The role of polysemy in masked semantic and translation prim(8)

8M.Finkbeineretal./JournalofMemoryandLanguage51(2004)1–22

categoryactsasakindofcontextwhichemphasizesthecategory-relevantpropertiesoftheprimeandtarget.Suchane ectwouldbeexpectedonlyforexemplarsofthecategory.Inwhatfollows,weprovideapossibleexplanationforthispatternof ndings.

TheSenseModel

Webeginwithanissuethathasnotreceivedade-quateattentionbeforeinthecross-languageprimingliterature:mostwordsarepolysemousandtherangeofsensesthatawordcanhavetendstobelanguagespeci c.3Forexample,theEnglishword‘‘black’’andtheJapaneseword‘‘kuroi’’aretranslationequivalents,despitehavinglittlebesidestheonesenseCOLORincommon.AccordingtoWordNet1.6(Fellbaum,1998),theword‘‘black’’has21di erentsenses.Forexample,‘‘black’’isusedtorefertoatypeofhumor,aswellasacalamitousdayonWallStreet(blackMonday).InJapanese,‘‘kuroi’’canbeusedtorefertothosewhoareevil-minded(blackbelly),aswellasthosewhoarewelltannedorguiltyofacrime.Hence,thelanguage-speci cshadesofmeaning,orsenses,forthesetwowordsextendwellbeyondthesinglesensedeterminingtranslationequivalency.Therelevanceofthisfortranslationprimingisthattheamountofprimingmaydependnotonlyontheoverlapinthesemanticsensesactivatedbytheprimeandthetarget,but,crucially,ontheratioofprimedtounprimedsensesassociatedwiththetarget.

Inthinkingabouthowtoformulateanaccountofthetranslationprimingasymmetrythatmakescentralthefactofpolysemy,wehavebegunwithasetofassump-tionsthat,thoughfundamentallydi erentfromthoseonwhichtheRevisedHierarchicalModelandtheSeparateMemorySystemsModelarebased,resonatefairlyclo-selywiththoseoftheDistributedConceptualFeatureModel(deGroot,1992).IndeGrootÕsdistributedcon-ceptualfeaturemodel(DCFM),lexicalnodesareasso-ciatedwithadistributedsetofconceptualfeatures,asopposedtolocalistconceptualrepresentations.Ac-cordingtotheDCFM,therearevaryingdegreesofoverlapofconceptualfeaturesbetweenL1andL2meanings,dependingonwhattypeofwordisrepre-sented.Forexample,deGroot(1992,1993)hasarguedthatthereismorefeaturaloverlapbetweentranslationequivalentsforconcretewordsthanabstractwordsbe-causeconcretereferentstendtohavethesameshape,size,andfunctioncross-linguistically.TheDCFMis

3Here,wedistinguishpolysemyfromambiguity.Apolyse-mousword(e.g.,‘‘call’’)hasseveralshadesofmeaningorusages(i.e.,senses)thatareallconsideredtobepartofthesameword(e.g.,‘‘telephonecall’’and‘‘tocallout’’).Anambiguouswordlike‘‘bark,’’though,canbeconsideredtobetwodistinctwords(‘‘tobark’’and‘‘treebark’’)thatjusthappentosharethesameorthographicandphonologicalform.

intuitivelyappealingonseveralfronts,notleastofwhichisitsstraightforwardaccountofhowtranslationequivalentscanhavelanguagespeci cmeanings.Infact,deGroot(1993)hasreferredtotheDCFMasa‘‘mixed’’modelbecauseitispossibleforthemeaningrepresen-tationstorangefromhavingcompletefeaturaloverlap(essentiallythearchitectureofacompoundbilingual)tohavingnofeaturaloverlap(i.e.,thearchitectureofacoordinatebilingual)dependinguponthetypeofwordinquestion.

AcriticalproblemwiththeDCFM,ascurrentlyspeci ed,though,isitsdi cultyinaccountingforasymmetriesintranslationperformance.Thatis,be-causethismodelattributestranslationperformancetotheconceptualfeaturescommontobothL1andL2,andbecausetheamountoffeaturaloverlapbetweentwotranslationequivalentsisconstantregardlessoftrans-lationdirection,anye ect(suchaspriming)foundinonedirectionshouldoccurequallystronglyintheotherdirection.Assuch,theDCFMhasaparticulardi cultyinaccountingforthemaskedtranslationprimingasymmetry.Intheirrevisionoftheconceptualfeaturemodel,KrollanddeGroot(1997)haveaddressedthisshortcomingbyproposingweakconnectionsbetweenL2lexicalnodesandconceptualfeatures.Asthe ndingsofExperiments1–3indicate,though,theconnection-strengthapproachmaynotbethecorrectone.Instead,wehavechosentopursueadi erentmodi cationtotheDCFM,oneinwhichlexicalsemanticrepresentationsareassumedtobeboundedandcomprisedofdistinct‘‘bundles’’offeaturescorrespondingtodistinctusages,whichwerefertoassemanticsenses.Thesemodi ca-tionstotheDCFMpresentuswiththehypothesisthatwhatiscriticalinobservingtranslationprimingisthedegreetowhichthecompletelexicalsemanticrepresen-tation(asopposedtojustthefeaturesincommonbe-tweentranslationequivalents)hasbeenactivatedbytheprime.

Thenumberofbundledfeatures,orsenses,presentinalexicalsemanticrepresentationcanrangefromonetoseveraldozen.Theexactnumberdependsonhowmanyusagesthatparticularwordhasand,importanttothediscussionofthebilinguallexicon,howknowledgeabletheindividualisofthoseusages.Becausetheshadesofmeaningandrangeofusagesthataparticularwordmayhavevaryfromlanguagetolanguage,itisassumedthatseveralofawordÕssenses(i.e.,thewayinwhichcon-ceptualfeaturesareboundtogether)willbelanguagespeci c,withanobviousexceptionforthesemanticsense(s)determiningtranslationequivalency.These-manticsensedeterminingtranslationequivalencyissharedbetweenL1andL2lexicalentries,andisthusconsideredtobelargely,ifnotcompletely,identicalbetweenL1andL2(seeFig.2).

Byassumingthateachsenseofawordconstitutesadistinctmentalrepresentationwithinalexicalsemantic