File size: 77,045 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 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 |
{
"paper_id": "2021",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T02:10:55.313921Z"
},
"title": "Domain adaptation in practice: Lessons from a real-world information extraction pipeline",
"authors": [
{
"first": "Timothy",
"middle": [],
"last": "Miller",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Harvard Medical School",
"location": {}
},
"email": "timothy.miller@childrens.harvard.edu"
},
{
"first": "Egoitz",
"middle": [],
"last": "Laparra",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Arizona",
"location": {}
},
"email": "laparra@arizona.edu"
},
{
"first": "Steven",
"middle": [],
"last": "Bethard",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Arizona",
"location": {}
},
"email": "bethard@arizona.edu"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "Advances in transfer learning and domain adaptation have raised hopes that oncechallenging NLP tasks are ready to be put to use for sophisticated information extraction needs. In this work, we describe an effort to do just that-combining state-of-the-art neural methods for negation detection, document time relation extraction, and aspectual link prediction, with the eventual goal of extracting drug timelines from electronic health record text. We train on the THYME colon cancer corpus and test on both the THYME brain cancer corpus and an internal corpus, and show that performance of the combined systems is unacceptable despite good performance of individual systems. Although domain adaptation shows improvements on each individual system, the model selection problem is a barrier to improving overall pipeline performance.",
"pdf_parse": {
"paper_id": "2021",
"_pdf_hash": "",
"abstract": [
{
"text": "Advances in transfer learning and domain adaptation have raised hopes that oncechallenging NLP tasks are ready to be put to use for sophisticated information extraction needs. In this work, we describe an effort to do just that-combining state-of-the-art neural methods for negation detection, document time relation extraction, and aspectual link prediction, with the eventual goal of extracting drug timelines from electronic health record text. We train on the THYME colon cancer corpus and test on both the THYME brain cancer corpus and an internal corpus, and show that performance of the combined systems is unacceptable despite good performance of individual systems. Although domain adaptation shows improvements on each individual system, the model selection problem is a barrier to improving overall pipeline performance.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Advances in machine learning methods and the release of annotated datasets of clinical texts (Uzuner et al., 2011; Styler IV et al., 2014) in the past decade has led to an increase of available clinical NLP systems for interesting tasks. Recent advances in pre-trained models (Devlin et al., 2019; Liu et al., 2019) have made ever more accurate clinical NLP systems possible. Unsupervised domain adaptation algorithms (e.g., Ziser and Reichart (2019) ) have made it possible to reduce performance degradation when applying trained models to new domains. The great promise of these developments is that these methods can be combined into pipelines that allow for sophisticated information extraction capabilities for downstream clinical use cases. Rather than building one-off datasets for each complex downstream task that arises, standard NLP components could potentially be used as \"Lego\"-style building blocks that allow for flexibly approaching new tasks as they arise.",
"cite_spans": [
{
"start": 93,
"end": 114,
"text": "(Uzuner et al., 2011;",
"ref_id": "BIBREF16"
},
{
"start": 115,
"end": 138,
"text": "Styler IV et al., 2014)",
"ref_id": "BIBREF15"
},
{
"start": 276,
"end": 297,
"text": "(Devlin et al., 2019;",
"ref_id": "BIBREF1"
},
{
"start": 298,
"end": 315,
"text": "Liu et al., 2019)",
"ref_id": null
},
{
"start": 425,
"end": 450,
"text": "Ziser and Reichart (2019)",
"ref_id": "BIBREF20"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "However, the existence of the building blocks alone does not solve this problem. Combining individual components into NLP pipelines can lead to cascading errors (Finkel et al., 2006) . The true error rate for structured extraction tasks is potentially as high as the sum of the component tasks' errors. For example, if the goal is to extract normalized concepts with assertion status, the concept error can come from normalization error, negation detection error, uncertainty detection error, etc, and the errors may not be correlated. These problems are exacerbated in the common case where individual components are trained on data from different domains, and tested on data from yet another domain.",
"cite_spans": [
{
"start": 161,
"end": 182,
"text": "(Finkel et al., 2006)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In this work, we quantitatively examine the issues described above in the context of extracting drug temporality signatures, with the goal of understanding drug start and stop events. We approach this task with the combination of three sub-tasks: 1) the temporal relation of these mentions to the document creation time (DocTimeRel), 2) negation status of the mention, and 3) aspectual link relations of the mention (e.g., is it being described as starting or stopping). Figure 1 shows an example sentence with a drug mention, that demonstrates how the three tasks work together to establish the status of that drug in that patient. Successfully solving this task is beneficial for understanding patient treatment course, and enabling more causal understanding in important tasks such as adverse drug event detection or relating medication courses to outcomes.",
"cite_spans": [],
"ref_spans": [
{
"start": 471,
"end": 479,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We first set state-of-the-art benchmarks for three tasks on the THYME corpus by fine-tuning large pre-trained transformer models (Devlin et al., 2019; Liu et al., 2019) . We then examine how the performance of individual systems degrades when moving from the training data to our target data (a pediatric cardiology cohort), and how the overall system performs when combining multiple systems Additionally , as the patient has preserved ejection fraction , no prior history of embolic phenomena , and no significant valvular disease , it would be acceptable for him to remain off e Coumadin /e for the interim . Figure 1 : An example sentence with highlighted drug name Coumadin to be classified for all three tasks. The gold standard has this drug mention classified as negated, with DocTimeRel=OVERLAP, and ALINK=CONTINUES. These three facts can be used to understand that the patient is not on the drug now or going forward, and was likely not on the drug prior to the note as well.",
"cite_spans": [
{
"start": 129,
"end": 150,
"text": "(Devlin et al., 2019;",
"ref_id": "BIBREF1"
},
{
"start": 151,
"end": 168,
"text": "Liu et al., 2019)",
"ref_id": null
}
],
"ref_spans": [
{
"start": 612,
"end": 620,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "with imperfect performance. Despite strong individual results, we find that performance suffers immensely due to both out-of-domain performance losses and the basic combinatorial math of integrating outputs from multiple systems. This is the case even though we use a metric, accuracy, that is forgiving to the worst-performing individual model.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "It is both formally and empirically understood that classifiers can suffer performance loss when the test data is drawn from a different distribution than the training data (sometimes called domain shift). This presents a difficult challenge in clinical NLP because data-sharing limitations make it difficult to create large and diverse training corpora. As a result, domain adaptation approaches have been applied to multiple tasks in clinical NLP (Miller et al., 2017; Liu et al., 2018; Hur et al., 2020) .",
"cite_spans": [
{
"start": 449,
"end": 470,
"text": "(Miller et al., 2017;",
"ref_id": "BIBREF12"
},
{
"start": 471,
"end": 488,
"text": "Liu et al., 2018;",
"ref_id": "BIBREF9"
},
{
"start": 489,
"end": 506,
"text": "Hur et al., 2020)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "Recent work in the general domain has made use of transfer learning, which can attack the problem of domain shift, but by a different mechanism than domain adaptation; by training on massive corpora, large pre-trained models both learn general features, and are able to learn from smaller new datasets without overfitting. The most prominent of these models are based on the transformer architecture (Vaswani et al., 2017) .",
"cite_spans": [
{
"start": 400,
"end": 422,
"text": "(Vaswani et al., 2017)",
"ref_id": "BIBREF17"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "BERT (Devlin et al., 2019) uses a transformer encoder, and has shown that pre-training with massive amounts of text on a language modeling task, then fine-tuning on a supervised task of interest, achieves large performance gains in multiple NLP tasks. 1 During fine-tuning for sentence classification tasks, a classification head with randomly initialized weights is attached to a special sentenceinitial token. Fine-tuning then proceeds in a standard supervised learning paradigm, with the goal of learning the weights of the classification head, but where the weights of all of the transformer encoder layers can also be updated. We use RoBERTabase, a 12-layer transformer encoder that provides excellent performance but manageable memory utilization for our hardware (Liu et al., 2019) .",
"cite_spans": [
{
"start": 5,
"end": 26,
"text": "(Devlin et al., 2019)",
"ref_id": "BIBREF1"
},
{
"start": 770,
"end": 788,
"text": "(Liu et al., 2019)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "The bigger vision of our current work is extracting temporally-aware medication mentions from electronic health records data. This would enable important downstream tasks including automatically extracting drug timelines to correlate with treatments, or extracting better causal information about drugs and potential adverse events. Some other recent work has also examined this topic (Ramirez et al., 2019) , but focused on a single drug class (proton pump inhibitors), was limited to the problem list section, and made the assumption that missing drug implied drug stoppage.",
"cite_spans": [
{
"start": 385,
"end": 407,
"text": "(Ramirez et al., 2019)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "We began this work by developing several NLP components necessary to extract drug temporality signatures, including negation detection, relation to document creation time (DocTimeRel), and aspectual link extraction (ALINK), all detailed below. Detecting negation helps us avoid false positives from mentions corresponding to, for example, decisions to not use a drug. DocTimeRel helps us distinguish mentions of drugs that are current from those that predate the current time period, or are being speculated about for future use. ALINK can model drug start, stop, and continuation events, which can help to distinguish whether a missing mention in the middle of a record corresponds to a stop and restart, or an incidentally omitted mention. Figure 1 shows an example instance of a drug mention to be classified for all three tasks.",
"cite_spans": [],
"ref_spans": [
{
"start": 742,
"end": 750,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Methods",
"sec_num": "3"
},
{
"text": "The THYME dataset (Styler IV et al., 2014) , released as part of Clinical TempEval , contains all three of these annotation types, on 1200 notes of patients with colon and brain cancer. We train all models on the colon cancer section (details on data are in Section 4). While our bigger project is specific to drug mentions, the problem is not limited to drug mentions, so we train and evaluate on all annotated events in the THYME corpus. We also assume that events are given, to allow a straightforward metric of how many events we \"get right\" when combining all property predictions. In the real world, events will have to be automatically detected, so our metric will be an upper bound on how often the combined models get everything correct.",
"cite_spans": [
{
"start": 18,
"end": 42,
"text": "(Styler IV et al., 2014)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Methods",
"sec_num": "3"
},
{
"text": "This is the task of finding whether a given event is being negated (e.g., statins is negated in not currently on statins). We model this as a spanin-context classification -given a sentence in a document with a marked event span, classify that span as being negated or not negated. We experiment with two different machine learning models. The first is a classical feature-based support vector machine that is the default model of Apache cTAKES (Savova et al., 2010) . Features include bag of words and part of speech tags in and around the event, negation cue words from lists and their relation to the event, and dependency parse features that relate negation cue words to events. Details of this system were presented by Wu et al. (2014) . For comparison we train a RoBERTa-based system, where the input representation is the sentence with special tokens indicating the event to be classified. We put a binary sigmoid layer as the output, with the \"[CLS]\" token representation from the final layer as the classifier input, and fine-tune the entire model. Hyperparameters such as learning rate and number of training epochs are optimized on the THYME colon development set. Our implementation uses the Huggingface Transformers library (Wolf et al., 2019) . DocTimeRel DocTimeRel classification is the task of relating an event to the document creation time. The categories are BEFORE, OVERLAP, AFTER, and BE-FORE/OVERLAP. As above, we model this as a span-in-context classification, and we again compare a feature-based approach with a RoBERTabased approach.",
"cite_spans": [
{
"start": 445,
"end": 466,
"text": "(Savova et al., 2010)",
"ref_id": "BIBREF14"
},
{
"start": 724,
"end": 740,
"text": "Wu et al. (2014)",
"ref_id": "BIBREF19"
},
{
"start": 1237,
"end": 1256,
"text": "(Wolf et al., 2019)",
"ref_id": "BIBREF18"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Negation detection",
"sec_num": null
},
{
"text": "The feature-based approach again uses the default cTAKES SVM-based implementation (Lin et al., 2016) , with features based on bags of words in and around the event, and verb tense information for verbs on either side of the event. We train a separate RoBERTa-based model with the same architecture as the negation model, with the only difference being that the output layer is a softmax over the four categories rather than a sigmoid.",
"cite_spans": [
{
"start": 82,
"end": 100,
"text": "(Lin et al., 2016)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Negation detection",
"sec_num": null
},
{
"text": "Aspectual link extraction (ALINK) is the task of classifying whether an event mention is related to an aspectual temporal modifier, for example, discontinued. This is annotated as a relation between an event and a modifier, but we model it as an event property classification task since each event can only participate in one type of relation. The set of possible labels is INITIATES, CONTINUES, TERMINATES, and REINITIATES.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Aspectual Link Extraction",
"sec_num": null
},
{
"text": "We are not aware of any existing open-source models for this task, so for our feature-based baseline we train a model with the same SVM classification approach and feature set as the DocTimeRel model in cTAKES. We did not perform extensive feature engineering for this task, so further gains in the SVM system are probably possible. For the RoBERTa-based model, we used the same architecture as both systems above, with a softmax over the 5 categories -the 4 ALINK categories above as well as NONE, indicating a drug mention does not participate in any ALINK relation. NONE is by far the most common category.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Aspectual Link Extraction",
"sec_num": null
},
{
"text": "The tasks described above are trained on a single source dataset, and must be combined into a pipeline that will run on data from a different target distribution. To adapt to the target domain, we use unsupervised domain adaptation methods, where we have access to only unlabeled target examples.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Domain Adaptation Methods",
"sec_num": "3.1"
},
{
"text": "Since large pre-trained transformer models have arrived, they have been shown to be quite robust to out-of-distribution examples (Hendrycks et al., 2020) , including on clinical tasks (Lin et al., 2020) , where it was shown that adding domain adaptation layers on top of BERT was no better than BERT itself for negation detection. One of the few effective methods for improving the out-of-distribution performance of pre-trained transformer models has been to continue to pre-train the language modeling objective on the target domain data, before any fine-tuning is done on the source data (Han and Eisenstein, 2019; Gururangan et al., 2020) . In this work, we focus on that method, since this is currently the most promising direction for adapting large pre-trained transformers. Specifically, to use this method, we run additional masked language model training steps on the target training data from the RoBERTa-base checkpoint, before fine-tuning on the labeled colon cancer data, and then testing on target test data. We tune the learning rate for the language model pre-training on target development set data, optimizing for perplexity.",
"cite_spans": [
{
"start": 129,
"end": 153,
"text": "(Hendrycks et al., 2020)",
"ref_id": "BIBREF5"
},
{
"start": 184,
"end": 202,
"text": "(Lin et al., 2020)",
"ref_id": null
},
{
"start": 591,
"end": 617,
"text": "(Han and Eisenstein, 2019;",
"ref_id": "BIBREF4"
},
{
"start": 618,
"end": 642,
"text": "Gururangan et al., 2020)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Domain Adaptation Methods",
"sec_num": "3.1"
},
{
"text": "For the three tasks of interest, we evaluate indomain (THYME colon cancer corpus), as well as one closely related out-of-domain corpus (THYME brain cancer corpus). We also use a second out-ofdomain corpus, an internal data set we annotated for all three tasks (pulmonary hypertension [PH] notes). This annotation was performed by an experienced annotator who has worked on clinical annotation projects in the past.",
"cite_spans": [
{
"start": 284,
"end": 288,
"text": "[PH]",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation",
"sec_num": "4"
},
{
"text": "We measure performance on negation with F1score, on DocTimeRel with accuracy (because the classes are relatively balanced), and on ALINK extraction with the average F1 score of all categories, macro-F1 (because the high frequency NONE label makes accuracy uninformative). In addition to system-level performance, we report an evaluation of mention-level accuracy: an event is counted as correct if all three systems made the correct prediction, and we report the percentage of events that were correct. This setting estimates how usable the entire pipeline is, given different system settings.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation",
"sec_num": "4"
},
{
"text": "The \"Colon\" columns of Table 1 show results on the THYME colon cancer data (in-domain). RoBERTa performance is stronger than the SVM on all three tasks. Negation performance is particularly strong, though we are not aware of any reported results on this dataset to compare against. DocTimeRel performance is 3 points better than the best result of Clinical TempEval 2016 . ALINK scores are lower than the other tasks, though again there are no published comparisons. It is likely this is a more difficult task, in particular because the RE-INITIATES category has relatively few examples and whose low performance skews the averaging of the macro-F1.",
"cite_spans": [],
"ref_spans": [
{
"start": 23,
"end": 30,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Evaluation",
"sec_num": "4"
},
{
"text": "The \"Brain\" and \"PH\" columns of Table 1 show out-of-domain performance of the same systems on the THYME brain cancer and our internal pulmonary hypertension data, respectively. On THYME brain cancer data, RoBERTa again out-performs SVM substantially on all sub-tasks, but surprisingly the SVM performs better on PH data for negation and DocTimeRel. Adapting the RoBERTa model (RoBERTa+LM) by performing additional language modeling in the target domain before fine-tuning on colon cancer data leads to gains only on DocTimeRel for the PH data and on ALINK for both corpora. However, the improvement to DocTimeRel from adapting RoBERTa still leaves it worse off than the SVM.",
"cite_spans": [],
"ref_spans": [
{
"start": 32,
"end": 39,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Evaluation",
"sec_num": "4"
},
{
"text": "Mention level accuracy (\"All\" column) is good for the in-domain data (THYME colon cancer), but drops off substantially even for the THYME brain cancer corpus from the same institution, created with the same guidelines and using the same annotators. The mention level accuracy for our internal PH data is unusable at an accuracy of 0.506 with RoBERTa+LM. This accuracy means that roughly one of every two drug mentions will have at least one of its attributes classified incorrectly.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation",
"sec_num": "4"
},
{
"text": "The results also show that combining NLP systems for new, complex, information needs is likely to run into issues even when individual systems perform well. In particular, our experiments raise questions about real-world use of domain adaptation. If we treated THYME colon and brain sets as representative in-domain and out-of-domain datasets we would select RoBERTa or RoBERTa+LM for everything. But an oracle optimizing PH performance would tell us to use the SVM for negation and DocTimeRel and RoBERTa+LM for ALINK. One of the difficulties in even studying domain adaptation is model selection -if labeled target data is not available, standard practices like tuning on held out data are impossible. But the reality our results suggest is that different algorithms work well on different tasks and datasets, and selecting the best model for each task is an unsolved and under-studied problem.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion and Conclusion",
"sec_num": "5"
},
{
"text": "One direction of research that may address these concerns is on better modeling of domains themselves. The problem has been exacerbated with the move from feature-based classifiers to pre-trained black box models, as it is now even more difficult to understand the cause of errors in new domains without interpretable features. Domain adaptation should leverage \"BERTology\" and interpretability research to help understand how different aspects of domains contribute to performance differences. For example, in clinical notes, variation in institutions, specialties, note types, authors, etc., all Table 1 : Performance on both the individual sub-tasks (Negation, DocTimeRel, and ALINK) and the complete task (All) for systems trained on the THYME colon cancer training set and tested on the in-domain THYME colon test set, the out-of-domain THYME brain test set, and the out-of-domain pulmonary hypertension (PH) test set.",
"cite_spans": [],
"ref_spans": [
{
"start": 598,
"end": 605,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Discussion and Conclusion",
"sec_num": "5"
},
{
"text": "probably contribute differently to domain shift, and these sources of variation should be empirically explored. Future work will explore this direction to develop unsupervised model selection algorithms that better predict target domain performance.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion and Conclusion",
"sec_num": "5"
},
{
"text": "The RoBERTa system that followed(Liu et al., 2019) found further gains by pre-training on even larger datasets and for more iterations.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "Research reported in this publication was supported by the National Library Of Medicine of the National Institutes of Health under Award Numbers R01LM012918 and R01LM012973. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgments",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "SemEval-2017 task 12: Clinical TempEval",
"authors": [
{
"first": "Steven",
"middle": [],
"last": "Bethard",
"suffix": ""
},
{
"first": "Guergana",
"middle": [],
"last": "Savova",
"suffix": ""
},
{
"first": "Martha",
"middle": [],
"last": "Palmer",
"suffix": ""
},
{
"first": "James",
"middle": [],
"last": "Pustejovsky",
"suffix": ""
}
],
"year": 2017,
"venue": "Proceedings of the 11th International Workshop on Semantic Evaluation (SemEval-2017)",
"volume": "",
"issue": "",
"pages": "565--572",
"other_ids": {
"DOI": [
"10.18653/v1/S17-2093"
]
},
"num": null,
"urls": [],
"raw_text": "Steven Bethard, Guergana Savova, Martha Palmer, and James Pustejovsky. 2017. SemEval-2017 task 12: Clinical TempEval. In Proceedings of the 11th International Workshop on Semantic Evalua- tion (SemEval-2017), pages 565-572, Vancouver, Canada. Association for Computational Linguistics.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"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
},
"BIBREF2": {
"ref_id": "b2",
"title": "Solving the problem of cascading errors: Approximate Bayesian inference for linguistic annotation pipelines",
"authors": [
{
"first": "Jenny",
"middle": [
"Rose"
],
"last": "Finkel",
"suffix": ""
},
{
"first": "Christopher",
"middle": [
"D"
],
"last": "Manning",
"suffix": ""
},
{
"first": "Andrew",
"middle": [
"Y"
],
"last": "Ng",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "618--626",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jenny Rose Finkel, Christopher D. Manning, and An- drew Y. Ng. 2006. Solving the problem of cascad- ing errors: Approximate Bayesian inference for lin- guistic annotation pipelines. In Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing, pages 618-626, Sydney, Aus- tralia. Association for Computational Linguistics.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Don't stop pretraining: Adapt language models to domains and tasks",
"authors": [
{
"first": "Ana",
"middle": [],
"last": "Suchin Gururangan",
"suffix": ""
},
{
"first": "Swabha",
"middle": [],
"last": "Marasovi\u0107",
"suffix": ""
},
{
"first": "Kyle",
"middle": [],
"last": "Swayamdipta",
"suffix": ""
},
{
"first": "Iz",
"middle": [],
"last": "Lo",
"suffix": ""
},
{
"first": "Doug",
"middle": [],
"last": "Beltagy",
"suffix": ""
},
{
"first": "Noah",
"middle": [
"A"
],
"last": "Downey",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Smith",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "8342--8360",
"other_ids": {
"DOI": [
"10.18653/v1/2020.acl-main.740"
]
},
"num": null,
"urls": [],
"raw_text": "Suchin Gururangan, Ana Marasovi\u0107, Swabha Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, and Noah A. Smith. 2020. Don't stop pretraining: Adapt language models to domains and tasks. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8342-8360, Online. Association for Computational Linguistics.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Unsupervised domain adaptation of contextualized embeddings for sequence labeling",
"authors": [
{
"first": "Xiaochuang",
"middle": [],
"last": "Han",
"suffix": ""
},
{
"first": "Jacob",
"middle": [],
"last": "Eisenstein",
"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 (EMNLP-IJCNLP)",
"volume": "",
"issue": "",
"pages": "4238--4248",
"other_ids": {
"DOI": [
"10.18653/v1/D19-1433"
]
},
"num": null,
"urls": [],
"raw_text": "Xiaochuang Han and Jacob Eisenstein. 2019. Unsu- pervised domain adaptation of contextualized em- beddings for sequence labeling. In Proceedings of the 2019 Conference on Empirical Methods in Nat- ural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 4238-4248, Hong Kong, China. Association for Computational Linguistics.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Pretrained transformers improve out-of-distribution robustness",
"authors": [
{
"first": "Dan",
"middle": [],
"last": "Hendrycks",
"suffix": ""
},
{
"first": "Xiaoyuan",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "Eric",
"middle": [],
"last": "Wallace",
"suffix": ""
},
{
"first": "Adam",
"middle": [],
"last": "Dziedzic",
"suffix": ""
},
{
"first": "Rishabh",
"middle": [],
"last": "Krishnan",
"suffix": ""
},
{
"first": "Dawn",
"middle": [],
"last": "Song",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "2744--2751",
"other_ids": {
"DOI": [
"10.18653/v1/2020.acl-main.244"
]
},
"num": null,
"urls": [],
"raw_text": "Dan Hendrycks, Xiaoyuan Liu, Eric Wallace, Adam Dziedzic, Rishabh Krishnan, and Dawn Song. 2020. Pretrained transformers improve out-of-distribution robustness. In Proceedings of the 58th Annual Meet- ing of the Association for Computational Linguistics, pages 2744-2751, Online. Association for Computa- tional Linguistics.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Domain adaptation and instance selection for disease syndrome classification over veterinary clinical notes",
"authors": [
{
"first": "Brian",
"middle": [],
"last": "Hur",
"suffix": ""
},
{
"first": "Timothy",
"middle": [],
"last": "Baldwin",
"suffix": ""
},
{
"first": "Karin",
"middle": [],
"last": "Verspoor",
"suffix": ""
},
{
"first": "Laura",
"middle": [],
"last": "Hardefeldt",
"suffix": ""
},
{
"first": "James",
"middle": [],
"last": "Gilkerson",
"suffix": ""
}
],
"year": 2020,
"venue": "Proceedings of the 19th SIGBioMed Workshop on Biomedical Language Processing",
"volume": "",
"issue": "",
"pages": "156--166",
"other_ids": {
"DOI": [
"10.18653/v1/2020.bionlp-1.17"
]
},
"num": null,
"urls": [],
"raw_text": "Brian Hur, Timothy Baldwin, Karin Verspoor, Laura Hardefeldt, and James Gilkerson. 2020. Domain adaptation and instance selection for disease syn- drome classification over veterinary clinical notes. In Proceedings of the 19th SIGBioMed Workshop on Biomedical Language Processing, pages 156-166, Online. Association for Computational Linguistics.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Farig Sadeque, Guergana Savova, and Timothy A Miller. 2020. Does BERT need domain adaptation for clinical negation detection",
"authors": [
{
"first": "Chen",
"middle": [],
"last": "Lin",
"suffix": ""
},
{
"first": "Steven",
"middle": [],
"last": "Bethard",
"suffix": ""
},
{
"first": "Dmitriy",
"middle": [],
"last": "Dligach",
"suffix": ""
}
],
"year": null,
"venue": "Journal of the American Medical Informatics Association",
"volume": "27",
"issue": "4",
"pages": "584--591",
"other_ids": {
"DOI": [
"10.1093/jamia/ocaa001"
]
},
"num": null,
"urls": [],
"raw_text": "Chen Lin, Steven Bethard, Dmitriy Dligach, Farig Sad- eque, Guergana Savova, and Timothy A Miller. 2020. Does BERT need domain adaptation for clinical negation detection? Journal of the American Medi- cal Informatics Association, 27(4):584-591.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Multilayered temporal modeling for the clinical domain",
"authors": [
{
"first": "Chen",
"middle": [],
"last": "Lin",
"suffix": ""
},
{
"first": "Dmitriy",
"middle": [],
"last": "Dligach",
"suffix": ""
},
{
"first": "Timothy",
"middle": [
"A"
],
"last": "Miller",
"suffix": ""
},
{
"first": "Steven",
"middle": [],
"last": "Bethard",
"suffix": ""
},
{
"first": "Guergana K",
"middle": [],
"last": "Savova",
"suffix": ""
}
],
"year": 2016,
"venue": "Journal of the American Medical Informatics Association",
"volume": "23",
"issue": "2",
"pages": "387--395",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Chen Lin, Dmitriy Dligach, Timothy A Miller, Steven Bethard, and Guergana K Savova. 2016. Multi- layered temporal modeling for the clinical domain. Journal of the American Medical Informatics Asso- ciation, 23(2):387-395.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Domain adaptation for disease phrase matching with adversarial networks",
"authors": [
{
"first": "Miaofeng",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "Jialong",
"middle": [],
"last": "Han",
"suffix": ""
},
{
"first": "Haisong",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Yan",
"middle": [],
"last": "Song",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the BioNLP 2018 workshop",
"volume": "",
"issue": "",
"pages": "137--141",
"other_ids": {
"DOI": [
"10.18653/v1/W18-2315"
]
},
"num": null,
"urls": [],
"raw_text": "Miaofeng Liu, Jialong Han, Haisong Zhang, and Yan Song. 2018. Domain adaptation for disease phrase matching with adversarial networks. In Proceedings of the BioNLP 2018 workshop, pages 137-141, Mel- bourne, Australia. Association for Computational Linguistics.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Roberta: A robustly optimized bert pretraining approach",
"authors": [],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1907.11692"
]
},
"num": null,
"urls": [],
"raw_text": "Roberta: A robustly optimized bert pretraining ap- proach. arXiv preprint arXiv:1907.11692.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Unsupervised domain adaptation for clinical negation detection",
"authors": [
{
"first": "Timothy",
"middle": [],
"last": "Miller",
"suffix": ""
},
{
"first": "Steven",
"middle": [],
"last": "Bethard",
"suffix": ""
},
{
"first": "Hadi",
"middle": [],
"last": "Amiri",
"suffix": ""
},
{
"first": "Guergana",
"middle": [],
"last": "Savova",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "165--170",
"other_ids": {
"DOI": [
"10.18653/v1/W17-2320"
]
},
"num": null,
"urls": [],
"raw_text": "Timothy Miller, Steven Bethard, Hadi Amiri, and Guer- gana Savova. 2017. Unsupervised domain adap- tation for clinical negation detection. In BioNLP 2017, pages 165-170, Vancouver, Canada,. Associ- ation for Computational Linguistics.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Extracting Drug Exposure Epochs and Drug Response Outcomes from Electronic Health Records",
"authors": [
{
"first": "Andrea",
"middle": [
"H"
],
"last": "Ramirez",
"suffix": ""
},
{
"first": "Yaping",
"middle": [],
"last": "Shi",
"suffix": ""
},
{
"first": "Elliot",
"middle": [],
"last": "Fielstein",
"suffix": ""
},
{
"first": "Jonathan",
"middle": [],
"last": "Schildcrout",
"suffix": ""
},
{
"first": "Henry",
"middle": [
"H"
],
"last": "Ong",
"suffix": ""
},
{
"first": "Josh",
"middle": [
"C"
],
"last": "Denny",
"suffix": ""
},
{
"first": "Josh",
"middle": [
"F"
],
"last": "Peterson",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of AMIA Annual Symposium (Podium Abstract)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Andrea H. Ramirez, Yaping Shi, Elliot Fielstein, Jonathan Schildcrout, Henry H. Ong, Josh C. Denny, and Josh F. Peterson. 2019. Extracting Drug Ex- posure Epochs and Drug Response Outcomes from Electronic Health Records. In Proceedings of AMIA Annual Symposium (Podium Abstract), Washington, DC.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Mayo clinical Text Analysis and Knowledge Extraction System (cTAKES): architecture, component evaluation and applications",
"authors": [
{
"first": "K",
"middle": [],
"last": "Guergana",
"suffix": ""
},
{
"first": "James",
"middle": [
"J"
],
"last": "Savova",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Masanz",
"suffix": ""
},
{
"first": "Jiaping",
"middle": [],
"last": "Philip V Ogren",
"suffix": ""
},
{
"first": "Sunghwan",
"middle": [],
"last": "Zheng",
"suffix": ""
},
{
"first": "Karin",
"middle": [
"C"
],
"last": "Sohn",
"suffix": ""
},
{
"first": "Christopher",
"middle": [
"G"
],
"last": "Kipper-Schuler",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Chute",
"suffix": ""
}
],
"year": 2010,
"venue": "Journal of the American Medical Informatics Association",
"volume": "17",
"issue": "5",
"pages": "507--513",
"other_ids": {
"DOI": [
"10.1136/jamia.2009.001560"
]
},
"num": null,
"urls": [],
"raw_text": "Guergana K Savova, James J Masanz, Philip V Ogren, Jiaping Zheng, Sunghwan Sohn, Karin C Kipper- Schuler, and Christopher G Chute. 2010. Mayo clin- ical Text Analysis and Knowledge Extraction Sys- tem (cTAKES): architecture, component evaluation and applications. Journal of the American Medical Informatics Association, 17(5):507-513.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Temporal annotation in the clinical domain",
"authors": [
{
"first": "F",
"middle": [],
"last": "William",
"suffix": ""
},
{
"first": "I",
"middle": [
"V"
],
"last": "Styler",
"suffix": ""
},
{
"first": "Steven",
"middle": [],
"last": "Bethard",
"suffix": ""
},
{
"first": "Sean",
"middle": [],
"last": "Finan",
"suffix": ""
},
{
"first": "Martha",
"middle": [],
"last": "Palmer",
"suffix": ""
},
{
"first": "Sameer",
"middle": [],
"last": "Pradhan",
"suffix": ""
},
{
"first": "Piet",
"middle": [
"C"
],
"last": "De Groen",
"suffix": ""
},
{
"first": "Brad",
"middle": [],
"last": "Erickson",
"suffix": ""
},
{
"first": "Timothy",
"middle": [],
"last": "Miller",
"suffix": ""
},
{
"first": "Chen",
"middle": [],
"last": "Lin",
"suffix": ""
},
{
"first": "Guergana",
"middle": [],
"last": "Savova",
"suffix": ""
},
{
"first": "James",
"middle": [],
"last": "Pustejovsky",
"suffix": ""
}
],
"year": 2014,
"venue": "Transactions of the Association for Computational Linguistics",
"volume": "2",
"issue": "",
"pages": "143--154",
"other_ids": {
"DOI": [
"10.1162/tacl_a_00172"
]
},
"num": null,
"urls": [],
"raw_text": "William F. Styler IV, Steven Bethard, Sean Finan, Martha Palmer, Sameer Pradhan, Piet C de Groen, Brad Erickson, Timothy Miller, Chen Lin, Guergana Savova, and James Pustejovsky. 2014. Temporal an- notation in the clinical domain. Transactions of the Association for Computational Linguistics, 2:143- 154.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "i2b2/VA challenge on concepts, assertions, and relations in clinical text",
"authors": [
{
"first": "Ozlem",
"middle": [],
"last": "Uzuner",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Brett",
"suffix": ""
},
{
"first": "Shuying",
"middle": [],
"last": "South",
"suffix": ""
},
{
"first": "Scott L",
"middle": [],
"last": "Shen",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Duvall",
"suffix": ""
}
],
"year": 2010,
"venue": "Journal of the American Medical Informatics Association",
"volume": "18",
"issue": "5",
"pages": "552--556",
"other_ids": {
"DOI": [
"10.1136/amiajnl-2011-000203"
]
},
"num": null,
"urls": [],
"raw_text": "Ozlem Uzuner, Brett R South, Shuying Shen, and Scott L DuVall. 2011. 2010 i2b2/VA challenge on concepts, assertions, and relations in clinical text. Journal of the American Medical Informatics Asso- ciation, 18(5):552-556.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Attention is all you need",
"authors": [
{
"first": "Ashish",
"middle": [],
"last": "Vaswani",
"suffix": ""
},
{
"first": "Noam",
"middle": [],
"last": "Shazeer",
"suffix": ""
},
{
"first": "Niki",
"middle": [],
"last": "Parmar",
"suffix": ""
},
{
"first": "Jakob",
"middle": [],
"last": "Uszkoreit",
"suffix": ""
},
{
"first": "Llion",
"middle": [],
"last": "Jones",
"suffix": ""
},
{
"first": "Aidan",
"middle": [
"N"
],
"last": "Gomez",
"suffix": ""
},
{
"first": "\u0141ukasz",
"middle": [],
"last": "Kaiser",
"suffix": ""
},
{
"first": "Illia",
"middle": [],
"last": "Polosukhin",
"suffix": ""
}
],
"year": 2017,
"venue": "Advances in neural information processing systems",
"volume": "",
"issue": "",
"pages": "5998--6008",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, \u0141ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in neural information pro- cessing systems, pages 5998-6008.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Huggingface's transformers: State-of-the-art natural language processing",
"authors": [
{
"first": "Thomas",
"middle": [],
"last": "Wolf",
"suffix": ""
},
{
"first": "Lysandre",
"middle": [],
"last": "Debut",
"suffix": ""
},
{
"first": "Victor",
"middle": [],
"last": "Sanh",
"suffix": ""
},
{
"first": "Julien",
"middle": [],
"last": "Chaumond",
"suffix": ""
},
{
"first": "Clement",
"middle": [],
"last": "Delangue",
"suffix": ""
},
{
"first": "Anthony",
"middle": [],
"last": "Moi",
"suffix": ""
},
{
"first": "Pierric",
"middle": [],
"last": "Cistac",
"suffix": ""
},
{
"first": "Tim",
"middle": [],
"last": "Rault",
"suffix": ""
},
{
"first": "R'emi",
"middle": [],
"last": "Louf",
"suffix": ""
},
{
"first": "Morgan",
"middle": [],
"last": "Funtowicz",
"suffix": ""
},
{
"first": "Jamie",
"middle": [],
"last": "Brew",
"suffix": ""
}
],
"year": 2019,
"venue": "ArXiv",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pier- ric Cistac, Tim Rault, R'emi Louf, Morgan Funtow- icz, and Jamie Brew. 2019. Huggingface's trans- formers: State-of-the-art natural language process- ing. ArXiv, abs/1910.03771.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Negation's not solved: generalizability versus optimizability in clinical natural language processing",
"authors": [
{
"first": "Stephen",
"middle": [],
"last": "Wu",
"suffix": ""
},
{
"first": "Timothy",
"middle": [],
"last": "Miller",
"suffix": ""
},
{
"first": "James",
"middle": [],
"last": "Masanz",
"suffix": ""
},
{
"first": "Matt",
"middle": [],
"last": "Coarr",
"suffix": ""
},
{
"first": "Scott",
"middle": [],
"last": "Halgrim",
"suffix": ""
},
{
"first": "David",
"middle": [],
"last": "Carrell",
"suffix": ""
},
{
"first": "Cheryl",
"middle": [],
"last": "Clark",
"suffix": ""
}
],
"year": 2014,
"venue": "PloS one",
"volume": "9",
"issue": "11",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Stephen Wu, Timothy Miller, James Masanz, Matt Coarr, Scott Halgrim, David Carrell, and Cheryl Clark. 2014. Negation's not solved: generalizabil- ity versus optimizability in clinical natural language processing. PloS one, 9(11):e112774.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Task refinement learning for improved accuracy and stability of unsupervised domain adaptation",
"authors": [
{
"first": "Yftah",
"middle": [],
"last": "Ziser",
"suffix": ""
},
{
"first": "Roi",
"middle": [],
"last": "Reichart",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "5895--5906",
"other_ids": {
"DOI": [
"10.18653/v1/P19-1591"
]
},
"num": null,
"urls": [],
"raw_text": "Yftah Ziser and Roi Reichart. 2019. Task refinement learning for improved accuracy and stability of un- supervised domain adaptation. In Proceedings of the 57th Annual Meeting of the Association for Com- putational Linguistics, pages 5895-5906, Florence, Italy. Association for Computational Linguistics.",
"links": null
}
},
"ref_entries": {
"TABREF0": {
"content": "<table><tr><td>System</td><td colspan=\"2\">Negation (F1)</td><td colspan=\"4\">DocTimeRel (Acc) ALINK (MacroF)</td><td/><td>All (Acc)</td></tr><tr><td>RoBERTa+LM</td><td>-</td><td>0.831 0.582</td><td>-</td><td>0.758 0.615</td><td>-</td><td>0.660 0.694</td><td>-</td><td>0.736 0.506</td></tr></table>",
"html": null,
"text": "Colon Brain PH Colon Brain PH Colon Brain PH Colon Brain PH SVM 0.924 0.705 0.625 0.842 0.703 0.694 0.502 0.338 0.345 ---RoBERTa 0.950 0.833 0.583 0.874 0.757 0.542 0.684 0.633 0.674 0.860 0.732 0.454",
"num": null,
"type_str": "table"
}
}
}
} |