File size: 67,294 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 1163 1164 1165 1166 |
{
"paper_id": "A00-1039",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T01:12:12.645894Z"
},
"title": "Unsupervised Discovery of Scenario-Level Patterns for Information Extraction",
"authors": [
{
"first": "Roman",
"middle": [],
"last": "Yangarber",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Ralph",
"middle": [],
"last": "Grishman",
"suffix": "",
"affiliation": {},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "Information Extraction (IE) systems are commonly based on pattern matching. Adapting an IE system to a new scenario entails the construction of a new pattern base-a timeconsuming and expensive process. We have implemented a system for finding patterns automatically from un-annotated text. Starting with a small initial set of seed patterns proposed by the user, the system applies an incremental discovery procedure to identify new patterns. We present experiments with evaluations which show that the resulting patterns exhibit high precision and recall.",
"pdf_parse": {
"paper_id": "A00-1039",
"_pdf_hash": "",
"abstract": [
{
"text": "Information Extraction (IE) systems are commonly based on pattern matching. Adapting an IE system to a new scenario entails the construction of a new pattern base-a timeconsuming and expensive process. We have implemented a system for finding patterns automatically from un-annotated text. Starting with a small initial set of seed patterns proposed by the user, the system applies an incremental discovery procedure to identify new patterns. We present experiments with evaluations which show that the resulting patterns exhibit high precision and recall.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "The task of Information Extraction (I-E) is the selective extraction of meaning from free natural language text. I \"Meaning\" is understood here in terms of a fixed set of semantic objects--entities, relationships among entities, and events in which entities participate. The semantic objects belong to a small number of types, all having fixed regular structure, within a fixed and closely circumscribed subject domain. The extracted objects are then stored in a relational database. In this paper, we use the nomenclature accepted in current IE literature; the term subject domain denotes a class of textual documents to be processed, e.g., \"business news,\" and scenario denotes the specific topic of interest within the domain, i.e., the set of facts to be extracted. One example of a scenario is \"management succession,\" the topic of MUC-6 (the Sixth Message Understanding Conference); in this scenario the system seeks to identify events in which corporate managers left 1For general references on IE, cf., e.g., (Pazienza, 1997; muc, 1995; muc, 1993) . their posts or assumed new ones. We will consider this scenario in detail in a later section describing experiments. IE systems today are commonly based on pattern matching. The patterns are regular expressions, stored in a \"pattern base\" containing a general-purpose component and a substantial domain-and scenario-specific component.",
"cite_spans": [
{
"start": 1017,
"end": 1033,
"text": "(Pazienza, 1997;",
"ref_id": null
},
{
"start": 1034,
"end": 1044,
"text": "muc, 1995;",
"ref_id": null
},
{
"start": 1045,
"end": 1055,
"text": "muc, 1993)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": null
},
{
"text": "Portability and performance are two major problem areas which are recognized as impeding widespread use of IE. This paper presents a novel approach, which addresses both of these problems by automatically discovering good patterns for a new scenario. The viability of our approach is tested and evaluated with an actual IE system.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": null
},
{
"text": "In the next section we describe the problem in more detail in the context of our IE system; sections 2 and 3 describe our algorithm for pattern discovery; section 4 describes our experimental results, followed by comparison with prior work and discussion, in section 5.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": null
},
{
"text": "1 The IE System",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": null
},
{
"text": "Our IE system, among others, contains a a backend core engine, at the heart of which is a regular-e~xpression pattern matcher. The engine draws on attendant knowledge bases (KBs) of varying degrees of domain-specificity. The KB components are commonly factored out to make the systems portable to new scenarios. There are four customizable knowledge bases in our IE system: the Lexicon contains general dictionaries and scenario-specific terms; the concept base groups terms into classes; the predicate base describes the logical structure of events to be extracted, and the pattern base contains patterns that catch the events in text.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": null
},
{
"text": "Each KB has a. substantial domain-specific component, which must be modified when mov-ing to new domains and scenarios. The system allows the user (i.e. scenario developer) to start with example sentences in text which contain events of interest, the candidates, and generalize them into patterns. However, the user is ultimately responsible for finding all the candidates, which amounts to manually processing example sentences in a very large training corpus. Should s/he fail to provide an example of a particular class of syntactic/semantic construction, the system has no hope of recovering the corresponding events. Our experience has shown that (1) the process of discovering candidates is highly expensive, and (2) gaps in patterns directly translate into gaps in coverage. How can the system help automate the process of discovering new good candidates? The system should find examples of all common linguistic constructs relevant to a scenario. While there has been prior research on identifying the primary lexical patterns of a sub-language or corpus (Grishman et al., 1986; Riloff, 1996) , the task here is more complex, since we are typically not provided in advance with a sub-corpus of relevant passages; these passages must themselves be found as part of the discovery process.",
"cite_spans": [
{
"start": 1063,
"end": 1086,
"text": "(Grishman et al., 1986;",
"ref_id": "BIBREF2"
},
{
"start": 1087,
"end": 1100,
"text": "Riloff, 1996)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": null
},
{
"text": "The difficulty is that one of the best indications of the relevance of the passages is precisely the presence of these constructs. Because of this circularity, we propose to acquire the constructs and passages in tandem.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": null
},
{
"text": "We outline our procedure for automatic acquisition of patterns; details are elaborated in later sections. The procedure is unsupervised in that it does not require the training corpus to be manually annotated with events of interest, nor a pro-classified corpus with relevance judgements, nor any feedback or intervention from the user 2. The idea is to combine IR-style document selection with an iterative relaxation process; this is similar to techniques used elsewhere in NLP, and is inspired in large part, if remotely, by the work of (Kay and RSscheisen, 1993) on automatic alignment of sentences and words in a bilingual corpus. There, the reasoning was: sentences that are translations of each 2however, it may be supervised after each iteration, where the user can answer yes/no questions to improve the quality of the results other are good indicators that words they contain are translation pairs; conversely, words that are translation pairs indicate that the sentences which contain them correspond to one another.",
"cite_spans": [
{
"start": 540,
"end": 566,
"text": "(Kay and RSscheisen, 1993)",
"ref_id": "BIBREF6"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Solution",
"sec_num": "2"
},
{
"text": "In our context, we observe that documents that are relevant to the scenario will necessarily contain good patterns; conversely, good patterns are strong indicators of relevant documents. The outline of our approach is as follows. (2) an initial set of trusted scenario patterns, as chosen ad hoc by the user--the seed; as will be seen, the seed can be quite small--two or three patterns seem to suffice. (3) an initial (possibly empty) set of concept classes",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Solution",
"sec_num": "2"
},
{
"text": "The pattern set induces a binary partition (a split) on the corpus: on any document, either zero or more than zero patterns will match. Thus the universe of documents, U, is partitioned into the relevant sub-corpus, R, vs. the non-relevant sub-corpus, R = U -R, with respect to the given pattern set. Actually, the documents are assigned weights which are 1 for documents matched by the trusted seed, and 0 otherwise. 3 2. Search for new candidate patterns: (a) Automatically convert each sentence in the corpus,into a set of candidate patterns, 4",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Solution",
"sec_num": "2"
},
{
"text": "(b) Generalize each pattern by replacing each lexical item which is a member of a concept class by the class name.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Solution",
"sec_num": "2"
},
{
"text": "(c) Working from the relevant documents, select those patterns whose distribution is strongly correlated with other relevant documents (i.e., much more 3R represents the trusted truth through the discovery iterations, since it was induced by the manually-selected seed.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Solution",
"sec_num": "2"
},
{
"text": "4Here, for each clause in the sentence we extract a tuple of its major roles: the head of the subject, the verb group, the object, object complement, as described below. This tuple is considered to be a pattern for the present purposes of discovery; it is a skeleton for the rich, syntactically transformed patterns our system uses in the extraction phase. densely distributed among the relevant documents than among the nonrelevant ones). The idea is to consider those candidate patterns, p, which meet the density, criterion:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Solution",
"sec_num": "2"
},
{
"text": "IHnRI IRI -->> IHnUI IUI",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Solution",
"sec_num": "2"
},
{
"text": "where H = H(p) is the set of documents where p hits.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Solution",
"sec_num": "2"
},
{
"text": "(d) Based on co-occurrence with the chosen patterns, extend the concept classes.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Solution",
"sec_num": "2"
},
{
"text": "classes to the user for review, retaining those relevant to the scenario.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Optional: Present the new candidates and",
"sec_num": "3."
},
{
"text": "The new pattern set induces a new partition on the corpus. With this pattern set, return to step 1. Repeat the procedure until no more patterns can be added.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "4.",
"sec_num": null
},
{
"text": "3 Methodology 3.1 Pre-proeessing: Normalization Before applying the discovery procedure, we subject the corpus to several stages of preprocessing. First, we apply a name recognition module, and replace each name with a token describing its class, e.g. C-Person, C-Company, etc. We collapse together all numeric expressions, currency values, dates, etc., using a single token to designate each of these classes.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "4.",
"sec_num": null
},
{
"text": "We then apply a parser to perform syntactic normalization to transform each clause into a common predicate-argument structure. We use the general-purpose dependency parser of English, based on the FDG formalism (Tapanainen and J~rvinen, 1997) and developed by the Research Unit for Multilingual Language Technology at the University of Helsinki, and Conexor Oy. The parser (modified to understand the name labels attached in the previous step) is used for reducing such variants as passive and relative clauses to a tuple, consisting of several elements.",
"cite_spans": [
{
"start": 211,
"end": 242,
"text": "(Tapanainen and J~rvinen, 1997)",
"ref_id": "BIBREF15"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Syntactic Analysis",
"sec_num": "3.2"
},
{
"text": "1. For each claus, the first element is the subject, a \"semantic\" subject of a non-finite sentence or agent of the passive. 5",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Syntactic Analysis",
"sec_num": "3.2"
},
{
"text": "2. The second element is the verb.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Syntactic Analysis",
"sec_num": "3.2"
},
{
"text": "3. The third element is the object, certain object-like adverbs, subject of the passive or subject complement 6 4. The fourth element is a phrase which refers to the object or the subject. A typical example of such an argument is an object complement, such as Company named John Smith president. Another instance is the so-called copredicatire (Nichols, 1978) , in the parsing system (J~irvinen and . A copredicative refers to a subject or an object, though this distinction is typically difficult to resolve automatically/ Clausal tuples also contain a locative modifier, and a temporal modifier. We used a corpus of 5,963 articles from the Wall Street Journal, randomly chosen. The parsed articles yielded a total of 250,000 clausal tuples, of which 135,000 were distinct.",
"cite_spans": [
{
"start": 344,
"end": 359,
"text": "(Nichols, 1978)",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Syntactic Analysis",
"sec_num": "3.2"
},
{
"text": "Because tuples may not repeat with sufficient frequency to obtain reliable statistics, each tuple is reduced to a set of pairs: e.g., a verbobject pair, a subject-object pair, etc. Each pair is used as a generalized pattern during the candidate selection stage. Once we have identified pairs which are relevant to the scenario, we use them to construct or augment concept classes, by grouping together the missing roles, (for example, a class of verbs which occur with a relevant subject-object pair: \"company (hire/fire/expel...} person\"). This is similar to work by several other groups which aims to induce semantic classes through syntactic co-occurrence analysis (Riloff and Jones, 1999; Pereira et al., 1993; Dagan et al., 1993; Hirschman et al., 1975) , although in .our case the contexts are limited to selected patterns, relevant to the scenario. SE.g., \"John sleeps\", \"John is appointed by Company\", \"I saw a dog which sleeps\", \"She asked John to buy a car\". 6E.g., \"John is appointed by Company\", \"John is the president of Company\", \"I saw a dog which sleeps\", The dog which I saw sleeps. 7For example, \"She gave us our coffee black\", \"Company appointed John Smith as president\".",
"cite_spans": [
{
"start": 668,
"end": 692,
"text": "(Riloff and Jones, 1999;",
"ref_id": "BIBREF13"
},
{
"start": 693,
"end": 714,
"text": "Pereira et al., 1993;",
"ref_id": "BIBREF12"
},
{
"start": 715,
"end": 734,
"text": "Dagan et al., 1993;",
"ref_id": "BIBREF0"
},
{
"start": 735,
"end": 758,
"text": "Hirschman et al., 1975)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Generalization and Concept Classes",
"sec_num": "3.3"
},
{
"text": "Here we present the results from experiments we conducted on the MUC-6 scenario, \"management succession\". The discovery procedure was seeded with a small pattern set, namely:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Pattern Discovery",
"sec_num": "3.4"
},
{
"text": "Subject Verb Direct Object C-Company C-Appoint C-Person C-Person C-Resign",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Pattern Discovery",
"sec_num": "3.4"
},
{
"text": "Documents are assigned relevance scores on a scale between 0 and 1. The seed patterns are accepted as ground truth; thus the documents they match have relevance 1. On subsequent iterations, the newly accepted patterns are not trusted as absolutely. On iteration number i q-1, each pattern p is assigned a precision measure, based on the relevance of the documents it matches: Pc(P) --Igl is the conditional probability of relevance. We further impose two support criteria: we distrust such frequent patterns where [HA U{ > a[U[ as uninformative, and rare patterns for which [H A R[ </3 as noise. \u00b0 At the end of each iteration, the system selects the pattern with the highest score, L(p), and adds it to the seed set. The documents which the winning pattern hits are added to the relevant set. The pattern search is then restarted.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Pattern Discovery",
"sec_num": "3.4"
},
{
"text": "Relevance The above is a simplification of the actual procedure, in several important respects.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Re-computatlon of Document",
"sec_num": "3.5"
},
{
"text": "Only generalized patterns are considered for candidacy, with one or more slots filled with wild-cards. In computing the score of the generalized pattern, we do not take into consideration all possible values of the wild-card role. We instead constrain the wild-card to those values which themselves in turn produce patterns with high scores. These values then become members of a new class, which is output in tandem with the winning pattern 1\u00b0 Ssimilarly to (Riloff, 1996) \u00b0U denotes the universe of documents. We used c~ = 0.i and ~-----2.",
"cite_spans": [
{
"start": 459,
"end": 473,
"text": "(Riloff, 1996)",
"ref_id": "BIBREF14"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Re-computatlon of Document",
"sec_num": "3.5"
},
{
"text": "1\u00b0The classes are currently unused by subsequent iterations; this important issue is considered in future work.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Re-computatlon of Document",
"sec_num": "3.5"
},
{
"text": "Preci+l(p) = 1 {H(p){ ~ Reli(d) (2) dEH(p)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Re-computatlon of Document",
"sec_num": "3.5"
},
{
"text": "where Reli(d) is the relevance of the document from the previous iteration, and H(p) is the set of documents where p matched. More generally, if K is a classifier consisting of a set of patterns, we can define H(K) as the set of documents where all of patterns p E K match, and the \"cumulative\" precision 11 of K as",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Re-computatlon of Document",
"sec_num": "3.5"
},
{
"text": "Preci+l(K) = 1 ~ Reli(d) (3) IH(K)[ riCH(K)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Re-computatlon of Document",
"sec_num": "3.5"
},
{
"text": "Once the new winning pattern is accepted, the relevance scores of the documents are readjusted as follows. For each document d which is matched by some (non-empty) subset of the currently accepted patterns, we can view that subset of patterns as a classifier K d = {py}. These patterns determine the new relevance score of the document",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Re-computatlon of Document",
"sec_num": "3.5"
},
{
"text": "Reli+l(d) = max (Rel~(d),Prec~+l(Kd)) (4)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Re-computatlon of Document",
"sec_num": "3.5"
},
{
"text": "This ensures that the relevance score grows monotonically, and only when there is sufficient positive evidence, as the patterns in effect vote \"conjunctively\" on the documents. The results which follow use this measure.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Re-computatlon of Document",
"sec_num": "3.5"
},
{
"text": "Thus in the formulas above, R is not simply the count of the relevant documents, but is rather their cumulative relevance. The two formulas, (3) and (4), capture the mutual dependency of patterns and documents; this recomputation and growing of precision and relevance scores is at the heart of the procedure.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Re-computatlon of Document",
"sec_num": "3.5"
},
{
"text": "11Of course, this measure is defined only when H(K) # 0.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Re-computatlon of Document",
"sec_num": "3.5"
},
{
"text": "An objective measure of goodness of a pattern o. 9 is not trivial to establish since the patterns cannot be used for extraction directly, without be-o. s ing properly incorporated into the knowledge base. Thus, the discovery procedure does not o. v lend itself easily to MUC-style evaluations, since 0.6 a pattern lacks information about which events it induces and which slots its arguments should 0.5 fill.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "1",
"sec_num": null
},
{
"text": "However, it is possible to apply some objec-o. a tive measures of performance. One way we evaluated the system is by noting that in addition o. to growing the pattern set, the procedure also grows the relevance of documents. The latter o. 2 can be objectively evaluated.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "1",
"sec_num": null
},
{
"text": "We used a test corpus of 100 MUC-6 formaltraining documents (which were included in the o main development corpus of about 6000 documents) plus another 150 documents picked at random from the main corpus and judged by hand. These judgements constituted the ground truth and were used only for evaluation, (not in the discovery procedure). Figure 1 shows the recall/precision measures with respect to the test corpus of 250 documents, over a span of 60 generations, starting with the seed set in table 3.4. The Seed patterns matched 184 of the 5963 documents, yielding an initial recall of .11 and precision of .93; by the last generation it searched through 982 documents with non-zero relevance, and ended with .80 precision and .78 recall. This facet of the discovery procedure is closely related to the MUC '%ext-filtering\" sub-task, where the systems are judged at the level of documents rather than event slots. It is interesting to compare the results with other MUC-6 participants, shown anonymously in figure 2. Considering recall and precision separately, the discovery procedure attains values comparable to those achieved by some of the participants, all of which were either heavily-supervised or manually coded systems. It is important to bear in mind that the discovery procedure had no benefit of training material, or any information beyond the seed pattern set. Figure 2 shows two evaluations of our discovery procedure, tested against the original MUC-6 corpus of 100 documents, and against our test corpus, which consists of an additional 150 documents judged manually. The two plots in the figure show a slight difference in results, indicating that in some sense, the MUC corpus was more \"random\", or that our expanded corpus was somewhat skewed in favor of more common patterns that the system is able to find more easily.",
"cite_spans": [],
"ref_spans": [
{
"start": 339,
"end": 347,
"text": "Figure 1",
"ref_id": null
},
{
"start": 1379,
"end": 1387,
"text": "Figure 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "0.1",
"sec_num": null
},
{
"text": "The graphs shown in Figures 1 and 2 are based on an \"objective\" measure we adopted during the experiments. This is the same measure of relevance used internally by the discovery procedure on each iteration (relative to the \"truth\" of relevance scores of the previous iteration), and is not quite the standard measure used for text filtering in IR. According to this measure, the system gets a score for each document based on the relevance which it assigned to the document. Thus if the system .assigned relevance of X percent to a relevant document, it only received X Figure 3 : Results on the MUC corpus percent on the recall score for classifying that document correctly. Similarly, if the system assigned relevance Y to an irrelevant document, it was penalized only for the mis-classified Y percent on the precision score. To make our results more comparable to those of other MUC competitors, we chose a cut-off point and force the system to make a binary relevance decision on each document. The cut-off of 0.5 seemed optimal from empirical observations. Figure 3 shows a noticeable improvement in scores, when using our continuous, \"objective\" measure, vs. the cut-off measure, with the entire graph essentially translated to the right for a gain of almost 10 percentage points of recall.",
"cite_spans": [],
"ref_spans": [
{
"start": 20,
"end": 35,
"text": "Figures 1 and 2",
"ref_id": null
},
{
"start": 570,
"end": 578,
"text": "Figure 3",
"ref_id": null
},
{
"start": 1062,
"end": 1070,
"text": "Figure 3",
"ref_id": null
}
],
"eq_spans": [],
"section": "Choice of Evaluation Metric",
"sec_num": "4.3"
},
{
"text": "i i i i i im~ ! ! ! i i DiE c ! i i i i i i i i 0 . '",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Choice of Evaluation Metric",
"sec_num": "4.3"
},
{
"text": "Another effective, if simple, measure of performanceis how many of the patterns the procedure found, and comparing them with those used by an extraction engine which was manually constructed for the same task. Our MUC-6 system used approximately 75 clause level patterns, with 30 distinct verbal heads. In one conservative experiment, we observed that the discovery procedure found 17 of these verbs, or 57%. However, it also found at least 8 verbs the manual system lacked, which seemed relevant to the scenario: company-bring-person- [as\u00f7officer] 12",
"cite_spans": [
{
"start": 536,
"end": 548,
"text": "[as\u00f7officer]",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluating Patterns",
"sec_num": "4.4"
},
{
"text": "person-rejoin-company- [as + o25cer] person- { ret , conti, e, remai, ,stay}-[as + o25cer] person-pursue-interest At the risk of igniting a philosophical debate over what is or is not relevant to a scenario, we note that the first four of these verbs are evidently essential to the scenario in the strictest definition, since they imply changes of post. The next three are \"staying\" verbs, and are actually also needed, since higher-level inferences required in tracking events for long-range merging over documents, require knowledge of persons occupying posts, rather than only assuming or leaving them. The most curious one is \"person-pursue-interesf'; surprisingly, it too is useful, even in the strictest MUC sense, cf., (muc, 1995) . Systems are judged on filling a slot called \"other-organization\", indicating from or to which company the person came or went. This pattern is consistently used in text to indi- cate that the person left to pursue other, undisclosed interests, the knowledge of which would relieve the system from seeking other information in order to fill this slot. This is to say that here strict evaluation is elusive.",
"cite_spans": [
{
"start": 23,
"end": 36,
"text": "[as + o25cer]",
"ref_id": null
},
{
"start": 45,
"end": 90,
"text": "{ ret , conti, e, remai, ,stay}-[as + o25cer]",
"ref_id": null
},
{
"start": 726,
"end": 737,
"text": "(muc, 1995)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "person-come-[to+eompanv]-[as+oZScer]",
"sec_num": null
},
{
"text": "Some of the prior research has emphasized interactive tools to convert examples to extraction patterns, cf. (Yangarber and Grishman, 1997) , while others have focused on methods for automatically converting a corpus annotated with extraction examples into such patterns (Lehnert et al., 1992; Fisher et al., 1995; Miller et al., 1998) . These methods, however, do not reduce the burden of finding the examples to annotate. With either approach, the portability bottleneck is shifted from the problem of building patterns to that of finding good candidates. The prior work most closely related to this study is (Riloff, 1996) , which, along with (Riloff, 1993) , seeks automatic methods for filling slots in event templates. However, the prior work differs from that presented here in several crucial respects; firstly, the prior work does not attempt to find entire events, after the fashion of MUC's highest-level scenario-template task. Rather the patterns produced by those systems identify NPs that fill individual slots, without specifying how these slots may be combined at a later stage into complete event templates. The present work focuses on directly discovering event-level, multi-slot relational patterns. Secondly, the prior work either relies on a set of documents with relevance judgements to find slot fillers where they are relevant to events, (Riloff, 1996) , or utilizes an un-classified corpus containing a very high proportion of relevant documents to find all instances of a semantic class, (Riloff and Jones, 1999) . By contrast, our procedure requires no relevance judgements, and works on the assumption that the corpus is balanced and the proportion of relevant documents is small. Classifying documents by hand, although admittedly easier than tagging event instances in text for automatic training, is still a formidable task. When we prepared the test corpus, it took 5 hours to mark 150 short documents.",
"cite_spans": [
{
"start": 108,
"end": 138,
"text": "(Yangarber and Grishman, 1997)",
"ref_id": "BIBREF16"
},
{
"start": 270,
"end": 292,
"text": "(Lehnert et al., 1992;",
"ref_id": "BIBREF7"
},
{
"start": 293,
"end": 313,
"text": "Fisher et al., 1995;",
"ref_id": "BIBREF1"
},
{
"start": 314,
"end": 334,
"text": "Miller et al., 1998)",
"ref_id": "BIBREF8"
},
{
"start": 610,
"end": 624,
"text": "(Riloff, 1996)",
"ref_id": "BIBREF14"
},
{
"start": 645,
"end": 659,
"text": "(Riloff, 1993)",
"ref_id": "BIBREF13"
},
{
"start": 1362,
"end": 1376,
"text": "(Riloff, 1996)",
"ref_id": "BIBREF14"
},
{
"start": 1514,
"end": 1538,
"text": "(Riloff and Jones, 1999)",
"ref_id": "BIBREF13"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion and Current Work",
"sec_num": "5"
},
{
"text": "The presented results indicate that our method of corpus analysis can be used to rapidly identify a large number of relevant patterns without pre-classifying a large training corpus. We are at the early stages of understanding how to optimally tune these techniques, and there are number of areas that need refinement. We are working on capturing the rich information about concept classes which is currently returned as part of our pattern discovery procedure, to build up a concept dictionary in tandem with the pattern base. We are also considering the proper selection of weights and thresholds for controlling the rankings of patterns and documents, criteria for terminating the iteration process, and for dynamic adjustments of these weights. We feel that the generalization technique in pattern discovery offers a great opportunity for combating sparseness of data, though this requires further research. Lastly, we are studying these algorithms under several unrelated scenarios to determine to what extent scenario-specific phenomena affect their performance.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion and Current Work",
"sec_num": "5"
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Contextual word similarity and estimation from sparse data",
"authors": [
{
"first": "Shaul",
"middle": [],
"last": "Ido Dagan",
"suffix": ""
},
{
"first": "Shaul",
"middle": [],
"last": "Marcus",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Markovitch",
"suffix": ""
}
],
"year": 1993,
"venue": "Proceedings of the 31st Annual Meeting of the Assn. for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "31--37",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ido Dagan, Shaul Marcus, and Shaul Markovitch. 1993. Contextual word simi- larity and estimation from sparse data. In Proceedings of the 31st Annual Meeting of the Assn. for Computational Linguistics, pages 31-37, Columbus, OH, June.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Description of the UMass system as used for MUC-6",
"authors": [
{
"first": "David",
"middle": [],
"last": "Fisher",
"suffix": ""
},
{
"first": "Stephen",
"middle": [],
"last": "Soderland",
"suffix": ""
},
{
"first": "Joseph",
"middle": [],
"last": "Mc-Carthy",
"suffix": ""
}
],
"year": 1995,
"venue": "Proc. Si;zth Message Understanding Conf. (MUC-6)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "David Fisher, Stephen Soderland, Joseph Mc- Carthy, Fang-fang Feng, and Wendy Lehnert. 1995. Description of the UMass system as used for MUC-6. In Proc. Si;zth Message Un- derstanding Conf. (MUC-6), Columbia, MD, November. Morgan Kaufmann.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Discovery procedures for sublanguage selectional patterns: Initial experiments",
"authors": [
{
"first": "R",
"middle": [],
"last": "Grishman",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Hirschman",
"suffix": ""
},
{
"first": "N",
"middle": [
"T"
],
"last": "Nhan",
"suffix": ""
}
],
"year": 1986,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "R. Grishman, L. Hirschman, and N.T. Nhan. 1986. Discovery procedures for sublanguage selectional patterns: Initial experiments.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Grammatically-based automatic word class formation",
"authors": [
{
"first": "Lynette",
"middle": [],
"last": "Hirschman",
"suffix": ""
},
{
"first": "Ralph",
"middle": [],
"last": "Grishman",
"suffix": ""
},
{
"first": "Naomi",
"middle": [],
"last": "Sager",
"suffix": ""
}
],
"year": 1975,
"venue": "Information Processing and Management",
"volume": "11",
"issue": "1/2",
"pages": "39--57",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Lynette Hirschman, Ralph Grishman, and Naomi Sager. 1975. Grammatically-based automatic word class formation. Information Processing and Management, 11(1/2):39-57.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "A dependency parser for English. Technical Report TR-1, Department of General Linguistics",
"authors": [
{
"first": "J",
"middle": [],
"last": "Timo",
"suffix": ""
},
{
"first": "Pasi",
"middle": [],
"last": "/Irvinen",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Tapanainen",
"suffix": ""
}
],
"year": 1997,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Timo J/irvinen and Pasi Tapanainen. 1997. A dependency parser for English. Technical Re- port TR-1, Department of General Linguis- tics, University of Helsinki, Finland, Febru- ary.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Text-translation alignment",
"authors": [
{
"first": "Martin",
"middle": [],
"last": "Kay",
"suffix": ""
},
{
"first": "Martin",
"middle": [],
"last": "Rsscheisen",
"suffix": ""
}
],
"year": 1993,
"venue": "Computational Linguistics",
"volume": "",
"issue": "1",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Martin Kay and Martin RSscheisen. 1993. Text-translation alignment. Computational Linguistics, 19(1).",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "University of massachusetts: MUC-4 test results and analysis",
"authors": [
{
"first": "W",
"middle": [],
"last": "Lehnert",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Cardie",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Fisher",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Mccarthy",
"suffix": ""
},
{
"first": "E",
"middle": [],
"last": "Riloff",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Soderland ; Mclean",
"suffix": ""
},
{
"first": "June",
"middle": [
"Morgan"
],
"last": "Va",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Kaufmann",
"suffix": ""
}
],
"year": 1992,
"venue": "Proc. Fourth Message Understanding Conf",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "W. Lehnert, C. Cardie, D. Fisher, J. McCarthy, E. Riloff, and S. Soderland. 1992. Univer- sity of massachusetts: MUC-4 test results and analysis. In Proc. Fourth Message Un- derstanding Conf., McLean, VA, June. Mor- gan Kaufmann.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Algorithms that learn to extract information; BBN: Description of the SIFT system as used for MUC-7",
"authors": [
{
"first": "Scott",
"middle": [],
"last": "Miller",
"suffix": ""
},
{
"first": "Michael",
"middle": [],
"last": "Crystal",
"suffix": ""
},
{
"first": "Heidi",
"middle": [],
"last": "Fox",
"suffix": ""
},
{
"first": "Lance",
"middle": [],
"last": "Ramshaw",
"suffix": ""
},
{
"first": "Richard",
"middle": [],
"last": "Schwartz",
"suffix": ""
},
{
"first": "Rebecca",
"middle": [],
"last": "Stone",
"suffix": ""
}
],
"year": 1993,
"venue": "Proceedings of the Fifth Message Understanding Conference (MUC-5)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Scott Miller, Michael Crystal, Heidi Fox, Lance Ramshaw, Richard Schwartz, Rebecca Stone, Ralph Weischedel, and the Annota- tion Group. 1998. Algorithms that learn to extract information; BBN: Description of the SIFT system as used for MUC-7. In Proc. of the Seventh Message Understanding Confer- ence, Fairfax, VA. 1993. Proceedings of the Fifth Message Un- derstanding Conference (MUC-5), Baltimore, MD, August. Morgan Kaufmann.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Proceedings of the Sixth Message Understanding Conference (MUC-6)",
"authors": [],
"year": 1995,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "1995. Proceedings of the Sixth Message Un- derstanding Conference (MUC-6), Columbia, M_D, November. Morgan Kaufmann.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Secondary predicates. Proceedings of the 4th Annual Meeting of Berkeley Linguistics Society",
"authors": [
{
"first": "Johanna",
"middle": [],
"last": "Nichols",
"suffix": ""
}
],
"year": 1978,
"venue": "",
"volume": "",
"issue": "",
"pages": "114--127",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Johanna Nichols. 1978. Secondary predicates. Proceedings of the 4th Annual Meeting of Berkeley Linguistics Society, pages 114-127.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Information Extraction",
"authors": [],
"year": 1997,
"venue": "Lecture Notes in Artificial Intelligence",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Maria Teresa Pazienza, editor. 1997. Infor- mation Extraction. Springer-Verlag, Lecture Notes in Artificial Intelligence, Rome.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Distributional clustering of English words",
"authors": [
{
"first": "Fernando",
"middle": [],
"last": "Pereira",
"suffix": ""
},
{
"first": "Naftali",
"middle": [],
"last": "Tishby",
"suffix": ""
},
{
"first": "Lillian",
"middle": [],
"last": "Lee",
"suffix": ""
}
],
"year": 1993,
"venue": "Proceedings of the 31st Annual Meeting of the Assn. for Computational Linguistics",
"volume": "",
"issue": "",
"pages": "183--190",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Fernando Pereira, Naftali Tishby, and Lillian Lee. 1993. Distributional clustering of En- glish words. In Proceedings of the 31st An- nual Meeting of the Assn. for Computational Linguistics, pages 183-190, Columbus, OH, June.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Learning dictionaries for information extraction by multi-level bootstrapping",
"authors": [
{
"first": "Ellen",
"middle": [],
"last": "Riloff",
"suffix": ""
},
{
"first": "Rosie",
"middle": [],
"last": "Jones",
"suffix": ""
}
],
"year": 1993,
"venue": "Proceedings of Sixteenth National Conference on Artificial Intelligence (AAAI-99)",
"volume": "",
"issue": "",
"pages": "811--816",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ellen Riloff and Rosie Jones. 1999. Learn- ing dictionaries for information extraction by multi-level bootstrapping. In Proceedings of Sixteenth National Conference on Artificial Intelligence (AAAI-99), Orlando, Florida, Ellen Riloff. 1993. Automatically construct- ing a dictionary for information extraction tasks. In Proceedings of Eleventh National Conference on Artificial Intelligence (AAAI- 93), pages 811-816. The AAAI Press/MIT Press.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Automatically generating extraction patterns from untagged text",
"authors": [
{
"first": "Ellen",
"middle": [],
"last": "Riloff",
"suffix": ""
}
],
"year": 1996,
"venue": "Proceedings of Thirteenth National Conference on Artificial Intelligence (AAAL96)",
"volume": "",
"issue": "",
"pages": "1044--1049",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ellen Riloff. 1996. Automatically generating extraction patterns from untagged text. In Proceedings of Thirteenth National Confer- ence on Artificial Intelligence (AAAL96), pages 1044-1049. The AAAI Press/MIT Press.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "A non-projective dependency parser",
"authors": [
{
"first": "Pasi",
"middle": [],
"last": "Tapanainen",
"suffix": ""
},
{
"first": "Timo",
"middle": [],
"last": "J~rvinen",
"suffix": ""
}
],
"year": 1997,
"venue": "Proceedings of the 5th Conference on Applied Natural Language Processing",
"volume": "",
"issue": "",
"pages": "64--71",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Pasi Tapanainen and Timo J~rvinen. 1997. A non-projective dependency parser. In Pro- ceedings of the 5th Conference on Applied Natural Language Processing, pages 64-71, Washington, D.C., April. ACL.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Customization of information extraction systems",
"authors": [
{
"first": "Roman",
"middle": [],
"last": "Yangarber",
"suffix": ""
},
{
"first": "Ralph",
"middle": [],
"last": "Grishman",
"suffix": ""
}
],
"year": 1997,
"venue": "International Workshop on Lexically Driven Information Extraction",
"volume": "",
"issue": "",
"pages": "1--11",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Roman Yangarber and Ralph Grishman. 1997. Customization of information extraction sys- tems. In Paola Velardi, editor, International Workshop on Lexically Driven Information Extraction, pages 1-11, Frascati, Italy, July. Universit?~ di Roma.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"text": "a large corpus of un-annotated and un-classified documents in the domain;",
"num": null,
"type_str": "figure",
"uris": null
},
"FIGREF1": {
"text": "Here C-Company and C-Person denote semantic classes containing named entities of the corresponding semantic types. C-Appoirlt denotes a class of verbs, containing four verbs { appoint, elect, promote, name}; C-Resign = { resign, depart, quit, step-down }. During a single iteration, we compute the score s, L(p), for each candidate pattern p: L(p) = Pc(P)\" log {H A R] (1) where R denotes the relevant subset, and H --H(p) the documents matching p, as above, and [gnR[",
"num": null,
"type_str": "figure",
"uris": null
},
"FIGREF2": {
"text": "....... ~ .......... ~ ......... ~ ........... ~ .......... ~ .......... ~ ......... ...... iiiiiiiiiiiiiilEi ........... ........ /... .......... ~ .......... '.\" ........ \": .......... ~ ........... r .......... ~ .......... ! .........",
"num": null,
"type_str": "figure",
"uris": null
},
"FIGREF3": {
"text": "Figure h Recall/Precision curves for Management Succession",
"num": null,
"type_str": "figure",
"uris": null
},
"FIGREF4": {
"text": "7 ................................................................... ~\" ............... ... ~ ........ , ........ ~ ....... ~ ........ ~ ....... ~ ........ ! ........ : ........ 4 .......",
"num": null,
"type_str": "figure",
"uris": null
},
"TABREF0": {
"text": "....... ........ ............... ---! ........ ....... ...... i ........ i ........ [ ....... ? ........ f ............ T ....... ...... J ........ i ........ i. ....... i ........ ~ ....... .; ........... ..: ........ i....: 0 6 ........................ .' 7 ....... ' ........ ~ ....... ~ ........ ~ ....................... ..... '. ........ , ........ ~\" ....... , ........ ~ ....... ~ ........ * ....... \"=. ........ , ....... ...... e ........ i ........ ! ........ ~ ........ ~... ' ........ ! ........ i .......",
"html": null,
"content": "<table><tr><td/><td/><td>II :</td><td>I :</td><td>I .</td><td>I ,</td><td/><td>I .</td><td/><td>I :</td><td>I \u2022</td><td>I</td></tr><tr><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/><td>i</td><td>.</td><td>i</td><td>~</td><td>iB</td><td>i</td></tr><tr><td colspan=\"2\">0.9</td><td/><td/><td/><td/><td/><td/><td/><td/><td/><td>0.9</td></tr><tr><td colspan=\"2\">0.8</td><td>i</td><td>i</td><td>i</td><td>::</td><td>i</td><td>i</td><td>i</td><td>i</td><td>i</td><td>~</td></tr><tr><td>0</td><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/></tr><tr><td colspan=\"2\">0.7</td><td/><td/><td/><td/><td/><td/><td/><td/><td/></tr><tr><td>(I)</td><td/><td/><td/><td/><td/><td/><td/><td/><td/><td/></tr><tr><td/><td/><td/><td/><td>i</td><td/><td/><td/><td/><td/><td/></tr><tr><td/><td/><td/><td/><td>i i</td><td/><td/><td/><td/><td/><td/></tr><tr><td>0</td><td>. 5</td><td>..i</td><td/><td>z</td><td/><td/><td/><td/><td/><td/></tr><tr><td colspan=\"2\">0.4</td><td>i</td><td/><td>I</td><td/><td/><td/><td/><td/><td/></tr><tr><td/><td>0</td><td colspan=\"8\">0.10.20.30.40.50.60.70.80.9</td><td/><td>1</td></tr><tr><td/><td/><td/><td/><td/><td colspan=\"2\">Recall</td><td/><td/><td/><td/></tr><tr><td/><td/><td colspan=\"7\">Figure 2: Precision vs. Recall</td><td/><td/></tr></table>",
"num": null,
"type_str": "table"
}
}
}
} |