File size: 106,480 Bytes
6fa4bc9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 | {
"paper_id": "C12-1012",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T13:23:44.304907Z"
},
"title": "Deriving Paraphrases for Highly Inflected Languages from Comparable Documents",
"authors": [
{
"first": "Kfir",
"middle": [],
"last": "Bar",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Tel Aviv University",
"location": {
"settlement": "Ramat Aviv",
"country": "Israel"
}
},
"email": "kfirbar@post.tau.ac.il"
},
{
"first": "Nachum",
"middle": [],
"last": "Dershowitz",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Tel Aviv University",
"location": {
"settlement": "Ramat Aviv",
"country": "Israel"
}
},
"email": "nachumd@tau.ac.il"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "We describe an automatic paraphrase-inference procedure for a highly inflected language like Arabic. Paraphrases are derived from comparable documents, that is, distinct documents dealing with the same topic. A co-training approach is taken, with two classifiers, one designed to model the contexts surrounding occurrences of paraphrases, and the other trained to identify significant features of the words within paraphrases. In particular, we use morpho-syntactic features calculated for both classifiers, as is to be expected when working with highly inflected languages. We provide some experimental results for Arabic, and for the simpler English, which we find to be encouraging. Our immediate interest is to incorporate such paraphrases within an Arabic-to-English translation system.",
"pdf_parse": {
"paper_id": "C12-1012",
"_pdf_hash": "",
"abstract": [
{
"text": "We describe an automatic paraphrase-inference procedure for a highly inflected language like Arabic. Paraphrases are derived from comparable documents, that is, distinct documents dealing with the same topic. A co-training approach is taken, with two classifiers, one designed to model the contexts surrounding occurrences of paraphrases, and the other trained to identify significant features of the words within paraphrases. In particular, we use morpho-syntactic features calculated for both classifiers, as is to be expected when working with highly inflected languages. We provide some experimental results for Arabic, and for the simpler English, which we find to be encouraging. Our immediate interest is to incorporate such paraphrases within an Arabic-to-English translation system.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Paraphrases are pairs of sequences of words, both in the same language, that have the same meaning in at least some contexts. Given a text, \"paraphrasing\" is the act of generating an alternate sequence of words that conveys the same meaning. Since the meaning of a text is determined only when its context is given, paraphrases are sometimes referred to as \"dynamic translations\" or \"semantic equivalents\". Identifying paraphrases is an important capability for many natural language processing applications, including machine translation, as a possible workaround for the problem of limited coverage inherent in a corpus-based translation approach (Callison-Burch et al., 2006; Marton et el., 2009) . Other applications of paraphrasing include automatic evaluation of summaries (Zhao et al., 2008) and question answering (Duboue and Chu-Carroll, 2006; Riezler et al., 2007) .",
"cite_spans": [
{
"start": 649,
"end": 678,
"text": "(Callison-Burch et al., 2006;",
"ref_id": "BIBREF8"
},
{
"start": 679,
"end": 699,
"text": "Marton et el., 2009)",
"ref_id": null
},
{
"start": 779,
"end": 798,
"text": "(Zhao et al., 2008)",
"ref_id": "BIBREF31"
},
{
"start": 822,
"end": 852,
"text": "(Duboue and Chu-Carroll, 2006;",
"ref_id": "BIBREF15"
},
{
"start": 853,
"end": 874,
"text": "Riezler et al., 2007)",
"ref_id": "BIBREF25"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "There are two main directions of work on paraphrases that one can find in this field: investigating an automatic approach for uncovering paraphrases in a given corpus and using paraphrases to improve the performance of a specific task. In this paper, we introduce a novel method for extracting Arabic paraphrases from a corpus of comparable documents as part of our work on improving Arabic-to-English machine translation. Comparable documents are ones that deal with the same topic, such as two newspaper reports of the same event. The extraction technique is based on a learning method, known as \"co-training\" (Blum and Mitchel, 1998) and inspired by the work of Barzilay and McKeown (2001) for finding paraphrases in a parallel monolingual English corpus. In order to validate our technique, we have applied it also to a similar English corpus.",
"cite_spans": [
{
"start": 612,
"end": 636,
"text": "(Blum and Mitchel, 1998)",
"ref_id": null
},
{
"start": 665,
"end": 692,
"text": "Barzilay and McKeown (2001)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Like many other Semitic languages, Arabic is highly inflected; therefore, data sparseness becomes even more noticeable than in English and extracting paraphrases from a corpus turns out to be even more complicated. Arabic words are derived from a root and a pattern (template), combined with prefixes, suffixes and circumfixes. Using the same root with different patterns may yield words with different meanings. Words are inflected for person, number and gender; prefixes and suffixes are then added to indicate definiteness, conjunction, various prepositions and possessive forms. We will list some of the morpho-syntactic features we use for identifying paraphrases in the corpus. Based on the definition, paraphrases are identified as part of the context they are mentioned in within the corpus. Paraphrase is in fact only one of the semantic relations that can be identified to hold between two word sequences in their contexts; it can be seen as a special case of textual entailment (Dagan and Glickman, 2004) , where each sequence entails the other.",
"cite_spans": [
{
"start": 989,
"end": 1015,
"text": "(Dagan and Glickman, 2004)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "There are several existing approaches for inferring paraphrases from a corpus, which differ from one another in the type of corpus they employ. Some require bilingual parallel corpora (Callison-Burch et al., 2006; Zhao et el., 2008) , some need monolingual parallel corpora (Barzilai and McKeown, 2001) , some need general monolingual corpora (Marton et al., 2009) and others need corpora of comparable documents (Rui and Callison-Burch, 2011; Dolan et al., 2004) . Bilingual parallel corpora, pairing Arabic with languages other than English, are very hard to obtain. In this paper we take the last approach, leaving the other directions for future investigation.",
"cite_spans": [
{
"start": 184,
"end": 213,
"text": "(Callison-Burch et al., 2006;",
"ref_id": "BIBREF8"
},
{
"start": 214,
"end": 232,
"text": "Zhao et el., 2008)",
"ref_id": null
},
{
"start": 274,
"end": 302,
"text": "(Barzilai and McKeown, 2001)",
"ref_id": null
},
{
"start": 343,
"end": 364,
"text": "(Marton et al., 2009)",
"ref_id": "BIBREF21"
},
{
"start": 413,
"end": 443,
"text": "(Rui and Callison-Burch, 2011;",
"ref_id": null
},
{
"start": 444,
"end": 463,
"text": "Dolan et al., 2004)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Section 2 cites some related work. Our proposal is described in Section 3 with some experimental results reported in Section 4. Conclusions are given in the last section.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "To the best of our knowledge, ours is the first work on paraphrasing in Arabic. In our previous work (Bar and Dershowitz, 2010) , we extracted Arabic synonyms, that is, single-word paraphrases, using the English glosses provided by SAMA (Maamouri, 2010) , along with WordNet for English (Fellbaum, 1998) . The inferred synonyms were used to improve a corpusbased translation system. Salloum and Habash (2011) developed a rule-based algorithm for generating Modern Standard Arabic (MSA) paraphrases for dialectical Arabic phrases given to a statistics-based automatic translation system. They focused only on input phrases that do not exist in the translation table used by the translation system, for the purpose of improving its coverage. The MSA paraphrases were generated mostly using different morphological variations of the input words. They reported a slight improvement in BLEU score (Papineni, 2002 ) over a baseline system that does not use their generated paraphrases. In another work, by Denkowski et al. (2010) , 726 Arabic paraphrases were manually generated and confirmed using Amazon's Mechanical Turk, from the NIST OpenMT 2002 development set (Garofolo, 2002) . That was mainly done with the purpose of improving the evaluation of an English-to-Arabic machine translation system.",
"cite_spans": [
{
"start": 101,
"end": 127,
"text": "(Bar and Dershowitz, 2010)",
"ref_id": "BIBREF1"
},
{
"start": 237,
"end": 253,
"text": "(Maamouri, 2010)",
"ref_id": "BIBREF20"
},
{
"start": 287,
"end": 303,
"text": "(Fellbaum, 1998)",
"ref_id": "BIBREF16"
},
{
"start": 383,
"end": 408,
"text": "Salloum and Habash (2011)",
"ref_id": "BIBREF27"
},
{
"start": 892,
"end": 907,
"text": "(Papineni, 2002",
"ref_id": "BIBREF22"
},
{
"start": 1000,
"end": 1023,
"text": "Denkowski et al. (2010)",
"ref_id": "BIBREF11"
},
{
"start": 1161,
"end": 1177,
"text": "(Garofolo, 2002)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related work",
"sec_num": null
},
{
"text": "There are also related works in other languages. We only mention a few. Marton et al. (2009) found paraphrases to improve Spanish-to-English and English-to-Chinese statistical machine translation (SMT). For each phrase (as defined in SMT) that was left without a translation, they looked for it in a monolingual corpus and recorded the contexts in which it appeared. They modeled the contexts using a vector that captured phrase occurrences with their context words, and searched for other phrases with the most similar vector of occurrences to improve the translation. Callison-Burch et al. (2006) measured the effect of using paraphrases on Spanish-to-English and French-to-English SMT. They reported a significant improvement in coverage and in the final translation. The paraphrases were automatically extracted following the technique developed by Bannard and Callison-Burch (2005) , using several parallel corpora of French and Spanish paired with other languages. This method is usually referred to as \"pivoting\". Both of these works claimed improvements in the translations. Callison-Burch (2008) and Zhao et al. (2008) developed this approach further by adding syntactic constraints to the extraction algorithms. In a recent work by Wang and Callison-Burch (2011) , English paraphrases were found in a corpus of comparable documents. Similar to what we have done, they started with a large English corpus to find comparable documents. Those documents were used to find comparable sentences from which they extracted sub-sentential comparable fragments, that is, paraphrases. They used a chunker for finding linguistically safe boundaries for the fragments they extracted, and matched fragments based on the n-gram alignment method.",
"cite_spans": [
{
"start": 72,
"end": 92,
"text": "Marton et al. (2009)",
"ref_id": "BIBREF21"
},
{
"start": 570,
"end": 598,
"text": "Callison-Burch et al. (2006)",
"ref_id": "BIBREF8"
},
{
"start": 853,
"end": 886,
"text": "Bannard and Callison-Burch (2005)",
"ref_id": "BIBREF0"
},
{
"start": 1083,
"end": 1104,
"text": "Callison-Burch (2008)",
"ref_id": "BIBREF7"
},
{
"start": 1109,
"end": 1127,
"text": "Zhao et al. (2008)",
"ref_id": "BIBREF31"
},
{
"start": 1242,
"end": 1272,
"text": "Wang and Callison-Burch (2011)",
"ref_id": "BIBREF30"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related work",
"sec_num": null
},
{
"text": "The most inspiring work for us is the one by Barzilay and McKeown (2001) , in which paraphrases are extracted from a corpus containing multiple English translations of the same source. Using this type of corpus allow them to mark initial aligned anchors, chosen based on the results of an alignment algorithm, and to train a classifier to identify the best context environments surrounding potential paraphrases. Based on the resulting contexts, another classifier was trained for finding new paraphrases. This \"co-training\" process was repeated until no new paraphrases were extracted. In our work, we follow the same idea, implemented on Arabic. Since there is no monolingual parallel corpus available for Arabic, we created a corpus of comparable documents and used it as a resource for paraphrasing. Considering that Arabic is a morphologically rich language, we incorporated morphological features of the surrounding words as well as the paraphrase patterns themselves.",
"cite_spans": [
{
"start": 45,
"end": 72,
"text": "Barzilay and McKeown (2001)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related work",
"sec_num": null
},
{
"text": "Inferring paraphrases",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "3",
"sec_num": null
},
{
"text": "As just mentioned, our approach to inferring paraphrases is based on the work of Barzilay and McKeown (2001) on finding paraphrases in different English translations of the same source text.",
"cite_spans": [
{
"start": 81,
"end": 108,
"text": "Barzilay and McKeown (2001)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Preparing the corpus",
"sec_num": "3.1"
},
{
"text": "Understanding how powerful such a resource can be for paraphrasing, but finding no such resource for Arabic, we built a corpus of comparable documents, that is, distinct documents dealing with the same topic or event. This corpus was extracted from the Arabic Gigaword 4.0 (Parker, 2009) , which contains newswire documents published by several news agencies, grouped by their publication date. Pairing documents, based on their topic, was done automatically using cosine similarity over the lemma-frequency vector of every document, with the lemma of every word extracted using MADA (Habash and Rambow, 2005; Roth et al., 2008) . We considered candidates for document pairs only when they were published by different news agencies on the same day. For every document published by one agency, we pair it with a document from the agency that maximizes the similarity score over all the other documents published by the same agency on the same day. Not only that, we require that the score be higher than a predefined threshold that was set, in our experiment settings, to make sure that every candidate pair is composed of two documents sharing at least one third of the largest one. We also tried using lower thresholds for which we retrieved additional pairs; however, precision decreased linearly. It is obvious, then, that this approach prefers precision to recall; in other words, we probably miss a large number of potential candidates, while the candidates that we do extract are likely correct.",
"cite_spans": [
{
"start": 273,
"end": 287,
"text": "(Parker, 2009)",
"ref_id": "BIBREF23"
},
{
"start": 584,
"end": 609,
"text": "(Habash and Rambow, 2005;",
"ref_id": "BIBREF18"
},
{
"start": 610,
"end": 628,
"text": "Roth et al., 2008)",
"ref_id": "BIBREF26"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Preparing the corpus",
"sec_num": "3.1"
},
{
"text": "All together, we created 690 document pairs, comprising about half a million words. Our corpus of comparable documents was manually evaluated by two Arabic speakers. We randomly selected 120 document pairs out of the 690 and, for each, asked the evaluators for a simple \"yes\" or \"no\" answer to the question, \"Do both documents discuss the same event?\" The results are encouraging: out of the 120 pairs, 100 were classified as correct by both evaluators. Of the other 20 instances, 5 were classified \"yes\" by one evaluator. The rest of the pairs actually dealt with the same general domain but were not specifically discussing the same event. This positive evaluation allowed us to use this corpus in the next step of our inference technique.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preparing the corpus",
"sec_num": "3.1"
},
{
"text": "Every document was pre-processed with AMIRAN before being given to the inference classifier, described in the next section. AMIRAN, an updated version of the AMIRA tools (Diab et al., 2004 (Diab et al., , 2007 , is a tool for finding the context-sensitive morpho-syntactic information. AMIRAN combines AMIRA output with morphological analyses provided by SAMA. AMIRAN is also enriched with Named-Entity-Recognition (NER) class tags provided by (Benajiba et al., 2008) . For every word, AMIRAN is capable of identifying the clitics, diacritized lemma, stem, full partof-speech tag (excluding case and mood), base-phrase chunks and NER tags. The corpus is obviously not annotated with paraphrasing-related information and there is no alignment indication included at any level.",
"cite_spans": [
{
"start": 170,
"end": 188,
"text": "(Diab et al., 2004",
"ref_id": "BIBREF13"
},
{
"start": 189,
"end": 209,
"text": "(Diab et al., , 2007",
"ref_id": null
},
{
"start": 444,
"end": 467,
"text": "(Benajiba et al., 2008)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Preparing the corpus",
"sec_num": "3.1"
},
{
"text": "To infer new paraphrases from the corpus, we follow the \"co-training\" technique, training two different classifiers: one for modeling the context of a potential paraphrase and another for modeling the features of the paraphrase pattern itself. The main idea of the co-training approach applied to unlabeled data is to use the two classifiers on different views of the data. In our case, the two views are the context (CX) and the pattern (PT), with one classifier labeling the most reliable unlabeled data items for training the second classifier. Then, the second classifier can label some of the data items for training the first one. This process is repeated several times, and the labeled data collected during the entire run is returned. The algorithm runs in iterations; each iteration increases the number of words a potential paraphrase may contain, that is, in the first iteration only single-word paraphrases are allowed to be found, in the second one, paraphrases composed of up to two words are allowed, and so on. The input of the algorithm is the pairs of documents that we found on the previous section, from which we extract pairs of word sequences. A pair of word sequences is composed of two sequences, one from each of a pair of comparable documents. Since alignment at any level does not exist for comparable documents, we consider all the possible pairs of word sequences, given a pair of documents. To avoid too much noise, we restrict a word sequence for consideration to be composed of at least one nonfunction word and it to not break a base-phrase in the middle, similar to (Wang and Callison-Burch, 2011) . Function words, in our case, are identified based on their part-of-speech and basephrase tags, as provided by AMIRAN. Otherwise, a huge number of pairs containing only function words, not too important for paraphrasing, would be considered. The number of iterations, and concomitantly, the maximum length of the output sequences, is a parameter we control. As implied before, we start with single-word sequences and increase this parameter with every iteration. During the entire run of the algorithm, we maintain two sets of pairs of word sequences:",
"cite_spans": [
{
"start": 1600,
"end": 1631,
"text": "(Wang and Callison-Burch, 2011)",
"ref_id": "BIBREF30"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Inference technique",
"sec_num": "3.2"
},
{
"text": "1. Labeled -containing pairs of word sequences with their label, \"true\" to indicate paraphrases and \"false\" to indicate that the word sequences are not paraphrases of each other. This set starts off empty. 2. Unlabeled -containing pairs of word sequences that are still waiting for their label assignment by the algorithm.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Inference technique",
"sec_num": "3.2"
},
{
"text": "In every iteration, the algorithm performs the following steps:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Inference technique",
"sec_num": "3.2"
},
{
"text": "1. deterministic labeling of potential paraphrases; 2. training the CX classifier using the labeled set as training data; 3. running CX on unlabeled pairs and labeling the most reliable ones; 4. training the PT classifier using the labeled set as training data; 5. running the PT classifier on the labeled set; 6. labeling some unlabeled pairs, based on the labels provided by both classifiers.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Inference technique",
"sec_num": "3.2"
},
{
"text": "We now describe these steps in greater detail.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Inference technique",
"sec_num": "3.2"
},
{
"text": "We cannot estimate in advance the weight of the selected features and their effect on the predictions of the classifiers; therefore, we chose to use support vector machine (SVM) classifiers (Vapnik and Cortes, 1995) because of their good generalization property. Technically, the classifiers are trained on the WEKA platform (Hall et al., 2009) running with the LibSVM library (Chang and Lin, 2011) . One drawback of using SVM in this kind of setting is the long running time of the training algorithm. Because we are running the trainer twice during every iteration, this drawback becomes even more pronounced.",
"cite_spans": [
{
"start": 190,
"end": 215,
"text": "(Vapnik and Cortes, 1995)",
"ref_id": "BIBREF29"
},
{
"start": 325,
"end": 344,
"text": "(Hall et al., 2009)",
"ref_id": "BIBREF19"
},
{
"start": 377,
"end": 398,
"text": "(Chang and Lin, 2011)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Inference technique",
"sec_num": "3.2"
},
{
"text": "The labeled pairs are used as training data for both classifiers, with every pair formatted as a feature vector. The features for the CX classifier capture some morpho-syntactic information expressed by the window-based context words. In the current experiment, we use a window of size three, that is, three words before each word sequence from the pair, and three words afterward. That gives us twelve words from which we extract features for representing a single pair and that number does not change during the entire learning process. Table 1 shows an example for a context. 1 The two main columns represent two Arabic sentences with their corresponding English translations, for easy reading. The emphasized texts are the actual paraphrases while the surrounding words are composing the context, which is described in the last row. In this case, the paraphrase pair is composed of a single identical lemma, inflected differently for person. The context of a paraphrase pair is composed of four parts: left and right words of each of the paired texts. The PT classifier makes its predictions based on the word sequences themselves; their number varies as the iteration number increases. For both classifiers, we use a quadratic kernel for capturing the common effect of all the features on prediction. Tables 2 and 3 summarize the features we currently use for building the feature vectors for the CX and PT classifiers, respectively. NER tags are assigned to persons, organizations, geo-political organizations and locations. The gloss-match rate is calculated for both sides of the context. In the example of Table 1 , there is no word that matches on the left side (note that proper nouns usually do not have glosses). However, on the right side \u202b\ufedf\ufed8\ufe8e\u0621\u202c \u202b\ufecb\ufee5\u0646\u202c \u202b\ufea7\ufe91\ufead\u0631\ufe8d\u0627\u202c xbrA En lqA' (\"a story about a meeting\") matches \u202b\ufecb\ufee5\u0646\u202c \u202b\ufea7\ufe91\ufead\u0631\ufe8d\u0627\u202c \u202b\ufedf\ufed8\ufe8e\ufe8b\ufeea\ufbab\u202c",
"cite_spans": [
{
"start": 579,
"end": 580,
"text": "1",
"ref_id": null
}
],
"ref_spans": [
{
"start": 539,
"end": 546,
"text": "Table 1",
"ref_id": null
},
{
"start": 1306,
"end": 1320,
"text": "Tables 2 and 3",
"ref_id": "TABREF1"
},
{
"start": 1615,
"end": 1622,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Inference technique",
"sec_num": "3.2"
},
{
"text": "xbrA En lqA}h (\"a story about his meeting\") with all three words on the gloss level; therefore, the left gloss-match rate is 0 and the right one is 1. The same calculation works with lemma-match rates on the lemma level. The morphological features we currently use in the PT classifier capture some common Arabic morphological variations. They are all Boolean values indicating whether the word expresses the feature or not. For example, the word \u202b\ufeed\u0648\ufe91\ufedb\ufe97\ufe8e\ufe91\ufeea\ufbab\u202c wbktAbh (\"and in his book\") expresses conjunction, preposition and possessive. When working with Arabic, a highly inflected language, morphological features may contribute to the classification performance. We intend to further explore this direction in the future. The n-gram score is a simple language model score for capturing the co-occurrence of the candidate sequence words. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Inference technique",
"sec_num": "3.2"
},
{
"text": "The number of words in each sequence TABLE 3 -The features we use for training the PT classifier on Arabic.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Sequence length",
"sec_num": null
},
{
"text": "The first time one of the classifiers is trained, it needs some labeled items. With \"co-training\", those items are usually provided by manual annotation of a relatively small fraction of the data or, in this case, by using an automatic deterministic annotation algorithm. Therefore, in the first step of every iteration, the algorithm enriches the labeled set with additional \"true\" labeled pairs following a deterministic approach. Since it is very difficult to obtain a word or sentence-level alignment of two given comparable documents, our algorithm simply adds all the pairs whose word sequences match on the lemma level, word by word. If the lemma does not exist, we use the word's surface form for matching. Lengths of word sequences are determined by the iteration number, so in the first iteration only sequences of size 1 are added, in the second iteration sequences of size 2 are added, and so forth. Such a pair, matched on the lemma level, is shown in Table 1 . Note that paraphrases work on the sense level, rather than on the surface form; however, our assumption is that, because we are using sequences from comparable documents, their senses may be the same with a reasonable high probability. Note that, since we are using the contextsensitive lemmas for matching, one can think of that as matching words on the sense level. However, AMIRAN was trained mostly with morpho-syntactic features and therefore achieves good performance in identifying the common lemma of a context-sensitive part-of-speech tag for every word. When a word may have two or more different lemmas for the same part-of-speech tag that have different senses, AMIRAN does not perform as well. For example, the word \u202b\ufe83\u0623\ufee3\ufe8e\ufee7\ufe94\u202c >mAnp has three different noun lemmas: >amAnap_1 (\"faithfulness\"), >amAnap_2 (\"secretariat\") and >amAnap_3 (\"deposit\").",
"cite_spans": [],
"ref_spans": [
{
"start": 965,
"end": 972,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Sequence length",
"sec_num": null
},
{
"text": "That approach leaves us with some deterministically selected \"true\" examples; however, it does not provide us with the necessary \"false\" examples. In the first iteration, we consider word sequences of size 1 only. Our assumption is that word pairs sharing some of their senses in common may be considered paraphrases, thus cannot be naturally selected as \"false\" examples. Currently we use the English gloss of every word, as provided by AMIRAN, to select word pairs with different gloss values as \"false\" examples. Therefore, under this condition, the Arabic word pair \u202b\ufee3\ufe9f\ufe8e\ufedd\u0644\u202c mjAl and \u202b\ufee3\ufee7\ufec1\u0637\ufed8\ufe94\u202c mnTqp is not considered as a \"false\" example because they share the same gloss value: \"area\". An alternative approach, which we plan to employ in the future, would be using Arabic WordNet (Black et al., 2006) . It implies that, in our first iteration, only word pairs that have the same English gloss and not the same Arabic lemma are put in the unlabeled set. That dramatically reduces the amount of paraphrases of size one, better known as \"synonyms\", that we can find. Since we are more interested in longer paraphrases, we can live with this limitation.",
"cite_spans": [
{
"start": 784,
"end": 804,
"text": "(Black et al., 2006)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Sequence length",
"sec_num": null
},
{
"text": "FIGURE 1 -An overview of the paraphrase inference co-training algorithm.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Sequence length",
"sec_num": null
},
{
"text": "In subsequent iterations, \"false\" examples will be assigned automatically by the classifiers of the previous one, in the following way: after training the CX classifier in step 2, we use the classifier to tag the unlabeled pairs in step 3. Some pairs are assigned with the \"true\" label and some with \"false\". Those for which the classifier has a \"good sense\" are added to the labeled set with their corresponding label. \"Good sense\" is measured with a confidence score that is provided by LibSVM along with every tested pair. Since this score is based on margin length calculations, one should use them carefully. Currently, we only set some threshold values for adding pairs to the labeled set, with a high score, empirically determined. The unlabeled set is also updated with additional examples of length not exceeding the iteration number. In that sense, the iteration number is actually an upper bound on the length of the examples, allowing the algorithm to select sequences of a lower length paired with longer sequences. For example, in the second iteration, the unlabeled set also contains examples that pair a sequence of one word with a sequence of two words. The labeled set after step 3 contains \"true\" as well as \"false\" pairs, added by both the deterministic algorithm and the CT classifier, for training the PT classifier. Steps 4 and 5 train and test the classifier PT on the labeled and unlabeled sets respectively. Finally, in step 6, pairs that receive the same label from both classifiers, with a confidence score higher than the predefined threshold, are added to the labeled set with their corresponding label and stay there forever. This labeled set is used as part of the training data in the next iteration. The number of iterations is manually configured upon initialization of the algorithm and at the end, the \"true\" pairs are deemed paraphrases. The entire process is summarized in Figure 1 .",
"cite_spans": [],
"ref_spans": [
{
"start": 1912,
"end": 1920,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Sequence length",
"sec_num": null
},
{
"text": "To get a feeling for the robustness of the methodology, we applied the same technique to the task of generating paraphrases in English. English has shallow morphology as compared with Arabic, on one hand, but, on the other hand, uses more words than Arabic to convey the same meaning. Based on this observation, for English, we changed the settings of the data for using a window of size 4 instead of 3 and removed most of the morphology-related features. The English set of features for the CX and PT classifiers are summarized in Tables 4 and 5, respectively.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Sequence length",
"sec_num": null
},
{
"text": "Lemma, POS, NER, BP of each context word Lemma-match rate The rate of lemma match on each side of the context Comparable documents were extracted using the same technique from a relatively small part of the English Gigaword (5th ed.) (Parker et al., 2011) . We preprocessed the documents using the OpenNLP library. For every word, we determined its part-of-speech, base-phrase and namedentity tags. The lemma of each word was retrieved from WordNet (Fellbaum et al., 1998) by providing it with the surface form and the part-of-speech tag as inferred by OpenNLP. Overall, we found 294 document pairs, containing about 220,000 words. Similar to the evaluation step of the Arabic corpus, we randomly selected 80 document pairs for vetting their correspondence to each other. Out of the selected 80 document pairs, 65 were classified as \"yes\" instances by both evaluators. Of the other 15 instances, 3 were classified as \"yes\" by only one evaluator. As for Arabic, the rest of the pairs were actually dealing with the same general domain but not specifically discussing the same event. The inference algorithm for English worked exactly as described above. Recall that in the first iteration on Arabic, we used a deterministic algorithm for labeling some of the data for training the classifiers for the first time. For Arabic, we used the English gloss values of the Arabic words for finding \"false\" examples; for English, we use WordNet for the same task in such a way that synonyms are not considered as \"false\" examples.",
"cite_spans": [
{
"start": 234,
"end": 255,
"text": "(Parker et al., 2011)",
"ref_id": null
},
{
"start": 449,
"end": 472,
"text": "(Fellbaum et al., 1998)",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Description",
"sec_num": null
},
{
"text": "Results and evaluation",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "4",
"sec_num": null
},
{
"text": "Our initial experiments perform only five iterations on both corpora (Arabic, as well as English), which means that we find paraphrases of no longer than five words. The two classifiers are configured with different thresholds. The confidence score given by LibSVM for every classification is a value between 0 and 1; therefore, we experimented with different threshold values and realized that the best settings in this case are obtained when using 0.85 for \"true\" pairs for the CX classifier and 0.75 for the PT classifier. For the \"false\" pairs, we use 0.75 for both classifiers. Since we noticed that the number of \"false\" pairs is much larger than the number of \"true\" ones in the training data of every iteration, we defined another parameter (currently 6) that limits the factor of \"false\" pairs allowed in the training data with respect to the \"true\" pairs.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiment settings",
"sec_num": "4.1"
},
{
"text": "In the next section, we show some results when running over 240 document pairs in Arabic, containing about 165,000 words, and 40 English document pairs containing about 11,000 words.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiment settings",
"sec_num": "4.1"
},
{
"text": "First, we give some statistics on the results obtained by the inference algorithm on both the Arabic and English corpora, in Tables 6 and 7, In both tables, the initialization row shows the number of \"true\" and \"false\" examples as was labeled by the deterministic algorithm and the size of the unlabeled examples set. In the following rows, the numbers refer to the results of the specific iteration. The numbers of \"true\" and \"false\" pairs reported on every line are the aggregated numbers collected from all previous iterations. Recall that at the beginning of every iteration, a deterministic algorithm adds pairs of word sequences that match on the lemma level, word by word; hence, the number of \"true\" pairs in every line is the sum of the pairs from the previous iterations, the pairs added by the deterministic algorithm for the next iteration and the paraphrase pairs inferred by the current iteration. The third column, unique paraphrase pairs, is merely the number of unique paraphrase pairs inferred during the current iteration. The parenthesized numbers indicate the difference in the quantity of \"true\" pairs from the previous iteration. So, the total number of extracted paraphrases is the number written on the total line in the unique paraphrases column. In Arabic, we found 1,773 paraphrase pairs and in English we found 525. This process can be scaled up for finding more paraphrases. We do not include paraphrases generated after the first iteration because, by definition, they are composed of synonymous words. Recall that, during initialization, the deterministic algorithm adds pairs to the unlabeled set if their paired words are synonyms in English or share the same English gloss, in Arabic. Table 8 shows some statistics for the entire inference process. The raw data corpus size is a rough estimation of the amount of words we had in the corpus at the beginning. Note that currently we did not use the entire Gigaword corpora: in Arabic we used about 30% of the entire set and in English we only used about 10% of the documents. The following column shows the number of comparable document pairs we found using the pairing algorithm described above. Since the pairing algorithm was designed to prefer recall over precision, the number of comparable documents is lower than might be expected considering the relatively large number of words we had in the raw corpus. We expect that this number will grow larger once we improve the pairing algorithm. The next column, number of words used in inference, sums up the number of words of the entire set of comparable document pairs from the previous column. The last column shows the number of paraphrase pairs extracted by the inference algorithm. Comparing the results to the results retrieved by other works is difficult because there is neither a shared task for paraphrase extraction nor common resources for comparison. Therefore, we show some manual evaluations of our results. The evaluation was performed by two Arabic-English speakers by going over the reported paraphrases one by one. For each pair, we assigned one label: P -indicating correct paraphrase, E -indicating unidirectional entailment, R -related (for other semantic relations except antonyms, e.g. San Diego/Los Angeles) and F -wrong (including antonyms). Table 9 and 10 summarizes our preliminary evaluation report on Arabic and English, respectively. The evaluation results reported in both tables are based on the agreement of the two evaluators; in other words, we report here only on pairs that were annotated by both evaluators with the same tag. Note that the first column, length, indicates the number of words of the largest phrase included in the evaluated paraphrase pair. Paraphrase pairs containing a single word in both phrases were not evaluated at all. In the last column, we calculate the precision, considering pairs tagged with P, E and R as positive instances. The last row summarizes the results. In Arabic, 66% of the generated paraphrase pairs are at least considered as semantically related; among them, about 43% are considered real paraphrases. In English, only 63% of the paraphrase pairs are considered related, out of which 30% are real paraphrases. As can be seen from the tables, there is no preferred length for the inference algorithm. We see a slight improvement in the precision of paraphrases up to length three; however, this improvement does not seem significant, considering the relatively small amount of evaluated pairs.",
"cite_spans": [],
"ref_spans": [
{
"start": 125,
"end": 140,
"text": "Tables 6 and 7,",
"ref_id": "TABREF4"
},
{
"start": 1720,
"end": 1727,
"text": "Table 8",
"ref_id": null
},
{
"start": 3304,
"end": 3311,
"text": "Table 9",
"ref_id": "TABREF7"
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "4.2"
},
{
"text": "When we increased the threshold on the confidence that is used by the PT classifier on English to 0.9, the number of paraphrases reported by the inference algorithm decreased to 330 and the average number of similar words in a pair, increased. As a results of that, the overall precision got improved to 72%, calculated over 250 evaluated pairs. These results helped us understand the effect of the PT classifier on performance. The pairs with a high confidence score, as reported by the PT classifier, are most likely to be real paraphrases; however, in most cases, the word sequences of such a pair share more words in common than do other pairs (e.g. \"the U.S. Air Forces\" \u00f3\uf0f3 \"the United States Air Force\"). Table 11 gives some examples of Arabic, as well as English, paraphrase pairs that were extracted by our inference algorithm. The third column is the evaluation score given by one of the evaluators. ",
"cite_spans": [],
"ref_spans": [
{
"start": 711,
"end": 719,
"text": "Table 11",
"ref_id": "TABREF9"
}
],
"eq_spans": [],
"section": "Length",
"sec_num": null
},
{
"text": "The method suggested here has demonstrated its potential for inferring paraphrases from a corpus of comparable documents, using \"co-training\". As we have seen, incorporating morphological features for a highly inflected language, such as Arabic, is very effective. SVM with its generalization property was a natural option for dealing with combination of features that can play an important role for identifying paraphrases. Finding more features that help to properly match the true senses of word sequences is definitely a direction for future investigation. In a similar experiment performed on English, we still saw encouraging results, despite the smaller corpus. In the next stage of research, we plan to scale up the experiments and use more raw data along with an improved document-pairing algorithm for inferring additional paraphrases. We also plan to use those paraphrases within an Arabic-to-English translation system so as to hopefully improve the quality of the translations.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": null
},
{
"text": "We use the Buckwalter transliteration scheme(Buckwalter, 2002) for rendering Arabic script in Romanization throughout this paper.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Paraphrasing with bilingual parallel corpora",
"authors": [
{
"first": "C",
"middle": [],
"last": "Bannard",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Callison-Burch",
"suffix": ""
}
],
"year": 2005,
"venue": "Proceedings of the Annual Meeting of Association of Computational Linguistics (ACL 2005)",
"volume": "",
"issue": "",
"pages": "597--604",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Bannard, C. and Callison-Burch, C. (2005). Paraphrasing with bilingual parallel corpora. In Proceedings of the Annual Meeting of Association of Computational Linguistics (ACL 2005), pages 597-604, Ann Arbor, MI.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Using synonyms for Arabic-to-English example-based translation",
"authors": [
{
"first": "K",
"middle": [],
"last": "Bar",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Dershowitz",
"suffix": ""
}
],
"year": 2010,
"venue": "Proceedings of the Ninth Conference of the Association for Machine Translation in the Americas",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Bar, K. and Dershowitz, N. (2010). Using synonyms for Arabic-to-English example-based translation. In Proceedings of the Ninth Conference of the Association for Machine Translation in the Americas (AMTA 9), Denver, CO.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Extracting paraphrases from a parallel corpus",
"authors": [
{
"first": "R",
"middle": [],
"last": "Barzilay",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Mckeown",
"suffix": ""
}
],
"year": 2001,
"venue": "Proceedings of the 39th Annual Meeting of Association of Computational Linguistics (ACL 2001)",
"volume": "",
"issue": "",
"pages": "50--57",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Barzilay, R. and McKeown, K. (2001). Extracting paraphrases from a parallel corpus. In Proceedings of the 39th Annual Meeting of Association of Computational Linguistics (ACL 2001), pages 50-57, Toulouse, France.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Arabic named entity recognition: An SVM-based approach",
"authors": [
{
"first": "Y",
"middle": [],
"last": "Benajiba",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Diab",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Rosso",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of the Arab International Conference on Information Technology (ACIT-2008)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Benajiba, Y., Diab, M., and Rosso, P. (2008). Arabic named entity recognition: An SVM-based approach. In Proceedings of the Arab International Conference on Information Technology (ACIT-2008), Hammamet, Tunisia.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Introducing the Arabic WordNet project",
"authors": [
{
"first": "W",
"middle": [],
"last": "Black",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Elkateb",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Rodriguez",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Alkhalifa",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Vossen",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Pease",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Fellbaum",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of the 3rd Global Wordnet Conference (GWC 2006)",
"volume": "",
"issue": "",
"pages": "295--299",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Black, W., Elkateb, S., Rodriguez, H., Alkhalifa, M., Vossen, P., Pease, A., and Fellbaum, C. (2006). Introducing the Arabic WordNet project. In Proceedings of the 3rd Global Wordnet Conference (GWC 2006), pages 295-299, Jeju Island, South Korea.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Combining labeled and unlabeled data with co-training",
"authors": [
{
"first": "A",
"middle": [],
"last": "Blum",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Mitchell",
"suffix": ""
}
],
"year": 1998,
"venue": "Proceedings of 11th Annual Conference on Computational Learning Theory (COLT)",
"volume": "",
"issue": "",
"pages": "92--100",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Blum, A. and Mitchell, T. (1998). Combining labeled and unlabeled data with co-training. In Proceedings of 11th Annual Conference on Computational Learning Theory (COLT), pages 92- 100, Madison, WI.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Buckwalter Arabic morphological analyzer Version 1.0. LDC catalog number LDC2002L49",
"authors": [
{
"first": "T",
"middle": [],
"last": "Buckwalter",
"suffix": ""
}
],
"year": 2002,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Buckwalter, T. (2002). Buckwalter Arabic morphological analyzer Version 1.0. LDC catalog number LDC2002L49.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Syntactic constraints on paraphrases extracted from parallel corpora",
"authors": [
{
"first": "C",
"middle": [],
"last": "Callison-Burch",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of the Conference for Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "196--205",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Callison-Burch, C. (2008). Syntactic constraints on paraphrases extracted from parallel corpora. In Proceedings of the Conference for Empirical Methods in Natural Language Processing (EMNLP 2008), pages 196-205, Honolulu, HI.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Improved statistical machine translation using paraphrases",
"authors": [
{
"first": "C",
"middle": [],
"last": "Callison-Burch",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Koehn",
"suffix": ""
},
{
"first": "Osborne",
"middle": [],
"last": "",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of the North American Association for Computational Linguistics (NAACL 2006)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Callison-Burch, C., Koehn, P., and Osborne, M. (2006). Improved statistical machine translation using paraphrases. In Proceedings of the North American Association for Computational Linguistics (NAACL 2006), New York City, NY.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "LIBSVM: A library for support vector machines",
"authors": [
{
"first": "C",
"middle": [
"C"
],
"last": "Chang",
"suffix": ""
},
{
"first": "C",
"middle": [
"J"
],
"last": "Lin",
"suffix": ""
}
],
"year": 2011,
"venue": "ACM Transactions on Intelligent Systems and Technology",
"volume": "2",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Chang, C. C. and Lin, C. J. (2011). LIBSVM: A library for support vector machines. ACM Transactions on Intelligent Systems and Technology, 2:27:1-27:27. Software available at http://www.csie.ntu.edu.tw/~cjlin/libsvm.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Probabilistic textual entailment: Generic applied modeling of language variability",
"authors": [
{
"first": "I",
"middle": [],
"last": "Dagan",
"suffix": ""
},
{
"first": "O",
"middle": [],
"last": "Glickman",
"suffix": ""
}
],
"year": 2004,
"venue": "PASCAL workshop on Learning Methods for Text Understanding and Mining",
"volume": "",
"issue": "",
"pages": "26--29",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Dagan, I. and Glickman, O. (2004). Probabilistic textual entailment: Generic applied modeling of language variability. In PASCAL workshop on Learning Methods for Text Understanding and Mining, pages 26-29, Grenoble, France.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Turker-assisted paraphrasing for English-Arabic machine translation",
"authors": [
{
"first": "M",
"middle": [],
"last": "Denkowski",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Al-Haj",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Lavie",
"suffix": ""
}
],
"year": 2010,
"venue": "Proceedings of the Workshop on Creating Speech and Language Data with Amazon's Mechanical Turk at the Conference of Human Language Technology-North American Association for Computational Linguistics (HLT-NAACL 2010)",
"volume": "",
"issue": "",
"pages": "66--70",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Denkowski, M., Al-Haj, H., and Lavie, A. (2010). Turker-assisted paraphrasing for English- Arabic machine translation. In Proceedings of the Workshop on Creating Speech and Language Data with Amazon's Mechanical Turk at the Conference of Human Language Technology- North American Association for Computational Linguistics (HLT-NAACL 2010), pages 66-70, Los Angeles, CA.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Second generation tools (AMIRA 2.0): Fast and robust tokenization, POS tagging, and base phrase chunking",
"authors": [
{
"first": "M",
"middle": [],
"last": "Diab",
"suffix": ""
}
],
"year": 2009,
"venue": "MEDAR 2nd International Conference on Arabic Language Resources and Tools",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Diab, M. (2009). Second generation tools (AMIRA 2.0): Fast and robust tokenization, POS tagging, and base phrase chunking. MEDAR 2nd International Conference on Arabic Language Resources and Tools, Cairo, Egypt.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Automatic tagging of Arabic text: From raw text to base phrase chunks",
"authors": [
{
"first": "M",
"middle": [],
"last": "Diab",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Hacioglu",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Jurafsky",
"suffix": ""
}
],
"year": 2004,
"venue": "Proceedings of Human Language Technology-North American Association for Computational Linguistics (HLT-NAACL 2004)",
"volume": "",
"issue": "",
"pages": "149--152",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Diab, M., Hacioglu, K., and Jurafsky, D. (2004). Automatic tagging of Arabic text: From raw text to base phrase chunks. In Proceedings of Human Language Technology-North American Association for Computational Linguistics (HLT-NAACL 2004), pages 149-152, Boston, MA.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Unsupervised construction of large paraphrase corpora: Exploiting massively parallel news sources",
"authors": [
{
"first": "B",
"middle": [],
"last": "Dolan",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Quirk",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Brockett",
"suffix": ""
}
],
"year": 2004,
"venue": "Proceedings of the 20th International Conference on Computational Linguistics",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Dolan, B., Quirk, C., and Brockett, C. (2004). Unsupervised construction of large paraphrase corpora: Exploiting massively parallel news sources. In Proceedings of the 20th International Conference on Computational Linguistics (COLING 2004), Geneva, Switzerland.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Answering the question you wish they had asked: The impact of paraphrasing for question answering",
"authors": [
{
"first": "P",
"middle": [
"A"
],
"last": "Duboue",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Chu-Carroll",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of Human Language Technology-North American Association for Computational Linguistics (HLT-NAACL 2006)",
"volume": "",
"issue": "",
"pages": "33--36",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Duboue, P. A. and Chu-Carroll, J. (2006). Answering the question you wish they had asked: The impact of paraphrasing for question answering. In Proceedings of Human Language Technology-North American Association for Computational Linguistics (HLT-NAACL 2006), pages 33-36, New York City, NY.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "WordNet: An Electronic Lexical Database",
"authors": [
{
"first": "C",
"middle": [],
"last": "Fellbaum",
"suffix": ""
}
],
"year": 1998,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Fellbaum, C. (1998). WordNet: An Electronic Lexical Database. MIT Press, Cambridge, MA.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Arabic tokenization, morphological analysis, and part-ofspeech tagging in one fell swoop",
"authors": [
{
"first": "N",
"middle": [],
"last": "Habash",
"suffix": ""
},
{
"first": "O",
"middle": [],
"last": "Rambow",
"suffix": ""
}
],
"year": 2005,
"venue": "Proceedings of the Conference of American Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "578--580",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Habash, N. and Rambow, O. (2005). Arabic tokenization, morphological analysis, and part-of- speech tagging in one fell swoop. In Proceedings of the Conference of American Association for Computational Linguistics, pages 578-580, Ann Arbor, MI.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "The WEKA data mining software: An update. SIGKDD Explorations",
"authors": [
{
"first": "M",
"middle": [],
"last": "Hall",
"suffix": ""
},
{
"first": "E",
"middle": [],
"last": "Frank",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Holmes",
"suffix": ""
},
{
"first": "B",
"middle": [],
"last": "Pfahringer",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Reutemann",
"suffix": ""
},
{
"first": "I",
"middle": [
"H"
],
"last": "Witten",
"suffix": ""
}
],
"year": 2009,
"venue": "",
"volume": "11",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann, P., and Witten, I. H. (2009). The WEKA data mining software: An update. SIGKDD Explorations, volume 11, issue 1.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Standard Arabic morphological analyzer (SAMA), Version 3.1. Linguistic Data Consortium",
"authors": [
{
"first": "M",
"middle": [],
"last": "Maamouri",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Graff",
"suffix": ""
},
{
"first": "B",
"middle": [],
"last": "Bouziri",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Krouna",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Bies",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Kulick",
"suffix": ""
}
],
"year": 2010,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Maamouri, M., Graff, D., Bouziri, B., Krouna, S., Bies, A., and Kulick, S. (2010). Standard Arabic morphological analyzer (SAMA), Version 3.1. Linguistic Data Consortium, Philadelphia, PA.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Improved statistical machine translation using monolingually-derived paraphrases",
"authors": [
{
"first": "Y",
"middle": [],
"last": "Marton",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Callison-Burch",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Resnik",
"suffix": ""
}
],
"year": 2009,
"venue": "Proceedings of the Conference for Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Marton, Y., Callison-Burch, C., and Resnik, P. (2009). Improved statistical machine translation using monolingually-derived paraphrases. In Proceedings of the Conference for Empirical Methods in Natural Language Processing (EMNLP 2009), Singapore.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "Bleu: A method for automatic evaluation of machine translation",
"authors": [
{
"first": "K",
"middle": [],
"last": "Papineni",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Roukos",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Ward",
"suffix": ""
},
{
"first": "W",
"middle": [
"J"
],
"last": "Zhu",
"suffix": ""
}
],
"year": 2002,
"venue": "Proceedings of the ACL 40th Annual Meeting",
"volume": "",
"issue": "",
"pages": "311--318",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Papineni, K., Roukos, S., Ward, T., and Zhu, W. J. (2002). Bleu: A method for automatic evaluation of machine translation. In Proceedings of the ACL 40th Annual Meeting, pages 311- 318, Philadelphia, PA.",
"links": null
},
"BIBREF23": {
"ref_id": "b23",
"title": "Arabic Gigaword, Fourth Edition. Linguistic Data Consortium",
"authors": [
{
"first": "R",
"middle": [],
"last": "Parker",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Graff",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Kong",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Maeda",
"suffix": ""
}
],
"year": 2009,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Parker, R., Graff, D., Chen, K., Kong, J., and Maeda, K. (2009). Arabic Gigaword, Fourth Edition. Linguistic Data Consortium, Philadelphia, PA.",
"links": null
},
"BIBREF25": {
"ref_id": "b25",
"title": "Statistical machine translation for query expansion in answer retrieval",
"authors": [
{
"first": "S",
"middle": [],
"last": "Riezler",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Vasserman",
"suffix": ""
},
{
"first": "I",
"middle": [],
"last": "Tsochantaridis",
"suffix": ""
},
{
"first": "V",
"middle": [],
"last": "Mittal",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Liu",
"suffix": ""
}
],
"year": 2007,
"venue": "Proceedings of Association for Computational Linguistics (ACL 2007)",
"volume": "",
"issue": "",
"pages": "464--471",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Riezler, S., Vasserman, A., Tsochantaridis, I., Mittal, V., and Liu, Y. (2007). Statistical machine translation for query expansion in answer retrieval. In Proceedings of Association for Computational Linguistics (ACL 2007), pages 464-471, Prague, Czech Repulic.",
"links": null
},
"BIBREF26": {
"ref_id": "b26",
"title": "Arabic morphological tagging, diacritization, and lemmatization using lexeme models and feature ranking",
"authors": [
{
"first": "R",
"middle": [],
"last": "Roth",
"suffix": ""
},
{
"first": "O",
"middle": [],
"last": "Rambow",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Habash",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Diab",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Rudin",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of Association for Computational Linguistics (ACL 2008)",
"volume": "",
"issue": "",
"pages": "117--120",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Roth, R., Rambow, O., Habash, N., Diab, M., and Rudin, C. (2008). Arabic morphological tagging, diacritization, and lemmatization using lexeme models and feature ranking. In Proceedings of Association for Computational Linguistics (ACL 2008), pages 117-120, Columbus, Ohio.",
"links": null
},
"BIBREF27": {
"ref_id": "b27",
"title": "Dialectal to standard Arabic paraphrasing to improve Arabic-English statistical machine translation",
"authors": [
{
"first": "W",
"middle": [],
"last": "Salloum",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Habash",
"suffix": ""
}
],
"year": 2011,
"venue": "Proceedings of the Dialects workshop at the Conference for Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Salloum, W. and Habash, N. (2011). Dialectal to standard Arabic paraphrasing to improve Arabic-English statistical machine translation. In Proceedings of the Dialects workshop at the Conference for Empirical Methods in Natural Language Processing (EMNLP 2011).",
"links": null
},
"BIBREF29": {
"ref_id": "b29",
"title": "Support vector networks",
"authors": [
{
"first": "V",
"middle": [],
"last": "Vapnik",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Cortes",
"suffix": ""
}
],
"year": 1995,
"venue": "Machine Learning",
"volume": "20",
"issue": "",
"pages": "273--297",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Vapnik, V. and Cortes, C. (1995). Support vector networks. Machine Learning, vol. 20, pages 273-297.",
"links": null
},
"BIBREF30": {
"ref_id": "b30",
"title": "Paraphrase fragment extraction from monolingual comparable corpora",
"authors": [
{
"first": "R",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Callison-Burch",
"suffix": ""
}
],
"year": 2011,
"venue": "Proceedings of Fourth Workshop on Building and Using Comparable Corpora (BUCC)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Wang, R. and Callison-Burch, C. (2011). Paraphrase fragment extraction from monolingual comparable corpora. In Proceedings of Fourth Workshop on Building and Using Comparable Corpora (BUCC), Istanbul, Turkey.",
"links": null
},
"BIBREF31": {
"ref_id": "b31",
"title": "Pivot approach for extracting paraphrase patterns from bilingual corpora",
"authors": [
{
"first": "S",
"middle": [],
"last": "Zhao",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Li",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of the Annual Meeting of Association of Computational Linguistics (ACL 2008)",
"volume": "",
"issue": "",
"pages": "780--788",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Zhao, S., Wang, H., Liu, T., and Li, S. (2008). Pivot approach for extracting paraphrase patterns from bilingual corpora. In Proceedings of the Annual Meeting of Association of Computational Linguistics (ACL 2008), pages 780-788, Columbus, OH.",
"links": null
}
},
"ref_entries": {
"TABREF1": {
"num": null,
"type_str": "table",
"html": null,
"text": "The features we use for training the CX classifier on Arabic.",
"content": "<table><tr><td/><td>Description</td></tr><tr><td>Lemma, POS, NER, BP</td><td>of each context word</td></tr><tr><td>Gloss-match rate</td><td>The rate of gloss match on each</td></tr><tr><td/><td>side of the context (left and right)</td></tr><tr><td>Lemma-match rate</td><td>The rate of lemma match on each</td></tr><tr><td/><td>side of the context</td></tr><tr><td>Feature</td><td>Description</td></tr><tr><td>n-gram score</td><td>Normalized n-gram frequency</td></tr><tr><td/><td>score for word sequences up to 4</td></tr><tr><td/><td>words (2-4 grams)</td></tr><tr><td>POS, NER, BP</td><td>of each sequence word</td></tr><tr><td>Boolean morphological</td><td>of each sequence word</td></tr><tr><td>features (exists / does not</td><td/></tr><tr><td>exist): Conjunction,</td><td/></tr><tr><td>Possessive, Determiner and</td><td/></tr><tr><td>Prepositions</td><td/></tr></table>"
},
"TABREF2": {
"num": null,
"type_str": "table",
"html": null,
"text": "The features we use for training the CX classifier on English. The features we use for training the PT classifier on English.",
"content": "<table><tr><td>Feature</td><td>Description</td></tr><tr><td>n-gram score</td><td>Normalized n-gram frequency</td></tr><tr><td/><td>score for word sequences up to</td></tr><tr><td/><td>4 words (2-4 grams)</td></tr><tr><td>POS, NER, BP</td><td>of each sequence word</td></tr><tr><td>Possessive form</td><td>of each sequence word</td></tr><tr><td>Sequence length</td><td>The number of words in each</td></tr><tr><td/><td>sequence</td></tr></table>"
},
"TABREF4": {
"num": null,
"type_str": "table",
"html": null,
"text": "Statistics and final results of the inference algorithm running on the Arabic corpus.",
"content": "<table/>"
},
"TABREF5": {
"num": null,
"type_str": "table",
"html": null,
"text": "Statistics and final results of the inference algorithm running on the English corpus.",
"content": "<table><tr><td/><td>\"false\" pairs</td><td>\"true\" pairs</td><td>Unique</td><td>Unlabeled pairs</td></tr><tr><td/><td/><td/><td>paraphrase</td><td/></tr><tr><td/><td/><td/><td>pairs</td><td/></tr><tr><td>Initialization</td><td>876,947</td><td>32,972</td><td/><td>3,597</td></tr><tr><td>After iteration 1</td><td>960,840</td><td>(+868) 33,840</td><td/><td>86,648</td></tr><tr><td>After iteration 2</td><td colspan=\"2\">1,058,970 (+1,633) 35,473</td><td>230</td><td>58,312</td></tr><tr><td>After iteration 3</td><td colspan=\"2\">1,109, 746 (+1,194) 36,667</td><td>177</td><td>21,332</td></tr><tr><td>After iteration 4</td><td>1,127,643</td><td>(+339) 37,006</td><td>94</td><td>6,677</td></tr><tr><td>After iteration 5</td><td>1,128,475</td><td>(+52) 37,058</td><td>24</td><td>1,490</td></tr><tr><td>Total</td><td/><td/><td>525</td><td/></tr></table>"
},
"TABREF7": {
"num": null,
"type_str": "table",
"html": null,
"text": "Manual evaluation summary for Arabic. P: paraphrases, E: unidirectional entailment, R: related, F: wrong, i.e. unrelated or antonyms.",
"content": "<table><tr><td/><td>Evaluated</td><td>P</td><td>E</td><td>R</td><td>F</td><td>Precision</td></tr><tr><td>2</td><td>120</td><td>49</td><td>12</td><td>25</td><td>34</td><td>71%</td></tr><tr><td>3</td><td>95</td><td>45</td><td>10</td><td>11</td><td>31</td><td>69%</td></tr><tr><td>4</td><td>70</td><td>26</td><td>4</td><td>5</td><td>35</td><td>50%</td></tr><tr><td>5</td><td>50</td><td>24</td><td>2</td><td>7</td><td>20</td><td>66%</td></tr><tr><td>Total</td><td>335</td><td>144</td><td>28</td><td>48</td><td>120</td><td>66%</td></tr></table>"
},
"TABREF8": {
"num": null,
"type_str": "table",
"html": null,
"text": "Manual evaluation results for English. P: paraphrases, E: unidirectional entailment, R: related, F: wrong, i.e. unrelated or antonyms.",
"content": "<table><tr><td/><td>Evaluated</td><td>P</td><td>E</td><td>R</td><td>F</td><td>Precision</td></tr><tr><td>2</td><td>120</td><td>23</td><td>11</td><td>37</td><td>49</td><td>59%</td></tr><tr><td>3</td><td>60</td><td>28</td><td>6</td><td>9</td><td>17</td><td>71%</td></tr><tr><td>4</td><td>50</td><td>15</td><td>8</td><td>8</td><td>21</td><td>62%</td></tr><tr><td>5</td><td>25</td><td>8</td><td>5</td><td>2</td><td>10</td><td>60%</td></tr><tr><td>Total</td><td>255</td><td>74</td><td>30</td><td>56</td><td>97</td><td>63%</td></tr></table>"
},
"TABREF9": {
"num": null,
"type_str": "table",
"html": null,
"text": "Example results in both Arabic and English.",
"content": "<table><tr><td/><td>Paraphrase pair</td><td>Evaluation score</td></tr><tr><td>Arabic</td><td>\u202b\ufe8d\u0627\ufedf\ufed4\ufee0\ufeb3\ufec1\u0637\ufef3\ufbfe\ufee7\ufef2\u202c \u202b\ufe8d\u0627\ufedf\ufead\u0631\ufe8b\ufef3\ufbfe\ufeb1\u0633\u202c Alr}ys AlflsTyny (\"the Palestinian</td><td>Related</td></tr><tr><td/><td>president\")</td><td/></tr><tr><td/><td>\u00f3\uf0f3</td><td/></tr><tr><td/><td>\u202b\ufe8d\u0627\ufedf\ufed4\ufee0\ufeb3\ufec1\u0637\ufef3\ufbfe\ufee7\ufef3\ufbfe\ufe94\u202c \u202b\ufe8d\u0627\ufedf\ufeed\u0648\ufec1\u0637\ufee7\ufef3\ufbfe\ufe94\u202c \u202b\ufe8d\u0627\ufedf\ufeb3\ufee0\ufec1\u0637\ufe94\u202c AlslTp AlwTnyp AlflsTynyp (\"the</td><td/></tr><tr><td/><td>Palestinian authority\")</td><td/></tr><tr><td>Arabic</td><td>\u202b\ufe91\ufeed\u0648\ufeb5\u0634\u202c \u202b\ufeed\u0648\ufeed\u0648\ufedb\ufead\u0631\u202c \u202b\ufe9f\ufeed\u0648\ufead\u0631\ufe9d\u062c\u202c jwrj wwkr bw$ (\"George Walker Bush\")</td><td>Paraphrases</td></tr><tr><td/><td>\u00f3\uf0f3</td><td/></tr><tr><td/><td>\u202b\ufe91\ufeed\u0648\ufeb5\u0634\u202c \u202b\ufe9f\ufeed\u0648\ufead\u0631\ufe9d\u062c\u202c jwrj bw$ (\"George Bush\")</td><td/></tr><tr><td>Arabic</td><td>\u202b\ufe8d\u0627\ufedf\ufeb3\ufe8e\ufea9\u062f\ufeb1\u0633\u202c \u202b\ufe8d\u0627\ufedf\ufee3\ufe85\u0624\ufe97\ufee3\ufead\u0631\u202c Alm&tmr AlsAds (\"the Sixth conference\")</td><td>Paraphrases</td></tr><tr><td/><td>\u00f3\uf0f3</td><td/></tr><tr><td/><td>\u202b\ufe8d\u0627\ufedf\ufeb3\ufe8e\ufea9\u062f\ufeb1\u0633\u202c \u202b\ufe8d\u0627\ufedf\ufeed\u0648\ufeaf\u0632\ufe8d\u0627\ufead\u0631\ufef1\u064a\u202c \u202b\ufe8d\u0627\ufefb\ufe9f\ufe97\ufee3\ufe8e\ufec9\u0639\u202c AlAjtmAE AlwzAry AlsAds (\"the Sixth</td><td/></tr><tr><td/><td>ministerial meeting\")</td><td/></tr><tr><td>Arabic</td><td>\u202b\ufe9f\ufefc\ufeaf\u0632\ufead\u0631\u202c \u202b\ufea9\u062f\ufe8d\u0627\ufee7\ufef3\ufbfe\ufef3\ufbfe\ufedd\u0644\u202c dAnyyl jlAzr (\"Daniel Glaser\")</td><td>Paraphrases</td></tr><tr><td/><td>\u00f3\uf0f3</td><td/></tr><tr><td/><td>\u202b\ufecf\ufefc\ufeb3\ufead\u0631\u202c \u202b\ufea9\u062f\ufe8d\u0627\ufee7\ufef3\ufbfe\ufe8b\ufedd\u0644\u202c dAny}l glAsr (\"Daniel Glaser\")</td><td/></tr><tr><td>Arabic</td><td>\u202b\ufeed\u0648\ufe8d\u0627\ufee7\ufe9f\ufeed\u0648\ufedf\ufeed\u0648\u202c \u202b\ufecf\ufeed\u0648\ufee7\ufeaf\u0632\ufe8d\u0627\ufedf\ufef3\ufbfe\ufeaf\u0632\u202c \u202b\ufedb\ufef3\ufbfe\ufee0\ufef2\u202c kyly gwnzAlyz wAngwlw (\"Kaylie Gonzales and Angelo\") \u00f0\uf0f0</td><td>Unidirectional entailment</td></tr><tr><td/><td>\u202b\ufe8d\u0627\ufedf\ufea7\ufec1\u0637\ufef3\ufbfe\ufead\u0631\ufef3\ufbfe\ufee5\u0646\u202c \u202b\ufe8d\u0627\ufefb\ufead\u0631\ufe9f\ufee7\ufe97\ufef3\ufbfe\ufee7\ufef3\ufbfe\ufef3\ufbfe\ufee5\u0646\u202c AlArjntynyyn AlxTyryn (\"the dangerous</td><td/></tr><tr><td/><td>Argentinians\")</td><td/></tr><tr><td>Arabic</td><td>\u202b\ufe8d\u0627\ufedf\ufe9f\ufea9\u062f\ufef3\ufbfe\ufea9\u062f\u202c \u202b\ufe8d\u0627\ufedf\ufe91\ufead\u0631\ufedf\ufee3\ufe8e\ufee5\u0646\u202c AlbrlmAn Aljdyd (\"the new Parliament\")</td><td>Paraphrases</td></tr><tr><td/><td>\u00f3\uf0f3</td><td/></tr><tr><td/><td>\u202b\ufe8d\u0627\ufedf\ufee3\u202c \u202b\ufecb\ufeb7\ufead\u0631\u202c \u202b\ufe8d\u0627\ufedf\ufeb3\ufe8e\ufe91\ufeca\u202c \u202b\ufe8d\u0627\ufedf\ufeed\u0648\ufec1\u0637\ufee7\ufef2\u202c \u202b\ufe9f\ufee0\ufeb1\u0633\u202c Almjls AlwTny AlsAbE E$r (\"the</td><td/></tr><tr><td/><td>Seventeenth Parliament\")</td><td/></tr><tr><td>Arabic</td><td>\u202b\ufe8d\u0627\ufedf\ufee0\ufe91\ufee7\ufe8e\ufee7\ufef3\ufbfe\ufe94\u202c \u202b\ufe8d\u0627\ufedf\ufeb3\ufeed\u0648\ufead\u0631\ufef3\ufbfe\ufe94\u202c \u202b\ufe8d\u0627\ufedf\ufea3\ufea9\u062f\ufeed\u0648\ufea9\u062f\u202c AlHdwd Alswryp AllbnAnyp (\"the</td><td>Unidirectional</td></tr><tr><td/><td>Syrian-Lebanese borders\")</td><td>entailment</td></tr><tr><td/><td>\u00f3\uf0f3</td><td/></tr><tr><td/><td>\u202b\ufe8d\u0627\ufedf\ufeb3\ufeed\u0648\ufead\u0631\ufef3\ufbfe\ufe94\u202c \u202b\ufe8d\u0627\ufedf\ufea3\ufea9\u062f\ufeed\u0648\ufea9\u062f\u202c AlHdwd Alswryp (\"the Syrian border\")</td><td/></tr><tr><td>English</td><td>could veto</td><td>Related</td></tr><tr><td/><td>\u00f3\uf0f3</td><td/></tr><tr><td/><td>threatened to veto</td><td/></tr><tr><td>English</td><td>the U.S. Naval Task Force</td><td>Paraphrases</td></tr><tr><td/><td>\u00f3\uf0f3</td><td/></tr><tr><td/><td>a US Naval Task Group</td><td/></tr></table>"
}
}
}
} |