File size: 62,755 Bytes
6fa4bc9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 |
{
"paper_id": "2021",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T14:47:08.725551Z"
},
"title": "Few-shot and Zero-shot Approaches to Legal Text Classification: A Case Study in the Financial Sector",
"authors": [
{
"first": "Rajdeep",
"middle": [],
"last": "Sarkar",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "National University of Ireland Galway",
"location": {
"country": "Ireland"
}
},
"email": "rajdeep.sarkar@insight-centre.org"
},
{
"first": "Atul",
"middle": [
"Kr"
],
"last": "Ojha",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "National University of Ireland Galway",
"location": {
"country": "Ireland"
}
},
"email": "atulkumar.ojha@insight-centre.org"
},
{
"first": "Jay",
"middle": [],
"last": "Megaro",
"suffix": "",
"affiliation": {
"laboratory": "FMR LLC",
"institution": "",
"location": {
"settlement": "Boston",
"country": "USA"
}
},
"email": "jay.megaro@fmr.com"
},
{
"first": "John",
"middle": [],
"last": "Mariano",
"suffix": "",
"affiliation": {
"laboratory": "FMR LLC",
"institution": "",
"location": {
"settlement": "Boston",
"country": "USA"
}
},
"email": "john.mariano@fmr.com"
},
{
"first": "Vall",
"middle": [],
"last": "Herard",
"suffix": "",
"affiliation": {
"laboratory": "FMR LLC",
"institution": "",
"location": {
"settlement": "Boston",
"country": "USA"
}
},
"email": "vall.herard@fmr.com"
},
{
"first": "John",
"middle": [
"P"
],
"last": "Mccrae",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "National University of Ireland Galway",
"location": {
"country": "Ireland"
}
},
"email": "john.mccrae@insight-centre.org"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "The application of predictive coding techniques to legal texts has the potential to greatly reduce the cost of legal review of documents, however, there is such a wide array of legal tasks and continuously evolving legislation that it is hard to construct sufficient training data to cover all cases. In this paper, we investigate few-shot and zero-shot approaches that require substantially less training data and introduce a triplet architecture, which for promissory statements produces performance close to that of a supervised system. This method allows predictive coding methods to be rapidly developed for new regulations and markets.",
"pdf_parse": {
"paper_id": "2021",
"_pdf_hash": "",
"abstract": [
{
"text": "The application of predictive coding techniques to legal texts has the potential to greatly reduce the cost of legal review of documents, however, there is such a wide array of legal tasks and continuously evolving legislation that it is hard to construct sufficient training data to cover all cases. In this paper, we investigate few-shot and zero-shot approaches that require substantially less training data and introduce a triplet architecture, which for promissory statements produces performance close to that of a supervised system. This method allows predictive coding methods to be rapidly developed for new regulations and markets.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Organizations that are governed by legal and regulatory statues concerning communications with the public are required to comply with principles-based content standards. As such, this involves a significant expense due to having to use highly qualified staff to review, iterate on communications internally and file content with regulators, externally. There is thus a substantial expense in terms of iteration time and specialized staff associated with this process. With recent advances in Natural Language Processing (NLP) technologies, it is increasingly becoming possible to automatically flag high-risk statements by predictive coding and thus reduce the cost of these manual reviews. However, each industry has specific regulatory requirements and modern NLP systems need large training sets to be effective, and as such it is challenging to develop such systems. In this paper, we focus on a single example of such a regulatory compliance in the financial domain under the US regulation FINRA 2210 1 , which states that \"no member may make any false, exaggerated, unwarranted, promissory or misleading statement or claim in any communication.\" We examine how we can train a system in the following settings: firstly in a traditional dataheavy supervised setting, where a large number of existing examples have been classified. Secondly, we investigate a zero-shot training situation, where we have asked a legal expert to provide only rough guidelines for what is not compliant with the legal code. Finally, we combine this in a few-shot setting and show that with comparatively little training data, we can achieve performance that is equivalent with the data-heavy supervised setting and thus enables text classification systems for regulatory compliance to be constructed quickly and with little effort allowing them to cover a wide range of industries and national regulatory frameworks.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "There has been some work in the area of legal text classification and the application of text classification techniques to legal texts has mostly been successful so far. Methods based on counting the words in the text and then classifying using machine learning approaches such as support vector machines (Cortes and Vapnik, 1995) for example by Sulea et al. (2017) , where they applied this method to the classification of texts according to the legal area, ruling and time span of the text. Deep learning methods such as Convolutional Neural Networks (CNNs) have been shown to further improve the performance of such systems (Wei et al., 2018) . More recently, the emergence of large pre-trained language models such as BERT (Devlin et al., 2019) has further increased the performance and Shaheen et al. (2020) showed that these models could be used to classify legal texts according to thousands of labels and even on multiple languages if sufficient training data exists.",
"cite_spans": [
{
"start": 305,
"end": 330,
"text": "(Cortes and Vapnik, 1995)",
"ref_id": "BIBREF1"
},
{
"start": 346,
"end": 365,
"text": "Sulea et al. (2017)",
"ref_id": "BIBREF12"
},
{
"start": 627,
"end": 645,
"text": "(Wei et al., 2018)",
"ref_id": "BIBREF13"
},
{
"start": 727,
"end": 748,
"text": "(Devlin et al., 2019)",
"ref_id": "BIBREF2"
},
{
"start": 791,
"end": 812,
"text": "Shaheen et al. (2020)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "A criticism of such NLP-based approaches to predictive coding, especially with the emergence of more sophisticated deep learning methods, is that they can appear to be 'black boxes', and thus there has been work in providing explainable systems (Mahoney et al., 2019) that can identify snippets and provides explanations for why they make certain predictions. Similarly, some work has gone into the investigation of specific complexities of legal texts, such as in Nallapati and Manning (2008) , who showed that for some legal texts the complex combination of negative and positive statements can confused machine learning approaches. They showed that by combining these machine learning approaches with propositional logic, text classification systems could handle intricate legal wording.",
"cite_spans": [
{
"start": 465,
"end": 493,
"text": "Nallapati and Manning (2008)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "To solve the problem of legal text classification, we approach this with a triplet architecture (Wei et al., 2021) where an input sentence, s, is compared with a positive example s + and a negative example s \u2212 as depicted in Figure 1 . We begin by describing the model architecture. Then we discuss the triplet loss used for training the network. Finally, we describe the classification model used for the final classification.",
"cite_spans": [
{
"start": 96,
"end": 114,
"text": "(Wei et al., 2021)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [
{
"start": 225,
"end": 233,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Methodology",
"sec_num": "3"
},
{
"text": "Most existing methods of text classification only consider the local features of the samples, and their experimental results show better performance than traditional non-deep learning methods. However, in these methods, the global features of the sample are usually ignored, and these ignored global features will affect the classification accuracy. These global features are key to the use-case presented. To solve this problem, a triplet capsule network framework is proposed for text classification, to optimize results.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model Architecture",
"sec_num": "3.1"
},
{
"text": "A triplet network consist of three instances of the same neural network with shared parameters. The network takes as input three examples in each sample. The three samples consists of the anchor, positive and negative examples. The anchor and positive examples belong to the same class, while the negative example belongs to a different class. The network outputs two values, the distance between the anchor and the positive example and the distance between the anchor and the negative example.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model Architecture",
"sec_num": "3.1"
},
{
"text": "We design a triplet network for the sentence classification task. The network encodes each incoming sentence using Sentence-BERT (Reimers and Gurevych, 2019) encoder. Sentence-BERT captures the contextual information in a sentence in a fixed-size vector representation. The contextual sentence representation is then fed to a two-layer perceptron. The hidden layer of the perceptron has ReLU (Nair and Hinton, 2010) activation for introducing non-linearity in the perceptron.",
"cite_spans": [
{
"start": 129,
"end": 157,
"text": "(Reimers and Gurevych, 2019)",
"ref_id": "BIBREF10"
},
{
"start": 392,
"end": 415,
"text": "(Nair and Hinton, 2010)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Model Architecture",
"sec_num": "3.1"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "e 1 s = S-BERT(s)",
"eq_num": "(1)"
}
],
"section": "Model Architecture",
"sec_num": "3.1"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "e 2 s = RELU (W \u03b8,1 e 1 s )",
"eq_num": "(2)"
}
],
"section": "Model Architecture",
"sec_num": "3.1"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "e 3 s = W \u03b8,2 e 2 s",
"eq_num": "(3)"
}
],
"section": "Model Architecture",
"sec_num": "3.1"
},
{
"text": "where W \u03b8,1 \u2208 R d e 2 \u00d7d e 1 h and W \u03b8,2 \u2208 R d e 3 \u00d7d e 2 and the parameter matrices to be learned during training. The Sentence-BERT model is also finetuned during the training procedure. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model Architecture",
"sec_num": "3.1"
},
{
"text": "S-Bert Encoder S-Bert Encoder S-Bert Encoder FC, ReLu FC, ReLu FC, ReLu FC FC FC d(s, s + ) d(s, s -) s + s - s",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Model Architecture",
"sec_num": "3.1"
},
{
"text": "Triplet loss (Hoffer and Ailon, 2015) has been used in few-shot classification methods. Although introduced for images, it has been successfully adapted in natural language processing (Wei et al., 2021; Lauriola and Moschitti, 2021) . Triplet loss enables the network to distinguish been positive and negative examples of a class. It is defined in Equation 4.",
"cite_spans": [
{
"start": 13,
"end": 37,
"text": "(Hoffer and Ailon, 2015)",
"ref_id": "BIBREF3"
},
{
"start": 184,
"end": 202,
"text": "(Wei et al., 2021;",
"ref_id": "BIBREF14"
},
{
"start": 203,
"end": 232,
"text": "Lauriola and Moschitti, 2021)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Triplet Loss",
"sec_num": "3.2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "L(D + , D \u2212 ) = D + , D \u2212 \u2212 1 2 2",
"eq_num": "(4)"
}
],
"section": "Triplet Loss",
"sec_num": "3.2"
},
{
"text": "where D + and D \u2212 are defined in Equation 7 and 8 respectively. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Triplet Loss",
"sec_num": "3.2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "d + (s, s + ) = e 3 s \u2212 e 3 s + 2 (5) d \u2212 (s, s \u2212 ) = e 3 s \u2212 e 3 s \u2212 2",
"eq_num": "(6)"
}
],
"section": "Triplet Loss",
"sec_num": "3.2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "D + = e d + (s,s + ) e d + (s,s + ) + e d \u2212 (s,s \u2212 )",
"eq_num": "(7)"
}
],
"section": "Triplet Loss",
"sec_num": "3.2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "D \u2212 = e d \u2212 (s,s \u2212 ) e d + (s,s + ) + e d \u2212 (s,s \u2212 )",
"eq_num": "(8)"
}
],
"section": "Triplet Loss",
"sec_num": "3.2"
},
{
"text": "(s,s + ) d \u2212 (s,s \u2212 ) \u2192 0, then L \u2192 0.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Triplet Loss",
"sec_num": "3.2"
},
{
"text": "We minimize L to learn the parameters of our model.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Triplet Loss",
"sec_num": "3.2"
},
{
"text": "The network learns sentence representations where examples of the same class are close together. The closeness of two sentences is measured by calculating the euclidean distance between their representations from Equation 3. For the final classification, we use a Support Vector Machine (SVM) with Radial Basis Function (RBF) kernel. We use SVM for classification as it learns by minimizing the hinge loss which is similar to the loss used for training the triplet network. Given the sentence representation from Equation 3, the SVM outputs a probability p of a sentence being a promissory sentence. The sentence is classified being promissory using Equation 9.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Classification",
"sec_num": "3.3"
},
{
"text": "class(s) = promissory p \u2265 \u03b1 not promissory p < \u03b1 (9)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Classification",
"sec_num": "3.3"
},
{
"text": "where \u03b1 is a hyperparameter to be set.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Classification",
"sec_num": "3.3"
},
{
"text": "We retrieved data from internal and external data sources in the financial services industry to create the initial data sets for the approach. After data setup, we cleaned the data to remove duplicate and irrelevant content to ensure data quality before review. Each data point was reviewed and labelled by both in-house licensed staff and contractors to confirm the interpretation of regulatory content standards.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Dataset",
"sec_num": "4.1"
},
{
"text": "We split the dataset into training, development and test set. The training set contains 2,016 promissory sentences and 3,260 non-promissory sentence. The test set contains 860 and 1,402 promissory and non-promissory examples, respectively. For our few-shot learning model, we sample 40 promissory and 190 non-promissory examples sentences from the training set and learn our model on this subset.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Dataset",
"sec_num": "4.1"
},
{
"text": "We compare the performance of our approach with the following supervised learning methods.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Baselines and Evaluation Metrics",
"sec_num": "4.2"
},
{
"text": "\u2022 Naive Bayes: We learn a Naive Bayes classification model using TF-IDF scores of the tokens in the sentence.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Baselines and Evaluation Metrics",
"sec_num": "4.2"
},
{
"text": "\u2022 Multi-Layer Perceptron (MLP): We learn a two-layer perceptron with ReLu activation in the hidden layer using the TF-IDF scores of the sentence tokens as input features to the model.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Baselines and Evaluation Metrics",
"sec_num": "4.2"
},
{
"text": "\u2022 SVM: Similar to the MLP model, we learn an SVM model for the classification task. We set the regularization parameter C and gamma to 1.0 and 0.1 respectively.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Baselines and Evaluation Metrics",
"sec_num": "4.2"
},
{
"text": "\u2022 Sentence-BERT (Reimers and Gurevych, 2019) : This setting is similar to our proposed approach. We encode each sentence into a fixed-sized vector using its Sentence-BERT embedding. The sentence embedding is then fed into a 3 layer fully connected neural network with ReLu activation in the first two layers. The model is learned by minimizing the cross-entropy loss of classification using the Adam optimizer.",
"cite_spans": [
{
"start": 16,
"end": 44,
"text": "(Reimers and Gurevych, 2019)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Baselines and Evaluation Metrics",
"sec_num": "4.2"
},
{
"text": "\u2022 Laser (Artetxe and Schwenk, 2019) : In this setting, we encode each sentence using its Laser embeddings. The remaining architecture remains the same as that used in the Sentence-BERT model.",
"cite_spans": [
{
"start": 8,
"end": 35,
"text": "(Artetxe and Schwenk, 2019)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Baselines and Evaluation Metrics",
"sec_num": "4.2"
},
{
"text": "In addition to the supervised approaches, we compare our few-shot learning approach against a zeroshot learning approach. Yin et al. (2019) suggested method for using pre-trained natural language inference models as sequence classifiers. Towards this end, we use BART model (Lewis et al., 2020) as our zero-shot learning model. We consider the sentences tagged as 'promissory' as the hypothesis. The probability of a sentence being the premise Sentence Model Result Gold Label 1 Stocks are an income source which main street is ignoring non-promissory promissory 2 It is going up in all currencies non-promissory promissory 3 Joe Smith picks the best stock in each sector for the fund non-promissory promissory 4 All rights reserved. promissory non-promissory 5 Save more now. promissory non-promissory 6 There is no action required on your part. promissory non-promissory for these tagged sentence is calculated using the BART model. We then consider the maximum of those scores, and if the maximum score is greater than 0.7, we classify the sentence as a promissory sentence.",
"cite_spans": [
{
"start": 122,
"end": 139,
"text": "Yin et al. (2019)",
"ref_id": "BIBREF15"
},
{
"start": 274,
"end": 294,
"text": "(Lewis et al., 2020)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Baselines and Evaluation Metrics",
"sec_num": "4.2"
},
{
"text": "For the task, we use the Sentence-BERT base model. It encodes a sentence into a fixed-size vector of length 768. We set d e 1 , d e 2 and d e 3 to 768, 300 and 10 respectively. For every positive sentence belonging to the promissory class, we sample three sentences from the non-promissory class as negative sentences. We use grid-search on the development set to set the values of hyperparameters. The batch size is set to 16 for the triplet network and the model trained using Adam optimizer (Kingma and Ba, 2015) with a learning rate of 1e-5 for 10 epochs. We set the cost parameter C of SVM to 0.03 and \u03b1 in Equation 9 to 0.005.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Implementation Details",
"sec_num": "4.3"
},
{
"text": "In section, we first perform a quantitative analysis of models. We then study a few examples where our approach produces results different from the gold standard dataset.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "In this work, we propose a few-shot learning method for legal text classification. Table 1 shows the performance of different model. Even when training with a limited number of examples, the few-shot learning model achieves better recall performance as compared to different supervised models. We find that the precision of our model is better than the zero-shot learning model but lower than the supervised models. Overall the F-Measure shows that similar results can be obtained with a few-shot approach and this enables the goal of rapid training of systems for different legal tasks. In the situation where the classifier is applied as a first filter, a high recall is preferable as we would rather create more work for a second manual annotation than miss some important texts.",
"cite_spans": [],
"ref_spans": [
{
"start": 83,
"end": 90,
"text": "Table 1",
"ref_id": "TABREF0"
}
],
"eq_spans": [],
"section": "Quantitative Analysis",
"sec_num": "5.1"
},
{
"text": "In Table 2 , we see some examples of misclassifications made by our algorithm. It is obvious that this is a very challenging task, with subtle changes in meaning being important for the classification. Examples 3 and 6 both appear to make factual statements, however, Example 3 is classed as promissory due to the context that 'Joe Smith' is likely an agent of the company. Similarly, Example 2 is difficult to classify without context and this shows that the introduction of further context is most likely to improve the effectiveness of the approach.",
"cite_spans": [],
"ref_spans": [
{
"start": 3,
"end": 10,
"text": "Table 2",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Qualitative Analysis",
"sec_num": "5.2"
},
{
"text": "We have investigated the use of few-shot and zeroshot text classification methods for the quick development of predictive coding systems for legal texts. We found that zero-shot systems have a substantial decrease in performance relative to a supervised approach. We then developed a few-shot approach based on a triplet architecture and showed that this model is within a few percentage points of the supervised system in performance but requires much less manual annotation in order to develop the system.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6"
}
],
"back_matter": [
{
"text": "This work has been funded by FMR LLC. Researchers at the Data Science Institute are supported by Science Foundation Ireland as part of Grant Number SFI/12/RC/2289_P2, Insight SFI Centre for Data Analytics.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgements",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Massively multilingual sentence embeddings for zeroshot cross-lingual transfer and beyond",
"authors": [
{
"first": "Mikel",
"middle": [],
"last": "Artetxe",
"suffix": ""
},
{
"first": "Holger",
"middle": [],
"last": "Schwenk",
"suffix": ""
}
],
"year": 2019,
"venue": "Trans. Assoc. Comput. Linguistics",
"volume": "7",
"issue": "",
"pages": "597--610",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mikel Artetxe and Holger Schwenk. 2019. Mas- sively multilingual sentence embeddings for zero- shot cross-lingual transfer and beyond. Trans. Assoc. Comput. Linguistics, 7:597-610.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Supportvector networks",
"authors": [
{
"first": "Corinna",
"middle": [],
"last": "Cortes",
"suffix": ""
},
{
"first": "Vladimir",
"middle": [],
"last": "Vapnik",
"suffix": ""
}
],
"year": 1995,
"venue": "Machine learning",
"volume": "20",
"issue": "3",
"pages": "273--297",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Corinna Cortes and Vladimir Vapnik. 1995. Support- vector networks. Machine learning, 20(3):273-297.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "BERT: Pre-training of deep bidirectional transformers for language understanding",
"authors": [
{
"first": "Jacob",
"middle": [],
"last": "Devlin",
"suffix": ""
},
{
"first": "Ming-Wei",
"middle": [],
"last": "Chang",
"suffix": ""
},
{
"first": "Kenton",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Kristina",
"middle": [],
"last": "Toutanova",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
"volume": "1",
"issue": "",
"pages": "4171--4186",
"other_ids": {
"DOI": [
"10.18653/v1/N19-1423"
]
},
"num": null,
"urls": [],
"raw_text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language under- standing. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Associ- ation for Computational Linguistics.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Deep metric learning using triplet network",
"authors": [
{
"first": "Elad",
"middle": [],
"last": "Hoffer",
"suffix": ""
},
{
"first": "Nir",
"middle": [],
"last": "Ailon",
"suffix": ""
}
],
"year": 2015,
"venue": "Similarity-Based Pattern Recognition -Third International Workshop",
"volume": "9370",
"issue": "",
"pages": "84--92",
"other_ids": {
"DOI": [
"10.1007/978-3-319-24261-3_7"
]
},
"num": null,
"urls": [],
"raw_text": "Elad Hoffer and Nir Ailon. 2015. Deep metric learning using triplet network. In Similarity-Based Pattern Recognition -Third International Workshop, SIM- BAD 2015, Copenhagen, Denmark, October 12-14, 2015, Proceedings, volume 9370 of Lecture Notes in Computer Science, pages 84-92. Springer.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Adam: A method for stochastic optimization",
"authors": [
{
"first": "P",
"middle": [],
"last": "Diederik",
"suffix": ""
},
{
"first": "Jimmy",
"middle": [],
"last": "Kingma",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Ba",
"suffix": ""
}
],
"year": 2015,
"venue": "3rd International Conference on Learning Representations",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In 3rd Inter- national Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Answer sentence selection using local and global context in transformer models",
"authors": [
{
"first": "Ivano",
"middle": [],
"last": "Lauriola",
"suffix": ""
},
{
"first": "Alessandro",
"middle": [],
"last": "Moschitti",
"suffix": ""
}
],
"year": 2021,
"venue": "Advances in Information Retrieval -43rd European Conference on IR Research, ECIR 2021, Virtual Event",
"volume": "12656",
"issue": "",
"pages": "298--312",
"other_ids": {
"DOI": [
"10.1007/978-3-030-72113-8_20"
]
},
"num": null,
"urls": [],
"raw_text": "Ivano Lauriola and Alessandro Moschitti. 2021. An- swer sentence selection using local and global con- text in transformer models. In Advances in Infor- mation Retrieval -43rd European Conference on IR Research, ECIR 2021, Virtual Event, March 28 - April 1, 2021, Proceedings, Part I, volume 12656 of Lecture Notes in Computer Science, pages 298-312. Springer.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "BART: denoising sequence-to-sequence pretraining for natural language generation, translation, and comprehension",
"authors": [
{
"first": "Mike",
"middle": [],
"last": "Lewis",
"suffix": ""
},
{
"first": "Yinhan",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "Naman",
"middle": [],
"last": "Goyal ; Abdelrahman Mohamed",
"suffix": ""
},
{
"first": "Omer",
"middle": [],
"last": "Levy",
"suffix": ""
},
{
"first": "Veselin",
"middle": [],
"last": "Stoyanov",
"suffix": ""
},
{
"first": "Luke",
"middle": [],
"last": "Zettlemoyer",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
"volume": "2020",
"issue": "",
"pages": "7871--7880",
"other_ids": {
"DOI": [
"10.18653/v1/2020.acl-main.703"
]
},
"num": null,
"urls": [],
"raw_text": "Mike Lewis, Yinhan Liu, Naman Goyal, Mar- jan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: denoising sequence-to-sequence pre- training for natural language generation, translation, and comprehension. In Proceedings of the 58th An- nual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 7871-7880. Association for Computational Linguistics.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "A framework for explainable text classification in legal document review",
"authors": [
{
"first": "J",
"middle": [],
"last": "Christian",
"suffix": ""
},
{
"first": "Jianping",
"middle": [],
"last": "Mahoney",
"suffix": ""
},
{
"first": "Nathaniel",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Peter",
"middle": [],
"last": "Huber-Fliflet",
"suffix": ""
},
{
"first": "Haozhen",
"middle": [],
"last": "Gronvall",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Zhao",
"suffix": ""
}
],
"year": 2019,
"venue": "2019 IEEE International Conference on Big Data (Big Data)",
"volume": "",
"issue": "",
"pages": "1858--1867",
"other_ids": {
"DOI": [
"10.1109/BigData47090.2019.9005659"
]
},
"num": null,
"urls": [],
"raw_text": "Christian J. Mahoney, Jianping Zhang, Nathaniel Huber-Fliflet, Peter Gronvall, and Haozhen Zhao. 2019. A framework for explainable text classifi- cation in legal document review. In 2019 IEEE International Conference on Big Data (Big Data), Los Angeles, CA, USA, December 9-12, 2019, pages 1858-1867. IEEE.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Rectified linear units improve restricted Boltzmann machines",
"authors": [
{
"first": "Vinod",
"middle": [],
"last": "Nair",
"suffix": ""
},
{
"first": "Geoffrey",
"middle": [
"E"
],
"last": "Hinton",
"suffix": ""
}
],
"year": 2010,
"venue": "Proceedings of the 27th International Conference on Machine Learning (ICML-10)",
"volume": "",
"issue": "",
"pages": "807--814",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Vinod Nair and Geoffrey E. Hinton. 2010. Rectified linear units improve restricted Boltzmann machines. In Proceedings of the 27th International Conference on Machine Learning (ICML-10), June 21-24, 2010, Haifa, Israel, pages 807-814. Omnipress.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Legal docket classification: Where machine learning stumbles",
"authors": [
{
"first": "Ramesh",
"middle": [],
"last": "Nallapati",
"suffix": ""
},
{
"first": "Christopher",
"middle": [
"D"
],
"last": "Manning",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "438--446",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ramesh Nallapati and Christopher D. Manning. 2008. Legal docket classification: Where machine learning stumbles. In Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 438-446, Honolulu, Hawaii. Association for Computational Linguistics.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Sentence-BERT: Sentence embeddings using Siamese BERTnetworks",
"authors": [
{
"first": "Nils",
"middle": [],
"last": "Reimers",
"suffix": ""
},
{
"first": "Iryna",
"middle": [],
"last": "Gurevych",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing",
"volume": "",
"issue": "",
"pages": "3980--3990",
"other_ids": {
"DOI": [
"10.18653/v1/D19-1410"
]
},
"num": null,
"urls": [],
"raw_text": "Nils Reimers and Iryna Gurevych. 2019. Sentence- BERT: Sentence embeddings using Siamese BERT- networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Nat- ural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019, pages 3980-3990. Association for Computational Linguis- tics.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Large scale legal text classification using transformer models",
"authors": [
{
"first": "Zein",
"middle": [],
"last": "Shaheen",
"suffix": ""
},
{
"first": "Gerhard",
"middle": [],
"last": "Wohlgenannt",
"suffix": ""
},
{
"first": "Erwin",
"middle": [],
"last": "Filtz",
"suffix": ""
}
],
"year": 2020,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Zein Shaheen, Gerhard Wohlgenannt, and Erwin Filtz. 2020. Large scale legal text classification using transformer models. CoRR, abs/2010.12871.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Exploring the use of text classification in the legal domain",
"authors": [
{
"first": "Octavia-Maria",
"middle": [],
"last": "Sulea",
"suffix": ""
},
{
"first": "Marcos",
"middle": [],
"last": "Zampieri",
"suffix": ""
},
{
"first": "Shervin",
"middle": [],
"last": "Malmasi",
"suffix": ""
},
{
"first": "Mihaela",
"middle": [],
"last": "Vela",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Liviu",
"suffix": ""
},
{
"first": "Josef",
"middle": [],
"last": "Dinu",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Van Genabith",
"suffix": ""
}
],
"year": 2017,
"venue": "Proceedings of the Second Workshop on Automated Semantic Analysis of Information in Legal Texts co-located with the 16th International Conference on Artificial Intelligence and Law",
"volume": "2143",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Octavia-Maria Sulea, Marcos Zampieri, Shervin Mal- masi, Mihaela Vela, Liviu P. Dinu, and Josef van Genabith. 2017. Exploring the use of text classifi- cation in the legal domain. In Proceedings of the Second Workshop on Automated Semantic Analysis of Information in Legal Texts co-located with the 16th International Conference on Artificial Intelli- gence and Law (ICAIL 2017), London, UK, June 16, 2017, volume 2143 of CEUR Workshop Proceedings. CEUR-WS.org.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Empirical study of deep learning for text classification in legal document review",
"authors": [
{
"first": "Fusheng",
"middle": [],
"last": "Wei",
"suffix": ""
},
{
"first": "Han",
"middle": [],
"last": "Qin",
"suffix": ""
},
{
"first": "Haozhen",
"middle": [],
"last": "Shi Ye",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Zhao",
"suffix": ""
}
],
"year": 2018,
"venue": "IEEE International Conference on Big Data, Big Data",
"volume": "",
"issue": "",
"pages": "3317--3320",
"other_ids": {
"DOI": [
"10.1109/BigData.2018.8622157"
]
},
"num": null,
"urls": [],
"raw_text": "Fusheng Wei, Han Qin, Shi Ye, and Haozhen Zhao. 2018. Empirical study of deep learning for text clas- sification in legal document review. In IEEE Inter- national Conference on Big Data, Big Data 2018, Seattle, WA, USA, December 10-13, 2018, pages 3317-3320. IEEE.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Few-shot text classification with triplet networks, data augmentation, and curriculum learning",
"authors": [
{
"first": "Jason",
"middle": [],
"last": "Wei",
"suffix": ""
},
{
"first": "Chengyu",
"middle": [],
"last": "Huang",
"suffix": ""
},
{
"first": "Soroush",
"middle": [],
"last": "Vosoughi",
"suffix": ""
},
{
"first": "Yu",
"middle": [],
"last": "Cheng",
"suffix": ""
},
{
"first": "Shiqi",
"middle": [],
"last": "Xu",
"suffix": ""
}
],
"year": 2021,
"venue": "Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2021",
"volume": "",
"issue": "",
"pages": "5493--5500",
"other_ids": {
"DOI": [
"10.18653/v1/2021.naacl-main.434"
]
},
"num": null,
"urls": [],
"raw_text": "Jason Wei, Chengyu Huang, Soroush Vosoughi, Yu Cheng, and Shiqi Xu. 2021. Few-shot text clas- sification with triplet networks, data augmentation, and curriculum learning. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2021, Online, June 6-11, 2021, pages 5493-5500. Association for Computational Linguistics.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Benchmarking zero-shot text classification: Datasets, evaluation and entailment approach",
"authors": [
{
"first": "Wenpeng",
"middle": [],
"last": "Yin",
"suffix": ""
},
{
"first": "Jamaal",
"middle": [],
"last": "Hay",
"suffix": ""
},
{
"first": "Dan",
"middle": [],
"last": "Roth",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing",
"volume": "",
"issue": "",
"pages": "3912--3921",
"other_ids": {
"DOI": [
"10.18653/v1/D19-1404"
]
},
"num": null,
"urls": [],
"raw_text": "Wenpeng Yin, Jamaal Hay, and Dan Roth. 2019. Benchmarking zero-shot text classification: Datasets, evaluation and entailment approach. In Proceedings of the 2019 Conference on Em- pirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019, pages 3912-3921. Association for Computational Linguistics.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"type_str": "figure",
"num": null,
"text": "Overall architecture of our approach.",
"uris": null
},
"TABREF0": {
"content": "<table><tr><td>Model</td><td colspan=\"3\">Precision Recall F1 Accuracy</td></tr><tr><td>Naive Bayes</td><td>0.78</td><td>0.48 0.60</td><td>0.75</td></tr><tr><td>MLP</td><td>0.66</td><td>0.70 0.68</td><td>0.75</td></tr><tr><td>SVM</td><td>0.76</td><td>0.67 0.71</td><td>0.79</td></tr><tr><td>S-BERT</td><td>0.72</td><td>0.69 0.70</td><td>0.78</td></tr><tr><td>Laser</td><td>0.75</td><td>0.68 0.71</td><td>0.79</td></tr><tr><td>Zero-Shot</td><td>0.48</td><td>0.75 0.59</td><td>0.60</td></tr><tr><td>Few-Shot(ours)</td><td>0.61</td><td>0.74 0.67</td><td>0.72</td></tr><tr><td colspan=\"4\">where \u2022 2 denotes the l 2 norm. The embeddings</td></tr><tr><td colspan=\"4\">e s , e s + and e s \u2212 denote the representation of the</td></tr><tr><td colspan=\"4\">anchor, positive and negative sentences from Equa-</td></tr><tr><td colspan=\"4\">tion 3 and e is Euler's number. The loss objective</td></tr><tr><td colspan=\"2\">ensures that when d +</td><td/><td/></tr></table>",
"html": null,
"text": "Performance of our few-shot learning model in comparison with other supervised and zero-shot learning methods.",
"type_str": "table",
"num": null
},
"TABREF1": {
"content": "<table/>",
"html": null,
"text": "Error Analysis: Examples where our few-shot model produces classification labels different from the gold labels.",
"type_str": "table",
"num": null
}
}
}
} |