File size: 82,768 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 | {
"paper_id": "D09-1036",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T16:38:17.720820Z"
},
"title": "Recognizing Implicit Discourse Relations in the Penn Discourse Treebank",
"authors": [
{
"first": "Ziheng",
"middle": [],
"last": "Lin",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "National University of Singapore",
"location": {
"addrLine": "13 Computing Drive",
"postCode": "117417",
"country": "Singapore"
}
},
"email": "linzihen@comp.nus.edu.sg"
},
{
"first": "Min-Yen",
"middle": [],
"last": "Kan",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "National University of Singapore",
"location": {
"addrLine": "13 Computing Drive",
"postCode": "117417",
"country": "Singapore"
}
},
"email": ""
},
{
"first": "Hwee",
"middle": [
"Tou"
],
"last": "Ng",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "National University of Singapore",
"location": {
"addrLine": "13 Computing Drive",
"postCode": "117417",
"country": "Singapore"
}
},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "We present an implicit discourse relation classifier in the Penn Discourse Treebank (PDTB). Our classifier considers the context of the two arguments, word pair information, as well as the arguments' internal constituent and dependency parses. Our results on the PDTB yields a significant 14.1% improvement over the baseline. In our error analysis, we discuss four challenges in recognizing implicit relations in the PDTB.",
"pdf_parse": {
"paper_id": "D09-1036",
"_pdf_hash": "",
"abstract": [
{
"text": "We present an implicit discourse relation classifier in the Penn Discourse Treebank (PDTB). Our classifier considers the context of the two arguments, word pair information, as well as the arguments' internal constituent and dependency parses. Our results on the PDTB yields a significant 14.1% improvement over the baseline. In our error analysis, we discuss four challenges in recognizing implicit relations in the PDTB.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "In the field of discourse modeling, it is widely agreed that text is not understood in isolation, but in relation to its context. One focus in the study of discourse is to identify and label the relations between textual units (clauses, sentences, or paragraphs). Such research can enable downstream natural language processing (NLP) such as summarization, question answering, and textual entailment. For example, recognizing causal relations can assist in answering why questions. Detecting contrast and restatements is useful for paraphrasing and summarization systems. While different discourse frameworks have been proposed from different perspectives (Mann and Thompson, 1988; Hobbs, 1990; Lascarides and Asher, 1993; Knott and Sanders, 1998; Webber, 2004) , most admit these basic types of discourse relationships between textual units.",
"cite_spans": [
{
"start": 656,
"end": 681,
"text": "(Mann and Thompson, 1988;",
"ref_id": "BIBREF6"
},
{
"start": 682,
"end": 694,
"text": "Hobbs, 1990;",
"ref_id": "BIBREF1"
},
{
"start": 695,
"end": 722,
"text": "Lascarides and Asher, 1993;",
"ref_id": "BIBREF4"
},
{
"start": 723,
"end": 747,
"text": "Knott and Sanders, 1998;",
"ref_id": "BIBREF3"
},
{
"start": 748,
"end": 761,
"text": "Webber, 2004)",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "When there is a discourse connective (e.g., because) between two text spans, it is often easy to recognize the relation between the spans, as most connectives are unambiguous (Miltsakaki et al., 2005; Pitler et al., 2008) . On the other hand, it is difficult to recognize the discourse relations when there are no explicit textual cues. We term these cases explicit and implicit relations, respectively.",
"cite_spans": [
{
"start": 175,
"end": 200,
"text": "(Miltsakaki et al., 2005;",
"ref_id": "BIBREF9"
},
{
"start": 201,
"end": 221,
"text": "Pitler et al., 2008)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "While the recognition of discourse structure has been studied in the context of explicit relations (Marcu, 1998) in the past, little published work has yet attempted to recognize implicit discourse relations between text spans.",
"cite_spans": [
{
"start": 99,
"end": 112,
"text": "(Marcu, 1998)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Detecting implicit relations is a critical step in forming a discourse understanding of text, as many text spans do not mark their discourse relations with explicit cues. Recently, the Penn Discourse Treebank (PDTB) has been released, which features discourse level annotation on both explicit and implicit relations. It provides a valuable linguistic resource towards understanding discourse relations and a common platform for researchers to develop discourse-centric systems. With the recent release of the second version of this corpus (Prasad et al., 2008) , which provides a cleaner and more thorough implicit relation annotation, there is an opportunity to address this area of work.",
"cite_spans": [
{
"start": 540,
"end": 561,
"text": "(Prasad et al., 2008)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In this paper, we provide classification of implicit discourse relations on the second version of the PDTB. The features we used include contextual modeling of relation dependencies, features extracted from constituent parse trees and dependency parse trees, and word pair features. We show an accuracy of 40.2%, which is a significant improvement of 14.1% over the majority baseline.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "After reviewing related work, we first give an overview of the Penn Discourse Treebank. We then describe our classification methodology, followed by experimental results. We give a detailed discussion on the difficulties of implicit relation classification in the PDTB, and then conclude the paper.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "One of the first works that use statistical methods to detect implicit discourse relations is that of Marcu and Echihabi (2002) . They showed that word pairs extracted from two text spans provide clues for detecting the discourse relation between the text spans. They used a set of textual patterns to automatically construct a large corpus of text span pairs from the web. These text spans were assumed to be instances of specific discourse relations. They removed the discourse connectives from the pairs to form an implicit relation corpus. From this corpus, they collected word pair statistics, which were used in a Na\u00efve Bayes framework to classify discourse relations. Saito et al. (2006) extended this theme, to show that phrasal patterns extracted from a text span pair provide useful evidence in the relation classification. For example, the pattern \"... should have done ...\" usually signals a contrast. The authors combined word pairs with phrasal patterns, and conducted experiments with these two feature classes to recognize implicit relations between adjacent sentences in a Japanese corpus.",
"cite_spans": [
{
"start": 102,
"end": 127,
"text": "Marcu and Echihabi (2002)",
"ref_id": "BIBREF7"
},
{
"start": 675,
"end": 694,
"text": "Saito et al. (2006)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "Both of these previous works have the shortcoming of downgrading explicit relations to implicit ones by removing the explicit discourse connectives. While this is a good approach to automatically create large corpora, natively implicit relations may be signaled in different ways. The fact that explicit relations are explicitly signaled indicates that such relations need a cue to be unambiguous to human readers. Thus, such an artificial implicit relation corpus may exhibit marked differences from a natively implicit one. We validate this claim later in this work. Wellner et al. (2006) used multiple knowledge sources to produce syntactic and lexico-semantic features, which were then used to automatically identify and classify explicit and implicit discourse relations in the Discourse Graphbank (Wolf and Gibson, 2005) . Their experiments show that discourse connectives and the distance between the two text spans have the most impact, and event-based features also contribute to the performance. However, their system may not work well for implicit relations alone, as the two most prominent features only apply to explicit relations: implicit relations do not have discourse connectives and the two text spans of an implicit relation are usually adjacent to each other.",
"cite_spans": [
{
"start": 569,
"end": 590,
"text": "Wellner et al. (2006)",
"ref_id": "BIBREF17"
},
{
"start": 803,
"end": 826,
"text": "(Wolf and Gibson, 2005)",
"ref_id": "BIBREF18"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "The work that is most related to ours is the forthcoming paper of Pitler et al. (2009) on implicit relation classification on the second version of the PDTB. They performed classification of implicit discourse relations using several linguistically informed features, such as word polar-ity, verb classes, and word pairs, showing performance increases over a random classification baseline.",
"cite_spans": [
{
"start": 66,
"end": 86,
"text": "Pitler et al. (2009)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "The Penn Discourse Treebank (PDTB) is a discourse level annotation (Prasad et al., 2008) over the one million word Wall Street Journal corpus. The PDTB adopts the predicate-argument view of discourse relations, where a discourse connective (e.g., because) is treated as a predicate that takes two text spans as its arguments. The argument that the discourse connective structurally attaches to is called Arg2, and the other argument is called Arg1. The PDTB provides annotations for explicit and implicit discourse relations. By definition, an explicit relation contains an explicit discourse connective. In the PDTB, 100 explicit connectives are annotated. Example 1 shows an explicit Contrast relation signaled by the discourse connective but.",
"cite_spans": [
{
"start": 67,
"end": 88,
"text": "(Prasad et al., 2008)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Overview of the Penn Discourse Treebank",
"sec_num": "3"
},
{
"text": "The last line shows the relation type and the file in the PDTB from which the example is drawn.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Overview of the Penn Discourse Treebank",
"sec_num": "3"
},
{
"text": "(1) Arg1: In any case, the brokerage firms are clearly moving faster to create new ads than they did in the fall of 1987. Arg2: But it remains to be seen whether their ads will be any more effective. (Contrast -wsj 2201)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Overview of the Penn Discourse Treebank",
"sec_num": "3"
},
{
"text": "In the PDTB, implicit relations are constrained by adjacency: only pairs of adjacent sentences within paragraphs are examined for the existence of implicit relations. When an implicit relation was inferred by an annotator, he/she inserted an implicit connective that best reflects the relation. Example 2 shows an implicit relation, where the annotator inferred a Cause relation and inserted an implicit connective so (i.e., the original text does not include so). The text in the box (he says) shows the attribution, i.e., the agent that expresses the arguments. The PDTB provides annotation for the attributions and supplements of the arguments.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Overview of the Penn Discourse Treebank",
"sec_num": "3"
},
{
"text": "(2) Arg1: \"A lot of investor confidence comes from the fact that they can speak to us,\" he says . Arg2: [so] \"To maintain that dialogue is absolutely crucial.\" (Cause -wsj 2201)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Overview of the Penn Discourse Treebank",
"sec_num": "3"
},
{
"text": "The PDTB provides a three level hierarchy of relation tags for its annotation. The first level consists of four major relation classes: Temporal, Contingency, Comparison, and Expansion. For each class, a second level of types is defined to provide finer semantic distinctions. A third level of subtypes is defined for only some types to specify the semantic contribution of each argument. Relation classes and types in the PDTB are reproduced in the first two columns of Table 1 .",
"cite_spans": [],
"ref_spans": [
{
"start": 471,
"end": 478,
"text": "Table 1",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Overview of the Penn Discourse Treebank",
"sec_num": "3"
},
{
"text": "We focus on implicit relation classification of the Level 2 types in the PDTB, as we feel that Level 1 classes are too general and coarse-grained for downstream applications, while Level 3 subtypes are too fine-grained and are only provided for some types. Table 1 shows the distribution of the 16 Level 2 relation types of the implicit relations from the training sections, i.e., Sections 2 -21. As there are too few training instances for Condition, Pragmatic Condition, Pragmatic Contrast, Pragmatic Concession, and Exception, we removed these five types from further consideration. We thus use the remaining 11 Level 2 types in our work. The initial distribution and adjusted distribution are shown in the last two columns of the table. We see that the three predominant types are Cause (25.63%), Conjunction (22.25%), and Restatement (19.23%). ",
"cite_spans": [],
"ref_spans": [
{
"start": 257,
"end": 264,
"text": "Table 1",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Overview of the Penn Discourse Treebank",
"sec_num": "3"
},
{
"text": "Our implicit relation classifier is built using supervised learning on a maximum entropy classifier. As such, our approach processes the annotated argument pairs into binary feature vectors suitable for use in training a classifier. Attributions and supplements are ignored from the relations, as our system does not make use of them. We chose the following four classes of features as they represent a wide range of information -contextual, syntactic, and lexical -that have been shown to be helpful in previous works and tasks. We now discuss the four categories of features used in our framework. Contextual Features. Lee et al. (2006) showed that there are a variety of possible dependencies between pairs of discourse relations: independent, fully embedded argument, shared argument, properly contained argument, pure crossing, and partially overlapping argument. They argued that the last three cases -properly contained argument, pure crossing, and partially overlapping argument -can be factored out by appealing to discourse notions such as anaphora and attribution. Moreover, we also observed from the PDTB corpus that fully embedded argument and shared argument are the most common patterns, which are shown in Figure 1. The top portion of Figure 1 shows a case where relation r 1 is fully embedded in Arg1 of relation r 2 , and the bottom portion shows r 1 and r 2 sharing an argument. We model these two patterns as contextual features. We believe that these discourse dependency patterns between a pair of adjacent relations are useful in identifying the relations. For example, if we have three items in a list, according to the PDTB binary predicate-argument definitions, there will be a List relation between the first item and the second item, and another List relation between the previous List relation and the third item, where the previous List relation is fully embedded in Arg1 of the current List relation. As we are using the gold standard argument segmentation from the PDTB, we can extract and leverage these dependency patterns. For each relation curr, we use the previous relation prev and the next relation next as evidence to fire six binary features, as defined in Table 2 .",
"cite_spans": [
{
"start": 621,
"end": 638,
"text": "Lee et al. (2006)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [
{
"start": 1222,
"end": 1228,
"text": "Figure",
"ref_id": null
},
{
"start": 1251,
"end": 1259,
"text": "Figure 1",
"ref_id": "FIGREF0"
},
{
"start": 2198,
"end": 2205,
"text": "Table 2",
"ref_id": "TABREF2"
}
],
"eq_spans": [],
"section": "Methodology",
"sec_num": "4"
},
{
"text": "Note that while curr is an implicit relation to be classified, both prev and next can be implicit or explicit relations. Pitler et al. (2008) showed that the type of a relation sometimes correlates to the type of its adjacent relation. When the adjacent relation is explicit, its type may be suggested by its discourse connective. Thus we include another two groups of contextual features representing the connectives of prev and next when they are explicit relations.",
"cite_spans": [
{
"start": 121,
"end": 141,
"text": "Pitler et al. (2008)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Methodology",
"sec_num": "4"
},
{
"text": "prev embedded in curr.Arg1 next embedded in curr.Arg2 curr embedded in prev.Arg2 curr embedded in next.Arg1 Shared argument:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Fully embedded argument:",
"sec_num": null
},
{
"text": "prev.Arg2 = curr.Arg1 curr.Arg2 = next.Arg1 (Jiang and Ng, 2006) and tree fragments (Moschitti, 2004) . From our observation of the PDTB relations, syntactic structure within one argument may constrain the relation type and the syntactic structure of the other argument. For example, the constituent parse structure in Figure 2 (a) usually signals an Asynchronous relation when it appears in Arg2, as shown in Example 3, while the structure in Figure 2(b) usually acts as a clue for a Cause relation when it appears in Arg1, as shown in Example 4. In both examples, the lexicalized parts of the parse structure are bolded.",
"cite_spans": [
{
"start": 44,
"end": 64,
"text": "(Jiang and Ng, 2006)",
"ref_id": "BIBREF2"
},
{
"start": 84,
"end": 101,
"text": "(Moschitti, 2004)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [
{
"start": 319,
"end": 327,
"text": "Figure 2",
"ref_id": null
},
{
"start": 444,
"end": 455,
"text": "Figure 2(b)",
"ref_id": null
}
],
"eq_spans": [],
"section": "Fully embedded argument:",
"sec_num": null
},
{
"text": "(3) Arg1: But the RTC also requires \"working\" capital to maintain the bad assets of thrifts that are sold For Arg1 and Arg2 of each relation, we extract the corresponding gold standard syntactic parse trees from the corpus. As an argument can be a single sentence, a clause, or multiple sentences, this results in a whole parse tree, parts of a parse tree, or multiple parse trees. From these parses, we extract all possible production rules. Although the structures shown in Figure 2 are tree fragments, tree fragments are not extracted as production rules act as generalization of tree fragments. As an example, Figure 3 shows the parse tree for Arg1 of an implicit discourse relation from the text wsj 2224. As Arg1 is a clause, the extracted tree is a subtree. We then collect all production rules from this subtree, with function tags (e.g., SBJ) removed from internal nodes. POS tag to word production rules are collected as well. The resulting production rules include ones such as: S \u2192 NP VP, NP \u2192 PRP, PRP \u2192 \"We\", etc. Each production rule is represented as three binary features to check whether this rule appears in Arg1, Arg2, and both arguments.",
"cite_spans": [],
"ref_spans": [
{
"start": 476,
"end": 484,
"text": "Figure 2",
"ref_id": null
},
{
"start": 614,
"end": 622,
"text": "Figure 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Fully embedded argument:",
"sec_num": null
},
{
"text": "Dependency Parse Features. We also experimented with features extracted from dependency trees of the arguments. We used the Stanford dependency parser (de Marneffe et al., 2006) , which takes in a constituent parse tree and produces a dependency tree. Again, for an argument, we may collect a whole dependency tree, parts of a tree, or multiple trees, depending on the span of the argument. The reason for using dependency trees is that they encode additional information at the word level that is not explicitly present in the constituent trees. From each tree, we collect all words with the dependency types from their dependents. Figure 4 shows the dependency subtree for the same example in Figure 3 , from which we collect three dependency rules: \"had\" \u2190 nsubj dobj, \"problems\" \u2190 det nn advmod, \"at\" \u2190 dep.",
"cite_spans": [
{
"start": 151,
"end": 177,
"text": "(de Marneffe et al., 2006)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [
{
"start": 633,
"end": 641,
"text": "Figure 4",
"ref_id": "FIGREF2"
},
{
"start": 695,
"end": 703,
"text": "Figure 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Fully embedded argument:",
"sec_num": null
},
{
"text": "Note that unlike the constituent parse features which are guaranteed to be accurate (as they are extracted from the gold parses of the corpus), the dependency parses occasionally contain errors. As with the constituent parse features, each dependency rule is represented as three binary features to check whether it appears in Arg1, Arg2, and both arguments. Lexical Features. Marcu and Echihabi (2002) demonstrated that word pairs extracted from the respective text spans are a good signal of the discourse relation between arguments. Thus we also consider word pairs as a feature class. We stemmed and collected all word pairs from Arg1 and Arg2, i.e., all (w i , w j ) where w i is a word from Arg1 and w j a word from Arg2. Unlike their study, we limit the collection of word pair statistics to occurrences only in the PDTB corpus.",
"cite_spans": [
{
"start": 377,
"end": 402,
"text": "Marcu and Echihabi (2002)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Fully embedded argument:",
"sec_num": null
},
{
"text": "For the collection of production rules, dependency rules, and word pairs, we used a frequency cutoff of 5 to remove infrequent features. From the implicit relation dataset of the training sections (i.e., Sec. 2 -21), we extracted 11,113 production rules, 5,031 dependency rules, and 105,783 word pairs in total. We applied mutual information (MI) to these three classes of features separately, resulting in three ranked lists. A feature f has 11 MI values with all 11 types (for example, M I(f, Cause) and M I(f, Restatement)), and we used the MI with the highest value for a feature to select features. In our experiments, the top features from the lists are used in the training and test phases.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Selection",
"sec_num": "4.1"
},
{
"text": "We experimented with a maximum entropy classifier from the OpenNLP MaxEnt package using various combinations of features to assess their efficacy. We used PDTB Sections 2 -21 as our training set and Section 23 as the test set, and only used the implicit discourse relations.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "5"
},
{
"text": "In the PDTB, about 2.2% of the implicit relations are annotated with two types, as shown in Example 7 in Section 6. During training, a relation that is annotated with two types is considered as two training instances, each with one of the types. During testing, such a relation is considered one test instance, and if the classifier assigns either of the two types, we consider it as correct. Thus, the test accuracy is calculated as the number of correctly classified test instances divided by the total number of test instances.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "5"
},
{
"text": "In our work, we use the majority class as the baseline, where all instances are classified as Cause. This yields an accuracy of 26.1% on the test set. A random baseline yields an even lower accuracy of 9.1% on the test set.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "5"
},
{
"text": "To check the efficacy of the different feature classes, we trained individual classifiers on all features within a single feature class (Rows 1 to 4 in Table 3 ) as well as a single classifier trained with all features from all feature classes (Row 5). Among the four individual feature classes, production rules and word pairs yield significantly better performance over the baseline with p < 0.01 and p < 0.05 respectively, while context features perform slightly better than the baseline. Interestingly, we noted that the performance with all dependency rules is slightly lower than the baseline (Row 2), and applying all feature classes does not yield the highest accuracy (Row 5), which we suspected were due to noise. To confirm this, we employed MI to select the top 100 production rules and dependency rules, and the top 500 word pairs (as word pairs are more sparse). We then repeated the same set of experiments, as shown in Table 4 (Row 4 of this table is repeated from Table 3 for consistency). With only the top features, production rules, dependency rules, and word pairs all gave significant improvement over the baseline with p < 0.01. When we used all feature classes, as in the last row, we obtained the highest accuracy of 40.2%. Table 4 also validates the pattern of predictiveness of the feature classes: production rules contribute the most to the performance individually, followed by word pairs, dependency rules, and finally, context features. A natural question to ask is whether any of these feature classes can be omitted to achieve the same level of performance as the combined classifier. To answer this question, we conducted a final set of experiments, in which we gradually added in feature classes in the or-der of their predictiveness (i.e., production rules word pairs dependency rules context features), with results shown in Table 5 . These results confirm that each additional feature class indeed contributes a marginal performance improvement, (although it is not significant) and that all feature classes are needed for optimal performance. Table 5 : Accuracy with feature classes gradually added in the order of their predictiveness.",
"cite_spans": [],
"ref_spans": [
{
"start": 152,
"end": 159,
"text": "Table 3",
"ref_id": "TABREF4"
},
{
"start": 935,
"end": 989,
"text": "Table 4 (Row 4 of this table is repeated from Table 3",
"ref_id": "TABREF4"
},
{
"start": 1250,
"end": 1257,
"text": "Table 4",
"ref_id": "TABREF6"
},
{
"start": 1864,
"end": 1871,
"text": "Table 5",
"ref_id": null
},
{
"start": 2084,
"end": 2091,
"text": "Table 5",
"ref_id": null
}
],
"eq_spans": [],
"section": "Results and Analysis",
"sec_num": "5.1"
},
{
"text": "Note that Row 3 of Table 3 corresponds to Marcu and Echihabi (2002) 's system which applies only word pair features. The difference is that they used a Na\u00efve Bayes classifier while we used a maximum entropy classifier. As we did not implement their Na\u00efve Bayes classifier, we compare their method's performance using the result from Table 3 Row 3 with ours from Table 5 Row 4, which shows that our system significantly (p < 0.01) outperforms theirs. Table 6 : Recall, precision, F 1 , and counts for 11 Level 2 relation types. \"-\" indicates 0.00. Table 6 shows the recall, precision, and F 1 measure for the 11 individual Level 2 relation types in the final experiment set up (Row 4 from Table 5). A point worth noting is that the classifier labels no instances of the Synchrony, Pragmatic Cause, Concession, and Alternative relation types. The reason is that the percentages for these four types are so small that the classifier is highly skewed towards the other types. From the distribution shown in Table 1 , there are just 4.76% training data for these four types, but 95.24% for the remaining seven types. In fact, only 30 test instances are labeled with these four types, as shown in the last column of Table 6 . As Cause is the most pre-dominant type in the training data, the classifier tends to label uncertain relations as Cause, thus giving Cause high recall but low precision. We see that the F measures correlate well with the training data frequency, thus we hypothesize that accuracy may improve if more training data for low frequency relations can be provided.",
"cite_spans": [
{
"start": 42,
"end": 67,
"text": "Marcu and Echihabi (2002)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [
{
"start": 19,
"end": 26,
"text": "Table 3",
"ref_id": "TABREF4"
},
{
"start": 333,
"end": 369,
"text": "Table 3 Row 3 with ours from Table 5",
"ref_id": "TABREF4"
},
{
"start": 450,
"end": 457,
"text": "Table 6",
"ref_id": null
},
{
"start": 547,
"end": 554,
"text": "Table 6",
"ref_id": null
},
{
"start": 1003,
"end": 1010,
"text": "Table 1",
"ref_id": "TABREF1"
},
{
"start": 1210,
"end": 1217,
"text": "Table 6",
"ref_id": null
}
],
"eq_spans": [],
"section": "Results and Analysis",
"sec_num": "5.1"
},
{
"text": "Our work differs from that of (Pitler et al., 2009 ) in that our system performs classification at the more fine-grained Level 2 types, instead of the coarse-grained Level 1 classes. Their system applies a Na\u00efve Bayes classifier whereas our system uses a maximum entropy classifier, and the sets of features used are also different. In addition, the data set of (Pitler et al., 2009) includes EntRel and AltLex, which are relations in which an implicit connective cannot be inserted between adjacent sentences, whereas ours excludes EntRel and AltLex.",
"cite_spans": [
{
"start": 30,
"end": 50,
"text": "(Pitler et al., 2009",
"ref_id": "BIBREF13"
},
{
"start": 362,
"end": 383,
"text": "(Pitler et al., 2009)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Results and Analysis",
"sec_num": "5.1"
},
{
"text": "6 Discussion: Why are implicit discourse relations difficult to recognize?",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results and Analysis",
"sec_num": "5.1"
},
{
"text": "In the above experiments, we have shown that by using the four feature classes, we are able to increase the classification accuracy from 26.1% of the majority baseline to 40.2%. Although we feel a 14.1 absolute percentage improvement is a solid result, an accuracy of 40% does not allow downstream NLP applications to trust the output of such a classification system. To understand the difficulties of the task more deeply, we analyzed individual training and validation data pairs, from which we were able to generalize four challenges to automated implicit discourse relation recognition. We hope that this discussion may motivate future work on implicit discourse relation recognition.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results and Analysis",
"sec_num": "5.1"
},
{
"text": "Ambiguity. There is ambiguity among the relations. For example, we notice that a lot of Contrast relations are mistakenly classified as Conjunction. When we analyzed these relations, we observed that Contrast and Conjunction in the PDTB annotation are very similar to each other in terms of words, syntax, and semantics, as Examples 5 and 6 show. In both examples, the same antonymous verb pair is used (fell and rose), different subjects are mentioned in Arg1 and Arg2 (net and revenue in the first example, and net and sales in the second), and these subjects are all compared to like items from the previous year. Moreover, the implicit discourse connective given by the annotators is while in both cases, which is an ambiguous connective as shown in (Miltsakaki et al., 2005 (Conjunction -wsj 1926) Relation ambiguity may be ameliorated if an instance is analyzed in context. However, according to the PDTB annotation guidelines, if the annotators could not disambiguate between two relation types, or if they felt both equally reflect their understanding of the relation between the arguments, they could annotate two types to the relation. In the whole PDTB corpus, about 5.4% of the explicit relations and 2.2% of the implicit relations are annotated with two relation types. Example 7 is such a case where the implicit connective meanwhile may be interpreted as expressing a Conjunction or Contrast relation. Inference. Sometimes inference and a knowledge base are required to resolve the relation type. In Example 8, to understand that Arg2 is a restatement of Arg1, we need a semantic mechanism to show that either the semantics of Arg1 infers that of Arg2 or the other way around. In the below example, I had calls all night long infers I was woken up every hour semantically, as shown in: receive call(I) \u2227 duration(all night) \u21d2 woken up(I) \u2227 duration(every hour). In fact, most relation types can be represented using formal semantics (PDTB-Group, 2007) , as shown in Table 7 , where |Arg1| and |Arg2| represent the semantics extracted from Arg1 and Arg2, respectively. This kind of formal semantic reasoning requires a robust knowledge base, which is still beyond our current technology. -Group, 2007) .",
"cite_spans": [
{
"start": 754,
"end": 778,
"text": "(Miltsakaki et al., 2005",
"ref_id": "BIBREF9"
},
{
"start": 779,
"end": 802,
"text": "(Conjunction -wsj 1926)",
"ref_id": null
},
{
"start": 1948,
"end": 1966,
"text": "(PDTB-Group, 2007)",
"ref_id": "BIBREF11"
},
{
"start": 2202,
"end": 2215,
"text": "-Group, 2007)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [
{
"start": 1981,
"end": 1988,
"text": "Table 7",
"ref_id": "TABREF11"
}
],
"eq_spans": [],
"section": "Results and Analysis",
"sec_num": "5.1"
},
{
"text": "Context. PDTB annotators adopted the Minimality Principle in argument selection, according to which they only included in the argument the minimal span of text that is sufficient for the interpretation of the relation. While the context is not necessary to interpret the relation, it is usually necessary to understand the meaning of the arguments. Without an analysis of the context, Arg1 and Arg2 may seem unconnected, as the following example shows, where the meaning of Arg1 is mostly derived from its previous context (i.e., West German ... technical reactions).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results and Analysis",
"sec_num": "5.1"
},
{
"text": "(9) Prev. Context: West German Economics Minister Helmut Haussmann said, \"In my view, the stock market will stabilize relatively quickly. There may be one or other psychological or technical reactions, Arg1: but they aren't based on fundamentals. Arg2: [in short] The economy of West Germany and the EC European Community is highly stable.\" (Conjunction -wsj 2210)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results and Analysis",
"sec_num": "5.1"
},
{
"text": "Sometimes the range of the context may easily extend to the whole text, which would require a system to possess a robust context modeling mechanism. In Example 10, in order to realize the causal relation between Arg2 and Arg1, we possibly need to read the whole article and understand what was happening: the machinist union was having a strike and the strike prevented most of its union members from working. (10) Arg1: And at the company's Wichita, Kan., plant, about 2,400 of the 11,700 machinists still are working, Boeing said. Arg2: [because] Under Kansas right-to-work laws, contracts cannot require workers to be union members. (Cause -wsj 2208)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results and Analysis",
"sec_num": "5.1"
},
{
"text": "World Knowledge. Sometimes even context modeling is not enough. We may also need world knowledge to understand the arguments and hence to interpret the relation. In the following example, from the previous sentence of Arg1, it is reported that \"the Senate voted to send a delegation of congressional staffers to Poland to assist its legislature\", and this delegation is viewed as a \"gift\" in Arg1. It is suggested in Arg2 that the Poles might view the delegation as a \"Trojan Horse\". Here we need world knowledge to understand that \"Trojan Horse\" is usually applied as a metaphor for a person or thing that appears innocent but has harmful intent, and hence understand that Arg2 poses a contrasting view of the delegation as Arg1 does.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results and Analysis",
"sec_num": "5.1"
},
{
"text": "(11) Arg1: Senator Pete Domenici calls this effort \"the first gift of democracy\". Arg2:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results and Analysis",
"sec_num": "5.1"
},
{
"text": "[but] The Poles might do better to view it as a Trojan Horse. (Contrast -wsj 2237)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results and Analysis",
"sec_num": "5.1"
},
{
"text": "These four classes of difficulties -ambiguity between relations, inference, contextual modeling, and world knowledge -show that implicit discourse relation classification needs deeper semantic representations, more robust system design, and access to more external knowledge. These obstacles may not be restricted to recognizing implicit relations, but are also applicable to other related discourse-centric tasks.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results and Analysis",
"sec_num": "5.1"
},
{
"text": "We implemented an implicit discourse relation classifier and showed initial results on the recently released Penn Discourse Treebank. The features we used include the modeling of the context of relations, features extracted from constituent parse trees and dependency parse trees, and word pair features. Our classifier achieves an accuracy of 40.2%, a 14.1% absolute improvement over the baseline. We also conducted a data analysis and discussed four challenges that need to be addressed in future to overcome the difficulties of implicit relation classification in the PDTB.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "7"
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Generating typed dependency parses from phrase structure parses",
"authors": [
{
"first": "Marie-Catherine",
"middle": [],
"last": "De Marneffe",
"suffix": ""
},
{
"first": "Bill",
"middle": [],
"last": "Maccartney",
"suffix": ""
},
{
"first": "Christopher",
"middle": [
"D"
],
"last": "Manning",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of the Fifth International Conference on Language Resources and Evaluation (LREC 2006)",
"volume": "",
"issue": "",
"pages": "449--454",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Marie-Catherine de Marneffe, Bill MacCartney, and Christopher D. Manning. 2006. Generating typed dependency parses from phrase structure parses. In Proceedings of the Fifth International Confer- ence on Language Resources and Evaluation (LREC 2006), pages 449-454.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Literature and cognition",
"authors": [
{
"first": "R",
"middle": [],
"last": "Jerry",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Hobbs",
"suffix": ""
}
],
"year": 1990,
"venue": "CSLI Lecture Notes Number 21",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jerry R. Hobbs. 1990. Literature and cognition. In CSLI Lecture Notes Number 21. CSLI Publications.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Semantic role labeling of NomBank: A maximum entropy approach",
"authors": [
{
"first": "Ping",
"middle": [],
"last": "Zheng",
"suffix": ""
},
{
"first": "Hwee Tou",
"middle": [],
"last": "Jiang",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Ng",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "138--145",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Zheng Ping Jiang and Hwee Tou Ng. 2006. Semantic role labeling of NomBank: A maximum entropy ap- proach. In Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing (EMNLP 2006), pages 138-145, Sydney, Australia.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "The classification of coherence relations and their linguistic markers: An exploration of two languages",
"authors": [
{
"first": "Alistair",
"middle": [],
"last": "Knott",
"suffix": ""
},
{
"first": "Ted",
"middle": [],
"last": "Sanders",
"suffix": ""
}
],
"year": 1998,
"venue": "Journal of Pragmatics",
"volume": "30",
"issue": "2",
"pages": "135--175",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Alistair Knott and Ted Sanders. 1998. The classifica- tion of coherence relations and their linguistic mark- ers: An exploration of two languages. Journal of Pragmatics, 30(2):135-175.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Temporal interpretation, discourse relations and commonsense entailment",
"authors": [
{
"first": "Alex",
"middle": [],
"last": "Lascarides",
"suffix": ""
},
{
"first": "Nicholas",
"middle": [],
"last": "Asher",
"suffix": ""
}
],
"year": 1993,
"venue": "Linguistics and Philosophy",
"volume": "16",
"issue": "5",
"pages": "437--493",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Alex Lascarides and Nicholas Asher. 1993. Temporal interpretation, discourse relations and commonsense entailment. Linguistics and Philosophy, 16(5):437- 493.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Complexity of dependencies in discourse: Are dependencies in discourse more complex than in syntax?",
"authors": [
{
"first": "Alan",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Rashmi",
"middle": [],
"last": "Prasad",
"suffix": ""
},
{
"first": "Aravind",
"middle": [],
"last": "Joshi",
"suffix": ""
},
{
"first": "Nikhil",
"middle": [],
"last": "Dinesh",
"suffix": ""
},
{
"first": "Bonnie",
"middle": [],
"last": "Webber",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of the 5th International Workshop on Treebanks and Linguistic Theories",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Alan Lee, Rashmi Prasad, Aravind Joshi, Nikhil Di- nesh, and Bonnie Webber. 2006. Complexity of dependencies in discourse: Are dependencies in dis- course more complex than in syntax? In Proceed- ings of the 5th International Workshop on Treebanks and Linguistic Theories, Prague, Czech Republic, December.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Rhetorical Structure Theory: Toward a functional theory of text organization",
"authors": [
{
"first": "C",
"middle": [],
"last": "William",
"suffix": ""
},
{
"first": "Sandra",
"middle": [
"A"
],
"last": "Mann",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Thompson",
"suffix": ""
}
],
"year": 1988,
"venue": "Text",
"volume": "8",
"issue": "3",
"pages": "243--281",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "William C. Mann and Sandra A. Thompson. 1988. Rhetorical Structure Theory: Toward a functional theory of text organization. Text, 8(3):243-281.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "An unsupervised approach to recognizing discourse relations",
"authors": [
{
"first": "Daniel",
"middle": [],
"last": "Marcu",
"suffix": ""
},
{
"first": "Abdessamad",
"middle": [],
"last": "Echihabi",
"suffix": ""
}
],
"year": 2002,
"venue": "Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics (ACL 2002)",
"volume": "",
"issue": "",
"pages": "368--375",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Daniel Marcu and Abdessamad Echihabi. 2002. An unsupervised approach to recognizing discourse re- lations. In Proceedings of the 40th Annual Meet- ing of the Association for Computational Linguistics (ACL 2002), pages 368-375, Morristown, NJ, USA.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "A surface-based approach to identifying discourse markers and elementary textual units in unrestricted texts",
"authors": [
{
"first": "Daniel",
"middle": [],
"last": "Marcu",
"suffix": ""
}
],
"year": 1998,
"venue": "Proceedings of the COLING-ACL 1998 Workshop on Discourse Relations and Discourse Markers",
"volume": "",
"issue": "",
"pages": "1--7",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Daniel Marcu. 1998. A surface-based approach to identifying discourse markers and elementary tex- tual units in unrestricted texts. In Proceedings of the COLING-ACL 1998 Workshop on Discourse Rela- tions and Discourse Markers, pages 1-7, Montreal, Canada, August.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Experiments on sense annotations and sense disambiguation of discourse connectives",
"authors": [
{
"first": "Eleni",
"middle": [],
"last": "Miltsakaki",
"suffix": ""
},
{
"first": "Nikhil",
"middle": [],
"last": "Dinesh",
"suffix": ""
},
{
"first": "Rashmi",
"middle": [],
"last": "Prasad",
"suffix": ""
},
{
"first": "Aravind",
"middle": [],
"last": "Joshi",
"suffix": ""
},
{
"first": "Bonnie",
"middle": [],
"last": "Webber",
"suffix": ""
}
],
"year": 2005,
"venue": "Proceedings of the Fourth Workshop on Treebanks and Linguistic Theories (TLT2005)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Eleni Miltsakaki, Nikhil Dinesh, Rashmi Prasad, Ar- avind Joshi, and Bonnie Webber. 2005. Experi- ments on sense annotations and sense disambigua- tion of discourse connectives. In Proceedings of the Fourth Workshop on Treebanks and Linguistic The- ories (TLT2005), Barcelona, Spain, December.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "A study on convolution kernels for shallow semantic parsing",
"authors": [
{
"first": "Alessandro",
"middle": [],
"last": "Moschitti",
"suffix": ""
}
],
"year": 2004,
"venue": "Proceedings of the 42nd Annual Meeting of the Association for Computational Linguistics (ACL 2004)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Alessandro Moschitti. 2004. A study on convolu- tion kernels for shallow semantic parsing. In Pro- ceedings of the 42nd Annual Meeting of the Asso- ciation for Computational Linguistics (ACL 2004), Barcelona, Spain.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "The Penn Discourse Treebank 2.0 Annotation Manual. The PDTB Research Group",
"authors": [
{
"first": "",
"middle": [],
"last": "Pdtb-Group",
"suffix": ""
}
],
"year": 2007,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "PDTB-Group, 2007. The Penn Discourse Treebank 2.0 Annotation Manual. The PDTB Research Group, December.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Easily identifiable discourse relations",
"authors": [
{
"first": "Emily",
"middle": [],
"last": "Pitler",
"suffix": ""
},
{
"first": "Mridhula",
"middle": [],
"last": "Raghupathy",
"suffix": ""
},
{
"first": "Hena",
"middle": [],
"last": "Mehta",
"suffix": ""
},
{
"first": "Ani",
"middle": [],
"last": "Nenkova",
"suffix": ""
},
{
"first": "Alan",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Aravind",
"middle": [],
"last": "Joshi",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of the 22nd International Conference on Computational Linguistics",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Emily Pitler, Mridhula Raghupathy, Hena Mehta, Ani Nenkova, Alan Lee, and Aravind Joshi. 2008. Eas- ily identifiable discourse relations. In Proceedings of the 22nd International Conference on Compu- tational Linguistics (COLING 2008), Manchester, UK, August.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Automatic sense prediction for implicit discourse relations in text",
"authors": [
{
"first": "Emily",
"middle": [],
"last": "Pitler",
"suffix": ""
},
{
"first": "Annie",
"middle": [],
"last": "Louis",
"suffix": ""
},
{
"first": "Ani",
"middle": [],
"last": "Nenkova",
"suffix": ""
}
],
"year": 2009,
"venue": "Proceedings of the Joint Conference of the 47th Annual Meeting of the Association for Computational Linguistics and the 4th International Joint Conference on Natural Language Processing of the Asian Federation of Natural Language Processing",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Emily Pitler, Annie Louis, and Ani Nenkova. 2009. Automatic sense prediction for implicit discourse relations in text. To appear in Proceedings of the Joint Conference of the 47th Annual Meeting of the Association for Computational Linguistics and the 4th International Joint Conference on Natural Lan- guage Processing of the Asian Federation of Natural Language Processing (ACL-IJCNLP 2009).",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "The Penn Discourse Treebank 2.0",
"authors": [
{
"first": "Rashmi",
"middle": [],
"last": "Prasad",
"suffix": ""
},
{
"first": "Nikhil",
"middle": [],
"last": "Dinesh",
"suffix": ""
},
{
"first": "Alan",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Eleni",
"middle": [],
"last": "Miltsakaki",
"suffix": ""
},
{
"first": "Livio",
"middle": [],
"last": "Robaldo",
"suffix": ""
},
{
"first": "Aravind",
"middle": [],
"last": "Joshi",
"suffix": ""
},
{
"first": "Bonnie",
"middle": [],
"last": "Webber",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of the 6th International Conference on Language Resources and Evaluation",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Rashmi Prasad, Nikhil Dinesh, Alan Lee, Eleni Milt- sakaki, Livio Robaldo, Aravind Joshi, and Bon- nie Webber. 2008. The Penn Discourse Treebank 2.0. In Proceedings of the 6th International Confer- ence on Language Resources and Evaluation (LREC 2008).",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Using phrasal patterns to identify discourse relations",
"authors": [
{
"first": "Manami",
"middle": [],
"last": "Saito",
"suffix": ""
},
{
"first": "Kazuhide",
"middle": [],
"last": "Yamamoto",
"suffix": ""
},
{
"first": "Satoshi",
"middle": [],
"last": "Sekine",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of the Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics (HLT-NAACL 2006)",
"volume": "",
"issue": "",
"pages": "133--136",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Manami Saito, Kazuhide Yamamoto, and Satoshi Sekine. 2006. Using phrasal patterns to iden- tify discourse relations. In Proceedings of the Hu- man Language Technology Conference of the North American Chapter of the Association for Computa- tional Linguistics (HLT-NAACL 2006), pages 133- 136, New York, USA, June.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "D-LTAG: Extending lexicalized TAG to discourse",
"authors": [
{
"first": "Bonnie",
"middle": [],
"last": "Webber",
"suffix": ""
}
],
"year": 2004,
"venue": "Cognitive Science",
"volume": "28",
"issue": "5",
"pages": "751--779",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Bonnie Webber. 2004. D-LTAG: Extending lex- icalized TAG to discourse. Cognitive Science, 28(5):751-779, September.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Classification of discourse coherence relations: An exploratory study using multiple knowledge sources",
"authors": [
{
"first": "Ben",
"middle": [],
"last": "Wellner",
"suffix": ""
},
{
"first": "James",
"middle": [],
"last": "Pustejovsky",
"suffix": ""
},
{
"first": "Catherine",
"middle": [],
"last": "Havasi",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of the 7th SIGdial Workshop on Discourse and Dialogue",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ben Wellner, James Pustejovsky, Catherine Havasi, Anna Rumshisky, and Roser Sauri. 2006. Clas- sification of discourse coherence relations: An ex- ploratory study using multiple knowledge sources. In Proceedings of the 7th SIGdial Workshop on Dis- course and Dialogue, Sydney, Australia, July.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Representing discourse coherence: a corpus-based analysis",
"authors": [
{
"first": "Florian",
"middle": [],
"last": "Wolf",
"suffix": ""
},
{
"first": "Edward",
"middle": [],
"last": "Gibson",
"suffix": ""
}
],
"year": 2005,
"venue": "Proceedings of the 20th International Conference on Computational Linguistics (COLING 2004)",
"volume": "",
"issue": "",
"pages": "134--140",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Florian Wolf and Edward Gibson. 2005. Representing discourse coherence: a corpus-based analysis. In Proceedings of the 20th International Conference on Computational Linguistics (COLING 2004), pages 134-140, Morristown, NJ, USA.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"num": null,
"text": "Two types of discourse dependency structures. Top: fully embedded argument, bottom: shared argument.",
"type_str": "figure",
"uris": null
},
"FIGREF1": {
"num": null,
"text": "[subsequently] That debt would be paid off as the assets are sold (Asynchronous -wsj 2200) (4) Arg1: It would have been too late to think about on Friday. Arg2: [so] We had to think about it ahead of time(a) constituent parse in Arg2 of Example 3, (b) constituent parse in Arg1 of Example 4. A gold standard subtree for Arg1 of an implicit discourse relation from wsj 2224.",
"type_str": "figure",
"uris": null
},
"FIGREF2": {
"num": null,
"text": "A dependency subtree for Arg1 of an implicit discourse relation from wsj 2224.",
"type_str": "figure",
"uris": null
},
"FIGREF3": {
"num": null,
"text": "Arg1: Sales surged 40% to 250.17 billion yen from 178.61 billion. Arg2: [meanwhile] Net income rose 11% to 29.62 billion yen from 26.68 billion. (Conjunction; Contrast -wsj 2242)",
"type_str": "figure",
"uris": null
},
"FIGREF4": {
"num": null,
"text": "Arg1: \"I had calls all night long from the States,\" he said. Arg2: \"[in fact] I was woken up every hour -1:30, 2:30, 3:30, 4:30.\"(Restatement -wsj 2205)",
"type_str": "figure",
"uris": null
},
"TABREF1": {
"type_str": "table",
"content": "<table/>",
"html": null,
"text": "Distribution of Level 2 relation types of implicit relations from the training sections (Sec.",
"num": null
},
"TABREF2": {
"type_str": "table",
"content": "<table/>",
"html": null,
"text": "Six contextual features derived from two discourse dependency patterns. curr is the relation we want to classify.Constituent ParseFeatures. Research work from other NLP areas, such as semantic role labeling, has shown that features derived from syntactic trees are useful in semantic understanding. Such features include syntactic paths",
"num": null
},
"TABREF4": {
"type_str": "table",
"content": "<table><tr><td>: Classification accuracy with all features</td></tr><tr><td>from each feature class. Rows 1 to 4: individual</td></tr><tr><td>feature class; Row 5: all feature classes.</td></tr></table>",
"html": null,
"text": "",
"num": null
},
"TABREF6": {
"type_str": "table",
"content": "<table><tr><td>:</td><td>Classification accuracy with top</td></tr><tr><td colspan=\"2\">rules/word pairs for each feature class. Rows 1</td></tr><tr><td colspan=\"2\">to 4: individual feature class; Row 5: all feature</td></tr><tr><td>classes.</td><td/></tr></table>",
"html": null,
"text": "",
"num": null
},
"TABREF11": {
"type_str": "table",
"content": "<table><tr><td>: Some examples of relation types with</td></tr><tr><td>their semantic representations, as taken from</td></tr><tr><td>(PDTB</td></tr></table>",
"html": null,
"text": "",
"num": null
}
}
}
} |