{ "paper_id": "M92-1031", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T03:13:02.155095Z" }, "title": "CRL/NMSU and Brandeis : Description of the MucBruce System as Used for MUC-4", "authors": [ { "first": "Jim", "middle": [], "last": "Cowie", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Louise", "middle": [], "last": "Guthrie", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Yorick", "middle": [], "last": "Wilks", "suffix": "", "affiliation": {}, "email": "" }, { "first": "James", "middle": [], "last": "Pustejovsky", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Scott", "middle": [], "last": "Waterma", "suffix": "", "affiliation": {}, "email": "" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "Through their involvement in the Tipster project the Computing Research Laboratory at New Mexic o State University and the Computer Science Department at Brandeis University are developing a method fo r identifying articles of interest and extracting and storing specific kinds of information from large volumes o f Japanese and English texts. We intend that the method be general and extensible. The techniques involve d are not explicitly tied to these two languages nor to a particular subject area. Development for Tipster ha s been going on since September, 1992. The system we have used for the MUC-4 tests has only implemented some of the features we pla n to include in our final Tipster system. It relies intensively on statistics and on context-free text markin g to generate templates. Some more detailed parsing has been added for a limited lexicon, but lack of fulle r coverage places an inherent limit on its performance. Most of the information produced in our MUC template s is arrived at by probing the text which surrounds `significant' words for the template type being generated , in order to find appropriately tagged fillers for the template fields .", "pdf_parse": { "paper_id": "M92-1031", "_pdf_hash": "", "abstract": [ { "text": "Through their involvement in the Tipster project the Computing Research Laboratory at New Mexic o State University and the Computer Science Department at Brandeis University are developing a method fo r identifying articles of interest and extracting and storing specific kinds of information from large volumes o f Japanese and English texts. We intend that the method be general and extensible. The techniques involve d are not explicitly tied to these two languages nor to a particular subject area. Development for Tipster ha s been going on since September, 1992. The system we have used for the MUC-4 tests has only implemented some of the features we pla n to include in our final Tipster system. It relies intensively on statistics and on context-free text markin g to generate templates. Some more detailed parsing has been added for a limited lexicon, but lack of fulle r coverage places an inherent limit on its performance. Most of the information produced in our MUC template s is arrived at by probing the text which surrounds `significant' words for the template type being generated , in order to find appropriately tagged fillers for the template fields .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "The overall system architecture is shown in Figure 1 . Three independent processes operate on an inpu t text . One, the Text Tagger, marks a variety of strings with semantic information . The other two, the Relevant Template Filter and the Relevant Paragraph Filter, perform word frequency analysis to determin e whether a text should be allowed to generate templates for particular incident types and which paragraph s are specifically related to each incident type . These predictions are used by the central process in th e system, the Template Constructor, which uses a variety of heuristics to extract template information fro m the tagged text . A skeleton template structure is then passed to the final process, the Template Formatter, which performs some consistency checking, creates cross references and attempts to expand any names foun d in the template to the longest form in which they occur in the text . Each of the above processes is described in more detail below .", "cite_spans": [], "ref_spans": [ { "start": 44, "end": 52, "text": "Figure 1", "ref_id": "FIGREF0" } ], "eq_spans": [], "section": "OVERVIEW OF THE TEMPLATE FILLING PROCES S", "sec_num": null }, { "text": "We have developed a procedure for detecting document types in any language . The system requires training texts for the types of documents to be classified and is developed on a sound statistical basis usin g probabilistic models of word occurrence [Guthrie and Walker 1991] . This may operate on letter grams o f appropriate size or on actual words of the language being targeted and develops optimal detection algorithm s from automatically generated \"word\" lists . The system depends on the availability of appropriate training texts . So far the method has been applied to English, discriminating between Tipster and MUC texts, an d to Japanese between Tipster texts and translations of ACM proceedings . In both cases the classification scheme developed was correct 99% of the time .", "cite_spans": [ { "start": 249, "end": 274, "text": "[Guthrie and Walker 1991]", "ref_id": "BIBREF2" } ], "ref_spans": [], "eq_spans": [], "section": "Relevancy Filters", "sec_num": null }, { "text": "The method has now been extended to the identification of relevant paragraphs and relevant templat e types for the MUC documents . This is a more complex problem due to the non-homogeneous nature of th e texts and the difficulty of deriving training sets of text . Each process uses two sets of words, one whic h occurs with high probability in the texts of interest, and the other which occurs in the `non-interesting ' texts . Due to the complexity of separating relevant from non-relevant information for the MUC texts w e actually use three filters, two trained on sets of non-relevant and relevant paragraphs and one trained o n sets of relevant and non-relevant texts . The lists of relevant and non-relevant paragraphs were derived using the templates of the 1300 text test corpus . Any paragraph which contributed two or more string fills to a particular template was used as part of the relevant training set ; paragraphs contributing only one string fill were regarded as of dubious accuracy and were not placed in either set and all other paragraphs wer e considered as non-relevant . Word lists were derived automatically by finding those words in the relevan t training set which occurred within a threshold of most frequently occurring words in the relevant paragraphs and not in the non-relevant paragraphs, and vice versa to obtain a set of non-relevant words .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Relevancy Filters", "sec_num": null }, { "text": "The relevant template marker consists of two processes, the first trained on a set of texts consistin g of paragraphs from the MUC corpus which produced two or more string fills against text consisting o f paragraphs which generated no string fills .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Relevancy Filters", "sec_num": null }, { "text": "These allow us to determine, based on word counts taken at paragraph level, whether the whole tex t should be checked for specific template types . The second stage is activated if any single paragraph in the text is found to be `relevant' . This stage is trained on the set of texts which generated a particular templat e type against texts which produced no templates . There are separate relevant and non-relevant lists of word s used to determine each template type .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Relevancy Filters", "sec_num": null }, { "text": "The result is a vector represented as a Prolog fact which determines whether the texts will be allowed t o generate templates of a particular type . Thus : ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Relevancy Filters", "sec_num": null }, { "text": "The relevant paragraph filter is the final stage and uses word lists which were derived from relevant an d non-relevant paragraphs for each template type .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "slot(4, ['NO', 'ARSON', 'NO', 'ATTACK', 'YES', 'BOMBING' , 'NO', 'KIDNAPPING', 'NO', 'ROBBERY', 'NO', 'DUMMY']) .", "sec_num": null }, { "text": "Once again this operates at the paragraph level and produces a list of paragraph numbers for eac h template type . These paragraph lists are only used if the relevant template filter has also predicted a template of that type . This stage produces a vector of relevant paragraphs . Thus : The two stages can be thought of as first distinguishing relevant texts for a particular template typ e from among all texts and second, given a relevant text, to distinguish between the relevant and non-relevan t paragraphs within that text for the template type .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "slot(4, ['NO', 'ARSON', 'NO', 'ATTACK', 'YES', 'BOMBING' , 'NO', 'KIDNAPPING', 'NO', 'ROBBERY', 'NO', 'DUMMY']) .", "sec_num": null }, { "text": "Partial word lists for relevant and non-relevant texts are given in Tables 1 and 2 . The full lists contain 124 and 117 words respectively . Partial relevant word lists for BOMBING at the text level (relevant template ) and the paragraph level are given in Tables 3 and 4 . The full lists contain 176 and 51 words respectively .", "cite_spans": [], "ref_spans": [ { "start": 68, "end": 82, "text": "Tables 1 and 2", "ref_id": "TABREF2" }, { "start": 257, "end": 271, "text": "Tables 3 and 4", "ref_id": "TABREF3" } ], "eq_spans": [], "section": "slot(4, ['NO', 'ARSON', 'NO', 'ATTACK', 'YES', 'BOMBING' , 'NO', 'KIDNAPPING', 'NO', 'ROBBERY', 'NO', 'DUMMY']) .", "sec_num": null }, { "text": "A key question for the Tipster and MUC tasks is the correct identification of place names, company an d organization names, and the names of individuals . We now have available to us several sources of geographic , company and personal name information . In addition the templates provided for MUC also supplied nam e information . These have been incorporated in a set of tagging files which provide lexical information as a pre-processing stage for every text .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Semantic Tagging", "sec_num": null }, { "text": "The details of the Text Tagger are shown in Figure 2 , which is a screen dump of an interface which allow s examination of the operation of each stage in the filter . The text window on the left shows the state of a text after the group dates process has converted dates to standard form and on the right after the temporary tags placed to identify date constituents have been removed . Each stage, apart from the last, marks the text with tags in the form :", "cite_spans": [], "ref_spans": [ { "start": 44, "end": 52, "text": "Figure 2", "ref_id": "FIGREF3" } ], "eq_spans": [], "section": "Semantic Tagging", "sec_num": null }, { "text": "Thus for example a date takes the form : In general each stage in the pipeline is only allowed to modify text which is not already marked, althoug h an examination of already marked text is allowed . Several stages also place temporary markers in the text For processing by the template constructor the final convert facts stage changes each sentence into a Prolog fact, containing sentence and paragraph numbers and a list of structures holding the marked item s Thus : . sen (3,3,[name(\"GARCIA ALVARADO\",null),',', num(\"86\",num(86)),',' , cs(\"WAS\",closed(was,[pastv] )), gls(\"KILLED\",action(killed,'ATTACK')) , cs (\"WHEN\",closed(when,[conj,pron] )), cs (\"A\",closed(a,[determiner] All the programs in the Tagger are written in `C' or Lex . We describe three of these components in mor e detail .", "cite_spans": [], "ref_spans": [ { "start": 477, "end": 570, "text": "(3,3,[name(\"GARCIA ALVARADO\",null),',', num(\"86\",num(86)),',' , cs(\"WAS\",closed(was,[pastv]", "ref_id": null }, { "start": 618, "end": 649, "text": "(\"WHEN\",closed(when,[conj,pron]", "ref_id": null }, { "start": 657, "end": 683, "text": "(\"A\",closed(a,[determiner]", "ref_id": null } ], "eq_spans": [], "section": "<\\TYPE> ACTUAL TEXT STRING {SEMANTIC INFORMATION} <\\ENDTYPE>", "sec_num": null }, { "text": "This program uses a large list of known strings which is held alphabetically . For each word in the text a binary search is performed on the list . When a match is found it will be with the longest string beginnin g with the word, subsequent words in the text are compared with the matched string . If the complete string i s matched then this portion of text is marked with the information associated with the string . If a complet e match is not achieved the word is checked against the previous item in the list, which may also match the word, and the process is repeated.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Known Item s", "sec_num": null }, { "text": "The strings and information in the file are derived from a variety of sources . The place name informatio n provided for MUC, organization, target and weapon names derived from the MUC templates and furthe r lists of human occupations and titles derived from Longman's .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Known Item s", "sec_num": null }, { "text": "The proper name filter uses a variety of methods to successfully identify a large majority of the huma n names found in a MUC text . It uses two data resources ; a complete word list of all the Longman Dictionar y headwords and a list of English and Spanish first and last names . In addition it uses the hidden Marko v Model algorithm described by BBN in MUC-3 to identify Spanish words . The first stage marks words no t in Longman's, Spanish words and known first and last names . The second stage decides whether a group of these items is indeed a name . Any group containing a Spanish word or a known name is recognized , unknown words on their own must be preceded by a title of some kind (identified by the Known Items step) . Once an unknown item is identified as a name, however, it is added temporarily to the list of first and las t names, so if it occurs in isolation later in the text it will be recognized correctly . A further complication to the problem of name recognition was found in several names which contained text which had already bee n identified as a place name . In this case the proper name marker over-rides the previous marking and marks the entire section of text as a human name .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Proper Names", "sec_num": null }, { "text": "The date marker uses a wide variety of patterns which have been identified in the MUC and Tipster texts a s referring to time . Each date is converted to a standard form and the identified text marked . Relative time expressions are always converted with reference to the headline date on the text . This assumption appears to be valid in the vast majority of cases we have examined .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Date Part s", "sec_num": null }, { "text": "The template constructor uses the tagged text and the list of relevant paragraphs for each template typ e to generate skeleton templates which are produced as a list of triples, SLOT NUMBER, SET FILL, STRIN G FILL . For example :", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Template Construction", "sec_num": null }, { "text": "[ [0 , 'TST2-MUC4-0048 ' ,null] ,", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Template Construction", "sec_num": null }, { "text": "A sequence of paragraphs is assumed to generate a new template . The sentences in these paragraphs are examined for a sentence containing a key verb for the template type . Sentences before this sentence ar e held in reverse order and sentences after in normal order . Each sentence is stripped of any prefatory claus e terminated by \"that\" (e .g . GOVERNMENT OFFICIALS REPORTED TODAY THAT) . The remainder of the sentence is reordered into lists containing texts marked with specific semantic types . These correspond to the appropriate fillers for the main sections of the template . The sentence is then marked as active or passive . A search is then made in the current sentence and either the previous or the succeeding ones fo r items satisfying the appropriate conditions to fill a template slot . Thus for an active sentence the perpetrator will be sought in the head of the sentence and then, if not found, in previous sentences . This provides a crude form of reference resolution as pronouns are not marked with any specific semantic information . The target is checked for in the tail of the sentence and then in subsequent sentences . This process is repeated for all the main fields of the template . It relies heavily on the fact that our text locating techniques are accurate . If no appropriate action word is found the template creation process is abandoned . The process is also abandoned if some of the template filling criteria are not satisfied (eg if the human target is a militar y officer) . The template construction program is written in Prolog and was compiled to run stand-alone usin g Quintus Prolog .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Template Construction", "sec_num": null }, { "text": "We obviously need to add more precise syntax and semantics at the sentence level and to provide a structure which allows the inter-relationship of a group of sentences to be captured . The advantage of the method we are using at the moment is that it is robust and can be used as a fall-back whenever the mor e precise methods fail . A limited amount of semantic parsing was implemented before the final MUC-4 test . This over-rode the robust method whenever an appropriate parse was found . Due to the limited number of lexical entries we were able to generate before the test, it was not possible to accurately assess the impac t of the more precise grammar .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Template Construction", "sec_num": null }, { "text": "Below are given sample entries of the lexical structures used in the MUC-4 tests. The transitive ver b murder and gerundive nominal killing illustrate the current state of the integration of lexical semanti c information (seen in the qualia field) with corpus-related information derived from tuning (seen in th e cospec field) [Pustejovsky 1991] . Cospecifacaiion is a semantic tagging of what collocational patterns th e lexical item may enter into . The sem field specifies directly how to map the qualia values into the appropriat e slots in the MUC templates . Parsing rules which allow indeterminate gaps are used to match the cospecification against the ke y sentences found . A parser-generator uses the cospec fields of the GLS's to construct the parsing rules, wit h type constraints obtained from the corresponding qualia fields . Certain operators within the rules (such as np() and \"*\") allow varying degrees of unspecified material to be considered in the constituents of the parse . The parsing rules can in this way be seen as specifying complex regular expressions . Because of thi s looseness, the parser will not break due to unknown items or intervening material .", "cite_spans": [ { "start": 328, "end": 346, "text": "[Pustejovsky 1991]", "ref_id": "BIBREF4" } ], "ref_spans": [], "eq_spans": [], "section": "Template Construction", "sec_num": null }, { "text": "These parsing rules are individually pre-compiled into compact Prolog code (each a small expressio n matching machine) before being included into the template constructor . The term-unification machinery of Prolog automatically relates the syntactic constituents of the parse with the type constraints from th e qualia and also with the arguments of the template semantics, avoiding the need for complex type matchin g and argument matching procedures .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Template Construction", "sec_num": null }, { "text": "Performance is degraded by the current partial implementation of the cospec field in the lexical structure definition . The statistical-based corpus-tuning program for the lexical structures was not included for th e MUC-4 test runs, but is on development-schedule for inclusion in the Tipster test run later this summer . The cospec for a lexical item ideally encodes corpus-based usage information for each semantic aspect of the word (e .g . its qualia, event type, and argument structure) . This is a statistically-encoded structure o f all admissible semantic collocations associated with the lexical item .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Template Construction", "sec_num": null }, { "text": "The initial seeding of the LS's is being done from lexical entries in the Longman Dictionary of Contemporary English [Proctor et al 1978] , largely using tools described in [Wilks et al 1990] . These are the n automatically adapted to the format of generative lexical structures . It is these lexical structures which ar e then statistically tuned against the corpus, following the methods outlined in [Pustejovsky 1992 ] and [Anic k and Pustejovsky 1990] . Semantic features for a lexical item which are missing or only partially specifie d from dictionary seeding are, where possible, induced from a semantic model of the corpus . ", "cite_spans": [ { "start": 117, "end": 137, "text": "[Proctor et al 1978]", "ref_id": "BIBREF3" }, { "start": 173, "end": 191, "text": "[Wilks et al 1990]", "ref_id": "BIBREF6" }, { "start": 402, "end": 419, "text": "[Pustejovsky 1992", "ref_id": "BIBREF5" }, { "start": 434, "end": 455, "text": "and Pustejovsky 1990]", "ref_id": "BIBREF1" } ], "ref_spans": [], "eq_spans": [], "section": "Template Construction", "sec_num": null }, { "text": "This final stage is also a Prolog program . This takes as input the lists of triples produced by the previou s stage and a list of every name found in the text . It then produces the final template, introducing cros s references between serially defined fields which are related to each other . The name list is used to attemp t to choose the fullest version of a name found in the text and substitute this for any shorter versions foun d in the template outline.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Template Formattin g", "sec_num": null }, { "text": "MucBruce generates four templates for this text . All are related to the vehicle bomb described at th e beginning of the text . The template and relevant paragraphs filters produce the following predictions : 4, ['NO', 'ARSON', 'NO', 'ATTACK', 'YES', 'BOMBING', 'NO' , 'KIDNAPPING', 'NO', 'ROBBERY', 'NO', 'DUMMY'] ) . rel_paras ([[1,3,5,6,13,18,19,20],'ARSON' , [1,2,3,4,5,6,7,8,9,10,11,12,13,14,16,17,18,19,20,21],'ATTACK' , [1,3,4,5,6,7,8,9,10,11,13,14,16,17,18,19,20],'BOMBING' , [1,3,6,7,16,17,20],'KIDNAPPING', [19,20],'ROBBERY', [],'DUMMY'] ) .", "cite_spans": [ { "start": 209, "end": 211, "text": "4,", "ref_id": null }, { "start": 212, "end": 218, "text": "['NO',", "ref_id": null }, { "start": 219, "end": 227, "text": "'ARSON',", "ref_id": null }, { "start": 228, "end": 233, "text": "'NO',", "ref_id": null }, { "start": 234, "end": 243, "text": "'ATTACK',", "ref_id": null }, { "start": 244, "end": 250, "text": "'YES',", "ref_id": null }, { "start": 251, "end": 261, "text": "'BOMBING',", "ref_id": null }, { "start": 262, "end": 268, "text": "'NO' ,", "ref_id": null }, { "start": 269, "end": 282, "text": "'KIDNAPPING',", "ref_id": null }, { "start": 283, "end": 288, "text": "'NO',", "ref_id": null }, { "start": 289, "end": 299, "text": "'ROBBERY',", "ref_id": null }, { "start": 300, "end": 305, "text": "'NO',", "ref_id": null }, { "start": 306, "end": 314, "text": "'DUMMY']", "ref_id": null } ], "ref_spans": [ { "start": 329, "end": 552, "text": "([[1,3,5,6,13,18,19,20],'ARSON' , [1,2,3,4,5,6,7,8,9,10,11,12,13,14,16,17,18,19,20,21],'ATTACK' , [1,3,4,5,6,7,8,9,10,11,13,14,16,17,18,19,20],'BOMBING' , [1,3,6,7,16,17,20],'KIDNAPPING', [19,20],'ROBBERY', [],'DUMMY']", "ref_id": "FIGREF0" } ], "eq_spans": [], "section": "TST2-MUC4-004 8", "sec_num": null }, { "text": "slot(", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "TST2-MUC4-004 8", "sec_num": null }, { "text": "This means that only 4 BOMBING templates will be produced . The first of these produces a reasonably complete match to the key ; details on the driver and bodyguards are omitted . The remaining three template s are incorrect, carrying only the information that a bombing has taken place. The attack on the home i s not identified by our naive method of multiple template generation, as it already occurs in a sequence o f paragraphs in which only the first event is found .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "TST2-MUC4-004 8", "sec_num": null }, { "text": "We feel that our present system, given its only partially completed state, shows potential . In particular th e following techniques seem generally useful :", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "CONCLUSION S", "sec_num": null }, { "text": "\u2022 The recognition of text types and sub-texts within a text using statistical techniques trained on larg e numbers of sample texts .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "CONCLUSION S", "sec_num": null }, { "text": "\u2022 The use of the key templates to derive system lexicons .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "CONCLUSION S", "sec_num": null }, { "text": "\u2022 The automatic seeding of lexical structures from machine readable dictionaries .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "CONCLUSION S", "sec_num": null }, { "text": "\u2022 The use of lexically-driven cospecification to provide a robust parsing method at the sentence level .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "CONCLUSION S", "sec_num": null }, { "text": "\u2022 The successful combination of a variety of techniques in the human name recognizer .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "CONCLUSION S", "sec_num": null }, { "text": "\u2022 The production of a number of independent tools for tagging texts .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "CONCLUSION S", "sec_num": null }, { "text": "The system is robust and provides a good starting point for the application of more sophisticated techniques . Given appropriate data it should be possible to produce a similar system for a different domain in a matter of weeks . The tagger software is already being adapted to Japanese and we have already establishe d that we can achieve similar performance with the statistical methods for Japanese texts using characte r bigrams .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "CONCLUSION S", "sec_num": null } ], "back_matter": [ { "text": "The system described here has been created using work funded by DARPA under contract number MDA904-91-C-9328 . The following colleagues at CRL and Brandeis have contributed time, ideas, programming ability and enthusiasm to the development of the MucBruce system ; Federica Busa, Peter Dilworth, Ted Dunning , Eric Eiverson, Steve Helmreich, Wang Jin, Fang Lin, Bill Ogden, Gees Stein, and Takahiro Waka o", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "ACKNOWLEDGEMENTS", "sec_num": null } ], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "THE FARABUNDO MARTI NATIONAL LIBERATION FRONT", "authors": [ { "first": "'el Salvador : San", "middle": [], "last": "Salvador", "suffix": "" } ], "year": null, "venue": "", "volume": "18", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "1, '6' ,null] , [4, 'ATTACK ',null] , [2,'19 APR 89',null] , [3,'EL SALVADOR : SAN SALVADOR (CITY)',null] , [6, 'null' ,\"BOMB\"] , [7, 'BOMB' ,null] , [18, 'null' , \"ROBERTO GARCIA ALVARADO\"] , [8, 'TERRORIST ACT' ,null] , [9, 'null' ,\"TERRORIST\"] , [10, 'null ' , \"THE FARABUNDO MARTI NATIONAL LIBERATION FRONT\"] , [12, 'null' ,\"VEHICLE\"] , [13, 'TRANSPORT VEHICLE' ,null] , [19, 'null' ,\"GENERAL\"] , [20, 'MILITARY' ,null] , [21, 'null' ,null] , [5, 'ACCOMPLISHED' ,null] , [16,'-',null] , [23, 'DEATH' ,null] ]", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "An Application of Lexical Semantics to Knowledge Acquisitio n from Corpora", "authors": [ { "first": "Peter", "middle": [], "last": "Anick", "suffix": "" }, { "first": "J", "middle": [], "last": "Pustejovsky", "suffix": "" } ], "year": 1990, "venue": "Proceedings of Coling 90", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Anick, Peter and Pustejovsky, J . (1990) . An Application of Lexical Semantics to Knowledge Acquisitio n from Corpora. Proceedings of Coling 90, Helsinki, Finland .", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "Some Comments on Document Classification by Machine", "authors": [ { "first": "Louise", "middle": [], "last": "Guthrie", "suffix": "" }, { "first": "Elbert", "middle": [], "last": "Walker", "suffix": "" } ], "year": 1991, "venue": "Computer and Cognitive Science", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Guthrie, Louise and Elbert Walker (1991) . Some Comments on Document Classification by Machine . Mem- orandum in Computer and Cognitive Science, MCCS-92-935, Computing Research Laboratory, New Mexic o State University, New Mexico .", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "Longman Dictionary of Contemporary English", "authors": [ { "first": "Paul", "middle": [], "last": "Proctor", "suffix": "" }, { "first": "Robert", "middle": [ "F" ], "last": "Ilson", "suffix": "" }, { "first": "John", "middle": [], "last": "Ayto", "suffix": "" } ], "year": 1978, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Proctor, Paul, Robert F . Ilson, John Ayto, et al . (1978) . Longman Dictionary of Contemporary English , Longman Group Limited : Harlow, Essex, England .", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "The Generative Lexicon", "authors": [ { "first": "James", "middle": [], "last": "Pustejovsky", "suffix": "" } ], "year": 1991, "venue": "Computational Linguistics", "volume": "17", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Pustejovsky, James (1991) \"The Generative Lexicon,\" Computational Linguistics, 17 .4, 1991 .", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "The Acquisition of Lexical Semantic Knowledge from Large Corpora", "authors": [ { "first": "James", "middle": [], "last": "Pustejovsky", "suffix": "" } ], "year": 1992, "venue": "Proceedings of the DARPA Spoken and Written Language Workshop", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Pustejovsky, James (1992) \"The Acquisition of Lexical Semantic Knowledge from Large Corpora \" , in Pro- ceedings of the DARPA Spoken and Written Language Workshop, Arden House, New York, February, 1992 , Morgan Kaufmann .", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "Providing Machin e Tractable Dictionary Tools", "authors": [ { "first": "Y", "middle": [], "last": "Wilks", "suffix": "" }, { "first": "D", "middle": [], "last": "Fass", "suffix": "" }, { "first": "", "middle": [], "last": "C-M", "suffix": "" }, { "first": "", "middle": [], "last": "Guo", "suffix": "" }, { "first": "J", "middle": [ "E" ], "last": "Mcdonald", "suffix": "" }, { "first": "T", "middle": [], "last": "Plate", "suffix": "" }, { "first": "B", "middle": [ "M" ], "last": "Slator", "suffix": "" } ], "year": 1990, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Wilks, Y ., Fass, D ., C-M ., Guo, McDonald, J . E ., Plate, T . and Slator, B .M . 1990 . \"Providing Machin e Tractable Dictionary Tools,\" in Machine Translation, 5 .1, 1990 .", "links": null } }, "ref_entries": { "FIGREF0": { "uris": null, "type_str": "figure", "num": null, "text": "MucBruce -System Overvie w" }, "FIGREF2": { "uris": null, "type_str": "figure", "num": null, "text": "<\\date> 5 DAYS AGO {date(\"14 APR 89\",890414)} <\\enddate >" }, "FIGREF3": { "uris": null, "type_str": "figure", "num": null, "text": "MucBruce -Tagging Pipeline to allow subsequent grouping by following stages . These temporary markers are removed by the filter stages . Each text is marked as follows : Known Items Places, organizations, physical targets, human occupations, weapons . Proper Names Human proper names . Dates All standard date forms and other references to time. Closed class Prepositions, determiners and conjunctions . Residue All other words are marked as unknown . The final tagged text looks like this : <\\name> GARCIA ALVARADO <\\endname>, <\\num> 56 {num(56)} <\\endnum> , <\\cs> WAS {closed(was,[pastv])} <\\endcs> <\\gls> KILLED {action(killed,'ATTACK')} <\\endgls> <\\cs> WHEN {closed(when,[conj,pron])} <\\endcs> <\\cs> A {closed(a,[determiner]) } <\\endcs> <\\weapon> BOMB {type(['BOMB'])} <\\endweapon> <\\res> PLACE D {atom(placed)} <\\endres > <\\cs> BY {closed(by,[prep])} <\\endcs> <\\res> URBAN {atom(urban)} <\\endres > <\\organ> GUERRILLAS {type(['TERRORIST', 'NOUN' ])} <\\endorgan> <\\cs > ON {closed(on,[prep])} <\\endcs> <\\cs> HI S {closed(his,[determiner,pron])} <\\endcs> <\\target> VEHICL E {type(['TRANSPORT VEHICLE'])} <\\endtarget> <\\gls> EXPLODE D {action(exploded,'BOMBING')} <\\endgls> <\\cs> A S {closed(as,[conj,pron,prep])} <\\endcs> <\\cs> IT {closed(it,[pron]) } <\\endcs> <\\res> CAME {atom(came)} <\\endres > <\\cs> TO {closed(to,[prep])} <\\endcs> <\\cs> A {closed(a,[determiner]) } <\\endcs> <\\res> HALT {atom(halt)} <\\endres > <\\cs> AT {closed(at,[prep])} <\\endcs> <\\cs> AN {closed(an,[determiner])} <\\endcs > <\\res> INTERSECTION {atom(intersection)} <\\endres> <\\cs> IN {closed(in,[prep]) } <\\endcs> <\\res> DOWNTOWN {atom(downtown)} <\\endres> <\\place> SAN SALVADO R {type([['CITY','EL SALVADOR'],['DEPARTMENT','EL SALVADOR']])} <\\endplace> ." }, "FIGREF4": { "uris": null, "type_str": "figure", "num": null, "text": "\",closed(his,[determiner,pron])), target(\"VEHICLE\",type(['TRANSPORT VEHICLE'])) , gis(\"EXPLODED\",action(exploded,'BOMBING')), cs(\"AS\",closed(as,[conj,pron,prep])) , cs(\"IT\",closed(it,[pron])), res(\"CAME\",atom(came)) , cs(\"TO\",closed(to,[prep])), cs(\"A\",closed(a,[determiner])) , res(\"HALT\",atom(halt)), cs(\"AT\",closed(at,[prep])) , cs(\"AN\",closed(an,[determiner])), res(\"INTERSECTION\",atom(intersection)) , cs(\"IN\",closed(in,[prep])), res(\"DOWNTOWN\",atom(downtown)) , place(\"SAN SALVADOR\",type([['CITY','EL SALVADOR'],['DEPARTMENT','EL SALVADOR']])),' .']) ." }, "FIGREF5": { "uris": null, "type_str": "figure", "num": null, "text": "self , \"*\" , \"WITH\" , np(I1)])]) , sem([type ('AMOK '),perp(H1),hum_tgt(H2),last (I1),hum_tgt_eff('DEATH')]" }, "FIGREF6": { "uris": null, "type_str": "figure", "num": null, "text": "TGT : EFFECT OF INCIDENT DEATH : \"ROBERTO GARCIA ALVARADO \" 24 . HUM TGT : TOTAL NUMBER * _Table 5 : One of Four Templates Generated for TST2-MUC4-004 8" }, "TABREF0": { "type_str": "table", "num": null, "text": "A FLAG FROM THE <\\organ> MANUEL RODRIGUEZ' PATRIOTIC FRONT Itvpef[TERRORIST '; ' NAME' DI <\\endorgan> (<\\organ> FPMR Itypef[TERRORIST ', ' NAME 111 <\\endorgan> ) WAS FOUN D AT THE SCENE OF THE EXPLOSION. THE<\\organ> FPMR Itypet(TERRORIST, 'NAME' DI <\\endorgan> IS A CLANDESTINE LEFTIS T <\\organ GROUP ltypei('OTHER NOUN Di <\\endorgan> THAT PROMOTES \"ALL FORMS O F STRUGGLE\"AGAINST THE <\\organ> MILITAR Y", "content": "
Input file: TST2-:MLC.-0002StartOverview : : Qui t
Tagger
i Know m
+ I Itemsz
Inpu t
Itype((' MILITARY: 'NOUN'. Dl <\\endorgan <\\organ> GOVERNMENT Itype(CGOVERNMENT' , 'NOUN ' <\\human> POLICE Itype(('LAW ENFORCEM E NT' REPORTED THAT THE EXPLOSION CAUSED SERIOUS .'NOUN'DI <\\endhuman> SOURCES HAV EA FLAG FROM THE <\\organ> MANUEL RODRIGUEZ PATRIOTIC FRONT Itypel[TERRORIST , ' NAME' DI <\\endorgan> (<\\organ> FP'I R \u00a3typet(TERRORIST. .'NAME' DI <\\endorgan> ) WAS FOUN D AT THE SCENE OF THE EXPLOSION. THE <\\organ > FPMR type((TERRORIST '.'NAME DI<\\endorgan > Properl IS A CLANDESTINE LEFTIST Names <\\organ> GROUP Itypeb1OTHEK, iNOUN`DI <\\endorgan> THAT PROMOTES 'ALL FORMS O F STRUGGLE\" AGAINST THE <\\organ>MILITARY \u00a3type(CM ILI TARY' . 'NOUN' DI <\\endorgan> <\\organ> GOVERNMENT ItypelrGOVERNMENT , 'NOUN' DI <\\endorgan> esuHEADED= BY <\\human GENERAL Itypef[' MILITARY', 'NOUN; ' RANK 'DI <\\endhuman> -.nAUOUSTO= =suPINOCHET ..
", "html": null }, "TABREF2": { "type_str": "table", "num": null, "text": "", "content": "
: Part of Non-Relevant Text Word Lis t
FREQUENCYWORD
BOM B
EXPLOSIO N
INJURE D
EXPLODED
DYNAMIT E
CA R
BOMBS
STREET
PLACE D
DAMAGED
", "html": null }, "TABREF3": { "type_str": "table", "num": null, "text": "", "content": "
: Part of Relevant Template Word List : BOMBIN G
", "html": null }, "TABREF4": { "type_str": "table", "num": null, "text": "", "content": "", "html": null }, "TABREF5": { "type_str": "table", "num": null, "text": "....... . ...... . ....... ................... .. ........ ....... ......", "content": "
\" MUCBr uca' [8L-NMSU/Brandei s
Releven t
Templates
Template
Tagge rFormate r
(click to view)
", "html": null } } } }