This paper describes a system for extracting typed dependency parses of English sentences from phrase structure parses. In order to capture inherent relations occurring in corpus texts that can be critical in real-world applications, many NP relations are
GeneratingTypedDependencyParsesfromPhraseStructureParses
Marie-CatherinedeMarneffe, BillMacCartney, andChristopherD.Manning
DepartmentofComputingScience,Universit´ecatholiquedeLouvain
B-1340Louvain-la-Neuve,Belgium
ComputerScienceDepartment,StanfordUniversity
Stanford,CA94305,USA
{mcdm,wcmac,manning}@stanford.edu
Abstract
ThispaperdescribesasystemforextractingtypeddependencyparsesofEnglishsentencesfromphrasestructureparses.Inordertocaptureinherentrelationsoccurringincorpustextsthatcanbecriticalinreal-worldapplications,manyNPrelationsareincludedinthesetofgrammaticalrelationsused.WeprovideacomparisonofoursystemwithMiniparandtheLinkparser.ThetypeddependencyextractionfacilitydescribedhereisintegratedintheStanfordParser,availablefordownload.
1.Introduction
Wedescribeasystemforautomaticallyextractingtypedde-pendencyparsesofEnglishsentencesfromphrasestruc-tureparses.Typeddependenciesandphrasestructuresaredifferentwaysofrepresentingthestructureofsen-tences:whileaphrasestructureparserepresentsnest-ingofmulti-wordconstituents,adependencyparserep-resentsdependenciesbetweenindividualwords.Atypeddependencyparseadditionallylabelsdependencieswithgrammaticalrelations,suchassubjectorindirectobject.Therehasbeenmuchlinguisticdiscussionofthetwofor-malisms.Thereareformalisomorphismsbetweencertainstructures,suchasbetweendependencygrammarsandonebar-level,headedphrasestructuregrammars(Miller,2000).Inmorecomplextheoriesthereissigni cantdebate:dom-inantChomskyantheories(Chomsky,1981)havede nedgrammaticalrelationsascon gurationsatphrasestructure,whileothertheoriessuchasLexical-FunctionalGrammarhasrejectedtheadequacyofsuchanapproach(Bresnan,2001).Ourgoalsherearemorepractical,thoughinessencewearefollowinganapproachwherestructuralcon gura-tionsareusedtode negrammaticalroles.
Recentyearshaveseentheintroductionofanumberoftreebank-trainedstatisticalparsers[Collins(Collins,1999),Charniak(Charniak,2000),Stanford(KleinandManning,2003)]capableofgeneratingparseswithhighaccuracy.Theoriginaltreebanks,inparticularthePennTreebank,wereforEnglish,andprovidedonlyphrasestructuretrees,andhencethisisthenativeoutputformatoftheseparsers.Atthesametime,therehasbeenincreasinginterestinusingdependencyparsesforarangeofNLPtasks,frommachinetranslationtoquestionanswering.Suchapplicationsbene- tparticularlyfromhavingaccesstodependenciesbetweenwordstypedwithgrammaticalrelations,sincethesepro-videinformationaboutpredicate-argumentstructurewhicharenotreadilyavailablefromphrasestructureparses.Per-hapspartlyasaconsequenceofthis,severalmorerecenttreebankshaveadopteddependencyrepresentationastheirprimaryannotationformat,evenifaconversiontoaphrasestructuretreeformisalsoprovided(e.g.,theDutchAlpinocorpus(vanderBeeketal.,2002)andtheDanishDepen-dencyTreebank(Kromann,2003)).However,existingde-
pendencyparsersforEnglishsuchasMinipar(Lin,1998)andtheLinkParser(SleatorandTemperley,1993)arenotasrobustandaccurateasphrase-structureparserstrainedonverylargecorpora.Thepresentworkremediesthisre-sourcegapbyfacilitatingtherapidextractionofgrammat-icalrelationsfromphrasestructureparses.Theextractionusesrulesde nedonthephrasestructureparses.
2.Grammaticalrelations
Thissectionpresentsthegrammaticalrelationsoutputbyoursystem.
Theselectionofgrammaticalrelationstoincludeinourschemawasmotivatedbypracticalratherthantheoreticalconcerns.Weusedasastartingpointthesetofgrammat-icalrelationsde nedin(Carrolletal.,1999)and(Kingetal.,2003).Thegrammaticalrelationsarearrangedinahi-erarchy,rootedwiththemostgenericrelation,dependent.Whentherelationbetweenaheadanditsdependentcanbeidenti edmoreprecisely,relationsfurtherdowninthehier-archycanbeused.Forexample,thedependentrelationcanbespecializedtoaux(auxiliary),arg(argument),ormod(modi er).Theargrelationisfurtherdividedintothesubj(subject)relationandthecomp(complement)relation,andsoon.ThewholehierarchyofourgrammaticalrelationsisgiveninFigure2.
Altogether,thehierarchycontains48grammaticalrela-tions.Whilethebackboneofthehierarchyisquitesimi-lartothatin(Carrolletal.,1999),overtimewehavein-troducedanumberofextensionsandre nementstofacil-itateuseinapplications.ManyNP-internalrelationsplayaveryminorroleintheoreticallymotivatedframeworks,butareaninherentpartofcorpustextsandcanbecriti-calinreal-worldapplications.Therefore,besidesthecom-monestgrammaticalrelationsforNPs(amod-adjectivemodi er,rcmod-relativeclausemodi er,det-determiner,partmod-participialmodi er,infmod-in nitivalmodi er,prep-prepositionalmodi er),ourhierarchyincludesthefollowinggrammaticalrelations:appos(appositivemodi- er),nn(nouncompound),num(numericmodi er),num-ber(elementofcompoundnumber)andabbrev(abbrevi-ation).Theexamplesentence“Billsonportsandimmi-grationweresubmittedbySenatorBrownback,Republican
Edges in a dependency tree may be c typed (...space of dependency parses and, with slight modi... extracted from treebank phrase-structure trees. ...
Edges in a dependency tree may be c typed (...space of dependency parses and, with slight modi... extracted from treebank phrase-structure trees. ...
热门文档