File size: 63,642 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 |
{
"paper_id": "I08-1009",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:41:58.238701Z"
},
"title": "A More Discerning and Adaptable Multilingual Transliteration Mechanism for Indian Languages",
"authors": [
{
"first": "Harshit",
"middle": [],
"last": "Surana",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Language Tech. Research Centre IIIT",
"location": {
"settlement": "Hyderabad",
"country": "India"
}
},
"email": "surana.h@gmail.com"
},
{
"first": "Anil",
"middle": [
"Kumar"
],
"last": "Singh",
"suffix": "",
"affiliation": {},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "Transliteration is the process of transcribing words from a source script to a target script. These words can be content words or proper nouns. They may be of local or foreign origin. In this paper we present a more discerning method which applies different techniques based on the word origin. The techniques used also take into account the properties of the scripts. Our approach does not require training data on the target side, while it uses more sophisticated techniques on the source side. Fuzzy string matching is used to compensate for lack of training on the target side. We have evaluated on two Indian languages and have achieved substantially better results (increase of up to 0.44 in MRR) than the baseline and comparable to the state of the art. Our experiments clearly show that word origin is an important factor in achieving higher accuracy in transliteration.",
"pdf_parse": {
"paper_id": "I08-1009",
"_pdf_hash": "",
"abstract": [
{
"text": "Transliteration is the process of transcribing words from a source script to a target script. These words can be content words or proper nouns. They may be of local or foreign origin. In this paper we present a more discerning method which applies different techniques based on the word origin. The techniques used also take into account the properties of the scripts. Our approach does not require training data on the target side, while it uses more sophisticated techniques on the source side. Fuzzy string matching is used to compensate for lack of training on the target side. We have evaluated on two Indian languages and have achieved substantially better results (increase of up to 0.44 in MRR) than the baseline and comparable to the state of the art. Our experiments clearly show that word origin is an important factor in achieving higher accuracy in transliteration.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Transliteration is a crucial factor in Cross Lingual Information Retrieval (CLIR). It is also important for Machine Translation (MT), especially when the languages do not use the same scripts. It is the process of transforming a word written in a source language into a word in a target language without the aid of a resource like a bilingual dictionary. Word pronunciation is usually preserved or is modified according to the way the word should be pronounced in the target language. In simple terms, it means finding out how a source word should be written in the script of the target languages such that it is acceptable to the readers of the target language.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "One of the main reasons of the importance of transliteration from the point of view of Natural Language Processing (NLP) is that Out Of Vocabulary (OOV) words are quite common since every lexical resource is very limited in practical terms. Such words include named entities, technical terms, rarely used or 'difficult' words and other borrowed words, etc. The OOV words present a challenge to NLP applications like CLIR and MT. In fact, for very close languages which use different scripts (like Hindi and Urdu), the problem of MT is almost an extension of transliteration.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "A substantial percentage of these OOV words are named entities Davis and Ogden, 1998) . It has also been shown that cross language retrieval performance (average precision) reduced by more than 50% when named entities in the queries were not transliterated .",
"cite_spans": [
{
"start": 63,
"end": 85,
"text": "Davis and Ogden, 1998)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Another emerging application of transliteration (especially in the Indian context) is for building input methods which use QWERTY keyboard for people who are more comfortable typing in English. The idea is that the user types Roman letters but the input method transforms them into letters of Indian language (IL) scripts. This is not as simple as it seems because there is no clear mapping between Roman letters and IL letters. Moreover, the output word should be a valid word. Several commercial efforts have been started in this direction due to the lack of a good (and familiar) input mech-anism for ILs. These efforts include the Google Transliteration mechanism 1 and Quilpad 2 . (Rathod and Joshi, 2002) have also developed more intuitive input mechanisms for phonetic scripts like Devanagari.",
"cite_spans": [
{
"start": 686,
"end": 710,
"text": "(Rathod and Joshi, 2002)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Our efforts take into account the type of the word, the similarities among ILs and the characteristics of the Latin and IL scripts. We use a sophisticated technique and machine learning on the source language (English) side, while a simple and light technique on the target (IL) side. The advantage of our approach is that it requires no resources except unannotated corpus (or pages crawled from the Web) on the IL side (which is where the resources are scarce). The method easily generalizes to ILs which use Brahmi origin scripts. Our method has been designed such that it can be used for more conventional applications (MT, CLIR) as well as for applications like building an input mechanism.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Much of the work for transliteration in ILs has been done from one Indian script to another. One of the major work is of Punjabi machine transliteration (Malik, 2006) . This work tries to address the problem of transliteration for Punjabi language from Shahmukhi (Arabic script) to Gurmukhi using a set of transliteration rules (character mappings and dependency rules). Om transliteration scheme (Ganapathiraju et al., 2005) also provides a script representation which is common for all Indian languages. The display and input are in human readable Roman script. Transliteration is partly phonetic. (Sinha, 2001) Aswani et. al (Aswani and Gaizauskas, 2005 ) have used a transliteration similarity mechanism to align English-Hindi parallel texts. They used character based direct correspondences between Hindi and English to produce possible transliterations. Then they apply edit distance based similarity to select the most probable transliteration in the English text. However, such method can only be appropriate for aligning parallel texts as the number of possible candidates is quite small.",
"cite_spans": [
{
"start": 153,
"end": 166,
"text": "(Malik, 2006)",
"ref_id": "BIBREF6"
},
{
"start": 397,
"end": 425,
"text": "(Ganapathiraju et al., 2005)",
"ref_id": "BIBREF3"
},
{
"start": 600,
"end": 613,
"text": "(Sinha, 2001)",
"ref_id": "BIBREF16"
},
{
"start": 628,
"end": 656,
"text": "(Aswani and Gaizauskas, 2005",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The paper is structured as follows. In Section-2, we discuss the problem of a high degree of variation in Indian words, especially when written in Latin script. In Section-3, we explain the idea of using information about the word origin for improving transliteration. Then in Section-4 we describe the method that we use for guessing the word origin. Once the word origin is guessed, we can apply one of the two methods for transliteration depending on the word origin. These two methods are described in Section-5 and Section-6, respectively. Fuzzy string matching, which plays an important role in our approach, is described in Section-7. In Section-8 we put together all the elements covered in the preceding sections and explain the Discerning Adaptable Transliteration Mechanism. Section-9 presents the evaluation of our approach in comparison with two baseline methods, one of which uses knowledge about word origin. Finally, in Section-10 we present the conclusions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Since the purpose of our work is not only to transliterate named entities but to be useful for applications like input mechanisms, we had to consider some other issues too which may not be considered directly related to transliteration. One of these is that there is a lot of spelling variation in ILs. This variation is much more when the IL words are written using the Latin script ( Table-1 ). In other words, the amount of ambiguity is very high when we try to build a system that can be used for purposes like designing input mechanisms, instead of just for transliteration of NEs etc. for MT or CLIR. One reason for very high variation in the latter case is that unlike Romaji for Japanese (which is taught in schools in Japan), there is no widely adopted transliteration scheme using the Latin script, although there are a number of standard schemes, which are not used by common users. At present the situation is that most Indians use Indian scripts while writing in ILs, but use the Latin script when communicating online. ILs are rarely used for official communication, except in government offices in some states.",
"cite_spans": [],
"ref_spans": [
{
"start": 386,
"end": 393,
"text": "Table-1",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Variation in Indian Words in Latin Script",
"sec_num": "2"
},
{
"text": "Previous work for other languages has shown that word origin plays a part in how the word should be transliterated (Oh and Choi, 2002; May et al., 2004) . Llitjos and Black (Llitjos and Black, 2001) had shown that the knowledge of language origin can substantially improve pronunciation generation accuracy. This information has been used to get better results (Oh and Choi, 2002) . They first checked whether the word origin is Greek or not before selecting one of the two methods for transliteration. This approach improved the results substantially. However, they had used a set of prefixes and suffixes to identify the word origin. Such an approach is not scalable. In fact, in a large number of cases, word origin cannot be identified by using list of affixes. For ILs, we also define two categories of words: words which can be roughly considered Indian and those which can be roughly considered foreign. Note that 'Indian' and 'foreign' are just loose labels here. Indian words, which include proper nouns and also common vocabulary words, are more relevant in applications like input methods. Two different methods are used for transliterating, as explained later.",
"cite_spans": [
{
"start": 115,
"end": 134,
"text": "(Oh and Choi, 2002;",
"ref_id": "BIBREF9"
},
{
"start": 135,
"end": 152,
"text": "May et al., 2004)",
"ref_id": "BIBREF7"
},
{
"start": 155,
"end": 198,
"text": "Llitjos and Black (Llitjos and Black, 2001)",
"ref_id": "BIBREF5"
},
{
"start": 361,
"end": 380,
"text": "(Oh and Choi, 2002)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Word Origin and Two Ways of Transliteration",
"sec_num": "3"
},
{
"text": "Previously (Llitjos and Black, 2001 ) used probabilities of all trigrams to belong to a particular language as an measure to disambiguate word origins. We use a more sophisticated method that has been successfully used for language and encoding identification (Singh, 2006a) .",
"cite_spans": [
{
"start": 11,
"end": 35,
"text": "(Llitjos and Black, 2001",
"ref_id": "BIBREF5"
},
{
"start": 260,
"end": 274,
"text": "(Singh, 2006a)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Disambiguating Word Origin",
"sec_num": "4"
},
{
"text": "We first prepare letter based 5-gram models from the lists of two kinds of words (Indian and foreign). Then we combine n-grams of all orders and rank them according to their probability in descending order. Only the top N n-grams are retained and the rest are pruned. Now we have two probability distributions which can be compared by a measure of distributional similarity. The measure used is symmetric cross entropy or SCE (Singh, 2006a) .",
"cite_spans": [
{
"start": 426,
"end": 440,
"text": "(Singh, 2006a)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Disambiguating Word Origin",
"sec_num": "4"
},
{
"text": "Since the accuracy of identification is low if test data is very low, which is true in our case because we are trying to identify the class of a single word, we had to extend the method used by Singh. One major extension was that we add word beginning and ending markers to all the words in training as well as test data. This is because n-grams at beginning, middle and end of words should be treated differently if we want to identify the 'language' (or class) of the word.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Disambiguating Word Origin",
"sec_num": "4"
},
{
"text": "For every given word, we get a probability about its origin based on SCE. Based on this probability measure, transliteration is performed using different techniques for different classes (Indian or foreign). In case of ambiguity, transliteration is performed using both methods and the probabilities are used to get the final ranking of all possible transliterations.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Disambiguating Word Origin",
"sec_num": "4"
},
{
"text": "These words include named entities (George Bush) and more common nouns (station, computer) which are regularly used in ILs. To generate transliteration candidates for such words, we first try to guess the word pronunciation or use a lookup dictionary (if available) to find it. Then we use some simple manually created mappings, which can be used for all Indian languages. Note that these mappings are very few in number ( Figure-1 and Figure-2 ) and can be easily created by non-linguistically trained people. They play only a small role in the method because other steps (like fuzzy string matching) do most of the work.",
"cite_spans": [],
"ref_spans": [
{
"start": 423,
"end": 431,
"text": "Figure-1",
"ref_id": null
},
{
"start": 436,
"end": 444,
"text": "Figure-2",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Transliteration of Foreign Words",
"sec_num": "5"
},
{
"text": "For our experiments, we used the CMU speech dictionary as the lookup, and also to train pronunciation estimation. If a word is not in the CMU dictionary, we estimate the word pronunciation, as explained later.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Transliteration of Foreign Words",
"sec_num": "5"
},
{
"text": "We directly map from English phonemes to IL letters. This is based on our observation that a foreign word is usually transliterated in almost the same way as it is pronounced. Almost all English phonemes can be roughly mapped to specific letters (representing phonemes, as IL scripts are phonetic in na-ture) in ILs. Similar observations have been made about Hindi by Su-Youn Yoon, Kyoung-Young Kim and Richard Sproat (Yoon et al., 2007) . We have prepared our own mappings with help from native speakers of the languages concerned, which is relatively quite a simple task since the letters in Indic scripts correspond closely with phonemes.",
"cite_spans": [
{
"start": 418,
"end": 437,
"text": "(Yoon et al., 2007)",
"ref_id": "BIBREF17"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Transliteration of Foreign Words",
"sec_num": "5"
},
{
"text": "These words include (mainly Indian) named entities of (e.g. Taj Mahal, Manmohan Singh) and common vocabulary words (common nouns, verbs) which need to be transliterated. They also include words which are spelled similar to the way Indian words are spelled when written in Latin (e.g. Baghdad, Husain). As stated earlier, this class of words are much more relevant for an input method using a QWERTY keyboard.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Transliteration of Indian Words",
"sec_num": "6"
},
{
"text": "Since words of Indian origin usually have phonetic spellings when they are written in English (Latin), the issue of pronunciation estimation or lookup is not important. However, there can be many possible vowel and consonant segments which can be formed out of a single word. For example 'ai' can be interpreted as a single vowel with sound AE (as in Husain), or as two vowels AA IH (as in Rai). To perform segmentation, we have a simple program which produces candidates for all possible segments. This program uses a few rules defining the possible consonant and vowel combinations. Now we simply map these segments to their nearest IL letters (or letter combinations). This is also done using a simple set of mappings, which do not contain any probabilities or contexts. This step generates transliteration candidates. These are then filtered and ranked using fuzzy string matching.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Transliteration of Indian Words",
"sec_num": "6"
},
{
"text": "The initial steps use simpler methods to generate transliteration candidates on the source as well as the target side. They also use no resources on the target (IL) side. The step of fuzzy string matching compensates for the lack of more language specific knowledge during the earlier phase. The transliteration candidates are matched with the words in the target language corpus (actually, words in the word list extracted from the corpus). The fuzzy string Figure 1 : Mappings for foreign words. The three columns are for Roman, Devanagari and Telugu matching algorithm we use is finely tuned for Indian Languages and performs much better than language independent approaches like edit distance (Singh et al., 2007) . This method can be used for all the languages which use Abugida scripts, e.g. Hindi, Bengali, Telugu, Amharic, Thai etc. It uses characteristics of a writing system for fuzzy search and is able to take care of spelling variation, which is very common in these languages. This method shows an improvement in F-measure of up to 30% over scaled edit distance.",
"cite_spans": [
{
"start": 697,
"end": 717,
"text": "(Singh et al., 2007)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [
{
"start": 459,
"end": 467,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Fuzzy String Matching",
"sec_num": "7"
},
{
"text": "The method for fuzzy string matching is based on the Computational Phonetic Model of Scripts or CPMS (Singh, 2006b) , which models scripts (specifically Indic scripts) in terms of phonetic (articulatory) and orthographic features. For calculating the distance between two letters it uses a Stepped Distance Function (SDF). Each letter is represented as a vector of features. Then, to calculate the distance between two strings, it uses an adapted version of the Dynamic Time Warping algorithm (My- ers, 1980) . In the fuzzy string matching method that we use (Singh et al., 2007) , an akshar (roughly a syllable) is used as the unit, instead of a letter.",
"cite_spans": [
{
"start": 101,
"end": 115,
"text": "(Singh, 2006b)",
"ref_id": "BIBREF14"
},
{
"start": 498,
"end": 508,
"text": "ers, 1980)",
"ref_id": null
},
{
"start": 559,
"end": 579,
"text": "(Singh et al., 2007)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Fuzzy String Matching",
"sec_num": "7"
},
{
"text": "We use the above mentioned steps to transliterate a given word based on its origin. In case of ambiguity of word origin both methods are used, and possible transliterations are ranked. Based on the class of the word, the possible pronunciations (for foreign words) and the possible segmentations (for Indian words) are generated. Then, for foreign words, English phonemes are mapped to IL segments. For Indian words, Latin segments are mapped to IL segments. Now, the transliteration candidates are matched with target language words, using the fuzzy text search method (Singh et al., 2007) . Possible transliterations are ranked based on three parameters: word frequency, text search cost and the probability of the word belonging to the class through which it Figure-3 . The ranks are obtained on the basis of a score which is calculated using the following formula:",
"cite_spans": [
{
"start": 570,
"end": 590,
"text": "(Singh et al., 2007)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [
{
"start": 762,
"end": 770,
"text": "Figure-3",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Discerning Adaptable Transliteration Mechanism (DATM)",
"sec_num": "8"
},
{
"text": "T t = log(f t ) * p(C | s) cost(c, t) + K (1)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discerning Adaptable Transliteration Mechanism (DATM)",
"sec_num": "8"
},
{
"text": "where T t is the transliteration score for the target word t, f t is the frequency of t in the target language corpus, C is the word class (foreign or Indian), s is the source word, c is a transliteration candidate which has been generated depending on the predicted class C, p(C|s) is the probability of the class C given s, cost(c, t) is the cost of fuzzy string matching between c and t, and finally K is a constant which determines how much weight is given to the cost of fuzzy string matching.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discerning Adaptable Transliteration Mechanism (DATM)",
"sec_num": "8"
},
{
"text": "We evaluate our method for two major languages of India: Hindi and Telugu. We compare our results with a very commonly used method (Oh and Choi, 2006) based on bilingual dictionary to learn translit- ",
"cite_spans": [
{
"start": 131,
"end": 150,
"text": "(Oh and Choi, 2006)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation",
"sec_num": "9"
},
{
"text": "We created 2000-word lists which consisted of both foreign and Indian words written in Latin script and their transliterations in Hindi and Telugu. This dictionary was created by people with professional knowledge in both English and the respective Indian language. We only use this list for training the baseline method, as our method does not need training data on the target side. The size of bilingual word lists that we are using is less than those used for experiments by some other researchers. But our approach focuses on developing transliterations for languages with resource scarcity. This setup is more meaningful for languages with scarce resources. Since, normal transliteration mechanisms do not consider word origin, we train the baseline using the set of 2000 words containing both foreign and Indian words. Alignments from English to respective Indian languages were learned by aligning these lists using GIZA++. The alignments obtained were fed into a maximum entropy classifier with a context window size of 2 (3 is generally considered better window size, but because the training size is not huge, a context window of 3 gave substantially worse results). This method is similar to the grapheme based model as described by Oh and Choi (Oh and Choi, 2006) . However, unlike in their approach, the candidate pairs are matched with words in the target language and are ranked based on edit distance (BL).",
"cite_spans": [
{
"start": 1244,
"end": 1250,
"text": "Oh and",
"ref_id": null
},
{
"start": 1256,
"end": 1275,
"text": "(Oh and Choi, 2006)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Setup",
"sec_num": "9.1"
},
{
"text": "For our method (DATM), we have used CMU dictionary and a collection of Indian named entities (written in Latin) extracted from web to train the language identification module. We have considered n-grams of order 5 and pruned them by 3500 frequency. In case the foreign word is not found in CMU Speech dictionary, we guess its pronunciation using the method described by Oh and Choi. However, in this case, the context window size is 3.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Setup",
"sec_num": "9.1"
},
{
"text": "We also use another method (DBL) to check the validity of our assumptions about word origin. We use the same technique as BL, but in this case we train two models of 1000 words each, foreign and Indian. To disambiguate which model to use, we use the same language identification method as in DATM.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Setup",
"sec_num": "9.1"
},
{
"text": "To evaluate our method we have created word lists of size 200 which were doubly checked by two individuals. These also contain both Indian and Foreign words. We use both precision and mean reciprocal rank (MRR) to evaluate our method against baseline (BL) and discerning baseline (DBL). MRR is a measure commonly used in information retrieval when there is precisely one correct answer (Kandor and Vorhees, 2000) . Results can be seen in Table- 2. The highest scores were obtained for Hindi using DATM. The MRR in this case was 0.87.",
"cite_spans": [
{
"start": 386,
"end": 412,
"text": "(Kandor and Vorhees, 2000)",
"ref_id": null
}
],
"ref_spans": [
{
"start": 438,
"end": 444,
"text": "Table-",
"ref_id": null
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "9.2"
},
{
"text": "One important fact that comes out from the results is that determining the class of a word and then using an appropriate method can lead to significant increase in performance. This is clear from the results for BL and DBL. The only difference between these two was that two different models were trained for the two classes. Then the class of the word was identified (in DBL) and the model trained for that class was used for transliteration.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "9.2"
},
{
"text": "It should be noted that Yoon et al. (Yoon et al., 2007) have also reported MRR score on Hindi. They have used a number of phonetic and pseudo features, and trained their algorithm on a winnow classifier. They tested their algorithm only for named entities. They have considered a relatively limited number of candidate words on the target language side (1,500) which leads to 150k pairs on which they have evaluated their method. They have reported the results as 0.91 and 0.89 under different test conditions. In case of our evaluation, we do not restrict the candidate words on the target side except that it should be available in the corpus. Because of this formulation, there are over 1000k words for Hindi and over 1800k words from Telugu. This leads to a extremely high number of pairs possible. But such an approach is also necessary as we want our algorithm to be scalable to bigger sizes and also because there are no high quality tools (like named entity recognizers) for Indian languages. This is one of the reason for relatively (compared to figures reported by other researchers) low baseline scores. Despite all these issues, our simpler approach yields similar results. Figure-4 shows how the number of correct words varies with the rank.",
"cite_spans": [
{
"start": 36,
"end": 55,
"text": "(Yoon et al., 2007)",
"ref_id": "BIBREF17"
}
],
"ref_spans": [
{
"start": 1186,
"end": 1194,
"text": "Figure-4",
"ref_id": "FIGREF2"
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "9.2"
},
{
"text": "Two possible issues are the out of vocabulary (OOV) words and misspelled or foreign words in the IL corpus. The OOV words are not handled right now by our method, but we plan to extend our method to at least partially take care of such words. The second issue is mostly resolved by our use of fuzzy string matching, although there is scope for improvement.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results",
"sec_num": "9.2"
},
{
"text": "We presented a more general and adaptable method for transliteration which is especially suitable for Indian languages. This method first identifies the class (foreign or Indian) of the word on the source side. Based on the class, one of the two methods is used for transliteration. Easily creatable mapping tables and a fuzzy string matching algorithm are then used to get the target word. Our evaluations shows that the method performs substantially better than the two baselines we tested against. The results are better in terms of both MRR (up to 0.44) and precision (45%). Our method is designed to be used for other applications like tolerant input methods for Indian languages and it uses no resources on the target languages side except an unannotated corpus. The results can be further improved if we consider context information too.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions and Further Work",
"sec_num": "10"
},
{
"text": "We have also shown that disambiguating word origin and applying an appropriate method could be critical in getting good transliterations. Currently we are assuming that the word to be transliterated is in the target language corpus. We plan to extend the method so that even those words can be transliterated which are not in the target language corpus. We are also working on using this method for building a tolerant input method for Indian languages and on integrating the transliteration mechanism as well as the input method with an open source NLP friendly editor called Sanchay Editor (Singh, 2008) .",
"cite_spans": [
{
"start": 592,
"end": 605,
"text": "(Singh, 2008)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions and Further Work",
"sec_num": "10"
},
{
"text": "www.google.co.in/press/pressrel/news transliteration.html 2 www.quillpad.com",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Statistical transliteration for english-arabic cross language information retrieval",
"authors": [
{
"first": "N",
"middle": [],
"last": "Abduljaleel",
"suffix": ""
},
{
"first": "L",
"middle": [
"S"
],
"last": "Larkey",
"suffix": ""
}
],
"year": 2003,
"venue": "Proceedings of the twelfth international conference on Information and knowledge management",
"volume": "",
"issue": "",
"pages": "139--146",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "N. AbdulJaleel and L.S. Larkey. 2003. Statistical transliteration for english-arabic cross language infor- mation retrieval. Proceedings of the twelfth interna- tional conference on Information and knowledge man- agement, pages 139-146.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "A hybrid approach to align sentences and words in English-Hindi parallel corpora",
"authors": [
{
"first": "N",
"middle": [],
"last": "Aswani",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Gaizauskas",
"suffix": ""
}
],
"year": 2005,
"venue": "Proceedings of the ACL Workshop on\" Building and Exploiting Parallel Texts",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "N. Aswani and R. Gaizauskas. 2005. A hybrid approach to align sentences and words in English-Hindi paral- lel corpora. Proceedings of the ACL Workshop on\" Building and Exploiting Parallel Texts.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Free resources and advanced alignment for cross-language text retrieval",
"authors": [
{
"first": "M",
"middle": [
"W"
],
"last": "Davis",
"suffix": ""
},
{
"first": "W",
"middle": [
"C"
],
"last": "Ogden",
"suffix": ""
}
],
"year": 1998,
"venue": "Proceedings of the 6th Text Retrieval Conference (TREC-6)",
"volume": "",
"issue": "",
"pages": "385--402",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "M.W. Davis and W.C. Ogden. 1998. Free resources and advanced alignment for cross-language text re- trieval. Proceedings of the 6th Text Retrieval Confer- ence (TREC-6), pages 385-402.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "OM: One Tool for Many",
"authors": [
{
"first": "M",
"middle": [],
"last": "Ganapathiraju",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Balakrishnan",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Balakrishnan",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Reddy",
"suffix": ""
}
],
"year": 2005,
"venue": "dian) Languages. ICUDL: International Conference on Universal Digital Library",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "M. Ganapathiraju, M. Balakrishnan, N. Balakrishnan, and R. Reddy. 2005. OM: One Tool for Many (In- dian) Languages. ICUDL: International Conference on Universal Digital Library, Hangzhou.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "What's in a Name? Proper Names in Arabic Cross-Language Information Retrieval",
"authors": [
{
"first": "L",
"middle": [],
"last": "Larkey",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Abduljaleel",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Connell",
"suffix": ""
}
],
"year": 2003,
"venue": "CIIR Technical Report",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "L. Larkey, N. AbdulJaleel, and M. Connell. 2003. What's in a Name? Proper Names in Arabic Cross- Language Information Retrieval. Technical report, CIIR Technical Report, IR-278.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Knowledge of language origin improves pronunciation of proper names. Proceedings of EuroSpeech-01",
"authors": [
{
"first": "A",
"middle": [],
"last": "Llitjos",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Black",
"suffix": ""
}
],
"year": 2001,
"venue": "",
"volume": "",
"issue": "",
"pages": "1919--1922",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "A. Llitjos and A. Black. 2001. Knowledge of language origin improves pronunciation of proper names. Pro- ceedings of EuroSpeech-01, pages 1919-1922.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Punjabi Machine Transliteration",
"authors": [
{
"first": "M",
"middle": [
"G A"
],
"last": "Malik",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of the 21st International Conference on Computational Linguistics and the 44th annual meeting of the ACL",
"volume": "",
"issue": "",
"pages": "1137--1144",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "M.G.A. Malik. 2006. Punjabi Machine Transliteration. Proceedings of the 21st International Conference on Computational Linguistics and the 44th annual meet- ing of the ACL, pages 1137-1144.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Surprise! What's in a Cebuano or Hindi Name?",
"authors": [
{
"first": "J",
"middle": [],
"last": "May",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Brunstein",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Natarajan",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Weischedel",
"suffix": ""
}
],
"year": 2004,
"venue": "ACM Transactions on Asian Language Information Processing (TALIP)",
"volume": "2",
"issue": "3",
"pages": "169--180",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J. May, A. Brunstein, P. Natarajan, and R. Weischedel. 2004. Surprise! What's in a Cebuano or Hindi Name? ACM Transactions on Asian Language Information Processing (TALIP), 2(3):169-180.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "A Comparative Performance Study of Several Dynamic Time Warping Algorithms for Speech Recognition",
"authors": [
{
"first": "C",
"middle": [
"S"
],
"last": "Myers",
"suffix": ""
}
],
"year": 1980,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "C. S. Myers. 1980. A Comparative Performance Study of Several Dynamic Time Warping Algorithms for Speech Recognition. Ph.D. thesis, M.I.T., Cambridge, MA, Feb. http://gate.ac.uk.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "An English-Korean transliteration model using pronunciation and contextual rules",
"authors": [
{
"first": "J",
"middle": [
"H"
],
"last": "Oh",
"suffix": ""
},
{
"first": "K",
"middle": [
"S"
],
"last": "Choi",
"suffix": ""
}
],
"year": 2002,
"venue": "Proceedings of the 19th international conference on Computational linguistics",
"volume": "1",
"issue": "",
"pages": "1--7",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J.H. Oh and K.S. Choi. 2002. An English-Korean transliteration model using pronunciation and contex- tual rules. Proceedings of the 19th international con- ference on Computational linguistics-Volume 1, pages 1-7.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "An ensemble of transliteration models for information retrieval",
"authors": [
{
"first": "J",
"middle": [
"H"
],
"last": "Oh",
"suffix": ""
},
{
"first": "K",
"middle": [
"S"
],
"last": "Choi",
"suffix": ""
}
],
"year": 2006,
"venue": "Information Processing and Management: an International Journal",
"volume": "42",
"issue": "4",
"pages": "980--1002",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J.H. Oh and K.S. Choi. 2006. An ensemble of translit- eration models for information retrieval. Information Processing and Management: an International Jour- nal, 42(4):980-1002.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "A Dynamic Text Input scheme for phonetic scripts like Devanagari. Proceedings of Development by Design (DYD)",
"authors": [
{
"first": "A",
"middle": [],
"last": "Rathod",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Joshi",
"suffix": ""
}
],
"year": 2002,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "A. Rathod and A. Joshi. 2002. A Dynamic Text Input scheme for phonetic scripts like Devanagari. Proceed- ings of Development by Design (DYD).",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "More accurate fuzzy text search for languages using abugida scripts",
"authors": [
{
"first": "Anil Kumar",
"middle": [],
"last": "Singh",
"suffix": ""
},
{
"first": "Harshit",
"middle": [],
"last": "Surana",
"suffix": ""
},
{
"first": "Karthik",
"middle": [],
"last": "Gali",
"suffix": ""
}
],
"year": 2007,
"venue": "Proceedings of ACM SI-GIR Workshop on Improving Web Retrieval for Non-English Queries",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Anil Kumar Singh, Harshit Surana, and Karthik Gali. 2007. More accurate fuzzy text search for languages using abugida scripts. In Proceedings of ACM SI- GIR Workshop on Improving Web Retrieval for Non- English Queries, Amsterdam, Netherlands.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Study of some distance measures for language and encoding identification",
"authors": [
{
"first": "Anil Kumar",
"middle": [],
"last": "Singh",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of ACL 2006 Workshop on Linguistic Distance",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Anil Kumar Singh. 2006a. Study of some distance mea- sures for language and encoding identification. In Pro- ceedings of ACL 2006 Workshop on Linguistic Dis- tance, Sydney, Australia.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "A computational phonetic model for indian language scripts",
"authors": [
{
"first": "Anil Kumar",
"middle": [],
"last": "Singh",
"suffix": ""
}
],
"year": 2006,
"venue": "Constraints on Spelling Changes: Fifth International Workshop on Writing Systems",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Anil Kumar Singh. 2006b. A computational phonetic model for indian language scripts. In Constraints on Spelling Changes: Fifth International Workshop on Writing Systems, Nijmegen, The Netherlands.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "A mechanism to provide language-encoding support and an nlp friendly editor",
"authors": [
{
"first": "Anil Kumar",
"middle": [],
"last": "Singh",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of the Third International Joint Conference on Natural Language Processing",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Anil Kumar Singh. 2008. A mechanism to provide language-encoding support and an nlp friendly editor. In Proceedings of the Third International Joint Con- ference on Natural Language Processing, Hyderabad, India.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Dealing with unknowns in machine translation. Systems, Man, and Cybernetics",
"authors": [
{
"first": "",
"middle": [],
"last": "Rmk Sinha",
"suffix": ""
}
],
"year": 2001,
"venue": "IEEE International Conference on",
"volume": "2",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "RMK Sinha. 2001. Dealing with unknowns in machine translation. Systems, Man, and Cybernetics, 2001 IEEE International Conference on, 2.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Multilingual Transliteration Using Feature based Phonetic Method",
"authors": [
{
"first": "S",
"middle": [
"Y"
],
"last": "Yoon",
"suffix": ""
},
{
"first": "K",
"middle": [
"Y"
],
"last": "Kim",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Sproat",
"suffix": ""
}
],
"year": 2007,
"venue": "Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics",
"volume": "",
"issue": "",
"pages": "112--119",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "S.Y. Yoon, K.Y. Kim, and R. Sproat. 2007. Multilingual Transliteration Using Feature based Phonetic Method. Proceedings of the 45th Annual Meeting of the Associ- ation of Computational Linguistics, pages 112-119.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"type_str": "figure",
"text": "Mappings for Indian Words",
"num": null,
"uris": null
},
"FIGREF1": {
"type_str": "figure",
"text": "Block Diagram of the Discerning Adaptive Transliteration Method (DATM) is transliterated. A block diagram describing the method is shown in",
"num": null,
"uris": null
},
"FIGREF2": {
"type_str": "figure",
"text": "Number of Correct Words vs. Rank. A significantly higher percentage of correct words occur at rank 1 for the DATM method, as compared to BL and DBL methods. This percentage indicates a more practical view of the accuracy transliteration algorithm.",
"num": null,
"uris": null
},
"TABREF1": {
"text": "",
"html": null,
"num": null,
"type_str": "table",
"content": "<table><tr><td>: Variations of a Hindi Word nOkarI (job).</td></tr><tr><td>The numbers are pages returned when searching on</td></tr><tr><td>Google.</td></tr></table>"
},
"TABREF4": {
"text": "",
"html": null,
"num": null,
"type_str": "table",
"content": "<table><tr><td>: Evaluation on English-Hindi and English-Telugu</td></tr><tr><td>erations. As there are no bilingual transliteration</td></tr><tr><td>dictionaries available for ILs, we had to create our</td></tr><tr><td>own resources.</td></tr></table>"
}
}
}
} |