File size: 83,398 Bytes
6fa4bc9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 | {
"paper_id": "D09-1047",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T16:40:05.756764Z"
},
"title": "Joint Learning of Preposition Senses and Semantic Roles of Prepositional Phrases",
"authors": [
{
"first": "Daniel",
"middle": [],
"last": "Dahlmeier",
"suffix": "",
"affiliation": {},
"email": "danielhe@comp.nus.edu.sg"
},
{
"first": "Hwee",
"middle": [
"Tou"
],
"last": "Ng",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Tanja",
"middle": [],
"last": "Schultz",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Karlsruhe",
"location": {}
},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "The sense of a preposition is related to the semantics of its dominating prepositional phrase. Knowing the sense of a preposition could help to correctly classify the semantic role of the dominating prepositional phrase and vice versa. In this paper, we propose a joint probabilistic model for word sense disambiguation of prepositions and semantic role labeling of prepositional phrases. Our experiments on the PropBank corpus show that jointly learning the word sense and the semantic role leads to an improvement over state-of-theart individual classifier models on the two tasks.",
"pdf_parse": {
"paper_id": "D09-1047",
"_pdf_hash": "",
"abstract": [
{
"text": "The sense of a preposition is related to the semantics of its dominating prepositional phrase. Knowing the sense of a preposition could help to correctly classify the semantic role of the dominating prepositional phrase and vice versa. In this paper, we propose a joint probabilistic model for word sense disambiguation of prepositions and semantic role labeling of prepositional phrases. Our experiments on the PropBank corpus show that jointly learning the word sense and the semantic role leads to an improvement over state-of-theart individual classifier models on the two tasks.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Word sense disambiguation (WSD) and semantic role labeling (SRL) are two key components in natural language processing to find a semantic representation for a sentence. Semantic role labeling is the task of determining the constituents of a sentence that represent semantic arguments with respect to a predicate and labeling each with a semantic role. Word sense disambiguation tries to determine the correct meaning of a word in a given context. Ambiguous words occur frequently in normal English text.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "One word class which is both frequent and highly ambiguous is preposition. The different senses of a preposition express different relations between the preposition complement and the rest of the sentence. Semantic roles and word senses offer two different inventories of \"meaning\" for prepositional phrases (PP): semantic roles distinguish between different verb complements while word senses intend to fully capture the preposition semantics at a more fine-grained level. In this paper, we use the semantic roles from the PropBank corpus and the preposition senses from the Preposition Project (TPP). Both corpora are explained in more detail in the following section. The relationship between the two inventories (PropBank semantic roles and TPP preposition senses) is not a simple one-to-one mapping, as we can see from the following examples:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "\u2022 She now lives with relatives [in sense1",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Alabama.] ARGM-LOC",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "\u2022 The envelope arrives [in sense1 the mail.] ARG4",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "\u2022 [In sense5 separate statements] ARGM-LOC the two sides said they want to have \"further discussions.\"",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In the first two examples, the sense of the preposition in is annotated as sense 1 (\"surrounded by or enclosed in\"), following the definitions of the TPP, but the semantic roles are different. In the first example the semantic role is a locative adjunctive argument (ARGM-LOC), while in the second example it is ARG4 which denotes the \"end point or destination\" of the arriving action 1 . In the first and third example, the semantic roles are the same, but the preposition senses are different, i.e., sense 1 and sense 5 (\"inclusion or involvement\"). Preposition senses and semantic roles provide two different views on the semantics of PPs. Knowing the semantic role of the PP could be helpful to successfully disambiguate the sense of the preposition. Likewise, the preposition sense could provide valuable information to classify the semantic role of the PP. This is especially so for the semantic roles ARGM-LOC and ARGM-TMP, where we expect a strong correlation with spatial and temporal preposition senses respectively.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In this paper, we propose a probabilistic model for joint inference on preposition senses and semantic roles. For each prepositional phrase that has been identified as an argument of the predicate, we jointly infer its semantic role and the sense of the preposition that is the lexical head of the prepositional phrase. That is, our model maximizes the joint probability of the semantic role and the preposition sense.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Previous research has shown the benefit of jointly learning semantic roles of multiple constituents (Toutanova et al., 2008; Koomen et al., 2005) . In contrast, our joint model makes predictions for a single constituent, but multiple tasks (WSD and SRL) .",
"cite_spans": [
{
"start": 100,
"end": 124,
"text": "(Toutanova et al., 2008;",
"ref_id": "BIBREF14"
},
{
"start": 125,
"end": 145,
"text": "Koomen et al., 2005)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Our experiments show that adding the SRL information leads to statistically significant improvements over an independent, state-of-the-art WSD classifier. For the SRL task, we show statistically significant improvements of our joint model over an independent, state-of-the-art SRL classifier for locative and temporal adjunctive arguments, even though the overall improvement over all semantic roles is small. To the best of our knowledge, no previous research has attempted to perform preposition WSD and SRL of prepositional phrases in a joint learning approach.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The remainder of this paper is structured as follows: First, we give an introduction to the WSD and SRL task. Then, in Section 3, we describe the individual and joint classifier models. The details of the data set used in our experiments are given in Section 4. In Section 5, we present experiments and results. Section 6 summarizes related work, before we conclude in the final section.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "This section gives an introduction to preposition sense disambiguation and semantic role labeling of prepositional phrases.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Task Description",
"sec_num": "2"
},
{
"text": "The task of word sense disambiguation is to find the correct meaning of a word, given its context. Most prior research on word sense disambiguation has focused on disambiguating the senses of nouns, verbs, and adjectives, but not on prepositions. Word sense disambiguation can be framed as a classification task. For each preposition, a classifier is trained on a corpus of training examples annotated with preposition senses, and tested on a set of unseen test examples.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preposition Sense Disambiguation",
"sec_num": "2.1"
},
{
"text": "To perform WSD for prepositions, it is necessary to first find a set of suitable sense classes.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preposition Sense Disambiguation",
"sec_num": "2.1"
},
{
"text": "We adopt the sense inventory from the Preposition Project (TPP) (Litkowski and Hargraves, 2005) that was also used in the SemEval 2007 preposition WSD task (Litkowski and Hargraves, 2007) . TPP is an attempt to create a comprehensive lexical database of English prepositions that is suitable for use in computational linguistics research. For each of the over 300 prepositions and phrasal prepositions, the database contains a set of sense definitions, which are based on the Oxford Dictionary of English. Every preposition has a set of fine-grained senses, which are grouped together into a smaller number of coarse-grained senses. In our experiments, we only focus on coarse-grained senses since better inter-annotator agreement can be achieved on coarse-grained senses, which also results in higher accuracy of the trained WSD classifier.",
"cite_spans": [
{
"start": 64,
"end": 95,
"text": "(Litkowski and Hargraves, 2005)",
"ref_id": "BIBREF6"
},
{
"start": 156,
"end": 187,
"text": "(Litkowski and Hargraves, 2007)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Preposition Sense Disambiguation",
"sec_num": "2.1"
},
{
"text": "The task of semantic role labeling in the context of PropBank is to label tree nodes with semantic roles in a syntactic parse tree.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Semantic Role Labeling",
"sec_num": "2.2"
},
{
"text": "The PropBank corpus adds a semantic layer to parse trees from the Wall Street Journal section of the Penn Treebank II corpus (Marcus et al., 1993) . There are two classes of semantic roles: core arguments and adjunctive arguments. Core arguments are verb sense specific, i.e., their meaning is defined relative to a specific verb sense. They are labeled with consecutive numbers ARG0, ARG1, etc. ARG0 usually denotes the AGENT and ARG1 the THEME of the event. Besides the core arguments, a verb can have a number of adjunctive arguments that express more general properties like time, location, or manner. They are labeled as ARGM plus a functional tag, e.g., LOC for locative or TMP for temporal modifiers. Prepositional phrases can appear as adjunctive arguments or core arguments.",
"cite_spans": [
{
"start": 125,
"end": 146,
"text": "(Marcus et al., 1993)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Semantic Role Labeling",
"sec_num": "2.2"
},
{
"text": "The standard approach to semantic role labeling is to divide the task into two sequential sub-tasks: identification and classification. During the identification phase, the system separates the nodes that fill some semantic roles from the rest. During the classification phase, the system assigns the exact semantic roles for all nodes that are identified as arguments. In this paper, we focus on the classification phase. That is, we assume that prepositional phrases that are semantic arguments have been identified correctly and concentrate on the task of determining the semantic role of prepositional phrases. The reason is that argument identification mostly relies on syntactic features, like the path from the constituent to the predicate (Pradhan et al., 2005) . Consider, for example, the phrase in the dark in the sentence: \"We are in the dark\", he said. The phrase is clearly not an argument to the verb say. But if we alter the syntactic structure of the sentence appropriately (while the sense of the preposition in remains unchanged), the same phrase suddenly becomes an adjunctive argument: In the dark, he said \"We are\". On the other hand, we can easily find examples, where in has a different sense, but the phrase always fills some semantic role:",
"cite_spans": [
{
"start": 747,
"end": 769,
"text": "(Pradhan et al., 2005)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Semantic Role Labeling",
"sec_num": "2.2"
},
{
"text": "\u2022 In a separate manner, he said . . .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Semantic Role Labeling",
"sec_num": "2.2"
},
{
"text": "\u2022 In 1998, he said . . .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Semantic Role Labeling",
"sec_num": "2.2"
},
{
"text": "\u2022 In Washington, he said . . . This illustrates that the preposition sense is independent of whether the PP is an argument or not. Thus, a joint learning model for argument identification and preposition sense is unlikely to perform better than the independent models.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Semantic Role Labeling",
"sec_num": "2.2"
},
{
"text": "This section describes the models for preposition sense disambiguation and semantic role labeling.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Models",
"sec_num": "3"
},
{
"text": "We compare three different models for each task: First, we implement an independent model that only uses task specific features from the literature. This serves as the baseline model. Second, we extend the baseline model by adding the most likely prediction of the other task as an additional feature. This is equivalent to a pipeline model of classifiers that feeds the prediction of one classification step into the next stage. Finally, we present a joint model to determine the preposition sense and semantic role that maximize the joint probability.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Models",
"sec_num": "3"
},
{
"text": "Our approach to building a preposition WSD classifier follows that of Lee and Ng (2002) , who evaluated a set of different knowledge sources and learning algorithms for WSD. However, in this paper we use maximum entropy models 2 (instead of support vector machines (SVM) reported in (Lee and Ng, 2002) ), because maximum entropy models output probability distributions, unlike SVM. This property is useful in the joint model, as we will see later. Maxent models have been successfully applied to various NLP tasks and achieve state-of-the-art performance. There are two training parameters that have to be adjusted for maxent models: the number of training iterations and the Gaussian smoothing parameter. We find optimal values for both parameters through 10-fold crossvalidation on the training set.",
"cite_spans": [
{
"start": 70,
"end": 87,
"text": "Lee and Ng (2002)",
"ref_id": "BIBREF5"
},
{
"start": 283,
"end": 301,
"text": "(Lee and Ng, 2002)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "WSD model",
"sec_num": "3.1"
},
{
"text": "For every preposition, a baseline maxent model is trained using a set of features reported in the state-of-the-art WSD system of Lee and Ng (2002) . These features encode three knowledge sources:",
"cite_spans": [
{
"start": 129,
"end": 146,
"text": "Lee and Ng (2002)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "WSD model",
"sec_num": "3.1"
},
{
"text": "\u2022 Part-of-speech (POS) of surrounding words",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "WSD model",
"sec_num": "3.1"
},
{
"text": "\u2022 Single words in the surrounding context",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "WSD model",
"sec_num": "3.1"
},
{
"text": "For part-of-speech features, we include the POS tags of surrounding tokens from the same sentence within a window of seven tokens around the target prepositions. All tokens (i.e., all words and punctuation symbols) are considered. We use the Penn Treebank II POS tag set.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "\u2022 Local collocations",
"sec_num": null
},
{
"text": "For the knowledge source single words in the surrounding context, we consider all words from the same sentence. The input sentence is tokenized and all tokens that do not contain at least one alphabetical character (such as punctuation symbols and numbers) and all words that appear on a stopword list are removed. The remaining words are converted to lower case and replaced by their morphological root form. Every unique morphological root word contributes one binary feature, indicating whether or not the word is present in the context. The position of a word in the sentence is ignored in this knowledge source.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "\u2022 Local collocations",
"sec_num": null
},
{
"text": "The third knowledge source, local collocations, encodes position-specific information of words within a small window around the target preposition. For this knowledge source, we consider unigrams, bigrams, and trigrams from a window of seven tokens. The position of the target preposition inside the n-gram is marked with a special character ' '. Words are converted to lower case, but no stemming or removal of stopwords is performed. If a token falls outside the sentence, it is replaced by the empty token symbol nil.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "\u2022 Local collocations",
"sec_num": null
},
{
"text": "During testing, the maxent model computes the conditional probability of the sense, given the feature representation of the surrounding context c. The classifier outputs the sense that receives the highest probability:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "\u2022 Local collocations",
"sec_num": null
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "s = argmax s P (s|\u03a8(c))",
"eq_num": "(1)"
}
],
"section": "\u2022 Local collocations",
"sec_num": null
},
{
"text": "where \u03a8(\u2022) is a feature map from the surrounding context to the feature representation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "\u2022 Local collocations",
"sec_num": null
},
{
"text": "To ensure that our model is competitive, we tested our system on the data set from the SemEval 2007 preposition WSD task (Litkowski and Hargraves, 2007) . Our baseline classifier achieved a coarse-grained accuracy of 70.7% (micro-average) on the official test set. This would have made our system the second best system in the competition, behind the MELB-YB system (Ye and Baldwin, 2007) .",
"cite_spans": [
{
"start": 121,
"end": 152,
"text": "(Litkowski and Hargraves, 2007)",
"ref_id": "BIBREF7"
},
{
"start": 366,
"end": 388,
"text": "(Ye and Baldwin, 2007)",
"ref_id": "BIBREF17"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "\u2022 Local collocations",
"sec_num": null
},
{
"text": "We also investigate the effect of the semantic role label by adding it as a feature to the baseline model. This pipeline model is inspired by the work of Dang and Palmer (2005) who investigated the role of SRL features in verb WSD. We add the semantic role of the prepositional phrase dominating the preposition as a feature to the WSD model. During training, the PropBank gold SRL label is used. During testing, we rely on the baseline SRL model (to be introduced in the next subsection) to predict the semantic role of the prepositional phrase. This is equivalent to first performing semantic role labeling and adding the output as a feature to the WSD classifier. In earlier experiments, we found that training on gold SRL labels gave better results than training on automatically predicted SRL labels (using crossvalidation). Note that our approach uses automatically assigned SRL labels during testing, while the system of Dang and Palmer (2005) only uses gold SRL labels.",
"cite_spans": [
{
"start": 154,
"end": 176,
"text": "Dang and Palmer (2005)",
"ref_id": "BIBREF1"
},
{
"start": 928,
"end": 950,
"text": "Dang and Palmer (2005)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "\u2022 Local collocations",
"sec_num": null
},
{
"text": "Our semantic role labeling classifier is also based on maxent models. It has been shown that maximum entropy models achieve state-of-the-art results on SRL (Xue and Palmer, 2004; Toutanova et al., 2008) . Again, we find optimal values for the training parameters through 10-fold crossvalidation on the training set.",
"cite_spans": [
{
"start": 156,
"end": 178,
"text": "(Xue and Palmer, 2004;",
"ref_id": "BIBREF15"
},
{
"start": 179,
"end": 202,
"text": "Toutanova et al., 2008)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "SRL model",
"sec_num": "3.2"
},
{
"text": "By treating SRL as a classification problem, the choice of appropriate features becomes a key issue. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "SRL model",
"sec_num": "3.2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "a = argmax a P (a|t, p, v) (2) = argmax a P (a|\u03a6(t, p, v))",
"eq_num": "(3)"
}
],
"section": "SRL model",
"sec_num": "3.2"
},
{
"text": "where \u03a6(\u2022, \u2022, \u2022) is a feature map to an appropriate feature representation. For our baseline SRL model, we adopt the features used in other state-of-the-art SRL systems, which include the seven baseline features from the original work of Gildea and Jurafsky (2002) , additional features taken from Pradhan et al. 2005, and feature combinations which are inspired by the system in Xue and Palmer (2004) . Table 1 lists the features we use for easy reference.",
"cite_spans": [
{
"start": 238,
"end": 264,
"text": "Gildea and Jurafsky (2002)",
"ref_id": "BIBREF2"
},
{
"start": 380,
"end": 401,
"text": "Xue and Palmer (2004)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [
{
"start": 404,
"end": 411,
"text": "Table 1",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "SRL model",
"sec_num": "3.2"
},
{
"text": "In the pipeline model, we investigate the usefulness of the preposition sense as a feature for SRL by adding the preposition lemma concatenated with the sense number (e.g., on 1) as a feature. During training, the gold annotated preposition sense is used. During testing, the sense is automatically tagged by the baseline WSD model. This is equivalent to first running the WSD classifier for all prepositions, and adding the output preposition sense as a feature to our baseline SRL system.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "SRL model",
"sec_num": "3.2"
},
{
"text": "The two previous models seek to maximize the probability of the semantic role and the preposition sense individually, thus ignoring possible dependencies between the two. Instead of maximizing the individual probabilities, we would like to maximize the joint probability of the semantic role and the preposition sense, given the parse tree, predicate, constituent node, and surrounding context.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Joint Inference Model",
"sec_num": "3.3"
},
{
"text": "(a, s) = argmax (a,s)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Joint Inference Model",
"sec_num": "3.3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "P (a, s|t, p, v, c)",
"eq_num": "(4)"
}
],
"section": "Joint Inference Model",
"sec_num": "3.3"
},
{
"text": "We assume that the probability of the semantic role is already determined by the syntactic parse tree t, the predicate p, and the constituent node v, and is conditionally independent of the remaining surrounding context c given t, p, and v. Likewise, we assume that the probability of the preposition sense is conditionally independent of the parse tree t, predicate p, and constituent v, given the surrounding context c and the semantic role a. This assumption allows us to factor the joint probability into an SRL and a WSD component:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Joint Inference Model",
"sec_num": "3.3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "(a, s) = argmax (a,s) P (a|t, p, v)\u00d7P (s|c, a)",
"eq_num": "(5)"
}
],
"section": "Joint Inference Model",
"sec_num": "3.3"
},
{
"text": "= argmax (a,s)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Joint Inference Model",
"sec_num": "3.3"
},
{
"text": "P (a|\u03a6(t, p, v))\u00d7P (s|\u03a8(c, a))(6)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Joint Inference Model",
"sec_num": "3.3"
},
{
"text": "We observe that the first component in our joint model corresponds to the baseline SRL model and the second component corresponds to the WSD pipeline model. Because our maxent models output a complete probability distribution, we can combine both components by multiplying the probabilities. Theoretically, the joint probability could be factored in the other way, by first computing the probability of the preposition sense and then conditioning the SRL model on the predicted preposition sense. However, in our early experiments, we found that this approach gave lower classification accuracy. During testing, the classifier seeks to find the tuple of semantic role and preposition sense that maximizes the joint probability. For every semantic role, the classifier computes its probability given the SRL features, and multiplies it by the probability of the most likely preposition sense, given the context and the semantic role. The tuple that receives the highest joint probability is the final output of the joint classifier. Test ARG0 28 15 13 ARG1 374 208 166 ARG2 649 352 297 ARG3 111 67 44 ARG4 177 91 86 ARGM-ADV 141 101 40 ARGM-CAU 31 23 ",
"cite_spans": [],
"ref_spans": [
{
"start": 1032,
"end": 1176,
"text": "Test ARG0 28 15 13 ARG1 374 208 166 ARG2 649 352 297 ARG3 111 67 44 ARG4 177 91 86 ARGM-ADV 141 101 40 ARGM-CAU 31 23",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Joint Inference Model",
"sec_num": "3.3"
},
{
"text": "The joint model uses the probability of a preposition sense, given the semantic role of the dominating prepositional phrase. To estimate this probability, we need a corpus which is annotated with both preposition senses and semantic roles. Unfortunately, PropBank is not annotated with preposition senses. Instead, we manually annotated the seven most frequent prepositions in four sections of the PropBank corpus with their senses from the TPP dictionary. According to Jurafsky and Martin (2008) , the most frequent English prepositions are: of, in, for, to, with, on and at (in order of frequency). Our counts on Sections 2 to 21 of PropBank revealed that these top 7 prepositions account for about 65% of all prepositional phrases that are labeled with semantic roles. The annotation proceeds in the following way. First, we automatically extract all sentences which have one of the prepositions as the lexical head of a prepositional phrase. The position of the preposition is marked in the sentence. By only considering prepositional phrases, we automatically exclude occurrences of the word to before infinitives and instances of particle usage of prepositions, such as phrasal verbs. The extracted prepositions are manually tagged with their senses from the TPP dictionary. Idiomatic usage of prepositions like for example or in fact, and complex preposition constructions that involve more than one word (e.g., because of, instead of, etc.) are excluded by the annotators and compiled into a stoplist.",
"cite_spans": [
{
"start": 470,
"end": 496,
"text": "Jurafsky and Martin (2008)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Data Set",
"sec_num": "4"
},
{
"text": "We annotated 3854 instances of the top 7 prepo- Preposition Total Training Test at 404 260 144 for 478 307 171 in 1590 1083 507 of 97 51 46 on 408 246 162 to 532 304 228 with 345 211 134 Total 3854 2462 1392 Table 3 : Number of annotated prepositional phrases for each preposition sitions in Sections 2 to 4 and 23 of the PropBank corpus. The data shows a strong correlation between semantic roles and preposition senses that express a spatial or temporal meaning. For the preposition in, 90.8% of the instances that appear inside an ARGM-LOC are tagged with sense 1 (\"surrounded by or enclosed in\") or sense 5 (\"inclusion or involvement\"). 94.6% of the instances that appear inside an ARGM-TMP role are tagged with sense 2 (\"period of time\"). Our counts furthermore show that about one third of the annotated prepositional phrases fill core roles and that ARGM-LOC and ARGM-TMP are the most frequent roles. The detailed breakdown of semantic roles is shown in Table 2 .",
"cite_spans": [],
"ref_spans": [
{
"start": 48,
"end": 250,
"text": "Preposition Total Training Test at 404 260 144 for 478 307 171 in 1590 1083 507 of 97 51 46 on 408 246 162 to 532 304 228 with 345 211 134 Total 3854 2462 1392 Table 3",
"ref_id": "TABREF1"
},
{
"start": 996,
"end": 1003,
"text": "Table 2",
"ref_id": "TABREF3"
}
],
"eq_spans": [],
"section": "Data Set",
"sec_num": "4"
},
{
"text": "To see how consistent humans can perform the annotation task, we computed the inter-annotator agreement between two annotators on Section 4 of the PropBank corpus. We found that the two annotators assigned the same sense in 86% of the cases. Although not directly comparable, it is interesting to note that this figure is similar to inter-annotator agreement for open-class words reported in previous work (Palmer et al., 2000) . In our final data set, all labels were tagged by the same annotator, which we believe makes our annotation reasonably consistent across different instances. Because we annotate running text, not all prepositions have the same number of annotated instances. The numbers for all seven prepositions are shown in Table 3 . In our experiments, we use Sections 2 to 4 to train the models, and Section 23 is kept for testing. Although our experiments are limited to three sections of training data, it still allows us to train competitive SRL models. Pradhan et al. (2005) have shown that the benefit of using more training data diminishes after a few thousand training instances. We found that the accuracy of our SRL baseline model, which is trained on the 5275 sentences of these three sections, is only an absolute 3.89% lower than the accuracy of the same model when it is trained on twenty sections (71.71% accuracy compared to 75.60% accuracy).",
"cite_spans": [
{
"start": 406,
"end": 427,
"text": "(Palmer et al., 2000)",
"ref_id": "BIBREF11"
},
{
"start": 974,
"end": 995,
"text": "Pradhan et al. (2005)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [
{
"start": 739,
"end": 746,
"text": "Table 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Data Set",
"sec_num": "4"
},
{
"text": "We evaluate the performance of the joint model on the annotated prepositional phrases in test section 23 and compare the results with the performance of the baseline models and the pipeline models. Figure 1 shows the classification accuracy of the WSD models for each of the seven prepositions in the test section. The results show that the pipeline model and the joint model perform almost equally, with the joint model performing marginally better in the overall score. The detailed scores are given in Table 4 . Both models outperform the baseline classifier for three of the seven prepositions: at, for, and to. For the prepositions in, of, and on, the SRL feature did not affect the WSD classification accuracy significantly. For the preposition with, the classification accuracy even dropped by about 6%.",
"cite_spans": [],
"ref_spans": [
{
"start": 198,
"end": 206,
"text": "Figure 1",
"ref_id": "FIGREF0"
},
{
"start": 505,
"end": 512,
"text": "Table 4",
"ref_id": "TABREF5"
}
],
"eq_spans": [],
"section": "Experiments and Results",
"sec_num": "5"
},
{
"text": "Performing the student's t-test, we found that the improvement for the prepositions at, for, and to is statistical significant (p < 0.05), as is the overall improvement. This confirms our hypothesis that the semantic role of the prepositional phrase is a strong hint for the preposition sense. However, our results also show that it is the SRL feature that brings the improvement, not the joint model, because the pipeline and joint model achieve about the same performance.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments and Results",
"sec_num": "5"
},
{
"text": "For the SRL task, we report the classification accuracy over all annotated prepositional phrases in the test section and the F 1 measure for the semantic roles ARGM-LOC and ARGM-TMP. Fig ure 2 shows the results. The joint model shows a small performance increase of 0.43% over the baseline in the overall accuracy. Adding the preposition sense as a feature, on the other hand, significantly lowers the accuracy by over 2%. For ARGM-LOC and ARGM-TMP, the joint model improves the F 1 measure by about 1.3% each. The improvement of the joint model for these roles is statistically significant (p \u2264 0.05, student's ttest). Simply adding the preposition sense in the pipeline model again lowers the F 1 measure. The detailed results are listed in Table 5 . Table 5 : F 1 measure and accuracy of the baseline, pipeline, and joint model on the SRL task in test section 23, statistically significant improvements over the baseline are marked with an (*)",
"cite_spans": [],
"ref_spans": [
{
"start": 183,
"end": 186,
"text": "Fig",
"ref_id": null
},
{
"start": 743,
"end": 750,
"text": "Table 5",
"ref_id": null
},
{
"start": 753,
"end": 760,
"text": "Table 5",
"ref_id": null
}
],
"eq_spans": [],
"section": "Experiments and Results",
"sec_num": "5"
},
{
"text": "Our SRL experiments show that a pipeline model degrades the performance. The reason is the relatively high degree of noise in the WSD classification and that the pipeline model does not discriminate whether the previous classifier predicts the extra feature with high or low confidence. Instead, the model only passes on the 1best WSD prediction, which can cause the next classifier to make a wrong classification based on the erroneous prediction of the previous step. In principle, this problem can be mitigated by training the pipeline model on automatically predicted labels using cross-validation, but in our case we found that automatically predicted WSD labels decreased the performance of the pipeline model even more. In contrast, the joint model computes the full probability distribution over the semantic roles and preposition senses. If the noise level in the first classification step is low, the joint model and the pipeline model perform almost identically, as we have seen in the previous WSD experiments. But if the noise level is high, the joint model can still improve while the pipeline model drops in performance. Our experiments show that the joint model is more robust in the presence of noisy features than the pipeline model.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments and Results",
"sec_num": "5"
},
{
"text": "There is relatively less prior research on prepositions and prepositional phrases in the NLP community. O'Hara and Wiebe (2003) proposed a WSD system to disambiguate function tags of prepositional phrases. An extended version of their work was recently presented in (O'Hara and Wiebe, 2009) . Ye and Baldwin (2006) extended their work to a semantic role tagger specifically for prepositional phrases. Their system first classifies the semantic roles of all prepositional phrases and later merges the output with a general SRL system. Ye and Baldwin (2007) used semantic role tags from surrounding tokens as part of the MELB-YB preposition WSD system. They found that the SRL features did not significantly help their classifier, which is different from our findings. Dang and Palmer (2005) showed that semantic role features are helpful to disambiguate verb senses. Their approach is similar to our pipeline WSD model, but they do not present results with automatically predicted semantic roles. Toutanova et al. (2008) presented a re-ranking model to jointly learn the semantic roles of multiple constituents in the SRL task. Their work dealt with joint learning in SRL, but it is not directly comparable to ours. The difference is that Toutanova et al. attempt to jointly learn semantic role assignment of different constituents for one task (SRL), while we attempt to jointly learn two tasks (WSD and SRL) for one constituent. Because we only look at one constituent at a time, we do not have to restrict ourselves to a re-ranking approach like Toutanova et al., but can calculate the full joint probability distribution of both tasks. Andrew et al. (2004) propose a method to learn a joint generative inference model from partially labeled data and apply their method to the problems of word sense disambiguation for verbs and determination of verb subcategorization frames. Their motivation is similar to ours, but they focus on learning from partially labeled data and they investigate different tasks.",
"cite_spans": [
{
"start": 266,
"end": 290,
"text": "(O'Hara and Wiebe, 2009)",
"ref_id": "BIBREF10"
},
{
"start": 293,
"end": 314,
"text": "Ye and Baldwin (2006)",
"ref_id": "BIBREF16"
},
{
"start": 534,
"end": 555,
"text": "Ye and Baldwin (2007)",
"ref_id": "BIBREF17"
},
{
"start": 767,
"end": 789,
"text": "Dang and Palmer (2005)",
"ref_id": "BIBREF1"
},
{
"start": 996,
"end": 1019,
"text": "Toutanova et al. (2008)",
"ref_id": "BIBREF14"
},
{
"start": 1639,
"end": 1659,
"text": "Andrew et al. (2004)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "6"
},
{
"text": "None of these systems attempted to jointly learn the semantics of the prepositional phrase and the preposition in a single model, which is the main contribution of our work reported in this paper.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "6"
},
{
"text": "We propose a probabilistic model to jointly classify the semantic role of a prepositional phrase and the sense of the associated preposition. We show that learning both tasks together leads to an improvement over competitive, individual models for both subtasks. For the WSD task, we show that the SRL information improves the classification accuracy, although joint learning does not significantly outperform a simpler pipeline model here. For the SRL task, we show that the joint model improves over both the baseline model and the pipeline model, especially for temporal and location arguments. As we only disambiguate the seven most frequent prepositions, potentially more improvement could be gained by including more prepositions into our data set.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "7"
},
{
"text": "http://verbs.colorado.edu/framesets/arrive-v.html",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "Zhang Le's Maximum Entropy Modeling Toolkit, http://homepages.inf.ed.ac.uk/s0450736/maxent toolkit.html",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "This research was supported by a research grant R-252-000-225-112 from National University of Singapore Academic Research Fund.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgements",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Verb Sense and Subcategorization: Using Joint Inference to Improve Performance on Complementary Tasks",
"authors": [
{
"first": "Galen",
"middle": [],
"last": "Andrew",
"suffix": ""
},
{
"first": "Trond",
"middle": [],
"last": "Grenager",
"suffix": ""
},
{
"first": "Christopher",
"middle": [
"D"
],
"last": "Manning",
"suffix": ""
}
],
"year": 2004,
"venue": "Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "150--157",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Galen Andrew, Trond Grenager, and Christopher D. Manning. 2004. Verb Sense and Subcategorization: Using Joint Inference to Improve Performance on Complementary Tasks. In Proceedings of the 2004 Conference on Empirical Methods in Natural Lan- guage Processing (EMNLP 2004), pages 150-157.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "The Role of Semantic Roles in Disambiguating Verb Senses",
"authors": [
{
"first": "Trang",
"middle": [],
"last": "Hoa",
"suffix": ""
},
{
"first": "Martha",
"middle": [],
"last": "Dang",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Palmer",
"suffix": ""
}
],
"year": 2005,
"venue": "Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics (ACL-05)",
"volume": "",
"issue": "",
"pages": "42--49",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hoa Trang Dang and Martha Palmer. 2005. The Role of Semantic Roles in Disambiguating Verb Senses. In Proceedings of the 43rd Annual Meet- ing of the Association for Computational Linguistics (ACL-05), pages 42-49.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Automatic Labeling of Semantic Roles",
"authors": [
{
"first": "Daniel",
"middle": [],
"last": "Gildea",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Jurafsky",
"suffix": ""
}
],
"year": 2002,
"venue": "Computational Linguistics",
"volume": "28",
"issue": "3",
"pages": "245--288",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Daniel Gildea and Daniel Jurafsky. 2002. Automatic Labeling of Semantic Roles. Computational Lin- guistics, 28(3):245-288.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Speech and Language Processing",
"authors": [
{
"first": "Daniel",
"middle": [],
"last": "Jurafsky",
"suffix": ""
},
{
"first": "James",
"middle": [
"H"
],
"last": "Martin",
"suffix": ""
}
],
"year": 2008,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Daniel Jurafsky and James H. Martin. 2008. Speech and Language Processing. Prentice-Hall, Inc. Up- per Saddle River, NJ, USA.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Generalized Inference with Multiple Semantic Role Labeling Systems",
"authors": [
{
"first": "Peter",
"middle": [],
"last": "Koomen",
"suffix": ""
},
{
"first": "Vasin",
"middle": [],
"last": "Punyakanok",
"suffix": ""
},
{
"first": "Dan",
"middle": [],
"last": "Roth",
"suffix": ""
},
{
"first": "Wen-Tau",
"middle": [],
"last": "Yih",
"suffix": ""
}
],
"year": 2005,
"venue": "Proceedings of the 9th Conference on Computational Natural Language Learning",
"volume": "",
"issue": "",
"pages": "181--184",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Peter Koomen, Vasin Punyakanok, Dan Roth, and Wen-tau Yih. 2005. Generalized Inference with Multiple Semantic Role Labeling Systems. In Pro- ceedings of the 9th Conference on Computational Natural Language Learning (CoNLL 2005), pages 181-184.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"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 the 2002 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "41--48",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yoong Keok Lee and Hwee Tou Ng. 2002. An Empir- ical Evaluation of Knowledge Sources and Learn- ing Algorithms for Word Sense Disambiguation. In Proceedings of the 2002 Conference on Empirical Methods in Natural Language Processing (EMNLP 2002), pages 41-48.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "The Preposition Project",
"authors": [
{
"first": "C",
"middle": [],
"last": "Kenneth",
"suffix": ""
},
{
"first": "Orin",
"middle": [],
"last": "Litkowski",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Hargraves",
"suffix": ""
}
],
"year": 2005,
"venue": "Proceedings of the 2nd ACL-SIGSEM Workshop on The Linguistic Dimensions of Prepositions and Their Use in Computational Linguistic Formalisms and Applications",
"volume": "",
"issue": "",
"pages": "171--179",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kenneth C. Litkowski and Orin Hargraves. 2005. The Preposition Project. In Proceedings of the 2nd ACL- SIGSEM Workshop on The Linguistic Dimensions of Prepositions and Their Use in Computational Lin- guistic Formalisms and Applications, pages 171- 179.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "SemEval-2007 Task 06: Word-Sense Disambiguation of Prepositions",
"authors": [
{
"first": "C",
"middle": [],
"last": "Kenneth",
"suffix": ""
},
{
"first": "Orin",
"middle": [],
"last": "Litkowski",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Hargraves",
"suffix": ""
}
],
"year": 2007,
"venue": "Proceedings of the 4th International Workshop on Semantic Evaluations",
"volume": "",
"issue": "",
"pages": "24--29",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kenneth C. Litkowski and Orin Hargraves. 2007. SemEval-2007 Task 06: Word-Sense Disambigua- tion of Prepositions. In Proceedings of the 4th In- ternational Workshop on Semantic Evaluations (Se- mEval 2007), pages 24-29.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Building a Large Annotated Corpus of English: The Penn Treebank",
"authors": [
{
"first": "Mitchell",
"middle": [
"P"
],
"last": "Marcus",
"suffix": ""
},
{
"first": "Mary",
"middle": [
"A"
],
"last": "Marcinkiewicz",
"suffix": ""
},
{
"first": "Beatrice",
"middle": [],
"last": "Santorini",
"suffix": ""
}
],
"year": 1993,
"venue": "Computational Linguistics",
"volume": "19",
"issue": "2",
"pages": "313--330",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mitchell P. Marcus, Mary A. Marcinkiewicz, and Beat- rice Santorini. 1993. Building a Large Annotated Corpus of English: The Penn Treebank. Computa- tional Linguistics, 19(2):313-330.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Preposition Semantic Classification via Penn Treebank and FrameNet",
"authors": [
{
"first": "Janyce",
"middle": [],
"last": "Tom O'hara",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Wiebe",
"suffix": ""
}
],
"year": 2003,
"venue": "Proceedings of the 7th Conference on Computational Natural Language Learning",
"volume": "",
"issue": "",
"pages": "79--86",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Tom O'Hara and Janyce Wiebe. 2003. Preposi- tion Semantic Classification via Penn Treebank and FrameNet. In Proceedings of the 7th Conference on Computational Natural Language Learning (CoNLL 2003), pages 79-86.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Exploiting Semantic Role Resources for Preposition Disambiguation",
"authors": [
{
"first": "Janyce",
"middle": [],
"last": "Tom O'hara",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Wiebe",
"suffix": ""
}
],
"year": 2009,
"venue": "Computational Linguistics",
"volume": "35",
"issue": "2",
"pages": "151--184",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Tom O'Hara and Janyce Wiebe. 2009. Exploiting Se- mantic Role Resources for Preposition Disambigua- tion. Computational Linguistics, 35(2):151-184.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Sense Tagging the Penn Treebank",
"authors": [
{
"first": "Martha",
"middle": [],
"last": "Palmer",
"suffix": ""
},
{
"first": "Hoa",
"middle": [
"Trang"
],
"last": "Dang",
"suffix": ""
},
{
"first": "Joseph",
"middle": [],
"last": "Rosenzweig",
"suffix": ""
}
],
"year": 2000,
"venue": "Proceedings of the 2nd International Conference on Language Resources and Evaluation",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Martha Palmer, Hoa Trang Dang, and Joseph Rosen- zweig. 2000. Sense Tagging the Penn Treebank. In Proceedings of the 2nd International Conference on Language Resources and Evaluation (LREC 2000).",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "The Proposition Bank: An Annotated Corpus of Semantic Roles",
"authors": [
{
"first": "Martha",
"middle": [],
"last": "Palmer",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Gildea",
"suffix": ""
},
{
"first": "Paul",
"middle": [],
"last": "Kingsbury",
"suffix": ""
}
],
"year": 2005,
"venue": "Computational Linguistics",
"volume": "31",
"issue": "1",
"pages": "71--105",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Martha Palmer, Daniel Gildea, and Paul Kingsbury. 2005. The Proposition Bank: An Annotated Cor- pus of Semantic Roles. Computational Linguistics, 31(1):71-105.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Support Vector Learning for Semantic Argument Classification",
"authors": [
{
"first": "Kadri",
"middle": [],
"last": "Sameer Pradhan",
"suffix": ""
},
{
"first": "Valerie",
"middle": [],
"last": "Hacioglu",
"suffix": ""
},
{
"first": "Wayne",
"middle": [],
"last": "Krugler",
"suffix": ""
},
{
"first": "James",
"middle": [
"H"
],
"last": "Ward",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Martin",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Jurafsky",
"suffix": ""
}
],
"year": 2005,
"venue": "Machine Learning",
"volume": "60",
"issue": "",
"pages": "11--39",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sameer Pradhan, Kadri Hacioglu, Valerie Krugler, Wayne Ward, James H. Martin, and Daniel Juraf- sky. 2005. Support Vector Learning for Semantic Argument Classification. Machine Learning, 60(1- 3):11-39.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "A Global Joint Model for Semantic Role Labeling",
"authors": [
{
"first": "Kristina",
"middle": [],
"last": "Toutanova",
"suffix": ""
},
{
"first": "Aria",
"middle": [],
"last": "Haghighi",
"suffix": ""
},
{
"first": "Christopher",
"middle": [
"D"
],
"last": "Manning",
"suffix": ""
}
],
"year": 2008,
"venue": "Computational Linguistics",
"volume": "34",
"issue": "2",
"pages": "161--191",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kristina Toutanova, Aria Haghighi, and Christopher D. Manning. 2008. A Global Joint Model for Se- mantic Role Labeling. Computational Linguistics, 34(2):161-191.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Calibrating Features for Semantic Role Labeling",
"authors": [
{
"first": "Nianwen",
"middle": [],
"last": "Xue",
"suffix": ""
},
{
"first": "Martha",
"middle": [],
"last": "Palmer",
"suffix": ""
}
],
"year": 2004,
"venue": "Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "88--94",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nianwen Xue and Martha Palmer. 2004. Calibrating Features for Semantic Role Labeling. In Proceed- ings of the 2004 Conference on Empirical Methods in Natural Language Processing (EMNLP 2004), pages 88-94.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Semantic Role Labeling of Prepositional Phrases",
"authors": [
{
"first": "Patrick",
"middle": [],
"last": "Ye",
"suffix": ""
},
{
"first": "Timothy",
"middle": [],
"last": "Baldwin",
"suffix": ""
}
],
"year": 2006,
"venue": "ACM Transactions on Asian Language Information Processing (TALIP)",
"volume": "5",
"issue": "3",
"pages": "228--244",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Patrick Ye and Timothy Baldwin. 2006. Seman- tic Role Labeling of Prepositional Phrases. ACM Transactions on Asian Language Information Pro- cessing (TALIP), 5(3):228-244.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "MELB-YB: Preposition Sense Disambiguation Using Rich Semantic Features",
"authors": [
{
"first": "Patrick",
"middle": [],
"last": "Ye",
"suffix": ""
},
{
"first": "Timothy",
"middle": [],
"last": "Baldwin",
"suffix": ""
}
],
"year": 2007,
"venue": "Proceedings of the 4th International Workshop on Semantic Evaluations",
"volume": "",
"issue": "",
"pages": "241--244",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Patrick Ye and Timothy Baldwin. 2007. MELB-YB: Preposition Sense Disambiguation Using Rich Se- mantic Features. In Proceedings of the 4th Interna- tional Workshop on Semantic Evaluations (SemEval 2007), pages 241-244.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"text": "Classification accuracy of the WSD models for the seven most frequent prepositions in test section 23",
"uris": null,
"type_str": "figure",
"num": null
},
"FIGREF1": {
"text": "F 1 measure of the SRL models for ARGM-LOC and ARGM-TMP, and overall accuracy on prepositional phrases in test section 23",
"uris": null,
"type_str": "figure",
"num": null
},
"TABREF0": {
"text": "Features are encoded as binary-valued functions. During testing, the maxent model computes",
"html": null,
"content": "<table><tr><td colspan=\"2\">Baseline Features (Gildea and Jurafsky, 2002)</td></tr><tr><td>pred</td><td>predicate lemma</td></tr><tr><td>path</td><td>path from constituent to predicate</td></tr><tr><td>ptype</td><td>syntactic category (NP, PP, etc.)</td></tr><tr><td>pos</td><td>relative position to the predicate</td></tr><tr><td>voice</td><td>active or passive voice</td></tr><tr><td>hw</td><td>syntactic head word of the phrase</td></tr><tr><td>sub-cat</td><td>rule expanding the predicate's parent</td></tr><tr><td colspan=\"2\">Advanced Features (Pradhan et al., 2005)</td></tr><tr><td>hw POS</td><td>POS of the syntactic head word</td></tr><tr><td>PP hw/POS</td><td>head word and POS of the rightmost</td></tr><tr><td/><td>NP child if the phrase is a PP</td></tr><tr><td>first/last word</td><td>first/last word and POS in the con-</td></tr><tr><td/><td>stituent</td></tr><tr><td>parent ptype</td><td>syntactic category of the parent node</td></tr><tr><td colspan=\"2\">parent hw/POS head word and POS of the parent</td></tr><tr><td>sister ptype</td><td>phrase type of left and right sister</td></tr><tr><td>sister hw/POS</td><td>head word and POS of left and right</td></tr><tr><td/><td>sister</td></tr><tr><td>temporal</td><td>temporal key words present</td></tr><tr><td>partPath</td><td>partial path predicate</td></tr><tr><td>proPath</td><td>projected path without directions</td></tr><tr><td colspan=\"2\">Feature Combinations (Xue and Palmer, 2004)</td></tr><tr><td>pred & ptype</td><td>predicate and phrase type</td></tr><tr><td>pred & hw</td><td>predicate and head word</td></tr><tr><td>pred & path</td><td>predicate and path</td></tr><tr><td>pred & pos</td><td>predicate and relative position</td></tr></table>",
"num": null,
"type_str": "table"
},
"TABREF1": {
"text": "",
"html": null,
"content": "<table/>",
"num": null,
"type_str": "table"
},
"TABREF3": {
"text": "",
"html": null,
"content": "<table><tr><td>: Number of annotated prepositional</td></tr><tr><td>phrases for each semantic role</td></tr></table>",
"num": null,
"type_str": "table"
},
"TABREF4": {
"text": "-",
"html": null,
"content": "<table><tr><td colspan=\"3\">Preposition Baseline Pipeline at 70.83 78.47 * for 41.52 49.12 *</td><td>Joint 78.47 * 49.12 *</td></tr><tr><td>in</td><td>62.33</td><td>61.74</td><td>61.93</td></tr><tr><td>of</td><td>43.48</td><td>43.48</td><td>43.48</td></tr><tr><td>on to</td><td>51.85 58.77</td><td>51.85 67.11 *</td><td>52.47 66.67 *</td></tr><tr><td>with Total</td><td>44.78 56.54</td><td>38.06 58.76 *</td><td>38.06 58.84 *</td></tr></table>",
"num": null,
"type_str": "table"
},
"TABREF5": {
"text": "",
"html": null,
"content": "<table><tr><td/><td/><td colspan=\"3\">: Classification accuracy of the baseline,</td></tr><tr><td colspan=\"5\">pipeline, and joint model on the WSD task in test</td></tr><tr><td colspan=\"5\">section 23, statistically significant improvements</td></tr><tr><td colspan=\"5\">over the baseline are marked with an (*)</td></tr><tr><td/><td>90%</td><td/><td/></tr><tr><td/><td/><td/><td>Baseline</td></tr><tr><td/><td/><td/><td>Pipeline</td></tr><tr><td/><td>85%</td><td/><td>Joint</td></tr><tr><td>f1\u2212measure</td><td>75% 80%</td><td/><td/></tr><tr><td/><td>70%</td><td/><td/></tr><tr><td/><td>65%</td><td>Argm\u2212LOC</td><td>Argm\u2212TMP</td><td>Overall</td></tr></table>",
"num": null,
"type_str": "table"
}
}
}
} |