File size: 86,934 Bytes
6fa4bc9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 | {
"paper_id": "D09-1049",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T16:39:43.627424Z"
},
"title": "Multi-Word Expression Identification Using Sentence Surface Features",
"authors": [
{
"first": "Ram",
"middle": [],
"last": "Boukobza",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Hebrew University of Jerusalem",
"location": {}
},
"email": "ram.boukobza@mail.huji.ac.il"
},
{
"first": "Ari",
"middle": [],
"last": "Rappoport",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Hebrew University of Jerusalem",
"location": {}
},
"email": "arir@cs.huji.ac.il"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "Much NLP research on Multi-Word Expressions (MWEs) focuses on the discovery of new expressions, as opposed to the identification in texts of known expressions. However, MWE identification is not trivial because many expressions allow variation in form and differ in the range of variations they allow. We show that simple rule-based baselines do not perform identification satisfactorily, and present a supervised learning method for identification that uses sentence surface features based on expressions' canonical form. To evaluate the method, we have annotated 3350 sentences from the British National Corpus, containing potential uses of 24 verbal MWEs. The method achieves an F-score of 94.86%, compared with 80.70% for the leading rule-based baseline. Our method is easily applicable to any expression type. Experiments in previous research have been limited to the compositional/non-compositional distinction, while we also test on sentences in which the words comprising the MWE appear but not as an expression.",
"pdf_parse": {
"paper_id": "D09-1049",
"_pdf_hash": "",
"abstract": [
{
"text": "Much NLP research on Multi-Word Expressions (MWEs) focuses on the discovery of new expressions, as opposed to the identification in texts of known expressions. However, MWE identification is not trivial because many expressions allow variation in form and differ in the range of variations they allow. We show that simple rule-based baselines do not perform identification satisfactorily, and present a supervised learning method for identification that uses sentence surface features based on expressions' canonical form. To evaluate the method, we have annotated 3350 sentences from the British National Corpus, containing potential uses of 24 verbal MWEs. The method achieves an F-score of 94.86%, compared with 80.70% for the leading rule-based baseline. Our method is easily applicable to any expression type. Experiments in previous research have been limited to the compositional/non-compositional distinction, while we also test on sentences in which the words comprising the MWE appear but not as an expression.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Multi-Word Expressions (MWEs) such as 'pull strings', 'make a face' and 'get on one's nerves' are very common in language. Such MWEs can be characterized as being non-compositional: the meaning of the expression does not transparently follow from the meaning of the words that comprise it. Much of the work on MWEs in NLP has been in MWE extraction -the discovery of new MWEs from a corpus, using statistical and other methods. Identification of known MWEs in text has received less attention, but is necessary for many NLP applications, for example in machine translation. The current work deals with the MWE identification task: deciding if a sentence contains a use of a known expression.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "MWE identification is not as simple as may initially appear, as will be shown by the performance of two rule-based baselines in our experiments. One source of difficulty is variations in expressions' usage in text. Although MWEs generally show less variation than single words, they show enough that it cannot be ignored. In a study on V+NP idioms, Riehemann (2001) found that the idioms' canonical form accounted for 75% of their appearances in a corpus. Additionally, expressions differ considerably in the types of variations they allow, which include passivization, nominalization and addition of modifying words (Moon, 1998) .",
"cite_spans": [
{
"start": 349,
"end": 365,
"text": "Riehemann (2001)",
"ref_id": "BIBREF19"
},
{
"start": 617,
"end": 629,
"text": "(Moon, 1998)",
"ref_id": "BIBREF17"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "A second source of difficulty is that expressions consisting of very frequent words will often cooccur in sentences in a non-MWE usage and in similar but distinct expressions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "MWE identification can be modeled as a two step process. Given a sentence and a known expression, step (1) is to decide if the sentence contains a potential use of the expression. This is a relatively simple step based on the appearance in the sentence of the words comprising the MWE.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Step (2) is to decide if the potential use is indeed non-compositional. Consider the following sentences with regard to the expression hit the road, meaning 'to leave on a journey': (a) 'At the time, the road was long and difficult with few travelers daring to take it.' (b) 'The headlights of the taxi-van behind us flashed as it hit bumps in the road.'",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "(c) 'The bullets were hitting the road and I could see them coming towards me a lot faster than I was able to reverse.' (d) 'Lorry trailers which would have been hitting the road tomorrow now stand idle.'",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Sentence (a) does not contain a potential use of the expression due to the missing component 'hit'. Each of (b)-(d) does contain a potential use of the expression. In (b) all of the expression components are present, but they do not form an expression. In (c), the words form an expression, but with a compositional (literal) meaning. Only (d) contains a non-compositional use of hit the road. The task we address in this paper is to identify whether or not we are in case (d), for a given expression in a given sentence.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "To date, most work in MWE identification has focused on manually encoding rules that identify expressions in text. The encodings, usually consisting of regular expressions and syntactic structures, are intended to contain all the necessary information for processing the MWE in text. Being manual, this is time-consuming work and requires expert knowledge of individual expressions. In terms of the above model, such encodings handle both MWE identification steps.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "A second approach is to use machine learning methods to learn an expression's behavior from a corpus. Studies taking this approach have focused on distinguishing between compositional and noncompositional uses of an expression (cases (c) and (d) above). As will be detailed in Section 2, existing methods are tailored to an expression's type, and experiment with a single MWE pattern. In addition, the training and test sets they used did not contain non-expression uses as in case (b), which can be quite common in practice.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Our approach is more general. Given a set of sentences with potential MWE uses, we use sentence surface features to create a Support Vector Machine (SVM) classifier for each expression. The classifier is binary and differentiates between non-compositional uses of the expression ((d) above) on the one hand, and compositional and non-expression uses ((b) and (c)) on the other. The experiments and results presented below focus on verbal MWEs, since verbal MWEs are quite common in language use and have also been investigated in related MWE research (e.g., (Cook et al., 2007) ). However, the developed features are not specific to a particular type of expression.",
"cite_spans": [
{
"start": 350,
"end": 354,
"text": "((b)",
"ref_id": null
},
{
"start": 558,
"end": 577,
"text": "(Cook et al., 2007)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The supervised method is compared with two simple rule-based baselines in order to test whether a simple approach is sufficient. In addition, the use of surface features is compared with the use of syntactic features (based on dependency parse trees of the sentences). Averaged over expressions in an independent test set, the supervised classifiers outperform the rule-based baselines, with F-scores of 94.86% (surface features) and 87.77% (syntactic features), compared with 80.70% for the best baseline.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Section 2 reviews previous work. Section 3 discusses the features used for the supervised classifier. Section 4 explains the experimental setting. The results and a discussion are given in sections 5 and 6.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The approach to handling MWEs in early systems was to employ a list of expressions, each with a quasi regular expression that encodes morphosyntactic variations. One example is Leech et al. (1994) who used this method for automatic part-of-speech tagging for the BNC. Another is a formalism called IDAREX (IDioms And Regular EXpressions) (Breidt et al., 1996) .",
"cite_spans": [
{
"start": 177,
"end": 196,
"text": "Leech et al. (1994)",
"ref_id": "BIBREF15"
},
{
"start": 338,
"end": 359,
"text": "(Breidt et al., 1996)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "MWE Lexical Encoding",
"sec_num": "2.1"
},
{
"text": "More recent research emphasizes the integration of MWE lexical entries into existing single word lexicons and grammar systems (Villavicencio et al., 2004; Alegria et al., 2004) . There is also an attempt to take advantage of regularities in morpho-syntactic properties across MWE groups, which allows encoding the behavior of the group instead of individual expressions (Villavicencio et al., 2004; Gr\u00e9goire, 2007) . Fellbaum (1998) discusses some difficulties in representing idioms, which are largely figurative in meaning, in Word-Net. More recent work (Fellbaum et al., 2006) focuses on German VP idioms.",
"cite_spans": [
{
"start": 126,
"end": 154,
"text": "(Villavicencio et al., 2004;",
"ref_id": "BIBREF21"
},
{
"start": 155,
"end": 176,
"text": "Alegria et al., 2004)",
"ref_id": "BIBREF0"
},
{
"start": 370,
"end": 398,
"text": "(Villavicencio et al., 2004;",
"ref_id": "BIBREF21"
},
{
"start": 399,
"end": 414,
"text": "Gr\u00e9goire, 2007)",
"ref_id": "BIBREF11"
},
{
"start": 417,
"end": 432,
"text": "Fellbaum (1998)",
"ref_id": "BIBREF10"
},
{
"start": 556,
"end": 579,
"text": "(Fellbaum et al., 2006)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "MWE Lexical Encoding",
"sec_num": "2.1"
},
{
"text": "As already mentioned, one issue with lexical encoding is that it is done manually, making lexicons difficult to create, maintain and extend. The use of regularities among different types of MWEs is one way of reducing the amount of work required. A second issue is that implementations tend to ignore the likelihood and even the possibility of compositional and other interpretations of expressions in text, which can be common for some expressions. For example, in an MWE identification study, Hashimoto et al. (2006) built an identification system using hand crafted rules for some 100 Japanese idioms. The results showed near perfect performance on expressions without compositional/noncompositional ambiguity but significantly poorer performance on expressions with ambiguity. Katz and Giesbrecht (2006) used a supervised learning method to distinguish between compositional and non-compositional uses of an expression (in German text) by using contextual information in the form of Latent Semantic Analysis (LSA) vectors. LSA vectors of compositional and non-compositional meaning were built from a training set of example sentences and then a nearest neighbor algorithm was applied on the LSA vector of one tested MWE. The technique was tested more thoroughly in Cook et al. (2007) . Cook et al. (2007) devised two unsupervised methods to distinguish between compositional (literal) and non-compositional (idiomatic) tokens of verb-object expressions. The first method is based on an expression's canonical form. In a previous study (Fazly and Stevenson, 2006) , the authors came up with a dozen possible syntactic forms for verb-object pairs (based on passivization, determiner, and object pluralization) and used a corpusbased statistical measure to determine the canonical form(s). The method classifies new tokens as idiomatic if they use a canonical form, and literal otherwise.",
"cite_spans": [
{
"start": 495,
"end": 518,
"text": "Hashimoto et al. (2006)",
"ref_id": "BIBREF12"
},
{
"start": 781,
"end": 807,
"text": "Katz and Giesbrecht (2006)",
"ref_id": "BIBREF14"
},
{
"start": 1269,
"end": 1287,
"text": "Cook et al. (2007)",
"ref_id": "BIBREF3"
},
{
"start": 1290,
"end": 1308,
"text": "Cook et al. (2007)",
"ref_id": "BIBREF3"
},
{
"start": 1539,
"end": 1566,
"text": "(Fazly and Stevenson, 2006)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "MWE Lexical Encoding",
"sec_num": "2.1"
},
{
"text": "The second method uses context as well as form. Co-occurrence vectors representing the idiomatic and literal meaning of each expression were computed based on corpus data. Idiomaticmeaning vectors were based on examples matching the expressions' canonical form. Literal meaning vectors were based on examples that did not match the canonical form. New tokens were classified as literal/idiomatic based on their (cooccurrence) vector's cosine similarity to the idiomatic and literal vectors. (Sporleder and Li, 2009 ) also attempted to distinguish compositional from non-compositional uses of expressions in text. Their assumption was that if an expression is used literally, but not idiomatically, its component words will be related semantically to several words in the surrounding discourse. For example, when the expression 'play with fire' is used literally, words such as 'smoke, 'burn', 'fire department', and 'alarm' tend to also be used nearby; when it is used idiomatically, they aren't (indeed, other words, e.g., 'danger' or 'risk' appear nearby but they are not close semantically to 'play' or to 'fire'). This property was used to distinguish literal and non-literal instances by measuring the semantic relatedness of an expression's component words to nearby words in the text. If one or more of the expression's components were sufficiently related to enough nearby words, forming a 'lexical chain', the usage was classified as literal. Otherwise it was idiomatic. Two classifiers based on lexical chains were devised. These were compared with a supervised method that trains a classifier for each expression based on surrounding context. The results showed that the supervised classifier method did much better (90% F-score on literal uses) than the lexical chain classifier methods (60% F-score).",
"cite_spans": [
{
"start": 491,
"end": 514,
"text": "(Sporleder and Li, 2009",
"ref_id": "BIBREF20"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "MWE Identification by ML",
"sec_num": "2.2"
},
{
"text": "In the above studies the focus is on the compositional/non-compositional expression distinction. The sentence data used contains examples of either one or the other. In (Sporleder and Li, 2009) the experimental data included only sentences in which the expressions were in canonical form (allowing for verb inflection). In (Cook et al., 2007) a syntactic parser was used to collect sentences containing the MWEs in the active and passive voice using heuristics. Thus, examples such as the following (from the BNC) would not be included in their sample:",
"cite_spans": [
{
"start": 169,
"end": 193,
"text": "(Sporleder and Li, 2009)",
"ref_id": "BIBREF20"
},
{
"start": 323,
"end": 342,
"text": "(Cook et al., 2007)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "MWE Identification by ML",
"sec_num": "2.2"
},
{
"text": "1. take a chance: 'While he still had a chance of being near Maisie, he would take it'.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "MWE Identification by ML",
"sec_num": "2.2"
},
{
"text": "2. face the consequences: '. . . she did not have to face, it appears, the possible serious or even fatal consequences of her decision'.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "MWE Identification by ML",
"sec_num": "2.2"
},
{
"text": "3. make a distinction: 'Logically, the distinction between the two aspects of the theory can and should be made'.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "MWE Identification by ML",
"sec_num": "2.2"
},
{
"text": "4. break the ice: 'The ice, if not broken, was beginning to soften a little'.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "MWE Identification by ML",
"sec_num": "2.2"
},
{
"text": ". settle a score: 'Morrissey had another score to settle'.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "5",
"sec_num": null
},
{
"text": "This means that their experiments have not included all types of sentences that might be encountered in practice when attempting MWE identifi-cation. Specifically, they would miss many examples in which the MWE words are present but are not used as an expression (case (b) in Section 1). Moreover, their heuristics are tailored to the Verb-Direct Object MWE type. Different heuristics would need to be employed for different MWE types.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "5",
"sec_num": null
},
{
"text": "In our approach there is no pre-processing stage requiring type-specific knowledge. Specifically, the above examples are used as training sentences in our experiments.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "5",
"sec_num": null
},
{
"text": "There exists an extensive body of research on MWE extraction (see Wermter and Hahn (2004) for a review), where the only input is a corpus, and the output is a list of MWEs found in it. Most methods collect MWE candidates from the corpus, score them according to some association measure between their components, and accept candidates with scores passing some threshold. The focus of research has been on developing association measures, including statistical, information-theoretic and linguistically motivated measures (e.g., Justeson and Katz (1995) , Wermter and Hahn (2006) , and Deane (2005) ).",
"cite_spans": [
{
"start": 66,
"end": 89,
"text": "Wermter and Hahn (2004)",
"ref_id": "BIBREF22"
},
{
"start": 541,
"end": 552,
"text": "Katz (1995)",
"ref_id": "BIBREF13"
},
{
"start": 555,
"end": 578,
"text": "Wermter and Hahn (2006)",
"ref_id": "BIBREF23"
},
{
"start": 585,
"end": 597,
"text": "Deane (2005)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "MWE Extraction",
"sec_num": "2.3"
},
{
"text": "Our method decides if a potential use of a known expression in a given sentence is noncompositional. The input to the method, for each MWE, is a labeled training set of sentences containing one or more potentially non-compositional uses of the MWE. The output, for each MWE, is a binary classifier, trained on those sentences. Thus, we target step (2) of MWE identification, which is the difficult one.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "MWE Identification Method",
"sec_num": "3"
},
{
"text": "The learning algorithm used is Support Vector Machine (SVM), which outputs a binary classifier, using Sequential Minimal Optimization (Platt, 1998) 1 in the Weka toolkit 2 (Witten and Frank, 2000) .",
"cite_spans": [
{
"start": 134,
"end": 147,
"text": "(Platt, 1998)",
"ref_id": "BIBREF18"
},
{
"start": 172,
"end": 196,
"text": "(Witten and Frank, 2000)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "MWE Identification Method",
"sec_num": "3"
},
{
"text": "For training, sentences are converted into feature vectors. Features depend on the assignment of the lexical components of the expression to specific tokens in the sentence. In some cases, there are several tokens in the sentence that match a single component in the expression, and this leads to 1 Using the PUK kernel (The Pearson VII functionbased Universal Kernel), with parameters omega=1.0 and sigma=1.0.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "MWE Identification Method",
"sec_num": "3"
},
{
"text": "2 Weka version 3.5.6; www.cs.waikato.ac.nz/ ml/ weka/ multiple (potential) assignments. So in the general case a sentence is converted to a set of feature vectors, each corresponding to a single assignment of the MWE's lexical components to sentence tokens.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "MWE Identification Method",
"sec_num": "3"
},
{
"text": "Training sentences are labeled positive if they contain a non-compositional use of the expression and negative if they do not (i.e., literal and other uses). If the sentence is positive, at least one of the assignments is the true assignment (there may be more than one, e.g., when an expression is used twice in the same sentence). The vector matching the true assignment is labeled positive. The others are labeled negative. If the sentence is negative, all of the vectors are labeled negative.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "MWE Identification Method",
"sec_num": "3"
},
{
"text": "As mentioned, the output of the method is a distinct binary classifier for each MWE. Although having a single classifier for all expressions would seem advantageous, the wide variation exhibited by MWEs (e.g., for some the passive is common, for other not at all) precludes this option and requires having a separate classifier for each expression.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "MWE Identification Method",
"sec_num": "3"
},
{
"text": "Surface features include order and distance, partof-speech and inflection of an expression's words in a sentence.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Features",
"sec_num": "3.1"
},
{
"text": "Use of surface features is intuitive and relatively cheap. In addition, many studies have shown the importance of order and distance in MWE extraction in English (two recent examples are (Dias, 2003; Deane, 2005) ). Thus, we develop a supervised classifier based on surface features.",
"cite_spans": [
{
"start": 187,
"end": 199,
"text": "(Dias, 2003;",
"ref_id": "BIBREF6"
},
{
"start": 200,
"end": 212,
"text": "Deane, 2005)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Features",
"sec_num": "3.1"
},
{
"text": "Many of the surface features make use of an expression's Canonical Form (CF), thus the learning algorithm assumes that it is given such a form. Formally defining the CF is difficult. Indeed, some researchers have concluded that some expressions do not have a CF (Moon, 1998) . For our purposes, CF can be informally defined as the most frequent form in which the expression appears. In practice, an approximation of this definition, explained in Section 4, is used.",
"cite_spans": [
{
"start": 262,
"end": 274,
"text": "(Moon, 1998)",
"ref_id": "BIBREF17"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Features",
"sec_num": "3.1"
},
{
"text": "1. Word Distance: The number of words between the leftmost and rightmost MWE tokens in the sentence.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Surface Features",
"sec_num": "3.1.1"
},
{
"text": "2. Ordered Gap List: A list of gaps, measured in number of words, between each pair of the expression's tokens in their canonical form order. For example, if the token locations (in canonical form order) are 10, 7 and 3, the ordered gap list would be (10 \u2194 7 = 2, 10 \u2194 3 = 6, 7 \u2194 3 = 3).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Surface Features",
"sec_num": "3.1.1"
},
{
"text": "3. Word Order: A boolean value indicating whether the expression's word order in the sentence matches the canonical form word order.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Surface Features",
"sec_num": "3.1.1"
},
{
"text": "4. Word Order Permutation: The permutation of word order relative to the canonical form. For example, the permutation (1,0,2) indicates that component words 1 and 0 have switched order in the sentence.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Surface Features",
"sec_num": "3.1.1"
},
{
"text": "The fraction of words in the expression that have undergone inflection relative to the canonical form.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Inflection Ratio:",
"sec_num": "5."
},
{
"text": "A list of the tokens in the sentence matching the expression's component words, ordered according to canonical form. For example, if the expression is 'make a distinction', a possible lexical values list is (made,no,distinction) in the sentence 'No possible distinction can be made between the two'.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Lexical Values:",
"sec_num": "6."
},
{
"text": "A boolean value indicating whether the expression's use in the sentence has the same part-of-speech pattern as the canonical form.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "POS Pattern:",
"sec_num": "7."
},
{
"text": "Two combinations of surface features are used in the experiments below. The first, named R1, uses all of the above features. The second, R2, uses only Word Distance, Ordered Gap List and Word Order Permutation. Using R2 the learner has only word order and distance information from which to create a classifier.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "POS Pattern:",
"sec_num": "7."
},
{
"text": "An expression's words may appear unrelated in a sentence, because of distance, order, part-ofspeech and other surface variations. However, the words will still be closely related syntactically. Syntactic analysis of the sentence in the form of a dependency parse tree directly gives the syntactic relationships between the expression's components. Thus, we also develop a classifier based on syntactic features. Dependency Parsing. A dependency parse tree is a directed acyclic graph in which the nodes represent tokens in the sentence and the edges represent syntactic dependencies between the words (e.g., direct-object, prepositional-object, nounsubject etc.). The Stanford Parser 3 (Marneffe et al., 2006) was used.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Syntactic Features",
"sec_num": "3.1.2"
},
{
"text": "Minimal Sub-Tree. To compute a syntactic feature, the dependency tree is computed and then the minimal sub-tree containing the expression's tokens is extracted.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Syntactic Features",
"sec_num": "3.1.2"
},
{
"text": "The features are:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Syntactic Features",
"sec_num": "3.1.2"
},
{
"text": "1. Sub-Tree Distance Sum: The number of edges in the minimal sub-tree. A large number of edges suggests a weaker dependency.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Syntactic Features",
"sec_num": "3.1.2"
},
{
"text": "2. Sub-Tree Distance List: A list of the distances of the MWE component nodes from the root of their sub-tree.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Syntactic Features",
"sec_num": "3.1.2"
},
{
"text": "ists if there is a directed path from one node (the ancestor) to the other (the descendant). Descendant relations are either direct (parentchild) or indirect. The list consists of the levels of descendant relations between the MWE component nodes, which can be none, indirect or direct.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Descendant Relations List: A list of descendant relations between each pair of MWE component nodes. A descendant relation between two nodes ex-",
"sec_num": "3."
},
{
"text": "A list of the directions of the descendant relations between each pair of MWE component nodes. If there are descendant relations between a pair of nodes, the direction of the dependency, indicating which is the modifying and which the modified node, is important. In the experiments reported below, the classifier using only the syntactic features is denoted by S, and the one using all surface and all syntactic features is denoted by C. We have experimented with additional feature combinations, with no improvement in results.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Descendant Direction List:",
"sec_num": "4."
},
{
"text": "Canonical form. As described, an expression's canonical form (CF) is used in many of the learning algorithm's features. The CF is taken from Collins COBUILD Advanced Learner's English Dictionary (2003) which is also used as our source for MWEs. COBUILD is an English-English dictionary based on the Bank of English (BOE) corpus (over 520 million words) with approximately 34,000 entries.",
"cite_spans": [
{
"start": 141,
"end": 201,
"text": "Collins COBUILD Advanced Learner's English Dictionary (2003)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Method",
"sec_num": "4"
},
{
"text": "Traditional single-word dictionaries are a good source for expressions because they usually list, as part of single-word entries, expressions in which the word is a component. The CF is not explicitly given in COBUILD, so an approximation is the form which appears in the expression's definition. This is a reasonable approximation since the COBUILD authors claim to have selected typical uses of the expressions in their definitions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Method",
"sec_num": "4"
},
{
"text": "Each CF also has a matching part-of-speech (POS) pattern, which is a list of the partsof-speech of the components in the CF. For example, 'walking on air' has the pattern (V erb, P reposition, N oun). COBUILD does not include part-of-speech information for expressions so this information was determined using the British National Corpus (BNC) (BNC, 2001 ), a (mostly) automatically POS tagged corpus (using the CLAWS tagger). For each MWE, the POS patterns of all instances of the CF in the corpus were counted. The most frequent pattern is the expression's POS pattern.",
"cite_spans": [
{
"start": 344,
"end": 354,
"text": "(BNC, 2001",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Method",
"sec_num": "4"
},
{
"text": "The expressions. A set of 17 verbal MWEs, the development set, was used for development of the surface and syntactic features described above. All of the development set MWEs had the POS pattern (V erb, Determiner, N oun). Another set of 24 verbal MWEs, the training/test set 4 , was then used to test the method. Because the method is not specific to the (V erb, Determiner, N oun) pattern, new POS patterns are included in the training/test set. The training/test set consists of 8 MWEs of the POS pattern (V erb, Determiner, N oun), 7 (V erb, P reposition, N oun) MWEs and and 9 (V erb, N oun, P reposition) MWEs. The list of MWEs was selected randomly from the corresponding POS pattern types. MWEs with a positive or negative percentage of under 5% in their data set were discarded 5 . The MWEs, in their canonical form, are: Development set:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Method",
"sec_num": "4"
},
{
"text": "(V erb, Determiner, N oun) [17]",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Method",
"sec_num": "4"
},
{
"text": ": break the ice, calls the shots, catch a cold, clear the air, face the consequences, fits the bill, hit the road, make a face, make a distinction, makes an impression, raise the alarm, set an example, sound the alarm, stay the course, take a chance, take the initiative, tie the knot. Training/test set: (V erb, Determiner, N oun) [8]: changes the subject, get a grip, get the picture, lead the way, makes the grade, sets the scene, take a seat, take the plunge; (V erb, P reposition, N oun) [7]: fall into place, goes to extremes, brought to justice, take to heart, gets on nerves, keep up appearances, comes to light; (V erb, N oun, P reposition) [9]: take aim at, make allowances for, takes advantage of, keep hands off, lay claim to, take care of, make contact with, gives rise to, wash hands of.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Method",
"sec_num": "4"
},
{
"text": "The sentences. As mentioned, the first step of MWE identification is to identify if the sentence contains a potential non-compositional use of the expression. In order to test our method, which targets step (2), a set of such sentences (for each expression) was collected from the BNC corpus and then labeled for use as training/test sentences 6 .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Method",
"sec_num": "4"
},
{
"text": "The collection method was intended to allow a wide range of variations in expression use. In practice, for each expression sentences containing all of the expression's CF components, in any of their inflections, were collected, but excluding common auxiliary words. So for example, when targeting the MWE 'make an impression' we allowed inflections of 'make' and 'impression' and did not require 'an', to allow for variations such as 'make no impression' and 'make some impression'. For some expressions, sentences were limited to those with a distance of up to 8 words between each expression component. Very long sentences (above 80 words) were discarded. The final set of sentences was then randomly selected.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Method",
"sec_num": "4"
},
{
"text": "Given this method, training/test sentences allow non-lexical variations: inflection, word order, part-of-speech, syntactic structure and other non-syntactic transformations. Lexical variations which involve a change in one of the expression's components are not allowed, except for common auxiliary words.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Method",
"sec_num": "4"
},
{
"text": "For the development set an average of 97 (40-137) sentences were collected per MWE, giving a total of 1663 sentences, with a micro average of 49% positive labels. For the training/test set there were 139 (73-150) sentences per MWE on average, totaling 3350, with a 40% average positive ratio.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Method",
"sec_num": "4"
},
{
"text": "The sentences were manually labeled as positive if they contained a non-compositional use of the MWE and negative if they contained a compositional or non-expression usage. Judgment was based on a single sentence, without wider context. The first method, CanonicalForm (CF), accepts a sentence use as a non-compositional MWE use if and only if the MWE is in canonical form (there are no intervening words between the MWE components, their order matches canonical-form order, and there is an inflection in at most one component word).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Method",
"sec_num": "4"
},
{
"text": "The second method, DistanceOrder (DO), ac- cepts a sentence use if and only if the number of words between the leftmost and rightmost MWE components is less than or equal to 2 (not counting the middle MWE component), and if the order matches the canonical form order.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Method",
"sec_num": "4"
},
{
"text": "The baseline methods (CF and DO) and the supervised methods (R1,R2,S,C) were run on the development and training/test sets. For the supervised methods, for each MWE we used 10-fold crossvalidation 7 . Tables 1 and 2 summarize the results for the development and test sets, respectively. For the development set, average results over all 17 MWEs and over the best 8 MWEs (on R1), a group size comparable to the test set, are shown. For the test set, results over all 24 MWEs and the three MWE types tested are shown.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "The tables show average overall accuracy and average precision, recall and F-score on positive instances, where the averages are taken over the results of the individual MWEs (i.e., microaveraged).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "Baselines. Baseline accuracy, (for DO) 82.27% on the development set and 87.2% on the test set (over all groups), is probably insufficient for many NLP applications.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "The baselines perform similarly in terms of average accuracy. CF does this with very high precision and low recall, while for DO recall improves at the expense of precision. Looking at individual MWEs reveals that for expressions which allow more variation in terms of intervening words Baselines versus supervised classifiers. As shown in the tables, R1 outperforms the best baseline in terms of accuracy in both test and development. Moreover, the supervised classifiers are more stable in their accuracy. For the development set the standard deviation of accuracy scores averages 22.58 for CF and DO, and 6.68 for R1, R2, S, and C. For the test set the baselines average 9.07 (Verb-Det-Noun), 11.11 (Verb-Prep-Noun) and 14.26 (Verb-Noun-Prep), and the supervised methods average 4.97 (Verb-Det-Noun), 7.66 (Verb-Prep-Noun) and 7.97. This stability means that the supervised classifiers are able to perform well on MWEs with different behavior. For example, R1 is able to perform well on expressions where order is strict, as DO does (e.g., make a face), while also performing well on those where order varies (e.g., make a distinction).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "Supervised classifiers. R1 and R2, based on surface features, show similar accuracy values, with R1 doing somewhat better in the Verb-Prep-Noun and Verb-Noun-Prep groups. This is due to the Lexical Values feature, which accounts for a change in preposition. A change in preposition (as in 'wash hands of some matter' versus 'wash hands in the sink') is more significant than a change in determiner in the Verb-Determiner-Noun group. This improves precision on negative instances, which are rejected more precisely based on the preposition value. Nevertheless, the relatively simple features in R2, essentially order and distance, perform quite well.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "The F-score result for R1, 94.86, is an improvement over the F-score result of the supervised classifier used in (Sporleder and Li, 2009) , 90.15. Although the sentence data is different (our data includes sentences with non-expression uses) the number of sentences used is similar.",
"cite_spans": [
{
"start": 113,
"end": 137,
"text": "(Sporleder and Li, 2009)",
"ref_id": "BIBREF20"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "S, based on syntactic features, performs worse than R1/2. It shows better accuracy than the baselines in all but the (Verb-Prep-Noun) group and is also more stable. C, a combination of surface and syntactic features, performs better than S and slightly worse than R1/2. Why do the syntactic features perform worse than surface features? An analysis of the S classifier errors reveals two important causes. First, there is substantial variation in the dependency tree structures of the non-compositional uses of the expressions as output by the parser. Thus, the syntactic feature classifier was more difficult to learn than the surface feature one, requiring a larger training set. This is not surprising, given that many MWEs exhibit an irregular syntactic behavior that might even seem strange at times. For example, in the sentence fragment \"and then he came to.\", 'came to' is an MWE. A parser might find it difficult to parse the sentence correctly, expecting a noun phrase to follow the 'to'.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "Second, as described above, the syntactic features consist of general syntactic relations extracted from the parse tree and not type-specific knowledge. As a result, literal or non-expression uses of the MWE's components, which have a close syntactic relation in a given sentence, appear as non-compositional uses of the expression to the classifier.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "5"
},
{
"text": "This study has addressed MWE identification: deciding if a potential use of an expression is a noncompositional one. Despite its importance in basic NLP tasks, the problem has been largely overlooked in NLP research, probably due to it presumed simplicity. However, as we have shown, simple methods for MWE identification, such as our baselines, do not perform consistently well across MWEs. This study serves to highlight this point and the need for more sophisticated methods for MWE identification.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6"
},
{
"text": "We have shown that using a supervised learning method employing surface sentence features based on canonical form, it is possible to improve performance significantly. Unlike previous research, our method is not tailored to specific MWE types, and we did not ignore non-expression uses in our experiments.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6"
},
{
"text": "Future research should experiment with nonverbal MWEs, since our features are not specific to verbal MWE types. Another direction is a more sophisticated corpus sampling algorithm. The current work ignored MWEs which had an unbalanced training set (usually too few positives). Methods for gathering enough positive instances of such MWEs will be useful for testing the methods proposed here, as well as for general MWE research.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "6"
},
{
"text": "http://www-nlp.stanford.edu/software/lex-parser.shtml",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "Using 10-fold cross validation.5 Initially there were 20 MWEs in the development set and 30 (10 per group) in the training/test set.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "The PyLucene software package, http://pylucene. osafoundation. org/, was used for building an index to the BNC and for searching.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "I.e., we ran 10 experiments where in each experiment we divided the corresponding annotated sentence sets into 90% training sentences and 10% test sentences, and the results reported are the average of the 10 experiments.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "and lexical change, DO outperforms CF. To name a few, make an impression, raise the alarm, take a chance and make allowances for. For example, for take a chance intervening words are quite common, as in: 'I'm taking a real chance on you.', or a change in determiner as in: 'I preferred to take my chances'. Indeed, CF showed poor precision only for MWEs with a common literal usage. Two such MWEs were present in the development set (break the ice and tie the knot ) and two in the test set (wash hands of and keep hands off).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "annex",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Representation and treatment of multiword expressions in Basque",
"authors": [
{
"first": "Olatz",
"middle": [],
"last": "I\u00f1ki Alegria",
"suffix": ""
},
{
"first": "Xabier",
"middle": [],
"last": "Ansa",
"suffix": ""
},
{
"first": "Nerea",
"middle": [],
"last": "Artola",
"suffix": ""
},
{
"first": "Koldo",
"middle": [],
"last": "Ezeiza",
"suffix": ""
},
{
"first": "Ruben",
"middle": [],
"last": "Gojenola",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Urizar",
"suffix": ""
}
],
"year": 2004,
"venue": "ACL '04 Workshop on Multiword Expressions",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "I\u00f1ki Alegria, Olatz Ansa, Xabier Artola, Nerea Ezeiza, Koldo Gojenola and Ruben Urizar. 2004. Repre- sentation and treatment of multiword expressions in Basque. ACL '04 Workshop on Multiword Expres- sions.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "The British National Corpus, version 2 (BNC World)",
"authors": [],
"year": 2001,
"venue": "Distributed by Oxford University Computing Services on behalf of the BNC Consortium",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "The British National Corpus. 2001. The British National Corpus, version 2 (BNC World). Dis- tributed by Oxford University Computing Ser- vices on behalf of the BNC Consortium. URL: http://www.natcorp.ox.ac.uk/",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Local grammars for the description of multi-word lexemes and their automatic recognition in texts. COMPLEX '96",
"authors": [
{
"first": "Elisabeth",
"middle": [],
"last": "Breidt",
"suffix": ""
},
{
"first": "Frederique",
"middle": [],
"last": "Segond",
"suffix": ""
},
{
"first": "Giuseppen",
"middle": [],
"last": "Valetto",
"suffix": ""
}
],
"year": 1996,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Elisabeth Breidt, Frederique Segond, and Giuseppen Valetto. 1996. Local grammars for the description of multi-word lexemes and their automatic recogni- tion in texts. COMPLEX '96. Budapest.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Pulling their weight: Exploiting syntactic forms for the automatic identification of idiomatic expressions in context",
"authors": [
{
"first": "Paul",
"middle": [],
"last": "Cook",
"suffix": ""
},
{
"first": "Afsaneh",
"middle": [],
"last": "Fazly",
"suffix": ""
},
{
"first": "Suzanne",
"middle": [],
"last": "Stevenson",
"suffix": ""
}
],
"year": 2007,
"venue": "ACL '07 Workshop on A Broader Perspective on Multiword Expressions",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Paul Cook, Afsaneh Fazly, and Suzanne Stevenson. 2007. Pulling their weight: Exploiting syntactic forms for the automatic identification of idiomatic expressions in context. ACL '07 Workshop on A Broader Perspective on Multiword Expressions.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Collins COBUILD Advanced Learner's English Dictionary",
"authors": [
{
"first": "Cobuild",
"middle": [],
"last": "Collins",
"suffix": ""
}
],
"year": 2003,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Collins COBUILD. 2003. Collins COBUILD Ad- vanced Learner's English Dictionary. Harper- Collins Publishers, 4th edition.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "A nonparametric method for extraction of candidate phrasal terms",
"authors": [
{
"first": "Paul",
"middle": [],
"last": "Deane",
"suffix": ""
}
],
"year": 2005,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Paul Deane. 2005. A nonparametric method for ex- traction of candidate phrasal terms. ACL '05.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Multiword unit hybrid extraction",
"authors": [
{
"first": "Gael",
"middle": [],
"last": "Dias",
"suffix": ""
}
],
"year": 2003,
"venue": "ACL '03 Workshop on Multiword Expressions: Analysis, Acquisition and Treatment",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Gael Dias. 2003. Multiword unit hybrid extrac- tion. ACL '03 Workshop on Multiword Expressions: Analysis, Acquisition and Treatment.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Automatically constructing a lexicon of verb phrase idiomatic combinations",
"authors": [
{
"first": "Afsaneh",
"middle": [],
"last": "Fazly",
"suffix": ""
},
{
"first": "Suzanne",
"middle": [],
"last": "Stevenson",
"suffix": ""
}
],
"year": 2006,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Afsaneh Fazly and Suzanne Stevenson. 2006. Auto- matically constructing a lexicon of verb phrase id- iomatic combinations. EACL '06.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Corpus-based studies of German idioms and light verbs",
"authors": [],
"year": null,
"venue": "International Journal of Lexicography",
"volume": "19",
"issue": "4",
"pages": "349--361",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Corpus-based studies of German idioms and light verbs. International Journal of Lexicography, 19(4):349-361.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Towards a representation of idioms in WordNet. COLING-ACL '98 Workshop on the Use of WordNet in Natural Language Processing Systems",
"authors": [
{
"first": "Christiane",
"middle": [],
"last": "Fellbaum",
"suffix": ""
}
],
"year": 1998,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Christiane Fellbaum. 1998. Towards a representation of idioms in WordNet. COLING-ACL '98 Workshop on the Use of WordNet in Natural Language Pro- cessing Systems.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Design and implementation of a lexicon of Dutch multiword expressions",
"authors": [
{
"first": "Nicole",
"middle": [],
"last": "Gr\u00e9goire",
"suffix": ""
}
],
"year": 2007,
"venue": "ACL '07 Workshop on A Broader Perspective on Multiword Expressions",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nicole Gr\u00e9goire. 2007. Design and implementation of a lexicon of Dutch multiword expressions. ACL '07 Workshop on A Broader Perspective on Multiword Expressions.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Japanese idiom recognition: Drawing a line between literal and idiomatic meanings",
"authors": [
{
"first": "Chikara",
"middle": [],
"last": "Hashimoto",
"suffix": ""
},
{
"first": "Satoshi",
"middle": [],
"last": "Sato",
"suffix": ""
},
{
"first": "Takehito",
"middle": [],
"last": "Utsuro",
"suffix": ""
}
],
"year": 2006,
"venue": "COLING-ACL '06",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Chikara Hashimoto, Satoshi Sato, and Takehito Utsuro. 2006. Japanese idiom recognition: Drawing a line between literal and idiomatic meanings. COLING- ACL '06, Poster Sessions.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Technical terminology: some linguistic properties and an algorithm for identification in text",
"authors": [
{
"first": "John",
"middle": [
"S"
],
"last": "Justeson",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Slava",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Katz",
"suffix": ""
}
],
"year": 1995,
"venue": "Natural Language Engineering",
"volume": "1",
"issue": "",
"pages": "9--27",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "John S. Justeson and Slava M. Katz. 1995. Technical terminology: some linguistic properties and an al- gorithm for identification in text. Natural Language Engineering, 1:9-27.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Automatic identification of non-compositional multiword expressions using latent semantic analysis",
"authors": [
{
"first": "Graham",
"middle": [],
"last": "Katz",
"suffix": ""
},
{
"first": "Eugenie",
"middle": [],
"last": "Giesbrecht",
"suffix": ""
}
],
"year": 2006,
"venue": "COLING-ACL '06 Workshop on Multiword Expressions: Identifying and Exploiting Underlying Properties",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Graham Katz and Eugenie Giesbrecht. 2006. Au- tomatic identification of non-compositional multi- word expressions using latent semantic analysis. COLING-ACL '06 Workshop on Multiword Expres- sions: Identifying and Exploiting Underlying Prop- erties..",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "CLAWS4: The tagging of the British National Corpus. COLING '94",
"authors": [
{
"first": "Geoffrey",
"middle": [],
"last": "Leech",
"suffix": ""
},
{
"first": "Roger",
"middle": [],
"last": "Garside",
"suffix": ""
},
{
"first": "Michael",
"middle": [],
"last": "Bryant",
"suffix": ""
}
],
"year": 1994,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Geoffrey Leech, Roger Garside and Michael Bryant. 1994. CLAWS4: The tagging of the British Na- tional Corpus. COLING '94.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Generating typed dependency parses from phrase structure parses",
"authors": [
{
"first": "Marie-Catherine",
"middle": [],
"last": "De Marneffe",
"suffix": ""
},
{
"first": "Bill",
"middle": [],
"last": "Maccartney",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Christopher",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Manning",
"suffix": ""
}
],
"year": 2006,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Marie-Catherine de Marneffe, Bill MacCartney and Christopher D. Manning. 2006. Generating typed dependency parses from phrase structure parses. LREC '06.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Fixed Expressions and Idioms in English",
"authors": [
{
"first": "Rosamund",
"middle": [],
"last": "Moon",
"suffix": ""
}
],
"year": 1998,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Rosamund Moon. 1998. Fixed Expressions and Id- ioms in English. Oxford: Clarendon Press.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Machines using sequential minimal optimization",
"authors": [
{
"first": "John",
"middle": [],
"last": "Platt",
"suffix": ""
}
],
"year": 1998,
"venue": "Advances in Kernel Methods -Support Vector Learning",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "John Platt. 1998. Machines using sequential minimal optimization. In In B. Schoelkopf and C. Burges and A. Smola, editors, Advances in Kernel Methods - Support Vector Learning.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "A Constructional Approach to Idioms and Word Formation",
"authors": [
{
"first": "Susanne",
"middle": [
"Z"
],
"last": "Riehemann",
"suffix": ""
}
],
"year": 2001,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Susanne Z. Riehemann. 2001. A Constructional Ap- proach to Idioms and Word Formation. Ph.D. Thesis. Stanford.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Unsupervised recognition of literal and non-literal use of idiomatic expressions",
"authors": [
{
"first": "Caroline",
"middle": [],
"last": "Sporleder",
"suffix": ""
},
{
"first": "Linlin",
"middle": [],
"last": "Li",
"suffix": ""
}
],
"year": 2009,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Caroline Sporleder and Linlin Li. 2009. Unsupervised recognition of literal and non-literal use of idiomatic expressions. EACL '09.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Benjamin Waldron, and Fabre Lambeau",
"authors": [
{
"first": "Aline",
"middle": [],
"last": "Villavicencio",
"suffix": ""
},
{
"first": "Ann",
"middle": [],
"last": "Copestake",
"suffix": ""
}
],
"year": 2004,
"venue": "Lexical encoding of MWE. ACL '04 Workshop on Multiword Expressions",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Aline Villavicencio, Ann Copestake, Benjamin Wal- dron, and Fabre Lambeau. 2004. Lexical encoding of MWE. ACL '04 Workshop on Multiword Expres- sions.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "Collocation extraction based on modifiability statistics",
"authors": [
{
"first": "Joachim",
"middle": [],
"last": "Wermter",
"suffix": ""
},
{
"first": "Udo",
"middle": [],
"last": "Hahn",
"suffix": ""
}
],
"year": 2004,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Joachim Wermter and Udo Hahn. 2004. Collocation extraction based on modifiability statistics. COL- ING '04.",
"links": null
},
"BIBREF23": {
"ref_id": "b23",
"title": "You can't beat frequency (unless you use linguistic knowledge) -a qualitative evaluation of association measures for collocation and term extraction",
"authors": [
{
"first": "Joachim",
"middle": [],
"last": "Wermter",
"suffix": ""
},
{
"first": "Udo",
"middle": [],
"last": "Hahn",
"suffix": ""
}
],
"year": 2006,
"venue": "COLING-ACL '06",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Joachim Wermter and Udo Hahn. 2006. You can't beat frequency (unless you use linguistic knowledge) - a qualitative evaluation of association measures for collocation and term extraction. COLING-ACL '06.",
"links": null
},
"BIBREF24": {
"ref_id": "b24",
"title": "Witten amd Eibe Frank",
"authors": [
{
"first": "H",
"middle": [],
"last": "Ian",
"suffix": ""
}
],
"year": 2000,
"venue": "Data Mining: Practical Machine Learning Tools and Techniques with Java Implementations",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ian H. Witten amd Eibe Frank. 2000. Data Mining: Practical Machine Learning Tools and Techniques with Java Implementations. Morgan Kaufmann.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"num": null,
"type_str": "figure",
"uris": null,
"text": "Baseline methods. Two baseline methods are used to test the intuitive notion that simple rulebased methods are sufficient for MWE identification as well as for comparison with the supervised learning methods."
},
"TABREF2": {
"type_str": "table",
"html": null,
"text": "",
"num": null,
"content": "<table><tr><td>: Development set: Average performance over all</td></tr><tr><td>MWEs and best 8. Supervised classifiers outperform base-</td></tr><tr><td>lines. A: Accuracy; P: Positive Precision; R: Positive Recall;</td></tr><tr><td>F: F-Score.</td></tr></table>"
}
}
}
} |