File size: 68,357 Bytes
6fa4bc9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 |
{
"paper_id": "1996",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T02:05:56.550571Z"
},
"title": "AN EXAMPLE-BASED APPROACH TO MACHINE TRANSLATION",
"authors": [
{
"first": "Brona",
"middle": [],
"last": "Collins",
"suffix": "",
"affiliation": {},
"email": "bcollins@cs.tcd.ie"
},
{
"first": "Padraig",
"middle": [],
"last": "Cunningham",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Tony",
"middle": [],
"last": "Veale",
"suffix": "",
"affiliation": {},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "In this paper we describe a methodological analysis of EBMT (Example-Based Machine Translation) based on a CBR (Case-Based Reasoning) perspective. This analysis focuses on adaptation. We argue that, just as in CBR, the overall power of an EBMT system is its ability to adapt examples retrieved to suit the new problem translation. Here we describe a technique whereby reusability is a function of the abstract 'adaptability' information stored in the cases. This information is exploited during both the adaptation and retrieval stages.",
"pdf_parse": {
"paper_id": "1996",
"_pdf_hash": "",
"abstract": [
{
"text": "In this paper we describe a methodological analysis of EBMT (Example-Based Machine Translation) based on a CBR (Case-Based Reasoning) perspective. This analysis focuses on adaptation. We argue that, just as in CBR, the overall power of an EBMT system is its ability to adapt examples retrieved to suit the new problem translation. Here we describe a technique whereby reusability is a function of the abstract 'adaptability' information stored in the cases. This information is exploited during both the adaptation and retrieval stages.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Example Based Machine Translation (EBMT) is the marriage of MT and Case Based Reasoning (CBR) techniques (see Nagao 1984; Sumita 1995) . It involves translating the source language (SL) into the target language (TL) via remindings from previous translation cases. Because of the sparse data problem in EBMT (Gale and Church, 1993; Moonjoo 1995) an EBMT system must have a sophisticated adaptation mechanism so that the previous translation may be suitably adapted to suit the new sentence at hand. However, adaptation is complicated by translation divergences (Dorr 1993) which cause the translation example to be brittle in reuse. This means that a translation example involving a target language segment that is structurally different to the source language segment can only be adapted with great care. What is required is a retrieval policy whereby the extent to which a previous translation may be adapted is made explicit and used as a deciding factor for choosing between several possible example translations. In CBR research, such a policy is called adaptation guided retrieval (AGR). An AGR policy in EBMT is useful because it helps retrieve cases that can be adapted safely. In this paper we present an approach to EBMT that quantifies this concept of 'safety' and makes it a key criterion in case selection. We describe ReVerb, an EBMT system that implements this idea. We illustrate that ReVerb retrieves adaptable cases in the translation of software manuals from English to German.",
"cite_spans": [
{
"start": 110,
"end": 121,
"text": "Nagao 1984;",
"ref_id": "BIBREF11"
},
{
"start": 122,
"end": 134,
"text": "Sumita 1995)",
"ref_id": null
},
{
"start": 307,
"end": 330,
"text": "(Gale and Church, 1993;",
"ref_id": "BIBREF6"
},
{
"start": 331,
"end": 344,
"text": "Moonjoo 1995)",
"ref_id": "BIBREF10"
},
{
"start": 560,
"end": 571,
"text": "(Dorr 1993)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "In section 2 we provide an overview of EBMT and present our model of EBMT that emphasises the transformations in EBMT that dictate the adaptability of a translation case in a particular context. In section 3 we describe the overall architecture of ReVerb and in section 4 we describe how adaptation-guided retrieval works in ReVerb. In section 5 we illustrate how ReVerb adapts cases, followed by an evaluation in section 6. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1."
},
{
"text": "Natural language translation can be seen as a heavily memory-dependent problem, especially when the text at hand is a sublanguage (a subset of language which is specific to one domain, e.g. the language of software documentation). Professional translators often get the feeling that much of their skill is wasted on translating very repetitive passages of text. The motivation for EBMT is thus clear: when faced with a novel sentence, chances are that the expert translator will have translated something very similar before, will recognise this (matching and retrieval), and will make the changes necessary (adaptation) to produce a good translation. Approaches differ with respect to the amount of linguistic processing performed on the example texts. At one end of the spectrum are string-based approaches (Nirenburg 1995; MacLean 1993) in which examples are retrieved on the basis of string similarity only and the input is covered by combining substrings retrieved from different cases, giving preference to longer substrings. Combining substrings in this linguistically unprincipled manner inevitably leads to the problem of 'boundary friction'-a well recognised problem in EBMT literature, but the technique has the advantage that a huge data base of examples can be built easily. Other approaches incorporate shallow syntactic indexing (Juola 1994; Cranias 1995) and then use clustering techniques or simulated annealing during retrieval. At the linguistic end of the scale, language is usually represented in tree forms which typically represent the 'dependency relations' between syntactic constituents and then tree operations (e.g. deleting and replacing subtrees) are performed in order to cover the input string (Sato 1990 ). The surrounding context of the subtree is taken into account, along with the number of tree operations which the template must undergo, when choosing candidate translation examples.",
"cite_spans": [
{
"start": 809,
"end": 825,
"text": "(Nirenburg 1995;",
"ref_id": "BIBREF12"
},
{
"start": 826,
"end": 839,
"text": "MacLean 1993)",
"ref_id": null
},
{
"start": 1344,
"end": 1356,
"text": "(Juola 1994;",
"ref_id": "BIBREF8"
},
{
"start": 1357,
"end": 1370,
"text": "Cranias 1995)",
"ref_id": "BIBREF3"
},
{
"start": 1726,
"end": 1736,
"text": "(Sato 1990",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "EBMT",
"sec_num": "2."
},
{
"text": "In ReVerb, translation examples consist of aligned sentences/ clauses which are extracted from a bilingual corpus, abstracted to a dependency level of syntactic representation, and then cross-linked during the indexing stage. Linking the respective SL and TL sentence sub-parts, or 'chunks' (where a 'chunk' is a syntactic constituent consisting of one or more words, e.g., \"on the screen\" is an adverbial chunk) in this way results in a translation process which accounts for structural and lexical divergences that may have occurred during initial human translation of the corpus. Thus, the ReVerb system uses shallower linguistic processing to that of Sato and Nagao, and, instead of calculating the exact tree operations necessary to adapt the chosen case for both the source and target language at run-time, ReVerb compares flat lists of features and values and chooses the case which will unify best with the input. ReVerb's templates contain variables which indicate the chunks that can be replaced, thus increasing the likelihood of the input and examples unifying because in these positions, only the features (syntactic functions), and not the values (actual words), are required to match. In Sato's approach, the context of the original case is taken into account when inserting it into a new tree. This means that subtrees are replaced by subtrees coming from a similar syntactic and semantic environment, which is desirable. However, it does not guarantee that the resulting tree is syntactically correct in the case of the target language. Hacking a tree on the basis of subtree similarity in the source language can never be guaranteed to result in a correspondingly hacked up target language tree with its syntactic structure still intact. This is the problem which our approach seeks to address in particular.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "EBMT",
"sec_num": "2."
},
{
"text": "One way of addressing this problem is by quantifying just how 'dependent' chunks within a sentence are on other chunks, not only with respect to other chunks in the sentence (within-language dependency) but also with respect to its mapping to the target text (betweenlanguages dependency). A high dependency of either kind is penalised and receives a low adaptability score.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "EBMT",
"sec_num": "2."
},
{
"text": "In general, an EBMT system architecture will contain the following data in the example base: SL Source Language in original string form. TL Target Language in original string form. SLA Source Language Abstraction. TLA Target language Abstraction.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Mappings in EBMT",
"sec_num": "2.1"
},
{
"text": "The SL is abstracted into the SLA by means of some standard linguistic processing, likewise the TL (see Figure. 1 ). The SLA and TLA representations are then linked at the sub-sentential level. The level of linguistic abstraction will determine how substrings can be linked. We link the SL and TL at the syntactic-functional level.",
"cite_spans": [],
"ref_spans": [
{
"start": 104,
"end": 113,
"text": "Figure. 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Mappings in EBMT",
"sec_num": "2.1"
},
{
"text": "In order to define a subsentential mapping between two natural languages it is not sufficient to link two nodes which have the same syntactic feature. Translators often find it necessary to express SL words as different syntactic functions in the TL in order to create a grammatically or even stylistically correct TL. For example, the verb, say, in the SLA, may be realised as the subject in the TLA. Such divergences are the main obstacles to MT and there have been several attempts to formalise the problem (see Dorr 1993; Streiter 1995) but they inevitably suffer from the knowledge-acquisition bottleneck problem. In CBR, the idea is to compare enough previous translations so that useful generalisations fall out of the data. Therefore, one needs firstly to link two constituents on the basis of their lexical meaning and then to determine how the corresponding TLA constituent was realised syntactically compared to the SLA. Each case therefore is a description of a TLA derivation. A scoring mechanism can be used to penalise SLA:TLA mappings where the syntactic function was changed. A case base of such derivations constitutes a subset of contrastive linguistic data, which, if suitably generalised, can be regarded as an abstract model of translation for the SLA and TLA in that particular domain. More importantly for our purposes, it represents how adaptable each translation case is, for only cases with good SLA : TLA mappings can be adapted to compensate for any differences between the SL and the SL'.",
"cite_spans": [
{
"start": 515,
"end": 525,
"text": "Dorr 1993;",
"ref_id": "BIBREF5"
},
{
"start": 526,
"end": 540,
"text": "Streiter 1995)",
"ref_id": "BIBREF18"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Fig.l. Multi-level correspondences between an input sentence (SL') and a given example (SL).",
"sec_num": null
},
{
"text": "EBMT Systems which base their retrieval mechanisms on both similarity and adaptability would then differentiate between the following scenarios given a case base of examples and an input sentence representation, SL'. Here, we ignore the surface similarity (of two sentences, SL':SL) but see Section 4 for a discussion of the Base Filtering Stage.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Fig.l. Multi-level correspondences between an input sentence (SL') and a given example (SL).",
"sec_num": null
},
{
"text": "\u2022 Good SLA' : SLA, Good SLA : TLA -Retrievable and adaptable \u2022 Poor SLA' : SLA, Good SLA : TLA -Difficult to retrieve, easy to adapt \u2022 Good SLA' : SLA, Poor SLA : TLA.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Fig.l. Multi-level correspondences between an input sentence (SL') and a given example (SL).",
"sec_num": null
},
{
"text": "-Easy to retrieve but adaptation is 'unsafe' \u2022 Poor SLA' : SLA Poor SLA : TLA.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Fig.l. Multi-level correspondences between an input sentence (SL') and a given example (SL).",
"sec_num": null
},
{
"text": "-The case will not be retrieved",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Fig.l. Multi-level correspondences between an input sentence (SL') and a given example (SL).",
"sec_num": null
},
{
"text": "Imitation of the retrieved example may involve nothing more than simple word substitution or it could involve integration of standard MT techniques (Kaji 1990; Sumita,1993) . A good way to proceed is to determine where simple chunk substitution is 'safe' with respect to the above named dependencies and to perform it as much as boundary friction will allow.",
"cite_spans": [
{
"start": 148,
"end": 159,
"text": "(Kaji 1990;",
"ref_id": null
},
{
"start": 160,
"end": 172,
"text": "Sumita,1993)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Fig.l. Multi-level correspondences between an input sentence (SL') and a given example (SL).",
"sec_num": null
},
{
"text": "The overall CBR model adopted for this work is best understood by consideration of Figure 2 following (adapted from Bergmann 1995) . While providing a coherent organisation for all the necessary components of a full EBMT system, we presently concentrate on only three components of this model, namely: Case Organisation, CaseRetrieval, and Case Adaptation (described in sections 3, 4 and 5 respectively).",
"cite_spans": [
{
"start": 116,
"end": 130,
"text": "Bergmann 1995)",
"ref_id": "BIBREF0"
}
],
"ref_spans": [
{
"start": 83,
"end": 91,
"text": "Figure 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Overall Architecture",
"sec_num": "3.1"
},
{
"text": "The current case-base is derived from a corpus of bilingual English/German CorelDRAW manuals. Each case is represented in a uniquely-labelled frame which in turn contains a number of 'chunk'-frame pointers. A frame-based language KRELL is used to represent the cases. KRELL is a generic frame-management system in which specific knowledge bases may be implemented. Since a case must be retrieved from memory on the basis of its string/word content, each such word is defined in ReVerb's frame-structured memory, providing pointers to the chunks and cases in which it is employed; this situation is illustrated in Figure 3 below. The links between case frames, chunk frames, and word frames are bidirectional, with associated demon procedures to ensure consistency after updates and deletions (see Cunningham and Veale 1991) . ",
"cite_spans": [
{
"start": 797,
"end": 823,
"text": "Cunningham and Veale 1991)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [
{
"start": 613,
"end": 621,
"text": "Figure 3",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "An Associative Model of Case-Memory",
"sec_num": "3.2."
},
{
"text": "While we are against case-decomposition in general, due to the boundary friction problem it causes, in some situations it is perfectly safe (and advisable) to split a given SL sentence into its component clauses if there is a direct mapping into its TL counterpart, as in the following example taken from the CorelDRAW corpus: Such separate clauses can then be combined with the clauses arising from other sentences in the corpus without experiencing the ill effects of boundary friction, because it has already been established that the clauses in question are independent of the sentence in which they originally occurred. In our test corpus of 200 examples, about 50% of the sentences consist of two or more clauses, and of those about 10% could be split up in this manner.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Aligning the Examples",
"sec_num": "3.3."
},
{
"text": "The model of EBMT described in Section 3.1 supports different levels of abstraction. At present, a surface syntactic structure is obtained for the English component of each case by passing it through the constraint based grammar engcg of Voutilanien (1995). This grammar tags its input with morphological and categorial information, while also annotating words with the correct syntactic function. It does not completely parse the input, as syntactic functions are not used for building a hierarchical representation of the sentence. Attachment ambiguities may thus be ignored completely. However this under-specification leaves us with a problem, namely that some fertile words have two syntactic functions. Postprocessing of the grammar output is currently necessary to eliminate this ambiguity of syntactic function.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstraction",
"sec_num": "3.4."
},
{
"text": "ReVerb uses a mappability scale of 0...3 to classify the different levels of correspondence that may exist between chunks (see Fig. 4 ). A mappability of 3 is granted only if the correspondence is perfect: i.e., all the words in an SLA chunk are mapped in 1:1 fashion to the TLA. A mappability of 2 indicates there is a morphological difference in the words surrounding the head (H) of the chunk, but that the syntactic functions are the same. Morphological variation would include many-to-one mappings in either direction, inflection, case-marking, additional particles, etc.). A mappability of 1 indicates that the SLA : TLA chunks differ in syntactic function, but have a lexical correspondence which would be awarded a 3 or 2 mapping had the syntactic function been the same, while zero is reserved for chunks which exhibit no SLA:TLA commonality. Because it cannot be discerned what these 'zero' chunks are linked to in the TL, it is dangerous to make any changes to them. In general, their function in the TL is incorporated into one of the chunks of mappability 1 or 2 so as to make a dictionary-based connection impossible. Fig. 4(b) . Linking the abstracted SL and TL.",
"cite_spans": [],
"ref_spans": [
{
"start": 127,
"end": 133,
"text": "Fig. 4",
"ref_id": null
},
{
"start": 1132,
"end": 1141,
"text": "Fig. 4(b)",
"ref_id": null
}
],
"eq_spans": [],
"section": "Linking",
"sec_num": "3.5."
},
{
"text": "Generalisation of cases (see Figure 2 ) in ReVerb is achieved by a process of case templatisation, whereby a case template is generated by combining either the source text of each chunk, or a representative variable encoding the syntactic function of that chunk. Variablisation is performed by substituting SL elements of a given level of mappability/adaptability (or higher) with their part-of-speech (POS) tags; the intuition at work here is that the higher a chunk's mappability, the safer it will be to adapt at a later stage. As an example, consider that the SL string \"Move allows you to move the active window with the direction keys on the keyboard\" which is generalisable at four levels of variablisation, each corresponding to a different mappability score (0...3):",
"cite_spans": [],
"ref_spans": [
{
"start": 29,
"end": 37,
"text": "Figure 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Generalisation of Cases for Indexed Retrieval",
"sec_num": "3.5"
},
{
"text": "Level 0: SUBJ FMAINV OBJ NFMAINV OBJ ADVL ADVL Level 1: SUBJ FMAINV OBJ NFMAINV OBJ ADVL ADVL Level 2: SUBJ allows OBJ NFMAINV OBJ ADVL on the keyboard Level 3: Move allows you to move the active window with the direction keys on the keyboard",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Generalisation of Cases for Indexed Retrieval",
"sec_num": "3.5"
},
{
"text": "The success of this indexing scheme in syntax-based retrieval was tested on a case base of 214 cases at different levels of chunk variablisation. The results, shown in Figure 5 following, give an indication as to the trade-off between precision and recall. One can observe that as the mappability threshold for generalisation increases, the recall potential, or average number of retrievable cases, decreases. At the very safe mappability level of 3, therefore, a case-base of 20,000 cases would by our calculations contain 20 very adaptable templates. Fig. 5 . This graph illustrates the probability of any two cases (from a 214 case memory) matching at different levels of variablisation. As we increase the threshold for variablisation, the percentage of matches drops considerably.",
"cite_spans": [],
"ref_spans": [
{
"start": 168,
"end": 176,
"text": "Figure 5",
"ref_id": null
},
{
"start": 553,
"end": 559,
"text": "Fig. 5",
"ref_id": null
}
],
"eq_spans": [],
"section": "Generalisation of Cases for Indexed Retrieval",
"sec_num": "3.5"
},
{
"text": "ReVerb currently utilises two different levels of case retrieval, using both surface and structural criteria; our aim is to compare the nature of the cases retrieved at each level independently, given the same input, to determine if string-based retrieval can act as a useful filter for subsequent structural retrieval. This arrangement is illustrated in Figure 6 . We shall now consider each of these levels in turn:",
"cite_spans": [],
"ref_spans": [
{
"start": 355,
"end": 363,
"text": "Figure 6",
"ref_id": "FIGREF2"
}
],
"eq_spans": [],
"section": "Retrieval in ReVerb",
"sec_num": "4."
},
{
"text": "At run time, the user inputs a sentence, or clause, for which examples must be selected from the corpus. No linguistic judgements are made by the system regarding this input: only exact words are matched, and near morphological neighbours (e.g. \"object\" and \"objects\") are not considered. Frame demons will activate all cases containing exact word matches, allocating the highest scores to those cases which have been activated the greatest number of times. This brute-force technique omits all consideration of part of speech criteria, syntactic dependency, or word count of the input, but any information stored in each case (e.g., POS, linear order) can be exploited if the user so wishes. The first pass is not reliable in and of itself, but is subsequently used as a loose-grain filter to select candidates for the structural matching stage, Phase 2. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "String Matching Retrieval: Phase 1",
"sec_num": "4.1"
},
{
"text": "In the loose-grain stage, retrieval can be regarded as a process that satisfies the constraints imposed by a memory probe (i.e., the input). Marker passing is a suitable technique for retrieval at this level, which is well documented in AI literature (see Charniak 1983; Hendlerl989) , and is used here to give any retrieved case a similarity score relative to the input. Activation passing is appropriate because it propagates simple numeric values throughout memory.",
"cite_spans": [
{
"start": 256,
"end": 270,
"text": "Charniak 1983;",
"ref_id": "BIBREF1"
},
{
"start": 271,
"end": 283,
"text": "Hendlerl989)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Activation Passing for Retrieval",
"sec_num": "4.2"
},
{
"text": "For structural retrieval, the input sentence is first pre-chunked, such that each chunk has an explicit head-word. The algorithm initiates activation from each word in the chunk, giving the head word an increased weighting to reflect its pivotal role in the chunk. The summation of activation proceeds as for the purely string based retrieval method, save that adaptability/mappability scores are integrated into the calculation of activation levels.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Activation Passing for Syntactic Retrieval: Phase 2",
"sec_num": "4.3"
},
{
"text": "By adaptation, we mean here on-line, or run-time adaptation of the TLA frame of a retrieved example to accommodate any differences between the input SL and the example SL'. We restrict our present discussion to the dependency-syntactic level of case representation. No changes are made to the syntactic structure (the linear ordering of syntactic functions) during adaptation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Adaptation",
"sec_num": "5."
},
{
"text": "If an SL' and an SL sentence are extremely similar on a syntactic level and yet semantically totally different, then they will almost always have different patterns of syntactic links to the target language. ReVerb will always choose an example which can be adapted easily at a syntactic level. So, if the input sentence (unlike the straightforward example) actually requires a divergence, caused by a certain verb, say, then the resultant translation will not be correct and the user will have to correct it and it will be stored as a new case. This stored example will have a poor mappability but it will be specialised for that particular verb. Now, if a new input sentence also has this divergency-causing verb in it and the surrounding context is sufficiently similar to render the example retrievable, then the poor mappability of the example is less likely to result in a poor translation this time around.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Adaptation",
"sec_num": "5."
},
{
"text": "If a chunk has been seen before by the system it will be stored in the system's bilingual lexicon, which is simply a table of word-frames which are linked, in turn, to their host cases and thus to their TL-equivalents within those cases. In other words, no effort goes into building the lexicon as it is a by-product of ReVerb's case-memory organisation.Chunks which have a high mappability score (3 being best) are preferred by the system, as are those which have the same syntactic function as the problem word (2+), and we are currently experimenting with incorporating a chunk's context as an added contributor to the score. The new target word (i.e., in German) is then inserted into the new TLA structure (a copy of the old TLA example), and assumes the same syntactic function, linear order and syntactic category as the original. Depending on whether the mappability of the substitution is 3,2 or l, the system awards a local translation quality of 100%, 67% and 33% respectively. The translation quality of a chunk which does not need to be substituted by lexical lookup is similarly determined on the basis of its SL': TL' mappability. In this way a global translation quality can be compositionally determined for the whole sentence.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Using the Corpus as a Bilingual Dictionary",
"sec_num": "5.1"
},
{
"text": "If the system has not previously 'seen' a particular SLA chunk to be adapted, then a bilingual dictionary outside the system has to be consulted. In this case, the word will be in its root form and so the system must consult the morphological markings on the original example's TL words to modify the dictionary entry correspondingly.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Using the Corpus as a Bilingual Dictionary",
"sec_num": "5.1"
},
{
"text": "Our experiments to adapt input structures have produced encouraging results in that the output translations are quite accurate and helpful to the user. The input sentences we have used are taken from the case base itself-that is the SLA components of the cases are used to index the case base. At this stage, we are primarily interested in the correct adaptation of each structure with respect to its nearest structural matches in the case base, provided there are any. For some 74% of cases at present, however, there are no exact structural matches even when all of the chunks are variablised (level 0). This indicates how varied English language structures can be, even in the subdomain of software documentation where sentences and clauses tend to be short. Of the 24% SLAs which do retrieve structurally similar examples, the results of the matching and adaptation modules are outlined in sections 6.2 and 6.3 respectively. The basefiltering (phase 1) is discussed in section 6.1 but it is clear that it will only prove its usefulness when the case base has been scaled up.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation",
"sec_num": "6."
},
{
"text": "The first retrieval phase acts as a base-filter which extracts a group of candidate translations on the basis of string level matching. These are then input into phase two. At present, the case base is too small for the retrieved examples to be significantly similar but when the data is scaled up to, say, 20,000 examples, then this stage will become more useful.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Phase 1 Retrieval : The Base Filter",
"sec_num": "6.1"
},
{
"text": "At this level of retrieval structurally similar templates are retrieved. The threshold adaptability level with respect to the SLA' can be set to 3,2,1 or 0 depending on whether the user wishes to include templates with only high-level chunk mappings (3) or to include all templates (0). When the adaptability level is set to zero it is equivalent to similarity-guided retrieval only. At present, the system's performance at each level of adaptability is evaluated on the basis of the quality of output sentences, that is, the number of mistakes found in the output translations, as judged by a human evaluator (see Figure 8 ). The 'translation score' is ReVerb's prediction regarding translation quality and is based on its knowledge about chunk similarity and mappability ",
"cite_spans": [],
"ref_spans": [
{
"start": 615,
"end": 623,
"text": "Figure 8",
"ref_id": null
}
],
"eq_spans": [],
"section": "Phase 2 Retrieval: Structural Comparison and Adaptability",
"sec_num": "6.2"
},
{
"text": "At this highly restrictive level, only those templates are retrieved whose SLA differences are all reconcilable by adapting a chunk of mappability 3, usually by simple NP substitution ( Figure 9 ).",
"cite_spans": [],
"ref_spans": [
{
"start": 186,
"end": 195,
"text": "Figure 9",
"ref_id": null
}
],
"eq_spans": [],
"section": "Retrieval at adaptability threshold: 3",
"sec_num": "6.2.1"
},
{
"text": "Result: 6 matches in 214 cases Quality: perfect translations Fig. 9 Adaptation at Level 3 of a highly similar example.",
"cite_spans": [],
"ref_spans": [
{
"start": 61,
"end": 67,
"text": "Fig. 9",
"ref_id": null
}
],
"eq_spans": [],
"section": "Retrieval at adaptability threshold: 3",
"sec_num": "6.2.1"
},
{
"text": "At this level, one or more of the SLA chunks which need adaptation is different in word-form to the TLA, but syntactic functionality is isomorphic. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Retrieval at adaptability threshold: 2",
"sec_num": "6.2.2"
},
{
"text": "In testing our experimental system, ReVerb, we found that it can accurately adapt exemplar TLA templates when the corresponding SLA exemplar has sufficient structural similarity to the input sentence. Indeed, the CorelDRAW corpus contains several examples of such nearmappings, which are typical of computer software manuals. We have implemented the system in such a way as to be extensible in the following important respects. Firstly, we wish to bring such near-misses within the reach of the system. For instance, the input sentence may differ from a memory exemplar only in the addition (or omission) of a trailing temporal adjunct such as \"yesterday\".Redundancy rules governing the organisation of the case base can easily generate new well-formed cases from old by deleting or adding such constituents (if they have a 3 mapping between SLA and TLA), thus increasing the coverage of the case-base while avoiding an over-complication of the system's matching algorithms.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions & Future Work",
"sec_num": "7."
},
{
"text": "Secondly, we wish to refine the on-line adaptation process so that mappings of 2 and 1 will cause specialised adaptation operators to come into effect which will refer to the POS knowledge stored in the cases, or will use a derivational analysis approach to reconciling SLA: TLA differences and transmitting them to the TL. Finally, the mechanisms are already in place for semi-automatic case creation from bilingual text, so our long-term aim is to increase the number of cases until such time that the process can proceed completely automatically. This will require that all possible structures will be present in the case base, perhaps in generalised form, and also most (contextualised) words which one may expect to find in the given application domain. Our results have lead us to believe that there is an upper limit to the number of generalised templates necessary for producing perfect translations with the help of specialised adaptation specialists. On reaching this limit, a system could perform fully automatic MT. In the mean time, however, we are confident AGR can produce translation tools which assist users-who are not necessarily bilingual-in producing high quality translations of software documentation texts.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions & Future Work",
"sec_num": "7."
}
],
"back_matter": [
{
"text": "Befehl loeschen loescht markierte Objekte. TL': Drehung dreht die Zwischenformen. Human evaluation: no mistakes ReVerb Evaluation: 0.889",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "TL:",
"sec_num": null
},
{
"text": "Here, the SLA chunks that need adaptation have different corresponding syntactic functions in the TLA.Result: 50 matches in 214 cases Quality: Varies from perfect (no mistakes) in some 36% of the cases to mediocre (3 errors or less) in most others. What is very interesting here is that almost two-thirds of the cases (62%) had only one error or less, which would be easy for even an inexperienced translator to correct. Only 2% of the cases had more than 3 errors. The most common problem is the verb positioning and the incorrect form of the verb appearing in the adapted case. Another common problem is that of prepositions vanishing because an ADVL chunk containing a preposition is suddenly replaced by a SUBJ or OBJ chunk then the preposition will be overwritten by a single noun. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Retrieval at adaptability threshold: 1",
"sec_num": "6.2.3"
},
{
"text": "One or more chunks in the SLA have no lexically determinable links to the TLA. Result: 104 matches in 214 cases Quality: Words which needed adaptation were often not recorded in the system dictionary. The translation quality suffered from extra TLA' words being present which were irrelevant to the SLA'. Also, linear order and word-forms of many chunks were incorrect, especially verbs. Note that if the SL and SL' are similar both syntactically and semantically, as in the example below, the corresponding TL and TL' structures will mirror each other. The system gives a poor evaluation score because many chunks with poor mappability have been adapted. However, due to the multilevel similarity between SL and SL', the adaptation was nonetheless 'safe'. This suggests that ReVerb's evaluation technique needs to be refined to take this into account. Example: SL': Use the Zoom Tool to bring all objects into view. SL: Use the Offset Command to specify the spacing between the shapes. TL: Mit der Option Abstand legen Sie den Abstand zwischen den Formen fest. TL': Mit der Option Zoom des Hilfmittels Zoom bringen Sie alle Objekte in den sichtbaren Bereich Human Evaluation: 0 mistakes ReVerb Evaluation: 0.583",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Retrieval at adaptability threshold: 0",
"sec_num": "6.2.4"
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Building and Refining Abstract Planning Cases by change of Representation Language",
"authors": [
{
"first": "R",
"middle": [],
"last": "Bergmann",
"suffix": ""
},
{
"first": "W",
"middle": [],
"last": "Wilke",
"suffix": ""
}
],
"year": 1995,
"venue": "Journal of AI Research",
"volume": "3",
"issue": "",
"pages": "53--118",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Bergmann R. and W. Wilke. 1995. Building and Refining Abstract Planning Cases by change of Representation Language. Journal of AI Research 3, 53-118.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Passing Markers: A Theory of Contextual Influence in Language Comprehension",
"authors": [
{
"first": "E",
"middle": [],
"last": "Charniak",
"suffix": ""
}
],
"year": 1983,
"venue": "Cognitive Science",
"volume": "7",
"issue": "",
"pages": "171--190",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Charniak, E. 1983. Passing Markers: A Theory of Contextual Influence in Language Comprehension. Cognitive Science 7. 171-190.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "A Methodology for EBMT. 4th Int. Conference on the Cognitive Science of Natural Language Processing",
"authors": [
{
"first": "B",
"middle": [],
"last": "Collins",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Cunningham",
"suffix": ""
}
],
"year": 1995,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Collins, B. and P. Cunningham. 1995. A Methodology for EBMT. 4th Int. Conference on the Cognitive Science of Natural Language Processing. Dublin.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "A Matching Technique in EBMT. The Computational Linguistics Archive",
"authors": [
{
"first": "L",
"middle": [],
"last": "Cranias",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Papageorgiou",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Piperdis",
"suffix": ""
}
],
"year": 1995,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Cranias, L., H. Papageorgiou, and S. Piperdis. 1995. A Matching Technique in EBMT. The Computational Linguistics Archive cmp-lg/9508005.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Organizational issues arising from the integration of the Concept Network and Lexicon in a Text Understanding System",
"authors": [
{
"first": "P",
"middle": [],
"last": "Cunningham",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Veale",
"suffix": ""
}
],
"year": 1991,
"venue": "Proceedings of IJCAI'91",
"volume": "",
"issue": "",
"pages": "986--991",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Cunningham, P., and T. Veale. 1991. Organizational issues arising from the integration of the Concept Network and Lexicon in a Text Understanding System, Proceedings of IJCAI'91, Morgan Kaufmann, 986-991.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Machine Translation: A View from the Lexicon",
"authors": [
{
"first": "B",
"middle": [
"J"
],
"last": "Dorr",
"suffix": ""
}
],
"year": 1993,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Dorr, B.J. (ed.) 1993. Machine Translation: A View from the Lexicon. MIT Press.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "A Program for Aligning Sentences in Bilingual Corpora",
"authors": [
{
"first": "W",
"middle": [
"A"
],
"last": "Gale",
"suffix": ""
},
{
"first": "K",
"middle": [
"W"
],
"last": "Church",
"suffix": ""
}
],
"year": 1993,
"venue": "Computational Linguistics",
"volume": "19",
"issue": "1",
"pages": "75--103",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Gale, W.A. and K.W. Church. 1993. A Program for Aligning Sentences in Bilingual Corpora. Computational Linguistics 19(1), 75-103.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Marker Passing over Micro-Features: Toward a Hybrid Symbolic/ Connectionist Model",
"authors": [
{
"first": "J",
"middle": [
"A"
],
"last": "Hendler",
"suffix": ""
}
],
"year": 1989,
"venue": "Cognitive Science",
"volume": "13",
"issue": "1",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hendler, J.A. 1989. Marker Passing over Micro-Features: Toward a Hybrid Symbolic/ Connectionist Model. Cognitive Science 13(1).",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Corpus-based Acquisition of Transfer functions using Psycholinguistic Principles",
"authors": [
{
"first": "P",
"middle": [],
"last": "Juola",
"suffix": ""
}
],
"year": 1994,
"venue": "Proceeding of NeMLaP-1, New Methods in Language Processing",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Juola, P. 1994. Corpus-based Acquisition of Transfer functions using Psycholinguistic Principles. Proceeding of NeMLaP-1, New Methods in Language Processing, UMIST.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Learning Translation Templates from Bilingual Text",
"authors": [
{
"first": "H",
"middle": [],
"last": "Kaji",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Kida",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Morimoto",
"suffix": ""
}
],
"year": 1992,
"venue": "",
"volume": "",
"issue": "",
"pages": "672--678",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kaji, H., Y. Kida, and Y. Morimoto. 1992. Learning Translation Templates from Bilingual Text. COLING, Nantes. 672-678.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Collocation Map for overcoming Data Sparseness",
"authors": [
{
"first": "K",
"middle": [],
"last": "Moonjoo",
"suffix": ""
},
{
"first": "S",
"middle": [
"H"
],
"last": "Young",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Choi",
"suffix": ""
}
],
"year": 1995,
"venue": "",
"volume": "",
"issue": "",
"pages": "53--59",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Moonjoo, K., S.H. Young and K. Choi. 1995. Collocation Map for overcoming Data Sparseness. EACL 1995, Dublin. 53-59",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "A framework for mechanical translation between Japanese and English by analogy principle",
"authors": [
{
"first": "M",
"middle": [],
"last": "Nagao",
"suffix": ""
}
],
"year": 1984,
"venue": "Artificial and Human Intelligence",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nagao, M. 1984. A framework for mechanical translation between Japanese and English by analogy principle. In A. Elithorn and R. Banerji (eds.), Artificial and Human Intelligence. NATO Publications.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "The Pangloss Mark III machine translation system",
"authors": [
{
"first": "S",
"middle": [],
"last": "Nirenburg",
"suffix": ""
}
],
"year": 1995,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nirenburg, S. (ed.). 1995. The Pangloss Mark III machine translation system (Tech. Rep. No. CMU-CMT-95-145). Pittsburgh: Carnegie Mellon University, Center for Machine Translation",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Toward Memory-based Translation",
"authors": [
{
"first": "S",
"middle": [],
"last": "Sato",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Nagao",
"suffix": ""
}
],
"year": 1990,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sato, S., and M. Nagao. 1990. Toward Memory-based Translation. COLING '90.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "A translation aid system Using Flexible Text retrieval Based on Syntax Matching",
"authors": [
{
"first": "E",
"middle": [],
"last": "Sumita",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Tsutsumi",
"suffix": ""
}
],
"year": 1988,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sumita E. and H. Tsutsumi. 1988. A translation aid system Using Flexible Text retrieval Based on Syntax Matching. TRL Research Report, IBM.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Example Based Machine Translation on Massively Parallel Processors. IJCAI",
"authors": [
{
"first": "E",
"middle": [],
"last": "Sumita",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Oi",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Iida",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Higuchi",
"suffix": ""
},
{
"first": "N",
"middle": [],
"last": "Takahashi",
"suffix": ""
},
{
"first": "H",
"middle": [],
"last": "Kitano",
"suffix": ""
}
],
"year": 1993,
"venue": "",
"volume": "",
"issue": "",
"pages": "1283--1288",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sumita, E., K. Oi, H. Iida, T. Higuchi, N. Takahashi, and H. Kitano. 1993. Example Based Machine Translation on Massively Parallel Processors. IJCAI 1993 1283-1288.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Case-Based Design",
"authors": [
{
"first": "B",
"middle": [],
"last": "Smyth",
"suffix": ""
}
],
"year": 1996,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Smyth, B. 1996. Case-Based Design. PhD Thesis, Trinity College Dublin.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Retrieving Adaptable Cases:Topics in Case-Based Reasoning",
"authors": [
{
"first": "B",
"middle": [],
"last": "Smyth",
"suffix": ""
},
{
"first": "M",
"middle": [
"T"
],
"last": "Keane",
"suffix": ""
}
],
"year": 1993,
"venue": "",
"volume": "",
"issue": "",
"pages": "209--220",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Smyth, B., and M.T. Keane. 1993. Retrieving Adaptable Cases:Topics in Case-Based Reasoning, Springer-Verlag. 209-220.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Patterns of Derivation, TMI-95",
"authors": [
{
"first": "O",
"middle": [],
"last": "Streiter",
"suffix": ""
}
],
"year": 1995,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Streiter, O. 1995. Patterns of Derivation, TMI-95, Leuven, 1995.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "A syntax-based part-of-speech analyser",
"authors": [
{
"first": "A",
"middle": [],
"last": "Voutilainen",
"suffix": ""
}
],
"year": 1995,
"venue": "Proceedings of the EACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Voutilainen, A. 1995. A syntax-based part-of-speech analyser. Proceedings of the EACL, Dublin.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"num": null,
"text": "1",
"uris": null,
"type_str": "figure"
},
"FIGREF1": {
"num": null,
"text": "Concepts linked in the ReVerb translation memory.",
"uris": null,
"type_str": "figure"
},
"FIGREF2": {
"num": null,
"text": "Retrieval levels in ReVerb.",
"uris": null,
"type_str": "figure"
},
"TABREF2": {
"text": "Rotate rotates the intermediate blend objects.",
"content": "<table><tr><td>Result: 8 matches in 214 cases</td></tr><tr><td>Quality: Same as for threshold 3,but is dependent on the chunk's translation being correct in the</td></tr><tr><td>lexicon for the context of the new case.</td></tr><tr><td>Example:</td></tr><tr><td>SL':</td></tr></table>",
"num": null,
"html": null,
"type_str": "table"
}
}
}
} |