The role of polysemy in masked semantic and translation prim(16)

16M.Finkbeineretal./JournalofMemoryandLanguage51(2004)1–22

Table2

Comparisonofpriminge ectsanderrorrates(inparenthesis)forall-Englishprime–targetpairsbydirection(many-to-onevs.one-to-many)andbytask(lexicaldecisionvs.semanticcategorization)inExperiments4–6

PrimeconditionPrime

LexicaldecisiontaskSemanticcategorizationtaskSemanticcategorizationtaskwithexemplarsascontrolprimes

Many-to-one(Experiment4a)One-to-many(Experiment4b)One-to-many(Experiment5a)Many-to-one(Experiment5b)One-to-many(Experiment6)

550538473532463

(5.70)(5.1)(3.3)(6.0)(3.3)

Control573528486552475

(5.75)(6.2)(7.5)(10.9)(5.5)

23)10132012Priminge ect

meaningsarerespondedtofasterthanunambiguouswords)isactuallyasenseadvantage.Theyfoundthatlexicaldecisiontimeswereactuallyslowerforambigu-ouswords(wordswithdistinctmeanings,like‘‘bank’’)thanforunambiguouswords(thoughnotsigni cantlyslower),whereasdecisiontimesforwordswithmanysenses(e.g.,‘‘hammer’’)weresigni cantlyfasterthanforwordswithfewsenses(e.g.,‘‘cake’’).Allwordswerematchedforfrequency,length,familiarityandcon-creteness.Similar ndingswereobtainedbyFinkbeiner(2002)usingalargerandrandomlygeneratedsetofitems.5ThisinvolvedusingWordNet(Fellbaum,1998)togeneratelistsoffew-sensewords(nomorethantwosenses)andmany-sensewords(15ormoresenses).Eightyitemsappearinginafrequencybandof20–80permillionwerethenrandomlyselectedfromeachlistofpossiblewordsforatotalof160experimentalitems(itemshadameanfrequencyof45.8andameanlengthof5.1lettersoneachlist).Usingtheseitemsinalexicaldecisiontask,areliabledi erencebetweenresponsetimesformany-sensewords(533ms)andfew-sensewords(564ms)wasfound(F1ð1;14Þ¼14:44,P¼:001;F2ð1;78Þ¼36:54,P<:001).Takentogether,these ndingssuggestthatlexicaldecisiontimesaresensitivetothenumberofsemanticsensesassociatedwithaparticulartarget.Itlogicallyfollowsfromthisthatinordertoshowpriming,alargeproportionofatargetwordÕssemanticsensesmustbepreactivatedbytheprime.

Insummary,themaskedtranslationprimingasym-metryisarguedtobethestraightforwardconsequenceofarepresentationalasymmetrybetweenL1andL2lexico-semanticrepresentations.L1wordsarearguedtohaverelativelymoresensesassociatedwiththemthanL2words,andtranslation-equivalentwordspresumablyonlysharearestrictednumberofsenses.Followingfromthis,itisarguedthattheproportionofL1sensesprimedbyanL2equivalentisgoingtobelowerthanthepro-

portionofL2sensesprimedbyanL1equivalent,whichleadstotheoftenobservedtranslationprimingasym-metryinlexicaldecision.

Taskdi erencesinmaskedtranslationpriming

Wereturnnowtothequestionofhowsemanticcategorizationmighteliminatetheprimingasymmetry.IfitiscorrecttoassumethatL2–L1primingdoesnotoccur(orisveryweak)becauseaninsu cientnumberofL1sensesarepreactivatedbytheL2prime,thentheremustbesomethingaboutthesemanticcategorizationtaskthatservestorestrictthenumberofsensesrecruitedwhenadecisionisgenerated.Essentially,wearguethatthesemanticcategorizationtaskturnsmany-sensetar-getsintoone-sensetargetsby‘‘ ltering’’outcategory-irrelevantsensesfromthedecisionmakingprocess.Thatis,insemanticcategorization,thesensethattriggersadecisionwillalwaysbethesensethatisrelevanttothecategory,andthisallowstheL2primetobemoreef-fective.Assumingthatthematerialshavebeendesignedappropriately,thecategory-relevantsensewillbethesensethatboththeL2primeandtheL1targetshare,meaningthatthesensethattriggersadecisionwillal-readyhavebeenactivatedbytheL2prime.

Thisaccountraisesseveralinterestingquestions.First,howisthecategoryabletoselectouttherelevantsense?Itisinterestingtonotethatasimilarphenomenonapparentlyoccursinthecontextofneighborhoode ects(Forster&Hector,2002).Theseinvestigatorsfoundthatinasemanticcategorizationtask(e.g.,‘‘Isitanani-mal?’’),non-wordsthatwereone-letterdi erentfrommanywords(e.g.,walley)wereclassi edjustasrapidlyasnon-wordswithfewneighbors(e.g.,braln),unlessoneofthoseneighborshappenedtobeananimalname(e.g.,turple),inwhichcaseasubstantialinhibitorye ectwasobserved.Evidently,aspellingcheckwastriggeredforturplebecauseofitssimilaritytoaword(whichhap-penedtobeanexemplar),butnotforanon-wordsuchascishop,despiteitssimilaritytoaword.Althoughthisishardlyasurprisingresult,itdoesraisethequestionofhowthedecisionsystemknowswhethertoignorethepresenceofawordneighbor.Indeed,non-wordsthathad

Forster(2000)hasarguedthatinthiskindofresearch,itisimperativethatitemsareselectedrandomlytoavoidinadver-tentexperimenterbias.

5