File size: 50,712 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 |
{
"paper_id": "2004",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:22:36.187148Z"
},
"title": "Minimum Error Training of Log-Linear Translation Models",
"authors": [
{
"first": "Mauro",
"middle": [],
"last": "Cettolo",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "ITC-irst -Centro per la Ricerca Scientifica e Tecnologica",
"location": {
"postCode": "I-38100",
"settlement": "Povo -Trento",
"country": "Italy"
}
},
"email": "cettolo@itc.it"
},
{
"first": "Marcello",
"middle": [],
"last": "Federico",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "ITC-irst -Centro per la Ricerca Scientifica e Tecnologica",
"location": {
"postCode": "I-38100",
"settlement": "Povo -Trento",
"country": "Italy"
}
},
"email": "federico@itc.it"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "Recent work on training of log-linear interpolation models for statistical machine translation reported performance improvements by optimizing parameters with respect to translation quality, rather than to likelihood oriented criteria. This work presents an alternative and more direct training procedure for log-linear interpolation models. In addition, we point out the subtle interaction between log-linear models and the beam search algorithm. Experimental results are reported on two Chinese-English evaluation sets, C-Star 2003 and Nist 2003, by using a statistical phrase-based model derived from Model 4. By optimizing parameters with respect to the BLUE score, performance relative improvements by 9.6% and 2.8% were achieved, respectively.",
"pdf_parse": {
"paper_id": "2004",
"_pdf_hash": "",
"abstract": [
{
"text": "Recent work on training of log-linear interpolation models for statistical machine translation reported performance improvements by optimizing parameters with respect to translation quality, rather than to likelihood oriented criteria. This work presents an alternative and more direct training procedure for log-linear interpolation models. In addition, we point out the subtle interaction between log-linear models and the beam search algorithm. Experimental results are reported on two Chinese-English evaluation sets, C-Star 2003 and Nist 2003, by using a statistical phrase-based model derived from Model 4. By optimizing parameters with respect to the BLUE score, performance relative improvements by 9.6% and 2.8% were achieved, respectively.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Log-linear interpolation models, which can be formally derived within the maximum entropy framework [1] , have been only recently applied to statistical machine translation (SMT) [2] . In addition, and similarly to what proposed for speech recognition [3] , optimization of interpolation parameters can directly address translation quality, rather than the usual maximum likelihood criterion [4] .",
"cite_spans": [
{
"start": 100,
"end": 103,
"text": "[1]",
"ref_id": "BIBREF0"
},
{
"start": 179,
"end": 182,
"text": "[2]",
"ref_id": "BIBREF1"
},
{
"start": 252,
"end": 255,
"text": "[3]",
"ref_id": "BIBREF2"
},
{
"start": 392,
"end": 395,
"text": "[4]",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "This paper goes along the direction of [4] , and proposes an alternative and more direct training procedure, but computationally more intensive. Moreover, a subtle relationship between the parameter optimization and the beam search algorithm is pointed out, which might have an important impact on the choice of optimal parameters.",
"cite_spans": [
{
"start": 39,
"end": 42,
"text": "[4]",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "Given a source string f and a target string e, the framework of maximum entropy [5] provides a mean to directly address the posterior probability Pr(e | f ). By introducing the hidden alignment variable a, the usual SMT optimization criterion is expressed by:",
"cite_spans": [
{
"start": 80,
"end": 83,
"text": "[5]",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Log-Linear Model for SMT",
"sec_num": "2."
},
{
"text": "e * = arg max e a",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Log-Linear Model for SMT",
"sec_num": "2."
},
{
"text": "Pr(e, a | f )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Log-Linear Model for SMT",
"sec_num": "2."
},
{
"text": "\u2248 arg max e,a",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Log-Linear Model for SMT",
"sec_num": "2."
},
{
"text": "Pr(e, a | f )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Log-Linear Model for SMT",
"sec_num": "2."
},
{
"text": "The conditional distribution Pr(e, a | f ) is determined through suitable real valued features functions h i (e, f , a), i = 1 . . . M , and takes the parametric form:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Log-Linear Model for SMT",
"sec_num": "2."
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "p \u03bb (e, a | f ) = exp{ i \u03bb i h i (e, f , a)} e,a exp{ i \u03bb i h i (e, f , a)}",
"eq_num": "(2)"
}
],
"section": "Log-Linear Model for SMT",
"sec_num": "2."
},
{
"text": "The maximum entropy criterion suggests to compute values \u03bb i , which maximize the log-likelihood over a training sample T :",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Log-Linear Model for SMT",
"sec_num": "2."
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "\u03bb * = arg max \u03bb (e,f ,a)\u2208T log p \u03bb (e, a | f )",
"eq_num": "(3)"
}
],
"section": "Log-Linear Model for SMT",
"sec_num": "2."
},
{
"text": "An interesting log-linear model results if the following feature functions derived from Model 4 [6] are used: h 1 (e, f , a) = log Pr(e) h 2 (e, f , a) = log Pr(\u03c6 | e) h 3 (e, f , a) = log Pr(\u03c4 | e, \u03c6) h 4 (e, f , a) = log Pr(\u03c0 | e, \u03c6, \u03c4 ), which explain f and a for e in terms of fertilities \u03c6, tablets \u03c4 and permutations \u03c0. In fact, after simple manipulations, the usual decoding criterion for Model 4 results, with the addition of four scaling factors: Pr(e)",
"cite_spans": [
{
"start": 96,
"end": 99,
"text": "[6]",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Log-Linear Model for SMT",
"sec_num": "2."
},
{
"text": "e * \u2248 arg max",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Log-Linear Model for SMT",
"sec_num": "2."
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "\u03bb 1 \u2022 Pr(\u03c6 | e) \u03bb 2 \u2022 Pr(\u03c4 | e, \u03c6) \u03bb 3 \u2022 Pr(\u03c0 | e, \u03c6, \u03c4 ) \u03bb 4",
"eq_num": "(5)"
}
],
"section": "Log-Linear Model for SMT",
"sec_num": "2."
},
{
"text": "To tackle the optimization problem of eq. (5), a search algorithm can be devised which incrementally extends partial translation hypotheses (\u1ebd,\u00e3) of the source string, until an optimal complete translation is found. A translation is said partial if its corresponding alignment a does not cover all positions in f . The complexity of the search algorithm mainly depends on the number of possible translations and of target positions to be considered for each source word. To avoid exponential complexity, constraints on both factors are generally introduced. Moreover, the so-called pruning of hypotheses is deployed, too. Hence, at each step (or target string length), only a beam with the most \"promising\" hypotheses is considered for extension. The following are two very popular pruning methods, which are usually applied to partial translations of the same length and/or covering the same source positions:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Log-Linear Model for SMT",
"sec_num": "2."
},
{
"text": "\u2022 threshold pruning: partial hypotheses (\u1ebd,\u00e3) whose score Q(\u2022) is smaller than the (local) optimum score Q * times a given factor T , i.e. Q(\u1ebd,\u00e3; \u03bb)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Log-Linear Model for SMT",
"sec_num": "2."
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "Q * < T ,",
"eq_num": "(6)"
}
],
"section": "Log-Linear Model for SMT",
"sec_num": "2."
},
{
"text": "are eliminated;",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Log-Linear Model for SMT",
"sec_num": "2."
},
{
"text": "\u2022 histogram pruning: hypotheses not among the top N best scoring ones are pruned.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Log-Linear Model for SMT",
"sec_num": "2."
},
{
"text": "In place of the criterion (3), [4] recently proposed to estimate parameters by minimizing the number of translation errors. We assume that a function E D (\u03bb) is available, which measures the translation errors made by running a model defined by parameter values \u03bb on a development set D. Hence, parameters are searched by:",
"cite_spans": [
{
"start": 31,
"end": 34,
"text": "[4]",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Minimum Error Training",
"sec_num": "3."
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "\u03bb * = arg min \u03bb E D (\u03bb)",
"eq_num": "(7)"
}
],
"section": "Minimum Error Training",
"sec_num": "3."
},
{
"text": "Unlike the log-likelihood criterion (3), the objective function E D (\u2022) might have many local minima. Hence, finding an optimal solution can be very hard. In this work, we use the simplex method [7] , an algorithm for multivariate function minimization which requires relatively few function evaluations. The same algorithm was already applied for the same task in [8] and for training log-linear language models in [1] .",
"cite_spans": [
{
"start": 195,
"end": 198,
"text": "[7]",
"ref_id": "BIBREF6"
},
{
"start": 365,
"end": 368,
"text": "[8]",
"ref_id": "BIBREF7"
},
{
"start": 416,
"end": 419,
"text": "[1]",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Minimum Error Training",
"sec_num": "3."
},
{
"text": "The optimization process, besides tuning the parameters of the statistical model, may also interfere with the beam search. The reason is in the following property of the scoring function (4):",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Interaction with Beam Search",
"sec_num": "3.1."
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "Q(\u1ebd,\u00e3; \u03b1\u03bb) = Q(\u1ebd,\u00e3; \u03bb) \u03b1",
"eq_num": "(8)"
}
],
"section": "Interaction with Beam Search",
"sec_num": "3.1."
},
{
"text": "for any positive real number \u03b1. As a consequence, the threshold criterion (6) is affected by any change of the parameter vector \u03bb which corresponds to a scaling transformation. For instance, a contraction of the parameter vector by a factor \u03b1 = 0.5 would implicitly determine the search to prune hypotheses according to the more relaxed constraint: Q(\u1ebd,\u00e3; \u03bb)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Interaction with Beam Search",
"sec_num": "3.1."
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "Q * < T 2",
"eq_num": "(9)"
}
],
"section": "Interaction with Beam Search",
"sec_num": "3.1."
},
{
"text": "Hence, we can expect that any optimization algorithm would be easily attracted by parameter values which relax the pruning threshold, and reduce the error rate at the expense of more computations.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Interaction with Beam Search",
"sec_num": "3.1."
},
{
"text": "To remove the impact of the pruning threshold, the simplex method is started from a parameter configuration inducing a loose threshold, so that any further widening of it does not give tangible effect on performance. A potential problem of this approach could be its high computational cost. In our implementation, the optimization remains feasible because the cost of search is also bounded by the histogram pruning. Moreover, optimization of parameters is no more influenced by the beam search, given that there is no relationship between the histogram pruning and the parameters. Another possibility could be to normalize the parameter vector (like in [2] ), or to fix one important parameter and let vary only the others.",
"cite_spans": [
{
"start": 655,
"end": 658,
"text": "[2]",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Simplex Initialization",
"sec_num": "3.2."
},
{
"text": "The core of the translation system is a statistical model, based on the IBM Model 4 and extended to deal with phrases rather than with single words [9] . The corresponding log-linear model is similar to that shown in eq. (5) with the addition of two terms, which explicitly scale the fertility and distortion probabilities of the null word. Search is performed by a decoder based on dynamic programming. Both in training and testing, sentences are pre-processed in order to reduce data sparseness. Pre-processing includes: Chinese word segmentation, separation of words from punctuation, handling of acronyms and abbreviations, number extraction, case normalization, etc.",
"cite_spans": [
{
"start": 148,
"end": 151,
"text": "[9]",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Baseline System",
"sec_num": "4.1."
},
{
"text": "In the following, we will refer to the baseline system when uniform parameters are assumed, which can be possibly scaled up or down to modify the beam-search pruning.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Baseline System",
"sec_num": "4.1."
},
{
"text": "We evaluated our approach on two Chinese-English translation tasks: the Nist 2003 MT evaluation task 1 , large-data case-insensitive conditions, and the C-Star 2003 evaluation campaign 2 . The first task concerns with translation of new agencies, while the second task concerns with basic traveling expressions [10] . Test sentences are provided with 4 and 16 human translations, respectively. Tables 1 and 2 report detailed statistics about the used training and test data. For parameter optimization, the Nist 2002 MT evaluation data and 1,000 sentences extracted from the C-Star training data were used, respectively.",
"cite_spans": [
{
"start": 311,
"end": 315,
"text": "[10]",
"ref_id": "BIBREF9"
}
],
"ref_spans": [
{
"start": 394,
"end": 408,
"text": "Tables 1 and 2",
"ref_id": "TABREF0"
}
],
"eq_spans": [],
"section": "Data",
"sec_num": "4.2."
},
{
"text": "It is worth noticing that the C-Star 2003 test set has been used as development set for the IWSLT-2004 evalu- ation campaign (see [9] ).",
"cite_spans": [
{
"start": 130,
"end": 133,
"text": "[9]",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Data",
"sec_num": "4.2."
},
{
"text": "Minimum error training was based on two well established MT objective quality measures, that adequately correlate with human subjective evaluations: namely the BLEU [11] and NIST 3 scores. Moreover, in order to compute time requirements independently from the hardware platform and CPU load, the number of hypotheses generated by the search algorithm was assumed as a reliable measure.",
"cite_spans": [
{
"start": 165,
"end": 169,
"text": "[11]",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Performance Evaluation",
"sec_num": "4.3."
},
{
"text": "Best achieved performance for each task are reported in Table 3 and Table 4 . In both cases consistent performance improvements were achieved with respect to baseline systems with comparable search complexity. On the Nist 2003 evaluation set, the BLEU score improved by 2.8% relative, while the NIST score by 1.7% relative. More significant gains were achieved on the C-Star task: 9.6% relative improvement on the BLEU score and 1.8% on the NIST score. Interestingly, for the C-Star task, a slightly better NIST score was obtained by optimizing parameters with respect to the BLEU score (see Table 4 ). In general, optimizing the NIST score results much harder than optimizing the BLEU score.",
"cite_spans": [],
"ref_spans": [
{
"start": 56,
"end": 63,
"text": "Table 3",
"ref_id": "TABREF2"
},
{
"start": 68,
"end": 75,
"text": "Table 4",
"ref_id": "TABREF3"
},
{
"start": 592,
"end": 599,
"text": "Table 4",
"ref_id": "TABREF3"
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "5."
},
{
"text": "A comparison considering many beam-search working points is shown, for the C-Star 2003 task, in Figure 1 . The two plotted curves represent performance of the baseline (using uniform parameters) and the log-linear model with parameters optimizing the BLEU score. In both cases, different working points of the search algorithm were obtained by linearly scaling the parameter 3 www.nist.gov/speech/tests/mt vector. Figure 2 is the analogous of Figure 1 , but refers to parameters optimal with respect to the NIST score. At equal computational costs, we observe performance gains between 7-10% on the BLEU score, and 1-2% on the NIST score, when optimized parameters are used instead of uniform ones. These achievements are significant, since parameter optimization on the test set allows for an improvement of 12.8% and 3.7%, respectively, which are fair approximations of the best achievable scores on that test set.",
"cite_spans": [
{
"start": 375,
"end": 376,
"text": "3",
"ref_id": "BIBREF2"
}
],
"ref_spans": [
{
"start": 96,
"end": 104,
"text": "Figure 1",
"ref_id": null
},
{
"start": 414,
"end": 422,
"text": "Figure 2",
"ref_id": null
},
{
"start": 443,
"end": 451,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "5."
},
{
"text": "In addition, we notice that best BLEU performance obtained by the baseline (best score of the baseline in Figure 1 ) is reached by the minimum error trained system with almost 50% less search cost.",
"cite_spans": [],
"ref_spans": [
{
"start": 106,
"end": 114,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Results",
"sec_num": "5."
},
{
"text": "The estimation of the translation probability Pr(e | f ) in the framework of maximum entropy was suggested by [2] . In [4] , parameter estimation of a log-linear translation model was done by optimizing the error rate instead of the likelihood. Error minimization relied on the availability of a set of n-best candidate translations for each input sentence, produced by the search algorithm. During training, optimal parameters were searched through the Powell's algorithm [7] . Since the n-best list can significantly change by modifying the parameters, the procedure is iterated until the n-best list remains stable. The author claims that, in practice, 5-7 iteration are enough for convergence.",
"cite_spans": [
{
"start": 110,
"end": 113,
"text": "[2]",
"ref_id": "BIBREF1"
},
{
"start": 119,
"end": 122,
"text": "[4]",
"ref_id": "BIBREF3"
},
{
"start": 473,
"end": 476,
"text": "[7]",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Previous Work and Discussion",
"sec_num": "6."
},
{
"text": "Unlike that in [4] , our optimization method exploits all possible solutions of the search algorithm and does not limit the search to n-best lists. As a consequence, our procedure is computationally more expensive. Each Although we measured consistent and stable improvements with two different MT scores and along many different beam-search configurations, our performance gains are significantly lower than those shown in [4] . Nevertheless, as pointed out in [8] , the simplex \"method has the advantage that it is not limited to the model scaling factors as the method described in\" [4] , but it also allows to optimize parameters of the models to be interpolated. Then, a direct comparison of the two approaches would be desirable: with this goal in mind, we are currently developing optimization methods exploiting n-best lists.",
"cite_spans": [
{
"start": 15,
"end": 18,
"text": "[4]",
"ref_id": "BIBREF3"
},
{
"start": 424,
"end": 427,
"text": "[4]",
"ref_id": "BIBREF3"
},
{
"start": 462,
"end": 465,
"text": "[8]",
"ref_id": "BIBREF7"
},
{
"start": 586,
"end": 589,
"text": "[4]",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Previous Work and Discussion",
"sec_num": "6."
},
{
"text": "www.nist.gov/speech/tests/mt 2 www.c-star.org",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "This work was partially financed by the European Commission under the project PF-STAR -Preparing Future Speech Translation Research (IST-2001-37599, http://pfstar.itc.it), and by the Province of Trento (FDR-PAT program) under the project WebFAQ.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgments",
"sec_num": "7."
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Log-linear interpolation of language models",
"authors": [
{
"first": "D",
"middle": [],
"last": "Klakow",
"suffix": ""
}
],
"year": 1998,
"venue": "Proc. of ICSLP",
"volume": "",
"issue": "",
"pages": "1695--1698",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "D. Klakow, \"Log-linear interpolation of language models,\" in Proc. of ICSLP, Sidney, Australia, 1998, pp. 1695-1698.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Discriminative training and maximum entropy models for statistical machine translation",
"authors": [
{
"first": "F",
"middle": [
"J"
],
"last": "Och",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Ney",
"suffix": ""
}
],
"year": 2002,
"venue": "Proc. of the ACL",
"volume": "",
"issue": "",
"pages": "295--302",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "F.J. Och and H. Ney, \"Discriminative training and maximum entropy models for statistical machine translation,\" in Proc. of the ACL, Philadelphia, PA, 2002, pp. 295-302.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Discriminative model combination",
"authors": [
{
"first": "P",
"middle": [],
"last": "Beyerlein",
"suffix": ""
}
],
"year": 1997,
"venue": "Proc. of IEEE ASRU",
"volume": "",
"issue": "",
"pages": "238--245",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "P. Beyerlein, \"Discriminative model combination,\" in Proc. of IEEE ASRU, S. Barbara, CA, 1997, pp. 238-245.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Minimum error rate training in statistical machine translation",
"authors": [
{
"first": "F",
"middle": [
"J"
],
"last": "Och",
"suffix": ""
}
],
"year": 2003,
"venue": "Proc. of the ACL",
"volume": "",
"issue": "",
"pages": "160--167",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "F.J. Och, \"Minimum error rate training in statistical machine translation,\" in Proc. of the ACL, Sapporo, Japan, 2003, pp. 160-167.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "A maximum entropy approach to natural language processing",
"authors": [
{
"first": "A",
"middle": [
"L"
],
"last": "Berger",
"suffix": ""
},
{
"first": "S",
"middle": [
"A"
],
"last": "Della Pietra",
"suffix": ""
},
{
"first": "V",
"middle": [
"J"
],
"last": "Della Pietra",
"suffix": ""
}
],
"year": 1996,
"venue": "Computational Linguistics",
"volume": "22",
"issue": "1",
"pages": "39--71",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "A.L. Berger, S.A. Della Pietra, and V.J. Della Pietra, \"A maximum entropy approach to natural language processing,\" Computational Linguistics, vol. 22, no. 1, pp. 39-71, 1996.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "The mathematics of statistical machine translation: parameter estimation",
"authors": [
{
"first": "P",
"middle": [
"F"
],
"last": "Brown",
"suffix": ""
},
{
"first": "S",
"middle": [
"A"
],
"last": "Della Pietra",
"suffix": ""
},
{
"first": "V",
"middle": [
"J"
],
"last": "Della Pietra",
"suffix": ""
},
{
"first": "Robert",
"middle": [
"L"
],
"last": "Mercer",
"suffix": ""
}
],
"year": 1993,
"venue": "Computational Linguistics",
"volume": "19",
"issue": "2",
"pages": "263--312",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "P.F. Brown, S.A. Della Pietra, V.J. Della Pietra, and Robert L. Mercer, \"The mathematics of statistical machine translation: parameter estimation,\" Com- putational Linguistics, vol. 19, no. 2, pp. 263-312, 1993.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Numerical recipes in C: the art of scientific computing",
"authors": [
{
"first": "W",
"middle": [
"M"
],
"last": "Press",
"suffix": ""
},
{
"first": "B",
"middle": [
"P"
],
"last": "Flannery",
"suffix": ""
},
{
"first": "S",
"middle": [
"A"
],
"last": "Teukolsky",
"suffix": ""
},
{
"first": "W",
"middle": [
"T"
],
"last": "Vetterling",
"suffix": ""
}
],
"year": 1992,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "W.M. Press, B.P. Flannery, S.A. Teukolsky, and W.T. Vetterling, Numerical recipes in C: the art of scientific computing, Cambridge Univ. Press, 1992.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Improvements in phrasebased statistical machine translation",
"authors": [
{
"first": "R",
"middle": [],
"last": "Zens",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Ney",
"suffix": ""
}
],
"year": 2004,
"venue": "Proc. of HLT/NAACL",
"volume": "",
"issue": "",
"pages": "257--264",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "R. Zens, and H. Ney, \"Improvements in phrase- based statistical machine translation\", in Proc. of HLT/NAACL, Boston, MA, 2004, pp. 257-264.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "The ITC-irst statistical machine translation system for IWSLT-2004",
"authors": [
{
"first": "N",
"middle": [],
"last": "Bertoldi",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Cattoni",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Cettolo",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Federico",
"suffix": ""
}
],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "N. Bertoldi, R. Cattoni, M. Cettolo, and M. Fed- erico, \"The ITC-irst statistical machine translation system for IWSLT-2004\", in these proceedings.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Toward a broad-coverage bilingual corpus for speech translation of travel conversations in the real world",
"authors": [
{
"first": "T",
"middle": [],
"last": "Takezawa",
"suffix": ""
},
{
"first": "E",
"middle": [],
"last": "Sumita",
"suffix": ""
},
{
"first": "F",
"middle": [],
"last": "Sugaya",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Yamamoto",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Yamamoto",
"suffix": ""
}
],
"year": 2002,
"venue": "Proc. of LREC",
"volume": "",
"issue": "",
"pages": "147--152",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "T. Takezawa, E. Sumita, F. Sugaya, H. Yamamoto, and S. Yamamoto, \"Toward a broad-coverage bilin- gual corpus for speech translation of travel conver- sations in the real world,\" in Proc. of LREC, Las Palmas, Spain, 2002, pp. 147-152.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"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.-J",
"middle": [],
"last": "Zhu",
"suffix": ""
}
],
"year": 2001,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "K. Papineni, S. Roukos, T. Ward, and W.-J. Zhu, \"Bleu: a method for automatic evaluation of ma- chine translation,\" Research Report RC22176, IBM Research Division, Thomas J. Watson Research Center, 2001.",
"links": null
}
},
"ref_entries": {
"FIGREF1": {
"uris": null,
"num": null,
"type_str": "figure",
"text": "BLEU score as function of the number of generated hypotheses. NIST score as function of the number of generated hypotheses. iteration of the simplex algorithm requires translating the full development set. Practically, each iteration on the Nist 2003 development data takes about 7 minutes with 12 CPUs, while a solution is found in about 100 steps."
},
"TABREF0": {
"type_str": "table",
"html": null,
"text": "Statistics of training data.",
"num": null,
"content": "<table><tr><td/><td/><td colspan=\"2\">Nist C-Star</td></tr><tr><td>vocabulary</td><td>source</td><td>148K</td><td>12K</td></tr><tr><td/><td>target</td><td>110K</td><td>11K</td></tr><tr><td>#words</td><td colspan=\"2\">source 13.1M</td><td>434K</td></tr><tr><td/><td>target</td><td>13.5M</td><td>45K</td></tr><tr><td>#sentence pairs</td><td/><td>687K</td><td>48K</td></tr><tr><td>#phrase pairs</td><td/><td>4.4M</td><td>381K</td></tr></table>"
},
"TABREF1": {
"type_str": "table",
"html": null,
"text": "Statistics of test data.",
"num": null,
"content": "<table><tr><td/><td colspan=\"3\">#sentences #words OOV rate</td></tr><tr><td>Nist 2003</td><td colspan=\"2\">919 27,254</td><td>1.89%</td></tr><tr><td>C-Star 2003</td><td>506</td><td>3,770</td><td>1.62%</td></tr></table>"
},
"TABREF2": {
"type_str": "table",
"html": null,
"text": "Best translation results on the Nist 2003 task with different parameter settings: optimizing BLEU, optimizing NIST, and uniform.",
"num": null,
"content": "<table><tr><td>Criterion</td><td>BLEU</td><td>NIST</td><td># hyp</td></tr><tr><td>BLEU</td><td>0.1854</td><td>7.2882</td><td>116M</td></tr><tr><td>NIST</td><td>0.1840</td><td>7.3362</td><td>115M</td></tr><tr><td>Baseline</td><td>0.1803</td><td>7.2115</td><td>116M</td></tr></table>"
},
"TABREF3": {
"type_str": "table",
"html": null,
"text": "",
"num": null,
"content": "<table><tr><td colspan=\"4\">Best translation results on the C-Star 2003 task</td></tr><tr><td colspan=\"4\">with different parameter settings: optimizing BLEU, op-</td></tr><tr><td colspan=\"2\">timizing NIST, and uniform.</td><td/><td/></tr><tr><td>Criterion</td><td>BLEU</td><td>NIST</td><td># hyp</td></tr><tr><td>BLEU</td><td>0.4614</td><td>8.4945</td><td>14.9M</td></tr><tr><td>NIST</td><td>0.4581</td><td>8.4675</td><td>14.9M</td></tr><tr><td>Baseline</td><td>0.4208</td><td>8.3169</td><td>14.8M</td></tr></table>"
}
}
}
} |