File size: 77,565 Bytes
6fa4bc9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 |
{
"paper_id": "2004",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:21:37.537306Z"
},
"title": "Phrase-based alignment combining corpus cooccurrences and linguistic knowledge",
"authors": [
{
"first": "Adri\u00e0",
"middle": [],
"last": "De Gispert",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Jos\u00e9",
"middle": [
"B"
],
"last": "Mari\u00f1o",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Josep",
"middle": [
"M"
],
"last": "Crego",
"suffix": "",
"affiliation": {},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "This paper introduces a phrase alignment strategy that seeks phrase and word links in two stages using cooccurrence measures and linguistic information. On a first stage, the algorithm finds high-precision links involving a linguistically-derived set of phrases, leaving word alignment to be performed in a second phase. Experiments have been carried out for an English-Spanish parallel corpus, and we show how phrase cooccurrence measures convey a complementary information to word cooccurrences, and a stronger evidence of a good alignment. Alignment Error Rate (AER) results are presented, being competitive with and even outperforming state-of-the-art alignment algorithms.",
"pdf_parse": {
"paper_id": "2004",
"_pdf_hash": "",
"abstract": [
{
"text": "This paper introduces a phrase alignment strategy that seeks phrase and word links in two stages using cooccurrence measures and linguistic information. On a first stage, the algorithm finds high-precision links involving a linguistically-derived set of phrases, leaving word alignment to be performed in a second phase. Experiments have been carried out for an English-Spanish parallel corpus, and we show how phrase cooccurrence measures convey a complementary information to word cooccurrences, and a stronger evidence of a good alignment. Alignment Error Rate (AER) results are presented, being competitive with and even outperforming state-of-the-art alignment algorithms.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Generally speaking, the automatic alignment task aims at revealing the relationship between bilingual units (be them words, subwords or phrases) in a given parallel corpus, detecting which words from each language are connected together in a given situation. This has many applications in natural language processing, such as bilingual dictionaries extraction or transfer rules learning. However, it is in the context of statistical machine translation where it becomes particularly crucial. As an essential block in the learning process of current statistical translation models (single-word or phrase-based, conditionalor joined-probability based), its correct production has a sound correlation with translation quality [1] .",
"cite_spans": [
{
"start": 723,
"end": 726,
"text": "[1]",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "This relevance has been corresponded by many previous works on the matter, including a shared task in the frame of the HLT-NAACL 2003 Workshop on Building and Using Parallel Texts [2] . Several competing systems were presented and evaluated against a manual reference using AER, the most widely used evaluation measure. Among the wide range of approaches presented, two basic trends should be highlighted.",
"cite_spans": [
{
"start": 180,
"end": 183,
"text": "[2]",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "On the one hand, the one based on the freely-available GIZA++ software [3] , implementing well-known IBM and HMM translation models [4, 5] , is considered to be the state-of-the-art. Nearly all current approaches to statistical translation rely on the results of Giza-based alignment processes to learn their models (see [6, 7, 8] among others). Although this approach, which is not solely focused on the alignment task (as alignment is considered to be a sort of hidden variable in a complete translation model), provides quite satisfactory results even with small corpora, it suffers from two structural flaws that limit its performance. Due to the model definition of word alignment as a function from positions in target sentence to positions in source sentence, it is strictly asymmetric, generating one-to-many word alignments that do not account for many translation phenomena. This effect has been tackled by several kinds of symmetrization heuristics (all of them linguistically blind), in search of a strategy to provide posterior phrase-based translation systems with the most accurate possible source. Moreover, the complexity of IBM models and their overload of parameters to estimate turn it very hard to introduce linguistic information into this setting in a reasonable way (some efforts being done in [9] ).",
"cite_spans": [
{
"start": 71,
"end": 74,
"text": "[3]",
"ref_id": "BIBREF2"
},
{
"start": 132,
"end": 135,
"text": "[4,",
"ref_id": "BIBREF3"
},
{
"start": 136,
"end": 138,
"text": "5]",
"ref_id": "BIBREF4"
},
{
"start": 321,
"end": 324,
"text": "[6,",
"ref_id": "BIBREF5"
},
{
"start": 325,
"end": 327,
"text": "7,",
"ref_id": "BIBREF6"
},
{
"start": 328,
"end": 330,
"text": "8]",
"ref_id": "BIBREF7"
},
{
"start": 1318,
"end": 1321,
"text": "[9]",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "On the other side of the spectrum, we find the approach based on word cooccurrences and link probabilities, presented in [10] . Its relative simplicity, its flexibility to introduce more knowledge sources, its symmetry and its promising results [2] make it appealing despite its dependence on empirical data and tuning strategies. However, the most important disadvantage of this approach is the one-to-one constrain, producing high precision alignments with low recall, what can represent a severe limitation to its use in practical translation systems.",
"cite_spans": [
{
"start": 121,
"end": 125,
"text": "[10]",
"ref_id": "BIBREF9"
},
{
"start": 245,
"end": 248,
"text": "[2]",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "We present in this paper an alignment strategy that is also based on bilingual cooccurrences, but aims at finding phrase-to-phrase alignment by using linguistic knowledge and, thereby, overcoming the one-to-one limitation. In section 2 this new strategy is described in detail, whereas sections 3, 4 and 5 describe our experimental framework and provide and discuss partial and complete alignment results. Finally, section 6 concludes and out-lines future research lines.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "Recent research efforts in statistical translation have clearly focused on improving translation quality by training models not only based on single words, but also on phrases. Typically, initial Giza-based alignments are generated and a symmetrization strategy is followed to obtain the core alignment from which bilingual phrases are built. These in turn are fed to the statistical translation model for estimation. However, current symmetrization strategies lack linguistic knowledge to decide ambiguities, and the translation model faces the task of learning translation probabilities from a noisy source. To face this problem, we present an alignment strategy that generates directly a phrase alignment from corpus cooccurrence counts. In contrast to previous work [11] , we do so by generating very high-confidence links between phrases before proceeding onto word alignment. This search for phrase links is limited to a small adequate set of possible phrases. In the following subsections our proposal is described in detail.",
"cite_spans": [
{
"start": 770,
"end": 774,
"text": "[11]",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "From word-based to phrase-based alignment",
"sec_num": "2."
},
{
"text": "Association or cooccurrence measures extracted from parallel corpora give strong evidence of so-called translation equivalence [12] , or simply alignment adequacy, between a pair of phrases or words. Among these measures we find Dice-score, \u03c6 2 score and some others, offering a similar performance. In this paper, we have used \u03c6 2 as presented in [13] , which is defined by the following equation:",
"cite_spans": [
{
"start": 127,
"end": 131,
"text": "[12]",
"ref_id": "BIBREF11"
},
{
"start": 348,
"end": 352,
"text": "[13]",
"ref_id": "BIBREF12"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Word and phrases association measures",
"sec_num": "2.1."
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "\u03c6 2 = (ad \u2212 bc) 2 (a + b)(a + c)(b + d)(c + d)",
"eq_num": "(1)"
}
],
"section": "Word and phrases association measures",
"sec_num": "2.1."
},
{
"text": "where a is the number of times two words (or phrases) cooccur, b and c are the number of times one occurs and the other does not, and finally d accounts for the number of times neither one nor the other occur in the data set. In our implementation, we have defined the cooccurrence of a word (or phrase) appearing x times in a sentence and a word (or phrase) occurring y times in its translation as min(x, y), for two reasons: on the one hand, the alternative option given by the product xy leads to confusing results when computing b and c, as these can be negative because the times a word cooccurs with another can overweight the total occurrences of the word. On the other hand, the word alignment algorithm used estimates link probabilities from existing one-to-one links (see section 2.2.3), being min(x, y) the maximum number of links that can be established between the two words (in which case their probability is the highest). This way we preserve stochastic consistency.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Word and phrases association measures",
"sec_num": "2.1."
},
{
"text": "Despite this score can be easily computed for each possible pair of words from both languages, computational problems arise when dealing with every bigram, trigram or, in general, phrase for each language. However, these scores can convey a useful complementary information in many cases, as in the examples of table 1, where the phrase-to-word score -in bold face-is comparatively much better than all word-word scores for all words involved in Spanish idioms 'por favor' and 'a lo mejor' (note that \u221210log will be assumed when referring to \u03c6 2 ). Furthermore, it is reasonable to expect that the longer the phrases considered, the stronger the evidence of a good alignment adequacy, so long as we have a reasonable number of occurrences of the phrase. The main problem is then the practical impossibility to compute all combinations for even relatively small corpora (see table 2 for the number of cooccurrence combinations considering all word, bigrams and trigrams in the relatively-small corpus used in section 3). To tackle this, we propose an algorithm that tries to extract as much useful information from these phrase cooccurrence measures by performing a selection of only a subset of all possible phrases. We have chosen to perform this selection using linguistic criteria, although statistics could also be used, as discussed in section 2.2.1. This way, we expect the algorithm to leverage linguistic knowledge and empirical evidence in a memory-efficient way. ",
"cite_spans": [],
"ref_spans": [
{
"start": 874,
"end": 881,
"text": "table 2",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Word and phrases association measures",
"sec_num": "2.1."
},
{
"text": "We propose a phrase alignment strategy in four stages, as shown in figure 1. Firstly, from all possible sets of words we select a small set of 'interesting' phrases for each language. This selection is linguistically guided and should produce a set of phrases containing words that play a unique semantic role. Secondly, a high-precision phrase alignment algorithm links these phrases together (with phrases or single words) using cooccurrence mea- sures, and discarding uncertain links. After these phrase alignments have been produced, we run a word-to-word alignment algorithm based on link probabilities and shallow syntactic information in the fashion of [10] . This stage takes advantage of the complexity reduction derived from the previous linking. Finally, in the fourth stage a postprocessing of the resultant alignment is done, disambiguating cases such as unaligned words, and making decisions at a sentence level. These blocks are now described in detail.",
"cite_spans": [
{
"start": 660,
"end": 664,
"text": "[10]",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "A phrase alignment strategy",
"sec_num": "2.2."
},
{
"text": "The objective of this stage is double. Given the exponential nature of the amount of different phrase cooccurrences shown above, which makes it impossible to work with cooccurrences between all combinations for each language, a first objective is the reduction of this huge space to those being 'interesting' from an alignment/translation point of view. Our criterion is one of so-called translational equivalence, so we define as interesting those phrases expressing a same concept or being semantically linked in one language, as it is reasonable to expect that these might be aligned to (or might translate into) a single word (or phrase) in another language. On the other hand, a semantic classification of these phrases should improve cooccurrence measures by adding different instantiations of the same concept to a same measure.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrases selection and classification",
"sec_num": "2.2.1."
},
{
"text": "As about the selection of phrases, we have followed a linguistically-guided strategy. Specifically, we have implemented two selection criteria using complementary knowledge. Firstly, we detect verb groups using deterministic automata that implement a few simple rules, as shown in example 2. These rules take as input word forms, POS-tags and word lemmas (or base forms), and map the resulting phrase to the lemma of the head verb. This way, the classification improves coocccurrence counts for verb groups no matter how their full form is expressed, as long as they share the same base form of the head verb. This way, forms like we have brought or will we bring are considered equivalent and add a cooccurrence count for the base form \"bring\", increasing its evidence and reducing evidence for function words like \"have\" and \"will\" that act as modifiers and may therefore be expressed in many ways in the other language, as they do not convey a stand-alone meaning in the sentence. We expect this to produce a special gain in languages that have important declination in verb forms (like languages belonging to the Romance family, as Spanish).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrases selection and classification",
"sec_num": "2.2.1."
},
{
"text": "After detection rules are applied, all remaining words with POS-tag of a Verb are substituted by their base form to enforce the verb's cooccurrence evidence. However, to avoid possible mistakes from the POS-tagger or lemmatizer and ensure we are dealing with a known verb, we filter out the resulting base form against two lists containing 25988 possible verb forms in English and 12668 possible verb lemmas in Spanish, available from the maco+ and relax tagging package [14] . We have also implemented a selection based on idiomatic expressions. Specifically, we match the corpus against a list of 1496 and 49 usual idiomatic expressions that we have available for Spanish and English, respectively (again from the maco+ and relax package). These expressions (among which we find example like 'on the other hand' for English, and 'sin embargo' or 'a lo mejor' for Spanish) tend to convey a single meaning and we can expect them to be aligned together to one or more words in the other language. At the moment, we have not used any kind of dictionary, so these expressions are not classified according to their meaning(s). Other possible linguistically-guided selection rules could include regular expressions such as numbers, dates or times of the day (that could also be classified) or even collocations and phrasal verbs. As this selection is language-dependent, every language will define its own adequate rules.",
"cite_spans": [
{
"start": 471,
"end": 475,
"text": "[14]",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Phrases selection and classification",
"sec_num": "2.2.1."
},
{
"text": "If no linguistic knowledge is available, statistical procedures can also be used to obtain a set of possible phrases. For example, we can select the N most frequent bigrams, trigrams and Ngrams in general, or the ones having a very high bigram, trigram or Ngram probability (defining groups of words that consistently appear together in the text). We plan to investigate this in the short term, as commented in section 6.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrases selection and classification",
"sec_num": "2.2.1."
},
{
"text": "At the moment, we have restricted our selection to phrases built by sequential groups of words. However, as word alignment is not always affected by this restriction, it should be eliminated, allowing phrases to contain words that are separated by other words (such as separable phrasal verbs in English or many separable verbs in German).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrases selection and classification",
"sec_num": "2.2.1."
},
{
"text": "It is important to note that we do not expect this selection to be exhaustive, nor does it imply that the selected phrase will necessarily be linked together at the next stage (it is not a hard decision in terms of alignment). It is the phrase alignment stage that decides whether a phrase should be linked together, or whether the words should be left free to be linked word-to-word.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrases selection and classification",
"sec_num": "2.2.1."
},
{
"text": "In this stage cooccurrence measures are computed for each selected phrase in one language against all selected phrases and single words in the other language. Then, a competitive linking strategy [12] is used, but not until all words or phrases are linked, but until a certain threshold is surpassed. Basically, we choose the link with best phrase-phrase or phrase-word cooccurrence measure as long as this is better than the threshold. This strategy relies on the fact that phrase cooccurrence measures are a stronger evidence of translational equivalence than word, and the threshold (which has to be empirically tuned) ensures that we generate only the highest-confidence links. This way, not all selected phrases will be linked, but only those having a high cooccurrence evidence in the data.",
"cite_spans": [
{
"start": 196,
"end": 200,
"text": "[12]",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Phrases alignment",
"sec_num": "2.2.2."
},
{
"text": "Once the linking of two phrases is decided, one can use several strategies to determine the internal links between words inside the phrases, if that is desired. For example, internal links can be solved using the general word alignment algorithm, but restricting the search inside the phrase positions. However, often selected phrases will contain function words that tend to depend on each language syntax and are not easily linked to the other language words. For this reason, we have decided to introduce all internal links between linked phrases.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrases alignment",
"sec_num": "2.2.2."
},
{
"text": "As about the word alignment algorithm, we have implemented an iterative algorithm similar to the one pre-sented in [10] . Basically, an initial alignment is generated using word cooccurrence measures, from which link probabilities are estimated. Then, a best first search is performed, following an heuristic function based on the global aligned sentence link probabilities. The search is further improved with a syntactic constrain (also called cohesion constrain [15] ) and can introduce features on the links, such as a dependence on adjacent links. Our implementation allows certain positions to be prohibited, so that previous phrase alignment is fixed, although its links also compute in link probability estimation at each iteration.",
"cite_spans": [
{
"start": 115,
"end": 119,
"text": "[10]",
"ref_id": "BIBREF9"
},
{
"start": 465,
"end": 469,
"text": "[15]",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Word alignment",
"sec_num": "2.2.3."
},
{
"text": "Given the enormous space of possible word alignments to choose from, the heuristic function becomes the key to efficiency, so long as it is correctly defined. Basic parameters are:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Word alignment",
"sec_num": "2.2.3."
},
{
"text": "\u2022 the initial null probability, or the prob. that a word links to null (no word), which is necessary to make fully-and partially-aligned solutions comparable",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Word alignment",
"sec_num": "2.2.3."
},
{
"text": "\u2022 and the minimum score to accept a link between two words (hereafter referred to as mscore)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Word alignment",
"sec_num": "2.2.3."
},
{
"text": "These parameters must be set empirically for the optimal performance of the algorithm. We also found that restricting the search of possible links to a window in the other language not only made the algorithm much more efficient (turning it from exponential time to linear time with input sentence lengths), but also improved results by discarding the ambiguities generated by the repetition of frequent words (mostly function words). We define this window in the neighbourhood of the diagonal defined by the division between both sentence lengths. Of course, this window is completely dependent on the pair of languages considered (might even be eliminated for certain pairs), but in our case (English-Spanish) turned out to be optimal considering 8 words.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Word alignment",
"sec_num": "2.2.3."
},
{
"text": "The postprocessing stage should take the final alignment decisions using sentence-level information (ie. deciding whether unlinked words should be linked, looking for long-distance links, reconsidering the links for a word/phrase given all its links in all sentence pairs, etc). Ideally, it should also feedback into the phrase selection/alignment blocks to reconsider previous decisions using global information of all sentences. However, this stage is strongly connected to the posterior translation modeling. Although alignment can be and must be evaluated separately, we are of the opinion that it is not completely independent from the translation model. At the moment, we have not implemented any postprocessing technique in our system.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Postprocessing",
"sec_num": "2.2.4."
},
{
"text": "To experiment with the alignment strategy described above, we have worked with an English-Spanish parallel corpus, namely the VerbMobil database, which has been translated to Spanish in the framework of the LC-Star project (IST-2001-32216) . This data is the transcription of spontaneous dialogues in the appointments and meeting-planning domain. Table 3 shows the main statistics of the data used, namely number of sentences, words, vocabulary, percentage of singletons, and maximum and mean sentence lengths for each language, respectively. ",
"cite_spans": [
{
"start": 223,
"end": 239,
"text": "(IST-2001-32216)",
"ref_id": null
}
],
"ref_spans": [
{
"start": 347,
"end": 354,
"text": "Table 3",
"ref_id": "TABREF2"
}
],
"eq_spans": [],
"section": "Parallel corpus",
"sec_num": "3.1."
},
{
"text": "All preprocessing that has been carried out is described as follows:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preprocessing",
"sec_num": "3.2."
},
{
"text": "\u2022 Normalization of contracted forms for English (ie. wouldn't = would not, we've = we have) and Spanish (del = de el)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preprocessing",
"sec_num": "3.2."
},
{
"text": "\u2022 English data has been tagged using freely-available TnT tagger [16] , and base forms have been obtained using wnmorph, included in the WordNet package [17] .",
"cite_spans": [
{
"start": 65,
"end": 69,
"text": "[16]",
"ref_id": "BIBREF15"
},
{
"start": 153,
"end": 157,
"text": "[17]",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Preprocessing",
"sec_num": "3.2."
},
{
"text": "\u2022 Spanish data has been tagged using maco+ and relax package already mentioned. This software also generates a lemma or base form for each input word.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preprocessing",
"sec_num": "3.2."
},
{
"text": "\u2022 Date and time expressions (which are numerous in the domain) have been substituted by a unified tag using a semi-automatic technique [18] .",
"cite_spans": [
{
"start": 135,
"end": 139,
"text": "[18]",
"ref_id": "BIBREF17"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Preprocessing",
"sec_num": "3.2."
},
{
"text": "\u2022 Finally, punctuation marks have been left out (as we expect to use the corpora for spoken language translation experiments).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preprocessing",
"sec_num": "3.2."
},
{
"text": "For evaluation purposes, we have randomly selected from this data two sets: a validation set of 100 sentences (for tuning of parameters) and a test set of 400 sentences. These have been manually aligned following the criterion of Sure and Possible links, in order to compute Alignment Error Rate (AER) as described in [19] and widely used in literature (including the above-mentioned HLT-NAACL 2003 Shared task).",
"cite_spans": [
{
"start": 318,
"end": 322,
"text": "[19]",
"ref_id": "BIBREF18"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation scheme",
"sec_num": "3.3."
},
{
"text": "It has been shown that the percentage of Sure and Possible links in the gold standard reference has a strong influence in the final AER result, favouring high-precision alignments when Possible links outnumber Sure links, and favouring high-recall alignments otherwise [20] . Our criterion has been to produce Possible links only when they allow combinations which are considered equally correct, as a reference with too many Possible links suffers from a resolution loss, causing several different alignments to be equally rated. This way, we have 80% Sure links and 20% Possible links. Evaluations are performed without taking links to NULL into account.",
"cite_spans": [
{
"start": 269,
"end": 273,
"text": "[20]",
"ref_id": "BIBREF19"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation scheme",
"sec_num": "3.3."
},
{
"text": "In this section, we evaluate separately the phrase selection, classification and alignment blocks described in sections 2.2.1 and 2.2.2. First we present results of the phrase alignment selecting only verb groups, to continue with results with only idiomatic expressions alignment. Finally, complete alignment results are presented, comparing performances for the isolated word-to-word alignment algorithm and the complete phrase-alignment strategy proposed against state-of-the-art alignments.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase alignment results",
"sec_num": "4."
},
{
"text": "Verb groups detection rules include 14 rules for English language and just 6 for Spanish, which usually employs declined verb forms omitting thus personal pronouns and using thus a single word. Verb groups rules have detected a total of:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Verb groups",
"sec_num": "4.1."
},
{
"text": "\u2022 1156 verb groups in English, classified into 238 different verb lemmas",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Verb groups",
"sec_num": "4.1."
},
{
"text": "\u2022 658 verb groups in Spanish, classified into 188 different verb lemmas",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Verb groups",
"sec_num": "4.1."
},
{
"text": "The classification of these phrases produces an efficient reduction of the cooccurrence table from 0.35M to 0.33M, we do not compute cooccurrence counts for all words internal to the phrase, but just for the lemma of its head verb. The results of the phrase alignment with these phrases are shown in the upper side of table 4, changing the value of the threshold to accept phrase links from more restrictive to less restrictive. A restriction that the linked pair cooccurs at least twice has also been used. Surprisingly, this relatively simple selection strategy provides very encouraging results, as Precision is consistently higher than 98 % for the three thresholds used, whereas Recall achieves around 9%. We have to keep in mind that these phrase links will necessarily boost Recall with respect to the isolated word aligner, as it is a one-toone algorithm, unable to produce these links. As about Precision, the high figures are due to the greater statistical evidence of phrase cooccurrence measures with respect to In this case, although Recall is much smaller than when considering verb groups, two points are worth raising. First, we have again a nearly error-free alignment using a relatively small set of phrases. And second, but not less important, that we expect this Recall to complement the previous experiment and further boost the global alignment Recall, as we find no verb groups among the idiomatic expressions considered. Results for the complete phrase alignment strategy follow in the next section.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Verb groups",
"sec_num": "4.1."
},
{
"text": "As baseline alignments, we have aligned our data using GIZA++ from English to Spanish and vice versa (performing 1 5 H 5 3 3 4 3 iterations), and have evaluated two symmetrization strategies, namely the union and the intersection. Their results are shown in the first four rows of table 5.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Complete alignment results and discussion",
"sec_num": "5."
},
{
"text": "We have also used the word alignment algorithm presented in section 2.2.3 to align the data without any kind of previous phrase selection and alignment, thus producing the one-to-one alignment shown in the fifth row. For this result, we have run three iterations with a mscore = 30, and three iterations further restricting it to 6 to achieve high precision, always using cohesion constrain and adjacency features up to two positions (initial NULL cost being set to 15). In contrast to giza++ intersection (the only baseline alignment that is also one-to-one and thus subject to a fair comparison), we observe a slight reduction in Precision and a slight increase in Recall, leading to an also slight AER reduction. However, both alignments skip around 30 % of good links (far below the other alignments recall), which make them unpractical for posterior statistical translation modeling. On the contrary, the giza++ union alignment, which rates the best AER, suffers from a severe decline in Precision as compared to all other alternatives, compensated by a very important boost in Recall.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Complete alignment results and discussion",
"sec_num": "5."
},
{
"text": "Remarkably, our phrase word aligner, whose results are shown in the last two rows (selecting verb and idiomatic phrases with the two less-restrictive \u03c6 2 thresholds shown in table 4), preserves as high a Precision level as the giza++ intersection, while providing a Recall increase of over 5 absolute points with respect to intersection, and of about 4 points with respect to the one-to-one aligner. The achieved Recall figure is absolutely comparable with all other alignments (that nevertheless offer worse Precision results), except for the giza++ union, still to be beaten in Recall.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Complete alignment results and discussion",
"sec_num": "5."
},
{
"text": "It is interesting to note the Precision improvement when comparing the word-to-word aligner to the phrase aligner proposed, which is due to two factors. On the one hand, the previous phrase alignment introduces links with a higher precision than that of the one-to-one aligner, and on the other hand, this previous linking results in a complexity reduction (less ambiguity) that simplifies the task of the one-to-one aligner, improving its performance. We find these results to be very promising: the alignment strategy achieves competitive results while still making a relatively small use of linguistic knowledge. However, this has already led to an important recall boost with respect to the word-to-word aligner, at no precision cost. As its architecture is open to an easy introduction of more information, many other knowledge sources could be used, as outlined in the following and concluding section.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Complete alignment results and discussion",
"sec_num": "5."
},
{
"text": "In this paper, we have presented a phrase alignment strategy that combines cooccurrence measures extracted from bilingual corpora and linguistic knowledge in detail. Re-sults have been reported, that show a very positive tendency, with a state-of-the-art Precision figure and a very high Recall, which could be improved in the short term. Besides, we do not find any alignment beating the rest in both Precision and Recall measures. However, when it comes to further research, and bearing in mind that we are interested in statistical machine translation, our two main research priorities are to evaluate this alignment strategy with other parallel corpora, and to evaluate it in the framework of a translation experiment, in addition to AER results. We plan to do so using several translation models in the very short term.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion and further research",
"sec_num": "6."
},
{
"text": "Future work also includes new and more linguistically-informed phrase selection schemes, such as allowing phrases containing non-consecutive words (ie., negative verb groups or expressions like 'i will probably leave') or searching for English phrasal verbs from available lists. Categorisation of numbers, dates and time expressions during the phrase selection and classification stage should also be investigated. It is our view that the alignment strategy, which does not imply a hard alignment decision when selecting the phrases, could help resolve the inherent ambiguity when detecting these expressions from a monolingual point of view (ie., 'a las tres' might refer to a time of the day or simply a number). Connected with this, another interesting point to investigate is to allow phrases to add a cooccurrence count for each of their meanings, in case they have more than one, so that alignment plays a sense disambiguation role. In this case, adequate dictionaries must be provided. Information from statistical chunkers could also be introduced in the medium term.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion and further research",
"sec_num": "6."
},
{
"text": "Regarding the alignment algorithm, we plan to investigate ways of allowing a certain degree of freedom to reconsider phrase links, which are currently fixed (although their precision is very high). Finally, postprocessing techniques discussed in section 2.2.4 will have to be tested, seeking a further boost in recall by using statistics of all aligned sentences. For example, new cooccurrence measures could be computed for all phrases that can be built in the neighbourhood of unaligned words, to consider whether these should be linked or left alone.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion and further research",
"sec_num": "6."
}
],
"back_matter": [
{
"text": "The authors want to thank Patrik Lambert, Javier P\u00e9rez and Antonio Bonafonte (members of the TALP Research Center) for their valuable contributions to this paper. We would also like to thank Colin Cherry for his kind input on the cooccurrences-based word alignment algorithm.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgements",
"sec_num": "7."
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "A systematic comparison of various statistical alignment models",
"authors": [
{
"first": "F",
"middle": [],
"last": "Och",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Ney",
"suffix": ""
}
],
"year": 2003,
"venue": "Computational Linguistics",
"volume": "29",
"issue": "1",
"pages": "19--51",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "F. Och and H. Ney, \"A systematic comparison of various statistical alignment models,\" Computa- tional Linguistics, vol. 29, no. 1, pp. 19-51, March 2003.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "An evaluation exercise for word alignment",
"authors": [
{
"first": "R",
"middle": [],
"last": "Mihalcea",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Pedersen",
"suffix": ""
}
],
"year": 2003,
"venue": "HLT-NAACL 2003 Workshop: Building and Using Parallel Texts: Data Driven Machine Translation and Beyond",
"volume": "",
"issue": "",
"pages": "1--10",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "R. Mihalcea and T. Pedersen, \"An evaluation ex- ercise for word alignment,\" in HLT-NAACL 2003 Workshop: Building and Using Parallel Texts: Data Driven Machine Translation and Beyond, R. Mihal- cea and T. Pedersen, Eds. Edmonton, Alberta, Canada: Association for Computational Linguis- tics, May 31 2003, pp. 1-10.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Giza++ software",
"authors": [],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Giza++ software, \"http://www-i6.informatik.rwth- aachen.de/\u02dcoch/software/giza++.html.\"",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "The mathematics of statistical machine translation",
"authors": [
{
"first": "P",
"middle": [],
"last": "Brown",
"suffix": ""
},
{
"first": "S",
"middle": [
"Della"
],
"last": "Pietra",
"suffix": ""
},
{
"first": "V",
"middle": [
"Della"
],
"last": "Pietra",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Mercer",
"suffix": ""
}
],
"year": 1993,
"venue": "Computational Linguistics",
"volume": "19",
"issue": "2",
"pages": "263--311",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "P. Brown, S. Della Pietra, V. Della Pietra, and R. Mercer, \"The mathematics of statistical machine translation,\" Computational Linguistics, vol. 19, no. 2, pp. 263-311, 1993.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Hmm-based word alignment in statistical translation",
"authors": [
{
"first": "S",
"middle": [],
"last": "Vogel",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Ney",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Tillmann",
"suffix": ""
}
],
"year": 1996,
"venue": "Proc. of the Int. Conf. on Computational Linguistics, COL-ING'96",
"volume": "",
"issue": "",
"pages": "836--841",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "S. Vogel, H. Ney, and C. Tillmann, \"Hmm-based word alignment in statistical translation,\" Proc. of the Int. Conf. on Computational Linguistics, COL- ING'96, pp. 836-841, August 1996.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Improved alignment models for statistical machine translation",
"authors": [
{
"first": "F",
"middle": [],
"last": "Och",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Tillmann",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Ney",
"suffix": ""
}
],
"year": 1999,
"venue": "Proc. of the Joint Conf. of Empirical Methods in Natural Language Processing and Very Large Corpora",
"volume": "",
"issue": "",
"pages": "20--28",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "F. Och, C. Tillmann, and H. Ney, \"Improved align- ment models for statistical machine translation,\" Proc. of the Joint Conf. of Empirical Methods in Natural Language Processing and Very Large Cor- pora, pp. 20-28, June 1999.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Using X-grams for speech-to-speech translation",
"authors": [
{
"first": "A",
"middle": [],
"last": "De Gispert",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Mari\u00f1o",
"suffix": ""
}
],
"year": 2002,
"venue": "Proc. of the 7th Int. Conf. on Spoken Language Processing",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "A. de Gispert and J. Mari\u00f1o, \"Using X-grams for speech-to-speech translation,\" Proc. of the 7th Int. Conf. on Spoken Language Processing, ICSLP'02, September 2002.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Statistical phrasebased translation",
"authors": [
{
"first": "P",
"middle": [],
"last": "Koehn",
"suffix": ""
},
{
"first": "F",
"middle": [],
"last": "Och",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Marcu",
"suffix": ""
}
],
"year": 2003,
"venue": "Proc. of the Human Language Technology Conference 2003 (HLT-NAACL'03)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "P. Koehn, F. Och, and D. Marcu, \"Statistical phrase- based translation,\" Proc. of the Human Language Technology Conference 2003 (HLT-NAACL'03), May 2003.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Extensions to hmm-based statistical word alignment models",
"authors": [
{
"first": "K",
"middle": [],
"last": "Toutanova",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Tolga Ilhan",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Manning",
"suffix": ""
}
],
"year": 2002,
"venue": "Proc. of the Conf. on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "K. Toutanova, H. Tolga Ilhan, and C. Manning, \"Ex- tensions to hmm-based statistical word alignment models,\" Proc. of the Conf. on Empirical Methods in Natural Language Processing, EMNLP'02, July 2002.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "A probability model to improve word alignment",
"authors": [
{
"first": "C",
"middle": [],
"last": "Cherry",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Lin",
"suffix": ""
}
],
"year": 2003,
"venue": "41st Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "C. Cherry and D. Lin, \"A probability model to im- prove word alignment,\" 41st Annual Meeting of the Association for Computational Linguistics, July 2003.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "A phrase-based, joint probability model for statistical machine translation",
"authors": [
{
"first": "D",
"middle": [],
"last": "Marcu",
"suffix": ""
},
{
"first": "W",
"middle": [],
"last": "Wong",
"suffix": ""
}
],
"year": 2002,
"venue": "Proc. of the Conf. on Empirical Methods in Natural Language Processing",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "D. Marcu and W. Wong, \"A phrase-based, joint probability model for statistical machine transla- tion,\" Proc. of the Conf. on Empirical Methods in Natural Language Processing, EMNLP'02, July 2002.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Empirical Methods for Exploiting Parallel Text",
"authors": [
{
"first": "D",
"middle": [],
"last": "Melamed",
"suffix": ""
}
],
"year": 2001,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "D. Melamed, Empirical Methods for Exploiting Parallel Text. Cambridge, MA: MIT Press, 2001.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Identifying word correspondences in parallel texts",
"authors": [
{
"first": "W",
"middle": [],
"last": "Gale",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Church",
"suffix": ""
}
],
"year": 1991,
"venue": "4th Speech and Natural Language Workshop",
"volume": "",
"issue": "",
"pages": "152--157",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "W. Gale and K. Church, \"Identifying word corre- spondences in parallel texts,\" 4th Speech and Natu- ral Language Workshop, pp. 152-157, 1991.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Morphosyntactic analysis and parsing of unrestricted spanish text",
"authors": [
{
"first": "J",
"middle": [],
"last": "Atserias",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Carmona",
"suffix": ""
},
{
"first": "I",
"middle": [],
"last": "Castell\u00f3n",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Cervell",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Civit",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "M\u00e0rquez",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Mart\u00ed",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Padr\u00f3",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Placer",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Rodr\u00edguez",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Taul\u00e9",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Turmo",
"suffix": ""
}
],
"year": 1998,
"venue": "1st Int. Conf. on Language Resources and Evaluation, LREC'98",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "J. Atserias, J. Carmona, I. Castell\u00f3n, S. Cervell, M. Civit, L. M\u00e0rquez, M. Mart\u00ed, L. Padr\u00f3, R. Placer, H. Rodr\u00edguez, M. Taul\u00e9, and J. Turmo, \"Mor- phosyntactic analysis and parsing of unrestricted spanish text,\" 1st Int. Conf. on Language Resources and Evaluation, LREC'98, 1998.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Word alignment with cohesion constraint",
"authors": [
{
"first": "C",
"middle": [],
"last": "Cherry",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Lin",
"suffix": ""
}
],
"year": 2003,
"venue": "Proceedings of HLT/NAACL'03",
"volume": "",
"issue": "",
"pages": "49--51",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "C. Cherry and D. Lin, \"Word alignment with co- hesion constraint,\" Proceedings of HLT/NAACL'03, pp. 49-51, Edmonton 2003.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "TnT -a statistical part-of-speech tagger",
"authors": [
{
"first": "T",
"middle": [],
"last": "Brants",
"suffix": ""
}
],
"year": 2000,
"venue": "Proc. of the Sixth Applied Natural Language Processing (ANLP-2000)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "T. Brants, \"TnT -a statistical part-of-speech tagger,\" in Proc. of the Sixth Applied Natural Language Processing (ANLP-2000), Seattle, WA, 2000. [Online]. Available: http://www.coli.uni- sb.de/\u02dcthorsten/tnt",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "WordNet: a lexical database for the English language",
"authors": [],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "WordNet: a lexical database for the English lan- guage, \"http://www.cogsci.princeton.edu/\u02dcwn.\"",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Experiments in word-ordering and morphological preprocessing for transducer-based statistical machine translation",
"authors": [
{
"first": "A",
"middle": [],
"last": "De Gispert",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Mari\u00f1o",
"suffix": ""
}
],
"year": 2003,
"venue": "IEEE Automatic Speech Recognition and Understanding Workhsop, ASRU'03",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "A. de Gispert and J. Mari\u00f1o, \"Experiments in word-ordering and morphological preprocessing for transducer-based statistical machine translation,\" IEEE Automatic Speech Recognition and Under- standing Workhsop, ASRU'03, November 2003.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Improved statistical alignment models",
"authors": [
{
"first": "F",
"middle": [],
"last": "Och",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Ney",
"suffix": ""
}
],
"year": 2000,
"venue": "38th Annual Meeting of the Association for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "440--447",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "F. Och and H. Ney, \"Improved statistical alignment models,\" 38th Annual Meeting of the Association for Computational Linguistics, pp. 440-447, Octo- ber 2000.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Alignment of parallel corpora exploiting asymmetrically aligned phrases",
"authors": [
{
"first": "P",
"middle": [],
"last": "Lambert",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Castell",
"suffix": ""
}
],
"year": 2004,
"venue": "4th Int. Conf. on Language Resources and Evaluation, LREC'04",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "P. Lambert and N. Castell, \"Alignment of parallel corpora exploiting asymmetrically aligned phrases,\" 4th Int. Conf. on Language Resources and Evalua- tion, LREC'04, May 2004.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"uris": null,
"type_str": "figure",
"text": "Phrase alignment stages.",
"num": null
},
"FIGREF1": {
"uris": null,
"type_str": "figure",
"text": "PrP + VB(L=have) {+ been} + VB{G} VB(L=have) + PrP {+ been} + VB{G} PrP + VB(L=be) +VBG VB(L=be) + PrP +VBG PrP + MD(L=will/would/...) + VB MD(L=will/would/...) + PrP + VB PrP + VB VB(L=do) + PrP + VB VB(L=be) + PrP PrP: Personal Pronoun VB / MD / VBG: Verb / Modal / Gerund (PennTree Bank POS) L: Lemma (or base form) { } / ( ): optionality / instantiation",
"num": null
},
"FIGREF2": {
"uris": null,
"type_str": "figure",
"text": "Verb group detection rules used for English.",
"num": null
},
"TABREF0": {
"content": "<table><tr><td colspan=\"2\">please</td><td/><td>maybe</td></tr><tr><td>por favor 1.2 22.4</td><td>0.9</td><td>a lo</td><td>23.1 18.2 8.0</td></tr><tr><td/><td/><td colspan=\"2\">mejor 12.2</td></tr></table>",
"html": null,
"type_str": "table",
"text": "Examples of \u03c6 2 between words and phrases.",
"num": null
},
"TABREF1": {
"content": "<table><tr><td/><td># Coocs</td></tr><tr><td>unigrams</td><td>0.35M</td></tr><tr><td>+bigrams</td><td>4.36M</td></tr><tr><td colspan=\"2\">+trigrams 10.42M</td></tr></table>",
"html": null,
"type_str": "table",
"text": "Number of different cooccurrence entries considering unigrams, bigrams and trigrams for VerbMobilEnglish-Spanish corpus.",
"num": null
},
"TABREF2": {
"content": "<table><tr><td colspan=\"4\">VMobil words vocab. singlets. Lmax Lmean</td></tr><tr><td>English 228328 3276</td><td>39 %</td><td>66</td><td>7.6</td></tr><tr><td>Spanish 219782 5084</td><td>43 %</td><td>66</td><td>7.3</td></tr></table>",
"html": null,
"type_str": "table",
"text": "Corpus used: 30054 sentences per language.",
"num": null
},
"TABREF3": {
"content": "<table><tr><td/><td colspan=\"2\">Recall Precision</td></tr><tr><td>Verbs \u03c6 2 < 8</td><td>8.07</td><td>99.02</td></tr><tr><td>Verbs \u03c6 2 < 10</td><td>9.00</td><td>99.12</td></tr><tr><td>Verbs \u03c6 2 < 15</td><td>9.68</td><td>98.69</td></tr><tr><td>Idioms \u03c6 2 < 5</td><td>2.01</td><td>98.48</td></tr><tr><td>Idioms \u03c6 2 < 10</td><td>3.06</td><td>99.00</td></tr><tr><td>Idioms \u03c6 2 < 15</td><td>3.50</td><td>97.41</td></tr><tr><td colspan=\"3\">single word cooccurrences. Interestingly, no error is in-</td></tr><tr><td colspan=\"3\">troduced from the first to the second case, whereas Recall</td></tr><tr><td colspan=\"2\">is increased by 1.2 absolute points.</td><td/></tr><tr><td>4.2. Idiomatic expressions</td><td/><td/></tr><tr><td colspan=\"3\">Regarding idiomatic expressions, a total of 20 English</td></tr><tr><td colspan=\"3\">and 99 Spanish expressions have been detected in the</td></tr><tr><td colspan=\"3\">parallel corpora, matching with the available lists pre-</td></tr><tr><td colspan=\"3\">sented in section 2.2.1. These have not been classi-</td></tr><tr><td colspan=\"3\">fied, leading to an increase in the cooccurrences table</td></tr><tr><td colspan=\"3\">from 0.35M to 0.37M entries. The results when align-</td></tr><tr><td colspan=\"3\">ing only idiomatic expressions phrase-phrase and phrase-</td></tr><tr><td colspan=\"3\">words links are shown in the lower side of table 4, again</td></tr><tr><td>for different thresholds.</td><td/><td/></tr></table>",
"html": null,
"type_str": "table",
"text": "Phrase alignment results considering verb groups only and idiomatic expressions only.",
"num": null
},
"TABREF4": {
"content": "<table><tr><td/><td colspan=\"3\">Recall Precision AER</td></tr><tr><td>giza++ eng2spa</td><td>76.99</td><td>93.15</td><td>15.51</td></tr><tr><td>giza++ spa2eng</td><td>78.75</td><td>94.19</td><td>13.94</td></tr><tr><td>giza++ union</td><td>84.47</td><td>90.85</td><td>12.30</td></tr><tr><td>giza++ intersection</td><td>71.27</td><td>97.58</td><td>17.52</td></tr><tr><td colspan=\"2\">one-to-one word aligner 72.56</td><td>96.69</td><td>16.96</td></tr><tr><td>phrase aligner \u03c6 2 < 10</td><td>76.31</td><td>97.48</td><td>13.36</td></tr><tr><td>phrase aligner \u03c6 2 < 15</td><td>76.88</td><td>97.35</td><td>13.20</td></tr></table>",
"html": null,
"type_str": "table",
"text": "Comparison of final alignment results.",
"num": null
}
}
}
} |