File size: 69,885 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 |
{
"paper_id": "I11-1013",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:31:23.560690Z"
},
"title": "Handling verb phrase morphology in highly inflected Indian languages for Machine Translation",
"authors": [
{
"first": "Ankur",
"middle": [],
"last": "Gandhe",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "IBM Research",
"location": {
"country": "India"
}
},
"email": ""
},
{
"first": "Rashmi",
"middle": [],
"last": "Gangadharaiah",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "IBM Research",
"location": {
"country": "India"
}
},
"email": "rashgang@in.ibm.com"
},
{
"first": "Karthik",
"middle": [],
"last": "Visweswariah",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "IBM Research",
"location": {
"country": "India"
}
},
"email": ""
},
{
"first": "Ananthakrishnan",
"middle": [],
"last": "Ramanathan",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "IBM Research",
"location": {
"country": "India"
}
},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "The phrase based systems for machine translation are limited by the phrases that they see during the training. For highly inflected languages, it is uncommon to see all the forms of a word in the parallel corpora used during training. This problem is amplified for verbs in highly inflected languages where the correct form of the word depends on factors like gender, number and tense aspect. We propose a solution to augment the phrase table with all possible forms of a verb for improving the overall accuracy of the MT system. Our system makes use of simple stemmers and easily available monolingual data to generate new phrase table entries that cover the different variations seen for a verb. We report significant gains in BLEU for English to Hindi translation.",
"pdf_parse": {
"paper_id": "I11-1013",
"_pdf_hash": "",
"abstract": [
{
"text": "The phrase based systems for machine translation are limited by the phrases that they see during the training. For highly inflected languages, it is uncommon to see all the forms of a word in the parallel corpora used during training. This problem is amplified for verbs in highly inflected languages where the correct form of the word depends on factors like gender, number and tense aspect. We propose a solution to augment the phrase table with all possible forms of a verb for improving the overall accuracy of the MT system. Our system makes use of simple stemmers and easily available monolingual data to generate new phrase table entries that cover the different variations seen for a verb. We report significant gains in BLEU for English to Hindi translation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Data driven approaches have become widely popular as they use little or no language specific knowledge. The main drawback of these approaches is the need for large amounts of data. (Koehn et al., 2003) have shown that the quality of the translations produced by data driven approaches mainly depends on the amount of parallel data available for the language-pair under consideration. Creation of a large bilingual corpus is expensive and time consuming if high quality manual translations are required. Hence, building MT systems for language-pairs with limited amounts of data is a big challenge.",
"cite_spans": [
{
"start": 181,
"end": 201,
"text": "(Koehn et al., 2003)",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Approaches have been suggested in the past to mine the world-wide-web to automatically obtain large amounts of parallel data. For example, news articles in two different languages describing the same event can be sentence-aligned to obtain a parallel corpus. Although this approach has shown improvements, this cannot be extended to languages that have little or no data on the world wide web.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The situation gets worse for languages that are rich in morphology. Clearly large amounts of parallel data are required to observe all variations of a word. Popovic and Ney (2004) applied transformations to verbs to reduce the number of outof-vocabulary words and showed improvements in translation quality when morphemes were considered. Yang and Kirchhoff (2006) used a back off model in a Phrase-based SMT system which translated word forms in the source language by hierarchical morphological abstractions. Unknown words in the test data were stemmed and phrasetable entries were modified such that words sharing the same root were replaced by their stems. Freeman et al. (2006) and Habash (2008) find in-vocabulary words for OOV words that could be morphological variants of the OOV words. Phrases in the phrase table containing these invocabulary words are then replaced by OOV words to create new entries. Vilar et al. (2007) used a letter-based MT system that treated the source and target sentences as a string of letters for translating unknown words.",
"cite_spans": [
{
"start": 157,
"end": 179,
"text": "Popovic and Ney (2004)",
"ref_id": "BIBREF5"
},
{
"start": 339,
"end": 364,
"text": "Yang and Kirchhoff (2006)",
"ref_id": "BIBREF6"
},
{
"start": 661,
"end": 682,
"text": "Freeman et al. (2006)",
"ref_id": "BIBREF3"
},
{
"start": 687,
"end": 700,
"text": "Habash (2008)",
"ref_id": "BIBREF4"
},
{
"start": 913,
"end": 932,
"text": "Vilar et al. (2007)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "All the above approaches handled OOV issues that arise when the source language is morphologically rich. Generation of the target sentence when the target language is morphologically rich from a source language that is not rich in morphology is non-trivial as the source language does not contain all the information for inflecting the target words. Minkov et. al (2007) predicted inflected forms of a sequence of word stems on languages that are morphologically rich using syntactic and rich morphological sources. This inflection generation model was then applied in MT by (Toutanova et al., 2008) while translating English into morphologically complex languages and showed improve-ment in translation quality. Their methods require a syntactic analyzer and a very rich morphological analyzer which may not be available for many rare or low-density languages. Also, their feature set includes bilingual features that require expensive and difficult to get bilingual corpora. We rely more on monolingual data and a small amount of parallel data. In cases of multi word compound words ( explained in section 1.1 ) , since inflections on the light verb might change with change in the root verb compounding with it, we need to predict these verbs together and not as separate words.",
"cite_spans": [
{
"start": 350,
"end": 370,
"text": "Minkov et. al (2007)",
"ref_id": "BIBREF0"
},
{
"start": 575,
"end": 599,
"text": "(Toutanova et al., 2008)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In this paper, we consider Indian languages which are considered as low density languages as they do not have rich knowledge sources such as parsers or complex morphological analyzers. These languages also suffer from data sparsity and hence form ideal languages for the analysis of our proposed method. We also consider only various forms of verbs and do not consider other words such as noun phrases and adjectives affected by inflections.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "India has fifteen official languages which originated from the Indo-Iranian branch of the Indo-European language family, the non-Indo-European Dravidian family, Austro-Asiatic, Tai-Kadai and the Sino-Tibetan language families (Microsoft Encarta Online Encyclopedia, 1997) . The languages that stem from the Dravidian family, are -Tamil, Kannada, Malayalam and Telugu, spoken in the South Indian states. Languages in North India, such as Hindi, Urdu, Punjabi, Gujarati, Bengali, Marathi, Kashmir, Sindhi, Konkani, Rajasthani, Assamese and Oriya, stem from Sanskrit and Pali. Indian languages are verb final i.e., verbs are placed at the end of the sentences. Verbs in these languages are inflected to contain information about gender (masculine and feminine), tense, aspect and number of the subject (singular or plural). A few examples showing inflections on the verbs in Hindi are shown below:",
"cite_spans": [
{
"start": 226,
"end": 271,
"text": "(Microsoft Encarta Online Encyclopedia, 1997)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Background on Indian Languages",
"sec_num": "1.1"
},
{
"text": "These languages also contain compound verbs (multi-word compound representing a single verb). They contain a light verb which receives inflections and another component that can be a noun or a verb responsible for conveying the meaning. For example, in Hindi, most commonly used light verbs are \"karna\" (to make), \"lena\" (to take), \"hona\" (to happen) and \"dena\" (to give).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Background on Indian Languages",
"sec_num": "1.1"
},
{
"text": "When translating from a morphologically poor language such as English to any of the Indian languages, finding the right translation along with the inflections on the verbs becomes difficult, especially when the amount of bilingual data available is scarce. We try to use the pattern behavior of verbs to tackle this problem. Table 1 gives an example of hindi verbs classified according to their light verbs. Hindi side is transliterated for the sake of clarity. It shows how the verb phrase of one compound verb (clean) can generate verb phrase for words in the same group (help and forgive) just by replacing the corresponding source and target root words. The suffixes (shown in bold) are separated from the word to show how the actual process takes place. This paper tries to automatically group the different kinds of verbs occurring in the language based on their light verbs and generates the variation for all the verbs in one group by looking at the variations of any one member. he will be clean ing -> vo saaph kar egaa he will be forgive ing -> vo maaph kar egaa he will be help ing -> vo madad kar egaa The novel concept of this paper is generation of verb phrases on the source side and their translations using a) source and target monolingual data, b) simple morphological segmentation on source and target side and c) Small amount of manual translations or d) Word alignments of parallel corpora. We have described two methods of generating these verb phrase translations in the following sections.",
"cite_spans": [],
"ref_spans": [
{
"start": 325,
"end": 332,
"text": "Table 1",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Motivation",
"sec_num": "2"
},
{
"text": "The idea here is to get the manual translations of source and target verb pairs which could capture the entire range of variations seen in the source and the target verb phrases. These translations can then be used to generate the variations for rest of the verb pairs. The entire flow of the method is shown in Figure 1 .",
"cite_spans": [],
"ref_spans": [
{
"start": 312,
"end": 320,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Manual Generation",
"sec_num": "3"
},
{
"text": "Given a language pair (e,f), we extract all verb phrases that occur in the source monolingual data using a verb phrase chunker. Part of speech (POS) tags can be used to extract verb phrases for languages having a good POS tagger. In our experiments for English-Hindi language pair, POS tags were used for English verb phrase chunking. Modals in English were included as a part of the verb phrase since their counterparts in Hindi appear as verbs. For Hindi, the verb phrase chunker was trained on a small set of 6000 sentences, where the reference markings were obtained by projecting the verb phrases from English. The 6000 sentences were hand aligned with the corresponding English sentences, hence helping with the accuracy of the projected verb phrases. On this data, we built a CRF based chunker (Lafferty et al., 2001 ) using word and POS tag features.",
"cite_spans": [
{
"start": 801,
"end": 823,
"text": "(Lafferty et al., 2001",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Verb Phrase Chunking",
"sec_num": "3.1"
},
{
"text": "Using a segmenter, the root verb is separated from its inflected suffix for all the extracted verb phrases. These extracted verb phrases are then clustered based on the root verb so that all the variations of a root verb '<verb>' are grouped together into one cluster. As an example, a part of the verb cluster for 'play' is shown below. Note that all possible variations of each verb (both source and target side) are under one cluster. play was play +ing should have play +ed ought be played would have been play +ed is play +ed cannot be play +ed is being play +ed The different variations within each verb cluster are normalized by replacing the root verb by a normalization tag '<verb>' so that similar root verb (Minnen et al., 2000) an open source stemming package, to get the root form of the head verb in the extracted English verb phrases. It was observed that all the root verb clusters had the same variations of verb phrases and hence all belonged to the same class. Thus, from the source (English) side, only one verb could be picked up to cover all the variations, which could be then replicated for all the others verbs. We call that class 'AE'.",
"cite_spans": [
{
"start": 718,
"end": 739,
"text": "(Minnen et al., 2000)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Verb Classing",
"sec_num": "3.2"
},
{
"text": "In Hindi, as explained in section 1.1, many verbs occur as compound verbs where a noun followed by a light verbs is considered as a verb and hence we also included these for our clustering and classification. The extracted verb phrases were segmented using a stemmer similar to one in (Ramanathan et al., 2003) . After clustering the verb phrases based on their root verbs into groups and classing them based on the different variations, the main classes depended on the a) whether the verb has a light verb or not and b) the type of light verb attached. Table 2 shows the different classes found along with the number of verbs within each class.",
"cite_spans": [
{
"start": 285,
"end": 310,
"text": "(Ramanathan et al., 2003)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [
{
"start": 555,
"end": 562,
"text": "Table 2",
"ref_id": "TABREF3"
}
],
"eq_spans": [],
"section": "Verb Classing",
"sec_num": "3.2"
},
{
"text": "There were more classes with a different auxiliary verb but we neglected them since the frequency of verbs in those classes was insignificant. 'lena' and 'dena' verb forms take the same vari-ations (differing only in one character), we could easily generate one from another. Overall, only 3 classes were used for manual translation on the Hindi side(AH,BH and CH). Note that we allow a verb to belong to more than one class, suggesting that a word can be used in more than just one way depending on the context.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Verb Classing",
"sec_num": "3.2"
},
{
"text": "Given a parallel corpus, it is possible to extract source root verb to target root verb translation. Since a parallel corpus will contain the inflected form of a verb, it is necessary to stem them to their root form before calculating the word translation probabilities. Hence, given a parallel corpus, sentences are machine aligned by a maxent model as described in (Ittycheriah and Roukos, 2005) and then the verbs on both the source and target side are stemmed to the respective root forms using a suitable stemmer. Given the list of possible source verbs and target words from the previous clustering step, the forward and reverse translation probabilities for these verbs is calculated from the alignments using by relative frequency:",
"cite_spans": [
{
"start": 367,
"end": 397,
"text": "(Ittycheriah and Roukos, 2005)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Root verb translation pairs",
"sec_num": "3.3"
},
{
"text": "P (f i /e j ) = count(f i , e j )/ f count(f, e j ) (1) P (e i /f j ) = count(e i , f j )/ e count(e, f j ) (2)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Root verb translation pairs",
"sec_num": "3.3"
},
{
"text": "Using these forward and reverse probability, a mapping file that maps the source root verb to the corresponding target root verb(s) is created by empirically combining the two probabilities. P tot (e i /f j ) = 0.5 * P (e i /f j )+0.5 * P (f j /e i ) (3)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Root verb translation pairs",
"sec_num": "3.3"
},
{
"text": "We allow one source verb mapping to multiple target verb, since the meaning can change due to context in the test sentence. TopM translations of source word are selected for translations, provided P tot > P thresh . We empirically found the P thresh = 0.2 and M=4 to work reasonably well. Two or more worded root verb, such as phrasal verbs 'take off', 'figure out', were not considered while creating the mapping since the meaning is often different that the individual words and the generation of verb phrases from these root verbs is more tricky. Such constructions, where one word verb may translate to multiple words, occurred for only 3% of the verbs in the test data and hence could be ignored without any significant loss in improvement. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Root verb translation pairs",
"sec_num": "3.3"
},
{
"text": "Given the root verb mapping and the classes to which these source and target root verb belong to, we create a 'source class' to 'target class' mapping, or a 'verb-pair class', by replacing the root verbs with their corresponding verb classes. This causes each of the verb pair to fall under some particular verb-pair class. If there are n classes in the source side and m on the target side, the maximum number of verb-pair classes are N = m*n By picking any one root verb pair from each of these verb-pair classes, we can cover all the possible variations of verb phrase translation pairs. These pairs can then be given for human translation, by creating all possible variations of either the source side or target side and asking humans to translate to the other.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Generation of Verb Phrase Dictionary",
"sec_num": "3.4"
},
{
"text": "Templates for each of the N verb-pair class are created from the manually translated data by segmenting the verb phrase pairs on both sides and replacing the root verb by the '<verb>' tag. An example of such translation pair for English-Hindi is shown was <verb> +ing == <verb> raha tha [Class AH-AE ] was <verb> +ing == <verb> kar raha tha [Class BH-AE ] Picking root verb pairs from each verb-pair class and replacing the <verb< tag with corresponding verbs, these templates are used to create new verb phrases which may not be present in the parallel data to a large extent. A reverse morphological tool or joiner is used to recombine the segmented verb phrases and create a verb phrase dictionary.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Generation of Verb Phrase Dictionary",
"sec_num": "3.4"
},
{
"text": "In our paper, English had one class and Hindi had 3. Thus, only 3 Hindi-English verb pairs needed to be translated, one from each of the verbpair classes AH-AE, BH-AE, and CH-AE. We created different variations of the English verbs, since it had only one class and could be easily generated using manually built rules. Grammar rules contain number (singular/plural), tense and aspect agreement between different auxiliary forms (for example: was, is, were, can, might, could not, wouldn't, etc) and verbs (for example: answer, punishing, cleaned, etc.). A unique Hindi-English verb pair is picked from each of the verb-pair classes obtained earlier and their English verbs are used in generating the English verb forms. For example, \"saaf\" belonged to the \"karna\" cluster, so its English translation, \"clean\" is used for creating verb forms. Gender information is also added to the Verb forms which will be required for the Hindi counterparts. About \u2248 970 verb forms were generated for each of the 3 verbs. Examples of a few Verb forms are given below: 'Not' is included as a part of the extracted verb phrases since it is the most common adverb that occurs within the verb phrases. Other adverbs such as 'now', 'also' have not been dealt with in this paper. These variations, along with the mapping of the English-Hindi root verb was given to the annotators for translation. The subject information within '[]' helps the annotators to decide on the number and gender inflections on the target (Hindi) side. These are removed before using in the machine translation system. The reverse morphology of the generated verb phrases is done using MorphG (Minnen et al., 2000) for English and a simple suffix joiner for hindi.",
"cite_spans": [
{
"start": 1648,
"end": 1669,
"text": "(Minnen et al., 2000)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Generation of Verb Phrase Dictionary",
"sec_num": "3.4"
},
{
"text": "[",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Generation of Verb Phrase Dictionary",
"sec_num": "3.4"
},
{
"text": "Although manual translation is a clean and effective way of generating these verb phrases, a human is still required in the loop to complete the setup. Instead, both side monolingual data can be employed to extract all the variations for each root verb, parallel corpus can be used to get the sourceverb to root-verb pairs, and finally a model can be learnt to align the source verb phrase to target verb phrase using the verb alignments from the hand alignments and machine alignments.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Automatic generation",
"sec_num": "4"
},
{
"text": "Using the technique described in section 3.3, a source to target root verb mapping are obtained.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Verb pairs",
"sec_num": "4.1"
},
{
"text": "Clustering of verb phrases on source and target side is done as explained in section 3.2 so that each cluster contains different variations of the same root verb form. The phrases within each cluster are segmented on both sides using the techniques described section 3.2 and are generalized by replacing the root verb in the segmented verb phrase by a '<verb>' tag as this will help while aligning the source verb phrase to its corresponding target verb phrase.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Clustering",
"sec_num": "4.2"
},
{
"text": "In order to learn a verb phrase alignment model, we need good quality verb phrase alignments from the parallel corpora. We concentrate on hand aligned data and accurate machine alignments. Machine aligned verb phrases that occurred less than three times were treated as inaccurate. The source side verb phrases are extracted using the scheme similar to one in section 3.1, and by looking at the target words they align to, verb phrase alignments are obtained. The aligned verb phrases are segmented on both the target and the source side using the strategy described in section 3.2 and then normalized by replacing the head word for both the source side verb phrase and the target side verb phrase by a '<verb>' tag. The '<verb>' tagged verb phrases act as templates for verb alignments. Since all the root forms of verb will not occur in the extracted verb alignments, it's necessary to normalize them to be able to learn a general model. This way, if the translation of a particular source verb phrase variation is known, its generalized form can be used to get the trans-lation of a different root verb for the same variation. This is similar to our claim in section 3.4 that translation of one root verb can generate translations for all other verbs belonging to the same class.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Verb Phrase Translations",
"sec_num": "4.3"
},
{
"text": "A simple word alignment model is used to learn the word translation probabilities. Please note that the suffixes of the verbs are also treated as words since they contain important information about tense, gender and number. We used GIZA++ model 4 to learn P(V si /V tj ), which is the probability of the i th source word/segment aligning to the j th target word/segment.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Verb Phrase Translations",
"sec_num": "4.3"
},
{
"text": "From the root-verb pairs obtained in section 4.1, each verb pair is picked and the best translation for a source verb phrase in source-side verb cluster is searched for in the target verb cluster. If a cluster for the source or target verb does not exist, that pair is ignored. Both the source and the target verb clusters contain the generalized verb phrase of the form '... aux \u22121 <verb> aux 1 aux 2 .. '. First, a perfect match of a source phrase and target verb phrase is searched in the hand aligned and machine aligned verb phrase pairs. If found, that phrase pair is treated as a valid verb phrase pair. If no perfect match is found, word alignment probabilities obtained in previous section are use to get the source to target verb phrase alignments. Any verb phrase alignment pair with score lower than a threshold score of 0.5 is ignored.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Automatic Alignment and Generation",
"sec_num": "4.4"
},
{
"text": "After obtaining all the valid verb phrase pairs, the tag <verb> is replaced by their corresponding root verbs and as in section 3.4 and the suffixes are joined to the root verb to get the automatically generated verb phrase dictionary which can be used in the MT system.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Automatic Alignment and Generation",
"sec_num": "4.4"
},
{
"text": "In this section, we report our experimental results on English -Hindi language pair. We first report on the coverage ratio, which gives an estimate of number of exact verb phrases covered by the baseline system and our method. In addition, we also report on English to Hindi Machine translation results for phrase based systems.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "5"
},
{
"text": "The data used for clustering and classification on source and target side, the parallel corpora and the test set details are shown in table 3. English Monolingual 6 million Hindi Monolingual 1.4 million Test set 1 4000 Test set 2 715 Training Data 280k Table 3 : Data used for experiments",
"cite_spans": [],
"ref_spans": [
{
"start": 143,
"end": 271,
"text": "English Monolingual 6 million Hindi Monolingual 1.4 million Test set 1 4000 Test set 2 715 Training Data 280k Table 3",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Discovery of new data",
"sec_num": "5.1"
},
{
"text": "The Hindi monolingual data was used to collect 4320 Hindi verb clusters belonging to 3 different classes ( section 3.2 ) and the English monolingual yielded 4872 clusters. Many of these clusters were false positives due to the bad quality of verb phrase chunker but were eliminated in the subsequent steps. The parallel data was aligned using a maxent model (Ittycheriah and Roukos, 2005) and gave us 2944 verb-pairs.",
"cite_spans": [
{
"start": 358,
"end": 388,
"text": "(Ittycheriah and Roukos, 2005)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "No of Sentences",
"sec_num": null
},
{
"text": "For manual generation method, 3 verb pairs were given to annotators for translation, with about 972 different English verb forms in each. The generated dictionary from the manual translations had 2.7 million verb phrases. The automatic method aligned the corresponding clusters of the 2944 verb-pairs and produced about 300k new verb phrases. The considerably lesser size of the automatically created verb phrase dictionary compared to the manual dictionary can be attributed to the fact that the manual dictionary contains variations that are not seen in our monolingual data. We claim that the manually and automatically created verb phrase dictionaries add new data to the system and have a higher chance of finding a matching source verb phrase in a given corpus than our phrase based system. We verify this claim by extracting verb phrases from two test sets and searching for them in:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "No of Sentences",
"sec_num": null
},
{
"text": "1. Baseline Phrase table 2. Base+Manual generated Dictionary 3. Base+Auto generated Dictionary Figure 2 shows the ratio of the number of verbs phrases found in the three cases to the total number of verbs searched. We call this as coverage ratio. The verb phrases are divided based on their lengths. The plot clearly shows that the coverage increases considerably by the addition of these generated verb phrases which may or may not be seen in the training data, especially as the length of the verb phrase increases. Verb phrases of length 1 are not shown since the coverage was almost same for the 3 settings. Table 3 shows the training data and the two test sets used for evaluation. The bilingual parallel data is split into training data and Test set 1. Test set 2 is a generic test set. All the data (training and test ) used is predominately contains news. We report are results on BLEU (Papineni et al., 2002) .",
"cite_spans": [
{
"start": 894,
"end": 917,
"text": "(Papineni et al., 2002)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [
{
"start": 95,
"end": 103,
"text": "Figure 2",
"ref_id": "FIGREF1"
},
{
"start": 612,
"end": 619,
"text": "Table 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "No of Sentences",
"sec_num": null
},
{
"text": "The verb phrase pairs were generated as explained in manual generation section and then added to the baseline system as a part of corpus(Base+manVPcorp). To emphasis on the improvement from generated verb phrases, an experiment where only the human translated verb phrases are added to the baseline corpus was also conducted(Base+humanVPcorp). Table 4 shows the results on Moses -a state of the art phrase based system, and on a phrase based system (PBMT) similar to (Tillman et al., 2006) on test sets 1 and 2. Both systems were trained on 280k parallel sentences. On the in-domain data, we had an improvement of 4.8 BLEU points for Moses and 0.9 for PBMT. On the more generic test set, Moses gave a BLEU score improvement of 1.1 whereas the PBMT performance was comparable. One reason for this difference in the BLEU score jumps is the better alignments in the PBMT system, aligned by a maxent model as described in (Ittycheriah and Roukos, 2005) . The PBMT system thus has a higher chances of having a good verb phrase in the baseline system than Moses and hence on adding generated verb phrases, we would see a lesser gain. Since the PBMT system had comparable results on the in-domain data with Moses and performed better on the out of domain (more generic) test set, the remaining experiments have been conducted with the PBMT system.",
"cite_spans": [
{
"start": 467,
"end": 489,
"text": "(Tillman et al., 2006)",
"ref_id": "BIBREF11"
},
{
"start": 918,
"end": 948,
"text": "(Ittycheriah and Roukos, 2005)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [
{
"start": 344,
"end": 351,
"text": "Table 4",
"ref_id": null
}
],
"eq_spans": [],
"section": "Machine translation Results",
"sec_num": "5.2"
},
{
"text": "PBMT Set 1 Set2 Set 1 Set2 Baseline 13.5 08.6 13.4 16.1 Base+humanVPcorp 14.1 08.6 14.0 16.1 Base+manVPcorp 18.3 9.7 14.3 16.0 Table 4 : BLEU score on test set 1 and 2 for different settings on moses and PBMT Adding the generated phrases as a parallel corpus can alter the translation probabilities of individual words and sub-phrases. This is one of the reasons for no improvement in the bleu score of the PBMT system when the generated verb phrases are added as corpus. A better method would be to add the verb phrases directly to the phrase table. We added the manual dictionary to the PBMT system and the results are tabulated in table 5.",
"cite_spans": [],
"ref_spans": [
{
"start": 127,
"end": 134,
"text": "Table 4",
"ref_id": null
}
],
"eq_spans": [],
"section": "Moses",
"sec_num": null
},
{
"text": "Baseline 13.4 16.1 Base+humanVP-PT 14.0 16.1 base+manVP-PT 14.9 16.5 base+autoVP-PT 14.8 16.3 Table 5 : BLEU score for PBMT system after adding verb phrases directly to Phrase Table (PT) Adding the verb phrases directly to the system keeps the rest of the phrases and their scores intact. Only phrases with matching source side phrase need to be re-normalized to adjust the translation probabilities. This would mean that the probability of only the verb phrases we add to the baseline phrase table would be affected while the rest of the translation model would be the same. Table 5 shows that addition of manually generated data to the phrase table(Base+manVP-PT), gives a good improvement of 1.5 points on the in-domain data and 0.4 on the out of domain data. A significant improvement of 1.4 BLEU points is seen even when the automatically generated verb phrases are added (Base+autoVP-PT), which was not seen when these were added as a corpus to the system. Figure 3 shows the variation of BLEU score with change in the corpus size. We should expect that the gain be higher in the case of low corpus size. However, note that the verb-pair list used to generate the verb phrases also changes with the change in corpus size, since decreasing the corpus size would decrease the quality of the overall alignments and the number of verbs seen. Thus, while the verb-pair list using 160k sentences had a total of 2499 verb pairs, the 20k corpus produced only 1347 verb pairs. So, for a smaller corpus size, the number of new verb phrases added to the table would also be lesser. This explains the rather constant gain in BLEU score throughout the graph. ",
"cite_spans": [],
"ref_spans": [
{
"start": 94,
"end": 101,
"text": "Table 5",
"ref_id": null
},
{
"start": 176,
"end": 186,
"text": "Table (PT)",
"ref_id": null
},
{
"start": 576,
"end": 583,
"text": "Table 5",
"ref_id": null
},
{
"start": 963,
"end": 971,
"text": "Figure 3",
"ref_id": "FIGREF2"
}
],
"eq_spans": [],
"section": "Set 1 Set2",
"sec_num": null
},
{
"text": "We showed an improvement of up to 1.5 bleu points on the in domain data set and an improvement of 0.4 bleu on the generic test set. However, there are still some errors in respect to the morphology of the verb phrases, which the Language Model is unable to tackle. These are primarily the long range dependencies which includes determining the gender and number of the subject or object to get the appropriate inflection. Having a dynamic feature based system, which does not require rich morphological resources, and predicts the suffixes and inflections would be able to solve this problem. Also, when a verb has more than one meaning, the contextual information is not captured efficiently in the current method and often produces a more literal translation than the the reference.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion and Future Work",
"sec_num": "6"
},
{
"text": "Apart from adding the verb phrases to the phrase table, filtering of poor verb phrase pairs from the original phrase table is another approach to consider. The two methods together can give a higher boost to the translation than just one of them. A more language independent method of extraction of verb phrases also needs to be constructed, which does not require building language dependent stemmers and verb phrase chunkers.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion and Future Work",
"sec_num": "6"
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Generating Complex Morphology for Machine Translation",
"authors": [
{
"first": "Einat",
"middle": [],
"last": "Minkov",
"suffix": ""
},
{
"first": "Kristina",
"middle": [],
"last": "Toutanova",
"suffix": ""
},
{
"first": "Hisami",
"middle": [],
"last": "Suzuki",
"suffix": ""
}
],
"year": 2007,
"venue": "Proc. 45th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "128--135",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Einat Minkov, Kristina Toutanova and Hisami Suzuki. 2007. Generating Complex Morphology for Ma- chine Translation, in Proc. 45th Annual Meeting of the Association for Computational Linguistics, 2007, pp. 128-135.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Applying Morphology Generation Models to Machine Translation",
"authors": [
{
"first": "Kristina",
"middle": [],
"last": "Toutanova",
"suffix": ""
},
{
"first": "Hisami",
"middle": [],
"last": "Suzuki",
"suffix": ""
},
{
"first": "Achim",
"middle": [],
"last": "Ruopp",
"suffix": ""
}
],
"year": 2008,
"venue": "Proc. 46th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kristina Toutanova, Hisami Suzuki and Achim Ruopp 2008. Applying Morphology Generation Models to Machine Translation, in Proc. 46th Annual Meet- ing of the Association for Computational Linguis- tics, 2008.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Can we translate letters?",
"authors": [
{
"first": "D",
"middle": [],
"last": "Vilar",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Peter",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Ney",
"suffix": ""
},
{
"first": "L",
"middle": [
"F"
],
"last": "Informatik",
"suffix": ""
}
],
"year": 2007,
"venue": "In Proceedings of Association Computational Linguistics Workshop on SMT",
"volume": "",
"issue": "",
"pages": "33--39",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "D. Vilar, J. Peter, H. Ney, and L. F. Informatik 2007. Can we translate letters?, In Proceedings of Associ- ation Computational Linguistics Workshop on SMT, pages 33-39, 2007.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Cross linguistic name matching in english and arabic: a \"one to many mapping\" extension of the levenshtein edit distance algorithm",
"authors": [
{
"first": "A",
"middle": [
"T"
],
"last": "Freeman",
"suffix": ""
},
{
"first": "S",
"middle": [
"L"
],
"last": "Condon",
"suffix": ""
},
{
"first": "C",
"middle": [
"M"
],
"last": "Ackerman",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of the main conference on Human Language Technology, Conference of the North American Chapter of the Association of Computational Linguistics",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "A. T. Freeman, S. L. Condon, and C. M. Ackerman 2006. Cross linguistic name matching in english and arabic: a \"one to many mapping\" extension of the levenshtein edit distance algorithm. , In Pro- ceedings of the main conference on Human Lan- guage Technology, Conference of the North Amer- ican Chapter of the Association of Computational Linguistics.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Four techniques for online handling of out-of-vocabulary words in arabic-english statistical machine translation",
"authors": [
{
"first": "N",
"middle": [],
"last": "Habash",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of Association for Computational Linguistics-08",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "N Habash 2008. Four techniques for online handling of out-of-vocabulary words in arabic-english statis- tical machine translation , In Proceedings of Asso- ciation for Computational Linguistics-08.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Towards the use of word stems and suffixes for statistical machine translation",
"authors": [
{
"first": "M",
"middle": [],
"last": "Popovic",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Ney",
"suffix": ""
}
],
"year": 2004,
"venue": "Proceedings of The International Conference on Language Resources and Evaluation",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "M. Popovic and H. Ney 2004. Towards the use of word stems and suffixes for statistical machine trans- lation, In Proceedings of The International Confer- ence on Language Resources and Evaluation.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Phrase-based backoff models for machine translation of highly inflected languages",
"authors": [
{
"first": "M",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Kirchhoff",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of the European Chapter of the ACL",
"volume": "",
"issue": "",
"pages": "41--48",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "M. Yang and K. Kirchhoff 2006 Phrase-based backoff models for machine translation of highly inflected languages, In Proceedings of the European Chapter of the ACL, pages 41-48, 2006.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Statistical Phrase-Based Translation",
"authors": [
{
"first": "Philipp",
"middle": [],
"last": "Koehn",
"suffix": ""
},
{
"first": "Franz",
"middle": [
"Josef"
],
"last": "Och",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Marcu",
"suffix": ""
}
],
"year": 2003,
"venue": "Proceedings of HLT-NAACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Philipp Koehn, Franz Josef Och, Daniel Marcu. 2003. Statistical Phrase-Based Translation, In Proceed- ings of HLT-NAACL 2003.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Enriching Morphologically Poor Languagesfor Statistical Machine Translation",
"authors": [
{
"first": "Eleftherios",
"middle": [],
"last": "Avramidis",
"suffix": ""
},
{
"first": "Philipp",
"middle": [],
"last": "Koehn",
"suffix": ""
}
],
"year": 2008,
"venue": "Proceedings of ACL-08",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Eleftherios Avramidis, Philipp Koehn 2008. Enrich- ing Morphologically Poor Languagesfor Statistical Machine Translation , In Proceedings of ACL-08, HLT.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "A Lightweight Stemmer for Hindi",
"authors": [
{
"first": "Ananthakrishnan",
"middle": [],
"last": "Ramanathan",
"suffix": ""
},
{
"first": "Durgesh",
"middle": [],
"last": "Rao",
"suffix": ""
}
],
"year": 2003,
"venue": "Workshop on Computational Linguistics for South-Asian Languages, EACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ananthakrishnan Ramanathan and Durgesh Rao 2003. A Lightweight Stemmer for Hindi 2003, Workshop on Computational Linguistics for South-Asian Lan- guages, EACL.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Simple Syntactic and Morphological Processing Can Help English-Hindi Statistical Machine Translation",
"authors": [
{
"first": "Ananthakrishnan",
"middle": [],
"last": "Ramanathan",
"suffix": ""
},
{
"first": "Pushpak",
"middle": [],
"last": "Bhattacharyya",
"suffix": ""
},
{
"first": "Jayprasad",
"middle": [],
"last": "Hegde",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Ritesh",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Shah",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Sasikumar",
"suffix": ""
}
],
"year": 2007,
"venue": "Proceedings of International Joint Conference on Natural Language Processing",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ananthakrishnan Ramanathan, Pushpak Bhat- tacharyya, Jayprasad Hegde, Ritesh M. Shah., Sasikumar M 2007 Simple Syntactic and Mor- phological Processing Can Help English-Hindi Statistical Machine Translation, In Proceedings of International Joint Conference on Natural Language Processing,2007.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Efficient Dynamic Programming Search Algorithms for Phrase-based SMT",
"authors": [
{
"first": "Christoph",
"middle": [],
"last": "Tillman",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of the Workshop CHPSLP at HLT'06",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Christoph Tillman 2006 Efficient Dynamic Program- ming Search Algorithms for Phrase-based SMT, In Proceedings of the Workshop CHPSLP at HLT'06.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Robust, applied morphological generation",
"authors": [
{
"first": "G",
"middle": [],
"last": "Minnen",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Carroll",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Pearce",
"suffix": ""
}
],
"year": 2000,
"venue": "Proceedings of the 1st International Natural Language Generation Conference",
"volume": "",
"issue": "",
"pages": "201--208",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Minnen, G., J. Carroll and D. Pearce 2000 Robust, ap- plied morphological generation, In Proceedings of the 1st International Natural Language Generation Conference, Mitzpe Ramon, Israel pp 201-208.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "A maximum entropy word aligner for arabic-english machine translation",
"authors": [
{
"first": "Abraham",
"middle": [],
"last": "Ittycheriah",
"suffix": ""
},
{
"first": "Salim",
"middle": [],
"last": "Roukos",
"suffix": ""
}
],
"year": 2005,
"venue": "Proceedings of HLT/EMNLP, HLT-05",
"volume": "",
"issue": "",
"pages": "89--96",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Abraham Ittycheriah and Salim Roukos 2005. A max- imum entropy word aligner for arabic-english ma- chine translation, In Proceedings of HLT/EMNLP, HLT-05, pages 89-96, Stroudsburg, PA, USA. Asso- ciation for Computational Linguistics.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Conditional random fields: Probabilistic models for segmenting and labeling sequence data",
"authors": [
{
"first": "John",
"middle": [],
"last": "Lafferty",
"suffix": ""
},
{
"first": "Andrew",
"middle": [],
"last": "Mccallum",
"suffix": ""
},
{
"first": "Fernando",
"middle": [],
"last": "Pereira",
"suffix": ""
}
],
"year": 2001,
"venue": "Proc. International Conference on Machine Learning (ICML)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "John Lafferty, Andrew McCallum, and Fernando Pereira 2002. Conditional random fields: Prob- abilistic models for segmenting and labeling se- quence data , in Proc. International Conference on Machine Learning (ICML), 2001.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "BLEU: a method for automatic evaluation of machine translation",
"authors": [
{
"first": "K",
"middle": [],
"last": "Papineni",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Roukos",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Ward",
"suffix": ""
},
{
"first": "W",
"middle": [
"J"
],
"last": "Zhu",
"suffix": ""
}
],
"year": 2002,
"venue": "Proc. ACL-2002: 40th Annual meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "311--318",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Papineni, K., Roukos, S., Ward, T., and Zhu, W. J. 2002. BLEU: a method for automatic evaluation of machine translation , in Proc. ACL-2002: 40th An- nual meeting of the Association for Computational Linguistics pp. 311 -318.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"uris": null,
"text": "Steps involved in manual generation of verb phrases",
"type_str": "figure",
"num": null
},
"FIGREF1": {
"uris": null,
"text": "Coverage percentage for different settings",
"type_str": "figure",
"num": null
},
"FIGREF2": {
"uris": null,
"text": "Change in BLEU for different corpus sizes",
"type_str": "figure",
"num": null
},
"TABREF1": {
"html": null,
"text": "",
"type_str": "table",
"content": "<table><tr><td>: Example of verbs belonging to different</td></tr><tr><td>groups based on their light verb</td></tr></table>",
"num": null
},
"TABREF3": {
"html": null,
"text": "Count of verbs belonging to different classes in Hindi clusters now contain exactly the same variations and can be aggregated together easily. These clusters are put in N different 'verb classes' so that all verbs occurring with the same variations are under",
"type_str": "table",
"content": "<table/>",
"num": null
}
}
}
} |