File size: 83,140 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 | {
"paper_id": "D09-1022",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T16:38:09.037486Z"
},
"title": "Extending Statistical Machine Translation with Discriminative and Trigger-Based Lexicon Models",
"authors": [
{
"first": "Arne",
"middle": [],
"last": "Mauser",
"suffix": "",
"affiliation": {
"laboratory": "Human Language Technology and Pattern Recognition Group Chair of Computer Science 6",
"institution": "RWTH Aachen University",
"location": {
"country": "Germany"
}
},
"email": ""
},
{
"first": "Sa\u0161a",
"middle": [],
"last": "Hasan",
"suffix": "",
"affiliation": {
"laboratory": "Human Language Technology and Pattern Recognition Group Chair of Computer Science 6",
"institution": "RWTH Aachen University",
"location": {
"country": "Germany"
}
},
"email": ""
},
{
"first": "Hermann",
"middle": [],
"last": "Ney",
"suffix": "",
"affiliation": {
"laboratory": "Human Language Technology and Pattern Recognition Group Chair of Computer Science 6",
"institution": "RWTH Aachen University",
"location": {
"country": "Germany"
}
},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "In this work, we propose two extensions of standard word lexicons in statistical machine translation: A discriminative word lexicon that uses sentence-level source information to predict the target words and a trigger-based lexicon model that extends IBM model 1 with a second trigger, allowing for a more fine-grained lexical choice of target words. The models capture dependencies that go beyond the scope of conventional SMT models such as phraseand language models. We show that the models improve translation quality by 1% in BLEU over a competitive baseline on a large-scale task.",
"pdf_parse": {
"paper_id": "D09-1022",
"_pdf_hash": "",
"abstract": [
{
"text": "In this work, we propose two extensions of standard word lexicons in statistical machine translation: A discriminative word lexicon that uses sentence-level source information to predict the target words and a trigger-based lexicon model that extends IBM model 1 with a second trigger, allowing for a more fine-grained lexical choice of target words. The models capture dependencies that go beyond the scope of conventional SMT models such as phraseand language models. We show that the models improve translation quality by 1% in BLEU over a competitive baseline on a large-scale task.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Lexical dependencies modeled in standard phrasebased SMT are rather local. Even though the decision about the best translation is made on sentence level, phrase models and word lexicons usually do not take context beyond the phrase boundaries into account. This is especially problematic since the average source phrase length used during decoding is small. When translating Chinese to English, e.g., it is typically close to only two words.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The target language model is the only model that uses lexical context across phrase boundaries. It is a very important feature in the log-linear setup of today's phrase-based decoders. However, its context is typically limited to three to six words and it is not informed about the source sentence. In the presented models, we explicitly take advantage of sentence-level dependencies including the source side and make non-local predictions for the target words. This is an important aspect when translating from languages like German and Chinese where long-distance dependencies are common. In Chinese, for example, tenses are often encoded by indicator words and particles whose position is relatively free in the sentence. In German, prefixes of verbs can be moved over long distances towards the end of the sentence.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In this work, we propose two models that can be categorized as extensions of standard word lexicons: A discriminative word lexicon that uses global, i.e. sentence-level source information to predict the target words using a statistical classifier and a trigger-based lexicon model that extends the well-known IBM model 1 (Brown et al., 1993) with a second trigger, allowing for a more finegrained lexical choice of target words. The loglinear framework of the discriminative word lexicon offers a high degree of flexibility in the selection of features. Other sources of information such as syntax or morphology can be easily integrated.",
"cite_spans": [
{
"start": 321,
"end": 341,
"text": "(Brown et al., 1993)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The trigger-based lexicon model, or simply triplet model since it is based on word triplets, is not trained discriminatively but uses the classical maximum likelihood approach (MLE) instead. We train the triplets iteratively on a training corpus using the Expectation-Maximization (EM) algorithm. We will present how both models allow for a representation of topic-related sentencelevel information which puts them close to word sense disambiguation (WSD) approaches. As will be shown later, the experiments indicate that these models help to ensure translation of content words that are often omitted by the baseline system. This is a common problem in Chinese-English translation. Furthermore, the models are often capable to produce a better lexical choice of content words.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The structure of the paper is as follows: In Section 2, we will address related work and briefly pin down how our models differentiate from previous work. Section 3 will describe the discriminative lexical selection model and the triplet model in more detail, explain the training procedures and show how the models are integrated into the decoder. The experimental setup and results will be given in Section 4. A more detailed discussion will be presented in Section 5. In the end, we conclude our findings and give an outlook for further research in Section 6.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Several word lexicon models have emerged in the context of multilingual natural language processing. Some of them were used as a machine translation system or as a part of one such system. There are three major types of models: Heuristic models as in (Melamed, 2000) , generative models as the IBM models (Brown et al., 1993) and discriminative models (Varea et al., 2001; Bangalore et al., 2006) .",
"cite_spans": [
{
"start": 251,
"end": 266,
"text": "(Melamed, 2000)",
"ref_id": "BIBREF10"
},
{
"start": 305,
"end": 325,
"text": "(Brown et al., 1993)",
"ref_id": "BIBREF2"
},
{
"start": 352,
"end": 372,
"text": "(Varea et al., 2001;",
"ref_id": "BIBREF14"
},
{
"start": 373,
"end": 396,
"text": "Bangalore et al., 2006)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "Similar to this work, the authors of (Varea et al., 2001 ) try to incorporate a maximum entropy lexicon model into an SMT system. They use the words and word classes from the local context as features and show improvements with nbest rescoring.",
"cite_spans": [
{
"start": 37,
"end": 56,
"text": "(Varea et al., 2001",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "The models in this paper are also related to word sense disambiguation (WSD). For example, (Chan et al., 2007 ) trained a discriminative model for WSD using local but also across-sentence unigram collocations of words in order to refine phrase pair selection dynamically by incorporating scores from the WSD classifier. They showed improvements in translation quality in a hierarchical phrase-based translation system. Another WSD approach incorporating context-dependent phrasal translation lexicons is given in (Carpuat and Wu, 2007) and has been evaluated on several translation tasks. Our model differs from the latter in three ways. First, our approach models word selection of the target sentence based on global sentence-level features of the source sentence. Second, instead of disambiguating phrase senses as in (Carpuat and Wu, 2007) , we model word selection independently of the phrases used in the MT models. Finally, the training is done in a different way as will be presented in Sections 3.1.1 and 3.2.1.",
"cite_spans": [
{
"start": 91,
"end": 109,
"text": "(Chan et al., 2007",
"ref_id": "BIBREF5"
},
{
"start": 513,
"end": 535,
"text": "(Carpuat and Wu, 2007)",
"ref_id": "BIBREF4"
},
{
"start": 821,
"end": 843,
"text": "(Carpuat and Wu, 2007)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "Recently, full translation models using discriminative training criteria emerged as well. They are designed to generate a translation for a given source sentence and not only score or disambiguate hypotheses given by a translation system. In (Ittycheriah and Roukos, 2007) , the model can predict 1-to-many translations with gaps and uses words, morphologic and syntactic features from the local context.",
"cite_spans": [
{
"start": 242,
"end": 272,
"text": "(Ittycheriah and Roukos, 2007)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "The authors of (Venkatapathy and Bangalore, 2007) propose three different models. The first one is a global lexical selection model which includes all words of the source sentence as features, regardless of their position. Using these features, the system predicts the words that should be included in the target sentence. Sentence structure is then reconstructed using permutations of the generated bag of target words. We will also use this type of features in our model. One of the simplest models in the context of lexical triggers is the IBM model 1 (Brown et al., 1993) which captures lexical dependencies between source and target words. It can be seen as a lexicon containing correspondents of translations of source and target words in a very broad sense since the pairs are trained on the full sentence level. The trigger-based lexicon model used in this work follows the training procedure introduced in (Hasan et al., 2008) and is integrated directly in the decoder instead of being applied in n-best list reranking. The model is very close to the IBM model 1 and can be seen as an extension of it by taking another word into the conditioning part, i.e. the triggering items. Thus, instead of p(f |e), it models p(f |e, e ). Furthermore, since the second trigger can come from any part of the sentence, there is a link to long-range monolingual triggers as presented in (Tillmann and Ney, 1997) where a trigger language model was trained using the EM algorithm and helped to reduce perplexities and word error rates in a speech recognition experiment. In (Rosenfeld, 1996) , another approach was chosen to model monolingual triggers using a maximum-entropy based framework. Again, this adapted LM could improve speech recognition performance significantly.",
"cite_spans": [
{
"start": 15,
"end": 49,
"text": "(Venkatapathy and Bangalore, 2007)",
"ref_id": "BIBREF15"
},
{
"start": 555,
"end": 575,
"text": "(Brown et al., 1993)",
"ref_id": "BIBREF2"
},
{
"start": 915,
"end": 935,
"text": "(Hasan et al., 2008)",
"ref_id": "BIBREF8"
},
{
"start": 1382,
"end": 1406,
"text": "(Tillmann and Ney, 1997)",
"ref_id": "BIBREF13"
},
{
"start": 1567,
"end": 1584,
"text": "(Rosenfeld, 1996)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "A comparison of a variant of the trigger-based lexicon model applied in decoding and n-best list reranking can be found in (Hasan and Ney, 2009) . In order to reduce the number of overall triplets, the authors use the word alignments for fixing the first trigger to the aligned target word. In general, this constraint performs slightly worse than the unconstrained variant used in this work, but allows for faster training and decoding.",
"cite_spans": [
{
"start": 123,
"end": 144,
"text": "(Hasan and Ney, 2009)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "2"
},
{
"text": "In this section, we present the extended lexicon models, how they are trained and integrated into the phrase-based decoder.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Extended Lexicon Models",
"sec_num": "3"
},
{
"text": "Discriminative models have been shown to outperform generative models on many natural language processing tasks. For machine translation, however, the adaptation of these methods is difficult due to the large space of possible translations and the size of the training data that has to be used to achieve significant improvements.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discriminative Lexicon Model",
"sec_num": "3.1"
},
{
"text": "In this section, we propose a discriminative word lexicon model that follows and integrate it into the standard phrasebased machine translation approach.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discriminative Lexicon Model",
"sec_num": "3.1"
},
{
"text": "The core of our model is a classifier that predicts target words, given the words of the source sentence. The structure of source as well as target sentence is neglected in this model. We do not make any assumptions about the location of the words in the sentence. This is useful in many cases, as words and morphology can depend on information given at other positions in the sentence. An example would be the character \u4e86 in Chinese that indicates a completed or past action and does not need to appear close to the verb.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discriminative Lexicon Model",
"sec_num": "3.1"
},
{
"text": "We model the probability of the set of target words in a sentence e given the set of source words f . For each word in the target vocabulary, we can calculate a probability for being or not being included in the set. The probability of the whole set then is the product over the entire target vocabulary V E :",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discriminative Lexicon Model",
"sec_num": "3.1"
},
{
"text": "P (e|f ) = e\u2208e P (e + |f ) \u2022 e\u2208V E \\e P (e \u2212 |f ) (1)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discriminative Lexicon Model",
"sec_num": "3.1"
},
{
"text": "For notational simplicity, we use the event e + when the target word e is included in the target sentence and e \u2212 if not. We model the individual factors p(e|f ) of the probability in Eq. 1 as a loglinear model using the source words from f as binary features",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discriminative Lexicon Model",
"sec_num": "3.1"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "\u03c6(f, f ) = 1 if f \u2208 f 0 else",
"eq_num": "(2)"
}
],
"section": "Discriminative Lexicon Model",
"sec_num": "3.1"
},
{
"text": "and feature weights \u03bb f,\u2022 :",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discriminative Lexicon Model",
"sec_num": "3.1"
},
{
"text": "P (e + |f ) = exp f \u2208f \u03bb f,e + \u03c6(f, f ) e\u2208{e + ,e \u2212 } exp f \u2208f \u03bb f,e \u03c6(f, f )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discriminative Lexicon Model",
"sec_num": "3.1"
},
{
"text": "(3) Subsequently, we will call this model discriminative word lexicon (DWL).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discriminative Lexicon Model",
"sec_num": "3.1"
},
{
"text": "Modeling the lexicon on sets and not on sequences has two reasons. Phrase-based MT along with n-gram language models is strong at predicting sequences but only uses information from a local context. By using global features and predicting words in a non-local fashion, we can augment the strong local decisions from the phrase-based systems with sentence-level information.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discriminative Lexicon Model",
"sec_num": "3.1"
},
{
"text": "For practical reasons, translating from a set to a set simplifies the parallelization of the training procedure. The classifiers for the target words can be trained separately as explained in the following section.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discriminative Lexicon Model",
"sec_num": "3.1"
},
{
"text": "Common classification tasks have a relatively small number of classes. In our case, the number of classes is the size of the target vocabulary. For large translation tasks, this is in the range of a hundred thousand classes. It is far from what conventional out-of-the-box classifiers can handle.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Training",
"sec_num": "3.1.1"
},
{
"text": "The discriminative word lexicon model has the convenient property that we can train a separate model for each target word making parallelization straightforward. Discussions about possible classifiers and the choice of regularization can be found in . We used the freely available MegaM Toolkit 1 for training, which implements the L-BFGS method (Byrd et al., 1995) . Regularization is done using Gaussian priors. We performed 100 iterations of the training algorithm for each word in the target vocabulary. This results in a large number of classifiers to be trained. For the Arabic-English data (cf. Section 4), the training took an average of 38 seconds per word. No feature cutoff was used.",
"cite_spans": [
{
"start": 346,
"end": 365,
"text": "(Byrd et al., 1995)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Training",
"sec_num": "3.1.1"
},
{
"text": "In search, we compute the model probabilities as an additional model in the log-linear model combination of the phrase-based translation approach. To reduce the memory footprint and startup time of the decoding process, we reduced the number of parameters by keeping only large values \u03bb f,e since smaller values tend to have less effect on the overall probability. In experiments we determined that we could safely reduce the size of the final model by a factor of ten without losing predictive power. In search, we compute the model probabilities as an additional model in the log-linear combination. When scoring hypotheses from the phrase-based system, we see the translation hypothesis as the set of target words that are predicted. Words from the target vocabulary which are not included in the hypothesis are not part of the set. During the search process, however, we also have to score incomplete hypotheses where we do not know which words will not be included. This problem is circumvented by rewriting Eq. 1 as",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Decoding",
"sec_num": "3.1.2"
},
{
"text": "P (e|f ) = e\u2208V E P (e \u2212 |f ) \u2022 e\u2208e P (e + |f ) P (e \u2212 |f )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Decoding",
"sec_num": "3.1.2"
},
{
"text": ".",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Decoding",
"sec_num": "3.1.2"
},
{
"text": "The first product is constant given a source sentence and therefore does not affect the search. Using the model assumption from Eq. 3, we can further simplify the computation and compute the model score entirely in log-space which is numerically stable even for large vocabularies. Experiments showed that using only the first factor of Eq. 1 is sufficient to obtain good results.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Decoding",
"sec_num": "3.1.2"
},
{
"text": "In comparison with the translation model from where a threshold on the probability is used to determine which words are included in the target sentence, our approach relies on the phrase model to generate translation candidates. This has several advantages: The length of the translation is determined by the phrase model. Words occurring multiple times in the translation do not have to be explicitly modeled. In , repeated target words are treated as distinct classes.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Decoding",
"sec_num": "3.1.2"
},
{
"text": "The main advantage of the integration being done in a way as presented here is that the phrase model and the discriminative word lexicon model are complementary in the way they model the translation. While the phrase model is good in predicting translations in a local context, the discriminative word lexicon model is able to predict global aspects of the sentence like tense or vocabulary changes in questions. While the phrase model is closely tied to the structure of word and phrase alignments, the discriminative word lexicon model completely disregards the structure in source and target sentences.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Decoding",
"sec_num": "3.1.2"
},
{
"text": "The triplets of the trigger-based lexicon model, i.e. p(e|f, f ), are composed of two words in the source language triggering one target language word. We chose this inverse direction since it can be integrated directly into the decoder and, thus, does not rely on a two-pass approach using reranking, as it is the case for (Hasan et al., 2008) . The triggers can originate from words of the whole source sentence, also crossing phrase boundaries of the conventional bilingual phrase pairs. The model is symmetric though, meaning that the order of the triggers is not relevant, i.e. (f, f \u2192 e) = (f , f \u2192 e). Nevertheless, the model is able to capture long-distance effects such as verb splits or adjustments to lexical choice of the target word given the topic-triggers of the source sentence. In training, we determine the probability of a target sentence e I 1 given the source sentence f J 1 within the model by",
"cite_spans": [
{
"start": 324,
"end": 344,
"text": "(Hasan et al., 2008)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Trigger-based Lexicon Model",
"sec_num": "3.2"
},
{
"text": "p(e I 1 |f J 1 ) = I i=1 p(e i |f J 1 ) = I i=1 2 J(J + 1) J j=0 J j =j+1 p(e i |f j , f j ), (4)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Trigger-based Lexicon Model",
"sec_num": "3.2"
},
{
"text": "where f 0 denotes the empty word and, thus, for f j = \u03b5, allows for modeling the conventional (inverse) IBM model 1 lexical probabilities as well. Since the second trigger f j always starts right of the current first trigger, the model is symmetric and does not need to look at all trigger pairs. Eq. 4 is used in the iterative EM training on all sentence pairs of the training data which is described in more detail in the following.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Trigger-based Lexicon Model",
"sec_num": "3.2"
},
{
"text": "For training the trigger-based lexicon model, we apply the Expectation-Maximization (EM) algorithm (Dempster et al., 1977) . The goal is to maximize the log-likelihood F trip of this model for a given bilingual training corpus",
"cite_spans": [
{
"start": 99,
"end": 122,
"text": "(Dempster et al., 1977)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Training",
"sec_num": "3.2.1"
},
{
"text": "{(f Jn 1 , e In 1 )} N 1",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Training",
"sec_num": "3.2.1"
},
{
"text": "consisting of N sentence pairs:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Training",
"sec_num": "3.2.1"
},
{
"text": "F trip := N n=1 log p(e In 1 |f Jn 1 ),",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Training",
"sec_num": "3.2.1"
},
{
"text": "where I n and J n are the lengths of the n-th target and source sentence, respectively. An auxiliary function Q(\u00b5;\u03bc) is defined based on F trip where\u03bc is the updated estimate within an iteration which is to be derived from the current estimate \u00b5. Here, \u00b5 stands for the entire set of model parameters, i.e. the set of all {\u03b1(e|f, f )} with the constraint e \u03b1(e|f, f ) = 1. The accumulators \u03b1(\u2022) are therefore iteratively trained on the training data by using the current estimate, i.e. deriving the expected value (E-step), and maximizing their likelihood afterwards to reestimate the distribution. Thus, the perplexity of the training data is reduced in each iteration.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Training",
"sec_num": "3.2.1"
},
{
"text": "In search, we can apply this model directly when scoring bilingual phrase pairs. Given a trained model for p(e|f, f ), we compute the feature score h trip (\u2022) of a phrase pair (\u1ebd,f ) as",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Decoding",
"sec_num": "3.2.2"
},
{
"text": "h trip (\u1ebd,f , f J 0 ) = (5) \u2212 i log 2 J \u2022 (J + 1) j j >j p(\u1ebd i |f j , f j ) ,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Decoding",
"sec_num": "3.2.2"
},
{
"text": "where i moves over all target words in the phras\u1ebd e, the second sum selects all source sentence words f J 0 including the empty word, and j > j incorporates the rest of the source sentence right of the first trigger. We take negative log-probabilities and normalize to obtain the final score (representing costs) for the given phrase pair. Note that in search, we can only use this direction, p(e|f, f ), since the whole source sentence is available for triggering effects whereas not all target words have been generated so far, as it would be necessary for the standard direction, p(f |e, e ).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Decoding",
"sec_num": "3.2.2"
},
{
"text": "Due to the enormous number of triplets, we trained the model on a subset of the overall training data. The subcorpus, mainly consisting of newswire articles, contained 1.4M sentence pairs with 32.3M running words on the English side. We trained two versions of the triplet lexicon, one using 4 EM iterations and another one that was trained for 10 EM iterations. Due to trimming of triplets with small probabilities after each iteration, the version based on 10 iterations was slightly smaller, having 164 million triplets but also performed slightly worse. Thus, for the experiments, we used the version based on 4 iterations which contained 291 million triplets.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Decoding",
"sec_num": "3.2.2"
},
{
"text": "Note that decoding with this model can be quite efficient if caching is applied. Since the given source sentence does not change, we have to calculate p(e|f, f ) for each e only once and can re-train (C/E) test08 (NW/WT) Sent. pairs 9.1M 480 490 Run. words 259M/300M 14.8K 12.3K Vocabulary 357K/627K 3.6K 3.2K trieve the probabilities from the cache for consecutive scorings of the same target word e. This significantly speeds up the decoding process.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Decoding",
"sec_num": "3.2.2"
},
{
"text": "In this section we evaluate our lexicon models on the GALE Chinese-English task for newswire and web text translation and additionally on the official NIST 2008 task for both Chinese-English and Arabic-English. The baseline system was built using a state-of-the art phrase-based MT system (Zens and Ney, 2008) . We use the standard set of models with phrase translation probabilities for source-to-target and target-to-source direction, smoothing with lexical weights, a word and phrase penalty, distance-based and lexicalized reordering and a 5-gram (GALE) or 6-gram (NIST) target language model. We used training data provided by the Linguistic Data Consortium (LDC) consisting of 9.1M parallel Chinese-English sentence pairs of various domains for GALE (cf. Table 1) and smaller amounts of data for the NIST systems (cf. Table 2). The DWL and Triplet models were integrated into the decoder as presented in Section 3.",
"cite_spans": [
{
"start": 289,
"end": 309,
"text": "(Zens and Ney, 2008)",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Evaluation",
"sec_num": "4"
},
{
"text": "For the GALE development and test set, we separated the newswire and web text parts and did separate parameter tuning for each genre using the corresponding development set which consists of 485 sentences for newswire texts and 533 sentences of web text. The test set has 480 sentences for newswire and 490 sentences for web text. For NIST, we tuned on the official 2006 eval set and used the 2008 evaluation set as a blind test set. Table 3 : Results on the GALE Chinese-English test set for the newswire and web text setting (case-insensitive evaluation).",
"cite_spans": [],
"ref_spans": [
{
"start": 434,
"end": 441,
"text": "Table 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Experimental Evaluation",
"sec_num": "4"
},
{
"text": "The translation results on the two GALE test sets are shown in Table 3 for newswire and web text. Both the discriminative word lexicon and the triplet lexicon can individually improve the baseline by approximately +0.6-0.9% BLEU and -0.5-0.8% TER. For the combination of both lexicons on the newswire setting, we observe only a slight improvement on BLEU but also an additional boost in TER reduction, arriving at +1% BLEU and -1.2% TER. For web text, the findings are similar: The combination of the discriminative and trigger-based lexicons yields +1% BLEU and decreases TER by -0.5%. We compared these results against an inverse IBM model 1 but the results were inconclusive which is consistent with the results presented in (Och et al., 2004) where no improvements were achieved using p(e|f ). In our case, inverse IBM1 improves results by 0.2-0.4% BLEU on the development set but does not show the same trend on the test sets. Furthermore, combining IBM1 with DWL or Triplets often even degraded the translation results, e.g. only 32.8% BLEU was achieved on newswire for a combination of the IBM1, DWL and Triplet model. In contrast, combinations of the DWL and Triplet model did not degrade performance and could benefit from each other.",
"cite_spans": [
{
"start": 728,
"end": 746,
"text": "(Och et al., 2004)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [
{
"start": 63,
"end": 70,
"text": "Table 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Translation Results",
"sec_num": "4.1"
},
{
"text": "In addition to the automatic scoring, we also did a randomized subjective evaluation where the hypotheses of the baseline was compared against the hypotheses generated using the discriminative word lexicon and triplet models. We evaluated 200 sentences from newswire and web text. In 80% of the evaluated sentences, the improved models were judged equal or better than the baseline.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Translation Results",
"sec_num": "4.1"
},
{
"text": "We tested the presented lexicon models also on another large-scale system, i.e. NIST, for two lan- guage pairs, namely Chinese-English and Arabic-English. Interestingly, the results obtained for Arabic-English are similar to the findings for Chinese-English, as can be seen in Table 4 . The overall improvements for this language pair are +1% BLEU and -1.4% TER. In contrast to the GALE Chinese-English task, the triplet lexicon model for the Arabic-English language pair performs slightly better than the discriminative word lexicon. These results strengthen the claim that the presented models are capable of improving lexical choice of the MT system. In the next section, we discuss the observed effects and analyze our results in more detail.",
"cite_spans": [],
"ref_spans": [
{
"start": 277,
"end": 284,
"text": "Table 4",
"ref_id": "TABREF4"
}
],
"eq_spans": [],
"section": "Translation Results",
"sec_num": "4.1"
},
{
"text": "In terms of automatic evaluation measures, the results indicate that it is helpful to incorporate the extended lexicon models into the search process. In this section, we will analyze some more details of the models and take a look at the lexical choice they make and what differentiates them from the baseline models. In Table 5 , we picked an example sentence from the GALE newswire test set and show the different hypotheses produced by our system. As can be seen, the baseline does not produce the present participle of the verb restore which makes the sentence somewhat hard to understand. Both the discriminative and the triggerbased lexicon approach are capable of generating this missing information, i.e. the correct use of restoring. Figure 1 gives an example how discontinuous triggers affect the word choice on the target side. Two cases are depicted where high probabilities of triplets including emergency and restoring on the target side influence the overall hypothesis selection. The non-local modeling advantages of the triplet model can be observed as well: The \u76ee\u524d , \u4e8b\u6545 \u62a2\u9669 \u7ec4 \u6b63\u5728 \u7d27\u6025 \u6062\u590d \u901a\u98ce \u7cfb\u7edf . triggering events do not need to be located next to each other or within a given phrase pair. They move across the whole source sentence, thus allowing for capturing of long-range dependencies. Table 6 shows the top ten content words that are predicted by the two models, discriminative word lexicon and triplet lexicon model. IBM model 1 ranks are indicated by subscripts in the column of the triplet model. Although the triplet model is similar to IBM1, we observe differences in the word lists. Comparing this to the visualization of the probability distribution for the example sentence, cf. Figure 2 , we argue that, although the IBM1 and Triplet distributions look similar, the triplet model is sharper and favors words such as the ones in Table 6 , resulting in different word choice in the translation process. In contrast, the DWL approach gives more distinct probabilities, selecting content words that are not chosen by the other models. Table 7 shows an example from the web text test set. Here, the baseline hypothesis contains an incorrect word, anna, which might have been mistaken for the name ying. Interestingly, the hypotheses of the DWL lexicon and the combination of DWL and Triplet contain the correct content word remarks. The triplet model makes an error by selecting music, an artifact that might come from words that co-occur frequently with the cor-responding Chinese verb to listen, i.e. \u542c , in the data. Although the TER score of the baseline is better than the one for the alternative models for this particular example, we still think that the observed effects show how our models help producing different hypotheses that might lead to subjectively better translations.",
"cite_spans": [],
"ref_spans": [
{
"start": 322,
"end": 329,
"text": "Table 5",
"ref_id": "TABREF5"
},
{
"start": 744,
"end": 752,
"text": "Figure 1",
"ref_id": "FIGREF0"
},
{
"start": 1305,
"end": 1312,
"text": "Table 6",
"ref_id": "TABREF7"
},
{
"start": 1707,
"end": 1715,
"text": "Figure 2",
"ref_id": "FIGREF1"
},
{
"start": 1857,
"end": 1864,
"text": "Table 6",
"ref_id": "TABREF7"
},
{
"start": 2060,
"end": 2067,
"text": "Table 7",
"ref_id": null
}
],
"eq_spans": [],
"section": "Discussion",
"sec_num": "5"
},
{
"text": "An Arabic-English translation example is shown in Table 8 . Here, the term incidents of murder in apartments was chosen over the baseline's killings inside the flats. Both translations are understandable and the difference in the wording is only based on synonyms. The translation using the discriminative and trigger-based lexicons better matches the reference translation and, thus, reflects a better lexical choice of the content words.",
"cite_spans": [],
"ref_spans": [
{
"start": 50,
"end": 57,
"text": "Table 8",
"ref_id": "TABREF8"
}
],
"eq_spans": [],
"section": "Discussion",
"sec_num": "5"
},
{
"text": "We have presented two lexicon models that use global source sentence context and are capable of predicting context-specific target words. The models have been directly integrated into the decoder and have shown to improve the translation quality of a state-of-the-art phrase-based machine translation system. The first model was a discriminative word lexicon that uses sentence-level features to predict if a word from the target vocabulary should be included in the translation or not. The second model was a trigger-based lexi-Source \u76ee\u524d , \u4e8b\u6545 \u62a2\u9669 \u7ec4 \u6b63\u5728 \u7d27\u6025 \u6062\u590d \u901a\u98ce \u7cfb\u7edf . Baseline at present, the accident and rescue teams are currently emergency recovery ventilation systems. DWL at present, the emergency rescue teams are currently restoring the ventilation system. Triplet at present, the emergency rescue group is in the process of restoring the ventilation system. DWL +Triplet at present, the accident emergency rescue teams are currently restoring the ventilation system. Reference right now, the accident emergency rescue team is making emergency repair on the ventilation system. con that uses triplets to model long-range dependencies in the data. The source word triggers can move across the whole sentence and capture the topic of the sentence and incorporate more finegrained lexical choice of the target words within the decoder. Overall improvements are up to +1% in BLEU and -1.5% in TER on large-scale systems for Chinese-English and Arabic-English. Compared to the inverse IBM model 1 which did not yield consistent improvements, the presented models are valuable additional features in a phrase-based statistical machine translation system. We will test this setup for other language pairs and expect that languages like German where long-distance effects are common can benefit from these extended lexicon models.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6"
},
{
"text": "In future work, we plan to extend the discriminative word lexicon model in two directions: extending context to the document level and feature engineering. For the trigger-based model, we plan to investigate more model variants. It might be interesting to look at cross-lingual trigger models such as p(f |e, f ) or constrained variants like p(f |e, e ) with pos(e ) < pos(e), i.e. the second trigger coming from the left context within a sentence which has already been generated. These Table 7 : Translation example from the GALE web text test set. In this case, the baseline has a better TER but we can observe a corrected content word (remark) for the extended lexicon models. The Chinese source sentence is shown in its segmented form.",
"cite_spans": [],
"ref_spans": [
{
"start": 488,
"end": 495,
"text": "Table 7",
"ref_id": null
}
],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6"
},
{
"text": "extensions could be integrated directly in search as well and would enable the system to combine both directions (standard and inverse) to some extent which was previously shown to help when applying the standard direction p(f |e, e ) as an additional reranking step, cf. (Hasan and Ney, 2009) .",
"cite_spans": [
{
"start": 272,
"end": 293,
"text": "(Hasan and Ney, 2009)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6"
},
{
"text": "Baseline some saudi newspapers have published a number of cases that had been subjected to imprisonment without justification, as well as some killings inside the flats and others. DWL +Triplet some of the saudi newspapers have published a number of cases which were subjected to imprisonment without justification, as well as some incidents of murder in apartments and others. Reference some saudi newspapers have published a number of cases in which people were unjustifiably imprisoned, as well as some incidents of murder in apartments and elsewhere. and Juri Ganitkevitch for their help training the extended lexicon models.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Source",
"sec_num": null
},
{
"text": "http://www.cs.utah.edu/\u02dchal/megam/",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "This material is partly based upon work supported by the Defense Advanced Research Projects Agency (DARPA) under Contract No. HR0011-06-C-0023, and was partly realized as part of the Quaero Programme, funded by OSEO, French State agency for innovation.The authors would like to thank Christian Buck",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgments",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Sequence classification for machine translation",
"authors": [
{
"first": "S",
"middle": [],
"last": "Bangalore",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Haffner",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Kanthak",
"suffix": ""
}
],
"year": 2006,
"venue": "Ninth International Conf. on Spoken Language Processing, Interspeech 2006 -ICSLP",
"volume": "",
"issue": "",
"pages": "1722--1725",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "S. Bangalore, P. Haffner, and S. Kanthak. 2006. Se- quence classification for machine translation. In Ninth International Conf. on Spoken Language Pro- cessing, Interspeech 2006 -ICSLP, pages 1722- 1725, Pitsburgh, PA, September.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Statistical machine translation through global lexical selection and sentence reconstruction",
"authors": [
{
"first": "S",
"middle": [],
"last": "Bangalore",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Haffner",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Kanthak",
"suffix": ""
}
],
"year": 2007,
"venue": "45th Annual Meeting of the Association of Computational Linguistics",
"volume": "",
"issue": "",
"pages": "152--159",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "S. Bangalore, P. Haffner, and S. Kanthak. 2007. Statis- tical machine translation through global lexical se- lection and sentence reconstruction. In 45th Annual Meeting of the Association of Computational Lin- guistics, pages 152-159, Prague, Czech Republic, June.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "The mathematics of statistical machine translation: parameter estimation",
"authors": [
{
"first": "P",
"middle": [
"F"
],
"last": "Brown",
"suffix": ""
},
{
"first": "V",
"middle": [
"J"
],
"last": "Della Pietra",
"suffix": ""
},
{
"first": "S",
"middle": [
"A"
],
"last": "Della Pietra",
"suffix": ""
},
{
"first": "R",
"middle": [
"L"
],
"last": "Mercer",
"suffix": ""
}
],
"year": 1993,
"venue": "Computational Linguistics",
"volume": "19",
"issue": "2",
"pages": "263--312",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "P. F. Brown, V. J. Della Pietra, S. A. Della Pietra, and R. L. Mercer. 1993. The mathematics of statistical machine translation: parameter estimation. Compu- tational Linguistics, 19(2):263-312, June.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "A limited memory algorithm for bound constrained optimization",
"authors": [
{
"first": "R",
"middle": [
"H"
],
"last": "Byrd",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Lu",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Nocedal",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Zhu",
"suffix": ""
}
],
"year": 1995,
"venue": "SIAM Journal on Scientific Computing",
"volume": "16",
"issue": "5",
"pages": "1190--1208",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "R. H. Byrd, P. Lu, J. Nocedal, and C. Zhu. 1995. A limited memory algorithm for bound constrained op- timization. SIAM Journal on Scientific Computing, 16(5):1190-1208.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Improving statistical machine translation using word sense disambiguation",
"authors": [
{
"first": "M",
"middle": [],
"last": "Carpuat",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Wu",
"suffix": ""
}
],
"year": 2007,
"venue": "Joint Conf. on Empirical Methods in Natural Language Processing and Computational Natural Language Learning",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "M. Carpuat and D. Wu. 2007. Improving statistical machine translation using word sense disambigua- tion. In Joint Conf. on Empirical Methods in Nat- ural Language Processing and Computational Nat- ural Language Learning (EMNLP-CoNLL 2007), Prague, Czech Republic, June.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Word sense disambiguation improves statistical machine translation",
"authors": [
{
"first": "Y",
"middle": [
"S"
],
"last": "Chan",
"suffix": ""
},
{
"first": "H",
"middle": [
"T"
],
"last": "Ng",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Chiang",
"suffix": ""
}
],
"year": 2007,
"venue": "45th Annual Meeting of the Association of Computational Linguistics",
"volume": "",
"issue": "",
"pages": "33--40",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Y. S. Chan, H. T. Ng, and D. Chiang. 2007. Word sense disambiguation improves statistical machine trans- lation. In 45th Annual Meeting of the Association of Computational Linguistics, pages 33-40, Prague, Czech Republic, June.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Maximum likelihood from incomplete data via the EM algorithm",
"authors": [
{
"first": "A",
"middle": [
"P"
],
"last": "Dempster",
"suffix": ""
},
{
"first": "N",
"middle": [
"M"
],
"last": "Laird",
"suffix": ""
},
{
"first": "D",
"middle": [
"B"
],
"last": "Rubin",
"suffix": ""
}
],
"year": 1977,
"venue": "Journal of the Royal Statistical Society, Series B",
"volume": "39",
"issue": "1",
"pages": "1--22",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "A. P. Dempster, N. M. Laird, and D. B. Rubin. 1977. Maximum likelihood from incomplete data via the EM algorithm. Journal of the Royal Statistical So- ciety, Series B, 39(1):1-22.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Comparison of extended lexicon models in search and rescoring for SMT",
"authors": [
{
"first": "S",
"middle": [],
"last": "Hasan",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Ney",
"suffix": ""
}
],
"year": 2009,
"venue": "NAACL HLT 2009, Companion Volume: Short Papers",
"volume": "",
"issue": "",
"pages": "17--20",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "S. Hasan and H. Ney. 2009. Comparison of extended lexicon models in search and rescoring for SMT. In NAACL HLT 2009, Companion Volume: Short Pa- pers, pages 17-20, Boulder, Colorado, June.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Triplet lexicon models for statistical machine translation",
"authors": [
{
"first": "J",
"middle": [],
"last": "Hasan",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Ganitkevitch",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Ney",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Andr\u00e9s-Ferrer",
"suffix": ""
}
],
"year": 2008,
"venue": "EMNLP",
"volume": "",
"issue": "",
"pages": "372--381",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hasan, J. Ganitkevitch, H. Ney, and J. Andr\u00e9s-Ferrer. 2008. Triplet lexicon models for statistical machine translation. In EMNLP, pages 372-381, Honolulu, Hawaii, October.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Direct translation model 2",
"authors": [
{
"first": "A",
"middle": [],
"last": "Ittycheriah",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Roukos",
"suffix": ""
}
],
"year": 2007,
"venue": "HLT-NAACL 2007: Main Conference",
"volume": "",
"issue": "",
"pages": "57--64",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "A. Ittycheriah and S. Roukos. 2007. Direct translation model 2. In HLT-NAACL 2007: Main Conference, pages 57-64, Rochester, New York, April.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Models of translational equivalence among words",
"authors": [
{
"first": "D",
"middle": [],
"last": "Melamed",
"suffix": ""
}
],
"year": 2000,
"venue": "Computational Linguistics",
"volume": "26",
"issue": "2",
"pages": "221--249",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "D. Melamed. 2000. Models of translational equiv- alence among words. Computational Linguistics, 26(2):221-249.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "A smorgasbord of features for statistical machine translation",
"authors": [
{
"first": "F",
"middle": [
"J"
],
"last": "Och",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Gildea",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Khudanpur",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Sarkar",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Yamada",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Fraser",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Kumar",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Shen",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Smith",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Eng",
"suffix": ""
},
{
"first": "V",
"middle": [],
"last": "Jain",
"suffix": ""
},
{
"first": "Z",
"middle": [],
"last": "Jin",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Radev",
"suffix": ""
}
],
"year": 2004,
"venue": "",
"volume": "",
"issue": "",
"pages": "161--168",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "F. J. Och, D. Gildea, S. Khudanpur, A. Sarkar, K. Ya- mada, A. Fraser, S. Kumar, L. Shen, D. Smith, K. Eng, V. Jain, Z. Jin, and D. Radev. 2004. A smor- gasbord of features for statistical machine transla- tion. pages 161-168, Boston, MA, May.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "A maximum entropy approach to adaptive statistical language modeling",
"authors": [
{
"first": "R",
"middle": [],
"last": "Rosenfeld",
"suffix": ""
}
],
"year": 1996,
"venue": "Computer Speech and Language",
"volume": "10",
"issue": "3",
"pages": "187--228",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "R. Rosenfeld. 1996. A maximum entropy approach to adaptive statistical language modeling. Computer Speech and Language, 10(3):187-228.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Word triggers and the EM algorithm",
"authors": [
{
"first": "C",
"middle": [],
"last": "Tillmann",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Ney",
"suffix": ""
}
],
"year": 1997,
"venue": "Proc. Special Interest Group Workshop on Computational Natural Language Learning (ACL)",
"volume": "",
"issue": "",
"pages": "117--124",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "C. Tillmann and H. Ney. 1997. Word triggers and the EM algorithm. In Proc. Special Interest Group Workshop on Computational Natural Lan- guage Learning (ACL), pages 117-124, Madrid, Spain, July.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Refined lexicon models for statistical machine translation using a maximum entropy approach",
"authors": [
{
"first": "Garc\u00eda",
"middle": [],
"last": "Varea",
"suffix": ""
},
{
"first": "F",
"middle": [
"J"
],
"last": "Och",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Ney",
"suffix": ""
},
{
"first": "F",
"middle": [],
"last": "Casacuberta",
"suffix": ""
}
],
"year": 2001,
"venue": "ACL '01: 39th Annual Meeting on Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "204--211",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Garc\u00eda Varea, F. J. Och, H. Ney, and F. Casacu- berta. 2001. Refined lexicon models for statistical machine translation using a maximum entropy ap- proach. In ACL '01: 39th Annual Meeting on Asso- ciation for Computational Linguistics, pages 204- 211, Morristown, NJ, USA.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Three models for discriminative machine translation using global lexical selection and sentence reconstruction",
"authors": [
{
"first": "S",
"middle": [],
"last": "Venkatapathy",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Bangalore",
"suffix": ""
}
],
"year": 2007,
"venue": "SSST, NAACL-HLT 2007 / AMTA Workshop on Syntax and Structure in Statistical Translation",
"volume": "",
"issue": "",
"pages": "96--102",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "S. Venkatapathy and S. Bangalore. 2007. Three models for discriminative machine translation us- ing global lexical selection and sentence reconstruc- tion. In SSST, NAACL-HLT 2007 / AMTA Workshop on Syntax and Structure in Statistical Translation, pages 96-102, Rochester, New York, April.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Improvements in dynamic programming beam search for phrase-based statistical machine translation",
"authors": [
{
"first": "R",
"middle": [],
"last": "Zens",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Ney",
"suffix": ""
}
],
"year": 2008,
"venue": "International Workshop on Spoken Language Translation",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "R. Zens and H. Ney. 2008. Improvements in dynamic programming beam search for phrase-based statis- tical machine translation. In International Work- shop on Spoken Language Translation, Honolulu, Hawaii, October.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"type_str": "figure",
"uris": null,
"text": "source target[...] the emergency rescue group is [...] restoring the ventilation system. p(restoring | \u6b63\u5728, \u6062\u590d) = 0.1572 p(emergency | \u7d27\u6025, \u62a2\u9669) = 0.3445 Triggering effect for the example sentence using the triplet lexicon model. The Chinese source sentence is shown in its segmented form. Two triplets are highlighted that have high probability and favor the target words emergency and restoring.",
"num": null
},
"FIGREF1": {
"type_str": "figure",
"uris": null,
"text": "Ranking of words for the example sentence for IBM1, Triplet and DWL model. Ranks are sorted at IBM1, darker colors indicate higher probabilities within the model.",
"num": null
},
"TABREF0": {
"html": null,
"content": "<table><tr><td/><td colspan=\"2\">train C/E -A/E</td><td>nist08 C/A</td></tr><tr><td>Sent. pairs</td><td>7.3M</td><td>4.6M</td><td>1357</td></tr><tr><td colspan=\"4\">Words (M) 185/196 142/139 36K/46K</td></tr><tr><td colspan=\"4\">Vocab. (K) 163/265 351/361 6.4K/9.6K</td></tr></table>",
"type_str": "table",
"num": null,
"text": "GALE Chinese-English corpus statistics including two test sets: newswire and web text."
},
"TABREF1": {
"html": null,
"content": "<table><tr><td>: NIST Chinese-English and Arabic-</td></tr><tr><td>English corpus statistics including the official</td></tr><tr><td>2008 test sets.</td></tr></table>",
"type_str": "table",
"num": null,
"text": ""
},
"TABREF4": {
"html": null,
"content": "<table><tr><td>: Results on the test sets for the NIST 2008</td></tr><tr><td>Chinese-English and Arabic-English task (case-</td></tr><tr><td>insensitive evaluation).</td></tr></table>",
"type_str": "table",
"num": null,
"text": ""
},
"TABREF5": {
"html": null,
"content": "<table/>",
"type_str": "table",
"num": null,
"text": "Translation example from the GALE newswire test set, comparing the baseline and the extended lexicon models given a reference translation. The Chinese source sentence is presented in its segmented form."
},
"TABREF7": {
"html": null,
"content": "<table><tr><td>Source</td><td>\u6211 \u542c \u4e86 \u83b9 \u7684\u8bdd , \u4e50 \u5f97 \u54c8\u54c8\u5927</td></tr><tr><td/><td>\u7b11 .</td></tr><tr><td>Baseline</td><td>i have listened to anna, happy and</td></tr><tr><td/><td>laugh.</td></tr><tr><td>DWL</td><td>i have listened to the remarks,</td></tr><tr><td/><td>happy and laugh.</td></tr><tr><td>Triplet</td><td>i have listened to the music, a roar</td></tr><tr><td/><td>of laughter.</td></tr><tr><td>DWL</td><td>i have listened to the remarks,</td></tr><tr><td>+Triplet</td><td>happy and laugh.</td></tr><tr><td colspan=\"2\">Reference hearing ying's remark, i laughed</td></tr><tr><td/><td>aloud happily.</td></tr></table>",
"type_str": "table",
"num": null,
"text": "The top 10 content words predicted by each model for the GALE newswire example sentence. Original ranks for the related IBM model 1 are given as subscripts for the triplet model."
},
"TABREF8": {
"html": null,
"content": "<table/>",
"type_str": "table",
"num": null,
"text": "Translation example from the NIST Arabic-English test set. The DWL and Triplet models improve lexical word choice by favoring incidents of murder in apartments instead of killings inside the flats. The Arabic source is shown in its segmented form."
}
}
}
} |