| { |
| "paper_id": "M95-1017", |
| "header": { |
| "generated_with": "S2ORC 1.0.0", |
| "date_generated": "2023-01-19T03:12:47.382870Z" |
| }, |
| "title": "UNIVERSITY OF SHEFFIELD : DESCRIPTION OF THE LaSIE SYSTE M AS USED FOR MUC-6", |
| "authors": [ |
| { |
| "first": "R", |
| "middle": [], |
| "last": "Gaizauskas", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "University of Sheffiel d", |
| "location": {} |
| }, |
| "email": "" |
| }, |
| { |
| "first": "T", |
| "middle": [], |
| "last": "Wakao", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "University of Sheffiel d", |
| "location": {} |
| }, |
| "email": "wkao@dcs.shef.ac.uk" |
| }, |
| { |
| "first": "K", |
| "middle": [], |
| "last": "Humphreys", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "University of Sheffiel d", |
| "location": {} |
| }, |
| "email": "" |
| }, |
| { |
| "first": "H", |
| "middle": [], |
| "last": "Cunningham", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "University of Sheffiel d", |
| "location": {} |
| }, |
| "email": "" |
| }, |
| { |
| "first": "Y", |
| "middle": [], |
| "last": "Wilks", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "University of Sheffiel d", |
| "location": {} |
| }, |
| "email": "" |
| } |
| ], |
| "year": "", |
| "venue": null, |
| "identifiers": {}, |
| "abstract": "The LaSIE (Large Scale Information Extraction) system has been developed at the University of Sheffiel d as part of an ongoing research effort into information extraction and, more generally, natural languag e engineering. LaSIE is a single, integrated system that builds up a unified model of a text which is then used t o produce outputs for all four of the MUC-6 tasks. Of course this model may also be used for other purposes aside from MUC-6 results generation, for example we currently generate natural language summaries of th e MUC-6 scenario results. Put most broadly, and superficially, our approach involves compositionally constructing semantic representations of individual sentences in a text according to semantic rules attached to phrase structure constituents which have been obtained by syntactic parsing using a corpus-derived context-free grammar. Th e semantic representations of successive sentences are then integrated into a `discourse model' which, once th e entire text has been processed, may be viewed as a specialisation of a general world model with which th e system sets out to process each text. LaSIE has a historical connection with the University of Sussex MUC-5 system [GCE93] from which it derives its approach to world modelling and coreference resolution and its approach to recombining fragmente d semantic representations which result from partial grammatical coverage. However, the parser and grammar differ significantly from those used in the Sussex system. In its approach to named entity identification LaSI E borrows to some extent from the approach adopted in the MUC-5 Diderot system [CGJ + 93]. Virtually al l of the code in LaSIE is new and has been developed since January 1995 with about 20 person-months o f effort. System Descriptio n Significant Feature s LaSIE is an integrated system, performing lexical, syntactic and semantic analysis to build a single, ric h representation of the text which is then used to produce the MUC-6 results for all four tasks. Features which distinguish the system are : \u2022 an integrated approach allowing knowledge at several linguistic levels to he applied to each MUC-6 task (e .g. coreference information is used in named entity recognition) ;", |
| "pdf_parse": { |
| "paper_id": "M95-1017", |
| "_pdf_hash": "", |
| "abstract": [ |
| { |
| "text": "The LaSIE (Large Scale Information Extraction) system has been developed at the University of Sheffiel d as part of an ongoing research effort into information extraction and, more generally, natural languag e engineering. LaSIE is a single, integrated system that builds up a unified model of a text which is then used t o produce outputs for all four of the MUC-6 tasks. Of course this model may also be used for other purposes aside from MUC-6 results generation, for example we currently generate natural language summaries of th e MUC-6 scenario results. Put most broadly, and superficially, our approach involves compositionally constructing semantic representations of individual sentences in a text according to semantic rules attached to phrase structure constituents which have been obtained by syntactic parsing using a corpus-derived context-free grammar. Th e semantic representations of successive sentences are then integrated into a `discourse model' which, once th e entire text has been processed, may be viewed as a specialisation of a general world model with which th e system sets out to process each text. LaSIE has a historical connection with the University of Sussex MUC-5 system [GCE93] from which it derives its approach to world modelling and coreference resolution and its approach to recombining fragmente d semantic representations which result from partial grammatical coverage. However, the parser and grammar differ significantly from those used in the Sussex system. In its approach to named entity identification LaSI E borrows to some extent from the approach adopted in the MUC-5 Diderot system [CGJ + 93]. Virtually al l of the code in LaSIE is new and has been developed since January 1995 with about 20 person-months o f effort. System Descriptio n Significant Feature s LaSIE is an integrated system, performing lexical, syntactic and semantic analysis to build a single, ric h representation of the text which is then used to produce the MUC-6 results for all four tasks. Features which distinguish the system are : \u2022 an integrated approach allowing knowledge at several linguistic levels to he applied to each MUC-6 task (e .g. coreference information is used in named entity recognition) ;", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Abstract", |
| "sec_num": null |
| } |
| ], |
| "body_text": [ |
| { |
| "text": "\u2022 the absence of any overt lexicon -lexical information needed for parsing is computed dynamicall y through part-of-speech-tagging and morphological analysis ;", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "\u2022 the use of a grammar derived semi-automatically from the Penn TreeBank corpus ;", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "\u2022 the use and acquisition of a world model, in particular for the coreference and scenario tasks ;", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "\u2022 a summarisation module which produces a brief natural language summary of scenario events . ", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "The high level structure of LaSIE is illustrated in Figure 1 . The system is a pipelined architectur e consisting of three principle processing stages : lexical preprocessing, parsing plus semantic interpretation , and discourse interpretation . Note that none of these high level system components corresponds directl y to any MUC-6 task . This reflects the fact that LaSIE has been designed as a general purpose informatio n extraction research system, initially geared towards, but not solely restricted to, MUC-6 tasks . Further , note that all MUC-6 results are generated after the most complete text representation has been built . Thi s reflects the desire to use information derived from all levels of linguistic processing in performing each of th e MUC-6 tasks .", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 52, |
| "end": 60, |
| "text": "Figure 1", |
| "ref_id": "FIGREF0" |
| } |
| ], |
| "eq_spans": [], |
| "section": "Architecture Overview", |
| "sec_num": null |
| }, |
| { |
| "text": "Input to the lexical preprocessor is a flat ASCII file containing a single Wall Street Journal article marke d up to the paragraph level in SGML . Output consists of two parts : a sequence of lexically seeded charts for input to the parser and a byte-offset/token representation of the initial text for later reconstruction with added markup in the results module .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Lexical Preprocessin g", |
| "sec_num": null |
| }, |
| { |
| "text": "Processing consists of tokenising and sentence-splitting the input, part-of-speech tagging the tokens , performing morphological analysis to obtain root forms, pattern-matching against precompiled lists of name d entities, and finally the creation of lexical or phrasal feature-structured edges for input to the parser .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Lexical Preprocessin g", |
| "sec_num": null |
| }, |
| { |
| "text": "Note that no conventional lexicon is used ; dynamic tagging and morphological analysis provide all of th e information required in the parser . Tokenization serves two purposes . Firstly token boundaries (as byte offsets into the text) are identifie d and each token is assigned an identifier (its number in the sequence that makes up the document) . These identifiers are preserved throughout the system and used with the associated offsets to facilitate allocatio n of SGML markup for NE and coreference annotation to byte positions in the MUC-6 results files .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Lexical Preprocessin g", |
| "sec_num": null |
| }, |
| { |
| "text": "Secondly, since the Brill tagger expects one sentence per line and expects each token to be separated b y white space, the text stream is changed into this format before it is fed to the tagger . BBN's POST program [WMS + 93] is run separately on the text as a sentence splitter to provide the sentence boundary information .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Lexical Preprocessin g", |
| "sec_num": null |
| }, |
| { |
| "text": "The Brill tagger [Bri94] is a rule-based part-of-speech tagger that has been extensively trained on Wall Street Journal Text . It uses the Penn Treebank tag set which consists of 48 part-of-speech tags [MSM93] . We hav e custom-configured the tagger in a number of ways . These changes include introducing new tags for dates , SGML markup, and for several punctuation symbols that are treated identically by the default tagger . We have also added several lexical and contextual rules to the tagger ' s rule base .", |
| "cite_spans": [ |
| { |
| "start": 17, |
| "end": 24, |
| "text": "[Bri94]", |
| "ref_id": "BIBREF0" |
| }, |
| { |
| "start": 202, |
| "end": 209, |
| "text": "[MSM93]", |
| "ref_id": "BIBREF8" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Part-of-Speech Tagging", |
| "sec_num": null |
| }, |
| { |
| "text": "Here is a sample from the walkthrough text of the transformed input stream after tokenizing and tagging . Each line corresponds to a single token and gives paragraph number, sentence number within paragraph , start byte-offset, end byte-offset, token, and part-of-speech tag of the token . ", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Part-of-Speech Tagging", |
| "sec_num": null |
| }, |
| { |
| "text": "Following tagging, all nouns and verbs are passed to a morphological analyser, which returns a root form fo r inclusion in the initial parser input . A set of 34 regular expression rules performs the analysis, in conjunctio n with a list of around 3000 irregular exceptions derived semi-automatically from the exception list used i n WordNet [Mi190] . The morphological analyser is implemented by compiling the irregulars and rules into a flex program which is translated to C .", |
| "cite_spans": [ |
| { |
| "start": 342, |
| "end": 349, |
| "text": "[Mi190]", |
| "ref_id": "BIBREF6" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Morphological Analysi s", |
| "sec_num": null |
| }, |
| { |
| "text": "Before parsing an attempt is made to identify and to tag named entity related phrases . This is done bot h by matching the input against pre-stored lists of proper names, date forms, currency names, etc . and by matching against lists of common nouns that act as reliable indicators or triggers for classes of named entity . These lists are compiled via flex program into a finite state recogniser . Each sentence is fed to the recognise r and all single and multi-word matches are used to associate token identifiers with named entity tags .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Named Entity Phrasal Taggin g", |
| "sec_num": null |
| }, |
| { |
| "text": "Lists of names are employed for locations, personal titles, organizations, dates/times and currencies . The following lists of names are used .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Named Entity Phrasal Taggin g", |
| "sec_num": null |
| }, |
| { |
| "text": "\u2022 Organization names : a cleaned up list originally from the Consortium for Lexical Research (CLR ) anonymous ftp site, containing about 2600 names .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Named Entity Phrasal Taggin g", |
| "sec_num": null |
| }, |
| { |
| "text": "\u2022 Company Designator : 'Co .','Ltd','PLC', etc . -94 designators based on the company designator list provided in the MUC6 reference resources .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Named Entity Phrasal Taggin g", |
| "sec_num": null |
| }, |
| { |
| "text": "\u2022 Titles : `President ' , `Mr .' -about 160 titles, manually collected .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Named Entity Phrasal Taggin g", |
| "sec_num": null |
| }, |
| { |
| "text": "\u2022 Human names : mainly first names, numbering about 500, based on a list of names in the Oxfor d Advanced Learner's Dictionary [Hor80] .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Named Entity Phrasal Taggin g", |
| "sec_num": null |
| }, |
| { |
| "text": "\u2022 Currency units : e .g . `dollars', `pounds', etc . -101 such unit names, taken from the MUC6 referenc e resources .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Named Entity Phrasal Taggin g", |
| "sec_num": null |
| }, |
| { |
| "text": "\u2022 Location names : names of major cities in the world as well as province/state and country names . Derived from a gazetteer list of about 150,000 place names by taking the highest level (`level 1') entrie s only -225 country, 1189 province, and 854 city names in total .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Named Entity Phrasal Taggin g", |
| "sec_num": null |
| }, |
| { |
| "text": "\u2022 Time expressions : phrases like `first quarter of' -49 phrases, manually constructed .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Named Entity Phrasal Taggin g", |
| "sec_num": null |
| }, |
| { |
| "text": "A trigger word is a word which indicates that the tokens surrounding it are probably a named entit y item and may reliably permit the type or even subtype of the named entity to be determined (e .g . company and government are subtypes of type organization) . The lists of trigger words were produced by hand .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Named Entity Phrasal Taggin g", |
| "sec_num": null |
| }, |
| { |
| "text": "\u2022 Location : 8 trigger words for location names, e .g . `Gulf',`Mountain' .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Named Entity Phrasal Taggin g", |
| "sec_num": null |
| }, |
| { |
| "text": "\u2022 Organization :", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Named Entity Phrasal Taggin g", |
| "sec_num": null |
| }, |
| { |
| "text": "-Government institutions : 7 trigger words for governmental institutions, e .g . `Agency',`Ministry' . For each sentence an initial chart is constructed which contains an edge for each lexical item and list-matche d named entity item in the sentence . The edges contain feature-value structures holding information abou t the items . The feature values are filled in from the information already present in the data stream (PO S tags, results of morphological analysis) . A P OS-tag-to-feature-value map database is used to associate on e or more feature-values with POS tags (e .g . the Penn Treebank tag NNS maps on to the features cat : noun , number :sing, person :3) .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Named Entity Phrasal Taggin g", |
| "sec_num": null |
| }, |
| { |
| "text": "The LaSIE parser is a simple bottom-up chart parser implemented in Prolog . The grammars it processes are unification-style feature-based context free grammars . During parsing semantic representations of constituents are constructed entirely using Prolog term unification . When parsing ceases, i .e . when the parser can generate no further edges, a `best parse selection' algorithm is run on the final chart to chose a singl e analysis . The semantics are then extracted from this analysis and passed on to the discourse interpreter .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Parsing", |
| "sec_num": null |
| }, |
| { |
| "text": "Parsing takes place in two passes, each using a separate grammar . In the first pass a special named entit y grammar is used, the sole purpose of which is to identify noun phrases relevant to the MUG-6 named entit y task . These constituents are then treated as unanalyzable during the second pass which uses a more genera l sentence' grammar .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Parsing", |
| "sec_num": null |
| }, |
| { |
| "text": "The grammar rules for Named Entity items constitute a subset of the system's noun phrase (NP) rules . All the rules were produced by hand . There are 206 such rules in total of which 94 are for organization, 54 for person, 11 for location, 18 for date/time, and 29 for money/percent expressions . The non-terminals LIST_LOCSIP, LIST_ORGAN1IP and CDGSIP are tags assigned to one or more input token s in the NE phrasal tagging stage of lexical preprocessing . The non-terminal NNP is the tag for proper name assigned to a single token by the Brill tagger .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Named Entity Grammar", |
| "sec_num": null |
| }, |
| { |
| "text": "The rule ORGANSIP --> NAMES_NP `&' NAMES_NP means that if an as yet unclassified or ambiguous prope r name (NAMES_NP) is followed by `&' and another ambiguous proper name, then it is an organization name . An example of this is \"Ammirati & Puris\" in the walkthrough text, which matches this pattern and is therefor e classified as an organization following NE parsing .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Named Entity Grammar", |
| "sec_num": null |
| }, |
| { |
| "text": "Nearly half of the NE rules are for organization names because they may contain any other proper name s (such as personal names, location names) as well as normal nouns, and their combinations . There are also a good number of rules for personal names since care needs to be taken of first names, family names, title s (e .g . `Mr .',`President'), and special lexical items such as `de' (as in ' J . Ignacio Lopez de Arriortua') an d etc . There are not so many rules for location names because they are recognized mainly in the previou s preprocessing stage by looking them up in the lists of city, province/state, country, and region names .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Named Entity Grammar", |
| "sec_num": null |
| }, |
| { |
| "text": "Rules for monetary and time expressions have been collected by analysing actual expressions in th e training texts .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Named Entity Grammar", |
| "sec_num": null |
| }, |
| { |
| "text": "The grammar used for parsing at the sentence level was derived from the Penn TreeBank-II (PTB-II ) [MSM93] , [MKM+ 95 ] . Since the PTB-II contains a large skeletally parsed corpus of Wall Street Journal articles, it seemed to us worth investigating as a potential source of a grammar for the MUC-6 tasks . Researc h into number and frequency distribution of rules in this corpus led to some surprising findings [Gai95a] . If a number of simplifying assumptions are made, a context-free grammar can be extracted from the PTB-I I WSJ corpus . Doing so led to an unmanageably large grammar : approximately 17,500 rules . However, only a small number rules account for the majority of rule occurrences . The following table illustrates the numbe r of rules in a grammar which accounts for the top n% of rule occurrences for each category which is eithe r an S or occurs as a nonlexical category on the right hand side of some other rule included in the grammar : Given the speed of our parser, the repair mechanisms for fragmentary parses in later parts of the system , and the difficulty of manually assigning semantic rules to large numbers of syntactic rules, we opted for th e 112 rule grammar representing 70 % of rule occurrences of the principal PTB constituent categories .", |
| "cite_spans": [ |
| { |
| "start": 99, |
| "end": 106, |
| "text": "[MSM93]", |
| "ref_id": "BIBREF8" |
| }, |
| { |
| "start": 109, |
| "end": 117, |
| "text": "[MKM+ 95", |
| "ref_id": null |
| }, |
| { |
| "start": 412, |
| "end": 420, |
| "text": "[Gai95a]", |
| "ref_id": "BIBREF2" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Sentence Grammar Rules", |
| "sec_num": null |
| }, |
| { |
| "text": "When parsing for a sentence is complete the resultant chart is analyzed to extract the `best parse' . Our algorithm for this was as follows : identify the set of syntactic categories for which useful standalone semantic s can be assigned -in our case S, NP, VP, and PP . Extract the set of shortest sequences of maximally spanning , non-overlapping edges of these categories . In the event of this set containing more than one member, pic k one arbitrarily and designate it the `best parse' . From the `best parse' the associated semantics are extracte d to be passed on to the discourse interpreter .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Sentence Grammar Rules", |
| "sec_num": null |
| }, |
| { |
| "text": "Semantic structures were assigned by hand to the set of rules automatically derived from the PTB-I I corpus . For simple verbs and nouns the morphological root is used as a predicate name in the semantics , and tense and number features are translated directly into the semantic representation where appropriate .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Sentence Grammar Rules", |
| "sec_num": null |
| }, |
| { |
| "text": "All NPs and VPs lead to the introduction of a unique instance constant in the semantics which serve s as an identifier for the object or event referred to in the text -e .g . company will map to something like company(e22) in the semantics and hired to hire (e34) , time(e34,past) . Each of these instance constant s is given a realisation property in the semantic representation, indicating, as a token range, the positio n in the text from which the semantics were derived . Nouns used as possessives or qualifiers also requir e realisation properties, necessitating the introduction of these categories into the grammar, although they did not occur in the original PTB-II ruleset . Each instance in the semantics is also augmented with furthe r realisation properties specifying the sentence number and paragraph number in which the instance' s surface token range occurs, and also whether the range is part of the header or the body of the article . These realisation properties provide back pointers from the semantics into the text and are necessary fo r writing out coreference markup and doing summarisation (they permit original surface forms to be used i n the summaries) . This requirement to go back from semantic representation to surface text was one of th e biggest innovations in our system that MUC-6 required .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Sentence Grammar Rules", |
| "sec_num": null |
| }, |
| { |
| "text": "A small set of hand constructed rules were also used in addition to those automatically derived from PTB -II, to extend the coverage of the grammar for particular constructions such as possessives . These additional rules are also used to combine verb-particle sequences into a compound form for use as a predicate name i n the semantics .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Sentence Grammar Rules", |
| "sec_num": null |
| }, |
| { |
| "text": "For example, the phrase \"stepping down as chief executive officer\" will be represented in the followin g form : step_down(e58) , time(e58,present) , realisation(e58,tokens(225,230)) , realisation(e58,sentence(10)), realisation(e58,section(4)), realisation(e58,type(body)) , as(e58,e60) , title(e60,'chief executive officer') , realisation(e60,tokens(228,230)) , realisation(e60,sentence(10)), realisation(e60,section(4)), realisation(e60,type(body) )", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Sentence Grammar Rules", |
| "sec_num": null |
| }, |
| { |
| "text": "The Discourse Interpreter module translates the semantic representation produced by the parser into a representation of instances, their ontological classes and their properties, using the XI knowledge representation language [Gai95b] . XI allows a straightforward definition of cross-classification hierarchies and th e association of arbitrary properties with classes or instances . These properties may be simple attribute-valu e associations, such as name(e1, 'PaineWebber ') or animat e(person(X) ,yes), or they may be rules which specify how the value of an attribute is to be derived from other information in the hierarchy, e .g . :", |
| "cite_spans": [ |
| { |
| "start": 226, |
| "end": 234, |
| "text": "[Gai95b]", |
| "ref_id": "BIBREF3" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Discourse Interpretatio n", |
| "sec_num": null |
| }, |
| { |
| "text": "orgsocale(el,e2) if in(el,e2) and e2 is an instance of the class locatio n XI provides a simple inheritance mechanism which allows properties to be inherited by classes or instance s lower in the hierarchy.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Discourse Interpretatio n", |
| "sec_num": null |
| }, |
| { |
| "text": "The definition of a cross-classification hierarchy is referred to as an ontology, and this together with an association of attributes with nodes in the ontology forms a world model. The basic ontology used for th e MUC-6 tasks is extremely simple, consisting of only 40 predefined object classes and 46 attribute types . For the scenario task, a hierarchy of 39 event classes and 9 additional attribute types were added . The The new nodes are currently only added as direct subclasses of either objects or events and so the numbe r of inheritable properties is extremely limited, but this mechanism does allow coreferences between instance s of a class previously unknown in the ontology, for instance two mentions of \"sailing\" in the walk-throug h article .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Ontology", |
| "sec_num": null |
| }, |
| { |
| "text": "As well as having attribute with atomic values, a node in the ontology may have attributes whose value s are specified by inference rules associated with it . The addition of an instance or property of a certain clas s to the world model while processing the text will lead to the evaluation of any inherited inference rules , potentially causing the addition of further instances or properties to the world model, or the reclassificatio n of existing instances . For example, the addition of name (e 1, 'PaineWebber') will cause the addition of e l as an instance of the object class, via the rule associated with the name property type which states that only objects can have name properties . We refer to this as presupposition expansion . A similar set of scenario specific inference rules are evaluated following coreference resolution in the consequence expansion stage .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Ontology", |
| "sec_num": null |
| }, |
| { |
| "text": "The semantic representation of a text is added to the world model sentence by sentence, with any presupposition expansion carried out immediately . After each sentence, all newly added instances are compared wit h previously added instances to determine whether any pair can be merged into a single instance, representin g a coreference in the text . The comparison of instances is carried out in several stages, terminating if a match is found during any one stage, as follows :", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Coreference Resolution", |
| "sec_num": null |
| }, |
| { |
| "text": "1. new instances with name properties are compared with all existing instances with name properties, i .e . named entity coreferences can range over the whole text ;", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Coreference Resolution", |
| "sec_num": null |
| }, |
| { |
| "text": "2. all new instances are compared with each other (intrasentential coreference resolution) ;", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Coreference Resolution", |
| "sec_num": null |
| }, |
| { |
| "text": "3. new pronoun instances are compared with existing instances from the same paragraph as the current input sentence, i .e . pronoun coreferences are intra-paragraph only, with an exception for paragrap h initial pronouns to allow reference to the previous paragraph ;", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Coreference Resolution", |
| "sec_num": null |
| }, |
| { |
| "text": "4. all other new instances are compared with existing instances from the current and previous paragraphs , i .e . all other coreferences are restricted to a span of two paragraphs .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Coreference Resolution", |
| "sec_num": null |
| }, |
| { |
| "text": "Each comparison involves first determining if there is a path between the instance s ' classes in the ontology. If no such path exists then the instances are on different branches of the ontology, and so such pairs are no t further considered for coreference . If a path does exist then the attributes of the instances are compare d to ensure no conflicts exist . Certain attributes, such as animate are defined as having unique fixed values for any instance and so instances with conflicting values for these attributes cannot be the same . If suc h conflicts are discovered then the comparison is abandoned . The name attribute is treated specially, usin g a semantic type specific name match (described below) to determine the compatibility of the newly inpu t instance's name with the longest known name of the existing instance .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Coreference Resolution", |
| "sec_num": null |
| }, |
| { |
| "text": "If no attribute conflicts are found between two instances, a similarity score is calculated based on th e number of common properties and on a semantic distance measure, determined simply in terms of the numbe r of nodes in the path between them . After a newly input instance has been compared with all others in a particular comparison set, it is merged in the world model with the instance with the highest similarity score, if one exists . In the case of equal scores for two or more previous instances, which is common in th e case of pronouns in the input, the most recent comparison, corresponding to the closest pair in the text, i s preferred .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Coreference Resolution", |
| "sec_num": null |
| }, |
| { |
| "text": "Coreference resolution for Named Entity items (proper names) is important in order to recognize variou s forms of proper names, especially organization names . For example, `Ford Motor Co.' may be used in a given text the first time the company is mentioned, and subsequently it may be referred to as `Ford' .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Name Matchin g", |
| "sec_num": null |
| }, |
| { |
| "text": "In order to determine whether given two proper names (organization, person, location) match or not , various heuristics are used, for example : two (multiword) names are judged the same if one name is an initial subsequence of the other .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Name Matchin g", |
| "sec_num": null |
| }, |
| { |
| "text": "There are 31 such heuristic rules for matching organization names, 11 heuristics for person names, an d three rules for location names .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Name Matchin g", |
| "sec_num": null |
| }, |
| { |
| "text": "Due to the use of capitalisation in article titles, the semantic representation produced by the parser is generally unreliable, with many capitalised words wrongly treated as proper names . For this reason processing of the header is delayed until after the body of the text, on the assumption that the true proper names in the header will also be mentioned, and more reliably detected, in the body. Proper names from the header which cannot be coreferred with anything in the body are then converted to normal predicate names and a further attempt made to find any coreferences .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Header Processing", |
| "sec_num": null |
| }, |
| { |
| "text": "The processing of a text acts to populate the initial bare world model with the various instances an d relations mentioned in the text . It is therefore converted into a world model specific to the particular text , i .e . a discourse model, containing the information necessary for the production of the results for all th e MUC-6 tasks, and other potential applications .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Discourse Model", |
| "sec_num": null |
| }, |
| { |
| "text": "The results for all four MUC-6 tasks are produced by scanning through the discourse model produced by th e discourse interpretation stage . Most of the semantic classes and property types in the predefined ontolog y are motivated by distinctions required by the various tasks, and for the TE and ST tasks specific ontologica l properties have been introduced for each slot required . The results generation therefore only involves th e retrieval from the discourse model of those instances which have all the required properties, and the correc t formatting of the property values .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Results Generatio n", |
| "sec_num": null |
| }, |
| { |
| "text": "All instances of the ontological classes of organisation, location, person, etc . are retrieved from the discourse model . For each of these that has an ne_tag property, the value of which is a token range, an entry i s added to an output file specifying the range and the required SGML markup type . The ne_tag propert y is introduced into the semantic representation of the text during the named entity stage of parsing . It is distinct from the realisat ion property, which also specifies a token range, because not all instances of th e required classes should be output for the NE result, for example location names within organisation names . The ne_tag property is only assigned to those instances that should be in the output . At the discours e interpretation stage, ambiguous names are also assigned ne_tag properties, but these names will only b e output in the NE result if they are subsequently classified as instances of one of the required classes .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Named Entity Result s", |
| "sec_num": null |
| }, |
| { |
| "text": "The file containing the list of token ranges to be tagged is then used in conjunction with the original tex t and the output of the lexical preprocessing stage, to produce a new version of the text with the require d SGML markup .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Named Entity Result s", |
| "sec_num": null |
| }, |
| { |
| "text": "The discourse model is searched for any instances of the object class which have more than one token rang e realisation property . This will only be the case where two distinct instances have been merged durin g discourse interpretation, resulting in a single instance with multiple realisations in the text . Instances of the event class also have realisat ion properties, but the MUC-6 task definition only requires coreferring nou n phrases to be identified .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Coreference Result s", |
| "sec_num": null |
| }, |
| { |
| "text": "An output file is written specifying all the token ranges included in each coreference chain . As with the NE result, this file is used in combination with the lexical preprocessor output to produce a new version o f the original text with the coreference SGML markup .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Coreference Result s", |
| "sec_num": null |
| }, |
| { |
| "text": "All organisation and person instances are retrieved from the discourse model, and those with name properties are formatted as required and written out directly to a results file . Property values for the other slots, such as ORG_LOCALE, are searched for by examining other related properties, such as being situated in a location , and the values output if found .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Template Element Result s", |
| "sec_num": null |
| }, |
| { |
| "text": "As for the template element result, this basically involves searching for instances which have values for th e required properties and then writing them directly to an output file in the required format . In this case w e require event instances of the type succession_event, which have values for the properties success ion_org , succession_post, etc . Each of these events is associated with at least two IN_AND_OUT objects (one I N and one OUT), also represented as instances in the discourse model . Output is only generated if at leas t one of these objects has values for its required properties .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Scenario Template Result s", |
| "sec_num": null |
| }, |
| { |
| "text": "The following table shows the scores for the four tasks . For the evaluation run the system processed 29 out of the 30 texts for the NE and CO tasks, and 98 out of 100 texts for the TE and ST tasks . This run produced the official scores, referred to here as ' incomplet e ' . Several texts were missed due to the omission of a trivial error trap that would have allowed the system to have continued at the next sentence on occurrence of a certain error, rather than at the next text . The system was re-run with the error trap included and th e results kindly scored by the MUC-6 scoring team . The results of this run are referred to in the table as complete' scores, and are, of course, unofficial .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "System Performanc e", |
| "sec_num": null |
| }, |
| { |
| "text": "Official and unofficial scores for the four tasks : Task NE CO TE S T R P P&R R P R P P&R While detailed evaluation of the contribution of system components to all tasks could not be undertake n before the conference, we have been able to partially analyse the behaviour of our NE subsystem . Th e table below illustrates the contribution of each of the system modules to the Named Entity task for the 3 0 (complete) NE texts . Setting 4 is the fully functional system setting .", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 52, |
| "end": 101, |
| "text": "Task NE CO TE S T R P P&R R P R P P&R", |
| "ref_id": null |
| } |
| ], |
| "eq_spans": [], |
| "section": "System Performanc e", |
| "sec_num": null |
| }, |
| { |
| "text": "Recall ", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "No Setting", |
| "sec_num": null |
| }, |
| { |
| "text": "The following table shows the scores of the LaSIE system for the walk through text . Our official scores for this text were well below our average across all texts due to failure to classify correctly one proper name , which led in turn to missing two of the succession events . After the evaluation we enhanced the discourse interpreter of the system with one specific feature, as described in the NE section below, and re-ran it o n the text . The scores for the enhanced system are also shown for comparison . Named Entity Tas k The score for NE task is 79 recall, 94 precision and 85 .91 for P&R . We missed one company name ('McCann-Erickson' and its abbreviation `McCann ' ) and one date expression . We recognized spuriously one company name ('Coca-Cola') and one person name . A company name was captured wrongly as person name P .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Walkthrough", |
| "sec_num": null |
| }, |
| { |
| "text": "Walter Thompson') .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Walkthrough", |
| "sec_num": null |
| }, |
| { |
| "text": "Coca-Cola' in `the prestigious Coca-Cola Classic account' was mistakenly recognized as company nam e since it is in the list of company names of the system and simply marked up as company at the list looku p stage .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Walkthrough", |
| "sec_num": null |
| }, |
| { |
| "text": "McCann-Erickson' was missed because the name itself does not have specific information to make it a company name and it remained an ambiguous proper name .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Walkthrough", |
| "sec_num": null |
| }, |
| { |
| "text": "The system correctly recognized the two shortened forms of `John J . Dooner Jr . ' ,`John Dooner ' and`Mr . Dooner' using the name matching algorithm in the coreference resolver .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Walkthrough", |
| "sec_num": null |
| }, |
| { |
| "text": "We enhanced our discourse interpreter so that it recognizes an ambiguous proper name as a compan y name when it is preceded by a post name(s) and `of', as in `chairman and chief executive officer of McCann-Erickson' . With this enhancement, which in turn permitted the correct coreference resolution with `McCann ' as its abbreviation, the score, especially recall went up : 94 recall, 95 precision and 94 .41 for P&R .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Walkthrough", |
| "sec_num": null |
| }, |
| { |
| "text": "Coreference scores for the walkthrough article are quite high compared to the overall system performance . The use of the enhanced system, which correctly recognises `McCann-Erickso n ' as a company name, rnake s little difference to the walkthrough score . One additional coreference is correctly made between the name s McCann-Erickson' and `McCann', due to the recognition as a company which invokes the organisatio n specific name matching, as described earlier .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Coreference Tas k", |
| "sec_num": null |
| }, |
| { |
| "text": "Obvious errors in the coreference include the omission of any appositions between person names an d post titles in the first paragraph . This is due to the simple fact that post titles were not specified as being instances of the class person in the ontology, and so they could not corefer because of the lack of any pat h between the two classes .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Coreference Tas k", |
| "sec_num": null |
| }, |
| { |
| "text": "Another error is the coreference of it at the beginning of the third paragraph, with Yesterday in th e previous paragraph . Only paragraph initial pronouns are allowed to have coreferents outside the curren t paragraph, but in this case the use of the pronoun is not to anything specifically mentioned in the text . There is a more general problem of the non-referring use of it, as in \"Mr . James says it is time . . . \", which is not treated specially in the system here, and the pronoun is simply coreferred with the closest potentia l candidate .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Coreference Tas k", |
| "sec_num": null |
| }, |
| { |
| "text": "Quotations also receive no specific treatment, leading to errors such as the I in \"I Can't Believe It's No t Butter\" being coreferred with the last person mentioned in the text, regardless of the fact that the quote d text is not attributed to anyone .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Coreference Tas k", |
| "sec_num": null |
| }, |
| { |
| "text": "Most of the heuristic rules for the coreference task, implemented via the properties of classes in the ontology, were produced from training on the MUC-6 dry-run articles, and few were subsequently modified . The restriction to only attempt coreferences within the two most recent paragraphs (apart from instance s with names), as described earlier, was introduced at a later stage to reduce the processing time of th e discourse interpretation phase, resulting in a predictable slight loss of recall with a corresponding increas e in precision .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Coreference Tas k", |
| "sec_num": null |
| }, |
| { |
| "text": "The score for TE task is 54 recall, 63 precision and 57 .89 for P&R . The score reflects, to large extent,th e successes and mistakes which are made at the NE task .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Template Element Tas k", |
| "sec_num": null |
| }, |
| { |
| "text": "TE task specific organization descriptors were not captured well in the system and this caused th e instantiation of two spurious organizations . As for locale and country slots, when a location name appears near an organization name, it will be associated with the organization . However, the two location names ar e missed for the text .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Template Element Tas k", |
| "sec_num": null |
| }, |
| { |
| "text": "As long as the names are correctly recognized, their aliases are all correctly recognized . `Coke' for `Coca-Cola ' is matched using a list of difficult alias names and `Mr . Dooner' is matched with its full name, `Joh n J . Dooner Jr .' accurately.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Template Element Tas k", |
| "sec_num": null |
| }, |
| { |
| "text": "One interesting name in the text is `J . Walter Thompson' . Clearly this is a person name however, it appears in the text as `. . . was hired from WPP Group ' s J. Walter Thompson last September' and here i t names a company. Our system does recognize it as person name but it does not change it to a compan y name .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Template Element Tas k", |
| "sec_num": null |
| }, |
| { |
| "text": "The system 's performance in this task for the walkthrough article is poor, producing a P&R score of onl y 21 .88 compared to the overall ST P&R of 49 .27 . However, with the enhanced version of the system P&R i s raised to 62 .65 for this article .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Scenario Template Task", |
| "sec_num": null |
| }, |
| { |
| "text": "The original poor result is due to the failure to identify the names `McCann-Erickson' and `J . Walter Thompson ' as company names . The use of the verb hire in the following piece of text Peter Kim was hired from WPP Group's J. Walter Thompson last September as vice chairman , chief strategy officer, world-wid e triggers the creation of a succession_event in the discourse model . This creates two in_and_out objects , with the IN object associated with `Peter Kim', because the verb hire is classified as an IN verb in th e ontology. The properties associated with this ontological class specify that the phrase \"as vice chairman \" following the verb indicates the succession post of this event . The properties of the hire event also specify that any \"by\" phrase would indicate the succession_org, but in this case no such phrase is found . Properties of the more general succession_event class are then used to introduce a dummy organisation int o the discourse model which the system then attempts to corefer in the same way as any other instance, in thi s case selecting the most recently mentioned company as the coreferent, and therefore the succession org , which, due to the omission of McCann, is in fact `Coca-Cola ' .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Scenario Template Task", |
| "sec_num": null |
| }, |
| { |
| "text": "succession events are also created in the discourse model for each of the verbs step_down, retire an d succeed in the second paragraph of the text . However, because McCann is not known to be a company , no succession_org is found for these events, and no output is produced, succession_org being one of the compulsory slots. In the enhanced version of the system, where McCann is recognised as a company , the succession_event created by retire does have values for all its properties, and it is output correctly. An attempt is made to merge this succession_event with the one created by the use of succeed, with th e intention of providing values for both the IN and OUT objects of the retire event, but the merge fails in thi s particular case .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Scenario Template Task", |
| "sec_num": null |
| }, |
| { |
| "text": "This approach to the scenario task involved the manual classification of 27 succession_event verbs , most of which had specific patterns associated with them, via the property mechanism, for establishing the person and organisation involved . A small set of general defaults could also be inherited by the events i f the specific patterns did not apply. The scenario patterns are heavily reliant on the accuracy of the name d entity recognition, although the ability is included to convert an ambiguous name in a specific verb pattern into a person name, and this could be extended further . This reliance results in the reasonable precisio n score of the system for this task, but it is at the expense of recall .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Scenario Template Task", |
| "sec_num": null |
| }, |
| { |
| "text": "An additional result produced by the system is a natural language summary of the text driven by the information contained in the scenario template . For each succession event found a simple sentence is produced , possibly referring to information in the discourse model which is not contained in the scenario templat e itself, for example: ", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Scenario Template Summarisatio n", |
| "sec_num": null |
| }, |
| { |
| "text": "Like most MUC systems, LaSIE was improving rapidly in performance up to and, as our enhancement s to the system for the walkthrough article show, after the final evaluation . So it is difficult to judge the limitations or still-to-be-realised potential of the underlying techniques .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Observations", |
| "sec_num": null |
| }, |
| { |
| "text": "What is clearly needed is yet more experimentation to determine just where the critical areas in performance are . In particular we need to assess the adequacy of the grammar and of our `best parse' selectio n algorithm : this could be done by doing parseval-style evaluations against the Penn Treebank . We also nee d to attempt to evaluate and improve the algorithms for automatically extending the ontology .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Observations", |
| "sec_num": null |
| }, |
| { |
| "text": "In our view MUC-6 has provided an extremely valuable increase in our understanding of informatio n extraction systems and the inter-relation of their components . We expect to be learning from our results for some time to come .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Observations", |
| "sec_num": null |
| } |
| ], |
| "back_matter": [ |
| { |
| "text": "This research has been made possible by the grants of the U .K . Department of Trade and Industry (Gran t Ref . YAE/8/5/1002) and the Engineering and Physical Science Research Council (Grant # GR/K25267) . We would like to thank BBN who have allowed us to use their POST program .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Acknowledgement s", |
| "sec_num": null |
| } |
| ], |
| "bib_entries": { |
| "BIBREF0": { |
| "ref_id": "b0", |
| "title": "Some advances in transformation-based part of speech tagging", |
| "authors": [ |
| { |
| "first": "E", |
| "middle": [], |
| "last": "Brill", |
| "suffix": "" |
| } |
| ], |
| "year": 1994, |
| "venue": "Proceedings of AAAI", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "E . Brill . Some advances in transformation-based part of speech tagging . In Proceedings of AAAI, 1994 .", |
| "links": null |
| }, |
| "BIBREF1": { |
| "ref_id": "b1", |
| "title": "Description of the Diderot system as used for MUC-5", |
| "authors": [ |
| { |
| "first": "]", |
| "middle": [ |
| "J" |
| ], |
| "last": "+ 93", |
| "suffix": "" |
| }, |
| { |
| "first": "L", |
| "middle": [], |
| "last": "Cowie", |
| "suffix": "" |
| }, |
| { |
| "first": "W", |
| "middle": [], |
| "last": "Guthrie", |
| "suffix": "" |
| }, |
| { |
| "first": "R", |
| "middle": [], |
| "last": "Jin", |
| "suffix": "" |
| }, |
| { |
| "first": "T", |
| "middle": [], |
| "last": "Wang", |
| "suffix": "" |
| }, |
| { |
| "first": "J", |
| "middle": [], |
| "last": "Wakao", |
| "suffix": "" |
| }, |
| { |
| "first": "S", |
| "middle": [], |
| "last": "Pustejovsky", |
| "suffix": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Waterman", |
| "suffix": "" |
| } |
| ], |
| "year": 1993, |
| "venue": "Proceedings of the Fifth Message Undersandin g Conference (MUC-5)", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "+ 93] J . Cowie, L . Guthrie, W . Jin, R. Wang, T . Wakao, J . Pustejovsky, and S . Waterman . Description of the Diderot system as used for MUC-5 . In Proceedings of the Fifth Message Undersandin g Conference (MUC-5). ARPA, Morgan Kaufmann, 1993 .", |
| "links": null |
| }, |
| "BIBREF2": { |
| "ref_id": "b2", |
| "title": "Investigations into the grammar underlying the Penn Treebank II", |
| "authors": [ |
| { |
| "first": "R", |
| "middle": [], |
| "last": "Gaizauskas", |
| "suffix": "" |
| } |
| ], |
| "year": 1995, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "R . Gaizauskas . Investigations into the grammar underlying the Penn Treebank II . Research Memorandum CS-95-25, Department of Computer Science, Univeristy of Sheffield, 1995 .", |
| "links": null |
| }, |
| "BIBREF3": { |
| "ref_id": "b3", |
| "title": "XI : A knowledge representation language based on cross-classification and inheritance", |
| "authors": [ |
| { |
| "first": "R", |
| "middle": [], |
| "last": "Gaizauskas", |
| "suffix": "" |
| } |
| ], |
| "year": 1995, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "R . Gaizauskas . XI : A knowledge representation language based on cross-classification and in- heritance . Research Memorandum CS-95-24, Department of Computer Science, Univeristy o f Sheffield, 1995 .", |
| "links": null |
| }, |
| "BIBREF4": { |
| "ref_id": "b4", |
| "title": "Description of the sussex system used for MUG -5", |
| "authors": [ |
| { |
| "first": "R", |
| "middle": [], |
| "last": "Gaizauskas", |
| "suffix": "" |
| }, |
| { |
| "first": "L", |
| "middle": [ |
| "J" |
| ], |
| "last": "Cahill", |
| "suffix": "" |
| }, |
| { |
| "first": "R", |
| "middle": [], |
| "last": "Evans", |
| "suffix": "" |
| } |
| ], |
| "year": 1993, |
| "venue": "Proceedings of the Fifth Message Undersanding Conference (MUC-5) . ARPA", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "R . Gaizauskas, L .J . Cahill, and R . Evans . Description of the sussex system used for MUG - 5 . In Proceedings of the Fifth Message Undersanding Conference (MUC-5) . ARPA, Morgan Kaufmann, 1993 .", |
| "links": null |
| }, |
| "BIBREF5": { |
| "ref_id": "b5", |
| "title": "Oxford Advanced Learner 's Dictionary of Current English", |
| "authors": [], |
| "year": 1980, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "A .S . Hornby, editor . Oxford Advanced Learner 's Dictionary of Current English . Oxford Uni- veristy Press, 1980 .", |
| "links": null |
| }, |
| "BIBREF6": { |
| "ref_id": "b6", |
| "title": "Wordnet : An on-line lexical database", |
| "authors": [ |
| { |
| "first": "G", |
| "middle": [], |
| "last": "Miller", |
| "suffix": "" |
| } |
| ], |
| "year": 1990, |
| "venue": "International Journal of Lexicography", |
| "volume": "3", |
| "issue": "4", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "G . Miller . Wordnet : An on-line lexical database . International Journal of Lexicography, 3(4) , 1990 .", |
| "links": null |
| }, |
| "BIBREF7": { |
| "ref_id": "b7", |
| "title": "The Penn Treebank : Annotating predicate argument structure . Distributed o n The Penn Treebank Release 2 CD-ROM by the Linguisitc Data Consortium", |
| "authors": [ |
| { |
| "first": "M", |
| "middle": [], |
| "last": "Marcus", |
| "suffix": "" |
| }, |
| { |
| "first": "G", |
| "middle": [], |
| "last": "Kim", |
| "suffix": "" |
| }, |
| { |
| "first": "M", |
| "middle": [ |
| "A" |
| ], |
| "last": "Marcinkiewicz", |
| "suffix": "" |
| }, |
| { |
| "first": "R", |
| "middle": [], |
| "last": "Maclntyre", |
| "suffix": "" |
| }, |
| { |
| "first": "A", |
| "middle": [], |
| "last": "Bies", |
| "suffix": "" |
| }, |
| { |
| "first": "M", |
| "middle": [], |
| "last": "Ferguson", |
| "suffix": "" |
| }, |
| { |
| "first": "K", |
| "middle": [], |
| "last": "Katz", |
| "suffix": "" |
| }, |
| { |
| "first": "B", |
| "middle": [], |
| "last": "Schasberger", |
| "suffix": "" |
| } |
| ], |
| "year": 1995, |
| "venue": "MKM + 95", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "MKM + 95] M . Marcus, G . Kim, M .A . Marcinkiewicz, R. Maclntyre, A . Bies, M . Ferguson, K .Katz, an d B . Schasberger . The Penn Treebank : Annotating predicate argument structure . Distributed o n The Penn Treebank Release 2 CD-ROM by the Linguisitc Data Consortium, 1995 .", |
| "links": null |
| }, |
| "BIBREF8": { |
| "ref_id": "b8", |
| "title": "Building a large annotated corpus o f english : The Penn Treebank", |
| "authors": [ |
| { |
| "first": "M", |
| "middle": [ |
| "P" |
| ], |
| "last": "Marcus", |
| "suffix": "" |
| }, |
| { |
| "first": "B", |
| "middle": [], |
| "last": "Santorini", |
| "suffix": "" |
| }, |
| { |
| "first": "M", |
| "middle": [ |
| "A" |
| ], |
| "last": "Marcinkiewicz", |
| "suffix": "" |
| } |
| ], |
| "year": 1993, |
| "venue": "Computational Linguistics", |
| "volume": "19", |
| "issue": "2", |
| "pages": "313--330", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "M .P. Marcus, B . Santorini, and M .A . Marcinkiewicz . Building a large annotated corpus o f english : The Penn Treebank . Computational Linguistics, 19(2) :313-330, 1993 .", |
| "links": null |
| }, |
| "BIBREF9": { |
| "ref_id": "b9", |
| "title": "Coping with ambiguity and unknown words through probabilistic models", |
| "authors": [ |
| { |
| "first": "M", |
| "middle": [], |
| "last": "Wms + 93] R . Weischedel", |
| "suffix": "" |
| }, |
| { |
| "first": "R", |
| "middle": [], |
| "last": "Meteer", |
| "suffix": "" |
| }, |
| { |
| "first": "L", |
| "middle": [], |
| "last": "Schwartz", |
| "suffix": "" |
| }, |
| { |
| "first": "J", |
| "middle": [], |
| "last": "Ramshaw", |
| "suffix": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Palmucci", |
| "suffix": "" |
| } |
| ], |
| "year": 1993, |
| "venue": "Computational Linguistics (Special Issue o n Using Large Corpora: II)", |
| "volume": "19", |
| "issue": "", |
| "pages": "359--382", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "WMS + 93] R . Weischedel, M . Meteer, R . Schwartz, L . Ramshaw, and J . Palmucci . Coping with ambiguity and unknown words through probabilistic models . Computational Linguistics (Special Issue o n Using Large Corpora: II), 19 :359-382, 1993 .", |
| "links": null |
| } |
| }, |
| "ref_entries": { |
| "FIGREF0": { |
| "type_str": "figure", |
| "uris": null, |
| "text": "LaSIE System Architectur e", |
| "num": null |
| }, |
| "FIGREF1": { |
| "type_str": "figure", |
| "uris": null, |
| "text": "Lexical Preprocessin gThe Tokenize r", |
| "num": null |
| }, |
| "FIGREF2": { |
| "type_str": "figure", |
| "uris": null, |
| "text": "16", |
| "num": null |
| }, |
| "FIGREF3": { |
| "type_str": "figure", |
| "uris": null, |
| "text": "Company : 138 trigger words for companies, e .g . `Airline', `Association' .The above names and key words are specially tagged as result of the list lookup stage, and are used in Named Entity grammar rules .", |
| "num": null |
| }, |
| "FIGREF4": { |
| "type_str": "figure", |
| "uris": null, |
| "text": "Parsing and Semantic InterpretationChart Seedin g", |
| "num": null |
| }, |
| "FIGREF5": { |
| "type_str": "figure", |
| "uris": null, |
| "text": "Here are some examples of the Named Entity grammar rules :", |
| "num": null |
| }, |
| "FIGREF6": { |
| "type_str": "figure", |
| "uris": null, |
| "text": "Discourse Interpretatio n ontology, together with the attributes associated with the nodes, were produced manually with the classe s and attributes being motivated mainly by the requirements of the MUC-6 tasks . During the processing o f a text, new classes of objects and events are automatically added to the ontology to enrich the hierarchy .", |
| "num": null |
| }, |
| "FIGREF7": { |
| "type_str": "figure", |
| "uris": null, |
| "text": "James steps down as chief executive officer of McCann-Erickson . Robert L . James will retire as chairman of McCann-Erickson . McCann-Erickson hired Peter Kim as vice chairman .", |
| "num": null |
| } |
| } |
| } |
| } |
