This paper presents a method of improving the quality of subcategorization frames (SCFs) acquired from corpora in order to augment a lexicon of a lexicalized grammar. We first estimate a confidence value that a word can have each SCF, and create an SCF con

GeneralizingSubcategorizationFramesAcquiredfromCorpora

UsingLexicalizedGrammars

NaokiYoshinaga UniversityofTokyo7-3-1Hongo,Bunkyo-ku,Tokyo,113-0033Japan

yoshinag@is.s.u-tokyo.ac.jp

Jun’ichiTsujii CREST,JST

4-1-8,Honcho,Kawaguchi-shi,Saitama,332-0012Japan

tsujii@is.s.u-tokyo.ac.jp

Abstract

Thispaperpresentsamethodofimprovingthequalityofsubcategorizationframes(SCFs)ac-quiredfromcorporainordertoaugmentalexi-conofalexicalizedgrammar.We rstestimateacon dencevaluethatawordcanhaveeachSCF,andcreateanSCFcon dence-valuevec-torforeachword.SincetheSCFcon dencevectorsobtainedfromthelexiconofthetar-getgrammarinvolveco-occurrencetendencyamongSCFsforwords,wecanimprovethequalityoftheacquiredSCFsbyclusteringvec-torsobtainedfromtheacquiredSCFlexiconandthelexiconofthetargetgrammar.Weap-plyourmethodtoSCFsacquiredfromcorporabyusingasubsetoftheSCFlexiconoftheXTAGEnglishgrammar.Acomparisonbe-tweentheresultingSCFlexiconandtherestofthelexiconoftheXTAGEnglishgrammarre-vealsthatwecanachievehigherprecisionandrecallcomparedtonaivefrequencycut-off.

withlexicalizedgrammars,becauseempiricalparsingef- ciencyandsyntacticambiguityoflexicalizedgrammarsareknowntobeproportionaltothenumberoflexicalen-triesusedinparsing(Sarkaretal.,2000).WethereforeneedsomemethodtoimprovethequalityoftheacquiredSCFs.

SchulteimWaldeandBrew(2002)andKorho-nen(2003)employedclusteringofverbSCF(probabil-ity)distributionstoinduceverbsemanticclasses.TheirstudiesarebasedontheassumptionthatverbSCFdistri-butionsarecloselyrelatedtoverbsemanticclasses.Con-versely,ifwecouldinducewordclasseswhoseelementwordshavethesamesetofSCFs,wecaneliminateSCFsacquiredinerrorfromthecorporaandpredictplausibleSCFsunseeninthecorpora.ThiskindofgeneralizationwouldbeusefultoimprovethequalityoftheacquiredSCFs.

Inthispaper,wepresentamethodofgeneralizingSCFsacquiredfromcorporainordertoaugmentalex-iconofalexicalizedgrammar.Forwordsintheac-quiredSCFlexiconandthelexiconofthetargetlexical-izedgrammar,we rstestimateacon dencevaluethatawordcanhaveeachSCF.WenextperformclusteringofSCFcon dence-valuevectorsinordertomakeuseofco-occurrencetendencyamongSCFsforwordsinthelex-iconofthetargetlexicalizedgrammar.Sinceeachcen-troidvalueoftheobtainedclustersindicatewhetherthewordsinthatclasshaveeachSCF,weeliminateimplausi-bleSCFsandaddunobservedbutpossibleSCFsaccord-ingtothatvalue.Inotherwords,wecangeneralizetheacquiredSCFsbythereliablelexiconofthetargetlexi-calizedgrammar.

WeappliedourmethodtoSCFsacquiredfrommo-bilephonenewsgroupscorpusbyamethoddescribedin(CarrollandFang,2004),inordertogeneralizetheacquiredSCFsbyusingatrainingportionoftheSCFlexiconoftheXTAGEnglishgrammar(XTAGResearchGroup,2001),alarge-scaleLexicalizedTreeAdjoiningGrammar(LTAG)(Schabesetal.,1988).WeevaluatedtheresultingSCFlexiconbycomparingittotherestof

1Introduction

Recently,avarietyofmethodshavebeenproposedforautomaticacquisitionofsubcategorizationframes(SCFs)fromcorpora(Brent,1993;Manning,1993;BriscoeandCarroll,1997;SarkarandZeman,2000;Korhonen,2002).Althoughtheseresearcheffortsaimedatenhanc-inglexiconresources,therehasbeenlittleworkonevalu-atingtheimpactofacquiredSCFsongrammarcoverageusinglarge-scalelexicalizedgrammarswiththeexcep-tionof(CarrollandFang,2004).

TheproblemwhenwecombineacquiredSCFswithexistinglexicalizedgrammarsislowerqualityoftheac-quiredSCFs,sincetheyareacquiredinanunsupervisedmanner,ratherthanbeingmanuallycoded.Ifweattempttocompensateforthelackofrecallbybeinglessstrictin lteringoutlesslikelySCFs,thenwewillendupwithalargernumberoflexicalentries.Thisisfatalforparsing

TAG+7: Seventh International Workshop on Tree Adjoining Grammar and Related Formalisms.