File size: 95,296 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 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 | {
"paper_id": "2020",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T13:27:18.382958Z"
},
"title": "Do Sentence Embeddings Capture Discourse Properties of Sentences from Scientific Abstracts ?",
"authors": [
{
"first": "Laurine",
"middle": [],
"last": "Huber",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "LORIA (UMR",
"location": {
"addrLine": "7503)",
"postCode": "F-54000",
"settlement": "Nancy",
"region": "Inria",
"country": "France"
}
},
"email": "laurine.huber@loria.fr"
},
{
"first": "Chaker",
"middle": [],
"last": "Memmadi",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "LORIA (UMR",
"location": {
"addrLine": "7503)",
"postCode": "F-54000",
"settlement": "Nancy",
"region": "Inria",
"country": "France"
}
},
"email": "chaker.memmadi@outlook.fr"
},
{
"first": "Mathilde",
"middle": [],
"last": "Dargnat",
"suffix": "",
"affiliation": {
"laboratory": "ATILF, Universit\u00e9 de Lorraine, CNRS (UMR 7118)",
"institution": "",
"location": {
"settlement": "Nancy",
"country": "France"
}
},
"email": "mathilde.dargnat@univ-lorraine.fr"
},
{
"first": "Yannick",
"middle": [],
"last": "Toussaint",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "LORIA (UMR",
"location": {
"addrLine": "7503)",
"postCode": "F-54000",
"settlement": "Nancy",
"region": "Inria",
"country": "France"
}
},
"email": "yannick.toussaint@loria.fr"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "We introduce four tasks designed to determine which sentence encoders best capture discourse properties of sentences from scientific abstracts, namely coherence between clauses of a sentence, and discourse relations within sentences. We show that even if contextual encoders such as BERT or SciBERT encodes the coherence in discourse units, they do not help to predict three discourse relations commonly used in scientific abstracts. We discuss what these results underline, namely that these discourse relations are based on particular phrasing that allow non-contextual encoders to perform well.",
"pdf_parse": {
"paper_id": "2020",
"_pdf_hash": "",
"abstract": [
{
"text": "We introduce four tasks designed to determine which sentence encoders best capture discourse properties of sentences from scientific abstracts, namely coherence between clauses of a sentence, and discourse relations within sentences. We show that even if contextual encoders such as BERT or SciBERT encodes the coherence in discourse units, they do not help to predict three discourse relations commonly used in scientific abstracts. We discuss what these results underline, namely that these discourse relations are based on particular phrasing that allow non-contextual encoders to perform well.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "This paper compares the ability of different sentence encoders at representing coherence of sentences from scientific abstracts, and more specifically discourse relations between clauses.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Our first hypothesis is that BERT (Devlin et al., 2019) and SciBERT (Beltagy et al., 2019) models enable to produce sentence representations that take coherence into account thanks to their training done on Next Sentence Prediction. Our second hypothesis is that this training should also enable to capture discourse relations between clauses.",
"cite_spans": [
{
"start": 34,
"end": 55,
"text": "(Devlin et al., 2019)",
"ref_id": "BIBREF7"
},
{
"start": 68,
"end": 90,
"text": "(Beltagy et al., 2019)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Discourse relations (DRs) represent the semantic and pragmatic links between discourse units (DUs), that are either clauses, sentences or groups thereof, within a hierarchical structure that represents the whole text. In this work, we focus on sentences that are defined as textual sequences separated by a period. Sentences may comprise one or more DUs, and when they have at least two DUs, the coherence links between them is represented through DRs. These DRs are either explicitly signaled, or left implicit. For example, the sentence \"By wearing a mask, we can protect the others.\" conveys an enablement relation between the action \"wearing a mask\" and the event \"protecting the others\", which is here lexicalized by the connective \"by\". DRs can be used to extract new knowledge, especially in scientific abstracts which are highly structured (Liddy, 1991) .",
"cite_spans": [
{
"start": 848,
"end": 861,
"text": "(Liddy, 1991)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Sentence embeddings (SE) represent the meaning of a sentence in a fixed-size vector space, and recent contextual approaches such as BERT have shown promising results for downstream tasks such as Semantic Textual Similarity (STS) or Natural Language Inference (NLI) (Reimers and Gurevych, 2019) . When further trained on downstream tasks, these models have shown promising results. However, their performance also rely on linguistic knowledge acquired at pre-training. For example, BERT and SciBERT are trained on both Masked Language Model and Next Sentence Prediction tasks, in order to capture general linguistic properties that are then transferred to learn more specific representations.",
"cite_spans": [
{
"start": 265,
"end": 293,
"text": "(Reimers and Gurevych, 2019)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In this work, we want to understand if discourse properties are embedded in sentence representations that are built before further training on downstream tasks.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We design probing tasks, that are classification tasks whose goals are to predict discourse properties of the sentences from their embedding. Our goal is to highlight if discourse properties of sentences are captured by those vectors without finetuning. We rely on the corpus SciDTB (Yang and Li, 2018) to build four datasets used to probe if embeddings capture some discourse properties of the sentences. The two first datasets probe the coherence of sentences, and the two others probe the presence of DRs. We use four different sentence encoders to produce sentence embeddings, that we then use as input vectors for two classification models. If the classifier succeeds, it means that the vectors stores the discourse property that is probed. We evaluate the classifiers, thus high-lighting encoders that best encode the properties we probe.",
"cite_spans": [
{
"start": 283,
"end": 302,
"text": "(Yang and Li, 2018)",
"ref_id": "BIBREF21"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "This paper is organized as follows. We first provide background to our work. Second, we detail the tasks that we design to detect discourse properties. Third, we present our experimental setup, including the different SE that we evaluate and the choices that are specific to the corpus on which we rely. Finally, we present our results and discuss some issues.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Coherence and cohesion are two key notions in the perception of a text as a unified whole.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "Coherence refers to logical and semantic relations between clauses and sentences in a text, while cohesion refers to grammatical or lexical devices such as pronouns, verb tense or connectives, that form external relations of a text (Halliday and Hasan, 1976) . While the former may stand without the latter and vice versa, the connectivity model of Renkema (2009) mixes coherence and cohesion cues. In this work we follow this approach, and will refer to as coherence the properties of both coherence and cohesion.",
"cite_spans": [
{
"start": 232,
"end": 258,
"text": "(Halliday and Hasan, 1976)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "DRs may either explicitly signal coherence relations by discourse connectives, or left them implicit. In this latter case, the reader infers the relations based on coherence links between clauses. For example, the marker \"to\" is frequently used to links an action X and a way Y to realize X, signaling that Y is the manner-means to do X. However, even if some relations are explicitly signaled with discourse connectives such as \"to\" or \"by\", others are less salient. For example in \"We propose a novel extension of this work using target context information.\", the ellipsis of \"by\" make it harder to infer the relation. Some of them (e.g enablement, manner-means or attribution) express a logical link between the content of the clauses they relate, while others (e.g elaboration or progression) only express a continuation or additive relation. In scientific abstracts especially, DRs that convey a logical link are often lexically marked.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "Several theories such as Rhetorical Structure Theory (RST) (Taboada and Mann, 2006) or Segmented Discourse Representation Theory (SDRT) (Lascarides and Asher, 2007) help to build text discourse structures, by providing both sets of DRs that are defined based on the content of the DUs and a framework for attaching the DUs by means of DRs. The RST defines a set of relations 1 that are either mononuclear (if for two DUs, one is more salient than the other) or multinuclear (if two or more DUs have the same importance). They serve to build the hierarchical discourse structure of the text, either as a constituency tree, or, more recently, with dependency trees (Morey et al., 2018) , which are used to annotate scientific abstracts from SciDTB (Yang and Li, 2018) . SE have shown promising results for a wide variety of NLP tasks (Conneau and Kiela, 2018) , but are not interpretable independently of the others, making it unclear what linguistic information they contain and what is the meaning that they represent exactly. Previous works (Shi et al., 2016; Adi et al., 2016; intended to clarify what linguistic information are contained in SE by designing auxiliary tasks that take a single sentence vector as input and tries to predict a simple linguistic property of the sentence. They showed that some syntactic properties (e.g word order or syntactic tree depth) and some properties that they defined as semantic 2 (e.g the tense of the main verb or the number of the subject) are well captured by SE. We then hypothesize that the latter may also capture discourse coherence and DRs in sentences from scientific abstracts, thus being the first step to extracting information at a semantic-pragmatic level.",
"cite_spans": [
{
"start": 59,
"end": 83,
"text": "(Taboada and Mann, 2006)",
"ref_id": "BIBREF17"
},
{
"start": 136,
"end": 164,
"text": "(Lascarides and Asher, 2007)",
"ref_id": "BIBREF10"
},
{
"start": 663,
"end": 683,
"text": "(Morey et al., 2018)",
"ref_id": "BIBREF12"
},
{
"start": 746,
"end": 765,
"text": "(Yang and Li, 2018)",
"ref_id": "BIBREF21"
},
{
"start": 832,
"end": 857,
"text": "(Conneau and Kiela, 2018)",
"ref_id": "BIBREF5"
},
{
"start": 1042,
"end": 1060,
"text": "(Shi et al., 2016;",
"ref_id": "BIBREF16"
},
{
"start": 1061,
"end": 1078,
"text": "Adi et al., 2016;",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background",
"sec_num": "2"
},
{
"text": "In this section, for each probing task, we explain how we build a dataset of sentences tagged with discourse properties. They are then encoded by various methods and used to train and evaluate classifiers. We introduce two tasks designed to check if coherence between clauses of a sentence is encoded (3.1, 3.2), and two tasks to check if DRs in sentences are encoded (3.3, 3.4).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Probing tasks",
"sec_num": "3"
},
{
"text": "Ordering of clauses fully participates to sentences coherence. We evaluate if coherence is captured by SE by evaluating the ability of a classifier to distinguish between coherent sentences and incoherent ones. We produce incoherent sentences by swap-ping two adjacent discourse units from sentences originally correct.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Swapped units detection",
"sec_num": "3.1"
},
{
"text": "The dataset for this probing task thus consists of original multi-clausal sentences and of their incoherent equivalent obtained by swapping two random adjacent DUs. By doing so, we however do not ensure that new sentences are always incoherent. For example, swapping units 0 and 1 in the sentence \"[Word alignment] [using recencyvector based approach] [has recently become popular.]\" produces the new coherent sentence \"[Using recency-vector based approach] [word alignment] [has recently become popular.]\".",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Swapped units detection",
"sec_num": "3.1"
},
{
"text": "The resulting task is a binary classification task into {yes, no}, where yes corresponds to swapped sentences, and no to the original sentences.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Swapped units detection",
"sec_num": "3.1"
},
{
"text": "Topic incoherence may emerge when a topic T 1 is involved in the context of another topic T 2 , leading to incoherence because of the incompatibility between T 1 and T 2 .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Scrambled sentence detection",
"sec_num": "3.2"
},
{
"text": "The dataset used to probe this property is built as follows. Starting from original multi-clausal sentences splitted in DUs, we replace a randomly chosen DU by another randomly chosen DU from a different document, thus changing the context of a DU. Here again, the new sentence might be still coherent. For example, replacing the second DU in the sentence \"[But these methods cannot be used][to obtain the estimates of causal effects-the quantity of interest for applied researchers.]\" produces a new coherent sentence \"[But these methods cannot be used] [using a two-phase approach.]\".",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Scrambled sentence detection",
"sec_num": "3.2"
},
{
"text": "The resulting task is a binary classification task into {yes, no}, where yes corresponds to scrambled sentences, and no to the original sentences.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Scrambled sentence detection",
"sec_num": "3.2"
},
{
"text": "A sentence may either be composed of a single DU, or of multiple DUs. In the last case, the DUs may be links through additive or continuation DRs (in bold in Table 1 ), or through logical relations (in italics in Table 1 ). We distinguish them and evaluate if SE can be used to predict whether a sentence contains a logical DR.",
"cite_spans": [],
"ref_spans": [
{
"start": 158,
"end": 165,
"text": "Table 1",
"ref_id": "TABREF1"
},
{
"start": 213,
"end": 220,
"text": "Table 1",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Relation detection",
"sec_num": "3.3"
},
{
"text": "We rely on the discourse annotations provided in SciDTB, from which we extract the subtrees corresponding to the discourse structures of the sentences. Sentences that contain only one DU and sentences made of multiple DUs whose subtree contains no logical DRs are classified as norel. Others are classified as rel regardless of the type of the logical DR they contain.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Relation detection",
"sec_num": "3.3"
},
{
"text": "The resulting task is a binary classification task into {rel, norel}.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Relation detection",
"sec_num": "3.3"
},
{
"text": "Being able to precisely identify which logical discourse relation is involved in a sentence may be useful as a first step toward knowledge extraction from texts. This task thus evaluates if the semantics of the DRs involved in sentences are represented in their embedding.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Relation semantics detection",
"sec_num": "3.4"
},
{
"text": "We rely again on the annotations to classify sentences based on the DR they contain. Sentences that contain only one DU and sentences made of multiple DU whose subtree contains no logical DRs are classified as norel. Others are classified based on the relation r they contain, if and only if they contain only one of these relations. The sentences that contain two or more logical relations are not tackled in our approach, and are thus not considered in the dataset.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Relation semantics detection",
"sec_num": "3.4"
},
{
"text": "The resulting task is a K-classes classification task where K is the number of relations that are considered.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Relation semantics detection",
"sec_num": "3.4"
},
{
"text": "SciDTB 3 is a corpus of 798 scientific abstracts from ACL Anthology 4 . Abstracts are segmented into DUs in a semi-automatic way, following the guidelines of (Carlson et al., 2003) , and annotated by discourse structure in dependency with DRs from the RST relations set, which was slightly modified and extended to be adapted to scientific abstracts 5 .",
"cite_spans": [
{
"start": 158,
"end": 180,
"text": "(Carlson et al., 2003)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Data",
"sec_num": "4.1"
},
{
"text": "The majority of documents contain between 5 and 7 sentences (minimum 2, maximum 14, mean 6), resulting in a set of 4196 sentences in total. Among them, 787 are made of a single DU, thus containing no relation, and 3409 contain more than one DU, and thus at least one DR. We use the segmentation into DUs to produce the datasets for coherence detection tasks (swapped and scrambled). We use the inter-clausal DRs to produce the datasets for DR detection tasks (binRel and semRel) 6 .",
"cite_spans": [
{
"start": 479,
"end": 480,
"text": "6",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Data",
"sec_num": "4.1"
},
{
"text": "For swapped and scrambled, each sentence that has more than one DU is classified as no, and is used to produce a new sentence that is classified as yes, resulting in a dataset of 6818 sentences. For each sentence (original and modified), we ensure that punctuation and case do not bias the experiment by removing capitalization, periods and commas.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data",
"sec_num": "4.1"
},
{
"text": "For binRel and semRel, we need to form classes that are broad enough to train the classifiers. It leads us to make choices because of the limited size of the corpus. Fig. 2 shows the number of sentences having one of the given DRs. Most of the DRs are involved in less than 200 sentences, which is not enough for training a classifier. Among the others, three relations are additive relations: Same-Unit links two segments of a DU broken into two parts, Joint, links two DUs which are in conjunction, and for Elaboration-addition one DU gives additional information to another DU 7 . Moreover, the latter are involved in most of sentences made of more than 3 DUs, and are thus often used together with logical DRs. We decide to ignore them, which allows us to build a sufficiently 6 We make the resulting datasets available at https://gitlab.com/laurinehu/scidtb_ expe/-/tree/master/probing_datasets. 7 We admit however that saying that Elaboration-addition does not contain an interesting semantics is somehow arbitrary, but can be understood by comparison with the semantics of other DRs. large dataset of sentences, tagged with the three relations: enablement, manner-means or attribution. We assume that the classifier will therefore learn to predict one of these relations. For binRel, each sentence which has no DR or only DRs from {joint, elaboration-addition, same-unit} is classified as norel, and the others as rel. For semRel, the norel classification is done in the same way, and sentences that involve one and only one relation from {attribution, enablement, manner-means} are classified according to the relation they contain. Table 2 summarizes the datasets used for each probing task, where the classes for each task are balanced. ",
"cite_spans": [
{
"start": 781,
"end": 782,
"text": "6",
"ref_id": null
},
{
"start": 901,
"end": 902,
"text": "7",
"ref_id": null
}
],
"ref_spans": [
{
"start": 166,
"end": 172,
"text": "Fig. 2",
"ref_id": "FIGREF1"
},
{
"start": 1641,
"end": 1648,
"text": "Table 2",
"ref_id": "TABREF3"
}
],
"eq_spans": [],
"section": "Data",
"sec_num": "4.1"
},
{
"text": "In this section, we present the different SE that we study, most of them being obtained by using pretrained models available on HuggingFace's (Wolf et al., 2019 ) website 8 . We compare SE obtained by averaging non-contextual word vectors (bag-ofvectors), and SE obtained by using more recent language contextual models based on transformers that have shown promising results for various tasks. Comparing the last to bag-of-vectors enables us to determine whether coherence or DRs are only captured because of linguistic cues, or because of implicit links between clauses.",
"cite_spans": [
{
"start": 142,
"end": 160,
"text": "(Wolf et al., 2019",
"ref_id": "BIBREF20"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Sentence embeddings",
"sec_num": "4.2"
},
{
"text": "Global word Vectors (Pennington et al., 2014) are word representations obtained from aggregated global word-word co-occurrence statistics from a large crawled corpus. Training objective is to learn word vectors such that their dot product is the log of the word's probability of co-occurrences. We use the largest pre-trained model available, constituted of a vocabulary of 2.2M words, and containing vectors of dimension 300. We build sentence vectors by computing the mean of the word vectors, thus producing bag-of-vectors (BoV) of dimension 300.",
"cite_spans": [
{
"start": 20,
"end": 45,
"text": "(Pennington et al., 2014)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Sentence embeddings",
"sec_num": "4.2"
},
{
"text": "Google USE (Cer et al., 2018 ) uses a Deep Averaging Network that is first trained in an unsupervised way as in Skip-Thoughts (Kiros et al., 2015) , and whose output is transferred to be further trained on a supervised way on Natural Language Inference task. This encoder was the state-of-theart before more recent contextual approaches with which we compare it. We use the largest model, which encodes sentences into vectors of dimension 512.",
"cite_spans": [
{
"start": 11,
"end": 28,
"text": "(Cer et al., 2018",
"ref_id": "BIBREF4"
},
{
"start": 126,
"end": 146,
"text": "(Kiros et al., 2015)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Sentence embeddings",
"sec_num": "4.2"
},
{
"text": "BERT (Devlin et al., 2019) is the current stateof-the-art for most NLP tasks. It produces contextual word representations, from training a bidirectional encoder on two tasks: the first called Masked LM (MLM) predicts a masked word from its left and right context, and the second called Next Sentence Prediction (NSP) predicts whether from two sentences A and B, B is the actual sentence that follows A. BERT thus represents the meaning of a word according to its immediate context, but also on the basis of relationships between sentences. We use BERT-base model, from which we recover the last hidden state after having processed the sentence in the model, thus producing vectors of dimension 768.",
"cite_spans": [
{
"start": 5,
"end": 26,
"text": "(Devlin et al., 2019)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Sentence embeddings",
"sec_num": "4.2"
},
{
"text": "SciBERT (Beltagy et al., 2019 ) is a BERT model specifically trained on full papers from the corpus of Semantic Scholar, thus seeking to provide a better representation of scientific vocabulary. We use SciBERT-base and proceed in the same way as for BERT, obtaining vectors of the same dimension.",
"cite_spans": [
{
"start": 8,
"end": 29,
"text": "(Beltagy et al., 2019",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Sentence embeddings",
"sec_num": "4.2"
},
{
"text": "For GloVe, BERT and SciBERT, we calculate the embedding e(S) of the sentence S by calculating the mean of each word vector as defined in 1.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Sentence embeddings",
"sec_num": "4.2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "e(S) = 1 |S| w\u2208S e(w)",
"eq_num": "(1)"
}
],
"section": "Sentence embeddings",
"sec_num": "4.2"
},
{
"text": "Because we deal with datasets of different size and vectors made of latent variables, we train and evaluate two different classifiers for each task, namely a Logistic Regression (LR) and a Multi-Layer-Perceptron (MLP) with one hidden layer and three hidden units. On the one hand, the datasets are small in size, which can make the training of the MLP hard and a good LR performance possible. On the other hand, the properties that we probe may not be linearly separable, which would make them hard to tackle with a LR classifier. The comparison of two classifiers thus gives us a fine-grained analysis of both the results obtained and the way in which the properties are encoded.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Classification",
"sec_num": "4.3"
},
{
"text": "Experimental set-up We use the evaluation toolkit Senteval 9 (Conneau and Kiela, 2018) , and keep the parameters as defined in it, namely the optimizer is RmsProp (Tijmen and Geoffrey, 2012) , the batch size is 164 and the loss function is crossentropy. Because the number of sentences that we have is quite small, we proceed with 5-fold cross validation. We compute the mean of accuracies at each fold. To compute the test accuracy, we keep the best model and test it on the testing set (10%) of the data. We compare both results to ensure that the model is not overfitting.",
"cite_spans": [
{
"start": 61,
"end": 86,
"text": "(Conneau and Kiela, 2018)",
"ref_id": "BIBREF5"
},
{
"start": 163,
"end": 190,
"text": "(Tijmen and Geoffrey, 2012)",
"ref_id": "BIBREF18"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Classification",
"sec_num": "4.3"
},
{
"text": "In this section, we comment the results obtained for each probing tasks. Table 3 shows the test accuracy and the mean accuracy of 5-fold cross validation. We only comment the latter, as it is more relevant due to the small datasets we have. The LR classifier obtains better results than the MLP for all tasks. Although losses in accuracy when using MLP can be explained by the lack of data, LR gives good results with BERT and SciB-ERT for binRel, swapped and scrambled, suggesting that those properties may be encoded linearly in the vectors.",
"cite_spans": [],
"ref_spans": [
{
"start": 73,
"end": 80,
"text": "Table 3",
"ref_id": "TABREF5"
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "Relation detection DR are almost as well encoded by BoV than by BERT and SciBERT, which can be explained by the fact that they are most of the time lexicalized. Contextual embeddings slightly improve the results for binRel (from \u2248 73% for BoV to 76.4 \u00b1 1.5), but not for semRel (61.3% for BoV to 57.9 \u00b1 2.2).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "By comparing two by two the results of BoV, BERT, and SciBERT encoders for binRel, we highlight that those models nearly give the same predictions (see Table. 4). Among the 290 sentences used as test, 94 are predicted differently by BERT and BoV, 85 by SciBERT and BoV, and 67 by SciBERT and BERT. Among the sentences that are similarly predicted, less than 20% correspond to common mistakes, and the others to common good predictions. For BERT and SciBERT, around 70% of the predictions are similar, showing that the models capture similar information for this task. For BERT and SciBERT with BoV respectively, more than 50% are common, but still they make a lot of different predictions, showing that they do not capture same aspects of the sentence.",
"cite_spans": [],
"ref_spans": [
{
"start": 152,
"end": 158,
"text": "Table.",
"ref_id": null
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "For semRel however, the models make very different predictions. Among the 144 examples used as test, 47 are predicted differently by BERT and BoV, 78 by SciBERT and BoV, and 71 by SciBERT and BERT. The results of Table 5 are consonant with this variation, showing that all encoders have in fact different abilities for the prediction.",
"cite_spans": [],
"ref_spans": [
{
"start": 213,
"end": 220,
"text": "Table 5",
"ref_id": "TABREF7"
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "We show both precision and recall obtained for SemRel in Table 5 . For attribution detection, BoV, BERT-uncased and SciBERT-cased gives the best recall (81%), for enablement it is BERT-uncased (67%), for manner-means it is BoV (78%), and for norel it is SciBERT-uncased (75%). These good recalls obviously result in lower precisions. However, as we want to find as many instances of a relation R, rather than maximizing the number of instances that are correctly found, we are still satisfied.",
"cite_spans": [],
"ref_spans": [
{
"start": 57,
"end": 64,
"text": "Table 5",
"ref_id": "TABREF7"
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "Coherence detection Representations build from BoV fail at predicting if units have been swapped (51% 10 ), and perform badly for predicting if units have been scrambled (\u2248 51.9%). Google Universal Sentence Encoder improves swapped only by 10% and scrambled by 4%. Representations built from contextual embeddings improve accuracy for swapped by almost 26% for BERT and 27.5% for SciBERT and for scrambled by almost 25% for BERT and 27.8% for SciBERT. This shows that training on STS, on which Google USE is trained, is not enough to capture coherence links between clauses, and that the \"Next Sentence Prediction\" (NSP) task on which both BERT and SciBERT are trained seems indeed to help a lot. This task enables somehow to account for the order of the units in the sentence. Even if BERT and SciBERT are not trained at clause level, the corpus on which they are trained is large and therefore contains a large number of sentences that are made of a single clause. This enables the model to learn clause level contingency relations, which are in fact the task that we probe here. binRel semRel BERT-BoV SciBERT-BoV SciBERT-BoV BERT-BoV SciBERT-BoV SciBERT -BoV # common predictions 196 205 223 97 66 73 # different predictions 94 85 67 47 78 71 # common errors 162 175 180 64 54 73 Table 4 : Two-by-two comparison of the predictions obtained by different encoders for binRel and semRel.",
"cite_spans": [],
"ref_spans": [
{
"start": 1158,
"end": 1314,
"text": "-BoV # common predictions 196 205 223 97 66 73 # different predictions 94 85 67 47 78 71 # common errors 162 175 180 64 54 73 Table 4",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "In this section, we discuss potential biases in the probing tasks that we designed, how we control them, and the possible improvements that could be further done.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6"
},
{
"text": "Relation prediction A bias that could affect DR detection is the length of the sentence. In particular, showed that BoV obtain 66.6% accuracy for predicting the length of the sentence, and get up to 99% with other more elaborated encoders such as BiLSTM or gated convolutional networks. We took that into account when selecting the different sentence sets. By including sentences that have more than one DU (sentences that only have relations with weak semantics) in the norel class , we somehow control the size of the sentences and ensure that the distributions of words per sentence are not specific to each class. The distributions are given in Fig. 3 , and are close enough to guarantee that the Sentence Length does not bias our predictions for binRel. Similarly, we ensured that the length of the sentence does not affect the detection of the relation semantics. The distributions which are given in Fig. 4 are close and, so, cannot be the only parameter involved in the good quality of the prediction.",
"cite_spans": [],
"ref_spans": [
{
"start": 649,
"end": 655,
"text": "Fig. 3",
"ref_id": "FIGREF2"
},
{
"start": 907,
"end": 914,
"text": "Fig. 4",
"ref_id": "FIGREF3"
}
],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6"
},
{
"text": "However, good performances of BoV for this task are influenced by the fact that most of them are linguistically signalled. Even if this bias is in fact related to the specific genre of the corpus that we study, we discuss here the linguistic properties that play a role in the various encodings and predictions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6"
},
{
"text": "Verbs such as \"we show\" or \"we demonstrate\" typically occur in sentences which contain an attribution relation. Conjunctions such as \"to\" or \"in order to\" express enablement relations. Finally, manner-means are often signalled by verb forms such as \"(by) +ing\". We think that these signals play an important role in the good performances of BoV especially for attribution and manner-means.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6"
},
{
"text": "Manner-means The sentence \"Most sentence embedding models typically represent each sentence only using word surface, which makes these models indiscriminate for ubiquitous homonymy and polysemy\" is a typical example of the sentences whose relation is predicted well with BoV encoders but neither by BERT, which predicts an attribution, nor by SciBERT, which predicts noRel. The verb \"using\" seem to be responsible for this good prediction. This underlies the good recall that BoV gets for the prediction of this DR, which gets a very bad recall for BERT and SciB-ERT. nated ] 4.\" is well predicted by all encoders, except SciBERT. Here, the captured relationship is a manner-means between units 3 and 4, which is in fact the deepest DR in the full structure of the sentence, and thus the less important. The second part (3+4) of the sentence is in conjunction with the first (1+2) and refers to it. The manner-means between 3 and 4 is thus not the more salient in the RST, but is however well predicted. This suggests that the choices that we did to ignore non-logical DRs work well and that the training made on sufficient data suffice to make the classifier focus on the relations that we consider.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6"
},
{
"text": "Enablement With BERT, most of the sentences that contain the connective \"to\" are well detected as an enablement relation. Among those that are not well classified, most of them are classified as attribution or norel. One common characteristic to the sentences that are classified as norel, is that they are often long and contain more than 2 DUs and several discourse connectives, including \"to\" or \"for\". It is thus harder for the model to understand and distinguish the role of the connectives, as they are drowned in a lot of lexical information, leading to errors because of the difficulty of the task. For sentences that are wrongly classified as attribution, we observe a similar behaviour. The sentence \"[Recent work has shown success] [in using continuous word embeddings] [to improve supervised NLP systems.]\" contains an enablement between segments 1 and 2, but is classified as an attribution. This errors seems to be due to the verb \"shown\" that appears a lot in sentences that contain an attribution. This shows that even if BERT improves a sort of coherence detection, it in fact relies a lot on lexical cues for detecting the semantics of relations.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6"
},
{
"text": "Swapped and scrambled We admit that the tasks designed for probing coherence are somehow arbitrary, and produces different biases that could be better controlled. We present two major issues, and propose solutions for future work. The first issue concerns the swapped task, where modifications of sentences may produce other coherent sentences, which are thus tagged as incoherent in the training set and used as it by the classifier. Those cases specifically correspond to sentences that are formed by two DUs and linked by a particle (such as \"by\", \"for\" or \"to\"), which are thus swapped and still coherent. For example, the sentence \"[By incorporating textual information,][RCM can effectively deal with data sparseness problem.]\" becomes \"[RCM can effectively deal with data sparseness problem][by incorporating textual information.]\". We expect such sentences to be poorly predicted, but it is not the case, showing that BERT and SciBERT probably rely on other indices. also created acceptable sentences for three of their probing tasks, and filtered them by crowd-sourcing, asking people to rate sentences according to their acceptability. This method would be the best suited to solve this issue, as it may be hard to control coherence automatically with syntactic parsers.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6"
},
{
"text": "Secondly, we did not ensure that the property captured does not rely on syntax, because sentence modifications may also produce sentences that are syntactically incorrect and that we did not filter out. Here the problematic cases mostly come from two clauses sentences containing a conjunction such as \"that\" or \"and\". For example the sentence \"[We present a human judgments dataset] [and an adapted metric for evaluation of Arabic machine translation.]\" becomes \"[And an adapted metric for evaluation of Arabic machine translation] [we present a human judgments dataset.]\", thus being obviously syntactically incorrect, and possibly captured as incoherent based on this syntactic property. A possible solution to that issue could be to forbid the swap of the first and last DU, thus inducing another different bias and in our case reducing the size of the corpus. Another possibility is to introduce rules based on the syntactic tree of new sentences in order to filter those cases. In this preliminary work, we introduced the second coherence detection task (scrambled) with the aim of reducing those biases.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6"
},
{
"text": "We introduced four tasks for probing the discourse properties involved in sentences from scientific abstracts. We evaluated the ability of a classifier to predict a discourse property of a sentence from its embedding. The performance of the classifier highlights the extent to which discourse properties are encoded in those representations.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "7"
},
{
"text": "We showed that coherence links are captured by vectors made from contextual models, as well as DRs, but that those models in fact do not encode the semantics of those DRs. We highlighted that BoV embeddings perform nearly as good as contextual embeddings for both DRs detection tasks, highlighting the fact that these relations are most of the time explicitly signalled in scientific abstracts.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "7"
},
{
"text": "This confirms our hypothesis that BERT and SciBERT training suffice to encode coherence to some extent. We think that the Next Sentence Prediction task is the reason for this performance, as it allows the model to learn adjacency properties of sentences, and thus of clauses. The second hypothesis however is not confirmed, as we concluded from the experiments that semantics of DRs are as well captured by BoV, as BERT or SciBERT.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "7"
},
{
"text": "The present work opens various possibilities, for improving the tasks as well as the analysis of the results. We decided here to focus on the detection of three specific relations that occur within sentences. We adopted strategies to make this possible, for example by ignoring DRs that are of continuation or additive and not logical. We could however go further by considering all relations involved in the sentences. A multi-label classification could be a solution to the problem of predicting all DRs of a sentence from its embedding. A Sequence-to-Sequence model could be a solution to the problem of predicting the discourse structure (or sequence of relations) from its embedding. For coherence prediction, we also plan to determine to what extent the tasks we introduced (swapped and scrambled) depend on syntactic properties. A statistical analysis of the syntactic structures involved could help to clarify the possible biases coming from the syntax. We also plan to further check the datasets that we introduce for those tasks, as we have raised that our method created coherent sentences, labeled as incoherent in our training dataset.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "7"
},
{
"text": "Other corpora such as RST-DT (Carlson and Okurowski, 2002) or PDTB (Webber et al., 2005) , which could be combined, could be used to produce other datasets for probing DRs, enabling to evaluate how other DRs (such as contrast or reason) are encoded in sentence embeddings.",
"cite_spans": [
{
"start": 29,
"end": 58,
"text": "(Carlson and Okurowski, 2002)",
"ref_id": null
},
{
"start": 67,
"end": 88,
"text": "(Webber et al., 2005)",
"ref_id": "BIBREF19"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "7"
},
{
"text": "https://www.sfu.ca/rst/01intro/ definitions.html 2 They admit however that the boundary they propose between syntactic and semantic tasks is somewhat arbitrary.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "https://github.com/PKU-TANGENT/SciDTB 4 https://www.aclweb.org/anthology/ 5 We refer the reader to the paper(Yang and Li, 2018) to get explanations on the adaptations and annotation procedure. The relations set is recalled inTable 1.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "https://huggingface.co/",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "https://github.com/facebookresearch/SentEval",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "We consider that an accuracy close or less than 50% corresponds to random prediction, because a classifier predicting each class with a probability of 0.5 would have an accuracy of 50%.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "We want to thank the reviewers for their constructive comments and suggestions.This work was supported partly by the french PIA project \"Lorraine Universit\u00e9 d'Excellence\", reference ANR-15-IDEX-04-LUE.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgments",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Fine-grained analysis of sentence embeddings using auxiliary prediction tasks",
"authors": [
{
"first": "Yossi",
"middle": [],
"last": "Adi",
"suffix": ""
},
{
"first": "Einat",
"middle": [],
"last": "Kermany",
"suffix": ""
},
{
"first": "Yonatan",
"middle": [],
"last": "Belinkov",
"suffix": ""
},
{
"first": "Ofer",
"middle": [],
"last": "Lavi",
"suffix": ""
},
{
"first": "Yoav",
"middle": [],
"last": "Goldberg",
"suffix": ""
}
],
"year": 2016,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yossi Adi, Einat Kermany, Yonatan Belinkov, Ofer Lavi, and Yoav Goldberg. 2016. Fine-grained anal- ysis of sentence embeddings using auxiliary predic- tion tasks. CoRR, abs/1608.04207.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "SciB-ERT: A pretrained language model for scientific text",
"authors": [
{
"first": "Iz",
"middle": [],
"last": "Beltagy",
"suffix": ""
},
{
"first": "Kyle",
"middle": [],
"last": "Lo",
"suffix": ""
},
{
"first": "Arman",
"middle": [],
"last": "Cohan",
"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": "3615--3620",
"other_ids": {
"DOI": [
"10.18653/v1/D19-1371"
]
},
"num": null,
"urls": [],
"raw_text": "Iz Beltagy, Kyle Lo, and Arman Cohan. 2019. SciB- ERT: A pretrained language model for scientific text. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Lan- guage Processing (EMNLP-IJCNLP), pages 3615- 3620, Hong Kong, China. Association for Computa- tional Linguistics.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Building a Discourse-Tagged Corpus in the Framework of Rhetorical Structure Theory. Current Directions in Discourse and Dialogue",
"authors": [
{
"first": "Lynn",
"middle": [],
"last": "Carlson",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Marcu",
"suffix": ""
},
{
"first": "Mary",
"middle": [
"Ellen"
],
"last": "Okurowski",
"suffix": ""
}
],
"year": 2003,
"venue": "",
"volume": "",
"issue": "",
"pages": "85--112",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Lynn Carlson, Daniel Marcu, and Mary Ellen Okurowski. 2003. Building a Discourse-Tagged Corpus in the Framework of Rhetorical Structure Theory. Current Directions in Discourse and Dia- logue, pages 85-112.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Chris Tar, Yun-Hsuan Sung, Brian Strope, and Ray Kurzweil",
"authors": [
{
"first": "Daniel",
"middle": [],
"last": "Cer",
"suffix": ""
},
{
"first": "Yinfei",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "Sheng-Yi",
"middle": [],
"last": "Kong",
"suffix": ""
},
{
"first": "Nan",
"middle": [],
"last": "Hua",
"suffix": ""
},
{
"first": "Nicole",
"middle": [],
"last": "Limtiaco",
"suffix": ""
},
{
"first": "Rhomni",
"middle": [],
"last": "St",
"suffix": ""
},
{
"first": "Noah",
"middle": [],
"last": "John",
"suffix": ""
},
{
"first": "Mario",
"middle": [],
"last": "Constant",
"suffix": ""
},
{
"first": "Steve",
"middle": [],
"last": "Guajardo-Cespedes",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Yuan",
"suffix": ""
}
],
"year": 2018,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Daniel Cer, Yinfei Yang, Sheng-yi Kong, Nan Hua, Nicole Limtiaco, Rhomni St. John, Noah Con- stant, Mario Guajardo-Cespedes, Steve Yuan, Chris Tar, Yun-Hsuan Sung, Brian Strope, and Ray Kurzweil. 2018. Universal sentence encoder. CoRR, abs/1803.11175.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "SentEval: An evaluation toolkit for universal sentence representations",
"authors": [
{
"first": "Alexis",
"middle": [],
"last": "Conneau",
"suffix": ""
},
{
"first": "Douwe",
"middle": [],
"last": "Kiela",
"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": "Alexis Conneau and Douwe Kiela. 2018. SentEval: An evaluation toolkit for universal sentence representa- tions. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA).",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "What you can cram into a single $&!#* vector: Probing sentence embeddings for linguistic properties",
"authors": [
{
"first": "Alexis",
"middle": [],
"last": "Conneau",
"suffix": ""
},
{
"first": "German",
"middle": [],
"last": "Kruszewski",
"suffix": ""
},
{
"first": "Guillaume",
"middle": [],
"last": "Lample",
"suffix": ""
},
{
"first": "Lo\u00efc",
"middle": [],
"last": "Barrault",
"suffix": ""
},
{
"first": "Marco",
"middle": [],
"last": "Baroni",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics",
"volume": "1",
"issue": "",
"pages": "2126--2136",
"other_ids": {
"DOI": [
"10.18653/v1/P18-1198"
]
},
"num": null,
"urls": [],
"raw_text": "Alexis Conneau, German Kruszewski, Guillaume Lam- ple, Lo\u00efc Barrault, and Marco Baroni. 2018. What you can cram into a single $&!#* vector: Probing sentence embeddings for linguistic properties. In Proceedings of the 56th Annual Meeting of the As- sociation for Computational Linguistics (Volume 1: Long Papers), pages 2126-2136, Melbourne, Aus- tralia. Association for Computational Linguistics.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"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
},
"BIBREF8": {
"ref_id": "b8",
"title": "Cohesion in english",
"authors": [
{
"first": "M",
"middle": [
"A"
],
"last": "Halliday",
"suffix": ""
},
{
"first": "Ruqaiya",
"middle": [],
"last": "Hasan",
"suffix": ""
}
],
"year": 1976,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "M.A.K Halliday and Ruqaiya Hasan. 1976. Cohesion in english. Longman.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Raquel Urtasun, and Sanja Fidler",
"authors": [
{
"first": "Ryan",
"middle": [],
"last": "Kiros",
"suffix": ""
},
{
"first": "Yukun",
"middle": [],
"last": "Zhu",
"suffix": ""
},
{
"first": "Ruslan",
"middle": [],
"last": "Salakhutdinov",
"suffix": ""
},
{
"first": "Richard",
"middle": [
"S"
],
"last": "Zemel",
"suffix": ""
},
{
"first": "Antonio",
"middle": [],
"last": "Torralba",
"suffix": ""
}
],
"year": 2015,
"venue": "Skip-thought vectors. CoRR",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ryan Kiros, Yukun Zhu, Ruslan Salakhutdinov, Richard S. Zemel, Antonio Torralba, Raquel Urta- sun, and Sanja Fidler. 2015. Skip-thought vectors. CoRR, abs/1506.06726.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Segmented Discourse Representation Theory: Dynamic Semantics With Discourse Structure",
"authors": [
{
"first": "Alex",
"middle": [],
"last": "Lascarides",
"suffix": ""
},
{
"first": "Nicholas",
"middle": [],
"last": "Asher",
"suffix": ""
}
],
"year": 2007,
"venue": "Computing Meaning",
"volume": "3",
"issue": "",
"pages": "",
"other_ids": {
"DOI": [
"http://link.springer.com/10.1007/978-1-4020-5958-2_5"
]
},
"num": null,
"urls": [],
"raw_text": "Alex Lascarides and Nicholas Asher. 2007. Segmented Discourse Representation Theory: Dynamic Seman- tics With Discourse Structure. In Harry Bunt and Reinhard Muskens, editors, Computing Meaning, volume 3. Springer Netherlands, Dordrecht.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "The discourse-level structure of empirical abstracts: An exploratory study",
"authors": [
{
"first": "Elizabeth",
"middle": [],
"last": "Duross",
"suffix": ""
},
{
"first": "Liddy",
"middle": [],
"last": "",
"suffix": ""
}
],
"year": 1991,
"venue": "Information Processing & Management",
"volume": "27",
"issue": "1",
"pages": "55--81",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Elizabeth DuRoss Liddy. 1991. The discourse-level structure of empirical abstracts: An exploratory study. Information Processing & Management, 27(1):55-81.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "A dependency perspective on RST discourse parsing and evaluation",
"authors": [
{
"first": "Mathieu",
"middle": [],
"last": "Morey",
"suffix": ""
},
{
"first": "Philippe",
"middle": [],
"last": "Muller",
"suffix": ""
},
{
"first": "Nicholas",
"middle": [],
"last": "Asher",
"suffix": ""
}
],
"year": 2018,
"venue": "Computational Linguistics",
"volume": "44",
"issue": "2",
"pages": "197--235",
"other_ids": {
"DOI": [
"10.1162/COLI_a_00314"
]
},
"num": null,
"urls": [],
"raw_text": "Mathieu Morey, Philippe Muller, and Nicholas Asher. 2018. A dependency perspective on RST discourse parsing and evaluation. Computational Linguistics, 44(2):197-235.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "GloVe: Global vectors for word representation",
"authors": [
{
"first": "Jeffrey",
"middle": [],
"last": "Pennington",
"suffix": ""
},
{
"first": "Richard",
"middle": [],
"last": "Socher",
"suffix": ""
},
{
"first": "Christopher",
"middle": [],
"last": "Manning",
"suffix": ""
}
],
"year": 2014,
"venue": "Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
"volume": "",
"issue": "",
"pages": "1532--1543",
"other_ids": {
"DOI": [
"10.3115/v1/D14-1162"
]
},
"num": null,
"urls": [],
"raw_text": "Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. GloVe: Global vectors for word representation. In Proceedings of the 2014 Confer- ence on Empirical Methods in Natural Language Processing (EMNLP), pages 1532-1543, Doha, Qatar. Association for Computational Linguistics.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Sentence-BERT: Sentence embeddings using Siamese BERTnetworks",
"authors": [
{
"first": "Nils",
"middle": [],
"last": "Reimers",
"suffix": ""
},
{
"first": "Iryna",
"middle": [],
"last": "Gurevych",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)",
"volume": "",
"issue": "",
"pages": "3982--3992",
"other_ids": {
"DOI": [
"10.18653/v1/D19-1410"
]
},
"num": null,
"urls": [],
"raw_text": "Nils Reimers and Iryna Gurevych. 2019. Sentence- BERT: Sentence embeddings using Siamese BERT- networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natu- ral Language Processing (EMNLP-IJCNLP), pages 3982-3992, Hong Kong, China. Association for Computational Linguistics.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "The Texture of Discourse",
"authors": [],
"year": 2009,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jan Renkema. 2009. The Texture of Discourse. John Benjamins Publishing Company.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Does string-based neural MT learn source syntax?",
"authors": [
{
"first": "Xing",
"middle": [],
"last": "Shi",
"suffix": ""
},
{
"first": "Inkit",
"middle": [],
"last": "Padhi",
"suffix": ""
},
{
"first": "Kevin",
"middle": [],
"last": "Knight",
"suffix": ""
}
],
"year": 2016,
"venue": "Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "1526--1534",
"other_ids": {
"DOI": [
"10.18653/v1/D16-1159"
]
},
"num": null,
"urls": [],
"raw_text": "Xing Shi, Inkit Padhi, and Kevin Knight. 2016. Does string-based neural MT learn source syntax? In Pro- ceedings of the 2016 Conference on Empirical Meth- ods in Natural Language Processing, pages 1526- 1534, Austin, Texas. Association for Computational Linguistics.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Rhetorical structure theory: Looking back and moving ahead. Discourse studies",
"authors": [
{
"first": "Maite",
"middle": [],
"last": "Taboada",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "William",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Mann",
"suffix": ""
}
],
"year": 2006,
"venue": "",
"volume": "8",
"issue": "",
"pages": "423--459",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Maite Taboada and William C Mann. 2006. Rhetorical structure theory: Looking back and moving ahead. Discourse studies, 8(3):423-459.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude. COURSERA: Neural networks for machine learning",
"authors": [
{
"first": "Tieleman",
"middle": [],
"last": "Tijmen",
"suffix": ""
},
{
"first": "Hinton",
"middle": [],
"last": "Geoffrey",
"suffix": ""
}
],
"year": 2012,
"venue": "",
"volume": "4",
"issue": "",
"pages": "26--31",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Tieleman Tijmen and Hinton Geoffrey. 2012. Lecture 6.5-rmsprop: Divide the gradient by a running aver- age of its recent magnitude. COURSERA: Neural networks for machine learning, 4:26-31.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "A short introduction to the penn discourse treebank",
"authors": [
{
"first": "Bonnie",
"middle": [],
"last": "Webber",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Aravind",
"suffix": ""
},
{
"first": "Eleni",
"middle": [],
"last": "Joshi",
"suffix": ""
},
{
"first": "Rashmi",
"middle": [],
"last": "Miltsakaki",
"suffix": ""
},
{
"first": "Nikhil",
"middle": [],
"last": "Prasad",
"suffix": ""
},
{
"first": "Alan",
"middle": [],
"last": "Dinesh",
"suffix": ""
},
{
"first": "Katherine",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Forbes",
"suffix": ""
}
],
"year": 2005,
"venue": "Copenhagen Working Papers in Language and Speech Processing",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Bonnie Webber, Aravind K Joshi, Eleni Miltsakaki, Rashmi Prasad, Nikhil Dinesh, Alan Lee, and Katherine Forbes. 2005. A short introduction to the penn discourse treebank. Copenhagen Working Pa- pers in Language and Speech Processing.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Huggingface's transformers: Stateof-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\u00e9mi",
"middle": [],
"last": "Louf",
"suffix": ""
},
{
"first": "Morgan",
"middle": [],
"last": "Funtowicz",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"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\u00e9mi Louf, Morgan Funtow- icz, et al. 2019. Huggingface's transformers: State- of-the-art natural language processing. ArXiv.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "SciDTB: Discourse dependency TreeBank for scientific abstracts",
"authors": [
{
"first": "An",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "Sujian",
"middle": [],
"last": "Li",
"suffix": ""
}
],
"year": 2018,
"venue": "Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics",
"volume": "2",
"issue": "",
"pages": "444--449",
"other_ids": {
"DOI": [
"10.18653/v1/P18-2071"
]
},
"num": null,
"urls": [],
"raw_text": "An Yang and Sujian Li. 2018. SciDTB: Discourse de- pendency TreeBank for scientific abstracts. In Pro- ceedings of the 56th Annual Meeting of the Associa- tion for Computational Linguistics (Volume 2: Short Papers), pages 444-449, Melbourne, Australia. As- sociation for Computational Linguistics.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"uris": null,
"type_str": "figure",
"num": null,
"text": "Example of dependency annotation from SciDTB."
},
"FIGREF1": {
"uris": null,
"type_str": "figure",
"num": null,
"text": "Distribution of relations in sentences."
},
"FIGREF2": {
"uris": null,
"type_str": "figure",
"num": null,
"text": "Distribution of the number of words per sentences for binRel."
},
"FIGREF3": {
"uris": null,
"type_str": "figure",
"num": null,
"text": "Distribution of the number words per sentences for semRel."
},
"TABREF1": {
"content": "<table/>",
"num": null,
"type_str": "table",
"html": null,
"text": "Discourse relations in SciDTB corpus"
},
"TABREF3": {
"content": "<table/>",
"num": null,
"type_str": "table",
"html": null,
"text": "Datasets description."
},
"TABREF5": {
"content": "<table/>",
"num": null,
"type_str": "table",
"html": null,
"text": "Test and 5-fold mean accuracies of LR and MLP classifiers on each task for each sentence representation."
},
"TABREF7": {
"content": "<table/>",
"num": null,
"type_str": "table",
"html": null,
"text": "Precision and Recall for SemRel."
}
}
}
} |