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/2401.13260v2 | http://arxiv.org/pdf/2401.13260v2 | MF-AED-AEC: Speech Emotion Recognition by Leveraging Multimodal Fusion, Asr Error Detection, and Asr Error Correction | 2024-01-24T06:55:55 | 2024-05-28T16:47:15 | Jiajun He, Xiaohan Shi, Xingfeng Li, Tomoki Toda | cs.CL | cs.CL, cs.MM, cs.SD, eess.AS | The prevalent approach in speech emotion recognition (SER) involves
integrating both audio and textual information to comprehensively identify the
speaker's emotion, with the text generally obtained through automatic speech
recognition (ASR). An essential issue of this approach is that ASR errors from
the text modality can worsen the performance of SER. Previous studies have
proposed using an auxiliary ASR error detection task to adaptively assign
weights of each word in ASR hypotheses. However, this approach has limited
improvement potential because it does not address the coherence of semantic
information in the text. Additionally, the inherent heterogeneity of different
modalities leads to distribution gaps between their representations, making
their fusion challenging. Therefore, in this paper, we incorporate two
auxiliary tasks, ASR error detection (AED) and ASR error correction (AEC), to
enhance the semantic coherence of ASR text, and further introduce a novel
multi-modal fusion (MF) method to learn shared representations across
modalities. We refer to our method as MF-AED-AEC. Experimental results indicate
that MF-AED-AEC significantly outperforms the baseline model by a margin of
4.1\%. | MF-AED-AEC: Speech Emotion Recognition by Leveraging Multimodal Fusion, Asr Error Detection, and Asr Error Correction: The prevalent approach in speech emotion recognition (SER) involves
integrating both audio and textual information to comprehensively identify the
speaker's emotion, with the text generally obtained through automatic speech
recognition (ASR). An essential issue of this approach is that ASR errors from
the text modality can worsen the performance of SER. Previous studies have
proposed using an auxiliary ASR error detection task to adaptively assign
weights of each word in ASR hypotheses. However, this approach has limited
improvement potential because it does not address the coherence of semantic
information in the text. Additionally, the inherent heterogeneity of different
modalities leads to distribution gaps between their representations, making
their fusion challenging. Therefore, in this paper, we incorporate two
auxiliary tasks, ASR error detection (AED) and ASR error correction (AEC), to
enhance the semantic coherence of ASR text, and further introduce a novel
multi-modal fusion (MF) method to learn shared representations across
modalities. We refer to our method as MF-AED-AEC. Experimental results indicate
that MF-AED-AEC significantly outperforms the baseline model by a margin of
4.1\%. | [
-0.0119507033,
0.0525838882,
0.0284637716,
-0.0004117851,
0.0088345297,
-0.0039761788,
0.0067427973,
-0.0620116591,
-0.0086683286,
-0.0799033195,
-0.0669557527,
-0.0201480873,
0.0325282253,
0.0027994288,
-0.0168921351,
-0.0513930246,
0.0039792587,
0.0206920318,
0.0177728478,
-0.0153074283,
0.0324278325,
0.0407745503,
-0.029429432,
0.0321543217,
-0.0614962354,
0.0077414894,
-0.0128043676,
0.0524622723,
0.0566487461,
-0.021283498,
0.0051400848,
0.0121023636,
0.0099437023,
0.0163698625,
0.0000021876,
-0.0253260825,
-0.0394020192,
0.0106071029,
-0.0129979448,
-0.0361544676,
0.0685920268,
-0.0022792413,
-0.0295692384,
0.0451557375,
0.0243706256,
0.0210229289,
0.0520841666,
0.0875540748,
0.0414438024,
0.0432134792,
-0.0259313546,
-0.004286374,
0.0151696764,
-0.0024620455,
-0.0417147055,
-0.0064612431,
-0.0223336779,
0.0002370682,
-0.010526632,
-0.0461858697,
-0.04814348,
0.0231994297,
-0.0157055836,
-0.006256822,
0.0254221167,
0.0081390776,
-0.0079914918,
-0.0398552231,
-0.0627843514,
0.0518940948,
-0.0061368975,
0.0136660095,
-0.0060686525,
-0.0088825636,
-0.0188830476,
0.0010123483,
0.0068674628,
0.0273579415,
-0.0345953107,
0.0149893155,
-0.0131931128,
0.0149581134,
0.0312169213,
0.0197232999,
-0.0230112225,
-0.0106693292,
-0.0340223908,
-0.0517796986,
0.0292687677,
-0.0252706911,
-0.0437170677,
0.0223598368,
0.0381787755,
-0.0215785149,
-0.0046178033,
-0.0013985892,
-0.0629289523,
-0.0745531246,
-0.0035043785,
-0.0536668301,
-0.0046910984,
0.0300478768,
-0.0978824794,
0.0460649319,
0.0063132625,
0.0221886765,
0.0480558798,
0.0106181642,
-0.031022843,
-0.0216878708,
-0.0596769154,
0.0332511701,
0.01502077,
0.0515851378,
-0.0283252466,
-0.0656731054,
-0.0821732208,
0.0050884048,
-0.0316425711,
0.0213371124,
0.0114517426,
-0.007797705,
-0.0008288892,
-0.0016732299,
0.0166353695,
0.0520654283,
0.0295708384,
0.0671743229,
0.0071940389,
-0.0337401405,
0.0161540546,
0.0294349249,
0.0200833846,
-0.0081895124,
0.0375996046,
0.0461487398,
-0.0342249721,
-0.0293778274,
-0.0391946137,
-0.0055495319,
-0.0570832714,
-0.0371036604,
-0.01627771,
0.0147616342,
0.0518110469,
-0.013836273,
0.0252349637,
0.0046806266,
0.0153797632,
-0.0436083041,
0.0087533863,
0.0867240801,
-0.0282958876,
-0.0044294908,
0.0273278765,
0.0038940122,
0.0055423439,
0.037406452,
0.0282345489,
0.0090462938,
0.0029193498,
-0.0076435455,
-0.0169428065,
-0.0150500126,
-0.0394843146,
-0.0200655051,
0.0036382801,
-0.0836699754,
-0.0200421195,
0.0046453006,
0.077501744,
0.0514063649,
-0.0307689086,
0.0159899052,
0.023818206,
0.0225776676,
0.0131124407,
0.0163870584,
-0.0252327025,
0.0067513082,
0.0167917293,
-0.0042865477,
-0.0252345782,
0.0052686892,
-0.0506646074,
-0.000204847,
0.0655250475,
0.0594703183,
-0.0105672805,
-0.0391652845,
-0.033929646,
-0.0491566695,
0.0058525275,
0.0445915908,
0.0389053486,
-0.0262135472,
-0.0500948727,
0.023699481,
0.008749431,
-0.0859713331,
0.015900692,
-0.0122755626,
-0.0073270081,
0.0443309881,
0.0097624697,
-0.073966071,
-0.0183964167,
-0.0279953629,
-0.0467268154,
0.0240647588,
0.003493835,
0.017252963,
0.0706941262,
-0.1040024236,
-0.0231559016,
0.0026531597,
0.0094082337,
0.0347704701,
-0.0933300629,
0.0234438702,
0.0116295973,
0.0239235535,
0.038470652,
-0.0083017545,
0.0369464606,
-0.0155898985,
-0.0177069679,
0.0066731195,
0.0519430116,
0.0364358686,
-0.0429502502,
-0.0104817711,
0.0270582028,
0.0119575299,
-0.050399702,
0.047991477,
-0.0650635511,
-0.0353560224,
0.0737221316,
-0.0293099359,
0.0110020125,
0.0126328059,
-0.0190480538,
-0.0740833282,
0.0658789277,
0.0061066984,
0.0520015284,
-0.022299435,
0.0520131513,
0.02760024,
0.024443971,
0.0318276323,
0.0130127817,
-0.0015042837,
-0.0120538492,
-0.017756762,
0.0105846627,
0.0030245106,
-0.0313190632,
-0.0326062664,
0.0150373317,
-0.017140314,
0.0495099761,
0.0063458704,
-0.0321559571,
-0.0332523584,
-0.0756359547,
-0.1353506446,
-0.014265541,
-0.0126387114,
0.0001693849,
-0.0308056977,
0.0077997847,
-0.0059654783,
-0.0220070723,
0.0000969099,
-0.0006976089,
0.0268526562,
-0.0269942898,
-0.045430664,
-0.020257324,
0.0074397828,
-0.0476210862,
0.0230718907,
-0.0324670784,
0.0069446554,
-0.0271342788,
-0.0303239748,
-0.0286026485,
-0.0096149147,
-0.0409195013,
0.0384536646,
-0.0086325463,
-0.0477057882,
-0.0091704549,
-0.0000347737,
0.0369419642,
0.0326108411,
0.0121870982,
0.0637043193,
0.0269846059,
-0.0294046886,
-0.0329113714,
0.0579365678,
0.0281324349,
0.048501797,
0.0045199995,
0.0289859846,
-0.0362153389,
0.0289873518,
0.0221882947,
0.0217451248,
-0.0114809759,
-0.0477376282,
0.0219176412,
-0.0596372858,
-0.0116248606,
-0.0515720025,
-0.013524319,
-0.0373859294,
-0.010704753,
-0.0270988513,
0.0161229521,
-0.049654793,
-0.0427088141,
-0.074744679,
-0.0669833049,
-0.0242039058,
0.0011445963,
-0.039552357,
0.0207503345,
-0.0162945632,
-0.0264127403,
-0.0029917043,
-0.0396873243,
-0.0024618157,
-0.0603844151,
0.0667112991,
-0.0089939348,
-0.021727737,
0.0181736089,
-0.0193566587,
-0.0003046869,
-0.0430973358,
-0.0548355244,
0.018863868,
-0.0212107804,
-0.0288948044,
-0.0468124673,
-0.0500394665,
0.0591574237,
-0.0141851828,
0.0344963633,
-0.0062773228,
-0.0105485106,
0.0518615507,
0.0010958044,
0.0661788285,
-0.0140880905,
-0.0106400568,
0.0304049719,
0.036954619,
0.0748262033,
-0.0181423258,
-0.0564094745,
0.010710692,
0.022559274,
0.0564203598,
-0.0112132933,
-0.0193907898,
0.0193400048,
-0.017820999,
0.0854469761,
0.0958619192,
0.0190307163,
-0.0118647851,
-0.0382970236,
-0.0370599031,
-0.0153932106,
0.0018728772,
-0.0230369344,
-0.0298676025,
-0.0600894578,
0.0673002377,
0.0218398087,
0.023145657,
0.0351216905,
0.0212268643,
-0.0700708404,
0.0123581477,
0.0302321874,
0.0154185751,
-0.0325427614,
-0.0358401537,
0.0056508314,
-0.030213898,
0.0333072729,
-0.054257255,
-0.0062835636,
-0.040224161,
0.0735310689,
0.0041285744,
-0.0794223323,
-0.0222974811,
-0.011458993,
0.0737020224,
0.0296178479,
0.0668294355,
-0.0403073095,
-0.0566659793,
0.04544539,
0.0083852205,
-0.0023286985,
-0.0190813635,
0.015451856,
-0.0227459036,
-0.0063202661,
-0.0025068447,
0.0196427666,
0.002736184,
-0.032472644,
-0.0296874829,
0.0674327612,
-0.0468374975,
0.006473125,
-0.0005435421,
0.0227886736,
0.0174007583,
0.032287728,
-0.0148994187,
-0.0151413083,
-0.0008478193,
0.021547759,
0.0654323846,
0.0034553341,
0.0165734701,
0.0397235788,
-0.0173272863,
-0.0290805437,
-0.0220061354,
-0.0489563681,
-0.0356007516,
0.0254895967,
0.0092194313,
-0.0225134213,
-0.0300031733,
0.0101383394,
-0.0142455678,
0.0357556082,
0.0337461121,
0.0207369551,
-0.0485284738,
0.0507400036,
0.0137085551,
-0.0430029333,
-0.0121804057,
0.0061858254,
0.0004717668,
-0.0652412176,
0.0201351158,
-0.011822029,
0.0044608698,
-0.0246848334,
0.0056901611,
-0.0286956597,
0.0565754622,
0.0430205017,
-0.0289955866,
0.0303673558,
0.0789301693,
0.0033413749,
-0.0372542925,
-0.0107461661,
-0.049093321,
0.0248505604,
-0.0435979739,
0.0207062718,
-0.0043944721,
-0.0716624632,
-0.009141353,
0.0335654803,
0.0577424727,
-0.013777365,
0.0241648909,
-0.0407860167,
0.0028601256,
0.0335009508,
0.0368864536,
-0.0342113674,
0.0066707092,
-0.0060712313,
-0.0033724839,
-0.0173316952,
0.0215178337,
0.002734598,
-0.0213001054,
0.0166443195,
0.0868509412,
0.046325352,
0.0018487955,
-0.0478653274,
0.0431186222,
0.006605539,
0.0143885743,
0.060313385,
-0.032283131,
0.0087990314,
0.0577459559,
-0.0019868128,
-0.0533552244,
-0.005395425,
0.0382610671,
-0.0118134981,
0.0470823497,
-0.0180689637,
0.0199244358,
0.049112834,
0.0022855354,
0.0704299212,
-0.0181335695,
0.0240053236,
-0.0383693837,
0.0260386802,
-0.0328229479,
-0.0038868005,
0.0408169366,
-0.0076636379,
-0.0123557774,
0.0082558794,
0.0301944278,
-0.0885352492,
0.0456699692,
-0.0426809974,
0.0122467279,
-0.0563433729,
-0.0184366573,
0.0622696728,
-0.00275757,
-0.0244898573,
0.0191132184,
-0.0291121006,
-0.0481921025,
0.0949953198,
-0.0249258839,
0.0053959922,
-0.0095865056,
-0.0250160936,
0.0100663351,
-0.0192504432,
0.0343461074,
-0.0050028297,
-0.0274793822,
-0.0369458683,
0.034300033,
0.0359401889,
-6.52899611e-33,
-0.0160541181,
0.0175198298,
-0.0437526293,
-0.0258492138,
-0.0559116341,
0.0461405292,
0.0271528382,
0.0275225788,
0.0331667922,
0.019621281,
-0.0042653698,
0.0165025033,
0.0284225959,
0.0129491584,
0.0154140629,
0.0117393984,
0.0567279384,
0.0226679035,
0.0266001169,
-0.0497435071,
0.0480123386,
0.0280607305,
0.0257034935,
-0.1043724641,
-0.0032262949,
0.0615914166,
-0.059730567,
-0.0092020547,
0.0347238965,
0.0600974001,
0.0070377751,
0.0481961444,
-0.0087178135,
0.0313151926,
-0.0179407895,
0.0322513171,
-0.0474701375,
-0.0019910708,
0.0009132359,
0.0284538213,
-0.0204898324,
-0.0114704166,
0.0584639162,
0.036650572,
0.0055732573,
-0.0437026396,
0.0222818051,
-0.02120387,
0.0111815007,
-0.037929628,
-0.0211033542,
0.0483912341,
-0.0215809774,
-0.0179058574,
0.0632367954,
0.0389714018,
0.0285035819,
-0.0016613794,
-0.0532785133,
-0.0104541741,
-0.0220197458,
0.0034186202,
0.0176128615,
0.0196362492,
-0.0490882546,
0.0683723241,
0.0248946585,
-0.0547758192,
-0.0315375589,
-0.0092845764,
0.0415422693,
0.0164024327,
0.0138296532,
0.0723091811,
0.0074024601,
-0.0044315169,
0.0085045109,
0.0222971942,
0.0275773779,
0.0314203501,
0.0155313369,
0.0636688918,
-0.0248706583,
0.0289865583,
-0.0516095199,
0.0143172164,
-0.0437378846,
-0.0088680638,
0.036929749,
0.0339323692,
0.0141118569,
0.0286840685,
0.0023739454,
0.0153244138,
0.0214690603,
0.0252086651,
0.0057675354,
0.0462508649,
-0.0167189632,
0.0156063298,
-0.0318149105,
-0.0386089683,
0.0716618523,
0.034901917,
0.0259980597,
-0.0334620699,
-0.0002371611,
0.0826858878,
-0.0726271123,
-0.0006309445,
0.0191639178,
0.0144826127,
0.0028007952,
-0.0034321316,
0.0021088319,
0.0069232038,
-0.0299006701,
-0.0137006007,
-0.0129310871,
0.0808235556,
-0.0149535788,
-0.0364113301,
-0.0145095773,
0.0228508636,
-0.0208674613,
0.0136213405,
0.0282648951,
-0.0733816251,
0.000765554,
0.0015129893,
0.002237875,
-0.0120190363,
3.061e-7,
-0.0486536063,
-0.0155884763,
0.0246182848,
-0.0809387043,
-0.0206793901,
0.0559460819,
-0.0347081609,
0.023315426,
0.0147038233,
-0.0027446002,
0.0299574733,
-0.0115061272,
0.070054099,
0.0406547487,
-0.0392455198,
-0.022785183,
-0.0255233999,
0.0068585277,
0.0228967313,
0.0515174456,
0.0820234641,
0.0814853832,
0.0241194535,
-0.0420677625,
-0.0048543988,
-0.0043850946,
0.0482941866,
-0.0092724087,
-0.0273009297,
-0.0045356778,
0.0031189416,
-0.0262982119,
-0.0230284985,
0.0150495805,
-0.0078082336,
-0.0020466635,
0.0641979799,
0.0082986485,
-0.017163353,
-0.0605725348,
-0.0010848737,
0.0865905061,
-0.042184785,
0.0341096148,
-0.0016989099,
-0.0020499604,
-0.027714802,
-0.0424840748,
-0.0499346964,
0.0206496343,
0.042161271,
-0.0052891634,
-0.0397704542,
0.0143780513,
-0.0096610878,
-0.0059440187,
-0.0609705746,
-0.1237239763,
-0.0178764276,
0.0549181327,
-0.0502010435,
0.0178107806,
0.0025666722,
0.0200713221,
0.0466322079,
-0.0806749314,
-0.0192491468,
2.177744252e-34,
0.0267989803,
-0.0454109684,
-0.010454637,
0.002999072,
0.0236210227,
-0.0121392747,
-0.0458406992,
-0.0173472576,
0.0244578477,
-0.0374254771,
0.000651205
] |
http://arxiv.org/abs/2401.13256v1 | http://arxiv.org/pdf/2401.13256v1 | UniMS-RAG: A Unified Multi-source Retrieval-Augmented Generation for Personalized Dialogue Systems | 2024-01-24T06:50:20 | 2024-01-24T06:50:20 | Hongru Wang, Wenyu Huang, Yang Deng, Rui Wang, Zezhong Wang, Yufei Wang, Fei Mi, Jeff Z. Pan, Kam-Fai Wong | cs.CL | cs.CL, cs.AI | Large Language Models (LLMs) has shown exceptional capabilities in many
natual language understanding and generation tasks. However, the
personalization issue still remains a much-coveted property, especially when it
comes to the multiple sources involved in the dialogue system. To better plan
and incorporate the use of multiple sources in generating personalized
response, we firstly decompose it into three sub-tasks: Knowledge Source
Selection, Knowledge Retrieval, and Response Generation. We then propose a
novel Unified Multi-Source Retrieval-Augmented Generation system (UniMS-RAG)
Specifically, we unify these three sub-tasks with different formulations into
the same sequence-to-sequence paradigm during the training, to adaptively
retrieve evidences and evaluate the relevance on-demand using special tokens,
called acting tokens and evaluation tokens. Enabling language models to
generate acting tokens facilitates interaction with various knowledge sources,
allowing them to adapt their behavior to diverse task requirements. Meanwhile,
evaluation tokens gauge the relevance score between the dialogue context and
the retrieved evidence. In addition, we carefully design a self-refinement
mechanism to iteratively refine the generated response considering 1) the
consistency scores between the generated response and retrieved evidence; and
2) the relevance scores. Experiments on two personalized datasets (DuLeMon and
KBP) show that UniMS-RAG achieves state-of-the-art performance on the knowledge
source selection and response generation task with itself as a retriever in a
unified manner. Extensive analyses and discussions are provided for shedding
some new perspectives for personalized dialogue systems. | UniMS-RAG: A Unified Multi-source Retrieval-Augmented Generation for Personalized Dialogue Systems: Large Language Models (LLMs) has shown exceptional capabilities in many
natual language understanding and generation tasks. However, the
personalization issue still remains a much-coveted property, especially when it
comes to the multiple sources involved in the dialogue system. To better plan
and incorporate the use of multiple sources in generating personalized
response, we firstly decompose it into three sub-tasks: Knowledge Source
Selection, Knowledge Retrieval, and Response Generation. We then propose a
novel Unified Multi-Source Retrieval-Augmented Generation system (UniMS-RAG)
Specifically, we unify these three sub-tasks with different formulations into
the same sequence-to-sequence paradigm during the training, to adaptively
retrieve evidences and evaluate the relevance on-demand using special tokens,
called acting tokens and evaluation tokens. Enabling language models to
generate acting tokens facilitates interaction with various knowledge sources,
allowing them to adapt their behavior to diverse task requirements. Meanwhile,
evaluation tokens gauge the relevance score between the dialogue context and
the retrieved evidence. In addition, we carefully design a self-refinement
mechanism to iteratively refine the generated response considering 1) the
consistency scores between the generated response and retrieved evidence; and
2) the relevance scores. Experiments on two personalized datasets (DuLeMon and
KBP) show that UniMS-RAG achieves state-of-the-art performance on the knowledge
source selection and response generation task with itself as a retriever in a
unified manner. Extensive analyses and discussions are provided for shedding
some new perspectives for personalized dialogue systems. | [
0.0645177886,
0.0341688879,
-0.0037039986,
0.025664892,
-0.0364239477,
-0.030926764,
0.00083485,
-0.0256254375,
-0.0049899756,
-0.0483137369,
-0.0547474176,
-0.0247516725,
-0.0396084189,
0.0329078212,
0.0244926717,
-0.0264297128,
0.0663090572,
-0.0133378422,
0.044849582,
-0.0107306717,
0.0047595999,
0.0174090266,
-0.0235877298,
0.0624149293,
-0.0229311697,
0.0063512004,
-0.0049903314,
0.0007757485,
0.013258853,
-0.0300532468,
0.0231111217,
0.0425395183,
0.0197424311,
0.0283080302,
0.0000017242,
-0.0444268025,
-0.0127849597,
-0.017601002,
-0.021468224,
-0.0295863412,
0.0377240144,
0.0397292152,
0.0221435577,
0.0247941297,
-0.06945429,
-0.0063192309,
0.0308401883,
0.0363673232,
0.070063971,
0.0892047137,
0.0038890142,
-0.021589227,
-0.00677648,
-0.0268147849,
0.0158943646,
-0.0130522633,
0.0068362448,
0.0006463568,
0.0232897438,
0.027033342,
0.0264465436,
0.0176874138,
0.0141610755,
-0.0370162092,
0.016478043,
0.0130816856,
-0.0505992025,
-0.0096764779,
-0.0301310029,
0.0859736353,
0.063830927,
0.0034957645,
-0.0118981777,
0.0049701934,
0.0260709003,
-0.0107875923,
-0.051151447,
0.0113880886,
-0.0036314947,
0.0052218526,
-0.015549235,
0.0093200598,
-0.0048100092,
0.0305788834,
-0.0350337923,
0.0619284809,
0.015090744,
-0.0267776828,
-0.0320330262,
-0.0060427031,
-0.0038904133,
-0.057458628,
-0.0020400002,
0.0425095186,
-0.0641312003,
0.0244845469,
-0.0384864621,
-0.0173955709,
0.0334761404,
-0.0037760027,
0.0109497346,
0.0318257697,
-0.0299517121,
0.0403557494,
-0.0049898471,
0.0382115059,
-0.0172782578,
-0.0015937441,
-0.063812755,
-0.0213700086,
-0.0427552983,
-0.0088192327,
-0.0277114958,
0.0448637679,
-0.080666177,
-0.0223606434,
-0.0171162914,
0.039420668,
0.0065686023,
-0.0042002876,
0.0307368934,
0.0449765064,
-0.0341564566,
0.0110590244,
-0.0466523208,
0.0320954993,
-0.0533062778,
-0.0001960703,
0.0475976095,
-0.0589575022,
-0.0286320224,
0.044993706,
0.0058771078,
-0.0161658619,
-0.0100246537,
0.0463208742,
-0.0328261666,
-0.0298767593,
-0.0210648645,
-0.0536863729,
-0.0536988638,
-0.0609968118,
0.0131653873,
-0.0364319384,
0.0649947152,
0.011823995,
-0.0581557453,
-0.0175928753,
0.0224432517,
0.02608376,
-0.0318958312,
0.00647027,
-0.0072175935,
-0.0052837534,
0.0231285561,
0.0033695018,
-0.0148991207,
0.0498253033,
0.0154419187,
0.0251964908,
0.0537209064,
0.0161648728,
-0.0256504752,
-0.0171306077,
0.0290919673,
0.0118116178,
0.0128943734,
0.0517131872,
-0.0149753969,
0.0158543084,
0.0199284945,
0.0512035452,
-0.064585261,
-0.016262779,
0.039283108,
0.0084783724,
0.0196237639,
0.0213427749,
0.0583327264,
0.0426916555,
0.0013351149,
0.0199798234,
-0.0091525558,
0.0229193736,
-0.0065435418,
0.0068518789,
0.0768414289,
0.0193365458,
-0.0255420972,
-0.0525271781,
0.0171955284,
0.0100828446,
-0.0119200181,
0.0182618704,
0.0330157354,
0.033387199,
-0.0389987119,
0.0500015914,
0.0151794814,
-0.0744308084,
-0.0025074473,
-0.0223782435,
0.0582627282,
0.0721086338,
0.0642620847,
0.0021700889,
-0.0335915126,
0.0003261955,
-0.0339627452,
0.0421632975,
0.0267306976,
-0.0087226229,
0.0221888162,
-0.0562629066,
-0.036095351,
-0.0348813422,
-0.027310729,
0.0549179092,
0.0047365087,
-0.0042610457,
0.0393816158,
0.0234351344,
0.0234676488,
0.0133492406,
-0.0078582512,
-0.0029590707,
0.0510346778,
-0.0197310541,
0.0561376177,
0.0573971905,
-0.0000466366,
0.0206108373,
0.0494653173,
-0.0244990233,
-0.0140770432,
0.0302767176,
-0.072424829,
-0.0637450889,
0.0343583263,
0.0232573394,
0.0354410633,
-0.0227632653,
-0.0113356356,
-0.086742878,
0.0372331068,
-0.0434526242,
0.0127956914,
-0.080711998,
-0.0181463454,
0.0677619278,
0.0308899172,
-0.0423404463,
0.0544018708,
-0.0017052197,
-0.0181455277,
0.0569502786,
0.0221798029,
0.0253621992,
-0.0745829493,
-0.0259829815,
0.00982833,
-0.0146203507,
0.0043535512,
0.0576101169,
0.0039522722,
-0.0435311347,
-0.0379365422,
-0.1144566312,
0.0231162906,
-0.0123829534,
-0.0034713626,
0.0019076469,
0.0150631303,
-0.000872344,
-0.0370453931,
-0.0557231456,
-0.0135533325,
0.075423114,
0.0147896111,
0.0258811265,
0.0673711076,
-0.0002860703,
-0.0558806248,
0.0121688312,
0.030288117,
0.0480205156,
-0.0149646169,
0.0404314511,
-0.0539662577,
-0.0514054894,
-0.0812303871,
0.0368771926,
-0.0047023296,
-0.0268970858,
0.0342128277,
-0.0277192574,
0.0473085903,
0.0066550705,
0.0518014207,
0.0257610921,
0.020288229,
-0.0583013892,
-0.0975141451,
0.0520424321,
-0.0452471189,
0.0065987902,
-0.0141108278,
0.0077027436,
0.0273737814,
0.0109121967,
0.0411455445,
0.0001235189,
-0.0520989373,
0.0448473506,
0.0578587092,
-0.0496956185,
-0.0148982303,
-0.0168951936,
-0.0189932417,
0.0189315062,
-0.0061294404,
-0.0274419729,
0.0690639168,
-0.0532868914,
-0.0188581869,
-0.0476782247,
0.0119599337,
-0.0372168534,
0.0332860835,
-0.0220516231,
0.0245747566,
-0.0155426571,
0.0085856579,
-0.0066912118,
-0.009351884,
0.0072012255,
-0.059607856,
0.0164978504,
-0.0233536623,
0.0052405759,
0.00686887,
-0.0413152874,
-0.0733088627,
-0.0370822251,
0.0188507382,
-0.0085152965,
0.0251992233,
-0.0110791102,
-0.0353273489,
0.0036507708,
0.0530156717,
0.0029357036,
-0.0233081952,
0.0070380117,
-0.0195914879,
0.0482998863,
0.0507449992,
0.0287651755,
-0.0071473108,
0.0204695892,
0.0052822656,
0.0453070775,
0.0245486461,
-0.0168977771,
-0.00714229,
-0.0201258436,
0.039228525,
-0.0246143509,
0.0007697919,
-0.0101196421,
0.0341198072,
-0.0272121932,
-0.0191354323,
0.0931504071,
0.026271658,
-0.0225583594,
-0.0256631617,
-0.0315899961,
0.0240306072,
0.0012654016,
0.0492923819,
-0.082140781,
-0.0191289596,
-0.0276659466,
0.0426660962,
-0.0085270014,
-0.0216195453,
-0.0333875157,
-0.0998935252,
0.0394128673,
0.0054320544,
-0.0515564531,
-0.0166842397,
-0.0155973323,
0.004312519,
0.0081152013,
0.028779339,
0.036710158,
0.0190192536,
0.0129459724,
0.0018661969,
0.0143131055,
-0.0464331992,
-0.0180236157,
-0.00007817,
0.0417318046,
0.0562445372,
0.0314093977,
0.0584903024,
-0.0537469201,
-0.0113901421,
-0.0105785336,
0.0288512446,
-0.0237951335,
0.0355562828,
-0.0187365431,
0.0048984136,
-0.0326824188,
0.0005336251,
0.0007036523,
-0.0163018666,
0.0152019644,
0.0162660666,
-0.0477479398,
0.0287974067,
-0.0366174802,
0.0327401347,
0.0446536988,
0.079030551,
0.0180089604,
-0.0327397473,
-0.0123420581,
-0.0074483114,
-0.0057959598,
-0.062532708,
-0.0015294551,
0.0175330359,
0.0107120834,
0.0104965884,
-0.118675366,
-0.0557183735,
-0.0060708341,
0.0345703624,
-0.0240279026,
-0.0004573868,
-0.0459793471,
-0.0007543272,
-0.0109880799,
0.0264456812,
-0.0287731905,
-0.0347198658,
-0.0491640568,
-0.028931167,
0.0443569571,
-0.0778814703,
-0.0111086462,
0.0512843467,
-0.0204483848,
-0.0623928867,
-0.047762394,
0.0848387554,
0.017174311,
0.0709791332,
0.0036295759,
-0.0102834972,
0.0164274033,
0.0272558406,
-0.0378914848,
-0.0137131,
0.0243334193,
0.0281388629,
-0.0266235471,
0.0302510131,
-0.0402694345,
0.0326878093,
-0.0209874641,
0.0859801769,
0.011322964,
-0.0706375912,
0.0158366859,
0.0156426616,
0.0447394699,
0.0009920341,
-0.010019294,
-0.0001102744,
0.0265886206,
-0.036312975,
-0.0388450027,
-0.0687442347,
0.0172494315,
-0.0112714786,
-0.0540045574,
0.0208604038,
-0.0357739888,
-0.0332047045,
0.0511642434,
-0.0163760018,
-0.0276618022,
0.0162328649,
0.0097723911,
0.0094394591,
-0.0126777729,
-0.0382132977,
0.0269757975,
-0.0065291994,
0.0063894941,
-0.0308265742,
0.0879456252,
0.0090861637,
-0.0493886918,
0.028325934,
0.0357690677,
-0.0732643157,
-0.0014484989,
-0.0268321764,
0.0155207468,
0.0243325904,
0.038073618,
0.0462520532,
-0.0054646339,
0.0134186856,
-0.0219658483,
0.0409107693,
0.0147445938,
-0.0397047438,
0.0066387616,
0.015344223,
-0.0390616544,
0.0179949682,
0.0680852756,
-0.0237324219,
-0.0080089988,
-0.0157543346,
-0.0055109896,
-0.0659564063,
-0.0594104193,
-0.0099041956,
0.0392889008,
-0.0221443269,
0.005411318,
-0.0047664447,
-0.0119593432,
-0.0044508236,
0.0225727335,
-0.0379514955,
0.0172951687,
0.0576793104,
0.0825144574,
-0.0276390258,
-0.0349672101,
-0.0383154489,
0.0352546163,
-0.0435183123,
-0.0241897162,
-0.0232250765,
-5.865211348e-33,
-0.0367723592,
-0.011724201,
0.0466025174,
-0.015991861,
-0.021177439,
-0.0513210297,
-0.050738249,
0.0060689221,
-0.0150706144,
-0.0251210835,
-0.0083305305,
-0.0318826288,
0.0122652138,
0.056350518,
-0.0240409374,
0.0105150333,
0.0489022397,
0.0122764921,
-0.0342537388,
0.0534586571,
0.0813039839,
0.0224647895,
0.0329325162,
-0.0752329528,
0.0294107217,
-0.040532589,
0.0051376252,
0.0322370417,
0.0025749872,
0.0755191296,
-0.0124679236,
0.0552947223,
-0.0025663413,
-0.038045384,
-0.0165309422,
0.0084321201,
-0.0516857803,
-0.0277887713,
0.048205737,
-0.0288992561,
-0.0614631251,
-0.0462765694,
0.0164688732,
-0.0366018265,
0.0105275875,
0.034550231,
0.0004451673,
-0.0168988965,
-0.0256286822,
-0.0109171569,
-0.0287917424,
0.0184709318,
-0.001590074,
-0.0202068929,
-0.044201225,
0.0863290057,
0.02412837,
-0.0232345574,
-0.0251620114,
0.0352934375,
-0.0090009784,
0.0030122516,
0.0415500887,
0.0199985728,
-0.0099127302,
0.0358624756,
0.0156272445,
-0.0210874174,
-0.0131625468,
-0.0626897439,
-0.0158458985,
0.0628432259,
0.0118479561,
0.0916818008,
-0.0241871588,
0.0070227059,
-0.0377383381,
0.0286841951,
0.0297718588,
0.0291554295,
0.0367707238,
0.0141141554,
-0.0087083951,
0.0084590726,
-0.0171611384,
-0.0103218891,
-0.0349479578,
-0.0505500622,
0.0523768663,
0.0595312007,
-0.0161033906,
-0.0117833344,
0.0599233136,
-0.0104375789,
0.0070889439,
0.0636517555,
-0.0432124287,
-0.0277116168,
-0.0197269972,
0.0098002041,
-0.0300962683,
-0.0049125724,
0.0801968053,
-0.0336021595,
-0.0450676531,
0.0244428888,
0.005717339,
-0.003323203,
-0.0145384036,
0.0149915777,
-0.0137852831,
0.0748564526,
-0.0697087869,
-0.0263428539,
0.0265636873,
0.0238794573,
-0.0462883711,
0.0452618785,
-0.0008029277,
0.0066294153,
-0.0431158654,
-0.0485013016,
0.0486301295,
0.0022561566,
-0.0169930868,
0.045696497,
0.0117326956,
0.0139014469,
-0.004027334,
-0.0053646225,
0.0009523423,
0.0441011898,
2.468e-7,
-0.0014353001,
0.0010300077,
0.021197889,
0.0007290169,
0.0077802888,
0.0110529363,
0.0020522205,
0.0323508754,
0.0392601378,
0.0056443987,
0.0064939065,
0.0328122564,
0.0405110307,
0.0246988591,
-0.0936652869,
-0.0295891315,
-0.0849055946,
-0.0229295585,
-0.0463001318,
-0.013453641,
0.0564818382,
0.0616010241,
-0.0038163376,
0.0190457478,
-0.0160672497,
-0.0493570901,
0.0723774284,
-0.0195798688,
-0.02336777,
0.0062873759,
0.0649448335,
-0.0025107891,
0.0044874279,
-0.0002757898,
0.0016519177,
0.051634144,
-0.0118523501,
0.0572912619,
0.0238628034,
-0.0814557821,
0.0120243328,
0.0143784666,
0.0287406538,
-0.0498990193,
0.048352737,
-0.0402794667,
0.0097818784,
-0.0107177207,
-0.020050928,
0.0040262467,
-0.0043714126,
-0.0329348333,
-0.0263753589,
0.0254514795,
0.0226161014,
0.0059381048,
0.0125960279,
-0.0887464806,
-0.0177778509,
-0.0038481031,
-0.0707198232,
-0.0228010993,
-0.0447474644,
-0.0303006563,
0.0053357952,
-0.0478143021,
0.0005349324,
1.880327307e-34,
0.0193100329,
-0.0023308324,
-0.00031829,
0.0127170105,
0.0374187753,
0.0041708406,
-0.0338255018,
0.0151748434,
0.0359415635,
-0.0884774178,
-0.0356810875
] |
http://arxiv.org/abs/2401.13246v3 | http://arxiv.org/pdf/2401.13246v3 | SEER: Facilitating Structured Reasoning and Explanation via Reinforcement Learning | 2024-01-24T06:10:51 | 2024-06-04T06:14:56 | Guoxin Chen, Kexin Tang, Chao Yang, Fuying Ye, Yu Qiao, Yiming Qian | cs.CL | cs.CL | Elucidating the reasoning process with structured explanations from question
to answer is crucial, as it significantly enhances the interpretability,
traceability, and trustworthiness of question-answering (QA) systems. However,
structured explanations demand models to perform intricately structured
reasoning, which poses great challenges. Most existing methods focus on
single-step reasoning through supervised learning, ignoring logical
dependencies between steps. Moreover, existing reinforcement learning (RL)
based methods overlook the structured relationships, underutilizing the
potential of RL in structured reasoning. In this paper, we propose SEER, a
novel method that maximizes a structure-based return to facilitate structured
reasoning and explanation. Our proposed structure-based return precisely
describes the hierarchical and branching structure inherent in structured
reasoning, effectively capturing the intricate relationships between different
reasoning steps. In addition, we introduce a fine-grained reward function to
meticulously delineate diverse reasoning steps. Extensive experiments show that
SEER significantly outperforms state-of-the-art methods, achieving an absolute
improvement of 6.9% over RL-based methods on EntailmentBank, a 4.4% average
improvement on STREET benchmark, and exhibiting outstanding efficiency and
cross-dataset generalization performance. Our code is available at
https://github.com/Chen-GX/SEER. | SEER: Facilitating Structured Reasoning and Explanation via Reinforcement Learning: Elucidating the reasoning process with structured explanations from question
to answer is crucial, as it significantly enhances the interpretability,
traceability, and trustworthiness of question-answering (QA) systems. However,
structured explanations demand models to perform intricately structured
reasoning, which poses great challenges. Most existing methods focus on
single-step reasoning through supervised learning, ignoring logical
dependencies between steps. Moreover, existing reinforcement learning (RL)
based methods overlook the structured relationships, underutilizing the
potential of RL in structured reasoning. In this paper, we propose SEER, a
novel method that maximizes a structure-based return to facilitate structured
reasoning and explanation. Our proposed structure-based return precisely
describes the hierarchical and branching structure inherent in structured
reasoning, effectively capturing the intricate relationships between different
reasoning steps. In addition, we introduce a fine-grained reward function to
meticulously delineate diverse reasoning steps. Extensive experiments show that
SEER significantly outperforms state-of-the-art methods, achieving an absolute
improvement of 6.9% over RL-based methods on EntailmentBank, a 4.4% average
improvement on STREET benchmark, and exhibiting outstanding efficiency and
cross-dataset generalization performance. Our code is available at
https://github.com/Chen-GX/SEER. | [
0.0259409752,
0.0170899238,
-0.002796765,
0.0401882008,
0.0157927945,
0.0339803286,
-0.0378322676,
-0.0009559853,
-0.0014088996,
0.0238269009,
0.0104138916,
0.0161706973,
-0.0092813885,
0.0449563488,
-0.0105962772,
-0.1088761315,
0.0442556702,
-0.0458057374,
0.0423872322,
0.0038386406,
0.0083829267,
0.0290784463,
-0.0168034118,
0.037429437,
0.010741909,
-0.0368982702,
-0.0364667438,
0.0222596359,
0.0217374004,
-0.0275836457,
0.0179344602,
0.0030009521,
0.0458003469,
0.0214583054,
0.0000022423,
-0.0237421263,
-0.0303236712,
0.0119945286,
-0.0365656614,
-0.0134377386,
-0.0902134851,
-0.0236833654,
0.0355321132,
0.0313926414,
-0.0236715879,
0.0059899618,
0.0388038903,
0.050907366,
0.0813694596,
0.0678131282,
0.0063746367,
-0.0502722524,
0.0105839074,
-0.0183917657,
0.0574377887,
-0.0356967375,
-0.0017860277,
0.0578474328,
0.0117117949,
0.0190381426,
-0.0109369857,
-0.0095103271,
0.0725167096,
-0.0248491447,
-0.0089283343,
0.0150400596,
0.0111944778,
0.0014549945,
-0.0043438277,
0.0719113648,
0.0165896211,
-0.0055039483,
0.0113100121,
-0.0288529024,
-0.015453306,
0.0007051735,
-0.0286532342,
0.0075873639,
-0.0472303182,
-0.0041774758,
0.0042920955,
0.0051659285,
0.0218369346,
0.0249202289,
-0.0229563732,
0.0951108858,
-0.0388779715,
-0.0105343703,
-0.0015251578,
0.0109030223,
0.0161034614,
-0.0563171729,
-0.0044556693,
0.0609827638,
-0.0758647546,
0.0313426927,
0.0123710725,
0.0236012414,
0.0063617351,
0.029397035,
0.0060319551,
0.020378707,
0.0227355454,
0.064873144,
-0.0375318229,
0.0448152162,
-0.046969872,
0.0419100262,
-0.0187457819,
-0.0137622673,
-0.070117116,
-0.0331010781,
-0.0219661128,
0.0517940372,
-0.037158031,
-0.0040473063,
-0.0308053605,
0.0316598974,
-0.0300281867,
0.0170887224,
0.0197412893,
0.0326111838,
0.0022067525,
-0.0019378571,
-0.0817486793,
-0.0076592956,
-0.125661999,
0.0075929556,
0.0234251525,
0.0128905,
-0.0376217738,
-0.0070074331,
-0.0003106387,
-0.0148569643,
-0.0128976721,
0.0477315038,
0.0076859812,
0.0176866185,
-0.0389218032,
-0.0588448271,
0.0215116031,
-0.05457348,
0.0039092116,
-0.0426465385,
0.0174225029,
-0.01305243,
-0.0208678618,
0.0149559118,
0.0077585615,
0.0285853855,
-0.0268562324,
0.0773227885,
-0.0305826403,
-0.0301531497,
-0.0240965411,
0.0301896017,
-0.0249549914,
0.046614971,
0.0148518411,
0.0161391329,
0.0251965392,
-0.0063131382,
0.0069534802,
-0.0535070077,
0.0353513174,
-0.0196785536,
0.0306567941,
0.0589116365,
-0.0534059294,
0.0206563286,
0.0065222126,
-0.0011274769,
-0.0474385768,
0.0029535752,
-0.0199590456,
0.0117470846,
0.0236062892,
-0.0364282504,
0.0152351772,
-0.0076046446,
-0.0012762788,
-0.0070902235,
-0.0457410067,
0.0567106269,
-0.0172578674,
0.0181426927,
0.0032558516,
-0.0315194316,
-0.0292917751,
-0.0631293654,
0.0065588444,
0.0279152095,
0.0174755715,
0.048454944,
0.0306723155,
0.0399597064,
-0.0355297327,
0.0253754053,
0.0438335426,
-0.0063512954,
-0.0071614487,
-0.0504006483,
0.0914042145,
0.0600902252,
0.0372865647,
-0.0446275175,
-0.0080823014,
-0.0191770475,
0.0280307457,
0.0139684938,
-0.0243501179,
-0.0223192088,
0.0335608646,
-0.0350568257,
-0.0498897582,
-0.0452238843,
-0.0076862667,
0.0050129662,
0.0633594543,
-0.0143624814,
0.0002688943,
-0.0488729142,
0.0267981682,
0.0142186796,
0.0313852243,
-0.0096022924,
0.0325660184,
0.0026375111,
0.017475836,
0.0360155925,
0.0194175709,
0.0198686812,
0.0053482833,
-0.0047075637,
-0.03494348,
-0.0419234037,
-0.0714340061,
-0.0564066358,
-0.031282194,
-0.0271012075,
-0.0324041173,
0.0028529088,
-0.065392822,
-0.0343413316,
-0.0158799551,
-0.0419852771,
0.021523172,
-0.0120908394,
-0.0410245582,
0.0293442551,
-0.0341376774,
-0.0304736458,
0.0791300535,
-0.002355725,
0.0098009231,
-0.0167791005,
-0.0118296845,
0.0086364979,
-0.0133893592,
0.0725811571,
0.0032276257,
-0.0226893146,
-0.0264109839,
0.0350988656,
0.0204139147,
-0.0668052286,
0.0163853597,
0.0090273749,
-0.0165133774,
-0.0151901692,
0.0324946456,
0.0378553756,
0.0233938266,
0.044458352,
-0.0538877733,
-0.0526047349,
0.0236664433,
-0.0099353772,
0.0147679308,
0.043792922,
0.0010304627,
-0.0213555098,
-0.0288076997,
-0.0188831035,
-0.0121587794,
0.026517598,
0.0084364871,
-0.0072393189,
-0.0670813993,
-0.0410985053,
-0.0346197225,
-0.0062800585,
-0.0156193068,
0.0103115113,
0.0444923043,
-0.0221805144,
-0.0306584146,
0.0566859357,
-0.0068993396,
0.0350076295,
0.0020938176,
0.0200224761,
-0.0631894171,
0.0333512984,
-0.0348645411,
0.0101342956,
-0.0274819918,
-0.040550448,
-0.0051071998,
0.0634660572,
-0.0124978246,
0.0331545174,
-0.0538616925,
0.0133702941,
0.0204024557,
-0.0160676651,
-0.0487136208,
-0.0060295518,
0.0208089836,
0.0713045821,
-0.0346671455,
-0.0429168716,
0.1056932434,
0.0316726379,
-0.0105747087,
-0.0185236931,
-0.0097003859,
-0.0587335154,
0.0818357915,
-0.001042155,
0.0110334745,
0.0149696926,
-0.0464710183,
-0.0216904897,
-0.0510884114,
0.0509966016,
0.0028866557,
0.0286926199,
-0.0301765744,
-0.0141986255,
0.0270752944,
-0.0409688167,
-0.0334339961,
-0.0016512802,
-0.0270599164,
0.0114489039,
-0.0153397433,
-0.0347491279,
0.0006289049,
-0.0252078902,
0.0490648076,
0.0162907671,
-0.0064138169,
-0.0276368931,
-0.0284314696,
0.0486557037,
0.0608972721,
0.022525765,
0.0589116625,
-0.0454843529,
0.0274144653,
0.0617296919,
0.0104051335,
-0.0141402157,
-0.0072927382,
0.0059797475,
-0.006573556,
-0.0360497013,
0.0152248517,
0.0202332549,
0.0398983918,
-0.0927151069,
-0.0268363543,
0.0864188895,
0.0369077958,
0.0047755064,
-0.0015540371,
-0.0451342575,
0.0657024384,
0.0281874835,
0.04395115,
-0.0684382766,
-0.046037063,
0.0037293194,
0.0292963907,
0.0368940644,
-0.0298730303,
0.0345171019,
-0.0931148231,
0.0191347748,
-0.0257469174,
-0.0436509885,
-0.042506095,
-0.0370926857,
-0.0185999386,
0.01870602,
-0.0129307322,
0.0178017281,
-0.0129790939,
-0.0038438179,
0.0163078494,
-0.0219239723,
0.051976338,
0.0488205031,
-0.0435522683,
-0.0185164344,
0.0881113857,
0.021549087,
0.0207775794,
-0.037953876,
-0.0376116596,
-0.0020317202,
-0.0008742949,
-0.0205642786,
0.0418555588,
0.0154827433,
-0.0092146546,
-0.0355298184,
0.0553626344,
-0.0525465421,
-0.0179957673,
0.0352069736,
0.0571246929,
-0.0149253821,
-0.0142770801,
-0.0231467616,
0.002665675,
0.0034613907,
0.0488721617,
0.004457965,
0.0099622821,
-0.0864009634,
-0.0435931198,
-0.0132144447,
-0.040054433,
0.0228194278,
0.0051098703,
0.0090638911,
0.0500902981,
-0.0570378378,
-0.0516916141,
0.0403185599,
0.0117147518,
-0.0541261658,
-0.018232746,
0.0057484391,
-0.0059132655,
-0.047163751,
0.0118850963,
-0.0323472321,
0.0323359817,
-0.0146867009,
-0.0083173728,
-0.0319379754,
-0.012696065,
-0.0301312469,
-0.0224595349,
-0.0199001059,
0.0046464037,
-0.0616172031,
0.0407026373,
0.0126884189,
0.0920402706,
-0.0150409779,
-0.0045512458,
0.0159226265,
-0.0575692356,
-0.0316854902,
-0.0305158533,
0.0473673232,
0.0170617681,
0.0048387456,
0.0344372094,
-0.0762621462,
-0.0069975778,
-0.0162919611,
0.0651036426,
-0.0398190916,
0.0158924852,
0.0233730935,
-0.0308076926,
0.0190318953,
0.0486314557,
-0.0120944064,
-0.0203946009,
0.0659043118,
-0.0294286422,
-0.0298419762,
-0.0193438102,
0.0323983021,
0.0148537131,
-0.0584698021,
0.0180376302,
-0.0007573935,
-0.0056539248,
0.0042872308,
-0.0065754107,
-0.0004589336,
0.0587006249,
0.0166855063,
0.0007875672,
-0.0180742946,
-0.0056454185,
0.0205952898,
0.0284407716,
0.0697140992,
0.004596876,
0.0870586559,
0.0329303592,
0.0035509595,
-0.0261022374,
-0.0264308341,
-0.0640099645,
-0.0511445254,
0.0187784601,
-0.0183946341,
0.0020454326,
0.028751608,
0.0233019162,
0.0195023362,
-0.0190121736,
0.0394267514,
0.0256059691,
0.0179778207,
-0.0239814017,
0.0632831678,
0.0503858589,
0.0065246997,
-0.0080697984,
0.0438698567,
-0.0064594736,
0.0193303097,
-0.0159423333,
-0.0203673057,
-0.0124478918,
-0.0234853718,
-0.0099559827,
0.0610130094,
0.0109036248,
0.0105915992,
-0.020739425,
-0.0507561192,
-0.0244744234,
0.0340097584,
-0.0167875905,
0.0209777243,
0.0459339097,
0.017324565,
0.0510368608,
-0.0602415241,
-0.0537405498,
-0.0296062417,
-0.0479920879,
-0.0360825621,
-0.0021520469,
-6.542058056e-33,
-0.0166534744,
-0.0451262668,
0.0234833937,
-0.0185406674,
0.0195278469,
0.0194416661,
-0.0643071979,
-0.032129366,
-0.0259052087,
0.0583906919,
0.0034912953,
0.0298125874,
0.0060192687,
0.0301743727,
0.0125425868,
-0.026485607,
-0.001711502,
0.0398987345,
-0.003939589,
0.0352187082,
0.0437413156,
-0.0048767929,
0.0288328771,
-0.0365873612,
0.0287002176,
0.0275575016,
-0.0302417427,
0.0477432385,
-0.0641359314,
0.0538926534,
0.0042323186,
-0.0060748653,
0.0196592901,
-0.0304305777,
0.0182597488,
0.023994511,
-0.1556432396,
-0.0362056009,
0.0223670527,
-0.0231578387,
-0.1585219502,
-0.0021877068,
-0.009611018,
-0.0113213537,
0.0917998552,
-0.0663735792,
0.0196463894,
0.0020380449,
0.0050196694,
-0.0557786748,
-0.0447019339,
0.0093900515,
-0.0363629721,
-0.02423108,
-0.0013704792,
0.0590660572,
0.0025650002,
0.0174461696,
-0.0305718631,
-0.0217711385,
0.0111689586,
-0.002751624,
0.018114714,
0.045812279,
-0.0196421146,
0.0317168534,
0.0142100789,
-0.0321909674,
-0.0373754166,
0.0263224803,
-0.0562970564,
0.0527735688,
0.0535404459,
0.0072610555,
0.0227096397,
0.0582962185,
-0.1046558321,
0.0077581601,
-0.0046006967,
0.09277419,
-0.0155477868,
0.0177572016,
-0.0475110635,
-0.0182656217,
-0.0511679724,
0.018570587,
-0.045592498,
-0.0290392023,
0.0239282697,
0.0223376863,
-0.0542989373,
-0.0011333822,
-0.0356171243,
-0.045925431,
0.0158408545,
0.0390355736,
-0.0003486039,
-0.0062092291,
0.0076139607,
0.0295127854,
-0.0171508566,
0.0135289095,
0.0221201237,
-0.0591542572,
0.0215581283,
0.0245542116,
0.046739392,
-0.0169348139,
-0.0511226654,
0.0010034825,
-0.0529810674,
0.0236599352,
-0.0324210301,
0.0032605738,
0.0330058001,
0.0172625389,
-0.0215480272,
0.0544344522,
-0.0046224813,
0.0127290795,
-0.0167468172,
0.0384808965,
-0.0118743069,
0.0314177051,
-0.007922003,
-0.0112034259,
-0.0337986089,
-0.0127724083,
0.0278278291,
-0.0643855855,
-0.0161171276,
0.0452493392,
2.757e-7,
0.0190646593,
0.0259305984,
0.0381851606,
0.0711846426,
0.0301464442,
0.0442430489,
0.0383861102,
-0.0135987978,
0.0027403892,
-0.0767047927,
-0.0004880497,
0.0229384378,
0.0365687087,
0.0364869125,
-0.0442105308,
-0.0116992611,
-0.0581658781,
-0.0464358442,
-0.0215967987,
-0.0210627913,
0.0452421755,
0.0468826815,
0.0254290234,
-0.0131551912,
0.0226796661,
0.0044353413,
0.0304426663,
0.0151795968,
0.0393888466,
0.0018592854,
-0.0326340348,
0.0052441587,
0.0259291176,
-0.0054481835,
-0.0029379104,
-0.0077247135,
-0.0075259609,
0.0602496937,
-0.0026860645,
-0.0375219882,
-0.0524108,
-0.0150368586,
0.0661618635,
-0.0488724187,
0.0706807896,
-0.0460115895,
-0.0156848934,
0.0190653615,
-0.0252678283,
-0.0310610179,
0.0046589742,
0.0015814556,
-0.0104906047,
0.0028810657,
0.005159589,
-0.0143989781,
0.01308994,
-0.0760288537,
-0.0190024786,
0.027143579,
0.0110025443,
0.0281150732,
-0.0085214358,
0.0239544176,
-0.0083107334,
-0.0553614758,
-0.036688339,
2.416335856e-34,
-0.0015447374,
-0.0146317361,
0.0189063251,
0.0316719003,
0.0725220069,
0.0196351539,
-0.0425912477,
-0.017277712,
0.0569590256,
-0.0520960316,
-0.0236157589
] |
http://arxiv.org/abs/2401.13229v1 | http://arxiv.org/pdf/2401.13229v1 | From Random to Informed Data Selection: A Diversity-Based Approach to Optimize Human Annotation and Few-Shot Learning | 2024-01-24T04:57:32 | 2024-01-24T04:57:32 | Alexandre Alcoforado, Thomas Palmeira Ferraz, Lucas Hideki Okamura, Israel Campos Fama, Arnold Moya Lavado, Bárbara Dias Bueno, Bruno Veloso, Anna Helena Reali Costa | cs.CL | cs.CL, cs.AI, cs.LG | A major challenge in Natural Language Processing is obtaining annotated data
for supervised learning. An option is the use of crowdsourcing platforms for
data annotation. However, crowdsourcing introduces issues related to the
annotator's experience, consistency, and biases. An alternative is to use
zero-shot methods, which in turn have limitations compared to their few-shot or
fully supervised counterparts. Recent advancements driven by large language
models show potential, but struggle to adapt to specialized domains with
severely limited data. The most common approaches therefore involve the human
itself randomly annotating a set of datapoints to build initial datasets. But
randomly sampling data to be annotated is often inefficient as it ignores the
characteristics of the data and the specific needs of the model. The situation
worsens when working with imbalanced datasets, as random sampling tends to
heavily bias towards the majority classes, leading to excessive annotated data.
To address these issues, this paper contributes an automatic and informed data
selection architecture to build a small dataset for few-shot learning. Our
proposal minimizes the quantity and maximizes diversity of data selected for
human annotation, while improving model performance. | From Random to Informed Data Selection: A Diversity-Based Approach to Optimize Human Annotation and Few-Shot Learning: A major challenge in Natural Language Processing is obtaining annotated data
for supervised learning. An option is the use of crowdsourcing platforms for
data annotation. However, crowdsourcing introduces issues related to the
annotator's experience, consistency, and biases. An alternative is to use
zero-shot methods, which in turn have limitations compared to their few-shot or
fully supervised counterparts. Recent advancements driven by large language
models show potential, but struggle to adapt to specialized domains with
severely limited data. The most common approaches therefore involve the human
itself randomly annotating a set of datapoints to build initial datasets. But
randomly sampling data to be annotated is often inefficient as it ignores the
characteristics of the data and the specific needs of the model. The situation
worsens when working with imbalanced datasets, as random sampling tends to
heavily bias towards the majority classes, leading to excessive annotated data.
To address these issues, this paper contributes an automatic and informed data
selection architecture to build a small dataset for few-shot learning. Our
proposal minimizes the quantity and maximizes diversity of data selected for
human annotation, while improving model performance. | [
0.0441006646,
0.040157143,
-0.0165786427,
0.0216296334,
-0.0245896876,
0.0200329665,
0.0332479738,
-0.003462143,
0.0037067046,
-0.0865636617,
-0.0265916046,
-0.017154621,
-0.0070529399,
-0.0524075143,
0.0238295328,
-0.0645403937,
0.025226485,
-0.0271341261,
0.0265409555,
0.0009350832,
0.0061218864,
0.010705296,
-0.0019223084,
0.043818064,
-0.0152836181,
0.0195080712,
0.0166685376,
0.0197681133,
-0.00421402,
-0.0441466942,
0.0082079032,
0.0980919749,
0.0013870687,
0.0339345187,
0.0000018305,
-0.0531445891,
-0.0016252308,
0.0256182961,
-0.0659705102,
-0.0440277867,
0.0233896356,
0.012390011,
0.0125588626,
-0.0161798559,
-0.0193092041,
-0.0433112234,
0.052016858,
0.0183719844,
0.0359471105,
0.0504924692,
0.0097234827,
-0.0208109822,
-0.0450025126,
-0.0081784017,
0.0395302474,
-0.0242189858,
0.0348032713,
-0.0244930647,
0.0417384878,
0.0289045963,
-0.0036805754,
-0.0054564294,
-0.0294041876,
-0.0131068183,
0.0658062696,
0.0209414512,
-0.1056521833,
-0.0420306213,
0.0002715414,
0.0999101773,
0.0071752742,
-0.0110822394,
0.0083614951,
-0.0404425785,
0.0140184667,
0.0036176592,
-0.0390209779,
0.0094757779,
-0.0017130636,
0.0409302898,
-0.0178361014,
0.0260529406,
0.0141182542,
0.0248061605,
0.0341883525,
-0.0115852524,
-0.0045998213,
-0.0081599532,
0.0180321056,
0.056335289,
-0.0193312429,
-0.0583897792,
-0.0351807661,
-0.0053251716,
-0.0340897143,
0.0106667476,
-0.0438232087,
-0.0743569806,
0.0283920318,
-0.0259665139,
0.0296643618,
-0.006863229,
-0.0133490451,
0.0746042728,
0.0386541523,
-0.0511981621,
0.0467749238,
0.0098173209,
0.0011726711,
-0.0194432195,
-0.0724333748,
0.0392964333,
-0.0472365171,
0.0398559682,
-0.0028475218,
0.014794155,
-0.0525007173,
-0.0236026067,
-0.0322940126,
-0.0146673359,
0.0189035293,
0.014035658,
-0.0403619669,
0.0060072369,
-0.0093585206,
0.0219255034,
-0.0413650014,
0.002538736,
-0.0009677726,
-0.0906127393,
0.0078193005,
0.0117995003,
0.020251466,
-0.0578183942,
-0.0127698025,
0.037300311,
0.0210251249,
-0.0619609244,
-0.0648669377,
-0.0028893098,
-0.0056175995,
-0.0240604207,
-0.0087426212,
-0.0716117546,
0.0336006694,
-0.0227178596,
0.0215838104,
-0.0137133561,
-0.0009021641,
0.0376183651,
-0.0454258621,
0.0894372687,
-0.0691182762,
-0.0091529079,
0.0266678538,
-0.0212970041,
-0.0127993282,
0.109879531,
0.0309867002,
-0.016612567,
0.0678364262,
0.0122974422,
0.0421151072,
0.0165191442,
-0.0013839932,
-0.0114485743,
0.0187487584,
-0.0054493058,
-0.0330359973,
0.0042636991,
0.0064121657,
0.0326272771,
-0.0414821655,
-0.0083249575,
0.0326709598,
0.0467556566,
-0.0220573246,
-0.0066263373,
0.0112559376,
0.0570772514,
0.0118621551,
-0.0202823263,
-0.0095410589,
0.0462955944,
-0.025061246,
-0.0284563545,
0.0112722348,
0.023434246,
-0.0301807746,
0.0026371789,
-0.0460294336,
-0.0077352421,
-0.0011968095,
0.0118729109,
0.0701212063,
0.0290210117,
0.0216735844,
0.0406374559,
0.0613139234,
-0.0255813096,
-0.013709222,
-0.0128186271,
0.0215981882,
0.05521762,
0.0278918073,
-0.0369894058,
-0.0032306097,
-0.0589573421,
0.0403913036,
0.0372651704,
0.0470886827,
-0.0355225503,
0.0152562102,
-0.0568790399,
-0.0560212582,
0.046733249,
0.0255678073,
0.0082780505,
0.0245724972,
-0.009575435,
-0.0131780151,
-0.0348867178,
-0.0406853631,
0.051784385,
0.0117852427,
-0.0340822637,
0.0393547565,
0.0046262173,
0.0307742581,
0.0106799025,
-0.0270339549,
-0.0063188085,
-0.0128768878,
-0.0211788658,
0.0109862881,
0.0035604713,
0.0136570521,
-0.035467755,
0.0075531085,
-0.0681142509,
-0.0455978028,
0.0151916817,
-0.0534687266,
-0.0304940846,
0.000664749,
-0.0403983369,
-0.0274901725,
-0.0535366014,
-0.0306736995,
-0.0239843689,
0.0215536691,
-0.0407746062,
0.0622095503,
0.0310064573,
-0.0012586765,
0.0068316166,
-0.0411374755,
0.0173287876,
-0.0833807066,
-0.0586072281,
-0.0195387248,
0.0108530112,
0.0203369856,
0.050255511,
0.0063082632,
-0.0339560322,
0.0232446138,
-0.070091702,
-0.007231262,
-0.0198663808,
-0.0042187879,
-0.0339194611,
0.0003830744,
-0.0154123297,
-0.0434679203,
0.0080013676,
0.0102845579,
-0.0323081538,
0.0462501459,
0.0400302447,
-0.0062244162,
0.0362408794,
-0.0085015055,
-0.0197530873,
0.0131593086,
0.0433331393,
0.0354932472,
-0.0412741527,
-0.0183723122,
-0.0215512794,
-0.0156742986,
0.0217471085,
0.0380686447,
-0.0076582846,
0.0112939421,
0.0107937744,
-0.0403837152,
0.0083192848,
0.0018420308,
0.023596514,
0.0382770486,
0.0188707914,
-0.0325121358,
-0.0005724706,
-0.0055111106,
0.0492579527,
-0.0029106692,
-0.0108596859,
-0.0041030352,
0.0452255197,
-0.0073842607,
-0.0323083065,
-0.0528204776,
0.0200585201,
-0.0423963107,
-0.0313134678,
-0.0653115511,
0.0029274295,
-0.0344857797,
0.0301371831,
-0.0247154105,
-0.0162917133,
0.06275177,
0.0354944952,
0.0113348644,
-0.0284388717,
0.0459204465,
-0.0180640258,
0.0349566638,
0.0027012352,
-0.0255425777,
0.0028825363,
0.0094401501,
-0.023407001,
-0.0131313726,
0.0664619133,
-0.0250737593,
0.0109499525,
-0.0521542691,
0.0372859314,
-0.0158056077,
-0.0498972721,
-0.0291359816,
-0.0534066185,
-0.0680592731,
0.0028472387,
0.0104358466,
0.0228082854,
-0.0062831803,
-0.0374986753,
0.000293098,
0.0499293469,
0.0013463356,
-0.0095186476,
0.003123987,
0.0353539363,
0.0317188315,
0.0090371845,
0.0118034529,
-0.0048405174,
0.0134483324,
0.0333869532,
0.023117749,
0.0132423323,
-0.0094488962,
0.0030916219,
0.036542818,
-0.0295393746,
-0.0310992338,
0.0114305792,
0.0599520132,
-0.0429978073,
-0.0076954686,
0.0751981363,
0.0022802642,
-0.0627444312,
0.0017513261,
-0.0008436972,
0.0368944183,
0.0381311774,
0.0498740226,
-0.1047950909,
-0.0593081005,
-0.0019415786,
0.0479224399,
-0.042179361,
-0.0697267503,
0.0002018879,
-0.1342710257,
0.074202165,
-0.0088743502,
-0.0656156465,
0.0489250124,
-0.0727778524,
0.0422109775,
-0.0771475211,
-0.0002400467,
-0.0288116988,
0.0117108338,
0.0137376208,
0.0610381439,
-0.0296204463,
-0.0852636024,
0.0245812368,
-0.0365547054,
0.0842122063,
0.0475679226,
0.0305652265,
-0.0118193729,
-0.0065587624,
0.0081365686,
0.0039659785,
0.0056223753,
-0.0215147845,
0.0995500088,
-0.0042945314,
-0.0326458476,
-0.0067509268,
0.041552268,
-0.0070061572,
-0.0227198228,
-0.0103662023,
0.0650381744,
0.0091215624,
0.0286245663,
-0.0135233877,
-0.0289448611,
0.0517724119,
0.0936900973,
0.0257504806,
-0.05969451,
-0.0146372095,
-0.0158877503,
-0.0233324692,
-0.0942751095,
-0.019657854,
0.0697406605,
-0.0327038765,
-0.0994082838,
-0.0562125258,
-0.0349289589,
0.0808095559,
0.0447097607,
0.0055353674,
0.0380325876,
0.003666956,
0.0615818799,
0.0236456804,
0.0258239377,
-0.0353042185,
0.003687273,
-0.0152118653,
-0.0447737873,
-0.0145495711,
-0.0621087551,
-0.0124794506,
0.0061091133,
-0.0998183265,
-0.036120113,
0.0419675224,
0.0635607615,
0.0179391671,
0.0575918779,
-0.0091448389,
-0.0912166163,
0.0300471857,
0.0819598809,
-0.0062768194,
-0.0161500368,
0.0069608577,
0.0045457883,
-0.0268366132,
0.0418421924,
-0.0459140316,
0.0407154188,
0.0593743175,
0.0097879488,
-0.0432670936,
0.0054280586,
0.0040814411,
0.0368006825,
0.0560470596,
-0.0551050864,
-0.032477174,
-0.0461448021,
-0.0103068976,
0.0053246086,
-0.0255078059,
-0.0576065518,
-0.0006163971,
-0.0063186307,
-0.0526451282,
-0.039468959,
-0.0081996694,
-0.011038241,
0.0457758605,
-0.0473172925,
0.019798357,
0.0657859817,
0.0132347317,
0.002284915,
0.0180528425,
-0.0041199732,
0.026956724,
-0.0045484202,
0.042013377,
0.0138576953,
0.0101219118,
0.0096987337,
-0.0113412235,
0.0031061156,
-0.0215143897,
-0.053636834,
0.0081849275,
-0.0222224798,
0.0590092614,
0.0232455656,
-0.0475511439,
0.0466040038,
0.0169641171,
-0.0241553579,
-0.0181784444,
0.0326417573,
-0.0030072175,
-0.0125338174,
-0.0085257431,
0.0122019826,
-0.0239870772,
0.0022207168,
0.0954390019,
-0.0101049282,
0.0327034183,
-0.0103677735,
-0.0158929769,
-0.0460671335,
0.0422646515,
0.0072177849,
0.0263653025,
0.0069899759,
0.0238040388,
-0.0790914074,
-0.0244914331,
-0.006815094,
-0.0073562763,
0.0030074497,
0.0355852768,
0.0688588396,
0.0199700389,
-0.0320943929,
0.0051877405,
-0.0542841516,
-0.011519134,
0.0007835569,
0.0103483079,
-0.0106676277,
-5.985638908e-33,
-0.0153634092,
-0.0177669693,
-0.034968704,
-0.0304282513,
0.0308926702,
-0.006838243,
-0.0024698728,
0.0050856019,
-0.0263129417,
0.0385938995,
-0.0187897962,
-0.0119729089,
0.0242625587,
0.0365551338,
0.0551057793,
-0.019217873,
0.0431354716,
-0.0256626587,
-0.0087259365,
-0.0000846735,
0.0389793366,
0.0273689907,
0.0425810628,
-0.0278871227,
0.0324312635,
0.0151291629,
0.0156204021,
0.0278066657,
0.0394053124,
0.047862649,
0.0001871717,
0.050754223,
0.0199200977,
-0.0285914484,
-0.0300797038,
-0.0241507068,
-0.0684392154,
-0.0606749505,
-0.027459437,
0.0500863232,
-0.0087655066,
-0.0115569215,
0.0229540616,
-0.0219764281,
0.0116363075,
-0.021674905,
0.0656932518,
-0.027178647,
-0.0483791269,
-0.0118487943,
-0.042536065,
0.0569785573,
-0.0115844505,
-0.0208601244,
0.0182809066,
0.0235456061,
-0.0191937182,
0.0339849256,
-0.0373708643,
0.0143011035,
0.0019968774,
-0.0215658564,
0.0167023875,
0.0161975194,
0.0173621736,
0.104061313,
0.039598532,
0.0074639502,
-0.0101179555,
-0.0282959752,
-0.0183049403,
0.0329778045,
0.0171844717,
0.1148799434,
0.0575719699,
0.0137850782,
-0.0622983314,
0.0197129268,
0.0233923718,
-0.0093868216,
0.0032134058,
0.0103201866,
-0.0624010898,
-0.0257332791,
0.000788279,
-0.0617810935,
-0.0229767803,
0.0012703693,
0.0011504,
0.0105228294,
-0.071440205,
0.0350733846,
0.0344767161,
-0.0296979751,
-0.0380483605,
-0.0063745081,
0.0137950787,
0.0118148653,
0.0233726781,
-0.0289731734,
-0.0539827794,
-0.0101071801,
0.0157727934,
0.0194285978,
0.0247733407,
0.0099235186,
0.0936535373,
0.0199153982,
-0.043771971,
-0.0093617169,
-0.0101431496,
0.0073297992,
0.0037283497,
0.0234619454,
0.0321463458,
0.0489149243,
-0.0095065711,
0.0544882491,
0.0218588263,
0.001351926,
0.0342018344,
-0.0135376705,
-0.034619391,
-0.013930941,
-0.0079279188,
-0.0006434156,
0.0140177105,
-0.0265789255,
0.0348903127,
-0.0278118029,
0.003157096,
-0.0088302484,
2.517e-7,
0.0545415916,
0.0535523184,
-0.0143834287,
-0.0411128476,
0.0166065246,
0.0326396748,
-0.0073255827,
0.0291797314,
-0.0080516152,
-0.0137863969,
0.0020707995,
0.0077515105,
0.0522669479,
0.0538473502,
-0.1149251387,
0.0066640219,
-0.0253184661,
0.0390353911,
-0.0284344479,
-0.0229771975,
0.0404167101,
0.0861815587,
0.0484271608,
-0.0189698488,
-0.0145602515,
-0.0059599141,
0.0310577322,
-0.0267811809,
-0.0064700721,
-0.0043596863,
0.046776861,
0.0275031831,
-0.0041922415,
0.0379147083,
-0.0524223335,
-0.0191001985,
0.0165463388,
0.0171484146,
-0.0351809412,
-0.0048163803,
-0.0034434015,
-0.0016857345,
0.0066599175,
-0.0034586673,
0.0684497878,
-0.07212805,
-0.0350101888,
-0.0443593822,
-0.0077920221,
0.0005133025,
0.0379628092,
-0.0068034767,
-0.0030625712,
-0.0091144051,
0.0373675749,
-0.0076708319,
-0.0417734869,
-0.0385896675,
-0.0051946575,
-0.0242928434,
-0.0332501978,
-0.0396861471,
-0.005166905,
-0.0286839548,
-0.0056640701,
-0.0631315708,
0.0202408396,
1.615881715e-34,
0.0180697907,
0.0188919939,
-0.0008673094,
0.0081004724,
-0.0287963673,
0.0047638654,
0.0109626753,
-0.0142277731,
0.0232884772,
-0.011760721,
-0.0121920714
] |
http://arxiv.org/abs/2401.13227v3 | http://arxiv.org/pdf/2401.13227v3 | LPNL: Scalable Link Prediction with Large Language Models | 2024-01-24T04:50:16 | 2024-02-20T03:53:06 | Baolong Bi, Shenghua Liu, Yiwei Wang, Lingrui Mei, Xueqi Cheng | cs.CL | cs.CL, cs.AI, cs.LG, cs.SI | Exploring the application of large language models (LLMs) to graph learning
is a emerging endeavor. However, the vast amount of information inherent in
large graphs poses significant challenges to this process. This work focuses on
the link prediction task and introduces $\textbf{LPNL}$ (Link Prediction via
Natural Language), a framework based on large language models designed for
scalable link prediction on large-scale heterogeneous graphs. We design novel
prompts for link prediction that articulate graph details in natural language.
We propose a two-stage sampling pipeline to extract crucial information from
the graphs, and a divide-and-conquer strategy to control the input tokens
within predefined limits, addressing the challenge of overwhelming information.
We fine-tune a T5 model based on our self-supervised learning designed for link
prediction. Extensive experimental results demonstrate that LPNL outperforms
multiple advanced baselines in link prediction tasks on large-scale graphs. | LPNL: Scalable Link Prediction with Large Language Models: Exploring the application of large language models (LLMs) to graph learning
is a emerging endeavor. However, the vast amount of information inherent in
large graphs poses significant challenges to this process. This work focuses on
the link prediction task and introduces $\textbf{LPNL}$ (Link Prediction via
Natural Language), a framework based on large language models designed for
scalable link prediction on large-scale heterogeneous graphs. We design novel
prompts for link prediction that articulate graph details in natural language.
We propose a two-stage sampling pipeline to extract crucial information from
the graphs, and a divide-and-conquer strategy to control the input tokens
within predefined limits, addressing the challenge of overwhelming information.
We fine-tune a T5 model based on our self-supervised learning designed for link
prediction. Extensive experimental results demonstrate that LPNL outperforms
multiple advanced baselines in link prediction tasks on large-scale graphs. | [
0.0259436239,
0.0040019858,
-0.0252743773,
0.0407232679,
-0.0017983143,
-0.0363957025,
-0.0015992571,
0.0204080492,
0.1111094728,
-0.0416018851,
-0.0485004894,
0.0017678199,
-0.0423272178,
0.0330800489,
0.0468258038,
-0.029521469,
-0.0010203099,
-0.0409141146,
-0.0005250385,
-0.0220961068,
-0.0170499533,
0.0634446815,
0.0680247471,
0.0120830294,
-0.0105780847,
0.0052354238,
0.0003098223,
0.0350361504,
-0.0236523654,
-0.0527795777,
-0.0242802668,
0.0360224955,
0.0149354357,
-0.0088077784,
0.0000016834,
-0.0526178665,
-0.006855329,
-0.0234458335,
-0.0262987725,
0.0050629941,
0.0184999816,
-0.0046425126,
0.0525924936,
0.0196881574,
-0.0212078393,
-0.031658411,
0.011344742,
0.0474436916,
0.0274330303,
0.0591370352,
-0.0017548444,
-0.021804003,
0.0020745839,
-0.0244267751,
0.003045412,
-0.0334769785,
0.0537915938,
-0.0026204328,
0.0164271425,
0.0323814079,
-0.0298845246,
0.0081860311,
0.0034839453,
-0.0142988479,
0.0512845032,
0.0497948527,
-0.0013316993,
0.0035478692,
-0.0102961846,
0.1294950843,
-0.0226629786,
-0.0337039568,
0.0212267786,
-0.0189391524,
-0.0005136765,
0.0411329791,
-0.036231067,
-0.0385723375,
0.0071639461,
0.0374061204,
-0.019608492,
-0.0217141621,
0.0711828098,
-0.0171326287,
0.0096039502,
0.0431849398,
-0.0158358216,
0.0076066139,
0.054335583,
0.0202307962,
-0.0016773056,
-0.0244286638,
0.0206044354,
0.0200645477,
-0.0460122898,
0.0009611213,
-0.0000290757,
0.0164875165,
-0.0009247851,
-0.0163116008,
0.0119916238,
-0.0525592417,
0.0219669845,
0.0463801697,
-0.0487722978,
-0.0255913176,
-0.0486634709,
-0.0271413382,
-0.020434415,
-0.0412209108,
-0.0305024087,
0.014553153,
-0.0350713506,
0.0309058204,
-0.018500274,
-0.0334901176,
-0.048369877,
0.0061291135,
-0.0407383554,
0.0356150158,
0.0363502614,
0.0188725367,
-0.0042281714,
-0.0286029316,
-0.02587598,
0.0357103981,
-0.040915709,
-0.010227493,
-0.0456658565,
-0.0327482037,
-0.0149164172,
0.0079674544,
0.034131173,
-0.0573836528,
0.0205303896,
0.1112374589,
0.0319921672,
-0.0092573268,
-0.0155044729,
-0.0596598051,
-0.0118698198,
-0.0569840074,
0.0253573973,
-0.067351535,
0.0245340616,
0.0451226607,
0.002100392,
-0.0187299643,
0.029766839,
0.0063769897,
-0.071126394,
0.0521992855,
0.0085635129,
-0.0137113137,
-0.0206706543,
-0.0393571071,
0.0108734109,
0.0308704823,
-0.0124316355,
0.0637512505,
0.0314806998,
0.0408188216,
-0.0137380408,
0.0602348708,
0.0552186295,
-0.0093157152,
0.0137791252,
0.0017449231,
0.0107788416,
0.0059860325,
0.0038873106,
0.0893281028,
-0.0085996883,
-0.0426201001,
0.0289760213,
0.0489758365,
0.0265837424,
0.0148542626,
0.07318528,
0.0236207359,
-0.0250937734,
-0.0545419231,
-0.0581273958,
0.0327976048,
-0.0087550152,
0.0000297898,
0.0314201936,
0.0227298718,
-0.022584457,
0.0083826035,
-0.0066374126,
-0.0169287808,
-0.0109375352,
0.0106836651,
0.0627007484,
0.0113966176,
-0.0313546173,
0.0532210954,
0.0324311443,
0.0251781251,
0.0461077876,
0.0439416282,
0.0011532397,
0.0714557618,
0.0265944786,
-0.0213824622,
0.0181286037,
-0.0312650427,
0.0074250214,
-0.0392388478,
0.0280853529,
-0.0422458425,
-0.0247074179,
-0.0312145874,
-0.0303341914,
0.02371707,
-0.0284068417,
0.0105377836,
0.0325103365,
0.0182738509,
0.0168189239,
-0.0759704709,
-0.0144343013,
-0.0050555901,
0.0001819671,
0.0033354852,
0.0388090611,
-0.0063460288,
-0.0054430775,
0.035297513,
0.0027468195,
-0.0153828934,
0.0417209826,
-0.010567693,
-0.0359844863,
-0.0285091065,
0.0107019357,
-0.0221676454,
0.0557584837,
-0.0357440263,
-0.0023561015,
-0.0190275107,
-0.0292886905,
-0.0252693165,
0.0733429715,
-0.03858402,
0.0141822817,
0.0249407403,
-0.0207684822,
0.0280382968,
0.0030365051,
-0.005285867,
0.0385288596,
0.008526912,
0.0072063175,
0.0036262253,
0.0070484588,
-0.019937342,
-0.0577151366,
-0.0315063223,
0.0042220308,
0.036596179,
-0.0150179127,
0.0526596978,
0.0080336966,
-0.0483987294,
0.0033351679,
-0.0639217645,
-0.0050474615,
-0.0059774062,
-0.0111454595,
-0.0333622918,
0.0072945352,
0.0177284256,
-0.0629202724,
-0.015745217,
0.0266815275,
-0.004731549,
0.0014559147,
0.0368422158,
0.0117593044,
-0.0064882264,
-0.0412229933,
-0.0639412701,
-0.0307138376,
0.0927537233,
0.0277705453,
-0.0665970743,
-0.0158331301,
-0.007960679,
-0.059965156,
0.0193288084,
-0.0226050559,
0.0398421623,
-0.0511646792,
-0.0322343558,
0.047827851,
-0.0000485763,
-0.026168935,
0.0146596404,
0.0528235026,
0.0288112946,
-0.0382696763,
0.0051148287,
-0.0624186695,
0.0019399998,
-0.0510702282,
0.0155190025,
-0.0032818483,
-0.0156364106,
0.0171513539,
-0.0102523761,
-0.0356571935,
0.005818048,
-0.0037697474,
0.0297323409,
-0.0155684697,
-0.0102753025,
0.008992617,
-0.0175784882,
-0.0042495485,
-0.0474526845,
0.0573165603,
0.0167022124,
0.017533863,
0.0420316868,
-0.0329228342,
-0.0528377816,
0.0879852176,
0.0494032279,
0.0358240604,
-0.0015412081,
-0.0305586513,
-0.0401757322,
0.0319588073,
0.0012354681,
-0.0213161893,
0.0384664424,
0.0030451061,
0.0512000136,
0.0558323301,
-0.0382474549,
-0.0139798252,
-0.0122296307,
-0.0753869042,
0.0047506033,
0.0076737567,
-0.0366720557,
0.033149831,
-0.0154041657,
-0.0257820357,
-0.0101494063,
-0.0078183394,
-0.0164294038,
-0.0045025903,
0.0432661809,
0.0520735532,
-0.0719311759,
0.0176277831,
-0.0172268543,
0.0079203723,
0.0241890103,
0.0240671858,
-0.0242203306,
-0.015265367,
0.0180752315,
0.0201957095,
-0.0045742858,
-0.0114962682,
0.0254344102,
0.0523236878,
0.0083548175,
0.0021400228,
0.110258244,
0.0339825191,
-0.0502770171,
-0.0030412958,
0.0258611422,
0.0354908668,
0.0302735046,
0.0272328537,
-0.0802641213,
-0.0316951647,
-0.0319211446,
-0.0098593207,
0.0508217141,
-0.084289901,
0.0105856769,
-0.1349232942,
0.0164903738,
0.0208540782,
-0.0000362287,
-0.0429753624,
-0.0936546177,
-0.0046012183,
-0.0374587066,
-0.0274239928,
0.0052690543,
-0.0507886894,
0.0164692067,
0.0284712706,
-0.0371003561,
-0.0060636955,
0.0528232045,
-0.068680048,
0.0482305326,
-0.0315492041,
0.0441567637,
-0.002863572,
0.0229195822,
0.0553254783,
-0.0136877783,
-0.0148448506,
-0.0416744612,
0.0833513588,
-0.010707329,
0.0119584929,
-0.0250911452,
0.032940574,
-0.0219441522,
-0.0193649661,
-0.0226581227,
0.1094892472,
0.0019748223,
-0.0761975348,
0.0007883023,
0.0066110105,
-0.0205587372,
0.0427271761,
0.0105175311,
-0.0185625665,
-0.0198433753,
-0.0411735177,
0.0124833724,
0.0071552885,
-0.0361644104,
-0.0004546278,
-0.0818504691,
0.0114033921,
-0.0406918153,
0.0047746277,
0.0494591072,
0.0197367985,
0.0249390658,
0.0332764201,
0.026117906,
0.0010869117,
-0.0501929373,
0.0176510084,
-0.0475652367,
0.0319182016,
-0.0004480197,
0.0092948759,
0.0301179625,
-0.0226682089,
-0.0191158801,
0.0632193312,
-0.0837690905,
-0.08318609,
-0.0594917014,
0.0193977803,
0.0181934889,
0.0669607893,
-0.0177513454,
-0.026888838,
0.0469469428,
0.0520149916,
-0.0150345685,
-0.0119137466,
-0.0223795958,
0.0093493648,
0.0116364686,
0.0615752488,
-0.0739389509,
-0.0145433331,
0.041732315,
0.0082198493,
-0.0372189991,
-0.0647314638,
-0.0031554899,
0.0549598075,
0.0342220627,
0.0059854882,
0.0337708667,
-0.0465137027,
-0.0409416072,
-0.0142433252,
-0.0382297039,
0.0003081513,
0.0001697524,
0.0268773939,
0.0122255227,
-0.0127631044,
-0.0697139949,
0.0052340096,
0.0231959671,
-0.0439463705,
-0.0254167244,
-0.0093322825,
-0.0161559489,
0.0744609311,
0.0120974155,
0.0157500785,
-0.0268994495,
0.0505838208,
0.0279620215,
0.0025849652,
-0.0088078715,
0.0211762879,
-0.0097412355,
-0.0926623568,
-0.0254434943,
-0.031615898,
-0.0328593068,
-0.0110934004,
0.0059444755,
0.0540036447,
-0.0145669598,
0.0239887219,
-0.0423696488,
-0.0020333331,
-0.0180702694,
0.0331682079,
-0.0247022454,
-0.0499971956,
0.0604787096,
-0.0063009025,
-0.023971973,
0.0519560128,
0.0552201569,
0.0070523466,
0.0407022946,
0.0222339723,
-0.0503632426,
-0.0186085738,
0.0042148451,
-0.0465511307,
0.0213612244,
-0.0031942725,
0.0059385421,
-0.0511651672,
-0.097179465,
0.0525628254,
-0.0091623012,
0.0070889341,
0.0130490288,
0.0497231185,
0.0236613154,
0.0363280922,
-0.013531643,
-0.0400542021,
0.0471914969,
0.0129717942,
-0.0438798293,
-0.0186002087,
-6.076576253e-33,
0.0094431499,
-0.0749468431,
0.0397238731,
0.0309857726,
-0.0467567146,
0.0305399001,
-0.0301852524,
-0.0141201895,
-0.0337488949,
0.025294045,
-0.0099973716,
-0.013724925,
-0.0021324831,
0.0258157421,
-0.0044469885,
-0.0468028598,
0.0349329486,
-0.0066338773,
0.0325186253,
0.0330275521,
0.0334170945,
0.025916785,
0.0268054325,
-0.0372403488,
0.028272083,
-0.0219749808,
0.0115419244,
0.033991795,
0.0533431247,
0.0287675653,
0.0055141803,
0.0630866438,
0.0134189827,
-0.0207733717,
0.0105202636,
0.0241889097,
-0.0750461221,
-0.0333735123,
0.0038795213,
-0.015870858,
0.0047508464,
0.0013687547,
0.0214455947,
0.0114927972,
0.0011874565,
-0.0339311324,
0.0408339985,
0.0453833379,
-0.0564458817,
-0.0093672229,
0.0128492601,
0.0191564709,
-0.0426377729,
0.0382105894,
0.0163997635,
0.0064433841,
-0.0033750178,
0.0169465113,
-0.0494385995,
0.0178401619,
-0.0235867538,
-0.0053248233,
0.0077432524,
0.0280263536,
0.0362209715,
0.0762720406,
0.0320399702,
-0.0054283435,
-0.0243019052,
0.0342889987,
-0.0274477266,
0.0392934382,
-0.0185247436,
0.0095349429,
-0.0521223955,
-0.0149100302,
-0.0995150208,
-0.0047243275,
0.0500374138,
0.0332423858,
-0.0111143934,
0.0180238225,
-0.0004747755,
-0.0433229208,
-0.0477997847,
-0.0509662479,
0.0079908231,
-0.0272486135,
-0.032849662,
0.0136596225,
-0.0003036289,
0.0401287153,
-0.0025495624,
-0.0494199954,
-0.069985725,
0.0597534068,
-0.013998488,
0.0094639678,
0.0085361525,
0.0108106621,
-0.0709631369,
0.0107443724,
0.00250479,
-0.0604055896,
-0.0063311365,
0.0008730766,
0.0175800323,
-0.0128078004,
-0.0721736103,
-0.0215008222,
-0.0352160633,
-0.0226598904,
-0.0187175218,
-0.0326626226,
0.053256195,
0.0502772629,
-0.0124931345,
0.0394023806,
0.0647886842,
0.005294391,
0.0215814319,
0.0490776598,
0.0138843935,
-0.0334440246,
-0.0445579141,
-0.0030116334,
-0.0405795723,
0.0295424238,
0.0383869708,
-0.0879886225,
-0.0288345199,
-0.0162395909,
2.496e-7,
0.0441664308,
0.0610524416,
0.0229464676,
0.0383358784,
0.0226591341,
0.0039425902,
0.0315322839,
0.0706916004,
0.0334664509,
0.0067887362,
0.0078158947,
0.0186394919,
-0.0441565067,
0.0373911262,
-0.1001639366,
0.0588524826,
-0.0526323989,
-0.0368808918,
-0.0468904935,
-0.0091708619,
-0.0203905012,
0.0689263791,
0.0630454347,
0.0095639853,
-0.0034765634,
-0.0392674766,
0.0549616292,
-0.0489398241,
-0.0323281661,
-0.0112742819,
0.0036282458,
-0.0213914104,
-0.0318538882,
0.0407039933,
-0.0196637698,
0.0258615743,
0.0232091211,
0.0207518507,
0.0157087389,
0.012245697,
0.0172285549,
-0.0375700183,
0.0019984413,
0.0006998103,
0.0612001866,
-0.0177097544,
-0.0248634461,
0.0075639128,
-0.0234270897,
-0.0412347578,
0.0101773217,
0.0016380688,
-0.0262224581,
0.0491959117,
0.0058961702,
0.0202170145,
-0.0043663792,
-0.0544384792,
0.0029538362,
0.0327489749,
-0.0115740579,
-0.0669962466,
0.020592127,
0.0041029919,
0.0106368056,
-0.0315718912,
0.0386463031,
1.908571658e-34,
0.0108219683,
-0.0071636466,
-0.0125472043,
0.0145133995,
0.0528455265,
-0.0481693782,
-0.0403526351,
0.0274367388,
0.0685289055,
-0.0890857726,
-0.0211270321
] |
http://arxiv.org/abs/2402.01696v2 | http://arxiv.org/pdf/2402.01696v2 | HiGen: Hierarchy-Aware Sequence Generation for Hierarchical Text Classification | 2024-01-24T04:44:42 | 2024-02-22T03:16:46 | Vidit Jain, Mukund Rungta, Yuchen Zhuang, Yue Yu, Zeyu Wang, Mu Gao, Jeffrey Skolnick, Chao Zhang | cs.CL | cs.CL, cs.LG | Hierarchical text classification (HTC) is a complex subtask under multi-label
text classification, characterized by a hierarchical label taxonomy and data
imbalance. The best-performing models aim to learn a static representation by
combining document and hierarchical label information. However, the relevance
of document sections can vary based on the hierarchy level, necessitating a
dynamic document representation. To address this, we propose HiGen, a
text-generation-based framework utilizing language models to encode dynamic
text representations. We introduce a level-guided loss function to capture the
relationship between text and label name semantics. Our approach incorporates a
task-specific pretraining strategy, adapting the language model to in-domain
knowledge and significantly enhancing performance for classes with limited
examples. Furthermore, we present a new and valuable dataset called ENZYME,
designed for HTC, which comprises articles from PubMed with the goal of
predicting Enzyme Commission (EC) numbers. Through extensive experiments on the
ENZYME dataset and the widely recognized WOS and NYT datasets, our methodology
demonstrates superior performance, surpassing existing approaches while
efficiently handling data and mitigating class imbalance. The data and code
will be released publicly. | HiGen: Hierarchy-Aware Sequence Generation for Hierarchical Text Classification: Hierarchical text classification (HTC) is a complex subtask under multi-label
text classification, characterized by a hierarchical label taxonomy and data
imbalance. The best-performing models aim to learn a static representation by
combining document and hierarchical label information. However, the relevance
of document sections can vary based on the hierarchy level, necessitating a
dynamic document representation. To address this, we propose HiGen, a
text-generation-based framework utilizing language models to encode dynamic
text representations. We introduce a level-guided loss function to capture the
relationship between text and label name semantics. Our approach incorporates a
task-specific pretraining strategy, adapting the language model to in-domain
knowledge and significantly enhancing performance for classes with limited
examples. Furthermore, we present a new and valuable dataset called ENZYME,
designed for HTC, which comprises articles from PubMed with the goal of
predicting Enzyme Commission (EC) numbers. Through extensive experiments on the
ENZYME dataset and the widely recognized WOS and NYT datasets, our methodology
demonstrates superior performance, surpassing existing approaches while
efficiently handling data and mitigating class imbalance. The data and code
will be released publicly. | [
0.0331642181,
0.072014764,
0.0160075799,
-0.0428161323,
-0.0131383305,
0.0093394518,
0.0009729902,
0.030558385,
0.0152211832,
-0.0988561213,
-0.0120779816,
-0.022930054,
0.0084307585,
0.0372859798,
0.018785987,
-0.0476431362,
0.0199069995,
0.0127299391,
0.0124218166,
-0.0007853299,
-0.0021856807,
0.0278899409,
0.0381524563,
0.0436953865,
0.0061152228,
-0.0221363939,
-0.0082882484,
0.0168871265,
0.0293595232,
-0.0503898896,
0.0732133314,
0.0506626926,
0.0595616773,
0.0309705734,
0.0000019671,
-0.0260907859,
-0.0272834692,
0.0031506524,
-0.0260259882,
-0.003683873,
0.0520383939,
-0.0439219996,
0.0250459388,
0.0201177076,
0.0165665075,
-0.0053331526,
0.0036561806,
0.0224297252,
0.0368463323,
0.0625878572,
-0.0071217851,
-0.0507798381,
0.0510116704,
-0.0364826098,
0.0442951098,
-0.0269413721,
0.0376136564,
-0.0008027843,
-0.0004809554,
0.0486550592,
0.0146213314,
0.0049998155,
0.0223111678,
-0.0079846373,
0.0621689409,
0.0382666737,
0.0130217625,
-0.0322752297,
0.0200320166,
0.058786232,
-0.0558154918,
-0.0425370596,
0.0121919224,
-0.0101208594,
0.0106501058,
-0.0387269296,
-0.0099899834,
0.0517726988,
-0.0387948938,
0.0210456792,
0.0474804491,
0.0327475443,
0.011029576,
-0.0201734342,
0.016498914,
0.0166570377,
0.0075035309,
-0.021721879,
-0.0032838278,
-0.0073766978,
0.0137828877,
-0.0536082946,
0.0244948361,
0.0306061208,
-0.02544659,
0.0080475323,
-0.0559245124,
-0.0491896942,
0.0705553368,
-0.1132650748,
-0.0125525631,
-0.0074375574,
0.0293813143,
-0.0068537197,
0.0004454934,
-0.014836981,
0.0191852394,
-0.0013142193,
-0.0193641987,
0.008936041,
-0.0606297255,
-0.0199847613,
-0.024061447,
0.0472679399,
0.0048084687,
0.0053409664,
-0.0721456781,
-0.0022847638,
0.0109207844,
-0.0051921927,
-0.0138819488,
0.0341542922,
-0.0638888553,
0.012070775,
0.0234609172,
0.0123943072,
0.0116804782,
0.0101692108,
0.0133390073,
0.0084633529,
0.0011204648,
0.0224122927,
0.004301805,
-0.0030383107,
0.0140995346,
0.0608921722,
-0.0105833523,
-0.0451823622,
-0.0640200973,
-0.0570314862,
-0.0230226181,
-0.0435366668,
-0.0112041673,
-0.0318748988,
0.0291913487,
-0.0160738472,
-0.0133028887,
-0.0214598924,
0.0267187357,
0.0216877516,
-0.0317561664,
0.0261631347,
-0.0162197817,
-0.0610658079,
0.0558734201,
0.0362460874,
0.0104302056,
0.0399954692,
0.0221445803,
0.0466360338,
0.0155858872,
-0.0120533286,
-0.0012852887,
0.0077415924,
0.0287004337,
-0.0339065231,
-0.0016248493,
0.0082623353,
-0.0169514976,
0.0574881844,
0.0343886577,
-0.0022392112,
0.0117856944,
-0.0011031001,
0.0474918336,
0.0333131552,
0.0017786872,
-0.0062840632,
0.0476818979,
0.0288601816,
-0.0375884324,
-0.0142020797,
0.0470443331,
-0.011569906,
-0.023615595,
0.0217576195,
0.005234221,
-0.0048785061,
-0.0213982854,
-0.0404173136,
-0.0145920599,
-0.002301814,
-0.0213154107,
-0.0408026502,
0.0476094931,
0.0355539508,
-0.0473035946,
0.0439609811,
0.0482837446,
-0.0521183759,
0.0445827618,
-0.0247534905,
0.0230952501,
0.0460255221,
0.0379493237,
-0.0327088982,
-0.0186440181,
-0.0084679564,
0.0110205635,
0.0346453264,
-0.0282106958,
-0.0483145118,
-0.0057069059,
-0.0331401862,
-0.016358871,
-0.0100881001,
0.0210464839,
-0.000424058,
-0.0614635088,
0.0066279736,
0.0413309038,
-0.0468197949,
-0.0448009036,
0.0428179726,
-0.0317987166,
-0.0379823335,
0.0411463194,
-0.0128077511,
0.0328780375,
0.0008851565,
-0.0230010953,
0.0112164905,
0.0373723097,
-0.0329010412,
-0.0591845028,
-0.083598502,
-0.0035233905,
-0.0778319538,
-0.0040506031,
-0.0100082736,
-0.00412563,
0.0034320683,
-0.0685812831,
-0.0169701278,
-0.0084737157,
-0.047477562,
0.0160168484,
-0.0185751356,
0.0049874936,
0.0090258243,
0.0184405632,
-0.0437705889,
0.0829143971,
-0.0042016329,
0.0099146869,
-0.0015637199,
-0.0372254252,
0.0052258051,
-0.0053449348,
0.0389089845,
0.0120833116,
0.0016345576,
0.0066467491,
0.0486601815,
0.0300562792,
-0.0310214441,
0.0073758452,
-0.0086475136,
-0.0045016007,
-0.0013090059,
0.0238125697,
-0.0161383953,
-0.0130015071,
0.0190150253,
-0.0327705927,
-0.0612391792,
0.0031909498,
-0.0395491309,
0.0045316969,
0.0328605063,
-0.011727172,
0.0009927553,
-0.010383673,
-0.0106955897,
-0.0242134724,
0.0825872272,
0.0059087318,
0.0019202749,
-0.0159028694,
-0.0068034967,
-0.0090385228,
0.005463358,
0.0028627948,
0.0190215725,
-0.0147354137,
0.013315226,
0.0176332649,
0.0316354707,
0.0456905402,
0.0555126593,
0.0323981568,
0.042633269,
-0.0202521663,
-0.0181699879,
-0.0570612773,
0.0325504504,
-0.0081429379,
-0.0500562266,
0.025831487,
0.1098395437,
0.0117382947,
0.0034666627,
-0.0385650732,
-0.0280574895,
0.0152207734,
0.0080470936,
-0.0600939207,
-0.0126413526,
0.0008083166,
0.0082801161,
0.0087717902,
-0.0425383821,
0.0584291741,
-0.0245974325,
0.0159307998,
-0.0617726743,
0.0034666993,
-0.0239693169,
0.0354602747,
0.0405666269,
0.0324374922,
0.0588759631,
-0.0246320721,
-0.0171218961,
0.0476227663,
0.0048102266,
-0.034986902,
-0.0123925665,
-0.0436354578,
-0.0061131464,
-0.0239586383,
-0.0093373153,
-0.0148959868,
0.0085476357,
0.0258667748,
0.0405685268,
0.0214024261,
0.001405786,
-0.0739116222,
0.0141622908,
-0.0026357754,
0.0746803805,
-0.0079548806,
-0.0243322831,
-0.0104218526,
0.0607872047,
0.061057359,
-0.0227345321,
-0.0254191551,
-0.0289827995,
0.0240294728,
0.0607366897,
-0.001716812,
0.0033656047,
0.0218609907,
-0.0051854062,
0.0225108415,
-0.0425221138,
0.0033026491,
0.0138114775,
-0.034126386,
-0.0565422773,
-0.0234519113,
0.0924221128,
0.0517827421,
-0.0865495801,
0.0437944345,
0.0150417211,
0.053995423,
-0.0038857413,
0.0028324143,
-0.0697810277,
-0.0285108816,
0.0042513181,
0.0391173773,
-0.0255288705,
-0.0262812059,
0.0191088784,
-0.0497433469,
0.0200256612,
0.0133181754,
-0.033290226,
-0.0105742449,
-0.0578399636,
-0.0585976653,
0.0016342929,
-0.0065218266,
0.0197553821,
-0.0095067592,
0.0067978408,
0.0466613322,
-0.0978448913,
-0.0712772384,
0.0343681164,
0.013798641,
-0.000075768,
0.0552881174,
0.034339577,
0.017112026,
0.0024318516,
0.0208031535,
-0.0246081427,
-0.0181290954,
-0.0344976895,
0.0531694815,
-0.0371688493,
0.0290677864,
-0.0339228846,
0.0601119734,
-0.0604297221,
-0.0169953816,
0.0204002131,
0.0680462047,
-0.0105917715,
-0.0378549471,
-0.0701771602,
-0.0150869461,
0.011810624,
0.0441820621,
0.0178190768,
-0.0607176609,
-0.016010087,
-0.0183807462,
0.0363287553,
-0.0126957996,
0.0109571703,
0.0514692999,
-0.0675463974,
-0.0439308323,
-0.0859085619,
-0.0478638113,
0.0050085071,
0.0016863419,
-0.0060794251,
-0.0063393372,
-0.0049659102,
-0.043980442,
-0.0375466198,
0.0107204486,
0.0322519056,
-0.0147966696,
-0.0295710135,
-0.0504605062,
0.0569463372,
-0.0166749358,
-0.0567781031,
0.0174151938,
-0.0312453415,
-0.0554787405,
-0.051313933,
0.0454468131,
0.0080094542,
0.0386010967,
0.0406290144,
-0.0443410501,
0.0798344687,
0.0380974747,
-0.0248554219,
-0.0333250389,
0.0265301727,
-0.0095991725,
-0.0310593788,
0.0359673165,
-0.0592580214,
0.0393673033,
-0.0034967996,
0.0527540185,
0.0239359997,
-0.0015654968,
0.041154366,
-0.0499058701,
0.0287400074,
-0.0599631406,
-0.0373533256,
-0.0324929617,
0.0537380017,
-0.0121033555,
-0.0273615737,
-0.0328508914,
0.0303272828,
0.0339329876,
0.0031319084,
-0.0194790661,
0.0160794482,
-0.0250651408,
0.034634009,
-0.0199053958,
0.010077049,
-0.0280064717,
0.0546733662,
0.0297756996,
-0.0173480082,
-0.0310969595,
-0.0308554675,
-0.0126130078,
0.006764336,
-0.0333978944,
0.0298629161,
0.0135727832,
0.0474934392,
-0.0301499125,
-0.0001546515,
-0.0432553738,
0.0770785138,
-0.0765302181,
-0.0309994761,
0.0427087098,
-0.0007375835,
0.0147639168,
-0.0370640457,
-0.0198482741,
-0.0113581792,
0.0759120733,
-0.017042907,
-0.0477749109,
0.0089862412,
-0.0124341641,
-0.0557495244,
-0.0284432229,
0.1124536097,
0.1133055389,
0.0027965975,
0.0010759198,
-0.0396365263,
-0.0252621323,
-0.0011870119,
0.0412980728,
0.0425416157,
-0.0086947894,
-0.0209746175,
-0.0484364592,
-0.0356501974,
0.0323312394,
-0.0133463107,
0.0144232213,
0.0291659702,
0.040883027,
0.0298236366,
-0.0095159952,
-0.0527047105,
-0.0464271121,
0.0284400918,
-0.0176962782,
-0.0401636511,
-0.0288094096,
-5.834125767e-33,
-0.0042205998,
-0.0407529883,
-0.0140524851,
0.0408150032,
-0.0059183445,
0.0376607105,
-0.0718355104,
-0.0180794299,
0.0485366732,
0.0307178516,
0.011079723,
-0.0193647146,
-0.0156715252,
0.0604114011,
0.0383089967,
0.011621573,
0.0394358709,
0.0003449995,
-0.0180408489,
0.0732620209,
0.0256014336,
0.0279541258,
0.0612698533,
-0.0324093103,
0.0054348004,
-0.0160179976,
0.0060489429,
-0.0279549938,
-0.00157086,
0.0190043394,
-0.0310458206,
0.0725245625,
0.0238600411,
-0.021871049,
0.0014649846,
0.0348969288,
-0.0527955666,
-0.0076991101,
-0.0642305762,
-0.0105723459,
-0.0339316726,
-0.041849114,
0.0289565474,
0.0252671689,
-0.0064666299,
-0.0249042064,
0.0944365039,
-0.0119029926,
-0.0658239275,
-0.1074386463,
-0.087689355,
0.0351090133,
-0.0065559465,
0.0040840995,
-0.0218574125,
0.0571984909,
0.0207246933,
0.0336866081,
-0.0381791517,
0.0013803982,
0.0143170487,
0.0251903217,
0.0233717132,
0.0220098495,
-0.0004945358,
0.0265718009,
0.0081174979,
0.0529964752,
-0.0035916474,
0.0477258228,
-0.0007505763,
-0.0113955056,
0.0352071673,
0.0005416179,
0.0613290705,
0.0041035358,
-0.1026303619,
0.0783644021,
-0.0436155982,
0.0066665616,
-0.0057833088,
0.0103881732,
0.0378600806,
-0.0103998352,
-0.0152874859,
0.0237914342,
0.0025295003,
-0.0327796265,
-0.000320743,
0.0077495738,
-0.0004231284,
0.0159465037,
0.0178732481,
0.0093597686,
-0.0611188337,
0.0134501001,
0.0134260738,
0.0032498445,
-0.002604879,
0.0058398466,
-0.042016305,
-0.0432294533,
0.0535959825,
-0.0192549136,
0.0092710638,
-0.0225543343,
0.0695255324,
0.0066014412,
-0.0172183588,
0.0074440138,
-0.0195325222,
0.002656268,
0.0222441666,
0.0823152289,
0.0110105611,
-0.0038555677,
0.0131839337,
0.0863014162,
0.0062725381,
-0.0233747158,
0.0051598684,
0.0427448042,
0.0118696624,
0.0325005911,
-0.031471055,
-0.0150357466,
-0.0264396742,
-0.000343545,
0.0199156944,
-0.085362643,
-0.0111433752,
-0.0231558643,
2.689e-7,
0.0492356196,
0.0313508399,
0.0014037872,
-0.0567951612,
0.0310432967,
0.0356012136,
0.010791651,
0.071811527,
-0.0663584396,
-0.0089399992,
0.0348677114,
-0.0644973516,
0.0563009977,
0.0173609555,
-0.0758635625,
-0.0183574501,
-0.0998369306,
-0.0086878249,
-0.0701173469,
0.0109407352,
0.045613002,
0.0760192499,
-0.0100240121,
-0.0286312308,
0.0324958637,
-0.0066619664,
0.0636828318,
0.0315476507,
0.0148186404,
0.0574704111,
-0.0359563418,
0.008314698,
-0.0137106702,
0.0210943483,
-0.0792171359,
-0.0185056832,
-0.0038024725,
0.0269274823,
-0.0478858165,
-0.0346839279,
0.0580576956,
0.0113243926,
0.0140120517,
-0.0345293544,
0.0194148142,
-0.0706916004,
-0.0491148457,
-0.0027724074,
-0.0172446519,
0.0278200898,
0.0067558344,
-0.0319262259,
-0.0167674422,
0.0327022374,
-0.0014194442,
0.0220317598,
-0.0007730967,
-0.0828385204,
0.0036920037,
0.036275357,
-0.0485107675,
0.0042219376,
-0.0163251758,
0.0314241499,
0.0092558749,
-0.0294095967,
0.0152799059,
2.905761883e-34,
0.0413692333,
-0.0454865843,
-0.0626447275,
0.0219085775,
-0.0028121136,
-0.0110161556,
0.0046677021,
0.0191080272,
0.016812155,
-0.0600653626,
-0.0178437494
] |
http://arxiv.org/abs/2401.13223v2 | http://arxiv.org/pdf/2401.13223v2 | TAT-LLM: A Specialized Language Model for Discrete Reasoning over Tabular and Textual Data | 2024-01-24T04:28:50 | 2024-02-22T13:36:56 | Fengbin Zhu, Ziyang Liu, Fuli Feng, Chao Wang, Moxin Li, Tat-Seng Chua | cs.CL | cs.CL, cs.AI | In this work, we address question answering (QA) over a hybrid of tabular and
textual data that are very common content on the Web (e.g. SEC filings), where
discrete reasoning capabilities are often required. Recently, large language
models (LLMs) like GPT-4 have demonstrated strong multi-step reasoning
capabilities. We then consider harnessing the amazing power of LLMs to solve
our task. We abstract a Step-wise Pipeline for tabular and textual QA, which
consists of three key steps, including Extractor, Reasoner and Executor, and
initially design an instruction to instantiate the pipeline and validate that
GPT-4 outperforms all existing methods. However, utilizing an online LLM like
GPT-4 holds various challenges in terms of cost, latency, and data security
risk, which motivates us to specialize smaller LLMs in this task. We develop a
TAT-LLM language model by fine-tuning LLaMA 2 with the training data generated
automatically from existing expert-annotated datasets following the Step-wise
Pipeline. The experimental results have verified that our TAT-LLM model can
outperform all baseline models, including the previous best fine-tuned models
and very large-scale LLMs like GPT-4 on FinQA, TAT-QA and TAT-DQA benchmarks. | TAT-LLM: A Specialized Language Model for Discrete Reasoning over Tabular and Textual Data: In this work, we address question answering (QA) over a hybrid of tabular and
textual data that are very common content on the Web (e.g. SEC filings), where
discrete reasoning capabilities are often required. Recently, large language
models (LLMs) like GPT-4 have demonstrated strong multi-step reasoning
capabilities. We then consider harnessing the amazing power of LLMs to solve
our task. We abstract a Step-wise Pipeline for tabular and textual QA, which
consists of three key steps, including Extractor, Reasoner and Executor, and
initially design an instruction to instantiate the pipeline and validate that
GPT-4 outperforms all existing methods. However, utilizing an online LLM like
GPT-4 holds various challenges in terms of cost, latency, and data security
risk, which motivates us to specialize smaller LLMs in this task. We develop a
TAT-LLM language model by fine-tuning LLaMA 2 with the training data generated
automatically from existing expert-annotated datasets following the Step-wise
Pipeline. The experimental results have verified that our TAT-LLM model can
outperform all baseline models, including the previous best fine-tuned models
and very large-scale LLMs like GPT-4 on FinQA, TAT-QA and TAT-DQA benchmarks. | [
0.0508748107,
0.0501911379,
0.0089107277,
0.0326524302,
-0.0123996837,
0.0371729471,
0.0052076415,
-0.0105063813,
0.0147106033,
-0.0447185189,
-0.0196699016,
-0.022231685,
-0.0424208082,
0.0281850491,
0.0311088581,
0.0089542065,
0.0325253196,
-0.0332876593,
0.0280040354,
-0.0074623879,
0.0160486139,
0.034001153,
0.002854926,
-0.0271414239,
0.0357100815,
-0.0271445028,
-0.0126130208,
-0.0187365375,
-0.0037335262,
0.0023821515,
0.0263842735,
0.0289860331,
0.0290499814,
0.0112669487,
0.0000019944,
-0.0222888086,
-0.0587702654,
0.0129671628,
-0.0456162095,
0.0208900739,
-0.051789619,
-0.0028409031,
0.0131227942,
0.0319953486,
-0.0505695455,
0.0150666274,
-0.0061860802,
0.0263022836,
0.1056005955,
0.0677877665,
-0.0015573467,
-0.0494158305,
0.0415391661,
-0.0130227525,
0.0181252863,
-0.065736711,
0.0343521535,
0.0126781035,
0.0306538995,
0.0008506362,
0.0176225267,
0.0241894331,
0.0379865542,
0.0252716374,
-0.0257958677,
0.0247092191,
-0.0231875423,
0.030207511,
-0.0129915224,
0.1054109633,
0.0357459784,
0.0123075452,
0.0201641843,
0.0172916669,
0.0106444675,
-0.0101661924,
-0.0513361171,
0.0400987603,
-0.0539524741,
-0.002735134,
0.0223304145,
-0.0317603946,
-0.0172573458,
0.0269855298,
-0.0581239723,
0.0746419579,
-0.0115599865,
-0.0384072326,
0.0335655473,
-0.005426466,
-0.0165537409,
-0.0355599485,
0.0420295522,
0.0854825452,
-0.0065642656,
0.0422727019,
0.0055257319,
-0.0144468704,
0.0328725837,
-0.0275589228,
0.0233434346,
0.0339758396,
0.0014315028,
0.0819239318,
-0.0743325651,
0.0520938076,
-0.0356660262,
0.0589117855,
-0.0251993649,
-0.0495531522,
-0.0437662266,
-0.0293889139,
-0.045632232,
0.0726831481,
-0.0595134981,
-0.0016100856,
0.0060828738,
0.0359167345,
0.0146163907,
0.0368028544,
0.0197015647,
-0.0051880064,
0.0055084135,
-0.0376598947,
-0.0452629849,
-0.0050826496,
-0.0683045015,
0.0255663563,
-0.0077494495,
-0.0760111287,
-0.0290405452,
-0.0110132918,
0.0026966305,
-0.0280101839,
0.0321204364,
0.0749372616,
-0.0272926912,
-0.0118085407,
-0.0527083948,
-0.0381650329,
0.0183533672,
-0.0810204819,
-0.0063385903,
-0.0541621633,
0.0623274744,
0.0095610209,
-0.0190616585,
-0.0245691277,
0.0144881215,
0.0518907122,
-0.0073543563,
0.066979222,
0.0083263936,
-0.0042043887,
0.0132904667,
0.0174734704,
-0.0216610525,
0.0582137257,
0.0355081894,
0.0139917172,
0.0008922002,
0.0094170561,
0.0070377225,
0.0255347062,
0.0388623998,
-0.007358951,
0.0266994685,
0.0939989612,
-0.0126700504,
0.0136902975,
0.0157837886,
0.0376841128,
-0.0155139668,
-0.0087972237,
0.0344430767,
0.0551082864,
0.0066859205,
0.0226036012,
-0.0070503782,
0.0284108985,
0.0362893343,
0.0040768315,
-0.0201782454,
0.0482343212,
-0.0328369252,
0.0159200039,
0.0418047197,
0.0306922924,
-0.0456350073,
-0.0642079934,
0.001644515,
0.0239153728,
-0.0311276447,
0.0368999317,
0.0293540806,
0.0160283279,
-0.0690258741,
0.0394681618,
-0.0066433786,
-0.0294936281,
-0.0419436321,
-0.0409619845,
0.06615334,
0.0540687777,
0.0340441354,
-0.0515667461,
-0.0480396673,
-0.0353548042,
0.0210268237,
0.0145215699,
-0.0015195805,
-0.0172663629,
-0.007095885,
-0.0423219651,
-0.005422248,
-0.0386322811,
0.0052212635,
0.0140759842,
0.0382200517,
0.0160282142,
-0.0054228087,
-0.0470283963,
0.0228539295,
0.0262235273,
0.0108319437,
-0.0617674105,
0.0019492325,
-0.004121731,
0.0122759342,
0.0159303788,
0.0157915391,
0.0102083571,
0.0196462795,
-0.0268046539,
-0.0334877111,
-0.0463994928,
-0.0056863669,
-0.0396731496,
0.0339950584,
0.009796422,
-0.0282824785,
0.0154116461,
-0.075111568,
-0.0170811415,
-0.0240131319,
-0.0247071963,
-0.0273695774,
-0.0496373735,
-0.0352527276,
0.0530071668,
0.0265925694,
-0.023220513,
0.0886484459,
-0.010635851,
0.0052838526,
-0.0091628088,
-0.0291297175,
0.0141463932,
-0.0435478166,
0.021527851,
0.0085035842,
-0.0212513302,
0.020649394,
-0.0249145329,
0.0157047175,
-0.0530996881,
-0.0249262843,
-0.0545291305,
0.0524775088,
-0.0280645676,
0.069289811,
0.0324587077,
-0.0047701942,
-0.0154208038,
-0.0211673304,
-0.0822476149,
-0.0353598557,
0.0181884244,
0.0192138944,
0.0586398654,
0.0077166068,
-0.0154833822,
-0.0237962697,
-0.0133185582,
-0.0168817565,
0.0149612678,
-0.0129229249,
-0.0269025303,
-0.0220119599,
-0.0281360466,
-0.0259230081,
0.0137039181,
0.0050463239,
0.0031317689,
-0.0072342712,
-0.0412416905,
-0.0240383931,
0.0758307204,
-0.0260916464,
0.0277409032,
-0.0026190472,
0.0053477408,
-0.0403232723,
0.0446763299,
0.0098775066,
-0.0022125251,
-0.0115929302,
-0.0275130644,
0.043663945,
0.0375011861,
-0.0170894098,
0.0009455211,
-0.0444546565,
0.0385030285,
0.006725315,
0.053070534,
-0.0261092819,
0.0130919376,
-0.0006376173,
0.0524448529,
-0.0200316999,
-0.0056275902,
0.0528205968,
-0.0219070055,
-0.0275559649,
-0.0588063076,
0.0042195865,
-0.0819105208,
0.0587455407,
-0.0199636072,
-0.0096560577,
-0.0262230318,
-0.0253756419,
-0.0226151589,
-0.0480551012,
0.0322988033,
-0.0632705837,
0.0588050149,
-0.0379721895,
0.0308801774,
-0.0264610741,
-0.0333824828,
-0.0259943549,
-0.0107124113,
-0.0286636595,
0.024427712,
-0.0403354205,
0.0029663045,
0.0002901415,
-0.0098985638,
0.0104763228,
0.036527317,
-0.0213752873,
-0.0274862051,
-0.0301658995,
0.082610935,
0.0446591601,
0.0156387407,
0.0199461486,
-0.0172901675,
0.0589031577,
0.0729145929,
0.0292334221,
-0.0123539241,
0.0047040591,
-0.0025416045,
0.013575023,
-0.0455973409,
0.0168232955,
0.0121174343,
0.0359517559,
-0.0839249939,
0.004746296,
0.0776898786,
0.0432285257,
0.0364238918,
-0.0291952658,
0.013606186,
-0.0070183272,
0.003245763,
0.0644899458,
-0.1011422053,
-0.0495611504,
-0.0055748792,
0.021634832,
0.0203805417,
-0.0213740245,
-0.0238507222,
-0.0951133519,
0.0459345467,
-0.03341287,
-0.0121229142,
-0.0605504066,
-0.0452012867,
-0.0025537144,
-0.006624897,
-0.0028394393,
0.0383785367,
-0.0225473419,
0.0121046957,
0.0140743013,
-0.014826104,
-0.0010778836,
0.0489378385,
-0.0027808805,
0.0000275311,
0.0414051004,
0.0203266088,
-0.0001702265,
-0.0429262333,
0.0139807612,
-0.0511234738,
-0.0140226213,
-0.0052109337,
0.0381314829,
-0.0228435118,
0.0095663853,
-0.0029573846,
0.0537141114,
-0.0501435846,
0.0425339341,
0.0259462893,
0.0774831772,
-0.0232959315,
-0.0168393627,
0.0112398518,
0.001828177,
-0.0469529815,
0.0790793821,
0.0199795961,
0.0218904745,
-0.0853639618,
-0.0404593311,
-0.052820392,
-0.072261855,
0.0271782968,
-0.000315704,
0.0260215085,
0.0337939784,
-0.0669663325,
-0.0443409607,
0.057488203,
0.0345138237,
-0.031005213,
0.0374721698,
-0.0111820465,
0.0124341734,
-0.0297668409,
0.0203281809,
-0.0463299155,
0.0207423139,
-0.0167772137,
-0.01522299,
0.0051143528,
-0.0257059522,
-0.0471740365,
0.0274446588,
-0.084775731,
-0.0282085724,
-0.0628191233,
0.0066083367,
0.0231890269,
0.0185042564,
-0.0250989739,
-0.0031091331,
0.0337211974,
-0.0052199946,
-0.0800897032,
-0.0353531651,
0.0575980991,
-0.0063826479,
0.0042466223,
0.0393178575,
-0.0471090116,
0.0047088424,
-0.0170432571,
0.0490292534,
-0.0211723521,
-0.0001768134,
0.0302653387,
0.0156972483,
-0.0329791419,
0.0656339005,
-0.0351670645,
-0.012748722,
0.0511242487,
-0.0104406476,
-0.0204324033,
-0.0216024257,
0.0173646323,
0.0120421425,
-0.0501854122,
-0.0238546692,
-0.0027921482,
0.0054644132,
0.0294896159,
-0.0250221249,
0.0187494438,
0.0425403602,
-0.0075028222,
-0.0036652291,
-0.0586252809,
-0.0266749077,
-0.0083886692,
0.0171740819,
0.0630541071,
-0.0138331568,
0.0620647073,
0.0365471728,
-0.020253608,
-0.013211783,
0.0290720798,
-0.0720794573,
-0.0839117318,
-0.0113990325,
0.0433563888,
0.009987724,
0.0248191468,
0.008947568,
-0.013067171,
-0.0160569176,
0.0276659261,
-0.0156342573,
0.0312713459,
-0.0028731972,
0.0539060645,
0.0340664238,
0.0077931914,
-0.0113535374,
0.0693520978,
0.0053130644,
0.0263713021,
-0.0189546719,
-0.0287710391,
-0.0535979271,
-0.0417699553,
-0.0417668521,
0.026252361,
0.0039178017,
-0.0214113463,
-0.0336915255,
-0.0361099839,
0.0206404254,
-0.0095509076,
-0.0359716304,
-0.0109658595,
0.0099321315,
0.0663262978,
0.0566464663,
-0.0322127007,
-0.0572167411,
0.0105833756,
-0.0048139701,
-0.0382730626,
-0.0076592127,
-6.4945218e-33,
-0.0310153775,
-0.0147519736,
0.045906961,
0.0108842971,
0.0192868691,
-0.0140785938,
-0.0484563261,
-0.0556458868,
-0.0477749445,
-0.0111090941,
-0.0341231525,
0.0319468565,
0.0127492798,
0.0101401387,
-0.0146723418,
-0.0472685061,
-0.00520139,
-0.0236003548,
-0.0307473838,
0.0161300078,
-0.0041895891,
0.0358919725,
0.0192559939,
0.0067391861,
0.0294467565,
0.0006651992,
-0.036158476,
0.011372895,
-0.0452170484,
0.0551379211,
-0.0333904661,
0.0126332929,
0.0435514376,
-0.0807061046,
0.0083773006,
0.0500647984,
-0.1419878453,
-0.0267662108,
0.0077588232,
-0.0109068993,
-0.0411237553,
-0.0168669261,
0.0145205846,
-0.0128761688,
0.0410120152,
-0.031181233,
0.0109594315,
-0.0319227688,
0.0319499411,
-0.0500071086,
-0.02342516,
0.0276602861,
-0.0299597289,
0.0047735721,
0.0283966996,
0.0422520079,
0.0086604757,
0.0211090446,
-0.0456024595,
-0.0208566599,
0.0190667715,
-0.000198285,
0.0418088585,
-0.0432700254,
-0.0069777207,
-0.0041294117,
-0.0105027175,
-0.0360691361,
-0.0479064099,
0.0163884256,
-0.037714202,
0.0301569235,
0.0483120978,
-0.0178535189,
0.0331347957,
0.0221565645,
-0.0802436918,
0.0389878191,
0.024759315,
0.05526907,
0.0325215422,
-0.0189349446,
-0.0437458158,
0.0014476466,
-0.0212624744,
0.0312747955,
-0.029622307,
0.0138487983,
0.0349398889,
0.0121822925,
0.0031888252,
-0.0026452478,
0.0054332889,
-0.0186356902,
-0.0034201969,
0.055971995,
-0.0118062301,
0.002728143,
0.0238769874,
0.0506044254,
-0.0242165662,
0.0227675196,
0.0080511374,
-0.0319023915,
-0.0044657984,
0.0075743129,
0.0405157283,
-0.0189628266,
-0.0819844007,
0.0093731061,
-0.0597995296,
0.0685750619,
-0.0173320379,
0.0433273055,
-0.0090477392,
0.0404651873,
-0.0017185423,
0.0618888251,
-0.018042855,
0.0167096667,
-0.0327041261,
0.0232123639,
-0.0241494682,
0.0065225214,
-0.0154714277,
0.0142317135,
-0.0047968398,
0.0282082185,
-0.0472559892,
-0.0311533846,
0.0209912211,
0.0598244481,
2.632e-7,
0.0573017076,
0.0170327891,
0.0335279591,
0.0268567186,
0.0171319786,
0.001997835,
0.0008765145,
0.0046671731,
-0.0075535644,
-0.0362158902,
0.0116515569,
0.0648992136,
0.0431098975,
0.0437885337,
-0.0518760085,
-0.0160074178,
-0.027665861,
-0.0265467279,
-0.0702998936,
-0.0169221219,
0.0761931092,
0.0707415864,
-0.017081866,
-0.0353618935,
-0.0188878905,
-0.0498820543,
0.0407187976,
0.0268578455,
0.0331070684,
0.0340496525,
0.0477421768,
-0.0111345705,
-0.0158570837,
0.0296091773,
0.0090654055,
0.023308998,
-0.003961809,
0.1041415036,
0.0245139636,
-0.0120410444,
-0.0126679512,
-0.0679490715,
0.0274071842,
-0.0529898666,
0.1339891255,
-0.0201501474,
-0.0200249199,
0.0189536884,
-0.0324854515,
-0.0304005165,
0.0041229702,
0.0038886743,
-0.0003315028,
0.0236572549,
-0.0054324414,
-0.0487957485,
0.0241563804,
-0.0896264166,
0.013913244,
0.0094966311,
-0.0194891561,
0.0077058543,
-0.035716895,
-0.0107429475,
0.0202068686,
-0.0657935292,
-0.0019975652,
2.659487322e-34,
0.0305652227,
-0.0311610866,
0.0309702866,
0.0540839247,
0.051162187,
0.0105655352,
-0.0079784133,
-0.0010281155,
0.0270586777,
-0.0944391936,
0.0000973088
] |
http://arxiv.org/abs/2401.13218v1 | http://arxiv.org/pdf/2401.13218v1 | ULTRA: Unleash LLMs' Potential for Event Argument Extraction through Hierarchical Modeling and Pair-wise Refinement | 2024-01-24T04:13:28 | 2024-01-24T04:13:28 | Xinliang Frederick Zhang, Carter Blum, Temma Choji, Shalin Shah, Alakananda Vempala | cs.CL | cs.CL | Structural extraction of events within discourse is critical since it avails
a deeper understanding of communication patterns and behavior trends. Event
argument extraction (EAE), at the core of event-centric understanding, is the
task of identifying role-specific text spans (i.e., arguments) for a given
event. Document-level EAE (DocEAE) focuses on arguments that are scattered
across an entire document. In this work, we explore the capabilities of open
source Large Language Models (LLMs), i.e., Flan-UL2, for the DocEAE task. To
this end, we propose ULTRA, a hierarchical framework that extracts event
arguments more cost-effectively -- the method needs as few as 50 annotations
and doesn't require hitting costly API endpoints. Further, it alleviates the
positional bias issue intrinsic to LLMs. ULTRA first sequentially reads text
chunks of a document to generate a candidate argument set, upon which ULTRA
learns to drop non-pertinent candidates through self-refinement. We further
introduce LEAFER to address the challenge LLMs face in locating the exact
boundary of an argument span. ULTRA outperforms strong baselines, which include
strong supervised models and ChatGPT, by 9.8% when evaluated by the exact match
(EM) metric. | ULTRA: Unleash LLMs' Potential for Event Argument Extraction through Hierarchical Modeling and Pair-wise Refinement: Structural extraction of events within discourse is critical since it avails
a deeper understanding of communication patterns and behavior trends. Event
argument extraction (EAE), at the core of event-centric understanding, is the
task of identifying role-specific text spans (i.e., arguments) for a given
event. Document-level EAE (DocEAE) focuses on arguments that are scattered
across an entire document. In this work, we explore the capabilities of open
source Large Language Models (LLMs), i.e., Flan-UL2, for the DocEAE task. To
this end, we propose ULTRA, a hierarchical framework that extracts event
arguments more cost-effectively -- the method needs as few as 50 annotations
and doesn't require hitting costly API endpoints. Further, it alleviates the
positional bias issue intrinsic to LLMs. ULTRA first sequentially reads text
chunks of a document to generate a candidate argument set, upon which ULTRA
learns to drop non-pertinent candidates through self-refinement. We further
introduce LEAFER to address the challenge LLMs face in locating the exact
boundary of an argument span. ULTRA outperforms strong baselines, which include
strong supervised models and ChatGPT, by 9.8% when evaluated by the exact match
(EM) metric. | [
0.0479930937,
0.0514130741,
0.0172036793,
0.083783187,
-0.0310161337,
0.0107809184,
-0.0286642108,
-0.0243429746,
0.0285225343,
-0.0697969496,
-0.0219819304,
0.000849906,
-0.042513743,
0.0299756918,
-0.0170344599,
-0.0636392832,
0.0335902981,
-0.0042484454,
0.0569972433,
0.0159664117,
0.0312358178,
0.0261337217,
-0.0080321673,
0.0423893258,
-0.0044899746,
-0.0444420427,
-0.0237661581,
0.0059046447,
0.0405289456,
-0.0128438706,
0.0480109528,
-0.0036225594,
0.0275964979,
0.0406349674,
0.0000019092,
-0.0544739924,
-0.0352731459,
-0.0305234268,
-0.0338186584,
0.0502723455,
0.0214230232,
0.0158278011,
-0.0201937389,
-0.0018772408,
-0.0362689309,
-0.0151671302,
0.0057146871,
0.0639764071,
0.0370545723,
0.1024636254,
-0.0174565315,
-0.0132773872,
0.0041568242,
-0.0283194613,
-0.0335402526,
-0.050801333,
0.0404110365,
-0.0215396564,
-0.0093815774,
0.0306768846,
0.0154136932,
0.0165413767,
0.0049890676,
-0.0254642889,
-0.0215395298,
0.0167668276,
-0.07712733,
0.0040370254,
-0.0097442418,
0.0699657053,
0.0020074598,
-0.0156517513,
0.0373389162,
-0.0051617785,
0.0078481026,
0.0098303473,
-0.0561852716,
0.0394381098,
-0.0119475657,
0.0399503857,
0.0831390396,
0.0178486984,
0.0160270594,
0.0110604027,
-0.0009280584,
0.0648792908,
-0.0112790177,
-0.0030334243,
-0.0036829442,
-0.0237341523,
-0.0319611505,
-0.0654156655,
0.0295512434,
0.0369166471,
-0.0012664035,
0.0274346787,
-0.015709985,
0.0321678892,
0.0042675571,
-0.0740037486,
-0.0030251485,
0.0047433651,
-0.0169532932,
0.067312859,
0.0148057854,
-0.0117040016,
-0.0073593045,
-0.0118683623,
-0.0071246512,
-0.0262033045,
-0.0334618166,
0.0009276068,
-0.0355065838,
0.0246423967,
-0.0677197278,
-0.0387837514,
0.0239789784,
0.0083300574,
0.0107649993,
0.0057333168,
0.0104008215,
-0.0066020116,
-0.023535585,
0.0012795334,
0.0235128012,
0.0355469957,
-0.0350354575,
0.0042719496,
-0.004480314,
-0.0433415137,
-0.0314905383,
0.0433390178,
0.0265620723,
-0.0257705785,
-0.0233094562,
0.0314365737,
-0.0124540394,
-0.0180540532,
-0.0344801061,
-0.0311413854,
-0.0280142259,
-0.0356304757,
-0.014352222,
-0.0526707061,
0.016880339,
-0.0303985514,
-0.003148048,
-0.0198053885,
0.0375494249,
-0.0006013334,
-0.0557180755,
0.0160410032,
0.0015278899,
-0.0077418345,
0.0221604202,
0.017283788,
-0.0212863255,
0.0621797591,
0.0511640087,
0.0252250042,
0.0857099295,
0.0345821753,
-0.0256982464,
0.0493199714,
0.003866117,
-0.0142310336,
0.0685967207,
-0.0026067439,
-0.0149929356,
-0.0056814845,
0.0377948098,
0.0074630599,
-0.0448987298,
-0.0561549217,
0.0497147068,
0.0361009017,
-0.0359485783,
-0.0167780407,
0.0459054336,
0.029078031,
0.0281728394,
0.029397428,
-0.0050632083,
-0.0504340716,
0.0188320801,
-0.0203163382,
0.0065329084,
-0.0268799029,
-0.0188611839,
-0.0143196657,
-0.0121294428,
-0.0436675139,
-0.0353203341,
0.0007964239,
0.0388166234,
0.0045114043,
-0.0285713878,
0.0554505251,
0.0490531921,
-0.0664697438,
-0.0150171667,
0.0057913018,
0.0234470051,
-0.0053683864,
0.0381676033,
-0.0491533652,
-0.0456296392,
-0.009243425,
0.0666896179,
0.0232929941,
0.0438414514,
-0.0485534482,
0.0703672692,
-0.0274886116,
0.016764218,
-0.036972329,
0.019769134,
0.0113507723,
0.0152921602,
-0.0385813899,
-0.0111208307,
-0.0091232574,
-0.0362739861,
0.0223069247,
-0.024085233,
-0.0381816775,
-0.0056129508,
-0.0111580677,
0.0257923771,
0.0239290688,
-0.0023278415,
-0.01284777,
0.0404056162,
-0.0427251011,
0.0169953313,
0.0273701176,
-0.0375670716,
-0.0354874097,
0.0448416881,
-0.0077921706,
-0.0419674516,
0.0139007913,
-0.0617153868,
-0.0387878083,
-0.0043162587,
0.0141299935,
0.0157395825,
-0.0053571966,
-0.0027923142,
0.017993886,
-0.0058157435,
-0.0252471082,
0.054088112,
0.0297237076,
-0.026746016,
-0.0134708527,
0.0285910685,
0.0198670235,
-0.0384563506,
-0.035813868,
0.0237552058,
0.0052093812,
-0.0296552479,
0.0419622809,
-0.0156777948,
-0.0398855433,
-0.0271623172,
-0.1020920053,
0.0453059077,
-0.032766141,
0.0485675298,
0.0591702461,
0.0027968811,
-0.010759226,
0.0029375961,
-0.0235180706,
-0.0630029589,
-0.037158411,
-0.0141164185,
0.0566675812,
0.0179218762,
0.0141588952,
-0.0810396448,
0.0251207799,
-0.0505174212,
0.0567147881,
0.0094971312,
-0.0293953586,
-0.0213331338,
-0.0088648144,
0.0079907412,
0.0448922999,
-0.0308724046,
-0.0224497356,
0.0028256846,
0.0025980291,
0.0219914932,
0.0954716355,
0.0099777216,
0.0793194622,
-0.0108085582,
-0.0442203954,
-0.0563337617,
0.0160981175,
-0.0247137863,
0.0129877133,
-0.0405489728,
0.0238779746,
0.0236031301,
0.0809268877,
-0.0165359378,
0.0157874934,
-0.0176265277,
-0.002030649,
0.0138707682,
0.0049076038,
0.0331149548,
-0.0399342924,
-0.0445857272,
0.0131532075,
-0.0379766524,
0.0019887097,
0.0410640612,
-0.0493446849,
-0.0231623072,
-0.0312744826,
-0.0116897374,
-0.0670231804,
0.0714901462,
-0.0059525343,
-0.0054467223,
-0.0051297606,
-0.0433375798,
-0.0013058868,
0.0001679842,
0.0484245829,
-0.0139735034,
0.0362730548,
-0.0556965806,
-0.0004212146,
-0.0154152056,
-0.0414340124,
-0.0393106677,
-0.0516058244,
-0.0183219351,
0.0063026901,
-0.0379627384,
0.0002406289,
-0.0349751636,
-0.0084643336,
-0.0145467352,
0.0422403254,
-0.0221362337,
0.0207555927,
-0.030990202,
0.0529060736,
0.0317251906,
-0.003853465,
-0.0048351395,
0.0070437738,
0.0202326365,
0.0940297469,
-0.0159733724,
0.0046582986,
-0.0362152606,
0.0046817348,
0.0041985447,
-0.0407535248,
-0.0221469067,
-0.0430470333,
0.079784669,
-0.077133812,
0.0212527737,
0.1097746268,
0.026874464,
-0.0157225244,
-0.0012168175,
-0.020555554,
0.0502030402,
-0.0017866477,
-0.003248296,
-0.0862597376,
-0.0314486697,
-0.0131061003,
0.0186240561,
-0.0226133112,
-0.0523070097,
0.0030613868,
-0.0815034956,
0.0319793113,
0.0031953405,
0.020675363,
0.0191441737,
-0.0294854045,
0.0355362855,
-0.0374012664,
-0.0177835375,
0.0104161017,
-0.0217740051,
-0.00647128,
0.0036519445,
-0.0834030956,
-0.0403644703,
0.0450635403,
-0.0467894264,
0.0254759751,
0.0052412269,
0.0692822635,
0.0442184769,
0.0246745981,
0.0432372093,
-0.0426819064,
-0.0346694998,
-0.0108757094,
0.0429501049,
-0.0316878259,
-0.000114555,
0.000660339,
0.0807070807,
-0.0128387203,
-0.0387788638,
0.0234729778,
0.0553561039,
-0.0050405175,
0.0012325268,
0.0161169916,
0.0130439159,
0.0244995877,
0.0700452477,
0.042600695,
0.0021996775,
-0.039117191,
-0.023983784,
-0.0058876211,
-0.0220539328,
-0.0089449231,
0.0004449475,
-0.0118827121,
-0.0205293242,
-0.0870252997,
-0.061014656,
0.0143332714,
0.0128222136,
-0.0264195483,
0.0219987761,
0.0179359317,
-0.0116802529,
-0.0183449779,
0.0481815971,
-0.0516187213,
-0.0244270097,
0.0041392567,
-0.0108651854,
0.0295874961,
0.0216081422,
-0.0262991711,
0.0811038092,
-0.0271578636,
-0.0753549859,
-0.0076472368,
0.0040699383,
0.0200385861,
0.0913602263,
-0.0211390723,
-0.0200597215,
-0.0254794713,
0.009020282,
0.0068369783,
-0.0091332821,
0.0593116321,
0.013602945,
-0.0146094002,
0.0500760637,
-0.0099132741,
0.0268362481,
-0.0271989759,
0.0477001183,
-0.062697418,
-0.0412640385,
0.0063443263,
-0.0054383036,
0.0689190701,
0.0240788069,
-0.0250584837,
-0.0385213308,
0.0072869882,
-0.0514894687,
-0.0131805716,
-0.0347952247,
0.0498732664,
-0.020754192,
-0.0250176899,
-0.02416116,
0.0062370459,
-0.0260424875,
0.0556426868,
-0.0211311504,
0.0278601293,
0.0255590249,
0.0214708839,
-0.0191832688,
-0.0159807559,
-0.0315901488,
0.0295070373,
0.0064211432,
0.0275862925,
-0.0323790945,
0.0359641388,
0.0323457383,
-0.0096372589,
0.0248451363,
0.0026603551,
-0.0695553422,
0.0023495227,
-0.0042346274,
-0.0174600389,
0.0632483959,
-0.0711714551,
0.1035165787,
-0.0162647497,
-0.0131629463,
-0.0188618656,
0.0492803641,
-0.0033915157,
-0.0059569404,
0.0240593906,
0.0302952938,
-0.0511856787,
0.0517734028,
0.0615850277,
-0.00480734,
0.0558859482,
0.0393560119,
-0.0003291423,
-0.040896222,
-0.0422784686,
0.042018462,
0.0341825187,
-0.0023886797,
-0.0102342032,
-0.0415840633,
0.0281935912,
0.0210420322,
0.0022706571,
-0.0334038995,
-0.0090325316,
0.0248193666,
0.0536468029,
0.0081515014,
0.0078203175,
-0.0608526058,
0.0081658782,
0.014151061,
-0.0355676115,
-0.0335070714,
-6.47415636e-33,
-0.0187753551,
-0.0111745708,
-0.0007074834,
0.1148333102,
-0.0346104763,
0.0318736173,
-0.0498324037,
-0.0113845924,
-0.0135396877,
0.0188840702,
0.005809966,
-0.0148866102,
0.0185448993,
-0.0022614452,
-0.0006521774,
-0.0189375822,
0.0364563204,
0.0045490675,
-0.023631271,
0.0398648307,
0.0651971474,
0.018869536,
0.0740364045,
-0.0965321288,
0.0363653228,
-0.0194201283,
0.0020248774,
0.0143077653,
0.0320234261,
0.0470264964,
0.0037149803,
0.002125127,
0.0136182569,
-0.0166149028,
0.0089039318,
0.0354511477,
-0.1005087048,
-0.0212058164,
0.0448822305,
-0.0303791109,
-0.0317222588,
-0.0390563942,
0.0126677025,
0.0105500203,
0.0418116711,
-0.0336735919,
0.0363517851,
-0.011321187,
-0.0080224592,
-0.0334440693,
-0.067500256,
0.0712840408,
-0.0260402597,
-0.0075819511,
0.0012427687,
0.0172092896,
0.0492093228,
0.0446492732,
-0.0525698401,
-0.0149912564,
-0.0386614837,
0.0199079085,
-0.0028771211,
-0.0325345993,
-0.0183913428,
0.080495961,
-0.0363869853,
0.0032503076,
-0.026439501,
0.0260754582,
-0.002267892,
0.0787825286,
0.0004080813,
0.0491532125,
0.0040655024,
0.0302690584,
-0.0949667692,
0.0120997271,
0.0707329214,
0.0294678025,
-0.0265897792,
0.0146157183,
-0.0236402396,
-0.0006824978,
-0.007141877,
-0.0162754245,
0.0021443279,
-0.0116429767,
0.0060465843,
-0.0001622074,
-0.0020094404,
0.05368701,
0.0299514942,
-0.0277266596,
-0.0223399121,
0.0096815303,
-0.0271590296,
0.0101086432,
0.0008920474,
0.0327258594,
-0.0446083434,
-0.0243453532,
0.0031755406,
-0.0233579203,
0.0464534573,
0.0173620246,
0.0404482111,
-0.0158988368,
-0.0233543832,
-0.010747605,
-0.0257282946,
0.0032186976,
-0.0702775195,
0.0399512574,
0.0176508129,
0.0090407925,
0.0065760054,
0.0780908689,
0.010593595,
0.0235492773,
-0.0401043147,
-0.0365493149,
-0.0189223792,
-0.0049721319,
-0.0246458091,
-0.01262639,
0.0018093453,
0.0365077741,
0.0028159542,
-0.0433442518,
0.0151621206,
0.0036716417,
2.598e-7,
0.0619612671,
0.0431588404,
0.0611627884,
0.0069155036,
0.0239768326,
0.0212633926,
-0.0146850031,
0.0294980891,
0.0079111224,
-0.0247665439,
-0.0189483389,
0.0058941073,
0.0554227158,
0.0391296558,
-0.1044222191,
-0.0529174358,
-0.1136637405,
-0.0492183566,
-0.0576152951,
-0.0135874655,
0.0667617321,
0.0799614117,
0.0455560349,
0.0136372363,
-0.0249980595,
-0.0165706184,
0.0582614057,
-0.0161863007,
-0.049122218,
0.0555027016,
0.0249000546,
-0.0190736484,
-0.0247988664,
0.0072625317,
-0.0187516958,
0.0590396523,
-0.0165811125,
0.0967819318,
-0.002176241,
-0.01573769,
0.0077973073,
-0.038556654,
0.0049377633,
-0.057718087,
0.0734882802,
-0.0506357737,
-0.0425560176,
-0.0230820235,
-0.0254976116,
0.0252509322,
-0.0431438275,
0.0024411911,
-0.0090860371,
0.0545703582,
-0.0131744249,
-0.037803933,
-0.0061519397,
-0.089696683,
-0.0286640767,
0.0109882122,
-0.0263892915,
-0.0250542946,
0.0142797679,
-0.0320974998,
-0.0021847975,
-0.0437270403,
0.0072837821,
2.416889624e-34,
0.0042642555,
-0.0076735034,
-0.0412412025,
0.0417069793,
0.0643029138,
-0.0111121554,
-0.0277090147,
0.0222966541,
-0.0062709241,
-0.0240478106,
0.0080882711
] |
http://arxiv.org/abs/2401.16212v1 | http://arxiv.org/pdf/2401.16212v1 | Better Call GPT, Comparing Large Language Models Against Lawyers | 2024-01-24T03:53:28 | 2024-01-24T03:53:28 | Lauren Martin, Nick Whitehouse, Stephanie Yiu, Lizzie Catterson, Rivindu Perera | cs.CY | cs.CY, cs.CL | This paper presents a groundbreaking comparison between Large Language Models
and traditional legal contract reviewers, Junior Lawyers and Legal Process
Outsourcers. We dissect whether LLMs can outperform humans in accuracy, speed,
and cost efficiency during contract review. Our empirical analysis benchmarks
LLMs against a ground truth set by Senior Lawyers, uncovering that advanced
models match or exceed human accuracy in determining legal issues. In speed,
LLMs complete reviews in mere seconds, eclipsing the hours required by their
human counterparts. Cost wise, LLMs operate at a fraction of the price,
offering a staggering 99.97 percent reduction in cost over traditional methods.
These results are not just statistics, they signal a seismic shift in legal
practice. LLMs stand poised to disrupt the legal industry, enhancing
accessibility and efficiency of legal services. Our research asserts that the
era of LLM dominance in legal contract review is upon us, challenging the
status quo and calling for a reimagined future of legal workflows. | Better Call GPT, Comparing Large Language Models Against Lawyers: This paper presents a groundbreaking comparison between Large Language Models
and traditional legal contract reviewers, Junior Lawyers and Legal Process
Outsourcers. We dissect whether LLMs can outperform humans in accuracy, speed,
and cost efficiency during contract review. Our empirical analysis benchmarks
LLMs against a ground truth set by Senior Lawyers, uncovering that advanced
models match or exceed human accuracy in determining legal issues. In speed,
LLMs complete reviews in mere seconds, eclipsing the hours required by their
human counterparts. Cost wise, LLMs operate at a fraction of the price,
offering a staggering 99.97 percent reduction in cost over traditional methods.
These results are not just statistics, they signal a seismic shift in legal
practice. LLMs stand poised to disrupt the legal industry, enhancing
accessibility and efficiency of legal services. Our research asserts that the
era of LLM dominance in legal contract review is upon us, challenging the
status quo and calling for a reimagined future of legal workflows. | [
0.0371563137,
0.0492595509,
-0.0150720235,
-0.0454031453,
-0.0756955892,
0.0180779286,
0.0314100981,
0.0622753501,
-0.0064591221,
-0.0507548563,
-0.0211985335,
-0.0141701577,
-0.0127900187,
0.0271750707,
-0.0028682814,
0.0016088787,
0.0139615713,
-0.012743758,
0.0321674645,
-0.0272320416,
0.0076127062,
0.006119214,
0.0068864338,
0.0154940402,
-0.0231041163,
-0.0012661773,
-0.0053094598,
0.0053739292,
-0.0388114341,
-0.0062961932,
-0.0403839536,
0.0855919346,
0.0013457228,
0.0581545271,
0.000002022,
-0.0637721047,
0.0097450772,
-0.0045566508,
-0.0257125441,
0.0957505554,
0.0299513917,
-0.0407591052,
-0.027780734,
0.0242021605,
-0.0079363808,
0.0372705236,
0.0071721626,
-0.0046862634,
0.0166641269,
0.056782607,
-0.0095907282,
-0.0182149671,
0.0151983751,
-0.0458654985,
-0.0284542385,
-0.0771198124,
0.0079889754,
0.0130251488,
0.0439915359,
-0.0404661298,
-0.000556219,
0.0166583415,
0.0300426371,
-0.0503999703,
-0.0398876928,
0.0103398617,
-0.0244864821,
0.0359131247,
-0.0145024639,
0.0165776312,
0.0536183156,
0.0010645364,
-0.0030716325,
-0.0194496494,
0.0318854414,
0.03848847,
-0.0201071501,
0.077457197,
-0.0052274675,
0.005014353,
-0.0284275226,
-0.0164502952,
0.0193319041,
0.0173053462,
-0.0062926752,
0.0668834075,
0.0006260511,
-0.022388136,
0.0464507751,
0.0604496114,
0.0473869853,
-0.0786214694,
0.0287038293,
0.0230585355,
-0.013228084,
0.0054975194,
-0.0038390178,
-0.0085233338,
0.0603231825,
-0.038175106,
0.0308957975,
-0.0048006438,
-0.0079082744,
0.0575706698,
-0.0350472406,
0.0050104326,
-0.0488863178,
0.016792696,
0.0054510925,
0.011600174,
-0.0159340762,
0.0047655595,
-0.026465429,
0.0421534255,
-0.085451059,
0.0093322974,
-0.0484793372,
0.0075927032,
-0.0029326361,
0.0445156544,
0.0424975194,
0.0030417172,
0.0265482683,
-0.0427841544,
-0.0897944421,
-0.0542691313,
-0.0341359116,
0.0526625142,
-0.0044591678,
0.0071719019,
0.0276286099,
0.0149565628,
0.0021057473,
0.0044373996,
0.0126480144,
0.1101773307,
0.0180974044,
0.0180514567,
-0.0216709673,
-0.0535588376,
0.0038425613,
-0.057341788,
-0.0727709606,
-0.0637723804,
-0.0315352939,
-0.0118912952,
-0.0441668332,
-0.058225669,
-0.0046689897,
0.0132065704,
-0.0070203524,
0.0375325792,
-0.0657230541,
0.0355143584,
-0.0274346974,
0.0149027146,
-0.0097444104,
-0.0024445993,
-0.0092062149,
0.0163414348,
0.0221031774,
-0.0274482053,
0.0743669942,
0.0693228915,
-0.068525292,
0.0579573661,
-0.0064368178,
0.0110043883,
0.0264176223,
0.02549267,
0.0364553183,
0.0301478747,
-0.0244459119,
0.0113275275,
0.0737592429,
0.0666224658,
-0.0042121322,
0.063775979,
0.010997803,
0.0151250809,
0.021448249,
-0.0363296978,
-0.0478663407,
0.0260558277,
-0.0222385563,
-0.0047980221,
-0.04995827,
0.0043172059,
-0.0325564593,
-0.0380048752,
0.0106198555,
0.0121332752,
-0.017175287,
0.0134888114,
0.0250122696,
-0.0243176166,
-0.0272941515,
0.046862714,
0.0121709332,
0.0028798701,
-0.0776833445,
-0.0289300177,
-0.0314510688,
0.0395192467,
0.0109471437,
0.0045748609,
0.0194610748,
-0.015961865,
-0.0263091922,
0.0629640147,
0.0493059345,
0.0309729818,
-0.0404412374,
-0.0128382985,
-0.0280405637,
0.0161852334,
0.0105586303,
0.0225547198,
0.0205604006,
0.0251386221,
0.0248420425,
-0.0642668754,
-0.0383525901,
-0.0095780101,
-0.0077059083,
-0.0432299562,
0.0175036751,
-0.009101945,
0.0236786585,
0.0108699342,
0.0025975555,
0.0003859279,
0.014348709,
0.0156832132,
-0.0025406864,
0.0199418347,
0.0418889225,
0.0727209598,
0.0098481439,
-0.0041607334,
-0.01996544,
-0.0220854543,
-0.0272275843,
-0.0013499792,
-0.0338184424,
-0.0077937022,
0.0231363308,
-0.0317719132,
0.0021493623,
-0.0151477298,
-0.0313673466,
0.0133913914,
0.0149236945,
0.018563075,
-0.0020621871,
-0.0159334373,
-0.0063835923,
0.0290616993,
-0.0266209878,
-0.0031074339,
0.0406748615,
-0.0026557762,
0.017047802,
0.000950394,
0.0463378094,
-0.0308857728,
-0.0352388248,
-0.0685622245,
-0.0225798748,
0.0061501483,
0.0205887761,
0.0099360095,
-0.0479133464,
-0.0144954156,
0.0067426432,
-0.0411851443,
-0.0159605406,
0.0293570422,
0.0064452048,
0.0133447871,
-0.0204452239,
-0.067422621,
-0.0134307863,
-0.0165361688,
-0.0036033446,
0.0170710254,
0.0701568648,
-0.0106516723,
-0.0530006923,
0.0615818389,
-0.0185041837,
-0.0270801987,
0.000172246,
-0.0373406485,
-0.0083776629,
0.0236483775,
-0.0799857602,
-0.0166666582,
-0.0133411232,
-0.0047441917,
0.0388968885,
-0.012661187,
-0.0271666367,
-0.012603241,
0.0574686974,
-0.0209679361,
-0.0356209092,
-0.0246236213,
0.0141294552,
0.0127954371,
-0.0332422256,
0.0091487654,
0.0131662702,
0.0359527133,
0.0181596894,
-0.0129098333,
0.00766506,
-0.0087177921,
-0.0288222898,
0.0877019092,
0.0169801377,
-0.0786297694,
0.0138513995,
-0.0006808883,
0.039556589,
-0.0211474355,
-0.0293414574,
-0.0352339447,
0.0632208809,
-0.0400607251,
0.0137412343,
-0.054869581,
-0.0361876264,
-0.0036206529,
0.0042799963,
0.1006072015,
-0.0152954897,
0.0812531784,
-0.0504716858,
0.0408786386,
-0.0147759607,
-0.0374547206,
0.0070040463,
-0.0199958235,
-0.0568370037,
0.0442067869,
0.0267824251,
-0.005306711,
-0.0341757387,
0.0015254709,
-0.0033574628,
0.0671517104,
0.0228011496,
-0.0119571248,
-0.0163601339,
-0.028849002,
-0.0532238521,
-0.036547713,
0.0300781503,
-0.0700477734,
0.0175684672,
0.0292756185,
-0.0058577592,
-0.0254388098,
0.0127344411,
0.0059206313,
-0.0113342917,
-0.027871849,
-0.0264241304,
0.0463923961,
-0.0132700959,
-0.0522582456,
0.0406292304,
0.0788430795,
-0.0068134139,
-0.0741301626,
-0.0040243054,
0.0153470859,
-0.0023504142,
0.0446653627,
0.0072184424,
-0.0815591216,
0.0315194279,
0.0032833999,
0.0219415929,
-0.0400096215,
-0.0508009233,
-0.0205287598,
-0.0664982647,
0.103376992,
0.0103201298,
-0.0008554171,
0.0104614571,
0.0357793756,
-0.0011794769,
0.000693114,
-0.0040057744,
-0.0037302789,
-0.0159938987,
0.0019075361,
-0.0042067925,
-0.0444999039,
0.0580117889,
0.049133528,
-0.0424880534,
0.0165293142,
0.0265057534,
-0.0304350946,
-0.0297603682,
0.0199287925,
0.0728927627,
0.0128043508,
-0.0081050536,
0.0306824073,
0.0263696015,
0.0552560873,
0.0220006257,
-0.0375456922,
-0.001012123,
-0.0532151684,
0.0173016433,
0.0523790531,
0.047505606,
-0.0269603226,
-0.0187826511,
0.0264560655,
-0.0334905125,
-0.0361010134,
0.0165696237,
-0.0149939489,
-0.0600189641,
-0.0687279105,
-0.0364320688,
-0.0410219021,
0.0474988781,
-0.0586361475,
-0.0012336636,
-0.0138392402,
-0.0130274445,
-0.0207024701,
-0.0439501852,
0.0157085005,
0.0651277006,
0.0126873273,
0.0041811764,
0.0231163502,
-0.0019886084,
0.0141546121,
0.0014576536,
-0.0292719454,
0.0351999365,
-0.0390600711,
0.0264552105,
0.0006199873,
0.0114600882,
-0.0081558917,
-0.0375994779,
-0.017486766,
-0.0024912797,
-0.0526471809,
0.010740215,
0.036059469,
-0.0331691504,
-0.0124869561,
-0.0232593436,
-0.0418012924,
0.0179498512,
-0.0251124837,
0.0202318579,
0.030761078,
0.0187934544,
-0.0119109405,
-0.0221168976,
-0.0156299844,
-0.0012648017,
0.0444891714,
-0.0125015173,
-0.0400039703,
-0.0324679166,
0.0506334864,
0.0792696625,
-0.0414024517,
0.0030600429,
0.0215388928,
0.0007015635,
0.0011810715,
-0.031405516,
0.0038076034,
-0.0288919862,
-0.044004783,
0.0286553521,
-0.0068956921,
-0.0241009556,
0.0856189206,
-0.0006970181,
0.0038062965,
-0.0258522388,
0.0474717282,
-0.0135592222,
-0.0305742119,
0.0492257848,
-0.0347719155,
0.0059823645,
0.0137588782,
-0.0109324744,
0.0413481444,
0.0054454631,
0.025325289,
-0.0047068982,
0.0118442047,
-0.0019788556,
-0.0012582586,
-0.0201321673,
-0.1487195492,
0.0225269943,
0.0644442663,
0.0456475243,
0.0016565296,
-0.0011261955,
-0.0520243384,
0.0141963568,
-0.0141492803,
0.007069834,
-0.0091168676,
0.0079563269,
0.0158050545,
0.025984183,
-0.0111437887,
0.0049196174,
0.0410702676,
-0.0027602764,
0.0274856072,
-0.060086336,
-0.0281595755,
0.0020845907,
-0.0467349254,
-0.0391318612,
-0.0004673435,
0.0038133976,
-0.0443680696,
-0.1000243127,
-0.0244169012,
-0.0465685278,
-0.0287633836,
-0.009799636,
-0.0212925915,
-0.0003738755,
0.0736113563,
0.029369032,
-0.0084962407,
-0.0211627856,
0.0174384601,
0.0080051925,
-0.0205401927,
-0.002884669,
-5.547788568e-33,
0.0071102753,
-0.0806256756,
0.0185653064,
0.0013243489,
0.0065550967,
-0.0168807711,
-0.0140735377,
0.036681883,
0.0158701017,
0.0407829806,
-0.0533468798,
0.0301900618,
0.0247350484,
-0.0330132507,
0.0795163065,
0.0260146502,
0.0185358617,
-0.0250319541,
0.0029560863,
-0.0392845906,
-0.0859778225,
-0.0056210449,
0.0365306921,
-0.0673367456,
0.0094942488,
0.0481367223,
-0.0352229252,
-0.0194020849,
-0.0639639348,
0.0189835876,
0.0070997765,
0.0323210955,
0.0094263507,
-0.045806583,
0.0133036822,
-0.0221655294,
-0.0052045649,
0.0436167493,
0.0012957597,
-0.0189130586,
0.0325612091,
-0.013290491,
0.045225367,
-0.0068359706,
-0.0962830484,
-0.008060948,
-0.0081950463,
-0.0538586564,
-0.0362914875,
-0.0516398512,
-0.0534905121,
0.0499182194,
-0.0238179173,
0.0531310365,
0.0449674502,
-0.006683744,
-0.0237909928,
0.0256429501,
-0.0690006763,
0.000782059,
0.0092802634,
-0.0383658223,
0.032276053,
-0.0920876041,
-0.0023824195,
0.0306869913,
0.0046598841,
0.0452817827,
-0.0427722968,
-0.0508211516,
0.0113742771,
0.1032472029,
0.0154536981,
0.0203516576,
-0.0161848497,
-0.0252259187,
-0.0383914113,
0.0093771107,
0.0468435511,
0.0375468023,
0.0240563024,
-0.0146727767,
-0.0178011023,
-0.0244124718,
-0.0315906331,
-0.0577655472,
0.0064596129,
-0.005065565,
0.0164568052,
-0.0037641164,
0.0936954618,
-0.0380335152,
-0.0121419271,
-0.0147070205,
-0.0207020584,
-0.0127873048,
-0.0182322487,
0.0553119704,
-0.003867883,
0.008328394,
-0.0745793283,
0.0018650149,
-0.0482512042,
0.0226296112,
0.0156536978,
0.0120678591,
0.0008297919,
-0.0259547886,
-0.0819618255,
0.0286148079,
0.0040240576,
0.0481359847,
0.0586102717,
0.0017108364,
0.0221805293,
-0.0404883437,
0.0321259201,
0.0320616774,
-0.003529476,
-0.0197785012,
-0.0347307771,
0.0234367289,
-0.0214455277,
0.0033099342,
-0.0106501784,
0.0116491821,
0.0147591913,
0.0198359285,
0.0759965628,
0.0301015973,
0.005011417,
0.0207579732,
2.672e-7,
0.0570637919,
0.0532280132,
-0.0325696245,
0.0376686901,
0.0202493146,
0.0568558946,
-0.0051852483,
0.0039000243,
0.0163736232,
-0.0575335436,
0.0403336398,
-0.0281748213,
0.0588081293,
0.032424666,
-0.0455417037,
-0.008298764,
-0.0662051588,
-0.0235586744,
-0.0652401745,
-0.0041800132,
-0.0081533799,
0.0302314311,
0.0161349345,
-0.0300529059,
-0.0057104751,
-0.0170837604,
0.0532166585,
-0.0639366508,
0.0044123242,
-0.0076394859,
0.0880855098,
-0.055895105,
-0.010975372,
0.0522432402,
-0.0118397353,
0.0428908616,
-0.0136790946,
0.0684986487,
-0.0231488999,
0.0701093227,
0.0101993373,
-0.0657612011,
0.0208014976,
-0.0278503951,
0.0183170084,
0.0307071768,
-0.0243971311,
-0.0199405253,
-0.0146366423,
-0.0843075663,
0.0368651785,
0.0353858508,
0.0244910847,
0.0038984732,
0.0093450267,
-0.0292999875,
0.0249235388,
-0.0469181985,
0.0343800113,
0.0923460722,
-0.0346751101,
-0.0107675772,
-0.0212590806,
0.0215839799,
0.06335897,
-0.0169379115,
0.0112968702,
2.021507048e-34,
0.0501540415,
-0.0200988073,
0.042752143,
0.0827144533,
0.0310252178,
0.0645129308,
0.0088528488,
-0.008811608,
-0.0093278326,
-0.0618713722,
0.0154159116
] |
http://arxiv.org/abs/2402.01695v1 | http://arxiv.org/pdf/2402.01695v1 | Language-Guided World Models: A Model-Based Approach to AI Control | 2024-01-24T03:11:36 | 2024-01-24T03:11:36 | Alex Zhang, Khanh Nguyen, Jens Tuyls, Albert Lin, Karthik Narasimhan | cs.CL | cs.CL, cs.AI, cs.LG | Installing probabilistic world models into artificial agents opens an
efficient channel for humans to communicate with and control these agents. In
addition to updating agent policies, humans can modify their internal world
models in order to influence their decisions. The challenge, however, is that
currently existing world models are difficult for humans to adapt because they
lack a natural communication interface. Aimed at addressing this shortcoming,
we develop Language-Guided World Models (LWMs), which can capture environment
dynamics by reading language descriptions. These models enhance agent
communication efficiency, allowing humans to simultaneously alter their
behavior on multiple tasks with concise language feedback. They also enable
agents to self-learn from texts originally written to instruct humans. To
facilitate the development of LWMs, we design a challenging benchmark based on
the game of MESSENGER (Hanjie et al., 2021), requiring compositional
generalization to new language descriptions and environment dynamics. Our
experiments reveal that the current state-of-the-art Transformer architecture
performs poorly on this benchmark, motivating us to design a more robust
architecture. To showcase the practicality of our proposed LWMs, we simulate a
scenario where these models augment the interpretability and safety of an agent
by enabling it to generate and discuss plans with a human before execution. By
effectively incorporating language feedback on the plan, the models boost the
agent performance in the real environment by up to three times without
collecting any interactive experiences in this environment. | Language-Guided World Models: A Model-Based Approach to AI Control: Installing probabilistic world models into artificial agents opens an
efficient channel for humans to communicate with and control these agents. In
addition to updating agent policies, humans can modify their internal world
models in order to influence their decisions. The challenge, however, is that
currently existing world models are difficult for humans to adapt because they
lack a natural communication interface. Aimed at addressing this shortcoming,
we develop Language-Guided World Models (LWMs), which can capture environment
dynamics by reading language descriptions. These models enhance agent
communication efficiency, allowing humans to simultaneously alter their
behavior on multiple tasks with concise language feedback. They also enable
agents to self-learn from texts originally written to instruct humans. To
facilitate the development of LWMs, we design a challenging benchmark based on
the game of MESSENGER (Hanjie et al., 2021), requiring compositional
generalization to new language descriptions and environment dynamics. Our
experiments reveal that the current state-of-the-art Transformer architecture
performs poorly on this benchmark, motivating us to design a more robust
architecture. To showcase the practicality of our proposed LWMs, we simulate a
scenario where these models augment the interpretability and safety of an agent
by enabling it to generate and discuss plans with a human before execution. By
effectively incorporating language feedback on the plan, the models boost the
agent performance in the real environment by up to three times without
collecting any interactive experiences in this environment. | [
0.0615069866,
0.0421916582,
-0.0154167507,
0.0065842662,
-0.0429906808,
-0.0170910731,
0.005224485,
-0.0142554902,
0.0280246269,
-0.0165097974,
-0.0208109524,
-0.0148126381,
0.0237575974,
0.0479475334,
0.0172523931,
-0.0732488707,
0.0485941991,
-0.0599200353,
-0.0391823426,
0.0121052517,
-0.0582411401,
-0.0369922854,
-0.0043071383,
0.061741855,
0.0318878219,
0.0186042301,
-0.0737970844,
0.0220759511,
0.0353519395,
-0.0049156956,
-0.0246572588,
0.0332474075,
0.0366072208,
-0.0203296058,
0.0000024241,
-0.0313869976,
0.0125347916,
-0.0393223129,
0.041879829,
-0.0530246012,
0.0317663141,
0.0037134157,
-0.0209515095,
0.0093441186,
-0.0579383746,
0.0030363838,
0.0629883334,
0.0493609682,
0.048047699,
0.0733329803,
-0.0063222102,
0.0285272337,
0.0010566788,
0.005462666,
-0.0058719204,
-0.0488279313,
0.0238332339,
-0.0081275003,
0.0153479828,
0.0150895966,
-0.0397171304,
0.0093667144,
0.0155178756,
-0.0212488938,
-0.024143327,
0.0528774075,
-0.0062142564,
-0.038855277,
-0.0238906965,
0.0442695059,
-0.0127868112,
-0.0396457799,
0.0229676832,
0.0014802064,
0.0101579539,
0.0231300686,
-0.0223270226,
0.0496493615,
-0.0068463348,
0.0076774387,
-0.0183145925,
0.0360675156,
0.0083762305,
-0.0070095179,
-0.0234334432,
0.1021286324,
0.0025862653,
0.0132733583,
-0.0303928014,
0.0240480732,
0.0473880693,
-0.0921899453,
0.0114483582,
0.0369129628,
-0.0738026053,
0.0114158364,
-0.0828583464,
0.0305980984,
-0.0451278798,
-0.0703542307,
0.0390206687,
0.061427407,
0.0061863912,
0.0468011834,
-0.0008884188,
0.078652449,
-0.0024593128,
0.031927634,
-0.0567762889,
0.0036864951,
-0.0659452006,
0.0020543851,
-0.0461204536,
0.0059770108,
-0.0219802428,
-0.031284295,
-0.0063042878,
0.0588293076,
-0.0361370668,
0.0290771481,
-0.0120999776,
0.0005439912,
-0.0548477583,
-0.0036687304,
-0.0402194932,
-0.030094685,
-0.03235678,
-0.01737931,
0.031246338,
-0.0745681748,
-0.0019559744,
-0.0032463786,
-0.0035671024,
0.02747803,
-0.0129297264,
0.1251356602,
0.0134096686,
-0.0339115523,
-0.009517719,
-0.0079887109,
0.0071462607,
-0.0332395881,
0.0127681885,
-0.0414719991,
0.0353245251,
-0.0091908909,
-0.021382533,
-0.0539687313,
0.0159497857,
0.0019283093,
-0.0048776637,
0.07691852,
-0.0006904475,
0.0250126068,
0.0247912742,
0.0152135342,
-0.0177661255,
0.0716922507,
0.0703498051,
0.0555961281,
0.027372919,
0.0360391289,
0.0504161194,
0.0625297874,
-0.032398209,
0.0133662401,
-0.0028169504,
-0.0211078469,
-0.0062580076,
0.0339211188,
0.0104083987,
0.0341248438,
-0.0286283791,
0.0154011864,
0.0880344957,
0.0415852591,
-0.0011778631,
0.0186680704,
0.0560868606,
0.0507450216,
-0.0134086469,
-0.0099116163,
-0.0527169108,
0.0464152507,
-0.0187626202,
-0.0028911196,
-0.0234178174,
0.0240692608,
-0.0155083574,
-0.0496777184,
0.029122578,
-0.0262962133,
0.0772806779,
0.0105256531,
0.0394861512,
0.0286776517,
-0.0418625921,
0.0433252864,
0.0444393903,
-0.0611068085,
-0.011314854,
-0.063756831,
0.0107744187,
0.1017666012,
0.0396741815,
-0.0269885194,
-0.0265899058,
-0.0016915262,
-0.0067525427,
0.0448593125,
0.0508009456,
0.0166241322,
0.0278260205,
0.0028630584,
-0.0563411005,
0.0017618412,
-0.0071092616,
0.0432089902,
0.0341180153,
-0.0076364209,
-0.0202015564,
0.0122013176,
-0.0362401903,
0.0392785892,
-0.0264352094,
-0.0094458107,
-0.0144168772,
-0.0124274557,
0.0154527482,
-0.008351651,
-0.0216362998,
-0.0050398842,
0.0043428373,
-0.0314670093,
-0.0546602421,
0.0708443522,
-0.0759137496,
0.0130498474,
0.0285866335,
-0.0440593325,
0.0036230288,
0.0240746215,
-0.0016318107,
-0.0383995697,
0.004083808,
-0.0226151012,
0.0478778556,
0.0328247808,
0.0108435098,
-0.0353139825,
-0.0202043075,
-0.0212106891,
0.0029174478,
0.0062194732,
-0.0581501611,
-0.0203029048,
0.0149713298,
0.0073870993,
-0.0146387964,
-0.0059598354,
0.0119934482,
-0.0084864637,
-0.0146726072,
0.0583187379,
0.0251232907,
-0.0447554924,
-0.0447497033,
-0.0881879255,
-0.0276291091,
-0.0286808778,
-0.0383428596,
0.02755199,
0.017730603,
-0.0085135009,
0.0146357613,
-0.0090011312,
0.0489393361,
-0.0144552141,
0.0060174409,
0.0256206784,
0.001583359,
-0.0409161821,
-0.0573775917,
0.0203285944,
-0.0056412588,
0.0232253801,
0.023395488,
-0.0226372667,
-0.095249854,
0.0298471432,
-0.0178049468,
0.038107004,
-0.0563256629,
-0.0721035153,
0.0296168234,
-0.0286594629,
0.0453138426,
0.0584003255,
0.0034355076,
0.0329854563,
0.1082820967,
-0.0260432623,
-0.0546265282,
0.0190798733,
-0.022657685,
-0.0283145979,
-0.0682865307,
-0.036959935,
0.0177794434,
0.0647267178,
-0.0151506951,
-0.0018426562,
-0.0569821037,
0.00942914,
-0.0326331742,
-0.0604076721,
-0.0119658327,
-0.0446725748,
-0.032368958,
0.0480677783,
-0.0104172295,
-0.0262883399,
0.0625638366,
-0.0123728365,
0.0001979618,
-0.0118545936,
-0.0210504867,
-0.0021583347,
0.0599346049,
0.0224033762,
0.0218895506,
0.0230318569,
0.0202034824,
-0.0432009026,
-0.0163490418,
0.0200697873,
-0.012327401,
-0.0012692752,
-0.0890771449,
-0.0012206429,
0.0086640893,
-0.0430086106,
-0.0533612296,
0.0430935025,
-0.0526921526,
0.00490251,
0.020011846,
-0.0524836592,
-0.0458699018,
0.0175491832,
0.0660899356,
0.0488356203,
-0.0212465934,
-0.0474224761,
-0.0339752994,
-0.0570289716,
0.0744011477,
0.0307048727,
-0.0082048774,
-0.040108759,
0.0223426279,
0.0403056182,
0.0190776866,
-0.0145737072,
-0.0396985374,
0.0095526278,
0.0282737855,
-0.0088720368,
-0.0010229619,
0.0554104447,
0.0466397032,
-0.0457721874,
-0.0497642793,
0.0966686532,
0.0164035894,
-0.0308534317,
0.0144161163,
-0.0657507479,
-0.0080428729,
0.0086376537,
0.0524874069,
-0.0210684743,
-0.0250053834,
0.0339411348,
-0.0083686616,
0.0002719557,
-0.0223105028,
0.0200629663,
-0.0563453697,
0.0238742568,
0.0098689971,
-0.0436118692,
-0.0148954745,
-0.0342573151,
-0.0193279088,
-0.0147650568,
-0.0274139661,
-0.0177786518,
0.0387271829,
-0.0221043546,
-0.037149746,
-0.0359409824,
-0.0090874499,
0.0206456054,
-0.0417885259,
0.0357855484,
0.0786493793,
0.0118727172,
-0.0134316934,
-0.0298822187,
-0.0091550248,
0.1035852209,
0.0422394946,
0.004646169,
0.0021546276,
-0.0079659596,
-0.0057787448,
0.007883125,
0.0150026912,
-0.0344077721,
-0.0431146026,
-0.0061898883,
0.0557262599,
-0.0084112464,
-0.0125428261,
-0.015754668,
0.0329579636,
0.0630010441,
0.0620518476,
-0.0033959611,
-0.0709282383,
-0.034780778,
-0.0463673323,
-0.0212478582,
-0.0267620012,
-0.0705788061,
-0.0224310234,
0.0063491054,
-0.0347354002,
-0.0583523698,
-0.0466052443,
0.018005114,
0.0158071145,
0.0223292354,
-0.0597884804,
0.0860514194,
0.0149965649,
-0.0220426302,
-0.0141219227,
-0.0342929028,
0.0090817967,
-0.0032140266,
-0.031835191,
0.0374480002,
-0.0492588654,
-0.0035509604,
0.0509352013,
-0.0518753752,
-0.0632742196,
0.0469874255,
0.0854024515,
0.0210183952,
0.0339667909,
0.004670952,
-0.0111545706,
0.0103093339,
-0.0066307914,
-0.0273487382,
0.040190082,
0.0322309397,
0.0454257131,
0.0152931223,
0.028777808,
-0.0716103315,
-0.0028175337,
0.0179472696,
0.0324105136,
-0.0169317462,
-0.0530594997,
0.0267498922,
-0.0330480896,
0.0121232681,
-0.0196925346,
-0.0120942993,
0.003258249,
-0.0037426781,
-0.0463017337,
-0.0110111488,
-0.0583801866,
0.030587513,
0.0029067865,
-0.0127718151,
0.0598023608,
0.0227014534,
-0.0296513606,
-0.0249176566,
0.0172450989,
-0.0022397961,
0.0161812715,
0.00383454,
0.0252188444,
0.0125296535,
0.0067102737,
-0.0000353566,
0.015619399,
0.0233360585,
0.004987095,
0.0365760028,
0.027932588,
-0.0353683978,
-0.0419916101,
0.0156131312,
-0.0756451786,
-0.0798983574,
0.0280280579,
0.0942526907,
0.0646339953,
0.025723571,
0.022258658,
-0.0793631449,
0.0206876341,
0.0070454618,
0.0082827983,
0.0219580811,
-0.0445111841,
0.0070436043,
0.0180495288,
-0.0031044676,
0.0050176219,
0.0393340997,
0.0028661629,
0.0514166839,
-0.003852932,
-0.0017948742,
-0.0523143783,
-0.0000622416,
0.0210661031,
0.0229827277,
-0.0147325182,
-0.0120664202,
-0.0439164713,
-0.0062830215,
0.0165366717,
0.0490596592,
0.0116227856,
-0.0462256148,
0.0288570914,
0.0416421406,
-0.0106575796,
-0.0489346758,
-0.0309056733,
-0.0179947522,
-0.0092961472,
0.0071413456,
-0.0461572818,
-6.517226473e-33,
-0.0333796926,
-0.0328297578,
0.0381128415,
0.0106478576,
-0.0010118106,
-0.0530791134,
-0.0482403412,
0.0176905841,
-0.0056089936,
0.0441339277,
-0.0248068795,
0.0257412605,
0.0171933845,
0.0070070857,
0.0179735869,
0.0031087608,
0.0407716408,
-0.005700591,
-0.0608702302,
0.0113662034,
0.0469761677,
0.0155743724,
0.0480069779,
-0.0845805928,
0.0017585548,
0.008379356,
-0.0256828591,
-0.0088954708,
0.0155794593,
0.0018541277,
-0.0097099822,
0.0021772673,
0.0055544227,
-0.0245922115,
-0.0309480876,
0.0242386013,
-0.030805625,
-0.0434377417,
-0.0054505114,
0.0178270452,
-0.0514213331,
-0.0329442509,
0.0221328046,
0.0046267756,
-0.0343045965,
-0.0014933618,
0.030373631,
-0.0184721723,
-0.033255022,
0.0242051966,
-0.0809138566,
0.0070682135,
-0.0071229958,
-0.037602827,
-0.013387382,
0.0344224796,
-0.0288371127,
-0.0205356441,
-0.0321878307,
0.0470457301,
0.0304351579,
-0.0059496742,
0.0512161888,
0.0117994761,
0.0303276926,
0.0617191903,
0.0199683029,
0.0367196575,
-0.0403217115,
-0.0317880698,
0.0262542255,
0.0447349772,
0.0093471901,
0.0543713868,
-0.0383942612,
-0.0074032699,
-0.0242442153,
0.0322766304,
-0.0230816863,
0.062659964,
0.0411008373,
0.0320422202,
-0.0338896662,
-0.0131691853,
-0.0282724742,
0.0138537306,
0.0131668868,
-0.0299970135,
0.031748198,
0.013245834,
-0.0174783524,
-0.0611547977,
0.0084731719,
-0.0571682192,
0.0210360028,
0.0056801704,
-0.0317490883,
0.0307824034,
-0.0045689582,
-0.0035434256,
-0.0242189839,
-0.0353883393,
-0.0013412744,
0.0232445803,
0.0082383994,
0.0331904553,
-0.0174503066,
0.0681304634,
-0.0309561808,
-0.0251144469,
-0.0368737094,
-0.0084169535,
0.0224273168,
-0.0254596043,
0.0287478156,
-0.0237053707,
-0.0303931236,
0.0243019424,
-0.0091636246,
-0.0539950505,
-0.0048464872,
0.0073335604,
-0.0082192048,
0.0374117307,
-0.0176814627,
0.0136473458,
-0.0961251557,
0.0534077436,
0.0479162224,
-0.0489547364,
-0.0141940834,
0.0063204519,
3.207e-7,
-0.02241401,
0.0712774321,
0.0102225672,
0.0500209406,
0.0023993438,
0.0624142699,
-0.0103052184,
0.0457905084,
-0.0126364632,
0.0075176884,
0.0280178562,
0.0514275171,
0.0277636442,
0.0059787249,
-0.0642168969,
0.0156617649,
-0.0425785445,
-0.0103754969,
-0.0225451328,
-0.0252950285,
0.0701664612,
0.0472610891,
0.0235847812,
0.0336828344,
0.0354599953,
-0.0240807422,
0.0591230392,
-0.0935667008,
0.0192275885,
0.0231989864,
0.0104254419,
0.0111744599,
-0.0178824048,
-0.0062564565,
-0.0162715316,
0.0110571347,
-0.0446488336,
0.0812587142,
-0.0116278417,
-0.0009357822,
-0.0121684019,
-0.0091902716,
-0.0007289134,
-0.0657233223,
0.0359603874,
-0.0875096172,
0.0028859086,
-0.0020698225,
0.0087514911,
0.0060008252,
-0.0287969336,
-0.00743818,
-0.0260239933,
-0.0140902121,
-0.0018637922,
0.0347579755,
0.0132600842,
-0.0498152785,
-0.0075394032,
-0.0242720805,
-0.0331986398,
-0.0336331949,
-0.046543166,
-0.0331846029,
0.005948789,
-0.0182419531,
-0.0032447758,
2.805718819e-34,
0.0059688939,
0.01035464,
-0.0058506285,
0.00829098,
0.0483914465,
0.0012343986,
-0.0010793756,
0.0431452505,
0.0349521376,
-0.0397343114,
-0.0399672538
] |
http://arxiv.org/abs/2401.13201v3 | http://arxiv.org/pdf/2401.13201v3 | MLLMReID: Multimodal Large Language Model-based Person Re-identification | 2024-01-24T03:07:26 | 2024-06-10T10:21:19 | Shan Yang, Yongfei Zhang | cs.CV | cs.CV, cs.AI, cs.CL | Multimodal large language models (MLLM) have achieved satisfactory results in
many tasks. However, their performance in the task of ReID (ReID) has not been
explored to date. This paper will investigate how to adapt them for the task of
ReID. An intuitive idea is to fine-tune MLLM with ReID image-text datasets, and
then use their visual encoder as a backbone for ReID. However, there still
exist two apparent issues: (1) Designing instructions for ReID, MLLMs may
overfit specific instructions, and designing a variety of instructions will
lead to higher costs. (2) When fine-tuning the visual encoder of a MLLM, it is
not trained synchronously with the ReID task. As a result, the effectiveness of
the visual encoder fine-tuning cannot be directly reflected in the performance
of the ReID task. To address these problems, this paper proposes MLLMReID:
Multimodal Large Language Model-based ReID. Firstly, we proposed Common
Instruction, a simple approach that leverages the essence ability of LLMs to
continue writing, avoiding complex and diverse instruction design. Secondly, we
propose a multi-task learning-based synchronization module to ensure that the
visual encoder of the MLLM is trained synchronously with the ReID task. The
experimental results demonstrate the superiority of our method. | MLLMReID: Multimodal Large Language Model-based Person Re-identification: Multimodal large language models (MLLM) have achieved satisfactory results in
many tasks. However, their performance in the task of ReID (ReID) has not been
explored to date. This paper will investigate how to adapt them for the task of
ReID. An intuitive idea is to fine-tune MLLM with ReID image-text datasets, and
then use their visual encoder as a backbone for ReID. However, there still
exist two apparent issues: (1) Designing instructions for ReID, MLLMs may
overfit specific instructions, and designing a variety of instructions will
lead to higher costs. (2) When fine-tuning the visual encoder of a MLLM, it is
not trained synchronously with the ReID task. As a result, the effectiveness of
the visual encoder fine-tuning cannot be directly reflected in the performance
of the ReID task. To address these problems, this paper proposes MLLMReID:
Multimodal Large Language Model-based ReID. Firstly, we proposed Common
Instruction, a simple approach that leverages the essence ability of LLMs to
continue writing, avoiding complex and diverse instruction design. Secondly, we
propose a multi-task learning-based synchronization module to ensure that the
visual encoder of the MLLM is trained synchronously with the ReID task. The
experimental results demonstrate the superiority of our method. | [
0.0013897239,
0.0685543939,
0.007674546,
0.0746136233,
-0.0435936376,
0.0493356846,
0.0496258028,
-0.005444393,
0.0212444626,
-0.0516607128,
-0.0246914662,
-0.0338310525,
-0.0015154928,
-0.017364366,
0.0284903832,
-0.0220134649,
-0.0010585199,
-0.0341270976,
-0.0149089871,
-0.0003926333,
0.0060932999,
0.0115716178,
-0.0245719496,
0.0228675269,
-0.0353547707,
-0.0121937282,
-0.0086771986,
0.0410266705,
0.0741314963,
-0.0066685271,
0.0040895827,
0.0430671684,
0.0581101887,
-0.0053321244,
0.0000019823,
-0.0238152649,
0.0034439412,
0.00005417,
-0.0095037948,
-0.0493855216,
0.017623635,
-0.0431885831,
0.0131075699,
-0.0173141807,
-0.0000459293,
-0.0261309631,
0.0731988251,
0.0427605286,
0.0526552051,
0.0793939456,
-0.0198744815,
-0.0911218077,
-0.0224001314,
-0.0217103604,
0.0208765715,
-0.042317491,
0.0600322261,
0.036478363,
-0.0421276875,
-0.0377816446,
-0.032015834,
0.0350048319,
0.0046361377,
-0.0049653924,
0.0262462646,
-0.0042827595,
-0.0030586268,
-0.0392198376,
0.0157531444,
0.0336935297,
-0.0581121966,
0.0270382483,
0.008610419,
0.0224292446,
-0.0241449233,
0.0114169726,
-0.0079612043,
0.0243335515,
-0.0144673781,
0.0106117744,
-0.0674770474,
-0.0405992903,
0.0324860252,
0.0004404233,
-0.0281288959,
-0.0173055716,
0.0406987369,
0.0350501277,
-0.0203877166,
0.0251056068,
-0.0158925522,
-0.0331371687,
-0.0226718299,
0.036616195,
-0.0315399878,
-0.0207499061,
-0.0557230152,
-0.0255328137,
-0.0369599685,
-0.0095784739,
0.0321477279,
0.0142248664,
-0.0296316836,
0.0435451269,
-0.0425222181,
0.0890534371,
-0.0219638664,
-0.0077165598,
-0.0161240697,
-0.0404075533,
-0.0441573076,
0.0134631358,
-0.0059888586,
-0.0211507808,
0.0047637452,
-0.0242855679,
-0.0374011211,
0.0262878444,
-0.0082077114,
0.0035831891,
-0.0494250581,
-0.0328480713,
-0.0287270993,
-0.0236640908,
0.0018983325,
0.0579850487,
0.0519420989,
0.0054239542,
-0.053197708,
-0.0324055105,
0.001922804,
0.039445933,
0.0274962168,
-0.0298189614,
0.0107093109,
0.062728025,
-0.0416943841,
-0.0399544984,
-0.0364708565,
0.035805136,
-0.0175202992,
-0.0762212574,
0.0242238026,
-0.0678047314,
0.0051126974,
-0.0106962733,
0.0120291337,
-0.0420449488,
-0.0099605285,
-0.0157707296,
0.0185985472,
0.1039465889,
0.0069340663,
0.0023779373,
0.0402849242,
-0.0030112241,
0.0289187562,
0.0489716046,
0.0189023521,
0.0246663373,
0.0080955504,
-0.0234759338,
0.0192788281,
0.069166407,
-0.0361170135,
0.017943956,
-0.0057457266,
-0.0238522012,
0.0603849441,
0.0516833328,
0.0227465574,
0.0341056101,
-0.0223223064,
-0.0233725514,
0.0567732714,
0.0207232442,
0.0174251311,
-0.0061214049,
0.0404818989,
-0.0136469956,
-0.008592899,
-0.0177859273,
-0.0453042537,
0.0565457195,
-0.0002477826,
-0.0266366601,
-0.0313578472,
0.0789220631,
-0.0668689162,
0.0290475637,
0.0214628689,
-0.0602144599,
0.0465184785,
0.0168888234,
0.043226324,
-0.0188496802,
-0.0190171022,
-0.0057074428,
0.0083015244,
-0.0220093261,
0.0094865113,
-0.0405378416,
-0.0262786187,
0.0626168922,
0.0049710362,
0.0153638618,
-0.0312667266,
-0.0380273052,
0.0275072157,
0.0255206954,
-0.036338184,
0.0005152606,
0.0395967364,
-0.090854004,
-0.0549530685,
0.0269028768,
-0.0342414603,
0.0148534179,
0.0079813898,
0.0168299619,
0.0213362202,
0.0329279751,
0.032086961,
0.0407040492,
-0.0185238775,
-0.0042698216,
-0.0191195961,
0.0318338685,
-0.0145440111,
-0.0032029767,
-0.0169694405,
-0.061803937,
0.0350583084,
-0.0259912256,
0.0057975571,
-0.0165845882,
0.0171199478,
-0.0695385709,
0.0350607447,
0.0013146059,
-0.0235142633,
0.0351892896,
-0.0345535241,
-0.0426957496,
-0.0262491927,
-0.0524388105,
-0.0140307499,
-0.0349005833,
-0.0015907102,
0.0239500292,
-0.0071459743,
-0.0330002233,
0.0418669321,
0.0013034235,
0.0022002973,
0.0293175988,
0.0077413372,
-0.0003446918,
-0.088362053,
-0.0459973328,
-0.0263931025,
0.0326793231,
0.040492475,
0.0054221931,
0.0437635593,
-0.0035728069,
-0.0517854057,
-0.0554697216,
-0.0270624422,
-0.0275795087,
-0.0150718959,
-0.0217512213,
-0.0017362335,
-0.0004622193,
-0.0659362897,
-0.0279982109,
0.0349793173,
-0.0094268424,
0.0252858456,
-0.0087723881,
0.0075175618,
-0.0686916858,
0.0288261231,
-0.0036409029,
0.0697602332,
0.0463576578,
-0.0090676863,
-0.0346938334,
-0.1041610241,
-0.0235269014,
-0.0374257974,
0.0176669713,
-0.0280595105,
0.0135786021,
0.0050674616,
0.0215511117,
0.0665031672,
0.0507400818,
0.0089252209,
-0.0158486664,
-0.0076952735,
0.0178795382,
-0.0145266084,
-0.0118590416,
-0.0045433156,
-0.0457796119,
-0.0165517703,
-0.0257914513,
-0.0429165959,
-0.0202100296,
-0.0274906307,
0.0202115122,
-0.0145651624,
-0.0124366265,
0.0311921909,
-0.0417396091,
-0.0806264132,
-0.008350715,
-0.014451012,
-0.0340385325,
-0.0290566739,
-0.0567869581,
-0.008831392,
-0.0300247967,
-0.0228008721,
0.0019844368,
-0.031382475,
-0.0198668391,
0.0713946968,
0.0539856106,
0.0409383848,
0.0099570649,
-0.0319902822,
-0.0308615137,
-0.0065408824,
-0.0069572753,
-0.0237538591,
0.0530467667,
-0.0201743636,
0.0219398979,
-0.0398957022,
-0.0359679498,
-0.0088882875,
-0.028738359,
-0.009403469,
-0.0205602013,
-0.0414563455,
-0.0277631152,
-0.0077589015,
0.0039292537,
0.0464787297,
-0.0070072999,
-0.016533101,
0.0297126602,
-0.0365332589,
0.0617313199,
0.0386234708,
0.0372370183,
-0.0408486538,
-0.0698775351,
-0.0256070998,
0.0177704524,
-0.0020744335,
0.003558465,
-0.0857904777,
0.0070779901,
0.0093610119,
0.0704015717,
-0.0112395203,
0.0146328006,
0.06552995,
-0.025685275,
0.023593545,
0.0689794719,
-0.0357100666,
-0.0648093969,
-0.0503149852,
0.0236486699,
0.0310503561,
-0.0110546676,
-0.0049778805,
-0.0538087003,
-0.0459482633,
0.0257574134,
0.0024594623,
0.0397120975,
-0.0077190427,
-0.0219557546,
-0.0357759781,
0.0585675053,
0.0195026807,
-0.0386303663,
-0.0556239039,
-0.019738432,
-0.0018669183,
-0.0031411611,
0.0298698153,
0.0415813178,
-0.0199164525,
-0.0048790411,
0.0159398485,
0.0313567258,
-0.0338887535,
-0.0232970249,
-0.0573319532,
0.036491584,
0.0456950366,
0.0425004847,
0.00931942,
-0.0521162748,
0.0296713673,
-0.0142913638,
0.0170795657,
-0.010584048,
0.0451955833,
-0.0298805274,
-0.0628324077,
-0.0344949812,
0.0403231718,
0.0080735693,
-0.0308344197,
-0.0031537742,
0.0172895808,
-0.0517472215,
0.0313308835,
0.0108197387,
-0.019690536,
0.0453484133,
0.0369743109,
-0.0091064116,
-0.04103484,
0.0001743017,
0.0320292786,
0.011113245,
-0.0003167811,
-0.0152143175,
0.0052045891,
-0.0688799843,
0.0233279895,
-0.019007314,
0.0073410599,
0.0104959626,
0.0545788743,
-0.0158651564,
0.0156400241,
0.0019527121,
-0.0024865088,
-0.0749998316,
0.0651633143,
-0.0706747398,
0.0436984003,
-0.0116836196,
0.0025160222,
0.03423528,
-0.0026622708,
0.0435092226,
-0.0219337903,
-0.1151735783,
-0.0127579225,
-0.0315385982,
0.0333149061,
0.0128161721,
0.0495732538,
0.0364664309,
-0.0884627253,
-0.0304611735,
0.0022331297,
0.0276725963,
0.0578462295,
-0.0117179444,
0.0106845461,
-0.0382163227,
0.017685812,
-0.0487764589,
0.012699238,
0.034314625,
-0.0291020703,
-0.0261135492,
-0.058410015,
-0.0419033356,
0.0219621304,
0.0092410613,
-0.0341040194,
0.0341809131,
-0.0151854362,
0.0118662426,
0.0013077644,
0.0076971836,
-0.0369450822,
-0.0158861652,
0.0274501164,
0.0512466207,
0.0305925347,
0.053600397,
0.0578288399,
0.0195993576,
-0.0734807402,
0.0197602119,
-0.0246790219,
-0.005468139,
0.0268628057,
0.0413157828,
-0.0139814327,
0.0436737426,
0.0247060619,
0.0051141828,
0.0117343524,
0.0518861264,
0.0399377309,
0.0085242456,
-0.0449764058,
0.0325915627,
0.0113467472,
-0.0505888835,
-0.0252556968,
0.0098538557,
0.0579000413,
0.0476878285,
0.0563845821,
0.0025676591,
-0.0123108216,
-0.0574537851,
0.0615587123,
-0.009771226,
-0.0297238342,
0.0370392203,
-0.0114380559,
-0.029856829,
0.0319153145,
0.0730539337,
0.013238078,
-0.0040776925,
-0.0358277857,
0.0191263072,
-0.0383610055,
-0.0343620852,
0.0368547104,
-0.0302249324,
-0.0105008977,
-0.0249493048,
0.0535390936,
-0.0204912052,
-0.0008992251,
-0.0298083536,
-0.0214518309,
0.0276857764,
0.02884713,
0.0002127779,
0.0155960796,
-0.0520500131,
-0.0112784188,
-0.012914787,
0.0408188924,
-0.0322416835,
0.0356980711,
-5.753971746e-33,
-0.0342821404,
-0.0193078015,
0.0580535121,
-0.0238389578,
-0.0451681763,
0.0075044613,
0.0108537078,
0.0012684818,
0.0066233072,
-0.0249064211,
0.0459755957,
-0.0150450859,
0.0269989781,
0.0510723554,
-0.0270242747,
0.0243753269,
0.0455062911,
-0.0115348743,
0.0220322423,
0.0141109591,
0.0391949788,
0.0264316089,
0.0124289505,
-0.0537553802,
-0.042782329,
-0.036273133,
-0.0528979711,
0.0336043648,
-0.0015479429,
0.0075272927,
-0.0230385363,
-0.0115273744,
0.0308816973,
0.0024123441,
-0.0368942656,
0.0781428292,
-0.039532762,
0.0313336886,
-0.0409370586,
0.0478489548,
-0.0325884782,
-0.0320598148,
0.0623091422,
-0.0241807941,
-0.0048012096,
0.0069447113,
0.0364263691,
-0.0411124378,
0.0021021694,
0.0146426782,
-0.0215589199,
0.0297125727,
0.0070921774,
-0.0229302067,
0.0403795503,
0.0136815887,
-0.0063062594,
-0.0571724735,
-0.0872733593,
-0.0139798308,
-0.0376472101,
0.0224322621,
0.0402070805,
0.0992203355,
-0.001527306,
0.0662740991,
0.08356902,
-0.0142368292,
-0.0199914556,
0.0305165518,
0.0311488993,
0.0722780898,
0.0397548862,
0.1022143513,
-0.0255706534,
0.0051452187,
-0.0853827372,
0.0292021409,
-0.0161210187,
0.0128453393,
0.0471586622,
0.0613519736,
0.0234615915,
-0.0017619611,
-0.0211804286,
0.03018043,
-0.0143816397,
-0.0077736508,
0.0486437269,
0.0285098925,
0.0020544557,
-0.0142742256,
0.0937245935,
-0.0301522538,
0.0067024631,
0.0055835452,
-0.035439305,
0.0579623543,
-0.004202052,
0.0119178612,
-0.0242647901,
-0.0744267404,
0.0786423609,
-0.0047245855,
0.0188553855,
0.0134587744,
0.0285300631,
0.0625342652,
-0.0563580878,
-0.0044734101,
-0.0220564809,
0.0237593278,
0.0261618011,
0.0339479037,
0.0033082196,
-0.0205717627,
0.0101288659,
0.0499445312,
0.032954827,
-0.0251011886,
0.0107977064,
-0.0306119118,
0.0213939492,
-0.0062823459,
0.0070171263,
0.0622998588,
-0.0073379916,
0.0053594913,
0.0359626263,
-0.0305147655,
0.0125554344,
0.014991207,
2.678e-7,
-0.0058874907,
-0.0182646867,
0.0325660966,
-0.0226496868,
0.040958304,
0.0492866226,
-0.0208429173,
0.0546276048,
0.0330215842,
-0.0334294289,
0.0084395846,
0.0304394644,
0.0447290652,
-0.0036841873,
-0.0325236209,
0.02600611,
0.0000628559,
-0.0011580338,
-0.0710051432,
0.0115211084,
-0.007773452,
0.0471994802,
0.0188197549,
-0.0351585634,
0.0399195291,
-0.0496960878,
0.0873578861,
-0.0283525418,
-0.0402469151,
-0.018131258,
0.0182832815,
0.0417679027,
0.012984788,
0.042293109,
0.0224161651,
0.0009537295,
0.0409183428,
0.0534605868,
-0.0413822196,
-0.0085575785,
-0.0089350427,
0.0438473262,
-0.056439396,
-0.0178237408,
0.0725816116,
-0.018071387,
-0.0221786965,
-0.0360593386,
-0.0152300326,
-0.0136902258,
0.0241137967,
0.000013559,
-0.0091643259,
0.043665152,
0.0503445119,
-0.0131450091,
-0.0607604086,
-0.0915725827,
0.0007705708,
0.0236670505,
-0.0982911065,
-0.0272653494,
-0.0494699143,
-0.0361531004,
0.0216143224,
-0.043964453,
-0.0274539795,
1.757675377e-34,
-0.0227016248,
-0.0135199558,
0.0220700949,
0.0274607986,
0.0138973799,
-0.0093217082,
-0.0851579532,
0.0247306377,
-0.0427490138,
-0.0874392539,
-0.0100662736
] |
http://arxiv.org/abs/2401.13178v1 | http://arxiv.org/pdf/2401.13178v1 | AgentBoard: An Analytical Evaluation Board of Multi-turn LLM Agents | 2024-01-24T01:51:00 | 2024-01-24T01:51:00 | Chang Ma, Junlei Zhang, Zhihao Zhu, Cheng Yang, Yujiu Yang, Yaohui Jin, Zhenzhong Lan, Lingpeng Kong, Junxian He | cs.CL | cs.CL, cs.AI, cs.LG | Evaluating large language models (LLMs) as general-purpose agents is
essential for understanding their capabilities and facilitating their
integration into practical applications. However, the evaluation process
presents substantial challenges. A primary obstacle is the benchmarking of
agent performance across diverse scenarios within a unified framework,
especially in maintaining partially-observable environments and ensuring
multi-round interactions. Moreover, current evaluation frameworks mostly focus
on the final success rate, revealing few insights during the process and
failing to provide a deep understanding of the model abilities. To address
these challenges, we introduce AgentBoard, a pioneering comprehensive benchmark
and accompanied open-source evaluation framework tailored to analytical
evaluation of LLM agents. AgentBoard offers a fine-grained progress rate metric
that captures incremental advancements as well as a comprehensive evaluation
toolkit that features easy assessment of agents for multi-faceted analysis
through interactive visualization. This not only sheds light on the
capabilities and limitations of LLM agents but also propels the
interpretability of their performance to the forefront. Ultimately, AgentBoard
serves as a significant step towards demystifying agent behaviors and
accelerating the development of stronger LLM agents. | AgentBoard: An Analytical Evaluation Board of Multi-turn LLM Agents: Evaluating large language models (LLMs) as general-purpose agents is
essential for understanding their capabilities and facilitating their
integration into practical applications. However, the evaluation process
presents substantial challenges. A primary obstacle is the benchmarking of
agent performance across diverse scenarios within a unified framework,
especially in maintaining partially-observable environments and ensuring
multi-round interactions. Moreover, current evaluation frameworks mostly focus
on the final success rate, revealing few insights during the process and
failing to provide a deep understanding of the model abilities. To address
these challenges, we introduce AgentBoard, a pioneering comprehensive benchmark
and accompanied open-source evaluation framework tailored to analytical
evaluation of LLM agents. AgentBoard offers a fine-grained progress rate metric
that captures incremental advancements as well as a comprehensive evaluation
toolkit that features easy assessment of agents for multi-faceted analysis
through interactive visualization. This not only sheds light on the
capabilities and limitations of LLM agents but also propels the
interpretability of their performance to the forefront. Ultimately, AgentBoard
serves as a significant step towards demystifying agent behaviors and
accelerating the development of stronger LLM agents. | [
-0.0075876047,
0.013895873,
-0.0160868447,
0.0460580848,
-0.0659503713,
-0.0052551632,
0.0023499995,
-0.003815884,
0.0002144449,
-0.0146817416,
-0.0434019379,
-0.0239155367,
-0.0157551747,
0.0523774363,
0.0121910088,
-0.0239096992,
0.0440395772,
-0.0552781858,
-0.0153478757,
0.0132517321,
-0.0097374003,
-0.0215315316,
-0.0249467008,
0.0738111734,
0.0129472381,
-0.0191539284,
-0.0374914669,
-0.0208834428,
0.0301803872,
-0.0153683778,
-0.0535690673,
0.032082513,
0.0202819686,
0.0728466809,
0.0000021197,
-0.0610943399,
-0.0088665476,
0.0197251122,
-0.0464878045,
0.0084266989,
0.0683932677,
0.0121818772,
-0.0377047844,
0.0185992215,
-0.0582715236,
-0.0062889666,
0.0416576043,
-0.0177311115,
0.0685203224,
0.0626363531,
0.0110911978,
0.0155370208,
0.0232853685,
0.0250029545,
-0.0134191122,
-0.0513559841,
-0.0144360308,
-0.0184099004,
0.0395800136,
-0.0499903262,
-0.0277501661,
0.0257558655,
0.0379167423,
-0.047526598,
-0.0231126267,
0.0385983475,
-0.0303562712,
-0.034664046,
-0.0313375071,
0.0656535849,
0.0043621375,
0.0063477638,
-0.0053937971,
-0.0135146417,
-0.0051832073,
0.0103842374,
-0.0361925103,
0.0069758841,
0.023141291,
-0.0100192102,
-0.0244740974,
-0.0031901272,
0.0209329799,
-0.0046145027,
-0.0569116175,
0.125560984,
0.012399043,
0.0191493742,
-0.0253310837,
-0.030622229,
0.0374126658,
-0.0790887475,
-0.0073258439,
0.0561344698,
-0.0800552741,
0.0294025205,
-0.0157200042,
0.030123556,
0.0201462265,
-0.0430154763,
0.0312202871,
0.0425301194,
0.0073460429,
0.0608640797,
-0.0097691575,
0.0623892136,
-0.0079867775,
0.0548481271,
-0.0035755311,
-0.0105809895,
-0.0313281603,
0.0132588781,
0.000787802,
0.0181732774,
0.0113187786,
-0.0452271029,
-0.0055825645,
0.0512881279,
-0.0119373146,
0.0534464158,
0.0076884395,
-0.0193262435,
0.0224392209,
0.0042802375,
-0.0439700074,
-0.002913563,
-0.0419761874,
0.0004801786,
0.0301186685,
-0.0940748677,
-0.0062015541,
-0.0031374348,
-0.0142138824,
-0.0353432409,
-0.0224210918,
0.1056167707,
0.0320054702,
-0.0309275445,
0.0013282654,
-0.0070784744,
0.021985637,
-0.0962079167,
-0.0238179117,
-0.0673806146,
0.0356366225,
0.0079723634,
-0.0210637935,
-0.0220755823,
0.0145427994,
-0.0215070453,
0.0191475265,
0.0669307336,
-0.0094209993,
0.0177627522,
-0.0127685769,
0.0254958309,
-0.042345047,
0.0874304548,
0.0535387918,
0.0243680533,
0.047434289,
-0.0064452318,
0.0448946506,
0.0705822781,
-0.0215395633,
0.024617672,
0.0174173992,
0.0213049706,
0.0044985749,
0.0453578681,
-0.0212520324,
0.0101521844,
-0.0280991066,
-0.024250051,
0.0632400438,
0.0434570462,
0.0151471896,
-0.0235132184,
0.0466020741,
0.0082322806,
0.0158774201,
-0.006483444,
-0.0478467643,
0.0531760231,
0.0034205429,
0.0100507606,
-0.0075523239,
0.0425279103,
-0.0131593021,
-0.0370094627,
-0.0092062633,
0.0025233792,
0.0299956258,
-0.0088116778,
0.0523182228,
-0.0017288667,
-0.0496330336,
0.0264596101,
0.0054197041,
-0.0759829506,
-0.0077488129,
-0.0939758196,
0.013159737,
0.0743035376,
0.0791425183,
-0.0290826503,
-0.0564742722,
0.0136028202,
-0.0261552967,
0.0560089462,
0.043878559,
-0.0091010323,
0.0187540334,
-0.010373475,
-0.0420628861,
0.0017536513,
0.019780647,
0.0713942051,
-0.0048684618,
0.0086058443,
-0.0254346542,
-0.0398688428,
-0.0064793779,
0.0354339741,
-0.0041389759,
-0.0112394216,
-0.0436072946,
0.0063709659,
0.0565412976,
0.0211206898,
-0.0181424282,
0.0085766753,
0.0626345798,
0.0112788426,
-0.0168130286,
0.0351581387,
0.0225137975,
-0.0003826288,
0.0396111645,
-0.0023694825,
-0.0057647368,
0.0410585627,
-0.0277162865,
-0.0440183356,
0.0393042713,
-0.0197940301,
0.0788590461,
-0.0006405349,
0.0284442008,
-0.0413605422,
0.0231452957,
0.0031214359,
0.0149611644,
-0.0014394218,
-0.0082249623,
0.0329027399,
0.038092453,
0.0114471335,
-0.0163692888,
-0.0059346464,
0.0433447137,
0.0136404075,
0.0039027107,
0.0187905543,
0.0075064269,
-0.0507202819,
-0.0288050063,
-0.0629430264,
-0.0158017967,
-0.0106042558,
0.029270919,
0.0228858273,
-0.0089950431,
-0.0029565287,
0.0316658095,
0.0074432362,
0.0253876429,
0.0200857092,
0.02288693,
0.0090645896,
0.0164233074,
-0.1234864295,
-0.0753412619,
0.0188968927,
-0.0142048402,
0.0317813307,
-0.0007227455,
-0.0028297368,
-0.0495915785,
0.0115480749,
-0.0267880298,
0.0160382017,
-0.0277417786,
-0.0642714649,
0.0599201322,
-0.0229296666,
-0.0341803469,
0.038412679,
-0.0211343262,
0.0385824963,
0.0703488365,
-0.0621838272,
-0.0920902863,
0.0068392418,
-0.0237654056,
0.0040508597,
-0.0653144196,
-0.023179451,
-0.0203502234,
0.0526969545,
-0.0148990955,
0.0168333463,
-0.0402472951,
-0.0117626172,
-0.00121574,
-0.0770065337,
-0.0058744084,
0.0052730236,
-0.0440331884,
0.0408512391,
-0.0058884011,
-0.0494129732,
0.0655755177,
-0.0559236072,
-0.0299737304,
0.0171854571,
-0.0251801424,
-0.0411576703,
0.0878586024,
0.0002494745,
-0.0066986098,
-0.0403989404,
-0.0089683738,
-0.0218767319,
-0.0196935274,
0.0073912581,
-0.0262895245,
0.0441514179,
-0.0701376051,
0.0026629181,
-0.0112554757,
0.0092565259,
-0.0103542367,
0.0145863472,
-0.0227366798,
0.0050568394,
0.0141584724,
-0.0458379947,
-0.0644004345,
0.0422114879,
0.0474464297,
0.0159683134,
-0.0143979769,
-0.0158291124,
-0.0474327728,
-0.0332597643,
0.0061755306,
0.034057226,
0.0532044247,
-0.0243724417,
0.0613296703,
0.0567981228,
0.0357059538,
-0.0192500204,
-0.035257604,
-0.0128690787,
0.0338372849,
-0.0255206116,
0.0027145683,
0.0373188742,
0.0384825468,
-0.0761710629,
-0.0133475242,
0.1135601327,
0.0291192029,
-0.0412038788,
0.0212499127,
-0.0514564328,
0.0356631204,
0.0309920982,
0.0394764654,
-0.0419562832,
-0.0064010732,
-0.0046296152,
0.0066533615,
-0.0091786822,
-0.0305680204,
-0.0025251659,
-0.092524454,
0.0580419376,
-0.0029511328,
-0.0418548584,
0.0080880793,
-0.0646787882,
0.0138506182,
-0.0202881321,
0.0365810134,
0.016847739,
0.0297884326,
-0.0405853875,
-0.0251995083,
-0.0697653219,
-0.017236542,
0.0526302904,
-0.0382155776,
0.0801174566,
0.0597099327,
0.0349882618,
-0.0155587681,
-0.0404272713,
-0.0027239814,
0.0129853878,
0.0053551765,
-0.0097556403,
0.0244400445,
0.0137782758,
0.0015780542,
0.0125641664,
-0.0294791572,
-0.0338303559,
-0.0114196362,
0.0095888125,
0.0627228096,
-0.037544813,
-0.0045595965,
-0.0184577275,
-0.0154987946,
0.0114031332,
0.0420345552,
0.0242074262,
-0.0299566537,
-0.056870129,
-0.0586772934,
-0.0307786912,
-0.0581590161,
-0.0564733855,
-0.0119674066,
-0.0011922255,
-0.0326442048,
-0.0242919046,
-0.0587250665,
0.0293077659,
0.0101420991,
-0.0082101906,
0.0078217452,
0.0057756254,
0.0102953305,
0.0225134864,
-0.0225263163,
-0.0074142036,
0.0614171438,
-0.0382639579,
-0.0125383195,
0.0251655933,
-0.0308266412,
-0.006634342,
0.0059128613,
-0.0396907553,
-0.0722862184,
0.0478526503,
0.052738212,
0.0323527157,
0.0129657332,
-0.0279223286,
-0.0207388103,
0.0223043226,
-0.0042726193,
-0.0469294786,
0.0207464732,
0.0192757137,
0.0238534492,
-0.0009964508,
0.0263452101,
-0.0514751486,
-0.0074363821,
-0.0175587907,
0.0407162048,
-0.0510673486,
-0.0407683626,
-0.0021712519,
-0.0131552638,
0.0034849849,
0.0028875135,
-0.0160149001,
-0.0017957146,
0.0251493007,
-0.0544325486,
0.0113572814,
-0.0507765822,
-0.011106791,
0.0020046534,
-0.0103507247,
0.040092323,
0.0336206555,
-0.0347904824,
-0.0041922983,
-0.0054286667,
0.0447213836,
0.0062321923,
-0.0192356426,
0.0276528373,
0.0136572188,
-0.0070027416,
0.00780275,
-0.0278865062,
0.0330011174,
0.0217841882,
0.0360686779,
-0.002228162,
-0.0216670446,
0.0100364126,
0.0578900799,
-0.0856839418,
-0.0446377657,
0.0118266065,
0.0444665253,
0.0711552873,
0.0581540242,
0.0846126229,
-0.0349692106,
0.0222323723,
0.0002539065,
0.0057745641,
-0.0043917643,
-0.0362429433,
0.0119723277,
0.0449636504,
0.004796287,
-0.008792094,
0.0399499498,
-0.0057266941,
0.0680641755,
-0.0311935619,
0.0139961056,
-0.0404559933,
-0.0598911308,
-0.0049456544,
-0.0102195535,
-0.030452298,
-0.0455739312,
-0.0790022314,
-0.0353848115,
-0.0012759205,
0.0821001306,
-0.0146471011,
-0.0387191102,
0.0323360302,
0.0324534737,
0.0257425662,
-0.0123745091,
-0.046593789,
-0.0479743034,
-0.0155680403,
0.0089934729,
0.0154940682,
-6.161287151e-33,
-0.0674018264,
-0.0395745412,
0.0675375238,
0.0092179617,
-0.0360576734,
-0.0263167955,
-0.0596220046,
0.006471294,
-0.016862439,
-0.0065884474,
-0.0097954413,
0.0076010344,
0.0185962487,
-0.0325415134,
0.0072806561,
-0.009819896,
0.0161358081,
0.0076790717,
-0.0458874591,
-0.004245712,
0.0099918274,
-0.0022139498,
0.0390067399,
-0.0755144209,
0.0387423337,
0.0059524793,
-0.0444930047,
0.0180655476,
-0.0599227063,
0.0657115728,
-0.0018646385,
0.0126487724,
-0.0050093574,
-0.0304472018,
-0.0300205909,
-0.0309311617,
0.0054742782,
-0.0113113746,
0.0103282342,
0.0402109697,
-0.0490276404,
-0.0015186657,
0.0015090897,
0.0155538348,
-0.0503943227,
0.0096997172,
0.0187647864,
-0.0146817444,
-0.0196678191,
0.05382891,
-0.0495962501,
0.0379837789,
-0.0145553574,
-0.0159364771,
0.0174641684,
-0.0079468992,
-0.0014675887,
-0.0228396393,
-0.0741736665,
0.0321110226,
-0.0075106812,
-0.0078827692,
0.1045613438,
0.0356783867,
0.0310448371,
0.0597858354,
0.0478915386,
0.0084194876,
-0.0074387211,
-0.0247359611,
0.0328669585,
0.0649032593,
0.0373061933,
0.050537549,
-0.0396647863,
0.0069505004,
-0.00729506,
0.0515892766,
0.0674950927,
0.0550635569,
0.0078351619,
0.0298781656,
-0.0212643221,
0.0037775745,
-0.0350172594,
0.0281289574,
-0.0157396104,
-0.046167478,
0.0351060852,
-0.0083954781,
0.0078374734,
-0.0625504106,
0.0469747372,
-0.0270406809,
-0.0199584849,
-0.0066070752,
-0.0133380508,
0.0624816343,
0.0138410395,
0.0111796539,
-0.0243929774,
-0.0134158609,
-0.0099322824,
0.0142558264,
0.0456679016,
-0.0205959678,
-0.0270607807,
0.0732555166,
-0.0243279282,
0.0093686441,
-0.0502238497,
0.00044825,
0.0159013793,
-0.0215230994,
0.0149253635,
-0.0623346157,
-0.0157833826,
0.0583246909,
-0.0163107533,
-0.0271111801,
-0.0029501806,
0.0339730307,
0.0114023006,
-0.0028390479,
-0.0514268167,
-0.019592721,
-0.0749610588,
0.0321941972,
0.0417824127,
-0.0205465835,
-0.0114926919,
-0.0032756699,
2.817e-7,
0.0329457261,
0.0498171821,
0.0110493554,
0.0619421676,
0.0156610534,
0.0343233272,
-0.0623573251,
0.0071272971,
-0.0036140098,
-0.0080028567,
0.0193773638,
0.0339619704,
0.0444169827,
0.0495043583,
-0.0436391905,
-0.02212069,
-0.0372246057,
-0.0123257171,
-0.0452744067,
-0.0126129091,
0.0496385358,
0.0607687086,
0.0299748108,
0.0194960628,
0.0079975231,
-0.0407675169,
0.0140162194,
-0.0540852174,
0.0151712336,
0.0294355359,
-0.0407942496,
0.0131080346,
-0.0057745986,
-0.0040470459,
0.0147985807,
0.0266086534,
-0.0381440371,
0.0787777901,
-0.0098229805,
-0.0530839041,
0.0206628181,
-0.0411611162,
0.0034151683,
-0.0421047471,
0.0543361902,
-0.021170374,
-0.021897085,
-0.006031902,
-0.0093072122,
-0.0101143494,
-0.0015749189,
-0.014193275,
-0.0231962707,
0.0325095356,
-0.0098342877,
-0.0434402712,
-0.0082031516,
-0.0492213033,
-0.0252131019,
-0.0518904813,
-0.0461529717,
-0.0774333477,
-0.0364301912,
0.0054904134,
-0.0133457053,
-0.042378705,
-0.0182747934,
2.71408031e-34,
0.0347781554,
0.0095543982,
0.0148694478,
0.0477143377,
0.0577662326,
0.0117033962,
-0.0041411854,
0.0075416267,
0.013379856,
-0.0643335432,
-0.0337391086
] |
http://arxiv.org/abs/2401.13170v4 | http://arxiv.org/pdf/2401.13170v4 | CFMatch: Aligning Automated Answer Equivalence Evaluation with Expert Judgments For Open-Domain Question Answering | 2024-01-24T01:30:25 | 2024-06-30T02:09:46 | Zongxia Li, Ishani Mondal, Yijun Liang, Huy Nghiem, Jordan Boyd-Graber | cs.CL | cs.CL | Question answering (QA) can only make progress if we know if an answer is
correct, but for many of the most challenging and interesting QA examples,
current evaluation metrics to determine answer equivalence (AE) often do not
align with human judgments, particularly more verbose, free-form answers from
large language models (LLM). There are two challenges: a lack of data and that
models are too big: LLM-based scorers can correlate better with human judges,
but this task has only been tested on limited QA datasets, and even when
available, update of the model is limited because LLMs are large and often
expensive. We rectify both of these issues by providing clear and consistent
guidelines for evaluating AE in machine QA adopted from professional human QA
contests. We also introduce a combination of standard evaluation and a more
efficient, robust, and lightweight discriminate AE classifier-based matching
method (CFMatch, smaller than 1 MB), trained and validated to more accurately
evaluate answer correctness in accordance with adopted expert AE rules that are
more aligned with human judgments. | CFMatch: Aligning Automated Answer Equivalence Evaluation with Expert Judgments For Open-Domain Question Answering: Question answering (QA) can only make progress if we know if an answer is
correct, but for many of the most challenging and interesting QA examples,
current evaluation metrics to determine answer equivalence (AE) often do not
align with human judgments, particularly more verbose, free-form answers from
large language models (LLM). There are two challenges: a lack of data and that
models are too big: LLM-based scorers can correlate better with human judges,
but this task has only been tested on limited QA datasets, and even when
available, update of the model is limited because LLMs are large and often
expensive. We rectify both of these issues by providing clear and consistent
guidelines for evaluating AE in machine QA adopted from professional human QA
contests. We also introduce a combination of standard evaluation and a more
efficient, robust, and lightweight discriminate AE classifier-based matching
method (CFMatch, smaller than 1 MB), trained and validated to more accurately
evaluate answer correctness in accordance with adopted expert AE rules that are
more aligned with human judgments. | [
0.0420360416,
0.0668857172,
0.0160599239,
0.0377545953,
-0.0422769822,
-0.0090857167,
-0.0206500031,
0.0061189998,
0.0061140978,
-0.0491941012,
-0.0421393067,
-0.0160114337,
-0.0252620857,
0.0111072827,
-0.0110139465,
0.0052756602,
0.0171600077,
-0.0339026004,
0.0160113573,
-0.0241119117,
0.0225971844,
0.0217273384,
-0.0089607192,
0.0272386931,
0.0383767895,
-0.0302981287,
0.0323030092,
-0.0172255933,
0.0211066306,
0.0030202623,
0.0445133671,
0.0822737962,
-0.0104464702,
0.0174011644,
0.0000019217,
-0.0629656315,
0.0187944006,
-0.0013992613,
-0.0322929025,
0.0413690247,
-0.0304755848,
0.0440178961,
0.0339631252,
0.0282725673,
-0.0683193207,
-0.0054776259,
0.0306179598,
0.0072304029,
0.0586539395,
0.0346381702,
0.0142768323,
-0.0090979813,
0.0218806416,
0.0086879591,
-0.0444923416,
-0.063123554,
0.0334894247,
0.0436169207,
-0.008240181,
-0.0144969337,
0.0037048589,
0.011244826,
0.0087562986,
-0.0067779096,
0.0236588959,
0.0232486445,
-0.013292579,
0.0278295279,
0.0009458209,
0.0810999572,
0.0063694627,
-0.0696105361,
-0.0197967719,
0.0026190525,
-0.0080088703,
0.0295167025,
-0.0257744137,
0.0092778951,
-0.0203750301,
0.0126552163,
0.0235562008,
0.0543464459,
-0.0144264633,
0.0485173985,
0.0002444204,
0.0766018927,
0.02633797,
-0.0375670567,
0.0052775061,
0.0069795465,
-0.0469717905,
-0.0795317665,
0.0255941637,
0.035088744,
-0.0296561122,
-0.0284036826,
-0.0413144529,
-0.0245033074,
0.0500182994,
-0.0514820777,
0.0275770556,
0.0193721484,
-0.0060372315,
0.0692572594,
-0.0425566845,
0.0654055327,
-0.0042501744,
0.0294419695,
-0.0821799412,
-0.0422351807,
-0.0349491909,
-0.0092229862,
-0.0000106472,
0.0098430114,
-0.0109463101,
0.0042111916,
-0.0428926982,
-0.0219993852,
-0.0551288947,
0.0048912866,
0.0166029111,
0.0503436811,
0.03488582,
0.0143012665,
-0.0711655468,
0.0249340013,
-0.0767491981,
0.0434905104,
-0.0216857921,
-0.058531411,
-0.0066745412,
0.0196984876,
0.0091182245,
-0.020997528,
0.0489818789,
0.0402402095,
-0.0244568288,
0.0266625714,
0.043350935,
0.0052996869,
0.0343817174,
-0.0337842368,
-0.0194007121,
-0.0801144913,
0.0476721153,
0.0069312672,
-0.0113487206,
-0.0302582141,
0.0029581871,
0.0445144065,
-0.0025372826,
0.0457761772,
-0.0191518888,
0.0247030091,
0.0471070036,
-0.0120194135,
-0.012061324,
0.0003522473,
0.0009612439,
-0.0275072455,
0.0568651706,
-0.0239443015,
0.0540758632,
0.02354577,
0.0020337864,
-0.0098402463,
0.0138359144,
0.1077652723,
0.0459596626,
0.0428439043,
0.0522767566,
0.0425695032,
-0.009303581,
-0.0787212029,
0.021716306,
0.0727619156,
-0.0216185562,
-0.0126664266,
0.0154838739,
0.0035248597,
0.0490359254,
0.0103921192,
-0.0268367063,
0.0029494942,
-0.0267083161,
-0.0160727091,
0.0174747221,
0.0390234068,
-0.0222292505,
-0.0198758114,
-0.0190956239,
0.0231005158,
-0.0044705383,
0.0454018936,
-0.0076560639,
-0.0271502975,
0.0196486358,
0.0484130308,
0.0122236526,
-0.0414103083,
0.0037112497,
-0.0669757426,
0.0358415283,
0.0478570387,
0.0013053027,
-0.0194450747,
-0.0103739351,
-0.001264867,
0.0161437858,
0.0205520727,
0.0515706725,
-0.0219642501,
-0.010199843,
-0.0249118656,
0.012590833,
-0.0015558567,
-0.0194627959,
-0.0064598578,
0.0468892418,
0.0154405721,
0.0063034971,
-0.0568651073,
-0.0383195281,
0.0481659621,
-0.0096629756,
-0.0404977575,
0.0161007196,
-0.0144046741,
0.0282347817,
0.0080826012,
-0.0036141968,
0.0383743197,
0.0003507719,
-0.0282782372,
-0.0392892882,
-0.0280424319,
-0.0121462224,
0.009044175,
0.0536799729,
-0.0306600165,
-0.0454657562,
0.0602665469,
-0.0333325006,
-0.0268353838,
-0.0296303369,
-0.0158695113,
0.0090063857,
-0.0626318902,
0.0108644264,
-0.0034169813,
0.0009396837,
-0.0463345982,
0.0428405516,
0.0125522567,
0.0321854614,
0.0072185025,
-0.0285000093,
0.0222911183,
-0.0211271252,
-0.0224212315,
-0.0181354862,
-0.011441363,
0.0274118222,
0.0278460402,
0.0046386095,
-0.0500211194,
-0.0227981694,
-0.0784773007,
0.0249534547,
-0.004128057,
0.0457023866,
0.0355041139,
-0.0363225564,
-0.0158774424,
0.0062103244,
-0.0978644863,
-0.0300343838,
0.0401570462,
0.012750891,
0.0188760944,
-0.0167174079,
-0.0109318541,
-0.0590421408,
-0.025151059,
-0.0233384464,
0.0584037043,
-0.001686101,
0.0371677727,
-0.0525232255,
-0.0016564218,
-0.0574380644,
0.0485238358,
0.0081612766,
-0.0353795812,
-0.0014016416,
0.0135672176,
-0.0197912175,
0.0639735907,
0.05038568,
-0.0477326401,
0.0551983602,
0.0220995639,
-0.0660758913,
-0.0110467495,
-0.0073526981,
-0.0447119847,
-0.0137352105,
-0.0533778779,
0.0373588093,
0.0302324537,
0.0118870465,
-0.0116710262,
-0.0245550778,
0.0381797887,
-0.0029830108,
-0.0227265488,
-0.0496657379,
-0.0261406396,
-0.0028854765,
0.0265128687,
-0.0059218975,
-0.0141828815,
0.0973084867,
-0.0445750616,
0.0267196689,
-0.0348462872,
-0.0353638455,
-0.0578129888,
0.040752925,
-0.0264863484,
0.0189048685,
-0.0134355044,
0.0201738849,
-0.0259305276,
-0.0218356494,
0.0092849201,
-0.0646113604,
0.0253963526,
-0.0461274274,
0.0266857762,
-0.0044999681,
-0.0404785536,
-0.0510707274,
-0.0631821305,
-0.0452475846,
0.0349894837,
-0.0047094058,
-0.0100800041,
-0.0072186324,
0.0367854461,
-0.0057629151,
-0.0094168009,
-0.0108979763,
0.0075864531,
-0.0072782552,
0.0885128826,
0.0639393106,
0.0796550438,
0.0251473356,
0.0001465898,
0.0036111902,
0.0619631223,
0.0309284143,
0.021708848,
0.0031636374,
-0.012155151,
0.0429078154,
-0.0565655679,
0.0118213883,
0.0135764712,
0.0517908558,
-0.0194001235,
0.0202625152,
0.1014535576,
0.0304627549,
-0.0078972252,
-0.0092635881,
-0.0283869207,
-0.0032209381,
0.0146226548,
0.0419095047,
-0.0746243373,
-0.0441162549,
-0.0067205201,
0.0438092053,
-0.0050914558,
-0.0115719577,
0.0226487704,
-0.1079073548,
-0.0147601459,
-0.0222576484,
-0.0409619734,
-0.0534838364,
-0.0539589003,
0.049449116,
-0.012875828,
-0.0336474255,
0.0238328129,
-0.0102918483,
0.0050368989,
0.0164022166,
-0.0012171508,
-0.0028452713,
0.0933048874,
-0.0699856728,
0.0862636492,
0.0374195315,
0.0003952948,
0.00390903,
-0.0702302158,
-0.0128296725,
-0.0642715394,
-0.0289860312,
0.0101363277,
0.0129204756,
0.0058100014,
0.0100782588,
-0.0374799296,
0.035556443,
-0.0081735328,
-0.0040644128,
0.0381277576,
0.065450348,
0.010809972,
-0.0309645124,
-0.0273668133,
0.0162822381,
-0.0312119406,
0.0883508399,
0.0198150855,
-0.0264960714,
-0.059802521,
-0.0059393924,
-0.0065999473,
-0.0158023424,
-0.0555806085,
0.0303098764,
0.0458706506,
0.0100259334,
-0.0655298606,
-0.0140951723,
0.0030728406,
-0.0171075724,
-0.0204282254,
-0.0208664574,
-0.0239467751,
0.0147695486,
-0.0442774743,
0.0679914355,
-0.07455571,
0.0588299707,
-0.0668881759,
0.0079024397,
0.0121940654,
0.0114686769,
-0.0508899875,
0.0303778388,
-0.060869012,
-0.0191996433,
-0.0222224463,
0.029793093,
0.0094169062,
0.0015792946,
-0.0150252553,
-0.0147389667,
0.0072519337,
-0.0030517373,
-0.0301501155,
0.0492925383,
0.0107043684,
-0.0116375322,
-0.0323369838,
0.0097551811,
-0.0291644018,
0.0097193578,
0.0185950287,
0.0611660704,
0.014214986,
-0.1287074387,
0.0382649079,
0.0389017425,
-0.010853949,
0.0466982946,
-0.0116144828,
0.0049883588,
0.0006520702,
0.0208742209,
-0.0106772315,
0.0111981453,
-0.0426750667,
-0.0050844601,
-0.0132973492,
0.0017919552,
0.0074029667,
-0.0342393257,
0.0279524792,
-0.0469723828,
-0.0044691502,
0.0321451575,
0.0178067368,
-0.0270251501,
-0.0302248839,
-0.0571556501,
-0.0356022194,
0.022067843,
0.060854923,
-0.0208894759,
0.0441330932,
0.0024168361,
0.0003371067,
-0.0128262574,
0.0154239545,
-0.0782225505,
-0.0260025449,
-0.050831344,
0.0286091641,
0.0546921194,
-0.0101724509,
0.0061745532,
0.0144533729,
-0.0076982602,
-0.0103098527,
-0.0705779493,
-0.0004492892,
0.020761244,
0.0707818344,
0.0266475473,
-0.0028081103,
0.0110064447,
0.0718871951,
0.0168908834,
0.0518039763,
-0.0146762514,
-0.0357835628,
-0.0007994553,
-0.0427134782,
-0.0167538971,
-0.0210696459,
-0.0013545583,
-0.019673761,
-0.039448373,
-0.0311980434,
0.0223621354,
0.0302360021,
-0.0158600677,
0.0130167799,
0.0236647092,
0.0522670969,
0.0164677575,
-0.0678389221,
-0.0228677411,
0.0343386419,
-0.0089431508,
-0.0146788647,
-0.0766405091,
-6.518620903e-33,
-0.0032111143,
-0.0455366895,
0.0071679875,
-0.0480123125,
-0.0107311914,
-0.0050230762,
0.0112463692,
-0.0434533022,
-0.0488919169,
-0.0004826181,
-0.0139881149,
-0.0072513297,
-0.0005640669,
-0.0010226479,
0.0253441632,
-0.05725329,
0.004278983,
-0.0073654605,
-0.0245836452,
-0.0152699435,
-0.0059842756,
-0.0332201533,
0.0159526579,
-0.0705184489,
0.0407996476,
-0.0159929693,
-0.0186913516,
0.0354063548,
-0.0252898596,
0.0440713651,
-0.0264433622,
0.0456773564,
0.041987855,
-0.0273216348,
0.0346321799,
0.0270148441,
-0.0578078069,
-0.0176972635,
0.0408832282,
-0.0497199968,
-0.0059750746,
-0.0190636516,
0.0253415741,
0.00727899,
-0.038010221,
-0.0057974327,
0.0414061658,
-0.0237659514,
0.0482103191,
-0.0211669672,
-0.0080024069,
0.0244579632,
0.0007202305,
-0.0195691194,
-0.033679422,
0.0618549325,
-0.0103866439,
0.0017986078,
-0.0333700292,
0.0193720199,
0.027833486,
0.0445663594,
0.0619216636,
0.0015969485,
-0.0068705613,
0.0633636191,
0.0196760651,
-0.0355810486,
-0.0257528666,
-0.0219533835,
-0.0231680684,
0.0159150902,
0.1063531414,
0.053192243,
0.0312365294,
0.0307221394,
-0.0367385782,
0.0578034185,
-0.0184005052,
0.0331953615,
-0.007111507,
-0.0096568335,
-0.0243216455,
-0.004724402,
-0.0683886632,
0.0442698635,
-0.0725221336,
0.0183344353,
0.0452132449,
-0.0409906469,
-0.0166114699,
-0.0289010871,
0.0116742468,
-0.0499816462,
0.0031968316,
0.0169674624,
0.0211871061,
0.0203819461,
0.0084312893,
0.0385429002,
-0.0392207131,
0.0094820615,
0.0208517741,
-0.0242549535,
0.0235593431,
0.0180472713,
-0.0148162246,
-0.0365584195,
-0.0401752479,
0.0272725057,
-0.0011983556,
-0.0158629604,
-0.0344972722,
-0.0166388769,
0.0111119803,
0.0472095497,
0.0181817636,
0.0165587589,
0.000064118,
0.0234257188,
-0.0352110527,
-0.0507256724,
0.0173932854,
0.0243504923,
-0.0125376377,
0.0593927838,
-0.0563835278,
0.0213968642,
-0.0700455755,
-0.0168534685,
0.0087961219,
0.0326466523,
2.515e-7,
0.0495106131,
0.0076266495,
0.0165127125,
0.0137683162,
-0.0191805586,
0.0276429392,
0.0289815441,
0.0117526399,
0.0429882035,
-0.0067975679,
0.002241632,
0.0455630273,
0.0611794777,
0.057653144,
-0.0368218869,
-0.0221120808,
-0.1089293286,
-0.0238763075,
-0.0328702219,
-0.0168076456,
0.0887357146,
0.050140813,
0.0065644099,
-0.0061598308,
-0.0373713598,
-0.0525944345,
0.0254607033,
-0.018077882,
0.0057792608,
0.0250713676,
0.0627601519,
-0.0471829921,
0.0437958017,
0.0673428252,
-0.0057852073,
0.0030447112,
0.0541376881,
0.0701592416,
-0.0128807267,
-0.0241182391,
-0.0765700266,
-0.0357571132,
0.0390874967,
-0.0630073622,
0.0810933858,
-0.0217522588,
0.0027010425,
-0.0673985928,
-0.0390839837,
-0.0366908349,
0.0262977853,
-0.0327883027,
-0.0235512964,
0.0426235609,
0.031937249,
0.020489186,
-0.0167903211,
-0.0388168804,
-0.0387965254,
0.0149795674,
-0.0458608568,
-0.0233107414,
-0.0280729868,
0.0004350824,
0.0153856818,
-0.0334748738,
0.0366543308,
2.133467605e-34,
0.013595161,
0.0122082457,
0.026112847,
0.0181900393,
0.037348289,
0.0387228727,
0.0001866293,
-0.0105892476,
0.0197478496,
-0.1114168018,
0.0184808187
] |
http://arxiv.org/abs/2401.13165v3 | http://arxiv.org/pdf/2401.13165v3 | Misgendering and Assuming Gender in Machine Translation when Working with Low-Resource Languages | 2024-01-24T00:58:30 | 2024-03-12T17:41:13 | Sourojit Ghosh, Srishti Chatterjee | cs.CL | cs.CL | This chapter focuses on gender-related errors in machine translation (MT) in
the context of low-resource languages. We begin by explaining what low-resource
languages are, examining the inseparable social and computational factors that
create such linguistic hierarchies. We demonstrate through a case study of our
mother tongue Bengali, a global language spoken by almost 300 million people
but still classified as low-resource, how gender is assumed and inferred in
translations to and from the high(est)-resource English when no such
information is provided in source texts. We discuss the postcolonial and
societal impacts of such errors leading to linguistic erasure and
representational harms, and conclude by discussing potential solutions towards
uplifting languages by providing them more agency in MT conversations. | Misgendering and Assuming Gender in Machine Translation when Working with Low-Resource Languages: This chapter focuses on gender-related errors in machine translation (MT) in
the context of low-resource languages. We begin by explaining what low-resource
languages are, examining the inseparable social and computational factors that
create such linguistic hierarchies. We demonstrate through a case study of our
mother tongue Bengali, a global language spoken by almost 300 million people
but still classified as low-resource, how gender is assumed and inferred in
translations to and from the high(est)-resource English when no such
information is provided in source texts. We discuss the postcolonial and
societal impacts of such errors leading to linguistic erasure and
representational harms, and conclude by discussing potential solutions towards
uplifting languages by providing them more agency in MT conversations. | [
0.0464410558,
0.0384524614,
-0.0473391563,
0.0194357093,
-0.0026620005,
0.0220491998,
0.0242437627,
-0.0277705267,
-0.0081463791,
-0.0118205994,
0.0326173268,
-0.0413866378,
0.0955312997,
-0.0497081392,
0.0294968672,
-0.0474630035,
0.0199474692,
0.0322850496,
-0.0103952605,
-0.0026361656,
0.0208993927,
0.0599232912,
0.0502169281,
0.064278014,
-0.0018395825,
-0.0253224038,
0.0226378944,
0.0335525684,
-0.0271414556,
0.0212397259,
-0.0138591975,
0.0237156022,
-0.0196539667,
0.0593637265,
0.000001858,
-0.0537388027,
0.0049110502,
0.0098602083,
-0.0421177782,
-0.0147150969,
0.0400126502,
-0.0289001781,
-0.0318739787,
0.0119776735,
-0.0137961935,
0.0220572148,
0.072422564,
-0.0190171059,
0.0228508078,
0.0210804474,
0.0015138217,
0.0128724566,
-0.0408630483,
-0.0139819933,
0.0173850227,
-0.0116531998,
0.0189333819,
-0.0139731867,
0.0120947836,
0.0240682159,
-0.0501491502,
0.0619177334,
-0.0080693252,
-0.0494690686,
-0.0529605485,
-0.0228799041,
-0.0151892845,
-0.0774974152,
0.0463676453,
0.031418398,
0.0056197238,
0.0035560287,
-0.0005968125,
0.048989851,
-0.0168710314,
-0.0413009748,
0.0135625983,
0.0404503793,
-0.0354986489,
-0.010249692,
0.040089272,
-0.0209955927,
-0.0027296625,
0.0220394507,
-0.0135041596,
0.020068761,
-0.0140578011,
-0.0151903611,
-0.0137799671,
0.0121836876,
-0.0155479852,
-0.0057921298,
-0.005665794,
-0.0250517353,
-0.0169953462,
0.0279387813,
0.0028266811,
-0.0672188848,
0.0119865704,
-0.0560245886,
0.0117730927,
0.0535354987,
-0.0103470767,
0.0003032616,
0.0093204034,
0.0124722868,
-0.0292015504,
-0.0155932009,
-0.0338083766,
-0.0060051107,
0.0417030305,
-0.0369727202,
-0.0114570856,
-0.0491177179,
-0.0591900311,
0.0673424676,
-0.0374783576,
0.0157897566,
-0.0228869673,
0.0179588571,
-0.0183685403,
0.003166554,
-0.048744671,
-0.0299701914,
0.0103599094,
0.0279945023,
-0.0071037812,
-0.0021524327,
-0.0294288732,
-0.1275140047,
0.0051816888,
0.006754939,
-0.0769145191,
-0.020616604,
-0.0308031458,
0.0339837223,
-0.0107988697,
0.0019800151,
-0.0126487082,
-0.0156072741,
-0.0486424156,
-0.0564391725,
-0.023379907,
-0.0242530573,
0.0305612795,
0.0752636418,
0.0060005989,
-0.0511494949,
0.0180887152,
0.0110902144,
-0.042910859,
0.0565184839,
0.0178775974,
0.00444355,
0.0890072882,
0.0485313796,
-0.0158716384,
0.0456769317,
0.0587728992,
-0.0202092603,
0.0283557195,
-0.0065505705,
-0.004979779,
0.0349217914,
0.0046985815,
-0.0053324145,
-0.0057591307,
0.0319610164,
-0.0348291062,
0.0303063244,
0.0747650713,
0.0212358646,
0.0260345303,
0.0532446355,
0.0091957934,
0.0717331246,
-0.0265289005,
0.1009775475,
0.0152472761,
0.0098936632,
-0.0263344012,
-0.0018650076,
-0.0317994878,
-0.003109976,
0.0051534539,
-0.0469705947,
0.0031198596,
0.0298963916,
0.0114312163,
-0.0317877829,
-0.0051223719,
0.0291246474,
0.0141390208,
-0.0259015094,
0.0178818721,
0.0210580081,
0.0200617053,
0.0432270132,
0.0148641104,
-0.0526949726,
0.0081703961,
-0.0452109389,
-0.020925317,
0.0147275049,
-0.0128932884,
-0.0158711709,
0.0787049085,
-0.0269008968,
0.0005187661,
0.0398988053,
0.0276147407,
0.0105035072,
-0.0029803456,
0.0023758956,
0.0352224149,
0.0892811865,
0.0095512485,
0.0260445289,
-0.019209981,
0.0613031872,
0.0385677144,
-0.0199262835,
-0.1081480682,
0.0098316818,
-0.0027696013,
-0.0223223604,
0.0080331424,
0.0074171438,
0.0059616626,
0.0080212392,
-0.0277134888,
0.0062297834,
-0.0293913968,
-0.0026473468,
0.0048766672,
0.0671390817,
-0.0384239182,
0.0093406737,
-0.029788414,
-0.0865215734,
-0.0383471586,
0.0228829738,
0.0003787714,
-0.0081567476,
0.0818710178,
-0.0429736786,
0.001483916,
-0.013561896,
0.0614933781,
-0.0105975047,
-0.0130770495,
-0.0296882819,
-0.0082624266,
-0.0226216614,
-0.0258613899,
-0.020010408,
-0.0403558239,
0.0459171869,
-0.0502789244,
-0.0382967852,
0.087311618,
0.0503916144,
0.0407997407,
0.0343609042,
0.0211375467,
-0.0239518303,
-0.0126622645,
-0.0401431173,
-0.0508491658,
0.001114076,
-0.0000279853,
-0.0235436596,
0.0628455728,
-0.0109911943,
-0.0109831328,
0.0304976162,
0.0027349207,
0.0102704978,
0.0167489741,
0.0166421235,
-0.0147115262,
0.0713477358,
0.0161947608,
0.0160767175,
-0.0099580958,
0.0543880872,
0.0346019603,
-0.0379558541,
-0.0425145477,
-0.0119967153,
-0.0292039309,
-0.049902875,
-0.0181996189,
-0.0379122272,
-0.002139257,
0.017077202,
0.0022154015,
0.0212609638,
-0.0203217827,
0.0243905876,
0.0104509918,
0.0307080559,
-0.009658548,
-0.0399842337,
0.0471093766,
-0.034204971,
0.0436283275,
-0.0630544573,
-0.0290949196,
0.0166446734,
0.0374893099,
-0.0072560646,
-0.0213056076,
0.0293452796,
-0.0556464903,
-0.0210533906,
-0.0340274945,
-0.0254095104,
-0.0083868653,
0.0490419269,
-0.0507031754,
-0.0264637936,
0.0430795401,
0.0230530966,
-0.0244420413,
-0.0394886136,
0.0246593338,
-0.0003525343,
0.0384706669,
-0.0457191467,
0.0333035924,
-0.0131308055,
-0.0366516523,
-0.0490830764,
-0.0007048982,
0.0313124023,
-0.0266802441,
0.0063585886,
-0.0913634822,
0.0388184004,
-0.0352035463,
-0.0403930135,
-0.0069768983,
-0.0084443558,
-0.0073142191,
-0.0184574872,
0.0045839115,
0.0161270909,
0.0066646156,
0.0319720097,
0.0264993012,
0.0047729975,
-0.0023620198,
0.0018344159,
0.0009869935,
-0.026646968,
-0.0201687608,
-0.0209602565,
0.0652634725,
-0.0275225826,
0.0176384877,
0.0170498118,
-0.0278945751,
-0.0259427484,
0.0542231388,
-0.0060404171,
0.0483984686,
0.0065474319,
-0.0452044494,
-0.0076480727,
0.037160743,
-0.0356073603,
0.0013962247,
0.0561360158,
-0.0100438986,
-0.0346841104,
0.0235473141,
-0.0760963559,
-0.0066025192,
-0.0014352298,
0.0127843553,
-0.07458435,
0.0448221453,
0.0429826379,
0.0267181415,
-0.0432646796,
-0.0432788953,
-0.0512803458,
-0.0858406425,
-0.0032875142,
0.0506865643,
-0.0448022932,
-0.0587635413,
0.0312365964,
-0.0322940983,
0.0231849886,
0.0736005232,
0.0311325882,
0.0058180639,
-0.008906276,
0.0118919909,
-0.0198742487,
-0.0158473104,
-0.0267824419,
-0.0343509391,
0.0122461561,
0.0211942829,
0.0371457748,
-0.0179569591,
0.0526712723,
0.0236193463,
-0.013360884,
-0.0323988348,
-0.0208708867,
-0.0279804133,
-0.0110200131,
-0.0248177182,
-0.0257213283,
0.0003591648,
0.0293609984,
0.0158536136,
0.0307720043,
0.0202926509,
-0.0255163666,
-0.0340979062,
0.0140063958,
-0.0224664621,
0.0362169109,
0.0268274657,
0.0209891889,
-0.0065834168,
-0.0194889642,
-0.0606945679,
-0.0148721579,
0.0242607277,
-0.0886233002,
0.0119026154,
-0.0416743085,
-0.0467340611,
-0.0242493208,
-0.1057817861,
0.0112115331,
-0.0026035924,
0.0189112555,
-0.039794948,
0.022375429,
0.0142075922,
-0.056726668,
0.0276417658,
-0.0298364982,
0.0099036051,
-0.0077098911,
0.0041440818,
-0.0330791958,
-0.0061557689,
-0.000976354,
0.0195133723,
0.002262224,
-0.0327352434,
0.0392333567,
0.0481594913,
0.0308580287,
-0.0178447738,
0.0069936058,
-0.0747532547,
0.0066099353,
0.0277670585,
-0.0123621635,
0.0236508008,
-0.0014979353,
0.0156817622,
-0.0181506705,
0.0699807629,
0.0018039531,
-0.0122759473,
-0.016261097,
-0.0264594778,
-0.0569165573,
-0.0221752301,
0.0229543224,
-0.0174627677,
-0.0233465638,
-0.0847477764,
-0.0303340871,
-0.0530910678,
-0.0358630419,
-0.0039368425,
-0.039056059,
-0.0067712725,
-0.0353520475,
0.0145267146,
0.0249564331,
0.0062335394,
-0.0240318123,
-0.0438045152,
-0.0008903645,
0.0335387401,
0.0389792025,
0.0174288917,
0.0991626605,
0.0228722617,
0.0477079749,
0.0414245762,
0.0167535506,
-0.0530923717,
0.0354518257,
0.0075331395,
0.0290088672,
-0.0060267081,
0.026707653,
0.0267629363,
0.0092961406,
0.0526967086,
-0.0369677544,
0.0127649065,
-0.0205443688,
0.0647933632,
-0.0384236276,
0.0831204653,
0.006879352,
-0.0137726292,
-0.0147697814,
0.003094929,
0.0109779257,
0.0019367419,
-0.0113973441,
-0.0196499638,
-0.0155451568,
0.0468132794,
0.0340313092,
-0.0047894535,
0.0232361387,
0.0382435098,
-0.0474428162,
0.0058846874,
-0.0522890054,
-0.0213998649,
-0.0706551075,
0.0041109812,
-0.0003433975,
-0.1096793711,
-0.0539883897,
-0.0139288232,
-0.0211139694,
0.0224698558,
-0.0023755417,
-0.0176017601,
0.0752301589,
-0.0079179406,
-0.0049693324,
-0.0835703611,
0.0545781627,
-0.0431332849,
0.0084957564,
0.0043019527,
-5.395379113e-33,
-0.0297668781,
-0.0615101755,
-0.0135654397,
-0.0442084447,
-0.0108237527,
0.0152287371,
-0.0377165526,
0.015425683,
-0.0013474867,
0.0341689624,
-0.021874195,
0.0097432267,
0.0138280187,
-0.0264740381,
0.0685129389,
0.0429671779,
0.0475129522,
0.0554181337,
-0.0217205584,
-0.045617152,
0.0392959155,
-0.0182775706,
0.061487738,
-0.082154654,
0.1440729648,
-0.0703363419,
-0.0486448668,
0.0326239355,
0.0064661736,
-0.0490005016,
-0.00551093,
0.0457842536,
-0.0048788446,
0.018536564,
0.0099909492,
0.0540027507,
-0.0723990202,
0.0074445424,
0.0122733423,
0.0598701388,
-0.0028592593,
-0.0477483794,
0.0241027251,
-0.0209726002,
-0.0673425645,
0.0341393463,
-0.0023839478,
0.0004562226,
0.0429947078,
0.0257777665,
-0.0497399941,
0.0016052695,
-0.0412396118,
0.0191679467,
0.0225230083,
0.0338195376,
-0.0132627487,
-0.0778619796,
-0.1035861224,
0.0169484168,
-0.0173641108,
0.015833227,
0.0498141013,
0.0184863377,
-0.019992644,
-0.0317983627,
0.0619135275,
0.0069971159,
0.0358281806,
0.0143159805,
0.0276531111,
-0.059323635,
-0.0121159395,
0.064569205,
0.0360121764,
-0.0600468963,
-0.0035078349,
-0.0256232861,
0.0164436419,
0.0565247908,
-0.0355878435,
-0.0083322274,
-0.0685292482,
-0.0627209842,
-0.0088371262,
-0.094707109,
-0.0063603395,
-0.0201151911,
-0.0100590261,
-0.0334778354,
-0.0065152594,
0.0338215977,
0.0241162293,
-0.0426577181,
-0.0099877529,
-0.062453229,
-0.0146030057,
-0.0337795317,
0.0380558893,
0.0535428487,
0.0039146277,
-0.0001041798,
-0.0205743704,
0.0336128771,
0.0218800865,
0.0237033926,
-0.001190289,
-0.0106903594,
-0.0346930549,
0.0004810165,
-0.0159630775,
0.0215542465,
-0.0027869653,
0.0532424077,
0.0153262643,
0.0186647587,
-0.0044465195,
-0.0380037986,
0.0841677934,
0.0621569566,
-0.0110019892,
0.0478542894,
-0.0412659384,
0.0031282308,
-0.0474620089,
0.0114603899,
-0.0318317041,
0.0608262084,
0.0703312084,
-0.0449375808,
-0.0020689568,
-0.033665359,
2.644e-7,
0.0150565384,
0.0836542845,
0.0131883817,
-0.0389501862,
0.0065368814,
0.0118496977,
0.0083068144,
0.0240677446,
-0.0027121396,
0.0505953729,
-0.0086919377,
0.0309313666,
0.0654954761,
0.0424116589,
-0.0540065393,
-0.0237200242,
-0.0434415489,
0.0147586456,
-0.0274321474,
0.0297283325,
-0.0484996177,
0.0377126187,
0.0180419646,
-0.021407932,
-0.0616014376,
-0.0159069579,
0.0177651979,
0.0038436465,
0.0160637423,
0.0061984006,
0.0383747183,
0.0406974033,
0.0045204554,
-0.0448047891,
0.0143496701,
-0.0285970774,
0.0424859338,
0.0521690473,
-0.0142057231,
0.0378407985,
0.0005163917,
-0.0071668127,
-0.0071734958,
-0.0352795199,
0.0407673232,
0.0525647663,
-0.0226262882,
0.0259445552,
0.0053852419,
0.0189629458,
0.0516278595,
-0.022428805,
-0.0378172956,
0.0205483455,
0.0363900065,
-0.0383615978,
0.0089683272,
-0.0171934851,
-0.0173485093,
0.0025945087,
0.0146502052,
-0.0531250425,
0.0290760733,
0.0791126639,
0.0119862491,
-0.0486175045,
0.0264896583,
2.033520279e-34,
0.0198244657,
0.0101330923,
0.0351506174,
-0.0235605296,
0.0212341622,
0.0187121537,
0.0272111055,
-0.0114435125,
0.0416979752,
0.0396481901,
-0.0092185568
] |
http://arxiv.org/abs/2401.13160v1 | http://arxiv.org/pdf/2401.13160v1 | SpacTor-T5: Pre-training T5 Models with Span Corruption and Replaced Token Detection | 2024-01-24T00:36:13 | 2024-01-24T00:36:13 | Ke Ye, Heinrich Jiang, Afshin Rostamizadeh, Ayan Chakrabarti, Giulia DeSalvo, Jean-François Kagy, Lazaros Karydas, Gui Citovsky, Sanjiv Kumar | cs.LG | cs.LG, cs.CL | Pre-training large language models is known to be extremely resource
intensive and often times inefficient, under-utilizing the information
encapsulated in the training text sequences. In this paper, we present SpacTor,
a new training procedure consisting of (1) a hybrid objective combining span
corruption (SC) and token replacement detection (RTD), and (2) a two-stage
curriculum that optimizes the hybrid objective over the initial $\tau$
iterations, then transitions to standard SC loss. We show empirically that the
effectiveness of the hybrid objective is tied to the two-stage pre-training
schedule, and provide extensive analysis on why this is the case. In our
experiments with encoder-decoder architectures (T5) on a variety of NLP tasks,
SpacTor-T5 yields the same downstream performance as standard SC pre-training,
while enabling a 50% reduction in pre-training iterations and 40% reduction in
total FLOPs. Alternatively, given the same amount of computing budget, we find
that SpacTor results in significantly improved downstream benchmark
performance. | SpacTor-T5: Pre-training T5 Models with Span Corruption and Replaced Token Detection: Pre-training large language models is known to be extremely resource
intensive and often times inefficient, under-utilizing the information
encapsulated in the training text sequences. In this paper, we present SpacTor,
a new training procedure consisting of (1) a hybrid objective combining span
corruption (SC) and token replacement detection (RTD), and (2) a two-stage
curriculum that optimizes the hybrid objective over the initial $\tau$
iterations, then transitions to standard SC loss. We show empirically that the
effectiveness of the hybrid objective is tied to the two-stage pre-training
schedule, and provide extensive analysis on why this is the case. In our
experiments with encoder-decoder architectures (T5) on a variety of NLP tasks,
SpacTor-T5 yields the same downstream performance as standard SC pre-training,
while enabling a 50% reduction in pre-training iterations and 40% reduction in
total FLOPs. Alternatively, given the same amount of computing budget, we find
that SpacTor results in significantly improved downstream benchmark
performance. | [
0.0335620344,
0.0789587125,
0.0138700707,
0.0523931272,
-0.0054697283,
0.0044944044,
-0.0035701804,
-0.008080448,
-0.0449675471,
-0.0426248424,
0.0054656328,
-0.0271814372,
0.0043753483,
0.0482898876,
0.0440921187,
-0.0351347961,
0.0745362118,
0.0003659899,
0.0163817145,
-0.0109753413,
0.0135821151,
-0.002883302,
-0.0004166113,
0.0060438155,
-0.0380543172,
0.0071158591,
-0.0071537066,
0.0342020504,
-0.0122539168,
-0.0515422449,
0.039193958,
0.0568171814,
0.0041142027,
0.0176995825,
0.0000021965,
-0.0015105429,
-0.028400531,
-0.0282913428,
-0.0189857911,
0.0714059398,
0.0420883745,
-0.0313419364,
-0.0105633186,
0.0289334394,
-0.0026445764,
-0.0371580832,
0.0367190726,
0.0916259736,
0.0803451911,
0.0866527334,
-0.0108446022,
0.0129622836,
0.0307585876,
-0.0125702573,
-0.0287493262,
-0.0468156151,
0.0377807505,
-0.038045954,
-0.0505462661,
0.0094635552,
-0.0366892517,
0.0137816807,
0.0280867051,
0.0136584779,
0.0245640296,
0.0354497209,
-0.0596605912,
-0.0227496829,
-0.0330900252,
0.0305156019,
-0.0594101511,
-0.031441994,
0.0064052283,
0.021421073,
-0.0031607877,
-0.0791256949,
0.0183685981,
-0.0008284675,
0.0248783473,
0.0463050343,
0.0109786987,
-0.0484235249,
0.0404830426,
0.0058710845,
-0.0322194286,
0.007005414,
0.0362694003,
-0.0668248013,
-0.0064267903,
0.0321135856,
-0.0125955204,
-0.0528755374,
0.0418399833,
-0.0141385784,
-0.0525989458,
0.0220847111,
-0.0789614841,
-0.039677877,
-0.0125175696,
-0.0368334986,
0.0058969106,
0.0516529158,
-0.0317605026,
0.0105575286,
-0.0550787523,
0.0418532602,
-0.0144553194,
0.0285420958,
-0.0337622575,
0.0173181258,
0.0044347667,
-0.0232575517,
-0.0417062901,
0.0446803384,
-0.0398835875,
-0.0494451225,
-0.0346117355,
0.0074428674,
-0.0054551852,
0.0074445759,
-0.015104739,
0.0266870391,
-0.0104964813,
0.0248867366,
0.0092185326,
0.0497367419,
0.0003829022,
-0.0052386634,
-0.002441155,
-0.0807874724,
-0.0490798652,
0.0006602453,
-0.0110427821,
-0.0466153398,
0.0151353898,
0.045164872,
-0.0257428158,
-0.0594082251,
-0.0523745045,
-0.0299393423,
-0.0318278931,
-0.0053261048,
-0.0205817707,
-0.0374229737,
0.0340079442,
-0.0158131681,
0.0031637931,
0.0122603858,
0.0353312753,
0.0229549576,
-0.0248880945,
0.0679237619,
-0.035723228,
-0.0229155757,
-0.0151705137,
-0.0303821545,
0.0443837978,
0.0740022287,
0.0107157575,
0.0345124379,
0.0338807888,
-0.0077937683,
-0.0319638848,
0.0542544201,
0.0423901156,
-0.0020612765,
-0.0070348759,
-0.0309081227,
0.0625832975,
0.0251726899,
0.0849610269,
0.0697090253,
-0.0458416305,
-0.0027936483,
0.0392927974,
0.0099699721,
0.0042897579,
-0.0014620102,
0.0122227734,
-0.0361413248,
-0.0088719726,
-0.0104561178,
-0.042010203,
-0.0252257586,
0.0064656241,
-0.0124546057,
0.0424096622,
0.0228153057,
-0.0315928459,
-0.0006030342,
-0.0450323671,
-0.0128686903,
-0.030474985,
-0.0209878199,
0.0429255851,
0.0257496499,
-0.056340117,
0.0944855362,
0.0189084038,
-0.0455906242,
0.0444131196,
-0.0294813719,
-0.0116718831,
0.0900881737,
0.0843795612,
0.0017996834,
-0.0132941362,
-0.0000084343,
-0.0033880265,
0.0132965883,
0.0449337363,
-0.022004554,
0.0243756995,
-0.0520050228,
-0.0122452462,
-0.0077107963,
0.0190443192,
0.027235264,
0.0271287728,
0.0076556429,
0.0015000051,
-0.0527187288,
-0.0414726175,
0.0577553175,
0.009584412,
-0.0184667818,
0.0170643702,
0.0031756107,
0.0322319232,
0.0109164268,
-0.0333736092,
-0.0152846919,
0.0334997885,
-0.0292167272,
-0.0433383919,
-0.009118991,
-0.1131114438,
-0.0748840421,
0.0588238351,
-0.0335232578,
-0.0086040702,
0.0746185035,
-0.0244385805,
-0.0459617116,
0.0375025868,
-0.0489855669,
0.0459306575,
-0.0432219319,
0.0489913337,
0.0055075251,
0.0297089443,
-0.0067906631,
0.0618899763,
-0.0311063714,
0.0152400555,
-0.0107312528,
0.0296687372,
0.0154662998,
-0.042525094,
-0.0469442382,
0.0072752628,
-0.0243481696,
-0.0042135655,
-0.0078043179,
0.0194082148,
-0.0231782719,
-0.0308325682,
-0.0763300732,
-0.0074124294,
-0.0451505184,
0.0197945666,
0.0212055147,
0.0051560849,
0.0164752789,
-0.0223111641,
-0.0009959958,
-0.0294998549,
-0.0004128324,
0.0309034232,
0.0264104009,
0.0011866906,
0.001416108,
0.0031132516,
-0.0172719248,
-0.0125215063,
0.041579932,
-0.0297022033,
-0.0220719427,
-0.0505172759,
0.0034441042,
0.0450524054,
0.0326630846,
0.0131970132,
-0.0247173514,
-0.0253672376,
0.001947147,
0.0332795568,
0.0218384005,
0.0233807303,
0.0813912973,
-0.0252102781,
0.0077477549,
-0.0338901021,
0.0043105823,
-0.0016407748,
0.011046648,
-0.0513650812,
-0.0076713385,
0.0410414189,
0.058119569,
-0.0104870256,
0.0229187086,
-0.0583402328,
-0.0247506183,
-0.0045514768,
-0.0443505161,
0.0041650175,
-0.0481182933,
-0.0297361743,
0.0196730494,
-0.0470401905,
0.0181810409,
0.0305425897,
-0.0516902357,
-0.0282904208,
-0.0412162319,
0.0204196796,
-0.0301497504,
-0.0110767707,
0.0184407234,
0.0147692384,
-0.0203198474,
-0.1032261699,
0.0033312733,
0.0341206975,
-0.0030970573,
-0.0107368743,
-0.0007725321,
-0.039575994,
0.0188036766,
-0.0158397611,
-0.0564236939,
-0.0380666442,
-0.0273015816,
0.0123259183,
-0.0036430508,
0.033686012,
0.0033841187,
-0.0304380432,
-0.0051770988,
-0.043995373,
0.0722165182,
-0.00823808,
0.0329573974,
-0.0305748209,
0.0655624941,
0.0317744091,
-0.0346323252,
0.0191536043,
-0.0721025467,
-0.0112454528,
0.0711399019,
0.0374415591,
0.0083673596,
-0.0366156884,
0.0321221612,
0.0580010153,
-0.0247081555,
0.0179825872,
-0.0037279218,
0.0284877848,
-0.0468364172,
-0.0500688925,
0.1150574908,
0.0185870081,
-0.0367123745,
0.0134927602,
0.0352790803,
0.031898234,
0.0175450966,
-0.0357429609,
-0.0761398152,
-0.0390106514,
0.0388657451,
0.0267087538,
0.0040403428,
-0.0104006082,
0.0091789085,
-0.105714485,
-0.0146704549,
-0.0026124544,
-0.0577807911,
-0.0099012693,
-0.0308415219,
-0.033029791,
-0.0308186095,
0.0820320621,
0.0067329584,
0.0078422958,
0.0106696524,
0.0366660394,
-0.0148417596,
-0.0824457183,
0.0005150837,
-0.0385576896,
0.0556507483,
0.0954877958,
-0.0203487184,
-0.0141526479,
-0.0249481816,
0.079317309,
-0.0015141064,
-0.0016586906,
-0.0207197014,
0.0618801415,
-0.0492510125,
-0.037066631,
0.0204559825,
0.0177680887,
-0.047565572,
0.018760236,
-0.0138639677,
0.0748789385,
0.0352054872,
-0.0157290213,
0.0170566086,
-0.0471940972,
-0.0059435316,
0.0715591609,
0.0196815077,
-0.0397126786,
-0.0690982491,
-0.0090994118,
0.0050979671,
-0.0949123055,
-0.0028815181,
0.0036250243,
-0.0031803355,
-0.0537811927,
-0.0439609997,
-0.0511791855,
0.0328032412,
0.0415459462,
-0.0006651009,
-0.0288865045,
-0.0135436002,
-0.0186579637,
-0.0406508856,
0.0054688295,
-0.0505808219,
0.0037166097,
-0.0543061271,
0.0224574264,
-0.0280954912,
0.0139780724,
-0.0166558512,
-0.0219826289,
-0.0788047239,
-0.0522383377,
0.0108531928,
0.0276258327,
0.0104618585,
0.0299305227,
-0.0141907921,
-0.0168938357,
0.0452051274,
0.0449901633,
-0.0616655163,
0.0222273655,
0.0018578846,
-0.0672546923,
-0.0029344556,
0.0171024967,
-0.0250753071,
-0.0212967899,
0.0331048071,
0.0459308438,
0.0284655541,
0.0337568559,
0.0454726405,
0.0202152804,
0.0535105206,
-0.0043514241,
-0.0164317321,
-0.0221541394,
0.0320303924,
-0.0345592313,
-0.005540045,
-0.0135947578,
0.0311372392,
0.0361281931,
0.0381387658,
-0.0138960863,
0.0275111888,
0.0045505334,
0.0430654548,
-0.029708134,
-0.0389619097,
0.002457635,
0.0052910764,
0.0494582951,
0.0121379877,
-0.0092638852,
-0.0363245234,
0.0354228988,
0.0122450013,
-0.0551115759,
-0.0266384371,
0.0541253127,
0.0197949726,
0.0078186709,
0.0293955468,
-0.0411421917,
0.0201292317,
-0.0681966394,
-0.0183947179,
0.0463211834,
-0.0284995474,
0.0438228957,
-0.0147294737,
-0.0191626381,
-0.0076351734,
-0.0122349598,
-0.0115605602,
-0.058095064,
0.0159656368,
0.0090481266,
-0.0161925834,
-0.0310981199,
0.0689955354,
-0.0207738541,
0.0630688518,
-0.0036301934,
-0.0010126765,
-0.0279031731,
-0.0115522193,
-0.0134344604,
0.0265924316,
-0.0345242918,
0.005868894,
-0.0016290082,
-0.0622117482,
0.0025174574,
-0.0219224431,
0.0261282474,
0.0215796903,
0.0209811274,
0.0605989397,
-0.0043509565,
0.0002766558,
0.0025161975,
-0.0124616195,
0.0741528496,
-0.0388000011,
-0.0275469366,
-6.681157242e-33,
0.0108339684,
0.001616988,
0.012395395,
-0.0432730205,
-0.0507006645,
0.0112187462,
-0.0172836985,
-0.0261482429,
0.003636912,
-0.008846825,
0.0027154195,
0.0234261323,
0.0149102826,
0.0554299317,
0.0060515204,
0.009883116,
0.1087915003,
-0.0057899682,
-0.0209479555,
-0.0091609554,
0.030918384,
0.0463444442,
0.0526691861,
0.0030731312,
-0.0375286043,
0.0190569013,
-0.0083556222,
0.005833589,
0.0453398377,
0.0095789814,
-0.0253110901,
0.0411586873,
0.0314189084,
-0.0368031189,
-0.0094753876,
-0.0208599139,
-0.1059905663,
0.0218057334,
-0.0145449983,
0.0892763212,
-0.0373748429,
-0.0434246026,
0.036756333,
0.0154906735,
-0.0130198952,
0.0261436459,
0.0120646432,
0.0105790617,
-0.0100815399,
0.0019313919,
-0.0361804217,
0.0116916383,
-0.0250803474,
0.0074394736,
0.0250431504,
0.0223935638,
0.0000736846,
0.0482824072,
-0.050736513,
-0.0014405453,
-0.0102915447,
0.0542176701,
0.0159816854,
-0.000056324,
-0.0277234185,
0.0616753586,
0.0531550869,
-0.0143921226,
0.0113882227,
-0.002808531,
0.0461457185,
-0.0093775392,
0.0471908487,
-0.0099049844,
0.0655378103,
-0.0517697409,
-0.0350721478,
-0.0142660644,
0.0042425576,
0.0405937135,
0.0295361914,
0.0267484691,
0.0069818981,
-0.0312305726,
-0.0299367551,
-0.0397905596,
-0.0554668233,
-0.0185619872,
0.0357554071,
0.0492707677,
0.010761736,
0.0357855633,
0.0139057236,
-0.0225912556,
0.010411201,
0.0009311519,
0.0050403709,
0.006551994,
-0.0145307826,
0.0597900823,
-0.039347358,
-0.0350168012,
0.0334384441,
-0.0104174651,
0.0006348743,
-0.0254346803,
0.0527996011,
0.0094157588,
-0.0803651959,
0.0038895125,
-0.024176348,
0.0390912257,
0.0124397054,
0.0363003053,
0.0467954651,
0.0061540324,
-0.0070519159,
0.0429765731,
0.0160755478,
0.0171091091,
-0.0479428507,
-0.0106415665,
0.0083490247,
0.0025575929,
-0.0037513019,
-0.0037956056,
-0.0506148301,
0.0130636171,
-0.0076745753,
-0.0755791664,
-0.0238770302,
0.0115262885,
2.828e-7,
0.0046944786,
0.0597069785,
0.0537470542,
-0.0260581225,
0.0299484897,
0.0135271512,
-0.0074196854,
0.0493586324,
-0.0211432911,
0.0279511176,
-0.0201099049,
-0.0199671686,
0.0483348295,
0.0274037458,
-0.0830283388,
0.0318409577,
-0.0788703933,
0.019598119,
-0.0537759997,
-0.009578215,
0.031157624,
0.0963000953,
-0.023942424,
-0.0274901893,
0.034208525,
-0.015942879,
0.0495421067,
0.0302017909,
-0.027094027,
0.0076992367,
-0.0358518846,
0.0170986634,
0.0144684101,
0.0251999591,
-0.0367172696,
0.0140652107,
-0.0147677055,
0.0491072834,
-0.0095338756,
-0.0573983639,
0.0071731368,
0.0099066375,
-0.0115971779,
-0.0209529027,
0.081386894,
0.0181334913,
-0.0110031655,
-0.0227262191,
-0.0513768606,
0.006591164,
-0.0095282821,
-0.02656563,
-0.0120389154,
0.03305934,
0.0407049321,
0.0288816001,
-0.04495731,
-0.090830788,
0.0270185564,
-0.0067911884,
-0.0326760821,
-0.0130133964,
-0.0225988645,
0.0325012654,
0.0524257794,
-0.0723355338,
-0.0009526752,
3.031535187e-34,
0.0483064353,
-0.04562518,
0.0328968875,
0.0471002609,
0.0242427811,
-0.0135921277,
-0.0049912455,
-0.0035860811,
-0.0272171851,
-0.0446541421,
-0.0278639421
] |