File size: 74,452 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 |
{
"paper_id": "I08-1021",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:40:39.719233Z"
},
"title": "Modeling Context in Scenario Template Creation",
"authors": [
{
"first": "Long",
"middle": [],
"last": "Qiu",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "National University of Singapore Singapore",
"location": {
"postCode": "117590"
}
},
"email": "qiul@comp.nus.edu.sg"
},
{
"first": "Min-Yen",
"middle": [],
"last": "Kan",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "National University of Singapore Singapore",
"location": {
"postCode": "117590"
}
},
"email": ""
},
{
"first": "Tat-Seng",
"middle": [],
"last": "Chua",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "National University of Singapore Singapore",
"location": {
"postCode": "117590"
}
},
"email": "chuats@comp.nus.edu.sg"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "We describe a graph-based approach to Scenario Template Creation, which is the task of creating a representation of multiple related events, such as reports of different hurricane incidents. We argue that context is valuable to identify important, semantically similar text spans from which template slots could be generalized. To leverage context, we represent the input as a set of graphs where predicate-argument tuples are vertices and their contextual relations are edges. A context-sensitive clustering framework is then applied to obtain meaningful tuple clusters by examining their intrinsic and extrinsic similarities. The clustering framework uses Expectation Maximization to guide the clustering process. Experiments show that: 1) our approach generates high quality clusters, and 2) information extracted from the clusters is adequate to build high coverage templates.",
"pdf_parse": {
"paper_id": "I08-1021",
"_pdf_hash": "",
"abstract": [
{
"text": "We describe a graph-based approach to Scenario Template Creation, which is the task of creating a representation of multiple related events, such as reports of different hurricane incidents. We argue that context is valuable to identify important, semantically similar text spans from which template slots could be generalized. To leverage context, we represent the input as a set of graphs where predicate-argument tuples are vertices and their contextual relations are edges. A context-sensitive clustering framework is then applied to obtain meaningful tuple clusters by examining their intrinsic and extrinsic similarities. The clustering framework uses Expectation Maximization to guide the clustering process. Experiments show that: 1) our approach generates high quality clusters, and 2) information extracted from the clusters is adequate to build high coverage templates.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Scenario template creation (STC) is the problem of generating a common semantic representation from a set of input articles. For example, given multiple newswire articles on different hurricane incidents, an STC algorithm creates a template that may include slots for the storm's name, current location, direction of travel and magnitude. Slots in such a scenario template are often to be filled by salient entities in the scenario instance (e.g., \"Hurricane Charley\" or \"the coast area\") but some can also be filled by prominent clauses, verbs or adjectives that describe these salient entities. Here, we use the term salient aspect (SA) to refer to any of such slot fillers that people would regard as important to describe a particular scenario. Figure 1 shows such a manuallybuilt scenario template in which details about important actions, actors, time and locations are coded as slots. STC is an important task that has tangible benefits for many downstream applications. In the Message Understanding Conference (MUC), manuallygenerated STs were provided to guide Information Extraction (IE). An ST can also be viewed as regularizing a set of similar articles as a set of attribute/value tuples, enabling multi-document summarization from filled templates.",
"cite_spans": [],
"ref_spans": [
{
"start": 749,
"end": 757,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Despite these benefits, STC has not received much attention by the community. We believe this is because it is considered a difficult task that requires deep NL understanding of the source articles. A problem in applications requiring semantic similarity is that the same word in different contexts may have different senses and play different roles. Conversely, different words in similar contexts may play similar roles. This problem makes approaches that rely on word similarity alone inadequate.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We propose a new approach to STC that incorporates the use of contextual information to address this challenge. Unlike previous approaches that concentrate on the intrinsic similarity of candidate slot fillers, our approach explicitly models contextual evidence. And unlike approaches to word sense disambiguation (WSD) and other semantic analyses that use neighboring or syntactically related words as contextual evidence, we define contexts by semantic relatedness which extends beyond sentence boundaries. Figure 2 illustrates a case in point with two excerpts from severe storm reports. Here, although the intrinsic similarity of the main verbs \"hit\" and \"land\" is low, their contextual similarity is high as both are followed by clauses sharing similar subjects (hurricanes) and the same verbs. Our approach encodes such contextual information as graphs, mapping the STC problem into a general graph overlay problem that is solvable by a variant of Expectation Maximization (EM).",
"cite_spans": [],
"ref_spans": [
{
"start": 509,
"end": 517,
"text": "Figure 2",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Our work also contributes resources for STC research. Until now, few scenario templates have been publicly available (as part of MUC), rendering any potential evaluation of automated STC statistically insignificant. As part of our study, we have compiled a set of input articles with annotations that we are making available to the research community. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "A natural way to automate the process of STC is to cluster similar text spans in the input article set. SAs then emerge through clustering; if a cluster of text spans is large enough, the aspects contained in it will be considered as SAs. Subsequently, these SAs will be generalized into one or more slots in the template, depending on the definition of the text span. Assuming scenarios are mainly defined by actions, the focus should be on finding appropriate clusters for text spans each of which represents an action. Most of the related work (although they may not directly address STC) shares this assumption and performs",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "Charley landed further south on Florida's Gulf coast than predicted, ... The hurricane ... has weakened and is moving over South Carolina.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "At least 21 others are missing after the storm hit on Wednesday. .... But Tokage had weakened by the time it passed over Japan's capital, Tokyo, where it left little damage before moving out to sea. Curved lines indicate similar contexts, providing evidence that \"land\" and \"hit\" from two articles are semantically similar.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "action clustering accordingly. While the target application varies, most systems that need to group text spans by similarity measures are verb-centric. In addition to the verb, many systems expand their representation by including named entity tags (Collier, 1998; Yangarber et al., 2000; Sudo et al., 2003; Filatova et al., 2006) , as well as restricting matches (using constraints on subtrees (Sudo et al., 2003; Filatova et al., 2006) , predicate argument structures (Collier, 1998; Riloff and Schmelzenbach, 1998; Yangarber et al., 2000; Harabagiu and Maiorano, 2002) or semantic roles).",
"cite_spans": [
{
"start": 249,
"end": 264,
"text": "(Collier, 1998;",
"ref_id": "BIBREF0"
},
{
"start": 265,
"end": 288,
"text": "Yangarber et al., 2000;",
"ref_id": "BIBREF16"
},
{
"start": 289,
"end": 307,
"text": "Sudo et al., 2003;",
"ref_id": "BIBREF14"
},
{
"start": 308,
"end": 330,
"text": "Filatova et al., 2006)",
"ref_id": "BIBREF2"
},
{
"start": 395,
"end": 414,
"text": "(Sudo et al., 2003;",
"ref_id": "BIBREF14"
},
{
"start": 415,
"end": 437,
"text": "Filatova et al., 2006)",
"ref_id": "BIBREF2"
},
{
"start": 470,
"end": 485,
"text": "(Collier, 1998;",
"ref_id": "BIBREF0"
},
{
"start": 486,
"end": 517,
"text": "Riloff and Schmelzenbach, 1998;",
"ref_id": "BIBREF13"
},
{
"start": 518,
"end": 541,
"text": "Yangarber et al., 2000;",
"ref_id": "BIBREF16"
},
{
"start": 542,
"end": 571,
"text": "Harabagiu and Maiorano, 2002)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "Given these representations, systems then cluster similar text spans. To our knowledge, all current systems use a binary notion of similarity, in which pairs of spans are either similar or not. How they determine similarity is tightly coupled with their text span representation. One criterion used is pattern overlap: for example, (Collier, 1998; Harabagiu and Lacatusu, 2005 ) judge text spans to be similar if they have similar verbs and share the same verb arguments. Working with tree structures, Sudo et al. and Filatova et al. instead require shared subtrees.",
"cite_spans": [
{
"start": 332,
"end": 347,
"text": "(Collier, 1998;",
"ref_id": "BIBREF0"
},
{
"start": 348,
"end": 376,
"text": "Harabagiu and Lacatusu, 2005",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "Calculating text span similarity ultimately boils down to calculating word phrase similarity. Approaches such as Yangarber's or Riloff and Schmelzenbach's do not employ a thesaurus and thus are easier to implement, but can suffer from over-or under-generalization. In certain cases, either the same actor is involved in different actions or different verbs realize the same action. Other systems (Collier, 1998; Sudo et al., 2003) do employ lexical similarity but threshold it to obtain binary judgments. Systems then rank clusters by cluster size and correlation with the relevant article set and equate top clusters as output scenario slots.",
"cite_spans": [
{
"start": 396,
"end": 411,
"text": "(Collier, 1998;",
"ref_id": "BIBREF0"
},
{
"start": 412,
"end": 430,
"text": "Sudo et al., 2003)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "Automating STC requires handling a larger degree of variations than most previous work we have surveyed. Note that the actors involved in actions in a scenario generally differ from event to event, which makes most related work on text span similarity calculation unsuitable. Also, action participants are not limited to named entities, so our approach needs to process all NPs. As both actions and actors may be realized using different words, a similarity thesaurus is necessary. Our approach to STC uses a thesaurus based on corpus statistics (Lin, 1998) for real-valued similarity calculation. In contrast to previous approaches, we do not threshold word similarity results; we retain their fractional values and incorporate these values holistically. Finally, as the same action can be realized in different constructions, the semantic (not just syntactic) roles of verb arguments must be considered, lest agent and patient roles be confused. For these reasons, we use a semantic role labeler (Pradhan et al., 2004) to provide and delimit the text spans that contain the semantic arguments of a predicate. We term the obtained text spans as predicate argument tuples (tuples) throughout the paper. The semantic role labeler reportedly achieves an F 1 measure equal to 68.7% on identificationclassification of predicates and core arguments on a newswire text corpus (LDC, 2002) . Within the confines of our study, we find it is able to capture most of the tuples of interest.",
"cite_spans": [
{
"start": 546,
"end": 557,
"text": "(Lin, 1998)",
"ref_id": "BIBREF9"
},
{
"start": 998,
"end": 1020,
"text": "(Pradhan et al., 2004)",
"ref_id": "BIBREF10"
},
{
"start": 1370,
"end": 1381,
"text": "(LDC, 2002)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Context-Sensitive Clustering (CSC)",
"sec_num": "3"
},
{
"text": "Our approach explicitly captures contextual evidence. We define a tuple's contexts as other tuples in the same article segment where no topic shift occurs. This definition refines the n-surrounding word constraint commonly used in spelling correction (for example, (Hirst and Budanitsky, 2005) ), Word Sense Disambiguation ( (Preiss, 2001) , (Lee and Ng, 2002) , for instance), etc. while still ensures the relatedness between a tuple and its contexts. Specifically, a tuple is contextually related to other tuples by two quantifiable contextual relations: argument-similarity and position-similarity. For our experiments, we use the leads of newswire articles as they normally summarize the news. We also assume a lead qualifies as a single article segment, thus making all of its tuples as potential contexts to each other. Figure 3: Being similar contexts, \"weakened\" and \"moving\" provide contextual evidence that \"land\" and \"hit\" are similar.",
"cite_spans": [
{
"start": 265,
"end": 293,
"text": "(Hirst and Budanitsky, 2005)",
"ref_id": "BIBREF5"
},
{
"start": 325,
"end": 339,
"text": "(Preiss, 2001)",
"ref_id": "BIBREF11"
},
{
"start": 342,
"end": 360,
"text": "(Lee and Ng, 2002)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Context-Sensitive Clustering (CSC)",
"sec_num": "3"
},
{
"text": "from A 2 from A 1 weakened(storm) v 2 1 hit(storm) v 2 2 moving(storm) v 2 3 weakened(hurricane) v 1 1 landed(hurricane) v 1 2 moving(hurricane) v 1 3 e 2 1,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Context-Sensitive Clustering (CSC)",
"sec_num": "3"
},
{
"text": "First, we split the input article leads into sentences and perform semantic role labeling immediately afterwards. Our system could potentially benefit from additional pre-processing such as co-reference resolution. Currently these pre-processing steps have not been properly integrated with the rest of the system, and thus we have not yet measured their impact.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Context-Sensitive Clustering (CSC)",
"sec_num": "3"
},
{
"text": "We then transform each lead A i into a graph",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Context-Sensitive Clustering (CSC)",
"sec_num": "3"
},
{
"text": "G i = {V i , E i }. As shown in Figure 3, vertices V i = {v i j }(j = 1, ..., N )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Context-Sensitive Clustering (CSC)",
"sec_num": "3"
},
{
"text": "are the N predicate argument tuples extracted from the ith article, and directed edges",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Context-Sensitive Clustering (CSC)",
"sec_num": "3"
},
{
"text": "E i = {e i m,n = (v i m , v i n )} reflect contextual relations between tuple v i m and v i n .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Context-Sensitive Clustering (CSC)",
"sec_num": "3"
},
{
"text": "Edges only connect tuples from the same article, i.e., within each graph G i . We differentiate between two types of edges. One is argument-similarity, where the two tuples have semantically similar arguments. This models tuple cohesiveness, where the edge weight is determined by the similarity score of the most similar inter-tuple argument pair. The other is positionsimilarity, represented as the offset of the ending tuple with respect to the other, measured in sentences. This edge type is directional to account for simple causality.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Context-Sensitive Clustering (CSC)",
"sec_num": "3"
},
{
"text": "Given this set of graphs, the clustering task is to find an optimal alignment of all graphs (i.e., superimposing the set of article graphs to maximize vertex overlap, constrained by the edges). We adapt Expectation Maximization (Dempster et al., 1977) to find an optimal clustering. This process assigns tuples to suitable clusters where they are semantically similar and share similar contexts with other tuples. Algorithm 1 outlines this alignment process. During initialization, tuples whose pairwise similarity higher than a threshold \u03c4 are merged to form highly cohesive seed clusters. To compute a continuous similarity Sim(t a , t b ) of tuples t a and t b , we use the similarity measure described in (Qiu et al., 2006) , which linearly combines similarities between the semantic roles shared by the two tuples. Some other tuples are related to these seed clusters by argument-similarity. These related tuples are temporarily put into a special \"other\" cluster. The cluster membership of these related tuples, together with those currently in the seed clusters, are to be further adjusted. The \"other\" cluster is so called because a tuple will end up being assigned to it if it is not found to be similar to any other tuple. Tuples that are neither similar to nor contextually related by argument-similarity to another tuple are termed singletons and excluded from being clustered.",
"cite_spans": [
{
"start": 228,
"end": 251,
"text": "(Dempster et al., 1977)",
"ref_id": "BIBREF1"
},
{
"start": 709,
"end": 727,
"text": "(Qiu et al., 2006)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Context-Sensitive Clustering (CSC)",
"sec_num": "3"
},
{
"text": "Algorithm 1 Graph Alignment(G) /*G is a set of graph {G i }*/ T \u2190 all tuples in G C \u2190",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Context-Sensitive Clustering (CSC)",
"sec_num": "3"
},
{
"text": "We then iteratively (re-)estimate clusters of tuples across the set of article graphs G. In the E-step of the EM algorithm, all contextual relations between each pair of clusters are collected as two set of edges.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Context-Sensitive Clustering (CSC)",
"sec_num": "3"
},
{
"text": "Here we assume argument-similarity and positionsimilarity are independent and thus we differentiate them in the computation. Accordingly, there are two sets: edges as and edges ps . For simplicity, we assume independent normal distributions for the strength of each set (inter-tuple argument similarity for edges as and sentence distance for edges ps ). The edge strength distribution parameters for both sets between each pair of clusters are re-estimated based on current edges in edges as and edges ps .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Context-Sensitive Clustering (CSC)",
"sec_num": "3"
},
{
"text": "In the M-step, we examine each tuple's fitness for belonging to its cluster and relocate some tuples to new clusters to maximize the likelihood given the latest estimated edge strength distributions. In the following equations, we denote the proposition that predicate argument tuple t a belongs to cluster c m as t a \u2208c m ; a typical tuple (the centroid) of the cluster c m as t cm ; and the cluster of t a as c ta . The objective function to maximize is:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Context-Sensitive Clustering (CSC)",
"sec_num": "3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "Obj(G) = X ta\u2208G log(P (ta\u2208ct a )),",
"eq_num": "(1)"
}
],
"section": "Context-Sensitive Clustering (CSC)",
"sec_num": "3"
},
{
"text": "where P (ta\u2208cm) = 2Ps(ta\u2208cm) Pc(ta\u2208cm) Ps(ta\u2208cm) + Pc(ta\u2208cm) .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Context-Sensitive Clustering (CSC)",
"sec_num": "3"
},
{
"text": "Equation 2 takes the harmonic mean of two factors: a contextual factor P c and and a semantic factor P s :",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Context-Sensitive Clustering (CSC)",
"sec_num": "3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "Pc(ta\u2208cm) = max{P (edges(ta, t b )| t b :edges(ta,t b ) =null edges(cm, ct b ))},",
"eq_num": "(3)"
}
],
"section": "Context-Sensitive Clustering (CSC)",
"sec_num": "3"
},
{
"text": "Ps(ta\u2208cm) = ( sim def ault , cm = c other , Sim(ta, tc m ), otherwise.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Context-Sensitive Clustering (CSC)",
"sec_num": "3"
},
{
"text": "Here the contextual factor P c models how likely t a belongs to c m according to the contextual information, i.e., the conditional probability of the contextual relations between c m and c t b given the contextual relations between t a and one particular context t b , which maximizes this probability. According to Bayes' theorem, it is computed as shown in Equation 3. In practice, we multiply two conditional probabilities: P (edge as (t a , t b )|edges as (c m , c t b )) and P (edge ps (t a , t b )|edges ps (c m , c t b )), assuming independence between edges as and edges ps . We assume there are still singleton tuples that are not semantically similar to another tuple and should belong to the special \"other\" cluster. Given that they are dissimilar to each other, we set sim def ault to a small nonzero value in Equation 4 to prevent the \"other\" cluster from expelling them based on their low semantic similarity. Tuples' cluster memberships are recalculated, and the parameters describing the contextual relations between clusters are reestimated. New EM iterations are performed as long as one or more tuple relocations occur. Once the EM halts, clusters of equivalent tuples are formed. Among these clusters, some correspond to salient actions that, together with their actors, are all SAs to be generalized into template slots. Cluster size is a good indicator of salience, and each large cluster (excluding the \"other\" cluster) can be viewed as containing instances of a salient action.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Context-Sensitive Clustering (CSC)",
"sec_num": "3"
},
{
"text": "Formulating the clustering process as a variant of iterative EM is well-motivated as we consider the similarity scores as noisy and having missing observations. Calculating semantic similarity is at best inaccurate. Thus it is difficult to cluster tuples correctly based only on their semantic similarity. Also to check whether a tuple shares contexts with a cluster of tuples, the cluster has to be relatively clean. An iterative EM as we have proposed naturally improve the cleanness of these tuple clusters gradually as new similarity information comes to light.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Context-Sensitive Clustering (CSC)",
"sec_num": "3"
},
{
"text": "For STC, we argue that it is crucial to cluster tuples with high recall so that an SA's various surface forms can be captured and the size of clusters can serve as a salience indicator. Meanwhile, precision should not be sacrificed, as more noise will hamper the downstream generalization process which outputs template slots. We conduct experiments designed to answer two relevant research questions: 1) Cluster Quality: Whether using contexts (in CSC) produces better clustering results than ignoring it (in the K-means baseline); and 2) Template Coverage: Whether slots generalized from CSC clusters cover human-defined templates.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation",
"sec_num": "4"
},
{
"text": "A straightforward evaluation of a STC system would compare its output against manually-prepared gold standard templates, such as those found in MUC.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data Set and Baseline",
"sec_num": "4.1"
},
{
"text": "Unfortunately, such scenario templates are severely limited and do not provide enough instances for a proper evaluation. To overcome this problem, we have prepared a balanced news corpus, where we have manually selected articles covering 15 scenarios. Each scenario is represented by a total of 45 to 50 articles which describe 10 different events.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data Set and Baseline",
"sec_num": "4.1"
},
{
"text": "Our baseline is a standard K-means clusterer. Its input is identical to that of CSC -the tuples extracted from relevant news articles and are not excluded from being clustered by CSC in the initialization stage (refer to Section 3) -and employs the same tuple similarity measure (Qiu et al., 2006) . The differentiating factor between CSC and K-means is the use of contextual evidence. A standard K-means clusterer requires a k to be specified. For each scenario, we set its k as the number of clusters generated by CSC for direct comparison.",
"cite_spans": [
{
"start": 279,
"end": 297,
"text": "(Qiu et al., 2006)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Data Set and Baseline",
"sec_num": "4.1"
},
{
"text": "We fix the test set for each scenario as ten randomly selected news articles, each reporting a different instance of the scenario; the development set (which also serves as the training set for determining the EM initialization threshold \u03c4 and sim def ault in Equation 4) is a set of ten articles from the \"Air-linerCrash\" scenario, which are excluded from the test set. Both systems analyze the first 15 sentences of each article, and sentences generate 2 to 3 predicate argument tuples on average, resulting in a total of 10 \u00d7 15 \u00d7 (2 to 3) = 300 to 450 tuples for each scenario.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data Set and Baseline",
"sec_num": "4.1"
},
{
"text": "This experiment compares the clustering results of CSC and K-means. We use the standard clustering metrics of purity and inverse purity (Hotho et al., 2003) . The first author manually constructed the gold standard clusters for each scenario using a GUI before conducting any experiments. A special cluster, corresponding to the \"other\" cluster in the CSC clusters, was created to hold the singleton tuples for each scenario. Table 1 shows this under the column \"#Gold Standard Clusters\".",
"cite_spans": [
{
"start": 136,
"end": 156,
"text": "(Hotho et al., 2003)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [
{
"start": 426,
"end": 433,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Cluster Quality",
"sec_num": "4.2"
},
{
"text": "Using the manual clusters as the gold standard, we obtain the purity (P) and inverse purity (IP) scores of CSC and K-means on each scenario. In Table 1 , we see that CSC outperforms K-means on 10 of 15 scenarios for both P and IP. For the remaining 5 scenarios, where CSC and K-means have comparable P scores, the IP scores of CSC are all significantly higher than that of K-means. This suggests clusters tend to be split apart more in K-means than in CSC when they have similar purity. One thing worth mentioning here is that the \"other\" cluster normally is relatively large for each scenario, and thus may skew the results. To remove this effect, we excluded tuples belonging to the CSC \"other\" cluster from the K-means input, generating one fewer cluster. Running the evaluation again, the resulting P-IP scores again show that CSC outperforms the baseline Kmeans. We only report the results for all tuples in our paper for simplicity. Table 1 : CSC outperforms K-means with respect to the purity (P) and inverse purity (IP) scores.",
"cite_spans": [],
"ref_spans": [
{
"start": 144,
"end": 151,
"text": "Table 1",
"ref_id": null
},
{
"start": 939,
"end": 946,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Cluster Quality",
"sec_num": "4.2"
},
{
"text": "A close inspection of the results reveals some problematic cases. One issue worth mentioning is that for certain actions both CSC and K-means produce split clusters. In the CSC case, we traced this problem back to the thesaurus, where predicates for one action seem to belong to two or more totally dissimilar semantic categories. The corresponding tuples are thus assigned to different clusters as their low semantic similarity forces the tuples to remain separate, despite the shared contexts trying to join them. One example is \"blast (off)\" and \"lift (off)\" in the \"Launch Event\" scenario. The thesaurus shows the two verbs are dissimilar and the corresponding tuples end up being in two split clusters. This can not be solved easily without an improved thesaurus. We are considering adding a prior to model the op-timal size for clusters, which may help to compact such cases.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Cluster Quality",
"sec_num": "4.2"
},
{
"text": "We also assess how well the resulting, CSCgenerated tuple clusters serve in creating good scenario template slots. We start from the top largest clusters from each scenario, and decompose each of them into six sets: the predicates, agents, patients, predicate modifiers, agent modifiers and patient modifiers. For each of the first three sets for each cluster, we create a generalized term to represent it using an extended version of a generalization algorithm (Tseng et al., 2006) . These terms are deemed output slots, and are put into the template with their agent-predicate-patient relations preserved. The size of the template may increase when more clusters are generalized, as new slots may result.",
"cite_spans": [
{
"start": 462,
"end": 482,
"text": "(Tseng et al., 2006)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Template Coverage",
"sec_num": "4.3"
},
{
"text": "We manually compare the slots that are output from the system with those defined in existing scenario templates in MUC. The results here are only indicative and not conclusive, as there are only two MUC7 templates available for comparison: Aviation Disaster and Launch Event. shows an excerpt of the automatically generated template \"AviationDisaster\" (\"Airliner-Crash\" in our corpus) where the semantic roles in the top two biggest clusters have been generalized. Their modifiers are quite semantically diverse, as shown in Table 2 . Thus, generalization (probably after a categorization operation) remains as a challenging problem.",
"cite_spans": [],
"ref_spans": [
{
"start": 525,
"end": 532,
"text": "Table 2",
"ref_id": "TABREF2"
}
],
"eq_spans": [],
"section": "Template Coverage",
"sec_num": "4.3"
},
{
"text": "Nonetheless, the information contained in these semantic roles and their modifiers covers human-semantic role modifier head samples agent:aircraft A, U.N., The, Swiss, Canadianbuilt, AN, China, military, Iranian, Air, refueling, US, ... action:crash Siberia, mountain, rain, Tuesday, flight, Sharjah, flames, Sunday, board, Saturday, 225 , Rockaway, approach, United, mountain, hillside patient:people all, 255, 71 Figure 5 : MUC-7 template coverage: asterisks marking all the slots that could be automatically generated.",
"cite_spans": [
{
"start": 183,
"end": 186,
"text": "AN,",
"ref_id": null
},
{
"start": 187,
"end": 193,
"text": "China,",
"ref_id": null
},
{
"start": 194,
"end": 203,
"text": "military,",
"ref_id": null
},
{
"start": 204,
"end": 212,
"text": "Iranian,",
"ref_id": null
},
{
"start": 213,
"end": 217,
"text": "Air,",
"ref_id": null
},
{
"start": 218,
"end": 228,
"text": "refueling,",
"ref_id": null
},
{
"start": 229,
"end": 232,
"text": "US,",
"ref_id": null
},
{
"start": 233,
"end": 258,
"text": "... action:crash Siberia,",
"ref_id": null
},
{
"start": 259,
"end": 268,
"text": "mountain,",
"ref_id": null
},
{
"start": 269,
"end": 274,
"text": "rain,",
"ref_id": null
},
{
"start": 275,
"end": 283,
"text": "Tuesday,",
"ref_id": null
},
{
"start": 284,
"end": 291,
"text": "flight,",
"ref_id": null
},
{
"start": 292,
"end": 300,
"text": "Sharjah,",
"ref_id": null
},
{
"start": 301,
"end": 308,
"text": "flames,",
"ref_id": null
},
{
"start": 309,
"end": 316,
"text": "Sunday,",
"ref_id": null
},
{
"start": 317,
"end": 323,
"text": "board,",
"ref_id": null
},
{
"start": 324,
"end": 333,
"text": "Saturday,",
"ref_id": null
},
{
"start": 334,
"end": 337,
"text": "225",
"ref_id": null
}
],
"ref_spans": [
{
"start": 415,
"end": 423,
"text": "Figure 5",
"ref_id": null
}
],
"eq_spans": [],
"section": "Template Coverage",
"sec_num": "4.3"
},
{
"text": "defined scenario templates quite well. The two MUC7 templates are shown as a list of slots in Figure 5 , where horizontal lines delimit slots about different semantic roles, and asterisks mark all the slots that could be automatically generated by our system once it has an improved generalizer. We can see substantial amount of overlap, indicating that a STC system powered by CSC is able to capture scenarios' important facts.",
"cite_spans": [],
"ref_spans": [
{
"start": 94,
"end": 102,
"text": "Figure 5",
"ref_id": null
}
],
"eq_spans": [],
"section": "Template Coverage",
"sec_num": "4.3"
},
{
"text": "We have introduced a new context-sensitive approach to the scenario template creation (STC) problem. Our method leverages deep NL processing, using semantic role labeler's structured semantic tuples as input. Despite the use of deeper semantics, we believe that intrinsic semantic similarity by itself is not sufficient for clustering. We have shown this through examples and argue that an approach that considers contextual similarity is necessary. A key aspect of our work is the incorporation of such contextual information. Our approach uses a notion of context that combines two aspects: positional similarity (when two tuples are adjacent in the text), and argument similarity (when they have similar arguments). The set of relevant articles are represented as graphs where contextual evidence is encoded. By mapping our problem into a graphical formalism, we cast the STC clustering problem as one of multiple graph alignment. Such a graph alignment is solved by an adaptation of EM, which handles contexts and real-valued similarity by treating both as noisy and potentially unreliable observations.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "5"
},
{
"text": "While scenario template creation (STC) is a difficult problem, its evaluation is arguably more difficult due to the dearth of suitable resources. We have compiled and released a corpus of over 700 newswire articles that describe different instances of 15 scenarios, as a suitable input dataset for further STC research. Using this dataset, we have evaluated and analyzed our context-sensitive approach. While our results are indicative, they show that considering contextual evidence improves performance.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "5"
}
],
"back_matter": [
{
"text": "The authors are grateful to Kathleen R. McKeown and Elena Filatova at Columbia University for their stimulating discussions and comments over different stages of the preparation of this paper.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgments",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Automatic Template Creation for Information Extraction",
"authors": [
{
"first": "Robin",
"middle": [],
"last": "Collier",
"suffix": ""
}
],
"year": 1998,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Robin Collier. 1998. Automatic Template Creation for Information Extraction. Ph.D. thesis, University of Sheffield, UK.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Maximum likelihood from incomplete data via the EM algorithm",
"authors": [
{
"first": "A",
"middle": [
"P"
],
"last": "Dempster",
"suffix": ""
},
{
"first": "N",
"middle": [
"M"
],
"last": "Laird",
"suffix": ""
},
{
"first": "D",
"middle": [
"B"
],
"last": "Rubin",
"suffix": ""
}
],
"year": 1977,
"venue": "JRSSB",
"volume": "39",
"issue": "",
"pages": "1--38",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "A.P. Dempster, N.M. Laird, and D.B. Rubin. 1977. Max- imum likelihood from incomplete data via the EM al- gorithm. JRSSB, 39:1-38.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Automatic creation of domain templates",
"authors": [
{
"first": "Elena",
"middle": [],
"last": "Filatova",
"suffix": ""
},
{
"first": "Vasileios",
"middle": [],
"last": "Hatzivassiloglou",
"suffix": ""
},
{
"first": "Kathleen",
"middle": [],
"last": "Mckeown",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of the COLING/ACL '06",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Elena Filatova, Vasileios Hatzivassiloglou, and Kathleen McKeown. 2006. Automatic creation of domain tem- plates. In Proceedings of the COLING/ACL '06.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Topic themes for multi-document summarization",
"authors": [
{
"first": "M",
"middle": [],
"last": "Sanda",
"suffix": ""
},
{
"first": "V",
"middle": [
"Finley"
],
"last": "Harabagiu",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Lacatusu",
"suffix": ""
}
],
"year": 2005,
"venue": "Proceedings of SIGIR '05",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sanda M. Harabagiu and V. Finley Lacatusu. 2005. Topic themes for multi-document summarization. In Proceedings of SIGIR '05.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Multidocument summarization with GISTEXTER",
"authors": [
{
"first": "M",
"middle": [],
"last": "Sanda",
"suffix": ""
},
{
"first": "S",
"middle": [
"J"
],
"last": "Harabagiu",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Maiorano",
"suffix": ""
}
],
"year": 2002,
"venue": "Proceedings of LREC '02",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sanda M. Harabagiu and S. J. Maiorano. 2002. Multi- document summarization with GISTEXTER. In Pro- ceedings of LREC '02.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Correcting real-word spelling errors by restoring lexical cohesion",
"authors": [
{
"first": "Graeme",
"middle": [],
"last": "Hirst",
"suffix": ""
},
{
"first": "Alexander",
"middle": [],
"last": "Budanitsky",
"suffix": ""
}
],
"year": 2005,
"venue": "Natural Language Engineering",
"volume": "",
"issue": "1",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Graeme Hirst and Alexander Budanitsky. 2005. Cor- recting real-word spelling errors by restoring lexical cohesion. Natural Language Engineering, 11(1).",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "WordNet improves text document clustering",
"authors": [
{
"first": "Andreas",
"middle": [],
"last": "Hotho",
"suffix": ""
},
{
"first": "Steffen",
"middle": [],
"last": "Staab",
"suffix": ""
},
{
"first": "Gerd",
"middle": [],
"last": "Stumme",
"suffix": ""
}
],
"year": 2003,
"venue": "Proceedings of the SIGIR 2003 Semantic Web Workshop",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Andreas Hotho, Steffen Staab, and Gerd Stumme. 2003. WordNet improves text document clustering. In Pro- ceedings of the SIGIR 2003 Semantic Web Workshop.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "The aquaint corpus of english news text, catalog no",
"authors": [
{
"first": "",
"middle": [],
"last": "Ldc",
"suffix": ""
}
],
"year": 2002,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "LDC. 2002. The aquaint corpus of english news text, catalog no. LDC2002t31.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "An empirical evaluation of knowledge sources and learning algorithms for word sense disambiguation",
"authors": [
{
"first": "Yoong Keok",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Hwee Tou",
"middle": [],
"last": "Ng",
"suffix": ""
}
],
"year": 2002,
"venue": "Proceedings of EMNLP '02",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yoong Keok Lee and Hwee Tou Ng. 2002. An empiri- cal evaluation of knowledge sources and learning algo- rithms for word sense disambiguation. In Proceedings of EMNLP '02.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Automatic retrieval and clustering of similar words",
"authors": [
{
"first": "Dekang",
"middle": [],
"last": "Lin",
"suffix": ""
}
],
"year": 1998,
"venue": "Proceedings of COLING/ACL '98",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Dekang Lin. 1998. Automatic retrieval and clustering of similar words. In Proceedings of COLING/ACL '98.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Shallow semantic parsing using support vector machines",
"authors": [
{
"first": "Sameer",
"middle": [],
"last": "Pradhan",
"suffix": ""
},
{
"first": "Wayne",
"middle": [],
"last": "Ward",
"suffix": ""
},
{
"first": "Kadri",
"middle": [],
"last": "Hacioglu",
"suffix": ""
},
{
"first": "James",
"middle": [],
"last": "Martin",
"suffix": ""
},
{
"first": "Dan",
"middle": [],
"last": "Jurafsky",
"suffix": ""
}
],
"year": 2004,
"venue": "Proceedings of HLT/NAACL '04",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sameer Pradhan, Wayne Ward, Kadri Hacioglu, James Martin, and Dan Jurafsky. 2004. Shallow semantic parsing using support vector machines. In Proceed- ings of HLT/NAACL '04.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Local versus global context for wsd of nouns",
"authors": [
{
"first": "Judita",
"middle": [],
"last": "Preiss",
"suffix": ""
}
],
"year": 2001,
"venue": "Proceedings of CLUK4",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Judita Preiss. 2001. Local versus global context for wsd of nouns. In Proceedings of CLUK4.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Paraphrase recognition via dissimilarity significance classification",
"authors": [
{
"first": "Long",
"middle": [],
"last": "Qiu",
"suffix": ""
},
{
"first": "Min-Yen",
"middle": [],
"last": "Kan",
"suffix": ""
},
{
"first": "Tat-Seng",
"middle": [],
"last": "Chua",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of EMNLP '06",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Long Qiu, Min-Yen Kan, and Tat-Seng Chua. 2006. Paraphrase recognition via dissimilarity significance classification. In Proceedings of EMNLP '06.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "An empirical approach to conceptual case frame acquisition",
"authors": [
{
"first": "Ellen",
"middle": [],
"last": "Riloff",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Schmelzenbach",
"suffix": ""
}
],
"year": 1998,
"venue": "Proceedings of WVLC '98",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ellen Riloff and M. Schmelzenbach. 1998. An empiri- cal approach to conceptual case frame acquisition. In Proceedings of WVLC '98.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "An improved extraction pattern representation model for automatic IE pattern acquisition",
"authors": [
{
"first": "Kiyoshi",
"middle": [],
"last": "Sudo",
"suffix": ""
},
{
"first": "Satoshi",
"middle": [],
"last": "Sekine",
"suffix": ""
},
{
"first": "Ralph",
"middle": [],
"last": "Grishman",
"suffix": ""
}
],
"year": 2003,
"venue": "Proceedings of ACL '03",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kiyoshi Sudo, Satoshi Sekine, and Ralph Grishman. 2003. An improved extraction pattern representation model for automatic IE pattern acquisition. In Pro- ceedings of ACL '03.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Toward generic title generation for clustered documents",
"authors": [
{
"first": "",
"middle": [],
"last": "Yuen-Hsien",
"suffix": ""
},
{
"first": "Chi-Jen",
"middle": [],
"last": "Tseng",
"suffix": ""
},
{
"first": "Hsiu-Han",
"middle": [],
"last": "Lin",
"suffix": ""
},
{
"first": "Yu-I",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Lin",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of AIRS '06",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yuen-Hsien Tseng, Chi-Jen Lin, Hsiu-Han Chen, and Yu- I Lin. 2006. Toward generic title generation for clus- tered documents. In Proceedings of AIRS '06.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Unsupervised discovery of scenario-level patterns for information extraction",
"authors": [
{
"first": "Roman",
"middle": [],
"last": "Yangarber",
"suffix": ""
},
{
"first": "Ralph",
"middle": [],
"last": "Grishman",
"suffix": ""
},
{
"first": "Pasi",
"middle": [],
"last": "Tapanainen",
"suffix": ""
},
{
"first": "Silja",
"middle": [],
"last": "Huttunen",
"suffix": ""
}
],
"year": 2000,
"venue": "Proceedings of ANLP '00",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Roman Yangarber, Ralph Grishman, Pasi Tapanainen, and Silja Huttunen. 2000. Unsupervised discovery of scenario-level patterns for information extraction. In Proceedings of ANLP '00.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"uris": null,
"type_str": "figure",
"text": "An example scenario template (filled).",
"num": null
},
"FIGREF1": {
"uris": null,
"type_str": "figure",
"text": "Contextual evidence of similarity.",
"num": null
},
"FIGREF2": {
"uris": null,
"type_str": "figure",
"text": "highly cohesive tuples clusters other \u2190 remaining tuples semantically connected with C C[C.length] \u2190 other repeat /*E step*/ for each i such that i < C.length do for each j such that j < C.length do if i == j then continue; re-estimate parameters[C[i], C[j]] /*distribution parameters of edges between two clusters*/ tupleReassigned = f alse /*reset*/ /*M step*/ for each i such that i < T.length do aBestLikelihood = T [i].likelihood; /*likelihood of being in its current cluster*/ for each tuple tcontxt that contextually related with T [i] do for each cluster c cand , any candidate cluster that contextually related with tcontxt.cluster do P (T [i] \u2208 c cand ) = comb(Ps, Pc) likelihood = log(P (T [i] \u2208 c cand )) if likelihood > aBestLikelihood then aBestLikelihood = likelihood T [i].cluster = c cand tupleReassigned = true until tupleReassigned == f alse /*alignment stable*/ return",
"num": null
},
"FIGREF3": {
"uris": null,
"type_str": "figure",
"text": "Figure 4: Automated scenario template of \"Avia-tionDisaster\".",
"num": null
},
"FIGREF4": {
"uris": null,
"type_str": "figure",
"text": "Figure 4 shows an excerpt of the automatically generated template \"AviationDisaster\" (\"Airliner-Crash\" in our corpus) where the semantic roles in the top two biggest clusters have been generalized. Their modifiers are quite semantically diverse, as shown in Table 2. Thus, generalization (probably after a categorization operation) remains as a challenging problem. Nonetheless, the information contained in these semantic roles and their modifiers covers human-",
"num": null
},
"TABREF0": {
"num": null,
"html": null,
"content": "<table><tr><td>Storm Name Charley Storm Action landed Location Florida's Gulf coast Time Friday at 1950GMT Speed 145 mph Victim Category 1 13 people Action died Victim Category 2 over one million Action affected</td></tr></table>",
"type_str": "table",
"text": ""
},
"TABREF2": {
"num": null,
"html": null,
"content": "<table><tr><td>AviationDisaster * AIRCRAFT * AIRLINE DEPARTURE POINT DEPARTURE DATE * AIRCRAFT TYPE * CRASH DATE * CRASH SITE CAUSE INFO * VICTIMS NUM</td><td>LaunchEvent * VEHICLE * VEHICLE TYPE * VEHICLE OWNER * PAYLOAD PAYLOAD TYPE PAYLOAD FUNC * PAYLOAD OWNER PAYLOAD ORIGIN * LAUNCH DATE * LAUNCH SITE MISSION TYPE MISSION FUNCTION MISSION STATUS</td></tr></table>",
"type_str": "table",
"text": "Sample automatically detected modifier heads of different semantic roles."
}
}
}
} |