantly
improves performance compared to standard few-shot prompting | [
0.0752133653,
0.0623065792,
-0.0146370782,
0.0344960801,
-0.0370072611,
0.0135391215,
-0.0153902415,
0.0050271018,
0.0123301093,
-0.0736736432,
0.0071079959,
-0.0484169275,
-0.055695489,
0.0256483927,
0.0050209933,
-0.0666702911,
0.0159962773,
-0.0318383873,
0.0444544591,
-0.013046124,
0.0318701789,
0.035934791,
-0.0017552958,
0.0289743096,
0.0194617752,
-0.0442845412,
-0.0043008034,
0.0449218675,
-0.0014812486,
-0.0153708952,
0.0634127408,
0.0594536886,
-0.0070272232,
0.0175053515,
0.0000019895,
-0.0544000156,
0.0187275503,
-0.0154636549,
-0.027839629,
-0.0197779592,
0.0772116706,
0.0044017476,
-0.013174708,
0.0555897988,
-0.0749019012,
0.0345184244,
0.0471131653,
0.0472414605,
0.0663437396,
0.0746163949,
-0.0026209459,
-0.0206172485,
-0.0232060421,
0.0077010444,
0.0197754279,
0.0155002475,
0.0044868127,
-0.0228768997,
0.0124673285,
0.0166501198,
0.0071685626,
0.0344494246,
0.014746367,
-0.0361947715,
0.0114590982,
0.0141121726,
-0.0186080281,
-0.0170244183,
-0.007335999,
0.0901248753,
0.0046762875,
-0.0368714146,
-0.0122149959,
-0.0285361391,
0.0139927585,
-0.0031753704,
-0.0456816517,
-0.0071969577,
-0.0070467959,
0.0189892426,
0.0144021893,
0.0180555191,
0.0039097853,
0.0276754741,
-0.0039237416,
0.0300127156,
0.0018921826,
-0.0315665752,
-0.0025751705,
0.001424281,
0.0130197499,
-0.0519456118,
-0.0051302887,
0.0456262194,
-0.0597037487,
0.0220884327,
-0.0470400192,
-0.028398348,
0.0485351682,
-0.0412136763,
0.0603509247,
0.0319004171,
-0.0011838952,
0.0335883871,
-0.003873386,
-0.0027231004,
-0.0122417025,
0.0078047733,
-0.0407201611,
-0.0201301686,
-0.05079915,
-0.0125335716,
-0.0205597579,
0.0195593238,
-0.0206138138,
0.0132725127,
0.0028431567,
0.059507817,
-0.0236152597,
0.0088086948,
0.0110467821,
0.0266468488,
-0.0290631633,
-0.0168768968,
-0.0453351736,
0.0135100484,
-0.0519444346,
-0.0024378556,
0.0171379615,
-0.0470789969,
0.0112401526,
0.0398835093,
0.0018314358,
-0.0152508821,
0.0151444003,
0.0865655988,
0.0019446617,
-0.0049431315,
-0.0398027971,
-0.01622146,
-0.0164696164,
-0.0354341008,
-0.0113725476,
-0.062273249,
-0.0014586191,
0.0368711613,
-0.0450759009,
-0.0347673371,
0.010017979,
0.0505019203,
-0.0111396536,
0.049456045,
-0.0437457412,
0.0056210649,
0.0302762836,
0.0083213914,
-0.0113557717,
0.0746541247,
0.0034003931,
0.0065148156,
0.0310844425,
-0.008855626,
0.0274712611,
0.0214531329,
0.0295408498,
0.013832272,
-0.0569479838,
0.0380877331,
-0.0303421114,
-0.0018416021,
0.0399880707,
0.0662055016,
0.0022472655,
-0.0295924209,
0.0706234798,
0.0425636061,
0.0044883625,
0.0050311037,
0.0242563058,
0.0465278551,
-0.0125468653,
0.0123319412,
-0.0253038052,
0.0344343334,
-0.0401931256,
0.0078368494,
0.024223363,
0.0095504932,
-0.0351296291,
-0.0526367463,
-0.014852697,
0.0004289129,
0.0295510609,
0.0260216463,
0.0244231652,
0.0311607514,
-0.0353734531,
0.0064059813,
0.0451622941,
-0.0330515802,
-0.009198294,
-0.0685856268,
0.052768115,
0.0777006745,
0.0340061672,
-0.0212991517,
-0.0192065854,
-0.0281078238,
0.0134716062,
0.0675747171,
0.0683945417,
-0.0224021878,
-0.011807261,
-0.0477672517,
-0.0125455866,
-0.0208963454,
-0.0117233731,
-0.0066112899,
0.025901394,
-0.0217348021,
0.0120580113,
-0.0173411444,
-0.0427316688,
0.0303350016,
0.0110609103,
-0.0393538997,
0.0382654704,
0.0022519643,
0.0309180655,
0.0284910351,
0.0224850997,
0.0125787258,
0.0130324019,
-0.0183411036,
-0.0219399761,
-0.0205811281,
-0.0566229708,
-0.0562597439,
0.024873402,
-0.0315544121,
-0.0301526207,
0.0359759517,
-0.039239727,
-0.0226150136,
-0.0063215722,
-0.0337969735,
0.0640802383,
-0.0330610946,
-0.0231939089,
0.0000406991,
-0.0023849064,
-0.0492340326,
0.043740835,
0.0080916146,
0.0024496291,
0.045484066,
0.0050300108,
-0.0188146867,
-0.0175964031,
-0.0127263106,
-0.0260024257,
0.0007743971,
0.0274891742,
0.0353432335,
0.0282738227,
-0.0607767776,
-0.0026162087,
-0.0504601896,
-0.029451048,
-0.0499049872,
0.0099859806,
0.0534782372,
0.0021136,
-0.0026921236,
-0.0341587923,
-0.0655418932,
-0.01613787,
0.0068310457,
0.0222557131,
0.0526443161,
0.0033286025,
-0.0079645747,
-0.0454937406,
0.0013625039,
-0.0062139677,
0.0258293729,
-0.0056395656,
0.0110103423,
-0.0637255684,
0.0370671824,
-0.0090388227,
0.0214044061,
-0.0078466805,
-0.0462496988,
0.0195721295,
-0.0111776982,
0.0316057205,
0.0608762279,
0.0583618991,
0.0319584981,
0.0445424281,
0.0437517054,
-0.0871249586,
0.0043611019,
-0.0267408974,
-0.0019688022,
-0.025879547,
-0.044852268,
0.0068688686,
0.0496987998,
0.005422838,
0.0198738184,
-0.0657917261,
0.0421740711,
0.0088645881,
-0.030542111,
-0.0122167328,
-0.0116552534,
-0.0292653609,
0.0515905842,
0.0044400156,
-0.0503446199,
0.0938123167,
0.0076903766,
-0.018408196,
-0.0200457592,
-0.0354946777,
-0.0190060008,
0.0415280089,
-0.0181481149,
0.0200581029,
0.0148836281,
0.0036401264,
-0.0444051102,
-0.0431055874,
0.0080938945,
-0.0293278899,
0.0459561944,
-0.0470486507,
-0.0039539393,
0.0071345791,
-0.0385357216,
-0.0832041875,
-0.062693432,
-0.0202024765,
0.0091577275,
-0.0051393937,
-0.0123714376,
0.0140710957,
0.0005500512,
0.005840708,
0.0343137607,
-0.027800817,
-0.0307875182,
-0.0049049486,
0.0266480166,
0.0967338085,
0.0177816935,
0.0011879596,
-0.0134152938,
0.0517891794,
0.0532785058,
0.0200735498,
-0.0217727572,
-0.0164709371,
-0.0338685438,
0.0541147515,
-0.0231861677,
-0.0004474186,
0.039551679,
0.0378173813,
-0.0401762389,
-0.0406353362,
0.1125038415,
0.0336169042,
-0.0597249456,
0.0096135763,
-0.0511903465,
0.0572991446,
0.0314612053,
0.0561817735,
-0.0739652738,
-0.0413406678,
-0.0141645139,
0.0525010154,
-0.0323096253,
0.0114823999,
-0.0190498419,
-0.1514342725,
0.0201375633,
-0.0049913358,
-0.028805539,
-0.0129590463,
-0.0264303032,
-0.0061291172,
-0.0073596975,
0.0140389986,
0.0122893648,
-0.0166982673,
-0.0061605871,
0.0306376144,
-0.0340743512,
-0.0684494227,
0.007872426,
-0.0141550209,
0.0614109598,
0.0596645586,
-0.0008697687,
0.0052725733,
-0.0312377829,
0.0160870682,
-0.0358459949,
0.027491264,
-0.0228401031,
0.0512858294,
-0.0014793716,
0.011073417,
-0.0455463529,
0.0535185151,
-0.034736149,
-0.0194042306,
0.0190707389,
0.059409868,
-0.0366959721,
0.0107283778,
-0.0543993711,
0.0105593577,
0.0583639033,
0.0832009688,
0.0584677942,
-0.1007523686,
-0.0233269446,
-0.0505688898,
-0.0054223747,
-0.0496103875,
-0.0232575051,
-0.0209217519,
0.0196226425,
0.030035682,
-0.0667568296,
-0.0604668632,
0.0024130885,
0.025134841,
-0.0047006109,
-0.0135978162,
-0.0319945179,
-0.0090121841,
-0.0173317418,
0.0041606175,
-0.0189875234,
0.0121796615,
-0.0611736439,
-0.0381373838,
0.0393610522,
-0.0452727564,
-0.0164111983,
-0.0066985241,
-0.0265961867,
-0.0833796039,
0.0038237087,
0.0601583384,
0.0238821656,
0.0671403036,
-0.0159333386,
-0.0711237639,
0.0487225465,
0.0697359592,
-0.0173035339,
-0.0293370504,
0.0515910834,
0.0282567609,
-0.0031179895,
0.0439400226,
-0.0532583743,
0.0238827392,
0.0275899414,
0.0176624991,
-0.0613386594,
-0.0493255407,
0.0227545332,
0.0068452721,
0.0338494927,
0.0015114931,
-0.0309273638,
-0.0080179386,
0.023952106,
-0.0184931904,
-0.0326142944,
-0.033423312,
0.0013716862,
-0.0156979468,
-0.0374733135,
-0.0095872721,
-0.0222015176,
-0.0687471926,
0.0296891555,
-0.0275711715,
0.0340816192,
0.0235030577,
0.0353042968,
0.0247663278,
-0.0059904945,
-0.0429642536,
0.0221616887,
0.0374466218,
0.028009329,
-0.008706918,
0.057080552,
0.02185807,
-0.0073691085,
0.0011610108,
0.0248636678,
-0.0647011846,
-0.0329198614,
-0.0288876165,
-0.0109780561,
0.0223646779,
-0.0179869272,
0.0353596546,
-0.0167725701,
-0.0188322794,
-0.0123819923,
0.0387421437,
0.0229673777,
-0.0410086736,
0.0441502556,
0.0227990951,
-0.0255701151,
-0.0005853559,
0.0546807647,
-0.033100754,
0.0234710611,
0.0022126532,
0.0270999447,
-0.0448067226,
0.0050076423,
0.0326188505,
0.0065630018,
-0.0039190231,
0.0106417229,
-0.0595843568,
-0.0756582692,
-0.0081907464,
0.0127375228,
-0.0072753788,
-0.022265967,
0.0473691449,
0.0472370423,
0.0087305456,
-0.0229236558,
-0.0604702756,
0.0133694448,
-0.0169230886,
-0.0200325772,
-0.0630288273,
-6.33645087e-33,
0.0076034046,
-0.0563351065,
0.0049563004,
0.0019171467,
0.0475867689,
-0.0159409307,
-0.0191671662,
0.0026588496,
-0.0237910859,
0.0082978467,
-0.0297539867,
0.0001119101,
0.0016717596,
0.0429294445,
-0.0047064265,
-0.0144960983,
0.0568450876,
-0.0055158436,
-0.0238253959,
0.0384480283,
0.0229049567,
0.006988598,
0.0117378086,
-0.0779182464,
0.0327370353,
-0.0308458116,
-0.0200933535,
0.0370210893,
-0.0164970513,
0.0739586502,
-0.0159617439,
0.0618003979,
0.01859322,
-0.0188062247,
-0.0236888155,
0.0380890109,
-0.0362868905,
-0.0386810675,
0.0359256901,
-0.02204437,
-0.0668551847,
-0.0386356674,
0.041657988,
-0.0264612585,
-0.014628781,
0.0070151677,
0.0483609103,
-0.0103491303,
-0.0033818274,
-0.0233042296,
-0.0719430968,
0.0111358343,
-0.0023974909,
-0.0267784353,
-0.0071457443,
0.0808190927,
0.0021047574,
0.0360359214,
-0.0368728302,
0.018895844,
0.0201945473,
0.0029380016,
0.0719514638,
0.0305679329,
-0.0257567894,
0.0649588853,
-0.0065712999,
0.0218175221,
0.0177601725,
-0.0408703983,
-0.0106379893,
0.0090465648,
0.0472906008,
0.0966900736,
0.0224836022,
-0.0114521636,
-0.0431349315,
0.0320444815,
0.0131196203,
-0.0082169855,
0.049002666,
-0.0071783359,
-0.0266860351,
-0.0061360681,
-0.0287291147,
-0.0263762623,
-0.038870696,
-0.0291418061,
0.0141953919,
0.0025072847,
-0.0536899418,
-0.023406174,
0.0120395273,
-0.0406047665,
-0.0141656538,
0.0409201421,
-0.0386200882,
-0.0008294532,
0.0195344295,
-0.004889119,
-0.0448002405,
0.0017947038,
0.0155788241,
-0.0074384958,
0.0080894418,
0.0352886841,
0.0565594435,
0.0029634284,
-0.0836860687,
0.0074165678,
-0.026795527,
0.0537743904,
-0.0299628731,
-0.0378112607,
0.0268351994,
-0.0016690425,
-0.0160279591,
0.0780779347,
0.0120317787,
-0.0290036593,
-0.0588555932,
-0.0298463944,
0.0025502271,
0.0104711968,
-0.0481717214,
-0.014656947,
0.014921613,
0.0242535826,
0.0423184372,
-0.0355420895,
-0.0177098364,
0.0019531455,
2.573e-7,
0.0465318039,
0.0456793532,
0.0047496539,
0.0512356311,
0.0161734037,
0.0655665249,
0.0110826176,
0.0094716959,
0.0275734477,
-0.0014852269,
0.0120917838,
0.011798772,
0.0420729481,
0.0344800688,
-0.0846320614,
0.0232213307,
-0.0900132954,
0.0035626686,
-0.0218995139,
0.0145321852,
0.108246021,
0.0828208178,
0.0102435658,
-0.0112878866,
0.0024296031,
-0.0175509285,
0.0500820354,
-0.0469990596,
0.0073490464,
-0.0057765315,
0.0116801327,
-0.0236198325,
0.0248101112,
0.0190484226,
-0.0213651285,
-0.0333620124,
-0.0461812764,
0.0808482617,
0.010235603,
-0.0733863786,
-0.0032904956,
-0.0002407595,
0.0441161878,
-0.0608815663,
0.0715988055,
-0.0549995005,
-0.0196887609,
-0.0719186068,
-0.0081157433,
-0.0223272052,
0.0000333428,
-0.042447824,
-0.030756956,
0.0039406144,
0.0004389255,
0.0112645943,
-0.0242628641,
-0.0791291147,
-0.0326153263,
0.0121508697,
-0.0125277322,
-0.0331394151,
-0.0280498751,
0.048198089,
0.0339549705,
-0.0428193547,
0.0353966616,
2.413666198e-34,
0.0522321165,
0.0083124619,
0.011519298,
0.0533741452,
0.0097060772,
0.0403549746,
-0.0065157204,
0.0065533239,
0.0698543563,
-0.0862984955,
-0.0310194492
] |
http://arxiv.org/abs/2401.17377v3 | http://arxiv.org/pdf/2401.17377v3 | Infini-gram: Scaling Unbounded n-gram Language Models to a Trillion Tokens | 2024-01-30T19:03:49 | 2024-04-04T17:28:38 | Jiacheng Liu, Sewon Min, Luke Zettlemoyer, Yejin Choi, Hannaneh Hajishirzi | cs.CL | cs.CL, cs.AI, cs.IR | Are $n$-gram language models still relevant in this era of neural large
language models (LLMs)? Our answer is yes, and we showcase their values in both
text analysis and improving neural LLMs. This was done by modernizing $n$-gram
LMs in two aspects. First, we train them at the same data scale as neural LLMs
-- 5 trillion tokens. This is the largest $n$-gram LM ever built. Second,
existing $n$-gram LMs use small $n$ which hinders their performance; we instead
allow $n$ to be arbitrarily large, by introducing a new $\infty$-gram LM with
backoff. Instead of pre-computing $n$-gram count tables (which would be very
expensive), we develop an engine named infini-gram -- powered by suffix arrays
-- that can compute $\infty$-gram (as well as $n$-gram with arbitrary $n$)
probabilities with millisecond-level latency. The $\infty$-gram framework and
infini-gram engine enable us to conduct many novel and interesting analyses of
human-written and machine-generated text: we find that the $\infty$-gram LM has
fairly high accuracy for next-token prediction (47%), and can complement neural
LLMs to greatly reduce their perplexity. When analyzing machine-generated text,
we also observe irregularities in the machine--$\infty$-gram agreement level
with respect to the suffix length, which indicates deficiencies in neural LLM
pretraining and the positional embeddings of Transformers. | Infini-gram: Scaling Unbounded n-gram Language Models to a Trillion Tokens: Are $n$-gram language models still relevant in this era of neural large
language models (LLMs)? Our answer is yes, and we showcase their values in both
text analysis and improving neural LLMs. This was done by modernizing $n$-gram
LMs in two aspects. First, we train them at the same data scale as neural LLMs
-- 5 trillion tokens. This is the largest $n$-gram LM ever built. Second,
existing $n$-gram LMs use small $n$ which hinders their performance; we instead
allow $n$ to be arbitrarily large, by introducing a new $\infty$-gram LM with
backoff. Instead of pre-computing $n$-gram count tables (which would be very
expensive), we develop an engine named infini-gram -- powered by suffix arrays
-- that can compute $\infty$-gram (as well as $n$-gram with arbitrary $n$)
probabilities with millisecond-level latency. The $\infty$-gram framework and
infini-gram engine enable us to conduct many novel and interesting analyses of
human-written and machine-generated text: we find that the $\infty$-gram LM has
fairly high accuracy for next-token prediction (47%), and can complement neural
LLMs to greatly reduce their perplexity. When analyzing machine-generated text,
we also observe irregularities in the machine--$\infty$-gram agreement level
with respect to the suffix length, which indicates deficiencies in neural LLM
pretraining and the positional embeddings of Transformers. | [
0.038265653,
0.0833051279,
0.001057053,
0.0469630584,
-0.0316481628,
0.0148742413,
0.0014574797,
0.0148383956,
0.0006811061,
-0.0174881723,
-0.008568502,
-0.0712972358,
-0.0136920949,
0.0399664156,
0.0566221811,
0.0032543074,
0.0443983488,
-0.0277106222,
-0.0284624472,
0.0094478317,
0.0270822085,
0.003042253,
0.0161560979,
0.0470655337,
-0.0384601057,
-0.0250157267,
-0.0125306249,
0.0119210249,
-0.01267948,
-0.0276839398,
0.0058555412,
0.0255738609,
0.0055218008,
0.0556725785,
0.0000016621,
-0.0184678528,
0.0162738636,
-0.01184035,
-0.0311873127,
0.0519907996,
0.0340633877,
-0.0209972654,
-0.0214644317,
0.0249593742,
-0.026820071,
-0.0264957324,
0.0286345761,
0.0384509712,
0.0467611291,
0.1054989025,
-0.0024913843,
-0.0339318961,
0.0714306533,
-0.0692972764,
-0.0022411698,
-0.0229127835,
0.0465084538,
-0.0555771776,
0.0079980502,
0.0137896072,
-0.0023735596,
0.0315237902,
0.0276025478,
-0.0243187677,
0.0080452925,
0.0391793139,
0.0312850289,
-0.0099615445,
-0.023289321,
0.0461528599,
-0.0457576886,
-0.0306010488,
0.0135213463,
0.0161186568,
-0.0133445663,
-0.0222211462,
-0.0449590832,
0.0447632447,
0.0151026165,
0.0448632464,
0.0456618853,
-0.0090889838,
0.0494804271,
0.036539387,
-0.0229105446,
0.0556817092,
0.0353136659,
-0.024362972,
-0.0377039388,
0.0188352615,
-0.0395815969,
-0.0617624708,
0.0353424735,
0.0224152636,
-0.0256818105,
0.0169512983,
-0.0458112136,
-0.0509411804,
0.0453561917,
-0.0402317308,
0.0416177846,
0.0122193443,
0.014508306,
0.0450255796,
-0.0236946307,
0.0363261178,
-0.03411806,
-0.0160290226,
-0.0475929081,
-0.0353087038,
-0.0009869505,
-0.0139359431,
-0.0289552473,
0.0197664723,
-0.0438758992,
-0.0205717459,
-0.042212598,
0.023966847,
0.0226545576,
0.0074231746,
0.0171268266,
0.0132634854,
-0.0161499865,
-0.0269395448,
-0.009861392,
0.0232630521,
-0.036010813,
0.002792344,
0.0183344726,
-0.0456497036,
-0.0250791684,
0.0427115746,
-0.0035782172,
-0.013535873,
0.02711384,
0.0577583723,
-0.0150955245,
-0.0253590811,
-0.0054034372,
-0.0124406116,
-0.0168019682,
-0.0578328669,
0.0107968627,
-0.078970246,
0.0154932663,
0.0147125395,
-0.0205619112,
-0.0410995483,
0.0074245827,
0.0718874708,
-0.0382102057,
0.0488323085,
0.0081519764,
0.022642279,
0.0076910066,
-0.0029951194,
0.0316709951,
0.0217811167,
0.0235138312,
0.0279624462,
0.0302325822,
0.005593277,
-0.0058770534,
0.0766484588,
0.0060366341,
0.0227972288,
-0.0256212875,
-0.0128266541,
0.0324631333,
0.0525040329,
0.04360798,
0.0940724313,
-0.0161794573,
-0.0235399492,
0.0726813227,
0.0567905977,
-0.0072744251,
0.0284380522,
0.0443183519,
-0.0085791023,
0.0080645271,
-0.0069710584,
-0.0445444435,
0.0207703877,
-0.0052189417,
0.0081104236,
0.0598311201,
0.047745347,
-0.0425965935,
-0.024712747,
-0.0143206324,
-0.0294866264,
-0.0099578984,
-0.0352078192,
0.0393008552,
-0.00574642,
-0.0457832143,
0.0656185672,
0.0382199436,
-0.0442482308,
0.0365811139,
-0.0433727615,
0.0435292423,
0.0484459698,
0.0547666699,
-0.0065919263,
-0.0065039168,
-0.0002979655,
-0.0000345043,
0.0776789486,
-0.0087621771,
0.0019112813,
-0.0077864695,
-0.0319758169,
0.0396963097,
0.031978976,
0.0219587367,
0.0222340282,
0.0219885614,
0.01392892,
0.0038631076,
-0.032796707,
0.0037087656,
0.0489571877,
-0.027177779,
-0.0025210157,
0.0013824484,
-0.0285792146,
0.0422374941,
0.0088110128,
-0.0183334108,
0.0339553431,
0.0609524585,
-0.059012264,
-0.0233079586,
-0.0512291789,
-0.0846367851,
-0.0313742049,
0.0573127791,
-0.0082941549,
-0.0152208805,
0.0321705304,
-0.0436804444,
-0.0302888732,
0.0103314258,
-0.0598745085,
0.0395179726,
-0.0809211358,
0.0297193024,
0.0274628773,
0.0309487805,
-0.0388534702,
0.0381295718,
-0.0204688571,
0.0333553553,
-0.0176781397,
0.0085216546,
0.0158283133,
-0.018365182,
-0.0571723506,
0.0080843056,
-0.0003063151,
0.0054720985,
-0.0111075444,
-0.0117634926,
-0.0103193754,
-0.0505598001,
-0.1257230639,
-0.0003704066,
-0.0173856243,
0.0182546452,
0.0070408727,
0.0094894245,
-0.0011656383,
-0.0252694245,
-0.0379491597,
0.00873743,
0.0394781493,
0.0117271459,
0.0499917418,
0.0025968291,
-0.0152521329,
-0.0342644565,
-0.0351728909,
-0.0053181504,
0.0703502446,
0.0002758045,
-0.0393940546,
-0.0760311708,
-0.0006877941,
-0.0296540391,
0.0175914951,
-0.0146840066,
-0.0740307346,
-0.0118132513,
-0.0467287488,
0.0432474539,
0.0482621081,
0.0336943455,
0.0075052669,
0.0222876277,
0.0078458032,
-0.0324068628,
-0.011073851,
-0.0045619593,
-0.0020243768,
-0.0269702226,
0.0154483244,
-0.0031485721,
0.1059020758,
-0.0176590569,
-0.012352353,
-0.0044243126,
0.0296890456,
-0.0037992867,
-0.0059342925,
0.0156632736,
-0.0137154842,
-0.046859134,
0.0154161928,
-0.0264758989,
-0.0623736754,
0.0445470251,
-0.0587599687,
-0.0200279374,
-0.0268504955,
-0.0309788622,
-0.0510174036,
0.0590543449,
0.0280701146,
0.0079723084,
-0.016114248,
-0.0262720566,
-0.0332240127,
0.0216146372,
-0.0606660731,
-0.0231859237,
0.0887050852,
-0.0294679776,
0.0171601437,
-0.0101415571,
-0.0564719588,
-0.0440204814,
-0.0761567652,
-0.0403360613,
0.0009628157,
0.0141604468,
-0.0305432845,
-0.0435842462,
-0.0225351956,
-0.0002772435,
0.0374585167,
-0.026658779,
0.018647654,
-0.0237662513,
0.0512135215,
0.0726687163,
-0.0026036224,
0.0194691047,
-0.0315425359,
-0.0005610765,
0.0768575296,
-0.030224815,
0.012965533,
-0.0217923317,
-0.0171712209,
0.014205453,
-0.0155117065,
-0.0270440932,
0.021273803,
0.025464803,
-0.0253360309,
-0.0362014435,
0.1220477149,
0.0141440006,
-0.0619652495,
0.0135831982,
0.0303354487,
0.0615661778,
0.0018037736,
0.0179921705,
-0.0503794588,
-0.0405981392,
0.0173892844,
-0.011896641,
0.0068718791,
-0.0321327858,
-0.0144169293,
-0.1091360971,
0.0239827931,
0.008091216,
-0.0038317458,
-0.0224676039,
-0.0284833461,
-0.0149694588,
0.012553216,
0.0393620767,
0.0197730046,
-0.0143117988,
-0.0032371106,
0.0162643064,
-0.007778863,
-0.0256026518,
0.013048253,
-0.0690511391,
0.0324346572,
0.0405191109,
0.0152875036,
-0.0070567522,
-0.0415254571,
0.0511246063,
-0.0017354902,
0.0086033493,
0.0027823092,
0.0616186,
-0.0070521217,
-0.0251043122,
-0.0456627011,
0.0171066597,
-0.050505206,
0.0052319062,
-0.0033033874,
0.0837881342,
-0.010703898,
-0.005740847,
-0.0191877093,
-0.0213678665,
-0.0054313857,
0.0897586942,
0.0298551507,
-0.0675643384,
-0.0484071374,
-0.0395981222,
-0.0021079583,
-0.0119645735,
0.0037269313,
0.00297056,
0.0173956696,
-0.0000500592,
-0.0853063539,
-0.0360274725,
-0.0019629004,
0.0231042765,
-0.006660054,
0.0021334752,
0.0326117352,
-0.0389196575,
-0.0459560789,
0.0144134108,
-0.036267668,
0.0058912351,
-0.0779887736,
0.0308711454,
0.0637169331,
0.0076092868,
-0.0148280868,
0.0349065028,
-0.1320362985,
-0.0644957125,
0.0012671885,
-0.0098871179,
0.0204581488,
0.0145934224,
0.0010502371,
-0.0641157553,
0.0600541942,
0.0517680682,
-0.04339936,
0.0404619165,
-0.0033456329,
0.0155032817,
-0.002849156,
0.0635584667,
-0.0166650061,
0.0087077208,
0.0142256543,
0.0120243467,
-0.0315807797,
-0.0320409313,
-0.0175465047,
-0.0012490759,
0.0562204868,
-0.0069894022,
0.0397705212,
-0.0292962976,
0.0103877019,
-0.0288229417,
0.0053010811,
-0.0398781188,
-0.0075343754,
0.0497759022,
-0.0021313725,
-0.0465024225,
0.0345901921,
0.0157517046,
0.0314543359,
0.0077189906,
0.0083240103,
-0.0478722602,
-0.0183816683,
0.0489838868,
0.0027209828,
-0.0173120424,
-0.0627174675,
-0.0186625998,
0.012017915,
-0.0368979871,
0.0215282962,
0.0227591321,
0.0086687244,
-0.0284108985,
0.0168747455,
-0.061081171,
-0.0157776773,
-0.0291628186,
0.0195102245,
0.0811947957,
-0.012819388,
0.0253808312,
-0.0217132457,
-0.0060091824,
-0.0474183299,
0.0247592945,
-0.0267830007,
-0.022173455,
0.0366865285,
0.0071451329,
-0.028492583,
0.0309632514,
0.0914313868,
0.0054060863,
0.030208163,
0.0424335226,
0.010757192,
-0.033740446,
-0.0201204941,
-0.0037365414,
0.0044032605,
-0.0175408833,
-0.0073622381,
-0.0089598373,
-0.0671548396,
0.0444231965,
0.0192359686,
-0.0181605835,
-0.0233006496,
0.0267547835,
0.0314089209,
-0.0020262678,
0.0271326918,
-0.0178611204,
0.025806183,
-0.0147271734,
-0.0688878894,
-0.0248656049,
-6.217225621e-33,
-0.005295421,
-0.0243772808,
0.003873833,
-0.0194596015,
-0.0294478182,
-0.008760185,
-0.0320895612,
-0.007906707,
0.0067911665,
0.0134346746,
-0.0464820564,
0.007970579,
-0.000273609,
0.0710709766,
0.0395238474,
-0.0087112272,
0.0455546975,
-0.0356525965,
0.0371414125,
0.0059600985,
0.0067206128,
0.0365115479,
0.0317115076,
-0.0267354362,
-0.0108967144,
-0.0289609395,
-0.0086924685,
0.045619674,
-0.0177792907,
0.0547653995,
-0.0336723514,
0.0250059441,
0.0351607166,
-0.0257730186,
-0.0303557012,
0.0407990888,
-0.0672221854,
-0.0285105743,
-0.0035476799,
-0.0043829139,
0.0076444144,
-0.0353702754,
0.047734119,
-0.022385465,
-0.0259349458,
0.0093125757,
0.0157283489,
0.0078144697,
-0.0144215189,
-0.0036830879,
-0.0137081454,
0.020075934,
0.0187008139,
0.0061910031,
-0.0175128561,
-0.0155922119,
0.009787376,
0.0308738686,
-0.0607168786,
-0.0044773961,
-0.0383960456,
0.0657165945,
0.0316994376,
0.0154089462,
-0.0054731136,
0.0762692839,
0.0495226458,
-0.0160271991,
-0.0440547615,
0.0202939529,
0.0175911617,
0.0027226943,
0.0738418102,
0.0533879325,
0.0613114834,
-0.0015535975,
-0.0869831964,
0.0103484308,
0.0350679457,
0.0418724418,
0.0248330496,
0.0117115844,
-0.0326611511,
-0.0271523595,
-0.0169105288,
-0.0365958326,
-0.027441198,
-0.0192496181,
0.0248234551,
0.0146304555,
0.0339536332,
-0.0084869191,
0.0251098257,
-0.0379705429,
-0.0323526971,
0.0196274612,
-0.0059250714,
0.0147486329,
0.0115847811,
0.0326568708,
-0.0521880277,
-0.018673921,
-0.0256852955,
0.0073070861,
-0.0046333489,
0.0072223358,
0.0493221432,
-0.0069764266,
-0.0886736661,
0.0089054909,
-0.0274928622,
0.0258126073,
0.0240777209,
0.0222801901,
0.0495030172,
-0.0041899332,
0.0028082416,
0.0545946434,
-0.0036075795,
0.0158022512,
-0.000779334,
-0.033945296,
0.0242340472,
0.006566348,
-0.0318826959,
0.0105158761,
-0.0669713095,
0.0206245091,
0.0467245728,
-0.0577729717,
0.0184446927,
0.0412355661,
2.383e-7,
0.0338296779,
0.0510554723,
0.0475521609,
0.008967882,
0.060603939,
0.0226824041,
-0.0203123651,
0.034703441,
-0.0072835241,
0.0076076766,
-0.0072802836,
0.0052349246,
0.0572639331,
-0.0056528258,
-0.1388415545,
0.0319352821,
-0.0958182514,
0.0233821608,
-0.0561273731,
-0.0048084632,
0.0203334764,
0.1370304227,
0.0112325968,
-0.0220477208,
-0.0084224688,
-0.0715074539,
0.0600877032,
-0.0153039144,
-0.0233436935,
0.0052247043,
0.022554189,
0.0024770708,
-0.0018308104,
-0.0103900405,
-0.0046096314,
0.0277380329,
0.0248587634,
0.0682174116,
0.009236943,
-0.0261172578,
0.0181620624,
-0.0343560427,
-0.0091384156,
-0.0536065139,
0.0535936467,
-0.0212629829,
-0.0149558783,
-0.0318185948,
-0.0070843096,
-0.0088666938,
0.0043306774,
-0.0133395521,
0.0006933246,
0.0221560746,
0.0375199728,
-0.0315040648,
-0.0120048905,
-0.06790106,
-0.0324389935,
-0.0149128288,
-0.0813766569,
-0.0200380385,
-0.0694444478,
0.0327810161,
0.0661236346,
-0.0631440803,
0.0027925859,
2.175319107e-34,
0.0421295427,
-0.0183266457,
0.0175257921,
-0.0116326893,
0.0304304436,
-0.0124367299,
-0.0002906824,
0.0281479899,
-0.0029018256,
-0.115431577,
-0.0182787199
] |
http://arxiv.org/abs/2401.17373v1 | http://arxiv.org/pdf/2401.17373v1 | Arabic Tweet Act: A Weighted Ensemble Pre-Trained Transformer Model for Classifying Arabic Speech Acts on Twitter | 2024-01-30T19:01:24 | 2024-01-30T19:01:24 | Khadejaa Alshehri, Areej Alhothali, Nahed Alowidi | cs.CL | cs.CL, cs.AI | Speech acts are a speakers actions when performing an utterance within a
conversation, such as asking, recommending, greeting, or thanking someone,
expressing a thought, or making a suggestion. Understanding speech acts helps
interpret the intended meaning and actions behind a speakers or writers words.
This paper proposes a Twitter dialectal Arabic speech act classification
approach based on a transformer deep learning neural network. Twitter and
social media, are becoming more and more integrated into daily life. As a
result, they have evolved into a vital source of information that represents
the views and attitudes of their users. We proposed a BERT based weighted
ensemble learning approach to integrate the advantages of various BERT models
in dialectal Arabic speech acts classification. We compared the proposed model
against several variants of Arabic BERT models and sequence-based models. We
developed a dialectal Arabic tweet act dataset by annotating a subset of a
large existing Arabic sentiment analysis dataset (ASAD) based on six speech act
categories. We also evaluated the models on a previously developed Arabic Tweet
Act dataset (ArSAS). To overcome the class imbalance issue commonly observed in
speech act problems, a transformer-based data augmentation model was
implemented to generate an equal proportion of speech act categories. The
results show that the best BERT model is araBERTv2-Twitter models with a
macro-averaged F1 score and an accuracy of 0.73 and 0.84, respectively. The
performance improved using a BERT-based ensemble method with a 0.74 and 0.85
averaged F1 score and accuracy on our dataset, respectively. | Arabic Tweet Act: A Weighted Ensemble Pre-Trained Transformer Model for Classifying Arabic Speech Acts on Twitter: Speech acts are a speakers actions when performing an utterance within a
conversation, such as asking, recommending, greeting, or thanking someone,
expressing a thought, or making a suggestion. Understanding speech acts helps
interpret the intended meaning and actions behind a speakers or writers words.
This paper proposes a Twitter dialectal Arabic speech act classification
approach based on a transformer deep learning neural network. Twitter and
social media, are becoming more and more integrated into daily life. As a
result, they have evolved into a vital source of information that represents
the views and attitudes of their users. We proposed a BERT based weighted
ensemble learning approach to integrate the advantages of various BERT models
in dialectal Arabic speech acts classification. We compared the proposed model
against several variants of Arabic BERT models and sequence-based models. We
developed a dialectal Arabic tweet act dataset by annotating a subset of a
large existing Arabic sentiment analysis dataset (ASAD) based on six speech act
categories. We also evaluated the models on a previously developed Arabic Tweet
Act dataset (ArSAS). To overcome the class imbalance issue commonly observed in
speech act problems, a transformer-based data augmentation model was
implemented to generate an equal proportion of speech act categories. The
results show that the best BERT model is araBERTv2-Twitter models with a
macro-averaged F1 score and an accuracy of 0.73 and 0.84, respectively. The
performance improved using a BERT-based ensemble method with a 0.74 and 0.85
averaged F1 score and accuracy on our dataset, respectively. | [
0.0313179083,
0.0575267673,
0.0116727259,
-0.0094073908,
-0.0233538523,
0.032184802,
0.0270562433,
-0.0150247607,
0.0026526358,
-0.0683131814,
0.008398816,
-0.095264107,
0.0042058644,
0.0245179702,
-0.0102114584,
-0.0081448872,
0.0068565924,
0.0149690015,
0.0192747731,
-0.0386269949,
0.0338395685,
0.030199239,
-0.0282012485,
0.0340638049,
-0.0199283157,
0.022656139,
-0.0298259016,
-0.0253382046,
0.0284350645,
0.0441740677,
0.057170026,
0.0867583677,
0.0174163841,
0.0676429942,
0.0000019867,
-0.0592707545,
-0.0155743342,
-0.0418933891,
-0.0106741926,
-0.0111476174,
-0.0081532057,
-0.0156258978,
-0.0208318867,
0.0260671787,
-0.0371747464,
-0.0034845434,
0.053864494,
0.0772181526,
0.0658846647,
0.0907157138,
-0.0046274266,
-0.0207992978,
0.0299625006,
-0.0121091194,
0.0088585541,
0.0021175337,
0.0188161843,
-0.0498496443,
-0.0380290113,
-0.0004992152,
-0.0013115699,
0.0193938352,
0.0154119488,
-0.0416949131,
-0.0072315922,
-0.0410453305,
-0.0883250311,
-0.0326066054,
0.022617247,
0.0258710366,
-0.0031585002,
-0.0354733691,
0.0044137994,
-0.0057167145,
0.0082404111,
0.0110610547,
-0.0099293627,
0.0320087224,
0.0221223515,
0.0152367251,
0.0035572136,
0.0336722881,
0.0178484693,
0.052903071,
-0.069826223,
0.0205418728,
-0.0340018719,
-0.0316544808,
0.0015919664,
0.0240635686,
-0.0203588828,
-0.0287586395,
0.0532562882,
0.0036436834,
-0.0258489531,
-0.0092976941,
-0.0092057297,
-0.0595307015,
0.016501572,
-0.0837945491,
0.0044616894,
0.0499469005,
-0.0209973603,
0.0178767256,
-0.0290358514,
0.0132076256,
0.0223409776,
0.0397826284,
-0.0086803511,
-0.0053060218,
-0.0688129663,
-0.0022370368,
0.011496596,
0.0639212579,
-0.0259246603,
0.0102690225,
-0.078629978,
-0.0037722432,
-0.0248993188,
-0.0151616689,
0.053975109,
0.0090206331,
-0.0049008965,
-0.0013300658,
0.0002393807,
0.0442492329,
0.0249312446,
0.0311208647,
-0.0468570739,
-0.0418917499,
0.0358379334,
0.0037881478,
0.0490416437,
0.0341247804,
0.0051948805,
-0.0131290453,
-0.014463475,
-0.064609386,
-0.0315554701,
0.0363795757,
-0.0194642786,
-0.0114083476,
-0.0140716303,
-0.0636563376,
0.0365265608,
0.000270088,
-0.0230596755,
0.0396583937,
0.0262802038,
0.0153715406,
0.0279498771,
0.0063188788,
0.0268996079,
0.018434817,
0.0451572239,
0.018105628,
0.0514789298,
0.0795244575,
0.0440118313,
0.0414574035,
0.0326788798,
-0.0145842489,
0.0160243288,
0.051372245,
0.0201518368,
-0.0024058844,
-0.0061145648,
-0.0469699837,
-0.0729782805,
0.0268054698,
-0.0320208408,
0.0174794775,
-0.0192621723,
-0.0022570682,
0.0451724604,
0.0148258163,
0.0132688591,
0.0311616547,
0.0053765001,
0.0679256842,
0.0106511572,
-0.0341048948,
0.0266196914,
0.0068717818,
0.0007078202,
-0.0268067643,
-0.0000371298,
0.0614948235,
-0.0453130901,
-0.0554874614,
0.0288164839,
-0.0374371409,
0.0464773141,
-0.0187583733,
0.0435696878,
-0.0097781932,
-0.044061821,
0.0602544323,
0.0233930964,
-0.0714579076,
0.0026789848,
-0.0623474009,
-0.0320705511,
0.059325505,
-0.0261115264,
-0.0007656264,
-0.0453123115,
-0.0711952969,
0.0190817863,
0.0341515392,
0.0069832215,
-0.0015451708,
-0.0037197801,
-0.0531046651,
-0.0230868608,
0.0421036705,
0.0079776198,
0.0080713239,
-0.0057969014,
0.0358783938,
-0.0221647751,
0.0159925912,
-0.033806365,
-0.0111393668,
0.0067735696,
-0.0247101504,
0.0312681235,
0.0331171751,
0.0217122398,
0.0779481456,
-0.0241655279,
-0.0420335308,
0.0305090677,
-0.0047004358,
-0.0699095502,
0.0271751434,
-0.0395104289,
-0.0308930706,
0.0373608358,
0.004677135,
-0.0348852016,
0.0425565466,
-0.029146038,
-0.0525766201,
0.0508471504,
-0.0276244264,
0.052492395,
-0.059688285,
0.03325928,
-0.0678961277,
0.0197561346,
0.0237072688,
0.0287702046,
-0.0451685414,
-0.0206930581,
-0.0348680653,
-0.0252473336,
0.0405025668,
0.0432551317,
-0.0246377345,
0.0134378457,
-0.0016228681,
0.0452984087,
0.0584654063,
0.0072205351,
-0.0171884019,
-0.0226880107,
-0.1096415073,
-0.02037872,
-0.0563566275,
0.0159521196,
-0.0259783734,
0.0545264408,
-0.0074148341,
0.0139994053,
-0.0215070639,
-0.044283621,
0.0111486958,
0.0034925416,
0.0098762382,
-0.0306689963,
-0.0002484381,
-0.0106758503,
-0.000655078,
-0.0259525198,
0.0168599393,
0.0028106009,
-0.0550004877,
0.0549410544,
-0.0629366189,
0.0138224242,
0.0194792859,
0.0130002005,
-0.0839119405,
0.0045498819,
-0.0070524607,
-0.0010150832,
0.0113203404,
0.0584138334,
0.0552646406,
-0.0189865958,
0.0288888142,
-0.0511055961,
0.0131134633,
-0.0261061061,
0.0031926876,
-0.0395550765,
-0.0398260169,
-0.024805909,
0.0428231098,
-0.0351627842,
-0.0044301227,
-0.0711844936,
-0.0473081805,
0.0022494041,
-0.0126329372,
0.0050476552,
-0.0159009732,
-0.036291074,
0.0111539979,
0.01311999,
-0.0252501331,
0.0175612252,
0.0087431148,
0.0011779909,
-0.0664086789,
-0.0178804565,
-0.0269909985,
0.0211613588,
0.015781017,
0.0124070263,
-0.0145483818,
-0.0442201607,
0.0189162381,
-0.016540207,
-0.0223280396,
-0.0388951451,
0.0281064212,
0.0359755158,
-0.0020038285,
-0.0018277008,
-0.0182536785,
-0.0109689869,
-0.0335096829,
-0.0346499644,
0.0073298924,
0.0216901973,
-0.0654972568,
-0.0093883956,
-0.0023970117,
0.0109747816,
-0.0024702416,
0.0037450381,
0.0178221595,
-0.0035793558,
-0.0103224562,
0.0156298857,
0.0065735863,
-0.0239410046,
0.0320526846,
-0.0000265859,
0.0117362756,
-0.0239516702,
-0.0115220491,
-0.0731218234,
-0.0360996388,
0.0430988967,
-0.0448391698,
-0.0020689135,
-0.037562836,
0.0195104592,
-0.0172534268,
0.0534968413,
0.1193297133,
0.0179785918,
-0.0186133776,
-0.0168378502,
0.0139885377,
-0.0131427348,
-0.021973677,
-0.0070212292,
-0.0872552767,
-0.0016759593,
0.0068068882,
-0.0251240712,
-0.0346761793,
-0.012937706,
0.0547604673,
-0.0879284143,
0.0101420004,
-0.0255271439,
0.0033942729,
-0.0303094164,
-0.0195867941,
-0.0051825354,
-0.0264993887,
0.0302146394,
-0.0681846067,
-0.0101768691,
-0.0064067752,
0.0602517575,
-0.0520052016,
-0.0837184042,
0.0191795658,
-0.0235760827,
0.0274044462,
0.0119684665,
0.0322476029,
-0.0696718171,
-0.0418585613,
0.0030326915,
-0.0151735414,
-0.0327981412,
0.0060611726,
0.0194519721,
0.0018760677,
-0.0094916988,
-0.0104844384,
-0.0186430905,
-0.058286909,
-0.0004100608,
-0.0615099631,
0.0556824654,
-0.0151001168,
0.0046669343,
-0.0255502332,
0.0148459226,
0.0172083005,
0.0396078378,
-0.002354569,
-0.0639409646,
0.0224020742,
-0.0029112401,
-0.0147106024,
-0.0942266732,
-0.0041839834,
-0.00974546,
0.0303697903,
-0.0400652252,
-0.042823419,
-0.0425455943,
0.0022627115,
0.0028405846,
-0.0212226585,
0.0071730386,
0.0417966209,
-0.0018706699,
-0.0352263488,
-0.0093991878,
0.0505462028,
0.0257272888,
0.0046540117,
-0.0523441769,
0.0391812995,
-0.0211055651,
0.0216492023,
-0.0737263635,
-0.0248389449,
-0.0340757445,
0.0368803106,
0.0007875451,
0.0093802651,
0.0115040913,
-0.0138788428,
-0.059066575,
0.0554655045,
0.0626157895,
0.0216372479,
-0.0338327996,
0.086018756,
0.0254978724,
-0.03768076,
0.0172370616,
0.0107319122,
0.0074233264,
-0.0083964784,
-0.04204778,
-0.0603201315,
-0.0282498579,
-0.0227465238,
-0.0546878837,
0.033323884,
0.0125847058,
-0.0725148246,
-0.0498964787,
-0.0051132995,
0.0205532704,
0.0249294993,
-0.0189195853,
0.0096358946,
0.0382368341,
0.0262340959,
-0.0331391282,
0.0297258627,
-0.0353371277,
-0.0387098119,
-0.0260797944,
0.0957683697,
0.0091684507,
0.0394064188,
-0.0161253605,
0.0222465228,
-0.0246709567,
-0.0409922674,
0.0607319213,
-0.0098652244,
0.0017373535,
-0.0059639374,
-0.0032018151,
0.0259095319,
0.03968779,
0.0607855916,
-0.004011794,
0.0165071953,
-0.0513584018,
0.036129076,
0.0364201479,
-0.066069372,
0.0554858223,
0.0231795851,
0.0132106561,
-0.0055049402,
0.0194399133,
-0.0460835434,
0.0003583016,
0.0003955523,
-0.0139304856,
-0.0149117513,
0.0578699484,
0.0268208906,
-0.0565452054,
0.0291002952,
0.0510598347,
0.0096885506,
-0.0146316709,
-0.0179845635,
0.0605473854,
-0.0026300915,
-0.0337528735,
0.02301763,
-0.0003847675,
-0.0570224077,
0.0634736866,
-0.0459367894,
0.0265598502,
-0.0003157812,
0.0251962934,
0.0768326297,
-0.0082042078,
0.0506626926,
-0.0588856451,
-0.0286093168,
-0.0022908016,
-0.0241590887,
0.0028227989,
-6.270219845e-33,
0.0493449457,
0.0085707065,
-0.0206858069,
0.0132640023,
-0.0189617649,
-0.0095659671,
-0.0084170979,
0.0139417928,
-0.0187452734,
0.0528297685,
-0.0207217094,
0.0359424911,
0.0107832467,
0.0590669662,
0.0287281889,
-0.0390918441,
0.0323570967,
0.0055886563,
-0.01093305,
0.0467718467,
0.0436799638,
0.0392141752,
0.0870108604,
-0.0951843038,
-0.0048923413,
-0.0134570934,
-0.030838348,
0.0057910071,
0.036437124,
0.0848445073,
-0.0009899908,
0.0980543047,
0.0381635278,
-0.0391860381,
0.038333822,
0.0614010952,
0.0023165806,
-0.0241785496,
-0.0224212613,
0.0424146876,
0.0150132542,
-0.0276249461,
-0.0057458202,
0.015777057,
-0.0689990297,
0.031537924,
0.0251648799,
0.0104538295,
0.038301833,
-0.0734559,
-0.0707479566,
0.0360743999,
-0.022063449,
-0.0335577875,
0.0221534669,
0.0432340093,
0.0210789703,
0.0156161124,
-0.0597438999,
0.0170602612,
0.0282520354,
0.0068479721,
0.0154254111,
-0.0112326751,
-0.0498519018,
0.0339477435,
-0.0127298245,
0.0041930862,
0.0168903749,
0.030704502,
0.0208385829,
-0.002346199,
0.0198393259,
0.079907991,
0.0116897915,
-0.0253956858,
-0.0741835982,
0.045040898,
0.0513533689,
0.0327607729,
0.0067781019,
-0.0109319091,
-0.0313040949,
-0.0017199778,
0.0174575616,
-0.0632058308,
-0.011568849,
-0.0015671509,
-0.0303659812,
0.0128701609,
-0.0168447345,
0.0693656728,
0.0310494695,
-0.0424765237,
-0.0124513544,
0.021173954,
-0.03319446,
0.009268729,
0.0185239445,
0.0505086742,
-0.0465407632,
-0.0496609919,
0.036538709,
0.0417116545,
0.0117698517,
0.0254484881,
0.0456241444,
0.0772047639,
-0.0522940978,
0.0178600047,
0.0114812218,
0.0293398406,
-0.0590666085,
0.0767051354,
-0.020261392,
0.0088919932,
-0.0079800487,
0.0428232737,
0.0221171044,
0.0178094041,
-0.0043121129,
-0.030022148,
-0.0341277979,
-0.0233863592,
-0.0133063206,
0.0075393589,
-0.0213634707,
0.0361346714,
-0.0010414969,
-0.0766890943,
0.005312949,
-0.0412781201,
2.807e-7,
0.0259100515,
0.055749137,
-0.0112922201,
-0.0531716458,
0.0042234012,
0.0393681265,
0.0281749237,
0.0243212525,
0.0189612638,
0.0076461402,
0.0168196633,
-0.0151576381,
0.0308224391,
-0.0095392894,
-0.0842322931,
-0.0305506457,
-0.0833331347,
0.0257219411,
-0.0281682741,
0.0352123976,
0.0445154868,
0.1309492439,
0.0040568835,
-0.0305934101,
-0.0262173098,
-0.0161153805,
0.0341157615,
-0.0384347439,
-0.0244026203,
0.034945108,
0.0630486757,
-0.0207962301,
0.0232397188,
-0.0048498069,
-0.0639857501,
0.0582019389,
0.0073060649,
0.0465696603,
0.01930275,
0.0141057754,
-0.0039465339,
-0.0309879687,
0.0143153677,
-0.0392825827,
0.0325629935,
-0.049142886,
-0.0542459749,
0.0116491085,
-0.0195737369,
0.0303425658,
0.0548077114,
-0.0294589642,
-0.0071287435,
-0.0001378274,
0.0353942923,
-0.0143057285,
-0.0206205584,
-0.0808840245,
0.038956292,
-0.0319962613,
-0.0216285307,
0.0007627154,
0.0155137787,
-0.0480025746,
0.0195370894,
-0.0017090933,
0.0428240821,
2.434159289e-34,
0.0471582301,
-0.0001354601,
0.0158865098,
0.0136798155,
-0.0009294078,
0.0083263675,
0.0283548161,
0.0042738141,
0.0084146177,
-0.0173281562,
-0.0053537465
] |
http://arxiv.org/abs/2401.17268v1 | http://arxiv.org/pdf/2401.17268v1 | Weaver: Foundation Models for Creative Writing | 2024-01-30T18:58:43 | 2024-01-30T18:58:43 | Tiannan Wang, Jiamin Chen, Qingrui Jia, Shuai Wang, Ruoyu Fang, Huilin Wang, Zhaowei Gao, Chunzhao Xie, Chuou Xu, Jihong Dai, Yibin Liu, Jialong Wu, Shengwei Ding, Long Li, Zhiwei Huang, Xinle Deng, Teng Yu, Gangan Ma, Han Xiao, Zixin Chen, Danjun Xiang, Yunxia Wang, Yuanyuan Zhu, Yi Xiao, Jing Wang, Yiru Wang, Siran Ding, Jiayang Huang, Jiayi Xu, Yilihamu Tayier, Zhenyu Hu, Yuan Gao, Chengfeng Zheng, Yueshu Ye, Yihang Li, Lei Wan, Xinyue Jiang, Yujie Wang, Siyu Cheng, Zhule Song, Xiangru Tang, Xiaohua Xu, Ningyu Zhang, Huajun Chen, Yuchen Eleanor Jiang, Wangchunshu Zhou | cs.CL | cs.CL, cs.AI, cs.LG | This work introduces Weaver, our first family of large language models (LLMs)
dedicated to content creation. Weaver is pre-trained on a carefully selected
corpus that focuses on improving the writing capabilities of large language
models. We then fine-tune Weaver for creative and professional writing purposes
and align it to the preference of professional writers using a suit of novel
methods for instruction data synthesis and LLM alignment, making it able to
produce more human-like texts and follow more diverse instructions for content
creation. The Weaver family consists of models of Weaver Mini (1.8B), Weaver
Base (6B), Weaver Pro (14B), and Weaver Ultra (34B) sizes, suitable for
different applications and can be dynamically dispatched by a routing agent
according to query complexity to balance response quality and computation cost.
Evaluation on a carefully curated benchmark for assessing the writing
capabilities of LLMs shows Weaver models of all sizes outperform generalist
LLMs several times larger than them. Notably, our most-capable Weaver Ultra
model surpasses GPT-4, a state-of-the-art generalist LLM, on various writing
scenarios, demonstrating the advantage of training specialized LLMs for writing
purposes. Moreover, Weaver natively supports retrieval-augmented generation
(RAG) and function calling (tool usage). We present various use cases of these
abilities for improving AI-assisted writing systems, including integration of
external knowledge bases, tools, or APIs, and providing personalized writing
assistance. Furthermore, we discuss and summarize a guideline and best
practices for pre-training and fine-tuning domain-specific LLMs. | Weaver: Foundation Models for Creative Writing: This work introduces Weaver, our first family of large language models (LLMs)
dedicated to content creation. Weaver is pre-trained on a carefully selected
corpus that focuses on improving the writing capabilities of large language
models. We then fine-tune Weaver for creative and professional writing purposes
and align it to the preference of professional writers using a suit of novel
methods for instruction data synthesis and LLM alignment, making it able to
produce more human-like texts and follow more diverse instructions for content
creation. The Weaver family consists of models of Weaver Mini (1.8B), Weaver
Base (6B), Weaver Pro (14B), and Weaver Ultra (34B) sizes, suitable for
different applications and can be dynamically dispatched by a routing agent
according to query complexity to balance response quality and computation cost.
Evaluation on a carefully curated benchmark for assessing the writing
capabilities of LLMs shows Weaver models of all sizes outperform generalist
LLMs several times larger than them. Notably, our most-capable Weaver Ultra
model surpasses GPT-4, a state-of-the-art generalist LLM, on various writing
scenarios, demonstrating the advantage of training specialized LLMs for writing
purposes. Moreover, Weaver natively supports retrieval-augmented generation
(RAG) and function calling (tool usage). We present various use cases of these
abilities for improving AI-assisted writing systems, including integration of
external knowledge bases, tools, or APIs, and providing personalized writing
assistance. Furthermore, we discuss and summarize a guideline and best
practices for pre-training and fine-tuning domain-specific LLMs. | [
0.0576822981,
0.0202822629,
-0.0077854688,
-0.0165999047,
-0.056119632,
-0.0247833468,
0.0089835934,
-0.0028528927,
0.0022567743,
-0.0489915051,
-0.0251477864,
-0.0229677893,
-0.019549692,
0.0508943684,
-0.006720202,
-0.0308610946,
0.0156855173,
-0.03053464,
0.0147794103,
0.0080839684,
0.0165758226,
0.0362620279,
0.002404148,
0.0619690083,
-0.0109864222,
-0.03766644,
-0.0116967056,
0.0321920812,
-0.0093926927,
-0.0110910218,
0.0005621881,
0.0546423793,
0.0675981939,
0.0396636091,
0.0000021284,
-0.0292395446,
0.0050233309,
-0.0075744763,
0.022067083,
-0.0117272837,
0.0319458656,
0.0063121826,
-0.0083967485,
0.0326239727,
-0.0236936659,
0.0188623257,
0.0513703376,
0.0418028794,
0.0586151741,
0.084292084,
-0.0161577649,
-0.0318956748,
0.0435279161,
-0.0227426123,
-0.0059695272,
0.0285856891,
-0.0066502467,
0.0169978421,
0.0552227013,
0.0116604529,
-0.0067068408,
0.0313571319,
0.0201117676,
-0.0467161946,
0.0399466492,
0.0317354724,
-0.0201196484,
-0.0075748074,
0.0031857279,
0.0258994829,
0.0377114415,
-0.0362497419,
-0.0059555564,
0.0102745872,
0.0013277145,
0.0187733211,
-0.0553509668,
0.0272970311,
-0.0076152263,
-0.0066085365,
0.0117035545,
0.0005288349,
-0.0167665537,
0.0222560856,
0.0410501175,
0.006155469,
0.0394104123,
-0.0038904196,
-0.02938403,
0.0058798878,
0.0168383494,
-0.0796146169,
0.0285219606,
0.0502411798,
-0.0019309829,
0.0030363947,
-0.0869597197,
-0.0418849327,
-0.0022253143,
-0.004518419,
0.0834877715,
0.0314783826,
0.043536935,
0.0181779228,
-0.0034546244,
0.0348662846,
-0.016124513,
-0.0023233425,
-0.0149078462,
0.0306303483,
-0.0561452992,
-0.0385866761,
-0.0663872585,
0.0476840585,
-0.044010397,
-0.0480745919,
-0.0409029312,
0.0457595512,
0.0189051516,
-0.0112957014,
0.0137522025,
0.0371469818,
-0.046079319,
0.0042922837,
-0.0200458206,
0.0051912214,
-0.0066812518,
0.0112209013,
0.0328158066,
-0.0112310527,
0.0047708098,
0.0243765172,
0.0305780917,
-0.0247219037,
0.0087017119,
0.0512969978,
-0.0364968404,
-0.0481792465,
-0.0389360376,
-0.0292420816,
0.0270849559,
-0.0643639416,
-0.0149253244,
-0.0767239556,
0.047830306,
0.0168531425,
-0.0216177665,
-0.0760652348,
0.002107186,
0.0534665249,
-0.0153230317,
-0.0094060544,
-0.0083131976,
0.0149006089,
0.0409549363,
-0.0032212671,
0.0131381722,
0.0141459936,
0.0871497542,
0.0762311742,
0.0141212279,
-0.0008718125,
0.0343054235,
0.002427815,
-0.0254492741,
-0.003983127,
-0.0291178506,
-0.0040982529,
-0.0352299176,
0.0622567534,
-0.0038567844,
0.0779811591,
-0.0283783264,
0.0414777845,
0.1063982472,
0.044018358,
0.045587305,
0.0529548116,
0.1063634753,
0.0390800759,
0.0206079017,
-0.0112157259,
-0.0399456434,
0.0506530479,
0.0053436328,
0.0374936499,
-0.0048579113,
0.017685743,
-0.0217738766,
-0.0396484472,
-0.0190580748,
-0.0164358392,
0.0018507444,
-0.0155453226,
0.0182261001,
0.0211542863,
-0.0787045062,
0.0036980023,
-0.0134962946,
-0.1013928652,
0.0153551092,
-0.032908503,
0.0370170139,
0.1083971635,
0.0920357332,
-0.0162925068,
-0.0012054482,
-0.0143521717,
0.0181606803,
0.0695721135,
0.0194160081,
0.0007157125,
-0.0097671561,
-0.0381232724,
0.0218439251,
-0.0404681452,
0.0008800955,
0.0165312029,
0.0130766872,
0.0109460419,
0.0173229575,
-0.0296724159,
-0.0849037319,
0.0514893383,
-0.0667932481,
-0.0710540488,
-0.0213280115,
-0.0057018986,
-0.0314494222,
0.0200964659,
-0.0549853109,
0.0256290622,
0.0236369949,
-0.0519099087,
-0.0261517446,
0.0148640666,
-0.0246245861,
0.0132900374,
0.0055215661,
-0.0249044504,
0.0402827822,
0.0178930964,
0.0158244129,
-0.0560927279,
0.0118568689,
-0.0389859229,
0.0530834384,
-0.0502044372,
0.0041186148,
0.0045979382,
-0.031602744,
-0.0189247541,
0.0384581797,
0.0100298375,
-0.0191457458,
0.0055242553,
0.0288040545,
-0.0062764939,
-0.0061827041,
-0.0125726862,
0.0040213792,
0.0114770317,
0.0385719687,
0.0201845355,
0.0106972503,
-0.0212696232,
-0.0666781515,
-0.0691818073,
0.0283632651,
-0.0166277532,
-0.0194511227,
0.0109876851,
0.006485654,
-0.0277268328,
0.0322995856,
-0.0493634827,
-0.0113597177,
0.0004519348,
0.0087360824,
0.0194591675,
0.0118914666,
0.0055292216,
-0.0398490131,
-0.0076640351,
-0.006425682,
0.0371148437,
0.0510121472,
0.0180481765,
-0.0569610596,
0.0583539084,
-0.0344130509,
-0.0044970056,
0.0032597694,
-0.0576019101,
0.012566424,
-0.0438418277,
0.0547276475,
0.0623058192,
0.0031762272,
0.029582087,
0.0747874975,
-0.0065562925,
-0.0146081522,
0.025338307,
-0.0031893924,
0.0159789212,
-0.0013062255,
-0.0403631888,
0.0187465698,
0.0614024252,
0.0249809306,
-0.0080752652,
-0.0569694079,
0.0306777712,
0.0162506979,
-0.0162988175,
0.0282287337,
-0.0501017086,
-0.0189675372,
0.0190170407,
-0.0267433655,
-0.0587492362,
0.034185186,
-0.0424370021,
-0.0058163167,
-0.0248141997,
0.0021190406,
-0.0358352698,
0.0615911447,
0.0159193873,
0.0351404957,
-0.0389144607,
0.0105870934,
-0.0597804971,
0.0109753218,
0.0399135463,
-0.0315962955,
0.026204804,
-0.0454972684,
0.0184120871,
-0.0227375589,
-0.0260998532,
-0.0453667156,
0.0073360866,
0.0000398022,
0.0099201873,
0.0225449651,
-0.0276578926,
-0.0212141052,
-0.0005990651,
0.035748858,
0.071053341,
-0.03299018,
-0.0062807878,
-0.046790354,
0.0024429322,
0.0523013659,
-0.028725272,
-0.0496137887,
-0.0380019583,
0.0389831774,
0.015898129,
0.0027683352,
-0.0104356101,
-0.0174095221,
-0.0157385413,
0.0263418294,
-0.0282304976,
-0.0210097451,
0.026536338,
0.0739017427,
-0.0465858467,
-0.0426229872,
0.1018965617,
0.0393654034,
-0.0562896058,
-0.0190609284,
0.0053975121,
0.043234434,
0.029382227,
0.0585704148,
-0.0326328091,
-0.0175979156,
0.014030735,
0.051532153,
-0.0095703565,
-0.0099756746,
-0.034448158,
-0.1048315465,
0.0545842983,
0.0095391599,
-0.0006885139,
-0.0319937058,
-0.0118866963,
-0.0410855636,
0.0550664663,
-0.0378720276,
0.0263020433,
-0.0388664007,
0.021721201,
-0.0128951855,
-0.0066479095,
-0.0338530876,
-0.0321437716,
-0.0337543041,
0.0125778755,
0.0601587556,
0.0547786839,
0.0082194572,
-0.0309405364,
-0.0210507773,
-0.0077582244,
0.0239333883,
-0.0087869819,
0.0356659703,
-0.0215050112,
0.0235117339,
-0.0539785065,
0.049594827,
-0.0221348051,
-0.0465785339,
-0.0047944137,
0.0270414297,
-0.033531867,
0.0154818408,
-0.0494280644,
0.0325621814,
0.0159261301,
0.0545260906,
0.0388310626,
-0.0831813738,
-0.0467442013,
-0.0323461443,
0.0090927491,
-0.0986154601,
-0.0299741104,
-0.0150412191,
-0.0368749052,
-0.0239836369,
-0.0893606246,
-0.027680669,
0.0148050515,
0.0728636086,
0.0466883592,
0.0094324443,
-0.0164910723,
-0.022046797,
-0.0893792585,
0.0124257272,
-0.0014236371,
0.0117462706,
-0.0654233396,
-0.027684439,
0.0032341185,
-0.0806833506,
0.007107222,
0.0244151056,
-0.106971547,
-0.0723291263,
0.0103164688,
0.0536284484,
0.0238532685,
0.0109061524,
0.005096477,
-0.0540891699,
0.0111529604,
0.0381180346,
-0.0265966896,
0.0428189598,
-0.0080078822,
-0.0083372248,
-0.011532221,
0.0166683011,
-0.0540207475,
0.0168684218,
0.0291504879,
0.0580726713,
0.0333415978,
0.0262353681,
0.021029871,
-0.0026472963,
-0.0298779085,
-0.0350133367,
0.0317308828,
-0.0768488646,
-0.0041899248,
-0.0574949197,
-0.0238594003,
-0.0503648706,
0.0200778022,
0.0189072601,
0.0057104831,
-0.0045600021,
0.0258367844,
-0.0139337089,
0.0313746557,
-0.0391253568,
0.0269414354,
-0.043726664,
0.0245818347,
0.0340408608,
-0.0091842152,
-0.0488617644,
0.0336146206,
-0.0666676909,
0.0405301228,
-0.0026001984,
0.0453014784,
0.01174164,
0.0120253228,
-0.0095220208,
0.0347927883,
-0.0668562353,
-0.0348887257,
-0.0222550984,
0.0172786489,
0.0577138625,
-0.0089823082,
0.0574953258,
-0.0437504053,
0.0103105363,
-0.0170940533,
0.0567102097,
0.0272562597,
-0.0362729914,
-0.0314002037,
0.0178442523,
0.0106527936,
-0.0332302563,
0.0801681206,
-0.0252530593,
0.0180193987,
-0.0022244796,
-0.012704134,
-0.0359391123,
-0.0301052202,
0.0385361873,
0.029194884,
0.0260788891,
-0.0025421144,
-0.0669519305,
-0.0457238331,
0.047482606,
0.0004876399,
-0.0089769186,
-0.0579604171,
0.0271292161,
0.057085298,
-0.0122865736,
-0.0157043375,
-0.0530746765,
0.012705883,
-0.0449452922,
-0.0267279726,
-0.0197731666,
-5.794688299e-33,
-0.037980888,
-0.0338265412,
0.0282095559,
-0.0023878997,
-0.0161256976,
0.0034799592,
-0.0250926241,
-0.0171264652,
0.0310762413,
-0.0105099855,
-0.0311673228,
-0.0044186446,
-0.0054136785,
0.0299019832,
-0.0128883161,
0.0552924685,
0.0668557361,
-0.0080163395,
-0.0563565306,
0.0257516149,
0.025994882,
0.0227573905,
0.021864038,
-0.0331475921,
-0.0122648217,
-0.0620422438,
-0.0120166214,
-0.0020058071,
0.0210377276,
0.0543532521,
-0.0157015081,
0.034322951,
0.0134960338,
0.0493087769,
-0.0007518578,
0.0041187583,
-0.0346447416,
-0.0204944555,
0.0077529987,
-0.0165885389,
-0.0299701039,
-0.0781283602,
0.0215719976,
-0.0353264287,
-0.0165649001,
0.0078906175,
0.0458701178,
0.004090081,
-0.0108517613,
-0.0391053744,
-0.0878130496,
-0.0026984948,
0.0137716532,
-0.0395213328,
-0.0295175556,
0.0297703948,
0.0181602538,
0.0109511893,
-0.0767774358,
0.028683085,
-0.0041122972,
0.0439317301,
0.0337441117,
0.0307122935,
-0.0097847823,
0.0604627617,
0.0021778122,
0.048226174,
-0.0158869419,
0.021119101,
0.0198433176,
0.0544832684,
0.0279323421,
0.0695066601,
0.0192058329,
-0.0054189521,
-0.0043013832,
0.0384331942,
-0.0152826291,
-0.0046817246,
0.0489923544,
-0.0010126923,
-0.0205037408,
-0.0196809415,
-0.0243086442,
-0.0112118954,
-0.0136943711,
-0.0174670164,
0.0590249002,
0.0224293899,
-0.0065249354,
-0.0389444306,
0.0165092461,
-0.009539213,
0.0123483362,
0.0065705781,
-0.0657935068,
0.0674216673,
-0.027511565,
0.0130798873,
-0.0182695631,
-0.0310403146,
0.0365258045,
-0.0149343107,
0.0046691429,
0.0396047868,
0.0227140244,
0.0628515035,
-0.0667019486,
0.0055602938,
-0.0266887844,
0.0521128923,
0.0367909223,
-0.0213187505,
0.0245731231,
0.0078773098,
-0.0084235137,
0.0376783796,
0.0068310336,
-0.0088755609,
-0.0317267291,
-0.0667263567,
0.0077483789,
0.0183886867,
-0.0234794598,
0.0262062121,
-0.0243014079,
0.0314915292,
0.0241471026,
-0.0619892329,
-0.0204111077,
0.0051183249,
2.875e-7,
0.0316069126,
0.032032311,
0.0225623362,
0.0570163019,
0.0040060892,
-0.0026881995,
-0.0025906032,
0.0509673096,
-0.0126178628,
0.0323784798,
0.0138449548,
0.0451852456,
0.0526676029,
-0.0021623846,
-0.103769742,
0.028742915,
-0.0336386785,
-0.0011027583,
-0.0880646408,
-0.0190664101,
0.0724149048,
0.0076197148,
0.0080473004,
-0.0228850879,
-0.0034270266,
-0.0632917732,
0.0742877945,
-0.0168012157,
-0.012431168,
0.0129376035,
-0.0162884519,
0.0189759452,
-0.0178850517,
-0.0253604949,
-0.0366232507,
0.0223008972,
0.0057901633,
0.0627784654,
-0.0385668576,
0.0075980499,
0.050099723,
0.0030638066,
-0.0138696739,
-0.0172917023,
0.0428067967,
0.0175160635,
-0.0089682201,
-0.0082834763,
-0.0067035318,
-0.0043561109,
-0.0335916728,
-0.0442706645,
-0.012889564,
0.0072172503,
0.0073074019,
-0.0239574816,
0.0246843435,
-0.0333279669,
-0.0300207529,
0.0015089504,
-0.0217434652,
-0.0215313882,
-0.0818893239,
-0.0497591868,
0.0118714143,
-0.0734657124,
0.0027744712,
3.186166647e-34,
0.0456660353,
-0.014119355,
0.0144560169,
0.0285472367,
0.0523406677,
0.0302043818,
-0.0001862529,
0.0156882517,
0.0083995853,
-0.0783054456,
-0.0380144529
] |
http://arxiv.org/abs/2401.17263v3 | http://arxiv.org/pdf/2401.17263v3 | Robust Prompt Optimization for Defending Language Models Against Jailbreaking Attacks | 2024-01-30T18:56:08 | 2024-06-05T23:39:54 | Andy Zhou, Bo Li, Haohan Wang | cs.LG | cs.LG, cs.AI, cs.CL, cs.CV | Despite advances in AI alignment, large language models (LLMs) remain
vulnerable to adversarial attacks or jailbreaking, in which adversaries can
modify prompts to induce unwanted behavior. While some defenses have been
proposed, they have not been adapted to newly proposed attacks and more
challenging threat models. To address this, we propose an optimization-based
objective for defending LLMs against jailbreaking attacks and an algorithm,
Robust Prompt Optimization (RPO) to create robust system-level defenses. Our
approach directly incorporates the adversary into the defensive objective and
optimizes a lightweight and transferable suffix, enabling RPO to adapt to
worst-case adaptive attacks. Our theoretical and experimental results show
improved robustness to both jailbreaks seen during optimization and unknown
jailbreaks, reducing the attack success rate (ASR) on GPT-4 to 6% and Llama-2
to 0% on JailbreakBench, setting the state-of-the-art. Code can be found at
https://github.com/lapisrocks/rpo | Robust Prompt Optimization for Defending Language Models Against Jailbreaking Attacks: Despite advances in AI alignment, large language models (LLMs) remain
vulnerable to adversarial attacks or jailbreaking, in which adversaries can
modify prompts to induce unwanted behavior. While some defenses have been
proposed, they have not been adapted to newly proposed attacks and more
challenging threat models. To address this, we propose an optimization-based
objective for defending LLMs against jailbreaking attacks and an algorithm,
Robust Prompt Optimization (RPO) to create robust system-level defenses. Our
approach directly incorporates the adversary into the defensive objective and
optimizes a lightweight and transferable suffix, enabling RPO to adapt to
worst-case adaptive attacks. Our theoretical and experimental results show
improved robustness to both jailbreaks seen during optimization and unknown
jailbreaks, reducing the attack success rate (ASR) on GPT-4 to 6% and Llama-2
to 0% on JailbreakBench, setting the state-of-the-art. Code can be found at
https://github.com/lapisrocks/rpo | [
0.0533210151,
0.0583376288,
-0.0232525058,
0.0393663421,
-0.0270352997,
-0.0172228534,
-0.0003160319,
0.004948548,
0.0306442492,
-0.0070166807,
-0.0021336742,
-0.0281733517,
0.0253624916,
-0.0067055579,
0.038208656,
-0.0031424118,
0.0632275492,
-0.0366282351,
-0.0074176053,
-0.0124283452,
0.0131782396,
-0.0206393003,
0.0171245821,
0.0292252172,
-0.0171494968,
-0.0072791893,
-0.0239347201,
0.0214620065,
0.0226026699,
-0.0385220721,
0.0071603893,
0.0479535051,
-0.0358208045,
0.0701887161,
0.0000022345,
-0.0378978699,
-0.060158357,
0.0233584512,
-0.0383144319,
0.0421314612,
0.0195491277,
-0.0556986146,
-0.0446851216,
0.035304416,
-0.033349108,
0.0421423987,
0.0373349451,
0.0178560633,
0.0685334206,
0.0850375295,
0.0073687108,
0.0118375672,
0.0649984181,
-0.0270632394,
-0.0153547898,
-0.077717334,
0.0216822866,
-0.0693224072,
0.0090795094,
0.0229350403,
0.001681417,
-0.0098819975,
0.0032925846,
-0.0154720815,
-0.0980204269,
-0.0156329889,
-0.0208845716,
-0.0230260752,
-0.017080551,
0.0468779318,
-0.038013678,
-0.0242346693,
0.0080470741,
-0.0086560156,
0.0464632697,
0.0084241722,
0.0053989491,
0.0107331425,
0.0019282351,
-0.0201411378,
0.0220330395,
-0.0529384837,
0.0512711741,
-0.0136646507,
0.0059511974,
0.0665969104,
0.0457682088,
-0.0276413001,
-0.0034924853,
-0.0536491796,
-0.0305987503,
-0.0685389936,
0.0007240684,
0.0039161541,
-0.026772216,
0.0295165759,
-0.0962069705,
-0.0400292352,
0.0224718172,
-0.0190449879,
-0.0029464432,
0.041442737,
-0.0018354341,
-0.0190127622,
-0.0077350885,
0.0454223938,
-0.0312989093,
0.0248064678,
0.0081655281,
0.0240537617,
-0.019371314,
-0.0182562731,
-0.0129608447,
0.0321899876,
-0.0293470733,
0.0151932249,
-0.0196988285,
0.0485724956,
0.001556202,
0.0271049477,
-0.0555466264,
-0.008320502,
-0.0004803885,
-0.0003321778,
0.0497003719,
-0.0321133919,
-0.0203882381,
-0.0094611738,
0.007122566,
-0.0804544315,
-0.0231578685,
-0.0002221406,
-0.0816149041,
-0.0310783647,
-0.0159384795,
0.1026716232,
0.0167432353,
-0.0086367903,
-0.0632924065,
-0.0666490495,
-0.0474095866,
-0.0264985561,
-0.0107798278,
-0.0669276491,
-0.017983187,
-0.0332881808,
-0.0100428313,
-0.0898095369,
0.0029445207,
0.0104534291,
-0.0374078304,
0.043008551,
-0.0063707386,
-0.0118583143,
0.0069544245,
0.0248282943,
0.0174143612,
0.0664900094,
0.0210342705,
0.0489393957,
0.0711947158,
-0.0045504537,
0.0174472872,
0.0558896177,
-0.0018425919,
0.0264123641,
-0.0032716021,
-0.0108629605,
0.0739370137,
0.0002608436,
-0.0310378764,
0.032661248,
0.0292018913,
-0.0571777783,
0.0798304081,
0.0519787371,
-0.0676061586,
0.039824944,
-0.0250505861,
-0.0417046882,
-0.0046455273,
-0.0081173265,
-0.0824030191,
-0.0217996705,
-0.1174607202,
-0.0165497493,
-0.0455134884,
0.052487921,
0.0139988344,
0.0100539709,
-0.0080616679,
-0.0092095928,
-0.0570600666,
0.0246520936,
-0.0282293744,
-0.007335179,
-0.0308413692,
0.0361010283,
0.0069353692,
-0.0090679461,
-0.0463183038,
-0.0870424137,
-0.0053069256,
0.0168958269,
0.0304075405,
-0.0189537387,
-0.0221798811,
0.0042350725,
-0.0025580782,
0.0143073834,
0.0097550247,
0.0084714592,
0.037915919,
-0.0003650179,
0.0004316552,
0.0057555982,
0.0299091563,
0.0627617314,
0.007974606,
0.0062738806,
0.0033322987,
-0.0001320437,
-0.0327125676,
0.0440500677,
0.0198171549,
-0.0369884558,
0.0344666801,
0.0255561937,
0.022846194,
0.0300172232,
-0.0226510186,
0.0189001784,
0.0441886559,
-0.0024393806,
0.0588327684,
-0.0084545361,
-0.0344838575,
-0.0326575711,
0.0508397222,
0.0190363601,
-0.0096652675,
0.0877421498,
-0.0465414971,
-0.0361181125,
0.0501304008,
-0.0291577019,
0.0306548513,
0.0111583853,
0.0426226929,
-0.0132617475,
0.0386376232,
-0.0476226248,
0.0111619886,
-0.0091239428,
-0.014787619,
-0.0039014393,
0.0434637778,
0.0370100699,
-0.0287235398,
-0.0633972436,
0.0388486534,
-0.038900312,
0.0003736816,
0.0238333698,
0.0157844443,
0.0012386261,
-0.043746978,
-0.0266499184,
0.0166894961,
-0.0270163883,
0.012017359,
0.0421943031,
0.0394830033,
-0.0511718728,
-0.022304615,
0.0060535292,
-0.021920044,
-0.0006911004,
0.0069269468,
-0.0062737339,
-0.0140754208,
-0.0596581027,
-0.0404077731,
-0.006692803,
-0.0218795016,
0.0296441019,
-0.0410317779,
-0.0227951352,
-0.0214859638,
-0.0264780838,
0.0008698145,
0.0449975729,
-0.0046730149,
-0.0297006108,
0.0308579262,
0.0174154583,
0.0321549363,
0.0850037411,
0.0171535201,
0.0188993104,
0.061645411,
0.0266889706,
-0.0272992197,
-0.0246364679,
0.0222323034,
-0.0618397743,
-0.031368915,
-0.0490083955,
0.0419835001,
0.0683615282,
0.0317025594,
-0.0104281772,
-0.0388411917,
-0.0006290032,
0.0273946952,
-0.0210088529,
-0.0039295112,
-0.0041951765,
-0.0734937415,
0.0217305198,
-0.0328546315,
-0.0555122942,
0.0805254057,
-0.0662932768,
0.0290461034,
-0.025082726,
-0.024037268,
0.018397592,
0.0225580037,
-0.0265314933,
0.0199789181,
0.0494421124,
-0.0374456123,
-0.0375660583,
-0.0231580492,
0.0323211215,
-0.0045526493,
0.0223624445,
-0.0407463051,
-0.0046355021,
-0.0041375523,
-0.0321759023,
-0.0328528732,
-0.051052887,
-0.014428746,
-0.0141571211,
0.0300448984,
-0.0281191189,
0.0062213629,
-0.0333365761,
0.0132422149,
0.0206036512,
-0.0345467366,
0.0203835014,
-0.0363059118,
0.0300566014,
0.0549607761,
-0.0493422784,
-0.0037075358,
-0.0127174538,
0.0603369586,
0.032423418,
0.0497046076,
0.0007493426,
-0.0105640972,
0.0235945955,
0.0346912742,
-0.0170469377,
-0.0144700455,
0.052384112,
0.0422285907,
-0.0121272244,
-0.0110476241,
0.0498013534,
0.0366784409,
-0.0039406158,
0.0351692289,
-0.021641098,
0.07483989,
0.0149083184,
0.0379232429,
-0.0374874026,
-0.0325173326,
0.0367824771,
0.0756858587,
0.0018602745,
-0.01442861,
-0.0468365885,
-0.1101436168,
-0.0205259621,
0.0203384794,
-0.028294947,
-0.0215702169,
-0.0410406962,
-0.0104690576,
0.033212781,
0.0514579229,
0.0143268239,
0.0114058368,
0.0010762493,
0.0136854155,
-0.0466629528,
0.0039103888,
-0.0049147052,
-0.0326709412,
0.0796145052,
0.1088010296,
-0.0083926581,
-0.009210364,
0.0316835493,
0.0566999726,
0.0101709422,
-0.0079024229,
0.0125398636,
-0.0394712165,
0.0104139028,
0.0034866601,
0.025075905,
0.0412734523,
0.0031211427,
-0.0225815605,
0.0611513555,
0.0384968333,
0.0032373401,
0.0137368711,
-0.0194009431,
-0.0360567868,
0.0157409664,
0.0346108228,
0.02813337,
-0.0317811407,
-0.0470254533,
-0.0787524208,
-0.0211843587,
-0.0635167882,
-0.0022733728,
0.0146537079,
0.036763791,
-0.0204695314,
-0.0518747456,
0.0140524376,
-0.0418320335,
0.0514044277,
0.0410964489,
-0.0147802308,
-0.0530671738,
0.0099681783,
-0.0193943009,
-0.0280845612,
-0.0134321377,
0.0371046402,
-0.0907368362,
0.0459714495,
-0.0047923136,
-0.0206646584,
-0.0134944813,
0.012206764,
-0.0478913821,
-0.110881187,
0.0362825282,
0.059996888,
0.0189134255,
0.0246058479,
-0.0484013744,
-0.0665951893,
0.0405231416,
0.0646693558,
-0.0220552925,
0.0210470576,
0.0505578071,
-0.0241738502,
0.0086493762,
0.0269711073,
-0.0420273356,
-0.0221579149,
0.0720566586,
0.0097235786,
-0.0042972886,
-0.0262518153,
0.0204380658,
0.0713810474,
0.0123692127,
-0.0443906449,
-0.0247398876,
0.0383115858,
0.0269834604,
-0.005976669,
-0.0132109337,
0.0232970268,
0.0319037773,
0.0136223352,
-0.0168702174,
0.0463175178,
-0.002680664,
-0.0077654603,
0.010962653,
-0.0015582334,
0.0377169326,
0.000479317,
0.0778628141,
0.0499123707,
0.0327116884,
-0.0119904755,
0.0838190913,
0.0017832047,
0.0385188833,
-0.0301530268,
0.045482602,
-0.0048850132,
-0.0148845669,
-0.0390585177,
0.0707608312,
-0.0792731866,
-0.0463419445,
-0.0018951809,
0.0902210847,
0.0131749427,
0.0108286468,
-0.019185381,
-0.0047456552,
-0.0023972066,
-0.0306848418,
-0.0383888744,
0.0021249782,
-0.0509922132,
0.0305840168,
-0.0006568577,
0.0022780027,
0.0380895734,
0.0633983612,
0.0216940008,
0.0575656816,
0.0176818781,
-0.003564249,
-0.0423721746,
0.0010165742,
0.0504023656,
0.0015133055,
-0.0738308281,
-0.0441690236,
-0.0430109501,
-0.076225169,
-0.0063939164,
0.0199906416,
0.045686122,
0.0280132405,
-0.0283964928,
0.0322036035,
-0.0134088537,
0.0408143774,
-0.0514680706,
0.0014449826,
0.0172598306,
-0.0601173006,
0.0318591632,
-6.048707118e-33,
0.0027425413,
0.0022614715,
-0.0604927801,
-0.047268033,
0.0040596607,
-0.0172016621,
-0.0309402216,
0.0330763087,
-0.0395208858,
-0.0551823415,
-0.0348026194,
-0.0140781347,
0.0031910343,
0.0020586508,
-0.0545730218,
-0.0159091707,
-0.007607135,
0.0096692294,
-0.0592860393,
0.0149940699,
0.0390514806,
0.0172411632,
-0.0027338089,
-0.0126767997,
0.050137829,
-0.019224925,
-0.0413941145,
0.0417341404,
-0.0406883173,
0.0461796075,
0.0093425438,
-0.0043174815,
-0.011800115,
-0.0262045674,
0.0053956904,
0.0644439533,
0.004197571,
-0.0178901423,
0.0243608523,
-0.0035592718,
0.0263717845,
-0.0711361319,
0.0183158163,
-0.0000870117,
-0.0476036258,
0.0151080098,
-0.0095511274,
-0.0102809081,
-0.023730319,
0.0140333381,
-0.0298449788,
-0.0092667351,
-0.0355094597,
-0.0124052223,
-0.0104322433,
-0.0253016297,
0.0092869727,
0.0141936727,
-0.0116809318,
0.0424106047,
0.0265603028,
0.0424760841,
0.0594694205,
0.0276325345,
-0.0219760817,
0.0327799655,
0.0027831218,
-0.0385519303,
0.0213073026,
-0.0111134145,
0.0097421762,
0.0159921274,
0.0599217154,
0.0821324736,
0.0500691608,
-0.0611171462,
-0.0617957972,
0.0164421871,
0.0511299372,
-0.0015847727,
0.0161170848,
-0.0044730022,
-0.0346540213,
0.0073399735,
-0.0240656026,
-0.0273598693,
-0.0215312168,
-0.0331861228,
0.0334358141,
-0.0139722498,
-0.0276759509,
-0.0306678787,
0.0130198747,
-0.0469788313,
0.008263628,
0.0151661439,
-0.021903405,
-0.0138918208,
-0.0044714594,
0.0363899656,
0.0197130013,
-0.0059864474,
0.0149246249,
-0.0580918267,
0.0435296521,
-0.0242465474,
0.0164522268,
0.0288919527,
-0.0641559064,
0.0020820925,
-0.0016442939,
0.0459304452,
0.015958963,
0.0083556958,
0.0226117019,
-0.0448530763,
0.0159730911,
0.007356531,
0.035720259,
-0.0134866834,
-0.0469885282,
0.0295836069,
-0.015058239,
0.0161549971,
-0.0065999548,
0.0174902361,
0.0334771313,
0.0447460748,
0.0085486006,
-0.0358615741,
-0.017218139,
-0.0412240401,
2.985e-7,
-0.0220045466,
0.0113260075,
0.0410964265,
-0.0251122788,
0.0076801484,
0.0595250279,
-0.0332251936,
0.0811773464,
-0.0240264758,
-0.0448969267,
-0.0239958614,
-0.0493578464,
0.0397382788,
0.0024624339,
-0.0602172911,
0.0471250303,
-0.0661322176,
0.0168076996,
-0.0764422938,
0.0104905982,
0.017529143,
0.0949919969,
-0.0007076535,
-0.0301673859,
-0.0264007412,
-0.0302356314,
-0.0139109641,
-0.0042063431,
0.0189680457,
-0.0122417826,
0.0233833119,
-0.0137580102,
0.0056634839,
0.0037429377,
-0.0371877365,
0.0172514599,
-0.047860764,
0.0425677784,
-0.0098282881,
-0.0284114126,
-0.0265626274,
-0.0141059533,
-0.0213725977,
-0.0530238263,
0.0601548366,
0.037886288,
-0.0257740356,
-0.035994757,
0.0212473962,
-0.0353574865,
-0.0112234531,
-0.0471014231,
0.0088368887,
0.0064526284,
0.0131242983,
0.0321330838,
0.0224258136,
-0.0590324476,
-0.0177151281,
0.0099047432,
-0.0081034256,
-0.0181203298,
-0.0141293025,
0.0492613241,
0.0512597449,
-0.0442479812,
0.022244744,
3.083919761e-34,
-0.0127491634,
-0.0457341932,
0.02651277,
0.0207268018,
0.0315529034,
0.0095073823,
0.0516580157,
0.0316193737,
0.0182298739,
-0.0244011749,
-0.0459439382
] |
http://arxiv.org/abs/2401.17256v2 | http://arxiv.org/pdf/2401.17256v2 | Weak-to-Strong Jailbreaking on Large Language Models | 2024-01-30T18:48:37 | 2024-02-05T18:19:46 | Xuandong Zhao, Xianjun Yang, Tianyu Pang, Chao Du, Lei Li, Yu-Xiang Wang, William Yang Wang | cs.CL | cs.CL | Large language models (LLMs) are vulnerable to jailbreak attacks - resulting
in harmful, unethical, or biased text generations. However, existing
jailbreaking methods are computationally costly. In this paper, we propose the
weak-to-strong jailbreaking attack, an efficient method to attack aligned LLMs
to produce harmful text. Our key intuition is based on the observation that
jailbroken and aligned models only differ in their initial decoding
distributions. The weak-to-strong attack's key technical insight is using two
smaller models (a safe and an unsafe one) to adversarially modify a
significantly larger safe model's decoding probabilities. We evaluate the
weak-to-strong attack on 5 diverse LLMs from 3 organizations. The results show
our method can increase the misalignment rate to over 99% on two datasets with
just one forward pass per example. Our study exposes an urgent safety issue
that needs to be addressed when aligning LLMs. As an initial attempt, we
propose a defense strategy to protect against such attacks, but creating more
advanced defenses remains challenging. The code for replicating the method is
available at https://github.com/XuandongZhao/weak-to-strong | Weak-to-Strong Jailbreaking on Large Language Models: Large language models (LLMs) are vulnerable to jailbreak attacks - resulting
in harmful, unethical, or biased text generations. However, existing
jailbreaking methods are computationally costly. In this paper, we propose the
weak-to-strong jailbreaking attack, an efficient method to attack aligned LLMs
to produce harmful text. Our key intuition is based on the observation that
jailbroken and aligned models only differ in their initial decoding
distributions. The weak-to-strong attack's key technical insight is using two
smaller models (a safe and an unsafe one) to adversarially modify a
significantly larger safe model's decoding probabilities. We evaluate the
weak-to-strong attack on 5 diverse LLMs from 3 organizations. The results show
our method can increase the misalignment rate to over 99% on two datasets with
just one forward pass per example. Our study exposes an urgent safety issue
that needs to be addressed when aligning LLMs. As an initial attempt, we
propose a defense strategy to protect against such attacks, but creating more
advanced defenses remains challenging. The code for replicating the method is
available at https://github.com/XuandongZhao/weak-to-strong | [
0.0176833924,
0.1043151245,
-0.0109675247,
0.0485561751,
-0.0368090086,
0.0119116139,
0.0173278395,
-0.0075182356,
0.0031536277,
-0.0099950498,
-0.0059940573,
-0.0605730116,
0.0571586564,
0.013995043,
0.057184007,
0.0092672054,
0.0210059416,
-0.0173074659,
-0.0018535019,
-0.0121201631,
0.0348479412,
0.0098363729,
0.0433575213,
0.0181593392,
0.0214675758,
-0.0247774199,
-0.0229565371,
0.0220969729,
0.0319988243,
-0.0173556749,
0.0056645796,
0.050698109,
-0.0348576084,
0.1064187065,
0.0000021516,
-0.0063504018,
-0.060341727,
0.0280124974,
0.0097130174,
0.0664936826,
0.0260299388,
-0.0823650807,
-0.0549125671,
0.0200566985,
-0.0367284343,
0.0131071648,
0.0315395966,
0.0219734926,
0.0717371702,
0.0986421704,
0.0053526214,
-0.023245994,
0.0414301641,
-0.041365426,
-0.0206839759,
-0.0691751391,
0.0349405818,
-0.0164698958,
-0.0630114079,
0.0202436689,
0.0027395312,
0.011398959,
0.0015261311,
-0.0114110084,
-0.0401273519,
-0.007765281,
-0.0255131591,
-0.0240433346,
-0.0009805523,
0.046006795,
-0.0467351712,
-0.0382578187,
0.0190818496,
-0.0132943261,
0.0325311646,
0.0001085175,
-0.0048203515,
0.0442824848,
-0.0024938136,
0.0172585044,
0.0270950105,
-0.0321799144,
0.0625956804,
-0.0075470894,
-0.0055940081,
0.0168146547,
0.0311727636,
-0.044265788,
-0.0288311392,
-0.0159626398,
-0.0102193067,
-0.046916496,
-0.0249578971,
-0.0237898137,
0.0141710052,
0.0304716732,
-0.1084879413,
-0.0445115976,
0.0193629358,
-0.0746709108,
0.0153272245,
0.0141718723,
0.0119976001,
0.0068019447,
-0.0159348752,
0.0046067345,
-0.0385665447,
0.0131304041,
-0.0173951481,
0.0297417976,
0.012898962,
-0.0103226537,
0.0318310596,
-0.0052623749,
-0.021987604,
0.0109156249,
-0.0395314805,
0.0202215537,
-0.0032119106,
0.0322535224,
-0.0313248262,
0.0194956828,
-0.0179833639,
0.0055668596,
0.0609359667,
-0.0012858219,
0.0163537059,
-0.0096377917,
-0.002064174,
-0.065132454,
-0.0268696547,
0.0309292227,
-0.0540388972,
-0.0111563914,
-0.0374710523,
0.0784716532,
0.0094106728,
0.0102309138,
-0.0590676591,
-0.0596058965,
-0.0589916147,
-0.0364287011,
-0.0056361943,
-0.0731677189,
-0.0144950803,
-0.0074780146,
-0.0235359836,
-0.0979554355,
0.0266655125,
-0.0096225115,
-0.0705489367,
0.0423502661,
-0.0062667364,
-0.0039265491,
0.006282426,
0.0060193655,
0.0136933457,
0.0958607122,
0.0273293331,
0.0506799631,
0.0602593459,
-0.0162280109,
0.0002019335,
0.0768900216,
0.0161029529,
0.0555683672,
-0.02263253,
-0.0177965462,
0.0374182649,
0.0042377776,
-0.0042982488,
0.037653666,
-0.0005053402,
-0.0260413196,
0.084267877,
0.0169031918,
-0.0295371972,
0.0471935757,
-0.0339070708,
-0.0189045258,
-0.0111924177,
0.0120715685,
-0.0736545324,
-0.0239394158,
-0.0859660506,
-0.0145783238,
0.005671001,
0.0456140153,
0.005888531,
0.0163984839,
-0.0067474926,
-0.0050173677,
0.0034297919,
-0.0120768808,
0.0098379962,
0.0128988912,
-0.0294746496,
0.06673076,
0.0228661764,
-0.0359578878,
-0.0200164244,
-0.0358504988,
-0.0179413501,
0.0365686417,
0.0470493138,
0.0150384391,
0.0138402535,
-0.0137582999,
-0.0007103153,
0.0196961425,
0.0181215014,
0.0118845273,
0.0010672346,
-0.0199548621,
-0.0204777773,
-0.0071762069,
0.0223762821,
0.0293361992,
0.0255918261,
-0.0051040333,
0.0076062786,
-0.0269861687,
-0.0526653342,
0.0099737663,
-0.0273440257,
-0.009206811,
0.0036180892,
0.0331142917,
0.0117624383,
0.0093995165,
-0.0486634523,
0.0138910627,
0.038203463,
0.0053267688,
0.0237115752,
0.0019350504,
-0.0251364801,
-0.0219129007,
0.0411312394,
-0.0391981788,
-0.0217562467,
0.0568628162,
-0.0452411249,
-0.0097756013,
0.0600359365,
-0.0309879389,
0.0196937136,
0.0080723716,
0.0487734862,
-0.0362975933,
0.0312075689,
-0.0359139182,
0.0524274632,
-0.0383118168,
0.0149201425,
-0.0168545134,
-0.000877893,
0.0302720014,
-0.0169977173,
-0.0419342518,
0.0043880278,
-0.0234436095,
0.0096570691,
0.0163062736,
0.0171277002,
0.0084094675,
-0.0559012741,
-0.0176567975,
-0.0101751825,
-0.0251269173,
0.0318323262,
0.0657188445,
0.0207642317,
-0.0233322047,
-0.0035864969,
-0.0266982522,
-0.019506732,
0.0025198965,
0.0097397771,
-0.0015687909,
-0.0206066947,
-0.0412540585,
-0.0179562494,
-0.013910138,
-0.004690744,
0.0730260089,
0.0048211999,
-0.032257542,
-0.0762943551,
0.0183552783,
0.013688107,
0.020452287,
-0.0066318223,
-0.052370172,
0.0216871686,
0.0163503196,
0.00568406,
0.077231437,
0.0151135726,
0.0606064238,
0.0312865227,
0.0242068861,
-0.0166104119,
-0.0502840132,
0.0246127676,
-0.0255223103,
-0.0539672486,
-0.0842210427,
0.0128217377,
0.0611208007,
0.0498148538,
-0.0019870808,
-0.038077075,
0.0318381749,
-0.002741843,
-0.0315244608,
0.020829374,
-0.0092923976,
-0.0436940938,
-0.0014197628,
-0.0322162323,
-0.079390496,
0.0549638867,
-0.0660831407,
0.054229524,
-0.0145980045,
-0.035727717,
0.0030776057,
0.0290235728,
-0.0261014812,
0.0408097953,
0.0124684814,
-0.027528638,
-0.0638184398,
0.0061938148,
0.051791165,
-0.013772306,
0.0463168919,
-0.0352206193,
0.003744731,
-0.0256861802,
-0.0357368141,
-0.0309440512,
-0.084005706,
-0.0070280195,
-0.01851275,
0.0627750233,
-0.0399819314,
-0.0322982259,
-0.0215018503,
0.0183692034,
-0.032107003,
-0.0274769142,
0.0376788601,
-0.0267860424,
0.0348865725,
0.0504872017,
-0.0655232966,
0.014117308,
-0.0517805889,
0.0263090692,
0.0660020187,
0.04355295,
0.0563197285,
-0.0184283983,
-0.0179272592,
0.0211105011,
0.0257295631,
-0.0048402599,
0.0059012175,
0.0585915521,
-0.0102800634,
-0.0014940551,
0.0510971807,
0.0031012851,
0.0060391463,
0.036535047,
0.0223572999,
0.0564516857,
0.0130077796,
0.0126956766,
-0.0703886375,
-0.0244747866,
0.0136573147,
0.0827431902,
0.0040210672,
-0.0331444591,
-0.0155679137,
-0.1125586629,
-0.0169504378,
0.0558703654,
-0.0129471049,
-0.0346348248,
-0.0210399888,
-0.0249360781,
-0.0012308835,
0.0267605018,
0.0151278283,
0.0218866151,
-0.0160694495,
0.01461278,
-0.013387816,
-0.0503143519,
-0.0297856536,
-0.0290088858,
0.0690470859,
0.0891907662,
0.0002930897,
-0.0368581638,
0.0106860222,
0.0304178577,
0.0073703299,
-0.0227779951,
0.032998804,
0.0070493254,
0.0222461466,
-0.0234415736,
0.0126954289,
0.0431536101,
-0.0035981482,
-0.0125428401,
0.0193482265,
0.0299997367,
0.0075857304,
-0.0106321443,
-0.0033024903,
-0.0121107595,
0.0239625815,
0.0542437211,
0.014109035,
-0.0331460051,
-0.0327083431,
-0.0976575762,
-0.0113576911,
-0.0394634232,
-0.0130477082,
0.0478561446,
0.0154818697,
-0.0222797804,
-0.0642338991,
0.0070320186,
-0.025185965,
0.0357335173,
0.0387983546,
0.0045521008,
-0.0176717378,
-0.0220005065,
-0.0252575818,
-0.0184726473,
-0.0184662137,
0.044696413,
-0.0402231105,
0.0014567872,
-0.0162573569,
-0.0114918919,
0.0096021341,
0.0089742504,
-0.110735029,
-0.116099745,
0.0274462961,
0.0347723737,
0.0178015791,
0.0452086553,
-0.0362149402,
-0.0635373145,
0.062011037,
0.0642509609,
-0.042755954,
0.0354058966,
0.0279351771,
-0.045832552,
0.0090807611,
0.0303514041,
-0.060697671,
-0.0494511463,
0.0887841433,
0.0035551686,
-0.0390702076,
-0.0020305305,
0.0032906404,
0.0375294387,
0.0053831399,
-0.0446343869,
-0.0172072742,
0.0141536612,
0.0278997235,
-0.0168224946,
-0.0409439541,
0.0026544521,
0.0100617697,
0.0413668379,
-0.0158141349,
0.0417529307,
0.0415338464,
0.0118747847,
0.0173735451,
-0.018396087,
0.0275512803,
0.002933258,
0.0543287881,
0.0761713684,
0.0251230597,
-0.0202161092,
0.0709254891,
-0.0315474942,
0.02845552,
-0.0060500018,
0.0160327014,
0.0087555759,
-0.0177189074,
-0.002565281,
0.0410210192,
-0.0670579597,
-0.0284226984,
-0.0368150286,
0.0331181772,
0.0076633408,
0.0005359151,
0.0370487943,
-0.0025390617,
0.0059927716,
-0.0312679,
-0.0102290288,
-0.0133166686,
-0.0519396812,
0.0061759311,
-0.012803819,
-0.018848734,
0.0428405553,
0.0796780288,
0.0449871309,
0.039926324,
-0.0000722203,
-0.0340820253,
-0.0274715573,
-0.0124159632,
0.0714615062,
-0.0026855164,
-0.0586874336,
-0.0485479608,
-0.0348571315,
-0.0759408623,
-0.0074008619,
0.0060636881,
0.0305963438,
0.0209985096,
-0.041741997,
0.050312072,
-0.0199724529,
0.060358569,
-0.0465221591,
0.0326247029,
0.0239619259,
-0.0687278211,
0.0193249993,
-6.424646007e-33,
0.0333692655,
-0.0182547607,
-0.0716895759,
-0.0722224414,
0.0201747362,
-0.0040645096,
-0.0231508743,
0.0130163431,
-0.0179926027,
-0.0362938605,
-0.0107407179,
0.010324466,
0.0020048337,
0.0311867669,
-0.0325566083,
-0.0340196528,
0.0140150543,
0.0037272349,
-0.026621338,
0.0239609089,
0.0456334576,
0.0216839127,
0.0407300591,
-0.0448795222,
0.0451136343,
-0.0416503586,
-0.0564123206,
0.0093439873,
-0.024304647,
0.0446026474,
0.022291908,
0.0028581484,
-0.0090045715,
-0.0129307788,
-0.0058352766,
0.0386584774,
0.0113430871,
0.0313479863,
0.0203394536,
0.0337315984,
-0.0041921008,
-0.0457048081,
0.041018188,
0.0001298682,
-0.0022034617,
-0.0130820489,
0.0021240404,
-0.0145533932,
-0.0182213541,
0.0070586805,
-0.0268190615,
0.0225214139,
-0.037365362,
0.0129502406,
-0.0084764455,
-0.0276666153,
0.027647892,
0.0046305023,
-0.0422071442,
0.021063922,
0.0545338206,
0.0343773328,
0.0670343637,
0.0191515405,
-0.0051559675,
0.0707857385,
-0.0085515529,
-0.0101887314,
0.0006823177,
-0.0012253646,
0.0056665144,
0.0359051675,
0.030519031,
0.1013473347,
0.0588343367,
-0.0390552208,
-0.0757082105,
0.0137652969,
0.0022384676,
-0.0108915586,
0.0290801153,
0.0035370942,
-0.0420534685,
-0.0053595956,
-0.0243117474,
-0.0627022907,
-0.0245567672,
-0.0372328795,
0.0158257224,
-0.0039444598,
-0.0565471277,
-0.0161867719,
0.0167402811,
-0.0814694464,
-0.0253464691,
0.0459308028,
-0.0388686769,
-0.0150850518,
-0.0203866232,
0.0431369357,
-0.0024611382,
-0.0023210207,
0.0216970872,
-0.0260308888,
0.0334467851,
0.006797072,
0.064304851,
0.0102842608,
-0.0690885186,
0.0015199309,
-0.0176391583,
0.0308621805,
0.0182043798,
0.0082674744,
0.0341080949,
-0.0327589959,
-0.0019443536,
-0.0329854377,
0.0579927489,
0.0424877815,
-0.0176245999,
0.0008242925,
-0.032327313,
0.0228715874,
-0.0124038812,
0.002099219,
0.0214020796,
0.0148261627,
0.0491219163,
-0.0636758804,
-0.0045509934,
-0.0096712066,
2.88e-7,
0.0134363202,
0.0256884415,
0.0294983275,
-0.0209489558,
0.022125259,
0.0633893907,
-0.0368082523,
0.0805903152,
-0.0398104936,
-0.0535246432,
-0.0452364422,
-0.0255161319,
0.0834044442,
0.0396088101,
-0.061103411,
0.0549533814,
-0.0905271843,
0.0274093822,
-0.0644899011,
-0.0192640666,
0.0187904034,
0.1191557571,
0.0090603391,
-0.0264943615,
-0.0076402393,
-0.0437820368,
0.0181597788,
0.0144520765,
-0.0196450818,
-0.0068405992,
-0.0054051103,
-0.003957517,
-0.0046395301,
-0.0059617818,
-0.0425322838,
-0.0015254208,
-0.0185712744,
0.0229108799,
-0.0002278022,
0.0153691899,
-0.0288016926,
-0.0410619564,
-0.0052146418,
-0.0243977141,
0.0513592027,
0.013930439,
-0.0331071243,
-0.0610297397,
-0.0035326041,
-0.0204599872,
0.0053662551,
-0.0263754781,
0.0136445556,
0.0165131055,
0.0124637363,
0.0264842343,
0.0077254465,
-0.0456128158,
-0.0366905257,
0.0137254484,
-0.0146768233,
-0.0375465788,
-0.0077883988,
0.0473190285,
0.0542325862,
-0.0501079969,
0.0222482961,
2.902639247e-34,
-0.0075198398,
-0.0355536081,
-0.0033771198,
0.0102792401,
0.0273875985,
-0.0206427518,
0.0329552814,
0.0216915086,
-0.0061339689,
-0.0348920226,
-0.0300832186
] |
http://arxiv.org/abs/2401.17244v2 | http://arxiv.org/pdf/2401.17244v2 | LLaMP: Large Language Model Made Powerful for High-fidelity Materials Knowledge Retrieval and Distillation | 2024-01-30T18:37:45 | 2024-06-02T07:50:21 | Yuan Chiang, Elvis Hsieh, Chia-Hong Chou, Janosh Riebesell | cs.CL | cs.CL, cond-mat.mtrl-sci, cs.AI | Reducing hallucination of Large Language Models (LLMs) is imperative for use
in the sciences, where reliability and reproducibility are crucial. However,
LLMs inherently lack long-term memory, making it a nontrivial, ad hoc, and
inevitably biased task to fine-tune them on domain-specific literature and
data. Here we introduce LLaMP, a multimodal retrieval-augmented generation
(RAG) framework of hierarchical reasoning-and-acting (ReAct) agents that can
dynamically and recursively interact with computational and experimental data
on Materials Project (MP) and run atomistic simulations via high-throughput
workflow interface. Without fine-tuning, LLaMP demonstrates strong tool usage
ability to comprehend and integrate various modalities of materials science
concepts, fetch relevant data stores on the fly, process higher-order data
(such as crystal structure and elastic tensor), and streamline complex tasks in
computational materials and chemistry. We propose a simple metric combining
uncertainty and confidence estimates to evaluate the self-consistency of
responses by LLaMP and vanilla LLMs. Our benchmark shows that LLaMP effectively
mitigates the intrinsic bias in LLMs, counteracting the errors on bulk moduli,
electronic bandgaps, and formation energies that seem to derive from mixed data
sources. We also demonstrate LLaMP's capability to edit crystal structures and
run annealing molecular dynamics simulations using pre-trained machine-learning
force fields. The framework offers an intuitive and nearly hallucination-free
approach to exploring and scaling materials informatics, and establishes a
pathway for knowledge distillation and fine-tuning other language models. Code
and live demo are available at https://github.com/chiang-yuan/llamp | LLaMP: Large Language Model Made Powerful for High-fidelity Materials Knowledge Retrieval and Distillation: Reducing hallucination of Large Language Models (LLMs) is imperative for use
in the sciences, where reliability and reproducibility are crucial. However,
LLMs inherently lack long-term memory, making it a nontrivial, ad hoc, and
inevitably biased task to fine-tune them on domain-specific literature and
data. Here we introduce LLaMP, a multimodal retrieval-augmented generation
(RAG) framework of hierarchical reasoning-and-acting (ReAct) agents that can
dynamically and recursively interact with computational and experimental data
on Materials Project (MP) and run atomistic simulations via high-throughput
workflow interface. Without fine-tuning, LLaMP demonstrates strong tool usage
ability to comprehend and integrate various modalities of materials science
concepts, fetch relevant data stores on the fly, process higher-order data
(such as crystal structure and elastic tensor), and streamline complex tasks in
computational materials and chemistry. We propose a simple metric combining
uncertainty and confidence estimates to evaluate the self-consistency of
responses by LLaMP and vanilla LLMs. Our benchmark shows that LLaMP effectively
mitigates the intrinsic bias in LLMs, counteracting the errors on bulk moduli,
electronic bandgaps, and formation energies that seem to derive from mixed data
sources. We also demonstrate LLaMP's capability to edit crystal structures and
run annealing molecular dynamics simulations using pre-trained machine-learning
force fields. The framework offers an intuitive and nearly hallucination-free
approach to exploring and scaling materials informatics, and establishes a
pathway for knowledge distillation and fine-tuning other language models. Code
and live demo are available at https://github.com/chiang-yuan/llamp | [
-0.00173806,
0.0935096517,
-0.0088151172,
0.0332788751,
-0.0025989693,
-0.0224283226,
-0.0125535158,
0.0013348184,
0.0254366975,
0.0089484528,
-0.0498224013,
-0.0369913727,
-0.0520798005,
0.0477481447,
-0.0002011959,
-0.0057506971,
0.0377385989,
-0.0594979078,
0.0353021026,
-0.0366670899,
0.0076697203,
-0.0393605903,
-0.0184744485,
0.0152898962,
0.0244741123,
-0.0075827842,
-0.0376753621,
0.0339557566,
0.0259850565,
-0.0342479423,
0.0241515599,
0.0543858148,
0.0316492841,
0.0184390172,
0.0000025038,
-0.0657791495,
0.0118244644,
0.0350556485,
0.0171436351,
-0.0193388369,
0.040479891,
-0.0443162061,
-0.0381777436,
-0.0115522752,
-0.0135674588,
0.0063190418,
0.0619800016,
-0.006237946,
0.018304415,
-0.0213839263,
0.0159161687,
0.002117116,
0.0542366989,
0.013190032,
0.0975217745,
-0.0122705447,
0.0223775711,
0.0364493914,
0.0197077021,
0.046759814,
0.0213829167,
0.032197319,
-0.0138431359,
-0.0319145322,
0.0483346507,
0.004246668,
-0.0519091934,
-0.008683512,
-0.0206463337,
0.084682852,
0.055177737,
-0.0251988508,
-0.063407369,
0.0021148061,
-0.0601088032,
0.029348759,
-0.0629748851,
0.02053014,
0.0090979533,
0.0030086397,
-0.0034784216,
-0.0215569176,
0.0562022924,
0.0262296293,
-0.0087190494,
0.0157660898,
-0.0021127397,
-0.0141855031,
0.0097573427,
0.005741009,
-0.0144384131,
0.0205197725,
-0.0222477,
0.0237679332,
-0.0062943771,
0.0094179977,
0.0282456819,
-0.0200774036,
0.0547789708,
-0.0276368707,
0.0408594273,
0.0633300245,
-0.0235096607,
0.0263651349,
-0.0184391253,
-0.0091870567,
-0.0046442403,
0.0228087772,
-0.0314975902,
-0.023866117,
-0.0394461453,
-0.0378415361,
-0.0178931709,
0.0008165585,
-0.0498733483,
-0.0400728062,
-0.0142426789,
0.026302347,
-0.031810794,
0.0078658937,
-0.0396089666,
0.0206862129,
-0.0341364443,
0.0043861372,
-0.0521871895,
0.0322113521,
-0.0104261711,
0.0030033826,
0.0489119776,
-0.0401206911,
-0.0438351817,
-0.0233136397,
-0.0347318687,
0.006029367,
0.0301888622,
0.0875041559,
-0.0230279788,
-0.0174399894,
-0.021822501,
-0.0375901572,
-0.0092895236,
-0.0364220254,
0.0438030064,
-0.0527725853,
0.0868331119,
0.0466955714,
-0.0338202119,
-0.0789232031,
0.0229488984,
-0.0083731078,
-0.0167464912,
0.0775102228,
0.0098133348,
-0.0543492511,
0.0267787054,
0.0327180214,
0.0069826194,
0.02897566,
-0.0093459236,
0.0425597765,
-0.0126023069,
0.041216854,
-0.0106828529,
0.0005229855,
0.036883615,
-0.015601255,
-0.0562695451,
-0.0107649481,
-0.0066399635,
0.0306878574,
0.019548256,
0.0147402212,
-0.0243147239,
-0.0007014234,
0.0219198652,
0.0037905548,
0.0272094291,
0.0239425022,
-0.0215063076,
0.0105366763,
0.044329714,
-0.0599268749,
-0.0156683736,
0.0860527307,
-0.0073187393,
-0.0337959379,
0.0568755344,
0.0179175362,
0.0092367772,
-0.0403022431,
-0.0453017466,
-0.0051521184,
-0.0055266791,
-0.0304108616,
0.0529533625,
-0.0217713434,
-0.0579473004,
0.0087469397,
0.0363309681,
-0.0708269998,
-0.0463859886,
-0.0222376604,
0.0248891525,
0.0723367184,
0.0514530241,
-0.0795860291,
-0.0460909903,
0.0747466609,
-0.0185414739,
0.0023478693,
-0.0043683983,
-0.009665709,
0.0214287452,
-0.0197381601,
-0.0344213471,
-0.0535818636,
0.0079930825,
0.0127987005,
-0.00868099,
-0.0684050694,
-0.0012242773,
-0.0380498394,
-0.037046805,
0.0178674702,
0.0101507828,
-0.028253179,
0.0040802713,
-0.0170311555,
0.0543884113,
-0.015939137,
0.0081897732,
0.0279079359,
0.0751545802,
-0.0240362566,
0.0026688389,
-0.0750638321,
-0.0100115193,
-0.0399334431,
0.0372186564,
-0.002512038,
0.0600853972,
-0.0417431258,
-0.0200106259,
-0.0218911208,
0.0724773109,
0.0229914952,
0.000708224,
-0.0386530049,
-0.0302552208,
0.0136776837,
0.0545775741,
-0.0417193286,
0.0427225195,
-0.0157338809,
-0.0169929229,
0.0269223433,
0.0837746859,
0.0118091693,
-0.0046480605,
0.0157271605,
0.0046468591,
0.0183468852,
0.0506196544,
0.0564671494,
-0.0278396066,
0.0168529917,
-0.0513156578,
-0.0067476272,
0.012408142,
-0.0151200695,
-0.0181863364,
0.029260356,
-0.0393252186,
-0.0149157038,
-0.0071649589,
0.0413978733,
0.0251873359,
-0.010451314,
-0.0260506701,
0.0391392149,
-0.0130755436,
-0.0331774093,
-0.0360019282,
-0.0378571525,
-0.0387744121,
0.0299768616,
-0.0080333352,
0.0083930073,
-0.0084804464,
-0.0019299793,
-0.0284686889,
0.0252812505,
-0.0470031351,
-0.0166953914,
0.0455919951,
-0.0232727863,
0.0382168591,
0.0160601083,
-0.0090358853,
0.0507939681,
0.0803753361,
-0.0205865949,
-0.0446110517,
-0.0021289105,
-0.0074631576,
-0.0056599788,
0.0401084349,
0.0030880938,
0.0076498413,
0.0625910237,
-0.0135387741,
0.0377428643,
0.0090013342,
0.0341387615,
0.0195167847,
-0.0583381616,
0.0043120137,
0.0027015135,
0.0158274621,
-0.002097622,
0.0026803375,
-0.0439581126,
0.0844924673,
-0.0595428087,
-0.0029847149,
-0.0152445016,
-0.0201526377,
0.0123919118,
0.0452835523,
0.0100514768,
0.0299228188,
0.0085577546,
0.0482934192,
-0.0309963878,
-0.0265699588,
0.0327579826,
-0.0332532823,
-0.0201535281,
-0.0358529538,
0.0229310188,
0.0079128677,
-0.0277437083,
-0.013339296,
0.0598840825,
0.0088027287,
0.0340871178,
0.0020979801,
-0.0454914533,
-0.0290741641,
-0.0049471953,
0.1000344679,
-0.0169036351,
-0.015349553,
-0.0214288812,
-0.0451100804,
0.027301427,
0.0040219831,
-0.0130082062,
0.0288243052,
-0.0233690962,
0.0166136865,
0.0682527423,
0.0248790178,
-0.0265696794,
-0.0738580599,
0.0040863487,
0.0438506231,
0.063519679,
0.02833567,
0.0249376427,
0.0113623338,
-0.0626915991,
-0.0143217668,
0.0164808724,
0.0118754897,
-0.0427221544,
0.009824018,
0.0092833852,
0.0602807663,
0.0144899487,
0.0277073961,
-0.0685864463,
-0.0325174741,
-0.0020587393,
0.028976934,
0.0237195846,
-0.0179201476,
0.0142301898,
-0.0646994486,
0.0734979585,
-0.0055342219,
-0.0379195139,
-0.042636767,
-0.0579921119,
-0.0073069301,
0.0285581145,
0.0021628647,
0.1024222076,
-0.0085191755,
-0.0223327298,
0.0200732872,
0.0366051719,
-0.0490605608,
0.0290255416,
0.0062178439,
0.0078911651,
0.0491625667,
0.0055598472,
-0.0376379602,
-0.0340187438,
0.0129906153,
0.0155662587,
0.0337818079,
-0.0407747775,
0.0131495213,
0.0426841192,
-0.0098133199,
-0.0192066357,
0.0242712889,
0.0053809746,
-0.0771624297,
-0.0242981166,
0.0847787485,
-0.038289804,
0.0272224247,
-0.0414190143,
-0.0283013452,
0.0225766674,
0.0425153151,
-0.0265380386,
0.0152602829,
-0.0995256826,
-0.0127048288,
0.020975396,
-0.0257676262,
0.0069592511,
0.0271290578,
0.0239203516,
0.042774342,
-0.0043161912,
0.0094827786,
-0.0090170773,
0.0035391164,
-0.0360368565,
-0.0227880944,
-0.0338253751,
-0.0128438547,
-0.0766517073,
0.0229577515,
-0.0325182974,
0.0096928133,
0.0009694163,
-0.0441897549,
0.007035763,
-0.064320229,
-0.0073608616,
0.0603720807,
-0.0334218033,
-0.0235078186,
-0.0056230612,
0.0624158718,
0.0220237244,
0.0308932289,
0.0142434919,
-0.0505205281,
0.0652560294,
0.0542118885,
-0.0544592664,
-0.0004264566,
-0.0226975437,
-0.0032642586,
-0.0309842695,
-0.0380659215,
-0.0371457636,
0.0306026861,
-0.0108264619,
0.050023213,
0.0372502096,
-0.0501524918,
0.0108497785,
0.0282466579,
0.0226991363,
0.0039380104,
0.0181978624,
-0.0345471166,
-0.0468450934,
0.0404911973,
0.0381896608,
-0.0460830368,
0.0163689703,
0.0297771227,
-0.0723949,
0.008101766,
-0.0197991412,
0.0332727768,
0.0267366357,
0.0044322903,
0.0362755135,
-0.0018002623,
0.0337785743,
0.089413099,
0.0034245029,
-0.0148234284,
0.0311995856,
-0.0554138236,
0.0761770383,
-0.0270303469,
0.0430614725,
0.0291949231,
0.0205387585,
-0.0371489711,
0.0349805839,
-0.047771398,
-0.106330581,
-0.0393378362,
0.01203172,
0.0319345891,
0.0465975776,
0.0459937602,
-0.0560742281,
-0.0093808845,
-0.0242161863,
0.0239839684,
0.0439549573,
-0.0447620675,
0.0619229376,
0.0182009209,
-0.0086507797,
-0.0268434361,
0.0387827791,
0.0052625919,
-0.0028504604,
-0.0515948236,
-0.0293757711,
-0.0202192776,
-0.0691286922,
-0.0526378602,
-0.0040792013,
-0.0227238946,
-0.0446189865,
-0.0455347374,
-0.0072309515,
0.0015677498,
0.0548608936,
-0.0625231341,
-0.0445464551,
0.0141753173,
0.1242234781,
-0.0084356265,
-0.0291899443,
-0.0063610924,
-0.0022910805,
-0.0098762838,
-0.015003941,
0.0278694183,
-6.527387887e-33,
-0.0212353896,
0.0318669304,
0.0253690761,
-0.0256673135,
-0.0215579905,
-0.0543162711,
-0.0882081836,
-0.0236535575,
0.0153737627,
0.038340684,
-0.0081961071,
-0.0018650544,
0.0182712413,
-0.0210260842,
0.0391411372,
0.033468958,
0.0005729544,
0.0531077832,
-0.013124668,
0.010441307,
0.0560363866,
0.0471269786,
0.0454832688,
-0.0338983089,
0.0254092142,
0.0319594555,
0.0271803197,
-0.0251639877,
-0.0418084115,
0.0476850607,
-0.0395693406,
0.0302704293,
-0.0591108501,
-0.0228464957,
-0.0154931145,
-0.0199682415,
-0.0623120144,
0.0059273322,
-0.0057635084,
-0.0229204986,
-0.0379220843,
0.0143489763,
0.004969052,
0.0139551926,
-0.0533681214,
-0.0083267959,
0.0244624335,
-0.0453523993,
-0.0114141284,
-0.0207245275,
-0.0884870514,
-0.0131992064,
0.0122134117,
0.0281723589,
-0.0253235642,
0.0148438551,
-0.0129113412,
0.0313307382,
-0.0758603811,
0.0258288812,
0.002064575,
0.0166902486,
0.0175114814,
0.0514790043,
0.0226706713,
0.0247909576,
0.0423836559,
0.0150543479,
-0.0132395411,
-0.0043188743,
0.024296442,
0.0899788365,
0.0224740449,
0.0606013127,
0.0553101711,
-0.0041360552,
-0.1326812208,
-0.0073071728,
0.0126113836,
0.0472180806,
0.0449651666,
0.0600742437,
0.0271129441,
-0.0267842319,
-0.0456836112,
0.0384919867,
0.0199169591,
-0.053626772,
-0.0146587752,
0.007008268,
-0.0707434192,
-0.0490318798,
0.0404850766,
-0.0241495222,
-0.0111200111,
0.0500058196,
-0.0157134086,
-0.0068429704,
-0.0341688693,
-0.00254889,
-0.0346372128,
-0.0225112438,
-0.0242104102,
-0.0187521372,
0.0301676057,
0.0231591109,
-0.023321772,
0.0326508284,
-0.0007079694,
0.003481793,
-0.0996380225,
-0.0044382694,
0.0172940232,
-0.0145632913,
0.0015421448,
-0.0106403828,
-0.0172787961,
0.0082967523,
-0.0237953365,
0.0504189655,
0.0157134626,
0.0350576378,
-0.0184705369,
0.0043167342,
-0.0096120518,
-0.0003493165,
-0.0978176296,
0.0031252925,
0.0272125602,
-0.0674360469,
0.0157637279,
0.0236637965,
3.303e-7,
0.0365066044,
0.0409024283,
0.0440144166,
0.0086787511,
-0.0126643237,
0.0236822311,
-0.0120699825,
-0.0429709814,
0.0078122537,
-0.0566184334,
0.0416523479,
0.049632024,
0.0175045505,
0.0470123626,
-0.0430661403,
-0.0186627433,
-0.0399574973,
-0.0320180841,
-0.0477456823,
-0.0023699342,
-0.0262129027,
0.0470980108,
-0.015237432,
-0.0608284958,
0.0169629715,
-0.0469603725,
0.0802189261,
-0.0140350461,
0.028280884,
0.0572992712,
-0.0020174077,
-0.06516321,
-0.0167102162,
0.0326872654,
-0.014980346,
-0.0419331603,
-0.0275381152,
0.0082318168,
0.0215758886,
0.0072033014,
0.020888228,
-0.0298942626,
0.0216038618,
-0.0270650275,
0.0195419323,
-0.0109115969,
-0.0102009634,
0.0411009081,
-0.0138641801,
-0.015255386,
0.0071610184,
-0.0248181801,
0.0178807247,
-0.009372239,
-0.0359195769,
0.0399138071,
0.0055820253,
-0.0373349302,
-0.0288014337,
0.0414205231,
-0.0342591107,
-0.0126493275,
-0.0424540564,
-0.0267911144,
-0.0050178543,
-0.0179225951,
0.0083452817,
3.273043486e-34,
0.0543834642,
0.0030417922,
0.0089730006,
-0.0040649623,
0.0448713377,
-0.0042750775,
-0.0065478631,
0.0143269133,
-0.0088134101,
-0.045117002,
-0.0202245638
] |
http://arxiv.org/abs/2401.17228v1 | http://arxiv.org/pdf/2401.17228v1 | Morality is Non-Binary: Building a Pluralist Moral Sentence Embedding Space using Contrastive Learning | 2024-01-30T18:15:25 | 2024-01-30T18:15:25 | Jeongwoo Park, Enrico Liscio, Pradeep K. Murukannaiah | cs.CL | cs.CL, cs.AI | Recent advances in NLP show that language models retain a discernible level
of knowledge in deontological ethics and moral norms. However, existing works
often treat morality as binary, ranging from right to wrong. This simplistic
view does not capture the nuances of moral judgment. Pluralist moral
philosophers argue that human morality can be deconstructed into a finite
number of elements, respecting individual differences in moral judgment. In
line with this view, we build a pluralist moral sentence embedding space via a
state-of-the-art contrastive learning approach. We systematically investigate
the embedding space by studying the emergence of relationships among moral
elements, both quantitatively and qualitatively. Our results show that a
pluralist approach to morality can be captured in an embedding space. However,
moral pluralism is challenging to deduce via self-supervision alone and
requires a supervised approach with human labels. | Morality is Non-Binary: Building a Pluralist Moral Sentence Embedding Space using Contrastive Learning: Recent advances in NLP show that language models retain a discernible level
of knowledge in deontological ethics and moral norms. However, existing works
often treat morality as binary, ranging from right to wrong. This simplistic
view does not capture the nuances of moral judgment. Pluralist moral
philosophers argue that human morality can be deconstructed into a finite
number of elements, respecting individual differences in moral judgment. In
line with this view, we build a pluralist moral sentence embedding space via a
state-of-the-art contrastive learning approach. We systematically investigate
the embedding space by studying the emergence of relationships among moral
elements, both quantitatively and qualitatively. Our results show that a
pluralist approach to morality can be captured in an embedding space. However,
moral pluralism is challenging to deduce via self-supervision alone and
requires a supervised approach with human labels. | [
0.0302124061,
0.0878767371,
-0.0237540249,
0.056839101,
-0.0428728089,
0.01239195,
0.0072543966,
-0.0359847061,
0.0010194794,
-0.0183509942,
-0.0166820027,
-0.0564108938,
-0.0193102099,
-0.0374007374,
0.0155407991,
-0.0339170918,
0.0512468517,
0.031190699,
0.0798551366,
-0.0249454193,
0.0150112547,
0.018825544,
0.0240858179,
0.0009467235,
0.0089320932,
0.0034749892,
0.0010292405,
-0.0535594709,
0.0598706082,
-0.0203390587,
0.0515397564,
0.046313189,
0.0452697985,
0.0121438522,
0.0000016179,
-0.0246101245,
0.0152483415,
-0.0002639283,
0.0273973923,
-0.013271071,
0.0379784144,
-0.0273831207,
0.0181887448,
-0.0158065408,
-0.0581560992,
0.0171350017,
0.0269130599,
0.101967603,
0.0338464677,
0.0477515385,
-0.003521705,
-0.0550909527,
0.0101726679,
-0.0088762902,
-0.0274673998,
0.0614382587,
0.015962122,
0.0083775017,
-0.0194678176,
0.0348244756,
-0.0313726254,
-0.0047221743,
0.0353930444,
0.0122039076,
-0.003792777,
0.0044740778,
0.0037465105,
-0.0372956693,
-0.0267223157,
0.045660831,
0.0281022917,
-0.0152477166,
0.0360324495,
0.0059797391,
0.0023919225,
-0.008140468,
0.0188579801,
0.0560301282,
-0.0073853056,
0.0469875485,
0.0531412102,
0.0042520734,
-0.0085786516,
0.0270006564,
0.0280089695,
-0.0270935353,
-0.0042741164,
-0.0205847546,
0.0225138571,
0.0131039713,
-0.0893792957,
0.0291461293,
0.0300779492,
-0.0030432176,
0.0019177018,
0.02665562,
-0.0163617432,
-0.0080968551,
0.057132896,
-0.0102503812,
0.0397452004,
-0.0050730612,
-0.0225398969,
0.0382094234,
-0.0078666033,
0.016010819,
-0.0018977482,
0.0247080084,
-0.0549217984,
0.0344748013,
-0.0010267356,
-0.0010980507,
0.0147660142,
0.0569560416,
0.0002080528,
0.0365494452,
-0.0293560307,
-0.0202036053,
-0.0734076425,
-0.0121566756,
-0.0086020371,
0.0145011311,
-0.0093953367,
0.0059143915,
-0.0046239793,
0.002323003,
-0.0702563599,
0.0075421911,
-0.0300320368,
0.0063585015,
0.0156397838,
-0.0327825397,
-0.0266211256,
0.0019122622,
0.0112177664,
0.0683346391,
-0.0114629017,
-0.0150342779,
-0.0081384061,
-0.0052422555,
-0.0415001251,
-0.0195692796,
-0.0305698551,
-0.0509269647,
0.022839861,
-0.0130851623,
-0.0123562627,
-0.0515802912,
0.0401665866,
0.0166157968,
-0.0302484948,
0.0851408765,
0.0192613639,
-0.0370865129,
-0.0320968516,
0.0103857331,
0.0374578461,
0.0547456332,
0.0199781079,
-0.0044241832,
0.0262217801,
0.0282397494,
-0.0273157097,
-0.0648104772,
0.0107394606,
-0.0047080219,
-0.058038611,
0.0200589187,
0.0238251686,
0.0033419463,
0.0073082778,
0.0721619353,
-0.006287944,
-0.0664740205,
0.0283298567,
0.0984462723,
0.0182914902,
0.0352699682,
0.0514811277,
-0.0183306467,
-0.0206111316,
0.0417842343,
-0.0568616875,
0.0108439242,
0.0270876009,
-0.0175718945,
-0.0200897492,
0.0629931241,
-0.0050684619,
-0.0508547239,
0.0204257239,
-0.0095596388,
0.0200674012,
-0.0515501313,
0.0200697519,
-0.0153522817,
0.0047531254,
-0.0633217767,
0.0840630382,
-0.08990293,
-0.0045961067,
0.0131682474,
0.1083991081,
0.0708635524,
0.0275915135,
-0.0251405891,
-0.01638286,
-0.0453219675,
-0.0404658988,
0.0183489323,
-0.0032452287,
-0.035839241,
0.0068495157,
-0.0455780402,
-0.0252230298,
-0.0246031005,
0.0066979341,
0.0093807802,
-0.0105910823,
-0.039850954,
0.0022290896,
0.0077647744,
0.0293547474,
0.0606950298,
-0.003307495,
0.0159755554,
-0.0237187613,
0.0442614779,
0.0094645843,
0.0093507003,
0.0147048589,
-0.0455878004,
0.0174047966,
-0.0208828561,
-0.0341116078,
-0.0016201264,
-0.0431108251,
-0.0650754422,
0.0424692035,
-0.1145187169,
-0.0212034099,
0.0452794693,
-0.0276567675,
0.0155915162,
-0.0472119041,
-0.022263743,
-0.0123550789,
-0.0491579399,
-0.0262742192,
-0.0241499189,
-0.0180761516,
-0.0082360003,
0.1015924439,
-0.043629542,
0.0435833074,
-0.0781407878,
-0.0185475927,
0.0540915653,
0.0076775225,
-0.0085089467,
-0.0597033091,
-0.044334244,
0.0398395248,
0.0371864587,
0.0507089198,
-0.0252635814,
-0.0409635343,
-0.058248736,
-0.0112461345,
0.0169014465,
0.0074097039,
-0.0079080481,
-0.0443253405,
0.0506619699,
0.0054579019,
-0.0507612899,
-0.0306369681,
-0.0271862727,
-0.0131488303,
0.080154568,
-0.0339615121,
-0.0035742717,
0.0028853407,
-0.0083668577,
-0.0557995625,
0.0027657733,
-0.0239148531,
0.0103903776,
-0.0276645217,
0.034631718,
-0.058718089,
-0.0032530064,
0.0068533979,
-0.0704389438,
0.0159808565,
0.0482879132,
0.0391630195,
0.0765263885,
0.0571645312,
0.0512885302,
0.0309016425,
0.0677287728,
0.0037965223,
-0.0065629189,
-0.0392762832,
0.0577130578,
-0.0063448665,
-0.0204167329,
-0.0329264626,
0.1174481213,
-0.0025714415,
-0.0307618491,
-0.0455018915,
0.0179028921,
-0.0336572193,
-0.0623959899,
-0.01524135,
-0.0372530185,
-0.0189216621,
0.0017656052,
-0.031391222,
-0.048394531,
0.0710150078,
0.0642481521,
0.0091908807,
-0.0062774378,
-0.0182639845,
0.0072678593,
0.033589799,
-0.0057074553,
0.0174085777,
0.0457635373,
-0.0151187927,
-0.0332333967,
0.0073698177,
0.0400706045,
-0.0204937942,
0.0931977406,
-0.021866437,
0.0034189946,
-0.0042735836,
-0.1080711782,
-0.0463989116,
-0.0621802658,
-0.0076136612,
-0.0020957214,
0.0037514733,
0.0087524867,
0.0322873779,
-0.0112652825,
0.0037954221,
-0.0435442701,
-0.0116698779,
-0.005891074,
-0.0254536998,
-0.0058477866,
0.0187085755,
0.0002296681,
0.0602477044,
-0.012889876,
-0.0038186545,
0.0557442196,
0.0146035664,
0.0118454779,
0.0064933542,
0.0235523339,
-0.0048628589,
-0.0099067828,
-0.012007107,
0.0114071006,
0.054955259,
0.0022210248,
0.0163848009,
0.0669181868,
0.0330311842,
-0.0474439114,
0.0181862917,
-0.0031286662,
-0.0259836894,
0.0160124004,
-0.0298507679,
-0.0906764641,
-0.002964929,
-0.0238938853,
-0.0292194672,
-0.0443468057,
-0.0460220948,
0.0176643953,
-0.0562192798,
-0.0027947209,
0.0483011566,
-0.0225853492,
-0.0466247723,
-0.010128797,
-0.0061470568,
-0.026715342,
0.0294420756,
0.020441452,
-0.0086385654,
-0.0020593393,
0.0025758671,
0.0587148778,
0.0154457809,
0.0128861004,
-0.0390908346,
0.0690801889,
0.014159156,
0.0581445172,
-0.0106284861,
-0.0314210802,
-0.0325964428,
-0.0370488912,
-0.0298051331,
0.0179437883,
-0.0007155446,
-0.0221466459,
-0.0075304625,
-0.0679182932,
0.0492712781,
-0.0197625626,
-0.0356690139,
-0.0017742413,
0.0664529875,
0.0234670974,
-0.0384984612,
-0.0213599056,
0.0288082249,
-0.0213374421,
0.0585702509,
0.0101777278,
-0.0304224901,
-0.0131732095,
-0.061728172,
0.0125043932,
-0.0035395648,
-0.0431788154,
-0.0487718731,
0.0076766568,
0.0095609901,
-0.0477000885,
-0.0288378429,
-0.0329820588,
0.0011762297,
0.0215774179,
-0.0747098178,
0.019102633,
-0.0052797869,
-0.0399933681,
-0.0225292705,
-0.0306054205,
-0.0213783532,
0.001176486,
0.0021079299,
0.0010424501,
-0.0016161783,
-0.0238372516,
-0.071672529,
0.0004987172,
-0.0409121588,
0.0253153369,
0.0443178006,
-0.0019210377,
-0.0078951549,
-0.0179081857,
0.0220447239,
0.0475228094,
0.0758712143,
-0.0905496702,
0.0260321721,
0.0844064429,
0.0501211509,
0.0003977414,
0.0372375958,
0.0035283184,
0.0173779726,
0.0360055678,
0.0143370153,
-0.0411213152,
0.0144253485,
0.0330084637,
-0.0076516126,
0.0354165286,
0.0050096377,
0.0275627971,
-0.0560799465,
-0.0577334538,
-0.0283584557,
-0.0505030937,
-0.0336724371,
0.0073492574,
0.0340287946,
0.0113190608,
-0.0284965988,
0.0246867333,
0.0003768417,
0.0211207811,
-0.0156798903,
0.0053025172,
0.0969962999,
0.0413228162,
-0.0255652796,
-0.0113567859,
-0.0325137079,
0.0262773279,
-0.0134044085,
0.0262459926,
-0.0284883864,
0.0476035066,
0.0364820845,
0.0073630763,
-0.015466433,
0.0096097719,
-0.0662260801,
0.0533350781,
-0.048743926,
-0.0541979782,
0.0033713777,
0.0344989151,
-0.0048682885,
0.0015241229,
-0.0209545009,
-0.0124262413,
0.052201882,
-0.0220936183,
-0.0025428936,
0.0147910556,
0.0372025296,
-0.0432745293,
-0.0110548856,
0.0432081223,
-0.0866192803,
0.0125444783,
0.0364946537,
-0.0016504705,
-0.0439440347,
-0.0289331228,
0.0341197364,
-0.014040947,
-0.0048207091,
0.0532248616,
-0.0595373362,
-0.0216453373,
-0.0132842744,
0.0265126936,
0.0170230959,
-0.0192018878,
-0.0247909203,
0.0393796228,
-0.0016338637,
0.0073778564,
-0.0289446414,
0.0224186461,
-0.0006994847,
0.0118337804,
-0.062077634,
-5.248710479e-33,
-0.0424439646,
-0.016521344,
-0.0528854802,
-0.0400193371,
0.0001594595,
-0.0217914153,
0.0245128404,
0.0201674365,
-0.046003893,
0.0937778726,
0.0182557702,
0.0268279761,
0.0009046777,
0.0099838236,
0.0351742283,
-0.0501961596,
0.0165545586,
0.0028411229,
-0.0152917961,
-0.0003770989,
-0.0086009717,
0.006041138,
0.0500628911,
-0.0710681155,
0.0242292825,
0.0256546661,
-0.010965609,
0.043308977,
-0.0027454833,
0.0273689646,
-0.0149124302,
0.0737727731,
0.0211626254,
-0.0143752079,
0.0127507709,
0.0475384854,
-0.0506394207,
-0.0545011275,
0.0383112803,
-0.0442351103,
-0.0539474338,
-0.0485923439,
0.0026405654,
-0.0142396996,
0.0004454383,
-0.0360422581,
0.0579785779,
-0.0079298224,
0.0163737219,
0.0020240147,
-0.068434827,
-0.0204236191,
-0.0278355386,
-0.0055043595,
-0.0296071637,
-0.0061186994,
-0.0043509132,
0.0252085701,
0.0120211365,
0.0316445753,
0.0016496881,
0.0085502854,
-0.0424457192,
-0.0579907782,
0.0289352499,
0.0851071626,
-0.0056470539,
0.0126052964,
-0.00282493,
-0.0063660555,
0.0210460592,
0.021772366,
0.0093871066,
0.0623770952,
-0.0134723214,
-0.0242464803,
-0.043247316,
-0.0343883485,
0.0876876041,
-0.0351502448,
0.0192814879,
0.0108569404,
0.0062648826,
0.0121891061,
-0.0087024858,
-0.1115402579,
-0.0140320696,
-0.0084432047,
-0.0229283366,
0.0001452764,
-0.0373315699,
-0.0383569933,
-0.0402603671,
-0.0651641712,
0.0166787319,
0.0574402772,
-0.004696724,
-0.0166740026,
0.0281234421,
0.0175421033,
-0.0717721283,
-0.0502547324,
-0.0051195016,
0.000550574,
0.0163415521,
-0.0072400263,
0.0712543428,
-0.017957611,
-0.0149537697,
-0.0190348756,
-0.0082932739,
-0.0185618456,
-0.0168625508,
-0.0092154862,
0.0323882475,
0.0362513959,
-0.0005244549,
0.0154859098,
0.0275355261,
-0.0031648155,
-0.0333007127,
-0.0200689659,
-0.0278561786,
0.053481672,
-0.01805293,
-0.0021440443,
-0.0339935757,
0.0566930324,
0.0227576587,
-0.0200235303,
-0.0326131321,
0.0173585191,
2.282e-7,
0.0227632578,
-0.0125230914,
0.0283877514,
0.0133604351,
-0.0201366898,
0.061107941,
0.0585503504,
0.0455421358,
-0.0127515336,
-0.0537539311,
-0.0091308337,
0.02257015,
0.0461064577,
-0.0040491344,
-0.0004964742,
0.0161855109,
-0.0094871726,
-0.0397330225,
0.0006679852,
0.0445161499,
0.0547976643,
0.085537158,
0.0655746758,
-0.0038562883,
0.0080253668,
0.046310693,
0.0160355568,
-0.0368751064,
-0.0316391289,
-0.0292759389,
-0.0293836221,
-0.0484452397,
0.0089871977,
-0.0021069215,
0.0434002653,
0.0396383516,
0.0014122949,
0.0631824359,
-0.0066207196,
-0.025046438,
-0.006940152,
-0.0301592555,
-0.0217174478,
0.0058613513,
0.0711789876,
-0.0533523671,
-0.0235262047,
-0.0418910906,
-0.0140895611,
-0.0150847472,
0.0169838592,
0.0022123731,
-0.0431571007,
0.0014855596,
0.0159339849,
0.0136346025,
0.0213575475,
-0.0554399826,
-0.0100908168,
0.0410320945,
0.0066007059,
-0.0472927205,
0.067435205,
0.0306639392,
0.0451802611,
-0.0456037857,
0.0125331804,
1.76637966e-34,
0.0475663915,
-0.0616077296,
0.0255142581,
0.0278404746,
0.0327470005,
0.0036518683,
-0.0358169824,
-0.0187364537,
0.0715105385,
-0.0045695631,
-0.003311198
] |
http://arxiv.org/abs/2401.17221v1 | http://arxiv.org/pdf/2401.17221v1 | MouSi: Poly-Visual-Expert Vision-Language Models | 2024-01-30T18:09:11 | 2024-01-30T18:09:11 | Xiaoran Fan, Tao Ji, Changhao Jiang, Shuo Li, Senjie Jin, Sirui Song, Junke Wang, Boyang Hong, Lu Chen, Guodong Zheng, Ming Zhang, Caishuang Huang, Rui Zheng, Zhiheng Xi, Yuhao Zhou, Shihan Dou, Junjie Ye, Hang Yan, Tao Gui, Qi Zhang, Xipeng Qiu, Xuanjing Huang, Zuxuan Wu, Yu-Gang Jiang | cs.CV | cs.CV, cs.AI, cs.CL, cs.LG | Current large vision-language models (VLMs) often encounter challenges such
as insufficient capabilities of a single visual component and excessively long
visual tokens. These issues can limit the model's effectiveness in accurately
interpreting complex visual information and over-lengthy contextual
information. Addressing these challenges is crucial for enhancing the
performance and applicability of VLMs. This paper proposes the use of ensemble
experts technique to synergizes the capabilities of individual visual encoders,
including those skilled in image-text matching, OCR, image segmentation, etc.
This technique introduces a fusion network to unify the processing of outputs
from different visual experts, while bridging the gap between image encoders
and pre-trained LLMs. In addition, we explore different positional encoding
schemes to alleviate the waste of positional encoding caused by lengthy image
feature sequences, effectively addressing the issue of position overflow and
length limitations. For instance, in our implementation, this technique
significantly reduces the positional occupancy in models like SAM, from a
substantial 4096 to a more efficient and manageable 64 or even down to 1.
Experimental results demonstrate that VLMs with multiple experts exhibit
consistently superior performance over isolated visual encoders and mark a
significant performance boost as more experts are integrated. We have
open-sourced the training code used in this report. All of these resources can
be found on our project website. | MouSi: Poly-Visual-Expert Vision-Language Models: Current large vision-language models (VLMs) often encounter challenges such
as insufficient capabilities of a single visual component and excessively long
visual tokens. These issues can limit the model's effectiveness in accurately
interpreting complex visual information and over-lengthy contextual
information. Addressing these challenges is crucial for enhancing the
performance and applicability of VLMs. This paper proposes the use of ensemble
experts technique to synergizes the capabilities of individual visual encoders,
including those skilled in image-text matching, OCR, image segmentation, etc.
This technique introduces a fusion network to unify the processing of outputs
from different visual experts, while bridging the gap between image encoders
and pre-trained LLMs. In addition, we explore different positional encoding
schemes to alleviate the waste of positional encoding caused by lengthy image
feature sequences, effectively addressing the issue of position overflow and
length limitations. For instance, in our implementation, this technique
significantly reduces the positional occupancy in models like SAM, from a
substantial 4096 to a more efficient and manageable 64 or even down to 1.
Experimental results demonstrate that VLMs with multiple experts exhibit
consistently superior performance over isolated visual encoders and mark a
significant performance boost as more experts are integrated. We have
open-sourced the training code used in this report. All of these resources can
be found on our project website. | [
0.0365186036,
0.028970547,
0.0194956232,
0.0545003377,
-0.0132619757,
-0.0093231807,
0.0290562678,
-0.0056311693,
-0.0065623843,
-0.0478995629,
-0.0019934746,
-0.0153759429,
0.0267365016,
0.0506753959,
0.0087481672,
-0.0492708385,
-0.0211178288,
-0.0401932113,
-0.0128501095,
0.0036923967,
0.0287914518,
0.0098769227,
0.0297796205,
0.0398485921,
-0.0145963999,
0.0127550848,
-0.0129602998,
0.0306986403,
0.0352228284,
0.0098979445,
-0.0437449291,
0.0312457476,
0.0762978345,
-0.0179566424,
0.0000021478,
-0.0236454252,
-0.0038157399,
-0.0012685455,
0.0053304611,
0.0032523109,
-0.0079739112,
-0.0484389216,
0.0049797613,
0.0000133236,
-0.0090378281,
0.0015372444,
0.093902044,
0.0267528687,
0.0434744544,
0.0579532757,
-0.0344090126,
-0.0345861949,
0.0381895564,
-0.017702885,
-0.0058827326,
-0.015463192,
0.0418586507,
0.0162799843,
-0.0159736834,
-0.0400622413,
-0.0454931706,
0.0496497303,
0.0238415934,
-0.0158937182,
0.0243792534,
0.0217181575,
-0.0006086023,
-0.0315399244,
-0.0457036234,
0.0284252521,
-0.080984883,
-0.0030568831,
-0.0315536484,
0.0009027416,
-0.0140283545,
0.0133964373,
-0.025748577,
0.0338891968,
-0.0167160798,
0.0031800496,
-0.0525933355,
0.0626006499,
-0.0015043929,
-0.0091672698,
-0.0414531268,
-0.0066997558,
0.0022580668,
0.0124171479,
-0.038724415,
0.0119770085,
-0.0366740003,
-0.0194477979,
0.0229238458,
0.0211568195,
0.0264663771,
-0.0015958718,
-0.0546021163,
-0.0338819548,
-0.0279163718,
-0.0320696905,
0.0264815539,
0.0434896909,
-0.0411057919,
0.069229573,
-0.0630973428,
0.0945950076,
-0.0241080709,
0.0322360396,
-0.0417023785,
-0.0604823865,
-0.0371751897,
-0.0207733288,
-0.0309502762,
-0.0103426892,
-0.0282231364,
-0.0447533727,
-0.0348211676,
0.0103664985,
-0.0295407698,
0.0019657162,
0.0315417573,
0.0210949983,
0.0205808524,
-0.0563245378,
-0.036054384,
0.0487761684,
0.0438745953,
-0.0028625871,
-0.0320317373,
-0.0551718101,
0.0273830015,
0.0684352964,
-0.0072543295,
-0.0020374688,
-0.0102125295,
0.0700044259,
-0.0548011139,
-0.0218424313,
0.0079977112,
0.0255617667,
-0.0237904172,
-0.0183331072,
0.0085327215,
-0.085732393,
0.0620922446,
-0.0061794473,
0.023220107,
-0.045686312,
0.0134924576,
-0.0207234994,
-0.0125625553,
0.0762973949,
-0.0511696041,
0.0115312999,
0.0246030409,
0.004607392,
0.0359198377,
0.0421698913,
0.0389517881,
0.0728257746,
-0.0002550155,
0.0120895263,
0.0374415666,
0.0660309494,
-0.0474040098,
-0.0256938674,
-0.0638003796,
-0.005656091,
0.0328780636,
0.0067436048,
0.0213787258,
0.057593897,
-0.0258884821,
-0.025770748,
0.0536986068,
0.0424907282,
0.0050662421,
-0.0022218849,
0.0328127109,
0.0041559786,
0.0128198005,
-0.0616690032,
-0.0109850336,
0.0493229143,
0.0008898419,
-0.0006266838,
-0.0273169409,
0.060002543,
-0.0338826589,
-0.0223073121,
-0.0200961065,
-0.0289622303,
-0.0236098748,
0.0328737795,
0.0639506802,
-0.0621289313,
-0.0685291588,
-0.0026552486,
-0.0298545957,
-0.0298804548,
0.0161966123,
-0.0344593525,
-0.0073382095,
0.1112050638,
0.0413088538,
0.0157533288,
0.0084096408,
0.0063431459,
-0.0167313647,
0.0308340788,
-0.0210299697,
0.0161191691,
0.0210177843,
-0.0925344154,
-0.0762176216,
0.0330637172,
-0.0380401947,
0.0382729322,
0.0254864916,
-0.0135305496,
0.0180544611,
0.0249443315,
-0.0292490497,
0.0464143567,
0.0026587173,
0.0247441586,
-0.044242546,
0.0345476717,
0.0035111581,
-0.0182978511,
-0.0300738458,
-0.0126684718,
0.0252159275,
-0.0621106997,
0.0236730017,
-0.0127989184,
-0.048660107,
-0.0164947938,
0.0514384843,
-0.0506114177,
0.0223915223,
-0.0011414849,
-0.0332174748,
-0.0448924676,
-0.0370848551,
-0.0296679828,
-0.0328191817,
-0.038519118,
0.0176648404,
0.0137293683,
0.0066393903,
-0.0452910326,
0.0457829796,
-0.0328827538,
-0.0618483461,
0.0051082219,
-0.0110087125,
0.0378629677,
-0.0399586,
0.0030843215,
-0.01900349,
0.0259301886,
0.0242167413,
0.0072959918,
-0.0013774975,
0.0068677021,
-0.0599729903,
-0.1057797596,
-0.0469729602,
-0.0315323584,
-0.0372750908,
-0.0099464767,
-0.0240195729,
-0.0035593817,
-0.0338495672,
-0.0182024669,
0.0298109502,
0.0102171209,
-0.0213963054,
-0.0149534419,
0.0240246467,
-0.0221152771,
-0.0192591529,
-0.0382324979,
0.0300383996,
0.0358082764,
0.0180147681,
-0.1182369217,
-0.0256027002,
-0.007717147,
-0.018881293,
0.0294799004,
-0.0254845563,
-0.0058343327,
-0.0276110359,
-0.0430311672,
-0.0024624248,
0.0640720129,
0.0166720338,
0.017394729,
0.0136539573,
0.0008222559,
-0.0283434708,
-0.0010914219,
-0.0081256004,
0.0046161353,
-0.0291221552,
0.0140175913,
-0.0525412671,
-0.0115594398,
-0.0182191879,
0.0384749621,
-0.068116352,
0.0115600638,
0.0488538966,
-0.0630342737,
-0.0699364245,
-0.0490687676,
0.0212164503,
0.0191953927,
-0.0716538131,
-0.0696012154,
0.0175204985,
-0.0417566635,
-0.0163331572,
0.0482490137,
-0.0080390871,
-0.0349507369,
0.0685250685,
0.0385563299,
0.0271305125,
-0.0416534357,
0.0159005243,
-0.0362791978,
-0.0242165569,
-0.0256440416,
-0.0143063944,
0.0558095612,
-0.0689575225,
0.0505108349,
-0.0338580981,
-0.0071344622,
-0.0227347091,
0.0246879682,
-0.0482182726,
0.0380877629,
-0.0435633808,
-0.0242387187,
-0.0488190763,
0.0688334703,
0.0212031249,
0.0497100875,
0.0115169594,
0.0238990281,
-0.0293642394,
0.0237528123,
0.0400642604,
0.0060744472,
-0.0434892029,
-0.066734314,
0.0016661281,
0.0548914112,
-0.0048992313,
-0.0186770782,
-0.0740678087,
0.0112737827,
0.0422532968,
0.0311144125,
0.0137093952,
0.0185712073,
0.0342213064,
-0.0720574036,
0.0296354424,
0.070413664,
0.0187920015,
-0.0360178351,
-0.0584668517,
0.0168543328,
0.0304129981,
0.0123587865,
0.0167436823,
-0.039589338,
-0.0383263342,
-0.0171207432,
-0.0374945253,
0.016249558,
0.0212235358,
-0.0414329283,
-0.0670013428,
0.0467750728,
0.0507953651,
-0.0053320345,
-0.0583952256,
0.0465916917,
-0.0160388499,
-0.0192318745,
-0.0215119552,
-0.0004161166,
0.0014939881,
-0.020898778,
0.0279697739,
0.0118494332,
-0.0117104845,
-0.0072714495,
-0.0569953658,
0.003199623,
0.1062425002,
-0.0221875347,
0.004840109,
-0.0587999895,
0.0166702718,
0.0039512906,
0.0549786501,
0.0018777063,
0.0495595224,
-0.0553581789,
-0.0259516072,
0.0296842437,
-0.0008198241,
-0.0210950114,
0.0069073131,
-0.0007710633,
0.0143657587,
-0.0158595871,
0.0007236228,
-0.0472605713,
-0.0357939973,
-0.0295664016,
0.0160902459,
-0.0128170932,
-0.0198166501,
-0.0239841007,
-0.0113381483,
-0.0214445181,
-0.023336526,
-0.0322158523,
0.0158240721,
0.0244350024,
-0.0561284982,
-0.0165901985,
-0.0073342822,
0.073819384,
0.0533602647,
-0.0099351415,
-0.0003758273,
-0.0146396263,
0.039500799,
-0.0708555356,
0.0577813275,
-0.0307329651,
0.0305082221,
-0.0542462431,
0.0416408256,
0.0090539567,
-0.0250206478,
0.0030701593,
0.0005622617,
-0.0430476293,
-0.0137475794,
-0.0242328215,
0.105728507,
0.0149932234,
0.035953477,
0.0132373227,
-0.0707252026,
0.0569627956,
0.0556135997,
0.0319884755,
0.0161833297,
-0.0265680458,
-0.0573828556,
-0.0174333025,
-0.0085454965,
-0.0231071208,
0.0201119389,
0.0154439304,
-0.0212102756,
-0.037218444,
-0.0502921641,
0.0087247305,
-0.0228506364,
0.0276230499,
-0.019702239,
0.0439748392,
-0.0172688551,
0.0239787661,
0.0064327829,
0.0117275212,
-0.081661962,
0.0155847371,
0.0197495203,
0.0220804829,
0.0712271407,
0.039305862,
0.0595198795,
-0.0372052565,
0.0080019608,
0.0101323202,
-0.0153176142,
-0.0020483134,
-0.0017565542,
0.0375547297,
-0.008756578,
0.0259773768,
-0.0075041745,
0.0090799117,
0.0070409705,
0.0285087619,
0.0161289796,
-0.0055719898,
-0.0166868567,
0.0645764321,
0.0027452491,
-0.0470774919,
-0.070964694,
0.0536443256,
0.0582798906,
0.0110519044,
0.0052441596,
-0.0193884261,
0.0050093029,
-0.0719057918,
0.015942689,
0.0167300459,
-0.0039657168,
0.0470285639,
0.0277880337,
-0.0106648318,
-0.0228303038,
0.0420515314,
0.0159256738,
-0.0032767567,
-0.0257213786,
-0.014859926,
-0.0369823687,
-0.0150219984,
0.0170584172,
-0.0106112445,
-0.0316930935,
0.0340866335,
0.0155120576,
-0.0027235313,
0.0118645094,
0.0237448756,
0.0013623,
0.0142357908,
0.0466096327,
0.0271539539,
-0.0078414539,
0.0066021327,
0.0085201645,
-0.0536003299,
0.0082709901,
0.009990206,
0.0149920369,
-5.948532224e-33,
-0.0172627028,
-0.0623813458,
0.0679678172,
-0.0466968343,
-0.0023267418,
0.0046812878,
0.0254907347,
0.0025683444,
-0.0056851511,
-0.0223693624,
0.0130369719,
-0.0110503528,
0.0220986493,
0.0441363156,
-0.0041198744,
0.0300869755,
0.0454671532,
0.012592352,
0.0274905711,
-0.0447084904,
0.0040309243,
0.0175755918,
0.0260448959,
-0.0845083371,
-0.0297831092,
0.0146569619,
-0.0157957096,
-0.0003755568,
0.0238422211,
0.0481789783,
0.0088119572,
0.0070247105,
0.014571514,
0.0133213839,
-0.0293232761,
0.0244599525,
-0.0177817717,
0.0199982077,
0.0019908047,
0.050561808,
-0.0621176995,
-0.0222085137,
0.0621189624,
-0.0067219902,
0.004365549,
-0.0365798771,
0.0609781034,
-0.0464296155,
-0.0224401597,
-0.0227951445,
-0.0112761147,
0.0347507782,
0.0209794734,
-0.0076139751,
0.0908381939,
0.0172707904,
0.0234992336,
0.0222753417,
-0.042621363,
0.0113590769,
0.0066553885,
0.050830815,
0.0971704125,
0.0766650736,
0.0168398283,
0.0441158749,
0.0084930453,
-0.0631080493,
-0.02481764,
-0.030214278,
0.0084239123,
0.0597628802,
0.0867205188,
0.0202128403,
-0.0031647531,
0.0295596067,
-0.016407622,
0.0174996778,
-0.0461677574,
0.0363871157,
0.0329277255,
0.0285985917,
-0.0058944048,
0.0063086119,
-0.0134246945,
-0.010133924,
-0.0112651652,
-0.0197062772,
0.0563128665,
0.0530318357,
0.0048906431,
0.0575330555,
0.078526251,
-0.0469015688,
0.0300583243,
-0.0021207556,
-0.0005698969,
0.0703866035,
0.0141970767,
0.0360097587,
-0.0726566017,
-0.0246562548,
0.0399089046,
-0.0013756629,
-0.0199463628,
0.0197044127,
0.0167735424,
0.0425336137,
-0.0344757624,
-0.0049796528,
-0.0288741644,
-0.020763414,
-0.0044389726,
0.0578026585,
0.0265922416,
0.0216160808,
-0.0243581478,
0.0359736644,
-0.0541190766,
-0.0271835979,
0.0126330275,
-0.0391225815,
0.0090681985,
0.0245552845,
-0.0028979683,
0.0310830344,
-0.0652265176,
-0.0412861295,
0.02110398,
0.0158391017,
0.0069094701,
0.0365147814,
2.749e-7,
0.0310896654,
0.0221553985,
0.055152528,
-0.0397228338,
0.008715285,
0.0337523185,
0.0036261159,
0.0112177907,
0.0031517905,
-0.0581999794,
0.0331906863,
0.0647767112,
0.0372273326,
-0.0385068171,
-0.0342127942,
0.032354366,
0.0012457442,
-0.0003031448,
-0.0597194955,
0.0100245504,
0.0498562865,
0.1089881063,
-0.0138083715,
-0.0147987632,
0.0326724164,
-0.039298173,
0.0722754672,
-0.0194469206,
0.0071446905,
0.0027292594,
0.0196293071,
0.0200609565,
-0.0525445379,
0.054247871,
0.0231105555,
0.0170114227,
0.0433790721,
0.0424561389,
-0.0732389539,
0.004099838,
-0.0066820695,
0.01721045,
-0.0527927466,
-0.0330292396,
0.0442490131,
-0.0455002822,
0.0163796917,
-0.0074390825,
-0.0298187044,
0.0219665319,
-0.0072147823,
0.0189544819,
-0.0070305788,
0.0334491767,
0.0361152254,
-0.0547396652,
-0.0412181504,
-0.066294387,
-0.0037123982,
-0.0633963346,
-0.0316818543,
-0.0000675118,
-0.0508959331,
0.0127258617,
0.0402533673,
-0.0024522417,
-0.0455854088,
2.252654696e-34,
0.0055537326,
-0.0020868077,
0.0003382348,
0.017804889,
0.0020526422,
-0.0389809497,
-0.0893249437,
0.0298847035,
-0.0309761837,
-0.0466509871,
-0.0281290486
] |
http://arxiv.org/abs/2401.17206v1 | http://arxiv.org/pdf/2401.17206v1 | Gazetteer-Enhanced Bangla Named Entity Recognition with BanglaBERT Semantic Embeddings K-Means-Infused CRF Model | 2024-01-30T17:47:07 | 2024-01-30T17:47:07 | Niloy Farhan, Saman Sarker Joy, Tafseer Binte Mannan, Farig Sadeque | cs.CL | cs.CL | Named Entity Recognition (NER) is a sub-task of Natural Language Processing
(NLP) that distinguishes entities from unorganized text into predefined
categorization. In recent years, a lot of Bangla NLP subtasks have received
quite a lot of attention; but Named Entity Recognition in Bangla still lags
behind. In this research, we explored the existing state of research in Bangla
Named Entity Recognition. We tried to figure out the limitations that current
techniques and datasets face, and we would like to address these limitations in
our research. Additionally, We developed a Gazetteer that has the ability to
significantly boost the performance of NER. We also proposed a new NER solution
by taking advantage of state-of-the-art NLP tools that outperform conventional
techniques. | Gazetteer-Enhanced Bangla Named Entity Recognition with BanglaBERT Semantic Embeddings K-Means-Infused CRF Model: Named Entity Recognition (NER) is a sub-task of Natural Language Processing
(NLP) that distinguishes entities from unorganized text into predefined
categorization. In recent years, a lot of Bangla NLP subtasks have received
quite a lot of attention; but Named Entity Recognition in Bangla still lags
behind. In this research, we explored the existing state of research in Bangla
Named Entity Recognition. We tried to figure out the limitations that current
techniques and datasets face, and we would like to address these limitations in
our research. Additionally, We developed a Gazetteer that has the ability to
significantly boost the performance of NER. We also proposed a new NER solution
by taking advantage of state-of-the-art NLP tools that outperform conventional
techniques. | [
0.0384581424,
-0.0290659629,
0.0157043971,
0.0247876011,
-0.016880326,
0.04605259,
-0.0248754416,
0.0222556423,
0.0283597391,
-0.0159428902,
0.0240199249,
-0.0293167848,
0.0222012401,
-0.0150294043,
-0.0093664695,
-0.0685679466,
0.0320140906,
0.0269234627,
0.0319524519,
0.041649662,
0.0159804821,
0.0129924314,
-0.0031812531,
-0.0104395663,
0.007331789,
0.0114856595,
-0.0122689363,
-0.016485814,
0.0163989924,
-0.0704165772,
-0.0385528654,
0.0720480531,
-0.0265896861,
0.0217642374,
0.0000017616,
-0.0607498847,
-0.0672911629,
0.009802415,
-0.0081363665,
-0.0321298614,
0.0148813436,
-0.0062081786,
-0.0271159858,
0.0052239792,
0.0171714146,
-0.063599579,
0.0504125021,
0.0766721219,
-0.0268422142,
0.0770490915,
0.0133919017,
-0.0790549591,
0.0370451659,
0.0255691335,
-0.0031003356,
-0.0349577628,
0.0038456246,
-0.0248573497,
0.0276557673,
0.0153771751,
-0.0297445431,
0.0273074824,
-0.0349089727,
-0.0040050382,
0.0352487043,
0.0230785422,
0.006762966,
-0.0012175069,
0.0260209143,
0.0363459848,
-0.0331574008,
-0.0210162941,
0.0482070372,
0.0039261249,
-0.0413453504,
-0.0125469686,
-0.0175285395,
0.0147143723,
-0.0740799531,
0.006645578,
0.0080584604,
0.041777961,
0.0322769471,
-0.0237398483,
0.0274126083,
0.0041473522,
0.0012820489,
-0.0458335914,
-0.0070628999,
0.0196803343,
0.0317530669,
0.0025797994,
0.0086377542,
0.0065810573,
0.0447923467,
0.0466982052,
0.0031820331,
-0.0187977087,
0.0087835612,
-0.0103232348,
-0.029846102,
0.0234898981,
0.0038436905,
0.0155192893,
-0.0495839827,
-0.0056565003,
0.0208220929,
0.0225692783,
0.0397801809,
-0.0190455727,
-0.0607413054,
-0.0322432145,
-0.0297851171,
0.0324111581,
-0.0159901995,
0.0116759269,
-0.0644540563,
0.0291739423,
-0.0129997768,
0.0080229267,
0.0154771954,
-0.0037073018,
0.0297745243,
-0.0143603459,
-0.0151121607,
-0.0113096945,
-0.0022191573,
-0.0493917726,
-0.0118112005,
-0.043551784,
0.024250783,
-0.033859808,
0.0252598971,
-0.0026558728,
-0.0110255657,
0.0089079216,
0.0165089816,
-0.1087731719,
-0.1111890748,
0.0296216868,
-0.0050897254,
-0.0165313724,
-0.0190834589,
-0.0403384604,
0.0543542579,
0.0147306705,
0.0181377493,
-0.0271401722,
0.0079079801,
0.0109706707,
-0.0216799546,
0.101115413,
-0.0362309963,
-0.0084405718,
0.0491935015,
0.0551267639,
0.0884148553,
0.0325687751,
0.0021165519,
-0.0048044431,
0.0600157194,
-0.0151069714,
-0.0218715612,
0.0089512803,
-0.0166241806,
-0.0366052873,
0.000278416,
-0.0017084716,
-0.0327519961,
0.0281089786,
0.0062575266,
0.0177073088,
0.0117029976,
0.0216431208,
-0.0110846646,
0.0452666543,
0.0381634384,
0.0046887351,
-0.0244506989,
0.0213351715,
0.016814962,
-0.0059398683,
0.0175897647,
-0.0318899862,
-0.0403660424,
0.0001253076,
0.0392908566,
0.041936934,
-0.0176475514,
0.025626611,
-0.009884865,
-0.0073084263,
-0.0160029531,
-0.0254898351,
0.0491187535,
0.0031687752,
-0.0291506276,
0.0181384813,
0.0564241633,
-0.0111228181,
0.0306098517,
-0.0025950805,
0.0416946933,
0.0266795885,
0.0771184415,
-0.019023532,
-0.0048201615,
0.0082974127,
-0.0239313971,
0.0800759867,
-0.0172290858,
-0.0384048745,
-0.0235603955,
-0.0389334187,
0.0464705341,
0.0683209151,
0.0077318726,
-0.0006928837,
0.0384382792,
0.0533228703,
0.0299980734,
-0.005426486,
-0.0247647464,
0.0468016751,
0.0201402549,
-0.0052351807,
-0.0018325615,
0.0210277662,
-0.00334647,
0.0319965966,
0.0120789502,
-0.0387087911,
0.0065538441,
-0.0062523531,
0.0266751964,
0.0007119828,
0.0237393398,
-0.0647334605,
-0.0218828022,
0.007676092,
0.0047119246,
0.0005340663,
0.0144263292,
-0.0593820736,
-0.0243853964,
0.0436552912,
0.0642215461,
-0.0308345594,
0.0671280399,
-0.015588399,
-0.003761797,
-0.0604468584,
0.0438579246,
-0.031083541,
0.0059550935,
0.0006525607,
0.0123208016,
-0.0066720163,
-0.0195298102,
-0.0527611598,
0.0463458039,
0.0144399684,
-0.0162027124,
-0.0417106524,
-0.0169833545,
-0.0461431891,
-0.0137070185,
-0.0909775421,
-0.028538473,
-0.0082654553,
0.0205844939,
0.027972525,
0.0825283304,
-0.0344466455,
-0.0140848495,
0.0160626043,
0.0174450837,
0.0425512753,
-0.003215173,
0.0183956996,
-0.0182612315,
0.0017769997,
0.0035882574,
0.0058562956,
-0.0172764044,
0.0234258324,
0.0263569523,
-0.0542591773,
-0.0283060186,
-0.0076304814,
0.0061250427,
-0.0318414085,
0.0119834095,
0.0578891858,
-0.0356838442,
-0.0097223008,
0.0009847156,
0.0383987352,
0.0088367667,
0.0343531705,
-0.0216746889,
-0.0215558074,
-0.0062535666,
0.0097744782,
-0.046024438,
0.0751989558,
-0.0005046786,
-0.0193153769,
0.031259466,
0.0388881303,
-0.0298245549,
0.0364232287,
-0.0205911472,
-0.0364281423,
-0.0264302567,
-0.0208745524,
-0.0625595823,
0.0503053367,
-0.042457059,
0.0333504193,
-0.0122622736,
-0.0177195836,
0.0374553278,
-0.0531793013,
-0.0032170082,
-0.0187540911,
0.0429132581,
-0.0354964659,
0.0125321206,
-0.0125832167,
0.0218570363,
-0.0257422999,
-0.0650350749,
0.0267188698,
0.0445157774,
-0.0172075015,
0.0003898046,
-0.0032646686,
-0.0045692436,
0.001338392,
-0.007815077,
0.0122480216,
-0.0138476817,
-0.0150151979,
-0.0110564791,
0.0073436224,
0.0381221659,
-0.0058128312,
0.0382977501,
0.115224503,
-0.0225759111,
-0.0216016471,
0.0106206043,
-0.0058757421,
0.0030427151,
0.0338279568,
0.0018144714,
0.0023158684,
0.0582263283,
0.0398598947,
-0.0495722592,
-0.0240561347,
0.0056360047,
-0.0613689534,
0.0244147666,
-0.0375444964,
0.0092604998,
-0.1057058722,
0.0190417618,
-0.0185008962,
-0.0181524828,
-0.0070315958,
-0.0352653898,
0.0848714635,
0.0265151169,
0.0181983747,
-0.0266730599,
-0.0333832689,
0.0120666064,
-0.0460137129,
0.0123210847,
-0.0261289757,
-0.0332120806,
0.0205194764,
-0.0056455364,
0.0249778125,
-0.0585816242,
0.003090973,
-0.0912834406,
-0.0238922965,
-0.0359876044,
-0.0448680557,
-0.0060581127,
-0.0480986498,
-0.0257568639,
-0.0151611473,
-0.000884579,
0.0305279996,
-0.0397216678,
-0.0174393486,
0.0367297456,
-0.0260599609,
-0.0535233766,
0.0156466514,
-0.0687691271,
0.0264190212,
0.0348918289,
0.0227573197,
-0.0285358727,
0.0379539393,
-0.0612252839,
-0.0263447873,
-0.0335117392,
-0.0141265197,
0.0009219981,
-0.0536451638,
-0.0086478721,
-0.0193524156,
-0.0476725623,
0.01705366,
0.0127825132,
0.0157153644,
0.027642794,
0.0060389545,
-0.0048254281,
-0.0391259082,
-0.0205765869,
0.0178721491,
0.0374659374,
-0.0039584478,
-0.0058453446,
-0.0349366367,
0.0079859523,
-0.0257924665,
-0.0114834523,
-0.0251582358,
0.0208047125,
-0.0618179105,
-0.0925179422,
-0.0188602637,
-0.1016515791,
0.0605359934,
0.0513246134,
0.0054268125,
0.0294620134,
0.0610071197,
0.012417566,
-0.0538071021,
-0.0134967556,
-0.057151027,
-0.0151606612,
-0.072624512,
0.0231250301,
0.0063300375,
0.0218688045,
-0.0100589581,
-0.0250254534,
-0.0312074926,
-0.0288961306,
0.0461418703,
-0.0222931579,
0.0076058679,
-0.0095708771,
-0.0672498643,
-0.0472252294,
0.064346984,
-0.0075717997,
-0.0340803266,
0.0059146862,
-0.0004145933,
0.0250293426,
-0.0133927101,
0.0261666253,
-0.0112361368,
0.019193355,
-0.0094012888,
0.0127521623,
-0.006245947,
0.0388435274,
-0.0261211079,
0.0114862844,
0.0620583259,
-0.0479745828,
-0.0444553755,
-0.0580375083,
0.00423253,
0.0360202044,
-0.006359308,
0.0242386088,
-0.0254704896,
0.0401925072,
0.0246512424,
-0.0003108802,
0.0394864529,
0.0053157839,
-0.003168321,
0.0202994328,
0.0461583734,
0.0367157646,
0.0563358255,
-7.823e-7,
-0.0130187692,
0.0297900159,
0.0389135778,
0.0290105622,
0.0142431334,
-0.0210959036,
-0.0139492191,
-0.0187276434,
-0.0075024725,
-0.0011614875,
0.0497403741,
0.0109001668,
0.070932366,
-0.0291538071,
0.058270406,
0.0553170815,
-0.0244570151,
0.0222074445,
-0.0049699126,
0.0087503931,
-0.0319308974,
-0.0192764625,
-0.0272252206,
-0.0072877533,
0.0325689949,
-0.0102233253,
-0.0201761443,
0.0062263636,
0.0097447084,
0.0320347361,
0.0010852698,
0.0389677994,
-0.0224144701,
-0.0244122334,
-0.0327427164,
0.0327885374,
0.0598839186,
-0.0160425212,
-0.001677028,
-0.0902071968,
-0.0000353157,
0.0651914105,
-0.0446032509,
-0.0149771022,
0.033948388,
0.0001518217,
0.0110445451,
0.0575205237,
0.036353942,
-0.0567148924,
-0.0158280898,
-0.042798914,
-0.0193449091,
0.024381252,
-5.393540934e-33,
-0.0527913235,
-0.0343395732,
0.0228505544,
-0.0648757815,
0.0209493693,
0.0294422954,
-0.017335996,
0.0153771322,
-0.0090416009,
-0.044450108,
-0.053757634,
-0.0005857061,
0.0292503219,
0.0116559705,
0.0199704207,
-0.0495517962,
0.0175176859,
0.0768293664,
-0.0281457249,
0.0235376954,
0.0713091344,
0.0848599598,
0.0476159155,
-0.0394652225,
0.0370819494,
-0.0174089614,
-0.0062949443,
-0.0235319715,
0.0724416077,
0.0489733592,
-0.0493917912,
0.0016972001,
0.0294281561,
-0.013462347,
-0.0047526713,
0.0358320624,
-0.1297563463,
0.0181407761,
0.0104834763,
0.0671534464,
-0.0216958579,
0.0104331337,
0.0042233095,
-0.0037369654,
-0.0012102462,
0.0588340238,
-0.0023025814,
0.0142213693,
-0.0338377357,
0.0129218902,
-0.0632860065,
0.0711192265,
0.0333634652,
0.0781821534,
0.0833566859,
-0.0838195756,
0.0225762706,
0.0107031073,
-0.0769586191,
-0.0158120338,
0.0539774857,
-0.0013870908,
0.0464212187,
0.0507910326,
-0.0244679358,
0.0182348564,
0.0050234259,
-0.0770726502,
0.004271022,
0.0431497619,
0.0393166803,
0.0034770756,
-0.0172220618,
0.0777435899,
0.0494815595,
-0.0675669461,
-0.0223508179,
0.0382662676,
0.0322184637,
0.0118876891,
-0.0021766184,
-0.0291145593,
-0.1093366966,
-0.030919699,
0.0174155086,
-0.0026779994,
-0.0324063338,
-0.0126267197,
0.0717552081,
-0.000713254,
-0.0743581802,
0.066911757,
0.0024417075,
-0.011310216,
-0.0084388489,
-0.0041281036,
0.0183816943,
0.0354776271,
-0.0087417904,
0.0154617438,
-0.0786289126,
-0.0577903949,
0.0051189954,
-0.0100975921,
-0.0608058982,
-0.0017282614,
0.0387661308,
-0.0511986092,
-0.047943648,
0.0244951155,
0.0111578889,
0.0103073511,
-0.0394292958,
0.0163948368,
-0.0107228579,
-0.0583412461,
-0.0000846466,
-0.0245855059,
0.0951879546,
0.0186979529,
0.0166505352,
-0.0145484218,
0.0079711452,
-0.0152330659,
0.0068221856,
0.0165706687,
0.0156776607,
-0.0263902266,
0.0130502582,
-0.0438625291,
0.025261756,
-0.0238579959,
2.356e-7,
0.0096726222,
0.0240833312,
0.0425684862,
-0.0400101095,
-0.0336674526,
-0.0155451456,
0.0326642767,
0.0393212959,
-0.0742899477,
0.0290702395,
0.0258727502,
-0.017510213,
-0.0076462477,
0.0144216307,
-0.1369778812,
-0.0123145757,
-0.0301245004,
-0.0148373712,
-0.0641533136,
0.005603136,
-0.0355609283,
0.0523768291,
0.0079630064,
-0.0233749561,
-0.0344744921,
-0.0111444741,
0.0176275466,
0.0215424616,
0.0334649123,
0.0272904076,
0.044622194,
0.0886138901,
-0.0087549472,
-0.0044914382,
0.0129865753,
0.0130430339,
0.0285566691,
0.0605655722,
-0.0254117046,
0.0315741338,
0.0152034247,
-0.0519467965,
0.0190969221,
-0.0258344859,
0.0899410322,
0.0194382221,
-0.0115160216,
0.0306307189,
-0.0161756203,
-0.0069978037,
-0.004809801,
0.0232630204,
-0.0422551818,
-0.0236944351,
0.0347530916,
-0.0723627433,
-0.0146815414,
-0.0841950998,
0.0330680721,
0.0343234725,
0.01806923,
-0.0106804157,
-0.0162807293,
0.0087685566,
0.0359178968,
-0.0780809447,
0.0021631331,
2.290177073e-34,
-0.0135253547,
0.070250839,
-0.0258966293,
0.0240605194,
-0.0014180285,
-0.0048172208,
-0.0128502902,
-0.0001498821,
-0.0127491299,
-0.0712104887,
0.0195164289
] |
http://arxiv.org/abs/2402.01761v1 | http://arxiv.org/pdf/2402.01761v1 | Rethinking Interpretability in the Era of Large Language Models | 2024-01-30T17:38:54 | 2024-01-30T17:38:54 | Chandan Singh, Jeevana Priya Inala, Michel Galley, Rich Caruana, Jianfeng Gao | cs.CL | cs.CL, cs.AI, cs.LG | Interpretable machine learning has exploded as an area of interest over the
last decade, sparked by the rise of increasingly large datasets and deep neural
networks. Simultaneously, large language models (LLMs) have demonstrated
remarkable capabilities across a wide array of tasks, offering a chance to
rethink opportunities in interpretable machine learning. Notably, the
capability to explain in natural language allows LLMs to expand the scale and
complexity of patterns that can be given to a human. However, these new
capabilities raise new challenges, such as hallucinated explanations and
immense computational costs.
In this position paper, we start by reviewing existing methods to evaluate
the emerging field of LLM interpretation (both interpreting LLMs and using LLMs
for explanation). We contend that, despite their limitations, LLMs hold the
opportunity to redefine interpretability with a more ambitious scope across
many applications, including in auditing LLMs themselves. We highlight two
emerging research priorities for LLM interpretation: using LLMs to directly
analyze new datasets and to generate interactive explanations. | Rethinking Interpretability in the Era of Large Language Models: Interpretable machine learning has exploded as an area of interest over the
last decade, sparked by the rise of increasingly large datasets and deep neural
networks. Simultaneously, large language models (LLMs) have demonstrated
remarkable capabilities across a wide array of tasks, offering a chance to
rethink opportunities in interpretable machine learning. Notably, the
capability to explain in natural language allows LLMs to expand the scale and
complexity of patterns that can be given to a human. However, these new
capabilities raise new challenges, such as hallucinated explanations and
immense computational costs.
In this position paper, we start by reviewing existing methods to evaluate
the emerging field of LLM interpretation (both interpreting LLMs and using LLMs
for explanation). We contend that, despite their limitations, LLMs hold the
opportunity to redefine interpretability with a more ambitious scope across
many applications, including in auditing LLMs themselves. We highlight two
emerging research priorities for LLM interpretation: using LLMs to directly
analyze new datasets and to generate interactive explanations. | [
0.017180739,
0.0801613331,
-0.0240182821,
0.0712525547,
-0.011390035,
0.027108049,
-0.004102027,
0.0120650278,
-0.0112857912,
-0.0009137728,
-0.01711604,
-0.0178178865,
0.0017224554,
0.0424459614,
0.0198800135,
-0.0249343216,
0.0044176574,
-0.056018874,
-0.0067007705,
0.0232518744,
-0.0210051071,
0.0128424326,
0.0186561011,
0.0635646433,
0.0180040412,
-0.0281939805,
-0.033921022,
-0.0031240047,
-0.0117277438,
0.0121618053,
-0.0643384531,
0.0226271134,
0.0539429262,
0.0950634629,
0.0000019797,
-0.0531527996,
-0.0077636186,
0.0036343026,
-0.0996050984,
0.0512999818,
0.0187218878,
-0.0558888577,
0.0163910706,
-0.0309783295,
0.0041013132,
0.0176689699,
0.0490296856,
0.030245699,
0.0386242196,
0.0451695807,
0.012094765,
-0.0372636914,
0.0013401728,
-0.0150983389,
0.020665355,
-0.0198531915,
0.0127425538,
0.0058763251,
-0.0249924418,
0.0055675213,
-0.0457240008,
0.021995455,
0.0501943827,
-0.0230251383,
0.0024832007,
-0.0033739985,
-0.00972238,
-0.0284157526,
0.0277888607,
0.058766067,
-0.0022352384,
-0.0189920161,
-0.0070298566,
-0.0143362377,
-0.0010506628,
0.023066042,
-0.0236194171,
0.0149973445,
0.0126429982,
0.0423620678,
0.0273643136,
0.002691342,
0.0991737172,
0.0398770384,
-0.0566515401,
0.0681103021,
-0.0343848914,
0.0068462258,
-0.0177180916,
0.0191333648,
-0.0052572247,
-0.0126757603,
0.0100489063,
0.0234832671,
-0.0280277338,
0.0337472111,
0.0088471035,
0.0262137149,
0.0298274234,
-0.0335562862,
0.0687616542,
-0.0073809456,
-0.0524599738,
0.057971511,
-0.0097457981,
0.0492199846,
-0.0401317962,
0.0322619975,
-0.023293592,
-0.0167872775,
-0.0439587645,
-0.042700462,
0.0298993867,
-0.01827682,
-0.0340485014,
0.0055910591,
-0.0269539859,
0.0057425876,
-0.0451003686,
0.0112596061,
-0.0042600925,
-0.0100273397,
-0.0095598791,
-0.0342674889,
-0.0403636247,
-0.0072811916,
-0.016966749,
0.0254226755,
-0.0089283446,
-0.0674313009,
-0.0275866836,
0.0286982935,
-0.0141604608,
-0.0016176432,
-0.0102876751,
0.0756910145,
0.04971141,
-0.0023642057,
-0.0141781652,
-0.0584547333,
-0.0418972336,
-0.0496890098,
-0.0094624171,
-0.0265903864,
0.0150376623,
0.0011887997,
-0.0043520466,
-0.0550306626,
-0.0110971816,
-0.0172189102,
-0.0372458436,
0.0697283,
-0.030350402,
0.0173930041,
-0.0373785198,
0.0173335336,
-0.0217997525,
0.0544539988,
0.0289506763,
0.0279493183,
0.037617404,
-0.0194999296,
0.0512424856,
0.0164499041,
0.0222903788,
0.0174390636,
-0.0320219733,
-0.0202232655,
-0.0357714556,
0.0146157071,
0.0395064242,
0.0093571218,
-0.0106164161,
-0.0256724041,
0.0353070386,
0.0346977152,
0.0312881656,
0.0173832066,
0.0420867987,
0.0097647281,
-0.0029317588,
0.0068953237,
-0.0549316444,
0.0016831602,
-0.0115725677,
-0.0210600775,
-0.007368254,
-0.0270467736,
0.0123181222,
-0.0841282308,
-0.0072854608,
0.0047059981,
0.0037935495,
0.0281815268,
0.0773456469,
-0.009624742,
-0.0144721195,
0.0302432943,
0.0358808897,
-0.0316493735,
0.0140907457,
-0.08660312,
0.048067309,
0.102391623,
0.0243236423,
-0.0232467595,
0.0411604531,
-0.004811896,
0.012057405,
0.011292411,
-0.0098397527,
0.0215810891,
0.0630494878,
-0.0333085731,
-0.0184563547,
-0.0311630145,
0.053608872,
0.0273475032,
-0.047196392,
-0.0436462462,
0.0062957113,
-0.0540287308,
-0.064793095,
0.0261569768,
-0.010675854,
-0.0244460274,
0.0012857803,
0.0057509062,
0.016157778,
0.0296763796,
-0.0011179477,
0.006435921,
0.024349371,
-0.0250843596,
-0.0009280933,
0.0228746142,
-0.0001423385,
0.0032766222,
-0.0188784823,
-0.0306741893,
-0.0374113582,
0.0114918593,
-0.0482724085,
-0.0191617999,
0.0145494593,
-0.0598741099,
0.0483486354,
-0.0224282276,
-0.002350407,
-0.0324127786,
0.047757104,
-0.0227082409,
0.0551854298,
-0.0262400191,
-0.0105813425,
0.0259397961,
-0.0039555975,
0.012827958,
-0.0192061365,
-0.0083835302,
0.0203635134,
0.0056723855,
-0.0208714809,
-0.0132181691,
-0.0181864444,
-0.0309587102,
-0.0002859909,
-0.0220260397,
-0.0209089369,
-0.0253361203,
0.0131964236,
0.0176767837,
-0.0369662344,
0.0100607332,
0.0026708446,
0.0092724152,
-0.0048536127,
0.0189476032,
0.0044359975,
0.033713188,
-0.0057663471,
-0.0625691935,
-0.0138848787,
-0.0234763715,
-0.0240022372,
0.0941130072,
0.0252834186,
0.0005966374,
-0.0857864246,
0.0073238141,
0.0048732939,
-0.026637407,
-0.0273563582,
-0.0230032727,
0.014539022,
-0.0158154666,
-0.04432391,
0.0545286238,
-0.0363486558,
0.1184190959,
0.0310555119,
-0.0016928023,
-0.0600378178,
-0.052616518,
-0.0013718159,
-0.0163745824,
-0.0499160588,
-0.0752612874,
-0.0055163214,
0.0735741779,
-0.0021704421,
0.0001229225,
-0.0454695038,
0.0277233273,
-0.0202719979,
-0.0610153116,
-0.0379073657,
-0.0298937596,
-0.0032228592,
0.0884622782,
-0.0071119047,
-0.08836326,
0.0522995181,
-0.034235172,
0.0014350603,
0.019042192,
-0.0586516485,
-0.0609563813,
0.1000095904,
0.0003174992,
-0.0066253846,
-0.00974085,
-0.0341013558,
0.0046245977,
-0.016556358,
-0.011073702,
-0.0144038266,
0.0657594353,
-0.0640406832,
-0.0176166706,
0.0195549112,
0.0175968986,
-0.0243701395,
-0.0393909328,
-0.0435256176,
-0.0136917951,
0.0180705786,
-0.0196778979,
-0.042060256,
-0.0154838581,
0.0112361219,
-0.032550849,
0.0004703487,
-0.003130286,
-0.0326815173,
-0.0232493021,
0.0085708061,
0.0401108935,
0.0458477475,
-0.0484427325,
0.0054966784,
0.0448446684,
-0.0144187668,
0.0092926631,
-0.0474575683,
0.0379190221,
-0.0024776948,
0.0029585462,
-0.0063024783,
-0.0026914128,
0.0474603064,
-0.1103995219,
-0.0071597956,
0.0963844955,
0.0105052907,
-0.0257036891,
0.0521222427,
-0.0104933223,
0.0126498705,
-0.0205393452,
0.000888647,
-0.0719075352,
-0.0315687396,
0.0105202561,
0.035259176,
-0.0217613243,
-0.030147396,
0.0047385697,
-0.0570983551,
0.0514979474,
-0.0001317141,
-0.0250669718,
-0.044449158,
-0.0557908863,
-0.0353498384,
-0.0435012132,
0.0180024337,
0.0029739644,
0.0269278791,
0.0190307144,
0.0197740775,
-0.0250942856,
-0.0397025868,
0.0409362242,
-0.0108192544,
0.0405735448,
0.0556592532,
0.0286262017,
-0.0253022183,
-0.0272650383,
-0.0176325496,
0.0126454309,
0.0190333277,
0.0007339279,
0.0440783687,
0.0314225778,
-0.0377947465,
-0.0362926088,
0.0428415947,
-0.0478727892,
-0.0274222046,
-0.0027368555,
0.0637326315,
-0.0382461324,
-0.0266589038,
0.0091632428,
-0.0049940576,
0.0087816101,
0.1079306006,
0.0552125759,
-0.0161577482,
-0.0416624546,
-0.0502208211,
0.00009288,
-0.0266830735,
-0.0688653663,
0.0466843247,
0.0334675312,
-0.0204649605,
-0.0002381741,
-0.0421900526,
-0.0018195955,
0.0054246583,
-0.0017107112,
0.0107465759,
-0.0208266005,
-0.012101246,
-0.0284217726,
0.076070331,
-0.0093652429,
0.0168140382,
-0.0158142392,
-0.0366651155,
0.0089224698,
-0.0312785096,
-0.0274595395,
0.034147881,
-0.0547262467,
-0.0171057582,
0.0268451888,
0.0060689338,
0.0354703814,
0.0604505986,
-0.0026901998,
-0.0633177385,
0.012313433,
0.0177865494,
-0.0175132938,
0.0101338169,
0.0348128341,
0.0346500762,
0.0107043087,
0.0641991124,
-0.0304153245,
-0.0027180086,
-0.0094385026,
-0.0038210866,
-0.0498780794,
-0.0039566969,
0.0082372278,
-0.0035687506,
0.0289309062,
0.00212904,
-0.0373600721,
-0.0063885618,
0.0034002068,
-0.0206559524,
0.0074068564,
-0.0594339333,
0.0028490333,
0.0297713578,
-0.0375233553,
-0.0218628831,
-0.0298021138,
-0.0233133193,
-0.0107206972,
-0.0475031212,
0.0603366122,
0.0112189148,
0.0103560248,
0.003975078,
0.0119327111,
-0.0091942865,
0.0486321226,
-0.04488777,
0.0442874581,
0.0394490212,
0.042059809,
0.0129065635,
0.0106409276,
-0.0367734209,
0.0052604787,
-0.0754640922,
-0.0861660391,
0.0500707366,
0.0087381219,
0.0038673785,
-0.013713412,
0.0542098731,
-0.0178071428,
0.0087608248,
-0.0052234852,
0.0732459128,
0.0162834469,
-0.0278088301,
0.0137146823,
0.0212625619,
-0.0422513895,
-0.0179223586,
0.0733462349,
0.0265343674,
0.0267935451,
-0.0230870489,
-0.0226719622,
-0.0107522663,
-0.0545015037,
0.0120406579,
0.0091952672,
-0.0233489983,
0.0035800121,
-0.0666310564,
-0.0671602786,
-0.0107616549,
0.000256067,
0.0047810031,
-0.0257859938,
-0.0273898896,
0.0597897843,
-0.0055406839,
-0.0229555536,
-0.0440310761,
0.0259862542,
-0.0090356292,
-0.0016940302,
-0.0053115985,
-5.905798964e-33,
0.0157027617,
-0.0567989759,
-0.0097013814,
-0.0452417769,
-0.001178649,
0.0172539987,
-0.0585082136,
-0.005659509,
-0.0044749426,
0.0481706746,
0.0357748903,
0.0378328785,
0.0138254641,
0.037821278,
0.0067980681,
-0.0615878776,
0.0420435071,
0.05017801,
0.0274822693,
0.0245631859,
0.0668296367,
0.0217486657,
0.051843375,
-0.0845757723,
0.1034053713,
0.0086245602,
-0.0104110502,
0.0179448575,
-0.0370907374,
0.0688677356,
-0.0205401108,
0.0374874473,
0.0262074415,
-0.0518984608,
-0.0181625243,
-0.0005704054,
-0.0317797661,
0.0172133241,
0.0502791554,
0.0553772375,
-0.0467539392,
-0.0527633652,
0.0703976154,
-0.0034394786,
0.0123092383,
-0.0049894401,
0.0124797421,
-0.0378477946,
-0.0027179203,
0.025367517,
-0.0291692484,
0.044187583,
-0.0533721894,
0.0016041604,
0.048351299,
0.0249036923,
-0.018487053,
0.0452248417,
-0.0607543178,
-0.0023312953,
0.0345562585,
-0.0230305772,
0.087012507,
0.0121528078,
-0.0035528648,
0.0593500175,
0.0568653978,
-0.033844918,
-0.0730005726,
-0.0068228617,
0.0244441219,
0.0689590722,
0.0440889522,
0.0241196342,
0.0159883536,
0.0421214588,
-0.0708059445,
-0.0028338183,
-0.0267533064,
0.0140348049,
0.0540119819,
0.0123757105,
-0.0737125799,
-0.0214998424,
-0.0306089632,
-0.0432066843,
-0.0265579354,
-0.012259949,
0.0168652181,
-0.0489298254,
-0.0582887158,
0.0114374831,
0.0360510796,
-0.0776410773,
-0.011990984,
0.0150123006,
0.0112689389,
0.0003914077,
0.0240039993,
-0.0092337569,
0.0120904017,
-0.0176405255,
0.0040786122,
0.0052406592,
0.0680359602,
-0.0035090644,
0.0292596724,
-0.0140632251,
0.0028356796,
0.0006834495,
-0.0530520603,
-0.0336916521,
-0.009806796,
0.0137211429,
0.0634506866,
-0.0003636506,
-0.0328398682,
0.0220779795,
0.0171919186,
0.0328679085,
0.0290242862,
0.0339782834,
-0.0376609005,
-0.0018461725,
-0.0380640775,
-0.0458984599,
-0.0112872617,
-0.0183560625,
0.0713004246,
-0.0561228953,
-0.0076056649,
-0.0038003195,
2.491e-7,
0.0632766485,
0.0383489057,
0.0116542568,
0.0174194854,
0.0444711335,
0.0353925973,
0.0034138986,
-0.0085548526,
-0.0273161437,
-0.033133626,
0.0063463706,
0.0482751653,
0.0622563325,
-0.002213279,
0.0009804266,
-0.0002382901,
-0.0541859046,
-0.0144240623,
-0.0599240549,
0.0260391794,
0.070407033,
0.1055538282,
0.0098326979,
-0.0058921413,
0.0653555542,
-0.0209965687,
0.0239874553,
-0.0806350634,
0.0226676278,
-0.0104675889,
0.0410721228,
-0.0083276788,
0.0315554962,
0.0068853702,
0.0121644977,
0.0050743269,
-0.0119020622,
0.0723459646,
0.0039616311,
0.0019063238,
-0.0480746888,
-0.0281057451,
0.0196012985,
-0.0719901025,
0.0784152746,
-0.032527294,
-0.0711854324,
-0.0049399091,
0.022345528,
-0.0610891245,
0.0187434237,
0.0147412475,
0.0329985507,
0.0275573041,
-0.0150866983,
-0.0535371751,
0.0023182591,
-0.0470510945,
-0.0137839802,
-0.0220141578,
-0.0521749891,
0.0105201621,
-0.0389701426,
-0.0015624819,
0.0479200706,
-0.0490646772,
-0.0085002417,
2.335689367e-34,
0.0218299162,
0.0137695698,
0.019694306,
0.0228630602,
0.0499271974,
0.0330324918,
-0.0004853107,
0.0267337952,
0.0337308533,
-0.0758534744,
-0.0054608593
] |
http://arxiv.org/abs/2401.17196v1 | http://arxiv.org/pdf/2401.17196v1 | Single Word Change is All You Need: Designing Attacks and Defenses for Text Classifiers | 2024-01-30T17:30:44 | 2024-01-30T17:30:44 | Lei Xu, Sarah Alnegheimish, Laure Berti-Equille, Alfredo Cuesta-Infante, Kalyan Veeramachaneni | cs.CL | cs.CL | In text classification, creating an adversarial example means subtly
perturbing a few words in a sentence without changing its meaning, causing it
to be misclassified by a classifier. A concerning observation is that a
significant portion of adversarial examples generated by existing methods
change only one word. This single-word perturbation vulnerability represents a
significant weakness in classifiers, which malicious users can exploit to
efficiently create a multitude of adversarial examples. This paper studies this
problem and makes the following key contributions: (1) We introduce a novel
metric \r{ho} to quantitatively assess a classifier's robustness against
single-word perturbation. (2) We present the SP-Attack, designed to exploit the
single-word perturbation vulnerability, achieving a higher attack success rate,
better preserving sentence meaning, while reducing computation costs compared
to state-of-the-art adversarial methods. (3) We propose SP-Defense, which aims
to improve \r{ho} by applying data augmentation in learning. Experimental
results on 4 datasets and BERT and distilBERT classifiers show that SP-Defense
improves \r{ho} by 14.6% and 13.9% and decreases the attack success rate of
SP-Attack by 30.4% and 21.2% on two classifiers respectively, and decreases the
attack success rate of existing attack methods that involve multiple-word
perturbations. | Single Word Change is All You Need: Designing Attacks and Defenses for Text Classifiers: In text classification, creating an adversarial example means subtly
perturbing a few words in a sentence without changing its meaning, causing it
to be misclassified by a classifier. A concerning observation is that a
significant portion of adversarial examples generated by existing methods
change only one word. This single-word perturbation vulnerability represents a
significant weakness in classifiers, which malicious users can exploit to
efficiently create a multitude of adversarial examples. This paper studies this
problem and makes the following key contributions: (1) We introduce a novel
metric \r{ho} to quantitatively assess a classifier's robustness against
single-word perturbation. (2) We present the SP-Attack, designed to exploit the
single-word perturbation vulnerability, achieving a higher attack success rate,
better preserving sentence meaning, while reducing computation costs compared
to state-of-the-art adversarial methods. (3) We propose SP-Defense, which aims
to improve \r{ho} by applying data augmentation in learning. Experimental
results on 4 datasets and BERT and distilBERT classifiers show that SP-Defense
improves \r{ho} by 14.6% and 13.9% and decreases the attack success rate of
SP-Attack by 30.4% and 21.2% on two classifiers respectively, and decreases the
attack success rate of existing attack methods that involve multiple-word
perturbations. | [
0.0435671359,
0.0649451837,
-0.008836492,
0.0306147523,
-0.0593628176,
-0.0122323362,
-0.0115962774,
0.0073303976,
-0.0087264096,
-0.0691389441,
0.0211082213,
0.000852021,
0.0316247456,
0.0162108652,
0.0551875345,
-0.0281458907,
0.0702951625,
0.0251859985,
-0.0186556242,
-0.0090828193,
0.0208978411,
-0.0322282948,
-0.0005993632,
0.0151420999,
-0.0043170177,
-0.0083091585,
-0.0155852353,
0.0061995471,
0.0421810783,
-0.0356781408,
0.0404048972,
0.0464797579,
-0.0027588243,
0.1072608829,
0.0000018561,
-0.0437492728,
-0.0440328792,
0.0329092667,
-0.0061454759,
0.0475204699,
0.0501358844,
-0.0378330387,
-0.0036952847,
0.0219907481,
-0.0434666947,
0.0467693768,
0.0322625265,
0.05746736,
0.0478826761,
0.1183796227,
-0.0171722062,
-0.0494073853,
0.0311148725,
-0.0383239165,
0.0032010588,
-0.0411779955,
0.0602532737,
-0.0431542397,
-0.0598833002,
-0.00509727,
-0.03715818,
0.0336857475,
-0.0035522184,
0.0209984016,
-0.0036094422,
0.0170996897,
-0.0336600058,
0.0150880609,
0.0245702546,
0.0534596033,
-0.027263429,
-0.0242127702,
-0.0226576533,
0.0079188682,
0.0485303439,
0.0126108769,
0.0119839134,
-0.0387960523,
0.0230943747,
-0.0031584885,
-0.0307564568,
-0.0367485546,
0.0585164465,
0.0534938723,
-0.0060495338,
0.007775466,
0.030786125,
-0.0357721597,
-0.0310268681,
-0.0490344241,
-0.0104435394,
-0.0220774487,
0.0035349438,
0.0254851431,
0.0119523285,
0.020372143,
-0.1122508869,
-0.0962880328,
0.0542392954,
-0.10906104,
0.0440492667,
-0.0113001456,
-0.0198676828,
0.0159424488,
-0.0260998104,
0.0060054916,
0.0049795946,
0.0489724092,
-0.0006947676,
0.0475347377,
-0.0348609015,
-0.0094360709,
-0.0181553643,
0.0403148122,
0.0063304161,
-0.0196952298,
-0.050168965,
0.0347932205,
-0.0090051079,
0.0295888595,
-0.0685051456,
0.0204910971,
-0.0207403675,
0.0269182995,
0.0289656408,
-0.0209166501,
-0.0157620143,
0.0237633362,
-0.0282975547,
-0.0512273386,
-0.0362452269,
0.0233379323,
-0.0231555272,
-0.0133624543,
0.0171266943,
0.082765311,
0.0128597151,
-0.0632117391,
-0.0157918595,
-0.0451329462,
-0.0573880188,
-0.0057450621,
-0.044402454,
-0.0309004728,
-0.0309464149,
-0.0306241624,
-0.0281794388,
-0.0613887943,
0.0143494811,
0.0096951639,
-0.0475421771,
0.0837609023,
-0.0141331153,
-0.0506120734,
0.0482888184,
0.0254513081,
0.0114296749,
0.039570976,
0.0542261899,
0.0257114489,
0.0838429332,
0.0044286973,
0.0157311168,
0.0252398103,
0.0325334705,
0.0093296813,
-0.0242091194,
-0.0131692067,
-0.009154411,
-0.0151272845,
-0.0362969637,
0.0262281634,
-0.0160944965,
-0.0646109954,
0.1038354337,
0.0106318686,
-0.005569248,
0.0404778235,
-0.030648632,
0.0088296765,
-0.0003521946,
0.0019902845,
-0.043296963,
-0.0274255145,
-0.1025225073,
-0.028299002,
0.0051551796,
-0.0219421554,
-0.0066125849,
-0.0357943624,
0.0004877859,
-0.0083344085,
0.039658349,
-0.0289451703,
0.0405440852,
0.024609426,
-0.0136769684,
0.0321609452,
0.0100008585,
-0.0276513491,
0.0038914431,
-0.0477537401,
-0.0265488401,
0.0225095209,
0.0688730404,
-0.0563594103,
0.01183182,
-0.0489662327,
-0.0167739224,
0.0135174468,
0.0597372055,
-0.0029082315,
0.0246257782,
-0.0155235212,
0.0071745738,
-0.0607920215,
0.0325468518,
0.0242376477,
0.0361691155,
0.0128283277,
0.0137748523,
0.0068959789,
-0.0404943042,
0.0533318073,
0.0565006249,
-0.0578517765,
0.0410879031,
0.0211299974,
0.0277092699,
0.0136947148,
-0.0334074832,
0.0181124844,
0.0195287503,
0.009272093,
-0.0273920428,
-0.0232590456,
0.0405525342,
-0.0412384197,
0.0270898342,
-0.0002210559,
0.011454762,
0.0599902757,
-0.0410236493,
0.0088997856,
0.0355153047,
-0.0548520535,
0.0110183246,
-0.0271647815,
0.0080902288,
-0.047916919,
0.0866531208,
-0.0361672863,
0.0521658026,
-0.011053226,
0.0032613906,
-0.069497861,
0.0241059009,
0.0473067425,
-0.0032022584,
-0.0628508776,
-0.0247772876,
-0.0462733544,
0.0119452672,
0.0252751671,
-0.0277999863,
-0.027608946,
-0.0245768428,
-0.0058891932,
-0.0323711485,
-0.0280588288,
0.0135449795,
0.0085066119,
0.0019797056,
-0.0018189077,
-0.0261854194,
-0.0007071061,
-0.0086658597,
-0.0464517623,
0.026043063,
0.0484190136,
-0.01445687,
-0.0261297096,
-0.0120015619,
0.0101721026,
-0.00070974,
0.0505445637,
-0.0134036662,
-0.0237847771,
-0.0333200172,
0.0085679768,
-0.0153836822,
0.0406669974,
-0.0052853483,
-0.0522264205,
0.0155160064,
0.0294023044,
0.0006383662,
0.0705138072,
0.0642587319,
0.0553309061,
0.0226142891,
0.0368908644,
0.0034372597,
-0.0145091703,
-0.0046949317,
0.012022974,
-0.0345784724,
0.0020727874,
0.0205782168,
0.0124482913,
0.0295424443,
-0.030606091,
0.0020538543,
0.0226460677,
0.0086151576,
-0.0433174521,
-0.0087404605,
-0.013072541,
-0.0447169207,
-0.057477545,
-0.0233633369,
-0.0698380694,
0.082377553,
-0.02198679,
0.0353496112,
-0.0326551162,
-0.0593911558,
0.0329367593,
0.0380807929,
-0.0142657328,
0.0275788456,
-0.0156556629,
0.0055540735,
-0.0288196728,
-0.0503774434,
0.0571616329,
-0.0180882253,
0.0145782176,
-0.0259566754,
0.0020334122,
-0.0004441052,
-0.0337790139,
-0.0545487814,
-0.0818284154,
0.0137767587,
0.0081592621,
0.0048977663,
-0.0400777496,
-0.0630114824,
-0.0484849028,
0.0169459283,
0.0294664893,
-0.0358324759,
-0.0102527328,
-0.0143823763,
0.0330204517,
0.035884399,
-0.0345588997,
-0.01278874,
-0.0260199662,
0.0328278244,
0.0666819811,
0.0038364313,
0.0066615855,
-0.0444566719,
-0.0178880133,
-0.0168986917,
-0.0430040285,
-0.0058918977,
0.0034327968,
0.0141475815,
-0.0306738783,
-0.0151612954,
0.0538227744,
0.0107109789,
0.0139129888,
0.0421008281,
-0.0001601151,
0.0288968068,
0.0044829752,
-0.01053706,
-0.0486715958,
-0.0150321117,
-0.0014011613,
0.0282083862,
-0.02834934,
-0.0135826254,
-0.0357039161,
-0.0488842167,
-0.0229968205,
0.0050286041,
-0.0411714427,
-0.0133940279,
-0.0134396497,
0.0020080833,
0.0071623484,
0.0048174094,
-0.0199196134,
0.0487832017,
-0.0060618911,
0.0058787852,
-0.0266399067,
-0.0593912341,
-0.0188659485,
-0.0164167117,
0.0032771244,
0.0842423439,
0.0299638528,
-0.0442406721,
-0.0286077913,
0.0279257782,
0.0340820737,
-0.0060071526,
0.0096025905,
0.1062025875,
-0.0157194734,
-0.0100434097,
0.0278049167,
0.0379607826,
0.0025251424,
-0.0196751952,
0.0098013552,
-0.0045353766,
0.0119284792,
0.0020538871,
-0.0239222925,
-0.0083973752,
0.0214265957,
0.0488580018,
0.0023587453,
-0.0614315644,
-0.0215467792,
-0.0494853482,
-0.0097687216,
-0.0482296571,
-0.0005579102,
-0.0039800815,
-0.0201935358,
-0.0207351185,
-0.0619145744,
-0.0290069263,
-0.0388388075,
0.0371945314,
0.0095127458,
0.0179500338,
-0.0600850806,
0.0145400502,
-0.0580207407,
-0.028950939,
0.0093026813,
0.0662940443,
-0.0043163821,
-0.0134162651,
0.0245299377,
0.0197573192,
-0.0210152976,
-0.0154916095,
-0.061187461,
-0.0750414208,
0.0266120415,
0.0208584219,
0.0119948154,
0.0268010423,
-0.0539129972,
-0.0816466138,
0.0807702243,
0.0700239316,
0.0048534591,
0.02416192,
0.0857644528,
-0.0247477423,
0.0059316037,
0.0354196392,
-0.01619686,
-0.0143201305,
0.0630161688,
-0.0127666984,
-0.0465136729,
0.0107187852,
-0.0121052749,
0.0274965074,
0.0675787553,
-0.0210777484,
0.0131337158,
0.0477369279,
0.031640064,
-0.0095228776,
-0.0255027264,
0.008132983,
0.0145851877,
0.0245994572,
-0.0241625533,
-0.0276850685,
0.0253481455,
-0.0048540412,
-0.0136633897,
-0.0106582996,
-0.0087307496,
0.0238471758,
0.045612976,
0.01322552,
0.0214494038,
-0.0238718074,
0.0141652375,
0.0329110473,
0.0067836354,
-0.0056726723,
0.0290921386,
0.0309553687,
0.0074466257,
-0.0324872956,
0.007195334,
-0.0747071877,
0.0399334729,
-0.0436314084,
0.0475370735,
0.0005030319,
0.0178745221,
-0.0001184724,
0.0022570712,
-0.0328500792,
-0.0208469331,
-0.0144571662,
-0.0184668191,
-0.0442583226,
-0.0043608896,
0.0054600053,
-0.0098872539,
0.0664655119,
0.0513652302,
-0.0091293892,
0.0470590703,
-0.0035852126,
-0.0488962866,
-0.0499571413,
0.0015355445,
0.0721328855,
0.0183084588,
-0.0165813714,
0.0080054663,
-0.0467104092,
-0.0647746176,
0.0343614034,
0.0182247274,
0.0789031163,
0.0335820541,
-0.0091562057,
0.0540540442,
-0.0170851406,
0.0191968586,
-0.0631910786,
0.0264704749,
0.0429870002,
-0.0410759859,
0.0278078653,
-5.541027638e-33,
0.0101401825,
-0.0172370132,
-0.0328318961,
0.0028231842,
0.0168286413,
-0.0131937852,
0.0075926436,
-0.0024509167,
-0.0296055563,
0.0205566213,
-0.0007761439,
0.0301043093,
-0.0196440332,
0.0190436095,
-0.0308561902,
-0.0473756157,
0.0367088541,
-0.0016945535,
-0.0628670454,
-0.0109119238,
0.0619433969,
0.0155512234,
0.0487048216,
-0.0693488866,
0.0599653013,
0.0117846178,
-0.0782037377,
-0.0188848339,
-0.0310254227,
0.0195589829,
0.0068520275,
0.0416352898,
0.011549714,
0.0335188061,
-0.0050309482,
0.0111061241,
-0.0233820491,
-0.0394215174,
-0.018039003,
0.0290053636,
-0.0207964908,
-0.0172216147,
0.0458087921,
-0.0091047501,
-0.0085120192,
0.0162051991,
0.01726784,
-0.035495121,
-0.0382100679,
-0.0138195008,
-0.0473115034,
0.0390969999,
-0.0209244322,
-0.0238939114,
-0.0152229061,
-0.0050589819,
0.0342372023,
-0.0147993881,
-0.0641867518,
0.0076968502,
0.0633004308,
0.0207041875,
0.0089599928,
0.0080883428,
-0.0105131269,
0.0684663802,
0.0347066931,
-0.0157351773,
0.0297040455,
-0.0098431678,
0.0371048823,
0.0108474884,
0.0776332915,
0.1057612598,
0.0699843466,
-0.0467782207,
-0.0801388994,
0.0012135274,
0.0122455694,
-0.010952468,
0.046337001,
0.0076646409,
-0.0532385223,
0.0047006612,
-0.0313377492,
-0.0506814495,
-0.034975525,
-0.0251466203,
0.0061809607,
0.038491752,
-0.0367526487,
0.0339417234,
-0.0333453007,
-0.0596792251,
-0.0016336433,
0.0098603712,
-0.0371871479,
-0.0332591869,
0.0096319616,
-0.0192019325,
0.0193754472,
0.0100775668,
0.0209821686,
-0.0088880938,
0.0233366806,
-0.0101079168,
0.1067627966,
0.0253014788,
-0.0621951036,
-0.003013958,
0.0334901847,
-0.0048746378,
0.0308265779,
0.0162914228,
0.0012746195,
-0.0291492715,
-0.0053456277,
0.0048079975,
0.0427034833,
0.013468218,
-0.0272371788,
0.0329373963,
-0.0122767892,
0.0020326136,
-0.0131355347,
-0.0234576184,
0.0203610081,
0.0240165275,
0.0313456841,
-0.0954055339,
0.0076753092,
-0.0550735146,
2.739e-7,
0.0279395934,
0.0606232993,
-0.001676382,
-0.0303953886,
0.029148465,
0.059457045,
-0.0161581207,
0.0882002339,
-0.0558278784,
0.0091712745,
0.0053729336,
-0.0137954224,
0.0227219872,
-0.011151664,
-0.028745804,
-0.009598515,
-0.0770237744,
0.0318329632,
-0.0213194303,
0.0075407294,
0.0625691041,
0.083084628,
0.0706396848,
-0.0240512248,
-0.0230018999,
-0.0624841414,
0.0293244384,
0.0093231993,
0.0155068412,
0.0178922638,
0.0356695056,
0.0169830564,
0.0232509971,
0.0230091754,
-0.0816463903,
0.0113543067,
-0.0043990142,
0.0365658961,
-0.0548883528,
-0.0095220618,
-0.0438727103,
-0.0333050117,
0.0219509006,
-0.0196119715,
0.0557170659,
-0.0225116648,
-0.0207936261,
-0.0226693992,
-0.0207388476,
0.0068129264,
0.0759826079,
-0.0488325357,
-0.0288635138,
-0.0110356221,
0.0034932878,
0.0294549614,
-0.0063249595,
-0.0962366313,
-0.0032107015,
-0.0091030989,
-0.0057134083,
-0.0097440872,
-0.0020259253,
0.0076981476,
0.0669897124,
-0.0235033408,
-0.0076021464,
2.219758074e-34,
0.0330795981,
-0.0402444042,
-0.0126259672,
0.011495335,
0.0232688952,
-0.0034341572,
0.0689517409,
-0.0094748866,
0.0212641861,
-0.0372897871,
-0.0254014023
] |
http://arxiv.org/abs/2401.17181v1 | http://arxiv.org/pdf/2401.17181v1 | Transfer Learning for Text Diffusion Models | 2024-01-30T17:11:56 | 2024-01-30T17:11:56 | Kehang Han, Kathleen Kenealy, Aditya Barua, Noah Fiedel, Noah Constant | cs.CL | cs.CL | In this report, we explore the potential for text diffusion to replace
autoregressive (AR) decoding for the training and deployment of large language
models (LLMs). We are particularly interested to see whether pretrained AR
models can be transformed into text diffusion models through a lightweight
adaptation procedure we call ``AR2Diff''. We begin by establishing a strong
baseline setup for training text diffusion models. Comparing across multiple
architectures and pretraining objectives, we find that training a decoder-only
model with a prefix LM objective is best or near-best across several tasks.
Building on this finding, we test various transfer learning setups for text
diffusion models. On machine translation, we find that text diffusion
underperforms the standard AR approach. However, on code synthesis and
extractive QA, we find diffusion models trained from scratch outperform AR
models in many cases. We also observe quality gains from AR2Diff -- adapting AR
models to use diffusion decoding. These results are promising given that text
diffusion is relatively underexplored and can be significantly faster than AR
decoding for long text generation. | Transfer Learning for Text Diffusion Models: In this report, we explore the potential for text diffusion to replace
autoregressive (AR) decoding for the training and deployment of large language
models (LLMs). We are particularly interested to see whether pretrained AR
models can be transformed into text diffusion models through a lightweight
adaptation procedure we call ``AR2Diff''. We begin by establishing a strong
baseline setup for training text diffusion models. Comparing across multiple
architectures and pretraining objectives, we find that training a decoder-only
model with a prefix LM objective is best or near-best across several tasks.
Building on this finding, we test various transfer learning setups for text
diffusion models. On machine translation, we find that text diffusion
underperforms the standard AR approach. However, on code synthesis and
extractive QA, we find diffusion models trained from scratch outperform AR
models in many cases. We also observe quality gains from AR2Diff -- adapting AR
models to use diffusion decoding. These results are promising given that text
diffusion is relatively underexplored and can be significantly faster than AR
decoding for long text generation. | [
0.0009294276,
0.0705077201,
-0.0091545209,
0.0056957207,
0.0215857457,
0.022750549,
0.007126031,
-0.0164784547,
-0.0292449482,
-0.012273089,
-0.0117003275,
-0.0325352438,
0.0051620496,
0.029027544,
0.0372941643,
-0.0155608673,
0.0311516486,
-0.0531742163,
-0.0523369014,
0.0092635341,
0.0071402737,
0.0333781727,
0.0083602825,
0.0100712404,
0.0276558679,
-0.0108106295,
-0.0129423803,
0.0119655645,
0.0254996531,
-0.0291549917,
-0.0004387749,
0.0382578336,
0.054976508,
-0.0034643842,
0.0000021592,
0.0057397434,
-0.0313954912,
0.0042805457,
0.0056183184,
0.0369389243,
0.0709426105,
-0.0335423648,
-0.0081490623,
0.0576944984,
-0.026943367,
-0.0284257792,
0.0642534196,
-0.0072504799,
0.0846486911,
0.0888557732,
-0.0162811559,
-0.0527783819,
0.0426474996,
-0.0483312793,
0.0102987587,
-0.0556745827,
0.0171434097,
0.0529932119,
-0.0133479768,
0.01508475,
-0.0479404852,
0.0499141179,
-0.0033066361,
-0.0136859268,
0.0562485382,
0.0389539748,
0.0299673658,
0.0082182642,
-0.0310790855,
0.0774479583,
-0.0168508235,
-0.0324332006,
0.0202142857,
-0.0038295465,
-0.0352094658,
-0.0092926528,
-0.0195428859,
-0.002085485,
0.0209741201,
0.0511980802,
0.016677618,
-0.0286069661,
0.0368252099,
-0.0251102913,
-0.0530904979,
0.0015351828,
0.0182560887,
0.0079631815,
0.0004298929,
0.0211725086,
0.0457647443,
-0.0407321155,
0.0301599503,
0.0358891115,
0.0238504894,
0.0182785224,
-0.0794119313,
-0.0520116389,
-0.0325462036,
-0.0235796366,
0.0419392064,
0.0297205113,
0.0353425853,
0.017481938,
-0.0810910687,
0.0323549807,
-0.0387121178,
-0.0287834257,
-0.0503690131,
0.0120027047,
-0.0681945905,
0.0114499237,
-0.0514159054,
0.0371130407,
-0.0105533246,
-0.0316331647,
-0.0366478711,
0.0288530663,
0.033752162,
-0.0228780527,
0.0332261957,
0.0060027828,
-0.0499538518,
-0.0167827401,
0.0242667552,
0.0531247295,
0.0147281298,
-0.0120724449,
0.0319363885,
-0.0207309462,
-0.0246078391,
0.0069968565,
0.0272220876,
-0.0368247852,
-0.0052758069,
0.0320107453,
-0.0595020168,
0.0054272013,
-0.0404547565,
0.0306391716,
0.0116393464,
-0.0070237392,
0.0052943313,
-0.0642062351,
0.0000653047,
0.0032918686,
0.0287321638,
-0.0463681258,
0.0211527776,
0.0043944218,
-0.0294933487,
0.0551547259,
0.0005162646,
0.0304447245,
0.0370288119,
-0.0181988198,
0.009054102,
0.0620249771,
0.0401593894,
0.0441372916,
-0.0251778718,
-0.0357909873,
0.00013844,
0.0322552472,
0.0263591427,
0.0540108457,
-0.0044030468,
-0.0203435738,
-0.011389521,
-0.0012376237,
0.0749146193,
0.0343386419,
-0.0156156057,
0.0127154533,
0.0685947537,
0.0339407735,
0.0269306321,
0.0162895285,
0.0543734208,
0.0286873598,
-0.0281003974,
-0.030874785,
0.0354652517,
0.0283623748,
0.0021259456,
-0.0021212366,
0.0183114205,
0.0138366921,
-0.0403447114,
-0.0194344204,
0.0266291201,
-0.0361308493,
0.0509251878,
-0.0158780143,
0.0530107059,
0.0356364176,
-0.0710856095,
0.028071139,
-0.002723525,
-0.0244083107,
0.007551875,
-0.0523855947,
0.0258955173,
0.030626148,
0.0563064367,
0.0088509349,
0.0115857078,
-0.0302652605,
0.0247362293,
0.008337195,
0.0520426147,
0.0104382811,
0.0067950198,
-0.0588327199,
-0.0016263077,
0.0155367842,
0.0209170319,
0.0481941923,
-0.0220616572,
0.0698664337,
0.0262705684,
-0.0145201078,
-0.0640555471,
0.0397448465,
-0.0768255368,
-0.048421815,
0.0440060459,
-0.0171282515,
0.0017726183,
-0.0259186774,
-0.0479582846,
0.0005159775,
-0.0176454484,
-0.0500289425,
-0.0311394446,
-0.0133264568,
-0.0619916543,
-0.042981565,
-0.0315400176,
-0.0134682767,
0.0240176339,
0.0367923155,
-0.0360002927,
-0.0182172973,
0.0467850156,
-0.0249911826,
0.0392750576,
-0.0194319915,
0.0345982723,
0.0162551831,
0.0486050099,
-0.0211686138,
0.0128337266,
0.0064515085,
-0.0421062522,
-0.0344707891,
0.0443316884,
0.0075732544,
0.0266421363,
-0.0062091667,
-0.0151668517,
-0.0149329379,
0.0307829361,
-0.0143776946,
0.0227499194,
0.0004156346,
-0.0155757498,
-0.049977757,
0.0350162908,
-0.0349086449,
0.0266821422,
0.0249076746,
0.0206086934,
0.0068309954,
-0.0041251658,
-0.0473926999,
0.0053685256,
-0.0235383101,
-0.0063476339,
0.0294413976,
-0.0188497175,
0.0014549615,
-0.0572234318,
-0.0119910585,
0.010768217,
0.0699139163,
-0.026773816,
-0.0396036319,
-0.0572447889,
0.0595900379,
0.0199294388,
-0.0016958254,
-0.0043850793,
-0.0681576952,
-0.0054680388,
-0.0116553875,
0.055850938,
0.0581466593,
0.0316032097,
0.0372944474,
-0.0074522216,
0.0018111651,
-0.0405357108,
0.0128692389,
-0.0176621862,
-0.008340627,
-0.0545906611,
-0.0687325746,
0.0243083071,
0.0768948942,
0.017176833,
-0.0105302138,
-0.0679232106,
-0.0214776378,
-0.0239748266,
-0.0423167534,
-0.0114763025,
0.0101397047,
-0.0064108195,
0.0055851322,
-0.0364049412,
-0.0521667153,
0.0316174217,
-0.0481008664,
-0.0050438815,
-0.0587033369,
-0.0620347485,
-0.0387541689,
0.0116687017,
0.0050903293,
0.0185252428,
-0.0160069652,
-0.0362086967,
-0.0066831065,
0.0103821354,
-0.0109826578,
-0.0249395836,
0.0331810899,
-0.0397498012,
0.0157588478,
-0.012670611,
-0.002354125,
-0.0106099229,
-0.0339971781,
0.0170336924,
-0.0050001387,
0.0216608718,
-0.027931219,
0.0072510312,
0.0398449823,
-0.0013973743,
0.0527734049,
0.0072671892,
-0.0067710727,
-0.0295730531,
0.0283239614,
0.0584944449,
-0.0200203694,
-0.0251203161,
-0.0559580252,
0.0256952941,
0.0200285502,
0.0296226367,
0.0147296898,
-0.0740978345,
0.0068056569,
0.0209648907,
0.0052782348,
-0.0132067865,
0.0246635042,
0.0841051117,
-0.0529387072,
-0.0389728099,
0.0791519359,
0.0199592058,
-0.0415760614,
-0.0103050089,
0.0077997865,
0.0694535747,
-0.0138994474,
0.0192926601,
-0.0641291738,
-0.0397401042,
0.001477089,
0.0201542322,
0.0285062827,
-0.0038439317,
-0.0315471105,
-0.1002307907,
0.033992365,
0.0282427613,
-0.0326002575,
-0.0322331786,
-0.0460500009,
-0.0368335135,
0.0001239555,
-0.0086730374,
-0.0251145922,
-0.0222117696,
0.0086909067,
0.0539497621,
-0.0508088619,
-0.0486999527,
-0.0258141663,
-0.0265723765,
0.0557983555,
0.1137610078,
0.0165218189,
-0.0125854071,
-0.0380057953,
0.0461358503,
-0.0229050964,
-0.0082593132,
0.0062861051,
0.0371723808,
-0.0056373002,
-0.0267456919,
0.0319856964,
0.004108354,
-0.0565135665,
0.0129880644,
0.0327366516,
0.088116765,
0.0102754608,
-0.0758859217,
-0.0081402892,
-0.0377856866,
0.0328305028,
0.0583025478,
0.0321319774,
-0.0666765347,
-0.0767279342,
-0.0312414225,
0.0139741246,
-0.0484897867,
-0.0451325923,
0.0061394987,
-0.0415416434,
-0.0671096072,
-0.1066709906,
-0.0139731513,
0.0078166556,
0.0460453592,
0.010318486,
0.0399093255,
0.0310391169,
-0.0367505252,
-0.035512574,
-0.0124986852,
-0.0261512194,
0.015646847,
-0.0489375144,
-0.0100672273,
0.0208399631,
-0.0289898552,
-0.0073410198,
-0.0188951958,
-0.11438936,
-0.0591212995,
-0.0290627312,
0.0173459165,
0.0126937674,
0.0357282236,
-0.043545302,
-0.0428025015,
0.0685079321,
0.0061909859,
0.0183516815,
0.0258314274,
0.0358139053,
-0.0284914114,
-0.0060613113,
0.0140876425,
-0.0248706061,
0.0294260122,
-0.0155908335,
0.0210487302,
-0.0140596936,
0.0258394759,
-0.0053228461,
-0.0159856286,
-0.0197585393,
0.0380947627,
-0.0154747125,
-0.0309009384,
0.0108284075,
0.0119273383,
0.0291663371,
-0.0646390617,
-0.0003935467,
0.0840504169,
0.0102058146,
-0.0153794652,
0.0258145593,
0.03725053,
0.039877858,
0.0056558559,
0.0318616293,
-0.0053362167,
0.0875687152,
0.0579033643,
0.0000344,
-0.0198403653,
-0.0302767586,
0.0192845985,
0.029551778,
-0.023633888,
-0.0014992357,
0.0024752729,
0.0174533278,
0.0205490347,
0.0090033747,
-0.0475515015,
-0.031097332,
-0.0272710882,
-0.0238768663,
0.0417281687,
0.0152960103,
0.1181679741,
-0.0105116302,
0.0173323937,
-0.028446272,
0.0439554118,
-0.0304484293,
0.0368356928,
0.0152307153,
-0.0145863183,
-0.006945421,
0.0245363843,
0.0584389865,
0.0168451667,
0.0228442643,
-0.0077290311,
-0.0190220922,
-0.0224634372,
-0.006439717,
0.0035562557,
-0.0382107347,
0.0110658007,
0.027650822,
0.0270343926,
-0.0661104992,
-0.0389167517,
-0.0483617708,
0.0111173587,
-0.0235832091,
0.0095890034,
0.0595506169,
-0.0056757675,
0.0149592729,
-0.0736358762,
0.0183441006,
-0.0178955421,
-0.0231921691,
-0.0009790007,
-6.822290032e-33,
-0.0134068141,
0.0100908615,
0.0321414731,
-0.0623294488,
-0.0667125285,
-0.0147718312,
-0.0582828224,
-0.0241387673,
-0.0252052825,
0.017615797,
-0.004790104,
0.0304491371,
0.0064708679,
0.0376944058,
-0.0110589098,
0.0308557861,
0.0704212859,
-0.0195883363,
-0.017357558,
0.0466301851,
0.0031746274,
0.0578518882,
0.0011202836,
-0.0301310644,
-0.0375693738,
0.0181172136,
0.0098019885,
-0.0382012241,
0.061056003,
0.0170674566,
-0.0323818214,
-0.0027315025,
0.0187022667,
-0.0253824349,
-0.0457888171,
0.0304616578,
-0.0682327598,
-0.0121862218,
-0.0082356641,
-0.0091509139,
-0.0124512687,
-0.042942211,
0.0445227288,
0.009706853,
0.0116784209,
0.0026871967,
0.034200035,
-0.0220849104,
-0.0364107564,
-0.0196401514,
-0.0617387667,
0.0416713171,
-0.0059962929,
0.0217591207,
0.0181375351,
0.0027551318,
-0.0281293672,
0.0580449998,
-0.0722947121,
-0.0227248687,
-0.0069361799,
0.0675894618,
0.0619233288,
0.0289455913,
-0.0125492299,
0.0086610885,
0.0178040173,
0.0140515007,
-0.0152701112,
0.0141957505,
0.0352204517,
0.0482479446,
0.018612491,
-0.0090685664,
0.0412908159,
-0.0589865632,
-0.0490229651,
0.0030338513,
-0.047605291,
0.014867248,
0.0542066917,
0.0316625498,
-0.0246232506,
-0.0005507855,
-0.0557929128,
-0.0036471386,
0.0012251509,
-0.0491990633,
0.06241972,
0.0244279671,
-0.0184590723,
-0.0215335134,
0.0668281913,
-0.0431670435,
-0.0365885347,
-0.0025885389,
-0.0439021848,
0.027989462,
-0.0130453911,
0.0431114137,
-0.0567644052,
-0.0408335365,
0.0137961814,
0.0144537073,
0.0255904961,
-0.0194940865,
-0.0013989198,
0.0154861286,
-0.0829937235,
-0.0181238614,
-0.0507117622,
0.0398974642,
0.0079217358,
0.0767065883,
-0.0070495727,
-0.0150606073,
-0.0064003267,
0.0731177628,
-0.0344229043,
0.0062425658,
-0.0010300314,
-0.0318156481,
0.0067716655,
0.0167496316,
-0.0561656468,
-0.01017047,
-0.0490829237,
0.090856567,
0.0679912418,
-0.0646083876,
-0.0523175634,
0.0385212377,
2.779e-7,
0.0349662565,
0.0370923094,
0.0061401352,
0.0196536146,
0.0014484361,
0.030004045,
-0.0416865237,
0.0631606132,
0.0066453568,
0.0069167884,
-0.0061640143,
0.0431309231,
0.0999380052,
0.0179024581,
-0.0611246079,
-0.0380778909,
-0.0410863012,
0.027696535,
-0.0370747708,
-0.0374861732,
0.0225095786,
0.0726081729,
-0.0053793984,
-0.0020472119,
0.0426738039,
-0.0040548327,
0.0658078864,
0.0021900006,
0.0140340449,
0.0493587069,
-0.0095854281,
-0.0096290428,
-0.0492068753,
-0.0359371156,
-0.0443002954,
0.0107198358,
0.0032232099,
0.0438271016,
-0.0112460013,
-0.0383242704,
0.0482484177,
0.0107059758,
0.0167878121,
-0.0256740134,
0.0666721538,
-0.0308632385,
-0.0033855543,
-0.0112872832,
-0.0635684505,
-0.0099358289,
0.0268758405,
-0.0116399974,
-0.0102237277,
0.0110933157,
0.0404048525,
0.0147429071,
-0.0561545193,
-0.0656060576,
0.0128266886,
-0.0004964099,
-0.0324478932,
-0.031784419,
-0.0629765615,
0.0375195369,
0.0235185474,
-0.0247378573,
0.0523354188,
2.444971541e-34,
0.0421326831,
-0.0403756462,
-0.0026713717,
0.0363729261,
0.0351172611,
-0.00027661,
-0.0234870762,
0.020699691,
0.0059938785,
-0.0845202953,
-0.0081473608
] |
http://arxiv.org/abs/2401.17169v1 | http://arxiv.org/pdf/2401.17169v1 | Conditional and Modal Reasoning in Large Language Models | 2024-01-30T16:56:54 | 2024-01-30T16:56:54 | Wesley H. Holliday, Matthew Mandelkern | cs.CL | cs.CL, cs.AI, cs.LO, 68T50, 03B65, I.2.7 | The reasoning abilities of large language models (LLMs) are the topic of a
growing body of research in artificial intelligence and cognitive science. In
this paper, we probe the extent to which a dozen LLMs are able to distinguish
logically correct inferences from logically fallacious ones. We focus on
inference patterns involving conditionals (e.g., 'If Ann has a queen, then Bob
has a jack') and epistemic modals (e.g., 'Ann might have an ace', 'Bob must
have a king'). These inference patterns have been of special interest to
logicians, philosophers, and linguists, since they plausibly play a central
role in human reasoning. Assessing LLMs on these inference patterns is thus
highly relevant to the question of how much the reasoning abilities of LLMs
match those of humans. Among the LLMs we tested, all but GPT-4 often make basic
mistakes with conditionals. Moreover, even GPT-4 displays logically
inconsistent judgments across inference patterns involving epistemic modals. | Conditional and Modal Reasoning in Large Language Models: The reasoning abilities of large language models (LLMs) are the topic of a
growing body of research in artificial intelligence and cognitive science. In
this paper, we probe the extent to which a dozen LLMs are able to distinguish
logically correct inferences from logically fallacious ones. We focus on
inference patterns involving conditionals (e.g., 'If Ann has a queen, then Bob
has a jack') and epistemic modals (e.g., 'Ann might have an ace', 'Bob must
have a king'). These inference patterns have been of special interest to
logicians, philosophers, and linguists, since they plausibly play a central
role in human reasoning. Assessing LLMs on these inference patterns is thus
highly relevant to the question of how much the reasoning abilities of LLMs
match those of humans. Among the LLMs we tested, all but GPT-4 often make basic
mistakes with conditionals. Moreover, even GPT-4 displays logically
inconsistent judgments across inference patterns involving epistemic modals. | [
0.0457569361,
0.0781166553,
-0.0334848985,
0.040206667,
-0.0802653432,
0.0326279216,
-0.0030260258,
0.0268339235,
0.0309083853,
-0.0433234982,
-0.0047279983,
-0.0098681375,
-0.0148744117,
-0.0076520648,
0.024193462,
0.0007126841,
0.0469981954,
-0.0530434437,
0.0368514694,
-0.0264898557,
-0.0349711888,
0.0034446053,
-0.0076919985,
-0.0047525638,
0.0180121846,
0.0031002404,
-0.0036450648,
-0.032863386,
0.0052833082,
-0.01433123,
-0.0607728846,
0.0399147905,
0.0099809989,
0.0261481423,
0.000001875,
-0.0587477982,
0.0393278897,
-0.0291529484,
-0.0313862078,
0.0365990661,
-0.0079491483,
-0.0244931169,
0.0125077553,
-0.0095370263,
-0.0094652232,
-0.0153900785,
0.0277054273,
0.0696561411,
-0.0055674664,
0.0309739709,
0.0080325073,
-0.0441930741,
0.0193086471,
-0.0169594083,
0.0322407372,
-0.0530144647,
0.0630323812,
0.054899618,
0.0169146545,
-0.0003269997,
0.008531997,
0.0295805912,
0.0421935022,
0.0358823575,
-0.1041369885,
0.0530284084,
-0.0180596448,
-0.0141489254,
0.0235929284,
0.0033030272,
0.0465411916,
-0.0490767136,
-0.0377594531,
0.0145258633,
0.0181010272,
0.0589683503,
-0.0138911353,
0.0798591152,
-0.0428263061,
0.0241735671,
0.0867210552,
-0.0107325539,
0.0500972606,
0.0244017318,
-0.0522027239,
0.1066593155,
0.0090830298,
-0.0258605015,
0.006199243,
0.0112616569,
-0.0669086576,
-0.0542735197,
-0.0290068462,
-0.0051394915,
0.0152223753,
0.0277685039,
-0.0127179921,
0.0504178219,
0.0410083234,
-0.0308179203,
0.0102624269,
0.0512921177,
0.0026373565,
0.0433894098,
-0.0456829146,
0.0074473461,
-0.047803767,
0.0427059084,
0.0108040441,
-0.0325248353,
-0.035057772,
-0.0227244962,
0.0179452412,
0.0273376182,
0.020381581,
0.0190323032,
-0.0545974523,
-0.0059888624,
-0.0119628813,
0.0293703508,
-0.0133651812,
-0.0100884531,
0.0162147004,
-0.0373629518,
-0.1058337316,
-0.0070873662,
-0.0280524995,
0.0726121292,
-0.030644929,
-0.0603777878,
-0.0051656314,
-0.0281070229,
-0.0028135986,
-0.0278065056,
0.0271838084,
0.0660145804,
0.024442656,
0.0036719372,
-0.0455557927,
-0.0246070493,
-0.0016927912,
-0.073842302,
-0.0194532778,
-0.058240667,
-0.0044049076,
0.0442716777,
0.0030320599,
-0.0266418625,
0.010202569,
0.0036856974,
-0.0195539482,
0.07648772,
0.0417659171,
0.0295251962,
0.0130801303,
0.0147155868,
-0.0351817012,
0.0556254387,
0.0118199522,
0.0049189697,
-0.0129979765,
0.0067237825,
0.0070230151,
0.0158942919,
-0.0135620032,
-0.0134819467,
-0.023764845,
0.0015845309,
-0.0141807701,
0.0049621658,
0.0103094541,
0.0732657537,
-0.0246772207,
-0.0592875741,
0.0544137359,
0.0479669385,
0.0222766902,
0.0299025085,
0.0867662057,
-0.0574293137,
0.0162410345,
0.0339676589,
-0.0909465328,
-0.0083095161,
-0.0410830677,
-0.0099319294,
-0.0293909758,
0.0296904333,
-0.0040609469,
-0.032658767,
-0.0255549531,
-0.0024466394,
-0.0246208794,
0.0084272055,
0.0588158779,
-0.0273554474,
0.015458717,
0.0140851699,
0.0398283228,
-0.0142981168,
-0.0228126701,
-0.0667439997,
0.0375696383,
-0.0015603938,
0.0051250118,
0.0021270274,
-0.0201325454,
0.0260957088,
-0.0405214839,
0.0040561655,
0.0708782896,
0.0099620204,
-0.0023428032,
-0.0090364199,
0.044778239,
0.0018971476,
0.0443312377,
0.0134948408,
-0.0245297998,
0.0196162947,
-0.0049915165,
-0.0375643037,
0.0453516953,
0.019940814,
-0.0299057718,
0.0143436892,
-0.0604930185,
-0.0042107613,
0.0364280418,
-0.0264066048,
0.0080011375,
0.0190986376,
0.0217068885,
-0.0084115406,
-0.0278609097,
0.0327342562,
-0.0509991981,
-0.0171498749,
0.0208717883,
-0.0459302887,
-0.0435029864,
0.0099277366,
-0.0057884888,
-0.013766489,
0.0270406175,
-0.0270813052,
0.0488976836,
-0.0308883041,
0.0310005751,
-0.0133734103,
-0.0042642113,
0.0097423745,
0.1106560901,
-0.0081265522,
0.0246374328,
0.0204267651,
0.0283963215,
0.0101109054,
-0.0586858541,
0.0024890855,
-0.0020298322,
0.0248782597,
-0.0317096896,
-0.0189094879,
-0.0257439613,
-0.0521953665,
-0.0385505185,
-0.0898210555,
-0.0004308156,
0.0131863756,
0.0181184858,
0.0061851135,
-0.0262275375,
-0.0304459482,
0.0227043219,
0.0158101805,
-0.0116787897,
0.0260630623,
0.0307217762,
0.0103163365,
-0.036647249,
-0.0185812023,
-0.0052638417,
-0.0027816521,
-0.0578312054,
0.0210674163,
0.0288089719,
0.0285791121,
-0.0430986062,
0.0262855887,
-0.0665560812,
0.0395542569,
-0.0177069642,
-0.0048129489,
0.0018668466,
-0.0213748235,
0.0243405122,
0.0879280046,
-0.0338423438,
0.0175956842,
0.0786764845,
-0.0260872692,
-0.0229365285,
0.0056034392,
-0.018023219,
-0.0166534986,
-0.0564664379,
-0.081356816,
-0.05316291,
0.0465214215,
0.0187511202,
-0.0034180942,
-0.0342986472,
0.0621718317,
0.0054325056,
-0.0145590818,
0.0191641971,
-0.0070553836,
0.0170348976,
0.084449023,
0.0077299546,
-0.021267239,
0.0596366413,
0.006750437,
0.0268555302,
-0.0128362533,
-0.0031854631,
-0.0329266489,
0.0604859255,
0.0251749009,
0.0021592593,
0.0013284544,
0.0023906711,
-0.0375863612,
-0.0207999479,
-0.0034170977,
0.001220213,
0.0385452956,
-0.0792798996,
0.0204624273,
-0.0214134231,
-0.0402811281,
-0.0377395228,
-0.0187446605,
-0.0994945467,
0.0026757652,
-0.0184233524,
-0.014650967,
-0.0168986302,
0.018666327,
0.0049228594,
-0.0212772377,
0.0098385392,
-0.0173788015,
-0.027124105,
-0.0084086796,
0.0096492348,
0.0454695895,
0.0256933421,
-0.0422395431,
0.0203925874,
0.0418590121,
-0.013753281,
0.0411372371,
-0.0290546529,
0.0374091975,
0.0052107582,
0.0126798218,
-0.0382176377,
0.0291339345,
0.0277389288,
-0.0694832951,
0.013669271,
0.1012924239,
0.0321378186,
-0.0095088212,
0.0015732171,
-0.0379991569,
0.0101083349,
-0.0026650557,
-0.00826017,
-0.0860661864,
-0.0530359149,
0.0298647732,
-0.0649318919,
-0.0666696653,
-0.0448290519,
0.0038306122,
-0.1088131741,
0.0509886369,
-0.0199092105,
-0.0193214547,
-0.0291850679,
-0.0313195027,
0.03652291,
-0.018824663,
-0.0104569588,
0.0293909367,
-0.0024707566,
-0.0014550186,
-0.0075552524,
0.0396865308,
-0.0532194413,
0.0440957211,
-0.026237838,
0.0599763542,
-0.0204875711,
0.0334035791,
-0.0526771992,
-0.0655519366,
0.0053518279,
-0.002174865,
-0.0101630241,
-0.0080778971,
-0.0274756905,
0.010610356,
0.0141802737,
-0.0406871177,
0.0011807996,
0.0159646086,
-0.0343862213,
0.0301266853,
0.0826275423,
-0.0310786925,
-0.066479519,
0.0005874789,
0.03896657,
-0.0724542141,
0.0779466107,
0.0033439568,
0.0198657196,
-0.0466007292,
-0.0043192613,
-0.0351354517,
0.0089397952,
-0.0450603701,
0.0410061143,
0.0400604978,
-0.0093178703,
-0.0254600924,
-0.034361437,
0.0270032678,
-0.0264925137,
-0.00115866,
0.0073586046,
0.0183843765,
0.0008525716,
0.0062974379,
0.0302115623,
-0.0451971367,
-0.0076478487,
-0.02000173,
0.0219556876,
0.0406471007,
-0.0409745798,
-0.0162535831,
0.00862341,
-0.0730430111,
-0.0405341052,
0.0072626201,
0.0143445414,
0.0319375247,
-0.019062547,
-0.0101686474,
-0.0183146484,
-0.0382264964,
0.0227106102,
-0.006469199,
0.0523147248,
0.0159201268,
0.0140781244,
0.0294624008,
0.0327729732,
0.0259227715,
0.0465110503,
0.0011050006,
0.0055093197,
-0.0241100322,
-0.0081154965,
0.0401625484,
0.0142005431,
-0.0039826194,
0.0306008533,
-0.0049655568,
-0.0028998978,
0.0067672143,
-0.0305024702,
-0.0134854093,
0.0171076916,
0.0146609517,
0.0149741853,
0.040468134,
-0.0209134854,
-0.036059238,
-0.0109281158,
0.0126347123,
-0.0020691771,
0.0146285491,
0.0467842519,
-0.0044644484,
0.0140166366,
-0.0465509184,
-0.0142466258,
-0.0083287423,
-0.0192613192,
0.0673997775,
-0.0021093069,
0.0749976411,
0.030536579,
-0.0241378155,
-0.0517741516,
-0.0266114976,
-0.0845970809,
-0.0745702684,
0.0821909159,
0.0415429138,
0.031519331,
-0.034326233,
0.0136329625,
-0.0782084465,
0.0073344428,
0.0040533887,
-0.0100813024,
0.033845678,
0.0086557744,
0.06267602,
0.0021048028,
-0.0257137995,
0.0129245026,
0.0176402517,
-0.0044217184,
0.0689723045,
-0.0065480662,
-0.0384231769,
-0.0148460409,
-0.054418277,
0.0000923252,
-0.0077918656,
-0.0184026863,
0.0207102392,
-0.0808947161,
-0.0947410688,
-0.0251102634,
0.0285250731,
0.0220948514,
-0.0270274915,
-0.0424476899,
0.0627799928,
0.0273873806,
-0.0446755663,
-0.0136234639,
0.0118030729,
-0.0128740901,
0.0277562272,
-0.0503326766,
-6.423017213e-33,
-0.0167354271,
-0.0193594676,
0.0280118268,
-0.0327238776,
0.0030364194,
-0.021848131,
-0.0297011621,
0.0027496435,
-0.0535623021,
0.0312357079,
-0.0144783054,
0.0175992921,
0.0147545449,
-0.0174415503,
0.0259590354,
-0.0146162659,
-0.0039620376,
0.0417130999,
-0.0158406831,
-0.0243218094,
0.0230510626,
-0.030968504,
0.0219325274,
-0.0562857576,
0.0291414876,
-0.0088282321,
0.0058471574,
0.0184271298,
-0.1150407419,
0.0186104868,
-0.0173173472,
-0.0126805305,
0.0332441255,
0.0066416613,
0.0433321148,
0.002845143,
-0.0291982125,
-0.0146285603,
0.0874879435,
-0.0303270444,
-0.015319257,
-0.0177955683,
0.0353184976,
0.0210913494,
-0.0442444943,
0.022380475,
0.029017387,
-0.0299429391,
0.0612945929,
0.0342530571,
-0.0504372865,
0.024558045,
0.0116443066,
0.0190297496,
-0.0170683023,
0.0042685736,
-0.0213470683,
0.003574443,
0.003445816,
0.0188761372,
0.0543073043,
-0.0203154441,
0.0432324111,
-0.0204113852,
-0.0021954372,
0.0465681218,
0.0365755893,
-0.0101283817,
-0.0460608788,
-0.0028386922,
0.0144956177,
0.072233282,
0.0225615688,
0.0415376462,
0.0304131452,
0.0351062305,
-0.0615615249,
-0.0441926494,
0.0312636383,
0.0502592064,
0.0643730164,
-0.018510012,
-0.0267935768,
-0.0016908774,
-0.0418086015,
0.001828588,
-0.0386766344,
-0.014941901,
0.0092446897,
-0.0353055932,
0.0158044994,
-0.0045214691,
0.0128216334,
-0.0284807198,
-0.0204837248,
-0.0118577527,
0.0097806649,
-0.0235991161,
-0.012784862,
0.0455092117,
-0.0202484168,
-0.0148870172,
0.0111106522,
-0.0113351336,
-0.0019376748,
0.0196787342,
-0.007348781,
0.0052473801,
-0.0148736695,
-0.0290169138,
-0.0539526455,
-0.0331692658,
-0.0110860635,
-0.0071082064,
0.0706926882,
0.042901326,
0.0038408446,
-0.0452561378,
0.0248923115,
0.0130635342,
-0.0503492989,
0.0118741309,
-0.0341907591,
0.0258213971,
-0.0113261016,
-0.0192060918,
-0.0486509837,
0.0142163318,
-0.0252949595,
-0.0334448777,
-0.0038987042,
0.0240067486,
2.736e-7,
0.0520636588,
0.0623502359,
-0.0314794071,
0.0198616143,
0.0203070082,
0.0356306508,
-0.0422340445,
-0.0277993511,
0.0703483,
-0.0233346839,
0.0159171037,
0.0505416282,
0.0068091056,
0.0682192072,
-0.0147702489,
0.0442082323,
-0.0301946495,
-0.0629621372,
-0.0194975678,
0.0411596783,
0.0786520019,
0.0662454739,
0.0501165837,
-0.0189519245,
0.0324745961,
-0.0562551357,
0.0290872063,
-0.0301219933,
-0.0015494513,
0.0010535573,
0.0000197576,
0.0091120303,
0.0341038443,
0.0918589756,
0.061409235,
0.0121822916,
-0.0277913343,
0.0345549583,
-0.0170879588,
0.0459303446,
-0.0420811512,
-0.031970337,
-0.0077078566,
-0.0731253773,
0.1002308428,
-0.0591929331,
-0.0588995591,
-0.0010659056,
0.0180693511,
-0.0248490591,
-0.0048004095,
0.0165613256,
-0.0196846742,
0.0160038769,
-0.0053572743,
0.0053352476,
-0.0242698565,
-0.066980131,
0.0084415711,
0.0109324856,
-0.059541136,
-0.0693186074,
-0.0076855822,
-0.0095906453,
0.0362414345,
-0.0699553564,
-0.0039615314,
2.277611451e-34,
0.0219071358,
0.0060504414,
0.0131787499,
0.0147807738,
0.0743627325,
0.0285475515,
0.0113220476,
0.0422115438,
0.0188551843,
-0.0896160156,
-0.0115696322
] |
http://arxiv.org/abs/2402.00891v1 | http://arxiv.org/pdf/2402.00891v1 | Large Language Models in Cybersecurity: State-of-the-Art | 2024-01-30T16:55:25 | 2024-01-30T16:55:25 | Farzad Nourmohammadzadeh Motlagh, Mehrdad Hajizadeh, Mehryar Majd, Pejman Najafi, Feng Cheng, Christoph Meinel | cs.CR | cs.CR, cs.AI, cs.CL, cs.LG | The rise of Large Language Models (LLMs) has revolutionized our comprehension
of intelligence bringing us closer to Artificial Intelligence. Since their
introduction, researchers have actively explored the applications of LLMs
across diverse fields, significantly elevating capabilities. Cybersecurity,
traditionally resistant to data-driven solutions and slow to embrace machine
learning, stands out as a domain. This study examines the existing literature,
providing a thorough characterization of both defensive and adversarial
applications of LLMs within the realm of cybersecurity. Our review not only
surveys and categorizes the current landscape but also identifies critical
research gaps. By evaluating both offensive and defensive applications, we aim
to provide a holistic understanding of the potential risks and opportunities
associated with LLM-driven cybersecurity. | Large Language Models in Cybersecurity: State-of-the-Art: The rise of Large Language Models (LLMs) has revolutionized our comprehension
of intelligence bringing us closer to Artificial Intelligence. Since their
introduction, researchers have actively explored the applications of LLMs
across diverse fields, significantly elevating capabilities. Cybersecurity,
traditionally resistant to data-driven solutions and slow to embrace machine
learning, stands out as a domain. This study examines the existing literature,
providing a thorough characterization of both defensive and adversarial
applications of LLMs within the realm of cybersecurity. Our review not only
surveys and categorizes the current landscape but also identifies critical
research gaps. By evaluating both offensive and defensive applications, we aim
to provide a holistic understanding of the potential risks and opportunities
associated with LLM-driven cybersecurity. | [
0.0121273911,
0.066539146,
-0.027170904,
0.0551909879,
-0.0712298006,
-0.031387791,
0.0321650282,
-0.0065014269,
0.0227887817,
-0.0443229787,
0.0066634389,
-0.056167081,
-0.0014449777,
0.0297031682,
0.0774928704,
0.0244311895,
0.0284677297,
-0.0660703704,
-0.0076041943,
-0.0093528619,
-0.0258544013,
-0.0081428606,
0.0150342518,
0.0473324917,
0.0169281531,
0.0140941245,
-0.0134826973,
-0.0112370187,
0.0148837073,
0.0015014914,
-0.0016049907,
-0.004367691,
0.0312736705,
0.0546738394,
0.0000018183,
-0.0356928855,
-0.0207776353,
0.0364992023,
-0.0318634771,
0.0436023846,
0.0128035489,
-0.066171132,
-0.0161960907,
0.0295030139,
-0.0336007141,
0.0098930057,
0.0238452237,
0.025720425,
0.0326482989,
0.1063787639,
0.0016388017,
-0.0469799377,
0.0603279993,
-0.0193354953,
-0.0595945902,
-0.0025860148,
0.0473857671,
0.0055813394,
0.0074940315,
0.0343169011,
0.0094780307,
-0.012001249,
-0.0129417554,
-0.0032821877,
-0.077148661,
0.0349461846,
0.0005323734,
0.0398248211,
0.0141337011,
0.0202039722,
-0.0447545946,
-0.0104458462,
0.0278764125,
0.0070503615,
-0.0013095887,
0.0034978751,
-0.017449325,
0.0081987092,
0.0148133496,
-0.0085688587,
0.0226608645,
-0.0233949553,
0.0778852701,
0.018420767,
-0.0470475554,
0.0480959304,
0.0084673502,
-0.0055314689,
-0.0219135564,
-0.0929334983,
-0.0258115679,
-0.0266866423,
-0.0099847568,
0.0433570929,
-0.0022430418,
0.0015250072,
-0.062706016,
-0.0264114887,
0.0438469872,
-0.0473112762,
0.0218402855,
-0.0093924906,
0.0309983138,
0.0113185793,
-0.014814171,
0.0428121127,
-0.0447582304,
-0.0070065861,
-0.0090928292,
-0.0330310203,
-0.0477635004,
-0.0450498872,
0.0077043422,
0.0319644995,
0.0437511504,
0.0135270413,
-0.0452589467,
0.0183254592,
-0.0460295789,
0.0555927269,
-0.0336796977,
-0.0399517864,
0.0265647043,
0.0005002401,
0.0492862798,
0.0030574636,
0.0052252072,
0.0078853108,
-0.0402023308,
-0.0524476431,
-0.015788842,
0.0318153985,
-0.0732812881,
-0.0649365187,
-0.0087981205,
0.1127481461,
0.0345501415,
-0.0177073441,
-0.0124198087,
-0.04166748,
-0.0401588082,
-0.0230653509,
-0.0007363161,
-0.0591953434,
-0.0468309075,
0.0134706479,
-0.0028212334,
-0.0687199757,
-0.0213962458,
-0.0388482027,
-0.0518284515,
0.0811357573,
-0.0053327414,
0.0041768351,
0.0237866212,
-0.0098919207,
0.0374437347,
0.0396424308,
0.0233134869,
0.0294948928,
0.0075409878,
-0.0281210113,
0.0358768515,
0.0625216961,
0.0161671638,
0.00899354,
-0.0659902841,
0.0165333841,
0.053646706,
-0.0289247073,
-0.0290003847,
0.0564360581,
0.0093542347,
-0.0398838669,
0.0644096509,
0.0254692882,
-0.0316489413,
0.0582699031,
0.0414129421,
0.0395119712,
0.0019094221,
-0.0164618343,
-0.0523023196,
-0.018566329,
-0.0623617433,
-0.0207480229,
-0.0104510691,
0.0253033843,
-0.0198122561,
-0.011394565,
-0.0202100165,
-0.0571319833,
0.0117674749,
-0.0136449896,
0.078423053,
0.0030325053,
0.0084931543,
-0.0064730155,
0.029366849,
0.0035535388,
0.0024692507,
-0.1010093763,
-0.0540111847,
0.0352217779,
0.0536044538,
-0.0489685759,
-0.0325382575,
-0.0142985778,
-0.0072966558,
-0.007591275,
0.0054772701,
0.0282538291,
-0.0285409987,
-0.0323022269,
0.0291524753,
-0.0370097905,
0.0214921571,
-0.0027227325,
-0.0094346991,
-0.0030414581,
0.0039630905,
-0.0023065242,
-0.0411838777,
0.0075888466,
0.004696887,
0.0017380997,
0.024799183,
0.0279664285,
0.0633276626,
-0.0151276784,
-0.0350998677,
0.0500012934,
0.0552197881,
-0.0424992181,
-0.0337667949,
-0.0360801294,
0.0235280003,
0.046853222,
0.0344482958,
0.0028739586,
0.0479186997,
0.0616219863,
-0.0345673487,
-0.0103526255,
0.0768804848,
-0.0726124048,
0.0127499867,
0.003580848,
-0.0088588968,
-0.0271918587,
0.0602758415,
-0.0331152156,
0.0395456478,
-0.0020415115,
-0.0298511889,
-0.0105003966,
0.0041104099,
0.0195189305,
-0.0195427239,
-0.0517352335,
0.0066016493,
-0.0332900546,
0.0131482435,
0.0041350015,
-0.0169136561,
0.0351740755,
-0.0361957997,
-0.0267941076,
-0.0329501182,
-0.0118965628,
0.0317733102,
0.0046057194,
-0.0363917835,
-0.0180745218,
-0.0061168559,
0.0099899024,
-0.0107333707,
-0.0037419838,
0.0070687602,
0.020926889,
-0.0025782906,
-0.0770431086,
-0.0099373171,
-0.0408406183,
-0.0188582521,
0.0801983625,
-0.0101760831,
-0.0171283185,
-0.0235396214,
0.0370539241,
-0.0229058191,
0.0186898969,
-0.0476877987,
-0.0580822416,
0.0253617726,
-0.0331791751,
0.0035166154,
0.0374793373,
0.0214051623,
0.0219418015,
0.0843710303,
0.0160109047,
-0.0113156857,
-0.0122013418,
-0.0197856538,
-0.044812832,
-0.0899849087,
-0.0216382034,
-0.0046859835,
0.0601518191,
0.0424777307,
-0.034248177,
-0.007087315,
0.0282950997,
-0.0023400632,
0.0110096009,
-0.0204177201,
-0.0685477927,
0.0212552343,
0.0035775206,
0.0094250804,
-0.0755036101,
0.1134176627,
-0.0494641624,
0.031338308,
-0.0118953688,
-0.0657142177,
-0.0252120197,
0.0256252885,
-0.0001284607,
0.0334265344,
0.0093178852,
-0.00377116,
-0.0394791849,
-0.0352297798,
-0.0117805721,
-0.0238763522,
0.0126707777,
-0.0053129359,
0.0243762005,
0.0233047791,
-0.0015640212,
-0.0368860438,
-0.0289501213,
-0.0484076627,
0.0349483155,
0.0224181954,
-0.0524034649,
-0.0308110025,
0.031653516,
0.0573619343,
-0.0255308058,
0.0298596323,
0.0030283099,
-0.0042907535,
-0.0131269721,
0.0351897404,
-0.0415219516,
0.0325536728,
0.0044198288,
0.0528719127,
0.0320129395,
0.0363305584,
0.0156131266,
-0.0279922299,
-0.0044172443,
0.0348394737,
-0.0123740034,
-0.0397111513,
0.0434913673,
0.0123183867,
-0.0260646045,
-0.0191151705,
0.075815931,
0.0091527747,
0.0033293632,
0.0387347639,
-0.0164915752,
-0.0085715307,
0.0051823193,
0.0192395411,
-0.0316996314,
-0.0147855151,
0.0160551183,
0.019358797,
-0.0212670416,
-0.0182172451,
-0.0073361145,
-0.0912874639,
-0.0056699524,
0.0347335376,
-0.0146659324,
-0.0499251895,
-0.0652514845,
-0.0327800401,
0.0168558788,
0.0523022935,
-0.0216835402,
0.0228466932,
-0.0331633016,
0.0077818483,
-0.0627558082,
-0.0328600034,
0.0323728733,
-0.0266932528,
0.0496778451,
0.0687090904,
0.0242142882,
-0.0578851178,
0.0497163497,
0.0288537703,
-0.0094884569,
0.006475579,
0.0147360591,
0.0063768602,
0.0379077047,
-0.0226211045,
0.0177590754,
0.0494407341,
-0.0308696758,
0.0003005576,
0.047655616,
0.0505414084,
-0.0616318509,
-0.0040687355,
-0.0095826285,
-0.0180342067,
0.0267277695,
0.0575804226,
-0.0030375603,
-0.0630477294,
-0.0162978228,
-0.0517423935,
0.0087241875,
0.0148126595,
-0.0615501963,
0.0243663825,
0.0222810227,
-0.0247254875,
0.0072537684,
-0.0056643924,
-0.0041835187,
-0.0225377176,
0.0497717969,
-0.0051325047,
0.0084287254,
0.0102911303,
-0.0361757986,
-0.038544625,
-0.0228767842,
0.0704242364,
-0.0427762866,
0.0065399483,
0.0308222808,
-0.0406835936,
-0.0294058416,
0.0186467934,
-0.073047854,
-0.0905685499,
0.0069684247,
0.0506157093,
0.0151709625,
-0.0019379421,
-0.0505911559,
-0.072923705,
0.0888842046,
0.0037896489,
-0.0037686226,
-0.0086350432,
0.0553777888,
-0.0262255203,
0.0225676801,
0.0428248085,
-0.024905974,
0.0124925328,
0.0605158992,
-0.0125314519,
-0.0505750701,
-0.0165152438,
0.0217508432,
0.0589875765,
0.0394441076,
-0.0004266745,
0.0069293915,
0.046434585,
0.0139030814,
-0.0104795415,
-0.0238662511,
0.0158750415,
-0.0139591377,
0.0056476234,
0.0037669323,
0.0368334316,
0.0085938722,
0.0151288891,
-0.0066020614,
0.0089380695,
0.0289627519,
0.0273984019,
0.0422639102,
0.0813218802,
-0.0170039982,
0.0072340318,
0.0668904707,
0.0201146621,
0.0036291732,
-0.004435936,
0.0249395911,
-0.0012909421,
0.0048042177,
-0.0418110006,
0.0188465565,
-0.0371552631,
-0.060753122,
0.0320002995,
0.0326905586,
0.0515805967,
0.0338326506,
0.0168453027,
-0.0277209505,
0.0114609897,
-0.0439196862,
0.0093766367,
-0.0121464292,
-0.0297966078,
0.0189077966,
-0.0024574674,
-0.0470429584,
0.0426945649,
0.0837501213,
0.028728487,
0.0567510687,
0.0046841218,
-0.0244563576,
-0.0102837058,
-0.0426840112,
-0.0058999555,
0.0005686596,
-0.0531759597,
0.0022896759,
-0.0573559925,
-0.0875279307,
-0.0007117019,
-0.0224354472,
0.0259853061,
-0.0171355214,
-0.0271358453,
0.0709423274,
-0.0137330005,
0.054635305,
-0.0836119503,
0.0370170027,
-0.000892421,
-0.0432720073,
0.0221103393,
-5.456324822e-33,
0.0401835218,
-0.0408974327,
-0.0090921195,
-0.006735845,
0.0167377051,
0.0166542605,
-0.010458135,
-0.0178720336,
-0.025275439,
-0.0299922489,
0.0167716835,
-0.0303092189,
-0.0133490032,
0.0636610165,
-0.023431994,
0.0051972037,
0.0075460454,
-0.007931564,
-0.0009767647,
0.0405561887,
0.0760004744,
-0.0277435761,
-0.0053637661,
-0.0544874109,
0.0423262082,
0.0002470315,
-0.0200192444,
0.0159860067,
-0.0457617342,
0.0375082046,
-0.0083904304,
0.0470016375,
0.0068467096,
0.0050966325,
-0.0062817093,
0.0118227694,
0.0271963868,
0.043248836,
-0.0225886069,
-0.0162303243,
0.0490223318,
-0.0674809515,
0.0300058667,
-0.0050447686,
-0.0023470786,
0.0271399561,
0.0042644101,
-0.0269622318,
-0.0042555779,
-0.02622444,
-0.0163529944,
0.0011984539,
-0.0661328658,
0.0093116015,
0.0125729628,
-0.0189794339,
0.0274235997,
0.0351873897,
-0.0454506353,
0.0920536518,
0.0230240971,
0.0408535562,
0.0761957914,
0.0218290128,
-0.0052844384,
0.0130701903,
-0.0006442627,
-0.0305032246,
-0.0115298089,
-0.0045015886,
0.0211432483,
0.0349448323,
0.0566368029,
0.023396682,
-0.0299076084,
-0.046716135,
-0.1009833589,
-0.0122404052,
0.0274317227,
0.0300252214,
0.0583931021,
-0.0042293468,
-0.0467644967,
-0.0188550632,
-0.0253832992,
-0.0175208803,
-0.0149520943,
-0.0153626231,
-0.0246166755,
-0.0269554928,
-0.0199747682,
-0.0014557516,
0.0220068078,
-0.0530761629,
-0.0190537684,
0.0456674919,
0.0021394743,
0.0033433035,
-0.0066360421,
0.0230543297,
-0.0311855078,
-0.0487331524,
-0.0318069682,
-0.0501739942,
0.008777895,
-0.0255644098,
0.0229420215,
0.0327622294,
-0.0359386057,
-0.0070322449,
-0.012825083,
0.0286704861,
0.0129123526,
0.0437302701,
0.0387250558,
0.0271207616,
0.01320545,
-0.0055816309,
0.0283958707,
-0.0388955064,
0.0193291605,
0.0412639603,
-0.0357114412,
0.0304052457,
-0.0190686658,
0.01477875,
-0.0454018936,
0.0319285616,
0.0395220742,
-0.0565139838,
-0.0158377625,
-0.0556600802,
2.686e-7,
0.0036867012,
0.042988155,
0.0150287459,
0.031273853,
0.0127772307,
0.0702793822,
-0.0296772588,
0.0600263774,
-0.0469166785,
0.0032584697,
-0.0072529777,
0.0447969474,
0.0369124264,
-0.0035175995,
-0.0085819354,
0.0493005961,
-0.0790377036,
-0.0259218197,
-0.0901194885,
0.0013727724,
0.045331195,
0.1101380587,
0.0108066192,
-0.0227278601,
-0.0315020829,
-0.0864800289,
0.0187730156,
-0.030239759,
-0.0244338699,
0.0361450054,
0.013895032,
0.0316947736,
0.0102203926,
0.0742567703,
-0.0086498931,
-0.0034358895,
-0.0089641688,
0.0365199745,
0.0599363782,
0.0017539383,
-0.0353800766,
-0.0107434439,
-0.0272571724,
-0.0334305502,
0.0726511255,
-0.0378690548,
-0.0724360198,
-0.004520732,
-0.011636178,
-0.0066694389,
-0.0110983727,
-0.0580950715,
0.0051084599,
-0.0040233675,
-0.0194909126,
-0.022748407,
0.0771237463,
-0.0678060353,
-0.007233406,
0.0033131316,
-0.0511815697,
-0.0130195972,
-0.0030525262,
0.0144860186,
0.066151455,
0.018202791,
0.0373184308,
2.580728741e-34,
0.0203995686,
-0.0493570715,
-0.0153560303,
-0.0087735774,
0.0335571356,
0.0210275184,
-0.0313222744,
0.0410083346,
-0.009624864,
-0.0665319934,
-0.0132680619
] |
http://arxiv.org/abs/2401.17167v3 | http://arxiv.org/pdf/2401.17167v3 | Planning, Creation, Usage: Benchmarking LLMs for Comprehensive Tool Utilization in Real-World Complex Scenarios | 2024-01-30T16:52:56 | 2024-06-03T11:28:29 | Shijue Huang, Wanjun Zhong, Jianqiao Lu, Qi Zhu, Jiahui Gao, Weiwen Liu, Yutai Hou, Xingshan Zeng, Yasheng Wang, Lifeng Shang, Xin Jiang, Ruifeng Xu, Qun Liu | cs.CL | cs.CL | The recent trend of using Large Language Models (LLMs) as tool agents in
real-world applications underscores the necessity for comprehensive evaluations
of their capabilities, particularly in complex scenarios involving planning,
creating, and using tools. However, existing benchmarks typically focus on
simple synthesized queries that do not reflect real-world complexity, thereby
offering limited perspectives in evaluating tool utilization. To address this
issue, we present UltraTool, a novel benchmark designed to improve and evaluate
LLMs' ability in tool utilization within real-world scenarios. UltraTool
focuses on the entire process of using tools - from planning and creating to
applying them in complex tasks. It emphasizes real-world complexities,
demanding accurate, multi-step planning for effective problem-solving. A key
feature of UltraTool is its independent evaluation of planning with natural
language, which happens before tool usage and simplifies the task solving by
mapping out the intermediate steps. Thus, unlike previous work, it eliminates
the restriction of pre-defined toolset. Through extensive experiments on
various LLMs, we offer novel insights into the evaluation of capabilities of
LLMs in tool utilization, thereby contributing a fresh perspective to this
rapidly evolving field. The benchmark is publicly available at
https://github.com/JoeYing1019/UltraTool. | Planning, Creation, Usage: Benchmarking LLMs for Comprehensive Tool Utilization in Real-World Complex Scenarios: The recent trend of using Large Language Models (LLMs) as tool agents in
real-world applications underscores the necessity for comprehensive evaluations
of their capabilities, particularly in complex scenarios involving planning,
creating, and using tools. However, existing benchmarks typically focus on
simple synthesized queries that do not reflect real-world complexity, thereby
offering limited perspectives in evaluating tool utilization. To address this
issue, we present UltraTool, a novel benchmark designed to improve and evaluate
LLMs' ability in tool utilization within real-world scenarios. UltraTool
focuses on the entire process of using tools - from planning and creating to
applying them in complex tasks. It emphasizes real-world complexities,
demanding accurate, multi-step planning for effective problem-solving. A key
feature of UltraTool is its independent evaluation of planning with natural
language, which happens before tool usage and simplifies the task solving by
mapping out the intermediate steps. Thus, unlike previous work, it eliminates
the restriction of pre-defined toolset. Through extensive experiments on
various LLMs, we offer novel insights into the evaluation of capabilities of
LLMs in tool utilization, thereby contributing a fresh perspective to this
rapidly evolving field. The benchmark is publicly available at
https://github.com/JoeYing1019/UltraTool. | [
0.0264696758,
0.0354562439,
-0.0319919176,
0.0243315119,
-0.0484662727,
-0.0125400461,
-0.0355148762,
-0.0089385565,
-0.0112409247,
-0.0165976454,
0.0159710199,
0.0012993601,
-0.0373566486,
0.0298830122,
-0.001693322,
-0.051404804,
0.0454896167,
-0.029865209,
-0.045243632,
0.0374489464,
-0.0105134575,
0.0196751971,
-0.0309185833,
0.0434178486,
-0.0216432586,
0.0160798095,
-0.0535686314,
0.0135027738,
-0.0198427141,
-0.0397441126,
-0.0246551372,
0.0429394133,
0.0117587009,
0.0295278151,
0.000002243,
-0.0834274366,
-0.0192831717,
-0.007378018,
-0.0376583338,
-0.0112372264,
0.0450611189,
-0.0155949472,
-0.0161933675,
0.0423849672,
-0.0677077323,
0.0148657272,
0.0227235146,
-0.0114159146,
0.0238551423,
0.0756557956,
-0.0084229074,
0.0091948919,
-0.0029911052,
0.0087753907,
-0.0211210456,
-0.0542078167,
-0.0227630492,
0.0105609251,
0.110104382,
-0.0057081855,
0.0369739197,
0.0190266967,
0.0265645757,
-0.0313402303,
-0.0221555196,
0.0057522273,
-0.0321054384,
-0.0330046564,
0.0114988983,
0.0608706288,
-0.0040188753,
0.0060261954,
0.0413167477,
-0.0155117894,
-0.0175165087,
0.0243276898,
-0.041680295,
0.0088960761,
-0.0050244215,
-0.00058408,
-0.0283274669,
-0.0148734776,
0.0239788759,
0.0141033176,
-0.003207966,
0.0852996036,
-0.0027176708,
0.0091748657,
0.0194227081,
-0.049268268,
-0.0048063351,
-0.0849275365,
0.0002325337,
0.0758550093,
-0.1025583297,
0.0234463345,
-0.0069678714,
0.03220753,
0.0235992931,
-0.0505450629,
0.0587915778,
0.0533796363,
0.0342409462,
0.0387071297,
0.0070597571,
0.0686553791,
0.0168682393,
-0.010725175,
-0.0500551164,
-0.0078714406,
-0.0790276527,
0.000983746,
-0.043175973,
0.0101978872,
-0.0734515414,
-0.0367053822,
0.0031066595,
0.0526067913,
-0.000003,
0.0281312075,
0.0237110332,
-0.0140935173,
0.0038382504,
-0.0241649915,
-0.0747587457,
-0.0356338397,
-0.0571220368,
0.0276955683,
-0.0071694115,
-0.0648794472,
0.0121812876,
0.0230973773,
-0.0070417873,
-0.0236088689,
-0.0117452256,
0.1125270128,
-0.0306688715,
0.0164696239,
-0.0594979525,
-0.0764603466,
0.0379489772,
-0.0684526041,
-0.0190260969,
-0.0769195259,
0.0328772105,
0.0229159854,
-0.0444952138,
-0.0120855607,
0.0118333297,
-0.0056212437,
0.0058414768,
0.0175002031,
-0.0030506179,
0.0481846705,
-0.0153576797,
0.0166221838,
-0.0000582958,
0.0840634704,
0.0262320898,
0.0598728508,
-0.004382418,
0.0304592829,
-0.0000596874,
0.0689586848,
-0.0437981673,
-0.0088253599,
-0.0280901454,
0.0328525342,
-0.0097387936,
0.0643475056,
-0.0104376851,
0.0296865124,
0.0004313159,
0.02651852,
0.051919058,
0.0238789413,
0.0360478461,
0.0160434823,
0.0644478947,
0.0389971696,
0.0032913454,
-0.0440304689,
-0.0507043935,
0.0115668084,
-0.0480490327,
0.0029359914,
-0.0353937596,
0.0040172767,
0.0028622367,
-0.0359540284,
-0.0175578203,
0.0063300291,
-0.0153600583,
-0.0292915739,
0.013264697,
0.0005916028,
-0.0631326959,
0.0292408392,
-0.0206885301,
-0.0145754451,
-0.0128083015,
-0.0442603007,
-0.009171891,
0.081287533,
0.0534556732,
-0.0572927333,
-0.063224189,
0.0403400175,
-0.0085184593,
0.0269152522,
0.0334071554,
-0.0188665669,
-0.0179938506,
-0.0188480429,
-0.0388599746,
-0.0199424569,
-0.005221481,
0.0068362029,
0.0627711788,
0.003818325,
0.0069112587,
-0.0648492873,
-0.0018415317,
0.0216767211,
-0.0329065472,
-0.0363992453,
-0.0245353878,
-0.0029147477,
0.0283073764,
0.0041717477,
-0.0277826358,
0.0345254987,
0.0373092964,
-0.0193636809,
-0.0542044751,
0.0124330251,
-0.0102968328,
0.0235379189,
0.0323576592,
-0.0055334629,
0.0044323374,
-0.0039182366,
-0.0295889378,
-0.0398635678,
0.0050640879,
-0.0084951045,
0.0250102133,
0.0012312386,
0.0280717593,
-0.0132875135,
-0.049871102,
-0.0402543731,
0.0498769693,
-0.0056311875,
-0.018362185,
0.0351742208,
-0.0126527809,
-0.0037080955,
-0.0189486165,
0.0075234389,
0.064956598,
0.0077903057,
0.0170496739,
0.0070138793,
0.0517037138,
-0.0447459705,
-0.0564368851,
-0.1039198786,
0.0232633129,
-0.018418774,
0.0113380346,
0.0113146584,
-0.0037069265,
-0.0205623377,
0.0102249244,
-0.0156326611,
0.0224095117,
-0.0040699509,
-0.0009226364,
0.0002768745,
0.0326839574,
-0.0244721286,
-0.0359053239,
0.0024426633,
-0.0279277097,
0.0198886544,
0.016972458,
-0.0174383596,
-0.0032870038,
0.062903963,
-0.0161991939,
0.002850546,
0.0055418466,
-0.0165159591,
0.0210508034,
-0.0235062838,
0.0134381885,
0.0286949407,
-0.028726887,
-0.0091352584,
0.0490541682,
-0.0366672911,
-0.0653201789,
0.0160986613,
-0.0081481561,
-0.025910303,
-0.0476469547,
-0.0244941302,
0.0299680699,
0.0305293649,
-0.0007823296,
0.0249415524,
0.0003250041,
0.01663992,
0.0107204383,
-0.0420520827,
0.0062429179,
-0.0066746399,
0.0000701793,
0.0739267319,
-0.0107034175,
-0.0460153334,
0.0208193846,
-0.0224316325,
-0.0034162325,
0.0066865464,
-0.0302092824,
-0.0474230982,
0.0878430977,
0.0300629996,
0.0192110632,
-0.0305381175,
0.0075000329,
-0.0416366681,
-0.0350991525,
0.0106556257,
-0.0171073321,
0.0183996689,
-0.0846247151,
0.0301761329,
-0.0122597907,
-0.0264500976,
-0.0368198603,
-0.0130292596,
-0.0285769701,
0.0642156973,
0.0425753631,
-0.0481646545,
-0.0070833592,
0.0086723715,
0.0164762121,
0.0603653863,
0.0116735352,
-0.0252972059,
-0.0265926514,
-0.0064574527,
0.0163735561,
0.0369888507,
0.0087362211,
0.0334957354,
0.0268642996,
0.0401321799,
0.0575791411,
-0.0170322191,
-0.0561718084,
0.0136519736,
0.0022950978,
-0.0230032951,
-0.0123557411,
0.0650561973,
0.0144230127,
-0.0536161885,
0.0047672931,
0.0953481421,
0.0386614986,
0.0226289481,
0.0249496549,
-0.0025852968,
0.0353725962,
0.0499822833,
0.0487266891,
-0.0736296028,
-0.0275913309,
0.0183452871,
0.0069566453,
-0.0045441296,
-0.0703742281,
-0.0084764436,
-0.1417854726,
0.0339546874,
-0.0185026526,
-0.0188596733,
0.05474158,
-0.0544953793,
-0.002827768,
0.0075524431,
-0.0198007245,
0.0177936703,
-0.0034307851,
-0.0122922678,
0.0087230178,
-0.0545989685,
-0.037220858,
0.0320344791,
-0.027757993,
0.0231343545,
0.0804223046,
0.0344992876,
0.0171570089,
0.044139605,
0.0749158636,
0.0739416182,
0.0374729522,
-0.0520152263,
-0.0093589425,
0.0340361893,
-0.0047255466,
0.0013912658,
0.0040011355,
-0.0556697622,
-0.0266439915,
0.0297140293,
0.1099074334,
-0.0204037689,
0.0227744151,
-0.0498472974,
-0.017931411,
0.0120204762,
0.0445714071,
-0.0015838183,
-0.006065994,
-0.0616137199,
-0.0212843604,
-0.0595781468,
-0.0442798957,
-0.0511439294,
0.0114546418,
0.0097678714,
0.0238062516,
-0.0371987745,
-0.0304409619,
0.0665547699,
0.041095458,
0.02899714,
-0.0068662814,
0.0256561786,
0.0061239721,
-0.0492393114,
-0.0058819992,
-0.013905392,
0.054720629,
-0.0338886529,
0.0009512628,
0.0148420408,
-0.007236857,
-0.0119843604,
0.028993804,
-0.0394881591,
-0.0734030604,
-0.0031977363,
0.0441559888,
0.0333297066,
-0.0432679355,
-0.0109116975,
-0.0460693203,
-0.0011720414,
0.0401196145,
-0.048478581,
0.0088496255,
-0.0046786522,
0.0387932658,
0.0067901607,
0.0518298373,
-0.0311004929,
0.0054833377,
-0.0241611712,
0.05207441,
-0.0295017734,
-0.0498439111,
0.0221097711,
0.0352532603,
-0.036251355,
0.0173277352,
0.0154077476,
-0.0084462026,
0.0098204659,
-0.0175419655,
-0.0475259461,
-0.0561210811,
0.0138153555,
-0.0102003021,
-0.0066663804,
0.054632917,
0.0448870286,
-0.002736643,
0.0157823842,
0.0270200353,
0.0089491708,
0.0118573317,
-0.0312910937,
0.0449730121,
0.020601932,
0.0078663826,
0.0207160357,
-0.0152458949,
0.0508510992,
-0.00727718,
0.0255458951,
-0.0058744983,
0.0090713203,
-0.0393958986,
-0.0131192384,
-0.0189558975,
-0.1350783408,
0.0269616749,
0.0212028921,
0.0500353426,
-0.0422874168,
0.0224100985,
-0.0625465736,
0.0090530282,
-0.0415750928,
0.0025898884,
0.0010217169,
-0.0013721526,
0.0160385743,
0.029337272,
-0.0201034546,
-0.0261490922,
0.0481779017,
0.0044185086,
0.0492336452,
0.0135473665,
-0.0376716256,
-0.0552602932,
-0.0531894788,
0.0381988175,
-0.0082504675,
-0.0077332556,
-0.0163219571,
-0.0782613531,
-0.0237959512,
-0.0148026701,
0.0196262132,
-0.0119968867,
-0.0242354535,
0.0639469549,
0.0570539385,
0.0254349131,
-0.0029964417,
-0.0284448657,
-0.0264825802,
-0.0344595946,
-0.012213313,
0.0001168756,
-5.863756674e-33,
-0.0329739749,
-0.0501634367,
0.074675113,
-0.0450144075,
0.0100440001,
-0.0302120205,
-0.0540194251,
0.0294896904,
-0.0182494707,
-0.0040414124,
-0.0282156691,
-0.0022427004,
0.0193192605,
-0.008920352,
0.012897918,
0.0083870133,
0.0331991874,
0.0025116184,
-0.040580593,
-0.0222008079,
0.0245065819,
-0.0060675479,
0.0107571669,
-0.0429240875,
-0.0392500609,
0.0273490455,
0.006898162,
0.0594875887,
-0.01071091,
0.0549503863,
-0.0008716049,
0.018142825,
0.0121881189,
-0.0032933417,
-0.0339917615,
-0.0049960944,
-0.0257737804,
-0.0291024819,
0.0321092457,
-0.0048436308,
-0.0186160039,
-0.0192915387,
0.0294335112,
0.0204628054,
-0.0502314344,
0.0011524522,
0.0112320259,
-0.013454766,
-0.0149190119,
-0.0132803796,
-0.0709186569,
0.033371985,
-0.0087523106,
-0.0270482339,
0.009437127,
0.0195205156,
0.020831624,
-0.0184582025,
-0.0244854372,
0.0143206129,
-0.0021366617,
0.0234081671,
0.0840249956,
-0.0042360183,
0.0016180037,
0.061103005,
0.045902729,
0.0017363346,
0.0090107657,
0.016067991,
0.0060906019,
0.1355585456,
-0.0097093266,
0.0402920321,
-0.0148213319,
0.055835072,
-0.032365039,
0.0386983939,
-0.0093407882,
0.0608724505,
0.0597243272,
0.037347272,
-0.0436796322,
-0.0170742217,
-0.0674259514,
0.0431981422,
-0.0391252898,
-0.0095440056,
0.0337159485,
0.023730021,
0.0330896117,
-0.0129915243,
0.0225698166,
-0.018292591,
-0.0127680451,
0.0274371505,
-0.0499501489,
0.019812189,
-0.0067654559,
0.0486309715,
-0.0405701734,
-0.0108505376,
-0.0159401167,
-0.004353724,
0.0350950286,
0.0059273578,
-0.0518126115,
0.0532224141,
-0.0573960245,
0.003673255,
-0.0632814318,
0.016572902,
-0.0166380517,
-0.0075473865,
0.0369845666,
-0.0419338085,
-0.0010822862,
0.0107455077,
-0.0205842089,
-0.0870377123,
-0.0149156433,
-0.0150863687,
0.0079621049,
-0.0126622822,
-0.0086011896,
0.0091521302,
-0.0448899195,
-0.0083771199,
0.049676422,
-0.0209416598,
-0.0035872892,
0.0034196051,
2.977e-7,
0.0276497528,
0.0455035344,
0.0150603028,
0.0814675912,
0.0057406873,
0.0321766585,
-0.0362701416,
-0.0015990089,
-0.0293870606,
-0.0035120437,
0.0154278921,
0.0518029965,
0.0406474285,
0.0555769019,
-0.0429387949,
0.0125526171,
-0.0019405122,
-0.0290953945,
-0.0947007611,
0.0019513884,
0.0422922075,
0.0531101488,
0.0142170331,
-0.0052494481,
-0.0163882151,
-0.093608059,
-0.004238646,
-0.0525272526,
-0.0209318008,
0.0468774922,
0.0076378291,
0.0175478291,
0.0139736095,
-0.0027020886,
0.0062722978,
0.0169540346,
-0.0145391375,
0.0874033645,
0.0050260252,
-0.0287785642,
0.0200262927,
-0.0917137787,
-0.0026237371,
-0.0312077627,
0.0450633615,
-0.0212194808,
-0.0117250374,
0.0043329978,
0.0278437994,
-0.0213795863,
-0.021855766,
-0.0090937428,
0.026251547,
0.0100509003,
0.0203112438,
-0.0084990207,
-0.0177640002,
-0.0387947857,
-0.0223147869,
0.068543531,
-0.0432268307,
-0.0730478168,
-0.0590237975,
0.0483551994,
-0.0143570099,
-0.0049261902,
-0.0172591973,
2.927586129e-34,
0.0020675373,
0.0538242571,
0.0230618231,
0.0493070409,
0.0693039522,
0.029255433,
-0.014992631,
0.0281802118,
0.0109244324,
-0.0791027695,
-0.0245891809
] |
http://arxiv.org/abs/2402.01759v1 | http://arxiv.org/pdf/2402.01759v1 | Systematic Literature Review: Computational Approaches for Humour Style Classification | 2024-01-30T16:21:47 | 2024-01-30T16:21:47 | Mary Ogbuka Kenneth, Foaad Khosmood, Abbas Edalat | cs.CL | cs.CL, cs.AI, cs.LG | Understanding various humour styles is essential for comprehending the
multifaceted nature of humour and its impact on fields such as psychology and
artificial intelligence. This understanding has revealed that humour, depending
on the style employed, can either have therapeutic or detrimental effects on an
individual's health and relationships. Although studies dedicated exclusively
to computational-based humour style analysis remain somewhat rare, an expansive
body of research thrives within related task, particularly binary humour and
sarcasm recognition. In this systematic literature review (SLR), we survey the
landscape of computational techniques applied to these related tasks and also
uncover their fundamental relevance to humour style analysis. Through this
study, we unveil common approaches, illuminate various datasets and evaluation
metrics, and effectively navigate the complex terrain of humour research. Our
efforts determine potential research gaps and outlined promising directions.
Furthermore, the SLR identifies a range of features and computational models
that can seamlessly transition from related tasks like binary humour and
sarcasm detection to invigorate humour style classification. These features
encompass incongruity, sentiment and polarity analysis, ambiguity detection,
acoustic nuances, visual cues, contextual insights, and more. The computational
models that emerge contain traditional machine learning paradigms, neural
network architectures, transformer-based models, and specialised models attuned
to the nuances of humour. Finally, the SLR provides access to existing datasets
related to humour and sarcasm, facilitating the work of future researchers. | Systematic Literature Review: Computational Approaches for Humour Style Classification: Understanding various humour styles is essential for comprehending the
multifaceted nature of humour and its impact on fields such as psychology and
artificial intelligence. This understanding has revealed that humour, depending
on the style employed, can either have therapeutic or detrimental effects on an
individual's health and relationships. Although studies dedicated exclusively
to computational-based humour style analysis remain somewhat rare, an expansive
body of research thrives within related task, particularly binary humour and
sarcasm recognition. In this systematic literature review (SLR), we survey the
landscape of computational techniques applied to these related tasks and also
uncover their fundamental relevance to humour style analysis. Through this
study, we unveil common approaches, illuminate various datasets and evaluation
metrics, and effectively navigate the complex terrain of humour research. Our
efforts determine potential research gaps and outlined promising directions.
Furthermore, the SLR identifies a range of features and computational models
that can seamlessly transition from related tasks like binary humour and
sarcasm detection to invigorate humour style classification. These features
encompass incongruity, sentiment and polarity analysis, ambiguity detection,
acoustic nuances, visual cues, contextual insights, and more. The computational
models that emerge contain traditional machine learning paradigms, neural
network architectures, transformer-based models, and specialised models attuned
to the nuances of humour. Finally, the SLR provides access to existing datasets
related to humour and sarcasm, facilitating the work of future researchers. | [
0.0927810967,
0.0477226861,
-0.045952823,
0.0241334829,
-0.0155752338,
0.0390716568,
0.0011647448,
0.0408885293,
0.0239471421,
-0.0542524196,
-0.0025016607,
0.0041816011,
0.0109389555,
0.0387181826,
-0.0023030313,
-0.0658337623,
-0.017226357,
0.0662488043,
0.0179146342,
-0.0380912498,
-0.0121051241,
0.0172364712,
-0.0258339401,
0.0358857065,
0.0136371898,
0.014747126,
-0.0018400989,
0.1029098257,
0.0168745983,
-0.0034210614,
0.0308572985,
0.0106316693,
-0.0281079933,
0.0066248155,
0.0000017098,
-0.0316410027,
0.0247335881,
-0.0427742675,
0.0073195226,
-0.0041529057,
0.0775272548,
-0.0364591181,
-0.0328677781,
0.0111527452,
-0.0300082862,
-0.0119723277,
0.0849399716,
0.0088858977,
0.0580611117,
0.0751832724,
-0.0101985866,
-0.0291959085,
-0.0023911104,
-0.014696694,
0.0222188327,
-0.0188827943,
0.0461036153,
0.0062684859,
-0.0265955925,
-0.011672819,
-0.0614976324,
-0.0015913012,
-0.0444207154,
0.0352571793,
-0.0485893637,
0.0109871253,
-0.0542936735,
-0.0171032194,
-0.0178306978,
0.0252400003,
0.0038309901,
-0.0052594505,
0.0165198818,
0.0317311436,
0.0054574446,
-0.0204641391,
-0.0146940444,
0.0050320406,
0.0030158772,
-0.0073409406,
0.0122803701,
0.0454674587,
-0.0065365196,
0.0551427267,
0.0407656208,
-0.0146858124,
-0.0063938694,
-0.0153511893,
-0.0198566914,
-0.0049186242,
-0.0344450846,
-0.0217262134,
0.0320290513,
0.0502555706,
0.0567394868,
0.0158664268,
-0.0020624045,
-0.0524251759,
0.0520393327,
-0.0567147247,
0.0165780764,
0.0176444892,
-0.0471441448,
0.0560341962,
0.0044243629,
-0.0124951424,
-0.0137642771,
0.0060497201,
-0.0258115567,
-0.0023599316,
-0.039753262,
0.0164745189,
0.0003545264,
-0.0024993445,
0.0136393122,
-0.0085000843,
-0.0764636025,
0.035085585,
-0.0300476067,
-0.0237364843,
0.0189928114,
0.0427763313,
-0.0078125801,
-0.020295959,
0.010348036,
0.0425773822,
-0.0231109038,
0.010257802,
-0.0890493467,
-0.0385840312,
0.0000998777,
0.0342580043,
0.0203842837,
-0.0335655212,
0.0467951,
0.058026202,
0.045951996,
-0.0508528389,
-0.045345135,
0.0088709565,
-0.0814411566,
-0.0206178389,
0.0446077883,
-0.0465164818,
0.0321484916,
-0.0058552576,
-0.0051384866,
-0.00885667,
0.0071927393,
-0.0153178619,
-0.0095783519,
0.0368913859,
-0.0417207405,
-0.0146128275,
0.0417614803,
0.001314666,
-0.030906748,
0.0220758747,
-0.0016791485,
-0.0029141374,
0.0227478296,
-0.0219575521,
-0.0281009395,
-0.0101432158,
0.0051569859,
-0.0355884582,
-0.0235071275,
-0.0203786455,
-0.0354183801,
0.0289477613,
0.0005444731,
0.0162610896,
-0.0237667877,
-0.0322840437,
-0.0025468899,
0.0010535953,
0.0289171059,
0.0573469885,
0.0376566574,
-0.0087129511,
0.0160397589,
0.0201563258,
0.0037596021,
-0.0000320561,
-0.0587981306,
-0.0245675668,
-0.0284808837,
0.0447825044,
-0.0150802573,
-0.0342682488,
-0.0151994359,
-0.0486146696,
0.0632239655,
-0.0460203141,
0.0545238741,
0.0277552549,
-0.0173953921,
-0.015832657,
0.0362246856,
-0.0295487661,
-0.058901038,
-0.0128113758,
-0.0309162252,
0.0490507819,
0.0270516165,
-0.0356386788,
-0.0131408516,
-0.0069076344,
-0.0253965724,
-0.0045450414,
-0.0380650461,
0.0209640954,
0.0372271277,
-0.0349124968,
0.0443052314,
-0.0106010111,
0.0164552517,
0.0043415795,
0.0040192101,
0.0457551107,
-0.0354442336,
0.0216167141,
0.0866649598,
0.0140530067,
-0.0658928901,
-0.0259310901,
0.022430703,
0.0324506201,
0.0532243215,
-0.002652884,
-0.0198117625,
-0.0377485678,
-0.0074291215,
-0.0414450355,
-0.037313018,
0.0520963259,
-0.0139481016,
-0.0582257323,
0.0374048389,
-0.0617249869,
-0.0274954494,
0.0353799649,
-0.0317969173,
-0.0192683786,
0.0376928039,
-0.0787449703,
0.008108383,
-0.0785613209,
-0.0206304435,
-0.0572653674,
0.032435108,
-0.0075606061,
0.0672899336,
-0.0170153528,
-0.0009637615,
-0.0146278972,
-0.0126718795,
0.010165751,
0.0302043092,
-0.0423140116,
-0.0308422633,
-0.0150482245,
0.0097066816,
-0.0406829119,
0.0095496727,
-0.0211012159,
-0.0634309649,
-0.1185387895,
-0.0042400425,
0.0061442261,
-0.0140703525,
0.0157911982,
-0.0020460295,
-0.0200236365,
-0.0161450636,
0.019584449,
-0.0053108246,
-0.0166623238,
-0.0026715219,
0.0077615972,
-0.0090338159,
-0.009519943,
-0.0239148978,
-0.038513314,
-0.0391995311,
0.0203186926,
-0.0045541362,
0.0272482708,
-0.0088253589,
0.0050772862,
-0.057945136,
-0.0050810762,
-0.0033481028,
-0.0770101398,
-0.0149869937,
0.0138511509,
-0.0216981005,
0.0640925393,
0.0759653524,
0.0517344326,
0.0070450539,
-0.0045874389,
0.0198090896,
0.0015461172,
-0.0098374858,
0.0676948428,
-0.0075554,
-0.0501989983,
-0.0329738259,
0.0091940453,
0.0276142769,
0.038114503,
-0.053188026,
0.031568557,
0.0320585743,
0.0049067782,
0.0286062621,
-0.0327521078,
-0.0212879814,
-0.0164940879,
0.0106055643,
0.0040838891,
0.0838538706,
-0.0022845981,
-0.0450667925,
-0.0193160772,
-0.0317672044,
-0.017995514,
0.0703135207,
-0.0254114531,
0.0569833741,
-0.0007290637,
-0.0206669848,
0.0065313019,
-0.0354601666,
0.0060187005,
-0.0349290222,
0.0438634902,
0.0178462826,
-0.0326519497,
0.004312878,
-0.0409606285,
-0.0191722214,
-0.0437721126,
-0.0440322571,
-0.0201094113,
-0.0059232474,
-0.0231236313,
-0.007751938,
0.0045414004,
-0.0260769818,
-0.0478331745,
-0.0232996251,
-0.011612243,
0.0101777771,
0.0197479352,
0.0195600837,
-0.0058904393,
0.0114557799,
0.0459290445,
-0.0607010201,
0.0195009708,
-0.0880870894,
0.0404797047,
-0.0245960522,
0.0022279909,
0.0062784753,
-0.0084785316,
-0.0094964216,
-0.0312129389,
0.0156195639,
0.0056424406,
-0.0280976165,
0.0812857747,
0.0104127293,
-0.0142634148,
0.0130320545,
-0.0227083359,
-0.0633412674,
0.0256589763,
0.0109844562,
-0.0287679564,
-0.0384001806,
-0.0080573959,
0.014228127,
-0.046824418,
0.0094997287,
0.0030587912,
-0.0705564022,
0.0345232747,
0.014608141,
-0.0028202767,
0.023960039,
0.0172307007,
0.0076210517,
-0.0137681086,
-0.040430285,
-0.0431428626,
-0.0084312167,
-0.0027871486,
0.0163108744,
-0.0157213844,
-0.0443476252,
0.0148051269,
0.0241483524,
0.0652968735,
0.0050043603,
0.0674904585,
-0.024518827,
-0.0614378452,
-0.0345831215,
-0.0668991953,
0.0124621568,
-0.0151546337,
0.0843811333,
-0.006900039,
0.005653691,
-0.0438538343,
0.001284169,
-0.016104551,
0.0461767726,
-0.0138408905,
0.0275673233,
-0.0297272783,
-0.0338694081,
-0.0644532144,
0.0446593277,
0.0143165216,
0.0312163047,
0.0299334917,
-0.0765063688,
-0.0189143792,
-0.0157433636,
0.0182723403,
-0.0549224764,
-0.0050935056,
-0.0039508287,
0.035703022,
-0.0143935587,
-0.0428262204,
0.0037773482,
-0.0304146595,
0.0021941415,
-0.0283170268,
-0.0489211716,
-0.0601562262,
0.0082093189,
-0.0462574288,
-0.0364467166,
0.048787497,
-0.0001083372,
0.0012971237,
0.0134797003,
0.0549285449,
-0.0869148895,
-0.0327179395,
-0.0013426731,
-0.0482306555,
-0.0526837967,
0.0670866594,
0.0668275207,
0.0130464956,
-0.0337952971,
-0.0049871863,
-0.0507561937,
0.0277004503,
0.0944397002,
0.0227462631,
0.027847534,
0.055171106,
0.0456899442,
0.0046791406,
0.0350871533,
-0.0111483065,
0.0559486263,
0.0231520925,
-0.0272674058,
-0.0356201828,
-0.014658683,
-0.0283844974,
-0.035718523,
0.0314554796,
-0.0036680435,
-0.0756591409,
0.07714995,
-0.0128936274,
0.0181905441,
0.0008112001,
-0.0109922076,
-0.0582581013,
0.0120286252,
0.0306765884,
-0.0555425994,
-0.045200184,
-0.0422932021,
0.0301157068,
-0.0136071211,
0.0313928165,
0.0830434114,
0.0147715556,
-0.0360865146,
0.0416643508,
-0.0504459366,
0.005254515,
0.0162180308,
-0.0184245575,
0.0261942651,
0.021645721,
0.0303785559,
0.0081015648,
-0.0377283581,
0.0552083738,
0.0211157687,
0.0470216312,
0.0576383136,
-0.0328601524,
0.0566097908,
-0.0012598441,
0.052754987,
-0.0051526511,
0.0061908918,
-0.001398467,
0.0310473628,
-0.0431852825,
-0.0648996234,
0.0049482421,
0.011081093,
0.0138424952,
-0.0442613773,
0.0294320285,
-0.0969910771,
0.0317261592,
0.019008847,
-0.0109398374,
-0.0066209626,
-0.0200429615,
0.0139752235,
0.0255303625,
0.0288918074,
0.0370266698,
-0.0315644071,
-0.0609514527,
0.0783672184,
-0.0607437529,
0.0313785151,
-0.0073941448,
0.0331575572,
0.0505479947,
0.0257396437,
0.0257163569,
-0.0204002112,
0.0490195714,
-0.0747348964,
0.0258791149,
-0.0294439793,
-5.518910345e-33,
-0.0060007088,
-0.0310326107,
-0.0657516196,
0.063631475,
0.0108657815,
-0.0015547281,
-0.0487067066,
0.009165898,
-0.0277393665,
0.0256373584,
0.0102240248,
0.0197975282,
-0.0138571011,
0.0047609084,
0.039448455,
-0.009205644,
0.0194480065,
0.0263907686,
-0.0045674578,
-0.0136775002,
0.0486607328,
0.0294750631,
0.0499372743,
-0.111838907,
-0.0168023333,
0.00246051,
-0.0244265161,
0.0117466524,
0.0637918413,
0.0597842038,
0.0293948501,
0.0912833363,
0.0402977206,
0.0559835173,
0.0205063038,
0.0081730699,
-0.0295058507,
-0.0232702978,
-0.0063093123,
-0.0046318974,
-0.0278373938,
-0.0171856172,
0.0765639693,
-0.0068393406,
-0.031879846,
-0.0161848068,
0.0470620096,
-0.0180423837,
0.0509731062,
-0.0611412339,
-0.0321815647,
-0.003132367,
-0.0337864198,
-0.0141608147,
0.0045162216,
0.0297289379,
0.0133980988,
0.0159801971,
-0.0612751059,
0.0223175343,
-0.1006179452,
-0.0426098704,
-0.0501077101,
-0.0256135128,
0.0166679919,
0.0586659797,
0.0124816168,
0.0342283882,
-0.0274913423,
-0.0068253009,
0.0252489075,
0.0157895386,
0.067670323,
0.036758963,
0.0197861511,
-0.047731474,
-0.0099944966,
0.0101577444,
0.0705777928,
0.0006774673,
0.021527471,
-0.0194062684,
-0.0405367948,
0.0117690237,
0.0485206321,
-0.0128550008,
0.0177448355,
-0.0685376227,
-0.0197192077,
-0.0003413423,
-0.0110345837,
0.0088531598,
0.017601274,
-0.0392267331,
0.003212512,
0.0007308859,
-0.0012052709,
0.079482615,
0.0253023691,
0.0056203357,
0.0028949401,
-0.0183494911,
0.041011069,
0.0378532633,
0.0445562862,
-0.0207358636,
0.0758390352,
0.0314698145,
-0.0318016298,
0.0110363718,
0.0183237512,
0.028745193,
-0.0172143374,
0.0182800405,
-0.027530266,
0.069681108,
-0.0231792685,
-0.0557380468,
0.0458853245,
0.0611875281,
-0.0456245877,
-0.0223678872,
-0.0017407972,
0.0268245209,
-0.0606655031,
-0.0431639217,
0.0058836578,
-0.0784988552,
-0.0301936958,
-0.0478237979,
0.0026242733,
-0.0414884649,
2.58e-7,
0.0538492277,
0.0639081448,
0.0342135355,
0.0277287532,
0.0725222826,
0.0533382259,
0.0037732308,
0.0227445494,
0.039320536,
0.0131792566,
0.0282874051,
0.0061069671,
-0.0023031868,
0.0086505208,
-0.0514321588,
0.0475699008,
-0.0583854429,
0.0558314472,
-0.0062631802,
0.0379248373,
0.0373615436,
0.0224608742,
0.0272252169,
-0.02983528,
0.020278303,
-0.0202634633,
0.0248169955,
-0.0147223938,
0.0670249313,
-0.022465935,
-0.0046740789,
-0.0315537155,
0.0221572686,
-0.0042481748,
0.0311012045,
-0.0399483219,
-0.0090928385,
0.0339650773,
-0.0507989824,
-0.0076839244,
0.0225135535,
0.031268917,
-0.009864389,
-0.0360090658,
0.0448479354,
0.0177748855,
-0.0486702472,
-0.0437360965,
-0.0226125866,
-0.0161831081,
0.1145249903,
0.0238216761,
-0.0124233617,
0.0228583943,
0.0380271077,
-0.0085551897,
-0.0075446139,
-0.0701217353,
-0.0241530854,
0.01180002,
-0.0515412576,
0.0153264832,
0.0348835029,
-0.0097033791,
0.0265076216,
-0.0843262076,
0.0301562827,
1.879802697e-34,
0.0159114227,
0.0130905062,
0.0243672971,
-0.0202241912,
-0.0149795264,
0.0374225862,
-0.0140446527,
-0.0129678519,
0.0134957004,
0.0057450943,
-0.0070028901
] |
http://arxiv.org/abs/2401.17139v1 | http://arxiv.org/pdf/2401.17139v1 | Large Language Model Evaluation via Matrix Entropy | 2024-01-30T16:19:55 | 2024-01-30T16:19:55 | Lai Wei, Zhiquan Tan, Chenghai Li, Jindong Wang, Weiran Huang | cs.LG | cs.LG, cs.AI, cs.CL, cs.IT, math.IT | Large language models (LLMs) have revolutionized the field of natural
language processing, extending their strong capabilities into multi-modal
domains. Thus, it is vital to define proper and diversified metrics for the
evaluation of LLMs.
In this paper, we introduce matrix entropy, a novel metric rooted in
information theory and geometry principles to quantify the data compression
proficiency in LLMs. It reflects the model's ability to extract relevant
information and eliminate unnecessary elements, thereby providing insight into
the language model's intrinsic capability. Specifically, we demonstrate its
applicability in both single-modal (language) and multi-modal settings. For
language models, our findings reveal that the matrix entropy of representations
follows a scaling law type reduction when the model scales up, serving as a
complement to the traditional loss scaling law. For the multi-modal setting, we
also propose an evaluation method based on matrix entropy for assessing
alignment quality and we find that modern large multi-modal models exhibit
great alignment performance. | Large Language Model Evaluation via Matrix Entropy: Large language models (LLMs) have revolutionized the field of natural
language processing, extending their strong capabilities into multi-modal
domains. Thus, it is vital to define proper and diversified metrics for the
evaluation of LLMs.
In this paper, we introduce matrix entropy, a novel metric rooted in
information theory and geometry principles to quantify the data compression
proficiency in LLMs. It reflects the model's ability to extract relevant
information and eliminate unnecessary elements, thereby providing insight into
the language model's intrinsic capability. Specifically, we demonstrate its
applicability in both single-modal (language) and multi-modal settings. For
language models, our findings reveal that the matrix entropy of representations
follows a scaling law type reduction when the model scales up, serving as a
complement to the traditional loss scaling law. For the multi-modal setting, we
also propose an evaluation method based on matrix entropy for assessing
alignment quality and we find that modern large multi-modal models exhibit
great alignment performance. | [
-0.0185893476,
0.0597656742,
-0.0080490215,
0.0266948827,
-0.0719614327,
0.0301527232,
-0.0142851816,
0.0329438448,
-0.0467917547,
-0.0572710633,
-0.0372099429,
-0.0738010108,
0.0615870655,
-0.0297936574,
0.0276954751,
-0.0212064143,
0.028030334,
-0.0226306822,
-0.0437720194,
0.0137532316,
-0.0174633525,
-0.0317983665,
0.0394947119,
0.0397334918,
-0.0011048667,
0.0002959738,
-0.0016303612,
0.011099318,
0.0045200777,
-0.0086852061,
-0.0631031469,
0.0571441203,
0.0235958491,
0.0382603519,
0.0000018431,
-0.0477504879,
-0.0104970867,
0.0268969219,
0.0151941096,
0.0459685698,
0.07389725,
-0.0623515919,
-0.0274047125,
-0.0346102566,
0.0053788498,
-0.0516326055,
0.0727935284,
0.0372321308,
0.0209966395,
0.0568799973,
-0.0179097895,
-0.0755628869,
0.0628459007,
-0.0155702345,
0.0026229646,
-0.0111819608,
0.0478396453,
0.0126538863,
-0.046809949,
-0.0163531806,
-0.0312214866,
0.0405434705,
-0.0026709959,
0.0178107675,
-0.0110768871,
0.0158968605,
0.0338080525,
-0.0024651003,
0.0131752072,
0.0756597817,
-0.0050024176,
-0.0081592575,
0.0033924649,
0.005163461,
0.0006625589,
0.0093886759,
-0.0566691272,
0.0250294264,
0.0407033302,
0.0338662826,
0.0050120456,
0.0261367131,
0.0364706106,
-0.0096653523,
-0.0136864791,
0.0480484329,
0.0425049774,
-0.0501116775,
-0.0263956022,
-0.0297379382,
-0.0153165292,
-0.0174148139,
-0.0389743261,
-0.0003706759,
0.0260208435,
0.021995537,
-0.0402369872,
-0.0584889241,
0.0474009737,
-0.0784354955,
0.0553781129,
0.0171270799,
0.0397845805,
0.0366790406,
-0.0207046773,
0.0845642239,
-0.0179871991,
0.0493278205,
-0.0452784933,
-0.0141625619,
-0.0233028308,
0.014205615,
-0.0024263812,
0.0210038982,
0.0170792881,
-0.0416507274,
-0.0344396383,
0.0278520193,
-0.0110333264,
0.0260609817,
0.0199075621,
-0.0213727653,
-0.0062818867,
-0.0401595905,
-0.0134864943,
0.0321900919,
0.0219125841,
0.0092527606,
-0.0152823897,
-0.0041978476,
0.0002413885,
0.0428976566,
-0.0057943487,
-0.0179541837,
0.0178327803,
0.0716257766,
0.0217926726,
-0.0592598431,
-0.0119957691,
0.0122795189,
-0.0521068349,
-0.0653212741,
-0.0169615448,
-0.0399712324,
0.0390411876,
-0.0026571441,
-0.0186242275,
-0.0436636806,
-0.0132925091,
-0.0104283728,
-0.0533181131,
0.049448438,
0.0136338696,
0.0278789345,
0.0171581414,
-0.0143666612,
0.071271129,
0.0719566569,
0.0568519868,
0.0222680327,
0.0503782779,
-0.0446090959,
-0.0228309967,
0.0371199176,
0.0002329143,
0.0336740799,
-0.0558302701,
-0.0160455219,
0.0424908772,
0.0245073326,
0.047824271,
0.0923515037,
-0.0141172055,
-0.0323684886,
0.0722934306,
0.0613298677,
0.0474264026,
0.0394383781,
0.0177245997,
-0.0122535126,
-0.0019373643,
0.0214369036,
-0.0198671594,
-0.0024302423,
-0.0237818472,
0.0407275669,
0.1019166484,
0.0558329299,
0.0009394026,
-0.0219269954,
-0.0351262651,
-0.0385321863,
0.0143143591,
-0.0339814685,
0.0542020909,
-0.0372639522,
-0.031068882,
0.0480320193,
0.0200116485,
-0.0589900166,
-0.0015566633,
-0.0202348158,
-0.0118231075,
0.0312736705,
0.0400085375,
0.0088881738,
0.0300484486,
-0.0136774294,
-0.0390448831,
0.0020186959,
0.0201133527,
0.0022783927,
-0.0099682761,
-0.0525014624,
0.0013746185,
0.0265167262,
0.0330803357,
-0.0074843504,
-0.0435593873,
0.0014209372,
0.0216636285,
-0.0266324915,
-0.0134486966,
0.0222905856,
0.0133263338,
0.0021708461,
0.0040664962,
-0.0068110181,
0.0573980398,
0.0037363812,
-0.0414569266,
0.0128587903,
0.047851447,
-0.0688380748,
0.0003978549,
-0.0115395309,
-0.0348286182,
-0.0180944223,
0.0558032617,
-0.0147926705,
0.0287083928,
0.0317798443,
-0.0032019978,
-0.0397738889,
0.0189965591,
-0.0544630289,
0.0345863923,
-0.0566358976,
0.0363929644,
-0.0459768921,
0.0180846453,
-0.0599401109,
0.054153651,
-0.0343826562,
0.0580387264,
0.0075670574,
-0.0425271168,
0.0197297912,
0.0042443359,
-0.0310933273,
0.0679743811,
-0.0150081674,
0.0152353244,
0.0184664913,
-0.0335997716,
0.0216063671,
-0.0474156328,
-0.0634477213,
0.0218506251,
-0.0199274868,
0.0322936848,
0.040663939,
-0.0004141972,
0.0074201897,
0.0292499233,
-0.0344656557,
0.0201349091,
0.0094024017,
-0.0044951774,
0.0356192142,
0.0055581136,
-0.0447341986,
-0.025862867,
-0.0314459875,
0.0013199131,
0.0880508199,
-0.0130611425,
0.0174086243,
-0.0629181638,
0.062885344,
-0.0264608581,
0.0020762265,
0.0073205712,
-0.0590717047,
-0.0462307222,
-0.0255700164,
0.0355562195,
0.0247301906,
0.0325681008,
-0.0084800767,
0.0904476643,
-0.004220875,
-0.0185471419,
-0.0250431374,
-0.0473279767,
0.0117272772,
-0.0577918738,
-0.0782267898,
0.0097624902,
0.0575515293,
0.0275863633,
0.016066581,
-0.0479258709,
-0.0648794919,
-0.0326060988,
-0.0119417161,
-0.0295358989,
-0.0126326811,
-0.0173155311,
0.0359706469,
-0.0145342406,
-0.0581456758,
0.0360048153,
-0.0740422606,
0.0122054927,
-0.0040365406,
-0.0220814273,
-0.0607330315,
0.0664027929,
0.0263985526,
0.0398720279,
0.0113845337,
-0.012207902,
-0.0231834985,
-0.0031820573,
0.0359922126,
-0.0362936519,
0.0292609073,
-0.0121929729,
0.0159019735,
-0.0235154666,
-0.0061382838,
0.0274709426,
-0.04804755,
0.0226017088,
-0.0090467939,
0.051167801,
-0.0471027792,
-0.0168001018,
0.0560408346,
0.0320244282,
-0.0022597851,
0.0025968461,
0.0029482448,
-0.0156595334,
0.0463819914,
0.048659537,
0.0490900762,
0.0411638878,
-0.0217576157,
0.0029202232,
0.0331425183,
0.0016197631,
0.0701421574,
-0.0346715823,
-0.0080417544,
0.0534752049,
0.015305981,
-0.0379656963,
0.0298562888,
0.0207282435,
-0.0212530959,
0.0239150506,
0.1230550706,
0.0244262014,
-0.0536050573,
0.0103371162,
0.012659586,
0.0687664524,
-0.0119742472,
-0.0110292593,
-0.0497044437,
-0.0240044221,
0.0095318845,
-0.0226919688,
-0.0037806812,
-0.0418224335,
0.005985029,
-0.0816910863,
0.0024325235,
0.0283664484,
0.0182100255,
-0.084241271,
-0.0657565594,
0.0138127478,
-0.0408854112,
0.0167089887,
-0.0133993905,
-0.0314646922,
-0.0219777301,
0.018935604,
-0.0255352017,
-0.0729648396,
0.0007926143,
-0.0135413818,
0.0819706172,
0.0361765884,
0.0413038097,
-0.0140109826,
-0.0734238699,
-0.0154648218,
-0.01703454,
-0.0258612204,
0.0141140362,
0.0276493188,
0.0048523382,
-0.0005969062,
-0.0095715337,
-0.0237467941,
-0.0245756377,
-0.0021872022,
-0.0538511053,
0.058759816,
0.0086070281,
0.00950603,
-0.0020950208,
-0.0280594733,
-0.0322701894,
0.1116690785,
0.038333308,
0.0095173167,
-0.043178726,
-0.0047254134,
0.016941905,
0.0149355866,
-0.0391811617,
0.0052653467,
0.0013790461,
-0.0513277054,
-0.045199085,
0.0141440732,
-0.0154610211,
0.0204577837,
-0.0185671337,
-0.0150961,
-0.0182155855,
0.0008836037,
-0.0400137119,
-0.0223059021,
-0.0184170455,
0.0119706634,
-0.0640178397,
0.0659202337,
0.0241080392,
0.0231440216,
-0.0230117999,
-0.0022044333,
-0.1031874195,
-0.0591228828,
-0.0237471219,
-0.0011237448,
0.0210427195,
0.0324445777,
-0.0091085834,
-0.0597485304,
0.0110272327,
0.0126107149,
-0.0587798804,
0.0526289158,
0.0194972474,
-0.0341941491,
0.0098404624,
0.0482145585,
-0.0092194621,
-0.0234284177,
-0.0351834595,
0.0248365197,
-0.0046203579,
-0.0832780227,
0.0231699627,
0.032192938,
0.0253249723,
-0.0183796957,
-0.0458442681,
0.0192121994,
0.0061600092,
-0.0338937305,
-0.006742598,
-0.0267904121,
-0.0164892934,
0.0597374924,
-0.0028396659,
-0.0012550117,
0.0540132076,
0.0162893254,
0.0036362119,
-0.0042335098,
0.0022540614,
-0.0409655869,
-0.0035822608,
0.0156017682,
0.0155224176,
-0.0249069817,
-0.0123211397,
-0.0030404523,
0.0001125217,
-0.0258878823,
0.0170272496,
0.0072915237,
0.0047769081,
-0.0218444616,
0.0686888993,
-0.0556928143,
-0.0152367875,
-0.0412871502,
0.0098171216,
0.0469623804,
0.0278198179,
0.066200316,
-0.0389463603,
-0.0032417451,
-0.0187329464,
0.0250977166,
0.0093582571,
-0.0000080679,
0.0433696583,
0.0032757677,
-0.0334285088,
0.0183539595,
0.0724406391,
0.0183406342,
0.0278413035,
-0.0431038775,
-0.0208097212,
-0.0141446032,
-0.0013033664,
-0.0181412213,
0.0286812503,
-0.0331697874,
-0.0123431301,
-0.0883510411,
-0.0662935376,
-0.0330117866,
0.040408086,
0.0079752561,
0.0041716099,
0.0150056537,
0.0566636622,
-0.0138959633,
0.0079286154,
-0.0235566776,
-0.0361469537,
-0.0121587347,
-0.0057083117,
-0.012662366,
-6.065688603e-33,
-0.0295100901,
0.017548319,
-0.0152960764,
-0.0282379389,
-0.036927525,
-0.0008714127,
-0.0170451719,
0.0205694325,
-0.016195301,
0.002819424,
0.037790522,
-0.0254149102,
-0.0142693166,
0.0033010936,
-0.0027926997,
-0.0216529574,
0.0320254453,
-0.0175458565,
-0.0179600306,
0.0142167853,
0.0294088461,
-0.0153506026,
0.0411914997,
-0.0381491818,
0.0251855645,
0.0029156003,
-0.0423822813,
0.0087132538,
0.0445549302,
0.0474922135,
-0.0482552946,
0.0135418056,
0.0348030291,
-0.0685356855,
-0.0079236692,
0.021870425,
-0.0344198309,
-0.0162036847,
0.0329667032,
-0.0103066256,
-0.0289442502,
-0.0565230437,
0.0481538102,
0.0072830762,
-0.0556842461,
-0.0168808755,
0.0343491919,
-0.0189151298,
-0.0436716042,
-0.040806856,
-0.0299973525,
0.037899185,
-0.0288919006,
-0.0031012718,
0.0440320484,
0.0272134952,
-0.0072233942,
0.0328930467,
-0.0245194361,
0.0350467227,
0.0046196259,
0.0537617467,
0.0524310023,
-0.0041834135,
0.0126641002,
0.0654931664,
0.0128812278,
-0.038930364,
-0.0218853597,
0.0063341088,
0.0340155922,
0.0301355049,
0.0132526429,
0.0716848522,
0.0386925302,
0.012077651,
-0.0585716292,
0.03949764,
-0.0052673859,
0.0274747144,
0.0005387419,
0.028959699,
-0.0385093577,
-0.0091615822,
-0.0523602106,
-0.0174220875,
-0.0082151592,
-0.0278777033,
0.0049358942,
-0.022963766,
-0.0033078871,
0.0166479778,
0.0417935215,
-0.0469187312,
-0.0252995063,
0.0473873094,
-0.0149547551,
0.0575834662,
0.0115528153,
0.0049643428,
-0.0379759148,
-0.0127372351,
0.010336671,
0.0026550225,
0.0166061297,
-0.0155548221,
0.0082773957,
0.0349261276,
-0.0614593811,
-0.0139223551,
0.0018325144,
-0.0663485527,
0.0112419091,
-0.0190504044,
0.033517994,
-0.007793603,
-0.017981356,
0.0360969752,
0.0119839991,
0.042507261,
-0.0167229008,
-0.0135133397,
0.0031132021,
0.0390616469,
-0.0084168399,
-0.0021326833,
-0.0284256749,
-0.0571603365,
0.0936702937,
-0.0207128897,
-0.0205001943,
0.0058155777,
2.592e-7,
0.0516777895,
0.046353016,
-0.0125781549,
-0.0299957823,
0.009905423,
0.0604470521,
-0.0200261064,
0.0291604716,
-0.0168734044,
-0.0310533233,
0.0190649983,
0.0014263602,
0.0272447541,
-0.00411293,
-0.0487312824,
0.0316852741,
-0.0601039752,
-0.0073165796,
-0.0383039564,
0.0382605977,
0.0499415658,
0.1349466294,
0.0096796108,
-0.0155417537,
-0.0028767411,
-0.0402916297,
0.0507259481,
-0.0774420127,
0.0019360312,
0.0367893018,
0.0357253887,
-0.0353566036,
-0.004890915,
0.0052016727,
-0.0037956797,
0.0181375314,
0.0049015237,
0.0599689372,
0.0078019062,
0.0146005396,
-0.0057777921,
-0.0416874588,
-0.0335117206,
-0.0666459203,
0.0794570968,
0.0112498319,
-0.0494852439,
-0.0568923913,
-0.0089266393,
0.0480448194,
0.0430117063,
0.0085585499,
-0.0104346368,
0.0191656258,
0.0145183299,
-0.0192345455,
0.0053650164,
-0.0714528263,
-0.0146493912,
0.0204745308,
-0.0377310291,
-0.014059782,
-0.0238525551,
-0.0319958553,
0.087228395,
-0.1023839414,
0.0033869408,
2.896452289e-34,
0.0012274034,
-0.0205148458,
0.0260428693,
0.0312787704,
-0.0104613733,
-0.0259537306,
-0.0082615372,
0.0185193606,
-0.0400249623,
-0.126559332,
0.0143862143
] |
http://arxiv.org/abs/2401.17099v1 | http://arxiv.org/pdf/2401.17099v1 | MT-Ranker: Reference-free machine translation evaluation by inter-system ranking | 2024-01-30T15:30:03 | 2024-01-30T15:30:03 | Ibraheem Muhammad Moosa, Rui Zhang, Wenpeng Yin | cs.CL | cs.CL, I.2.7 | Traditionally, Machine Translation (MT) Evaluation has been treated as a
regression problem -- producing an absolute translation-quality score. This
approach has two limitations: i) the scores lack interpretability, and human
annotators struggle with giving consistent scores; ii) most scoring methods are
based on (reference, translation) pairs, limiting their applicability in
real-world scenarios where references are absent. In practice, we often care
about whether a new MT system is better or worse than some competitors. In
addition, reference-free MT evaluation is increasingly practical and necessary.
Unfortunately, these two practical considerations have yet to be jointly
explored. In this work, we formulate the reference-free MT evaluation into a
pairwise ranking problem. Given the source sentence and a pair of translations,
our system predicts which translation is better. In addition to proposing this
new formulation, we further show that this new paradigm can demonstrate
superior correlation with human judgments by merely using indirect supervision
from natural language inference and weak supervision from our synthetic data.
In the context of reference-free evaluation, MT-Ranker, trained without any
human annotations, achieves state-of-the-art results on the WMT Shared Metrics
Task benchmarks DARR20, MQM20, and MQM21. On a more challenging benchmark,
ACES, which contains fine-grained evaluation criteria such as addition,
omission, and mistranslation errors, MT-Ranker marks state-of-the-art against
reference-free as well as reference-based baselines. | MT-Ranker: Reference-free machine translation evaluation by inter-system ranking: Traditionally, Machine Translation (MT) Evaluation has been treated as a
regression problem -- producing an absolute translation-quality score. This
approach has two limitations: i) the scores lack interpretability, and human
annotators struggle with giving consistent scores; ii) most scoring methods are
based on (reference, translation) pairs, limiting their applicability in
real-world scenarios where references are absent. In practice, we often care
about whether a new MT system is better or worse than some competitors. In
addition, reference-free MT evaluation is increasingly practical and necessary.
Unfortunately, these two practical considerations have yet to be jointly
explored. In this work, we formulate the reference-free MT evaluation into a
pairwise ranking problem. Given the source sentence and a pair of translations,
our system predicts which translation is better. In addition to proposing this
new formulation, we further show that this new paradigm can demonstrate
superior correlation with human judgments by merely using indirect supervision
from natural language inference and weak supervision from our synthetic data.
In the context of reference-free evaluation, MT-Ranker, trained without any
human annotations, achieves state-of-the-art results on the WMT Shared Metrics
Task benchmarks DARR20, MQM20, and MQM21. On a more challenging benchmark,
ACES, which contains fine-grained evaluation criteria such as addition,
omission, and mistranslation errors, MT-Ranker marks state-of-the-art against
reference-free as well as reference-based baselines. | [
0.0335578397,
0.0465316214,
0.0106175337,
0.0434240848,
0.0046031899,
0.0270527881,
0.0203625839,
0.0008187392,
0.0056137866,
-0.0398839861,
-0.0742910653,
-0.0077560865,
0.034827657,
-0.0307287816,
0.0455775559,
-0.0295135975,
0.016505491,
0.0007495064,
-0.0190436151,
0.0244530402,
-0.0128645338,
0.0044657765,
0.008016645,
0.0298771616,
0.0102505842,
0.0199699309,
0.0304863993,
0.0350486785,
0.0802154019,
-0.011422554,
-0.0242929123,
0.0385284498,
0.0020881973,
0.0673488081,
0.0000019251,
-0.0300784875,
-0.0409021117,
0.0158093181,
0.0096130595,
-0.003402909,
0.0129203768,
-0.0260400288,
-0.0164363831,
-0.0018073037,
-0.0372536965,
-0.0149997724,
0.0415299013,
0.0010002516,
0.094796963,
0.0628465712,
-0.0117852064,
0.0081446944,
-0.0396021046,
-0.0436904393,
0.0060639242,
-0.0687671155,
-0.0435828976,
0.0340414792,
-0.0455327369,
0.0323871933,
0.0277544968,
0.0401985273,
0.0009124161,
-0.0153090814,
0.0146248359,
-0.0059086252,
-0.073989585,
-0.0201640148,
0.0141692217,
0.1039663181,
-0.0256743357,
-0.0060059419,
-0.0152954804,
-0.0005048444,
0.0026052194,
-0.0324794166,
-0.0108973216,
0.0051617138,
0.0062124953,
0.0129571781,
-0.0162794106,
0.0723659024,
0.0291191228,
0.0115482826,
-0.0265517682,
0.0154800275,
0.0136609627,
-0.0344963521,
0.0281247795,
0.0423102342,
-0.0272484161,
-0.0021807393,
0.0260331705,
-0.0321759991,
-0.0653310046,
-0.0007055519,
-0.0082613109,
-0.04958231,
0.0141801033,
-0.0755508319,
0.0432002544,
0.0578912273,
-0.0262840185,
0.0598361082,
-0.0467349291,
0.1025964171,
0.0065620951,
0.0182468332,
-0.0646801144,
-0.0025595732,
-0.018825911,
0.0455784276,
0.0032116377,
0.0037032682,
-0.0694887862,
0.0067590666,
-0.0461207218,
-0.0150957089,
-0.0144126918,
0.019053502,
0.0508557707,
-0.009185696,
-0.058822237,
0.0061382405,
-0.0209283289,
0.0296989121,
-0.0145178149,
-0.0205847435,
-0.0116446353,
-0.0263174083,
-0.0215031244,
0.0324558839,
-0.0318987332,
-0.0153375175,
0.0142033873,
0.1033948064,
-0.0336013734,
-0.0124157919,
0.0407486148,
-0.0029802592,
-0.0178555604,
-0.0685603768,
-0.0036805489,
-0.0560372919,
0.0552998483,
0.0241223257,
0.019961698,
0.0009269605,
0.0270341523,
0.0091224611,
-0.0105482927,
0.0507477894,
0.0206801277,
-0.0379648879,
-0.0137357712,
0.0051185931,
-0.0402227864,
0.0513690449,
0.0463863835,
0.048484765,
0.022499755,
0.0416835286,
-0.0218318999,
0.0313288681,
-0.0180587545,
0.0247367676,
-0.0054799989,
0.0072924378,
-0.0083988803,
0.0294872727,
0.0180900749,
0.0596974865,
-0.0209746677,
0.0159813501,
0.0491620935,
-0.0177282989,
0.0029460315,
0.0239128172,
0.0801499411,
0.020105999,
-0.0018878299,
0.0182073768,
-0.0393966399,
0.0376851782,
-0.0003248053,
-0.0036676142,
0.0768572316,
0.0174367055,
0.0081038065,
-0.0863008052,
0.0042642998,
-0.0237704776,
-0.0022768227,
-0.0303961877,
0.0431896858,
0.0365995653,
-0.0609391704,
0.0133845042,
-0.0197373834,
-0.0622683018,
0.000870792,
-0.0292303711,
-0.0307857189,
0.066805236,
0.0297362376,
-0.0214571785,
-0.0179815851,
-0.0003056717,
-0.0229407437,
0.0518186688,
0.0352492929,
0.0020734584,
0.0485442765,
-0.0478863157,
-0.0340649933,
0.0109457439,
0.0070849741,
0.0339928046,
-0.0131565174,
0.0114407465,
0.0455764756,
-0.0636143833,
-0.091088146,
0.0231448337,
0.0363211147,
-0.0322584137,
0.0394891314,
0.005241958,
0.0293542705,
0.021872662,
-0.0190706588,
0.0143820401,
0.0256507341,
-0.032690879,
-0.0334460996,
0.0223880894,
-0.0415830724,
-0.0514451228,
0.0029778273,
-0.0606590696,
-0.0271179806,
0.032518357,
-0.0766538978,
-0.0352633372,
-0.0229124147,
-0.0296834316,
-0.025147019,
-0.0184156299,
0.0254057236,
0.021275688,
-0.0334924944,
-0.048174955,
0.0653183684,
0.0018326773,
0.0378353857,
0.0319593213,
-0.0288018808,
0.0269006658,
-0.0023541851,
-0.0398838297,
0.0105797527,
0.0136099821,
0.0313578174,
0.031953875,
0.0368931182,
-0.0467185713,
-0.0271145646,
-0.0771647096,
0.0264543705,
-0.0268027317,
0.0380636305,
0.0285681561,
0.0172455944,
0.0215937514,
-0.0166515708,
0.0023647,
-0.0291508269,
0.014596642,
0.0045824838,
0.037401326,
0.0187191181,
0.0647559762,
-0.0550766289,
-0.0080288369,
0.0038233411,
0.0406332053,
0.0286955386,
0.0002632357,
-0.0659134388,
0.0322276168,
-0.0705163553,
-0.0114139207,
0.0087188855,
-0.0599100888,
0.0289740097,
-0.0028484794,
0.0339269042,
0.0550005846,
0.0028632374,
-0.0055084885,
0.0528738908,
0.0124465656,
-0.0671935603,
-0.0029363469,
0.0179202687,
-0.0407753512,
0.0101526482,
-0.0417425111,
-0.0155469421,
0.019311605,
0.0206954163,
-0.005265899,
-0.0495518781,
0.0471161529,
-0.0244502872,
-0.0287515093,
-0.0474098437,
0.0251510069,
0.0158821642,
0.0203621835,
-0.0233929008,
-0.0527404621,
0.1003613025,
-0.0410554931,
-0.0099754864,
-0.0564859621,
0.0254764017,
-0.0385050476,
0.0165584069,
-0.0200646464,
0.0165925995,
-0.0479270257,
0.0001383944,
-0.0440210588,
0.019756116,
0.0293750651,
-0.0411848053,
0.003898877,
-0.0695190504,
-0.0185464285,
-0.0301509611,
-0.0144367628,
-0.0246108193,
0.011218179,
-0.0148922512,
0.0027933053,
0.0242256746,
-0.0068159942,
-0.00441175,
-0.0000725328,
0.0539092943,
0.0467133969,
-0.007034875,
-0.0179464538,
-0.0146586196,
0.0207257867,
-0.0211120192,
-0.0069805598,
0.0230059326,
-0.026413301,
0.0157266725,
0.0105849029,
0.0280362852,
-0.0272583012,
-0.027934676,
-0.0296465307,
0.0277184509,
0.0199349113,
-0.0366516635,
-0.0190733336,
0.0402827859,
-0.044022318,
0.0020248932,
0.0980323479,
0.0357513241,
-0.0230752807,
-0.0268824417,
0.0016754935,
0.0229507908,
0.0233211052,
0.0221140366,
-0.0690206885,
-0.0178125203,
0.0007141795,
0.0304455534,
0.0101508768,
0.0060140528,
-0.0055021681,
-0.0748659745,
0.0197017305,
0.0369156748,
-0.0478236526,
-0.0198601596,
-0.0260921549,
-0.018750295,
-0.0524843819,
-0.0072389827,
-4.86e-8,
0.0268666688,
0.0190428458,
-0.0102511831,
-0.0028144643,
-0.038857244,
0.0046986928,
-0.0611977093,
0.0040517277,
0.0806958303,
0.0590118691,
0.0148263099,
-0.0421314202,
0.0133758299,
-0.0535666682,
0.0108604301,
-0.0160176493,
0.0341764614,
-0.0171004087,
-0.0175682679,
-0.001893739,
-0.0344236679,
-0.0054812562,
0.0482873879,
-0.0177729055,
0.0420037843,
-0.0054608067,
0.0157462806,
0.0251052137,
-0.0066829789,
-0.0065749022,
0.0534808598,
0.0055578062,
-0.0278713778,
0.0044849585,
-0.0341362767,
-0.0055440059,
-0.0633079335,
-0.0398607999,
0.0025242558,
-0.0473983996,
-0.0038497073,
-0.1179703847,
-0.0613359511,
-0.00081245,
0.0197936688,
-0.0126383658,
-0.003453078,
-0.0365651995,
-0.0026081663,
0.0058705052,
0.0345890261,
-0.0071547641,
-0.0003148692,
-0.0222822763,
0.0195066743,
0.0349156037,
-0.001788004,
-0.0231336597,
0.045854114,
-0.0114279101,
-0.043644961,
-0.0547261722,
0.109595336,
0.0193049274,
0.0150472382,
-0.0370262824,
0.0058732801,
0.002383559,
0.0212329626,
-0.0092280107,
-0.0044295602,
0.0668907091,
-0.012000259,
-0.021492999,
0.0440590158,
-0.0459830277,
0.0002444159,
-0.0124966297,
0.0441246219,
-0.0077969781,
-0.033330556,
0.0291882381,
0.0140084149,
0.0503965318,
-0.0466643721,
-0.0238242559,
-0.0072772317,
0.0197288264,
-0.0365079567,
-0.0472143553,
-0.053767655,
0.0105232038,
0.0284992531,
0.017633291,
-0.0220194515,
0.0159466863,
-0.0284635555,
-0.0119574768,
0.0004161197,
-0.0102815386,
0.0090205986,
0.0088358475,
-0.0244361665,
0.0129539454,
-0.0463340022,
-0.0186595172,
0.0117304148,
0.0243650489,
-0.0323235244,
0.0235646833,
0.0137252165,
0.0043568262,
0.0272448212,
0.0699574947,
-0.0264781956,
-0.001229789,
-0.047332827,
0.0123858806,
0.0771816671,
-0.0008230031,
0.1392697096,
0.0276477896,
-0.0026121186,
-0.001494924,
0.0398215279,
0.0325600393,
-0.0125539145,
-0.0532251857,
0.0186331831,
0.0167783983,
0.0726228282,
0.0593472607,
-0.0103870416,
-0.0311591141,
-0.0220801551,
-0.0270602256,
-0.0514330082,
-0.0019409714,
-0.0422095582,
-0.0275423974,
0.0087274248,
-0.0314203463,
-0.0391586758,
-0.0526739284,
0.0135125695,
0.0488200523,
-0.0027100027,
-0.0112478537,
0.0145982699,
0.0690026283,
-0.0383748896,
0.0088419933,
-0.0465147309,
-0.0370226763,
0.0085153524,
-0.0222367384,
-0.0450415164,
-6.126172931e-33,
-0.051852569,
-0.0478047915,
-0.0236596763,
-0.0462759323,
-0.0126952985,
0.0001333384,
-0.0222915392,
-0.0135993389,
0.0084668789,
0.0270365663,
0.0075674974,
-0.0090056006,
0.0298866462,
-0.033011429,
0.006556849,
-0.0019699258,
0.0563707612,
0.0140480958,
-0.0695562288,
0.0062570735,
0.0473315679,
0.001400223,
0.0770536587,
-0.0687832683,
0.0780816451,
-0.0181645527,
-0.0250633322,
0.0121196462,
0.0006304022,
0.0311511327,
0.0175233874,
0.0217325706,
0.0262883939,
-0.0405552834,
-0.0143660549,
-0.0211202875,
-0.080220446,
-0.002060977,
0.0157526955,
-0.0114840288,
-0.0304636043,
-0.0071279141,
0.0530754738,
-0.0431865007,
-0.0071436861,
-0.0025222984,
0.0335861146,
-0.0617954135,
0.0021670936,
-0.0241935682,
-0.0413214155,
0.023502104,
-0.0217367299,
-0.0188710373,
0.015096968,
0.0986175835,
-0.0216596313,
0.0061994437,
-0.0699889883,
0.0116855558,
-0.0592187457,
0.0477315038,
0.0835654214,
-0.0191019122,
-0.0439544655,
0.011432766,
-0.0005844697,
-0.028344566,
0.0526393093,
-0.0467327423,
-0.0082313409,
0.0181347076,
0.099055253,
0.0566928126,
0.0226623584,
0.0336777493,
-0.0789779499,
0.0941736773,
-0.0327294543,
0.0231359992,
-0.0177778732,
0.0233854875,
-0.0435267724,
0.012906678,
-0.0458676852,
-0.0440794677,
-0.0466723777,
-0.0397452042,
0.0075764139,
0.0255744159,
-0.0423707142,
0.003082091,
0.0574571081,
-0.0007672986,
0.0343541577,
-0.0071936203,
-0.0147598023,
0.0098965829,
0.0078491922,
0.0058904407,
-0.0080789067,
0.0036339934,
0.0285518821,
0.0590556003,
0.0426967405,
-0.0419003144,
0.0564954169,
-0.0075404327,
-0.0352713168,
0.030065285,
-0.0306509957,
-0.0284423996,
-0.0599703267,
-0.0070161112,
-0.0012127985,
-0.0288063455,
-0.0274467785,
0.0389102474,
-0.0140963383,
0.0272412058,
-0.0101565504,
0.0217692629,
0.0119309556,
0.0223259423,
-0.0311556794,
0.019639628,
0.0037602789,
0.0458037145,
0.0228502136,
-0.0385148004,
-0.0309160352,
-0.0043616337,
2.65e-7,
0.0553094037,
0.0531633459,
0.0078399247,
0.0054146592,
-0.0211025048,
0.0306779351,
-0.0308967959,
0.0400347598,
0.0281401798,
0.0346379653,
0.0338716991,
0.0097082546,
0.0616609901,
0.0395707227,
-0.0767240897,
-0.0447520092,
-0.0679476708,
-0.0151639218,
-0.0243038274,
-0.0347519964,
0.0740297064,
0.0612114817,
-0.0327890329,
0.0379465669,
-0.0237292573,
-0.0178533718,
0.0690466017,
-0.0140217077,
-0.0361045748,
-0.0294940658,
0.0744379833,
0.0043592411,
-0.0518289097,
0.0019882005,
-0.032863494,
0.0132534206,
0.0974459425,
0.0702983141,
0.0141695077,
-0.0330913402,
0.0156511758,
0.0021343145,
0.0252020843,
-0.0212496724,
0.0452034697,
-0.0484576263,
-0.0144665344,
0.0454512164,
-0.0580647029,
0.0275481008,
0.044127591,
-0.0206760857,
-0.0503207557,
0.0275050141,
0.0264413301,
0.008705575,
-0.0210550744,
-0.0298967585,
-0.051618617,
-0.0105952192,
0.0143804736,
-0.0550730005,
0.0109155234,
-0.0065384344,
-0.0033219596,
-0.0325653963,
0.0003682825,
2.156682929e-34,
0.0572473593,
-0.0277476218,
0.0435384251,
-0.0433974601,
0.0470036082,
0.0001382118,
-0.0508939922,
-0.0142207211,
0.0082659675,
-0.0283111427,
-0.0192016382
] |
http://arxiv.org/abs/2401.17093v1 | http://arxiv.org/pdf/2401.17093v1 | StrokeNUWA: Tokenizing Strokes for Vector Graphic Synthesis | 2024-01-30T15:20:26 | 2024-01-30T15:20:26 | Zecheng Tang, Chenfei Wu, Zekai Zhang, Mingheng Ni, Shengming Yin, Yu Liu, Zhengyuan Yang, Lijuan Wang, Zicheng Liu, Juntao Li, Nan Duan | cs.CV | cs.CV, cs.CL | To leverage LLMs for visual synthesis, traditional methods convert raster
image information into discrete grid tokens through specialized visual modules,
while disrupting the model's ability to capture the true semantic
representation of visual scenes. This paper posits that an alternative
representation of images, vector graphics, can effectively surmount this
limitation by enabling a more natural and semantically coherent segmentation of
the image information. Thus, we introduce StrokeNUWA, a pioneering work
exploring a better visual representation ''stroke tokens'' on vector graphics,
which is inherently visual semantics rich, naturally compatible with LLMs, and
highly compressed. Equipped with stroke tokens, StrokeNUWA can significantly
surpass traditional LLM-based and optimization-based methods across various
metrics in the vector graphic generation task. Besides, StrokeNUWA achieves up
to a 94x speedup in inference over the speed of prior methods with an
exceptional SVG code compression ratio of 6.9%. | StrokeNUWA: Tokenizing Strokes for Vector Graphic Synthesis: To leverage LLMs for visual synthesis, traditional methods convert raster
image information into discrete grid tokens through specialized visual modules,
while disrupting the model's ability to capture the true semantic
representation of visual scenes. This paper posits that an alternative
representation of images, vector graphics, can effectively surmount this
limitation by enabling a more natural and semantically coherent segmentation of
the image information. Thus, we introduce StrokeNUWA, a pioneering work
exploring a better visual representation ''stroke tokens'' on vector graphics,
which is inherently visual semantics rich, naturally compatible with LLMs, and
highly compressed. Equipped with stroke tokens, StrokeNUWA can significantly
surpass traditional LLM-based and optimization-based methods across various
metrics in the vector graphic generation task. Besides, StrokeNUWA achieves up
to a 94x speedup in inference over the speed of prior methods with an
exceptional SVG code compression ratio of 6.9%. | [
-0.0380753204,
0.0278366376,
0.006183221,
0.0375157818,
-0.006516241,
-0.0008044195,
0.0021910712,
0.0445893072,
-0.0732413754,
0.0167849232,
0.0031872361,
0.0066629481,
-0.0189189725,
-0.032110922,
0.0045049442,
-0.1107439995,
-0.0381908789,
-0.02637982,
-0.0354565643,
-0.026218297,
-0.0080889445,
-0.0082155569,
0.0093289036,
0.0214772485,
-0.0916833058,
-0.0187511854,
-0.0197014976,
0.0067130071,
0.0259770565,
-0.0514291599,
-0.073044464,
0.0075549823,
0.0445335694,
-0.005396469,
0.0000022425,
-0.0286819208,
-0.0048323087,
0.0175030269,
0.0351311155,
0.0479939766,
0.0649448112,
-0.0201798007,
-0.0239304416,
-0.0745979548,
0.0219983011,
0.0760007501,
0.0332451649,
0.0750825703,
0.0761756226,
-0.0305637643,
-0.0024291009,
-0.0700576901,
-0.0264150593,
0.0011688975,
0.0039786147,
0.0252462719,
-0.035454236,
-0.0164731611,
-0.0107923998,
0.0018729311,
-0.0270916689,
0.0384596176,
0.0136067579,
0.0152275963,
0.0389100425,
0.0316080526,
0.0427991152,
0.0153381433,
0.0191841293,
0.0152258044,
-0.1005182117,
0.023906583,
0.0636653975,
-0.0215534102,
-0.0340838954,
0.0034401091,
-0.0699874014,
0.0266126227,
-0.0550621673,
0.006212878,
-0.0334729478,
-0.004137319,
-0.0016104934,
-0.0780471489,
-0.022765195,
-0.0164494589,
0.0282960907,
0.0192947797,
-0.0762823671,
-0.0279198196,
-0.0642274469,
-0.0599710457,
0.0385832265,
0.0101484088,
0.0515425429,
-0.0048092082,
-0.0232745782,
-0.0235274788,
0.0433327518,
-0.0157305393,
0.0988419876,
0.0152763873,
-0.0330910757,
0.0309807695,
-0.0401636176,
0.0191981141,
-0.0673377663,
0.0380501598,
-0.0632377937,
0.0162704326,
0.0077685011,
-0.0088917632,
-0.0654246658,
0.0278902445,
-0.0642428994,
0.0410364419,
-0.0102156689,
0.0335399732,
0.0118084308,
-0.021314051,
0.0213249866,
-0.0063475221,
-0.0188463684,
0.0368550792,
-0.0393033326,
-0.0239396673,
-0.0671223328,
0.0155690713,
-0.015816018,
0.0340372138,
-0.0031081275,
-0.0086600641,
0.029642757,
0.0075207823,
0.0161743611,
0.0656546205,
-0.0358198546,
0.0294741746,
-0.0210840311,
-0.0079801539,
0.0206267983,
0.0078460425,
0.0753771961,
-0.0272825249,
0.0390192978,
-0.0163285844,
0.0236068163,
0.0309989881,
0.0083608078,
0.0086163338,
-0.0062551126,
-0.0267739799,
-0.0509532131,
0.0241634939,
0.0162938125,
-0.0085962815,
0.0208344832,
-0.0014007062,
0.0414191373,
0.075673297,
-0.0166336577,
-0.0418402702,
-0.0101930322,
-0.0043730582,
0.0418639816,
0.0139341662,
-0.0476794653,
-0.0337192304,
0.0003218323,
0.060330499,
0.0679798126,
0.0308404285,
-0.0976935774,
-0.0614499673,
0.0088851796,
0.0656453371,
0.0564150028,
0.0889905989,
0.0834310055,
-0.0052957879,
0.0072676078,
-0.018363541,
0.0107911956,
0.0304551516,
0.0672055855,
0.0261709895,
-0.0161114559,
0.0406781845,
-0.0045646271,
-0.0517900363,
0.0402731933,
-0.0071098986,
-0.0013865877,
-0.0525292717,
0.0252793394,
0.0291274041,
-0.0361966491,
0.0474397615,
-0.0096772211,
-0.0577410161,
-0.0345778167,
-0.0289060138,
0.0319035575,
0.0541844778,
0.0273957588,
0.0080682356,
0.0036717094,
-0.0124372523,
0.0426086262,
0.0318429694,
-0.0327124931,
0.0061298017,
0.0533781387,
0.0064112553,
0.0170390923,
-0.0263259634,
0.0002511779,
0.0934759527,
-0.0194264092,
-0.0032858669,
-0.0288557634,
-0.0755093321,
0.000919136,
0.0368212238,
0.0499618426,
0.0225199908,
0.007569971,
0.0004545435,
-0.01860944,
0.0233057719,
-0.0364507362,
-0.023579333,
0.0279077198,
-0.0884540677,
-0.0177614894,
-0.0087299366,
-0.0078146132,
-0.0139662763,
-0.0086708767,
-0.0286874529,
-0.0212961528,
-0.0443027988,
-0.0185071342,
0.0315005258,
-0.0017447037,
0.0179664213,
0.0146274148,
-0.0280150343,
-0.0337211639,
0.0184231363,
-0.0102566183,
-0.0217988063,
0.0530554242,
-0.0652549714,
0.0111114895,
0.0337805822,
0.0224387161,
-0.0038416293,
-0.0046750451,
0.0233299062,
0.0242014863,
-0.007879504,
-0.0253707841,
-0.0514048748,
-0.0118894046,
0.0030055402,
0.0286942031,
-0.0357035398,
-0.0376866758,
0.0421348661,
0.0103941467,
-0.0370506272,
-0.0268209055,
0.0073349653,
0.0282970723,
0.00289166,
0.0043261871,
-0.0158541091,
-0.0281843599,
0.0712992027,
-0.0266112648,
-0.028827928,
0.0316338353,
-0.0015619484,
-0.0005034714,
-0.0174447037,
0.0172933992,
-0.0127936294,
-0.0333008803,
0.0242873523,
-0.0470961183,
0.0006848615,
-0.0206858627,
-0.0077310805,
-0.0127112586,
-0.0325494334,
-0.0522094443,
0.0504791774,
0.0146027012,
-0.0064814775,
0.0431769527,
-0.0355956592,
0.0002477118,
0.0554173402,
-0.0171695333,
0.0847004578,
0.0305075124,
0.0054042181,
0.0686535537,
0.0147362733,
0.0156845711,
0.045262564,
0.0053051109,
-0.010702326,
0.0476750322,
-0.0211283788,
-0.0244134255,
-0.0145270852,
0.0021170275,
-0.0038989941,
-0.0320445672,
-0.0458122082,
0.006045125,
-0.0637561381,
-0.0142281642,
0.060592629,
0.0182100348,
-0.0139913559,
0.0273049399,
0.0429123417,
0.013933979,
-0.0351957902,
0.0282989573,
0.0238239113,
0.0427578799,
0.0158169735,
-0.0241599474,
0.0101971915,
-0.0232937541,
-0.042082265,
-0.0349717103,
-0.0043061893,
-0.0489759967,
0.0133173522,
0.0117130829,
-0.009775077,
0.0111455712,
-0.0162320118,
0.026084207,
0.0405403599,
-0.0289265066,
-0.0152406646,
0.0188889727,
-0.0331614986,
-0.0041775806,
0.0199452285,
-0.0131048365,
-0.0068174144,
-0.0484511666,
0.0077972398,
-0.0014294873,
0.0688973069,
-0.0019385491,
0.0015627466,
-0.0443919674,
0.0129508413,
0.0295394808,
0.0223713238,
-0.0484627225,
0.0365865827,
0.0676698312,
-0.0433770046,
0.0056832968,
0.0292186942,
0.0201643314,
0.0526601635,
-0.0096132262,
-0.0233396329,
0.0479578488,
0.0192330591,
0.021733528,
-0.0879258439,
-0.0069830441,
-0.000043502,
-0.0141848056,
0.0069619194,
-0.0337560326,
-0.0084534725,
-0.0051940442,
0.0957345888,
0.036863368,
-0.0652213246,
0.0659015477,
-0.014261323,
0.0190898944,
0.0393220708,
-0.0654127672,
-0.0278009437,
-0.0057182568,
-0.0106267035,
0.017568348,
-0.0358185582,
0.0344872549,
-0.0374627486,
0.0022404124,
-0.0224458016,
0.0377595834,
0.0384175926,
0.0243659727,
-0.0716705844,
0.0038665256,
-0.0034046643,
0.0505254231,
-0.0689693242,
0.0207597017,
0.0358389616,
-0.0001890053,
-0.0071151373,
-0.0143406587,
0.0461915135,
-0.0407950319,
-0.0105875134,
0.0014584963,
-0.0395369008,
-0.0053194221,
-0.0642299503,
-0.0396520346,
-0.0477893092,
0.0196621921,
0.0294530801,
0.0204312168,
-0.0635781437,
-0.0610092357,
0.0113556031,
-0.0304133724,
-0.0642306656,
-0.0382837318,
-0.0288564544,
-0.0609603785,
-0.0241664089,
0.033707466,
0.0111873988,
0.0373254791,
-0.0324420594,
0.0083263619,
-0.0202593636,
-0.0349704996,
-0.0704291165,
0.0020474999,
-0.0221876465,
0.0122580789,
-0.0087275775,
0.0408285521,
-0.0193571057,
-0.0240991283,
-0.0177371819,
0.0517924204,
-0.0515419841,
-0.012648195,
-0.0075928085,
0.0840998217,
0.0009067905,
0.0045557441,
0.0041163932,
-0.0329119191,
0.0639844239,
0.1056785807,
0.0708412454,
-0.021768,
0.010278766,
0.0784429461,
0.016642822,
-0.0195816401,
-0.0233262442,
0.0179354325,
0.0011904564,
0.0190195385,
-0.0112945195,
0.0394168831,
-0.015115832,
0.0214602295,
0.0196135137,
-0.027023606,
0.010010507,
-0.0129570113,
0.0067190542,
-0.0317054652,
-0.0172479153,
-0.0392891169,
0.0157081448,
-0.0290625878,
-0.0173850562,
0.0053032874,
-0.0057404819,
0.0206797477,
0.0130210752,
-0.0495512448,
-0.0226715021,
-0.033593148,
-0.0270900317,
-0.0073227268,
0.0112034027,
0.0475945994,
-0.0302677173,
-0.0232012328,
0.01690064,
0.0008342337,
0.0146090277,
0.0045680711,
-0.0319678597,
-0.0357150622,
0.0607815534,
-0.0171027221,
-0.0305897333,
-0.0248272959,
0.0599786006,
0.0667421892,
0.0627033859,
0.0449636839,
-0.0327017792,
-0.0042086681,
-0.0285724029,
0.0170596521,
0.068702884,
-0.042052839,
0.0476864502,
-0.0077553331,
-0.0236996971,
-0.1319714785,
0.0000207651,
-0.0502926297,
0.0078673791,
-0.0456610695,
-0.0260044616,
-0.0287219137,
-0.0197830889,
0.0028793276,
0.0287948456,
0.0188822448,
-0.000495493,
-0.0427931473,
-0.0194575004,
0.0711131915,
0.0339459367,
0.0036954565,
-0.0234282445,
0.0495199971,
0.0055212164,
-0.0295349676,
-0.0478892662,
0.0374977887,
-0.0138659934,
-0.0231755488,
-0.0276547074,
0.0049166987,
-5.585354061e-33,
-0.0275349393,
-0.0074489154,
0.0262222216,
-0.0313555859,
-0.0515019037,
-0.0345274098,
0.0005879902,
-0.0155911874,
0.059669517,
-0.0326397531,
-0.0155385695,
-0.0449403003,
-0.0236061271,
-0.0094084255,
-0.0077372105,
0.0096665211,
0.0829092041,
0.0537941419,
-0.0556428283,
-0.023475945,
-0.0424872935,
0.032724373,
0.0167644117,
-0.0383930877,
-0.0685802102,
-0.018396439,
0.017339306,
-0.0057030609,
0.0410378501,
0.0513212904,
-0.0342574939,
0.005359102,
0.0310853235,
-0.006505134,
0.0007666894,
-0.021212643,
-0.0878345519,
-0.0483251959,
0.0318144709,
0.0166117176,
-0.0739083961,
-0.0820304379,
0.0294087995,
0.012242184,
-0.0065874844,
-0.0315566845,
0.0100771245,
-0.0366093442,
-0.0033971602,
-0.0229040068,
-0.0062727188,
-0.0071674287,
-0.0428135209,
-0.0142326783,
0.0471121594,
-0.0295404792,
0.0518061034,
0.0373689495,
-0.0815686956,
0.0339717679,
0.0014287268,
-0.0246489681,
0.0443499796,
0.017657429,
0.0527760163,
0.0371476561,
0.0699467734,
-0.0452630483,
-0.0562530681,
0.0798821524,
0.0129370745,
0.0625542253,
0.0098051773,
0.0225731675,
0.0550860949,
0.0341066644,
-0.0050692917,
0.0650465116,
-0.061468441,
0.0685361624,
0.0107899709,
0.0731378943,
-0.0021181034,
-0.0045518074,
0.0080396943,
-0.0596163906,
-0.0373409614,
0.0218444131,
0.0251340736,
0.0006227178,
0.0042965501,
-0.0243589655,
-0.005138725,
0.0195334777,
0.0025969308,
-0.0218811743,
-0.0068223034,
0.0187363829,
-0.0072548408,
-0.0602237098,
0.0728530958,
-0.0347690806,
0.0422907919,
-0.0440121517,
0.0223771874,
-0.0018115611,
-0.0030354522,
-0.0035379133,
0.0046505551,
0.0238497965,
-0.0246117003,
-0.0129055781,
0.0290909111,
-0.0324745998,
0.0182926748,
-0.0119501324,
-0.0029507244,
0.0240256712,
0.0013231799,
0.0419945084,
-0.0103726359,
-0.0068830079,
-0.0542168319,
0.0158796366,
-0.0403604992,
0.0061795628,
0.0250155758,
0.0339105949,
0.051236093,
-0.089191094,
-0.0686026961,
-0.0126333702,
2.783e-7,
0.0293568093,
0.0193727091,
-0.0140145104,
0.0533127636,
-0.0099166976,
0.0152597306,
-0.0175963659,
-0.0192339402,
-0.0052390723,
0.0119364941,
0.0310705677,
0.0520419106,
0.0258924831,
-0.0192808788,
-0.0507213436,
0.0468817241,
0.0234175697,
0.0108161913,
-0.0259696804,
0.056063652,
0.0585876592,
0.0227228645,
0.0099237068,
-0.0509676151,
0.0615664013,
0.0086049689,
0.0096684331,
-0.0436498448,
0.0418956131,
-0.0074261366,
-0.0043880651,
0.0030606634,
-0.0046045249,
-0.0185876414,
0.0224351864,
-0.0183219779,
0.0067047291,
0.0359559692,
-0.0113813691,
0.0582906455,
0.0495272428,
-0.0043474687,
0.0043892534,
-0.011608067,
0.0197354108,
0.0102754356,
0.0102752233,
-0.0196715761,
0.0419339016,
0.0290788431,
0.012425052,
-0.0035422458,
-0.0287853908,
-0.0024522012,
0.0204235688,
-0.0026475047,
0.0201286748,
0.006406175,
-0.0156107219,
-0.0028955254,
-0.0307332519,
-0.0292099584,
-0.0653439164,
0.0802927092,
0.0004882335,
-0.0165447425,
-0.0403549448,
2.860244078e-34,
0.0035481588,
0.0112730525,
0.0029661383,
0.0668322891,
0.0260400567,
0.0041564316,
0.024816012,
-0.0220510643,
-0.0164416023,
-0.0258548371,
-0.006336581
] |
http://arxiv.org/abs/2401.17092v1 | http://arxiv.org/pdf/2401.17092v1 | NNOSE: Nearest Neighbor Occupational Skill Extraction | 2024-01-30T15:18:29 | 2024-01-30T15:18:29 | Mike Zhang, Rob van der Goot, Min-Yen Kan, Barbara Plank | cs.CL | cs.CL | The labor market is changing rapidly, prompting increased interest in the
automatic extraction of occupational skills from text. With the advent of
English benchmark job description datasets, there is a need for systems that
handle their diversity well. We tackle the complexity in occupational skill
datasets tasks -- combining and leveraging multiple datasets for skill
extraction, to identify rarely observed skills within a dataset, and overcoming
the scarcity of skills across datasets. In particular, we investigate the
retrieval-augmentation of language models, employing an external datastore for
retrieving similar skills in a dataset-unifying manner. Our proposed method,
\textbf{N}earest \textbf{N}eighbor \textbf{O}ccupational \textbf{S}kill
\textbf{E}xtraction (NNOSE) effectively leverages multiple datasets by
retrieving neighboring skills from other datasets in the datastore. This
improves skill extraction \emph{without} additional fine-tuning. Crucially, we
observe a performance gain in predicting infrequent patterns, with substantial
gains of up to 30\% span-F1 in cross-dataset settings. | NNOSE: Nearest Neighbor Occupational Skill Extraction: The labor market is changing rapidly, prompting increased interest in the
automatic extraction of occupational skills from text. With the advent of
English benchmark job description datasets, there is a need for systems that
handle their diversity well. We tackle the complexity in occupational skill
datasets tasks -- combining and leveraging multiple datasets for skill
extraction, to identify rarely observed skills within a dataset, and overcoming
the scarcity of skills across datasets. In particular, we investigate the
retrieval-augmentation of language models, employing an external datastore for
retrieving similar skills in a dataset-unifying manner. Our proposed method,
\textbf{N}earest \textbf{N}eighbor \textbf{O}ccupational \textbf{S}kill
\textbf{E}xtraction (NNOSE) effectively leverages multiple datasets by
retrieving neighboring skills from other datasets in the datastore. This
improves skill extraction \emph{without} additional fine-tuning. Crucially, we
observe a performance gain in predicting infrequent patterns, with substantial
gains of up to 30\% span-F1 in cross-dataset settings. | [
0.0247645359,
0.081003271,
-0.0033466497,
-0.009147061,
-0.0246098842,
-0.0028808594,
0.0257102977,
0.0194387808,
0.027404217,
-0.0116517609,
-0.0306399371,
0.000565809,
0.0007638396,
-0.0077544483,
0.0359395631,
0.0042827395,
-0.0075167418,
0.0241753384,
-0.0389739312,
0.0090295617,
0.008832803,
0.0474873036,
0.0007242498,
0.0535485968,
-0.0054455511,
0.0044503845,
0.0488147736,
-0.036896985,
0.0359332301,
0.0271536428,
0.0696655884,
-0.0024072961,
0.0112766456,
-0.0038391566,
0.0000019031,
0.0011276586,
-0.0368508026,
0.0255246814,
-0.0031872324,
-0.0971414223,
0.0388781764,
-0.0630908087,
0.0338207148,
0.0405110866,
-0.0222228914,
-0.0649299473,
0.0558274277,
-0.0177016854,
0.0244229548,
0.0593672767,
-0.0126750115,
0.0219044462,
-0.0105589461,
-0.020798346,
0.0067288401,
-0.0035089166,
0.0405353196,
-0.0466050543,
-0.0031521071,
0.0047959629,
0.0142427627,
0.0152692096,
-0.00129254,
-0.0090960711,
0.0284864474,
0.041984994,
0.0212096833,
0.0445601121,
0.0037632021,
0.0224554818,
-0.0533761084,
-0.0282152072,
-0.0045038275,
-0.0230681635,
0.0069168797,
0.0764157027,
-0.0449278131,
0.0015401006,
-0.0232056193,
-0.0087107345,
0.0082860515,
0.0157547798,
-0.0063290931,
0.029025279,
0.0127380034,
0.0500495285,
0.0058957143,
-0.0490575284,
0.0277614985,
0.014419863,
0.0427557193,
-0.0246235691,
0.0421514511,
0.0274309833,
-0.0409303233,
0.0182877444,
-0.0015733178,
-0.0097603258,
0.0064011929,
-0.0280572083,
0.008364087,
0.0140951145,
-0.056794472,
0.0465357341,
-0.0015448012,
-0.0301415306,
0.0397972055,
0.0039783195,
-0.0493872091,
-0.0457073264,
-0.1157053262,
-0.0188501421,
-0.0303391367,
-0.0146941552,
-0.019543739,
0.0192153379,
0.0181569587,
-0.0122490134,
-0.0518824086,
-0.0074231634,
0.0450616851,
0.0363411419,
-0.0409463793,
0.0015364203,
-0.0126939919,
0.027529696,
0.0041358257,
0.0244415067,
-0.0155864498,
-0.016636597,
-0.0146125229,
0.0010840467,
0.0571277402,
-0.0575049557,
0.0518884771,
0.0103200404,
0.0156760216,
0.0026819974,
0.0026437826,
-0.0248012822,
-0.0196348745,
-0.0236596651,
-0.0341209173,
-0.0741910189,
0.0580896065,
-0.0275991671,
-0.017796699,
-0.0744853467,
-0.0069475789,
0.0318779834,
-0.0750607625,
0.0378255881,
-0.0530060492,
0.0234205984,
0.0108988052,
-0.0390974805,
0.0079155313,
0.0284713022,
-0.0045117489,
-0.0056892601,
0.0479084961,
-0.0180081148,
0.0515118912,
0.0446145274,
-0.0337467529,
-0.0305150915,
-0.0347937532,
0.0226145945,
0.0120467646,
-0.0048544784,
0.0614064038,
-0.0509683341,
-0.0161553938,
0.0229811314,
0.0196038503,
0.0769681185,
0.0517661422,
-0.0140472734,
0.0490699746,
0.0379948877,
-0.0195577443,
-0.0252725128,
-0.0117774764,
-0.0071622082,
0.0262387041,
-0.0153300921,
-0.0197504107,
0.0540902726,
-0.0408280939,
-0.0169790741,
0.0171539914,
-0.0182532184,
-0.0106647527,
0.0609284751,
0.0433514006,
0.00077945,
0.0190201011,
0.0474685133,
0.0531576835,
-0.038166929,
0.0139483353,
0.0015977601,
-0.0252536498,
0.100735262,
0.0042023822,
-0.046369683,
-0.0262326542,
-0.0024852238,
0.0240640379,
0.0095211603,
0.0071836207,
-0.0478347726,
0.0342524983,
-0.0445092991,
-0.0546383858,
0.0455708653,
0.0280244611,
-0.0100912983,
0.0082463697,
0.0180200506,
0.0318700261,
-0.0283260848,
-0.0653561205,
0.0268735513,
-0.0448146276,
0.0004293255,
0.0369742922,
0.0220470242,
-0.050029397,
0.0177424569,
-0.0178118814,
0.0046426668,
0.0092729051,
-0.035460107,
-0.0252270494,
0.0380792022,
0.0035744526,
-0.014039957,
0.0379779227,
-0.0070825135,
-0.0021121311,
0.0319150686,
-0.0430400446,
-0.0150743509,
-0.0007625957,
-0.0145830624,
0.0167036392,
-0.0271295048,
-0.0030748365,
-0.0246572644,
-0.011770363,
-0.0319101103,
0.0811479166,
0.0398806892,
0.0496925525,
-0.0656063408,
-0.0122546852,
0.0079256948,
-0.0595239624,
0.0165406074,
-0.0495763086,
0.0060037673,
0.0658473074,
0.0408043116,
0.0071271313,
-0.033504799,
-0.0408909619,
-0.0449738838,
0.0536424741,
-0.0090101929,
0.0189330168,
0.0371203274,
-0.0023671479,
-0.0113253752,
-0.0122394068,
-0.0509882122,
-0.0320257582,
0.0185478926,
0.0134643745,
0.0318724923,
-0.0121803991,
-0.0201194994,
-0.0081329718,
-0.0186130982,
-0.0125037152,
0.0218414515,
0.00225724,
0.0117370868,
-0.0484786853,
0.0493051112,
-0.0288700592,
0.0308672301,
0.0077786339,
0.0033886202,
-0.0416854732,
0.0006445852,
-0.0395755656,
0.0052976487,
-0.0359776542,
-0.0347934254,
0.0441101342,
-0.0067769503,
-0.0339035317,
-0.001456701,
-0.0379669927,
0.0082776221,
-0.0105433222,
-0.0305541493,
-0.0292684641,
-0.0011403552,
0.0229386035,
-0.0571686327,
0.0026716869,
0.0068940097,
0.0084327627,
-0.0125953518,
-0.0466090627,
0.0376048647,
0.0064538806,
0.0092389677,
-0.0423204787,
-0.0287266318,
0.0003377469,
0.0015438992,
0.0073955376,
-0.0236446615,
-0.0648652166,
-0.0494280905,
0.0343231931,
-0.0237547532,
0.0039017373,
-0.0094249314,
-0.0494034626,
-0.0177881997,
0.0095490739,
0.0041603609,
-0.046537172,
-0.0127411969,
0.0036059751,
0.0101317754,
0.0387062654,
-0.0404347032,
-0.0373203196,
-0.0413257778,
0.0188917909,
0.0276510939,
0.0089095691,
-0.0009188588,
-0.0426167771,
-0.0106898034,
0.0185956899,
0.0496737845,
-0.0053201546,
-0.009304449,
-0.0139312614,
0.0593583956,
0.0661913976,
0.0523581654,
-0.068608813,
-0.0127443299,
-0.0480306484,
0.0341593176,
-0.005395418,
-0.0036394,
-0.0064241346,
0.0066926405,
0.0317185968,
-0.0513386428,
0.0048399079,
-0.0012446806,
-0.0010483107,
-0.0481828116,
-0.0295101441,
0.0831272006,
0.018256465,
-0.0329206549,
0.0129779605,
0.0442173369,
-0.0312492549,
0.0541586801,
0.0527234077,
-0.0347097516,
0.0129830185,
-0.0067804623,
0.0487940125,
0.0129953604,
-0.0460613109,
-0.0093316883,
-0.1292771101,
0.078551814,
-0.0163070355,
-0.0482157134,
-0.0420357883,
-0.0545638353,
0.0178321134,
-0.0525389016,
-0.028307246,
-0.0020145646,
-0.01305239,
0.0134485997,
0.053762164,
-0.0246981177,
-0.0819175392,
0.0370759107,
-0.0499050766,
0.0210062191,
0.0030955393,
0.0128499931,
-0.0312678479,
-0.0355368555,
-0.0298271906,
-0.0131341275,
-0.026840549,
0.0123358816,
0.0911223516,
-0.0209647715,
-0.0256126467,
0.0055619222,
0.0221356787,
-0.0289574545,
0.009034236,
0.0238655768,
0.0644434094,
0.0128580295,
0.0392497256,
0.0120286336,
-0.0239250194,
0.030224612,
0.0738243312,
-0.039627213,
-0.0474294834,
0.0081906738,
0.0105050858,
-0.0041271988,
-0.0093063693,
-0.0266286526,
0.0235959217,
-0.0444838926,
-0.0360021926,
-0.0154592274,
-0.0658377931,
0.0763288066,
0.0479247943,
-0.0003631516,
0.022570271,
-0.0107654789,
-0.0177208409,
-0.0805380121,
0.0362014323,
-0.0264234617,
0.0443978496,
0.0317451395,
-0.0246595964,
-0.0039519486,
-0.0231497139,
-0.0534977987,
-0.0417425372,
-0.0949629098,
-0.0052692085,
-0.0420789719,
-0.0176542569,
0.0103472965,
0.0610004403,
-0.0036155353,
-0.0484793521,
0.035885334,
0.0830077082,
-0.0741959512,
-0.0048349579,
0.0016811937,
-0.0453303978,
0.010154739,
0.0344923772,
-0.0399161354,
0.029422367,
0.0449525714,
0.0137932133,
-0.008859504,
0.0198510569,
0.0209143795,
0.0037454322,
0.0195867792,
0.0130902417,
-0.0033575275,
-0.1015606076,
-0.0119415773,
0.0065204571,
-0.0395720936,
-0.0497531183,
0.0243092608,
0.0728013292,
0.0019873246,
0.0249388553,
0.029484963,
0.0146608185,
0.0422578603,
-0.1226510182,
0.0283687282,
-0.0445115231,
0.0349319652,
-0.0232906509,
0.0122426832,
-0.0213040691,
-0.0524197593,
0.0056502884,
0.0395852588,
0.0042431606,
-0.0362041518,
-0.001414335,
0.0022908663,
-0.0049095438,
0.013635275,
-0.0682001784,
-0.0413061082,
-0.0465388559,
-0.0060122558,
0.0702452883,
-0.0983300805,
0.0289780889,
0.0216843616,
-0.0200439356,
-0.0118811214,
-0.0085098883,
0.0324190371,
0.0007417054,
0.0019765142,
0.0153094213,
-0.018998092,
-0.0225232244,
0.0866168216,
-0.0107846064,
0.0550614484,
-0.0029042859,
-0.0619212687,
-0.0303874407,
-0.0109469602,
0.0965535343,
0.0286114551,
-0.0231081787,
0.0002201264,
-0.0240513124,
-0.0614015609,
0.030659141,
-0.028622603,
-0.0353111811,
0.0313207954,
0.0160259791,
0.0375785194,
-0.0018348469,
0.025897814,
-0.0353815183,
-0.040044412,
0.0299416464,
-0.0013535932,
-0.0263317451,
-5.989301307e-33,
0.0072420542,
-0.0243645124,
0.0318096504,
0.0039097667,
0.0012138207,
0.0364577211,
0.0314048603,
-0.0451527573,
-0.0042662229,
0.0037541371,
-0.0046433914,
0.0151219275,
0.0139815714,
0.0302136224,
0.0567631461,
0.0299660154,
0.0206460822,
0.0126885669,
0.0314458236,
0.0113347918,
0.0252009463,
0.0397380143,
0.1235552505,
-0.1020489782,
0.0251796581,
0.0080161579,
-0.0160649996,
0.0323006213,
0.066176638,
0.0571976192,
0.0006078215,
0.0649188906,
0.0323206186,
-0.0690307468,
-0.011116365,
-0.0342399776,
-0.0528490134,
-0.0162882563,
-0.0004447579,
0.001978677,
0.0326642133,
0.004965819,
0.0193667617,
-0.0133011155,
0.0042940103,
-0.0427845232,
0.0745259374,
-0.058518786,
0.0072050132,
0.0447247289,
0.0069204033,
0.0283018593,
-0.0110028898,
-0.0123485178,
0.0223521944,
0.0184485745,
0.0171521492,
-0.0408741049,
-0.0920034572,
0.0114159128,
0.0109012863,
0.0563835874,
0.0152341966,
-0.0084962072,
0.0006886182,
0.0271953698,
0.0646324828,
-0.0225523189,
0.0523601845,
-0.006381758,
0.0075246235,
0.0550123379,
-0.0312695913,
0.0245274696,
0.0743858591,
-0.027737366,
-0.0563481934,
0.0494967885,
0.0041794945,
-0.0018199487,
-0.0066736802,
-0.0112329777,
-0.0487129316,
-0.0381686389,
0.0176772866,
-0.0071744937,
-0.0039344085,
0.0025833661,
0.0344473682,
0.0437986627,
0.0179771278,
0.0647194609,
-0.0329276696,
-0.0471449643,
-0.0141763259,
0.049310416,
-0.042776674,
-0.0071553718,
0.0163830984,
-0.0263762046,
-0.0549582243,
-0.0548779145,
-0.019307483,
0.0153584015,
0.0171855558,
0.0419406332,
0.059658885,
0.0478789471,
-0.0959909037,
-0.000142851,
-0.0268335883,
0.010053087,
-0.0363378599,
0.0185506865,
0.033918757,
-0.0092298891,
0.0059395805,
0.0238247924,
0.0505770259,
-0.0393478423,
-0.0099177398,
0.0034967388,
0.0330274031,
-0.0318184011,
0.0211921111,
-0.007937843,
0.0038949952,
-0.053004384,
0.0296418518,
-0.00064699,
0.0059968745,
-0.0256344024,
2.536e-7,
0.0610315278,
-0.0104868105,
-0.0170203112,
0.0421799459,
-0.0071564317,
-0.0257035587,
-0.0468327105,
0.0396398976,
0.0041608857,
0.0434774607,
0.0223450661,
0.0008448149,
0.0380039737,
0.0173901897,
-0.0675523505,
0.0047381446,
-0.0405029394,
-0.0387769938,
-0.0538274385,
-0.0203594435,
0.0478936583,
0.0645503476,
0.0365467593,
-0.0423548669,
-0.0235481318,
-0.0670069903,
0.0400037877,
-0.0124410791,
0.0315583311,
0.0176279377,
-0.0002459135,
0.056191992,
-0.0017215065,
0.0362781882,
-0.0231123976,
0.0666408315,
0.054792583,
0.0564873144,
-0.0584667362,
0.0377024114,
-0.0017699222,
-0.0034729212,
-0.0119831618,
-0.0250489507,
0.0572017059,
-0.0204005484,
-0.0019260145,
-0.0317100137,
-0.0548951142,
0.0402552783,
-0.0292873625,
0.0326448865,
-0.0416888967,
-0.0193065144,
0.045330286,
-0.0050662947,
-0.044979874,
-0.0282208323,
-0.0146467751,
0.0640892982,
-0.0258240253,
-0.0710682571,
0.0502226241,
0.022605069,
-0.0043327701,
-0.044374574,
-0.000103877,
2.542364462e-34,
0.0094989697,
0.0736603439,
0.014758016,
0.021795895,
0.0169201158,
-0.0187721513,
0.0246461518,
0.0067469231,
-0.02139185,
-0.0282247011,
-0.0276402347
] |
http://arxiv.org/abs/2401.17072v2 | http://arxiv.org/pdf/2401.17072v2 | SemScore: Automated Evaluation of Instruction-Tuned LLMs based on Semantic Textual Similarity | 2024-01-30T14:52:50 | 2024-02-05T10:53:21 | Ansar Aynetdinov, Alan Akbik | cs.CL | cs.CL | Instruction-tuned Large Language Models (LLMs) have recently showcased
remarkable advancements in their ability to generate fitting responses to
natural language instructions. However, many current works rely on manual
evaluation to judge the quality of generated responses. Since such manual
evaluation is time-consuming, it does not easily scale to the evaluation of
multiple models and model variants. In this short paper, we propose a
straightforward but remarkably effective evaluation metric called SemScore, in
which we directly compare model outputs to gold target responses using semantic
textual similarity (STS). We conduct a comparative evaluation of the model
outputs of 12 prominent instruction-tuned LLMs using 8 widely-used evaluation
metrics for text generation. We find that our proposed SemScore metric
outperforms all other, in many cases more complex, evaluation metrics in terms
of correlation to human evaluation. These findings indicate the utility of our
proposed metric for the evaluation of instruction-tuned LLMs. | SemScore: Automated Evaluation of Instruction-Tuned LLMs based on Semantic Textual Similarity: Instruction-tuned Large Language Models (LLMs) have recently showcased
remarkable advancements in their ability to generate fitting responses to
natural language instructions. However, many current works rely on manual
evaluation to judge the quality of generated responses. Since such manual
evaluation is time-consuming, it does not easily scale to the evaluation of
multiple models and model variants. In this short paper, we propose a
straightforward but remarkably effective evaluation metric called SemScore, in
which we directly compare model outputs to gold target responses using semantic
textual similarity (STS). We conduct a comparative evaluation of the model
outputs of 12 prominent instruction-tuned LLMs using 8 widely-used evaluation
metrics for text generation. We find that our proposed SemScore metric
outperforms all other, in many cases more complex, evaluation metrics in terms
of correlation to human evaluation. These findings indicate the utility of our
proposed metric for the evaluation of instruction-tuned LLMs. | [
0.0226009637,
0.0332173035,
0.0140579287,
0.0265597999,
-0.0335584581,
0.002144004,
0.0159776602,
0.0039543118,
-0.0149660278,
-0.0686354339,
-0.0639666617,
-0.0317754857,
-0.0066704354,
0.0136217289,
0.0177877061,
-0.0127961654,
0.0230436232,
-0.0217651986,
-0.0494175591,
0.0161135588,
0.0017483755,
0.0385077111,
0.0051943972,
0.0530350357,
0.0167570785,
-0.0063137384,
0.0182915628,
0.0300092436,
0.0340380482,
-0.0373969674,
0.0202154592,
0.0424747057,
0.0238040742,
0.0327303857,
0.0000018868,
-0.0587259941,
0.0029469505,
-0.0014653044,
0.0145310676,
0.0382816866,
0.0751402751,
0.0110697243,
0.0030411207,
0.0197990797,
-0.0233561918,
0.0099107334,
0.0346030705,
-0.0006629695,
0.0628908947,
0.0772226825,
0.0058863848,
-0.0078838756,
0.0143225985,
-0.008735152,
0.0026070501,
-0.0080415709,
-0.0091015575,
0.013959446,
0.01342443,
0.0168673471,
0.0263926648,
0.0216346215,
0.0081144553,
-0.0113202743,
-0.000208169,
0.0263271965,
0.0175736379,
0.004981162,
-0.0203479398,
0.0621827766,
-0.0131316539,
-0.026695203,
-0.0053113322,
-0.0164259505,
-0.0282740034,
-0.0124822529,
-0.0490195006,
0.0524091497,
0.0248158555,
0.0302252546,
0.0197368022,
0.0460866466,
0.0122611318,
0.0299513675,
0.0062283105,
0.0390722342,
0.0437108353,
-0.0343224779,
-0.0078745568,
-0.0020117308,
-0.0396791436,
-0.0530553237,
0.0137974415,
0.039798528,
-0.0632618517,
0.0013256004,
-0.0922811851,
-0.0061862525,
0.0333996192,
-0.0276526883,
0.111783281,
0.0538718477,
0.0150197838,
0.00626087,
-0.0451607034,
0.0929495692,
0.001667181,
0.0048514265,
-0.0649147853,
0.0331826508,
-0.0646171495,
-0.0025259135,
0.001352156,
0.0381535105,
-0.0355315879,
-0.055763863,
-0.0504314974,
0.0199298393,
0.0115219541,
-0.0013895441,
0.0301908106,
0.0142864231,
-0.0323819146,
-0.0073823864,
-0.0353767909,
-0.0151583506,
-0.043503888,
0.0026311162,
0.035520792,
-0.0414235592,
-0.0018872555,
0.0500400774,
-0.0064549237,
-0.0129897753,
0.0261632372,
0.0616110787,
-0.0267798807,
-0.0204769149,
-0.0426530093,
-0.0428832211,
-0.0187588148,
-0.0605889894,
-0.0168096237,
-0.0788537562,
0.025058182,
0.0095792618,
-0.0099398214,
-0.0257396642,
0.0222997367,
0.0334232971,
-0.0069737611,
0.0096461847,
0.0030137913,
0.0340542383,
0.0199144725,
0.0090598846,
0.0198122952,
0.0394816846,
0.0319255851,
0.0370626897,
0.0280905068,
-0.0010910437,
-0.0143107697,
0.0256831739,
-0.009225118,
0.0055211242,
-0.0195124503,
0.0108648511,
-0.0060893991,
0.0572315678,
0.0467519797,
0.0709858388,
-0.0154280588,
0.0055466229,
0.0694807991,
0.0476490483,
0.0026486788,
0.0185599644,
0.0983198956,
0.0352513231,
-0.0220285226,
0.0261666905,
-0.0213928241,
0.0288443305,
-0.0407011099,
0.0433301516,
0.0780361965,
0.0353163593,
-0.0220186599,
-0.0280066319,
-0.0116292536,
-0.0109168999,
-0.0133487228,
-0.0120141935,
-0.0103962198,
0.0203243736,
-0.0772928894,
0.0318970419,
0.0225300621,
-0.0617999882,
0.0176129211,
-0.0599717237,
0.0237422939,
0.0820692778,
0.0644596368,
0.0363927148,
-0.0210713539,
0.0156155555,
0.0168277379,
0.074898921,
0.0406996049,
0.0101738218,
-0.0065390253,
-0.0670036823,
-0.010764882,
-0.0403494313,
0.0019392896,
0.0137887876,
0.0362469926,
0.010827654,
0.0622551106,
-0.0295323599,
-0.0545468628,
0.0555473864,
-0.0271607526,
-0.0265608858,
0.00620045,
-0.0255732592,
0.0195534155,
0.0014820393,
-0.0309352707,
0.0635916367,
0.0088243168,
-0.0379082188,
-0.017784318,
-0.031773258,
-0.0633337796,
-0.0260542538,
0.0063148616,
-0.045794569,
0.023758186,
0.0279250816,
-0.0225189384,
-0.0364484638,
-0.0009294288,
-0.0557463244,
0.0088365385,
-0.040725816,
0.0584506914,
0.0162635017,
-0.0072764917,
-0.0452765301,
0.0424493551,
-0.001357511,
0.0284721423,
0.0361872353,
0.0418719426,
-0.0052976357,
-0.0390202925,
-0.0069001024,
-0.0293802004,
-0.0124102607,
0.0071319644,
0.0339179486,
0.0610196926,
-0.0369483456,
-0.0642288402,
-0.0710882694,
0.0432124697,
-0.0144009702,
0.0307889618,
0.0570663624,
-0.0042814952,
-0.01352234,
0.0344850644,
-0.0444972031,
-0.0203979537,
0.0501035564,
0.0124873435,
0.0420070551,
-0.033346232,
-0.0270418301,
-0.0680140257,
0.0087443637,
0.0094499029,
0.0324604064,
0.0216518883,
0.0585518554,
-0.0640086234,
0.0692343637,
-0.0514540523,
-0.0303591732,
0.0212729294,
-0.0625409111,
0.0252119675,
0.0115514128,
0.0391864739,
0.0387826189,
0.0441106856,
0.0050224662,
0.0397475548,
-0.0055603776,
-0.0621982105,
-0.0236476809,
-0.0332339592,
-0.0382809117,
-0.0101381876,
-0.0542955659,
0.0548855439,
0.0482334495,
0.0369048342,
0.0102892015,
-0.0648575276,
0.0219447613,
-0.0132989511,
-0.0387331545,
-0.0087042749,
-0.0195570048,
-0.0078724027,
-0.0065575163,
-0.0256678425,
-0.0668772608,
0.0492596179,
-0.0694444478,
0.0177310295,
-0.0319712609,
-0.0092678471,
-0.0397310369,
0.0387092195,
-0.0035306748,
0.0392434485,
-0.0228649769,
0.0148241948,
-0.0390297733,
0.0262284726,
0.0104343118,
-0.0251820236,
0.0400381051,
-0.0527778491,
0.0029922316,
0.0152601115,
0.003682337,
-0.0424650349,
-0.0266716871,
-0.0038765226,
0.0038431147,
0.0559093878,
-0.0105661768,
-0.0415003151,
0.0351988673,
0.0074090725,
0.0119037805,
-0.0148451738,
-0.0257847812,
-0.0363018177,
0.0137044331,
0.0604599826,
-0.0076651019,
0.0065958374,
-0.0377419926,
0.0443988442,
0.0293872412,
0.0137060303,
0.0218085833,
-0.0415620133,
0.0209010523,
0.0240880344,
-0.034791287,
0.0141637288,
0.055643335,
0.0679302737,
-0.0461992249,
-0.026199773,
0.1404108703,
0.0420088619,
0.0155551946,
0.021129854,
-0.0065912181,
0.0684309751,
0.0085966503,
0.0068832957,
-0.0375320576,
-0.0339779966,
-0.0135752484,
0.0403509513,
-0.0113527682,
0.0069489484,
-0.0172723569,
-0.1163914278,
0.0212568678,
0.0327997319,
-0.0431483313,
-0.0350487232,
-0.0123988781,
-0.014051497,
0.0120175593,
0.0002535527,
0.0116598075,
-0.0404938236,
0.007344272,
0.008157324,
-0.0210766196,
-0.0537827276,
0.0397228263,
-0.055682756,
0.0447538644,
0.0818247572,
0.0150738191,
0.0289054625,
-0.0321235359,
0.0232983883,
-0.0257117637,
0.0260290001,
-0.0108411349,
-0.0083113108,
-0.0089893136,
0.0199739579,
-0.0404609069,
0.0365157984,
-0.0172611214,
0.0075236899,
0.01894073,
0.0276962966,
-0.0221279562,
-0.0301726814,
-0.0310927257,
0.0309654009,
0.0027310748,
0.0925064608,
0.0596194714,
-0.0411012508,
-0.0395353734,
-0.0009576261,
0.0099221691,
-0.0882416517,
-0.0348773524,
0.0100427587,
-0.0281611979,
0.0095952898,
-0.0980652347,
-0.0321460739,
-0.0042762114,
0.0092057269,
0.0011024221,
-0.0323715732,
-0.0217337478,
-0.0231961627,
-0.0118593955,
0.0264690109,
-0.0275926664,
0.0291640051,
-0.0966900513,
-0.0089606885,
0.0208748616,
-0.0231155138,
-0.0176467802,
0.0015586866,
-0.0855732858,
-0.0604690984,
-0.0497181378,
0.0499265604,
0.0149314478,
0.0119569413,
0.0173973907,
-0.0392781459,
0.0294307321,
0.0368290655,
-0.0288132504,
0.0257000979,
-0.0036515109,
0.0171461701,
-0.0097150188,
0.0297241695,
-0.0246324837,
0.0489887111,
0.0182147231,
0.0931294113,
0.0285955388,
-0.0568830147,
0.0232455339,
0.0318899192,
-0.0099871196,
-0.0274315719,
-0.0069900341,
-0.0153577691,
-0.0168178044,
-0.0125837354,
-0.0209358949,
-0.0617760681,
-0.0036412524,
0.0187047813,
0.002050142,
-0.0051893792,
-0.0056985882,
-0.0344583839,
0.0597335212,
-0.0585589297,
0.0175589267,
-0.040458817,
0.0318388753,
0.0020952548,
-0.01670691,
-0.0597670078,
-0.0002964268,
-0.0332608148,
0.020045165,
-0.0109055052,
0.0673844442,
-0.0070275874,
0.020422833,
0.0070676827,
0.0609492511,
-0.006716168,
-0.0228049532,
-0.0781297162,
0.0044039516,
0.0581011772,
0.016099114,
0.0802960023,
-0.0187716614,
0.0138707925,
-0.0328836255,
0.0403353162,
0.0126891453,
-0.0204435978,
0.0442300215,
0.005795849,
-0.0102170976,
-0.0293125641,
0.0672022179,
-0.0084170597,
0.0199928712,
0.0185310226,
-0.0129100932,
-0.0343381092,
-0.0542570651,
0.0115460893,
0.0043833796,
0.0015521902,
-0.0155827254,
-0.0254862253,
-0.0382010564,
-0.0155459046,
0.0438148007,
-0.0041685957,
-0.0129693905,
0.0606298037,
0.0474744961,
-0.0432918221,
-0.0026430103,
-0.023904942,
-0.027535405,
-0.0502725393,
-0.0232373495,
-0.0592149161,
-6.214514637e-33,
-0.0514606014,
-0.0322221927,
0.0183978882,
-0.0332941413,
-0.037493147,
-0.0237968918,
-0.0467406176,
-0.0042993114,
0.0198805649,
-0.0091655254,
0.0307783596,
-0.0562657863,
0.0046096304,
0.0316903815,
-0.0138509702,
-0.0103492942,
0.0391630083,
-0.0064953626,
-0.0673669428,
0.0280109718,
0.0152365193,
-0.0175112467,
0.0025400235,
-0.0227430649,
0.0077124853,
-0.0286192223,
-0.0266680084,
-0.0142949913,
0.0146535262,
0.0792357326,
-0.01794759,
0.0396304689,
0.0178901665,
-0.0188476481,
-0.0395492278,
0.0247956999,
0.0110586593,
-0.0255558454,
0.0343377739,
-0.0597472228,
-0.0232934449,
-0.0362569876,
0.0277033672,
-0.0187985152,
-0.0674477741,
-0.0064927214,
0.0186916552,
-0.0152278533,
-0.003404214,
-0.0197282303,
-0.0716546029,
0.020235531,
-0.00979055,
-0.0399588309,
-0.0398791879,
0.0790994465,
-0.0190018304,
0.0209553223,
-0.004483392,
0.0387372226,
0.0236029048,
0.0532538369,
0.0846949369,
0.0184898507,
-0.0030562885,
0.0481117181,
-0.0055436124,
-0.0256783832,
-0.0098343696,
0.0161140244,
0.0006406329,
0.0533950962,
0.0483365878,
0.0371452421,
-0.0046226904,
0.0182714872,
-0.037569508,
0.0548105463,
-0.0095893452,
0.0227371212,
0.0589903668,
-0.0000254857,
-0.0407528915,
-0.0046094907,
-0.0606512278,
0.0313474648,
-0.0444255657,
-0.0405754521,
0.0850621611,
0.0073038018,
-0.0355384052,
-0.0319865867,
0.0229643155,
-0.0113101676,
0.0147615513,
-0.0223374795,
-0.0293989126,
0.0081046866,
-0.0229728315,
0.0075770351,
-0.0415117517,
-0.0539010242,
0.0634788871,
-0.0164978467,
0.0129575823,
-0.0051965732,
-0.0361066014,
0.0012545559,
-0.0547738262,
0.021388825,
-0.0376824439,
-0.0080248537,
-0.0338655747,
-0.0395430364,
-0.0062978533,
-0.0328960903,
-0.0276885331,
0.053385824,
-0.0166785754,
-0.036230851,
-0.0190377459,
-0.0348733328,
0.0151556917,
0.0400832556,
-0.0320048034,
0.0634358749,
-0.0393103212,
-0.0294185989,
0.0006609531,
-0.0140083134,
-0.0316866301,
0.0181741416,
2.51e-7,
0.0251081791,
0.036902681,
0.0063938643,
0.0488897264,
-0.0178096723,
0.0173352994,
-0.0027352036,
0.0217389353,
0.0337193646,
0.0161436293,
0.0020338641,
0.0141700758,
0.0725101978,
-0.0000390832,
-0.033940345,
0.0247972719,
-0.1049955115,
0.0144043677,
-0.0711467564,
-0.0296154097,
0.0668979585,
0.0362320505,
-0.0264311396,
0.0037281034,
0.0035948367,
-0.0215073433,
0.0572633073,
-0.0452309139,
-0.0132358521,
0.0296541881,
0.0153733985,
-0.006766778,
-0.0129804257,
0.0161822848,
-0.0101326937,
0.011203344,
0.0354906358,
0.0712187067,
0.0270165596,
-0.0375365801,
0.0173719432,
-0.0230228677,
0.0091592204,
-0.0815598518,
0.060620565,
-0.0446982644,
0.0087659368,
-0.0056871246,
-0.0111793978,
-0.0008598875,
-0.0059244079,
-0.0542292036,
-0.0409596115,
0.0219932236,
0.0499954373,
0.0156211853,
-0.0145832626,
-0.0112808524,
-0.0328024141,
-0.0053075417,
-0.0485384166,
-0.0569548979,
-0.0614480339,
0.0290673431,
0.0035920441,
-0.0542365946,
-0.0226154644,
2.495303362e-34,
0.0249658376,
-0.0069306414,
0.0276339315,
-0.0177342892,
0.0526081845,
0.0425369702,
-0.0003168852,
0.0069141565,
0.0038235195,
-0.1001873761,
-0.0174072441
] |
http://arxiv.org/abs/2401.17043v2 | http://arxiv.org/pdf/2401.17043v2 | CRUD-RAG: A Comprehensive Chinese Benchmark for Retrieval-Augmented Generation of Large Language Models | 2024-01-30T14:25:32 | 2024-02-19T03:14:30 | Yuanjie Lyu, Zhiyu Li, Simin Niu, Feiyu Xiong, Bo Tang, Wenjin Wang, Hao Wu, Huanyong Liu, Tong Xu, Enhong Chen, Yi Luo, Peng Cheng, Haiying Deng, Zhonghao Wang, Zijia Lu | cs.CL | cs.CL | Retrieval-Augmented Generation (RAG) is a technique that enhances the
capabilities of large language models (LLMs) by incorporating external
knowledge sources. This method addresses common LLM limitations, including
outdated information and the tendency to produce inaccurate "hallucinated"
content. However, the evaluation of RAG systems is challenging, as existing
benchmarks are limited in scope and diversity. Most of the current benchmarks
predominantly assess question-answering applications, overlooking the broader
spectrum of situations where RAG could prove advantageous. Moreover, they only
evaluate the performance of the LLM component of the RAG pipeline in the
experiments, and neglect the influence of the retrieval component and the
external knowledge database. To address these issues, this paper constructs a
large-scale and more comprehensive benchmark, and evaluates all the components
of RAG systems in various RAG application scenarios. Specifically, we have
categorized the range of RAG applications into four distinct types-Create,
Read, Update, and Delete (CRUD), each representing a unique use case. "Create"
refers to scenarios requiring the generation of original, varied content.
"Read" involves responding to intricate questions in knowledge-intensive
situations. "Update" focuses on revising and rectifying inaccuracies or
inconsistencies in pre-existing texts. "Delete" pertains to the task of
summarizing extensive texts into more concise forms. For each of these CRUD
categories, we have developed comprehensive datasets to evaluate the
performance of RAG systems. We also analyze the effects of various components
of the RAG system, such as the retriever, the context length, the knowledge
base construction, and the LLM. Finally, we provide useful insights for
optimizing the RAG technology for different scenarios. | CRUD-RAG: A Comprehensive Chinese Benchmark for Retrieval-Augmented Generation of Large Language Models: Retrieval-Augmented Generation (RAG) is a technique that enhances the
capabilities of large language models (LLMs) by incorporating external
knowledge sources. This method addresses common LLM limitations, including
outdated information and the tendency to produce inaccurate "hallucinated"
content. However, the evaluation of RAG systems is challenging, as existing
benchmarks are limited in scope and diversity. Most of the current benchmarks
predominantly assess question-answering applications, overlooking the broader
spectrum of situations where RAG could prove advantageous. Moreover, they only
evaluate the performance of the LLM component of the RAG pipeline in the
experiments, and neglect the influence of the retrieval component and the
external knowledge database. To address these issues, this paper constructs a
large-scale and more comprehensive benchmark, and evaluates all the components
of RAG systems in various RAG application scenarios. Specifically, we have
categorized the range of RAG applications into four distinct types-Create,
Read, Update, and Delete (CRUD), each representing a unique use case. "Create"
refers to scenarios requiring the generation of original, varied content.
"Read" involves responding to intricate questions in knowledge-intensive
situations. "Update" focuses on revising and rectifying inaccuracies or
inconsistencies in pre-existing texts. "Delete" pertains to the task of
summarizing extensive texts into more concise forms. For each of these CRUD
categories, we have developed comprehensive datasets to evaluate the
performance of RAG systems. We also analyze the effects of various components
of the RAG system, such as the retriever, the context length, the knowledge
base construction, and the LLM. Finally, we provide useful insights for
optimizing the RAG technology for different scenarios. | [
0.0793792903,
0.0142362062,
-0.0032128571,
0.0200406015,
-0.0378936566,
-0.0177497044,
-0.0404534452,
-0.0050479085,
-0.0134197641,
-0.0684594288,
-0.0104026804,
-0.0323459357,
-0.0199661851,
-0.0112199634,
-0.0019282197,
-0.0345947444,
0.0607123673,
-0.0033312975,
-0.0094449343,
0.0058686342,
0.0252344478,
0.0190236121,
0.0284664035,
0.0160436388,
-0.0032196951,
-0.0060829357,
-0.0192329101,
0.0268499907,
0.0078137349,
0.0078804027,
0.02693085,
0.0379140712,
0.0230216105,
0.0504416712,
0.0000020049,
-0.0717801675,
-0.0375680476,
0.0187561419,
-0.0336582512,
0.0377990678,
0.0488406345,
0.0257471316,
-0.0088625606,
0.0401248746,
-0.0351982638,
-0.0094100963,
0.011604635,
0.037073236,
0.0721505135,
0.0747135803,
-0.0003807227,
-0.0360490419,
0.0417708755,
-0.0152208945,
0.0347530097,
-0.048148375,
0.0132959848,
0.0614750832,
0.0391490385,
0.0149081834,
0.0081886174,
0.0464641787,
0.0251604617,
-0.0344713293,
0.004126397,
0.0005071488,
-0.0661370307,
0.0127681242,
-0.0092856986,
0.0657535791,
0.0382702872,
-0.0321488045,
-0.0018233254,
0.0410610996,
0.0128278919,
0.0462279692,
-0.0575237796,
0.0157644022,
-0.033718314,
-0.0080861431,
-0.0053941133,
-0.0418108478,
0.0032515812,
0.0346171223,
-0.0293607973,
0.0209536646,
0.0276806913,
-0.0416643545,
-0.0318712033,
-0.0169464666,
0.0006528638,
-0.0621701814,
-0.0055097155,
0.0730093643,
-0.030884916,
0.0374097489,
-0.0255737547,
-0.0404885262,
0.0494737886,
-0.0255943909,
0.0099764755,
0.0289799869,
-0.0044489163,
0.0577957928,
-0.0365195721,
0.0550684296,
-0.0128386095,
-0.016619131,
-0.0716290623,
-0.0232305527,
-0.0797365084,
-0.0517097116,
-0.0591982529,
0.0721541867,
-0.0431445837,
-0.027647078,
-0.0251569543,
0.049993135,
0.012738131,
0.0044707512,
0.0061977128,
0.0110144317,
-0.0306177046,
0.0010722221,
-0.0727406442,
0.037050806,
-0.0111411465,
-0.0001766811,
0.0341003388,
-0.0768150762,
0.001673204,
-0.0049375659,
0.0160698984,
-0.0094314134,
-0.0002654197,
0.0324587487,
-0.0325785466,
-0.0067309779,
-0.028757127,
-0.0547653548,
-0.0153388102,
-0.0571026839,
-0.0104541434,
-0.0818592682,
0.0271321442,
0.0139275976,
-0.0310383495,
-0.0213103555,
0.0197802372,
0.0501508862,
-0.0206869394,
0.0031767723,
-0.0224177632,
0.0368963964,
0.0548161827,
0.0142793301,
0.0162013154,
0.0453744531,
0.0066531869,
0.0127200708,
0.0494133607,
0.004159952,
-0.0011807382,
0.0099411728,
0.0103671532,
0.0072666891,
-0.020093739,
0.0619616583,
0.0051293289,
0.0161841512,
-0.0021293871,
0.085491918,
-0.0481692776,
0.0115186404,
0.0441694781,
0.0055860239,
0.0238964129,
0.0049209273,
0.0501012839,
0.0241490714,
0.0401266031,
0.0239432361,
-0.0080053108,
0.0246296749,
-0.0531360954,
0.0586187355,
0.0611685179,
0.0139509039,
-0.0343443491,
-0.0510061607,
-0.0276596453,
0.0222452134,
-0.0230675302,
0.0220715106,
0.0425723977,
0.0313223377,
-0.0401822962,
0.0431772433,
0.006612279,
-0.0613537282,
0.0279086027,
-0.0521319434,
0.0248720516,
0.085802339,
0.0819821432,
-0.0455768704,
-0.0313683823,
0.0076854029,
-0.0123007819,
0.0371651202,
0.0770857781,
-0.0246725567,
-0.0572410971,
-0.0299236607,
-0.0219473429,
-0.0162665788,
-0.0015365359,
0.0007478377,
0.0659229383,
-0.0130487578,
0.0413477272,
-0.0426941998,
-0.0497174747,
0.0195920691,
-0.0540852062,
-0.0119713452,
-0.0234026872,
-0.0276039634,
-0.0035812391,
0.0386168584,
-0.0396452621,
0.0436331965,
0.0317454934,
-0.0134327924,
-0.0500958562,
-0.0161820278,
-0.0182587989,
-0.0265493449,
0.0493889824,
0.0033061656,
-0.0034103924,
0.0182587188,
-0.0229981225,
-0.0536191985,
-0.012927765,
-0.0411241315,
0.0225504171,
-0.0996203423,
0.0136662591,
0.0115653882,
-0.0114243245,
-0.0495472029,
0.0735284984,
0.012388261,
0.0272980407,
0.0490858704,
0.0017475043,
0.0133257871,
-0.0531350374,
-0.0239822287,
-0.008902289,
-0.0119296229,
0.0320481807,
0.047161404,
-0.0208883155,
-0.0487971492,
-0.0399520583,
-0.0784238651,
0.0259498209,
-0.0419803038,
-0.0001971961,
0.0532630756,
0.0301577281,
-0.0039084558,
-0.0103258165,
-0.038023252,
-0.0092064003,
0.0366433673,
0.0469546169,
0.0102978423,
0.0405183285,
0.0225872193,
-0.0323062576,
0.0027805157,
0.0237150285,
0.0362310037,
0.0098136971,
0.013582781,
-0.047131747,
0.0445542708,
-0.0743576288,
0.0076016183,
-0.0018118384,
-0.0230082944,
-0.010531472,
-0.0371384248,
0.0570158474,
0.000120453,
0.0311456639,
-0.0010376107,
0.0327370688,
-0.0283903647,
-0.050731983,
0.0097303903,
-0.0273385104,
-0.041879449,
-0.0200796742,
-0.0406357162,
0.012344298,
0.015785329,
0.0110270754,
-0.0195634998,
-0.0274233241,
0.0039001433,
-0.0023150472,
-0.031583637,
-0.0459566601,
-0.021472225,
-0.0183948707,
0.0272934027,
-0.0470314138,
-0.0224381555,
0.0507700928,
-0.0244165007,
0.0196606964,
-0.0554481,
0.0166964326,
-0.0506670438,
0.0555146635,
-0.0167006738,
0.0338042788,
-0.048835706,
0.001357189,
-0.0556570143,
0.0070237173,
0.0214917138,
-0.0717668384,
0.0330416113,
-0.0420579165,
0.0312759653,
-0.0135477968,
-0.0115074273,
-0.0470694788,
-0.0251719654,
0.0189185403,
0.0215991251,
0.0003184853,
0.020397719,
0.0002072708,
0.0173352938,
0.0545650721,
0.0423766114,
-0.0233893488,
-0.0061150915,
-0.0005271204,
0.0390076861,
0.0791346878,
0.0145526957,
0.0164809842,
-0.0155576086,
0.0533629693,
0.0722951815,
0.0419124477,
-0.0187105965,
0.014531848,
-0.0273893867,
0.0009741953,
-0.0506819375,
-0.0166256092,
0.0088742403,
0.0354814045,
-0.0659889206,
-0.0241633579,
0.0935143456,
0.0502272695,
0.0055626826,
-0.0340297557,
-0.0229696427,
0.0183909722,
0.0324875675,
0.0541014113,
-0.0764520019,
0.0033224928,
0.0090090651,
0.0191347729,
-0.0066890959,
-0.0137561355,
-0.0006759014,
-0.1552500129,
0.0375459343,
-0.0209553242,
-0.0427905582,
-0.0057939272,
-0.0328948945,
0.0139488475,
0.0566438958,
0.0191744678,
0.0688273311,
-0.0296604298,
0.0055329148,
-0.0041554091,
-0.0103445826,
-0.0279037729,
-0.0047701178,
-0.0080589969,
-0.0048497631,
0.0770200491,
0.0223631319,
0.0043225703,
0.0187749267,
0.0242574587,
-0.0220848378,
0.0214700513,
0.000466903,
0.0506823994,
0.036981836,
0.0154007617,
-0.0345449932,
-0.0090053855,
0.0158015806,
0.002303384,
0.0166047104,
0.053385634,
-0.0189050883,
0.0176898669,
-0.0338139385,
-0.011067736,
-0.0075755776,
0.1000039205,
0.0126217501,
-0.0109400926,
-0.0391204171,
0.0035839565,
-0.0178562477,
-0.0495284647,
-0.0171067901,
0.0451758206,
-0.0400612131,
-0.0009741185,
-0.106420964,
-0.0638244301,
-0.0001008007,
0.0155412629,
0.0217012521,
0.0435796343,
0.0058822981,
-0.000067625,
-0.0588917434,
-0.0155040985,
-0.0403325036,
-0.0040037269,
-0.0236468762,
-0.0195117332,
0.0348959826,
-0.063394323,
-0.0017664861,
-0.0130715128,
-0.0421258211,
-0.0546713993,
0.0082325283,
0.013912227,
0.0222855359,
0.0368681513,
0.0136181563,
-0.029642744,
0.0404726639,
-0.0002179019,
-0.0429398343,
0.0301240571,
-0.0297225583,
0.0037856901,
-0.0302132517,
0.0199721716,
-0.0980726182,
0.0029207089,
0.0039323876,
0.0659922957,
-0.0086351372,
-0.0360239074,
0.0389332213,
0.0450314321,
-0.0202080589,
0.030610729,
-0.0015485636,
-0.0213897265,
0.0335533321,
-0.0599514171,
-0.0388496555,
-0.0142164798,
0.0305297654,
0.0251676738,
-0.0154400319,
0.0303612165,
0.0588964261,
-0.0008785983,
0.0502847359,
-0.0309156217,
-0.0483973436,
0.0171773415,
0.019162925,
0.025880605,
-0.0041631237,
-0.0380684361,
-0.013197626,
0.001379429,
0.0448607579,
-0.0147556122,
0.0600148477,
0.016627863,
-0.0194263719,
0.0027910776,
0.0032557433,
-0.0984978229,
-0.0495699123,
-0.0203919671,
0.0162631217,
0.0300943069,
0.0185913108,
0.0608789995,
-0.0208448283,
-0.0366209634,
-0.0596553348,
0.0022264956,
0.0463646241,
-0.0359044187,
0.0055849347,
0.0391826667,
-0.010856648,
0.0222228859,
0.0557710007,
-0.0161473099,
0.0517573655,
-0.0061868075,
-0.0236237291,
-0.0349785574,
-0.048224058,
-0.0032266953,
0.0025493731,
-0.0154687623,
-0.0116157904,
-0.0651713461,
-0.0419518687,
0.0121397935,
0.0200369172,
-0.012910326,
0.0032310444,
0.0428358652,
0.0829419419,
0.0170872137,
-0.0004896507,
-0.0690012649,
0.0308292862,
-0.0962006971,
-0.0111020682,
0.0070566679,
-6.469198713e-33,
-0.0220416989,
-0.0151930638,
0.041076947,
-0.0652258322,
-0.029005155,
0.0075659351,
-0.0460157953,
-0.0220561512,
-0.0383899584,
-0.0092808791,
-0.0289253537,
0.0067381863,
0.0026119763,
0.0303796548,
-0.0316357799,
0.0091768177,
0.0471978709,
0.0068643382,
-0.0176056288,
0.0179437604,
0.0798758715,
0.0229195524,
0.0263398159,
-0.0142749539,
0.0105577419,
-0.0196872242,
0.0149224112,
0.0440455377,
0.0004758494,
0.0788560584,
-0.010025776,
0.0518481731,
0.027865408,
-0.0205192305,
-0.0247778278,
-0.0033578835,
-0.0917414278,
-0.009516146,
0.0151367728,
-0.0131407175,
-0.0343067721,
-0.020207677,
0.0203277469,
-0.0295500923,
0.0201930143,
-0.0050570108,
0.0424234457,
-0.0443617478,
-0.0098970663,
-0.0346081741,
-0.0459803753,
0.017744774,
0.0217595473,
-0.0253509246,
-0.036707297,
0.0446562544,
0.0114930505,
0.0203195773,
-0.0610879026,
0.004534076,
-0.0065684081,
0.0709125325,
0.0399767049,
0.0200640298,
0.0147584379,
0.0439510122,
0.0088812429,
0.0089688133,
-0.0047649909,
-0.0166193731,
0.0018320881,
0.0286070071,
0.0221573468,
0.0771206468,
0.0368944257,
0.0192041285,
-0.0588493273,
0.0395436473,
-0.0141793322,
0.0597324558,
0.0293174181,
-0.0021113907,
-0.0289080292,
0.0036484802,
-0.0542031638,
0.0473929495,
-0.0436179452,
-0.0344950743,
0.0581071563,
0.0452776998,
-0.040290352,
-0.0160477776,
0.0427860171,
-0.0319750644,
-0.0081875073,
0.0465370566,
-0.0394008011,
0.0199548211,
-0.0292178839,
0.026142709,
-0.0102233049,
0.0108533474,
0.0458609797,
-0.0044671637,
-0.0374115966,
0.0255845021,
0.0156664643,
0.0098072421,
-0.0482296161,
0.0326444842,
-0.0402075499,
0.0411034077,
-0.025457317,
-0.024155708,
0.012062924,
0.0150922956,
0.005155697,
0.03285731,
0.0333764143,
0.0043711937,
-0.0468682088,
-0.0119560845,
0.0244782101,
-0.0131958015,
-0.0135750268,
0.0625182539,
-0.0244309604,
0.0458848625,
0.0045328578,
-0.021097444,
-0.0015239111,
0.0414242744,
2.751e-7,
0.0548509508,
0.0095483484,
0.0625144243,
0.0284757912,
-0.002709049,
-0.0167732891,
-0.0062516429,
-0.0036049124,
0.041210752,
0.0046098744,
-0.0062620058,
0.032333836,
0.0501850098,
0.0269806292,
-0.100094609,
-0.0307997912,
-0.0510284752,
-0.0305777267,
-0.0322134905,
-0.0011116653,
0.0535526685,
0.047617577,
-0.0341235362,
-0.0126096923,
-0.0358780026,
-0.0550624207,
0.0599355586,
0.0062037418,
-0.0401292518,
0.0632672384,
0.042201072,
0.0360758416,
0.0101888143,
-0.0216651764,
-0.012245222,
0.009530304,
0.0092219068,
0.0359820314,
-0.0008045743,
-0.0454977825,
-0.0193382762,
-0.0244943444,
0.0190023053,
-0.0452561975,
0.0733456984,
-0.0192403086,
0.0201518033,
0.0007418643,
-0.0311755501,
-0.0085881045,
-0.0254146233,
-0.0253965724,
-0.0194184035,
0.0272629466,
-0.0073768143,
0.0216461215,
-0.004231167,
-0.0668159723,
-0.0390432179,
0.0119303446,
-0.0360576734,
-0.0706352443,
-0.020061139,
-0.0150453039,
-0.0272570029,
-0.0685044006,
0.0116744963,
2.732996126e-34,
0.0331364498,
0.0019085658,
0.0066591767,
0.022033371,
0.038067773,
-0.0079322737,
-0.0457572043,
0.0039525325,
-0.0161331594,
-0.1042498052,
-0.0277963541
] |
http://arxiv.org/abs/2401.17039v1 | http://arxiv.org/pdf/2401.17039v1 | Taking Action Towards Graceful Interaction: The Effects of Performing Actions on Modelling Policies for Instruction Clarification Requests | 2024-01-30T14:18:31 | 2024-01-30T14:18:31 | Brielen Madureira, David Schlangen | cs.CL | cs.CL | Clarification requests are a mechanism to help solve communication problems,
e.g. due to ambiguity or underspecification, in instruction-following
interactions. Despite their importance, even skilful models struggle with
producing or interpreting such repair acts. In this work, we test three
hypotheses concerning the effects of action taking as an auxiliary task in
modelling iCR policies. Contrary to initial expectations, we conclude that its
contribution to learning an iCR policy is limited, but some information can
still be extracted from prediction uncertainty. We present further evidence
that even well-motivated, Transformer-based models fail to learn good policies
for when to ask Instruction CRs (iCRs), while the task of determining what to
ask about can be more successfully modelled. Considering the implications of
these findings, we further discuss the shortcomings of the data-driven paradigm
for learning meta-communication acts. | Taking Action Towards Graceful Interaction: The Effects of Performing Actions on Modelling Policies for Instruction Clarification Requests: Clarification requests are a mechanism to help solve communication problems,
e.g. due to ambiguity or underspecification, in instruction-following
interactions. Despite their importance, even skilful models struggle with
producing or interpreting such repair acts. In this work, we test three
hypotheses concerning the effects of action taking as an auxiliary task in
modelling iCR policies. Contrary to initial expectations, we conclude that its
contribution to learning an iCR policy is limited, but some information can
still be extracted from prediction uncertainty. We present further evidence
that even well-motivated, Transformer-based models fail to learn good policies
for when to ask Instruction CRs (iCRs), while the task of determining what to
ask about can be more successfully modelled. Considering the implications of
these findings, we further discuss the shortcomings of the data-driven paradigm
for learning meta-communication acts. | [
0.0695846602,
0.0089846309,
-0.0410426296,
0.0046514105,
-0.0276721232,
0.0267126355,
-0.0019316599,
-0.0179125704,
-0.0413661152,
-0.0228853878,
0.0116231674,
-0.0371591635,
0.0031276345,
-0.0062254439,
-0.0087268315,
-0.0864087269,
0.0478649922,
-0.0322248153,
0.0299326554,
0.0096888812,
0.0043062204,
-0.0373976603,
-0.0095806895,
0.0587599762,
0.0026259022,
-0.0215027276,
-0.079012841,
-0.0221830942,
0.010383077,
0.0004338941,
0.0245858729,
0.046338141,
0.0459471196,
0.0101531474,
0.0000019443,
-0.0444339775,
0.0126724467,
-0.0366724692,
-0.0291536693,
-0.0308405422,
-0.0107396469,
-0.0407408401,
0.0201124698,
0.0193067603,
-0.0534820966,
0.019346768,
0.0291236453,
0.0292298142,
0.0601143502,
0.0606614351,
0.008766219,
0.0176610593,
-0.0255679525,
-0.0085909329,
0.0564424209,
-0.036358986,
0.0168231372,
0.0046204198,
0.020329237,
0.0166489407,
-0.0577945374,
0.0175030641,
0.0309892297,
-0.0366859622,
-0.048962716,
0.0144213513,
0.038643118,
-0.0629918426,
-0.038314525,
0.0316027813,
0.0481707267,
0.0149795068,
0.0125152366,
-0.0322783403,
0.0581690148,
0.0601297878,
-0.0319154337,
-0.0104700532,
-0.0166988783,
-0.0247549955,
-0.0236772299,
-0.0172374435,
0.0070480532,
-0.010932304,
0.0286185592,
0.0406311303,
-0.0643527657,
-0.0501321927,
-0.0079554804,
-0.002270262,
0.0111133866,
-0.0495876297,
-0.0190298688,
0.0827201381,
-0.0695160553,
0.024009591,
-0.0001075969,
0.0573466271,
0.0262903832,
0.022874685,
0.0361288115,
0.0651427731,
0.0140253371,
0.0089028273,
0.0451225601,
0.0502228253,
-0.0087503344,
0.0516478717,
0.0030542025,
0.0574834198,
-0.0849622414,
-0.027096834,
-0.0383346081,
0.0340732224,
-0.0313873589,
-0.0037981006,
0.0042388337,
0.0375218317,
-0.0285998974,
0.0311136115,
-0.0004836373,
0.003997731,
-0.049122151,
0.0194529649,
-0.0792089328,
0.0170006603,
-0.0212540887,
0.0269720294,
0.0321435779,
-0.0087208692,
-0.0163964331,
-0.0312516242,
-0.0007071112,
0.0203175917,
-0.020375099,
0.074633792,
-0.0156513583,
-0.0193501264,
0.011493138,
-0.0271623433,
-0.0017894161,
-0.0454195067,
0.0392499939,
0.0012988644,
0.0409486741,
0.0118880924,
-0.0234230105,
-0.0202755779,
0.0063449703,
-0.024370946,
-0.0099883564,
0.0791712999,
0.0456662551,
0.0296641532,
-0.0203689821,
0.0373019949,
-0.0495649874,
0.057226222,
0.0183664802,
-0.0255043041,
0.011117748,
-0.0286889672,
0.078891933,
0.0047575361,
0.0061266725,
0.0099176904,
0.0654422492,
-0.0181955639,
-0.0295228772,
0.0217198562,
0.0346991383,
0.0109550133,
-0.0302471314,
0.001364934,
0.0253533907,
0.0344918706,
-0.0230704192,
-0.006689298,
0.035352096,
0.0259729158,
-0.0160817634,
0.0525198281,
-0.0632139295,
0.0462931804,
0.027898673,
0.0282089487,
-0.0674509108,
-0.0169815756,
-0.0403313413,
-0.0663739592,
0.026491357,
0.0148897022,
0.0022603914,
0.0286797993,
0.0029730373,
0.023463957,
-0.0373956859,
0.0117163016,
0.0141611807,
-0.0219313111,
0.0028259242,
-0.0956528261,
0.0365080498,
0.0746751204,
0.0165005047,
-0.023105463,
-0.0619715713,
-0.0328298099,
0.0257470328,
0.0289140828,
0.0309899207,
-0.0459577069,
0.0388418138,
-0.0230266638,
-0.0435707793,
-0.0276348628,
0.0112096891,
0.011731185,
-0.0162420291,
-0.0719509572,
-0.0215973146,
0.0643379316,
-0.0367884375,
0.0343483798,
-0.0119835697,
-0.011913049,
-0.0543349199,
0.0111254994,
0.0083780317,
-0.0188247543,
0.0328406468,
-0.0148290815,
0.0284414254,
0.0350588374,
0.0153871682,
0.0417752229,
-0.0528746247,
-0.0201854073,
0.0054550306,
0.0170772634,
-0.0512390323,
-0.0254406221,
-0.0260053016,
-0.0091000851,
-0.0036876935,
-0.014136428,
0.0665453747,
0.0220528785,
0.0482932217,
-0.0228096638,
-0.0815145969,
-0.0054686773,
-0.0263834801,
-0.0539643131,
-0.0071033929,
0.0438817479,
-0.019351067,
-0.0225765221,
-0.0254866034,
0.0060376837,
-0.0056831264,
-0.0446312614,
0.013307984,
0.0048212316,
0.0259798802,
-0.0505600348,
-0.0386605635,
0.0060148244,
-0.0561202392,
-0.0007740085,
-0.0160669126,
0.0118568176,
0.0060077212,
0.0015625224,
0.0002364114,
0.0258406494,
-0.0036698848,
0.0235818457,
0.0199948717,
0.0292590074,
0.0523841009,
-0.0472889394,
-0.0508472361,
-0.0007869883,
-0.0066768983,
-0.015598801,
0.0191244613,
-0.0098301275,
-0.0766691566,
0.0305560771,
0.010256662,
0.009754817,
-0.0470066108,
-0.0622780398,
0.0658401251,
0.0121327899,
0.0197919533,
0.0630345866,
-0.0080047827,
0.0780233145,
0.036849197,
-0.0251992829,
-0.0876988843,
0.0357859321,
-0.070413053,
-0.0376462005,
-0.0756071061,
-0.0537490286,
-0.0047757053,
0.06642019,
-0.0210173056,
0.0473105945,
-0.0660048574,
-0.0227697119,
-0.005825764,
-0.0651539415,
-0.0239425711,
-0.023146851,
-0.0421827435,
0.0276782159,
-0.0213089138,
-0.0324895531,
0.0746942461,
0.0438584611,
-0.0285008084,
0.003999006,
0.0255124252,
0.0186036732,
0.0538591892,
-0.0103286738,
-0.0144205568,
0.0300714429,
-0.0535307154,
-0.0628661513,
-0.048429817,
0.0596147031,
0.0052353651,
0.0275540631,
-0.0681495592,
-0.0246879831,
0.0290724412,
-0.0297828764,
-0.0364806689,
0.0109058898,
0.0043019517,
-0.0109584965,
0.0207518544,
-0.0400416665,
-0.0175909866,
-0.0080255531,
0.0279938299,
-0.0003006214,
-0.0014718398,
-0.0474363454,
-0.0349332839,
-0.0570119545,
0.0243039429,
0.040145766,
-0.0114946617,
-0.0464351028,
0.0438700914,
0.0238909367,
0.0311376229,
-0.0300042313,
-0.0049788607,
0.0158210229,
0.0098378779,
0.0070329169,
0.0068099853,
0.059680067,
0.0666075945,
-0.0193014629,
-0.0744307712,
0.0767786652,
0.0269895475,
-0.0204510707,
-0.0015708513,
-0.0469548814,
0.0004199525,
-0.0069536366,
0.0155771933,
-0.0418372825,
-0.004967968,
0.0110387085,
0.0342330635,
-0.0353544727,
-0.0020107587,
0.0130856484,
-0.0581963658,
0.0525977463,
-0.0248355102,
-0.0153679904,
-0.0331826992,
-0.0121998647,
0.0144308135,
-0.0101815425,
0.0521458536,
0.0022587336,
-0.0136937713,
-0.0166796278,
0.0569385551,
-0.0001399733,
-0.0443509072,
-0.0124428375,
-0.0091529582,
0.0262997281,
0.0551383235,
0.0069428268,
0.0089570219,
-0.0366207846,
0.030592693,
0.0559359938,
0.035959024,
-0.0168199837,
0.0404532664,
-0.0090049589,
0.0364983119,
-0.0228928179,
0.0188267268,
0.0155259063,
-0.0380226113,
0.0293581616,
0.0475255512,
-0.0445404574,
-0.020849051,
-0.0104509341,
0.0292354915,
0.0528861843,
0.0430613905,
0.0216058139,
-0.0442673415,
-0.0020476643,
-0.0068145371,
-0.0348660015,
-0.032087069,
-0.0284163039,
-0.003882814,
-0.0091756731,
0.0027356979,
-0.0157952812,
-0.089149572,
0.0347778983,
0.0420659855,
0.0295170527,
-0.0524049178,
0.0338570997,
0.0237076692,
0.010176816,
-0.0080205882,
0.0171610042,
0.0009071181,
-0.0574400909,
-0.0362227485,
-0.0301131289,
-0.0025053683,
-0.0104166577,
-0.0668528676,
-0.0161066949,
-0.0581347495,
0.0196283497,
0.102350302,
0.0446330719,
0.0535490625,
0.0344625078,
-0.017167069,
-0.0255835392,
0.0362242237,
-0.0064974176,
-0.0002203857,
0.0669967756,
-0.0090780854,
-0.0356924534,
0.004913718,
-0.011324971,
-0.003661908,
0.0139604053,
0.0293981936,
-0.0564083122,
-0.028591264,
0.0500809513,
-0.0461100861,
0.038724836,
0.0009940331,
-0.0362661518,
-0.0338509195,
-0.0132637629,
0.0015008261,
0.0060821678,
-0.0348315835,
-0.0052721961,
-0.0007745391,
0.0055595823,
0.0177685041,
-0.0247609559,
-0.046152465,
-0.0010607496,
-0.0084596388,
0.0217296351,
0.0120242927,
0.0317930169,
0.025686536,
0.0252932794,
0.0091512492,
0.0138401799,
0.0517427847,
0.0113208257,
0.0724483654,
0.1117701456,
0.0559789687,
0.0240614843,
-0.0150684332,
-0.0013552483,
-0.0632149205,
-0.073283717,
0.0023371815,
-0.0014483689,
-0.0179569498,
0.0095220646,
-0.0112092383,
-0.0343965814,
0.027521437,
0.0179458931,
0.056524273,
0.0324868597,
-0.0291464012,
0.0750612989,
0.0335586295,
0.0198377948,
-0.0262429155,
0.0363195613,
-0.056526389,
0.0689965934,
-0.0368307792,
-0.0007535451,
-0.0424772054,
-0.0508257523,
0.03299715,
-0.0141958706,
-0.0209587403,
-0.0009885185,
-0.068988286,
-0.0244396515,
-0.0594251789,
0.0314408317,
0.0072790189,
-0.0660808906,
0.0137119005,
0.020587787,
0.0086919749,
-0.0564239547,
-0.0256850105,
0.0031627046,
-0.0151659418,
0.0069892551,
-0.0647520497,
-5.409866714e-33,
-0.0353033766,
-0.031217007,
0.037196137,
0.0662845746,
-0.0021970675,
-0.0235651601,
-0.0527899787,
0.0146079361,
-0.0122563252,
0.0201778114,
-0.0000330277,
0.0519499481,
0.0186212193,
-0.0304835867,
0.0676455125,
-0.020258747,
0.0450367779,
0.0016789142,
-0.0496155769,
0.011001884,
0.0067303567,
-0.0146523453,
0.0067549027,
-0.0536616407,
0.0442901365,
-0.0064682942,
-0.0081458427,
0.0237043258,
-0.0338006169,
0.0462663211,
0.0169668589,
0.0002687787,
0.0129033104,
-0.0138288531,
0.0079334201,
0.0587786548,
-0.0645598844,
-0.0456358008,
0.0353155881,
-0.0075567085,
-0.0522899628,
-0.023945516,
-0.0071935207,
0.0183792431,
-0.0330228992,
-0.0219381787,
0.0190502796,
0.023318192,
0.0049004448,
0.0132900476,
-0.0872970372,
-0.025704205,
-0.0364929996,
-0.0761815682,
0.0149319191,
0.1042838171,
-0.0227503702,
-0.0209970288,
-0.027778713,
0.0451030359,
0.0198766049,
0.0108491024,
0.0353257582,
0.0547687784,
0.0085696587,
0.0300990492,
-0.0006897322,
0.0494516455,
0.0356305949,
-0.0208880343,
0.0154601131,
0.0432904363,
-0.0138772763,
0.0305238105,
0.0064403894,
-0.0112744197,
-0.0067878747,
0.017652506,
0.0267395992,
0.0789567456,
-0.012716624,
-0.0108207148,
-0.0370871611,
-0.047961399,
-0.0160673037,
-0.0045807431,
0.0051832306,
-0.039961651,
0.0375027396,
0.0356156342,
0.0137913339,
0.0108350432,
0.0166174825,
-0.0072330832,
-0.0091206729,
-0.0215625316,
-0.0093434611,
0.0438584127,
-0.0184402484,
-0.0091124047,
-0.0791922063,
-0.0642422214,
-0.0098880408,
-0.0002300636,
0.037580125,
0.0480465665,
0.0269747097,
0.0449531823,
-0.0109231556,
-0.0427398086,
-0.0365086347,
0.0463432111,
-0.001668858,
-0.0729050115,
0.0126346275,
0.0330154486,
-0.045439899,
0.0102125425,
-0.0179964844,
-0.0534616821,
-0.0568721518,
0.0473813787,
-0.0186487939,
0.0192011371,
-0.0355366245,
-0.0370628089,
-0.0382505767,
0.0153114079,
-0.001005022,
-0.0503417924,
-0.016844077,
-0.0116534429,
2.733e-7,
-0.0125542972,
0.0070032836,
0.026361037,
0.0431701541,
0.0097561469,
0.0398844406,
-0.0303923339,
0.0030455492,
0.0976778343,
-0.0294616967,
0.0518982038,
-0.0039761923,
0.0068857837,
-0.0161954761,
-0.0254878402,
-0.0287907571,
-0.0167470407,
-0.0419142842,
0.0089386376,
0.0114739286,
0.0515867062,
-0.0139940409,
-0.0072691133,
0.0134690022,
-0.0210990869,
0.0343184024,
0.0421320572,
-0.0653779432,
0.033264339,
0.0233421382,
0.0043956297,
-0.0305591896,
0.0422932021,
0.0639231503,
0.0161377974,
-0.0215584114,
-0.0657331496,
0.0635151565,
-0.0075424952,
-0.0093035633,
-0.0084641296,
0.009881366,
0.0284611806,
-0.0162685495,
0.003625578,
-0.1347622722,
-0.003010144,
0.0632398352,
-0.0057065864,
0.0110947983,
-0.0381918214,
0.0145941293,
-0.0019576533,
0.0109631494,
-0.0173464194,
0.0270750802,
0.0234998297,
-0.0373780355,
0.0249910653,
0.0114289969,
-0.0131432507,
-0.0705186203,
0.0156941004,
-0.0304485206,
-0.0037980557,
-0.0476827696,
-0.0205908697,
1.875219531e-34,
0.0044659683,
-0.0016107648,
0.0607746579,
0.0457249135,
0.0247081928,
0.013399668,
0.0067938436,
0.062799044,
0.0711150691,
-0.030117698,
-0.0320604928
] |
http://arxiv.org/abs/2402.07913v2 | http://arxiv.org/pdf/2402.07913v2 | QACP: An Annotated Question Answering Dataset for Assisting Chinese Python Programming Learners | 2024-01-30T13:11:23 | 2024-02-23T02:35:41 | Rui Xiao, Lu Han, Xiaoying Zhou, Jiong Wang, Na Zong, Pengyu Zhang | cs.CL | cs.CL, cs.AI, cs.HC | In online learning platforms, particularly in rapidly growing computer
programming courses, addressing the thousands of students' learning queries
requires considerable human cost. The creation of intelligent assistant large
language models (LLMs) tailored for programming education necessitates distinct
data support. However, in real application scenarios, the data resources for
training such LLMs are relatively scarce. Therefore, to address the data
scarcity in intelligent educational systems for programming, this paper
proposes a new Chinese question-and-answer dataset for Python learners. To
ensure the authenticity and reliability of the sources of the questions, we
collected questions from actual student questions and categorized them
according to various dimensions such as the type of questions and the type of
learners. This annotation principle is designed to enhance the effectiveness
and quality of online programming education, providing a solid data foundation
for developing the programming teaching assists (TA). Furthermore, we conducted
comprehensive evaluations of various LLMs proficient in processing and
generating Chinese content, highlighting the potential limitations of general
LLMs as intelligent teaching assistants in computer programming courses. | QACP: An Annotated Question Answering Dataset for Assisting Chinese Python Programming Learners: In online learning platforms, particularly in rapidly growing computer
programming courses, addressing the thousands of students' learning queries
requires considerable human cost. The creation of intelligent assistant large
language models (LLMs) tailored for programming education necessitates distinct
data support. However, in real application scenarios, the data resources for
training such LLMs are relatively scarce. Therefore, to address the data
scarcity in intelligent educational systems for programming, this paper
proposes a new Chinese question-and-answer dataset for Python learners. To
ensure the authenticity and reliability of the sources of the questions, we
collected questions from actual student questions and categorized them
according to various dimensions such as the type of questions and the type of
learners. This annotation principle is designed to enhance the effectiveness
and quality of online programming education, providing a solid data foundation
for developing the programming teaching assists (TA). Furthermore, we conducted
comprehensive evaluations of various LLMs proficient in processing and
generating Chinese content, highlighting the potential limitations of general
LLMs as intelligent teaching assistants in computer programming courses. | [
0.0619833209,
-0.0034823997,
-0.0265947282,
0.0468649678,
0.0083604315,
0.0384695753,
0.0195856243,
-0.0353926755,
-0.0009058784,
-0.0459156372,
-0.0108935395,
-0.0060486617,
-0.0285387766,
0.0019560643,
0.0340918303,
-0.0380419232,
0.0028963722,
-0.0050873845,
0.01121864,
0.0328840576,
0.025379993,
0.0354529507,
-0.0365081392,
0.0707063973,
0.0054932781,
-0.0412350893,
-0.0029354459,
-0.0278572999,
-0.0198905226,
-0.0401894674,
0.0344763249,
0.0168890171,
-0.0046381238,
-0.0023251376,
0.0000020088,
-0.0610999949,
-0.0195003543,
-0.0076624248,
-0.0430965573,
0.0355017446,
-0.0134665519,
0.0607130788,
0.0330799371,
0.0464559346,
-0.0763804913,
-0.0107460897,
0.0320388302,
0.0075450283,
0.1041989475,
0.0428639129,
0.0194512624,
-0.0047991551,
0.0305095855,
0.041597601,
0.0417682827,
-0.0225427207,
0.0542467125,
0.0697318986,
0.0159484427,
-0.0009131617,
0.0019723917,
-0.0131128654,
0.0071895579,
-0.0369686522,
0.0107241599,
-0.0047289222,
-0.0677698404,
-0.0130348653,
0.0055299285,
0.0734446719,
0.0140323481,
-0.0346112996,
-0.0171790142,
-0.009033327,
0.021797603,
0.0247217044,
-0.0692623183,
0.0323895812,
-0.014015438,
-0.0153568797,
0.0199781898,
-0.0415640809,
-0.0380417779,
0.0324960425,
-0.0200819913,
0.0381148495,
-0.0186330657,
-0.0439763665,
-0.043168433,
0.0351837315,
-0.0235846695,
-0.0127996681,
0.023856461,
0.0825490579,
0.0112865074,
0.0309859943,
-0.0010256671,
-0.0023401948,
0.001145654,
-0.0331441313,
0.0215139315,
0.0232525412,
-0.019435307,
0.0302331988,
-0.0178299677,
0.0501939505,
0.0226167757,
0.0247321595,
-0.0514317416,
-0.0119709419,
-0.074960418,
-0.0306379832,
-0.0410676152,
0.0772381127,
0.0270255003,
0.028526973,
-0.0384759121,
0.0609897487,
-0.0339549892,
0.0287977606,
-0.0240044743,
0.0087891473,
0.0046073943,
0.0178062581,
-0.0578748956,
0.0455456302,
-0.0389093384,
0.0085511291,
-0.0213265922,
-0.06258066,
0.0217837803,
0.025391316,
0.0319957249,
-0.0372197181,
0.0238540694,
0.0697157606,
0.0137109663,
-0.0109219979,
-0.0401224531,
-0.0323534459,
0.0459125564,
0.0282303337,
-0.001316709,
-0.0955360308,
-0.0233369879,
0.031389039,
0.0093613164,
-0.0127908513,
0.0137916962,
0.041441258,
0.0395247377,
0.0127185592,
-0.0743914619,
0.0187867191,
0.0613603331,
-0.0164147876,
0.0046471581,
0.014338701,
0.0315471515,
-0.0182429645,
0.0776820555,
-0.0418458283,
0.0434708409,
-0.0228387285,
-0.0002627299,
0.0023957265,
0.0092919515,
0.0820832625,
0.0282590687,
-0.0043726983,
-0.0410980396,
0.014105333,
0.0073607001,
0.008428866,
0.0030405265,
0.0479282849,
-0.02930752,
0.0315526426,
0.0049306774,
0.0410742834,
0.0480765738,
-0.0036670258,
-0.0107264696,
0.0641493872,
-0.0145847881,
0.0068692751,
0.0181016084,
0.0077635623,
-0.0215729903,
-0.1070510969,
-0.040179953,
0.0135906385,
-0.0341968387,
0.048671484,
0.0463574491,
-0.0016190087,
0.0189095289,
-0.0005519819,
0.0343005545,
-0.0463093035,
0.0036693462,
-0.0707325786,
0.054698877,
0.0671978891,
0.0361048132,
-0.0933713689,
-0.0591955557,
-0.0191854294,
0.0406105369,
-0.0011557139,
0.028235808,
-0.0480837673,
-0.0396803394,
-0.0236668941,
-0.0225723013,
-0.0139214844,
-0.0295774564,
-0.0241191033,
0.0310686976,
-0.0162927806,
0.0039366675,
-0.0031462908,
-0.0534745939,
0.0527798533,
-0.0152612822,
-0.0431160778,
-0.0215889513,
0.0132546974,
-0.0065646702,
0.0534378737,
-0.014677085,
0.0272095669,
0.0339383632,
0.0001465237,
-0.0031904667,
-0.0445228554,
-0.0233063791,
-0.0177679136,
0.1021989509,
-0.0530595593,
-0.0283961855,
0.0253223963,
-0.0339740142,
-0.0398929119,
0.0047200271,
-0.0354638211,
0.0216575097,
-0.0214955788,
0.0152773913,
-0.0222315155,
0.0028825754,
-0.0143792285,
0.0298463516,
0.0681159273,
0.0713262558,
0.0271281339,
0.0337898843,
0.0306146741,
-0.0213689096,
-0.0391549952,
-0.0150338216,
-0.0328539871,
0.0465400815,
0.038175486,
0.0331749134,
-0.0592852868,
0.0202875119,
-0.0319427364,
0.003150468,
0.0057290355,
0.0201916154,
0.0382509381,
-0.0120140174,
-0.0140684228,
-0.032629732,
-0.0669606179,
-0.020077195,
-0.0382522643,
0.0106083723,
0.0434875302,
-0.0097409533,
0.000090435,
-0.0381188951,
-0.0159662347,
-0.0356983952,
0.0106116999,
0.0555933155,
-0.0307870582,
-0.0426181518,
0.0077796434,
-0.0315261856,
-0.0243227798,
0.0160327516,
-0.0569133498,
-0.0138131706,
-0.0440638587,
-0.0078858566,
0.0744271725,
0.0112489359,
-0.0106984507,
0.040268369,
0.0173826907,
-0.0466479324,
0.0261441078,
0.0048113484,
-0.0345837697,
0.0226030741,
-0.047239054,
-0.0111835254,
0.0300828982,
0.0091668963,
-0.0293083098,
-0.0032768613,
0.0228922851,
-0.0263422504,
0.0382054821,
-0.0562709607,
0.0173511766,
0.0070330352,
0.0066755973,
-0.0018081511,
-0.0273540355,
0.0979945064,
-0.0050858539,
-0.0232404619,
-0.0535281077,
-0.0520229302,
-0.0550978966,
0.0352120772,
0.0060713212,
0.0032416522,
-0.0291427895,
-0.0183017161,
-0.0458426476,
0.0323098004,
0.018482687,
-0.0501187518,
0.0149729783,
-0.0264479481,
0.0030246573,
-0.0200151764,
0.0210561845,
-0.0345498398,
0.0138040576,
-0.0280449465,
-0.022510441,
-0.0013566528,
-0.0045460504,
0.0567720458,
0.0011680848,
0.0456129201,
0.0063183336,
-0.016628826,
-0.0193892103,
-0.0000095596,
0.0303493608,
0.030174138,
0.0264916774,
-0.0168290678,
-0.0149056828,
0.0745356679,
0.0382041819,
0.0427899659,
0.0144203659,
0.0194628183,
-0.0044003925,
0.0221526995,
-0.0577941239,
-0.0042251963,
0.026828561,
0.0118293418,
-0.084924683,
-0.0146127772,
0.0975393578,
0.0478522256,
-0.0269982852,
-0.0297398809,
-0.041531533,
-0.0296781026,
0.0608198531,
0.0459603332,
-0.1087506562,
-0.0225233007,
-0.0613461174,
0.0474474542,
-0.0256720372,
-0.0243455693,
-0.0026830002,
-0.1377336681,
0.0025197002,
-0.0282634608,
-0.0249825884,
-0.0257167257,
-0.066529803,
-0.0048871166,
-0.008917084,
0.0142458221,
-0.0042617065,
-0.0297986381,
0.0166071765,
0.077187188,
-0.0360132493,
-0.0070262603,
0.0176224615,
0.0032876313,
0.0256058332,
0.046380166,
0.0290611535,
-0.0426795669,
0.0012067852,
-0.0020047917,
-0.0334490165,
-0.0362885296,
0.003142166,
0.0365021154,
-0.0139036346,
-0.0226359386,
0.0039109266,
0.0250601321,
-0.0659404621,
0.0240457188,
0.1062841788,
0.0532171316,
-0.0118346559,
0.0289373621,
-0.0406894833,
-0.0111355027,
-0.0294768438,
0.078427963,
0.0518415868,
-0.0358261727,
-0.0092132902,
-0.0640775561,
-0.0309743937,
-0.0370231457,
0.0057037654,
0.0386631526,
-0.0286823064,
0.0273506772,
-0.0866435245,
-0.0143014444,
0.0866345838,
-0.0146555621,
-0.0008781373,
0.0240959376,
0.0337258466,
0.041582413,
-0.0652275458,
-0.005080577,
-0.0170852263,
-0.0011754023,
-0.0639041513,
-0.0663325414,
0.0232337341,
-0.0481591821,
-0.0185999684,
-0.0142673301,
-0.0575466231,
-0.0416481085,
-0.0115634529,
0.0261878576,
0.0103786336,
-0.0007920719,
0.0065644546,
-0.0245190915,
-0.0197594166,
0.0228438452,
-0.0213781986,
-0.0320501104,
-0.0075487387,
0.0312698074,
0.0118868146,
-0.0225813147,
-0.0686785802,
-0.0050031301,
-0.0023335542,
0.0019800849,
-0.0479016937,
-0.0190731212,
0.0432342589,
0.0666578487,
-0.0335080214,
0.0613243021,
-0.0233382545,
-0.0312904157,
0.0456736125,
-0.0149569381,
-0.0303585995,
-0.0200487357,
0.0136487465,
0.0134081757,
-0.0358967558,
0.0258079115,
0.0109993527,
-0.0018050707,
0.0056345775,
-0.0292457324,
0.0620266013,
0.00179246,
0.0426758602,
-0.013483258,
-0.0144957369,
-0.0312546715,
-0.0101487013,
0.0289927013,
0.0567962676,
0.0190655347,
0.0511914678,
-0.0012302435,
0.0057856631,
0.0422778577,
0.0158077069,
-0.0616002344,
-0.1015459821,
-0.0224512499,
0.0468861312,
0.0258206856,
0.0086008152,
0.0089223096,
-0.0264189485,
-0.0041160425,
0.0038250091,
0.0348045491,
0.0135397576,
-0.0463837534,
0.0201049522,
0.0339079946,
-0.0076878476,
-0.0040836083,
0.08873602,
-0.0084060412,
0.0084660845,
0.0188973919,
-0.0268222913,
-0.0238876492,
0.0217318125,
0.0344378687,
0.0224253759,
0.0242070686,
0.0124483909,
0.0048688501,
-0.0797949433,
-0.0383722112,
0.0174031872,
0.0173114929,
-0.010163812,
0.0408655666,
0.0535148792,
0.007313096,
0.0066087949,
-0.0565741919,
0.0279509332,
-0.0274276249,
-0.0166902393,
0.0011809563,
-5.75550944e-33,
-0.0252622422,
-0.0319684595,
0.0101417117,
0.0336455591,
0.0479677543,
0.0274994578,
-0.0128668165,
0.0086590862,
-0.0725529194,
-0.0251402017,
0.0049683931,
0.0467495881,
0.0014651947,
0.0358753242,
0.0152624082,
-0.024913026,
0.0047629909,
0.0111882426,
0.0265566856,
0.0555996969,
0.0376420617,
-0.0016410477,
0.0566936545,
-0.0261892341,
0.0199668538,
-0.0571224764,
0.0221372042,
0.0028409865,
-0.009113865,
0.0821953416,
-0.0101332832,
0.0530175343,
0.0462923832,
-0.0302708652,
-0.0123387044,
0.0050625456,
-0.0510430187,
-0.0664179549,
-0.0342268236,
-0.004062939,
-0.0091129504,
-0.011249003,
0.0403388366,
-0.0126630897,
0.0435162522,
0.0223539583,
0.0967270732,
-0.042025663,
0.0250604544,
-0.013457383,
-0.0344012864,
0.0103542022,
-0.0734089017,
-0.0041051367,
-0.0261039976,
0.0285036732,
-0.0087991683,
0.0107952291,
-0.0074093039,
0.0239719171,
0.0360149518,
0.0601158701,
0.0379545428,
-0.0016010817,
0.0004786618,
-0.0081279762,
0.0332977511,
0.0030692662,
0.0120024337,
0.0057271901,
-0.013311184,
0.022736894,
0.0567783192,
0.0394879133,
0.0075366604,
0.0026497366,
-0.0827115253,
0.0486098416,
0.056965258,
-0.0178082902,
0.0026936529,
-0.0352950729,
-0.052436009,
0.0071908934,
-0.0118273273,
0.0376902521,
-0.0305611249,
-0.0225155763,
0.0250176154,
-0.0613123775,
-0.0475870706,
0.0065036458,
-0.000498235,
-0.047332786,
-0.017805716,
0.0175330695,
-0.0233645439,
0.037956275,
-0.0231097601,
0.0231940858,
-0.0454021506,
-0.0030388068,
-0.0173706785,
-0.0615110621,
0.0230064243,
0.0226484835,
0.0244636424,
-0.0107970992,
-0.0508556589,
-0.0155336056,
-0.0452229604,
-0.0213412307,
0.0024410274,
0.0349103659,
0.001974286,
0.0374939442,
0.00360325,
0.0564132296,
0.0235826187,
-0.0208274014,
-0.0630622953,
-0.0681545064,
-0.0247649252,
-0.0387052819,
-0.0477712303,
-0.0115066338,
-0.0704311132,
0.0473325402,
-0.0329251736,
-0.0641185343,
-0.0132283559,
0.0125447689,
2.61e-7,
0.0490403175,
0.0192218032,
0.0328621268,
0.0492741689,
-0.0310785864,
-0.0035348528,
0.0148890624,
0.0068966579,
0.0399462208,
0.0182290077,
-0.0022856316,
0.0140354959,
0.0205885936,
0.0408933163,
-0.0843752772,
0.0026396581,
-0.0408809297,
-0.0181757156,
-0.064602837,
0.059435226,
0.0942474008,
0.0508014299,
0.0019483707,
0.0155453142,
-0.0216747727,
-0.0471035875,
-0.0283089895,
0.0133792041,
0.0167884994,
0.0693977326,
0.0017973956,
0.0264635086,
0.0143699376,
0.0054709036,
-0.0057833474,
-0.0061014197,
0.0183750447,
-0.0019458993,
0.0185060278,
-0.0057729403,
0.0160694886,
-0.0176927336,
0.0331757553,
-0.0615511388,
0.048587136,
-0.0050301398,
-0.0079711163,
0.0218109582,
-0.0163652711,
-0.038981881,
-0.0179197341,
-0.0257917549,
0.0172779895,
0.0108528696,
-0.0213501006,
-0.0577691756,
-0.0111353565,
-0.0315126739,
-0.0049132747,
-0.0188196599,
0.0089182826,
-0.0499617457,
-0.0128327124,
-0.035650935,
-0.004489488,
-0.0182455257,
0.0301921107,
2.024726342e-34,
-0.0196694955,
-0.0375791453,
0.0435693413,
0.0582503527,
0.0336428322,
0.0087023545,
-0.0094174519,
0.003784077,
0.0106557338,
-0.0153682362,
-0.0229408592
] |
http://arxiv.org/abs/2401.16968v1 | http://arxiv.org/pdf/2401.16968v1 | Distinguishing Fictional Voices: a Study of Authorship Verification Models for Quotation Attribution | 2024-01-30T12:49:40 | 2024-01-30T12:49:40 | Gaspard Michel, Elena V. Epure, Romain Hennequin, Christophe Cerisara | cs.CL | cs.CL | Recent approaches to automatically detect the speaker of an utterance of
direct speech often disregard general information about characters in favor of
local information found in the context, such as surrounding mentions of
entities. In this work, we explore stylistic representations of characters
built by encoding their quotes with off-the-shelf pretrained Authorship
Verification models in a large corpus of English novels (the Project Dialogism
Novel Corpus). Results suggest that the combination of stylistic and topical
information captured in some of these models accurately distinguish characters
among each other, but does not necessarily improve over semantic-only models
when attributing quotes. However, these results vary across novels and more
investigation of stylometric models particularly tailored for literary texts
and the study of characters should be conducted. | Distinguishing Fictional Voices: a Study of Authorship Verification Models for Quotation Attribution: Recent approaches to automatically detect the speaker of an utterance of
direct speech often disregard general information about characters in favor of
local information found in the context, such as surrounding mentions of
entities. In this work, we explore stylistic representations of characters
built by encoding their quotes with off-the-shelf pretrained Authorship
Verification models in a large corpus of English novels (the Project Dialogism
Novel Corpus). Results suggest that the combination of stylistic and topical
information captured in some of these models accurately distinguish characters
among each other, but does not necessarily improve over semantic-only models
when attributing quotes. However, these results vary across novels and more
investigation of stylometric models particularly tailored for literary texts
and the study of characters should be conducted. | [
0.0221035443,
0.04850531,
-0.0276928227,
0.0407915749,
-0.0334001407,
0.0189161077,
0.0690929294,
-0.0098960623,
0.0204648878,
-0.0371642597,
-0.0089372294,
-0.021362111,
0.0530824326,
-0.007043907,
0.0232961476,
0.0262859538,
-0.0030211012,
-0.0191101134,
0.0620256849,
0.0003089014,
0.0129604042,
0.0409000888,
0.0152897295,
0.0591740683,
-0.0040964037,
-0.0619539022,
0.0352887437,
0.0142978625,
0.0313331336,
0.0161021724,
0.0296060964,
0.0400121845,
-0.0273708533,
0.0230079312,
0.0000015584,
-0.0350530595,
0.031067498,
-0.0450690091,
-0.0187948775,
-0.005315823,
0.0495255925,
0.0754378289,
0.0050704065,
-0.0083709769,
0.0084923571,
-0.0183847528,
0.0144692138,
0.0444400385,
0.0109628234,
0.0626882762,
-0.032178618,
-0.0603704266,
0.0056563616,
0.0287693124,
-0.0072048418,
0.013042016,
0.0344768204,
-0.0375001915,
-0.0197529625,
0.0185882002,
-0.0019299192,
0.0255681649,
-0.0453513078,
0.0030527683,
-0.0012810319,
0.0071344692,
-0.0262874179,
0.0486924313,
-0.0001668033,
0.0148009304,
0.0292782225,
-0.0497085899,
0.0054763341,
0.007096523,
0.0029128224,
0.0273054801,
0.0319705643,
-0.0015038332,
-0.0709208697,
-0.013125388,
0.0285859685,
0.0133386571,
0.0124406544,
0.0195446275,
-0.0096929874,
0.0025564863,
0.0173486285,
0.0110760899,
-0.0121965855,
-0.0090005975,
-0.024049798,
-0.0470099896,
0.0283489842,
0.0052419417,
0.0349331535,
-0.017088173,
-0.0296567027,
-0.0231927428,
-0.0147652803,
-0.0307688527,
0.0345053226,
-0.007076025,
-0.0381428786,
0.0025615846,
-0.0198194291,
0.0191266257,
-0.028124515,
0.0209200513,
0.0046815858,
-0.0331926979,
-0.0443872437,
-0.0415391624,
-0.0543466881,
-0.0061002253,
0.0100916671,
-0.0073110461,
-0.0312470011,
-0.0179550927,
0.0228768177,
0.0032157709,
-0.0506371632,
0.0094611049,
-0.0111353528,
0.0110599957,
-0.0273733959,
0.0046865088,
-0.0164656062,
0.0004169816,
-0.0383755341,
-0.0048348792,
-0.0112637999,
0.0209874306,
0.0435449034,
-0.0374962091,
0.0471365117,
0.0513520613,
-0.0069995909,
-0.0308717545,
-0.0202358235,
-0.0241325349,
-0.0477431044,
-0.0255462583,
-0.0476857089,
-0.0108307526,
0.0142862927,
-0.0318143107,
0.0342629552,
-0.0530640632,
-0.0073130876,
0.0052076001,
-0.025959136,
0.0514761396,
0.0186933279,
0.0203992799,
0.0502728447,
-0.0292928368,
-0.0059130453,
-0.0313656256,
0.0888424888,
0.0116717415,
0.0240042489,
-0.0038636832,
0.0447267853,
0.0500918031,
-0.0248248465,
-0.0185379554,
-0.0055228956,
-0.0248615742,
0.002183046,
0.019914981,
0.014601348,
0.0412146449,
-0.0305609256,
-0.0310405642,
0.0966613293,
0.0782300532,
0.0350088626,
0.0361119285,
0.055108171,
0.0226461813,
0.0581223108,
-0.0086195953,
-0.0096245697,
-0.0158039723,
-0.0083203223,
0.0014997182,
-0.101407446,
0.0171234049,
-0.0020461453,
0.0269535072,
0.0007118033,
0.0078848572,
-0.0083338199,
-0.0136400182,
-0.0022731449,
-0.0044843443,
-0.0346113816,
-0.0202549398,
-0.0091472613,
-0.0624575131,
-0.0146544902,
-0.0189684909,
0.0045197541,
0.0571369268,
0.0709046498,
-0.0333740935,
-0.004486864,
-0.0565807447,
-0.0316826664,
0.0194595698,
-0.0194504838,
-0.0442943014,
0.0666756481,
-0.0364391543,
0.0470441133,
0.0062161637,
0.0093170581,
0.0719863698,
-0.04492753,
0.0082903616,
-0.0170293003,
0.00250279,
-0.0262291078,
0.0426770672,
-0.0105294036,
-0.0508513972,
-0.0105356192,
0.0380152576,
-0.0022545944,
-0.007049385,
-0.0628313199,
-0.0328703262,
-0.0209631082,
-0.0787395388,
0.0040971562,
0.0861722454,
-0.0302434564,
-0.0192012321,
0.0457222052,
-0.028029073,
-0.0049389522,
-0.0114811519,
0.0199350324,
-0.0465704203,
0.0167180374,
-0.0257004369,
0.0105055394,
-0.0491320342,
0.030089898,
-0.0313546471,
-0.0055625681,
-0.0554692112,
0.0491558909,
0.0710701272,
-0.0030551814,
0.0026145296,
0.0064375415,
0.0530686453,
-0.0302040763,
-0.0289536528,
-0.0240575001,
0.0352215469,
0.0149718774,
0.020385338,
-0.0514235049,
0.0106289983,
-0.0884966031,
-0.0689899176,
-0.0128399199,
0.0233790576,
0.0100621879,
-0.006981811,
-0.0132965175,
-0.0302267894,
0.0222197846,
-0.036597617,
-0.0665117353,
-0.0418332852,
-0.0441271178,
0.0465328619,
-0.0006536954,
-0.0107534109,
-0.0332775153,
0.016714219,
-0.0452113636,
0.0565147363,
0.0251336731,
-0.0073438832,
-0.0133798942,
-0.0095414985,
0.00602993,
0.0076380544,
-0.0162632093,
-0.0282314233,
0.0131763304,
0.001774668,
0.0218943004,
0.0780098587,
0.0078894431,
-0.019928826,
-0.0003618198,
-0.0474361219,
-0.0232987162,
0.019801572,
0.0396374762,
0.0123552112,
-0.0283472519,
-0.0298064332,
0.0211904682,
0.0309347361,
0.0131504508,
-0.009961538,
-0.0215538498,
0.0572941601,
0.054338485,
-0.043102663,
0.0358042866,
-0.0418218523,
-0.032032378,
0.0207581315,
-0.0187728256,
0.0338134505,
0.0393923596,
-0.0565279759,
-0.0256868713,
-0.0057343403,
0.0024814634,
-0.0301586054,
0.0244581811,
0.0160315502,
0.030994365,
0.055167634,
-0.0498740189,
-0.0245102439,
-0.0107122241,
0.0737407207,
-0.0282945,
0.0442094319,
0.0141831487,
-0.0032470322,
0.0169752184,
0.002934912,
-0.0344427824,
-0.0793775916,
-0.0483381823,
-0.0250563156,
0.0019580342,
-0.0495367795,
0.0031078199,
-0.0272060428,
-0.0129401935,
0.0337593295,
0.0081194043,
-0.005260753,
-0.0224118959,
0.0170087386,
-0.0182973165,
-0.0001485587,
-0.0333194919,
-0.0303905737,
0.0398176387,
-0.0394478813,
-0.0629454181,
0.0236157812,
-0.0517264381,
0.0125955772,
0.0233136062,
-0.0320581906,
-0.0372786075,
-0.0180455931,
0.0728751495,
-0.0506492294,
-0.003762508,
0.1192327216,
-0.0218116827,
-0.011987987,
-0.0430162214,
0.0074932361,
0.0032096957,
0.0214060247,
0.007771424,
-0.0536800735,
-0.0169020128,
0.0042097,
0.0135587966,
-0.0286175013,
-0.0411800966,
-0.0235534757,
-0.0611401349,
0.0443273261,
0.005107597,
-0.0326580964,
0.0228277706,
0.0310661886,
-0.0299439803,
-0.0122239143,
-0.0391005054,
-0.0090536429,
-0.0137109878,
-0.0035174959,
-0.0155036151,
-0.005776661,
0.0459555537,
0.030896537,
-0.0480241589,
0.0221011117,
0.0261867046,
0.0944696218,
-0.0321732052,
0.0181395803,
-0.0748204589,
-0.0423943922,
-0.0098263333,
0.0040270844,
0.0884310082,
-0.0558066256,
-0.0253476333,
-0.0492154174,
0.0344195962,
0.0255511142,
-0.0242508613,
0.0347948931,
0.009924476,
-0.010367427,
-0.0036106799,
0.0003295085,
0.0429348089,
0.0038796889,
0.0388361774,
0.0284809787,
-0.033754427,
-0.0602420866,
-0.0167242177,
0.0511541106,
-0.0329187065,
-0.045514632,
0.0092651015,
-0.0103543783,
-0.0305750072,
-0.0731136426,
-0.0048016668,
-0.0308823343,
0.0365718715,
0.0802918822,
0.0246834159,
0.0094802082,
0.0099173794,
-0.0134075126,
-0.0098795798,
-0.0405218638,
0.005332191,
-0.0022629437,
0.0159821585,
0.0513964742,
-0.0574803837,
0.0374135934,
0.0163850207,
-0.0423262492,
-0.0492896922,
0.0290309694,
-0.0178108048,
0.0247609671,
0.0420234166,
-0.0443855561,
-0.0673625097,
0.0575410351,
-0.0009794751,
-0.0022306812,
0.0663663894,
0.0352949612,
0.0154127991,
-0.0630024523,
-0.0023871544,
0.0062773996,
0.0615766384,
-0.0045075179,
-0.0628759488,
-0.0325330608,
-0.0378506258,
-0.0041548312,
0.0287126917,
-0.0060905511,
-0.0462858602,
0.0352009311,
-0.0337382182,
-0.0154982302,
-0.0072189956,
-0.0044924784,
0.0012913853,
0.0155024221,
0.016030252,
-0.0059052962,
-0.0376618244,
-0.0265783351,
-0.0164431371,
0.0958739668,
-0.0044579008,
0.007731664,
0.0504149795,
0.051175788,
-0.0629547611,
0.0275082942,
-0.0214543063,
0.0313054249,
-0.0236256924,
0.0710800141,
0.0398855247,
0.0272993091,
-0.0006794505,
0.0053829351,
-0.0418068394,
0.0156852733,
-0.1042362526,
-0.0118248453,
0.0111122942,
-0.0206676666,
0.036345385,
-0.0430538505,
0.0470779985,
-0.0153758731,
0.0104563804,
0.0004653451,
0.0376195684,
0.0300401859,
0.0630644113,
-0.0543528497,
0.0225561112,
-0.02393323,
0.015731765,
0.021600198,
0.0254484229,
0.0397070274,
0.0241100807,
-0.0554858297,
-0.0435582474,
-0.0539841056,
0.0078291399,
0.0162943024,
0.0386719517,
-0.0085571511,
-0.0758970305,
-0.0346026979,
0.1217464283,
-0.0014500709,
0.066047892,
-0.0384906754,
-0.0150275473,
0.0155304205,
0.0077916575,
-0.0028623657,
-0.0327977724,
0.0512152165,
0.0462926142,
0.0170494337,
-0.0192575399,
-4.435100513e-33,
-0.0396491922,
0.0195739139,
-0.0243625715,
0.0834522992,
-0.0439599678,
-0.0196448397,
-0.0064197779,
-0.045286037,
0.0083162934,
0.0361439288,
-0.021056857,
0.0114814388,
0.0032019331,
-0.0292651802,
0.0148811396,
0.008415225,
0.0674264356,
0.0136850029,
-0.0401489139,
-0.0154801123,
0.0768289566,
0.0035331736,
0.0383279473,
-0.0592062138,
0.0241710786,
-0.0817085057,
-0.0452355705,
-0.0306798499,
0.0243382826,
0.018377291,
-0.0046636318,
0.0357067659,
-0.0032715462,
0.0051217102,
-0.0190006346,
0.0430446714,
-0.0830145776,
-0.0369194783,
0.0105755376,
0.0111538172,
0.0926634148,
-0.0676376745,
0.0014876182,
-0.0107394382,
-0.0057551395,
-0.0224984121,
0.0847380906,
-0.0163503531,
0.0127761196,
-0.0568780489,
-0.0822933093,
-0.0004461894,
-0.0308456216,
0.0017201874,
0.0677102804,
0.0145097263,
0.0054211929,
-0.020957429,
-0.0267193206,
-0.0214862432,
-0.0373200588,
-0.0253589936,
-0.0613070987,
0.0779788643,
0.0108468672,
0.0776534379,
0.0470629707,
0.0204161145,
-0.0181654301,
0.0313841179,
0.0478584059,
-0.0099835349,
0.0121694868,
0.1261975318,
-0.0032724466,
0.0062433248,
-0.0315988846,
-0.0239323098,
0.0477371626,
-0.0204857048,
0.0142058488,
-0.0190178137,
-0.0329301953,
-0.0472417511,
-0.0383157283,
-0.0469545722,
0.0033745815,
0.0039104051,
-0.0142894965,
-0.0566982403,
-0.020776445,
0.0083849579,
0.0404584147,
-0.0533996187,
0.0170778576,
-0.0096933898,
-0.0375595503,
0.0760837644,
-0.0017159879,
0.0424911939,
0.0186479893,
-0.0400629677,
0.0267260522,
0.0264315326,
0.0565587431,
0.0141856382,
0.0317478962,
0.018728368,
-0.0534198731,
-0.0209128726,
0.0199691039,
0.0179170575,
-0.0259443149,
0.0571201108,
0.0006796918,
0.0122460807,
0.027890237,
0.0204716455,
-0.0134345619,
0.0033522127,
-0.0253595319,
-0.0268870499,
0.0228126422,
-0.0084731774,
-0.0531652085,
-0.0023748695,
0.0343350656,
-0.0323554017,
-0.0062591252,
-0.033084318,
0.0129900184,
-0.0506847575,
2.265e-7,
0.0259874798,
-0.0140292151,
0.080591999,
-0.0034793688,
0.0086097959,
0.0535203367,
0.0232068449,
0.0234170649,
0.0233247224,
-0.0419149399,
0.0257498492,
0.0133995144,
0.0518287979,
0.0224128049,
-0.0793735534,
-0.0031371498,
-0.0199980605,
0.0019068421,
-0.0425368883,
0.0042176959,
0.043181479,
0.0141347926,
0.0617875457,
-0.0284831673,
0.005511119,
-0.0568204448,
0.0572298951,
-0.0218895171,
-0.014906018,
0.0671489239,
0.0465689301,
-0.0236675702,
0.003521326,
0.0580834784,
0.0069674584,
0.0450257286,
0.0011631898,
0.0385583676,
-0.0436523706,
0.0044616079,
-0.0371207632,
-0.0338708982,
-0.0293234624,
-0.0011134582,
0.0169100258,
0.0544320941,
0.0033862176,
-0.0450224727,
-0.019328244,
-0.0098058172,
0.0263387207,
0.0150283678,
-0.0124273552,
0.0243155248,
0.0056651114,
-0.0107339611,
0.0216321461,
-0.0933115482,
-0.0169157702,
0.0198012721,
-0.0441393703,
-0.0286866371,
0.0399468616,
-0.0144452546,
0.0455487408,
-0.1010348648,
0.0376121812,
1.992725574e-34,
0.0085461987,
0.0194609724,
0.0062416564,
0.0939006135,
0.0298045389,
0.0061690402,
0.0424137041,
0.0104868542,
0.0396492593,
-0.0252022985,
0.007855772
] |
http://arxiv.org/abs/2401.16960v1 | http://arxiv.org/pdf/2401.16960v1 | Two Heads Are Better Than One: Integrating Knowledge from Knowledge Graphs and Large Language Models for Entity Alignment | 2024-01-30T12:41:04 | 2024-01-30T12:41:04 | Linyao Yang, Hongyang Chen, Xiao Wang, Jing Yang, Fei-Yue Wang, Han Liu | cs.CL | cs.CL, cs.AI | Entity alignment, which is a prerequisite for creating a more comprehensive
Knowledge Graph (KG), involves pinpointing equivalent entities across disparate
KGs. Contemporary methods for entity alignment have predominantly utilized
knowledge embedding models to procure entity embeddings that encapsulate
various similarities-structural, relational, and attributive. These embeddings
are then integrated through attention-based information fusion mechanisms.
Despite this progress, effectively harnessing multifaceted information remains
challenging due to inherent heterogeneity. Moreover, while Large Language
Models (LLMs) have exhibited exceptional performance across diverse downstream
tasks by implicitly capturing entity semantics, this implicit knowledge has yet
to be exploited for entity alignment. In this study, we propose a Large
Language Model-enhanced Entity Alignment framework (LLMEA), integrating
structural knowledge from KGs with semantic knowledge from LLMs to enhance
entity alignment. Specifically, LLMEA identifies candidate alignments for a
given entity by considering both embedding similarities between entities across
KGs and edit distances to a virtual equivalent entity. It then engages an LLM
iteratively, posing multiple multi-choice questions to draw upon the LLM's
inference capability. The final prediction of the equivalent entity is derived
from the LLM's output. Experiments conducted on three public datasets reveal
that LLMEA surpasses leading baseline models. Additional ablation studies
underscore the efficacy of our proposed framework. | Two Heads Are Better Than One: Integrating Knowledge from Knowledge Graphs and Large Language Models for Entity Alignment: Entity alignment, which is a prerequisite for creating a more comprehensive
Knowledge Graph (KG), involves pinpointing equivalent entities across disparate
KGs. Contemporary methods for entity alignment have predominantly utilized
knowledge embedding models to procure entity embeddings that encapsulate
various similarities-structural, relational, and attributive. These embeddings
are then integrated through attention-based information fusion mechanisms.
Despite this progress, effectively harnessing multifaceted information remains
challenging due to inherent heterogeneity. Moreover, while Large Language
Models (LLMs) have exhibited exceptional performance across diverse downstream
tasks by implicitly capturing entity semantics, this implicit knowledge has yet
to be exploited for entity alignment. In this study, we propose a Large
Language Model-enhanced Entity Alignment framework (LLMEA), integrating
structural knowledge from KGs with semantic knowledge from LLMs to enhance
entity alignment. Specifically, LLMEA identifies candidate alignments for a
given entity by considering both embedding similarities between entities across
KGs and edit distances to a virtual equivalent entity. It then engages an LLM
iteratively, posing multiple multi-choice questions to draw upon the LLM's
inference capability. The final prediction of the equivalent entity is derived
from the LLM's output. Experiments conducted on three public datasets reveal
that LLMEA surpasses leading baseline models. Additional ablation studies
underscore the efficacy of our proposed framework. | [
0.031435959,
0.0402547829,
0.0040615876,
0.0461779013,
-0.0091684256,
0.0310538076,
-0.0164357293,
-0.0204328876,
0.0357793011,
-0.0340640135,
0.0080805728,
0.0226449873,
-0.0210991427,
-0.0308269877,
0.0706890002,
0.0247278381,
0.035973344,
0.0523647815,
-0.0248551313,
0.0028105585,
0.0302615426,
0.0308586322,
0.0028213609,
0.0127162235,
0.0067173583,
-0.0193812158,
-0.0014343782,
0.0172051545,
0.065615803,
-0.0127280029,
-0.0150998421,
0.0201792363,
0.0167564601,
0.0644008964,
0.0000016922,
-0.0489603579,
-0.0088641476,
0.0079181772,
-0.0511134379,
-0.0125752874,
-0.02315812,
-0.0302358568,
0.0360127203,
-0.0170079973,
0.0010119105,
-0.0477272943,
0.047438655,
0.0207955167,
0.0322513729,
0.0388099104,
-0.0084123211,
-0.0284883305,
0.0201084893,
-0.0203990992,
-0.013770408,
-0.0319766589,
0.0341146141,
0.0241132025,
-0.0210448466,
0.0171188526,
0.0045440597,
0.0589836352,
-0.0112100318,
-0.0148470486,
0.0641952679,
0.0356781371,
-0.0074078343,
0.0376670845,
0.0231047515,
0.0947347358,
-0.0136229713,
-0.030575214,
0.0086830733,
-0.0056645661,
0.0265150182,
0.0048564808,
-0.0229738615,
-0.0215246659,
-0.0224465262,
0.0082311416,
0.0090015307,
-0.0304900091,
0.0486318506,
-0.0088538164,
-0.0094814133,
0.0209754203,
0.0440009199,
-0.0330901779,
-0.0042444784,
0.0252071414,
-0.0060213697,
-0.0641134605,
0.0030982443,
0.0354394168,
-0.0446903035,
0.0254372265,
-0.0636562854,
0.0034697233,
0.0201217122,
-0.0560781993,
-0.0010491064,
0.0474890061,
0.0352016501,
0.0946065485,
-0.0467495434,
0.0318909921,
-0.0400812514,
0.0159649774,
-0.0456603952,
-0.0494893193,
-0.0458453894,
0.0192544386,
0.0099876821,
0.0262734704,
0.0117303822,
-0.0197927654,
-0.0409306027,
0.0299034771,
-0.0106190173,
0.0772869885,
-0.0111733815,
0.0112495786,
0.0241136644,
-0.0199355446,
-0.0553782322,
0.0336989909,
-0.0384580642,
0.0116041778,
-0.0183598734,
-0.0024145355,
-0.016876664,
-0.0087121921,
0.0171821415,
-0.0633460283,
-0.0003886459,
0.0487957299,
0.0173825249,
-0.003332749,
-0.0392390601,
-0.0110641979,
0.0257932004,
-0.0413111895,
-0.0116471397,
-0.0496623069,
0.0925408229,
0.0155552076,
-0.0283795372,
-0.0517704114,
0.0256595872,
-0.0075985449,
-0.0510228984,
0.0798548758,
-0.02543802,
0.0314867795,
0.0277489759,
-0.0268632956,
0.0029371649,
0.0284647848,
0.0107871257,
-0.0469446518,
0.020417368,
-0.0106771551,
-0.0639513284,
0.0446743779,
0.0393245891,
0.0082138591,
0.0199889727,
0.037956994,
0.009611873,
0.0566926971,
0.0548387729,
0.029707972,
-0.0091342181,
-0.0292870309,
0.0625753552,
0.0269970335,
-0.0005919391,
0.0117812855,
0.0451523848,
0.001636286,
-0.0148395589,
-0.0180024225,
-0.0568981916,
0.0356368572,
-0.021011902,
0.0209896378,
0.0519345812,
0.0484192558,
-0.0293159019,
0.0467258543,
-0.001348424,
0.010302566,
-0.0010976304,
0.0395840444,
0.011045319,
-0.0153327128,
-0.0286049359,
0.0874256045,
0.0230408125,
-0.007567368,
-0.0009143203,
0.0276298169,
0.0027647675,
0.0587459244,
0.0455439873,
-0.0053556757,
-0.0567780584,
-0.0339786187,
-0.0288835075,
0.0446725935,
0.0449951328,
-0.0261674281,
-0.0191699229,
-0.0255108643,
-0.0377696827,
0.0117980763,
-0.0216299146,
0.0034882985,
0.0368633084,
0.0257627685,
0.0178160854,
-0.0671004057,
0.0000634902,
0.0304230079,
-0.0444899537,
-0.0080009885,
-0.0425052121,
0.0561083294,
-0.0040740394,
0.0346952975,
0.0201402102,
0.0277449898,
0.024331471,
-0.0304880477,
-0.0164864752,
-0.0311033707,
-0.0004128057,
-0.0352676176,
0.0795948133,
0.0007195391,
-0.0696731433,
-0.0160181541,
-0.0507644117,
-0.0273435004,
-0.0159827266,
-0.0220452789,
0.0602381043,
-0.0067872102,
0.0248825923,
0.0022316787,
-0.0262169633,
-0.0253522862,
0.0754520148,
-0.0089322813,
0.0579993315,
0.0059562908,
-0.0127418302,
0.0222160835,
-0.024046462,
-0.0387542434,
-0.0312897116,
0.0091820117,
-0.0013107134,
0.0076544415,
-0.0471147709,
-0.0867242813,
0.0148379961,
-0.1050855815,
0.0533739179,
0.0235885158,
0.0333660543,
0.0147778178,
0.0252791941,
0.0106323715,
0.0065789903,
-0.0344598368,
-0.0163762476,
-0.0151007567,
0.0098974938,
0.0353542008,
-0.0239472426,
-0.0299608521,
-0.0441112071,
-0.0078818481,
-0.0472943895,
0.0391789377,
0.0189723298,
-0.0170398988,
-0.0456745587,
0.0209654607,
-0.079714939,
0.011266076,
0.0101786992,
0.0176126659,
-0.0291612037,
0.0359918959,
0.0543010458,
0.0529488623,
-0.0324545763,
-0.03528497,
0.0065463041,
-0.0121981613,
-0.0688976645,
0.0077161738,
-0.0087976549,
-0.0004602638,
-0.0172963291,
-0.006957035,
-0.0244330913,
0.0318970084,
0.0072561726,
-0.0346479416,
-0.0514313653,
0.0414049476,
0.0317944251,
0.0446952656,
-0.0310692769,
0.0137996934,
0.0243319254,
-0.0174892209,
-0.0686392859,
0.0303068738,
0.1098093539,
-0.0241749417,
0.0210014954,
0.0278241243,
0.0167458523,
-0.0335059986,
0.0304312948,
-0.0052837231,
0.0240198039,
-0.0267024282,
-0.0219100844,
-0.0405747928,
0.0462810881,
0.0332170613,
-0.0418003239,
0.0796786621,
0.0249428283,
0.0571270213,
-0.0086286226,
-0.0099763703,
-0.0424865857,
0.0010725622,
-0.0208143406,
0.0233005378,
0.0136696016,
0.0041220859,
-0.0698928684,
0.0062860618,
0.0194721464,
-0.0264193211,
0.0359960757,
-0.0102037629,
-0.0335033797,
0.0320259817,
0.0491970032,
0.0134266904,
0.0438046902,
0.0122333802,
0.0309446584,
0.0222537313,
0.036527954,
-0.0178963225,
0.0073000914,
-0.0332798176,
0.007156047,
-0.0859365389,
0.0243529584,
-0.0335491076,
0.0383615419,
-0.0149257006,
0.0174385961,
0.0997417048,
0.0057409769,
0.0046976958,
-0.0114499116,
0.0126037057,
-0.0095354673,
0.051987417,
0.0209973566,
-0.0910992697,
-0.0242690966,
-0.0092518367,
-0.0102504063,
0.0504549816,
-0.0868461356,
0.0098800473,
-0.1302293539,
-0.0245564859,
-0.0058469819,
-0.0311684981,
-0.0623532608,
-0.0324062854,
0.0495993681,
-0.000057609,
-0.006704947,
0.0024010886,
0.0017595208,
-0.0164552592,
-0.0299328361,
0.0378256105,
-0.0147682708,
0.0371389315,
-0.0921609551,
0.0483223572,
0.0147982715,
0.0539491847,
0.0330726281,
-0.0106349206,
-0.031368047,
-0.0373855308,
-0.0660577267,
0.0319207273,
0.0260646995,
-0.070547089,
0.0062350808,
-0.0350629725,
0.0115597472,
-0.0169356335,
0.005292559,
0.0069308802,
0.0842716843,
0.0220080279,
0.0080093406,
-0.0001221061,
-0.0194954742,
-0.0020682353,
0.0844302922,
0.0168854576,
-0.0005516153,
0.0144812018,
0.0091105774,
0.013162069,
0.0031000152,
-0.0019096656,
0.0475002304,
-0.0417446233,
0.0202337876,
-0.0299365241,
-0.063308008,
0.0471351966,
-0.0137339868,
0.0540303998,
0.0323000886,
0.0382468589,
-0.0668512359,
-0.0655217096,
0.0341148712,
-0.0513197556,
0.0297655035,
0.0238772742,
0.0208084043,
0.0036436599,
0.0129966075,
-0.0413683616,
0.0053390372,
-0.0918820053,
-0.0755256116,
-0.0474653244,
-0.0224104989,
0.0034480796,
0.0649294332,
-0.0364139453,
-0.0007432561,
0.0209744312,
-0.014816273,
-0.0385159925,
-0.0124941282,
-0.0426029451,
0.0101591544,
0.0110182837,
0.0535999872,
-0.0682731718,
-0.0351665542,
0.0463452227,
0.0587497577,
-0.0278432351,
-0.0728037655,
0.0131099056,
0.0256435536,
-0.0030503899,
0.0076355198,
0.0106113739,
-0.0409428217,
-0.0002732097,
0.0085765235,
-0.043219924,
0.0045400024,
-0.043487329,
0.0212711561,
-0.0159807708,
0.0154214595,
0.026521014,
0.0124034975,
0.047468435,
-0.0535791814,
-0.0278100241,
-0.0266936291,
0.0004536311,
0.0354157872,
-0.0248558037,
-0.0501809567,
0.042068582,
0.0152609423,
0.0298405625,
-0.0417379104,
0.0171750076,
0.037954133,
0.0172858816,
-0.0242482722,
-0.0152190235,
-0.0148004843,
-0.0251527727,
-0.0041591865,
0.0080708545,
0.0833987966,
-0.0259238128,
0.0439003929,
0.0008867149,
-0.0282887258,
0.0133415228,
0.0140349828,
-0.0335518345,
-0.0227296054,
0.0493863411,
0.0102113867,
-0.0152301136,
0.0091896206,
0.042919226,
-0.0269016977,
0.0754050463,
-0.0428927541,
-0.0471007042,
-0.050315585,
-0.0364016928,
0.0245618708,
0.0109674772,
-0.0349384286,
-0.0203068051,
-0.0388107933,
-0.0909660384,
0.0166677181,
0.0084933266,
-0.0933634788,
0.0240817647,
-0.004447401,
0.003131086,
0.0252617374,
-0.0436970033,
0.0006105655,
0.0382064469,
-0.0548751056,
-0.0449838601,
-0.0153325265,
-5.512783815e-33,
-0.0548277684,
-0.029277442,
0.0324825682,
-0.0413420871,
-0.0638654232,
-0.0064459182,
-0.0299765822,
-0.0193365421,
-0.060891021,
0.0020771462,
-0.0015764948,
0.0041320436,
-0.0140004922,
-0.0158146918,
0.002337944,
-0.0411345325,
0.0111180879,
0.0355387703,
-0.0087193893,
0.0036594819,
0.0025622563,
0.0148218283,
0.0386530869,
-0.0534789898,
0.0582870841,
-0.0229300726,
0.0122538479,
0.0163427982,
0.0120334281,
0.0615589656,
-0.0108515006,
0.0115210135,
0.014132428,
-0.0187062752,
0.0035148151,
0.0282687135,
-0.110063307,
-0.0148211131,
0.0396084636,
-0.0002061741,
0.0072653331,
-0.0366883539,
0.0472146608,
-0.0015575622,
0.0434297249,
-0.0349877104,
0.0425425097,
-0.0188404173,
-0.0142692421,
-0.0421922542,
-0.0488033667,
0.0058389688,
-0.0190934241,
0.031287089,
0.0009051222,
-0.0034410746,
0.0619262606,
0.0136951087,
-0.1094243377,
0.0046452773,
-0.028454164,
-0.0117492788,
0.0620209835,
0.0134720253,
0.01373454,
0.0497146249,
-0.0325756446,
-0.0423709601,
-0.0337667838,
0.0047868085,
-0.0037419747,
0.0165099334,
0.0024208706,
0.0436929315,
-0.0308493227,
0.013565436,
-0.0664825812,
0.0305117741,
-0.0225446075,
0.0185115486,
-0.000256033,
0.0093375547,
-0.0246015023,
-0.0067500952,
-0.0625763163,
0.0026056934,
-0.0399030447,
-0.0278270729,
0.0352598093,
-0.035907466,
-0.0017046069,
0.0292548835,
0.0119763138,
-0.0469139963,
-0.0383146629,
0.0310504287,
0.01711905,
-0.0016700667,
-0.0149088772,
0.0157970544,
-0.0632098764,
-0.0095182834,
0.0076265186,
-0.0359103754,
0.0120472023,
0.0449287184,
0.0392875485,
-0.0503955632,
-0.0399839543,
0.0006954829,
-0.0158448648,
-0.0495660342,
-0.0813788995,
-0.0520670228,
0.0513448827,
0.0441453271,
0.0330487601,
0.0081018135,
0.1028515771,
0.0119404662,
-0.0059627891,
0.0176473502,
0.0276316553,
0.033179339,
-0.0051120478,
0.012551493,
-0.0253434833,
-0.0010707742,
-0.0120820096,
-0.0434513316,
-0.0084695006,
0.0089523131,
2.453e-7,
0.0554918945,
0.0032869284,
0.0346341208,
0.0324903615,
-0.0122755701,
0.0318069346,
0.0208464433,
0.0339529403,
0.0072582141,
-0.0350459851,
0.0364194363,
0.0218717661,
0.0172101744,
0.0256481972,
-0.0557753406,
-0.0182875413,
-0.0279746335,
0.0059765591,
-0.0611010678,
0.0102438629,
0.04155799,
0.0704637393,
0.0555740707,
0.0117673902,
-0.0222666077,
-0.0266053211,
0.0511984713,
0.0059764264,
-0.0172257498,
0.0287859514,
0.021323638,
0.0154293934,
-0.0164176598,
0.0506803468,
-0.0041634548,
0.0390907079,
0.0308306385,
0.0347333886,
-0.0190761704,
0.0095881466,
-0.0694542602,
-0.0516550019,
0.0187059734,
-0.0735724196,
0.0581719466,
-0.0418604836,
-0.0049537672,
0.0180170201,
-0.0512581281,
0.0165268108,
-0.0087177185,
0.0255400576,
0.0020983422,
0.0057646856,
0.0211227965,
0.0240083672,
-0.0281941202,
-0.0675109327,
0.0044227266,
0.0397025123,
-0.0403432287,
-0.0279315859,
0.0001499205,
-0.0365654342,
0.033243835,
-0.0633033291,
0.0038516598,
2.120033449e-34,
-0.0390400216,
0.0285072979,
0.0231377799,
0.0206803642,
0.0597867519,
-0.049964413,
-0.0193924587,
0.0071189827,
0.0063796211,
-0.0784824789,
0.0152877746
] |
http://arxiv.org/abs/2402.01758v1 | http://arxiv.org/pdf/2402.01758v1 | Aalap: AI Assistant for Legal & Paralegal Functions in India | 2024-01-30T12:39:58 | 2024-01-30T12:39:58 | Aman Tiwari, Prathamesh Kalamkar, Atreyo Banerjee, Saurabh Karn, Varun Hemachandran, Smita Gupta | cs.CY | cs.CY, cs.AI, cs.CL | Using proprietary Large Language Models on legal tasks poses challenges due
to data privacy issues, domain data heterogeneity, domain knowledge
sophistication, and domain objectives uniqueness. We created Aalalp, a
fine-tuned Mistral 7B model on instructions data related to specific Indian
legal tasks. The performance of Aalap is better than gpt-3.5-turbo in 31\% of
our test data and obtains an equivalent score in 34\% of the test data as
evaluated by GPT4. Training Aalap mainly focuses on teaching legal reasoning
rather than legal recall. Aalap is definitely helpful for the day-to-day
activities of lawyers, judges, or anyone working in legal systems. | Aalap: AI Assistant for Legal & Paralegal Functions in India: Using proprietary Large Language Models on legal tasks poses challenges due
to data privacy issues, domain data heterogeneity, domain knowledge
sophistication, and domain objectives uniqueness. We created Aalalp, a
fine-tuned Mistral 7B model on instructions data related to specific Indian
legal tasks. The performance of Aalap is better than gpt-3.5-turbo in 31\% of
our test data and obtains an equivalent score in 34\% of the test data as
evaluated by GPT4. Training Aalap mainly focuses on teaching legal reasoning
rather than legal recall. Aalap is definitely helpful for the day-to-day
activities of lawyers, judges, or anyone working in legal systems. | [
0.0478453524,
0.0333530717,
-0.0296843629,
-0.0091520809,
-0.0714632422,
0.0206060316,
0.0281854663,
-0.0164151769,
0.0530262142,
-0.0533352196,
-0.0047793672,
0.0144423163,
-0.0434696972,
0.0389681123,
-0.0063581192,
-0.0614171177,
0.0275706295,
-0.0021787446,
0.0747269243,
0.0020987359,
0.0197199043,
0.014387398,
-0.0324489251,
0.024269212,
-0.0236443672,
-0.0585182458,
-0.0209349263,
-0.0123550678,
-0.0097973682,
0.0000025609,
-0.0146900993,
0.0563478023,
0.031706091,
-0.0088811899,
0.0000022433,
-0.059805125,
-0.0096118301,
0.0134866573,
-0.0404208712,
0.0019471439,
-0.0002162137,
0.0031918865,
0.008529718,
0.0190943293,
-0.0568913594,
0.0737028569,
0.0093496041,
0.0306709297,
0.0498370826,
0.1093435809,
0.0024106307,
-0.060376972,
0.0437292904,
-0.0218216348,
-0.0154847093,
-0.0733271912,
-0.0028924956,
0.0041786744,
0.0460436419,
-0.0222160276,
0.028641345,
0.0042343922,
0.0352505147,
-0.0467719994,
-0.0254039336,
-0.0251457654,
-0.0164698772,
0.0089844512,
-0.0062933369,
0.0248903278,
0.0613739565,
-0.0357191153,
-0.0084308032,
0.020697128,
0.0670509264,
0.0187072121,
-0.0148763526,
0.0941878855,
-0.0283519849,
-0.0066913324,
0.067109555,
0.0365203805,
-0.0276616383,
0.0016930226,
0.0350470357,
0.0073611457,
-0.021354422,
-0.0249041785,
-0.0033696715,
-0.0020901198,
0.0208697766,
-0.0510821119,
0.0451468676,
0.0741961151,
-0.0212266296,
0.0228770319,
-0.0441306867,
-0.0326879732,
0.0497084111,
-0.0446780249,
0.0847122744,
0.0194966346,
0.0013629029,
0.0440399759,
-0.0400519967,
0.0648636892,
0.0144641055,
0.0373333171,
0.0171271432,
-0.0329914801,
-0.0438472629,
0.0026312554,
-0.0277676247,
0.0753171742,
-0.1119917482,
0.0293454044,
-0.0278522,
0.0176443141,
-0.050207898,
0.034109354,
0.0415605232,
0.0002921797,
0.0056666238,
-0.0485799089,
-0.0498779528,
-0.082232371,
-0.0610297322,
0.0476055481,
0.0267138835,
-0.0337774009,
0.0074604489,
-0.0100757582,
-0.0119603481,
0.0094266739,
-0.0108343055,
0.1062040403,
-0.0095466105,
-0.0257896744,
-0.0878973678,
-0.0506783277,
0.0294839926,
-0.0914093554,
-0.03116557,
-0.0531761348,
0.0368193425,
-0.0219295751,
-0.0466839373,
-0.0418561734,
0.0030555758,
0.058233507,
0.0181902535,
0.0567454696,
-0.0237183012,
0.0111153796,
0.011178405,
0.0469119325,
-0.0077327453,
0.0068965107,
-0.0319889002,
0.0282766409,
0.0147521002,
-0.0162873492,
0.0688598081,
-0.0205420218,
-0.0037707642,
-0.0124423113,
0.0157360807,
0.0451950021,
-0.0542259365,
-0.0068680285,
-0.0027265402,
0.0361885652,
-0.0086455029,
0.0031184682,
0.0230332278,
0.0508474,
0.0068674581,
0.0845523775,
-0.0023063261,
-0.0199080873,
0.0473634675,
-0.0146151092,
-0.0791023225,
0.0443898216,
-0.0303472672,
-0.003872449,
-0.0769331679,
0.0202006046,
-0.0292982105,
-0.0538379326,
0.0036235664,
-0.0298510939,
-0.0225959998,
0.004484877,
-0.0073051569,
0.0188279022,
-0.0290568564,
-0.0109290732,
0.0183101408,
-0.0461516269,
-0.058071956,
-0.0484534614,
0.0085232817,
0.031925302,
0.0127379037,
-0.0418943018,
-0.0185094085,
-0.0406368934,
-0.0007874548,
0.0207826793,
0.0235608369,
-0.0084370831,
-0.0139918905,
-0.0313620977,
-0.0304172542,
0.0017064494,
0.0276668128,
0.0221778043,
0.0128591685,
0.0243163519,
0.0176321082,
-0.0233520307,
-0.0015885136,
0.0202797782,
-0.0313573517,
-0.0589320622,
-0.0084801838,
0.0196318626,
0.0493729822,
0.059750516,
0.0268179774,
-0.0231719334,
0.0503305756,
0.0073392801,
0.006469795,
-0.0041606361,
-0.0177548341,
0.0253087468,
0.0106402263,
-0.0328311548,
0.0101142749,
0.0118546374,
-0.0158548672,
-0.0300158747,
-0.0294951741,
-0.0315382369,
0.0331934653,
-0.0376701094,
0.018083673,
0.0053699156,
-0.0526427142,
-0.0064359326,
0.0511107855,
-0.0014058837,
-0.0689075217,
-0.0089874892,
0.0079161851,
0.0426724628,
-0.0153617663,
0.0029627399,
-0.0441946387,
-0.0161092039,
0.0435456112,
0.044612892,
0.0135291191,
-0.0395334661,
-0.0153538035,
-0.0757778436,
-0.0079289144,
-0.0220823567,
-0.0169610418,
-0.0142298294,
-0.0133109316,
-0.0251183007,
-0.0368850119,
-0.0331922174,
-0.0214327145,
0.0048859473,
-0.0104765166,
-0.0038381862,
-0.0065969597,
-0.0228498932,
0.0030940706,
-0.0094805537,
-0.023015501,
-0.0094253868,
0.0337587185,
-0.0434212498,
-0.0341992751,
0.0544768721,
0.0092170676,
0.0154608544,
-0.0236416683,
-0.0186147615,
-0.0208496992,
0.0138615314,
-0.0419825427,
0.0607991889,
-0.0120417345,
0.0259077325,
0.0090035042,
0.0119940657,
-0.0508885644,
0.0091874711,
0.0243479125,
-0.0020949487,
0.0004268105,
0.0073117437,
-0.0003161253,
-0.0039012567,
-0.0083396537,
0.0268389322,
-0.014867139,
0.0176364668,
0.0095202774,
0.0083314143,
-0.009081834,
-0.0057696896,
-0.0141389053,
0.0517022014,
0.0052182036,
-0.0766310021,
0.0627166927,
0.0241334252,
0.0710368901,
-0.0758517087,
-0.0287225544,
-0.0235893298,
0.0782687739,
-0.0247428957,
0.0545627289,
-0.0368234925,
-0.0504641049,
0.007537574,
-0.0100146132,
0.045588322,
-0.0541537292,
0.0555073358,
-0.0527945049,
0.0274480693,
-0.0216753185,
-0.0465140566,
-0.0587123074,
0.0041162344,
0.011906404,
0.0531095639,
-0.0201715138,
0.0082239192,
-0.0517847613,
-0.0233674832,
-0.0122796735,
0.0194872394,
-0.0067381389,
-0.0202173572,
-0.0229083989,
-0.0391128249,
0.0216924492,
-0.0265654586,
-0.0021242455,
0.0158607922,
0.0177973062,
0.005149018,
0.0465380885,
-0.0347125903,
0.0116038732,
-0.0067833583,
0.0041482365,
-0.0570604652,
-0.0137345884,
0.0424375087,
0.0196646098,
-0.0401495099,
-0.0019792859,
0.074668847,
0.0160917193,
-0.0581791513,
-0.033704713,
-0.035877943,
-0.0359392576,
0.0278238486,
0.0492030382,
-0.1047939211,
0.0113846157,
0.0088830432,
0.0159815792,
-0.040404167,
-0.0468766242,
0.0043837922,
-0.0897508934,
0.0717425272,
0.0109831011,
0.0188002903,
-0.0685899407,
-0.0114376787,
-0.0133111868,
0.0015117254,
0.0412684679,
-0.0155463247,
-0.0438654982,
-0.029233709,
0.0484395027,
-0.00982494,
0.0027236145,
0.0223820638,
-0.0567119755,
0.0650875047,
0.0827097744,
-0.00073285,
-0.0222530123,
-0.0411975496,
-0.0383337885,
0.0279248096,
0.0221024733,
0.0153959803,
0.0167610999,
0.0072546625,
0.01786853,
-0.0248866063,
0.0580202714,
-0.0163664278,
0.0111925565,
0.0809716657,
0.0563676804,
-0.0376731567,
0.0127540911,
-0.0282243639,
-0.0661296472,
0.0052793534,
0.0239775293,
0.0038854964,
0.0200915616,
-0.0287782978,
-0.0324377269,
-0.0040773894,
0.0111057721,
0.015654454,
0.0112682246,
-0.0464806817,
-0.0124678509,
-0.0159129538,
-0.0609745868,
0.025397623,
0.0472558402,
0.0378223248,
0.0266449656,
0.0120697534,
-0.0060765571,
-0.044878345,
-0.0170749519,
-0.0613744967,
0.016850628,
-0.0402952321,
-0.0080100205,
-0.0136138098,
0.0259334333,
-0.004349913,
-0.0099945925,
0.0527903102,
0.0019341557,
0.0140800523,
0.0590882823,
0.0252885185,
0.0079402905,
-0.0315351784,
-0.0479325317,
0.0317282006,
-0.013127014,
-0.0144445393,
-0.0183815341,
0.0411947891,
-0.0300232619,
-0.0110833338,
-0.0017494544,
-0.0400062688,
0.019202251,
0.0207045563,
-0.0077237659,
-0.0111609306,
0.0136179421,
0.065764986,
0.0724192858,
-0.0071691545,
0.0260769241,
0.0136980172,
-0.016064411,
0.0023051384,
-0.0303999521,
0.0105380192,
-0.0154210143,
-0.0023761713,
0.032590311,
0.0069903047,
0.0328816809,
0.0129979383,
-0.0018235381,
-0.0210812371,
-0.0745090991,
0.0561024621,
0.0063592624,
0.0515405275,
-0.0226933621,
-0.0171962306,
-0.0221016202,
0.0056181434,
0.0099018496,
0.059252914,
-0.0206965189,
0.0938592628,
-0.0080294283,
0.0250378568,
0.0034479008,
0.0438220687,
0.0227183886,
-0.0661420524,
0.0122293383,
0.079026103,
0.0651064292,
0.0098049594,
-0.0160832461,
-0.0556230322,
0.0270318259,
0.0100418795,
0.0250032935,
0.0121281752,
-0.0131910881,
0.01893791,
0.0288977399,
-0.0073307687,
0.0069617759,
0.0383485593,
-0.0435870625,
0.0344775021,
-0.0365539454,
-0.0234732591,
-0.0394505449,
-0.0390690453,
0.0006765252,
0.0454347432,
0.0035058893,
-0.0350527689,
-0.0661912709,
-0.0221532173,
-0.0363962017,
-0.015043579,
-0.0128871128,
0.0017783686,
0.0260133669,
0.0623495653,
0.031803526,
-0.0567778759,
-0.0644212291,
-0.0125960791,
-0.0168436579,
0.0045260638,
0.0002249278,
-5.696934189e-33,
-0.0416594967,
-0.0258287583,
0.0483582765,
-0.0132365664,
0.0187554527,
-0.0184956044,
-0.0009438492,
-0.010091559,
-0.0545745753,
0.0457622781,
-0.0401900746,
0.0190201737,
0.0213369913,
-0.0255726893,
0.0295695104,
-0.0130595593,
0.0068378299,
0.0137956524,
0.0187118556,
0.0141883399,
-0.0190843381,
-0.0028020514,
0.0272391532,
-0.1179741174,
0.0219113864,
0.0233336054,
-0.0415978506,
-0.0169375781,
-0.1100472137,
0.0495482944,
-0.0073492411,
0.0434795842,
0.0126918536,
-0.0082796104,
0.0043664942,
0.0239801742,
-0.0856491402,
-0.01182165,
0.0053139739,
0.0070321346,
-0.0116473874,
-0.0465719737,
0.02331792,
-0.0238691773,
-0.0336357765,
0.0048914496,
0.0242242906,
-0.0068618041,
-0.0215638075,
-0.035567455,
-0.0689221919,
0.0262849666,
0.0008893452,
-0.0051305685,
0.0235431828,
0.0030377728,
-0.0089356406,
-0.0253382139,
-0.0863754228,
-0.0184323695,
0.0251003783,
-0.0356742218,
0.0206064731,
-0.0604218133,
-0.0281142741,
0.0205030181,
0.0444276147,
-0.0001536965,
-0.0224636048,
0.0048790169,
-0.0065944828,
0.1140239313,
0.0322984718,
-0.0200730041,
0.0097689088,
-0.0521988161,
-0.0323250778,
0.0111098243,
0.058679238,
0.0592009984,
0.0132905087,
-0.0150543926,
-0.1086621881,
0.0033554195,
0.0006221812,
-0.0321284086,
-0.0207073633,
-0.0160753615,
0.0352368839,
-0.0013322647,
0.0481279232,
0.0014627886,
-0.0272821728,
-0.0319553167,
0.0401867479,
0.0016513033,
0.0343838334,
-0.0065951906,
-0.0186700635,
0.0301747043,
-0.0666930303,
-0.0321413465,
-0.0020720966,
-0.0167173296,
0.0243124999,
0.0048191017,
0.0454853773,
-0.0128811579,
-0.0472211726,
-0.0058123018,
0.004042373,
0.0599185117,
0.0627351478,
0.0111885956,
0.0047724536,
0.0106065366,
0.0150906583,
0.0527684763,
0.0263464339,
-0.0261804983,
-0.0181606747,
0.0162353646,
0.0077155628,
-0.0091270851,
0.008973767,
0.0125809032,
-0.0207740571,
0.0373074971,
0.0062739747,
-0.0112813031,
0.0212396551,
0.0385503322,
2.929e-7,
0.0283723697,
0.0456022508,
-0.0203025416,
-0.0065999664,
-0.047728505,
0.0417975672,
-0.0001193183,
0.0350513756,
0.0054453122,
-0.0090085305,
-0.0057341149,
-0.0213673357,
0.0414676368,
0.0548825078,
-0.0514129363,
-0.046036765,
-0.0365994051,
-0.0418063328,
-0.0723757073,
0.065107435,
0.0113307647,
0.0667111725,
0.0567678884,
-0.0252736378,
0.0119601469,
0.0030035686,
0.0237057656,
0.0287357029,
-0.0185852312,
-0.0161051042,
0.0781400502,
-0.0194919836,
-0.0092241792,
0.0510742031,
0.008592735,
0.0631631091,
-0.0353692882,
0.0860221758,
-0.0293583926,
0.041504804,
0.0384546295,
-0.0590358302,
0.0260387454,
-0.0008561947,
0.0453477316,
0.0337907486,
-0.0315765217,
0.034632802,
-0.0543426126,
-0.0445514396,
0.014540012,
0.0221417323,
0.0155878542,
-0.029016437,
0.0127199376,
-0.0565195046,
0.0003866028,
-0.0625019595,
0.0209487695,
0.0107811503,
0.0079161609,
0.0145112053,
0.0066869096,
-0.0066927127,
0.0463485382,
-0.0786051601,
0.0244573876,
2.544396782e-34,
0.0078858081,
-0.0126189794,
0.0201099869,
0.0610686243,
0.0495111831,
0.0478462726,
-0.004710386,
0.0194599535,
-0.0016885011,
-0.0321968235,
0.0090878336
] |
http://arxiv.org/abs/2401.16895v2 | http://arxiv.org/pdf/2401.16895v2 | Cross-Lingual Transfer from Related Languages: Treating Low-Resource Maltese as Multilingual Code-Switching | 2024-01-30T11:04:36 | 2024-02-03T07:26:47 | Kurt Micallef, Nizar Habash, Claudia Borg, Fadhl Eryani, Houda Bouamor | cs.CL | cs.CL | Although multilingual language models exhibit impressive cross-lingual
transfer capabilities on unseen languages, the performance on downstream tasks
is impacted when there is a script disparity with the languages used in the
multilingual model's pre-training data. Using transliteration offers a
straightforward yet effective means to align the script of a resource-rich
language with a target language, thereby enhancing cross-lingual transfer
capabilities. However, for mixed languages, this approach is suboptimal, since
only a subset of the language benefits from the cross-lingual transfer while
the remainder is impeded. In this work, we focus on Maltese, a Semitic
language, with substantial influences from Arabic, Italian, and English, and
notably written in Latin script. We present a novel dataset annotated with
word-level etymology. We use this dataset to train a classifier that enables us
to make informed decisions regarding the appropriate processing of each token
in the Maltese language. We contrast indiscriminate transliteration or
translation to mixing processing pipelines that only transliterate words of
Arabic origin, thereby resulting in text with a mixture of scripts. We
fine-tune the processed data on four downstream tasks and show that conditional
transliteration based on word etymology yields the best results, surpassing
fine-tuning with raw Maltese or Maltese processed with non-selective pipelines. | Cross-Lingual Transfer from Related Languages: Treating Low-Resource Maltese as Multilingual Code-Switching: Although multilingual language models exhibit impressive cross-lingual
transfer capabilities on unseen languages, the performance on downstream tasks
is impacted when there is a script disparity with the languages used in the
multilingual model's pre-training data. Using transliteration offers a
straightforward yet effective means to align the script of a resource-rich
language with a target language, thereby enhancing cross-lingual transfer
capabilities. However, for mixed languages, this approach is suboptimal, since
only a subset of the language benefits from the cross-lingual transfer while
the remainder is impeded. In this work, we focus on Maltese, a Semitic
language, with substantial influences from Arabic, Italian, and English, and
notably written in Latin script. We present a novel dataset annotated with
word-level etymology. We use this dataset to train a classifier that enables us
to make informed decisions regarding the appropriate processing of each token
in the Maltese language. We contrast indiscriminate transliteration or
translation to mixing processing pipelines that only transliterate words of
Arabic origin, thereby resulting in text with a mixture of scripts. We
fine-tune the processed data on four downstream tasks and show that conditional
transliteration based on word etymology yields the best results, surpassing
fine-tuning with raw Maltese or Maltese processed with non-selective pipelines. | [
-0.0167632345,
0.0643786043,
-0.0179834645,
0.0250380244,
-0.0221589487,
0.058726117,
0.0148409922,
0.0035782447,
-0.0345972925,
-0.008027303,
-0.0455458723,
-0.0366860703,
0.0490618795,
0.015141177,
0.054376591,
0.0007272126,
0.0360041112,
-0.0614362657,
-0.038673494,
0.0187430866,
-0.002762204,
0.0438328236,
0.0320569053,
-0.0030707605,
-0.0132547431,
0.0175206475,
0.0287234019,
0.0129514225,
0.0262461845,
0.0535373576,
-0.0269440208,
0.013103433,
-0.0490100831,
0.0649953857,
0.0000018061,
-0.0478953011,
-0.0456588902,
-0.026817346,
-0.0300837494,
-0.0098334169,
0.0516417958,
-0.0884676278,
-0.035021551,
-0.0101089031,
0.0135366153,
0.0068240291,
0.090859808,
-0.000217048,
0.0007833117,
0.0673391968,
-0.0288236178,
-0.0334539898,
0.0241129901,
-0.0342630483,
0.0018585471,
-0.0272876583,
0.0103753833,
-0.018394256,
-0.0419769511,
-0.0101412861,
-0.0437729172,
0.0600846186,
-0.0255786981,
0.0124153122,
-0.0512354635,
0.0013447972,
-0.0014724645,
-0.0154939843,
0.0185624603,
0.0239336789,
-0.1072514877,
-0.0362959392,
-0.0314796455,
-0.0134968702,
-0.0016261637,
0.0027067917,
-0.017613776,
0.0050375056,
-0.0395814404,
0.0198458154,
0.0623160191,
0.0086427722,
0.043198783,
-0.0236997195,
-0.0735850111,
0.0283040721,
0.0525096357,
-0.0231409371,
0.0276132971,
0.0471627899,
0.0649810284,
-0.0383492112,
0.0084864274,
-0.009489798,
0.0150825735,
0.0041835289,
-0.038892813,
-0.0320903845,
-0.0113192433,
-0.0177734476,
0.0676077828,
0.0772168338,
-0.0373674259,
0.022465229,
-0.0687977299,
0.0192587655,
-0.0219993964,
-0.0049582212,
-0.0258566532,
-0.0253336504,
-0.0375870578,
0.0190907791,
-0.0481937081,
-0.0104826232,
-0.0433517955,
0.0532248579,
-0.0364631191,
0.0166250728,
0.022089744,
-0.012701069,
0.015936492,
-0.0897359848,
-0.0444127172,
0.0011222258,
0.0350252204,
0.0437822342,
-0.0107935024,
0.0023006035,
-0.0464590676,
-0.0753420368,
-0.006697116,
-0.0062656654,
0.0171391368,
-0.0165717714,
0.0171258822,
0.1153498143,
-0.0304169245,
-0.0429213308,
-0.0591365807,
0.017198965,
-0.022066094,
-0.034123145,
0.0105602583,
-0.0322280116,
0.0641482845,
-0.0059152218,
0.0360279046,
-0.0292778388,
-0.0330451243,
-0.0015838607,
-0.0384785384,
0.0714651793,
0.0186632015,
0.0010872709,
0.0136653539,
-0.0362693332,
-0.0370268188,
0.0952323452,
0.0492163822,
-0.0272846669,
0.0236540455,
-0.0340506956,
-0.0217219889,
0.0772644579,
0.0154829789,
0.0185844358,
-0.0212943926,
-0.0336774066,
0.0407718755,
0.0159853864,
0.057884194,
0.0209611617,
-0.0115019931,
-0.0072220857,
0.0343636721,
0.0252135545,
0.0425161384,
0.0282267202,
0.0602174141,
0.0061351378,
-0.0254617613,
0.0063082394,
0.0193322115,
0.0130784884,
0.0223272536,
-0.0033499058,
0.0121253775,
0.0882207453,
-0.0287515782,
0.0126604587,
0.0308446623,
-0.0356571227,
0.0383369029,
-0.0062602069,
0.0659827963,
0.0301945005,
-0.0483682379,
0.0047924998,
0.0435113832,
-0.044810269,
0.021123467,
-0.0236789286,
0.0027617875,
0.0724062994,
0.0154057117,
0.0141525017,
-0.0101609491,
-0.015526481,
0.0112061854,
0.0004947534,
0.0118935239,
-0.0113230282,
0.081531778,
-0.0462922864,
0.0168609302,
0.0007553597,
-0.0108802477,
0.0573040359,
-0.0356240496,
0.0631249622,
-0.0042721666,
-0.0252712034,
-0.0259728432,
0.0512082651,
-0.0704575554,
0.0255937334,
0.0024566429,
0.0106819328,
0.0091794441,
-0.005183361,
-0.0328747928,
-0.0077421092,
-0.0390075259,
-0.0608043745,
0.0125870891,
0.0144166779,
-0.0255339947,
-0.000460556,
0.0443999991,
-0.0062234518,
0.0122198239,
0.0018544977,
-0.0026373437,
-0.0340240188,
0.0419643931,
-0.0268148798,
0.0204596259,
-0.0189899094,
0.0403913856,
-0.0071126348,
0.0252492279,
-0.0127255712,
0.0234181471,
0.0407331996,
0.0192593168,
-0.0026104585,
-0.0476733968,
0.0395621248,
0.017658323,
0.0182208847,
0.0461140648,
-0.0191462077,
0.0053442945,
-0.0017886893,
-0.0059385253,
0.0115005365,
-0.0441204272,
-0.0519711114,
0.0057725054,
-0.012290475,
0.0475191437,
-0.0068950527,
0.0547651201,
-0.0244157352,
-0.0128282364,
-0.0116978483,
0.001294461,
-0.0161069743,
0.023807358,
0.0094607985,
-0.0085327309,
0.0309029538,
-0.0342855006,
-0.0125971781,
0.0289024245,
0.0732868016,
0.0106123555,
-0.0066882367,
-0.0403376818,
-0.0069493107,
-0.0210652947,
-0.0048969183,
-0.0021467318,
0.0048855585,
0.0206033271,
0.0069465954,
0.0264640041,
-0.0034158975,
-0.031099882,
-0.0683590397,
0.0439928435,
-0.0552853011,
-0.0355953015,
-0.0193236675,
0.0478376821,
-0.031877473,
-0.0908862203,
-0.0486952849,
-0.02573069,
0.0791634396,
-0.0389466994,
-0.0059041181,
-0.0963662863,
-0.0178807694,
-0.0213128515,
-0.0162488334,
0.0102284811,
0.002910156,
0.0194145162,
0.0729928911,
-0.0388347916,
-0.0147140911,
0.0407991484,
-0.0530566797,
-0.023612693,
-0.0268794,
0.0149973528,
-0.0337740332,
0.0173004698,
-0.0431240126,
0.0226278566,
-0.0489456095,
-0.0316601545,
-0.056611456,
0.0152022168,
0.0060404083,
-0.0219044164,
0.0431365408,
-0.0375888832,
0.0232147854,
-0.0249168407,
-0.015176178,
-0.011938761,
0.0203722566,
-0.0208576936,
-0.0065542394,
0.0235831458,
-0.0305254888,
0.0153936753,
0.0157430284,
-0.0070697465,
0.0835170448,
0.02995052,
-0.004168516,
-0.0126572121,
-0.0073668542,
-0.0304343794,
-0.0256843101,
-0.030519899,
-0.0168071277,
0.0242971722,
0.0001137699,
-0.0176400524,
-0.0245949123,
-0.0223984122,
0.0104185566,
0.0305189323,
-0.0048749843,
-0.0347305909,
-0.0576835275,
0.0326086469,
0.0067149908,
-0.0102301482,
0.0500966907,
-0.002413491,
-0.0595990941,
-0.0007844183,
-0.0295441803,
0.0197723918,
-0.0008205453,
0.0299037807,
-0.0707688406,
-0.0068052649,
0.0244232081,
0.0435944013,
0.0322393589,
-0.0176305696,
0.0071371486,
-0.0839592591,
0.0095750215,
0.0250122119,
-0.0920031592,
-0.0884599462,
-0.0262434557,
-0.0081342105,
-0.0346316583,
-0.0230033547,
-0.0157575998,
-0.0147491684,
-0.0060968758,
0.011400125,
0.0383992903,
-0.0288458671,
-0.010355629,
-0.041121006,
0.0287186392,
0.017654039,
0.0150118284,
-0.0288062189,
0.0194034949,
-0.0001546992,
-0.0488766506,
-0.0386601239,
0.0007357612,
0.0635742322,
-0.006642865,
-0.0277173668,
0.0378668606,
-0.0353709459,
-0.0481539182,
-0.0346225724,
0.0004832871,
0.037329454,
0.0048416094,
-0.0064251116,
0.0474982895,
0.0011309028,
0.0110258469,
0.0293564126,
0.0190787334,
-0.0513909385,
-0.0253743157,
-0.0308579765,
-0.0054036621,
-0.1014730483,
-0.0106361425,
0.0220207814,
-0.0107167941,
-0.0854324773,
-0.0924569964,
-0.0344985016,
0.0241475031,
0.0309160892,
0.0207393188,
-0.0030909432,
-0.0038995259,
0.0047133248,
-0.0244110916,
-0.0167468283,
-0.0318597779,
0.0245580468,
-0.0230064541,
-0.0297836401,
0.0134554571,
-0.0603532232,
0.0179322306,
0.0079316553,
-0.0941038877,
-0.049165491,
0.00648442,
0.0090768551,
0.0158276912,
0.0432940945,
-0.0099307895,
-0.0093384478,
0.012293024,
-0.0502698347,
-0.0325656906,
0.0362921916,
0.0627975538,
0.0048945886,
-0.0203938801,
0.0332175605,
-0.0229892321,
-0.0017839537,
0.0240993779,
-0.0016614031,
-0.0515102446,
-0.0292440467,
0.0007801026,
-0.0366053209,
-0.0343151614,
-0.0537392944,
-0.0426609591,
-0.0592004843,
0.0239813104,
-0.01655812,
0.0375674628,
-0.0281764008,
-0.026710948,
0.0652908012,
0.0043744082,
0.0272575393,
-0.023608014,
-0.0065392242,
-0.0265287422,
-0.0186953824,
0.0044105379,
0.0192731731,
0.0368903913,
0.0285191387,
0.0450113341,
0.0120415781,
0.0161694866,
-0.0179910194,
0.0296171475,
-0.0730910152,
-0.0199121535,
0.0115352729,
0.0176905375,
0.0246606264,
0.0283738691,
-0.0057505383,
-0.0440832339,
-0.0009632825,
-0.008673843,
0.0460906737,
-0.0283092763,
0.1059761718,
0.0032862106,
-0.0064423201,
-0.0200789962,
0.0092010377,
-0.0224588159,
0.0157266427,
-0.0380129404,
-0.0113324802,
-0.0124064609,
0.0004904068,
0.0775134116,
0.0073319576,
0.0211621542,
-0.025502298,
0.0181114804,
-0.0012870844,
-0.0214995611,
0.0109174056,
-0.0116891321,
-0.0102528725,
-0.0043286961,
-0.0472164303,
-0.0861981213,
0.0155197736,
-0.0167762917,
-0.003279147,
-0.0085566128,
0.0079063624,
0.0776382908,
0.002294285,
0.0616159551,
-0.0707003176,
0.0229838602,
0.0788659528,
-0.0238218885,
-0.0146841686,
-5.303649411e-33,
-0.0028801644,
0.0557050668,
0.0075597228,
0.0176278371,
0.0040803631,
-0.0053498205,
-0.0223649275,
-0.0278253052,
-0.0304413084,
-0.0214400291,
-0.019756915,
0.0051717209,
0.0300623886,
0.0096025122,
0.0746077001,
-0.0126385316,
0.043436911,
-0.0045862519,
-0.0233790409,
0.0300724786,
0.0509983264,
0.0589545779,
0.0733348504,
-0.0244697016,
0.0118998233,
0.0040311138,
0.0189776979,
-0.0398957394,
0.0328407623,
0.0512581468,
0.0052190227,
0.0107741272,
0.0293595344,
-0.0263356604,
-0.0005010778,
0.0319548994,
-0.0580892116,
0.0006011712,
0.019798981,
0.0513302162,
-0.0044539678,
-0.0651876926,
-0.0053049549,
-0.0075152414,
0.0114980787,
-0.0200666618,
0.0037978101,
-0.0345572904,
-0.00726353,
-0.0145686865,
-0.1164486036,
0.0414836258,
-0.0083868336,
-0.0031723231,
0.084239088,
0.0666214302,
0.0346571952,
0.0178681314,
-0.0388894677,
0.0227222685,
-0.0660313666,
0.0381496772,
0.0561791547,
0.0297551788,
-0.0769205689,
0.0000043416,
0.0299036894,
-0.000770656,
0.0704425573,
0.0482190661,
0.0480496213,
-0.0283752494,
0.0384847336,
0.0258458797,
0.0193541683,
-0.0452142507,
-0.0420235917,
0.0004273187,
0.0261788648,
-0.0161803477,
-0.0065365899,
0.0057075331,
-0.0050031897,
-0.0133436536,
-0.011441499,
-0.0551270284,
-0.0015189742,
0.0040094545,
0.0115594668,
-0.0140621401,
0.0001630162,
0.0419912115,
0.07802926,
-0.0055097109,
-0.0003810752,
0.0075945002,
-0.0076861428,
0.0417779572,
-0.0111172181,
0.0665528774,
-0.0330665894,
-0.0579770543,
0.0356052704,
-0.0261270013,
0.0366678163,
-0.0017390494,
0.065070495,
0.031966351,
-0.0790417865,
0.0123238638,
-0.0274711791,
0.0311097633,
0.007361229,
0.0897845626,
0.0077630272,
0.0083402749,
0.0220698491,
-0.009342039,
0.0477194749,
-0.0250807144,
0.0016552383,
-0.0084524238,
-0.0046891486,
0.0195396766,
-0.0360735618,
-0.0103221321,
-0.0239369273,
0.0233149286,
0.0344676375,
-0.0344964303,
-0.0354103334,
0.0287189316,
2.634e-7,
0.0889316648,
-0.0154079869,
0.0185781363,
-0.0082348837,
0.00610048,
-0.0006480829,
-0.0300050061,
0.0270627085,
0.011863119,
0.0400401466,
0.0115915053,
0.0025043788,
0.1040979922,
0.0409899428,
-0.0461685322,
-0.0830097497,
0.0344141647,
0.0334718078,
-0.0423034653,
0.014319323,
0.0044695493,
0.1267003864,
-0.0263244174,
-0.0148023516,
-0.0327241495,
-0.0390134752,
0.0369889587,
-0.0323225558,
0.0011186956,
-0.002657006,
-0.0010025362,
0.0078456178,
-0.0356114469,
0.0310879704,
-0.0468675829,
0.0247405153,
0.0613366961,
0.039218802,
-0.0265978705,
0.0008489178,
0.0148855662,
-0.0241112486,
-0.0430290289,
-0.0436240584,
0.0555644371,
0.0222255755,
0.0070785838,
-0.0174372531,
-0.0266294181,
0.0559427999,
0.0175392982,
-0.0259690527,
-0.0401962735,
0.0376268812,
0.0974353403,
0.0359410197,
-0.0421131514,
-0.0335088372,
0.0032893745,
0.0228104312,
-0.0055008559,
-0.0209249742,
-0.0169455819,
-0.0452487208,
0.0414053239,
-0.0136828087,
0.0161777344,
2.258739257e-34,
0.0181815606,
-0.005820476,
0.0293581393,
0.0127347326,
0.0092244344,
-0.0463275872,
-0.0239217766,
0.0534008332,
-0.057820715,
-0.0168404374,
-0.0019679139
] |
http://arxiv.org/abs/2401.16862v1 | http://arxiv.org/pdf/2401.16862v1 | State Value Generation with Prompt Learning and Self-Training for Low-Resource Dialogue State Tracking | 2024-01-30T10:05:03 | 2024-01-30T10:05:03 | Ming Gu, Yan Yang, Chengcai Chen, Zhou Yu | cs.CL | cs.CL | Recently, low-resource dialogue state tracking (DST) has received increasing
attention. First obtaining state values then based on values to generate slot
types has made great progress in this task. However, obtaining state values is
still an under-studied problem. Existing extraction-based approaches cannot
capture values that require the understanding of context and are not
generalizable either. To address these issues, we propose a novel State VAlue
Generation based framework (SVAG), decomposing DST into state value generation
and domain slot generation. Specifically, we propose to generate state values
and use self-training to further improve state value generation. Moreover, we
design an estimator aiming at detecting incomplete generation and incorrect
generation for pseudo-labeled data selection during self-training. Experimental
results on the MultiWOZ 2.1 dataset show that our method which has only less
than 1 billion parameters achieves state-of-the-art performance under the data
ratio settings of 5%, 10%, and 25% when limited to models under 100 billion
parameters. Compared to models with more than 100 billion parameters, SVAG
still reaches competitive results. | State Value Generation with Prompt Learning and Self-Training for Low-Resource Dialogue State Tracking: Recently, low-resource dialogue state tracking (DST) has received increasing
attention. First obtaining state values then based on values to generate slot
types has made great progress in this task. However, obtaining state values is
still an under-studied problem. Existing extraction-based approaches cannot
capture values that require the understanding of context and are not
generalizable either. To address these issues, we propose a novel State VAlue
Generation based framework (SVAG), decomposing DST into state value generation
and domain slot generation. Specifically, we propose to generate state values
and use self-training to further improve state value generation. Moreover, we
design an estimator aiming at detecting incomplete generation and incorrect
generation for pseudo-labeled data selection during self-training. Experimental
results on the MultiWOZ 2.1 dataset show that our method which has only less
than 1 billion parameters achieves state-of-the-art performance under the data
ratio settings of 5%, 10%, and 25% when limited to models under 100 billion
parameters. Compared to models with more than 100 billion parameters, SVAG
still reaches competitive results. | [
-0.0076343063,
0.0771003142,
0.0133007532,
0.0090949293,
0.00499412,
-0.0044042249,
0.0154497409,
-0.0218592342,
-0.010209715,
-0.047877647,
-0.0244367551,
-0.0475632101,
-0.0073005762,
0.0662638322,
0.0029661732,
-0.0592554584,
0.0522070751,
-0.0437932499,
-0.0085440734,
0.0103026768,
0.002017217,
-0.0262575708,
-0.0591373853,
0.0406246968,
-0.0295014568,
-0.0192190614,
-0.0631992891,
0.0004263455,
0.0051353807,
-0.0490808152,
-0.017251173,
0.0258973967,
0.0291015785,
-0.0028280835,
0.0000021041,
-0.0399037115,
-0.0383093432,
-0.025439661,
-0.0709958524,
-0.0279428121,
-0.0308375433,
0.0215827916,
-0.0349223725,
0.05105244,
-0.0350487418,
-0.0266117528,
0.0174150355,
0.0169590712,
0.0683307499,
0.0695862025,
-0.0061154244,
0.0225387216,
-0.0009912745,
-0.0314156562,
0.007627619,
-0.0775990933,
0.016416572,
-0.0039611012,
-0.0332780145,
-0.0126413675,
-0.0036642053,
0.0225330722,
0.0186352469,
-0.0425910279,
-0.0395794511,
0.0106146932,
-0.0058978959,
-0.0132198343,
-0.0488850996,
0.0801175609,
0.0535519272,
0.0081926696,
0.0430198647,
-0.0212579016,
0.0121540576,
-0.0203385353,
-0.0096369768,
-0.0138145555,
-0.0392607749,
0.0162511654,
-0.0035099743,
0.0699458867,
-0.0140557233,
-0.0194509756,
-0.0247730073,
0.0428078361,
-0.003231504,
-0.0422412492,
-0.0321832672,
-0.0278923884,
0.0314017534,
-0.076764226,
0.0052123214,
0.021421887,
-0.040797621,
0.0317727923,
-0.0701262206,
-0.0242546704,
-0.0135607356,
-0.0278225131,
0.0090299947,
0.0493380725,
-0.0423859619,
0.0307072923,
0.0248540509,
0.0374524295,
0.0015014788,
-0.0234642271,
-0.0215763375,
0.0124576725,
-0.0406139009,
0.0080052009,
-0.0764141306,
0.0277659316,
-0.028871201,
-0.030434832,
-0.0170745198,
0.0073447912,
-0.0084149484,
0.0371421054,
-0.0323824994,
0.0101864422,
-0.0460763015,
0.0084427577,
-0.0295572504,
0.0374086499,
-0.0000530078,
-0.0090833912,
0.027047988,
-0.0382724479,
-0.017014185,
0.0158068947,
0.0467862897,
-0.0073968042,
-0.0270295199,
0.0439075567,
-0.0141685652,
-0.0273947325,
-0.0421187542,
-0.0365204923,
-0.0168184526,
-0.0596598908,
0.0063015823,
-0.034338627,
0.0272099543,
-0.0030638147,
-0.0103715612,
-0.0348365009,
0.0097571407,
0.0085441954,
0.0396830067,
0.0250560828,
-0.0454342104,
-0.0083383387,
-0.0006765441,
-0.0144510418,
0.0440982059,
0.0620834827,
0.0544063412,
0.0247197449,
0.0300911255,
-0.0060875947,
0.0240592267,
0.0231072307,
0.062418513,
0.0244641248,
-0.0302518848,
0.0279278848,
0.024437746,
-0.0019963928,
0.0212500449,
0.0156961195,
-0.0386968479,
-0.006031977,
0.0528234504,
0.010118789,
0.0326775946,
-0.0051271082,
0.0463977717,
0.0370595977,
0.0259107612,
0.0060300548,
0.0102471625,
0.0182539839,
0.013816338,
-0.0040815561,
0.0637650341,
0.0133399284,
-0.0355285406,
-0.0602376126,
0.0140664643,
-0.0186303239,
-0.0104731685,
0.0123375403,
0.0642100796,
0.0291650891,
-0.0513486564,
0.0322806835,
0.0114861261,
-0.0374732092,
-0.0333037302,
-0.0418961309,
0.0084078452,
0.0529585108,
0.0219528805,
-0.0319270305,
-0.0524002723,
-0.0519846529,
0.0162399001,
0.0393689275,
0.0370637365,
-0.0310083199,
0.0647234321,
-0.0438410491,
-0.0397131741,
0.0097955065,
0.0051599736,
0.0262912232,
-0.015458826,
-0.0162739176,
0.0239923056,
0.0233859289,
0.0347989164,
0.0067547173,
0.0191978402,
-0.0508149192,
0.0373785347,
0.0059450045,
0.065723598,
0.0438207984,
0.0025263049,
-0.0057849241,
0.0030592128,
0.0035005619,
0.0073335962,
-0.0006577802,
-0.0980925336,
-0.0734931156,
0.0063993027,
0.0647775605,
-0.0049295183,
-0.0077265152,
-0.0255707055,
-0.096592389,
-0.0070436052,
-0.0266221575,
0.0148632107,
-0.0563495047,
-0.0041721505,
0.0708598718,
0.0101727387,
-0.0093172304,
0.0403507315,
0.0077375122,
-0.0289630983,
0.0363176651,
0.0595163293,
0.047692243,
-0.0705511495,
-0.0005012165,
0.0044798427,
-0.066257678,
0.0166644827,
0.0024959017,
0.0010225674,
-0.0454239026,
0.0322878323,
-0.0565554984,
0.0123799313,
-0.0289397072,
0.0092680315,
0.0127174808,
0.012540305,
-0.0198176,
-0.0233030338,
-0.0414977856,
-0.0409764349,
0.0340595841,
-0.031963475,
-0.0009981859,
0.0741480142,
-0.0151177179,
-0.0458673611,
0.0410315841,
-0.0141120264,
0.0054824892,
-0.0408795439,
-0.0217265468,
-0.0723276138,
0.047968898,
-0.0080348738,
0.0598065555,
-0.0238590762,
-0.0401104689,
0.0318598561,
0.0128636127,
-0.0040604961,
-0.0049018143,
0.0200839192,
0.0372702181,
0.0050106798,
-0.0325363539,
-0.0855486766,
0.0534572862,
-0.0333562829,
0.0348474532,
-0.0237563699,
0.0473162234,
0.0281953495,
0.0481101573,
0.03587934,
0.0424121283,
-0.0520221591,
-0.0005277927,
0.0075443676,
-0.0346123502,
-0.0136722839,
0.0207264666,
-0.0671188906,
0.0496260561,
0.0052092467,
-0.051892262,
0.0432638228,
-0.0442817733,
-0.0276319925,
-0.0309841,
-0.0191046838,
0.0043939776,
0.0323046483,
-0.0290210936,
0.0125942752,
0.0652314126,
-0.0354845412,
0.0022995267,
-0.0208187215,
0.0233447272,
-0.0351579636,
0.0387225412,
-0.0151914163,
0.0474322289,
-0.0046584313,
-0.0661156923,
-0.0780047327,
-0.0325155631,
0.0070288298,
-0.0154591985,
0.0017135657,
-0.0156943984,
0.0326310955,
0.0201755092,
0.0245672986,
0.0771718547,
-0.0059551182,
-0.0031204433,
-0.0202380698,
0.0552099496,
0.06525819,
0.0072803074,
-0.0230601951,
0.0004090234,
0.069539316,
0.0496879369,
0.0799285993,
0.0148431789,
-0.0100400271,
-0.0109597035,
0.0109754121,
-0.0151212402,
0.0064442819,
0.0269271452,
0.0569813102,
-0.0436817892,
-0.0129880589,
0.1037354022,
0.0263336115,
-0.0359380357,
-0.0140615655,
-0.0184525046,
-0.0044472921,
0.0089812949,
0.0242935643,
-0.0737667084,
-0.0169652086,
0.0019513608,
0.0675021484,
-0.0425499044,
-0.0271051005,
-0.0090409173,
-0.1030418202,
0.0297458787,
-0.0037650072,
-0.0515483022,
-0.0156114511,
-0.0178668927,
0.0329676643,
0.0457920842,
-0.0056289877,
0.0274063405,
0.0041339337,
-0.0234761033,
0.0548178069,
-0.0226957109,
-0.029480394,
0.002897141,
0.0058438326,
0.0888807848,
0.0725323781,
0.0312310569,
0.0081552062,
0.0068769571,
-0.0066390787,
-0.044962097,
-0.0140482346,
-0.0181276146,
0.0462098606,
-0.0099022416,
-0.0202642754,
-0.0003059681,
0.049459707,
0.0231251065,
-0.0120138992,
-0.0015203051,
0.005665632,
-0.0743170455,
0.0475642793,
0.0176335983,
-0.0016524622,
0.0566463657,
0.0345819555,
0.0256017931,
-0.0488460511,
0.0017762564,
-0.0105821239,
0.0049984143,
-0.0442141332,
0.0133905904,
-0.0223394763,
0.0141407065,
-0.0083104856,
-0.0793409124,
-0.0356288925,
0.0412858538,
0.1005777717,
-0.0250922609,
0.0198453348,
0.0257072523,
0.0089245038,
0.0024071524,
-0.0034929139,
-0.0440756008,
-0.0102693969,
-0.0520472117,
-0.0729270279,
0.0500509478,
-0.0205631852,
-0.0184863005,
0.0378357768,
-0.0301650763,
-0.059636049,
0.0078403018,
0.0797514021,
0.0278710015,
0.0452106372,
0.0031976898,
0.0284240283,
-0.0051251547,
0.0332532898,
-0.0018744228,
-0.0035685587,
0.0619256385,
0.0726654455,
-0.0396722853,
0.0248207971,
-0.0270806011,
0.0147846332,
-0.007128288,
0.0347917192,
-0.0612905324,
-0.0665048286,
0.0100464365,
-0.0227559935,
0.0605902411,
-0.0054008723,
-0.008588789,
-0.0254902616,
0.0400949046,
-0.0165473241,
-0.0034578335,
-0.08052136,
0.0466121882,
-0.0233878102,
-0.0060295295,
0.0204458646,
-0.056052655,
-0.0112945773,
0.0195200592,
-0.0029876628,
-0.0213351604,
-0.0120452307,
0.0040464331,
0.0197987743,
-0.0210732613,
0.0074001895,
0.0344804674,
0.0851549879,
0.0275746286,
0.0012508965,
0.1092881262,
0.0246229414,
-0.0197623968,
0.0144863473,
0.0255405027,
-0.0335564204,
-0.050462015,
-0.0603953153,
0.0180942379,
0.0323771313,
0.0205794796,
0.0290535204,
0.002138183,
0.0185538922,
-0.0006376206,
0.0735026598,
0.0182868056,
-0.01450013,
0.0137008233,
0.0169086512,
-0.0257089287,
0.017219631,
0.0677170455,
0.050513614,
0.0392823443,
0.0088614151,
0.0343373492,
-0.0731938705,
0.0011412019,
-0.0158853959,
0.0451812595,
-0.0232151859,
-0.0049783387,
-0.0603609346,
-0.0423150063,
0.0014239495,
0.0115669277,
-0.0156242298,
-0.0143163111,
0.0594685897,
0.0051039192,
-0.0331677906,
0.0085545136,
-0.0483203121,
-0.0543765724,
0.0136419218,
-0.0102967275,
0.0019301631,
-6.123684923e-33,
-0.0618393794,
-0.0377679765,
0.0352738202,
-0.0042621912,
-0.0257217903,
-0.022850018,
-0.0415951908,
0.0165873095,
0.022953175,
-0.0154952826,
-0.0250459295,
-0.0083397711,
0.0350518599,
0.0189137198,
-0.0222239532,
-0.0368380919,
0.0709748864,
-0.0235293917,
-0.0357426293,
0.0379061364,
0.068509236,
0.0243969802,
0.039458897,
-0.0998705104,
-0.0155596854,
0.0106118219,
0.0018724821,
0.0115237618,
0.0005917579,
0.0318414681,
0.0008240517,
0.0091737444,
-0.0032729364,
-0.0687605515,
-0.0129416957,
0.0273847338,
-0.0342406742,
-0.0402708277,
0.0025150068,
-0.0116426442,
0.0074963975,
0.0138795162,
-0.0214395467,
0.0090995776,
0.0123527786,
0.0500809327,
0.0079877228,
0.0062277755,
-0.0266162157,
-0.0304400772,
-0.0473081879,
0.0321234986,
0.0124638248,
-0.0843101516,
0.0085608652,
0.0779756829,
0.0197486114,
-0.0774555728,
-0.0508812703,
0.0004351974,
-0.004723228,
0.0151415737,
0.0406176522,
0.0349263549,
0.029101925,
0.0865718722,
0.0563951358,
0.0221165195,
-0.0033037534,
-0.0210768729,
-0.018601818,
0.0405664146,
-0.0044464488,
0.0039852066,
-0.0101339016,
-0.0330622345,
-0.0294891838,
0.0118993167,
0.0209493842,
0.0067688939,
-0.0138317375,
-0.0105124945,
-0.0595200099,
-0.0070074797,
-0.0096217878,
-0.011614359,
0.0223403536,
-0.0312068928,
0.0480896272,
0.0512011088,
-0.0509550162,
0.0263811033,
0.0941848233,
-0.0113844173,
-0.0064727608,
0.0007255768,
-0.0073278924,
0.0125823952,
-0.0316577516,
-0.0332104936,
-0.050729122,
-0.0524693653,
0.0437125713,
0.0188616179,
0.0347570814,
-0.0020047878,
-0.0287557002,
0.0248001497,
-0.0534191206,
0.0173780434,
-0.0536158755,
0.0641790926,
-0.036506746,
0.02926125,
0.0433567204,
-0.0233547594,
-0.0181969665,
0.0793649331,
-0.0015834769,
-0.0515076779,
-0.0337859988,
-0.0159249566,
0.0042464463,
0.0194027089,
-0.0141156828,
-0.012403531,
-0.0409523398,
-0.0107823648,
0.0222586431,
-0.0192277208,
0.0300514903,
0.0195448976,
2.892e-7,
-0.0337626822,
0.0170114525,
0.0408191346,
-0.0133881327,
0.0260295421,
0.0854870528,
-0.028325092,
-0.0074095735,
0.0355862156,
-0.042448543,
0.0040956265,
-0.0127347726,
0.0529267937,
0.0320665054,
-0.123715885,
-0.0357176773,
-0.016780166,
-0.0238013417,
-0.0404657982,
0.0156920534,
0.0423148386,
0.0533565991,
0.0098051876,
0.0106436126,
0.0689337552,
0.0038423173,
0.0399610586,
0.0198915768,
-0.0079581123,
0.0187378563,
0.0266463086,
0.0092628803,
-0.0369882546,
-0.0214531291,
-0.0105160205,
0.0528926551,
-0.0367613807,
0.0930861309,
0.0283330753,
-0.0651557893,
-0.0127544729,
0.0245732274,
0.0076866667,
-0.0113419993,
0.0444767289,
-0.0608270951,
-0.0037609749,
-0.0264960062,
0.0075284587,
0.0033889287,
-0.0210354459,
0.0261093397,
0.0016290218,
0.0051617478,
-0.0113289608,
-0.0022846649,
0.0004762949,
-0.0698005706,
-0.0085288612,
-0.0416822433,
-0.0536348708,
-0.045535069,
-0.0307813715,
-0.0168851893,
0.0031018425,
-0.0280143023,
-0.0400754102,
2.727071359e-34,
0.0107308673,
-0.0455922298,
-0.0006901875,
0.0439235717,
-0.0022834435,
-0.0141093656,
0.0024481032,
0.0090085706,
0.0315537788,
-0.0238712318,
-0.020956276
] |
http://arxiv.org/abs/2402.03362v1 | http://arxiv.org/pdf/2402.03362v1 | NanoNER: Named Entity Recognition for nanobiology using experts' knowledge and distant supervision | 2024-01-30T09:10:53 | 2024-01-30T09:10:53 | Martin Lentschat, Cyril Labbé, Ran Cheng | cs.IR | cs.IR, cs.AI, cs.CL | Here we present the training and evaluation of NanoNER, a Named Entity
Recognition (NER) model for Nanobiology. NER consists in the identification of
specific entities in spans of unstructured texts and is often a primary task in
Natural Language Processing (NLP) and Information Extraction. The aim of our
model is to recognise entities previously identified by domain experts as
constituting the essential knowledge of the domain. Relying on ontologies,
which provide us with a domain vocabulary and taxonomy, we implemented an
iterative process enabling experts to determine the entities relevant to the
domain at hand. We then delve into the potential of distant supervision
learning in NER, supporting how this method can increase the quantity of
annotated data with minimal additional manpower. On our full corpus of 728
full-text nanobiology articles, containing more than 120k entity occurrences,
NanoNER obtained a F1-score of 0.98 on the recognition of previously known
entities. Our model also demonstrated its ability to discover new entities in
the text, with precision scores ranging from 0.77 to 0.81. Ablation experiments
further confirmed this and allowed us to assess the dependency of our approach
on the external resources. It highlighted the dependency of the approach to the
resource, while also confirming its ability to rediscover up to 30% of the
ablated terms. This paper details the methodology employed, experimental
design, and key findings, providing valuable insights and directions for future
related researches on NER in specialized domain. Furthermore, since our
approach require minimal manpower , we believe that it can be generalized to
other specialized fields. | NanoNER: Named Entity Recognition for nanobiology using experts' knowledge and distant supervision: Here we present the training and evaluation of NanoNER, a Named Entity
Recognition (NER) model for Nanobiology. NER consists in the identification of
specific entities in spans of unstructured texts and is often a primary task in
Natural Language Processing (NLP) and Information Extraction. The aim of our
model is to recognise entities previously identified by domain experts as
constituting the essential knowledge of the domain. Relying on ontologies,
which provide us with a domain vocabulary and taxonomy, we implemented an
iterative process enabling experts to determine the entities relevant to the
domain at hand. We then delve into the potential of distant supervision
learning in NER, supporting how this method can increase the quantity of
annotated data with minimal additional manpower. On our full corpus of 728
full-text nanobiology articles, containing more than 120k entity occurrences,
NanoNER obtained a F1-score of 0.98 on the recognition of previously known
entities. Our model also demonstrated its ability to discover new entities in
the text, with precision scores ranging from 0.77 to 0.81. Ablation experiments
further confirmed this and allowed us to assess the dependency of our approach
on the external resources. It highlighted the dependency of the approach to the
resource, while also confirming its ability to rediscover up to 30% of the
ablated terms. This paper details the methodology employed, experimental
design, and key findings, providing valuable insights and directions for future
related researches on NER in specialized domain. Furthermore, since our
approach require minimal manpower , we believe that it can be generalized to
other specialized fields. | [
0.0896399468,
0.0121548008,
0.0071545821,
-0.0388113149,
-0.0284996722,
0.0221150257,
0.0484723449,
0.0355378501,
0.0571090877,
-0.0174603108,
-0.0137888668,
0.0058245929,
-0.0431935079,
0.0453833044,
-0.0005803158,
0.0123962583,
0.0407889374,
0.0196236838,
0.0247198418,
0.0177526306,
-0.005758021,
0.0082429247,
0.0057243519,
-0.0014484698,
0.0118005527,
0.0273871217,
0.007909406,
-0.0167726986,
0.0826130509,
-0.0637038425,
0.036582157,
0.0150440102,
0.0006333556,
0.0555161014,
0.0000020853,
-0.0482241251,
-0.101183176,
0.0273094419,
-0.0314148813,
-0.0405553021,
0.0499733761,
-0.0341222137,
0.0119488789,
0.0038778207,
-0.0062063364,
-0.0150511721,
0.0470573232,
0.070668906,
0.0056413058,
0.0555291325,
-0.0099791316,
-0.0221464857,
0.0630623847,
0.0265666172,
0.0143509936,
-0.0025153365,
0.0612900592,
0.0170887075,
0.0393614732,
0.0292194448,
0.0184696168,
0.0541305691,
-0.012570166,
0.0062101735,
0.0612399727,
0.0411211438,
0.0324078538,
0.0003852056,
-0.0093057007,
0.0612999015,
0.0038701766,
-0.038483955,
0.0315039791,
-0.0166880246,
-0.0454144999,
0.0468803048,
-0.0091381343,
-0.0329443254,
-0.0518891215,
0.024764061,
0.0313817263,
0.0280423071,
0.0175269805,
0.0126950443,
0.018085178,
0.0513155311,
-0.0232344158,
-0.057357423,
0.0083895652,
-0.0093451412,
-0.0087847142,
-0.0136482082,
0.0793886185,
0.044967711,
0.034569677,
0.0156597234,
-0.0003171458,
-0.0176540185,
0.0252883863,
-0.0405548885,
-0.0337148607,
-0.0205629189,
-0.0136430059,
0.0631301552,
-0.0452366285,
-0.0177191179,
0.0002169208,
0.01887317,
0.0132794725,
-0.0393089317,
-0.1120496467,
-0.0294774082,
-0.006229707,
0.0926016793,
-0.0082361205,
-0.040549729,
-0.0210801139,
-0.0497641824,
0.0232549813,
-0.0117908325,
-0.0053715659,
0.0152695803,
-0.0087365489,
0.0206681099,
-0.0255950186,
0.0337487198,
-0.0028749635,
-0.0313479193,
-0.0378012732,
-0.0343596749,
-0.0236793403,
-0.0374649651,
0.0294738524,
-0.0333965048,
0.023417227,
0.1067751348,
0.0211970303,
-0.0980399922,
-0.0211139284,
-0.0322003178,
-0.0057992865,
-0.0241979621,
-0.0047861878,
-0.0135979392,
0.0520801432,
0.0054300553,
0.0109974723,
-0.0847973228,
0.0151791573,
-0.0330251902,
-0.0677127987,
0.0718615502,
-0.0870310441,
-0.0621247217,
0.050440304,
0.0291325469,
0.0270559732,
-0.0300305057,
0.0085598035,
0.0197706427,
0.0004364326,
-0.0125960447,
0.0080602542,
-0.0359418727,
0.0006846382,
-0.0672434717,
0.011261493,
-0.0034584489,
-0.0088929432,
0.0281876344,
-0.0090119941,
0.0153291496,
0.0110764503,
0.0241263378,
0.0136006903,
0.0734694004,
-0.0077060629,
-0.0109023396,
-0.0033210954,
0.0261654593,
0.0041072499,
0.0036216753,
-0.0129750175,
0.0348044299,
-0.0109939454,
0.0163412895,
0.0290036723,
0.0137478244,
0.0206035301,
-0.014702105,
-0.0090027265,
0.0248532444,
-0.0306153446,
-0.0386423245,
0.069496423,
-0.0031178936,
-0.0434648767,
0.0377532169,
0.0869924799,
-0.0238926597,
0.0361455008,
-0.0256418027,
0.0416265838,
0.0766565129,
0.0539712161,
-0.0795283541,
-0.0764289126,
-0.0095879501,
-0.0549028479,
0.0646671802,
-0.0286847148,
-0.0286455024,
-0.0080481162,
-0.004501191,
-0.0262922887,
0.0039500548,
0.0256442651,
0.0471443236,
0.0355269276,
0.0270751361,
0.0279648546,
-0.0299241934,
-0.0397783257,
0.0511179566,
-0.0197227225,
-0.0461919345,
-0.0363855995,
0.0068120603,
0.0090280613,
0.0485474207,
-0.0139716025,
-0.0114175761,
0.0510493182,
-0.0418103859,
-0.030381443,
-0.0514569879,
0.0298283603,
-0.014517785,
0.0031516044,
-0.0264114141,
-0.0084238192,
0.0017010571,
-0.017244447,
-0.0353741311,
-0.02357601,
0.0231731739,
0.0237939917,
-0.0152128628,
-0.0079969568,
-0.0144886365,
-0.0030899753,
-0.0259376615,
0.0571549386,
0.0131331617,
0.0322338864,
-0.0198803451,
-0.0373064093,
-0.0146255484,
-0.0344051681,
0.0022762087,
-0.0454480276,
0.0116538592,
-0.018019123,
0.0291560926,
-0.030873267,
-0.0667876825,
-0.009068924,
-0.0371220075,
0.0011695637,
0.0222466271,
0.0626486093,
0.0105874036,
0.0005499816,
-0.0032431493,
-0.0260868594,
0.0027948548,
-0.0321304947,
-0.0246931165,
-0.0356839113,
0.0303287152,
0.0112363342,
-0.0156592578,
-0.027775988,
0.0131995808,
-0.0615976118,
0.02380931,
0.0469380282,
-0.0731647164,
-0.0277622547,
0.0248812567,
-0.0359734744,
0.0001290262,
0.0089563793,
0.0057997685,
-0.009469172,
0.0002474769,
-0.0329866,
0.0400526486,
-0.0245660208,
0.0319639333,
0.0262390692,
-0.0025692289,
-0.0362396613,
0.0068610935,
-0.0424794964,
0.0538081415,
0.0146815274,
0.0029315967,
0.0279014762,
0.0837856531,
-0.0204269774,
0.0051283548,
-0.0452935509,
0.0707607791,
-0.0358674861,
-0.0398475267,
0.022329472,
0.0312623493,
0.0110263648,
0.0244227387,
0.0008606222,
-0.0245556049,
0.094600454,
-0.0109409858,
0.0201442689,
-0.0256925877,
-0.028018659,
0.0213215947,
0.0277577136,
0.0314869843,
0.0211472977,
0.0615855493,
-0.0576347299,
0.045635093,
0.0366211049,
-0.0063713044,
-0.0076957471,
-0.058483161,
-0.0156643353,
0.0241938196,
0.0311596487,
0.0131699378,
-0.0144413356,
0.0276443809,
-0.0272864718,
0.0438285917,
0.0142097911,
-0.0274610631,
-0.0333537869,
0.0846448541,
0.0407917686,
0.009432856,
0.0159775298,
-0.0093818484,
-0.0440227389,
0.0528091937,
0.0293013286,
-0.0587398224,
0.0287233517,
0.0091737173,
0.0049087312,
0.0521115847,
-0.0022655684,
-0.0364781395,
-0.0009953012,
-0.0126539916,
-0.0344910435,
-0.0587156489,
0.0354021601,
-0.0295630079,
0.0257159993,
-0.0412906334,
-0.013891656,
0.0961881578,
0.0329366475,
0.016766144,
-0.012781309,
0.0338826701,
-0.0818918794,
-0.0116621898,
0.0280964077,
-0.0755149797,
-0.0251599215,
-0.0132076936,
-0.0273553599,
-0.0272040647,
-0.0956610516,
0.0111375405,
-0.0410801619,
0.0180379637,
-0.0055686878,
-0.0981958508,
-0.014333644,
-0.0805017054,
-0.0189267416,
-0.0281487741,
-0.0140619138,
-0.0044781971,
-0.0165815372,
0.0142200841,
0.0314632095,
0.0299317446,
-0.0565785356,
0.0220308471,
-0.0246321745,
0.012149781,
0.0513891019,
0.013749999,
-0.0069383658,
0.0410989262,
-0.0275294986,
-0.0194867477,
0.0208190326,
-0.0287111495,
0.0763066262,
-0.0281324126,
0.0010310982,
-0.0145596536,
0.0086638769,
-0.0197987445,
-0.017486807,
0.0162538588,
0.0815871134,
-0.0009869606,
0.0060459976,
-0.0036968011,
-0.0260293875,
0.0049131126,
0.0274265297,
0.0117755644,
0.0034632466,
-0.054296352,
0.0210965462,
0.0356737897,
-0.0266776904,
-0.0161687322,
-0.0057470701,
-0.0714227632,
-0.0107849548,
-0.0594313219,
-0.034981288,
0.0410030335,
0.0324037671,
0.0321115814,
0.0183235239,
0.0388613604,
0.0121378675,
-0.0733749345,
0.0240905285,
-0.0179900136,
-0.0208910834,
-0.0434873849,
-0.0479467846,
0.0664741844,
0.0222037733,
-0.0364898145,
0.0235583689,
-0.0077285799,
-0.0760032088,
-0.0191186722,
-0.0037996212,
0.0195560437,
0.0074385256,
0.0032996293,
-0.0383977666,
0.0949861705,
0.0418192148,
-0.0565771312,
-0.0414892174,
0.0089701507,
0.0023159555,
-0.0287533738,
0.0112905959,
0.0050156694,
0.0444666781,
0.0313180387,
0.0229003727,
-0.033153262,
-0.0238808151,
0.0468128212,
0.0076336213,
0.0687466115,
0.0112697957,
0.0030676364,
-0.0530818626,
0.0138203502,
0.0590473786,
0.0118861431,
0.0013968125,
-0.0305883773,
0.0206615254,
0.0130837644,
-0.0019811513,
0.0405011252,
-0.0089479955,
0.0180738997,
-0.0413300358,
-0.0209276713,
-0.0082863923,
0.0075920136,
0.044660911,
-0.0186658278,
-0.0171617325,
0.0020550613,
0.0140767777,
0.0220196173,
-0.0255310908,
0.0126328552,
0.0122260163,
0.0432987772,
-0.0544122197,
0.0128589347,
-0.0719830245,
0.0278710779,
-0.051209271,
0.0788083076,
0.0405350253,
-0.0176795144,
0.0275192726,
-0.024193367,
-0.0229726266,
-0.0010387747,
0.0312094297,
-0.0300584771,
0.0124876015,
0.0274450853,
0.026964847,
-0.013072703,
-0.0115695354,
0.0487697013,
0.0312686674,
0.0479420163,
0.0140759749,
-0.0345860161,
-0.0345545374,
-0.0106879408,
0.0022469785,
-0.0088500697,
-0.0220512059,
0.0236346126,
-0.0045291856,
-0.0198434629,
0.0488541462,
-0.0135281626,
-0.0221046284,
-0.0002216474,
0.0148183824,
0.0050041126,
0.0253827758,
-0.0463069715,
-0.0338759311,
0.0203385595,
0.0273168515,
-0.023134239,
0.0172980111,
-6.494819347e-33,
-0.0032339378,
-0.0558977611,
0.0076288595,
0.056665834,
0.0164275523,
0.0178576894,
-0.0433802307,
-0.0467303544,
0.0416321196,
0.0020394423,
-0.0122919939,
-0.017465055,
0.0225108154,
-0.0213619582,
0.0159398615,
-0.0073540113,
0.0429518558,
0.0503546745,
0.0008217191,
0.0290379655,
0.0568282381,
0.088472344,
0.085735403,
-0.0537212864,
0.0350659564,
0.0328980349,
0.0546546951,
-0.0274341423,
-0.0095895892,
0.0467677042,
-0.0419727638,
-0.0065490305,
-0.0025451861,
-0.0799736455,
0.0022612,
-0.0174568873,
-0.1365077198,
0.0137222521,
-0.0405348763,
0.0039616465,
-0.0191347897,
-0.047534883,
0.0090160696,
-0.0242389385,
-0.0192625467,
0.0196047053,
0.0678224415,
-0.0065984051,
-0.0237056687,
-0.0401756354,
-0.0732302368,
0.033078935,
0.0243671685,
0.0451194197,
0.0754374042,
-0.0428967699,
0.0243800152,
0.002225609,
-0.0472788922,
0.0208855048,
0.0730381683,
0.0269865766,
0.0198444631,
-0.010117108,
-0.0194047503,
0.0262567662,
0.0289234035,
-0.0323314071,
0.0471810289,
0.0018964878,
-0.028514836,
-0.0050564851,
-0.0173934586,
0.0157587286,
0.0627671033,
-0.0072992346,
-0.0729377121,
0.010439666,
-0.0369123556,
0.0460120738,
-0.0136967907,
0.0073681707,
-0.0171246566,
-0.0042135366,
0.0102773029,
-0.0068259169,
-0.0263561029,
-0.0624000691,
-0.0166931618,
-0.0182287004,
-0.0737949535,
0.0204388406,
-0.0162181426,
-0.0447270125,
-0.0329229608,
-0.030399818,
0.0631878078,
0.0136981802,
-0.0256694034,
-0.0086018592,
-0.0893142745,
-0.0599220544,
-0.0042239414,
-0.0492827892,
0.034580946,
-0.0319419391,
0.0158229545,
-0.0186055526,
-0.0488918796,
0.0174919851,
-0.0194025375,
-0.0001684299,
-0.000941026,
0.0283831619,
-0.012585368,
0.0177867953,
0.0041871145,
0.0162542351,
0.0672389939,
-0.0124219218,
0.0073509528,
0.0593386292,
0.0175176244,
-0.0030176099,
-0.0489472188,
-0.0234915838,
-0.0088323709,
0.0156293865,
-0.0355678126,
-0.0459266677,
0.0062039476,
0.0103919683,
2.952e-7,
0.0352181159,
0.0556070283,
0.0421067066,
-0.068344906,
-0.0240586381,
-0.0017145177,
-0.0119282436,
0.0245076958,
-0.0607519075,
-0.0197355263,
0.0068609882,
-0.0304920264,
-0.0129772061,
0.028353136,
-0.0920220166,
-0.0331955403,
0.0026121482,
-0.0521310754,
-0.0493216291,
0.0712868869,
-0.0191514213,
0.034668155,
0.01048104,
-0.0262920801,
-0.0186950397,
-0.0335644372,
0.027194418,
0.0108963959,
0.0160785802,
0.0543008968,
-0.0062130825,
0.0411007665,
-0.0393714048,
0.0502673052,
-0.0000872976,
0.0027781492,
-0.0077552893,
0.0280512106,
-0.0121442052,
0.0023260696,
0.0093360376,
0.006143345,
0.0327605009,
-0.0064161019,
0.0322329886,
-0.0605850033,
-0.0352654904,
0.0526946709,
0.0142433764,
-0.0189362597,
-0.0270021576,
0.0087180873,
-0.005957067,
0.0010827812,
-0.0320721604,
0.0139652798,
-0.0100011993,
-0.0643260777,
0.043951191,
0.0493807569,
0.0003431603,
-0.0512182601,
0.0076321145,
-0.028747933,
0.0053749662,
-0.0705815554,
0.0059296712,
2.387123673e-34,
-0.004895499,
0.0024192119,
-0.0392028801,
0.0356534906,
-0.0174266566,
-0.0029332032,
-0.0522180907,
0.0093367165,
-0.0212749578,
-0.0576664619,
-0.0097809369
] |
http://arxiv.org/abs/2401.16818v2 | http://arxiv.org/pdf/2401.16818v2 | H2O-Danube-1.8B Technical Report | 2024-01-30T08:45:08 | 2024-04-15T17:58:01 | Philipp Singer, Pascal Pfeiffer, Yauhen Babakhin, Maximilian Jeblick, Nischay Dhankhar, Gabor Fodor, Sri Satish Ambati | cs.CL | cs.CL, cs.LG | We present H2O-Danube, a series of small 1.8B language models consisting of
H2O-Danube-1.8B, trained on 1T tokens, and the incremental improved
H2O-Danube2-1.8B trained on an additional 2T tokens. Our models exhibit highly
competitive metrics across a multitude of benchmarks and, as of the time of
this writing, H2O-Danube2-1.8B achieves the top ranking on Open LLM Leaderboard
for all models below the 2B parameter range. The models follow core principles
of LLama 2 and Mistral, and we leverage and refine various techniques for
pre-training large language models. We additionally release chat models trained
with supervised fine-tuning followed by direct preference optimization. We make
all models openly available under Apache 2.0 license further democratizing LLMs
to a wider audience economically. | H2O-Danube-1.8B Technical Report: We present H2O-Danube, a series of small 1.8B language models consisting of
H2O-Danube-1.8B, trained on 1T tokens, and the incremental improved
H2O-Danube2-1.8B trained on an additional 2T tokens. Our models exhibit highly
competitive metrics across a multitude of benchmarks and, as of the time of
this writing, H2O-Danube2-1.8B achieves the top ranking on Open LLM Leaderboard
for all models below the 2B parameter range. The models follow core principles
of LLama 2 and Mistral, and we leverage and refine various techniques for
pre-training large language models. We additionally release chat models trained
with supervised fine-tuning followed by direct preference optimization. We make
all models openly available under Apache 2.0 license further democratizing LLMs
to a wider audience economically. | [
0.0242344048,
0.0507000275,
0.0080124736,
0.0418996885,
-0.0022693735,
0.01003512,
-0.028810991,
0.0100280438,
0.0615792423,
-0.0472183526,
-0.0828597397,
-0.065104343,
-0.0266982019,
0.028331859,
0.0284548886,
-0.0459532887,
0.0614134371,
-0.0413173251,
-0.0214896332,
-0.0185572673,
0.0189400613,
0.0006072419,
-0.0032605256,
0.0338185132,
0.0302825626,
-0.0020552136,
-0.0161230154,
0.0168008059,
-0.0193203744,
-0.027334474,
0.0204870515,
0.0559507795,
0.04112982,
0.0179867502,
0.0000019861,
-0.0512866378,
0.0098024737,
-0.0248814318,
-0.0383743457,
0.0308102686,
0.0630234182,
0.0168628581,
-0.0309629664,
0.0376776569,
-0.0272236001,
0.0011618084,
0.0454571433,
0.0342930071,
0.0517259128,
0.0851079375,
0.0027969833,
-0.0378380381,
0.0453679785,
-0.0454878472,
0.0173617397,
-0.045638144,
0.0364993103,
-0.0159599856,
-0.0021881366,
0.0297464523,
-0.0023637393,
0.0255031139,
-0.0061769192,
-0.0046312078,
-0.0108644646,
0.0332190245,
-0.0381583422,
-0.0345888995,
-0.0235514808,
0.0802568197,
-0.0174040683,
-0.0335567333,
0.0149588576,
0.0194378998,
-0.0481336936,
-0.0498777963,
-0.0441175252,
0.0054204836,
0.0356746092,
0.0238972288,
0.0359509662,
-0.005370622,
0.0377128869,
0.0165934712,
-0.0032954908,
0.0359053835,
0.0133315958,
-0.0033372189,
0.0024234294,
-0.0139906751,
-0.0158355143,
-0.0784922987,
0.0168851819,
0.03449421,
-0.0134967705,
0.0060803853,
-0.0254800953,
0.0344711356,
0.0471287072,
-0.0809738934,
0.0414290018,
0.0342963785,
-0.029187534,
0.0503241867,
-0.0079311905,
0.0263611395,
-0.0366117544,
0.0397626795,
-0.0227022935,
-0.005823771,
0.017074706,
0.0260092374,
-0.0487667136,
0.0162191931,
-0.0350358337,
-0.0238556545,
-0.0556002669,
0.0048386203,
0.0495521389,
-0.00467858,
0.0431267023,
-0.0041899681,
-0.021714028,
-0.010130153,
-0.0154810986,
0.0021245312,
0.0081848707,
0.0197242852,
0.0080210716,
-0.0990153328,
0.0013116893,
0.0305280462,
-0.0127834603,
-0.0118487207,
0.0126234619,
0.1006435379,
-0.0018038501,
-0.0387058035,
-0.0395077989,
0.0292682368,
0.0004539233,
-0.0752096474,
-0.0047744908,
-0.0521718897,
0.014900961,
-0.0155803403,
-0.0039733052,
-0.0234397519,
0.0220426042,
0.0069817193,
-0.0391972363,
0.0460936166,
-0.0359974019,
-0.0201618187,
0.0080815386,
0.0028901426,
0.0021510208,
0.0904656425,
0.0374863185,
0.0541354977,
0.0375024341,
-0.0054696016,
-0.0091751125,
0.0644170865,
0.0204414167,
0.0131222922,
-0.0199648403,
0.0218872204,
0.0057116942,
0.0316016413,
0.0559213944,
0.0839828625,
-0.0469361432,
-0.0476048253,
0.0816963837,
0.0603404157,
0.0365654677,
0.0099437656,
0.0454921313,
0.0082019512,
0.0182670336,
-0.0221822467,
-0.0461411066,
0.0262634158,
0.0114801992,
0.0005126835,
0.0317372009,
0.0280786511,
-0.0497044548,
-0.0495053083,
-0.025524104,
-0.0283693168,
-0.0197738558,
-0.0054403222,
0.0315795764,
-0.0069029033,
-0.0613646545,
0.116363667,
0.0484802127,
-0.0459849313,
-0.020328816,
-0.0601383038,
0.0460211188,
0.0837613568,
0.029494144,
-0.0189627111,
-0.0059957337,
-0.0201917328,
0.0163754169,
0.0338975154,
0.0751027241,
0.0024205919,
0.0342112593,
-0.0930437073,
0.0027945333,
-0.0014993473,
0.0290527921,
0.0282400213,
0.0007256951,
0.0095583089,
-0.0029582377,
-0.0099432273,
-0.0127728814,
0.0225719213,
-0.0168546606,
0.0010316318,
0.0041816351,
-0.0038520943,
0.0312553719,
-0.010575586,
0.0096964044,
-0.0012655949,
0.0421299078,
-0.0261969268,
0.0152393328,
0.0104809171,
-0.034890458,
-0.0404937118,
0.038112212,
0.0276513528,
0.0320714042,
0.0481262505,
-0.035277117,
-0.0228087157,
0.0205834191,
-0.0110740308,
0.0626915321,
-0.0494536832,
0.0366254859,
-0.0051214462,
0.0433594026,
-0.0241849162,
0.0229092706,
-0.0419589095,
0.0132003641,
-0.0232179128,
0.0028474722,
-0.0065667848,
-0.0500722975,
-0.0167173911,
0.0298554786,
-0.0092809908,
0.0257791057,
-0.0266703367,
0.028999202,
-0.0200705491,
-0.0403894112,
-0.0953916758,
0.0196246579,
-0.0412780084,
0.030389037,
0.016749572,
0.0086043207,
-0.0114827231,
-0.0194780529,
-0.033894293,
-0.0096401041,
-0.0027770412,
0.0204357971,
0.0241225976,
-0.0101344883,
-0.0401839949,
-0.086636968,
-0.0309446957,
0.0037501026,
0.02188663,
-0.0310118832,
-0.0316771008,
-0.0281572547,
0.0314225964,
-0.0351913758,
0.038517274,
-0.0263315812,
-0.0568432361,
-0.0057287766,
-0.0100962128,
0.0455119684,
0.0421054251,
0.0196148232,
0.0347750783,
0.0330377296,
-0.0042402945,
-0.0511224195,
-0.0008825714,
-0.018927481,
-0.0206912961,
-0.0557874478,
-0.009698105,
-0.0026587227,
0.0988903567,
-0.0521213412,
0.0022114853,
-0.0307611246,
-0.0303341933,
0.0050227758,
-0.0144737316,
-0.0037647136,
0.0013529132,
-0.031897489,
0.0420735776,
-0.0235832985,
-0.0625850782,
0.0213668067,
-0.058555197,
0.0013421996,
-0.0286631528,
-0.0211109594,
-0.0548396446,
-0.0053842259,
-0.0005615639,
0.0130970581,
-0.0468797684,
-0.0394867174,
-0.0155888842,
-0.0179501921,
-0.015752092,
-0.0360907763,
0.039607048,
-0.0090728868,
0.0166317672,
-0.0223673359,
-0.0505317189,
-0.0321851112,
-0.0362211429,
-0.0491685532,
0.0086411955,
0.0459285043,
-0.0230472982,
-0.0437063947,
0.0252073053,
-0.0316892527,
0.0314753801,
-0.0151761919,
0.0232738014,
-0.0371701084,
0.0435175002,
0.0482880138,
0.0066470494,
0.0443428867,
-0.0404988229,
0.0371226072,
0.0420517661,
0.0488583036,
-0.0130150765,
-0.0042401319,
0.0119406478,
0.0675648078,
-0.0233197995,
-0.0301911645,
0.0057851588,
0.0709114671,
-0.0106217107,
0.0011133492,
0.0997205451,
0.0037558028,
-0.0457013957,
0.0126392422,
-0.026074715,
0.0557455532,
0.0250439923,
0.0146582779,
-0.0897180438,
-0.0333826244,
0.0168491788,
-0.0099111153,
0.0055147936,
-0.0144782662,
-0.0266859122,
-0.0707131848,
0.0191530902,
0.0136736613,
-0.0326949283,
0.0238129906,
-0.0468812175,
-0.0124247679,
-0.0400896594,
0.0463505872,
0.0094108274,
-0.0134939533,
-0.03540859,
0.0317602679,
-0.015886886,
-0.028121924,
0.0484078079,
-0.0196264181,
0.0640589148,
0.0286085512,
0.0255010594,
-0.0010668804,
-0.040657986,
0.0704599321,
0.0034676408,
-0.0186064653,
-0.0251432043,
0.0224543083,
-0.0006733515,
-0.0277022291,
0.0016223125,
0.0171921439,
-0.0034391386,
0.010573864,
-0.0340775996,
0.0169569198,
0.0046606869,
-0.0265233554,
0.0016143468,
0.0270379726,
-0.0068721143,
0.0551885664,
0.0451177806,
-0.023870999,
-0.0264996439,
-0.0341502875,
-0.0070041697,
-0.0298738778,
-0.0370019004,
-0.0136012044,
0.0087314481,
-0.0568849817,
-0.0610507242,
-0.0250362083,
-0.0102554569,
0.008874096,
-0.0172668044,
0.0341071039,
0.0176548902,
0.000761196,
-0.0316942669,
0.0475479737,
-0.044135388,
0.0313109048,
-0.1024258882,
0.0092306668,
0.0639707446,
-0.0217841212,
-0.0331249759,
0.044234328,
-0.1230318174,
-0.0801436529,
0.0045473007,
0.0669316947,
0.0275074989,
0.0354159549,
-0.0172117595,
-0.0699774846,
0.0205708016,
0.0155046694,
-0.0120998258,
0.0200067945,
0.0479977168,
0.0248364434,
0.0032465621,
0.0606136471,
-0.0274443254,
0.0188396275,
0.0206198636,
0.0129824905,
-0.0246739909,
-0.0265255664,
-0.0136647094,
0.009385515,
0.0318059176,
0.0029603515,
-0.0162177142,
-0.0267456099,
0.0625195727,
-0.0048748958,
0.0503937118,
-0.0443264544,
0.0200259499,
0.0191924777,
-0.0162976403,
-0.0183572788,
-0.0221697502,
0.0012247119,
0.0016797724,
-0.0101424269,
0.0348703004,
0.0040313937,
0.0112726782,
0.0378602222,
0.009628443,
-0.0319461152,
0.0404951237,
0.0151607608,
0.0535092615,
-0.0244027134,
0.0229054913,
-0.022274144,
0.0009373616,
-0.0266794879,
0.0508309193,
-0.03252846,
-0.0402509384,
-0.0369818471,
0.0777358189,
0.0782357082,
0.0274730828,
0.054417856,
-0.0028021846,
0.0006458096,
-0.0304181948,
0.0397446826,
-0.0413438305,
-0.0524047427,
0.049842868,
0.0101557951,
-0.0427771136,
0.0457429215,
0.0676585585,
0.0295310002,
0.0522655547,
0.0162807424,
0.0360672884,
-0.0155047951,
-0.0128449406,
0.0237596929,
0.0399689302,
-0.0244673379,
-0.0273196436,
-0.0281215105,
-0.0673772991,
0.003588815,
0.0230243541,
-0.0285289455,
-0.0132722789,
0.0280160755,
0.0517652296,
0.0131798582,
-0.0047880122,
-0.0440764911,
-0.0565736778,
-0.0045865364,
-0.0248629637,
0.002768253,
-6.80363861e-33,
-0.0239473749,
0.0475450307,
0.0345275067,
-0.0357028805,
-0.0166633017,
-0.0131875556,
-0.0280974731,
-0.0227745119,
-0.0743960217,
-0.006144966,
-0.0279470291,
0.0041753519,
0.0293331221,
0.0082513308,
-0.0184719879,
0.0053457613,
0.046497006,
-0.0110067092,
-0.0127865775,
0.0420113392,
0.0154416375,
0.0318563916,
0.0734148547,
-0.0116050625,
-0.0275877547,
-0.0254445244,
-0.0316968597,
0.0351784229,
-0.0217980891,
0.0573294349,
-0.0558543131,
0.0338025466,
-0.0059784157,
-0.0197145883,
-0.0298037324,
-0.0077002356,
-0.0588238575,
-0.0105715431,
0.0303512048,
0.0226808637,
-0.0069720154,
-0.0418082736,
0.0508868881,
0.0097590182,
-0.0275269356,
0.0005450521,
-0.0013709944,
0.0100925229,
-0.037143331,
-0.0181803349,
-0.0312109832,
0.024439238,
-0.0029320132,
-0.0274928901,
0.0215124507,
-0.0023659123,
-0.0088412622,
0.0231453869,
-0.0589147881,
0.0445094444,
0.0202670451,
0.0519816577,
0.0596317574,
-0.0076620467,
0.0017743737,
0.0566518866,
0.0445252471,
0.0082384078,
0.0145389521,
0.0016566286,
0.0488053858,
0.0412056744,
0.0061976579,
-0.0177106094,
0.0085125193,
-0.0522263795,
-0.0855055824,
0.0015838513,
0.0798842236,
0.0275684334,
-0.0161006097,
0.0266166274,
-0.0433878265,
-0.0195278823,
-0.0354069769,
-0.0238283779,
-0.0184659176,
-0.0399909168,
0.0171342436,
0.0176499151,
0.0303326938,
-0.0280970596,
0.0379621126,
-0.047028482,
-0.0494728722,
0.0073661269,
0.0041677016,
0.0237262566,
-0.0169383399,
0.0555375293,
-0.0787883475,
-0.0370601006,
0.0063522873,
-0.0102630993,
-0.024557367,
0.0049784724,
0.0123560773,
0.0081666131,
-0.0764079913,
0.0060729003,
-0.0316045694,
0.0062896777,
0.0272525996,
0.0565876663,
0.0331612937,
-0.0405362435,
-0.0042057894,
0.0698919669,
-0.0229007304,
-0.0089977235,
-0.0110998908,
-0.0041641109,
0.0053278413,
0.0653652847,
-0.0240937956,
0.0161113124,
-0.0603194982,
0.0336381979,
0.0050332639,
-0.0295055788,
-0.0114681395,
0.0100743333,
2.684e-7,
0.0052442388,
0.0497094207,
-0.0045850212,
0.0145730451,
0.0049200668,
0.0162892565,
-0.0366073884,
0.0085622724,
0.009526304,
0.0005107694,
-0.0035510254,
0.0344524346,
0.0478910953,
-0.0109937787,
-0.1029566377,
0.0004570401,
-0.1043377817,
-0.0249260366,
-0.0276036914,
-0.0604352653,
0.0061383997,
0.1461217552,
-0.0014777402,
-0.0175917856,
-0.0112866564,
-0.003275265,
0.0557983667,
-0.0283882339,
-0.0369546525,
0.0531085692,
0.0460616983,
0.0217505451,
-0.0278102551,
-0.0362804644,
-0.0170520395,
0.037625324,
-0.0195332859,
0.0810568556,
-0.0074118399,
-0.0441644378,
0.0316233076,
-0.0072030025,
-0.0501277968,
-0.0524179302,
0.0658004507,
-0.0075811888,
-0.0329204425,
-0.0521356612,
0.0042232391,
-0.0215950161,
0.0162672084,
-0.0345175266,
-0.0199829079,
0.0290523805,
-0.00089445,
-0.0423545279,
-0.0171146858,
-0.0691497549,
-0.0198045727,
-0.0231198668,
-0.042920839,
-0.0410895832,
-0.0373619124,
0.0444153808,
0.0123271579,
-0.0138347428,
0.0098311128,
2.271916962e-34,
0.0557267107,
-0.0148165487,
0.0026599257,
0.0647657812,
0.021671487,
-0.0124879098,
0.0360449739,
0.0069594369,
0.0003938691,
-0.0851702988,
-0.0052291979
] |
http://arxiv.org/abs/2402.01752v1 | http://arxiv.org/pdf/2402.01752v1 | Identifying False Content and Hate Speech in Sinhala YouTube Videos by Analyzing the Audio | 2024-01-30T08:08:34 | 2024-01-30T08:08:34 | W. A. K. M. Wickramaarachchi, Sameeri Sathsara Subasinghe, K. K. Rashani Tharushika Wijerathna, A. Sahashra Udani Athukorala, Lakmini Abeywardhana, A. Karunasena | eess.AS | eess.AS, cs.AI, cs.CL, cs.LG, cs.SD | YouTube faces a global crisis with the dissemination of false information and
hate speech. To counter these issues, YouTube has implemented strict rules
against uploading content that includes false information or promotes hate
speech. While numerous studies have been conducted to reduce offensive
English-language content, there's a significant lack of research on Sinhala
content. This study aims to address the aforementioned gap by proposing a
solution to minimize the spread of violence and misinformation in Sinhala
YouTube videos. The approach involves developing a rating system that assesses
whether a video contains false information by comparing the title and
description with the audio content and evaluating whether the video includes
hate speech. The methodology encompasses several steps, including audio
extraction using the Pytube library, audio transcription via the fine-tuned
Whisper model, hate speech detection employing the distilroberta-base model and
a text classification LSTM model, and text summarization through the fine-tuned
BART-Large- XSUM model. Notably, the Whisper model achieved a 48.99\% word
error rate, while the distilroberta-base model demonstrated an F1 score of
0.856 and a recall value of 0.861 in comparison to the LSTM model, which
exhibited signs of overfitting. | Identifying False Content and Hate Speech in Sinhala YouTube Videos by Analyzing the Audio: YouTube faces a global crisis with the dissemination of false information and
hate speech. To counter these issues, YouTube has implemented strict rules
against uploading content that includes false information or promotes hate
speech. While numerous studies have been conducted to reduce offensive
English-language content, there's a significant lack of research on Sinhala
content. This study aims to address the aforementioned gap by proposing a
solution to minimize the spread of violence and misinformation in Sinhala
YouTube videos. The approach involves developing a rating system that assesses
whether a video contains false information by comparing the title and
description with the audio content and evaluating whether the video includes
hate speech. The methodology encompasses several steps, including audio
extraction using the Pytube library, audio transcription via the fine-tuned
Whisper model, hate speech detection employing the distilroberta-base model and
a text classification LSTM model, and text summarization through the fine-tuned
BART-Large- XSUM model. Notably, the Whisper model achieved a 48.99\% word
error rate, while the distilroberta-base model demonstrated an F1 score of
0.856 and a recall value of 0.861 in comparison to the LSTM model, which
exhibited signs of overfitting. | [
0.0704364404,
0.0208912268,
-0.0240513682,
0.0292236563,
-0.002620199,
0.0241372269,
0.0311458558,
0.0082857283,
-0.0202237219,
-0.0528034046,
-0.0099969478,
-0.0697786808,
0.0236190036,
0.0417536311,
-0.0344849825,
-0.070073016,
0.032008037,
0.0148236044,
0.0777415559,
-0.0066105439,
0.0436013229,
0.0210106522,
-0.0363500044,
-0.0073935548,
-0.0528626591,
0.0228805095,
0.0412903503,
-0.0115516819,
0.0403080732,
-0.0142489476,
0.0079232957,
0.0210658759,
0.0201808661,
0.0648073629,
0.0000018531,
-0.0421498604,
0.0134714022,
-0.0259938799,
-0.0035319913,
0.0142001128,
0.0525649264,
0.0380466506,
-0.0421783961,
0.081941843,
-0.0128772128,
0.0501928665,
0.0424897149,
0.0382489674,
0.0734423026,
0.0638046786,
-0.0175280161,
-0.0670992061,
0.0074645877,
0.0112284506,
0.044153057,
0.057617791,
0.0251235608,
-0.0132464888,
-0.0133982673,
-0.0130696557,
-0.0192003269,
0.0378850624,
0.0078649893,
-0.0094294688,
-0.0077302358,
-0.0482929051,
-0.0298908614,
-0.078801021,
0.0486123711,
0.0176777039,
-0.0296684224,
0.0138208857,
0.016480539,
0.0360786207,
-0.0347423293,
0.0067753526,
-0.0307893232,
0.0231211539,
-0.0130654061,
0.020171294,
0.0283641126,
0.0287165884,
0.0336490944,
0.026394451,
0.0587914884,
-0.0069675688,
0.001770656,
-0.0286980104,
-0.0213701557,
-0.0525732785,
0.0132679138,
-0.0059872614,
0.0638101995,
0.0156048192,
-0.0020311642,
0.0132156983,
-0.0264610704,
-0.0574422255,
-0.0212706346,
-0.1245161518,
0.0287480541,
0.0463576317,
-0.0303472877,
0.0179074183,
-0.0175139401,
0.0584680699,
0.0256134719,
0.0119467061,
-0.0037444457,
0.031675972,
-0.0181647372,
0.0360028446,
0.0231433492,
0.0060182586,
-0.0014246515,
0.0222422332,
-0.0741873458,
-0.011546392,
-0.0662149861,
-0.0086043151,
-0.0267138816,
-0.0159866046,
0.0141905202,
-0.0038492933,
-0.0249222219,
0.0391722471,
-0.0372360758,
0.0484066717,
-0.0424211994,
-0.0659782216,
0.0160167012,
0.0398509912,
-0.0216212794,
-0.0046311938,
0.0615538135,
0.0229781158,
-0.0025767318,
-0.0217245407,
-0.0363824591,
-0.0259539392,
-0.0546228588,
-0.0316730179,
-0.0176975578,
-0.0852912664,
0.0451666154,
-0.0115871951,
-0.0387788117,
-0.0465001203,
0.0132424263,
0.0112621728,
-0.0155926263,
0.0461553298,
0.0371020883,
0.0073805437,
0.0860156342,
0.0520849489,
0.0277979784,
0.026677778,
0.0463566333,
0.0554156452,
0.0534179918,
0.0058907848,
0.0145729743,
0.0451593436,
0.0196500421,
0.0519690774,
-0.0623476095,
0.0007467906,
-0.0548879206,
-0.0098706894,
0.0076444019,
0.0218607765,
0.0161606241,
-0.035426449,
0.0458104424,
0.048074469,
-0.0444968157,
0.0099333758,
0.0328205638,
0.0570440665,
0.0139634628,
-0.0118818739,
0.0216270369,
0.0268337019,
-0.0011106908,
0.0057625724,
0.0887292475,
0.0725593418,
-0.0263109002,
-0.0317701362,
-0.03059626,
-0.014688883,
0.0395561121,
-0.0237309448,
0.0372928791,
0.0031297111,
-0.0314970911,
0.0295696463,
0.0544952601,
-0.0196814537,
0.0374444425,
-0.0662928894,
0.0669300705,
0.0233315844,
-0.0141008534,
-0.0136623401,
0.0106963655,
-0.0864493623,
-0.0373729244,
0.0405678377,
-0.0116974013,
-0.0140123209,
-0.0380082205,
-0.0354365446,
-0.0304603092,
0.0247837827,
-0.0026139484,
0.0231043175,
0.0048987935,
0.0402131416,
0.0485239774,
0.0618314371,
-0.0377592482,
-0.0132789044,
0.0532051399,
-0.0110397907,
-0.0353347734,
0.0132066654,
0.1067029834,
0.0357439928,
-0.0422180668,
-0.0393216237,
0.0298731793,
-0.0010414708,
-0.0427952036,
-0.0035400153,
-0.0274545997,
-0.0023597158,
0.0272689909,
-0.0788817182,
0.0401220769,
0.0881022438,
0.0154560953,
0.0017945393,
0.082501173,
-0.0145861199,
-0.0230542235,
-0.0486828983,
0.0069188364,
-0.045691669,
-0.0119618317,
0.0155168511,
0.0221807044,
0.0086249737,
0.0098937312,
0.0096289786,
-0.0336060263,
0.0477294475,
0.0046555409,
-0.0678529441,
-0.0034743561,
-0.0518031679,
0.0358729474,
-0.0078278566,
-0.0249766391,
-0.0139378803,
-0.0013984051,
-0.039194528,
-0.005347271,
-0.0333189517,
-0.0085153654,
0.0120489728,
0.0023681091,
-0.0435529575,
0.0164507013,
-0.0002290063,
-0.0933223367,
0.0263736565,
0.0012349972,
-0.0360724181,
-0.0339672714,
-0.0195355825,
-0.0003972448,
-0.0124057643,
-0.02799966,
0.0113511439,
0.0170988366,
-0.0806816444,
0.0627112836,
0.016771961,
-0.0394712463,
-0.0099311564,
-0.016324101,
-0.0605302751,
-0.0019995242,
0.0353088081,
-0.0131254308,
0.0103370845,
0.0700273514,
0.0429425202,
0.0242439657,
0.0154557368,
0.0227032173,
-0.0176692326,
-0.0165380817,
0.0146782575,
-0.0073885177,
-0.0350040495,
-0.0191183183,
0.0298329908,
0.0751500949,
0.0200604275,
-0.0344525836,
-0.0016455445,
-0.0855364725,
0.0029857301,
-0.0362311937,
-0.0740456879,
-0.0099698342,
0.0299264044,
0.0063285427,
-0.0414441824,
0.0456911661,
-0.0118021937,
0.041162122,
-0.0579812489,
-0.0232791584,
-0.0110148713,
-0.006410941,
-0.0170835648,
0.066218175,
-0.0335756652,
-0.0257072616,
-0.0112225506,
-0.0208815262,
0.0425056703,
-0.0480094291,
-0.0268878248,
0.008094403,
0.0066655749,
-0.0066495351,
0.0014259962,
-0.0365227796,
-0.062082734,
-0.0219529439,
-0.0556268692,
-0.0042700302,
0.0118452786,
-0.0207995269,
0.0625795275,
0.0478759818,
-0.0222585034,
-0.0117207384,
0.046759706,
0.0507969856,
0.0139696402,
0.0198050942,
-0.0418918841,
-0.0102217877,
0.0020031468,
-0.0020350674,
0.0128425555,
0.0118418243,
0.0078991093,
0.0009612886,
0.0111013018,
0.006157435,
0.0072569083,
0.0133819534,
0.0122998599,
0.0118236467,
-0.0005647718,
0.0069499747,
0.1099169105,
-0.000452083,
-0.0009038402,
-0.047507342,
-0.001162936,
-0.0103671886,
0.010688521,
-0.0179879554,
-0.0747398585,
0.0312126987,
0.0090310834,
0.0149653917,
-0.0501848049,
0.0055282908,
-0.0087208245,
-0.0291752256,
0.0050439062,
0.0065531153,
-0.0132450666,
-0.0334682241,
-0.0177993029,
-0.0250268281,
0.0127531206,
0.0210926291,
0.0050001154,
-0.0070417356,
0.0030706765,
0.0441011712,
-0.0602561347,
-0.0181915089,
0.0286435094,
-0.0276674014,
0.0424354784,
0.0395145528,
0.0351902395,
-0.0853173062,
-0.0047570299,
-0.0134299556,
-0.0076768533,
-0.0249622911,
-0.0290579405,
0.0327569395,
-0.0054240022,
-0.0353348516,
-0.0046750866,
0.0240348466,
0.0330387913,
0.0047260113,
0.0020490959,
0.0476726554,
0.0406965949,
0.0016073857,
-0.0649902597,
0.0671942607,
-0.0009690061,
0.0703641623,
0.0361262448,
-0.0424340405,
0.072562851,
-0.0189686473,
-0.0570919812,
-0.0002980142,
-0.0779369846,
-0.0236841105,
-0.0165995229,
-0.014545178,
-0.02204201,
0.0002341733,
-0.013804392,
-0.0128111131,
0.0240872093,
0.027943844,
0.0187169779,
0.0000550849,
-0.0665309057,
-0.0119013088,
0.0292152185,
0.0411466546,
-0.0258392543,
0.0155299585,
0.0072608986,
-0.0433763899,
-0.0476518348,
0.0105950963,
-0.0517514311,
-0.0184914917,
0.0820366666,
0.0326465108,
0.0293667633,
0.0107262284,
0.0007073908,
-0.0172167718,
0.1041638777,
0.0150134331,
-0.010848674,
-0.0132996822,
0.0277037155,
0.016572684,
0.0056227897,
-0.002766687,
-0.0039175963,
0.0525153317,
0.0175636616,
-0.0691424981,
-0.0439533927,
-0.0242833272,
-0.0452777892,
-0.0254560914,
0.0182047132,
0.0057912543,
-0.0428777002,
-0.0651742145,
-0.0301589444,
0.016682066,
0.0095493002,
0.0007719815,
0.0299165249,
0.0305529032,
0.0062164045,
-0.0304499287,
-0.0197223425,
-0.054743778,
-0.0465730019,
-0.0451583862,
0.0498775169,
0.0123829488,
-0.0162408687,
0.0092881806,
0.0040203794,
-0.0093903877,
0.0320758447,
0.014772471,
-0.0295020118,
0.0206096582,
0.0262020007,
-0.0055768634,
0.0107582565,
0.0813288018,
0.0276913736,
0.0334148258,
-0.0008195246,
-0.0696667433,
-0.0227899589,
-0.0020866236,
-0.0262570549,
0.0254556499,
0.0060728998,
0.0078370515,
-0.0041002585,
0.0183489788,
-0.0525264181,
-0.0323149264,
0.0363574512,
0.0229261145,
-0.0161096901,
0.0232910831,
0.0248922613,
0.0029083283,
0.0004926343,
0.0778077394,
-0.0394869074,
-0.015371263,
0.008512239,
-0.0163048841,
0.0544566214,
-0.0107245892,
0.0031510512,
-0.0445163883,
-0.0436088406,
0.0204038266,
-0.0718380511,
0.0618763529,
0.0049467878,
0.0358875766,
0.0293055363,
-0.0106936591,
0.0553862266,
0.0106870309,
-0.0679679364,
-0.0063989954,
-0.0385308638,
0.0400345773,
-5.681620804e-33,
-0.0248828214,
-0.0047772462,
-0.042191606,
0.0520602912,
-0.0151206078,
-0.0357375965,
0.0208644923,
0.0262549128,
-0.0355874337,
0.0073817731,
0.0042798342,
0.0372047983,
0.0188287608,
0.0173113421,
0.0114680203,
-0.0356456563,
0.0298805311,
-0.0114694135,
-0.0097659444,
0.0017752984,
0.0772362873,
0.0096085146,
0.0621045083,
-0.0999047831,
0.0095132748,
-0.0133033544,
-0.0339242443,
-0.0319318995,
0.019992169,
0.080503732,
-0.0150322095,
0.0661893338,
0.0105170459,
-0.0268360302,
0.0033415635,
0.0175514892,
-0.0030485571,
0.0107524926,
0.0326873735,
0.0051546977,
0.0173442569,
0.0012599484,
0.0541992672,
0.0072945105,
-0.0598430112,
-0.0241361577,
0.0418740027,
-0.0435973071,
0.0185222309,
-0.069574371,
-0.0744857267,
0.0061380332,
-0.0138688534,
-0.0401240624,
0.0057508629,
0.0234979801,
0.0374023914,
-0.0247861389,
-0.0745716542,
0.0286514927,
0.0040148837,
0.0212853588,
0.0375038721,
-0.0170767363,
-0.0445863046,
0.0470793247,
0.0211233236,
-0.0672478452,
-0.0137272431,
-0.0157718044,
0.0153424377,
0.010414714,
0.041509714,
0.0304129999,
0.034815684,
-0.0955237448,
-0.0635309815,
0.0130365202,
0.0309626311,
0.0140138613,
0.0427432321,
-0.0138017796,
-0.0516645052,
-0.0305652414,
-0.0339303277,
-0.0476869419,
-0.0367372148,
-0.0306798294,
0.0074775815,
-0.0320545956,
-0.012316227,
-0.0195530001,
0.023912834,
-0.0147781195,
0.0729714483,
0.0040999609,
-0.0406548902,
0.031758666,
0.0275627952,
0.0305905528,
-0.0328505114,
0.0155927092,
0.0084224567,
0.0108082928,
0.0121085411,
-0.0242076088,
0.0096860062,
0.0522465967,
-0.0860147104,
-0.0006970939,
0.0341370367,
0.0261256844,
0.011851917,
0.0598885827,
-0.0130675389,
0.0478870384,
-0.0307965204,
-0.0198195204,
0.0240181014,
-0.0229530651,
-0.0079329861,
0.0032799479,
-0.0357241258,
-0.0226129796,
-0.032369785,
-0.0014889918,
-0.0050540464,
0.0066384235,
0.015168529,
-0.0863254443,
-0.0035706649,
-0.0432560258,
2.707e-7,
-0.0367591716,
0.0411712714,
-0.0221493244,
-0.0365233496,
0.028258564,
0.0378760807,
0.0258241501,
-0.0047654412,
0.0135467649,
-0.0113496259,
0.0003919734,
0.005764449,
0.0265518613,
0.0521278344,
-0.0480906852,
-0.0301006697,
-0.092248477,
-0.0226779617,
-0.0409976654,
-0.0031556364,
-0.0189557765,
0.0842870325,
0.0659404695,
-0.0621212125,
-0.0280169304,
-0.0145957153,
0.0165399872,
0.0073126615,
-0.033932846,
-0.0333581939,
0.0198153257,
0.0279583558,
0.0160415433,
-0.0364134945,
-0.0036798697,
0.0012921558,
0.0507250726,
0.055655241,
0.0179845896,
0.0055374457,
-0.0156844445,
0.0309485681,
0.0057907943,
-0.0553291216,
0.0199292954,
0.0614848025,
-0.0367734879,
-0.0519343056,
-0.0286630169,
0.0012821672,
0.0522780754,
-0.0156653132,
-0.0008341206,
-0.02736043,
0.0096681295,
0.0031008902,
-0.0229356848,
-0.0618002489,
-0.0022077623,
-0.0194027387,
-0.0256923568,
-0.0414921567,
-0.0187675394,
0.0064499439,
0.0181008093,
0.0215613637,
0.0245626383,
2.428558472e-34,
0.0506765246,
-0.0444073007,
-0.03035645,
-0.0722494125,
0.0161096361,
0.041952882,
-0.0389330313,
-0.0262372755,
0.0265749041,
-0.0182440951,
-0.0359794237
] |
http://arxiv.org/abs/2402.01751v1 | http://arxiv.org/pdf/2402.01751v1 | Performance Assessment of ChatGPT vs Bard in Detecting Alzheimer's Dementia | 2024-01-30T07:55:43 | 2024-01-30T07:55:43 | Balamurali B T, Jer-Ming Chen | cs.CL | cs.CL, cs.AI | Large language models (LLMs) find increasing applications in many fields.
Here, three LLM chatbots (ChatGPT-3.5, ChatGPT-4 and Bard) are assessed - in
their current form, as publicly available - for their ability to recognize
Alzheimer's Dementia (AD) and Cognitively Normal (CN) individuals using textual
input derived from spontaneous speech recordings. Zero-shot learning approach
is used at two levels of independent queries, with the second query
(chain-of-thought prompting) eliciting more detailed than the first. Each LLM
chatbot's performance is evaluated on the prediction generated in terms of
accuracy, sensitivity, specificity, precision and F1 score. LLM chatbots
generated three-class outcome ("AD", "CN", or "Unsure"). When positively
identifying AD, Bard produced highest true-positives (89% recall) and highest
F1 score (71%), but tended to misidentify CN as AD, with high confidence (low
"Unsure" rates); for positively identifying CN, GPT-4 resulted in the highest
true-negatives at 56% and highest F1 score (62%), adopting a diplomatic stance
(moderate "Unsure" rates). Overall, three LLM chatbots identify AD vs CN
surpassing chance-levels but do not currently satisfy clinical application. | Performance Assessment of ChatGPT vs Bard in Detecting Alzheimer's Dementia: Large language models (LLMs) find increasing applications in many fields.
Here, three LLM chatbots (ChatGPT-3.5, ChatGPT-4 and Bard) are assessed - in
their current form, as publicly available - for their ability to recognize
Alzheimer's Dementia (AD) and Cognitively Normal (CN) individuals using textual
input derived from spontaneous speech recordings. Zero-shot learning approach
is used at two levels of independent queries, with the second query
(chain-of-thought prompting) eliciting more detailed than the first. Each LLM
chatbot's performance is evaluated on the prediction generated in terms of
accuracy, sensitivity, specificity, precision and F1 score. LLM chatbots
generated three-class outcome ("AD", "CN", or "Unsure"). When positively
identifying AD, Bard produced highest true-positives (89% recall) and highest
F1 score (71%), but tended to misidentify CN as AD, with high confidence (low
"Unsure" rates); for positively identifying CN, GPT-4 resulted in the highest
true-negatives at 56% and highest F1 score (62%), adopting a diplomatic stance
(moderate "Unsure" rates). Overall, three LLM chatbots identify AD vs CN
surpassing chance-levels but do not currently satisfy clinical application. | [
0.023171708,
0.0226322133,
-0.0222195312,
0.00043914,
-0.0694360659,
0.0257271305,
0.0468594581,
0.0486006215,
0.1033392921,
-0.0647864565,
-0.0099872816,
-0.0191296153,
0.0185917877,
0.0167299882,
-0.0189224835,
-0.0281880032,
0.0641901568,
0.0065732393,
0.0863454938,
-0.0139637683,
-0.0136018284,
-0.0094866809,
-0.0682286769,
0.0750205442,
-0.0343909524,
-0.0174069852,
0.0111251008,
-0.0314954259,
0.0010981275,
-0.0266348813,
-0.0303390101,
0.0445700698,
-0.0081937741,
0.0452931821,
0.0000020725,
-0.0476545729,
0.0078103524,
0.0353076644,
-0.0232768711,
-0.0575595424,
0.0074361274,
0.0544346124,
0.0250929855,
0.0398763418,
-0.0476871245,
-0.0340956151,
0.0256377012,
0.0302241296,
0.0186625924,
0.07484065,
-0.0117869908,
-0.0047962312,
0.0517570525,
-0.0029507105,
0.0371153094,
-0.0526852459,
0.048374746,
-0.0008128486,
-0.0232538432,
-0.037767794,
0.0026472493,
0.0282073859,
-0.0457956009,
0.0182315428,
0.0025010509,
0.0477043949,
-0.0491026901,
-0.0437529981,
-0.0237215664,
0.029024519,
0.006831245,
0.0353134312,
-0.0070289047,
0.0368012078,
0.0096781794,
0.0217281301,
-0.0672899336,
-0.0067584859,
0.000136805,
0.0006347081,
-0.0413299575,
-0.004955559,
0.0054642865,
-0.0161355883,
-0.0074454369,
0.0138977002,
-0.0065039494,
-0.015496498,
0.0304305181,
-0.0149570201,
-0.046897985,
-0.0279007796,
0.0204199795,
0.055234734,
-0.0473740287,
0.0170125663,
0.0064689005,
0.0023608403,
0.0484582894,
-0.0353320539,
-0.0029376457,
0.0333696604,
-0.0656254888,
0.0553149842,
-0.0238801576,
0.0670497045,
0.0293921977,
0.0191872362,
-0.0122465082,
0.0121083744,
-0.0574765801,
0.0023013537,
-0.0500994883,
0.0323781781,
0.0199062321,
-0.0429052077,
-0.0915654898,
-0.0204147696,
0.027868459,
-0.0142400982,
0.0215517115,
0.0138881756,
0.0329311378,
0.041391518,
-0.046625834,
0.0173590165,
0.0093355719,
0.0365682244,
0.0072272578,
-0.0891064852,
0.0172977839,
0.0446328856,
-0.0229460951,
-0.0337501168,
0.0109295417,
0.0937920436,
-0.0041909385,
-0.0631775185,
-0.02143443,
-0.0040065646,
-0.0702009499,
-0.0555654801,
0.0294804033,
-0.0202784277,
0.0037877033,
0.0213717483,
0.0087165227,
-0.0035101918,
0.009946147,
0.018630933,
-0.0438685566,
0.0428982452,
0.0006480808,
0.018333042,
0.0382228047,
0.0048845424,
-0.0552444868,
0.0309556406,
0.0312507302,
-0.0382618979,
0.0066815028,
-0.0169857759,
0.0068561677,
0.0299972333,
-0.0041242172,
-0.0157663003,
-0.0185905993,
-0.0009203309,
0.0234311093,
0.0426640883,
-0.015278317,
0.0631955639,
0.0181780886,
-0.0301988758,
0.0691709071,
0.0689628646,
-0.0055390433,
0.0034046953,
0.006001737,
-0.0104375444,
0.0282520764,
-0.0061017824,
-0.0200464446,
-0.0153581584,
-0.0582289174,
-0.0503074788,
0.0228589177,
0.0037140665,
-0.016377233,
-0.0216044523,
-0.0203469153,
-0.0342282355,
0.0255228039,
0.005603876,
0.0265930817,
-0.0417457893,
-0.0092315543,
-0.0163326319,
0.0031052029,
-0.0081459517,
-0.0669739172,
-0.093927294,
0.0230299179,
0.0690224171,
0.0325443856,
-0.032945402,
0.0135314465,
0.0013366254,
-0.0222691037,
0.0276929568,
-0.0110537084,
0.0348178782,
0.0283965226,
-0.0323767066,
0.0068134563,
0.0334013663,
0.0155900829,
-0.0237061474,
-0.0274850577,
0.0374233052,
-0.0138729298,
0.0088417158,
0.0229434911,
0.0308205783,
-0.007005021,
-0.0041104644,
-0.0199985057,
-0.0080448361,
0.0525717437,
0.0283201616,
-0.0050840462,
0.0173123963,
0.0113684116,
-0.0372663885,
-0.0442974381,
0.0044584861,
-0.0512868166,
-0.0624438003,
0.0205092579,
0.0411141738,
0.0088147698,
-0.0179107282,
-0.002497944,
-0.0350991227,
-0.0221328288,
-0.0641258359,
0.0327400714,
-0.0357621349,
0.0436698161,
-0.0381722711,
-0.0018421655,
-0.0298802219,
0.0636946708,
-0.0050024469,
-0.0385764092,
0.007089288,
-0.0131601794,
0.0241558496,
-0.0533051603,
-0.0727233589,
0.0062077702,
-0.0087875733,
-0.0107002221,
-0.0160099957,
0.0176689085,
-0.0385847576,
0.0009639371,
-0.135774076,
-0.015483289,
0.019810183,
-0.0244473275,
-0.0210268218,
-0.0195047986,
-0.0176768769,
0.0099812988,
-0.0521752276,
0.0659830719,
0.0151495654,
0.054360114,
-0.0038529797,
0.0109115373,
-0.0628627017,
-0.0162719265,
0.0163405053,
-0.0269611292,
0.0381634869,
0.0217493419,
-0.0720357224,
-0.0099429367,
0.0357997753,
-0.0393783487,
0.0208080001,
-0.0224893074,
-0.0305692293,
0.0140982093,
-0.0293073654,
0.0225335956,
0.0865291432,
0.0147035569,
0.0188484397,
0.0310613587,
0.0241555274,
-0.0213042703,
-0.0104876515,
-0.0193940159,
-0.0119428672,
-0.0274140444,
-0.0260061845,
-0.0117259277,
0.0180432517,
0.0627169609,
0.0397615954,
-0.0599598885,
0.0230647698,
-0.0126369148,
-0.0210397672,
-0.0394989848,
0.0007689476,
-0.038852755,
0.0139062619,
0.007099614,
-0.06628225,
0.0585865043,
-0.0384777784,
0.0126286214,
-0.0516812168,
-0.0291257408,
-0.0193765145,
0.0307167917,
0.0041358694,
0.0257031303,
-0.001849076,
-0.0208917521,
-0.0288414992,
-0.0259158537,
-0.0145200836,
-0.0567887723,
0.0360752717,
-0.014812693,
0.0333414041,
0.0077867955,
-0.0019917961,
-0.0491814017,
0.0042259176,
-0.0454598777,
0.0396771319,
0.0095320083,
-0.0180336237,
-0.1011410132,
0.0500210486,
0.0212841388,
0.0192453768,
-0.0187237933,
-0.0022330834,
0.0080023315,
0.0256540161,
-0.0070387446,
0.009189304,
-0.0106936349,
0.0160974413,
0.0412063822,
0.0467424355,
-0.0390447862,
0.0395974591,
-0.0251217484,
-0.0343375988,
-0.0133779887,
-0.006388349,
-0.0277401395,
0.0094158826,
0.0245582983,
0.0042962036,
-0.0120391743,
0.1417856067,
0.0366009623,
-0.0975544453,
-0.0203676447,
-0.0493579321,
-0.0071545066,
-0.0076912926,
-0.0262953117,
-0.0571826659,
-0.0581421852,
-0.0046774056,
0.0333193168,
-0.0558767691,
-0.0034423268,
-0.0072460989,
-0.0726625919,
0.0220239684,
-0.0246371739,
-0.0404564999,
0.0076632053,
-0.0220511612,
0.0165227074,
-0.0361874066,
0.0409115478,
-0.0056557488,
-0.0075787534,
-0.0332750604,
0.0161549691,
-0.0866295248,
-0.0409218408,
0.0652161762,
-0.0212855041,
0.0841888785,
0.0777562112,
0.041191265,
0.0066779978,
-0.0118455524,
0.0048433468,
-0.0111880023,
0.047738459,
-0.0297864787,
0.0810106546,
0.006315147,
0.0249345098,
-0.0526548587,
0.0054558627,
0.0248202793,
0.0013027637,
0.0222869106,
0.0580935217,
-0.0764633939,
-0.0077034943,
-0.0255573429,
0.042036552,
0.0131183546,
0.0748143792,
0.0011206414,
-0.0308684874,
-0.0224906802,
0.0280381497,
0.0163529851,
0.0476099737,
-0.0286770836,
-0.0018659323,
-0.0303866901,
-0.0264280699,
-0.0815101117,
-0.0055061658,
0.0714479759,
0.0143792545,
0.0042136498,
0.0376253873,
0.0214504302,
0.0558634028,
-0.0157135092,
0.0104053281,
0.0112248035,
0.0058144922,
-0.0814486519,
-0.0419164188,
0.0477456488,
-0.0204363689,
-0.0276730247,
0.0017926526,
-0.0043395679,
-0.0107624838,
0.042541258,
0.0284404699,
0.0219625216,
-0.0203221887,
-0.0073899543,
-0.0327515379,
0.0356089845,
0.0559739769,
0.0105269598,
0.0221384075,
0.0057015019,
0.0851587802,
-0.0108378399,
0.0284384005,
-0.007990567,
0.0204401556,
-0.0095885918,
0.0010051146,
-0.0175700262,
-0.011915884,
0.0339885354,
0.0470245779,
0.0419705808,
-0.0071571339,
0.0171745922,
0.0130751133,
0.0161124505,
-0.0053712269,
0.0318069048,
-0.029953951,
-0.0438811816,
0.0151064079,
-0.0164733343,
0.0056037395,
-0.0350337476,
-0.0808005631,
-0.0345770344,
-0.0270915274,
0.0015305427,
0.0068379347,
0.0197798852,
0.011733328,
-0.0231341552,
-0.006360326,
0.0251466595,
-0.0210159495,
-0.0114988638,
0.0230925363,
0.1128221229,
-0.0214217771,
-0.0339724347,
0.0072482168,
0.0099821184,
-0.0342078991,
-0.0860186666,
0.024883572,
0.0803629234,
0.0413288213,
0.0409325622,
0.0336779207,
-0.004832475,
0.0276119057,
-0.0354547128,
-0.0126642007,
0.0176297836,
-0.0038330038,
0.0388896242,
-0.0071283476,
-0.0302577559,
0.0067783934,
0.0751184374,
-0.0051322593,
0.0136252576,
-0.0133653171,
-0.0626717657,
-0.0200086515,
-0.0448397361,
0.0747892037,
-0.0183586013,
-0.0550870784,
-0.0082600284,
-0.0336348042,
-0.0069876509,
0.0055786176,
-0.0351649225,
0.0247708764,
0.0074102301,
0.0471773446,
0.0357713997,
-0.0128683457,
-0.0413805172,
-0.0473413728,
0.0090908585,
-0.0675834715,
-0.0052711261,
-0.0121214036,
-6.332989039e-33,
-0.0301374383,
-0.0095418273,
0.0170249194,
-0.0231966805,
0.0103584323,
0.0072777555,
-0.0120572494,
0.0257954635,
-0.0256526452,
0.0142823327,
0.0075045107,
-0.0088392319,
0.0185595397,
0.0195352361,
0.0068090372,
-0.0547229499,
0.0180176757,
-0.0043088947,
-0.0062451637,
0.0138582317,
0.0409912691,
0.0169940833,
0.0298892967,
-0.0592699498,
-0.0000683194,
0.0244387276,
-0.0180702619,
0.0463814065,
-0.0046288492,
0.0132755293,
-0.0297417063,
0.0413410179,
0.0305015948,
0.0530229062,
0.00578993,
0.0384854972,
-0.0191558488,
-0.0047864751,
0.0408504419,
0.0181949884,
0.0149426265,
0.0092202164,
-0.0023136917,
-0.0196898673,
-0.0280192979,
0.0239168294,
0.0397331864,
-0.059946429,
-0.0003187445,
0.0033060643,
-0.0124808345,
-0.0058201035,
0.0160463564,
-0.0243067462,
-0.026956331,
0.0416799821,
-0.0284640249,
0.0349284261,
-0.0482872687,
0.0251603276,
-0.027179813,
-0.0062063169,
0.0220493656,
-0.0102982018,
0.0103476057,
0.0945983529,
0.0694920793,
-0.0253662989,
-0.0143206678,
-0.0355499946,
0.0086117238,
0.0597529262,
0.0830363929,
-0.0284374505,
-0.0148587916,
-0.0345552228,
-0.0667338744,
-0.0129900845,
0.0737305284,
0.0091829952,
0.0554970205,
-0.0127407312,
-0.0426643826,
-0.0025194366,
-0.0127715711,
0.0459343754,
-0.0451539755,
-0.0275832452,
0.0247612726,
-0.008283617,
0.0332843475,
-0.065761812,
0.0441084243,
-0.0197588578,
0.0405219384,
-0.0150276367,
0.0075296322,
0.0105620585,
-0.0173917972,
0.0089019127,
-0.0390588753,
-0.055531051,
0.0656869039,
0.0194167309,
0.0265686978,
-0.0075010685,
0.0329767913,
-0.0046949266,
-0.0312771164,
-0.0194237884,
0.0046057501,
0.0183744747,
-0.0244331025,
-0.0023332159,
-0.0194273703,
0.007974756,
-0.0136912866,
0.0680431053,
-0.0166131388,
0.0154915769,
-0.0128911342,
0.0018908795,
-0.028197946,
-0.017546352,
-0.0290321261,
0.0038951722,
-0.0203893743,
-0.0206496492,
-0.0144707272,
-0.0452985577,
-0.0150447674,
0.0101418346,
2.838e-7,
0.0370535627,
0.0872011185,
-0.0383713804,
0.0114500262,
0.0205792021,
0.0219468679,
-0.019915482,
0.0106275612,
0.0045519965,
0.007267511,
0.0130305057,
0.0265779775,
0.0329281688,
0.024124125,
-0.0592097528,
-0.0429849327,
-0.0877108574,
-0.0569465347,
-0.0526922531,
0.0445398763,
0.0613323227,
0.0550063327,
-0.0161389355,
-0.0395213291,
0.0337075368,
-0.0702846497,
0.0410747603,
-0.0190050136,
-0.0126026683,
0.0092489514,
0.0362342335,
0.0156227062,
0.0522978567,
0.0755016059,
0.0288621988,
0.0118532069,
-0.0068838862,
0.0689135566,
-0.0365120433,
-0.026844522,
0.0200004447,
0.0523847118,
-0.0097461063,
-0.0171298012,
0.0830483958,
-0.0428720862,
-0.0749818683,
-0.0052416897,
-0.0108494963,
-0.0340176597,
0.094247289,
-0.041233018,
-0.0125973793,
0.0017104151,
0.0033501526,
0.0130639328,
-0.0143060181,
-0.0314087197,
0.0261885673,
0.0092939949,
-0.0773455948,
-0.0738279447,
0.006052095,
-0.0194071922,
-0.0165518019,
-0.0854482502,
-0.0007161637,
2.400464386e-34,
0.0372501463,
0.010288991,
0.0027002168,
0.0239789411,
0.0008291638,
0.0293851383,
0.0683730245,
0.0166345593,
-0.0247784238,
-0.0605496988,
0.0136514846
] |
http://arxiv.org/abs/2401.16788v1 | http://arxiv.org/pdf/2401.16788v1 | Can Large Language Models be Trusted for Evaluation? Scalable Meta-Evaluation of LLMs as Evaluators via Agent Debate | 2024-01-30T07:03:32 | 2024-01-30T07:03:32 | Steffi Chern, Ethan Chern, Graham Neubig, Pengfei Liu | cs.CL | cs.CL, cs.AI | Despite the utility of Large Language Models (LLMs) across a wide range of
tasks and scenarios, developing a method for reliably evaluating LLMs across
varied contexts continues to be challenging. Modern evaluation approaches often
use LLMs to assess responses generated by LLMs. However, the meta-evaluation
conducted to assess the effectiveness of these LLMs as evaluators is typically
constrained by the coverage of existing benchmarks or requires extensive human
annotation. This underscores the urgency of methods for scalable
meta-evaluation that can effectively, reliably, and efficiently evaluate the
performance of LLMs as evaluators across diverse tasks and scenarios,
particularly in potentially new, user-defined scenarios. To fill this gap, we
propose ScaleEval, an agent-debate-assisted meta-evaluation framework that
leverages the capabilities of multiple communicative LLM agents. This framework
supports multi-round discussions to assist human annotators in discerning the
most capable LLMs as evaluators, which significantly eases their workload in
cases that used to require large-scale annotations during meta-evaluation. We
release the code for our framework, which is publicly available at:
\url{https://github.com/GAIR-NLP/scaleeval}. | Can Large Language Models be Trusted for Evaluation? Scalable Meta-Evaluation of LLMs as Evaluators via Agent Debate: Despite the utility of Large Language Models (LLMs) across a wide range of
tasks and scenarios, developing a method for reliably evaluating LLMs across
varied contexts continues to be challenging. Modern evaluation approaches often
use LLMs to assess responses generated by LLMs. However, the meta-evaluation
conducted to assess the effectiveness of these LLMs as evaluators is typically
constrained by the coverage of existing benchmarks or requires extensive human
annotation. This underscores the urgency of methods for scalable
meta-evaluation that can effectively, reliably, and efficiently evaluate the
performance of LLMs as evaluators across diverse tasks and scenarios,
particularly in potentially new, user-defined scenarios. To fill this gap, we
propose ScaleEval, an agent-debate-assisted meta-evaluation framework that
leverages the capabilities of multiple communicative LLM agents. This framework
supports multi-round discussions to assist human annotators in discerning the
most capable LLMs as evaluators, which significantly eases their workload in
cases that used to require large-scale annotations during meta-evaluation. We
release the code for our framework, which is publicly available at:
\url{https://github.com/GAIR-NLP/scaleeval}. | [
0.0263663046,
0.0481446385,
-0.028069973,
0.0442139357,
-0.0370774344,
0.0029694978,
0.0241081491,
-0.0142908674,
-0.026495818,
-0.0707070529,
-0.0538763255,
-0.0039243535,
-0.0053958204,
0.0143776657,
0.02530526,
-0.032407891,
0.019794248,
-0.0545004085,
0.0120425764,
0.002129216,
0.0208082106,
-0.0230058562,
0.0009852361,
0.0558734126,
0.0270632431,
-0.0060228794,
-0.0127138188,
-0.0193451513,
0.0150508201,
-0.0333295874,
-0.0195008833,
0.0584465526,
0.0069297571,
0.0485834107,
0.0000020521,
-0.0512770861,
0.0053583039,
0.0337423198,
-0.0201527588,
0.0210340098,
0.0798616037,
0.0471651293,
-0.0339019708,
0.0086636124,
-0.0492233597,
-0.0059420262,
0.05218428,
-0.0527028032,
0.0696897209,
0.0538494922,
0.0098625002,
0.0110719781,
-0.016467385,
-0.0020827202,
-0.0164180938,
-0.0302173514,
-0.0114655541,
-0.0033644305,
0.0242914949,
0.0295401346,
0.00634772,
0.0297187679,
0.0319701545,
-0.0496852286,
-0.0064950241,
-0.0024385969,
-0.0638005286,
-0.0213566199,
-0.0332905576,
0.0568163283,
0.053228341,
-0.0085312203,
-0.0232545007,
-0.0132809766,
0.0223953426,
0.0503087789,
-0.0462201461,
0.0259325691,
0.0132035101,
-0.0060813618,
-0.0163895339,
0.0113460384,
0.0347800888,
0.0205742437,
-0.0148895821,
0.1092102826,
0.0000198173,
-0.00641408,
0.0149883656,
0.0075904173,
0.0008449062,
-0.061121475,
-0.0142197143,
0.0276214294,
-0.0562941246,
0.0227255486,
0.0221272856,
-0.0179026518,
0.0639185384,
-0.0610789768,
0.0637272745,
0.012209001,
0.0014058854,
0.0456056632,
0.029615337,
0.0561422929,
0.0137079684,
0.0100415042,
-0.005394883,
0.0014041672,
-0.0582891107,
0.0080575757,
0.0021656381,
0.0154090216,
-0.0113780629,
0.0205876417,
-0.013491624,
0.0250646602,
-0.0430626199,
0.0486098826,
-0.0059656301,
-0.0268290527,
-0.0057609687,
-0.0175606441,
-0.048573941,
0.033361014,
-0.0378752239,
0.0302864797,
0.0389099345,
-0.0753639787,
0.0188772753,
0.0387112573,
-0.0433708616,
-0.0252993722,
0.0080863917,
0.0747946277,
0.0208080839,
-0.0053544277,
0.0001990559,
-0.0126172397,
-0.0148048569,
-0.078240484,
-0.0146693094,
-0.1005299315,
-0.0074062846,
0.0045940536,
-0.0017592796,
-0.0229562446,
0.0060439166,
0.0113303913,
0.0326638035,
0.0776303411,
-0.025302941,
0.0064970385,
0.0199657809,
0.0200806689,
-0.0602912679,
0.0972993448,
0.0476623476,
0.0269048754,
0.039728649,
0.0009807409,
0.0197600722,
0.0490013622,
-0.0117812436,
0.0295877103,
0.0267820843,
0.0324382968,
-0.0110542187,
0.0204325579,
-0.0186406355,
0.0798082128,
-0.0183950905,
-0.0088540977,
0.0712514967,
0.070663251,
-0.0119239315,
-0.0162271708,
0.0479065329,
0.0136702741,
0.0205992442,
-0.016336076,
-0.0286705382,
0.0180041213,
-0.01130265,
-0.018872425,
0.0107138008,
0.02830429,
-0.0111411717,
-0.0683388039,
-0.0210615844,
0.0067078024,
0.0223778114,
0.0240584537,
0.0250024218,
0.0054741208,
0.004587322,
0.0249385703,
0.0058802268,
-0.0584280975,
-0.0110388966,
-0.0728288218,
0.0091996379,
0.0491610803,
0.0638402775,
-0.0242717117,
-0.0396857224,
-0.0180847347,
0.027516447,
0.07268776,
0.0775631219,
-0.0331207439,
-0.0303704534,
-0.0331995711,
-0.0217467509,
-0.0125732645,
0.0051676468,
0.0132567128,
0.0199279003,
-0.0089445589,
-0.0089258645,
-0.0721064955,
-0.0652132258,
0.0237663649,
-0.0185165983,
-0.0082414765,
-0.03235,
0.0040630396,
0.041114457,
0.027688114,
-0.0226012822,
0.0335677378,
0.0201371983,
-0.0116395671,
-0.0083955433,
0.0223075226,
0.0132787544,
0.0146849109,
0.0518688969,
-0.0371671692,
-0.0242881328,
0.0146766696,
-0.0219778027,
-0.0261839051,
0.015723126,
-0.0285753105,
0.0099521782,
-0.0267320517,
0.0251765009,
-0.0457858853,
-0.0192520414,
-0.0586208403,
0.0206291843,
0.0063561122,
-0.004232049,
0.0315554067,
0.0165420324,
0.0189373512,
-0.0078529501,
-0.0705741346,
0.0412984528,
0.0225080512,
0.0378179178,
0.0446353592,
0.0213694759,
-0.0541456752,
-0.0429656692,
-0.1110966355,
-0.0069303741,
-0.0316717066,
0.0443354361,
0.0050013764,
-0.0128789432,
0.0170338694,
0.0384930074,
-0.0410055779,
0.0044919336,
0.0260198228,
0.0452835113,
0.0365318358,
0.0079261046,
-0.0284480639,
-0.050283026,
0.0067825001,
-0.0149735045,
0.058855392,
0.0401506014,
0.0268480908,
-0.0293926839,
0.0132647753,
-0.037327528,
-0.0128617277,
-0.0063831778,
-0.0221129768,
0.0621895418,
0.0069991215,
-0.0134026408,
0.0188688394,
-0.0041183359,
0.0336473919,
0.0563383959,
-0.0157852005,
-0.088602528,
-0.0207740068,
0.0405313745,
-0.0174660422,
-0.0512556136,
-0.0194149595,
0.0202351417,
0.0622001365,
-0.0325471647,
-0.0023909509,
-0.0256098509,
0.0341342129,
-0.0521650016,
-0.0557139106,
-0.0153152756,
-0.0216774177,
-0.0365880355,
0.0710020885,
-0.0138324536,
-0.0523113571,
0.0789036676,
-0.0429254137,
-0.0170851499,
-0.0494917929,
-0.0193311665,
-0.0651625618,
0.0643968955,
-0.0368809327,
-0.0048635444,
-0.0207589213,
0.0022510986,
-0.0502176918,
-0.035279952,
0.0456621088,
-0.0481063239,
0.0660269633,
-0.0730384812,
0.0264565125,
-0.0093623335,
-0.034174338,
-0.0206061695,
0.001661845,
-0.0497059971,
0.0292103346,
0.0092841284,
-0.0062807747,
-0.0455822088,
0.0043622539,
0.0155783799,
0.0063120262,
0.004641199,
-0.0182354525,
-0.0102962349,
-0.0152859837,
0.0048056492,
0.041803319,
0.0388558842,
-0.0200764406,
0.0527122654,
0.067736201,
0.032609731,
-0.0247852206,
-0.0642014742,
-0.0253470223,
0.0250826776,
-0.0486860238,
0.017632721,
0.0202498082,
0.0506355874,
-0.0753048137,
0.0016965676,
0.0935780108,
0.0371725745,
-0.0589817651,
0.0030092357,
-0.0616777167,
0.0516732931,
0.0250652507,
0.0220203549,
-0.0766867846,
-0.015218717,
0.0058770627,
0.0107754264,
-0.0148305148,
-0.0117455106,
-0.0000378782,
-0.1138937324,
0.0552493706,
0.0037465082,
-0.008329683,
0.0307463631,
-0.0460770987,
0.0115860403,
-0.0306014344,
0.0521824583,
0.0259522852,
-0.0081828143,
0.0114219096,
0.0114994692,
-0.0331772454,
-0.0393852293,
0.0442268997,
-0.0520770028,
0.0978624299,
0.0593204461,
0.0208226517,
-0.0132867163,
-0.0357626379,
0.0724869967,
0.0191126745,
0.0037359404,
0.0032598767,
0.0334482566,
0.0019476899,
0.0038104383,
-0.0176370163,
-0.0278641004,
-0.0274581984,
-0.0318020396,
0.0414309166,
0.0593675785,
-0.0245837476,
0.0088168569,
-0.008814293,
0.0042220419,
-0.005312182,
0.0590447597,
0.0571063682,
-0.0455785766,
-0.0515932292,
-0.0441845804,
-0.0453731157,
-0.0410046428,
-0.0560243241,
0.030727718,
0.0260638613,
0.005376413,
-0.0499191433,
-0.0757791921,
0.0309723299,
0.0135407597,
0.0124565577,
0.0084767751,
0.0143054696,
0.0152362604,
-0.0078630848,
0.0204307549,
-0.0335648842,
0.0438428111,
-0.04188025,
0.0055174655,
0.0307355709,
0.003834439,
-0.0008198616,
0.0398526527,
-0.067704469,
-0.0809839368,
-0.0149682835,
0.0609079637,
0.033988554,
0.0078409715,
-0.0131567204,
-0.0705006868,
0.0324264802,
0.0168204941,
-0.0361184813,
0.0045034313,
0.0377560817,
0.0144658592,
-0.0155936368,
0.0284823943,
-0.0238843672,
0.004029782,
-0.0002678157,
0.0414946824,
-0.0245647449,
-0.0397217274,
0.0076249945,
0.0269160178,
-0.0015099238,
-0.0281022992,
-0.0147567783,
-0.0068465956,
-0.0055379192,
-0.023934003,
-0.0309300106,
-0.0455474593,
-0.0054589827,
-0.0225163288,
-0.0332859568,
0.030437069,
0.0257916972,
-0.0333661325,
0.0453763157,
0.0064795339,
0.0751452595,
-0.0022791757,
0.019067347,
-0.0049016136,
0.0140531957,
-0.0130686164,
0.0007970206,
-0.0222440157,
0.0412884504,
0.0128094256,
0.0289292932,
0.0003905531,
-0.005271561,
0.0260379799,
0.0533225462,
-0.0603746399,
-0.0503795817,
0.0129197175,
0.0295431949,
0.026347531,
0.0714063272,
0.0605468191,
-0.0192900244,
0.0005824009,
-0.0242291111,
-0.0092616705,
-0.0000080171,
-0.018547006,
0.0207700748,
0.0269289967,
-0.0248456374,
-0.021968659,
0.0611118227,
-0.0143439481,
0.0031272289,
-0.0097881323,
-0.0040635876,
-0.0487844832,
-0.0696422458,
-0.0532965548,
-0.0075825187,
-0.0237883385,
-0.034171816,
-0.0860687047,
-0.0631892905,
-0.0165453292,
0.0547582172,
-0.0052930145,
-0.0228560008,
0.0132812057,
0.0613402352,
0.0184149668,
-0.0008502941,
-0.0645127445,
-0.0061224201,
-0.0158024505,
0.0049730684,
-0.0123210736,
-6.055742819e-33,
-0.0382062905,
-0.0441588387,
0.0637772381,
-0.0017222613,
-0.0103544118,
0.0023024892,
-0.0375512429,
0.0337010808,
-0.0117056696,
0.0225236882,
-0.0132766599,
-0.0025627441,
0.0136331199,
-0.0108848475,
0.0322705247,
-0.0347058885,
0.033691559,
0.0015182613,
-0.047076337,
0.0016723453,
-0.006962195,
-0.0410511605,
0.0113861775,
-0.0486280322,
0.0720744506,
0.026730977,
-0.0272108037,
0.0046532792,
-0.0300010908,
0.0836603492,
0.0108517464,
0.0502124652,
0.0197261106,
-0.0419073403,
-0.01942073,
-0.0057252082,
-0.0314376913,
-0.0191193018,
0.0426578261,
0.0043804571,
-0.0295358673,
-0.0351836979,
0.037994571,
-0.0029417234,
-0.0418733731,
0.0072372002,
0.0185117181,
-0.0289706867,
-0.0008171991,
0.0170734879,
-0.0580572486,
0.0555946268,
-0.0149170356,
0.0203786474,
0.0208810922,
0.0183080304,
-0.0074302675,
0.0120054046,
-0.0488908216,
0.0427690633,
0.0066072457,
0.0103349118,
0.0929475576,
0.0163180865,
-0.0155298766,
0.0724668726,
-0.0095200036,
0.0499495976,
-0.0004731807,
-0.0309813656,
0.0146353533,
0.0498122238,
0.0559191443,
0.0911012441,
-0.0306023061,
0.0290431846,
-0.0306638163,
0.053426709,
0.0461970232,
0.0360794812,
0.0111488039,
0.0154846469,
-0.0408318415,
-0.0101734772,
-0.0796909481,
0.0039637689,
-0.0456168316,
-0.0500273705,
0.0382558294,
-0.022876475,
-0.0298742726,
-0.0032479039,
0.0364237018,
-0.0619631447,
-0.0174306706,
-0.0178667121,
-0.0204840135,
0.0387110561,
0.0157220028,
0.0133539289,
-0.0408132933,
0.0128252078,
0.0003002554,
0.0543575324,
0.0419540703,
-0.0074291239,
-0.0059745098,
0.0123972781,
-0.0544079095,
0.0062018358,
-0.0561439767,
0.0055808839,
-0.0089120381,
-0.0115898317,
-0.0035848496,
-0.0102217253,
-0.0130630108,
0.0445802882,
-0.0073856651,
-0.0214869883,
-0.0177628454,
0.0173934959,
-0.0148005439,
-0.0101912692,
-0.0523701645,
-0.0024086528,
-0.0371814482,
0.0196640808,
0.0590443276,
-0.0099011511,
-0.0071910564,
-0.0277806409,
2.68e-7,
0.0657183453,
0.0393439382,
0.0003286618,
0.0132923536,
0.0205169301,
0.0303124618,
-0.0324922875,
0.0045197411,
0.0219230372,
-0.0317968018,
0.011238426,
0.0380012281,
0.0605115555,
0.0612205416,
-0.0392429456,
-0.0219535101,
-0.0563172549,
-0.0076356875,
-0.0441946276,
-0.0145581523,
0.0183399692,
0.0359204523,
-0.0010616623,
0.0050850264,
-0.0405459292,
-0.053060703,
0.0417317078,
-0.0548075922,
-0.0199085623,
0.0531630255,
0.007740201,
-0.0109916255,
0.0239968728,
0.0155212004,
0.005973042,
0.0071044583,
-0.0112773236,
0.072214894,
0.0128769279,
-0.0329778083,
-0.0076205833,
-0.0541261584,
0.0117522962,
-0.0495848171,
0.0435323045,
-0.0418792441,
-0.052527871,
-0.0531222411,
-0.0297654252,
-0.0253359638,
0.0320824496,
-0.0199646186,
-0.0103905797,
0.0275893435,
-0.0142255966,
-0.0157104265,
-0.018281851,
-0.0415141545,
-0.0414682962,
-0.0293482486,
-0.0318496302,
-0.0984952301,
-0.0275178663,
-0.0072081545,
-0.0375526994,
-0.0624359921,
0.0179231633,
2.260033219e-34,
0.0515755564,
-0.0050675981,
0.0142042059,
0.0412838645,
0.0436494686,
0.0367699303,
-0.0732793435,
-0.0027964974,
0.0288203899,
-0.071610868,
0.0051868367
] |
http://arxiv.org/abs/2402.01750v1 | http://arxiv.org/pdf/2402.01750v1 | PACE: A Pragmatic Agent for Enhancing Communication Efficiency Using Large Language Models | 2024-01-30T06:55:17 | 2024-01-30T06:55:17 | Jiaxuan Li, Minxi Yang, Dahua Gao, Wenlong Xu, Guangming Shi | cs.CL | cs.CL, cs.AI | Current communication technologies face limitations in terms of theoretical
capacity, spectrum availability, and power resources. Pragmatic communication,
leveraging terminal intelligence for selective data transmission, offers
resource conservation. Existing research lacks universal intention resolution
tools, limiting applicability to specific tasks. This paper proposes an image
pragmatic communication framework based on a Pragmatic Agent for Communication
Efficiency (PACE) using Large Language Models (LLM). In this framework, PACE
sequentially performs semantic perception, intention resolution, and
intention-oriented coding. To ensure the effective utilization of LLM in
communication, a knowledge base is designed to supplement the necessary
knowledge, dedicated prompts are introduced to facilitate understanding of
pragmatic communication scenarios and task requirements, and a chain of thought
is designed to assist in making reasonable trade-offs between transmission
efficiency and cost. For experimental validation, this paper constructs an
image pragmatic communication dataset along with corresponding evaluation
standards. Simulation results indicate that the proposed method outperforms
traditional and non-LLM-based pragmatic communication in terms of transmission
efficiency. | PACE: A Pragmatic Agent for Enhancing Communication Efficiency Using Large Language Models: Current communication technologies face limitations in terms of theoretical
capacity, spectrum availability, and power resources. Pragmatic communication,
leveraging terminal intelligence for selective data transmission, offers
resource conservation. Existing research lacks universal intention resolution
tools, limiting applicability to specific tasks. This paper proposes an image
pragmatic communication framework based on a Pragmatic Agent for Communication
Efficiency (PACE) using Large Language Models (LLM). In this framework, PACE
sequentially performs semantic perception, intention resolution, and
intention-oriented coding. To ensure the effective utilization of LLM in
communication, a knowledge base is designed to supplement the necessary
knowledge, dedicated prompts are introduced to facilitate understanding of
pragmatic communication scenarios and task requirements, and a chain of thought
is designed to assist in making reasonable trade-offs between transmission
efficiency and cost. For experimental validation, this paper constructs an
image pragmatic communication dataset along with corresponding evaluation
standards. Simulation results indicate that the proposed method outperforms
traditional and non-LLM-based pragmatic communication in terms of transmission
efficiency. | [
0.0632160157,
-0.0088788755,
-0.0160782784,
0.0110915778,
-0.0543784983,
0.0088573573,
-0.0152098201,
0.0152937844,
-0.0388952307,
-0.0284180511,
0.0023574068,
-0.0091774594,
-0.0365600847,
0.0072513167,
0.0054633236,
-0.0618628487,
0.034576565,
-0.0338945203,
0.079435803,
0.0133169703,
0.0385118611,
-0.0014942894,
0.0104020033,
0.0275302269,
-0.0493956916,
-0.02678244,
-0.0672147796,
0.0503885224,
-0.0184187628,
-0.033396706,
-0.0189636461,
0.0319347903,
0.0491640605,
0.0024758405,
0.0000021733,
-0.0429785326,
0.0023479238,
-0.0277734529,
-0.0607433878,
0.0121256718,
0.0865644738,
-0.0404909886,
-0.0291879177,
0.0112036709,
-0.0268096849,
0.0028720016,
0.068450436,
0.051902853,
0.0373267792,
0.0622888282,
-0.0202692226,
-0.0021812851,
0.0247484986,
-0.0046988102,
0.0002771569,
-0.0425210483,
0.0220772736,
0.0326221585,
-0.0029090613,
0.0245749746,
-0.0101571735,
0.0363861807,
0.0281997286,
-0.0313758105,
-0.0008508364,
0.021285085,
0.015043918,
-0.0503226221,
-0.0541805103,
0.0267176237,
0.0475654788,
0.0004554025,
0.015130573,
0.0000218871,
-0.0244938564,
-0.0169744045,
-0.0572284125,
0.0079237418,
0.00299462,
0.0165476892,
-0.0311463326,
0.0247336123,
-0.016348606,
0.0194890574,
-0.0388889387,
0.0549309961,
-0.006167328,
-0.0087158727,
-0.0163974073,
-0.0116743166,
0.044771269,
-0.0743652508,
-0.0079403054,
0.0547308326,
-0.0162209012,
0.0428921916,
-0.0268139802,
-0.0007656649,
0.007372519,
-0.0317162387,
0.0126407295,
0.0338577516,
0.0318160579,
0.0405796506,
0.018840123,
0.0163714979,
0.0095120883,
0.0279813204,
-0.0021313184,
0.0067829569,
-0.036084041,
-0.0333928317,
-0.0706829727,
0.0362769924,
-0.0494925901,
-0.0543502942,
-0.0332474783,
0.0573124029,
-0.0157134887,
0.0047472399,
0.0004743059,
-0.0295506213,
-0.0478698276,
-0.0187858306,
-0.0433988348,
0.0574404784,
-0.0416903086,
0.0236792229,
-0.013726023,
-0.0691809654,
0.0354006551,
0.0342790112,
-0.0008554189,
-0.0032167388,
-0.0292729475,
0.1053283513,
-0.0182754062,
-0.0340297855,
-0.0303162746,
-0.0316816755,
0.0448475629,
-0.0577776209,
0.0067593278,
-0.0554682575,
0.0100083342,
0.0059422948,
-0.0183037426,
-0.0264483299,
0.0013602737,
-0.0013376746,
-0.0132189626,
0.0577216819,
-0.0425404385,
0.0506872423,
0.0361041017,
0.0272400379,
0.0106322588,
0.0894992575,
0.0093539059,
0.0307154246,
0.045039542,
0.005997736,
-0.0222063214,
0.0197698008,
0.020389434,
0.0035926246,
0.0073404987,
-0.0104922708,
-0.0055642398,
0.0083142286,
0.0518291779,
0.0590511486,
-0.0218862463,
0.0064062276,
0.0462951064,
0.0328602083,
0.019527765,
-0.0332173333,
0.0281076301,
0.0849678889,
-0.0179434139,
0.0051363003,
-0.00730641,
0.0655816197,
-0.0104180593,
-0.0058313683,
-0.0133564258,
0.0078770351,
-0.0637674704,
-0.030965792,
0.0128468946,
-0.0410321876,
0.0386489816,
-0.0067203799,
0.0629210249,
-0.0262741353,
-0.0273827128,
-0.0003375408,
-0.0065846457,
-0.018330168,
-0.027408151,
-0.0302505195,
0.0012844333,
0.0913815275,
0.0110472497,
0.0457212329,
0.0129293716,
0.0233536549,
0.0415655747,
0.0256206915,
-0.0058081681,
-0.024425501,
-0.0050817141,
-0.0477027446,
0.0008959589,
0.0143075166,
-0.0194791313,
0.0166865829,
-0.0092621995,
0.0359253772,
-0.0191460531,
0.0384684093,
0.0093412045,
0.0206520408,
0.0393716209,
0.0305247381,
-0.0345732346,
0.0198918581,
0.0156950988,
0.0137277478,
-0.0382669903,
-0.0129975248,
0.0243871883,
-0.008393893,
0.0276130587,
0.0402795672,
-0.0070516495,
0.0128392037,
0.0086081484,
-0.0332402401,
-0.0056632417,
-0.0055287792,
0.0150451772,
-0.0301557407,
0.0204864386,
-0.0679833665,
0.0585607551,
-0.0778260306,
0.0386693664,
-0.0047677173,
0.0002580272,
0.0131826196,
0.0371968783,
0.0251457561,
-0.0540292002,
0.0489997901,
-0.0032091017,
0.0201908778,
-0.0500926003,
0.0031818033,
0.0287144203,
0.0000615076,
0.0140931215,
0.0472041368,
0.0669959635,
-0.0320823081,
-0.008783469,
-0.1130319983,
-0.0638516098,
-0.0185817666,
-0.0154080894,
0.0212533753,
0.0076013044,
0.0031002678,
-0.0152627164,
-0.0065495358,
0.0697630569,
0.0074029882,
0.0300309602,
-0.0000690671,
0.0245408434,
-0.0711449757,
0.0004584052,
-0.0006921641,
0.0324894637,
-0.0022965239,
0.0278075058,
-0.0396390185,
-0.0677213967,
0.0341860168,
-0.0158192683,
-0.0100010205,
-0.0025524576,
-0.0902336389,
0.013422559,
-0.0297313351,
-0.024492899,
0.0216164123,
-0.0004217203,
0.0426732339,
0.0025291617,
-0.025406342,
-0.0636155158,
-0.0036597773,
-0.0170496572,
0.0077918386,
-0.0513566993,
-0.0711245537,
-0.0034440982,
0.0182913654,
-0.0159950405,
0.0703118667,
-0.0721608773,
-0.0195969772,
-0.0152659696,
-0.0861329362,
-0.072105445,
-0.0001002345,
-0.006809656,
0.0606996715,
-0.0007064639,
-0.04507697,
0.0526590794,
-0.0149892531,
-0.0368531346,
0.0258976836,
-0.0161339808,
0.0017999844,
0.0740161538,
-0.0051081437,
0.0575605147,
-0.0276121367,
-0.0129696699,
-0.0296504442,
-0.0443026461,
-0.0224980917,
0.0041352571,
0.03131897,
-0.0862431005,
0.0326425284,
0.0045883423,
-0.0071132034,
-0.0117287626,
-0.0308780037,
-0.0286435466,
-0.0090600848,
-0.0223496649,
-0.0180556476,
0.0236962922,
0.0482688285,
0.0345911533,
0.0081524216,
0.0365795195,
-0.0247608609,
-0.0233348608,
-0.0353981741,
0.0310922675,
-0.0011122397,
-0.0343986303,
0.0121411616,
0.033245109,
0.0450450704,
-0.0004065652,
0.0072192745,
-0.0058852406,
0.0135060716,
0.0309356283,
0.0197077002,
-0.0168390609,
0.0625970736,
0.0835907012,
-0.0056281583,
-0.0077684554,
0.0842629373,
0.0209356546,
-0.0542119741,
-0.005625423,
-0.0304245912,
0.0149032855,
-0.0151357781,
0.0365717709,
-0.0430660583,
-0.0231700074,
-0.0191641115,
-0.0371920578,
-0.0481743813,
-0.001273252,
-0.0108200861,
-0.0596501268,
0.0483463593,
0.0066388762,
-0.0082294727,
-0.0027625854,
-0.006567562,
0.0015776873,
0.0555088483,
0.0046312683,
-0.0119314119,
0.0101433014,
-0.0458208211,
0.0001864918,
-0.030798506,
-0.0550879166,
0.0145818209,
-0.0260365568,
0.1059305221,
0.0675218478,
0.0214571171,
-0.0096928999,
-0.0686700419,
0.001658651,
0.0207317341,
0.0248599406,
-0.0162770711,
-0.0309876464,
0.0115941772,
-0.0012028533,
-0.0134384604,
-0.0142095909,
-0.0341384634,
-0.0123179322,
-0.0034290703,
0.0621436872,
-0.0537380241,
-0.0033831659,
-0.0313373879,
0.0275923572,
0.0574235022,
0.0501254089,
0.0094625335,
-0.0336966887,
-0.0552901551,
-0.0170847327,
-0.030505823,
-0.0004899437,
-0.0042849784,
0.0303952415,
-0.0292519759,
-0.0779118463,
-0.0406091996,
-0.0478056148,
0.03571219,
0.0328177251,
0.01816836,
0.0215491205,
0.0502339043,
0.0460228324,
-0.0482563823,
-0.063850753,
-0.0085675539,
0.0339079984,
-0.0673412532,
-0.0218295287,
0.0016143238,
-0.0242961068,
0.0187825225,
-0.0085220682,
-0.0397802293,
-0.0430896431,
0.0073354896,
0.0279061068,
0.0321825929,
0.0093453145,
-0.0276615098,
-0.0834276825,
-0.0542834625,
0.0115609867,
-0.0291979183,
0.0849272609,
0.0062352684,
0.0178790167,
0.0130386464,
0.0461388528,
-0.0375660434,
0.0414529443,
-0.0402837843,
0.0280839019,
-0.0682780072,
-0.0268972851,
0.0124055352,
0.0331253149,
-0.0385856442,
0.0091403956,
-0.0645803809,
0.031566076,
0.0088026505,
-0.0489490367,
0.0294746514,
-0.0370937474,
-0.0307102539,
-0.0512042269,
0.0383221097,
0.0462695435,
-0.0021268877,
-0.0354213752,
-0.0606065355,
0.0707837641,
0.0007128952,
-0.0210346673,
0.0258746818,
0.0000883605,
0.0212480854,
0.0072069759,
-0.0017412787,
-0.0123030227,
-0.0003085369,
0.0064012432,
0.0718022138,
0.0460470244,
-0.0239457674,
-0.0134121021,
0.0231490582,
0.0222509056,
-0.059488833,
0.0696818382,
0.0850010142,
0.0686289594,
0.063343443,
-0.0130944299,
-0.101479739,
0.0155096892,
-0.0338095613,
0.0378722511,
0.0157221984,
-0.05592186,
0.0572663397,
0.0185617544,
-0.0097136777,
-0.0275236368,
-0.003839822,
-0.0343192108,
0.004359833,
-0.0321880803,
-0.0147648463,
-0.0137186553,
-0.0022372305,
-0.0127609512,
-0.0012482029,
0.0031492962,
0.0389271826,
-0.0600868501,
-0.0395951085,
-0.0265380796,
0.020247113,
-0.0087423222,
-0.0750707388,
0.0701820254,
0.0389096625,
0.0038585227,
-0.0108346678,
-0.045881547,
-0.0363665596,
-0.0743093193,
0.0154965324,
-0.0295872875,
-6.682088821e-33,
-0.0132294586,
-0.0481351018,
0.0572480932,
-0.0052891192,
-0.0080193374,
-0.0058154445,
-0.0438975133,
-0.0181010235,
-0.0078091463,
0.0273690373,
-0.0193628818,
0.0327075981,
0.019824367,
0.007587729,
-0.0132916709,
0.0210819971,
0.0332079232,
-0.0076760966,
0.0028721695,
0.029987203,
0.0412501059,
0.0120909372,
-0.0094418293,
-0.0720819235,
-0.0114782341,
0.0193174612,
-0.0162166338,
0.0132107018,
0.0524593294,
0.0406814963,
-0.0108100818,
0.0297415704,
0.0292589962,
0.0163087379,
-0.0108097773,
0.0750464424,
-0.0154779823,
-0.0564837717,
0.048096139,
0.031510625,
-0.0591795817,
-0.0231274404,
-0.0140146213,
-0.0140661607,
-0.0624774173,
0.0567355305,
-0.002138939,
0.0041566836,
-0.0300010424,
0.0419391841,
-0.054775551,
-0.0035757078,
0.0158059746,
-0.0172322318,
0.0212042984,
-0.0313404165,
-0.0091238003,
-0.0290957652,
-0.0534015708,
0.0448173098,
-0.0093696667,
-0.0054842196,
0.0502879471,
0.0119218705,
0.0139529901,
0.0502718017,
0.0177487507,
-0.0234357677,
-0.002864342,
0.0163250677,
0.0051675988,
0.0201015882,
-0.0297886673,
0.0502913222,
-0.0262183342,
-0.0076339054,
-0.0327397399,
-0.0036423306,
0.0646593571,
0.05067157,
0.0356312543,
0.0170334112,
-0.0587473214,
-0.0054904041,
-0.0027076229,
0.0465424694,
-0.0243599117,
-0.0777081922,
0.0297789685,
0.0138053149,
0.025363626,
-0.0197697952,
0.051310882,
-0.026628904,
-0.0155378841,
-0.0001942366,
-0.0092748348,
0.057495635,
0.0227601491,
0.0195150096,
-0.0323416144,
-0.0022470476,
0.0316268466,
0.015761517,
0.0176518671,
0.0248447284,
-0.0138473008,
0.0400758013,
0.0081084901,
0.0005037541,
-0.0231251512,
0.0327547826,
0.0082704108,
-0.0688392222,
0.0150853302,
-0.0020798042,
-0.0438238382,
0.0421608798,
-0.0323246643,
-0.0915180966,
-0.0213693772,
-0.0180410333,
0.0095050512,
0.0229377914,
-0.053572651,
0.0167647824,
-0.0669522062,
0.0189150311,
0.0293899029,
-0.0341591872,
-0.0160841085,
0.0552316383,
3.1e-7,
-0.0291065648,
-0.0144256121,
-0.0111690778,
0.066846773,
0.0269571487,
0.0285611637,
-0.0370379798,
0.0309176017,
-0.038750995,
-0.0184595063,
0.0310066752,
0.0178138353,
0.0264635552,
-0.0175277852,
-0.0831009224,
0.0147052156,
0.0108877905,
-0.0015660777,
-0.0456023477,
0.030512847,
0.0920222104,
0.0547053404,
-0.0164612662,
-0.0007669131,
0.0385741852,
0.0012262028,
0.0450096875,
-0.0647454783,
0.0255114269,
-0.0041125105,
-0.0142048625,
0.0034708404,
-0.0105757965,
0.0380219854,
0.0322521478,
-0.033326745,
-0.0285345297,
0.0948888212,
-0.0251440462,
-0.0208819248,
-0.0029915445,
-0.0041338434,
-0.0277467314,
-0.0172432866,
0.0116353333,
-0.0519380718,
-0.0446759127,
0.0704421476,
0.0243929084,
-0.0186137483,
-0.0589504838,
-0.0077531827,
-0.0262355525,
-0.0144688934,
0.0106836362,
-0.0463209897,
-0.0046719401,
-0.0678748488,
0.0744097605,
-0.0256139748,
0.0114498753,
-0.0614771023,
-0.0758820176,
-0.0023898096,
0.0555328876,
-0.0522969775,
0.0004027403,
2.432636657e-34,
0.0148359453,
0.0161477625,
-0.0121145546,
0.0467186943,
-0.00109686,
0.0026903972,
0.012681257,
0.0617984124,
0.0154872555,
-0.0599070005,
0.0093070762
] |
http://arxiv.org/abs/2401.16748v1 | http://arxiv.org/pdf/2401.16748v1 | Detecting Racist Text in Bengali: An Ensemble Deep Learning Framework | 2024-01-30T04:56:55 | 2024-01-30T04:56:55 | S. S. Saruar, Nusrat, Sadia | cs.CL | cs.CL | Racism is an alarming phenomenon in our country as well as all over the
world. Every day we have come across some racist comments in our daily life and
virtual life. Though we can eradicate this racism from virtual life (such as
Social Media). In this paper, we have tried to detect those racist comments
with NLP and deep learning techniques. We have built a novel dataset in the
Bengali Language. Further, we annotated the dataset and conducted data label
validation. After extensive utilization of deep learning methodologies, we have
successfully achieved text detection with an impressive accuracy rate of
87.94\% using the Ensemble approach. We have applied RNN and LSTM models using
BERT Embeddings. However, the MCNN-LSTM model performed highest among all those
models. Lastly, the Ensemble approach has been followed to combine all the
model results to increase overall performance. | Detecting Racist Text in Bengali: An Ensemble Deep Learning Framework: Racism is an alarming phenomenon in our country as well as all over the
world. Every day we have come across some racist comments in our daily life and
virtual life. Though we can eradicate this racism from virtual life (such as
Social Media). In this paper, we have tried to detect those racist comments
with NLP and deep learning techniques. We have built a novel dataset in the
Bengali Language. Further, we annotated the dataset and conducted data label
validation. After extensive utilization of deep learning methodologies, we have
successfully achieved text detection with an impressive accuracy rate of
87.94\% using the Ensemble approach. We have applied RNN and LSTM models using
BERT Embeddings. However, the MCNN-LSTM model performed highest among all those
models. Lastly, the Ensemble approach has been followed to combine all the
model results to increase overall performance. | [
0.0593302175,
0.0286098942,
0.0097437771,
0.0598503351,
-0.0132240439,
0.042694632,
0.0093406038,
0.0008519561,
0.0123552121,
-0.0134257572,
-0.0053295544,
-0.0680910349,
-0.0021429677,
0.007934425,
-0.0453728549,
-0.0281667039,
0.0271921903,
0.0612492077,
0.0177837275,
-0.0010044125,
0.057744287,
0.0173777454,
-0.0163651519,
0.0284705572,
-0.0485629775,
0.0065858327,
0.0259554069,
0.0178090855,
0.0162922479,
-0.0313032456,
0.0304266997,
0.0535212755,
-0.0040473109,
0.0839456767,
0.0000019293,
-0.0362587608,
0.0042420737,
-0.0083741611,
-0.0051550814,
-0.0262082815,
0.0298961755,
0.0131324492,
-0.0347906761,
0.0312854424,
-0.0345215201,
0.0159327351,
0.0690323114,
0.0951720327,
0.0569310598,
0.0940566435,
0.0045542209,
-0.0371607617,
0.0088937311,
0.0125328889,
0.011234818,
0.0239602271,
0.0434188358,
-0.0585730933,
-0.0065004542,
-0.0370903984,
-0.0126250973,
0.0243757628,
0.0470193848,
-0.031584274,
0.0424824283,
-0.0331179313,
-0.02574596,
-0.0162520055,
0.0209016223,
0.0619456694,
-0.0775062889,
0.0244147126,
-0.004512364,
0.0008294227,
-0.024051806,
-0.0687829554,
-0.0025670296,
0.0324288532,
-0.0379683673,
0.0474781469,
-0.0224601459,
0.0142318094,
0.0502741113,
0.0652051792,
0.0227826294,
0.0231693313,
-0.0298281386,
-0.0456635989,
-0.0304989051,
-0.0025173463,
0.0116278818,
0.0122102015,
0.0631636754,
0.0139651028,
0.0066078594,
0.0089351423,
0.000970615,
-0.0716270357,
0.0130565036,
-0.0670743287,
0.0229707211,
0.0375553742,
-0.0544972718,
0.0248661581,
-0.0436515585,
0.0168554913,
-0.0399555899,
-0.0037755396,
0.012656359,
-0.0107374806,
-0.0115970019,
-0.005064066,
0.0153900366,
0.0378441922,
-0.0273811333,
-0.0034290866,
-0.0789274201,
0.001869456,
-0.0189329647,
-0.0074717803,
0.0342877544,
0.026460899,
-0.0268656909,
-0.0114012631,
-0.0003230578,
-0.0044964137,
-0.0248826463,
-0.0071940254,
-0.0118338298,
-0.1000957489,
0.0063999533,
0.0484050885,
0.0004989045,
-0.0123084709,
0.0163638145,
0.0380131304,
0.0175986532,
-0.0377386026,
-0.0455337241,
0.0245655607,
-0.0323136076,
-0.031383656,
0.0078592375,
-0.0791982114,
0.0279617216,
-0.0376778208,
-0.0326117501,
0.0109512815,
0.0106631648,
0.0374261513,
-0.0040882342,
0.0935073346,
-0.030778924,
-0.0408824235,
0.0678380653,
0.0295847282,
0.0377449952,
0.0638010353,
0.0240647234,
0.0548590049,
0.0249697231,
0.0091302209,
0.0046157534,
0.0314386562,
-0.015636934,
-0.0169456936,
-0.0202012155,
-0.0020902201,
-0.0565738939,
0.0108393272,
-0.0122590242,
0.0138908392,
-0.0019982283,
-0.0122828372,
0.0131226201,
0.0348957889,
-0.0086986274,
0.0625052825,
0.0221428145,
0.0099941473,
-0.0125119593,
-0.0056826505,
-0.0282027423,
0.0052093402,
-0.029554503,
-0.0539362319,
0.0266975444,
0.0566431358,
-0.0345217399,
-0.0177831817,
0.0134718567,
-0.0315128975,
0.0239157621,
-0.0273780916,
0.0566995144,
-0.0031936152,
-0.0329890475,
-0.0099253561,
0.0488074906,
-0.0319943801,
-0.0048540398,
-0.0391784944,
0.0476029776,
0.0444505177,
-0.0044024605,
-0.0417278484,
0.0179507621,
-0.0442104377,
-0.0137213385,
0.0898466259,
-0.0046528303,
0.0345162936,
0.0362292379,
-0.0705213174,
0.0026269476,
0.051958736,
0.0003321361,
0.0507215336,
0.0258051362,
0.0389317386,
0.0148663865,
0.0531360097,
-0.0296406522,
0.0374878533,
0.0216376893,
-0.0048390669,
0.011683072,
0.0269724932,
0.054088451,
0.0139922816,
-0.0232614018,
-0.0514495075,
0.0130682457,
-0.0427835584,
-0.0771754757,
0.0007690174,
-0.0834969431,
-0.0829625055,
0.0074159862,
-0.0648009181,
-0.0098904381,
0.052522257,
-0.01449899,
-0.0025722268,
0.0383252092,
-0.0347915962,
-0.0135475546,
-0.0315937437,
-0.0059101158,
-0.0369723104,
0.053137958,
0.0109071117,
0.0466931723,
-0.0250145886,
-0.0127158882,
-0.0021680046,
-0.0023400439,
0.0736850351,
-0.0142168691,
-0.0549107492,
0.0134291565,
0.0042748726,
0.002780718,
0.0143768797,
0.044878494,
-0.0107362969,
0.0097610662,
-0.1048289537,
0.0095483558,
-0.0244777352,
-0.0332519822,
0.0066279368,
0.0360834785,
-0.008127667,
-0.022174431,
-0.0089752274,
-0.0117303859,
0.0638840795,
-0.020335976,
0.0190649033,
-0.0286507979,
-0.0090430612,
-0.0112512307,
-0.0271630213,
-0.0236748308,
0.0545619987,
0.0068697804,
-0.0901237652,
0.0248748176,
-0.0725047961,
-0.0087838201,
0.0333560742,
0.0114638461,
-0.0760241151,
0.0042253602,
-0.0048441556,
-0.0295770168,
0.0700446293,
0.0428771749,
0.036602553,
-0.016150726,
0.0205379147,
-0.0053906832,
0.0091621717,
-0.0107320268,
0.0897166282,
0.0229812842,
-0.0388309583,
-0.0180353504,
0.0460884422,
0.0069422065,
0.0273083858,
-0.0347019583,
-0.0286335051,
-0.0142300017,
-0.0395537764,
-0.0349338539,
-0.0429705605,
-0.0400817245,
-0.0177451801,
-0.0393641591,
-0.0339289978,
0.0501595624,
0.0150731523,
0.0073513645,
-0.0600855313,
-0.00665246,
-0.0314881466,
0.0259510707,
0.0070526297,
0.023120366,
-0.0639718622,
-0.0247925576,
0.0171117391,
-0.0295677017,
0.0073924186,
-0.0203737728,
-0.0215658005,
-0.0125733837,
0.0165710915,
-0.0279991943,
-0.0157740712,
-0.0700421333,
-0.0382858068,
-0.0234496351,
-0.0143315224,
0.0166454334,
-0.0264759846,
-0.0243353806,
0.0235954151,
0.0192040596,
0.0080538243,
-0.0220566001,
0.0241251141,
-0.0054084174,
0.0116455052,
-0.0143509628,
-0.0309228748,
0.0312034767,
-0.0139250122,
0.0142703243,
0.0161921065,
-0.0211453307,
0.0174250081,
-0.0745089352,
-0.0141245434,
-0.0064139762,
-0.0008531504,
0.0294393692,
-0.0385676287,
0.0301905423,
-0.0262976121,
-0.0079146372,
0.0916383415,
-0.0041694865,
-0.0275350362,
-0.0329625346,
-0.037665002,
-0.0268980749,
-0.0297552329,
-0.0148416562,
-0.051602155,
-0.030189611,
-0.0093070893,
-0.0356755517,
-0.0269105341,
-0.0413471647,
-0.0064581479,
-0.0491097718,
-0.0296689346,
-0.022473909,
-0.0483637713,
-0.058497116,
0.0045311479,
0.0074357525,
0.0118830521,
0.0704682395,
-0.0275283083,
0.0218524355,
-0.0322468914,
0.0412808508,
-0.0828853995,
-0.0154397227,
-0.004914403,
-0.0690598637,
0.0047097108,
0.088046141,
0.0492931865,
-0.0463342108,
0.011991201,
0.0030169557,
-0.0126145072,
-0.023998389,
-0.0154961571,
0.0367627405,
-0.0266354587,
-0.040674217,
0.0017215402,
-0.0135793909,
-0.0335615501,
0.0209461618,
0.0002718923,
0.0358544588,
0.0092806509,
-0.0287831333,
-0.0491886735,
0.0179544073,
0.0127601279,
0.0584455356,
0.0506072044,
-0.0713787973,
-0.0086271735,
-0.0351974368,
-0.0498590842,
-0.0936411768,
-0.0077293413,
0.0339289047,
-0.0364896804,
-0.0431934707,
-0.0377923883,
-0.0548605137,
0.0130792707,
0.0001178381,
0.0138220955,
0.0143150641,
0.028846655,
0.0020067978,
-0.0408656262,
0.0010953103,
-0.0255321376,
0.0430561937,
-0.0287055001,
0.0345626138,
-0.0144284833,
-0.0159524195,
-0.0144706247,
-0.0234680157,
-0.0641738176,
-0.0115431333,
0.065732725,
0.015141408,
0.0077240481,
-0.007330555,
-0.0325192511,
-0.0880543068,
0.0337715037,
0.0738601983,
0.0120880995,
-0.044458963,
0.0049388688,
-0.0092610316,
-0.006298848,
0.0230323654,
-0.005042376,
0.0770143643,
0.0079806047,
-0.0444533266,
-0.0341393091,
0.0149488514,
-0.0360786878,
-0.0436533168,
0.0291489661,
-0.0056941458,
-0.0848853216,
-0.0396750532,
-0.0340496935,
0.0461131185,
-0.0126273613,
-0.0135892602,
0.0160991009,
0.0407935418,
-0.0031259242,
-0.0169258453,
0.0123449527,
0.0206521917,
-0.0141707966,
0.0084374892,
0.088754341,
0.0354703739,
0.0583450273,
0.0256365333,
0.0164620001,
-0.0391405523,
0.0294986274,
-0.0066545177,
-0.0181372613,
0.0125089474,
0.0327696986,
0.0271627251,
-0.0118287923,
0.0475656167,
0.0241804011,
0.0253275763,
0.0224820245,
-0.071162641,
0.0034275951,
0.0687806532,
-0.0334455967,
0.0246375501,
-0.0028440373,
0.0117866043,
-0.0182723179,
0.0103318999,
-0.0176888928,
-0.0010414475,
0.0397475064,
0.0045264461,
-0.0230555478,
0.0136388028,
0.010209118,
-0.0495251901,
0.01727264,
0.0815837085,
-0.0260041934,
-0.0518073402,
0.0170205832,
0.0410719812,
0.0351315551,
-0.0099619497,
0.040521048,
0.0413181707,
-0.0191561263,
0.011215982,
-0.0419358537,
0.0341203585,
0.0423028506,
0.0473654494,
0.0395347849,
0.0297690108,
0.0263062045,
0.0068809334,
0.0474376082,
-0.0126716644,
-0.0279800426,
-0.0253458098,
-5.897227773e-33,
-0.0048751421,
-0.0296991114,
-0.0502676144,
-0.0630007312,
0.009256253,
-0.0188862979,
-0.0027606336,
-0.0153335612,
-0.0395983346,
-0.0007241635,
-0.0123863788,
0.0498888157,
0.024623299,
0.0371611379,
-0.0162836704,
-0.0236562099,
0.0333986804,
0.0201634876,
0.0250841901,
0.0212542415,
0.0275500771,
0.0557700433,
0.076244995,
-0.0370203666,
0.0015520229,
-0.0385336354,
-0.0134525895,
0.0119333426,
-0.0091354409,
0.0510359593,
-0.0627373308,
0.0606475435,
0.0507991537,
-0.0012592148,
-0.0053360234,
0.0873639509,
-0.0324147083,
-0.0043805293,
-0.023055518,
0.0762023553,
-0.0605007447,
-0.0255017541,
0.0583566315,
0.0121874604,
-0.0254762787,
0.0529881604,
0.0482692681,
-0.0255545247,
0.0012628048,
-0.0208933856,
-0.0555136316,
0.059989512,
0.0083880145,
0.0123894392,
-0.0059694331,
0.0099121956,
0.0294975508,
-0.025936598,
-0.0919575617,
-0.0234216861,
-0.0002899281,
0.0039102994,
-0.0176849049,
0.0388186984,
-0.013048891,
0.0398405157,
0.02478013,
-0.0108861951,
-0.005686352,
-0.0005237607,
0.0387333743,
0.0090778833,
0.0505034477,
0.0514098965,
0.0671523735,
-0.0426613204,
-0.0841114521,
0.0301008038,
0.0695873722,
0.0192597229,
0.058490973,
-0.0171179269,
-0.0655363277,
-0.0081515247,
0.0575559847,
-0.0243716389,
-0.0165302232,
-0.0318396017,
0.0332963727,
0.0061690658,
-0.0254235864,
0.0324308872,
0.0345557109,
-0.0355817974,
0.0093332892,
0.006321514,
-0.0073666391,
0.0034746369,
0.0418780707,
0.0560295917,
-0.0598888956,
-0.0229122844,
-0.0061665806,
0.0232885778,
-0.0136479614,
0.0094281249,
0.0613597892,
-0.0000744862,
-0.059253592,
0.0153212091,
-0.0074373991,
0.0326298028,
0.0092718704,
0.119497411,
-0.0092247948,
-0.0239722263,
-0.0071387482,
0.023101382,
0.0009061203,
0.0034623202,
-0.0072858301,
-0.0119184498,
-0.0009308131,
-0.0136381146,
-0.0253880601,
0.0065762629,
-0.0475407206,
0.0204852279,
-0.0120627806,
-0.0759952441,
0.0236430205,
-0.0160780121,
2.549e-7,
-0.0236936081,
0.0301924478,
0.0345260352,
-0.0512664616,
0.0121849747,
0.0251550563,
0.0191071685,
0.0211791731,
-0.0333504714,
0.003174315,
-0.0074179103,
0.0195959993,
0.0226724669,
0.0213730447,
-0.0720096529,
-0.0279289633,
-0.0801394284,
0.0055263764,
-0.0403231531,
0.0167336836,
-0.0283324234,
0.1096825227,
0.0458789952,
-0.0411380045,
-0.0133105507,
-0.0172348842,
0.0115356762,
-0.0127078565,
0.0390102379,
0.0036527975,
0.0478827693,
0.0436374284,
-0.0164642092,
-0.0164908841,
0.0106097292,
0.0215114243,
0.0307795778,
0.0231380723,
-0.0290410593,
0.0266119409,
0.0433347635,
0.0086142337,
0.0075396248,
-0.0443071462,
0.038341701,
0.015354149,
-0.0335840099,
-0.0373328328,
0.0116440048,
-0.0108061386,
0.0546588488,
0.0155658573,
-0.0068573514,
-0.0164830219,
0.0277713314,
-0.0903799981,
-0.0338068046,
-0.091401346,
0.0021355641,
-0.0133554125,
-0.0613530651,
-0.007928432,
-0.0351419933,
0.0129224863,
0.0277246553,
-0.0250304453,
0.0399179235,
1.963549207e-34,
0.02745419,
-0.0086299917,
0.0021543428,
-0.0164011847,
0.0163733158,
0.0237534773,
-0.0208875984,
-0.0474821478,
0.0419266373,
-0.0351531617,
-0.0033866821
] |
http://arxiv.org/abs/2401.16745v1 | http://arxiv.org/pdf/2401.16745v1 | MT-Eval: A Multi-Turn Capabilities Evaluation Benchmark for Large Language Models | 2024-01-30T04:50:28 | 2024-01-30T04:50:28 | Wai-Chung Kwan, Xingshan Zeng, Yuxin Jiang, Yufei Wang, Liangyou Li, Lifeng Shang, Xin Jiang, Qun Liu, Kam-Fai Wong | cs.CL | cs.CL | Large language models (LLMs) are increasingly relied upon for complex
multi-turn conversations across diverse real-world applications. However,
existing benchmarks predominantly focus on single-turn evaluations, overlooking
the models' capabilities in multi-turn interactions. To address this gap, we
introduce MT-Eval, a comprehensive benchmark designed to evaluate multi-turn
conversational abilities. By analyzing human-LLM conversations, we categorize
interaction patterns into four types: recollection, expansion, refinement, and
follow-up. We construct multi-turn queries for each category either by
augmenting existing datasets or by creating new examples with GPT-4 to avoid
data leakage. To study the factors impacting multi-turn abilities, we create
single-turn versions of the 1170 multi-turn queries and compare performance.
Our evaluation of 11 well-known LLMs shows that while closed-source models
generally surpass open-source ones, certain open-source models exceed
GPT-3.5-Turbo in specific tasks. We observe significant performance degradation
in multi-turn settings compared to single-turn settings in most models, which
is not correlated with the models' fundamental capabilities. Moreover, we
identify the distance to relevant content and susceptibility to error
propagation as the key factors influencing multi-turn performance. MT-Eval is
released publicly to encourage future research towards more robust
conversational models. | MT-Eval: A Multi-Turn Capabilities Evaluation Benchmark for Large Language Models: Large language models (LLMs) are increasingly relied upon for complex
multi-turn conversations across diverse real-world applications. However,
existing benchmarks predominantly focus on single-turn evaluations, overlooking
the models' capabilities in multi-turn interactions. To address this gap, we
introduce MT-Eval, a comprehensive benchmark designed to evaluate multi-turn
conversational abilities. By analyzing human-LLM conversations, we categorize
interaction patterns into four types: recollection, expansion, refinement, and
follow-up. We construct multi-turn queries for each category either by
augmenting existing datasets or by creating new examples with GPT-4 to avoid
data leakage. To study the factors impacting multi-turn abilities, we create
single-turn versions of the 1170 multi-turn queries and compare performance.
Our evaluation of 11 well-known LLMs shows that while closed-source models
generally surpass open-source ones, certain open-source models exceed
GPT-3.5-Turbo in specific tasks. We observe significant performance degradation
in multi-turn settings compared to single-turn settings in most models, which
is not correlated with the models' fundamental capabilities. Moreover, we
identify the distance to relevant content and susceptibility to error
propagation as the key factors influencing multi-turn performance. MT-Eval is
released publicly to encourage future research towards more robust
conversational models. | [
0.0223008972,
0.0486597233,
-0.0218620952,
0.0522068664,
-0.0436685309,
0.010385897,
-0.0065396866,
-0.01492819,
-0.0135060577,
-0.0310834032,
-0.0434248634,
-0.0503016971,
0.0176340938,
0.0281399619,
0.0244979504,
-0.0520356633,
0.0396937653,
-0.038940452,
-0.0127559416,
0.0135542322,
-0.0089588398,
0.0043100533,
-0.0257948637,
0.0690612048,
-0.0041868784,
-0.0127511956,
-0.0228958093,
0.018737454,
0.002197704,
-0.0100714844,
-0.0055985549,
0.052554816,
0.0139490915,
0.042201113,
0.0000021657,
-0.0180561393,
0.0134829599,
-0.0234700684,
-0.0295015648,
0.0070958561,
0.0442923345,
-0.0144811468,
0.0058568339,
0.0308528952,
-0.040055342,
-0.0224390291,
0.0223759338,
0.0391394161,
0.0567326024,
0.0887001306,
-0.0062002982,
-0.0052099936,
0.0344894081,
-0.0361877494,
0.0160719026,
-0.0939439759,
0.007797638,
-0.0085044121,
0.0026127659,
0.0141782667,
-0.0080127232,
0.0383440442,
0.0326500908,
-0.021486586,
-0.0267623328,
0.0338125229,
-0.0423710681,
-0.016362993,
-0.008624061,
0.0807510242,
-0.0071549187,
0.015045139,
0.0058944123,
0.0064916462,
0.0044909641,
-0.0187577587,
-0.0713696554,
-0.0142784556,
0.0111737577,
-0.0227197558,
-0.0732588023,
0.0046334835,
0.0191058274,
0.0224996861,
-0.0371244103,
0.0666960105,
0.0011280394,
-0.0204538517,
-0.0008801431,
-0.0160093978,
0.0160381589,
-0.041687455,
-0.0073287766,
0.0076680593,
-0.0811803937,
0.0310637914,
-0.0362296067,
0.0380403064,
0.0380157866,
-0.0294058472,
0.0215408001,
0.0302088112,
0.0048340429,
0.0525132939,
0.0299366172,
0.0406316705,
-0.0297856294,
-0.0109058619,
-0.0133852987,
-0.0487167165,
-0.0467915721,
-0.0055595296,
-0.0305126887,
0.0112391077,
-0.0596214794,
-0.0594369844,
-0.0213744044,
0.0549338348,
0.0221484806,
0.0183976367,
0.0198832992,
-0.0084353974,
-0.0114792287,
-0.0039143935,
-0.0108913658,
0.0312156379,
-0.0184483398,
-0.003379778,
0.0410190523,
-0.1251812279,
-0.0176052954,
0.0027156398,
-0.0066160336,
-0.0549850091,
-0.0185350832,
0.0678876638,
-0.0167570841,
-0.0488546006,
0.009996783,
-0.0190771148,
-0.0206883438,
-0.0923458189,
-0.0230814181,
-0.0410454087,
0.0680778921,
0.0145979812,
-0.0092440266,
-0.0183068775,
0.0070276596,
0.0309566613,
-0.0358804725,
0.0337682739,
0.033333201,
0.0463930778,
0.0177759808,
0.031761732,
-0.0188797023,
0.074355267,
0.0233750865,
0.0026002601,
0.0550183915,
-0.0178312436,
0.0164909121,
0.0515539534,
0.0100241303,
-0.0077449796,
0.0114937434,
0.0293529835,
0.007494844,
0.0527407974,
0.0147734825,
0.0629210323,
-0.0598376878,
-0.0393128693,
0.0756246001,
0.0323382579,
0.0033955337,
0.0129336016,
0.0244220644,
0.0362811387,
0.0043561733,
-0.0005138766,
-0.0787680671,
0.0231090914,
-0.0281657055,
-0.0060116979,
0.0573673099,
0.0212001707,
-0.0330397263,
-0.0187009797,
0.0160494912,
0.0119136572,
0.0134804603,
0.0074192276,
0.0254973993,
-0.0255818423,
-0.0582462735,
0.0529010333,
0.000817491,
-0.0605804361,
-0.0426836535,
-0.0391991213,
-0.0090167988,
0.0765013397,
0.0482155383,
0.0081163095,
-0.0223249756,
-0.0031231658,
-0.0121655017,
0.0303879846,
0.0608471818,
0.0110388501,
0.0664418489,
-0.0398652479,
0.0075464365,
-0.0031546857,
0.0024546976,
0.0408210158,
0.01714845,
0.0143402992,
0.0100674499,
-0.0258902498,
0.0345218554,
0.0048977961,
-0.0123447198,
0.0142832305,
-0.016700197,
-0.0043750587,
0.0120918509,
0.0155710746,
-0.0152069312,
0.0148924338,
0.0309549253,
0.0031221765,
-0.0048753615,
0.0255812816,
-0.0658303574,
-0.0340158977,
0.022245001,
-0.0094655268,
-0.0291800648,
0.0314314365,
-0.0096205389,
-0.0614111647,
0.0119386557,
-0.0299146697,
0.0706439242,
-0.0306479298,
0.0516238511,
0.0086252606,
0.0198225994,
0.0106777586,
0.0368454605,
-0.0273953099,
0.0314275138,
0.0178403594,
-0.0138203604,
-0.0099108964,
-0.0623489656,
-0.0412091501,
0.0461392775,
-0.0071450206,
-0.0077970442,
0.0125844013,
-0.0013800275,
-0.0664089397,
-0.0503052957,
-0.1275974065,
0.0055711358,
-0.0403669477,
0.0005335115,
0.0008468388,
0.027982913,
-0.0149477106,
-0.0084320353,
-0.0240737218,
-0.025081329,
0.019925613,
0.0421226434,
-0.0043337303,
0.0321323201,
-0.0331338756,
-0.0441243947,
-0.0020842534,
0.0273814742,
0.0486280099,
-0.0237708762,
-0.0032384011,
0.0053022658,
-0.0104873776,
-0.0315463096,
0.0344842114,
-0.0479149967,
-0.0432549939,
0.0257519577,
-0.0130809182,
0.040391773,
0.0540130846,
0.0329016335,
0.0023626746,
-0.0092850178,
-0.0627894402,
-0.0648760274,
0.035118062,
-0.0256731696,
-0.0232344102,
-0.069278881,
-0.0171036534,
0.0175353941,
0.0465649851,
0.0001246356,
0.0442908369,
-0.0462355837,
-0.0350676365,
0.0359023027,
-0.0559750088,
-0.0180651844,
0.0173468031,
-0.0360399522,
0.0326655358,
-0.0301739592,
-0.0368599854,
0.0427603461,
-0.0634004325,
-0.0483776703,
0.0137684373,
-0.0040827785,
-0.0182978809,
0.0749067366,
0.0063193706,
0.0060823341,
-0.0335645676,
-0.0363023244,
-0.0160377678,
-0.0040554004,
0.0178850088,
-0.0450746939,
0.085204497,
-0.0547555611,
0.0048033372,
0.0136303473,
-0.0457478017,
-0.0245385282,
-0.0193979014,
-0.0226651188,
0.0110009657,
0.0362679288,
-0.0594002418,
-0.0521648452,
0.0163578112,
0.0281106643,
0.0494126454,
-0.0064637801,
-0.0024415331,
-0.0170825031,
-0.002899155,
0.0579467341,
0.0077855806,
-0.0069161719,
-0.018755246,
0.0672864988,
0.0631281286,
0.060100738,
-0.011526729,
-0.0017993749,
-0.0132876523,
0.0136490148,
-0.0161612742,
-0.0482005328,
0.0338304378,
0.0380927697,
-0.046977859,
-0.0143826446,
0.1120685115,
0.0051951674,
-0.0495170914,
-0.0176576339,
-0.0243423991,
0.0460882634,
0.0246847626,
0.0118068121,
-0.0707198009,
-0.0272413641,
0.0148344152,
0.009990179,
-0.002941245,
-0.0475013293,
-0.0120475711,
-0.1401878595,
0.0295031946,
-0.0188908242,
-0.0560062081,
-0.0152546596,
-0.0254630577,
0.0144599769,
-0.0074495869,
0.0414718315,
0.0126042999,
-0.0017800721,
-0.0379690193,
0.0004486049,
-0.0296774488,
-0.0422084257,
0.0673178956,
0.0100304987,
0.0388401672,
0.0436709598,
0.0380428843,
0.0215855576,
-0.032088086,
0.0488517769,
-0.0216864664,
0.006640438,
-0.0151390452,
0.0446897708,
-0.030733142,
0.0086772954,
-0.0114314295,
0.0072844629,
0.0077026347,
-0.0096995449,
0.018377265,
0.0427139215,
-0.043984253,
0.0061148079,
-0.0096117286,
-0.0014577038,
0.0322974809,
0.0807565823,
0.0387827717,
-0.0261738915,
-0.0495641679,
0.0100447387,
-0.0182031225,
-0.0475269891,
-0.0255417861,
-0.02252388,
0.0581340417,
0.0111553483,
-0.0383481197,
-0.0421866626,
0.0086593172,
0.0169862472,
-0.0474462286,
0.0234703794,
0.0214157812,
0.0145858815,
-0.0012251313,
-0.0015175072,
-0.0168639198,
0.038300585,
-0.0748487711,
0.0049350196,
0.0338661,
-0.0075364481,
-0.0127217425,
0.0133613767,
-0.0857699439,
-0.061312329,
0.0007789449,
0.0373946019,
0.0299579725,
0.043057967,
-0.0186706744,
-0.053603664,
-0.0281519406,
0.0069950651,
-0.0768821537,
0.0266863126,
0.0354960412,
0.0374706611,
-0.0128960004,
0.0688131973,
-0.0343848988,
0.0069466257,
-0.02996815,
0.0432290286,
-0.0520048775,
-0.0926736593,
-0.0125342663,
-0.0147537133,
0.0719204992,
0.0385134183,
0.0063984427,
0.0135855032,
0.0300851986,
-0.0348354727,
-0.0032867347,
-0.0664918795,
0.0078200363,
0.0014990148,
-0.0114419479,
0.0065003778,
-0.015513937,
-0.0330799446,
0.0176927596,
-0.0465313867,
-0.009379318,
0.0035982784,
-0.0307798386,
0.0313171372,
-0.0014544871,
-0.0206530597,
-0.0183298755,
0.0077106194,
0.0343269184,
0.002236279,
0.080652222,
0.0108838314,
-0.0408035703,
0.0252618715,
0.0246966872,
-0.0934744999,
-0.043566823,
0.008811783,
0.042734392,
0.0308008455,
0.0179054216,
0.0653108209,
-0.0104113976,
0.0188043322,
-0.0434323363,
0.0524971895,
0.0054540443,
-0.0394958071,
0.0487426147,
0.0325113758,
-0.0212464929,
0.0322669297,
0.0699822903,
-0.0081834411,
0.0183357038,
0.0046420936,
0.003895845,
-0.042368345,
-0.0646650866,
0.0158579405,
0.0130916778,
-0.0076787062,
0.0116397087,
-0.0610345714,
-0.0396817364,
0.0037961747,
-0.0036079935,
0.0041778809,
-0.0312010273,
0.041024901,
0.0812855363,
0.0089666592,
-0.0117493393,
-0.0492498688,
-0.0144449966,
-0.0305327158,
-0.0054304092,
-0.0117389793,
-6.785400815e-33,
-0.0199905429,
-0.0177279804,
0.0341758505,
-0.005563553,
-0.0574586764,
-0.0248651374,
-0.0495571978,
0.0298254527,
-0.0561261773,
-0.0057683042,
-0.0373542719,
0.0368295163,
0.0254873596,
0.0367655344,
-0.001992936,
-0.0178995673,
0.0372506753,
0.0090121347,
-0.0384032577,
0.0466749892,
0.0253146049,
0.0219267216,
0.0266222525,
-0.0662784502,
0.0100267762,
-0.0334303007,
-0.0193135571,
0.0332001522,
-0.0053925235,
0.0531242378,
0.0012952451,
0.0255843755,
0.0243005957,
-0.0666661561,
-0.0194284096,
0.0240580961,
-0.0334853604,
-0.0084508089,
0.0437413342,
0.0096184202,
-0.0202728119,
0.0162550882,
0.0101187034,
0.0039055001,
0.0089141428,
-0.0136122713,
-0.0000465239,
-0.0090000173,
-0.054599043,
0.0257082507,
-0.0219049528,
0.0362382419,
-0.0179590937,
-0.0488279015,
-0.0055171498,
0.0717131346,
0.0134997303,
-0.00062077,
-0.0718082711,
0.0206149202,
-0.023571426,
0.0206375569,
0.0553499982,
0.0247987006,
0.0253366902,
0.0604890957,
0.0303007998,
0.0139495777,
-0.0005299677,
-0.0203407444,
0.0172334854,
0.0806602761,
-0.0065122792,
0.0462477319,
-0.0165321529,
-0.004881382,
-0.0154250041,
-0.0048059383,
0.0353170633,
0.0655096695,
0.0241325423,
0.0072003384,
-0.0520884357,
-0.0112875663,
-0.0505654924,
0.0237281043,
-0.0353450924,
-0.0189464688,
0.0478542484,
0.0315148905,
0.0314078182,
0.0062609152,
0.0641703904,
-0.0191737581,
-0.0198102221,
0.053305693,
-0.0450695492,
0.0227339547,
-0.0117813786,
0.0271899719,
-0.0332763195,
-0.0222038869,
0.0104753664,
0.0202946737,
0.010720293,
0.0020302678,
-0.011578978,
0.01500306,
-0.0772148147,
-0.0208764467,
-0.0337650776,
0.0451861806,
-0.0640912428,
-0.0048025614,
0.0352836549,
-0.0228151064,
-0.0376374871,
0.0822843909,
-0.0339285769,
0.0120366877,
-0.0446065143,
-0.0427037552,
0.0336650722,
-0.0010201404,
-0.0345540829,
0.0011391868,
-0.0208809525,
-0.033085905,
0.0484473817,
0.011011539,
-0.0143743306,
0.0138060218,
2.883e-7,
0.0025811908,
0.0302827861,
0.0185782183,
0.0353731588,
0.010827058,
0.0319811776,
-0.035946656,
0.0101367654,
0.0232534707,
-0.0004353993,
0.0140500832,
0.0322926715,
0.0823018476,
0.0171043221,
-0.0736515298,
-0.0294308737,
-0.095422104,
-0.0188343525,
-0.0400008224,
-0.0391571,
0.0388165526,
0.0829373673,
-0.0095513314,
0.0103510376,
-0.002722661,
-0.0031720076,
0.0768940598,
-0.0747567862,
-0.0266318806,
0.0095408205,
0.0154317757,
0.0337059498,
0.0154812662,
0.0125940163,
0.0018625492,
0.0337386802,
0.0099596605,
0.1060497314,
0.0139872693,
-0.0616427585,
0.0231827628,
-0.0178487562,
-0.0089985905,
-0.0591499768,
0.0553928204,
-0.0280849803,
-0.0072098207,
0.0064807814,
-0.0032650197,
-0.0021998512,
0.0347751789,
-0.0028998125,
-0.0231185332,
0.0229585283,
0.0203541145,
-0.0170887224,
0.009781017,
-0.0870041624,
-0.0207427461,
-0.0181041006,
-0.071214959,
-0.052469302,
-0.052977588,
-0.0429207496,
0.0123698981,
-0.0113086803,
-0.0131754829,
2.318192639e-34,
0.0574555695,
-0.0064090695,
0.0355959646,
0.0323844217,
0.0422850922,
-0.0063865124,
0.0096017197,
0.0256393384,
0.0109054241,
-0.0971848816,
-0.0054322262
] |
http://arxiv.org/abs/2401.16736v3 | http://arxiv.org/pdf/2401.16736v3 | Engineering A Large Language Model From Scratch | 2024-01-30T04:29:48 | 2024-02-03T16:34:46 | Abiodun Finbarrs Oketunji | cs.CL | cs.CL, cs.CY, cs.LG, cs.SE, I.2.7, I.2.7 | The proliferation of deep learning in natural language processing (NLP) has
led to the development and release of innovative technologies capable of
understanding and generating human language with remarkable proficiency.
Atinuke, a Transformer-based neural network, optimises performance across
various language tasks by utilising a unique configuration. The architecture
interweaves layers for processing sequential data with attention mechanisms to
draw meaningful affinities between inputs and outputs. Due to the configuration
of its topology and hyperparameter tuning, it can emulate human-like language
by extracting features and learning complex mappings. Atinuke is modular,
extensible, and integrates seamlessly with existing machine learning pipelines.
Advanced matrix operations like softmax, embeddings, and multi-head attention
enable nuanced handling of textual, acoustic, and visual signals. By unifying
modern deep learning techniques with software design principles and
mathematical theory, the system achieves state-of-the-art results on natural
language tasks whilst remaining interpretable and robust. | Engineering A Large Language Model From Scratch: The proliferation of deep learning in natural language processing (NLP) has
led to the development and release of innovative technologies capable of
understanding and generating human language with remarkable proficiency.
Atinuke, a Transformer-based neural network, optimises performance across
various language tasks by utilising a unique configuration. The architecture
interweaves layers for processing sequential data with attention mechanisms to
draw meaningful affinities between inputs and outputs. Due to the configuration
of its topology and hyperparameter tuning, it can emulate human-like language
by extracting features and learning complex mappings. Atinuke is modular,
extensible, and integrates seamlessly with existing machine learning pipelines.
Advanced matrix operations like softmax, embeddings, and multi-head attention
enable nuanced handling of textual, acoustic, and visual signals. By unifying
modern deep learning techniques with software design principles and
mathematical theory, the system achieves state-of-the-art results on natural
language tasks whilst remaining interpretable and robust. | [
0.0198990032,
0.0278801844,
-0.0281699747,
0.0625864416,
-0.0140350992,
0.0093791327,
-0.0094190501,
-0.0238374118,
-0.0030703158,
-0.0367777832,
-0.0460778289,
-0.0311589148,
0.0251429696,
0.0632579327,
0.0423898473,
-0.0335682519,
0.0293473452,
-0.0131870257,
-0.0433688313,
0.0227161627,
-0.0056674881,
0.0100839734,
0.018241398,
0.0652175546,
-0.0337258913,
-0.012241831,
-0.0280159842,
-0.0125740934,
0.0111067966,
-0.0031368595,
0.0067102029,
0.0199492574,
0.0085339826,
0.0626321733,
0.0000022399,
0.0115331123,
0.0326333754,
-0.0163693856,
-0.0155446948,
-0.0094447266,
0.0377574638,
-0.0370627157,
-0.0110796606,
-0.0028153108,
-0.0536460057,
-0.0242131148,
0.0436906777,
0.0474018268,
0.0819714591,
0.0682733655,
-0.0223628785,
-0.0710671246,
0.0526530743,
-0.0334663019,
-0.051174257,
-0.0238458589,
0.0042576282,
-0.0571424104,
0.0256119464,
0.0028988777,
-0.0176392123,
0.0508106872,
0.0535299815,
-0.013557462,
0.0446387529,
0.026132971,
-0.0131867295,
-0.061082691,
-0.0345171504,
0.1101061627,
-0.0758025721,
-0.0103793899,
0.0033457342,
-0.0275514759,
0.0165074207,
-0.0326892696,
-0.0111080054,
0.0255353898,
0.0512760133,
0.0248052943,
-0.0012595484,
0.0002259189,
0.0294810906,
0.0176542085,
-0.0154494131,
0.0380616672,
0.0188199338,
0.0142584946,
-0.0176537819,
0.0023134025,
-0.0163104627,
-0.0331566297,
0.0229273625,
0.0514716655,
-0.0255873892,
-0.0005640627,
-0.0427149981,
-0.0803353712,
0.0038484728,
-0.0666292235,
0.06580735,
0.0230243057,
-0.0565544665,
0.0309046693,
-0.0231620539,
0.0432707556,
-0.0708981156,
0.0176657327,
-0.0244422294,
-0.0257206243,
-0.0190822966,
-0.0264091287,
-0.0340637602,
0.0222866647,
-0.0429407991,
-0.0543461218,
-0.0423426144,
0.0238055792,
-0.0010794041,
-0.0138672749,
0.0289006513,
0.0300556868,
-0.0610427335,
-0.0376136675,
-0.0031949682,
0.0524829291,
-0.0284664836,
0.0267858636,
0.001816736,
-0.046736531,
-0.0024535113,
0.0467337444,
0.0012920167,
-0.0162648801,
0.0026332305,
0.0722093582,
-0.0211627781,
-0.0160491392,
-0.0195339881,
-0.0049999342,
-0.0214014519,
-0.0408206284,
0.0062960889,
-0.0602930263,
0.0605349131,
-0.0118511366,
-0.0236129425,
-0.0474696159,
0.0074236505,
0.0167619921,
-0.0091704074,
0.0532256663,
0.000892101,
-0.0072477204,
0.0326005481,
0.0045404211,
0.0316289067,
0.0547602512,
0.0532428585,
0.0447991341,
-0.0086898571,
-0.0158044267,
0.0143072288,
0.0622563586,
-0.0094445962,
-0.033999268,
-0.0202275738,
-0.0297141746,
0.0023371249,
0.0060044588,
0.0251412466,
0.0496230796,
-0.0122115044,
-0.0034663337,
0.0704268143,
0.0265689231,
0.0009017285,
0.0828257576,
0.0351163894,
0.0203259978,
-0.02386971,
-0.0577797592,
-0.0392929874,
0.0770792365,
0.0204218626,
-0.016614601,
-0.0489181131,
0.065786168,
-0.0302131902,
-0.0317659378,
0.0199485142,
-0.0201046355,
0.0162405483,
0.0036502064,
0.0463614725,
0.0341823697,
-0.0646761507,
0.0814688876,
0.0142738372,
-0.01167503,
0.009554795,
-0.0348012038,
-0.0229400788,
0.0476964898,
0.0071792505,
-0.0224229731,
0.0130130388,
-0.0162383839,
-0.0022444548,
0.0670055747,
0.0264554042,
0.0158927627,
0.0127907544,
-0.0513744503,
0.0028892867,
-0.00375071,
0.0061327438,
0.0312478244,
-0.0250048451,
0.0167011321,
0.0192445256,
0.0370981544,
-0.0096809436,
0.0561458357,
-0.0218411535,
-0.0245959498,
0.0031883689,
0.0217186436,
0.0484687313,
-0.0073890002,
0.023031693,
0.008386272,
0.0458976887,
-0.0728769302,
-0.0126507804,
-0.0054546455,
-0.0401466116,
0.0240710322,
0.0353919156,
-0.0725513995,
-0.0195337925,
0.0629586354,
-0.024921475,
-0.0514150076,
0.0177322347,
-0.073262237,
0.0087279072,
-0.0431482978,
0.0256986693,
-0.0120331515,
0.0729509369,
-0.017454097,
0.0480480902,
0.0156812146,
-0.0328064933,
0.005089107,
-0.0367123969,
-0.003419322,
-0.0132731767,
-0.0374508016,
0.0055527589,
-0.0260104351,
-0.0043781977,
-0.0206476655,
-0.0049907546,
-0.0206304323,
-0.031563662,
-0.046232678,
-0.0238381084,
-0.0069284579,
-0.0110321371,
-0.0000419388,
0.0346203484,
0.0273089465,
-0.0168352444,
-0.0407404751,
0.0519058183,
0.0112010967,
-0.0085974298,
0.0433055386,
-0.0022598961,
-0.0397484452,
-0.0473025441,
-0.0445560701,
-0.0071226354,
0.0463335812,
-0.004056856,
0.0065162154,
-0.0619299635,
-0.057585638,
-0.0008928641,
0.0329793356,
0.0085711302,
-0.0721323937,
-0.0134688653,
-0.0333357453,
0.0370181724,
0.0726956949,
0.0176092219,
0.0451125912,
0.0419103429,
-0.0054832986,
-0.0151997562,
0.0395156853,
-0.0314292647,
0.0312731043,
-0.0320974216,
0.0255426839,
-0.0233503263,
0.0819201022,
-0.0142995194,
0.0139566092,
-0.0521767661,
0.0002998313,
0.0034398101,
-0.0277750008,
0.0108788945,
-0.0182112623,
-0.0310708638,
0.0304935295,
-0.0316701606,
0.001109665,
0.0681610107,
-0.0002749347,
-0.0297042504,
-0.0218625348,
0.0017124892,
-0.0076975119,
0.057425864,
0.0068398998,
0.0070225932,
-0.0242656898,
-0.0067841979,
-0.0445617996,
0.0219270941,
-0.0101985754,
-0.0257214047,
0.0595352724,
-0.0237726923,
0.002362866,
-0.014799946,
-0.0217818338,
-0.0467581525,
-0.0177704953,
-0.0399237536,
0.0072163432,
0.0359433666,
0.0011178878,
-0.05938958,
-0.0075052385,
0.0108474726,
0.0322307236,
0.0034546268,
0.020828845,
-0.043380294,
0.0463547781,
0.042818062,
0.0399920307,
0.0313840099,
-0.0022583893,
0.0248580575,
0.0271051358,
-0.0017696796,
0.0155022396,
-0.0535083041,
0.0067566624,
0.040009059,
-0.0166099332,
-0.0209740773,
0.0166013855,
0.0191433597,
-0.1073245704,
-0.0513743013,
0.0659079626,
0.001537696,
-0.0504326969,
0.0305118393,
0.0104052536,
0.0312090758,
0.0149224969,
0.0650292188,
-0.0574537106,
-0.0209632274,
0.0074488698,
0.0015541454,
0.0319062397,
0.0085347472,
0.0237368047,
-0.0974010229,
0.0246112905,
-0.004903357,
-0.0257200394,
-0.0275758039,
-0.0255580153,
-0.0082969768,
-0.000148136,
0.0271788817,
0.0169186499,
0.0247848518,
-0.0549941398,
-0.0173615795,
0.0035789192,
-0.0254248101,
0.063594386,
0.0040924847,
0.0335333943,
0.0515822321,
0.0458689965,
-0.0231504925,
-0.0688532442,
0.0008314725,
0.005688577,
0.0025798653,
0.0161625668,
0.0275814738,
0.0145018799,
-0.0517558008,
-0.0463530459,
0.0407900214,
-0.0502267219,
0.0058781146,
-0.0516261719,
0.0909927711,
-0.0137583455,
-0.004499787,
-0.0186580326,
-0.0354323275,
-0.0038675331,
0.0459104925,
0.0061961454,
-0.0718997195,
-0.0621203445,
-0.001686363,
-0.0064898264,
-0.0531517565,
-0.082994096,
0.0502889156,
0.0228727479,
-0.064320378,
-0.0713421404,
-0.0295304842,
-0.01239629,
0.0028240674,
0.0422090366,
-0.0750595182,
0.0761699155,
-0.008165651,
-0.0483947769,
0.0389569551,
-0.006669587,
0.025222173,
-0.0499613918,
-0.0004909015,
0.0501941666,
-0.0386831947,
-0.0008897264,
0.0669674426,
-0.0486501977,
-0.0531107597,
0.0475837924,
0.0498387851,
0.0145875877,
0.0214031078,
-0.0300697405,
-0.0391738787,
0.024710713,
0.005488304,
-0.0563072935,
-0.0173159875,
0.0236598663,
-0.0385818072,
-0.0019702853,
0.0553542599,
-0.0350323021,
0.0431343578,
0.0075354911,
0.0246133618,
-0.0089974748,
-0.0396577306,
0.003208929,
-0.0334001854,
0.0485930406,
0.0084499288,
-0.0165105034,
-0.0183867104,
3.3e-7,
-0.0165212657,
0.0108767133,
-0.0496266745,
-0.0124053294,
0.0553973466,
0.0116533963,
0.0189745054,
-0.0064449101,
0.0465362333,
0.0093550822,
-0.0112931784,
0.0520717204,
0.008155033,
0.0014800304,
-0.0172698777,
0.0131714055,
-0.0074669188,
0.0039527938,
-0.0112797078,
0.0429476202,
-0.0425090827,
0.0216894839,
0.0073745507,
-0.0106623601,
-0.06048676,
0.0241536666,
-0.0464529134,
-0.0163659137,
-0.0198473409,
0.0146459257,
0.0256463476,
-0.0222224612,
0.0589930937,
-0.007585505,
-0.0212829486,
0.0176857952,
0.0467579849,
-0.0401486792,
-0.0284981374,
-0.0092822928,
0.0228310451,
-0.0103865452,
-0.0183281619,
0.0353699103,
0.0079949554,
0.0049915337,
0.0085475156,
0.0077236765,
-0.0638526008,
-0.0126979677,
0.0107591394,
0.0145583609,
-0.0306389239,
0.0067999852,
-0.024765892,
-0.0729781836,
0.0025582146,
-0.000861234,
-0.0063841525,
-0.0207833368,
0.051242087,
0.0215655081,
0.0045936746,
-0.0365644954,
-0.0251826681,
0.0224857945,
0.0342075303,
-0.0222802814,
0.0001229136,
-6.727427638e-33,
-0.0295687541,
-0.0033232255,
-0.0043843254,
0.0193728637,
-0.0350892991,
-0.013480668,
-0.0396772027,
0.0007013911,
-0.0487840548,
0.0089426273,
-0.0166889261,
0.0208496321,
0.0059749684,
0.051063478,
0.0475565717,
0.0110227354,
0.0430073664,
0.0028371699,
0.0184915178,
0.0213732999,
0.0145061594,
0.0449904688,
0.0556872785,
-0.0386160985,
-0.0294713601,
0.002196976,
-0.0018677878,
0.0026294012,
-0.020164134,
0.0474246852,
-0.040033143,
0.0495574698,
0.0097030746,
-0.0735810548,
-0.0163712949,
0.0330070667,
-0.1169346198,
0.0006540382,
-0.0289024562,
0.0486406609,
-0.0118085146,
-0.022577282,
0.0913911089,
-0.0063316193,
0.0232940633,
-0.0381068885,
0.0513645224,
-0.0138677713,
0.0009640452,
-0.0303901117,
-0.0888037756,
-0.0177392736,
-0.0230464321,
-0.0342501476,
0.0436154716,
0.0103247957,
-0.0370267741,
0.0172441192,
-0.0984492153,
-0.009862205,
-0.0156693794,
0.037308991,
0.0649122968,
0.0597592555,
0.0052005127,
0.0345075838,
0.0301444512,
-0.0058945925,
-0.038947437,
0.0090754023,
0.0663314313,
0.0606166832,
0.0715629309,
0.0254538376,
0.0205933619,
-0.0180545002,
-0.0737914219,
-0.0355992168,
0.0210660622,
0.0783818886,
0.0380707793,
0.0279849451,
-0.0129605625,
-0.0043086773,
0.0111647295,
-0.0169178229,
-0.0315522961,
-0.0032940677,
0.0190727562,
0.0081840679,
-0.0482647941,
0.0528942235,
0.0324591249,
-0.0427404381,
0.0095049301,
0.0141228205,
0.002142488,
0.0270193685,
0.0099908402,
0.0248539001,
-0.0289163124,
-0.0557560138,
-0.009991724,
0.0254498273,
0.0173333492,
0.0206357352,
0.0231462084,
0.0364011601,
-0.0379506163,
-0.0108540198,
-0.0142348064,
-0.0036912602,
0.0283924378,
0.0598647855,
0.0270160213,
-0.0043693553,
-0.0268953163,
0.022168491,
-0.0014485245,
0.0185874328,
-0.0191913545,
-0.000131553,
0.0206120275,
0.0156884119,
-0.05888303,
-0.0018766923,
-0.0710193738,
0.0200684387,
0.0417817496,
-0.0671517774,
-0.0079818172,
0.0165602583,
2.972e-7,
0.0459075086,
0.0578326061,
0.0363161378,
0.0103844833,
0.0187602043,
0.0097102253,
0.0200723931,
0.0293670706,
-0.0015812702,
-0.0403316878,
0.0188259184,
0.0498050712,
0.0425797924,
-0.0095588323,
-0.0969892591,
-0.0092901886,
-0.0466021113,
0.014983532,
-0.0506995991,
-0.0378822461,
0.0224568546,
0.1502168477,
0.0120322295,
-0.0200388506,
0.0011556448,
-0.0053091873,
0.0483417884,
0.0026259478,
0.0354365148,
0.0157788545,
0.0246784184,
0.0096021984,
-0.0215538591,
0.0101116411,
-0.0138505949,
0.0232606679,
0.0196647998,
0.0592789128,
-0.0294370502,
-0.051205419,
-0.003532751,
-0.002763096,
-0.0050758258,
-0.0807595626,
0.0429959595,
-0.1099820137,
-0.0439356677,
-0.0198842231,
-0.0172076281,
-0.0037662452,
0.0110627478,
-0.0267739743,
0.0102732042,
0.0144836502,
0.030783616,
-0.0387627073,
-0.0164918527,
-0.0698898137,
-0.0214203503,
-0.0169705991,
-0.0649022758,
0.013958104,
-0.0653678775,
-0.0405755974,
0.0485438257,
-0.0248370729,
-0.0025136948,
2.487521911e-34,
0.007752338,
0.0278128441,
0.0121828737,
0.0555173606,
-0.0072630518,
-0.0080906637,
0.0143706398,
0.0210864581,
0.0221319906,
-0.0901203081,
-0.0455629602
] |
http://arxiv.org/abs/2401.16731v1 | http://arxiv.org/pdf/2401.16731v1 | Towards Generating Informative Textual Description for Neurons in Language Models | 2024-01-30T04:06:25 | 2024-01-30T04:06:25 | Shrayani Mondal, Rishabh Garodia, Arbaaz Qureshi, Taesung Lee, Youngja Park | cs.CL | cs.CL, cs.AI | Recent developments in transformer-based language models have allowed them to
capture a wide variety of world knowledge that can be adapted to downstream
tasks with limited resources. However, what pieces of information are
understood in these models is unclear, and neuron-level contributions in
identifying them are largely unknown. Conventional approaches in neuron
explainability either depend on a finite set of pre-defined descriptors or
require manual annotations for training a secondary model that can then explain
the neurons of the primary model. In this paper, we take BERT as an example and
we try to remove these constraints and propose a novel and scalable framework
that ties textual descriptions to neurons. We leverage the potential of
generative language models to discover human-interpretable descriptors present
in a dataset and use an unsupervised approach to explain neurons with these
descriptors. Through various qualitative and quantitative analyses, we
demonstrate the effectiveness of this framework in generating useful
data-specific descriptors with little human involvement in identifying the
neurons that encode these descriptors. In particular, our experiment shows that
the proposed approach achieves 75% precision@2, and 50% recall@2 | Towards Generating Informative Textual Description for Neurons in Language Models: Recent developments in transformer-based language models have allowed them to
capture a wide variety of world knowledge that can be adapted to downstream
tasks with limited resources. However, what pieces of information are
understood in these models is unclear, and neuron-level contributions in
identifying them are largely unknown. Conventional approaches in neuron
explainability either depend on a finite set of pre-defined descriptors or
require manual annotations for training a secondary model that can then explain
the neurons of the primary model. In this paper, we take BERT as an example and
we try to remove these constraints and propose a novel and scalable framework
that ties textual descriptions to neurons. We leverage the potential of
generative language models to discover human-interpretable descriptors present
in a dataset and use an unsupervised approach to explain neurons with these
descriptors. Through various qualitative and quantitative analyses, we
demonstrate the effectiveness of this framework in generating useful
data-specific descriptors with little human involvement in identifying the
neurons that encode these descriptors. In particular, our experiment shows that
the proposed approach achieves 75% precision@2, and 50% recall@2 | [
0.027781466,
0.0896677673,
-0.0124925449,
0.033969488,
0.0208286326,
0.0310576074,
0.0189038981,
-0.0269855522,
-0.0329144001,
-0.0229166355,
-0.0080495868,
-0.0089796633,
0.0082966369,
0.0404983237,
0.0524253957,
-0.0299483687,
0.0053107338,
-0.0138408467,
-0.0023376637,
0.0040855869,
-0.0033778318,
0.0459370315,
-0.0034528368,
0.0624755658,
-0.0103786308,
0.0195223819,
-0.0281451214,
0.0126783475,
0.0050354516,
0.0235909391,
0.0033681972,
0.0415489003,
0.0501443446,
0.038945552,
0.0000018212,
-0.0214662813,
0.0116790645,
0.0042474591,
-0.0466682613,
0.0056203338,
-0.0031204484,
-0.0403591394,
0.0138362059,
-0.0107115069,
0.0018870872,
-0.0235018134,
0.0612635948,
0.0271868128,
0.0381953157,
0.0636551902,
0.002211056,
-0.0520456992,
0.0191194471,
-0.0171076674,
0.0292311646,
-0.0133893806,
0.0056827138,
-0.0282636657,
-0.0119186575,
-0.005295679,
-0.0233358033,
0.0138730612,
0.0263604559,
-0.0479441807,
0.0372102931,
0.0407972746,
-0.0139197148,
-0.0341825001,
0.023324592,
0.0559140965,
-0.0131386127,
-0.0668308064,
0.0035914492,
0.008227828,
-0.0040696445,
0.0023150935,
-0.0351522788,
0.0207144003,
-0.0023447664,
0.0336175486,
-0.0145990727,
0.0006651093,
0.0431823656,
0.006002333,
-0.0226351693,
0.0085426299,
0.0159878079,
0.0254646316,
-0.0588959008,
0.0017749731,
-0.0084946677,
-0.0422742851,
0.0196698327,
0.0657042563,
0.0016187056,
0.0264585838,
-0.0217198357,
-0.0266596973,
-0.008490256,
-0.0279864687,
0.0456619561,
0.0451275967,
-0.0174381696,
0.0137587199,
-0.0318109989,
0.0678710863,
-0.0198649205,
0.0419464447,
-0.0205593407,
-0.0444761887,
-0.0947787464,
-0.0411895066,
-0.0273269061,
0.0063959006,
-0.0081973225,
-0.0159264319,
-0.0008528437,
0.0355571285,
0.016204061,
-0.0286110025,
0.0244042873,
0.0490403287,
-0.0513223559,
0.0062982244,
-0.0377742536,
0.0166331157,
-0.0304089505,
-0.0324604362,
0.0013089085,
-0.0281431489,
-0.0077574281,
0.0529233031,
0.0130017828,
-0.0320905074,
0.0067754574,
0.0712434947,
-0.0141076352,
-0.0312485676,
-0.0317136496,
0.0208907574,
-0.0223867595,
-0.0037660536,
0.0073144734,
-0.021824345,
0.0282551609,
-0.0041449098,
-0.0283933897,
-0.0438201427,
0.0171691608,
0.0100377435,
-0.0131634315,
0.0847746655,
-0.0479087606,
0.003653117,
0.0071953875,
-0.003559802,
0.0154261682,
0.0136665674,
0.0659317449,
0.0150535544,
0.004295141,
-0.0226075873,
0.0401741154,
-0.0138587076,
0.0353953019,
-0.0258140806,
-0.0473454483,
-0.0127249397,
-0.0370183438,
0.0255674105,
0.0549762174,
0.0445066169,
-0.0065675634,
-0.0054625338,
0.0869365931,
0.0713787451,
0.0150429541,
0.0409423746,
0.0642171726,
0.0042161858,
-0.0029242523,
-0.0011726343,
-0.0531904399,
0.0605396852,
0.0287302323,
0.0233130343,
-0.0148856668,
0.0215992127,
-0.0349494331,
-0.0457387865,
-0.0250356663,
-0.006784847,
0.003544875,
0.0289028455,
0.045150049,
0.0313796625,
-0.0132955089,
0.086581409,
0.0309077427,
-0.0648434088,
0.0244005099,
-0.0991836414,
0.0194812715,
0.0834939256,
0.0417498313,
-0.0221123938,
-0.0339129567,
0.001140301,
-0.0281747282,
0.0713861659,
0.0079837022,
-0.0117863975,
0.0227092709,
-0.0309704728,
-0.0143487677,
-0.0468453206,
0.0075083612,
0.0067141494,
0.0046464945,
-0.0805075467,
0.0166850649,
-0.0809145868,
-0.0562514029,
0.0571282953,
-0.0089399675,
-0.0297035966,
0.0240149163,
0.0097911777,
-0.0001654898,
-0.0031936585,
0.0157741047,
0.0291817356,
0.0207891371,
-0.0460438542,
0.0165809803,
0.0412722193,
-0.034249872,
-0.0192520116,
0.0285419282,
-0.0105874743,
0.0006135431,
0.0580899008,
-0.0198703464,
-0.0583876558,
0.0120430905,
-0.0790120363,
0.0472259782,
-0.0496333539,
-0.0207418967,
-0.0057154037,
0.0016202086,
-0.0068667638,
0.075176537,
0.0174719077,
0.0118753258,
-0.0093482779,
0.0019005229,
0.0148326056,
-0.0147987669,
-0.0417235605,
-0.0078103258,
-0.0340143144,
-0.0043556103,
0.0240905154,
-0.0283106621,
0.0042846054,
-0.001327795,
-0.0640309602,
-0.0283962581,
-0.0006494593,
0.0083363708,
0.0634615645,
-0.0178586934,
0.0281639975,
0.0104677128,
-0.0598613322,
-0.0070667872,
-0.0233913548,
-0.0149265463,
0.0733343959,
-0.0240193978,
-0.032753069,
-0.048300717,
-0.0198993608,
-0.0581563376,
0.0965445936,
0.0417637043,
-0.0164831039,
-0.0683955699,
-0.0099985721,
-0.0147969918,
0.0058045848,
-0.0240235757,
-0.0622354746,
0.0450562984,
0.0081049362,
-0.0051267296,
0.0522583984,
0.0172284804,
0.0660689622,
0.0494668707,
0.0136513561,
-0.0306240954,
-0.0144140553,
-0.0561968088,
0.0073022088,
-0.0425849184,
-0.0391954631,
-0.0028268106,
0.1070180312,
0.0020554836,
-0.049956318,
-0.0578189269,
-0.0015471681,
0.0218695607,
0.0030540146,
-0.0369714573,
-0.0208855569,
-0.0162146334,
0.0662494376,
-0.046241086,
-0.0693702176,
0.0988621339,
-0.0533821248,
-0.0098938104,
-0.0033715849,
-0.0091490429,
-0.0344607309,
0.0844855532,
0.016541725,
0.0141455038,
0.0240689628,
-0.0170914233,
-0.0174068194,
-0.0000687814,
-0.0035289896,
-0.0047462261,
0.0175742302,
-0.0424579345,
0.0042226831,
0.0341456868,
-0.0168245062,
-0.0345426574,
-0.0208369177,
-0.0069097718,
0.0592970252,
0.0636124983,
-0.0528737344,
-0.0549227148,
0.0045711761,
0.0482135639,
0.0607409403,
-0.0210713558,
-0.0103208972,
-0.0366695188,
0.0243001729,
0.0107306866,
0.0006249205,
0.0387605019,
-0.0376009569,
-0.0116948355,
0.0033975344,
0.0196450111,
0.0403859951,
-0.0698895007,
0.0069845361,
0.0039373958,
-0.0245634485,
-0.0009247195,
0.0132743381,
0.031538371,
-0.0515945368,
-0.0484024584,
0.12861076,
0.0200041365,
-0.0488452837,
0.0163186863,
0.0052161803,
0.0092161773,
0.032281924,
0.0551886223,
-0.0518269539,
-0.0064615849,
0.0063027246,
0.0354164876,
0.0590223782,
0.0118326936,
-0.0009221821,
-0.0685798675,
0.0409298837,
-0.0255233813,
0.0034174186,
-0.0292377919,
-0.0252512861,
-0.0221763384,
-0.0115654748,
-0.0276814196,
0.0121030407,
0.0243258588,
0.0142735178,
0.0391295515,
0.0130851148,
-0.0457231775,
0.0451962985,
-0.0231064465,
0.033726573,
0.0689223856,
0.0657822713,
-0.0003845436,
-0.0537890568,
-0.0174539723,
-0.0046231151,
0.0133380638,
0.0289490446,
0.0697721988,
0.0095353313,
-0.0306128617,
-0.0425874665,
0.0431849584,
-0.0466337502,
-0.0347477496,
-0.0990790278,
0.0231125001,
-0.023654677,
-0.0021153032,
0.0229843967,
-0.0081386203,
-0.0191493239,
0.0723741576,
0.0783459321,
-0.0360216089,
-0.012366524,
-0.0077746408,
0.0094300872,
-0.0692031607,
-0.0446676984,
0.0106787272,
0.0379907154,
-0.0563299768,
-0.0546317212,
-0.0139892278,
0.0280141868,
0.0116464971,
0.0252656396,
-0.0529379845,
0.0112350313,
-0.036832273,
-0.0605786704,
0.0529959686,
-0.0036102314,
-0.0066324119,
-0.0450285077,
-0.0630063713,
0.0529389121,
-0.0437864177,
-0.0291807391,
0.0050275829,
-0.0449597836,
-0.0390112996,
-0.0005812906,
0.0547113121,
0.0258957744,
0.0177698582,
-0.0165591873,
-0.0411537066,
0.0163122993,
0.0577234253,
-0.0177401863,
-0.0293008778,
0.0042770975,
0.0314234532,
0.0028295068,
0.0253321677,
-0.0246814359,
0.0308072101,
0.0241572,
0.0116451075,
-0.0118267201,
-0.028799491,
0.0441328436,
-0.011335751,
0.0492089242,
-0.0121145099,
0.0101609649,
-0.0596474595,
-0.0162625294,
0.0239340886,
-0.0125146369,
-0.0559050106,
0.0101329461,
0.0382095799,
-0.018099552,
0.0133986007,
-0.0217471831,
0.003908284,
0.0104938075,
-0.0350323506,
-0.0121240783,
0.0055007916,
-0.0114353085,
-0.043221131,
0.0203894544,
0.0090988539,
0.0186851881,
-0.0200590752,
0.0512135662,
0.0026586435,
0.0292712897,
0.0071234996,
0.0207527522,
-0.0207846947,
0.0144045996,
-0.0632770285,
-0.0356262513,
-0.0390159748,
-0.0108684227,
0.0235156026,
-0.0086302944,
0.0430898257,
0.0144115081,
-0.0025288924,
0.0286291409,
0.0494048558,
0.027745543,
-0.0276041962,
0.0172649268,
0.0210735444,
-0.0202152524,
-0.0676101819,
0.0765839741,
0.0400466807,
0.0271433052,
0.0053029973,
-0.0178095941,
-0.0316279754,
-0.0203459822,
0.0180564802,
-0.0024531453,
-0.0428719409,
0.012125331,
-0.0208599605,
-0.0709099621,
0.0096588796,
0.0197802912,
0.0111653227,
-0.0036867172,
0.0410395861,
0.0213193949,
-0.0092328116,
-0.045166187,
-0.0354094803,
-0.0014008307,
0.0071204482,
0.0139761642,
-0.032717403,
-5.890667413e-33,
-0.0464883074,
-0.0551566407,
-0.0213761441,
-0.0295846891,
0.0310120303,
0.0042671259,
-0.0702770576,
-0.0426467247,
-0.0014872977,
0.0311981458,
0.0131200748,
0.0234922785,
-0.0008706121,
0.0541212,
-0.012815957,
-0.0262671895,
0.0301512033,
-0.000534149,
-0.0344659016,
0.0213676579,
0.0688320994,
0.0384723842,
0.0834839717,
-0.0392932296,
0.0399767086,
-0.0101778526,
-0.0171302967,
0.064674668,
-0.0108735012,
0.0356579125,
-0.0371834338,
0.0179822948,
0.0183481928,
-0.0204666704,
-0.0185809713,
0.0117501086,
-0.0820765346,
0.0006581019,
0.0278423764,
0.0123642161,
-0.0203197002,
-0.035777431,
0.0592505559,
-0.0477567092,
0.0189343002,
-0.0047630719,
0.0554883741,
-0.012770961,
-0.0319490731,
-0.0223626774,
-0.0757331699,
-0.0106478538,
-0.0114854751,
-0.0671928748,
0.011805607,
0.0063879406,
-0.0023210084,
0.0218908694,
-0.0501572341,
-0.0169273503,
0.0013496854,
0.0519928709,
0.0613040477,
0.0115903271,
0.0193074569,
0.0689907447,
0.0246263891,
-0.007521417,
-0.0705824867,
-0.0158696789,
0.0140360249,
0.046815943,
0.0708307177,
0.0127616506,
0.0129127949,
0.0028665257,
-0.046028208,
0.0088678719,
-0.0196160264,
0.0162497424,
0.0412808843,
0.0085168714,
0.0127675338,
-0.0369060338,
-0.0299839769,
-0.0269831009,
-0.0460061952,
0.0082995817,
-0.0154774962,
-0.0174564198,
-0.0634347051,
-0.0043249968,
0.0040303338,
-0.0388445742,
0.030931564,
0.0364409238,
0.0079566818,
0.0253572222,
0.0043691732,
-0.0269102193,
-0.0120144906,
-0.0347125381,
0.0240829792,
0.0317498371,
0.0418693945,
0.0349198096,
0.0583461635,
0.046145808,
-0.0018954352,
0.0006882757,
-0.0877121538,
-0.0236087795,
-0.0303497426,
0.0528748669,
-0.0072098281,
0.0155105079,
-0.043993894,
0.002043121,
0.0134853162,
0.0082765706,
0.0129211806,
0.0328131616,
-0.017901836,
0.0351961702,
-0.0306736138,
-0.0284318831,
-0.0460541137,
0.0208096672,
0.0488149896,
-0.0471933819,
0.0044226209,
0.0172438305,
2.536e-7,
0.0785946846,
0.0604307167,
0.0506641455,
0.0100411316,
0.0404526331,
0.018518487,
0.0216409732,
0.0008063581,
-0.0190204382,
-0.04862177,
0.0094796289,
0.0305613093,
0.0581673384,
0.0294558126,
-0.0615163371,
0.0045199175,
-0.0953084752,
-0.0113697695,
-0.0332314484,
-0.0003453223,
0.0363539495,
0.0875054747,
0.0089421337,
-0.0256157424,
0.0466166697,
-0.0498590916,
0.0744473264,
-0.044348672,
0.0249978248,
0.0237437617,
0.0342175514,
0.0060365028,
-0.0214844868,
0.0067469045,
-0.0150917722,
0.0061571612,
-0.016382046,
0.0471842699,
-0.0392484292,
-0.0377328247,
-0.0251439624,
-0.0358545855,
-0.0038505399,
-0.0779992938,
0.0187980849,
-0.0631535277,
-0.0427605435,
0.0233476423,
0.0252975114,
-0.019240465,
-0.0232164711,
-0.0370901339,
-0.0121401642,
0.0117700556,
-0.0210529659,
-0.029080119,
0.0156744476,
-0.0895561948,
-0.0553755127,
-0.0149981193,
-0.0508051701,
0.0297737736,
-0.0379781984,
-0.058808811,
0.0622217618,
-0.0496488772,
0.0149998935,
2.167706173e-34,
-0.0033582349,
0.0360287353,
-0.05156314,
0.0145488884,
0.0413219221,
0.0140884854,
0.0190781653,
-0.0192104243,
0.0259663817,
-0.064222984,
-0.0458923914
] |
http://arxiv.org/abs/2402.00888v1 | http://arxiv.org/pdf/2402.00888v1 | Security and Privacy Challenges of Large Language Models: A Survey | 2024-01-30T04:00:54 | 2024-01-30T04:00:54 | Badhan Chandra Das, M. Hadi Amini, Yanzhao Wu | cs.CL | cs.CL, cs.AI, cs.CR | Large Language Models (LLMs) have demonstrated extraordinary capabilities and
contributed to multiple fields, such as generating and summarizing text,
language translation, and question-answering. Nowadays, LLM is becoming a very
popular tool in computerized language processing tasks, with the capability to
analyze complicated linguistic patterns and provide relevant and appropriate
responses depending on the context. While offering significant advantages,
these models are also vulnerable to security and privacy attacks, such as
jailbreaking attacks, data poisoning attacks, and Personally Identifiable
Information (PII) leakage attacks. This survey provides a thorough review of
the security and privacy challenges of LLMs for both training data and users,
along with the application-based risks in various domains, such as
transportation, education, and healthcare. We assess the extent of LLM
vulnerabilities, investigate emerging security and privacy attacks for LLMs,
and review the potential defense mechanisms. Additionally, the survey outlines
existing research gaps in this domain and highlights future research
directions. | Security and Privacy Challenges of Large Language Models: A Survey: Large Language Models (LLMs) have demonstrated extraordinary capabilities and
contributed to multiple fields, such as generating and summarizing text,
language translation, and question-answering. Nowadays, LLM is becoming a very
popular tool in computerized language processing tasks, with the capability to
analyze complicated linguistic patterns and provide relevant and appropriate
responses depending on the context. While offering significant advantages,
these models are also vulnerable to security and privacy attacks, such as
jailbreaking attacks, data poisoning attacks, and Personally Identifiable
Information (PII) leakage attacks. This survey provides a thorough review of
the security and privacy challenges of LLMs for both training data and users,
along with the application-based risks in various domains, such as
transportation, education, and healthcare. We assess the extent of LLM
vulnerabilities, investigate emerging security and privacy attacks for LLMs,
and review the potential defense mechanisms. Additionally, the survey outlines
existing research gaps in this domain and highlights future research
directions. | [
0.0557115413,
0.099706836,
-0.0121498322,
0.0555217564,
-0.0534904562,
-0.0034286568,
0.0630935431,
-0.0179601032,
0.0217647403,
-0.0504054278,
-0.0105214929,
-0.0509933271,
0.0234647691,
0.0083200624,
0.0392580293,
-0.0071716183,
0.0424343534,
-0.0605361052,
-0.0130569804,
-0.0127880406,
-0.0096876938,
0.0118278563,
0.0382650271,
0.0561029501,
-0.0049302531,
-0.0017965857,
-0.0257600993,
-0.0178894475,
0.0137671009,
-0.040210098,
-0.0059975623,
0.0570753515,
0.0072246478,
0.0628847629,
0.000001894,
-0.0280002356,
-0.0546085984,
0.0455601253,
-0.0173493531,
0.0238912739,
0.0537715033,
-0.0407597572,
-0.0133106196,
0.0162856523,
-0.0340303853,
0.0204713587,
0.0258707311,
0.0543457121,
0.07865832,
0.0742281452,
-0.0053206361,
-0.0313702263,
0.0207515452,
-0.0191574078,
-0.0098500913,
-0.0356373861,
0.0513085648,
-0.0493347161,
-0.0201969035,
0.0270607583,
0.0327893682,
0.0222928673,
-0.0081879515,
0.001739027,
-0.033012595,
0.0021636728,
-0.0252434555,
-0.0144541888,
0.0090276133,
0.0469934866,
-0.0164074805,
-0.0153821036,
0.0066809873,
0.0473427325,
0.0166667663,
0.0317146368,
-0.0118237725,
0.0540880561,
-0.0080094086,
0.0153170237,
0.0421079658,
-0.0292122923,
0.0228232387,
0.0158753842,
-0.0131583549,
0.0344118588,
0.0238251835,
-0.0249878652,
-0.0252352078,
-0.0332294554,
0.0083903056,
-0.0773992911,
0.0311533064,
0.0172529407,
0.029512573,
0.0262508597,
-0.0923102051,
-0.0530052558,
0.0419356599,
-0.0707376376,
-0.0027947621,
-0.004894597,
0.035524752,
0.0154556893,
0.022278443,
0.0166211389,
-0.0143290861,
0.0266200285,
-0.0086506605,
-0.0223881248,
-0.0304314438,
-0.0194308348,
0.0126086986,
0.0686545968,
-0.0450397767,
0.0122425128,
-0.05667172,
0.056894701,
-0.004679576,
0.0483098552,
-0.0454471894,
0.0038739562,
0.0159514043,
0.0063721156,
0.0061315917,
-0.0189483818,
0.0245312322,
-0.0104014222,
-0.0099656638,
-0.0531713814,
-0.0157669,
-0.0070664552,
-0.0157883503,
-0.0109197842,
-0.0350951627,
0.0742873102,
0.0188776013,
-0.0365316011,
-0.0489327908,
-0.1197914034,
-0.0373736359,
-0.0228832606,
-0.0112584392,
-0.0875307396,
0.0038760635,
-0.0214624852,
-0.0201809481,
-0.0624198839,
0.0232869592,
0.0527021363,
-0.0751538202,
0.0507144369,
-0.0154771181,
-0.000993029,
0.0149237141,
0.0435794629,
0.0464307778,
0.0383828692,
0.0204106849,
0.0324681029,
0.055679936,
-0.0001292818,
0.0234800242,
0.0294742007,
0.0331264772,
0.0033496355,
-0.0167114176,
0.0226750951,
0.0184247736,
-0.0481222644,
-0.0452626608,
0.0880949199,
0.0089589581,
-0.0261285044,
0.1078270897,
0.1122816429,
-0.0340603516,
0.0309758354,
-0.0333249532,
-0.0011059871,
0.0160549618,
-0.0084148627,
-0.0661468655,
-0.0300549865,
-0.0804416239,
-0.0331044309,
0.0093867118,
0.0336288847,
0.0128668342,
0.0015465205,
0.0050500007,
-0.0597038493,
0.0137885008,
0.0156060578,
0.0097742872,
0.0040306617,
0.05349347,
0.0644527376,
0.01724484,
-0.0562961549,
-0.0001911401,
-0.0579072423,
-0.0285280403,
0.0861387402,
0.0355512463,
-0.0084234839,
0.016371388,
-0.0335799344,
-0.0076521896,
0.0105469869,
0.0048007588,
0.0092331683,
0.0205520671,
-0.0043708053,
-0.0161730237,
0.0372817181,
0.0126197059,
-0.0142794186,
0.0231905673,
0.0314794518,
0.0309619904,
-0.0152173424,
-0.0369706117,
0.0197261833,
0.0026585497,
-0.0205141734,
0.0445920527,
0.025598105,
0.0262892283,
0.0496538393,
-0.0236734934,
0.0264758803,
0.0466566347,
0.0043068375,
0.016748419,
-0.0337122567,
-0.0203469042,
0.0098239649,
0.0236033928,
-0.0144363651,
0.0175665282,
0.0464661121,
-0.0346238203,
-0.026131466,
0.0051967911,
-0.0480713323,
0.0347162336,
-0.010501313,
0.0316393152,
-0.0021945201,
0.0375896804,
-0.0490167849,
0.0304023344,
-0.0405980423,
0.001242489,
-0.0310999677,
0.0090415636,
0.0019479801,
-0.0249167532,
-0.0491491891,
0.0149864219,
-0.0131324772,
0.0112159653,
0.013753322,
-0.0140640447,
-0.0121342707,
-0.0279218089,
-0.0416322649,
-0.0148717677,
-0.0132420892,
0.0428723544,
0.0107376501,
0.0323080644,
-0.0215500724,
-0.0409211963,
-0.0472389907,
-0.0547429696,
-0.0046577291,
0.0284614395,
0.0162926652,
-0.0005130734,
-0.0384692624,
-0.000849519,
-0.0285900328,
-0.0093907537,
0.0663049147,
-0.0212152209,
-0.038227085,
-0.0320869014,
0.0235218946,
-0.0250930078,
0.0414814353,
0.0115643879,
-0.0170609392,
0.0110514238,
0.0061818692,
0.0036338423,
0.0402135067,
0.0260537211,
0.0271204673,
0.0607157648,
-0.0220522434,
-0.0216944758,
-0.0299338568,
0.0043859947,
-0.0167710129,
-0.0369346887,
-0.0404868759,
0.0236138366,
0.0122045707,
0.0408617146,
-0.0221875235,
0.00813263,
0.0257507432,
-0.0124553284,
-0.0260951966,
-0.0189498961,
-0.0162789561,
-0.0177615527,
0.0197951626,
0.0112719703,
-0.0597628802,
0.0811581686,
-0.0550766401,
0.0524621084,
-0.0338074937,
-0.0628047213,
-0.0046224319,
0.0629488453,
0.0032848096,
0.060658969,
0.0358116999,
-0.045131322,
-0.0076399245,
-0.0069834106,
-0.0046678586,
-0.0600258298,
0.05234579,
-0.0459556021,
-0.0105160363,
-0.0159093179,
-0.0417460538,
-0.0495326295,
-0.0699604228,
-0.0144324554,
0.0022860758,
0.0181457419,
-0.0419149734,
-0.0023142351,
-0.0209384188,
0.0463519096,
-0.0119903078,
-0.0167584717,
0.0016181598,
-0.0227475613,
0.0469956324,
0.0324476622,
-0.0307593029,
0.0077845203,
-0.0017031917,
0.0169133171,
0.1060415432,
0.049661696,
0.0160326306,
0.01459214,
-0.0187355448,
0.0174275357,
-0.0310021248,
-0.0123517504,
0.0250086766,
0.0306582078,
-0.0407350361,
-0.0089478595,
0.0776161999,
0.0372489132,
-0.0512356795,
0.0000381283,
0.0496937335,
0.0012448751,
-0.0251513254,
-0.0046127941,
-0.0873134956,
0.004485846,
0.0185812917,
0.076687105,
-0.0402120352,
-0.0270442516,
-0.0297876801,
-0.097325854,
-0.0110311965,
0.0172363184,
-0.0222054478,
-0.0372286402,
-0.0453806892,
-0.0159040131,
0.0526794121,
0.0311999172,
-0.0466105342,
-0.0118162613,
-0.0207638312,
0.0101322737,
-0.0519196205,
-0.0601443686,
-0.0130135966,
-0.0309051741,
0.0452551097,
0.0708781108,
-0.0036353506,
-0.0245353542,
0.0158333723,
0.0239457469,
0.0256702844,
-0.0312730893,
-0.0054181544,
0.0365374945,
-0.0063538202,
-0.0057942476,
-0.0384628698,
0.0640974045,
-0.0099892057,
-0.0294928886,
0.0226031542,
0.0421891436,
-0.0506558903,
0.021415757,
-0.0275886003,
-0.0055386573,
0.0215311348,
0.060649544,
-0.0039865072,
-0.0401033834,
0.0000187067,
-0.0592992678,
-0.0011247572,
-0.0130367121,
-0.0062113404,
0.033904247,
0.0184580013,
-0.0433173701,
-0.0408343114,
-0.004654312,
-0.0036292025,
-0.0111111775,
0.0280711446,
0.0221556444,
0.0428873859,
-0.007382195,
-0.031711746,
0.0066718394,
-0.0438229665,
0.0353929065,
-0.0488798171,
0.013365156,
0.0174254458,
0.0046271901,
-0.0061222007,
0.0091695832,
-0.0952437893,
-0.0870500654,
0.0244182367,
0.0351305678,
0.0174960326,
0.0381948352,
-0.0281245448,
-0.0656654537,
0.0727568269,
0.0632134005,
-0.0488856323,
0.0125543103,
0.0132239508,
0.0005031447,
0.0074306503,
0.027605027,
-0.0416172519,
-0.0168690477,
0.056668587,
-0.0359642915,
-0.0568796545,
-0.0531316474,
-0.0104689328,
0.0614968762,
0.0166142676,
0.027375726,
0.0444577783,
-0.0059821662,
0.0399522185,
-0.0548923463,
-0.0193572454,
-0.0101007558,
-0.000375617,
0.0289043132,
-0.0057929372,
0.0482502691,
-0.0117130792,
0.0208810512,
-0.0018913771,
-0.0093930494,
0.0308162142,
-0.0005197221,
0.012223525,
0.0347301513,
0.0150357261,
-0.0434542485,
0.0497607738,
-0.0145371156,
0.0382838696,
-0.018192986,
0.0323805958,
-0.0237440523,
-0.0275798645,
-0.0141003458,
0.0563561954,
-0.0761087015,
-0.0306604505,
-0.0084247198,
0.0543521196,
0.0028161083,
0.0150490729,
0.0042245984,
-0.0457681455,
0.0022783957,
-0.0320026688,
-0.017063411,
-0.003186777,
-0.0287481621,
-0.001977033,
-0.0065642288,
-0.015499983,
0.0381857194,
0.0907011628,
0.0399343409,
0.0177682061,
0.0069342391,
-0.0604338162,
-0.0275358278,
-0.0111969663,
0.0189888105,
0.0264124107,
-0.0604755953,
-0.020673262,
-0.047798451,
-0.077966243,
-0.0052356287,
-0.0201525502,
0.0231102519,
-0.0024904676,
0.0148341423,
0.0228207037,
0.0138236945,
0.0030877891,
-0.0906978324,
0.0169935618,
-0.0084765507,
-0.0503121577,
0.009697807,
-5.599317465e-33,
0.0081670796,
-0.0412144065,
-0.0541044809,
0.0100924177,
0.0251105838,
0.0115495203,
0.00826695,
0.0124343662,
-0.0262208991,
-0.0494684093,
-0.0265672691,
-0.0022479277,
0.005222938,
0.0768555105,
-0.0253449883,
-0.0191932395,
0.0215378031,
0.0002202048,
0.0089322673,
0.0357561968,
0.0762075707,
0.0118712895,
0.0281344242,
-0.0877296552,
0.0559032746,
-0.0123043191,
-0.0302820373,
0.0211595595,
-0.0333619937,
0.0670887455,
-0.0293739699,
0.013821275,
0.0114426985,
-0.0010368584,
-0.0010712247,
0.078546606,
-0.0378130116,
-0.013076351,
-0.0244436786,
0.0015161505,
0.0218727291,
-0.0815809146,
0.0583337583,
-0.0421011746,
-0.0089705307,
0.0082239341,
-0.0038827609,
-0.023831157,
-0.0202780776,
-0.0129582966,
0.0165102705,
0.0012460662,
-0.0156259034,
0.014413666,
0.0317162573,
-0.0110434545,
0.0199963767,
-0.0210759155,
-0.0391542241,
0.0357837416,
0.0709247068,
0.0096656987,
0.0511346981,
0.0226952191,
0.0149635915,
0.048548054,
0.0207809359,
-0.0196277741,
-0.0260489434,
-0.000600316,
-0.0399107635,
0.0264212359,
0.0331875123,
0.0449559093,
0.0114488294,
-0.0143864909,
-0.0672678798,
0.0102513768,
-0.0027950881,
0.0025292686,
0.055279091,
-0.037499208,
-0.0740767494,
-0.0014340477,
-0.0170598943,
0.0117517076,
-0.0343570411,
-0.0044880873,
0.0012937854,
-0.0034192577,
-0.0375410914,
0.003924795,
0.0131492652,
-0.0721321926,
-0.0440565012,
0.0191730876,
0.0483881868,
-0.0259762984,
-0.0044462481,
0.0134404162,
-0.0210146438,
0.029967729,
-0.0050295703,
-0.0697965696,
0.0094354581,
0.0079869153,
0.0377803743,
0.0079005025,
-0.0815278441,
-0.0081314659,
0.0093391687,
0.0489498265,
0.022481164,
0.0346676596,
0.047661826,
-0.0238375813,
0.0189089179,
0.0226666499,
0.0267615151,
0.0363742672,
-0.0327873453,
-0.0016376236,
-0.0236078575,
0.0107613541,
0.0088940542,
-0.0011656884,
-0.0123347808,
0.0379533768,
0.0795212239,
-0.0109236399,
0.0186788235,
-0.0308316145,
2.745e-7,
0.0308358185,
0.0178840943,
0.0076997653,
-0.0175323486,
0.0141130155,
0.0588268228,
0.0049018133,
0.0634930581,
-0.0109887039,
0.0207335409,
-0.0173880216,
-0.0145162344,
0.057181038,
0.0480447896,
-0.0829570219,
-0.0039146412,
-0.048897218,
-0.0072530662,
-0.0847726539,
0.0214567278,
0.065242976,
0.1189500317,
0.0122274673,
-0.0312965997,
-0.0407592617,
-0.069380179,
0.0111605786,
-0.0136332298,
-0.0436038934,
-0.0209125392,
0.0314490274,
0.0709828585,
0.0170999169,
0.0048402906,
-0.0345454812,
0.0098617123,
-0.0507432297,
0.0602502972,
0.0135098463,
-0.0114879031,
-0.0159639474,
-0.0807813779,
-0.0100498395,
-0.0310081225,
0.0703516826,
-0.0079678176,
-0.0312211588,
-0.0220329482,
0.0395944268,
-0.0619725026,
0.0104595562,
-0.0290409327,
-0.0060768719,
0.0194518361,
0.0153469862,
-0.0032383727,
0.037109036,
-0.0724352375,
-0.0275550615,
-0.0188745968,
0.0028198292,
-0.0180148594,
0.0043543675,
0.0339737087,
-0.0037123093,
-0.0130301379,
0.0287535004,
2.769901139e-34,
-0.0247771461,
-0.0308566205,
-0.0319826826,
0.0075508347,
0.0406158492,
-0.0008339744,
-0.0009235516,
0.0164241996,
-0.0039041368,
-0.0798424929,
-0.0225806274
] |
http://arxiv.org/abs/2401.16727v2 | http://arxiv.org/pdf/2401.16727v2 | Recent Advances in Hate Speech Moderation: Multimodality and the Role of Large Models | 2024-01-30T03:51:44 | 2024-02-02T04:07:25 | Ming Shan Hee, Shivam Sharma, Rui Cao, Palash Nandi, Tanmoy Chakraborty, Roy Ka-Wei Lee | cs.CL | cs.CL | In the evolving landscape of online communication, moderating hate speech
(HS) presents an intricate challenge, compounded by the multimodal nature of
digital content. This comprehensive survey delves into the recent strides in HS
moderation, spotlighting the burgeoning role of large language models (LLMs)
and large multimodal models (LMMs). Our exploration begins with a thorough
analysis of current literature, revealing the nuanced interplay between
textual, visual, and auditory elements in propagating HS. We uncover a notable
trend towards integrating these modalities, primarily due to the complexity and
subtlety with which HS is disseminated. A significant emphasis is placed on the
advances facilitated by LLMs and LMMs, which have begun to redefine the
boundaries of detection and moderation capabilities. We identify existing gaps
in research, particularly in the context of underrepresented languages and
cultures, and the need for solutions to handle low-resource settings. The
survey concludes with a forward-looking perspective, outlining potential
avenues for future research, including the exploration of novel AI
methodologies, the ethical governance of AI in moderation, and the development
of more nuanced, context-aware systems. This comprehensive overview aims to
catalyze further research and foster a collaborative effort towards more
sophisticated, responsible, and human-centric approaches to HS moderation in
the digital era. WARNING: This paper contains offensive examples. | Recent Advances in Hate Speech Moderation: Multimodality and the Role of Large Models: In the evolving landscape of online communication, moderating hate speech
(HS) presents an intricate challenge, compounded by the multimodal nature of
digital content. This comprehensive survey delves into the recent strides in HS
moderation, spotlighting the burgeoning role of large language models (LLMs)
and large multimodal models (LMMs). Our exploration begins with a thorough
analysis of current literature, revealing the nuanced interplay between
textual, visual, and auditory elements in propagating HS. We uncover a notable
trend towards integrating these modalities, primarily due to the complexity and
subtlety with which HS is disseminated. A significant emphasis is placed on the
advances facilitated by LLMs and LMMs, which have begun to redefine the
boundaries of detection and moderation capabilities. We identify existing gaps
in research, particularly in the context of underrepresented languages and
cultures, and the need for solutions to handle low-resource settings. The
survey concludes with a forward-looking perspective, outlining potential
avenues for future research, including the exploration of novel AI
methodologies, the ethical governance of AI in moderation, and the development
of more nuanced, context-aware systems. This comprehensive overview aims to
catalyze further research and foster a collaborative effort towards more
sophisticated, responsible, and human-centric approaches to HS moderation in
the digital era. WARNING: This paper contains offensive examples. | [
0.0821358562,
0.0718919709,
-0.0435819514,
0.0167332683,
-0.0145503618,
0.0147431223,
0.0649869218,
0.0193708166,
0.0095656039,
-0.0818435103,
-0.0223724209,
-0.0350326076,
-0.0202366281,
0.0103073651,
0.0106640002,
-0.0621486306,
-0.00723906,
0.010496757,
0.0531412251,
0.0111185741,
0.0787359998,
0.0196417067,
0.006817807,
0.0503481701,
-0.0747897848,
0.0007149727,
0.0490658209,
-0.0049975673,
0.0128874481,
0.0065017687,
0.0076681124,
0.0190478638,
0.0369458124,
0.0606768839,
0.0000020204,
-0.0325608663,
0.0288388226,
-0.0129908193,
0.0203284509,
-0.0281066559,
0.0793968663,
-0.0453180335,
-0.0284792893,
0.0121036423,
-0.0239185374,
0.0333640873,
0.0469317734,
0.0340390056,
0.0524337813,
0.0433259569,
-0.0209001489,
-0.0298989564,
0.0181905925,
0.0335964113,
-0.0293585062,
0.0448416807,
0.0270625073,
-0.0381773189,
-0.0408234932,
-0.0040233876,
-0.0159921069,
0.0257388502,
0.0089851869,
-0.0017485229,
-0.0202818625,
-0.0533595383,
-0.0176275466,
-0.0585971698,
-0.0190157779,
0.0433156863,
-0.0095017366,
0.0004158278,
0.0224014409,
0.0107059479,
-0.0090241544,
0.0235471316,
-0.0408258215,
0.0488329418,
-0.0264793206,
-0.0001332143,
-0.0210322663,
-0.0009366887,
0.0494160727,
0.0754106343,
0.0539779589,
0.0191239938,
-0.0139049366,
-0.0317205414,
0.0067083845,
-0.0156744905,
-0.0084842844,
0.0005841269,
0.0365113318,
0.0306045953,
-0.0244768653,
0.0157540534,
-0.033409059,
-0.0539594218,
0.0021357974,
-0.0809236765,
0.0464538969,
0.0310539249,
-0.0335243866,
0.007554845,
0.045147974,
0.0192946065,
-0.0329496227,
0.0044799903,
0.0297874548,
0.0149142882,
-0.0104208831,
0.0119720139,
0.0069856043,
-0.0325877964,
-0.0037860151,
0.0144550223,
-0.0167310927,
-0.0196916237,
-0.0821148977,
-0.0051747416,
-0.0460953005,
0.0023567129,
0.014910494,
0.0086665871,
-0.0239734799,
0.0375217013,
-0.0384317823,
0.0207224824,
-0.0367645957,
-0.0594721064,
0.0439673997,
0.0506692715,
-0.0373772271,
-0.0242779814,
0.0297956057,
0.0758019239,
0.0154941408,
-0.0018878112,
-0.0234615151,
-0.052463755,
-0.0394060835,
-0.04564723,
0.0203649811,
-0.0430297107,
0.0146113811,
-0.0218959805,
-0.0297436006,
-0.0649978667,
0.0163781997,
0.0111105246,
-0.0196862072,
0.078236483,
0.017168723,
0.0187502056,
0.0588215925,
0.0419692732,
-0.0050510587,
0.0406148136,
0.0108462237,
0.0126662124,
0.0581002347,
-0.0091267461,
0.0057645203,
0.0500591695,
0.0183867216,
0.0192224998,
-0.0699945316,
-0.0225352235,
-0.0565901399,
-0.0228856262,
0.0127912136,
0.0687043667,
-0.0063704108,
-0.0359176509,
0.0726053491,
0.0820413977,
-0.0679741204,
0.0189593621,
0.0712306425,
0.0126616424,
-0.0142018152,
0.0060545448,
-0.0347967967,
-0.0111250496,
-0.0018712305,
-0.0564081892,
0.0316643417,
0.0628389865,
-0.0224852264,
-0.0468433537,
-0.0218480248,
-0.0243637934,
0.0625987798,
-0.0043761786,
0.034186475,
0.0108949477,
0.017518552,
-0.0332335904,
0.052649647,
-0.0120963566,
0.0065991352,
-0.0661096871,
0.0129270591,
0.0694645047,
0.0151119763,
-0.007019585,
0.0406066515,
-0.0723370835,
0.0088444017,
0.0657523051,
-0.003447579,
-0.0164004732,
-0.0486688428,
-0.0383496173,
-0.0247062482,
0.0183593314,
0.009210784,
0.0219474398,
0.012670869,
0.0173018966,
-0.029654365,
0.0568007976,
-0.0163428169,
-0.0096872635,
0.0304423999,
-0.0208009649,
-0.0498583876,
0.0359800234,
0.0480627827,
0.0388911553,
-0.0280184448,
-0.047647506,
0.0016773879,
-0.0001833586,
-0.0370979533,
0.0154784564,
-0.0309408884,
0.0691989064,
0.0503930561,
-0.0391921401,
-0.0105438987,
0.0320245959,
0.04228466,
0.0475981794,
0.0730632618,
-0.0069459537,
-0.0171347708,
-0.0457171351,
-0.0315756276,
-0.0659014508,
0.0018012236,
-0.0044715619,
0.0381367803,
-0.0219253208,
-0.0151953101,
0.0559502803,
-0.0640618056,
0.0244621858,
0.0438706577,
-0.0552794561,
-0.0114198346,
-0.0237316024,
0.0516285785,
0.0109038213,
-0.0134578068,
-0.0077082394,
0.0017659833,
-0.0630449355,
-0.043462012,
-0.029720109,
-0.0266381148,
0.0007311358,
-0.0244664457,
-0.0005113458,
0.0169162136,
0.001698994,
-0.0438877419,
0.0593846664,
0.0352105983,
-0.0149752479,
-0.0482468307,
-0.0350452624,
-0.0104441615,
-0.0075654201,
-0.025990529,
0.0234931894,
0.0076259351,
-0.0749813393,
0.0365413949,
-0.0368511193,
-0.0671472102,
0.0309522115,
-0.0180958174,
-0.0438590534,
0.0479747131,
0.0113435164,
0.0122854831,
0.0207986571,
0.0454948582,
0.1047979072,
0.0625398383,
0.0233477335,
-0.014128048,
-0.016665129,
0.0167355109,
0.0211148728,
-0.0273798518,
-0.0255303439,
-0.0153192924,
0.0585998036,
0.0499482192,
0.0077332971,
-0.0417727754,
0.0662994608,
-0.093287617,
-0.0771680102,
-0.0275596697,
-0.098299861,
-0.0139653888,
0.008676447,
-0.0236282311,
-0.0691896752,
0.0570732579,
0.0105486093,
0.0308976304,
0.0224667806,
-0.0560885929,
-0.0012534758,
0.0397495329,
-0.0054962747,
0.0772783756,
-0.0326773636,
-0.0019912114,
0.0096677989,
-0.032473214,
0.0447892658,
-0.0364827625,
-0.0097355889,
-0.0176890865,
-0.0013692291,
0.0236767717,
-0.0338747874,
-0.0195514336,
-0.0281534661,
-0.0521304421,
-0.0199050438,
0.0243875142,
0.017145684,
-0.0227514915,
0.030732302,
0.0603218898,
-0.0308938418,
0.0340774804,
0.0082127135,
0.0248444788,
-0.0253288615,
-0.0026112874,
0.0133241424,
0.0211157594,
-0.0079699215,
-0.0066383788,
0.074317269,
-0.019943824,
-0.0351706296,
0.0217537582,
0.0258974507,
0.0317728892,
0.060495995,
0.011860271,
0.0176685061,
0.0472270399,
-0.0224029869,
0.0196762998,
0.0591611341,
0.0157234557,
-0.0354338661,
-0.0139952153,
-0.0256620999,
-0.0246141944,
-0.0053903945,
-0.0333140641,
-0.0823126733,
0.0011659384,
-0.0157231614,
0.00877403,
-0.1003239825,
-0.0303208623,
0.0205887053,
-0.0334571674,
0.0390940942,
-0.0048775645,
-0.0262348298,
-0.0397521481,
0.0122061837,
0.0025276789,
0.032037504,
0.0465863533,
-0.0360472463,
0.026316246,
-0.0317498259,
0.0154957892,
-0.0369338281,
-0.0000158208,
-0.0172312427,
-0.0287403762,
0.0408672132,
0.028463142,
0.049037952,
-0.0366428196,
-0.0084700724,
0.026011914,
0.0215176195,
-0.0152730951,
-0.0386563875,
0.0098261684,
0.006134857,
-0.0301203858,
-0.0497731492,
0.0404645652,
0.0573031418,
-0.0158511586,
0.0142012592,
0.0640919134,
0.0161041841,
-0.0024121506,
-0.0705641583,
0.0370992683,
0.0274413712,
0.0425249524,
0.0220510047,
-0.0845433921,
0.027518291,
-0.0672913119,
0.0030280456,
0.0088951373,
-0.1058605835,
-0.0087503474,
0.007192268,
-0.0044230134,
-0.0200391859,
-0.0171587244,
-0.0226436183,
0.0001086112,
0.0549150705,
-0.013992249,
0.0108780628,
-0.0166611336,
-0.0036088396,
-0.011070664,
0.0093073584,
0.0538322292,
-0.013726607,
0.0227913596,
-0.0422750711,
-0.0410865434,
-0.0014994737,
0.0416354239,
-0.0195431206,
-0.0541700572,
0.0574855022,
0.05519595,
0.0155218095,
0.0041139051,
0.0174257979,
-0.081555292,
0.0588675402,
0.0396907069,
-0.0073112943,
0.0181392971,
0.0571559332,
0.0188400429,
-0.0063021341,
0.0174110383,
0.005558637,
0.0423773415,
0.0484736599,
-0.0341768265,
-0.0518869571,
0.0033546728,
-0.0067624021,
-0.0326997451,
-0.0079861889,
-0.0254352279,
-0.0506647043,
-0.0217500515,
-0.0449934043,
0.0058155167,
-0.0385645553,
-0.0240699537,
0.0200114269,
0.0040709008,
-0.0017802272,
-0.0325843394,
-0.0129389409,
-0.0324457139,
0.0220442563,
-0.0573802106,
0.0536344349,
0.0377249196,
0.0064970637,
-0.0012259955,
0.0242804531,
0.0052366583,
0.0887257829,
-0.0099817682,
-0.0311168861,
0.0023620704,
0.0476805903,
0.0216447655,
-0.0111783901,
0.0226032734,
-0.0193646308,
0.0140505787,
-0.0271119196,
0.0012443669,
-0.0035723648,
0.0131535297,
-0.0143937115,
0.0214864481,
0.0178620275,
-0.0372556895,
-0.0061414693,
0.0077806236,
-0.0192974657,
-0.0411490425,
0.0168414544,
0.0160079785,
-0.0255380999,
0.0264080483,
0.0024610807,
-0.050973732,
0.0545431077,
0.0388598852,
-0.0547775514,
-0.009010178,
-0.022825934,
0.0216706824,
0.0216628555,
0.0045836475,
0.0362502858,
-0.0640768334,
-0.02904962,
-0.0165639985,
-0.0517764352,
0.0197286382,
-0.0434080623,
-0.0138985887,
0.0478899404,
0.0102163497,
0.0186083578,
0.0016268803,
0.002699849,
-0.0165641755,
-0.0353691466,
0.0119287949,
-5.801072335e-33,
-0.0010468558,
-0.0605533049,
-0.011322761,
-0.0115648666,
-0.0027621279,
-0.0152378799,
0.0430437662,
0.0186823476,
-0.0026632331,
0.0343507528,
0.0064108851,
-0.0057491157,
-0.0065570571,
-0.0206674244,
0.0019951055,
-0.0187846422,
0.0315448008,
-0.0475935377,
-0.0225788653,
-0.0019944441,
0.0420815796,
-0.0248295609,
-0.0178276207,
-0.0974369273,
0.0424664915,
0.0155441416,
-0.0477127135,
0.0199849382,
-0.004035987,
0.0861087069,
-0.0275829379,
0.0912172422,
0.03763083,
-0.0042134924,
0.0055882614,
0.0893748775,
-0.0290419981,
0.0188129786,
0.0498627014,
-0.0046737823,
0.0155418133,
-0.0203726087,
0.0370760374,
-0.0132623846,
-0.0509000197,
0.0073543228,
0.0379785933,
-0.0294894073,
0.0186551958,
-0.0732872337,
-0.0733630657,
0.0078000552,
-0.0406680442,
-0.0157879274,
0.0052466467,
0.0313212983,
0.0195805486,
-0.004925468,
-0.0748746842,
0.0491523966,
-0.0155238714,
-0.048956044,
0.0216292664,
-0.0110866753,
0.0037574123,
0.0566954724,
0.0116113089,
0.0314237177,
-0.0500188246,
0.0048958678,
0.0247455165,
0.0358213633,
0.0342008695,
0.0369636528,
-0.0315993726,
-0.0916885063,
-0.0332820788,
-0.0285963882,
0.0228504185,
0.0158239044,
0.024865346,
-0.0116867209,
-0.0652119145,
-0.031409435,
-0.0250591058,
-0.0936789215,
-0.0355637185,
-0.0196986087,
0.025566278,
-0.0047613052,
-0.0112673193,
-0.0325244293,
0.0370216332,
-0.0482111536,
-0.016467806,
0.0242758635,
0.009956141,
0.0248084981,
0.0456970856,
0.0241446737,
-0.0293101612,
0.007155458,
0.0116244182,
-0.0449074432,
0.0371384919,
-0.0120952139,
0.0381769948,
0.0605089143,
-0.0396197662,
0.0182453431,
0.0116067128,
0.0554098152,
0.0186093803,
0.0301163141,
0.0202738103,
0.0442044884,
-0.0124217365,
-0.0315339155,
0.0487485304,
-0.0300601069,
-0.0139203602,
-0.0079249404,
-0.0608708002,
0.0049338802,
-0.0483147353,
0.0075284359,
-0.026312707,
-0.0000048709,
0.0511592627,
-0.0780583546,
0.0031653761,
-0.0536739379,
2.625e-7,
-0.0180570539,
0.0345529541,
-0.0086468784,
0.0043522855,
0.0427875929,
0.0373719484,
0.0268205982,
0.0022383544,
-0.0091036521,
-0.022007525,
-0.0047828583,
0.0465263948,
0.0258047935,
0.0460273363,
0.0203925036,
-0.012149008,
-0.056977734,
0.0023500333,
-0.0579956584,
0.0447537936,
-0.0102812387,
0.0466293618,
0.0380596593,
-0.0539759025,
-0.0410548933,
-0.0237327032,
0.0173467286,
-0.0565156974,
-0.0158489309,
-0.0371317901,
0.05376973,
0.0127666052,
-0.0017068234,
-0.0243284442,
0.0196806304,
0.0045521092,
0.0242073871,
0.0667460039,
0.0331817344,
0.0207591988,
-0.0034352615,
0.0125088589,
-0.0098541016,
-0.0134537034,
0.0284572728,
0.0237617977,
-0.0276083332,
-0.0644981787,
0.0056457752,
-0.0583801195,
0.0312193576,
-0.0308212712,
-0.0355979688,
-0.0275255349,
0.0250020549,
0.0013474443,
0.0330065712,
-0.0539700016,
-0.0155008417,
-0.0170081835,
-0.0238395259,
-0.0468869023,
-0.0286549926,
-0.0206228439,
0.0090993596,
0.0155929774,
0.031039698,
2.262445278e-34,
0.0376238897,
-0.0477843173,
0.0139821377,
-0.0440939963,
0.0249851625,
0.0392857045,
-0.018781215,
-0.0343714207,
0.0531324409,
-0.0071486933,
-0.0212741774
] |
http://arxiv.org/abs/2401.16713v1 | http://arxiv.org/pdf/2401.16713v1 | Prospects for inconsistency detection using large language models and sheaves | 2024-01-30T03:17:45 | 2024-01-30T03:17:45 | Steve Huntsman, Michael Robinson, Ludmilla Huntsman | cs.CY | cs.CY, cs.CL, math.AT | We demonstrate that large language models can produce reasonable numerical
ratings of the logical consistency of claims. We also outline a mathematical
approach based on sheaf theory for lifting such ratings to hypertexts such as
laws, jurisprudence, and social media and evaluating their consistency
globally. This approach is a promising avenue to increasing consistency in and
of government, as well as to combating mis- and disinformation and related
ills. | Prospects for inconsistency detection using large language models and sheaves: We demonstrate that large language models can produce reasonable numerical
ratings of the logical consistency of claims. We also outline a mathematical
approach based on sheaf theory for lifting such ratings to hypertexts such as
laws, jurisprudence, and social media and evaluating their consistency
globally. This approach is a promising avenue to increasing consistency in and
of government, as well as to combating mis- and disinformation and related
ills. | [
0.0345429853,
0.0931687057,
-0.0023886687,
0.0076094279,
-0.079120025,
-0.0103768278,
-0.0234444961,
0.0290871821,
0.0002989458,
-0.0345398821,
0.040428225,
-0.0432500802,
0.0037745426,
0.0093938764,
-0.0082514714,
-0.0323576406,
0.0592006594,
-0.0140213929,
0.0278969798,
-0.0213991124,
0.0257116109,
-0.0120098246,
-0.0005914287,
-0.008740169,
-0.0210871324,
-0.0034099065,
-0.0032681173,
-0.0250398628,
0.0059500658,
-0.0102904439,
-0.0244557355,
0.0300909448,
-0.025345942,
0.0868043751,
0.0000016827,
-0.0385387875,
-0.0054693199,
0.043007832,
-0.0139946872,
0.0286765043,
0.0012187795,
-0.0390358008,
-0.0292280782,
-0.019761838,
0.0134543413,
0.0436192565,
0.0376508236,
0.0140949134,
0.0710196346,
0.0462523326,
0.0049314182,
-0.0270507839,
0.0497303531,
-0.0232713781,
0.0226570945,
-0.0483830199,
0.0464682542,
0.0745717958,
-0.0226372052,
0.024868371,
0.006807629,
0.0291873608,
0.0033658508,
-0.0492917895,
-0.0388271473,
0.0120193623,
-0.0801293999,
0.0280196182,
0.0218866896,
0.0498548895,
0.0457579643,
-0.0212500915,
0.0455633067,
0.0089882389,
-0.015941998,
0.0767059997,
-0.0106452769,
0.070501551,
-0.0326731242,
0.0002335076,
0.0612055734,
-0.0179153085,
0.0483384356,
0.0246937331,
-0.0301495474,
0.1108456329,
0.0022509336,
-0.0623648055,
0.0056694453,
-0.0260450691,
-0.0259611961,
-0.0929529071,
-0.0003672067,
-0.0062043653,
-0.045817215,
0.0446777195,
-0.0356805883,
-0.0012892773,
0.0653317869,
-0.1179061383,
0.0288714468,
0.0233952068,
-0.0501977876,
0.0597392544,
0.0376848355,
0.0263958462,
-0.0266766269,
-0.0196854901,
0.0316120535,
-0.0219191145,
-0.0344970748,
0.013936149,
-0.0047724736,
0.0208800845,
-0.0437613465,
0.007620512,
-0.0412261747,
-0.0011012675,
-0.0840249136,
0.0474455729,
-0.0279106535,
-0.0609330796,
0.0009430551,
-0.0214148443,
-0.061084725,
-0.0043424536,
-0.0143152392,
0.0237302352,
-0.0431793779,
-0.024145944,
0.0430966467,
0.0032794366,
-0.0371349603,
0.0221092496,
-0.0066259056,
0.0293974001,
0.0456685498,
0.0336707197,
-0.0118976869,
-0.06312491,
-0.0754169002,
-0.0645484254,
-0.0531217679,
-0.0860812142,
-0.0226934124,
-0.0294523742,
0.0259503536,
-0.0460909493,
-0.0146462247,
0.0088268891,
-0.048541192,
0.0247613788,
-0.0350424759,
-0.0057630893,
-0.0026640533,
0.0001696584,
-0.0355116501,
-0.0095577696,
-0.027890753,
0.0243990365,
0.039694801,
0.0355107188,
0.0152610065,
-0.026592996,
-0.0139164608,
0.0566300377,
-0.0072378921,
0.0203744583,
0.0130433477,
-0.0075448756,
-0.0435489826,
0.0919207111,
-0.0153855011,
-0.0455367565,
0.0385028496,
0.0132464776,
0.0150179528,
-0.007691904,
0.0320461728,
-0.0073407521,
0.0320291221,
0.0479023345,
-0.0320134535,
-0.0024568331,
-0.063106142,
-0.0040951292,
0.0279746167,
0.0124031221,
-0.0172281861,
-0.0394313,
-0.0170990136,
0.0297222603,
-0.0571670234,
-0.020097388,
0.0138400029,
-0.0216840357,
0.0289664734,
0.0442876071,
0.0536859185,
-0.0296231974,
0.0023339388,
0.0022409447,
0.0500247814,
0.0227269121,
0.01812982,
-0.0070232837,
0.0032633389,
-0.0047897841,
0.0321281068,
0.0125416694,
0.0248885136,
-0.0199760869,
-0.0125668859,
0.0159264952,
0.0052831792,
-0.0095683383,
0.0401652791,
-0.0016321403,
-0.0154765239,
0.0253737122,
-0.0074408627,
-0.0459126793,
-0.030449016,
0.0074767782,
0.0014274369,
0.0311510731,
-0.0412595756,
-0.0113253882,
0.0477189943,
0.0117146159,
-0.050589025,
-0.0176816527,
0.0316717587,
0.0251152776,
-0.0080360873,
-0.0175105836,
-0.0062251957,
0.0705601051,
0.0749422535,
-0.0501915663,
-0.0249536335,
-0.005953582,
-0.0442110337,
0.0133324899,
0.0077419467,
-0.0509223379,
-0.0369211473,
-0.0428236425,
0.0415548496,
-0.0554285608,
-0.040913213,
0.0109541202,
0.0360702798,
-0.020533748,
0.0439735129,
0.0133206323,
-0.0280621629,
0.044963304,
-0.0066114296,
-0.0262985639,
0.0425993055,
0.0233257338,
0.0221722014,
0.0369981229,
-0.025832735,
-0.0218348578,
-0.0496515669,
-0.0244400818,
-0.0045729517,
-0.0246319585,
0.0158892125,
0.0085293613,
-0.0159713514,
0.0136286048,
0.0023622268,
0.005945602,
-0.038941469,
0.0045168814,
0.0035945782,
0.0316985846,
-0.0016138244,
-0.03191543,
0.0000567972,
-0.0206633732,
-0.0276916344,
0.0464189686,
0.0093065053,
-0.0131369531,
-0.0741335154,
0.00031527,
-0.0724292099,
0.0315298326,
-0.0184168052,
0.0364648588,
-0.0072853132,
0.0217448734,
-0.010426146,
-0.0148252081,
0.0186024066,
0.0469672121,
0.0567974411,
0.0182019044,
-0.0041020014,
-0.0553734042,
-0.0197527222,
-0.0075645256,
-0.0301402621,
-0.029905403,
0.0371252634,
0.0075909547,
-0.0006078081,
0.004800579,
-0.024310438,
0.0275230967,
-0.0514134727,
0.0150919724,
-0.0052529993,
-0.0504991598,
-0.0058576972,
0.056407854,
-0.018988708,
-0.0488781966,
0.0255232155,
-0.0291378461,
0.1098455042,
-0.0557792522,
0.0026023411,
-0.0297861714,
0.0233241785,
-0.0039624991,
0.0306600798,
0.0418483391,
0.0081130872,
0.0228054207,
-0.047546342,
0.0775173306,
-0.0277936105,
0.0156136239,
-0.0513979793,
0.0387008972,
0.0206889473,
-0.0667048246,
-0.0225377157,
-0.0622784644,
0.0160040837,
-0.0024039084,
-0.0128170187,
-0.0043336991,
-0.027781887,
-0.0000942951,
0.0041462472,
0.0006824009,
0.0260113813,
-0.0177381616,
-0.0273238234,
0.0431521498,
-0.0109108556,
-0.0519777499,
0.0263178032,
-0.0279195644,
-0.0502750278,
0.0517730117,
-0.0035645303,
0.0506779067,
-0.0129661784,
-0.0006403646,
-0.0235761404,
-0.0416794866,
-0.0071777184,
0.0231243577,
-0.0262076855,
-0.0276687834,
0.0402200148,
0.0892771035,
0.0402310751,
0.0368834957,
-0.002645181,
0.0111273089,
-0.048711095,
0.0167568102,
-0.0154850101,
-0.1050988659,
0.049234271,
0.0121387877,
-0.0327399299,
-0.0365816616,
-0.0190986414,
0.031822037,
-0.0529696159,
0.0392940976,
0.0362989791,
0.0149441492,
-0.0284444243,
-0.0052221236,
0.0420530811,
-0.0024298751,
0.003033869,
0.029953083,
0.0109244557,
-0.010701579,
0.0092198439,
-0.0148676354,
-0.0255014487,
0.0465590917,
-0.0564916469,
0.0462642685,
0.0220490992,
0.0484613739,
-0.0470669754,
-0.0998509303,
0.0168729927,
0.051088836,
-0.0605033785,
0.0188873336,
0.0327835344,
0.0063404394,
0.0025317017,
-0.0468133241,
-0.0037688981,
0.0572106913,
-0.0043961117,
-0.0238224808,
0.0383895822,
-0.0344730094,
-0.0247972179,
-0.0017977908,
-0.0055338521,
0.0122634405,
0.046849478,
-0.0396407694,
0.0365541801,
-0.0293407645,
-0.0285328645,
-0.0519005395,
0.0108071687,
0.0238135755,
0.0561464541,
0.0344785154,
-0.0390944704,
-0.0265712719,
0.0304135587,
-0.0550471097,
0.0100723282,
0.009999888,
0.0158832725,
0.088324964,
-0.0313662589,
-0.0234119911,
-0.0002768548,
-0.0682857782,
0.0326550379,
-0.0050277743,
0.020301342,
0.0055048778,
0.0341578573,
-0.0279421508,
0.018686818,
-0.0147811128,
-0.0232118368,
0.0087030558,
-0.0306233373,
0.0261828657,
0.0162844174,
-0.0519623421,
0.0033164266,
-0.0096631972,
0.0403944179,
-0.0684782341,
0.0524711572,
0.0170005858,
-0.0425076373,
0.0140367122,
0.0119333826,
0.0064181914,
0.0071490114,
0.0191466883,
0.0106630195,
-0.014303416,
0.0089826938,
0.0302026793,
0.0502837673,
0.0030750828,
0.046676062,
-0.0027073086,
0.0053154207,
-0.0077101518,
-0.0116542093,
-0.0354882181,
0.0214936715,
0.0110214977,
0.0186856762,
-0.015853947,
-0.0360404328,
0.0624421127,
0.0033416227,
0.0473190658,
-0.0327412263,
-0.0241944697,
-0.0068829632,
0.0547686256,
0.0437930152,
-0.0524759889,
-0.0017931294,
-0.0200399458,
-0.0265235621,
0.049843505,
-0.0021393863,
0.0053345715,
0.0472731739,
-0.0148294447,
-0.0203345791,
-0.0090880152,
-0.0803733766,
-0.038687814,
-0.0230979249,
0.002970783,
0.0435028039,
0.0314881578,
0.0094257426,
-0.0304783098,
-0.0249555595,
-0.0238669589,
-0.0007622999,
0.0620049685,
-0.0215739217,
0.0339503624,
0.0149883935,
-0.0176285915,
0.0439244434,
0.0148053039,
0.0057966872,
0.033983916,
0.0179145131,
-0.059485577,
0.0037067775,
-0.0365002342,
-0.0763805807,
0.0697082281,
-0.0296753775,
-0.0410853289,
-0.0997362584,
-0.0226512849,
0.0574963577,
0.0582905114,
0.0319930725,
0.0446217544,
-0.0366835967,
0.0832476839,
0.0377523005,
0.0095939077,
-0.0618433878,
-0.0087386109,
-0.0003167627,
0.0015208977,
-0.0086236773,
-5.476601365e-33,
-0.0237647127,
-0.0315042473,
-0.024615515,
0.0181850325,
-0.0435882472,
-0.0018890991,
-0.0225087162,
-0.0103606796,
-0.0233317539,
0.0510661341,
0.0392669328,
0.0721575022,
0.019314779,
-0.035633754,
-0.0151004074,
0.0060640518,
0.0371600948,
0.035421811,
0.0054379958,
-0.017274633,
0.0508603863,
-0.0645920709,
0.0291993171,
-0.1115389466,
0.0837594047,
0.0231173243,
-0.0291478317,
-0.0197796859,
-0.0216047298,
0.0484562889,
-0.0261301436,
0.0552867427,
-0.0047682323,
-0.0153636392,
0.006486597,
-0.0036612602,
-0.0483231768,
0.0047113458,
0.043205753,
-0.0396106504,
-0.0107375812,
-0.0053845569,
0.0099060144,
0.0432751253,
-0.0384163111,
0.0116538554,
0.0202925671,
-0.0390691757,
-0.0133068739,
-0.0380814187,
-0.0500181019,
0.0671418607,
-0.0103137093,
0.0556944832,
-0.0284136198,
-0.0041634487,
0.0172599833,
0.0192951672,
-0.0420897678,
0.0113069313,
-0.0199888758,
-0.0207448602,
-0.000268932,
-0.0689289048,
-0.0226213578,
0.0676436201,
-0.0275765993,
0.0118642356,
-0.0032349522,
0.006859621,
-0.0240040347,
0.0447942093,
0.0433885753,
0.0710962564,
0.0466522649,
0.0057458589,
-0.0943120643,
0.0019097212,
-0.0138258524,
0.0447434522,
0.060117919,
-0.0110676792,
-0.0438096896,
-0.0056620855,
-0.0514173359,
-0.0277921725,
-0.0357175395,
0.0316918306,
-0.0045323549,
0.0088000102,
0.0407796837,
0.059433233,
-0.002687165,
-0.0638006255,
-0.0240082312,
0.0309955161,
-0.0406544879,
0.0112458831,
0.012508085,
0.0285411868,
-0.0217035022,
-0.0029829661,
-0.0288574211,
0.0508703887,
0.0036408079,
-0.0201221481,
0.0523168668,
-0.012574655,
-0.0629780814,
-0.0135074882,
-0.0044983998,
-0.0058853496,
0.0158888809,
-0.0023852391,
0.0490000471,
0.0090327095,
-0.0021887829,
0.015826121,
-0.0273723099,
-0.0444996655,
-0.0216177143,
-0.006550319,
-0.0061717001,
0.0204312261,
0.0010091645,
-0.0109740747,
-0.0153041957,
0.012768846,
0.0530011319,
-0.0156480316,
0.0163169559,
0.0377264619,
2.504e-7,
0.0518708825,
0.0265083089,
-0.003904911,
0.0306717288,
0.0081245527,
0.0000880501,
-0.0159301385,
0.0173074808,
-0.0635220408,
-0.0562003516,
-0.0339463912,
-0.0036272034,
0.051225882,
0.0247117765,
-0.0085823704,
-0.0067703421,
-0.0615055449,
-0.0688169226,
-0.0437459461,
0.0297027808,
0.0330025479,
0.025985267,
0.0520798378,
-0.007090522,
0.0221455637,
-0.009544529,
0.0656874627,
0.0408810489,
-0.0095214043,
0.0105419708,
0.026951436,
-0.0371167213,
0.0234567635,
0.039378386,
0.0147443749,
-0.0073967073,
-0.027175257,
0.0525262691,
-0.0025898058,
0.0778981671,
-0.056946829,
-0.0740438476,
0.0108629735,
-0.034911219,
0.0427866913,
0.043934267,
-0.0286936592,
-0.0589471944,
-0.061292883,
-0.0103455465,
0.0603343174,
0.0335421488,
-0.0392607376,
0.011039299,
-0.0068467222,
0.0359930545,
-0.0157514103,
-0.0548564121,
-0.0007510833,
0.0323978439,
-0.0017901886,
-0.0487056151,
0.0329058319,
-0.0209233332,
0.0467809401,
-0.096894525,
0.0311143007,
2.151457719e-34,
0.0606701411,
-0.0200721361,
0.010048788,
-0.0315767862,
0.0273694899,
0.0187404808,
-0.0113736037,
-0.020460356,
-0.0059919963,
-0.0472458415,
0.0366340838
] |
http://arxiv.org/abs/2401.16678v1 | http://arxiv.org/pdf/2401.16678v1 | The Detection and Understanding of Fictional Discourse | 2024-01-30T01:57:17 | 2024-01-30T01:57:17 | Andrew Piper, Haiqi Zhou | cs.CL | cs.CL, cs.LG | In this paper, we present a variety of classification experiments related to
the task of fictional discourse detection. We utilize a diverse array of
datasets, including contemporary professionally published fiction, historical
fiction from the Hathi Trust, fanfiction, stories from Reddit, folk tales,
GPT-generated stories, and anglophone world literature. Additionally, we
introduce a new feature set of word "supersenses" that facilitate the goal of
semantic generalization. The detection of fictional discourse can help enrich
our knowledge of large cultural heritage archives and assist with the process
of understanding the distinctive qualities of fictional storytelling more
broadly. | The Detection and Understanding of Fictional Discourse: In this paper, we present a variety of classification experiments related to
the task of fictional discourse detection. We utilize a diverse array of
datasets, including contemporary professionally published fiction, historical
fiction from the Hathi Trust, fanfiction, stories from Reddit, folk tales,
GPT-generated stories, and anglophone world literature. Additionally, we
introduce a new feature set of word "supersenses" that facilitate the goal of
semantic generalization. The detection of fictional discourse can help enrich
our knowledge of large cultural heritage archives and assist with the process
of understanding the distinctive qualities of fictional storytelling more
broadly. | [
0.0643092766,
0.0545233414,
-0.0149988979,
0.0713191628,
-0.0591202527,
0.0111446138,
0.0062910165,
0.0306499526,
0.022547476,
-0.0294641349,
0.0384490415,
-0.0112198908,
0.0683485568,
0.0336830579,
-0.0154818334,
-0.0004882877,
0.0315128565,
0.0065808119,
0.0523467399,
0.0025022142,
-0.0014998551,
0.0738222301,
-0.0381059311,
0.0347942002,
-0.0300865248,
-0.0423240066,
0.0337690301,
-0.0031565838,
0.0274453126,
-0.0000128462,
0.0419069678,
0.0191364679,
0.0473858677,
0.0395250246,
0.0000014019,
-0.037052352,
0.0515348911,
-0.043686673,
0.0166667104,
0.0295807458,
0.0411309823,
0.0440886281,
0.0204269551,
-0.0158715062,
-0.0116019705,
0.0056931628,
0.0191949233,
0.0791699886,
0.018575402,
0.0693199113,
-0.015400541,
-0.055547528,
-0.0269043334,
0.0306798946,
-0.0208073109,
0.0045010136,
0.0265636221,
0.0275655519,
0.0032843,
0.0030885211,
-0.0087070279,
0.0160238203,
-0.0380924791,
0.0324009433,
-0.0212323554,
0.0315412395,
-0.0094535975,
0.0280077606,
-0.0047843265,
0.0204050001,
0.0323045813,
-0.053618636,
0.0136029804,
0.0272163805,
-0.0365220271,
0.0414899513,
0.0059075421,
0.011768085,
-0.0457701869,
0.0189752541,
0.0588234365,
0.0122334296,
0.0316878222,
0.0279143695,
0.0349691585,
0.0153211271,
-0.0019556468,
0.0351053849,
-0.0164860897,
-0.0060322098,
-0.0445156507,
-0.0513678677,
0.0238026958,
0.0133988978,
-0.0109792966,
0.0074175019,
-0.0340723507,
0.0040048687,
0.0279710665,
-0.0768702775,
0.0010576352,
0.0038850661,
-0.0222364422,
0.050022725,
0.0183122028,
-0.0070906105,
-0.0022181054,
-0.0392516963,
-0.0039735809,
0.000821585,
-0.0377893411,
-0.0118614435,
-0.0225455314,
0.0240425747,
-0.0677741691,
-0.0416128151,
0.0002368112,
-0.0340385549,
0.0502362326,
-0.0044975006,
-0.0357306525,
0.0155201945,
-0.0134498803,
0.0181372259,
-0.0518238991,
0.0126934126,
-0.0798588693,
0.0078588761,
-0.0576482974,
-0.0025767982,
0.0005816601,
0.0013921772,
0.0184396524,
-0.0216911845,
0.024266459,
0.048310928,
-0.0160355885,
-0.0560889728,
-0.0478787087,
-0.0124604087,
0.0111529296,
-0.0380467139,
-0.0060184584,
-0.0110709872,
0.0075712749,
-0.0246673208,
0.0338274613,
-0.0276639964,
0.0128234411,
0.0068144733,
-0.0230953116,
0.0258894879,
-0.0337436981,
-0.0144815706,
0.0778376162,
0.0259078611,
0.0061679585,
0.0081907501,
0.0689299852,
0.0234235544,
0.0773230717,
0.0280363597,
0.0054990086,
0.0086005209,
-0.0307286084,
-0.0070264554,
0.0116742365,
0.0111701665,
-0.0162270945,
0.0425963663,
-0.0147365341,
0.0218093041,
-0.0236098357,
-0.0549293794,
0.0464155897,
0.0982458293,
-0.0308712386,
0.0361378677,
0.0588967279,
0.0298824925,
0.0372769535,
0.0444214866,
0.0222272202,
-0.0280946195,
-0.0113719068,
-0.0404264517,
-0.0630567148,
-0.0087004984,
0.018863041,
0.0052346722,
-0.000854196,
0.0114312312,
0.0003825457,
-0.0373086445,
0.0018465366,
-0.0085484376,
-0.0759108961,
-0.0373308882,
0.0518554561,
-0.0535366237,
0.0314483829,
-0.0290009435,
0.0610864833,
0.0642703846,
0.0707947463,
-0.0513608567,
0.0464525595,
-0.0290925056,
0.0366946273,
0.049627021,
-0.0227199513,
-0.0312280599,
0.0467379428,
-0.0324793793,
0.0194316134,
-0.0477888733,
0.0243494976,
0.0350098945,
-0.0230119396,
-0.0007666539,
-0.0147043271,
-0.0156150078,
-0.0294335596,
0.0362108499,
-0.0685718507,
-0.0043937713,
-0.0461454466,
0.0016133566,
-0.0424807109,
0.0017828245,
-0.0582954921,
-0.0676603392,
0.0241445657,
-0.0400678739,
-0.0108878762,
0.0312933996,
0.0005760342,
-0.0164729673,
0.0143804671,
-0.1202741638,
0.0034880817,
0.0178583991,
0.0128898434,
-0.00010275,
0.011583305,
-0.0379941799,
0.0179728679,
-0.0277008228,
-0.0181646068,
-0.0520954579,
-0.0129395593,
-0.0374932177,
0.0657282621,
0.021808153,
0.0362943634,
-0.0310674794,
-0.0581304207,
0.0456870757,
0.014658412,
-0.01308366,
0.0158524178,
0.0447829105,
0.023983255,
0.0560799241,
-0.0609588623,
-0.0057424507,
-0.0180091057,
-0.0800417587,
0.0054582697,
-0.0118691465,
0.0046608564,
0.0246080831,
-0.0298756007,
-0.0444928184,
0.0342419185,
-0.0204878654,
-0.0394499935,
-0.0231683161,
-0.0166431107,
0.0413417108,
-0.0149504123,
0.031969253,
-0.0161806829,
0.010622873,
-0.0111536216,
0.0862835422,
0.0275697429,
-0.0045012506,
-0.0248477012,
-0.0401429087,
-0.0258804057,
-0.0165968556,
-0.0318909213,
-0.0527763851,
-0.0056209415,
-0.0049661105,
-0.0192601364,
0.0202728566,
0.0445914716,
0.0130863218,
0.0231966879,
-0.0552354977,
0.0269358102,
-0.0235271864,
0.0175807253,
0.0278911255,
-0.045784492,
0.0077403528,
0.0161320679,
0.0412190892,
0.0194604211,
-0.0105131371,
-0.027636582,
0.0316247009,
0.0176375173,
-0.0104238559,
0.0233190786,
-0.0346144959,
-0.0041440791,
0.0692661479,
-0.0057634995,
0.0071422569,
0.0397159345,
-0.0245738775,
0.0110052871,
-0.0270618871,
-0.0401627831,
-0.0166326575,
0.0576151349,
-0.0046133841,
0.0447950587,
0.0384196267,
-0.0318791158,
0.0181368776,
-0.0004221546,
0.0120301545,
-0.0260892157,
0.0360742062,
0.0044920058,
-0.013307848,
0.0377048478,
0.0057398994,
-0.0502084345,
-0.0553144217,
-0.038115304,
-0.0634799376,
-0.0541893989,
-0.0407385789,
-0.026019657,
-0.0393029004,
-0.0313315019,
0.0269535594,
0.0048036096,
-0.0159682315,
0.0025405567,
0.0323812515,
0.0234355014,
0.0058068112,
-0.0264534615,
0.0063982923,
-0.0201401263,
0.0526724569,
-0.07163506,
0.0039338009,
0.001015104,
-0.0000054685,
0.0180777274,
-0.0540087186,
-0.0129891783,
-0.0384530164,
0.0161596518,
-0.0721717179,
0.0034705778,
0.0866156593,
0.0501924753,
0.0232849605,
0.0136848427,
0.0244521163,
-0.009750234,
0.0158091616,
-0.0398014039,
-0.0563531257,
0.0019833373,
-0.0053164414,
-0.0099000391,
-0.0285064261,
-0.072886169,
-0.0148771135,
-0.0342504941,
0.0152520826,
0.0071320147,
0.02133568,
-0.0075130244,
0.0076088617,
0.0100480178,
0.0082481345,
-0.0594593063,
-0.0245798361,
0.0079946127,
0.0123602562,
-0.0696296915,
0.0183533505,
0.0187422931,
0.0217442606,
0.0046369219,
-0.0049039936,
0.0130416797,
0.0872473419,
-0.0103304042,
0.0040906649,
-0.0456431322,
-0.0467394814,
-0.0153494757,
-0.0083029093,
0.0903117284,
-0.0144408494,
0.0063870233,
-0.0195788443,
0.0401397012,
-0.0068162442,
-0.0472567119,
0.0600783974,
0.0317698792,
0.025844425,
-0.0389702767,
-0.0377386138,
0.0389320105,
0.0503739305,
0.0619646758,
0.0394026376,
-0.0035283577,
-0.0124139832,
-0.0586674698,
0.0125848837,
-0.014393298,
-0.0521750003,
-0.0426083654,
0.0237676222,
0.0115843453,
-0.0859260336,
-0.0382919125,
-0.0287964568,
0.0364164673,
0.0404026844,
0.0180909131,
-0.0133355716,
-0.0117601231,
-0.0551868938,
0.0015843792,
-0.0203494877,
0.0046814727,
0.0094757956,
0.0191135518,
0.0522200949,
-0.034671627,
-0.000242262,
0.0405527242,
-0.0447940789,
-0.0627221391,
0.0268080626,
-0.0482380129,
0.0201361608,
0.0130983163,
-0.004329781,
-0.041720584,
0.0785817653,
0.0304373559,
-0.0827674866,
0.0236869361,
-0.002839701,
0.0149816433,
-0.0200030878,
0.0114364903,
0.0155063802,
0.0321595483,
-0.0456918739,
-0.0284859538,
-0.0360043272,
0.0223948173,
-0.0103781065,
-0.0051537016,
0.0347474478,
-0.0486208871,
0.0171515569,
-0.0590770021,
-0.0194737185,
-0.0164876692,
-0.0216250364,
0.0013656281,
0.0297480971,
0.0099109169,
-0.0208788645,
-0.0544019081,
0.0102593051,
-0.0344250053,
0.06088572,
-0.0280260108,
0.0126968762,
0.0587569587,
0.0396922156,
-0.0154289138,
0.0100030117,
-0.0182985365,
0.0204672068,
-0.0155676948,
0.0589005686,
0.0062203621,
0.0011392927,
0.0214153677,
-0.0189927686,
-0.0021991353,
0.0036337434,
-0.0608524457,
0.0053055431,
0.031753201,
0.0376230106,
0.0747215375,
-0.0560649484,
0.0560290329,
-0.003742154,
-0.0068995613,
0.0089514991,
0.0528897084,
-0.0032777917,
0.0250729788,
-0.0706863031,
0.0389877222,
-0.0373304747,
0.0488037951,
0.0256887171,
-0.0509931706,
0.0401657149,
0.1129780784,
-0.0485969633,
-0.0375206359,
-0.057624843,
-0.0668475628,
0.0563926883,
0.0624433048,
0.0498991981,
-0.0539747998,
-0.0271665119,
0.0589506663,
-0.0133681055,
0.0393577851,
-0.0387016274,
-0.0044242507,
0.0205441248,
0.0161753688,
0.0288844462,
-0.0462272614,
0.0072655864,
0.0119689619,
0.0120777925,
-0.0132010616,
-4.889352305e-33,
-0.001967517,
-0.0278512258,
-0.0285236388,
0.0952617154,
-0.0503623001,
-0.0076442421,
-0.0129564088,
-0.0290613361,
0.002661369,
0.0506876893,
0.0277808961,
0.0228977073,
0.0020425487,
-0.0138553642,
-0.0492920578,
-0.0037039057,
0.0393122919,
0.0288419928,
-0.0757386833,
0.0062624747,
0.0815199837,
0.0287790317,
0.0673180893,
-0.0922857597,
0.0244464632,
-0.0369667783,
-0.0276993252,
0.0226276424,
0.0390553735,
0.0414052121,
-0.0312531404,
0.0896465406,
0.0098999878,
0.038545832,
0.0030089952,
0.0059732106,
-0.0583693646,
-0.0282781366,
0.053330984,
-0.0139330653,
0.0597692356,
-0.0229294319,
-0.0167773552,
0.0193641447,
-0.0183411613,
-0.0576738715,
0.0764744356,
-0.0127950329,
0.0189996045,
0.0019933449,
-0.0891649425,
0.005986406,
-0.0388761722,
0.0186439008,
0.0493234508,
0.0198509041,
0.0637526885,
-0.0032140191,
-0.0459002778,
0.033716958,
-0.0081128189,
0.0111640096,
-0.0563812964,
-0.0077529973,
-0.0226897243,
0.0741782859,
0.0267566554,
0.0086339964,
-0.0359091796,
0.0951870158,
-0.0107718837,
-0.0232555103,
0.0069187642,
0.1219718605,
0.0226352941,
-0.0298840348,
-0.0396613702,
-0.0103847301,
0.0230006948,
-0.0005712008,
0.017692592,
-0.0465103872,
-0.0240963809,
-0.0357609577,
-0.0279379115,
-0.027834814,
0.0051702452,
-0.0050090025,
-0.0386631303,
-0.055211179,
-0.0127953514,
0.0464325137,
-0.0228174943,
-0.009945333,
-0.0025498953,
0.0551574342,
-0.0465405062,
0.0286455732,
0.0155628584,
0.0229253042,
0.0014554997,
0.0129003972,
0.0139976097,
0.0174358282,
0.0610357709,
-0.0246340781,
-0.0005876943,
0.0210100021,
-0.0428079255,
0.0016623924,
0.01501894,
-0.0137963425,
-0.0341591537,
0.0065771393,
-0.0254577994,
-0.0199118406,
0.0035498997,
0.0061478997,
-0.0227442887,
0.0072680148,
-0.0254336074,
-0.0520435832,
-0.0188657679,
0.028481612,
-0.0287602246,
-0.0131745823,
0.0107878642,
-0.0343150273,
0.03826195,
-0.0734607577,
-0.00435807,
-0.0631779134,
2.229e-7,
-0.0125956656,
0.0197955593,
0.0357562006,
0.0269929692,
-0.0129108345,
0.0336335413,
-0.0136868246,
0.041362308,
0.0020685948,
-0.0645606667,
-0.0165067334,
0.0090693245,
0.0317914002,
0.0617201105,
-0.083087936,
-0.0231556445,
-0.0420148708,
-0.0137679931,
-0.0725369081,
0.0203457791,
0.0646036938,
-0.0005455134,
0.0744611248,
-0.047267016,
0.0046534524,
-0.0146026975,
0.0139263328,
-0.0382120833,
-0.0489598587,
0.0209304336,
0.0009721591,
0.0045853043,
-0.0009851573,
-0.0156025775,
-0.0098031936,
0.0325755961,
0.0160648599,
0.0293526296,
-0.0031486924,
0.0008663181,
-0.0040393327,
0.0014208254,
-0.0281147957,
0.0087735588,
0.0226199981,
0.0582556017,
-0.0236516036,
-0.0116189625,
0.0085250773,
0.0014520258,
0.0371357091,
-0.0118312985,
-0.0387537479,
0.0041229115,
0.0186083224,
0.0065793581,
0.0076778997,
-0.0853143707,
0.0039753923,
0.0399418846,
-0.0667938739,
-0.0124001969,
0.0381763205,
-0.0295906179,
-0.0294141825,
-0.0660595372,
0.0393316895,
2.110359727e-34,
0.0344369374,
-0.0193125587,
-0.049695801,
0.0359505378,
0.0674852133,
-0.0125285853,
0.0385087766,
-0.0250604115,
0.0323658958,
-0.0143170878,
-0.0022799571
] |
http://arxiv.org/abs/2401.16659v3 | http://arxiv.org/pdf/2401.16659v3 | History-Aware Conversational Dense Retrieval | 2024-01-30T01:24:18 | 2024-05-28T13:18:29 | Fengran Mo, Chen Qu, Kelong Mao, Tianyu Zhu, Zhan Su, Kaiyu Huang, Jian-Yun Nie | cs.IR | cs.IR, cs.CL | Conversational search facilitates complex information retrieval by enabling
multi-turn interactions between users and the system. Supporting such
interactions requires a comprehensive understanding of the conversational
inputs to formulate a good search query based on historical information. In
particular, the search query should include the relevant information from the
previous conversation turns. However, current approaches for conversational
dense retrieval primarily rely on fine-tuning a pre-trained ad-hoc retriever
using the whole conversational search session, which can be lengthy and noisy.
Moreover, existing approaches are limited by the amount of manual supervision
signals in the existing datasets. To address the aforementioned issues, we
propose a History-Aware Conversational Dense Retrieval (HAConvDR) system, which
incorporates two ideas: context-denoised query reformulation and automatic
mining of supervision signals based on the actual impact of historical turns.
Experiments on two public conversational search datasets demonstrate the
improved history modeling capability of HAConvDR, in particular for long
conversations with topic shifts. | History-Aware Conversational Dense Retrieval: Conversational search facilitates complex information retrieval by enabling
multi-turn interactions between users and the system. Supporting such
interactions requires a comprehensive understanding of the conversational
inputs to formulate a good search query based on historical information. In
particular, the search query should include the relevant information from the
previous conversation turns. However, current approaches for conversational
dense retrieval primarily rely on fine-tuning a pre-trained ad-hoc retriever
using the whole conversational search session, which can be lengthy and noisy.
Moreover, existing approaches are limited by the amount of manual supervision
signals in the existing datasets. To address the aforementioned issues, we
propose a History-Aware Conversational Dense Retrieval (HAConvDR) system, which
incorporates two ideas: context-denoised query reformulation and automatic
mining of supervision signals based on the actual impact of historical turns.
Experiments on two public conversational search datasets demonstrate the
improved history modeling capability of HAConvDR, in particular for long
conversations with topic shifts. | [
0.0363875814,
0.0259793419,
-0.0166244246,
0.0137087898,
-0.0217747018,
-0.0171992406,
0.00089866,
0.0158023369,
-0.0031261852,
-0.0216456149,
-0.0200952031,
-0.0254071094,
-0.0417157896,
0.0339529663,
-0.0420176797,
-0.0432606488,
0.0549256653,
0.0128437076,
0.0180039443,
0.050983306,
0.0120724645,
-0.0199432932,
-0.0051378985,
0.0336937495,
-0.0339519456,
0.0271747764,
-0.0745780766,
0.0110647678,
-0.0034795729,
-0.0033342885,
0.0390834771,
0.0441593975,
0.058357086,
0.0153668663,
0.0000016267,
-0.0037756078,
-0.0328308754,
-0.0332451575,
-0.0283067431,
-0.0277340505,
0.0333696343,
0.0302227698,
0.0210531559,
0.0010716395,
-0.0277461782,
0.0122504458,
-0.0117929047,
0.0748190731,
0.0399964713,
0.0615831912,
-0.011062704,
-0.0015252254,
0.0194275286,
-0.0117166452,
0.0359774493,
-0.0650915354,
0.0205089394,
0.000159506,
0.0107500944,
0.0606342703,
0.0248770211,
0.0445762165,
-0.0044139093,
-0.0583064407,
0.0320748761,
0.0008511768,
-0.006245031,
0.0191317722,
0.0022541194,
0.0596538335,
-0.0027826424,
-0.0107939346,
0.013775236,
0.0209101029,
-0.0263799727,
-0.0200477615,
-0.0707090348,
-0.0501208119,
0.0010800523,
-0.0107850228,
-0.0344629437,
0.0148614058,
0.0095683634,
0.0064420258,
-0.0110041136,
0.0385625064,
-0.0136114899,
-0.0523949452,
-0.0159497056,
-0.0183581095,
0.0348095,
-0.0177345611,
0.0514744967,
0.040321894,
-0.0597846694,
0.0090031018,
0.0369647928,
0.0181090087,
0.0023286853,
-0.0583024286,
-0.0609368607,
0.0023585341,
-0.0257544667,
0.0699778274,
0.0448945425,
0.0079742149,
-0.0086886277,
-0.0448430292,
-0.0286394227,
-0.0401972868,
-0.051177226,
0.0005763835,
-0.0426400974,
0.0291184243,
-0.0548729263,
-0.0134337898,
-0.0054543526,
0.0443662181,
0.0358364806,
-0.0387572348,
-0.0127009563,
0.0461413674,
-0.0379546992,
0.0111559452,
-0.0234556533,
0.0253118873,
-0.0026847809,
0.0192505866,
0.0197777152,
-0.0495624356,
-0.0193274878,
0.0136946198,
0.0304367468,
-0.041643057,
-0.0463044159,
0.00433048,
-0.0272592176,
-0.0612864308,
-0.0215728991,
-0.0258956552,
-0.0143857747,
0.0079233646,
-0.0064622038,
-0.0038881367,
0.0294450615,
-0.0281667188,
-0.029273184,
-0.0001582594,
-0.0039863754,
0.047046572,
-0.0107674925,
0.0206723977,
0.009544366,
0.0045042154,
0.0576242879,
0.0359543227,
0.0060225921,
0.0516405143,
-0.0006158583,
0.0512359105,
0.0542995967,
0.0159117803,
0.0373467207,
-0.0090771345,
0.0404886641,
0.0015001773,
-0.0103964247,
0.053242255,
-0.0223690998,
0.0197699778,
0.0087341713,
0.0388678499,
-0.0583775043,
-0.0251563527,
0.0072943065,
0.0499409176,
0.0367975496,
-0.0149345798,
-0.0033820826,
-0.0091946246,
-0.0264712591,
0.0131572736,
-0.0358713418,
-0.0193776302,
-0.0606839359,
-0.0507221669,
0.0141170826,
-0.0265493039,
-0.0334818587,
-0.0454759523,
0.058269497,
0.0353979692,
-0.0122819673,
0.0294016227,
0.0305925403,
-0.006816241,
-0.0213066209,
0.0102043841,
0.0194192585,
-0.0501471758,
-0.0007257627,
-0.0175580177,
0.0293032657,
0.0748596713,
0.0884845927,
-0.0317748673,
0.0045033391,
-0.0252373274,
0.0307765268,
-0.012379678,
0.0373293497,
-0.0331924297,
0.0529223159,
-0.0228047241,
-0.0632542893,
0.057302665,
-0.0007677356,
-0.0014800611,
0.0007676524,
0.0168606453,
0.0376982875,
0.0056369496,
-0.0038582727,
-0.0000672522,
0.0232508406,
-0.028469231,
0.0125491945,
0.00923876,
-0.0025322987,
0.0322773419,
0.0018279987,
0.0010732306,
0.0741758049,
0.0248712115,
-0.0360383689,
-0.0111763617,
-0.0928122029,
-0.035838332,
-0.0052217129,
0.005704931,
0.0104094511,
-0.0452422574,
0.0100121154,
-0.0598884039,
0.0117317513,
-0.0119131058,
0.0681996197,
-0.024984546,
0.0139337806,
0.0322712325,
-0.0024745602,
-0.058767207,
0.0500393845,
0.002284714,
-0.0102272881,
0.0199848972,
-0.004099024,
-0.0058654253,
-0.0747618452,
-0.0247405488,
0.0021187793,
0.0047468198,
-0.0015867044,
0.0660789832,
-0.0119011002,
-0.059969157,
0.0027604655,
-0.11709591,
0.0096740704,
-0.0438687839,
-0.0391642377,
-0.0114156837,
0.0335273519,
0.007546402,
-0.0509661213,
-0.0275010634,
-0.0028422873,
0.0175762735,
0.0357543305,
0.0167256314,
0.0753645897,
-0.0277736019,
-0.0725627318,
-0.0145578124,
0.0150147229,
0.0249011125,
-0.0102899671,
-0.0161369368,
0.0070204521,
0.0042758314,
-0.0082013477,
0.0496954657,
-0.05534219,
0.0091223791,
0.0255508926,
-0.0750286877,
0.0310307723,
0.0406369418,
0.0488872379,
0.0736691505,
-0.0266890861,
-0.0679166317,
-0.0813309997,
0.0394683257,
-0.0601327382,
0.0071536205,
-0.0355934352,
0.0261032972,
0.023950899,
0.0077610617,
-0.0242716521,
0.0454069823,
-0.0338046923,
-0.0776463374,
0.02342318,
-0.0263256729,
-0.0740908161,
0.0225826949,
-0.0005828177,
0.0306816511,
-0.0229343157,
-0.0331866518,
0.0458198488,
0.0058856905,
-0.0216354746,
0.0032789826,
-0.0674827844,
-0.0073404685,
0.0753785521,
-0.0462728553,
-0.0059469212,
-0.0077321059,
-0.0541100092,
-0.0230036397,
-0.017629683,
0.0238888767,
-0.0389269814,
0.0159309711,
-0.03061557,
0.025058277,
0.0511081927,
-0.0448228493,
-0.0527265854,
-0.0395967104,
0.0076696081,
-0.0096066585,
0.004751707,
-0.0434092283,
-0.0848790705,
0.0184933227,
0.0025170627,
0.0363873579,
-0.0055296454,
-0.0267738588,
0.0108972099,
0.0223094244,
0.0581638478,
0.0006305236,
-0.0534213744,
0.0151417376,
0.0340841226,
0.0434431173,
0.0340928212,
-0.0817768052,
-0.0167925116,
-0.0263584871,
0.0158712734,
-0.0167185664,
-0.0166197568,
0.005286511,
0.0161505695,
-0.0041977861,
-0.0219393559,
0.0755929425,
0.0384841189,
-0.1042782143,
-0.0277814623,
0.0147501612,
0.0193685964,
0.0031408758,
0.0156148411,
-0.037184678,
-0.0142332772,
0.0073025199,
0.0354081132,
0.0124208229,
-0.0491434634,
0.0009249336,
-0.0563577376,
0.0187929925,
-0.0096248761,
0.0270729754,
-0.0275709424,
-0.0550909899,
0.0221047886,
-0.0253800955,
0.0130456295,
0.0233894773,
-0.006060333,
-0.0289461352,
0.0273979697,
0.0206366125,
-0.0480790362,
0.016555544,
0.0261847563,
0.0224774629,
0.021990465,
0.0424241759,
0.0774250627,
-0.0007409165,
0.0328741334,
-0.0232335124,
0.013423943,
-0.0294924863,
0.0688233003,
-0.0402262881,
0.0279653445,
-0.0341006964,
-0.0132199144,
0.0181302819,
-0.0146588078,
-0.0006757989,
0.0324248932,
-0.0835472494,
0.039528951,
-0.0338374116,
-0.0058810329,
0.0478741787,
0.0529131442,
0.018890813,
0.0288042426,
-0.0263785385,
0.0053854198,
0.0163745824,
0.0190894883,
-0.0282401219,
-0.0519885272,
0.0262304675,
0.016990561,
-0.0428995341,
-0.0612292215,
-0.020964887,
0.0372924656,
-0.0325485431,
0.0090981601,
0.0098387506,
0.0019651598,
-0.0206237156,
0.0224415157,
-0.0266991314,
0.0027381717,
-0.0683757141,
-0.0030767978,
0.0089428211,
-0.0156343263,
0.0068635722,
0.0012761672,
-0.0324453898,
-0.0275715925,
0.0018521975,
0.0014977971,
0.0149965174,
0.0590247214,
-0.0239374787,
-0.039724052,
0.0121059725,
0.0152815748,
-0.0558732376,
-0.0665793493,
0.0299730133,
0.0460049547,
-0.0172923468,
0.0322962813,
-0.0501975715,
0.0351357609,
-0.0420172922,
0.0051161451,
-0.0907069594,
-0.076793097,
0.0037747347,
-0.0235795639,
0.1020928547,
0.0351364724,
0.0130634364,
-0.002661681,
0.0375856198,
-0.0572895333,
-0.0186247882,
-0.0290322173,
0.0341223516,
-0.0191458687,
-0.041947186,
0.0085478788,
0.0240736101,
-0.0367432535,
0.0242041275,
-0.0662722811,
0.0042250259,
0.0169416629,
-0.0080073383,
0.0156032303,
0.0218621381,
-0.0020484179,
0.0339787602,
0.0420775004,
0.0727429911,
-0.0226709936,
0.0323516801,
0.0180502199,
-0.0145357559,
-0.0026789133,
0.026759861,
-0.1476836652,
-0.0148015097,
-0.0001067424,
-0.0025741663,
0.0469362438,
-0.0223050304,
0.036389593,
0.0036018761,
0.0242309496,
-0.04532912,
0.0345046893,
0.0158629548,
-0.0151480427,
0.0248787273,
0.0366070867,
-0.014349632,
0.0678614452,
0.0497712903,
-0.0151326647,
0.0477080494,
-0.0181350224,
-0.0566350259,
-0.0237191822,
-0.0160857141,
0.0561436936,
0.0453231856,
-0.0370922051,
-0.0067814374,
-0.0053840014,
0.013638773,
-0.0049725808,
-0.0569414347,
-0.0051410892,
0.0030423896,
0.0758949146,
0.0827210173,
-0.0045251609,
-0.0263583548,
-0.0309640579,
0.0113366125,
0.0193527527,
-0.0088369492,
-0.0585615784,
-6.196529574e-33,
0.0016691294,
-0.0128590558,
0.0275725666,
-0.0173998177,
-0.0228533987,
-0.0009388247,
-0.0452623591,
-0.0157821588,
-0.0295668673,
0.0165685471,
-0.0343511365,
0.0648482814,
0.0161806419,
0.0117531735,
-0.024903195,
-0.0289508607,
0.079763934,
-0.0101228561,
0.0027630019,
0.0559444167,
0.0546580255,
0.0727585927,
0.0086016683,
-0.0924924612,
0.0256153066,
-0.0299614146,
0.0038008846,
0.0620643049,
0.0261268076,
0.0326311328,
-0.0093757194,
0.0413118489,
0.0070924796,
0.0120916534,
0.0040151752,
-0.0154574644,
-0.0746623948,
-0.0097765354,
-0.0207352228,
0.0004365078,
0.0148772178,
0.0272781588,
0.0142992111,
0.0054864888,
0.0150026632,
0.0209079031,
0.0122719975,
0.0127068544,
-0.0753171518,
-0.0117717469,
0.0371895283,
0.0499520898,
-0.0470267236,
-0.0475000627,
0.0116446484,
0.0530184098,
0.0429844968,
-0.007010085,
-0.0250193253,
0.0658424199,
-0.0384410322,
0.032445617,
-0.0128244851,
0.0013146502,
-0.0215197187,
0.0433875881,
0.0230018683,
0.0153676914,
-0.0085581513,
-0.0295664091,
-0.0047240271,
0.047918085,
-0.0601241887,
0.0019497186,
-0.0156787876,
-0.0197720975,
0.0011894857,
0.0204852466,
0.0470573567,
0.0609227307,
-0.0365772285,
0.0262370873,
-0.0781842396,
-0.0291950684,
0.0049951668,
0.0046080728,
-0.0067949453,
-0.0017245264,
0.0343424045,
0.0747121423,
0.0331092551,
-0.0267381351,
0.041426722,
0.0042463872,
0.0043727206,
0.0357584655,
-0.0134672541,
-0.0284523591,
0.0192282759,
-0.0029057474,
-0.0307527948,
-0.0230682753,
0.0650520623,
-0.0018853456,
-0.0279343296,
-0.0076389234,
-0.0274246149,
0.0345298126,
-0.031409435,
-0.0277076773,
0.0189798493,
-0.0058915084,
-0.0411612839,
0.000789517,
0.0126118008,
0.0075122816,
-0.0132788047,
0.0901167765,
0.003308357,
0.0765311494,
-0.0391039364,
-0.0773085579,
0.045266673,
-0.0002161798,
-0.001637294,
-0.0108459778,
-0.0077465824,
0.0259581879,
0.0467198342,
-0.0248633213,
0.0146517996,
0.0148425763,
2.557e-7,
-0.0042370744,
0.0379585214,
0.0249754544,
0.0011922916,
0.0038660509,
0.0146319177,
-0.0078497678,
-0.022282755,
0.0253026318,
0.0051658498,
0.049352739,
0.00850977,
0.0293481424,
0.0077535212,
-0.0551914126,
-0.0991137028,
-0.069857195,
-0.0638396516,
-0.0529192276,
-0.0424661711,
0.0653820261,
0.0791779384,
0.0108967824,
0.0088390075,
0.0031943861,
-0.0003943543,
0.0339636691,
-0.0135223838,
-0.0246119034,
-0.026587218,
0.0606902502,
0.0171166696,
-0.0199253,
-0.0275575109,
-0.0382159911,
0.0360468514,
0.0127898259,
0.0782470703,
-0.0151187098,
-0.0680948868,
-0.0117762871,
0.0120986635,
-0.0074362489,
0.0033611532,
0.0525754616,
0.0024256997,
0.0413733087,
0.0129534807,
-0.0141277285,
0.0131677222,
0.0372204818,
0.031374611,
-0.0435298346,
0.0027078928,
-0.0128636966,
0.0634912848,
0.0151020773,
-0.1350646764,
0.0279115327,
-0.0164679065,
-0.0503160357,
-0.0058668754,
-0.0238947235,
-0.0382459313,
-0.0024176035,
0.0079706386,
0.0378571525,
1.875883043e-34,
0.0549310595,
0.0220007226,
-0.0140532795,
0.0064055142,
0.0412515402,
-0.0002044067,
0.0205400456,
0.0004700121,
0.0586214103,
-0.084462367,
-0.0291336775
] |
http://arxiv.org/abs/2401.16657v1 | http://arxiv.org/pdf/2401.16657v1 | Recovering Mental Representations from Large Language Models with Markov Chain Monte Carlo | 2024-01-30T01:22:18 | 2024-01-30T01:22:18 | Jian-Qiao Zhu, Haijiang Yan, Thomas L. Griffiths | cs.AI | cs.AI, cs.CL | Simulating sampling algorithms with people has proven a useful method for
efficiently probing and understanding their mental representations. We propose
that the same methods can be used to study the representations of Large
Language Models (LLMs). While one can always directly prompt either humans or
LLMs to disclose their mental representations introspectively, we show that
increased efficiency can be achieved by using LLMs as elements of a sampling
algorithm. We explore the extent to which we recover human-like representations
when LLMs are interrogated with Direct Sampling and Markov chain Monte Carlo
(MCMC). We found a significant increase in efficiency and performance using
adaptive sampling algorithms based on MCMC. We also highlight the potential of
our method to yield a more general method of conducting Bayesian inference
\textit{with} LLMs. | Recovering Mental Representations from Large Language Models with Markov Chain Monte Carlo: Simulating sampling algorithms with people has proven a useful method for
efficiently probing and understanding their mental representations. We propose
that the same methods can be used to study the representations of Large
Language Models (LLMs). While one can always directly prompt either humans or
LLMs to disclose their mental representations introspectively, we show that
increased efficiency can be achieved by using LLMs as elements of a sampling
algorithm. We explore the extent to which we recover human-like representations
when LLMs are interrogated with Direct Sampling and Markov chain Monte Carlo
(MCMC). We found a significant increase in efficiency and performance using
adaptive sampling algorithms based on MCMC. We also highlight the potential of
our method to yield a more general method of conducting Bayesian inference
\textit{with} LLMs. | [
0.0240622666,
0.073135592,
-0.0521553382,
0.0248731896,
-0.0600310564,
0.0056959437,
-0.0301031657,
0.010135388,
-0.0020448989,
-0.0493792519,
-0.0391848199,
-0.0723754689,
0.0316339582,
-0.0324603915,
0.0436503068,
-0.084140785,
-0.0161542278,
-0.071253404,
0.0003295397,
-0.023077907,
-0.0164164472,
0.0059374175,
-0.0249602888,
0.0325599909,
0.0374039188,
-0.0057688784,
-0.0300172437,
-0.0240296442,
0.0069682566,
-0.0268834829,
-0.0507119857,
0.0536280498,
0.0483635999,
-0.0039248415,
0.0000018451,
-0.0456306152,
0.0150556117,
-0.0388261564,
-0.0219274666,
-0.0054303333,
0.0567726828,
-0.0076601841,
0.0222032685,
0.0006732707,
-0.021899879,
0.0096205641,
0.0208475254,
0.0374324434,
0.0063490211,
0.0108189769,
0.0073933126,
0.008242134,
-0.0197709594,
-0.0026864912,
0.0256251246,
-0.045182921,
0.0533961058,
0.0400240384,
0.0497392491,
-0.0111239953,
-0.0248274673,
-0.0000065933,
-0.017125465,
0.0293338802,
-0.0298613757,
0.0373388007,
0.0368267894,
0.0152940741,
-0.0191749986,
0.0430249088,
0.0329296701,
0.0003425776,
0.0116017042,
-0.007554072,
-0.0061625955,
0.0731993243,
-0.0662391558,
0.0482177474,
-0.0205611233,
0.0470369533,
0.0130643668,
-0.0117013371,
0.026116604,
-0.0133760842,
-0.0351182558,
0.0985635892,
0.0095143262,
-0.0068009864,
0.052049499,
0.0235839281,
-0.0025272176,
-0.0655827895,
0.0057439175,
-0.0164993573,
0.0078659933,
0.038262181,
-0.0377038606,
-0.0259133186,
0.0000296777,
-0.0276240725,
0.0705011636,
0.0259126667,
0.0115434602,
0.0342621803,
0.0174121782,
0.0126482937,
-0.0327226371,
-0.0031285649,
-0.0185321737,
0.0319787934,
-0.0574746057,
-0.005788757,
-0.0195574742,
0.0164611023,
0.0108501064,
0.0005503095,
-0.0283182245,
0.0155122047,
-0.0058129039,
-0.0514880903,
-0.0470655523,
0.0062715518,
-0.0277673397,
-0.0075055854,
-0.065680325,
-0.0058864243,
-0.0510530025,
0.0341325812,
0.0177761558,
-0.0426179357,
-0.0024205635,
-0.0050753751,
0.0215596706,
-0.0229956899,
-0.0234644078,
0.0696362033,
0.010947899,
-0.0207513217,
-0.0351694115,
-0.0177133884,
-0.0266953837,
-0.0467881821,
0.0577182807,
-0.0382083282,
0.0069945403,
0.0051169698,
0.0095001953,
-0.0257040113,
0.0195972286,
0.0004818819,
-0.0683910996,
0.0876857713,
0.0203285087,
-0.0080704829,
0.0125166001,
-0.0060542487,
-0.013707283,
0.068231985,
0.0244829766,
-0.0005876591,
0.0094455136,
0.0180715453,
0.0478269942,
0.0502574109,
-0.0088291205,
0.0091181491,
-0.016487062,
-0.048704464,
0.0077631553,
0.0526549406,
0.0840774328,
0.0472055189,
-0.0417269878,
-0.023397414,
0.0636159107,
0.0569406822,
-0.0493687987,
-0.012154161,
0.0528965332,
0.0285178777,
0.0392675325,
0.0572643243,
-0.0489998311,
0.0203876793,
-0.018795412,
0.007809951,
0.0377697684,
-0.0174372122,
-0.0034564438,
0.0104012117,
-0.0148246391,
0.005059924,
0.0038057661,
0.0178447422,
0.0800882801,
-0.0138528226,
-0.0061400789,
0.0320424438,
0.0626476407,
-0.005375748,
-0.0031470298,
-0.0675507709,
0.0135658998,
0.0427574143,
0.0138330087,
-0.0141836302,
0.0452606753,
-0.0077171051,
-0.0482558869,
0.0358440168,
0.020265175,
0.0082489569,
0.0023113708,
-0.008553748,
-0.0010934821,
0.0220076628,
0.0376335718,
0.0482697114,
-0.026201902,
-0.0144648245,
-0.0123596862,
-0.06742201,
0.0168690048,
0.0383686386,
-0.0314930789,
-0.0029013946,
-0.0093122711,
-0.0203888845,
0.0630604103,
-0.0064270571,
-0.0194848273,
0.0327513441,
0.009838446,
-0.0231961161,
0.012580934,
0.0063883974,
-0.0375082232,
-0.0411759652,
0.029919209,
-0.0427165069,
-0.0039212713,
-0.0110535379,
0.0181238279,
-0.0316286199,
0.0527427979,
-0.0526684113,
0.0299317595,
-0.0169241894,
0.0207771175,
-0.0017077806,
0.027510101,
-0.0141751682,
0.0605564639,
-0.0193809979,
0.0026370785,
0.0523716286,
-0.0253241714,
-0.0169853345,
-0.0799683556,
-0.0063219089,
-0.0266662855,
0.0245187152,
0.0006452093,
0.019425353,
-0.0440948382,
-0.0448281951,
-0.0111087104,
-0.0951792523,
-0.031500496,
0.0092837261,
-0.0006409334,
0.0069673685,
-0.0164625961,
-0.0170785338,
0.0243748911,
-0.0328746364,
0.0154714053,
-0.0128297657,
0.0314694867,
-0.0136319706,
-0.0179182403,
-0.0307035111,
-0.0310574081,
0.0084194886,
-0.0350219347,
0.0604268238,
0.0014067837,
-0.0102809528,
-0.0527807996,
-0.0183211826,
0.0043732016,
0.0269921888,
-0.0277567655,
-0.0190723147,
0.0391832143,
-0.046208594,
0.0801870376,
0.0617816783,
-0.0040620882,
0.0608947277,
0.0782345906,
-0.020075528,
-0.0203916896,
-0.0168785863,
-0.0171004273,
-0.0072075138,
-0.0467261225,
-0.0822226256,
-0.010493245,
0.0891551897,
0.0484237783,
0.0188950598,
-0.0337615386,
0.0451777279,
-0.0062684659,
-0.0678276718,
-0.0238169637,
0.02106346,
-0.0433161147,
0.0775687993,
-0.0055535287,
-0.0331838466,
0.0580076165,
-0.0195984058,
0.0097507974,
-0.0028820001,
0.0217280071,
0.0333010219,
0.0611307248,
0.0264792517,
-0.0000219727,
0.0398764051,
-0.0019108179,
-0.0612145402,
-0.0040576323,
0.0059152665,
-0.0449185818,
-0.0177257173,
-0.0898001194,
0.0117796445,
0.0369267687,
0.0152530903,
-0.0377209224,
-0.0049919263,
-0.0782877579,
0.0032487481,
0.0182036944,
0.0055569531,
-0.0253801979,
0.0135797178,
0.0390939377,
-0.0393468887,
-0.0086763408,
-0.0139960377,
-0.0146875903,
0.0351437181,
0.0620593429,
0.002067341,
0.0485341735,
-0.0684823841,
0.0040323362,
0.0504116155,
0.0653281361,
0.0541764088,
-0.0184136014,
0.0083425557,
0.0292180423,
0.003900921,
0.0043605403,
0.0587980524,
0.019820638,
-0.018628221,
-0.0358496867,
0.0973248556,
0.0218512211,
-0.0598707199,
0.0206773728,
-0.0045793643,
0.0016523609,
-0.0052665719,
0.0229907017,
-0.0640397146,
-0.0335624218,
0.02457528,
0.0067180437,
-0.0471408628,
-0.0649612173,
0.0126422411,
-0.1188191026,
0.014464872,
0.0078473277,
-0.0347830206,
-0.0061038784,
-0.0470643565,
0.0443578884,
0.0260557048,
0.0323530845,
-0.0155608598,
0.0120861493,
0.023989547,
0.0042906119,
-0.0071249288,
-0.0298441909,
0.031691812,
-0.0205327664,
0.0762372911,
0.0556337126,
0.0338484421,
-0.0039426116,
-0.0472628213,
-0.0075351959,
-0.004766596,
-0.0016978022,
0.0033191401,
0.0270359162,
0.0144634526,
0.0018162507,
-0.0175110493,
0.0587607399,
-0.0191030856,
-0.0433533192,
-0.0029159721,
0.0765127093,
-0.0465907529,
-0.0447001867,
-0.0041299411,
0.038916599,
0.024528021,
0.0849251524,
0.0594753325,
-0.0874240249,
0.0118175354,
-0.0398796648,
0.0141153578,
-0.000481774,
-0.0587770194,
0.0383003093,
-0.0004682479,
0.0066493414,
-0.0713120773,
-0.0128980707,
0.056415949,
0.0157939699,
0.0261425171,
-0.0053059817,
0.0687839836,
-0.0185930133,
-0.0055845408,
0.0073290947,
-0.075838618,
-0.0019577858,
0.0036510034,
-0.0303591136,
0.0062759244,
-0.0672373623,
-0.0123740342,
0.0191585142,
-0.0865873843,
-0.0757819787,
0.026269557,
0.0312152319,
0.0254419204,
0.0146795316,
0.0195386037,
-0.0659280419,
0.0076369178,
-0.0236264784,
-0.0114421593,
0.0522585325,
-0.0217474662,
0.0605725311,
-0.009820235,
0.0511438809,
-0.0073652407,
-0.0002656354,
0.0217215177,
0.0406441838,
-0.0423949994,
-0.0489921384,
0.0309538562,
0.0043605524,
0.0554078035,
-0.0261622686,
-0.0115876934,
-0.0215366893,
-0.0469796546,
-0.0353219807,
-0.022080509,
-0.0376334004,
-0.0401392318,
0.0343740545,
-0.0005095318,
0.0004151219,
-0.0921606943,
-0.0295244697,
0.0237574857,
0.0161904804,
-0.0367858522,
0.0030651793,
0.0042780046,
0.0354619771,
-0.0008015,
0.0188222211,
0.0116154039,
-0.0175893251,
0.0507358126,
0.0337713547,
0.0906050578,
0.0180240031,
-0.0502901934,
-0.0364863873,
-0.0506588407,
-0.0652693808,
-0.0620202385,
0.0260010287,
0.0631137788,
0.0527405068,
-0.015502709,
0.0399898812,
-0.0542849153,
0.0276156757,
-0.0104250945,
0.0291993674,
0.0204962883,
-0.0023795713,
0.0526807532,
0.0085721891,
-0.0177527666,
-0.0158697665,
0.0726964995,
0.0011505208,
0.0162287652,
-0.0056462153,
-0.0230477732,
-0.0194819327,
-0.0033894419,
-0.0136581436,
0.0404541269,
-0.0064374171,
0.023090262,
-0.0611966774,
-0.060260877,
-0.0222946648,
0.0186751727,
-0.029705571,
-0.0559751019,
0.0386306085,
0.047005441,
-0.0291062351,
-0.0067890715,
-0.0524790175,
0.0360713378,
-0.0001935465,
-0.0138299251,
-0.0391914137,
-6.442893353e-33,
-0.0334957801,
-0.0333853662,
-0.0175987277,
-0.0251892526,
-0.0120193679,
-0.0490005203,
-0.053098768,
0.0398767367,
0.019975001,
0.0351967253,
-0.0137508167,
-0.0015360572,
0.0167564843,
0.0176483933,
0.0170048308,
0.0076677888,
0.0070286873,
-0.0553230681,
-0.0085127931,
-0.009754641,
0.0270996075,
0.0123265916,
0.0487229563,
-0.0529028885,
0.0403335169,
-0.0217991024,
0.0048796162,
0.0248765443,
0.0111356387,
0.0246562324,
-0.0412005223,
0.0163776521,
0.0287881251,
-0.0275615659,
-0.0090982672,
0.0244326256,
-0.0352983326,
-0.0182634275,
0.0572323501,
0.0115491943,
-0.0354457535,
-0.0315171368,
0.01595103,
-0.0021148783,
-0.0067973849,
-0.002868715,
0.0268800929,
-0.02998708,
-0.0254971236,
0.0104220826,
-0.0265424587,
-0.0095878718,
-0.0078385696,
0.0304994993,
0.0111289946,
-0.004727812,
-0.0520840809,
-0.0007486775,
-0.0024309251,
0.0050643282,
0.0389322117,
0.0034540996,
0.0505627319,
0.0756477863,
0.0630985126,
0.085374333,
0.0291921292,
-0.0189457554,
-0.0215657987,
-0.0050105606,
0.0655065477,
0.0133642722,
0.0010008574,
0.0459870473,
0.0319212936,
0.0448603518,
-0.0169937573,
-0.0073536774,
0.0468768962,
0.0098954225,
0.0679940879,
0.0056885551,
-0.0196592342,
-0.0479517914,
-0.0610612594,
-0.0513622425,
-0.0147080719,
-0.0173324272,
0.0091381883,
-0.0334401205,
-0.0229440853,
-0.0359387062,
0.0572429262,
-0.0430434085,
-0.0396872945,
-0.0736594871,
0.0015419971,
0.0261245165,
0.0008151918,
0.0127803124,
-0.0483727753,
-0.0243341159,
-0.0111471452,
0.0073813149,
0.0066038687,
0.0209386479,
-0.0364939161,
-0.0406809188,
-0.0388548709,
-0.0022544793,
-0.0465583429,
0.008658397,
0.0190996081,
0.0007630747,
0.0562856421,
-0.0028913922,
-0.0035263391,
-0.0281075072,
0.0209000316,
-0.054268077,
-0.0026624338,
0.0041791061,
-0.0346523188,
0.0126120495,
-0.0521097407,
-0.0105677051,
-0.0225163326,
-0.0137633048,
0.0762341693,
-0.0419580042,
-0.0169062801,
0.029324159,
2.588e-7,
0.0481820293,
0.0339774452,
-0.0154503072,
0.0089954911,
0.0342290103,
0.0476043858,
-0.0248810705,
0.0050137904,
0.0556842908,
-0.0611106679,
0.0011138193,
0.0567499213,
0.0521689951,
0.0378841013,
-0.083373189,
-0.0184926856,
-0.0074660913,
0.0126607222,
0.001793667,
-0.0170684997,
0.0272783265,
0.0631413534,
0.0036879347,
0.005342376,
0.0077782259,
-0.0515631065,
0.0766532868,
-0.04524979,
-0.00277185,
0.0389195345,
-0.0069940654,
-0.0171567388,
-0.0079347827,
0.0399324037,
0.0263967253,
-0.0001301106,
0.0011201764,
0.0264016446,
0.0111596128,
-0.018557474,
-0.0016456214,
-0.0069543258,
0.0097960802,
-0.0169480238,
0.0360531807,
-0.088144049,
-0.0388706177,
-0.033369232,
0.0017490727,
-0.0410344526,
-0.0126921898,
0.0124048153,
-0.0041565346,
0.02658193,
0.0052159084,
0.0184227061,
0.0148091661,
-0.0436758511,
-0.0015447872,
-0.0204444528,
-0.0584769212,
-0.0688539371,
-0.011508815,
-0.0352337658,
0.0122826872,
-0.0902471989,
-0.0071451413,
2.205864511e-34,
0.0101323463,
0.0189697016,
-0.0207512379,
0.0503603406,
0.0347374752,
-0.0391011648,
-0.0034278911,
0.0326913595,
0.0574745201,
-0.112560302,
-0.0136519875
] |
http://arxiv.org/abs/2401.16658v2 | http://arxiv.org/pdf/2401.16658v2 | OWSM v3.1: Better and Faster Open Whisper-Style Speech Models based on E-Branchformer | 2024-01-30T01:22:18 | 2024-06-16T20:33:38 | Yifan Peng, Jinchuan Tian, William Chen, Siddhant Arora, Brian Yan, Yui Sudo, Muhammad Shakeel, Kwanghee Choi, Jiatong Shi, Xuankai Chang, Jee-weon Jung, Shinji Watanabe | cs.CL | cs.CL, eess.AS | Recent studies have highlighted the importance of fully open foundation
models. The Open Whisper-style Speech Model (OWSM) is an initial step towards
reproducing OpenAI Whisper using public data and open-source toolkits. However,
previous versions of OWSM (v1 to v3) are still based on standard Transformer,
which might lead to inferior performance compared to state-of-the-art speech
encoder architectures. This work aims to improve the performance and efficiency
of OWSM without additional data. We present a series of E-Branchformer-based
models named OWSM v3.1, ranging from 100M to 1B parameters. OWSM v3.1
outperforms its predecessor, OWSM v3, in most evaluation benchmarks, while
showing an improved inference speed of up to 25%. We further reveal the
emergent ability of OWSM v3.1 in zero-shot contextual biasing speech
recognition. We also provide a model trained on a subset of data with low
license restrictions. We will publicly release the code, pre-trained models,
and training logs. | OWSM v3.1: Better and Faster Open Whisper-Style Speech Models based on E-Branchformer: Recent studies have highlighted the importance of fully open foundation
models. The Open Whisper-style Speech Model (OWSM) is an initial step towards
reproducing OpenAI Whisper using public data and open-source toolkits. However,
previous versions of OWSM (v1 to v3) are still based on standard Transformer,
which might lead to inferior performance compared to state-of-the-art speech
encoder architectures. This work aims to improve the performance and efficiency
of OWSM without additional data. We present a series of E-Branchformer-based
models named OWSM v3.1, ranging from 100M to 1B parameters. OWSM v3.1
outperforms its predecessor, OWSM v3, in most evaluation benchmarks, while
showing an improved inference speed of up to 25%. We further reveal the
emergent ability of OWSM v3.1 in zero-shot contextual biasing speech
recognition. We also provide a model trained on a subset of data with low
license restrictions. We will publicly release the code, pre-trained models,
and training logs. | [
-0.0462234691,
0.0448204726,
0.0281156637,
0.0448988304,
-0.0104362555,
-0.0229409784,
0.0447400473,
0.0020217486,
-0.019989321,
-0.0148065863,
-0.0691181421,
-0.0091983732,
0.0165372007,
0.0725360364,
-0.0073021026,
-0.0193523597,
0.0107835783,
-0.0621932261,
-0.0062195375,
-0.0151875662,
-0.008086618,
-0.0040864497,
-0.0506714694,
0.0667406097,
0.0116142072,
-0.0083371447,
-0.0236122068,
0.016121693,
-0.0012726499,
-0.0053999797,
-0.0145588573,
0.0159896128,
0.0328033082,
0.0264404975,
0.0000023071,
-0.0242509712,
-0.024143843,
0.0158246066,
-0.1006216705,
0.0011251455,
0.0186841376,
0.0045817774,
-0.033005394,
0.0569983199,
-0.0102340868,
-0.0705680698,
0.0581448115,
0.0153253293,
0.0343836248,
0.0800111517,
-0.0006140649,
-0.0445693992,
0.0447148532,
-0.0080908407,
-0.0129800439,
-0.018229533,
0.0067923875,
-0.0781683624,
-0.0371927544,
-0.0289891157,
-0.0453208871,
0.0169817433,
0.039515242,
0.0048344899,
-0.0017614005,
0.0367884822,
-0.0164295305,
-0.0437992997,
-0.013616248,
0.0539179891,
0.0021514404,
-0.0099336607,
-0.0090670651,
-0.0214239918,
-0.0665168464,
-0.026829483,
-0.0466618426,
0.0679920688,
0.0274181776,
0.0016485215,
-0.0138938334,
0.0640938058,
0.0220583268,
0.0050173001,
-0.0136587918,
0.0629277229,
0.0093195224,
-0.0009454179,
0.0045172716,
-0.0322607979,
-0.009257745,
-0.0607637912,
-0.0029841517,
0.0259987377,
-0.0162367783,
-0.0102046924,
-0.0610036254,
-0.0143828019,
-0.0309847947,
-0.0150622502,
0.0286831297,
0.0591156147,
-0.0522858389,
-0.0000654714,
-0.0103229256,
0.0311524551,
-0.0132488711,
0.0024737346,
-0.0238573551,
-0.043454051,
0.036750976,
0.0237644315,
-0.0281092953,
0.0094292993,
-0.0725321248,
-0.0522732362,
-0.0299962498,
0.0095099974,
-0.0134667167,
-0.0252336152,
0.0611479245,
-0.009217659,
-0.0376664326,
-0.023220988,
0.017325256,
0.0302412063,
0.0266692676,
0.0366068371,
0.0264435783,
-0.0932256877,
-0.0480228662,
0.0482307859,
0.0207200982,
-0.004270738,
-0.0099212257,
0.0394447446,
0.0250253547,
-0.0078971349,
-0.0383100696,
-0.0136165703,
-0.0212389193,
-0.034670122,
-0.0010596699,
-0.0432620682,
0.0147438934,
-0.0039730915,
0.0199718196,
-0.0014542011,
0.0034915952,
-0.0111307986,
0.0001090208,
0.0038333093,
-0.0263901018,
-0.0145681519,
0.0228784848,
-0.0074107237,
0.0046335254,
0.0490258262,
0.0519878194,
0.0047518914,
0.0098816967,
-0.0264696423,
0.0189118218,
0.0706595182,
0.0287973061,
0.0086582061,
0.033204861,
-0.0012547151,
0.0114793489,
0.0173008647,
0.0516291298,
0.030842524,
-0.0329599977,
-0.0362490639,
0.0363539681,
0.0168323144,
0.0668316782,
-0.017643448,
0.0453228466,
0.0060775075,
-0.0290796943,
-0.0546208918,
0.0001398473,
0.0365969278,
-0.011332049,
0.0221187528,
0.0180386603,
0.020300936,
-0.0546361506,
0.0072774626,
-0.0014406112,
-0.0432079323,
-0.0302622449,
0.0145276114,
0.0577061102,
0.0200400148,
-0.066808641,
0.0797045827,
0.0542386249,
-0.0973248258,
-0.0000676704,
-0.057719741,
0.0305611938,
0.0335214697,
0.0250908658,
0.0217589289,
-0.0044687972,
-0.0348885246,
0.0146516906,
0.0437916182,
-0.0142455995,
0.0003723217,
0.0733882338,
-0.0938455313,
-0.0436789729,
0.0162422135,
-0.0084962761,
0.0247954763,
0.0241823364,
-0.011199099,
0.0045736646,
-0.022995472,
0.0205311067,
0.0022936929,
-0.0299330112,
-0.0047072698,
0.0353736281,
0.0123530924,
0.0560679734,
0.0307997651,
-0.0230440926,
-0.0025119949,
0.022106953,
-0.0454048775,
-0.0134517113,
0.0187570527,
-0.0648190156,
-0.0033540926,
0.0130482949,
0.0170529652,
0.0666002259,
0.0152555322,
-0.0229894742,
-0.0788237453,
0.029216053,
0.0141491406,
0.0800329372,
-0.0403958559,
0.0242219493,
0.0085617425,
0.0134659884,
0.0204259846,
-0.0074891597,
-0.0115596205,
-0.0373492502,
0.004118707,
0.0408828072,
-0.0070101381,
-0.0458995774,
-0.0244616829,
-0.0090262629,
-0.0300040524,
-0.0064833304,
-0.0480544604,
-0.0114303418,
0.0019612201,
0.0053747408,
-0.1187173948,
0.0036230218,
-0.0340487994,
0.0055309855,
-0.004082792,
0.0487471893,
-0.0170377288,
0.0147713963,
0.0047551277,
0.0258347671,
-0.0398211963,
-0.0307704005,
-0.0664311349,
-0.0121018114,
-0.0213927198,
-0.0135339489,
-0.0044448259,
0.0120298453,
0.0390919633,
-0.0312643386,
-0.0071409666,
-0.0199727546,
0.0224925708,
0.0308056567,
0.048113402,
-0.0084681027,
-0.0049160495,
-0.0044004819,
0.0087912949,
0.0395040996,
0.0391069911,
-0.0011177179,
0.0182447974,
-0.0034843555,
-0.0154761691,
-0.0348702855,
0.0034902103,
0.0006195986,
-0.0050565889,
-0.0208437555,
0.0168036465,
0.0006185896,
0.0308798645,
0.0212875009,
0.0317245871,
-0.0627976879,
-0.0254579186,
0.006670191,
-0.0189715438,
0.0464208052,
-0.0468750596,
-0.0133338254,
-0.0128068291,
-0.017370807,
0.002690013,
-0.0089131622,
-0.103290543,
-0.0327369496,
0.0010054491,
0.0029588207,
-0.0338626355,
0.0086767972,
-0.0371817276,
-0.0000418534,
0.0635297447,
0.0034905174,
-0.0005781145,
0.0221258271,
-0.0158699788,
-0.019651087,
-0.0002606955,
0.0000443568,
-0.0050631296,
-0.0128093893,
-0.0180309266,
-0.0239090156,
0.0088339271,
-0.0437683202,
0.0516720004,
-0.0025621885,
-0.0136412457,
-0.0277100056,
0.0551770926,
-0.0012522223,
0.0817784071,
-0.0030671856,
0.0297841318,
-0.0102634495,
0.035830684,
0.020526506,
0.0111034764,
0.0219935924,
-0.0331852138,
0.0575231761,
0.0433744118,
0.0637174621,
0.0411055237,
-0.0642196685,
-0.030941952,
0.0261940677,
-0.0010941662,
0.0330428705,
0.036369767,
0.0606425442,
-0.0317042917,
0.0012476154,
0.1223659217,
-0.0016263264,
-0.0473133884,
-0.0111640617,
0.0040479917,
0.033586178,
0.0016785109,
-0.0303951986,
-0.078312017,
-0.0510261618,
0.0457900241,
0.0335948877,
-0.0058374694,
0.0038311451,
0.0016419159,
-0.1466114372,
0.0588856041,
0.0057615577,
-0.0557234995,
-0.0082301749,
-0.0118462592,
-0.0235992074,
-0.0673161596,
0.0175200794,
0.036613483,
0.0046144556,
-0.0242485795,
0.0323636457,
-0.0383249782,
-0.0465087257,
0.0319781676,
-0.0744436234,
0.0633070022,
0.0532946773,
-0.0512843579,
0.0157675259,
-0.0251857024,
0.0803009793,
-0.0287743881,
0.0010989702,
-0.0287669469,
0.016082231,
-0.0372832566,
0.0003090969,
0.013511382,
0.0543597639,
-0.0178913586,
-0.0090206182,
-0.0050440244,
0.0333323814,
-0.0179481208,
0.0137953581,
-0.0038194184,
0.0129598929,
-0.0018445032,
0.0511029549,
0.0141666131,
-0.0262715146,
0.0069053019,
0.0000163587,
0.0162380431,
0.0160757471,
-0.0126702748,
0.0053793513,
0.0427529179,
-0.0931274667,
-0.0212125611,
0.028033888,
0.0205792505,
0.0581845455,
0.0161926262,
-0.0549775921,
0.0322395675,
0.0034341584,
-0.0320831686,
-0.0265563708,
-0.0153499367,
0.024869835,
-0.046032019,
0.0177432932,
0.0567968898,
-0.0109359706,
-0.0124265384,
0.0070591387,
-0.1082080007,
-0.0378616676,
-0.0008451966,
0.0566152744,
0.0189427733,
0.0396766588,
-0.030279221,
-0.0310856551,
0.0552100278,
-0.0104350382,
-0.0364907011,
0.0727046728,
0.0550727844,
0.0012334122,
0.0044701328,
0.0271735583,
-0.0496788807,
0.0154566467,
0.0425313786,
0.0101349466,
-0.0238750838,
0.0238128435,
-0.0089762658,
-0.028148016,
0.061041683,
0.0130904773,
0.0088258293,
-0.0293503851,
0.0700126439,
-0.0303222574,
0.0359127074,
-0.0524023287,
-0.0296132602,
0.0079669459,
0.0126427058,
0.0287457295,
0.0357281379,
-0.0273099896,
-0.0028106016,
-0.0309259314,
0.0627013892,
0.0453272201,
0.0182338394,
-0.0267177019,
-0.0175466016,
-0.0193885546,
0.0470960513,
0.0094488729,
0.0229021329,
-0.0044158283,
0.0540474653,
-0.0183567517,
0.0297054555,
0.0147411237,
0.013631383,
-0.0468751378,
-0.0420119949,
-0.0552751496,
0.0358778425,
-0.0027674315,
0.0204863492,
0.0267674029,
0.0240573343,
0.0170902684,
-0.015085225,
0.0327944346,
-0.03583223,
-0.0277804434,
0.0501653925,
0.0114197508,
-0.0095683094,
0.0296418909,
0.0183559991,
0.0190830622,
0.0365663804,
-0.0071530649,
0.0657906979,
-0.0145779531,
-0.0403959192,
0.009113133,
0.0046769376,
-0.0273032151,
-0.0066419155,
-0.0256838314,
-0.0169675015,
0.0407274961,
0.0091020614,
0.0426732972,
-0.0194079019,
0.0508031286,
-0.0121870097,
-0.031272009,
-0.0344646275,
0.0117707262,
-0.0830345973,
-0.022681661,
-0.0276968516,
0.0246398188,
-6.956132554e-33,
-0.0468830653,
0.0339926444,
0.0382236764,
0.0189057123,
-0.0461959727,
0.0248998906,
-0.0485370904,
0.0008878976,
-0.0319395587,
-0.045580633,
0.0057825362,
0.0390583649,
0.0396635085,
0.0518678389,
-0.0302896667,
-0.0117908781,
0.0774370804,
-0.0306309666,
-0.0065591135,
-0.0131610828,
0.0138015989,
0.0279528555,
0.0044457833,
-0.0414117612,
-0.0356874317,
-0.0081008831,
-0.0471438505,
-0.0226829909,
0.0157706477,
0.0349390879,
-0.0090964781,
0.0083504124,
-0.032518357,
-0.0477033816,
-0.0036054205,
-0.0201920941,
-0.0353021286,
0.0603016019,
-0.0279249381,
0.0313945711,
0.0068566403,
-0.0023150658,
0.0370656401,
0.0600971654,
-0.0480994396,
-0.0384210236,
-0.0062217624,
-0.0111461142,
-0.0177054554,
-0.0206734017,
-0.0209061317,
0.029234346,
-0.0028869014,
-0.0462023169,
0.0442057475,
0.0391278565,
-0.0293607432,
-0.0016462083,
-0.0595370047,
0.0234228,
0.0378845036,
0.0010618137,
0.0390045047,
0.0189842544,
-0.0087334104,
0.0400097482,
0.0651048571,
-0.0607816987,
-0.0211172942,
-0.0079169944,
0.0488038175,
0.0529912561,
0.0623929128,
0.0079011014,
-0.0073269443,
-0.001576086,
-0.056754522,
-0.0105044292,
0.0262732357,
0.0370571092,
0.0272836592,
0.032815177,
-0.0317559987,
-0.0208719596,
-0.032270059,
-0.138705641,
-0.042106472,
0.0251366403,
0.0349219479,
0.0264924914,
0.000484929,
-0.0472811162,
0.0851007998,
-0.0317314267,
0.0852560252,
-0.0077799717,
0.0047229589,
0.0462439172,
-0.0240602754,
0.0103208553,
-0.067398265,
-0.0384405144,
0.0854676813,
0.0435003489,
0.008964153,
-0.0068738251,
-0.0065022744,
0.0703700706,
-0.0373419002,
-0.0233667418,
-0.0050646714,
0.0201427918,
0.0176124014,
0.0536573343,
0.0438571833,
0.0010238354,
0.0090509253,
0.0591732003,
-0.0513985567,
-0.0052405442,
-0.0240993481,
-0.0256037619,
0.0432661623,
0.0031327226,
-0.0438750125,
0.0180819388,
0.0322021432,
-0.0336354338,
0.0064228009,
-0.0379862413,
-0.0286754109,
0.0274984911,
3.035e-7,
-0.0243316013,
-0.0139839444,
0.0360730141,
0.0015959117,
0.0322085209,
0.0298055038,
-0.0046236585,
0.0071626389,
-0.0119112264,
0.0161125734,
-0.007617305,
0.0185835157,
0.0830756202,
-0.016163094,
-0.0836803615,
-0.0161903966,
-0.0625271052,
-0.0077448403,
-0.0307896771,
-0.0107976766,
-0.0365507528,
0.0972034261,
-0.0114210481,
-0.0040820399,
0.0348040536,
0.0536711812,
0.0168754533,
-0.0146262059,
0.0172211099,
-0.0165528897,
-0.0883855671,
0.0283158086,
-0.0281068143,
-0.0020591044,
-0.0231981408,
0.0300853625,
-0.0196532607,
0.0722717941,
0.0003592989,
-0.0940624252,
0.0198702123,
0.0508723073,
-0.0294153541,
-0.0327589214,
0.0676234663,
-0.0167059563,
-0.0296541508,
-0.0320575759,
-0.0400768444,
0.0234461371,
-0.0420372523,
-0.0020456908,
0.0255974662,
-0.0085417256,
0.0048306123,
-0.0291214418,
-0.0186521746,
-0.042342145,
-0.0139469178,
-0.0560228676,
-0.0442729928,
-0.0370940082,
-0.004674627,
-0.022031175,
0.0314006396,
0.0286502782,
-0.0126349786,
2.702734034e-34,
0.0407039449,
-0.0162922554,
0.0107120825,
0.0417941995,
-0.0187674649,
-0.0365164652,
0.0399522185,
-0.0152487997,
-0.014271982,
-0.0911659896,
-0.0610330738
] |
http://arxiv.org/abs/2401.16656v1 | http://arxiv.org/pdf/2401.16656v1 | Gradient-Based Language Model Red Teaming | 2024-01-30T01:19:25 | 2024-01-30T01:19:25 | Nevan Wichers, Carson Denison, Ahmad Beirami | cs.CL | cs.CL | Red teaming is a common strategy for identifying weaknesses in generative
language models (LMs), where adversarial prompts are produced that trigger an
LM to generate unsafe responses. Red teaming is instrumental for both model
alignment and evaluation, but is labor-intensive and difficult to scale when
done by humans. In this paper, we present Gradient-Based Red Teaming (GBRT), a
red teaming method for automatically generating diverse prompts that are likely
to cause an LM to output unsafe responses. GBRT is a form of prompt learning,
trained by scoring an LM response with a safety classifier and then
backpropagating through the frozen safety classifier and LM to update the
prompt. To improve the coherence of input prompts, we introduce two variants
that add a realism loss and fine-tune a pretrained model to generate the
prompts instead of learning the prompts directly. Our experiments show that
GBRT is more effective at finding prompts that trigger an LM to generate unsafe
responses than a strong reinforcement learning-based red teaming approach, and
succeeds even when the LM has been fine-tuned to produce safer outputs. | Gradient-Based Language Model Red Teaming: Red teaming is a common strategy for identifying weaknesses in generative
language models (LMs), where adversarial prompts are produced that trigger an
LM to generate unsafe responses. Red teaming is instrumental for both model
alignment and evaluation, but is labor-intensive and difficult to scale when
done by humans. In this paper, we present Gradient-Based Red Teaming (GBRT), a
red teaming method for automatically generating diverse prompts that are likely
to cause an LM to output unsafe responses. GBRT is a form of prompt learning,
trained by scoring an LM response with a safety classifier and then
backpropagating through the frozen safety classifier and LM to update the
prompt. To improve the coherence of input prompts, we introduce two variants
that add a realism loss and fine-tune a pretrained model to generate the
prompts instead of learning the prompts directly. Our experiments show that
GBRT is more effective at finding prompts that trigger an LM to generate unsafe
responses than a strong reinforcement learning-based red teaming approach, and
succeeds even when the LM has been fine-tuned to produce safer outputs. | [
0.0224654302,
0.0546105839,
-0.0006216744,
0.0158346966,
-0.0217252597,
0.0049848496,
0.0172616392,
-0.0108385207,
0.0060093217,
-0.0383039489,
-0.0348207094,
-0.0317032263,
-0.0031198177,
0.0236639809,
0.0362294093,
-0.0578217432,
0.0747928172,
-0.0438105203,
-0.0049861353,
-0.0177853014,
0.0154290376,
0.0267625079,
-0.0080752652,
0.0594508462,
-0.0301861037,
-0.0377431102,
-0.0049403836,
0.0424470827,
0.0272999387,
-0.0305145234,
0.0458387807,
0.0650464669,
-0.0057681333,
0.062067546,
0.0000022822,
-0.0317923017,
-0.0348072797,
-0.0161835607,
0.0018569967,
0.0042339894,
0.0026803762,
-0.032106474,
-0.0334965698,
0.0430040285,
-0.033576034,
-0.0339338481,
0.0344336368,
0.0496121272,
0.0646107495,
0.0857969597,
0.0173112657,
0.0193505082,
0.0300624408,
-0.0558052957,
-0.0185594261,
-0.063855961,
0.0247819666,
-0.0283724666,
-0.0101320446,
0.0174591877,
-0.0240355115,
0.0258829724,
0.0325664319,
-0.0305173248,
-0.0476715751,
0.0388442911,
0.0042039324,
-0.0269453675,
-0.0204447992,
0.0871270671,
-0.079269208,
-0.0079854662,
0.0034576957,
-0.0239224359,
0.0080734203,
-0.0089454241,
0.0102618244,
0.0150289442,
-0.0092303958,
0.0275769327,
-0.0053223642,
-0.0034942375,
0.0460781567,
0.0175017938,
0.0403706282,
0.0367395133,
0.0080702351,
-0.0340455361,
-0.0257990677,
0.0031366614,
-0.0220592674,
-0.0429985225,
-0.0219573341,
-0.0049248659,
-0.0614179485,
0.0322025195,
-0.0784582198,
-0.0436279848,
0.0553598069,
0.0046182219,
0.0470787622,
0.0027839055,
-0.0182204656,
0.0413296781,
-0.0286019463,
0.0507851206,
0.0039456533,
0.0518886931,
-0.0304455366,
0.0211333521,
-0.0001259079,
-0.0144645134,
-0.0279445481,
-0.0300843827,
-0.0124453362,
-0.0056572957,
-0.0194807518,
0.0124501213,
-0.0272881091,
0.0464060493,
-0.0029299883,
0.0159581862,
-0.0475771762,
-0.0248948615,
-0.0214647949,
-0.0275986064,
-0.0241403375,
-0.0034596652,
0.009740321,
-0.0841973945,
-0.0285534952,
0.0204473287,
-0.0471945368,
-0.0196439642,
-0.0190022495,
0.0748593807,
-0.0064871721,
-0.0053326613,
-0.011808997,
-0.0208854042,
-0.0436407924,
-0.0495754071,
-0.0247657821,
-0.0496086553,
0.0163664352,
-0.0276845228,
-0.0207866002,
-0.0804302171,
0.025315959,
0.0129900277,
-0.0251520779,
0.0959157348,
-0.002636234,
-0.0186155066,
0.0200443063,
0.0235304721,
-0.004644034,
0.068181552,
0.006423899,
0.0348067395,
0.0156696588,
0.0008510494,
0.0504461639,
0.0422051288,
0.0406330451,
0.0183118526,
-0.019020658,
-0.017456159,
0.0525028929,
0.0358192325,
0.0299739111,
0.0926784649,
-0.006828981,
-0.0373165011,
0.0874265656,
0.008744102,
0.0087535158,
0.009466501,
0.048182372,
0.0136723127,
-0.0540065356,
0.0276539735,
-0.0604323037,
0.0045345156,
-0.0421709158,
-0.0096843746,
0.0377757847,
0.0336651579,
-0.0191567969,
-0.0381401628,
-0.0122919362,
0.001967598,
-0.0053927666,
0.0092247399,
0.0070777475,
0.0370378643,
-0.0897543952,
-0.0310703963,
0.0224499032,
-0.0235823747,
0.0057162326,
-0.0581266955,
0.0477487929,
0.0715827867,
0.0502474755,
-0.005351786,
-0.021138873,
-0.0524249934,
-0.0034165659,
0.040606942,
0.0349048115,
0.006408697,
0.0061529656,
-0.068243131,
-0.0388244651,
-0.0014941299,
0.0159393977,
0.0362892114,
0.0188125614,
-0.0022737766,
0.005503241,
0.023570057,
-0.0349418446,
0.0568392426,
-0.0016875586,
-0.0474735387,
0.0042903321,
0.0090934327,
0.0214751232,
-0.0312510952,
-0.0134613803,
0.0118927881,
0.0271323025,
-0.0095151616,
-0.00700976,
-0.0141185811,
-0.0537266918,
-0.0588090941,
-0.0007806442,
-0.001531173,
-0.042181503,
0.0487446785,
-0.0582443662,
-0.0068314318,
-0.0053860261,
-0.0234523006,
-0.0258274078,
-0.0454416275,
0.0180916339,
0.018030867,
0.0299142748,
-0.0529159494,
0.0737735033,
-0.0358756669,
0.0119508151,
-0.0237440951,
0.0328576267,
-0.0128919762,
-0.075808771,
-0.0270783119,
0.0219762586,
-0.0039957757,
0.0545104519,
0.0204759408,
0.0217397735,
-0.022200672,
-0.0234042257,
-0.0013394143,
0.0108058415,
-0.0438041911,
-0.0046295696,
0.0217514075,
-0.0351517573,
-0.0210401379,
-0.0200956222,
-0.0153694991,
-0.0086718379,
0.0152514605,
-0.0047693267,
0.0222192388,
-0.0130652552,
0.0109431697,
-0.0412236713,
0.0022891061,
-0.0043347417,
0.0631614178,
0.0035886425,
0.0045767925,
-0.0598632246,
0.0133111784,
-0.0098284716,
0.0432371497,
-0.0116234822,
-0.0671295002,
0.020910373,
0.0011334358,
0.0455885269,
0.0696116611,
0.040073432,
0.0406808816,
0.0660914779,
0.0355059505,
-0.0626118332,
0.0214885604,
-0.0386898294,
-0.0411151163,
0.0152112981,
-0.0251403563,
0.0131269051,
0.0725988075,
0.0304301381,
-0.0024251335,
-0.0514813587,
0.0221966784,
-0.00048728,
-0.0525638647,
-0.034945,
-0.0365063287,
-0.0384644419,
0.0078494344,
-0.0555168018,
-0.0572823361,
0.081726186,
-0.0326240472,
0.016493151,
-0.028815316,
-0.0240615085,
0.0202547312,
0.03743295,
0.0094806449,
0.0245466754,
0.0258777328,
-0.015238327,
-0.0523716398,
-0.0259791967,
0.0272446536,
-0.0264140945,
0.0349919125,
-0.0464964807,
-0.0035649228,
-0.0059442087,
-0.0291495565,
-0.1042295545,
-0.0465642549,
-0.0043466603,
0.0154429693,
0.0044148066,
0.0009584412,
-0.0162479505,
-0.0012262282,
0.0039968505,
0.0312938057,
-0.0197437126,
0.0043240977,
-0.014583909,
0.0313466266,
0.0674152076,
-0.0327285826,
0.0140530756,
-0.0552380383,
0.0498232283,
0.0697378293,
0.0059688534,
0.0421770997,
-0.0495344102,
-0.0225080047,
0.0210528821,
0.0177113116,
-0.0039773663,
0.0234747082,
0.0165282823,
-0.0616270937,
-0.0097349705,
0.0673463494,
0.007481223,
-0.0281265788,
0.0055371649,
0.0187919047,
0.0610688403,
-0.0037736192,
-0.0007439272,
-0.0434863158,
-0.0200011041,
0.016049169,
0.0427471213,
-0.0050475979,
-0.0000701233,
-0.0177330449,
-0.1286785007,
0.0176281165,
-0.0078924224,
-0.0317191146,
-0.0018850403,
0.0029913445,
0.0110270167,
0.0138137843,
0.0688140169,
-0.0022936836,
0.0131409662,
0.0064565563,
0.0321329832,
-0.0161113366,
-0.0691549107,
-0.0229266956,
-0.0511268862,
0.0763504282,
0.1194173619,
0.0270862971,
-0.0043851426,
-0.0357931443,
0.0440215766,
-0.0040248595,
0.0039269947,
-0.0019254222,
0.0448778383,
-0.0301393885,
-0.0247662533,
0.0064493231,
0.0432867073,
0.0166793503,
-0.0043726554,
0.0243750643,
0.0265962798,
-0.0138484761,
-0.0218141545,
-0.0188516639,
0.0333018117,
0.0645508468,
0.0937953517,
0.028236106,
-0.0486311615,
-0.0171413254,
-0.0510294773,
0.0349183269,
-0.0739963278,
-0.0481503606,
0.0132776806,
0.0092929117,
-0.0226849727,
-0.0716000348,
-0.0258699134,
0.0109150214,
-0.0024920965,
0.0088564679,
-0.0060736765,
-0.0875571817,
0.01196496,
-0.0107517773,
0.0356371664,
-0.0055162096,
0.0310170017,
-0.0851704553,
0.0171858165,
0.0126112401,
-0.0469422676,
-0.0225812886,
0.0530489124,
-0.0573153459,
-0.0861656517,
0.0094972262,
0.0948302373,
0.0074743568,
0.0215207655,
-0.0462576151,
-0.0607812442,
0.0803386942,
0.0786450803,
-0.0203023106,
0.0077861049,
0.0352101475,
-0.0181912836,
-0.0064007151,
0.0248904079,
-0.021696724,
0.0233076662,
0.0416294485,
0.0141975367,
-0.0509178005,
-0.0319366455,
0.0141502405,
-0.0123522729,
0.08231318,
-0.0350257754,
-0.0065593785,
-0.0330034383,
-0.0045245895,
0.0176321249,
-0.0332037285,
-0.0329329148,
0.0257895738,
0.0437849984,
-0.023522675,
0.0018044976,
0.0157768652,
-0.0112734931,
0.044626642,
-0.0397607274,
0.0294767991,
0.0388059169,
0.0469971858,
0.0366729051,
0.0254762825,
-0.0115247266,
0.0390163586,
-0.0143492259,
0.0050590802,
0.0149017507,
0.0387253053,
0.0298880171,
-0.0065456508,
0.0203284528,
0.0392973572,
-0.0350521356,
-0.0622959062,
-0.0188000109,
0.0463171676,
0.0487025604,
0.0220220368,
0.0325174741,
-0.012952893,
-0.0195407178,
-0.0257374328,
0.0222613532,
0.0250827037,
-0.0590813272,
0.0389683805,
0.0229593273,
-0.0145461205,
0.0121776862,
0.0662969723,
-0.0108949486,
0.0597503223,
0.010029193,
0.0099167163,
-0.039234098,
-0.0073394156,
0.0120830722,
-0.0052750674,
-0.0212361403,
0.0028469064,
-0.0363290533,
-0.0589259416,
-0.0194800384,
0.0561816059,
0.006623697,
0.0160460342,
0.0185042564,
0.0249952991,
-0.0240768231,
0.0298860818,
-0.0416574739,
0.0170492232,
-0.0121739702,
-0.0495590977,
-0.0263668429,
-6.62411197e-33,
-0.0001894321,
-0.0485643037,
-0.0204042215,
-0.037862841,
-0.0134248696,
0.0340249315,
-0.0141229741,
-0.0144883115,
-0.0184667073,
0.016392421,
-0.0023581178,
-0.0122162085,
-0.0045739757,
0.0208845623,
-0.0391892605,
-0.0447414331,
0.0634882003,
-0.0102074193,
-0.0347669758,
0.0093494551,
0.0436684862,
0.0277986974,
0.0100812176,
-0.0566041954,
0.0182066709,
-0.0395496227,
-0.0817098841,
0.0165259968,
-0.0387432761,
0.0518157594,
-0.007521281,
0.0199138783,
-0.012845004,
-0.0545610264,
-0.0167332832,
0.0705278292,
-0.0480541103,
-0.0068307552,
-0.0007179336,
-0.0085086422,
-0.0929055139,
-0.0415143073,
0.0408889987,
-0.0241470113,
0.0020456896,
-0.0083290283,
0.0160613507,
-0.0082762744,
-0.0423521549,
-0.0235185567,
-0.0356074944,
0.0150375823,
0.0110355569,
-0.0770494342,
0.01896278,
0.0355416536,
0.0246504787,
-0.0131007396,
-0.0332183354,
-0.0096189994,
-0.0013786237,
0.0076056011,
0.0613436997,
0.0396224,
-0.0223924685,
0.0820276886,
-0.0200399403,
-0.0116457529,
-0.0485374741,
-0.0184904002,
0.00049206,
0.0348227583,
0.04497087,
0.0772309899,
0.0212579034,
-0.0202990286,
-0.0511418842,
0.0337726176,
0.0181772448,
0.0305302013,
0.0392444357,
0.0106631583,
-0.0407548137,
-0.0079913652,
-0.0471995957,
-0.0532113649,
-0.0200334676,
-0.0373315923,
0.0305387378,
0.0251497235,
-0.055083923,
-0.0281510111,
-0.0070875939,
-0.0148640713,
0.0119777257,
0.0506041981,
-0.0042049196,
-0.036010921,
-0.0058645089,
0.0264377985,
-0.0221834518,
0.0086878687,
0.0449104756,
-0.0343138725,
0.0164373256,
0.0134180551,
0.065845646,
0.018050842,
-0.062054392,
0.024854701,
-0.0264311004,
0.0821969137,
-0.0038968993,
0.0065298476,
0.049974598,
-0.0217963811,
-0.0165252145,
0.0148785179,
0.0282952823,
-0.0120705245,
-0.019864656,
0.0161325131,
-0.0660077929,
0.0305984523,
-0.0440923423,
0.0043447968,
-0.0002140848,
-0.0053042537,
0.0562377498,
-0.0575824231,
-0.0028672162,
-0.0103003811,
2.855e-7,
0.0037454085,
0.0283610597,
0.0412737317,
0.0118291965,
-0.0017925183,
0.0483008027,
-0.0154071236,
0.0646079555,
0.0071546268,
-0.0071781087,
-0.0126568303,
0.0259182528,
0.0620978735,
0.0586597212,
-0.0550412275,
0.0296500456,
-0.0929485187,
0.0482287072,
-0.0440709293,
-0.0223161634,
0.0416781008,
0.0887271315,
0.0400051475,
-0.0224164668,
0.0350035839,
-0.0322749615,
0.036509268,
-0.0441186167,
-0.0072053992,
0.0016199437,
0.0381870866,
0.0106413085,
-0.0116996113,
0.0176624134,
-0.0145031298,
-0.0304244999,
0.0080852387,
0.0519570857,
0.0206672587,
-0.0456167497,
-0.0183848981,
0.0060589081,
0.0079143513,
-0.054150112,
0.0708518401,
-0.012764384,
0.0168893225,
-0.0434272699,
0.0041899215,
-0.0185405537,
0.0645922348,
-0.0460629538,
-0.0295089856,
0.0098098814,
0.0196001399,
0.0134333493,
-0.0121740242,
-0.0618604571,
-0.0354059897,
-0.0308378916,
-0.0222675987,
-0.0374977365,
-0.0085197641,
0.0418494083,
0.0066268253,
-0.0586656667,
0.0018997457,
2.56187675e-34,
0.0234422255,
-0.0370949022,
-0.0081350822,
-0.0220084116,
0.020510288,
0.0129923606,
-0.0083184615,
0.0141652096,
0.0365629829,
-0.0135376249,
-0.0184805598
] |
http://arxiv.org/abs/2401.16646v1 | http://arxiv.org/pdf/2401.16646v1 | Incoherent Probability Judgments in Large Language Models | 2024-01-30T00:40:49 | 2024-01-30T00:40:49 | Jian-Qiao Zhu, Thomas L. Griffiths | cs.CL | cs.CL, cs.AI | Autoregressive Large Language Models (LLMs) trained for next-word prediction
have demonstrated remarkable proficiency at producing coherent text. But are
they equally adept at forming coherent probability judgments? We use
probabilistic identities and repeated judgments to assess the coherence of
probability judgments made by LLMs. Our results show that the judgments
produced by these models are often incoherent, displaying human-like systematic
deviations from the rules of probability theory. Moreover, when prompted to
judge the same event, the mean-variance relationship of probability judgments
produced by LLMs shows an inverted-U-shaped like that seen in humans. We
propose that these deviations from rationality can be explained by linking
autoregressive LLMs to implicit Bayesian inference and drawing parallels with
the Bayesian Sampler model of human probability judgments. | Incoherent Probability Judgments in Large Language Models: Autoregressive Large Language Models (LLMs) trained for next-word prediction
have demonstrated remarkable proficiency at producing coherent text. But are
they equally adept at forming coherent probability judgments? We use
probabilistic identities and repeated judgments to assess the coherence of
probability judgments made by LLMs. Our results show that the judgments
produced by these models are often incoherent, displaying human-like systematic
deviations from the rules of probability theory. Moreover, when prompted to
judge the same event, the mean-variance relationship of probability judgments
produced by LLMs shows an inverted-U-shaped like that seen in humans. We
propose that these deviations from rationality can be explained by linking
autoregressive LLMs to implicit Bayesian inference and drawing parallels with
the Bayesian Sampler model of human probability judgments. | [
0.0232499503,
0.088204287,
-0.0182175711,
-0.0188047457,
-0.0403374545,
0.0321198292,
-0.0194970686,
-0.020788582,
-0.0146774547,
-0.0340391882,
-0.0202035718,
-0.0634357259,
0.0322374739,
0.0276415646,
0.0275163278,
-0.0543864071,
0.0106762014,
-0.0767696649,
0.0672881454,
-0.0270384755,
-0.01165024,
-0.0243219864,
-0.0274114627,
0.0291155055,
0.0451231971,
0.000284641,
-0.003171816,
-0.0389202684,
-0.0008098561,
-0.0477670319,
-0.0438148379,
0.0587238371,
-0.0181323513,
0.0277114436,
0.0000015998,
-0.0585688911,
0.0217188168,
-0.0500866622,
-0.0137444669,
0.0216334667,
-0.0313641652,
0.0186489988,
-0.0012551631,
0.0067589534,
-0.0199647583,
-0.0024305626,
0.0249455478,
0.0590645187,
-0.0477890745,
0.027309468,
0.0172814075,
-0.0202073883,
0.02368889,
-0.050462842,
0.0323088914,
-0.0382677242,
0.0350142606,
0.0305856317,
-0.0457039364,
0.0082370322,
-0.0155300675,
-0.0119709866,
-0.0011253443,
0.0227732491,
-0.0218332261,
0.066873908,
0.0601539388,
-0.0234141331,
-0.0152023751,
-0.0049383794,
0.0158394445,
-0.0197834894,
-0.0025569911,
-0.0244952384,
-0.0085230805,
0.0578450933,
-0.046495568,
0.0630576611,
-0.0126904584,
0.0547253564,
0.0363624059,
0.0310731437,
0.0252600983,
0.0003744503,
-0.0157144144,
0.0948209614,
0.0115154795,
-0.0347890779,
0.0153239993,
0.0476002097,
-0.0281339604,
-0.0724152997,
-0.000058979,
-0.0090346364,
0.0049837921,
0.0180405304,
-0.0034511674,
0.025014095,
0.0046265321,
-0.0512530617,
0.0288103539,
0.0268984549,
-0.0045370911,
0.0365988016,
-0.0010631633,
0.0279728062,
0.0043154103,
-0.023743894,
-0.0195876211,
0.0423062593,
-0.0162677318,
0.0021485796,
0.0176578406,
0.0016571352,
-0.0011177139,
-0.0196529385,
-0.0718478933,
0.008725455,
-0.005647155,
-0.0138802622,
-0.0004893015,
0.0484289564,
-0.026443569,
-0.0372912921,
-0.1096062735,
0.0157820396,
-0.0220882334,
0.0267936867,
-0.017743459,
-0.0284668338,
-0.0044537974,
0.0132723721,
-0.0124183884,
-0.0170867089,
-0.0175957624,
0.0499674417,
-0.0131211281,
-0.0045653824,
-0.0374214165,
-0.0348572731,
-0.0259369891,
-0.0493944101,
0.0171882361,
-0.0151957097,
-0.0263419412,
0.0267177504,
-0.0055810553,
-0.0541605242,
0.0139505779,
-0.000534054,
-0.057795871,
0.0837028772,
0.0397116244,
0.0206437744,
-0.0135943629,
0.0028586807,
-0.0019035236,
0.0696043745,
0.0287741125,
0.0243707597,
-0.0052557783,
-0.0058308132,
0.0173385665,
0.0072586052,
-0.0510255918,
0.001393487,
-0.0709213838,
-0.0059804888,
0.0086260177,
0.025552189,
0.1105980575,
0.0452678464,
-0.001286842,
-0.0230486356,
0.0742088482,
0.0285088886,
-0.0233624801,
-0.0242693052,
0.0452281833,
-0.034596242,
-0.000009566,
0.0843744874,
-0.0601419024,
0.0263982266,
-0.0083039133,
0.0299665052,
0.0613025315,
0.0186364762,
-0.0091588758,
-0.0374074578,
-0.0224461369,
-0.0122801317,
0.0104671372,
0.035700608,
0.0574735515,
-0.0479513407,
-0.0212540813,
-0.0060206335,
0.0420843959,
0.0358502939,
0.0013664159,
-0.0624334626,
0.066532895,
0.0561442971,
0.0079573561,
0.0389083661,
0.0295486134,
0.0013187054,
-0.0340511203,
0.0292635262,
0.072709173,
0.0075906911,
-0.0177155659,
-0.0122525468,
-0.0145308776,
0.0232165344,
0.0498480648,
0.0217161514,
-0.0447892025,
0.0299083423,
0.022349054,
-0.0143319052,
0.0073803808,
0.0406618118,
-0.0430290438,
-0.0195821095,
-0.0387456417,
-0.0395064503,
0.0247799382,
-0.0417415574,
-0.0045353617,
0.050501246,
0.0318455547,
0.0146791358,
-0.0079318583,
0.0163618047,
-0.0483554713,
0.0037324347,
0.0340692028,
-0.0066431058,
-0.0131520201,
-0.0586112551,
0.00932121,
-0.0333563611,
0.0214040186,
-0.0162175726,
0.0454365537,
-0.0072801523,
0.0518047921,
0.0253242813,
0.0205307454,
-0.0104899574,
0.0661232919,
-0.0088580651,
-0.0027390604,
-0.0032483297,
-0.0015485088,
-0.0030579709,
-0.0535033494,
0.0083015915,
-0.0030820204,
0.0356673077,
-0.0249769725,
0.0281904396,
-0.0390804969,
-0.0394349135,
-0.0159678552,
-0.0339877605,
-0.0187269859,
0.001173363,
0.0137267336,
-0.0258054659,
-0.0246939864,
-0.0186938252,
-0.0043579922,
0.0080787642,
-0.016808698,
0.0032012598,
0.0578484721,
0.0208632015,
-0.0471731201,
0.0115155512,
-0.0457850434,
-0.010492756,
-0.0414203294,
0.0528641604,
0.0362729765,
0.0129504437,
-0.0886434987,
0.0357470699,
-0.0431183092,
0.0039674193,
-0.014942511,
-0.0260208976,
0.0071139857,
-0.0122479526,
0.0944110751,
0.0865445733,
0.0313138478,
0.0473840758,
0.1142269149,
-0.0111089926,
-0.0161486398,
-0.0267495234,
-0.0570601411,
-0.0145880254,
-0.0311595984,
-0.0627724901,
-0.0252109151,
0.0731516406,
0.0414015688,
0.0200722851,
-0.0264721811,
0.0496922582,
-0.0081257243,
-0.0600330122,
0.002246757,
-0.0311497264,
-0.0175729487,
0.0833962038,
0.0234286357,
-0.0550918132,
0.0762994736,
-0.0024143069,
0.0309647433,
-0.0109457299,
0.0013144519,
0.0127936592,
0.0562332831,
0.0427148901,
-0.0221292712,
0.0240250807,
0.0108577814,
-0.0553729609,
-0.0007648336,
-0.0018590048,
-0.0141851399,
0.0194088947,
-0.0895354748,
-0.0115033044,
0.0258283298,
-0.0127447713,
-0.0495854318,
-0.0460768975,
-0.089175038,
-0.0221400391,
0.002012894,
-0.0128606828,
-0.0283543579,
0.0584136732,
0.042034965,
-0.0339400694,
-0.0163377076,
-0.0375829302,
0.0172421597,
0.0215077363,
0.0876265541,
-0.0204403177,
0.0697497129,
-0.0456082299,
-0.0130148856,
0.0172914173,
0.0356315672,
0.0457441881,
-0.0231880974,
0.0025278633,
0.0124692312,
-0.005793287,
-0.0303960536,
0.0791352987,
0.0195147712,
-0.0208219234,
0.0182948057,
0.1052234322,
0.0101259761,
-0.0057435129,
-0.0104292603,
-0.0316975452,
0.0141180232,
-0.0015934809,
-0.0287662353,
-0.0477358699,
-0.0375320278,
0.0229280181,
-0.0375132822,
-0.0397274308,
-0.0585859045,
0.0032664274,
-0.0761355236,
0.0104074199,
-0.0054245442,
0.0377220884,
-0.0262267403,
-0.0493394956,
0.0121919587,
-0.0265547056,
-0.0066815196,
0.0225945208,
-0.0322275087,
0.0490970202,
0.0064699603,
-0.0204677712,
-0.0089403968,
0.0064037959,
-0.0019750309,
0.0406293385,
0.0483575985,
0.0066159521,
-0.0192396753,
-0.0576329269,
-0.0006309418,
-0.0453574993,
0.0012830293,
0.0361954682,
0.0024303978,
0.0041345828,
0.0167884324,
-0.0419594273,
0.0177808348,
-0.0047833445,
-0.0285272375,
-0.0100397784,
0.062984243,
-0.0624145307,
-0.1154394001,
-0.0098112468,
0.036857307,
-0.0065979897,
0.1251279563,
0.0016805931,
-0.0122917052,
-0.045495268,
-0.009426997,
0.0214199238,
0.0352739096,
-0.0001438373,
0.0164361019,
0.0124218008,
-0.0574476495,
-0.0449337624,
-0.049818635,
-0.0148191303,
-0.0270922054,
0.007357989,
0.0172188338,
0.0481155179,
-0.0217304118,
-0.0152979707,
0.0202653538,
-0.0156109855,
-0.0292890389,
-0.0270165894,
0.0120881684,
0.0209114198,
-0.0225893538,
-0.0390380062,
0.023050487,
-0.0747517869,
-0.0637360662,
0.0446625464,
0.0023317419,
0.0374411717,
-0.0232878495,
0.0153302029,
-0.0381564051,
0.0128067499,
-0.0250592269,
-0.0098261032,
0.0680435747,
-0.0003105752,
0.0705136955,
0.012885768,
0.0351494588,
-0.0035671678,
0.0246404726,
-0.0229111854,
0.011631473,
-0.0259373244,
-0.099432826,
0.0055559981,
-0.0228222087,
0.0951797068,
-0.032684464,
0.0248910449,
-0.0182350911,
-0.0320000164,
-0.0065249009,
-0.0043739546,
0.0263393745,
-0.0059223012,
0.0583452657,
0.0323270187,
-0.0324208289,
-0.0052475692,
-0.0071810656,
0.014821318,
0.0160103962,
-0.0230587125,
0.0515417233,
-0.0197274704,
0.0517750531,
-0.0318523869,
0.0005866797,
0.0302267671,
-0.0302373283,
0.0429647304,
0.0541962534,
0.0450874157,
0.0147364987,
-0.0211852826,
-0.0114100454,
-0.068112053,
-0.1076325923,
-0.0647045597,
0.0332517587,
0.0174599476,
0.007108754,
-0.0148793552,
0.0491399691,
-0.076784201,
0.0207941774,
-0.0107558537,
0.0024896371,
0.0141432313,
0.0142737152,
0.0476810709,
-0.0037051837,
-0.0483135283,
0.0117843309,
0.0889972746,
0.0136987325,
0.0404531658,
-0.0068071671,
-0.0525136217,
0.0063921758,
-0.0636257082,
-0.0221149139,
0.0235722046,
-0.0232639555,
-0.0053990069,
-0.0626116395,
-0.0390123464,
0.0014780828,
0.0248471852,
0.0243837051,
-0.057732489,
0.0046482952,
0.0617205165,
-0.0238455068,
0.006506531,
-0.0202355012,
0.0501758792,
0.0019846282,
0.0262408126,
-0.0876953155,
-6.262244116e-33,
0.0132853314,
-0.0258815382,
0.0057263616,
-0.0561039299,
0.0077211373,
-0.0457043462,
-0.0302933156,
0.0208065324,
0.0095172822,
0.0364057869,
-0.0085306857,
0.0243002549,
0.031428922,
0.0093553849,
0.0764768347,
-0.0251549333,
0.0425632223,
-0.0057128454,
0.0092474408,
-0.0380315073,
0.0336312391,
0.0029787363,
0.0343632028,
-0.0267019216,
0.0416610427,
0.0109796077,
0.0141155785,
0.0224062093,
0.0170528572,
0.0091981273,
-0.0574917868,
0.0225206502,
0.0232967623,
0.0018851254,
-0.0161108077,
0.0172116086,
0.0114513068,
-0.0347486511,
0.039666675,
-0.0021113325,
-0.0110084582,
0.0106447078,
0.0099648181,
0.0108910166,
-0.0470139273,
-0.0086991768,
0.0415011533,
0.0010432197,
-0.010508569,
0.0105734197,
-0.0166312326,
0.014396633,
0.0068724072,
0.0228030607,
0.012013305,
-0.0219150838,
0.0052879807,
-0.0122469226,
-0.0176833738,
0.0229159426,
-0.0026390329,
0.0242900122,
0.046911858,
0.040394783,
0.0121691655,
0.0635576844,
0.0838962048,
-0.0152772637,
-0.0559645593,
0.009973892,
-0.0081759803,
0.0105474824,
-0.0278718919,
0.0205947421,
0.0642548129,
0.002172498,
-0.0215813164,
-0.0040982245,
0.0275568794,
0.0327982232,
0.0403067991,
0.0195944123,
-0.0340085551,
-0.0455883518,
-0.0407418758,
-0.0651247203,
-0.0322830975,
-0.0070665302,
0.0029039059,
-0.0049023628,
0.0430397987,
-0.0196044073,
0.0189508703,
0.0193685014,
-0.0534191355,
-0.0110640442,
-0.0224332847,
-0.0188933033,
-0.0044418359,
0.0442552157,
-0.0166099519,
0.0113169821,
-0.0011578657,
0.023439141,
-0.0293857306,
0.0256912485,
-0.0295114201,
-0.0145882908,
-0.0572781712,
-0.0132078994,
-0.0350866653,
-0.0217864458,
0.0500759967,
-0.0382000357,
0.0733467564,
0.0139590129,
-0.0078815091,
-0.020397665,
0.0076664384,
-0.0006683936,
0.0103634913,
-0.0328509584,
-0.011413849,
0.0091070253,
-0.0190171115,
-0.0206627175,
-0.0321112759,
0.0099991942,
0.0282976199,
-0.0484733172,
-0.0323873311,
0.0602982007,
2.512e-7,
0.0530912541,
0.056993138,
-0.0055798572,
0.0167791545,
0.0202761684,
0.043597769,
-0.0096993418,
-0.002153083,
0.0613265671,
-0.0202962216,
-0.0099229449,
0.0234251544,
0.0191392824,
0.0054701134,
-0.0131368935,
-0.019661814,
-0.0973720178,
0.0035933,
-0.000327999,
0.0038161445,
0.0493801013,
0.0491947606,
0.0313125551,
-0.0077231559,
0.0363508016,
-0.0120405825,
0.0424449965,
-0.0158283599,
-0.0046725594,
0.0161972418,
-0.0110130105,
-0.0094386199,
-0.0096339872,
0.0480258204,
0.007452575,
-0.0222085826,
-0.0176680516,
0.0234638639,
0.0125484141,
0.0105993524,
-0.0249371026,
-0.0156733729,
0.0143982172,
-0.0270579308,
0.0768228397,
-0.0459932312,
-0.0633106157,
-0.0151514476,
-0.0335393548,
-0.0272295009,
0.0375606492,
0.006260165,
-0.0400401913,
-0.0039156242,
-0.0087466799,
0.0276319832,
-0.0164589416,
-0.0748013258,
-0.043534968,
0.0249265544,
-0.0552049764,
-0.0597161613,
-0.0512996204,
-0.0202523116,
0.059340395,
-0.0779846832,
0.0007816413,
2.120041255e-34,
0.0568507686,
0.0243603121,
-0.0285231359,
-0.0452700481,
0.0331887342,
-0.0276478082,
0.020034261,
0.0355883203,
0.0431544036,
-0.0996246412,
-0.0324074179
] |
http://arxiv.org/abs/2401.16640v3 | http://arxiv.org/pdf/2401.16640v3 | TeenyTinyLlama: open-source tiny language models trained in Brazilian Portuguese | 2024-01-30T00:25:54 | 2024-05-17T12:36:21 | Nicholas Kluge Corrêa, Sophia Falk, Shiza Fatimah, Aniket Sen, Nythamar de Oliveira | cs.CL | cs.CL, cs.LG | Large language models (LLMs) have significantly advanced natural language
processing, but their progress has yet to be equal across languages. While most
LLMs are trained in high-resource languages like English, multilingual models
generally underperform monolingual ones. Additionally, aspects of their
multilingual foundation sometimes restrict the byproducts they produce, like
computational demands and licensing regimes. In this study, we document the
development of open-foundation models tailored for use in low-resource
settings, their limitations, and their benefits. This is the TeenyTinyLlama
pair: two compact models for Brazilian Portuguese text generation. We release
them under the permissive Apache 2.0 license on GitHub and Hugging Face for
community use and further development. See
https://github.com/Nkluge-correa/TeenyTinyLlama | TeenyTinyLlama: open-source tiny language models trained in Brazilian Portuguese: Large language models (LLMs) have significantly advanced natural language
processing, but their progress has yet to be equal across languages. While most
LLMs are trained in high-resource languages like English, multilingual models
generally underperform monolingual ones. Additionally, aspects of their
multilingual foundation sometimes restrict the byproducts they produce, like
computational demands and licensing regimes. In this study, we document the
development of open-foundation models tailored for use in low-resource
settings, their limitations, and their benefits. This is the TeenyTinyLlama
pair: two compact models for Brazilian Portuguese text generation. We release
them under the permissive Apache 2.0 license on GitHub and Hugging Face for
community use and further development. See
https://github.com/Nkluge-correa/TeenyTinyLlama | [
0.0175120682,
0.0725168809,
-0.0065805018,
0.0477547161,
-0.0109552871,
-0.0112536699,
0.0456876121,
0.0071447082,
-0.0158677958,
-0.0306905266,
-0.0611960068,
-0.0980150104,
0.0076007089,
0.030432418,
0.0366002843,
-0.0427955799,
0.0217191391,
-0.0511628091,
-0.0571742207,
-0.0236765724,
-0.0043567214,
0.0178570468,
0.0117441965,
0.049955029,
0.0287645701,
-0.0203585736,
-0.0110712992,
0.0001590212,
-0.011979281,
0.0061981804,
0.0156620666,
0.0518110283,
-0.0047750408,
0.0663365275,
0.0000020189,
-0.0330369174,
-0.0089576552,
-0.0144334612,
0.0334065296,
0.0227624942,
0.087993145,
-0.0178271886,
-0.027556235,
0.0334289148,
-0.0466464683,
-0.0342061706,
0.0408242531,
0.0013923561,
0.0450454764,
0.0734399036,
-0.0018157746,
-0.0352301523,
-0.0129904393,
-0.0635645762,
-0.0029219456,
0.0123321312,
0.0133663015,
-0.0274387207,
0.0018794138,
-0.0027765385,
0.0223342646,
0.0115064261,
0.0016300565,
-0.0225391928,
-0.0338467136,
0.0385696739,
0.0038729638,
-0.0641358793,
-0.0097991712,
0.0420404822,
-0.0152132558,
-0.0279894639,
0.0069872397,
0.0109161502,
-0.0355338342,
-0.040512953,
-0.0779420882,
0.0870163664,
0.0223719552,
0.0187471677,
0.0633288771,
0.0085956641,
0.0302800927,
0.0403674245,
0.017451806,
0.0016116736,
0.0500730164,
-0.0272602458,
-0.0041499105,
0.0026604529,
-0.0199520346,
-0.0598279908,
0.0144204097,
0.042651318,
0.0318554193,
0.0086516468,
-0.0694773272,
-0.0144531345,
0.0641890839,
-0.0660143122,
0.02184291,
-0.0079380404,
0.0065599806,
0.0184955187,
-0.0503108166,
0.0132144811,
-0.0449913926,
-0.0127563952,
-0.0393538661,
-0.00302246,
0.0165391676,
-0.0130065531,
-0.0582113415,
0.0339069515,
-0.0864008665,
-0.0036532322,
-0.0257355925,
0.0181478988,
0.0082609858,
-0.0181306694,
0.0085811736,
0.0357322805,
-0.0190722309,
-0.0135230562,
0.01640434,
0.036229331,
-0.0260804743,
0.0170486756,
0.0229593124,
-0.0698861927,
-0.0017652763,
0.0391592719,
-0.0117201759,
-0.0250762198,
0.0143511109,
0.0255986825,
0.0080817789,
-0.0161195826,
-0.0537953563,
-0.0592731088,
-0.0287813675,
-0.1120865196,
-0.0242660679,
-0.0839640126,
0.0034550382,
0.0119883353,
0.0309595875,
-0.0535039976,
0.0170764178,
0.0240485314,
-0.0053845905,
0.0090320539,
-0.0396465696,
-0.0115624638,
0.0139216557,
-0.0040065441,
0.0429604016,
0.0729880184,
0.0457666107,
0.0138403736,
0.0193309654,
-0.006103558,
-0.0454862118,
0.0589288101,
0.0142966257,
0.015074213,
-0.0294439532,
-0.0159649458,
-0.0067402474,
0.0531774759,
0.0720745698,
0.0969497338,
-0.0104852915,
0.0139513072,
0.0691478252,
0.0222055763,
0.0039214296,
0.0659725368,
0.1089692935,
-0.0446490012,
0.0155396499,
-0.0037766886,
0.017315397,
-0.0029978852,
-0.0107725924,
0.0163170397,
0.0133395661,
0.085509941,
-0.0395882353,
0.0005426144,
0.0120733622,
-0.0267436095,
-0.0229627695,
-0.0552924983,
0.041211687,
0.0158047229,
-0.0352935977,
0.0435092263,
0.0337131508,
-0.0701432303,
0.0248502232,
-0.0477654673,
0.045952592,
0.0255506504,
0.0364256017,
0.0250415131,
0.0044684019,
-0.0160358194,
-0.0074311532,
0.0617818013,
0.052448757,
-0.0161013696,
-0.0084716119,
-0.0765978023,
0.0395075455,
-0.0124220001,
0.0083690137,
0.0405786671,
-0.0149049312,
0.0353934653,
0.0508221835,
0.0032300625,
-0.0172814466,
0.0327614062,
-0.0695306733,
-0.0083697308,
0.0162665285,
-0.0066420985,
0.0292624403,
0.0251760781,
0.0151880281,
0.0292315222,
0.0465945713,
-0.0367693342,
0.0186292026,
0.0093134819,
-0.0214015041,
-0.0336802788,
0.0250143968,
0.0068611423,
0.0430323519,
0.0368081406,
0.0083157457,
-0.0510219075,
0.0347661227,
-0.0166173503,
0.0159111973,
-0.0513908453,
0.0273280945,
0.0152668441,
0.0343843736,
-0.0084770462,
0.0410231315,
-0.0042475141,
0.02444591,
0.0472655371,
-0.0006834837,
0.0159265902,
-0.0681374967,
-0.0179414731,
0.0014018689,
-0.0047225794,
0.0043630619,
0.0050750915,
0.05963061,
-0.0080479383,
-0.0512955934,
-0.0851333812,
0.0255681332,
-0.032300733,
0.0146597689,
0.0265190154,
0.0442684181,
-0.0062508127,
0.0024107096,
-0.0240006372,
-0.0019767957,
0.0433002934,
-0.0001881544,
0.0188204031,
0.0112436563,
-0.0205772836,
-0.0423495956,
-0.0094649177,
0.0416036919,
0.0476369597,
0.0282329898,
-0.0477137528,
-0.0667533726,
0.0464909226,
-0.0613708086,
0.0170479976,
0.0170854423,
-0.053612642,
-0.0068503795,
0.0108059123,
0.0765688866,
0.043462012,
0.0163239967,
0.0146430433,
0.04283797,
0.0260079484,
-0.0168095529,
-0.0145683382,
-0.0179694984,
-0.0203152094,
-0.0488161556,
-0.0352564268,
0.0296661016,
0.0944957733,
0.0029393542,
-0.0048714997,
-0.0356441922,
0.0281135645,
-0.0083819684,
-0.0408093818,
0.023437323,
0.0234143324,
-0.0150691262,
0.0102232238,
-0.0010427544,
-0.0361668281,
0.0265734531,
-0.0640713498,
-0.0063289613,
-0.0310907625,
-0.0133076636,
-0.0588004068,
0.0355066098,
0.0094128177,
0.0308627188,
0.0202854052,
0.0038756395,
-0.0426054746,
0.0564804971,
-0.0028930034,
-0.0151993912,
0.0309758633,
-0.0499532819,
0.0183605552,
-0.0310215298,
-0.0601938479,
-0.011230777,
-0.0617940724,
-0.029183656,
0.0107091404,
0.0464470722,
0.0181396399,
-0.0319005139,
0.0319715515,
0.0268486757,
0.0549806617,
-0.0375729054,
0.0054060477,
-0.0027684805,
0.0557271615,
0.054444015,
-0.0221202224,
0.0200161487,
-0.0571487546,
0.0163216759,
0.0332838073,
0.0355332568,
0.0199644528,
-0.0330090374,
0.0171947517,
0.032648582,
-0.0045390883,
-0.0613611974,
-0.0151456119,
0.0381591544,
-0.037212152,
-0.0354245752,
0.0887205601,
0.0053851893,
-0.0183618665,
0.0091266027,
-0.0416984372,
0.0934517458,
0.012151801,
0.0140844686,
-0.0689541623,
-0.0170836281,
-0.0049086111,
0.0369699821,
-0.0249471366,
-0.0081322733,
-0.0505754016,
-0.1135961786,
0.0535710976,
0.041666612,
-0.0391126722,
0.0078323698,
-0.0171512384,
-0.0389999487,
-0.0403720215,
-0.0025988566,
0.0384396501,
-0.0343034305,
0.0131977545,
0.0532698333,
-0.0008045357,
-0.0653972775,
-0.0172544606,
-0.018376993,
0.0474601351,
0.0245709382,
-0.0051974468,
0.0105314627,
-0.0054804143,
0.0441114604,
-0.0162647944,
0.0112970788,
-0.0469419584,
0.003527442,
0.0114494637,
-0.0204703882,
-0.0198269691,
0.0443646982,
-0.0077102706,
0.010842694,
0.0110384682,
0.0568202846,
0.0033075174,
-0.0110373162,
-0.0100985952,
-0.0066951294,
-0.0029283371,
0.0396531373,
0.0310001206,
-0.0374181829,
-0.0458653793,
-0.0063446416,
-0.0121362004,
-0.0318261944,
-0.03057619,
0.0208609495,
-0.0009032578,
-0.0691259652,
-0.0663904995,
-0.0565089248,
0.0146216964,
0.0401060469,
-0.0074938214,
-0.006963091,
-0.0235908329,
-0.0216810107,
-0.034574233,
0.0190042388,
-0.0196365099,
0.0037054652,
-0.0952061489,
0.0034842829,
0.0299968477,
-0.0118157584,
-0.0070643919,
0.0033066231,
-0.1313388199,
-0.0646109805,
-0.0196263902,
0.0597864576,
0.0330562703,
0.0151778655,
-0.0001403032,
-0.0779220983,
-0.0015953957,
0.0290846042,
-0.0434346199,
0.0408134647,
0.0024889489,
-0.0142809683,
-0.0056273243,
0.0648647398,
-0.0043466557,
0.0157128852,
-0.0259021278,
0.0761439577,
-0.0063550808,
-0.0340605341,
-0.0020586671,
0.0204106327,
0.0047465079,
-0.0614517927,
0.009712033,
-0.0500939414,
0.0072577666,
-0.0214854553,
0.0181646943,
-0.0594033897,
0.0133271366,
0.0404746644,
0.0030163846,
-0.0131453536,
-0.0222894475,
-0.0234224629,
0.0457298271,
-0.0212306622,
0.0064047645,
0.0019230996,
0.0094896164,
0.0269635934,
-0.0014462196,
-0.0444162562,
0.0001677828,
-0.0552611612,
0.0187723171,
-0.0419637002,
0.0415631011,
-0.0200652108,
-0.0056197513,
0.0081566181,
0.0295447949,
0.016296858,
-0.0034803206,
-0.0563867539,
0.0213703085,
0.0651489571,
0.0252090152,
0.0896282643,
-0.0322259329,
-0.0027263884,
-0.0462284274,
0.0360060446,
0.0141137214,
-0.067984499,
0.0374501646,
-0.0109479735,
-0.0266726408,
0.0124641564,
0.1156349778,
0.034757521,
0.0278776865,
0.0180344153,
0.0290703848,
-0.0269659627,
-0.0110445116,
0.0134467306,
-0.0143468333,
-0.0387625061,
-0.0123472232,
-0.0690387413,
-0.0403385349,
0.0112293549,
0.0082544405,
-0.0085053528,
-0.0220599379,
0.027806703,
0.0446478948,
-0.0219877698,
0.0160534848,
-0.0506362617,
0.0233685598,
-0.0429102182,
-0.0108152563,
-0.0285108872,
-6.20343046e-33,
-0.0402006842,
-0.0103152739,
0.035641551,
-0.0330909118,
-0.0377092212,
0.007109182,
-0.0445406437,
0.0131743634,
0.0136853782,
-0.0143185332,
0.0070024705,
-0.0346503407,
0.0150194764,
0.030579295,
0.0118397875,
0.028907571,
0.0484977439,
-0.013800119,
-0.0165451728,
0.0325697027,
0.0181419738,
0.0272488073,
0.0478719175,
-0.0100374827,
-0.002531952,
-0.0360025726,
-0.040633034,
0.0247316677,
-0.0134816468,
0.0652875975,
-0.0242249202,
0.0156582948,
0.0190526173,
-0.0021734349,
-0.0237817802,
0.0083094388,
-0.0403595902,
0.0016775841,
0.0456291959,
0.0083910841,
-0.0004800518,
-0.0426225252,
0.0279482249,
-0.0065710237,
0.0030229064,
0.0019692488,
0.004344563,
-0.0663208887,
-0.0461884178,
-0.001428768,
-0.0457918793,
0.0380385816,
0.0091858525,
-0.0128856394,
0.0071478281,
0.0618839674,
-0.0460386723,
-0.0183996949,
0.0193110555,
0.0250762831,
0.0123905996,
0.0709695369,
0.0646531284,
0.0164074283,
-0.0029511231,
0.039421726,
0.0344135612,
-0.0232804269,
-0.0161046367,
0.0201219916,
0.038406752,
0.0600497127,
0.03566093,
0.0783795714,
-0.0119180679,
0.0117896991,
-0.0895679668,
0.0014133147,
0.0128176929,
0.007318269,
0.0160007048,
-0.042052988,
0.006025068,
-0.0057476168,
-0.030142609,
-0.0452281125,
-0.0086470833,
-0.0177455768,
0.022617856,
0.0090554524,
-0.0289233606,
-0.0081254067,
0.0309330299,
-0.0092194462,
-0.0306580178,
0.0131880511,
-0.024468055,
0.0358022377,
-0.0163629763,
0.0375481471,
-0.0713969544,
-0.0463667773,
-0.0031739983,
0.0339034088,
-0.0050360491,
0.0120932097,
0.0191355478,
0.0073370375,
-0.0592028908,
0.0011756351,
-0.0492540151,
0.0553016141,
-0.008860698,
0.0242311638,
0.0224178433,
-0.0321783386,
-0.0042311042,
0.0409497432,
-0.0032985597,
-0.0197063442,
-0.0175504573,
-0.0445271581,
0.0244375467,
0.0540164486,
-0.0451896302,
0.0231876858,
-0.0274564475,
0.0483165719,
0.0279415399,
-0.0130359419,
-0.010801794,
-0.0105033321,
2.631e-7,
0.0016839019,
0.0050906911,
0.0197094325,
0.0002546451,
0.024408577,
0.0195461046,
-0.0186144691,
0.0325805917,
-0.0195182282,
0.021314282,
-0.0025788553,
-0.0133480923,
0.0799060687,
0.0057487497,
-0.0623564124,
0.0232317988,
-0.0667649508,
0.0096556153,
-0.052373264,
-0.0505784787,
-0.0233563464,
0.08239232,
-0.0043689846,
-0.0307412297,
-0.0201199576,
-0.0314534046,
0.101596795,
0.0049308473,
-0.0375468433,
-0.002154598,
0.0395588651,
0.017005248,
-0.0138393166,
-0.0448311269,
-0.0036483349,
0.0026042715,
-0.0182474405,
0.086799182,
0.0013133814,
0.0437662154,
0.0852283314,
-0.0423738733,
-0.0441962294,
-0.0752320215,
0.0386191979,
-0.0139286825,
-0.0163821559,
-0.0372886136,
0.0228642989,
-0.0045118784,
-0.0132541005,
-0.0677114874,
-0.0570129231,
0.0171921011,
0.0228651874,
-0.0476042368,
-0.0206190441,
-0.0097209252,
-0.0264078006,
-0.0041069915,
-0.0459589772,
-0.0344128199,
-0.0428940058,
0.0441212133,
0.0362118296,
-0.0491904728,
0.0072200256,
2.455206837e-34,
0.0425148346,
-0.0304986965,
-0.0192126706,
0.0163642578,
0.0070614,
-0.0256230496,
0.0321715176,
0.0118512101,
-0.0109877726,
-0.0490675345,
-0.0342670195
] |
http://arxiv.org/abs/2401.16638v1 | http://arxiv.org/pdf/2401.16638v1 | Breaking Free Transformer Models: Task-specific Context Attribution Promises Improved Generalizability Without Fine-tuning Pre-trained LLMs | 2024-01-30T00:23:29 | 2024-01-30T00:23:29 | Stepan Tytarenko, Mohammad Ruhul Amin | cs.CL | cs.CL, cs.AI, I.2.7; I.2.4 | Fine-tuning large pre-trained language models (LLMs) on particular datasets
is a commonly employed strategy in Natural Language Processing (NLP)
classification tasks. However, this approach usually results in a loss of
models generalizability. In this paper, we present a framework that allows for
maintaining generalizability, and enhances the performance on the downstream
task by utilizing task-specific context attribution. We show that a linear
transformation of the text representation from any transformer model using the
task-specific concept operator results in a projection onto the latent concept
space, referred to as context attribution in this paper. The specific concept
operator is optimized during the supervised learning stage via novel loss
functions. The proposed framework demonstrates that context attribution of the
text representation for each task objective can improve the capacity of the
discriminator function and thus achieve better performance for the
classification task. Experimental results on three datasets, namely HateXplain,
IMDB reviews, and Social Media Attributions, illustrate that the proposed model
attains superior accuracy and generalizability. Specifically, for the
non-fine-tuned BERT on the HateXplain dataset, we observe 8% improvement in
accuracy and 10% improvement in F1-score. Whereas for the IMDB dataset,
fine-tuned state-of-the-art XLNet is outperformed by 1% for both accuracy and
F1-score. Furthermore, in an out-of-domain cross-dataset test, DistilBERT
fine-tuned on the IMDB dataset in conjunction with the proposed model improves
the F1-score on the HateXplain dataset by 7%. For the Social Media Attributions
dataset of YouTube comments, we observe 5.2% increase in F1-metric. The
proposed framework is implemented with PyTorch and provided open-source on
GitHub. | Breaking Free Transformer Models: Task-specific Context Attribution Promises Improved Generalizability Without Fine-tuning Pre-trained LLMs: Fine-tuning large pre-trained language models (LLMs) on particular datasets
is a commonly employed strategy in Natural Language Processing (NLP)
classification tasks. However, this approach usually results in a loss of
models generalizability. In this paper, we present a framework that allows for
maintaining generalizability, and enhances the performance on the downstream
task by utilizing task-specific context attribution. We show that a linear
transformation of the text representation from any transformer model using the
task-specific concept operator results in a projection onto the latent concept
space, referred to as context attribution in this paper. The specific concept
operator is optimized during the supervised learning stage via novel loss
functions. The proposed framework demonstrates that context attribution of the
text representation for each task objective can improve the capacity of the
discriminator function and thus achieve better performance for the
classification task. Experimental results on three datasets, namely HateXplain,
IMDB reviews, and Social Media Attributions, illustrate that the proposed model
attains superior accuracy and generalizability. Specifically, for the
non-fine-tuned BERT on the HateXplain dataset, we observe 8% improvement in
accuracy and 10% improvement in F1-score. Whereas for the IMDB dataset,
fine-tuned state-of-the-art XLNet is outperformed by 1% for both accuracy and
F1-score. Furthermore, in an out-of-domain cross-dataset test, DistilBERT
fine-tuned on the IMDB dataset in conjunction with the proposed model improves
the F1-score on the HateXplain dataset by 7%. For the Social Media Attributions
dataset of YouTube comments, we observe 5.2% increase in F1-metric. The
proposed framework is implemented with PyTorch and provided open-source on
GitHub. | [
0.0397121795,
0.0499980561,
-0.0022051858,
0.0501924194,
-0.0004333053,
0.0093413517,
0.0257819947,
0.0124557475,
-0.0377861224,
-0.087608628,
-0.0862519816,
-0.0109851612,
-0.0297742989,
0.0243707523,
0.0048456597,
0.0078909695,
-0.0122503089,
0.0398765914,
0.0260171499,
0.0070000342,
0.0113962442,
0.0212628767,
0.0344115756,
0.0266132075,
0.0000587271,
0.0031270112,
0.0736247227,
0.0081522064,
0.0402282961,
0.032944072,
0.0536707528,
0.0618020259,
0.0451834984,
0.0156605747,
0.0000020579,
-0.0254515987,
-0.0079026828,
-0.0242908839,
-0.0084197028,
-0.0059526036,
0.056910295,
-0.0240965895,
0.0227473155,
0.0058149197,
-0.0318458863,
-0.0214945152,
0.0426075757,
0.0924251527,
0.0389423706,
0.0768503621,
-0.02850301,
-0.052464284,
0.0221246853,
-0.0368802249,
0.0144193321,
-0.031718269,
0.026304869,
-0.0204409026,
-0.0324047431,
0.0336204432,
-0.0216053482,
0.0279393308,
-0.0010680071,
0.011459576,
0.0559513234,
-0.0080805188,
-0.0332991555,
-0.0039827493,
0.0106271077,
0.0809996501,
-0.0212846473,
-0.0616385713,
-0.0122852419,
-0.015874669,
-0.0104481997,
0.0127879707,
-0.0446670465,
0.0216103233,
0.0113891065,
-0.009624958,
0.0127594974,
0.0212162286,
0.0366292149,
0.038459789,
-0.0115294922,
-0.0057406477,
0.0065600364,
-0.0191595107,
0.0144206164,
0.0281777661,
-0.0737219155,
-0.0051334435,
0.0245597642,
0.0022322985,
0.0270299129,
0.0079672467,
-0.033129096,
-0.0458208136,
-0.0115246354,
-0.0612075888,
0.0487443507,
0.0121566672,
-0.0387863144,
0.0380345061,
-0.025280321,
0.0577164255,
-0.0329941064,
0.0378044248,
-0.0183665715,
0.0043171458,
-0.0700627342,
-0.0106120845,
-0.0474485084,
0.0426833071,
-0.0036613808,
-0.0222588796,
-0.0423582047,
0.0414075181,
-0.0297255442,
-0.0415300988,
0.0459911004,
0.0245958716,
-0.0252072383,
0.0201602597,
-0.0015121781,
0.0420074537,
0.001944267,
0.060439501,
-0.0394545496,
-0.0142644616,
-0.0120921917,
0.041434288,
0.0216858741,
-0.0351783,
0.0433204025,
0.0354173146,
0.0264385175,
-0.0943254679,
-0.0409952737,
0.011986454,
-0.0478086956,
-0.0597814284,
-0.0176381413,
-0.0450638123,
0.0422037169,
0.0299222339,
-0.0181719605,
-0.0287431367,
-0.0108028641,
-0.0019165786,
-0.0303888395,
0.0890453681,
0.0546378344,
-0.0068128053,
0.0303206611,
0.0154588232,
-0.0141500635,
0.0584583618,
0.0090740314,
0.0575980395,
0.0478838533,
-0.0041957637,
0.0132818203,
0.0506737754,
-0.0018835388,
-0.0040453812,
-0.0249552969,
0.0162849613,
-0.0304900762,
-0.0007622433,
0.0244892668,
0.032081183,
-0.0156622715,
0.0311899129,
0.0477839857,
0.0105845835,
0.0352446511,
0.0483922213,
0.041585654,
0.0231889524,
-0.0033501147,
-0.0341071971,
-0.0375652798,
0.0138036907,
0.0133921588,
-0.0413237028,
-0.0172547195,
-0.0002549112,
-0.0599516965,
-0.0076865316,
0.025028212,
-0.0358746462,
0.0359406024,
-0.0035222012,
0.0684807897,
0.0279278811,
-0.0474857241,
0.0970599279,
0.0218983088,
-0.0413610935,
0.0321004279,
-0.0668320656,
0.0064560589,
0.0384455062,
0.0228659045,
-0.0355235711,
-0.006910637,
-0.0413132943,
0.0096653001,
0.0004507762,
0.0113436272,
-0.021641992,
-0.0175734777,
-0.0399944894,
-0.0477745906,
-0.0186573435,
0.0231331643,
0.024654923,
0.00297155,
0.024445964,
0.0419512615,
-0.0506627522,
-0.0135421604,
0.0152047463,
0.0235501416,
-0.0496779829,
0.0090411315,
0.0119005283,
0.032972835,
0.0431738608,
-0.0184298176,
-0.0228057429,
0.0253039319,
-0.032365799,
-0.0317726806,
-0.0121958852,
-0.0412305072,
-0.0245057922,
-0.0009649163,
-0.0315602869,
0.0414686054,
0.1008837745,
-0.048920881,
-0.0359269567,
0.015736619,
-0.014328626,
0.0516587272,
-0.0202999227,
0.0069722501,
-0.0439758748,
0.007857901,
-0.042148333,
0.0591261089,
0.0065466114,
0.0188236833,
-0.0204187743,
-0.0081408452,
0.0150432959,
-0.0009743863,
-0.0502222143,
-0.0234851036,
-0.0166957155,
0.0523889512,
0.050588984,
-0.0050649201,
0.0133287814,
-0.0411057808,
-0.0810841396,
0.0002458163,
-0.0491303727,
0.0041546421,
0.0451174639,
-0.0020926706,
0.0167939626,
0.0110270353,
-0.0341367163,
-0.0199616998,
0.0026617204,
0.0006767219,
-0.0014702702,
0.0134107955,
-0.0034757676,
0.0008367256,
-0.02051761,
-0.0034009677,
0.0367696956,
-0.009058835,
-0.0276178289,
-0.0311049316,
0.0314435549,
0.0102754431,
0.0105524501,
-0.0271555372,
-0.0118385842,
-0.0332844034,
0.0032015874,
-0.0033296493,
0.0179372206,
0.0159369521,
0.045682583,
0.002164681,
0.0254978277,
-0.03874119,
-0.0080877244,
-0.036836829,
-0.0411832631,
-0.0419406705,
-0.0117600067,
0.0233964212,
0.0670260414,
0.0176910758,
-0.0163619705,
-0.0768105388,
0.0110883815,
-0.0030056536,
0.0006662905,
0.0168577731,
-0.020327691,
0.0007791297,
-0.0198117662,
-0.0263018031,
-0.0229329448,
0.0283076186,
-0.0216917545,
0.0213483479,
-0.0321598873,
-0.0129151801,
-0.045084212,
0.0472153798,
0.0006479249,
0.0213042591,
0.0299030226,
-0.0327596702,
-0.0406578742,
0.0177755542,
0.0182605814,
-0.0320079811,
0.0742264986,
0.013363204,
0.0146520548,
0.0004869752,
-0.0218483638,
-0.0268530119,
-0.0526180454,
-0.007142934,
-0.0075853467,
0.041184254,
-0.0172856711,
-0.0425812341,
0.0173663031,
-0.0136478469,
0.0266756453,
-0.0287679005,
0.0018477269,
0.0008749188,
0.0436533913,
0.0285488945,
-0.0221760683,
-0.0196074173,
0.017388707,
-0.0342804231,
0.0299897119,
-0.006214221,
-0.0008849687,
-0.0371655002,
0.0057312087,
0.0324427709,
-0.0214533377,
-0.0145633956,
-0.0219756979,
0.066037558,
-0.0141679356,
-0.0080629578,
0.118555814,
0.0141669558,
-0.0619865544,
-0.0296799932,
0.0423961543,
-0.0283167288,
0.0306283571,
-0.0028226355,
-0.091828987,
-0.0435884073,
-0.0320353135,
0.0099002132,
0.0266818795,
0.0003469854,
0.0087115597,
-0.078261748,
0.0137389069,
-0.0091750808,
-0.0123933125,
-0.0368684568,
-0.0405415967,
-0.01048645,
-0.0599109009,
-0.0106921103,
-0.0442826264,
-0.0107050454,
-0.001656994,
0.0197878703,
0.0203035176,
-0.0762032643,
0.0640671551,
-0.0322163329,
0.0190961882,
0.0422240384,
0.0464263409,
-0.0243252218,
-0.0254830811,
-0.0039562397,
-0.0546449088,
-0.0084933201,
-0.0104062287,
0.0503069647,
-0.0395959429,
0.001716046,
0.014199107,
0.0009409803,
-0.0445841551,
-0.0098988302,
-0.0404812992,
0.0670822412,
-0.0071638427,
-0.010350395,
-0.0272606369,
-0.0247657169,
0.0052708313,
0.0857558921,
-0.0058259373,
-0.0592678934,
-0.0066620777,
0.0158513673,
0.0041711247,
-0.0244534854,
-0.0236873906,
-0.0104440674,
-0.0326202996,
-0.0481935292,
-0.043868497,
-0.0022678664,
0.0219474975,
0.0108582191,
0.0209644586,
-0.0205410626,
-0.03055319,
-0.0149613954,
-0.0588543564,
0.0483520366,
-0.0186868589,
0.0213715862,
-0.0297054667,
-0.0198442172,
0.0771794096,
-0.0248004403,
-0.0306432378,
0.0098947445,
-0.0924292952,
-0.033970546,
0.0001819683,
0.0320913531,
0.0191236641,
0.0358776413,
-0.0041811285,
-0.0517367385,
0.0918751135,
0.0422701128,
-0.0511918031,
-0.0065958672,
0.047998026,
-0.0095846541,
-0.027834218,
0.0401353016,
-0.0682863742,
0.0529812053,
0.0470722504,
0.01669956,
-0.037438944,
-0.0461333729,
-0.0048982026,
-0.0174789354,
0.0777707547,
-0.0157839544,
0.0001670323,
-0.0505794771,
-0.0058442117,
0.0175648723,
-0.0048081097,
-0.0575519688,
-0.0113940937,
0.0708984137,
0.0009812049,
-0.0509530939,
0.0264896229,
-0.018776387,
0.0391846448,
-0.0668268949,
0.0469145738,
0.0110708466,
-0.017324999,
0.0228192825,
0.0485523157,
-0.0194978546,
-0.0216291901,
0.057367973,
-0.0104303975,
-0.0126765529,
0.0521916896,
0.0149394348,
0.027569402,
-0.021539513,
0.0415558666,
-0.0488237031,
0.0052975044,
-0.0565498099,
-0.0376482345,
0.0332983173,
-0.0553373508,
0.0003625444,
-0.0107361916,
-0.0249510538,
-0.0109964171,
0.0419238359,
-0.002208218,
0.0067932853,
0.0564955026,
0.0065036477,
-0.0047193537,
0.0096333558,
0.0659545064,
0.0048900205,
0.0261752717,
0.0115383873,
-0.0318141654,
-0.0535601862,
-0.0067945421,
0.0187150743,
-0.0041069975,
-0.0046319203,
0.0261754151,
-0.0101768365,
-0.0661042556,
0.0550453924,
-0.0786205083,
0.0041074865,
-0.0112860976,
0.0403404757,
0.0765329897,
0.0085157501,
0.0037178302,
-0.0089716911,
-0.0475302637,
0.0303872209,
-0.0152443117,
-0.0186393578,
-6.311049907e-33,
0.006009527,
0.0179958735,
-0.026881557,
0.030054586,
-0.0216483418,
0.0104873152,
0.0090615107,
-0.0106574045,
0.0084929559,
-0.0054643312,
0.002240327,
0.0111886403,
0.0085887313,
0.0337713696,
-0.0236484502,
-0.0496180058,
0.0662043616,
0.0047573401,
-0.0179245509,
0.0410087593,
0.0277947281,
0.0420505814,
0.0773204416,
-0.0720949769,
0.0334031731,
-0.0291617159,
-0.0165764261,
0.0069950856,
0.061745014,
0.0546951182,
-0.0081894882,
0.0832857937,
0.0288173165,
-0.0771191567,
0.0083091846,
0.0636100844,
-0.1077107266,
-0.0023176018,
0.0066776704,
0.0196446907,
0.0314678252,
-0.0166094005,
0.0601507016,
0.0106998151,
-0.0217548348,
-0.0204695836,
0.0607645847,
-0.0161108859,
0.0284171775,
-0.0709403455,
-0.0482355468,
0.0091873072,
0.0011825273,
-0.0273076408,
0.0333254896,
0.014350039,
-0.0002627533,
0.0587239824,
-0.0841290951,
-0.0049439971,
-0.0087647643,
0.0061287927,
0.0330488235,
0.0292227045,
-0.0201108828,
0.0169354435,
-0.0255576167,
0.0054730014,
-0.03920497,
-0.0478692353,
0.0222394187,
0.0449752659,
0.022859294,
0.0369573161,
0.0251904335,
-0.0432242267,
-0.0802255347,
0.0136985462,
-0.0097414972,
0.0143660512,
0.045956336,
0.0172294918,
-0.0296514593,
-0.0222081076,
-0.0275378097,
-0.0387526974,
-0.04067038,
-0.0301825143,
0.0087151993,
0.0289179683,
-0.0116607817,
0.0162374862,
0.0236384179,
-0.0807362348,
-0.0071022445,
0.030655859,
-0.0421098359,
0.0229618996,
0.0095506161,
-0.0063944566,
-0.0794528723,
-0.0169257224,
0.0265463069,
0.0084177013,
0.0122814057,
0.0056104022,
0.0743051767,
0.0647627637,
-0.07112737,
-0.0059706289,
-0.0196773857,
-0.0085007362,
-0.0195020419,
0.0760824904,
-0.0073112492,
0.0435772501,
-0.0365726985,
0.0520793609,
0.0299041662,
0.0158209577,
-0.0015610341,
0.019146204,
0.0118082399,
0.0200575721,
-0.040625941,
-0.036186032,
0.0094090551,
0.0054535903,
0.0304290317,
-0.0901180133,
-0.0237483811,
-0.0392815135,
2.744e-7,
0.0565289631,
0.0233218186,
-0.0203701109,
-0.0176093131,
0.0223181285,
0.0492594317,
0.0296065621,
0.047149878,
-0.005648932,
-0.0047738948,
0.0170166157,
0.0190212522,
-0.0116810752,
-0.0263594128,
-0.0730748326,
0.0272605605,
-0.0628861859,
0.0299044382,
-0.0193081498,
0.0205602013,
0.0284097027,
0.1201862544,
0.0438290983,
-0.0376997851,
-0.0328858197,
-0.0484282896,
0.0839877501,
-0.0547896586,
-0.008922074,
0.0320863575,
0.0351186767,
0.0402719453,
-0.0274995659,
0.0350560993,
-0.0433770493,
0.0344799459,
0.0386832096,
0.0665602833,
-0.0082988888,
-0.0491301082,
-0.0139675913,
-0.0107119931,
0.0058314013,
-0.0332752317,
0.062643446,
-0.0793506354,
-0.0367092788,
-0.0118519124,
-0.0220821761,
0.0014270536,
0.001557839,
-0.0114152031,
0.0223679449,
0.0055346838,
0.0184707232,
0.0125980545,
-0.0279310066,
-0.1446655989,
0.007591764,
0.0316931605,
-0.0344548821,
-0.0032082731,
-0.0217219219,
-0.0436481573,
0.0515110008,
-0.0313801579,
0.0548895001,
2.554069585e-34,
0.0194246527,
0.0175997261,
-0.0067429887,
-0.0252045859,
-0.010624513,
0.0098013058,
-0.0134470994,
0.0010288138,
0.016534498,
-0.0575830676,
0.0029226993
] |
http://arxiv.org/abs/2402.01748v2 | http://arxiv.org/pdf/2402.01748v2 | Large Multi-Modal Models (LMMs) as Universal Foundation Models for AI-Native Wireless Systems | 2024-01-30T00:21:41 | 2024-02-07T17:55:11 | Shengzhe Xu, Christo Kurisummoottil Thomas, Omar Hashash, Nikhil Muralidhar, Walid Saad, Naren Ramakrishnan | cs.NI | cs.NI, cs.AI, cs.CL, cs.LG | Large language models (LLMs) and foundation models have been recently touted
as a game-changer for 6G systems. However, recent efforts on LLMs for wireless
networks are limited to a direct application of existing language models that
were designed for natural language processing (NLP) applications. To address
this challenge and create wireless-centric foundation models, this paper
presents a comprehensive vision on how to design universal foundation models
that are tailored towards the deployment of artificial intelligence (AI)-native
networks. Diverging from NLP-based foundation models, the proposed framework
promotes the design of large multi-modal models (LMMs) fostered by three key
capabilities: 1) processing of multi-modal sensing data, 2) grounding of
physical symbol representations in real-world wireless systems using causal
reasoning and retrieval-augmented generation (RAG), and 3) enabling
instructibility from the wireless environment feedback to facilitate dynamic
network adaptation thanks to logical and mathematical reasoning facilitated by
neuro-symbolic AI. In essence, these properties enable the proposed LMM
framework to build universal capabilities that cater to various cross-layer
networking tasks and alignment of intents across different domains. Preliminary
results from experimental evaluation demonstrate the efficacy of grounding
using RAG in LMMs, and showcase the alignment of LMMs with wireless system
designs. Furthermore, the enhanced rationale exhibited in the responses to
mathematical questions by LMMs, compared to vanilla LLMs, demonstrates the
logical and mathematical reasoning capabilities inherent in LMMs. Building on
those results, we present a sequel of open questions and challenges for LMMs.
We then conclude with a set of recommendations that ignite the path towards
LMM-empowered AI-native systems. | Large Multi-Modal Models (LMMs) as Universal Foundation Models for AI-Native Wireless Systems: Large language models (LLMs) and foundation models have been recently touted
as a game-changer for 6G systems. However, recent efforts on LLMs for wireless
networks are limited to a direct application of existing language models that
were designed for natural language processing (NLP) applications. To address
this challenge and create wireless-centric foundation models, this paper
presents a comprehensive vision on how to design universal foundation models
that are tailored towards the deployment of artificial intelligence (AI)-native
networks. Diverging from NLP-based foundation models, the proposed framework
promotes the design of large multi-modal models (LMMs) fostered by three key
capabilities: 1) processing of multi-modal sensing data, 2) grounding of
physical symbol representations in real-world wireless systems using causal
reasoning and retrieval-augmented generation (RAG), and 3) enabling
instructibility from the wireless environment feedback to facilitate dynamic
network adaptation thanks to logical and mathematical reasoning facilitated by
neuro-symbolic AI. In essence, these properties enable the proposed LMM
framework to build universal capabilities that cater to various cross-layer
networking tasks and alignment of intents across different domains. Preliminary
results from experimental evaluation demonstrate the efficacy of grounding
using RAG in LMMs, and showcase the alignment of LMMs with wireless system
designs. Furthermore, the enhanced rationale exhibited in the responses to
mathematical questions by LMMs, compared to vanilla LLMs, demonstrates the
logical and mathematical reasoning capabilities inherent in LMMs. Building on
those results, we present a sequel of open questions and challenges for LMMs.
We then conclude with a set of recommendations that ignite the path towards
LMM-empowered AI-native systems. | [
0.0144933732,
0.0307072327,
-0.0176384542,
-0.0097419312,
-0.0504443422,
-0.0438354947,
0.0304749496,
-0.0710272491,
-0.0156155024,
-0.0573417768,
0.0294570569,
0.025750827,
-0.0383653976,
0.0181557052,
0.0113217961,
0.0447319485,
0.0441278331,
-0.0402841009,
-0.0192038417,
0.0086690886,
-0.0178586338,
0.0166460406,
0.006318423,
0.0051426035,
-0.0605248548,
0.0140711991,
-0.0215463005,
0.0094715087,
-0.0219609048,
-0.0647570416,
-0.0332657062,
-0.0093678776,
0.0811286047,
0.011921783,
0.00000224,
-0.0352781415,
-0.0370575115,
0.0105857002,
0.0003937207,
-0.0356903821,
0.0069371909,
-0.0020084949,
0.0100319553,
0.032590095,
-0.0259996597,
0.0159767941,
0.0528775528,
0.0582977757,
0.035068851,
0.0616503432,
-0.0140304659,
-0.0542140491,
0.0399331301,
-0.0127811851,
0.0244852547,
-0.0156962052,
0.0081272246,
0.0241842065,
0.0354334749,
-0.0045486595,
0.0253141001,
0.0211879518,
0.0650616214,
0.0338165946,
0.0013947318,
-0.0012287056,
0.0154846655,
0.022693228,
-0.059920907,
0.0097122164,
0.0534924865,
-0.0702969208,
0.0253537633,
0.0513486341,
-0.0371082127,
0.0232734662,
-0.05936433,
-0.020562876,
-0.0092303446,
0.0255859513,
0.02948438,
0.0257628765,
0.0483662561,
0.0478446037,
-0.0651100799,
0.1403499395,
0.0263415948,
-0.0122512728,
-0.042575147,
-0.0157422386,
0.0383944511,
-0.0290710498,
0.0425736308,
0.0778108612,
-0.0155677032,
0.0059608594,
-0.0478360504,
0.0412720665,
-0.0233539194,
-0.054757297,
0.0328466631,
0.0302976873,
0.0183426365,
0.0341627225,
-0.0117678326,
-0.0367539711,
-0.0510601625,
0.0184983276,
-0.0360682532,
-0.0388164893,
-0.0945019871,
-0.0598269925,
-0.0017137296,
0.0632859766,
-0.0402339809,
-0.0431043468,
0.0137550756,
0.0327379778,
-0.0083479527,
-0.0448514596,
-0.0423139445,
-0.0079862056,
-0.0099399704,
-0.0296380874,
-0.0002679368,
0.0104685714,
-0.020380931,
0.0714452043,
-0.0010510171,
-0.1024360955,
0.0081943814,
0.0139609007,
-0.0382812954,
-0.0124049447,
-0.0149419475,
0.1111145392,
0.0189225208,
-0.0241598114,
-0.0293363649,
-0.0594907068,
0.0369929262,
-0.0418558456,
0.0302075204,
-0.0761311576,
-0.0141414357,
0.0233938787,
-0.0242012255,
-0.0155340889,
0.0160758141,
-0.0071439589,
0.0070301187,
0.0438355803,
-0.0337614715,
0.04149995,
0.0544309467,
-0.0010610175,
0.0488710925,
0.0282354373,
0.0116394153,
-0.0129458178,
0.0300581995,
-0.0092257829,
-0.0022193617,
0.0071307248,
0.0477229357,
-0.0140657695,
-0.0143835945,
-0.0260359645,
-0.0138368113,
0.0268818308,
-0.062337961,
0.0170265753,
0.0091698747,
0.0394261666,
0.0460105538,
0.0677735209,
0.0732722878,
0.0051346952,
0.0180207938,
0.0319882892,
-0.0086709494,
-0.0745871812,
-0.061218515,
0.0258050803,
-0.0985091925,
-0.0312216226,
0.005043366,
0.0600605719,
-0.002451991,
-0.0526499413,
0.0217577517,
-0.0347574875,
0.0330734886,
-0.005627776,
0.0497088209,
-0.0487496033,
0.0241716616,
0.0135835474,
-0.0022562486,
-0.0206824243,
-0.0041329507,
-0.0347910523,
0.0121220658,
0.0263229143,
0.0245834198,
-0.0074204723,
-0.023152478,
0.0365418792,
-0.0130885439,
0.018482931,
0.020518668,
0.0094065275,
-0.0141200554,
-0.0232789963,
-0.0021272253,
0.0292206183,
0.0252547804,
0.0076508825,
-0.0425342061,
0.0148834717,
-0.0074325344,
0.0005375538,
-0.0199531503,
0.0398999155,
0.0461936556,
0.0203311685,
-0.0031608751,
0.0210631359,
0.0344262347,
0.0352845192,
0.0028810101,
0.0285383109,
0.0282705463,
-0.0502575226,
-0.0530369058,
0.0220549516,
-0.0459158234,
0.0680629313,
0.0224114396,
0.0029515047,
0.0633956641,
0.0431061499,
0.0072252494,
-0.0614447966,
0.0487372167,
-0.0416412018,
0.0247824229,
-0.0105480477,
-0.0389765799,
0.0076632993,
-0.0195514672,
-0.0163418707,
0.008512374,
0.0417668931,
-0.0687903464,
0.0182815976,
0.0361644588,
0.0518956892,
-0.013587133,
0.0252450611,
-0.0029729526,
-0.0266532246,
0.0233147591,
0.0319379568,
0.0035446272,
0.0332660116,
-0.0210053064,
-0.0594430938,
-0.0034607565,
-0.0363740362,
0.0369408205,
0.0084553082,
0.0128062274,
0.0040463805,
-0.0038717187,
-0.0031550899,
0.0054581161,
0.0118365521,
-0.0206403453,
-0.0032832962,
0.0067840889,
-0.0549533889,
-0.0137305362,
-0.0342969373,
-0.0360548422,
0.0458019413,
0.0254714489,
-0.0238438435,
-0.0391827412,
-0.0185562149,
-0.0499876663,
0.0311422218,
-0.0102647897,
-0.019000303,
0.0608897842,
-0.0298562218,
-0.0290398486,
0.0191952549,
0.0084963832,
0.0340982303,
0.0058237445,
-0.0402401276,
-0.0654400736,
0.0074313157,
-0.01402664,
0.0412104279,
-0.0508090593,
0.0158867128,
0.015781533,
0.0507590175,
-0.0221439879,
0.0029090086,
-0.0168391354,
0.0094963284,
0.0077685178,
-0.0297728181,
-0.0239671301,
0.0012307797,
0.0206961576,
0.0761828721,
0.0179528631,
-0.0433858782,
0.0707852691,
-0.0098023349,
0.0262845922,
-0.0474001914,
-0.0406806692,
-0.0504514463,
0.0194318071,
-0.0287794806,
0.0270646345,
-0.0136844199,
0.0132462401,
-0.0139760589,
-0.008736182,
-0.0264319275,
-0.0358210355,
-0.0083582457,
-0.0523418337,
0.066005148,
0.009943313,
-0.0362480022,
-0.0013026968,
0.0355680846,
-0.0346525982,
0.0060532438,
-0.0713431761,
0.0195305608,
0.0005649773,
0.0249806065,
0.0373140723,
-0.0102033317,
-0.0229391586,
0.0112930201,
-0.0409013331,
-0.0152933476,
0.0311974883,
0.0129175987,
0.0158822443,
0.0345219709,
-0.0150809167,
-0.0200596675,
0.0170612615,
0.0403580666,
-0.0230806526,
0.0489957966,
0.0298460182,
0.0400669686,
-0.0549018793,
0.0813083947,
0.0525780097,
-0.0466897413,
-0.006467971,
0.0586989969,
0.0451946408,
0.0054990221,
0.0068853907,
-0.0371334851,
-0.037321575,
0.0207591951,
0.0347581841,
-0.0564396903,
-0.0343143865,
-0.0081202453,
-0.022898782,
0.0009509526,
0.0075886347,
-0.0094513977,
-0.0904826522,
-0.0046547824,
0.0383310802,
-0.0347703733,
-0.0214007348,
-0.0223489646,
-0.0459184498,
0.0237923618,
-0.0163472034,
0.0007074546,
0.0410892442,
-0.0022655178,
0.0103132734,
-0.015008253,
-0.1109771058,
0.0134915197,
-0.0079532089,
0.0764165297,
0.0290911663,
-0.0061605638,
-0.0226193536,
-0.0508222692,
0.0214586612,
0.0434394628,
0.0098523311,
-0.0292554237,
-0.0808610618,
0.009735927,
0.0111794313,
-0.0670420751,
0.0291890334,
-0.0507826842,
-0.0558514744,
-0.034887664,
0.0711155906,
-0.0591146238,
0.041965194,
-0.0615834706,
0.0018152919,
0.0440313108,
0.0143579086,
-0.030653825,
-0.0311185904,
-0.0411623344,
-0.0059294752,
-0.0215078853,
-0.0121787796,
-0.0643469617,
0.0111451708,
0.02001472,
-0.028902417,
-0.0232573319,
0.006469762,
-0.0089298747,
0.022196861,
0.0425206944,
0.043141596,
0.0735941902,
0.0065260581,
-0.0379437692,
-0.000402318,
-0.0123487208,
0.0116403503,
-0.0315502994,
0.0443068109,
0.0501278676,
-0.0701302141,
0.006405571,
0.0502266958,
-0.0534599088,
-0.0428620242,
0.0198212396,
0.0367374755,
0.0354550183,
-0.017983105,
-0.0348132923,
-0.0530801788,
-0.0004425164,
0.0284654163,
-0.0760347173,
0.0119297113,
0.0168017969,
0.0043316511,
0.0187989604,
0.0188055262,
-0.0621641055,
0.0713571161,
0.0023950422,
0.033353664,
-0.0001765453,
-0.0018926215,
0.0598572195,
0.0226755049,
-0.0224113688,
0.0448241644,
-0.0108871385,
0.008426344,
-0.0035739758,
-0.0387256555,
-0.0073155384,
-0.0430675037,
0.0258537475,
-0.0268470235,
0.0502759144,
0.0616340525,
-0.052355025,
0.0398533158,
0.0439469106,
0.0341017619,
0.000093363,
-0.0032968863,
0.0293632485,
-0.003753565,
0.0106320642,
0.0126160486,
-0.0047490369,
-0.0045923521,
0.0135775097,
-0.063842915,
0.0460159145,
0.0477394983,
-0.0131950695,
-0.0813816562,
0.0070746434,
-0.0280446764,
-0.0565230586,
0.0505461581,
0.0558029376,
0.0372243151,
0.0846016929,
-0.0253933407,
-0.0650878772,
0.0068678008,
-0.021463275,
0.0139840264,
0.0299638715,
0.0067170523,
0.0662207529,
0.0338148735,
-0.0101811523,
-0.0159052759,
-0.0335859619,
-0.0205751956,
0.0207926482,
-0.0221970584,
-0.0117501291,
-0.0077532819,
0.004976565,
0.0004972043,
0.0406809524,
-0.0369731188,
0.0280445386,
-0.0854792297,
-0.0154845864,
0.0089890165,
0.0415750891,
0.0133705605,
-0.0517347008,
0.030384589,
0.0733276084,
-0.0262445323,
-0.0268255789,
0.0282366537,
-0.0259864684,
-0.0978066102,
-0.0075385752,
0.0072532371,
-6.535800018e-33,
0.0099019697,
0.0215992313,
0.0558081381,
-0.0133166909,
-0.0247649606,
-0.01186259,
-0.0127006266,
-0.0224269498,
-0.0258702058,
-0.0328721851,
-0.0176610332,
0.017179681,
0.0130394064,
0.0494264923,
0.0026340273,
0.0016284733,
0.0299798921,
-0.0075789141,
-0.0379655175,
-0.0032015571,
0.034018252,
-0.0033438499,
0.0092639681,
-0.0863129422,
-0.0075546987,
-0.0045988467,
-0.0361536667,
0.0197232142,
-0.0694371611,
0.0444375537,
-0.0205249526,
0.0074112383,
0.0362877883,
-0.0172768608,
-0.003908359,
0.0652207434,
-0.0183808077,
-0.0033343274,
0.0328251645,
0.0185873564,
-0.0419177487,
-0.0571907647,
0.023241384,
0.0023515273,
-0.058699958,
0.0693976358,
0.0223305244,
-0.0152178872,
0.010142453,
-0.0170370936,
-0.0311126225,
-0.0034465129,
-0.0075376937,
-0.0311090983,
-0.032653261,
-0.0483158156,
-0.0451484174,
0.0195296947,
-0.0150101492,
0.0181946568,
-0.0394327827,
-0.0036640305,
0.0460772179,
-0.014103665,
-0.0148770353,
0.0352089591,
0.0107130399,
-0.0375234932,
-0.0424895771,
-0.0201149583,
0.0146111455,
0.0379366837,
0.0067310901,
0.0459053777,
0.0292517692,
0.0215493347,
-0.1015918851,
-0.0129896086,
-0.0000050153,
0.0497924164,
0.0391682722,
0.0242043752,
-0.0185331348,
-0.0017569391,
-0.0219590291,
-0.037598297,
-0.0284306221,
-0.0197579134,
0.0168118253,
-0.0025156853,
-0.0391566865,
0.0287983324,
0.0477115512,
-0.0468082987,
-0.0015748946,
0.0390870012,
0.0321050361,
0.0155790905,
-0.0356403701,
0.0117119299,
-0.0582744814,
-0.0113310972,
-0.0101710558,
-0.0096945073,
0.0260857791,
0.0107645774,
-0.030915264,
0.0856433883,
0.0024572795,
-0.0291830655,
-0.0196771286,
-0.0127226943,
0.0319027081,
-0.0161990952,
0.0850657895,
0.0138452677,
0.0078072264,
0.0444869138,
-0.0128358537,
-0.0324229077,
0.0162473004,
-0.0378281549,
0.0008523543,
0.0314810872,
-0.030534016,
0.0046263644,
-0.0992084667,
0.0166134723,
0.0410947502,
-0.0448772945,
-0.037909016,
0.0378235094,
3.312e-7,
0.0033593026,
0.0219077934,
0.0517709963,
-0.002584883,
0.0254343748,
0.0264364351,
-0.0048390925,
0.0244641993,
-0.0181055889,
0.0164394975,
-0.0319845192,
0.0246430729,
0.0433014147,
-0.01078186,
-0.0336570218,
0.0745682716,
-0.0300992876,
-0.0010857324,
-0.0489223301,
0.0334062018,
0.0944778696,
0.054260511,
0.0089366613,
-0.0058865538,
0.0140564786,
-0.0205897782,
0.0361586176,
-0.0273534488,
0.0019712446,
-0.0066567091,
-0.013133456,
0.0103785172,
-0.0471602343,
-0.0115895458,
0.0283110309,
0.0293047242,
-0.0083701806,
0.0333328769,
-0.0188214704,
-0.0130666299,
-0.0415972844,
-0.044721175,
-0.0334110633,
-0.014832831,
0.0330694541,
-0.0604850054,
-0.0316246375,
-0.0021508336,
0.0108376108,
-0.0125147067,
-0.0586743206,
0.00612634,
-0.011484121,
-0.0616157539,
-0.0278825089,
0.0188967921,
0.0065029627,
-0.0500940196,
0.0239513703,
0.0251938161,
0.0014970363,
-0.0108823376,
-0.0517728366,
0.0160133783,
0.1057008728,
-0.068134889,
0.0341447592,
2.77316635e-34,
0.0080558099,
0.0023347938,
0.0217096973,
0.0147217428,
-0.0002513161,
-0.0138339121,
-0.0664159209,
0.0334096849,
0.0062589017,
-0.0663484186,
-0.0030947025
] |
http://arxiv.org/abs/2401.16635v2 | http://arxiv.org/pdf/2401.16635v2 | Improving Reinforcement Learning from Human Feedback with Efficient Reward Model Ensemble | 2024-01-30T00:17:37 | 2024-05-21T22:21:16 | Shun Zhang, Zhenfang Chen, Sunli Chen, Yikang Shen, Zhiqing Sun, Chuang Gan | cs.LG | cs.LG, cs.AI, cs.CL | Reinforcement Learning from Human Feedback (RLHF) is a widely adopted
approach for aligning large language models with human values. However, RLHF
relies on a reward model that is trained with a limited amount of human
preference data, which could lead to inaccurate predictions. As a result, RLHF
may produce outputs that are misaligned with human values. To mitigate this
issue, we contribute a reward ensemble method that allows the reward model to
make more accurate predictions. As using an ensemble of large language
model-based reward models can be computationally and resource-expensive, we
explore efficient ensemble methods including linear-layer ensemble and
LoRA-based ensemble. Empirically, we run Best-of-$n$ and Proximal Policy
Optimization with our ensembled reward models, and verify that our ensemble
methods help improve the alignment performance of RLHF outputs. | Improving Reinforcement Learning from Human Feedback with Efficient Reward Model Ensemble: Reinforcement Learning from Human Feedback (RLHF) is a widely adopted
approach for aligning large language models with human values. However, RLHF
relies on a reward model that is trained with a limited amount of human
preference data, which could lead to inaccurate predictions. As a result, RLHF
may produce outputs that are misaligned with human values. To mitigate this
issue, we contribute a reward ensemble method that allows the reward model to
make more accurate predictions. As using an ensemble of large language
model-based reward models can be computationally and resource-expensive, we
explore efficient ensemble methods including linear-layer ensemble and
LoRA-based ensemble. Empirically, we run Best-of-$n$ and Proximal Policy
Optimization with our ensembled reward models, and verify that our ensemble
methods help improve the alignment performance of RLHF outputs. | [
0.0384686813,
0.0616222546,
0.0037935725,
0.0120640052,
-0.0098954616,
0.0011696131,
-0.030595921,
-0.025984183,
-0.0259951241,
-0.0014625487,
-0.0524238162,
-0.0601215996,
0.0289807413,
-0.0182199758,
0.005145146,
-0.0758403987,
0.0069080372,
-0.0485367067,
-0.0045845634,
-0.0367765576,
0.0153032653,
0.0011440009,
0.0006103893,
0.0695049465,
0.0421182588,
0.0011379339,
-0.0174974222,
-0.0009070532,
0.0717473999,
-0.0104423082,
0.0261473097,
0.047767859,
0.0135475472,
0.0463173687,
0.0000019837,
-0.0306719784,
-0.0238776784,
-0.014015845,
-0.0073954435,
-0.0302868206,
0.036542505,
0.0681028143,
-0.0364938602,
0.0321490057,
-0.0191271305,
0.0101103196,
0.0525677912,
0.0512653217,
0.023452403,
0.0533879064,
0.0151849864,
0.0074757244,
-0.021532679,
-0.0308021437,
0.0148453359,
-0.0256181303,
0.0223468505,
-0.0215716213,
0.0185896754,
-0.0466464311,
-0.0227114353,
-0.0056286515,
0.0219644643,
-0.0162370689,
0.0503526777,
0.0599065013,
0.01036637,
-0.0012484085,
-0.0456505492,
0.0702496022,
-0.0276508424,
-0.0052902345,
-0.0214476176,
-0.0229104646,
0.0035201558,
-0.0077719735,
-0.0008058774,
0.0277146418,
-0.0048625642,
0.0412590168,
-0.0110114971,
0.0373198465,
0.0339264199,
0.0336394235,
-0.0099535538,
0.0960514173,
-0.0053997943,
-0.0225629751,
0.0256484505,
0.0153103499,
-0.0287430491,
-0.0395520553,
0.0168605503,
0.0001403835,
-0.0762797371,
0.0043183798,
-0.0361196622,
-0.0059252423,
-0.0048811515,
0.0001054388,
0.0712272748,
0.0331390835,
0.0208636876,
0.0288957972,
0.0017668745,
0.0646501258,
-0.0168836899,
-0.0088558281,
-0.0230018273,
0.0147022195,
-0.0212070197,
-0.0011045672,
0.0101470891,
0.047168646,
-0.027832076,
-0.026154153,
-0.0948285013,
0.0231425073,
-0.0397224501,
0.0596973002,
0.0271587651,
0.0641714185,
-0.0512067303,
-0.0121096922,
-0.0416736603,
0.0324351415,
-0.0138915926,
0.0280297119,
0.0601309724,
0.01337987,
0.0101093613,
0.017498469,
-0.0134283276,
-0.0251233988,
-0.0064151096,
0.0640549436,
-0.0054674423,
-0.0025536201,
0.0346998237,
-0.0020118598,
-0.0203718115,
-0.0483485907,
-0.0442801937,
-0.0874222219,
0.0051228399,
-0.0279856641,
-0.0042163855,
-0.0349205844,
0.010194296,
-0.0017772152,
-0.0104966303,
0.0637528598,
0.036946509,
-0.0128793065,
0.0294101294,
-0.0279663354,
-0.0162706748,
0.0457427576,
0.0693425387,
0.0929156393,
-0.005708091,
0.0067179939,
0.0052389554,
0.0400887392,
0.0098389415,
0.0032643306,
-0.0330466367,
0.0060214978,
0.0212004166,
0.060759116,
0.0461759605,
0.0244286507,
-0.0330658853,
0.0225696694,
0.0902026221,
0.0363841951,
-0.0236061979,
-0.0461884886,
-0.0211478081,
0.0058133518,
-0.0143324174,
-0.0366776623,
-0.058731515,
0.0703547075,
-0.012494687,
0.002394476,
0.0599750169,
0.0432331115,
-0.0103560137,
-0.0500738919,
-0.002535102,
-0.0166270733,
-0.0006455981,
0.0533715673,
0.0437643975,
-0.0571527369,
-0.0568462089,
-0.0080662435,
0.0608450919,
-0.0341178812,
0.02571675,
-0.0320818461,
0.0628212467,
0.1032793969,
0.0206388012,
-0.0366409495,
0.0105293356,
-0.0490823835,
-0.0304049421,
0.0413112491,
0.0227010585,
-0.030897202,
-0.0042073359,
-0.0243954025,
-0.0617981292,
0.0651568919,
-0.0010712424,
0.0368409939,
0.0511416234,
-0.0111185536,
-0.0340576954,
0.0249047969,
-0.071212545,
0.0303848758,
-0.0031877551,
-0.0124534303,
0.0352648906,
0.0288262833,
0.0471715331,
-0.0049593467,
-0.0089671528,
0.0051312973,
0.0201977286,
-0.0126635041,
-0.0690372586,
0.0028917145,
-0.0584841184,
-0.0687435344,
0.0179876629,
0.000609817,
0.0121648498,
0.0281454064,
-0.0360987559,
-0.0362142511,
-0.0056184768,
-0.0323071219,
0.0068655028,
-0.0001238554,
0.0300656725,
-0.0114719495,
-0.0112993196,
-0.0740391538,
0.0283326861,
-0.0236954819,
0.0137650343,
-0.0577276908,
0.0035277426,
-0.0028246762,
-0.0896382481,
-0.0170744993,
0.0396629497,
0.015937265,
0.0505130216,
0.0698540732,
0.0048446818,
-0.0333235301,
-0.0400664769,
-0.0346362181,
-0.015320044,
-0.0157358162,
0.012071236,
0.0038648953,
0.0105038211,
-0.0079443306,
-0.0285331085,
-0.0379379243,
-0.008915293,
-0.0001118058,
0.0023449771,
0.0225153081,
0.0106864544,
-0.0373273715,
-0.0659709424,
-0.0102103529,
0.0076931785,
0.0094401678,
-0.0245488528,
0.008125497,
-0.0445144996,
-0.021145815,
-0.0241049882,
0.0061365208,
-0.0102374768,
-0.0138214612,
0.0242672637,
-0.0078287618,
0.002553304,
0.097898677,
0.0225752946,
0.0429025143,
0.046445597,
0.0549260266,
-0.1064974144,
0.0018971666,
-0.0609406047,
-0.0333567224,
-0.0314564556,
-0.0083787721,
0.0014955946,
0.0473790467,
0.0101388656,
0.0285062902,
-0.070269607,
0.0276540443,
-0.0166962482,
-0.0561932921,
-0.0750284791,
-0.0063608694,
-0.0421924554,
-0.0032022495,
-0.0348920785,
-0.0598669238,
0.0965472236,
0.0141609609,
0.0043495665,
-0.0461851582,
-0.0225430913,
-0.0260665659,
0.0128863873,
-0.0148434136,
0.0341613665,
-0.0254063588,
-0.0183656681,
-0.0459825173,
-0.0378222764,
0.038332209,
-0.0395334028,
0.0259806421,
-0.000513176,
0.0189098939,
-0.0040524346,
-0.092065312,
-0.0532350093,
-0.0497562811,
-0.0146507388,
0.0107722748,
-0.0125806434,
-0.042814672,
-0.0366731584,
-0.0046035489,
0.050028909,
-0.0075525707,
-0.0467030816,
-0.0050496506,
-0.0051280786,
0.0184126608,
0.0827628076,
0.0174827259,
-0.0044402047,
-0.0434619337,
0.0184556246,
0.0247149728,
0.0730703473,
0.0047036614,
-0.0227107666,
0.0137246484,
0.0492309444,
-0.0432224311,
0.0289126709,
0.0439745896,
0.0400865935,
-0.0205753893,
0.0071806689,
0.091555275,
0.0148416953,
-0.0775108635,
-0.0025946766,
-0.0082072066,
0.0808772668,
0.0092675528,
0.0439396575,
-0.0125302728,
-0.0253363103,
0.003918285,
0.0457720459,
0.0057116058,
-0.0082274899,
0.0321864448,
-0.1138257757,
-0.0000516582,
0.0183140095,
-0.0552263521,
-0.0198497456,
-0.0277102813,
0.0150115071,
-0.0401234403,
0.0060148053,
0.0266777202,
0.020477226,
-0.0013164207,
0.0382870473,
0.0167766493,
-0.036065422,
0.0199988056,
-0.0668512806,
0.0529483929,
0.1533590406,
0.0221070834,
-0.0206021648,
-0.011930122,
-0.0597215742,
0.0248086862,
-0.0067753545,
0.0238492265,
0.0056825699,
-0.0063575897,
0.0034741657,
-0.0076230578,
0.0771583617,
-0.0210666209,
0.0141579555,
0.0029507717,
0.0533579923,
-0.0262666624,
-0.0104804188,
0.0015296512,
-0.0204622559,
0.0624007769,
0.060331054,
-0.0111830309,
-0.0662809461,
-0.0150101138,
-0.0263514947,
-0.0816346258,
-0.0409149788,
0.0036316963,
0.004571917,
0.0290552881,
0.01111688,
-0.0611198545,
-0.0545031242,
0.0239163227,
0.0216932353,
-0.0135323703,
0.0243862737,
-0.02394793,
0.0236881152,
-0.0081538362,
0.0171928611,
-0.0096416455,
0.0568541996,
-0.0464765765,
-0.0097698187,
-0.0195920449,
-0.0406938754,
-0.0323964208,
-0.0237588957,
-0.0437821969,
-0.0715737045,
-0.0071067,
0.0734584406,
0.0114450073,
0.0660210252,
-0.0025089979,
-0.0609169751,
0.0906130001,
0.0780509114,
0.0196091104,
0.0243284963,
0.001152117,
0.0094600152,
-0.0278084315,
0.0378402621,
-0.0401291288,
0.0448683612,
0.0362744369,
0.0306949206,
0.0143516324,
-0.0258400291,
0.0307887234,
-0.0280093253,
0.0285988655,
-0.0676047727,
-0.0628944114,
-0.0192758515,
0.0070793894,
0.004999131,
-0.0328049064,
-0.0025534697,
-0.0225249138,
0.0513858646,
-0.0272118133,
0.0008084318,
-0.0083202198,
0.0420055054,
-0.0079127569,
-0.0257733986,
0.0234935042,
0.032269828,
0.0249804538,
0.0542423353,
0.0595676228,
-0.0059617795,
0.0028924015,
0.0190670304,
0.0405990779,
0.0305819549,
0.0233042352,
0.036295794,
0.0006829168,
-0.0089329164,
0.0247025862,
-0.0539291091,
-0.034281712,
-0.0599654987,
-0.0017086522,
0.0601711087,
0.0241214894,
0.0431219675,
-0.0199917797,
-0.0069384938,
-0.0139333056,
0.0121824341,
-0.0424012318,
-0.0003876335,
0.0555447973,
0.0168832354,
0.0020007808,
-0.0084200511,
0.0604713522,
0.0130001046,
0.0280699339,
-0.0219663046,
0.022102125,
-0.0286904406,
0.0124577973,
0.0067606503,
-0.0031571658,
-0.0404856391,
0.0003407277,
-0.0129874218,
-0.022717895,
-0.0064312685,
0.0414383262,
0.0157385692,
-0.0304438788,
0.0349681899,
0.0112728076,
-0.0404592268,
0.020701427,
-0.0558807366,
-0.0138043445,
-0.0686752051,
-0.0132319042,
-0.0660640001,
-6.325216818e-33,
-0.0392914936,
-0.0345828086,
0.0286493134,
-0.0133693675,
-0.0213087909,
-0.0375963375,
-0.0148011968,
0.0049114632,
0.0427746922,
0.0179338343,
-0.0268247519,
-0.0115641616,
0.0152178165,
0.0051233163,
0.0504838675,
0.0063316496,
0.0427176133,
-0.0283892658,
-0.0745643899,
0.0125422981,
0.0400871411,
0.0201182459,
0.0358049981,
-0.051707197,
0.0267567746,
0.0348467901,
-0.0552510768,
0.0126764281,
-0.0153191807,
0.0646167099,
-0.0081914216,
0.0378387347,
-0.017241884,
-0.0608018301,
-0.0066093821,
0.0145247011,
-0.0469049327,
-0.0184763633,
-0.0108782183,
0.005610703,
-0.0114945266,
-0.0111360839,
0.0409017429,
-0.0056503089,
-0.0511836074,
0.0031747597,
0.0394019112,
-0.015490232,
-0.0383335687,
-0.0021411269,
-0.0541600138,
0.0063312105,
0.0142462598,
0.010659229,
0.0138758626,
0.0342705473,
-0.0311099198,
0.0298689734,
-0.045210693,
0.0214505326,
-0.0060369656,
0.043267794,
0.0559396744,
0.0656637773,
0.0197080486,
0.0593324602,
0.0127155203,
-0.0240640324,
-0.0322505459,
-0.0209528729,
0.0084056007,
0.0201994814,
0.0361040756,
0.0496007353,
-0.0001878541,
0.0030802682,
-0.0128337704,
0.0447892249,
0.0094905868,
0.0064294096,
0.0377760865,
0.0343752131,
-0.0673088804,
-0.0292963851,
-0.0444197543,
-0.0431456268,
-0.0194818992,
-0.0306376424,
0.0314532034,
0.0247371383,
-0.0157986991,
-0.0296068899,
0.0106263161,
-0.0405467525,
-0.0412542857,
-0.0117555382,
-0.0263424832,
0.0145614017,
0.0110924738,
-0.0070039192,
-0.0875110403,
-0.0119593814,
-0.0306887794,
-0.0052140616,
-0.0211310852,
0.0360576957,
0.0453399867,
0.0315143131,
-0.0451827273,
0.0078189038,
-0.020489376,
0.0186585877,
0.0052584358,
0.0129182152,
0.0485613793,
-0.0162777398,
-0.0006695889,
0.0209646896,
-0.0149708148,
-0.0229176786,
0.0097956834,
-0.0403507054,
-0.007926519,
-0.0001586592,
-0.0221706256,
0.0180769619,
-0.0369422473,
-0.0185593572,
0.011724323,
-0.0239806026,
-0.0030678576,
-0.0130150486,
2.645e-7,
-0.0262086671,
0.054871235,
0.0185976569,
-0.017419612,
-0.0004919968,
0.0681383163,
-0.0194519311,
0.0106276246,
0.0128368512,
0.0004509132,
0.0155210868,
0.0254486576,
0.0446563922,
-0.0172333661,
-0.0793270096,
-0.0194115769,
-0.0821562558,
0.0217418764,
-0.0145503609,
-0.0385469235,
0.0067969183,
0.0842317492,
0.0579812564,
-0.0212327614,
0.0093977787,
0.0417023152,
0.0656499863,
-0.0292825755,
0.0105825383,
-0.0061700898,
0.0255462974,
-0.0397436768,
-0.0390633792,
-0.0082756607,
0.0030299942,
0.0213104635,
-0.0246432554,
0.0380140804,
-0.0365740582,
-0.0082855234,
-0.0355320834,
0.0203071814,
0.0366451479,
-0.0515209399,
0.0159251355,
-0.055024676,
-0.0042632688,
-0.0378712304,
0.0211663991,
-0.0205745194,
0.0264276434,
-0.0202837847,
0.0181844216,
0.0008850062,
0.0011960971,
0.0109941252,
0.0157569014,
-0.0727882683,
-0.0433897786,
0.0083981371,
-0.0427916162,
-0.0400869921,
-0.0324164927,
0.0688544437,
-0.016268827,
-0.0644039512,
-0.0222206898,
2.152755125e-34,
0.0132109327,
0.0041911369,
-0.0005546528,
-0.0034666143,
0.0108902724,
0.01906096,
0.0133380312,
0.0070619253,
0.0642937869,
-0.0530567765,
-0.0183894727
] |
http://arxiv.org/abs/2401.16589v2 | http://arxiv.org/pdf/2401.16589v2 | ToPro: Token-Level Prompt Decomposition for Cross-Lingual Sequence Labeling Tasks | 2024-01-29T21:44:27 | 2024-03-13T09:45:02 | Bolei Ma, Ercong Nie, Shuzhou Yuan, Helmut Schmid, Michael Färber, Frauke Kreuter, Hinrich Schütze | cs.CL | cs.CL | Prompt-based methods have been successfully applied to multilingual
pretrained language models for zero-shot cross-lingual understanding. However,
most previous studies primarily focused on sentence-level classification tasks,
and only a few considered token-level labeling tasks such as Named Entity
Recognition (NER) and Part-of-Speech (POS) tagging. In this paper, we propose
Token-Level Prompt Decomposition (ToPro), which facilitates the prompt-based
method for token-level sequence labeling tasks. The ToPro method decomposes an
input sentence into single tokens and applies one prompt template to each
token. Our experiments on multilingual NER and POS tagging datasets demonstrate
that ToPro-based fine-tuning outperforms Vanilla fine-tuning and Prompt-Tuning
in zero-shot cross-lingual transfer, especially for languages that are
typologically different from the source language English. Our method also
attains state-of-the-art performance when employed with the mT5 model. Besides,
our exploratory study in multilingual large language models shows that ToPro
performs much better than the current in-context learning method. Overall, the
performance improvements show that ToPro could potentially serve as a novel and
simple benchmarking method for sequence labeling tasks. | ToPro: Token-Level Prompt Decomposition for Cross-Lingual Sequence Labeling Tasks: Prompt-based methods have been successfully applied to multilingual
pretrained language models for zero-shot cross-lingual understanding. However,
most previous studies primarily focused on sentence-level classification tasks,
and only a few considered token-level labeling tasks such as Named Entity
Recognition (NER) and Part-of-Speech (POS) tagging. In this paper, we propose
Token-Level Prompt Decomposition (ToPro), which facilitates the prompt-based
method for token-level sequence labeling tasks. The ToPro method decomposes an
input sentence into single tokens and applies one prompt template to each
token. Our experiments on multilingual NER and POS tagging datasets demonstrate
that ToPro-based fine-tuning outperforms Vanilla fine-tuning and Prompt-Tuning
in zero-shot cross-lingual transfer, especially for languages that are
typologically different from the source language English. Our method also
attains state-of-the-art performance when employed with the mT5 model. Besides,
our exploratory study in multilingual large language models shows that ToPro
performs much better than the current in-context learning method. Overall, the
performance improvements show that ToPro could potentially serve as a novel and
simple benchmarking method for sequence labeling tasks. | [
0.0153301349,
-0.0296381749,
0.0161166061,
0.0150359226,
-0.0247495491,
0.0434887297,
0.032056503,
-0.0249343812,
0.0135144694,
-0.0530305766,
-0.0726953447,
-0.0265945457,
0.0047531119,
0.0215620156,
0.0223520789,
-0.0364844985,
0.0731112361,
-0.0282601044,
0.0234716516,
0.0313075334,
-0.0189625639,
-0.0023609037,
-0.0058136801,
0.0275501925,
-0.0077365297,
-0.0237336028,
-0.0255646724,
0.0032007871,
0.0286459327,
-0.0422927067,
-0.0136775644,
0.0643925369,
-0.0148187438,
0.0483180434,
0.0000018228,
-0.0430680625,
-0.0576385148,
-0.0085122623,
-0.0182590764,
0.0087872073,
0.0062964708,
-0.0125220232,
-0.0129231969,
0.0091701569,
-0.0015184368,
-0.0636248216,
0.027171392,
0.0662898496,
0.0445142686,
0.0803404897,
-0.001211687,
-0.0212286711,
0.0077192858,
-0.0249828361,
0.0067109866,
-0.0286238696,
0.013233955,
-0.064937599,
0.0106729781,
0.0367787071,
-0.0244467873,
0.0281644017,
0.0088558355,
0.0111967856,
-0.0198442619,
0.0267945826,
-0.0222844072,
-0.0384255722,
-0.028332971,
0.0691443533,
-0.020727504,
0.0049759694,
0.0289462488,
-0.0024362912,
-0.0172128603,
-0.076364994,
-0.0148598356,
0.0171786994,
-0.0035349433,
0.0128721315,
0.0557049699,
-0.0115837287,
0.0189753305,
-0.0056328373,
-0.0463765338,
0.0696195662,
0.0299266018,
-0.0274171997,
0.0019946257,
-0.0106640989,
-0.0181931704,
-0.0759823173,
0.0327836946,
0.0056208451,
-0.0530386977,
0.0072516198,
-0.0572919361,
-0.0126602147,
0.0205419939,
0.023788983,
0.0027801679,
0.0484406389,
-0.0367648192,
0.0491122566,
-0.0359323993,
0.0142326141,
-0.0081290025,
0.0284519084,
-0.0183139984,
-0.0486783832,
-0.0594930947,
0.0003725931,
-0.0452705845,
0.0333210342,
-0.0726610124,
-0.0092877978,
-0.0273774695,
0.0067988858,
0.0361895971,
0.0122256633,
-0.0089386487,
-0.0555493869,
-0.0082919998,
0.0108510852,
0.0018245474,
0.0565548353,
-0.0479713455,
-0.0017811292,
-0.0030548305,
-0.0644053072,
-0.0316430144,
0.0077227238,
0.0093496945,
-0.0427670516,
0.0092329085,
0.0273138918,
-0.0128797172,
-0.0698661581,
-0.0247211102,
-0.0363427512,
-0.0257541519,
-0.04534439,
-0.0193686113,
-0.0403909311,
0.0437351391,
-0.0130429221,
0.0109190866,
0.0129172327,
0.0031191956,
0.0071451762,
-0.0317899175,
0.0826653987,
-0.0243978202,
-0.0404111408,
-0.0155062936,
-0.0049508498,
0.0192033127,
0.0586623847,
0.013055169,
0.030678371,
0.0535319746,
0.0087820115,
-0.0471447296,
0.0567476526,
0.0358494483,
-0.0004139299,
0.0023654036,
-0.0014335958,
0.0664262101,
0.0091460282,
0.0231559649,
0.028815804,
-0.0068896404,
0.0270300861,
0.0189129915,
0.019995423,
-0.0359507278,
-0.036540173,
0.0678703934,
-0.0048304931,
-0.0281655677,
0.0188398231,
0.0001033529,
-0.0046918271,
-0.007423338,
0.0276273619,
0.0307398252,
0.0313261151,
-0.0534473695,
0.0016410119,
0.0040827421,
-0.0118398946,
-0.0293871593,
0.0141202919,
0.0594222657,
0.0258076619,
-0.0730866715,
0.0748919323,
0.0300043914,
-0.0245791618,
0.0329876393,
-0.0173456799,
0.021972945,
0.0160496272,
0.0007065178,
-0.0437998511,
-0.0278146379,
-0.0466386713,
0.0461554416,
0.0329055898,
0.0198814329,
-0.0193698704,
0.0292330775,
-0.0425984301,
-0.0262252875,
-0.008045679,
-0.0139517784,
0.0488978066,
-0.0452920161,
0.0195613336,
0.0162632558,
-0.0041215722,
0.0164446514,
0.0450397618,
-0.0441635884,
-0.0330878347,
0.0466660261,
0.039307002,
0.019633716,
0.07440155,
0.0078812316,
-0.006359864,
0.0481482297,
-0.0342520177,
0.0511297174,
-0.0072997175,
-0.0370706357,
-0.0844286084,
0.0214830376,
0.0003797832,
-0.0331426337,
-0.0253175274,
-0.0545408204,
-0.0571535081,
-0.0371153094,
0.0401332714,
0.0087711681,
-0.0477671213,
0.0216319337,
0.0172211863,
0.0349733606,
-0.0215370804,
0.0263077449,
0.0080651408,
0.0466406308,
0.0613200404,
-0.0085731437,
0.0097500393,
-0.0346992724,
-0.0557632148,
0.0420551524,
-0.0110059911,
-0.0547438487,
0.0231814459,
0.0241002105,
-0.0572761521,
0.0044319155,
-0.1278738827,
0.0239146147,
-0.0338894762,
0.0532297976,
0.0102012455,
0.0530112535,
-0.0188391432,
-0.0331105031,
-0.0125538697,
-0.0368509814,
-0.0048994846,
0.0282194391,
0.0478694141,
0.0183454491,
-0.0215317011,
-0.0043631438,
-0.0040343888,
-0.0346015692,
0.0300609116,
-0.0175254457,
-0.0593057275,
-0.0461258106,
0.0053815264,
-0.0123791853,
0.0115759121,
0.0131660104,
-0.0113026155,
0.0030250857,
0.0293379016,
0.072446391,
0.0234331135,
-0.0250888988,
0.0054559452,
-0.0030574084,
-0.0235123783,
-0.0248542968,
0.0184455477,
0.0018705965,
-0.0073378016,
0.0027320227,
-0.0283457842,
0.0408338681,
0.0828241259,
-0.0196665954,
0.0095341243,
-0.0524892285,
-0.0099602435,
-0.0128006004,
-0.0272261947,
-0.0369214863,
0.0198921375,
-0.0164263099,
0.02806825,
-0.0297193918,
0.0241498724,
0.0558626018,
-0.0780012086,
-0.0338939466,
-0.0494408384,
0.0137721309,
-0.0541556366,
0.0393051803,
0.0130350394,
0.0111073116,
0.0455539711,
-0.0732068419,
0.0003957449,
0.0082757287,
-0.0177787822,
0.0054483754,
-0.0223027561,
-0.0084336447,
-0.0187950023,
-0.027084887,
0.0016566135,
-0.005184161,
-0.03803223,
-0.0166850332,
0.0606199242,
-0.0296870619,
0.0222827867,
-0.0018094091,
0.0086289141,
0.0373673216,
0.0537099466,
-0.0099617522,
0.0223815218,
-0.0396557152,
0.0689880326,
0.0202772189,
-0.0281110313,
-0.0026633684,
-0.0328556672,
0.0412194245,
-0.0043619922,
0.0448197983,
-0.0064211204,
0.022682555,
0.0090298811,
0.0380036533,
-0.0621540919,
-0.0334081538,
0.0040244772,
0.0509125367,
-0.0339677408,
-0.0048400778,
0.1238416508,
0.0248977263,
-0.029610863,
-0.0200803746,
0.0082972357,
0.0020952458,
0.0208337028,
0.0119898394,
-0.0926220194,
-0.0635434762,
0.0195921343,
-0.0139554879,
0.0046267663,
-0.058679305,
-0.0092057306,
-0.1223653331,
0.0038812775,
-0.0322786383,
-0.0984150842,
-0.010655066,
-0.0820157155,
-0.0078031905,
0.0067077857,
0.0194419511,
-0.030204989,
-0.0048801512,
0.0028781923,
0.0257372614,
0.0409673788,
-0.041416958,
0.022696279,
-0.0634871572,
0.0347299427,
0.0430907942,
0.0275544804,
-0.0005435522,
-0.0144755309,
0.0408536717,
-0.0279212091,
-0.0073911371,
-0.0275434926,
-0.0012383497,
-0.0517973937,
-0.0256425999,
0.0305014048,
0.0033115267,
-0.0150092039,
-0.0101067126,
0.0029512201,
0.0851783305,
0.0486715399,
0.0475367159,
0.0194006413,
-0.0179547295,
0.0343463197,
0.0782691166,
0.0377876051,
-0.0288731139,
-0.0219185222,
0.0082892589,
0.0091123264,
-0.0309648011,
0.0301910006,
0.011896044,
-0.0725059062,
-0.0606582277,
-0.0361798778,
-0.06016564,
0.0768548176,
0.0446800552,
0.0023912967,
0.0187923741,
-0.0401461758,
-0.0035885551,
-0.0251069292,
0.0435093231,
-0.0181530509,
0.0198072884,
-0.1085295677,
0.0120244166,
0.0237246267,
-0.0064773271,
-0.0219955798,
-0.0025971751,
-0.0235249028,
-0.050133191,
-0.0130076595,
0.0323002487,
-0.0014362154,
0.0305928644,
-0.0166817755,
-0.0231808387,
0.033605136,
0.069412984,
-0.0463185981,
-0.0177781787,
0.0307964645,
0.0291045588,
-0.028999038,
0.0319253318,
-0.0328857414,
0.0083391936,
0.0376982316,
0.0558706746,
0.0174671225,
-0.0394022651,
-0.0142931286,
0.0507667474,
0.0932591707,
-0.0364468247,
-0.0306439269,
-0.0288881082,
0.0283824522,
-0.0174208228,
-0.0000383958,
-0.0025572809,
-0.0135134561,
0.0109849591,
-0.0403507389,
-0.001859767,
0.0497288145,
-0.0390818939,
0.0296723675,
-0.0174898636,
0.0246048234,
0.013240966,
0.0331977792,
-0.0069555342,
-0.0022773601,
-0.0157040954,
0.0088509517,
0.0428944081,
0.0285605788,
-0.0505693592,
0.0383662879,
0.0018701674,
0.0309459157,
-0.0170448627,
0.0627271757,
-0.0181879289,
0.0115043363,
0.0024635093,
0.0404108092,
0.0658431724,
-0.0123582017,
0.0422031246,
0.0314538404,
-0.0303215645,
-0.0175035596,
0.0422024503,
-0.0131254876,
-0.0239128973,
0.0167643167,
-0.0013474302,
-0.0334258303,
0.0049879462,
0.0619502962,
-0.0059613348,
0.0320869461,
0.0260697398,
0.0002399348,
-0.0267834291,
0.0343483053,
0.0170304775,
-0.0073864167,
-0.0332638249,
-0.0017750375,
-0.041209463,
-0.035877768,
0.0271729212,
-0.001269719,
-0.0230467748,
0.0534471646,
0.0233417265,
0.035971351,
0.0016211692,
-0.0170190576,
-0.0381352901,
0.0261197202,
0.0733005702,
-0.0577890389,
0.0182592627,
-5.808456644e-33,
-0.0506016389,
0.0131812217,
0.0005816004,
-0.008531549,
0.0053019817,
0.0262841247,
-0.0262340251,
-0.0193624496,
0.0157869384,
-0.0251213964,
-0.0348234661,
-0.0473019704,
0.045179978,
0.0340424962,
0.0009144665,
-0.0235892236,
0.0861444846,
0.0098556206,
-0.0360082798,
0.012873196,
0.0700507909,
0.0597351007,
0.1057610363,
0.033746127,
0.0528492816,
-0.0103197331,
0.0216183439,
0.0175373871,
0.066151686,
0.0490499996,
-0.0127151655,
-0.0063324729,
0.009530291,
-0.0660928115,
-0.0156668331,
0.0627454892,
-0.1373878866,
-0.0235337485,
0.0231881849,
0.0165122226,
-0.0030288333,
-0.0538213514,
-0.0291516632,
-0.0062889014,
0.0302213077,
0.0207426958,
0.0171479415,
0.0172971059,
-0.0074865804,
-0.0371642075,
-0.0678665638,
0.023140464,
-0.0121426294,
-0.0040831114,
0.0689277351,
0.0795116723,
0.006032276,
0.0415780395,
-0.059690699,
0.0021665562,
-0.0228759497,
0.0231682975,
0.0497212037,
0.0181880184,
-0.0106448866,
0.0243068617,
0.0507321656,
-0.0162882153,
0.0099129463,
0.0274137184,
0.0311177932,
-0.0079150274,
-0.0098813307,
0.0395336524,
0.022684766,
-0.0116233118,
-0.0810364783,
0.0204516239,
0.0393724367,
0.0071574464,
-0.0091992794,
-0.0200237371,
-0.0005336903,
0.000896911,
-0.0305180438,
-0.0671934336,
-0.0476501323,
-0.0088857971,
0.0273773391,
0.0028803272,
-0.0472349301,
0.0026707705,
0.0829384699,
0.0121539999,
-0.0330348201,
0.0169254187,
0.0215037875,
0.0068872366,
-0.0162978154,
0.0538854152,
-0.0981157497,
-0.0545696802,
0.0257123113,
-0.0180457421,
0.0193465762,
-0.0180377346,
0.0539251938,
-0.0052195317,
-0.059889935,
0.0271218847,
-0.0152501715,
0.0549161024,
-0.0306767412,
0.0388449281,
0.0258266367,
0.000847193,
-0.0042281672,
0.0485513993,
0.0304369368,
-0.0426561199,
0.0044923117,
-0.0003254797,
0.0104097864,
0.0381547511,
0.0019595174,
-0.003902205,
-0.0030765589,
0.0577168837,
0.0083644725,
-0.0210407712,
-0.0059189308,
0.0267270468,
2.561e-7,
-0.0031628229,
0.018791113,
0.0286566708,
-0.0245897602,
-0.0141282026,
0.051992815,
-0.0168192089,
0.0464396887,
0.0065635713,
-0.0184283871,
-0.0107263057,
-0.0452680402,
0.0161040891,
0.015852632,
-0.1310568899,
-0.0259388704,
-0.0374077559,
0.0208297819,
-0.0618131123,
0.036394503,
0.0284925494,
0.0874091089,
-0.0049414076,
0.0045695389,
-0.0338488333,
0.0097763073,
0.0339530893,
-0.0070346785,
-0.0522816256,
0.0119921388,
0.0089677405,
-0.0086577414,
-0.0144795515,
0.0215843171,
-0.0340169333,
0.0278854147,
0.0064912951,
0.0529598221,
0.0006769517,
-0.0195725337,
0.0139011424,
-0.0092562726,
-0.0180184394,
-0.0445598178,
0.092318207,
-0.0277266279,
-0.0142420428,
-0.0272618029,
-0.0307686981,
-0.0009697875,
0.0459956042,
-0.037227992,
-0.0346436761,
0.077206865,
0.0379500687,
-0.0146588944,
-0.0490449108,
-0.0458642729,
0.0099443756,
-0.0151610728,
-0.0093727876,
-0.0064678276,
-0.0000041259,
-0.01768969,
0.0064932322,
-0.0606269203,
-0.0034749559,
2.455476833e-34,
0.0074689514,
-0.0200527031,
-0.0201781075,
0.0170565862,
-0.0018579422,
-0.0326584168,
-0.035574723,
0.006587611,
-0.007342739,
-0.0531082638,
-0.0271453951
] |
http://arxiv.org/abs/2401.16587v3 | http://arxiv.org/pdf/2401.16587v3 | A Linguistic Comparison between Human and ChatGPT-Generated Conversations | 2024-01-29T21:43:27 | 2024-04-26T01:16:35 | Morgan Sandler, Hyesun Choung, Arun Ross, Prabu David | cs.CL | cs.CL, cs.AI, cs.CY | This study explores linguistic differences between human and LLM-generated
dialogues, using 19.5K dialogues generated by ChatGPT-3.5 as a companion to the
EmpathicDialogues dataset. The research employs Linguistic Inquiry and Word
Count (LIWC) analysis, comparing ChatGPT-generated conversations with human
conversations across 118 linguistic categories. Results show greater
variability and authenticity in human dialogues, but ChatGPT excels in
categories such as social processes, analytical style, cognition, attentional
focus, and positive emotional tone, reinforcing recent findings of LLMs being
"more human than human." However, no significant difference was found in
positive or negative affect between ChatGPT and human dialogues. Classifier
analysis of dialogue embeddings indicates implicit coding of the valence of
affect despite no explicit mention of affect in the conversations. The research
also contributes a novel, companion ChatGPT-generated dataset of conversations
between two independent chatbots, which were designed to replicate a corpus of
human conversations available for open access and used widely in AI research on
language modeling. Our findings enhance understanding of ChatGPT's linguistic
capabilities and inform ongoing efforts to distinguish between human and
LLM-generated text, which is critical in detecting AI-generated fakes,
misinformation, and disinformation. | A Linguistic Comparison between Human and ChatGPT-Generated Conversations: This study explores linguistic differences between human and LLM-generated
dialogues, using 19.5K dialogues generated by ChatGPT-3.5 as a companion to the
EmpathicDialogues dataset. The research employs Linguistic Inquiry and Word
Count (LIWC) analysis, comparing ChatGPT-generated conversations with human
conversations across 118 linguistic categories. Results show greater
variability and authenticity in human dialogues, but ChatGPT excels in
categories such as social processes, analytical style, cognition, attentional
focus, and positive emotional tone, reinforcing recent findings of LLMs being
"more human than human." However, no significant difference was found in
positive or negative affect between ChatGPT and human dialogues. Classifier
analysis of dialogue embeddings indicates implicit coding of the valence of
affect despite no explicit mention of affect in the conversations. The research
also contributes a novel, companion ChatGPT-generated dataset of conversations
between two independent chatbots, which were designed to replicate a corpus of
human conversations available for open access and used widely in AI research on
language modeling. Our findings enhance understanding of ChatGPT's linguistic
capabilities and inform ongoing efforts to distinguish between human and
LLM-generated text, which is critical in detecting AI-generated fakes,
misinformation, and disinformation. | [
0.0834342465,
0.0468146838,
-0.0387851633,
0.0294172093,
-0.0380027629,
0.0193017665,
-0.0233286843,
0.0135681769,
0.0579495952,
-0.0241454076,
-0.0286469068,
-0.0199099779,
-0.0046073105,
0.02716312,
0.0245488975,
-0.0337594636,
0.0278081167,
0.0611827224,
0.0752659813,
-0.0091049755,
0.0286765136,
0.051703386,
-0.0048919874,
0.0455316827,
-0.0503110439,
0.0050183642,
-0.0201898534,
0.0350022875,
0.0231076311,
-0.0078900959,
0.0316889547,
0.0461655259,
0.0208587591,
0.0518841781,
0.0000018537,
-0.0170248039,
0.0158944707,
-0.0071167029,
-0.0339641795,
-0.0149208652,
0.1100600958,
-0.0169637296,
-0.0141294813,
0.0355591029,
-0.0532164723,
0.0045986245,
0.028066719,
0.0813119039,
0.0664827004,
0.0976965204,
-0.0303132664,
0.0105063124,
0.0357415117,
0.0063362075,
0.0031119911,
-0.0062534311,
0.0015880113,
0.0155450087,
0.0068257228,
-0.0077099716,
0.0211264156,
0.0097906506,
-0.0014608043,
-0.0196858644,
-0.0013980068,
0.0305673238,
-0.0519149005,
-0.0579351373,
-0.0076257903,
0.0599640571,
0.0257684346,
-0.0134863732,
0.0188225079,
0.00960063,
0.0217040572,
-0.0148819732,
-0.0294350162,
0.0357219651,
-0.0166346636,
0.0156451967,
-0.0025439668,
0.0099261636,
0.0037959139,
0.0255683381,
0.0381923243,
0.0241074152,
-0.0090818871,
-0.0203694664,
0.00944064,
-0.003572843,
-0.011328754,
-0.0611465201,
-0.0156128937,
0.0014425751,
-0.0127815492,
0.0230352767,
-0.050219547,
0.0261457469,
0.0379853435,
-0.061200764,
-0.0110065639,
0.0316076241,
-0.0151098166,
0.0184890162,
0.0244145691,
0.0309819989,
0.0218302347,
0.0417770967,
-0.0192056615,
-0.008450916,
-0.044482477,
-0.0122329248,
-0.0154997241,
0.0157594513,
0.0077550439,
-0.0016565138,
-0.0592951365,
0.0199096631,
-0.0029472453,
-0.0005080768,
-0.0044741444,
0.0033795496,
-0.0170107819,
0.0100373402,
-0.0097758668,
-0.0221232995,
-0.0388356447,
0.0129912812,
-0.0004761212,
-0.1110684946,
0.0150819113,
0.0522548556,
-0.0332227461,
-0.0097111911,
-0.0081467507,
0.0374397449,
0.010204629,
-0.050233569,
-0.0176038407,
-0.0087806191,
-0.0125929387,
-0.0646037385,
-0.0048501622,
-0.0181742478,
0.0343736932,
-0.0190282222,
-0.0148372734,
-0.0150655303,
0.0311922673,
0.0202955008,
-0.0488675497,
0.0479892902,
-0.0124317501,
0.0128849493,
0.0465859026,
0.0285733324,
-0.0143389115,
0.0396494344,
0.0423325226,
0.0385775119,
0.0803694353,
0.0057573481,
0.0341255181,
-0.0007013252,
-0.0007059646,
0.0074188332,
0.0118711218,
-0.0090375161,
-0.034528587,
0.023040073,
0.0212498922,
0.059328936,
-0.0271543562,
-0.068361111,
0.0696607158,
0.0705575943,
-0.0187545866,
0.0543467402,
0.0065986374,
0.0510918275,
0.0442397296,
0.0247331113,
-0.0291152652,
0.0125603909,
-0.0277642477,
-0.0444600284,
0.0075630927,
-0.0173770487,
-0.0131508913,
-0.0341583826,
-0.0294377375,
-0.0157252941,
0.0407772586,
-0.0367667936,
-0.0117700296,
-0.043989718,
-0.0251793973,
0.0002205337,
0.0278286096,
-0.0811677128,
-0.0505694225,
-0.055373922,
0.0230600294,
0.0580041073,
0.0344471745,
0.0032544143,
-0.0376120023,
-0.02879427,
-0.0095458627,
0.076914385,
0.0498508401,
0.0072125555,
0.0004441578,
-0.020595124,
0.0231318623,
-0.0296585988,
-0.0011665605,
0.0040753633,
0.0146989347,
0.0367261097,
-0.0134305591,
-0.0092262644,
-0.0209980719,
0.0340710059,
-0.0494209006,
0.0144939916,
-0.0079997452,
-0.0143668121,
0.0279139783,
0.022461107,
-0.0254460797,
-0.0128937066,
0.0311525967,
0.0047207335,
-0.0633333325,
0.0363160074,
-0.0126446169,
-0.017401129,
0.0246733017,
-0.0249257982,
-0.0290556736,
0.006464771,
-0.0225814264,
0.0096504949,
0.0031298955,
-0.0718647465,
0.0416569375,
-0.0677228794,
0.0022542495,
-0.0258197095,
0.0548059382,
0.0070538502,
0.0278703962,
-0.0305248629,
-0.0194183774,
-0.0074704248,
-0.008637066,
0.0107376454,
-0.0339160636,
-0.0027161124,
0.0332151987,
0.024422301,
0.000945949,
0.0203843117,
0.0098245172,
-0.0748939961,
-0.0448172577,
-0.1422561854,
0.0006934046,
-0.0016942231,
0.0145236831,
0.002571211,
-0.0247975867,
-0.0358737782,
0.0420252346,
-0.0564561263,
0.0052673277,
0.0234851725,
0.0002778272,
0.0393860973,
-0.0434532017,
0.0063103754,
-0.0496393703,
0.0431705229,
-0.0154330013,
0.0056088064,
0.013032238,
-0.0293957163,
-0.0338606536,
0.0117799873,
-0.0704759881,
-0.0023083843,
-0.0308842547,
-0.1033906713,
0.0144305034,
0.0217807312,
0.0074307448,
0.0533887036,
0.0646622926,
0.0878027305,
-0.0139068039,
0.0264161173,
-0.0367497653,
0.0084424065,
-0.0037115796,
0.0133182667,
-0.0121577922,
-0.0633930564,
-0.0067096571,
0.03093227,
0.0163430087,
0.0141458474,
-0.0570756346,
0.0198024884,
0.0124874609,
-0.0526062176,
0.0183159839,
-0.0211992823,
-0.0332934745,
0.0070462725,
0.0064138146,
-0.0790079087,
0.0903320983,
-0.0111323036,
-0.0122271487,
-0.0343835726,
-0.0408824943,
-0.0059978585,
0.0351782516,
-0.0433116779,
-0.0017606325,
0.0027814147,
0.0021099346,
-0.0385746546,
-0.0372809358,
0.0083076889,
-0.0612638816,
-0.0029105083,
-0.0225380901,
-0.01219643,
0.0382602252,
-0.0192337204,
-0.0671431348,
-0.0743386,
-0.013248736,
-0.0047673401,
0.0328163654,
-0.0465502888,
-0.030331023,
-0.0319990218,
0.0214555245,
-0.0120729888,
-0.0081418222,
-0.0172921252,
0.0230108537,
0.0142843779,
0.0296424422,
0.0167002082,
0.0542291105,
-0.0102454145,
0.0570216849,
0.0745642707,
-0.0213145744,
0.02183358,
0.0251496173,
-0.0345646404,
-0.0233979058,
0.0164799914,
-0.0112605598,
-0.0229772031,
0.0301648993,
0.001108894,
0.0028207789,
0.1324660778,
0.037751127,
-0.0532444566,
0.0171802137,
-0.0446225256,
-0.0031634956,
0.0023944459,
0.0020721732,
-0.0664713234,
-0.0007881098,
0.0046338933,
0.0140067646,
-0.0471762903,
-0.0241463315,
-0.0183133595,
-0.0824398249,
0.0425817817,
-0.0020651012,
0.0224619024,
-0.01349294,
-0.0222019516,
-0.001283108,
0.0325059481,
0.0212246589,
-0.0312883668,
0.0215602759,
-0.0095202643,
-0.008822822,
-0.0499422662,
-0.0240349714,
0.0292752553,
-0.009753664,
0.0848019794,
0.0205480345,
0.1066832468,
0.0013939996,
-0.0142894015,
0.0407446362,
0.0403579883,
-0.0162404254,
0.0027238145,
0.0689997151,
-0.0324387588,
0.0014423383,
-0.0429136865,
0.0497826673,
-0.0262439325,
0.011227305,
0.0003946745,
0.0437814556,
-0.0497085154,
-0.0344301984,
-0.0248151962,
0.0439535156,
0.0304571148,
0.0848733783,
0.0179608203,
-0.0226994846,
-0.0005646409,
-0.0023366788,
0.0324294604,
0.0425777696,
-0.0635308549,
0.0303905606,
-0.0030325311,
-0.0292740874,
-0.0842911825,
-0.08335343,
-0.0106904097,
-0.0232863724,
-0.0382350907,
-0.0333887525,
0.0139934253,
0.0064347666,
-0.0254364163,
0.0048170825,
-0.018351227,
0.0320189372,
-0.0041190386,
-0.0198490769,
0.0145369079,
-0.0292187445,
0.0203505605,
0.0157186333,
-0.0432112031,
-0.050258562,
0.0051657888,
-0.0021603717,
0.0160282236,
0.0275201965,
-0.0041764285,
-0.0175940916,
0.0006711507,
0.0603763871,
-0.0514836982,
0.0347907767,
0.0320944712,
0.0716045871,
0.0102538941,
0.0402909294,
-0.0108563891,
0.0167655163,
0.0031636846,
-0.0003847439,
-0.0145339379,
-0.0730754063,
0.0062191826,
0.0418364443,
0.0235633459,
-0.0036536865,
0.0077908565,
-0.0266964715,
-0.0334093794,
-0.038690228,
0.0036740345,
-0.0500105023,
-0.0380582102,
-0.0101395855,
0.0062004016,
-0.050058648,
-0.0399265066,
-0.0901186168,
0.0234814677,
-0.0090036131,
0.0151019404,
-0.0173773449,
0.017046351,
-0.020849485,
0.0018334141,
-0.0037681381,
0.0370570831,
-0.0304702744,
0.0229638722,
-0.0141586373,
0.0682030767,
0.0021807074,
-0.0327643156,
0.0404130295,
0.0420179293,
-0.0336478427,
-0.0229713134,
-0.004190899,
0.0466474406,
0.0267139543,
0.0135569414,
0.050607726,
-0.0577618666,
0.0047941017,
-0.0169865135,
0.0315750055,
-0.0322657451,
-0.0138228899,
-0.0010655482,
-0.0098518822,
-0.070317328,
0.0376502834,
0.048050236,
-0.019008752,
0.0371008106,
-0.0014081779,
-0.036057245,
-0.038067352,
-0.0170320626,
0.0660346225,
-0.0326378457,
0.0076807053,
0.0257133562,
-0.0786572322,
-0.0761569738,
0.03106617,
-0.010813213,
0.0194085725,
-0.0361706316,
0.0114389854,
0.0563451089,
-0.0299968328,
0.0343245342,
-0.0565735586,
0.0362056568,
-0.0391508937,
-0.0200112239,
-0.0012623179,
-6.235250358e-33,
0.0099125253,
-0.0625492334,
-0.0334955789,
0.0378500782,
-0.0582789257,
0.0208319928,
-0.0112634199,
0.0138623202,
-0.0164030381,
0.0682476535,
-0.0281790607,
0.0700282082,
-0.0040753824,
-0.0159676205,
-0.0160353892,
-0.0247723013,
0.0141048403,
-0.0259309504,
-0.008935716,
0.0272697844,
0.0130002666,
0.0231822487,
0.0586331934,
-0.105910778,
0.04986066,
-0.0023946459,
-0.0188836362,
0.0125735225,
0.0163986161,
0.032179419,
-0.034281332,
0.0649053529,
0.0178736355,
0.0158876237,
0.0095976535,
0.0169942491,
-0.000740089,
-0.0376332402,
0.0205853358,
-0.0168215763,
-0.0352820866,
-0.0282949172,
0.0161382128,
-0.0305605642,
-0.0386507884,
0.0028434242,
0.0045892494,
-0.0078922193,
0.0261258744,
-0.0308223069,
-0.0365376472,
0.013673869,
-0.0110144727,
-0.039750129,
-0.0185326822,
0.0207049455,
0.0018535915,
0.0368559025,
-0.0527647361,
0.0086609023,
-0.0610639043,
-0.0421920344,
0.0095014516,
-0.0637334585,
0.0181778185,
0.0838547722,
0.0430712067,
0.0440785885,
-0.0597949997,
-0.006173823,
0.0062189708,
0.0207071379,
0.0595921203,
0.0744514316,
-0.0427786522,
-0.0618429668,
-0.0447367355,
0.0146683706,
0.0741527081,
0.0146088833,
0.0239343457,
0.0040420769,
-0.0409289859,
0.0003961319,
0.0056294529,
-0.0273297895,
-0.0147261322,
-0.0266621765,
0.0040286314,
-0.0157860015,
-0.0010398088,
-0.0600783862,
0.0054343389,
-0.0597818568,
0.0376241431,
0.0061339326,
-0.0470311381,
0.0022413153,
0.0168792754,
0.0263102483,
-0.00853376,
0.0058545414,
0.0200124476,
0.0152655542,
0.0515963845,
-0.0077931196,
0.0208394192,
0.0095076775,
-0.0394321233,
0.0043331767,
-0.0118304985,
0.053596925,
-0.023032682,
-0.0278485883,
-0.0173676088,
0.0176660866,
-0.0206924621,
0.0261412263,
0.0131535856,
0.0525404923,
-0.0413980857,
-0.0199801326,
-0.0087654106,
0.0046717394,
-0.0461240374,
0.0044828719,
-0.0187888462,
-0.0221984815,
0.0081834607,
-0.048918236,
-0.009820722,
-0.0482494794,
2.758e-7,
-0.0107845254,
0.0817657486,
0.0248644724,
0.0116130384,
0.0304341372,
0.0516880192,
0.0004712766,
0.0379378907,
-0.0080769313,
-0.0672881901,
0.0243313648,
0.0176491421,
0.0879010037,
0.0427253358,
-0.0735031739,
0.0210146382,
-0.0737486258,
-0.0506406575,
-0.0265889447,
0.0180567764,
0.1044890285,
0.075617075,
0.0456053242,
-0.0113486098,
-0.0258552469,
0.0008140829,
0.040755298,
-0.0454830788,
-0.0441239551,
0.0210545957,
0.0495020859,
-0.0176331177,
0.0215044469,
0.0007509189,
0.0039823093,
0.0037216928,
0.021140432,
0.0726468563,
-0.0115517341,
-0.0430021286,
0.0107702203,
-0.0214134213,
-0.0078349626,
-0.0407058895,
0.0406788401,
-0.0497590564,
-0.0484541655,
-0.0441967919,
0.0084244786,
-0.0267150588,
0.069339186,
-0.0253743213,
-0.042516429,
0.0578050949,
0.0191568881,
-0.0022773545,
0.0126160812,
-0.0495330244,
0.0117874024,
-0.0227538869,
-0.0672341511,
-0.0532115884,
-0.0024382491,
0.0427930206,
0.0448765457,
0.0085582146,
0.0386872143,
2.100764525e-34,
0.0351851732,
0.0036258411,
0.0076648896,
0.0201182701,
0.0740224794,
0.0270663034,
0.027342055,
0.0030532849,
0.0395736992,
-0.0535804182,
0.0218660198
] |
http://arxiv.org/abs/2401.16582v1 | http://arxiv.org/pdf/2401.16582v1 | Massively Multilingual Text Translation For Low-Resource Languages | 2024-01-29T21:33:08 | 2024-01-29T21:33:08 | Zhong Zhou | cs.CL | cs.CL | Translation into severely low-resource languages has both the cultural goal
of saving and reviving those languages and the humanitarian goal of assisting
the everyday needs of local communities that are accelerated by the recent
COVID-19 pandemic. In many humanitarian efforts, translation into severely
low-resource languages often does not require a universal translation engine,
but a dedicated text-specific translation engine. For example, healthcare
records, hygienic procedures, government communication, emergency procedures
and religious texts are all limited texts. While generic translation engines
for all languages do not exist, translation of multilingually known limited
texts into new, low-resource languages may be possible and reduce human
translation effort. We attempt to leverage translation resources from
rich-resource languages to efficiently produce best possible translation
quality for well known texts, which are available in multiple languages, in a
new, low-resource language. To reach this goal, we argue that in translating a
closed text into low-resource languages, generalization to out-of-domain texts
is not necessary, but generalization to new languages is. Performance gain
comes from massive source parallelism by careful choice of close-by language
families, style-consistent corpus-level paraphrases within the same language
and strategic adaptation of existing large pretrained multilingual models to
the domain first and then to the language. Such performance gain makes it
possible for machine translation systems to collaborate with human translators
to expedite the translation process into new, low-resource languages. | Massively Multilingual Text Translation For Low-Resource Languages: Translation into severely low-resource languages has both the cultural goal
of saving and reviving those languages and the humanitarian goal of assisting
the everyday needs of local communities that are accelerated by the recent
COVID-19 pandemic. In many humanitarian efforts, translation into severely
low-resource languages often does not require a universal translation engine,
but a dedicated text-specific translation engine. For example, healthcare
records, hygienic procedures, government communication, emergency procedures
and religious texts are all limited texts. While generic translation engines
for all languages do not exist, translation of multilingually known limited
texts into new, low-resource languages may be possible and reduce human
translation effort. We attempt to leverage translation resources from
rich-resource languages to efficiently produce best possible translation
quality for well known texts, which are available in multiple languages, in a
new, low-resource language. To reach this goal, we argue that in translating a
closed text into low-resource languages, generalization to out-of-domain texts
is not necessary, but generalization to new languages is. Performance gain
comes from massive source parallelism by careful choice of close-by language
families, style-consistent corpus-level paraphrases within the same language
and strategic adaptation of existing large pretrained multilingual models to
the domain first and then to the language. Such performance gain makes it
possible for machine translation systems to collaborate with human translators
to expedite the translation process into new, low-resource languages. | [
0.0413192771,
0.0622310564,
0.0000103401,
0.0330081247,
-0.0574801564,
0.0225104187,
0.0250079613,
0.0140456613,
-0.0571088046,
0.0014908392,
-0.0173942391,
-0.039409101,
0.0129461093,
-0.0746701285,
0.0037455603,
-0.0878178105,
0.0181123931,
-0.0081209997,
-0.0164598748,
-0.0143014519,
-0.0209894348,
0.0548957959,
0.036570549,
0.0109980004,
0.0350572802,
0.0004418953,
-0.0084766652,
0.0423624925,
0.0612249449,
-0.0127735585,
-0.009690077,
-0.011114127,
0.0190013424,
0.034395583,
0.0000021069,
-0.0143586872,
-0.0432384312,
0.0067636762,
-0.0063707139,
0.0092056757,
0.0654701367,
-0.1261273026,
-0.0057107653,
0.0001197364,
-0.0185849015,
-0.0331130363,
0.0696858317,
-0.0018685141,
0.058733657,
0.075818263,
-0.0128484434,
-0.0477471538,
-0.0135891484,
-0.0177448951,
-0.0432461128,
-0.0368163288,
-0.0189139917,
-0.0209878217,
0.0239786394,
0.0168557223,
0.0211837608,
0.081533052,
0.0060405624,
0.0315178446,
-0.0269027632,
-0.0250838622,
-0.0548365302,
-0.0242292695,
0.0078253821,
0.0500375926,
-0.0540733077,
-0.0364597738,
0.0234765727,
0.0044085402,
-0.0338243693,
0.0070889969,
-0.034665294,
0.0271563791,
-0.0360266156,
0.0046226108,
0.0357096754,
0.0527780466,
0.0414622873,
0.0033838407,
-0.030641228,
-0.0337304547,
0.0178488642,
-0.027258534,
-0.0115665672,
0.0152809275,
-0.0476709679,
0.0020366497,
0.0264767371,
-0.009109335,
-0.0249256752,
0.0214966591,
-0.0506100096,
-0.0906761512,
0.0289408304,
-0.0721961036,
0.035881184,
0.0402513705,
0.0000540466,
0.0390068963,
0.0123609323,
-0.0029647318,
-0.0592246167,
-0.0188462622,
-0.07616207,
-0.0295309592,
-0.0178510472,
0.0509447195,
-0.0528476164,
-0.0257869866,
-0.0591332726,
0.0220416654,
-0.0225211754,
0.0051438529,
0.0488663167,
-0.0047011441,
0.0153150903,
-0.0168942753,
-0.0724456012,
-0.0450384133,
0.0301204976,
0.0559208207,
-0.0106288884,
-0.0072998414,
-0.0498042665,
-0.0626039729,
0.0043169539,
0.0067642564,
0.0325218514,
-0.0267826468,
0.0080566797,
0.0572712496,
-0.032732591,
0.0176129229,
-0.0524996705,
0.0047810189,
-0.0140552334,
-0.0414708443,
0.0018490124,
-0.0520266667,
0.0567978173,
0.046338439,
0.0384645648,
-0.0227176808,
0.0119151408,
-0.0142190615,
-0.082478784,
0.0330460072,
0.0116173588,
0.0005814395,
0.0421004854,
-0.0062216423,
0.0119466027,
0.0633638203,
0.0238357261,
0.0064765159,
0.0535578057,
0.0084719434,
-0.0589925125,
0.0407608524,
-0.0024465199,
0.0531122312,
-0.0267737843,
0.0143126827,
0.04020622,
0.0292681083,
0.0134011582,
0.0496021546,
-0.0110970931,
0.0335364081,
0.0626933798,
0.0172247607,
0.0030273721,
0.0456534624,
0.0736918747,
0.0268495753,
-0.0152292987,
-0.0064550969,
0.0134714209,
-0.0025363027,
0.0895235538,
-0.0076175239,
0.037333034,
0.0609012805,
-0.0069243922,
-0.0114987511,
0.0056697042,
-0.0340989865,
-0.0386558436,
-0.0108524123,
0.0572559871,
0.0041966368,
-0.0588247254,
0.0220425501,
0.031901639,
-0.0632322729,
0.0148093244,
-0.0047343858,
-0.0562702492,
0.0829265714,
0.0333299972,
-0.0203532167,
-0.0048558451,
-0.0044656484,
0.0560848825,
0.0302674565,
0.0312618464,
0.0213685818,
0.0736732334,
-0.0515507683,
-0.0189198498,
0.0341320187,
0.0051341252,
0.04960832,
-0.0118341157,
0.0393083207,
0.0280662235,
-0.0015794256,
-0.0660123751,
0.0088186823,
-0.0436729342,
0.0031589828,
0.0188193191,
0.0323534496,
0.0271465946,
0.027195923,
0.0062696212,
-0.0117527926,
0.0041952496,
-0.0728966966,
-0.0091268532,
0.0583228804,
-0.0270018391,
0.0088460203,
-0.01975164,
-0.0543614253,
0.009822865,
0.0218874607,
-0.0226395093,
-0.0421668552,
0.0111796102,
-0.0147005487,
0.0419633426,
-0.0159612596,
0.0692712441,
-0.0174414273,
-0.0070458059,
-0.0225730427,
0.0438654162,
0.0165545382,
0.0081079174,
0.008928325,
-0.0648636669,
0.0549333021,
-0.0291733332,
0.0095736394,
-0.0008640723,
0.0318250693,
0.0359597392,
0.0320281871,
0.0090104435,
-0.0030162048,
-0.0290690493,
-0.1005674303,
0.0074335933,
-0.0149075342,
0.03088733,
0.0313071832,
0.0827787071,
0.0181323569,
0.0101043079,
0.0140734762,
0.0213704668,
0.004779513,
0.0048270142,
0.005290675,
0.0132885901,
0.0333893895,
-0.003375421,
0.0173351318,
0.0522945225,
0.0455915034,
0.0471411981,
-0.0332069844,
-0.0470312946,
0.0024673073,
-0.0168811381,
0.0064431718,
0.0031096896,
-0.0747504607,
-0.0081464322,
-0.0138909901,
0.0173407737,
-0.0117416056,
-0.0140789943,
-0.0465839095,
0.0119318478,
0.032314226,
-0.0399578102,
0.0149462651,
0.0020907714,
-0.0504544489,
-0.0388718769,
-0.0491409227,
0.0013934975,
0.0042003333,
0.0175332148,
-0.0149429711,
-0.0582753494,
-0.0509597324,
-0.0328312144,
-0.0176359769,
-0.0027940159,
0.0022223615,
0.0099609224,
0.0284056868,
-0.0138503853,
-0.0334900133,
0.0207860991,
0.0051066494,
-0.0038226852,
-0.069060877,
-0.01652872,
-0.0665245056,
0.0185391288,
-0.0369849093,
0.0453825518,
-0.0534627736,
-0.0002646468,
-0.022446597,
0.0657941401,
0.0108846985,
-0.0033766439,
0.0357652307,
-0.055612579,
0.0099598709,
-0.0196226239,
-0.0539812781,
-0.0281224586,
-0.0501369052,
0.0027249779,
-0.0367671624,
0.0327335559,
-0.0060691014,
-0.0489244647,
-0.0019924177,
0.0489016101,
0.0866933167,
0.0334263146,
-0.0015833462,
-0.0182574354,
0.0243165996,
-0.0568489172,
-0.0354898348,
0.000662072,
-0.0352299958,
0.0063112276,
0.0268975794,
0.0586357154,
-0.045665808,
-0.0028615764,
-0.0219142754,
0.0392795578,
0.0127870981,
-0.0572751611,
-0.0327244177,
0.0444359332,
-0.0256010797,
0.0195056498,
0.0843390673,
0.0019148232,
-0.0654498786,
0.0030476216,
-0.0478571057,
-0.0222338177,
-0.0124909906,
0.0275395904,
-0.0640761405,
0.0087726517,
-0.0035567537,
0.0141314762,
-0.0137984753,
-0.0018332175,
-0.0087791253,
-0.0434343331,
0.0116852978,
0.0427747518,
-0.0621941872,
-0.013049894,
-0.0195990968,
-0.0234184582,
0.0004769651,
0.0008825143,
-0.0080204504,
0.0019415065,
-0.005250012,
-0.0239567608,
0.0404523835,
-0.0525747016,
-0.0339868255,
-0.0455311723,
0.0152105708,
0.0476649031,
0.0491347201,
-0.0043289983,
0.0063753305,
0.0499110036,
-0.0088537717,
0.0029508546,
-0.0094247069,
0.0162405819,
0.0367501862,
-0.0319213383,
0.0280011371,
0.0100806085,
-0.0591842197,
0.0142821176,
0.0541212521,
0.0463843457,
-0.0130926147,
-0.0117916716,
0.0350414552,
-0.0073433747,
0.0615011454,
0.0242616069,
-0.0263387114,
-0.0555339903,
0.0329870731,
0.0013545596,
0.010258927,
-0.0785812289,
-0.014176419,
0.0234447587,
-0.1112649813,
-0.0864305496,
-0.0735090524,
-0.0092704371,
-0.0276416745,
0.0590629801,
0.0273475926,
0.0011413534,
0.0241908859,
0.0050013913,
-0.0022808856,
-0.0088666948,
-0.0355933532,
0.002246927,
-0.0391618051,
-0.0110734738,
0.0034009353,
-0.0294323247,
0.0051433034,
0.0204719733,
-0.0420820527,
-0.0320274979,
0.0053247008,
0.0062515126,
0.0172039457,
0.01837538,
0.0209197309,
-0.0175334252,
0.0110151349,
0.0064166449,
-0.0572854094,
0.0134831611,
0.0114238085,
-0.0056898338,
-0.0213029664,
0.0649913624,
-0.000037971,
0.0338838026,
0.0249882881,
0.0172128696,
-0.0187511425,
-0.0362697393,
-0.0060946406,
0.0110248476,
-0.0105191004,
-0.0308845267,
-0.0457449369,
-0.0126595441,
0.025279386,
-0.0131595675,
-0.0230222307,
-0.0458521433,
0.0039237449,
0.0386662297,
0.0123540889,
-0.0007381255,
0.0247110464,
-0.0075948178,
-0.0431919768,
-0.0139589813,
0.0136084175,
0.006987127,
0.0736054182,
0.0177023504,
0.0232599117,
-0.0091884369,
0.0436244197,
-0.0378315039,
0.000850101,
-0.0512090065,
0.0029008391,
0.0009421899,
0.015004525,
0.0144046014,
0.0146461297,
-0.0323096178,
-0.044863984,
-0.0238129869,
0.0116834091,
0.0868926868,
0.0276637413,
0.0892704278,
-0.0095214471,
-0.0231052209,
-0.0183262415,
0.0415577106,
0.020977933,
-0.0111911027,
-0.0747954845,
-0.0147744212,
-0.0016207594,
0.0600572266,
0.0553960614,
-0.00422377,
0.0323427096,
0.0277274922,
-0.0209883749,
-0.0053352667,
-0.0319421813,
0.0142756356,
0.0008854473,
-0.0237162989,
0.0053738644,
-0.0336350724,
-0.0168284699,
0.0427178666,
0.0075021531,
-0.0487697534,
-0.031887494,
-0.0005305447,
0.0554461107,
-0.0329129025,
0.0272857547,
-0.0494217463,
0.0564767569,
-0.0349963941,
-0.0042995946,
0.0160303134,
-5.977215756e-33,
0.0340464637,
-0.0063372161,
-0.0021653844,
-0.0840361565,
0.0037988944,
0.0327302702,
0.0009805569,
-0.031564001,
-0.0191575531,
0.0012730149,
-0.0317293778,
0.0026228796,
0.0403013751,
0.0084449602,
0.0401584283,
0.0442096666,
0.0465645865,
0.0174889266,
-0.0208662506,
0.0256330818,
0.0464316346,
0.0182478987,
0.0837343186,
-0.0879451483,
0.0371633694,
-0.0554912649,
-0.0129705584,
-0.0030653321,
0.0740624815,
-0.0092317108,
0.0056364853,
0.0072768643,
0.0138629861,
-0.0480036475,
-0.004532807,
-0.0041665547,
-0.0890441909,
0.0208909474,
0.0550809093,
0.022481557,
-0.034885671,
-0.0888657421,
0.0073579042,
-0.0200180188,
0.0052423226,
-0.00523595,
0.001076995,
-0.0532321446,
-0.0014976469,
-0.0185631979,
-0.0697586238,
0.024362905,
0.0010941131,
0.022345772,
-0.009684057,
0.0727264956,
-0.0417350307,
0.0321129821,
-0.0310795736,
0.0089621106,
-0.0541812666,
0.0377702713,
0.0450148918,
-0.0384886302,
-0.0343624428,
-0.0255353469,
-0.0226839297,
-0.0179725494,
0.0316763259,
0.0037998792,
-0.023835713,
-0.0034617237,
0.0152775589,
0.1211384758,
0.0216165483,
-0.0209905393,
-0.0758498833,
0.0505311564,
-0.0275465939,
0.0179546438,
0.0272428468,
0.0186254755,
-0.0023397824,
-0.0015017011,
-0.0296899714,
-0.0241328888,
-0.0133993849,
-0.0033702066,
-0.0079386663,
0.0285728481,
-0.0218949541,
0.0427723527,
0.0589040443,
0.0027830235,
0.025850378,
0.0252030324,
0.0069420254,
0.0051596025,
-0.0258135311,
0.0594776198,
-0.0357316695,
-0.0279470459,
0.0130904634,
0.0254874118,
0.0157969315,
-0.0213711206,
0.0501584895,
0.0099348724,
-0.0442002341,
0.0125519838,
-0.0142826503,
-0.0063736076,
-0.0290567372,
0.0268830825,
0.0069456883,
-0.0253120232,
0.0233636107,
-0.018240517,
0.0455419645,
-0.0587212779,
0.0154629499,
-0.0034528775,
0.02383353,
0.0088743391,
-0.0548232831,
0.0250377152,
-0.0380433053,
0.0194033328,
0.0228410047,
-0.035151083,
-0.0271223243,
-0.0185141303,
2.852e-7,
0.0649236813,
0.0046618786,
0.0191590711,
-0.0110623054,
0.002158189,
0.0037417745,
0.0009866433,
0.0409054048,
-0.080506146,
0.0385127179,
0.0567852519,
0.0219505448,
0.1026813462,
0.0234628506,
-0.0915568396,
-0.0457633883,
-0.030898653,
0.0077888737,
-0.0534776002,
-0.0181493741,
-0.000257948,
0.0898052007,
-0.0070817037,
-0.0168399047,
-0.0249554217,
-0.0530637354,
0.0768068135,
-0.0480505228,
-0.0379986949,
0.0106329201,
-0.0014420648,
0.0090319309,
-0.0646651015,
-0.0172405057,
-0.0159879643,
-0.0123339659,
0.054637976,
0.1115838587,
0.0087860851,
0.0260400176,
0.0016113215,
-0.0365031473,
-0.0374412313,
-0.0237551127,
0.0418538004,
0.0050321342,
-0.0117033077,
0.0296329949,
0.0078064324,
0.0377512649,
0.0124169886,
-0.0073295212,
-0.0376969352,
0.0375625975,
0.0488971435,
-0.0133720795,
-0.0448345281,
-0.0383632518,
-0.0142914038,
-0.0106223598,
-0.0401057079,
-0.057218913,
0.0083033061,
0.0177285653,
0.0336072817,
-0.0385716371,
0.0470539182,
2.332311198e-34,
0.0466726162,
-0.0250952002,
-0.0157757439,
-0.0405552238,
0.0144186467,
-0.0263689049,
-0.0391475521,
0.0129950084,
-0.0053679491,
-0.0097713489,
0.0059235259
] |
http://arxiv.org/abs/2401.16578v3 | http://arxiv.org/pdf/2401.16578v3 | Leveraging Professional Radiologists' Expertise to Enhance LLMs' Evaluation for Radiology Reports | 2024-01-29T21:24:43 | 2024-02-17T03:07:00 | Qingqing Zhu, Xiuying Chen, Qiao Jin, Benjamin Hou, Tejas Sudharshan Mathai, Pritam Mukherjee, Xin Gao, Ronald M Summers, Zhiyong Lu | cs.CL | cs.CL, cs.AI | In radiology, Artificial Intelligence (AI) has significantly advanced report
generation, but automatic evaluation of these AI-produced reports remains
challenging. Current metrics, such as Conventional Natural Language Generation
(NLG) and Clinical Efficacy (CE), often fall short in capturing the semantic
intricacies of clinical contexts or overemphasize clinical details, undermining
report clarity. To overcome these issues, our proposed method synergizes the
expertise of professional radiologists with Large Language Models (LLMs), like
GPT-3.5 and GPT-4 1. Utilizing In-Context Instruction Learning (ICIL) and Chain
of Thought (CoT) reasoning, our approach aligns LLM evaluations with
radiologist standards, enabling detailed comparisons between human and AI
generated reports. This is further enhanced by a Regression model that
aggregates sentence evaluation scores. Experimental results show that our
"Detailed GPT-4 (5-shot)" model achieves a 0.48 score, outperforming the METEOR
metric by 0.19, while our "Regressed GPT-4" model shows even greater alignment
with expert evaluations, exceeding the best existing metric by a 0.35 margin.
Moreover, the robustness of our explanations has been validated through a
thorough iterative strategy. We plan to publicly release annotations from
radiology experts, setting a new standard for accuracy in future assessments.
This underscores the potential of our approach in enhancing the quality
assessment of AI-driven medical reports. | Leveraging Professional Radiologists' Expertise to Enhance LLMs' Evaluation for Radiology Reports: In radiology, Artificial Intelligence (AI) has significantly advanced report
generation, but automatic evaluation of these AI-produced reports remains
challenging. Current metrics, such as Conventional Natural Language Generation
(NLG) and Clinical Efficacy (CE), often fall short in capturing the semantic
intricacies of clinical contexts or overemphasize clinical details, undermining
report clarity. To overcome these issues, our proposed method synergizes the
expertise of professional radiologists with Large Language Models (LLMs), like
GPT-3.5 and GPT-4 1. Utilizing In-Context Instruction Learning (ICIL) and Chain
of Thought (CoT) reasoning, our approach aligns LLM evaluations with
radiologist standards, enabling detailed comparisons between human and AI
generated reports. This is further enhanced by a Regression model that
aggregates sentence evaluation scores. Experimental results show that our
"Detailed GPT-4 (5-shot)" model achieves a 0.48 score, outperforming the METEOR
metric by 0.19, while our "Regressed GPT-4" model shows even greater alignment
with expert evaluations, exceeding the best existing metric by a 0.35 margin.
Moreover, the robustness of our explanations has been validated through a
thorough iterative strategy. We plan to publicly release annotations from
radiology experts, setting a new standard for accuracy in future assessments.
This underscores the potential of our approach in enhancing the quality
assessment of AI-driven medical reports. | [
0.0625570938,
0.0284858532,
0.0113691604,
-0.0222324226,
-0.0354405455,
0.0317639261,
0.0154818716,
0.0113275908,
-0.0018126105,
-0.0114152916,
-0.0145273339,
-0.0330836475,
0.0163232293,
0.0287861209,
0.0135716395,
-0.0826315656,
0.0114430282,
0.015246572,
0.0289247613,
0.0235212166,
0.0104380259,
0.0364397764,
-0.0339589193,
0.0416872501,
-0.008403712,
-0.0042468519,
0.0304751638,
0.0006588252,
0.0366700627,
-0.0548841879,
-0.0342519619,
0.0411671847,
0.0114939772,
0.0006317519,
0.0000020167,
-0.0534992218,
-0.0244742893,
0.042033039,
-0.0168311764,
-0.0284777507,
0.0601972155,
-0.060084518,
0.006951109,
0.0246721935,
-0.0254459158,
-0.0090911118,
0.024535019,
0.0269440059,
0.0805216357,
0.0644770637,
-0.0188217591,
0.0059058964,
0.0249734093,
-0.0060168761,
0.0016252373,
-0.0400095135,
-0.0187292024,
0.0619827434,
0.009344154,
-0.00250002,
0.0138928955,
0.0366516374,
0.0256090984,
0.0029826348,
0.015634032,
0.049690593,
0.0002712191,
-0.0228686798,
0.0049020932,
0.0336639211,
0.0240280498,
-0.0027323661,
-0.0015093262,
-0.0125584751,
0.008489917,
-0.0127014844,
-0.0421609022,
0.0277593676,
-0.0194570906,
-0.0084341634,
0.0219566692,
0.0133382147,
0.0074131503,
0.0420060828,
0.0035571805,
0.004617699,
0.0000163071,
-0.0323776305,
0.0313309766,
0.0438474976,
-0.0207068939,
-0.0336809866,
0.0194152482,
0.0458162352,
-0.0791396126,
0.0236071553,
-0.0064770803,
-0.0066870358,
0.0612603724,
-0.0193298701,
0.0008519968,
0.0124385366,
-0.0014761346,
0.0345405415,
-0.0148735549,
0.0400237329,
0.0064140721,
0.0491387099,
-0.0397779755,
0.018612491,
-0.0756152794,
-0.0373238325,
-0.0219768044,
0.0246320385,
0.0271678977,
-0.001042323,
-0.0687073544,
-0.0392473936,
-0.0294710174,
0.031158004,
0.0374295078,
0.0575379357,
-0.0488210656,
0.012520398,
-0.0472501218,
0.0274335463,
-0.0166822262,
0.0097408015,
0.0200688392,
-0.0049786442,
0.0170266274,
0.0314222723,
-0.0232296791,
0.035241276,
-0.0051914281,
0.0322999321,
-0.0160804559,
0.0105733247,
0.0203913581,
-0.0955896378,
-0.0294783749,
-0.0615316741,
0.0325180702,
-0.0693421885,
0.0143404203,
0.0253769141,
0.0100189922,
-0.0310370382,
0.0023781778,
0.0230523702,
-0.0194570553,
0.0477375537,
-0.0684185028,
0.0157658681,
0.0231684539,
0.0436791703,
0.0118789915,
-0.0021054379,
0.0373734497,
0.0012203077,
-0.0177192427,
0.004231377,
0.0227535348,
-0.0338369645,
-0.0033433291,
-0.0372017249,
0.0033698922,
0.0169330612,
-0.0533701628,
0.0915127248,
-0.0110146264,
0.0648827553,
-0.0169244222,
0.0391905718,
0.019856628,
0.0536637306,
0.0293854401,
0.0245381761,
0.0812836811,
-0.0079996008,
0.0094680293,
0.0241776593,
-0.0126428874,
0.040010374,
0.0215903316,
0.008585508,
-0.0146766081,
-0.0147520797,
-0.018543046,
-0.0640309751,
-0.039589908,
0.0159732178,
-0.0194763746,
0.0039782836,
0.0125290873,
-0.0041022617,
-0.0468064584,
0.0364925601,
0.0265717041,
-0.0323588476,
0.0049214996,
-0.0465548225,
-0.0150660127,
0.1101522744,
0.0687027574,
-0.0266075265,
-0.0303841028,
-0.0311050806,
0.0393870957,
0.0577438958,
0.0337273479,
-0.0105214948,
-0.0106496839,
-0.0404616855,
-0.0353338234,
0.0107547669,
0.0225816835,
-0.0186326634,
-0.0061685396,
0.0138853854,
-0.0003177925,
-0.0350870602,
-0.0831770152,
0.0437222868,
-0.0436280258,
-0.0228692628,
-0.0214096047,
0.0037859594,
0.0186478719,
0.0257091597,
-0.0188651495,
0.0344750471,
-0.0091753192,
0.015579843,
-0.0429789573,
0.0004117498,
-0.0681756288,
-0.0278764572,
0.0174592584,
-0.0808349773,
-0.0198065341,
-0.006169037,
-0.0423080884,
-0.0179378521,
-0.048817832,
-0.0653335601,
0.019389553,
-0.0197337624,
0.0216986258,
0.0400277451,
-0.0331547149,
-0.0325101838,
0.0457561314,
-0.0230676737,
0.0109087862,
0.0366741717,
-0.0122650601,
0.0262840949,
-0.0283337682,
0.0010113559,
0.0055294959,
0.0229045656,
0.0224260427,
0.0035569037,
0.0448768586,
-0.0536956191,
-0.0268821884,
-0.1077544317,
-0.0273612496,
0.0035572122,
-0.0089592058,
0.0375694372,
-0.0478334762,
0.0105343303,
0.0233327579,
-0.0466297865,
-0.056275174,
0.0552808903,
0.0056036781,
0.0624356866,
0.0396819189,
0.0048994678,
-0.0402923003,
-0.0150981816,
-0.0594501868,
0.0012400275,
0.0879076794,
-0.0089645237,
-0.0348565988,
0.0302770454,
-0.0509086922,
-0.0386443101,
0.0115472097,
-0.0574591383,
0.0625165552,
-0.0143499533,
-0.0296176597,
0.0562163293,
0.0143457018,
0.0331284069,
0.0652140081,
-0.0053088767,
-0.0245603267,
0.0002186882,
-0.0108094523,
-0.0342857018,
0.0184599031,
-0.0660744011,
-0.0077264262,
0.0233276878,
0.0156200845,
0.0145972837,
-0.0434952825,
-0.0049570389,
-0.0357490294,
-0.0183819141,
-0.0161147192,
-0.0200947989,
0.0002120196,
0.0601999946,
0.0000780552,
-0.0342070274,
0.1121816263,
-0.0078938473,
0.0364204384,
-0.0947314501,
0.016903786,
-0.0487695485,
0.0415208712,
-0.021430755,
0.0349972323,
0.006648364,
-0.002418658,
-0.0363011137,
0.0114839729,
0.0391296335,
-0.0513048917,
0.0224764477,
-0.0508353561,
-0.0147769218,
0.0166131854,
-0.0211344827,
-0.0333203673,
-0.0081195757,
-0.0096576931,
0.0204839148,
0.013324433,
-0.0119197648,
-0.0517732464,
0.0203480031,
0.041385293,
0.0190565661,
0.0016316846,
-0.0431654975,
0.0102273058,
-0.0020856303,
0.0271156039,
-0.0022508122,
-0.003153204,
-0.0344598331,
0.0299521554,
0.0532877296,
-0.0031534287,
0.0379095785,
-0.0249646958,
-0.0050995522,
-0.0243716408,
-0.0181384645,
-0.0021793111,
0.0299882814,
0.0209798552,
-0.0570083857,
-0.0299840681,
0.1230992526,
0.0417667963,
-0.0560008027,
-0.0089681419,
0.0019180602,
-0.0044165412,
-0.0054611727,
0.017376896,
-0.0162160918,
-0.0222255606,
-0.024650408,
0.0458249561,
-0.0321141146,
-0.0270463899,
-0.0014690228,
-0.0706084222,
0.0503499135,
-0.0039841048,
-0.0321861245,
-0.0458805598,
-0.0372469872,
-0.0274253227,
0.0208749007,
-0.0295280386,
-0.0132180387,
-0.0151234716,
0.014881745,
0.0196528211,
-0.0838131681,
-0.0854604617,
0.0465636551,
-0.0368153974,
0.0249977391,
0.0702542588,
0.0567164347,
-0.0241973903,
-0.033504799,
-0.0389033854,
0.0313694216,
0.0311052091,
0.0104637919,
0.1035931185,
0.0067135687,
0.0317173153,
-0.0628853217,
0.0196960177,
0.00795596,
0.004952136,
-0.0032789321,
0.0720409453,
-0.0373824872,
-0.0079750819,
-0.0444466732,
0.0733861178,
0.016177468,
0.0651016012,
0.0148297278,
0.0062358561,
-0.0194983203,
-0.007300416,
0.060939882,
-0.0215284452,
-0.0334161818,
0.0342852473,
-0.0183621328,
-0.0419883616,
-0.08229509,
-0.0488271713,
0.0311858524,
0.0300067794,
0.0070419433,
0.0163730923,
-0.0102422107,
-0.005112994,
-0.0381024182,
0.0374652036,
0.0267797336,
0.0181094818,
-0.0359904282,
-0.0186696853,
0.0396151096,
-0.0068987533,
-0.0182349347,
0.0586096272,
-0.0292417891,
0.0150048705,
-0.0271699522,
0.0556644499,
0.0289525166,
-0.0012388622,
0.0045492705,
-0.0147495428,
0.043640364,
0.0627811849,
-0.0388445854,
-0.0153266927,
0.0089820074,
-0.002784485,
-0.0066492902,
0.0067020683,
-0.0286921635,
-0.0089428136,
0.0056624836,
0.0364066996,
-0.027064601,
-0.0064737238,
0.0517478101,
0.0400245748,
0.0164986514,
-0.0471690185,
0.0068394463,
-0.0119244643,
-0.0140547883,
-0.0540203191,
-0.0275177062,
-0.0321674496,
0.0266190656,
0.0309523121,
-0.006422332,
-0.0066082617,
0.0186870992,
-0.0638994649,
-0.0392955728,
-0.0291875582,
0.0279119257,
0.0167103726,
0.0070166113,
0.0071089901,
-0.0187222902,
-0.0549252853,
-0.0754554272,
-0.0106750671,
0.0181127749,
0.0170449279,
0.0941777155,
0.0292250626,
0.0411610827,
0.0088302875,
0.0498165973,
-0.0415823907,
-0.0459599607,
-0.0351522304,
0.0123765711,
0.0418784879,
0.0180262998,
0.0135935573,
-0.0679496676,
0.012250714,
-0.010206054,
0.0545997657,
-0.0033966706,
-0.0094756344,
0.0297221746,
0.0028422645,
0.0298891738,
-0.0300170518,
0.0798596591,
-0.0021089213,
0.0087889181,
-0.038890563,
-0.0828216523,
-0.0472285151,
-0.0504383668,
-0.0214996282,
0.007337667,
-0.0123737073,
0.0026714376,
-0.0701736435,
-0.0566009842,
-0.0184456855,
0.0012875251,
0.0104966182,
-0.0443766899,
0.0593278408,
0.0758033991,
-0.0235794205,
-0.034515705,
-0.0204319526,
0.0279939901,
-0.0122364294,
-0.0028768042,
-0.0192512143,
-5.729478485e-33,
0.0030145142,
-0.0692792609,
0.0467188805,
-0.0229695197,
0.0101181939,
0.0482475907,
-0.0502905697,
-0.0107866451,
0.0037299125,
0.0413955376,
-0.0142398933,
-0.0128019275,
0.0071154176,
0.0248271879,
0.0136964917,
-0.0121705839,
0.0263935421,
0.0027249798,
-0.0308302883,
0.0134659158,
-0.041436445,
0.0038912243,
0.0315520465,
-0.0950817987,
0.0327513069,
0.0026116113,
-0.0215298329,
-0.0293762013,
-0.0350464918,
0.0613560751,
-0.0018437083,
0.0103619574,
0.0482943021,
-0.0428852513,
0.0004384581,
-0.0088289259,
-0.0538149886,
-0.0204104241,
0.0347072743,
-0.0333168507,
-0.073842898,
-0.0049111862,
0.0478642657,
-0.0257409196,
-0.0158692729,
-0.0321077555,
0.0506921224,
-0.0075228801,
0.0209302958,
-0.034827631,
-0.057274688,
0.0446328335,
-0.0141479298,
-0.0242386889,
-0.0278572887,
0.0143062091,
-0.0387653373,
0.0422534309,
-0.0327808149,
-0.0067185615,
0.0335607827,
0.0092156604,
0.0545423217,
-0.0297046062,
-0.0342100561,
0.0364698097,
0.0250125919,
0.0008641596,
-0.0493308976,
0.0181497913,
-0.0484327972,
0.0478180423,
0.0782315433,
0.0435099639,
0.0244016293,
-0.001154419,
-0.0964202732,
0.0531443171,
-0.044990655,
0.079665795,
0.0196686778,
-0.0421830155,
-0.0891446397,
-0.0119492244,
-0.0141530018,
-0.0383013859,
-0.0129439458,
-0.043472372,
0.0422366858,
0.0120759383,
-0.0131437844,
0.0175067186,
0.0001236281,
0.0096857054,
0.0314802416,
-0.0290560313,
-0.0156443734,
0.024130052,
-0.0280431099,
0.0043441085,
-0.0463576578,
0.0292748511,
0.0312097091,
0.0378422402,
0.0654175133,
0.0122129684,
0.0523672774,
0.0069466061,
-0.0295061897,
0.0155419987,
-0.0157568119,
-0.0104257483,
-0.0339482389,
-0.0282965768,
0.0258583818,
0.0161989331,
-0.0302148089,
0.0421668142,
-0.0341555476,
0.0041789236,
-0.0277229566,
-0.0741204172,
-0.0054450468,
0.0015678863,
-0.0520528182,
-0.008927403,
-0.0131989494,
-0.015962733,
0.048092775,
-0.0154668549,
-0.0158084091,
0.0107223997,
2.765e-7,
0.0764631853,
0.0554600991,
0.0108719356,
0.0097767878,
0.0100617325,
0.004627367,
-0.0140428292,
-0.0030282347,
0.0282102693,
-0.0266561173,
0.0449692272,
0.0181251839,
0.0243375171,
0.0481385812,
-0.0597110502,
-0.0477830805,
-0.0544489995,
-0.0296828188,
-0.0340609103,
-0.0152903674,
0.1095057949,
0.0188948717,
-0.0295804366,
-0.0492915548,
-0.0033767163,
-0.0194903146,
0.0312597044,
-0.0110023953,
0.0148122469,
0.0197090786,
0.0551114976,
0.029911099,
0.0234071631,
0.077154018,
-0.0148769207,
-0.0447244868,
0.0331524014,
0.0801823959,
-0.0389017314,
-0.0015896006,
0.0230138563,
-0.0113700209,
0.0236540139,
-0.050623212,
0.0430620648,
-0.04564422,
-0.024655316,
0.0266610309,
-0.0328337923,
-0.0138837779,
0.0690923184,
-0.0298179258,
-0.0390643366,
0.0367515795,
0.0003517572,
0.0153871682,
-0.0259274673,
-0.0733033791,
-0.0113595277,
0.0277114697,
-0.0106229624,
-0.025380997,
0.027322717,
-0.0140355946,
0.0042692758,
-0.0764435902,
-0.0053834459,
2.495175252e-34,
0.0293884985,
-0.0047568684,
0.036543034,
-0.0291781407,
0.0636178181,
0.0335242637,
-0.0140562458,
0.0136911804,
0.0304887835,
-0.0404446349,
-0.0188151449
] |
http://arxiv.org/abs/2401.16577v1 | http://arxiv.org/pdf/2401.16577v1 | LLMs as On-demand Customizable Service | 2024-01-29T21:24:10 | 2024-01-29T21:24:10 | Souvika Sarkar, Mohammad Fakhruddin Babar, Monowar Hasan, Shubhra Kanti Karmaker | cs.CL | cs.CL, cs.AI | Large Language Models (LLMs) have demonstrated remarkable language
understanding and generation capabilities. However, training, deploying, and
accessing these models pose notable challenges, including resource-intensive
demands, extended training durations, and scalability issues. To address these
issues, we introduce a concept of hierarchical, distributed LLM architecture
that aims at enhancing the accessibility and deployability of LLMs across
heterogeneous computing platforms, including general-purpose computers (e.g.,
laptops) and IoT-style devices (e.g., embedded systems). By introducing a
"layered" approach, the proposed architecture enables on-demand accessibility
to LLMs as a customizable service. This approach also ensures optimal
trade-offs between the available computational resources and the user's
application needs. We envision that the concept of hierarchical LLM will
empower extensive, crowd-sourced user bases to harness the capabilities of
LLMs, thereby fostering advancements in AI technology in general. | LLMs as On-demand Customizable Service: Large Language Models (LLMs) have demonstrated remarkable language
understanding and generation capabilities. However, training, deploying, and
accessing these models pose notable challenges, including resource-intensive
demands, extended training durations, and scalability issues. To address these
issues, we introduce a concept of hierarchical, distributed LLM architecture
that aims at enhancing the accessibility and deployability of LLMs across
heterogeneous computing platforms, including general-purpose computers (e.g.,
laptops) and IoT-style devices (e.g., embedded systems). By introducing a
"layered" approach, the proposed architecture enables on-demand accessibility
to LLMs as a customizable service. This approach also ensures optimal
trade-offs between the available computational resources and the user's
application needs. We envision that the concept of hierarchical LLM will
empower extensive, crowd-sourced user bases to harness the capabilities of
LLMs, thereby fostering advancements in AI technology in general. | [
0.0224889088,
0.0199202392,
-0.0378918201,
0.0416223146,
-0.0321549661,
-0.0043683536,
0.0426122285,
-0.0137732504,
0.0317359902,
-0.0079615423,
-0.0362450853,
-0.0300135147,
-0.0089489799,
0.032303609,
0.0033878901,
-0.0373323448,
0.0406049713,
-0.0356686376,
-0.0192588698,
-0.0199382827,
-0.0009031558,
0.019877024,
0.0009119118,
0.0301738828,
0.0203596484,
-0.000080891,
-0.0336862393,
0.0269913822,
-0.0119442018,
-0.0247078408,
-0.0105369482,
0.018564133,
0.0585151352,
0.043252565,
0.0000020597,
-0.043718487,
-0.0294589121,
0.0234948788,
-0.0406031422,
0.0000463612,
0.0695411935,
0.0256708562,
-0.0355909392,
0.024863435,
-0.0368538387,
0.02720231,
0.0562577285,
-0.0162681732,
0.0580526777,
0.0420701094,
0.0099438662,
-0.0133580398,
-0.0054146848,
-0.0029734306,
-0.0287751611,
-0.0063270093,
0.0034973191,
-0.0199665893,
0.0493310057,
-0.0052663707,
0.0218779333,
0.0183255691,
0.0538452193,
0.0293842498,
0.018930167,
0.0300151296,
-0.0325011685,
-0.0140976412,
-0.057695806,
0.0420320034,
0.0005778226,
-0.0225502178,
0.0392817967,
0.0112877591,
-0.0271981321,
-0.0071461545,
-0.0261930935,
0.0319374204,
0.0063731298,
0.0212585237,
0.0252216216,
-0.0528627709,
0.0078760125,
0.025444068,
0.0036787265,
0.0945163891,
0.0219083596,
-0.0089617828,
-0.0424707383,
-0.0089055672,
0.032277029,
-0.0610312447,
0.0263528302,
0.0686487556,
-0.035814371,
0.0219772179,
-0.0780276656,
-0.0356726162,
-0.0081658252,
-0.0241805725,
0.071285747,
0.0092659453,
-0.0004946343,
0.0374304093,
0.0082640667,
0.0053785341,
-0.0249826573,
-0.0197444409,
-0.0359317437,
-0.0149100674,
-0.0468695201,
-0.0094428724,
-0.0551508963,
0.0408591963,
-0.0754958764,
-0.0238242578,
-0.0249802433,
0.0612554476,
-0.0099217994,
0.0087242797,
-0.0048242328,
0.0133885723,
-0.0176402628,
-0.0599787459,
-0.0059173992,
0.0036337744,
-0.009613255,
0.006767672,
0.0206162781,
-0.0958218053,
0.031193167,
0.0403636955,
-0.0237122942,
0.0293032695,
0.0058327224,
0.0895752534,
0.0002014326,
-0.0237764549,
-0.0479424782,
-0.0404432341,
0.0460381284,
-0.0615076981,
-0.0105280289,
-0.0853744149,
0.0215028413,
0.0045363451,
-0.0033608926,
-0.0114740366,
0.0034547036,
0.0213905443,
-0.006410969,
0.0258062314,
-0.0415074825,
0.0339503549,
0.0682047009,
0.007648685,
0.0099737635,
0.0394450389,
0.0688180998,
0.0285805129,
0.0328981802,
0.0065293512,
0.005417882,
0.0587875806,
-0.0208320841,
-0.0079127541,
-0.0325967707,
0.0190111306,
-0.0545151867,
0.03206167,
-0.0152254384,
0.0427610949,
-0.0044827228,
0.0262182187,
0.0808460265,
0.0810047239,
-0.0022763668,
0.0348367207,
0.0272068623,
0.0520351119,
-0.0055282549,
-0.0615155436,
-0.0504769385,
0.0363670215,
-0.0052481312,
0.0083390353,
0.0037344431,
0.0474087149,
-0.0085140746,
-0.0131244836,
-0.025755439,
-0.048615519,
-0.0395228043,
0.0138044935,
0.0591278486,
-0.0215538256,
-0.0068806461,
0.0530703962,
0.0282566529,
-0.0474111065,
0.0201061256,
-0.0785195008,
0.0335892737,
0.1269759685,
0.0481979847,
-0.0115178237,
0.0020276466,
0.0348812304,
-0.0102502108,
0.0553296059,
0.0225492064,
-0.0225090217,
-0.0123571297,
-0.0421468988,
-0.0116982684,
0.0355189368,
0.0039939499,
0.0269365292,
0.0068590371,
-0.0204741154,
-0.008968316,
-0.0270637255,
-0.0378688984,
0.0049619409,
-0.0130147524,
0.0068271938,
0.0036786732,
0.0209802035,
0.0410590135,
0.024952352,
-0.0114973187,
-0.0149804065,
0.0215723421,
-0.0431920849,
-0.024189055,
-0.0050952369,
-0.0273648165,
0.0763238743,
0.0246176179,
-0.0199311785,
0.014111653,
0.0144308591,
-0.024513416,
-0.03640908,
0.0396336466,
-0.0379198007,
0.039418105,
-0.0259079281,
0.0100537529,
-0.0107503748,
-0.0025303892,
-0.0122389793,
0.053324867,
-0.0084597096,
-0.0601148419,
0.0402157046,
0.0142551567,
0.011554596,
-0.0517957807,
-0.0050634136,
-0.0131856175,
0.0260338262,
0.0326737389,
-0.0126753747,
0.0075595174,
-0.0425267406,
0.0007218951,
-0.0982197374,
0.0090610553,
-0.0220413152,
0.0483565666,
0.0228693392,
0.0231684539,
0.0026534793,
-0.0273893122,
-0.0284154452,
0.0486931056,
0.0274407919,
0.0198814869,
0.0112799089,
0.0170392562,
-0.0515000857,
-0.0204968043,
0.007768563,
0.0244036205,
0.0011138145,
-0.0238867868,
0.0079562813,
-0.095191218,
-0.0234094914,
-0.0215466265,
0.0426681712,
0.0037427405,
-0.0255647246,
0.0325461514,
-0.0216751769,
0.0379587337,
0.045390632,
-0.0145489275,
0.0107613541,
0.0731432959,
-0.0075797294,
-0.059049353,
0.0106994761,
0.0083518354,
-0.0116262836,
-0.0366643779,
-0.022416871,
-0.0095854029,
0.0617920421,
-0.0194172524,
0.0085825482,
-0.0339415185,
0.0292230491,
-0.0194103755,
-0.0263348613,
-0.0107754674,
0.0091669289,
-0.0453731567,
0.0627617016,
-0.0095966505,
-0.0949045643,
0.0659977719,
-0.0502520278,
-0.001760238,
-0.0019230683,
-0.0748293549,
-0.0364199467,
0.0646903515,
-0.0051585739,
0.0351568162,
-0.0295371246,
0.006158405,
-0.028697731,
0.04074746,
-0.0153860617,
-0.0406665802,
0.0144248744,
-0.0912350789,
-0.0040195207,
-0.0060198009,
-0.0383225381,
-0.002322058,
0.0249107629,
-0.0667371675,
-0.007040651,
-0.0016483216,
-0.0381897949,
-0.0107914191,
0.0114295911,
0.0288806353,
0.0461818203,
0.0004523284,
-0.003781691,
-0.0361087322,
-0.0014835749,
0.0294565577,
0.0472408906,
-0.0009419608,
0.0059580472,
-0.000040917,
0.038017597,
0.0622865967,
-0.0377748609,
-0.028672507,
-0.0203660876,
0.0678472742,
-0.0212309137,
-0.018893538,
0.0428590551,
0.0332419947,
-0.0747760087,
-0.0205698516,
0.0997155905,
0.0034451096,
-0.0654674545,
0.0194978099,
-0.0200533979,
0.0425774194,
0.0313192457,
0.0575160608,
-0.0977528244,
-0.005704626,
0.0269805025,
0.0657224357,
-0.0181815568,
-0.0087153288,
-0.038246613,
-0.1021945029,
0.0407422744,
0.0221034773,
-0.0083903177,
-0.0161010344,
-0.0279896446,
0.0053194128,
0.0314086676,
0.0149611644,
0.019161161,
-0.0050299517,
-0.0026435119,
-0.0155008221,
-0.0423867032,
-0.0591468066,
0.0118371788,
-0.0157676917,
0.075890854,
0.0703287944,
-0.0073636021,
0.0166452732,
-0.0207688008,
0.0152195822,
0.0374026522,
0.0359171331,
-0.0401049368,
-0.0365829393,
0.0049293321,
-0.0104786567,
-0.0439698622,
0.0472386181,
-0.0379745588,
0.0057914043,
0.0010676594,
0.0959273726,
-0.0239681769,
0.0090472698,
-0.013128872,
-0.0709768906,
0.0260114726,
0.0611637123,
-0.0006127158,
-0.0521911122,
-0.0384401567,
-0.0235601962,
-0.0503206477,
-0.0056015644,
-0.0391619354,
0.0360177793,
-0.0544107817,
-0.0547642708,
-0.0505390912,
-0.0036441302,
0.0025295764,
0.0081048058,
0.0299262777,
-0.0201926008,
0.0467584655,
0.0134104434,
-0.0430536531,
0.0217453316,
-0.057931263,
0.0068884995,
-0.0864040032,
-0.0542398207,
0.0070947753,
-0.0776225403,
-0.0011450934,
0.0137160243,
-0.0392727107,
-0.0786823481,
0.0284120291,
0.0544575565,
0.0338049755,
0.0353266522,
-0.0017189257,
-0.087417759,
0.0577448346,
0.0336585343,
-0.0500594489,
0.0418570191,
-0.0022145568,
0.0059390655,
0.0104355616,
0.0402114242,
-0.0819195211,
0.015781777,
0.0019655253,
-0.0072671599,
-0.0193211231,
-0.0300431661,
0.0074811238,
0.0310513992,
-0.0079927221,
-0.0352993086,
-0.000292534,
-0.0198913887,
0.0473495424,
-0.0240325816,
-0.0016181224,
-0.0571904108,
-0.0222340785,
-0.0384155437,
-0.0257404279,
0.0397341326,
-0.0204434935,
0.0229438599,
-0.0022118092,
0.0018552297,
0.0530770272,
-0.0160071105,
0.0219522566,
0.0493403934,
0.0128247729,
-0.0140925311,
0.0240731966,
-0.0505923107,
0.0384049825,
-0.0305172373,
0.0394058451,
-0.0230942313,
-0.0293715075,
-0.0125553748,
0.0577243231,
-0.0136702815,
-0.1192570105,
0.0151419751,
0.0507545844,
0.0276100803,
0.0681773275,
0.0468227863,
-0.063455455,
0.0016930727,
-0.0372641273,
0.0131964767,
0.000867321,
-0.0540814698,
-0.0096842861,
0.0243678372,
-0.0094938576,
0.0133964382,
0.0492200553,
0.0230379719,
0.018581057,
-0.0230167266,
-0.0163241476,
-0.0488116071,
-0.0197998285,
-0.0363823101,
0.0184844192,
-0.0576811582,
-0.0196447521,
-0.0473093092,
-0.0513314791,
-0.0177716557,
0.0539236143,
-0.0079125408,
-0.0606071241,
0.0549895056,
0.0586154349,
-0.009200532,
-0.0266711321,
-0.0356644131,
0.0351490341,
-0.0506766625,
-0.0005356474,
0.0093988543,
-6.407341261e-33,
-0.0132312654,
-0.0292171128,
0.0674685389,
0.0076524895,
0.0056439326,
0.0072491099,
-0.024969494,
0.0238578711,
0.0122892233,
-0.0303730518,
-0.0041213008,
-0.0076453742,
0.0254184231,
0.0333370939,
0.0187857747,
0.0270553511,
0.0432056747,
-0.0161509048,
-0.0041592307,
0.0443514287,
0.019715732,
0.0238905046,
0.0293452907,
-0.0717238486,
0.0040921974,
-0.0368251503,
0.0225355197,
0.0205114,
-0.023409741,
0.0373318084,
-0.0173427537,
0.0436150804,
0.025392592,
0.0029051031,
-0.0114335883,
0.0108589362,
-0.0334815681,
0.0295360032,
0.0188388731,
0.0400697514,
-0.0417908542,
-0.0478321128,
0.0460224636,
0.0014035606,
-0.0339237824,
0.031394016,
0.0112256631,
-0.0160316546,
-0.0219790246,
0.008717875,
-0.0396268219,
0.0241999812,
0.0398560427,
0.0073383898,
0.0054754284,
-0.0187423136,
-0.0357258394,
0.0338970721,
-0.0488862246,
0.01656642,
-0.010008716,
0.0053866175,
0.09978237,
0.0038378653,
0.0022267853,
0.019744236,
0.0444024019,
-0.0237172544,
-0.0116072865,
-0.0038491529,
-0.0103724441,
0.0696459562,
0.0121748131,
0.0327566415,
0.0005885517,
-0.0128834015,
-0.0169941504,
-0.0170261804,
0.062733762,
0.0153350476,
0.0240866821,
0.029586995,
-0.005138706,
0.0195801109,
-0.0035713173,
-0.0342838503,
-0.0340188034,
-0.0373133346,
0.0073008011,
0.0235125106,
-0.0183370337,
-0.0559345745,
0.0171677787,
-0.010960537,
0.0049840682,
-0.0186790209,
0.0244138464,
0.0257966369,
0.0146833556,
0.0295149833,
-0.0508958176,
-0.0388049185,
-0.0557189137,
-0.0128439125,
0.0113962535,
0.0026970312,
-0.0155092608,
0.0396252759,
-0.0462158434,
0.0443203226,
-0.0352476463,
0.0324183591,
0.0115851955,
0.0201387368,
0.0425043404,
-0.0210350715,
-0.0006505293,
0.0575293005,
-0.0227847304,
-0.0132926041,
-0.0140860248,
0.0005208693,
-0.0061225579,
-0.0089007616,
-0.0280192662,
0.0317216367,
-0.0778219402,
0.0454348922,
0.0412828736,
-0.0296797082,
-0.0156661514,
0.0105249118,
2.986e-7,
-0.0142103825,
0.0332409739,
-0.0014747127,
0.0044634701,
0.0650688857,
0.0215415787,
0.0229027197,
0.0222630501,
-0.0233003516,
0.0040533193,
-0.0191252176,
0.0392544083,
0.0717506483,
0.0114391437,
-0.0674492642,
-0.0155806858,
-0.0360321105,
0.0424093753,
-0.0531290956,
0.0279994048,
0.0271459296,
0.057445664,
0.0128666786,
-0.0088448366,
0.0184499547,
-0.016475914,
0.0118029406,
-0.0289132781,
0.0118496176,
-0.0027667063,
-0.0092490222,
0.0629457831,
-0.0445619412,
-0.0059986641,
-0.0102067022,
0.0210639648,
-0.0822686329,
0.0700321645,
0.00125485,
-0.0003082105,
0.0381029397,
-0.0505761877,
-0.0511685982,
-0.0564114153,
0.0493374877,
-0.0304591209,
-0.0105039515,
-0.0565601476,
0.0279196631,
-0.0316348188,
-0.0601394698,
-0.0438347682,
0.0067030936,
-0.0044396762,
-0.0106172459,
-0.0776783973,
-0.0107944207,
-0.0647974536,
0.0241377521,
-0.0240767598,
-0.055982668,
-0.0472064875,
-0.0937941745,
0.0072986912,
-0.00180561,
-0.0039198562,
0.0174266584,
2.417467269e-34,
-0.0123793716,
0.0072589898,
0.0100289602,
0.0731744692,
0.0174806137,
0.0183032192,
-0.0160484668,
0.0002113053,
0.0073011182,
-0.0746921748,
-0.0320633538
] |
http://arxiv.org/abs/2401.16575v1 | http://arxiv.org/pdf/2401.16575v1 | Beyond Image-Text Matching: Verb Understanding in Multimodal Transformers Using Guided Masking | 2024-01-29T21:22:23 | 2024-01-29T21:22:23 | Ivana Beňová, Jana Košecká, Michal Gregor, Martin Tamajka, Marcel Veselý, Marián Šimko | cs.CL | cs.CL, cs.CV | The dominant probing approaches rely on the zero-shot performance of
image-text matching tasks to gain a finer-grained understanding of the
representations learned by recent multimodal image-language transformer models.
The evaluation is carried out on carefully curated datasets focusing on
counting, relations, attributes, and others. This work introduces an
alternative probing strategy called guided masking. The proposed approach
ablates different modalities using masking and assesses the model's ability to
predict the masked word with high accuracy. We focus on studying multimodal
models that consider regions of interest (ROI) features obtained by object
detectors as input tokens. We probe the understanding of verbs using guided
masking on ViLBERT, LXMERT, UNITER, and VisualBERT and show that these models
can predict the correct verb with high accuracy. This contrasts with previous
conclusions drawn from image-text matching probing techniques that frequently
fail in situations requiring verb understanding. The code for all experiments
will be publicly available https://github.com/ivana-13/guided_masking. | Beyond Image-Text Matching: Verb Understanding in Multimodal Transformers Using Guided Masking: The dominant probing approaches rely on the zero-shot performance of
image-text matching tasks to gain a finer-grained understanding of the
representations learned by recent multimodal image-language transformer models.
The evaluation is carried out on carefully curated datasets focusing on
counting, relations, attributes, and others. This work introduces an
alternative probing strategy called guided masking. The proposed approach
ablates different modalities using masking and assesses the model's ability to
predict the masked word with high accuracy. We focus on studying multimodal
models that consider regions of interest (ROI) features obtained by object
detectors as input tokens. We probe the understanding of verbs using guided
masking on ViLBERT, LXMERT, UNITER, and VisualBERT and show that these models
can predict the correct verb with high accuracy. This contrasts with previous
conclusions drawn from image-text matching probing techniques that frequently
fail in situations requiring verb understanding. The code for all experiments
will be publicly available https://github.com/ivana-13/guided_masking. | [
0.0117595308,
0.0299506988,
-0.023615865,
0.0925420225,
-0.0488662384,
-0.0074215266,
0.0018714606,
0.0384019203,
-0.0378854573,
-0.0948505327,
-0.0520977862,
-0.0546022281,
0.0124096908,
0.0384576358,
0.0291795097,
-0.0304953549,
0.0425881296,
-0.0169695299,
-0.0829490051,
0.0233367868,
0.0236668549,
0.0074378964,
-0.0138038453,
0.0514990315,
-0.0397636667,
-0.0028355783,
-0.0015890544,
-0.0011749413,
0.0025244197,
0.0010632627,
-0.0221794453,
0.020957496,
0.0385037325,
0.0228837691,
0.0000018423,
-0.0445340835,
-0.0246070679,
-0.0069499831,
0.0227153134,
-0.0536280721,
0.0600169264,
-0.074092187,
0.0303348638,
-0.0088866167,
-0.0231874492,
0.0168204904,
0.0705978125,
0.0357591361,
0.0137443263,
0.0401306637,
-0.038193386,
-0.0225676931,
0.0116875926,
0.0013598404,
0.0043247538,
-0.0009937438,
0.0008915112,
-0.0231430773,
-0.0194048714,
-0.0029160245,
-0.1072732508,
0.0423348136,
0.0013441421,
-0.0044619632,
-0.0186223965,
0.0388007797,
0.0044722361,
-0.0559671186,
-0.0322667994,
0.0500112623,
-0.0533869751,
-0.0137482779,
-0.0089635961,
-0.0043027257,
-0.0376981199,
0.0540133379,
-0.0192015003,
0.0359522626,
0.0017467008,
0.0217214096,
-0.0513141118,
-0.0218496826,
-0.0132157495,
-0.006398032,
-0.010914498,
-0.0033852435,
0.0404626019,
0.0218635462,
-0.013351609,
0.010310594,
-0.0379958116,
-0.0426683389,
0.0009391223,
0.0085423859,
0.0038131657,
0.0142051745,
-0.0328619368,
-0.0579727516,
-0.0163007211,
0.0108991927,
0.0261419024,
0.0544627085,
-0.0756135508,
0.048796922,
-0.04311141,
0.0608404092,
-0.015450404,
0.075740248,
-0.0378821865,
-0.0263125505,
-0.0419916324,
-0.0002953773,
-0.0337795876,
-0.0179170445,
-0.0391116329,
-0.0589242578,
-0.0008476705,
-0.0195701178,
0.0200192742,
-0.0289196018,
-0.0162268318,
-0.0034870256,
-0.0368313827,
-0.011020422,
-0.0178084429,
0.0612820908,
0.0160136428,
0.041731149,
-0.0422372557,
-0.0639741421,
-0.0137316929,
0.0569675826,
0.0238792859,
-0.0544649735,
-0.0203782972,
0.0687832385,
-0.0652559176,
-0.0857667401,
0.0196469724,
-0.043247208,
-0.0463833064,
-0.0288756657,
0.0664767846,
-0.0160635803,
0.0350966416,
0.0069338353,
0.0341047682,
-0.0258917902,
0.0063956617,
0.0122592989,
-0.0021806033,
0.0996614024,
0.0477660038,
-0.0091081904,
-0.0275461432,
0.0187396351,
-0.0416667275,
0.0490666963,
-0.0041106879,
0.0487778597,
0.0388357639,
0.0012877306,
0.0379354022,
0.0589060858,
-0.0088793226,
-0.0119780572,
-0.0289550927,
-0.031280376,
0.0343041718,
0.0024468168,
0.0693231896,
0.0338136032,
-0.0273773503,
-0.0422239155,
0.0940124691,
0.0563769527,
-0.0182392001,
-0.0564492084,
0.0471937582,
-0.0077046743,
-0.0016810009,
0.0001211219,
-0.0328501947,
0.0452367663,
-0.0314953029,
-0.0290343352,
-0.0525691435,
0.0464663431,
-0.0215480812,
0.0287566949,
-0.0440318175,
-0.0129643735,
0.0243338346,
0.01586424,
0.069417119,
0.0418476015,
-0.0394153073,
0.0219376069,
-0.0104934508,
-0.0222573131,
0.0004386141,
-0.0549620502,
0.0147724478,
-0.0192424338,
0.0400335379,
0.0223444197,
0.0202531479,
0.0058446145,
0.0305362865,
0.0404189192,
-0.0181717221,
0.0337442458,
0.0449657477,
-0.0525031015,
-0.0329119004,
-0.0095332311,
-0.0045935083,
-0.0191902518,
-0.0077616535,
-0.0200701281,
-0.0117574381,
-0.0444055237,
0.0591440201,
0.0699382797,
0.0254897103,
0.0005272412,
-0.0547822267,
0.0173515435,
0.0019858393,
-0.0389062651,
0.0069406996,
0.005935905,
0.0058069895,
-0.0241104849,
0.0082302075,
0.0367804617,
0.0108470917,
-0.0585200526,
0.0114491936,
-0.0602253824,
-0.0208404921,
0.0584818944,
0.0045114215,
-0.021865448,
-0.0058015282,
-0.0518085361,
0.0291487239,
-0.0369983986,
-0.0032430333,
0.0320012681,
0.0516119339,
-0.0141602485,
0.0250382256,
-0.0181791428,
-0.00172062,
-0.0142478608,
-0.0272591598,
0.0300103128,
0.0107126683,
-0.0454702489,
0.0032253659,
-0.025788337,
-0.0266984999,
0.0041965721,
-0.0181427542,
-0.0424494632,
-0.0137981372,
-0.064590469,
-0.0220658779,
-0.0308909807,
-0.025376616,
0.0067441165,
-0.0087916208,
0.0263948012,
-0.0451146513,
-0.0127910525,
0.0438475013,
0.0072972612,
0.0263953544,
0.0132508297,
-0.0164868962,
-0.0179101061,
-0.0312904641,
0.0020583235,
-0.0142538734,
0.0764198676,
-0.0087832715,
-0.0272408724,
0.0016447551,
0.0130287502,
-0.0059168041,
0.0186963584,
-0.0234069284,
-0.0332785472,
0.0195323769,
0.0030697954,
0.0602385998,
0.0438938253,
-0.0086693764,
0.0391636267,
0.0334692188,
-0.0152902082,
-0.0275952294,
-0.0352228358,
-0.0183496345,
-0.039132569,
-0.0108007677,
-0.0158839561,
0.0056186174,
0.0692608431,
0.0384916849,
0.0109237144,
-0.0513615459,
-0.0033628999,
-0.0253228284,
-0.0511648804,
-0.0511805005,
0.0299751963,
-0.0116276555,
0.025853809,
-0.0605610609,
-0.0382053368,
0.0378667302,
-0.0366048291,
-0.0443111733,
-0.0193325561,
0.0169060417,
-0.0018365021,
0.0893081576,
0.0574418344,
0.0348575301,
0.023760939,
0.0048348894,
-0.0708929598,
-0.0083709368,
0.0165989064,
-0.0083521232,
0.0708626583,
-0.0476034135,
0.0234252736,
-0.0141844684,
-0.0097049512,
-0.0157263819,
-0.0621636212,
-0.0214330014,
-0.0197873004,
-0.013425733,
0.0001799106,
-0.0108775487,
0.0301010385,
0.0416601598,
0.0238590539,
0.0122402804,
-0.0269744862,
-0.0200499501,
0.0484256558,
0.0484744534,
0.0025400305,
-0.0297076497,
-0.0486275554,
-0.0118994638,
0.059804529,
0.0023791767,
0.0053178994,
-0.0503728874,
0.0456830859,
0.0546400845,
0.0470782369,
-0.0084356638,
0.0102828061,
0.0866599753,
-0.0119831413,
-0.0115337064,
0.0335701779,
0.0183427203,
-0.0298034009,
0.0125644207,
-0.0037606433,
0.0138668241,
0.034361057,
-0.0131144663,
-0.0976110399,
-0.0743756071,
-0.0044491449,
-0.0350506753,
0.0353663303,
0.0083831307,
-0.0223867111,
-0.0858618543,
0.0398686491,
-0.0337153114,
-0.0303641167,
-0.0138698528,
0.0119241243,
0.0331292115,
-0.0320484117,
0.0098800613,
-0.022534458,
-0.0062244055,
-0.0051255273,
0.0098609589,
0.0361342914,
-0.0429588892,
0.02929819,
-0.04793166,
-0.0029802106,
0.01875256,
0.0283506606,
0.0087045096,
-0.0695594028,
0.0156866629,
0.0259711202,
0.0226672888,
0.0160347484,
0.0637372434,
-0.0174535979,
-0.0422771983,
0.0269773733,
0.0112682283,
0.0087771025,
-0.0783073604,
-0.0567527637,
0.0230651144,
-0.0089837853,
-0.0060630962,
-0.0254111513,
0.0019134511,
-0.0244012345,
0.0516348183,
0.0211887117,
-0.0183566697,
0.0105987927,
0.0160039347,
0.0257681832,
-0.0278548058,
-0.0565018803,
0.0045686825,
-0.0417611375,
0.0113745695,
-0.0487258472,
-0.003996341,
0.0521814711,
0.0249142982,
-0.00688475,
-0.0169974025,
-0.0143481242,
0.0189939458,
-0.0450499207,
0.0786050931,
-0.0392525233,
0.0787578821,
-0.0137721421,
0.0173215102,
0.0112738861,
-0.0251883306,
-0.0111290254,
0.0474953204,
-0.0864462256,
-0.0232150592,
0.0084481463,
0.0375904441,
0.0179835446,
-0.0019553211,
0.0182227436,
-0.0299438145,
-0.0490596667,
-0.0202836599,
-0.0124976197,
0.0442564152,
-0.0021658735,
-0.0339021869,
-0.0166523252,
0.0193914622,
-0.0277601127,
0.0381143056,
0.0069336644,
0.0201091431,
-0.0199839808,
0.0040038018,
0.0345868729,
-0.022168057,
0.0805510804,
-0.0283562429,
0.0530955307,
-0.0312683955,
-0.0253869426,
0.017080823,
0.0070411465,
-0.0646938682,
-0.002307514,
0.0294471569,
0.024341831,
-0.01295208,
0.0344075523,
-0.0081968401,
0.0314093754,
-0.0218260773,
-0.0124981627,
0.0328878425,
-0.0295261927,
-0.0153409708,
0.0580078363,
0.0161159746,
0.01389011,
-0.0105620809,
0.0184894092,
0.0089293579,
0.0827224702,
0.0268645547,
-0.0248496514,
-0.0197285693,
0.0144042037,
-0.0070809666,
0.013420077,
-0.0213176981,
-0.0171050709,
0.0322000869,
-0.0357349999,
0.0248995051,
-0.0196963008,
-0.0080461511,
0.0093879541,
0.0448444076,
-0.0351324454,
-0.0099395029,
0.0900344998,
-0.0046289251,
-0.0042166258,
-0.0312061105,
0.0649679899,
-0.0283832029,
0.040123295,
0.0063922922,
-0.0051902225,
-0.0282893665,
0.0113698021,
0.0949849412,
0.0034035402,
0.011126793,
0.0046358258,
-0.0346791483,
-0.0502154045,
-0.0140273077,
-0.0121327452,
0.0003889485,
0.0118034445,
0.0401359163,
0.0376633219,
0.0167615376,
-0.0105932169,
-0.0409307815,
-0.0220243223,
0.0029126352,
0.0364315584,
0.0006613871,
-5.868582078e-33,
-0.0420195572,
-0.0305104572,
0.01287656,
-0.0015133158,
-0.0036402806,
0.0334256589,
-0.0285385419,
-0.0004351256,
-0.0154170599,
0.0293910857,
0.0493016019,
-0.0097644841,
0.0135449534,
0.0301480647,
0.0091330176,
-0.001160663,
0.0220645908,
-0.0138873952,
-0.0269337986,
-0.0577934682,
0.0223465227,
0.0167854261,
0.076079756,
-0.0416153967,
-0.0183046553,
0.0476407185,
-0.0561391786,
0.018664645,
0.0044347984,
0.0707293823,
-0.0459464677,
0.0239361301,
0.0034976925,
-0.0405793525,
-0.0350260399,
0.0428077616,
-0.0314251333,
0.003341157,
0.0336023048,
0.0445704125,
-0.0683481097,
0.0350897424,
0.0446251146,
-0.0427880138,
0.0101863919,
0.0014892193,
0.0409771502,
0.0103384834,
0.0277378522,
-0.009895674,
-0.0680935457,
0.0030863283,
0.0141801573,
-0.0558951609,
0.0475456417,
0.0426250659,
0.0386388935,
0.0190237984,
-0.0582306832,
0.0123674301,
-0.0419344157,
0.0446114428,
0.0811374635,
0.0510482006,
-0.0207006298,
0.071108073,
0.0310280137,
-0.049139034,
0.0094836447,
0.0066842344,
0.0531983972,
0.0661394075,
0.0464252941,
0.0671416298,
0.0262928512,
0.0255272221,
0.0127851767,
0.0022838598,
-0.0359101258,
0.0464114137,
0.0182013065,
0.0334548578,
0.0016962263,
0.0164259952,
-0.0446765013,
0.0071814149,
-0.0562105924,
0.0114063593,
0.0618271641,
0.020814281,
0.0074541816,
0.0382304899,
0.0982219353,
-0.0302693453,
-0.0084982738,
-0.0233945958,
-0.0366587564,
0.0737847909,
0.003974061,
0.0277877972,
-0.0302278101,
-0.046460703,
0.0768154934,
0.0018139592,
0.0237064324,
0.0112974253,
0.0194536895,
0.0367027409,
-0.0428498685,
-0.0187221412,
-0.0546062738,
-0.0371860005,
-0.0175779369,
0.0136048095,
0.0265615378,
0.033384338,
-0.0734893531,
-0.0014096692,
0.0094270622,
-0.0288820025,
-0.0460656211,
-0.0491592102,
-0.0190906916,
0.0144779943,
-0.0170616973,
0.0177686792,
-0.0308702365,
-0.0374588743,
0.0617604591,
-0.0418900549,
-0.0029415879,
0.0363524929,
2.6e-7,
0.0818054229,
0.0301037934,
-0.0126649328,
-0.0014671237,
-0.0207275283,
0.0434852652,
-0.0221629329,
0.0703574196,
-0.0084477579,
-0.0465976186,
-0.0332862996,
0.0175547116,
0.0320354812,
-0.002937376,
0.0018049793,
-0.0073454678,
0.0073224814,
-0.0159155838,
-0.0344532356,
0.0088116182,
0.0636097342,
0.0851195306,
-0.006631799,
-0.0098469574,
-0.0020332078,
-0.0154020339,
0.0914178863,
-0.0288621951,
-0.0187670887,
-0.0234558862,
0.0308899879,
-0.0107393255,
0.0069643976,
0.0304730348,
-0.0143823177,
0.0029673802,
0.0597649701,
0.0006176296,
-0.0545323305,
-0.0283881389,
-0.0421982706,
0.0031176191,
-0.0464457087,
-0.0496888272,
0.06523332,
-0.0644454807,
-0.0118696745,
-0.0218220521,
-0.0118557047,
0.0353290476,
0.0175162628,
0.007789663,
-0.0421810448,
0.0294040106,
0.0560398549,
0.0200642552,
-0.0845603272,
-0.0762966201,
-0.0473070294,
0.0051157707,
-0.0433283523,
-0.0047561903,
-0.0548810214,
-0.0383551568,
0.0238113292,
-0.0681971163,
-0.0937924609,
2.051399043e-34,
-0.0158690736,
-0.0135150552,
0.0075413701,
0.0348296501,
0.0203329884,
-0.0250230022,
-0.0114381416,
0.0271132365,
-0.0214123875,
-0.0461608842,
-0.0113160769
] |
http://arxiv.org/abs/2401.16569v1 | http://arxiv.org/pdf/2401.16569v1 | Autoencoder-Based Domain Learning for Semantic Communication with Conceptual Spaces | 2024-01-29T21:08:33 | 2024-01-29T21:08:33 | Dylan Wheeler, Balasubramaniam Natarajan | cs.LG | cs.LG, cs.AI, cs.CL, eess.IV | Communication with the goal of accurately conveying meaning, rather than
accurately transmitting symbols, has become an area of growing interest. This
paradigm, termed semantic communication, typically leverages modern
developments in artificial intelligence and machine learning to improve the
efficiency and robustness of communication systems. However, a standard model
for capturing and quantifying the details of "meaning" is lacking, with many
leading approaches to semantic communication adopting a black-box framework
with little understanding of what exactly the model is learning. One solution
is to utilize the conceptual spaces framework, which models meaning explicitly
in a geometric manner. Though prior work studying semantic communication with
conceptual spaces has shown promising results, these previous attempts involve
hand-crafting a conceptual space model, severely limiting the scalability and
practicality of the approach. In this work, we develop a framework for learning
a domain of a conceptual space model using only the raw data with high-level
property labels. In experiments using the MNIST and CelebA datasets, we show
that the domains learned using the framework maintain semantic similarity
relations and possess interpretable dimensions. | Autoencoder-Based Domain Learning for Semantic Communication with Conceptual Spaces: Communication with the goal of accurately conveying meaning, rather than
accurately transmitting symbols, has become an area of growing interest. This
paradigm, termed semantic communication, typically leverages modern
developments in artificial intelligence and machine learning to improve the
efficiency and robustness of communication systems. However, a standard model
for capturing and quantifying the details of "meaning" is lacking, with many
leading approaches to semantic communication adopting a black-box framework
with little understanding of what exactly the model is learning. One solution
is to utilize the conceptual spaces framework, which models meaning explicitly
in a geometric manner. Though prior work studying semantic communication with
conceptual spaces has shown promising results, these previous attempts involve
hand-crafting a conceptual space model, severely limiting the scalability and
practicality of the approach. In this work, we develop a framework for learning
a domain of a conceptual space model using only the raw data with high-level
property labels. In experiments using the MNIST and CelebA datasets, we show
that the domains learned using the framework maintain semantic similarity
relations and possess interpretable dimensions. | [
0.0111331837,
0.0588651821,
-0.0437348969,
0.0460770838,
-0.0188777614,
0.0273757875,
-0.0084710149,
-0.0057579884,
-0.0722800195,
-0.0364912003,
-0.0208841395,
-0.0235300865,
0.023707686,
0.0748724788,
0.0177901164,
-0.0701475441,
-0.0171548706,
0.0157237798,
-0.0229467526,
-0.0062916316,
-0.0359204859,
-0.006692146,
-0.0175928529,
0.0540890507,
-0.0537178181,
0.010038672,
-0.0400775783,
-0.015545913,
0.0033417356,
-0.0535295606,
-0.0420931578,
0.0020301773,
0.0847051218,
-0.0164104309,
0.0000022221,
-0.023732109,
0.0036067346,
-0.0251432806,
-0.0359845161,
-0.0518390387,
0.0226027668,
-0.0722342879,
0.0207166802,
-0.0131355841,
-0.0323105119,
-0.0735853463,
0.0507735461,
0.0357188471,
0.025283413,
0.0213032458,
-0.0272371564,
-0.030000614,
0.0447943173,
-0.0516551286,
0.0195932686,
-0.0609179735,
0.0447611548,
0.0204408653,
-0.0472956263,
0.0078209732,
-0.0677917525,
0.021767281,
0.0271437112,
0.0095784292,
0.0264550149,
0.0298373923,
0.0247732755,
-0.032230176,
-0.0000486333,
0.0632582828,
0.0320307203,
0.0174808279,
0.0078072972,
-0.0151394969,
-0.0184465777,
0.0155236851,
-0.048169788,
-0.008495572,
0.0187838953,
-0.0154852672,
-0.0683483332,
0.069338426,
0.0235524084,
-0.0298503805,
-0.0501063466,
0.0591797456,
-0.0033368007,
-0.0040484359,
-0.0189487115,
-0.0306653343,
-0.0112083275,
-0.003173298,
-0.0287882797,
0.0117065161,
0.0179477725,
0.0112792701,
-0.0232142154,
-0.0103744613,
-0.0061855195,
-0.0246520527,
0.0337146707,
0.0177826006,
0.0260222889,
0.0038046613,
0.0004833639,
0.0182148125,
-0.0207219981,
0.0307671148,
-0.0379996523,
-0.0335159823,
-0.0640705675,
-0.0214806609,
-0.0302289017,
0.0369956382,
0.0087317722,
-0.0691199377,
-0.0317913741,
0.0463290401,
-0.0369253792,
0.0235420782,
-0.0375533663,
0.0095000491,
-0.0624864921,
0.0295017697,
-0.0067454986,
0.0015826368,
-0.0531960391,
0.0334084071,
-0.019873824,
-0.0343073569,
0.0224924888,
0.0259829871,
0.0302720685,
-0.0139004467,
-0.0321974121,
0.0985877886,
0.0606659949,
-0.0102488454,
-0.0007309152,
-0.0408340506,
-0.0265543405,
0.0046533337,
-0.0121755823,
-0.032321427,
0.0187364314,
-0.0059451847,
0.000119702,
-0.0998162031,
-0.0207833294,
-0.0637518913,
0.0037401246,
0.1136046946,
-0.0103531405,
0.0111880843,
0.0775478557,
0.0187660344,
-0.0324837603,
0.0159805268,
0.0397449732,
0.0476536863,
0.028724134,
0.0088340361,
-0.0297033675,
-0.0010223506,
0.0190422665,
-0.0254174378,
-0.0448205061,
-0.0113835856,
-0.0055415058,
0.0744739175,
0.0055753505,
-0.0037659435,
-0.0370657258,
-0.0328499489,
-0.0050875205,
0.0533079989,
-0.0011364407,
-0.0025257962,
0.071032472,
0.0343302451,
-0.0519256741,
-0.029271204,
-0.0097712623,
0.011187694,
0.0296466947,
0.0201465748,
0.0324784629,
-0.0227512568,
-0.0383977443,
-0.0495400056,
0.0342231467,
-0.0539176352,
0.0048942696,
-0.0029231487,
0.0986836702,
0.0137135871,
0.0145093566,
0.0577203035,
0.0479503237,
-0.0049402793,
0.0330400951,
-0.0026185922,
0.0101013416,
0.0178272948,
0.0208958127,
-0.0025445558,
0.0173828192,
-0.0476735905,
0.0458768867,
0.0401552804,
-0.0117245968,
-0.0273089595,
0.0267348718,
-0.0372240506,
-0.0584839061,
-0.0412932299,
0.0195218083,
0.0190165993,
-0.0255355742,
0.019234268,
0.0348322839,
-0.0387836583,
0.0311675426,
0.021128688,
0.0274994187,
0.0115609178,
0.0211443305,
0.0450260788,
0.0339239351,
0.02820432,
-0.0195682887,
-0.0164799541,
0.0178023055,
-0.0560838245,
-0.0486534946,
-0.0049782693,
-0.0394250676,
-0.0127907209,
0.0119091747,
-0.0539878309,
-0.0258996412,
0.0205766056,
-0.01983013,
0.0075324993,
0.0793086514,
-0.0294802021,
0.0031066602,
-0.0158624034,
-0.0325665623,
-0.0208215211,
0.001477483,
0.000370953,
0.0339754596,
-0.008020387,
-0.0142130284,
-0.0168439802,
0.0464479662,
0.0592364669,
0.0034249879,
0.0216184463,
-0.0111981984,
0.0029460965,
-0.0103769433,
0.0379757546,
-0.0160617549,
0.0029928561,
-0.0065701716,
0.0115962708,
-0.0371516235,
-0.0231068749,
-0.0131858028,
0.0181530453,
0.0388518833,
0.016465731,
-0.0065775272,
0.006030031,
0.0390184559,
-0.0451368988,
-0.0006838526,
0.063437447,
0.0206148792,
-0.0323987231,
0.0149595607,
-0.0242180862,
-0.0075973524,
0.0834577456,
0.0316111818,
-0.0064060525,
-0.0033616526,
0.0243208259,
-0.040841125,
0.0098088235,
-0.014909328,
-0.0176635813,
0.0303977411,
-0.0104818763,
-0.0425723307,
0.013139477,
0.0250501148,
0.0962597877,
0.037091136,
0.0268209595,
-0.0370262526,
-0.0250578877,
-0.084367618,
0.019842932,
-0.0380940959,
-0.024858987,
0.0377568491,
0.0333497636,
0.0238077473,
0.0272099469,
-0.0571252145,
0.0215823911,
-0.0040508686,
-0.0298082754,
-0.0575441234,
0.0057060667,
-0.0031232713,
0.0828993544,
0.000798404,
-0.0211526137,
0.0577372313,
0.0072808242,
0.0252342429,
-0.0144766187,
-0.0128885908,
-0.0232275352,
0.0487884767,
0.0035510208,
0.0348585211,
0.046700526,
-0.0273536798,
0.0164379068,
0.0178738236,
0.0310418885,
0.0366329513,
0.0056067957,
-0.0322394222,
-0.0027530356,
0.0565517433,
-0.056062568,
-0.0341991447,
0.0455799885,
-0.0079028849,
-0.0093245963,
-0.027747089,
-0.00600554,
-0.036635872,
0.0090775341,
-0.0278650075,
-0.0059817368,
0.0167419221,
-0.0444499105,
-0.0363776013,
0.0358027406,
0.0711100772,
0.0011532301,
-0.0337112881,
-0.0211746488,
-0.0317107029,
0.0159510635,
-0.0454961769,
0.023157971,
-0.036314968,
0.0531640463,
0.0086903227,
0.0261853766,
0.0094164452,
0.0296277888,
-0.0012916873,
0.0124736959,
-0.0617477857,
0.1204994395,
0.0342243463,
0.00503489,
0.0138788801,
-0.0309946984,
0.0044465708,
0.0041109673,
-0.0174161438,
-0.053226836,
-0.0430174135,
-0.0178871192,
-0.0042242231,
0.0167249534,
-0.0065903636,
0.0143453078,
-0.0527749211,
-0.0024561225,
-0.003616164,
-0.0886594504,
-0.0746252909,
-0.0400363281,
0.0032673522,
-0.0235427096,
-0.0342076048,
-0.0793213844,
0.0472173691,
-0.0092626596,
0.0082622943,
-0.0264958683,
-0.0667569712,
0.0059455736,
-0.0487720966,
0.0403727256,
0.043083705,
0.0290400814,
0.0000327853,
-0.051268857,
-0.0344707891,
-0.0187430847,
0.0040229126,
-0.0302682351,
0.0671135113,
-0.0313689448,
0.0091240667,
0.0366433375,
0.0362955295,
-0.0825660899,
0.0032251407,
-0.0090627139,
0.0309447534,
-0.033317294,
-0.0152086606,
-0.0309316628,
-0.0041557206,
0.0284870714,
0.0409736708,
0.0553110018,
0.0024521283,
-0.0217756387,
-0.034872625,
0.0313660949,
0.0100879846,
-0.0049350299,
-0.041959729,
0.032023482,
-0.0505378768,
-0.0214407798,
-0.0024621454,
0.0120546464,
-0.0019963456,
0.0116804205,
-0.0287541971,
0.0316376388,
-0.0241387431,
-0.0952222794,
-0.0040747765,
0.0107094925,
0.0134779075,
0.0016654286,
-0.0241740886,
0.0552990809,
-0.0066466462,
-0.0429464318,
0.00953025,
-0.0240457021,
-0.0248665977,
-0.0094150566,
0.0146804284,
0.0206959862,
0.0508865602,
0.0009256701,
0.0020470582,
0.030572176,
0.0234706905,
-0.0518884473,
0.0014577305,
0.0258298889,
0.0075462088,
0.0036903988,
0.0432387367,
-0.0531417839,
0.0668531805,
-0.012441162,
-0.0054851016,
-0.0216083415,
-0.0466146283,
0.0025632267,
-0.0371021815,
0.0426439233,
0.0131480228,
-0.0627501085,
-0.0208908003,
-0.0229380522,
-0.0113895386,
0.0110974712,
-0.0423433445,
-0.004911398,
-0.0000063138,
0.0336744376,
0.007476327,
-0.0140995346,
0.0111906724,
0.0187500138,
0.0256328434,
-0.0214069039,
0.020123817,
0.032255467,
0.0051406622,
0.0089036459,
0.0063136565,
-0.0032720438,
0.0137978857,
-0.0007763106,
-0.0058726207,
0.0165121686,
0.0490812026,
0.0288646519,
-0.0229270235,
0.0579879545,
-0.0080313217,
-0.03983647,
-0.0045002783,
0.0209880471,
0.0126800202,
0.0295289233,
-0.0008311679,
-0.0444717407,
-0.003315516,
-0.0155177796,
0.0649508685,
-0.0093083149,
-0.0596725643,
0.0750036761,
0.0070465398,
-0.0101596834,
-0.0457254685,
0.0499727204,
0.0038665941,
0.0353289284,
-0.0285754614,
0.0070777433,
-0.0355364121,
0.0283140484,
0.0304040015,
0.0454182141,
-0.0005354962,
0.060677018,
-0.0547681414,
-0.0663212836,
0.0300697573,
-0.0142350839,
0.0011807188,
0.0298652686,
0.0839754045,
0.0440696962,
-0.0017212315,
-0.0164007302,
-0.0083564101,
-0.0204700064,
-0.011066746,
-0.0279430449,
-0.0232774056,
-6.027876991e-33,
-0.0061922455,
-0.03766625,
-0.0205307212,
0.0550598055,
-0.003040164,
-0.0134614669,
-0.0371432751,
-0.0321531817,
0.0125208972,
0.0384002067,
0.0547241084,
0.0107713481,
-0.0092927227,
0.0331489779,
0.0302159488,
-0.0310181007,
0.0500248671,
-0.00456492,
-0.0256902594,
0.0201997627,
0.0267152619,
0.030843446,
-0.0082036294,
-0.115129441,
-0.0120230746,
0.0309620295,
0.0467955321,
0.024533134,
0.0850823,
0.024797542,
-0.0311111528,
0.0767614469,
0.0140451202,
-0.0294714589,
-0.0098249083,
0.0115979882,
-0.0967996046,
-0.0747714117,
-0.058284834,
0.0133769941,
-0.0320731513,
-0.0200187564,
0.0506048948,
0.0010405169,
0.0098385941,
-0.0389104448,
0.0691541284,
-0.0192747191,
-0.0279823486,
-0.0280416608,
-0.0342443213,
0.0149611589,
-0.0298968703,
-0.037591286,
0.05993836,
-0.0256371927,
0.0379381776,
0.0081221564,
-0.0562801398,
0.0649839267,
-0.039213378,
-0.024165431,
0.0339036137,
-0.0167731848,
0.0266588628,
0.0568086877,
0.0326602906,
-0.0436848216,
-0.0041753957,
0.0030500048,
0.0050455858,
0.0142255491,
-0.0141809694,
0.0587691516,
0.0314007141,
-0.0585340783,
-0.0686582401,
0.0344391987,
-0.0527011603,
0.0397705883,
-0.0228404496,
0.0091137914,
-0.0509155802,
-0.036764387,
-0.014443648,
0.0098972302,
-0.0164927561,
-0.0336499549,
0.0174230356,
-0.0473195538,
-0.0148082571,
0.0339053087,
-0.0121442787,
-0.0735636353,
-0.0451869667,
0.0141216842,
0.0412913673,
0.0356339887,
0.0280416124,
-0.0159148648,
0.0045431638,
-0.0326177999,
0.0132531673,
0.025306467,
0.0282937158,
0.0103333937,
0.0300638452,
0.0244551413,
0.0018647959,
-0.0385969132,
-0.0120249102,
-0.0623112321,
0.0148409903,
-0.051755663,
-0.0265557859,
0.0455302298,
-0.0378133915,
0.0271398779,
-0.0140253585,
-0.0421644635,
0.0075786621,
-0.0066419486,
0.0199084468,
0.0285019297,
-0.0811628401,
-0.0483097546,
-0.0418079793,
-0.0352667049,
0.1030122787,
-0.0488211699,
-0.0439524315,
-0.0548664592,
2.941e-7,
0.051293388,
0.0068888171,
0.0444441065,
0.0488470607,
0.0172868315,
0.0546654835,
-0.0055758501,
0.0596298538,
-0.0343878493,
-0.0090505183,
0.0400408357,
0.0479175337,
0.0030579425,
-0.0235753171,
-0.0598664321,
0.0561924316,
0.0253517702,
-0.0187693425,
-0.0177992731,
0.0343412608,
0.0801220536,
0.05521027,
0.0344618,
0.0023503718,
0.0386311151,
0.02137197,
0.0589897521,
-0.0367720872,
0.0795118287,
0.0377340391,
-0.0330337957,
0.0035232208,
-0.0098020202,
0.051810313,
0.0126622114,
-0.0137084303,
0.037841171,
0.0132748662,
-0.0124553386,
-0.0001533672,
-0.021871483,
-0.0204181112,
-0.0070868749,
-0.0191757791,
0.0146579053,
-0.1410079002,
-0.088446334,
0.0117289461,
0.0261319689,
0.0069389855,
0.0417257212,
0.0291595943,
-0.0375024118,
-0.0237189438,
0.0323187597,
0.0117309224,
0.048402369,
-0.0576544255,
0.0644254535,
-0.0060134823,
-0.0442217402,
0.0258772522,
-0.0158523601,
-0.0199299771,
0.0397505425,
-0.0338711739,
0.006661627,
2.042035037e-34,
0.0014619051,
0.0286966041,
-0.0520921312,
0.0187905915,
0.0038535567,
-0.0574490763,
0.0298380367,
-0.0073898262,
0.0135542899,
-0.009211543,
0.0066452534
] |
http://arxiv.org/abs/2401.16561v1 | http://arxiv.org/pdf/2401.16561v1 | Multi-class Regret Detection in Hindi Devanagari Script | 2024-01-29T20:58:43 | 2024-01-29T20:58:43 | Renuka Sharma, Sushama Nagpal, Sangeeta Sabharwal, Sabur Butt | cs.CL | cs.CL, cs.AI | The number of Hindi speakers on social media has increased dramatically in
recent years. Regret is a common emotional experience in our everyday life.
Many speakers on social media, share their regretful experiences and opinions
regularly. It might cause a re-evaluation of one's choices and a desire to make
a different option if given the chance. As a result, knowing the source of
regret is critical for investigating its impact on behavior and
decision-making. This study focuses on regret and how it is expressed,
specifically in Hindi, on various social media platforms. In our study, we
present a novel dataset from three different sources, where each sentence has
been manually classified into one of three classes "Regret by action", "Regret
by inaction", and "No regret". Next, we use this dataset to investigate the
linguistic expressions of regret in Hindi text and also identify the textual
domains that are most frequently associated with regret. Our findings indicate
that individuals on social media platforms frequently express regret for both
past inactions and actions, particularly within the domain of interpersonal
relationships. We use a pre-trained BERT model to generate word embeddings for
the Hindi dataset and also compare deep learning models with conventional
machine learning models in order to demonstrate accuracy. Our results show that
BERT embedding with CNN consistently surpassed other models. This described the
effectiveness of BERT for conveying the context and meaning of words in the
regret domain. | Multi-class Regret Detection in Hindi Devanagari Script: The number of Hindi speakers on social media has increased dramatically in
recent years. Regret is a common emotional experience in our everyday life.
Many speakers on social media, share their regretful experiences and opinions
regularly. It might cause a re-evaluation of one's choices and a desire to make
a different option if given the chance. As a result, knowing the source of
regret is critical for investigating its impact on behavior and
decision-making. This study focuses on regret and how it is expressed,
specifically in Hindi, on various social media platforms. In our study, we
present a novel dataset from three different sources, where each sentence has
been manually classified into one of three classes "Regret by action", "Regret
by inaction", and "No regret". Next, we use this dataset to investigate the
linguistic expressions of regret in Hindi text and also identify the textual
domains that are most frequently associated with regret. Our findings indicate
that individuals on social media platforms frequently express regret for both
past inactions and actions, particularly within the domain of interpersonal
relationships. We use a pre-trained BERT model to generate word embeddings for
the Hindi dataset and also compare deep learning models with conventional
machine learning models in order to demonstrate accuracy. Our results show that
BERT embedding with CNN consistently surpassed other models. This described the
effectiveness of BERT for conveying the context and meaning of words in the
regret domain. | [
0.0544035733,
0.0860615894,
-0.0106506106,
-0.0038844822,
-0.0077211359,
0.0182380565,
0.0055726231,
0.0389349274,
0.053516008,
-0.0351261981,
-0.0055059907,
-0.048222933,
-0.0306938104,
0.0005459767,
-0.0062484541,
-0.0162013881,
0.0131924581,
0.0561095029,
0.0946720615,
-0.0243921001,
0.0251203869,
0.03377451,
-0.0100928741,
0.0399534926,
-0.083372362,
-0.0031193991,
0.0145976851,
0.0223327707,
0.0226692073,
-0.0339859352,
0.0602420717,
0.0600275807,
0.0202471316,
0.0521562658,
0.0000019333,
-0.0403448939,
-0.0272734519,
-0.0463154428,
-0.0177726876,
0.0347567834,
0.0622183941,
0.0378859006,
-0.0085451631,
0.0086721499,
-0.0335706808,
0.0334224328,
0.0613532737,
0.0577009693,
0.0378011465,
0.0916643739,
-0.0339633487,
-0.022881994,
0.0063860351,
-0.0375562981,
0.0100637861,
0.0357531384,
0.0071645975,
-0.033367563,
-0.0621444806,
-0.0556848198,
-0.0271527432,
0.0104074106,
-0.0017906517,
-0.009257976,
0.0058420007,
0.0027598704,
0.0018650775,
-0.0259628817,
0.0206244551,
-0.0018137463,
-0.0192725267,
-0.0240883008,
0.0046724211,
-0.0092990994,
-0.0378092863,
-0.0272864606,
-0.0248072967,
0.0054609934,
0.0015867188,
0.0540520549,
-0.0004558822,
0.0256991629,
0.0092587145,
0.0365059003,
0.0290044546,
-0.0074379905,
-0.0147367362,
-0.0196282025,
0.0129398182,
-0.0073957173,
-0.02504736,
-0.0263743363,
0.0545168221,
-0.0045048324,
-0.0159074403,
0.0291867126,
-0.0148308463,
-0.051673796,
0.0271214508,
-0.0556453615,
-0.0075506452,
0.0077621397,
-0.0619922765,
-0.0154499663,
-0.0135476161,
0.0331976749,
0.0030083947,
0.0208890792,
-0.0028578411,
0.0177747365,
0.0029252903,
-0.0459441654,
0.0025041264,
0.0359994136,
0.0431717746,
0.0196410958,
-0.0668438524,
0.0042316881,
-0.0389793999,
0.0175040755,
-0.0344651006,
0.0168512892,
-0.0117174555,
-0.0176302698,
-0.0408612154,
0.047992114,
-0.0213017296,
-0.0016896337,
0.0158277508,
-0.0525447838,
0.0585544631,
-0.0219416637,
0.027417196,
-0.0264317393,
0.0260897502,
0.0482150763,
-0.0133958133,
-0.0242313091,
-0.0480658747,
0.003437161,
-0.0405038483,
-0.027673034,
-0.0395153314,
-0.0585014224,
0.0257084146,
0.0004874327,
-0.002971027,
-0.0366447195,
0.0472003706,
0.0184549447,
-0.0358434431,
0.0451418906,
0.0302800853,
-0.0126587376,
0.0459473729,
0.0150819253,
0.0570763201,
0.0020917167,
-0.0065844571,
0.1067151576,
0.0417137444,
-0.0038910864,
0.0204633959,
-0.0042608134,
0.0110046286,
-0.0184445735,
-0.0850200057,
-0.0096426494,
-0.0501051396,
0.072877124,
-0.0037750616,
0.0462853014,
-0.0212792084,
0.0076110135,
0.0362891257,
0.0746446401,
-0.0016057107,
0.0961859077,
-0.0252583139,
0.0304153357,
0.0228120107,
-0.0186745338,
-0.0165311545,
0.0309650581,
0.0048870509,
-0.0297498833,
0.0087988814,
0.005530918,
-0.0115114748,
-0.0133072445,
-0.0327887237,
-0.0347904004,
-0.0053632711,
-0.021443706,
0.0248118695,
-0.005028856,
-0.0549549721,
0.0060920361,
0.0438975766,
-0.0946250036,
-0.020664297,
-0.0138178859,
0.0415307991,
0.0861149877,
0.0227034967,
-0.0315371342,
-0.0307207219,
-0.0032079211,
-0.0269399676,
0.0000477351,
0.0361937843,
-0.0100515541,
-0.0210304819,
-0.0577209033,
0.0214232914,
0.0289733559,
-0.005137912,
0.046084296,
-0.0186540503,
0.062988922,
-0.0192792192,
-0.0070186821,
0.0198250134,
0.0171679333,
0.0216362979,
-0.0050652064,
0.0004078364,
0.0481248796,
0.0138373803,
0.0127982805,
-0.0119829345,
-0.0176741183,
0.0198248923,
0.0150834927,
-0.0752884373,
0.0291839633,
-0.0273683351,
-0.0440664962,
-0.0066589219,
-0.0683861673,
-0.0174500402,
0.0672728866,
-0.0172835272,
-0.0548571646,
0.0133596305,
0.0162421558,
0.0933275297,
-0.0619089603,
0.005912838,
-0.0635406151,
-0.0046410165,
-0.0209189877,
0.0652930811,
-0.0365832373,
0.009096209,
-0.0357082225,
-0.0587983355,
0.0461937748,
-0.0490980782,
-0.0571733862,
-0.0054399469,
-0.001242506,
0.0306468271,
0.0153121212,
0.0400881618,
-0.0422291905,
-0.0448954292,
-0.1123130918,
-0.0079200193,
0.007633328,
-0.0194012895,
-0.0196531732,
0.0003489901,
-0.0023897537,
-0.0180141144,
-0.0507837161,
0.0323308222,
0.0082111564,
0.0161011759,
0.0449491441,
-0.0357142724,
0.0590648763,
-0.0628092885,
-0.0356253535,
-0.0420234688,
0.0346059427,
-0.0239562895,
-0.0359568261,
0.046377413,
0.0028665706,
-0.0357929692,
0.0236862209,
-0.0048735919,
-0.0797015876,
-0.0181363858,
-0.0474990048,
0.0434547476,
0.0168735366,
0.0349157304,
0.1059996039,
-0.0030953782,
0.0120728686,
-0.0281259753,
0.0094429692,
-0.009862287,
0.006743575,
-0.0290130861,
0.0062990603,
-0.0455669425,
0.0060061989,
0.0070973225,
0.0134643,
-0.0277677868,
-0.0956122205,
-0.0205264464,
-0.0033105898,
-0.0377340056,
0.0085229445,
0.0021043883,
0.0398031324,
0.0219078436,
-0.0475082099,
0.03049252,
0.0560520999,
-0.0015989796,
0.0059222667,
1.78e-8,
0.0143313026,
0.0511205085,
-0.0131358644,
0.027406076,
0.0131678777,
-0.0463852063,
-0.0251548477,
0.0084619289,
-0.0174397323,
-0.0572439209,
0.0041255229,
0.0004443843,
-0.0078850761,
0.0140618039,
-0.0337559618,
-0.0386859253,
-0.0831154138,
-0.023891788,
-0.070907712,
0.0242034253,
-0.0379458778,
0.0027164137,
-0.0193093941,
-0.0041384269,
-0.0515496172,
0.0033729228,
-0.0238907486,
0.0401322246,
0.0292990319,
0.0690335631,
-0.0323426723,
0.0614295788,
0.0313308723,
0.0118787335,
-0.0006404479,
0.0747908205,
0.0064086551,
-0.0317254029,
-0.0078118169,
0.009267007,
-0.0171097182,
-0.0035328392,
-0.0071645221,
0.0240590125,
0.0274728965,
-0.0005879669,
0.0728692934,
0.0134583665,
-0.0728525817,
-0.0065831076,
-0.0560835525,
-0.0072166272,
-0.0004994339,
-0.0145423785,
-0.0807707235,
-0.0209443998,
0.0008639623,
-0.0474276319,
-0.0065303771,
0.0127552766,
0.0294846669,
-0.0942028537,
-0.0069231824,
-0.0273320116,
0.0207311902,
0.0068062092,
0.010650672,
-0.0410125218,
-0.0177075695,
0.0300022066,
-0.0206945892,
-0.0083983801,
-0.0229694452,
0.0191003848,
-0.0557024367,
-0.0132165747,
0.0525648445,
0.0062023047,
0.0522626378,
0.0027671282,
0.0768760815,
-0.013256589,
-0.0354202464,
0.010841934,
-0.0064974492,
-0.0234335177,
0.0136304898,
0.0509090424,
-0.0046940423,
-0.0219641365,
-0.0461448431,
0.0387427099,
-0.0262629874,
-0.0002188574,
-0.0351383612,
0.0433058925,
-0.0174941104,
-0.017589774,
0.0158473533,
-0.0165989324,
-0.0005097326,
0.0155840283,
0.002516031,
-0.0580165088,
-0.0499439389,
-0.0038329188,
0.0294059236,
-0.0214774571,
-0.0148190632,
0.0096839182,
-0.0060489657,
-0.0330643803,
-0.0190195423,
-0.05120508,
-0.0009595654,
-0.0032397006,
0.0182708595,
-0.0265049096,
0.010509911,
-0.0235034432,
-0.0793726668,
-0.0202249922,
-0.0071489443,
0.0329865888,
-0.0089833159,
0.0082823811,
-0.0075270622,
0.0046616988,
0.007617848,
-0.0531403944,
-0.0511261187,
0.006019169,
0.0426315442,
0.0153582515,
0.0108177867,
-0.0262296069,
-0.0336552784,
-0.0743882656,
0.0534647219,
0.0098040625,
-0.0089606224,
-0.0269503146,
0.0483834557,
0.078110151,
-0.0029650324,
0.0481317863,
-0.0155017572,
-0.0188062619,
-0.017134428,
-0.0292723421,
-0.0022000747,
-0.006343883,
0.0007537207,
-0.0216930024,
0.0364718698,
-0.0477502868,
-0.0092945471,
-0.0464491993,
-0.0309200305,
0.0254103784,
-0.0121751111,
-0.0078832423,
-0.0124346269,
0.0779575258,
0.0042690481,
-0.0343507081,
-0.0882828757,
-0.001442179,
-0.0145170894,
-0.0436849743,
0.046166081,
0.0662767589,
0.0726202503,
0.0100457203,
0.0390423648,
-0.0246278942,
0.0009551015,
-0.0095648626,
-0.0179704092,
-0.0343671478,
0.0395571366,
-0.0023021363,
-0.0421366505,
0.0542358942,
0.0168307684,
0.0193034299,
0.018867664,
-0.0287984628,
0.0242698658,
0.0924466997,
-0.056576591,
0.0615197122,
-0.0393074341,
0.0294756033,
-0.0135136303,
0.0347253159,
-0.046542339,
-0.0579922758,
0.0125724711,
-0.0126210535,
-0.0292102396,
0.0177628864,
0.037729945,
-0.025052188,
0.0261321012,
0.0353545398,
-0.0216282103,
-0.0207791477,
0.0001096345,
0.070906058,
0.0133990478,
0.0192468222,
-0.0183578543,
-0.043754261,
-0.0405771211,
0.0201982073,
-0.0418157652,
0.0318212733,
-0.0170457046,
0.0083303992,
0.0235489961,
0.0185108557,
0.0356202684,
-0.0511865802,
0.0551133826,
-0.0428241789,
-0.0222674087,
-0.0499381758,
-5.743914657e-33,
0.0364140235,
-0.0084562432,
-0.0510093868,
-0.0009045894,
-0.0259902012,
0.0222638734,
0.0007670732,
-0.0103355572,
0.0100586871,
0.0770712718,
-0.0490247384,
0.0461000912,
0.0139168724,
0.0346305817,
0.0720480382,
-0.0304475222,
0.002997393,
0.0100023653,
0.0185716748,
0.0078093056,
0.0667759627,
0.0454234742,
0.094019033,
-0.0205674414,
0.0111893434,
-0.0053487951,
-0.0524238944,
0.0122604165,
0.0323797725,
0.0284223277,
-0.0399123132,
0.0498819873,
0.0213202685,
-0.0346698165,
0.0132962568,
0.0450599343,
-0.0553907491,
-0.0399864018,
-0.019501213,
0.0641321167,
0.0018213709,
-0.0238859188,
-0.0076216231,
-0.0192967653,
-0.0000719344,
0.0409470983,
0.0370918848,
0.0460314415,
-0.0074405475,
-0.0272550713,
-0.0286022909,
0.0016546506,
-0.0299525317,
-0.0112510705,
-0.0112480829,
0.0372481458,
0.0147008942,
-0.0421134718,
-0.0644720718,
-0.0041014976,
-0.0308822505,
0.0050695869,
-0.0178550612,
-0.0087267514,
-0.0343522802,
0.0768822879,
0.0130592277,
-0.021817131,
0.0276499111,
0.0684884861,
0.0216474477,
0.0312388502,
-0.0032807274,
0.0400968455,
0.0332285613,
-0.1145503744,
-0.0002377332,
0.0203005765,
0.0160267092,
0.0428273454,
0.0375443436,
0.05054399,
-0.0052884948,
0.0015967715,
0.016088549,
-0.0357364863,
-0.0396490805,
-0.0050318795,
0.0093333544,
-0.0461671017,
-0.0108898357,
0.0183191299,
0.0081872549,
-0.037499506,
-0.0178589337,
0.0386389233,
-0.0555868186,
0.0552770458,
0.0094890734,
0.0355972536,
-0.0142036472,
-0.0402034968,
-0.0023029039,
0.0600132234,
0.0243319888,
0.0274303239,
0.070540145,
0.0249964893,
-0.0620794706,
0.009035198,
-0.0207386035,
-0.0323885009,
-0.0130025186,
-0.0228952058,
-0.0051113931,
0.0026210952,
-0.0081840437,
-0.0059198602,
0.0187017564,
0.0653878525,
-0.0196680259,
-0.0236428902,
0.0058778501,
-0.0269606486,
-0.0119167902,
0.0080065178,
-0.0662572011,
0.0126558272,
0.019649243,
0.0022446765,
0.0141283553,
-0.0324373133,
2.597e-7,
-0.0152744902,
0.0541229807,
0.0465661064,
-0.0724105239,
0.0053837947,
0.0632200539,
0.0134153245,
0.0205280334,
-0.0203651935,
0.0604869947,
0.0567260087,
-0.0123175113,
0.0488305874,
-0.0042069978,
-0.0368577093,
0.005178812,
-0.0836857781,
-0.0477725416,
-0.0280958898,
0.005831718,
0.0576206595,
0.0766350999,
0.0820437595,
-0.0395488851,
-0.0087757176,
0.0316440873,
0.0463425592,
-0.0453863628,
0.0019499549,
0.0143781584,
0.0265528057,
-0.0111378906,
-0.0076383008,
-0.0775412396,
0.0051281797,
-0.0460112914,
0.037965849,
0.0449121259,
-0.0201908369,
-0.0455548242,
0.0248127282,
0.0341161303,
-0.0207040422,
-0.0221837703,
0.0217864085,
0.0261847526,
-0.0285649933,
-0.0523196124,
0.0445365682,
0.0014592834,
0.0841360465,
-0.0151463961,
-0.0205139499,
-0.0136328135,
0.0022235182,
0.007112009,
-0.0174282715,
-0.0669767931,
-0.0075610392,
0.0682801679,
0.0039641396,
0.0282052346,
-0.0284865778,
0.0012638307,
0.0524377488,
-0.0767835528,
0.0380595997,
1.842459858e-34,
0.021411946,
0.0040464238,
0.0106028281,
-0.0219955649,
0.0268544275,
0.0301197227,
0.0061930283,
0.0031449497,
0.0223226864,
-0.0105983308,
-0.012276398
] |
http://arxiv.org/abs/2401.16558v1 | http://arxiv.org/pdf/2401.16558v1 | Diverse, but Divisive: LLMs Can Exaggerate Gender Differences in Opinion Related to Harms of Misinformation | 2024-01-29T20:50:28 | 2024-01-29T20:50:28 | Terrence Neumann, Sooyong Lee, Maria De-Arteaga, Sina Fazelpour, Matthew Lease | cs.CY | cs.CY, cs.CL | The pervasive spread of misinformation and disinformation poses a significant
threat to society. Professional fact-checkers play a key role in addressing
this threat, but the vast scale of the problem forces them to prioritize their
limited resources. This prioritization may consider a range of factors, such as
varying risks of harm posed to specific groups of people. In this work, we
investigate potential implications of using a large language model (LLM) to
facilitate such prioritization. Because fact-checking impacts a wide range of
diverse segments of society, it is important that diverse views are represented
in the claim prioritization process. This paper examines whether a LLM can
reflect the views of various groups when assessing the harms of misinformation,
focusing on gender as a primary variable. We pose two central questions: (1) To
what extent do prompts with explicit gender references reflect gender
differences in opinion in the United States on topics of social relevance? and
(2) To what extent do gender-neutral prompts align with gendered viewpoints on
those topics? To analyze these questions, we present the TopicMisinfo dataset,
containing 160 fact-checked claims from diverse topics, supplemented by nearly
1600 human annotations with subjective perceptions and annotator demographics.
Analyzing responses to gender-specific and neutral prompts, we find that GPT
3.5-Turbo reflects empirically observed gender differences in opinion but
amplifies the extent of these differences. These findings illuminate AI's
complex role in moderating online communication, with implications for
fact-checkers, algorithm designers, and the use of crowd-workers as annotators.
We also release the TopicMisinfo dataset to support continuing research in the
community. | Diverse, but Divisive: LLMs Can Exaggerate Gender Differences in Opinion Related to Harms of Misinformation: The pervasive spread of misinformation and disinformation poses a significant
threat to society. Professional fact-checkers play a key role in addressing
this threat, but the vast scale of the problem forces them to prioritize their
limited resources. This prioritization may consider a range of factors, such as
varying risks of harm posed to specific groups of people. In this work, we
investigate potential implications of using a large language model (LLM) to
facilitate such prioritization. Because fact-checking impacts a wide range of
diverse segments of society, it is important that diverse views are represented
in the claim prioritization process. This paper examines whether a LLM can
reflect the views of various groups when assessing the harms of misinformation,
focusing on gender as a primary variable. We pose two central questions: (1) To
what extent do prompts with explicit gender references reflect gender
differences in opinion in the United States on topics of social relevance? and
(2) To what extent do gender-neutral prompts align with gendered viewpoints on
those topics? To analyze these questions, we present the TopicMisinfo dataset,
containing 160 fact-checked claims from diverse topics, supplemented by nearly
1600 human annotations with subjective perceptions and annotator demographics.
Analyzing responses to gender-specific and neutral prompts, we find that GPT
3.5-Turbo reflects empirically observed gender differences in opinion but
amplifies the extent of these differences. These findings illuminate AI's
complex role in moderating online communication, with implications for
fact-checkers, algorithm designers, and the use of crowd-workers as annotators.
We also release the TopicMisinfo dataset to support continuing research in the
community. | [
0.0513271876,
0.0558063015,
-0.0088068536,
0.0092923185,
-0.0171686392,
0.0074844779,
0.031221183,
-0.0225790907,
-0.002236726,
-0.0810925215,
0.0278853681,
-0.0431666449,
0.0278074946,
0.0092681907,
0.0280707907,
-0.0184394382,
-0.0053409245,
-0.0003675586,
0.1072795019,
-0.0307305399,
0.0511442088,
0.0294488929,
-0.0088554015,
0.0617605038,
-0.0290349051,
-0.022370575,
0.0556359813,
0.0064504123,
0.0401365906,
-0.0346169621,
0.0166489519,
0.0590677112,
0.0207626186,
0.0156419147,
0.0000020322,
-0.0533124171,
0.0191316176,
0.0322782807,
-0.0306645203,
0.0142611191,
0.0329407379,
0.0225317776,
-0.024964571,
0.0165018905,
-0.021322716,
0.0041464712,
0.0318155587,
0.0300547089,
0.0672958419,
0.0330163948,
-0.0128215207,
0.0255226977,
-0.0346820988,
0.0049225511,
-0.000269804,
-0.0190300904,
0.0470370129,
-0.0257192273,
-0.0549461134,
0.041752506,
0.0111723281,
0.0218695775,
-0.0266951583,
-0.0321942233,
-0.0185588095,
0.0177098755,
-0.0282719098,
-0.0448747538,
-0.0038474267,
0.0621354319,
0.0516700707,
-0.0006501234,
0.0235500075,
-0.0114393253,
0.036381688,
0.0508992188,
-0.0264904872,
0.0721862316,
-0.0483632348,
-0.0305288639,
0.0014051307,
-0.0236653928,
0.0350128151,
0.0496449322,
0.0272716545,
0.1000854969,
-0.0205393173,
-0.0287557691,
0.0228606667,
-0.0087392954,
0.0271877721,
-0.0611419082,
0.0025376121,
-0.0434624739,
-0.0124238059,
-0.0065635266,
0.025935201,
-0.0329643451,
0.0521858856,
-0.0936514139,
-0.0208594613,
0.012899789,
0.0010795811,
0.0281151216,
0.0309092905,
-0.0082621314,
-0.0021003925,
0.0167770758,
0.0236035287,
-0.0190911219,
-0.0188300144,
0.0137665141,
0.0156800635,
-0.0013905216,
0.0148757566,
0.0581442006,
-0.0324732065,
-0.0304970145,
-0.0896587223,
0.0137492158,
-0.0017644241,
-0.0010046454,
0.0029964813,
-0.0195838623,
-0.0409795642,
0.0533487983,
-0.0663314015,
0.0408061855,
-0.0023215348,
-0.0754880831,
0.0067531965,
0.0165808816,
-0.0204366688,
0.0134768328,
0.000885695,
-0.0166077707,
0.02925759,
0.0106588136,
-0.0185768083,
-0.012145442,
-0.0231501143,
-0.0398987643,
-0.0028591566,
-0.061212942,
0.0173551757,
0.0389240086,
0.0084630363,
-0.0261496771,
0.0061670439,
0.0293139331,
-0.0404184647,
0.0484862179,
0.02530347,
0.0246148817,
0.0795238018,
0.0252885502,
-0.01524786,
0.0715171546,
0.0363923386,
0.0369717516,
0.0310501363,
0.0276660174,
0.0073078158,
0.0024999795,
0.0097929062,
0.0589279309,
0.0240586493,
0.0152337253,
-0.0425025821,
-0.0130230319,
0.0324636847,
0.0491290763,
-0.0086303335,
-0.0208683349,
0.074420616,
0.0441484191,
-0.0564803854,
0.0248488225,
0.023451332,
0.0225264058,
-0.0091318702,
-0.0122832367,
-0.0507054813,
0.0129795093,
-0.0316299498,
-0.0646810308,
0.0170051474,
0.0536900796,
-0.0189249218,
-0.0367168821,
-0.0256420206,
0.0173577704,
0.0318473503,
0.0339037776,
-0.0174452942,
-0.0136017753,
-0.0134045193,
0.0325999036,
0.0169220436,
-0.0344455056,
-0.0337789543,
-0.0504400618,
0.0212307815,
0.0336229987,
-0.0253456254,
-0.0195374116,
-0.0020084395,
-0.0706253648,
0.0276908707,
0.0260707065,
0.0625734776,
-0.0047336309,
-0.0581636578,
-0.0202569254,
-0.034458179,
0.0072009503,
0.0314277224,
0.049548395,
0.0618768185,
0.0223768838,
-0.0014967876,
-0.0212906357,
-0.0654228106,
-0.0001336505,
0.0227495376,
-0.0052647251,
-0.0022757016,
0.0154073657,
0.0433057435,
0.0176235382,
-0.0248838905,
0.0098432433,
0.0095211528,
-0.000242086,
-0.0370517857,
-0.0247497279,
0.0180904828,
0.0796157271,
0.0092010666,
-0.0559065416,
-0.0937413797,
0.0165430233,
-0.0109897256,
0.0273753852,
0.0267952103,
-0.028957054,
-0.0563593358,
0.0038664569,
-0.0206269789,
-0.0060438868,
-0.0096526854,
-0.01245174,
0.0408688672,
-0.0562009737,
-0.0315323733,
0.006642608,
-0.049779404,
-0.0151996203,
-0.0417697392,
-0.0401274562,
-0.0061167195,
0.0317746364,
0.0274109021,
0.0554635115,
-0.0547219329,
-0.0117022479,
-0.0091685224,
-0.0731836036,
-0.0428256914,
-0.0109008523,
-0.0051119528,
0.0063950005,
-0.0058079283,
-0.0135321645,
0.0067653526,
-0.0199297927,
0.0013287568,
-0.0216878932,
0.0487413667,
0.0284557734,
0.0010222369,
0.0167158917,
-0.0250676926,
0.0028910055,
-0.040340703,
0.0119195022,
0.0819149315,
-0.0574950092,
-0.0379479416,
-0.03217775,
-0.0713997483,
-0.0003185342,
-0.0338052623,
-0.0093825702,
0.0330662169,
0.0330306515,
0.0239762273,
0.0414091907,
0.0470784679,
0.0407370292,
0.0749013275,
0.044121366,
-0.0301794186,
-0.0291793793,
0.0149401501,
0.0160189644,
0.019571187,
-0.0898505822,
0.0247101095,
0.0238450058,
0.0074967397,
-0.0103899548,
-0.0338499434,
0.0588244125,
-0.0684276447,
-0.0433309004,
-0.0227107592,
-0.0617003366,
0.0117619401,
0.0549329892,
-0.0294100381,
-0.0353907607,
0.0570895001,
0.0017462651,
0.0444276817,
-0.0089736525,
-0.0080819158,
-0.0389156491,
-0.002412844,
-0.0337587744,
0.0262343865,
0.0086773271,
0.0280958358,
-0.0370621309,
-0.0485537164,
0.0636561885,
-0.0423415415,
0.0216669887,
-0.0547561981,
0.0387413241,
0.0320643745,
-0.0556248911,
-0.0127101578,
-0.0841869712,
-0.0370875448,
-0.0250775758,
0.018047208,
-0.0141961006,
-0.0520679802,
0.0004086349,
0.0666489676,
0.015223355,
0.0197521914,
-0.0055752043,
0.0257848948,
-0.0337268896,
0.0113332467,
0.0178185012,
0.0849265829,
-0.0236054137,
-0.0216192529,
0.059602607,
-0.0026182313,
-0.0084330197,
0.0002120539,
-0.0353726521,
0.0197992921,
-0.0081533268,
-0.0027498507,
0.0051635364,
0.0236480385,
-0.0448369868,
0.0290392973,
0.0564724728,
-0.004027036,
-0.0514645875,
-0.0222051293,
-0.0076562129,
0.0085343309,
0.041241657,
-0.0052033602,
-0.067277208,
0.0314674117,
0.0146427853,
0.0005079488,
-0.0618524663,
-0.0449982509,
-0.0234410316,
-0.099826999,
0.0349833369,
0.0032673329,
0.0062010926,
-0.0232734382,
-0.0393784158,
0.0030786749,
-0.0006069305,
0.0064496463,
-0.0045141359,
0.0411952846,
-0.0131719979,
-0.0103781894,
-0.0409320295,
-0.0379573666,
0.050735645,
-0.028820293,
0.0585802794,
0.0275981594,
0.0540189296,
-0.0391175076,
-0.0305478275,
0.0412986465,
0.0353834629,
-0.0428489409,
-0.0084501747,
0.0425302051,
-0.0338453464,
0.0088141114,
-0.0605162308,
0.0066334573,
0.0292045847,
0.0019753219,
0.0077748057,
0.0176866762,
0.0342371315,
-0.0452232324,
-0.0396253057,
0.0100410488,
0.0469623171,
0.057915207,
-0.0033234423,
-0.0181296375,
-0.0226306282,
-0.0664145648,
0.0067457724,
0.0372273512,
-0.0775833204,
0.0161749721,
0.0312015824,
-0.0760752261,
-0.0415785462,
-0.0095052412,
-0.0029335716,
-0.0240444224,
0.0118194502,
-0.0440395474,
0.0811107308,
0.0266359225,
-0.0166965406,
0.0456309132,
-0.0200395584,
0.0287405234,
-0.0093644634,
-0.0290992837,
-0.0177063625,
0.005467386,
-0.018957857,
0.0678269938,
-0.0438549928,
-0.0332637392,
0.042948354,
0.0244221371,
0.0209941193,
0.0328165665,
-0.0093689216,
-0.0369453467,
0.0324044339,
0.046034459,
-0.0229163524,
0.0160639044,
0.0122261774,
0.0234679431,
0.0180099327,
0.0076636383,
-0.0340581462,
0.010847073,
0.0157381948,
-0.0200977623,
-0.0635656416,
-0.0234937146,
0.0393064059,
0.0334350467,
0.0026174793,
-0.0135009121,
0.0092036119,
-0.0471940227,
-0.0382651426,
-0.000639172,
-0.0449290499,
0.0082011418,
-0.0187553838,
0.0175790489,
-0.0448648259,
-0.0039796182,
0.0254026633,
-0.0232982207,
0.0001134678,
-0.0263155755,
0.0189096369,
0.0073843002,
0.0549610406,
0.0051350393,
-0.0293182489,
0.0091911722,
0.0112972055,
-0.023134863,
0.0214770548,
0.0289491992,
0.0046743704,
-0.0032181817,
-0.0006242813,
0.0131482016,
-0.0167669207,
-0.0501865372,
-0.0480931662,
-0.0018830828,
0.0293804538,
-0.0213655923,
0.0043854513,
0.0395882949,
0.0082446504,
-0.0349867344,
-0.0029866858,
-0.0286507197,
0.0013600972,
0.0140113551,
0.0273554567,
0.0118832821,
-0.0218409039,
0.0116442023,
0.0389242247,
0.018913785,
0.0333895423,
0.0080923932,
-0.0858830884,
-0.0140915066,
-0.0518949144,
-0.0437897481,
0.0097996313,
0.0122725898,
0.0013819899,
-0.1329508424,
-0.038465254,
0.0020154975,
0.0034899635,
0.0311152954,
-0.0150968609,
-0.0167841464,
0.0632980615,
-0.0073782383,
-0.0433798991,
-0.0633043498,
0.0202540215,
-0.0236308984,
-0.007946372,
-0.0157422964,
-6.339295567e-33,
-0.0124249784,
-0.0770619512,
-0.0344130583,
-0.0014123084,
0.0019593108,
0.0517511144,
0.0103426194,
0.0127461394,
-0.0333782695,
0.0576323941,
0.009487398,
0.0490623638,
0.0111806504,
-0.0375608876,
-0.0000812827,
-0.0199226867,
0.0383146182,
-0.0222962704,
-0.0423953347,
-0.0056016995,
0.0063694096,
-0.0345609151,
-0.0165640041,
-0.1083795726,
0.1138600484,
-0.009526832,
-0.0296965893,
-0.0061222361,
-0.0409219712,
0.0104451105,
0.0107685653,
0.0673026666,
0.0282506756,
-0.0602583028,
0.0169021804,
0.0622812025,
-0.0453324616,
-0.0080197453,
0.0314883702,
-0.0260050744,
0.0147933597,
0.003881155,
0.0204217453,
0.0257921834,
-0.0590517856,
-0.0157937054,
0.0464497954,
0.0046312236,
0.0711319596,
0.0093988888,
-0.0304279737,
0.0042127725,
-0.0254367217,
0.020126462,
-0.0235817153,
0.0413837805,
-0.0159810558,
0.0419900902,
-0.1092658043,
0.0712235048,
-0.00498359,
-0.0251594037,
0.0220748987,
-0.0137122134,
-0.0015431158,
0.0433273613,
0.0184563603,
0.0681282356,
-0.0875914246,
-0.0599926934,
-0.0337272249,
-0.0078705186,
0.032908801,
0.1012354717,
0.0128611615,
-0.0182557832,
-0.0748264939,
-0.0118036419,
0.0251350887,
0.020980401,
0.0447333083,
-0.0448246226,
-0.0657133162,
-0.0415313952,
-0.0335478932,
-0.0363403335,
-0.038430538,
0.0317829214,
0.0061273039,
0.0233170539,
-0.0731816068,
0.0077885021,
-0.0254900753,
-0.0366601907,
-0.0054756561,
0.0163110718,
0.0060245791,
-0.0125733363,
0.0429552458,
0.0101241805,
-0.0601605475,
0.0739724338,
-0.0232572444,
0.0240505319,
0.0273655914,
0.0170253571,
0.0131483972,
-0.0099944193,
-0.0576287769,
-0.0199006367,
-0.0041829953,
0.0342974998,
-0.0004489482,
-0.0041209324,
0.0358006507,
0.0381818712,
-0.0003219605,
-0.0112552876,
0.0399224423,
0.042016007,
0.0108504212,
-0.0104868216,
-0.0554546863,
0.0208957456,
-0.0287041254,
-0.0056181871,
0.0052312464,
0.0300799571,
0.0153327342,
-0.0435390137,
-0.0164630413,
-0.0413072929,
2.726e-7,
0.0320113339,
0.0666580051,
-0.0132003464,
0.0191132557,
0.0319994874,
0.0445533916,
0.0259509217,
0.022878157,
0.0472639911,
-0.0859500244,
-0.0275581423,
0.050976228,
-0.0053700777,
0.0519167855,
-0.041545663,
0.009219137,
-0.0366927758,
0.0187593568,
-0.0166786406,
0.0316559002,
-0.0136766406,
0.0293886401,
0.0297692604,
0.0011966582,
-0.0425463617,
0.0170738455,
0.0745054781,
-0.0178473033,
-0.0243177023,
0.0269590486,
0.0564366058,
-0.0105837071,
0.0143736601,
0.072373271,
0.0114017641,
-0.0267501753,
0.0169916525,
0.0368079878,
0.0042480333,
0.0061654346,
-0.0231757686,
0.0051559373,
0.0143716102,
-0.0661827177,
0.039668344,
0.0086455625,
-0.0493566543,
-0.0677167848,
-0.0080192517,
0.0012017889,
0.0698591471,
-0.0181659888,
-0.0334777236,
-0.0134709394,
0.0010987109,
0.023801988,
-0.0021445926,
-0.0135962674,
-0.0037927167,
-0.0232092142,
-0.0479055606,
-0.0983381048,
-0.0102226026,
0.0129060093,
0.0109584657,
-0.0276080035,
0.0541956313,
2.270277469e-34,
0.009947476,
0.0268892832,
-0.0048073176,
-0.0375389233,
0.048564557,
0.045174405,
-0.0314617939,
-0.0242106225,
0.030651303,
0.0095308255,
-0.0036110315
] |
http://arxiv.org/abs/2401.16553v5 | http://arxiv.org/pdf/2401.16553v5 | SelectLLM: Can LLMs Select Important Instructions to Annotate? | 2024-01-29T20:44:10 | 2024-04-18T01:35:12 | Ritik Sachin Parkar, Jaehyung Kim, Jong Inn Park, Dongyeop Kang | cs.CL | cs.CL, cs.AI | Instruction tuning benefits from large and diverse datasets, however creating
such datasets involves a high cost of human labeling. While synthetic datasets
generated by large language models (LLMs) have partly solved this issue, they
often contain low-quality data. One effective solution is selectively
annotating unlabelled instructions, especially given the relative ease of
acquiring unlabeled instructions or texts from various sources. However, how to
select unlabelled instructions is not well-explored, especially in the context
of LLMs. Further, traditional data selection methods, relying on input
embedding space density, tend to underestimate instruction sample complexity,
whereas those based on model prediction uncertainty often struggle with
synthetic label quality. Therefore, we introduce SelectLLM, an alternative
framework that leverages the capabilities of LLMs to more effectively select
unlabeled instructions. SelectLLM consists of two key steps: Coreset-based
clustering of unlabelled instructions for diversity and then prompting a LLM to
identify the most beneficial instructions within each cluster. Our experiments
demonstrate that SelectLLM matches or outperforms other state-of-the-art
methods in instruction tuning benchmarks. It exhibits remarkable consistency
across human and synthetic datasets, along with better cross-dataset
generalization, as evidenced by a 10% performance improvement on the Cleaned
Alpaca test set when trained on Dolly data. All code and data are publicly
available (https://github.com/minnesotanlp/select-llm). | SelectLLM: Can LLMs Select Important Instructions to Annotate?: Instruction tuning benefits from large and diverse datasets, however creating
such datasets involves a high cost of human labeling. While synthetic datasets
generated by large language models (LLMs) have partly solved this issue, they
often contain low-quality data. One effective solution is selectively
annotating unlabelled instructions, especially given the relative ease of
acquiring unlabeled instructions or texts from various sources. However, how to
select unlabelled instructions is not well-explored, especially in the context
of LLMs. Further, traditional data selection methods, relying on input
embedding space density, tend to underestimate instruction sample complexity,
whereas those based on model prediction uncertainty often struggle with
synthetic label quality. Therefore, we introduce SelectLLM, an alternative
framework that leverages the capabilities of LLMs to more effectively select
unlabeled instructions. SelectLLM consists of two key steps: Coreset-based
clustering of unlabelled instructions for diversity and then prompting a LLM to
identify the most beneficial instructions within each cluster. Our experiments
demonstrate that SelectLLM matches or outperforms other state-of-the-art
methods in instruction tuning benchmarks. It exhibits remarkable consistency
across human and synthetic datasets, along with better cross-dataset
generalization, as evidenced by a 10% performance improvement on the Cleaned
Alpaca test set when trained on Dolly data. All code and data are publicly
available (https://github.com/minnesotanlp/select-llm). | [
0.0448706672,
0.0419287197,
-0.0321513303,
0.0233561955,
0.0116612874,
0.0575901754,
0.0100444462,
0.0132091688,
0.0162252393,
-0.0361144133,
-0.0549578518,
0.0007701381,
0.0241712425,
-0.044398848,
0.0453339554,
-0.0120516885,
0.0571237803,
-0.0472766757,
-0.0319558755,
0.0110359862,
-0.0043849838,
0.0423074849,
0.0143555421,
0.0314898454,
-0.0172102172,
-0.0130625116,
-0.0041185752,
0.0107572749,
0.0263034757,
-0.0082631875,
0.0199808795,
0.0750222579,
0.0069089131,
0.0171547588,
0.0000019967,
-0.0070173801,
-0.0304900128,
-0.0100132376,
-0.033641424,
0.0547131188,
0.0323160253,
-0.0010178826,
-0.0198004674,
0.012707931,
0.0381633192,
0.000506836,
0.0644965619,
-0.0041097268,
0.0222561453,
0.0822705403,
0.0048572877,
0.0079402402,
-0.0048554051,
-0.0100088567,
0.0536533222,
-0.0063024629,
0.0360067226,
0.0039452091,
-0.0067037316,
0.0410729833,
-0.0101662297,
0.0256746281,
-0.0018906608,
-0.0079802526,
0.01205712,
0.050515864,
-0.0319571607,
-0.0300440695,
0.0158369355,
0.0856710821,
-0.0270173401,
-0.0112395752,
0.0270399805,
-0.0070440713,
0.0046564448,
-0.0086745368,
-0.0274698734,
0.0407807343,
-0.0171510503,
0.0186197739,
0.0213752836,
-0.0280551836,
0.0382540338,
-0.0174717847,
0.0288267862,
0.08359503,
0.032972306,
-0.012166311,
0.0118296668,
0.0092370724,
-0.0411887206,
-0.0510253236,
0.0029141253,
0.0024954465,
-0.0245090798,
0.0241063386,
-0.0485727414,
-0.0173272211,
0.0258386433,
-0.0450785421,
0.0753792077,
0.009854015,
-0.0394052118,
0.0192454197,
-0.0321483202,
0.0151467947,
0.0215339288,
0.0758526102,
-0.0289268494,
0.0270576105,
-0.0515279695,
0.0219184309,
-0.0212742407,
0.0396950655,
-0.0400788151,
0.026651321,
-0.0299241319,
0.0212712344,
-0.0140972817,
-0.0088727642,
0.0370986238,
-0.0300839674,
-0.050679449,
0.0341832377,
-0.000749872,
-0.0135891018,
-0.0460608415,
-0.0049089026,
-0.0377214067,
-0.0746814534,
-0.0226774421,
0.0587317646,
-0.0173206851,
-0.0348449461,
0.0276783016,
0.0568790212,
-0.0124939932,
-0.0403834321,
-0.0816634372,
-0.0286603104,
-0.0159426555,
-0.0356694609,
0.0201905388,
-0.0559819452,
0.0153340083,
-0.015101118,
0.0268851165,
-0.043950282,
0.0046170675,
0.0401087403,
-0.028589109,
0.0429898314,
-0.0187100861,
0.0119603537,
0.0044246446,
0.0137988832,
-0.0036123921,
0.0401081294,
0.0170441568,
0.0306564234,
0.0700600967,
-0.0242031775,
0.0027004138,
0.0356846973,
-0.007047168,
0.0162875988,
-0.0095381206,
0.0040113991,
0.0512994528,
0.0036961921,
0.0676334426,
0.0307918303,
-0.0109839244,
-0.0397645049,
0.0458470099,
0.0394314714,
-0.0230298005,
-0.0298300795,
0.0595367104,
0.0334271751,
-0.008463406,
0.0119772004,
-0.0312352404,
0.0201145299,
-0.0428783372,
-0.0094195502,
0.0631008074,
0.0040742094,
-0.0305062719,
-0.035819076,
-0.0283808336,
-0.0107076941,
0.0087161483,
0.0019900084,
0.0159116518,
-0.0099957567,
-0.0619059317,
0.0704342276,
0.0481765084,
-0.0306403544,
-0.0333522335,
-0.0435338654,
0.0536075719,
0.0450143814,
0.0307572167,
-0.0347258858,
-0.0297615305,
0.0081286253,
0.0464820713,
0.0167375933,
0.015854083,
-0.0466731004,
0.0358435586,
-0.074316673,
-0.040371269,
-0.0103142792,
0.0116174361,
0.0121668419,
0.0560784303,
-0.0417469256,
0.0148128094,
-0.0395542756,
-0.0605116375,
0.0807536468,
-0.0150288306,
-0.0369372331,
0.0088282609,
0.0025727763,
0.0157316178,
-0.0010504365,
-0.02051723,
-0.0027003936,
0.0435013734,
0.0026191105,
0.0210869908,
-0.0759391934,
-0.036019899,
-0.0416186303,
0.012089923,
-0.0277034212,
-0.0588951111,
0.0081343725,
-0.0671975985,
-0.0264145602,
-0.0299952347,
-0.0719966739,
0.0119480006,
-0.0092812236,
0.0236685816,
-0.0046988823,
0.0072260383,
-0.0415695608,
0.0408646576,
-0.0227169171,
0.0217239168,
0.0370852724,
0.0063865823,
0.0015523925,
-0.0251652841,
-0.0057359389,
-0.0010938192,
0.0029442825,
0.0013904331,
0.0234702192,
0.0516824163,
-0.0416334122,
-0.0355812274,
-0.0920029581,
0.0211770795,
-0.0352037549,
0.0124112265,
0.0390800424,
-0.0225003846,
-0.0055853636,
0.0029298349,
0.0023369703,
-0.0345670544,
-0.0636672154,
0.0333340764,
0.0244828556,
0.0224421378,
-0.0451804027,
-0.0320010707,
-0.0269613564,
0.0105567118,
0.0184021406,
-0.0129009495,
-0.010485163,
-0.0569319054,
0.0275198296,
-0.0055366806,
-0.0047530751,
-0.0040603797,
-0.039842058,
0.0232046042,
0.0052283336,
-0.0075288494,
0.0625094995,
-0.018874405,
0.0080374246,
0.0630403608,
-0.0095780138,
-0.0325460508,
0.0003986981,
-0.0147766434,
-0.0081416732,
-0.0359842107,
-0.0619775541,
0.0027314634,
0.1038246825,
-0.0019187974,
-0.0071290717,
-0.0526535064,
-0.0353143662,
-0.0039292504,
-0.0102428049,
-0.0325986817,
0.0079562543,
-0.0108207865,
0.0097513702,
-0.047295399,
-0.0796283633,
0.0534938313,
-0.0287849344,
-0.0160788447,
-0.0497583039,
0.0027088013,
-0.0167707093,
0.0144714182,
-0.0032239708,
0.0087384265,
0.0168463141,
-0.0125950957,
-0.0265523363,
0.03966875,
-0.0118654482,
-0.0083328504,
0.0383750163,
-0.0796786621,
0.0098175844,
0.0028313603,
0.0034773673,
-0.0326711759,
0.0113739511,
0.0266085435,
-0.0056981593,
0.0843869895,
-0.0410377309,
-0.0275539383,
-0.0214001965,
0.0228368882,
0.074413918,
-0.0040813782,
0.0036183614,
-0.0141781922,
-0.0280830953,
0.0307305604,
-0.0523773432,
-0.0113854902,
-0.0298327599,
-0.0127081452,
0.044877924,
0.0084590977,
-0.0309795607,
0.0040603057,
0.0334764309,
0.0459082611,
-0.0226985812,
-0.0407107547,
0.0636761561,
0.082026042,
-0.0318334736,
-0.0303958766,
0.1140090749,
0.0516741574,
-0.0378052816,
0.0223131105,
0.0165036302,
0.0452581048,
0.0249258149,
0.0405213237,
-0.0425606668,
-0.037080314,
0.0048185969,
0.0455812886,
-0.0113724638,
-0.0148696834,
0.0155867143,
-0.1025072262,
0.0497799963,
0.0268249847,
-0.0850397125,
-0.0257958565,
-0.0639062822,
-0.0251220949,
-0.0272510108,
0.0196961146,
-0.0058484506,
-0.0418053828,
-0.0004406199,
0.0352474935,
-0.0524316095,
-0.0654855222,
0.0514494069,
0.019716816,
0.0671035945,
0.0679094717,
0.0109939547,
0.0464948565,
-0.0298671462,
0.05822099,
-0.0145277036,
0.0005623404,
-0.0598157197,
0.0879964828,
-0.006863473,
-0.0213340484,
0.0061939354,
0.036715392,
-0.0316167399,
-0.0120855477,
-0.0117946984,
0.0471616611,
-0.0133717973,
0.0139316441,
0.0045594168,
-0.0159464162,
0.0069649657,
0.0579768717,
0.0589421503,
-0.0271637551,
-0.0097522894,
-0.043486461,
-0.0241735447,
-0.0905674845,
-0.0154704405,
0.0515309013,
-0.075766772,
-0.063520439,
-0.0792822167,
-0.0443053208,
0.07624387,
0.0010506934,
0.0141295912,
-0.0026255713,
0.0082793329,
-0.0038687382,
-0.0503179207,
0.0490863137,
-0.0570931435,
0.0110221216,
-0.0876850411,
-0.0893640071,
0.004544172,
0.0216037966,
-0.0143119972,
0.027292816,
-0.0623574778,
-0.0810451955,
0.0025990733,
0.0853939429,
0.0104392413,
0.046707239,
0.0194991566,
-0.0617827699,
0.0364173986,
0.0306115672,
0.0017709556,
0.0153707229,
0.0513839424,
-0.0497849844,
-0.0149727007,
0.0062056244,
-0.0321402214,
0.0405732952,
0.0475290269,
0.0089338478,
-0.0463645495,
-0.0004637128,
0.0159451663,
0.0028978419,
0.0415223129,
-0.0192663707,
0.0247790609,
-0.0460254848,
0.0254075862,
0.019529812,
0.0102240006,
-0.0237933621,
0.0109004993,
0.0144124636,
-0.0201813839,
-0.0200397372,
0.0239971783,
-0.0410008207,
0.0454952307,
-0.0173397604,
0.0338831134,
0.0176609121,
0.0460021906,
0.0091832066,
0.0218779221,
-0.0200922359,
0.0323289447,
0.0096077323,
0.0524852797,
0.0179541316,
0.0236800015,
-0.0178271737,
0.049279727,
-0.0036639643,
0.03059683,
-0.0158075634,
0.0044224458,
-0.0821044147,
0.0124588767,
0.0270512961,
-0.0604618341,
0.0676201433,
0.0052084909,
-0.0153172212,
-0.0572582819,
0.0146510433,
-0.0131348688,
-0.0271283537,
-0.0193595085,
0.0406808406,
-0.0312473644,
-0.0262559466,
0.104656972,
0.0228627026,
0.0620299019,
-0.0076666055,
-0.0278143454,
-0.0502777174,
-0.0206307694,
0.0275555979,
0.0047163037,
-0.022489354,
-0.0020129599,
-0.0212749578,
-0.0607588738,
-0.0087055545,
0.0021410557,
0.0217591133,
0.0196484383,
0.047278557,
-0.0015456586,
-0.0061893775,
-0.0170493852,
-0.0536482856,
-0.0334379822,
0.0099282032,
-0.0300815925,
-0.0060850307,
-6.362460887e-33,
-0.0493519306,
-0.025702456,
-0.0256271213,
-0.0386887863,
0.0289945137,
0.0224833153,
-0.0427511521,
-0.0065307799,
0.0099324267,
0.0139394393,
0.0020658807,
-0.0156347342,
0.023021061,
0.0107985092,
0.03415104,
0.012003541,
0.040299762,
-0.0185734238,
-0.0298034921,
0.0512764268,
0.0312887393,
0.0188253503,
0.0448846184,
-0.0385225378,
0.0169050694,
-0.0004084924,
0.0148849944,
-0.0164714847,
-0.0132846572,
0.0884704515,
-0.0045220205,
0.0244299583,
0.0086469976,
-0.0375778638,
-0.0518156029,
-0.01326501,
-0.0477827229,
-0.034963984,
0.0087121893,
-0.0032316539,
-0.0158219114,
-0.0860763788,
0.0043655322,
-0.0094238324,
-0.026397014,
-0.0100261774,
0.0395689979,
-0.0328190736,
-0.0306021161,
0.004555868,
-0.0574460439,
0.0317717977,
-0.0189617258,
-0.0168506932,
0.0294547733,
0.0204170011,
0.0302558877,
0.0160191525,
-0.0431325883,
0.0352115221,
0.0169711672,
0.0986189544,
0.0750930384,
0.0324984416,
0.0237911586,
0.039475292,
0.0224075653,
0.0231916178,
0.0412257127,
0.0220842343,
0.0120454887,
0.0502456091,
0.0164099857,
-0.024440879,
0.0356187597,
0.0147703988,
-0.0594695136,
0.0429623425,
-0.0231203735,
-0.0089597339,
0.0093659498,
0.0216531493,
-0.0575021431,
-0.0152870165,
-0.0152757773,
0.0089541283,
-0.0125595573,
-0.0105715944,
0.038247671,
0.013889771,
-0.0852187723,
0.0257611275,
0.0243274868,
-0.0052094441,
-0.0044405907,
-0.0331801027,
-0.0237902924,
0.0060985466,
0.010750561,
-0.0221849941,
-0.0555624925,
-0.0973362848,
0.0727390349,
-0.0511972606,
0.0628995746,
-0.0030210279,
0.0540723465,
0.0269162972,
-0.0681240559,
-0.0015225648,
-0.0389986634,
0.0172058549,
-0.0128824497,
0.0195526704,
-0.007836719,
-0.0004207128,
0.000177738,
0.0662115514,
0.0047208988,
-0.0038324879,
-0.0159689188,
-0.0348584056,
-0.0133831361,
0.01624812,
-0.0271358788,
0.0080146417,
-0.0154978102,
0.0170572903,
0.0110638393,
-0.0334548727,
-0.0386024639,
0.0060664415,
2.669e-7,
0.0384541415,
0.05496471,
0.0400637724,
0.0201355368,
0.0099167684,
0.0263434835,
-0.0251533855,
0.0180207584,
-0.0261917673,
-0.0804407969,
0.0186099485,
0.0209059361,
0.069320932,
0.0130952187,
-0.0489254668,
-0.0056784856,
-0.0359504148,
0.0272130985,
-0.0674811229,
-0.020770831,
-0.0140870977,
0.097950615,
0.0008077285,
-0.0065165595,
0.0054808124,
0.0093491869,
0.00799945,
-0.0362742767,
0.0155245382,
0.036304418,
0.0293960907,
0.0218093116,
-0.0375415049,
0.0120070362,
-0.0581716448,
-0.0420056842,
0.0082599856,
0.1057614759,
-0.0311853271,
-0.0310221463,
0.0402071215,
0.0025793023,
-0.0031842308,
-0.0169938356,
0.0506749265,
-0.015275036,
0.00921431,
-0.0109358998,
0.016628338,
0.0012458408,
0.0252890382,
-0.0155227752,
0.021742804,
0.0414765142,
0.0290118419,
-0.0241571292,
-0.0176614374,
-0.0759870335,
-0.0033557615,
0.0177631024,
-0.0465967618,
-0.0632935762,
-0.0481910929,
0.0269197095,
-0.0071416381,
-0.0371434055,
-0.0025868379,
2.80274817e-34,
0.0210156292,
-0.0332349315,
0.0138871064,
0.0127223544,
0.0402954929,
-0.0129138641,
-0.0420318693,
0.0145471012,
-0.0316466093,
-0.0526861213,
-0.0186307598
] |
http://arxiv.org/abs/2401.16541v1 | http://arxiv.org/pdf/2401.16541v1 | GuReT: Distinguishing Guilt and Regret related Text | 2024-01-29T20:20:44 | 2024-01-29T20:20:44 | Sabur Butt, Fazlourrahman Balouchzahi, Abdul Gafar Manuel Meque, Maaz Amjad, Hector G. Ceballos Cancino, Grigori Sidorov, Alexander Gelbukh | cs.CL | cs.CL, cs.AI | The intricate relationship between human decision-making and emotions,
particularly guilt and regret, has significant implications on behavior and
well-being. Yet, these emotions subtle distinctions and interplay are often
overlooked in computational models. This paper introduces a dataset tailored to
dissect the relationship between guilt and regret and their unique textual
markers, filling a notable gap in affective computing research. Our approach
treats guilt and regret recognition as a binary classification task and employs
three machine learning and six transformer-based deep learning techniques to
benchmark the newly created dataset. The study further implements innovative
reasoning methods like chain-of-thought and tree-of-thought to assess the
models interpretive logic. The results indicate a clear performance edge for
transformer-based models, achieving a 90.4% macro F1 score compared to the
85.3% scored by the best machine learning classifier, demonstrating their
superior capability in distinguishing complex emotional states. | GuReT: Distinguishing Guilt and Regret related Text: The intricate relationship between human decision-making and emotions,
particularly guilt and regret, has significant implications on behavior and
well-being. Yet, these emotions subtle distinctions and interplay are often
overlooked in computational models. This paper introduces a dataset tailored to
dissect the relationship between guilt and regret and their unique textual
markers, filling a notable gap in affective computing research. Our approach
treats guilt and regret recognition as a binary classification task and employs
three machine learning and six transformer-based deep learning techniques to
benchmark the newly created dataset. The study further implements innovative
reasoning methods like chain-of-thought and tree-of-thought to assess the
models interpretive logic. The results indicate a clear performance edge for
transformer-based models, achieving a 90.4% macro F1 score compared to the
85.3% scored by the best machine learning classifier, demonstrating their
superior capability in distinguishing complex emotional states. | [
0.0356477387,
0.0935479254,
-0.0063939719,
0.0053823045,
-0.0477687381,
0.0419471338,
0.015597159,
0.0084048267,
0.0781338513,
-0.0118094906,
-0.0215102676,
-0.007958414,
-0.0261401962,
0.0089297844,
-0.0553887337,
-0.0248161517,
0.0071913004,
0.0612532198,
0.0584503822,
0.0047265482,
-0.0310406592,
0.0394716784,
0.0071353712,
0.017569324,
-0.0597432032,
-0.0377354249,
-0.0109826503,
0.0308975969,
0.0674359649,
-0.0085383365,
0.0160464011,
0.0335001908,
-0.0076063075,
0.0301259812,
0.0000023976,
-0.023031868,
-0.0382588506,
-0.0192626063,
-0.041825667,
-0.0277558882,
0.0549247302,
0.0054004975,
0.0042179548,
0.0080132158,
-0.0250625554,
0.0152706271,
0.0283447597,
0.1041826531,
0.0563674457,
0.0848057866,
-0.0485903211,
-0.0577992536,
0.0215234272,
-0.0352943763,
0.0268036015,
0.0164527278,
0.0177305602,
0.0052597402,
-0.0478303619,
-0.056061767,
-0.0164079126,
0.0000619878,
0.0277853552,
0.0199420545,
-0.0376014598,
-0.0105907898,
0.0199600309,
-0.0524517894,
-0.0177601185,
0.0150170131,
-0.028428426,
-0.0114641599,
0.0004149049,
-0.0313274786,
0.0087638991,
-0.0293185636,
0.014138028,
0.0288967378,
-0.0298543405,
0.027406048,
0.0320744477,
0.0318272635,
0.0109020881,
-0.0052583241,
-0.0294477008,
0.0131641105,
-0.0439859964,
-0.0159880165,
0.0582771935,
0.0098974239,
-0.0383353196,
-0.0146128573,
0.0225710459,
-0.016857598,
-0.0119167371,
0.0021336088,
-0.0088764671,
-0.0027698092,
-0.0158782452,
-0.0307388958,
-0.0004543943,
0.0194105078,
-0.0501613207,
0.0206095632,
0.0095753828,
0.0258813798,
0.0064005842,
-0.0017119679,
-0.0098288041,
0.0009886657,
-0.038231425,
0.0091819111,
0.0063325083,
0.0389545709,
0.0113248425,
0.0134498226,
-0.0694703087,
-0.0337337963,
-0.0450352952,
0.0171969533,
0.0040262626,
0.0226421524,
0.0141495103,
0.0036794443,
-0.0319272093,
0.0319822952,
-0.0120923165,
0.0217364635,
0.0135805123,
-0.0221636221,
0.040388152,
-0.0253194217,
0.0061590248,
-0.0191012323,
0.0558802187,
0.0636190996,
0.0082826475,
-0.0115164854,
-0.0496563166,
-0.027800709,
-0.0231317524,
-0.0596346259,
-0.0105378423,
-0.0382525288,
0.0081099514,
-0.0022116059,
0.0009200246,
-0.0599552281,
0.033073958,
-0.0011580932,
-0.0377364121,
0.0353849456,
0.0271684136,
-0.0465214439,
-0.0040912703,
0.0131121036,
-0.0027216447,
0.0369712412,
0.0163726583,
0.124809131,
0.0243750326,
-0.0019023184,
0.0255986471,
-0.0330124237,
-0.0279142503,
-0.061605677,
-0.0187132303,
-0.0444248095,
-0.051133465,
0.0527438447,
-0.023688782,
0.034588661,
-0.0241192132,
-0.0118551049,
0.0115757547,
0.0367035866,
0.0280539691,
0.0935952961,
-0.0205948092,
0.0015103171,
-0.000673804,
-0.0421769358,
-0.0469666198,
0.0488375314,
0.0006823139,
-0.0403381959,
-0.0478075482,
0.0373283662,
-0.0062753232,
-0.0613212362,
-0.0080871815,
-0.0206071623,
-0.0031655263,
0.0030812728,
0.0342226811,
-0.0160040446,
-0.0590378344,
0.015493596,
0.0233916715,
-0.0748616755,
-0.0481556617,
-0.022675449,
0.0254697204,
0.0447366498,
0.0156770796,
-0.0520471036,
-0.0146667231,
-0.0426080227,
-0.0239087138,
-0.0217271056,
-0.004679082,
0.0291007217,
0.0206421837,
-0.0593330003,
-0.0121286009,
0.0051669255,
0.0029532225,
0.0376709923,
-0.0574607812,
0.0228508301,
0.0123077305,
0.0069491891,
0.0220452659,
0.0081431726,
0.0153086018,
0.0184243191,
0.0205939021,
0.0381252319,
0.0099843461,
0.0073956507,
0.0148968883,
-0.0286113806,
0.0229387395,
-0.0109767243,
-0.0508132204,
0.0213305131,
-0.0330652259,
-0.0145818675,
0.0132880602,
-0.0737148821,
-0.0457366034,
0.058453232,
-0.0368148014,
-0.013847095,
-0.0223967675,
0.0049171345,
0.0081485715,
-0.008825779,
0.0188566223,
-0.0489949472,
0.0127309076,
-0.0085046487,
0.0894828513,
-0.032891728,
0.0139695518,
-0.0438118652,
-0.0825956911,
0.0416033603,
0.0136662005,
-0.0371751599,
-0.0469818823,
-0.0256133415,
0.0037574591,
-0.0042330124,
0.0339797474,
-0.0482776016,
-0.0716218725,
-0.0771208629,
0.0157259442,
-0.0088468418,
-0.0049634124,
-0.030901622,
-0.0262998417,
0.0097265337,
0.0253998972,
-0.051195465,
0.0579006262,
0.001675677,
-0.0081491331,
0.0192110818,
-0.0521132648,
0.0273693632,
-0.0387124978,
-0.0265342016,
-0.0510187298,
0.0245792251,
0.0098636625,
-0.0101694129,
0.0145805283,
0.0412164778,
-0.0342191756,
0.020769868,
-0.0078156479,
-0.0623084866,
0.012515747,
-0.0213977322,
0.0240886211,
0.0829686224,
0.0205706079,
0.081382513,
0.004093098,
0.0013729972,
-0.0357485339,
0.0217193067,
-0.0460587852,
0.0213970151,
-0.0438513756,
-0.0236710757,
-0.0636698529,
0.0753708631,
-0.0060315058,
0.0009873296,
-0.0216219481,
-0.0296133757,
0.0030144178,
-0.0235346816,
0.012544699,
-0.0046457811,
0.0490021631,
0.054581359,
0.0136895375,
-0.0376621783,
0.0470568724,
0.0578923263,
0.0046512317,
-0.0084688021,
-0.0057850252,
-0.0069578532,
0.0861088783,
0.0068023433,
0.0037835131,
0.0255694091,
-0.0032843309,
-0.0115431594,
0.0039668051,
-0.0133496653,
-0.035058856,
0.0263028499,
-0.0221276227,
-0.029910611,
0.0066593532,
-0.0440650061,
-0.0545045845,
-0.0319256112,
-0.061528489,
-0.0407387875,
0.0322621986,
-0.0137880193,
-0.0239796452,
-0.0355145,
0.009415349,
-0.0384323299,
-0.0091352845,
-0.0542868711,
0.0118453298,
-0.0100642052,
0.0453112014,
-0.0065469476,
0.0877837837,
0.0137765938,
0.0224183574,
0.0347959027,
0.0158428345,
0.0107159074,
-0.0248739198,
0.0540334508,
-0.0054226131,
0.0082314434,
-0.0137495752,
-0.0066537582,
-0.0088716913,
-0.0108226212,
0.0019715121,
0.1074531451,
0.0335115679,
-0.0889791772,
0.0092385672,
-0.0643739402,
-0.0141389249,
0.011570516,
-0.0309206825,
-0.0935179889,
-0.0438211858,
-0.0017985569,
-0.0217833146,
-0.0277509838,
-0.0425159782,
0.063195087,
-0.0612142049,
0.0399606638,
-0.025233753,
0.0292944722,
-0.0068997554,
-0.0052827923,
-0.0270857383,
-0.0557729714,
-0.040576715,
-0.0569703504,
0.024021063,
-0.0294613559,
0.0121518299,
-0.0270714983,
0.0081858085,
0.0681411326,
-0.0129193375,
0.0189878587,
-0.0171489883,
0.0969275534,
0.0036358174,
-0.0316444337,
0.0256702676,
0.0022719239,
-0.0099293999,
-0.011847863,
0.0410959721,
-0.0032401527,
-0.0102326674,
-0.0333035067,
0.035661608,
-0.0347763114,
0.0130187385,
0.0030367854,
0.0967748463,
0.003976705,
-0.0233192015,
0.0144665912,
-0.003489713,
-0.0141331339,
0.0408901758,
-0.0035544354,
-0.0176812764,
-0.0668095723,
-0.0397129506,
0.0324089192,
-0.0028725367,
-0.0386586711,
0.0096381735,
0.0195642151,
-0.0296791457,
-0.0068392758,
-0.0536409393,
-0.0066744769,
-0.0017904657,
0.0026400939,
-0.0807954669,
0.0082184616,
0.0023481718,
-0.1000877097,
0.0063450043,
0.0116079086,
0.008706904,
0.0082751848,
0.0336396508,
-0.0047454173,
0.0090175457,
-0.0377399102,
-0.0138724046,
-0.0559825078,
0.0259041637,
0.0513100997,
0.0386463292,
0.0224876534,
-0.0044921287,
-0.0193127021,
-0.0104606999,
-0.0060467939,
0.0629610419,
-0.0263138842,
-0.0221520606,
0.0703627169,
0.0583918877,
0.0030149305,
0.0186531767,
-0.0301049501,
0.0345490277,
0.0281896703,
-0.0291484781,
-0.0312932953,
-0.0329418629,
0.0125009865,
-0.0234729778,
0.0306945257,
0.0257412419,
-0.0253135581,
-0.0195637308,
0.0051078275,
0.0065504853,
0.0079174284,
-0.0455653295,
-0.0236023981,
0.0595653728,
0.0121553317,
-0.0310298968,
-0.0461683236,
0.0069619156,
-0.0459135212,
-0.0201367177,
0.0058135558,
0.0877336636,
0.0451633334,
-0.0268834345,
0.0115816519,
-0.0363629572,
-0.0414568894,
0.0313877687,
0.0259494781,
-0.0193768032,
0.0593408346,
0.0252194479,
0.0028196585,
0.006338696,
0.0129836099,
-0.0524222068,
-0.0144459428,
-0.0040095896,
-0.0670157373,
0.0552658923,
-0.0186654218,
0.0391757302,
-0.0314306393,
0.0283979811,
0.0081959907,
0.0424042977,
-0.0314647593,
-0.042216897,
-0.0301691163,
-0.0123799704,
-0.0141413724,
-0.0060424539,
0.0136340708,
-0.0554322638,
0.0317556076,
0.0186706726,
-0.0228480883,
-0.042699486,
-0.0260358602,
0.0518962219,
-0.040730428,
0.0345636494,
-0.0230764132,
-0.0637630746,
-0.0439048633,
0.0433017872,
-0.0187461786,
0.0492904149,
-0.0231780242,
0.0014951881,
0.0458460785,
-0.0105064586,
-0.0031335831,
-0.0338171348,
0.0424271636,
-0.0430982038,
-0.0108541409,
-0.0800204277,
-6.156923128e-33,
0.023611201,
0.0074875518,
-0.0347122736,
0.0083344607,
-0.0170967914,
0.023555195,
-0.0097793695,
-0.0115867015,
-0.0158919953,
0.0756300017,
-0.0061008288,
0.0619326457,
0.0186753888,
0.0168930963,
0.0400582999,
-0.0352165885,
0.0148952641,
0.0225853417,
0.0212412607,
0.0130591756,
0.0200818945,
0.0456778258,
0.0541410521,
-0.084311083,
0.0027556261,
0.0449339449,
-0.0308512431,
-0.0121749202,
-0.0014711559,
0.0174257234,
-0.0135952299,
0.0324724838,
0.0284346305,
-0.0400305465,
0.0267484374,
0.0045684129,
-0.060025841,
-0.0280615352,
-0.0326901041,
0.0026326627,
-0.0069017988,
-0.035206534,
-0.0059092157,
0.0372999348,
0.0095272725,
-0.0708990395,
0.0449949987,
0.0270374045,
0.0201141685,
-0.0903245732,
-0.0222210661,
0.0082536144,
-0.0393278934,
0.0002124674,
-0.0140333762,
0.0744743347,
-0.0185308401,
0.0278983209,
-0.0568165742,
-0.0046591014,
-0.0219786875,
-0.0121093243,
-0.0103152702,
-0.0481404401,
-0.0199450869,
0.0773582831,
0.0252469257,
0.0074820244,
-0.0699400455,
0.0920814499,
0.0161717832,
0.0510362312,
0.0971657634,
0.0316320024,
0.0179721508,
-0.0352305807,
-0.0601764135,
0.0198839456,
0.047824882,
0.0303551033,
0.0286769755,
0.0047174529,
0.022707073,
0.0213645007,
-0.0151823731,
-0.0462344363,
-0.0457587838,
0.0024464894,
0.0025190411,
-0.0511286668,
-0.0173364021,
-0.0225750487,
0.0067673274,
-0.0394435339,
0.0104448171,
0.0584650673,
0.0157858972,
0.0704243556,
0.016182201,
0.0377624705,
-0.016917685,
-0.0538919531,
0.0294772778,
0.0298424624,
0.105482921,
0.0016694731,
0.0437541865,
0.0191309024,
-0.0255557876,
-0.021267239,
-0.028854724,
0.0137838293,
-0.0146656074,
0.0110209594,
-0.0255136508,
0.0225907229,
0.0049306871,
-0.0196826831,
0.0006547515,
0.0594377965,
-0.0192514565,
0.0028331641,
0.0155042466,
0.0199552868,
-0.0114532756,
-0.0284028463,
-0.0369441845,
-0.0192684922,
0.0120598609,
-0.0478295162,
-0.0136432257,
-0.0317995772,
3.012e-7,
-0.0024043862,
0.0398723297,
0.0441235714,
-0.0321332663,
0.0046051103,
0.0767586306,
0.0453903601,
0.0120193576,
0.0386952832,
-0.0040801689,
0.0621688403,
0.0002448424,
0.0700923949,
0.0030641237,
-0.010495441,
-0.0142673338,
0.0044813971,
-0.0247249026,
-0.0298138671,
0.0494224876,
0.1137549281,
0.1102320775,
0.0630359277,
-0.0316286758,
0.0416970924,
0.0818094611,
0.0123327319,
-0.0718675181,
0.0299574453,
-0.0058087842,
0.05514431,
-0.0419220403,
0.0067652664,
-0.0143067427,
0.0366754644,
-0.0345546715,
-0.0026298573,
0.0185050722,
-0.0133774132,
-0.0417021364,
0.0281219538,
0.015919935,
-0.0160607938,
-0.0335954987,
0.0477405228,
-0.0597964004,
-0.0231883451,
-0.0248256605,
0.0016212187,
-0.0064446772,
0.0620938428,
0.0353690572,
-0.0053312327,
-0.0035912499,
0.0069815167,
-0.0040327609,
0.0007777496,
-0.0733394399,
-0.0010380886,
0.0588659234,
-0.0067128222,
0.0323438868,
-0.0291346144,
0.0414985754,
0.06620235,
-0.03369838,
0.0085910754,
2.358235441e-34,
0.0200485885,
-0.0096125193,
0.0024340362,
-0.0070030745,
0.0578231364,
0.0186596774,
-0.0671113133,
0.0117422501,
-0.0170094427,
-0.0382544845,
-0.0074768332
] |
http://arxiv.org/abs/2401.16475v2 | http://arxiv.org/pdf/2401.16475v2 | InfoLossQA: Characterizing and Recovering Information Loss in Text Simplification | 2024-01-29T19:00:01 | 2024-06-04T13:36:07 | Jan Trienes, Sebastian Joseph, Jörg Schlötterer, Christin Seifert, Kyle Lo, Wei Xu, Byron C. Wallace, Junyi Jessy Li | cs.CL | cs.CL | Text simplification aims to make technical texts more accessible to laypeople
but often results in deletion of information and vagueness. This work proposes
InfoLossQA, a framework to characterize and recover simplification-induced
information loss in form of question-and-answer (QA) pairs. Building on the
theory of Question Under Discussion, the QA pairs are designed to help readers
deepen their knowledge of a text. We conduct a range of experiments with this
framework. First, we collect a dataset of 1,000 linguist-curated QA pairs
derived from 104 LLM simplifications of scientific abstracts of medical
studies. Our analyses of this data reveal that information loss occurs
frequently, and that the QA pairs give a high-level overview of what
information was lost. Second, we devise two methods for this task: end-to-end
prompting of open-source and commercial language models, and a natural language
inference pipeline. With a novel evaluation framework considering the
correctness of QA pairs and their linguistic suitability, our expert evaluation
reveals that models struggle to reliably identify information loss and applying
similar standards as humans at what constitutes information loss. | InfoLossQA: Characterizing and Recovering Information Loss in Text Simplification: Text simplification aims to make technical texts more accessible to laypeople
but often results in deletion of information and vagueness. This work proposes
InfoLossQA, a framework to characterize and recover simplification-induced
information loss in form of question-and-answer (QA) pairs. Building on the
theory of Question Under Discussion, the QA pairs are designed to help readers
deepen their knowledge of a text. We conduct a range of experiments with this
framework. First, we collect a dataset of 1,000 linguist-curated QA pairs
derived from 104 LLM simplifications of scientific abstracts of medical
studies. Our analyses of this data reveal that information loss occurs
frequently, and that the QA pairs give a high-level overview of what
information was lost. Second, we devise two methods for this task: end-to-end
prompting of open-source and commercial language models, and a natural language
inference pipeline. With a novel evaluation framework considering the
correctness of QA pairs and their linguistic suitability, our expert evaluation
reveals that models struggle to reliably identify information loss and applying
similar standards as humans at what constitutes information loss. | [
0.0272770245,
0.0408105217,
-0.0271955132,
0.0365291908,
-0.0576127917,
0.0006826447,
-0.026546916,
0.03899258,
-0.0175539609,
-0.0016380697,
-0.005001782,
0.0223191511,
0.0554793924,
-0.0241093244,
0.0437755734,
-0.0667213574,
0.039024204,
-0.0093812216,
0.0127842603,
0.0012060662,
0.0297071971,
0.0112625677,
0.0172889847,
0.0323653892,
0.0413712338,
-0.0236680228,
-0.0013414719,
0.0139557431,
0.0013228231,
0.0109799746,
0.0122093391,
0.0716063902,
-0.0105539206,
-0.0054710922,
0.0000020979,
-0.049389109,
-0.0154884486,
0.0279455464,
-0.0276410338,
0.0567355119,
0.0523443632,
-0.0037162912,
0.0120105045,
0.0146705769,
-0.0342393182,
-0.0123710828,
0.0211410243,
0.0312497746,
0.0532188676,
0.0536049232,
-0.0085690971,
-0.0235385448,
0.0242613014,
0.0075641167,
0.0356452316,
-0.0111413877,
0.0254613161,
0.0039475104,
0.0230954383,
0.0175430123,
0.003668962,
0.0698780194,
0.0315656848,
-0.037026789,
-0.0011653169,
0.0120389042,
-0.0099406261,
-0.0229655821,
0.010528977,
0.0674545616,
0.0011417981,
-0.0584777221,
0.0048739272,
-0.0103444709,
0.0200395118,
0.0485761054,
-0.0760488585,
-0.0196568146,
-0.0145211853,
-0.0135059776,
0.005988162,
-0.0294020548,
0.007069537,
0.0228931345,
0.0176289119,
0.0200005397,
0.0224344078,
-0.0060950206,
-0.0201824233,
-0.0261473767,
-0.0189967807,
-0.0695502982,
0.0429109186,
0.0520264097,
0.0090449285,
0.0307156947,
-0.0231536031,
-0.0974657238,
0.0346316323,
-0.0346789695,
-0.0292805061,
0.0194586925,
-0.0202820301,
0.0386703536,
-0.0455254577,
0.0307967495,
0.0179720521,
0.0159341823,
-0.0481475331,
0.011886647,
0.0146321245,
-0.0385295041,
-0.0305279288,
0.0484667569,
-0.0269333441,
-0.0040740282,
0.0061405352,
-0.0136776287,
0.009801236,
-0.0098619079,
-0.0330817737,
0.0072126123,
-0.0246851221,
-0.0366290733,
-0.0449513197,
0.0570472293,
-0.0256410856,
0.0407807566,
-0.0082855886,
-0.0799823627,
-0.0256198104,
0.0337458439,
0.0139071336,
0.0107666766,
-0.0107351737,
0.0194840543,
-0.0370048881,
-0.0077235429,
-0.025204232,
-0.041303087,
0.0296788383,
-0.0348548703,
-0.0105581544,
-0.0673109517,
0.0528612547,
0.0099554714,
-0.0314399078,
-0.0774755329,
-0.0046640015,
0.0318022482,
-0.0474854782,
0.0661668926,
-0.0793854892,
0.0382382572,
0.0637962595,
0.0149181867,
0.0128088286,
0.0323686264,
0.0275272168,
-0.0480737053,
0.0376098529,
-0.0242818706,
0.0444258228,
0.0294029936,
0.0371859483,
0.0072057317,
0.0389053263,
0.0152918082,
-0.0524655133,
0.046829436,
0.0296748374,
0.0906692073,
-0.0604649186,
0.0024704949,
0.0427934937,
0.0657990202,
0.0018550631,
0.0211262219,
0.0162586533,
0.0305565987,
0.0597458035,
0.0490964614,
0.0048890137,
0.0193680879,
-0.0192605909,
0.0131874969,
0.0491789319,
-0.0329906158,
-0.0226376206,
-0.0716535226,
0.006179221,
-0.0015277353,
0.0425285883,
-0.0091388524,
0.0307377372,
0.0462714322,
-0.0400751084,
0.0422062427,
0.0163484421,
-0.0606440008,
0.0363910682,
-0.0263200384,
0.0337784663,
0.0419980437,
0.046353735,
-0.0355398916,
-0.031321872,
-0.0021443919,
0.0048107207,
-0.00577405,
0.0327007659,
-0.0323403627,
0.0110034402,
-0.0356922634,
0.0061796932,
0.0061702593,
0.0187215228,
-0.0283110123,
0.0139972093,
-0.035616342,
0.0050228671,
-0.059915904,
-0.0434640385,
0.0359363332,
-0.0421857238,
-0.0056228223,
0.0002109464,
-0.014701196,
-0.01669994,
0.0211350899,
0.0250086486,
0.0463870391,
0.0386447273,
-0.0184206087,
-0.0233487803,
0.0091935443,
-0.0012667108,
-0.0483245477,
0.0194921531,
-0.0676453114,
-0.0037341646,
0.0342437327,
-0.0371861868,
-0.0321158543,
-0.0064955223,
-0.0747215673,
0.008324828,
-0.0219531879,
0.018143123,
0.0261932798,
-0.0287006591,
-0.0453190655,
0.0371902101,
0.0094886012,
0.0535349771,
0.0566559061,
-0.0547287166,
0.0088682016,
-0.0313920192,
0.0378942974,
-0.0368485451,
-0.0298759528,
-0.0140974624,
0.0429845452,
-0.0014502283,
-0.0390978903,
0.0140023921,
-0.091233775,
-0.0165804047,
-0.011723293,
0.056868393,
0.053109061,
0.0134792794,
-0.0060816598,
-0.0209457967,
-0.0385650992,
-0.0411339514,
-0.0041590598,
0.017587252,
0.0756143853,
0.0218928419,
0.0226934403,
-0.0420260318,
-0.0102088582,
-0.0153802987,
0.0409033932,
0.0388335176,
-0.0534751341,
-0.0641805306,
0.0249721445,
-0.0223535411,
-0.0288086962,
0.0012193013,
0.0004586612,
-0.0180160273,
0.0006397421,
0.0490721501,
0.0719126761,
0.0059358515,
-0.0104692876,
0.0337397642,
0.0326163806,
-0.0671862364,
0.0050175223,
-0.0380157344,
-0.0209046528,
0.0064845416,
-0.0750789493,
0.0431109965,
0.0342672542,
-0.0041465126,
0.0151310731,
-0.0527870841,
0.0209098235,
-0.0357831307,
-0.0232500434,
-0.0377102979,
-0.0345369242,
-0.0191096645,
0.0815721005,
-0.0122266812,
-0.0356094092,
0.0723154917,
0.0099059856,
0.0004459807,
-0.0843224451,
0.0182585381,
-0.0475350991,
0.0131352879,
-0.0419301949,
0.0120692486,
0.0789906606,
-0.0198048949,
-0.0840041488,
-0.0043166643,
0.009725607,
-0.0334995799,
0.0358215719,
-0.0946592912,
0.0033149442,
-0.0148801934,
-0.0281746909,
0.0156870708,
-0.0355107524,
0.009217469,
-0.0163327865,
0.0458838157,
-0.0080688419,
0.0035107606,
-0.0373616032,
0.0428845473,
0.0564852767,
-0.0180716012,
-0.0010793882,
-0.0106327832,
0.0627313256,
0.0621492676,
0.0172336213,
0.055501312,
-0.0568977483,
0.0081456127,
0.0725393593,
-0.0349513777,
0.0146431727,
0.0279552974,
-0.0484111421,
0.0006345817,
-0.03361829,
-0.0284251943,
0.0126853706,
0.0544832498,
-0.074595809,
0.004563509,
0.0813702568,
0.0154999709,
-0.0190067776,
0.0081306659,
-0.0445763469,
0.0200019982,
-0.0247911382,
0.0060101212,
-0.1082237065,
-0.0265768357,
0.015902482,
0.0213992763,
0.0072523127,
0.0003527527,
-0.0009995605,
-0.0507409908,
0.0198257156,
-0.0037507163,
-0.0456072539,
-0.0182519183,
-0.0615159161,
0.024522705,
0.0085511906,
-0.005718933,
0.0267372895,
-0.0121260621,
0.0373273268,
-0.0017525873,
-0.0908168033,
-0.0457091965,
-0.0521738306,
-0.0049550985,
0.009059187,
0.020340398,
0.0343306288,
0.0074147154,
-0.0228441097,
0.0503437482,
-0.0149461636,
-0.0182400942,
-0.0311012324,
0.0697160438,
0.0369062424,
0.0217474625,
-0.0275856499,
0.0732307807,
-0.0154140815,
-0.0053365142,
0.0068872194,
0.0422609709,
0.0209002811,
-0.0317672193,
-0.0143675692,
-0.0009018746,
-0.0251341667,
0.100195162,
0.0588035285,
0.011859538,
-0.0403705463,
-0.0204265248,
0.02694677,
-0.0301801227,
-0.057915356,
-0.0170007721,
0.0024123646,
-0.0448003039,
-0.0931546837,
-0.0457223579,
0.0352664031,
-0.0465711169,
0.0225836728,
0.0335963257,
0.0160032436,
-0.0123762665,
-0.0350633822,
0.0301836748,
-0.0065487423,
0.0193527285,
-0.0305467956,
0.0102635203,
-0.0124813942,
0.0013892951,
-0.0010743986,
0.0334711485,
-0.0928704292,
-0.0651940405,
-0.0412339643,
0.0464436039,
0.0286667496,
0.0376294702,
-0.001928906,
-0.0458844081,
0.0092727933,
0.0343735702,
-0.0439805016,
0.0487753898,
0.0238505974,
0.0146554243,
-0.0105867106,
0.0382573642,
-0.0185565259,
-0.0423897468,
-0.0251136832,
0.0237476509,
-0.0441940762,
-0.0292639975,
0.0262834374,
0.0486010015,
-0.003455132,
0.013397322,
0.0157617815,
-0.0080613047,
0.0201747622,
-0.0376070291,
-0.0027989396,
-0.0413573422,
0.0293312389,
0.0105160801,
0.0171104614,
-0.0081248125,
0.0304307546,
-0.0209815148,
-0.005618419,
-0.0138308974,
-0.0056604645,
0.006422373,
0.017191546,
0.0270881951,
-0.0261541121,
0.0001875991,
-0.0252071023,
-0.0513064042,
0.0213021431,
0.0045854831,
-0.0005186262,
0.0266097169,
0.016860703,
0.0056745117,
0.0183254275,
-0.024657933,
-0.1003805175,
-0.0001747077,
0.0169700626,
0.0021032398,
0.0449387096,
0.0735734403,
-0.0219534673,
-0.0208040345,
-0.0358696729,
0.0462542884,
0.0061765993,
0.0058081709,
0.0637455061,
0.006198003,
-0.0018682104,
0.003266756,
0.0948809907,
-0.0311646666,
0.0351871774,
-0.0139840776,
-0.0079136901,
-0.0283504296,
-0.0574889556,
-0.0146264154,
0.0300876461,
-0.0216986258,
-0.0242295936,
-0.0624820217,
-0.0385288708,
-0.0085916501,
0.0075137024,
0.0004569787,
-0.0484959669,
0.0388593487,
0.0413526744,
-0.0129268644,
-0.0495289229,
-0.0202882867,
0.0436884426,
-0.0424298681,
-0.0376505144,
-0.0456498861,
-6.415156094e-33,
0.0135334842,
-0.0222434159,
0.0037616971,
-0.0076734964,
-0.0348729081,
0.0356604643,
-0.0864261165,
0.0009756667,
-0.0132590998,
0.0424962528,
-0.0184790418,
0.0241112094,
-0.0080479076,
-0.0051769628,
-0.0213450938,
-0.025008833,
0.0466974191,
0.0020060327,
-0.0344914459,
0.0233768579,
-0.0138600366,
0.0083054956,
0.0343124457,
-0.0403666981,
0.0402432829,
-0.0148454886,
-0.0242959727,
0.0207866021,
-0.0163463391,
0.0374991745,
-0.0071851574,
0.0314544849,
0.0068273405,
-0.0402829722,
-0.0316502526,
0.0558789335,
-0.0983122066,
-0.0210459959,
0.0556182824,
-0.0063864021,
-0.116167441,
-0.054228723,
0.0868904367,
-0.0029216893,
0.0146984356,
-0.0010938968,
0.0458893254,
-0.0266651306,
0.0014786565,
-0.0536892191,
-0.0265256464,
0.0314182937,
-0.0326284952,
-0.0345052667,
-0.0079787485,
0.0534252636,
-0.0341937095,
-0.0065317699,
-0.0113438945,
0.0210437309,
0.0827905685,
0.0529148765,
0.0341985784,
-0.0636319444,
0.0121275168,
0.0084009459,
0.0064099161,
0.008027182,
0.0007000557,
-0.0062160506,
0.0110342382,
-0.0012675952,
0.0754958466,
0.0254625753,
0.0400262289,
0.0168906599,
-0.0453920364,
0.0364716612,
-0.0347808562,
0.0894566923,
-0.0290380269,
0.0172210187,
-0.0014895465,
-0.0360025465,
-0.0592866689,
0.0334182866,
-0.0087229479,
-0.0330183469,
-0.0151149798,
-0.0147641785,
-0.036356885,
-0.013152875,
-0.0025962184,
-0.0249159168,
0.0049149441,
0.0000559922,
0.0061720405,
0.0498416424,
0.0179425385,
0.0033770816,
-0.073410742,
0.0615531988,
0.0059432513,
0.0204756223,
0.0361925662,
0.0222437978,
0.0138684446,
0.004603941,
-0.0400789045,
0.0066218507,
0.0000040332,
-0.0347602963,
-0.0004925511,
-0.0212863702,
0.0189560764,
0.0736726001,
0.0111072231,
0.0269027818,
0.0162856672,
-0.0066749281,
-0.0550639145,
0.010344957,
0.0096930629,
0.0187964607,
-0.0278678183,
-0.0058031017,
-0.0284036808,
0.0292443261,
0.0324459188,
-0.0554128215,
-0.0098710814,
0.0111112446,
2.789e-7,
0.0894642547,
0.0469730683,
0.059197966,
0.0025140403,
0.0136431009,
0.0019392647,
0.0276221149,
-0.0099422187,
0.0183012616,
-0.0007453898,
0.0333379395,
-0.0044241673,
0.0338610411,
0.0019232841,
-0.1184761673,
-0.0095714852,
-0.0467446484,
-0.0297360756,
-0.0415013395,
-0.0147201838,
0.0522775762,
0.0281534828,
-0.0245062634,
0.0043962407,
-0.0242877267,
0.0015037779,
0.0465074889,
-0.0467264056,
-0.0006544113,
0.0582641661,
0.0490103438,
-0.0188281145,
0.0333096199,
0.0135654872,
-0.0296527483,
-0.0431387648,
0.0243639592,
0.0626140237,
0.0084334882,
0.0104482099,
-0.0154989781,
-0.0290495045,
0.0076752873,
-0.0589981377,
0.0764794201,
-0.0503463708,
0.0037101971,
0.0068510389,
-0.0458665974,
-0.0196148697,
0.029812796,
0.006400275,
-0.0582696684,
0.0419125818,
0.0033975893,
0.0225397777,
-0.0162957069,
-0.0540172532,
-0.0474488139,
0.0103324689,
-0.00148264,
-0.0325841419,
-0.0335683934,
-0.0162104778,
-0.0206719507,
-0.0816958919,
0.0053644013,
2.477499904e-34,
0.0206141267,
-0.0116322599,
0.0100382781,
0.0526401326,
0.0126760621,
-0.0286527108,
-0.0024934213,
0.0325795673,
0.0279966984,
-0.0541061424,
0.0032124233
] |
http://arxiv.org/abs/2401.16421v2 | http://arxiv.org/pdf/2401.16421v2 | Two Stones Hit One Bird: Bilevel Positional Encoding for Better Length Extrapolation | 2024-01-29T18:59:07 | 2024-06-17T06:18:13 | Zhenyu He, Guhao Feng, Shengjie Luo, Kai Yang, Liwei Wang, Jingjing Xu, Zhi Zhang, Hongxia Yang, Di He | cs.LG | cs.LG, cs.AI, cs.CL, stat.ML | In this work, we leverage the intrinsic segmentation of language sequences
and design a new positional encoding method called Bilevel Positional Encoding
(BiPE). For each position, our BiPE blends an intra-segment encoding and an
inter-segment encoding. The intra-segment encoding identifies the locations
within a segment and helps the model capture the semantic information therein
via absolute positional encoding. The inter-segment encoding specifies the
segment index, models the relationships between segments, and aims to improve
extrapolation capabilities via relative positional encoding. Theoretical
analysis shows this disentanglement of positional information makes learning
more effective. The empirical results also show that our BiPE has superior
length extrapolation capabilities across a wide range of tasks in diverse text
modalities. | Two Stones Hit One Bird: Bilevel Positional Encoding for Better Length Extrapolation: In this work, we leverage the intrinsic segmentation of language sequences
and design a new positional encoding method called Bilevel Positional Encoding
(BiPE). For each position, our BiPE blends an intra-segment encoding and an
inter-segment encoding. The intra-segment encoding identifies the locations
within a segment and helps the model capture the semantic information therein
via absolute positional encoding. The inter-segment encoding specifies the
segment index, models the relationships between segments, and aims to improve
extrapolation capabilities via relative positional encoding. Theoretical
analysis shows this disentanglement of positional information makes learning
more effective. The empirical results also show that our BiPE has superior
length extrapolation capabilities across a wide range of tasks in diverse text
modalities. | [
-0.0085867327,
0.0689692423,
-0.0194169357,
0.0123907635,
-0.0268283896,
-0.0054463632,
-0.0291394275,
0.0439854972,
-0.026553968,
-0.0170564391,
0.0120249288,
-0.0344179086,
0.0353846401,
0.0234251786,
0.04889635,
-0.0843513831,
0.0287469234,
-0.0186877828,
0.0266895909,
0.0289149955,
0.0212377273,
-0.0223429482,
0.0044353795,
0.0472592823,
-0.0094671519,
0.0256792214,
-0.052756764,
-0.0486357585,
0.0319955684,
-0.0310506783,
0.0075988858,
0.0164333377,
0.0462270193,
-0.0017488759,
0.0000019259,
-0.0058561834,
-0.0452720337,
-0.0120075084,
0.0035517607,
-0.0311796572,
-0.0086418837,
-0.087521404,
0.0204580985,
-0.0095035275,
-0.0366850756,
-0.0076418542,
0.0457938053,
0.0146279102,
0.0023062772,
0.0631936342,
-0.0314652584,
-0.0410628021,
0.0100195659,
-0.0378209315,
0.0293533728,
0.0173441395,
0.0144928005,
0.0228671655,
0.0036461516,
0.0713498741,
-0.0456346981,
0.0097957449,
0.035667792,
0.0312750116,
0.0149459764,
0.0713922828,
0.0180738568,
-0.0149292899,
-0.0125679998,
0.0314709544,
-0.0700115263,
-0.0271189995,
-0.0223866515,
0.0119822249,
0.032629814,
-0.0654361323,
-0.0741807669,
0.0298590232,
-0.0006520034,
-0.0009562947,
0.013959337,
0.0680777654,
0.0002934555,
-0.0295647029,
-0.0393398069,
0.0310986582,
-0.0063276216,
-0.0508431122,
-0.0262874439,
-0.0005959099,
-0.0415180661,
-0.0047948612,
0.0610291921,
0.0462391004,
0.005570814,
0.0527437963,
-0.03884238,
-0.013802141,
0.0069366689,
-0.044580657,
0.033419136,
0.054571975,
-0.0116615202,
-0.0034257406,
-0.0777498186,
0.0539868772,
-0.0047637159,
0.0500817858,
-0.0311472677,
0.0249380935,
-0.0218178965,
-0.018746784,
-0.046912469,
0.0162013341,
-0.0631835237,
-0.05108165,
-0.0173365735,
0.0288585313,
0.021825714,
0.0224794671,
0.03592062,
0.0292124785,
-0.0181828197,
0.0015523314,
0.0068613403,
0.0060741277,
-0.0745121688,
0.0005398136,
0.0029085174,
0.0246775057,
-0.0308497492,
0.0139956735,
0.0153799988,
-0.0451583825,
-0.015652094,
0.0620857328,
-0.0259760395,
-0.0389032885,
0.0232278015,
-0.0335787795,
-0.0432137586,
-0.0352829248,
0.0041136798,
-0.0469283201,
0.0293403286,
0.0095042819,
-0.0238405913,
-0.0324392654,
0.0126486188,
0.0199317094,
-0.0184850339,
0.0838903636,
-0.029915357,
-0.0142782237,
-0.0141931111,
0.0342193432,
0.0383202992,
0.0610970668,
0.0288794264,
0.0331912562,
0.0418757685,
0.0128833596,
-0.0063423091,
0.0322174914,
0.0280290265,
-0.0446509197,
-0.0409473032,
-0.006571719,
0.0130069051,
0.0179260504,
0.0203150138,
0.038762372,
-0.0418426804,
-0.0531416126,
0.0316697992,
0.081515193,
0.0214220416,
0.0234725699,
0.0543224737,
-0.0162027311,
-0.0142171206,
0.0199840236,
-0.0130815553,
0.0441591814,
0.0383248515,
0.0512665734,
0.015701402,
0.0123273125,
0.006350568,
-0.0314072706,
-0.0448678546,
-0.0225322209,
0.0660889596,
-0.0415831842,
0.022921687,
0.0027499115,
-0.0670701191,
0.0690661818,
-0.0235338509,
-0.0332903676,
-0.0265050698,
0.0028238161,
-0.002025208,
0.0520933978,
0.010435598,
-0.0413663425,
-0.0282830372,
0.0004033488,
-0.0290000066,
0.0109863346,
-0.0210759044,
-0.0226636734,
0.0237600803,
-0.0522738397,
-0.0226441436,
0.0162739214,
0.0056680017,
0.0387816951,
0.064299278,
0.0066989027,
0.0386149846,
-0.059657868,
0.0251832772,
0.0317820832,
-0.0408870429,
-0.0155700678,
-0.0422386564,
0.0141577199,
-0.0327628888,
0.0009587381,
0.0346365832,
0.0154991737,
0.0706370622,
-0.0415695943,
0.0300748982,
-0.0109077049,
-0.0998268053,
-0.0653466433,
0.0648432449,
-0.0192322545,
0.0045678136,
-0.0132277748,
-0.0129692089,
-0.02391183,
0.0262495764,
-0.0340226442,
0.0417554826,
-0.0086101191,
0.0021086158,
0.0189579688,
0.0227444991,
-0.0197839085,
0.0580276474,
-0.0325909778,
0.0303550269,
0.0036861603,
-0.0199497975,
-0.0098717818,
-0.0176758636,
0.0120662907,
0.0202599764,
0.0064359498,
-0.0160636026,
0.020000089,
-0.0189792067,
-0.0382694378,
-0.0328445323,
-0.1168742776,
-0.0362586714,
-0.0130549259,
0.0088982014,
0.0134500302,
0.0088774459,
0.0078213392,
-0.0268129464,
-0.064665854,
-0.0420089215,
-0.0299736559,
0.0046101739,
0.0252925679,
0.0483387746,
-0.0164354015,
-0.0291156881,
-0.018305324,
-0.0156430919,
0.0369722433,
-0.0366298445,
-0.0356661715,
-0.088180095,
-0.037077006,
-0.0264136791,
0.0191799514,
-0.0327551849,
-0.0451572873,
-0.0182694215,
-0.0581852049,
0.0243135914,
0.0307376478,
0.0153367957,
0.0448695943,
0.0400495194,
-0.0159342401,
-0.0303564686,
-0.0018575146,
-0.0524226725,
0.0687607005,
-0.0623198599,
0.0140013471,
-0.0503231585,
0.081698291,
-0.0262826364,
0.0109432731,
-0.0492233485,
0.0002851686,
0.0067683593,
-0.0450191237,
-0.0363057032,
-0.0413467027,
-0.0406937227,
0.023764547,
-0.049956724,
-0.0558199584,
0.0438317843,
-0.0363900661,
-0.0188960452,
-0.0430908278,
-0.0286094137,
-0.0385641344,
0.0923723429,
0.0339984633,
0.0304582529,
-0.0286478028,
-0.0442638174,
-0.0552294627,
0.0486637503,
-0.0427757129,
-0.0328535959,
0.0483913757,
-0.0274542589,
-0.0182634108,
0.0073897019,
-0.0145030329,
-0.0456270278,
-0.0115137184,
0.0245312527,
0.0216079559,
0.0447789691,
-0.0467515253,
0.0157298651,
0.0410136543,
0.0141376955,
0.0628487691,
-0.0233731102,
0.0160482526,
-0.0366616845,
0.0375061929,
0.0227175076,
-0.0027935903,
-0.0210552979,
-0.0439275652,
-0.0163675006,
0.0465034842,
0.0149809029,
-0.0296357721,
-0.0348948352,
0.026417857,
0.010177617,
-0.0423794352,
-0.0141618503,
0.0133431517,
0.0339160971,
-0.0445240662,
-0.0568283834,
0.1095126793,
0.035156332,
-0.0406125076,
0.0171787087,
-0.0087985229,
0.0647747219,
0.015744919,
0.0137820998,
-0.0277139097,
-0.0122114355,
0.0109807402,
-0.0234271511,
0.0034024832,
-0.0161367524,
-0.0300990343,
-0.1018110141,
-0.0210269503,
0.0103081614,
-0.0184977241,
-0.0475082211,
-0.0105621517,
-0.0128256185,
0.0512375385,
0.0120953023,
0.0175709147,
-0.0242040213,
-0.0011983789,
0.0417705998,
-0.0639016554,
-0.0072449204,
-0.0348408855,
0.024777092,
-0.0324931704,
0.0210514013,
0.0394212157,
0.0064267679,
-0.001356674,
0.0330650508,
-0.0154281156,
0.0210927986,
0.0113893626,
0.0751389861,
-0.0059817042,
0.0138411149,
0.0380635932,
0.0205438901,
-0.0084249079,
-0.012151951,
-0.0148435244,
0.078449361,
0.0079049012,
0.0275905747,
-0.0351362862,
-0.0015522449,
-0.0122450795,
0.0247584265,
0.0311320461,
-0.0122486055,
-0.022453256,
-0.0207041912,
0.0085697239,
-0.0423442833,
0.0072405902,
0.0150663536,
0.0013093678,
-0.0092493612,
-0.0715129524,
-0.0472098626,
0.0403654426,
0.0910898745,
0.0098642502,
-0.0051015979,
-0.0139229838,
-0.0303216223,
-0.0871507376,
-0.0487785786,
0.0155615741,
0.0165258627,
0.0166816339,
0.0238235798,
-0.0061739855,
-0.0085024647,
-0.0178056974,
0.0239648651,
-0.062025059,
-0.0514230318,
-0.0067599933,
0.0702883154,
0.018820839,
0.098867774,
-0.0050560101,
-0.0004066354,
0.0841791928,
0.0103697628,
-0.0445423201,
-0.0004488495,
0.0379275903,
0.0075796498,
0.0105833309,
0.0122678867,
-0.0020333428,
-0.0161700416,
-0.0076175006,
0.0385550037,
-0.0355461389,
0.0053752046,
0.0506026335,
-0.0212908126,
0.0473366566,
-0.0106230043,
0.0525310338,
-0.0029787712,
0.0226659421,
-0.0549592748,
0.0200602077,
-0.0175313149,
0.0261831749,
0.0369979739,
0.0051279967,
0.003819319,
0.026001567,
0.0292889345,
-0.0324025415,
-0.0012159768,
0.0394652039,
0.0086093135,
-0.0065661618,
0.0195211377,
0.0439467952,
0.0041091945,
-0.0072214538,
-0.0219492167,
0.0542214587,
-0.0501507446,
0.0341940895,
0.0248253886,
0.0151290689,
-0.0320486091,
0.0402921885,
-0.0664526522,
0.0375496261,
-0.0445331857,
0.0016562889,
0.0080874478,
-0.0298417266,
0.0375864059,
-0.0174789503,
0.004379014,
0.0120173311,
0.0175258834,
-0.0338901989,
-0.065734379,
-0.006431011,
0.0181431007,
-0.0254882779,
-0.0096635902,
0.0609293804,
-0.038241975,
0.0245482363,
-0.0442482196,
-0.0438058153,
-0.0367347114,
-0.0034389591,
0.0151211666,
0.0169080254,
0.0035703639,
0.0003924877,
-0.000107193,
-0.0413700268,
0.0089354459,
0.0036901964,
0.0047633103,
0.014194346,
0.0299793482,
0.0433599129,
-0.0258658323,
0.0001039867,
-0.0289355014,
-0.0202391259,
0.0131694172,
-0.0377264097,
-0.0020020097,
-5.631871679e-33,
-0.0172616318,
-0.0198480375,
0.0134503543,
-0.0913401693,
-0.0531609468,
0.0232457537,
-0.0205288976,
0.0262572132,
0.0553394109,
0.0050540897,
0.0098765409,
-0.0408335514,
0.0209170394,
0.0218181312,
0.0236598272,
-0.0005639481,
0.0242495462,
0.0197138097,
0.0013783721,
-0.0252884161,
0.0640266165,
0.0570219792,
0.082839027,
-0.0181486495,
0.0297268704,
0.0303175915,
0.0158505738,
-0.0066514709,
0.0449633375,
0.090275839,
-0.0009812576,
0.0087201903,
0.0235349834,
-0.0361479819,
-0.0457691662,
0.0287757553,
-0.0938181803,
-0.0240819491,
0.0017308004,
0.0221617632,
-0.039269954,
-0.0088386228,
0.0429224744,
-0.0403801873,
0.0086053964,
-0.0299484283,
0.0433833338,
-0.040022824,
-0.0080708154,
-0.0222426355,
-0.0643427446,
0.0136917988,
-0.008230296,
0.0335807502,
0.0817033574,
0.0592943244,
0.0297280289,
-0.0698275492,
-0.0197385475,
0.0054836161,
0.0062454008,
0.0760296211,
0.0192444734,
0.0194469225,
-0.0073402394,
0.0393697917,
0.0532218888,
-0.0851956308,
0.0011226755,
0.0550059229,
0.0330537073,
-0.0194960721,
0.0226508956,
-0.0042090924,
0.0178858563,
0.003621022,
-0.014434346,
0.007878582,
0.0022255615,
-0.0060065156,
0.0097575104,
0.0470639952,
-0.0316268392,
-0.0059159431,
-0.0126728667,
-0.0306150205,
0.0016169009,
-0.0745170861,
0.04469328,
0.0448936298,
0.0175755173,
0.054102648,
0.0686921403,
-0.0187642034,
-0.0139520932,
0.0186076425,
-0.0297428519,
0.0360951759,
0.0133854924,
0.0145919202,
-0.0098182661,
-0.0456245132,
0.0203593802,
0.0013691325,
0.0139494631,
-0.0256874748,
0.0519762821,
-0.0060767732,
-0.0542239174,
0.0029997085,
-0.0367276222,
-0.0154372621,
0.0039487961,
0.0488316789,
0.0768325552,
0.0024582176,
-0.0265273973,
0.0719093308,
0.0041615078,
0.0225202963,
0.001892008,
-0.0437397622,
-0.0032492625,
0.0413565114,
-0.0443465896,
-0.0250451937,
-0.0278887954,
-0.0122498246,
-0.0046012443,
-0.0548775904,
-0.0066395532,
0.0306721795,
2.671e-7,
0.0495006181,
0.0463532172,
0.0947815552,
-0.0310927257,
0.0199821945,
-0.0177825205,
0.0017153852,
0.038517423,
-0.0476517379,
-0.0660575628,
0.0089379987,
0.0009607836,
-0.0067271977,
0.0124704847,
-0.0769867897,
-0.021649031,
-0.0407550111,
0.0089449044,
-0.0398242772,
0.035957586,
0.0751704648,
0.0928857848,
0.0002373964,
0.0302197486,
-0.0100693768,
-0.0140156746,
0.0389484614,
0.0033028231,
0.0053227777,
-0.0106212953,
-0.016607644,
0.0136481514,
0.0029240777,
-0.0222238936,
-0.0124348458,
0.0253534149,
0.0464533716,
0.0364663601,
-0.0238954686,
-0.0324605927,
0.014304501,
0.0257373154,
-0.0059389696,
-0.0604339913,
0.04367036,
-0.0391913168,
-0.0023628261,
-0.0088959839,
-0.0330356359,
0.0293984041,
-0.0274172407,
0.0111986091,
0.0092931483,
0.0466629937,
0.0331111476,
0.008261363,
-0.0108801667,
-0.1329528391,
-0.0393376015,
0.0392124802,
-0.0218644906,
0.0161691755,
-0.0536158048,
0.0569971576,
0.0292137582,
-0.0580435656,
-0.0458439253,
2.773762592e-34,
-0.0193552002,
-0.0059916116,
0.0127307344,
0.0380458944,
0.012914734,
-0.0500224866,
0.014331081,
0.031883724,
0.0383110978,
-0.0471726097,
-0.0711055472
] |
http://arxiv.org/abs/2401.16420v1 | http://arxiv.org/pdf/2401.16420v1 | InternLM-XComposer2: Mastering Free-form Text-Image Composition and Comprehension in Vision-Language Large Model | 2024-01-29T18:59:02 | 2024-01-29T18:59:02 | Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, Wenwei Zhang, Yining Li, Hang Yan, Yang Gao, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, Jiaqi Wang | cs.CV | cs.CV, cs.CL | We introduce InternLM-XComposer2, a cutting-edge vision-language model
excelling in free-form text-image composition and comprehension. This model
goes beyond conventional vision-language understanding, adeptly crafting
interleaved text-image content from diverse inputs like outlines, detailed
textual specifications, and reference images, enabling highly customizable
content creation. InternLM-XComposer2 proposes a Partial LoRA (PLoRA) approach
that applies additional LoRA parameters exclusively to image tokens to preserve
the integrity of pre-trained language knowledge, striking a balance between
precise vision understanding and text composition with literary talent.
Experimental results demonstrate the superiority of InternLM-XComposer2 based
on InternLM2-7B in producing high-quality long-text multi-modal content and its
exceptional vision-language understanding performance across various
benchmarks, where it not only significantly outperforms existing multimodal
models but also matches or even surpasses GPT-4V and Gemini Pro in certain
assessments. This highlights its remarkable proficiency in the realm of
multimodal understanding. The InternLM-XComposer2 model series with 7B
parameters are publicly available at
https://github.com/InternLM/InternLM-XComposer. | InternLM-XComposer2: Mastering Free-form Text-Image Composition and Comprehension in Vision-Language Large Model: We introduce InternLM-XComposer2, a cutting-edge vision-language model
excelling in free-form text-image composition and comprehension. This model
goes beyond conventional vision-language understanding, adeptly crafting
interleaved text-image content from diverse inputs like outlines, detailed
textual specifications, and reference images, enabling highly customizable
content creation. InternLM-XComposer2 proposes a Partial LoRA (PLoRA) approach
that applies additional LoRA parameters exclusively to image tokens to preserve
the integrity of pre-trained language knowledge, striking a balance between
precise vision understanding and text composition with literary talent.
Experimental results demonstrate the superiority of InternLM-XComposer2 based
on InternLM2-7B in producing high-quality long-text multi-modal content and its
exceptional vision-language understanding performance across various
benchmarks, where it not only significantly outperforms existing multimodal
models but also matches or even surpasses GPT-4V and Gemini Pro in certain
assessments. This highlights its remarkable proficiency in the realm of
multimodal understanding. The InternLM-XComposer2 model series with 7B
parameters are publicly available at
https://github.com/InternLM/InternLM-XComposer. | [
0.0208359286,
0.049867034,
0.0101362364,
0.0434321687,
-0.0341743343,
0.0083013354,
0.0084850555,
-0.0342792384,
-0.0134386644,
-0.0590969473,
-0.026993366,
-0.0465397164,
0.0380889997,
0.0840139911,
0.0219644327,
-0.0776123032,
0.0071838023,
-0.0638113841,
-0.0364695899,
0.0264313482,
0.0486662462,
0.0201573223,
0.0080291564,
0.0231583696,
-0.0110258749,
-0.0448771492,
-0.0441024117,
0.0275154207,
0.0213557202,
-0.0104806973,
-0.0484897755,
0.0218905527,
0.0450200215,
0.0102819242,
0.0000019852,
-0.013460381,
0.0109248189,
0.0187056009,
0.0377563946,
-0.0293954052,
0.0605298281,
-0.075231187,
0.0207483638,
-0.0373726003,
-0.0120411599,
-0.0028910255,
0.0524060428,
0.0754106417,
0.0802136585,
0.0512953959,
-0.0225395169,
-0.0240808669,
0.0093601812,
-0.018226359,
0.0185614508,
-0.0175567567,
-0.0018384798,
-0.018372139,
0.0019689456,
0.0188000146,
-0.044655852,
0.0341585428,
0.0523093119,
-0.0595480949,
-0.0040178141,
0.039598614,
0.0111976229,
-0.049041748,
-0.0136533892,
0.0825052783,
0.0011558743,
-0.0172112472,
-0.0262385737,
-0.0048302584,
0.0012294924,
-0.0181834828,
-0.0512999184,
0.0533406064,
-0.0245138295,
0.0099897413,
-0.038373746,
-0.0175709054,
-0.007509843,
-0.0167268962,
0.0185961891,
-0.0048928219,
0.0232463442,
0.0086534359,
0.0030163745,
-0.0217915867,
-0.022417644,
-0.070422411,
-0.0018789512,
0.063496016,
0.000739227,
-0.0072003445,
-0.0868838131,
-0.0557776652,
0.0177334193,
-0.0058176601,
0.0598826595,
0.0168761481,
-0.0699580759,
0.0727729648,
-0.0307064392,
0.0692466795,
-0.0024346309,
0.0332459398,
-0.050350897,
-0.0231006723,
-0.0099132983,
-0.0217263363,
-0.0474868566,
-0.0047672405,
-0.0823383704,
-0.0575784929,
0.0086983424,
0.0279439948,
0.0128530115,
0.0373134799,
0.0023261388,
0.005685661,
-0.0413858108,
-0.0328915007,
-0.0364293568,
0.0387217924,
0.0119958632,
0.0175471231,
-0.000540024,
-0.0112709859,
-0.0027654902,
0.0537217148,
0.0122444155,
-0.0100457901,
-0.0274076164,
0.0755306333,
-0.0493340269,
-0.0222720876,
-0.0060322131,
-0.0308167692,
0.0030068082,
-0.0909696892,
0.0204413235,
-0.0421574563,
0.0371105373,
-0.0146167791,
-0.0039452715,
-0.006976244,
-0.012228922,
0.0358374752,
0.0054761772,
0.0939249694,
-0.0308801942,
-0.0171543099,
-0.0110849729,
0.0150155304,
-0.0060494044,
0.0812474415,
0.018301107,
0.0524600483,
0.0045259031,
0.0197397601,
0.0016348356,
0.0076099001,
-0.0211791955,
-0.0095905373,
-0.046114374,
0.0114216516,
0.0149883935,
-0.001212075,
0.0438518152,
0.0457701273,
-0.0308696777,
-0.0085288351,
0.091498524,
-0.0032343676,
0.0336662307,
0.0061184829,
0.0779516622,
0.0138799092,
0.0287671555,
0.0079134023,
-0.0525790565,
0.0826074928,
0.0242910553,
0.0277235713,
-0.0108800177,
0.0317919105,
-0.0297504924,
-0.0446385331,
-0.0154990396,
-0.0175800603,
-0.0030466199,
-0.0200714115,
0.0263864975,
0.0488720089,
-0.0573997572,
0.0217114668,
0.0030442493,
-0.025404444,
0.030367652,
-0.0140599087,
0.0581340119,
0.0638000295,
0.0083260853,
0.0014424234,
-0.0018228847,
-0.0168513246,
0.057720378,
0.0884199291,
-0.010857746,
0.0173200499,
0.0277861748,
-0.0731009841,
-0.0399936698,
-0.0603188574,
-0.0052356198,
0.0076301484,
-0.0359161831,
0.0231342707,
0.0067157536,
-0.0279451367,
0.0031516405,
0.0511683747,
0.0193383116,
-0.0021238336,
-0.0559229888,
0.0189697519,
-0.0134672364,
-0.0172646735,
0.0010300775,
0.0026376629,
0.0052063749,
-0.0377386771,
-0.0032749188,
0.0299933515,
-0.0334012099,
-0.0459322631,
0.0229902118,
-0.0833043531,
-0.0031375221,
-0.0040057031,
-0.0237628948,
-0.0433077775,
-0.0005579919,
-0.0505548306,
0.0099733332,
-0.0391940102,
-0.0266648717,
0.046328105,
0.0138073247,
-0.0062098289,
0.0345194191,
-0.018061256,
-0.02204003,
0.0261072293,
0.0156495534,
0.0242421199,
0.0197772328,
-0.0148583502,
-0.0002381927,
-0.0034900729,
-0.0038680879,
-0.0222097784,
0.0333954245,
-0.0375229903,
-0.0399536565,
-0.0757990554,
-0.0196182653,
-0.0269247722,
-0.0409007445,
0.0111659355,
-0.0227826014,
0.0197931267,
-0.0341654047,
-0.0357760862,
0.0322078578,
-0.0031986302,
-0.0033156364,
0.0300129559,
0.0148920016,
0.0203075018,
-0.0007868636,
0.0024123904,
0.0140381372,
0.0672951117,
0.0481551103,
-0.0134520419,
-0.0458646826,
-0.005367294,
-0.0330891274,
-0.0071967826,
-0.0457571484,
-0.0265144911,
0.0275390614,
-0.0277737007,
0.0546113253,
0.0520924889,
0.0151600316,
-0.0138928387,
0.0308019016,
0.0018201187,
-0.0069372142,
0.0207089111,
0.0143440124,
0.0054538012,
-0.0068956176,
-0.0006762312,
-0.0361012816,
0.0392045341,
-0.0003139845,
0.0180852097,
-0.0672226325,
0.0142008457,
-0.0014675938,
-0.057575345,
-0.0339544788,
0.000558212,
0.0028767118,
0.0220831949,
-0.0719880238,
-0.0554875694,
0.0404721759,
-0.0616845042,
-0.0425647721,
-0.0024548445,
0.0207509343,
-0.0525190979,
0.0814189911,
0.0300121438,
0.0387773998,
-0.0116862925,
0.0010447663,
-0.0495114513,
-0.0030476234,
-0.0024703159,
0.0147945853,
0.0344232991,
-0.0347625986,
0.025350282,
-0.0576234199,
-0.0107811484,
0.0011567685,
-0.0059054983,
0.0007504251,
0.0112281805,
-0.0236515701,
-0.0042121285,
-0.0158624034,
0.0349806547,
0.031021025,
0.0402813628,
-0.0145309949,
-0.0059139682,
-0.0299488269,
0.0330316275,
0.0787839293,
0.031349346,
-0.0094880331,
-0.0685381666,
0.0017006736,
0.0397606976,
-0.009165586,
-0.001345682,
-0.0374538824,
0.0726113394,
0.0001216545,
0.0356350206,
0.0206428226,
0.0058094542,
0.0733590648,
-0.0715112388,
-0.005864576,
0.0378953405,
-0.0125026461,
-0.049058456,
-0.0401090533,
-0.0079270154,
0.0447583646,
0.0526817776,
-0.0087447008,
-0.0940467119,
-0.0382423364,
-0.0073059332,
-0.0531372838,
0.0351963863,
0.0033445598,
-0.0631040558,
-0.057150472,
0.0477826893,
-0.0070903548,
-0.0609055236,
-0.0209766589,
0.0066177631,
-0.0066806288,
0.018775519,
-0.0003739911,
0.0101965992,
-0.0220740587,
-0.0233989041,
-0.006724006,
0.0323512852,
-0.0007252816,
-0.0204592384,
-0.018941015,
-0.0222982075,
0.0088656647,
0.0314001106,
0.0022998201,
-0.062520735,
0.0425786376,
0.014323948,
-0.0019275496,
0.0031972132,
0.0561754629,
-0.0110366829,
-0.033086434,
-0.0185220912,
0.0346732996,
-0.0193366334,
-0.0323902816,
-0.0326285549,
0.0131318439,
-0.0009327335,
-0.0144613655,
-0.0533729121,
0.0189062767,
-0.0080383131,
0.044624757,
0.0395907052,
-0.0398095921,
-0.0397411436,
-0.0155575285,
-0.0041257963,
-0.0762245953,
0.0101091433,
0.0306429639,
0.0152582014,
0.03056092,
-0.0387868546,
-0.0368112251,
0.0420586653,
0.0754342601,
-0.0057695545,
0.0051438492,
-0.048047062,
-0.0160509422,
-0.0434401259,
0.0660028979,
-0.0015335059,
0.0189680159,
-0.0414187834,
-0.0144901443,
-0.0117928246,
-0.0625968575,
-0.019519249,
0.0334955081,
-0.0883098021,
-0.0179639086,
-0.0194320567,
0.0694826245,
0.0212426763,
0.0346526951,
0.022698123,
-0.046742741,
0.05219578,
0.0176688079,
0.0322038531,
0.0383887999,
-0.0063277651,
-0.0251799095,
0.007791766,
0.0323084481,
-0.0254548118,
0.0116096688,
0.0267366003,
0.0215768144,
0.0343820229,
-0.0176485758,
0.0183497109,
-0.0011831318,
-0.0262756795,
-0.0160541199,
0.0338076353,
-0.0476450734,
0.0130528882,
-0.0392495915,
-0.0146920858,
-0.1035658419,
0.0047769067,
0.0115829585,
0.0125458343,
0.005063667,
0.0133037427,
0.0125004677,
0.0273500066,
-0.0108281402,
0.0123235565,
0.0321363322,
-0.0149892466,
0.0203985721,
0.0292260405,
-0.0148044676,
0.022588376,
-0.0743016899,
0.0199082382,
0.0034711235,
0.0631925464,
0.0108169587,
-0.0331059508,
-0.0218155179,
0.0412029475,
-0.0145643763,
-0.0241168234,
-0.0369681939,
0.0275851656,
0.0245045032,
0.0262687393,
0.0610929839,
-0.0095509049,
-0.021158535,
0.0026052194,
0.0622179843,
0.0134135252,
-0.0441300422,
0.0551510081,
0.0031105254,
0.012818207,
-0.0120211961,
0.0675796941,
0.0225803088,
0.0350450575,
0.003240417,
0.0451676548,
-0.0447325073,
-0.0142742693,
0.0354822949,
0.014379032,
0.0069772005,
0.0075313454,
-0.0140116895,
-0.0471988209,
0.0372766256,
0.0478683934,
-0.0171006862,
0.040898528,
0.0493788458,
0.050732173,
0.0085903592,
-0.0286848117,
-0.0062637427,
-0.0318761542,
0.0270498842,
0.0088400887,
0.0196609534,
-5.864686784e-33,
-0.0517639704,
-0.0269946046,
0.0492215715,
-0.0088513149,
-0.0510869399,
0.0281124655,
-0.0354386456,
-0.0052534323,
0.0201773811,
0.0121907247,
0.0558123775,
-0.0191073269,
0.0022511072,
0.0621168017,
-0.0243053772,
0.0136129605,
0.0327384993,
0.0144998655,
-0.0112582222,
0.0063636717,
0.0220866092,
0.0142986653,
0.0351869129,
-0.004564744,
-0.0042659859,
-0.0046665994,
-0.0340746492,
0.0088046975,
-0.0208102446,
0.0961062312,
-0.0348317064,
0.0130871478,
0.0225617364,
0.0107525587,
-0.0410147607,
0.0212576874,
-0.0575259775,
-0.0165059194,
0.0421293899,
0.0293116067,
-0.1106475741,
-0.012957505,
0.0731029809,
-0.0445264503,
0.0422750525,
-0.0612669326,
0.0662953854,
0.0062213405,
0.0063072429,
0.0096279169,
-0.0747187138,
-0.0184467062,
-0.0004213757,
-0.0166536029,
0.0432592705,
0.0160250477,
0.0186518319,
0.0391094871,
-0.0527669862,
-0.0146505805,
-0.0184113421,
0.0210602004,
0.0468193255,
0.0332706943,
-0.0072230059,
0.0416349992,
0.0233791359,
-0.0096887155,
-0.0061088968,
0.0072524464,
0.0241647307,
0.0570941977,
0.0268549845,
0.0934653431,
0.0050932672,
0.0445464589,
-0.0413291901,
0.0002242537,
-0.0568942875,
0.0562087744,
-0.0017182615,
0.0057831639,
0.0231643897,
0.0314039104,
-0.0555333234,
-0.0412369594,
-0.0572998449,
-0.0006898714,
0.0577154607,
0.0465916619,
0.0140597019,
0.0075624241,
0.0613979064,
-0.0083148992,
0.0033769475,
-0.0214022677,
-0.0432914346,
0.0743701085,
0.0015266383,
0.0292681437,
-0.0171975084,
-0.029773213,
0.0676886216,
0.012007962,
0.0021148084,
0.0093973009,
0.0654560849,
0.0452368632,
-0.038464915,
0.0048687393,
-0.0519646406,
0.0322548784,
-0.0038340141,
0.0427447483,
0.0390119329,
-0.0106789852,
-0.0870895609,
0.0143802948,
-0.0231889989,
-0.0006210168,
-0.0186887104,
-0.0464376211,
-0.0018562813,
0.0332439095,
-0.0255823489,
0.0056087435,
-0.0336083882,
-0.0378400236,
0.0346910283,
-0.0556410514,
0.0108083719,
0.0738726929,
2.481e-7,
0.0086887283,
0.0058430429,
0.0210702829,
-0.0139538068,
0.020176122,
0.0202791207,
-0.0550898165,
0.0493819527,
-0.0257118177,
-0.0211544055,
-0.0174445771,
0.0428290777,
0.0375453085,
-0.0087106908,
-0.0419875234,
0.0361211263,
-0.0450456329,
-0.0254995208,
-0.0629100204,
0.0024051205,
0.0638148859,
0.0640946776,
-0.034493167,
-0.0406898297,
0.0313385911,
-0.0550292172,
0.1173875928,
-0.0089772725,
-0.0182118472,
-0.0146696987,
0.0197803732,
0.0032323918,
-0.0218669791,
0.0064685522,
0.0035930562,
-0.0415932946,
0.0546201728,
0.0443802401,
-0.0260490626,
-0.0105433576,
0.0240947958,
0.0225639008,
-0.0303435866,
-0.0651614293,
0.0790099278,
-0.0136140063,
-0.0213473625,
-0.0434259176,
-0.0479426607,
-0.0049501481,
0.0065120803,
-0.0256507508,
-0.0292529035,
0.0366376638,
0.0387518294,
-0.0204061046,
-0.0216090884,
-0.065327771,
-0.0439203307,
-0.0273729563,
-0.0322774611,
0.0184427127,
-0.0998984352,
-0.01019846,
0.0442355238,
-0.0571564883,
-0.04843463,
2.190225385e-34,
0.0228629969,
-0.0406632349,
0.0138337268,
-0.0121340435,
0.0125961956,
0.0017945323,
-0.003358698,
0.0046990816,
-0.0257015042,
-0.0773732588,
-0.0211270265
] |
http://arxiv.org/abs/2401.16467v2 | http://arxiv.org/pdf/2401.16467v2 | ReGAL: Refactoring Programs to Discover Generalizable Abstractions | 2024-01-29T18:45:30 | 2024-06-06T17:31:07 | Elias Stengel-Eskin, Archiki Prasad, Mohit Bansal | cs.SE | cs.SE, cs.AI, cs.CL, cs.LG, cs.PL | While large language models (LLMs) are increasingly being used for program
synthesis, they lack the global view needed to develop useful abstractions;
they generally predict programs one at a time, often repeating the same
functionality. Generating redundant code from scratch is both inefficient and
error-prone. To address this, we propose Refactoring for Generalizable
Abstraction Learning (ReGAL), a gradient-free method for learning a library of
reusable functions via code refactorization, i.e., restructuring code without
changing its execution output. ReGAL learns from a small set of existing
programs, iteratively verifying and refining its abstractions via execution. We
find that the shared function libraries discovered by ReGAL make programs
easier to predict across diverse domains. On five datasets -- LOGO graphics
generation, Date reasoning, TextCraft (a Minecraft-based text-game) MATH, and
TabMWP -- both open-source and proprietary LLMs improve in accuracy when
predicting programs with ReGAL functions. For CodeLlama-13B, ReGAL results in
absolute accuracy increases of 11.5% on LOGO, 26.1% on date understanding, and
8.1% on TextCraft, outperforming GPT-3.5 in two of three domains. Our analysis
reveals ReGAL's abstractions encapsulate frequently-used subroutines as well as
environment dynamics. | ReGAL: Refactoring Programs to Discover Generalizable Abstractions: While large language models (LLMs) are increasingly being used for program
synthesis, they lack the global view needed to develop useful abstractions;
they generally predict programs one at a time, often repeating the same
functionality. Generating redundant code from scratch is both inefficient and
error-prone. To address this, we propose Refactoring for Generalizable
Abstraction Learning (ReGAL), a gradient-free method for learning a library of
reusable functions via code refactorization, i.e., restructuring code without
changing its execution output. ReGAL learns from a small set of existing
programs, iteratively verifying and refining its abstractions via execution. We
find that the shared function libraries discovered by ReGAL make programs
easier to predict across diverse domains. On five datasets -- LOGO graphics
generation, Date reasoning, TextCraft (a Minecraft-based text-game) MATH, and
TabMWP -- both open-source and proprietary LLMs improve in accuracy when
predicting programs with ReGAL functions. For CodeLlama-13B, ReGAL results in
absolute accuracy increases of 11.5% on LOGO, 26.1% on date understanding, and
8.1% on TextCraft, outperforming GPT-3.5 in two of three domains. Our analysis
reveals ReGAL's abstractions encapsulate frequently-used subroutines as well as
environment dynamics. | [
-0.0044053611,
0.0517294444,
-0.0309803803,
0.0383396111,
-0.0095106671,
-0.0008448958,
-0.0608561076,
-0.0003036765,
0.0114931241,
0.0090107527,
-0.0202303473,
0.0327511653,
-0.004319387,
0.0243495721,
0.016504366,
-0.044265952,
0.0329264812,
-0.0320610069,
-0.0089742988,
0.0117201442,
-0.0361778513,
0.0091996267,
0.028599713,
0.0827225372,
0.0629811287,
-0.0362059809,
-0.0138618648,
0.029288251,
-0.0182069633,
0.0189284086,
0.0012713439,
0.015798388,
0.0118540805,
0.0361331068,
0.0000024352,
-0.0173439998,
-0.0133542344,
-0.0006328797,
0.0447343327,
0.0204479378,
0.0058872718,
-0.0015034968,
0.0197498277,
0.0272953119,
-0.0369488485,
-0.0358324088,
0.0359701663,
-0.0431823619,
0.0766833946,
0.04904994,
0.0113027329,
-0.0459317975,
-0.0078034126,
0.0076012188,
0.0363634937,
0.0052343127,
-0.0221817121,
0.0413699076,
-0.0152964927,
-0.0601990335,
-0.0184291657,
0.0393508337,
0.0171591789,
0.03570446,
0.0240858309,
0.0381999724,
-0.0524987057,
0.0000714658,
0.015698988,
0.0386124849,
-0.0982471257,
-0.0625980496,
0.0378068201,
0.0094666909,
-0.0145012876,
-0.0210641231,
-0.0464878082,
0.0058767726,
0.0087241055,
0.0051931501,
0.0152110048,
-0.0217750072,
0.0342847444,
-0.0091061275,
-0.0376219451,
0.0672436506,
-0.0001745534,
0.029858252,
-0.0478629023,
-0.018637307,
-0.0047233058,
-0.0442845263,
-0.0199686438,
0.0253033899,
-0.0161354803,
0.0158392675,
-0.0526753776,
-0.0135815283,
-0.0107021118,
0.0017491435,
0.0305223241,
0.0588172711,
0.0398273803,
0.0040210676,
-0.0778897628,
0.0655779243,
-0.0327411518,
-0.0157127343,
-0.036232423,
0.0177187752,
-0.0537169389,
-0.0115682948,
-0.0342868939,
0.0822830126,
-0.0115233883,
0.0405237079,
-0.0396760926,
0.0523636527,
-0.0199723374,
0.0532713458,
-0.0081879562,
-0.0096825957,
-0.0031985214,
-0.0340104327,
0.0018370707,
-0.0079043172,
0.0001622095,
-0.0441974886,
-0.0186478142,
-0.0387390852,
0.0057526957,
-0.0152582526,
-0.0627943501,
0.0104686385,
0.0285922494,
0.1024685204,
-0.0363374539,
0.0384475626,
-0.0520731248,
-0.0337713324,
0.0432799794,
-0.03197705,
0.0274847075,
-0.1192278117,
-0.0104897097,
0.0137921078,
0.0236938447,
-0.015672449,
0.0079390816,
-0.0267461818,
-0.0078776926,
0.0502696559,
-0.0166682936,
0.0522090755,
0.0364228934,
0.0228311699,
-0.0138879539,
0.0449505337,
0.0319108739,
-0.0032089802,
-0.0007582809,
-0.0329153799,
0.039348416,
0.0326033719,
-0.0120120738,
-0.0300691668,
0.0286358185,
0.0050412551,
-0.0202933773,
-0.0040252698,
0.015410156,
0.007404929,
-0.0212455429,
-0.0117976293,
0.0087988935,
0.0222209953,
-0.0135538904,
0.063002035,
0.0864599869,
0.0276335962,
-0.0001214953,
-0.0014799855,
-0.0553309917,
0.0401464254,
0.0098534049,
0.0440310016,
-0.0260504279,
0.0241227504,
-0.0158268921,
-0.0825290531,
0.0144038731,
-0.01522341,
0.0038288829,
-0.0194093585,
0.0200867411,
0.0050430885,
-0.0724499077,
0.0134766083,
-0.0037849715,
-0.0414302535,
-0.0515532307,
-0.0695607588,
0.0593455322,
0.0728094876,
0.0503905341,
-0.0730053559,
-0.005081363,
-0.0150374295,
0.0059717526,
-0.044151891,
0.0311186686,
-0.008306724,
0.025584247,
-0.0460572131,
0.0112831779,
-0.036675062,
0.0147424098,
0.0981830284,
0.0457120165,
-0.0240819417,
0.0198168587,
-0.037927866,
-0.0348298438,
0.0278909225,
-0.006921547,
-0.0327952094,
0.0374583155,
0.0035147923,
0.0182210002,
-0.0711655766,
-0.0327571258,
0.0068543656,
0.0353653952,
-0.0546009354,
-0.0402882993,
-0.0225167777,
-0.0716817006,
-0.0146927489,
0.023236515,
-0.0519988015,
-0.002879533,
0.0405366197,
-0.0638840348,
-0.0192402452,
0.1232557148,
0.0167295616,
0.0941109434,
0.0234578848,
0.0327557251,
-0.0264089871,
-0.0370247252,
-0.0082417987,
0.0395150855,
-0.0008297865,
-0.0143260211,
0.007368525,
0.0209006425,
0.0316957645,
0.0592281558,
-0.0150565971,
-0.0069843777,
-0.0407267064,
0.0403445587,
-0.0186793711,
0.0191996619,
-0.0266088173,
-0.0228256993,
-0.0656227395,
0.0125081642,
-0.0190126672,
0.0127795935,
0.0242788177,
0.0199959837,
0.0177729037,
-0.0147907566,
0.0143695138,
-0.0265527871,
-0.0385446809,
-0.0012615521,
0.0239262395,
-0.0220804177,
-0.0816652402,
-0.0143023627,
-0.0341383815,
-0.0337008908,
0.0693656802,
-0.0670877025,
-0.0086996825,
-0.0461498462,
0.044120349,
0.010093376,
0.0433525629,
-0.0476484336,
-0.0119286785,
-0.0004733615,
-0.0389876738,
0.0813058764,
0.0603510812,
-0.0488859899,
-0.0239008442,
0.0366973132,
0.0029810704,
-0.050390251,
0.0352478288,
0.0251767226,
-0.044826813,
0.0236353278,
-0.0283609219,
0.0207789093,
0.0728218704,
-0.0081078224,
-0.0264395997,
-0.055801034,
0.000943462,
0.0179808903,
-0.0073182755,
-0.021555325,
-0.0322451293,
0.0228738263,
-0.0158592314,
-0.0247809906,
-0.0997958779,
0.0498393327,
-0.0117892176,
-0.0235514827,
-0.0162116271,
-0.0215814915,
-0.0286234114,
0.0644545704,
-0.0026275418,
0.0022508467,
0.0215167217,
-0.0312165935,
-0.0477948301,
-0.0051242616,
0.0054213307,
-0.0388385653,
0.0878363252,
-0.0711476952,
-0.0074310862,
0.0174122564,
-0.0411361419,
-0.033308357,
0.0326509476,
-0.0154646914,
0.0018543024,
-0.0081027113,
-0.0232049301,
-0.0191441756,
0.0026250291,
0.0125817675,
0.0229413416,
0.0108056385,
-0.0486344211,
-0.003747368,
-0.0365897045,
-0.0185973104,
-0.0535215326,
-0.0273630805,
-0.0228000209,
0.014604914,
-0.0005234923,
0.0322451815,
-0.0135656698,
-0.0681266487,
-0.0238264445,
0.0142668476,
-0.0119020166,
-0.0536024272,
0.0423928611,
0.0509930663,
-0.071130313,
-0.0159240831,
0.0615707859,
0.0179831292,
-0.0088872397,
0.0075163157,
-0.0189212766,
0.0939612985,
0.0770503953,
0.0668485314,
-0.0384070128,
-0.0434049554,
-0.0060910159,
0.0175590236,
-0.0002888355,
-0.0068341475,
-0.008400593,
-0.1006469578,
0.0136973532,
0.02293179,
-0.0475616679,
-0.0143753458,
-0.020104846,
-0.0303217229,
0.0173882693,
-0.0727695227,
0.0270275306,
-0.0102850711,
0.0264086556,
-0.0115963751,
-0.1238553897,
0.0258555915,
0.0216661636,
0.0013427595,
-0.0216707774,
0.0391671807,
0.0161944758,
0.0093353027,
-0.0133675402,
0.0622007847,
-0.0025410536,
-0.0016851653,
-0.0313101709,
0.0318099819,
0.04026559,
-0.0219673198,
0.0465791598,
0.0304712709,
-0.0261738244,
-0.0296378061,
0.0220654141,
0.0719042942,
0.0035708928,
-0.0396376513,
-0.0129899271,
-0.0294135883,
-0.0053541674,
0.0199330393,
0.0165286642,
-0.0003721881,
-0.0556188338,
-0.067485176,
-0.014462769,
-0.0436116606,
-0.0157835446,
-0.051606331,
0.0031985908,
0.0021272439,
-0.039006792,
0.1012715474,
0.0265511293,
0.0297340341,
0.0261131227,
-0.017621737,
0.0256196223,
-0.0218630638,
-0.0888002068,
-0.0018942995,
-0.026737608,
0.0203261822,
-0.0008425069,
-0.036412403,
-0.0056333099,
-0.0149207572,
-0.0205011908,
0.0834310278,
-0.0527622849,
-0.0450931974,
-0.0364368968,
0.1056550965,
0.0212384369,
0.0358827077,
-0.0197714716,
0.0059721614,
0.0176771656,
0.0238350034,
0.0236011725,
0.0081247324,
0.0232125316,
-0.0065879528,
0.0399727784,
-0.0079162372,
-0.0413439609,
-0.0097355861,
0.0566921495,
-0.0274806768,
0.0296549797,
0.1030397713,
0.0141105475,
0.0342031233,
-0.0048069977,
0.0054620188,
0.0041517387,
-0.0025337236,
-0.0055509736,
0.0220950209,
-0.0205919351,
-0.0394239463,
0.0546475537,
0.0219912399,
0.0015606323,
0.0166899152,
-0.0055643539,
-0.0052542011,
0.0378477611,
-0.0068693445,
0.0053571523,
-0.018726686,
0.052520074,
0.0778499693,
-0.0084037017,
-0.012197203,
0.0037744732,
0.0101656485,
0.0335300192,
-0.0031461213,
0.005599516,
0.0180281065,
0.0626702532,
-0.049580127,
0.036746677,
-0.0339237265,
-0.1024188399,
-0.0150494548,
0.0377642252,
0.0342234112,
0.0244485214,
0.0337955169,
-0.0260000098,
0.0121286428,
-0.0162761025,
0.0471539348,
0.0038593735,
-0.0267429836,
-0.000253549,
0.0290329214,
-0.0201903824,
-0.0611498691,
0.0662753135,
-0.0333243012,
0.0273691323,
0.0010984442,
-0.0011728627,
-0.0326429345,
-0.0007786327,
0.0283264257,
0.0041864421,
-0.019522775,
-0.0524947383,
-0.0380292125,
-0.0421050675,
-0.0017912593,
0.0472651459,
0.019421678,
0.0020435145,
0.0331741385,
0.0868345201,
0.019479908,
-0.0488074683,
-0.0107484721,
0.0094529614,
-0.0137226339,
-0.0295597501,
-0.0324781612,
-6.127701441e-33,
-0.0057937438,
0.0314630829,
0.005993824,
0.0368919186,
-0.0098068016,
0.0193435531,
-0.0635403693,
0.0052419426,
0.0067299474,
-0.01421923,
0.0056218631,
-0.0062032822,
-0.0097922422,
-0.0407316759,
0.036697574,
0.0128836604,
0.0183911249,
0.0282608215,
-0.0873182043,
0.0218792036,
-0.0154351341,
0.0154993683,
0.0394921713,
-0.0902423784,
-0.078021206,
0.0150747467,
0.0125346705,
0.0053346795,
0.0012873366,
0.0011326213,
0.0072972248,
0.0026554163,
-0.0105452016,
-0.0239363257,
-0.0405337662,
-0.0129509438,
-0.033588998,
-0.0304497089,
0.0460996293,
0.0106186466,
-0.0375213549,
-0.0301588234,
0.0847617313,
0.0404498167,
-0.0152852163,
-0.0208934322,
0.0168635435,
-0.0049664364,
0.0140092783,
-0.0271829348,
-0.0641648322,
0.0003011501,
-0.0209098291,
-0.0149870617,
-0.0218138192,
0.02196417,
-0.0200468823,
0.0444535017,
0.0187465921,
0.012709287,
0.0120620895,
0.0428638719,
0.0820915475,
0.0196243133,
0.0502364561,
0.0355081409,
0.0397787169,
0.051373262,
-0.0167063139,
0.0149440244,
0.040496733,
0.0498544201,
0.0044309078,
-0.0165685546,
0.0222170949,
0.0013750745,
-0.0435398184,
0.0076808869,
-0.0277889129,
0.0302622896,
0.0283613577,
0.0292197298,
0.0128528113,
-0.0077814581,
-0.0299082864,
0.0197755061,
-0.0152388345,
-0.0086934073,
0.0108322483,
0.015512933,
-0.0113361608,
-0.0177591648,
-0.0109321885,
-0.0439481623,
-0.0143171065,
0.0074884812,
-0.0550975464,
0.0104763331,
-0.0066769053,
0.0070671453,
0.0176557135,
-0.032014668,
-0.0112236049,
-0.0231274609,
0.0747247562,
-0.0015181717,
0.0177987292,
0.0412598848,
-0.0726830512,
0.0040875771,
-0.0633518919,
-0.0080406489,
-0.0113586504,
0.0265491456,
-0.0229113176,
-0.0072222683,
0.0009166661,
0.0140670557,
-0.0017137076,
-0.0281098913,
-0.0258858092,
0.0321762785,
0.0263420995,
0.0430823788,
-0.0413216949,
-0.0484333411,
-0.0492440835,
0.0299717709,
0.0448498242,
-0.0209463853,
-0.0693645999,
0.0285082739,
3.164e-7,
0.0090326462,
0.0135356607,
-0.0224281661,
0.0405848697,
-0.0028610199,
0.0445043594,
-0.0299433097,
-0.0057626972,
0.0185026471,
0.0051983111,
-0.0014067251,
0.0233706869,
0.043180488,
-0.0112922546,
-0.1048021838,
0.0376944505,
0.004623455,
-0.0063746353,
-0.0429330617,
-0.0146485902,
-0.0120844617,
0.0611917377,
0.0637386739,
0.0086569665,
0.0365915298,
-0.0257071294,
0.0043563163,
-0.0440928712,
0.0178064015,
0.0393958502,
-0.0522365533,
0.0281033292,
-0.0096755559,
-0.0225892086,
-0.0090655694,
-0.0130988974,
-0.0033130627,
0.0675092489,
0.0038071908,
0.0678275973,
0.0504109003,
-0.0320489556,
0.0088746529,
-0.0647426471,
0.0419186838,
-0.0234719273,
0.0083174575,
-0.0065510604,
-0.0239686463,
0.0201878976,
0.000894928,
-0.003094296,
0.0172663387,
0.0105018253,
-0.0142330257,
-0.0008627706,
-0.0224096067,
-0.0172852017,
-0.065878354,
0.0269606896,
-0.0252810884,
-0.0247003213,
-0.0439441167,
0.0215213615,
0.000699651,
-0.0631550327,
0.0152004166,
3.244787999e-34,
0.0451257788,
-0.0260687731,
0.0043197623,
0.0486593135,
0.0496495143,
-0.0265222527,
-0.0138577381,
0.0120000727,
-0.0000051638,
0.0258474816,
-0.0175124779
] |
http://arxiv.org/abs/2401.16405v2 | http://arxiv.org/pdf/2401.16405v2 | Scaling Sparse Fine-Tuning to Large Language Models | 2024-01-29T18:43:49 | 2024-02-02T14:53:14 | Alan Ansell, Ivan Vulić, Hannah Sterz, Anna Korhonen, Edoardo M. Ponti | cs.CL | cs.CL, cs.AI, cs.LG | Large Language Models (LLMs) are difficult to fully fine-tune (e.g., with
instructions or human feedback) due to their sheer number of parameters. A
family of parameter-efficient sparse fine-tuning methods have proven promising
in terms of performance but their memory requirements increase proportionally
to the size of the LLMs. In this work, we scale sparse fine-tuning to
state-of-the-art LLMs like LLaMA 2 7B and 13B. We propose SpIEL, a novel sparse
fine-tuning method which, for a desired density level, maintains an array of
parameter indices and the deltas of these parameters relative to their
pretrained values. It iterates over: (a) updating the active deltas, (b)
pruning indices (based on the change of magnitude of their deltas) and (c)
regrowth of indices. For regrowth, we explore two criteria based on either the
accumulated gradients of a few candidate parameters or their approximate
momenta estimated using the efficient SM3 optimizer. We experiment with
instruction-tuning of LLMs on standard dataset mixtures, finding that SpIEL is
often superior to popular parameter-efficient fine-tuning methods like LoRA
(low-rank adaptation) in terms of performance and comparable in terms of run
time. We additionally show that SpIEL is compatible with both quantization and
efficient optimizers, to facilitate scaling to ever-larger model sizes. We
release the code for SpIEL at https://github.com/AlanAnsell/peft and for the
instruction-tuning experiments at https://github.com/ducdauge/sft-llm. | Scaling Sparse Fine-Tuning to Large Language Models: Large Language Models (LLMs) are difficult to fully fine-tune (e.g., with
instructions or human feedback) due to their sheer number of parameters. A
family of parameter-efficient sparse fine-tuning methods have proven promising
in terms of performance but their memory requirements increase proportionally
to the size of the LLMs. In this work, we scale sparse fine-tuning to
state-of-the-art LLMs like LLaMA 2 7B and 13B. We propose SpIEL, a novel sparse
fine-tuning method which, for a desired density level, maintains an array of
parameter indices and the deltas of these parameters relative to their
pretrained values. It iterates over: (a) updating the active deltas, (b)
pruning indices (based on the change of magnitude of their deltas) and (c)
regrowth of indices. For regrowth, we explore two criteria based on either the
accumulated gradients of a few candidate parameters or their approximate
momenta estimated using the efficient SM3 optimizer. We experiment with
instruction-tuning of LLMs on standard dataset mixtures, finding that SpIEL is
often superior to popular parameter-efficient fine-tuning methods like LoRA
(low-rank adaptation) in terms of performance and comparable in terms of run
time. We additionally show that SpIEL is compatible with both quantization and
efficient optimizers, to facilitate scaling to ever-larger model sizes. We
release the code for SpIEL at https://github.com/AlanAnsell/peft and for the
instruction-tuning experiments at https://github.com/ducdauge/sft-llm. | [
0.0113734649,
0.0586781837,
-0.0077219545,
0.0499644354,
0.023423342,
0.0211237259,
-0.0094137164,
0.0075420197,
-0.058833003,
-0.0409677252,
-0.0703488961,
-0.0308130533,
0.0137613071,
0.0040003578,
0.0505621843,
-0.028270172,
0.0374831706,
-0.0641817525,
-0.0238282047,
-0.0161003135,
0.003616472,
0.0101273051,
0.0082475664,
0.0271017812,
0.0480001606,
0.030407114,
0.0013155222,
0.0323942825,
0.0336331949,
0.0259956606,
0.0237976555,
0.0917941108,
0.0454710089,
0.0287786424,
0.0000021187,
-0.0161823444,
0.0181053858,
-0.0164056681,
-0.0174706113,
0.0736662373,
0.0478142276,
0.001924733,
-0.0299491864,
0.013970783,
-0.0056371945,
-0.0172106866,
0.0823583081,
-0.0163561739,
0.0609755889,
0.092524305,
0.0008374978,
-0.029683359,
0.0100713931,
-0.0603338145,
-0.016179774,
-0.0371986926,
0.0340106189,
-0.0278383587,
-0.0205469094,
0.0074272337,
-0.0335306078,
0.0021669632,
-0.0135957217,
0.0173352454,
-0.0028029727,
0.0076975543,
-0.0471843481,
-0.0196424425,
-0.0089403307,
0.0600216463,
-0.0616442114,
-0.0355056375,
0.0203059763,
0.0024430752,
-0.0202034265,
-0.0460964814,
-0.0231552273,
-0.0159117468,
0.0316319056,
0.0054986537,
0.0146362139,
-0.0684776232,
0.0433004498,
0.0313815288,
-0.0321051516,
0.0005425386,
0.0347458757,
-0.0007039313,
-0.0309359059,
-0.0299114678,
-0.0438049771,
-0.0222002435,
-0.0125415353,
0.0129481815,
-0.057385873,
-0.0170642547,
-0.0925689861,
-0.0462910309,
0.0127142519,
0.0228256602,
0.1060807854,
0.0553187504,
-0.0006748202,
0.0304616597,
-0.026758872,
0.0291971564,
-0.0033632279,
0.0422344655,
-0.0488087833,
0.0212366153,
-0.0131002404,
0.0430656224,
-0.0255167317,
0.0044258344,
-0.0320492461,
0.0148448618,
-0.0381940715,
0.0170723721,
0.0059924577,
-0.0100463973,
0.0043661362,
-0.0268101823,
-0.0525114462,
-0.0076992433,
0.0229452997,
0.0470023826,
-0.0301185008,
0.0317878276,
0.0061363666,
-0.0514760762,
-0.0276877936,
0.0199821983,
0.0315195508,
-0.0161315892,
0.0508215986,
0.0222700872,
-0.0185026564,
-0.0485835299,
-0.0439110249,
0.0245564263,
0.0061003119,
-0.0836997405,
-0.0260097831,
-0.0671673417,
0.0131184133,
-0.0185573734,
0.0367154069,
-0.0612973198,
-0.0048658303,
-0.0185292941,
-0.0543730296,
0.0358945541,
-0.0320715271,
0.0163349733,
0.015747441,
-0.0009692166,
0.0094801895,
0.0783463344,
0.0420591794,
0.0402690656,
0.0027153746,
-0.0186491888,
-0.0350186303,
0.0432318598,
0.0221453141,
0.0591673553,
-0.0118367225,
-0.0126712453,
0.009291348,
0.0182363614,
0.0482797697,
0.0423167534,
-0.0416364484,
-0.0318235531,
0.0451281667,
0.0344519205,
0.0322769023,
0.0455506369,
0.0317301117,
-0.0176080633,
0.0044383514,
-0.0083568273,
-0.0442850329,
0.0303283762,
-0.0434984751,
-0.0127606001,
0.0502665676,
0.0805612355,
-0.0563005283,
-0.0250833146,
-0.0056924205,
0.0168866795,
-0.0255778227,
0.0253704172,
0.035788741,
-0.0236424059,
-0.0417371318,
0.0284162033,
0.0586164147,
-0.0281474069,
-0.0168768335,
-0.0657147691,
0.0133906445,
0.0619761087,
0.0211210418,
-0.0142516475,
-0.0427632257,
-0.0108005004,
0.0339386426,
-0.0102032321,
0.0393158644,
0.0126414178,
0.0348654389,
-0.0672062337,
-0.0079538068,
0.0048402776,
0.0229660794,
0.0449373238,
0.0148867825,
0.0122169387,
0.024482891,
-0.0150448922,
-0.015774345,
0.0299315974,
-0.0091311643,
-0.020824464,
0.0188676622,
-0.0133094955,
0.0324426033,
-0.0176044721,
-0.0269559734,
-0.019565301,
0.0284844786,
-0.0284503251,
0.0628321841,
-0.0639323145,
-0.0079792831,
-0.0030869986,
0.036591474,
0.0236679371,
0.0298731104,
0.0391554832,
-0.0537028462,
-0.0478882268,
0.0434172601,
-0.0356890187,
-0.0039053271,
-0.0424826927,
0.0180238541,
-0.0445569493,
0.0689141303,
-0.031672813,
0.0168906804,
-0.0071675624,
-0.0035140766,
0.0231097024,
0.0136534004,
-0.0003275823,
-0.052584067,
-0.0464194492,
0.0062542963,
0.0021456203,
0.0509966649,
-0.0081738364,
0.0278827157,
-0.0085400417,
-0.0386012085,
-0.0940988958,
0.0403160825,
-0.0641445369,
0.0309599992,
0.031032959,
0.0305989534,
-0.0231455769,
-0.001872245,
0.0042116595,
0.0366997942,
-0.017758213,
0.0268421154,
-0.0129951984,
0.0170490667,
-0.050881736,
-0.0644369945,
-0.0348781198,
0.0296626501,
0.0085451184,
-0.0460953303,
-0.0088769877,
-0.0692345947,
0.0326447934,
-0.024216136,
0.0303174835,
-0.016007809,
-0.0558538921,
-0.0137587599,
-0.0357487574,
0.0501146205,
0.0762846842,
-0.0041036992,
0.0288782567,
0.0657145903,
0.0083568143,
-0.0115290591,
0.0121766673,
-0.0110160718,
-0.056528192,
-0.0462226756,
-0.0158336759,
-0.0135195674,
0.068750374,
-0.0243858006,
-0.0238187332,
-0.065330863,
0.0229719318,
0.0027818701,
-0.0107737659,
-0.0293918923,
0.0464979,
-0.011397602,
0.0247127898,
-0.0113410568,
-0.0575647615,
0.0621898584,
-0.0444622785,
-0.0132880481,
-0.0084105162,
-0.0360933356,
-0.0368050151,
-0.0085442578,
-0.0136747463,
0.0152193876,
-0.030333491,
0.0133683179,
-0.0523851439,
-0.0203097947,
0.0170323104,
-0.0171593595,
0.0390970111,
-0.0390686244,
0.0250534061,
-0.0413834602,
-0.0326693244,
0.0232761968,
-0.0435193889,
-0.0191268064,
0.0272549335,
0.0238730349,
-0.0511906557,
-0.0240215063,
0.0084388414,
0.0021327117,
0.062607199,
-0.034744285,
0.0059810677,
-0.0358963236,
-0.0046680355,
0.0759076849,
0.0614473335,
0.0141276224,
-0.0398082323,
-0.0052439468,
-0.0067416518,
0.0459571406,
0.0039257989,
-0.056043379,
0.0290694237,
0.0292252079,
0.0116813267,
-0.0019278793,
0.0276707727,
0.0613311715,
-0.0385291278,
-0.0006721099,
0.0839965567,
-0.0005677262,
-0.053147912,
0.0262138508,
0.018102454,
0.0872226134,
0.0251534376,
0.0489319116,
-0.0533605143,
-0.0370440036,
0.0014469753,
0.0443728082,
0.0234860517,
-0.0040840418,
0.0068299458,
-0.0545942113,
0.0359520838,
0.0489220023,
-0.0177673567,
-0.0523652509,
-0.0284570269,
0.0271943342,
-0.0116167571,
0.0287953466,
0.0271247327,
-0.0178649016,
-0.02072783,
0.0209944323,
0.0425859578,
-0.0985121056,
0.0236079358,
-0.0179590154,
0.0488041341,
0.023706777,
0.0060347677,
0.0094668241,
-0.0288083069,
0.0325781815,
-0.0345492102,
-0.0107311746,
-0.0423343405,
-0.0124711413,
0.0133176232,
-0.0271841083,
0.0470850058,
0.0417768657,
-0.0213015154,
0.0078877909,
-0.0101400642,
0.0576021001,
-0.0292936489,
-0.0048424569,
0.0236767288,
-0.044106327,
-0.0182228349,
0.0199564155,
0.0351682268,
-0.0306895021,
0.0085307891,
-0.0202843789,
-0.026324302,
-0.0292739049,
-0.0076275473,
-0.0233467482,
-0.028389154,
-0.0011698225,
-0.0236485917,
0.0188001189,
0.0440727472,
0.0415444896,
-0.0092331218,
0.0038222992,
-0.0179323293,
0.0018276015,
-0.032137081,
0.040853925,
-0.0533710383,
0.0063806521,
-0.0610423051,
-0.0566893891,
0.033744026,
-0.0057215751,
0.0090271896,
-0.0236259717,
-0.1329498738,
-0.0689369589,
-0.0087923314,
0.0541383289,
0.0151618766,
0.0385196432,
-0.0066909948,
-0.0960937291,
0.0506294556,
0.0265037678,
0.0066269571,
0.0777423829,
0.0228558555,
-0.0583913028,
-0.0157363638,
0.0365996957,
-0.0466928557,
0.0152919535,
0.031549193,
0.0302945673,
-0.018907493,
-0.0509111062,
-0.0358508714,
0.0051125628,
0.0306363255,
-0.0075506773,
-0.0198789202,
-0.0288084615,
0.0296118557,
0.0345842429,
0.0180887282,
-0.0435803011,
-0.0358661637,
0.0241345782,
-0.0064433436,
-0.023803493,
-0.008544812,
0.0143140284,
0.025261471,
-0.0349350758,
0.0465991646,
0.0002396377,
0.0138305835,
0.0547720604,
0.009174196,
-0.0077142958,
0.0514629558,
0.0228578448,
0.0335500017,
-0.0350050889,
0.0142991515,
0.0100048976,
0.0189956874,
-0.0686798915,
0.0535679795,
-0.0303894002,
-0.0773559213,
-0.0616476387,
0.0254536327,
0.068280369,
0.0155737912,
0.0541122556,
0.0131873237,
-0.0172351897,
-0.032884609,
0.004827023,
-0.007193631,
0.0123124309,
0.0161087681,
0.0260283165,
-0.00272084,
0.0503522716,
0.0629318506,
0.033210773,
0.0453200117,
0.0105485842,
0.0283505339,
-0.0240289085,
-0.0220566504,
-0.0337645747,
0.0100091556,
-0.0034864875,
0.0057008397,
0.0277220476,
-0.0597941726,
-0.0130786579,
-0.0143051753,
-0.0023689035,
-0.0254777111,
0.0202753991,
0.0287489854,
-0.0226945635,
0.0265868008,
-0.0269292183,
-0.0415728316,
0.00099292,
-0.0337758064,
0.0052028117,
-7.097698808e-33,
-0.0484660119,
0.0329709426,
0.0117420005,
0.0153093524,
-0.0267391168,
-0.0371055044,
-0.0371667817,
0.0145466672,
-0.0362865441,
-0.0045559127,
-0.0296211988,
-0.0037827361,
0.0159240663,
0.0010884914,
-0.0235170964,
-0.0056228898,
0.067288354,
-0.0161565561,
-0.0354343727,
0.0313250609,
-0.0034139189,
0.0220567491,
0.0430193804,
-0.0096201235,
-0.0457711481,
0.0260269493,
-0.0293465126,
-0.0005398946,
-0.0202848874,
0.0337625667,
-0.0033623111,
0.045651,
-0.0043412573,
0.0181547403,
-0.0348199159,
-0.0083757304,
-0.0675576702,
-0.0189864803,
0.0366483256,
-0.015251697,
0.0120492997,
-0.0487624705,
0.0407604687,
0.006146132,
-0.0348145403,
-0.0051186457,
-0.0048900894,
-0.0339508019,
-0.0374480262,
-0.0394089185,
-0.0468549579,
0.0258449875,
0.0006639664,
-0.0283598192,
0.0275187101,
-0.0178335607,
-0.0169368926,
0.0015445021,
-0.0172985848,
-0.0076473523,
0.0425639935,
0.1137314513,
0.0403161645,
0.0509698018,
-0.0016824071,
0.0136983478,
-0.0046178773,
0.0271045938,
0.0380398259,
0.0250336882,
0.0941701382,
0.0257181078,
0.0228153076,
0.0183243379,
0.0074846991,
-0.0319768265,
-0.0653875768,
-0.0089192865,
0.0018946372,
0.0378760323,
0.0425057411,
0.0507448502,
0.0163913779,
-0.0197273418,
-0.0598148666,
0.0001869731,
-0.0674489811,
0.0081134411,
0.014701942,
0.0344326571,
-0.0490206853,
-0.0091224136,
0.0468765311,
-0.0422966033,
0.0151952617,
-0.0004273832,
-0.0180811398,
0.0118622081,
0.0118124522,
0.0171354599,
-0.0487448312,
-0.0925296694,
0.05236388,
-0.0099822534,
0.017668074,
-0.0233677793,
0.0111655854,
0.05025886,
-0.0495739877,
-0.0175226983,
-0.0133952117,
0.0251186229,
0.0026114844,
0.060603492,
-0.0025911091,
-0.0061706933,
0.0150931999,
0.0490975864,
0.0008274082,
0.0026988222,
-0.0052606789,
-0.0201972257,
0.0248754993,
0.0568124726,
-0.0400890112,
0.0235511009,
-0.0407683998,
-0.0253791921,
0.0303915609,
-0.0511704423,
-0.0251002796,
0.0150291715,
2.824e-7,
0.0016313872,
0.0592070296,
-0.0393881239,
-0.0350443535,
0.0205891002,
0.0207022224,
-0.0186587721,
-0.0101972083,
0.0297465753,
-0.0120337596,
0.0357498527,
0.0419764929,
0.064214766,
-0.0092940927,
-0.061486721,
0.0224031955,
-0.0518983342,
0.0613433719,
-0.0573205538,
-0.0457499847,
-0.0255514625,
0.1254179925,
0.0237160195,
-0.0146363126,
0.0279662628,
-0.0074228528,
0.0441727638,
-0.0250300206,
-0.0053083627,
0.0310014244,
0.0350097865,
0.0208574962,
-0.0289643817,
-0.0207263865,
-0.0093853334,
0.0521356687,
-0.0417840108,
0.0983116329,
-0.028538337,
0.0030079943,
0.0411616787,
0.010744513,
-0.00632045,
-0.0401446335,
0.0522527471,
-0.0254794024,
0.0024575531,
-0.0575814359,
0.0093602017,
-0.0200829133,
0.0096290465,
-0.0701677427,
-0.0068775457,
0.037617512,
0.0415997058,
-0.0174348913,
-0.0281596724,
-0.0750144348,
0.0100752311,
0.03356985,
-0.0462333225,
-0.0541235842,
-0.0551365055,
-0.0246483535,
0.0271389671,
-0.0289210584,
0.031511046,
2.597764685e-34,
0.062401291,
-0.0305261705,
0.0232178289,
0.004544544,
0.0376627035,
0.0026655879,
-0.0225240681,
0.0270418078,
-0.0033733465,
-0.1143539175,
-0.0263125449
] |
http://arxiv.org/abs/2401.16403v2 | http://arxiv.org/pdf/2401.16403v2 | ViLexNorm: A Lexical Normalization Corpus for Vietnamese Social Media Text | 2024-01-29T18:41:39 | 2024-01-31T07:59:16 | Thanh-Nhi Nguyen, Thanh-Phong Le, Kiet Van Nguyen | cs.CL | cs.CL | Lexical normalization, a fundamental task in Natural Language Processing
(NLP), involves the transformation of words into their canonical forms. This
process has been proven to benefit various downstream NLP tasks greatly. In
this work, we introduce Vietnamese Lexical Normalization (ViLexNorm), the
first-ever corpus developed for the Vietnamese lexical normalization task. The
corpus comprises over 10,000 pairs of sentences meticulously annotated by human
annotators, sourced from public comments on Vietnam's most popular social media
platforms. Various methods were used to evaluate our corpus, and the
best-performing system achieved a result of 57.74% using the Error Reduction
Rate (ERR) metric (van der Goot, 2019a) with the Leave-As-Is (LAI) baseline.
For extrinsic evaluation, employing the model trained on ViLexNorm demonstrates
the positive impact of the Vietnamese lexical normalization task on other NLP
tasks. Our corpus is publicly available exclusively for research purposes. | ViLexNorm: A Lexical Normalization Corpus for Vietnamese Social Media Text: Lexical normalization, a fundamental task in Natural Language Processing
(NLP), involves the transformation of words into their canonical forms. This
process has been proven to benefit various downstream NLP tasks greatly. In
this work, we introduce Vietnamese Lexical Normalization (ViLexNorm), the
first-ever corpus developed for the Vietnamese lexical normalization task. The
corpus comprises over 10,000 pairs of sentences meticulously annotated by human
annotators, sourced from public comments on Vietnam's most popular social media
platforms. Various methods were used to evaluate our corpus, and the
best-performing system achieved a result of 57.74% using the Error Reduction
Rate (ERR) metric (van der Goot, 2019a) with the Leave-As-Is (LAI) baseline.
For extrinsic evaluation, employing the model trained on ViLexNorm demonstrates
the positive impact of the Vietnamese lexical normalization task on other NLP
tasks. Our corpus is publicly available exclusively for research purposes. | [
0.039101053,
0.0468002856,
0.0178660192,
0.0032023843,
-0.0239543226,
-0.0045860941,
-0.0367881916,
0.0208714698,
-0.0013112321,
-0.043088913,
-0.0108568221,
-0.0333608724,
0.0019808479,
-0.0097735142,
-0.0148718236,
-0.0602654293,
0.0346016362,
0.0279477797,
-0.0233062506,
0.0224690698,
0.0519460402,
-0.0033326331,
0.0591696203,
0.0354912952,
0.0048444131,
0.0095784478,
0.0407742225,
0.0306858607,
0.0161207635,
-0.0478431247,
-0.0025398741,
0.0871047601,
-0.0311594922,
0.0249661561,
0.0000018404,
-0.0335875452,
-0.0295243673,
-0.0308931973,
0.0414708294,
0.0380192846,
0.0784597695,
-0.0740676001,
-0.003422007,
0.0378204584,
-0.0473011099,
-0.0540558249,
0.0363140181,
0.0605972297,
0.0758656561,
0.0687990263,
-0.0277808961,
-0.0512903109,
-0.004178571,
-0.0280240364,
-0.0377224423,
-0.0312576257,
0.0074044405,
0.0153320972,
-0.0054944404,
-0.022484215,
-0.0100293458,
0.0393472277,
0.0073768394,
-0.0155144874,
0.0334488451,
-0.0100920489,
0.0009873168,
-0.0135508943,
0.0007056756,
0.0567943305,
-0.003975695,
-0.0459097661,
-0.0029561033,
0.0278227497,
0.0149023477,
0.002106654,
-0.0208348762,
-0.0022526253,
0.0054557156,
0.0542865098,
0.0287609082,
-0.0086588552,
0.037520349,
0.0326169915,
0.0456314422,
-0.0224890336,
-0.0075122132,
-0.037535239,
0.0033253178,
0.0165274814,
-0.0710142329,
-0.015972713,
0.0725054294,
0.0111747002,
-0.016772015,
-0.0009840672,
-0.0156312939,
-0.0530522689,
0.0426975265,
-0.091904372,
-0.015713552,
0.0092559448,
-0.0297406279,
0.0081121922,
-0.0178652648,
0.0227525663,
0.0009849533,
0.0197104253,
-0.0601774827,
-0.0164214075,
-0.0492464714,
0.0322430134,
-0.0593520515,
0.0488647372,
-0.0031738228,
-0.0371505432,
-0.0476981029,
-0.0303496365,
-0.0019055996,
-0.0172627307,
0.0492726974,
0.0435438827,
-0.0286096539,
-0.0039372914,
-0.0100158108,
0.055533953,
-0.0412056632,
0.0355339609,
0.0017726772,
-0.0252131373,
0.0401714221,
0.02866528,
0.0540864132,
0.0037122199,
0.0045709098,
-0.0005032314,
-0.0164540038,
-0.016474802,
-0.0435131416,
-0.0113134244,
0.0044838777,
-0.0059709027,
-0.0665694848,
-0.0802540779,
0.0043807495,
-0.0022305306,
0.0321155414,
-0.0046558562,
0.0229928903,
0.0278212633,
-0.0324511044,
0.1027948931,
-0.0075672227,
0.0099486606,
0.0357082747,
0.0264339168,
0.0395364277,
0.0334038399,
0.0203069057,
0.0218744371,
0.0563939251,
-0.024630351,
0.0246533621,
0.0476099364,
-0.0254615266,
-0.0424183384,
-0.0075264326,
-0.0071562435,
-0.0551491454,
0.0383454747,
0.0176970121,
0.0978317633,
-0.0278712194,
-0.0016300497,
0.017465204,
0.0325565487,
-0.0373775773,
-0.0149614802,
0.0332164057,
0.0044182236,
0.0057749925,
-0.0067023467,
-0.0654453412,
-0.0233943108,
-0.0238640532,
-0.0163068864,
0.0230128914,
-0.0079441322,
-0.025409963,
-0.0374212451,
0.0061869365,
-0.0509663671,
0.0636357591,
-0.0399347022,
0.0295740031,
0.038098406,
-0.0167599693,
0.0418470278,
0.0141482735,
-0.0211833864,
0.0106810983,
-0.0138454139,
-0.0349804573,
0.0245698802,
0.0544301756,
0.0120897871,
0.0003902653,
-0.0348596349,
-0.0104748346,
0.0536797345,
0.0132934451,
-0.0341175012,
-0.0155577501,
-0.0491506904,
0.0134113012,
0.0909040868,
0.0224068202,
-0.0187668335,
0.0399968475,
-0.0071715945,
0.0263152122,
-0.0453204848,
-0.0919503793,
0.053338889,
0.0163418651,
-0.0303195659,
0.0007645365,
0.01055454,
-0.0080135027,
0.0578005165,
-0.0056349388,
0.0003029896,
0.0115613835,
-0.0462610088,
-0.0219555385,
0.0190500487,
-0.01896099,
-0.005778071,
0.0442728996,
-0.098796621,
-0.0029901776,
0.0863696262,
-0.0254432596,
-0.0476575904,
0.0039029554,
-0.0039416957,
-0.0066964319,
-0.0663465112,
0.0289257914,
-0.0411169194,
0.0475101843,
-0.0180906337,
0.01424977,
-0.0336378254,
-0.0021549731,
-0.0131433653,
-0.0625563413,
0.0172599275,
-0.0044520083,
-0.0309726354,
-0.0022906943,
-0.0143306376,
0.0456482247,
0.0368078388,
0.0259197876,
-0.0341976285,
-0.0088800332,
-0.1390756518,
0.0243003499,
-0.0374296159,
0.0396409817,
0.0556549691,
0.0188993122,
0.0001438483,
-0.0491971374,
-0.0096825771,
0.0008482264,
0.0520593598,
0.0091473069,
0.0465199836,
-0.0220104586,
0.0098105026,
-0.0511438139,
-0.002614148,
0.0108479019,
-0.009979927,
0.0022753677,
-0.0631525517,
-0.0068066395,
0.0248730611,
0.0139245456,
-0.0118207792,
0.0380444564,
-0.0632836372,
-0.0189860519,
0.0239724889,
-0.0161088873,
0.0337665193,
0.0114034023,
0.0119565194,
0.0084570702,
0.0241538584,
-0.0203758664,
-0.0345822759,
0.0034480891,
-0.0235301144,
-0.0377773792,
-0.0477410406,
-0.0230254624,
0.0691009536,
-0.0067314729,
0.0288562141,
-0.0229271781,
-0.0289648939,
-0.0582047813,
-0.0096961502,
0.0097074173,
0.0048231725,
-0.0320007876,
0.0181000438,
-0.0326253735,
-0.0173729956,
0.0613363497,
0.0200932492,
-0.0082307924,
-0.0384513214,
0.006433534,
-0.0160221998,
0.0151585275,
0.0000978307,
0.0067690536,
-0.0310724266,
-0.0615171194,
0.0243867915,
-0.0108832074,
0.0256776474,
-0.0379986092,
0.0499539562,
0.0023383512,
0.0054287692,
-0.008689384,
-0.0478400402,
-0.0186547264,
-0.0296327565,
-0.0550510921,
0.0293304473,
-0.0033342282,
0.0111699756,
-0.003318937,
0.0258752014,
-0.0290543195,
-0.0077083292,
0.0078048846,
-0.0106876241,
-0.014371248,
0.0790815875,
0.0293082539,
0.0265229773,
-0.0132683786,
0.0073251226,
0.0465176702,
0.0189948846,
0.0330856368,
0.0217347629,
-0.0260972138,
-0.0318357758,
-0.0006423951,
-0.0336787254,
-0.0033452883,
-0.0201376379,
0.0585558563,
-0.0114656305,
0.0209841542,
0.1172779053,
0.0254460238,
-0.060431242,
-0.0034034296,
0.000094259,
-0.010331694,
0.0117874229,
-0.0060137729,
-0.0753232092,
-0.0392412208,
0.0111263143,
-0.0104308566,
0.035113927,
-0.0393908694,
0.0324281566,
-0.0814579874,
-0.0139931217,
0.0086580813,
-0.0179881807,
-0.0367970578,
-0.0464377366,
0.0051707779,
-0.0618719272,
0.0726260841,
-0.0052672005,
-0.0010012002,
-0.0102089243,
0.0229455195,
-0.035360869,
-0.0485626794,
0.0239369981,
-0.0861357227,
0.0631823838,
0.043979615,
0.0235252939,
-0.013704462,
0.0027310648,
0.0516517535,
0.0208999664,
-0.0042917468,
0.0019713896,
0.0219016746,
0.0271525867,
-0.014931282,
-0.0264805332,
0.0102847954,
-0.0628700703,
0.0043923729,
-0.0056341086,
0.0840308517,
0.0188700538,
-0.025316773,
-0.0050475793,
-0.00592505,
-0.0058843591,
0.052237127,
0.0248920191,
-0.0581181571,
-0.0314769037,
-0.0237800907,
-0.037487369,
-0.053876467,
-0.0508350618,
0.0255846474,
-0.035617426,
-0.0423284732,
-0.0900349244,
-0.0851328969,
0.0008641687,
0.000698091,
-0.0049999589,
-0.014006963,
0.0383054242,
-0.0069175246,
-0.0174749866,
0.0200109314,
-0.0261285435,
0.0045163413,
-0.0286114905,
0.0271830726,
-0.0184887089,
0.011279746,
0.0187313762,
0.0207756907,
-0.0319670998,
-0.0309988279,
0.0200201329,
-0.0230708234,
0.0169738922,
-0.0024485616,
-0.0426588655,
-0.0601948611,
-0.0028512613,
0.0997732952,
-0.0251909718,
-0.0210909527,
-0.0012497096,
0.0225546118,
-0.0060926988,
0.0711978674,
0.0193732884,
0.0054454384,
-0.0150395921,
0.0524415337,
-0.0090400819,
-0.0219910461,
0.0130266165,
0.0506645665,
0.057972528,
-0.0585412942,
0.0047130533,
-0.0350131951,
-0.0691771358,
-0.0566154495,
-0.0043801297,
-0.0015508771,
0.0237922911,
0.0691643134,
0.0291257184,
-0.041153647,
0.0153968539,
-0.0126821958,
-0.012711809,
0.001870439,
0.0239201766,
0.0402080007,
0.0394308493,
-0.0137962606,
0.0036047634,
0.0074005644,
-0.0125355283,
0.0168478768,
0.0224270616,
-0.0332349651,
0.0035036895,
-0.0113181425,
0.0422606207,
-0.0347129442,
0.0113632213,
0.010487847,
0.0183470249,
-0.0408113115,
0.0283681024,
0.0957141221,
-0.0351812951,
0.0929355994,
-0.0089168195,
0.0008477837,
-0.0248317625,
0.0051187021,
-0.0260486584,
-0.0104645332,
-0.0051223924,
-0.0119359195,
-0.0195394717,
0.0380666479,
0.0725858957,
-0.0695891827,
0.0300180633,
0.0094696609,
-0.0211503897,
-0.0263017491,
0.0177918244,
0.0071001546,
0.0378605835,
-0.0487180799,
-0.0062237112,
-0.0037872796,
-0.0742358938,
0.0053708237,
-0.001776234,
0.0007779943,
0.0443986356,
-0.0063482914,
0.0494574532,
0.0231210571,
0.0244742595,
-0.0391220339,
-0.0121241743,
-0.0244056638,
-0.0207323208,
-0.0097708711,
-6.404223997e-33,
-0.0198243521,
-0.0017716191,
-0.033892367,
-0.0098885782,
-0.0001389818,
-0.0153392162,
-0.0140890582,
0.0469893701,
-0.0142100351,
0.0246198922,
-0.0477556549,
0.0119523117,
0.0312142055,
0.004463247,
0.069925338,
-0.0063253106,
0.0146333901,
-0.0090729147,
-0.0127406269,
0.0105942944,
0.0150452731,
0.0448677391,
0.0862711295,
-0.117582716,
0.0428062193,
0.0280916877,
-0.0303500667,
0.0249638259,
0.0765120313,
0.0493218713,
-0.0385455303,
0.0670448095,
0.0441469215,
-0.0011079396,
-0.0286627114,
0.0206703264,
-0.1111295074,
-0.028053252,
0.03477237,
0.0406837389,
-0.0520493574,
-0.016531378,
0.0727537572,
-0.0394590087,
-0.0447688662,
0.0374217667,
0.0191704091,
-0.0306417197,
0.0273729861,
0.0358526595,
-0.0681328773,
0.037551444,
0.0168141779,
-0.0011162369,
0.046697326,
0.0156592075,
-0.0739456341,
-0.0171549525,
-0.0613383837,
-0.0120131616,
-0.0350710303,
0.0449605845,
0.0200795997,
0.008961238,
-0.0357402265,
0.0113755725,
0.0471529216,
-0.0204407349,
-0.0313771255,
0.0201391596,
0.0369883925,
0.0027306369,
0.0335409269,
0.0717423186,
0.0558430739,
-0.0469990149,
-0.0293693729,
0.0371562466,
0.0468222238,
0.0145548852,
-0.0212543085,
0.0140751805,
-0.0172050707,
-0.022468172,
-0.0199935213,
-0.0497713536,
-0.0333862603,
-0.0365113206,
-0.0006378801,
0.0231586881,
-0.0142911263,
0.0298144426,
0.0424009115,
-0.0746226311,
-0.0213694442,
-0.0048870724,
0.0136114974,
0.0195227712,
-0.004298504,
-0.0078172442,
-0.030249102,
-0.033290673,
0.0000838728,
0.0032122037,
-0.001402869,
-0.0084615713,
0.0319548771,
-0.0144236414,
-0.0680001751,
0.017211752,
-0.0104087321,
-0.019754678,
-0.0245657749,
0.0242386237,
0.0169426389,
0.0071112365,
-0.0042317477,
0.0050607878,
0.0456743799,
0.0336114019,
-0.0379226878,
-0.0113833398,
0.0231190752,
-0.0140372952,
-0.0147834029,
0.0448557623,
-0.0486369282,
0.0134048993,
0.0081565576,
0.0025874844,
-0.0044993055,
-0.011609978,
2.538e-7,
0.083514519,
0.0472417548,
0.0770655125,
-0.0002391255,
-0.0075442768,
0.0179513525,
0.0141620347,
0.0240185298,
0.0000858169,
-0.0435848795,
0.0471208952,
-0.0062280935,
0.0478823073,
0.0004660215,
-0.0746990964,
-0.014256889,
-0.0690057874,
-0.0303514767,
-0.0121714678,
-0.0066179722,
0.0084837992,
0.086122185,
-0.0093025416,
-0.0197397862,
-0.0356156789,
-0.0698854327,
0.0289948452,
-0.0171659123,
-0.0181040447,
0.0308329761,
0.0663518906,
0.0049388977,
0.019076148,
-0.0341522805,
0.0133876381,
0.0010268716,
0.0702160448,
0.0697287619,
-0.0204966627,
0.0321569517,
-0.0222808775,
-0.0258966312,
-0.0001424308,
-0.0404968709,
0.0350963213,
-0.000519098,
-0.0189372618,
-0.0562263466,
-0.0025957206,
0.0480282344,
0.068758063,
0.000333821,
-0.0562469102,
0.0478056297,
0.0633536801,
0.0400959104,
-0.0599032044,
-0.0415821746,
-0.0707496405,
-0.0000660789,
-0.0082335938,
0.012190029,
-0.0008392992,
0.0299698822,
-0.0057625654,
-0.0267218724,
0.0276950393,
1.846805284e-34,
0.0451769792,
-0.0116884522,
0.0146254832,
-0.0058135386,
0.0347151794,
-0.0189488269,
-0.0447942205,
0.0038711552,
-0.0259812307,
-0.0653352886,
0.0254831407
] |
http://arxiv.org/abs/2401.16380v1 | http://arxiv.org/pdf/2401.16380v1 | Rephrasing the Web: A Recipe for Compute and Data-Efficient Language Modeling | 2024-01-29T18:19:08 | 2024-01-29T18:19:08 | Pratyush Maini, Skyler Seto, He Bai, David Grangier, Yizhe Zhang, Navdeep Jaitly | cs.CL | cs.CL | Large language models are trained on massive scrapes of the web, which are
often unstructured, noisy, and poorly phrased. Current scaling laws show that
learning from such data requires an abundance of both compute and data, which
grows with the size of the model being trained. This is infeasible both because
of the large compute costs and duration associated with pre-training, and the
impending scarcity of high-quality data on the web. In this work, we propose
Web Rephrase Augmented Pre-training ($\textbf{WRAP}$) that uses an
off-the-shelf instruction-tuned model prompted to paraphrase documents on the
web in specific styles such as "like Wikipedia" or in "question-answer format"
to jointly pre-train LLMs on real and synthetic rephrases. First, we show that
using WRAP on the C4 dataset, which is naturally noisy, speeds up pre-training
by $\sim3x$. At the same pre-training compute budget, it improves perplexity by
more than 10% on average across different subsets of the Pile, and improves
zero-shot question answer accuracy across 13 tasks by more than 2%. Second, we
investigate the impact of the re-phrasing style on the performance of the
model, offering insights into how the composition of the training data can
impact the performance of LLMs in OOD settings. Our gains are attributed to the
fact that re-phrased synthetic data has higher utility than just real data
because it (i) incorporates style diversity that closely reflects downstream
evaluation style, and (ii) has higher 'quality' than web-scraped data. | Rephrasing the Web: A Recipe for Compute and Data-Efficient Language Modeling: Large language models are trained on massive scrapes of the web, which are
often unstructured, noisy, and poorly phrased. Current scaling laws show that
learning from such data requires an abundance of both compute and data, which
grows with the size of the model being trained. This is infeasible both because
of the large compute costs and duration associated with pre-training, and the
impending scarcity of high-quality data on the web. In this work, we propose
Web Rephrase Augmented Pre-training ($\textbf{WRAP}$) that uses an
off-the-shelf instruction-tuned model prompted to paraphrase documents on the
web in specific styles such as "like Wikipedia" or in "question-answer format"
to jointly pre-train LLMs on real and synthetic rephrases. First, we show that
using WRAP on the C4 dataset, which is naturally noisy, speeds up pre-training
by $\sim3x$. At the same pre-training compute budget, it improves perplexity by
more than 10% on average across different subsets of the Pile, and improves
zero-shot question answer accuracy across 13 tasks by more than 2%. Second, we
investigate the impact of the re-phrasing style on the performance of the
model, offering insights into how the composition of the training data can
impact the performance of LLMs in OOD settings. Our gains are attributed to the
fact that re-phrased synthetic data has higher utility than just real data
because it (i) incorporates style diversity that closely reflects downstream
evaluation style, and (ii) has higher 'quality' than web-scraped data. | [
0.0523084551,
0.0581015013,
-0.0007867271,
0.0285827611,
-0.0264507215,
0.0117889885,
-0.021156162,
0.0310645644,
-0.0170674231,
-0.0890326947,
-0.0410288721,
-0.0271266326,
-0.0287640858,
-0.0205738228,
0.028772505,
-0.0398650728,
0.0282822549,
-0.0301397759,
0.0063544963,
-0.0020930611,
0.0213338044,
0.0032087136,
-0.0015890019,
0.0115864035,
0.0137075242,
-0.0003596772,
-0.0112584457,
0.0408201739,
-0.0363321155,
-0.0031920203,
0.0573680177,
0.0694754943,
0.0217514075,
0.0352910236,
0.0000018574,
-0.0618210658,
-0.0204481818,
-0.0052852454,
-0.0394426621,
0.0568443462,
0.0601669773,
-0.002906228,
0.0057703317,
0.0280647092,
-0.04722137,
-0.0231693443,
0.0328306742,
0.0203131996,
0.0850730315,
0.1233300492,
-0.0065165623,
-0.0278651733,
0.0409101769,
-0.0189105365,
0.0138653293,
-0.0198991336,
0.0162739083,
-0.0171441808,
-0.0216187388,
0.0271206461,
-0.0251338985,
0.0185865685,
0.0005843277,
0.0271255989,
0.0125575326,
-0.0102086673,
-0.0541439429,
0.0082406495,
-0.0123731401,
0.1083554551,
-0.0163872596,
-0.0243056528,
0.0146902949,
0.0151921241,
0.0083830496,
0.0340889171,
-0.0519900173,
0.0317662545,
0.0254019853,
0.0032578446,
-0.0288609825,
-0.0177070592,
0.0219229553,
0.0359164253,
-0.0435802229,
0.0291317496,
0.024176443,
-0.0440359898,
-0.0302989539,
0.012468854,
-0.0277577452,
-0.0288110953,
0.0000744132,
0.021846313,
-0.0089504262,
0.0239903573,
-0.0974119827,
-0.0474472716,
0.0295714363,
-0.011140639,
0.0724299848,
0.009583489,
0.0040821168,
0.0388103984,
-0.0253177285,
0.0261132922,
-0.016929552,
0.0423543043,
-0.058885362,
0.0194327664,
-0.0462089851,
-0.0049400609,
-0.0397189707,
0.0586264133,
-0.0156782344,
-0.0150379362,
-0.0214193575,
0.0328672938,
0.0256223399,
-0.0157193244,
0.003013727,
-0.0022017465,
-0.0365673527,
-0.0215349961,
-0.0208427571,
0.029642567,
0.0053157327,
0.0454447493,
0.0210199859,
-0.0784032568,
-0.0173677411,
0.0037303332,
0.0362068266,
-0.0285800602,
0.038451422,
0.060662806,
-0.0224545188,
-0.0165407397,
-0.0257992242,
-0.0399573408,
0.0001987972,
-0.0249566063,
0.0106724119,
-0.0766339451,
0.0466976874,
-0.0049892059,
-0.0073268246,
-0.0402641594,
0.0053562555,
0.0357916616,
-0.0409592465,
0.0662234202,
-0.0398750603,
0.0175874177,
-0.0233860593,
-0.0055743293,
0.0290871579,
0.0549768433,
0.0432832018,
0.0029450119,
0.076394923,
0.0007252183,
0.0115085635,
0.0316504203,
0.038921345,
0.0295379702,
0.0111074997,
0.0228459891,
0.0167952012,
0.0323308557,
0.0399473347,
0.0477395281,
-0.0447668396,
-0.0130880894,
0.0881357118,
0.0196745899,
0.0274568815,
0.0188099425,
0.0246270262,
0.0434368402,
0.0224920604,
0.0196636152,
-0.0562974401,
-0.018335687,
-0.0499982014,
0.0593013428,
0.0315944962,
0.0123790298,
-0.0303316098,
-0.0452534668,
0.0034554948,
-0.0214957558,
-0.0007522387,
0.0218929294,
0.0198096391,
0.0389217511,
-0.0569335222,
0.0565201193,
0.0344797149,
-0.0739998892,
-0.0027415787,
-0.059969075,
-0.0204503108,
0.0696255267,
0.0631147698,
-0.0183045678,
-0.0525686927,
-0.0295930486,
0.0450005122,
0.0311580393,
0.0619952828,
-0.0014195252,
0.0101699382,
-0.0655218735,
-0.0186799783,
-0.0367523283,
0.0123732397,
-0.0186370444,
0.0703265816,
-0.0228488557,
0.0297758318,
-0.0699752867,
-0.0468300097,
0.0414678194,
-0.0077621932,
-0.0305860676,
0.001583724,
0.0059518763,
0.0091576399,
-0.0100831138,
-0.0325332657,
0.029158188,
0.024800716,
-0.0380332284,
-0.018252233,
-0.0402399413,
-0.0287636574,
-0.0485969186,
0.061593391,
-0.0322456919,
0.0070839338,
0.0488797314,
-0.0570075028,
-0.0379924662,
0.0051288595,
-0.0502901487,
0.0499708317,
-0.0485497601,
0.0207658615,
-0.0048526735,
0.029443847,
-0.061953973,
0.0413546115,
0.0054638847,
0.0151845794,
0.0214531627,
-0.0070296135,
-0.0001346134,
-0.0056152535,
-0.0319535062,
-0.0283657871,
-0.0051741698,
0.0270707142,
-0.0086932927,
0.0262297802,
-0.064180553,
-0.0306960121,
-0.0951331183,
0.003629806,
-0.0316868909,
0.0344674252,
0.0582301356,
0.0129363285,
-0.019231379,
-0.0161301848,
-0.022794826,
-0.0410693847,
-0.0256628599,
0.0398624837,
0.0488718003,
-0.0036656363,
-0.0131997745,
-0.0233791955,
-0.0456794128,
-0.0139022665,
0.0266400296,
-0.0119419582,
0.0103714606,
-0.0347017869,
0.0374026969,
-0.0443264209,
-0.0059522646,
0.0012375163,
-0.0464503728,
-0.0218801461,
-0.0457540117,
0.0509919114,
0.0754298791,
0.0275226571,
-0.0079955086,
0.0338544957,
0.0019475494,
-0.0362257026,
-0.0056098499,
-0.0099577168,
-0.0403865725,
-0.0372560956,
-0.037503846,
0.0108477706,
0.0426407382,
-0.0059977053,
-0.012852706,
-0.0147776939,
0.0159396101,
0.0049848924,
-0.0171176102,
-0.0250343122,
0.0235588327,
-0.0091952533,
0.047790125,
-0.0197984781,
-0.02712938,
0.0776561201,
-0.0181839466,
-0.0026039793,
-0.045060914,
-0.0002895225,
-0.0342073478,
0.0348345861,
-0.0098442119,
0.0109108947,
-0.009758749,
-0.0014043179,
-0.0740647912,
-0.0151762757,
0.0388080589,
-0.0647158101,
0.0479982607,
-0.0540662259,
0.0280972254,
-0.0159765184,
-0.0544119366,
-0.0290278308,
-0.0847105086,
0.006121499,
-0.0119214971,
-0.0004194689,
-0.0070621422,
0.0075316755,
-0.0011659202,
0.0226397756,
0.0652854517,
-0.0281210858,
0.0096549634,
-0.0292696729,
0.0616713203,
0.0730893612,
0.0336195193,
0.026351558,
-0.0322903283,
0.0135674058,
0.0426833369,
0.0613374114,
-0.0292302892,
-0.008204394,
0.0128721669,
0.0171874221,
-0.0641753078,
-0.0270573571,
0.0310365278,
0.0471180156,
-0.0528291799,
-0.0200050063,
0.1263609976,
0.0215708595,
-0.0461154394,
-0.006900188,
0.0165902916,
0.0416909009,
0.0312677696,
0.0374914408,
-0.0735818595,
-0.0422392115,
0.0071926448,
0.0158685911,
0.0123636769,
-0.0169818345,
0.0011412482,
-0.1236290187,
0.0221158378,
-0.011266836,
-0.0150646735,
0.0009353033,
-0.042867709,
-0.0006246009,
0.0141926445,
0.0156814102,
0.0186104532,
-0.0468203612,
0.0141332895,
0.0057440395,
-0.017117884,
-0.0594901182,
-0.0061127315,
-0.0325275958,
0.0257271733,
0.0370602347,
0.0042049177,
-0.0147425486,
-0.0437765047,
0.0533464178,
-0.0201506522,
-0.0119621279,
-0.0087594474,
0.0743010938,
0.0007359915,
-0.0011935441,
-0.0077536595,
0.0286952145,
-0.0450757071,
0.0249818526,
0.0056028534,
0.0552182533,
-0.0124493698,
0.0143579757,
-0.0034758679,
-0.0341478102,
-0.0038117711,
0.0597012639,
0.0459562168,
-0.0401209407,
-0.055726178,
-0.025914602,
-0.0019647181,
-0.0730212852,
-0.0330130123,
0.0010944526,
-0.0264063496,
0.0073457453,
-0.085457541,
-0.0341669843,
0.0118373102,
0.039875716,
0.0197097622,
-0.0045236954,
-0.0019128406,
0.0155656701,
-0.0872991234,
0.0276492797,
-0.0659050718,
0.025821045,
-0.0479950719,
-0.0374685787,
0.0349133536,
-0.0115952734,
-0.0055645918,
-0.0128400773,
-0.0897730365,
-0.0619211644,
0.0037504972,
0.0639335513,
0.0217375383,
0.0621256568,
0.0013672366,
-0.0369093418,
0.0391706005,
0.0251842514,
-0.0660305768,
0.0376018211,
0.0558640622,
0.001621471,
0.0193897858,
0.0585548691,
-0.0422083251,
-0.004841045,
0.0246529058,
0.0440785587,
-0.028218383,
-0.0135149313,
0.0469641201,
0.0439019836,
0.0167940632,
0.0030309074,
0.0122704962,
-0.0000364963,
0.0629067644,
-0.0210469309,
-0.0175763536,
-0.0365907252,
-0.0045628762,
0.0216698702,
-0.0005690222,
0.0002883873,
0.0345995203,
-0.0179634262,
0.0267667305,
-0.0420048013,
-0.0076551172,
-0.0138584189,
-0.013800594,
0.0278222486,
-0.0044690683,
-0.0164435748,
0.0330233946,
0.0066592367,
0.0308383759,
-0.0111600552,
0.0071032871,
0.0221405234,
0.007925543,
-0.050492119,
0.0475733727,
-0.0851885006,
-0.0364558697,
-0.0061206766,
-0.0042292578,
0.0267011672,
-0.0189886391,
0.0657890141,
-0.0066130334,
-0.0272767339,
-0.0096669672,
0.0363995582,
-0.0126801059,
-0.0059908279,
0.0309776124,
0.017719971,
-0.0123673305,
0.0137362676,
0.1293749809,
0.0272084717,
0.0457237624,
0.0220030099,
0.017084334,
-0.0452106968,
-0.0430259518,
-0.0023907193,
0.0510969423,
-0.0069846846,
-0.0185259227,
-0.011943765,
-0.0674137771,
-0.0138859833,
-0.006011853,
0.0055453111,
-0.0337711088,
0.0416879468,
0.0704896152,
-0.0144171249,
-0.0151448073,
-0.0426185504,
0.0344847627,
0.0253157746,
-0.0339183249,
-0.0154614272,
-6.257129246e-33,
0.0026063316,
0.0158906206,
0.0136450147,
-0.0077151977,
-0.0317026339,
0.0053656632,
-0.0292518176,
0.0076551926,
-0.0472614206,
-0.0141691165,
-0.0478342883,
0.0162206776,
0.0129616456,
0.0629794374,
-0.0062281983,
-0.025030788,
0.0413110927,
-0.0208435599,
-0.0240582395,
0.0059987633,
-0.0175945833,
0.0327455476,
0.06337291,
-0.0083941352,
-0.0097552557,
-0.032702446,
-0.0220534429,
0.0040090363,
0.0180522557,
0.0415826589,
-0.0152423419,
0.0233766884,
0.0165892933,
-0.0147903888,
-0.0287078843,
-0.0040631057,
-0.0814314112,
-0.0043143942,
0.0223133639,
0.0125002535,
-0.0471325666,
-0.0247520804,
0.0612826869,
-0.0159813575,
0.0136963595,
-0.0158812162,
0.0308459289,
-0.030401025,
-0.0286460631,
-0.0055030165,
-0.0138702113,
0.0068962537,
0.003556133,
-0.0249425955,
0.0122706164,
0.0289365835,
-0.0121034579,
0.0455068722,
-0.0484177805,
0.0284402668,
0.0111409687,
0.0772492588,
0.0383262523,
0.0181977581,
0.0101210745,
0.0412035696,
-0.0152726788,
0.0195572451,
-0.016273452,
0.0194031633,
0.0448740534,
0.0276833437,
0.0189812668,
0.0407782495,
0.0538294241,
0.0130378399,
-0.0539724603,
0.0105598141,
0.0218256228,
0.0021043895,
-0.0013021199,
0.0002345046,
-0.0224500187,
-0.0223416667,
-0.038973961,
0.0221047848,
-0.0451955386,
-0.0211264305,
0.0249071103,
0.0241818726,
-0.0406827927,
-0.00423632,
0.0104003269,
-0.0610977188,
-0.0356263816,
-0.0035621612,
-0.0188310612,
0.0412410796,
-0.0161358025,
-0.0101737734,
-0.0279188063,
-0.0104861315,
-0.0022831589,
0.0075307689,
0.0000900068,
-0.0080702584,
0.0626270026,
0.0266851857,
-0.0793655962,
-0.0023857127,
-0.021800397,
0.0022961327,
0.0163231008,
0.0023843846,
0.0252163429,
0.0329041965,
-0.0064346101,
0.0274134371,
0.0399519466,
-0.0056902785,
-0.0444995016,
-0.0366462991,
0.010697295,
0.0276724491,
-0.0223075598,
-0.0084375432,
-0.0351323262,
0.0027953316,
-0.0101021379,
-0.0742094293,
-0.0211330727,
0.0087267961,
2.562e-7,
0.0535539389,
0.0587094463,
0.0235735942,
0.0238055326,
0.0099199377,
0.0147419134,
0.0193515625,
0.0090765553,
-0.0106523717,
-0.0240605809,
0.0344628207,
0.0266897026,
0.0618022457,
-0.0257829595,
-0.1206066161,
-0.0142797856,
-0.056792099,
0.0305464603,
-0.0492647178,
-0.0411643237,
0.0399709418,
0.1060965955,
0.0141471485,
-0.0364311263,
-0.0175661072,
-0.02306021,
0.0603120625,
-0.0369212963,
-0.0169977769,
0.0471413136,
0.013650015,
0.0322217681,
-0.0014261107,
0.0237152036,
-0.0381239094,
0.0053529241,
0.009179363,
0.0767124221,
0.0153200356,
-0.0160188712,
-0.0117891543,
-0.0625100136,
-0.0004382022,
-0.044417683,
0.0901458487,
0.0014396652,
-0.0023496663,
-0.0132063525,
-0.0389804691,
-0.0327991173,
0.0083931079,
-0.0551611558,
-0.0201042276,
0.0418980233,
0.025704613,
-0.0056537208,
-0.0055015841,
-0.0666639283,
-0.0570567176,
0.0095771709,
-0.0122823454,
-0.0320870392,
-0.0298766997,
-0.010123712,
0.0465223268,
-0.0617415532,
0.0129991947,
2.825391819e-34,
0.0645437986,
-0.0101014841,
0.0354541577,
0.0602423996,
0.0161409862,
-0.007305807,
0.0168897845,
-0.0163137857,
-0.008023547,
-0.1145900786,
-0.0149324909
] |
http://arxiv.org/abs/2401.16367v1 | http://arxiv.org/pdf/2401.16367v1 | TQCompressor: improving tensor decomposition methods in neural networks via permutations | 2024-01-29T18:07:56 | 2024-01-29T18:07:56 | V. Abronin, A. Naumov, D. Mazur, D. Bystrov, K. Tsarova, Ar. Melnikov, I. Oseledets, S. Dolgov, R. Brasher, M. Perelshtein | cs.LG | cs.LG, cs.AI, cs.CL | We introduce TQCompressor, a novel method for neural network model
compression with improved tensor decompositions. We explore the challenges
posed by the computational and storage demands of pre-trained language models
in NLP tasks and propose a permutation-based enhancement to Kronecker
decomposition. This enhancement makes it possible to reduce loss in model
expressivity which is usually associated with factorization. We demonstrate
this method applied to the GPT-2$_{small}$. The result of the compression is
TQCompressedGPT-2 model, featuring 81 mln. parameters compared to 124 mln. in
the GPT-2$_{small}$. We make TQCompressedGPT-2 publicly available. We further
enhance the performance of the TQCompressedGPT-2 through a training strategy
involving multi-step knowledge distillation, using only a 3.1% of the
OpenWebText. TQCompressedGPT-2 surpasses DistilGPT-2 and KnGPT-2 in comparative
evaluations, marking an advancement in the efficient and effective deployment
of models in resource-constrained environments. | TQCompressor: improving tensor decomposition methods in neural networks via permutations: We introduce TQCompressor, a novel method for neural network model
compression with improved tensor decompositions. We explore the challenges
posed by the computational and storage demands of pre-trained language models
in NLP tasks and propose a permutation-based enhancement to Kronecker
decomposition. This enhancement makes it possible to reduce loss in model
expressivity which is usually associated with factorization. We demonstrate
this method applied to the GPT-2$_{small}$. The result of the compression is
TQCompressedGPT-2 model, featuring 81 mln. parameters compared to 124 mln. in
the GPT-2$_{small}$. We make TQCompressedGPT-2 publicly available. We further
enhance the performance of the TQCompressedGPT-2 through a training strategy
involving multi-step knowledge distillation, using only a 3.1% of the
OpenWebText. TQCompressedGPT-2 surpasses DistilGPT-2 and KnGPT-2 in comparative
evaluations, marking an advancement in the efficient and effective deployment
of models in resource-constrained environments. | [
-0.0266378764,
0.100330621,
-0.0031069857,
0.0200113561,
-0.0206134208,
0.0262322538,
-0.0039783879,
0.0380873308,
-0.0272693094,
-0.0065362025,
-0.0251499712,
-0.01796964,
0.0136934239,
0.0083895512,
0.0185573455,
-0.0144025395,
0.003175887,
-0.0148936063,
-0.0395653136,
0.0004992429,
0.0045731743,
0.0250431634,
-0.0005951959,
0.013024101,
-0.0263070781,
-0.0274559464,
-0.01166509,
-0.0288913902,
-0.0257275086,
0.0199708864,
-0.0115899388,
0.0584917739,
0.0226323251,
0.0226464532,
0.0000023848,
0.0518895499,
0.0092096636,
0.0128121804,
-0.0500192419,
0.0701699257,
0.0322871916,
0.0053809523,
0.013088136,
0.0075777718,
-0.0598905347,
-0.0578036085,
0.043456763,
0.0746249259,
0.0804699659,
0.0417985693,
-0.0118433507,
0.0074795573,
0.0182527732,
-0.0190804955,
-0.0425568372,
-0.045271758,
-0.0209681857,
-0.0482866243,
-0.0648072362,
0.0252437685,
-0.007096027,
0.0252814721,
0.0193939973,
0.0561993048,
-0.0261029601,
0.0369938239,
-0.0151179368,
-0.0231778789,
-0.0204339046,
0.0579868965,
-0.0141572077,
-0.0235076007,
-0.010472578,
0.0155241657,
-0.0048632878,
-0.0057629202,
-0.0437330753,
0.0027302837,
0.0451394692,
0.0254133549,
-0.0037989044,
0.0010583815,
0.0360945202,
-0.0082159573,
-0.0330200046,
0.0366091914,
0.0422082879,
-0.0094585335,
-0.0315454453,
-0.0149848126,
-0.0199369863,
-0.0068632131,
-0.0030896172,
-0.01187489,
0.0338496454,
-0.0311989337,
-0.044917468,
-0.0580646172,
-0.018073637,
-0.0655854046,
0.0577152185,
0.0479582325,
0.0227438435,
0.0364634246,
-0.0429061726,
0.0603332557,
-0.0044684159,
-0.0261510648,
-0.0199727155,
-0.0087344972,
-0.0507181734,
-0.0269789826,
-0.0508708358,
0.0124336891,
-0.0521560311,
-0.028223265,
-0.0135815246,
-0.0048464392,
-0.0254583005,
-0.0464588851,
0.0630789623,
0.0321192406,
0.0210228693,
0.0179581326,
-0.0140224621,
0.0503029227,
-0.0118073775,
-0.0069025364,
-0.0327290632,
-0.0230196826,
-0.0466908813,
0.013938373,
0.0115952902,
-0.0251837149,
0.0422197133,
0.0440097898,
0.0015238703,
-0.0598565079,
0.0143154059,
0.0035072812,
0.0156037593,
-0.056911964,
0.0158375576,
-0.0308365077,
0.1077547595,
-0.0408530869,
-0.0087636309,
-0.0554551892,
-0.0141293686,
0.0035690621,
-0.069697246,
0.0407697447,
0.0013045026,
0.0190889705,
-0.0109095508,
0.0146196121,
0.0479967594,
0.0408969373,
0.0142822349,
0.0003447472,
-0.0021451954,
-0.0543635003,
-0.0023061356,
0.0203690901,
0.048820775,
-0.0189878624,
-0.0212240666,
-0.0347450562,
0.0167193282,
0.009366204,
0.0435864218,
0.0817847326,
-0.0819447488,
0.0146781858,
0.0492355041,
0.0494697727,
0.0527943894,
0.0785998404,
0.0041764029,
-0.0108131021,
0.0126468521,
-0.0235119369,
-0.0612077974,
0.0136083309,
-0.0176565312,
-0.0366515219,
0.0311361775,
0.0591794327,
-0.0460833497,
-0.0140258316,
0.0228151381,
-0.0168655477,
0.0690131634,
-0.01115283,
0.0356758125,
-0.0047079641,
-0.031062074,
0.0919479504,
0.0032654942,
-0.0604586154,
0.0121489624,
-0.0361919925,
-0.0373473987,
0.0176903587,
0.0443417914,
0.0155580733,
-0.0082627386,
0.0045306752,
-0.0444931984,
-0.0486528352,
0.0365675762,
0.0003018213,
0.0392885841,
-0.0652614608,
-0.0132924626,
-0.0130851511,
0.0303364135,
-0.0408443101,
0.0247182455,
0.0196868088,
0.0347741209,
-0.0310913846,
0.0092098219,
0.0290020332,
-0.0084619382,
-0.0726164654,
0.0301235784,
0.0089413123,
0.0125010097,
0.025715271,
0.0017194261,
-0.0109811844,
0.0401682071,
-0.0644285008,
0.0057077771,
-0.0369401313,
-0.0213737283,
-0.0729612336,
0.0655074492,
0.0119533716,
0.0357324742,
0.0655688122,
-0.0638006851,
-0.0539431907,
0.0602994189,
-0.0053764018,
0.03966498,
-0.0479079857,
-0.0276378188,
0.000845906,
0.0219077263,
-0.0050024488,
0.0268495623,
-0.0403569415,
-0.0178216211,
-0.0063362238,
-0.0719252154,
-0.0039940081,
0.021946812,
-0.0273599364,
0.0056901532,
0.0081349127,
0.0154519575,
0.0020888629,
0.0229343381,
0.0444963612,
-0.014941833,
-0.076247178,
0.0243387558,
-0.0263769459,
0.0323235951,
0.0479857661,
0.0270391423,
0.0244098138,
0.0183850937,
-0.0298729744,
-0.0057459548,
0.0076778806,
-0.0091635948,
0.0042735697,
-0.0231724624,
0.0064085103,
-0.00508297,
-0.0740714744,
-0.0026056645,
0.0598406382,
-0.0830492303,
-0.0016176691,
-0.0087250071,
0.019938441,
-0.0165263806,
0.029350806,
-0.0160773695,
-0.04828972,
-0.0147597408,
-0.0070509454,
0.0528077222,
0.0323645063,
-0.0032993876,
0.0494429804,
0.0481215306,
0.0311543085,
-0.0233313013,
0.0110690286,
-0.0417492539,
-0.0312323719,
-0.0112709738,
-0.0310570113,
-0.0195004717,
0.0621063709,
-0.0026665614,
-0.0185576081,
-0.0682387426,
-0.020696599,
-0.0323754735,
0.0101520875,
-0.0646795481,
-0.0152637511,
-0.0040704971,
0.0321882144,
0.0185729302,
0.0141406106,
0.0441073924,
-0.0305840746,
-0.0009554585,
-0.0429038182,
-0.0503772274,
-0.0271466505,
0.0214443374,
-0.0055502984,
0.0464274622,
-0.0649557486,
0.0347135849,
-0.0611931123,
-0.003048257,
-0.0022545576,
-0.0354904719,
-0.0231960844,
0.0318921022,
-0.0304901376,
-0.0103391074,
-0.0239112116,
0.0197375547,
-0.0191500988,
0.0122183664,
0.0335030183,
0.0416571572,
-0.0082610762,
0.0261498913,
0.0454335548,
-0.0023613912,
0.0586799569,
-0.0136454143,
0.0372884385,
-0.0065956712,
0.0438935421,
0.0332570337,
-0.0118180737,
0.0559825636,
-0.0101102162,
-0.0018585358,
0.0791053623,
0.0517471507,
0.0038603819,
-0.0327843875,
0.0027002685,
0.0120597091,
-0.0014892844,
-0.0452004783,
-0.0076541454,
-0.0280823335,
-0.0455067419,
0.0098560816,
0.0711179674,
0.0180184338,
-0.0570435263,
0.0125714848,
0.0685887635,
0.0231455564,
0.0042886515,
0.0317617469,
-0.0536547564,
-0.0060916552,
0.0040001962,
0.0000813354,
0.0506071188,
-0.0124372318,
0.0144335069,
-0.0526438467,
0.0288952664,
0.024689544,
0.0281337555,
-0.0417835936,
-0.0495766476,
-0.0226582512,
-0.025921898,
0.0107889008,
0.0255142767,
0.0117009738,
0.0084908456,
-0.006570891,
-0.0015102869,
-0.0561802499,
0.0028904704,
0.0070823915,
0.0337780304,
0.0025124454,
0.0271937419,
-0.0159112588,
-0.0692839548,
0.0413021594,
0.0079205716,
-0.0079004923,
0.0183668956,
0.0162460357,
0.0200916603,
-0.0167691689,
-0.0216497146,
0.03312001,
-0.0574934594,
-0.0234539974,
-0.0581411198,
0.1123890877,
0.0206300151,
-0.0606392063,
-0.0233937055,
-0.0387866385,
-0.0275303442,
0.0732885972,
0.0487133153,
-0.0293581579,
-0.0598249435,
0.0469292775,
0.0288136993,
0.0069289566,
-0.0370875187,
-0.0116394674,
0.0458016992,
-0.0363368914,
-0.0260772742,
0.0184474513,
0.0020247297,
0.0287106857,
-0.0276731271,
-0.003986117,
-0.0074306368,
-0.011765697,
-0.0726386383,
0.0450501628,
0.0108240256,
-0.0274281204,
-0.027261477,
0.0222142413,
0.0183907151,
0.0153978374,
-0.0289050229,
0.0085436059,
-0.0898115113,
-0.0352161713,
-0.0109251468,
0.0539318956,
0.0120173264,
-0.0096451519,
0.0105099306,
-0.0398239568,
0.076334551,
0.0173579156,
-0.0681524128,
0.056266658,
-0.0006300356,
-0.0651209503,
-0.0011195206,
0.024425,
0.0028472894,
0.0048411344,
-0.0129985269,
0.0463426374,
-0.0138006527,
0.0102846855,
0.0247103423,
0.0213670582,
0.0463729464,
-0.0104258647,
0.0543977767,
-0.0272416156,
-0.0025632936,
0.0159257799,
0.0169329513,
-0.0689397305,
0.0271917041,
0.0371147431,
0.0392073132,
-0.0161689892,
-0.0391955376,
0.0725128651,
0.0095071131,
0.027429346,
-0.0336855911,
0.0006050137,
-0.0275663715,
0.0476665162,
0.0189406946,
-0.0184990894,
-0.0134415096,
-0.034964826,
-0.010596877,
-0.0590915792,
-0.0058639371,
0.0100170989,
0.0084375059,
-0.0569017008,
0.0758678317,
-0.0395299979,
-0.0847397149,
-0.0308755245,
-0.0221236944,
0.0223524496,
0.0147085665,
0.0113706253,
-0.0434848741,
-0.0294880159,
0.0009556038,
-0.0008903825,
0.0083739636,
0.0432808176,
0.0025963401,
0.0091231409,
-0.0153587703,
-0.0074869329,
0.1041154936,
0.0128704039,
0.0319185928,
0.0026187927,
0.0208819415,
-0.0267523192,
-0.0263792295,
-0.0569275469,
0.0819568485,
0.0005405307,
0.0158929806,
-0.0145195238,
-0.0695342571,
-0.0341632701,
0.0030957626,
-0.0262097027,
-0.03704964,
0.0582821034,
0.0274343211,
0.0034410788,
0.0111097367,
0.0097300252,
-0.0272416547,
0.0198409967,
-0.0646276027,
-0.0213867836,
-6.829365039e-33,
-0.0050265733,
0.0402636006,
0.0237185862,
-0.0141028101,
-0.0343249477,
0.0090840142,
-0.003797448,
-0.0175433811,
-0.0064360038,
0.0166403186,
-0.0405389071,
0.0241192859,
-0.0225048866,
0.0230603945,
0.0049504125,
0.0024806596,
0.054603871,
-0.0594432019,
-0.0057762489,
0.0584868193,
0.0161723625,
0.0207896698,
0.0776425824,
-0.0572038293,
-0.0772226825,
0.0180682261,
0.0059813093,
-0.0032265924,
0.0259960573,
0.0256921779,
-0.0721986815,
-0.0034178845,
0.029153049,
-0.0379996784,
0.0026124706,
-0.0169353504,
-0.058236476,
-0.0286472924,
0.0105633028,
-0.0232587513,
-0.0454747528,
-0.0311006438,
0.0277995095,
0.0181370992,
-0.0170913767,
-0.0161113404,
-0.0226844121,
0.0285744499,
-0.0282988716,
-0.0233783498,
0.0163538493,
-0.0056215897,
-0.0089416141,
-0.0061709145,
0.0671380758,
0.035274826,
-0.0584980994,
-0.0228070952,
-0.0856954828,
-0.0022702147,
0.0218121689,
0.0394767299,
0.0321474858,
-0.0138733229,
0.0355549008,
0.0219807271,
0.0117579019,
-0.002847834,
-0.0624081492,
0.0017867524,
0.0459784456,
0.0411364809,
0.0547042154,
0.0192304812,
0.0467992425,
-0.04033361,
-0.0543321967,
-0.0216906667,
0.0044748778,
-0.0050325724,
0.0251549911,
-0.0067527276,
0.0046478491,
-0.0454087481,
-0.0452423021,
-0.0371195376,
-0.0266810693,
-0.0133508556,
0.0100068571,
0.037954025,
-0.0524085723,
-0.0358660072,
0.0338988751,
-0.080946371,
0.0279694181,
0.0664827004,
-0.0278580319,
0.0337127559,
0.0419186093,
0.015697293,
-0.0475334302,
-0.0116399173,
-0.0234951936,
-0.0396499448,
0.0227159895,
0.0107912635,
0.0069793882,
0.044461105,
-0.0559903942,
-0.01569256,
-0.0061165416,
0.0115160383,
0.0362828448,
-0.0008227064,
0.0192308035,
0.0233566817,
0.0105401156,
-0.0033194833,
0.0489115119,
0.056894403,
0.0078062727,
-0.0089397114,
0.0555295534,
0.0259624198,
0.020777015,
-0.0220476743,
-0.0570523478,
0.0087211858,
0.0728987455,
-0.0643119663,
-0.027712021,
0.0229977425,
3.024e-7,
0.0401721485,
0.0334013477,
-0.0369452126,
-0.0231630728,
0.0301046576,
0.0209043268,
0.041785378,
0.0247931127,
-0.0081258314,
0.0159422103,
0.050241217,
0.0402667448,
0.0045732171,
0.0005241648,
-0.0458873063,
-0.0217206124,
-0.0697757006,
-0.0237239376,
-0.0411564261,
-0.0252649765,
0.0252059586,
0.1164871156,
0.0398388654,
-0.0258959215,
0.0093177827,
0.0207277387,
0.0553829558,
-0.0409891382,
0.0244799592,
0.0044878614,
0.022520503,
-0.006928707,
-0.0227904581,
-0.001187058,
-0.0078228237,
0.0467547365,
-0.0045906077,
0.0784735531,
-0.0137113687,
-0.0559938625,
0.0253319722,
-0.0242252629,
-0.0134121887,
-0.0524495281,
0.0743863806,
0.0044398499,
-0.0121023813,
-0.0628191009,
-0.0601382032,
-0.0361877866,
0.0458978638,
-0.0147372251,
0.0369227827,
0.0499904379,
0.0036161449,
-0.0105943372,
0.0031522885,
-0.0945151895,
0.0146223474,
0.0129766986,
-0.0216129553,
-0.0250726026,
-0.0196596663,
-0.0258890633,
0.0846180469,
-0.0038076821,
-0.0185266826,
3.162122048e-34,
0.0323584154,
-0.0089832628,
0.0595192052,
0.0595342368,
0.0099259838,
-0.0387528799,
0.0161430724,
-0.0121990982,
-0.0202426538,
-0.1061476842,
-0.0363366976
] |
http://arxiv.org/abs/2401.16349v1 | http://arxiv.org/pdf/2401.16349v1 | ConFit: Improving Resume-Job Matching using Data Augmentation and Contrastive Learning | 2024-01-29T17:55:18 | 2024-01-29T17:55:18 | Xiao Yu, Jinzhong Zhang, Zhou Yu | cs.CL | cs.CL, cs.CY | A reliable resume-job matching system helps a company find suitable
candidates from a pool of resumes, and helps a job seeker find relevant jobs
from a list of job posts. However, since job seekers apply only to a few jobs,
interaction records in resume-job datasets are sparse. Different from many
prior work that use complex modeling techniques, we tackle this sparsity
problem using data augmentations and a simple contrastive learning approach.
ConFit first creates an augmented resume-job dataset by paraphrasing specific
sections in a resume or a job post. Then, ConFit uses contrastive learning to
further increase training samples from $B$ pairs per batch to $O(B^2)$ per
batch. We evaluate ConFit on two real-world datasets and find it outperforms
prior methods (including BM25 and OpenAI text-ada-002) by up to 19% and 31%
absolute in nDCG@10 for ranking jobs and ranking resumes, respectively. | ConFit: Improving Resume-Job Matching using Data Augmentation and Contrastive Learning: A reliable resume-job matching system helps a company find suitable
candidates from a pool of resumes, and helps a job seeker find relevant jobs
from a list of job posts. However, since job seekers apply only to a few jobs,
interaction records in resume-job datasets are sparse. Different from many
prior work that use complex modeling techniques, we tackle this sparsity
problem using data augmentations and a simple contrastive learning approach.
ConFit first creates an augmented resume-job dataset by paraphrasing specific
sections in a resume or a job post. Then, ConFit uses contrastive learning to
further increase training samples from $B$ pairs per batch to $O(B^2)$ per
batch. We evaluate ConFit on two real-world datasets and find it outperforms
prior methods (including BM25 and OpenAI text-ada-002) by up to 19% and 31%
absolute in nDCG@10 for ranking jobs and ranking resumes, respectively. | [
0.0406825617,
0.079913877,
-0.0084490059,
-0.0176154412,
-0.0178318713,
-0.0265288986,
-0.0273170508,
0.0134641388,
-0.0069274744,
-0.0202861354,
-0.0171509739,
-0.0385757573,
0.0124583459,
-0.0069265752,
0.026916014,
-0.0588373765,
0.0015905503,
0.04120766,
-0.0405139625,
0.0212488938,
0.0132293655,
0.0710506588,
-0.0024040879,
0.0536179431,
0.0205468107,
-0.0058525945,
0.0081339562,
0.0260502473,
0.0725075752,
-0.0078016217,
0.032208398,
0.0101849195,
0.0171121825,
0.0153510058,
0.0000019505,
-0.0296246745,
0.0044446043,
0.0397894233,
-0.0222916994,
-0.0289686173,
0.0439566895,
-0.0159807783,
0.0449391529,
0.0484962873,
-0.0302441046,
0.0090453848,
0.016868934,
0.0231557358,
0.0393191613,
0.0471434444,
-0.0015628013,
-0.0279529057,
0.0182569195,
-0.0485664159,
0.0171061307,
0.0017485819,
0.0595770665,
-0.0714520365,
0.0391580909,
0.0155815436,
0.0032514434,
-0.0097371852,
-0.0284360163,
-0.0523907579,
0.043478407,
0.0523918644,
0.0046665925,
0.0347095691,
0.0345227383,
0.0817256793,
-0.0001460807,
-0.008514341,
-0.0275366381,
-0.0408681743,
0.0248908158,
0.014548908,
-0.0316125304,
-0.0345753804,
-0.0228354428,
0.0048983591,
-0.0175272133,
0.0328531563,
0.0165577196,
-0.0065534543,
0.0082914354,
0.0341268331,
-0.0199200436,
-0.0633790866,
0.0137584694,
0.0019445278,
0.0380481631,
-0.0390412994,
0.0420587137,
0.0187178086,
-0.0518210866,
-0.0075667389,
-0.0525216945,
-0.0440217033,
0.0238030851,
-0.0109666232,
-0.0052239704,
0.015809834,
-0.039157141,
0.0576602742,
-0.0078451289,
-0.0002620822,
-0.0073404759,
-0.0267061703,
-0.0436872914,
-0.0490709767,
-0.0779525638,
0.0005560818,
-0.0320968851,
0.0221704189,
-0.0382603444,
0.0276566707,
-0.0641767532,
-0.0240086112,
-0.0472701639,
-0.0064274636,
0.049751509,
0.0456987731,
-0.0437425785,
0.0377879292,
-0.0857536793,
-0.0179054514,
0.0213632155,
0.0139745623,
-0.0229198802,
-0.0146487411,
0.0239133444,
-0.000434427,
0.0517356768,
-0.0536910221,
0.0064897877,
0.0332218483,
0.043548245,
-0.0194812249,
-0.028241884,
0.0026729763,
0.0066118306,
-0.0233350489,
-0.069853045,
-0.0324578844,
0.0691426247,
-0.0059239781,
-0.0400110371,
-0.0565784648,
0.031665042,
0.0397307053,
-0.0543469302,
0.0713098422,
-0.0144359441,
0.0039439076,
0.0078307688,
-0.0220570844,
-0.0364530571,
0.0008662,
0.0227098893,
0.0638921335,
0.0023732777,
-0.0081728203,
0.0462333858,
0.0236097947,
0.0070903879,
-0.0071799923,
-0.0087768352,
0.0713784173,
-0.0039133462,
0.0343272835,
0.0608391911,
0.0053180889,
-0.0473917499,
-0.008632699,
0.0381869189,
0.0125333555,
0.0622139908,
0.0216480475,
0.0575516522,
0.0674659684,
-0.0309320912,
-0.0154802436,
-0.0193968993,
0.0491643995,
-0.0089778965,
-0.0157849006,
0.0102898655,
0.0364335291,
-0.0299438778,
-0.0010069172,
0.0236791689,
-0.0087001063,
-0.0061840345,
0.0077940039,
0.0234088041,
0.0028451092,
-0.0527456179,
-0.0297137666,
0.0430155173,
-0.0025629695,
0.0021236229,
-0.0168616064,
-0.0522459969,
0.1119443551,
0.046972122,
-0.0375774205,
-0.0278013796,
-0.0585981943,
-0.013269959,
0.0135136144,
0.0558814593,
-0.0112743918,
0.0481395125,
-0.0445852317,
-0.0324411429,
-0.0152890049,
-0.0128613543,
0.0401132256,
0.0288183335,
0.044336345,
0.0213477574,
0.0085631814,
-0.028892858,
0.0278470218,
-0.0437596552,
-0.0013112059,
0.0365046076,
0.0290270578,
-0.03012776,
0.0112161506,
0.0027853951,
0.00375406,
0.0137321977,
-0.0282996334,
-0.0069852229,
-0.0253256578,
-0.0063339546,
-0.0357060134,
0.0164581444,
-0.0058084195,
-0.0314498022,
0.0266886447,
-0.0541450568,
-0.012550219,
0.005690278,
0.0090158479,
0.01516065,
-0.0341643915,
-0.0029110378,
-0.0078273406,
0.0290669352,
-0.0336285867,
0.0809632689,
0.0267788377,
0.0232856125,
0.0075608715,
-0.0106782736,
0.008210226,
-0.1361936778,
-0.0120514873,
-0.0746054277,
0.0141330296,
0.0100572612,
-0.0252456106,
-0.0311837345,
-0.0390571691,
-0.0173727609,
-0.050254561,
0.0450742692,
0.0374850258,
0.0063972133,
0.0151789375,
0.0033336896,
0.0181235112,
-0.0172109753,
-0.0545535721,
-0.0152406935,
0.0245662071,
-0.0174532887,
0.0484768227,
0.0283400808,
-0.0421827361,
-0.050024461,
-0.0171439573,
-0.0154143972,
0.0110157132,
0.001278556,
0.0171213038,
-0.0661722943,
0.0251921546,
-0.0347545445,
0.0340218991,
0.003696972,
0.0031337016,
-0.0243695509,
0.0130444132,
-0.0382538922,
0.0319798626,
-0.0095594674,
0.0070030075,
0.0097140241,
0.041180186,
-0.0248152297,
0.0179108419,
0.0052461689,
0.0628975704,
-0.0122808358,
-0.0584644787,
-0.0445391275,
-0.0029023932,
0.0171856061,
-0.0293783862,
-0.0112005584,
0.0073589929,
0.040431425,
-0.0000598872,
-0.0406780466,
-0.0020885153,
0.0036522821,
0.0162500273,
-0.0505199842,
0.0020059147,
0.0568382591,
-0.0011691229,
0.0148034254,
0.0195716135,
-0.0867031515,
-0.0640751943,
0.0275641903,
-0.0034547637,
-0.0065570283,
-0.0230613071,
-0.0219387338,
-0.0506502129,
-0.0400481112,
-0.0151654771,
-0.025652675,
-0.010189753,
0.0151655739,
0.0060962061,
0.0077115134,
-0.0727954656,
-0.0641076937,
-0.0455608629,
-0.0072672511,
0.0521250404,
0.0417275392,
-0.008586471,
-0.0639723763,
-0.0154437097,
-0.0165581685,
0.0476118289,
0.0097431969,
-0.0419199727,
-0.0283422489,
0.0594175495,
0.0749199763,
0.0362211242,
-0.0021474508,
-0.0110689588,
-0.0318250358,
-0.003780433,
-0.0015802577,
-0.0190765951,
-0.0425132439,
-0.0489210673,
0.0078760739,
-0.0492730141,
-0.0006168414,
-0.0100364424,
-0.0403360352,
-0.0155118732,
0.0246176831,
0.0630030707,
-0.005528545,
-0.0402048454,
0.0220639147,
0.0101618404,
-0.0290906988,
0.0357290059,
0.0786507279,
-0.0348243788,
0.0000148416,
-0.0150312614,
0.0442029946,
0.0222411454,
-0.0431769527,
-0.0099274842,
-0.113775745,
0.0501217209,
-0.0359703377,
-0.0277882311,
-0.0181323476,
-0.0548222736,
0.0073754238,
-0.0472616628,
-0.0102886781,
-0.0163785089,
-0.0265502259,
-0.0055633266,
0.0220743734,
0.0728524327,
-0.0574658848,
0.02134973,
-0.0415871926,
0.0129764322,
0.0020779278,
0.0648528636,
-0.0008644291,
0.0039595356,
-0.0004711763,
-0.0029247322,
0.0240086261,
0.0408841446,
0.069501482,
-0.0295181442,
-0.0132839913,
-0.0247306526,
-0.0184563324,
-0.0286234245,
0.0154690454,
0.0273860246,
0.0575778224,
-0.0068205534,
0.0268223919,
-0.0334814712,
-0.0140242167,
0.0262314808,
0.0829578042,
0.0078346441,
-0.0546403714,
0.0178581327,
-0.0074984101,
-0.0032269189,
-0.0564449653,
0.0075275348,
-0.0087305568,
-0.0178724639,
0.0601882711,
-0.0405104235,
-0.0522454381,
0.0120583158,
0.0012355471,
0.0124496771,
0.0494588688,
-0.0416167006,
-0.0301736295,
-0.0406222828,
-0.0123805869,
-0.0398571119,
0.0751334354,
-0.0179119464,
-0.0237060077,
-0.0125653287,
-0.0737430453,
-0.0559089109,
-0.0724730641,
-0.0801833272,
-0.0016089779,
-0.0619093291,
0.0037162667,
0.0207509324,
0.0693745613,
-0.0408073962,
-0.0414041989,
0.0080095427,
0.0941316634,
-0.0324299634,
-0.0145796873,
-0.0335568115,
-0.0081206905,
-0.0118525047,
0.0108685382,
-0.0454422086,
0.0200892556,
0.0272377711,
-0.0146639012,
0.0107904235,
-0.0504109487,
-0.0139388647,
0.0077052065,
0.0349082388,
0.030861197,
-0.0066377819,
-0.0773912147,
0.0276939459,
-0.0149173131,
0.0098138424,
-0.0425716899,
0.0015026432,
0.0665130317,
0.0387381166,
0.0077826832,
-0.0375419855,
-0.0190317668,
0.0608471036,
-0.085913755,
-0.0127167599,
-0.0422835238,
0.0242520124,
0.0086539332,
0.0005387992,
-0.0042542904,
0.0091095157,
0.0323086977,
0.0557985082,
-0.0162309892,
-0.0142884944,
0.0221740212,
0.02475008,
-0.0070825443,
0.031086592,
-0.0533657856,
-0.0101422006,
-0.0425425842,
0.0093068397,
0.0478290655,
-0.0642711073,
0.0193589553,
0.0031329978,
0.0016016805,
-0.0133072641,
0.0014747328,
0.002562026,
-0.0140697332,
0.0079920851,
0.0165857282,
0.014604182,
0.0458541363,
0.0621586964,
0.02419677,
0.0227996185,
-0.0304272361,
-0.0326484181,
-0.0160282664,
0.0040069111,
0.0537866019,
0.1012890562,
-0.0336874798,
0.0111212144,
-0.0376457535,
-0.0450456589,
0.0235372446,
-0.0340676196,
-0.0150140179,
-0.0169155784,
0.0237752646,
-0.0038831735,
0.0213984717,
0.0097695375,
-0.0400290787,
0.0106535088,
0.0064034038,
-0.0247225445,
-0.0595390201,
-5.907350616e-33,
-0.0081275329,
-0.0240553617,
0.0205768663,
0.0116254054,
-0.0128183616,
0.0151983313,
0.0376540758,
-0.0725911558,
-0.0251636356,
-0.0086373044,
-0.0305186491,
0.0164310597,
0.0035041161,
0.0129690077,
0.0036825216,
0.0090314094,
0.0479055047,
0.0394284725,
0.0157023482,
0.0041123219,
-0.0057082856,
0.0663102493,
0.0344517455,
-0.0887931138,
-0.0042857951,
-0.0209973436,
-0.0077242353,
0.0338342786,
0.015112225,
0.0757508576,
-0.0147166839,
0.0294393953,
0.0416803509,
-0.0276828632,
0.0312316045,
-0.0297457092,
-0.0701904073,
-0.0051774997,
0.0140913296,
-0.003280432,
0.0004166673,
0.0016525129,
-0.0146392258,
-0.0119715529,
0.0127634471,
0.008200339,
0.0751804188,
-0.0399628803,
-0.006259643,
0.0590605661,
0.0147413071,
0.0047755879,
-0.0161334481,
0.0020729136,
-0.0041072499,
0.0260945577,
0.0142448284,
0.1019901112,
-0.1312822253,
-0.0224727467,
-0.0178819057,
0.0432702042,
-0.0055512637,
-0.0055725891,
-0.0238350034,
0.0501701161,
0.0630696192,
0.0379697308,
0.0419192165,
-0.0027238289,
-0.0110081527,
0.0113435816,
-0.0084914826,
0.0631036088,
0.0788608715,
-0.0255346633,
-0.0528035499,
0.0530718081,
0.0128065655,
-0.0084159784,
-0.0069294819,
0.0464073606,
-0.0223900434,
-0.0315904208,
0.0032109902,
-0.0173906796,
-0.0477151647,
-0.0007271827,
0.0244758297,
0.0394765101,
0.0151876556,
0.0550341383,
-0.0244105104,
-0.0507801585,
-0.0051514101,
0.026511116,
-0.0469062328,
-0.0291904397,
0.0195644498,
-0.0257076547,
-0.0773761794,
-0.0379446708,
0.0152912624,
0.0161860678,
0.035087321,
0.0766897351,
0.0625282153,
-0.0033306538,
-0.0516622774,
-0.0061803721,
-0.017248502,
-0.0032921438,
-0.0444382019,
0.0042553181,
0.0079591004,
0.0265788697,
0.0217078105,
0.0517370217,
0.0638037547,
0.0258571729,
-0.0015723112,
-0.0912685618,
0.0479996912,
0.0134423422,
0.0178806707,
0.0012872871,
-0.0196957979,
-0.0144013688,
0.0043762652,
-0.0587577671,
-0.0247026607,
0.0081565725,
2.625e-7,
0.0284043122,
-0.008080923,
-0.0125585636,
0.0341554433,
0.0433612466,
-0.0050176554,
-0.0128648123,
0.0273876786,
0.0236935746,
0.040357478,
0.021731656,
0.0019329954,
0.020073317,
-0.022147676,
-0.0854343474,
-0.0134062525,
-0.0344724879,
-0.0564430617,
-0.0590838157,
-0.0087773046,
0.0896078274,
0.0854157731,
0.081919089,
-0.0116756428,
0.0043909671,
-0.0053024362,
0.0568489619,
0.0129763493,
0.0086576864,
0.016418878,
0.0190068055,
0.0399967656,
-0.0390840285,
0.0564158149,
-0.0276890434,
0.0648562759,
0.0521945655,
0.0449667834,
-0.0443317257,
0.0278610475,
0.0146511476,
-0.0003125662,
-0.0246448219,
-0.022259593,
0.0547834784,
-0.0529994927,
0.0093621602,
0.0100686038,
-0.0731513947,
0.0248426609,
0.0084051378,
0.0274111498,
-0.0204353109,
-0.0350801274,
0.0259000715,
0.0052766497,
-0.0764811188,
-0.0461427495,
0.0214849487,
0.0234121662,
-0.0287711788,
-0.0486080572,
0.0054826913,
-0.0105820531,
0.0348014981,
-0.032027889,
-0.0103514576,
2.275385589e-34,
0.034935914,
0.0222148206,
0.0032647101,
-0.0189212039,
0.0381658487,
-0.0014959029,
0.0150907077,
0.0062382054,
0.0172944963,
-0.0467314571,
-0.0332669467
] |
http://arxiv.org/abs/2401.16348v2 | http://arxiv.org/pdf/2401.16348v2 | Improving the TENOR of Labeling: Re-evaluating Topic Models for Content Analysis | 2024-01-29T17:54:04 | 2024-02-20T03:10:58 | Zongxia Li, Andrew Mao, Daniel Stephens, Pranav Goel, Emily Walpole, Alden Dima, Juan Fung, Jordan Boyd-Graber | cs.CL | cs.CL, cs.CY, cs.HC | Topic models are a popular tool for understanding text collections, but their
evaluation has been a point of contention. Automated evaluation metrics such as
coherence are often used, however, their validity has been questioned for
neural topic models (NTMs) and can overlook a models benefits in real world
applications. To this end, we conduct the first evaluation of neural,
supervised and classical topic models in an interactive task based setting. We
combine topic models with a classifier and test their ability to help humans
conduct content analysis and document annotation. From simulated, real user and
expert pilot studies, the Contextual Neural Topic Model does the best on
cluster evaluation metrics and human evaluations; however, LDA is competitive
with two other NTMs under our simulated experiment and user study results,
contrary to what coherence scores suggest. We show that current automated
metrics do not provide a complete picture of topic modeling capabilities, but
the right choice of NTMs can be better than classical models on practical task. | Improving the TENOR of Labeling: Re-evaluating Topic Models for Content Analysis: Topic models are a popular tool for understanding text collections, but their
evaluation has been a point of contention. Automated evaluation metrics such as
coherence are often used, however, their validity has been questioned for
neural topic models (NTMs) and can overlook a models benefits in real world
applications. To this end, we conduct the first evaluation of neural,
supervised and classical topic models in an interactive task based setting. We
combine topic models with a classifier and test their ability to help humans
conduct content analysis and document annotation. From simulated, real user and
expert pilot studies, the Contextual Neural Topic Model does the best on
cluster evaluation metrics and human evaluations; however, LDA is competitive
with two other NTMs under our simulated experiment and user study results,
contrary to what coherence scores suggest. We show that current automated
metrics do not provide a complete picture of topic modeling capabilities, but
the right choice of NTMs can be better than classical models on practical task. | [
0.022659367,
-0.0407800265,
-0.0114573576,
-0.0038280883,
-0.0416426472,
0.0138582652,
0.0333439857,
-0.0436283983,
-0.0182615947,
-0.0608384795,
-0.0276848767,
0.0129663609,
-0.0227737296,
0.0329542235,
-0.0090330811,
-0.0552189946,
0.0309630893,
-0.0111539662,
0.0299875923,
0.0543961078,
-0.0071716527,
-0.0129910326,
-0.0283612181,
0.0322918631,
-0.0330271497,
-0.0534782335,
-0.0102472259,
-0.0188855212,
-0.0007152724,
-0.0489775054,
0.0050723553,
0.053235434,
0.0194718912,
-0.0122638894,
0.0000021647,
-0.0221374929,
-0.0063976985,
-0.0064970599,
0.0034883693,
0.0200473648,
0.0785887539,
0.0069485609,
0.0222898386,
-0.0159755815,
-0.0131271491,
0.0409293287,
-0.0242807474,
0.0785938874,
0.0724752843,
0.0316944122,
-0.0224419609,
-0.0159184616,
0.0504713431,
-0.0216853116,
0.0249354001,
-0.0499576144,
0.0098892832,
0.0501219481,
0.0135811446,
0.0644697025,
0.0239884965,
0.0447596796,
0.0082661705,
-0.0475219563,
0.0525370836,
0.0328512043,
0.0091068158,
-0.0435690172,
0.0048426203,
0.0585213453,
0.0336078145,
0.0361779183,
0.0435067937,
-0.0110556837,
-0.0225849617,
0.0162106231,
-0.0255349614,
-0.001774905,
-0.0201472007,
-0.0039048803,
0.054524228,
0.0383637398,
0.0258422624,
-0.0016135981,
0.0579178631,
0.0558455326,
-0.0108266464,
-0.0243683308,
-0.0094397068,
0.0310877971,
-0.0091295512,
-0.047006458,
0.0409855992,
0.02190735,
0.0077746534,
0.0230588689,
-0.0071795285,
-0.0626368374,
0.0468169414,
-0.0687089041,
-0.0325746313,
-0.0034869406,
-0.0503003784,
0.0758287385,
0.037354324,
0.0146139748,
0.0274324026,
-0.0128897969,
-0.0211334229,
0.0078498572,
-0.0541834533,
0.0160962697,
-0.0047576907,
0.0563130192,
-0.0035908236,
-0.0493307151,
-0.0028111809,
0.0378254466,
-0.00497786,
0.0103950938,
-0.0254023001,
0.0112458076,
-0.0245810952,
0.0208831746,
-0.02871079,
-0.0144594312,
-0.0516274907,
0.0418842696,
-0.0103422841,
0.0240500979,
0.0346030816,
0.0467359014,
-0.0133287022,
-2.25e-7,
-0.0182004962,
0.0121210814,
0.0072914022,
-0.0315177515,
-0.0617693812,
-0.0201589614,
-0.0267768111,
-0.0083339149,
-0.0036216455,
-0.0642428026,
-0.0159662589,
0.0101162158,
-0.006620849,
-0.0292914882,
-0.0198426694,
0.02157115,
-0.0126359807,
0.0356265306,
-0.0327282287,
0.0037595029,
0.1331511587,
0.0488663651,
-0.0164604262,
0.0303822216,
0.0191851556,
0.0502277985,
0.0426261537,
0.0282881204,
0.0202796739,
-0.0204721577,
-0.0254198927,
-0.0059875962,
0.0071214633,
0.0115680704,
-0.0124430042,
0.019920269,
0.0071602841,
0.0491901077,
-0.0434475206,
-0.0063874889,
0.0291536804,
0.0875678658,
0.0335026793,
-0.0176906791,
0.0002479049,
0.0268409904,
0.0227492861,
0.0010226975,
0.0214217883,
0.0293876603,
-0.0153223025,
0.0134910885,
0.0597700663,
-0.0262819398,
-0.0470711999,
-0.0855599418,
-0.029198559,
-0.0000156483,
0.0179155413,
-0.0463200472,
0.0634841919,
-0.0226163212,
0.0306000169,
0.0096054217,
0.0392013155,
-0.0161685757,
0.0265643056,
-0.0111673595,
0.075348787,
0.0488071591,
0.0371684209,
-0.0705158785,
-0.0475823246,
-0.0151306577,
0.0339698754,
0.0264721457,
-0.0006631153,
-0.0323577784,
-0.0156062497,
-0.0321356542,
-0.0002717893,
-0.051798705,
0.0273262281,
0.0043990719,
-0.0415125154,
-0.0099797724,
0.0190766957,
-0.0639196932,
-0.0295972619,
0.0609197207,
-0.0124497348,
0.0069924598,
0.004858518,
-0.0073882104,
0.0247733295,
0.0553122349,
0.013880726,
0.0115182977,
0.045252081,
-0.0297184587,
-0.0002172599,
-0.0486998186,
-0.0238508247,
-0.0178488698,
0.0222090147,
-0.0471677259,
-0.0194156822,
-0.0343663618,
-0.0200167838,
-0.0392688103,
0.0285081305,
-0.0319268815,
-0.0536497869,
-0.0450505279,
-0.0200794321,
-0.0290596895,
-0.0448959917,
-0.0476630405,
0.0874077231,
0.0065430924,
0.0310432948,
0.0299892202,
0.0088486737,
0.0191469695,
-0.0552589819,
-0.0211461596,
-0.0222700816,
0.008866014,
0.0146242529,
0.0155927474,
0.0095439954,
-0.0234246664,
-0.0320023857,
-0.0946496949,
0.0039714454,
-0.0480796956,
0.005595854,
0.0380366035,
-0.0383656695,
0.0105903186,
-0.0169767067,
-0.0284515228,
-0.0324194953,
-0.0371070616,
-0.0060365675,
0.0547500961,
0.0118279057,
0.0363814048,
-0.0774518102,
-0.0099612866,
-0.0256484672,
0.0580688603,
0.0674338788,
-0.04587337,
-0.038610816,
0.0294087827,
-0.0107459594,
-0.0305678789,
-0.0151955523,
0.0416486375,
0.0439101495,
0.005817988,
0.0522898771,
0.0326920897,
0.0186057463,
0.0629539192,
0.0579643622,
0.0016650994,
-0.0306377374,
0.0012675668,
-0.0064450428,
0.0725983307,
0.013159113,
-0.0556092374,
0.0736393407,
0.0668850616,
-0.0020545993,
0.0338486843,
-0.0404388197,
-0.0142186135,
-0.0709991083,
0.0040379171,
-0.0905990601,
-0.0291225798,
0.009505949,
0.0176148955,
-0.0428053029,
-0.0140908938,
0.0638710186,
-0.0250003189,
0.0311325956,
-0.0715385899,
-0.0336222388,
-0.0539446995,
0.0324995108,
-0.0048250924,
-0.0084768273,
-0.0141493967,
-0.0158227552,
0.020068299,
0.0361982025,
-0.0068908487,
0.0000207773,
0.041809015,
-0.0510231145,
0.0152542852,
0.008970052,
-0.0032590479,
-0.0119011616,
-0.0428709984,
0.0103831803,
0.0278976373,
0.0212243274,
0.0383763425,
-0.0088712601,
0.0366804972,
0.0601571389,
0.0243276842,
-0.0135429492,
-0.0396709368,
-0.0019389404,
0.0494040251,
0.0462204926,
-0.0286034215,
-0.0252882671,
-0.0119633274,
0.0121910973,
0.0323107354,
0.0058751362,
0.0163396187,
-0.0630295798,
0.0113733774,
-0.033876352,
-0.0426215008,
-0.0392694809,
0.003760108,
0.0092374021,
-0.0114447996,
-0.00072767,
0.1295323372,
0.0384936109,
-0.0776124001,
-0.0322121494,
-0.042419441,
-0.0257603545,
0.0047559892,
-0.0032199367,
-0.0302205142,
-0.0246975757,
-0.0201132447,
0.0502387621,
0.0321439579,
-0.0239903778,
0.027435597,
-0.0245349128,
0.0110217789,
-0.0379801206,
-0.032749474,
0.0033048247,
-0.0892129168,
0.0305043235,
-0.0327359587,
-0.0130520239,
-0.0159270819,
0.0167796593,
0.0016193824,
0.0483723171,
-0.0227345359,
-0.0440594181,
-0.0122102089,
-0.0343554877,
0.0406769961,
0.0331059061,
0.028141167,
0.0075698253,
-0.0031974264,
0.0203877222,
-0.0031247493,
0.0304530989,
0.0217031632,
0.0556912273,
-0.0311680697,
0.0271551982,
-0.014458687,
0.0138937039,
-0.0292380396,
-0.0441703983,
0.0185351297,
0.0592357628,
-0.0395988636,
0.0216570739,
-0.0262354091,
0.0158284418,
0.0092361402,
0.0704220682,
0.0517819002,
-0.0382918678,
0.0172180794,
-0.0286964066,
0.0705908462,
0.03665426,
0.0143184336,
-0.016106477,
-0.0156714767,
-0.0252489392,
-0.0762655511,
-0.0607121959,
0.0153818699,
-0.0187150519,
0.0077354764,
0.0060307155,
0.0387599431,
-0.04148129,
-0.0117909219,
0.0549289584,
0.0084073348,
-0.01555329,
-0.077618964,
-0.016531365,
0.0438644178,
-0.0221357998,
-0.0244934317,
-0.0006352956,
-0.0442167595,
-0.0281954184,
0.0227480344,
-0.0191622302,
0.0076656234,
0.0007249083,
-0.0586973354,
-0.0815092698,
-0.0073881079,
0.0336888097,
-0.0842211396,
0.0148368804,
-0.0179089811,
0.0166634414,
-0.0111806877,
-0.0125398338,
-0.0611587018,
0.0650833845,
-0.0361854248,
0.0044559091,
-0.0892137811,
-0.0671220198,
-0.0323447846,
-0.0095137171,
0.081789121,
0.0312043633,
0.0044957921,
-0.0575403459,
-0.0456291176,
0.0003902577,
-0.0414550491,
0.0092568472,
-0.0062716841,
0.0041620899,
-0.0188000109,
-0.0401918888,
0.0638671592,
-0.0255837943,
0.0524345785,
-0.0320108607,
0.0129263178,
0.0297402143,
0.0228851978,
-0.0799650177,
0.0098231081,
-0.0193427466,
-0.0116989557,
-0.0288010556,
0.02065631,
-0.006291688,
0.0562450476,
0.0021666454,
-0.0061746808,
-0.0023491085,
0.0401816033,
-0.061825864,
0.0210951045,
-0.0102630518,
0.0301363133,
0.0148296943,
0.0057132812,
0.0272122677,
-0.0314094499,
-0.0118111921,
-0.0285248198,
0.0289322231,
0.0482792221,
-0.0094038201,
-0.0014607963,
0.0496604256,
0.0020656933,
0.0114762001,
0.0658284202,
-0.0382932238,
-0.0277733468,
0.0245690588,
-0.0476265363,
-0.0392307118,
0.0419720598,
0.0135169551,
0.0593776032,
-0.0271389335,
-0.0142878005,
-0.0433501676,
0.028872814,
0.0744739398,
-0.0149950776,
0.015519036,
0.0030958413,
0.053478092,
0.0712843165,
-0.0064209891,
0.0252754688,
-0.061702583,
0.0098156082,
-0.0304994602,
-0.0106202504,
-0.0595475994,
-6.823335488e-33,
-0.0465870909,
-0.0390045904,
0.0141289392,
0.0281390864,
0.0268863514,
0.0331688449,
-0.068209298,
0.0151306419,
0.0255528782,
0.021678146,
0.0212403797,
0.0391826555,
-0.0049314667,
-0.0060017765,
0.0117675941,
-0.0133922147,
0.0443195291,
0.0018349456,
-0.012973466,
0.026982585,
0.0342924707,
0.000121566,
-0.0062475479,
-0.1170647144,
0.0678506866,
0.0076272557,
0.019269146,
-0.0121763954,
0.0659029782,
0.0359268822,
-0.0467710346,
0.0246997271,
0.0598769486,
-0.0606006086,
0.0203615408,
0.0291532688,
-0.0610912852,
-0.0808037445,
-0.0191170257,
-0.0510759056,
-0.0293389689,
-0.0469327234,
0.041656293,
0.0384232551,
-0.0128012514,
0.032779824,
0.1074144989,
0.019883072,
-0.0387056768,
-0.0357696787,
-0.0534161329,
0.0587845482,
-0.0303646289,
-0.0013634462,
0.0036066622,
-0.0135325594,
-0.0216308925,
0.0548234656,
-0.0810249075,
0.0221529678,
-0.004053886,
0.0027008792,
0.0230839364,
-0.0142569756,
-0.0201492198,
0.0594445281,
-0.0130087333,
0.0468738899,
-0.0435181074,
0.0057411063,
0.0168628469,
0.0421230644,
0.0592761338,
0.0028661836,
0.0000577625,
-0.0004692395,
-0.0382658727,
0.094857499,
-0.0306633972,
0.0051035592,
0.0337050222,
0.0106930714,
-0.0372721627,
-0.0198568441,
-0.043255467,
-0.0088607948,
-0.0423682518,
-0.0088820094,
-0.0109872818,
-0.0187356602,
-0.0375373475,
-0.0108308354,
-0.0251910649,
0.0292206611,
0.0243543535,
0.0030334445,
-0.0237058848,
0.053655006,
0.0169908777,
-0.0058951951,
-0.0859618783,
-0.0103375036,
0.0154328719,
0.0248477552,
0.0262640193,
-0.0226946846,
0.0168580785,
-0.0313109048,
0.0064563057,
-0.0127709983,
0.0178745557,
-0.0278764348,
-0.0099451244,
0.0082083661,
0.0043370053,
0.0000105155,
-0.0184834469,
0.0756080076,
0.0221740287,
0.0390653461,
0.0106490245,
-0.0736976191,
-0.0029189298,
-0.0164713152,
-0.0184923392,
0.0076870895,
0.0011704937,
0.0330024175,
0.0372147933,
-0.0358818956,
-0.0011719295,
0.0038194056,
2.82e-7,
0.0688415989,
0.0622510351,
0.0016783866,
-0.0282634869,
0.0251478143,
0.0438374095,
-0.0206689946,
-0.0043215528,
-0.0050774622,
-0.0372290201,
0.0091371881,
0.0289291367,
0.023249805,
0.0612422451,
-0.0390627347,
-0.0089379707,
-0.0467000231,
0.0057875696,
-0.0484473743,
0.022959372,
0.0089515084,
0.0177441705,
0.0086856075,
0.0090699419,
-0.0138432942,
-0.0387763456,
0.0406374447,
-0.0325799026,
-0.0184362195,
0.0219081882,
0.0814385638,
0.0438146703,
-0.0000111986,
-0.0118539603,
-0.0138547067,
-0.0163733363,
0.0581206717,
0.0463409945,
0.0021653452,
-0.0321349874,
0.0020285528,
0.0017502632,
0.0107642459,
-0.0263642687,
0.0515919775,
-0.0304787774,
-0.0179636013,
-0.0229371227,
-0.0169221256,
0.0263452325,
0.0241661835,
-0.0018610463,
-0.0054407334,
-0.0001485889,
-0.0327579863,
-0.0352093279,
-0.0178172905,
-0.0443064198,
-0.0282766949,
0.0168755949,
-0.0290269032,
0.0002938866,
-0.0347553566,
0.0057968595,
-0.0029002216,
-0.0470100604,
0.0419999287,
2.040142766e-34,
0.020365335,
-0.0097330511,
-0.0400875285,
-0.0307472255,
-0.0203541294,
0.0271313339,
-0.1052232906,
-0.0194174238,
-0.0210732222,
-0.063439779,
0.0335828997
] |
http://arxiv.org/abs/2401.16335v1 | http://arxiv.org/pdf/2401.16335v1 | Iterative Data Smoothing: Mitigating Reward Overfitting and Overoptimization in RLHF | 2024-01-29T17:43:42 | 2024-01-29T17:43:42 | Banghua Zhu, Michael I. Jordan, Jiantao Jiao | cs.LG | cs.LG, cs.AI, cs.CL, stat.ML | Reinforcement Learning from Human Feedback (RLHF) is a pivotal technique that
aligns language models closely with human-centric values. The initial phase of
RLHF involves learning human values using a reward model from ranking data. It
is observed that the performance of the reward model degrades after one epoch
of training, and optimizing too much against the learned reward model
eventually hinders the true objective. This paper delves into these issues,
leveraging the theoretical insights to design improved reward learning
algorithm termed 'Iterative Data Smoothing' (IDS). The core idea is that during
each training epoch, we not only update the model with the data, but also
update the date using the model, replacing hard labels with soft labels. Our
empirical findings highlight the superior performance of this approach over the
traditional methods. | Iterative Data Smoothing: Mitigating Reward Overfitting and Overoptimization in RLHF: Reinforcement Learning from Human Feedback (RLHF) is a pivotal technique that
aligns language models closely with human-centric values. The initial phase of
RLHF involves learning human values using a reward model from ranking data. It
is observed that the performance of the reward model degrades after one epoch
of training, and optimizing too much against the learned reward model
eventually hinders the true objective. This paper delves into these issues,
leveraging the theoretical insights to design improved reward learning
algorithm termed 'Iterative Data Smoothing' (IDS). The core idea is that during
each training epoch, we not only update the model with the data, but also
update the date using the model, replacing hard labels with soft labels. Our
empirical findings highlight the superior performance of this approach over the
traditional methods. | [
0.0488424487,
0.0557742268,
0.0119201588,
-0.0093236789,
0.0114213787,
0.0197408404,
-0.0812110528,
0.0125914663,
-0.0108606135,
-0.0017322662,
-0.0171808079,
-0.009900107,
0.0178193264,
-0.037481308,
-0.0361372195,
-0.1006315127,
0.0509787798,
-0.0407745466,
0.0578285158,
-0.0357925557,
-0.0150757264,
-0.0253969245,
0.0172961783,
0.0102214199,
0.0042465832,
0.0146529265,
0.0118229082,
-0.0163622685,
0.0508792251,
-0.0558636896,
0.0269303434,
0.0323951878,
0.0307739582,
0.0543124154,
0.0000022629,
-0.024291737,
-0.0008757181,
-0.0275734942,
-0.0195736159,
-0.0212283283,
0.0506798141,
0.0436992906,
-0.0360786766,
0.0114126056,
-0.0446993075,
0.0070461291,
0.0088529745,
0.0245534908,
0.0519045666,
0.0482133999,
0.0031125788,
-0.0197657645,
-0.0367290117,
-0.0446701124,
0.0273790415,
-0.0645306557,
0.0158478636,
0.0398110226,
0.0395779908,
-0.0015367726,
-0.0353567526,
0.0281418301,
-0.0007604745,
0.0062964014,
0.0525149517,
0.0517647676,
0.05132173,
-0.0020830156,
-0.0343016088,
0.0507185943,
-0.0444961078,
-0.0145232268,
0.0005002024,
0.0017383263,
0.0061113657,
-0.0358084068,
-0.004602002,
0.0105161099,
0.0112447804,
0.0381226726,
0.0450703762,
-0.0042057014,
0.0228958521,
0.0463112816,
-0.0059462315,
0.0799094588,
-0.01664708,
-0.0669597611,
0.0269955099,
-0.0099289371,
0.0103971213,
-0.0278029349,
-0.0024716188,
0.0455395095,
-0.0706926435,
0.0193722434,
-0.0127809988,
0.0101082511,
-0.0089755161,
-0.0264937114,
0.0186532158,
0.0173381194,
0.0089150472,
0.0318069234,
-0.0245524589,
0.0154573144,
0.0357388183,
0.0423878469,
-0.0679380894,
0.0605833195,
-0.0154553857,
-0.000940846,
-0.007358979,
0.0461385436,
0.0050437832,
0.007458936,
-0.0777496323,
0.0127254706,
-0.0176745597,
0.0411017835,
0.0248133037,
0.0006147537,
-0.0715509728,
0.0445241407,
-0.0677574426,
0.0106433649,
-0.0624325052,
-0.0136755137,
0.0085442076,
0.025849089,
0.0129250623,
-0.0314609483,
0.0125480471,
-0.0055890074,
0.0023085645,
0.0427935086,
-0.0496052727,
0.0155082205,
-0.035398338,
-0.0293142665,
0.029737018,
-0.0282814037,
-0.000777978,
-0.0339545123,
-0.0062476886,
-0.0346640982,
-0.0079939123,
-0.009617405,
-0.0006100389,
0.0090096714,
-0.0579653084,
0.0666844249,
-0.0113769975,
0.0195448026,
0.0404962115,
-0.0227658637,
0.0282491054,
0.034980651,
0.0417313017,
0.018865509,
0.0056773187,
0.0103328768,
-0.0130124781,
0.0200791899,
0.0444043875,
-0.0004614917,
-0.0273701474,
-0.0186201967,
0.011033996,
0.0372446887,
0.0541256815,
0.026624212,
-0.0354740359,
0.0268784277,
0.077170454,
-0.0035966407,
-0.0076277666,
-0.0077545322,
-0.01779457,
0.0326696523,
-0.0084233163,
-0.0009274246,
-0.0428592339,
0.0361684375,
0.0104760593,
-0.0096817473,
-0.0288627949,
-0.0401457623,
-0.0298469514,
-0.0480718873,
0.0151800467,
-0.0097742407,
0.0839485601,
0.0336539224,
0.040686883,
0.0048042545,
-0.0427901931,
-0.0393253975,
0.0525397696,
0.0064899251,
0.015685115,
-0.027261911,
0.0537414104,
0.0818152353,
0.0200527515,
-0.0387228094,
-0.0057408828,
-0.0297779199,
0.0035980802,
0.000992398,
0.0439989083,
-0.0603081249,
-0.0302177742,
-0.0272231586,
-0.0547512136,
0.0653565601,
-0.0128170801,
0.0431688689,
0.0023183518,
-0.0127226515,
-0.0115280515,
0.0188657157,
-0.096305795,
0.0260371715,
0.0192243531,
-0.026968468,
0.0353305638,
0.011460566,
0.028233448,
-0.0134723606,
0.029360801,
-0.0120217372,
0.0314670615,
0.0013525666,
-0.0044631823,
0.0282169208,
-0.0598066933,
-0.0283189882,
0.0433034524,
0.0145695545,
0.0132294903,
0.0671913847,
-0.0362822935,
-0.0267682914,
0.0191735476,
0.0095120221,
0.0101634422,
-0.0112248808,
0.0373565368,
-0.0051468569,
-0.00737812,
-0.0937513709,
0.052662354,
-0.0167870764,
0.0285711847,
-0.0286767706,
-0.0141534861,
-0.0143902078,
-0.078563869,
-0.0120247444,
0.0649406835,
-0.0018685758,
0.0351307243,
0.0372502096,
0.0188891031,
-0.0523794219,
-0.0031181993,
-0.0190145168,
-0.055529274,
-0.0074377125,
-0.0036329515,
0.0031376986,
-0.0159491766,
0.0041788379,
-0.0519559681,
-0.0308184046,
-0.0178458784,
0.0175055843,
0.0147875361,
0.0377296843,
0.0516541302,
0.0038394977,
-0.0264980458,
-0.0373738445,
-0.0040326817,
-0.0219419971,
0.0068778298,
-0.0152919833,
-0.068132326,
0.0003691289,
-0.0690564737,
0.0064218175,
-0.0159374122,
0.0099446224,
0.0102207279,
-0.0103216171,
0.0502479151,
0.1031051129,
0.04164562,
0.0182063915,
0.0117561761,
-0.001147281,
-0.0752881318,
0.0253853463,
-0.0529250577,
-0.0593457632,
-0.008046045,
-0.0013291585,
-0.0095477309,
0.0404691845,
-0.0227070153,
0.0237478111,
-0.0657223016,
0.021042604,
-0.0706236362,
-0.0339750014,
-0.0752560198,
0.0246281531,
-0.0284656491,
-0.0191809498,
-0.0264242273,
-0.0428143404,
0.0733074769,
0.0707494542,
0.0070732995,
-0.0468269363,
0.0039607063,
0.0018415512,
-0.0224539917,
-0.0106230145,
0.039065782,
-0.0570408255,
-0.0520006977,
-0.0511133038,
-0.0642006621,
0.0383634865,
-0.0539792664,
0.0207183287,
-0.0019937337,
-0.0128009953,
-0.0137929739,
-0.0596199855,
-0.042914059,
-0.0475650094,
0.0009792086,
-0.0154922428,
-0.0120819341,
0.0103602801,
0.0110685276,
-0.0337594077,
0.0136704342,
0.0151881194,
-0.0145282131,
-0.026888309,
-0.0016320627,
-0.0271687265,
0.0500061251,
0.0044720061,
0.0023570301,
-0.0400424935,
0.0225915425,
-0.0062726582,
0.1120203361,
0.0023306375,
0.0149375731,
0.0558452159,
0.0482066311,
-0.0602206998,
0.0096480809,
0.0674238503,
-0.0137666985,
-0.011235863,
0.0090632737,
0.0770089924,
0.0112680085,
-0.0774407312,
0.0369663462,
-0.0256483927,
0.0911783725,
0.0134931607,
0.0313478895,
-0.0483164564,
0.004520792,
0.0065439348,
0.0637066737,
-0.0152581818,
-0.039004039,
0.0382281169,
-0.075423874,
-0.0055484087,
0.0183201954,
-0.0253758822,
0.012356041,
-0.0179078504,
-0.0139007857,
-0.0092298184,
-0.0013906321,
0.0511021502,
-0.0175791942,
-0.004298625,
0.05229038,
0.0369602926,
-0.0662629828,
0.0254543219,
-0.0252673328,
0.0468811318,
0.1081165969,
0.0449350961,
-0.0362902768,
-0.0154542318,
-0.0339465141,
0.0634229705,
-0.0181214716,
-0.0353722945,
0.008952192,
-0.0205711722,
-0.0130678806,
-0.0148415379,
0.0391400978,
0.0354634374,
0.0219301395,
-0.0158174057,
0.0530534163,
-0.0037661665,
-0.0040805531,
0.0199310649,
-0.0251704026,
0.0416317247,
0.0459777601,
-0.0437328182,
-0.0574919358,
0.0026040215,
-0.0111968126,
0.0172311272,
-0.0292050485,
-0.0206807274,
0.0294449758,
-0.0349078849,
-0.0060921344,
-0.0725744665,
-0.0554903857,
-0.0331438221,
0.0193090048,
0.0140808634,
0.0032305843,
0.0078696413,
0.0256897397,
-0.0375675745,
0.0004643319,
-0.0426098071,
0.0318131112,
-0.0795121789,
-0.0302444566,
-0.032280419,
-0.0523821302,
-0.0218891073,
-0.0079000378,
-0.0139236664,
-0.0233075544,
0.0291860681,
0.0512130074,
0.0142226517,
0.0876083672,
0.001919257,
-0.0444419123,
0.090011023,
0.0351295695,
0.0042468212,
0.0205328632,
0.0358602144,
0.0176881235,
-0.0173019134,
0.0176915079,
-0.067784436,
0.0388120152,
0.0306022111,
0.0657976642,
0.0174115747,
0.002141733,
0.0434244312,
0.035643395,
0.0178243387,
-0.0174898244,
-0.0272502098,
-0.0215045717,
0.0039400891,
-0.007036177,
-0.0402365215,
-0.0171171147,
0.0276241675,
0.036391601,
0.0253777094,
-0.027367454,
-0.0289358161,
0.0389883183,
-0.0122572975,
-0.0529185496,
-0.019696584,
-0.0281000901,
0.021425141,
0.0102643603,
0.0379109792,
0.0077861561,
0.0058590979,
0.021108808,
0.0308160968,
0.0205680095,
0.0532586612,
0.0502371788,
0.0303020645,
-0.0300701894,
0.0168199502,
-0.0567699857,
-0.0104802633,
-0.0599555075,
-0.0704354942,
0.0717287436,
0.0132342279,
0.0428869948,
-0.0279728584,
0.0091828713,
0.0078696767,
-0.0205882341,
-0.022223074,
-0.0200275704,
0.0101694763,
0.0208003651,
-0.007855488,
0.0119868172,
0.0525702015,
-0.0009463135,
0.0176645666,
-0.0393187031,
-0.0062697888,
-0.0239828769,
-0.0092293937,
0.0221469849,
0.006291437,
-0.0290361959,
-0.0251700617,
-0.0191012342,
-0.0408755541,
-0.0251696892,
0.0322824791,
0.0203068834,
-0.0391810238,
0.0384659395,
0.0243074801,
-0.0485643558,
-0.0080175968,
-0.0923302025,
-0.0315941013,
-0.0118651986,
-0.0055487482,
-0.0584850311,
-6.614503773e-33,
-0.0318443403,
0.0234145448,
-0.004806424,
0.0258135479,
-0.056568142,
0.0027460693,
0.0008212802,
0.0013969503,
0.0020778177,
0.0252093188,
-0.0476774685,
-0.0243952479,
0.0132683972,
-0.0180277601,
0.0417769141,
-0.0244276095,
0.0382410884,
-0.0268704426,
-0.0712529272,
0.0133564118,
0.0156979542,
0.0315177292,
0.0769809708,
-0.0713506117,
0.0354032777,
0.0557484217,
-0.0293040685,
0.0467926338,
0.0203171354,
0.0112256724,
0.0094602471,
0.0349038988,
0.0064271903,
-0.0650655851,
-0.0326924324,
0.0048284382,
-0.1349921674,
-0.0208145641,
-0.0058292141,
0.0102719422,
-0.0059386776,
0.0173614584,
0.0286609624,
-0.0284555815,
-0.0423319265,
0.0354022607,
0.0226105172,
-0.0109637929,
-0.0515030958,
-0.00944605,
-0.0199241359,
0.0175863728,
0.0035813849,
-0.0087095415,
0.0540549792,
0.0476563834,
-0.0200046692,
0.0371080451,
-0.0921918526,
0.0563537367,
-0.0100681754,
0.0439633243,
0.0268926676,
0.0787104741,
0.022290105,
0.0141135147,
-0.0166578405,
0.0398955345,
0.0222044829,
0.0081407363,
0.0537821762,
-0.0085696792,
0.0143895149,
-0.0234776158,
0.0479667038,
-0.0470914766,
-0.0065377145,
0.0359008461,
0.0091505535,
0.0349077247,
0.0007048197,
0.0379502289,
-0.0653468892,
-0.0373468734,
-0.0336774401,
0.0158878937,
-0.0148005337,
-0.0350142308,
0.0211964957,
0.0533418357,
-0.0248243045,
0.0288333017,
0.0060486947,
-0.0233979411,
-0.0397548303,
0.0082655568,
-0.0247270837,
0.0067665763,
0.00318238,
-0.0328216739,
-0.1150464416,
-0.0413749777,
-0.02296675,
-0.0032080053,
-0.0127976965,
0.022712823,
0.0185815431,
0.0463845432,
-0.0227638967,
0.0116948513,
0.0133140841,
0.001815505,
0.0054629203,
-0.0151133426,
0.0491667911,
-0.0389319211,
0.0136964601,
0.0406793915,
0.0302848574,
0.0323060229,
-0.0281426758,
-0.020475449,
0.009261556,
0.0108893951,
-0.0015466086,
-0.0109680686,
-0.0335405767,
0.013880454,
0.0167215317,
-0.0442125238,
-0.0042563933,
-0.0087695783,
2.997e-7,
-0.0482402518,
0.0517952815,
-0.0152976895,
0.0214969888,
0.0032956931,
0.0327406414,
0.0045378152,
0.0192987919,
0.0552486964,
0.0330985487,
0.0597474203,
0.0023754148,
0.0233819708,
-0.0039634882,
-0.0627103597,
-0.1052100658,
-0.094767794,
0.0043726512,
-0.0008212832,
-0.0253138635,
-0.0015439084,
0.0527949482,
0.0670734048,
0.0017288608,
0.0186083093,
-0.0053528859,
0.0303138029,
-0.0316988938,
-0.0114033381,
0.0258410051,
0.0385697521,
0.0166105088,
-0.016644787,
0.0098702274,
-0.0250509605,
0.0169366449,
0.0148919001,
0.0258104764,
-0.0538238846,
0.0074698799,
-0.0191955194,
0.0293661766,
0.0209903326,
-0.0454463847,
0.0206081588,
-0.0736849755,
-0.0251075067,
0.0042047929,
-0.0109654814,
-0.003566338,
0.0121598011,
-0.0249894913,
-0.0459718779,
-0.0020523465,
0.0200414658,
0.0655108392,
-0.0036861689,
-0.0793850869,
-0.0219852887,
0.0376849286,
-0.0198118575,
-0.0528767742,
-0.0342789441,
0.0366560146,
-0.0450675711,
-0.082593739,
-0.0031084497,
2.587500691e-34,
0.0240533426,
-0.0209117867,
0.0122318882,
-0.0117944488,
0.0287451595,
0.0429333635,
0.0323187821,
0.0237972718,
0.0657841265,
-0.0214343127,
-0.0268390849
] |
http://arxiv.org/abs/2401.16332v3 | http://arxiv.org/pdf/2401.16332v3 | Tradeoffs Between Alignment and Helpfulness in Language Models with Representation Engineering | 2024-01-29T17:38:14 | 2024-05-26T16:07:55 | Yotam Wolf, Noam Wies, Dorin Shteyman, Binyamin Rothberg, Yoav Levine, Amnon Shashua | cs.CL | cs.CL, cs.AI | Language model alignment has become an important component of AI safety,
allowing safe interactions between humans and language models, by enhancing
desired behaviors and inhibiting undesired ones. It is often done by tuning the
model or inserting preset aligning prompts. Recently, representation
engineering, a method which alters the model's behavior via changing its
representations post-training, was shown to be effective in aligning LLMs (Zou
et al., 2023a). Representation engineering yields gains in alignment oriented
tasks such as resistance to adversarial attacks and reduction of social biases,
but was also shown to cause a decrease in the ability of the model to perform
basic tasks. In this paper we study the tradeoff between the increase in
alignment and decrease in helpfulness of the model. We propose a theoretical
framework which provides bounds for these two quantities, and demonstrate their
relevance empirically. First, we find that under the conditions of our
framework, alignment can be guaranteed with representation engineering, and at
the same time that helpfulness is harmed in the process. Second, we show that
helpfulness is harmed quadratically with the norm of the representation
engineering vector, while the alignment increases linearly with it, indicating
a regime in which it is efficient to use representation engineering. We
validate our findings empirically, and chart the boundaries to the usefulness
of representation engineering for alignment. | Tradeoffs Between Alignment and Helpfulness in Language Models with Representation Engineering: Language model alignment has become an important component of AI safety,
allowing safe interactions between humans and language models, by enhancing
desired behaviors and inhibiting undesired ones. It is often done by tuning the
model or inserting preset aligning prompts. Recently, representation
engineering, a method which alters the model's behavior via changing its
representations post-training, was shown to be effective in aligning LLMs (Zou
et al., 2023a). Representation engineering yields gains in alignment oriented
tasks such as resistance to adversarial attacks and reduction of social biases,
but was also shown to cause a decrease in the ability of the model to perform
basic tasks. In this paper we study the tradeoff between the increase in
alignment and decrease in helpfulness of the model. We propose a theoretical
framework which provides bounds for these two quantities, and demonstrate their
relevance empirically. First, we find that under the conditions of our
framework, alignment can be guaranteed with representation engineering, and at
the same time that helpfulness is harmed in the process. Second, we show that
helpfulness is harmed quadratically with the norm of the representation
engineering vector, while the alignment increases linearly with it, indicating
a regime in which it is efficient to use representation engineering. We
validate our findings empirically, and chart the boundaries to the usefulness
of representation engineering for alignment. | [
0.0128264688,
0.0569539033,
-0.0371575393,
0.0289254207,
-0.0327418931,
0.0122278212,
-0.0126230121,
-0.0078034266,
0.0056776106,
-0.0223032981,
-0.0451260097,
-0.0419678278,
0.0106815929,
-0.0030386981,
0.0515381508,
-0.0439920276,
0.0808123499,
-0.0128686614,
0.0252092555,
-0.0123721883,
0.0300335437,
-0.0232234076,
0.008515168,
0.031271968,
-0.0072075822,
-0.0213324688,
-0.0266482234,
0.0008426004,
0.0292014293,
0.0195842348,
-0.0010395346,
0.0258261077,
0.0134108048,
0.0525030382,
0.0000019445,
-0.0573165677,
-0.0207261983,
0.0052208425,
-0.0216603316,
-0.0077276202,
0.0285206567,
-0.0772213414,
-0.0201389808,
0.0088797929,
-0.0715685487,
0.0298105739,
0.0670619607,
0.0417996161,
0.0504657403,
0.0910210684,
-0.0038710549,
-0.0188461021,
0.0392993502,
-0.0064652753,
-0.0208862871,
-0.0461577736,
0.0049604778,
-0.017548399,
-0.0360995866,
0.0042678318,
-0.0289149191,
0.032332927,
-0.0095137432,
0.0035317275,
0.0259434655,
0.0171995368,
-0.0066720443,
-0.0289093088,
0.0056244517,
0.0589484014,
-0.0262701437,
-0.0358712673,
-0.0101496307,
-0.0184790585,
0.0521645062,
0.0125703951,
0.0012551866,
-0.0139287114,
0.0273211543,
0.0297663342,
0.027080236,
-0.0218208116,
0.059444759,
0.0039053413,
0.0199146587,
0.0573013872,
0.0180334765,
-0.0361489542,
-0.0210223738,
-0.010428844,
-0.0262461845,
-0.0391570777,
-0.0310726911,
0.0144462297,
-0.0404499806,
0.0247801375,
-0.0939089879,
-0.0248747338,
0.0139354104,
-0.0506032519,
0.035777349,
0.0585502647,
0.0220660083,
0.0054758741,
-0.0098275738,
0.0578580983,
-0.0306749102,
0.0512223728,
-0.0268995445,
0.0262967404,
-0.0156889223,
-0.0352782905,
-0.0017992943,
-0.0140911303,
0.0141950948,
0.0337785035,
-0.0432769135,
0.0454401933,
-0.0520984046,
0.056871932,
-0.0405477621,
-0.0134515185,
-0.034281496,
-0.0050165267,
-0.0075615724,
-0.0022080911,
-0.0395772941,
-0.0015366802,
-0.0106885368,
-0.0534397028,
-0.028277358,
0.0162629671,
-0.027341146,
-0.0485741235,
-0.0397247709,
0.1169335395,
0.004040835,
-0.023563087,
-0.0104546845,
-0.0188404266,
-0.0142421136,
-0.0391108543,
-0.0018749526,
-0.0271414276,
0.0391724221,
-0.0101234745,
-0.0432371125,
-0.1198543459,
-0.0216422547,
-0.0333520994,
-0.0293689929,
0.0782452077,
-0.0092704222,
0.0245585553,
0.024204798,
-0.0024658372,
-0.0145517318,
0.0726386011,
0.0470939986,
-0.0234680697,
0.0331136547,
-0.0446582176,
0.0239149872,
0.0439494103,
0.0282414258,
0.0078713205,
-0.0288078934,
0.0065603186,
0.0327159278,
0.0200228281,
0.0214499105,
0.0332034044,
0.0071856151,
-0.0563894548,
0.0832011551,
0.0453958139,
-0.0278089922,
0.042872455,
-0.0186202023,
-0.0044907532,
-0.0175600667,
0.0356303789,
-0.0949141905,
0.0164977144,
-0.0785697252,
0.006873705,
-0.0243568961,
0.0341006592,
-0.033858519,
-0.0375403911,
-0.0169293247,
-0.0305738952,
0.0750126243,
0.0005001535,
0.0195362996,
-0.0085025355,
-0.023255026,
0.0483006723,
0.0172681585,
-0.0518610142,
0.0025510646,
-0.0602675751,
0.0086757736,
0.0179999303,
0.0148981949,
-0.0532175526,
-0.0470751375,
-0.0014514548,
-0.0186217465,
0.0328973942,
0.0778116435,
0.0031672393,
0.0138700716,
-0.0354779996,
-0.0137237925,
-0.012059832,
0.010859346,
-0.0084191291,
0.0278343707,
0.000333829,
-0.0166253988,
-0.0105146738,
-0.0808671266,
0.0590410046,
0.0072462782,
-0.0307484418,
-0.0465480313,
0.048356615,
0.0330445282,
0.0058065108,
-0.0203540679,
0.0157259554,
0.0343760662,
0.0004786247,
0.0058020363,
0.0362661816,
-0.0378164575,
0.0254036151,
0.0511647351,
-0.0031581495,
-0.0396963619,
0.0503686406,
-0.0379513651,
-0.0162069798,
0.0348307937,
-0.0456742086,
0.0354569964,
0.0174255166,
0.0543659292,
-0.050353989,
0.0009091625,
-0.0614775158,
0.0486782193,
-0.0000572144,
-0.011129735,
-0.0048859925,
-0.0238441993,
0.0078056133,
-0.0465716235,
-0.038973514,
0.0043928074,
0.0069789798,
0.0459816605,
0.0596639365,
-0.0295682661,
-0.0154559789,
-0.0211222637,
-0.039493572,
-0.0227862392,
-0.0138227027,
0.0193866566,
0.0430679508,
-0.0192559045,
-0.0069824108,
-0.0077722706,
-0.0587120727,
0.005444488,
-0.0014383295,
0.0077092936,
0.063079603,
-0.0170199145,
-0.0713260323,
-0.0490494557,
0.0049962373,
-0.0250040479,
0.0031086719,
-0.0050091101,
-0.0157862622,
-0.0799919963,
0.0252465717,
-0.0353863649,
0.0599412695,
-0.0169150084,
-0.0691878572,
0.0229844805,
0.0338155515,
0.0167659372,
0.1040122136,
0.0148296477,
0.0456636772,
0.0922579914,
0.0712470114,
-0.0507801473,
0.0025925636,
0.007084127,
-0.0267492421,
-0.0513741896,
-0.0345760025,
-0.0049131401,
0.0687279403,
0.0119453222,
-0.0097094513,
-0.0761950612,
0.0265707504,
0.0187220369,
-0.049118869,
0.0139089357,
-0.0579144508,
-0.0365918241,
0.0232504513,
-0.0576683693,
-0.0977508128,
0.084068127,
-0.0202345718,
0.0015018968,
0.0038523683,
-0.0260757543,
0.0095604733,
0.0472283103,
0.0220019072,
0.0385011211,
0.0165228732,
-0.0202560853,
-0.0966786966,
-0.0217785388,
0.0404284783,
-0.0286064744,
0.0308402758,
-0.0394584723,
0.0014803156,
-0.0125093162,
-0.0331097394,
-0.0497684963,
-0.0502077602,
0.0087396372,
-0.000943318,
-0.0053355386,
-0.0278231222,
-0.0189602319,
-0.0065928088,
0.0259648245,
-0.0288482569,
-0.0063402387,
0.0258046407,
-0.0162162613,
-0.0272152741,
0.0475344136,
0.0473268628,
0.0133378049,
-0.0359422266,
0.0357828885,
0.0420623682,
0.0053291991,
0.0097094448,
-0.0427260399,
-0.020113511,
0.0451624431,
-0.0337521061,
-0.0292873811,
0.0268111322,
0.0742447078,
-0.0024244755,
-0.000143726,
0.0807756782,
0.0181992725,
-0.0132107642,
0.0462679975,
-0.0128494697,
0.0540742502,
0.0167799033,
0.0215427876,
-0.033805985,
0.0080015091,
0.0204274356,
0.0201683417,
-0.0099216318,
-0.0034254934,
-0.0015057443,
-0.0972332209,
-0.0123587195,
0.0521046035,
-0.0434903614,
-0.0582505204,
0.0047226031,
-0.0000081571,
-0.0230495017,
0.0720449537,
0.0130099207,
0.0082079954,
-0.0279216785,
0.0144562973,
-0.0234759226,
-0.039746251,
-0.0294959638,
-0.0575930476,
0.0842203945,
0.0893827602,
0.0331211798,
-0.0126630403,
-0.0523091182,
0.0046011372,
0.0814523622,
0.0169239193,
0.0358935148,
0.0172486193,
-0.0133051286,
-0.0308947675,
0.0188973956,
0.0372240581,
0.0034530559,
-0.0092562716,
0.0046707904,
0.0443718731,
-0.0192326251,
0.0090309959,
-0.008850174,
-0.0121886609,
0.0198298059,
0.0770927146,
0.0300287195,
-0.0369350649,
0.0055436776,
-0.029445203,
0.0008698573,
0.0117224632,
-0.0608997867,
0.0307143629,
-0.0091211135,
-0.0368586779,
-0.0637355223,
-0.0386952385,
-0.0264605097,
0.0311116129,
0.0369290598,
-0.0326902866,
-0.0061048232,
0.0012770935,
-0.006170122,
0.0267061777,
-0.0195588265,
0.0195808411,
-0.0261803754,
0.0115553783,
-0.0147170732,
-0.0312300716,
0.0114648426,
-0.0345696993,
-0.0270005614,
-0.1167271659,
0.0228553806,
0.0452905595,
0.0290073827,
0.0711292922,
-0.0183103997,
-0.0741091147,
0.0152831553,
0.0252960753,
-0.0146057997,
0.0325030163,
0.0378443301,
-0.0267635621,
0.011012841,
0.0482302606,
-0.0397869721,
-0.0352079757,
0.0501743294,
0.0027467599,
-0.063130796,
-0.0264681857,
0.0289632082,
0.0541031286,
0.0387306213,
-0.0241544489,
0.0160343461,
0.0079881782,
-0.025802698,
-0.027259836,
-0.0331571922,
-0.0003633338,
-0.0221049115,
0.0252082236,
0.0075420821,
0.0255915448,
0.0000918336,
-0.020550916,
0.0263152663,
-0.0285497233,
0.0129232369,
-0.0239589792,
0.0267396197,
0.0168279763,
0.0612947382,
-0.016040951,
0.058264561,
-0.0220515262,
0.0140636256,
0.017355945,
0.026742382,
0.0211178474,
0.0465032235,
-0.0384490788,
0.0393643603,
-0.0659276247,
-0.0711651444,
0.0070870169,
0.0103086224,
0.0261194501,
-0.0160795283,
0.0503710583,
-0.0147251952,
-0.006930233,
-0.0055460902,
0.0253062006,
0.0091269771,
-0.0380399749,
0.0459067672,
0.0119634531,
-0.0307302959,
-0.0116349496,
0.0437753536,
-0.0236674696,
0.0303776432,
-0.0462822467,
0.000568929,
-0.0295984652,
-0.0439713337,
0.0566968359,
-0.0044766092,
-0.0537370406,
-0.0208245236,
-0.064622879,
-0.0760011524,
-0.017382497,
0.0259701647,
0.0008262861,
-0.0056648031,
0.0247516371,
0.0454674661,
-0.0203352235,
0.0221346635,
-0.0141148334,
-0.0205362737,
0.0173205491,
-0.0173569191,
-0.0237725545,
-5.868480324e-33,
-0.0163182858,
0.0008002704,
-0.0158291683,
-0.0335349068,
-0.0301206261,
-0.0223407932,
-0.0129111521,
0.0432998165,
-0.0569592379,
-0.0102039091,
-0.0037928284,
0.0095704282,
-0.0110814767,
-0.0164096579,
-0.0144578284,
-0.0260856971,
0.0087966258,
-0.0259855818,
-0.0416107252,
-0.0328432098,
0.0139383432,
-0.0035633915,
0.0292301998,
-0.0682513565,
0.0728510916,
0.0070666289,
-0.0380079113,
0.0203555264,
-0.03933293,
0.0602663904,
0.0055733104,
0.0313438848,
0.0077210823,
-0.000825265,
-0.0013497005,
0.0309229698,
-0.0045859115,
-0.0223250315,
0.0382005014,
0.0243785959,
-0.0371353701,
-0.0672474727,
0.0771365389,
-0.0364441425,
-0.0564406998,
0.0085044736,
0.0159081873,
-0.0100213746,
-0.005299042,
0.0355833285,
-0.0904486477,
-0.0051303566,
-0.0399107113,
-0.0199348256,
0.035068769,
0.009318498,
0.0385887921,
0.010299149,
-0.0454571098,
-0.0010394765,
-0.0130417217,
0.009575082,
0.0866631493,
0.0123888999,
0.0091283722,
0.0723681524,
0.0270611756,
0.005242507,
-0.0104551716,
-0.0312667266,
0.0459278338,
0.0074715577,
0.0452339984,
0.0642475039,
0.0198531989,
-0.05085418,
-0.0519954078,
0.0117845125,
0.0100885117,
0.0317528173,
-0.004559827,
0.037715327,
-0.0885093063,
-0.01705507,
-0.0202329662,
-0.037072055,
-0.0249569863,
-0.0382234938,
-0.0014843434,
-0.0201170687,
-0.0333452672,
-0.0390817337,
0.0007148768,
-0.0453704335,
0.0399105698,
0.0238385201,
-0.0073028323,
-0.0088720955,
0.0062593296,
0.0146719925,
-0.0220949613,
-0.0477137938,
-0.0336330496,
-0.0092886426,
0.0437649786,
0.0152713889,
0.0607912056,
0.0460239314,
-0.0415517911,
0.0104353968,
-0.0234537721,
0.0021313843,
0.0359986238,
-0.0001025675,
0.0718523189,
-0.0002515267,
-0.0045144344,
0.0066332947,
0.0255979467,
-0.0275181327,
-0.0425938591,
0.0446507968,
-0.0252860785,
0.0222340655,
-0.0240828265,
-0.0176558271,
0.0177156273,
0.0596325509,
0.056316074,
-0.0476111621,
-0.0261646621,
-0.0455671027,
2.74e-7,
0.0307037197,
0.0117947031,
0.0321607292,
0.0448607244,
0.0161044803,
0.0685816035,
-0.027802268,
0.0794805288,
-0.0156311337,
-0.0044423239,
0.0124468692,
0.0298497137,
0.0334446616,
0.0297124404,
-0.0314869024,
0.023420658,
-0.0745939463,
0.0222769268,
-0.0074415295,
0.031368725,
0.0644748285,
0.0810720921,
0.0162809044,
0.0110124061,
-0.0280029122,
-0.0395787582,
0.0368682668,
-0.0612498745,
0.0228511859,
0.0138791306,
0.0166230835,
0.0209542941,
-0.006870477,
0.0228512324,
0.0069328207,
0.031261865,
-0.0234339535,
0.0657934025,
-0.0456165709,
-0.0065418375,
-0.0193339381,
-0.0188974142,
-0.0104868608,
-0.0308296941,
0.0453942306,
-0.0342368186,
-0.0418413319,
-0.0318816043,
-0.0141856214,
0.0087872446,
0.0170401726,
-0.0028441078,
-0.014729145,
0.0228536334,
0.0256719757,
0.008701941,
0.0017150124,
-0.0482689328,
-0.0339617319,
0.0076175928,
0.0027706996,
-0.0553814992,
0.0091069341,
0.0509051606,
0.0294639152,
-0.0436702333,
-0.0095118694,
2.260453367e-34,
-0.0126057984,
-0.0299562812,
0.0537913851,
0.0363788195,
0.0269951355,
-0.0162050221,
0.0035016548,
0.0199138932,
0.0607831925,
-0.0266462285,
-0.0246862862
] |
http://arxiv.org/abs/2401.16313v1 | http://arxiv.org/pdf/2401.16313v1 | Machine Translation Meta Evaluation through Translation Accuracy Challenge Sets | 2024-01-29T17:17:42 | 2024-01-29T17:17:42 | Nikita Moghe, Arnisa Fazla, Chantal Amrhein, Tom Kocmi, Mark Steedman, Alexandra Birch, Rico Sennrich, Liane Guillou | cs.CL | cs.CL | Recent machine translation (MT) metrics calibrate their effectiveness by
correlating with human judgement but without any insights about their behaviour
across different error types. Challenge sets are used to probe specific
dimensions of metric behaviour but there are very few such datasets and they
either focus on a limited number of phenomena or a limited number of language
pairs. We introduce ACES, a contrastive challenge set spanning 146 language
pairs, aimed at discovering whether metrics can identify 68 translation
accuracy errors. These phenomena range from simple alterations at the
word/character level to more complex errors based on discourse and real-world
knowledge. We conduct a large-scale study by benchmarking ACES on 50 metrics
submitted to the WMT 2022 and 2023 metrics shared tasks. We benchmark metric
performance, assess their incremental performance over successive campaigns,
and measure their sensitivity to a range of linguistic phenomena. We also
investigate claims that Large Language Models (LLMs) are effective as MT
evaluators by evaluating on ACES. Our results demonstrate that different metric
families struggle with different phenomena and that LLM-based methods fail to
demonstrate reliable performance. Our analyses indicate that most metrics
ignore the source sentence, tend to prefer surface-level overlap and end up
incorporating properties of base models which are not always beneficial. We
expand ACES to include error span annotations, denoted as SPAN-ACES and we use
this dataset to evaluate span-based error metrics showing these metrics also
need considerable improvement. Finally, we provide a set of recommendations for
building better MT metrics, including focusing on error labels instead of
scores, ensembling, designing strategies to explicitly focus on the source
sentence, focusing on semantic content and choosing the right base model for
representations. | Machine Translation Meta Evaluation through Translation Accuracy Challenge Sets: Recent machine translation (MT) metrics calibrate their effectiveness by
correlating with human judgement but without any insights about their behaviour
across different error types. Challenge sets are used to probe specific
dimensions of metric behaviour but there are very few such datasets and they
either focus on a limited number of phenomena or a limited number of language
pairs. We introduce ACES, a contrastive challenge set spanning 146 language
pairs, aimed at discovering whether metrics can identify 68 translation
accuracy errors. These phenomena range from simple alterations at the
word/character level to more complex errors based on discourse and real-world
knowledge. We conduct a large-scale study by benchmarking ACES on 50 metrics
submitted to the WMT 2022 and 2023 metrics shared tasks. We benchmark metric
performance, assess their incremental performance over successive campaigns,
and measure their sensitivity to a range of linguistic phenomena. We also
investigate claims that Large Language Models (LLMs) are effective as MT
evaluators by evaluating on ACES. Our results demonstrate that different metric
families struggle with different phenomena and that LLM-based methods fail to
demonstrate reliable performance. Our analyses indicate that most metrics
ignore the source sentence, tend to prefer surface-level overlap and end up
incorporating properties of base models which are not always beneficial. We
expand ACES to include error span annotations, denoted as SPAN-ACES and we use
this dataset to evaluate span-based error metrics showing these metrics also
need considerable improvement. Finally, we provide a set of recommendations for
building better MT metrics, including focusing on error labels instead of
scores, ensembling, designing strategies to explicitly focus on the source
sentence, focusing on semantic content and choosing the right base model for
representations. | [
0.0368510596,
0.0477107316,
0.0029444946,
0.033320386,
-0.0281822477,
0.0095623853,
0.0296990089,
-0.0019954946,
-0.0337228216,
-0.0403263792,
-0.0432742424,
-0.0360428207,
0.0728846416,
-0.0090350639,
0.0388230011,
-0.0455272458,
0.0243891515,
-0.0000036614,
-0.0115086129,
0.0157633219,
-0.0065286066,
0.0019361345,
-0.0061567831,
0.0469980165,
0.0113611808,
0.0003277215,
0.0270648599,
0.0455615334,
0.0621091314,
-0.0289511401,
0.0235669333,
0.0554181188,
0.0019471385,
0.1187703311,
0.0000020364,
-0.0371040516,
-0.0162381902,
0.0212768577,
0.0086752195,
0.0438262187,
0.0492603965,
-0.0493543036,
-0.0135704745,
0.0118484935,
-0.0115787694,
-0.0492979586,
0.0583736897,
-0.0139715364,
0.0651507378,
0.059504915,
-0.0084021064,
0.0186778605,
0.0088525461,
-0.0461447015,
-0.0132655744,
-0.0569211766,
-0.009472128,
0.0158872642,
-0.0596227944,
0.0188741889,
-0.0198448468,
0.0388794206,
-0.0050803758,
-0.0105714863,
-0.0087301591,
-0.0061586997,
-0.05055172,
-0.027834028,
0.0010611948,
0.0691241547,
-0.0050004269,
-0.0086086085,
-0.0360271595,
-0.0101590175,
-0.0152049363,
-0.0354775153,
-0.0022159801,
0.0144335032,
-0.0053203669,
-0.0036756543,
-0.0169000942,
0.0477307215,
0.0517750457,
0.018140303,
-0.0298288465,
0.0088084759,
0.0246715322,
-0.0422016159,
-0.0013636227,
0.0406503379,
-0.0199332591,
-0.0208981056,
0.0319385119,
-0.0344254598,
-0.0372190773,
0.0039786217,
-0.0315746665,
-0.0531121269,
0.0253095832,
-0.0490418635,
0.0728845894,
0.0527931377,
-0.0401537567,
0.0580317751,
-0.0078598158,
0.1182634383,
-0.0067981333,
0.0093241306,
-0.0588155948,
-0.0141750127,
-0.0200173501,
0.0399333686,
0.0030739945,
-0.0189841893,
-0.0580053404,
0.026742937,
-0.0509577282,
-0.0019124665,
-0.0219590552,
0.0321087055,
0.0181200523,
-0.0492405482,
-0.0337445624,
-0.0304872338,
-0.019760441,
0.0513730869,
-0.0109067308,
0.005122181,
0.0073611131,
-0.0495515279,
-0.0032035119,
0.0279881787,
-0.0386570208,
-0.0309107676,
0.0194566902,
0.0906365067,
-0.0000283284,
-0.045357652,
0.0325241759,
0.002549489,
-0.044285804,
-0.0612020418,
-0.0247349031,
-0.0589837693,
0.0445595495,
-0.009777233,
0.0159559362,
-0.0215698648,
0.0085447831,
-0.007901663,
-0.0067888149,
0.0492585599,
0.0095759397,
0.0157186445,
0.0282685906,
-0.0005551597,
-0.029040603,
0.0449988693,
0.0581943654,
0.0189669337,
0.0528520271,
0.0133744832,
-0.011609178,
0.0377242751,
0.0128528941,
0.0348731838,
-0.0135608716,
-0.0183998402,
-0.0064788433,
0.0357074253,
0.027030265,
0.0855177045,
-0.0287439208,
0.0325613096,
0.0766128749,
0.0337305591,
-0.0289102346,
0.0223356504,
0.0390158482,
-0.0165777765,
0.0171553344,
-0.0210473929,
-0.0348196737,
-0.0067546391,
-0.0149052041,
-0.0243955813,
0.0753269792,
0.0341598392,
0.0002321472,
-0.0733700916,
-0.0171878468,
-0.0101208724,
0.0020270946,
-0.0239023734,
0.0281114653,
-0.0145255672,
-0.0302901398,
0.0170880761,
-0.0162185952,
-0.0686652735,
0.0181368366,
-0.0197227094,
-0.0220655203,
0.0687648579,
0.0587682128,
-0.0044483575,
-0.0288825929,
-0.0106463842,
0.0136044575,
0.063202776,
0.0833458677,
0.0082427664,
0.0281179231,
-0.0287728664,
-0.0138584767,
0.0201992113,
0.0397228114,
0.0157987811,
0.0089142499,
0.0117741507,
0.0232580043,
-0.0860005617,
-0.0913741812,
0.0388903134,
0.0034288652,
0.0027382455,
0.0198867768,
0.0012764266,
0.0424080342,
0.0267772321,
-0.0276841968,
0.0021813246,
-0.0057958988,
-0.0392140113,
-0.0215323418,
0.0262140222,
-0.0190362521,
-0.0203502718,
0.0450173318,
-0.0755442381,
-0.0267381705,
0.0290490389,
-0.0618290864,
-0.0407546163,
-0.0130959498,
-0.0353368819,
-0.0290103313,
-0.0467617288,
0.044612132,
-0.0346095301,
0.0041250065,
-0.050490804,
0.0279880669,
-0.018834047,
0.0329063758,
0.0002345467,
-0.0464519821,
0.0302256774,
0.0061894339,
-0.0548842177,
0.02605257,
0.0101863956,
0.0510954224,
0.0518904254,
0.0342153758,
-0.0441943742,
-0.044875551,
-0.0835229903,
0.0142049771,
-0.025810929,
0.0416261517,
0.0479584783,
0.0168871097,
-0.0009826469,
0.0207459219,
-0.0047419025,
-0.0410138778,
0.0134589663,
0.0000888968,
0.0335151106,
-0.0062297042,
0.0402770378,
-0.0605935007,
-0.0083233546,
0.0083916085,
0.0447805673,
0.0091563826,
0.0141844163,
-0.0545700975,
0.0281751286,
-0.0417291783,
-0.0198565461,
0.0148847029,
-0.0611638464,
0.0032121048,
-0.0040392503,
0.0110460408,
0.0337399393,
0.0074068559,
0.0237201042,
0.0420545787,
0.0056475908,
-0.0580111556,
-0.0148198763,
-0.0076701818,
-0.0406836905,
-0.0232378319,
-0.0581390932,
0.0177328885,
0.0440464653,
0.0023715666,
0.0020453664,
-0.0589718334,
0.0032461956,
-0.0429258645,
-0.0331376418,
-0.0113014067,
0.0010738791,
0.0004030646,
0.053222537,
-0.0530670583,
-0.0446290076,
0.0808266997,
-0.040686205,
-0.0059510376,
-0.0288512874,
0.0244124681,
-0.0230799932,
0.0378315896,
-0.0101083852,
0.0057784379,
-0.0398050137,
-0.0095592951,
-0.0405810289,
-0.0126390988,
0.0418289378,
-0.0544444844,
0.0464114286,
-0.0870587751,
-0.0007974294,
-0.0144879399,
-0.0106369294,
-0.0247304011,
-0.0183683522,
-0.0095129218,
-0.0168548804,
-0.0130981253,
-0.0286091994,
-0.0510292053,
-0.0002563364,
0.0167386103,
0.041285187,
-0.0088940477,
-0.0069378582,
-0.0089630391,
0.0026356194,
-0.0196021665,
0.0069029015,
0.0308762509,
-0.0514491424,
0.0370509215,
0.0282816235,
-0.0048651737,
0.0178382881,
-0.0272595342,
-0.0151680941,
0.0378401503,
0.0109082302,
-0.0255255699,
-0.0167543609,
0.0877491757,
-0.0473619327,
-0.0171202347,
0.1160730645,
0.0481224395,
-0.0401149541,
-0.0100837769,
-0.0045681889,
0.0304517634,
0.0200201888,
-0.00486951,
-0.0602302477,
-0.0153332921,
0.0172414649,
0.0553071499,
0.0093193296,
-0.0039185267,
-0.0057666064,
-0.0920213759,
0.0094006443,
0.0587500185,
-0.0533123873,
-0.0593214668,
0.0007439565,
0.0040401081,
-0.0416449159,
0.0428364947,
0.0010062696,
0.0332660824,
0.0015661112,
0.0280574784,
-0.0172550008,
-0.0980081335,
0.0277659651,
-0.0352118835,
0.0474325567,
0.0740567893,
0.0287447143,
-0.0143718598,
-0.0284693465,
0.0400728434,
-0.0210743155,
-0.0396886282,
0.0080760866,
0.0535918251,
-0.0053428533,
-0.0111596528,
0.0014645718,
-0.001538948,
-0.0212338679,
-0.0009389082,
0.0070648217,
0.0667246878,
0.0008662136,
0.0098112617,
0.0237136111,
-0.0242470503,
-0.000387531,
0.0627904534,
0.0306376275,
-0.0418208279,
-0.0078801615,
-0.0229418688,
0.0069387569,
-0.0751501024,
-0.01545313,
0.0289170593,
-0.0229613632,
-0.013379734,
-0.086508438,
-0.0767167062,
-0.0073131952,
-0.0030313618,
0.0037461277,
-0.0202835575,
-0.0179204587,
0.0006196548,
-0.0258260965,
0.0354930535,
-0.0206827056,
0.0350675434,
-0.0413008183,
0.044240471,
-0.0186522175,
0.0010529223,
-0.012350101,
0.0447525308,
-0.0810544714,
-0.0593025386,
-0.0154018542,
0.0501885898,
0.0235950407,
0.0077869417,
-0.0496593229,
-0.0214954522,
0.0206716266,
0.0195754115,
-0.0132477861,
0.0136740003,
0.0405701958,
-0.0176610555,
-0.0257957373,
0.0343562067,
-0.0340266153,
-0.0093170535,
0.0017691586,
0.0629437119,
0.0074186777,
-0.0421850719,
0.0306735858,
0.0218065083,
0.0496965125,
-0.0556290634,
-0.0424642153,
-0.0250265505,
-0.0073954528,
-0.000650264,
-0.0438195392,
-0.0378127843,
-0.0207833219,
0.0580799207,
-0.0133579448,
-0.0118828258,
0.0255478714,
-0.0356202386,
0.0016888032,
0.0058653099,
0.018700555,
-0.0356295221,
0.0852874666,
-0.0141100381,
0.0315597728,
-0.0338723846,
-0.0181667693,
0.0001912742,
0.0076879105,
-0.0275146198,
0.008735965,
0.0148277925,
0.0237812009,
0.0232094135,
0.0507415719,
-0.0310655236,
0.0278722662,
-0.0227682628,
0.0237964597,
0.0942148194,
0.0214239266,
0.1182563901,
-0.0026301641,
-0.0123116607,
-0.045438312,
0.0388434604,
0.010483237,
-0.0061956896,
-0.0153614152,
-0.0027242696,
0.0053272792,
0.0132957371,
0.0536288247,
-0.0397944637,
0.0162370056,
0.0025577825,
-0.0215130169,
-0.0230047982,
-0.0274411794,
-0.0500609726,
-0.0400981084,
-0.0307306591,
-0.030839419,
-0.0700367317,
-0.0638542101,
0.0373600051,
0.0498883612,
0.0047416799,
0.0054470119,
-0.0299560595,
0.0705987066,
-0.0332169272,
0.0369988456,
-0.0539794415,
-0.0137812141,
-0.0019723894,
-0.0039167977,
-0.0450366065,
-6.188199727e-33,
-0.0432522371,
-0.0219790451,
-0.0079118144,
-0.0512540452,
-0.0504683964,
0.006085468,
-0.030560324,
0.0018293192,
-0.004701789,
0.0384761356,
0.010989476,
0.0012791712,
0.0171203669,
-0.00192745,
0.0186488032,
-0.0054091043,
0.0460799932,
0.0098916395,
-0.0673085973,
-0.0085697938,
0.0243040211,
-0.0138165765,
0.0754573569,
-0.0758072659,
0.0939537063,
0.0016474371,
-0.0389610492,
0.0163969919,
0.0134271309,
0.0382930487,
0.0112919146,
0.0504669659,
0.0189052168,
0.0135102402,
-0.0295413844,
-0.0430472717,
-0.0877166912,
-0.019878855,
0.0382476188,
0.0392111167,
-0.0343324356,
-0.0318133086,
0.0304913502,
-0.0226444602,
-0.0370369665,
-0.0293314885,
0.0162576605,
-0.074274905,
-0.0087044779,
0.0173048023,
-0.031641949,
0.0328590423,
-0.027756352,
0.0018513429,
0.0161646791,
0.0776381791,
-0.0022414438,
0.0286290143,
-0.0720316991,
0.0196156837,
-0.0268598422,
0.027795678,
0.0680377334,
-0.0084118163,
-0.0247719027,
0.032758005,
-0.0001348823,
-0.0030052296,
0.0338822082,
-0.0113928756,
0.0275593698,
-0.0310750864,
0.0718527213,
0.0820995942,
0.0086852955,
-0.0129269511,
-0.0523213968,
0.0619295761,
-0.0305650756,
0.020026153,
0.0160823092,
-0.0089284405,
-0.0377888791,
0.0066202311,
-0.0737034529,
-0.0486901626,
-0.0462132543,
-0.013637634,
0.0141798835,
0.0112746162,
-0.0047984691,
-0.0180444457,
0.0518730134,
-0.021981027,
0.0166822355,
-0.0101751331,
-0.0206502303,
0.0115796914,
-0.007007272,
0.0518468022,
-0.0185581539,
-0.0166731235,
0.0130611649,
0.0682567209,
0.0213181898,
-0.039940916,
0.0502539203,
0.0118819028,
-0.0607216842,
0.0288775247,
-0.028887568,
-0.0066410857,
-0.0433089845,
-0.0074710906,
0.0046911971,
-0.0099106161,
-0.0205680262,
0.0151421111,
0.0188818295,
0.0131077962,
-0.0159636587,
-0.0036836958,
-0.0013915333,
0.0335056148,
-0.0446193404,
0.0281133559,
-0.008415482,
0.0052442271,
0.0506469458,
-0.0194023754,
-0.0113513647,
-0.0130637353,
2.636e-7,
0.0675670505,
0.0689745918,
0.0194841698,
0.0166777801,
-0.0016093201,
0.0133410888,
-0.0561622009,
0.0120831169,
0.0025050661,
0.0383789092,
0.0497767441,
0.0080259293,
0.1142405123,
0.0397132076,
-0.0186986886,
-0.0037980929,
-0.0589147434,
-0.0000940846,
-0.0297900401,
-0.0053898864,
0.0604559518,
0.0655851811,
0.0125142112,
0.0147054344,
-0.020959828,
-0.0273719095,
0.0885673091,
-0.0342927761,
-0.0205635689,
0.0094294334,
0.0279423557,
0.0240780711,
-0.0126308212,
0.0054537663,
-0.0145358313,
0.0004187503,
0.0557953008,
0.0559049137,
0.0182389151,
-0.0070578577,
-0.0433836877,
0.0023888049,
0.0056687435,
-0.0316574723,
0.030252181,
-0.0125582479,
-0.0214040838,
-0.0008241178,
-0.0369434394,
0.0226768553,
0.0856632367,
-0.024123637,
-0.0484527983,
0.0141135873,
0.0354779251,
0.0218596626,
-0.0223163683,
-0.0368990749,
-0.0417960212,
-0.0134019069,
0.0081917411,
-0.0698867813,
0.0033764455,
0.0055473265,
-0.0015214756,
-0.0590612032,
0.0073424154,
2.408283949e-34,
0.0581762865,
-0.0300764572,
0.01634969,
-0.0420504436,
0.0317765921,
-0.0002679974,
-0.023458723,
-0.0168703198,
-0.001669329,
-0.0396482572,
0.0133909248
] |
http://arxiv.org/abs/2401.16293v1 | http://arxiv.org/pdf/2401.16293v1 | Textual Entailment for Effective Triple Validation in Object Prediction | 2024-01-29T16:50:56 | 2024-01-29T16:50:56 | Andrés García-Silva, Cristian Berrío, José Manuel Gómez-Pérez | cs.CL | cs.CL, cs.AI, cs.DL | Knowledge base population seeks to expand knowledge graphs with facts that
are typically extracted from a text corpus. Recently, language models
pretrained on large corpora have been shown to contain factual knowledge that
can be retrieved using cloze-style strategies. Such approach enables zero-shot
recall of facts, showing competitive results in object prediction compared to
supervised baselines. However, prompt-based fact retrieval can be brittle and
heavily depend on the prompts and context used, which may produce results that
are unintended or hallucinatory.We propose to use textual entailment to
validate facts extracted from language models through cloze statements. Our
results show that triple validation based on textual entailment improves
language model predictions in different training regimes. Furthermore, we show
that entailment-based triple validation is also effective to validate candidate
facts extracted from other sources including existing knowledge graphs and text
passages where named entities are recognized. | Textual Entailment for Effective Triple Validation in Object Prediction: Knowledge base population seeks to expand knowledge graphs with facts that
are typically extracted from a text corpus. Recently, language models
pretrained on large corpora have been shown to contain factual knowledge that
can be retrieved using cloze-style strategies. Such approach enables zero-shot
recall of facts, showing competitive results in object prediction compared to
supervised baselines. However, prompt-based fact retrieval can be brittle and
heavily depend on the prompts and context used, which may produce results that
are unintended or hallucinatory.We propose to use textual entailment to
validate facts extracted from language models through cloze statements. Our
results show that triple validation based on textual entailment improves
language model predictions in different training regimes. Furthermore, we show
that entailment-based triple validation is also effective to validate candidate
facts extracted from other sources including existing knowledge graphs and text
passages where named entities are recognized. | [
0.0455289893,
-0.0018392412,
0.0280601066,
0.0437708721,
-0.0033911988,
0.0139501281,
-0.0212962199,
0.013305462,
0.0338842571,
-0.0811009109,
-0.0196624734,
-0.0300460104,
-0.0274295844,
0.0216981787,
0.0312206,
-0.003469886,
0.0678615496,
0.0262846891,
0.0386470594,
-0.0164044723,
0.0202705674,
0.0163957682,
0.010123698,
-0.0090430966,
-0.0210725423,
0.0201593935,
0.016691478,
0.0018911517,
-0.0158623476,
-0.0320733301,
0.0081709661,
0.0349589847,
0.0363509245,
-0.0023939055,
0.0000019904,
-0.099878408,
-0.0253956188,
-0.0223245136,
-0.0544288941,
-0.0295834951,
-0.03055395,
-0.0169974584,
0.0522321276,
-0.0143894963,
-0.0160417184,
-0.0227866136,
0.0131026143,
0.0604930408,
0.0432352945,
0.0739050955,
-0.0328205451,
0.0046862727,
0.0119208032,
0.0045142132,
0.0218485408,
0.0143565927,
0.0295225959,
0.004353859,
-0.0060464512,
0.0287285913,
0.0236471407,
0.0264890604,
0.0153675256,
0.020181112,
0.0331193618,
0.0414465815,
-0.0100557925,
-0.0291963574,
-0.0021118859,
0.0529743433,
0.0060523222,
-0.0500809401,
-0.0023163587,
-0.0149751017,
0.0029653839,
0.0404386185,
-0.0218172558,
0.0292950012,
-0.0582889318,
-0.0040361537,
-0.0115998909,
0.01439365,
0.0355282947,
0.0073935832,
-0.0149968537,
0.0488651097,
0.0101210661,
-0.0436921343,
0.0364255682,
0.0325368308,
-0.0348139554,
-0.0340190679,
-0.0012754586,
0.013874894,
0.0152635407,
0.0131937237,
-0.0334016867,
0.0198207628,
0.0362709053,
-0.0507802777,
-0.0092432545,
0.0038257728,
-0.028180683,
0.0636479631,
-0.040336322,
0.0594729111,
0.010678865,
0.0059127347,
-0.0214161202,
-0.0258148946,
-0.040224798,
0.015677873,
-0.0287790168,
0.0197567549,
0.0387512818,
-0.0090123573,
-0.0147544863,
-0.0124635268,
-0.0050243363,
0.077830404,
-0.0254075211,
0.0126384906,
-0.0076046209,
-0.0165570639,
-0.0740204602,
0.0307754595,
-0.0285631567,
0.0023868114,
-0.0319696032,
-0.0375430696,
-0.0212987289,
0.0188068356,
0.0204486866,
-0.0327775478,
0.0547310673,
0.0669138804,
0.0022911155,
-0.028873533,
-0.0269510373,
-0.0568835512,
-0.0058144904,
-0.0459502228,
0.0455201566,
-0.0681558251,
0.0158761255,
-0.0085121244,
0.0027864396,
-0.0142082376,
0.0087081734,
0.0171953961,
0.0067947023,
0.0397811234,
-0.043971125,
-0.028873181,
-0.0076732063,
0.012878485,
0.0213263016,
0.0472308397,
0.0131831681,
0.0121692615,
0.0339403935,
0.0184658859,
0.014551023,
0.0224642381,
0.0407595597,
0.0010346256,
-0.0146970842,
0.0359029658,
0.054878898,
0.0000306626,
-0.01667854,
0.0241930895,
-0.0238358788,
-0.0513677895,
0.0498163924,
0.0398206897,
-0.0174985286,
-0.0359490365,
0.0244314894,
-0.0308813211,
0.0306148268,
0.0183153078,
-0.0509254076,
0.0126202852,
-0.0569682978,
0.0385905281,
-0.020773774,
0.0195920318,
-0.0152609153,
-0.0280916896,
-0.058392141,
-0.0095735453,
-0.0242353901,
-0.0331988297,
0.0592492558,
0.0354269259,
-0.0327706896,
0.0492686518,
0.0461463295,
-0.0653720424,
-0.0244174916,
-0.019339703,
0.0573917069,
0.0624088682,
0.0444166288,
-0.0209174268,
-0.0544375405,
0.0038506784,
0.0184227154,
0.0464047194,
0.0142719662,
-0.0222665761,
0.0226799138,
-0.0102753015,
-0.0017189498,
0.0118026379,
0.0139210941,
0.0101850862,
0.0429873504,
0.0204569884,
0.0215043221,
-0.0176472608,
-0.0097494936,
0.0513075329,
0.0004819072,
0.0038585039,
-0.023204878,
0.0325999223,
0.0021444841,
-0.0041307583,
-0.0162458252,
-0.0086795827,
0.0244281683,
-0.0169849116,
-0.0504512861,
0.002398158,
-0.0271638837,
-0.0888883322,
0.0324010551,
-0.0129601452,
-0.028487226,
-0.038150873,
-0.0085980017,
-0.034599483,
-0.027201172,
-0.0199166872,
0.0614500009,
-0.0417108871,
-0.0096709058,
-0.0319866538,
-0.0351125225,
-0.0089210412,
0.0813305825,
0.0105690826,
0.0316719413,
-0.0408864655,
-0.0081671178,
0.0289535038,
-0.0179010704,
-0.0260762889,
0.009138423,
0.0185212307,
0.0061738743,
0.0156682599,
-0.0018466541,
-0.0549483299,
0.0306261461,
-0.0807447508,
0.0189916734,
0.0085056517,
0.0003527341,
-0.0077321054,
-0.0243611205,
-0.0022766509,
-0.026803907,
-0.0213696677,
-0.0034552764,
0.0069435197,
0.0150569417,
0.0619087443,
-0.0262722392,
-0.0085013602,
-0.032030236,
-0.0040442953,
-0.0494526774,
0.0883090943,
0.051912237,
-0.0277346242,
0.0015259382,
0.0774556696,
-0.0811566636,
0.0301673431,
-0.0171139482,
0.0094322721,
0.0104840267,
-0.0312049892,
0.0080165565,
0.0473297462,
-0.0229329765,
-0.0108574992,
0.0549042039,
0.0066359481,
-0.0291921403,
0.0072032292,
-0.0270439293,
-0.0117843281,
0.0233803429,
0.0241154302,
0.0064496854,
0.0230134167,
0.0084585361,
-0.0248630997,
-0.0502217896,
-0.002183368,
0.0053923633,
0.035540089,
-0.0597041622,
-0.0336960629,
0.0336476415,
0.0413116105,
-0.0425249301,
-0.0297792815,
0.1048986241,
-0.0677016824,
-0.0265056584,
-0.0113003217,
-0.0180815961,
-0.0113852751,
0.0760243833,
0.0355937369,
0.0078578778,
0.0446790121,
0.0162347779,
-0.0266202465,
-0.0217104573,
0.0213769339,
-0.0326802619,
0.0384181775,
-0.0146535179,
0.0208729487,
0.0077643134,
-0.02180361,
-0.0437838286,
-0.0309494305,
-0.0192269422,
-0.0163162127,
-0.0778072998,
-0.0265483074,
-0.0236700028,
-0.0081006447,
0.0132275773,
0.0419887602,
-0.0150834257,
-0.0292461365,
-0.0165152848,
-0.0012112817,
0.0612074584,
0.0227905121,
-0.0136879981,
-0.0513893887,
-0.0091601349,
0.0501505248,
-0.0236129109,
-0.024775967,
-0.0194955915,
0.0104663316,
-0.0097452216,
-0.0166971311,
-0.0259540938,
-0.0009352558,
0.0052741226,
-0.0731913745,
-0.0160472319,
0.121878162,
0.012768548,
-0.0005130089,
-0.0294308867,
0.0088739041,
0.030465059,
0.0497275554,
-0.0001541996,
-0.1008908898,
-0.0740775615,
-0.0241846405,
-0.0100555848,
-0.0054612821,
-0.0135926595,
-0.0388687588,
-0.1399950832,
0.0261577126,
-0.0175723862,
-0.0291502643,
-0.0457105488,
-0.0910889208,
0.0073878285,
0.000292407,
-0.0524378419,
0.0137557508,
-0.0165962242,
0.0033337015,
0.010823274,
-0.0261857156,
-0.0307526048,
0.0504653938,
-0.0453962684,
-0.0110551538,
0.011713949,
0.0281197131,
-0.0079145897,
-0.036317274,
-0.0221789293,
-0.0053085727,
-0.0125087807,
0.0136969993,
0.0608535223,
-0.0369902961,
0.0131258145,
-0.0178583786,
-0.0048023476,
0.0019461166,
-0.0238692071,
0.0263179168,
0.0673796907,
-0.0046000285,
-0.053415373,
-0.0051744399,
0.0066519817,
0.0124346912,
0.0939684287,
0.0114314035,
0.0066389535,
0.0039324183,
-0.0259503406,
-0.0223996937,
0.031457983,
-0.0086166058,
0.0002755967,
-0.060223598,
-0.0315785408,
-0.1008358151,
-0.0189262945,
0.0131274667,
0.0795003548,
0.0219392646,
-0.0156343207,
0.0181963462,
0.0017623563,
-0.0514482819,
0.0396277718,
-0.0326146856,
-0.0179639198,
0.0026472828,
0.0290006325,
0.051327534,
-0.0336814374,
-0.0256095324,
0.0306721777,
-0.0485904068,
-0.0479144119,
-0.0673969314,
0.0342744812,
0.0217838325,
0.077241078,
-0.0202082675,
0.0241215937,
0.0835769325,
0.0731804818,
-0.0293283761,
0.0041545504,
0.0345660262,
-0.0216257256,
0.0206761751,
0.0096038571,
-0.0458394662,
0.0291034449,
-0.0085267248,
-0.0221953355,
-0.0500531793,
-0.0580977984,
-0.0050172079,
0.0440877303,
0.0728187338,
0.0246092286,
-0.0002908237,
-0.0562614538,
0.0169429723,
0.0336671881,
-0.045154769,
-0.0113931699,
0.0312478859,
-0.007986851,
0.0113335522,
-0.006993107,
-0.0126225427,
0.0008728536,
0.0335815214,
-0.0034196612,
-0.0631909817,
0.055868417,
0.0356551073,
0.0061970698,
-0.0289777704,
-0.0391577557,
0.0396763906,
0.0073255878,
0.0329728313,
-0.0296938885,
0.0151424035,
0.0213571787,
0.0036914372,
-0.0221128892,
-0.0410959274,
-0.0854874104,
0.0442118272,
-0.0366921462,
0.0284958221,
0.0356382541,
-0.0389539897,
0.0604424775,
-0.0314461142,
0.0091163144,
0.0090095922,
0.0377910808,
-0.0183068682,
-0.0521522611,
0.0090031754,
0.0177790932,
-0.0320843868,
-0.0149368839,
0.0469162576,
0.0053375373,
0.0841363519,
0.0176046789,
-0.035713166,
-0.039671395,
-0.0464766324,
-0.0085167354,
0.0069363941,
-0.0047129514,
0.0030067933,
-0.0235497616,
-0.0589014366,
0.0546700209,
0.0436552763,
0.003501876,
0.0280616898,
0.0357621089,
0.0580283888,
0.0339954868,
-0.0436362103,
-0.0191128757,
0.017646037,
-0.0125678414,
0.0267979093,
-0.0077747079,
-6.514388389e-33,
-0.0661738813,
-0.047698155,
0.0235475935,
0.0190570857,
0.0165493898,
0.0125759793,
-0.0376142971,
-0.0154598569,
-0.0214558598,
0.0362915322,
0.0164432917,
0.0368338861,
-0.0017740902,
-0.0120399483,
0.0155449566,
-0.011126007,
0.0267652981,
0.0046398379,
-0.0472457223,
0.0128758727,
0.0929178745,
0.0221792199,
0.0858035535,
-0.0261428785,
0.0373128876,
-0.0254829824,
-0.0064345668,
0.0400289781,
-0.0464572385,
0.0403801203,
-0.012659004,
0.0333078355,
0.0100095123,
-0.0353540182,
0.0158163775,
0.0207556207,
-0.0814822093,
-0.0797107294,
0.0092311893,
-0.0006837072,
0.0043339753,
0.0492489785,
0.0353660025,
-0.0262782499,
0.0198240597,
-0.0094396621,
0.0687348694,
-0.0071370625,
-0.0118868798,
-0.0432725772,
-0.0574991629,
0.0150114922,
0.00709775,
0.0058110408,
-0.0109008867,
0.0264144,
0.0374335162,
0.0010382284,
-0.0614574514,
0.0206556655,
-0.0224574842,
0.008931445,
0.0114208106,
0.0088149682,
0.0097652376,
0.1310884356,
-0.0530198254,
-0.0302966554,
-0.0788144693,
0.0345695466,
0.0617964827,
0.0468584113,
0.0259477515,
0.0749761835,
0.0430265591,
0.0138123101,
-0.1093818471,
0.0577250533,
-0.032413125,
0.0288678017,
-0.0029754899,
-0.047427319,
-0.0396020599,
-0.0302701779,
0.0011115273,
0.0143313073,
-0.025147561,
0.0184495635,
0.0037252943,
-0.0168445725,
-0.0193528477,
0.0161917023,
-0.0085827317,
0.0034395261,
-0.013104138,
0.0709535331,
-0.0408446752,
-0.0184228066,
-0.0110118715,
0.0306602102,
-0.0457635,
-0.0282981209,
0.0255428758,
0.0086181806,
0.0148853967,
0.0269998331,
0.08591865,
-0.0121974275,
-0.0539880097,
-0.0095030367,
-0.0385372601,
-0.010781209,
-0.0061079669,
-0.0263565965,
0.0431688428,
0.0164124183,
-0.0036088566,
0.0433908477,
0.0270803589,
-0.0018937755,
-0.0584411211,
0.018334832,
0.028791802,
0.0174537394,
-0.0183530767,
-0.0103240199,
-0.0246728454,
-0.0096147051,
0.0030941474,
-0.0580339134,
0.0202085115,
0.0165304821,
2.741e-7,
0.0276567172,
0.0356442891,
0.0049081338,
0.0202410705,
-0.0293449312,
0.0591247752,
-0.0357303359,
0.0282222833,
0.0076465681,
-0.073766537,
0.0089685982,
0.0478769392,
0.02157229,
0.069282718,
-0.1145588756,
0.0028313769,
-0.0810663849,
-0.0483341031,
-0.0446117334,
0.0113264965,
0.0459629558,
0.0498951711,
0.0734175518,
-0.0397132188,
0.0119483871,
-0.0776057616,
0.0621414557,
0.0074962801,
0.01732531,
0.0202907119,
0.0594774112,
0.0603842027,
0.0038733815,
0.0647742152,
0.0086201122,
0.0314168185,
0.0021943715,
0.0314489603,
-0.0184879322,
0.0156361125,
-0.0233786833,
0.0072581493,
-0.0081611359,
-0.0149878524,
0.079901956,
0.0334717669,
0.019495165,
-0.0418700278,
0.0078633083,
-0.0187656917,
0.037816979,
-0.0021293464,
-0.0059294757,
-0.0204417408,
-0.0087660467,
-0.0207283981,
-0.0028187502,
-0.0707462355,
-0.0491888635,
0.0268458575,
-0.006268959,
-0.0604915656,
0.0145843234,
-0.0344897136,
0.0123779606,
-0.0867073387,
0.0124899484,
2.592983499e-34,
0.0035315428,
-0.021006586,
0.0016617749,
-0.0174511205,
0.0398837104,
-0.0028887859,
-0.0130869346,
-0.0152740693,
0.0056521017,
-0.0224191565,
-0.001123755
] |
http://arxiv.org/abs/2401.16287v1 | http://arxiv.org/pdf/2401.16287v1 | GAPS: Geometry-Aware Problem Solver | 2024-01-29T16:48:34 | 2024-01-29T16:48:34 | Jiaxin Zhang, Yinghui Jiang, Yashar Moshfeghi | cs.AI | cs.AI, cs.CL | Geometry problem solving presents a formidable challenge within the NLP
community. Existing approaches often rely on models designed for solving math
word problems, neglecting the unique characteristics of geometry math problems.
Additionally, the current research predominantly focuses on geometry
calculation problems, while overlooking other essential aspects like proving.
In this study, we address these limitations by proposing the Geometry-Aware
Problem Solver (GAPS) model. GAPS is specifically designed to generate solution
programs for geometry math problems of various types with the help of its
unique problem-type classifier. To achieve this, GAPS treats the solution
program as a composition of operators and operands, segregating their
generation processes. Furthermore, we introduce the geometry elements
enhancement method, which enhances the ability of GAPS to recognize geometry
elements accurately. By leveraging these improvements, GAPS showcases
remarkable performance in resolving geometry math problems. Our experiments
conducted on the UniGeo dataset demonstrate the superiority of GAPS over the
state-of-the-art model, Geoformer. Specifically, GAPS achieves an accuracy
improvement of more than 5.3% for calculation tasks and an impressive 41.1% for
proving tasks. Notably, GAPS achieves an impressive accuracy of 97.5% on
proving problems, representing a significant advancement in solving geometry
proving tasks. | GAPS: Geometry-Aware Problem Solver: Geometry problem solving presents a formidable challenge within the NLP
community. Existing approaches often rely on models designed for solving math
word problems, neglecting the unique characteristics of geometry math problems.
Additionally, the current research predominantly focuses on geometry
calculation problems, while overlooking other essential aspects like proving.
In this study, we address these limitations by proposing the Geometry-Aware
Problem Solver (GAPS) model. GAPS is specifically designed to generate solution
programs for geometry math problems of various types with the help of its
unique problem-type classifier. To achieve this, GAPS treats the solution
program as a composition of operators and operands, segregating their
generation processes. Furthermore, we introduce the geometry elements
enhancement method, which enhances the ability of GAPS to recognize geometry
elements accurately. By leveraging these improvements, GAPS showcases
remarkable performance in resolving geometry math problems. Our experiments
conducted on the UniGeo dataset demonstrate the superiority of GAPS over the
state-of-the-art model, Geoformer. Specifically, GAPS achieves an accuracy
improvement of more than 5.3% for calculation tasks and an impressive 41.1% for
proving tasks. Notably, GAPS achieves an impressive accuracy of 97.5% on
proving problems, representing a significant advancement in solving geometry
proving tasks. | [
0.0122665046,
0.0809565559,
-0.0235341191,
0.0356468447,
-0.0711924136,
0.0168174263,
0.0368564427,
0.0180651993,
-0.0063347998,
-0.0490838587,
0.0475951843,
0.0265843254,
0.0276059937,
0.0232634414,
-0.0164064523,
-0.0802644417,
0.0129187694,
-0.0218005069,
0.0216371715,
0.0214128271,
-0.0188987609,
0.0106609734,
-0.0561216585,
0.0142223518,
-0.0616581813,
-0.0441867672,
-0.0290756449,
-0.0448787063,
0.0124554308,
-0.0641943365,
0.0258710533,
0.0600369796,
-0.0187191274,
0.0378517993,
0.0000020527,
-0.019967733,
-0.0524877422,
0.0324008912,
-0.0069072894,
-0.0123962276,
-0.0440009721,
-0.0556753054,
-0.0011252299,
-0.0147342924,
-0.0626659617,
0.0251906458,
0.0144389169,
0.0312122293,
0.0582462437,
0.0656314716,
-0.0244669504,
-0.0437905118,
0.0166654047,
0.0309552588,
0.0324300714,
-0.0306453872,
-0.0183227435,
0.012268289,
0.0135664782,
0.0131583083,
-0.0194082689,
0.0301902108,
0.0012892092,
0.01718143,
-0.0066167903,
-0.0330883116,
-0.054593239,
0.0377311222,
-0.0188797731,
0.0480114631,
-0.0136206886,
-0.0069911098,
0.0149641056,
-0.0145345312,
-0.0055581573,
0.0312456414,
-0.0589542091,
-0.0086792605,
-0.031263642,
-0.0145715401,
-0.0333213881,
-0.0071513276,
-0.0137277646,
0.0042175478,
-0.0378034674,
0.0685072765,
-0.0181697663,
-0.0668907985,
-0.0256023519,
0.0265662391,
-0.0704260617,
-0.0637678355,
0.0750117451,
0.0582317561,
-0.0036385558,
0.0182265099,
0.0012033605,
-0.0644950718,
-0.0103048263,
-0.0593300126,
0.0516342446,
0.0256463457,
0.0569906496,
0.0284173805,
-0.0471123196,
0.0556020215,
0.02103048,
0.0497238338,
-0.0059991726,
0.0594309717,
-0.0268696379,
-0.0107708471,
-0.0622773767,
0.0362769105,
0.0124374917,
0.0149818333,
0.0463158004,
0.0183203556,
-0.0220719296,
0.0154101467,
0.0070198867,
0.0120562585,
0.0122769345,
-0.0353671312,
-0.0282328278,
-0.0031826664,
-0.0421409383,
0.009770521,
-0.0261943899,
0.0077105039,
0.0213840492,
-0.0042752135,
0.0551651269,
-0.0241113957,
0.0389599763,
0.0256432835,
-0.003275689,
-0.0089524584,
-0.0396165214,
-0.0463904254,
0.0132892802,
-0.077304624,
-0.0247004479,
-0.0891284719,
0.0245927591,
0.0181901027,
-0.0491423123,
-0.023888357,
0.0174729172,
0.0103993621,
0.0146348821,
0.0529798195,
0.0065173008,
0.0100233853,
0.0706498995,
0.013230036,
0.025049245,
0.0538482107,
0.0605490394,
0.016449891,
0.0358469374,
0.0194969382,
0.0282015521,
0.0197835993,
0.0183334183,
-0.033753626,
-0.0240457077,
0.0010272246,
0.0009818901,
0.0504007563,
-0.0542000271,
0.0409806035,
-0.0139332367,
0.0137206605,
-0.0127016027,
0.0304264352,
0.0165386945,
0.0335930809,
0.0275502875,
0.0308255721,
0.0176524166,
-0.0012146682,
-0.0427008271,
-0.0163747463,
-0.0451633185,
0.0102016255,
-0.0214532651,
0.0286503341,
0.0077634058,
-0.0854062885,
-0.0544939712,
0.026488943,
-0.0058251028,
-0.0044247457,
0.0909767151,
-0.0026479352,
-0.0093571143,
0.0453001671,
-0.0086585237,
-0.0308680609,
-0.0121401139,
0.0098005757,
0.0240082424,
-0.0198992528,
0.0607908294,
-0.0634346008,
-0.0540160276,
0.0002417146,
0.0117400773,
-0.0251032151,
0.0690865591,
-0.0121702487,
-0.0501142852,
-0.0100894822,
-0.0375552103,
0.0217132438,
0.0181584768,
-0.0248658042,
0.0834729224,
0.0281329174,
0.0244103074,
-0.0531500541,
0.0172309652,
0.0348270498,
0.0049619507,
-0.0292463806,
-0.0166166183,
0.0377954803,
-0.051974345,
0.012761334,
0.0244149696,
0.0187588446,
0.0212650783,
-0.0251504853,
-0.0171926897,
-0.0241369177,
-0.0351922251,
-0.0547159538,
0.0536354519,
-0.0634425208,
-0.0359356292,
-0.0247092545,
-0.044346083,
0.0031349908,
0.0308160018,
-0.0391277708,
0.0139346672,
-0.0194020551,
0.0352625065,
0.0006115726,
-0.0534912646,
-0.0476393551,
0.0907063186,
0.0326638967,
0.0290942658,
-0.036254283,
0.0015117695,
0.0399878174,
-0.0046316031,
0.00389735,
-0.0210033227,
-0.0189093482,
0.0186215974,
-0.0253614504,
0.0405024998,
0.0096803755,
-0.0651792958,
-0.0402295999,
0.0016937305,
0.0022377139,
0.0558350272,
0.0169997495,
-0.042753689,
-0.0379585363,
-0.0202184208,
-0.0394809246,
-0.0521406457,
0.002108387,
-0.0119915409,
0.031872984,
0.0060865064,
0.0115222484,
-0.0163507164,
-0.0228388458,
-0.0516278818,
0.0280916989,
0.0443702377,
-0.0087271519,
-0.0283960477,
0.0423239991,
-0.0285685658,
0.0152349724,
0.041254282,
0.0137516242,
0.007635511,
-0.0195271969,
0.0175419785,
0.0357358791,
0.0120163383,
0.0113563482,
0.0204368904,
-0.030877972,
-0.0754468814,
-0.0019879134,
-0.057034377,
0.0228263903,
0.014391073,
-0.0398670956,
0.0018160653,
0.0448213108,
-0.0172908865,
-0.0194525998,
0.0180730689,
0.0464980751,
-0.0120541677,
0.0415330678,
-0.0408331677,
0.0173713416,
0.01034417,
0.0466167144,
0.0025895624,
-0.0153169241,
0.050674662,
-0.0177404508,
0.009396024,
-0.0748993903,
-0.0008522821,
-0.0973810479,
0.0483082868,
-0.0042604152,
0.0233070981,
0.0048800027,
-0.0403533615,
0.0055471975,
0.0318642445,
0.0556402802,
-0.0001537803,
0.0244772434,
-0.0338990353,
0.0122350687,
-0.0017562334,
-0.0781559721,
-0.0378021412,
0.0531441495,
0.0099205188,
0.0172890685,
-0.0057339822,
-0.0195231307,
-0.0107952915,
0.0587643571,
0.0147255296,
0.0299460832,
-0.0536628477,
-0.0129251452,
-0.0486247465,
0.0397809595,
0.0204994623,
0.0433964469,
-0.0106896283,
0.0112606101,
-0.0505600534,
0.0409214608,
0.0031567162,
0.0098454403,
0.0259121023,
0.0446654744,
0.0283773318,
-0.0520435683,
-0.0370530561,
-0.0088771228,
-0.003409093,
-0.0781995729,
-0.0293960162,
0.0609287806,
0.0902809799,
0.005582747,
-0.031179186,
-0.0649449155,
-0.0154402964,
0.0200931802,
0.0414408334,
-0.0613766834,
0.025965536,
0.0213080328,
-0.0233596917,
0.0132048223,
-0.0040555266,
-0.0018933571,
-0.074726373,
0.059508495,
0.0222193152,
-0.0380694233,
-0.0253945086,
-0.0232902728,
-0.0141802151,
0.0018512021,
-0.0673725829,
-0.0045967824,
0.0032260895,
0.019062683,
0.0328689516,
-0.0468710326,
0.0070346734,
-0.0291009396,
-0.0284120906,
-0.0719242916,
-0.0040709167,
0.0262191407,
-0.0019773077,
-0.1232229993,
0.0445050374,
0.0626570657,
-0.0110722231,
-0.0370969698,
0.0158306751,
0.0332298279,
0.0177918412,
0.0029329122,
-0.0072344081,
-0.0853697881,
-0.045795273,
0.0418948792,
0.1038216427,
0.0045958399,
0.0438946374,
-0.0377081223,
-0.0143496767,
0.008740793,
-0.0014833745,
-0.0063051251,
0.0168401469,
-0.1170691922,
-0.038711071,
0.0234011374,
-0.0665955693,
0.0022415058,
-0.0026555797,
0.0336866677,
-0.0146665052,
-0.0483108424,
0.0024764217,
0.0568437874,
0.0921462849,
-0.0018071744,
-0.0700097755,
0.003958445,
-0.0148224086,
-0.062183626,
0.0133362366,
-0.0059706206,
0.0554859862,
-0.0405430533,
0.0171096046,
-0.043685399,
0.0046365666,
-0.0358032696,
-0.0161705744,
-0.0638084114,
-0.0333654433,
-0.0105159646,
0.0175916776,
0.0134690348,
-0.0032648928,
-0.0202214997,
0.0151451938,
-0.0144836074,
0.0502950214,
-0.0415220484,
-0.0301765744,
-0.007702555,
0.0219797306,
0.0095907003,
-0.0245591234,
0.0498371758,
0.0017153139,
-0.0040002097,
0.0353492796,
-0.0563684888,
0.0086594094,
0.0270617642,
0.0373137817,
0.0242641121,
0.0317891985,
0.0020810312,
-0.0218943823,
-0.0108095864,
-0.0386481136,
-0.0315215252,
-0.0146772079,
0.0535713732,
-0.0457045585,
-0.004683604,
0.0260608345,
0.0248339139,
0.0032772231,
0.0468938202,
0.0072438358,
0.0403313227,
0.0221744776,
0.0021563875,
0.0382757634,
0.0167197101,
0.0197187569,
-0.0243618712,
0.0101705445,
0.0391192846,
-0.0141820107,
0.0181422047,
0.0078963228,
0.0087074786,
-0.0501734242,
-0.0220336057,
-0.0170420948,
-0.0540470295,
-0.0747812539,
0.0268971305,
0.0276390947,
0.0354072005,
-0.0758840367,
-0.0068229325,
-0.0387769267,
0.0298403129,
-0.007812771,
0.0268229228,
-0.0100705279,
0.0106788101,
0.0287216268,
0.0032732734,
-0.0782571509,
0.0175266657,
-0.056673985,
0.0204174537,
0.0154074654,
-0.0475796647,
-0.0211968347,
-0.004190349,
0.0206048284,
0.0475336947,
0.0077620442,
0.0193227176,
-0.0732352436,
-0.0414416492,
-0.0002148993,
0.0204294417,
0.0026023653,
0.079259716,
0.0340735167,
0.0574673004,
-0.0193006825,
-0.0470903069,
-0.0011724258,
0.0248679947,
-0.0553696863,
-0.0127019212,
0.0179712996,
-5.303866878e-33,
-0.0398506485,
-0.0776150972,
0.0360808112,
0.0091426186,
0.0130314976,
-0.028290337,
-0.0085089365,
0.0497617982,
0.014370379,
-0.01221928,
-0.0064556757,
-0.017440265,
0.0299147889,
0.001043844,
-0.0086564971,
0.0216617752,
0.013229467,
0.002816553,
-0.0312435143,
-0.0400336683,
0.0063457773,
0.0099652242,
0.058011122,
-0.0857596248,
0.0121584022,
0.0308894124,
-0.0136258574,
0.0483049229,
-0.0049515618,
0.0654176548,
0.0150212636,
0.0210982133,
0.0114328219,
0.0227155723,
-0.031662073,
0.0098528238,
-0.0625654012,
-0.0589340515,
-0.0280750133,
0.0350424647,
-0.0270676091,
-0.0274388827,
0.0175311379,
0.0027448945,
0.0082800454,
-0.0332845002,
0.050314784,
-0.0136452885,
-0.0070367078,
0.0220046174,
-0.0517860875,
0.0535353534,
-0.0117312353,
-0.0105712973,
0.0448426902,
0.0088643311,
0.0033530192,
-0.0479725935,
-0.0029613301,
0.0392314047,
0.0652707368,
0.018773336,
0.0513141565,
-0.0312999338,
-0.045968689,
0.0197353065,
0.0159813315,
-0.0598622933,
-0.010642115,
0.0075098448,
0.0394619778,
0.0697782785,
0.0377714299,
-0.0147994068,
0.042930916,
-0.0142181264,
0.0171526819,
0.0258612949,
-0.0715303719,
-0.0015056571,
-0.0215784628,
0.0045457813,
-0.0692287758,
-0.0223701037,
-0.0113568474,
-0.0175658185,
-0.0408116579,
-0.0151245845,
0.0170743689,
0.0284499805,
0.0153526682,
0.0843018517,
-0.0121636838,
-0.0276166312,
-0.0313316062,
0.0074902521,
0.0172376279,
0.0183669087,
-0.0178889912,
0.0264961943,
-0.0236187968,
-0.0244631991,
-0.0171199553,
0.0407087915,
0.0197646804,
-0.0010770934,
0.0100660315,
0.0179647487,
-0.0588216148,
0.0051464648,
-0.0424524881,
-0.0146351196,
0.0277185291,
0.0051299506,
-0.0032505656,
-0.0139513602,
-0.015177032,
0.0459014475,
0.0063853962,
-0.0243422389,
-0.0352668725,
0.0455788374,
-0.0407558903,
0.0305684786,
0.0091525503,
-0.0189875215,
0.0092974901,
-0.0101275351,
0.0505674072,
-0.005614961,
0.021192994,
-0.0289819594,
2.645e-7,
0.043565765,
0.0670770034,
0.0373795182,
-0.0024121201,
0.0056533692,
0.014892214,
0.0188101549,
-0.0216717944,
-0.0154243764,
0.077367112,
0.0400303751,
0.0269391201,
0.012953409,
0.0522504784,
-0.0525127761,
0.1030403599,
0.0247603972,
-0.0624445938,
-0.1112992242,
0.0135402512,
0.1033372954,
-0.0048518805,
0.0333481766,
0.0050685471,
-0.0453315936,
-0.0291857086,
-0.0130395945,
0.0255386811,
0.0059224847,
0.0528699905,
0.0428726673,
-0.0420085825,
0.0344598964,
-0.0009060274,
-0.0099687912,
0.0264691263,
0.0595808811,
0.0489522032,
0.0067483676,
0.0319543034,
0.006215666,
-0.051369939,
-0.018390242,
-0.0409248285,
0.0204709414,
-0.0400682278,
-0.0080364244,
-0.014858136,
0.0109747816,
-0.0225297064,
0.000207377,
0.0028895708,
0.0389137529,
0.0024488191,
0.0495561995,
-0.0170098133,
-0.0103354519,
-0.0765705556,
-0.0318827368,
0.0708800554,
-0.0188449305,
0.0035754815,
-0.0053103603,
0.0279021934,
0.0077108275,
-0.0096926391,
-0.0362267196,
2.544322625e-34,
-0.0040014386,
0.0451788083,
-0.0053902222,
0.0480943248,
0.037015833,
0.0129306791,
0.0143322134,
0.0179812536,
-0.0242640991,
-0.0221871156,
0.0061172382
] |
http://arxiv.org/abs/2401.16285v1 | http://arxiv.org/pdf/2401.16285v1 | Capturing Pertinent Symbolic Features for Enhanced Content-Based Misinformation Detection | 2024-01-29T16:42:34 | 2024-01-29T16:42:34 | Flavio Merenda, José Manuel Gómez-Pérez | cs.CL | cs.CL, cs.AI | Preventing the spread of misinformation is challenging. The detection of
misleading content presents a significant hurdle due to its extreme linguistic
and domain variability. Content-based models have managed to identify deceptive
language by learning representations from textual data such as social media
posts and web articles. However, aggregating representative samples of this
heterogeneous phenomenon and implementing effective real-world applications is
still elusive. Based on analytical work on the language of misinformation, this
paper analyzes the linguistic attributes that characterize this phenomenon and
how representative of such features some of the most popular misinformation
datasets are. We demonstrate that the appropriate use of pertinent symbolic
knowledge in combination with neural language models is helpful in detecting
misleading content. Our results achieve state-of-the-art performance in
misinformation datasets across the board, showing that our approach offers a
valid and robust alternative to multi-task transfer learning without requiring
any additional training data. Furthermore, our results show evidence that
structured knowledge can provide the extra boost required to address a complex
and unpredictable real-world problem like misinformation detection, not only in
terms of accuracy but also time efficiency and resource utilization. | Capturing Pertinent Symbolic Features for Enhanced Content-Based Misinformation Detection: Preventing the spread of misinformation is challenging. The detection of
misleading content presents a significant hurdle due to its extreme linguistic
and domain variability. Content-based models have managed to identify deceptive
language by learning representations from textual data such as social media
posts and web articles. However, aggregating representative samples of this
heterogeneous phenomenon and implementing effective real-world applications is
still elusive. Based on analytical work on the language of misinformation, this
paper analyzes the linguistic attributes that characterize this phenomenon and
how representative of such features some of the most popular misinformation
datasets are. We demonstrate that the appropriate use of pertinent symbolic
knowledge in combination with neural language models is helpful in detecting
misleading content. Our results achieve state-of-the-art performance in
misinformation datasets across the board, showing that our approach offers a
valid and robust alternative to multi-task transfer learning without requiring
any additional training data. Furthermore, our results show evidence that
structured knowledge can provide the extra boost required to address a complex
and unpredictable real-world problem like misinformation detection, not only in
terms of accuracy but also time efficiency and resource utilization. | [
0.0395415537,
0.0306382533,
-0.0397072807,
0.0570574105,
-0.0531964935,
0.0308552105,
-0.006269455,
0.0106616039,
0.0195971262,
-0.0490147136,
-0.0147568658,
-0.0202951189,
0.0165605117,
0.0595017225,
0.01344823,
-0.0121009871,
0.0650339648,
0.0336657763,
0.0695064813,
-0.0105584133,
0.0269081499,
0.040591985,
0.014110445,
0.0162729472,
-0.061214529,
-0.0348356441,
0.0436305627,
0.0067485711,
0.0508562289,
-0.0153443841,
0.0323943458,
0.0330954269,
0.0349170081,
0.0665742457,
0.000001739,
-0.0045536249,
-0.0160410237,
0.0147122964,
-0.03192598,
-0.0049978583,
0.0747747347,
-0.0764626265,
0.0138244955,
0.0184069723,
-0.0270655993,
0.0190328304,
0.0478575639,
0.076964587,
0.0787692443,
0.1061728373,
-0.0164764747,
0.0113251898,
0.0004635459,
-0.0018281525,
0.0084912311,
-0.0193703268,
0.0704961047,
-0.0030033484,
-0.0458440557,
0.0507310852,
0.0174027272,
0.013869687,
0.0012169518,
-0.005584646,
-0.0240984242,
-0.0125520658,
-0.0109810736,
-0.0168499071,
-0.0198899563,
0.0608188212,
-0.03879866,
-0.0558379218,
0.010503076,
-0.0111406436,
0.008049177,
0.0338861085,
0.0036985481,
0.0160583146,
-0.0265556164,
0.037490584,
0.0295673274,
0.0016243872,
0.0540335402,
0.0324259028,
0.0134238135,
0.0367607288,
-0.0196888912,
-0.0280347504,
0.0369354524,
-0.0221771132,
-0.0078715915,
-0.0489663742,
-0.0133224679,
0.0145897474,
0.0025653483,
0.0266387388,
-0.0981791243,
-0.0643354058,
0.0155427288,
-0.1185713708,
0.0026996348,
0.0038699971,
0.0221905652,
0.048437532,
-0.0054041818,
0.0212675761,
0.0043823044,
0.0660819635,
-0.0175761599,
-0.0433753692,
-0.0395467505,
-0.0010897779,
-0.0108265504,
0.0610262193,
-0.0003003861,
-0.0051711272,
-0.0380624384,
-0.0142372018,
-0.0528771244,
0.0112312902,
-0.0060727387,
0.0180216096,
-0.0294443499,
-0.0211852156,
-0.001644465,
0.0096403779,
0.0200863555,
0.0395022407,
-0.0275816526,
-0.033148516,
0.0051824497,
0.0227783471,
-0.0328962058,
0.0004739938,
-0.0044238013,
0.0476643778,
0.0250647627,
0.0027918944,
-0.0374042429,
-0.0281238537,
-0.0415548421,
-0.0400802605,
-0.0062147742,
-0.0277864095,
0.0078147231,
-0.0390317999,
-0.020863004,
-0.0574504025,
-0.0023193681,
0.0342227109,
-0.0641413704,
0.0507301055,
-0.0164586604,
-0.009056014,
0.0385017879,
0.0194527972,
-0.0275881216,
0.0628688633,
0.0169984009,
0.1016543433,
0.0479741581,
-0.0025912735,
-0.0092803501,
-0.0142401187,
0.0800884292,
0.021818554,
0.0169759076,
0.0041398993,
-0.0137712657,
-0.0421130769,
-0.0037269967,
0.0456008874,
0.0168674942,
-0.0702877715,
0.065687567,
-0.0125041027,
-0.0339489318,
-0.0234762374,
0.0389690772,
0.0416547731,
-0.0034905863,
-0.0433497168,
-0.0603220426,
-0.0016674194,
-0.083395198,
-0.0389391594,
-0.0354707949,
0.0438501164,
0.0059431898,
-0.0165020023,
-0.0090572732,
-0.0093880119,
0.0215552058,
0.0078437598,
0.0059773941,
0.0105532212,
-0.0340394676,
0.0687777027,
0.0697856843,
0.0269501358,
-0.0133308051,
-0.0088444417,
0.0225226972,
0.0288176797,
0.020560218,
-0.0159347672,
0.0092414329,
-0.047792241,
0.0049673175,
0.011364826,
0.0318191983,
-0.0172710754,
-0.0104792165,
0.0355288349,
-0.0136824055,
-0.028778594,
0.0213404316,
0.0213216152,
0.0193335172,
0.0580628887,
0.0202676561,
-0.0315440781,
-0.0545816533,
0.034756653,
0.0240975227,
-0.0174142495,
0.0056043332,
0.035910964,
0.0396793,
0.0092180148,
0.0009859448,
-0.0259706769,
0.0485150702,
-0.0247463211,
-0.0519577265,
-0.0415634848,
-0.0109672667,
-0.0076369909,
0.0193323772,
-0.0268974826,
-0.0201228634,
0.0515237302,
-0.0553273149,
0.0298812315,
0.0636157691,
-0.0627704561,
-0.0124667687,
-0.019663468,
-0.0345379747,
-0.0088650128,
0.0477585346,
0.0233725272,
0.0744049549,
-0.0360519662,
-0.039978236,
0.0127199609,
-0.0397593118,
-0.0024747411,
0.0122204842,
0.0473413691,
-0.0161171127,
-0.0131939566,
0.0024259819,
0.0139673231,
-0.0186720248,
0.0327433646,
0.0118988361,
-0.0274158884,
-0.058679387,
-0.0291447956,
-0.0000792493,
0.0441932306,
-0.0232755858,
0.0202602241,
0.0120487623,
-0.0299046859,
0.0086732553,
-0.0020350073,
-0.0057687713,
0.0157704558,
-0.0171843562,
-0.0029047525,
-0.0069471137,
0.0136043811,
-0.0233289339,
0.0416204147,
0.0237689987,
-0.0729255825,
-0.0311378986,
-0.0141343316,
-0.0288666934,
0.0344619378,
-0.0287220161,
-0.0021414733,
-0.0254632495,
0.0295373723,
-0.0558668636,
-0.0114813242,
0.0464749336,
0.0539108887,
0.0024951098,
0.0355692282,
-0.0259648282,
-0.0162905119,
-0.0408298336,
0.0604385138,
-0.04792732,
-0.0606164597,
0.0625001043,
0.0464477502,
0.0586471744,
-0.0170392673,
-0.0428503044,
0.0458601825,
-0.0123083843,
-0.0512718037,
-0.0231030751,
-0.0449855961,
0.0092425225,
0.0079687936,
-0.0414503478,
-0.0550832674,
0.1005739719,
-0.0290000122,
0.0715359747,
-0.0200792439,
-0.056599997,
-0.0110633783,
0.0048206025,
-0.0094473083,
0.0381788351,
0.0537916347,
0.0045427638,
-0.0071432921,
-0.0058467053,
0.0409371108,
-0.0143105956,
0.0023553807,
-0.048504971,
0.0430704206,
0.0362144336,
-0.0332142748,
-0.0535738394,
-0.1067228839,
0.0170233026,
-0.0135412756,
-0.0054510301,
-0.0231719427,
-0.0728423595,
-0.0698428378,
-0.0006880827,
-0.0182013866,
0.0088198325,
0.0038002741,
0.0088083548,
0.0099481046,
0.0476450957,
-0.0135884071,
0.0619532838,
-0.0281480402,
-0.0492690876,
0.0517816544,
-0.043968305,
0.0078714993,
-0.0053956183,
-0.0475755334,
-0.021096075,
-0.0441225916,
0.0134194214,
0.0092687272,
-0.0031156829,
-0.0419079885,
0.0023378883,
0.0738149136,
-0.0064927898,
0.0020556706,
0.022395974,
0.0314327218,
-0.041182965,
0.0016995731,
0.0031533395,
-0.0787364244,
-0.0162911378,
-0.0107345935,
-0.0061720195,
-0.0303853825,
-0.0141680995,
-0.0336890072,
-0.0638623759,
-0.0103905518,
0.0085952785,
-0.0589543395,
-0.086137943,
-0.0212622788,
-0.0207031779,
0.0027247537,
0.0137904976,
-0.0420460477,
0.0194705352,
-0.0347740203,
-0.0098882951,
-0.037844561,
-0.0675088689,
0.0140335858,
0.0261777733,
0.0316470414,
0.0225768518,
0.0506391302,
-0.0085878018,
-0.0495984219,
-0.0066945641,
0.0208813976,
-0.0307724942,
-0.0049073729,
0.0800061673,
-0.0492695384,
-0.0303180236,
-0.0052692736,
0.023367947,
-0.0292876922,
-0.015873326,
-0.0173629392,
0.0693010241,
0.014034125,
-0.0589397065,
-0.0506152846,
-0.006447989,
0.0871991366,
0.0696989819,
-0.0283824913,
0.0106047392,
-0.0080756228,
-0.0344893597,
-0.0073073171,
0.0225343667,
-0.0068134973,
0.0443679653,
0.0175659526,
-0.0114055034,
-0.0702938661,
-0.0120081427,
-0.0012532773,
-0.0233305674,
0.0090285838,
-0.0054179402,
0.0295027569,
-0.0317273401,
-0.0514197834,
-0.0087638488,
-0.0486657657,
0.0775065646,
0.00681719,
-0.0122433258,
-0.0400254764,
-0.0250001214,
-0.0130339237,
-0.0024444095,
-0.0948907956,
-0.0551196672,
-0.0098580066,
0.0216310322,
0.0062523275,
0.0430588014,
-0.0470315181,
-0.0096691269,
0.0685256645,
0.0395467207,
-0.0340299606,
-0.0130657507,
0.0348952375,
-0.0021404033,
0.0397648253,
0.053742148,
-0.0293311197,
0.0372385308,
0.0129680606,
-0.0417301618,
-0.0646894351,
0.0198044963,
0.0306843575,
0.0221826881,
0.0251378417,
0.0312663987,
0.0053038797,
-0.0347947478,
-0.0095345117,
-0.0176687017,
-0.0358448513,
-0.0048848656,
0.013464082,
0.0423368774,
-0.0079034297,
-0.0085818442,
0.0089723133,
-0.0127782952,
0.000852261,
-0.0173542649,
-0.0407082587,
0.0220897682,
0.0450554304,
0.0272844154,
-0.0009677889,
-0.010982384,
0.0159581583,
0.0092262682,
0.0266608205,
-0.0003397756,
-0.025341915,
0.0098773465,
-0.0180342589,
-0.0092598498,
-0.0480953865,
-0.0039998549,
-0.0079688542,
-0.0267060623,
-0.0134885311,
0.0128989667,
0.005592477,
0.0018896308,
-0.0183556918,
-0.0224338062,
-0.0100015523,
-0.0075924532,
0.008708789,
-0.0071841204,
0.0260445382,
0.0067066513,
-0.031200584,
0.0230994131,
0.0760001093,
-0.0099408478,
0.0321727358,
0.014363572,
-0.0677709281,
-0.0311969854,
0.0077067129,
0.0267467033,
0.0477039777,
-0.0080547025,
0.0220764503,
-0.0859788805,
-0.0650528595,
0.0596969053,
-0.0196089614,
0.0186466724,
0.0308151543,
-0.0091723586,
0.0559926033,
0.0218634531,
0.0125557417,
-0.0787875652,
-0.0218763612,
0.0086737508,
-0.0194070451,
-0.0110291624,
-5.860644552e-33,
0.0116792163,
-0.0432430916,
-0.0360713862,
0.0269891135,
0.0171189774,
0.0328428447,
0.0078591025,
-0.055042211,
-0.0225079525,
0.0725921392,
0.0488226041,
0.039369788,
-0.0206961799,
0.0210229401,
-0.0586730242,
-0.0261950362,
0.0292109959,
0.0200968403,
-0.0139062749,
0.0043856902,
0.0954766273,
0.0128284562,
0.0452650301,
-0.0951145589,
0.0458347723,
0.0119848205,
-0.0253276322,
-0.002653454,
-0.0165741537,
0.0284369253,
0.0198468678,
0.0742155761,
0.0349902362,
-0.0183406174,
-0.0009958281,
0.0240571816,
-0.0240530465,
-0.0338675939,
0.0323669799,
-0.0153887691,
-0.0018427743,
-0.0003109237,
0.0295607168,
-0.0111566279,
0.0072736745,
-0.0156155499,
0.026464887,
-0.0201821439,
0.0017927049,
-0.0435771793,
-0.0747032315,
-0.000399275,
0.0027144267,
0.050546363,
-0.0292253159,
0.0634661019,
0.0113581996,
0.0209560543,
-0.0811521336,
-0.0012915995,
-0.0310526099,
-0.0216282625,
0.019106254,
-0.0502877533,
-0.0263289679,
0.0801040977,
0.005630062,
0.0190388877,
-0.050286077,
0.0465843193,
-0.0113331759,
0.0334674083,
0.0697610527,
0.082895346,
0.0230034031,
-0.0515819341,
-0.0932697952,
-0.0010058556,
0.0172683746,
0.0029755812,
0.0125262933,
-0.0389371738,
-0.0546843112,
0.0037238144,
-0.0246542357,
0.0229855087,
-0.0326199271,
-0.0174117666,
0.0079594878,
0.0234088916,
-0.0294485018,
0.0394122116,
-0.0273566134,
-0.0815216973,
-0.0127815343,
0.0451392867,
0.0105727073,
-0.0331395976,
0.0329734609,
0.0291437395,
-0.0742728263,
-0.0097845299,
0.0420077592,
-0.0334814414,
0.0191268586,
-0.0201565381,
0.0907319337,
0.019908661,
-0.0315673687,
-0.0035416195,
-0.0215047076,
0.0413875692,
0.0237274468,
-0.0012927206,
0.0364151299,
0.0248605777,
-0.0283619575,
-0.0077242567,
-0.0134040313,
0.0097653642,
0.0344154686,
0.0449226312,
-0.0140232705,
0.0124731679,
-0.0318445452,
-0.0318321772,
-0.0265303906,
0.0026773263,
0.0205718484,
-0.0840699226,
0.0053095804,
0.0124377115,
2.639e-7,
0.02164823,
0.0537404977,
-0.0365483873,
0.0076070367,
0.0272234157,
0.0449537523,
0.0202463493,
0.0250403807,
-0.0356010348,
-0.0851778984,
-0.0286498331,
0.0417060219,
0.015282521,
0.0403481312,
-0.0404126868,
0.0043633766,
-0.0276396219,
-0.0360649116,
-0.0322794989,
0.0098729553,
0.027631335,
0.1023675501,
0.0567399338,
-0.0473193936,
-0.0066729067,
0.0099057499,
0.0488913283,
0.0145423515,
-0.0173856281,
0.0004846653,
0.0278478507,
-0.017615987,
-0.0026309032,
0.0329599902,
-0.0304127727,
-0.0016820073,
-0.0261548422,
0.0106436536,
-0.0031661072,
-0.002699225,
-0.0533815511,
-0.0360500813,
0.0119435284,
-0.0171643049,
0.0626431182,
-0.0549866073,
-0.0280902628,
-0.0623664446,
-0.0318799876,
0.0249730907,
0.0180960614,
0.0136538129,
-0.0005884848,
0.0263337288,
0.0034485098,
0.0209184065,
0.0072987163,
-0.0459369645,
0.0553223528,
0.0321444273,
-0.0339486115,
0.0217932668,
0.0101132449,
-0.004765186,
0.0724753812,
-0.0377900377,
0.0559138991,
2.100408892e-34,
0.0327388011,
-0.0418927334,
-0.013167249,
-0.0140051767,
0.027895838,
0.0050638826,
-0.0574861243,
-0.0223222971,
0.0092943041,
-0.0046726507,
0.01899557
] |
http://arxiv.org/abs/2401.16282v1 | http://arxiv.org/pdf/2401.16282v1 | MAPLE: Micro Analysis of Pairwise Language Evolution for Few-Shot Claim Verification | 2024-01-29T16:39:39 | 2024-01-29T16:39:39 | Xia Zeng, Arkaitz Zubiaga | cs.CL | cs.CL, cs.AI, cs.LG | Claim verification is an essential step in the automated fact-checking
pipeline which assesses the veracity of a claim against a piece of evidence. In
this work, we explore the potential of few-shot claim verification, where only
very limited data is available for supervision. We propose MAPLE (Micro
Analysis of Pairwise Language Evolution), a pioneering approach that explores
the alignment between a claim and its evidence with a small seq2seq model and a
novel semantic measure. Its innovative utilization of micro language evolution
path leverages unlabelled pairwise data to facilitate claim verification while
imposing low demand on data annotations and computing resources. MAPLE
demonstrates significant performance improvements over SOTA baselines SEED, PET
and LLaMA 2 across three fact-checking datasets: FEVER, Climate FEVER, and
SciFact. Data and code are available here: https://github.com/XiaZeng0223/MAPLE | MAPLE: Micro Analysis of Pairwise Language Evolution for Few-Shot Claim Verification: Claim verification is an essential step in the automated fact-checking
pipeline which assesses the veracity of a claim against a piece of evidence. In
this work, we explore the potential of few-shot claim verification, where only
very limited data is available for supervision. We propose MAPLE (Micro
Analysis of Pairwise Language Evolution), a pioneering approach that explores
the alignment between a claim and its evidence with a small seq2seq model and a
novel semantic measure. Its innovative utilization of micro language evolution
path leverages unlabelled pairwise data to facilitate claim verification while
imposing low demand on data annotations and computing resources. MAPLE
demonstrates significant performance improvements over SOTA baselines SEED, PET
and LLaMA 2 across three fact-checking datasets: FEVER, Climate FEVER, and
SciFact. Data and code are available here: https://github.com/XiaZeng0223/MAPLE | [
0.0048382943,
0.0835341737,
0.0074228169,
0.0335491113,
-0.0488886312,
0.0188874416,
-0.0023629719,
-0.0077180308,
-0.0230661742,
-0.0459122434,
0.0216953177,
0.0210158508,
0.0243244786,
0.0528116226,
0.0251989365,
-0.0094403056,
0.0789947286,
0.0067490353,
0.0875068381,
-0.0111554284,
-0.0018073618,
-0.0000135695,
-0.0288210586,
0.0055630137,
-0.0191396363,
-0.022840431,
0.0105376374,
-0.0035905265,
0.0675153136,
-0.0570913255,
-0.0055129402,
0.0431128778,
0.0118805505,
0.1056627184,
0.0000016168,
-0.0781007111,
0.004787636,
0.0645796433,
-0.0147863114,
0.0254044682,
-0.002549283,
0.0319854803,
-0.0206184201,
0.0080779446,
-0.0151965348,
-0.0533616804,
0.0766039789,
-0.0254634339,
0.039082855,
0.068206057,
-0.0103075141,
-0.0407938957,
0.0131030157,
0.0083971182,
0.0078182193,
0.0152884647,
0.069421716,
-0.0103724599,
-0.0006811074,
-0.0022099162,
0.0405628309,
0.0453982241,
-0.0179007053,
-0.0243565235,
-0.0750205889,
0.0645656735,
-0.0566724464,
-0.0162236132,
0.014326768,
0.0072225202,
0.0021997867,
-0.0623826273,
-0.0029029197,
-0.0065285605,
0.0148168122,
0.021479167,
-0.0100766374,
0.0317866839,
-0.0303732734,
-0.0039088493,
0.0673876256,
0.0090276012,
0.0206048116,
0.0134822372,
-0.0038540508,
0.0546029545,
0.0085211275,
-0.0634633228,
-0.0139957918,
-0.0282497499,
0.0187244304,
-0.0611643046,
-0.0081921555,
0.0103886863,
-0.0175939165,
0.018827524,
0.009409735,
-0.0047007604,
0.0829766244,
-0.0727133825,
0.0128962332,
0.0036447162,
-0.0367534123,
0.0655144826,
-0.0321575888,
-0.0244901162,
0.028796846,
-0.025838729,
0.0239075869,
-0.0300370157,
-0.0251094066,
-0.0135330064,
-0.0232166238,
-0.0105174063,
-0.0314245895,
0.0348160714,
-0.0239287484,
-0.0286506023,
0.0096645262,
0.0315201357,
-0.0112031028,
-0.0248451978,
-0.0236368142,
0.0305322725,
-0.0163973514,
0.0149358204,
-0.0288018417,
0.028422432,
-0.0019687633,
-0.0438252799,
-0.0301660486,
0.0443890207,
0.0204295609,
-0.0124891289,
0.010476131,
-0.0045993654,
0.0408073217,
-0.0115139578,
-0.0166615471,
-0.0378887616,
-0.0585563853,
-0.0604612269,
-0.0250654072,
-0.0732531473,
0.0049976897,
-0.017542582,
-0.0161607433,
0.0060401666,
-0.0394212678,
-0.0068858853,
-0.0440060347,
0.0259107202,
-0.09412916,
-0.0211779717,
0.0095990207,
-0.019377064,
0.035508059,
0.0618749559,
0.0348418467,
0.027169127,
0.0456803702,
0.0078082629,
-0.0142732169,
-0.007681868,
0.0459735543,
0.0570802949,
0.0148956003,
0.0016769824,
0.0263143182,
0.0093915816,
-0.0207443684,
0.0347913653,
-0.0200477783,
-0.0282616392,
0.0461025462,
0.0469992943,
0.0065222182,
-0.0280342512,
0.0584012344,
-0.0276163314,
0.0092809489,
-0.0170440041,
-0.0159612242,
0.0219951905,
-0.0566089265,
-0.0243031755,
0.034286052,
0.0386698693,
0.0056934454,
-0.0353202783,
-0.0267836489,
0.034811534,
-0.0636824444,
-0.025647724,
-0.017104337,
0.0246727709,
0.0206338149,
0.0144430902,
0.055504933,
-0.046575509,
-0.0393254757,
-0.0553434044,
0.0525381789,
-0.0232467689,
0.0182823855,
-0.0098244045,
-0.0211213268,
0.0001660476,
-0.0198598802,
0.0430337675,
0.0547673479,
-0.0070171617,
0.0031097278,
0.020502951,
0.0069865379,
-0.0443699397,
0.0518002957,
-0.0224661417,
0.0299698636,
0.0220601745,
-0.0074223359,
-0.0088611394,
-0.0327858031,
0.0248413347,
-0.0017719478,
-0.0043580388,
0.0197448079,
0.0040167347,
0.0625747666,
-0.0142939333,
-0.0163877942,
0.0036139383,
0.0403944328,
0.0197649449,
-0.0140416948,
0.0604488887,
0.0708361194,
-0.0711327046,
0.0519780777,
-0.0142637473,
-0.0286654625,
-0.0299173314,
-0.0309197735,
-0.0093322974,
-0.0255977996,
-0.0497562177,
-0.0139680766,
-0.0343944468,
-0.0193633679,
-0.0332100578,
0.0105587607,
-0.0077543715,
0.056979347,
0.0071761929,
-0.0007874906,
-0.0476616248,
-0.0068603246,
0.0386599116,
-0.0218042005,
-0.0227687284,
-0.0219514426,
0.0036277531,
-0.0444774814,
0.0286721326,
-0.0497018658,
-0.0323345177,
0.0118349651,
-0.0468612611,
0.0300475135,
0.0146404933,
0.0583396517,
0.0226213038,
-0.0287000183,
-0.0231365431,
0.0353911072,
-0.022346247,
-0.0100539466,
0.0069132908,
0.0305449404,
0.0801941529,
-0.0180607028,
-0.0223882459,
-0.0415713117,
0.0265418924,
-0.0540364049,
0.0425646082,
0.0312988795,
-0.0239788759,
-0.0390576534,
-0.0199299604,
-0.0591747835,
0.0019918233,
-0.0021946537,
-0.0202765614,
0.0294900257,
0.0187766198,
0.0024378968,
0.0216525253,
-0.0079956958,
-0.0172235332,
0.0479113646,
0.0020508585,
-0.0056886137,
-0.003237477,
0.0001573698,
0.0168264676,
0.018217802,
-0.0339604057,
0.0222582091,
0.0471329615,
-0.0249727424,
-0.0269080102,
-0.0196019076,
0.0565995052,
0.0053810277,
-0.0277274549,
0.0591578782,
0.0007011571,
-0.0148898503,
0.032910306,
-0.0311732758,
-0.0693717524,
0.0818599463,
-0.0425028801,
0.0231782161,
-0.0733282715,
-0.0216018353,
-0.0212893821,
0.0533747748,
-0.0208908338,
0.0132870236,
0.0122215897,
0.0382587537,
0.0026381176,
-0.0506005324,
0.0461576395,
0.0019092659,
-0.0163723752,
-0.0535552837,
0.047338292,
0.0054347883,
-0.008214009,
0.0104332669,
-0.04129925,
-0.0340052284,
0.0105229532,
-0.0337655731,
-0.0280395392,
-0.0902722031,
0.0084153172,
0.0564266182,
0.0426826663,
0.0089134108,
-0.0049587875,
-0.0049750847,
-0.0104717566,
0.0327132791,
0.008385133,
0.0300949998,
0.0124909701,
0.0325833894,
0.0670104325,
-0.036510136,
0.0606827177,
-0.0680636689,
-0.0454304144,
-0.0498314835,
-0.0248214733,
0.0145376828,
-0.0100376895,
0.0480037816,
-0.048410207,
-0.0036591033,
0.05739066,
0.0207903571,
0.0389399752,
0.0356174186,
-0.0137613872,
0.0006699052,
0.0206403602,
0.0245449096,
-0.1160520241,
0.0036283587,
-0.0213404372,
-0.0365235284,
-0.0454940088,
-0.0379444025,
-0.0328882635,
-0.0504435077,
0.0628012121,
0.0028934029,
-0.0462769903,
0.0001682256,
-0.10709133,
0.0224380977,
-0.002077261,
-0.0156231467,
0.0111539345,
0.0222789664,
-0.0032792646,
-0.004593112,
0.0023172798,
-0.0294475388,
0.0594598018,
-0.0476394966,
0.0441809632,
0.0096420432,
0.0455324017,
-0.0191937257,
-0.0392883196,
0.0042518317,
-0.0130681293,
-0.0432094708,
-0.0011020111,
0.0414542221,
-0.0037118448,
0.015367534,
-0.0125653502,
0.0198975168,
-0.0043483684,
-0.008608955,
0.0179442782,
0.0855982602,
-0.0067803068,
-0.066238597,
-0.0327242762,
-0.0027817721,
0.0046671783,
0.0652406514,
0.0163762532,
0.0453774557,
0.0058359341,
0.0005517491,
-0.054817412,
-0.0245699733,
-0.0345226787,
0.0401360095,
-0.0214878153,
-0.0644128919,
-0.0689603165,
0.0208676904,
0.0156740379,
0.0461661108,
-0.0401796289,
-0.0091871126,
0.0018914595,
-0.0142134381,
-0.0491397642,
0.0165949035,
-0.0367611013,
0.039947398,
0.0019648371,
0.0128127765,
0.0631957129,
0.0774003044,
-0.0214764792,
0.0479294397,
-0.056008175,
-0.0649683475,
-0.0373142511,
0.0291215945,
0.023410948,
0.0060249362,
0.0027451026,
-0.0275109354,
0.0480181985,
0.1012880579,
0.0011752237,
0.011612392,
-0.0074991658,
-0.0436095558,
-0.000077457,
0.0047927471,
0.0120895514,
0.0369574614,
0.0030576487,
0.0244779699,
-0.0390412509,
-0.0251513887,
0.0315529108,
0.0514960736,
0.0880278051,
-0.0276991129,
-0.0668239668,
-0.0014882794,
0.0143299578,
-0.0245575346,
-0.0360478163,
-0.005716464,
0.0043191845,
-0.053975679,
-0.0190554373,
-0.0186177827,
-0.0069797523,
-0.0515959337,
0.0142888511,
-0.0447573662,
0.0191736966,
0.0037876519,
0.0591410249,
0.0106861629,
-0.0100986995,
-0.0175044984,
-0.0071201543,
-0.0122152809,
0.0270912778,
0.0085099172,
-0.0291387346,
0.0545797534,
0.0165951159,
0.0057400134,
-0.0113704568,
-0.0331829302,
0.0153855113,
-0.0131555684,
0.0149487872,
0.0128408996,
0.0188979097,
0.0173806399,
-0.007172592,
-0.0297979545,
-0.0344381705,
0.0147100268,
-0.035693232,
-0.0241046492,
0.0263067018,
0.0137545224,
-0.0532449894,
0.0086360164,
0.0273375176,
0.0102123246,
0.0738017187,
0.0001793015,
-0.0168803986,
-0.0243421402,
-0.0602567196,
-0.0615309551,
-0.008168594,
-0.0200105403,
-0.0246851705,
-0.0861845538,
-0.0460951515,
-0.001169814,
0.056180045,
0.0211488139,
0.0568399094,
-0.0210658051,
0.0607907064,
0.006549113,
-0.0550128482,
-0.0040374571,
0.025391927,
0.0133393006,
-0.0373618379,
0.0325564183,
-6.084606716e-33,
0.0026575113,
-0.0564873442,
0.0109869903,
0.0328087807,
0.0236023385,
0.039087981,
-0.0473074578,
-0.022890551,
-0.0268561598,
0.0648088604,
0.047109928,
0.0150160138,
0.0080570141,
-0.0519601218,
0.023458153,
0.0244301874,
0.0473233983,
0.017423939,
-0.0378625132,
-0.0026004834,
0.0374061726,
-0.0379883125,
0.0548109077,
-0.071441412,
0.0384202115,
0.0410959907,
-0.0242207367,
0.0210184809,
-0.0932786018,
0.049488239,
0.0274604149,
0.0122329211,
0.0017451708,
0.006058835,
0.0261554662,
-0.0499368273,
-0.0421983413,
-0.0194490608,
0.054234948,
-0.0605726317,
-0.0475810096,
0.0025409551,
0.005601068,
0.0333780646,
-0.0322892256,
0.0078923618,
0.0176584609,
-0.0423064604,
0.0036952905,
-0.0612966605,
-0.0550956465,
0.0718392506,
0.0074000903,
0.0087013617,
-0.0224602818,
-0.026719138,
0.0180746429,
0.027226653,
-0.0381919183,
0.0468101948,
0.0025342442,
-0.0146373017,
0.031132428,
-0.0512577146,
-0.013989252,
0.056298729,
0.0254834536,
0.0257889368,
0.0238194298,
0.0158412438,
0.0139000351,
0.0295398086,
0.0430280678,
0.0627787039,
0.0338165946,
0.015837783,
-0.1048834175,
0.0451577753,
0.0638250411,
0.0326898359,
-0.0177738946,
-0.0068273996,
-0.0056805178,
0.0153990826,
-0.0573963784,
0.0004172841,
-0.0368094034,
0.0439364798,
-0.0316683426,
-0.0021315652,
-0.0306826364,
0.0819222555,
0.0115965679,
-0.0093557285,
-0.0044790423,
-0.0358255133,
-0.0178689118,
-0.0054140263,
0.0061737937,
0.0428452827,
-0.062812902,
-0.0079679228,
0.0176805276,
0.0419494398,
0.0430921391,
0.0097820144,
0.0710508898,
-0.0146428533,
-0.0885231867,
0.0342266373,
-0.0183288436,
-0.0133451466,
-0.0070227184,
-0.052977439,
0.0057581686,
0.000173035,
0.038023822,
0.0034872091,
0.0318635404,
0.0034291847,
-0.0269028693,
0.0623646975,
-0.0435238965,
0.0091248313,
-0.0158267319,
-0.0136818802,
0.0130869485,
-0.0139497602,
-0.0007958972,
-0.0705535784,
0.0274497196,
0.0252083112,
2.57e-7,
0.041161608,
0.0733811855,
0.0105026681,
-0.0210398976,
0.0221584141,
0.0163940266,
-0.0241953898,
0.0022693388,
0.0042029223,
-0.0692057759,
-0.0335469134,
-0.000780562,
0.0121631958,
0.0611047149,
-0.0449465737,
0.0230936557,
-0.0895995274,
-0.0243021585,
-0.0771716088,
-0.0011815599,
-0.0007506367,
0.0448367409,
0.0477827825,
0.0037471964,
0.0057481616,
-0.0062720813,
0.0451220647,
0.0490074567,
-0.0183961224,
0.0367650762,
0.0176891834,
-0.0273289625,
0.026402222,
0.0338550434,
-0.0441322736,
0.0327646919,
-0.070174329,
0.0381614715,
-0.0091297859,
0.0403521992,
-0.0060592704,
-0.0324633047,
0.0212864596,
-0.0594569296,
0.0410721712,
-0.0423961431,
-0.0420302078,
-0.0654721111,
-0.0349231921,
0.0547027849,
0.0203892365,
0.0032794587,
-0.023482766,
0.0298108868,
-0.0316357724,
0.0263572093,
-0.0356974714,
-0.0348026194,
-0.0203844681,
0.010044598,
-0.0224745963,
-0.0646794438,
0.0347485058,
-0.0276964307,
0.020407591,
-0.0536231995,
0.0555885658,
2.180011212e-34,
0.0469038561,
-0.013405744,
-0.0147779761,
-0.0605818331,
0.0308353975,
-0.0035408274,
-0.0055384268,
0.0197437108,
0.0366115682,
-0.0292475745,
-0.0011724524
] |
http://arxiv.org/abs/2402.01742v1 | http://arxiv.org/pdf/2402.01742v1 | Towards Optimizing the Costs of LLM Usage | 2024-01-29T16:36:31 | 2024-01-29T16:36:31 | Shivanshu Shekhar, Tanishq Dubey, Koyel Mukherjee, Apoorv Saxena, Atharv Tyagi, Nishanth Kotla | cs.CL | cs.CL, cs.AI, cs.LG | Generative AI and LLMs in particular are heavily used nowadays for various
document processing tasks such as question answering and summarization.
However, different LLMs come with different capabilities for different tasks as
well as with different costs, tokenization, and latency. In fact, enterprises
are already incurring huge costs of operating or using LLMs for their
respective use cases.
In this work, we propose optimizing the usage costs of LLMs by estimating
their output quality (without actually invoking the LLMs), and then solving an
optimization routine for the LLM selection to either keep costs under a budget,
or minimize the costs, in a quality and latency aware manner. We propose a
model to predict the output quality of LLMs on document processing tasks like
summarization, followed by an LP rounding algorithm to optimize the selection
of LLMs. We study optimization problems trading off the quality and costs, both
theoretically and empirically. We further propose a sentence simplification
model for reducing the number of tokens in a controlled manner. Additionally,
we propose several deterministic heuristics for reducing tokens in a quality
aware manner, and study the related optimization problem of applying the
heuristics optimizing the quality and cost trade-off. We perform extensive
empirical validation of our methods on not only enterprise datasets but also on
open-source datasets, annotated by us, and show that we perform much better
compared to closest baselines. Our methods reduce costs by 40%- 90% while
improving quality by 4%-7%. We will release the annotated open source datasets
to the community for further research and exploration. | Towards Optimizing the Costs of LLM Usage: Generative AI and LLMs in particular are heavily used nowadays for various
document processing tasks such as question answering and summarization.
However, different LLMs come with different capabilities for different tasks as
well as with different costs, tokenization, and latency. In fact, enterprises
are already incurring huge costs of operating or using LLMs for their
respective use cases.
In this work, we propose optimizing the usage costs of LLMs by estimating
their output quality (without actually invoking the LLMs), and then solving an
optimization routine for the LLM selection to either keep costs under a budget,
or minimize the costs, in a quality and latency aware manner. We propose a
model to predict the output quality of LLMs on document processing tasks like
summarization, followed by an LP rounding algorithm to optimize the selection
of LLMs. We study optimization problems trading off the quality and costs, both
theoretically and empirically. We further propose a sentence simplification
model for reducing the number of tokens in a controlled manner. Additionally,
we propose several deterministic heuristics for reducing tokens in a quality
aware manner, and study the related optimization problem of applying the
heuristics optimizing the quality and cost trade-off. We perform extensive
empirical validation of our methods on not only enterprise datasets but also on
open-source datasets, annotated by us, and show that we perform much better
compared to closest baselines. Our methods reduce costs by 40%- 90% while
improving quality by 4%-7%. We will release the annotated open source datasets
to the community for further research and exploration. | [
0.039317932,
0.0288127344,
0.0179859102,
0.0311197098,
-0.0341662653,
-0.0051442147,
-0.0032088552,
0.0257791765,
-0.0096764769,
-0.0288496315,
0.0040248339,
0.0124630369,
0.0008926389,
-0.001068228,
0.0246278979,
-0.0451376773,
0.0405555442,
-0.0010801534,
0.012160996,
-0.0159328841,
0.0171608292,
0.0045721447,
0.0044470327,
0.0234800447,
-0.019428622,
-0.016853705,
0.0080372738,
0.0042175809,
0.01903891,
0.0146089066,
-0.0141044157,
0.0014133938,
-0.0152944569,
-0.0002393272,
0.0000020295,
-0.0640229881,
-0.0325238183,
0.0006127154,
-0.0244095773,
0.0049261381,
0.0601333268,
0.0128286518,
-0.0023252734,
0.061237067,
-0.0338143632,
0.0284546632,
0.0053324136,
0.0273019932,
0.0674715415,
0.0827848166,
0.0012350265,
-0.0080673909,
0.0156432502,
-0.016126398,
-0.0023036574,
-0.0438170843,
-0.0002933466,
-0.0008866134,
0.0005925713,
-0.0347372852,
-0.0152273243,
0.0379482768,
0.0250072684,
-0.0365396813,
0.0154120345,
0.0207450297,
-0.0436934344,
0.0216691494,
0.0021005182,
0.069332622,
-0.0101336306,
-0.0184181631,
0.0352219269,
-0.0089544067,
0.0297905244,
-0.0163157675,
-0.0582927205,
0.0388625748,
-0.0102035273,
0.0242721401,
-0.0016810044,
-0.0189388283,
-0.0024405231,
0.0354222842,
-0.0156747028,
0.0309614334,
0.0037102667,
-0.0369328596,
-0.0244913083,
0.0119752781,
-0.0409099311,
-0.0685770735,
0.021745421,
0.0567141175,
-0.0336320363,
0.0097550387,
-0.0669732541,
-0.0687065795,
0.0177626349,
0.0032422768,
-0.0020923405,
0.0217764284,
0.0181887411,
0.0506365784,
-0.0577324256,
0.0335040018,
0.0007596569,
0.0183029007,
-0.0527960062,
-0.0116269123,
-0.0057891235,
-0.0074839382,
-0.0391479135,
0.091696851,
-0.0428786241,
-0.0127732661,
-0.0304114521,
0.049014952,
0.0251088943,
0.0127035193,
0.0362084508,
0.0172433797,
0.0130319716,
-0.0101987505,
-0.0398950391,
-0.0058175703,
-0.0482063927,
0.0255032852,
0.0077116014,
-0.0414719805,
-0.0161314402,
0.0534659624,
0.028770715,
0.0199411195,
0.0148105081,
0.0695273802,
-0.0122561846,
-0.0088112177,
-0.0635615736,
-0.0376033038,
0.0240568928,
-0.0366874188,
-0.0274996012,
-0.0741948038,
0.0023722872,
0.0040579536,
-0.0234032366,
-0.0046294737,
-0.0196709577,
0.0238561947,
-0.0212321971,
0.0320286639,
-0.0336661078,
0.0328812487,
0.0414547734,
-0.0133460341,
-0.0000319693,
0.0486528873,
0.0622885153,
0.0410434939,
0.0678828582,
0.0134204486,
0.0251912884,
0.0412297659,
0.0134883374,
0.0031987424,
0.0281821247,
-0.0010899894,
-0.061299108,
0.0266125835,
0.0713726506,
0.0670808107,
-0.0391581431,
0.0203041472,
0.0579632483,
0.0826912746,
0.0167221092,
0.0032379306,
0.0409415439,
0.0170497093,
0.0194664933,
0.0224963371,
-0.0256385393,
0.0466187149,
-0.0154233174,
0.0485906079,
0.078449294,
-0.0039242688,
-0.0326406695,
-0.0917690545,
0.0082169138,
-0.004034196,
-0.0440852493,
0.0104270587,
0.0046285638,
0.0044768201,
-0.0807448849,
0.0493811369,
-0.020804869,
-0.0551464707,
-0.0166956298,
-0.0814812332,
0.0450325422,
0.0776767582,
0.0690708309,
-0.0262634736,
-0.0045798216,
-0.0198880211,
-0.0073604477,
0.036801178,
0.0305170752,
-0.0121334381,
-0.0144324843,
-0.0655466914,
-0.0195168871,
-0.0220396686,
-0.0056798616,
-0.0023815285,
0.0279810838,
0.0130236913,
0.0252624806,
-0.0788030103,
-0.0053344574,
0.0343916044,
0.0020548631,
-0.0455564596,
0.0291035064,
0.0032672125,
0.0359901041,
0.0317165963,
-0.0030298603,
0.0335633866,
0.011594668,
-0.0339114591,
-0.0334532186,
-0.0072628576,
-0.0193015356,
-0.0261689015,
0.0406437963,
-0.0543182343,
-0.0186732151,
0.0238492824,
-0.0802505463,
-0.055695083,
0.0018237879,
-0.0706048906,
0.0165910069,
-0.0586909056,
0.0187262055,
0.0186312832,
0.0062526255,
-0.0329442695,
0.0677342266,
-0.0211007744,
0.0215084031,
0.0374545902,
0.0027216908,
0.019498542,
-0.0562398247,
-0.0010507967,
-0.0297397245,
-0.0070841559,
0.0261645224,
0.0402903445,
0.0258123912,
-0.0376770198,
-0.0283473935,
-0.1036957428,
0.024065353,
-0.0084051723,
0.0424691364,
0.027093282,
-0.0070830006,
-0.0253437385,
-0.0036130056,
-0.0588094555,
-0.0066473829,
0.0363813452,
0.0069871536,
0.0860479623,
0.0003214102,
0.0009729127,
-0.0332930945,
-0.0084190704,
0.0109004686,
0.0042743403,
-0.0025386524,
-0.0211190749,
-0.0638033152,
0.0525808819,
-0.0250017326,
-0.0130910715,
0.0231547784,
-0.0111977253,
-0.0134631842,
-0.0182373263,
0.0535406172,
0.0844488591,
0.018985929,
-0.0146332104,
0.0306607503,
0.0393680669,
-0.0887430906,
0.0136006176,
-0.00001599,
0.0155491773,
0.0187126976,
-0.0354762189,
0.0772927403,
0.0351314694,
0.0016485426,
-0.007976233,
-0.0144101754,
0.0370129831,
0.0120440302,
-0.0406626426,
-0.0106659401,
0.0323636383,
-0.0163527299,
0.0443618856,
0.0001180763,
-0.059051577,
0.0496101528,
-0.003860811,
0.0254870187,
-0.0334176533,
0.0080128247,
-0.0701235086,
-0.0043381997,
-0.0121534886,
0.0037038769,
-0.0443276316,
-0.0414287113,
-0.0636797249,
0.0051068864,
0.0410394631,
-0.0269614812,
0.0721079186,
-0.0384360813,
0.00955017,
-0.0437042378,
-0.0397696421,
-0.0120911589,
-0.0001476211,
-0.0055992659,
0.0130403005,
0.0246272516,
0.0217195507,
-0.0343006402,
-0.0061847852,
-0.0015750997,
0.0730365068,
-0.0258952789,
-0.0233218092,
-0.0171223078,
0.0593827516,
0.05211474,
0.030373387,
0.0464070365,
-0.0071609663,
0.0216473583,
0.0561186522,
0.0388494059,
-0.014295998,
0.0176518261,
-0.0298816487,
0.0216922276,
-0.0864732489,
0.000121865,
0.0080065774,
0.0255383961,
-0.0485923588,
0.0177290142,
0.1133064702,
0.0164518803,
-0.0377350859,
-0.0055439696,
-0.0122834127,
-0.0050665587,
0.0093242321,
0.0355187394,
-0.0998424664,
-0.0214785673,
0.017277658,
0.0740131438,
-0.0051248926,
-0.0325440541,
-0.0094197672,
-0.1237267554,
0.0528178662,
-0.0277158786,
-0.0158882383,
-0.0328808799,
-0.031894777,
-0.0144559527,
-0.0093409382,
0.0267289337,
0.0116387345,
-0.0196229238,
0.0159868635,
0.0055623967,
-0.0570024028,
-0.0149822515,
-0.0393268391,
-0.0036718713,
0.005569262,
0.0398271345,
0.0383378267,
-0.0001684913,
-0.02432511,
0.0614340678,
0.0029688724,
-0.0143405749,
-0.0345472209,
0.0340841077,
0.0051150522,
0.0212812517,
-0.0150282755,
0.0493109114,
-0.0026361838,
0.0133412378,
0.0361733027,
0.0300229862,
-0.0043947357,
-0.0325162224,
-0.0301912185,
0.0002323795,
-0.0342331901,
0.0879394487,
0.0462610163,
-0.0221376251,
-0.0718412101,
-0.029323101,
-0.0087473011,
-0.1022889689,
0.0098164249,
-0.0345729478,
-0.0305652861,
0.0223028082,
-0.113449432,
-0.0116551081,
0.032842733,
0.0128958616,
-0.0309721325,
-0.0050897961,
-0.0490082838,
0.0162366312,
-0.0322093815,
-0.0342706963,
-0.0417087972,
0.0042988602,
-0.080691956,
0.0099510495,
0.0164133031,
-0.0219984427,
-0.0150519665,
-0.0004783071,
-0.0968154296,
-0.0598883778,
-0.0497123189,
0.0599592477,
0.0177999269,
0.0601174496,
-0.0081180735,
-0.0310527273,
0.0725171641,
0.0763393268,
-0.0213864986,
0.0435456,
0.0126525741,
0.0510177389,
-0.0152006103,
0.0296496749,
-0.0688315704,
0.0033706853,
0.0018457243,
0.0528911985,
-0.0330656096,
-0.0488429554,
-0.0016221164,
0.0905411094,
-0.0413780361,
-0.0331043415,
0.0237056836,
-0.0148240095,
0.0221002754,
-0.0019105311,
-0.0101146875,
-0.0318832248,
0.0408583917,
-0.0323198065,
0.0044985251,
-0.0114579853,
0.0511377379,
-0.0199440867,
0.0096405493,
-0.0210155938,
0.0212499164,
-0.039942231,
-0.0250961259,
-0.0017677927,
-0.0143809291,
-0.0336913317,
-0.0350891054,
-0.0116027016,
0.0271492954,
0.005896355,
0.0059367595,
0.0192467924,
0.0536661819,
-0.0138903791,
0.0465188771,
-0.0492627248,
-0.0525817648,
-0.0163548496,
0.0618954189,
0.0419176146,
0.0098093627,
0.042845767,
-0.0154816685,
-0.0234881602,
-0.0128366528,
0.0243034102,
0.0355918407,
0.0035801029,
-0.0031548846,
0.0192687903,
-0.0151751423,
-0.0299287643,
0.091641143,
-0.037154872,
0.0283971559,
-0.0356191397,
-0.0095400838,
-0.0605813637,
-0.0461347699,
-0.0328789875,
-0.004368464,
0.0155735528,
-0.0349946134,
-0.0358739085,
-0.0595671646,
-0.0011162151,
0.0261330679,
0.0113582416,
-0.021860607,
0.0374507122,
0.051017154,
0.0302755777,
-0.0020558208,
-0.0417734459,
-0.0084925424,
-0.0230318177,
-0.0302869547,
0.0011857854,
-5.906827889e-33,
-0.0113516031,
-0.0384886898,
0.0025028826,
0.0266279466,
-0.005599895,
0.0129820844,
-0.0458103009,
-0.0282992534,
0.0049116872,
-0.017972447,
-0.0515765958,
0.0157664791,
0.0179635063,
0.0133157177,
0.0006607827,
0.0068424232,
0.0400273427,
-0.0263387486,
-0.0592404269,
-0.006688538,
0.0138568971,
0.0032431041,
0.0142134149,
-0.0058108536,
-0.0064586122,
-0.0033374105,
-0.035012126,
0.0313533843,
0.023791207,
0.0458289497,
0.0138043836,
0.0377609469,
0.038117785,
-0.0457680263,
-0.0384921208,
0.0271712542,
-0.0895071477,
-0.0490707792,
0.0307190381,
-0.0266552996,
-0.0361293517,
-0.0829136893,
0.0551267453,
0.0018373608,
-0.0114363683,
-0.0092410007,
0.0356491096,
-0.0436571836,
-0.026716508,
-0.0957583562,
-0.0148829045,
0.0211615991,
-0.021761898,
0.0133003909,
0.0036007755,
0.0012969985,
-0.0055048936,
0.0258809924,
-0.0394832231,
0.0144280037,
0.0228104107,
0.041445408,
0.0693832859,
-0.0206473432,
0.0054956749,
0.0049685105,
-0.0267724898,
-0.014414479,
-0.0351492353,
0.0029103383,
-0.0006153589,
0.0064426283,
0.0515453406,
0.0291192438,
0.0574969761,
-0.0061988668,
-0.0486267172,
0.0821921825,
0.0113700302,
0.0342593156,
-0.0156042725,
0.0111574745,
0.0037989675,
-0.0019741743,
-0.0129515994,
0.0478467979,
-0.0172301102,
-0.021578243,
0.0031601735,
0.0272888299,
-0.0290166363,
-0.0621857047,
0.011114954,
-0.0366886668,
0.0262916926,
-0.0038692805,
0.0110825198,
0.0263174661,
0.0197059959,
-0.0283732098,
-0.0500268377,
0.0170139931,
-0.00378029,
0.0043674666,
0.0114150513,
0.0081483694,
0.0404288471,
-0.0179728698,
-0.0730369389,
0.0457578227,
-0.0024389604,
0.0466457382,
0.0045872754,
-0.0133348433,
-0.0129019618,
-0.0174376145,
0.0163197629,
0.0047846674,
-0.0158836637,
0.021746546,
-0.0148340361,
0.0163775869,
0.0194045864,
0.0270843618,
-0.0522027873,
0.0312762633,
-0.0281685703,
0.0550068505,
0.0206275079,
-0.0395156741,
-0.0284497663,
0.0135224639,
2.65e-7,
0.0580844097,
0.022197105,
0.052728381,
0.0486189686,
-0.0109441616,
0.0075545274,
0.0362865627,
0.0042558829,
-0.016610153,
0.0259115156,
0.0426797196,
0.0231129881,
0.0527813062,
0.022006575,
-0.1400482208,
0.0083572138,
-0.0583774969,
0.006873101,
-0.0161640048,
-0.018306056,
0.0428171679,
0.0387048237,
-0.0213989839,
-0.019694427,
-0.0165733825,
-0.0431713946,
0.0269830581,
-0.0049782475,
-0.0024319652,
0.0296664853,
0.0828258246,
-0.0029189067,
0.010327274,
-0.0224236902,
-0.0241610482,
0.0007256958,
0.0353853926,
0.0719371215,
0.0243457723,
-0.0431744941,
0.0175951272,
-0.0672642216,
0.0335677974,
-0.0029505263,
0.0800770149,
-0.0062108757,
0.0035343133,
-0.0293572824,
-0.0048760721,
-0.0112308143,
-0.0137285301,
-0.0358773619,
-0.0185367111,
0.0459738933,
0.0066627478,
-0.0217708275,
0.0056019081,
-0.0750460327,
-0.0022718292,
-0.0129482225,
-0.0107564824,
-0.0612516738,
-0.0274589844,
0.0445499383,
0.0414824113,
-0.0446666554,
-0.0294537321,
2.454145219e-34,
0.0629896075,
-0.0157906935,
0.012426828,
0.058720924,
0.0139055122,
0.0006067769,
-0.0289449375,
-0.002005239,
0.0165038854,
-0.0704292506,
-0.029581666
] |
http://arxiv.org/abs/2401.16265v1 | http://arxiv.org/pdf/2401.16265v1 | CO2: Efficient Distributed Training with Full Communication-Computation Overlap | 2024-01-29T16:12:31 | 2024-01-29T16:12:31 | Weigao Sun, Zhen Qin, Weixuan Sun, Shidi Li, Dong Li, Xuyang Shen, Yu Qiao, Yiran Zhong | cs.CL | cs.CL, cs.DC | The fundamental success of large language models hinges upon the efficacious
implementation of large-scale distributed training techniques. Nevertheless,
building a vast, high-performance cluster featuring high-speed communication
interconnectivity is prohibitively costly, and accessible only to prominent
entities. In this work, we aim to lower this barrier and democratize
large-scale training with limited bandwidth clusters. We propose a new approach
called CO2 that introduces local-updating and asynchronous communication to the
distributed data-parallel training, thereby facilitating the full overlap of
COmunication with COmputation. CO2 is able to attain a high scalability even on
extensive multi-node clusters constrained by very limited communication
bandwidth. We further propose the staleness gap penalty and outer momentum
clipping techniques together with CO2 to bolster its convergence and training
stability. Besides, CO2 exhibits seamless integration with well-established
ZeRO-series optimizers which mitigate memory consumption of model states with
large model training. We also provide a mathematical proof of convergence,
accompanied by the establishment of a stringent upper bound. Furthermore, we
validate our findings through an extensive set of practical experiments
encompassing a wide range of tasks in the fields of computer vision and natural
language processing. These experiments serve to demonstrate the capabilities of
CO2 in terms of convergence, generalization, and scalability when deployed
across configurations comprising up to 128 A100 GPUs. The outcomes emphasize
the outstanding capacity of CO2 to hugely improve scalability, no matter on
clusters with 800Gbps RDMA or 80Gbps TCP/IP inter-node connections. | CO2: Efficient Distributed Training with Full Communication-Computation Overlap: The fundamental success of large language models hinges upon the efficacious
implementation of large-scale distributed training techniques. Nevertheless,
building a vast, high-performance cluster featuring high-speed communication
interconnectivity is prohibitively costly, and accessible only to prominent
entities. In this work, we aim to lower this barrier and democratize
large-scale training with limited bandwidth clusters. We propose a new approach
called CO2 that introduces local-updating and asynchronous communication to the
distributed data-parallel training, thereby facilitating the full overlap of
COmunication with COmputation. CO2 is able to attain a high scalability even on
extensive multi-node clusters constrained by very limited communication
bandwidth. We further propose the staleness gap penalty and outer momentum
clipping techniques together with CO2 to bolster its convergence and training
stability. Besides, CO2 exhibits seamless integration with well-established
ZeRO-series optimizers which mitigate memory consumption of model states with
large model training. We also provide a mathematical proof of convergence,
accompanied by the establishment of a stringent upper bound. Furthermore, we
validate our findings through an extensive set of practical experiments
encompassing a wide range of tasks in the fields of computer vision and natural
language processing. These experiments serve to demonstrate the capabilities of
CO2 in terms of convergence, generalization, and scalability when deployed
across configurations comprising up to 128 A100 GPUs. The outcomes emphasize
the outstanding capacity of CO2 to hugely improve scalability, no matter on
clusters with 800Gbps RDMA or 80Gbps TCP/IP inter-node connections. | [
0.0243728328,
0.0761960074,
-0.0008910258,
0.0616369285,
-0.0262399893,
0.0224788059,
0.0056957612,
-0.001255848,
-0.0019865499,
-0.0107196923,
-0.0110220406,
-0.0553037189,
0.0247153882,
0.018108001,
-0.0007026161,
-0.0471750572,
0.0590323694,
-0.0295740031,
-0.0344239771,
-0.042605184,
0.029470278,
0.0084800525,
0.0007805139,
0.004255991,
-0.0289296564,
-0.0489014015,
-0.0167669617,
-0.0045047253,
0.01854847,
-0.0158009231,
-0.0089140674,
0.0347649716,
0.050177183,
0.0564390123,
0.000002149,
-0.0010892682,
0.0061567859,
-0.0032909955,
-0.0220448319,
0.0390447825,
0.0094440235,
-0.0449041016,
-0.0348235779,
-0.0079274485,
-0.0135001065,
-0.0106252246,
0.0659793094,
-0.0126587665,
0.0648306161,
0.0426825881,
-0.0033558079,
-0.0377826132,
0.0356471837,
-0.0539335236,
-0.056716878,
-0.0340464599,
0.0446949005,
0.028456362,
-0.0236313902,
-0.0315972008,
-0.0405990593,
0.0290628634,
0.0280502215,
-0.0065928209,
0.0256650951,
0.0548243821,
-0.0409338474,
-0.0188211389,
-0.0516269915,
0.0341351889,
-0.0335713141,
-0.0361807048,
-0.0142763406,
0.0147250658,
-0.048880633,
-0.0258129332,
-0.0080762953,
0.0229663234,
0.0078505641,
0.0320644677,
-0.0103166243,
-0.0353775732,
0.052239649,
0.0234062206,
0.0024823467,
0.0284179803,
0.0157846678,
-0.0061032027,
-0.0056043933,
0.0321074836,
-0.0869740546,
-0.0376919359,
0.0378018208,
0.0145761399,
-0.0109611051,
0.0067505133,
-0.0536673442,
-0.0322902352,
-0.0080869878,
-0.0429311059,
0.0314380191,
0.0151870865,
0.0391682722,
0.0875749961,
-0.0262817983,
-0.0046457979,
-0.0124585219,
-0.042849306,
-0.0137832109,
0.000213861,
-0.0420473143,
-0.0375288837,
-0.0253817197,
0.019404389,
-0.0343912132,
-0.0441479124,
-0.0087665459,
0.0264901817,
-0.0333775021,
-0.0019440907,
0.0011167143,
0.0267652143,
-0.0273974836,
-0.0028070144,
0.0063669281,
0.0398460217,
-0.0168672372,
0.0080156615,
0.004722225,
-0.0505362414,
-0.0168312378,
0.0152257383,
-0.0082645267,
0.0222393665,
0.0220754556,
0.0320329629,
-0.0428807512,
-0.0214360487,
-0.0157569014,
0.0063542617,
0.0066550747,
-0.0686406046,
-0.0566992275,
-0.050747335,
0.0264505129,
0.0040428177,
0.0252910964,
-0.0189379882,
0.0236802511,
-0.0104289632,
-0.0035931265,
0.0819001868,
-0.0392968357,
0.0042876611,
0.049461104,
0.0060158637,
0.0336864516,
0.0880209282,
0.0058212709,
0.004016927,
0.029740816,
0.0015065328,
-0.0408495702,
0.0369763188,
0.0307672136,
0.0444726758,
-0.0491685569,
0.0119270002,
0.0495796278,
0.063696526,
0.0196777191,
0.0726652294,
-0.0612878166,
-0.0500198677,
0.0818852261,
0.0599769168,
-0.0425467379,
0.010626534,
0.0027796552,
-0.007067983,
-0.0065963161,
-0.0176634677,
-0.0146620451,
-0.0216657873,
0.0311379042,
-0.0144519843,
0.0375605375,
0.0503589287,
-0.0471141599,
-0.0043161716,
-0.0100011965,
0.0075850412,
0.0001506229,
0.0121765714,
0.032285586,
-0.0176089015,
-0.0362635851,
0.021228049,
0.0446850248,
-0.0215836503,
0.0067605022,
-0.0075944159,
0.0095803197,
0.0628713667,
0.0371118821,
-0.0083634323,
-0.0298318006,
-0.0240792166,
0.0237983074,
0.0409985781,
0.0021455865,
-0.0097825127,
0.0256314166,
-0.0517235696,
0.0040700203,
0.0451526418,
-0.0273412503,
0.0139701515,
0.0610750914,
0.0179199204,
-0.0125231296,
0.0530563742,
-0.0309969857,
0.0286535844,
0.0122867357,
-0.0059079956,
-0.0082662683,
-0.0018610274,
0.0573432781,
-0.0151921734,
-0.031348981,
-0.041591581,
0.0527803972,
-0.0185846034,
-0.0083165392,
0.0088432757,
-0.078255631,
-0.0140847452,
0.0455621444,
0.0080506271,
-0.0073579964,
0.0452395342,
-0.0507471822,
-0.0404814743,
0.0146409879,
-0.0254012831,
0.0412780866,
-0.0589432083,
0.012212499,
0.0020043582,
0.0493679792,
-0.0128449658,
0.0213920269,
-0.0247723833,
-0.0278357547,
-0.0009740133,
-0.0007590528,
0.0032326079,
-0.0647338554,
-0.0125194369,
0.0166874845,
0.0190744009,
0.0330824926,
-0.0082978038,
-0.0320544876,
-0.0168251023,
0.0109569859,
-0.0932473466,
0.01191728,
-0.034429729,
0.0233675614,
0.0404292531,
-0.0013222937,
-0.0021681078,
-0.0393119305,
0.0289277174,
0.0449757539,
0.0094254613,
0.0418667309,
0.0096599832,
0.0270969979,
-0.0431035981,
-0.0039198296,
-0.0563946925,
0.0429680534,
0.0629319549,
0.0173488185,
-0.0559438802,
-0.0411485285,
0.0101340013,
-0.0242534578,
0.0713985339,
-0.022873152,
0.0164698288,
0.0073389253,
-0.0041943211,
0.0088844318,
0.0450161994,
0.0184970237,
0.0005892132,
-0.006910543,
-0.0005111791,
-0.0286714826,
0.0172139835,
-0.0511280186,
-0.0114268949,
-0.0506185219,
0.0504528433,
-0.0252995379,
0.0779999644,
-0.0390032753,
-0.0313039273,
-0.0358463712,
-0.048581779,
-0.0177096799,
-0.0673585385,
-0.0757198855,
-0.0219769776,
-0.0034727694,
-0.0398780815,
-0.046574112,
-0.0783064589,
0.0463276282,
-0.0331249647,
-0.0153203802,
-0.0135804992,
-0.0734798238,
-0.0337133668,
0.082570076,
0.0151968803,
0.0344849452,
-0.0615968294,
-0.0250749961,
0.0082717985,
0.0084498748,
-0.0078315884,
0.0385899171,
-0.0084356321,
-0.0378693938,
0.0511600673,
-0.0331856906,
-0.0897702053,
0.0028647021,
0.0123734288,
-0.0374687202,
0.0214575604,
0.0170867108,
-0.0458357111,
-0.0014947405,
0.0534025319,
-0.0063376226,
0.0666887015,
-0.0247277636,
-0.0140803233,
-0.0227015,
0.0079882042,
0.0372335948,
0.0032704032,
-0.0293438062,
-0.0417170711,
-0.0218920931,
0.0786652714,
0.0418466665,
-0.0288100634,
-0.0609416924,
-0.0183090735,
0.0113953417,
-0.0554887839,
-0.0147971874,
-0.0116650695,
0.0267917197,
-0.0721198767,
-0.0136915138,
0.0386277251,
-0.0039554909,
-0.0940725282,
-0.0227647964,
0.0336973928,
0.0489334539,
0.0114432173,
0.0044698291,
-0.059531372,
-0.0247891061,
0.0013475373,
-0.0035744177,
0.0393578447,
-0.0160409752,
-0.0217494015,
-0.0993225425,
0.0173764117,
0.0053767608,
-0.0502966233,
-0.0136559727,
-0.0631015003,
-0.0094355708,
-0.0061574364,
-0.0025476366,
0.014510856,
0.0035296322,
-0.0020776012,
0.0284411963,
0.0095040342,
-0.0224174801,
0.0253825095,
-0.0460277982,
0.0040681097,
0.0430207402,
-0.0212782342,
0.0157854278,
-0.023106873,
0.0255628955,
0.0183919594,
0.0194348469,
-0.0280701891,
-0.0160959437,
-0.0452911481,
-0.0162221715,
0.0015860982,
0.0115496283,
0.0052946708,
-0.0100642368,
0.0193688776,
0.104138881,
0.0112959417,
-0.0029911867,
0.0188344046,
-0.007901459,
0.0363893993,
0.045254793,
-0.018354822,
-0.0279007833,
-0.0389715582,
0.0046409708,
0.0030514968,
0.0161864944,
0.0191866439,
-0.0470224097,
0.0252551939,
-0.0253062379,
-0.0349444263,
0.0043925848,
0.0319762267,
0.0318930037,
0.0151697891,
0.0082559958,
0.0223845989,
0.0140083991,
-0.0904898047,
0.0307919066,
-0.0656686351,
0.0065031503,
-0.0806560591,
0.0150710419,
0.0143319732,
0.0073348167,
-0.0210715998,
0.0276230201,
-0.0983219743,
-0.0809408352,
0.013614635,
0.01529977,
0.0145127373,
0.0913296193,
0.0258069113,
-0.0734711811,
0.0420865268,
0.0781212747,
-0.0615780689,
0.0586177818,
0.017385684,
-0.0439981036,
0.0036003804,
0.0427252725,
-0.0040012333,
0.0195438284,
0.020788407,
-0.0166727398,
-0.0442287102,
-0.0082611386,
-0.0371772461,
0.0076264204,
0.0384845361,
-0.0065672179,
0.0217316505,
-0.0124645634,
0.0296186116,
0.0421019606,
0.0150617585,
-0.0469030924,
-0.0056799543,
0.0333733484,
-0.0009978412,
0.0167817734,
0.0413207002,
0.0884890109,
-0.0243438762,
-0.0089794798,
-0.0267942473,
0.018784564,
-0.0236698017,
0.0636838526,
0.0299399514,
0.0110577801,
-0.0132062808,
0.0007087071,
0.0198930576,
-0.0168457702,
0.0056842808,
0.0315855108,
-0.0111933751,
-0.0119441077,
0.0205048043,
-0.0539026149,
-0.0654110387,
-0.0263440106,
0.0376984365,
0.1053638384,
0.0506364442,
0.008472099,
-0.0244574361,
-0.0274631083,
-0.0263345987,
0.0121200588,
-0.0195862968,
-0.0516620278,
0.0261750817,
0.0362423472,
-0.0062109246,
0.0806576312,
0.0507242344,
-0.0373728313,
0.0333669037,
0.0201707315,
0.0067310021,
-0.0320443399,
-0.0386808738,
-0.0097406721,
0.0120756682,
-0.0138915237,
-0.0134430761,
0.0470479205,
-0.0336673856,
0.0217066258,
0.0166900121,
0.0117216762,
-0.0165575277,
0.0494078062,
0.0384418257,
-0.001387162,
-0.0088088633,
-0.0318431854,
0.0376966558,
0.0156434346,
0.0078414092,
0.008100789,
-7.046907166e-33,
0.0125413975,
0.025729496,
0.0345612988,
-0.000546768,
-0.0271745101,
0.0437630154,
-0.0148288831,
-0.0100288382,
0.0065815193,
0.0186821092,
-0.0329052992,
0.0434619449,
0.0237533879,
0.0372685865,
-0.0063679414,
-0.0003993869,
0.05504971,
-0.013543996,
-0.0019605746,
0.0486342497,
0.0354609862,
0.0591919795,
0.0404604822,
-0.0102672214,
-0.0239155069,
-0.0149294017,
-0.0097039212,
0.0376084,
0.0074768183,
0.0028228615,
-0.0723020807,
0.0303738154,
-0.006289226,
-0.084073782,
-0.0228203703,
-0.0037283765,
-0.0654576495,
0.0329411589,
0.002329475,
0.0029019227,
-0.0626943707,
0.0207686275,
0.0060227513,
0.0290841833,
0.0062491917,
-0.0050834133,
-0.0067544593,
-0.0019717806,
-0.0508010164,
-0.0006827763,
-0.0023739287,
0.0290256962,
0.0071009709,
0.0436802469,
0.049198471,
-0.0630830303,
-0.0025473074,
0.009365893,
-0.0697888061,
0.0111239646,
-0.0033068182,
0.0651700273,
0.0515753701,
0.0477164835,
0.0570349917,
0.0642841831,
0.0440538675,
0.0047909743,
-0.0347701348,
0.0445952378,
0.0095391124,
0.029520819,
0.0188343227,
0.0145767163,
0.0223951731,
-0.011010699,
-0.0753864422,
-0.0688973293,
0.0244496949,
0.0190645643,
-0.0085999705,
0.0283934139,
-0.0239850972,
-0.042558182,
-0.0321325101,
-0.0193269625,
-0.0012978036,
0.0090538124,
0.0188599508,
0.0775820538,
-0.0319707654,
0.0584198087,
0.0162928589,
-0.0367757678,
-0.0053085489,
0.0068958472,
0.0097230515,
0.0160080716,
0.0327515379,
0.0256781876,
-0.0980777666,
-0.023586493,
-0.044523675,
-0.0013695451,
0.0100020617,
-0.0123209404,
0.0511270091,
0.0381795056,
-0.0506938174,
-0.0035419136,
-0.0328917876,
0.0490286425,
0.0201420244,
0.0400187932,
0.0579993539,
0.0072292965,
0.0248002317,
0.0614712127,
-0.0037206353,
0.0257955026,
0.0204693619,
-0.008776594,
-0.0109632984,
0.0327673517,
0.0143006677,
0.0114000952,
-0.0585499592,
-0.0059414436,
0.0466723852,
-0.0866150409,
-0.0052726483,
0.0045450237,
2.917e-7,
-0.0303464979,
0.0221738499,
0.0220332853,
-0.0081707733,
0.0349493138,
-0.0038577763,
0.013218319,
0.020001689,
0.0275189932,
0.0252174102,
-0.0147657953,
0.0712076351,
0.0385191664,
-0.0004030721,
-0.0843108967,
-0.0255484544,
-0.1099020317,
0.0371732339,
-0.0520712249,
-0.0317164883,
-0.0832459182,
0.0929343253,
0.0224024113,
-0.0256110281,
0.0468559638,
-0.0354708731,
0.0458885059,
-0.027547637,
0.0148483403,
0.0135202166,
0.0105508668,
0.0350758769,
-0.0416733883,
0.0148691572,
-0.023340147,
-0.0059401332,
-0.0282033347,
0.0638130158,
-0.0058083665,
0.0540851131,
0.0190707296,
0.028124772,
-0.0596954487,
-0.0082610855,
0.0816123635,
-0.0284390356,
0.0035860059,
-0.0727079734,
0.0291107818,
-0.0516878441,
0.0368018821,
-0.0007750027,
0.0199567005,
0.008418222,
0.0091207726,
-0.0087832119,
-0.0305397548,
-0.0750253275,
-0.0049895942,
-0.0369388126,
-0.0655892938,
-0.0410657674,
-0.0646288618,
0.0018861877,
0.0165922493,
0.0029197396,
-0.001688196,
2.365175908e-34,
0.053937912,
0.0089246305,
0.0018926564,
-0.0022340603,
-0.0054721567,
-0.0538590103,
-0.0240342785,
-0.0153863337,
-0.026454065,
-0.0767394826,
-0.0235584751
] |
http://arxiv.org/abs/2402.01741v2 | http://arxiv.org/pdf/2402.01741v2 | Development and Testing of a Novel Large Language Model-Based Clinical Decision Support Systems for Medication Safety in 12 Clinical Specialties | 2024-01-29T16:03:29 | 2024-02-17T21:13:16 | Jasmine Chiat Ling Ong, Liyuan Jin, Kabilan Elangovan, Gilbert Yong San Lim, Daniel Yan Zheng Lim, Gerald Gui Ren Sng, Yuhe Ke, Joshua Yi Min Tung, Ryan Jian Zhong, Christopher Ming Yao Koh, Keane Zhi Hao Lee, Xiang Chen, Jack Kian Chng, Aung Than, Ken Junyang Goh, Daniel Shu Wei Ting | cs.CL | cs.CL, cs.AI | Importance: We introduce a novel Retrieval Augmented Generation (RAG)-Large
Language Model (LLM) framework as a Clinical Decision Support Systems (CDSS) to
support safe medication prescription.
Objective: To evaluate the efficacy of LLM-based CDSS in correctly
identifying medication errors in different patient case vignettes from diverse
medical and surgical sub-disciplines, against a human expert panel derived
ground truth. We compared performance for under 2 different CDSS practical
healthcare integration modalities: LLM-based CDSS alone (fully autonomous mode)
vs junior pharmacist + LLM-based CDSS (co-pilot, assistive mode).
Design, Setting, and Participants: Utilizing a RAG model with
state-of-the-art medically-related LLMs (GPT-4, Gemini Pro 1.0 and Med-PaLM 2),
this study used 61 prescribing error scenarios embedded into 23 complex
clinical vignettes across 12 different medical and surgical specialties. A
multidisciplinary expert panel assessed these cases for Drug-Related Problems
(DRPs) using the PCNE classification and graded severity / potential for harm
using revised NCC MERP medication error index. We compared.
Results RAG-LLM performed better compared to LLM alone. When employed in a
co-pilot mode, accuracy, recall, and F1 scores were optimized, indicating
effectiveness in identifying moderate to severe DRPs. The accuracy of DRP
detection with RAG-LLM improved in several categories but at the expense of
lower precision.
Conclusions This study established that a RAG-LLM based CDSS significantly
boosts the accuracy of medication error identification when used alongside
junior pharmacists (co-pilot), with notable improvements in detecting severe
DRPs. This study also illuminates the comparative performance of current
state-of-the-art LLMs in RAG-based CDSS systems. | Development and Testing of a Novel Large Language Model-Based Clinical Decision Support Systems for Medication Safety in 12 Clinical Specialties: Importance: We introduce a novel Retrieval Augmented Generation (RAG)-Large
Language Model (LLM) framework as a Clinical Decision Support Systems (CDSS) to
support safe medication prescription.
Objective: To evaluate the efficacy of LLM-based CDSS in correctly
identifying medication errors in different patient case vignettes from diverse
medical and surgical sub-disciplines, against a human expert panel derived
ground truth. We compared performance for under 2 different CDSS practical
healthcare integration modalities: LLM-based CDSS alone (fully autonomous mode)
vs junior pharmacist + LLM-based CDSS (co-pilot, assistive mode).
Design, Setting, and Participants: Utilizing a RAG model with
state-of-the-art medically-related LLMs (GPT-4, Gemini Pro 1.0 and Med-PaLM 2),
this study used 61 prescribing error scenarios embedded into 23 complex
clinical vignettes across 12 different medical and surgical specialties. A
multidisciplinary expert panel assessed these cases for Drug-Related Problems
(DRPs) using the PCNE classification and graded severity / potential for harm
using revised NCC MERP medication error index. We compared.
Results RAG-LLM performed better compared to LLM alone. When employed in a
co-pilot mode, accuracy, recall, and F1 scores were optimized, indicating
effectiveness in identifying moderate to severe DRPs. The accuracy of DRP
detection with RAG-LLM improved in several categories but at the expense of
lower precision.
Conclusions This study established that a RAG-LLM based CDSS significantly
boosts the accuracy of medication error identification when used alongside
junior pharmacists (co-pilot), with notable improvements in detecting severe
DRPs. This study also illuminates the comparative performance of current
state-of-the-art LLMs in RAG-based CDSS systems. | [
0.0685763657,
0.005471155,
-0.0044262647,
-0.0371897034,
0.0108762728,
0.0283711385,
0.0232627559,
-0.0045283781,
0.0463315174,
-0.0286988728,
0.044054009,
-0.0330252945,
-0.0368695855,
0.0327604003,
0.0190575607,
0.022096511,
0.0151519096,
0.0223356988,
0.062037427,
0.0270192847,
0.0152528789,
-0.0039740917,
-0.0462999307,
0.0595670976,
-0.0197337903,
-0.0019728667,
0.03729349,
-0.0345818251,
0.0270392243,
-0.091967389,
0.01041988,
0.0229264218,
0.0306690522,
0.0084397923,
0.0000021321,
-0.0752164796,
-0.0117072742,
0.0286975596,
-0.0929280519,
-0.0881542489,
0.0492171347,
0.0114312004,
-0.0017047101,
0.0257322434,
-0.0191696752,
-0.0326868892,
0.0338937119,
0.0483997948,
0.0129702734,
0.0568294041,
-0.00874859,
0.0034509983,
-0.0041585318,
0.0060354955,
0.0055863895,
-0.0513402633,
0.0436361022,
0.0481764115,
0.0533221103,
-0.0119206849,
0.0400098711,
0.0152840689,
-0.0210941322,
-0.0132594658,
0.0073779724,
0.0049052965,
-0.0305100121,
-0.029248314,
-0.0066827261,
0.0115990518,
0.0485391654,
0.0201698765,
-0.0051889597,
0.0029826502,
0.016854208,
0.0268281102,
-0.0182856731,
-0.0228979345,
-0.0427891426,
0.0070557375,
-0.0127693098,
0.0074431757,
0.013658368,
-0.0110744815,
-0.0359312743,
0.0310326051,
-0.0256780628,
-0.0249231178,
0.0307779945,
0.0020784154,
-0.011805743,
0.0286990106,
0.0537460968,
0.0348826125,
-0.0801315382,
0.0131221702,
0.0267247353,
-0.0411144458,
0.1017972454,
-0.067344822,
0.0554117709,
-0.0159893762,
0.0117790904,
-0.0110838311,
-0.0033999204,
0.0334561579,
0.0031564997,
0.0208454505,
-0.0730298385,
0.0295082331,
-0.0474367402,
-0.0454199463,
-0.0330393463,
0.0149682416,
0.0110122627,
-0.0129805282,
-0.1008673981,
-0.0200471506,
-0.0501756035,
0.0285807494,
0.0130237984,
0.0571825467,
-0.0134366844,
0.0123465061,
-0.0257810708,
0.0216152128,
0.0216762051,
0.0497645624,
0.0136100203,
-0.0065037715,
0.065346159,
0.0095122503,
-0.0067533189,
0.0298732538,
0.0177922621,
0.0814779624,
-0.0136885187,
-0.0163286552,
-0.0323346183,
-0.0627936348,
-0.0279825255,
-0.0216568671,
-0.0128289433,
-0.0574088879,
0.03124189,
0.0169862006,
-0.0292199589,
-0.0284539778,
0.0239418168,
0.0253708959,
-0.0280107297,
0.0243007932,
-0.030205315,
0.0201958772,
0.0420407578,
0.0109924162,
0.0118331965,
-0.0040782844,
-0.014866923,
-0.0203895774,
-0.0082546892,
-0.0192457084,
0.0386617258,
-0.0170570929,
-0.0438269265,
-0.0338031948,
-0.0205661077,
0.0244122352,
-0.0044937753,
0.0402908251,
-0.0692067966,
0.0613601096,
-0.0162120536,
0.0342886746,
0.0198853519,
0.0048367023,
0.0598302335,
-0.0029176902,
0.0277537275,
-0.0186262634,
-0.0142196743,
-0.0347819403,
-0.0039681713,
-0.0322503336,
-0.0384318531,
-0.019186005,
-0.0089365281,
-0.0084606465,
-0.0177639816,
-0.0183082391,
-0.0493717678,
-0.0227965135,
-0.0445556715,
0.0177632645,
0.0512437299,
-0.0454398617,
-0.0315745212,
0.0469162762,
0.0162984915,
-0.0209716782,
-0.0535441861,
-0.0381966457,
-0.0069011711,
0.1257572919,
0.0341615789,
-0.0453064404,
-0.0386150256,
-0.0345184021,
-0.022653684,
-0.0420836434,
-0.0075137159,
-0.0308975559,
-0.0379727297,
-0.0305167455,
-0.0125361327,
0.079767637,
0.0071584699,
0.0002114334,
-0.0271500871,
0.031074997,
0.0003012791,
0.0053797835,
-0.0360592157,
0.0165647641,
-0.0293432418,
-0.0506793559,
-0.0228139497,
-0.0040888349,
0.0057963151,
0.0609665886,
0.0108314827,
0.0336823054,
0.0558195077,
-0.0383402333,
-0.0465204716,
-0.0201880317,
0.0085987542,
0.0127203679,
0.0306147803,
0.015247087,
0.0126395375,
-0.0520283654,
-0.0324306823,
-0.0364546627,
-0.0101672141,
-0.0664901361,
-0.0026949171,
-0.0135925105,
0.0282057691,
0.0265774671,
0.0059700385,
-0.0540047511,
-0.0090921372,
0.0138334604,
0.0151134022,
0.0102513982,
-0.0010706757,
0.0176734775,
-0.0156565644,
0.0149947135,
0.0265255217,
-0.0163542535,
0.0594902188,
0.0159223862,
0.007274963,
-0.0011942454,
-0.0180944242,
-0.057837151,
-0.0285112411,
0.0386276394,
-0.0043782145,
-0.0031318218,
-0.0202047322,
-0.0037768178,
-0.0145809855,
-0.0082425643,
0.0135628721,
0.0355980545,
-0.0031901011,
0.001411058,
0.0847712457,
-0.0162697304,
-0.0088138552,
-0.0141647058,
-0.0294345766,
0.023744965,
0.0071565178,
-0.0147857144,
-0.0596749783,
0.023919927,
-0.0538960807,
0.0045207767,
0.0018661236,
-0.0084127476,
0.0287795067,
-0.0410841741,
0.0004361286,
0.0107957758,
-0.0450860411,
0.001600309,
0.0407870933,
-0.0452767164,
-0.0370067731,
0.0084309587,
-0.0512132421,
-0.0765436813,
0.0554098897,
-0.0304030553,
-0.0440576151,
0.0226149205,
0.0050760307,
-0.0079611698,
0.0022915772,
-0.0171688385,
-0.0259141196,
0.0023220412,
-0.0478388965,
-0.0072150696,
-0.0174807273,
0.0522100218,
0.0286572501,
-0.0430611707,
0.0692515522,
0.0047547887,
0.0358508565,
-0.087606445,
-0.0359806232,
-0.0189572051,
-0.0556863472,
-0.0040960833,
0.0190367438,
-0.007403126,
-0.0260521434,
0.0181168132,
-0.0181663837,
-0.0113856345,
-0.0650301054,
0.0599035807,
-0.0390583538,
0.0366569459,
0.0008826643,
-0.0262386426,
-0.0388183221,
0.0060527525,
-0.00195286,
0.001717355,
-0.0070512122,
0.0159629323,
-0.0570287704,
0.0303177647,
-0.0008352122,
0.032870356,
0.0331231616,
-0.0302982777,
0.0279498305,
-0.0517290719,
0.0184435453,
-0.0342652872,
0.0530002229,
-0.005465223,
0.0584923811,
0.0528977364,
0.0110326232,
-0.0207501333,
-0.0256554224,
0.0207146406,
0.0178126935,
0.0071444153,
-0.0165260248,
0.0548664369,
-0.0182707198,
-0.0373065732,
0.0141186267,
0.0767156407,
0.0361865163,
-0.0543598831,
0.0011623951,
-0.036862988,
-0.090283595,
-0.0037720986,
0.0072118351,
-0.0194978956,
-0.0049536298,
-0.0162373278,
0.0620140657,
-0.0503988564,
-0.02400296,
-0.0051777493,
-0.0496935211,
0.05363952,
0.0302172061,
-0.0399068706,
-0.025244778,
0.0020600678,
0.0024278187,
0.0017130991,
0.0163852368,
0.0213546306,
-0.0009799093,
-0.0297356192,
0.0294440817,
-0.0732872412,
-0.085833542,
0.0014329915,
-0.0129277995,
0.0338728689,
0.0499693677,
-0.0145151578,
-0.0121570723,
-0.0154238734,
-0.0094022686,
-0.0162921771,
0.0201100819,
0.0013317937,
0.0322022103,
0.0426127166,
0.0438500717,
-0.0558685288,
0.0101847062,
-0.0541568249,
-0.017925838,
-0.0286993627,
0.08810018,
-0.0536982194,
-0.0106407767,
-0.0554578006,
0.0402779467,
0.0465441719,
0.0057188608,
-0.0076531111,
0.0333233364,
-0.0443641581,
0.016469473,
0.0137227653,
0.021744173,
0.0245473199,
0.0360319652,
0.0012775162,
-0.0110195009,
-0.0432324298,
-0.0101818219,
0.0300563555,
0.0062652798,
0.0035615244,
0.0301627703,
-0.0051370244,
-0.0079151662,
-0.0145294797,
0.0029270493,
0.0468249246,
0.0503379777,
-0.0618105531,
0.0233191662,
0.0368975475,
0.0221662484,
-0.0105213868,
0.0054162666,
0.0384391844,
-0.0408503711,
0.0537704006,
0.0517706312,
0.0298909377,
-0.0730854198,
-0.0142010283,
-0.0361434333,
0.0357342698,
0.0711939558,
0.0150879864,
-0.0487373061,
-0.0438853242,
-0.0071234517,
-0.0029330519,
-0.0300926641,
-0.041530326,
0.034871079,
-0.0083879009,
0.0345397405,
0.0021211193,
-0.027423298,
0.0288326126,
0.0355548933,
0.024744818,
-0.0131766936,
0.019322874,
-0.0263593085,
0.0059881215,
-0.0312243402,
0.0122484639,
-0.0018672878,
0.0108798342,
0.0047302283,
-0.0448821709,
0.0444688462,
-0.0120703857,
-0.0534073114,
-0.0528448932,
-0.0002898321,
0.0371689387,
0.0004651989,
0.0305584427,
0.0770630613,
-0.0069687432,
0.003158652,
-0.0215667319,
-0.019651413,
0.0375801548,
-0.0052494695,
0.1079800278,
-0.0477074981,
0.0024784908,
-0.0373875238,
0.0367590524,
-0.0043354845,
-0.1052558199,
-0.0103841806,
0.0382811539,
0.0720515549,
-0.0126415482,
-0.001654882,
-0.0372788087,
0.0007898968,
-0.0759330615,
0.0134662092,
0.0213629901,
0.02885383,
0.0144879464,
0.0305389408,
-0.0173524488,
-0.0299035758,
0.0851740167,
-0.0232199095,
-0.0081365174,
-0.0686835498,
-0.0806719959,
-0.0219332725,
-0.0194080062,
0.0218670145,
0.0081178183,
-0.0739248395,
-0.0135121709,
-0.0967359617,
-0.0338172093,
-0.0217486005,
0.0204381626,
-0.0086508095,
-0.0162550565,
0.0315799378,
0.0547301359,
0.0085965283,
-0.023578709,
-0.039318759,
0.0258164424,
-0.0441788658,
0.0310754254,
-0.0235934108,
-5.160697911e-33,
0.0149495974,
-0.0342155658,
0.0122244433,
-0.0440592393,
0.0170462895,
-0.0009870837,
-0.0075614038,
0.0503496826,
0.0339972302,
0.0381423496,
0.0036737535,
-0.0107909655,
0.0061779311,
0.0072112833,
0.0495944284,
-0.01623003,
0.000548206,
0.0392736755,
0.0118900379,
0.0116074719,
0.0401790738,
0.0542636327,
0.0159611646,
-0.1016805917,
0.0223689545,
-0.0073422827,
0.0588527359,
0.024123529,
-0.0207186546,
0.0644917637,
0.0145930061,
0.0538893417,
0.0207873303,
-0.0279489923,
-0.0277605653,
0.014253282,
0.0006549442,
-0.0201797374,
-0.0088065108,
-0.0435276031,
0.0203265473,
-0.0113824252,
0.0443717204,
-0.024353059,
-0.0776111111,
0.0519528016,
0.0082087135,
-0.0356111117,
-0.0269259959,
-0.0368175991,
-0.018677352,
0.0195709337,
0.0035334001,
0.002541115,
-0.0272337683,
0.0638774857,
-0.0303804222,
0.0210856143,
-0.0314956121,
0.0022366759,
-0.0055260253,
0.0032825319,
0.0163828097,
-0.0033563883,
-0.0476246923,
0.036485523,
0.0034858435,
-0.0159600694,
-0.0075971265,
-0.0193030871,
-0.0694159716,
0.0694547147,
0.0607421771,
0.0314699188,
0.025280226,
-0.0203456227,
-0.0720707402,
0.0355192162,
-0.1098996997,
0.0832505897,
0.0481557995,
-0.0083749574,
-0.0979821011,
-0.0342470072,
0.0090118852,
0.0482435003,
-0.038272839,
-0.0450628139,
0.0467240922,
-0.0029531303,
-0.0064859516,
0.0245386027,
0.020178711,
-0.0081304144,
0.0143547785,
-0.0390183739,
0.0279230829,
0.011844594,
-0.0357360579,
-0.0188376252,
-0.0636416152,
-0.0155411102,
0.0090897446,
-0.0209009573,
0.0096406769,
0.0270828716,
0.0051806709,
-0.0228001829,
-0.0129955923,
0.0290117469,
0.0435055383,
-0.002883821,
0.0002301505,
0.0181676522,
0.0385582224,
0.0348339565,
0.0118910233,
0.0588156655,
0.0147368945,
-0.0369977541,
-0.0092365686,
-0.0646521524,
0.0085139507,
-0.0000402473,
-0.0306564625,
0.0444840677,
-0.0649903715,
0.0723031834,
0.041197639,
0.0039934651,
0.0228200145,
0.0074618068,
2.967e-7,
0.0795413554,
0.0320882052,
0.0045673433,
-0.0355040096,
0.0188243575,
0.0136683686,
-0.0001766038,
-0.0099780858,
0.0052996827,
0.0108651537,
0.0649822056,
-0.0335496701,
0.0130705917,
0.0482541174,
0.0088067278,
-0.0553892888,
-0.0243998785,
-0.0474826954,
-0.0503135845,
-0.0055125668,
0.0974439085,
-0.0067662685,
-0.0254152976,
-0.0509632789,
0.0215811729,
-0.0512554571,
0.0118066743,
-0.0382702276,
0.0222843643,
0.0239074212,
0.0692609996,
0.0120124482,
0.0228888895,
0.0719756261,
0.0098647652,
-0.0202327315,
-0.0378572643,
0.0513246842,
-0.0752512366,
0.0353406221,
0.0336170383,
-0.0170047209,
0.0406383947,
-0.0380679816,
0.0043423683,
-0.039399676,
-0.0287673976,
0.0142681757,
-0.0126625802,
-0.0418972112,
0.0781717524,
-0.0215342753,
-0.0222651344,
0.0303068887,
-0.0065755565,
0.0517496802,
-0.0083665792,
-0.0545239039,
0.0267946292,
0.0817541331,
-0.0266635474,
-0.0433014855,
0.0607528798,
-0.0185988769,
-0.0173120834,
-0.0628444999,
0.0262621976,
2.651409931e-34,
0.0305921063,
0.0228558574,
0.0169212315,
-0.0219749119,
0.0595592633,
0.0177145284,
-0.028615376,
0.0183227751,
0.0070697386,
-0.0379033908,
-0.0125716915
] |
http://arxiv.org/abs/2401.16247v1 | http://arxiv.org/pdf/2401.16247v1 | Towards Red Teaming in Multimodal and Multilingual Translation | 2024-01-29T15:49:40 | 2024-01-29T15:49:40 | Christophe Ropers, David Dale, Prangthip Hansanti, Gabriel Mejia Gonzalez, Ivan Evtimov, Corinne Wong, Christophe Touret, Kristina Pereyra, Seohyun Sonia Kim, Cristian Canton Ferrer, Pierre Andrews, Marta R. Costa-jussà | cs.CL | cs.CL, cs.CY, I.2.7 | Assessing performance in Natural Language Processing is becoming increasingly
complex. One particular challenge is the potential for evaluation datasets to
overlap with training data, either directly or indirectly, which can lead to
skewed results and overestimation of model performance. As a consequence, human
evaluation is gaining increasing interest as a means to assess the performance
and reliability of models. One such method is the red teaming approach, which
aims to generate edge cases where a model will produce critical errors. While
this methodology is becoming standard practice for generative AI, its
application to the realm of conditional AI remains largely unexplored. This
paper presents the first study on human-based red teaming for Machine
Translation (MT), marking a significant step towards understanding and
improving the performance of translation models. We delve into both human-based
red teaming and a study on automation, reporting lessons learned and providing
recommendations for both translation models and red teaming drills. This
pioneering work opens up new avenues for research and development in the field
of MT. | Towards Red Teaming in Multimodal and Multilingual Translation: Assessing performance in Natural Language Processing is becoming increasingly
complex. One particular challenge is the potential for evaluation datasets to
overlap with training data, either directly or indirectly, which can lead to
skewed results and overestimation of model performance. As a consequence, human
evaluation is gaining increasing interest as a means to assess the performance
and reliability of models. One such method is the red teaming approach, which
aims to generate edge cases where a model will produce critical errors. While
this methodology is becoming standard practice for generative AI, its
application to the realm of conditional AI remains largely unexplored. This
paper presents the first study on human-based red teaming for Machine
Translation (MT), marking a significant step towards understanding and
improving the performance of translation models. We delve into both human-based
red teaming and a study on automation, reporting lessons learned and providing
recommendations for both translation models and red teaming drills. This
pioneering work opens up new avenues for research and development in the field
of MT. | [
0.0512057766,
0.0412780009,
-0.0270245522,
0.0273593739,
-0.0440784693,
0.0106210308,
0.0478303693,
-0.0015338092,
-0.0326092504,
-0.0595886223,
-0.0615465827,
-0.0200789478,
0.0234865323,
0.0093929917,
0.0091501391,
-0.0777597874,
0.0348531231,
-0.0106912507,
-0.0466836169,
-0.0009371189,
-0.0190020762,
0.0193060692,
-0.0078367144,
0.0358453691,
-0.0242949668,
0.0100091649,
0.0188795906,
0.0587032475,
0.0426736474,
-0.0099492483,
-0.0382844657,
0.0571695454,
-0.0010088995,
0.050479129,
0.0000021918,
-0.0472603478,
-0.0302000083,
-0.0048153074,
0.0394432098,
-0.0304420423,
0.0306646321,
-0.0364417695,
-0.014518287,
-0.0028389809,
-0.049326621,
-0.0211553853,
0.0536534525,
-0.0009049756,
0.0526209027,
0.0626617149,
-0.017784765,
0.0029106156,
-0.0179065689,
-0.0327047594,
0.0018444138,
-0.0665939227,
-0.0047049085,
0.0094747022,
-0.0133503908,
-0.0186765641,
-0.0318367817,
0.046001561,
0.048979722,
0.0018478255,
-0.031529326,
0.0224168207,
-0.0314174294,
-0.0605090111,
-0.0380764715,
0.0669328719,
-0.0323128179,
0.0182707049,
-0.0392475501,
-0.0054115481,
-0.0225099362,
0.0183273107,
0.0001897434,
0.0298946425,
-0.0155940317,
0.0101552671,
-0.0027669801,
0.0398861691,
0.0152269499,
0.0004887134,
-0.0305007882,
0.0355422124,
0.0074713454,
-0.0271927398,
-0.0148659321,
0.0625989586,
-0.0098559093,
-0.0192929525,
0.024383219,
-0.02855324,
0.0157977659,
0.0119269695,
-0.0491990559,
-0.0805307701,
0.0122486008,
-0.0241845436,
0.044852823,
0.0343479998,
-0.0223599561,
0.052168835,
-0.0119298445,
0.0490803868,
-0.0104800947,
0.0363119096,
-0.0464851409,
-0.0009519485,
-0.0558038279,
0.0528146103,
-0.0508222543,
-0.0382755771,
-0.0241218843,
-0.0158159155,
-0.0081245927,
0.0038117764,
-0.0132723385,
0.0082131028,
0.0142278811,
-0.0198586788,
-0.0682942346,
-0.0268067122,
-0.0150439851,
0.0039240606,
-0.0253553595,
0.0044507678,
-0.0042267404,
-0.0681745633,
-0.005999594,
0.0283802673,
-0.0431081764,
-0.0459644943,
-0.0089717787,
0.1081422269,
-0.0075974925,
-0.0284974109,
0.0198995024,
-0.0390832499,
-0.016238194,
-0.074659124,
0.0260326825,
-0.0415835343,
0.0578466021,
0.002146614,
0.0336597897,
-0.009737798,
-0.0148308799,
-0.0078069218,
0.0127318501,
0.0874236971,
0.0048263315,
0.0025833347,
-0.0087169176,
0.0151625453,
-0.0266015809,
0.0793625265,
0.0607207939,
0.0289148595,
0.0395241678,
0.0162107665,
0.0196157638,
0.0611136295,
-0.0197337586,
0.0114946058,
-0.0272070505,
-0.0381376781,
0.0270219855,
0.0310676154,
0.0459124893,
0.0474327691,
-0.019790465,
-0.0028288439,
0.049929779,
0.0131410602,
0.0139573356,
0.0331871696,
0.0894139409,
-0.0064991773,
-0.0214405376,
0.0582632311,
-0.0173331443,
0.0326630436,
-0.0118440408,
-0.0221536793,
0.0155957211,
0.0616565533,
-0.0017308872,
-0.0404013805,
-0.0078834612,
-0.0107509885,
0.0287792962,
-0.0293908324,
0.0596223176,
0.0329158567,
-0.0433801189,
-0.0441124476,
-0.0137052909,
-0.0493528247,
0.0139020337,
-0.0676142871,
-0.0337269306,
0.04595083,
0.0623079166,
-0.0141977984,
-0.0066677411,
0.014018748,
-0.0321156718,
0.0395591408,
0.052065596,
0.0237715207,
0.0113056134,
-0.0539050065,
-0.0361289755,
0.0350609235,
0.0012535873,
0.0456336774,
0.0186871439,
0.0185484271,
0.0021550185,
-0.0198824592,
-0.0446313061,
0.0466620997,
0.022823764,
0.0190349799,
0.0247697253,
0.0196662955,
0.0294636674,
0.0036596241,
-0.0533397235,
-0.0034168621,
0.0164060164,
-0.0261895228,
-0.0324603431,
0.0701675192,
-0.00341509,
-0.0519818775,
-0.0160668846,
-0.0855462104,
-0.0261275377,
0.0416519679,
-0.0414352752,
-0.027874155,
-0.0046802075,
-0.0367877483,
0.000614494,
-0.0420750193,
0.0256965272,
0.0210600905,
0.0383525342,
-0.0375401676,
0.0452405736,
-0.0261220653,
-0.0169018619,
-0.0235829372,
-0.0332142822,
0.0273262896,
-0.0733399317,
-0.0127789546,
0.0275866408,
0.0515877753,
0.0380657353,
0.0391884185,
0.0332356729,
-0.0164821688,
-0.0345070027,
-0.0664444193,
-0.0087388996,
-0.0176239889,
0.0161568914,
0.0222829506,
-0.0062580509,
-0.0096894894,
-0.0142078819,
0.0225134026,
0.0086311959,
0.047856953,
0.0052413493,
0.0193282943,
-0.0024000362,
-0.0030220619,
-0.0190209337,
-0.0061274124,
0.0088770436,
0.0571204908,
0.066079706,
-0.0353766866,
-0.067573145,
0.0277690981,
-0.0238092691,
0.0214212257,
0.0126874736,
-0.0622473694,
0.0164086558,
-0.0141749084,
0.0128464224,
0.0433872379,
-0.0296817068,
0.0044902395,
0.0460112542,
0.0088577718,
-0.0204512775,
-0.0034971228,
-0.0021315382,
-0.0226438027,
0.0177461114,
-0.0595913157,
-0.0111843236,
0.0315700062,
0.0381922424,
0.0027693112,
-0.0487948395,
-0.0141017949,
-0.0040953262,
-0.0772514567,
0.009742179,
-0.0074009579,
0.0289788023,
0.0676734149,
-0.0484986976,
-0.0569854341,
0.0829716846,
-0.0216719452,
-0.0140453652,
-0.0538746938,
0.0323605649,
-0.0394966975,
0.0724567175,
0.0190941896,
0.0237541199,
-0.0315449983,
0.0193388816,
-0.0227460619,
0.0144355064,
0.0005722797,
-0.0070206183,
0.0129727963,
-0.085564591,
-0.0001362471,
-0.0107818693,
0.0024272234,
-0.029238373,
-0.0000254715,
-0.0243486147,
-0.016129395,
0.0142879682,
-0.0029880765,
-0.031620156,
0.0190978255,
0.0142903384,
0.0376744345,
0.0214956272,
-0.019910207,
0.0023423047,
0.0355523825,
-0.0180693362,
-0.019373849,
0.0028099988,
-0.0943082497,
0.0187150706,
0.0341698192,
0.0512153432,
0.0095140887,
-0.0595025383,
-0.0118541671,
0.0262804125,
0.0359837748,
-0.0181973074,
-0.0222240593,
0.0182626266,
-0.0612932183,
0.0227255132,
0.0691217855,
0.0282012485,
-0.0642364621,
-0.0233912673,
0.0013221699,
0.0060248938,
0.00324001,
0.012965004,
-0.0880779102,
-0.0386001542,
-0.0029632468,
0.0192919802,
-0.0028926507,
-0.0108010946,
-0.0088158222,
-0.1119071171,
0.049636595,
0.0210504569,
-0.0593050309,
-0.015800491,
-0.0105393035,
-0.0167929251,
-0.0391351357,
0.0149367219,
-0.0252165496,
0.018545974,
0.0087301098,
-0.0006236539,
-0.0004041508,
-0.0766792223,
0.0130905686,
-0.0676782355,
0.0284702033,
0.0881301165,
0.0291113276,
-0.0080517065,
-0.0302715544,
0.0022157293,
0.0050525735,
0.0468183011,
-0.0031408465,
0.0474389866,
-0.0252500493,
-0.0054118396,
-0.0018217171,
-0.021815436,
0.0072023799,
0.0013397518,
0.0147218267,
0.0425539613,
-0.066727519,
-0.0084438054,
-0.0135239102,
0.0174622871,
0.0305743609,
0.0679094791,
0.0059704157,
-0.0645571277,
-0.0122222099,
0.0114594856,
0.0476287529,
-0.0995108485,
-0.0251362771,
0.0062413351,
-0.0659055412,
-0.0480258726,
-0.0896128342,
-0.0662944615,
0.0262996946,
0.0547151491,
0.0260428376,
-0.0213319454,
-0.043010246,
0.0205726288,
0.0025682934,
0.0664760321,
0.0013043068,
0.0123175187,
-0.0568450615,
-0.0048767352,
-0.0144671658,
-0.0479580201,
0.0103014121,
0.0569303893,
-0.0626577809,
-0.0476910137,
-0.0056612892,
0.0859891474,
0.0235064384,
0.0125371953,
-0.0263255332,
-0.0563359931,
-0.0221455209,
0.0273780599,
-0.0116842855,
0.0197178666,
0.0371025391,
-0.0040906202,
-0.031697724,
0.0150556453,
0.0044233161,
0.0373800546,
-0.009045138,
0.0305480473,
-0.0524859354,
-0.0596608929,
0.0052984869,
0.0046400712,
0.0547605902,
-0.0844246596,
-0.0336437523,
-0.0429255441,
-0.0039564534,
-0.0117103821,
-0.0257815123,
-0.0751652792,
-0.0055943849,
0.0267772246,
0.0165119972,
-0.0027646385,
0.0309379026,
-0.0469946451,
-0.0318290666,
-0.0398693196,
0.0345377848,
0.043077834,
0.0350323543,
-0.0198470261,
0.0391185656,
-0.0015801097,
0.0205093008,
-0.0146029312,
0.014259181,
0.0084390268,
0.0719183013,
0.0143055459,
0.0008207461,
0.0195371155,
0.0489581637,
-0.0201588385,
-0.0308012012,
0.0165762194,
0.0415787548,
0.0863272473,
0.0200269111,
0.0771163926,
-0.0245624948,
-0.0117256101,
-0.012533105,
0.0455042906,
0.0168956351,
-0.0112747224,
-0.0060571972,
-0.0036054465,
0.0146582881,
0.0466056578,
0.052476678,
-0.0569437444,
0.0367487743,
-0.0586136319,
-0.0210534558,
-0.014064556,
-0.0184645802,
0.0182259735,
-0.0401404537,
0.0009535244,
-0.0046148268,
-0.0651790127,
-0.0632188395,
-0.001074726,
0.072296083,
0.0382753462,
0.0174386036,
0.0255381204,
0.0940874889,
-0.0394961499,
0.0214888025,
-0.0543947332,
-0.0267965309,
-0.0199796464,
-0.0087787202,
-0.0265649725,
-6.439935515e-33,
-0.0368964933,
-0.0451671369,
0.0181076545,
-0.0138885267,
-0.0070088264,
0.0369323604,
-0.0082429135,
-0.0034411931,
0.0047763353,
0.0237086266,
0.0130336797,
-0.0288399234,
0.0210967399,
-0.0203677136,
0.0118350722,
0.0279840082,
0.0375459827,
-0.0031054385,
-0.0465008579,
-0.0282087885,
0.0284892432,
0.0039351881,
0.0520337224,
-0.084823072,
0.0224603862,
-0.0098852785,
-0.0574011207,
0.0097674448,
-0.0131602315,
0.0302673485,
0.0038087741,
-0.0067154872,
0.0349056348,
-0.0317214914,
-0.0175392888,
0.0094750561,
-0.0751357302,
0.0140842944,
0.0057830224,
0.0432984196,
-0.0721074641,
0.0098784408,
0.0517147928,
-0.02537762,
-0.0378260761,
0.0075813946,
0.0438101739,
-0.0385907628,
0.0000245857,
0.0169027299,
-0.0689906403,
0.0315096602,
-0.0030041288,
-0.037115939,
0.0521267503,
0.0774242431,
-0.0028506655,
-0.039906662,
-0.0576040149,
0.0174514931,
-0.0426763743,
-0.011096851,
0.1014498472,
0.0132262567,
-0.0247724671,
0.0417370722,
0.0313236974,
-0.0136163961,
0.0014998247,
-0.0181514602,
0.0035043643,
0.0236504655,
0.0380879678,
0.0588148274,
-0.0107339071,
0.0248730984,
-0.032802999,
0.0705668479,
-0.0089625074,
0.0246793944,
0.0078645218,
0.0219815709,
-0.0587125644,
0.023421621,
-0.0100328391,
-0.0416352376,
-0.0296287369,
-0.020891957,
0.0046118842,
0.0245775189,
-0.0503198318,
0.0136504713,
0.0714575797,
0.0028242155,
0.0428520814,
-0.0030338466,
-0.0157147199,
0.0333451815,
-0.0219341554,
0.0075946576,
0.0136555061,
-0.0085231122,
0.0565694235,
0.0369795077,
0.0521877073,
-0.0297520179,
0.0359104089,
0.0314810388,
-0.0340372883,
0.0065571284,
-0.0444683433,
0.0226279702,
-0.000268874,
-0.0042028944,
0.0385418981,
0.0038502871,
-0.0383427888,
0.0336028524,
0.0239105523,
0.0161341671,
-0.032183554,
0.0118268915,
-0.0165369269,
0.0181069057,
-0.0565567948,
0.0286897235,
0.0118634291,
-0.0521197245,
0.0373575799,
-0.0223837122,
-0.0459042415,
0.020682374,
2.845e-7,
0.0548827685,
0.0327844918,
0.0376044139,
0.0109434603,
-0.0060467036,
0.04829669,
-0.0583698563,
0.0256778859,
0.0209917165,
0.056066256,
0.0324044302,
0.0470352247,
0.0561184846,
0.0456627719,
-0.0021450517,
-0.0244104359,
-0.0106369983,
0.0339420438,
-0.0314528644,
-0.0140870381,
0.0626603067,
0.0544213541,
0.0147337252,
-0.0328435935,
-0.0124201328,
-0.0433310717,
0.0647868663,
-0.0189969819,
-0.0231267195,
-0.0043437667,
0.0332846493,
0.0244679898,
-0.0490538478,
0.0520477593,
-0.0068006255,
-0.0312276743,
0.0663139522,
0.0303881932,
-0.0331434235,
-0.0039850185,
0.0189484134,
0.0073409039,
-0.0317697227,
-0.0335017331,
0.0624643676,
-0.0175189003,
-0.0020495255,
-0.0031562471,
-0.0143324034,
0.019180594,
0.0554334037,
-0.0221810807,
-0.0272068046,
0.0515635237,
0.038913656,
0.004625245,
-0.061972592,
-0.0374514945,
-0.0184249599,
-0.0301743317,
-0.0396685489,
-0.0403807163,
0.0061514634,
-0.0007240816,
-0.0265902746,
-0.0688493699,
-0.0258262623,
2.530031415e-34,
0.035139516,
-0.0043470105,
-0.0080102263,
-0.0357969217,
0.0556451567,
-0.0095023336,
-0.0670645908,
-0.0043873433,
-0.0025122438,
-0.0263462607,
-0.025482161
] |
http://arxiv.org/abs/2402.01740v3 | http://arxiv.org/pdf/2402.01740v3 | Reducing Selection Bias in Large Language Models | 2024-01-29T15:43:23 | 2024-06-15T13:23:21 | J. E. Eicher, R. F. Irgolič | cs.CL | cs.CL, cs.AI, I.2.0 | Large Language Models (LLMs) like gpt-3.5-turbo-0613 and claude-instant-1.2
are vital in interpreting and executing semantic tasks. Unfortunately, these
models' inherent biases adversely affect their performance Particularly
affected is object selection from lists; a fundamental operation in digital
navigation and decision-making. This research critically examines these biases
and quantifies the effects on a representative list selection task. To explore
these biases, we experiment manipulating temperature, list length, object
identity, object type, prompt complexity, and model. We isolated and measured
the influence of the biases on selection behavior. Our findings show that bias
structure is strongly dependent on the model, with object type modulating the
magnitude of the effect. With a strong primacy effect, causing the first
objects in a list to be disproportionately represented in outputs. The usage of
guard rails, a prompt engineering method of ensuring a response structure,
increases bias and decreases instruction adherence when to a selection task.
The bias is ablated when the guard rail step is separated from the list
sampling step, lowering the complexity of each individual task. We provide LLM
applications and theoretically suggest that LLMs experience a form of cognitive
load that is compensated for with bias. | Reducing Selection Bias in Large Language Models: Large Language Models (LLMs) like gpt-3.5-turbo-0613 and claude-instant-1.2
are vital in interpreting and executing semantic tasks. Unfortunately, these
models' inherent biases adversely affect their performance Particularly
affected is object selection from lists; a fundamental operation in digital
navigation and decision-making. This research critically examines these biases
and quantifies the effects on a representative list selection task. To explore
these biases, we experiment manipulating temperature, list length, object
identity, object type, prompt complexity, and model. We isolated and measured
the influence of the biases on selection behavior. Our findings show that bias
structure is strongly dependent on the model, with object type modulating the
magnitude of the effect. With a strong primacy effect, causing the first
objects in a list to be disproportionately represented in outputs. The usage of
guard rails, a prompt engineering method of ensuring a response structure,
increases bias and decreases instruction adherence when to a selection task.
The bias is ablated when the guard rail step is separated from the list
sampling step, lowering the complexity of each individual task. We provide LLM
applications and theoretically suggest that LLMs experience a form of cognitive
load that is compensated for with bias. | [
0.0345193297,
0.0021172685,
-0.0442335196,
0.0202034079,
-0.0300877411,
0.0179678295,
0.0193820857,
-0.0041358159,
0.0533381552,
-0.0483701527,
-0.0151793379,
0.010570108,
-0.0013104966,
-0.0360919908,
-0.0128077585,
-0.0790820643,
0.0874013081,
-0.0878354162,
0.0263629556,
-0.0147008346,
-0.037998978,
0.0019222184,
0.0041748737,
0.0329086632,
0.028330151,
0.0016901689,
0.0025190616,
0.0179094337,
-0.0254416801,
-0.0895053446,
-0.0299006253,
0.0757387504,
0.0134713557,
-0.0238586292,
0.0000025193,
-0.0627388433,
-0.0029448657,
-0.0280436799,
-0.0664548576,
0.0035528457,
0.0404024273,
0.0048130299,
0.0021142892,
-0.0093058301,
0.0047984654,
-0.0181419943,
0.0594733879,
0.0121354172,
-0.0243856106,
0.0056358054,
-0.0026720313,
0.0305135604,
-0.022028828,
-0.0070742061,
0.0396147035,
-0.037802197,
0.0180277452,
-0.0071247173,
0.0499982312,
0.0407141969,
-0.0132732624,
-0.0120446626,
-0.005500921,
0.0161654875,
-0.0028353068,
0.0474114902,
-0.0171702318,
-0.033607129,
-0.0190045312,
0.0500966161,
-0.0168946981,
-0.0195878651,
0.0223830361,
-0.0239512399,
0.0243044589,
-0.0247049611,
-0.0596038438,
0.0228261724,
-0.0299796201,
-0.0031333773,
0.0168014131,
0.0182573535,
0.0201674234,
0.0046352739,
-0.0287147462,
0.0625508428,
0.0100738397,
-0.0415939614,
0.0462387092,
0.0232306607,
-0.0021512851,
-0.0875523537,
-0.0107621681,
-0.0042139068,
-0.1132847965,
0.0141204912,
-0.0161540024,
0.0104950024,
-0.002989783,
-0.0329200178,
0.0458293967,
0.064894706,
-0.000176578,
0.0250060465,
0.0026451543,
0.0545124784,
-0.0307665616,
0.0148951998,
-0.0464433208,
0.0350582153,
-0.0179347601,
-0.0154759157,
-0.0478896536,
0.0490763523,
-0.0136289978,
-0.027555583,
-0.0387155972,
0.017639365,
0.0025081788,
0.0056800107,
-0.0347978733,
-0.0198631529,
-0.0052873716,
0.0354894772,
-0.0573525093,
0.0278480649,
-0.0461694896,
0.00551533,
-0.0141345439,
-0.0840435177,
0.0029642051,
-0.0224837773,
-0.0182211827,
-0.0882389173,
-0.0214419048,
0.0894136429,
-0.0112747997,
-0.0354828276,
-0.0498331673,
-0.0017821884,
0.0146185327,
-0.0454258658,
0.0152407214,
-0.0383620784,
0.0005471735,
0.0494055636,
0.0047328565,
-0.0132472944,
-0.0004971404,
0.0419132598,
-0.0412965938,
0.0283725336,
0.0201353356,
0.0073659574,
-0.0000369737,
0.0086401673,
-0.0115781743,
0.0700758845,
0.0095397532,
0.0439118519,
0.024248736,
-0.0370053947,
0.0316444859,
0.0378755853,
-0.0282197986,
0.0023192565,
-0.0608193018,
0.0199926328,
0.0112259109,
0.0059347493,
0.0456644855,
0.0662995726,
-0.0311353598,
0.0299619809,
0.0842540562,
0.0330998972,
-0.0328488462,
0.0135709848,
0.0011931412,
0.0083121732,
-0.0113378083,
0.0357631184,
-0.0685249865,
0.0806063116,
-0.0250895582,
-0.0060571847,
-0.0208336003,
0.0180301871,
-0.0276916213,
-0.0019252719,
-0.0187750887,
0.0188863687,
0.020078985,
0.0320357755,
0.0243326128,
-0.015586406,
-0.0459786765,
0.0449190997,
0.0145692537,
-0.0247650947,
-0.0396360159,
-0.0410458744,
0.035760574,
0.0666232035,
0.0193136148,
-0.0103868339,
-0.0183709059,
-0.0090843933,
0.0155473705,
0.0206138082,
0.1121357381,
0.0150145367,
-0.0327454731,
-0.0386642255,
-0.0584887341,
0.0242731776,
-0.0147166215,
0.0371072851,
0.0052718311,
0.0050176173,
0.0047237328,
-0.0128328428,
-0.0319332629,
0.0228338819,
0.0474584997,
-0.0292586163,
-0.0401007459,
-0.0054037506,
0.0206789132,
-0.0391293801,
-0.0138537576,
-0.0048503913,
0.0199771971,
0.012184863,
0.0491505861,
-0.0067447848,
-0.0562447719,
0.0325083211,
-0.0099157067,
0.0099044126,
-0.0504444726,
0.0255969577,
0.0013215661,
-0.0061991331,
0.022880245,
-0.0196818244,
0.0280302446,
0.03087941,
0.0286137443,
0.0039050824,
-0.0079057664,
-0.0499002971,
0.0856073275,
0.0205138456,
0.0133878682,
0.0749144778,
0.0102114696,
0.0104196602,
-0.0905261189,
-0.0066865524,
0.0471709222,
0.0250223354,
-0.0162262432,
0.0410951488,
-0.017807046,
-0.0293953437,
-0.0367315635,
-0.0814037845,
0.0061585945,
-0.0158495195,
0.0100827822,
-0.0352353193,
0.0003524604,
-0.0145495096,
0.0073681413,
0.0217322502,
0.004220265,
-0.0089562554,
0.0756035,
-0.013278163,
-0.0419708118,
-0.0323526151,
-0.0411211438,
-0.0486228876,
-0.0164249595,
0.0182754248,
0.0279054381,
-0.0034384907,
-0.0225951541,
0.0245087054,
-0.0126813715,
0.0214958582,
-0.00868324,
0.0229765419,
0.039406728,
-0.0361682177,
0.0352023654,
0.0704348162,
-0.0152444122,
0.0494838282,
0.0616486371,
-0.0373498201,
-0.0514681078,
0.0066953506,
0.0083516305,
-0.0078004887,
-0.013580624,
-0.0103530809,
-0.0166346077,
0.0618275702,
0.0063508837,
0.0188996885,
-0.0078724707,
0.032838542,
-0.038209904,
-0.0201221518,
-0.0237486847,
-0.0504359044,
-0.0306513086,
0.087105453,
-0.0288975332,
-0.0431498475,
0.0349643007,
-0.0032485432,
0.0068052365,
-0.0386648849,
0.0315944329,
0.0057945773,
0.0621761605,
0.0300918482,
-0.0114682131,
0.0150080267,
-0.0106909638,
-0.028477978,
-0.0309934989,
-0.0087538669,
-0.0194853954,
0.0472367778,
-0.0719720721,
0.0149385771,
-0.0328371711,
-0.0388860442,
-0.0313314162,
0.0010628186,
-0.0299517661,
-0.0008920918,
0.0327985361,
0.0011040611,
-0.0224161744,
0.0039714645,
0.0603788346,
0.0581411049,
-0.0040433081,
-0.0213792492,
-0.0336577557,
-0.0297532715,
0.042700626,
0.0107469605,
0.0494011939,
-0.0298082661,
0.0247912221,
0.0121022146,
0.0585291833,
-0.0009755305,
-0.0397964902,
0.0482938439,
0.0651076511,
-0.040663328,
-0.0716360584,
0.0850102156,
0.0219217706,
-0.0553676523,
-0.0079693198,
0.1035231352,
0.0256199986,
-0.0754758269,
0.0142944064,
-0.0104997931,
0.0525781773,
0.0327268764,
0.0345594771,
-0.0493191555,
-0.0166642759,
0.0414182432,
-0.00530631,
-0.0419417396,
-0.0251214262,
0.0095528578,
-0.1280139238,
0.0608619377,
-0.0423047505,
-0.0522049069,
0.0192314945,
-0.0324222632,
0.0109224105,
0.0124115767,
0.0268457104,
0.0182531849,
-0.0429932773,
-0.0001456033,
0.0211777426,
-0.0102804871,
0.0149895092,
0.0315849371,
0.0077662221,
0.0224139038,
0.0963587016,
-0.0107329711,
-0.047208827,
-0.0017437838,
0.0986238793,
-0.028124176,
0.010470883,
-0.0599607676,
0.0137124695,
0.0274055116,
-0.0024969336,
-0.0351284295,
-0.0199454129,
0.0166624729,
-0.0201420467,
0.0100263851,
0.0616884343,
-0.0121998498,
-0.0282287672,
-0.0279189236,
-0.0573951937,
0.0058891987,
0.0597020127,
0.0018067688,
-0.0233124215,
-0.0526694655,
-0.0534191653,
-0.0639862269,
-0.0121503687,
-0.0409182347,
0.0267099515,
-0.0303543396,
-0.0244401302,
-0.0149295423,
-0.0262888819,
0.024369888,
0.0347694196,
0.0288221743,
0.0211354494,
0.0106753074,
0.0067494516,
0.0298910271,
0.0329368785,
-0.0296276994,
0.0230815671,
-0.0328144096,
0.0156065747,
-0.0221112259,
-0.0520951524,
-0.0017887251,
0.0207971912,
-0.0108655207,
-0.0303106569,
0.0326332711,
0.0667326301,
0.0412760712,
-0.0030313258,
0.0223049466,
-0.0565882176,
-0.013336054,
0.0139220636,
-0.0055377125,
0.0428516939,
-0.0216985121,
-0.0218303539,
0.004353459,
-0.0138243837,
-0.0354577899,
0.0514794365,
0.0409674495,
0.0192457065,
-0.008707813,
-0.0451964475,
0.0019989107,
0.0337391086,
0.0410892181,
0.0105402963,
0.0369645767,
-0.0102239614,
0.0103167975,
-0.0188533552,
-0.0060176365,
-0.0369755365,
-0.0079694409,
-0.0058094962,
0.0251331199,
0.0169416759,
-0.0474566482,
-0.0135912988,
0.0110017387,
-0.051902093,
0.0127649987,
-0.0079977186,
0.009146505,
0.0466316156,
0.0059148562,
-0.0455102921,
0.0664610192,
0.0223089047,
0.055186417,
0.0014134906,
0.072148934,
0.019644497,
0.0035677883,
-0.0426104702,
-0.0260605011,
-0.0435022675,
-0.0400356576,
0.0284958631,
0.0298944395,
0.0112967454,
-0.0892981812,
0.047817938,
-0.0439368039,
0.0339481346,
-0.0495791733,
-0.0001588735,
0.0463013686,
-0.0272856653,
0.0114745433,
0.0299100149,
-0.0066973022,
-0.0669691563,
0.0451655947,
-0.0081373621,
0.019252358,
0.0090267872,
-0.0361983292,
-0.0220035445,
-0.023097109,
0.0244466886,
-0.0078158816,
-0.02064633,
-0.0269986298,
-0.0798124224,
-0.0415677093,
-0.0243842266,
-0.0115864603,
0.0527243055,
-0.085574463,
0.0611974783,
0.0578230619,
-0.0044359695,
-0.0710879937,
-0.0079548527,
-0.0293066334,
-0.0375243686,
-0.0015134033,
-0.0406621732,
-6.729082881e-33,
-0.0444679894,
-0.0096873324,
0.042305842,
0.0002992077,
-0.0201605409,
-0.0257390868,
-0.0214697067,
0.0322341099,
-0.0083177397,
-0.0293118078,
-0.0295470599,
-0.0352457128,
0.0202386901,
-0.0245466717,
0.0518023223,
-0.0158226248,
0.0334163606,
0.0320913531,
-0.051712174,
-0.0098059103,
-0.0059757256,
-0.0124220559,
-0.0059054121,
-0.0338897929,
0.0222485736,
-0.0078588435,
0.0592354797,
-0.0361631736,
-0.0278636273,
0.0641011745,
-0.0042338064,
0.0394053832,
0.0266055539,
-0.0681546181,
-0.0316187367,
0.0557790101,
-0.0278137699,
-0.0586728416,
0.0458294041,
0.0310456641,
0.0408634357,
-0.0406809747,
0.0153580299,
0.0308197159,
-0.0455048755,
0.0156947747,
0.024858145,
-0.0191124696,
-0.0700651929,
0.0532277562,
-0.0810022876,
0.0162472017,
0.0090913298,
0.017240148,
0.0464822277,
0.0456173681,
-0.0390672162,
0.0240979865,
0.0011176788,
0.0145510929,
0.0260032639,
0.0350533761,
0.0447710976,
0.0502565652,
0.0027982213,
0.0277088787,
0.0456293933,
0.0307213031,
-0.0108961165,
-0.0206012912,
-0.0010491653,
0.0584848821,
-0.0118706366,
-0.0030251134,
0.0627456158,
-0.0072980481,
-0.0726436228,
-0.002471687,
0.0414054058,
0.0532539673,
0.0594049618,
0.0171160325,
-0.0337373167,
-0.0458225384,
-0.0287119318,
0.024633823,
-0.0323544927,
-0.0283303522,
0.0639356747,
-0.0146872029,
0.0090279933,
-0.0009265636,
0.046720095,
0.0222934876,
-0.0680198297,
-0.0281008407,
-0.0375528745,
0.0258509386,
-0.0023009651,
0.0103086643,
-0.0616344586,
-0.0457555689,
-0.0075210636,
-0.00196651,
0.0300459098,
0.0531103835,
-0.0301696304,
-0.0029898109,
-0.0414612293,
-0.00747474,
-0.0303842705,
-0.0249110963,
0.021460196,
0.0171747766,
0.0580089539,
-0.0298470967,
0.0111479722,
0.0103660738,
-0.0111789089,
-0.0912653506,
-0.0385501161,
0.0166049767,
-0.0275718924,
-0.0025357048,
-0.0135609638,
-0.0208341293,
-0.0578938201,
0.039163202,
0.0252812337,
-0.0555049032,
-0.0291239973,
0.0149868373,
3.291e-7,
0.0090819541,
0.0309221856,
-0.0139704356,
0.0628859475,
-0.0115734804,
0.0300673172,
-0.0004835752,
0.0307336096,
0.0558867082,
-0.0030270731,
-0.0075928485,
0.0468337201,
0.007964286,
0.050173752,
-0.0698397979,
-0.0146729182,
-0.0008246386,
0.063605167,
-0.0480217524,
0.036346063,
0.0395552069,
0.0085283816,
-0.0122745698,
0.0154952807,
-0.0157314241,
-0.0061410139,
0.0094905514,
-0.0497910306,
0.0429247059,
-0.0106635457,
0.0368578471,
0.0018092181,
-0.0235225298,
0.0787737519,
0.0364175439,
-0.0365756527,
0.0074434793,
0.0334644727,
-0.0070880651,
0.020927798,
0.0312377457,
-0.0245997459,
-0.0043096291,
-0.0524756461,
0.0662855431,
-0.0585119836,
0.0234059505,
0.0077876281,
-0.0115061728,
-0.0217511076,
-0.0166235864,
0.0181561299,
0.0311395191,
-0.0012439024,
0.0385388546,
0.0175081138,
-0.030143274,
-0.0307828095,
-0.0293857586,
0.0179049056,
-0.0337996632,
-0.093388699,
-0.0600157864,
0.0275788419,
-0.0066466196,
-0.0313918255,
-0.0406253263,
3.343531591e-34,
0.0158852227,
-0.0057835742,
0.038430322,
0.0055454681,
0.0078068594,
-0.0046756677,
-0.0124833044,
0.0531600565,
0.0183046479,
-0.0582314022,
-0.0207430478
] |
http://arxiv.org/abs/2401.16240v2 | http://arxiv.org/pdf/2401.16240v2 | Combining Hierachical VAEs with LLMs for clinically meaningful timeline summarisation in social media | 2024-01-29T15:42:57 | 2024-02-16T12:42:28 | Jiayu Song, Jenny Chim, Adam Tsakalidis, Julia Ive, Dana Atzil-Slonim, Maria Liakata | cs.CL | cs.CL, cs.AI | We introduce a hybrid abstractive summarisation approach combining
hierarchical VAE with LLMs (LlaMA-2) to produce clinically meaningful summaries
from social media user timelines, appropriate for mental health monitoring. The
summaries combine two different narrative points of view: clinical insights in
third person useful for a clinician are generated by feeding into an LLM
specialised clinical prompts, and importantly, a temporally sensitive
abstractive summary of the user's timeline in first person, generated by a
novel hierarchical variational autoencoder, TH-VAE. We assess the generated
summaries via automatic evaluation against expert summaries and via human
evaluation with clinical experts, showing that timeline summarisation by TH-VAE
results in more factual and logically coherent summaries rich in clinical
utility and superior to LLM-only approaches in capturing changes over time. | Combining Hierachical VAEs with LLMs for clinically meaningful timeline summarisation in social media: We introduce a hybrid abstractive summarisation approach combining
hierarchical VAE with LLMs (LlaMA-2) to produce clinically meaningful summaries
from social media user timelines, appropriate for mental health monitoring. The
summaries combine two different narrative points of view: clinical insights in
third person useful for a clinician are generated by feeding into an LLM
specialised clinical prompts, and importantly, a temporally sensitive
abstractive summary of the user's timeline in first person, generated by a
novel hierarchical variational autoencoder, TH-VAE. We assess the generated
summaries via automatic evaluation against expert summaries and via human
evaluation with clinical experts, showing that timeline summarisation by TH-VAE
results in more factual and logically coherent summaries rich in clinical
utility and superior to LLM-only approaches in capturing changes over time. | [
0.0481345095,
0.0827252418,
-0.0040334417,
-0.0318348482,
-0.0047370368,
0.0369102471,
-0.0234156158,
0.0032978454,
0.0043896493,
0.0298223589,
0.0460988209,
-0.0719424859,
0.0060406551,
0.0965980589,
0.0027849746,
-0.1377013028,
0.0191349313,
0.0366807692,
0.0159120224,
-0.0093430411,
-0.0035482217,
-0.0019246381,
-0.0172941517,
0.0195787959,
0.0110800825,
0.0370246246,
0.0055999774,
-0.0023422171,
0.0752543882,
-0.0543937087,
-0.0390773788,
0.0028670777,
0.0484016724,
-0.0261537097,
0.0000019104,
-0.0103577664,
0.0014622533,
-0.0348897055,
-0.0204972085,
-0.0457460806,
0.0331524834,
-0.0663679615,
0.0245837644,
-0.0224536005,
-0.0394774228,
-0.0293267947,
-0.0104847085,
0.0359955728,
0.078617245,
0.0596362092,
-0.0234627519,
-0.0259897597,
0.0615943111,
-0.0567687824,
-0.0007029905,
-0.0467331707,
0.0164879635,
0.0108675975,
0.0428374857,
0.0055102035,
-0.0142265987,
0.0439186022,
0.0357102342,
0.0002814409,
0.0289424416,
0.0537891611,
0.0400312021,
-0.0108476067,
-0.0109727001,
0.0524681136,
0.0349931233,
0.0286000911,
0.0193145219,
-0.0037638985,
-0.0333291925,
0.000018142,
-0.0499871112,
0.0464103185,
-0.0098281493,
0.0075990651,
0.015746871,
0.0706875101,
-0.0141832549,
-0.0031068779,
0.0073236497,
-0.0235082768,
0.0146173621,
-0.0052128551,
-0.0252731554,
0.0027655445,
0.0005741976,
-0.0691134408,
0.0223781504,
0.0426818244,
-0.0602447353,
-0.0107128993,
-0.0054466673,
-0.0465332381,
-0.0010284538,
-0.0386006497,
-0.013989632,
0.0330974348,
-0.0024199542,
-0.0197253525,
0.028327398,
0.0138774384,
-0.0260322932,
-0.0132221775,
-0.0364516079,
0.0052210195,
-0.0326101072,
-0.0002351935,
0.013134174,
0.0113476673,
-0.0148297511,
-0.015377651,
-0.0623474233,
0.0355725661,
0.0097970972,
0.0150900604,
-0.0391716249,
0.0615732707,
-0.0286655724,
0.0157081485,
-0.0308508109,
0.0283783805,
-0.0712844059,
-0.019258434,
0.0136656193,
-0.0018531096,
0.0161505286,
0.0273142327,
0.0425932147,
0.006005906,
-0.0198869687,
0.0098529086,
-0.0032014421,
0.017832391,
-0.0278642792,
-0.0547761358,
-0.0085898517,
-0.006142362,
-0.0094020898,
-0.0423246473,
0.0665067434,
-0.012009955,
-0.0009651212,
-0.0146547547,
-0.0129418774,
-0.0247991048,
-0.0662441403,
0.0278466735,
-0.0416186787,
0.0010134722,
0.0241411626,
0.0301966798,
0.0575698093,
0.0122361677,
0.0435309745,
0.0466066152,
-0.006677676,
0.0031546275,
-0.0077390331,
-0.0093693621,
-0.0012806098,
-0.0562640801,
-0.0424774215,
-0.0328689925,
-0.0652475432,
0.0471222512,
0.0105582867,
0.0567820668,
-0.014583285,
0.020880634,
0.0281695742,
0.0574998967,
0.0171807464,
-0.0194369089,
0.0629203916,
-0.0148137165,
0.0013230561,
0.0387636274,
0.0035283731,
0.0705714971,
0.0144824535,
0.0418111421,
0.0131604364,
-0.0062460057,
-0.0194275044,
-0.0555963628,
-0.0021861889,
-0.0013148637,
-0.0337690674,
-0.0043795425,
-0.0169358384,
-0.0150388284,
-0.0190904606,
0.0421407558,
0.047814589,
-0.0062018535,
0.0339943282,
-0.0117225694,
-0.0000898084,
0.105648011,
0.0281838086,
0.0045634531,
-0.0150653059,
-0.0295984261,
0.0177728888,
0.0338300802,
0.0217188969,
-0.0257402677,
0.0217137244,
-0.0233166367,
-0.0487051494,
-0.0052039525,
0.0048239748,
-0.0151147675,
-0.0266694836,
-0.0211504363,
-0.0130743207,
0.019117035,
0.0047345334,
0.0248963274,
-0.0714030191,
0.0198009815,
0.0152195189,
0.0162454788,
0.0160726439,
0.0499254242,
0.0183903202,
0.017201094,
0.0401216261,
-0.0584519245,
-0.0337454267,
0.0197075456,
-0.05625787,
-0.0700385049,
-0.0071706572,
-0.0414054953,
-0.0045705875,
0.0036338535,
-0.0478377901,
-0.0693682656,
0.0347446986,
-0.0603213124,
-0.0106528215,
-0.0078639062,
-0.0354054347,
0.0474996455,
-0.0041771815,
-0.0296432935,
0.0635346025,
-0.0683971196,
-0.0168826487,
0.0310452525,
-0.0139566045,
-0.0380059965,
-0.0024575358,
0.0008671121,
0.0079781162,
0.018232774,
-0.0176205002,
0.0329356305,
-0.0314073674,
-0.0411666892,
-0.0364647284,
-0.1117712408,
0.0078243408,
-0.029623447,
-0.0068631913,
0.04007544,
0.0462438129,
-0.0148706604,
-0.0284172446,
-0.050548505,
0.0177397095,
0.0428388081,
0.0307956859,
0.0285606142,
0.0652163178,
0.0527193993,
-0.0482957959,
-0.0243008658,
-0.0349799134,
0.0573188402,
0.01088282,
-0.0528652184,
0.0120863635,
-0.0692036077,
-0.0548599921,
-0.0395514853,
-0.0043846145,
0.0115528051,
0.0177052673,
0.0061507146,
0.0790973902,
0.0111878673,
0.0447590649,
0.083445482,
0.0193639416,
0.0100288885,
0.0102985958,
0.0109465448,
-0.0734405965,
-0.0111029409,
-0.0114965262,
-0.0221266765,
0.0025567156,
0.0299011301,
0.0615427308,
0.0349665694,
-0.0419670083,
-0.0004548217,
0.0076149041,
0.010043175,
0.0009426399,
-0.0135232648,
-0.0378448106,
0.0319662988,
-0.0408460014,
0.0007797414,
0.0912526548,
0.0336981677,
0.0187678877,
0.0334508717,
-0.0111580435,
-0.045729164,
0.0382840298,
0.033255972,
0.0564330369,
-0.0135177374,
-0.0398106985,
0.0188225508,
0.0262343455,
-0.0028916078,
-0.0081685279,
0.041707553,
-0.025950294,
-0.0155209014,
0.0270156041,
-0.001898363,
-0.0302984715,
-0.0514526963,
-0.0154945785,
0.0312301423,
-0.0304175653,
0.0350416899,
-0.0737276748,
-0.0219138507,
-0.0002440856,
0.0408175625,
-0.0324729271,
-0.0299131032,
-0.0134454221,
0.0893803462,
0.0429297984,
0.0063630193,
0.0060868603,
-0.0516169146,
0.0345823243,
0.034400668,
-0.0132690556,
0.0220059678,
0.0115502644,
0.0477205999,
-0.0403500944,
-0.0453127362,
0.0250324365,
0.0257189609,
0.027923543,
-0.0289758407,
0.0110684838,
0.0928281322,
-0.0214254633,
-0.0496837534,
0.0282447003,
-0.0013038387,
-0.0071314839,
-0.0112480391,
-0.0187761523,
-0.0404251702,
0.0190828405,
-0.0496348888,
0.0294162743,
-0.018808201,
-0.0516584143,
0.0103678172,
0.0033109963,
0.0116590261,
0.0149772428,
-0.0080104638,
0.0333723575,
-0.0357703567,
0.0245162845,
-0.0084630828,
0.0036902938,
-0.0136902798,
-0.0149587793,
-0.0068465187,
0.0268271454,
-0.0617740564,
-0.0586050861,
0.0034966618,
-0.0074717626,
0.0387940817,
0.0485125333,
0.0599529669,
-0.0031204852,
-0.0177087653,
-0.0462281555,
0.0109240366,
0.0373032317,
0.0060573062,
0.0853189304,
-0.0041955463,
0.0238660183,
-0.0495615341,
0.0662262216,
0.0000096624,
-0.016933376,
-0.0478248,
0.0392281674,
-0.0455614477,
-0.0379779525,
-0.0679690018,
0.0379728898,
0.0171971098,
0.0057727578,
0.0608668029,
-0.0022498509,
-0.0020892555,
0.0078936424,
0.0524237193,
-0.0683135241,
-0.0409796387,
-0.0154309934,
-0.0055324738,
-0.0455045737,
-0.0494386815,
-0.0309172645,
-0.0044694971,
-0.0397924632,
-0.0471252836,
-0.0285673309,
0.0349934064,
-0.028784126,
-0.076270014,
0.0189187657,
0.0285198875,
-0.0033352943,
-0.0062873662,
-0.0172425322,
0.0378519893,
0.0030126388,
-0.0045563979,
-0.0080499323,
-0.0613331459,
-0.0078516426,
0.0052256896,
0.0479660071,
0.012305351,
0.048031494,
0.0395330787,
-0.0172158852,
0.0840105638,
-0.0036769928,
0.0121078454,
0.0122692762,
-0.0275358651,
0.060033977,
-0.0444160998,
0.02465735,
-0.0143246483,
0.0148900347,
-0.0304023698,
0.0182654746,
-0.0037287085,
-0.0337166898,
-0.0192014743,
-0.027966097,
0.0124882916,
-0.0641517565,
-0.027566554,
-0.0078173568,
-0.0513097495,
-0.0344139487,
0.0146930264,
0.0041456223,
0.0042287223,
0.02139109,
0.0064319591,
-0.0349357016,
0.0142630031,
-0.02126709,
-0.033501748,
-0.0816089287,
-0.0077343727,
-0.0266259052,
-0.0077200113,
-0.0231797937,
0.0345656984,
-0.0483158603,
0.0090611558,
0.0028723557,
0.0220010746,
-0.0276056156,
0.0173267126,
-0.0099164173,
0.0461881086,
0.0104087936,
0.0697731897,
-0.100595355,
-0.031851057,
-0.0412025265,
-0.0420648232,
0.0431120396,
-0.0103812804,
0.0676034391,
0.0161621626,
0.0313696861,
-0.0146818506,
0.0065135751,
-0.0266693719,
-0.0367362835,
-0.0132189067,
-0.0085890312,
0.0263805389,
0.0115105044,
0.0624176376,
-0.0350660868,
0.0270007998,
0.0052217301,
0.001568927,
-0.0336769596,
-0.0326795578,
-0.0208197962,
0.0331638716,
-0.0307589304,
-0.0251572933,
-0.0423993953,
0.0006397608,
0.0850161016,
-0.0175613612,
-0.029020559,
-0.0046127047,
0.0365059227,
0.0064768842,
0.0358163603,
-0.015718326,
-0.0339155123,
0.0326138251,
0.0261872355,
-0.008702727,
-0.0373030975,
-6.139306876e-33,
-0.0279653165,
-0.0335896276,
-0.0157296415,
-0.0378078967,
-0.0620699301,
0.0436922051,
-0.1025032103,
0.006602515,
0.0426388383,
0.0141179906,
0.0118220178,
0.0082446076,
0.0047888029,
0.0581945404,
0.033819057,
0.0033198418,
0.0411453806,
-0.0093343677,
-0.0209423006,
0.0744499713,
0.0038340765,
0.0403576717,
0.0289905537,
-0.0698084831,
0.0457226411,
0.0151556255,
0.0315891877,
0.0256755631,
0.0339466557,
0.0262854695,
-0.0298668034,
0.03640173,
0.0376451686,
0.0201674346,
-0.0409442075,
-0.0091647841,
-0.0809886456,
-0.0554704107,
0.0049450914,
-0.0507012792,
-0.076646097,
-0.0232996326,
0.0568046384,
-0.0327711664,
0.0149199739,
-0.0506411232,
0.0390318781,
-0.0089776972,
0.0024538187,
-0.1011293381,
-0.0096006878,
0.0098583642,
-0.0295189489,
0.0057449169,
-0.002800646,
-0.011117612,
0.0056105615,
0.0215206798,
-0.0442110263,
0.0198068712,
-0.028836105,
-0.0173193775,
0.0175943449,
-0.034584187,
-0.0324573815,
0.0589294359,
0.0124605047,
0.0290570706,
-0.0812062845,
0.0614540279,
-0.0345337391,
0.0274810754,
0.0220505279,
-0.0049730968,
-0.0036340198,
-0.0009874521,
-0.0429910794,
0.0717407018,
-0.0384681672,
0.0602670275,
-0.0291203819,
0.014257879,
0.0238761269,
-0.0146606062,
-0.0023715859,
-0.0258201454,
-0.0013118084,
-0.0338860787,
-0.0411399342,
-0.012887883,
0.0074861776,
-0.0036739071,
0.0010645658,
-0.0142765604,
0.0182637256,
-0.0172232427,
0.0248048473,
0.0518288948,
0.0044627595,
0.0072354935,
-0.0107813524,
0.052170556,
0.0422293916,
-0.0096065234,
0.0227288473,
-0.0128113478,
0.0118468972,
0.0274126939,
0.0037293469,
-0.0206140503,
0.0180157237,
-0.0124909896,
-0.0043848925,
0.0057047419,
-0.0102045769,
0.0249064844,
-0.0065432256,
0.0640678108,
-0.0311682988,
0.0295185857,
-0.0360278115,
-0.0241501164,
0.0479972772,
0.0389605761,
-0.0220125802,
-0.0044802604,
-0.0441586152,
0.0586971268,
0.039250657,
-0.0450571254,
-0.0067570321,
0.014094254,
2.626e-7,
0.0492309965,
0.0344932005,
0.0809263289,
-0.027680289,
0.037140213,
-0.0072688665,
0.0067754588,
0.0479947627,
-0.0171820316,
-0.0329009779,
0.0495769344,
0.0603524782,
0.0389689431,
0.0451733656,
0.0251954477,
-0.0309161153,
-0.0813055634,
-0.0504685119,
-0.0519582406,
0.0033280631,
0.1146249622,
0.0199107919,
0.0029965623,
-0.0263420586,
0.0453360714,
0.0396512672,
0.069820933,
0.0237636082,
-0.0437361337,
0.0500508435,
0.0257587079,
0.0099344952,
0.0040023569,
-0.0238185264,
-0.0061903032,
-0.0425344519,
-0.0000476201,
0.042397242,
-0.0141009819,
-0.0480647013,
-0.0185369626,
0.005346179,
0.0258476287,
0.0246190634,
0.034015689,
-0.0607166439,
-0.0635094866,
-0.0038025256,
-0.0138779487,
0.0392840467,
0.0506561957,
-0.0273544639,
-0.0564533211,
0.0657868609,
0.0019073142,
0.0406094566,
-0.0205583759,
-0.0965219438,
-0.0189727452,
0.0290479846,
-0.055054117,
0.0229077339,
0.0214838441,
-0.0087651294,
-0.0067017279,
-0.0343844257,
0.0035811728,
2.100552156e-34,
0.0452858247,
0.0538170636,
-0.0347201116,
-0.020632416,
-0.0024817931,
-0.0101425238,
0.0026639709,
-0.0092532579,
0.0055704974,
-0.0645251498,
-0.0354690887
] |
http://arxiv.org/abs/2401.16209v1 | http://arxiv.org/pdf/2401.16209v1 | MultiMUC: Multilingual Template Filling on MUC-4 | 2024-01-29T15:02:24 | 2024-01-29T15:02:24 | William Gantt, Shabnam Behzad, Hannah YoungEun An, Yunmo Chen, Aaron Steven White, Benjamin Van Durme, Mahsa Yarmohammadi | cs.CL | cs.CL, cs.AI | We introduce MultiMUC, the first multilingual parallel corpus for template
filling, comprising translations of the classic MUC-4 template filling
benchmark into five languages: Arabic, Chinese, Farsi, Korean, and Russian. We
obtain automatic translations from a strong multilingual machine translation
system and manually project the original English annotations into each target
language. For all languages, we also provide human translations for sentences
in the dev and test splits that contain annotated template arguments. Finally,
we present baselines on MultiMUC both with state-of-the-art template filling
models and with ChatGPT. | MultiMUC: Multilingual Template Filling on MUC-4: We introduce MultiMUC, the first multilingual parallel corpus for template
filling, comprising translations of the classic MUC-4 template filling
benchmark into five languages: Arabic, Chinese, Farsi, Korean, and Russian. We
obtain automatic translations from a strong multilingual machine translation
system and manually project the original English annotations into each target
language. For all languages, we also provide human translations for sentences
in the dev and test splits that contain annotated template arguments. Finally,
we present baselines on MultiMUC both with state-of-the-art template filling
models and with ChatGPT. | [
0.0262981858,
-0.0142505253,
0.0087551465,
-0.008416621,
-0.0429639295,
0.0459134318,
0.0058831358,
-0.0100084031,
-0.0451508798,
-0.0249568652,
0.0216087624,
-0.0223898627,
0.0177540034,
-0.0632351488,
0.0401233658,
-0.0822885484,
-0.0030691898,
-0.0469958745,
-0.0306099113,
0.0147411246,
-0.0040905611,
0.0663021356,
-0.0259776153,
0.0158248115,
-0.0368526578,
-0.0706229284,
-0.0001511267,
0.0419747792,
0.0267681014,
0.0002710271,
0.0451692007,
0.0090386439,
-0.0338650867,
-0.0074751102,
0.0000017298,
-0.0365761071,
-0.0279707313,
0.0269059595,
0.0050961524,
0.0469805449,
0.0279151574,
-0.0755776316,
-0.0215710178,
0.0252820626,
-0.0061352174,
-0.030503137,
0.0263152961,
0.0622623712,
0.0545543842,
0.0505420715,
-0.0360139646,
-0.0540704355,
0.0199159738,
-0.0280998796,
-0.0160006359,
-0.0162153803,
-0.0096542016,
0.0301912259,
0.0058245263,
0.0032257342,
0.0255244914,
0.0387439951,
0.0204166882,
-0.0190591943,
-0.0378880613,
0.0098451879,
-0.0355306193,
-0.0232698228,
0.0007342734,
0.0499605499,
-0.0325561874,
-0.0205680113,
-0.0162074137,
-0.0037160171,
0.0123559879,
0.0163328238,
-0.0244301707,
-0.0114428997,
-0.0472435653,
-0.0036211426,
0.0732894316,
-0.0242332481,
0.0211420469,
-0.0032501835,
-0.0145272398,
0.0286937337,
0.0389572307,
-0.0682221651,
-0.0024886006,
-0.0041973973,
0.0094263544,
-0.0921218991,
0.0346410088,
-0.0234082248,
-0.0048589488,
0.0038350576,
-0.0579032898,
-0.0658028796,
0.0287868176,
-0.0357881486,
0.0093257446,
0.0817651749,
-0.0397507697,
0.0701874048,
-0.0579213947,
0.0120097697,
-0.0602166168,
-0.0226224884,
-0.0797905475,
-0.0350790359,
-0.0231888872,
0.0060283891,
-0.0556146838,
-0.0167307835,
-0.038418293,
-0.0218763296,
-0.0354875065,
0.0340540856,
0.0018738764,
0.0344006233,
0.0272865407,
-0.060772866,
-0.0421056785,
0.0128048463,
0.0251610875,
0.0391970202,
-0.0645644143,
-0.041068282,
-0.0316158868,
-0.0657207221,
-0.0055989115,
0.0359948538,
-0.007973603,
-0.0577329658,
0.0049202358,
0.0561062582,
-0.0113543561,
-0.0096631246,
0.016790498,
0.0020947841,
0.0147993024,
-0.0720061362,
0.0073944465,
-0.080561772,
0.0245160628,
0.0442385934,
0.025643941,
0.030044511,
0.0099458667,
0.0030540642,
-0.0180746783,
0.078300491,
0.0176348891,
-0.0198483989,
0.0332132205,
-0.0330123007,
0.0075105797,
0.0848307163,
-0.0091406321,
0.0305112768,
0.0381653868,
0.0274888314,
0.0027174144,
0.0648274571,
-0.0200186521,
0.0397885852,
-0.015260742,
0.0202843267,
0.0665743127,
0.05782599,
0.0184311196,
0.0321562253,
-0.0605439767,
0.0165060125,
0.0161806382,
-0.0218453947,
0.0070617734,
0.0581909902,
0.0740966499,
0.0161682963,
0.0169761218,
0.0497212186,
0.0027477727,
0.0161242913,
0.0492157303,
0.0486733876,
0.0394212157,
0.0512651913,
-0.0247164983,
0.0019319137,
-0.0401828364,
-0.033355929,
-0.010397925,
-0.054497432,
0.0340848975,
0.0290548876,
-0.0574844591,
0.1012804657,
0.0047473814,
-0.0163413715,
-0.0126449317,
0.0424434617,
-0.0168655328,
0.0473692752,
0.0448005795,
0.0010052847,
-0.0922582299,
-0.0418815315,
-0.01139443,
0.0214881171,
0.0684601516,
-0.0076761236,
0.0196439661,
-0.0666806623,
-0.0017391326,
-0.0072760344,
-0.0105534457,
0.0455983728,
0.0171257574,
-0.0057609398,
0.0052120178,
-0.0380621329,
-0.0177999977,
0.0345580205,
0.0113029359,
0.0008704825,
0.0007390199,
0.0018042021,
0.0347759202,
0.0204210058,
-0.0219656136,
-0.0180573612,
0.0103099775,
-0.0476142615,
-0.0077893529,
0.0779335052,
-0.0463202782,
-0.0463796593,
0.0369734243,
-0.0694875419,
-0.052472569,
0.0286034551,
-0.0150154019,
-0.0435526632,
-0.0209037941,
-0.0043810094,
0.0530178063,
-0.0377489701,
0.0367063582,
0.0252451878,
0.0234436058,
0.0322396308,
0.0571620725,
-0.0115738343,
0.0141290892,
0.0368842892,
-0.0520155653,
0.0210510977,
-0.0145570589,
0.0125271138,
0.0152612245,
0.0120348521,
-0.0170535613,
-0.0033400052,
0.0093688406,
-0.0213666148,
-0.008912053,
-0.1211692095,
0.0178861674,
-0.0206558872,
0.0422177091,
0.0278626159,
0.0325620398,
0.0007111273,
-0.015701754,
-0.0227407105,
-0.0873948038,
-0.0266398527,
0.0379956588,
0.0114616901,
-0.0055206786,
0.0424868837,
0.03902613,
-0.0045670657,
-0.0021404042,
0.0438008271,
0.0048496276,
-0.0577932596,
-0.0617622547,
-0.0365190431,
-0.0143875508,
0.0159623045,
0.0180113036,
-0.047024712,
0.0213513039,
-0.0054339236,
0.0722536743,
0.03392075,
-0.0022361539,
-0.0141523741,
0.0054455968,
0.026851818,
-0.0245933272,
0.0501682721,
0.036187198,
0.036626976,
-0.0228377301,
-0.0219167434,
-0.0267834663,
0.0463316739,
-0.0062068654,
0.0044726389,
-0.0606216192,
-0.0650189593,
-0.0170279387,
-0.0027228706,
-0.0087624546,
0.0249219108,
-0.0427438803,
0.0285827611,
-0.0101374323,
-0.0135777844,
0.0454445556,
-0.0379428826,
-0.0316211246,
-0.0443571955,
0.0650158525,
-0.0494631752,
-0.0171217415,
-0.0077602048,
0.0182981957,
-0.0477159657,
-0.0202217866,
0.0008950877,
-0.0027174915,
0.0188010726,
0.0077758394,
-0.0154530173,
-0.0144402254,
-0.0053754942,
-0.0191048793,
-0.0452124029,
-0.0056990399,
-0.0017443661,
-0.0283821262,
0.0151006309,
0.0486007072,
-0.0204123389,
-0.0258576218,
0.0127076693,
-0.0144160679,
0.0544521995,
-0.0116209397,
0.0107298214,
-0.0139114009,
0.0546999052,
0.0058241771,
-0.0153293256,
0.0142430756,
-0.0328377523,
-0.0034200344,
0.047827635,
0.0070802784,
0.0141036464,
-0.0311194919,
0.0036762867,
0.039767202,
-0.0264067706,
-0.0383864231,
-0.0431500711,
0.0255366825,
-0.0201899465,
0.0475061163,
0.0910977125,
0.0078650517,
-0.0285347831,
-0.0315006897,
-0.0178606324,
0.0258759167,
0.0210918039,
0.0331885703,
-0.1295300573,
-0.0032876106,
0.0334810056,
0.0104253171,
0.0105396984,
0.004677813,
-0.0012478152,
-0.0645361096,
0.0397910923,
-0.0019715196,
-0.0371687301,
0.021577742,
-0.0181298032,
0.0212458242,
-0.0169283114,
0.0005685833,
-0.0045804754,
0.0006486549,
0.0042664697,
0.0211128201,
-0.0275228079,
-0.044702746,
-0.0170420408,
-0.033762686,
0.0121551203,
0.0139466068,
0.053677652,
-0.0151146203,
-0.0327790231,
0.0364678912,
-0.0134434402,
-0.0154195055,
-0.013781175,
0.0128420619,
-0.0128533496,
-0.0309010539,
0.01488089,
0.0248761084,
-0.0292103384,
0.011814327,
-0.0028732289,
0.0734087378,
0.0237631593,
0.0512580425,
0.0069010337,
-0.0154231777,
0.027051067,
0.0458365418,
-0.0165647976,
-0.0556696951,
-0.0110481791,
-0.0385256708,
0.0035482685,
-0.0724508837,
0.0419112183,
0.0115848435,
-0.075396426,
-0.0295608882,
-0.0597904511,
-0.0580890663,
0.0545545854,
0.0700306967,
0.0148635209,
0.0030919928,
-0.0424784534,
-0.0268463828,
-0.0123412348,
-0.0205519162,
-0.014269392,
0.0365033112,
-0.0361516215,
-0.0458509699,
-0.0179392602,
-0.0536195301,
0.0015491922,
0.0423549823,
-0.097197853,
-0.028635174,
0.0178808346,
0.0387923457,
0.0139846895,
0.0674358383,
-0.0465203747,
-0.013564378,
0.023200145,
-0.0039382102,
-0.0556573495,
0.0097611789,
0.0324045904,
-0.016321538,
-0.0093857758,
0.0249504726,
0.0042890217,
-0.0182037484,
-0.0165722687,
0.0326189548,
0.026436571,
-0.0144612044,
0.0450891815,
0.014240196,
0.0077951984,
-0.0354284234,
0.0047789831,
0.0093967477,
0.0571066476,
-0.0274352375,
-0.0285349973,
-0.0126684187,
0.0164844748,
0.0320565812,
-0.0007393067,
-0.0008325501,
0.0144377556,
-0.0062634251,
0.0148119144,
-0.0213280395,
-0.0305553786,
-0.0200298317,
0.014758761,
0.0304673165,
-0.0116175441,
-0.0089916168,
0.0238998216,
-0.0378145687,
0.0018246705,
-0.0496248379,
0.0308039449,
0.0181881227,
0.0236134045,
0.0434263013,
0.0246484838,
0.0083710831,
-0.028738739,
-0.0298531242,
0.0146020371,
0.0594235025,
-0.0148325516,
0.0847112536,
-0.0467123576,
-0.049923256,
0.016352348,
0.012238292,
-0.0068289689,
-0.0053971894,
-0.027571572,
0.0150870932,
-0.0061222385,
0.0278794114,
0.0625600889,
-0.038064599,
0.0518037751,
0.0593902431,
0.0057345764,
-0.0454451144,
-0.0000165931,
0.0000549387,
0.02154365,
0.0064995997,
-0.0389493257,
-0.0321283974,
-0.041452799,
0.0264594015,
0.0274416823,
-0.0474703014,
0.0228936952,
-0.008851856,
0.0698103309,
-0.0062218076,
-0.0231241398,
-0.0812001675,
0.0082322042,
-0.0200775247,
-0.0208893698,
0.0135205882,
-5.936848913e-33,
-0.0006810505,
0.0276210103,
-0.0278906729,
-0.0247414652,
-0.016419081,
-0.0099204862,
0.0167702232,
-0.0263093654,
0.0373889618,
0.0211725868,
-0.0195646156,
0.0086574126,
0.0384789295,
0.000248687,
0.0233909041,
0.0257853381,
0.0732168108,
0.0258741491,
-0.0316241346,
0.0062477831,
0.0360715985,
0.0269177761,
0.0946264789,
-0.0352832824,
0.0241252966,
-0.0590939894,
-0.0263106301,
0.0168457851,
0.0135970693,
0.0532518253,
0.0160089806,
0.0034270855,
0.0206473321,
-0.0303966496,
-0.0063239797,
-0.0420638286,
-0.0962083116,
-0.0020542073,
0.0638955832,
0.045947995,
-0.0922611207,
-0.0707460567,
-0.0011625985,
-0.0275452528,
0.036712151,
0.0086114006,
0.0058548194,
-0.0295480452,
-0.021284163,
-0.0164481439,
-0.0286076311,
0.0194420014,
-0.0377536118,
-0.015580981,
-0.0005110608,
0.0636126474,
-0.0026814512,
-0.0063365768,
-0.0281726476,
0.0071930587,
-0.0468152016,
0.0332323126,
0.0368855372,
-0.0179323647,
-0.0122679947,
0.0087546492,
-0.0251415111,
0.0121824918,
0.0431439914,
-0.0220159292,
-0.0368287228,
0.0178878512,
-0.0010809855,
0.0620884411,
0.0200627726,
0.0131902974,
0.0201459695,
0.0513581559,
0.0236959383,
0.0758127421,
0.0004039686,
0.0266379025,
-0.0188693982,
-0.0140927089,
-0.0069225882,
-0.0620809533,
-0.0292343646,
-0.030010568,
0.0008115372,
-0.0209391061,
-0.055560831,
0.0549020059,
0.0446925871,
0.019402504,
0.0188805293,
0.0419405699,
-0.0399775356,
0.0220416207,
-0.0478876643,
0.0596069582,
-0.0058406233,
-0.0348472334,
0.0043353802,
0.0068863644,
0.0484604128,
0.0018367441,
0.0376732461,
0.0440975092,
-0.0746840835,
0.0058671948,
-0.030520102,
0.008735233,
-0.0233212765,
0.0111392038,
0.064978987,
-0.0105356555,
-0.0096845897,
-0.0054106303,
0.0575364269,
-0.0196112376,
-0.0481528305,
-0.0158570781,
0.0082183471,
0.0188163593,
-0.0257127881,
0.0242800042,
-0.0379350521,
-0.0238935817,
0.0545070209,
-0.0474041365,
-0.0017753228,
0.0427006781,
2.45e-7,
0.0389870107,
0.0380162261,
0.0453464463,
-0.0137797007,
0.0208926834,
-0.0058665434,
-0.0070932931,
0.0453253314,
-0.0019555376,
0.0541806705,
0.0247953776,
-0.0167231876,
0.0706354752,
0.0514903404,
-0.1121921912,
0.0477607548,
-0.0615652576,
0.047181312,
-0.0857325941,
-0.0029251103,
0.0215177704,
0.0835720897,
0.0120586259,
0.0102708992,
-0.0101239244,
-0.0075218789,
0.0479491688,
-0.0097445548,
-0.0487300642,
-0.0072711562,
0.0267649703,
0.0011161072,
-0.0323792845,
0.0329905786,
-0.0067492835,
-0.0129185459,
0.0791151449,
0.0625719279,
0.0520160235,
0.0386124291,
0.0021746373,
-0.0368060879,
-0.0151613019,
-0.0740298778,
0.0470157303,
0.0182780288,
-0.0066955374,
-0.0632924065,
-0.0533291921,
0.0142390756,
0.0246430747,
0.0086774398,
-0.0164264962,
0.0511032902,
0.0378459282,
-0.0004212022,
-0.0535709523,
-0.0419363528,
-0.0464189388,
-0.0278213602,
-0.0282085761,
-0.0227283463,
0.0017030487,
-0.0476759411,
0.0337751247,
-0.0419364721,
-0.025819907,
1.936486205e-34,
0.0554303005,
-0.0051197829,
0.0200023744,
0.0275563039,
0.0368282311,
-0.0527711026,
-0.03204358,
0.0513467006,
-0.0258175842,
-0.0608143993,
-0.0198737811
] |
http://arxiv.org/abs/2401.16184v5 | http://arxiv.org/pdf/2401.16184v5 | Realizing Disentanglement in LM Latent Space via Vocabulary-Defined Semantics | 2024-01-29T14:29:48 | 2024-05-26T13:12:35 | Jian Gu, Aldeida Aleti, Chunyang Chen, Hongyu Zhang | cs.CL | cs.CL, cs.LG | Understanding the latent space of language models (LMs) is important for
improving the performance and interpretability of LMs. Existing analyses often
fail to provide insights that take advantage of the semantic properties of
language models and often overlook crucial aspects of language model
adaptation. In response, we introduce a pioneering approach called
vocabulary-defined semantics, which establishes a reference frame grounded in
LM vocabulary within the LM latent space. We propose a novel technique to
compute disentangled logits and gradients in latent space, not entangled ones
on vocabulary. Further, we perform semantical clustering on data
representations as a novel way of LM adaptation. Through extensive experiments
across diverse text understanding datasets, our approach outperforms
state-of-the-art methods of retrieval-augmented generation and
parameter-efficient finetuning, showcasing its effectiveness and efficiency. | Realizing Disentanglement in LM Latent Space via Vocabulary-Defined Semantics: Understanding the latent space of language models (LMs) is important for
improving the performance and interpretability of LMs. Existing analyses often
fail to provide insights that take advantage of the semantic properties of
language models and often overlook crucial aspects of language model
adaptation. In response, we introduce a pioneering approach called
vocabulary-defined semantics, which establishes a reference frame grounded in
LM vocabulary within the LM latent space. We propose a novel technique to
compute disentangled logits and gradients in latent space, not entangled ones
on vocabulary. Further, we perform semantical clustering on data
representations as a novel way of LM adaptation. Through extensive experiments
across diverse text understanding datasets, our approach outperforms
state-of-the-art methods of retrieval-augmented generation and
parameter-efficient finetuning, showcasing its effectiveness and efficiency. | [
0.0163426176,
0.0835168734,
-0.0313729122,
0.0868887231,
-0.0190381836,
0.0219898801,
-0.0247016922,
-0.0121691981,
-0.0052667782,
-0.0360289328,
-0.0117559675,
-0.0424299799,
-0.0112186447,
0.0465322994,
0.0162065495,
-0.0560874343,
0.039344348,
-0.0268415846,
-0.0561696067,
-0.021635741,
0.0050853556,
-0.000997711,
0.0130541436,
-0.0186965764,
0.0361828357,
-0.0260652471,
-0.0155380368,
-0.0176714566,
0.0112549737,
0.0098017678,
-0.0064574294,
0.0355629884,
0.0725284815,
0.0462149233,
0.0000020688,
-0.0131131336,
-0.0202033892,
-0.0313214064,
-0.0079440083,
-0.0099310242,
0.0186869558,
-0.0502282083,
0.0332082063,
-0.0278671142,
-0.0178397018,
-0.0403735191,
0.0636549443,
0.0118298903,
0.057845585,
0.0526463538,
-0.0133504709,
-0.034629792,
0.0177956987,
-0.0434689745,
0.0241314583,
-0.0012928497,
0.0192516875,
-0.0001258745,
-0.0463465117,
0.0039302511,
-0.0715036616,
0.0307830032,
0.0072891088,
0.0271973424,
0.0143957362,
0.0288834888,
0.0166819934,
-0.0287475884,
0.0093567297,
0.0770682693,
0.0116819264,
-0.0257154908,
0.0021767633,
-0.0012130967,
-0.0424195528,
0.0667056516,
-0.0586132035,
0.0145662436,
0.0298686177,
0.0399427153,
0.0463574268,
-0.021689659,
0.0712401122,
0.0282440465,
-0.0666907057,
0.0515460745,
0.0436750874,
-0.0016061226,
-0.0496616103,
-0.0023676245,
-0.031756077,
-0.0002685683,
-0.0146343429,
-0.0005963225,
-0.0142937563,
0.0319837444,
-0.0013479833,
-0.0311039127,
-0.0113876862,
-0.003162808,
0.0001581011,
0.0379468463,
0.029148398,
0.0461431965,
-0.0294460803,
0.0350844301,
-0.0349658914,
0.0153093413,
-0.0522605553,
-0.0402402058,
-0.0361662246,
-0.0501951016,
-0.0077677811,
0.0410025828,
-0.0465135723,
-0.0223344639,
0.0184024815,
0.0455836728,
-0.0250863414,
-0.0153669231,
-0.0596355721,
0.0193231199,
-0.0260184873,
0.0012051199,
-0.020356575,
-0.0143493861,
-0.0153292315,
0.0086913183,
-0.0094114803,
-0.0070823571,
-0.0352609865,
0.0268701389,
-0.01182245,
-0.0500866845,
-0.0285879187,
0.0442126915,
-0.0086876201,
-0.0315387174,
0.0031492519,
-0.0136124957,
-0.020520879,
-0.0197978914,
0.0377855636,
-0.0746833906,
0.03489637,
0.0280301962,
0.0229602996,
-0.0750383437,
0.0025755004,
-0.0222658142,
-0.0634898245,
0.094596304,
-0.040515177,
-0.007098252,
0.0032012812,
0.0375252366,
0.0592939816,
0.0833484158,
0.0131476717,
0.0457558595,
0.050159283,
-0.0080449851,
-0.0128977811,
-0.0004154838,
0.0270413607,
0.0212636404,
-0.0923774838,
0.0125463838,
0.011757778,
0.0014875518,
0.0444871299,
0.0427394174,
-0.0553157218,
-0.0112028969,
0.0343442634,
0.066401504,
0.0610141493,
-0.0016481106,
-0.011782798,
0.0076595936,
-0.0556560867,
0.0234015658,
-0.0293202847,
0.0213534553,
-0.0178833231,
0.0052196854,
0.0754881501,
0.0026271064,
-0.0394639969,
-0.015064721,
-0.0031003628,
-0.0005010098,
0.0432330929,
-0.0035230494,
0.0528739952,
0.0232637823,
0.0146838333,
0.0654852241,
0.0625888631,
-0.0383694619,
0.051117368,
-0.0394304581,
0.0024227586,
0.056478072,
0.0645327643,
-0.0281411652,
-0.0071546366,
0.0116289491,
0.004795149,
0.0601472594,
-0.0138378087,
-0.0111367879,
0.025489768,
-0.0590120964,
-0.0215032659,
-0.0431895629,
0.0179754905,
0.0074942252,
0.0163581129,
-0.0096615506,
0.0177384354,
-0.0495749712,
0.0273152962,
0.0342624411,
-0.0143582439,
-0.0230514575,
-0.0020532997,
-0.0010244857,
0.0486611724,
0.0147899576,
-0.0018514284,
0.0033046261,
0.0380059853,
-0.046922978,
-0.0408227555,
-0.0129143959,
-0.0039955289,
-0.054982014,
0.0091781635,
-0.0488180146,
0.0235098526,
0.0241619255,
-0.0407770127,
-0.0629093796,
0.0784642324,
-0.0046944804,
0.067362763,
-0.0462114625,
-0.0218660571,
-0.0207322799,
0.0332531966,
-0.0628761947,
0.0607695244,
-0.025110539,
0.0412572846,
-0.012537065,
0.0587384515,
0.0055932384,
0.036953833,
-0.030274177,
0.0463464856,
0.0222397484,
0.0232775677,
0.0401536003,
-0.0168419275,
-0.0255713407,
-0.0684776604,
-0.0208384041,
0.009032839,
-0.0352242328,
-0.0037847734,
0.0604717173,
0.0182702392,
0.0183434803,
0.0019291037,
-0.0129946787,
0.0237175841,
0.0155571839,
0.0328150503,
0.022360187,
-0.0082525183,
-0.0255607758,
-0.023364475,
-0.0098546529,
0.0019333009,
0.0442661904,
-0.0441356935,
-0.0192597173,
-0.053396076,
-0.0192117114,
-0.0472441986,
-0.0075518596,
-0.0333517231,
0.0115325376,
0.0319909826,
-0.0424250849,
0.0213785302,
0.0111352894,
-0.0034728525,
0.1036423817,
0.0952237993,
-0.0313222706,
-0.0276871081,
-0.0163311176,
-0.054314997,
-0.0124419676,
-0.0088034067,
-0.0119020985,
-0.0075230054,
0.0696872026,
0.0120632807,
-0.0271246452,
-0.0639512017,
-0.0024709783,
-0.0112329032,
-0.0369387791,
-0.0614160635,
0.023855919,
-0.0140422322,
0.0686165765,
-0.0429240093,
-0.0609068796,
0.0435631089,
-0.0134082306,
0.0077239936,
-0.015654508,
-0.0243389625,
-0.0353250429,
0.099156864,
0.016447233,
0.0040749828,
-0.0530443713,
-0.025409326,
-0.0640051439,
0.0150640653,
0.0167708099,
-0.0283993632,
0.0570763275,
-0.0465429202,
-0.0003968937,
0.0207613446,
0.0213090815,
-0.0302654915,
-0.0348220505,
0.0162429269,
0.0036139269,
0.029273143,
0.0078907358,
-0.0403572731,
0.0108569432,
0.0095006544,
-0.0107745696,
-0.0165625811,
-0.0062052789,
-0.0520242192,
0.0232337564,
0.0964787602,
-0.0152555993,
0.0020839588,
-0.056733422,
0.0243767668,
0.0470409431,
-0.0218705814,
-0.0191103388,
-0.0044247541,
0.0298884641,
0.0591400042,
0.0143028144,
-0.017695073,
-0.0211490281,
0.0051015615,
-0.0112961084,
-0.0388975665,
0.1065642089,
0.0245259609,
-0.0677758157,
0.0114861801,
-0.0194499716,
0.0608808212,
0.0226214733,
-0.0078758514,
-0.0517105795,
-0.0223234277,
-0.008711502,
0.0185154062,
0.0482933261,
-0.0139568066,
-0.0156074194,
-0.0214766618,
-0.0165008269,
0.014936137,
-0.0556611083,
-0.0592533387,
-0.0984834358,
0.0271328799,
-0.0135236494,
0.0477335565,
0.0035356337,
0.0081619285,
0.0310945064,
-0.0218001381,
0.0302087627,
-0.0317695253,
-0.0098856715,
0.0140461707,
0.0415495783,
0.0461421125,
0.0375862829,
-0.0048149009,
-0.0123382825,
-0.0397228636,
-0.0651681498,
-0.0281486865,
-0.0287272558,
0.0363875553,
0.0004240649,
-0.0089239543,
-0.0037874281,
0.0412224717,
-0.0410249345,
-0.0546969883,
-0.0160606243,
0.0272352751,
-0.0340042263,
-0.0165963005,
-0.0107460683,
-0.0216054562,
-0.0061556338,
0.0648423582,
0.030148834,
-0.0373190083,
0.0003149119,
-0.0322048999,
0.0163103864,
0.070564203,
-0.0151593946,
-0.0327572525,
0.0000895968,
-0.0108568128,
-0.0414175615,
-0.0168859903,
0.0118806092,
0.024908375,
0.016465025,
0.0535702258,
0.0045538121,
-0.0581459403,
-0.0698601007,
0.0427887216,
-0.0308854803,
-0.0200596079,
-0.0055911555,
-0.0089256167,
0.0166012645,
-0.0341056585,
-0.0266982913,
0.0055325706,
-0.0619558021,
-0.0941932127,
0.0177907124,
-0.0318541229,
0.0209374465,
0.0581691824,
-0.0116350809,
-0.030342577,
0.0270687081,
0.0377120376,
-0.0162529312,
0.024457572,
-0.0098281736,
0.0226289891,
-0.0003383209,
0.04195996,
-0.0519798845,
-0.0075717065,
0.037335895,
0.044065658,
-0.045430351,
-0.0580678619,
0.0093628177,
0.0017846961,
0.1053509489,
-0.0033074538,
-0.0193640813,
-0.0564314984,
-0.0286462214,
-0.0174943618,
-0.0200929344,
-0.0311120711,
-0.0217457917,
0.0473814532,
-0.0342035517,
-0.015095152,
0.0494976379,
0.019281894,
0.0398129672,
-0.0324502215,
-0.0256266277,
-0.0322372951,
0.0203628875,
0.0019245367,
0.0282252543,
0.0169067942,
-0.0018173415,
-0.0222245697,
0.0167401992,
-0.0181249455,
0.0216566976,
0.0378826372,
-0.0080146538,
-0.0430046581,
0.0342321508,
-0.0910823718,
-0.0286292266,
-0.0136651732,
-0.0060591018,
0.0134641277,
-0.010982845,
0.0650410578,
-0.0318260565,
-0.0198858269,
-0.0033365437,
0.0377352424,
0.0433675647,
-0.0125723388,
0.0427688546,
0.0329941697,
-0.019260928,
0.0115415948,
0.0709456876,
0.0075629815,
0.0133836241,
0.0108563779,
0.0236732308,
-0.0327154435,
-0.0256301071,
0.027374072,
0.0819927827,
-0.0011474437,
0.0506393425,
-0.032526575,
-0.058643654,
-0.0372778289,
0.0050686994,
-0.0401507057,
0.0072235698,
0.0301571265,
0.0683948845,
0.0114958864,
0.0055603487,
-0.0511316881,
0.0207300968,
0.0168365315,
-0.0246742945,
-0.019040281,
-6.887636498e-33,
-0.0521416888,
-0.0251712874,
-0.011727456,
0.0131053226,
-0.0391522385,
0.0118564609,
-0.0254283771,
0.0095102135,
-0.0151584586,
0.0145506104,
0.0079827439,
0.0564387701,
-0.0161078889,
0.0499666296,
-0.0324600264,
-0.0064972523,
0.0253418777,
0.0260724816,
-0.0021857615,
0.0187338106,
0.0723601356,
0.0156811327,
0.0301049929,
-0.0582703017,
0.0558466688,
-0.0204942841,
0.0091094021,
0.0404323675,
0.019895412,
-0.0041329325,
-0.0553855412,
0.0495793149,
0.0199383236,
-0.0203133952,
-0.0306480657,
-0.0211245697,
-0.0696765557,
-0.0681811124,
0.0483236276,
-0.0289003439,
-0.0335528962,
-0.0449682251,
0.0563582182,
-0.0025779512,
0.0400572792,
-0.0112067629,
0.0312962607,
-0.0137426462,
-0.042930007,
0.0281859506,
-0.0450174101,
0.0109113706,
-0.0107415598,
-0.0015395985,
0.0615074411,
-0.0267580058,
0.0124827866,
0.0282001048,
-0.0757001415,
0.028199397,
-0.0153130544,
0.0390364788,
0.0569168888,
0.0812032446,
0.0327756964,
0.0703792199,
0.0198277477,
-0.0299466327,
-0.0474008508,
0.0056395992,
0.0405094065,
0.0092588961,
0.0113879601,
0.0431739278,
0.0387612768,
-0.0047115819,
-0.0625711232,
0.0140248779,
-0.052390866,
0.0271793734,
0.0237070136,
0.0518165864,
0.0028239642,
-0.0541640036,
-0.0837602243,
0.0077572796,
-0.0433693528,
-0.0091862874,
0.0012053696,
-0.0183443148,
-0.0477455147,
-0.0551408269,
-0.0127864089,
-0.0409918055,
-0.0124062225,
0.0250205155,
0.0107393814,
-0.0166116748,
0.0311184935,
0.0161268562,
0.0075471294,
-0.0512134358,
0.0294465683,
0.0048461901,
0.0057177003,
0.0100867422,
-0.0092306938,
-0.0023555749,
-0.0220426004,
-0.0170652941,
-0.0459980071,
-0.0495467,
-0.0329532251,
-0.0322035253,
0.0213997271,
0.015017638,
-0.0085809333,
0.0401732214,
0.0442556851,
0.0018704205,
-0.0033845545,
-0.0154066384,
-0.0159826111,
0.0330970399,
-0.0450202413,
0.0037591986,
-0.004330162,
-0.0013136229,
0.125479579,
-0.0349482149,
-0.0095835133,
0.0071080308,
2.727e-7,
0.0747053698,
0.0396755189,
-0.0116943298,
0.0222837012,
0.0202262998,
0.0251274705,
-0.0197839662,
0.0196916312,
-0.0020476729,
-0.0247731879,
0.0049527325,
0.0708536357,
0.022256976,
0.0130994953,
-0.0114232088,
0.0413682088,
-0.0395842791,
-0.0033795754,
-0.0054642558,
0.0111302892,
0.0466176271,
0.084630616,
0.0396836251,
0.0092174085,
-0.0003482191,
-0.0312398784,
0.0558320917,
-0.0593939461,
0.0112265218,
0.0020498848,
0.0052926396,
0.0354596339,
0.0272431951,
0.0004122927,
0.0142596411,
0.0290886518,
0.0126931677,
0.0391778238,
-0.0083898278,
-0.0183670148,
-0.0305174813,
-0.0237447713,
0.0016409699,
-0.0696938857,
0.0854342878,
-0.0596930943,
-0.0492581315,
-0.0249659959,
-0.0288287047,
0.035484802,
0.0214496292,
0.0046683913,
-0.0476170033,
0.0003127631,
-0.0028233798,
0.0194442645,
0.0124789029,
-0.0738107562,
-0.0169791952,
0.0072389757,
-0.0322201736,
-0.0211123973,
-0.0169551242,
-0.0081499321,
0.0282376837,
-0.0265181996,
0.0176270157,
2.701926571e-34,
0.0159865282,
0.0234558322,
-0.0182880629,
-0.0153565686,
0.0323876366,
-0.0563002415,
-0.0221819021,
0.0154531896,
0.0177299101,
-0.1128756031,
-0.0000746429
] |
http://arxiv.org/abs/2401.16182v1 | http://arxiv.org/pdf/2401.16182v1 | LLaMandement: Large Language Models for Summarization of French Legislative Proposals | 2024-01-29T14:23:51 | 2024-01-29T14:23:51 | Joseph Gesnouin, Yannis Tannier, Christophe Gomes Da Silva, Hatim Tapory, Camille Brier, Hugo Simon, Raphael Rozenberg, Hermann Woehrel, Mehdi El Yakaabi, Thomas Binder, Guillaume Marie, Emilie Caron, Mathile Nogueira, Thomas Fontas, Laure Puydebois, Marie Theophile, Stephane Morandi, Mael Petit, David Creissac, Pauline Ennouchy, Elise Valetoux, Celine Visade, Severine Balloux, Emmanuel Cortes, Pierre-Etienne Devineau, Ulrich Tan, Esther Mac Namara, Su Yang | cs.CL | cs.CL, cs.AI | This report introduces LLaMandement, a state-of-the-art Large Language Model,
fine-tuned by the French government and designed to enhance the efficiency and
efficacy of processing parliamentary sessions (including the production of
bench memoranda and documents required for interministerial meetings) by
generating neutral summaries of legislative proposals. Addressing the
administrative challenges of manually processing a growing volume of
legislative amendments, LLaMandement stands as a significant legal
technological milestone, providing a solution that exceeds the scalability of
traditional human efforts while matching the robustness of a specialized legal
drafter. We release all our fine-tuned models and training data to the
community. | LLaMandement: Large Language Models for Summarization of French Legislative Proposals: This report introduces LLaMandement, a state-of-the-art Large Language Model,
fine-tuned by the French government and designed to enhance the efficiency and
efficacy of processing parliamentary sessions (including the production of
bench memoranda and documents required for interministerial meetings) by
generating neutral summaries of legislative proposals. Addressing the
administrative challenges of manually processing a growing volume of
legislative amendments, LLaMandement stands as a significant legal
technological milestone, providing a solution that exceeds the scalability of
traditional human efforts while matching the robustness of a specialized legal
drafter. We release all our fine-tuned models and training data to the
community. | [
0.0239885151,
0.0869808346,
0.0281894468,
0.0051912791,
-0.0477577746,
0.0010368469,
-0.023719335,
0.0078107021,
-0.031091474,
-0.0673209429,
0.0223776363,
-0.0555374809,
-0.0182590764,
0.0277815778,
0.0302759688,
-0.0874980465,
0.0218251944,
-0.0177821852,
-0.0012255919,
-0.0072266292,
0.0450488962,
0.0174524095,
0.0106902495,
0.0359239168,
0.0143844392,
0.0326078087,
-0.0615792722,
0.0090992637,
-0.0194133725,
0.0167021137,
-0.0370228775,
0.0521654859,
0.0005789101,
0.0453132503,
0.0000016679,
-0.0343555808,
-0.0306895543,
0.0035130368,
0.0205306411,
0.0218932088,
0.0754462928,
-0.0997444689,
-0.0521697775,
0.0240358356,
-0.0145096881,
0.0355447605,
0.0086787585,
0.0157448258,
0.0595163368,
0.0773909017,
-0.0109915994,
-0.0394555964,
-0.0296721235,
-0.0630677491,
-0.0151549308,
-0.007647106,
-0.000347192,
-0.0218337271,
-0.0043446892,
-0.010588794,
0.0172678623,
0.0069437334,
-0.0025004935,
-0.0670641363,
-0.0361175574,
-0.0007326385,
-0.0124524571,
-0.0185704082,
0.0173294675,
0.0505408198,
0.0421817005,
-0.0235461053,
0.0409064181,
-0.0043631843,
0.0016514761,
0.0218476299,
-0.0881398618,
0.0590640157,
-0.0056084231,
0.0138132246,
0.0339641795,
-0.0160937831,
-0.0096322717,
-0.0122958394,
-0.0345202126,
0.0233073365,
0.0047057532,
-0.0534102507,
-0.0032525486,
-0.0020961429,
0.006750843,
-0.0703120232,
0.0410389192,
0.0445353612,
0.0012244931,
0.0195797794,
-0.0092087593,
-0.0408947915,
0.0713918507,
-0.0643173456,
-0.0060959547,
0.0190232061,
0.0075096777,
0.089051798,
-0.014975721,
0.0142419562,
-0.0084527954,
-0.0012711361,
-0.0460142158,
-0.0187946316,
-0.0321478844,
0.0254353303,
-0.0459512621,
0.018282149,
-0.0919981226,
-0.0071901507,
-0.0222719498,
0.0112880385,
0.0249918606,
0.0429084301,
0.0507097207,
-0.0133131752,
-0.005646456,
-0.0393083692,
-0.0525288284,
0.0309690852,
-0.0358666778,
0.0102895163,
-0.0165000651,
0.03130408,
0.0123541383,
0.0313565433,
0.0049654855,
0.0174379144,
-0.0058386922,
0.0532120578,
-0.0352898948,
0.0469774194,
0.025686143,
-0.0566443615,
-0.0118928077,
-0.0686732903,
-0.0207632184,
-0.097488299,
0.0134984124,
-0.031702932,
-0.0097351419,
-0.0417831391,
0.0089604259,
-0.0074758683,
-0.0181103814,
0.011493043,
-0.0412480347,
-0.0011991097,
0.0095883943,
0.0133663686,
-0.004973839,
0.0544565283,
0.0149248717,
0.0338005871,
0.0143846758,
0.0070846109,
0.0055830036,
0.0641394407,
0.0016306592,
0.048538167,
-0.0512547269,
-0.0047697774,
0.0100894608,
-0.0358997509,
0.0417882539,
0.0825047493,
-0.0023869341,
0.0049140053,
0.0417654738,
0.0559742227,
-0.0090421354,
0.0335113369,
-0.0023121878,
0.0293670706,
0.0475051887,
-0.0005561009,
0.0033671011,
0.0495889783,
-0.010490532,
0.0159565676,
0.0287682712,
-0.0259629078,
-0.028144896,
-0.0420645289,
0.0088414261,
0.004851691,
-0.00340317,
-0.0145077286,
0.046252504,
0.040070042,
-0.0660158172,
0.0086828312,
0.0075098216,
-0.059072528,
0.0064312108,
0.0044159479,
0.0080006048,
0.0796673372,
0.0271988213,
-0.0467841327,
-0.0481967703,
-0.0424683131,
0.0100011826,
0.0495591499,
0.0689024776,
-0.0572802462,
-0.0342805088,
-0.0292698722,
-0.049314227,
0.0080569917,
0.014223285,
0.0182726402,
-0.0280046854,
-0.0065683401,
0.0058844234,
-0.016533941,
-0.0065635401,
0.0460664891,
-0.0118202744,
-0.0088836327,
-0.0274254028,
-0.0032211659,
0.0219390653,
0.0252624601,
-0.0126516568,
0.0116449017,
0.0551702715,
-0.0137717286,
0.0101311235,
0.0105315754,
0.0007556851,
0.0124026462,
0.0581533723,
-0.0075675705,
0.0005601592,
-0.0082234377,
-0.0728469938,
-0.0245499294,
0.0366117954,
-0.0102363853,
-0.0154941827,
-0.0373983607,
0.0476773456,
0.0258609429,
-0.0541483909,
-0.0191729069,
0.0451489389,
-0.0189527124,
0.0147701893,
0.0384956561,
-0.0219348036,
0.0242722724,
-0.0444689058,
-0.0186927244,
0.0177706629,
-0.0104644131,
0.0238829739,
0.032463368,
0.0240676682,
-0.0153278233,
-0.0654330328,
-0.0805228278,
-0.0145413382,
-0.0390757509,
0.0419115648,
0.0411156267,
-0.0106382631,
0.005141892,
-0.0350972563,
-0.0335618779,
-0.0176240485,
0.0047027939,
-0.0206107888,
0.0321959257,
0.01781285,
-0.030661134,
-0.0330225676,
-0.0197396725,
-0.0043864944,
0.0211969912,
0.0364034325,
-0.0429779142,
-0.014970053,
0.0269147493,
-0.0406126752,
0.0068324832,
-0.0004953233,
-0.0116095431,
-0.0003408736,
-0.0492806323,
0.0116373515,
0.0052087549,
0.0323733576,
0.0169548187,
0.0386056565,
0.0287677478,
-0.0354414806,
-0.0164126623,
-0.0410825834,
0.0380278938,
-0.0651103854,
-0.0423460752,
0.0336337723,
0.0301680714,
-0.0255282577,
0.0283468775,
-0.0150660314,
0.0711189285,
0.0053371629,
0.0023730246,
-0.0277891438,
0.0037402585,
-0.0152449636,
0.0509904176,
-0.0066492329,
-0.0145056201,
0.0254864059,
-0.0032822993,
0.0313692056,
-0.0324839577,
-0.0086189061,
-0.0322133191,
0.0975749716,
-0.0066455076,
0.0360292979,
-0.0364686698,
-0.0131335491,
-0.0257595312,
-0.0473296046,
0.0500755608,
-0.0248217527,
0.0986325815,
-0.0690907985,
0.0458800271,
-0.0501766466,
-0.0588790402,
0.0012213038,
-0.0439392328,
-0.0173207708,
0.0198467877,
0.0042600078,
0.0031673913,
-0.0696934387,
0.0404647142,
-0.0092055341,
0.1017654091,
0.0493115336,
-0.0097021274,
-0.0073182466,
0.0086291321,
-0.0158510488,
-0.0527199581,
-0.0108551653,
-0.0548981614,
0.0301170982,
0.057433214,
0.026040256,
0.0235037897,
-0.0014261663,
-0.0153537123,
-0.0062774341,
-0.0792344287,
-0.0165244509,
-0.0021200723,
0.0082417317,
-0.0434711315,
0.0141888047,
0.0965375155,
0.0141106239,
-0.0312490594,
-0.0032540315,
0.0018353687,
0.0277026221,
0.0364845544,
0.014815337,
-0.0725416765,
0.0444815531,
0.0238874089,
0.0124512296,
-0.0179092139,
-0.0179845877,
-0.0130038038,
-0.0573381931,
0.0886073261,
0.0114590693,
-0.039474763,
0.0055780183,
-0.0316778943,
0.0149756549,
-0.0136699155,
-0.0359042175,
0.0248358306,
-0.0240726508,
-0.0190946665,
0.0222671814,
-0.0800222382,
-0.0047789845,
-0.0430711843,
-0.0295970924,
0.0374270305,
0.0432703048,
0.049594868,
-0.0266982969,
0.0553831644,
0.0792532563,
0.0059345681,
-0.0234558471,
0.0037633807,
-0.0026761268,
0.0362888165,
0.0053116051,
-0.0185570419,
0.0192053877,
0.0348757245,
-0.0133332331,
0.0630827993,
0.0529779755,
-0.0632725582,
0.0278854184,
-0.0027968052,
0.0125358701,
0.0051889005,
0.0560846813,
0.021917941,
0.0115371859,
-0.0354782231,
-0.0492615402,
-0.0377283096,
-0.0737417489,
-0.0102273878,
-0.0395082161,
-0.0466072448,
-0.0038324036,
-0.0611769892,
-0.0168607328,
0.0422166251,
0.0680300742,
0.021070648,
0.0406148732,
0.0424146652,
-0.0140818525,
-0.0244045034,
-0.0170122255,
-0.019450767,
-0.0497023687,
-0.0636680871,
-0.0834071338,
-0.0067139282,
0.0365541391,
-0.058643654,
-0.0328038856,
-0.0485987328,
-0.0425128527,
-0.0249658115,
0.0170649569,
0.0387180597,
0.0245882254,
0.0120093152,
-0.0111104427,
0.0301312823,
0.0277838111,
0.0262456257,
-0.0208942611,
0.0301793814,
0.0005344423,
-0.0144706843,
0.0327095799,
-0.044136595,
0.0267916992,
-0.0038171527,
0.0247807968,
-0.0195747297,
-0.0336949825,
-0.0027743599,
0.048331093,
-0.040245872,
-0.0236360114,
-0.0480637439,
-0.0521090999,
-0.0009682793,
-0.0325827338,
-0.0105706239,
-0.0187945906,
0.0281880591,
0.0150143243,
-0.0085634608,
0.0136320964,
0.0473002642,
-0.0118822465,
0.0286337659,
-0.0340946801,
0.0333320834,
-0.0123270527,
-0.0003448667,
0.0573177338,
-0.0037470269,
-0.0103437286,
0.0490577221,
-0.0159141961,
0.038371522,
-0.0333091244,
0.0122037465,
0.0242596902,
0.0195433311,
-0.0078527927,
0.0102465525,
-0.0368279666,
-0.0499274693,
-0.0634035841,
0.1086995378,
0.0551534705,
-0.0048977211,
0.0504224487,
-0.0572206974,
-0.0089869546,
-0.0486831181,
0.0181549266,
-0.0066498308,
-0.0214508232,
-0.0124879554,
0.0129515575,
-0.0117340488,
0.0189516079,
0.0511652455,
-0.0222906247,
0.0196536146,
0.0629090369,
-0.0097410912,
-0.0577071421,
-0.0591637045,
-0.0133501207,
-0.0241867453,
-0.0609384552,
-0.0748048797,
-0.0480329394,
0.0296123512,
0.009982558,
0.0247263629,
-0.0288304221,
0.0036200499,
-0.0031386821,
0.0124213174,
0.0116983764,
0.0598482452,
-0.0626999959,
0.0355031677,
-0.043143481,
-0.0349091552,
0.0041544028,
-5.318476436e-33,
-0.0395140573,
-0.0221153572,
0.0150647322,
-0.0007058593,
-0.0157108158,
-0.0403626077,
-0.0412384458,
-0.0216676127,
0.036464382,
0.0070397048,
-0.0083799968,
-0.0063574216,
0.0427501313,
-0.0003326986,
0.0265947971,
0.0027329628,
0.0652714744,
0.0048488528,
-0.0248416029,
0.0084374202,
0.0022821443,
0.0131152142,
0.0214382801,
-0.0440853462,
0.04091217,
-0.0028470631,
-0.0364937186,
-0.028518185,
-0.0014290013,
0.058332257,
-0.0073344535,
-0.0113641387,
0.0174226854,
-0.0301256161,
-0.0354621783,
0.0508012772,
-0.0637648702,
-0.0068121566,
-0.0227183197,
0.0081143361,
0.0130898003,
-0.0249452535,
0.0467279553,
-0.0031180452,
-0.0503212735,
0.0199904963,
0.0240738392,
-0.0342405997,
-0.021287445,
-0.0534942932,
-0.0706825778,
0.0261494163,
0.0117722079,
0.0241227522,
0.0478121601,
0.0426970199,
0.0387456417,
-0.0384721458,
-0.0782916769,
0.0176441409,
-0.0123768076,
0.032671202,
0.0429537855,
-0.0518036298,
0.0013630447,
0.034470208,
0.0386216231,
0.0201499462,
-0.0181545392,
0.0248915832,
0.0074726385,
0.00602159,
0.0483585335,
0.0294218995,
0.0602797046,
0.0153560452,
-0.0659456998,
0.0460824706,
0.0271241888,
0.0352544598,
0.0474166125,
-0.0047740941,
0.0271261763,
-0.0000297293,
-0.0606259517,
-0.0254969504,
0.024128383,
-0.000163491,
0.0241119098,
0.0276007392,
0.0502424054,
-0.0108184032,
0.0475939997,
-0.0183533877,
-0.0148004461,
0.0109568555,
-0.0201973543,
0.0278213415,
-0.0137905134,
0.0188404173,
-0.0684268549,
-0.0221051797,
0.023133615,
0.0332678147,
0.0192517322,
0.0104538854,
0.0259682443,
0.0028977843,
-0.0677286163,
0.0185893252,
-0.0064814813,
0.0472415239,
0.0337835923,
0.0958684757,
0.0285075475,
-0.0399288386,
0.0473041758,
0.014660053,
-0.0157769676,
-0.0387781374,
-0.0048280372,
0.0239041504,
0.0174062829,
0.0108376322,
-0.0417688936,
0.0191442594,
-0.014285828,
-0.0192027856,
0.0674631968,
0.0318224132,
-0.0065262471,
0.0379523151,
2.276e-7,
0.0678997263,
0.0311859567,
0.0484076925,
-0.015806308,
-0.0089607751,
0.0109651526,
-0.0365491994,
0.0280797929,
-0.0460331216,
-0.0014233656,
0.0431612544,
-0.0074933697,
0.0406087637,
0.0512862206,
-0.11544884,
-0.0211150479,
-0.0467608161,
-0.0249155927,
-0.0544401258,
0.0277010854,
-0.0160294604,
0.0918807983,
-0.0059612948,
-0.0291395988,
-0.0218367316,
-0.0150425779,
0.0580514371,
0.004297893,
-0.0436379574,
0.01537205,
0.103002131,
0.0101672485,
-0.0243016761,
0.0008334788,
0.0155474022,
-0.0587802678,
-0.0291261431,
0.0931732655,
0.0006944514,
0.0537174866,
-0.0028944286,
-0.071151644,
0.0089018662,
0.0178151559,
0.0107423728,
0.0096668527,
0.007987299,
-0.0269621946,
-0.0329994857,
-0.0040867669,
0.0215271153,
0.0030437403,
0.0157744959,
0.015485554,
0.0309810992,
-0.0043442361,
0.0081215147,
-0.071190089,
-0.0121112121,
-0.0107902018,
-0.0046900143,
-0.0257769041,
0.0297502484,
0.0394916274,
-0.0024150165,
-0.0760433897,
-0.0009469547,
2.221527055e-34,
0.0691453591,
0.0046976735,
-0.0601782054,
0.017670529,
0.0328842439,
-0.0190757774,
-0.0008819007,
-0.0091358637,
-0.0336726382,
-0.057732489,
-0.0075536999
] |
http://arxiv.org/abs/2401.16167v2 | http://arxiv.org/pdf/2401.16167v2 | "You tell me": A Dataset of GPT-4-Based Behaviour Change Support Conversations | 2024-01-29T13:54:48 | 2024-04-03T09:02:52 | Selina Meyer, David Elsweiler | cs.HC | cs.HC, cs.CL | Conversational agents are increasingly used to address emotional needs on top
of information needs. One use case of increasing interest are counselling-style
mental health and behaviour change interventions, with large language model
(LLM)-based approaches becoming more popular. Research in this context so far
has been largely system-focused, foregoing the aspect of user behaviour and the
impact this can have on LLM-generated texts. To address this issue, we share a
dataset containing text-based user interactions related to behaviour change
with two GPT-4-based conversational agents collected in a preregistered user
study. This dataset includes conversation data, user language analysis,
perception measures, and user feedback for LLM-generated turns, and can offer
valuable insights to inform the design of such systems based on real
interactions. | "You tell me": A Dataset of GPT-4-Based Behaviour Change Support Conversations: Conversational agents are increasingly used to address emotional needs on top
of information needs. One use case of increasing interest are counselling-style
mental health and behaviour change interventions, with large language model
(LLM)-based approaches becoming more popular. Research in this context so far
has been largely system-focused, foregoing the aspect of user behaviour and the
impact this can have on LLM-generated texts. To address this issue, we share a
dataset containing text-based user interactions related to behaviour change
with two GPT-4-based conversational agents collected in a preregistered user
study. This dataset includes conversation data, user language analysis,
perception measures, and user feedback for LLM-generated turns, and can offer
valuable insights to inform the design of such systems based on real
interactions. | [
0.0843554363,
0.016562026,
-0.0288735423,
-0.001017633,
-0.032212723,
0.0144299883,
0.0076120598,
-0.00100575,
0.0554682799,
-0.0125383241,
0.0127778687,
0.0012186542,
-0.0090189287,
0.0519951619,
-0.0038789473,
-0.0824652538,
-0.0030805552,
0.0157594923,
0.0629298687,
0.0227711201,
0.0593743622,
0.0250077359,
-0.0027990153,
0.0619737469,
-0.0226438604,
-0.0337874517,
-0.0054389178,
0.0384258032,
0.0264210086,
-0.0690627918,
0.0156338643,
0.0485313237,
0.0009860833,
0.0380606279,
0.0000021174,
-0.0104109477,
0.0093652913,
0.0020727413,
-0.0342179947,
0.0096126022,
0.0681235343,
-0.0246525351,
-0.0186812747,
0.008390232,
-0.0188865829,
0.015461945,
0.0482191294,
0.0946454257,
0.0759487823,
0.0499501787,
-0.0395969562,
0.0003972991,
-0.0052985316,
-0.0108788647,
0.0095109437,
-0.0111044804,
0.0175046641,
0.0056236102,
0.0450236537,
0.0072883097,
-0.0025367411,
0.0192793272,
-0.0249201953,
-0.0000423066,
-0.0482346974,
0.004498594,
-0.0400655866,
-0.0447533838,
-0.0034024741,
0.0831796825,
0.0446434431,
0.025997607,
0.0054332023,
-0.0318696536,
0.0232098028,
0.030770231,
-0.0338860042,
-0.0040958873,
-0.0068848794,
0.0101382285,
-0.024955688,
-0.005751668,
0.0226622224,
-0.0077605699,
-0.0003207967,
0.0342419222,
0.0018474263,
-0.0312879719,
0.0255377665,
-0.0156931784,
-0.0081533194,
-0.0582137741,
0.0031982637,
-0.0030878729,
-0.0387270786,
0.0352901258,
-0.0099429982,
0.018257454,
0.0211396106,
-0.0541792773,
-0.0014108865,
0.0613375604,
-0.0261027366,
-0.014190576,
0.0537929311,
0.02383627,
0.0090159317,
-0.006123255,
-0.0079589142,
0.0263916384,
-0.0558614135,
-0.0333321951,
-0.0628535897,
-0.0142736025,
-0.0247426704,
-0.0431875698,
-0.0619442724,
0.064717114,
-0.0222184118,
0.0064244024,
-0.0234228279,
-0.0338898934,
-0.0046894937,
0.011691561,
-0.0032114096,
0.0464234687,
-0.0243045855,
0.0110005401,
0.0239435639,
-0.0988555551,
0.0486920215,
-0.0095188906,
-0.0218292866,
-0.026306089,
-0.000622268,
0.050947506,
0.0198937412,
-0.0287074372,
-0.0382611975,
-0.0078078159,
-0.0479888469,
-0.038548816,
-0.042906709,
-0.0528118648,
0.0129280863,
-0.0017978929,
0.0236373022,
-0.0441689938,
0.0435833596,
0.021666795,
-0.0500783511,
0.073942557,
-0.045939371,
0.0004798477,
0.0157101154,
0.0353567638,
0.0392379127,
0.0497395135,
0.0323621184,
0.018214209,
0.0551122576,
-0.0480921306,
0.0446074605,
0.0396852344,
-0.030996874,
-0.0104059568,
0.0202045832,
-0.0222074389,
-0.0108824512,
0.0443582311,
-0.0277622994,
0.043000944,
-0.0304040276,
0.0423444659,
0.1132381484,
0.0743155107,
0.0093282647,
0.0202018004,
-0.0104050618,
0.0517954007,
-0.0147556113,
-0.009900324,
-0.0105847986,
-0.0120036891,
-0.0245728437,
-0.0089462921,
0.0034313966,
0.0073883771,
-0.0398871563,
-0.0295486003,
-0.0133343861,
-0.0240207966,
0.0304872096,
0.0024063119,
-0.0059042103,
-0.023604827,
-0.0135388141,
0.0079467082,
0.0169997122,
-0.0499980822,
-0.0305939633,
-0.030131286,
0.0352578685,
0.1151776686,
0.0330772437,
-0.0097650597,
-0.0138012189,
-0.0242644772,
0.0111388182,
0.0370074958,
0.0171764735,
0.0150693431,
-0.0168059468,
-0.0370269343,
0.007984817,
0.0040546926,
-0.0238991622,
-0.0306950342,
-0.030795997,
0.0336666815,
-0.0015741272,
0.0472444817,
-0.0442573801,
0.0028341,
0.0040159845,
-0.0325921811,
0.0059288442,
-0.0052957498,
0.0529803783,
0.0524314158,
-0.003918088,
-0.0115191685,
-0.0033446571,
0.0367675982,
-0.0066397525,
0.0721117929,
-0.0044296477,
-0.0124832178,
0.022096334,
-0.0161068849,
0.0198532958,
0.0296165179,
0.0008241001,
-0.0175415166,
-0.0108691566,
-0.0668684393,
0.0684091896,
-0.0167177916,
0.0169372372,
0.0179983005,
0.0266869273,
-0.0193936396,
-0.0125044566,
0.0083388081,
-0.0243672598,
-0.0203193221,
0.0016177556,
-0.021484198,
-0.051174365,
-0.0138714071,
0.0358908176,
0.0226380378,
-0.0045650839,
0.0236251187,
0.0176203735,
-0.0471440367,
-0.0401221178,
-0.1330880076,
-0.0291668475,
-0.0083304942,
-0.0243221615,
0.0352566056,
0.020905504,
-0.0304763224,
-0.0228076018,
-0.0540459864,
0.0026073067,
0.0451733544,
0.0432372317,
-0.0052148635,
0.0287069567,
-0.026321901,
-0.0306364615,
0.0048220865,
-0.0145171937,
0.0043133232,
0.0073004989,
-0.0475333445,
-0.0418888107,
-0.0178894792,
-0.0367043726,
-0.0079370253,
-0.0148442835,
-0.0497384332,
0.0424050502,
0.0097657545,
0.0917434841,
-0.0115078101,
0.0184717402,
0.0657225177,
0.0014467068,
-0.0382247195,
-0.0353735909,
-0.0014513076,
-0.0222430825,
-0.0236483365,
-0.0657689944,
-0.0928480849,
0.0522105098,
0.025865728,
0.0352913588,
0.0424984284,
-0.048580043,
-0.0140886987,
0.0206220448,
-0.0173719693,
-0.0093350736,
-0.0121480217,
-0.0338480324,
0.0183709711,
0.0160385594,
-0.0623258203,
0.0315998569,
-0.0175496843,
-0.0329143107,
-0.0378525406,
-0.0381927118,
-0.0043186429,
0.0361255556,
0.001271117,
0.0303825289,
0.0319581181,
-0.0179518275,
0.0653132275,
-0.0713906661,
-0.0269801077,
-0.0064854333,
0.0473853089,
-0.1016980633,
0.0090965163,
0.0405722633,
-0.0118124941,
0.0013302348,
-0.0588078499,
-0.0611355193,
-0.0471734442,
0.0187299848,
-0.0263463762,
-0.0364838578,
-0.0212813001,
0.0561853498,
0.0720745027,
0.002359601,
-0.0305807348,
-0.0013576571,
-0.0435349457,
0.0652439818,
-0.0091651259,
0.0067514479,
-0.0169435833,
0.0712336227,
0.0393808149,
-0.0161712151,
-0.0136395898,
0.0033415256,
-0.0244870465,
0.0050173588,
0.0112524908,
-0.0092074517,
0.0344131328,
0.0480801165,
-0.0142921237,
-0.0207712371,
0.0908616558,
-0.0166096762,
-0.0642331168,
0.036106132,
-0.0565115102,
0.0013376571,
-0.0112360725,
0.0016012889,
-0.0635865554,
-0.0189256128,
-0.01865598,
0.0325280428,
-0.0596929342,
0.0067352573,
-0.0047045927,
-0.030192554,
0.043226596,
0.0194547568,
-0.0094002541,
0.010181766,
0.0244443454,
0.0152429044,
0.0005958795,
0.0045981733,
-0.0291860793,
-0.0153009696,
0.0006004394,
-0.0091083394,
-0.0733980909,
-0.030369442,
0.0236005858,
0.0080020344,
0.0752453431,
0.0618020557,
0.0493405201,
-0.0156935118,
-0.0328767262,
0.0310490131,
0.0360765718,
0.0023940296,
0.0091531696,
0.0441099629,
-0.0084155472,
0.0104486225,
-0.0470990278,
0.0636872277,
-0.0372662693,
-0.0177895408,
0.0475569405,
0.0642495751,
-0.0182105936,
-0.0060515217,
-0.0092358021,
0.0468991585,
0.0642093495,
0.0495809764,
0.0280018579,
-0.075677298,
-0.0044931606,
-0.0232982058,
0.0031939393,
0.0459298901,
-0.0682522729,
-0.0405060984,
-0.0140971672,
0.0326509178,
-0.023733465,
-0.1073243991,
0.0412944518,
-0.0186752267,
-0.0116053065,
0.0042230058,
0.0307040699,
0.0470297858,
0.0373407826,
0.0320313387,
-0.0095008276,
0.0500494987,
-0.0800925419,
-0.0154525684,
0.0173547156,
-0.0266610831,
0.0339049064,
-0.0034071484,
0.0089222379,
-0.0334523804,
0.0326050296,
0.0416800901,
0.0224875752,
0.008105767,
0.0465048589,
-0.0350851789,
-0.0311963093,
0.0166496485,
-0.0306788888,
0.0195320603,
0.0188413896,
0.0884309784,
0.0063033942,
0.0184852686,
0.0066182413,
0.0181426778,
-0.0083319014,
0.0092074582,
-0.0633410588,
-0.0420122519,
0.0134395976,
0.0280339085,
0.0044107102,
-0.0126639325,
-0.010859712,
-0.00676797,
-0.0069631762,
-0.0498715304,
0.0047174641,
-0.0502806306,
-0.0084085912,
-0.0009469534,
-0.0285491068,
-0.0182557385,
-0.0439448096,
-0.0991535038,
-0.0252758637,
-0.0284855217,
0.0013925286,
0.0017422199,
0.018988058,
0.002237041,
-0.0023695885,
-0.011518362,
0.0237357598,
-0.0118978694,
0.0153650297,
0.0188820045,
0.1153681949,
0.0139147583,
-0.0361473784,
0.0522881933,
0.0457455441,
-0.0464979,
-0.0843128487,
0.0032908281,
-0.0458463654,
-0.0088833673,
0.0066447547,
0.0671951249,
-0.0565802045,
0.0020966472,
-0.0538026206,
0.0299549419,
-0.0006568041,
-0.0300071184,
-0.0039318376,
-0.0188413821,
-0.0343772434,
-0.0361674912,
0.0406348929,
-0.0343939476,
0.0416801311,
0.0076382784,
-0.0450439751,
-0.0349513367,
-0.0281392429,
0.0163264703,
0.0135597438,
-0.0308510661,
-0.0269692577,
-0.059630733,
-0.0310476273,
-0.0250387993,
-0.0388872661,
-0.0060965144,
-0.0651943162,
0.0390581638,
0.0338070244,
0.0074886405,
-0.0198276099,
-0.0408352874,
0.0200042091,
-0.0495096669,
0.0149448412,
-0.0514791049,
-5.457569009e-33,
0.0032476566,
-0.0045988942,
0.0084670344,
0.0392824635,
-0.031656947,
-0.034280315,
-0.0311834496,
0.0293599218,
0.0207361989,
0.047003869,
0.0034048774,
0.0467392281,
0.0040261885,
0.0136227515,
0.009754098,
0.0190378185,
0.0502535142,
0.0052813939,
-0.0250951368,
0.0566095337,
0.0343067087,
0.0317342952,
0.0393785425,
-0.1286557317,
0.0594553947,
0.0025387399,
-0.0179904904,
0.0108397016,
0.0017269743,
0.0218939446,
-0.0318919271,
0.062393643,
0.0263007563,
0.0165519565,
-0.0062871212,
0.0506927259,
0.0249016918,
-0.0378404595,
0.0489154607,
-0.0056111016,
-0.021888312,
-0.0403293259,
-0.0011986702,
-0.0329588577,
-0.0312938616,
0.0197632164,
0.0141371125,
-0.0260123089,
0.0128243463,
0.0201081429,
-0.062684834,
-0.0090130642,
-0.0281188134,
0.0044026566,
-0.0476000346,
0.0653285235,
-0.0553569794,
0.0046326974,
-0.0231176913,
0.0314030461,
0.0481241159,
-0.0144997584,
0.0045596319,
-0.0074745161,
0.0095744794,
0.0356451124,
-0.014094349,
0.0122202868,
-0.029971607,
0.0074615013,
-0.018740721,
0.0184002221,
0.0317254178,
0.0277788639,
-0.0322276019,
-0.0212030336,
-0.0164233577,
-0.005755581,
0.0397343524,
0.0001442751,
0.0327886492,
0.0208073352,
-0.0423120223,
-0.0369671583,
-0.0161858127,
0.0163320526,
-0.019153107,
-0.0868625641,
0.0186219718,
-0.0193201266,
0.0167213045,
0.0099793877,
-0.0189559925,
-0.005475007,
-0.0092052603,
-0.0082407333,
-0.015339057,
0.022159975,
-0.0021605629,
-0.0230459087,
-0.0611471683,
0.0231361631,
-0.0080464538,
0.0349424221,
0.0503704473,
0.0101866089,
0.0131048663,
0.0217862185,
-0.0421300754,
0.0021359674,
0.0071016974,
0.0522108376,
-0.0145537211,
-0.0196711365,
-0.0296125524,
0.010825756,
-0.0307717472,
0.0605299994,
-0.0179293118,
-0.0234356578,
-0.0606888011,
-0.041142568,
-0.0058191237,
0.0239883419,
-0.034267962,
-0.0210709516,
-0.0308687519,
0.0449905023,
0.0918887109,
-0.0169345699,
0.0206446089,
-0.0312989578,
2.764e-7,
-0.013296891,
0.0481989421,
-0.0036907641,
-0.0030543408,
0.0241515767,
0.0344049707,
0.0083137648,
0.0387731493,
0.0264713597,
0.0213048141,
0.0279712044,
0.0029788213,
0.0359078832,
0.0282053575,
-0.0047670924,
-0.0057797143,
-0.052500084,
-0.0361471586,
-0.0810977891,
0.0319773704,
0.1012375653,
0.0486687496,
0.0181439128,
-0.0080837226,
-0.0030822556,
0.0273875427,
0.0359433927,
-0.098516725,
-0.005311972,
0.0160498507,
0.0113999071,
-0.0311790276,
0.0010663344,
-0.0139608644,
0.0177086983,
-0.0327648334,
-0.0258267913,
0.074893102,
-0.0301408116,
0.0023470579,
0.0001710548,
0.0217453167,
-0.0118151,
-0.0345712751,
0.009559405,
-0.0724095032,
-0.0548480377,
0.0244170595,
-0.0141308242,
-0.0928506851,
0.0105885379,
-0.0460160561,
-0.0361764655,
0.0308710467,
0.0100849802,
0.0084835235,
0.0143210609,
-0.0397286788,
0.0377981663,
0.0034341982,
-0.0442751199,
-0.0422488041,
-0.0598694943,
0.0129487058,
-0.039229542,
-0.06933669,
0.0385863408,
1.998932735e-34,
0.0256278478,
0.0474041998,
0.0173041336,
0.0141835492,
0.0471207872,
0.0522298515,
0.0507084876,
0.0545471907,
0.0190723985,
-0.0972557142,
-0.0228712428
] |
http://arxiv.org/abs/2401.16158v2 | http://arxiv.org/pdf/2401.16158v2 | Mobile-Agent: Autonomous Multi-Modal Mobile Device Agent with Visual Perception | 2024-01-29T13:46:37 | 2024-04-18T06:53:38 | Junyang Wang, Haiyang Xu, Jiabo Ye, Ming Yan, Weizhou Shen, Ji Zhang, Fei Huang, Jitao Sang | cs.CL | cs.CL, cs.CV | Mobile device agent based on Multimodal Large Language Models (MLLM) is
becoming a popular application. In this paper, we introduce Mobile-Agent, an
autonomous multi-modal mobile device agent. Mobile-Agent first leverages visual
perception tools to accurately identify and locate both the visual and textual
elements within the app's front-end interface. Based on the perceived vision
context, it then autonomously plans and decomposes the complex operation task,
and navigates the mobile Apps through operations step by step. Different from
previous solutions that rely on XML files of Apps or mobile system metadata,
Mobile-Agent allows for greater adaptability across diverse mobile operating
environments in a vision-centric way, thereby eliminating the necessity for
system-specific customizations. To assess the performance of Mobile-Agent, we
introduced Mobile-Eval, a benchmark for evaluating mobile device operations.
Based on Mobile-Eval, we conducted a comprehensive evaluation of Mobile-Agent.
The experimental results indicate that Mobile-Agent achieved remarkable
accuracy and completion rates. Even with challenging instructions, such as
multi-app operations, Mobile-Agent can still complete the requirements. Code
and model will be open-sourced at https://github.com/X-PLUG/MobileAgent. | Mobile-Agent: Autonomous Multi-Modal Mobile Device Agent with Visual Perception: Mobile device agent based on Multimodal Large Language Models (MLLM) is
becoming a popular application. In this paper, we introduce Mobile-Agent, an
autonomous multi-modal mobile device agent. Mobile-Agent first leverages visual
perception tools to accurately identify and locate both the visual and textual
elements within the app's front-end interface. Based on the perceived vision
context, it then autonomously plans and decomposes the complex operation task,
and navigates the mobile Apps through operations step by step. Different from
previous solutions that rely on XML files of Apps or mobile system metadata,
Mobile-Agent allows for greater adaptability across diverse mobile operating
environments in a vision-centric way, thereby eliminating the necessity for
system-specific customizations. To assess the performance of Mobile-Agent, we
introduced Mobile-Eval, a benchmark for evaluating mobile device operations.
Based on Mobile-Eval, we conducted a comprehensive evaluation of Mobile-Agent.
The experimental results indicate that Mobile-Agent achieved remarkable
accuracy and completion rates. Even with challenging instructions, such as
multi-app operations, Mobile-Agent can still complete the requirements. Code
and model will be open-sourced at https://github.com/X-PLUG/MobileAgent. | [
0.0316514,
-0.0302124266,
-0.0307924934,
-0.0073985732,
-0.0372666307,
0.0070663188,
0.0479800962,
-0.0047885496,
0.0556805395,
-0.0951013267,
-0.0311850924,
0.0050643566,
0.0050507747,
-0.0117686996,
-0.008453846,
-0.0260684099,
-0.0001876184,
-0.0228365399,
-0.0262209438,
0.0331638865,
0.013650991,
0.0057240869,
0.0182334352,
0.0387548357,
-0.0134576429,
-0.0384143628,
-0.0305152535,
0.0547956936,
-0.0184296314,
-0.0326527171,
-0.0346166529,
0.031672895,
0.0003382344,
0.0008889751,
0.0000021217,
-0.0362376161,
-0.0032684433,
-0.0177635606,
0.0247420911,
-0.0004046675,
0.0409650803,
0.02573663,
-0.0238689817,
0.0208716709,
-0.032935515,
-0.007729758,
0.0286105145,
0.0090198833,
0.0459335521,
0.038249135,
-0.0128792208,
0.0249776654,
-0.0189219434,
0.058310058,
0.0434868224,
-0.0395909473,
-0.0301971119,
0.0120206941,
0.0328815691,
-0.0228102375,
0.0296340976,
-0.0037220814,
0.0221301634,
0.0270453915,
-0.044273667,
0.0176336281,
0.0340819582,
-0.0585743301,
-0.0422786064,
0.0224980991,
-0.0278613437,
-0.000035855,
0.0354931056,
0.0079703452,
-0.0173217859,
-0.0086778756,
-0.016439382,
0.0218498185,
-0.0387019999,
-0.0463550352,
-0.0266747121,
-0.0342497416,
0.0102783991,
-0.0094440989,
-0.0510974303,
0.0076483474,
0.0297493096,
-0.0266942512,
-0.01729865,
-0.0237046219,
0.0007813482,
-0.0180438273,
0.0510151573,
0.0576507896,
-0.0541276187,
0.0173997227,
-0.0445830896,
-0.064962171,
-0.017392071,
-0.0696332529,
0.0236777849,
0.014697602,
-0.0433921516,
0.0396720842,
-0.0152546763,
0.0668337196,
-0.0050337589,
-0.0215075128,
-0.0351639725,
0.0045269872,
-0.0451227166,
0.0137914419,
-0.0048002382,
-0.0166254379,
-0.0779685676,
-0.0202986263,
-0.0132407034,
0.0578759052,
-0.0037266868,
0.0076760002,
-0.0320479199,
-0.0020331915,
0.0090120425,
-0.0154993339,
-0.002608005,
-0.0052141845,
-0.0452140234,
-0.0003354005,
-0.0140800383,
-0.0630120784,
0.0706884935,
0.0069339019,
0.0062801172,
0.0119437892,
-0.0396864302,
0.0995771885,
-0.0024555936,
-0.0381782204,
-0.0664947554,
-0.0213510636,
0.0270250309,
-0.0323677175,
-0.0323078372,
-0.0786402896,
0.0169265755,
-0.0024804147,
-0.0084512075,
0.0030695135,
0.0256939046,
0.0381674729,
0.0125733931,
0.040918231,
0.0145094758,
0.0256116409,
0.044796396,
0.0209955461,
-0.0225425996,
0.0287350211,
0.016898917,
0.0974657312,
0.0479826257,
-0.0084017431,
0.0317647234,
0.0500573106,
-0.0763198584,
-0.0298937894,
0.0560726449,
-0.0259874538,
0.0082330154,
-0.0056505473,
-0.0607924722,
0.030216258,
0.0096810125,
0.032959044,
0.0833050236,
0.0755844489,
0.0103305345,
-0.0127247646,
0.0018854231,
0.0341540836,
-0.0223068483,
-0.0387019664,
-0.0679727942,
0.0263310764,
-0.0500855818,
-0.0176484473,
-0.0251859389,
0.0575305857,
-0.0146982651,
-0.0030066075,
-0.02037297,
-0.0393495969,
0.0327749774,
-0.0001161771,
0.0458956733,
-0.0137259495,
-0.0001634184,
-0.0087773753,
-0.0003473976,
-0.0008103513,
-0.050054349,
-0.0667206347,
0.0272482708,
0.0576058738,
0.0330389403,
0.0203663204,
-0.0079633398,
0.0074581257,
-0.0038877197,
-0.0372531824,
0.0217186064,
-0.0254766066,
0.0128170038,
-0.0275878143,
-0.0342381895,
0.0208173376,
0.031489674,
0.0067956015,
-0.008683729,
0.0107709356,
0.0362075083,
-0.0472838916,
0.0061541223,
0.0387934111,
0.0175639428,
-0.0340762772,
0.0155701851,
0.0540298745,
-0.0024661284,
-0.0044089686,
0.0135981552,
-0.0404583924,
0.0313528404,
-0.0164875221,
0.0751885548,
0.0020373396,
0.0244379472,
0.0411409885,
0.0411085188,
-0.0181899816,
0.0207559131,
0.0478796661,
0.0283748135,
-0.0158490743,
0.0059372559,
-0.0014874424,
0.0302873496,
0.0157969948,
0.069666855,
-0.0156481192,
-0.0051902761,
-0.0262475815,
0.0300576612,
0.0409696065,
-0.0106387269,
0.032253433,
0.0213110372,
-0.0019167186,
0.0435187593,
0.0015843932,
0.0028512811,
-0.0041787741,
0.0408334471,
0.0094331196,
0.0120530967,
-0.0150967287,
-0.0355767198,
-0.1138335466,
0.0201253723,
-0.0516316034,
-0.0085875681,
0.0701403394,
0.0845928341,
0.0055630757,
-0.0369267873,
0.0225834176,
0.0656805038,
0.0015099579,
0.0409678705,
-0.0257417839,
-0.0331513807,
-0.0358011834,
-0.0099026179,
0.0096291583,
-0.0142791299,
0.0259846598,
-0.0022429877,
-0.0080385124,
-0.0408019722,
-0.0288816877,
0.0048876638,
0.0039662798,
-0.0229412541,
-0.0432702489,
0.0214583892,
-0.0212401096,
0.0188471396,
0.0896469727,
-0.0093780635,
0.0528991446,
0.0503837392,
-0.031382557,
-0.0329245478,
-0.0127904331,
0.0429477394,
-0.0024632229,
-0.0827146918,
-0.0351021513,
0.0186726097,
0.0723079592,
-0.0254966766,
0.0522315167,
-0.0030368424,
-0.0393060632,
-0.0398795456,
-0.0306522436,
-0.0078129554,
0.0731207728,
-0.056483943,
0.0679618046,
-0.0400322191,
-0.0473900251,
0.035627611,
-0.0272442792,
-0.0008568444,
-0.0050495951,
-0.0324817896,
0.020362407,
0.1034678966,
-0.0095234457,
0.032557413,
0.0035604469,
-0.0260624494,
-0.0232685506,
0.0241798274,
0.0058794045,
-0.03266792,
0.021215504,
-0.0828232765,
-0.0187426973,
-0.014066929,
-0.0119330259,
0.0085992292,
-0.0145469829,
-0.0320902057,
-0.0533071533,
-0.0225545429,
-0.0063704029,
-0.074690029,
0.0358222835,
0.0517142601,
0.0678114295,
0.017060319,
-0.0192239694,
-0.022977842,
-0.0210278947,
0.0144952489,
0.0220319126,
-0.0506091602,
-0.0225395374,
0.0248176549,
0.0291411206,
0.0277594998,
-0.0227661133,
-0.0440664813,
0.0123837553,
0.03715856,
0.0236503109,
-0.0385532491,
0.0978743732,
0.0396776646,
-0.070558928,
0.0111461189,
0.0777623579,
0.0432066992,
-0.051843863,
0.0020892601,
-0.0157074314,
-0.0186085198,
0.0305040814,
0.0413488187,
-0.0670319274,
-0.0633650348,
-0.0085177012,
-0.0181908514,
0.0542631857,
0.0213731322,
0.0294717271,
-0.0019782849,
-0.0021128627,
0.0034711312,
0.0175655335,
-0.0025579366,
-0.038233798,
0.039344795,
0.0067465012,
-0.0224770997,
0.0154077727,
-0.0062932726,
-0.0346147679,
-0.0453914888,
-0.0316813253,
-0.0242402591,
0.0346758217,
-0.0332413651,
-0.0162560549,
0.1126105487,
0.019008372,
0.0330454968,
-0.0348024033,
0.0227563251,
0.0282298904,
0.0106688561,
-0.0009278617,
-0.0361171924,
0.0081025176,
0.0221212376,
-0.013586198,
0.0253211223,
-0.0482465103,
-0.0333698951,
0.0287124496,
0.0722940862,
-0.0108643444,
0.0064893533,
-0.0419037901,
-0.0621345043,
0.0102071119,
-0.0049417317,
-0.0206014533,
-0.0520888381,
-0.0764766783,
-0.0413252227,
-0.051333081,
0.0322993882,
-0.0093904603,
0.0503150746,
-0.0148505494,
-0.0678535476,
-0.0267761033,
-0.0095066056,
0.033860147,
0.058980152,
0.0399078205,
-0.0084880786,
0.0330155715,
0.051469937,
-0.0064953747,
-0.0134917982,
0.0182389058,
0.0465159267,
-0.0143913776,
0.0310557745,
0.0467449017,
-0.0855170786,
0.0079484507,
-0.0179040506,
-0.0490766466,
-0.0388479717,
-0.0064618997,
0.0224476084,
0.0166289769,
0.0297485124,
-0.0576994233,
-0.0640673041,
-0.0001797698,
-0.006295064,
-0.090913482,
0.0362127274,
-0.0059465594,
-0.0127025219,
-0.0227187853,
0.0008324984,
-0.0831270888,
-0.0057991114,
-0.0194251016,
0.014130339,
-0.0456805825,
-0.033092346,
0.0365754664,
0.0159684643,
0.0107040145,
0.0032930248,
-0.0235947017,
0.0096643595,
0.029166894,
0.0243404116,
-0.034725517,
-0.0293779969,
0.0149747161,
-0.0380381718,
0.026850909,
0.0630087554,
-0.01212015,
0.0352235176,
-0.028946666,
0.0063608941,
0.0530729853,
0.0158659499,
0.014084422,
0.0254108496,
0.0144274561,
-0.003349242,
0.034957774,
0.0069156601,
0.0569241419,
-0.0497064479,
0.0492134318,
-0.0315627903,
-0.0448296405,
-0.0375373662,
0.0420169719,
-0.0108503597,
-0.0148912035,
0.0068567027,
0.0331905819,
0.0472971573,
0.0082587553,
0.0417603925,
-0.0414502248,
0.0239861161,
-0.0671539456,
0.0143603226,
-0.0084592197,
0.000211955,
-0.0072050937,
0.0199384782,
0.0189995356,
-0.0194598511,
0.0161736794,
-0.0357732251,
0.0124447551,
-0.0012812708,
-0.048122678,
-0.0618742853,
0.0015020918,
0.0336032771,
0.0462152027,
-0.0448853374,
-0.0111615555,
-0.0131804077,
-0.014285598,
-0.0377076529,
0.0345478319,
0.0034898329,
-0.0541309267,
-0.0131283039,
0.1022570655,
0.0428677388,
-0.0104804728,
-0.0665093586,
-0.0445289165,
-0.0328742191,
0.0456906483,
-0.0279562976,
-5.213893071e-33,
-0.0128222769,
-0.010819803,
0.0302119739,
0.0066830004,
0.0333217494,
-0.0323906913,
-0.007353799,
0.0332427099,
0.0148872444,
-0.0312154256,
-0.0181750916,
-0.0222666282,
0.0106602022,
0.0018135185,
0.0031009382,
0.0525614098,
0.0282940529,
-0.0121213906,
-0.0657375306,
-0.0062356396,
0.0013632632,
0.029772155,
-0.0064882324,
-0.072920084,
-0.0343422443,
-0.0022333728,
0.007752371,
-0.0203728937,
0.0035906681,
0.0712844878,
-0.021185251,
0.0017086765,
0.0436661839,
0.0105462875,
-0.0233206917,
0.0578980707,
-0.0244786795,
-0.0004483085,
-0.0032847447,
0.0982467756,
-0.0557282194,
-0.0133643439,
0.0374918878,
-0.0079549188,
-0.0358667746,
0.0524869412,
0.0578055345,
-0.0344443768,
-0.0269776545,
0.0222879145,
-0.0482307859,
0.0290863384,
0.0434518605,
-0.0454614721,
0.0287777092,
0.0259597395,
-0.0422290415,
-0.0088258181,
-0.019033432,
0.0092574963,
-0.0510660298,
0.0177775212,
0.074312225,
0.0642181784,
0.0099146422,
0.0281628072,
0.0276050214,
0.0003676417,
0.0035772016,
0.0206590723,
0.0104330257,
0.0784038231,
-0.0454358906,
0.040168643,
-0.003383575,
-0.0196465589,
0.0102985092,
0.0193623882,
0.0387884863,
0.01594989,
0.030014541,
0.0329419822,
-0.0132009229,
0.0298783574,
-0.0175812226,
0.0374163426,
-0.0384609066,
-0.0556234941,
0.0472581498,
0.0493796729,
-0.0065702591,
-0.0309130475,
0.0409366749,
-0.0069092726,
0.0073977672,
-0.0654566363,
-0.0047572167,
0.0552244112,
-0.0181524865,
0.0077251056,
-0.0154644884,
-0.0365096219,
-0.040662393,
-0.0218010619,
0.0526978932,
0.0297598802,
-0.0681313127,
0.0305691417,
-0.0249578524,
0.0038906678,
-0.0348532796,
-0.0262536816,
0.007984031,
0.0066733435,
0.0366937481,
-0.0609983318,
-0.0353353731,
0.0609065592,
-0.0484875776,
-0.0792734399,
-0.0176547207,
-0.0733178779,
0.037882816,
-0.0159917679,
-0.0654067546,
-0.0435631275,
-0.116937615,
0.014745093,
0.0205661487,
0.0092732487,
-0.0073699025,
0.0390506238,
2.873e-7,
-0.0392787904,
-0.0095133837,
-0.0508941039,
-0.0151202129,
0.0356608257,
0.0769492313,
-0.0062657436,
0.0219425578,
-0.03337273,
-0.0575148463,
-0.0395081379,
-0.0188781079,
0.0325338542,
0.0364873856,
-0.0130627491,
0.0423115008,
0.0441123098,
0.0370076969,
-0.0467081703,
0.0543635264,
0.0548995323,
0.0755550042,
0.0041124397,
0.0012427381,
0.0452428013,
-0.0262787882,
-0.0178644322,
-0.0207452755,
0.009198918,
-0.0078853415,
-0.0195472967,
0.0340654813,
-0.02815325,
-0.0093229385,
0.0012089398,
-0.007871558,
-0.0137015041,
0.0542577952,
-0.0203612056,
0.0122322356,
0.0231919941,
-0.0461177714,
-0.022963278,
-0.0201552585,
0.0533775091,
0.0254210867,
-0.0055022785,
-0.0200976022,
-0.009855574,
-0.019382244,
-0.0599604808,
0.0016334917,
0.0053026052,
0.003610932,
0.0122404806,
-0.0290536359,
-0.0527708568,
-0.0670446604,
-0.0037741382,
-0.0042613521,
0.0050760647,
-0.0450005792,
-0.0797659755,
0.0104078948,
0.0518073514,
-0.0354801491,
0.0034146488,
2.479781557e-34,
-0.0192246027,
-0.0109225307,
-0.0073953676,
0.0548110642,
0.0354929082,
0.0221131388,
-0.0140190283,
0.0175001845,
0.0330046043,
-0.0417991579,
-0.0055990834
] |
http://arxiv.org/abs/2401.16137v1 | http://arxiv.org/pdf/2401.16137v1 | X-PEFT: eXtremely Parameter-Efficient Fine-Tuning for Extreme Multi-Profile Scenarios | 2024-01-29T13:13:32 | 2024-01-29T13:13:32 | Namju Kwak, Taesup Kim | cs.LG | cs.LG, cs.AI, cs.CL | Parameter-efficient fine-tuning (PEFT) techniques, such as adapter tuning,
aim to fine-tune a pre-trained language model (PLM) using a minimal number of
parameters for a specific task or profile. Although adapter tuning provides
increased parameter efficiency compared to full-model fine-tuning, it
introduces a small set of additional parameters attached to a PLM for each
profile. This can become problematic in practical applications with multiple
profiles, particularly when a significant increase in the number of profiles
linearly boosts the total number of additional parameters. To mitigate this
issue, we introduce X-PEFT, a novel PEFT method that leverages a multitude of
given adapters by fine-tuning an extremely small set of compact tensors for a
new profile, which serve as binary masks to adaptively select the given
adapters. To efficiently validate our proposed method, we implement it using a
large number of trained or untrained (random) adapters. We evaluate the
performance of X-PEFT through LaMP and GLUE tasks and demonstrate that it
either matches or surpasses the effectiveness of conventional adapter tuning,
despite reducing the memory requirements per profile by a factor of 10,000
compared to it. | X-PEFT: eXtremely Parameter-Efficient Fine-Tuning for Extreme Multi-Profile Scenarios: Parameter-efficient fine-tuning (PEFT) techniques, such as adapter tuning,
aim to fine-tune a pre-trained language model (PLM) using a minimal number of
parameters for a specific task or profile. Although adapter tuning provides
increased parameter efficiency compared to full-model fine-tuning, it
introduces a small set of additional parameters attached to a PLM for each
profile. This can become problematic in practical applications with multiple
profiles, particularly when a significant increase in the number of profiles
linearly boosts the total number of additional parameters. To mitigate this
issue, we introduce X-PEFT, a novel PEFT method that leverages a multitude of
given adapters by fine-tuning an extremely small set of compact tensors for a
new profile, which serve as binary masks to adaptively select the given
adapters. To efficiently validate our proposed method, we implement it using a
large number of trained or untrained (random) adapters. We evaluate the
performance of X-PEFT through LaMP and GLUE tasks and demonstrate that it
either matches or surpasses the effectiveness of conventional adapter tuning,
despite reducing the memory requirements per profile by a factor of 10,000
compared to it. | [
0.0173273683,
0.0326589383,
0.0079307826,
0.0217692517,
-0.0102082575,
0.0158220939,
0.0354688689,
-0.0068689436,
-0.0312551595,
-0.0392387062,
-0.0506863743,
-0.0342465192,
0.0303712972,
0.0427919738,
0.0414737985,
-0.0523151979,
0.0290272851,
-0.0510715954,
-0.0075370753,
0.0117843384,
-0.0127449892,
0.0304841064,
0.0026889895,
0.0328283086,
0.0053912331,
0.0240258407,
-0.0155056296,
0.0383151211,
0.0299612619,
0.0444859304,
0.038985692,
0.0872222185,
0.0102992104,
0.0067842999,
0.0000021543,
-0.0050197686,
-0.032154981,
-0.0014847289,
0.0131191239,
0.0422790274,
0.0556548461,
0.0147565566,
-0.0322575308,
0.0438493416,
-0.0183445644,
-0.0189724341,
0.0538179167,
0.0222812332,
0.0601227582,
0.0749521926,
-0.0129170539,
-0.0118499286,
-0.0230875202,
-0.0382467285,
-0.043008294,
-0.0194930639,
0.0120917121,
-0.0702218413,
-0.0244687274,
-0.0375999957,
-0.0284473188,
0.0034948997,
0.0013275407,
0.0202812944,
-0.0275989827,
0.0096164066,
-0.0680347756,
-0.0286388472,
-0.0011121091,
0.0737763941,
-0.0914220363,
-0.0369757973,
0.0081870258,
-0.0430013947,
-0.0629816949,
-0.0263445135,
0.0091085741,
-0.0140315974,
0.016535949,
0.0130422916,
0.0171966739,
-0.0266309101,
0.0283511411,
0.0083234739,
-0.0760537535,
0.0118013555,
0.0534599163,
0.0025814287,
0.0082851639,
-0.0242024604,
-0.0029141726,
-0.0100976918,
-0.0095206946,
-0.0155742168,
-0.0354155265,
0.0032873433,
-0.0813808963,
-0.0731780902,
0.0079190563,
0.0029210539,
0.0996595919,
0.0623092949,
-0.009788787,
0.0430591106,
-0.078144066,
0.0203943849,
0.0063659982,
0.0232337993,
-0.0429075584,
-0.0060316878,
-0.0063696867,
0.0262302235,
-0.075920485,
-0.0158700272,
-0.0707672536,
-0.0295799281,
0.0073723542,
0.0058059292,
0.018125996,
-0.0106372358,
0.0141900359,
-0.0093456954,
-0.0516526587,
-0.0163883232,
0.0259124786,
0.0493376069,
-0.0214164089,
-0.0018587494,
-0.0377728231,
-0.0522763878,
-0.0016321297,
0.0417988934,
0.0173637867,
-0.0371941477,
0.0405324288,
0.076288484,
-0.0149648981,
-0.0545632467,
-0.0092473486,
0.0067493031,
0.0083765825,
-0.0742732361,
0.0078070159,
-0.0480086952,
0.020883752,
0.0121078743,
0.0137986252,
-0.0501841679,
-0.0141077498,
-0.0181001276,
-0.0439967178,
0.0478995852,
0.0031266685,
0.0131190121,
0.002547493,
-0.0085080778,
0.0355974659,
0.077453047,
0.0378950648,
0.0261438005,
0.0234207511,
-0.0252383184,
-0.0211020727,
0.06238655,
0.0107141379,
0.0199266821,
-0.0225809738,
-0.0178519525,
0.0467375591,
0.0027811099,
0.0336861126,
0.0430559851,
-0.0341353975,
-0.0083589824,
0.0378139131,
0.0168316886,
0.0602306612,
0.02183396,
0.0396947116,
-0.0017199159,
0.0307357088,
-0.0105968527,
-0.0501665138,
0.0216837693,
-0.0251632612,
0.0111878347,
0.0337560289,
0.0861020759,
-0.0452522114,
0.0006007667,
0.0087917196,
-0.0112894773,
0.00869554,
-0.0009821064,
0.0386598185,
-0.0147555321,
-0.0672709122,
0.0713513643,
0.040405374,
-0.0254783407,
-0.0091448529,
-0.0453372709,
-0.0394227169,
0.0630087107,
0.0256779604,
-0.0051257866,
-0.0494736508,
0.0264045056,
0.0284162201,
0.0468226001,
0.0554558747,
0.020431295,
0.0508729331,
-0.0581230894,
-0.0426797494,
-0.015054333,
0.0044654268,
0.0253894664,
0.0238644332,
0.0001029672,
0.0306438953,
-0.023292765,
-0.0365333483,
0.0315505043,
-0.0483552814,
-0.0204405393,
0.0224280022,
0.0159639288,
-0.0003056724,
-0.0040036445,
-0.0238733795,
-0.0189364552,
0.0273352396,
-0.0142913442,
0.0686185062,
-0.0168236587,
-0.0267695934,
0.0519553125,
0.021150222,
-0.0145998802,
0.0320456512,
0.0128595242,
-0.0492467135,
-0.0780628771,
0.0223748665,
0.0048572319,
0.0604327284,
-0.0288412012,
0.005988434,
-0.0568413287,
0.0332455449,
-0.0217697416,
0.0019549183,
0.0344228558,
-0.0018496891,
0.0197452158,
0.0102126328,
0.0199087411,
-0.0039338754,
-0.0497037135,
-0.0016904909,
0.031949304,
0.0162205454,
0.0086313551,
0.0442687534,
-0.0229443777,
-0.0371501707,
-0.1030784771,
0.0383624695,
-0.0569460727,
0.0055849273,
0.0294661112,
0.024109643,
-0.0118071437,
-0.0295621119,
-0.012786217,
0.0030500886,
-0.0108341165,
0.0120584965,
-0.0183015522,
0.0097078197,
-0.028393263,
-0.0296472479,
-0.0233417116,
0.012424483,
0.0283641331,
-0.0295441188,
-0.0577079207,
-0.0583474785,
0.0325168297,
-0.0020190962,
0.0473593287,
-0.005123816,
-0.0549297519,
-0.0032543412,
-0.0303815026,
0.0573035181,
0.0595064871,
0.0003898431,
-0.0224775169,
0.0504543856,
0.0072487551,
-0.0182937048,
0.0043605408,
-0.0147682335,
-0.023537511,
-0.0492557213,
0.0209918972,
-0.0258284584,
0.0490509868,
-0.0408648886,
0.0044060852,
-0.0769834891,
0.0017779811,
-0.0168309808,
-0.0130065745,
0.0090041431,
0.0225229245,
-0.0148293301,
0.0116211455,
-0.0048440197,
-0.0544118136,
0.0524462238,
-0.0546507128,
-0.0180791803,
-0.0468221568,
-0.0246703345,
-0.0079141362,
0.0162144508,
0.0322867446,
0.0428756252,
0.0086851958,
-0.0132887913,
-0.0601491705,
-0.0197582319,
-0.005858101,
-0.009758682,
0.0434122011,
-0.0405014679,
0.0071477378,
-0.0425034799,
-0.0377211049,
-0.0174235236,
-0.0485539362,
-0.0050326553,
0.0247022342,
-0.0023180982,
-0.0702472255,
-0.0206001829,
0.0180124287,
-0.0108502451,
0.0794238821,
-0.0224219263,
0.0222511888,
-0.0340208225,
-0.0022136129,
0.0718699396,
0.0349031612,
0.0418015532,
-0.0382166803,
-0.0436784886,
0.0039353585,
0.022284314,
0.0090102702,
-0.057938315,
0.0258673765,
0.0236938428,
-0.012848421,
-0.0261906926,
0.0167816635,
0.0402428396,
-0.0585376732,
0.0092714382,
0.0936120152,
0.0200467575,
-0.0465730689,
-0.0152856642,
0.045394998,
0.0564172678,
0.0179985445,
0.0656615421,
-0.0570742264,
-0.0273570288,
-0.0028422258,
0.0433557965,
0.0206168629,
0.0170268789,
-0.0009062756,
-0.0871436596,
0.0030104893,
0.0431131534,
-0.0005699779,
-0.0234702528,
-0.0293595679,
0.0146985054,
-0.0151023623,
0.0475536622,
0.0105603095,
-0.0115646124,
-0.0185336601,
0.0105624441,
0.0204272326,
-0.0903112516,
0.0257208254,
-0.0193529241,
0.0518587455,
0.0155496271,
-0.0405098386,
-0.0045616925,
-0.0110360458,
0.0233836789,
-0.0203232281,
0.0185936466,
-0.028516762,
-0.0024822357,
0.0107448073,
-0.0261244867,
0.0344234146,
0.0216693897,
-0.0354251675,
0.0040765535,
-0.056191273,
0.0377070867,
-0.0174134634,
-0.0009492363,
0.0199916977,
-0.024933463,
-0.0029050524,
0.0460328311,
0.0250765067,
-0.0576178692,
-0.0223840829,
0.0050202571,
-0.022934014,
-0.0845671892,
-0.0208381098,
-0.0094010551,
-0.0079811709,
-0.0010201719,
-0.0123321079,
0.0425241664,
0.0418818109,
0.0666078702,
-0.0084158219,
-0.0177276656,
-0.0176686347,
0.0061939084,
-0.0425143391,
0.0502231233,
-0.0353245288,
0.006472677,
-0.0775402412,
-0.0382704251,
0.0499875024,
-0.023817163,
-0.0202832799,
0.0258201659,
-0.0737612695,
-0.0690104216,
0.0273051336,
0.0454473533,
0.0190972593,
0.0456470139,
-0.0297363922,
-0.0756828412,
0.0523361079,
0.0064069228,
-0.0072004306,
0.0608961321,
0.0370908678,
-0.0712446198,
-0.0356933437,
0.0287076719,
-0.0410545841,
-0.0052562249,
0.0410462432,
0.0319075435,
0.002997929,
-0.095978789,
-0.0239434894,
-0.0119817741,
0.0517490432,
-0.0322854035,
-0.0083698146,
-0.0470854454,
0.0469708852,
0.0164062176,
-0.0030294997,
-0.0374681354,
-0.0189744495,
-0.0006897229,
0.0134992804,
-0.02120235,
-0.0111385724,
-0.0014646639,
0.0417465493,
-0.0248314682,
0.0456616133,
-0.0298501477,
-0.0215237699,
0.0395780653,
0.0080535403,
-0.0116596529,
0.0442767367,
0.007919699,
0.0327291228,
-0.0723802894,
0.0186499339,
0.0186816938,
0.0173221994,
-0.0946445987,
0.0637618676,
-0.0020344614,
-0.0518539697,
-0.0661426708,
0.0258040261,
0.04419313,
0.0038286147,
0.0578970425,
-0.0238439571,
-0.0246810801,
-0.0185126048,
0.0184034761,
0.0212192517,
0.0268638115,
0.0058007119,
0.0087444941,
-0.0223634858,
0.0283999424,
0.0365417339,
0.0052982075,
0.0404355414,
0.0142417867,
0.0235515442,
-0.0128204301,
-0.0086535243,
-0.0145534789,
0.0133679705,
-0.0323025286,
0.0011806034,
-0.0119559513,
-0.0455787107,
-0.0079621682,
0.0055161836,
-0.0262161177,
-0.0007715101,
0.0225547422,
0.045188345,
-0.0006871044,
0.0301025361,
-0.005806861,
-0.056647215,
0.0348727591,
-0.005471033,
0.0120854517,
-6.637140856e-33,
-0.0430449024,
0.0174371824,
0.0011167774,
0.0044750366,
-0.0194853898,
-0.0004781905,
-0.039133843,
0.0079438528,
-0.0207585562,
-0.0337390192,
-0.0194715206,
-0.0270992424,
0.0331348926,
0.0213802885,
-0.016671896,
0.0357498974,
0.0665644929,
-0.006123744,
-0.0150321992,
0.0336968005,
0.0392142199,
0.0264076814,
0.0446710102,
-0.0276490971,
-0.0734720528,
0.0035079299,
-0.021642454,
0.0036529517,
-0.0089962399,
0.0716949701,
0.0080203908,
0.0195083208,
-0.0155243687,
0.0055377372,
-0.0374266952,
-0.0069834744,
-0.0920867845,
-0.0250499435,
0.0257201772,
0.0110022957,
-0.0082620652,
-0.0193921942,
0.0385628454,
-0.0167381335,
0.0029853976,
-0.0038496314,
0.0145058297,
-0.0118559822,
-0.0069039091,
-0.0071360385,
-0.0656512678,
0.0254422985,
0.0236646421,
0.0115463324,
0.0504105948,
0.0051812483,
-0.0155831659,
-0.00027133,
-0.0577063225,
-0.0009367177,
0.0282052588,
0.0766713619,
0.0473567657,
0.0424851067,
0.0021073401,
0.0179565419,
0.014810225,
-0.0122950226,
0.0497041866,
0.02054126,
0.0983635411,
0.0221446324,
0.0052749594,
0.0177813061,
-0.0070038387,
-0.0418656245,
-0.0397280008,
-0.0161336828,
0.0412059464,
0.0166820586,
0.0386360586,
0.0278722346,
0.0044112736,
-0.0091431038,
-0.0674617514,
-0.0012789025,
-0.0500861779,
-0.0290760584,
0.0155863566,
0.0461048968,
-0.059460748,
0.0171912648,
0.0655093193,
-0.0352808647,
0.0562392026,
0.0120214252,
-0.0400646068,
0.0252220128,
-0.0046053799,
0.0437512398,
-0.0281702392,
-0.0979501605,
0.0444305837,
-0.0311506763,
0.018319374,
-0.0396683253,
0.0204085708,
0.0834620371,
-0.0697614029,
0.0049054376,
-0.0129688047,
0.0084867775,
0.0199076086,
0.097907111,
0.0191163272,
0.0014520974,
-0.0061765187,
-0.0034987479,
-0.0167713482,
-0.0230854023,
-0.013398028,
-0.0170223005,
0.0390166081,
0.0402770974,
-0.0353953578,
0.0353830978,
-0.0446701646,
-0.0089956177,
0.0594220534,
-0.0364665538,
-0.0095378486,
0.0490882732,
2.862e-7,
0.0346849486,
0.0489477664,
-0.0029244653,
-0.0144833783,
0.0157253053,
0.0111859292,
-0.0118551897,
0.006046067,
-0.0010055267,
0.0250373762,
0.0033231322,
0.0460284725,
0.0360944718,
0.0080076735,
-0.0533364601,
0.0151224853,
-0.0009767342,
0.0457767062,
-0.0417403802,
-0.0196764041,
-0.0040461011,
0.1159011349,
0.0128031084,
-0.024761932,
0.0397615172,
-0.0157984067,
0.0394328646,
-0.0323895812,
0.000816692,
0.0362899341,
-0.0273120031,
0.0528591871,
-0.0197069999,
0.0240984634,
-0.0256743934,
0.0389096588,
-0.0302155782,
0.0976918042,
-0.0227901414,
-0.0377379693,
0.038296584,
0.0021485006,
-0.0079650981,
-0.048813656,
0.0737972036,
-0.0264236573,
-0.0003696241,
-0.0547761321,
-0.0039238264,
-0.016289223,
0.0171505995,
-0.0535740331,
-0.036291182,
0.0542035922,
0.0471123829,
-0.0230198279,
-0.0265311766,
-0.0950309113,
-0.006490265,
0.0448385477,
-0.0212061536,
-0.0260064527,
-0.042382706,
-0.0582097284,
0.0579097308,
-0.0553656705,
0.0001454831,
3.182322184e-34,
0.0456670448,
-0.0321907029,
0.0346622914,
-0.0205781665,
0.0212415606,
0.0026000857,
-0.0286468379,
0.021000959,
-0.0155483559,
-0.0987187177,
-0.0458306335
] |
http://arxiv.org/abs/2401.16107v1 | http://arxiv.org/pdf/2401.16107v1 | Beyond Direct Diagnosis: LLM-based Multi-Specialist Agent Consultation for Automatic Diagnosis | 2024-01-29T12:25:30 | 2024-01-29T12:25:30 | Haochun Wang, Sendong Zhao, Zewen Qiang, Nuwa Xi, Bing Qin, Ting Liu | cs.CL | cs.CL, cs.AI | Automatic diagnosis is a significant application of AI in healthcare, where
diagnoses are generated based on the symptom description of patients. Previous
works have approached this task directly by modeling the relationship between
the normalized symptoms and all possible diseases. However, in the clinical
diagnostic process, patients are initially consulted by a general practitioner
and, if necessary, referred to specialists in specific domains for a more
comprehensive evaluation. The final diagnosis often emerges from a
collaborative consultation among medical specialist groups. Recently, large
language models have shown impressive capabilities in natural language
understanding. In this study, we adopt tuning-free LLM-based agents as medical
practitioners and propose the Agent-derived Multi-Specialist Consultation
(AMSC) framework to model the diagnosis process in the real world by adaptively
fusing probability distributions of agents over potential diseases.
Experimental results demonstrate the superiority of our approach compared with
baselines. Notably, our approach requires significantly less parameter updating
and training time, enhancing efficiency and practical utility. Furthermore, we
delve into a novel perspective on the role of implicit symptoms within the
context of automatic diagnosis. | Beyond Direct Diagnosis: LLM-based Multi-Specialist Agent Consultation for Automatic Diagnosis: Automatic diagnosis is a significant application of AI in healthcare, where
diagnoses are generated based on the symptom description of patients. Previous
works have approached this task directly by modeling the relationship between
the normalized symptoms and all possible diseases. However, in the clinical
diagnostic process, patients are initially consulted by a general practitioner
and, if necessary, referred to specialists in specific domains for a more
comprehensive evaluation. The final diagnosis often emerges from a
collaborative consultation among medical specialist groups. Recently, large
language models have shown impressive capabilities in natural language
understanding. In this study, we adopt tuning-free LLM-based agents as medical
practitioners and propose the Agent-derived Multi-Specialist Consultation
(AMSC) framework to model the diagnosis process in the real world by adaptively
fusing probability distributions of agents over potential diseases.
Experimental results demonstrate the superiority of our approach compared with
baselines. Notably, our approach requires significantly less parameter updating
and training time, enhancing efficiency and practical utility. Furthermore, we
delve into a novel perspective on the role of implicit symptoms within the
context of automatic diagnosis. | [
0.0587791651,
0.0016875833,
-0.0126910172,
0.0124716619,
-0.0082764039,
-0.0017069959,
0.0120844729,
0.0205462109,
0.0461386107,
-0.0327262245,
-0.0101024555,
-0.0246681646,
-0.0099017331,
0.0070558116,
-0.0021939594,
-0.0429869704,
0.0340017751,
-0.0087710675,
0.0632330626,
-0.0024111609,
-0.0080675669,
-0.0051612207,
-0.0467525609,
0.0287378188,
-0.0060098921,
-0.0039590234,
0.0031142146,
-0.043171715,
0.0628799498,
-0.0387059376,
-0.0194142833,
0.0012313013,
-0.0041461391,
-0.0264676027,
0.0000019063,
-0.0368160866,
0.0034138998,
0.0115670422,
-0.000625268,
-0.0365901925,
0.0221754443,
-0.0124455392,
0.0244518723,
0.0360138193,
-0.0490359478,
0.0038445757,
-0.0053292885,
0.0502403006,
0.0601548217,
0.0766080394,
-0.0118374247,
-0.0120987417,
0.0024782391,
-0.0107356999,
-0.0024688658,
-0.0876506865,
0.0599684417,
-0.0102893477,
0.0481301658,
-0.0311427526,
0.001941281,
0.0222392809,
-0.0285734255,
0.0005567035,
-0.0039699608,
0.0533680134,
0.0284859687,
-0.0331965387,
-0.0293347668,
0.0598373935,
0.0164682213,
0.0366054066,
0.0381862037,
-0.0210605562,
0.0144095616,
0.0287819598,
0.0055221836,
0.0066733081,
-0.0497563891,
0.0124088312,
0.0797681734,
0.0081414366,
0.03654553,
0.0027690188,
0.000983361,
0.0880604535,
-0.0143780261,
-0.0152910147,
0.0153159024,
-0.0148148909,
-0.0088310912,
-0.0671393499,
0.0356692038,
0.0883295164,
-0.1203253493,
0.0159163605,
0.0102701997,
-0.0067444318,
0.0423946232,
-0.0497891679,
0.0218436271,
0.0192696974,
0.0257646814,
0.0263435673,
-0.0054053189,
-0.0221332386,
-0.0305710901,
0.0352294594,
-0.0433586426,
-0.0028514976,
-0.0647880286,
0.005876164,
-0.0476975963,
0.0303680934,
0.0175086968,
0.0002987168,
-0.0526991263,
0.0282406956,
-0.0275835358,
-0.0063715195,
0.0437548272,
0.0462697372,
-0.0223925691,
0.0260203797,
-0.0570190027,
0.0329210572,
-0.0359223597,
0.0251849871,
-0.0099655977,
-0.0125690317,
0.0225929469,
0.0097960429,
-0.0104709929,
0.0153707452,
0.0048629241,
0.0544795245,
-0.0098237237,
-0.0086814975,
-0.0097349016,
-0.0428971536,
-0.0259146504,
-0.0399102159,
-0.0543111674,
-0.0454498343,
0.015908597,
0.0139686828,
0.0087759458,
-0.0548089966,
0.0200821757,
-0.0371449217,
-0.0462444797,
0.0789716318,
-0.0544636287,
0.026986884,
0.0266173221,
0.0336342528,
-0.0099914363,
0.027407594,
0.0185816325,
-0.0274581239,
-0.0176294558,
-0.0140460553,
0.0477842204,
0.003403774,
-0.0221313667,
0.0152510963,
0.0206537098,
-0.0129086552,
-0.0410702191,
0.1049411818,
-0.0338729806,
0.0404596739,
0.0306106173,
-0.0168272182,
0.0055960091,
0.0397703499,
0.0513650663,
0.0087191267,
0.0505089946,
-0.0119873891,
-0.0264502168,
-0.003075443,
0.0001010757,
0.00627359,
-0.0386306047,
-0.0125623979,
-0.0173598044,
-0.0138655398,
-0.0318394154,
-0.0788832977,
-0.0567953624,
-0.0058052861,
-0.004874636,
-0.0025110538,
0.0352244079,
-0.0484040156,
-0.0136196576,
0.0572342426,
0.0503193252,
0.008563065,
-0.0330727436,
-0.0927942395,
0.0145082073,
0.0916463062,
0.0125766397,
-0.0744162574,
-0.043295119,
0.0035339345,
-0.00144814,
0.0104586408,
0.0463406034,
-0.013013294,
0.0195700042,
-0.0296044275,
-0.018905988,
0.0422790796,
0.0172662754,
0.0167116821,
-0.0524214096,
0.0063382038,
-0.0229177959,
-0.0254014768,
-0.0286590327,
0.0469210558,
-0.0714427158,
-0.008523535,
0.0085242903,
-0.0045242142,
0.0690485612,
0.0111962454,
-0.0260731727,
0.0272517111,
0.0118121961,
-0.0424657948,
-0.0825805962,
0.0234447047,
-0.0014901764,
-0.0047881557,
0.0211620815,
0.0088133691,
-0.0049662092,
-0.0490553938,
-0.0433252342,
-0.0397166349,
0.0062831179,
-0.0292331614,
0.0288095959,
-0.0007908139,
-0.0105896369,
0.0520634167,
-0.0076050819,
0.0017509435,
0.0418934152,
-0.0258661807,
-0.0379131734,
0.0067851809,
-0.003621249,
-0.0044000996,
-0.0543611646,
-0.0094464347,
-0.0208188873,
0.0362972096,
0.0317301452,
-0.0131605966,
-0.0184700117,
-0.001414885,
-0.0381969512,
-0.1248070821,
-0.014979776,
0.0254015177,
-0.0290461741,
-0.0114843408,
0.0185497832,
-0.0130204866,
0.0032291012,
-0.056928739,
0.0495635495,
0.0370837487,
0.0193370879,
0.0557381846,
0.0578790419,
-0.0446280427,
-0.0508757569,
0.019781217,
-0.0754107237,
-0.0341993384,
0.0546748042,
-0.016939614,
-0.0480368771,
-0.0119740665,
-0.0636947826,
0.0315334313,
-0.0126766842,
-0.0418480486,
-0.0077473107,
-0.0637422353,
-0.0649638623,
0.0626308173,
-0.0047153407,
0.0223398358,
0.0532531664,
-0.0078181261,
-0.0538319685,
0.0484613143,
-0.0423993729,
-0.0246468857,
-0.0375844575,
-0.0401982851,
-0.0058039161,
0.0923549458,
0.0257454999,
-0.0048527936,
-0.0301634111,
0.0112618487,
0.0200857725,
-0.0080520641,
-0.0081664314,
-0.0014394366,
-0.0011084855,
0.0487609915,
-0.0010180078,
-0.0250766538,
0.1251809746,
0.0046901051,
0.0343575664,
-0.0596607998,
-0.0655209944,
-0.0477554873,
-0.0202729777,
-0.0225995164,
0.0265034381,
0.0141213965,
0.0127858818,
0.0044788877,
-0.021707179,
0.0218719598,
-0.0416897088,
0.0352187864,
-0.0421767198,
0.0572528839,
0.0363256373,
-0.0215155222,
-0.0280376244,
-0.0477864519,
-0.0399026908,
0.0187187809,
-0.025837481,
0.0105687473,
-0.0760838538,
0.0443320014,
-0.0227206219,
0.0348772891,
0.0378450081,
-0.0208945759,
-0.0072312932,
-0.0327103771,
-0.0036474236,
0.0282741264,
-0.0044263387,
-0.0324593484,
0.0298513472,
0.0300130565,
-0.0138992714,
0.0019959116,
-0.0270066466,
0.0085124411,
-0.0351112597,
-0.0389103778,
-0.0009759435,
0.0422981493,
-0.0186327677,
-0.0255956631,
0.0090602217,
0.1149958521,
0.0248190165,
-0.0643976927,
0.0207712147,
-0.0235215407,
-0.0461162254,
-0.0109797725,
0.0030969928,
-0.0443911962,
-0.0335932225,
-0.0208826885,
0.0132566448,
-0.0193981696,
-0.0610351935,
0.0361449756,
-0.0166593399,
0.0225496832,
-0.0122316685,
0.018141944,
0.0088866726,
-0.030783264,
0.0004974319,
-0.0323144346,
0.0009250156,
-0.0296047572,
0.0246555414,
-0.0158377849,
0.0126581211,
-0.0497576781,
-0.0499264151,
0.0384873487,
-0.0274013244,
0.0614825897,
0.0478726402,
0.01733559,
-0.0151342591,
-0.0383325554,
-0.0234168191,
0.0422287099,
0.0594261736,
-0.0222496223,
0.0323389918,
-0.0125273829,
0.0197724234,
-0.0247894209,
0.0147418734,
-0.0737692788,
0.0159440842,
0.0313958302,
0.0758466497,
-0.0523338951,
-0.0456884727,
-0.0394897498,
0.0440728962,
0.0641614944,
0.0648266897,
-0.0337271579,
-0.0154720815,
-0.0187637284,
0.0155872963,
0.0052055265,
0.0689588636,
-0.0334733352,
0.0262689553,
-0.0353358053,
-0.0353324264,
-0.0321752802,
-0.0290818438,
0.0306185093,
-0.0024372695,
0.02543105,
0.0187201165,
0.010425495,
0.0172134656,
-0.0285911914,
0.0227476694,
0.053681843,
0.0362560339,
-0.0490738675,
-0.0312933549,
0.0808364898,
0.0197294392,
-0.0037947891,
0.0148747461,
-0.0358945914,
-0.042271398,
-0.0132570229,
0.0519130118,
0.0258757882,
0.0074815368,
-0.0273560211,
-0.0259908084,
0.0202123187,
0.0071493564,
-0.023331644,
0.0279923324,
-0.0021513903,
0.0378239527,
0.0075626997,
0.0449339971,
-0.0348552577,
0.0349276662,
-0.0097015202,
0.0374706462,
-0.0699656978,
-0.052549541,
0.0144598046,
0.0114531647,
0.0127053726,
-0.0121718496,
-0.0255435295,
-0.0202646106,
0.0068126414,
-0.0084769884,
-0.0152293919,
-0.0128985252,
-0.0007419553,
-0.0079027871,
-0.0378078185,
0.033383809,
-0.0147728724,
-0.0637743175,
-0.0490623005,
-0.001728244,
0.0519615747,
0.0456625335,
-0.0637939796,
0.0490324832,
-0.0057229805,
0.0083707077,
0.0182827562,
0.0239061639,
0.0204755142,
0.0408099517,
0.0552113988,
0.0081814546,
0.0570180714,
-0.0353487916,
0.004008594,
-0.0139673175,
-0.1008368731,
0.0232163295,
0.0489177629,
0.0238202307,
0.0385273136,
0.0186581053,
-0.0360184982,
0.0340068787,
-0.0043637599,
0.0145755215,
-0.0138500202,
0.0111469701,
0.0781378746,
0.0047018016,
0.0039034579,
-0.0023255514,
0.0688153952,
-0.0122831585,
0.024783507,
-0.0072534503,
-0.0491403714,
-0.0570101328,
-0.0170562062,
-0.0083884401,
0.0386050418,
-0.0411906876,
0.0294806883,
-0.0678743422,
-0.0392679982,
0.0043480503,
-0.0268038716,
-0.0303312726,
-0.0377716497,
0.0337300189,
0.0373062305,
0.0250111576,
-0.0183018297,
-0.0277220663,
0.0223755781,
-0.0474976152,
-0.0147901615,
0.0038700835,
-5.672546722e-33,
-0.0287389252,
-0.0357965194,
0.0348328687,
-0.0145346597,
0.0408282764,
0.036945533,
-0.0329212174,
0.0090193748,
0.0046281475,
0.036180649,
-0.0043300111,
0.016273303,
0.0127583342,
-0.0079506636,
0.0297450647,
-0.0200578794,
0.0372009315,
0.0011514942,
-0.0106907589,
-0.0100603122,
-0.0194236822,
0.0194247644,
0.0083672181,
-0.0799990892,
0.0090560941,
0.0054802634,
-0.0058406801,
0.0222222619,
-0.0402063429,
0.0485757627,
-0.0024389308,
0.0627007633,
0.0519159622,
0.0341913775,
0.0085766576,
-0.0128276693,
-0.0377786532,
-0.0266686175,
0.0285334531,
-0.0333290771,
-0.0384187773,
-0.0185087211,
0.0591918193,
0.0101882275,
-0.024058599,
0.0001690777,
0.0542268306,
-0.0313328169,
0.0072962884,
-0.0091752,
-0.0073162112,
0.0314944945,
-0.0055845482,
0.0066208155,
-0.0050211344,
-0.0331710875,
-0.0400116183,
-0.0059657008,
-0.0290604904,
-0.0052822446,
0.0711729899,
0.0031612257,
0.0173956379,
-0.0196378753,
-0.0519085042,
0.0280511584,
0.0543959737,
-0.0267049056,
-0.0754975677,
-0.0080991304,
-0.0564082824,
0.0369479544,
-0.0019741384,
0.0164230019,
-0.0183127671,
0.0076382724,
-0.0637509525,
0.0371186323,
0.0274195354,
0.0570504442,
0.0403279886,
0.0072094817,
-0.0441770926,
-0.0282185227,
-0.0324628875,
-0.0148118176,
-0.0369322747,
-0.034719564,
0.040664658,
0.021083286,
0.0304705352,
-0.0245826952,
0.0027079226,
-0.0050712661,
0.0134670092,
-0.0379625745,
0.0571893193,
0.0156035004,
-0.024779886,
0.0181042533,
-0.055508025,
-0.0335386954,
-0.0300406851,
-0.0128879603,
0.0461287349,
0.0185785685,
-0.0342631638,
-0.0273284446,
-0.0158662088,
-0.0005597582,
0.0298489965,
-0.0095817903,
0.0031533141,
0.065897502,
0.0172581971,
0.010883729,
-0.0214140788,
0.0622946024,
-0.0163663533,
-0.0058621336,
-0.0125409365,
-0.011251268,
0.0064586243,
-0.0130456528,
-0.0731812567,
-0.0188729204,
-0.0084600598,
-0.0127917482,
0.0626614615,
-0.0377297886,
-0.0124399019,
-0.0217693709,
2.678e-7,
0.0687848777,
0.0499035865,
0.0086712167,
-0.0100738946,
0.0382947624,
0.0391530506,
0.0027854419,
0.0274452288,
-0.0022337376,
-0.0246475562,
0.0348747596,
0.0347772576,
-0.0042545805,
0.0447043851,
-0.0308710989,
-0.0362387039,
-0.0324347429,
0.0124917179,
-0.0250406936,
-0.0133397924,
0.0450033657,
0.0727762356,
-0.0038778784,
-0.0248078369,
0.030804595,
-0.0303586069,
0.0501382984,
-0.0356573686,
0.0214956459,
0.080032222,
0.0368016846,
0.0242163334,
0.0222167633,
0.0835811198,
0.0098802336,
0.0047493502,
-0.0313113369,
0.0943610892,
-0.0783038512,
-0.0035493588,
0.0121057741,
-0.0278289132,
0.0487302504,
-0.0023076064,
0.0303678326,
-0.0719191357,
-0.0605681203,
0.0370678641,
0.0016851665,
-0.0353736952,
0.031299863,
-0.0222923551,
-0.0249433815,
0.0322523005,
-0.0038503585,
0.018140927,
-0.0204228163,
-0.0811822265,
-0.0126749491,
0.021714367,
-0.0479834341,
-0.0285761878,
0.0637778342,
-0.0213029906,
0.0327952132,
-0.0935964584,
0.0214498192,
2.193551667e-34,
0.0409576967,
0.0486228168,
-0.0252200309,
0.0545614362,
0.0340327136,
0.0142890196,
-0.0392750092,
0.0134918997,
0.0595349222,
-0.0955069289,
-0.0377597511
] |
http://arxiv.org/abs/2402.01739v2 | http://arxiv.org/pdf/2402.01739v2 | OpenMoE: An Early Effort on Open Mixture-of-Experts Language Models | 2024-01-29T12:05:02 | 2024-03-27T10:21:24 | Fuzhao Xue, Zian Zheng, Yao Fu, Jinjie Ni, Zangwei Zheng, Wangchunshu Zhou, Yang You | cs.CL | cs.CL, cs.AI, cs.DC, cs.LG | To help the open-source community have a better understanding of
Mixture-of-Experts (MoE) based large language models (LLMs), we train and
release OpenMoE, a series of fully open-sourced and reproducible decoder-only
MoE LLMs, ranging from 650M to 34B parameters and trained on up to over 1T
tokens. Our investigation confirms that MoE-based LLMs can offer a more
favorable cost-effectiveness trade-off than dense LLMs, highlighting the
potential effectiveness for future LLM development.
One more important contribution of this study is an in-depth analysis of the
routing mechanisms within our OpenMoE models, leading to three significant
findings: Context-Independent Specialization, Early Routing Learning, and
Drop-towards-the-End. We discovered that routing decisions in MoE models are
predominantly based on token IDs, with minimal context relevance. The
token-to-expert assignments are determined early in the pre-training phase and
remain largely unchanged. This imperfect routing can result in performance
degradation, particularly in sequential tasks like multi-turn conversations,
where tokens appearing later in a sequence are more likely to be dropped.
Finally, we rethink our design based on the above-mentioned observations and
analysis. To facilitate future MoE LLM development, we propose potential
strategies for mitigating the issues we found and further improving
off-the-shelf MoE LLM designs. | OpenMoE: An Early Effort on Open Mixture-of-Experts Language Models: To help the open-source community have a better understanding of
Mixture-of-Experts (MoE) based large language models (LLMs), we train and
release OpenMoE, a series of fully open-sourced and reproducible decoder-only
MoE LLMs, ranging from 650M to 34B parameters and trained on up to over 1T
tokens. Our investigation confirms that MoE-based LLMs can offer a more
favorable cost-effectiveness trade-off than dense LLMs, highlighting the
potential effectiveness for future LLM development.
One more important contribution of this study is an in-depth analysis of the
routing mechanisms within our OpenMoE models, leading to three significant
findings: Context-Independent Specialization, Early Routing Learning, and
Drop-towards-the-End. We discovered that routing decisions in MoE models are
predominantly based on token IDs, with minimal context relevance. The
token-to-expert assignments are determined early in the pre-training phase and
remain largely unchanged. This imperfect routing can result in performance
degradation, particularly in sequential tasks like multi-turn conversations,
where tokens appearing later in a sequence are more likely to be dropped.
Finally, we rethink our design based on the above-mentioned observations and
analysis. To facilitate future MoE LLM development, we propose potential
strategies for mitigating the issues we found and further improving
off-the-shelf MoE LLM designs. | [
0.0479329079,
0.0801927447,
0.0159294512,
0.0166552141,
-0.0295353048,
-0.0271498077,
0.0341922492,
-0.0117264995,
-0.044884529,
-0.0312136356,
-0.0507591479,
-0.0107616382,
0.0018612839,
0.0176929887,
0.0094392393,
-0.0340249762,
0.0213821717,
-0.072464861,
0.0025287669,
-0.0213993844,
-0.0237760749,
0.0256885309,
-0.0340169892,
0.0799491704,
0.0081217345,
0.011247742,
-0.0063607651,
0.0155012524,
0.0018905146,
0.0044666156,
-0.0078140916,
0.0837827772,
0.0273065586,
0.0519481897,
0.0000020244,
-0.0374095254,
-0.0103430077,
0.0013810799,
-0.05957219,
0.0133050606,
0.0254309811,
0.010176315,
-0.0213772245,
0.0398637988,
0.0177913122,
-0.0255005173,
0.0520229675,
0.0212871321,
0.0261421893,
0.0935258195,
0.0035994353,
-0.0067427643,
0.0637831315,
-0.0248368867,
0.0134994518,
-0.0528758839,
0.0141219208,
-0.0206689015,
-0.0475887023,
0.0083691645,
-0.0216084681,
0.0333289243,
0.0211751852,
0.0236423295,
0.0248265229,
0.0557619259,
0.0034856314,
-0.0329309292,
-0.0691522583,
0.0803425983,
-0.0285944045,
-0.0012264259,
-0.0063878018,
-0.0032967727,
-0.0423274338,
-0.0085910046,
-0.0414550044,
0.0366555415,
0.0111333793,
0.0186633691,
0.0013965797,
0.0304616969,
0.0245030094,
0.0246707797,
-0.0824067369,
0.0368402079,
0.0167990681,
-0.0315189809,
-0.0089942105,
-0.0237082448,
-0.0108493464,
-0.0442228802,
0.0342032798,
0.0060815043,
-0.0311318841,
0.015187622,
-0.0287039112,
-0.0262189154,
-0.0026955896,
0.0136682084,
-0.0045148446,
0.0733709559,
-0.037774168,
0.0390944891,
-0.0043688798,
-0.0045151426,
-0.0220813137,
0.019372372,
-0.0313119665,
-0.0329589769,
-0.0433425456,
-0.0027276203,
-0.0180173349,
0.0137150036,
-0.0532913916,
-0.0501310639,
-0.0000753958,
0.0201886445,
0.0078457613,
-0.0122531299,
0.0592085049,
0.0044749225,
-0.0355870984,
-0.0253257379,
-0.0058321874,
0.0566022508,
-0.0082405368,
0.0214119926,
0.0337042138,
-0.0840546861,
-0.0001091149,
0.0402416959,
-0.0125497645,
-0.0378289223,
-0.0072874427,
0.0575330406,
-0.0503746755,
-0.0270651393,
-0.0028874082,
-0.0189865865,
-0.0198174585,
-0.0504423864,
-0.0132240821,
-0.0587218329,
0.0368065462,
0.0229218937,
0.0391698107,
-0.027250506,
-0.0179113429,
-0.0052479613,
-0.0522507429,
0.0122242877,
-0.0129277138,
0.0355482474,
0.0241102781,
0.0083890492,
-0.0175912529,
0.0598663054,
0.004512751,
-0.0059878351,
0.0203337651,
-0.0184257682,
0.0135025568,
0.0429204777,
0.0015505274,
0.0106030079,
0.0328510664,
0.0109993629,
0.0096690319,
0.048947528,
0.066054523,
0.0683477074,
-0.0675418898,
-0.0243424736,
0.0527226776,
0.0528530329,
0.0307143796,
0.0223816056,
0.0240858961,
0.0526286885,
0.0069805062,
-0.0126264496,
-0.0200224426,
0.0047036158,
-0.0239421241,
-0.0115062613,
0.096252799,
0.0463203974,
-0.0441844203,
0.0044398801,
0.0149044255,
-0.0185708944,
-0.054118365,
0.0177729279,
0.0577662252,
-0.0274864789,
-0.0412866659,
0.0627201125,
0.0175609943,
-0.0417371057,
-0.0105091091,
-0.0517853871,
0.005386868,
0.0561874844,
0.0509605855,
0.0089262817,
-0.0094383443,
-0.0314299986,
-0.0067901863,
0.020299416,
0.0275758766,
0.0092723724,
0.0418184064,
-0.08353623,
-0.0280089043,
0.0139896646,
0.0033314943,
0.0374753699,
0.0107438769,
0.0052197548,
0.0009280993,
-0.032317441,
0.021837594,
0.0215458218,
-0.0081149759,
0.010541263,
0.0331507362,
0.0016607596,
0.0346764363,
0.0328949764,
-0.0265439656,
0.0057218219,
0.0250680801,
-0.0136137307,
0.0021262998,
-0.0041918913,
-0.1208335534,
-0.007014499,
0.0287881158,
-0.0024178172,
-0.0223881099,
-0.0142349852,
-0.0178679693,
-0.0609641522,
0.0433732606,
-0.0105509926,
0.0567994267,
-0.0382165499,
0.0134981098,
0.0253697429,
-0.0139416344,
-0.0023877451,
0.02928482,
-0.043492116,
0.0028548164,
0.0393192098,
0.0172573123,
0.0131116733,
-0.0573053136,
-0.0087785376,
0.0001825834,
-0.0177023634,
0.0393315814,
-0.0181638449,
0.0119243516,
-0.0167254042,
-0.0705609322,
-0.1160326749,
0.0018158092,
-0.042518869,
0.0103963045,
0.0112603828,
0.0051474553,
-0.0256658327,
0.0153169222,
0.0021920484,
-0.0327424593,
-0.0268482268,
-0.0189016238,
-0.0451018661,
0.0090770284,
-0.019613605,
-0.0105727548,
-0.0297364239,
0.0208107755,
0.0549852028,
-0.0419513322,
-0.0021305729,
-0.0287647787,
-0.0201430954,
0.0039458391,
0.0122688999,
-0.0272464231,
-0.0234176647,
0.0184783563,
-0.0090594329,
0.0957524255,
0.008087934,
0.006682313,
0.0179302916,
0.0234407969,
-0.029625522,
-0.053943947,
0.0021085897,
0.0053066183,
-0.0462871343,
-0.0658610538,
-0.010590205,
-0.002045658,
0.0472169891,
-0.0054677804,
0.0142292175,
-0.0742869079,
-0.0429432094,
0.0078907274,
-0.069925867,
-0.0160917956,
-0.0072998125,
-0.0241472758,
0.0154078249,
-0.0188727621,
-0.0226706155,
0.0496849529,
-0.0415374078,
-0.0061365901,
-0.0411036238,
-0.0147717716,
-0.0422745496,
-0.0027183613,
-0.0436674953,
0.0199155305,
-0.0182465278,
-0.0538643338,
-0.0218638927,
0.0214145295,
-0.0399624668,
-0.0201155171,
0.0310674123,
-0.0507201888,
0.0057250718,
0.0001176737,
-0.030128736,
-0.0128550595,
0.0011454341,
-0.0442989543,
0.0385558344,
0.0582810827,
-0.0311550926,
-0.0559773743,
0.0150260832,
0.022959305,
0.0548012219,
0.0064199367,
0.0198078137,
-0.0209246259,
0.0238348637,
0.0114074377,
0.0282885879,
-0.0130322129,
-0.0581905879,
0.0188370664,
0.0245838687,
0.082781598,
-0.0433217846,
-0.0244525354,
0.0089725116,
0.0388340726,
-0.0027497883,
-0.011392124,
0.0428646319,
0.0496853329,
-0.0319863893,
0.0154550979,
0.1180632189,
0.0147920363,
-0.1074702516,
-0.0225637462,
-0.0257017408,
0.0507735536,
-0.0022443826,
0.0027552054,
-0.0960653275,
-0.046331007,
0.0211745296,
0.0398047864,
-0.001974371,
-0.0043587387,
0.0003977223,
-0.1041187569,
0.0251017455,
0.0086404998,
-0.1035190299,
-0.0149798263,
-0.0271424633,
-0.0363879502,
-0.0337164402,
0.0707193911,
0.0224288329,
0.0154474247,
-0.0330228768,
0.0575667694,
-0.0187038556,
-0.0793552026,
0.0294933543,
-0.0354426503,
0.1010320187,
0.0265656952,
-0.0420600139,
0.0120662637,
-0.0152479745,
0.0326545425,
-0.0052438886,
0.0182672758,
-0.0219256971,
0.0249264464,
-0.0668165237,
0.0194001924,
-0.0191532075,
0.0163953975,
-0.0112348599,
-0.0013445476,
0.0047483146,
0.0783412978,
-0.0037876237,
-0.011897767,
-0.0017849274,
-0.0241350215,
0.0084047513,
0.076794453,
0.0016070575,
-0.0440254658,
-0.0328782275,
-0.0051601622,
0.0178225003,
-0.0108592343,
0.002962234,
0.0212197453,
0.021343125,
-0.0794841871,
-0.011220431,
0.0264702495,
0.0677174777,
0.0534485206,
0.0046709492,
0.0338157676,
-0.0187710375,
0.0065240781,
-0.0229171012,
0.0558098704,
-0.0482119247,
-0.0007821493,
-0.1468103975,
-0.0196169354,
0.024128601,
-0.0328523554,
0.0029245438,
0.0113327857,
-0.0724603608,
-0.0675451383,
-0.0163168255,
0.0407839194,
0.0267494861,
0.0515260212,
-0.0029274244,
-0.1148128957,
0.0602565259,
0.0182107072,
-0.0711238086,
0.0200950857,
0.0597212538,
0.0094913468,
-0.0006696706,
0.0215478111,
-0.0405342914,
-0.0168644544,
0.0070082047,
0.0520655252,
-0.0168490987,
-0.018134091,
0.0021464638,
-0.0191968419,
0.0489180312,
-0.0109238252,
0.016166715,
-0.0296979267,
0.0504965223,
0.0063096411,
0.0252765082,
-0.0251063835,
0.0055497168,
0.0460512042,
0.0208973903,
0.0306907855,
0.017069893,
-0.001760877,
0.0362476371,
0.0059501119,
0.0232265852,
-0.0008460214,
0.0025157593,
0.0369092301,
-0.0197183471,
-0.0319431238,
-0.0008812739,
-0.0099735186,
-0.009541722,
-0.0132168913,
0.0521042384,
-0.0060931602,
-0.0102640018,
0.0071367533,
0.0278608818,
-0.0482354574,
-0.0481510051,
0.0109474827,
0.0561432578,
0.03982272,
0.0307968445,
0.0502842292,
-0.0213004835,
0.0226178523,
-0.0448793881,
0.025198333,
-0.0411825404,
-0.0038364679,
0.0264983829,
0.0353066064,
-0.006524038,
0.0343438834,
0.0652805865,
0.0234476533,
0.0008223,
-0.0274241753,
0.0205947161,
-0.0234217532,
-0.0060764356,
0.016925456,
0.0603022613,
-0.0558303408,
0.0024138226,
-0.0389680602,
-0.0201734528,
-0.0252181645,
-0.0131493723,
0.0100097749,
-0.0016855972,
0.0490899235,
0.0460842662,
-0.0464265607,
-0.035732016,
-0.0088332519,
-0.0336118005,
0.0184509624,
-0.0293267127,
0.0099149663,
-7.271610258e-33,
0.0178564768,
0.0150876511,
0.0474763289,
-0.0622431263,
-0.0152418809,
-0.003372856,
-0.0304954983,
0.0261423998,
-0.0305746272,
-0.0320663974,
-0.0175539833,
0.005434853,
0.0313382074,
0.0578353852,
-0.013141037,
0.031478364,
0.0741109923,
-0.0234744288,
0.0041493117,
0.0289612319,
0.0098126968,
0.0265380982,
0.0303804632,
-0.0439053103,
-0.0158431046,
0.01935976,
-0.0007881755,
0.0074900873,
-0.0103710117,
0.058294557,
-0.0354403071,
0.0461650304,
0.0071399552,
-0.0085370122,
-0.0291396156,
-0.0147836488,
-0.055051107,
0.018375827,
0.0162295736,
-0.0047283336,
-0.0094130198,
-0.0409587249,
0.0247576497,
0.0187793467,
-0.015523064,
-0.0096541233,
0.0110237421,
0.0065474729,
-0.0477959998,
0.0122704627,
0.0025584297,
0.0473548733,
-0.0238146782,
-0.0148690026,
0.0290299468,
-0.0112550016,
-0.0028491379,
0.000295867,
-0.0229756963,
0.0299245343,
0.0042741667,
0.0197323784,
0.0792442784,
0.0435160436,
0.0098107904,
0.040478155,
0.0404203273,
-0.0085500507,
-0.027616607,
-0.0207223315,
0.0528494529,
0.0634415373,
0.0158231836,
-0.0117602767,
0.0136093311,
-0.0131976055,
-0.0372354947,
-0.0239616949,
0.0250034221,
0.0262121409,
0.0643020049,
0.049801182,
-0.0257887784,
-0.007638731,
-0.0308709759,
-0.0240795687,
-0.0392570719,
-0.0148780877,
0.0440888219,
0.0393534414,
-0.0270985067,
-0.0495437086,
0.084762983,
-0.0231398623,
0.0056899269,
0.0086787483,
-0.0043900069,
0.0523239598,
-0.031398572,
0.0433673374,
-0.1357145011,
-0.0407162011,
0.067372568,
-0.026936898,
-0.014662697,
-0.0189629178,
-0.0252737645,
0.0176352933,
-0.0568408892,
-0.0200481359,
-0.0150441751,
0.0286155902,
-0.0097674318,
0.0360236578,
0.0442880578,
-0.0015045528,
-0.0007457752,
0.0315293297,
-0.0172963031,
0.0046503828,
-0.0126132313,
-0.0377509333,
0.0007730479,
-0.0115513969,
-0.0186887756,
0.0304488353,
-0.0191799644,
-0.0335702859,
0.0369682722,
-0.0086555341,
-0.0371164791,
0.0277986266,
2.799e-7,
0.0198428705,
0.0477734208,
0.030482363,
-0.0220250078,
0.0117083173,
0.0081670498,
-0.0235885829,
0.0163399503,
0.0133274244,
-0.0201486666,
0.0158806052,
0.0383206941,
0.0565517582,
-0.0026557522,
-0.1008958966,
-0.0064502591,
-0.0496313386,
0.0335251391,
-0.067840755,
-0.0154869938,
0.0056133326,
0.0908346176,
-0.0047735474,
0.0031082248,
-0.0042896331,
0.0043581994,
0.0375010595,
-0.0540253371,
-0.0036223403,
0.0247188788,
0.0541468635,
0.0251539294,
0.0002393012,
0.0271398537,
0.0246183425,
0.0368353836,
0.0186512843,
0.0705284774,
0.0070833717,
-0.0910214111,
0.0368624926,
-0.016788993,
-0.0182908271,
-0.0338662304,
0.0454130955,
-0.0228342991,
-0.0298551843,
-0.0055967886,
-0.0357828774,
-0.0067940294,
0.0044297972,
-0.0325952321,
-0.0424327739,
0.0245276205,
0.0150621589,
-0.0113033503,
0.0046912134,
-0.0719063878,
0.0193284862,
-0.0252397768,
-0.0624482594,
-0.0798386484,
-0.0183501057,
0.0225958917,
0.0367615856,
-0.0070292368,
0.0011154942,
2.462604416e-34,
0.0552995875,
0.0091532059,
0.0146952095,
0.0484691858,
0.0239651706,
-0.0324852653,
-0.0367287882,
0.0175347049,
-0.0056599365,
-0.0991137773,
-0.0317268819
] |
http://arxiv.org/abs/2401.16092v3 | http://arxiv.org/pdf/2401.16092v3 | Multilingual Text-to-Image Generation Magnifies Gender Stereotypes and Prompt Engineering May Not Help You | 2024-01-29T12:02:28 | 2024-05-15T15:29:00 | Felix Friedrich, Katharina Hämmerl, Patrick Schramowski, Manuel Brack, Jindrich Libovicky, Kristian Kersting, Alexander Fraser | cs.CL | cs.CL, cs.CY, cs.LG | Text-to-image generation models have recently achieved astonishing results in
image quality, flexibility, and text alignment, and are consequently employed
in a fast-growing number of applications. Through improvements in multilingual
abilities, a larger community now has access to this technology. However, our
results show that multilingual models suffer from significant gender biases
just as monolingual models do. Furthermore, the natural expectation that
multilingual models will provide similar results across languages does not hold
up. Instead, there are important differences between languages. We propose a
novel benchmark, MAGBIG, intended to foster research on gender bias in
multilingual models. We use MAGBIG to investigate the effect of multilingualism
on gender bias in T2I models. To this end, we construct multilingual prompts
requesting portraits of people with a certain occupation or trait. Our results
show that not only do models exhibit strong gender biases but they also behave
differently across languages. Furthermore, we investigate prompt engineering
strategies, such as indirect, neutral formulations, to mitigate these biases.
Unfortunately, these approaches have limited success and result in worse
text-to-image alignment. Consequently, we call for more research into diverse
representations across languages in image generators, as well as into
steerability to address biased model behavior. | Multilingual Text-to-Image Generation Magnifies Gender Stereotypes and Prompt Engineering May Not Help You: Text-to-image generation models have recently achieved astonishing results in
image quality, flexibility, and text alignment, and are consequently employed
in a fast-growing number of applications. Through improvements in multilingual
abilities, a larger community now has access to this technology. However, our
results show that multilingual models suffer from significant gender biases
just as monolingual models do. Furthermore, the natural expectation that
multilingual models will provide similar results across languages does not hold
up. Instead, there are important differences between languages. We propose a
novel benchmark, MAGBIG, intended to foster research on gender bias in
multilingual models. We use MAGBIG to investigate the effect of multilingualism
on gender bias in T2I models. To this end, we construct multilingual prompts
requesting portraits of people with a certain occupation or trait. Our results
show that not only do models exhibit strong gender biases but they also behave
differently across languages. Furthermore, we investigate prompt engineering
strategies, such as indirect, neutral formulations, to mitigate these biases.
Unfortunately, these approaches have limited success and result in worse
text-to-image alignment. Consequently, we call for more research into diverse
representations across languages in image generators, as well as into
steerability to address biased model behavior. | [
0.0178602599,
0.0714630932,
-0.0282678362,
0.0331581384,
-0.0127751278,
0.0246256962,
0.0990120023,
0.007697891,
0.0056040711,
-0.0180586372,
-0.0111751659,
-0.0941395685,
0.0421190411,
0.0308714248,
0.0471154749,
-0.0933328271,
-0.003457861,
-0.0059249373,
-0.0133680813,
-0.0425993353,
0.0167486239,
0.0425709784,
0.0105776424,
0.0564711988,
-0.040717192,
-0.0490785651,
0.0156591162,
0.0547875613,
0.0098416572,
0.0104724951,
0.0006407885,
0.0537752137,
-0.0114009231,
0.0159958862,
0.0000020188,
-0.0634856299,
0.0142988889,
0.0126227662,
0.0238763504,
0.0099172443,
0.0523388982,
-0.0209254827,
-0.0386287123,
-0.0233318787,
-0.0170674324,
0.025102172,
0.0939799026,
0.0350051634,
0.0433551408,
0.0086054429,
-0.0117653152,
0.0066379574,
-0.0703932941,
-0.0234622434,
0.0290892795,
0.01613429,
-0.0097562848,
-0.0436687134,
-0.0342513695,
0.0020865279,
-0.0168818068,
0.0403909497,
-0.0014428882,
-0.0035188773,
-0.0157786757,
0.0512951016,
0.020459028,
-0.0879327729,
-0.0054192231,
0.0560203232,
0.009788882,
0.0282024927,
-0.021219464,
-0.0130801685,
0.0053775916,
0.0358044989,
-0.0307888184,
0.0493415333,
-0.0216125827,
-0.0298847966,
-0.0397664681,
-0.0005290539,
0.0034063228,
-0.0022720799,
0.0002123048,
0.018323442,
0.0380128883,
-0.0093827778,
-0.0200084969,
0.0289309025,
-0.0033238099,
-0.0688854158,
0.0244698804,
-0.0096533634,
0.0017925918,
0.0042343377,
-0.0347707048,
-0.0450180583,
0.0279354416,
-0.0295003038,
0.0189496279,
0.0452802218,
-0.0275710896,
0.0003813246,
-0.0069201817,
0.0814909562,
-0.0094713317,
-0.0009716048,
-0.0453015864,
-0.0255592242,
0.0080641964,
-0.0392661355,
-0.0161108337,
-0.0542865992,
-0.0368531197,
0.0107778618,
-0.0624829084,
0.0071852617,
-0.0394915603,
0.0138973547,
-0.0051440713,
0.0437700972,
-0.0797765628,
0.0169108436,
-0.030089464,
0.0347204432,
-0.0562770553,
-0.0142750507,
-0.0153618539,
-0.0864104629,
0.0047772578,
0.0100106318,
-0.0000613511,
-0.0349077731,
-0.0134685766,
0.047649771,
-0.0199546441,
-0.0070512346,
0.0219202787,
-0.0237403549,
-0.0404882208,
-0.060026288,
0.0214978289,
-0.0744407475,
0.0356603637,
0.0355431326,
0.0313265733,
-0.0594335347,
-0.0053636404,
0.0209941026,
-0.0031742172,
0.0471715927,
0.0506988838,
0.0027596096,
0.0616379827,
0.0113557233,
-0.0482952073,
0.0867653564,
0.0393370092,
0.0329237059,
0.0503069833,
-0.008475448,
-0.0045449152,
0.0694324449,
0.007005387,
0.0206284467,
-0.0143652773,
0.0168079566,
0.0261538755,
0.0401673131,
0.0741386861,
0.0224681012,
0.0145618301,
0.0079936944,
0.0830225646,
0.0616691336,
0.006474203,
0.017750416,
0.0640656501,
-0.0114790648,
-0.0098563768,
0.0330620222,
-0.0379465818,
0.0458694547,
0.0135926334,
0.0195147023,
-0.0286671147,
0.1040938944,
-0.0423964672,
-0.0218546018,
0.0329387598,
0.0081814788,
0.0204209406,
-0.0028426449,
0.0140188839,
0.0459647477,
-0.0271719936,
0.0323498473,
-0.0033060201,
-0.0811133385,
0.006328979,
-0.0229098015,
0.015671242,
0.057532005,
0.0025265981,
0.0269042738,
0.0228870157,
-0.0129463328,
-0.0036820956,
0.0923939496,
0.0498713106,
0.0261879843,
0.0279323068,
-0.0487990603,
-0.0266221836,
0.0032815004,
0.0048693377,
0.0692017972,
0.027416449,
0.0602855384,
0.0116052181,
-0.0013621994,
-0.0454884768,
0.0117062489,
-0.0222035181,
-0.0236761495,
0.0010304269,
0.0460576527,
-0.0193365347,
0.0178539474,
-0.0445405431,
0.0077603278,
-0.0090072416,
-0.0219633728,
0.0156541131,
0.0589778386,
-0.0427401625,
-0.0096295699,
-0.0002372169,
-0.0801154897,
-0.0726123378,
-0.0040469714,
0.0075773089,
-0.0080932016,
0.0304474384,
-0.0237700678,
-0.0237708185,
-0.0052562333,
0.0063173496,
0.0046329,
0.0149395335,
-0.0042457422,
0.0342086293,
-0.0151458001,
-0.0129073886,
0.0375689641,
-0.0049519846,
-0.0235238615,
-0.0693521127,
-0.0382635854,
0.0456949882,
0.0224326551,
0.0013725897,
0.0302569233,
0.0139390407,
-0.0261447243,
-0.012478902,
-0.0688082203,
-0.0256551541,
0.0012230247,
-0.0261086896,
-0.0246539451,
0.0205010921,
-0.0097436924,
-0.0060186153,
-0.07821922,
-0.0236072373,
-0.0001321258,
0.0319045819,
0.0414127335,
-0.0123586422,
0.013517729,
-0.0295954645,
-0.012981602,
-0.0019586126,
0.0532149337,
0.0419644304,
-0.0126315681,
-0.0432588495,
0.0054605035,
-0.0383482985,
-0.0116621638,
-0.0061750486,
-0.0172871426,
0.0350601897,
0.0393220112,
0.0716814399,
0.0790219977,
0.0068332166,
-0.0120487167,
0.0045792181,
0.0040755803,
-0.0448460355,
-0.0025294798,
0.0257835407,
0.0046873521,
0.0565217473,
-0.1326675564,
-0.0016521142,
0.011426311,
-0.0024517693,
0.0274866335,
-0.0780653581,
0.0544328243,
-0.0448648483,
-0.0524255484,
-0.0404896177,
0.0202771071,
0.0165118091,
0.052834522,
-0.0516684316,
-0.0554755293,
0.0507061072,
-0.0291348342,
-0.03157828,
0.0283170305,
0.0283488128,
-0.017960785,
0.0448381566,
-0.001801222,
0.066251412,
0.0482385196,
0.0356146581,
-0.0858513191,
-0.0055832253,
0.0548727028,
-0.0309341811,
0.028603524,
-0.0181912854,
-0.0289882757,
-0.0314987488,
-0.0523183234,
-0.0240909122,
-0.0104320273,
0.0081219161,
-0.0188661795,
0.0054367427,
-0.0144721288,
0.006706432,
0.006629921,
0.014076259,
-0.007006438,
-0.0415339209,
-0.0062949783,
-0.0078825438,
0.019668594,
0.0473220088,
-0.014239735,
-0.0126513559,
-0.0644584075,
0.0076181726,
0.034122508,
-0.0381454155,
-0.0019971591,
-0.0042509073,
0.0383151211,
0.0319311656,
0.0296242144,
-0.0440350026,
0.0086559346,
0.0786291435,
-0.0317223817,
0.0187818911,
0.0639952049,
-0.0054895277,
-0.0301838592,
-0.0150478259,
0.0124106715,
0.0539710969,
0.013573641,
0.0090057598,
-0.0733476952,
0.0148694441,
0.0075445394,
-0.002241957,
-0.0422286466,
-0.0137036676,
-0.0290602557,
-0.1348468959,
0.0341039114,
0.0293998662,
-0.0413143598,
-0.0244047269,
0.0114715211,
-0.0080431085,
0.0169871282,
0.0103242053,
-0.0045153927,
0.0065618544,
0.0153252892,
-0.0045014718,
-0.0033623134,
-0.0321451351,
-0.0090648197,
-0.0644549429,
0.0296371505,
0.0585189573,
0.0436656401,
0.0188699439,
-0.0571304485,
-0.0062545664,
0.0337602049,
0.035301432,
-0.0148402257,
0.0361587405,
-0.0069310698,
-0.0273176748,
-0.0398209877,
0.0209524129,
0.0550399348,
0.0145838875,
0.0029792206,
-0.0023175972,
0.0251109619,
-0.0106537119,
-0.0531835221,
-0.0094063515,
0.0044599399,
0.0485443473,
0.0176546946,
-0.062100362,
-0.051035881,
-0.0452000573,
0.0299526658,
-0.0545473807,
-0.052428063,
0.0181554556,
-0.0131769311,
-0.0665782839,
-0.052845221,
-0.0508345924,
0.0033281827,
0.0169651452,
0.0060835821,
-0.0712972879,
-0.0499698855,
0.0174965356,
-0.041609548,
0.036447566,
-0.0126653826,
0.0407860763,
-0.0126218535,
-0.0188016649,
-0.0462398417,
-0.0529094227,
0.0017423562,
0.0261088088,
-0.0498930886,
-0.0441122316,
-0.0148148453,
0.0751696154,
0.0233210139,
0.0389347523,
0.0331420451,
-0.0397279114,
0.0116947712,
0.0433350392,
0.039196603,
0.0075915782,
0.0121015524,
0.0314194895,
-0.0099043623,
0.0284720603,
-0.0348489843,
0.0093840789,
0.0186539236,
-0.0075990087,
0.0069836928,
-0.044254899,
-0.0009234127,
0.0310372561,
-0.0202656314,
-0.0848550871,
0.0103389798,
-0.0324674994,
-0.0230575968,
-0.0469574034,
-0.0287071429,
-0.0317749418,
-0.0375006534,
0.0235138778,
0.0246703848,
0.0271687005,
-0.0113337366,
-0.0115089817,
0.0073303278,
-0.0540756732,
0.0002204347,
0.0157478917,
0.037371397,
0.0019203682,
0.0338686593,
0.001745529,
0.0389241464,
-0.0352025144,
0.0322262496,
0.011233951,
0.0468246602,
0.0056169084,
0.0025823817,
-0.0109290723,
0.025072068,
0.0030569518,
-0.0326480456,
-0.0120255658,
0.0224785227,
0.0575625636,
-0.0259768814,
0.0759332553,
-0.0080826422,
-0.0052189399,
0.0120717781,
0.011453541,
0.0487380996,
-0.0307598859,
0.0310753193,
-0.0234817341,
0.0208455101,
-0.0531941615,
0.0500674285,
0.0004296722,
0.0263933539,
-0.038257882,
-0.032627929,
-0.0244227778,
-0.0093199406,
0.0400030501,
-0.0638273433,
-0.0084844939,
-0.0077082594,
-0.0938707441,
-0.0555543415,
0.0086480668,
0.014969442,
-0.0068394844,
0.0069232159,
0.0277816392,
0.0487185791,
-0.0112307388,
-0.0183239821,
0.0044927592,
0.0022173272,
-0.0374424942,
-0.0126402108,
-0.0197700933,
-5.902332358e-33,
-0.0509148613,
-0.0228114445,
-0.0116012618,
-0.0354490541,
-0.007382798,
-0.0011348759,
-0.0256039742,
0.0134388227,
0.0113836201,
0.0008050525,
-0.0052680401,
-0.0292005613,
0.0154213198,
0.008406424,
-0.0009813276,
0.0459605604,
0.0690891445,
0.0138405329,
-0.0624943785,
-0.022626603,
0.01009811,
-0.0036432545,
0.0297732037,
-0.050083492,
0.0690734163,
-0.0460126773,
-0.0239403751,
-0.0381629355,
0.0076583182,
0.0392084643,
0.0033895262,
0.033057522,
0.0182150006,
-0.0248189308,
-0.0019341403,
0.0727978721,
-0.050066974,
-0.0314877741,
-0.0159850027,
0.0573402308,
-0.0213509556,
-0.0474878848,
0.011292439,
-0.0146300448,
-0.0555447526,
-0.0280130971,
0.0658511668,
-0.0094465949,
0.0441801138,
0.0250632428,
-0.0708338544,
-0.0129957907,
-0.045187965,
-0.0153659424,
0.0100920163,
0.0999184474,
0.0113471942,
0.0021417008,
-0.0633549094,
0.0481109731,
-0.0018424904,
-0.018348353,
0.0782834888,
0.0337875299,
0.0276596639,
0.0118037648,
0.0698189139,
0.0327432491,
-0.0160682146,
-0.040509969,
0.0372456908,
0.0048093968,
0.0385006443,
0.0986490473,
0.0350093208,
-0.0123116923,
-0.0256584808,
0.0378362909,
0.0114567131,
0.007813273,
0.0192773249,
-0.0148532894,
-0.0434321687,
-0.0179777443,
-0.0320445262,
-0.079257369,
-0.0303986948,
-0.0392562672,
0.0227168873,
-0.0167192649,
-0.022155514,
0.0063281944,
0.0313598625,
-0.0260274392,
-0.0227401946,
-0.0338649638,
-0.0518945791,
0.0166441817,
0.0064426209,
0.0126519399,
-0.012912062,
-0.014174765,
0.02427141,
0.0590178892,
0.0306474846,
0.0483918786,
0.0114053357,
-0.0098083699,
-0.0640637502,
-0.0061456868,
-0.0445479862,
0.0165325571,
0.0168942884,
0.0279387478,
0.0398092009,
-0.0132131176,
-0.049089361,
-0.0285497215,
0.0194248594,
-0.0003901022,
-0.0608928017,
-0.0341682844,
-0.0083547225,
0.019224273,
-0.0347451903,
0.0219693519,
-0.0301275775,
0.0126815876,
0.0561096072,
-0.0466948822,
-0.0446539037,
-0.0387749262,
2.636e-7,
-0.0063852593,
0.0134268235,
0.0044959201,
0.047111962,
0.0337764919,
0.0521149337,
-0.0194561202,
0.0406731665,
0.0395321697,
-0.001694749,
-0.0368664823,
0.0252118446,
0.029115336,
0.016340157,
-0.0170344114,
0.0223623402,
-0.050491292,
0.0617254302,
-0.0204472728,
0.0481626168,
0.0213234164,
0.0374948196,
-0.0117831919,
-0.0320109352,
0.0059813643,
-0.0169091281,
0.0863604918,
-0.0346799083,
0.0131603992,
-0.0254242141,
-0.0085195284,
-0.008721414,
0.0112846363,
0.0250201058,
0.0107970592,
-0.0183444116,
0.0423239581,
0.0179016702,
-0.0308867544,
0.0452596955,
0.0145921409,
0.0119034098,
-0.0455504134,
-0.0739280805,
0.0543507785,
0.0190776065,
-0.019062154,
-0.031340979,
-0.0190007649,
0.0233037155,
0.0976265594,
-0.0415636227,
-0.0521709733,
0.0069627627,
0.043432802,
0.0037073076,
-0.0190202072,
-0.0088431854,
-0.0503232181,
-0.0351812467,
-0.0071097328,
-0.0524671562,
-0.0776756927,
0.0204166919,
0.0192864127,
-0.0310961101,
-0.0090018613,
2.163301743e-34,
-0.0118581215,
0.0236007199,
0.0388375781,
-0.0271579791,
0.0110090915,
-0.0296382625,
0.0055260835,
-0.0115431175,
0.0422128141,
0.0009515171,
-0.0294971839
] |
http://arxiv.org/abs/2403.08783v1 | http://arxiv.org/pdf/2403.08783v1 | Image-Text Out-Of-Context Detection Using Synthetic Multimodal Misinformation | 2024-01-29T11:55:14 | 2024-01-29T11:55:14 | Fatma Shalabi, Huy H. Nguyen, Hichem Felouat, Ching-Chun Chang, Isao Echizen | cs.CV | cs.CV, cs.AI, cs.CL | Misinformation has become a major challenge in the era of increasing digital
information, requiring the development of effective detection methods. We have
investigated a novel approach to Out-Of-Context detection (OOCD) that uses
synthetic data generation. We created a dataset specifically designed for OOCD
and developed an efficient detector for accurate classification. Our
experimental findings validate the use of synthetic data generation and
demonstrate its efficacy in addressing the data limitations associated with
OOCD. The dataset and detector should serve as valuable resources for future
research and the development of robust misinformation detection systems. | Image-Text Out-Of-Context Detection Using Synthetic Multimodal Misinformation: Misinformation has become a major challenge in the era of increasing digital
information, requiring the development of effective detection methods. We have
investigated a novel approach to Out-Of-Context detection (OOCD) that uses
synthetic data generation. We created a dataset specifically designed for OOCD
and developed an efficient detector for accurate classification. Our
experimental findings validate the use of synthetic data generation and
demonstrate its efficacy in addressing the data limitations associated with
OOCD. The dataset and detector should serve as valuable resources for future
research and the development of robust misinformation detection systems. | [
0.016039256,
0.0186166503,
0.0032717807,
0.0912080333,
-0.0054715867,
0.0205941107,
-0.0051356247,
0.0512915961,
-0.0039689671,
-0.0674441382,
0.0083729075,
-0.0324226059,
0.0159415957,
0.0525278747,
0.0103320219,
-0.0302302148,
0.0152343623,
0.0111374287,
0.0004592067,
0.0030594571,
0.0378209874,
0.0198468417,
-0.01456967,
0.0114319166,
-0.0716499165,
-0.0499475561,
0.0075369161,
0.0421233028,
0.0474239178,
-0.0245485753,
0.0212099459,
0.0143691162,
0.0545810573,
0.0493184887,
0.000001999,
-0.0224732794,
0.0043386491,
0.0098451246,
-0.0368984118,
-0.0062663183,
0.0737348199,
-0.0512470417,
-0.0214211643,
-0.0030703696,
-0.0127843237,
0.057883665,
0.0697896704,
0.0565362275,
0.0540342592,
0.1002600417,
-0.0240720175,
0.0027360904,
0.0073963352,
0.0224291254,
0.0452175178,
0.0091202902,
0.0150441732,
-0.0085497126,
-0.0386795178,
0.0355661847,
-0.0006442896,
0.0193278044,
-0.0112009989,
-0.0146900183,
-0.0402441397,
0.0260443594,
0.010616703,
-0.0402708426,
-0.0251442045,
0.0527700931,
-0.0596971437,
-0.0280744918,
-0.01124267,
-0.0304832291,
-0.0003288121,
0.0438585095,
0.012060157,
0.0447864756,
-0.0486447439,
0.0278026834,
-0.0313949399,
0.0111516565,
0.0180264842,
0.0038733133,
0.0229756404,
-0.0072684041,
-0.0042635277,
-0.0182749238,
-0.0548117794,
-0.0082480703,
-0.0119750006,
-0.0326537006,
0.0047043329,
-0.011962411,
0.0295618121,
-0.0055756816,
-0.0968840495,
-0.0869391188,
0.0378305689,
-0.1089527681,
0.0027921405,
-0.0125899212,
-0.0504785702,
0.0296327788,
-0.017151678,
0.0155537529,
-0.0312770866,
0.0260892697,
-0.0102278292,
-0.0043439553,
-0.0137754418,
0.0014773965,
-0.008583474,
-0.017812185,
-0.060040839,
-0.0104928352,
-0.0411578715,
-0.0078869574,
-0.0234400649,
0.0127499299,
-0.0334137306,
0.0127503062,
-0.0700666532,
-0.010527066,
-0.0267215725,
0.0207485575,
0.0371510647,
-0.0063905367,
-0.0283644274,
-0.0646654293,
0.0115850139,
0.0833249167,
-0.0180633385,
0.0150492257,
-0.0101967705,
0.0418737084,
-0.0220240876,
0.0053808736,
-0.0577680096,
-0.0496132337,
-0.0705180913,
-0.0212140568,
0.0382446051,
-0.0526303537,
-0.0143025415,
-0.0333412178,
0.0024315561,
0.0029242788,
0.01383805,
0.0295345467,
-0.0293946788,
0.0607035011,
-0.0924217552,
-0.0397776738,
0.0436131507,
0.0143696582,
0.0107760392,
0.1003888026,
-0.0012226175,
0.0837370381,
0.0749810934,
0.0060178121,
0.0248571914,
0.0209936965,
0.0277696811,
0.0076880306,
0.0179750491,
-0.0106675783,
-0.0083449818,
-0.0134518687,
0.0308714136,
-0.0092987856,
-0.0223218761,
-0.0392472669,
0.0651705936,
-0.0349217504,
-0.0152973933,
-0.0503040738,
0.0217180066,
0.0504063927,
-0.0267150458,
0.0057772533,
-0.0127370702,
0.0228060801,
-0.0630455688,
-0.0085162856,
-0.0036259426,
0.0532525033,
-0.0010599847,
-0.0167712122,
-0.004376099,
-0.0141876815,
-0.0020716551,
-0.0159443635,
0.0337902643,
0.0420142114,
-0.0501795895,
0.0313713364,
0.0431269109,
0.0052969195,
-0.0225303788,
-0.0092656687,
0.0033297094,
0.0699664503,
0.0325709991,
-0.0212370995,
0.0370670855,
-0.0280080196,
0.0345769748,
0.0368777364,
0.0008127238,
0.006532765,
0.0127934869,
-0.0333821028,
-0.0228035673,
-0.0390758365,
-0.0066850754,
-0.0003711475,
0.0323051848,
0.0289631821,
0.0001047762,
-0.0109347785,
-0.0145095792,
0.03017007,
0.0697932914,
0.0059501389,
-0.0186440237,
0.0244818851,
0.0310337394,
0.0117152706,
-0.0289801136,
-0.0282178968,
0.0472399928,
-0.0123568941,
-0.0095277801,
-0.0232603792,
-0.0254409704,
-0.0557204559,
0.0174754634,
-0.0512002632,
-0.0592100322,
-0.0075199185,
-0.053399723,
0.0507967025,
0.0188185349,
-0.0506478809,
0.0115983002,
-0.0693247095,
-0.0203214698,
0.0154955797,
0.069830969,
0.0395622142,
0.070662424,
-0.0354853347,
-0.0410085805,
0.0200708807,
0.0320464447,
0.0266013816,
-0.0026931968,
0.0486271083,
-0.0133712078,
-0.0237100068,
0.0054188762,
-0.0270507894,
0.0241107363,
-0.0036056917,
0.0048303278,
-0.0173144899,
-0.0642014593,
-0.029735785,
-0.0395675302,
0.0548313893,
-0.0346037336,
0.0102355313,
-0.0200494565,
0.0143450145,
0.0255899746,
0.000413939,
-0.0034443173,
0.0141630778,
-0.0256863795,
0.0116101671,
0.0353643224,
0.0227954295,
0.0049458812,
0.0674266219,
0.0392033644,
-0.0818038583,
-0.0771039873,
-0.021683313,
-0.0195046701,
0.0386722237,
-0.0377940722,
-0.0243929792,
0.004743869,
0.0346992873,
0.0025339429,
-0.00280688,
0.0614575706,
0.059773095,
0.021087477,
0.0265391003,
0.019640971,
-0.0198400673,
0.0022040075,
0.0381047316,
-0.0015396556,
-0.0740386918,
0.0265212003,
-0.0058140163,
0.0267644413,
0.0318503045,
-0.0178210195,
0.0018718215,
0.0045130872,
-0.0705747902,
-0.0437507704,
-0.037366908,
-0.025279019,
0.032079801,
-0.0415355451,
-0.0485466681,
0.0403640494,
-0.0294236746,
0.0158120673,
-0.0279223528,
-0.0157660898,
0.043239098,
-0.0089208148,
-0.0096851606,
0.0586991273,
0.0508304872,
0.0231943727,
0.0170618575,
-0.0218237061,
0.0381540582,
-0.005574598,
0.0003820662,
-0.0520182587,
0.0209301449,
0.0291634146,
-0.0094160046,
-0.0490520783,
-0.0909898728,
0.0380557775,
-0.0534553379,
-0.0158333089,
-0.0350210518,
-0.0441097952,
-0.0313265845,
0.0014156211,
0.0345209129,
0.0298486743,
-0.0043628137,
0.0175210014,
0.0880101547,
0.0201240275,
-0.0365433879,
0.0041037095,
-0.0654219463,
-0.0123701878,
0.0837136656,
-0.0265997145,
-0.0229601432,
-0.0118948426,
-0.0253047608,
-0.0087155756,
0.0595807694,
0.0219709054,
0.0303057525,
0.0409659706,
-0.043421451,
-0.0026473654,
0.0322504863,
0.0000472815,
-0.0310438778,
-0.0027797977,
0.0140326228,
-0.0293703526,
0.0198526345,
-0.0409486443,
-0.0684310049,
-0.0441982858,
-0.0065808473,
-0.0099046016,
-0.0599257834,
0.0244685877,
-0.0270512365,
-0.0750094205,
0.0603540689,
0.011684834,
-0.0322365239,
-0.0782372728,
0.003903541,
-0.0297475532,
0.056969963,
-0.0038516745,
-0.0024243384,
0.0329858772,
-0.0461158492,
-0.0168894436,
-0.0045118844,
-0.0627173707,
-0.0251459554,
0.0204335917,
0.0123738488,
0.0451825969,
0.0247809086,
-0.0031370805,
-0.0591770597,
0.0103767179,
-0.0096482737,
0.0115392758,
-0.0101725804,
0.1044813842,
-0.0294184498,
-0.0271215681,
-0.0361888967,
0.0120750125,
0.0262673441,
-0.0217435453,
-0.029033741,
0.0242130291,
0.0054995106,
-0.0340180472,
-0.0469302759,
0.0327360556,
0.070011124,
0.0101639722,
-0.0228604469,
-0.0059774537,
-0.0272689871,
-0.0263986047,
0.0331764072,
-0.0126902806,
0.0325258859,
0.0525958948,
-0.0091010015,
-0.0519583076,
-0.0581437387,
0.0057980474,
-0.0276017953,
0.0150228431,
-0.0074402755,
-0.0041532656,
-0.0131287528,
-0.0015243747,
-0.0564840659,
-0.0100438697,
-0.0219541974,
0.0567000695,
0.0026611784,
-0.0037070664,
-0.0241932385,
-0.0485432781,
0.0047152326,
0.0872794315,
-0.0727133676,
-0.0311331172,
0.000706254,
0.0364083499,
0.0064930143,
0.0430338085,
0.0007883173,
-0.0058207964,
0.0895434394,
0.0537987463,
-0.0042014411,
-0.0005854712,
0.0227890741,
0.0192052256,
-0.0003217036,
-0.0093306573,
-0.012027191,
0.039100334,
0.0124599887,
-0.0517076924,
-0.0830126777,
0.02653124,
0.0295986291,
-0.0226925928,
0.0190649014,
-0.058209151,
-0.0058131353,
-0.012567048,
0.0105106877,
-0.0324168205,
-0.0192776006,
0.0029006277,
0.0511543974,
0.0102949748,
-0.0223497283,
0.0437984243,
0.0489394478,
-0.0286533125,
-0.0057137273,
0.0261890404,
-0.034188658,
0.0571253821,
0.0439978763,
0.0197822563,
0.0165307987,
0.0065118787,
0.0176333971,
-0.0074640717,
0.0123359961,
0.0209129248,
0.0692863241,
0.033371374,
-0.0448065549,
-0.0119115226,
-0.0118620973,
-0.0165089574,
0.084088631,
-0.0787898824,
0.0771661028,
0.0493235327,
0.0362584479,
0.0210656952,
-0.0392930582,
-0.023907952,
-0.0080165518,
0.0237080343,
-0.0134338234,
-0.0463989526,
-0.0020771353,
-0.0200752392,
0.0011975553,
0.0234311465,
0.0651516616,
-0.0084904609,
0.0190597251,
0.0044331509,
-0.044194486,
-0.0545093752,
0.0202124119,
-0.0038371037,
0.0520496331,
0.008327039,
0.0180923995,
-0.0370858237,
-0.0471017957,
0.0356653966,
0.0335528627,
0.0187047161,
0.0534681007,
0.0144779515,
0.0524492934,
0.0015026542,
-0.0176856834,
0.0011527907,
-0.012200729,
0.0032985201,
0.0113579379,
-0.001475114,
-6.449523876e-33,
0.0055291746,
-0.0892554522,
-0.0367616415,
-0.0210866313,
0.0189297684,
0.0245915186,
0.0182148535,
-0.0449163355,
0.029113695,
0.0354869142,
0.0459298976,
0.015614463,
0.0037769773,
0.0400266796,
-0.067262575,
0.0084502613,
0.0305883568,
-0.0366707705,
-0.006682198,
-0.0086285844,
0.0808718055,
-0.0068017966,
0.0363900326,
-0.0956251323,
0.0438204668,
-0.000321263,
-0.06565696,
0.0030392935,
0.0004019077,
0.0238077771,
-0.0360081568,
0.0727792382,
0.037513148,
0.0068906331,
-0.021671176,
0.0412338153,
-0.0328747667,
-0.0228172988,
0.0056981482,
0.0163748078,
-0.0217552967,
0.0047775782,
0.0195945837,
-0.0162115991,
-0.0178878158,
-0.014304758,
0.0435819179,
-0.0256169699,
0.0178284012,
-0.0005043575,
-0.0731435642,
0.0121759297,
0.01097738,
-0.0014058673,
-0.0148471035,
0.0571060367,
0.0551083609,
-0.0124490112,
-0.0771771148,
0.0071619907,
-0.0195741206,
-0.0438758694,
0.0334287174,
-0.0027834112,
-0.0028549901,
0.0840263292,
0.0306880306,
0.0287031513,
-0.0312209167,
-0.0106877415,
-0.0110463835,
0.0154440356,
0.0488521867,
0.1370534748,
0.0626459792,
-0.0377730168,
-0.0666867122,
0.0242673662,
-0.0309172459,
-0.0414391309,
0.0374312177,
-0.0255443659,
-0.0526646264,
-0.0041281511,
0.0090333633,
-0.0120421043,
-0.0188117959,
-0.005285799,
0.0311260335,
0.0530582443,
-0.0299062971,
0.0493040346,
0.011105625,
-0.0342824198,
0.0177351106,
-0.0327772461,
0.0078363009,
0.0261986814,
0.0137860551,
-0.0022462138,
-0.0146813886,
-0.0024606977,
0.0932471082,
0.0111830598,
-0.0137346638,
-0.0291356072,
0.0671363398,
-0.0089677293,
-0.0466535613,
0.0217271876,
-0.0228158813,
0.0556107424,
0.0286231972,
-0.0324287154,
0.0073459353,
-0.0220310148,
-0.0566872768,
-0.0172806308,
-0.0068909405,
0.032277599,
0.0049210233,
-0.0344441347,
-0.0492941551,
0.0499197356,
-0.035012342,
0.0227643251,
-0.0159635134,
-0.0472971387,
0.0261021554,
-0.0746584162,
-0.0069487267,
0.0237754751,
2.681e-7,
0.0162327625,
0.0276154932,
0.0430832095,
-0.021326609,
0.0397287831,
0.0453621596,
-0.0063088695,
-0.0155747589,
-0.0356381685,
-0.0940134525,
-0.0270046592,
0.029709179,
-0.0013540043,
0.0232831277,
0.004951369,
-0.0131071461,
-0.0040621744,
-0.0242943056,
-0.0399552844,
0.004700263,
0.0350595452,
0.0590457469,
0.0137263238,
-0.0752984434,
0.0312122907,
-0.010673902,
0.0435509831,
0.0256851725,
0.0085887816,
-0.0531758778,
0.0158912688,
-0.0087544881,
0.0011845735,
0.0455131754,
0.0039040553,
-0.0772954598,
-0.0243820511,
-0.0177458078,
-0.0046122731,
-0.0176986046,
-0.0456495844,
0.0132037234,
-0.0150737436,
-0.0264514852,
0.0186635982,
0.0046796943,
-0.029880885,
-0.0513942689,
-0.0301230699,
-0.000744619,
0.0103455242,
-0.0178482421,
-0.0384677239,
0.0110598607,
0.0340552367,
0.0250670463,
-0.0274110064,
-0.038300138,
0.0126543157,
0.0204507653,
-0.052545473,
-0.0062660924,
-0.0029199817,
-0.0031354313,
0.0720796064,
-0.015574296,
0.0092642624,
2.040612505e-34,
0.0370883085,
-0.0040249424,
-0.0134306177,
-0.0551474318,
0.00788817,
-0.0104404269,
-0.0347926095,
-0.0616557747,
0.0064580301,
0.0279845763,
0.0101602292
] |
http://arxiv.org/abs/2401.16086v1 | http://arxiv.org/pdf/2401.16086v1 | Non-Fluent Synthetic Target-Language Data Improve Neural Machine Translation | 2024-01-29T11:52:45 | 2024-01-29T11:52:45 | Víctor M. Sánchez-Cartagena, Miquel Esplà-Gomis, Juan Antonio Pérez-Ortiz, Felipe Sánchez-Martínez | cs.CL | cs.CL | When the amount of parallel sentences available to train a neural machine
translation is scarce, a common practice is to generate new synthetic training
samples from them. A number of approaches have been proposed to produce
synthetic parallel sentences that are similar to those in the parallel data
available. These approaches work under the assumption that non-fluent
target-side synthetic training samples can be harmful and may deteriorate
translation performance. Even so, in this paper we demonstrate that synthetic
training samples with non-fluent target sentences can improve translation
performance if they are used in a multilingual machine translation framework as
if they were sentences in another language. We conducted experiments on ten
low-resource and four high-resource translation tasks and found out that this
simple approach consistently improves translation performance as compared to
state-of-the-art methods for generating synthetic training samples similar to
those found in corpora. Furthermore, this improvement is independent of the
size of the original training corpus, the resulting systems are much more
robust against domain shift and produce less hallucinations. | Non-Fluent Synthetic Target-Language Data Improve Neural Machine Translation: When the amount of parallel sentences available to train a neural machine
translation is scarce, a common practice is to generate new synthetic training
samples from them. A number of approaches have been proposed to produce
synthetic parallel sentences that are similar to those in the parallel data
available. These approaches work under the assumption that non-fluent
target-side synthetic training samples can be harmful and may deteriorate
translation performance. Even so, in this paper we demonstrate that synthetic
training samples with non-fluent target sentences can improve translation
performance if they are used in a multilingual machine translation framework as
if they were sentences in another language. We conducted experiments on ten
low-resource and four high-resource translation tasks and found out that this
simple approach consistently improves translation performance as compared to
state-of-the-art methods for generating synthetic training samples similar to
those found in corpora. Furthermore, this improvement is independent of the
size of the original training corpus, the resulting systems are much more
robust against domain shift and produce less hallucinations. | [
0.0255881939,
0.0395752974,
-0.0078783995,
0.0261708219,
-0.0273958277,
0.0094907815,
0.073353,
-0.0070987507,
-0.0406590216,
-0.0270043183,
-0.0627769381,
-0.0496118777,
0.0048944703,
-0.0083812978,
0.0294987671,
-0.0455267243,
0.0081134578,
-0.0002443916,
-0.0471417755,
0.0038750945,
-0.0286072604,
0.0306472499,
-0.00784628,
0.0105523383,
-0.0341852531,
-0.0209059734,
0.0166628901,
0.04017419,
0.0259438083,
-0.007255658,
-0.0048860782,
0.029056862,
0.0054438314,
0.0432092473,
0.0000019917,
-0.0052119954,
-0.0102314716,
-0.0175689347,
0.0210165232,
-0.0110884001,
0.070089139,
-0.0516332164,
-0.0052891755,
0.0113457106,
-0.0520817377,
-0.0230042953,
0.0381306037,
0.04087786,
0.0295627285,
0.111938186,
-0.0400872007,
-0.043278195,
-0.0253621973,
-0.0481662042,
-0.0295438208,
0.0043150852,
-0.0384119339,
-0.0116829742,
-0.0521512963,
-0.007218563,
0.0200965237,
0.0523170605,
0.0485955477,
0.0061090207,
-0.014090511,
0.0160712115,
-0.0318185501,
-0.041214563,
0.0059491745,
0.0615738928,
-0.0822231174,
-0.027110653,
-0.0192923844,
-0.0011797801,
0.030925639,
-0.0292500965,
-0.0232384931,
0.0316362865,
-0.0114694238,
0.0308144055,
0.0303586405,
0.0447771922,
0.03758334,
-0.0463625267,
-0.041029647,
-0.0194501914,
0.0184421744,
-0.0372896828,
0.042998556,
0.0295799281,
-0.0292485803,
0.0157838184,
0.0356191993,
-0.0365341231,
-0.039435368,
0.0063471249,
-0.0709231049,
-0.0697827265,
-0.0048617735,
-0.0928123742,
0.0481383614,
0.0433559008,
-0.0299282558,
0.0102385385,
-0.0338920169,
0.0210743286,
-0.0304652974,
0.0474372469,
-0.0731482357,
-0.0152008962,
-0.0259381253,
0.0296313372,
-0.0632492974,
-0.0132817868,
-0.0735882968,
-0.0081907166,
-0.0156638864,
0.0084034828,
0.0109420652,
-0.0078637805,
0.0312854797,
-0.0022333779,
-0.0787730217,
-0.0091804592,
0.0167111754,
0.0139959436,
-0.0233776011,
-0.0111772427,
0.0215722118,
-0.0971176401,
-0.0211253818,
0.0336886942,
-0.0207824558,
-0.0329351388,
-0.0007132912,
0.0974759907,
-0.0036471908,
-0.0083535798,
-0.0452870168,
0.0079630082,
-0.0074474472,
-0.0664983988,
0.0405104645,
-0.0361410901,
0.0807112455,
-0.0010906187,
0.0125026358,
-0.0394457728,
-0.0007588412,
0.013945193,
-0.0014843761,
0.0386933014,
-0.0387991779,
-0.0182179715,
0.0103913117,
-0.006174074,
0.012835836,
0.0784458742,
0.0467286445,
0.0090334397,
0.0207343381,
-0.0075280671,
0.0060767615,
0.0366068594,
0.0063878698,
0.0018296741,
-0.0493799709,
0.0044173114,
0.0061677378,
0.0416117311,
0.039814271,
0.0438271128,
-0.0160239376,
0.0162262581,
0.0619532168,
-0.0100930156,
0.0070427763,
0.0433273949,
0.0976756886,
-0.024728559,
-0.0305879191,
0.0425209552,
-0.0297442172,
0.0300812405,
0.0519862063,
-0.0051178657,
0.0176387429,
0.0401272401,
-0.0365982279,
-0.0408288836,
0.0203635991,
-0.0085243899,
-0.0157511272,
-0.0262075644,
-0.0029336689,
0.0635526478,
-0.0583239533,
0.0442008637,
0.0191420894,
-0.0295396149,
0.0076202992,
0.0302552935,
-0.0720177069,
0.0651895925,
0.0110222921,
0.0224160552,
-0.0056310422,
-0.038480036,
0.0172378086,
0.0948274061,
0.0749861076,
0.0295931697,
0.053984832,
-0.0635555759,
0.007552376,
-0.0268551596,
0.0005990901,
0.0286851134,
0.0033620992,
0.0368872806,
0.0349697098,
-0.0196993183,
-0.0647457317,
0.0415707342,
-0.0243395623,
-0.0172103196,
0.0506462902,
0.0232234392,
0.0250777435,
0.0096550863,
-0.0091283079,
-0.0061573964,
0.0159930158,
-0.0620149337,
-0.0457915775,
0.0565877482,
-0.0597794764,
-0.0524423085,
0.0070216041,
-0.0687013492,
-0.0165404025,
0.0310744215,
-0.0270459373,
-0.0382562838,
0.0039262883,
-0.0278983396,
-0.0070452942,
-0.0375236794,
0.0345365033,
0.0203983914,
0.055964753,
-0.0242963973,
0.0373058654,
0.0229858067,
0.0030863364,
-0.0031620343,
-0.0118406853,
0.0393077545,
-0.0347301885,
0.0155884819,
0.021062227,
0.0303600505,
0.0331514552,
-0.0047850721,
0.0076076952,
-0.0338546894,
-0.0227920748,
-0.0316933095,
0.0228396431,
-0.0368285663,
0.0097031267,
0.0358698815,
0.0389332995,
-0.0006526388,
-0.0181787629,
-0.0473232344,
0.0431099124,
-0.0101602962,
0.0098980973,
0.006583441,
0.0056697056,
0.0223915391,
0.0191003922,
0.0123413159,
0.0232555158,
0.0624773242,
0.0034362844,
-0.0274242889,
-0.0838571489,
0.0281146709,
-0.0002920937,
0.0397767648,
0.0231755599,
-0.0458757132,
-0.0170288924,
-0.0039504226,
0.0568501502,
0.0266387668,
-0.0008282695,
-0.0135239735,
0.0437689908,
0.0290320199,
-0.0292762648,
0.0252584703,
-0.0330487527,
-0.0029047532,
-0.0344485603,
-0.0332392268,
0.029630959,
0.0331738032,
0.0219210982,
-0.0075185224,
-0.0540419035,
-0.0148775866,
0.0136248283,
-0.0485120378,
0.0108016199,
0.0317338966,
0.0108473906,
0.0148240142,
-0.0251891389,
-0.0229043998,
0.0573385619,
-0.0118690403,
-0.0145619307,
-0.0686841905,
-0.0186297931,
-0.0082709463,
-0.0195718724,
-0.0080446033,
0.022640679,
-0.0448306762,
0.0002004456,
-0.0227653403,
0.0186522827,
-0.0417017788,
0.0128709404,
0.0144578004,
-0.0276755858,
-0.0283930898,
-0.01199672,
-0.019426031,
-0.0469812416,
-0.0236452036,
0.0032906476,
-0.0086994078,
0.0145586599,
0.0054637562,
-0.0549363978,
-0.022543801,
-0.011573541,
0.1031228974,
-0.0144432373,
-0.0052661975,
-0.0031997431,
0.016003076,
0.0101399524,
-0.0201745816,
0.0113218213,
-0.0590485074,
0.047950726,
0.0296731703,
0.0027631561,
-0.0010978046,
-0.059221115,
-0.0309183132,
0.0023679694,
-0.0099988542,
-0.0235818308,
-0.0028255926,
0.0058372552,
-0.0310102887,
-0.0213696007,
0.0792625695,
0.0141594186,
-0.0554275028,
0.0146385431,
-0.0112742009,
0.0345807001,
0.0246661175,
0.0255214833,
-0.0652747303,
-0.0115177538,
-0.0219370387,
0.0106186233,
0.0111642964,
0.0447997525,
-0.0268119331,
-0.1002181992,
0.0227214638,
0.0378621034,
-0.0845337734,
-0.0398294106,
0.0029996964,
-0.0469128713,
0.0008534489,
0.0630888045,
-0.0243047401,
0.0188362151,
-0.0284304097,
0.0071011037,
0.0607275739,
-0.0670237914,
-0.045134183,
-0.0612052158,
0.0225990936,
0.0449060723,
0.0296043903,
0.0078122164,
-0.0200407393,
0.0539248735,
-0.0071090814,
0.02558548,
0.0087297484,
0.0383412205,
0.0144848488,
-0.0328182392,
0.0477781631,
-0.0051069553,
-0.0624877363,
0.0262543373,
0.0255347583,
0.0507433675,
-0.0347614363,
-0.0316119231,
0.0006148024,
0.0007403318,
0.0076095047,
0.0371988751,
0.0215413589,
-0.0563212186,
0.0138038136,
0.0039631976,
-0.0283835456,
-0.1122741103,
-0.0147091961,
0.0570199341,
-0.0106732026,
-0.0560915209,
-0.0966260508,
-0.0883048177,
0.0015339811,
0.0722933412,
0.0268865507,
0.0193228498,
-0.018270446,
-0.0190991759,
-0.0036596358,
0.0460268371,
-0.0385131985,
-0.0063500009,
-0.014164865,
0.0136806294,
0.0152923455,
-0.0323207006,
0.028690232,
0.0151920337,
-0.0751874819,
-0.059680663,
-0.0183944385,
0.0494716838,
0.0136127155,
0.0627104491,
-0.0292387269,
-0.0141745899,
0.0340765156,
0.0450885184,
-0.0371000953,
-0.024928255,
0.0264804475,
-0.0149777578,
0.0023621933,
0.0391334295,
0.001047741,
0.0097205108,
0.0236836448,
0.0324293785,
0.0078170812,
-0.0016451729,
0.0501101874,
-0.0146861244,
0.0297611728,
-0.0253200997,
0.007674431,
-0.0221853908,
0.0317034051,
-0.0315908492,
-0.0156895258,
-0.0503302477,
-0.0061909799,
0.0264604706,
0.0296332221,
0.0213074163,
-0.0173250474,
-0.0440528803,
-0.0226472933,
0.0133274114,
-0.0451761931,
-0.046149388,
0.0303507056,
0.0001218344,
0.0388981923,
-0.0249003917,
0.0227067489,
0.0141015565,
0.0232505221,
-0.027707573,
0.0386140347,
0.0175127946,
0.018549066,
-0.0091292188,
0.0185047835,
-0.0322581567,
0.0127825467,
-0.0401754864,
-0.0124899978,
0.0739639774,
-0.0199328642,
0.0901427865,
-0.011235157,
-0.0021795142,
0.0101345517,
0.0220577661,
0.0220514927,
-0.0373783484,
-0.0277212597,
-0.0228113085,
-0.0121759698,
-0.0060340846,
0.0782283247,
-0.003869575,
0.0168701764,
0.0103519177,
0.0225697979,
-0.0547826588,
-0.0327816643,
-0.0287654009,
-0.0233990531,
-0.014080504,
-0.02450612,
0.0041578868,
-0.0590975322,
0.0083725089,
0.0058063627,
0.0026665733,
0.008673477,
0.0344583169,
0.061494682,
-0.0501020402,
-0.000218355,
-0.0340964831,
0.0126382597,
0.0381813906,
-0.0208880045,
-0.034635894,
-5.787064851e-33,
0.0009773326,
0.0157486945,
0.0262122713,
-0.0410867967,
-0.0211454351,
-0.0461009964,
-0.018302355,
-0.0233942885,
0.0159381684,
0.0300254133,
-0.0075122113,
0.0245395452,
0.0408149399,
0.016718477,
-0.0081937909,
0.0054251631,
0.0777062476,
-0.0164641757,
-0.001309022,
0.0145121915,
0.0401565954,
0.0234904177,
0.0990294293,
-0.0836553052,
0.0081605799,
-0.040301159,
-0.0042457455,
0.0193399563,
-0.0059839059,
0.0052435724,
-0.0006820842,
0.0356755108,
0.0166528858,
-0.0556171499,
-0.0191354603,
-0.0215529576,
-0.0437310562,
-0.0009767542,
0.0223886613,
0.0674031004,
-0.0259904973,
-0.0728322193,
0.0455803126,
-0.0306046456,
0.0326256305,
0.0022657206,
0.013132046,
-0.0757624209,
-0.0149711976,
0.003175868,
-0.0637184754,
-0.0161576886,
0.0179174449,
0.0306590348,
0.0215879623,
0.1154870316,
-0.0370571502,
-0.0500697605,
-0.0623876229,
0.0143333282,
-0.05244679,
0.0318442099,
0.0405714139,
0.0038637195,
-0.0177958868,
0.0224383622,
0.0661865324,
-0.0020648788,
0.0299465954,
0.0315541364,
0.01959984,
0.0007072942,
0.0642858967,
0.0665888861,
0.0777851939,
-0.007629293,
-0.0790596828,
0.0318547487,
-0.0525632948,
-0.0419265218,
0.0412155725,
-0.0301398747,
-0.0234230664,
-0.0069966754,
-0.0101559181,
-0.0229754131,
-0.0346196108,
-0.0375934355,
0.017634498,
0.0304636583,
0.003714686,
0.0150638511,
0.0481030829,
-0.0079797544,
0.0474642776,
0.0112462202,
-0.0346513018,
0.003051152,
-0.0135495495,
-0.0007041531,
0.0114101879,
-0.0659128949,
0.0325113945,
0.051870741,
0.0260128099,
0.0132270986,
0.0789873376,
0.0003857075,
-0.0450922847,
0.0041520875,
-0.0388968587,
0.0247989688,
-0.0004514966,
0.0290507488,
0.0179842021,
0.0042043286,
-0.0323492885,
0.0265702233,
0.0321490653,
-0.0406309105,
0.0091911647,
0.0469580851,
0.0271547511,
0.0538023263,
-0.0403248183,
-0.0033054901,
-0.0274743792,
-0.0203348491,
0.0366103798,
-0.0832980499,
-0.0340309031,
-0.0235275831,
2.61e-7,
0.0079633119,
0.0223714728,
-0.0032758261,
0.0063661919,
-0.0012030805,
-0.0023497902,
-0.0200828481,
0.0492016859,
0.0030250696,
0.02399767,
0.0135641033,
0.0345976017,
0.0917392373,
0.0350789502,
-0.1376894265,
-0.0436038896,
0.0006171654,
0.0269597266,
-0.0681641176,
-0.0005831305,
0.0066246348,
0.080290921,
-0.0102093024,
-0.030336665,
0.0204797462,
-0.0260915216,
0.0528978109,
0.004544796,
-0.0219361242,
0.0059625483,
0.0203146115,
0.0348717608,
-0.0456405059,
0.0345285609,
-0.0351568013,
0.0115818391,
0.0492265932,
0.0380894691,
0.0218428224,
-0.0212473087,
0.0220274031,
-0.0178892128,
-0.0210707244,
0.0214263983,
0.0145791033,
-0.063749209,
-0.015947951,
0.0524847247,
-0.0228771549,
0.0287301298,
0.0421617925,
-0.038283851,
-0.0230546445,
0.0106399935,
0.0599500872,
0.0100713689,
-0.0382366404,
-0.0201679077,
-0.0392044894,
-0.015187107,
-0.0147733269,
-0.0041870819,
-0.0107997935,
0.0464186668,
0.0348209962,
-0.0216107816,
0.0134825697,
2.132573127e-34,
0.0385253504,
0.0207716692,
0.0448583812,
-0.027334176,
-0.0069354256,
-0.0424949974,
0.0358163007,
-0.0160985626,
-0.009375954,
-0.030882882,
-0.0277859345
] |
http://arxiv.org/abs/2401.16078v1 | http://arxiv.org/pdf/2401.16078v1 | Understanding the effects of word-level linguistic annotations in under-resourced neural machine translation | 2024-01-29T11:39:46 | 2024-01-29T11:39:46 | Víctor M. Sánchez-Cartagena, Juan Antonio Pérez-Ortiz, Felipe Sánchez-Martínez | cs.CL | cs.CL | This paper studies the effects of word-level linguistic annotations in
under-resourced neural machine translation, for which there is incomplete
evidence in the literature. The study covers eight language pairs, different
training corpus sizes, two architectures, and three types of annotation: dummy
tags (with no linguistic information at all), part-of-speech tags, and
morpho-syntactic description tags, which consist of part of speech and
morphological features. These linguistic annotations are interleaved in the
input or output streams as a single tag placed before each word. In order to
measure the performance under each scenario, we use automatic evaluation
metrics and perform automatic error classification. Our experiments show that,
in general, source-language annotations are helpful and morpho-syntactic
descriptions outperform part of speech for some language pairs. On the
contrary, when words are annotated in the target language, part-of-speech tags
systematically outperform morpho-syntactic description tags in terms of
automatic evaluation metrics, even though the use of morpho-syntactic
description tags improves the grammaticality of the output. We provide a
detailed analysis of the reasons behind this result. | Understanding the effects of word-level linguistic annotations in under-resourced neural machine translation: This paper studies the effects of word-level linguistic annotations in
under-resourced neural machine translation, for which there is incomplete
evidence in the literature. The study covers eight language pairs, different
training corpus sizes, two architectures, and three types of annotation: dummy
tags (with no linguistic information at all), part-of-speech tags, and
morpho-syntactic description tags, which consist of part of speech and
morphological features. These linguistic annotations are interleaved in the
input or output streams as a single tag placed before each word. In order to
measure the performance under each scenario, we use automatic evaluation
metrics and perform automatic error classification. Our experiments show that,
in general, source-language annotations are helpful and morpho-syntactic
descriptions outperform part of speech for some language pairs. On the
contrary, when words are annotated in the target language, part-of-speech tags
systematically outperform morpho-syntactic description tags in terms of
automatic evaluation metrics, even though the use of morpho-syntactic
description tags improves the grammaticality of the output. We provide a
detailed analysis of the reasons behind this result. | [
0.0196756087,
0.0156242521,
0.0088307709,
0.0378964543,
-0.0006282261,
0.0191867743,
0.0043426007,
0.0014666077,
-0.043995548,
-0.0587159395,
-0.0353258476,
-0.0356662348,
0.0598703772,
-0.0331718698,
0.0583624914,
-0.0586132035,
0.0264517516,
-0.0142689226,
-0.0234484617,
0.0295706466,
0.0048541715,
-0.0040070806,
0.0231613796,
0.0421388708,
-0.0694373995,
0.0387309939,
-0.0028688123,
0.0694978461,
0.0102566704,
-0.0323418938,
-0.0076637035,
0.0186648648,
-0.0382653065,
0.0433225632,
0.0000016935,
-0.0359468311,
-0.000354419,
0.0397907794,
-0.003789833,
0.0298200399,
0.0636944324,
-0.0407570414,
-0.0304927379,
0.022997763,
-0.0024791653,
-0.0382015444,
0.0688712373,
0.0303412899,
0.0670834258,
0.0676610693,
-0.0267150979,
-0.0181161314,
-0.0043071755,
-0.0411061272,
-0.0050198706,
-0.0158704221,
-0.0453568324,
-0.0275875963,
-0.0237667952,
0.0175184309,
0.0084427735,
0.0480145179,
-0.0170823243,
-0.0254380591,
0.0013282092,
0.0129400846,
-0.0390771627,
-0.0442207605,
0.0179386456,
0.0385018289,
-0.0580293089,
0.0319208987,
-0.0238501094,
-0.0245447736,
0.01317757,
-0.071958974,
-0.011561119,
0.0450135283,
0.0112302788,
0.0276235752,
0.0483659916,
0.0570062734,
0.0303399749,
-0.0139643345,
-0.0065142289,
0.0562405884,
0.0179183166,
-0.0192405395,
-0.0403671525,
0.0309712011,
0.0091795959,
-0.0498714373,
0.0155829061,
-0.0033694399,
0.033136908,
0.0186864696,
-0.0659372807,
-0.0629082695,
0.0173125565,
-0.0674467012,
0.0471119732,
0.0323690102,
-0.0379444212,
0.0380637981,
-0.0410148129,
0.049383115,
0.0060713193,
-0.0032241943,
-0.0719805211,
-0.019611368,
-0.0094827805,
0.0224462915,
-0.0462021232,
0.0111551275,
-0.0135527626,
0.0143177258,
-0.0425265841,
0.0031213029,
0.0013476578,
0.0027913472,
0.0178395323,
-0.0309405979,
-0.0796713904,
0.0177190024,
-0.0428342745,
0.0222207531,
-0.0213731937,
-0.0177434888,
0.0403400585,
-0.0717581213,
-0.037164744,
0.0605276339,
-0.021917697,
-0.0419434048,
-0.0137499273,
0.0833411068,
0.0142668951,
-0.0486252829,
-0.020587733,
-0.0258756336,
-0.0393591151,
-0.023555195,
0.0163098909,
-0.0361586586,
0.0485546701,
0.0026886868,
0.0109547544,
-0.0131595498,
0.0097250743,
0.0021178718,
-0.0090691596,
0.0580284148,
-0.0483357608,
-0.0105224634,
-0.0050685639,
-0.0146307377,
0.0108002685,
0.053807877,
0.0645798594,
-0.0131903579,
0.0143593196,
0.0050223996,
-0.0439041257,
0.0191702545,
-0.0014816823,
-0.0000107984,
-0.0336662456,
-0.0431931764,
-0.0012043586,
0.0433269255,
0.0794219747,
0.0607211031,
-0.0234133713,
0.0229103845,
0.0454635732,
0.0599349998,
0.0149109708,
0.0350058004,
0.0578705408,
-0.0054462021,
0.0081766201,
0.035236299,
-0.0125232358,
0.0330459736,
-0.0154252574,
0.0136271762,
0.0331671201,
0.0192159787,
-0.0269291233,
-0.0754425004,
-0.0452968366,
-0.0428885035,
0.0174700562,
-0.0645792186,
0.0138378777,
0.045229353,
-0.0133534102,
0.0544443466,
0.0105323223,
-0.0381576233,
0.0464182347,
-0.0410440601,
-0.0696131736,
0.0495744981,
0.0493499897,
0.0002739399,
-0.0633982047,
0.0058972496,
-0.0224773735,
0.0611203276,
0.0702743381,
0.03033874,
0.0153723713,
-0.0391320139,
0.0049145082,
-0.0449090078,
0.0372072943,
0.0018410202,
-0.0158563349,
-0.0109819844,
0.0451214015,
-0.0119252857,
-0.1023457497,
0.0385488756,
0.0130493473,
-0.017795153,
0.0204293989,
-0.0016393675,
0.0306121185,
0.0117277298,
-0.0194683652,
0.0404623151,
0.004098698,
-0.0209819786,
-0.0304874387,
0.0252767261,
-0.0358426385,
-0.0429805331,
0.0233626664,
-0.0047819861,
0.0036752957,
-0.0301540382,
-0.0197460689,
-0.0583363175,
-0.0144083984,
-0.0849915519,
0.0376395099,
-0.0480949543,
0.0493699908,
0.0286780559,
0.0163755342,
-0.031909015,
0.0296497792,
0.0081652859,
0.0078120986,
-0.0029885422,
-0.03585805,
0.0313870236,
-0.005871234,
-0.0541164093,
0.00789358,
0.0169408843,
0.0116295535,
0.0155897355,
-0.0002435455,
-0.0136422478,
-0.0103454217,
-0.0668782294,
0.0145354774,
-0.0236245357,
0.0535889566,
0.0427872092,
0.0494983569,
0.02274739,
-0.0084430929,
-0.0364978872,
0.0462226085,
-0.0224633832,
0.0088850968,
0.0658104941,
0.0323062092,
0.0319090672,
-0.0404850654,
-0.004778293,
0.0102829775,
0.0366378427,
0.0576158129,
-0.0632189885,
-0.0823228434,
0.0121715507,
-0.0444672741,
0.0084188953,
0.0338093266,
-0.0874137133,
0.0061239623,
-0.0019136735,
0.0272402409,
0.0004298176,
-0.0121141654,
0.008397161,
0.0649276823,
0.0122741517,
-0.0476559326,
-0.0030545872,
-0.0243203193,
-0.0189455189,
-0.0018708523,
-0.0239581726,
-0.0017930674,
0.0917394981,
-0.0137775894,
-0.0142411385,
-0.0184678864,
-0.0004031976,
-0.0373405144,
-0.0000516858,
0.0122809,
-0.0065743271,
-0.0052952254,
0.0301960949,
-0.0109677371,
-0.0458734818,
0.0815069303,
-0.0472028702,
-0.0352012813,
-0.0485101081,
0.0033156939,
-0.0254002232,
0.0211625956,
-0.0193575118,
0.0269399527,
0.0143553009,
-0.0579207167,
-0.0194386821,
0.0110062715,
0.0168096162,
-0.0028045112,
0.013884482,
-0.0502460897,
-0.0135321114,
-0.0524020344,
0.0054313778,
-0.0104262196,
-0.0462393872,
-0.0095118415,
0.0173897743,
0.0282496549,
-0.0403417051,
0.0253486447,
0.0163751654,
0.0049706125,
0.0767330527,
-0.030481372,
-0.0011722217,
0.001612202,
0.0128249219,
-0.0502000153,
-0.0011710899,
0.0047449465,
-0.0097612254,
0.0206063539,
0.0029465824,
-0.0248294491,
0.0307268389,
-0.0427920856,
-0.0330301896,
0.034424305,
-0.0165154506,
-0.0810781717,
-0.0014027491,
0.0510950349,
-0.0611157082,
-0.0286930706,
0.1066525206,
0.0296484865,
-0.0195465628,
0.0001742216,
0.0050122165,
0.060120251,
0.0211963933,
0.0306900628,
-0.0369632393,
-0.0191820897,
0.0137397284,
0.0603458621,
0.0316954702,
0.0249844156,
-0.023083616,
-0.0853431672,
0.010621734,
0.0393690616,
-0.048053734,
-0.0141384816,
0.0000798909,
-0.0206026342,
0.0053693322,
0.0176394992,
-0.0386265516,
0.0061545232,
-0.0070985686,
0.0314324461,
-0.0433575884,
-0.0765778273,
-0.0314841121,
-0.0526661053,
0.0540875681,
0.0915556699,
0.0227663107,
-0.025946781,
-0.0151610291,
0.09632916,
0.0049053263,
0.0222617742,
0.0080053946,
0.0448323339,
-0.0106206201,
-0.0296404269,
0.0355345719,
-0.0414805077,
0.0206001177,
0.0303292908,
-0.0456523709,
0.0516145416,
-0.0126290526,
0.0135476058,
-0.0006343101,
-0.0111540547,
-0.0031177034,
0.0503544211,
0.0732911304,
-0.0494767874,
-0.0057255062,
-0.01155922,
-0.0072329771,
-0.0400600098,
0.0112750642,
0.0244269744,
-0.0527418777,
-0.062784113,
-0.1060443521,
-0.1076460779,
0.0060586031,
0.0107486341,
0.002152184,
-0.0242639501,
0.0027237884,
-0.0267678648,
0.0084891226,
0.0478786752,
0.0157808606,
0.0005513037,
-0.0142791588,
-0.0121472245,
0.0552253947,
0.0281043444,
0.0159457531,
0.0051496252,
-0.0542717203,
-0.0442074351,
-0.0062311566,
0.0684499666,
0.0199562553,
0.0287711322,
-0.0622712597,
-0.0818084627,
-0.0298497286,
0.0471603684,
0.0036347064,
-0.0026060054,
0.0476222411,
0.0105494121,
-0.0126838246,
0.0456143506,
-0.021599967,
-0.0035448398,
-0.0211515371,
0.0465435609,
-0.0123918746,
-0.0395633839,
0.0323954448,
0.0189016052,
0.0536785908,
-0.0520402901,
0.003045412,
-0.0292345639,
-0.0115995444,
-0.0288392324,
-0.0289389417,
-0.0609717332,
-0.0148160327,
0.0385241546,
-0.0023093002,
0.0044057234,
0.009345565,
-0.0500082374,
0.0098080272,
0.0354668498,
-0.0194586217,
0.0252414849,
0.0032814988,
-0.0364898369,
0.0408796631,
-0.0287749842,
-0.0131578119,
0.0050649694,
-0.0278968234,
-0.0602898411,
0.0096891839,
-0.0152267199,
0.021745177,
-0.0061629042,
0.0380868949,
0.0066496627,
0.0301255081,
-0.0269695483,
0.0041088513,
0.0597043186,
-0.0139975315,
0.1116215214,
-0.0200739931,
-0.0051200679,
-0.0278013032,
0.048365131,
-0.0173315257,
-0.0423466451,
0.0067367447,
-0.0050922665,
-0.0155883851,
0.0000213082,
0.0620656051,
0.0156560242,
-0.0019710185,
-0.0210156254,
-0.0021152347,
0.0069662901,
-0.0043124272,
-0.0185457859,
-0.0028962891,
-0.0347754844,
-0.0273852162,
-0.0399741568,
-0.0624163821,
-0.0071643777,
0.0283145066,
0.0231033713,
0.0112209283,
0.0368251391,
0.0210494604,
-0.0388653539,
0.0002168432,
-0.043626219,
-0.02941758,
0.0329577737,
0.0113668349,
-0.0347960778,
-5.877294695e-33,
-0.0636355504,
-0.0228537172,
-0.0120254783,
-0.0350761674,
-0.0268796869,
0.0214969274,
-0.0457808189,
0.0040236195,
0.0079831211,
0.0192062017,
0.0022860973,
0.0139408503,
0.0267860573,
-0.0054330034,
0.0206149835,
-0.0170992874,
0.0513381474,
-0.0025666673,
-0.027760148,
-0.0014073827,
0.0228354726,
-0.005296987,
0.095246315,
-0.053994745,
0.0770558789,
-0.0062870155,
-0.0280245114,
0.0038499015,
0.0361038186,
0.0228345506,
0.0049299747,
0.0191836581,
0.0158424769,
-0.0087323142,
-0.04106424,
-0.0372314602,
-0.0701989755,
-0.0317137204,
0.0092064692,
0.0342231393,
-0.0605648011,
-0.014261663,
0.0515793785,
-0.0370015316,
-0.0079169227,
0.0002044524,
0.022134481,
-0.0207999684,
-0.0283402875,
-0.0079570413,
-0.0614520386,
0.0412545912,
0.0154540231,
0.0081511913,
0.039122615,
0.0489972495,
-0.0136795649,
-0.0466310531,
-0.0606656075,
-0.0002283748,
-0.0122771524,
0.0582935438,
0.0443812124,
0.0237204246,
-0.0494424291,
0.0090060346,
0.0464358255,
0.0449372903,
0.0499148518,
0.0291143823,
0.0441719331,
-0.0336940773,
0.0906998813,
-0.0080137448,
0.0247175265,
0.0122147398,
-0.043346677,
0.0305885989,
-0.0121109392,
0.0248911362,
-0.0142559158,
-0.0052855001,
-0.0468611009,
-0.0262136906,
-0.0324771889,
-0.0637431368,
-0.0284530148,
-0.035388995,
-0.0073267929,
0.0133602759,
-0.0219087899,
0.029865576,
0.0484578498,
0.018604923,
0.0380213149,
-0.0186435618,
-0.0020817218,
0.0375042707,
-0.0250666551,
0.0296414383,
0.0108437669,
-0.0449834354,
0.0106077231,
0.0286966618,
0.0533659048,
-0.0081841694,
0.0581723079,
0.0736361668,
-0.0291947909,
-0.0002830203,
-0.0413763635,
-0.0052872216,
-0.0390313491,
0.0007304725,
0.0239794478,
-0.016132066,
-0.0241746493,
0.0224639941,
0.0399522558,
0.0336875468,
-0.0159774367,
-0.0082319817,
0.0413917899,
0.0239189863,
-0.0428505577,
0.0336309113,
-0.0000353479,
0.0187313948,
0.0309986379,
-0.0263372231,
-0.0325249918,
0.0014399398,
2.58e-7,
0.0620810911,
0.0438477732,
0.0101364655,
-0.0113952663,
0.0083374092,
0.0090287272,
-0.0443580374,
0.027070161,
-0.0046246122,
0.0339528583,
0.0070031644,
-0.0238304939,
0.0526615642,
0.051910162,
-0.0930205658,
-0.048574876,
-0.0297236405,
0.0498743206,
-0.0213470235,
-0.0274305269,
0.0478812046,
0.0894759744,
0.0012928634,
0.0228423551,
-0.0134905931,
-0.0376276039,
0.0605871566,
0.0092864484,
-0.0187470838,
0.0000888392,
0.0847127214,
0.0979715958,
-0.0200460479,
-0.0120592518,
-0.0557317585,
-0.0031071573,
0.0520799533,
0.0536445752,
-0.027264189,
-0.0202060938,
0.0249975566,
0.0029777891,
-0.0031383208,
-0.0228867047,
0.0201264061,
-0.0334094577,
-0.0102378242,
-0.0183829218,
0.008661638,
0.0206852127,
0.0379822105,
-0.0328927152,
-0.0560196564,
0.0098749772,
0.0523153208,
-0.0161840655,
-0.0644756928,
-0.0582761057,
-0.0456203744,
0.0203857087,
0.0250728671,
-0.0332301557,
-0.007939063,
0.0458921492,
0.0542887636,
-0.0337330475,
-0.0233852174,
2.45651159e-34,
0.0431558415,
-0.030694684,
0.0347900316,
0.0122930473,
0.0011460518,
0.0164895747,
-0.0092054568,
0.0116356779,
-0.0021074526,
-0.0218504407,
-0.0439510457
] |