File size: 58,727 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 |
{
"paper_id": "I05-1018",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:25:05.861217Z"
},
"title": "Adapting a Probabilistic Disambiguation Model of an HPSG Parser to a New Domain",
"authors": [
{
"first": "Tadayoshi",
"middle": [],
"last": "Hara",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Tokyo",
"location": {
"addrLine": "Hongo 7-3-1, Bunkyo-ku",
"postCode": "113-0033",
"settlement": "Tokyo",
"country": "Japan"
}
},
"email": ""
},
{
"first": "Yusuke",
"middle": [],
"last": "Miyao",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Tokyo",
"location": {
"addrLine": "Hongo 7-3-1, Bunkyo-ku",
"postCode": "113-0033",
"settlement": "Tokyo",
"country": "Japan"
}
},
"email": ""
},
{
"first": "Jun",
"middle": [
"'"
],
"last": "Ichi Tsujii",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Tokyo",
"location": {
"addrLine": "Hongo 7-3-1, Bunkyo-ku",
"postCode": "113-0033",
"settlement": "Tokyo",
"country": "Japan"
}
},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "This paper describes a method of adapting a domain-independent HPSG parser to a biomedical domain. Without modifying the grammar and the probabilistic model of the original HPSG parser, we develop a log-linear model with additional features on a treebank of the biomedical domain. Since the treebank of the target domain is limited, we need to exploit an original disambiguation model that was trained on a larger treebank. Our model incorporates the original model as a reference probabilistic distribution. The experimental results for our model trained with a small amount of a treebank demonstrated an improvement in parsing accuracy.",
"pdf_parse": {
"paper_id": "I05-1018",
"_pdf_hash": "",
"abstract": [
{
"text": "This paper describes a method of adapting a domain-independent HPSG parser to a biomedical domain. Without modifying the grammar and the probabilistic model of the original HPSG parser, we develop a log-linear model with additional features on a treebank of the biomedical domain. Since the treebank of the target domain is limited, we need to exploit an original disambiguation model that was trained on a larger treebank. Our model incorporates the original model as a reference probabilistic distribution. The experimental results for our model trained with a small amount of a treebank demonstrated an improvement in parsing accuracy.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Natural language processing (NLP) is being demanded in various fields, such as biomedical research, patent application, and WWW, because an unmanageable amount of information is being published in unstructured data, i.e., natural language texts. To exploit latent information in these, the assistance of NLP technologies is highly required. However, an obstacle is the lack of portability of NLP tools. In general, NLP tools specialized to each domain were developed from scratch, or adapted by considerable human effort. This is because linguistic resources for each domain, such as a treebank, have not been sufficiently developed yet. Since dealing with various kinds of domains is an almost intractable job, sufficient resources can not be expected.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The method presented in this paper is the development of disambiguation models of an HPSG parser by combining a disambiguation model of an original parser with a new model adapting to a new domain. Although the training of a disambiguation model of a parser requires a sufficient amount of a treebank, its construction requires a considerable human effort. Hence, we exploit the original disambiguation model that was trained with a larger, but domain-independent treebank. Since the original disambiguation model contains rich information of general grammatical constraints, we try to use its information in developing a disambiguation model for a new domain.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Our disambiguation model is a log-linear model into which the original disambiguation model is incorporated as a reference distribution. However, we cannot simply estimate this model, because of the problem that has been discussed in studies of the probabilistic modeling of unification-based grammars [1, 2] . That is, the exponential explosion of parse candidates assigned by the grammar makes the estimation intractable. The previous studies solved the problem by applying a dynamic programming algorithm to a packed representation of parse trees. In this paper, we borrow their idea, and define reference distribution on a packed structure. With this method, the log-linear model with a reference distribution can be estimated by using dynamic programming.",
"cite_spans": [
{
"start": 302,
"end": 305,
"text": "[1,",
"ref_id": "BIBREF0"
},
{
"start": 306,
"end": 308,
"text": "2]",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In the experiments, we used an HPSG parser originally trained with the Penn Treebank [3] , and evaluated a disambiguation model trained with the GE-NIA treebank [4] , which consisted of abstracts of biomedical papers. First, we measured the accuracy of parsing and the time required for parameter estimation. For comparison, we also examined other possible models other than our disambiguation model. Next, we varied the size of a training corpus in order to evaluate the size sufficient for domain adaptation. Then, we varied feature sets used for training and examined the parsing accuracy. Finally, we compared the errors in the parsing results of our model with those of the original parser.",
"cite_spans": [
{
"start": 85,
"end": 88,
"text": "[3]",
"ref_id": "BIBREF2"
},
{
"start": 161,
"end": 164,
"text": "[4]",
"ref_id": "BIBREF3"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In Section 2, we introduce the disambiguation model of an HPSG parser. In Section 3, we describe a method of adopting reference distribution for adapting a probabilistic disambiguation model to a new domain. In Section 4, we examine our method through experiments on the GENIA treebank.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The HPSG parser used in this study is Enju [5] . The grammar of Enju was extracted from the Penn Treebank [3] , which consisted of sentences collected from The Wall Street Journal [6] . The disambiguation model of Enju was trained on the same treebank. This means that the parser has been adapted to The Wall Street Journal, and would be difficult to apply to other domains such as biomedical papers that include different distribution of words and their constraints.",
"cite_spans": [
{
"start": 43,
"end": 46,
"text": "[5]",
"ref_id": "BIBREF4"
},
{
"start": 106,
"end": 109,
"text": "[3]",
"ref_id": "BIBREF2"
},
{
"start": 180,
"end": 183,
"text": "[6]",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "An HPSG Parser",
"sec_num": "2"
},
{
"text": "In this study, we attempted the adaptation of a probabilistic disambiguation model by fixing the grammar and the disambiguation model of the original parser. The disambiguation model of Enju is based on a feature forest model [2] , which is a maximum entropy model [7] on packed forest structure. The probability, p E (t|s), of producing the parse result t for a given sentence s is defined as To avoid an exponential explosion, Enju represented T (s) in a packed form of HPSG parse trees [5] . In chart parsing, partial parse candidates are stored in a chart, in which phrasal signs are identified and packed into an equivalence class if they are determined to be equivalent and dominate the same word sequence. A set of parse trees is then represented as a set of relations among equivalence classes. Figure 1 shows a chart for parsing \"he saw a girl with a telescope\", where the modifiee (\"saw\" or \"girl\") of \"with\" is ambiguous. Each feature structure expresses an equivalence class, and the arrows represent immediate-dominance relations. The phrase, \"saw a girl with a telescope\", has two ambiguous subtrees (A in the figure). Since the signs of the top-most nodes are equivalent, they are packed into the same equivalence class. The ambiguity is represented as two pairs of arrows that come out of the node.",
"cite_spans": [
{
"start": 226,
"end": 229,
"text": "[2]",
"ref_id": "BIBREF1"
},
{
"start": 265,
"end": 268,
"text": "[7]",
"ref_id": "BIBREF6"
},
{
"start": 489,
"end": 492,
"text": "[5]",
"ref_id": "BIBREF4"
}
],
"ref_spans": [
{
"start": 803,
"end": 811,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "An HPSG Parser",
"sec_num": "2"
},
{
"text": "p E (t|s) = 1 Z s exp i \u03bb i f i (t, s) Z s = t \u2208T (s) exp i \u03bb i f i (t , s) ,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "An HPSG Parser",
"sec_num": "2"
},
{
"text": "A packed chart can be interpreted as an instance of a feature forest [2] . A feature forest represents a set of exponentially-many trees in an \"and/or\" graph of a tractable size. A feature forest is formally defined as a tuple C, D, R, \u03b3, \u03b4 , where C is a set of conjunctive nodes, D is a set of disjunctive nodes, R \u2286 C is a set of root nodes 1 , \u03b3 : D \u2192 2 C is a conjunctive daughter function, and Based on the definition, parse tree t of sentence s can be represented as the set of conjunctive nodes in the feature forest. The probability p E (t|s) is then redefined as",
"cite_spans": [
{
"start": 69,
"end": 72,
"text": "[2]",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "An HPSG Parser",
"sec_num": "2"
},
{
"text": "\u03b4 : C \u2192 2 D is",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "An HPSG Parser",
"sec_num": "2"
},
{
"text": "p E (t|s) = 1 Z s exp c\u2208t i \u03bb i f i (c) Z s = t \u2208T (s) exp c\u2208t i \u03bb i f i (c) ,",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "An HPSG Parser",
"sec_num": "2"
},
{
"text": "where f i (c) are alternative feature functions assigned to conjunctive nodes c \u2208 C. By using this redefined probability, a dynamic programming algorithm can be applied to estimate p(t|T (s)) without unpacking the packed chart [2] . Feature functions in feature forest models are designed to capture the characteristics of a conjunctive node. In HPSG parsing, it corresponds to a tuple of a mother and its daughters. Enju uses features that are combinations of the atomic features listed in Table 1 . The following combinations are used for representing the characteristics of the binary/unary rule applications. where suffix h and n means a head daughter and a non-head daughter, respectively. ",
"cite_spans": [
{
"start": 227,
"end": 230,
"text": "[2]",
"ref_id": "BIBREF1"
}
],
"ref_spans": [
{
"start": 491,
"end": 498,
"text": "Table 1",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "An HPSG Parser",
"sec_num": "2"
},
{
"text": "In addition, the following feature is used for expressing the condition of the root node of the parse tree.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Fig. 3. Example features",
"sec_num": null
},
{
"text": "f root = sym,word,pos,le Figure 3 shows example features: f root is the feature for the root node, in which the phrase symbol is S and the surface form, part-of-speech, and lexical entry of the lexical head are \"saw\", VBD, and a transitive verb, respectively. The f binary is the feature for the binary rule application to \"saw a girl\" and \"with a telescope\", in which the applied schema is the Head-Modifier Schema, the head daughter is VP headed by \"saw\", and the non-head daughter is PP headed by \"with\", whose part-of-speech is IN and the lexical entry is a VP-modifying preposition.",
"cite_spans": [],
"ref_spans": [
{
"start": 25,
"end": 33,
"text": "Figure 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Fig. 3. Example features",
"sec_num": null
},
{
"text": "The method of domain adaptation is to develop a new maximum entropy model with incorporating an original model as a reference probabilistic distribution. The idea of adaptation using a reference distribution has already been presented in several studies [8, 9] . When we have a reference probabilistic model p 0 (t|s) and are making a new model p M (t|s), the probability is defined as",
"cite_spans": [
{
"start": 254,
"end": 257,
"text": "[8,",
"ref_id": "BIBREF7"
},
{
"start": 258,
"end": 260,
"text": "9]",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Re-training of Disambiguation Models",
"sec_num": "3"
},
{
"text": "p M (t|s) = 1 Z s p 0 (t|s) exp \u239b \u239d j \u03c1 j g j (t , s) \u239e \u23a0 where Z s = t \u2208T (s) p 0 (t |s) exp \u239b \u239d j \u03c1 j g j (t , s) \u239e \u23a0 .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Re-training of Disambiguation Models",
"sec_num": "3"
},
{
"text": "Model parameters, \u03c1 j , are estimated so as to maximize the likelihood of the training data as in ordinary maximum entropy models. The maximization of the likelihood with the above model is equivalent to finding the model p M that is closest to the reference probability p 0 in terms of the Kullback-Leibler distance.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Re-training of Disambiguation Models",
"sec_num": "3"
},
{
"text": "However, we cannot simply apply the above method to our task because the parameter estimation requires the computation of the above probability for all parse candidates T (s). As discussed in Section 2, the size of T (s) is exponentially related to the length of s. This imposes a new problem, that is, we need to enumerate p 0 (t|s) for all candidate parses. Obviously, this is intractable.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Re-training of Disambiguation Models",
"sec_num": "3"
},
{
"text": "Since Enju represented a probabilistic disambiguation model in a packed forest structure, we exploit that structure to represent our probabilistic model. That is, we redefine p M with feature functions g j on conjunctive nodes as ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Re-training of Disambiguation Models",
"sec_num": "3"
},
{
"text": "p M (t|s) = 1 Z s p 0 (t|s) exp \u239b \u239d c\u2208t j \u03c1 j g j (c) \u239e \u23a0 where Z s = t \u2208T (s) p 0 (t|s) exp \u239b \u239d c\u2208t j \u03c1 j g j (c) \u239e \u23a0 .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Re-training of Disambiguation Models",
"sec_num": "3"
},
{
"text": "t 1 selected t 2 selected \u2211 j c gj j ) ( 1 \u03c1 \u2211 i c fi i ) ( 1 \u03bb \u2211 j c gj j ) ( 4 \u03c1 \u2211 j c gj j ) ( 3 \u03c1 \u2211 j c gj j ) ( 2 \u03c1 \u2211 i c fi i ) ( 2 \u03bb \u2211 i c fi i ) ( 3 \u03bb \u2211 i c fi i ) ( 4 \u03bb",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Re-training of Disambiguation Models",
"sec_num": "3"
},
{
"text": "As described in Section 2, the original model, p E (t|s), is expressed in a packed structure as",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Fig. 4. Example of importing a reference distribution into each conjunctive node",
"sec_num": null
},
{
"text": "p E (t|s) = 1 Z s exp c\u2208t i \u03bb i f i (c)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Fig. 4. Example of importing a reference distribution into each conjunctive node",
"sec_num": null
},
{
"text": "where",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Fig. 4. Example of importing a reference distribution into each conjunctive node",
"sec_num": null
},
{
"text": "Z s = t \u2208T (s) exp c\u2208t i \u03bb i f i (c) .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Fig. 4. Example of importing a reference distribution into each conjunctive node",
"sec_num": null
},
{
"text": "Then, p 0 (t|s) is substituted by p E (t|s), and p M (t|s) is formulated as",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Fig. 4. Example of importing a reference distribution into each conjunctive node",
"sec_num": null
},
{
"text": "p M (t|s) = 1 Z s 1 Z s exp c\u2208t i \u03bb i f i (c) exp \u239b \u239d c\u2208t j \u03c1 j g j (c) \u239e \u23a0 = 1 Z s \u2022 Z s exp \u239b \u239d c\u2208t i \u03bb i f i (c) + c\u2208t j \u03c1 j g j (c) \u239e \u23a0 = 1 Z s exp \u23a7 \u23a8 \u23a9 c\u2208t \u239b \u239d i \u03bb i f i (c) + j \u03c1 j g j (c) \u239e \u23a0 \u23ab \u23ac \u23ad where Z s = Z s \u2022 Z s = t\u2208T (s) exp \u23a7 \u23a8 \u23a9 c\u2208t \u239b \u239d i \u03bb i f i (c) + j \u03c1 j g j (c) \u239e \u23a0 \u23ab \u23ac \u23ad .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Fig. 4. Example of importing a reference distribution into each conjunctive node",
"sec_num": null
},
{
"text": "With this form of p M (t|s), a dynamic programing algorithm can be applied. For example, we show how to obtain probabilities of parse trees in the case of ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Fig. 4. Example of importing a reference distribution into each conjunctive node",
"sec_num": null
},
{
"text": "p M (t 1 |s)= 1 Z s exp \u23a7 \u23a8 \u23a9 \u239b \u239d i \u03bb i f i (c 1 ) + j \u03c1 j g j (c 1 ) \u239e \u23a0 + \u239b \u239d i \u03bb i f i (c 2 ) + j \u03c1 j g j (c 2 ) \u239e \u23a0 + \u239b \u239d i \u03bb i f i (c 3 ) + j \u03c1 j g j (c 3 ) \u239e \u23a0 + \u2022 \u2022 \u2022 \u23ab \u23ac \u23ad p M (t 2 |s)= 1 Z s exp \u23a7 \u23a8 \u23a9 \u239b \u239d i \u03bb i f i (c 1 ) + j \u03c1 j g j (c 1 ) \u239e \u23a0 + \u239b \u239d i \u03bb i f i (c 2 ) + j \u03c1 j g j (c 2 ) \u239e \u23a0 + \u239b \u239d i \u03bb i f i (c 4 ) + j \u03c1 j g j (c 4 ) \u239e \u23a0 + \u2022 \u2022 \u2022 \u23ab \u23ac \u23ad .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Fig. 4. Example of importing a reference distribution into each conjunctive node",
"sec_num": null
},
{
"text": "We implemented the method described in Section 3. The original parser, Enju, was developed on Section 02-21 of the Penn Treebank (39,832 sentences) [5] . For the training of our model, we used the GENIA treebank [4] , which consisted of 500 abstracts (4,446 sentences) extracted from MEDLINE. We divided the GENIA treebank into three sets of 400, 50, and 50 abstracts (3, 524, 455 , and 467 sentences), and these sets were used respectively as training, development, and final evaluation data. The method of Gaussian MAP estimation [10] was used for smoothing. The meta parameter \u03c3 of the Gaussian distribution was determined so as to maximize the accuracy on the development set. In the following experiments, we measured the accuracy of predicate-argument dependencies on the evaluation set. The measure is labeled precision/recall (LP/LR), which is the same measure as previous work [11, 5] that evaluated the accuracy of lexicalized grammars on the Penn Treebank.",
"cite_spans": [
{
"start": 148,
"end": 151,
"text": "[5]",
"ref_id": "BIBREF4"
},
{
"start": 212,
"end": 215,
"text": "[4]",
"ref_id": "BIBREF3"
},
{
"start": 368,
"end": 371,
"text": "(3,",
"ref_id": "BIBREF2"
},
{
"start": 372,
"end": 376,
"text": "524,",
"ref_id": null
},
{
"start": 377,
"end": 380,
"text": "455",
"ref_id": null
},
{
"start": 532,
"end": 536,
"text": "[10]",
"ref_id": "BIBREF9"
},
{
"start": 886,
"end": 890,
"text": "[11,",
"ref_id": "BIBREF10"
},
{
"start": 891,
"end": 893,
"text": "5]",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "4"
},
{
"text": "First, we measured the accuracy of parsing and the time required for parameter estimation. Table 2 compares the results of the following estimation methods. method gave significantly lower accuracy. We expect that the method clearly lacked the amount of the training corpus for obtaining generic grammatical information.",
"cite_spans": [],
"ref_spans": [
{
"start": 91,
"end": 98,
"text": "Table 2",
"ref_id": "TABREF3"
}
],
"eq_spans": [],
"section": "Experiments",
"sec_num": "4"
},
{
"text": "The \"Combined\" method achieved the accuracy close to our method. However, it is notable that our method took much less time for the training of the model since ours did not need to handle the Penn Treebank. Instead, our method exploited the original model of Enju, which was trained on the Penn Treebank, and this resulted in much less cost of training.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experiments",
"sec_num": "4"
},
{
"text": "Next, we changed the size of the GENIA treebank for training: 40, 80, 120, 160, 200, 240, 280, 320, 360, and 400 abstracts. Figure 5 shows the accuracy when the size of the training data was changed. We can say that, for those feature sets giving remarkable accuracy in the experiments, the accuracy edged upwards with the size of the training corpus, and the trend does not seem to converge even if more than 400 abstracts exist. If we choose more complex feature sets for higher accuracy, data sparseness will occur and an even larger corpus will be needed. These findings indicate that we can further improve the accuracy by using a larger treebank and a proper feature set. Table 3 shows the accuracy of models with only atomic feature templates. The bottom of the table gives the accuracy attained by the original parser. When we focus on the WORD features, we can see the combination of WORD h and WORD n improved the accuracy significantly, although each of the features by itself did not improve so much. DIST, SYMBOL, and POS feature templates lowered the accuracy. The other feature templates improved the accuracy, though not as well as the WORD templates. Table 4 shows that the RULE feature combined with one or more other features often gave a little higher accuracy than the RULE feature gave by itself, though not as well as the WORD features. Table 5 shows that the WORD features combined with one or more other features gave remarkable improvement to the accuracy as a whole. RULE and COMMA features gave even higher accuracy than with only the WORD features. Our results revealed that the WORD features were crucial for the adaptation to the biomedical domain. We expect that this was because the biomedical domain had a different distribution of words, while more generic grammatical constraints were not significantly different from other domains. Table 6 shows the comparison of the number of errors of our model with those of the original model in parsing the GENIA corpus. Though our model gave less errors than the original model, our model introduced a certain amount of new errors. In future work, we need to investigate manually those errors to find more suitable feature templates without losing the information in the original model.",
"cite_spans": [],
"ref_spans": [
{
"start": 124,
"end": 132,
"text": "Figure 5",
"ref_id": "FIGREF5"
},
{
"start": 678,
"end": 685,
"text": "Table 3",
"ref_id": null
},
{
"start": 1168,
"end": 1175,
"text": "Table 4",
"ref_id": null
},
{
"start": 1360,
"end": 1367,
"text": "Table 5",
"ref_id": "TABREF4"
},
{
"start": 1869,
"end": 1876,
"text": "Table 6",
"ref_id": null
}
],
"eq_spans": [],
"section": "Experiments",
"sec_num": "4"
},
{
"text": "We have presented a method of adapting a domain-independent HPSG parser to a biomedical domain. Since the treebank of the new domain was limited, we exploited an original disambiguation model. The new model was trained on a biomedical treebank, and was combined with the original model by using it as a reference distribution of a log-linear model. The experimental results demonstrated our new model was adapted to the target domain, and was superior to other adaptation methods in accuracy and the cost of training time. With our model, the parsing accuracy for the target domain improved by 1.77 point with the treebank of 3,524 sentences. Since the accuracy did not seem to saturate, we will further improve the accuracy by increasing the size of the domain-dependent treebank. In addition, the experimental results showed that the WORD feature significantly contributed to the accuracy improvement.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "5"
},
{
"text": "We examined only a few feature templates, and we must search for further more feature templates. Not only the new combinations of the atomic features but also new types of features, which may be domain-dependent such as named entities, will be possible.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusions",
"sec_num": "5"
},
{
"text": "R. Dale et al. (Eds.): IJCNLP 2005, LNAI 3651, pp. 199-210, 2005. c Springer-Verlag Berlin Heidelberg 2005",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "For the ease of explanation, the definition of root node is slightly different from the original.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "Our method: training with our method Combined: training Enju model with the training corpus replaced by the combination of the GENIA corpus and the Penn Treebank GENIA only: training Enju model with the training corpus replaced by the GENIA corpus only Original Model: training an original Enju modelThe table shows the accuracy and the parsing time for the GENIA corpus and the Penn Treebank Section 23, and also shows the time required for the training of the model. The additional feature used in our method was RULE+WORD h + WORD n , which will be explained later. In the \"Combined\" method, we could not train the model with the original training parameters (n = 20, = 0.98 in [5] ) because the estimator ran out of memory. Hence, we reduced the parameters to n = 10, = 0.95.For the GENIA corpus, our model gave the higher accuracy than the original model and the other estimation methods, while for the Penn Treebank, our model gave a little lower accuracy than the original model. This result indicates that our model was more adapted to the specific domain. The \"GENIA only\"",
"cite_spans": [
{
"start": 681,
"end": 684,
"text": "[5]",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "annex",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Dynamic programming for parsing and estimation of stochastic unification-based grammars",
"authors": [
{
"first": "S",
"middle": [],
"last": "Geman",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Johnson",
"suffix": ""
}
],
"year": 2002,
"venue": "Proc. 40th ACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Geman, S., Johnson, M.: Dynamic programming for parsing and estimation of stochastic unification-based grammars. In: Proc. 40th ACL. (2002)",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Maximum entropy estimation for feature forests",
"authors": [
{
"first": "Y",
"middle": [],
"last": "Miyao",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Tsujii",
"suffix": ""
}
],
"year": 2002,
"venue": "Proc. HLT 2002",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Miyao, Y., Tsujii, J.: Maximum entropy estimation for feature forests. In: Proc. HLT 2002. (2002)",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "The Penn Treebank: Annotating predicate argument structure",
"authors": [
{
"first": "M",
"middle": [],
"last": "Marcus",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Kim",
"suffix": ""
},
{
"first": "M",
"middle": [
"A"
],
"last": "Marcinkiewicz",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Macintyre",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Bies",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Ferguson",
"suffix": ""
},
{
"first": "K",
"middle": [],
"last": "Katz",
"suffix": ""
},
{
"first": "B",
"middle": [],
"last": "Schasberger",
"suffix": ""
}
],
"year": 1994,
"venue": "ARPA Human Language Technology Workshop",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Marcus, M., Kim, G., Marcinkiewicz, M.A., MacIntyre, R., Bies, A., Ferguson, M., Katz, K., Schasberger, B.: The Penn Treebank: Annotating predicate argument structure. In: ARPA Human Language Technology Workshop. (1994)",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Genia corpus -a semantically annotated corpus for bio-textmining",
"authors": [
{
"first": "J",
"middle": [
"D"
],
"last": "Kim",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Ohta",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Teteisi",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Tsujii",
"suffix": ""
}
],
"year": 2003,
"venue": "Bioinformatics",
"volume": "19",
"issue": "",
"pages": "180--182",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kim, J.D., Ohta, T., Teteisi, Y., Tsujii, J.: Genia corpus -a semantically annotated corpus for bio-textmining. Bioinformatics 19 (2003) i180-i182",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Probabilistic disambiguation models for wide-coverage HPSG parsing",
"authors": [
{
"first": "Y",
"middle": [],
"last": "Miyao",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Tsujii",
"suffix": ""
}
],
"year": 2005,
"venue": "Proc. ACL 2005",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Miyao, Y., Tsujii, J.: Probabilistic disambiguation models for wide-coverage HPSG parsing. In: Proc. ACL 2005. (2005)",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Corpus-oriented grammar development for acquiring a Head-driven Phrase Structure Grammar from the Penn Treebank",
"authors": [
{
"first": "Y",
"middle": [],
"last": "Miyao",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Ninomiya",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Tsujii",
"suffix": ""
}
],
"year": 2004,
"venue": "Proc. IJCNLP-04",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Miyao, Y., Ninomiya, T., Tsujii, J.: Corpus-oriented grammar development for acquiring a Head-driven Phrase Structure Grammar from the Penn Treebank. In: Proc. IJCNLP-04. (2004)",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "A maximum entropy approach to natural language processing",
"authors": [
{
"first": "A",
"middle": [
"L"
],
"last": "Berger",
"suffix": ""
},
{
"first": "S",
"middle": [
"A D"
],
"last": "Pietra",
"suffix": ""
},
{
"first": "V",
"middle": [
"J D"
],
"last": "Pietra",
"suffix": ""
}
],
"year": 1996,
"venue": "Computational Linguistics",
"volume": "22",
"issue": "",
"pages": "39--71",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Berger, A.L., Pietra, S.A.D., Pietra, V.J.D.: A maximum entropy approach to natural language processing. Computational Linguistics 22 (1996) 39-71",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Statistical Methods for Speech Recognition",
"authors": [
{
"first": "F",
"middle": [],
"last": "Jelinek",
"suffix": ""
}
],
"year": 1998,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jelinek, F.: Statistical Methods for Speech Recognition. The MIT Press (1998)",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Exploiting auxiliary distributions in stochastic unificationbased grammars",
"authors": [
{
"first": "M",
"middle": [],
"last": "Johnson",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Riezler",
"suffix": ""
}
],
"year": 2000,
"venue": "Proc. 1st NAACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Johnson, M., Riezler, S.: Exploiting auxiliary distributions in stochastic unification- based grammars. In: Proc. 1st NAACL. (2000)",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "A gaussian prior for smoothing maximum entropy models",
"authors": [
{
"first": "S",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Rosenfeld",
"suffix": ""
}
],
"year": 1999,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Chen, S., Rosenfeld, R.: A gaussian prior for smoothing maximum entropy models. Technical Report CMUCS-99-108, Carnegie Mellon University (1999)",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Parsing the WSJ using CCG and log-linear models",
"authors": [
{
"first": "S",
"middle": [],
"last": "Clark",
"suffix": ""
},
{
"first": "J",
"middle": [
"R"
],
"last": "Curran",
"suffix": ""
}
],
"year": 2004,
"venue": "Proc. 42nd ACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Clark, S., Curran, J.R.: Parsing the WSJ using CCG and log-linear models. In: Proc. 42nd ACL. (2004)",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"type_str": "figure",
"text": "Chart for parsing \"he saw a girl with a telescope\"where T (s) is the set of parse candidates assigned to s. The feature function f i (t, s) represents the characteristics of t and s, while the corresponding model parameter \u03bb i is its weight. Model parameters were estimated so as to maximize the log-likelihood of the training data.Estimation of the above model requires a set of training pairs t s , T (s) , where t s is the correct parse for the sentence s. While t s is provided by a treebank, T (s) is computed by parsing each s in the treebank. However, the simple enumeration of T (s) is impractical because the size of T (s) is exponential to the length of s.",
"num": null,
"uris": null
},
"FIGREF1": {
"type_str": "figure",
"text": "Packed representation of HPSG parse trees inFigure 1",
"num": null,
"uris": null
},
"FIGREF2": {
"type_str": "figure",
"text": "shows (a part of) the HPSG parse trees inFigure 1represented as a feature forest. Square boxes are conjunctive nodes, dotted lines express a disjunctive daughter function, and solid arrows represent a conjunctive daughter function.",
"num": null,
"uris": null
},
"FIGREF3": {
"type_str": "figure",
"text": "f binary = rule,dist,comma, span h , sym h , word h , pos h , le h , spann, symn, wordn, posn, len funary = rule,sym,word,pos,le",
"num": null,
"uris": null
},
"FIGREF4": {
"type_str": "figure",
"text": "For ease, we assume that there are only two disjunctive daughters (dotted lines) that are of the top conjunctive node. The left disjunctive node introduces a parse tree t 1 that consists of conjunctive nodes {c 1 , c 2 , c 3 , . . . }, and the right one, t 2 that consists of {c 1 , c 2 , c 4 , . . . }. To each conjunctive node c k , a weight from the reference distribution i \u03bb i f i (c k ) is assigned. Probability p M (t 1 |s) and p M (t 2 |s) are then given as",
"num": null,
"uris": null
},
"FIGREF5": {
"type_str": "figure",
"text": "Corpus size vs. Accuracy",
"num": null,
"uris": null
},
"TABREF0": {
"type_str": "table",
"text": "a disjunctive daughter function.",
"num": null,
"html": null,
"content": "<table><tr><td/><td/><td/><td>c 1</td><td>HEAD verb</td></tr><tr><td/><td/><td/><td/><td>SUBCAT <></td></tr><tr><td/><td/><td/><td colspan=\"2\">HEAD noun</td><td>HEAD verb</td></tr><tr><td/><td/><td/><td colspan=\"2\">SUBCAT <></td><td>SUBCAT <NP></td></tr><tr><td/><td>c 3</td><td colspan=\"2\">HEAD verb SUBCAT <NP></td><td>c 4</td><td>SUBCAT <NP> HEAD verb</td></tr><tr><td/><td colspan=\"2\">HEAD verb SUBCAT <NP></td><td colspan=\"2\">HEAD prep MOD VP SUBCAT <></td><td>HEAD verb SUBCAT <NP,NP></td><td>HEAD noun SUBCAT <></td></tr><tr><td colspan=\"2\">he SUBCAT <> HEAD noun c 2</td><td colspan=\"3\">HEAD prep MOD VP SUBCAT <NP> HEAD prep MOD VP HEAD noun SUBCAT <> SUBCAT <> c 5</td><td>HEAD prep MOD NP SUBCAT <> HEAD noun SUBCAT <> HEAD noun SUBCAT <> c 6</td></tr><tr><td>c 7</td><td colspan=\"2\">HEAD verb</td><td/><td>c 8</td></tr><tr><td colspan=\"4\">SUBCAT <NP> HEAD verb SUBCAT <NP,NP> HEAD noun SUBCAT <></td><td>HEAD verb SUBCAT <NP,NP> saw</td></tr></table>"
},
"TABREF1": {
"type_str": "table",
"text": "Templates of atomic features rule the name of the applied schema dist the distance between the head words of the daughters comma whether a comma exists between daughters and/or inside of daughter phrases span the number of words dominated by the phrase sym the symbol of the phrasal category (e.g. NP, VP) word the surface form of the head word pos the part-of-speech of the head word le the lexical entry assigned to the head word",
"num": null,
"html": null,
"content": "<table><tr><td/><td/><td/><td>f</td><td>root =</td><td>S,</td><td>saw,</td><td>VBD,</td><td>transitive</td></tr><tr><td/><td>c 1</td><td colspan=\"2\">HEAD verb</td><td/><td/></tr><tr><td/><td/><td colspan=\"3\">SUBCAT <></td><td/></tr><tr><td/><td colspan=\"2\">HEAD noun</td><td colspan=\"3\">HEAD verb</td></tr><tr><td/><td colspan=\"2\">SUBCAT <></td><td colspan=\"4\">SUBCAT <NP></td></tr><tr><td>he SUBCAT <> HEAD noun c 2</td><td colspan=\"4\">HEAD verb SUBCAT <NP> HEAD verb SUBCAT <> SUBCAT <NP> HEAD prep MOD VP c 3</td><td colspan=\"2\">HEAD verb SUBCAT <NP,NP> HEAD verb HEAD noun SUBCAT <> SUBCAT <NP> c 4</td></tr><tr><td/><td>f</td><td>binary =</td><td colspan=\"4\">vp VBD, with,IN, saw, mod,3,0, -PP, VP, head 3, 3,</td><td>prep , transitive -mod -</td></tr></table>"
},
"TABREF3": {
"type_str": "table",
"text": "Accuracy and time cost for various estimation methods",
"num": null,
"html": null,
"content": "<table><tr><td/><td/><td>F-score</td><td/><td>Training</td><td colspan=\"2\">Parsing time (sec.)</td></tr><tr><td/><td colspan=\"6\">GENIA Corpus Penn Treebank time (sec.) GENIA Corpus Penn Treebank</td></tr><tr><td>Our method</td><td colspan=\"2\">86.87</td><td>86.81</td><td>2,278</td><td>611</td><td>3,165</td></tr><tr><td>Combined</td><td colspan=\"2\">86.32</td><td>86.09</td><td>29,421</td><td>424</td><td>2,757</td></tr><tr><td>GENIA only</td><td colspan=\"2\">85.72</td><td>42.49</td><td>1,694</td><td>332</td><td>8,183</td></tr><tr><td>Original model</td><td colspan=\"2\">85.10</td><td>87.16</td><td>137,038</td><td>515</td><td>2,554</td></tr><tr><td/><td/><td>87</td><td/><td/><td/></tr><tr><td/><td/><td>86.8</td><td/><td/><td/></tr><tr><td/><td/><td>86.6</td><td/><td/><td/></tr><tr><td/><td/><td>86.4</td><td/><td/><td/></tr><tr><td/><td>F-score</td><td>85.8 86 86.2</td><td/><td/><td/></tr><tr><td/><td/><td>85.6</td><td/><td/><td/></tr><tr><td/><td/><td>85.4</td><td/><td/><td/></tr><tr><td/><td/><td>85.2</td><td/><td/><td/></tr><tr><td/><td/><td>85</td><td/><td/><td/></tr><tr><td/><td/><td colspan=\"4\">0 500 1000 1500 2000 2500 3000 3500</td></tr><tr><td/><td/><td/><td colspan=\"2\">training sentences</td><td/></tr></table>"
},
"TABREF4": {
"type_str": "table",
"text": "Accuracy with the combination of WORD and another feature Features LP LR F-score diff. WORD h +WORDn+RULE 87.12 86.62 86.87 +1.77 WORD h +WORDn+DIST 86.41 85.86 86.14 +1.04 WORD h +WORDn+COMMA 86.91 86.38 86.64 +1.54 WORD h +WORDn+SPAN h +SPANn 85.77 85.22 85.49 +0.39 WORD h +WORDn+SYMBOL h +SYMBOLn 86.58 85.70 86.14 +1.04 WORD",
"num": null,
"html": null,
"content": "<table><tr><td colspan=\"3\">h +WORDn+POS h +POSn</td><td>86.53 85.99 86.26 +1.16</td></tr><tr><td colspan=\"2\">WORD h +WORDn+LE h +LEn</td><td/><td>86.16 85.68 85.92 +0.82</td></tr><tr><td colspan=\"2\">None</td><td/><td>85.39 84.82 85.10</td></tr><tr><td colspan=\"4\">Table 6. Errors in our model and Enju</td></tr><tr><td/><td colspan=\"3\">Total errors Common errors Errors not in</td></tr><tr><td/><td/><td/><td>the other model</td></tr><tr><td>Our model</td><td>1179</td><td>1050</td><td>129</td></tr><tr><td>Original model</td><td>1338</td><td>1050</td><td>288</td></tr></table>"
}
}
}
} |