File size: 84,941 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 |
{
"paper_id": "O09-1015",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T08:10:50.962189Z"
},
"title": "Improving Translation Fluency with Search-Based Decoding and a Monolingual Statistical Machine Translation Model for Automatic Post-Editing",
"authors": [
{
"first": "Jing-Shin",
"middle": [],
"last": "Chang",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Sheng-Sian",
"middle": [],
"last": "Lin",
"suffix": "",
"affiliation": {},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "The BLEU scores and translation fluency for the current state-of-the-art SMT systems based on IBM models are still too low for publication purposes. The major issue is that stochastically generated sentences hypotheses, produced through a stack decoding process, may not strictly follow the natural target language grammar, since the decoding process is directed by a highly simplified translation model and n-gram language model, and a large number of noisy phrase pairs may introduce significant search errors. This paper proposes a statistical post-editing (SPE) model, based on a special monolingual SMT paradigm, to \" translate\"disfluent sentences into fluent sentences. However, instead of conducting a stack decoding process, the sentence hypotheses are searched from fluent target sentences in a large target language corpus or on the Web to ensure fluency. Phrase-based local editing, if necessary, is then applied to correct weakest phrase alignments between the disfluent and searched hypotheses using fluent target language phrases; such phrases are segmented from a large target language corpus with a global optimization criterion to maximize the likelihood of the training sentences, instead of using noisy phrases combined from bilingually wordaligned pairs. With such search-based decoding, the absolute BLEU scores are much higher than automatic post editing systems that conduct a classical SMT decoding process. We are also able to fully correct a significant number of disfluent sentences into completely fluent versions. The BLEU scores are significantly improved. The evaluation shows that on average 46% of translation errors can be fully recovered, and the BLEU score can be improved by about 26%.",
"pdf_parse": {
"paper_id": "O09-1015",
"_pdf_hash": "",
"abstract": [
{
"text": "The BLEU scores and translation fluency for the current state-of-the-art SMT systems based on IBM models are still too low for publication purposes. The major issue is that stochastically generated sentences hypotheses, produced through a stack decoding process, may not strictly follow the natural target language grammar, since the decoding process is directed by a highly simplified translation model and n-gram language model, and a large number of noisy phrase pairs may introduce significant search errors. This paper proposes a statistical post-editing (SPE) model, based on a special monolingual SMT paradigm, to \" translate\"disfluent sentences into fluent sentences. However, instead of conducting a stack decoding process, the sentence hypotheses are searched from fluent target sentences in a large target language corpus or on the Web to ensure fluency. Phrase-based local editing, if necessary, is then applied to correct weakest phrase alignments between the disfluent and searched hypotheses using fluent target language phrases; such phrases are segmented from a large target language corpus with a global optimization criterion to maximize the likelihood of the training sentences, instead of using noisy phrases combined from bilingually wordaligned pairs. With such search-based decoding, the absolute BLEU scores are much higher than automatic post editing systems that conduct a classical SMT decoding process. We are also able to fully correct a significant number of disfluent sentences into completely fluent versions. The BLEU scores are significantly improved. The evaluation shows that on average 46% of translation errors can be fully recovered, and the BLEU score can be improved by about 26%.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Translation fluency of Machine Translation systems is a serious issue in the current SMT research works. With the research efforts for the past tens of years, the performances are still far from satisfactory. In translating English to Chinese, for instance, the BLEU scores [16] range only between 0.21 and 0.29 [22, 5, 17] , depending on test sets and numbers of reference translations. Such translation quality is extremely disfluent for human readers. We therefore propose a statistical post-editing (SPE) model, based on a special monolingual SMT framework, for improving the fluency and adequacy of translated sentences.",
"cite_spans": [
{
"start": 274,
"end": 278,
"text": "[16]",
"ref_id": "BIBREF15"
},
{
"start": 312,
"end": 316,
"text": "[22,",
"ref_id": "BIBREF21"
},
{
"start": 317,
"end": 319,
"text": "5,",
"ref_id": "BIBREF4"
},
{
"start": 320,
"end": 323,
"text": "17]",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Fluency Problems with Statistical Machine Translations",
"sec_num": "1.1"
},
{
"text": "The classical IBM SMT models [1, 2] formulate the translation problem of a source sentence F as finding the best translation E* from some stack decoded hypotheses, E, such that: The arg max E operation implies to generate candidate target sentences E of F so that the SMT model can score each one, based on the TM and LM scores and select the best candidate. The process of candidate generation is known as the decoding process. The conventional decoding process is significantly affected by the TM and LM scores; only those candidates that satisfy the underlying criteria of the TM and LM will receive high scores.",
"cite_spans": [
{
"start": 29,
"end": 32,
"text": "[1,",
"ref_id": null
},
{
"start": 33,
"end": 35,
"text": "2]",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Fluency Problems with Statistical Machine Translations",
"sec_num": "1.1"
},
{
"text": "\uf028",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Fluency Problems with Statistical Machine Translations",
"sec_num": "1.1"
},
{
"text": "Unfortunately, to make the SMT computationally feasible, the TM and LM are highly simplified. Therefore, the candidates are not really generated based on target language grammar, but based on the model constraints. For instance, the classical SMT model does not prefer word re-ordering with long distance movement. Such candidates are then not generated regardless of the possibility that the target grammar might prefer them.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Fluency Problems with Statistical Machine Translations",
"sec_num": "1.1"
},
{
"text": "There are three directions to improve the translation fluency with the classical SMT model, Equation (1) . Firstly, we can improve the Translation Model (TM) to fit the source-target transfer process. Secondly, we can improve the Language Model (LM) to respect the target language grammar. Finally, we could try to generate better and much more fluent candidates in the decoding process so that the TM and LM can select the real best one from fluent candidates, rather than from junk sentences.",
"cite_spans": [
{
"start": 101,
"end": 104,
"text": "(1)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "LM and Decoding",
"sec_num": "1.2"
},
{
"text": "The research communities normally focus on the TM and LM components by assuming that there are good ways to generate good candidates for scoring. Actually, most attention is paid to the Translation Model (TM); LM and decoding were not gaining the same weight. In particular, people tend to think that the candidate generation process guided by the highly simplified TM and LM will eventually generate good candidates.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "LM and Decoding",
"sec_num": "1.2"
},
{
"text": "Unfortunately, to make the computation feasible, the classical SMT models have very low expressive power in the Translation Model (TM) and Language Model (LM) components. It formulates the TM in terms of the fertility probability, lexical translation probability and distortion probability [1, 2] . A word-based 3-gram model is usually used as the language model (LM). Longer n-grams are used at higher training cost and severe data sparseness.",
"cite_spans": [
{
"start": 290,
"end": 293,
"text": "[1,",
"ref_id": null
},
{
"start": 294,
"end": 296,
"text": "2]",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "LM and Decoding",
"sec_num": "1.2"
},
{
"text": "In fact, the candidates of the target sentence, which are hidden in the arg max E operator, are generated as a stochastic process in most SMT today. Starting from a particular state, the next word is predicted based on a local n-gram window within a distance allowed by the distortion criterion; the possible paths are exploited using stack decoding, beam search or other searching algorithms. The candidates generated in this way thus may be only \" piecewise\"consistent with the target language grammar, but may not be really globally grammatical or fluent. This means that the TM and LM are not scoring a complete sentence but some segments pasted by the n-gram LM. It is then not likely to be fluent all the time.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "LM and Decoding",
"sec_num": "1.2"
},
{
"text": "This decoding process therefore sometimes falls into the \" garbage-in and garbage-out\" situation. No matter how well-formulated the TM and LM may be, if the stochastically generated candidates do not include the correct and fluent translation, the system will eventually deliver a garbage output, that is, a disfluent sentence, as the best one. This kind of error is known as searching error. Because the TM and LM have limited expressive power to describe the real criteria that carry the generation process, the decoding process might only generate noisy sentence segments and thus disfluent sentences for scoring. This could lead to bad performance in terms of BLEU score or human judgments.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "LM and Decoding",
"sec_num": "1.2"
},
{
"text": "Phrase-based SMT had partially resolved the expressive power issue of TM and LM by using longer word sequences. However, the acquisition of \" phrases\"has its own problems. In particular, most phrase-based SMT acquires the phrase pairs by conducting bilingual word alignment first. Adjacent words are then connected in some heuristic ways [12, 13, 14, 15] , which do not have direct link with the source or target grammar, to form the \" phrases\" . The phrases generated in this way normally do not satisfy any global optimization criteria related to the target grammar, such as maximizing the likelihood of the target language sentences. The quality of such phrases is therefore greatly affected by the word alignment accuracy; and, the phrases for the target language side may not really respect the target grammar. Under such circumstances, a huge number of noisy \" phrases\"will be introduced and significantly enlarge the searching space. The stochastically generated phrase sequences thus may not correspond to good candidate sentences either.",
"cite_spans": [
{
"start": 338,
"end": 342,
"text": "[12,",
"ref_id": "BIBREF11"
},
{
"start": 343,
"end": 346,
"text": "13,",
"ref_id": "BIBREF12"
},
{
"start": 347,
"end": 350,
"text": "14,",
"ref_id": "BIBREF13"
},
{
"start": 351,
"end": 354,
"text": "15]",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "LM and Decoding",
"sec_num": "1.2"
},
{
"text": "To summarize, the application of word-for-word or phrase-to-phrase translation (with \" noisy\"phrases) plus a little bit local word/phrase re-ordering in classical SMT might not generate fluent target sentences that respect the target grammar. In particular, many target specific lexical items and morphemes cannot be generated through this kind of models. If they do, they may be generated in very special ways. This could be a significant reason why the SMT models do not work well after the long period of research.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "LM and Decoding",
"sec_num": "1.2"
},
{
"text": "The implication is that we might have to examine the arg max E operation, that is, the decoding or searching process, in the classical SMT models more carefully. We should try decoding method that respect target grammar more, instead of following the criteria set forth by the TM and LM of the SMT model, which encode highly simplified version of the target grammar. Only with a decoding process that respect the target grammar, will the system generate fluent candidates at the first place before submitting the candidates to the TM and LM for scoring.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "LM and Decoding",
"sec_num": "1.2"
},
{
"text": "Furthermore, a phrase-based language model, instead of word-based n-gram model for the target side may improve the fluency of machine translation further since more context words can be consulted, if the \" phrases\"are not noisy. To avoid a huge number of noisy source-dependent phrases that might be harmful for fluency and searching, such phrases may better be trained from a target corpus, instead of being acquired from bilingually wordaligned chunks.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "LM and Decoding",
"sec_num": "1.2"
},
{
"text": "Instead of developing new models for the TM and LM, an alternative to improve the translation fluency is to cascade an Automatic Post-Editing (APE) module to the translation output of an MT/SMT system. While the classical SMT models may not be suitable for directly generating fluent translation, due to the limited expressive power of the TM and LM and search errors of the decoding process, an SMT or its variant may be sufficient for reranking hypotheses in the automatic post editing purposes, if appropriate hypotheses generation mechanism is available. Actually, we can regard a post-editing process as a translation process from disfluent sentence to fluent sentence. This is particularly true if the disfluency is limited to local editing operations like insertion of target specific morphemes, deletion of source-specific function words, and lexical substitution from many possible lexical choices. These kinds of errors are often seen in MT/SMT systems. Inspired by the above ideas, this paper propose a statistical post-editing (SPE) model based on a monolingual SMT paradigm for improving the translation fluency of an MT system, instead of improving the TM directly.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Statistical Post-Editing Model Based on Monolingual SMT",
"sec_num": "1.3"
},
{
"text": "In this SPE model, the searching or decoding is a fluency-based search. We search fluent translations, based on the lexical hints of the disfluent sentence, from a large target text corpus or from the Web. Therefore, all candidates will be fluent ones. The best hypotheses reranked best by the SPE model will then serve as the post-edited version of the disfluent sentence. Sometimes, a searched sentence may not have a high translation score to justify itself as an appropriate translation. For instance, the target sentence pattern may be correct but different lexical choices have been made. In this case, automatic local editing is applied to the weakest alignments to incrementally patch the target sentence pattern with right target lexical items. By combining the grammatical (and fluent) sentence pattern of the searched sentence and the right lexical items from the disfluent sentence, the disfluent translation could be repaired to a fluent one incrementally. This may include some local insertion, deletion and lexical substitution operations over phrase pairs that are unlikely to be translation of each other.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Statistical Post-Editing Model Based on Monolingual SMT",
"sec_num": "1.3"
},
{
"text": "To really improve the fluency incrementally, the local editing process is applied in a manner that will monotonically increase the likelihood of the incrementally repaired sentence. To respect the target grammar further, the repair is phrase-based. In other words, phrasebased n-gram language model (n=1) is used in the translation score so that the likelihood of the repaired target sentence is incrementally increased during the local editing process.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Statistical Post-Editing Model Based on Monolingual SMT",
"sec_num": "1.3"
},
{
"text": "In parallel with the development of our work, a few APE systems were also proposed [7, 20, 21, 8] with good results. Publicly available SMT systems (like Portage PBMT, Moses, etc.) are used directly as the post-editing module. They are trained using human post-edited target sentences with their un-edited MT outputs to learn the translation knowledge between disfluent (' source' ) and fluent (' target' ) sentences [20] . Alternatively, they may be trained using standard parallel corpora (Europarl, News Commentary, Job Bank, Hansard, etc.) where the disfluent sentences are generated using a rule-based MT (like SYSTRAN) or other SMT [21] .",
"cite_spans": [
{
"start": 83,
"end": 86,
"text": "[7,",
"ref_id": "BIBREF6"
},
{
"start": 87,
"end": 90,
"text": "20,",
"ref_id": "BIBREF19"
},
{
"start": 91,
"end": 94,
"text": "21,",
"ref_id": "BIBREF20"
},
{
"start": 95,
"end": 97,
"text": "8]",
"ref_id": "BIBREF7"
},
{
"start": 417,
"end": 421,
"text": "[20]",
"ref_id": "BIBREF19"
},
{
"start": 638,
"end": 642,
"text": "[21]",
"ref_id": "BIBREF20"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Statistical Post-Editing Model Based on Monolingual SMT",
"sec_num": "1.3"
},
{
"text": "Therefore, these works require substantial human post-editing costs to train the SMT. Or they need a sizable parallel corpus for training, which may not be available to many language pairs. In addition, it requires an RBMT or SMT pre-trained for translating the source corpus, which may not be available to many language pairs. Most importantly, these frameworks use the same decoding process as well as the TM and LM of the original SMT to generate their post-editing hypotheses. Therefore, the previously discussed performance issues that apply to classical SMT will also apply to such APE modules. The cascade of an SMT as an APE module might imply the use of a system with low BLEU performance to correct the outputs with low BLEU scores. The improvement could thus be substantially limited. This may be seen from the fact that the contribution of the APE becomes negligible as the training data is increased [21] .",
"cite_spans": [
{
"start": 913,
"end": 917,
"text": "[21]",
"ref_id": "BIBREF20"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Statistical Post-Editing Model Based on Monolingual SMT",
"sec_num": "1.3"
},
{
"text": "In contrast, we discard the stochastic decoding process, which might generate disfluent hypotheses, but search a large corpus for highly similar sentences to the disfluent sentence, and thus will have raw hypotheses with high BLEU scores. Additional local editing will further improve the fluency. Furthermore, our proposal can generate interesting error patterns automatically using the target language corpus alone. Therefore, the APE module can be constructed without a real MT system (although it would be better to have one in order to correct the specific errors of a specific system.). The following sections will discuss the formulation in more details.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Statistical Post-Editing Model Based on Monolingual SMT",
"sec_num": "1.3"
},
{
"text": "In our work, we propose to adopt a Statistical Post-Editing (SPE) Model to translate disfluent s e n t e n c e s i n t o f l u e n t v e r s i o n s . S u c h a s y s t e m c a n b e r e g a r d e d a s a \" d i s f l u e n t -to-f l u e n t \" S MT . As will be seen later, it can be trained with a Monolingual SMT Model. Given a disfluent sentence E' translated form a source sentence F, the automatic post-editing problem can be formulated as finding the most fluent sentence E* from some candidate sentences E such that:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Problem Formulation for SPE",
"sec_num": "2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "\uf028 \uf029 \uf028 \uf029 \uf028\uf029 E E E E E E E E Pr | ' Pr max arg ' | Pr max arg * \uf03d \uf03d",
"eq_num": "(2)"
}
],
"section": "Problem Formulation for SPE",
"sec_num": "2"
},
{
"text": "As usual, we will refer Pr(E' |E) as the translation model (TM), and Pr(E) as the language model (LM) of the SPE model. We thus encountered the same SMT problems to formulate the TM, LM and the decoding (or searching) process.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Problem Formulation for SPE",
"sec_num": "2"
},
{
"text": "The automatic post-editing problem is intuitively easier than SMT since we can assumes that the disfluency is due to some local editing errors, such as mis-insertion or mis-deletion of function words, and wrong lexical choices. Under this assumption, we can formulate the TM as:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Order-Preserved Translation Model",
"sec_num": "2.1"
},
{
"text": "\uf028 \uf029 \uf028 \uf029 \uf028 \uf029 \uf028 \uf029 \uf028 \uf029 \uf028 \uf029 ' ' Pr ' | Pr ', | max Pr ', | Pr ', | Pr | p s p A A s p p E A E E E E A E E A E E A E E E \uf03d \uf03d \uf0bb \uf0bb \uf03d \uf0e5 \uf0d5 (3)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Order-Preserved Translation Model",
"sec_num": "2.1"
},
{
"text": "In Eqn. (3), phrase-aligned phrase pairs are represented by E' p and Ep for the disfluent and fluent versions, respectively. We assume that the most likely alignment A s , among all generic alignment pattern A, between E'and E is an \" order-preserved\"or \" sequential\" alignment between their constituents. We further assume that this most likely alignment has much higher probability than other alignments such that we don' t have to sum over all generic alignment patterns. In the post-editing context, this assumption may be reasonable if the disfluency results from simple local editing operations. In particular, if we are using phrase-based alignment, the word order within the phrases can be ignored. The order preservation assumption will be even more reasonable. We therefore assume that the TM is the product of the probabilities of sequentially aligned target phrase pairs. The phrase segmentation model for dividing E or E'into phrases will be further detailed later when discussing the target phrase-based LM. Given the segmented phrases, the best sequential alignment can easily be found using a standard dynamic programming algorithm for finding the \" shortest path\" .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Order-Preserved Translation Model",
"sec_num": "2.1"
},
{
"text": "The TM for the SPE model is special in that the training corpus can be easily acquired from a large monolingual corpus with fluent target sentences. Generating a disfluent version of the fluent monolingual corpus automatically based on some error model of the translation process will make this possible. One can then easily acquire the model parameters for translating disfluent sentences into fluent ones through a similar training process for a standard SMT. In comparison with standard SMT training, which requires a parallel bilingual corpus, the monolingual corpus is much easier to acquire.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Order-Preserved Translation Model",
"sec_num": "2.1"
},
{
"text": "To respect the fluency of the target language in the decoding process, the language model score Pr(E) should be evaluated based on long target language phrases, Ep, instead of target words. The \" phrases\"should also be defined independent of source-language in order not to introduce a huge number of noisy phrases as PBSMT normally did. The proposed LM for the current SPE, which is responsible for selecting fluent target segments, is therefore a phrasebased unigram model, instead of the widely used word-based n-gram model. In other words, we have",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Target Phrase-Based Language Model",
"sec_num": "2.2"
},
{
"text": "\uf028 \uf029 \uf028 \uf029 Pr Pr Ep E E Ep \uf0ce \uf03d \uf0d5 .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Target Phrase-Based Language Model",
"sec_num": "2.2"
},
{
"text": "To avoid source-language dependency, we also decided not to define target phrases in terms of chunks of bilingually aligned words. Instead, the best target phrases are directly trained from the monolingual target corpus by optimizing the phrase-based unigram model. In other words, the best phrase sequence * p \uf072 for an n-word sentence 1 n w , will be the sequence, among all possible phrase segmentation, 1 m p , such that:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Target Phrase-Based Language Model",
"sec_num": "2.2"
},
{
"text": "\uf028 \uf029 \uf028 \uf029 1 1 * 1 1 arg max Pr | arg max Pr m m m n i p p i p p w p \uf03d \uf03d \uf0d5 \uf072 .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Target Phrase-Based Language Model",
"sec_num": "2.2"
},
{
"text": "Fortunately, extracting monolingual phrases using the phrase-based uni-gram model can be done easily. The training method is just like the word based uni-gram word segmentation model [4] , which was frequently used in Chinese word segmentation tasks. Unsupervised training is easy for this. Upon convergence, a set of well-formed phrases can be acquired. (This set of phrases will be called a phrase example base, PEB. Phrases in the PEB will be used later in the Local Editing Algorithm for post-editing.)",
"cite_spans": [
{
"start": 183,
"end": 186,
"text": "[4]",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Target Phrase-Based Language Model",
"sec_num": "2.2"
},
{
"text": "Since a phrase trained in this way can be longer than a 3-gram pattern, the modeling error could be reduced to some extend. Furthermore, the number of such phrases will be much smaller than those randomly combined phrases acquired from word-aligned word chunks. As a result, the estimation error due to data sparseness will be significantly reduced too. Unlike the rare parallel bilingual training corpus, the amount of such target language corpora is extremely large. Therefore, fluent phrases can be extracted easily. With phrases as the basic lexical unit, SPE model will reduces to Since a phrase can cover more than 3 words, the selected phrases might be more fluent than word trigrams. Such phrases will fit target grammar better and therefore will prefer more fluent target sentences in general.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Target Phrase-Based Language Model",
"sec_num": "2.2"
},
{
"text": "One key issue that causes disfluency is the decoding process used in classical SMT. Most decoding process regard target sentence generation as a stochastic process, and only local context of finite length window is consulted while decoding. Therefore, the target sentences generated in this way are usually not fluent. Our work proposes to search fluent translation candidates from a huge target sentence base or from web documents, instead of using traditional decoding methods to generate the translation candidates. Since the large corpus and the Web documents are produced by native speakers, the target sentences thus searched are most likely fluent with high BLEU scores.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Search-Based Decoding for Fluency",
"sec_num": "2.3"
},
{
"text": "Our current work simply used a heuristic matching score to extract a set of candidate sentences for a disfluent sentence. The candidates are then re-ranked using the translation score defined by the SPE model. The best candidate will be regarded as the post-edited version of the disfluent sentence if the translation score is higher than a threshold. Otherwise, it will be locally edited to incrementally increase its translation score. The matching score is simply the number of identical word tokens in two sentences, which is normalized by the average length of the two sentences. In other words, it is the percentage of word matches between two sentences.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Search-Based Decoding for Fluency",
"sec_num": "2.3"
},
{
"text": "We searched the candidate translations from the Academia Sinica Word Segmentation Corpus, ASWSC-2001 [6] , as well as Chinese webpages indexed by Google. (We assume that the target language is Chinese.) Different query strings will result in different returned pages. Totally, we have tried 4 models for searching:",
"cite_spans": [
{
"start": 101,
"end": 104,
"text": "[6]",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Search-Based Decoding for Fluency",
"sec_num": "2.3"
},
{
"text": "(1) Model C: search the corpus (only) for Top-N hypotheses (N=20). (The length difference must not be greater than two words.) (2) Model C+W: search the corpus and the web for additional N hypotheses by submitting the complete disfluent target sentence as-is to Google. (3) Model C+W+P: including partial matches against substrings of the disfluent target sentence, where 1~L-1 words in the disfluent sentence are successively deleted and then submitted as query strings to the search engine. (L: number of words in disfluent sentence) (4) Model C+W+Q: adjacent words in the deleted disfluent sentence are quoted as a single query token before submission so that the search engine will match more exactly.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Search-Based Decoding for Fluency",
"sec_num": "2.3"
},
{
"text": "Even with such a heuristic search, a substantial number of fluent sentences similar to the disfluent sentences can be found for re-ranking and local editing.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Search-Based Decoding for Fluency",
"sec_num": "2.3"
},
{
"text": "If exact translation is found during searching, the searching process itself is exactly a perfect translation process. If highly similar sentences are found, simple lexical substitution or automatic post-editing [9, 11] might patch the searched fluent sentences into correct translations. Some previous works for automatic post editing have been restricted to special function words, such as the English article ' the/a' [9, 10] , the Japanese case markers and Chinese classifier or particle ' de' [18] . The automatic post-editing model here is intended to resolve general editing errors that are frequently made by a machine translation system. Briefly, the best sentence eb E * in the searched candidates will be output as the translation of the disfluent translation E'if the translation score associated with the SPE model is higher than a threshold. (The set of candidate translation sentences is called its example base, thus the subscript ' eb' .) Otherwise, the automatic local editing algorithm will find the weakest phrase alignments and fix them one-by-one to maximize the translation score.",
"cite_spans": [
{
"start": 212,
"end": 215,
"text": "[9,",
"ref_id": "BIBREF8"
},
{
"start": 216,
"end": 219,
"text": "11]",
"ref_id": "BIBREF10"
},
{
"start": 421,
"end": 424,
"text": "[9,",
"ref_id": "BIBREF8"
},
{
"start": 425,
"end": 428,
"text": "10]",
"ref_id": "BIBREF9"
},
{
"start": 498,
"end": 502,
"text": "[18]",
"ref_id": "BIBREF17"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Local Editing",
"sec_num": "2.4"
},
{
"text": "An alignment phrase pair <Ep' , Ep> is said to be \" weak\"if its local alignment score Pr(Ep' |Ep) x Pr(Ep) is small and thus contributes little to the global translation score for the sentence pair <E' , E>. When the weakest pair, (Ep' -| Ep-) with the lowest local alignment score is identified, we should try to replace Ep-, the \" most questionable phrase\"in the fluent (yet incorrect) example sentence E, with some candidates that would make the patched example sentence more likely to be the translation of E' .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Local Editing",
"sec_num": "2.4"
},
{
"text": "There are some reasons why the alignment (Ep' -| Ep-) is the weakest. First of all, Epmight not be the right phrase, and should be replaced by Ep' -to make the fluent sentence E also the correct translation of E' . Second, Ep' -might not be the correct translation of some source phrase. In this case, the most likely translation(s) of Ep' -, called Ep+, should be used to replace Ep-. Third, Ep-is a more appropriate phrase than Ep+. In this case, it should be retained and next weakest alignment pair be repaired.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Local Editing",
"sec_num": "2.4"
},
{
"text": "As a result, potential candidates for replacing Ep-will include Ep' -, Ep+ and Ep-itself. The best substitution will be the phrase that maximizes Pr(Ep' |Ep) x Pr(Ep). Actually, many phrases in the PEB can be a more fluent version of Ep' -. Currently, the 20 best matches will play the role of Ep+ during local editing. And the local editing algorithm will successively edit weaker alignments until the (monotonically increasing) translation score is above some threshold. The algorithm is outlined as follows. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Local Editing",
"sec_num": "2.4"
},
{
"text": "Note that, local editing is applied only to a local region of the example sentence based on the disfluent sentence. Intuitively, those sentences searched from a text corpus or from the Web corpus will be much more fluent than stochastically combined sentences from the SMT decoding module. Even if local editing is required, the repair will be quite local. The search space for repairing will be significantly constrained by words in the most likely example sentence. Such a searching and local editing combination can thus be regarded as a constrained decoding. The searching error can thus be reduced significantly in comparison with the large search space of the decoding process of a typical SMT.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Constrained Decoding",
"sec_num": null
},
{
"text": "The TM parameters can actually be trained from an E'-to-E monolingual Machine Translation System, where E' can be derived by applying to E some commonly found editing operations in the SMT translation process. The operations might include the insertion of target specific lexicon, deletion of source specific lexicon, local reordering of words and substitution of lexical items.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Generating Faulty Sentences",
"sec_num": "2.5"
},
{
"text": "In the current work, we apply three kinds of editing operations to the fluent sentences in a monolingual corpus to simulate frequently found errors in an MT system. The fluent and its disfluent versions are then phrase segmented so that the sentences are represented by phrase tokens (instead of word tokens). Such fluent-disfluent (E-E' ) target sentence pairs are then trained using the GIZA++ alignment tools [12, 13, 14, 15] . Upon convergence, the translation model between the sentences to be post-edited and their correct translation can readily be acquired.",
"cite_spans": [
{
"start": 412,
"end": 416,
"text": "[12,",
"ref_id": "BIBREF11"
},
{
"start": 417,
"end": 420,
"text": "13,",
"ref_id": "BIBREF12"
},
{
"start": 421,
"end": 424,
"text": "14,",
"ref_id": "BIBREF13"
},
{
"start": 425,
"end": 428,
"text": "15]",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Generating Faulty Sentences",
"sec_num": "2.5"
},
{
"text": "The three editing operations include:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Generating Faulty Sentences",
"sec_num": "2.5"
},
{
"text": "(1) Insertion: The insertion errors will occur when an MT system translates a source word into a target word while it should not be translated. For instance, the English infinitive \" to\"need not be translated into any Chinese word most of the time. But the bilingual dictionary may indicate the possibility to translate it into \" \u53bb \"(chu). We therefore automatically insert the Chinese words to simulate such an error.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Generating Faulty Sentences",
"sec_num": "2.5"
},
{
"text": "(2) Deletion: The deletion error occurs when a target specific word is not generated in the translation. For instance, the Chinese classifiers have no correspondence in the English language. We therefore delete the following classifiers from fluent Chinese sentences to create instances with deletion errors:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Generating Faulty Sentences",
"sec_num": "2.5"
},
{
"text": "' \u500b' , ' \u96bb' , ' \u679d' , ' \u4f4d' , ' \u9846' , ' \u68f5' .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Generating Faulty Sentences",
"sec_num": "2.5"
},
{
"text": "(3) Substitution: When a translation system chooses a wrong lexical item, a typical substitution error will occur. To simulate the substitution errors, Chinese words in the fluent sentences are lookup against an English-Chinese dictionary. Chinese words that are also the translation of the English word are then substituted to simulate the substitution error. For instance, ' \u554f\u984c' is a Chinese translation for the English word ' problem' . But ' problem' also has other translations, like ' \u7fd2\u984c' and ' \u7591\u96e3' . These words are therefore used to simulate the substitution errors. In our simulation, the top-30 most frequently used Chinese words are adopted to simulate the substitution errors.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Generating Faulty Sentences",
"sec_num": "2.5"
},
{
"text": "With disfluent sentences created from fluent sentences with the above frequently encountered translation errors, an automatic statistical post-editing model can readily be trained using state-of-the-art alignment tools.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Generating Faulty Sentences",
"sec_num": "2.5"
},
{
"text": "To see the performance of the current SMT-based SPE model, about 300,000 word segmented Chinese sentences from the Academia Sinica [6] was used as our target sentence corpus. The corpus has about 2,450,000 word tokens, and the vocabulary size is about 83,000 word types. 10% of the sentences are used as the test set and 90% are used for training. The 3 types of errors are applied to the testing sentences independently. For each error type, 100 sentences are randomly selected for evaluating automatic post editing.",
"cite_spans": [
{
"start": 131,
"end": 134,
"text": "[6]",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "3"
},
{
"text": "The performance is evaluated in terms of two criteria. The first criterion is the number (percentage) of fully corrected disfluent sentences from the test set. By fully corrected, we mean that the sentence corrected by the statistical post editing (SPE) system is completely the same as its original fluent version. translation. With the SPE, the local editing algorithm tries to maximize the translation score for each local editing. It therefore improves the translation fluency incrementally. Since the TM can be trained from an automatically generated fluent-disfluent parallel corpus, training such a system is easy. The evaluation shows that, on average, 46% of translation errors can be fully recovered, and the BLEU score can be improved by about 26%. The absolute BLEU is also high with the search-based decoding process in comparison with conventional decoding process.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "3"
}
],
"back_matter": [
{
"text": "Note that, even with the very simple minded searching method, the SPE was able to correct, on average, about 48% of the faulty sentences to their fluent version if the search space is sufficiently large (with the C+W+Q searching model). The performance increases with the search space. And the performance is increased at most by 62%, 121% and 17.5 %, respectively for the substitution, deletion and insertion errors when the Web corpus is included to the search space. Obviously, the substitution is the hardest to resolve while insertion error seems to be easier to resolve.The second evaluation criterion is the improvement in the BLUE score with respect to the un-corrected test sentence. Table 2 shows the BLEU scores for the various searching models. The first column labeled as E' (ts) lists the BLEU scores for the test sentences that has not been post-edited. By searching for fluent translation and applying local editing, the BLEU scores are improved with increasing search space. The best performance is to increase the BLEU scores by 15%, 38% and 26% respectively for the three types of errors. On average, the improvement is about 26%, which is substantial. On the other hand, the absolute changes are 9. Note that, with search-based decoding, the absolute BLEU scores are much higher than automatic post editing systems that simply cascade a classical SMT module to the output of an MT/SMT [20, 21, 8] . Although the experiment settings are not the same and thus cannot be compared directly, the results to have higher absolute BLEU scores can be expected since searched sentences are almost always fluent, whether they are post-edited or not.Obviously, with the same training corpus, the search space and the searching method play important roles in improving the performance. The inclusion of the web corpus does improve the performance significantly. It was reported in [19] that well formulated query strings can effectively improve searching accuracy. Therefore, by using better searching strategy, part of the translation problems for fluent translation might be resolved as a searching and automatic post-editing problems. Currently, a statistical searching model specific for the fluency-based decoding is being developed.",
"cite_spans": [
{
"start": 1405,
"end": 1409,
"text": "[20,",
"ref_id": "BIBREF19"
},
{
"start": 1410,
"end": 1413,
"text": "21,",
"ref_id": "BIBREF20"
},
{
"start": 1414,
"end": 1416,
"text": "8]",
"ref_id": "BIBREF7"
},
{
"start": 1888,
"end": 1892,
"text": "[19]",
"ref_id": "BIBREF18"
}
],
"ref_spans": [
{
"start": 693,
"end": 700,
"text": "Table 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "annex",
"sec_num": null
},
{
"text": "In this paper, we propose not to generate sentence hypotheses for APE systems by using conventional SMT decoding process, since such a decoding process tends to lead to an openended search space. It is not easy to generate fluent sentence hypotheses under such circumstances due to the large search error. We propose to search sentence hypotheses, from a large target text corpus or from the web, based on the words in the disfluent translations, since the potential candidates will mostly be fluent. A statistical post-editing model is also proposed to re-rank the searched sentences, and a local editing algorithm is proposed to automatically recover the translation errors when the searched sentence is not a good",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Concluding Remarks",
"sec_num": "4"
}
],
"bib_entries": {
"BIBREF1": {
"ref_id": "b1",
"title": "T h e mathematics of statistical machine translation: Parameter e s t i ma t i o n",
"authors": [
{
"first": "Peter",
"middle": [
"F"
],
"last": "Brown",
"suffix": ""
},
{
"first": "A",
"middle": [
"Della"
],
"last": "Stephen",
"suffix": ""
},
{
"first": "Vincent",
"middle": [
"J"
],
"last": "Pietra",
"suffix": ""
},
{
"first": "R",
"middle": [
"L"
],
"last": "Della Pietra",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Mercer",
"suffix": ""
}
],
"year": 1993,
"venue": "Computational Linguistics",
"volume": "19",
"issue": "2",
"pages": "263--311",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Brown, Peter F., Stephen A. Della Pietra, Vincent J. Della Pietra, and R. L. Mercer, \" T h e mathematics of statistical machine translation: Parameter e s t i ma t i o n . \" Computational Linguistics, 19(2):263-311, 1993.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "A Chinese-to-Chinese Statistical Machine Translation Model for Mining Synonymous Simplified-T r a d i t i o n a l C h i n e s e T e r ms",
"authors": [
{
"first": "Jing-Shin",
"middle": [],
"last": "Chang",
"suffix": ""
},
{
"first": "Chun-Ka I",
"middle": [],
"last": "Ku N G",
"suffix": ""
}
],
"year": 2007,
"venue": "Proceedings of Machine Translation Summit XI",
"volume": "",
"issue": "",
"pages": "10--14",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Chang, Jing-Shin and Chun-Ka i Ku n g , \" A Chinese-to-Chinese Statistical Machine Translation Model for Mining Synonymous Simplified-T r a d i t i o n a l C h i n e s e T e r ms , \" Proceedings of Machine Translation Summit XI, pages 81-88, Copenhagen, Denmark, 10-14, September, 2007.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Statistical Models for Word Segmentation and Unknown Word Resolution",
"authors": [
{
"first": "Tung-Hui",
"middle": [],
"last": "Chiang",
"suffix": ""
},
{
"first": "Jing-Shin",
"middle": [],
"last": "Chang",
"suffix": ""
},
{
"first": "Ming-Yu",
"middle": [],
"last": "Lin",
"suffix": ""
},
{
"first": "Keh-Yih",
"middle": [],
"last": "Su",
"suffix": ""
}
],
"year": 1992,
"venue": "Proceedings of ROCLING-V",
"volume": "",
"issue": "",
"pages": "123--146",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Chiang, Tung-Hui, Jing-Shin Chang, Ming-Yu Lin and Keh-Yih Su, \" Statistical Models for Word Segmentation and Unknown Word Resolution,\"Proceedings of ROCLING-V, pp. 123-146, Taipei, Taiwan, R.O.C., 1992.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "A Hi e r a r c h i c a l P h r a s e -Based Model for Statistical Machine Translation",
"authors": [
{
"first": "David",
"middle": [],
"last": "Chiang",
"suffix": ""
}
],
"year": 2005,
"venue": "Proc. ACL-2005",
"volume": "",
"issue": "",
"pages": "263--270",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Chiang, David, \" A Hi e r a r c h i c a l P h r a s e -Based Model for Statistical Machine Translation,\" Proc. ACL-2005, pages 263-270, 2005.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Academia Sinica Word Segmentation Corpus, ASWSC-2001, (\u4e2d\u7814\u9662\u4e2d\u6587 \u5206 \u8a5e \u8a9e\u6599\u5eab )",
"authors": [],
"year": 2001,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "CKIP 2001, Academia Sinica Word Segmentation Corpus, ASWSC-2001, (\u4e2d\u7814\u9662\u4e2d\u6587 \u5206 \u8a5e \u8a9e\u6599\u5eab ), Chinese Knowledge Information Processing Group, Acdemia Sinica, Tiapei, Taiwan, ROC, 2001.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "S t a t i s t i c a l P o s t -Editing on SYSTRANS's Rule-Based Translation System",
"authors": [
{
"first": "L",
"middle": [],
"last": "Dugast",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Senellart",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Koehn",
"suffix": ""
}
],
"year": 2007,
"venue": "Proceedings of the Second Workshop on Statistical Machine Translation, 2nd WSMT",
"volume": "",
"issue": "",
"pages": "220--223",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Dugast, L., J. Senellart, P. Koehn, \" S t a t i s t i c a l P o s t -Editing on SYSTRANS's Rule-Based Translation System,\" Proceedings of the Second Workshop on Statistical Machine Translation, 2nd WSMT, pp. 220-223, Prague, Czech Republic, June 2007.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Do ma h Automatic Post-Editing",
"authors": [
{
"first": "P",
"middle": [],
"last": "Isabelle",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Goutter",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Simard",
"suffix": ""
}
],
"year": 2007,
"venue": "Proceedings of MT Summit XI",
"volume": "",
"issue": "",
"pages": "10--14",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Isabelle, P., G. Goutter, M. Simard, \" Do ma h Automatic Post-Editing,\"Proceedings of MT Summit XI, pp. 255-261, Copenhagen, Denmark, 10-14 Sept. 2007.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Au t o ma t e d Post-Editing of Documents",
"authors": [
{
"first": "Kevin",
"middle": [],
"last": "Knight",
"suffix": ""
},
{
"first": "Ishwar",
"middle": [],
"last": "Chander",
"suffix": ""
}
],
"year": 1994,
"venue": "Proceedings of the Twelfth National Conference on Artificial Intelligence",
"volume": "",
"issue": "",
"pages": "779--784",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Knight, Kevin, and Ishwar Chander, \" Au t o ma t e d Post-Editing of Documents,\"in Proceedings of the Twelfth National Conference on Artificial Intelligence, pp. 779-784, CA, USA, 1994.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Automatic Article Restoratio n",
"authors": [
{
"first": "J",
"middle": [],
"last": "Lee",
"suffix": ""
}
],
"year": 2004,
"venue": "Proc. HLT-NAACL 2004 Student Research Workshop",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Lee, J., \" Automatic Article Restoratio n , \"in Proc. HLT-NAACL 2004 Student Research Workshop, Boston, MA, 195-200, May, 2004.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Au t o ma t i n g P o s t -Editing to Improve MT S y s t e ms",
"authors": [
{
"first": "Ariadna",
"middle": [],
"last": "Llitj\u00f3s",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Font\u00f3s",
"suffix": ""
}
],
"year": 2006,
"venue": "AMTA",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Llitj\u00f3s, Ariadna Font\u00f3s, a n d J a i me C a r b o n e l l , \" Au t o ma t i n g P o s t -Editing to Improve MT S y s t e ms , \" i n Automated Post-Editing Workshop, AMTA, Boston, USA, August 12, 2006.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Improved Alignment Models for Statistical Machine Translation",
"authors": [
{
"first": "Franz",
"middle": [],
"last": "Och",
"suffix": ""
},
{
"first": "Christoph",
"middle": [],
"last": "Josef",
"suffix": ""
},
{
"first": "Hermann",
"middle": [],
"last": "Tillmann",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Ney",
"suffix": ""
}
],
"year": 1999,
"venue": "Proc. EMNLP/WVLC",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Och, Franz Josef, Christoph Tillmann, and Hermann Ney, \" Improved Alignment Models for Statistical Machine Translation,\" in Proc. EMNLP/WVLC, 1999.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "I n Proc. COLING ' 0 0 : T h e 1 8 t h I n t",
"authors": [
{
"first": "Franz",
"middle": [
"Josef"
],
"last": "Och",
"suffix": ""
},
{
"first": "Hermann",
"middle": [],
"last": "Ney",
"suffix": ""
}
],
"year": 2000,
"venue": "",
"volume": "",
"issue": "",
"pages": "1086--1090",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Och, Franz Josef and Hermann Ney, \" Acomparison of alignment models for statistical ma c h i n e t r a n s l a t i o n . \" I n Proc. COLING ' 0 0 : T h e 1 8 t h I n t e r n a t i o n a l C o n f e r e n c eon Computational Linguistics, pages 1086-1090, Saarbr\u00fccken, Germany, August, 2000.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "I mp s",
"authors": [
{
"first": "Franz",
"middle": [
"Josef"
],
"last": "Och",
"suffix": ""
},
{
"first": "Hermann",
"middle": [],
"last": "Ney",
"suffix": ""
}
],
"year": 2000,
"venue": "I n Proceedings of the 38th Annual Meeting of the ACL",
"volume": "",
"issue": "",
"pages": "440--447",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Och, Franz Josef and Hermann Ney, \" I mp s . \" I n Proceedings of the 38th Annual Meeting of the ACL, pages 440-447, 2000.",
"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": [
"J"
],
"last": "Zhu",
"suffix": ""
}
],
"year": 2002,
"venue": "Proceedings of ACL-2002, 40th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "311--318",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Papineni, K., S. Roukos, T. Ward, and W. J. Zhu, \" BLEU: a method for automatic evaluation of machine translation,\"In Proceedings of ACL-2002, 40th Annual Meeting of the Association for Computational Linguistics pp. 311-318, 2002.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "T h e MI T -LL/AFRL IWSLT-2006 MT System",
"authors": [
{
"first": "",
"middle": [],
"last": "Shen",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Wa D E",
"suffix": ""
}
],
"year": 2006,
"venue": "Proc. of the International Workshop on Spoken Language Translation (IWSLT)",
"volume": "",
"issue": "",
"pages": "71--76",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Shen, Wa d e , B n , \" T h e MI T -LL/AFRL IWSLT-2006 MT System,\" Proc. of the International Workshop on Spoken Language Translation (IWSLT) 2006, pp. 71-76, Kyoto, Japan, 27 November 2006.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Using Phrase Structure and Fluency to Improve Statistical Machine Translation",
"authors": [
{
"first": "Min",
"middle": [
"-"
],
"last": "Shia",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Shiang",
"suffix": ""
}
],
"year": 2006,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Shia, Min-Shiang, Using Phrase Structure and Fluency to Improve Statistical Machine Translation, Master Thesis, Computer Science and Information Engineering, National Cheng Kung University, Tainan, Taiwan, ROC, June, 2006.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Augmentation Model for Answering Well-Defined Questions",
"authors": [
{
"first": "",
"middle": [],
"last": "Shih",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Shu-Fan",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Query",
"suffix": ""
}
],
"year": 2007,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Shih, Shu-Fan, A Query Augmentation Model for Answering Well-Defined Questions, Master Thesis, Department of Computer Science and Information Engineering, National Chi Nan University, Taiwan, ROC, July, 2007.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Proceedings of NAACL-HLT 2007",
"authors": [
{
"first": "M",
"middle": [],
"last": "Simard",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Goutter",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Isabelle",
"suffix": ""
}
],
"year": 2007,
"venue": "",
"volume": "",
"issue": "",
"pages": "508--515",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Simard, M., G. Goutter, P. Isabelle, \" g \" . Proceedings of NAACL-HLT 2007, pp. 508-515, Rochester, NY, April 2007.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "R u l e -Based Translation with Statistical Phrase-Based Post-E d i t i n g",
"authors": [
{
"first": "M",
"middle": [],
"last": "Simard",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Ueffing",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Isabelle",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Kuhn",
"suffix": ""
}
],
"year": 2007,
"venue": "Proceedings of the Second Workshop on Statistical Machine Translation, 2nd WSMT",
"volume": "",
"issue": "",
"pages": "203--206",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Simard, M., N. Ueffing, P. Isabelle, R. Kuhn, \" R u l e -Based Translation with Statistical Phrase-Based Post-E d i t i n g \" . Proceedings of the Second Workshop on Statistical Machine Translation, 2nd WSMT, pp. 203-206, Prague, Czech Republic, June 2007.",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Proceedings of IEEE International Conference on Systems, Man & Cybernetics (SMCC2004)",
"authors": [
{
"first": "Yu",
"middle": [],
"last": "Zhou",
"suffix": ""
},
{
"first": "Chengqing",
"middle": [],
"last": "Zong",
"suffix": ""
},
{
"first": "Bo",
"middle": [],
"last": "Xu",
"suffix": ""
}
],
"year": 2004,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Zhou, Yu, Chengqing Zong, and Bo Xu, \" B i l i n g u a l C h u n k Al i g n me n t i n S t a t i s t i c a l Ma c h i n e T r a n s l a t i o n , \"In Proceedings of IEEE International Conference on Systems, Man & Cybernetics (SMCC2004), Hague, Netherlands, 2004.",
"links": null
}
},
"ref_entries": {
"TABREF2": {
"text": "Input\uff1a E\uf0a2 and eb E Find the weakest alignment entry in E\uf0a2 from the < E\uf0a2 , eb E * > alignment.",
"num": null,
"type_str": "table",
"content": "<table><tr><td/><td/><td>Ep</td><td>'</td><td colspan=\"3\">' ' arg min Pr Ep E \uf0ce \uf02d\uf03d</td><td>\uf028 Ep Ep ' |</td><td>\uf029 Pr( ) Ep</td></tr><tr><td colspan=\"2\">Step 2\uff1a Identify</td><td colspan=\"5\">\uf02d Ep that is the phrase in eb E * aligned with \uf028 \uf029 \uf02d \uf03d \uf02d ' align Ep Ep</td><td>\uf02d Ep . '</td></tr><tr><td colspan=\"7\">Step 3\uff1aFind the fluent phrase Ep \uf02bof \uf03d \uf02b PEB Ep ' Ep \uf02dfrom PEB . \uf028 \uf029 \uf02d ' Ep</td></tr><tr><td colspan=\"7\">Step 4\uff1aSelect the best substitution among Ep' -, Ep+ and Ep-which maximize the</td></tr><tr><td colspan=\"5\">translation score: * ps E E \uf03d</td><td>ps</td><td>\uf07b arg max Pr ' | Pr( ) \uf07d \uf028 \uf029 ' , , Ep Ep Ep E E E \uf0ce \uf02d \uf02b \uf02d</td></tr><tr><td>Step 5\uff1aCut</td><td colspan=\"6\">\uf02d Ep from eb E * eb E E \uf03d</td><td>eb</td><td>( \uf02d \uf02d\uf02b ) ( Ep</td><td>Eps</td><td>)</td></tr><tr><td colspan=\"7\">(Repeat until the translation score Pr(E' | E ) xPr(E) reaches some threshold.)</td></tr></table>",
"html": null
},
"TABREF3": {
"text": "indicates the performance in terms of the error correction capability.",
"num": null,
"type_str": "table",
"content": "<table><tr><td>Error types</td><td>C</td><td colspan=\"3\">Searching Models C+W C+W+P C+W+Q</td></tr><tr><td>Substitution</td><td>21</td><td>23</td><td>32</td><td>34</td></tr><tr><td>Deletion</td><td>28</td><td>39</td><td>46</td><td>62</td></tr><tr><td>Insertion</td><td>40</td><td>43</td><td>47</td><td>47</td></tr><tr><td>Average</td><td>30</td><td>35</td><td>42</td><td>48</td></tr><tr><td colspan=\"4\">Table 1. Number of fully corrected sentences with</td><td/></tr><tr><td colspan=\"4\">different searching models (N=100)</td><td/></tr></table>",
"html": null
}
}
}
} |