File size: 69,396 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 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 | {
"paper_id": "2020",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T13:05:09.113031Z"
},
"title": "TermEval 2020: TALN-LS2N System for Automatic Term Extraction",
"authors": [
{
"first": "Amir",
"middle": [],
"last": "Hazem",
"suffix": "",
"affiliation": {
"laboratory": "LS2N -UMR CNRS 6004",
"institution": "Universit\u00e9 de Nantes",
"location": {
"country": "France"
}
},
"email": "amir.hazem@ls2n.fr"
},
{
"first": "M\u00e9ri\u00e8me",
"middle": [],
"last": "Bouhandi",
"suffix": "",
"affiliation": {
"laboratory": "LS2N -UMR CNRS 6004",
"institution": "Universit\u00e9 de Nantes",
"location": {
"country": "France"
}
},
"email": "merieme.bouhandi@ls2n.fr"
},
{
"first": "Florian",
"middle": [],
"last": "Boudin",
"suffix": "",
"affiliation": {
"laboratory": "LS2N -UMR CNRS 6004",
"institution": "Universit\u00e9 de Nantes",
"location": {
"country": "France"
}
},
"email": "florian.boudin@ls2n.fr"
},
{
"first": "B\u00e9atrice",
"middle": [],
"last": "Daille",
"suffix": "",
"affiliation": {
"laboratory": "LS2N -UMR CNRS 6004",
"institution": "Universit\u00e9 de Nantes",
"location": {
"country": "France"
}
},
"email": "beatrice.daille@ls2n.fr"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "Automatic terminology extraction is a notoriously difficult task aiming to ease effort demanded to manually identify terms in domain-specific corpora by automatically providing a ranked list of candidate terms. The main ways that addressed this task can be ranged in four main categories: (i) rule-based approaches, (ii) feature-based approaches, (iii) context-based approaches, and (iv) hybrid approaches. For this first TermEval shared task, we explore a feature-based approach, and a deep neural network multitask approach-BERT-that we fine-tune for term extraction. We show that BERT models (RoBERTa for English and CamemBERT for French) outperform other systems for French and English languages.",
"pdf_parse": {
"paper_id": "2020",
"_pdf_hash": "",
"abstract": [
{
"text": "Automatic terminology extraction is a notoriously difficult task aiming to ease effort demanded to manually identify terms in domain-specific corpora by automatically providing a ranked list of candidate terms. The main ways that addressed this task can be ranged in four main categories: (i) rule-based approaches, (ii) feature-based approaches, (iii) context-based approaches, and (iv) hybrid approaches. For this first TermEval shared task, we explore a feature-based approach, and a deep neural network multitask approach-BERT-that we fine-tune for term extraction. We show that BERT models (RoBERTa for English and CamemBERT for French) outperform other systems for French and English languages.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Automatic terminology extraction (ATE) is a very challenging task beneficial to a broad range of natural language processing applications, including machine translation, bilingual lexicon induction, thesauri construction (Lin, 1998; Wu and Zhou, 2003; van der Plas and Tiedemann, 2006; Hagiwara, 2008; Andrade et al., 2013; Rigouts Terryn et al., 2019) , to cite a few. Traditionally, this task is conducted by a terminologist, but hand-operated exploration, indexation, and maintenance of domain-specific corpora and terminologies is a costly enterprise. The automatization aims to ease effort demanded to manually identify terms in domain-specific corpora by automatically providing a ranked list of candidate terms. Despite being a well-established research domain for decades, NLP methods still fail to meet human standards, and ATE is still considered an unsolved problem with considerable room for improvement. If it is generally admitted that terms are single words or multiword expressions representing domain-specific concepts and that terminologies are the body of terms used with a particular domain, the lack of annotated data and agreement between researchers make ATE evaluation very difficult (Terryn et al., 2018) . In order to gather researchers around a common evaluation scheme, TermEval shared task (Rigouts Terryn et al., 2019) offers a unified framework aiming a better ATE's comprehension and analysis 1 . The shared task provides four data sets: Corruption, dressage, wind energy and heart failure; in three languages: English, French and Dutch. With the advance of neural network language models and following the current trend and excellent results obtained by transformer architecture on other NLP tasks, we have decided to experiment and compare two classification methods, one feature-based and the BERT-based. We show that BERT models (RoBERTa for English and Camem-BERT for French) outperform other systems for French and English languages. Also, the feature-based approach shows competitive results.",
"cite_spans": [
{
"start": 221,
"end": 232,
"text": "(Lin, 1998;",
"ref_id": "BIBREF10"
},
{
"start": 233,
"end": 251,
"text": "Wu and Zhou, 2003;",
"ref_id": "BIBREF22"
},
{
"start": 252,
"end": 285,
"text": "van der Plas and Tiedemann, 2006;",
"ref_id": "BIBREF17"
},
{
"start": 286,
"end": 301,
"text": "Hagiwara, 2008;",
"ref_id": "BIBREF6"
},
{
"start": 302,
"end": 323,
"text": "Andrade et al., 2013;",
"ref_id": "BIBREF1"
},
{
"start": 324,
"end": 352,
"text": "Rigouts Terryn et al., 2019)",
"ref_id": "BIBREF14"
},
{
"start": 1208,
"end": 1229,
"text": "(Terryn et al., 2018)",
"ref_id": "BIBREF15"
},
{
"start": 1319,
"end": 1348,
"text": "(Rigouts Terryn et al., 2019)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "1 https://termeval.ugent.be/",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "The shared task provides four data sets. Three of them are dedicated to the training phase: corruption, dressage and wind energy, and one to the test phase: heart failure. All the corpora are provided in three languages: English, French and Dutch. The data sets are described in detail in (Rigouts Terryn et al., 2019) . Five teams have participated in the TermEval shared task. All teams submitted results for English, three submitted for French and two for Dutch. We submitted results for the French and English data sets. The Precision, recall, and F1-score were calculated twice: once including and once excluding Named Entities.",
"cite_spans": [
{
"start": 289,
"end": 318,
"text": "(Rigouts Terryn et al., 2019)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Task Description",
"sec_num": "2."
},
{
"text": "We present in this section the two experimented approaches during the training phase that is: (i) the feature-based and, (ii) the BERT-based approaches. For the test phase, the submitted results are those of BERT approach only. However, we also report the obtained results of the feature-based approach for comparison.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Proposed System",
"sec_num": "3."
},
{
"text": "3.1. Feature-based Approach 3.1.1. Feature Extraction Classical methods for extracting terms from corpora often consist of three major steps: the first one uses some linguistic filtering, the second one consists of describing the candidates through different features in order to give them a weight indicating the degree of confidence that they are indeed a term, and the third is more of a selection phase. As for the first step, we know that, often, the first requirement is for a term to be a noun phrase, and our main morphosyntactic pattern is defined (primarily by observing recurrent patterns in the given reference lists of terms): a noun or nouns (or proper nouns), which might be preceded or followed by adjectives (vertical axis wind turbine), or of-genitives (United States of America). These patterns are then passed to spaCy's rule-matching engine 2 to extract a list of candidate terms. Once our candidate terms are extracted, we processed to the second step, and we assign to each one of them linguistic, stylistic, statistic, and distributional descriptors that might help us get insights as to the nature of terms (Table 1) . In this work, beyond the common statistical descriptors, we wanted to focus on different measures of specificity and termhood, since we know that a term is much more common and essential in a specialized corpus than it is in a general domain corpus. Termhood is defined by (Kageura and Umino, 1996) as \"the degree to which a linguistic unit is related to domain-specific context\":",
"cite_spans": [
{
"start": 1417,
"end": 1442,
"text": "(Kageura and Umino, 1996)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [
{
"start": 1132,
"end": 1141,
"text": "(Table 1)",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Proposed System",
"sec_num": "3."
},
{
"text": "\u2022 Measures of specificity and termhood a) with a the term, f D (a) the term frequency in the specialized corpus and f G (a) its outof-domain frequency.",
"cite_spans": [
{
"start": 39,
"end": 41,
"text": "a)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Proposed System",
"sec_num": "3."
},
{
"text": "-Specificity (Specif icity): Specif icity(a) = 2 \u2022 f D (a)\u00d7f G (a) f D (a)+f G (",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Proposed System",
"sec_num": "3."
},
{
"text": "-Term's relation to Context (W rel ): W rel (a) = (0.5 + ((W L \u2022 T F (a) M axT F ) + P L)) + (0.5 + ((W R \u2022 T F (a) M axT F ) + P R))",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Proposed System",
"sec_num": "3."
},
{
"text": "with T F (a) the term frequency in the document, M axT F the frequency of the most frequent word, W L (or [W R]) is the ratio between the number of different words that co-occur with the candidate term (on the left [right] side) and the total number of words that it co-occurs with. P L (or [P R]) is the ratio between the number of different words that co-occur with the candidate term (on the left [right] side) and the M axT F . W rel measures the singularity of the term a in the corpus and quantifies the extent to which a term resembles the characteristics of a stopword. The more a candidate word co-occurs with different words, the more likely it is to be unimportant in the document.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Proposed System",
"sec_num": "3."
},
{
"text": "-Cvalue (Cval): Cval(a) = log 2 |a| \u2022 (f (a) \u2212 1 P (Ta)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Proposed System",
"sec_num": "3."
},
{
"text": "n\u2208Ta f (n)) with f (a) the frequency of term a, |a| the number of words in a, T a is the set of extracted candidate terms that contain a, P (T a ) is the total number of longer candidate terms that contain a.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Proposed System",
"sec_num": "3."
},
{
"text": "-Termhood (T H): W (a) = f 2 a n \u2022 n 1 (log f n,D N D \u2212 f n,R N R ) with f 2",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Proposed System",
"sec_num": "3."
},
{
"text": "a the absolute frequency of the word in the domain-specific corpus, n the number of words in a, As for the last step, classification is conducted to select the terms using these features.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Proposed System",
"sec_num": "3."
},
{
"text": "Boosting is a classification method that consists of iteratively learning several classifiers whose individual weights are corrected as they go along to better predict difficult values. The classifiers are then weighted according to their performance and aggregated iteratively. We use the XGBoost model (eXtreme Gradient Boosting) (Chen and Guestrin, 2016) , and we feed it our feature vectors after being normalized using sklearn 3 standard scaler, which transforms an x value into a z = x\u2212u s value, with u being Table of Features the mean of the x and s its standard deviation. However, these features can be more or less essential to characterize our terms. After several tests, we have empirically determined that only the elements that correlate at more than a certain threshold (mean correlation) with our target class are retained for classification (bolded in 1).",
"cite_spans": [
{
"start": 332,
"end": 357,
"text": "(Chen and Guestrin, 2016)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [
{
"start": 516,
"end": 533,
"text": "Table of Features",
"ref_id": null
}
],
"eq_spans": [],
"section": "Classification",
"sec_num": "3.1.2."
},
{
"text": "BERT has proven to be efficient in many downstream NLP tasks (Devlin et al., 2018) including next sentence prediction, question answering and named entity recognition (NER). It can also be used for feature extraction or classification. Prior to the emergence of transformer-based architectures like BERT, several deep learning architectures for terminology extraction have been proposed. Wang et al. (2016) introduce a weakly-supervised classification-based approach. Amjadian et al. (2016) leverage local and global embeddings to encapsulate the meaning and behavior of the term for the classification step, although they only work with unigram terms. We must note that exploring these architectures is not the focus of this work; we mainly want to observe how BERTbased models can be used for ATE and how they perform in comparison to more traditional feature-based methods.",
"cite_spans": [
{
"start": 61,
"end": 82,
"text": "(Devlin et al., 2018)",
"ref_id": "BIBREF5"
},
{
"start": 388,
"end": 406,
"text": "Wang et al. (2016)",
"ref_id": "BIBREF20"
},
{
"start": 468,
"end": 490,
"text": "Amjadian et al. (2016)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "BERT",
"sec_num": "3.2."
},
{
"text": "In order to do that, we use different versions of BERT as a binary classifier for term prediction. For English, we use RoBERTa (Liu et al., 2019) , which is a model built based on BERT but modifies key hyperparameters in the original BERT model, eliminating its nextsentence pretraining objective and training the model with much larger mini-batches and more substantial learning rates, leading to more solid downstream task performance. For French, we use CamemBERT (Martin et al., 2019) , the French version of the BERT model. For both languages, we Table 2 : Terminology extraction scores (%) obtained on the training data sets. BERT3 for instance, stands for BERT using ngrams of length 3 for training.",
"cite_spans": [
{
"start": 127,
"end": 145,
"text": "(Liu et al., 2019)",
"ref_id": "BIBREF11"
},
{
"start": 467,
"end": 488,
"text": "(Martin et al., 2019)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [
{
"start": 552,
"end": 559,
"text": "Table 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "BERT",
"sec_num": "3.2."
},
{
"text": "will use pre-trained models, and both of them are fine-tuned during the classification. The general objectives BERT is trained on gives the model an innate sentence classification capability. The main idea is to associate each term with its context. Hence, by analogy to the next sentence prediction, the first sentence given to BERT is the one which contains the term, and the sentence to predict is the term itself. For training, we feed the model with all the context/term pairs that appear in the corpus as positive examples. The negative examples are generated randomly. Given the following sentence: \"this is the first global instrument in the fight against corruption\", corruption is annotated as a positive example (term) and a randomly chosen word or n-gram, global for instance, is annotated as a negative example. It is important to highlight the fact that the negative examples are all the n-grams that do not appear in the training evaluation term list. Also, the number of negative examples is equal to the number of positive ones.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "BERT",
"sec_num": "3.2."
},
{
"text": "Hand-engineering features is a challenging assignment, even more so for a task as challenging as extracting terms from domain-specific corpora and finding features to capture the right characteristics for each term and stay relevant with any corpora in hand. We can observe, from our results (table 2) , that we often fail to find a good trade-off between recall and precision. As a matter of fact, with features as strict as these, we often find ourselves with correct precision and quite a weak recall.",
"cite_spans": [],
"ref_spans": [
{
"start": 292,
"end": 301,
"text": "(table 2)",
"ref_id": null
}
],
"eq_spans": [],
"section": "Experiments and Results",
"sec_num": "4."
},
{
"text": "For the fine-tuning phase of BERT, we used the simpletransformers 4 library and its default parameters setting. For English, we used RoBERTa with n-gram size of four while for French, we used CamemBERT with n-gram size of five.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "BERT Settings",
"sec_num": "4.1."
},
{
"text": "We started with the hypothesis that the features of a term noun phrase must be different from the features of a nonterm noun phrase and that the features that characterize these terms must be valid from one corpus to the other. However, we can clearly see that the main problem encountered with the feature-based method is that the features learned by the model are hardly transferable from one corpus to another, as the notion of the relevance of each candidate term changes from one application area to another, and from one domain to another. Hard-coded features learned on one corpus do not transfer well to another during classification, since not only are the texts and domains vary greatly, but even the range of the values for the noun phrases features in the different corpora can vary enormously (see figures 1, 2, 3) . Going for a feature-less method seems to be a nice direction to explore (table 4) . Our experiments with BERT, even if they were somewhat successful, were a bit abrupt, since we consider all the n-grams as potential candidates, without prior filtering. We end up, after classification, with false positives in our list, such as phrases beginning or ending with pronouns or conjunctions. One of the reasons that pushed us first to test this configuration without prior noun phrases filtering was our fear of losing potential positive candidates (we can in table 4 see that recall post-filtering is average). Future work will incorporate syntactic information into this BERT process in order to get better precision. Table 3 : Official results on the heart failure test set(%). The results on the test set are consistent with the results on the training corpora. The same patterns can be observed, and results on the test set are in the same range. Based on the F1-score, our approach represented by TALN-LS2N using BERT obtained the best results of the competition. However, we see that in terms of precision, the NYU team obtained the best results for English. Overall, feature-based and BERT-based approaches exhibit similar performance on the French test set while for English, BERT is more accurate. Further experiments are certainly needed to improve Table 4 : Results on the heart failure test set(%) using BERT with different ngram's size. BERT3 for instance, stands for BERT using ngrams of length 3 for training. both methods. However, the capability of BERT (certainly thanks to its attention mechanism) to learn hidden features suggests less effort is needed compared to the feature-based approach, which requires more efforts in the design of the features. Also, the n-gram size used in BERT was fixed empirically based on the development data sets. Further anal- ysis is needed to make our approach n-gram independent for better term length coverage. Indeed, we limited our system outputs to 4-grams for English and 5-grams for French, which did not allow the extraction of longer terms. Finally, recent work has shown several improvements of BERT such as StructBERT (Wang et al., 2020) and T5 (Raffel et al., 2019) . These recent state-of-the-art approaches can, in the future, be used to further improve the results of ATE.",
"cite_spans": [
{
"start": 3009,
"end": 3028,
"text": "(Wang et al., 2020)",
"ref_id": "BIBREF21"
},
{
"start": 3036,
"end": 3057,
"text": "(Raffel et al., 2019)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [
{
"start": 811,
"end": 827,
"text": "figures 1, 2, 3)",
"ref_id": "FIGREF1"
},
{
"start": 902,
"end": 911,
"text": "(table 4)",
"ref_id": null
},
{
"start": 1545,
"end": 1552,
"text": "Table 3",
"ref_id": null
},
{
"start": 2185,
"end": 2192,
"text": "Table 4",
"ref_id": null
}
],
"eq_spans": [],
"section": "Experiments on the Training Data Sets",
"sec_num": "4.2."
},
{
"text": "Term extraction has been a very active field of research for many decades. Methods based solely on linguistic analysis and patterns have given way to new statistical, machine, and deep learning methods. We conducted several experiments using classical hand-engineered features-based methods in order to find the best way to extract terms in several specialized domains. These models that combine linguistic, statistical and distributional descriptors suggest that the relation between test and training corpora are of central importance. Moreover, we have seen that it is only natural for the very notion of termhood in different domains to be more pragmatic than theoretical. We then proposed a BERT-based classification approach that outperformed classical methods on this shared task. This contribution is setting a new, simple and strong baseline for terminology extraction. However, the overall results of this task are average at best, and much room is left for improvement.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "5."
},
{
"text": "https://spacy.io/",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "https://github.com/ThilinaRajapakse/ simpletransformers",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Local-global vectors to improve unigram terminology extraction",
"authors": [
{
"first": "E",
"middle": [],
"last": "Amjadian",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Inkpen",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Paribakht",
"suffix": ""
},
{
"first": "F",
"middle": [],
"last": "Faez",
"suffix": ""
}
],
"year": 2016,
"venue": "Proceedings of the 5th International Workshop on Computational Terminology (Com-puterm2016)",
"volume": "",
"issue": "",
"pages": "2--11",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Amjadian, E., Inkpen, D., Paribakht, T., and Faez, F. (2016). Local-global vectors to improve unigram termi- nology extraction. In Proceedings of the 5th Interna- tional Workshop on Computational Terminology (Com- puterm2016), pages 2-11, Osaka, Japan, December. The COLING 2016 Organizing Committee.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Synonym acquisition using bilingual comparable corpora",
"authors": [
{
"first": "D",
"middle": [],
"last": "Andrade",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Tsuchida",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Onishi",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Ishikawa",
"suffix": ""
}
],
"year": 2013,
"venue": "International Joint Conference on Natural Language Processing (IJCNLP'13)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Andrade, D., Tsuchida, M., Onishi, T., and Ishikawa, K. (2013). Synonym acquisition using bilingual compara- ble corpora. In International Joint Conference on Natu- ral Language Processing (IJCNLP'13), Nagoya, Japan.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Keyword extraction using auto-associative neural networks",
"authors": [
{
"first": "G",
"middle": [],
"last": "Aquino",
"suffix": ""
},
{
"first": "W",
"middle": [],
"last": "Hasperu\u00e9",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Lanzarini",
"suffix": ""
}
],
"year": 2014,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Aquino, G., Hasperu\u00e9, W., and Lanzarini, L. (2014). Key- word extraction using auto-associative neural networks.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "A text feature based automatic keyword extraction method for single documents",
"authors": [
{
"first": "R",
"middle": [],
"last": "Campos",
"suffix": ""
},
{
"first": "V",
"middle": [],
"last": "Mangaravite",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Pasquali",
"suffix": ""
},
{
"first": "A",
"middle": [
"M"
],
"last": "Jorge",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Nunes",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Jatowt",
"suffix": ""
}
],
"year": 2018,
"venue": "Advances in Information Retrieval",
"volume": "",
"issue": "",
"pages": "684--691",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Campos, R., Mangaravite, V., Pasquali, A., Jorge, A. M., Nunes, C., and Jatowt, A. (2018). A text feature based automatic keyword extraction method for single docu- ments. In Gabriella Pasi, et al., editors, Advances in In- formation Retrieval, pages 684-691, Cham. Springer In- ternational Publishing.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Xgboost: A scalable tree boosting system",
"authors": [
{
"first": "T",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Guestrin",
"suffix": ""
}
],
"year": 2016,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Chen, T. and Guestrin, C. (2016). Xgboost: A scalable tree boosting system. CoRR, abs/1603.02754.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Bert: Pre-training of deep bidirectional transformers for language understanding",
"authors": [
{
"first": "J",
"middle": [],
"last": "Devlin",
"suffix": ""
},
{
"first": "M.-W",
"middle": [],
"last": "Chang",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Toutanova",
"suffix": ""
}
],
"year": 2018,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1810.04805"
]
},
"num": null,
"urls": [],
"raw_text": "Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. (2018). Bert: Pre-training of deep bidirectional trans- formers for language understanding. arXiv preprint arXiv:1810.04805.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "A supervised learning approach to automatic synonym identification based on distributional features",
"authors": [
{
"first": "M",
"middle": [],
"last": "Hagiwara",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of the ACL-08: HLT Student Research Workshop",
"volume": "",
"issue": "",
"pages": "1--6",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hagiwara, M. (2008). A supervised learning approach to automatic synonym identification based on distributional features. In Proceedings of the ACL-08: HLT Student Research Workshop, pages 1-6, Columbus, Ohio, June. Association for Computational Linguistics.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Automatic keyphrase extraction: A survey of the state of the art",
"authors": [
{
"first": "K",
"middle": [
"S"
],
"last": "Hasan",
"suffix": ""
},
{
"first": "V",
"middle": [],
"last": "Ng",
"suffix": ""
}
],
"year": 2014,
"venue": "Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics",
"volume": "1",
"issue": "",
"pages": "1262--1273",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hasan, K. S. and Ng, V. (2014). Automatic keyphrase ex- traction: A survey of the state of the art. In Proceedings of the 52nd Annual Meeting of the Association for Com- putational Linguistics (Volume 1: Long Papers), pages 1262-1273, Baltimore, Maryland, June. Association for Computational Linguistics.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "A statistical interpretation of term specificity and its application in retrieval",
"authors": [
{
"first": "K",
"middle": [
"S"
],
"last": "Jones",
"suffix": ""
}
],
"year": 2004,
"venue": "Journal of Documentation",
"volume": "60",
"issue": "5",
"pages": "493--502",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jones, K. S. (2004). A statistical interpretation of term specificity and its application in retrieval. Journal of Documentation, 60(5):493-502.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Methods of automatic term recognition: A review",
"authors": [
{
"first": "K",
"middle": [],
"last": "Kageura",
"suffix": ""
},
{
"first": "B",
"middle": [],
"last": "Umino",
"suffix": ""
}
],
"year": 1996,
"venue": "Terminology. International Journal of Theoretical and Applied Issues in Specialized Communication",
"volume": "3",
"issue": "2",
"pages": "259--289",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kageura, K. and Umino, B. (1996). Methods of automatic term recognition: A review. Terminology. International Journal of Theoretical and Applied Issues in Specialized Communication, 3(2):259-289.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Automatic retrieval and clustering of similar words",
"authors": [
{
"first": "D",
"middle": [],
"last": "Lin",
"suffix": ""
}
],
"year": 1998,
"venue": "Proceedings of the 36th Annual Meeting of the Association for Computational Linguistics and 17th International Conference on Computational Linguistics",
"volume": "2",
"issue": "",
"pages": "768--774",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Lin, D. (1998). Automatic retrieval and clustering of simi- lar words. In Proceedings of the 36th Annual Meeting of the Association for Computational Linguistics and 17th International Conference on Computational Linguistics -Volume 2, ACL '98, pages 768-774, Stroudsburg, PA, USA. Association for Computational Linguistics.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Roberta: A robustly optimized bert pretraining approach",
"authors": [
{
"first": "Y",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Ott",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Goyal",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Du",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Joshi",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "O",
"middle": [],
"last": "Levy",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Lewis",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Zettlemoyer",
"suffix": ""
},
{
"first": "V",
"middle": [],
"last": "Stoyanov",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1907.11692"
]
},
"num": null,
"urls": [],
"raw_text": "Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., and Stoyanov, V. (2019). Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "CamemBERT: a Tasty French Language Model. arXiv e-prints",
"authors": [
{
"first": "L",
"middle": [],
"last": "Martin",
"suffix": ""
},
{
"first": "B",
"middle": [],
"last": "Muller",
"suffix": ""
},
{
"first": "P",
"middle": [
"J"
],
"last": "Ortiz Su\u00e1rez",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Dupont",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Romary",
"suffix": ""
},
{
"first": "\u00c9",
"middle": [],
"last": "Villemonte De La Clergerie",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Seddah",
"suffix": ""
},
{
"first": "B",
"middle": [],
"last": "Sagot",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1911.03894"
]
},
"num": null,
"urls": [],
"raw_text": "Martin, L., Muller, B., Ortiz Su\u00e1rez, P. J., Dupont, Y., Ro- mary, L., Villemonte de la Clergerie,\u00c9., Seddah, D., and Sagot, B. (2019). CamemBERT: a Tasty French Lan- guage Model. arXiv e-prints, page arXiv:1911.03894, Nov.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Exploring the limits of transfer learning with a unified textto",
"authors": [
{
"first": "C",
"middle": [],
"last": "Raffel",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Shazeer",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Roberts",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Narang",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Matena",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Zhou",
"suffix": ""
},
{
"first": "W",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "P",
"middle": [
"J"
],
"last": "Liu",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., and Liu, P. J. (2019). Ex- ploring the limits of transfer learning with a unified text- to-text transformer. arXiv e-prints.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "In no uncertain terms: a dataset for monolingual and multilingual automatic term extraction from comparable corpora",
"authors": [
{
"first": "A",
"middle": [],
"last": "Rigouts Terryn",
"suffix": ""
},
{
"first": "V",
"middle": [],
"last": "Hoste",
"suffix": ""
},
{
"first": "E",
"middle": [],
"last": "Lefever",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Rigouts Terryn, A., Hoste, V., and Lefever, E. (2019). In no uncertain terms: a dataset for monolingual and multilin- gual automatic term extraction from comparable corpora. In Language Resources and Evaluation.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "A Gold Standard for Multilingual Automatic Term Extraction from Comparable Corpora: Term Structure and Translation Equivalents",
"authors": [
{
"first": "A",
"middle": [
"R"
],
"last": "Terryn",
"suffix": ""
},
{
"first": "V",
"middle": [],
"last": "Hoste",
"suffix": ""
},
{
"first": "E",
"middle": [],
"last": "Lefever",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Terryn, A. R., Hoste, V., and Lefever, E. (2018). A Gold Standard for Multilingual Automatic Term Extraction from Comparable Corpora: Term Structure and Trans- lation Equivalents. In Proceedings of the Eleventh Inter- national Conference on Language Resources and Eval- uation (LREC 2018), Miyazaki, Japan, May 7-12, 2018.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Finding synonyms using automatic word alignment and measures of distributional similarity",
"authors": [
{
"first": "L",
"middle": [],
"last": "Van Der Plas",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Tiedemann",
"suffix": ""
}
],
"year": 2006,
"venue": "21st International Conference on Computational Linguistics and 44th Annual Meeting of the Association for Computational Linguistics ACL'06",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "van der Plas, L. and Tiedemann, J. (2006). Finding syn- onyms using automatic word alignment and measures of distributional similarity. In 21st International Con- ference on Computational Linguistics and 44th Annual Meeting of the Association for Computational Linguis- tics ACL'06, Sydney, Australia.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Bilingual term recognition revisited: The bag-of-equivalents term alignment approach and its evaluation",
"authors": [
{
"first": "S",
"middle": [],
"last": "Vintar",
"suffix": ""
}
],
"year": 2010,
"venue": "Terminology. International Journal of Theoretical and Applied Issues in Specialized Communication",
"volume": "16",
"issue": "2",
"pages": "141--158",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Vintar, S. (2010). Bilingual term recognition revisited: The bag-of-equivalents term alignment approach and its evaluation. Terminology. International Journal of The- oretical and Applied Issues in Specialized Communica- tion, 16(2):141-158.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Term extraction through unithood and termhood unification",
"authors": [
{
"first": "T",
"middle": [],
"last": "Vu",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Aw",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Zhang",
"suffix": ""
}
],
"year": 2008,
"venue": "IJCNLP",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Vu, T., Aw, A., and Zhang, M. (2008). Term extraction through unithood and termhood unification. In IJCNLP.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Featureless domain-specific term extraction with minimal labelled data",
"authors": [
{
"first": "R",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "W",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Mcdonald",
"suffix": ""
}
],
"year": 2016,
"venue": "Proceedings of the Australasian Language Technology Association Workshop",
"volume": "",
"issue": "",
"pages": "103--112",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Wang, R., Liu, W., and McDonald, C. (2016). Feature- less domain-specific term extraction with minimal la- belled data. In Proceedings of the Australasian Lan- guage Technology Association Workshop 2016, pages 103-112, Melbourne, Australia, December.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Structbert: Incorporating language structures into pre-training for deep language understanding",
"authors": [
{
"first": "W",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "B",
"middle": [],
"last": "Bi",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Yan",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Wu",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Xia",
"suffix": ""
},
{
"first": "Z",
"middle": [],
"last": "Bao",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Peng",
"suffix": ""
},
{
"first": "Si",
"middle": [],
"last": "",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "",
"suffix": ""
}
],
"year": 2020,
"venue": "International Conference on Learning Representations",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Wang, W., Bi, B., Yan, M., Wu, C., Xia, J., Bao, Z., Peng, L., and Si, L. (2020). Structbert: Incorporating language structures into pre-training for deep language understanding. In International Conference on Learning Representations.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "Optimizing synonym extraction using monolingual and bilingual resources",
"authors": [
{
"first": "H",
"middle": [],
"last": "Wu",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Zhou",
"suffix": ""
}
],
"year": 2003,
"venue": "Proceedings of the second international workshop on Paraphrasing",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Wu, H. and Zhou, M. (2003). Optimizing synonym ex- traction using monolingual and bilingual resources. In In Proceedings of the second international workshop on Paraphrasing, page 72.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"text": "f n,D N D the frequency of each constituant of the term in the domain-specific corpus ( f n,R N R for the general domain) relative of the size of the corpora (in tokens).",
"type_str": "figure",
"num": null,
"uris": null
},
"FIGREF1": {
"text": "Range of the TFIDF values on all the corpora for English",
"type_str": "figure",
"num": null,
"uris": null
},
"FIGREF2": {
"text": "Range of the Specificity values on all the corpora for English",
"type_str": "figure",
"num": null,
"uris": null
},
"FIGREF3": {
"text": "Range of the Relevance values on all the corpora for English",
"type_str": "figure",
"num": null,
"uris": null
},
"TABREF1": {
"type_str": "table",
"text": "",
"content": "<table/>",
"html": null,
"num": null
},
"TABREF2": {
"type_str": "table",
"text": ".42 76.2 5.60 6.60 68.4 11.7 1.50 76.1 2.40 2.60 61.5 5.10 6.80 50.2 10.7 1.20 55.8 2.20 Features 40.6 16.4 23.7 38.7 19.4 25.5 51.1 10.9 17.2 39.4 17.6 24.4 38.7 19.1 25.4 51.2 10.48.4 41.7 9.08 78.1 16.2 9.11 82.5 16.4 26.6 43.5 33.1 7.27 76.8 13.2 7.21 81.8 13.2",
"content": "<table><tr><td/><td/><td/><td/><td/><td/><td/><td/><td colspan=\"2\">English</td><td/><td/><td/><td/><td/><td/><td/><td/></tr><tr><td/><td/><td/><td/><td>NES</td><td/><td/><td/><td/><td/><td/><td/><td/><td>ANN</td><td/><td/><td/><td/></tr><tr><td>Tools</td><td>Corp</td><td/><td/><td>Equi</td><td/><td/><td>Wind</td><td/><td/><td>Corp</td><td/><td/><td>Equi</td><td/><td/><td>Wind</td><td/></tr><tr><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td></tr><tr><td colspan=\"18\">Patterns 28 17.4</td></tr><tr><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td></tr><tr><td colspan=\"18\">BERT3 27.1 41.4 32.8 28.4 82.0 42.2 22.2 81.1 34.8 18.4 35.6 24.2 20.5 80.6 32.7 16.0 82.4 26.9</td></tr><tr><td colspan=\"18\">BERT4 28.5 38.9 32.9 26.5 85.0 40.4 21.3 83.8 33.9 17.8 30.7 22.5 19.1 83.5 31.1 15.4 85.6 26.2</td></tr><tr><td colspan=\"18\">BERT5 25.5 42.9 32.0 27.3 80.5 40.8 19.9 93.5 32.8 16.7 35.7 22.8 19.4 78.1 31.1 14.7 90.5 25.4</td></tr><tr><td colspan=\"18\">BERT6 25.6 57.6 35.5 27.6 84.3 41.6 16.9 89.9 28.5 18.7 53.4 27.7 19.8 82.6 32.0 12.4 93.0 22.0</td></tr><tr><td/><td/><td/><td/><td/><td/><td/><td/><td colspan=\"2\">French</td><td/><td/><td/><td/><td/><td/><td/><td/></tr><tr><td/><td/><td/><td/><td>NES</td><td/><td/><td/><td/><td/><td/><td/><td/><td>ANN</td><td/><td/><td/><td/></tr><tr><td>Tools</td><td>Corp</td><td/><td/><td>Equi</td><td/><td/><td>Wind</td><td/><td/><td>Corp</td><td/><td/><td>Equi</td><td/><td/><td>Wind</td><td/></tr><tr><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td></tr><tr><td colspan=\"18\">Patterns 3.08 74.6 5.93 5.26 67.9 9.76 1.69 77.2 3.31 3.69 72.8 7.08 6.75 71.3 12.3 2.09 76.3 4.08</td></tr><tr><td colspan=\"18\">Features 30.9 25.1 27.7 54.3 11.5 19.6 46.4 16.4 24.2 31.5 25.9 28.4 54.3 11.5 19.1 45.4 16.4 24.1</td></tr><tr><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td></tr><tr><td colspan=\"18\">BERT3 41.5 23.1 29.6 25.8 61.7 36.4 18.2 58.9 27.8 23.2 15.9 18.9 19.8 58.1 29.5 13.6 55.5 21.9</td></tr><tr><td colspan=\"18\">BERT4 27.9 48.5 35.4 24.8 63.0 35.6 17.9 67.4 28.2 20.8 44.9 28.4 18.9 59.1 28.7 13.7 64.8 22.6</td></tr><tr><td colspan=\"18\">BERT5 30.1 57.2 39.5 20.1 71.2 31.4 11.3 76.9 19.8 23.1 54.4 32.5 15.7 68.6 25.6 8.89 75.2 15.9</td></tr><tr><td>BERT6 36.7</td><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/></tr></table>",
"html": null,
"num": null
}
}
}
} |