File size: 63,752 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 |
{
"paper_id": "I08-1032",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:41:34.729616Z"
},
"title": "What Prompts Translators to Modify Draft Translations? An Analysis of Basic Modification Patterns for Use in the Automatic Notification of Awkwardly Translated Text",
"authors": [
{
"first": "Takeshi",
"middle": [],
"last": "Abekawa",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Tokyo",
"location": {}
},
"email": "abekawa@p.u-tokyo.ac.jp"
},
{
"first": "Kyo",
"middle": [],
"last": "Kageura",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Tokyo",
"location": {}
},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "In human translation, translators first make draft translations and then modify them. This paper analyses these modifications, in order to identify the features that trigger modification. Our goal is to construct a system that notifies (English-to-Japanese) volunteer translators of awkward translations. After manually classifying the basic modification patterns, we analysed the factors that trigger a change in verb voice from passive to active using SVM. An experimental result shows good prospects for the automatic identification of candidates for modification.",
"pdf_parse": {
"paper_id": "I08-1032",
"_pdf_hash": "",
"abstract": [
{
"text": "In human translation, translators first make draft translations and then modify them. This paper analyses these modifications, in order to identify the features that trigger modification. Our goal is to construct a system that notifies (English-to-Japanese) volunteer translators of awkward translations. After manually classifying the basic modification patterns, we analysed the factors that trigger a change in verb voice from passive to active using SVM. An experimental result shows good prospects for the automatic identification of candidates for modification.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "We are currently developing an English-to-Japanese translation aid system aimed at volunteer translators mainly working online (Abekawa and Kageura, 2007) , As part of this project, we are developing a module that notifies (inexperienced) translators of awkwardly translated expressions that may need refinement or editing.",
"cite_spans": [
{
"start": 127,
"end": 154,
"text": "(Abekawa and Kageura, 2007)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In most cases, translators first make draft translations, and then examine and edit them later, often repeatedly. Thus there are normally at least two versions of a given translation, i.e. a draft and the final translation. In commercial translation environments, it is sometimes the case that texts are first translated by inexperienced translators and then edited by experienced translators. However, this does not apply to voluntary translation. In addition, volunteer translators tend to be less experienced than commercial translators, and devote less time to editing. It would therefore be of great help to these translators if the CAT system automatically pointed out awkward translations for possible modification. In order to realise such a system, it is necessary to first clarify (i) the basic types of modification made by translators to draft translations, and (ii) what triggers these modifications.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In section 2 we introduce the data used in this study. In section 3, we clarify the nature of modification in the translation process. In section 4, we identify the actual modification patterns in the data. In section 5, focusing on \"the change from the passive to the active voice\" pattern, we analyse and clarify the triggers that may lead to modification. Section 6 is devoted to an experiment in which machine learning methods are used to detect modification candidates. The importance of the various triggers is examined, and the performance of the system is evaluated.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The data used in the present study is the Japanese translation of an English book about the problem of peak oil (Leggett, 2005) . The book is aimed at a popular audience and is relevant to the sort of texts we have in mind, because the majority of texts volunteer translators translate deal with current affairs, social issues, politics, culture and sports, and/or economic issues for a popular audience 1 . The data consists of the English original (henceforth \"English\"), the draft Japanese translation (\"Draft\") and the final Japanese translation (\"Final\"). The \"Draft\" was made by two translators (one with two years' experience and the other with five years' experience), and 1 Software localisation is another area of translation in which volunteers are heavily involved. We do not include it in our target because it has different characteristics. Figure 1 : An example of word alignment using GIZA++ the \"Final\" was made by a translator with 12 years' experience. Table 1 gives the quantities of the data. \"English\" \"Draft\" \"Final\" As little research has been carried out into the process by which translators modify draft translations, we manually analysed a part of the data in which modifications were made, in consultation with a translator. In the modification process, the translator first recognises (though often not consciously) one of a number of states in a draft translation and the underlying cause of the state. S/he then modifies the draft translation if necessary. Table 2 shows the basic classification of states and possible causes. Although the states are conceptually clear, it is not necessarily the case that translators can judge the state of a given translation consistently, because judging a sentence as being \"natural\" or \"confusing\" is not a binary process but a graded one, and the distinction between different states is often not immedi-ately clear. Many concrete modification patterns found in the data are covered in translation textbooks (Anzai, 1995; Nakamura, 2003) . However, although it is obvious in some cases that a section of translated text needs to be modified, in other cases it is less clear, and judgments will vary according to the translator. The task that automatic notification addresses, therefore, is essentially an ambiguous one, even though the actual system output may be binary.",
"cite_spans": [
{
"start": 112,
"end": 127,
"text": "(Leggett, 2005)",
"ref_id": "BIBREF9"
},
{
"start": 1980,
"end": 1993,
"text": "(Anzai, 1995;",
"ref_id": "BIBREF1"
},
{
"start": 1994,
"end": 2009,
"text": "Nakamura, 2003)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [
{
"start": 855,
"end": 863,
"text": "Figure 1",
"ref_id": null
},
{
"start": 972,
"end": 979,
"text": "Table 1",
"ref_id": null
},
{
"start": 1489,
"end": 1496,
"text": "Table 2",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "The data",
"sec_num": "2"
},
{
"text": "We also identified the distinction between two types of modification: (i) \"generative\" modification, in which the modified translation is generated on the spot, with reference to the English original; and (ii) \"considered\" modification, in which alternate expressions (phrases, collocations, etc.) are retrieved from the depository of useful, elegant, or conventional expressions in the translator's mind. These two types of modification can be activated in the face of one token of modification at once.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The data",
"sec_num": "2"
},
{
"text": "The most natural way to classify modification patterns is by means of basic linguistic labels such as \"change of voice\" or \"change from nominal modification to adverbial modification\" (cf. Anzai, 1995) . These modification patterns consist of one or more primitive operations. For instance, a \"change of voice\" may consist of such primitive operations as \"changing the case-marker of the subject,\" \"swapping the position of subject and object,\" etc.",
"cite_spans": [
{
"start": 189,
"end": 201,
"text": "Anzai, 1995)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Modification patterns",
"sec_num": "4"
},
{
"text": "As preparation, we extracted modification patterns from the data 2 . In order to do so, we first aligned the \"Draft\" and the \"Final\" at the sentence level using DP matching, and then at the morpheme level using GIZA++ (Och and Ney, 2003) . Figure 1 illustrates an example of word/morpheme level English: If it was perceived to be true by the majority of Thinkers, ... \"Draft\": JINRUI-NO TASUU-NIYOTTE SORE-GA SINJITU-DE-ARU-TO NINSIKI-SA-RERE-BA (thinkersgenitive) (majority ablative ) (it subject ) (to be true) (be perceived) \"Final\": JINRUI-NO TASUU-GA SORE-WO SINJITU-TO NINSIKI-SURE-BA (thinkersgenitive) (majority subject ) (it object ) (to be true) (perceive) Primitive replace(\"NIYOTTE\", replace(\"GA\", delete(\"DE\") delete(\"RARERU\") operations:",
"cite_spans": [
{
"start": 218,
"end": 237,
"text": "(Och and Ney, 2003)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [
{
"start": 240,
"end": 249,
"text": "Figure 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Modification patterns",
"sec_num": "4"
},
{
"text": "\"GA\") \"WO\") delete(\"ARU\") Figure 1 , and the \"Final\" and the \"Draft\" are not completely parallel at the word or morpheme level. As a result, GIZA++ sometimes misaligns the units. From the aligned \"Draft\" and \"Final\" data, we identified the primitive operations. We limited these operations to syntactic operations and semantic operations such as the changing of content words, because the latter is hard to generalise with a small amount of data. Primitive operations were extracted by calculating the difference between corresponding bunsetsu, which basically consist of a content word and postpositions/suffixes, in the \"Draft\" and in the \"Final\". An example is given in Table 3 . Table 4 shows the five most frequent changes in verb inflections and case markers, which are two dominant classes of primitive operation. In addition, we observed deletions and insertions of Sahen verbs.",
"cite_spans": [],
"ref_spans": [
{
"start": 26,
"end": 34,
"text": "Figure 1",
"ref_id": null
},
{
"start": 673,
"end": 680,
"text": "Table 3",
"ref_id": "TABREF2"
},
{
"start": 683,
"end": 690,
"text": "Table 4",
"ref_id": "TABREF4"
}
],
"eq_spans": [],
"section": "Modification patterns",
"sec_num": "4"
},
{
"text": "Modification patterns were identified by observing the degree of co-occurrence among these primitive operations. We used Cabocha 3 to identify the syntactic dependencies and used the log-likelihood ratio (LLR) to calculate the degree of co-occurrence of primitive operations that occupy syntactically dependent positions. Three main modification patterns were identified: (i) a change from the passive to the active voice (226 cases); (ii) a change from a Sahen verb to a Sahen noun (208 cases); and (iii) a change from nominal modification to clausal structure. These patterns have been discussed in studies of paraphrases (Inui and Fujita, 2004) and in translation textbooks (Anzai, 1995; Nakamura, 2003) . We focus on \"the change from the passive to the active voice\". It is one of the most important and interesting modification patterns because (i) it is mostly concerned with the main clausal structure in which other modifications are embedded; and (ii) the use of active and passive voices differs greatly between English and Japanese and thus there will be much to reveal.",
"cite_spans": [
{
"start": 624,
"end": 647,
"text": "(Inui and Fujita, 2004)",
"ref_id": "BIBREF8"
},
{
"start": 677,
"end": 690,
"text": "(Anzai, 1995;",
"ref_id": "BIBREF1"
},
{
"start": 691,
"end": 706,
"text": "Nakamura, 2003)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Modification patterns",
"sec_num": "4"
},
{
"text": "Given a draft translation, an experienced translator will be able to recognise any problematic states in it (see Table 2 ), identify the causes of these states and deal with them. As computers (and inexperienced translators) cannot do the same (cf. Sun et al., 2007) , it is necessary to break these causes down into computationally tractable triggers. Keeping in mind the nature of the modification process discussed in section 3, we analysed the actual data, this time with the help of a translator and a linguist.",
"cite_spans": [
{
"start": 249,
"end": 266,
"text": "Sun et al., 2007)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [
{
"start": 113,
"end": 120,
"text": "Table 2",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Triggers that lead to modification",
"sec_num": "5"
},
{
"text": "At the topmost level, two types of triggers were identified: (i) \"pushing\" triggers that are identified as negative characteristics of the draft translation expressions themselves; and (ii) \"pulling\" triggers that come from outside (from the depository of expressions in the translator's mind) and work as concrete \"model translations\". The distinction is not entirely clear, because a model is needed in order to identify negative characteristics, and some sort of negative impression is needed for the \"model translation\" to be called up. Table 5 : Five of the most frequent co-occurrence patterns between two primitive operations important, both theoretically and practically. Theoretically, it corresponds to the types of modification observed in section 3. From the practical point of view, the first type is related to the general structural modelling (in its broad sense) of language, while the second is closely related to the status of individual lexicalised expressions. Correspondingly, an NLP system that addresses the first type needs to assume a language model, while a system that addresses the second type needs to call on the relevant external data on the spot. We address the first type of trigger, because we can hypothesise that the modification by change of voice is mainly related to the structural nature of expressions. It should also be noted that, from the machine learning point of view, there are positive and negative features which respectively promote and restrict the modification.",
"cite_spans": [],
"ref_spans": [
{
"start": 541,
"end": 548,
"text": "Table 5",
"ref_id": "TABREF3"
}
],
"eq_spans": [],
"section": "Triggers that lead to modification",
"sec_num": "5"
},
{
"text": "We classified the features that may represent potential triggers into five groups: (A) Features related to the readability of the English, because the complexity of English sentences (cf. Fry, 1968; Gunning, 1959) can affect the quality of draft translations. Thus the number of words in a sentence, length of words, number of verbs in a sentence, number of commas, etc. can be used as tractable features for automatic treatment. (B) Features reflecting the correspondence between the English and the draft Japanese translation. Translations that are very literal, either lexically or structurally, are often also awkward. On the other hand, a high degree of word order correspondence can be a positive sign (cf. Anzai, 1995) , because it indicates that the information flow in English is maintained and the Japanese translation is well examined. (C) Features related to the Japanese target verbs. The characteristics of the target verbs should affect the environments in which they occur. (D) Features related to the \"naturalness\" of the Japanese. Repetitions or redundancies of elements or sound patterns may lead to unnatural Japanese sentences. (E) Features related to the complexity of the Japanese. If a draft translation is too complex, it may be confusing or hard to read. Structural complexity, the length of a sentence, the number of commas, etc. can be used as triggers that reflect the complexity of the Japanese translation. Table 6 shows the computationally tractable features we defined within this framework. Features with '#' in their name are numeric features and the others are binary features (taking either 0 or 1).",
"cite_spans": [
{
"start": 188,
"end": 198,
"text": "Fry, 1968;",
"ref_id": "BIBREF5"
},
{
"start": 199,
"end": 213,
"text": "Gunning, 1959)",
"ref_id": "BIBREF6"
},
{
"start": 713,
"end": 725,
"text": "Anzai, 1995)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [
{
"start": 1438,
"end": 1445,
"text": "Table 6",
"ref_id": null
}
],
"eq_spans": [],
"section": "Triggers that lead to modification",
"sec_num": "5"
},
{
"text": "Using these features, we carried out an experiment of automatic identification of modification candidates. As a machine learning method, we used SVM (Vapnik, 1995) . The aim of the experiment was twofold: (i) to observe the feasibility of automatic notification of modification candidates, and (ii) to examine the factors that trigger modifications in more detail.",
"cite_spans": [
{
"start": 149,
"end": 163,
"text": "(Vapnik, 1995)",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Detecting modification candidates",
"sec_num": "6"
},
{
"text": "In the application of SVM, we reduced the number of binary features by using those that have higher correlations with positive and negative examples, using mutual information (MI). Table 7 shows features that have high correlations with positive and negative features (eight for each). SVM settings: The liner kernel was used. For a numeric feature X, the value x is normalized by zscore, norm(x) = x\u2212avg(X) \u221a",
"cite_spans": [],
"ref_spans": [
{
"start": 181,
"end": 188,
"text": "Table 7",
"ref_id": null
}
],
"eq_spans": [],
"section": "Experimental setup",
"sec_num": "6.1"
},
{
"text": ", where avg(x) is the empirical mean of X and var(X) is the variance of X. Data: The numbers of positive and negative cases in the data are 226 and 894, respectively (1120 in total). In order to balance the positive and negative examples, we used an equal number of examples for training. translation probability between the source and target language sentences (C) F suffix : a suffix following the target verb F particle : a particle following the target verb F pause park : a pause mark following the target verb D modifying case : case marker of the element that modifies the target verb D modifying agent : case marker of the element that modifies the target verb, if its case element has an AGENT attribute D functional : functional noun which is modified by the target verb D modified case : case marker of the element that is modified by the target verb S first agent : first case element in the sentence has an AGENT attribute S before passive : Is there a passive verb before the target verb in the sentence? S after passive : Is there a passive verb after the target verb in the sentence? (D) N modifying voice : the voice of the verb that modifies the target verb N modifying voice : the voice of the verb that is modified by the target verb N grandparent voice : the voice of the grandparent verb of the target verb N grandchild voice : the voice of the grandchild verb of the target verb N case adjacency ; bigram consists of a particle of the target verb and a particle of the adjacency bunsetsu chunk (E) J #morpheme : the number of morphemes in the target Japanese sentence J #pause : the number of pause marks in the target Japanese sentence J #verb : the number of verbs in the target Japanese sentence J #passive : the number of verbs with passive voice in the target Japanese sentence J #depth : depth of the modifier which modifies the target verb Table 6 : Features",
"cite_spans": [],
"ref_spans": [
{
"start": 1870,
"end": 1877,
"text": "Table 6",
"ref_id": null
}
],
"eq_spans": [],
"section": "var(X)",
"sec_num": null
},
{
"text": "We used (i) 10-fold cross validation to check the power of classifiers for unknown data and (ii) a partially closed test in which the 226 positive and negative examples were used for training and 1120 data were evaluated, in order to observe the realistic prospects for actual use. Table 8 shows the results. Though they are reasonable, the overall accuracy, especially for the partially closed test, shows that the method is in need of improvement. In order to evaluate the effectiveness of the feature sets, we carried out experiments only using and without using each feature set. Table 9 shows that how efficient is each feature set defined in Table 6 . The left-hand column in Table 9 shows the result with all feature sets except focal feature set, and the right-hand column shows the result when only the focal feature set was used.",
"cite_spans": [],
"ref_spans": [
{
"start": 282,
"end": 289,
"text": "Table 8",
"ref_id": "TABREF7"
},
{
"start": 584,
"end": 591,
"text": "Table 9",
"ref_id": "TABREF8"
},
{
"start": 648,
"end": 655,
"text": "Table 6",
"ref_id": null
},
{
"start": 682,
"end": 689,
"text": "Table 9",
"ref_id": "TABREF8"
}
],
"eq_spans": [],
"section": "Methods of evaluation:",
"sec_num": null
},
{
"text": "The experiment showed that the feature set that contributed most was C (features related to the Japanese target verbs). We also carried out an experiment to check which features are effective among this set, in the same manner as the experiments for checking the effectiveness of the feature sets. The result showed that the feature D modifying case is the feature that contributed the most by far. In Japanese, case markers are strongly correlated with the voice of verbs, and the coverage of this feature for tokens related to voice is high because it is common for a verb to be modified by the case element with the case marker.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Result of experiment and feature analysis",
"sec_num": "6.2"
},
{
"text": "It became clear that the numeric features A and E contribute little to the overall accuracy. Table 10 shows the correlation coefficient between the numeric features and correct answers. The table shows that there is no noticeable relation between the nu- Table 7 : Features which have high correlation with positive and negative examples meric features and the correct results. We introduced most numeric features based on the study of readability. In readability studies, however, these features are defined in terms of the overall document, and not in terms of individual sentences or of verb phrases. It would be preferable to develop numerical features that can properly reflect the nature of individual sentences or smaller constructions. Table 10 : The correlation coefficient between each feature and correct answer precision of all the feature sets. This mean that there are not many occasions on which the feature set D can be applied, but when it is applied, the result is reliable. The feature set D thus is efficient as a trigger once it is applied, and the different treatment of the tokens that contain this feature set may contribute to the performance improvement.",
"cite_spans": [],
"ref_spans": [
{
"start": 93,
"end": 101,
"text": "Table 10",
"ref_id": null
},
{
"start": 255,
"end": 262,
"text": "Table 7",
"ref_id": null
},
{
"start": 744,
"end": 752,
"text": "Table 10",
"ref_id": null
}
],
"eq_spans": [],
"section": "Result of experiment and feature analysis",
"sec_num": "6.2"
},
{
"text": "The critical cases from the point of view of improving the performance are the false positives and false negatives. We thus manually analysed the false positives and false negatives obtained in the partially closed experiment (in the actual application environment, as much training data as available should be used; we thus used the results of the partially closed experiment here). For the false positive, we extracted 100 sample sentences from 504 sentences. For the false negative we used all 33 sentences. We asked two translators to judge whether (i) it would be better to modify the draft translations or (ii) it would not be necessary to modify the draft translations.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Diagnosis",
"sec_num": "6.3"
},
{
"text": "From the 100 sample sentences, we excluded 23 cases, 18 of which were judged as in need of modification by one of the translators and 5 of which were judged as in need of modification by both of the translators. We manually analysed the remaining 77 cases. Rather than the problems with the features that we used, we identified the potential factors that would contribute to the restriction of modification. Three types of restricting factor were recognised:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "False positives",
"sec_num": "6.3.1"
},
{
"text": "1. The nature of individual verbs allows or requires the passive voice. Within the data, three subtypes were identified, i.e. (i) the use of the passive is natural irrespective of context, as in \" (consumed)\" (48 cases); (ii) the use of the passive is natural within certain fixed syntactic patterns, as in \"X Y (Y called X)\" (10 cases); and (iii) the passive is used as part of a common collocation, as in \" (attacked by anxiety)\" (2 cases);",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "False positives",
"sec_num": "6.3.1"
},
{
"text": "2. The use of the active voice is blocked by selectional restrictions, as in \" (a sediment made by ...)\" (1 case); and 3. The structure of the sentence requires the passive, as in \" (The biggest companies were all companies making cars, in which most of the oil was consumed)\" (16 cases).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "False positives",
"sec_num": "6.3.1"
},
{
"text": "Together they cover 73 cases (in 4 out of 77 cases we could not identify the factor, and in 4 of the 73 cases two of the above factors were identified). It is anticipated that the first type (60 cases; about 85%) could be dealt with by introducing \"pulling\" triggers, i.e. using large corpora to identify the characteristics of the use of voice for individual verbs, in order to enable the system to judge the desirability of given expressions vis-\u00e0-vis the conventional alternatives. To deal with the second type requires a detailed semantic description of nouns, which is difficult to achieve, though in some cases it could be approximated by collocational tendencies. In regards to the third type of false positive, we expected that the type of features used in the experiment would have been sufficient to eliminate them, but this was not the case. In fact, many of the features require discourse level information, such as the choice of subject within the flow of discourse, in order to function properly, which we did not take into account. Although high-performance discourse processing is still in an embryonic stage, in the setting of the present study the correspondence between key information in English and that in Japanese could be used to deal with this type of false positive.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "False positives",
"sec_num": "6.3.1"
},
{
"text": "Here, it is necessary to find factors that would promote modification. Among the 33 false negatives, 4 were judged as not in need of modification by both the translators. We thus examined the remaining 29 cases. In 13 cases, the verb was replaced by another verb. Including these cases, we identified four basic factors that are related to triggering modification:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "False negatives",
"sec_num": "6.3.2"
},
{
"text": "1. The nature of the individual verbs strongly requires the active voice, either independently or within the particular context, as in \" (was asked by)\" (9 cases);",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "False negatives",
"sec_num": "6.3.2"
},
{
"text": "2. The structure of the sentence is rendered rather awkward by the use of passives, as in \" (a report published in ...... by analysts)\" (4 cases); 3. A given lexical collocation is unnatural or awkward, as in \" (that all investments be screened is collectively insisted)\" (2 cases); and 4. A lexicalised collocation in the draft was subtly awkward and there is a better collocation or expression that fits the situation (14 cases).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "False negatives",
"sec_num": "6.3.2"
},
{
"text": "Together they cover 26 cases. We could not identify features in 3 cases. As in false positives, the first, second and fourth types (22 cases or about 85% are fully covered by these three types) could be dealt with by introducing \"pulling\" triggers, using large external corpora.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "False negatives",
"sec_num": "6.3.2"
},
{
"text": "For the overall data, we would expect that around 85% of 388 (77% of 504 cases) false positives (330 cases) could be dealt with by introducing \"pulling\" triggers. If these false positives could be removed completely, the precision would become well over 0.5 (193/(697-330) ) and the ratio of notified cases would become about one third ((697-330)/1120) of the total relevant cases. Though it is unreasonable to assume this ideal case, this indicates that the features we defined and introduced in this studythough limited to those related to \"pushing\" triggers -were effective, and that what we have achieved by using these features is very promising in terms of realising a system that notifies users of awkward translations.",
"cite_spans": [
{
"start": 258,
"end": 272,
"text": "(193/(697-330)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "False negatives",
"sec_num": "6.3.2"
},
{
"text": "In this paper, we examined the factors that trigger modifications when translators are revising draft translations, and identified computationally tractable features relevant to the modification. We carried out an experiment for automatic detection of modification candidates. The result was highly promising, though it revealed several issues that need to be addressed further.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "7"
},
{
"text": "Following the results reported in this paper, we are currently working on.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "7"
},
{
"text": "(i) extending the experiment by introducing outside data to carry out open experiments (we have obtained draft and final translations of three more books);",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "7"
},
{
"text": "(ii) introducing the degree of necessity for modifications by asking translators to judge the data; and (iii) further examining the features used in the experiment for the improvement of performance.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "7"
},
{
"text": "In addition, we are experimenting with a method for making use of large-scale external corpora in order to deal with \"pulling\"-type triggers, with additional features taken from large external corpora.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "7"
},
{
"text": "This task is similar to the acquisition of paraphrase knowledge(Barzilay and McKeown, 2001;Shinyama et al., 2002;Quirk et al. 2004;Barzilay and Lee, 2003;Dolan et al., 2004). However, our aim here is to clarify basic modification patterns and not automatic identification.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "This research is partly supported by grant-in-aid (A) 17200018 \"Construction of online multilingual reference tools for aiding translators\" by the Japan Society for the Promotion of Sciences (JSPS), and also by grant-in-aid from The HAKUHO FOUNDA-TION, Tokyo.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgement",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "A translation aid system with a stratified lookup interface",
"authors": [
{
"first": "T",
"middle": [],
"last": "Abekawa",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Kageura",
"suffix": ""
}
],
"year": 2007,
"venue": "Proc. of ACL 2007 Demos and Poster Sessions",
"volume": "",
"issue": "",
"pages": "5--8",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Abekawa, T. and Kageura, K. 2007. A translation aid system with a stratified lookup interface. In Proc. of ACL 2007 Demos and Poster Sessions, p. 5-8.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Eibun Hon'yaku Jutu",
"authors": [
{
"first": "T",
"middle": [],
"last": "Anzai",
"suffix": ""
}
],
"year": 1995,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Anzai, T. 1995. Eibun Hon'yaku Jutu (in Japanese). Tokyo: Chikuma.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Extracting paraphrases from a parallel corpus",
"authors": [
{
"first": "R",
"middle": [],
"last": "Barzilay",
"suffix": ""
},
{
"first": "K",
"middle": [
"R"
],
"last": "Mckeown",
"suffix": ""
}
],
"year": 2001,
"venue": "Proc. of ACL",
"volume": "",
"issue": "",
"pages": "50--57",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Barzilay, R. and McKeown, K. R. 2001. Extracting para- phrases from a parallel corpus. In Proc. of ACL 2001, p. 50-57.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Learning to paraphrase: An unsupervised approach using multiple-sequence alignment",
"authors": [
{
"first": "R",
"middle": [],
"last": "Barzilay",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Lee",
"suffix": ""
}
],
"year": 2003,
"venue": "Proc. of HLT-NAACL 2003",
"volume": "",
"issue": "",
"pages": "16--23",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Barzilay, R. and Lee, L. 2003. Learning to paraphrase: An unsupervised approach using multiple-sequence alignment. In Proc. of HLT-NAACL 2003, p. 16-23.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Unsupervised construction of large paraphrase corpora: Exploiting massively parallel news sources alignment",
"authors": [
{
"first": "B",
"middle": [],
"last": "Dolan",
"suffix": ""
}
],
"year": 2004,
"venue": "Proc. of COLING",
"volume": "",
"issue": "",
"pages": "350--356",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Dolan, B. et. al. 2004. Unsupervised construction of large paraphrase corpora: Exploiting massively paral- lel news sources alignment. In Proc. of COLING 2004, p. 350-356.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "A readability formula that saves time",
"authors": [
{
"first": "E",
"middle": [],
"last": "Fry",
"suffix": ""
}
],
"year": 1968,
"venue": "Journal of Reading",
"volume": "11",
"issue": "",
"pages": "575--578",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Fry, E. 1968. A readability formula that saves time. Journal of Reading, 11, p. 513-516, 575-578.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "The Technique of Clear Writing",
"authors": [
{
"first": "R",
"middle": [],
"last": "Gunning",
"suffix": ""
}
],
"year": 1959,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Gunning, R. 1959. The Technique of Clear Writing. New York: McGraw-Hill.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "High-performance bilingual text alignment using statistical and dictionary information",
"authors": [
{
"first": "M",
"middle": [],
"last": "Haruno",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Yamazaki",
"suffix": ""
}
],
"year": 1996,
"venue": "Proc. of ACL",
"volume": "",
"issue": "",
"pages": "131--138",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Haruno, M. and Yamazaki, T. 1996. High-performance bilingual text alignment using statistical and dictionary information. In Proc. of ACL 1996, p. 131-138.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "A survey on paraphrase generation and recognition",
"authors": [
{
"first": "K",
"middle": [],
"last": "Inui",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Fujita",
"suffix": ""
}
],
"year": 2004,
"venue": "Journal of Natural Language Processing",
"volume": "11",
"issue": "5",
"pages": "131--138",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Inui, K. and Fujita, A. 2004. A survey on paraphrase generation and recognition. Journal of Natural Lan- guage Processing, 11(5), p. 131-138.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Half Gone. London: Portobello. [Masuoka, K. et. al. trans. 2006. Peak Oil Panic",
"authors": [
{
"first": "J",
"middle": [],
"last": "Leggett",
"suffix": ""
}
],
"year": 2005,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Leggett, J. 2005. Half Gone. London: Portobello. [Ma- suoka, K. et. al. trans. 2006. Peak Oil Panic. Tokyo: Sakuhinsha.]",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Eiwa Hon'yaku no Genri Gihou",
"authors": [
{
"first": "Y",
"middle": [],
"last": "Nakamura",
"suffix": ""
}
],
"year": 2003,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nakamura, Y. 2003. Eiwa Hon'yaku no Genri Gihou (in Japanese). Tokyo: Nichigai Associates.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "A systematic comparison of various statistical alignment models",
"authors": [
{
"first": "F",
"middle": [
"J"
],
"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": "Och, F. J. and Ney, H. 2003. A systematic comparison of various statistical alignment models. Computational Linguistics, 29(1), p. 19-51.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Monolingual machine translation for paraphrase generation",
"authors": [
{
"first": "C",
"middle": [],
"last": "Quirk",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Brocktt",
"suffix": ""
},
{
"first": "W",
"middle": [
"B"
],
"last": "Dolan",
"suffix": ""
}
],
"year": 2004,
"venue": "Proc. of EMNLP",
"volume": "",
"issue": "",
"pages": "142--149",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Quirk, C., Brocktt, C. and Dolan, W. B. 2004 Monolin- gual machine translation for paraphrase generation. In Proc. of EMNLP 2004, p. 142-149.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Probabilistic part-of-speech tagging using decision trees",
"authors": [
{
"first": "H",
"middle": [],
"last": "Schmid",
"suffix": ""
}
],
"year": 1994,
"venue": "Proc. of NeMLAP",
"volume": "",
"issue": "",
"pages": "44--49",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Schmid, H. 1994. Probabilistic part-of-speech tagging using decision trees. In Proc. of NeMLAP, p. 44-49.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Automatic paraphrase acquisition from news articles",
"authors": [
{
"first": "Y",
"middle": [],
"last": "Shinyama",
"suffix": ""
}
],
"year": 2002,
"venue": "Proc. of HLT",
"volume": "",
"issue": "",
"pages": "40--46",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Shinyama, Y. et. al. 2002. Automatic paraphrase acqui- sition from news articles. In Proc. of HLT 2002, p. 40-46.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Detecting erroneous sentences using automatically mined sequential patterns",
"authors": [
{
"first": "",
"middle": [],
"last": "Sun",
"suffix": ""
}
],
"year": 2007,
"venue": "Proc. of ACL",
"volume": "",
"issue": "",
"pages": "81--88",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sun, et. al. 2007. Detecting erroneous sentences using automatically mined sequential patterns. In Proc. of ACL 2007, p. 81-88.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "The Nature of Statistical Learning Theory",
"authors": [
{
"first": "V",
"middle": [
"N"
],
"last": "Vapnik",
"suffix": ""
}
],
"year": 1995,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Vapnik, V. N. 1995. The Nature of Statistical Learning Theory. New York: Springer.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"num": null,
"text": "Cross validation 0.646 (291/452) 0.656 (138/214) 0.614 (138/226) 0.643 (153/238) 0.677 (153/226) Partially closed 0.521 (583/1120) 0.277 (193/697) 0.854 (193/226) 0.922 (390/423) 0.436 (390/894)",
"uris": null,
"type_str": "figure"
},
"TABREF1": {
"html": null,
"text": "",
"type_str": "table",
"num": null,
"content": "<table/>"
},
"TABREF2": {
"html": null,
"text": "",
"type_str": "table",
"num": null,
"content": "<table><tr><td>: An example of a primitive modification operation</td></tr><tr><td>alignment. Changes in word order occur frequently,</td></tr><tr><td>as is shown in</td></tr></table>"
},
"TABREF3": {
"html": null,
"text": "shows the top five pairwise co-occurrence patterns. inflection del. ins. case marker del. ins.",
"type_str": "table",
"num": null,
"content": "<table><tr><td>DA</td><td>379 291 NI</td><td>476 384</td></tr><tr><td>TE</td><td>269 358 GA</td><td>387 502</td></tr><tr><td>TA</td><td>247 306 NO</td><td>366 204</td></tr><tr><td colspan=\"2\">RARERU 224 122 WO</td><td>293 421</td></tr><tr><td>IRU</td><td>197 267 DE</td><td>203 193</td></tr></table>"
},
"TABREF4": {
"html": null,
"text": "Frequent primitive operations",
"type_str": "table",
"num": null,
"content": "<table/>"
},
"TABREF6": {
"html": null,
"text": "EN #word : the number of words in the English sentence EN #pause : the number of delimiters in the English sentence EN #verb : the number of verbs in the English sentence EN #VVN : the number of VNN verbs in the English sentence EN #word len : the average number of characters in a word OS : a bigram of E POS before and E POS E POS:POS after : a bigram of E POS and E POS after EJ #translation :",
"type_str": "table",
"num": null,
"content": "<table><tr><td>(A)</td><td/></tr><tr><td>(B)</td><td/></tr><tr><td>E POS :</td><td>POS of the English word corresponding to the target Japanese verb</td></tr><tr><td>E POS before :</td><td>POS of a word before the English word corresponding to the target Japanese verb</td></tr><tr><td>E POS after :</td><td>POS of a word after the English word cor-responding to the target Japanese verb</td></tr><tr><td>E P OS bef ore:P</td><td/></tr></table>"
},
"TABREF7": {
"html": null,
"text": "The accuracy of classification",
"type_str": "table",
"num": null,
"content": "<table><tr><td/><td colspan=\"2\">without this feature set</td><td colspan=\"2\">using only this feature set</td></tr><tr><td colspan=\"2\">feature set accuracy (+)precision</td><td>(+)recall</td><td>accuracy (+)precision</td><td>(+)recall</td></tr><tr><td>(A)</td><td colspan=\"4\">0.638 0.638 (144/226) 0.639 (144/226) 0.521 0.541 (62/115) 0.277 (62/226)</td></tr><tr><td>(B)</td><td colspan=\"4\">0.634 0.649 (132/203) 0.584 (132/226) 0.563 0.549 (159/290) 0.705 (159/226)</td></tr><tr><td>(C)</td><td colspan=\"4\">0.579 0.576 (136/237) 0.604 (136/226) 0.610 0.620 (128/207) 0.570 (128/226)</td></tr><tr><td>(D)</td><td colspan=\"3\">0.645 0.654 (138/212) 0.615 (138/226) 0.523 0.679 (19/29)</td><td>0.087 (19/226)</td></tr><tr><td>(E)</td><td colspan=\"4\">0.629 0.666 (117/175) 0.518 (117/226) 0.492 0.491 (101/205) 0.447 (101/226)</td></tr></table>"
},
"TABREF8": {
"html": null,
"text": "",
"type_str": "table",
"num": null,
"content": "<table><tr><td/><td colspan=\"2\">: The evaluation result for each feature set</td></tr><tr><td>feature</td><td>MI</td><td>f(+) f(-)</td></tr><tr><td colspan=\"2\">D modifying agent =NIYOTTE 0.843</td><td>15 17</td></tr><tr><td>E POS:POS after =VVN:NN</td><td>0.656</td><td>14 22</td></tr><tr><td>E POS before =IN</td><td>0.536</td><td>10 19</td></tr><tr><td>E POS before =JJ</td><td>0.530</td><td>12 23</td></tr><tr><td>D modified case =GA</td><td>0.428</td><td>13 29</td></tr><tr><td>N grandparent voice =passive</td><td>0.408</td><td>17 39</td></tr><tr><td>N grandchild voice =passive</td><td>0.368</td><td>14 34</td></tr><tr><td>E POS =VVZ</td><td>0.368</td><td>14 34</td></tr><tr><td>F suffix =NARU</td><td>0.225</td><td>0 23</td></tr><tr><td>N case adjacency =GA:TO</td><td>0.225</td><td>0 12</td></tr><tr><td>F suffix =SHIMAU</td><td>0.225</td><td>0 16</td></tr><tr><td>E POS =RB</td><td>0.225</td><td>0 10</td></tr><tr><td>E POS:POS after =VVG:DT</td><td>0.225</td><td>0 10</td></tr><tr><td>E POS:POS after =VVN:TO</td><td>0.179</td><td>2 42</td></tr><tr><td colspan=\"2\">E POS:POS after =VVN:SENT 0.159</td><td>3 44</td></tr><tr><td>D modifying agent =NI</td><td>0.154</td><td>4 54</td></tr></table>"
},
"TABREF9": {
"html": null,
"text": "",
"type_str": "table",
"num": null,
"content": "<table><tr><td>shows that the result when only using the</td></tr><tr><td>feature set D has a very low recall, but the highest</td></tr></table>"
}
}
}
} |