File size: 30,591 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 |
{
"paper_id": "2004",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:15:56.610769Z"
},
"title": "An EBMT System Based on Word Alignment",
"authors": [
{
"first": "Hou",
"middle": [],
"last": "Hongxu",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Chinese Academy of Sciences",
"location": {
"settlement": "Beijing",
"country": "PR China"
}
},
"email": ""
},
{
"first": "Dan",
"middle": [],
"last": "Deng",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Chinese Academy of Sciences",
"location": {
"settlement": "Beijing",
"country": "PR China"
}
},
"email": ""
},
{
"first": "Zou",
"middle": [],
"last": "Gang",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Chinese Academy of Sciences",
"location": {
"settlement": "Beijing",
"country": "PR China"
}
},
"email": ""
},
{
"first": "Y",
"middle": [
"U"
],
"last": "Hongkui",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Chinese Academy of Sciences",
"location": {
"settlement": "Beijing",
"country": "PR China"
}
},
"email": ""
},
{
"first": "Liu",
"middle": [],
"last": "Yang",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Chinese Academy of Sciences",
"location": {
"settlement": "Beijing",
"country": "PR China"
}
},
"email": ""
},
{
"first": "Xiong",
"middle": [],
"last": "Deyi",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Chinese Academy of Sciences",
"location": {
"settlement": "Beijing",
"country": "PR China"
}
},
"email": ""
},
{
"first": "",
"middle": [],
"last": "Liu Qun",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Chinese Academy of Sciences",
"location": {
"settlement": "Beijing",
"country": "PR China"
}
},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "This system is an experiment of examples based approach. It is based on a corpus containing 220 thousand sentence pairs with word alignment. The system contains four parts: matching and search, fragment matching, fragment assembling, evaluation and post processing. We use word alignment information to find and combine fragments.",
"pdf_parse": {
"paper_id": "2004",
"_pdf_hash": "",
"abstract": [
{
"text": "This system is an experiment of examples based approach. It is based on a corpus containing 220 thousand sentence pairs with word alignment. The system contains four parts: matching and search, fragment matching, fragment assembling, evaluation and post processing. We use word alignment information to find and combine fragments.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "This system is our first experiment of example based approach. It is based on corpus with word alignment. The corpus contains 220 thousands of news, literal, dictionaries, dialog sentence pairs. All sentence pairs are POS tagged and word aligned.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "The system has two parts: corpus and program. The corpus includes 220 thousand sentences pairs and a 460 thousand words and phrases dictionary. The program has four parts: matching and search, fragment matching, fragment assembling, evaluation and post processing.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "System Architecture",
"sec_num": "2."
},
{
"text": "Search for the most similar sentences from corpus.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Matching and searching",
"sec_num": "2.1."
},
{
"text": "Find out all matching and non-matching fragment of example sentence and corresponding fragment of translation of example sentence.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Fragment matching",
"sec_num": "2.2."
},
{
"text": "Assemble the fragment into a full sentence.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Fragment assembling",
"sec_num": "2.3."
},
{
"text": "Evaluation the result of translation and determine keep or discard a non-aligned part.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation",
"sec_num": "2.4."
},
{
"text": "Process spaces, cases and punctuations.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Post processing",
"sec_num": "2.5."
},
{
"text": "This system uses a corpus having 220 thousand sentences pair. It contains news, literal, dictionaries, dialog, etc. The sentences in corpus are POS tagged and word aligned.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Corpus",
"sec_num": "3."
},
{
"text": "The source language of the corpus is Chinese, and target language is English. For Chinese POS tagging, we use the ICTCLS 2.0. It's developed by ICT. It uses Multi Layer HMM. In former test of Chinese High Technology Development and Research, it got 97.58% of accuracy and could process 31.5Kb characters per second.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "POS tag",
"sec_num": "3.1."
},
{
"text": "Word alignment is the base of our system. The arithmetic of word alignment is based on dictionary. It uses large scale bilingual dictionary, Word-Net and other human-readable dictionary. It is inspired by Ker's method [4] . This method mainly depends on similarity measured by bilingual dictionary, relative distortion information and Part-of-Speech information to align words. By setting alignment window it acquires many-to-many word alignments. On a test set of 650 translation sentence pairs of Chinese and English, in which Chinese sentence has 24.8 words in average and English 34.5, the word alignment system gets a result of recall 62.9 \uff05 at the precision of 84.0%. Our algorithm is improved on Ker's in these aspects: (1) The computation of relative distortion of Ker is improved, and the initial alignment anchors chosen by dictionary-based word similarity is added to improve alignment.",
"cite_spans": [
{
"start": 218,
"end": 221,
"text": "[4]",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Word alignment",
"sec_num": "3.2."
},
{
"text": "(2) Proposed a concept of 'alignment window'. By setting alignment window in the aligning process, many-to-many word alignments can be found.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Word alignment",
"sec_num": "3.2."
},
{
"text": "The system uses a 460 thousand words and phrases bilingual dictionary. The dictionary is the base of word alignment and translation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Dictionary",
"sec_num": "3.3."
},
{
"text": "This step is searching for the most similar sentence pairs from example base. There are two problems: how to measure the similarity of two sentences and how to find out the most similar sentence from a large scale corpus.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Matching and searching",
"sec_num": "4."
},
{
"text": "The measurement of similarity determines whether or not find out the most fitted example for translation. Because our system is based on word, we must find out the longest match fragment for translation. We use follow formula to measure the similarity,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Measure of similarity",
"sec_num": "4.1."
},
{
"text": ") ( 2 * ) ( * )) ( ( i w i match i pos w m \u2211 =",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Measure of similarity",
"sec_num": "4.1."
},
{
"text": "The first item w is the weight of POS, different POS has different weight. In our system, verbs have the largest weight and stop words and named entities have the lowest weight. The second item match is 1 if the corresponding words are matched and 0 if not. The last item w2 is the measure of concatenation of words. If there are more concatenated words matched, w2 value is larger. In our system the value of w2 calculated as follow, Figure 1 give out a example of similarity, for source sentence S, there are two most similar sentences S1 and S2, and the similarity of S1 is 0.1651 and the similarity of S2 is 0.1547. For the similarity calculation is bi-direction, the largest value is 2.0 and the lowest values is 0.",
"cite_spans": [],
"ref_spans": [
{
"start": 435,
"end": 443,
"text": "Figure 1",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Measure of similarity",
"sec_num": "4.1."
},
{
"text": "This step is the most time consuming step of the system. Actually, the efficiency of the step determines the efficiency of translation. In our system, an index is created for very words appeared in corpus. So, we can find all sentences which contain certain words.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Searching",
"sec_num": "4.2."
},
{
"text": "In the system, we search the example base for each word in source sentence orderly, and the searching results are joined into a set. Through some experiments, we know some words are not helpful for finding out the most similar example and consume much long time. So, we exclude highest frequency words from searching. After searching, the most similar examples are selected as final candidate set.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Searching",
"sec_num": "4.2."
},
{
"text": "The key of EBMT system is how to split sentence into fragments and how to assemble the fragments into sentence. The familiar methods are based on parsing or word. We adopt word-based method.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Fragment matching and assembling",
"sec_num": "5."
},
{
"text": "This step is finding out all matching or non-matching fragments from source sentence and example sentence. First, we find out all matched words, and then find all mismatched words but have same part of speech. Figure 2 shows a procedure of matching. S is the source sentence, and S1 is the examples. First, the matching word string \"was washed away\" is matched (labeled as 1), then \"bridge\" (labeled as 2) is not matched, but there is a noun \"web\" is funded, so \"bridge\" is matched. In this example, there is only one matched fragment, because (1) and (2) are concatenated.",
"cite_spans": [],
"ref_spans": [
{
"start": 210,
"end": 218,
"text": "Figure 2",
"ref_id": "FIGREF2"
}
],
"eq_spans": [],
"section": "Fragment matching",
"sec_num": "5.1."
},
{
"text": "After finding out the matched fragments, we must find out their corresponding fragment in target example. The word alignment is the guild of this step. Figure 3 shows an example of word alignment. The upper sentence is source example, and lower sentence is corresponding target sentence. There are 3 pairs, (\"qiao\", bridge), (\"chongzou\", washed) and (\"chongzou\", away). As an example, we assume that \"qiao\" and \"bei chongzou le\" are different matched fragments. The first fragment \"qiao\" has one aligned pair, then \"The bridge was\" is the corresponding target fragment. The second fragment \"bei chongzou le\" has two aligned pairs, then \"was washed away\" is the corresponding target fragment. In this example, the word \"was\" is contained in both target fragments. It will be processed in later step.",
"cite_spans": [],
"ref_spans": [
{
"start": 152,
"end": 160,
"text": "Figure 3",
"ref_id": "FIGREF3"
}
],
"eq_spans": [],
"section": "Target fragment",
"sec_num": "5.2."
},
{
"text": "After finding out the target fragments, there are several matched target fragments and non-matched target fragments. For non-matched fragments, it must be searched in candidate set again for most similar examples. If can not find more example, it will be translated using dictionary. Actually, word alignment determines the positions of target fragments. Figure 4 shows an example of fragment assembling. Word matched fragment \"was washed away\" is placed in the target sentence directly, and the POS match fragment \"bridge\" is replaced by \"web\"--the translation of \"jiaopu\".",
"cite_spans": [],
"ref_spans": [
{
"start": 355,
"end": 363,
"text": "Figure 4",
"ref_id": "FIGREF4"
}
],
"eq_spans": [],
"section": "Fragment assembling",
"sec_num": "5.3."
},
{
"text": "In the last example, we could find the target sentence T contains many extra words, such as \"The\", \"was\". They may be the correct parts of the target sentence or wrong. So we must make choices. We use N-gram to determine which words will be hold and which words will be discarded. Figure 5 shows an example of N-gram. The first sentence contains a wrong word \"the\", and in the second sentence it be removed. The perplexities of two sentences are 938.48 and 858.60, clearly, the second sentence is better than the first one. Then, the second sentence will be the last translation result.",
"cite_spans": [],
"ref_spans": [
{
"start": 281,
"end": 289,
"text": "Figure 5",
"ref_id": "FIGREF5"
}
],
"eq_spans": [],
"section": "Make choices",
"sec_num": "5.4."
},
{
"text": "Because of carelessness in submitting translation result, the official result of our system is very bad. Later, we have submitted the result with correct format. .0820 These two result are generated by same corpus and engine, but the correct result processed cases, punctuations etc. We can divided the translation result into four classes, (1) Existed in corpus There are about 8% of test sentences have already existed in the corpus. These sentences are translated extremely well.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation results",
"sec_num": "6."
},
{
"text": "The test sentence is a part of example, and well aligned, so we can easily get the translation sentence from examples. The translations of these sentences are very good. The sentence is combined by several phrases that could be found in different examples. The most errors of translations are lost or add words. (6) Small fragments Theses sentences have no similar examples in corpus, so they are often translated using dictionary. The translations of these sentence are often very bad.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "(2) Good fragments of a example",
"sec_num": null
},
{
"text": "We have only completed a rudiment of an EBMT system. It has main elements of EBMT, but there are many things to be improved.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion",
"sec_num": "7."
},
{
"text": "In our system, a fragment splitting is totally blindfolded. So there are many half-balked fragments. To avoid this, we need phrase recognition to determine which positions may be divided and which position may not. Because there are many phrases in the dictionary, phrase recognition should improve the efficient of phrase in the dictionary.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phrase Recognition",
"sec_num": "7.1."
},
{
"text": "However, the accurate and recall of word alignment is still low, especially recall. So, there are large number of words need to be choose by N-gram. We must improve the arithmetic of word alignment.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Word alignment",
"sec_num": "7.2."
},
{
"text": "Word cluster will be helpful for rising the accurate of sentence matching.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Word cluster",
"sec_num": "7.3."
},
{
"text": "Actually, divided examples into different class will reduce the time consuming of searching and rise the accurate of sentence matching. We can divide examples into questions, negatives and so on.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Sentence classify",
"sec_num": "7.4."
},
{
"text": "Our corpus is not complete checked for correction and fitness. There are many repeated or out of season contents.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Corpus",
"sec_num": "7.5."
},
{
"text": "After completed this experiment system, we must improve our system. An idea is combining EBMT and SMT. We have begun some attempt of SMT and other improvement.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "8."
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "A Hybrid Rule and Example-based Method for Machine Translation",
"authors": [
{
"first": "Satoshi",
"middle": [],
"last": "Shirai",
"suffix": ""
},
{
"first": "Francis",
"middle": [],
"last": "Bond",
"suffix": ""
},
{
"first": "Yamato",
"middle": [],
"last": "Takahashi",
"suffix": ""
}
],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Satoshi Shirai, Francis Bond and Yamato Takahashi. \"A Hybrid Rule and Example-based Method for Machine Translation\".",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "A Matching Technique in Example-Based Machine Translation",
"authors": [
{
"first": "Lambros",
"middle": [],
"last": "Cranias",
"suffix": ""
},
{
"first": "Harris",
"middle": [],
"last": "Papageorgiou",
"suffix": ""
},
{
"first": "Stelios",
"middle": [],
"last": "Piperidis",
"suffix": ""
}
],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Lambros Cranias, Harris Papageorgiou, Stelios Piperidis. \"A Matching Technique in Example-Based Machine Translation\".",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Adapting an Example-Based Translation System to Chinese",
"authors": [
{
"first": "Ying",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Ralf",
"middle": [],
"last": "Brown",
"suffix": ""
},
{
"first": "Robert",
"middle": [
"E"
],
"last": "Frederking",
"suffix": ""
}
],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ying Zhang, Ralf Brown, Robert E. Frederking. \"Adapting an Example-Based Translation System to Chinese\".",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Align more words with high precision for small bilingual corpora",
"authors": [
{
"first": "Sue",
"middle": [
"J"
],
"last": "Ker",
"suffix": ""
},
{
"first": "Jason",
"middle": [
"S"
],
"last": "Chang",
"suffix": ""
}
],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sue J. Ker, and Jason S. Chang. \"Align more words with high precision for small bilingual corpora\"[J].",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Computational Linguistics and Chinese Language Processing",
"authors": [],
"year": 1997,
"venue": "",
"volume": "2",
"issue": "",
"pages": "63--96",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Computational Linguistics and Chinese Language Processing, 1997, 2(2):63-96",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"uris": null,
"text": "is the length of matching string.",
"num": null,
"type_str": "figure"
},
"FIGREF1": {
"uris": null,
"text": "An example of similarity.",
"num": null,
"type_str": "figure"
},
"FIGREF2": {
"uris": null,
"text": "An example of fragment matching.",
"num": null,
"type_str": "figure"
},
"FIGREF3": {
"uris": null,
"text": "Word alignment.",
"num": null,
"type_str": "figure"
},
"FIGREF4": {
"uris": null,
"text": "Fragment assembling.",
"num": null,
"type_str": "figure"
},
"FIGREF5": {
"uris": null,
"text": "An example of N-gram.",
"num": null,
"type_str": "figure"
},
"FIGREF6": {
"uris": null,
"text": "Replace some words These sentences are very similar to the examples, and because there are only a few words are different, the examples can be found out from corpus. The most errors of translations are the chosen of words. (4) Part of example Like (2), the sentence is the part of example, but the example is not well aligned or lost some important part. The most errors of translations are lost or add words. (5) Combine of phrases",
"num": null,
"type_str": "figure"
},
"TABREF0": {
"type_str": "table",
"num": null,
"html": null,
"text": "The score of official and correct results",
"content": "<table><tr><td/><td>Official</td><td>Correct result</td></tr><tr><td/><td>result</td><td/></tr><tr><td>BLEU</td><td>0.0798</td><td>0.2013</td></tr><tr><td>GTM</td><td>0.3862</td><td>0.6380</td></tr><tr><td>NIST</td><td>3.6443</td><td>6.4716</td></tr><tr><td>WER</td><td>0.8466</td><td>0.6275</td></tr><tr><td>PER</td><td>0.7650</td><td>0.5187</td></tr><tr><td>Fluency</td><td>2.7180</td><td/></tr><tr><td>Adequacy</td><td>3</td><td/></tr></table>"
}
}
}
} |