File size: 83,134 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 | {
"paper_id": "R11-1013",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T15:03:52.448278Z"
},
"title": "Enriching a Statistical Machine Translation System Trained on Small Parallel Corpora with Rule-Based Bilingual Phrases",
"authors": [
{
"first": "V\u00edctor",
"middle": [
"M"
],
"last": "S\u00e1nchez-Cartagena",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Felipe",
"middle": [],
"last": "S\u00e1nchez-Mart\u00ednez",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Juan",
"middle": [],
"last": "Antonio P\u00e9rez-Ortiz",
"suffix": "",
"affiliation": {},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "In this paper, we present a new hybridisation approach consisting of enriching the phrase table of a phrase-based statistical machine translation system with bilingual phrase pairs matching structural transfer rules and dictionary entries from a shallowtransfer rule-based machine translation system. We have tested this approach on different small parallel corpora scenarios, where pure statistical machine translation systems suffer from data sparseness. The results obtained show an improvement in translation quality, specially when translating out-of-domain texts that are well covered by the shallow-transfer rule-based machine translation system we have used.",
"pdf_parse": {
"paper_id": "R11-1013",
"_pdf_hash": "",
"abstract": [
{
"text": "In this paper, we present a new hybridisation approach consisting of enriching the phrase table of a phrase-based statistical machine translation system with bilingual phrase pairs matching structural transfer rules and dictionary entries from a shallowtransfer rule-based machine translation system. We have tested this approach on different small parallel corpora scenarios, where pure statistical machine translation systems suffer from data sparseness. The results obtained show an improvement in translation quality, specially when translating out-of-domain texts that are well covered by the shallow-transfer rule-based machine translation system we have used.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Statistical machine translation (SMT) (Koehn, 2010) is currently the leading paradigm in machine translation research. SMT systems are very attractive because they may be built with little human effort when enough monolingual and bilingual corpora are available. However, bilingual corpora large enough to build competitive SMT systems are not always easy to harvest, and they may not even exist for some language pairs. On the contrary, rule-based machine translation systems (RBMT) may be built without any parallel corpus; however, they need an explicit representation of linguistic information whose coding by human experts requires a considerable amount of time.",
"cite_spans": [
{
"start": 38,
"end": 51,
"text": "(Koehn, 2010)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "When both parallel corpora and linguistic information exist, hybrid approaches (Thurmair, 2009) may be followed in order to make the most of such resources. We focus on alleviating the data sparseness problem suffered by phrase-based statistical machine translation (PBSMT) systems (Koehn, 2010, ch. 5) when trained on small parallel corpora. We present a new hybrid approach which enriches a PBSMT system with resources from shallow-transfer RBMT. Shallow-transfer RBMT systems, which are described in detail below, do not perform a complete syntactic analysis of the input sentences, but rather work with much simpler intermediate representations. Hybridisation between shallow-transfer RBMT and SMT has not yet been explored. Existing hybridisation strategies involve more complex RBMT systems (Eisele et al., 2008) which are usually treated as black boxes; in contrast, our approach directly uses the RBMT dictionaries and rules.",
"cite_spans": [
{
"start": 79,
"end": 95,
"text": "(Thurmair, 2009)",
"ref_id": "BIBREF19"
},
{
"start": 282,
"end": 302,
"text": "(Koehn, 2010, ch. 5)",
"ref_id": null
},
{
"start": 797,
"end": 818,
"text": "(Eisele et al., 2008)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "We provide an exhaustive evaluation of our hybridisation approach with two different language pairs: Breton-French and Spanish-English. While the first one suffers from actual resource scarceness, many different parallel corpora are available for the second one, which allows us to test our approach on different domains and check if it is able to improve the poor performance of PBSMT systems when translating texts from a domain not covered by the bilingual training data.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The rest of the paper is organised as follows. Next section overviews the two systems we combine in our approach. Then, section 3 outlines related hybrid approaches, whereas our approach is described in section 4. Sections 5 and 6 present the experiments conducted and discuss the results achieved, respectively. The paper ends with our conclusions and future research lines.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "PBSMT systems (Koehn, 2010, ch. 5) translate sentences by maximising the translation probability as defined by the log-linear combination of a number of feature functions, whose weights are chosen to optimise translation quality (Och, 2003) . A core component of every PBSMT system is the phrase table, which contains bilingual phrase pairs extracted from a bilingual corpus after word alignment (Och and Ney, 2003) . The set of translations from which the most probable one is chosen is built by segmenting the source sentence in all possible ways and then combining the translation of the different source segments according to the phrase table. Common feature functions are: source-totarget and target-to-source phrase translation probabilities, source-to-target and target-to-source lexical weightings (calculated by using a probabilistic bilingual dictionary), reordering costs, number of words in the output (word penalty), number of phrase pairs used (phrase penalty), and likelihood of the output as given by a target-language model.",
"cite_spans": [
{
"start": 14,
"end": 34,
"text": "(Koehn, 2010, ch. 5)",
"ref_id": null
},
{
"start": 229,
"end": 240,
"text": "(Och, 2003)",
"ref_id": "BIBREF14"
},
{
"start": 396,
"end": 415,
"text": "(Och and Ney, 2003)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-Based Statistical Machine Translation",
"sec_num": "2.1"
},
{
"text": "The RBMT process (Hutchins and Somers, 1992) can be split into three different steps: analysis of the source language (SL) text to build a SL intermediate representation; transfer from that SL intermediate representation to a target language (TL) intermediate representation; and generation of the final translation from the TL intermediate representation.",
"cite_spans": [
{
"start": 17,
"end": 44,
"text": "(Hutchins and Somers, 1992)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Shallow-Transfer Rule-Based Machine Translation",
"sec_num": "2.2"
},
{
"text": "Shallow-transfer RBMT systems use relatively simple intermediate representations, which are based on lexical forms consisting of lemma, part of speech and morphological inflection information of the words in the input sentence, and simple shallow-transfer rules that operate on sequences of lexical forms: this kind of systems do not perform a complete syntactic analysis. Apertium (Forcada et al., 2011) , the shallow-transfer RBMT platform used to evaluate our approach, splits the transfer stage into structural and lexical transfer. The lexical transfer is done by using a bilingual dictionary which, for each SL lexical form, provides a single TL lexical form; thus, no lexical selection is performed. It is worth noting that multi-word expressions, such as on the other hand (which acts as a single adverb), may be analysed to (or generated from) a single lexical form.",
"cite_spans": [
{
"start": 382,
"end": 404,
"text": "(Forcada et al., 2011)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Shallow-Transfer Rule-Based Machine Translation",
"sec_num": "2.2"
},
{
"text": "Structural transfer is done by applying a set of rules in a left-to-right, longest-match fashion to prevent the translation to be performed word for word in those cases in which this would result in an incorrect translation. Structural transfer rules process sequences of lexical forms by performing operations such as reorderings and gender and number agreements. For the translation between non-related language pairs, the structural transfer may be split into three levels in order to facilitate the writing of rules by linguists. The first level performs short-distance operations (such as gender and number agreement between nouns and adjectives) and groups word sequences into chunks; the second one performs inter chunk operations; and the third one generates a sequence of lexical forms from each chunk. Note that, although this multi-stage shallow transfer allows performing operations between words which are distant in the source sentence, shallowtransfer RBMT systems are less powerful that the ones which perform full parsing.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Shallow-Transfer Rule-Based Machine Translation",
"sec_num": "2.2"
},
{
"text": "Bilingual dictionaries are the most reused resource from RBMT. They have been added to SMT systems since its early days (Brown et al., 1993) . One of the simplest strategies, which has already been put into practice with the Apertium bilingual dictionaries (Tyers, 2009) , consists of adding the dictionary entries directly to the parallel corpus. In addition to the obvious increase in lexical coverage, Schwenk et al. (2009) state that the quality of the alignments obtained is also improved when the words in the bilingual dictionary appear in other sentences of the parallel corpus. However, it is not guaranteed that, following this strategy, multi-word expressions from the bilingual dictionary that appear in the SL sentences are translated as such by the SMT decoder because they may be split into smaller units by the phrase-extraction algorithm. Our strategy differs from these approaches in that we ensure the proper translation of multi-word expressions, but also add the dictionary entries to the training corpus with the aim of improving word alignment. Other approaches go beyond adding a dictionary to the parallel corpus: dictionary entries may constrain the decoding process (Langlais, 2002) , or may be used in conjunction with handcrafted rules to reorder the SL sentences to match the structure of the TL (Popovi\u0107 and Ney, 2006) .",
"cite_spans": [
{
"start": 120,
"end": 140,
"text": "(Brown et al., 1993)",
"ref_id": "BIBREF0"
},
{
"start": 257,
"end": 270,
"text": "(Tyers, 2009)",
"ref_id": "BIBREF20"
},
{
"start": 405,
"end": 426,
"text": "Schwenk et al. (2009)",
"ref_id": "BIBREF17"
},
{
"start": 1193,
"end": 1209,
"text": "(Langlais, 2002)",
"ref_id": "BIBREF10"
},
{
"start": 1326,
"end": 1349,
"text": "(Popovi\u0107 and Ney, 2006)",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "3"
},
{
"text": "Although RBMT transfer rules have also been reused in hybrid systems, they have been mostly used implicitly as part of a complete RBMT engine. For instance, Dugast et al. (2008) show how a PBSMT system can be bootstrapped using only monolingual data and an RBMT engine. Another remarkable study (Eisele et al., 2008 ) presents a strategy based on the augmentation of the phrase table to include information provided by an RBMT system. In this approach, the sentences to be translated by the hybrid system are first translated with an RBMT system and then a small phrase table is obtained from the resulting parallel corpus. Phrase pairs are extracted following the usual procedure (Koehn, 2010, sec. 5.2. 3) which generates the set of all possible phrase pairs that are consistent with the word alignments. In order to obtain reliable word alignments, they are computed using an alignment model previously built from a large parallel corpus. Finally, the RBMT-generated phrase table is added to the original one. On the contrary, our approach directly generates phrase pairs which match either an entry in the bilingual dictionary or a structural transfer rule; thus preventing them from being split into smaller phrase pairs even if they would be consistent with the word alignments. In addition, our approach does not require a large parallel corpus from which to learn an alignment model. Preliminary experiments show that our hybrid approach outperforms Eisele et al.'s (2008) strategy when translating from Spanish to English.",
"cite_spans": [
{
"start": 157,
"end": 177,
"text": "Dugast et al. (2008)",
"ref_id": "BIBREF2"
},
{
"start": 295,
"end": 315,
"text": "(Eisele et al., 2008",
"ref_id": "BIBREF3"
},
{
"start": 681,
"end": 704,
"text": "(Koehn, 2010, sec. 5.2.",
"ref_id": null
},
{
"start": 1458,
"end": 1480,
"text": "Eisele et al.'s (2008)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "3"
},
{
"text": "Other strategies involving neither transfer rules nor bilingual dictionaries may alleviate the data sparseness problem in PBSMT. For example, paraphrases may be derived from a SL monolingual corpus (Marton et al., 2009) and verb forms may be substituted by their lemma when translating into highly-inflected languages (de Gispert et al., 2005) .",
"cite_spans": [
{
"start": 198,
"end": 219,
"text": "(Marton et al., 2009)",
"ref_id": "BIBREF12"
},
{
"start": 318,
"end": 343,
"text": "(de Gispert et al., 2005)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "3"
},
{
"text": "Our hybridisation strategy modifies two elements of a standard PBSMT system: the word alignments and the phrase translation model.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Enhancing Phrase-Based SMT With Shallow-Transfer Linguistic Resources",
"sec_num": "4"
},
{
"text": "As improving the quality of the word alignments in a PBSMT system could lead to improvements in translation performance (Lopez and Resnik, 2006) , in our approach we add to the original corpus all the entries, after suitably inflecting them, from the Apertium bilingual dictionary, to help the word aligment process. Recall that some multi-word expressions are encoded as single lexical forms in the Apertium dictionaries; therefore, the entries generated from Apertium may contain multi-word parallel segments. Once word alignments have been computed and the probabilistic bilingual dictionary used to compute the lexical weightings of the phrase pairs has been learned, dictionary entries are ignored and no phrase pair are extracted from them. In contrast to Schwenk et al. (2009) , we avoid extracting phrase pairs which do not preserve the translation of multi-word expressions as such by including the dictionary entries directly in the phrase table, as discussed next.",
"cite_spans": [
{
"start": 120,
"end": 144,
"text": "(Lopez and Resnik, 2006)",
"ref_id": "BIBREF11"
},
{
"start": 762,
"end": 783,
"text": "Schwenk et al. (2009)",
"ref_id": "BIBREF17"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Improving Word Alignment with RBMT Bilingual Dictionaries",
"sec_num": "4.1"
},
{
"text": "As already mentioned, the Apertium structural transfer detects sequences of lexical forms which need to be translated together to prevent them from being translated word for word, which would result in an incorrect translation. Therefore, adding to the phrase table of a PBSMT system all the bilingual phrase pairs which either match one of these sequences of lexical forms in the structural transfer or an entry in the bilingual dictionary ensures that all the linguistic information of Apertium is encoded with the minimum amount of phrase pairs.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Enriching the Phrase Translation Model",
"sec_num": "4.2"
},
{
"text": "Generating a phrase pair from every entry in the bilingual dictionary is straightforward: it only involves the inflection of source and target lexical forms. The generation of phrase pairs from the structural transfer rules is performed by finding sequences of SL words in the sentences to be translated that match a structural transfer rule. Each of these sequences constitute the SL side of a bilingual phrase pair; the corresponding TL phrase is obtained by translating the SL side with Apertium. It is worth noting that the generation of bilingual phrase pairs from the shallow-transfer rules is guided by the test corpus. We decided to do it in this way in order to avoid meaningless phrases and also to make our approach computationally feasible. Consider, for instance, a rule which is triggered every time a determiner followed by a noun and an adjective is detected. Generating phrase pairs from this rule would involve combining all the determiners in the dictionary with all the nouns and all the adjectives, causing the generation of many meaningless phrases, such as el ni\u00f1o inal\u00e1mbricothe wireless boy. In addition, the number of combinations to deal with would become unmanageable as the length of the rule grows.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase Pair Generation",
"sec_num": "4.2.1"
},
{
"text": "State-of-the-art PBSMT systems usually attach 5 scores to every phrase pair in the translation table: source-to-target and target-to-source phrase translation probabilities, source-to-target and target-tosource lexical weightings, and phrase penalty.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Scoring the New Phrase Pairs",
"sec_num": "4.2.2"
},
{
"text": "To calculate the phrase translation probabilities of the new phrase pairs obtained from the shallowtransfer RBMT resources we simply add them once to the list of corpus-extracted phrase pairs, and then compute the probabilities by relative frequency as it is usually done (Koehn, 2010, sec. 5.2.5) . In this regard, it is worth noting that as RBMT-generated phrase pairs are added only once, if one of them happens to share its source side with many other corpus-extracted phrase pairs, or even with a single, very frequent one, the RBMT-generated phrase pair will receive lower scores, which penalises its use. To alleviate this without adding the same phrase pair an arbitrary amount of times, we introduce an additional boolean score to flag phrase pairs obtained from the RBMT resources.",
"cite_spans": [
{
"start": 272,
"end": 297,
"text": "(Koehn, 2010, sec. 5.2.5)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Scoring the New Phrase Pairs",
"sec_num": "4.2.2"
},
{
"text": "To calculate the lexical weightings (Koehn, 2010, sec. 5.3 .3) of the RBMT-generated phrase pairs the alignments between the words in the source side and those in the target side are needed. They are computed by tracing the operations carried out in the different stages of the shallow-transfer RBMT system. Only those words which are neither split nor joint with other words by the RBMT engine are included in the alignments; thus, multi-word expressions are left unaligned. This is done for convenience since, in this way, the number of lexical probabilities to take into account is reduced, and, as a result, phrase pairs containing multi-word expressions receive higher scores.",
"cite_spans": [
{
"start": 36,
"end": 58,
"text": "(Koehn, 2010, sec. 5.3",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Scoring the New Phrase Pairs",
"sec_num": "4.2.2"
},
{
"text": "We evaluated our RBMT-SMT hybridisation approach on two different language pairs, namely Breton-French and Spanish-English, and with different small training corpus sizes. While the Breton-French language pair suffers from actual resource scarceness (there are only around 30 000 parallel sentences available), Spanish-English was chosen because it has a wide range of parallel corpora available, which allows us to perform both in-domain and out-of-domain evaluations.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Settings",
"sec_num": "5"
},
{
"text": "SMT systems for Spanish-English were trained from the Europarl v5 parallel corpus (Koehn, 2005) , collected from the proceedings of the European Parliament. Its whole target side, except for the Q4/2000 portion, was used to train the TL model used in the experiments. We learned the translation model from corpora of different sizes; more precisely, we used fragments of the Europarl corpus consisting of 2 000, 5 000, 10 000, 20 000, 40 000 and 80 000 parallel sentences. The sentences in each training set were randomly chosen (avoiding the Q4/2000 portion) in such a way that larger corpora include the sentences in the smaller ones.",
"cite_spans": [
{
"start": 82,
"end": 95,
"text": "(Koehn, 2005)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Settings",
"sec_num": "5"
},
{
"text": "Regarding Breton-French, the translation model was built using the only freely-available parallel corpus for such language pair (Tyers, 2009) , which contains short sentences from the tourism and computer localisation domains split in different sections for training, tuning and testing. We also used dif- Tyers (2009) 5 000 10k Tyers (2009) 10 000 20k Tyers (2009) 20 000 \u2248 27k Tyers (2009) 26 835 In-domain tuning Tyers (2009) 2 000 In-domain test Tyers (2009) 2 000 ferent training corpora sizes, namely 2 000, 5 000, 10 000, 20 000, and 26 835 parallel sentences, the last one corresponding to the whole training section of the corpus. As in the Spanish-English pair, sentences were randomly chosen and larger corpora include the sentences in the smaller ones. The TL model was learnt from a monolingual corpus built by concatenating the target side of the whole bilingual training corpus and the French monolingual data from the Europarl corpus provided for the WMT 2011 shared translation task. 1 The weights of the different feature functions were optimised by means of minimum error rate training (MERT; Och, 2003) . Breton-French systems were tuned using the tuning section of the parallel corpus by Tyers (2009) and evaluated using the devtest section of the same corpus. Note that we can only perform in-domain evaluation for this language pair.",
"cite_spans": [
{
"start": 128,
"end": 141,
"text": "(Tyers, 2009)",
"ref_id": "BIBREF20"
},
{
"start": 306,
"end": 318,
"text": "Tyers (2009)",
"ref_id": "BIBREF20"
},
{
"start": 329,
"end": 341,
"text": "Tyers (2009)",
"ref_id": "BIBREF20"
},
{
"start": 353,
"end": 365,
"text": "Tyers (2009)",
"ref_id": "BIBREF20"
},
{
"start": 379,
"end": 391,
"text": "Tyers (2009)",
"ref_id": "BIBREF20"
},
{
"start": 416,
"end": 428,
"text": "Tyers (2009)",
"ref_id": "BIBREF20"
},
{
"start": 450,
"end": 462,
"text": "Tyers (2009)",
"ref_id": "BIBREF20"
},
{
"start": 1001,
"end": 1002,
"text": "1",
"ref_id": null
},
{
"start": 1112,
"end": 1122,
"text": "Och, 2003)",
"ref_id": "BIBREF14"
},
{
"start": 1209,
"end": 1221,
"text": "Tyers (2009)",
"ref_id": "BIBREF20"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Settings",
"sec_num": "5"
},
{
"text": "Regarding Spanish-English, we have carried out both in-domain and out-of-domain evaluations. The former was performed by tuning the systems with 2 000 parallel sentences randomly chosen from the Q4/2000 portion of Europarl v5 corpus (Koehn, 2005) and evaluating them with 2 000 random parallel sentences from the same corpus; special care was taken to avoid the overlapping between the test and development sets. The outof-domain evaluation was performed by using the newstest2008 set for tuning and the newstest2010 test for testing; both sets belong to the news domain and are distributed as part of the WMT 2010 shared translation task. 2 Tables 1 and 2 summarise the data about the corpora used in the experiments.",
"cite_spans": [
{
"start": 233,
"end": 246,
"text": "(Koehn, 2005)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Settings",
"sec_num": "5"
},
{
"text": "We used the free/open-source PBSMT system Moses 3 (Koehn et al., 2007) SRILM language modeling toolkit (Stolcke, 2002) , which was used to train a 5-gram language model using interpolated Kneser-Ney discounting (Goodman and Chen, 1998) . Word alignments from the training parallel corpus were computed by means of GIZA++ (Och and Ney, 2003) . The Apertium (Forcada et al., 2011) engine and the linguistic resources for Spanish-English and Breton-French were downloaded from the Apertium Subversion repository. 4 The Apertium linguistic data contains 326 228 entries in the bilingual dictionary, 106 firstlevel rules, 31 second-level rules, and 7 third-level rules for Spanish-English; and 21 593, 169, 79 and 6, respectively, for Breton-French (see section 2.2 for a description of the different rule levels).",
"cite_spans": [
{
"start": 50,
"end": 70,
"text": "(Koehn et al., 2007)",
"ref_id": "BIBREF7"
},
{
"start": 103,
"end": 118,
"text": "(Stolcke, 2002)",
"ref_id": "BIBREF18"
},
{
"start": 211,
"end": 235,
"text": "(Goodman and Chen, 1998)",
"ref_id": "BIBREF5"
},
{
"start": 321,
"end": 340,
"text": "(Och and Ney, 2003)",
"ref_id": "BIBREF13"
},
{
"start": 510,
"end": 511,
"text": "4",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Settings",
"sec_num": "5"
},
{
"text": "We have tested the following configurations:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Settings",
"sec_num": "5"
},
{
"text": "\u2022 a state-of-the-art PBSMT system with the feature functions discussed in section 2.1 (baseline);",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Settings",
"sec_num": "5"
},
{
"text": "\u2022 the Apertium shallow-transfer RBMT engine, from which the dictionaries and transfer rules have been taken (Apertium);",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Settings",
"sec_num": "5"
},
{
"text": "\u2022 the hybridisation approach described along this paper (phrase-rules) and a variation in which only dictionary-matching bilingual phrases are included in the phrase table (phrase-dict); and",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Settings",
"sec_num": "5"
},
{
"text": "\u2022 a reduced version of our approach in which the entries in the bilingual dictionary are only added to the training corpus for the computation of the word alignments and the probabilistic bilingual dictionary, as explained in section 4.1 (alignment).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Settings",
"sec_num": "5"
},
{
"text": "6 Results and Discussion Table 3 reports the translation performance as measured by BLEU (Papineni et al., 2002) for the dif-ferent configurations and language pairs described in section 5. Statistical significance of the differences between systems has been computed by performing 1 000 iterations of paired bootstrap resampling (Zhang et al., 2004) with a p-level of 0.05. In addition, table 4 presents the optimal weight obtained with MERT for the feature function that flags whether a phrase pair has been obtained from the Apertium bilingual resources (dictionaries and rules). Table 5 shows the proportion of RBMTgenerated phrases used to perform each translation. The results show that our hybrid approach outperforms both pure RBMT and PBSMT systems in terms of BLEU. However, the difference is statistically significant only under certain circumstances.",
"cite_spans": [
{
"start": 89,
"end": 112,
"text": "(Papineni et al., 2002)",
"ref_id": "BIBREF15"
},
{
"start": 330,
"end": 350,
"text": "(Zhang et al., 2004)",
"ref_id": "BIBREF21"
}
],
"ref_spans": [
{
"start": 25,
"end": 32,
"text": "Table 3",
"ref_id": null
},
{
"start": 583,
"end": 590,
"text": "Table 5",
"ref_id": null
}
],
"eq_spans": [],
"section": "Experimental Settings",
"sec_num": "5"
},
{
"text": "The in-domain evaluation shows that the statistical significance only holds in the smallest corpus scenarios (i.e., when the training corpus contains at most 40 000 sentences for Spanish-English, and for all the training corpus sizes except 20 000 for Breton-French 5 ), and the difference between the baseline PBSMT system and our hybrid approach is reduced as the parallel training corpus grows. Apertium data has been developed bearing in mind the translation of general texts (mainly news) whereas the in-domain test sets come from the specialised domains of parliament speeches (Spanish-English) or tourism and computer localisation (Breton-French). Thus, as soon as the PBSMT system learns reliable information from the parallel corpus, Apertium phrases become useless. On the contrary, the out-of-domain Spanish-English tests, performed on a general (news) domain, show a statistically-significant improvement with all the training corpus sizes tested. In this case, Apertium-generated phrases, which contain handcrafted knowledge from a general domain, cover more sequences of words in the input text which are not covered, or are sparsely found, in the original training corpora. The data reported in tables 4 and 5 support these hypotheses; in the in-domain evaluation, the proportion of phrases generated from Apertium included in the translations drops abruptly as the corpus grows. On the contrary, when evaluating the Spanish-English systems in a different domain, the proportion of Apertium- 5 Our results do not agree with those by Tyers (2009) , who reported a substantial improvement in BLEU when adding dictionaries to the training corpus. In a personal communication, the author stated that in Tyers (2009) a baseline in which the feature weights were optimised with MERT was compared to a system enriched with the Apertium dictionaries using the default (not optimised) feature weights. Incidentally, not optimising the feature weights provided better results. If the feature weights are optimised in both cases the results obtained are in the line of those reported in this paper. Table 3 : BLEU score achieved by the different configurations listed in section 5. Hybrid system scores in bold mean that they outperform both Apertium and the PBSMT baseline, and that the improvement is statistically significant. The score of the hybrid system built with the Apertium rules and dictionaries is underlined if it outperforms its dictionary-based counterpart by a statistically significant margin. The \u2248 27k corpus size is only tested with the Breton-French language pair because it corresponds to the full Breton-French training corpus size. generated phrases is higher and falls smoothly, and the value of the feature function is higher than in the in-domain tests. The inclusion of shallow-transfer rules provides a statistically-significant improvement over the dictionaries for all the training corpus sizes in the Spanish-English out-of-domain evaluation scenario and for the smallest ones in the in-domain tests. That is, shallow-transfer rules are effective when the decoder chooses a high proportion of Apertium-generated phrase pairs.",
"cite_spans": [
{
"start": 1507,
"end": 1508,
"text": "5",
"ref_id": null
},
{
"start": 1548,
"end": 1560,
"text": "Tyers (2009)",
"ref_id": "BIBREF20"
},
{
"start": 1714,
"end": 1726,
"text": "Tyers (2009)",
"ref_id": "BIBREF20"
}
],
"ref_spans": [
{
"start": 2103,
"end": 2110,
"text": "Table 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Experimental Settings",
"sec_num": "5"
},
{
"text": "Finally, the addition of the bilingual dictionaries to the training corpus before the computation of the word alignments and the probabilistic bilingual dictionary results in a small performance drop. It remains to be studied whether the dictionaries improve alignments but not translation performance.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Settings",
"sec_num": "5"
},
{
"text": "In this paper we have described a new hybridisation approach consisting of enriching a PBSMT system by adding to its phrase table bilingual phrase pairs matching structural transfer rules and dictionaries from a shallow-transfer RBMT system. The experiments conducted show an improvement of the translation quality when only a small parallel corpus is available. Our approach also helps when training on larger parallel corpora and the texts to translate come from a general (news) domain that is well covered by the RBMT system; in this case, shallow-transfer rules have a greater impact on translation quality than dictionaries.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions and Future Work",
"sec_num": "7"
},
{
"text": "Our future plans include evaluating the presented hybridisation strategy with more language pairs and bigger training corpora, focusing on test corpora from the news domain, which seems to be the scenario in which our approach better fits. We also plan to further investigate the negative impact that adding the entries in the Apertium bilingual dictionary to the corpus has on translation performance. 2k 5k 10k 20k \u2248 27k 40k 80k es-en phrase-dict 0.0025 0.0010 \u22120.0003 \u22120.0003 -0.0007 0.0067 in-domain phrase-rules 0.0029 \u22120.0061 0 \u22120.0023 -\u22120.0054 \u22120.0094 es-en phrase-dict 0.0202 0.0138 0.0073 0.0162 -0.0211 0.0259 out-of-domain phrase-rules 0.0227 0.0219 0.0288 0.0156 -0.0092 0.0230 br-fr phrase-dict 0.0106 0.0052 0.0098 0.0079 0.0115 -in-domain phrase-rules 0.0024 0.0103 0.0020 0.0044 0.0029 -- Table 4 : Relative weights assigned to the binary feature function that flags whether a phrase pair has been obtained from the Apertium bilingual resources (dictionaries and rules) or from the training parallel corpus in the different evaluation set-ups.",
"cite_spans": [],
"ref_spans": [
{
"start": 805,
"end": 812,
"text": "Table 4",
"ref_id": null
}
],
"eq_spans": [],
"section": "Conclusions and Future Work",
"sec_num": "7"
},
{
"text": "Weigths have been normalised by dividing them by the highest weight assigned to a feature of its corresponding set-up. Table 5 : Proportion of phrase pairs used in the translation of the test set that have been generated from the Apertium bilingual resources (dictionaries and rules).",
"cite_spans": [],
"ref_spans": [
{
"start": 119,
"end": 126,
"text": "Table 5",
"ref_id": null
}
],
"eq_spans": [],
"section": "Conclusions and Future Work",
"sec_num": "7"
},
{
"text": "http://www.statmt.org/wmt11/ translation-task.html 2 http://www.statmt.org/wmt10/ translation-task.html3 Revision 3739, downloaded from https: //mosesdecoder.svn.sourceforge.net/ svnroot/mosesdecoder/trunk.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "Revisions 24177, 22150 and 28674, respectively.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "Work funded by the Spanish Ministry of Science and Innovation through project TIN2009-14009-C02-01 and by Generalitat Valenciana through grant ACIF/2010/174 (VALi+d programme).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgments",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "But dictionaries are data too",
"authors": [
{
"first": "P",
"middle": [
"F"
],
"last": "Brown",
"suffix": ""
},
{
"first": "S",
"middle": [
"A D"
],
"last": "Pietra",
"suffix": ""
},
{
"first": "V",
"middle": [
"J D"
],
"last": "Pietra",
"suffix": ""
},
{
"first": "M",
"middle": [
"J"
],
"last": "Goldsmith",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Hajic",
"suffix": ""
},
{
"first": "R",
"middle": [
"L"
],
"last": "Mercer",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Mohanty",
"suffix": ""
}
],
"year": 1993,
"venue": "Proceedings of the workshop on Human Language Technology",
"volume": "",
"issue": "",
"pages": "202--205",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "P. F. Brown, S. A. D. Pietra, V. J. D. Pietra, M. J. Gold- smith, J. Hajic, R. L. Mercer, and S. Mohanty. 1993. But dictionaries are data too. In Proceedings of the workshop on Human Language Technology, pages 202-205.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Improving statistical machine translation by classifying and generalizing inflected verb forms",
"authors": [
{
"first": "A",
"middle": [],
"last": "De Gispert",
"suffix": ""
},
{
"first": "J",
"middle": [
"B"
],
"last": "Mari\u00f1o",
"suffix": ""
},
{
"first": "J",
"middle": [
"M"
],
"last": "Crego",
"suffix": ""
}
],
"year": 2005,
"venue": "Proceedings of the 9th European Conference on Speech Communication and Technology",
"volume": "",
"issue": "",
"pages": "3185--3188",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "A. de Gispert, J.B. Mari\u00f1o, and J.M. Crego. 2005. Im- proving statistical machine translation by classifying and generalizing inflected verb forms. In Proceed- ings of the 9th European Conference on Speech Com- munication and Technology, pages 3185-3188.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Can we Relearn an RBMT System?",
"authors": [
{
"first": "L",
"middle": [],
"last": "Dugast",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Senellart",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Koehn",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of the 3rd Workshop on Statistical Machine Translation",
"volume": "",
"issue": "",
"pages": "175--178",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "L. Dugast, J. Senellart, and P. Koehn. 2008. Can we Relearn an RBMT System? In Proceedings of the 3rd Workshop on Statistical Machine Transla- tion, pages 175-178.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Using Moses to integrate multiple rule-based machine translation engines into a hybrid system",
"authors": [
{
"first": "A",
"middle": [],
"last": "Eisele",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Federmann",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Saint-Amand",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Jellinghaus",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Herrmann",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Chen",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of the 3rd Workshop on Statistical Machine Translation",
"volume": "",
"issue": "",
"pages": "179--182",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "A. Eisele, C. Federmann, H. Saint-Amand, M. Jelling- haus, T. Herrmann, and Y. Chen. 2008. Us- ing Moses to integrate multiple rule-based machine translation engines into a hybrid system. In Pro- ceedings of the 3rd Workshop on Statistical Machine Translation, pages 179-182.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Apertium: a free/open-source platform for rule-based machine translation",
"authors": [
{
"first": "M",
"middle": [
"L"
],
"last": "Forcada",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Ginest\u00ed-Rosell",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Nordfalk",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "O'regan",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Ortiz-Rojas",
"suffix": ""
},
{
"first": "J",
"middle": [
"A"
],
"last": "P\u00e9rez-Ortiz",
"suffix": ""
},
{
"first": "F",
"middle": [],
"last": "S\u00e1nchez-Martnez",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Ram\u00edrez-S\u00e1nchez",
"suffix": ""
},
{
"first": "F",
"middle": [
"M"
],
"last": "Tyers",
"suffix": ""
}
],
"year": 2011,
"venue": "Machine Translation",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"DOI": [
"10.1007/s10590-011-9090-0"
]
},
"num": null,
"urls": [],
"raw_text": "M. L. Forcada, M. Ginest\u00ed-Rosell, J. Nordfalk, J. O'Regan, S. Ortiz-Rojas, J. A. P\u00e9rez-Ortiz, F. S\u00e1nchez-Martnez, G. Ram\u00edrez-S\u00e1nchez, and F. M. Tyers. 2011. Apertium: a free/open-source platform for rule-based machine translation. Machine Trans- lation. doi: 10.1007/s10590-011-9090-0.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "An empirical study of smoothing techniques for language modeling",
"authors": [
{
"first": "J",
"middle": [],
"last": "Goodman",
"suffix": ""
},
{
"first": "S",
"middle": [
"F"
],
"last": "Chen",
"suffix": ""
}
],
"year": 1998,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J. Goodman and S. F. Chen. 1998. An empirical study of smoothing techniques for language model- ing. Technical Report TR-10-98, Harvard Univer- sity, August.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "An introduction to machine translation",
"authors": [
{
"first": "W",
"middle": [
"J"
],
"last": "Hutchins",
"suffix": ""
},
{
"first": "H",
"middle": [
"L"
],
"last": "Somers",
"suffix": ""
}
],
"year": 1992,
"venue": "",
"volume": "362",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "W. J. Hutchins and H. L. Somers. 1992. An introduc- tion to machine translation, volume 362. Academic Press, New York.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Moses: Open Source Toolkit for Statistical Machine Translation",
"authors": [
{
"first": "P",
"middle": [],
"last": "Koehn",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Hoang",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Birch",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Callison-Burch",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Federico",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Bertoldi",
"suffix": ""
},
{
"first": "B",
"middle": [],
"last": "Cowan",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Shen",
"suffix": ""
},
{
"first": "W",
"middle": [],
"last": "Moran",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Zens",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Dyer",
"suffix": ""
},
{
"first": "O",
"middle": [],
"last": "Bojar",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Constantin",
"suffix": ""
},
{
"first": "E",
"middle": [],
"last": "Herbst",
"suffix": ""
}
],
"year": 2007,
"venue": "Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics, demonstration session",
"volume": "",
"issue": "",
"pages": "177--180",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "P. Koehn, H. Hoang, A. Birch, C. Callison-Burch, M. Federico, N. Bertoldi, B. Cowan, C. Shen, W.and Moran, R. Zens, C. Dyer, O. Bojar, A. Con- stantin, and E. Herbst. 2007. Moses: Open Source Toolkit for Statistical Machine Translation. In Pro- ceedings of the 45th Annual Meeting of the Associ- ation for Computational Linguistics, demonstration session, pages 177-180.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Europarl: A parallel corpus for statistical machine translation",
"authors": [
{
"first": "P",
"middle": [],
"last": "Koehn",
"suffix": ""
}
],
"year": 2005,
"venue": "",
"volume": "5",
"issue": "",
"pages": "12--16",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "P. Koehn. 2005. Europarl: A parallel corpus for statis- tical machine translation. MT summit, 5:12-16.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Statistical Machine Translation",
"authors": [
{
"first": "P",
"middle": [],
"last": "Koehn",
"suffix": ""
}
],
"year": 2010,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "P. Koehn. 2010. Statistical Machine Translation. Cam- bridge University Press.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Improving a general-purpose Statistical Translation Engine by terminological lexicons",
"authors": [
{
"first": "P",
"middle": [],
"last": "Langlais",
"suffix": ""
}
],
"year": 2002,
"venue": "Second International Workshop on Computational Terminology",
"volume": "",
"issue": "",
"pages": "1--7",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "P. Langlais. 2002. Improving a general-purpose Sta- tistical Translation Engine by terminological lexi- cons. In Second International Workshop on Com- putational Terminology, pages 1-7.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Word-based alignment, phrase-based translation: Whats the link",
"authors": [
{
"first": "A",
"middle": [],
"last": "Lopez",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Resnik",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of the 7th Conference of the Association for Machine Translation in the Americas",
"volume": "",
"issue": "",
"pages": "90--99",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "A. Lopez and P. Resnik. 2006. Word-based alignment, phrase-based translation: Whats the link. In Pro- ceedings of the 7th Conference of the Association for Machine Translation in the Americas, pages 90-99.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"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 2009 Conference on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "381--390",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Y. Marton, C. Callison-Burch, and P. Resnik. 2009. Improved statistical machine translation using monolingually-derived paraphrases. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 381-390.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "A systematic comparison of various statistical alignment models",
"authors": [
{
"first": "F",
"middle": [
"J"
],
"last": "Och",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Ney",
"suffix": ""
}
],
"year": 2003,
"venue": "Computational Linguistics",
"volume": "29",
"issue": "",
"pages": "19--51",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "F. J. Och and H. Ney. 2003. A systematic comparison of various statistical alignment models. Computa- tional Linguistics, 29:19-51, March.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Minimum error rate training in statistical machine translation",
"authors": [
{
"first": "F",
"middle": [
"J"
],
"last": "Och",
"suffix": ""
}
],
"year": 2003,
"venue": "Proceedings of the 41st Annual Meeting on Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "160--167",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "F. J. Och. 2003. Minimum error rate training in statis- tical machine translation. In Proceedings of the 41st Annual Meeting on Association for Computational Linguistics, pages 160-167.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"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": [],
"last": "Zhu",
"suffix": ""
}
],
"year": 2002,
"venue": "Proceedings of the 40th Annual Meeting on Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "311--318",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "K. Papineni, S. Roukos, T. Ward, and W. Zhu. 2002. BLEU: a method for automatic evaluation of ma- chine translation. In Proceedings of the 40th Annual Meeting on Association for Computational Linguis- tics, pages 311-318.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Statistical machine translation with a small amount of bilingual training data",
"authors": [
{
"first": "M",
"middle": [],
"last": "Popovi\u0107",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Ney",
"suffix": ""
}
],
"year": 2006,
"venue": "LREC workshop on Minority Languages",
"volume": "",
"issue": "",
"pages": "25--29",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "M. Popovi\u0107 and H. Ney. 2006. Statistical machine translation with a small amount of bilingual training data. In LREC workshop on Minority Languages, pages 25-29.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "SMT and SPE machine translation systems for WMT'09",
"authors": [
{
"first": "H",
"middle": [],
"last": "Schwenk",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Abdul-Rauf",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Barrault",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Senellart",
"suffix": ""
}
],
"year": 2009,
"venue": "Proceedings of the 4th Workshop on Statistical Machine Translation",
"volume": "",
"issue": "",
"pages": "130--134",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "H. Schwenk, S. Abdul-Rauf, L. Barrault, and J. Senel- lart. 2009. SMT and SPE machine translation sys- tems for WMT'09. In Proceedings of the 4th Work- shop on Statistical Machine Translation, pages 130- 134.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "SRILM -an extensible language modeling toolkit",
"authors": [
{
"first": "A",
"middle": [],
"last": "Stolcke",
"suffix": ""
}
],
"year": 2002,
"venue": "7th International Conference on Spoken Language Processing",
"volume": "",
"issue": "",
"pages": "901--904",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "A. Stolcke. 2002. SRILM -an extensible language modeling toolkit. In 7th International Conference on Spoken Language Processing, pages 901-904.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Comparing different architectures of hybrid Machine Translation systems",
"authors": [
{
"first": "G",
"middle": [],
"last": "Thurmair",
"suffix": ""
}
],
"year": 2009,
"venue": "Proceedings MT Summit XII",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "G. Thurmair. 2009. Comparing different architectures of hybrid Machine Translation systems. In Proceed- ings MT Summit XII.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Rule-based augmentation of training data in Breton-French statistical machine translation",
"authors": [
{
"first": "F",
"middle": [
"M"
],
"last": "Tyers",
"suffix": ""
}
],
"year": 2009,
"venue": "Proceedings of the 13th Annual Conference of the European Association for Machine Translation",
"volume": "",
"issue": "",
"pages": "213--217",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "F. M. Tyers. 2009. Rule-based augmentation of train- ing data in Breton-French statistical machine transla- tion. In Proceedings of the 13th Annual Conference of the European Association for Machine Transla- tion, pages 213-217.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Interpreting BLEU/NIST scores: How much improvement do we need to have a better system",
"authors": [
{
"first": "Y",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Vogel",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Waibel",
"suffix": ""
}
],
"year": 2004,
"venue": "Proceedings of the 4th International Conference on Language Resources and Evaluation",
"volume": "",
"issue": "",
"pages": "2051--2054",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Y. Zhang, S. Vogel, and A. Waibel. 2004. Interpret- ing BLEU/NIST scores: How much improvement do we need to have a better system. In Proceedings of the 4th International Conference on Language Re- sources and Evaluation, pages 2051-2054.",
"links": null
}
},
"ref_entries": {
"TABREF1": {
"text": "Description of the Breton-French parallel corpora used in the experiments.",
"num": null,
"content": "<table/>",
"html": null,
"type_str": "table"
},
"TABREF3": {
"text": "Description of the Spanish-English parallel corpora used in the experiments.",
"num": null,
"content": "<table/>",
"html": null,
"type_str": "table"
},
"TABREF4": {
"text": "24.24 26.46 28.45 -29.86 30.88 12.59 14.90 16.92 18.63 -20.32 21.80 alignment 19.31 23.71 25.89 28.10 -29.73 30.83 12.06 14.55 16.88 18.66 -20.34 21.68 es-en phrase-dict 24.29 26.39 27.93 29.30 -30.36 31.14 19.76 20.48 21.26 21.89 -22.67 23.20 phrase-rules 24.68 26.81 28.28 29.40 -30.41 31.02 20.97 21.36 22.20 22.77",
"num": null,
"content": "<table><tr><td/><td/><td/><td/><td>In-domain</td><td/><td/><td/><td/><td colspan=\"3\">Out-of-domain</td></tr><tr><td/><td>2k</td><td>5k</td><td>10k</td><td colspan=\"3\">20k \u2248 27k 40k 80k</td><td>2k</td><td>5k</td><td>10k</td><td colspan=\"2\">20k \u2248 27k 40k</td><td>80k</td></tr><tr><td>baseline</td><td colspan=\"11\">20.74 -</td><td>23.29 23.76</td></tr><tr><td>Apertium</td><td/><td/><td/><td>18.00</td><td/><td/><td/><td/><td/><td>20.30</td></tr><tr><td>baseline</td><td colspan=\"4\">18.86 24.17 28.26 33.17 34.69</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td colspan=\"5\">alignment 17.53 23.56 27.82 32.17 34.76</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td colspan=\"5\">br-fr phrase-dict 21.57 26.39 29.66 33.42 35.50</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td colspan=\"5\">phrase-rules 22.67 26.42 29.60 33.14 35.83</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Apertium</td><td/><td/><td/><td>17.56</td><td/><td/><td/><td/><td/><td>-</td></tr></table>",
"html": null,
"type_str": "table"
}
}
}
} |