File size: 72,680 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 |
{
"paper_id": "I08-1030",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:42:34.149014Z"
},
"title": "Chinese Unknown Word Translation by Subword Re-segmentation",
"authors": [
{
"first": "Ruiqiang",
"middle": [],
"last": "Zhang",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "National Institute of Information and Communications Technology",
"location": {}
},
"email": "ruiqiang.zhang@atr.jp"
},
{
"first": "Eiichiro",
"middle": [],
"last": "Sumita",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "National Institute of Information and Communications Technology",
"location": {}
},
"email": "eiichiro.sumita@nict.go.jp@atr.jp"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "We propose a general approach for translating Chinese unknown words (UNK) for SMT. This approach takes advantage of the properties of Chinese word composition rules, i.e., all Chinese words are formed by sequential characters. According to the proposed approach, the unknown word is re-split into a subword sequence followed by subword translation with a subwordbased translation model. \"Subword\" is a unit between character and long word. We found the proposed approach significantly improved translation quality on the test data of NIST MT04 and MT05. We also found that the translation quality was further improved if we applied named entity translation to translate parts of unknown words before using the subword-based translation.",
"pdf_parse": {
"paper_id": "I08-1030",
"_pdf_hash": "",
"abstract": [
{
"text": "We propose a general approach for translating Chinese unknown words (UNK) for SMT. This approach takes advantage of the properties of Chinese word composition rules, i.e., all Chinese words are formed by sequential characters. According to the proposed approach, the unknown word is re-split into a subword sequence followed by subword translation with a subwordbased translation model. \"Subword\" is a unit between character and long word. We found the proposed approach significantly improved translation quality on the test data of NIST MT04 and MT05. We also found that the translation quality was further improved if we applied named entity translation to translate parts of unknown words before using the subword-based translation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "The use of phrase-based translation has led to great progress in statistical machine translation (SMT). Basically, the mechanism of this approach is realized by two steps:training and decoding. In the training phase, bilingual parallel sentences are preprocessed and aligned using alignment algorithms or tools such as GIZA++ (Och and Ney, 2003) . Phrase pairs are then extracted to be a phrase translation table. Probabilities of a few pre-defined features are computed and assigned to the phrase pairs. The final outcome of the training is a translation table consisting of source phrases, target phrases, and lists of probabilities of features. In the decoding phase, the translation of a test source sentence is made by reordering the target phrases corresponding to the source phrases, and searching for the best hypothesis that yields the highest scores defined by the search criterion.",
"cite_spans": [
{
"start": 326,
"end": 345,
"text": "(Och and Ney, 2003)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "However, this mechanism cannot solve unknown word translation problems. Unknown words (UNK) point to those unseen words in the training or nonexisting words in the translation table. One strategy to deal with translating unknown words is to remove them from the target sentence without translation on assumption of fewer UNKs in the test data. Of course, this simple way produces a lower quality of translations if there are a lot of UNKs in the test data, especially for using a Chinese word segmenter that produces many UNKs. The translation of UNKs need to be solved by a special method.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The translation of Chinese unknown words seems more difficult than other languages because Chinese language is a non-inflected language. Unlike other languages (Yang and Kirchhoff, 2006; Nie\u00dflen and Ney, 2000; Goldwater and McClosky, 2005) , Chinese UNK translation cannot use information from stem and inflection analysis. Using machine transliteration can resolve part of UNK translation (Knight and Graehl, 1997) . But this approach is effective for translating phonetically related unknown words, not for other types. No unified approach for translating Chinese unknown words has been proposed.",
"cite_spans": [
{
"start": 160,
"end": 186,
"text": "(Yang and Kirchhoff, 2006;",
"ref_id": "BIBREF11"
},
{
"start": 187,
"end": 209,
"text": "Nie\u00dflen and Ney, 2000;",
"ref_id": "BIBREF6"
},
{
"start": 210,
"end": 239,
"text": "Goldwater and McClosky, 2005)",
"ref_id": "BIBREF2"
},
{
"start": 390,
"end": 415,
"text": "(Knight and Graehl, 1997)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In this paper we propose a novel statistics-based approach for unknown word translation. This approach uses the properties of Chinese word composition rules -Chinese words are composed of one or more Chinese characters. We can split longer unknown words into a sequence of smaller units: characters or subwords. We train a subword based translation model and use the model to translate the sub-word sequence. Thus we get the translation of the UNKs. We call this approach \"subword-based unknown word translation\".",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In what follows, section 2 reviews phrase-based SMT. section 3 describes the dictionary-based CWS, that is the main CWS in this work. Section 4 describes our named entity recognition approach. Section 5 describes the subword-based approach for UNK translation. Section 7 describes the experiments we conducted to evaluate our subword approach for translating Chinese unknown words. Section 8 describes existing methods for UNK translations for other languages than Chinese. Section 9 briefly summarizes the main points of this work.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Phrase-based SMT uses a framework of log-linear models (Och, 2003) to integrate multiple features. For Chinese to English translation, source sentence C is translated into target sentence E using a probability model:",
"cite_spans": [
{
"start": 55,
"end": 66,
"text": "(Och, 2003)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "P \u039b (E|C) = exp( M i=1 \u03bb i f i (C, E)) E exp( M i=1 \u03bb i f i (C, E )) \u039b = {\u03bb M 1 , } (1) where f i (C, E)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "is the logarithmic value of the i-th feature, and \u03bb i is the weight of the i-th feature. The candidate target sentence that maximizes P(E|C) is the solution.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "Obviously, the performance of such a model depends on the qualities of its features. We used the following features in this work.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "\u2022 Target language model: an N-gram language model is used.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "\u2022 Phrase translation model p(e| f ): gives the probability of the target phrases for each source phrase.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "\u2022 Phrase inverse probability p( f |e): the probability of a source phrase for a given target phrase. It is the coupled feature of the last one.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "\u2022 Lexical probability lex(e| f, a): the sum of the target word probabilities for the given source words and the alignment of the phrase pairs.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "\u2022 Lexical inverse probability lex( f |e, a): the sum of the source word probabilities for the given target words and alignment.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "\u2022 Target phrase length model #(p): the number of phrases included in the translation hypothesis.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "\u2022 Target word penalty model: the number of words included in the translation hypothesis.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "\u2022 Distance model #(w): the number of words between the tail word of one source phrase and the head word of the next source phrase.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "In general, the following steps are used to get the above features.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "1. Data processing: segment Chinese words and tokenize the English.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "2. Word alignment: apply two-way word alignment using GIZA++.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "3. Lexical translation: calculate word lexical probabilities.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "4. Phrase extraction: extract source target bilingual pairs by means of union, intersection, et. al.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "5. Phrase probability calculation: calculate phrase translation probability.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "6. Lexical probability: generate word lexical probabilities for phrase pairs. 7. Minimal error rate training: find a solution to the \u03bb's in the log-linear models.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "3 Dictionary-based Chinese word segmentation",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "For a given Chinese character sequence, C = c 0 c 1 c 2 . . . c N , the problem of word segmentation is addressed as finding a word sequence,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "W = w t 0 w t 1 w t 2 . . . w t M , where the words, w t 0 , w t 1 , w t 2 , . . . , w t M ,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "are pre-defined by a provided lexicon/dictionary, which satisfy",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "w t 0 = c 0 . . . c t 0 , w t 1 = c t 0 +1 . . . c t 1 w t i = c t i\u22121 +1 . . . c t i , w t M = c t M\u22121 +1 . . . c t M t i > t i\u22121 , 0 \u2264 t i \u2264 N, 0 \u2264 i \u2264 M",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "This word sequence is found by maximizing the function below,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "W = arg max W P(W|C) = arg max W P(w t 0 w t 1 . . . w t M )",
"eq_num": "(2)"
}
],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "We applied Bayes' law in the above derivation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "P(w t 0 w t 1 . . . w t M )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "is a language model that can be expanded by the chain rule. If trigram LMs are used, it is approximated as",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "P(w 0 )P(w 1 |w 0 )P(w 2 |w 0 w 1 ) \u2022 \u2022 \u2022 P(w M |w M\u22122 w M\u22121 )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "where w i is a shorthand for w t i .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "Equation 2 indicates the process of the dictionarybased word segmentation. Our CWS is based on it. We used a beam search algorithm because we found that it can speed up the decoding. Trigram LMs were used to score all the hypotheses, of which the one with the highest LM scores is the final output.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "As the name indicates, the word segmentation results by the dictionary-based CWS are dependent on the size and contents of the lexicon. We will use three lexicons in order to compare effects of lexicon size to the translations. The three lexicons denoted as Character, Subword and Hyperword are listed below. An example sentence, (HuangYingChun lives in Beijing City), is given to show the segmentation results of using the lexicons.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "\u2022 Character:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "Only Chinese single characters are included in the lexicon. The sentence is split character by character. / / / / / / /",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "\u2022 Subword: A small amount of most frequent words (10,000) are added to the lexicon. Choosing the subwords are described in section 5. / / / / / /",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "\u2022 Hyperword: A big size of lexicon is used, consisting of 100,000 words. / / / / /",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase-based statistical machine translation",
"sec_num": "2"
},
{
"text": "Named entities in the test data need to be treated separately. Otherwise, a poor translation quality was found by our experiments. We define four types of named entities: people names (nr), organization names (nt), location names (ns), and numerical expressions (nc) such as calendar, time, and money. Our NER model is built according to conditional random fields (CRF) methods (Lafferty et al., 2001) , by which we convert the problem of NER into that of sequence labeling. For example, we can label the last section's example as, \" /B nr /I nr /I nr /O /O /B nt /I nt /I nt\", where \"B\" stands for the first character of a NE; \"I\", other than the first character of a NE; \"O\", isolated character. \"nr\" and \"nt\" are two labels of NE.",
"cite_spans": [
{
"start": 378,
"end": 401,
"text": "(Lafferty et al., 2001)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Named entity recognition (NER)",
"sec_num": "4"
},
{
"text": "We use the CRF++ tools to train the models for named entity recognition 1 . The performance of our NER model was shown in Table 4 . We use the Peking University (PKU) named entity corpus to train the models. Part of the data was used as test data.",
"cite_spans": [],
"ref_spans": [
{
"start": 122,
"end": 129,
"text": "Table 4",
"ref_id": "TABREF3"
}
],
"eq_spans": [],
"section": "Named entity recognition (NER)",
"sec_num": "4"
},
{
"text": "We stick to the results of CWS if there are ambiguities in the segmentation boundary between CWS and NER.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Named entity recognition (NER)",
"sec_num": "4"
},
{
"text": "The NER was used only on the test data in translations. It was not used on the training data due to the consideration of data sparseness. Using NER will generate more unknown words that cannot be found a translation in the translation table. That is why we use a subword-based translation approach.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Named entity recognition (NER)",
"sec_num": "4"
},
{
"text": "We found there were two reasons accounting for producing untranslatable words. The first is the size of lexicon. We proposed three size of lexicons in section 3, of which the Hyperword type uses 100,000 words. Because of a huge lexical size, some of the words cannot be learned by SMT training because of limited training data. The CWS chooses only one candidate segmentation from thousands in splitting a sentence into word sequences. Therefore, the use of a candidate will block other candidates. Hence, many words in the lexicon cannot be fully trained if a large lexicon is used. The second is our NER module. The NER groups a longer sequence of characters into one entity that cannot be translated. We have analyzed this points in the last section. Therefore, in order to translate unknown words, our approach is to split longer unknown words into smaller pieces, and then translate the smaller pieces by using Character or Subword models. Finally, we put the translations back to the Hyperword models. We call this method subword-based unknown word translation regardless of whether a Character model or Subword model is used.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Subword-based translation model for UNK translation",
"sec_num": "5"
},
{
"text": "As described in Section 3, Characters CWS uses only characters in the lexicon. So there is no tricks for it. But for the Subword CWS, its lexicon is a small subset of the Hyperword CWS. In fact, we use the following steps for generating the lexicon. In the beginning, we use the Hyperword CWS to segment the training data. Then, we extract a list of unique tokens and calculate their counts from the results of segmentation. Next, we sort the list as the decreasing order of the counts, and choose N most frequent words from the top of the list. We restrict the length of subwords to three. We use the N words as the lexicon for the subword CWS. N can be changed. Section 7.4 shows its effect to translations. The subword CWS uses a trigram language model to disambiguate. Refer to (Zhang et al., 2006) for details about selecting the subwords.",
"cite_spans": [
{
"start": 782,
"end": 802,
"text": "(Zhang et al., 2006)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Subword-based translation model for UNK translation",
"sec_num": "5"
},
{
"text": "We applied Subword CWS to re-segment the training data. Finally, we can train a subword-based SMT translation model used for translating the unknown words. Training this subword translation model was done in the same way as for the Hyperword translation model that uses the main CWS, as described in the beginning of Section 2.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Subword-based translation model for UNK translation",
"sec_num": "5"
},
{
"text": "The subword-based UNK translation approach can be applied to all the UNKs indiscriminately. However, if we know an UNK is a named entity, we can translate this UNK more accurately than using the subword-based approach. Some unknown words can be translated by named entity translation if they are correctly recognized as named entity and fit a translation pattern. For example, the same words with different named entities are translated differently in the context. The word, \" \", is translated into \"nine\" for measures and money, \"September\" for calendar, and \"jiu\" for Chinese names.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Named entity translation",
"sec_num": "6"
},
{
"text": "As stated in Section 4, we use NER to recognize four types of named entities. Correspondingly, we created the translation patterns to translate each type of the named entities. These patterns include patterns for translating numerical expressions, patterns for translating Chinese and Japanese names, and patterns for translating English alphabet words. The usages are described as follows.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Named entity translation",
"sec_num": "6"
},
{
"text": "Numerical expressions are the largest proportion of unknown words. They include calendar-related terms (days, months, years), money terms, measures, telephone numbers, times, and addresses. These words are translated using a rule-based approach. For example, \" \", is translated into \"at 3:15\".",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Named entity translation",
"sec_num": "6"
},
{
"text": "Chinese and Japanese names are composed of two, three, or four characters. They are translated into English by simply replacing each character with its spelling. The Japanese name, \" \", is translated into \"Shinzo Abe\".",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Named entity translation",
"sec_num": "6"
},
{
"text": "English alphabets are encoded in different Chinese characters. They are translated by replacing the Chinese characters with the corresponding English letters.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Named entity translation",
"sec_num": "6"
},
{
"text": "We use the above translation patterns to translate the named entities. Using translation patterns produce almost correct translation. Hence, we put the named entity translation to work before we apply the subword translation model. The subword translation model is used when the unknown words cannot be translated by named entity translation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Named entity translation",
"sec_num": "6"
},
{
"text": "We used LDC Chinese/English data for training. We used two test data of NIST MT04 and NIST MT05. The statistics of the data are shown in Table 6 . We used about 2.4 million parallel sentences extracted from LDC data for training. Experiments on both the MT04 and MT05 test data used the same translation models on the same training data, but the min- We used a Chinese word segmentation tool, Achilles, for doing word segmentation. Its word segmentation accuracy was higher than the stanford word segmenter (Tseng et al., 2005) in our laboratory test (Zhang et al., 2006) .",
"cite_spans": [
{
"start": 507,
"end": 527,
"text": "(Tseng et al., 2005)",
"ref_id": "BIBREF10"
},
{
"start": 551,
"end": 571,
"text": "(Zhang et al., 2006)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [
{
"start": 137,
"end": 144,
"text": "Table 6",
"ref_id": "TABREF5"
}
],
"eq_spans": [],
"section": "SMT experiments 7.1 Data",
"sec_num": "7"
},
{
"text": "The average length of a sentence for the test data MT04 and MT05 after word segmentation is 37.5 by using the Subword CWS, and 27.9 by using the Hyperword CWS. Table 6 shows statistics of unknown words in MT04 and MT05 using different word segmentation. Obviously, character-based and subword-based CWS generated much fewer unknown words, but sentences are over-segmented. The CWS of Hyperword generated many UNKs because of using a large size of lexicon. However, if named entity recognition was applied upon the segmented results of the Hyperword, more UNKs were produced. Take an example for MT04. There are 1,305 UNKs in which numeric expressions amount to 35.2%, people names at 11.2%, organization names at 19.2%, location names at 17.6%, and others at 16.8%. Analysis of these numbers helps to understand the distribution of unknown words.",
"cite_spans": [],
"ref_spans": [
{
"start": 160,
"end": 167,
"text": "Table 6",
"ref_id": "TABREF5"
}
],
"eq_spans": [],
"section": "SMT experiments 7.1 Data",
"sec_num": "7"
},
{
"text": "As described in section 3, we used three lexicon size for the dictionary-based CWS. Therefore, we had three CWS denoted as: Character, Subword and Hyperword. We used the three CWS in turn to do word segmentation to the training data, and then built the translation models respectively. We tested the performance of each of the translation models on the test data. The results are shown on Table 4 . The translations are evaluated in terms of BLEU score (Papineni et al., 2002) . This experiment was just testing the effect of the three CWS. Therefore, all the UNKs of the test data were not translated, simply removed from the results.",
"cite_spans": [
{
"start": 453,
"end": 476,
"text": "(Papineni et al., 2002)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [
{
"start": 389,
"end": 396,
"text": "Table 4",
"ref_id": "TABREF3"
}
],
"eq_spans": [],
"section": "Effect of the various CWS",
"sec_num": "7.2"
},
{
"text": "We found the character-based CWS yielded the lowest BLEU scores, indicating the translation quality of this type is the worst. The Hyperword CWS achieved the best results. If we relate it to Table 6, we found while the Hyperword CWS produced many more UNKs than the Character and Subword CWS, its translation quality was improved instead. The fact proves the quality of translation models play a more important role than the amount of unknown word translation. Using the Hyperword CWS can generate a higher quality of translation models than the Character and Subword CWS. Therefore, we cannot use the character and subword-based CWS in Chinese SMT system due to their overall poor performance. But we found their ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Effect of the various CWS",
"sec_num": "7.2"
},
{
"text": "The experiments in this section show the effect of using the subword translation model for UNKs. We compared the results of using subword translation with those of without using it. We also used named entity translation together with the subword translation. Thus, we could compare the effect of subword translation under conditions of with or without named entity translation. We listed four kinds of results to evaluate the performance of our approach in Table 5 where the symbols indicate:",
"cite_spans": [],
"ref_spans": [
{
"start": 457,
"end": 464,
"text": "Table 5",
"ref_id": "TABREF4"
}
],
"eq_spans": [],
"section": "Effect of subword translation for UNKs",
"sec_num": "7.3"
},
{
"text": "\u2022 Baseline: this is the results made by the Hyperword CWS of Table 4 . No subword translation for UNKs and named entity translations were used. Unknown words were simply removed from the output.",
"cite_spans": [],
"ref_spans": [
{
"start": 61,
"end": 68,
"text": "Table 4",
"ref_id": "TABREF3"
}
],
"eq_spans": [],
"section": "Effect of subword translation for UNKs",
"sec_num": "7.3"
},
{
"text": "\u2022 Baseline+Subword: the results were made under the same conditions as the first except all of the UNKs were extracted, re-segmented by the subword CWS and translated by the subword translation models. However, the named entity translation was not used.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Effect of subword translation for UNKs",
"sec_num": "7.3"
},
{
"text": "\u2022 Baseline+NER: this experiment did not use subword-based translation for UNKs. But we used named entity translation. Part of UNKs was labeled with named entities and translated by pattern match of section 6.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Effect of subword translation for UNKs",
"sec_num": "7.3"
},
{
"text": "\u2022 Baseline+NER+Subword: this experiment used the named entity translation and the subword-based translation. The difference from the second one is that some UNKs were translated by the translation patterns of section 6 at first and the remaining UNKs were translated using the subword model (the second one translated all of the UNKs using the subword model).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Effect of subword translation for UNKs",
"sec_num": "7.3"
},
{
"text": "The results of our experiments are shown in Table 5. We found the subword models improved translations in all of the experiments. Using the subword models on the MT04 test data improved translations in terms of BLEU scores from 0.280 to 0.283, and from 0.236 to 0.244 on the MT05 test data. While only small gains of BLEU were achieved by UNK translation, this improvement is sufficient to prove the effectiveness of the subword models, given that the test data had only a low proportion of UNKs.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Effect of subword translation for UNKs",
"sec_num": "7.3"
},
{
"text": "The BLEU scores of \"Baseline+NER\" is higher than that of \"Baseline\", that proves using named entity translation improved translations, but the effect of using named entity translation was worse than using the subword-based translation. This is because the named entity translation is applicable for the named entities only. However, the subword-based translation is used for all the UNKs.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Effect of subword translation for UNKs",
"sec_num": "7.3"
},
{
"text": "When we applied named entity translation to translate some of recognized named entities followed by using the subword models, we found BLEU gains over using the subword models uniquely, 0.2% for MT04 and 0.2% for MT05. This experiment proves that the best way of using the subword models is to separate the UNKs that can be translated by named entity translation from those that cannot, and let the subword models handle translations of those not translated.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Effect of subword translation for UNKs",
"sec_num": "7.3"
},
{
"text": "Analysis using the bootstrap tool created by Zhang et al. (Zhang et al., 2004) showed that the results made by the subword translations were significantly better than the ones not using it.",
"cite_spans": [
{
"start": 45,
"end": 78,
"text": "Zhang et al. (Zhang et al., 2004)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Effect of subword translation for UNKs",
"sec_num": "7.3"
},
{
"text": "We have found a significant improvement by using the subword models. The essence of the approach Table 6 , we used three classes of subword lists: character, 10K subwords and 20K subwords. The \"character\" class used only single-character words, about 5,000 characters. The other two classes, \"10K\" and \"20K\", used 10,000 and 20,000 subwords. The method for choosing the subwords was described in Section 5. We have used \"10K\" in the previous experiments. We did not use named entity translation for this experiment. We found that using \"character\" as the subword unit brought in nearly no improvement over the baseline results. Using 20K subwords yielded better results than the baseline but smaller gains than that of using the 10K subwords for MT05 data. It proves that using subword translation is an effective approach but choosing a right size of subword lexicon is important. We cannot propose a better method for finding the size. We can do more experiments repeatedly to find this value. We found the size of 10,000 subwords achieved the best results for our experiments.",
"cite_spans": [],
"ref_spans": [
{
"start": 97,
"end": 104,
"text": "Table 6",
"ref_id": "TABREF5"
}
],
"eq_spans": [],
"section": "Effect of changing the size of subword lexicon",
"sec_num": "7.4"
},
{
"text": "Unknown word translation is an important problem for SMT. As we showed in the experiments, appropriate handling of this problem results in a significant improvement of translation quality. As we have known, there exists some methods for solving this problem. While these approaches were not proposed in aim to unknown word translation, they can be used for UNK translations indirectly.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related work",
"sec_num": "8"
},
{
"text": "Most existing work focuses on named entity translation (Carpuat et al., 2006) because named entities are the large proportion of unknown words. We also used similar methods for translating named entities in this work. Some used stem and morphological analysis for UNKs such as (Goldwater and McClosky, 2005) . Morphological analysis is effective for inflective languages but not for Chinese. Using unknown word modeling such as backoff models was proposed by (Yang and Kirchhoff, 2006) .",
"cite_spans": [
{
"start": 55,
"end": 77,
"text": "(Carpuat et al., 2006)",
"ref_id": "BIBREF1"
},
{
"start": 277,
"end": 307,
"text": "(Goldwater and McClosky, 2005)",
"ref_id": "BIBREF2"
},
{
"start": 459,
"end": 485,
"text": "(Yang and Kirchhoff, 2006)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related work",
"sec_num": "8"
},
{
"text": "Other proposed methods include paraphrasing (Callison-Burch et al., 2006) and transliteration (Knight and Graehl, 1997) that uses the feature of phonetic similarity. However, This approach does not work if no phonetic relationship is found.",
"cite_spans": [
{
"start": 44,
"end": 73,
"text": "(Callison-Burch et al., 2006)",
"ref_id": "BIBREF0"
},
{
"start": 94,
"end": 119,
"text": "(Knight and Graehl, 1997)",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related work",
"sec_num": "8"
},
{
"text": "Splitting compound words into translatable subwords as we did in this work have been used by (Nie\u00dflen and Ney, 2000) and (Koehn and Knight, 2003) for languages other than Chinese where detailed splitting methods are proposed. We used forward maximum match method to split unknown words. This splitting method is relatively simple but works well for Chinese. The splitting for Chinese is not as complicated as those languages with alphabet.",
"cite_spans": [
{
"start": 93,
"end": 116,
"text": "(Nie\u00dflen and Ney, 2000)",
"ref_id": "BIBREF6"
},
{
"start": 121,
"end": 145,
"text": "(Koehn and Knight, 2003)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related work",
"sec_num": "8"
},
{
"text": "We made use of the specific property of Chinese language and proposed a subword re-segmentation to solve the translation of unknown words. Our approach was tested under various conditions such as using named entity translation and varied subword lexicons. We found this approach was very effective. We are hopeful that this approach can be applied into languages that have similar features as Chinese, for example, Japanese.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion and conclusion",
"sec_num": "9"
},
{
"text": "While the work was done on a SMT system which is not the state-of-the-art 2 , the idea of using subword-based translation for UNKs is applicable to any systems because the problem of UNK translation has to be faced by any system.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion and conclusion",
"sec_num": "9"
},
{
"text": "http://chasen.org/\u02dctaku/software/CRF++/",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "The BLEU score of the top one system is about 0.35 for MT05 (http://www.nist.gov/speech/tests/mt/).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "The authors would like to thank Dr.Michael Paul for his assistance in this work, especially for evaluating methods and statistical significance test.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgement",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Improved statistical machine translation using paraphrases",
"authors": [
{
"first": "Chris",
"middle": [],
"last": "Callison",
"suffix": ""
},
{
"first": "-",
"middle": [],
"last": "Burch",
"suffix": ""
}
],
"year": 2006,
"venue": "HLT-NAACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Chris Callison-Burch, Philipp Koehn, and Miles Os- borne. 2006. Improved statistical machine translation using paraphrases. In HLT-NAACL-2006.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Toward Integrating Word Sense and Entity Disambiguation into Statistical Machine Translation",
"authors": [
{
"first": "Marine",
"middle": [],
"last": "Carpuat",
"suffix": ""
},
{
"first": "Yihai",
"middle": [],
"last": "Shen",
"suffix": ""
},
{
"first": "Xiaofeng",
"middle": [],
"last": "Yu",
"suffix": ""
},
{
"first": "Dekai",
"middle": [],
"last": "Wu",
"suffix": ""
}
],
"year": 2006,
"venue": "Proc. of the IWSLT",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Marine Carpuat, Yihai Shen, Xiaofeng Yu, and Dekai Wu. 2006. Toward Integrating Word Sense and Entity Disambiguation into Statistical Machine Translation. In Proc. of the IWSLT.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Improving statistical MT through morphological analysis",
"authors": [
{
"first": "Sharon",
"middle": [],
"last": "Goldwater",
"suffix": ""
},
{
"first": "David",
"middle": [],
"last": "Mcclosky",
"suffix": ""
}
],
"year": 2005,
"venue": "Proceedings of the HLT/EMNLP",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sharon Goldwater and David McClosky. 2005. Improv- ing statistical MT through morphological analysis. In Proceedings of the HLT/EMNLP.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Machine transliteration",
"authors": [
{
"first": "Kevin",
"middle": [],
"last": "Knight",
"suffix": ""
},
{
"first": "Jonathan",
"middle": [],
"last": "Graehl",
"suffix": ""
}
],
"year": 1997,
"venue": "Proc. of the ACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kevin Knight and Jonathan Graehl. 1997. Machine transliteration. In Proc. of the ACL.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Empirical methods for compound splitting",
"authors": [
{
"first": "Philipp",
"middle": [],
"last": "Koehn",
"suffix": ""
},
{
"first": "Kevin",
"middle": [],
"last": "Knight",
"suffix": ""
}
],
"year": 2003,
"venue": "EACL-2003",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Philipp Koehn and Kevin Knight. 2003. Empirical meth- ods for compound splitting. In EACL-2003.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Conditional random fields: probabilistic models for segmenting and labeling sequence data",
"authors": [
{
"first": "John",
"middle": [],
"last": "Lafferty",
"suffix": ""
},
{
"first": "Andrew",
"middle": [],
"last": "Mccallum",
"suffix": ""
},
{
"first": "Fernando",
"middle": [],
"last": "Pereira",
"suffix": ""
}
],
"year": 2001,
"venue": "Proc. of ICML-2001",
"volume": "",
"issue": "",
"pages": "591--598",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "John Lafferty, Andrew McCallum, and Fernando Pereira. 2001. Conditional random fields: probabilistic models for segmenting and labeling sequence data. In Proc. of ICML-2001, pages 591-598.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Improving smt quality with morpho-syntactic analysis",
"authors": [
{
"first": "Sonja",
"middle": [],
"last": "Nie\u00dflen",
"suffix": ""
},
{
"first": "Hermann",
"middle": [],
"last": "Ney",
"suffix": ""
}
],
"year": 2000,
"venue": "Proc. of COLING",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sonja Nie\u00dflen and Hermann Ney. 2000. Improving smt quality with morpho-syntactic analysis. In Proc. of COLING.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "A systematic comparison of various statistical alignment models",
"authors": [
{
"first": "Josef",
"middle": [],
"last": "Franz",
"suffix": ""
},
{
"first": "Hermann",
"middle": [],
"last": "Och",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Ney",
"suffix": ""
}
],
"year": 2003,
"venue": "Computational Linguistics",
"volume": "29",
"issue": "1",
"pages": "19--51",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Franz Josef Och and Hermann Ney. 2003. A system- atic comparison of various statistical alignment mod- els. Computational Linguistics, 29(1):19-51.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Minimum error rate training for statistical machine trainslation",
"authors": [
{
"first": "F",
"middle": [
"J"
],
"last": "Och",
"suffix": ""
}
],
"year": 2003,
"venue": "Proc. ACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "F. J. Och. 2003. Minimum error rate training for statisti- cal machine trainslation. In Proc. ACL.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "BLEU: a method for automatic evaluation of machine translation",
"authors": [
{
"first": "K",
"middle": [],
"last": "Papineni",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Roukos",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Ward",
"suffix": ""
},
{
"first": "W",
"middle": [],
"last": "Zhu",
"suffix": ""
}
],
"year": 2002,
"venue": "Proc. of the 40th ACL",
"volume": "",
"issue": "",
"pages": "311--318",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "K. Papineni, S. Roukos, T. Ward, and W. Zhu. 2002. BLEU: a method for automatic evaluation of machine translation. In Proc. of the 40th ACL, pages 311-318, Philadelphia, USA.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "A conditional random field word segmenter for Sighan bakeoff",
"authors": [
{
"first": "Huihsin",
"middle": [],
"last": "Tseng",
"suffix": ""
},
{
"first": "Pichuan",
"middle": [],
"last": "Chang",
"suffix": ""
},
{
"first": "Galen",
"middle": [],
"last": "Andrew",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Jurafsky",
"suffix": ""
},
{
"first": "Christopher",
"middle": [],
"last": "Manning",
"suffix": ""
}
],
"year": 2005,
"venue": "Proceedings of the Fourth SIGHAN Workshop on Chinese Language Processing",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Huihsin Tseng, Pichuan Chang, Galen Andrew, Daniel Jurafsky, and Christopher Manning. 2005. A condi- tional random field word segmenter for Sighan bake- off 2005. In Proceedings of the Fourth SIGHAN Work- shop on Chinese Language Processing, Jeju, Korea.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Phrase-based backoff models for machine translation of highly inflected languages",
"authors": [
{
"first": "Mei",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "Katrin",
"middle": [],
"last": "Kirchhoff",
"suffix": ""
}
],
"year": 2006,
"venue": "EACL-2006",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mei Yang and Katrin Kirchhoff. 2006. Phrase-based backoff models for machine translation of highly in- flected languages. In EACL-2006.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "terpreting bleu/nist scores: How much improvement do we need to have a better system? In Proceedings of the LREC",
"authors": [
{
"first": "Ying",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Stephan",
"middle": [],
"last": "Vogel",
"suffix": ""
},
{
"first": "Alex",
"middle": [],
"last": "Waibel",
"suffix": ""
}
],
"year": 2004,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ying Zhang, Stephan Vogel, and Alex Waibel. 2004. In- terpreting bleu/nist scores: How much improvement do we need to have a better system? In Proceedings of the LREC.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Subword-based tagging by conditional random fields for chinese word segmentation",
"authors": [
{
"first": "Ruiqiang",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Genichiro",
"middle": [],
"last": "Kikui",
"suffix": ""
},
{
"first": "Eiichiro",
"middle": [],
"last": "Sumita",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of the HLT-NAACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ruiqiang Zhang, Genichiro Kikui, and Eiichiro Sumita. 2006. Subword-based tagging by conditional random fields for chinese word segmentation. In Proceedings of the HLT-NAACL.",
"links": null
}
},
"ref_entries": {
"TABREF0": {
"text": "NER accuracy type Recall Precision F-score nr 85.32% 93.41% 89.18% ns 87.80% 90.46% 89.11% nt 84.50% 87.54% 85.99% all 84.58% 90.97% 87.66%",
"type_str": "table",
"html": null,
"content": "<table/>",
"num": null
},
"TABREF1": {
"text": "Statistics of data for MT experiments",
"type_str": "table",
"html": null,
"content": "<table><tr><td>Chinese</td><td>English</td></tr></table>",
"num": null
},
"TABREF2": {
"text": "Statistics of unknown words of test data using different CWS",
"type_str": "table",
"html": null,
"content": "<table><tr><td/><td colspan=\"3\">Hyperword+Named entities</td><td/><td colspan=\"3\">Hyperword Subwords Characters</td></tr><tr><td/><td colspan=\"4\">Numerics People Org. Loc. other</td><td/><td/><td/></tr><tr><td>MT04</td><td>460</td><td>146</td><td>250 230</td><td>219</td><td>650</td><td>18</td><td>2</td></tr><tr><td>MT05</td><td>414</td><td>271</td><td>311 146</td><td>323</td><td>680</td><td>23</td><td>2</td></tr><tr><td colspan=\"4\">imum error rate training was different. The MT04</td><td/><td/><td/><td/></tr><tr><td colspan=\"4\">and MT05 test data were also used as development</td><td/><td/><td/><td/></tr><tr><td colspan=\"2\">data for cross experiments.</td><td/><td/><td/><td/><td/><td/></tr></table>",
"num": null
},
"TABREF3": {
"text": "Compare the translations by different CWS (BLEU scores)",
"type_str": "table",
"html": null,
"content": "<table><tr><td/><td>MT04 MT05</td></tr><tr><td>Character</td><td>0.253 0.215</td></tr><tr><td>Subword</td><td>0.265 0.229</td></tr><tr><td colspan=\"2\">Hyperword 0.280 0.236</td></tr></table>",
"num": null
},
"TABREF4": {
"text": "",
"type_str": "table",
"html": null,
"content": "<table><tr><td colspan=\"2\">: Effect of subword and named entity translation</td></tr><tr><td>(BLEU)</td><td/></tr><tr><td/><td>MT04 MT05</td></tr><tr><td>Baseline(Hyperword)</td><td>0.280 0.236</td></tr><tr><td>Baseline+Subword</td><td>0.283 0.244</td></tr><tr><td>Baseline+NER</td><td>0.283 0.242</td></tr><tr><td colspan=\"2\">Baseline+NER+Subword 0.285 0.246</td></tr><tr><td>usage for UNK translation.</td><td/></tr></table>",
"num": null
},
"TABREF5": {
"text": "BLEU scores for changing the subword lexicon size",
"type_str": "table",
"html": null,
"content": "<table><tr><td colspan=\"2\">subword size MT04 MT05</td></tr><tr><td>character</td><td>0.280 0.237</td></tr><tr><td>10K</td><td>0.283 0.244</td></tr><tr><td>20K</td><td>0.283 0.240</td></tr><tr><td colspan=\"2\">is to split unknown words into subword sequences</td></tr><tr><td colspan=\"2\">and use subword models to translate the subword</td></tr><tr><td colspan=\"2\">sequences. The choices are flexible in choosing the</td></tr><tr><td colspan=\"2\">number of subwords in the subword lexicon. If a</td></tr><tr><td colspan=\"2\">different subword list is used, the results of the sub-</td></tr><tr><td colspan=\"2\">word re-segmentation will be changed. Will choos-</td></tr><tr><td colspan=\"2\">ing a different subword list have a large impact on</td></tr><tr><td colspan=\"2\">the translation of UNKs? As shown in</td></tr></table>",
"num": null
}
}
}
} |