File size: 79,711 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 |
{
"paper_id": "I05-1040",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:24:46.370143Z"
},
"title": "An Ensemble of Grapheme and Phoneme for Machine Transliteration",
"authors": [
{
"first": "Jong-Hoon",
"middle": [],
"last": "Oh",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "KAIST/KORTERM/BOLA",
"location": {
"addrLine": "373-1 Guseong-dong, Yuseong-gu",
"postCode": "305-701",
"settlement": "Daejeon",
"country": "Republic of Korea"
}
},
"email": ""
},
{
"first": "Key-Sun",
"middle": [],
"last": "Choi",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "KAIST/KORTERM/BOLA",
"location": {
"addrLine": "373-1 Guseong-dong, Yuseong-gu",
"postCode": "305-701",
"settlement": "Daejeon",
"country": "Republic of Korea"
}
},
"email": "kschoi@world.kaist.ac.kr"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "Machine transliteration is an automatic method to generate characters or words in one alphabetical system for the corresponding characters in another alphabetical system. There has been increasing concern on machine transliteration as an assistant of machine translation and information retrieval. Three machine transliteration models, including \"grapheme-based model\", \"phonemebased model\", and \"hybrid model\", have been proposed. However, there are few works trying to make use of correspondence between source grapheme and phoneme, although the correspondence plays an important role in machine transliteration. Furthermore there are few works, which dynamically handle source grapheme and phoneme. In this paper, we propose a new transliteration model based on an ensemble of grapheme and phoneme. Our model makes use of the correspondence and dynamically uses source grapheme and phoneme. Our method shows better performance than the previous works about 15~23% in English-to-Korean transliteration and about 15~43% in English-to-Japanese transliteration.",
"pdf_parse": {
"paper_id": "I05-1040",
"_pdf_hash": "",
"abstract": [
{
"text": "Machine transliteration is an automatic method to generate characters or words in one alphabetical system for the corresponding characters in another alphabetical system. There has been increasing concern on machine transliteration as an assistant of machine translation and information retrieval. Three machine transliteration models, including \"grapheme-based model\", \"phonemebased model\", and \"hybrid model\", have been proposed. However, there are few works trying to make use of correspondence between source grapheme and phoneme, although the correspondence plays an important role in machine transliteration. Furthermore there are few works, which dynamically handle source grapheme and phoneme. In this paper, we propose a new transliteration model based on an ensemble of grapheme and phoneme. Our model makes use of the correspondence and dynamically uses source grapheme and phoneme. Our method shows better performance than the previous works about 15~23% in English-to-Korean transliteration and about 15~43% in English-to-Japanese transliteration.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Machine transliteration is an automatic method to generate characters or words in one alphabetical system for the corresponding characters in another alphabetical system. For example, English word data is transliterated into Korean 'deita' 1 and Japanese 'deeta'. Transliteration is used to phonetically translate proper names and technical terms especially from languages in Roman alphabets to languages in non-Roman alphabets such as from English to Korean, Japanese, and Chinese and so on. There has been increasing concern on machine transliteration as an assistant of Machine Translation (MT) [2] , [10] , mono-lingual information retrieval (MLIR) [8] , [11] and cross-lingual information retrieval (CLIR) [6] . In the area of MLIR and CLIR, machine transliteration bridges the gap between a transliterated localized form and its original form by generating all possible transliterated forms from each original form. Especially for CLIR, machine transliteration gives a help to query translation where proper names and technical terms frequently appear in source language queries. In the area of MT, machine transliteration prevents translation failure when translations of proper names and technical terms are not registered in a translation dictionary. A machine transliteration system, therefore, may affect the performance of MT, MLIR, and CLIR system.",
"cite_spans": [
{
"start": 240,
"end": 241,
"text": "1",
"ref_id": "BIBREF0"
},
{
"start": 598,
"end": 601,
"text": "[2]",
"ref_id": "BIBREF1"
},
{
"start": 604,
"end": 608,
"text": "[10]",
"ref_id": "BIBREF9"
},
{
"start": 653,
"end": 656,
"text": "[8]",
"ref_id": "BIBREF7"
},
{
"start": 659,
"end": 663,
"text": "[11]",
"ref_id": "BIBREF10"
},
{
"start": 711,
"end": 714,
"text": "[6]",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Three machine transliteration models have been studied: called \"grapheme 2 -based transliteration model (\u03c8 G )\" [7] , [8] , [9] , [11] , [12] , [13] , \"phoneme 3 -based transliteration model (\u03c8 P )\" [10] , [12] , and \"hybrid transliteration model (\u03c8 H )\" [2] , [4] , [12] . \u03c8 G and \u03c8 P are classified in terms of units to be transliterated. \u03c8 G is referred to the direct model because it directly transforms source language graphemes to target language graphemes without any phonetic knowledge of source language words. \u03c8 P is called the pivot model because it makes use of phonemes as a pivot during a transliteration process. Therefore \u03c8 P usually needs two steps; the first step is to produce phonemes from source language graphemes, and the second step is to produce target language graphemes from phonemes. \u03c8 H combines \u03c8 G and \u03c8 P with the linear interpolation style. Hereafter, we will use a source grapheme for a source language grapheme and a target grapheme for a target language grapheme.",
"cite_spans": [
{
"start": 112,
"end": 115,
"text": "[7]",
"ref_id": "BIBREF6"
},
{
"start": 118,
"end": 121,
"text": "[8]",
"ref_id": "BIBREF7"
},
{
"start": 124,
"end": 127,
"text": "[9]",
"ref_id": "BIBREF8"
},
{
"start": 130,
"end": 134,
"text": "[11]",
"ref_id": "BIBREF10"
},
{
"start": 137,
"end": 141,
"text": "[12]",
"ref_id": "BIBREF11"
},
{
"start": 144,
"end": 148,
"text": "[13]",
"ref_id": "BIBREF12"
},
{
"start": 199,
"end": 203,
"text": "[10]",
"ref_id": "BIBREF9"
},
{
"start": 206,
"end": 210,
"text": "[12]",
"ref_id": "BIBREF11"
},
{
"start": 255,
"end": 258,
"text": "[2]",
"ref_id": "BIBREF1"
},
{
"start": 261,
"end": 264,
"text": "[4]",
"ref_id": "BIBREF3"
},
{
"start": 267,
"end": 271,
"text": "[12]",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Though transliteration is the phonetic process (\u03c8 P ) rather than the orthographic one (\u03c8 G ) [10] , we should consider both source grapheme and phoneme to achieve high performance in machine transliteration because the standard transliterations are not restricted to phoneme-based transliterations 4 . However, many previous works make use of either source grapheme or phoneme. They simplify a machine transliteration problem into either \u03c8 G or \u03c8 P assuming that one of \u03c8 G and \u03c8 P is able to cover all transliteration behaviors. However, transliteration is a complex process, which does not rely on either source grapheme or phoneme. For example, the standard Korean transliterations of amylase and data are grapheme-based transliteration 'amillaaje' and phoneme-based transliteration 'deiteo', respectively. A machine transliteration model, therefore, should reflect the dynamic transliteration behaviors in order to produce the correct transliterations. \u03c8 H has the limited power for producing the correct transliterations because it just combines \u03c8 G and \u03c8 P with the linear interpolation style. \u03c8 H does not consider correspondence between source grapheme and phoneme during the transliteration process. However the correspondence plays important roles in machine transliteration. For example, phoneme /AH/ 5 produces high ambiguities since it can be mapped to almost every single vowels in source language and target language (the underlined grapheme corresponds to /AH/: cinema, hostel, holocaust in English, 'sinema', 'hostel', 'hollokoseuteu' in their Korean counterparts, and 'sinema', 'hoseuteru', 'horokoosuto' in their Japanese counterparts). If we know the correspondence between source grapheme and phoneme in this context, then we can more easily infer the correct transliteration of /AH/, since a target grapheme of /AH/ usually depends on a source grapheme corresponding to /AH/. Korean transliterations of source grapheme a is various such as 'a', 'ei', 'o', 'eo' and so on. Like the previous example, correspondence makes it possible to reduce transliteration ambiguities like Table 1 . In Table 1 , the underlined source grapheme a in the example column is pronounced as the phoneme in the phoneme column. The correct Korean transliterations of source grapheme a can be more easily found, like in the Korean grapheme column, by means of phonemes in the phoneme column. In this paper, we propose a new machine transliteration model based on an ensemble of source grapheme and phoneme, symbolized as \u03c8 C (\"correspondence-based transliteration model\"). \u03c8 C has two strong points over \u03c8 G , \u03c8 P , and \u03c8 H . First, \u03c8 C can produce transliterations by considering correspondence between source grapheme and phoneme. As described above, correspondence is very useful for reducing transliteration ambiguities. From the viewpoint of reducing the ambiguities, \u03c8 C has an advantage over \u03c8 G , \u03c8 P , and \u03c8 H because \u03c8 C can more easily reduce the ambiguities by considering the correspondence. Second, \u03c8 C can dynamically handle source grapheme and phoneme according to their contexts. Because of this property, \u03c8 C can produce grapheme-based transliterations as well as phoneme-based transliterations. It can also produce a transliteration, where one part is a grapheme-based transliteration and the other part is a phoneme-based transliteration. For example, the Korean transliteration of neomycin, 'neomaisin', where 'neo' is a grapheme-based transliteration and 'maisin' is a phoneme-based transliteration.",
"cite_spans": [
{
"start": 94,
"end": 98,
"text": "[10]",
"ref_id": "BIBREF9"
},
{
"start": 299,
"end": 300,
"text": "4",
"ref_id": "BIBREF3"
}
],
"ref_spans": [
{
"start": 2098,
"end": 2105,
"text": "Table 1",
"ref_id": "TABREF0"
},
{
"start": 2111,
"end": 2118,
"text": "Table 1",
"ref_id": "TABREF0"
}
],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Correspondence-based transliteration model (\u03c8 C ) is composed of two component functions (\u03c8 C : \u03b4 p \u00d7\u03b4 t ). In this paper, we refer to \u03b4 p as a function for \"producing pronunciation\" and \u03b4 t as a function for \"producing target grapheme\". First, \u03b4 p produces pronunciation and then \u03b4 t produces target graphemes with correspondence between source grapheme and phoneme produced by \u03b4 p . The goal of the \u03b4 p is to produce the most probable sequence of phonemes corresponding to source graphemes. For example, \u03b4 p produces /B/, /AO/, /~/ 6 , /R/, and /D/ for each source grapheme, b, o, a, r, and d in board (see \"The result of \u03b4 p \" in the right side of Fig 1) . In this step, pronunciation is generated through two ways; pronunciation dictionary search and pronunciation estimation. A pronunciation dictionary contains the correct pronunciation corresponding to English words. Therefore, English words are first investigated whether they are registered in the dictionary otherwise their pronunciation is estimated by pronunciation estimation. The goal of \u03b4 t is to produce the most probable sequence of target graphemes with correspondence between source grapheme and phoneme, which is the result of \u03b4 p . For example, \u03b4 t produces 'b', 'o', '~', '~', and 'deu' using the result of \u03b4 p , b-/B/, o-/AO/, a-/~/, r-/R/, and d-/D/ (see \"The result of \u03b4 t \" in the right side of Fig 1) . Finally, the target language transliteration, such as the Korean transliteration 'bodeu' for board, can be acquired by concatenating the sequence of target graphemes in the result of \u03b4 t . Table 2 . Feature types used for correspondence-based transliteration model: where S is a set of source graphemes (e.g. English alphabets), P is a set of phonemes defined in ARPABET, T is a set of target graphemes. Note that f S,GS is a symbol for indicating both f S and f GS . f P,GP is a symbol for indicating both f P and f GP . ",
"cite_spans": [
{
"start": 534,
"end": 535,
"text": "6",
"ref_id": "BIBREF5"
}
],
"ref_spans": [
{
"start": 651,
"end": 657,
"text": "Fig 1)",
"ref_id": "FIGREF0"
},
{
"start": 1372,
"end": 1378,
"text": "Fig 1)",
"ref_id": "FIGREF0"
},
{
"start": 1570,
"end": 1577,
"text": "Table 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Correspondence-Based Machine Transliteration Model",
"sec_num": "2"
},
{
"text": "board /D/ /R/ /~/ /AO/ /B/ d r a o b /D/ /R/ /~/ /AO/ /B/ 'deu' 'o' 'b' d r a o b p p p p t t Result of p",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Correspondence-Based Machine Transliteration Model",
"sec_num": "2"
},
{
"text": "(s i ) uses (f S , f GS , f P ) and \u03b4 t (s i , \u03b4 p (s i )) does (f S , f P , f GS , f GP , f T ).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Correspondence-Based Machine Transliteration Model",
"sec_num": "2"
},
{
"text": "Producing pronunciation (\u03b4 p :S\u2192P) is a function that finds phonemes in a set P for each source grapheme, where P is a set of phonemes defined in ARPABET, and S is a set of source graphemes (e.g. English alphabets). The results of this step can be represented as a sequence of correspondences between source grapheme and phoneme. We will denote it as GP={gp 1 ,gp 2 ,\u2026,gp n ; gp i =(s i ,\u03b4 p (s i ))} where s i is the i th source grapheme of SW=s 1 ,s 2 ,...,s n . Producing pronunciation is composed of two steps. The first step involves a search in the pronunciation dictionary, which contains English words and their pronunciation. This paper uses The CMU Pronouncing Dictionary 7 , which contains 120,000 English words and their pronunciation. The second step involves pronunciation estimation. If an English word is not registered in the pronunciation dictionary, we must estimate its pronunciation. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Producing Pronunciation (\u03b4 p )",
"sec_num": "2.1"
},
{
"text": "\u03b4 p (C0) f S $ $ $ b o a r f GS $ $ $ C V V C /B/ f P $ $ $ Let SW=s 1 ,s 2 ,.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Producing Pronunciation (\u03b4 p )",
"sec_num": "2.1"
},
{
"text": "..,s n be an English word, and P SW = p 1 ,p 2 ,...,p n be SW's pronunciation, where s i represents the i th grapheme and p i =\u03b4 p (s i ). Pronunciation estimation is a task to find the most relevant phoneme among a set of all possible phonemes, which can be derived from source grapheme s i . Table 3 shows an example of pronunciation estimation for b in board. In Table 3 , L1~L3 and R1~R3 represent the left contexts and right contexts, respectively. C0 means the current context (or focus). \u03b4 p (C0) means the estimated phoneme of C0. $ is a symbol for representing the start of words. The result can be interpreted as follows. The most relevant phoneme of b, /B/, can be produced with the context, f S , f GS , and f P in contexts of L1~L3, C0, and R1~R3. Other phonemes for o, a, r, and d in board are produced in the same manner. Thus, we can get the pronunciation of board as /B AO R D/ by concatenating the phoneme sequence.",
"cite_spans": [],
"ref_spans": [
{
"start": 294,
"end": 301,
"text": "Table 3",
"ref_id": "TABREF4"
},
{
"start": 366,
"end": 373,
"text": "Table 3",
"ref_id": "TABREF4"
}
],
"eq_spans": [],
"section": "Producing Pronunciation (\u03b4 p )",
"sec_num": "2.1"
},
{
"text": "Producing target graphemes (\u03b4 t :S\u00d7P\u2192T) is a function that finds the target grapheme in T for each gp i that is a result of \u03b4 p . A result of this step, GT, is represented by a sequence of gp i and its corresponding target graphemes generated by \u03b4 t , like GT={gt 1 , gt 2 ,\u2026, gt n ; gt i =(gp i ,\u03b4 t (gp i ))}. ) , phoneme type (f GP ) and \u03b4 t 's previous output (f T ) in the context window. Table 4 shows an example of \u03b4 t for b in board. \u03b4 t produces the most probable sequence of target graphemes (e.g. Korean), like \u03b4 t (gp",
"cite_spans": [],
"ref_spans": [
{
"start": 312,
"end": 313,
"text": ")",
"ref_id": null
},
{
"start": 394,
"end": 401,
"text": "Table 4",
"ref_id": "TABREF5"
}
],
"eq_spans": [],
"section": "Producing Target Graphemes (\u03b4 t )",
"sec_num": "2.2"
},
{
"text": "\u03b4 t (C0) f S $ $ $ b o a r 'b' f P $ $ $ /B/ /AO/ /~/ /R/ f GS $ $ $ C V V C f GP $ $ $ C V /~/ C f T $ $ $ Let",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Producing Target Graphemes (\u03b4 t )",
"sec_num": "2.2"
},
{
"text": "1 )= 'b', \u03b4 t (gp 2 )= 'o', \u03b4 t (gp 3 )='~', \u03b4 t (gp 4 )='~'",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Producing Target Graphemes (\u03b4 t )",
"sec_num": "2.2"
},
{
"text": ", and \u03b4 t (gp 5 )='deu' for board. Finally, the target language transliteration of board as 'bodeu' can be acquired by concatenating the sequence of produced target graphemes.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Producing Target Graphemes (\u03b4 t )",
"sec_num": "2.2"
},
{
"text": "In this section we will describe a way of modeling component functions using three machine learning algorithms (maximum entropy model, decision tree, and memorybased learning).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Machine Learning Algorithms for Each Component Function",
"sec_num": "3"
},
{
"text": "The maximum entropy model (MEM) is a widely used probability model that can incorporate heterogeneous information effectively [3] . In the maximum entropy model, an event ev is usually composed of a target event (te) and a history event (he), say ev=<te, he>. Event ev is represented by a bundle of feature functions, fe i (ev), which represent the existence of a certain characteristic in event ev. A feature function is a binary valued function. It is activated (fe i (ev)=1) when it meets its activating condition, otherwise it is deactivated (fe i (ev)=0) [ ",
"cite_spans": [
{
"start": 126,
"end": 129,
"text": "[3]",
"ref_id": "BIBREF2"
},
{
"start": 560,
"end": 561,
"text": "[",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Maximum Entropy Model",
"sec_num": "3.1"
},
{
"text": "f T (i-3,i-1)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Maximum Entropy Model",
"sec_num": "3.1"
},
{
"text": "where i is a index of the current source grapheme and phoneme to be transliterated and f X(l,m) represents features of feature type f X located from position l to position m. Target events are a set of target graphemes (phonemes) derived from history events of \u03b4 t (\u03b4 p ). Given history events, \u03b4 t (\u03b4 p ) finds the most probable target grapheme (phoneme), which maximizes formula (1). One important thing in designing a model based on the maximum entropy model is to determine feature functions which effectively support certain decision of the model. Our basic philosophy of feature function design for each component function is that context information collocated with the unit of interest is an important factor. With the philosophy, we determined the history events (or activating conditions) of the feature functions by combinations of features in feature types. Possible feature combinations for history events are between features in the same feature type and between features in different feature types. The used feature combinations in each component function are listed in Table 5 . ",
"cite_spans": [],
"ref_spans": [
{
"start": 1085,
"end": 1092,
"text": "Table 5",
"ref_id": "TABREF8"
}
],
"eq_spans": [],
"section": "Maximum Entropy Model",
"sec_num": "3.1"
},
{
"text": "= = i i GS S i i P i i p i i GP P i i GS S i i T i i p i t f f p p s f f f t p s s \u03b4 \u03b4 \u03b4 (1)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Maximum Entropy Model",
"sec_num": "3.1"
},
{
"text": "Decision tree learning is one of the most widely used and well-known methods for inductive inference [15] . ID3, which is a greedy algorithm and constructs decision trees in a top-down manner, adopts a statistical measure called information gain that measures how well a given feature (or attribute) separates training examples according to their target class [15] . We use C4.5 [15] , which is a well-known tool for decision tree learning and implementation of Quinlan's ID3 algorithm.",
"cite_spans": [
{
"start": 101,
"end": 105,
"text": "[15]",
"ref_id": "BIBREF14"
},
{
"start": 360,
"end": 364,
"text": "[15]",
"ref_id": "BIBREF14"
},
{
"start": 379,
"end": 383,
"text": "[15]",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Decision Tree",
"sec_num": "3.2"
},
{
"text": "Training data for each component function is represented by features of feature types in the context of L3~L1, C0, and R1~R3 as described in Table 3 . Fig. 2 shows a fraction of our decision trees for \u03b4 p and \u03b4 t in English-to-Korean transliteration (note that the left side represents the decision tree for \u03b4 p and the right side represents the decision tree for \u03b4 t ). A set of the target classes in the decision tree for \u03b4 p will be a set of phonemes and that for \u03b4 t will be a set of target graphemes. In Fig. 2, rectangles indicate a leaf node and circles indicate a decision node. In order to simplify our examples, we just use f S and f P in Fig. 2 . Intuitively, the most effective feature for \u03b4 p and \u03b4 t may be located in C0 among L3~L1, C0, and R1~R3 because the correct outputs of \u03b4 p and \u03b4 t strongly depend on source grapheme or phoneme in the C0 position.",
"cite_spans": [],
"ref_spans": [
{
"start": 141,
"end": 148,
"text": "Table 3",
"ref_id": "TABREF4"
},
{
"start": 151,
"end": 157,
"text": "Fig. 2",
"ref_id": null
},
{
"start": 509,
"end": 527,
"text": "Fig. 2, rectangles",
"ref_id": null
},
{
"start": 649,
"end": 655,
"text": "Fig. 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Decision Tree",
"sec_num": "3.2"
},
{
"text": "As we expected, the most effective feature in the decision trees is located in the C0 position like C0(f S ) for \u03b4 p and C0(f P ) for \u03b4 t (Note that the first feature to be tested is the most effective feature). In Fig. 2 , the decision tree for \u03b4 p outputs phoneme /AO/ for the instance x(SP) by retrieving the decision nodes C(f S )=o, R1(f S )=a, and R2(f S )=r represented with '*'. With the similar manner, the decision tree for \u03b4 t produces target grapheme (Korean grapheme) 'o' for the instance x(SPT) by retrieving the decision nodes from C0(f P )=/AO/ to R1(f P )=/~/ represented with '*'. ",
"cite_spans": [],
"ref_spans": [
{
"start": 215,
"end": 221,
"text": "Fig. 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Decision Tree",
"sec_num": "3.2"
},
{
"text": "C0(f S ): i C0(f S ): i L2(f S ): a L2(f S ): a L2(f S ): r L2(fS): r L2(f S ): $ L2(f S ): $ \u2026\u2026 'o' \u2192 d r a o b $ $ f S x(SPT) f P Feature type /D/ /R/ /~/ /AO/ /B/ $ $ \u03b4t R3 R2 R1 C0 L1 L2 L3",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Decision Tree",
"sec_num": "3.2"
},
{
"text": "Decision tree for \u03b4 t Decision tree for \u03b4 t Fig. 2 . Decision tree for \u03b4 p and\u03b4 t",
"cite_spans": [],
"ref_spans": [
{
"start": 44,
"end": 50,
"text": "Fig. 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Decision Tree",
"sec_num": "3.2"
},
{
"text": "Memory-based learning (MBL) is an example-based learning method. It is also called instance-based learning and case-based learning method. It is based on a k-nearest neighborhood algorithm [1] , [5] . MBL represents a training data as a vector. In the training phase, MBL puts all training data as examples in memory, and clusters some examples with a k-nearest neighborhood principle. It then outputs a target class using similarity-based reasoning between test data and examples in the memory. Let test data be x and a set of examples in a memory be Y, the similarity between x and Y is estimated by a distance function, \u2206(x,Y). MBL selects an example y i or a cluster of examples that are most similar to x, then assign a target class of the example to x's class. We use a memory-based learning tool called TiMBL (Tilburg memory-based learner) version 5.0 [5] .",
"cite_spans": [
{
"start": 189,
"end": 192,
"text": "[1]",
"ref_id": "BIBREF0"
},
{
"start": 195,
"end": 198,
"text": "[5]",
"ref_id": "BIBREF4"
},
{
"start": 859,
"end": 862,
"text": "[5]",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Memory-Based Learning",
"sec_num": "3.3"
},
{
"text": "Training data for each component function is represented by features of feature types in the context of L3~L1, C0, and R1~R3 as described in Table 4 . ",
"cite_spans": [],
"ref_spans": [
{
"start": 141,
"end": 148,
"text": "Table 4",
"ref_id": "TABREF5"
}
],
"eq_spans": [],
"section": "Memory-Based Learning",
"sec_num": "3.3"
},
{
"text": "<-3, +3> X GPC O X X Unbounded O GMEM O X X <-3, +3> O HWFST O O X - - Ours O O O <-3, +3> O",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Memory-Based Learning",
"sec_num": "3.3"
},
{
"text": "From the viewpoint of information usage, if a transliteration model adopts wide context window and considers previous outputs, it tends to show better performance. For example, GMEM that satisfies the conditions gives more accurate results than GDT which does not satisfy one of them. Because machine transliteration is sensitive to context, wider contexts give more powerful transliteration ability to machine transliteration systems. Note that the previous works, however, limit their context window size to 3, because the context window size over 3 degrades the performance [8] or does not change the performance of their transliteration model [9] . Determining reasonable context window size, therefore, is very important for machine transliteration.",
"cite_spans": [
{
"start": 577,
"end": 580,
"text": "[8]",
"ref_id": "BIBREF7"
},
{
"start": 647,
"end": 650,
"text": "[9]",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Memory-Based Learning",
"sec_num": "3.3"
},
{
"text": "For \"Context window size test\", we use \u03c8 C based on MBL, which shows the best performance among three machine learning algorithms in Table 6 . Experiments are performed by changing the context window size from 1 to 5. Table 8 shows results of context window size test. The results indicate that the best performance is shown when the context window size is 3. When the context window size is 1, there are many cases where the correct transliterations are not produced due to lack of information. For example, in order to produce the correct target grapheme of t in -tion, we need the right three graphemes of t, -ion. When the context window size is over 3, it is difficult to generalize the training data because of increase of variety of the training data. With the two reasons, our system shows the best performance when the context window size is 3. Table 8 also shows that context size should be at least 2 to avoid significant decrease of performance due to lack of contextual information. In summary, our method shows significant performance improvement, about 15%~23%, in English-to-Korean transliteration, and about 15%~ 43% in English-to-Japanese transliteration. Experiments show that a good transliteration system should consider; 1) source grapheme and phoneme along with their correspondence simultaneously and 2) reasonable context size and previous output. Our transliteration model satisfies the two conditions, thus it shows higher performance than the previous works.",
"cite_spans": [],
"ref_spans": [
{
"start": 133,
"end": 140,
"text": "Table 6",
"ref_id": null
},
{
"start": 218,
"end": 225,
"text": "Table 8",
"ref_id": "TABREF11"
},
{
"start": 854,
"end": 861,
"text": "Table 8",
"ref_id": "TABREF11"
}
],
"eq_spans": [],
"section": "Memory-Based Learning",
"sec_num": "3.3"
},
{
"text": "This paper has described a correspondence-based machine transliteration model (\u03c8 C ). Unlike the previous transliteration models, \u03c8 C uses correspondence between source grapheme and phoneme. The correspondence makes it possible for \u03c8 C to effectively produce both grapheme-based transliterations and phoneme-based transliterations. Moreover, the correspondence helps \u03c8 C to reduce transliteration ambiguities more easily. Experiments show that \u03c8 C is more powerful transliteration model than the previous transliteration models (\u03c8 C shows significant performance improvement, about 15%~23%, in English-to-Korean transliteration, and about 15%~ 43% in English-to-Japanese transliteration).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "5"
},
{
"text": "In future work, we will apply our transliteration model to English-to-Chinese transliteration model. In order to prove usefulness of our method in NLP applications, we need to apply our system to applications such as automatic bi-lingual dictionary construction, information retrieval, machine translation, speech recognition and so on.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "5"
},
{
"text": "In this paper, target language transliterations are represented with their Romanization form in a quotation mark ('') .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "Graphemes refer to the basic units (or the smallest contrastive units) of written language: for example, English has 26 graphemes or letters, Korean has 24, and German has 30.3 Phonemes are the simplest significant unit of sound (or the smallest contrastive units of the spoken language): for example, the /M/, /AE/, and /TH/ in math.4 In an English-to-Korean transliteration test set[14], we find that about 60% are phonemebased transliterations, while about 30% are grapheme-based ones. The others are transliterations generated by combining \u03c8 G and \u03c8 P .5 ARPAbet symbol will be used for representing phonemes. ARPAbet is one of the methods used for coding phonemes into ASCII characters (www.cs.cmu.edu/~laura/pages/arpabet.ps).In this paper, we will denote phonemes and pronunciation with two slashes like so : /AH/. Pronunciation represented in this paper is based on The CMU Pronunciation Dictionary and The American Heritage(r) Dictionary of the English Language.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "In this paper, '/~/' represents silence and '~' represents null target grapheme.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "Available at http://www.speech.cs.cmu.edu/cgi-bin/cmudict",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "http://www.csse.monash.edu.au/~jwb/j_edict.html",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "This work was supported by the Korea Ministry of Science and Technology, the Korea Ministry of Commerce, Industry and Energy, and the Korea Science and Engineering Foundation (KOSEF).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgement",
"sec_num": null
},
{
"text": "Training instances in a memory (\u03b4 t ) Training instances in a memory (\u03b4 t ) Fig. 3 . Memory-based learning for \u03b4 p and \u03b4 t",
"cite_spans": [],
"ref_spans": [
{
"start": 76,
"end": 82,
"text": "Fig. 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "annex",
"sec_num": null
},
{
"text": "We perform experiments for English-to-Korean and English-to-Japanese transliteration. English-to-Korean test set (EKSet) [14] consists of 7,185 English-Korean pairsthe number of training data is 6,185 and that of test data is 1,000. EKSet contains no transliteration variations. English-to-Japanese test set (EJSet), which is an Englishkatakana pair in EDICT 8 , consists of 10,398 -1,000 for test and the rest for training. EJSet contains transliteration variations, like (micro, 'maikuro') and (micro, 'mikuro'); the average number of Japanese transliterations for an English word is 1.15. Evaluation is performed by word accuracy (W. A.) in formula (2) . We perform two experiments called \"Comparison test\" and \"Context window size test\". In the \"Comparison test\", we compare our \u03c8 C with the previous works. In \"Context window size test\", we evaluate the performance of our transliteration model depending on context window size. Table 6 shows results of \"Comparison test\". MEM, DT, and MBL represent \u03c8 C based on maximum entropy model, decision tree, and memory-based learning, respectively. GDT [8] , GPC [9] , GMEM [7] and HWFST [4] , which are one of the best machine transliteration methods in English-to-Korean transliteration and English-to-Japanese transliteration, are compared with \u03c8 C . Table 7 shows the key feature of each method in the viewpoint of information type (SG, PH, COR) and information usage (Context size, POut). Information type indicates that each transliteration method belongs to which transliteration model. For example, GDT, GPC, and GMEM will belong to \u03c8 G because they use only the source grapheme; while HWFST belongs to \u03c8 H . Information usage gives information about what kinds of information each transliteration method can deal with. From the viewpoint of information type, phoneme and correspondence, which most previous works do not consider, is the key point of the performance gap between our method and the previous works.",
"cite_spans": [
{
"start": 121,
"end": 125,
"text": "[14]",
"ref_id": "BIBREF13"
},
{
"start": 359,
"end": 360,
"text": "8",
"ref_id": "BIBREF7"
},
{
"start": 652,
"end": 655,
"text": "(2)",
"ref_id": "BIBREF1"
},
{
"start": 1101,
"end": 1104,
"text": "[8]",
"ref_id": "BIBREF7"
},
{
"start": 1111,
"end": 1114,
"text": "[9]",
"ref_id": "BIBREF8"
},
{
"start": 1122,
"end": 1125,
"text": "[7]",
"ref_id": "BIBREF6"
},
{
"start": 1136,
"end": 1139,
"text": "[4]",
"ref_id": "BIBREF3"
}
],
"ref_spans": [
{
"start": 934,
"end": 941,
"text": "Table 6",
"ref_id": null
},
{
"start": 1302,
"end": 1309,
"text": "Table 7",
"ref_id": null
}
],
"eq_spans": [],
"section": "Experiments",
"sec_num": "4"
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Lazy learning: Special issue editorial",
"authors": [
{
"first": "D",
"middle": [
"W"
],
"last": "Aha",
"suffix": ""
}
],
"year": 1997,
"venue": "Artificial Intelligence Review",
"volume": "11",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Aha, D. W. Lazy learning: Special issue editorial. Artificial Intelligence Review, 11:710, (1997).",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Translating Named Entities Using Monolingual and Bilingual Resources",
"authors": [
{
"first": "Y",
"middle": [],
"last": "Al-Onaizan",
"suffix": ""
},
{
"first": "Kevin",
"middle": [],
"last": "Knight",
"suffix": ""
}
],
"year": 2002,
"venue": "the Proceedings of ACL 2002",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Al-Onaizan Y. and Kevin Knight, \"Translating Named Entities Using Monolingual and Bilingual Resources\", In the Proceedings of ACL 2002, (2002)",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "A maximum entropy approach to natural language processing",
"authors": [
{
"first": "A",
"middle": [],
"last": "Berger",
"suffix": ""
},
{
"first": "S",
"middle": [
"Della"
],
"last": "Pietra",
"suffix": ""
},
{
"first": "V",
"middle": [],
"last": "Della Pietra",
"suffix": ""
}
],
"year": 1996,
"venue": "Computational Linguistics",
"volume": "22",
"issue": "1",
"pages": "39--71",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Berger, A., S. Della Pietra, and V. Della Pietra. , A maximum entropy approach to natural language processing. Computational Linguistics, 22(1), (1996), 39-71",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Improving Back-Transliteration by Combining Information Sources",
"authors": [
{
"first": "Bilac",
"middle": [],
"last": "Slaven",
"suffix": ""
},
{
"first": "Hozumi",
"middle": [],
"last": "Tanaka",
"suffix": ""
}
],
"year": 2004,
"venue": "Proc. of IJC-NLP2004",
"volume": "",
"issue": "",
"pages": "542--547",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Bilac Slaven and Hozumi Tanaka. \"Improving Back-Transliteration by Combining Infor- mation Sources\". In Proc. of IJC-NLP2004, (2004) 542-547",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Timble TiMBL: Tilburg Memory Based Learner",
"authors": [
{
"first": "W",
"middle": [],
"last": "Daelemans",
"suffix": ""
},
{
"first": "Jakub",
"middle": [],
"last": "Zavrel",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Ko Van Der",
"suffix": ""
},
{
"first": "Antal",
"middle": [],
"last": "Sloot",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Van Den",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Bosch",
"suffix": ""
}
],
"year": 2002,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Daelemans, W., Jakub Zavrel, Ko van der Sloot, and Antal van den Bosch, 2002, Timble TiMBL: Tilburg Memory Based Learner, version 4.3, Reference Guide, ILK Technical Report 02-10, (2002).",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Japanese/English Cross-Language Information Retrieval: Exploration of Query Translation and Transliteration",
"authors": [
{
"first": "Atsushi",
"middle": [],
"last": "Fujii",
"suffix": ""
},
{
"first": "Ishikawa",
"middle": [],
"last": "Tetsuya",
"suffix": ""
}
],
"year": 2001,
"venue": "Computers and the Humanities",
"volume": "35",
"issue": "4",
"pages": "389--420",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Fujii, Atsushi and Tetsuya, Ishikawa. Japanese/English Cross-Language Information Re- trieval: Exploration of Query Translation and Transliteration. Computers and the Humani- ties, Vol.35, No.4, (2001) 389-420",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Transliteration Considering Context Information Based on the Maximum Entropy Method",
"authors": [
{
"first": "I",
"middle": [],
"last": "Goto",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Kato",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Uratani",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Ehara",
"suffix": ""
}
],
"year": 2003,
"venue": "Proceedings of MT-Summit IX",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Goto, I., N. Kato, N. Uratani and T. Ehara, Transliteration Considering Context Informa- tion Based on the Maximum Entropy Method, In Proceedings of MT-Summit IX, (2003)",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Automatic Transliteration and Back-transliteration by Decision Tree Learning",
"authors": [
{
"first": "B",
"middle": [
"J"
],
"last": "Kang",
"suffix": ""
},
{
"first": "K-S",
"middle": [],
"last": "Choi",
"suffix": ""
}
],
"year": 2000,
"venue": "Proceedings of the 2nd International Conference on Language Resources and Evaluation",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kang B.J. and K-S. Choi, \"Automatic Transliteration and Back-transliteration by Decision Tree Learning\", In Proceedings of the 2nd International Conference on Language Re- sources and Evaluation, (2000)",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "English-to-Korean Transliteration using Multiple Unbounded Overlapping Phoneme Chunks",
"authors": [
{
"first": "I",
"middle": [
"H"
],
"last": "Kang",
"suffix": ""
},
{
"first": "G",
"middle": [
"C"
],
"last": "Kim",
"suffix": ""
}
],
"year": 2000,
"venue": "Proceedings of the 18th International Conference on Computational Linguistics",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kang, I.H. and G.C. Kim, \"English-to-Korean Transliteration using Multiple Unbounded Overlapping Phoneme Chunks\", In Proceedings of the 18th International Conference on Computational Linguistics, (2000).",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Machine Transliteration",
"authors": [
{
"first": "K",
"middle": [],
"last": "Knight",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Graehl",
"suffix": ""
}
],
"year": 1997,
"venue": "Proceedings. of the 35th Annual Meetings of the Association for Computational Linguistics (ACL)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Knight, K. and J. Graehl, \"Machine Transliteration\". In Proceedings. of the 35th Annual Meetings of the Association for Computational Linguistics (ACL), (1997)",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "English to Korean Statistical transliteration for information retrieval",
"authors": [
{
"first": "J",
"middle": [
"S"
],
"last": "Lee",
"suffix": ""
},
{
"first": "K",
"middle": [
"S"
],
"last": "Choi",
"suffix": ""
}
],
"year": 1998,
"venue": "Computer Processing of Oriental Languages",
"volume": "12",
"issue": "1",
"pages": "17--37",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Lee, J. S. and K. S. Choi, English to Korean Statistical transliteration for information re- trieval. Computer Processing of Oriental Languages, 12(1), (1998), 17-37.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "An English-Korean transliteration and Retransliteration model for Cross-lingual information retrieval",
"authors": [
{
"first": "J",
"middle": [
"S"
],
"last": "Lee",
"suffix": ""
}
],
"year": 1999,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Lee, J.S., An English-Korean transliteration and Retransliteration model for Cross-lingual information retrieval, PhD Thesis, Computer Science Dept., KAIST, (1999)",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "A Joint Source-Channel Model for Machine Transliteration",
"authors": [
{
"first": "Li",
"middle": [],
"last": "Haizhou",
"suffix": ""
},
{
"first": "Min",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Jian",
"middle": [],
"last": "Su",
"suffix": ""
}
],
"year": 2004,
"venue": "",
"volume": "",
"issue": "",
"pages": "159--166",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Li Haizhou, Min Zhang and Jian Su , A Joint Source-Channel Model for Machine Trans- literation , ACL 2004, (2004), 159-166",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Foreign dictionary, Sung-An-Dang publisher",
"authors": [
{
"first": "Y",
"middle": [
"S"
],
"last": "Nam",
"suffix": ""
}
],
"year": 1997,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nam, Y.S., Foreign dictionary, Sung-An-Dang publisher, (1997)",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "C4.5: Programs for Machine Learning",
"authors": [
{
"first": "J",
"middle": [
"R"
],
"last": "Quinlan",
"suffix": ""
}
],
"year": 1993,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Quinlan, J.R., \"C4.5: Programs for Machine Learning\", Morgan Kauffman, (1993)",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Maximum Entropy Modeling Toolkit for Python and C++",
"authors": [
{
"first": "Le",
"middle": [],
"last": "Zhang",
"suffix": ""
}
],
"year": 2004,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Zhang, Le. Maximum Entropy Modeling Toolkit for Python and C++. http://www.nlplab.cn/zhangle/, (2004)",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"num": null,
"text": "The overall system architecture",
"type_str": "figure",
"uris": null
},
"FIGREF1": {
"num": null,
"text": "S ): y R1(fS): y R1(f S ): e or q R1(fS): e or q R1(f S ): a(S ): x R1(fS): x \u2026\u2026 R2(f S ): d R2(f S ): d R2(f S ): r(*) R2( R2(f f S S): r(*) ): r(*) R2(f S ): others R2(f S ): others R2(f S ): $ R2(f S ): $ /OW/ /OW/ /OW/ /OW/ /AO/(*) /AO/(*) /AO/(*) R1(f S ): b R1(fS): b L2(f S ): a L2(f S ): a L2(f S ): r L2(f S ): r L2(f S ): $ L2(f S ): $ \u2026\u2026 S ): a C0(f S ): a C0(f S ): e C0(f S ): e C0(f S ): o(S ): others C0(fS): others \u2026\u2026 R1(f P ): /R/ R1(f P ): /R/ R1(f S ): /~/(*) R1( R1(f f S S): /~/(*) ): /~/(*) R1(f P ): others R1(f P ): others",
"type_str": "figure",
"uris": null
},
"FIGREF2": {
"num": null,
"text": "shows examples of \u03b4 p and \u03b4 t based on MBL in English-to-Korean transliteration. In order to simplify our examples, we just use f S and f P in Fig. 3. All training data are represented with their features in the context of L3~L1, C0, and R1~R3 and their target classes for \u03b4 p and \u03b4 t . They are stored in the memory through a training phase. Feature weighting for dealing with features of differing importance is also performed in the training phase. In Fig. 3, \u03b4 p based on MBL outputs the phoneme /AO/ for x(SP) by comparing the similarities between x(SP) and Y using distance metric \u2206(x(SP),Y). With the similar manner, \u03b4 t based on MBL outputs the target grapheme 'o'.",
"type_str": "figure",
"uris": null
},
"TABREF0": {
"type_str": "table",
"html": null,
"num": null,
"text": "Examples of Korean graphemes derived from source grapheme a and its correspond-",
"content": "<table><tr><td colspan=\"3\">ing phoneme: the underline indicates source graphemes corresponding to each phoneme in the</td></tr><tr><td>phoneme column</td><td/><td/></tr><tr><td>Korean grapheme</td><td>Phoneme</td><td>Example</td></tr><tr><td>'a'</td><td>/AA/</td><td>adagio, safari, vivace</td></tr><tr><td>'ae'</td><td>/AE/</td><td>advantage, alabaster, travertine</td></tr><tr><td>'ei'</td><td>/EY/</td><td>chamber, champagne, chaos</td></tr><tr><td>'i'</td><td>/IH/</td><td>advantage, average, silage</td></tr><tr><td>'o'</td><td>/AO/</td><td>allspice, ball, chalk</td></tr></table>"
},
"TABREF3": {
"type_str": "table",
"html": null,
"num": null,
"text": "Pronunciation estimation in \u03b4 p and \u03b4 t are trained by machine learning algorithms. To train each component function, we need features that represent training instance and data.Table 2shows five feature types, f S , f P , f GS , f GP, and f T that our model uses. Depending on component functions, different feature types are used. For example, \u03b4 p",
"content": "<table><tr><td colspan=\"2\">Feature Type</td><td>Description</td><td>Possible feature values</td></tr><tr><td>f S,GS</td><td>f S</td><td>Source graphemes</td><td>Source grapheme in S; 26 alphabets</td></tr><tr><td/><td/><td/><td>for English</td></tr><tr><td/><td>f GS</td><td>Source grapheme type</td><td>Consonant (C), and Vowel (V)</td></tr><tr><td>f P,GP</td><td>f P</td><td>Phonemes</td><td>Phonemes in P (/AA/, /AE/, etc.)</td></tr><tr><td/><td>f GP</td><td>Phoneme type</td><td>Consonant (C), Vowel (V), Semi-</td></tr><tr><td/><td/><td/><td>vowel (SV) and silence (/~/)</td></tr><tr><td/><td>f T</td><td>Target graphemes</td><td>Target graphemes in T</td></tr></table>"
},
"TABREF4": {
"type_str": "table",
"html": null,
"num": null,
"text": "An example of pronunciation estimation for b in board",
"content": "<table><tr><td>Feature type L3</td><td>L2</td><td>L1</td><td>C0 R1</td><td>R2</td><td>R3</td></tr></table>"
},
"TABREF5": {
"type_str": "table",
"html": null,
"num": null,
"text": "An example of \u03b4 t for b in board",
"content": "<table><tr><td>Feature type</td><td>L3</td><td>L2</td><td>L1</td><td>C0</td><td>R1</td><td>R2</td><td>R3</td></tr></table>"
},
"TABREF6": {
"type_str": "table",
"html": null,
"num": null,
"text": "SW=s 1 ,s 2 ,...,s n be a source language word, P SW = p 1 ,p 2 ,...,p n be SW's pronunciation and T SW = t 1 , t 2 ,...,t n be a target language word of SW, where s i , \u03b4 p (s i )=p i and \u03b4 t (gp i ) = t i represent the i th source grapheme, phoneme corresponding to s i , and target graph-",
"content": "<table/>"
},
"TABREF7": {
"type_str": "table",
"html": null,
"num": null,
"text": "3].\u03b4 p and \u03b4 t based on the maximum entropy model can be represented as formula(1).",
"content": "<table><tr><td>History events in each component function are made from the left, right and current context. For example, history events for \u03b4 t are composed of f S,GS (i-3,i+3) , f P,GP (i-3,i+3) , and</td></tr></table>"
},
"TABREF8": {
"type_str": "table",
"html": null,
"num": null,
"text": "Used feature combinations for history eventsIn formula (1), history events of \u03b4 p and \u03b4 t are defined by the conditions described inTable 5. Target events of \u03b4 t are all possible target graphemes derived from its history events; while those of \u03b4 p are all possible phonemes derived from its history events. In order to model each component function based on MEM, Zhang's maximum entropy modeling tool is used[16].",
"content": "<table><tr><td>\u03b4 p</td><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td colspan=\"2\">\u03b4 t</td><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/></tr><tr><td colspan=\"12\">Between features in the same feature</td><td/><td/><td colspan=\"13\">Between features in the same feature</td></tr><tr><td>type</td><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td colspan=\"3\">type</td><td/><td/><td/><td/><td/><td/><td/><td/><td/></tr><tr><td colspan=\"12\">Between features in different feature</td><td/><td/><td colspan=\"13\">Between features in different feature</td></tr><tr><td>types</td><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td colspan=\"5\">types</td><td/><td/><td/><td/><td/><td/><td/></tr><tr><td colspan=\"3\">f S,GS and f P</td><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td colspan=\"6\">f S,GS and f P,GP</td></tr><tr><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td colspan=\"6\">f S,GS and f T</td></tr><tr><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td colspan=\"6\">f P,GP and f T</td></tr><tr><td>(</td><td>,</td><td>(</td><td colspan=\"2\">))</td><td>arg</td><td colspan=\"2\">max</td><td>(</td><td>|</td><td/><td>\u2212</td><td>3</td><td>,</td><td>1 \u2212</td><td>,</td><td/><td>,</td><td/><td>3 \u2212</td><td>,</td><td>+</td><td>3</td><td>,</td><td>,</td><td>\u2212</td><td>3</td><td>,</td><td>+</td><td>3</td><td>)</td></tr><tr><td>(</td><td>)</td><td colspan=\"2\">arg</td><td colspan=\"2\">max</td><td>(</td><td>|</td><td>3 \u2212</td><td>,</td><td>1 \u2212</td><td>,</td><td/><td>,</td><td/><td>\u2212</td><td>, 3</td><td>+</td><td>3</td><td>)</td><td/><td/><td/><td/><td/><td/></tr></table>"
},
"TABREF9": {
"type_str": "table",
"html": null,
"num": null,
"text": "",
"content": "<table><tr><td>y i</td><td>Feature type</td><td>L3</td><td>L2</td><td>L1</td><td>C0</td><td>R1</td><td/><td>R2</td><td>R3</td><td colspan=\"2\">\u03b4 p (C0)</td><td>\u0394 (x(SP),y i )</td></tr><tr><td>1*</td><td>f S</td><td>$</td><td>a</td><td>b</td><td>o</td><td>a</td><td/><td>r</td><td>d</td><td colspan=\"2\">/AO/</td><td>0.93</td></tr><tr><td>2</td><td>f S</td><td>h</td><td>a</td><td>b</td><td>o</td><td>$</td><td/><td>$</td><td>$</td><td colspan=\"2\">/OW/</td><td>0.38</td></tr><tr><td>3</td><td>fS</td><td>$</td><td>$</td><td>b</td><td>o</td><td>a</td><td/><td>s</td><td>t</td><td colspan=\"2\">/OW/</td><td>0.81</td></tr><tr><td>4</td><td>fS</td><td>$</td><td>$</td><td>b</td><td>o</td><td>a</td><td/><td>t</td><td>$</td><td colspan=\"2\">/OW/</td><td>0.81</td></tr><tr><td>5</td><td>f S</td><td>$</td><td>$</td><td>c</td><td>o</td><td>a</td><td/><td>r</td><td>s</td><td colspan=\"2\">/AO/</td><td>0.73</td></tr><tr><td>6</td><td>f S</td><td>$</td><td>e</td><td>d</td><td>o</td><td>a</td><td/><td>r</td><td>d</td><td colspan=\"2\">/W/</td><td>0.75</td></tr><tr><td>7</td><td>fS</td><td>c</td><td>k</td><td>s</td><td>a</td><td>w</td><td/><td>$</td><td>$</td><td colspan=\"2\">/AO/</td><td>0.16</td></tr><tr><td>8</td><td>fS</td><td>e</td><td>a</td><td>b</td><td>o</td><td>u</td><td/><td>t</td><td>$</td><td colspan=\"2\">/UW/</td><td>0.51</td></tr><tr><td/><td>x(SP)</td><td colspan=\"2\">Feature type</td><td>L3</td><td>L2</td><td>L1</td><td>C0</td><td>R1</td><td>R2</td><td>R3</td><td>\u03b4 p</td></tr><tr><td/><td/><td>f S</td><td/><td>$</td><td>$</td><td>b</td><td>o</td><td>a</td><td>r</td><td>d</td><td>\u2192</td><td>/AO/</td></tr></table>"
},
"TABREF10": {
"type_str": "table",
"html": null,
"num": null,
"text": "Evaluation results of \"Comparison test\" Key features of our machine transliteration model and the previous works: SG, PH, COR and POut represent source grapheme, phoneme, correspondence and previous output, respectively",
"content": "<table><tr><td>Method</td><td/><td>EKSet</td><td/><td/><td>EJSet</td></tr><tr><td/><td/><td>W.A</td><td colspan=\"2\">Chg %</td><td>W.A</td><td>Chg %</td></tr><tr><td>GDT</td><td/><td>51.4%</td><td>23.2%</td><td/><td>50.3%</td><td>43.5%</td></tr><tr><td>GPC</td><td/><td>55.1%</td><td>17.6%</td><td/><td>53.2%</td><td>35.7%</td></tr><tr><td>GMEM</td><td/><td>55.9%</td><td>16.4%</td><td/><td>56.2%</td><td>28.5%</td></tr><tr><td>HWFST</td><td/><td>58.3%</td><td>14.7%</td><td/><td>62.5%</td><td>15.5%</td></tr><tr><td>DT</td><td/><td>62.0%</td><td>7.3%</td><td/><td>66.8%</td><td>8.1%</td></tr><tr><td>MEM</td><td/><td>63.3%</td><td>5.4%</td><td/><td>67.0%</td><td>7.8%</td></tr><tr><td>MBL</td><td/><td>66.9%</td><td>0%</td><td/><td>72.2%</td><td>0%</td></tr><tr><td>Method</td><td>SG</td><td>PH</td><td>COR</td><td colspan=\"2\">Context size</td><td>POut</td></tr><tr><td>GDT</td><td>O</td><td>X</td><td>X</td><td/><td/></tr></table>"
},
"TABREF11": {
"type_str": "table",
"html": null,
"num": null,
"text": "Evaluation results of \"Context window size test\"",
"content": "<table><tr><td>Context Size</td><td>EKSet</td><td>EJSet</td></tr><tr><td>1</td><td>54.5%</td><td>62.7%</td></tr><tr><td>2</td><td>63.3%</td><td>70.0%</td></tr><tr><td>3</td><td>66.9%</td><td>72.2%</td></tr><tr><td>4</td><td>63.9%</td><td>70.7%</td></tr><tr><td>5</td><td>63.8%</td><td>69.3%</td></tr></table>"
}
}
}
} |