File size: 52,092 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 |
{
"paper_id": "M92-1030",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T03:13:04.202136Z"
},
"title": "MITRE-Bedford : Description of the ALEMBIC System as Used for MUC-4",
"authors": [
{
"first": "John",
"middle": [],
"last": "Aberdeen",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "John",
"middle": [],
"last": "Burger",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Dennis",
"middle": [],
"last": "Connolly",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Susan",
"middle": [],
"last": "Roberts",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "Marc",
"middle": [],
"last": "Vilai",
"suffix": "",
"affiliation": {},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "The ALEMBIC text understanding system fielded at MUC-4 by MITRE-Bedford is primarily based on natura l language techniques. ALEMBIC 1 is a research prototype that is intended to explore several major areas o f investigation : \u2022 Error recovery, involving primarily issues of semi-parsing and recovery of plausible attachments. \u2022 Robustness, involving primarily issues of uncertain reasoning and tractable inference. \u2022 Self-extensibility, focusing primarily on machine learning of natural language and userconfigurable semantics. \u2022 System integration, through SGML (the Standard Generalized Markup Language), both at the leve l of meaning analysis and at the overall application level. This investigation is part of an internally-funded research initiative towards processing open source texts (i .e. , free natural language texts drawn from broadcast transcripts, news wires, etc .). This initiative has been underway for just over half a year, prior to which our group was focusing nearly exclusively on natural language interfaces t o expert systems. We are thus newcomers to the MUC data extraction task, and our system is still very much in earl y phases of development. The system details we present here should thus be taken as preliminary .",
"pdf_parse": {
"paper_id": "M92-1030",
"_pdf_hash": "",
"abstract": [
{
"text": "The ALEMBIC text understanding system fielded at MUC-4 by MITRE-Bedford is primarily based on natura l language techniques. ALEMBIC 1 is a research prototype that is intended to explore several major areas o f investigation : \u2022 Error recovery, involving primarily issues of semi-parsing and recovery of plausible attachments. \u2022 Robustness, involving primarily issues of uncertain reasoning and tractable inference. \u2022 Self-extensibility, focusing primarily on machine learning of natural language and userconfigurable semantics. \u2022 System integration, through SGML (the Standard Generalized Markup Language), both at the leve l of meaning analysis and at the overall application level. This investigation is part of an internally-funded research initiative towards processing open source texts (i .e. , free natural language texts drawn from broadcast transcripts, news wires, etc .). This initiative has been underway for just over half a year, prior to which our group was focusing nearly exclusively on natural language interfaces t o expert systems. We are thus newcomers to the MUC data extraction task, and our system is still very much in earl y phases of development. The system details we present here should thus be taken as preliminary .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "The system ' s underlying architecture, shown in Figure 1 , follows a task breakdown used in several other system s that have recently participated in Muc (e .g ., PLUM [10] or NLTOOLSET [4] ) . Processing occurs in three distinct phases : preprocessing, natural language analysis, and application-specific output generation . One of the way s ALEMBIC differs from other Muc systems, however, is in exploiting SGML as the interchange lingua franca betwee n these three processing phases. The intention is to allow system modules whose invocation occurs early in the analysis of a document to record processing results directly in the document through SGML markup . This information then becomes available to subsequent modules as meta-data.",
"cite_spans": [
{
"start": 169,
"end": 173,
"text": "[10]",
"ref_id": "BIBREF9"
},
{
"start": 187,
"end": 190,
"text": "[4]",
"ref_id": "BIBREF3"
}
],
"ref_spans": [
{
"start": 49,
"end": 57,
"text": "Figure 1",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "OVERALL ARCHITECTUR E",
"sec_num": null
},
{
"text": "As a result of this sGML-based architecture, the system's overall flow of control is governed from an objectoriented document manager built on top of a Common Lisp port of Goldfarb's public domain SGML parser . For MUC-4, the pre-processing phase thus takes a FBIS message file and normalizes it by recoding it in SGML Th e document manager then builds an internal document object by parsing the resulting sGML . The actual conten t l alembic 1 : an alchemical apparatus used for distillation 2 : something that refines or transmutes as if by distillation . ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "OVERALL ARCHITECTUR E",
"sec_num": null
},
{
"text": "Application-specific phase Doc manage r (SGML parser) f \u25ba analysis of the document is performed by invoking the natural language analysis modules on the internal documen t object, and the results of these analyses are stored as attributes of the document. The system's output is normall y just another SGML file, in which the content analysis is simply encoded as additional (semantic) markup . For MUC-4, we also provided selective output that consisted solely of filled templates .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NL parser & interpreter",
"sec_num": null
},
{
"text": "As an example of this overall flow of control, and its corresponding encoding in SGML, consider the first paragraph of message TST2-MUC4-0048 :",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "NL parser & interpreter",
"sec_num": null
},
{
"text": "The SGML normal\u00bbation of this paragraph produced by the pre-processor begins as follows .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "SALVADORAN PRESIDENT-ELECT ALFREDO CRISTIANI CONDEMNED THE TERRORIS T KILLING OF ATTORNEY GENERAL ROBERTO GARCIA ALVARADO AND ACCUSED TH E FARABUNDO MARTI NATIONAL LIBERATION FRONT (FMLN) OF THE CRIME ,",
"sec_num": null
},
{
"text": "<p><s>SALVADORAN PRESIDENT<punct 1oc=\"midword\" type=\"hyphen\"> -</punct>ELECT ALFREDO CRISTIANI CONDEMNED THE TERRORIST KILLING . . .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "SALVADORAN PRESIDENT-ELECT ALFREDO CRISTIANI CONDEMNED THE TERRORIS T KILLING OF ATTORNEY GENERAL ROBERTO GARCIA ALVARADO AND ACCUSED TH E FARABUNDO MARTI NATIONAL LIBERATION FRONT (FMLN) OF THE CRIME ,",
"sec_num": null
},
{
"text": "The p and s tags stand respectively for paragraph and sentence delimiters, and the punct tag encodes normalized punctuation . In SGML parlance, the text bracketed by the <punct . . .> and </punct> delimiters is a punct element, and the equated terms in the punct tag are attributes of the overall element . For other details on SGML, see, e.g., [8] .",
"cite_spans": [
{
"start": 345,
"end": 348,
"text": "[8]",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "SALVADORAN PRESIDENT-ELECT ALFREDO CRISTIANI CONDEMNED THE TERRORIS T KILLING OF ATTORNEY GENERAL ROBERTO GARCIA ALVARADO AND ACCUSED TH E FARABUNDO MARTI NATIONAL LIBERATION FRONT (FMLN) OF THE CRIME ,",
"sec_num": null
},
{
"text": "Turning to the natural language phase, the structural markup for sentences, paragraphs, and quotes is exploite d straightforwardly to dispatch text chunks to the linguistic parser . More interestingly, punctuation markup can also appear as part of the actual definitions of lexical items, e.g ., the possessive marker ('s) or hyphenated words. Th e lexicon entry for the title modifier -elect, for example, is the sequence (*mw-hyphen* elect), in which *mw-hyphen * matches any SGML punct element with loc and type attributes respectively set to midword and hyphen .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "SALVADORAN PRESIDENT-ELECT ALFREDO CRISTIANI CONDEMNED THE TERRORIS T KILLING OF ATTORNEY GENERAL ROBERTO GARCIA ALVARADO AND ACCUSED TH E FARABUNDO MARTI NATIONAL LIBERATION FRONT (FMLN) OF THE CRIME ,",
"sec_num": null
},
{
"text": "As mentioned, when the natural language phase has been completed, ALEMBIC records its analysis of th e document as further annotation . In the case of the MUC-4 version of the system, this markup simply encodes th e templates that the system has produced, e .g . , <p><template > <slotname>0 . MESSAGE :ID</slotname> <slotval>TST2-MUC4-0048</slotval > <slotname>l . MESSAGE : TEMPLATE</slotname> <slotval>1</slotval > </template > <s>SALVADORAN PRESIDENT <punct loc=\"midword\" type=\"hyphen> -</punct>ELECT . ..",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "SALVADORAN PRESIDENT-ELECT ALFREDO CRISTIANI CONDEMNED THE TERRORIS T KILLING OF ATTORNEY GENERAL ROBERTO GARCIA ALVARADO AND ACCUSED TH E FARABUNDO MARTI NATIONAL LIBERATION FRONT (FMLN) OF THE CRIME ,",
"sec_num": null
},
{
"text": "As we alluded to above, this phase of processing is intended to normalize aspects of a document that are simpl y too messy to be dealt with during linguistic parsing . In the version of ALEMBIC used in MUC-4, this include s document structure, especially header structure, punctuation, and numeral strings . By handling the document structure in this preliminary phase, we gain all the usual advantages of separating tokenization from parsing, and additionally can introduce special-purpose error recovery strategies . These strategies address such problems as missing quote marks, missing periods at the end of paragraphs, and so forth. One advantage of using SGML is that it actually simplifies implementing these error recovery strategies . SGML allows the preprocessor to ommit issuin g many structural tags, in particular some that are keyed directly off of punctuation . The document manager treats th e missing markup as implicit, and fills it in from a document grammar instead .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "INDIVIDUAL PROCESSING MODULE S Pre-processing phas e",
"sec_num": null
},
{
"text": "A further motivation for using SGML is that it readily allows us to extend the scope of pre-processing through incremental addition of further modules . Among the modules that we have considered adding to the pre-processor are an sGML-native part-of-speech tagger, and a coarse subject identifier (based on Amsler's FORCE4 algorithm) . Both of these have been implemented by our colleagues in MITRE Washington .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "INDIVIDUAL PROCESSING MODULE S Pre-processing phas e",
"sec_num": null
},
{
"text": "The document manager provides an object-oriented framework for working with SGML documents . The manager is entirely cLOS-based, and SGML elements are thus made available as instances of cLOS objects . A sentence element (corresponding to the string bracketed by matching <s> and </s> tags) is mapped into an instance of the S object, and any S-specific code (e .g ., the linguistic parser) is thus made applicable to the element.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Document Manage r",
"sec_num": null
},
{
"text": "As mentioned, the document manager is built around a public domain SGML parser/tokenizer written by Goldfarb , the godfather of SGML. The parser consists of C language routines that were made available through the Consortiu m for Lexical Research . On the Lisp side, there are several ways to use the parser . At the lowest level, one can simpl y get token types and character strings out of an SGML document . At the highest level, one can get a stream of cLO s objects representing SGML tokens . The parser takes care of canonicalizing a document by, e .g., inserting any tags left implicit by the preprocessor, or filling in the default attribute values of attributes.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Document Manage r",
"sec_num": null
},
{
"text": "The design of our lexicon is hierarchical, with specific lexical classes (e .g ., auxiliary verbs or mono-transitive verbs) being subsumed in the hierarchy by more abstract lexical classes (e .g ., verbs in general). This allows for significant sharing of lexical information between lexical classes, and reduces the complexity of the lexicon .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Linguistic Lexicon",
"sec_num": null
},
{
"text": "Lexicon entries correspond to base stems, which are automatically expanded into the lexical forms that ar e actually used by the system . Our syntactic framework closely follows Steedman's combinatory categorial grammar s (CCG's), and as a result the expansion of a stem occurs along two dimensions .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Linguistic Lexicon",
"sec_num": null
},
{
"text": "\u2022 Lexical word forms, that is, the surface forms of the stem . For count nouns this is just the singular and plural form of the noun; for main verbs, this includes the usual participial forms, tensed forms , and infinitive, as well as adjectival forms, and various nominalizations .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Linguistic Lexicon",
"sec_num": null
},
{
"text": "\u2022 Lexical category types, that is, the underlying syntactic categories that are sanctioned by a give n word form. In the case of a mono-transitive verb's past tense/participle (e.g ., \"condemned,\" the first verb in TST2-MUC4-0048), this includes the active voice (e .g ., \"Cristiani . condemned the terrorist killing\"), the passive voice, and ancillary forms such as the detransitivized active voice and verbal adjectives.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Linguistic Lexicon",
"sec_num": null
},
{
"text": "In our variant of CCU's, lexical categories are treated either as complex vectors of features, or as mathematica l functions over such vectors . For example, stripping away all syntactic and semantic features, the syntactic category corresponding to a transitive verb is the function SWP/NP, i.e ., a function that will produce an S given an object NP on its right (the /NP term) and a subject NP on its left (the \\NP term) . To accomodate syntactic and semanti c features, categories are actually implemented in a standard unification framework (as in [11] ). Some features can b e themselves category-valued, and so the overall category structure is actually a reentrant graph that can become fairl y involved, as attested to by a partial expansion of \"condemned\" in the active voice : This encoding is based on Pareschi and Steedman's category notation [6] , wherein the res, arg, and dir features are used to encode a syntactic function' s result, argument, and direction of application . To reduce the complexity of defining tens of thousands of such lexical entries, we associate to each category type (such as the active voice of a transitive verb) a lexical skeleton, that is, a partial lexical entry that is fleshed out during the expansion of a stem . The fleshing out of skeletons does not actually occur until run time, when a word form is actually found in a document. Since category data structures are fairly substantial, this yields tremendous memory savings .",
"cite_spans": [
{
"start": 553,
"end": 557,
"text": "[11]",
"ref_id": null
},
{
"start": 856,
"end": 859,
"text": "[6]",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Linguistic Lexicon",
"sec_num": null
},
{
"text": "[[",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Linguistic Lexicon",
"sec_num": null
},
{
"text": "The lexicon was populated in part by adapting knowledge bases provided to us by Richard Tong of ADS .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Linguistic Lexicon",
"sec_num": null
},
{
"text": "For Muc4, we used a number of strategies for handling open classes of proper names . For geographical names, we relied primarily on a listing of such names that had been compiled by previous MUC participants, and which was forwarded to us by ADS . As a back up, we also encoded a small grammar of Spanish geographical names-for example, \"san\" has a reading as a functor that produces geographical names given a personal name on its right.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Geographical Names, Personal Names, and Unknown Words",
"sec_num": null
},
{
"text": "For personal names, we relied primarily on a cross-cultural database of 15,000 names obtained from variou s public domain databases. Most of these are first names, with only about 2,000 family names covered by th e database . In order to fill inevitable gaps in the database, we allowed unknown words to be given, among others, a definition as a personal proper name . Separately, we provided a grammatical type-shifting rule that turns personal names into functors that combine with any personal name on their right. In non-ccG terms, this amounts to a grammar rule of form :",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Geographical Names, Personal Names, and Unknown Words",
"sec_num": null
},
{
"text": "All the names in TST2-MUC4-0048 turned out to be in our database, in part because we had already extended i t with a list of VIP names provided by ADS .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "PERS-NAME -> PERS-NAME PERS-NAM E",
"sec_num": null
},
{
"text": "We chose to use categorial grammars in ALEMBIC for a number of reasons. First and foremost, we expected ou r syntactic coverage to be necessarily incomplete, and knew that we would have to rely heavily on bottom-up parsing . In this light, categorial grammars are effectively the uhr-bottom-up linguistic framework, as one canno t meaningfully speak of top-down parsing in this framework. We also wanted a framework that was strongly lexically governed, as in C(G's, in order to reduce the parsing search space. Finally, in anticipation of eventually wanting to provide fairly comprehensive coverage of syntax, we chose one of the recent mildly context sensitive frameworks, in the hope that we could exploit the linguistic strengths of the framework at some future point.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "CCG Parser",
"sec_num": null
},
{
"text": "Our current ccG parser is based upon Pareschi and Steedman's algorithm for left-to-right parsing of combinatorial categorial grammars [6] . Their approach is nominally only intended to produce full chart-based parses . Because we anticipated our syntactic coverage to be incomplete, we extended the basic algorithm into a heuristi c semi-parser . The semi-parser heuristics are used to provide a spanning segmentation of the input string .",
"cite_spans": [
{
"start": 134,
"end": 137,
"text": "[6]",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "CCG Parser",
"sec_num": null
},
{
"text": "In addition, we extended the CCG framework per se in order to cover optional and displaced arguments, which ar e typically weaknesses of traditional categorial grammar frameworks. The approach we've taken involves introducing a pair of features for each optional argument, one feature to encode the type of argument that is expected, and th e second to encode the disposition of the argument's semantics . For instance, consider the passive voice of a transitive verb-kill is a canonical example . A partial encoding would be as follows : The by-pp-prep feature indicates that the category can be modified by a PP headed by by; the by-pp-obj feature indicates that the embedded semantics of the PP is then indirectly unified with the semantic agent of the sentence .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "CCG Parser",
"sec_num": null
},
{
"text": "[[",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "CCG Parser",
"sec_num": null
},
{
"text": "At the time we fielded our system for MUC-4, both our syntactic coverage and semi-parsing heuristics were still very preliminary, and our overall parses were thus extremely fragmentary . For example, the first sentence in TST2-MUC4-0048 ended up being bracketed roughly as follows :",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "CCG Parser",
"sec_num": null
},
{
"text": "This particular bracketing illustrates several early shortcomings of our grammar, many of which have bee n addressed in the months since the MUC-4 evaluation. First, the MUC-4 version of the title sub-grammar was weak: title modifiers such as -elect or general were simply absent from the grammar . Second, prepositional phrase coverage was incomplete: PP's that appeared as optional arguments of categories would parse, but those that shoul d be treated as modifiers failed to do so. In addition, many verbs simply lacked the appropriate subcategorization frame for PP arguments, as in this case with accused. Finally, as with many semi-parsers, ALEMBIC currently punts on coordination.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "[SALVADORAN PRESIDENT] [-] [ELECT] [[ALFREDO CRISTIANI] CONDEMNE D [THE TERRORIST KILLING OF ATTORNEY]] [GENERAL ROBERTO GARCI A ALVARADO] [AND] [ACCUSED [THE FARABUNDO MARTI NATIONAL LIBERATION FRONT] ] [( ] [FMLN] [ ) ] [OF] [THE CRIME] [ . 1",
"sec_num": null
},
{
"text": "As might be gleaned from the category definitions given above, ALEMBIC produces semantic interpretations concurrently with parsing . The meaning representation language that we use is directly descended from our earlie r work on the King Kong interface [2] , whose underlying approach is similar to that in the core language engine [1] . Meaning representations are given at the so-called interpretation level, where quantifiers are not scoped with respec t to each other, but are simply left \"in place,\" i .e., attached to their associated noun phrases . For example, the interpretation of the fragment \"the terrorist killing\" in message TST2-MUC4-0048 is : ",
"cite_spans": [
{
"start": 253,
"end": 256,
"text": "[2]",
"ref_id": "BIBREF1"
},
{
"start": 332,
"end": 335,
"text": "[1]",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Semantic Interpretations",
"sec_num": null
},
{
"text": "[[",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Semantic Interpretations",
"sec_num": null
},
{
"text": "The approach we have taken towards reference resolution [3] attempts to integrate several disparate approache s towards the problem . The literature on reference resolution identifies a number of sources of linguistic evidence tha t can be applied towards resolving anaphoric references, but few attempts have been made at combining thes e evidence sources in a principled way (for an exception, see [5] ) . The approach embodied in our system attempts to perform the integration by exploiting a Bayesian belief network .",
"cite_spans": [
{
"start": 56,
"end": 59,
"text": "[3]",
"ref_id": "BIBREF2"
},
{
"start": 400,
"end": 403,
"text": "[5]",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Reference Resolution",
"sec_num": null
},
{
"text": "The network combines a number of evidence sources that bear upon whether an anaphor (either a definit e reference or a pronoun) can be resolved to a particular candidate referent . Because of the fragmentary nature of ou r parses, the reference resolution network only considered non-grammatical features of the anaphor and candidate . In particular, these included : We experimented with a number of such networks prior to the MUC-4 evaluation run, including hand-buil t networks and networks derived by machine learning algorithms . We ended up selecting a simple flat network i n which all evidence sources were directly allowed to bear upon the root node (which stood for coreference of th e anaphor and candidate) .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Reference Resolution",
"sec_num": null
},
{
"text": "To apply the network, our system first collects a set of mentions from the parsed document : these amount roughly to noun phrases and to event verbs with any arguments that might have been attached by the parser. Anaphoric mentions are then compared to mentions preceding them in the document . The comparison is performed by populating the evidence nodes of the network according to the characteristics of the anaphor and candidate . Mentions that are found to co-refer are grouped together and assigned to a unique discourse entity (called a peg) .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Reference Resolution",
"sec_num": null
},
{
"text": "In the case of our actual run on TST2-MUC4-0048, for example, the bracketing of the first sentence produced b y the semi-parser lead to identifying as mentions (among others) Alfredo Cristiani, and the murder event introduced b y the nominalization of \"kill .\" The second of these phrases was then taken as potentially anaphoric and compared t o earlier mentions in the sentence, including that for Cristiani . In the case, of Cristiani the mentions were found not to co-refer, reflecting the importance of KR compatibility . Nevertheless, the fragmentary nature of the parses, couple d with the relative lack of grammatical features in the Bayesian network, led to disapointing reference resolutio n performance overall . As we describe below and elsewhere in these proceedings, this led indirectly to our relativel y low precision scores .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Reference Resolution",
"sec_num": null
},
{
"text": "Once reference resolution has been performed, the system enters a MUC-4-specific processing phase . The first step towards producing templates consists of identifying significant violent events, which is performed by searching the document for event pegs whose semantic heads are subsumed by the KR node for violence. In our actual run on TST2-MUC4-0048, two such pegs were found in the first paragraph : one for \"killing\" and one for \"the crime.\" The fact that two separate pegs were found for these phrases reflects a failure on the part of our reference resolutio n mechanism, as these two phrases should properly have been determined to be co-referential .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Extraction of Significant Events and Template Generation",
"sec_num": null
},
{
"text": "In the MUC-4 version of ALEMBIC, the actual generation of templates is keyed off of the pegs for violent events .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Extraction of Significant Events and Template Generation",
"sec_num": null
},
{
"text": "Each such peg is taken to indicate a separate act of violence for which a template should be generated . This strategy is very straightforward, but it places a tremendous burden on the system's ability to identify coreferential events . As reference resolution is actually a weak point in the MUC-4 version of the system, this leads us to generate multipl e templates for what is really the same event. As a result, we paid a significant toll in our precision scores.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Extraction of Significant Events and Template Generation",
"sec_num": null
},
{
"text": "Turning to the slot-filling mechanism for a particular template, once a significant event peg has been identified , ALEMBIC then attempts to locate the participants in the event. Preferentially, the system attempts to use the syntactic arguments of the event phrase, if the parser succeeded in identifying them . In those cases where the parser failed to provide arguments to a verb or a nominalized event, the system assumes that the parse must have been fragmented , and attempts to locate potential arguments elsewhere in the sentence . This search is clearly heuristic and application-specific. In the case of the actual perpetrator of the event, the system attempts to fmd phrases with the appropriate agentive heads-this includes military organizations, terrorist organizations, and known terrorists. A similar process is performed to identify entities that might bear on other template slots . For example, targets an d instruments are identified by searching for phrases headed by KR relations denoting damage, injury, or weapons .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Extraction of Significant Events and Template Generation",
"sec_num": null
},
{
"text": "The heuristic nature of this process yields mixed results. At times it reunites arguments that had been separate d from their verbs due to fragmentary parsing, and at times it simply results in unprincipled filling of slots (wit h results that can be incorrect, and even humorous . )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Extraction of Significant Events and Template Generation",
"sec_num": null
},
{
"text": "The final step taken by the system towards analyzing a message is also the most uninteresting . Once the message has been fully analyzed, it is simply dumped back out to a file, along with any relevant markup of meaning analysis . In the case of the MUC-4 task, this amounts to associating SGML template tags to relevant paragraphs of text, or mor e simply, to ignoring the production of an SGML output file, and just printing the templates on their own .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Extraction of Significant Events and Template Generation",
"sec_num": null
},
{
"text": "As we mentioned at the beginning of this note, ALEMBIC is still in a very early stage of development . Although we are satisfied with the system's (fairly humble) performance given its relative youth, we are also painfully aware of the compromises performed in fielding a message processing system on such a tight development schedule . Many of the shortcuts we took are unsatisfyingly heuristic . In addition, many interesting ideas that seemed promising in paper studies were never included in the fielded system . It was our original intent, for example, to us e a completely different parsing algorithm that supports nearly-semantic parse rules ; this class of rules is related to the nearly-syntactic extraction rules of some recent Muc-class systems, e .g., FASTUS and CIRCUS . We had also intended to extend the semantic interpretation process with a terminological inference mechanism based on [9] . These modules were not implemented in time for Muc--nor were a host of other improvements detailed in ou r companion results and analysis paper . It is to these ideas that we now turn, in the expectation that the next versio n of ALEMBIC that we apply to the MUC data extraction task will dramatically outperform the version presented here .",
"cite_spans": [
{
"start": 902,
"end": 905,
"text": "[9]",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "CONCLUDING THOUGHTS",
"sec_num": null
}
],
"back_matter": [
{
"text": "Principal funding for this work was provided by the MITRE Corporation . Funding for our participation in MUC-4 was provided by the Defense Advanced Research Projects Agency through the Muc conference committee . W e would also like to express our gratitude to Richard Tong and his colleagues at ADS for providing us with thei r helpful lexicon and taxonomies . Finally, we would like to thank Beth Sundheim for her ongoing support .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgments",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Logical forms in the core language engine",
"authors": [
{
"first": "H",
"middle": [],
"last": "Alshawi",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Van Eijck",
"suffix": ""
}
],
"year": 1989,
"venue": "Proceedings of the 27th Annua l Meeting of the ACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Alshawi, H and van Eijck, J, \"Logical forms in the core language engine,\" in Proceedings of the 27th Annua l Meeting of the ACL, Vancouver, British Columbia. 1989\"",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "The relation-based knowledge representation of King Kong",
"authors": [
{
"first": "S",
"middle": [],
"last": "Bayer",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Vilain",
"suffix": ""
}
],
"year": 1991,
"venue": "SIGART Bulletin",
"volume": "2",
"issue": "3",
"pages": "15--21",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Bayer, S, and Vilain, M, \"The relation-based knowledge representation of King Kong,\" SIGART Bulletin 2(3) , 15-21 . 1991 .",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Probabilistic resolution of anaphoric reference",
"authors": [
{
"first": "",
"middle": [],
"last": "Burger",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Connolly",
"suffix": ""
}
],
"year": 1992,
"venue": "To appear in Proceedings of the 1992 AAA' Fall Symposium on Probabilistic Approaches to Natural Language",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Burger, J and Connolly, D, \"Probabilistic resolution of anaphoric reference, \" To appear in Proceedings of the 1992 AAA' Fall Symposium on Probabilistic Approaches to Natural Language, Boston, MA. 1992 .",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "GE : Description of the NLTOOLSET system as used in MUC -3",
"authors": [
{
"first": "G",
"middle": [],
"last": "Krupka",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Jacobs",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Rau",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Iwanska",
"suffix": ""
}
],
"year": 1991,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Krupka, G, Jacobs, P, Rau, L, and Iwanska, L, \"GE : Description of the NLTOOLSET system as used in MUC - 3,\" in [7] . 1991 .",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "A Computational Analysis of Context-dependent Referring Expressions , doctoral dissertation",
"authors": [
{
"first": "S",
"middle": [],
"last": "Luperfoy",
"suffix": ""
},
{
"first": "Discourse",
"middle": [],
"last": "Pegs",
"suffix": ""
}
],
"year": 1991,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Luperfoy, S, Discourse Pegs : A Computational Analysis of Context-dependent Referring Expressions , doctoral dissertation, Dept. of Linguistics, University of Texas at Austin . 1991 .",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "A lazy way to chart-parse with categorial grammars",
"authors": [
{
"first": "R",
"middle": [],
"last": "Pareschi",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Steedman",
"suffix": ""
}
],
"year": 1987,
"venue": "Proceedings of the 25th Annual Meeting of the ACL",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Pareschi R and Steedman, M, \"A lazy way to chart-parse with categorial grammars,\" in Proceedings of the 25th Annual Meeting of the ACL, Stanford, CA . 1987.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Proceedings of the Third Message Understanding Conference",
"authors": [
{
"first": "B",
"middle": [],
"last": "Sudheim",
"suffix": ""
}
],
"year": 1991,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sudheim, B, ed, Proceedings of the Third Message Understanding Conference, Morgan Kaufman. 1991 .",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Deduction as parsing: tractable classification in the KL-ONE framework",
"authors": [
{
"first": "M",
"middle": [],
"last": "Vilain",
"suffix": ""
}
],
"year": 1991,
"venue": "Proceedings of th e Ninth National Conference on Artificial Intelligence (AAA'91)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Vilain, M, \"Deduction as parsing: tractable classification in the KL-ONE framework,\" in Proceedings of th e Ninth National Conference on Artificial Intelligence (AAA'91), Anaheim, CA . 1991 .",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "BBN : Description of the PLUM system as used in MUC-3",
"authors": [
{
"first": "R",
"middle": [],
"last": "Weischedel",
"suffix": ""
},
{
"first": "D",
"middle": [],
"last": "Ayuso",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Boisen",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Ingria",
"suffix": ""
},
{
"first": "J",
"middle": [],
"last": "Palmucci",
"suffix": ""
}
],
"year": 1991,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Weischedel, R, Ayuso, D, Boisen, S, Ingria, R, and Palmucci J, \"BBN : Description of the PLUM system as used in MUC-3,\" in [7] . 1991 .",
"links": null
}
},
"ref_entries": {
"FIGREF1": {
"type_str": "figure",
"text": "Overall System Architectur e",
"uris": null,
"num": null
},
"FIGREF2": {
"type_str": "figure",
"text": "Agreement on number, person, and gender \u2022 Compatibility vis-a-vis the semantic hierarchie s \u2022 Recency \u2022 Reflexivity \u2022 Phrase type (pronominal, definite, or otherwise )",
"uris": null,
"num": null
},
"TABREF2": {
"type_str": "table",
"html": null,
"num": null,
"content": "<table><tr><td colspan=\"2\">head :KILL ]</td><td/></tr><tr><td colspan=\"4\">[args #( [[head :TERRORISM-AGENT ]</td></tr><tr><td/><td colspan=\"2\">[quant NIL] ]</td></tr><tr><td/><td colspan=\"2\">OBJ-VAR ) ]</td></tr><tr><td colspan=\"3\">[quant :DEFINITE] ]</td></tr><tr><td colspan=\"2\">[[head :ATTACK ]</td><td/></tr><tr><td>[arg s</td><td>#([[hea d</td><td colspan=\"2\">:TERRORISM-AGENT]] ... ) ]</td></tr><tr><td>[proxy</td><td colspan=\"2\">PROX179]</td></tr><tr><td colspan=\"2\">[mods ([[hea d</td><td>:TIME-OF ]</td></tr><tr><td/><td>[args</td><td>#(PROX17 9</td></tr><tr><td/><td/><td>[[head</td><td>:DATE ]</td></tr><tr><td/><td/><td>[proxy</td><td>PROX180 ]</td></tr><tr><td/><td/><td>[mods</td><td>([head :BEFORE ]</td></tr><tr><td/><td/><td/><td>[args #(PROX:180 *NOW* [[head :DAY ]</td></tr><tr><td/><td/><td/><td>[quant 5]])]}]])]])]]</td></tr></table>",
"text": "In addition, the representation maintains an implicitly Davidsonian representation of events and other relations . That is, aside from their underlying arguments, the relations may be modified through a proxy variable, as in th e following encoding of a later sentence in the message, \"guerillas attacked . . . five days ago . \""
}
}
}
} |