MichaelW.Harm
CarnegieMellonUniversity
MarkS.Seidenberg
UniversityofWisconsin-Madison
Arewordsreadvisually(bymeansofadirectmappingfromorthographytosemantics)orphonolog-ically(bymappingfromorthographytophonologytosemantics)?Weaddressedthislongstandingdebatebyexamininghowalarge-scalecomputationalmodelbasedonconnectionistprincipleswouldsolvetheproblemandcomparingthemodel’sperformancetopeoples’.Incontrasttopreviousmod-els,thepresentmodelemploysanarchitectureinwhichmeaningsarejointlydeterminedbythetwocomponents,withthedivisionoflaborbetweenthemaffectedbythenatureofthemappingsbetweencodes.Themodelisconsistentwithavarietyofbehavioralphenomena,includingtheresultsofstudiesofhomophonesandpseudohomophonesthoughttosupportothertheories,andillustrateshowefficientprocessingcanbeachievedusingmultiplesimultaneousconstraints.
1.INTRODUCTION
Althoughhumanshavebeenreadingforseveralthou-sandyearsandstudyingreadingformorethanacentury,themechanismsgoverningtheacquisition,useandbreakdownofthisskillcontinuetobethesubjectofconsiderableinter-estandcontroversy(seeAdams,1990;Rayner,Foorman,Perfetti,Pesetsky,&Seidenberg,2001;NationalReadingPanel,2000,forreviews).Thepresentarticlefocusesonacentralaspectofreading,theprocessesinvolvedindeter-miningthemeaningsofwordsfromprint.
Inprinciple,askilledreadercoulddeterminethemean-ing(ormeanings)ofaworddirectlyfromknowledgeofitsspelling.However,alphabeticorthographies,inwhich
2
HARM,SEIDENBERG
trate.Wedevelopedamodelofthecomputationofwordmeaningfromprintbasedongeneralcomputationalprinci-plesthathavebeenexploredinpreviousresearchonread-ing(Plaut,McClelland,Seidenberg,&Patterson,1996;Sei-denberg&McClelland,1989;Harm&Seidenberg,1999)andotherphenomena.Whereasourearlierreadingmodelsfocusedonthetranslationfromprinttosound,thepresentmodeladdressesreadingformeaning.Conversely,(Hinton&Shallice,1991;Plaut&Shallice,1993)addressedcom-plementaryissuesconcerningthecomputationfromorthog-raphytomeaningintheirworkonacquireddeepdyslexia,anunusualreadingimpairmentobservedfollowingsometypesofbraininjury.Ourmodelbuildsonthisworkbutdiffersfromitinsofarasitisthefirstmodeltoaddresshowmeaningiscomputedinasysteminwhichbothvisual(orth-sem)andphonologically-mediated(orth-phon-sem)path-waysareavailable.Theimplementedmodelwasthenas-sessedagainstabodyofcriticalfindingsfrombehavioralstudies.
Asitturnsout,theproposedmodelisconsistentwithmanyaspectsofearlieraccounts,butdiffersfromtheminimportantrespectsbecauseofspecificpropertiesofthecomputationalmechanismsthatareemployed.Withinthisframework,themeaningofawordisapatternofac-tivationoverasetofsemanticunitsthatdevelopsovertimebasedoncontinuousinputfrombothorthsemandorthphonsemcomponentsofthetriangle(seeFig-ure1).Themaintheoreticalissueconcernsthecomputa-tionalconsiderationsthatdeterminehowthemodel(andbyhypothesisthereader)arrivesatanefficientdivisionofla-borbetweenthesesourcesofinput.Thustheconceptofindependentvisualandphonologicalrecognitionroutines,oneofwhich(e.g.,thefastest-finishing)providesaccesstomeaning(e.g.,McCuskeretal.,1981;Carr&Pollatsek,1985;Caplan,1992;Frost,1998)isreplacedbyacoop-erativecomputationinwhichsemanticpatternsreflectthejointeffectsofinputfromdifferentsources.Themannerinwhichthedivisionoflaboremergesinthemodelre-lateswelltofindingsconcerningtheprimacyofphonolog-icalcodesinreadingacquisition.Themodelisalsocon-sistentwithandprovidesinsightaboutanumberofimpor-tantempiricalfindingsconcerningtheprocessingofhomo-phones(e.g.,BARE-BEAR)andpseudohomophones(e.g.,BAIR)thathavefiguredprominentlyinpreviousaccounts.Thestructureofthearticleisasfollows.Wefirstreviewthepretheoreticalargumentsandcriticalempiricaldatathatledtodifferingconclusionsabouttheimportanceofdi-rectvs.phonologically-mediatedaccess.Therearegoodargumentsonbothsidesofthedebate,astheinconclu-sivestateofcurrenttheorizingwouldpredict.Wethende-scribeanapproachtothisissuebasedongeneralcomputa-tionalprinciplesconcerningknowledgerepresentation,ac-quisitionandprocessingderivedfromtheconnectionistorparalleldistributedprocessingapproach(Rumelhart,Mc-Clelland,&PDPResearchGroup,1986).Acomputational
ContextSemanticsOrthographyPhonologyMAKE/mAk/Figure1.The“triangle”modelofSeidenbergandMcClelland(1989).Theimplementedmodelexaminedhowphonologicalcodesarecomputedfromorthography.Thepresentresearchex-aminedprocessesinvolvedincomputingsemanticcodesfromor-thography,giventheavailabilityofbothdirect(orthsem)andphonologically-mediated(orthphonsem)pathways.
modelembodyingtheseprinciplesandotherassumptionsaboutcriticalcharacteristicsofreadingandtheconditionsunderwhichchildrenlearntoreadwillbeintroducedandanalyzed,anditsbehaviorlinkedtoempiricalfindings.Inthegeneraldiscussionwesummarizetheimportantproper-tiesofthemodelandconsidersomelimitationsofthework,unresolvedissuesanddirectionsforfutureresearch.
IntuitionsandEvidence
Thissectionprovidesanoverviewofpreviousresearchonvisualandphonologicalprocessesinreading.Beforeproceeding,aterminologicalissueneedstobeaddressed.Basicprocessesinreadingareoftendiscussedintermsof“models”thatillustratetheoreticalclaims(e.g.,LaBerge&Samuels,1974;Marshall&Newcombe,1973;Mor-ton,1969;Coltheart,Curtis,Atkins,&Haller,1993;Sei-denberg&McClelland,1989).Modelsthatincorporatebothdirect-visualandphonologically-mediatedcomputa-tionsfromprinttomeaningareoftentermed“dual-routemodels”(seeFrost,1998,forarecentexampleanddiscus-sionofthisuseoftheterm).However,thisusageispo-tentiallyconfusing,becausethetermhasalsobeenexten-sivelyusedinthereadingliteratureinreferencetoadiffer-entissue,themechanism(s)involvedingeneratingpronun-ciationsfromprint(e.g.,Coltheartetal.,1993).
Thattherearebothdirect-visualandphonologically-mediatedmappingsfromprinttomeaningisnotatheo-reticalclaimspecifictoanyparticularmodelofreading.Rather,thebasicdesignfeatureofalphabeticwritingsys-temsisthatalthoughstringsofletterscanbedirectlyasso-
MULTICOMPONENTMODELOFREADING
3
ciatedwithmeanings(ascanothervisualstimulisuchas+),thelettersalsorepresentthesoundsofwords,whichisinturnassociatedwithoneormoremeanings.Wherethe-oriesdifferiswithrespecttohowtheselexicalcodesandrelationsbetweenthemarestructured,howthisknowledgeisacquiredandrepresentedinmemory,andwhatrolesthesetypesofinformationplayinreading.1
Whereastheabovesenseof“dual-routemodel”referstomechanismsfortranslatingfromprinttomeaning,thetermalsoreferstoaspecifictheoreticalproposal,stud-iedformanyyearsbyColtheartandothers(e.g.,McCann&Besner,1987),concerningmechanismsfortranslatingfromprinttosound.Accordingtothistheory,pronouncingletterstringsinEnglish(wordsandpseudowordssuchasNUST)requirestwomechanisms,oneinvolvingknowledgeofwholewords,andoneinvolvingrulesgoverningthecor-respondencesbetweengraphemesandphonemes.Alltheo-riesofreadingarenot“dual-route”inthissense;inpartic-ular,connectionistmodelsdatingfromSeidenbergandMc-Clelland(1989)havesuggestedthatthefunctionsthattheColtheartmodelattributestotwoseparatemechanismsarisefromasingleconnectionistmechanism(seealsoGlushko,1979).Thesealternativetheoriesarethesubjectofcontinu-ingresearchanddebate,butarenotthefocusofthepresentarticle.2
termsthathadbeenused,suchasthe“dual-encodinghypothesis”(Meyeretal.,1974)or“parallelcodingsystemsmodels”(Carr&Pollatsek,1985).Asrecentlyas2000Colthearthasusedthisterminreferencetoboththecomputationofmeaning(directvs.phonologically-mediated)andthecomputationofphonology(lex-icalvs.sublexicalprocedures;Coltheart,2000).Wethinkthisus-ageisconfusing,however,forthefollowingreason:evidencethatthereare“dual”visualandphonologically-mediatedmappingstomeaning,whichistrueofallalphabets,oftenregistersasevidenceforthe“dualroutemodel”anditsmorespecificclaimthattherearetwomechanismsforpronouncingletterstrings.Becauseofthisambiguity,because“dualroutemodel”isusedindifferentwaysindifferentcontexts,andbecauseourmodeldiffersfromtheColtheartpronunciationmodelwithwhichthetermisstronglyassociated,weavoiditintheremainderofthisarticle.
3The
followingnotationalconventionsareusedinthisarticle.
Thewrittenformofawordisshowninsmallcaps,thephonolog-icalformiscodedinInternationalPhoneticAlphabetnotationbe-tweenslashes,thesemanticconceptfortheitemisshowninbracesandthesemanticfeaturescomprisingthatconceptaredenotedinbrackets.HencethevisualformCATcorrespondstothephonolog-icalrepresentation/
/andthesemanticconcept{cat},whichconsistsofsemanticfeaturessuchas[feline],[has-fur],[living-thing],etc.
4
HARM,SEIDENBERG
icalrecodingplaysacausalroleintheaccessofmean-ingwasequivocal(Perfetti&McCutchen,1983;McCuskeretal.,1981).Itwasverydifficulttocreateconditionsthatshowednotmerelythatreadersactivatedphonologicalin-formationbutthattheyusedthisinformationinaccessingmeaning.Severalmodelsthatemphasizedvisually-basedrecognitionprocedureswereproposed(e.g.Baron,1973;Paap,Newsome,McDonald,&Schvanevelt,1982;Mc-Clelland&Rumelhart,1981).Coltheart(1978)alsoarguedstronglyfordirectvisualaccess.
Evidenceforphonologicalmediation.Overthepast20yearsthedirectvisualaccessviewhasbeenstronglycalledintoquestion.Thedirectaccessviewhasanairofparadoxaboutit:thedevelopmentofwritingsystemssinceabout2500B.C.hasbeentowardsymbolsthatrepresentsoundsratherthanmeanings(Hung&Tzeng,1981).Whyaretherealphabeticwritingsystemsifphonologicalinforma-tionplaysnousefulroleinreading?ThereisnowstrongevidencefortheextensiveuseofphonologyinreadingformeaninginEnglishandotherlanguages(e.g.,Perfetti,Bell,&Delaney,1988;VanOrden,Johnston,&Hale,1988),de-rivedfrombehavioralstudiesofchildrenandadults,andfromobservationsaboutdifferencesbetweenthemappingsbetweenspellingandsoundvs.spellingandmeaningthataffectlearning.Wewillsummarizethisevidenceandre-latedargumentsbriefly(seeRayner&Pollatsek,1989;VanOrden,Pennington,&Stone,1990;Frost,1998,forfullerdiscussion).
Childrenhavelargespoken-wordvocabulariesbythetimereadinginstructionbegins.Reading,onthisview,in-volveslearninghowwrittensymbolsrelatetoknownspo-kenwordforms.InalphabeticorthographiessuchastheoneforEnglish,writtensymbolsrepresentsounds,specif-icallyphonemicsegments.Thus,successfulreadingac-quisitionrequiresdevelopingsegmentalrepresentationsofspeechandgraspingthe“alphabeticprinciple”concerningthemappingbetweenletters(orcombinationsofletters)andphonemes(Liberman,Shankweiler,&Liberman,1989;Gathercole&Baddeley,1993).
JormandShare(1983)furtherobservedthattheabil-itytosoundoutwords(eitherovertlyorcovertly)givesthechildaself-teachingmechanismthatfacilitateslearn-ingtoread:Thechildcansoundoutawordanddeterminewhetheritmatchesaknownspokenword.Connectionistmodelsprovideamechanisticinterpretationofthistypeoflearning.Thecomparisonbetweentheself-generatedpro-nunciationandinformationaboutaword’ssoundcanbeseenasthebasisforcomputinganerrorsignalthatallowsadjustmentoftheweightsonconnectionsmediatingtheorthphonmapping.
VanOrdenandcolleagues(VanOrden,1987;VanOr-denetal.,1988,1990)presentedasomewhatdifferentargu-ment.TheyobservedthatinEnglish,orthographyandse-manticsarelargelyuncorrelated,whereasorthographyand
phonologyarehighlycorrelated;thus,theformershouldbehardertolearnthanthelatter.AsVanOrdenetal.(1990)stated,“Weproposethattherelativelyinvariantcorrespondencebetweenorthographicrepresentationsandphonologicrepresentationsexplainswhywordidentifica-tionappearstobemediatedbyphonology.”Flesch(1955)alsomadethisargumentwithgreaterpolemicalfervor,as-sertingthatteachingchildrentoreadbyrotememorizationoftheassociatedmeaningsofwordformsratherthanlog-icaldeductionofthesoundsofwords“...consistsessen-tiallyoftreatingchildrenasiftheyweredogs...It’sthemostinhuman,mean,stupidwayoffoistingsomethingonachild’smind.”
Mergingtheseargumentsyieldsatheoreticalstanceinwhichorthphoniseasiertolearnthanorthsem,andphonsemisalreadyknownformanywords.Hence,earlyreadingreliesonorthphonsemmuchmorethanorthsem.
Thereisconsiderableevidencethatchildrenusephono-logicalinformationinreading(Liberman&Shankweiler,1985)andthatthequalityofphonologicalrepresentationsisstronglyrelatedtoreadingachievement(Snowling,1991).Themostcompellingevidencederivesfromstudiesshow-ingthatpre-readers’knowledgeofphonologicalstructureispredictiveofreadingachievementseveralyearslater(Bradley&Bryant,1983;Lundberg,Olofsson,&Wall,1980).Thereisalsoevidencethatimpairmentsintherepre-sentationofphonologyareoftenobservedindyslexics(seeHarm&Seidenberg,1999,forasummaryandacomputa-tionalmodeloftheseeffects).Thesedevelopmentalresultsfindanaturalinterpretationwithinatheorywhichstatesthatorthographicpatternsactivatephonologicalrepresentationsearlyintheprocessofreadingwordsformeaning.
Giventheextensiveevidencefortheuseofphonolog-icalinformationinbeginningreading,itisbeenoftenas-sumedthatthisstrategycarriesovertoskilledadultread-ing.Frost(1998)termsthisthe“strongphonology”theory.Manystudiesofadultreaderssupportthisview;herewereviewsomecriticalfindingsthatarerelevanttothesimu-lationsreportedbelow.
AclassicstudybyVanOrden(1987)yieldeddirectev-idencethatphonologicalinformationhasacausalroleintheaccessofmeaning.Subjectsperformedasemanticde-cisiontaskinwhichtheyhadtodecideifatargetwordwasanexemplarofaspecifiedcategory.Forexample,forthe“food”category,thetargetswereeithertrueex-emplars(e.g.,MEAT),homophonousfoils(e.g.,MEET),ornon-homophonousspellingcontrols(e.g.,MOOT).VanOr-denfoundthatsubjectsmadeahighnumberoffalseposi-tivesonphonologicalfoilsrelativetoorthographiccontrols.Subjectswouldnotmakefalsepositivesunlesstheywereactivatingphonologicalinformationandusingittoaccessmeaning.Laterexperiments(e.g.,VanOrdenetal.,1988)yieldedthesameeffectforpseudohomophonestimuli(e.g.,category:clothing;target:SUTE).Theseresultsweretaken
MULTICOMPONENTMODELOFREADING
5
toindicatethatwordrecognitionprogressesfromspellingtosoundtomeaning,withhomophonessuchasBEARorPLANEdisambiguatedbyalate“spellingcheck”procedureaftermeaningshavebeenaccessed.
Perfettietal.(1988)demonstratedeffectsofthephono-logicalformofwordsataveryearlystageofprocessing.Theyfoundthatwhenawordispresentedverybrieflyandthenmaskedbyahomophonouswordmask,identificationofthetargetwordisfacilitatedrelativetoaneutralmask.Theseresultsalsosuggestthatthephonologicalformofawordisactivatedautomaticallyataveryearlystageinpro-cessing.
LeschandPollatsek(1993)andLukatelaandTurvey(1994b)extendedthesefindingsusinghomophonesinase-manticprimingparadigm.Iftheaccessofmeaningisini-tiallyphonological,andhomophonesaredisambiguatedbyasubsequentspellingcheck,thereshouldbeapointearlyinprocessingatwhichhomophonesactivatemultiplemean-ings;later,afterthespellingcheckhasoccurred,onlytheappropriatemeaningshouldbeactive.LeschandPollat-sek(1993)andLukatelaandTurvey(1994b)usedamaskedprimingparadigmtoexplorethishypothesis.Inthecriticalconditions,atargetsuchasFROGwasprecededbyarelatedprimesuchasTOADoritssemantically-unrelatedhomo-phoneTOWED.Theprimewordwaspresentedforeitherashort(50ms)orlong(250ms)durationandthenmaskedbythetarget,whichwastobenamed.Semanticallyrelatedprime-targetpairs(e.g.,TOAD-FROG)producedfacilitationcomparedtoanunrelatedcontrolconditionatbothprimedurations.Inappropriateprimes(e.g.,TOWED-FROG)pro-ducedfacilitationonlyintheshortcondition.Thustheef-fectswereconsistentwithVanOrden’saccountinwhichmeaningisinitiallyactivatedviaphonologywithhomo-phonessubsequentlydisambiguatedbyaspellingcheck.Maskingthestimuliatanearlystageinprocessing(50ms)removestheorthographicinformationthatnormallysup-portsthespellingcheck.
ReconcilistTheories
Althoughconsiderableattentionhasfocusedondirectvisualaccessandphonologically-mediatedaccessascom-petingalternatives,othertheorieshaveassumedthatreadersmakeuseofboth,withseveralfactorsdeterminingwhichpathwaywillbedominantinagivensituation(e.g.,Baron&Strawson,1976;Carr&Pollatsek,1985;seeSeiden-berg,1995,fordiscussion).Thestrongdirectvisualac-cesspositionadvocatedbySmithcannotbecorrect;therearetoomanystudiesshowingunambiguousphonologicaleffectsinreadingformeaning.Moreover,Smith’sargu-mentaboutthedifficultiesinvolvedinusingphonologicalmediationrestsontheassumptionthatspelling-soundcor-respondencesareencodedbyrules.ConnectionistmodelssuchasSeidenbergandMcClelland’s(1989)subsequentlyprovidedanalternativeinwhichthecorrespondencesare
encodedbyweightsonconnectionsbetweenunitsinvolvedintheorthography-phonologymapping.Suchsystemscanencodedifferentdegreesofconsistencyintheorthphonmappingoperatingovermanydifferentorthographicandphonologicalsubunits.Thusthemodelinstantiatedathe-oryofhowreaderscouldefficientlyactivatephonologicalcodesforallwords,includingonesthatinvolveatypicalmappings.
Thestrongversionofthephonologicalmediationthe-oryhasalsobeenquestioned,however.Everynormalin-dividualcanrecognizeandaccessconceptualinformationassociatedwithobjectswithoutanintermediatephonolog-icalrecodingstep;whywouldn’tthisbepossiblewhentheobjectsinquestionhappentobefamiliarletterstrings?Individualswhoareprofoundlydeaffrombirthandhavenotreceivedspeechtrainingcandeterminethemeaningsofprintedwords,aprocessthatcannotrelyonphonologicalinformation.Thisobservationsuggeststhatmeaningscanbecomputeddirectlyfromprint,butleavesopentheextenttowhichthisprocessisusedbyindividualswhoalsohaveaccesstophonology.
Otherquestionsariseconcerningtheprocessingofho-mophones.ThemanyhomophonesinEnglishpresentacomplicationforasysteminwhichmeaningsareexclu-sivelyactivatedthroughphonology;thesewordswillhavetobedisambiguatedeverytimetheyareread,whichwouldseemtoimposeaconsiderableburdenonthereadingsys-tem,aburdenthatwouldbeavoidedifmeaningswereac-cesseddirectlyfromprint.ThesolutionthatVanOrden(1987)proposedwasaveryrapidspellingcheckfollow-ingtheinitial,phonologically-drivenactivationofmeaning,i.e.,comparingtheactivatedmeaningsagainstthespellingofthewordtodeterminewhichiscorrect.Thespellingcheckideaseemstoentailthatthereaderbeabletocom-putethearbitraryassociationbetweenaword’smeaninganditsspelling.Ifreadersareabletocomputefrommean-ingtospelling,itisnotclearwhytheywouldnotbeabletocomputefromspellingtomeaning.Thus,arealisticimple-mentationofthespellingcheckprocedureseemstorequiremasteringthekindofarbitrarymappingthatisproscribedin“strongphonology”theories(Seidenberg,1995).
Thereisalsoanempiricalquestion:JaredandSei-denberg(1991)providedevidencethattheextenttowhichphonologyentersintotheactivationofmeaningvariesasafunctionofwordfrequency.TheyreplicatedtheVanOr-den(1987)results,butalsoexperimentallymanipulatedthefrequenciesoftheexemplars(e.g.,ROSE)andhomophonefoils(e.g.,ROWS).Intheirstudies,onlyhomophoneswithtwolowfrequencymeaningsgeneratedsignificantfalsepositives.Higherfrequencywordsdidnotyieldsignificantfalsepositives.Insofarasthepresenceoffalsepositiveef-fectshasprovidedthebasisfordiagnosingtheuseofpho-nologicalinformation,theabsenceoftheseeffectscouldbetakenasevidencethatthisinformationwasnotused.
TheJaredandSeidenbergresultshavegeneratedcontro-6
HARM,SEIDENBERG
versy,focusedonthepossibilitythatthefailuretoobservesignificantfalsepositivesinthehigherfrequencyconditionswasaTypeIIerror.LeschandPollatsek(1993)didnotexplicitlymanipulateprimefrequencyintheirstudy,buttheyreportedapost-hocanalysisthatrevealednoeffectoffrequencyonthemagnitudeofpriming.LukatelaandTurveydidmanipulateprimefrequencybutfoundthatithadnoeffectinsofarasbothhighandlowfrequencycondi-tionsyieldedevidenceforphonologically-basedactivationofmeaning.However,asdiscussedbelow,thefrequencymanipulationinthisstudywasquiteweak,andotheras-pectsofthestimuliandanalysisraisequestionsabouttheresults.
Theprocessingofhomophonesandpseudohomophonesisamajorfocusofthemodelingdescribedbelow.Tofore-shadowtheresults,themodelbehavessomewhatdifferentlythanbothVanOrdenetal.(1990)andJaredandSeidenberg(1991)proposedandprovidesareconciliationoftheirfind-ings.
Logicalandobservationalargumentsabouttherelativeeaseoflearningtheorthphonandorthsemmappingsalsoneedtobeexaminedcarefully.Therelationshipbe-tweenspellingandmeaningisoftensaidtobearbitraryandthereforedifficulttolearnbecausethereisnothingaboutthespellingofawordsuchasDOGthatdemandsthatit,ratherthansomeotherspellingpattern,beassoci-atedwiththeconcept{domesticcanine}.However,En-glishandsomeotheralphabeticwritingsystemsexhibitnonarbitraryform-meaningcorrespondences.Forexample,DOGmakessimilarsemanticcontributionstomanyrelatedwords(DOGS,DOGLEG,DOGHOUSE,etc.),word-final-Softenindicatesplurality,word-final-EDusuallyindicatespastness,andsoon.Thereareothercorrelationsbetweensound(andhencespelling)andmeaning,illustratedbywordssuchasGLITTER,GLISTEN,GLEAM,GLINT,GLAREandSLIP,SLIDE,SLITHER(seeMarchand,1969,formanyexamples).Further,asChomskyandHalle(1968)noted,Englishspellingpreservesmorphologicalinformationoverphonologicalinmanycases,suchasSIGN-SIGNATUREandBOMB-BOMBARD.ShalloworthographiessuchastheoneforSerbiansacrificethismorphologicalinformationinfa-vorofpreservingspelling-soundconsistency.SeidenbergandGonnerman(2000)discusstheroleofsuchnonarbi-traryform-meaningcorrespondencesinthedevelopmentofmorphologicalrepresentations.AlthoughthemappingfromspellingtomeaningislesssystematicthanfromspellingtosoundinEnglish,itisfarfromarbitrary(seealsoKelly,1992).
Itisalsoclearthatwithsufficienttrainingconnectionistmodelscanlearnarbitrarymappings(despiteassertionstothecontrary;Forster,1994).Moreover,itshouldbenotedthatevenwordswithhighlyunusualpronunciationsarenotwhollyarbitraryandthereforepartiallyoverlapwithotherwords.Higherfrequencywordsmaybeencounteredoftenenoughfortheorthsemmappingtobecomeestablished
relativelyquicklyregardlessofthedegreeofinconsistencyinpronunciation.Ingeneral,conjecturesabouttherelativeeaseoflearningdifferenttypesofmappingsneedstobeexaminedusingexplicitmodelsofthesecomputations.
Considerationssuchasthesesupportatheoryincorpo-ratingbothdirectvisualandphonologically-mediatedpro-cesses.Whichpathwayprovidesaccesstomeaningforagivenwordisthoughttodependonfactorssuchastherelativespeedofthetwomechanisms,wordfrequency,orthographic-phonologicalregularity,andthedepthoftheorthography(Frostetal.,1987;Henderson,1982;Seiden-berg,1995).
Summary
Theliteraturetodatehasfocusedonempiricalevidenceandtheoreticalargumentsconcerningtherelativepromi-nenceofthedirect-visualandphonologically-mediatedmechanisms.Eachalternativecontinuestohavestrongpro-ponents:Researcherswhomainlystudyissuesconcerningreadingacquisitionanddyslexiatendtoemphasizetheim-portanceofphonologicalcoding(e.g.,Wagner&Torgesen,1987),whereasmanyresearcherswhomainlystudyvisualwordrecognitioninadultshavefocusedontheroleofor-thography(e.g.,Grainger&Jacobs,1996).Themodelingworkdescribedbelowrepresentsanattempttoendthisim-passebytreatingtheissueasacomputationalone.Wedidnotbuildthemodelwithaparticularanswertothedivi-sionoflaborquestioninmind;rather,weasked,givenamodelofthecomputationofmeaningbasedontheprinci-plesexploredinourpreviouswork,howwoulditsolvetheproblem?Inparticular,whatarethecomputationalfactorsthatdeterminethedivisionoflaborgivenanarchitectureinwhichbothpathwayscanactivatesemantics?Wethenaskedwhetherthemodelwasconsistentwithfactsaboutreadingacquisitionandskilledperformanceandwhetheritprovidedfurtherinsightaboutthesephenomena.
Themodelthatwedescribehasanaffinitytotherecon-cilistmodelsinthesensethatbothvisualandphonologicalprocessescanactivatelexicalsemantics;importantly,how-ever,thesecomponentsarenotindependent.Ratherthanparallelprocessingroutesthatdevelopindependentlyandoperateinparallel,withoneortheotherprovidingaccesstomeaning,ourmodelemphasizesthedependencebetweenthetwoandthewaytheyjointlyandco-operativelyachieveanefficientsolutioninthecourseoflearningtomasterthetask.4
OurmodeliscloserinspirittoVanOrdenetal.’s(1990)discussionofalexicalsysteminwhichallpartsareoper-atingsimultaneouslyandthereforecontributingtotheacti-
MULTICOMPONENTMODELOFREADING
7
vationofmeaning.Ourworkdiffersfromtheiraccountinsomeways,however.AlthoughVanOrdenetal.discusseda“resonance”theoryinwhichallcomponentsofthelexicalsystemarecontinuouslyinteracting,theyalsoemphasizedtheprimacyoftheorthphonsemcomponentandsug-gestedthattheroleoforthsemwasminimalbecauseofthearbitrarinessofthemapping.Inimplementingacompu-tationalmodel,wefoundthatitbehavedinwaysthatsug-gestasomewhatdifferentpictureoftheroleoforthsemcomponent.Ourworkalsoplacesgreateremphasisonthemutualdependenceofthetwocomponents:whateachcom-ponentcontributestotheactivationofsemanticsdependsonwhattheothercontributes.Thisdivisionoflabordevelopsinthecourseoflearningtomasterthetaskandwedevoteconsiderableattentiontothefactorsthataffectitanditsrel-evancetoreadingbehavior.
Theremainderofthepaperisstructuredasfollows:Section2summarizestheprinciplesandassumptionsthatguidedthedevelopmentofthemodel.Wethendescribethesimulations,whichwereconductedintwophases.Phase1(Section3)involvedtrainingthephonologicalandse-manticattractorsandthemappingsbetweenthem;thiswasintendedtoapproximatethekindsoflexicalknowl-edgethatchildrenpossessinadvanceoflearningtoread.Phase2(Section4)involvedintroducingorthography.Forbothphasesweprovidedetailsaboutthemodel’sarchitec-tureandtraining,summarizeoverallperformance,andthencomparethemodel’sperformancetobehavioraldata.Sec-tions5through7describesimulationsofcentralbehavioralphenomena.Weconcludebydiscussinglimitationsofthemodelandfuturedirections.
Ourpresentationnecessarilygoesintoconsiderablede-tailconcerningthemotivationfortheapproach;thestruc-tureofthemodel,whichincorporatessometechnicalinno-vations;descriptionsandanalysesofthemodel’sbehavior;andcomparisonstobehavioraldata.Thismaterialisintheserviceofaddressingfourcentralissues:
1.Co-operativecomputationofmeaning.Oneprincipalgoalwastoexaminethefeasibilityofasysteminwhichsemanticactivityisdeterminedbycomputationsinvolvingbothorthsemandorthphonsemandtoexploretheextenttowhichsuchamodeliscompatiblewithevidenceconcerninghumanperformance.
2.Transitionfrombeginningtoskilledreading.Themajorfeatureofthistransitionisthatwhereasbegin-ningreadingreliesheavilyonphonologicalinformation,inskilledreadingtheroleofthevisualprocessincreasesgreatly.Themodeladdresseswhythisdevelopmentalse-quenceoccurs.
3.Processingofhomophonesandpseudohomophones.Studiesofthesestimulihaveprovidedcriticalevidenceconcerningtheroleofphonologicalinformationinwordreading.DecidingthatBEARmeans{ursineanimal}not{naked}requiresusinginformationabouttherelationshipbetweenspellingandmeaning.Pseudohomophonessuch
asBAIRprovideawaytodiagnoseifphonologicalinfor-mationhasbeenactivatedandraisequestionsabouttheroleoforthographyindeterminingthattheyarenotactualwords.Themodeladdressesthenatureofthecomputationsinvolvedinprocessingsuchstimuli.
4.Differentialeffectsofmasking.Weusedthemodeltostudyeffectsofthemaskingprocedureusedinmanystudiesinthisarea.Themodelsuggeststhatmaskinghassomewhatdifferenteffectsthanstandardlyassumedininterpretingtheresultsofsuchstudiesandinvalidatessomeoftheconclu-sionsstandardlydrawnfromsuchdata.
2.DESIGNCONSTRAINTS
Thepresentresearchispartoftheon-goingdevelop-mentofatheoryofwordreading.Instudyingthecompu-tationofmeaningsfromprint,weusedthesameresearchstrategyasinourpreviousworkonthecomputationofphonology(Seidenberg&McClelland,1989;Plautetal.,1996;Harm&Seidenberg,1999).Weworkbackandforthbetweenahigh-leveltheoryofhowpeoplereadandcompu-tationalmodelsthatinstantiatepartsofthesystem.Thethe-oryisbasedonprinciplesconcerningknowledgerepresen-tation,learningandprocessingthatarecomponentsoftheparalleldistributedprocessingapproach(Rumelhartetal.,1986).Theseprinciplesaregeneral–thoughttounderliemanyaspectsofperceptionandcognition–ratherthanspe-cifictoreading.Thisisconsistentwiththeobservationthatreading,atechnologyinventedrelativelyrecentlyinhumanhistory,makesuseofcapacitiesthatdidnotevolvespecifi-callyforthispurpose.Thetheoryalsoincorporatesconsid-erationsthatarereading-specific(e.g.,concerningthecon-ditionsunderwhichchildrenlearntoread).Thecomputa-tionalmodelisanimplementationofimportantaspectsofthetheory;itactsbothasatestoftheadequacyofproposedmechanismsandasadiscoveryprocedure,thatis,asourceofadditionalinsightaboutthebehaviorinquestion.Theresultsofthemodelingcanleadtomodificationsorexten-sionsofboththereadingtheoryandthegeneralcomputa-tionalapproach.
Inthissectionwediscussthefactorsthatdeterminedtheformoftheimplementedmodel.Thesedesignconstraintsinvolvedthreekindsofconsiderations:
(a)computationalconsiderations.Theprinciplesunder-lyingparalleldistributedprocessingmodelsandtheirratio-nalehavebeendiscussedelsewhere(e.g.,Rumelhartetal.,1986;O’Reilly&Munakata,2000;McLeod,Plunkett,&Rolls,1998);belowwefocusonpropertiesthatplayedthemostimportantrolesindeterminingourmodel’sbehavior.
(b)factsaboutreadingacquisition.Thewaythemodelwasstructuredandtrainedreflectedobservationsaboutthecapacitiesthatchildrenbringtobearonlearningtoreadandcriticalaspectsoftheirearlyreadingexperience.
(c)practicalandtheoreticalconsiderationsthatledustofocusonspecificaspectsofthetaskandmakesimplifying
8
HARM,SEIDENBERG
assumptionsaboutothers.
1.ArchitecturalHomogeneity
Standarddual-mechanismapproaches(e.g.,Coltheart,Rastle,Perry,Langdon,&Ziegler,2001)assumethatthereareseparatemechanismsinvolvingdifferenttypesofknowledgeandprocesses.Thephonologicalmechanismisusuallyassumedtoinvolverulesgoverningspelling-soundcorrespondences,whereasthedirectvisualroutein-volveslexicallookuporaninteractive-activationprocedure.Themechanismsbehavedifferentlybecausetheyarecon-structedoutofdifferentelementsandgovernedbydiffer-entprinciples.Thesystemthatweimplemented(likeotherPDPmodels)ishomogeneousinthesensethatallcomputa-tionsinvolvethesamekindsofstructures(distributedrep-resentationsoforthographic,phonological,andsemanticcodes),andcomputations(equationsgoverningthespreadofactivationalongweightedconnectionsbetweenunits).Thisisacentraltenetofthereadingtheory,onethatdis-tinguishesitfromotherapproaches.Thehomogeneityas-sumptionismotivatedbytwomainconsiderations.First,wewantedthemodel’sdivisionoflabortoemergeinthecourseoflearningtoperformthetask,notasaconsequenceofbuilt-indifferencesbetweenthetwomechanisms,be-causewethinkthisishowchildrensolvetheproblem.Sec-ond,weassumethatthebrainusesthesamebasicmech-anismstoencodedifferentlexicalcodesandthemappingsbetweenthem.Thereisnoindependentevidence,forexam-ple,thatthedifferentbrainstructuresthatsupportorthogra-phytophonologyconversionandphonologytosemanticsconversion,respectively,haveintrinsicallydifferentcompu-tationalproperties(e.g.,temporaldynamics).Thesecom-putationsenduphavingdifferentcharacteristicsbecausetheyinvolvedifferenttypesofinformationandbecausethecodesrelatetoeachotherindifferentways,butnotbe-causetheyinvolvedifferenttypesofcomputationalorneu-ralmechanisms.
2.DistributedRepresentations
Themodelutilizesdistributedrepresentations,meaningthateachcode(orthography,phonology,semantics)isrep-resentedbyasetofunitsandeachunitparticipatesintherepresentationofmanywords.Thiscontrastswith“local-ist”systemsinwhichindividualunitsareusedtorepresentthespelling,soundandmeaningofawordortheword’s“lexicalentry.”Importantadvanceshavebeenmadeus-ingbothtypesofrepresentation(e.g.,localist:McClelland&Rumelhart,1981;Dell,1986;Joanisse&Seidenberg,1999;distributed:Gaskell&Marslen-Wilson,1997;Plaut&Booth,2000).Ouruseofdistributedrepresentationswasmotivatedbyseveralconsiderations:
a.thistypeofrepresentationistiedtootheraspectsofthecomputationalframeworkweemployed,includingtheuseofmultilayernetworksthatincorporateunderlying,
“hidden”units,andtheuseofaweight-adjustinglearningalgorithm;
b.thedesiretomaintaincontinuitywithourpreviouswork,inwhichmodelsthatusedsuchrepresentationspro-videdinsightaboutotheraspectsofwordreading;
c.inrecognitionofevidencethatthebrainwidelyem-ploysdistributedrepresentations(see,e.g.,Rolls,Critch-ley,&Treves,1996;Ishai,Ungerleider,Martin,&Haxby,2000;Andersen,1999).Althoughmuch-simplifiedwithrespecttotheunderlyingneuralmechanisms,theuseoftheserepresentationsrepresentsasteptowardincorporatingbiologically-motivatedconstraintsoncognitivemodels;
d.theuseoftheserepresentationsfiguresinseveralofthereadingphenomenathatarethefocusofthework(e.g.,theeffectsofmaskingdiscussedinSection7).
Thus,theuseofdistributedrepresentationsispartofthetheoryofwordreadingthatisproposedhere.Therearenoimplementedlocalistmodelsthataddressthebehavioralphenomenadiscussedbelowwithwhichtocompareourap-proach;whetheralocalistmodelcouldexhibitthesamebe-haviorisnotclearinadvanceofattemptingtoimplementone.Anysuchmodelwouldtreatthebehaviorashavingaverydifferentbasisthanoursdoes,however.5
Becauseitusesdistributedrepresentations,ourmodeldepartsfromthecommonmetaphorof“accessing”themeaningofaword(seeSeidenberg,1987;Seidenberg&McClelland,1989,fordiscussion).Thelexicalaccessideaaroseinthecontextofearlymodelsinwhichawordwassaidtoberecognizedwhenitsentryinlexicalmemorywascontacted,throughanactivation(e.g.,Morton,1969)orsearch(Forster,1994)process,creatingwhatBalota(1990)calledthe“magicmoment”oflexicalaccess.Thelexicalentryactedasanindexforwheretofindassociatedtypesofinformation,includingaword’sspelling,sound,andmean-ing.Therepresentationsfordifferentwordsweredistinct
MULTICOMPONENTMODELOFREADING
9
fromeachother,andthereforeisolable,asinadictionary.
Ourmodelhasadifferentcharacter.Processingdoesnotinvolveaccessingthelexicalrepresentationforawordbecausetherearenone.Allweightsonconnectionsbe-tweenunitsareusedinprocessingallwords.Thehid-denunitsthatmediatethesecomputationsallowthemodeltoencodecomplexrelationsbetweencodes,butindividualhiddenunits(orsubsetsofthem)arenotdedicatedtoindi-vidualwords(theycannotbebecausetherearemanyfewerhiddenunitsthanwordsinthemodel’svocabulary).Therepresentationofawordisnotisolable;thus,itcouldnotbecutoutofthenetworkwithoutaffectingperformanceonallotherwords.Ratherthanattemptingtoaccessthestoredlexicalentryforaword,themodeltakesaspellingpatternasinputandcomputesitssemanticandphonologicalcodesondemand.Thereisnomagicmoment;themodelisady-namicalsystemthatsettlesintoastablepatternofsemanticactivationoverseveraltimesteps,basedoncontinuousbuttime-varyinginputfromorthsemandorthphonsem(asdetailedbelow).Thus,theweightsinthemodelallowameaningtobecomputedfromanorthographicinputpat-tern;meaningsarenot“accessed”inthestandardsense.6Althoughtheknowledgethatpermitsthenetworktocom-putethemeaningofeachwordisstoredinthenetwork,meaningsarenotthemselvesaccessedinthestandardsense.
10
HARM,SEIDENBERG
namicalsystemwhosestatevariesincomplexwaysovertime.Propertiesoftheattractorbasinshaveimportantef-fectsonthereadingsystem’sdynamics.Forexample,PlautandShallice(1991)examinedhowthesemanticdimensionofabstractness/concretenesswasrelatedtotheparaphasiasofdeepdyslexicpatientsandthedynamicsofaseman-ticattractorthatencodesthistypeofinformation.Furtherconstraintsontheformationofsemanticattractorsaredis-cussedinSection3.Ourmodelingbuildsonthisearlierresearchimplicatingattractorstructuresintheexplanationofreadingandotherphenomena.
5.Pre-existingKnowledge
Themodelisconcernedwiththecentraltaskcon-frontingabeginningreader:learningtocomputemeaningsfromprint.Inlearningtoreadchildrenutilizepre-existingperceptual,learning,andmemorycapacitiesthatarenotreading-specific,aswellaspre-existingknowledge(e.g.,ofspokenformsofwordsandtheirmeanings,andknowl-edgeoftheworld).Ourmodelisnotageneralaccountofperceptual,cognitiveorlinguisticdevelopment;rather,itaddressesthequestion,givensuchpre-existingcapacitiesandtypesofknowledge,howisthetaskoflearningtomapfromprinttomeaningaccomplished?Thusitfocusesonwhatisnovelaboutreading,thefactthatitinvolveslearningaboutorthography,andinparticularhowcharacteristicsoftherelationshipsbetweenthewritten,spoken,andsemanticrepresentationsaffectlearningandskilledperformance.
Ingeneral,thedesignofthemodelinvolvedmak-ingminimalassumptionsaboutthenatureoftheortho-graphic,phonologicalandsemanticcodes,whileincor-poratingstrongassumptionsabouttherelationsbetweenthem.Considerfirstthemodel’sphonologicalrepresen-tations.Phonologicalinformationplaysacriticalroleinlearningtoread(seeRayneretal.,2001,foranoverview);thequalityofpre-readers’phonologicalrepresentationsisrelatedtotheirsuccessinlearningtoreadandtosomeformsofdyslexia.Manyissuesconcerningtheroleofphonologyinreadingacquisitionanddyslexiawerediscussedinourpreviouswork(Harm&Seidenberg,1999).Weassumethatphonologydevelopsasanunderlyingrepresentationthatmediatesbetweentheproductionandcomprehensionofspokenlanguage,butdidnotattempttomodelthis(seePlaut&Kello,1999,however).Rather,wegavethemodelthecapacitytoencodephoneticfeaturesandthentraineditonthemappingsbetweenthephonologicalandseman-ticpatternsformanywords.Thepretrainedphonology-semanticscomponentwastheninplacewhenthemodelwasintroducedtoorthography.
Thisdesignfeatureisimportantforouraccountofthebehavioralphenomenaaddressedbelow.Thesephenomenaconcerntherelativecontributionsoftheorthphonsemandorthsempathwaysoverthecourseofreadingac-quisition.Theformer(thephonologically-mediatedpath-way)doesinvolveanextra“step”comparedtothedirect(orthsem)pathway,asmanyhaveobserved;howeverthechild’slearningtousethemediatedpathwayisfacilitatedbythefactthatpartofitisalreadyknown.Henceitwasimportanttorecreatethisconditioninthemodeling.
Thephonologicalrepresentationthatweuseddoesnotcaptureallaspectsofphonologicalknowledge,norhaveweattemptedtosimulatethecourseofphonologicalac-quisition,issuesthatareclearlybeyondthescopeofthepresentproject.Leavingasidethispragmaticissue,theuseofthisrepresentationisjustifiableonindependentgrounds.Thefeaturesetweusedwasdrawnfromphoneticresearch,wheresuchrepresentationsareoftenemployeddespitetheirinherentlimitations,becausetheycapturegeneralizationsatalevelthatisappropriateforanimportantrangeofphe-nomena.Ouruseofthistypeofrepresentationhasasimilarbasis:itispitchedatalevelthatisusefulandappropriategiventhetypeandgrainofthebehavioraldatathataread-dressed.Themainlimitationsofthisfeatureschemeariseinconnectionwithfactsaboutmultisyllabicwords(e.g.,as-signmentofsyllabicstress),butthepresentmodelislimitedtomonosyllables.Similarly,althoughphonologicalknowl-edgecontinuestodevelopthroughtheearlyyearsofschool-ing(Vihman,1996;Locke,1995),muchofthesystemisinplacebyaboutage5.Theadditionallearningthatoccursagainmainlyinvolvesmorecomplexwordsthanusedinthecurrentmodel.Thus,phonologicalacquisitionissimilartotheacquisitionofsyntaxinsofarasbothsystemsarelargelyinplacebythestartofschooling,althoughbothcontinuetoberefinedwithadditionalexperience.7
Insummary,theheuristicvalueofphoneticfeaturerep-resentationsisclearfrompreviousresearch.Weassumewithmanyothersthatthefeaturesareapproximationsthatwilleventuallybeexplainedintermsofmorebasicper-ceptualandarticulatory-motormechanismsthatgiverisetothem(see,forexample,Browman&Goldstein,1990).
Thesemanticfeaturesthatwereusedhadasimilarra-MULTICOMPONENTMODELOFREADING
11
tionale.Thegoalwasnottoaddressissuesaboutthestruc-tureofconcepts,orthecontributionsofinnateandexperien-tialfactorstotheirdevelopment.Norwouldweclaimthatknowledgeofwordmeaningsisexclusivelyrepresentedintermsoffeaturalprimitivesorthatsuchafeatureschememerelyneedstobescaledupinordertoaccountforabroaderrangeofsemanticphenomena.Rather,therepre-sentationentailedmakingminimalassumptionsaboutthebeginningreaders’knowledgeofwordmeaningsinordertoexamineamorebasicissue,theeffectsofthedifferingmappingsbetweencodesonhowthereadingsystemde-velops.Thus,themodel’ssemanticrepresentationsreflecttheassumptionthatmeaningsarecomposedoutofelementsthatrecurinmanywords;thatdifferentmeaningshavedif-ferentrepresentations(e.g.,themeaningsofhomophonessuchasPEAR/PARE/PAIRweredistinct);andthatmeaningsarecomputedovertimeratheraccessedataninstantaneousmoment.Inaddition,thereadingmodelwastrainedinamannerconsistentwiththeobservationthatchildrenknowthemeaningsofmanywordsfromspokenlanguageattheonsetofreadinginstruction.Furtherdetailabouttheprop-ertiesofthesemanticrepresentationsisprovidedbelowandinHarm(2002).Likethephoneticfeatures,thesemanticfeaturesalsohaveheuristicvalue:theyhavebeenshowntoprovideagoodapproximationtothekindsofinformationthatareinitiallyactivatedwhenwordsareread,asindexedbymeasuressuchassemanticpriming(McRae&Boisvert,1998;McRae,deSa,&Seidenberg,1997;Plaut&Booth,2000).Theserepresentationshavealsobeenusedtounder-standselectivepatternsofsemanticimpairmentfollowingbraininjury,theprogressivelossofsemanticinformationduetodegenerativeneuropathology,andtheneuralbasesofsemantics(Patterson&Hodges,1992;Hinton&Shallice,1991;Patterson,LambonRalph,Hodges,&McClelland,2001;Gainotti,2000).Asinthecaseofphoneticfeatures,weassumethatthefeaturalsemanticrepresentationsareap-proximate;thatsemanticphenomenawillultimatelybeex-plainedintermsofmorebasicbiologicalandexperientialfactors;andthatsuchatheorywillexplainthefeaturesqueaspectsofbehavioridentifiedinstudiessuchastheafore-mentionedones.
Finally,wegavethemodelthecapacitytoencodeletterstringseventhoughinrealitychildrenhaveonlypartiallymasteredthisbythestartofformalinstruction.Apropertreatmentofthenatureofletterrecognitionandhowthisskillisacquiredgoesfarbeyondtheissuesaddressedhere.Weassumethatthissimplificationhadasimilarimpactonboththeorthsemandorthphonsemcomponentsofthesystemandthereforehadlittlebiasingeffectonthere-sults.
Insummary,weapproximatedsomeaspectsofthechild’sknowledgeandexperienceinordertoexploreacen-tralissueinconsiderabledetail.Everycomputationalmodelnecessarilyinvolvessuchsimplifications;forfurtherdiscus-sionseeSeidenberg(1993).Theparticularsimplifications
wemadewereappropriatebecausemoregeneralpropertiesofthetaskandnetworkexertmuchgreaterinfluenceonthetargetphenomena.Thusthegrainofthesimulationmatchesthegrainofthebehavioralphenomenatobeexplained.
6.Learning
Themodelinstantiatestheideathatlearningtoreadinvolveslearningthemappingsbetweenlexicalcodesandthatthisastatisticallearningproblem,solvedusingastatis-ticallearningprocedure.Thecorrespondencesbetweenthecodesdifferinthedegreetowhichtheyarecorrelated,andnoneofthecorrelationsareperfect.Thechildhastolearnthat-AVEisalwayspronounced//exceptinthecontextofH-,whereasOUGHispronounceddifferentlyinthecon-textsR-,C-,D-,PL-,THR-andcoda-T.Similarly,BEAKandBEAKSoverlapinmeaningwhereasBEATandBEASTdonot.Therelationsbetweencodesareprobabilistic,andlearningisstatisticalinthesenseofbeingdrivenbythefrequencyandsimilarityofpatterns.Theweightsreflecttheaggregateeffectsofexposuretomanypatternsratherthanlearningasetofrulesorexemplars.Thereisgoodevidencethatpeople(includingbabies;Saffran,Aslin,&Newport,1996)andotherspeciesengageinthistypeoflearning,anditsneurobiologicalbasesarebeginningtobeunderstood(O’Reilly&Munakata,2000).
Likeotheraspectsofthemodel,weattemptedtocap-turecorecomponentsofthistypeoflearningandmadesim-plifyingassumptionsaboutothers.Threeaspectsoflearn-ingneedtobeconsidered:thenatureofthelearningpro-cedureitself,thenatureoftheinput(“experience”)fromwhichthemodellearns,andtherelationshipbetweenthistrainingprocedureandthechild’sexperience.Earlymod-elssuchasSeidenbergandMcClelland(1989)usedasu-pervisedlearningprocedurecalledbackpropagation,whichissuitablefortrainingstrictlyfeedforwardnetworks.Inthepresentmodelweusedavariantofbackpropagationthatissuitablefortrainingattractornetworksthatsettleintopatternsovertime.Detailsofthelearningprocedureareprovidedbelow.Heretheimportantpointisthatlearninginvolvedpresentingaletterpatterntothemodel;lettingitcomputesemanticoutput;comparingthecomputedoutputtothecorrect,targetpattern;andusingthediscrepancytomakesmalladjustmentstotheweights.Throughmanysuchexperiencestheweightsgraduallyassumevaluesthatyieldaccurateperformance.
Theprimarymotivationforusingbackpropagationisitsapparentrelevancetothebehaviorinquestion.Thedemandsofthereadingtaskappeartoexceedthelimitedcomputationalcapacitiesofnetworkstrainedusingotherprinciples(e.g.,Hebbianorreinforcementlearning).Thenetworkhastobothlearnthewordsinthetrainingsetandrepresentthisknowledgeinawaythatsupportsgeneraliza-tion.Thetaskthereforerequiresthecomputationalpowerprovidedbymultilayernetworkstrainedusingalgorithms
12
HARM,SEIDENBERG
suchasbackpropagation.Thefactthatthisalgorithmissufficientlypowerfultolearnthetaskandthefactthatmod-elstrainedusingthisproceduresimulatedetailedaspectsofhumanperformanceisconsistentwiththeconclusionthattheprinciplesbywhichpeoplelearnhavesimilarproper-ties.Thebrainmayachievethistypeofperformancebyus-ingbackpropagationoranotherlearningprincipleorcombi-nationofprinciplesthathavesimilareffects,althoughthisissueisunresolved(seeO’Reilly&Munakata,2000,fordiscussion).
Asecondcomputationalconsiderationisthatthebackpropagationprocedureresultsincooperativelearningacrossdifferentpartsofthesystem:thus,theperformanceofeachcomponentissubjectnotonlytoitsownintrinsiccapabilitiesbutalsotothesuccessesandfailuresofothercomponents.Inpracticethispressuresthesystemtopro-ducethecorrectoutputusingwhatevermeansareavailable.Ifonecomponentofthesystem(e.g.,orthphonsemororthsem)failsorisslowforagivenitem,thisgenerateserror.Thiserrorcanarisefrommanysources:itmayarisebecausethemodelhasreceivedinsufficienttrainingtohavelearnedamapping;becausethemappingisadifficultone,suchasspellingtomeaning;orbecauseofambiguitiesinthetrainingsetthatlimitperformance(e.g.,homophonyinthemappingfromsoundtomeaning).Giventhenatureofthelearningprocedure,theerrorthatonecomponentissloworunabletoreducecreatespressureforthesystemtomakeupthedifferencesomewhereelse.Hence,eachcomponentofthesystemissensitivetothesuccessesandfailuresofothercomponents.
Thistypeoflearningcontrastswithmechanismsthatarecorrelativeratherthandrivenbyerror,theclassicex-amplebeingHebbianlearning(Hebb,1949).Insuchsys-tems,learningofanitembyonecomponent(again,forex-ample,orthsem)wouldbeindependentofthesuccessorfailureoforthphonsemforthatitem.However,itwillbeshowninsubsequentsectionsthatthedivisionoflaborthatresultsfromusingtheerror-correctinglearningalgo-rithmplaysanimportantroleinaccountingforbehavioralphenomena.Weviewthemutualdependencebetweendif-ferentcomponentsofthesystemasacentralpropertyofthereadingsystemwhichemergesinthecourseoflearning.
Inourmodel,then,thecomputationofmeaningfromorthographyisaconstraintsatisfactionproblem:Thecom-putedmeaningistheoutputpatternthatbestsatisfiestheconstraintsencodedbytheweightsonconnectionsinthenetwork.Inreading,theweightsincludethosemediatedbyboththeorthsemandorthphonsemcomponents.Learninginvolvesfindingasetofweightsthatyieldsthebestperformancepossiblegiventhecapacityofthenet-workandthestructureoftheinput.SeeRumelhartetal.(1986)fordiscussionofconstraintsatisfactionprocessesinPDPmodels,andSeidenbergandMacDonald(1999)foranoverviewoftheroleofconstraintsatisfactioninseveralaspectsoflanguageuse.
Thefactthatourmodelinvolvesacooperativedivisionoflaborutilizinginputfromallpartsofthesystemcanbecontrastedwithotherrecentmodels.InColtheartetal.’s(2001)DRCmodel,twoprocedures(oneinvolvingrules,theotheralocalistconnectionistnetwork)passactivationtoacommonsetofphonologicaloutputunits.Thiscapturestheideathatthecomputedoutputisdeterminedbyinputfromdifferentsources,anditcontrastswithearlierpronun-ciationmodelsinwhichtheroutesoperateinparallelwitharacebetweenthem(seePaap&Noel,1991;Henderson,1982,fordiscussion).Asidefromthefactthatitiscon-cernedwiththecomputationofpronunciationratherthanmeaning,theColtheartetal.modeldoesnotincorporatetheideathatthecontributionsofdifferentpartsofthesystemaremutuallydependentandemergeinthecourseoflearn-ing.Inourmodel,whatonesetofweightscontributestotheoutputdependsonwhatothersetsofweightscontribute,asdescribedabove.Incontrast,thecontributionsoftheroutesinDRCareindependentlydeterminedbytheintrinsiccom-putationalcharacteristicstheyareassigned.Theseintrinsiccharacteristicsincludethefactthattherulesareformulatedsothattheygeneratecorrectpronunciationsforonlysomewords(e.g.,MINTandLINTbutnotPINT),andtheroute-specificparametersthatdeterminetheirspeeds.Coltheartetal.’simplementationofasysteminwhichtwopathwaysjointlydetermineoutputisamajorsteptowardaconstraintsatisfactionsystem,butdoesnotincorporatetheideaofmutualdependencebetweendifferentcomponentsarisingthroughacommonlearningmechanism.
MorecloselyrelatedtoourmodelistheworkbyPlautetal.(1996),whichlikeDRCaddressedmechanismsin-volvedingeneratingpronunciationsfromprint.Plautetal.proposedthatpronunciationsaredeterminedbyinputfrombothorthphonandorthsemphoncomponentsofthelexicaltriangle(Figure1).Specifically,theyas-sumedthatthedivisionoflaborinpronunciationissuchthatthecontributionfromorthsemphonisgreaterforwordswithatypicalpronunciations(suchasPINT)thanforwordswithmoreconsistentspelling-soundcorrespon-dences,whichwereencodedbytheorthphonpathway.Theyimplementedamodeloftheorthphoncomputa-tionandsimulatedthecontributionoforthsemphonbymeansofanequationspecifyingthatitsinputincreasesgraduallyovertimeandisstrongerforhigherfrequencywords.Themodelwasthenusedtoaddressissuesconcern-ingthepronunciationerrorsthatoccurinsurfacedyslexia,atypeofreadingimpairmentfollowingbraininjury.
OurmodeloriginatedwithsomeobservationsbySei-denberg(1992a)concerningthecomputationofmeaningindifferentwritingsystems.Seidenbergintroducedtheideathatsemanticscouldbepartiallyactivatedbybothdirect-visualandphonologically-mediatedprocesseswithinthe“triangle”framework:“Accordingtothistheory,codesarenotaccessed,theyarecomputed;semanticactivationac-cruesovertime,andtherecanbepartialactivationfrom
MULTICOMPONENTMODELOFREADING
13
bothorthographicandphonologicalsources.”(p.105).Sei-denbergdiscussedpropertiesofdifferentwritingsystemsthatwouldaffectwhathetermedthe“equitabledivisionoflabor”thatwouldemergeinsuchasystem.Thepresentmodelisanextendedexplorationofthefeasibilityandpsy-chologicalplausibilityofthisidea.UnlikeboththePlautetal.andColtheartetal.models,thedivisionoflaborbe-tweencomponentsdevelopedthroughlearningratherthanexternalspecification.ConsistentwithPlautetal.(1996),orthsemdevelopedmoreslowlythanorthphonseminourmodel.However,Plautetal.’sanalysisofthedivi-sionoflaborwaslimitedandleftopenabroadrangeofpossibilitiesforhowthesystemwouldsolvethecompu-tationofmeaningproblem.Itwasnotclearinadvance,forexample,whetherthemodelwoulddivideuptheprob-lembyassigningsomewordstoorthsemandotherstoorthphonsem(asinsomepreconnectionistaccounts,e.g.,Baron&Strawson,1976)oronthebasisofotherstruc-turalcharacteristics.Asdiscussedbelow,thedivisionoflabortosemanticswasgreatlyaffectedbyfactorssuchashomophonyandvisualsimilarity(whichwerenotrelevanttoearliermodelsofpronunciation)andthetwopathwaysjointlydeterminedthemeaningsofmostwords.
BasesforDerivingSignal
theError
Inbackpropagation,learningdependsonthespecifica-tionofthecorrecttargetor“teacher”inordertogenerateanerrormeasure.Asinpreviousmodels,wemerelyprovidedthetargetoneverytrial,ratherthanattemptingtomodelthesourcesforitorotheraspectsofthechild’sexperience.Sev-eralpointsshouldbenotedinconsideringhowthistrainingprocedurerelatestothechild’sexperience.
Fortaskssuchasreading,forwhichthereisexplicitin-struction,thereoftenisanactualtargetprovidedbyaliteralteacher.Infactchildrentypicallyreceivemoretypesofex-plicitfeedbackthanweusedintrainingthemodel,includ-inginstructionaboutthepronunciationofletters,digraphs,onsetsandrimes,andsyllables.Inthisrespectthemodel’s“experience”ismoreimpoverishedthanthechild’s,makingthelearningtaskmoredifficult.
Inothercasesthechildcanbethoughtofasusingvar-iousstrategiestoderiveateachingsignalratherthanus-inganextrinsically-providedone.Forexample,theremaybepragmaticorcontextualinformationprovidingevidenceaboutthecorrectmeaningsofwordsonsomeoccasions,towhichchildrencancomparetheirowncomputedmean-ings.Theteachingsignalmayalsobeinternallygeneratedonsomelearningtrials.Forexample,thechildmaygen-erateatargetbycomparingthemeaningcomputedonthebasisoforthographytotheonecomputedonthebasisofsayingthewordtooneself(i.e.,throughthespokenwordrecognitionpathway).Thisisaversionoftheself-teachingmechanismdescribedbyJormandShare(1983).Thechildwilloftenremembertheidentityofawordfromprevious
exposuretothetextinwhichitoccurs,orbeabletopiecetogetherthecorrecttargetbyusingaconjunctionofvisualandcontextualclues.
Finally,thehippocampusisthoughttoprovideanim-portantinternalsourcefortheerrorsignal.Briefly,thereisevidencethattherearetwoprincipalformsoflearninginhumansandsomeotherspecies(McClelland,McNaughton,&O’Reilly,1995).Corticallearningisthoughttobegrad-ual,requirerepeatedexperiences,andbesensitivetosimi-laritiesamonginputpatterns.Learninginthehippocampalformationisrelativelyrapid,requiresfewexposures(possi-blyonlyone),andisitem-specific.Accordingtothistheory,therepresentationsofwordsencodedinthehippocampusactasteachersforthecorticalsystem.Thatis,thehip-pocampalrepresentationofawordmaybeplayedbacktothecortexmultipletimes,providingtheteachingsignalforthegraduallearningprocedure.Again,ratherthanmod-elingthiscomponentofhumanlearningandmemorywemerelyprovidedtheteachingsignal.8
Onothertrialsthefeedbacktothechildisincompleteorwhollyabsent.Sometimesthechildmayknowthatacomputedmeaningdoesnotfitinagivencontextbutnotexactlywhatthediscrepancyis;thusthechildreceivespos-itiveornegativefeedbackforaresponseratherthanthecor-rectanswer(”reinforcementlearning”;Barto,1985;Sutton,1988).Incaseswherethereisnointernalorexternalbasisfortheteachingsignal,thechild’sowncomputedresponsemayprovidethebasisforlearning(e.g.,inanunsupervised,Hebbian,manner).Inthenearfutureitshouldbepossibletoimplementamorerealisticlearningprocedureinwhichthespecificityandaccuracyofthefeedbackvariesacrosstrials.Hereitshouldbenotedthatitcannotbeassumedthatprovidingfull,explicitfeedbackoneverytrialnecessar-ilyyieldsfasterlearningorbetterasymptoticperformancecomparedtothemorevariablesituationcharacteristicofchildren’slearning.Thereissomeevidencethatprovidingmorevariable,lessprecisefeedbackmayleadtomorero-bustperformancethanmerelyprovidingthecorrecttargetoneverytrial(Bishop,1995).Theuseofvariabletypesoffeedbackmaydiscouragethedevelopmentofoverlyword-specificrepresentationsinfavorofrepresentationsthatcap-turestructurethatissharedacrosswords,improvinggener-alization,butthisissueneedstobeinvestigatedfurther.
Insummary,theclaimthatlearningthemappingsbe-tweenlexicalcodesisastatisticalproblemiscentraltothetheoryanddifferentiatesitfromtheoriesinwhichlearninginvolvesruleinductionorencodingexemplars.Weusedanerror-correctinglearningalgorithmthatissensitivetodif-
14
HARM,SEIDENBERG
ferencesinthecorrelationsbetweencodesandthuscap-turestherelativedifficultyoflearningtheorthsemvs.orthphonmappings.Italsocreatescooperationbetweendifferentcomponentsofthenetwork,givingrisetothedi-visionoflabordescribedbelow.
Itshouldbeclearfromthispresentationthatthemodelattemptstocapturemuchofwhatthechildlearnsaboutre-lationsbetweenlexicalcodeswithoutaddressingdetailedaspectsoftheirclassroomexperience.Childrentypicallylearntoreadthroughexplicitinstruction,whichrarelyre-semblesatrialofbackpropagationlearning.Ourmodelat-temptstocaptureaformofstatisticallearningthatisim-plicitinthesenseofrecentmodelsoflearningandmemory(seeCleermans,1997,foranoverview).Theostensivegoalofovertinstructionistopromoteexplicitlearning,whichoccursinmanydomainsandmaycontributetothechild’sknowledgeofthelexicon.Ourmodeldoesaddressthisformoflearning.However,itshouldalsobenotedthattherelationshipbetweentheteacher’sexplicitinstructionandhowthechildlearnsfromitiscomplexandnotfullyun-derstood.Whenateacherexplicitlydrawsachild’satten-tiontothesimilaritiesamongBAT,CATandSAT,thechild’slearningmaybemediatedbyanimplicitstatisticalmecha-nismliketheonewehaveemployed.Similarly,whereasateachermaythinks/heisteachingachildapronunciationrule,theeffectofthisexperiencemaybetotunetherepre-sentationofstatisticalregularities.Thereareimportantun-resolvedquestionsabouthowexplicitinstructionalexperi-encestranslateintobrain-basedlearningeventswhichneedtobeaddressedinfutureresearch.Inthepresentcontext,weonlyintendtoshowthatmuchofwhatthechildknowsabouttherelationshipsbetweenlexicalcodesisstatisticalinnatureandcloselyapproximatedbyourmodel,includingthelearningprocedureweemploy.
7.PressuretoComputeRapidly
Themodelincorporatestheassumptionthatthereader’staskistocomputemeaningsbothquicklyandaccurately.Asidefromtheobviouspracticalimportanceofrapidread-ing,datafromeyemovementstudiessuggestthatreadingskillisdeterminedmorebytheefficiencyofcognitivepro-cessesinvolvedincomprehendingwordsintextsthanbytheefficiencyofoculomotorprocessessuchasmakingsac-cades(seeRayner,1998,forareview).Thusweassumedthatthemodelshouldnotonlybedrivenbytheneedtobeasymptoticallyaccurate,buttorecognizeawordrapidlyusingwhateverresourcesareavailable.Thistenetresultsinasystemthatisgreedy:itdemandsactivationfromallavailablesourcestothemaximumdegree.Thisassumptionwasoperationalizedbypenalizingthenetworknotonlyforproducingincorrectresponses,butforbeingslow;errorwasinjectedintothenetworkearlyinprocessingtoencouragethequickrampupofactivity.
Thedecisiontoemphasizebothspeedandaccuracyin
trainingthemodelwasprincipallymotivatedbyobserva-tionsaboutreadingbehavior.However,aswithaspectsofthetrainingregimediscussedinthenextsection,ade-signdecisionthatwasbasedonbehavioralconsiderationsalsocontributedimportantlytothemodel’scapacitytoper-formthetaskandsimulatehumanperformance.Bullinaria(1996)implementedamodelthat,likeours,examinedthedivisionoflaborbetweenvisual(orthsem)andphonolog-ical(orthphonsem)componentsoftheSeidenbergandMcClelland(1989)trianglemodel.Bullinariatrainedthemodelonasmallvocabulary(300words)inwhichseman-ticcodeswererepresentedbyrandombitpatterns.Bulli-naria’smodellearnedtocomputephonologicalcodesfromorthographyandsemanticcodesviatheorthphonsempathway.However,almostnolearningoccurredwithintheorthsempathway.Bullinariaconcludedfromtheseresultsthatreadingproceedsbyorthphonsem,withorthsemcontributinglittle.Inpilotsimulationsweob-tainedverysimilarresults(Harm,1998).
Learningdidnotoccurwithintheorthsempath-wayinBullinaria’smodel(orinourpilotsimulations)be-causetherewasnosourceoferrorthatwouldforceitto.Thesemodelswerenottrainedwithpressuretocomputerapidly.Thephonsempathwayhadbeenpretrained,leav-ingonlyorthphonandorthsemtobelearned.Be-causeorthphonarecorrelatedandorthsemarenot,themodelslearnedtoproducecorrectsemanticoutputviaorthphonsem.Thiswasadequatebecausethesesimu-lationshadvirtuallynohomophones.Ineffect,orthsemhadnothingfurthertocontributeandsolearningdidnotoccurwithinthispathway.
Thesituationchangeswhenthepressuretocomputerapidlyisintroduced.Nowtheorthsempathwayhasachancetolearnbecauseitisashorterpathwaythanorthphonsem.Aswedetailbelow,thisresultsinanel-egantsharingofresponsibilitybetweenthetwopathways.Thissharingisparticularlyrelevanttodisambiguatingthemanyhomophonesinthelanguage,whichwereincludedinthemuchlargertrainingsetusedinthesimulationsde-scribedbelow.
Insummary,thetrainingprocedureemphasizedbothspeedandaccuracy;thisdesignfeaturewasmotivatedbyobservationsaboutthenatureofskilledreadingbutalsobypreliminarysimulationsofBullinaria’sandourownindicat-ingthatspeedpressurepromoteslearningwithintheorth-sempathway.
8.TrainingRegime
Finally,weneedtoconsiderthewaythemodelwastrainedandhowthisprocedurerelatestochildren’sexpe-rience.Childrenlearntoreadinthecontextofotherlin-guisticandnonlinguisticexperiences.Thevarioususesoflanguageareinterspersed:thechildlearnstobothproduceandcomprehendlanguage;learningtoreadisintermixed
MULTICOMPONENTMODELOFREADING
15
withusingspokenlanguage,andsoon.Thewaythemodelwastrainedreflectedthisbasicfactaboutthechild’sexpe-rience.
Asdetailedbelow,thefirstphaseofthesimulationsin-volvedtrainingthemodelonthemappingfromphonol-ogytosemantics(asinlistening)andfromsemanticstophonology(asinspeechproduction).Duringthisphasethemodelwasalsotrainedontasksrelatedtolearningaboutthestructureofphonologyandsemantics.Whichtaskthemodelwastrainedonvariedquasi-randomlyfromtrialtotrial.Thisprocedure(whichHetherington&Seidenberg,1989,termed“interleaving”)contrastswithblockedtrain-ingproceduresinwhichasingletask(orsetofpatterns)islearnedtosomecriterion,atwhichpointtrainingonthattaskendsandtrainingbeginsonasecondtask(McCloskey&Cohen,1989).Thesecondphaseofthemodeling,inwhichorthographywasintroduced,followedthesamelogicalthoughthetrainingprocedurewassomewhatdifferent.Theweightsthatresultedfromthefirstphasewerefrozenandthemodelwastrainedtomapfromorthographytose-manticsandphonology.Asdiscussedbelow,freezingtheweightshasmuchthesameeffectasinterleavingreadingandspokenlanguagetasksbutrequiresmuchlesscomputertime,whichwasasignificantconsiderationgiventhesizeofthemodel.
Themainreasonforusingthisprocedurewastheob-servationthatchildren’sexperiencewithlanguageisnotstrictlyblocked.Althoughwedidnotattempttocloselymodelthechild’spre-readingexperience,thephase1train-ingprocedurewasbroadlyconsistentwiththefactthatpriortotheonsetofreadinginstructionchildrenhaveac-quiredconsiderableknowledgeofphonologicalandseman-ticstructureandthemappingsbetweenthem,andthatthedifferentusesoflanguagethroughwhichthisknowledgeisacquiredareintermixed.
Aswiththepressureforspeeddiscussedabove,al-thoughtheintermixingoftrialswaslargelymotivatedbyfactsaboutchildren’sexperience,thisdesignfeaturealsohadabeneficialeffectonnetworkperformance:usingablockedprocedurecancreatetheeffectthatMcCloskeyandCohen(1989)termed“catastrophicinterference.”Inbrief,McCloskeyandCohen(1989)foundthattrainingasimplefeedforwardnetworkononesetofpatterns(e.g.,arandomlistofwords),followedbytrainingthenetworkonasecondsetofpatternsresultedinunlearningofthefirstset.Thiseffectwasthoughttobeunlikehumanperformanceandtoreflectalimitationonthecapacityofthistypeofnetwork.However,catastrophicinterferenceisrelatedtothestrictblockingoftrials,whichoccursinsomeverballearningparadigmsbutnotinlearningalanguageorlearningtoread.HetheringtonandSeidenberg(1989)foundthatrelaxingthestrictblockingoftrainingtrials(e.g.,providingoccasionaltrialstorefreshlearningonthefirstsetwhilethetrainingthesecond)eliminatedtheinterferenceeffect.Thusthechild’sexperienceinlearninglanguagecoincideswithconditions
thatfacilitatelearninginconnectionistnetworks.9
Thefinalissueconcernsthewayinwhichwordswerepresentedtothemodelduringtraining.Asinpreviousmod-els(Seidenberg&McClelland,1989,Plautetal.,1996),themodelwastrainedonalargevocabularyofwords,withtheprobabilitythatawordwouldbepresentedafunctionofitsfrequency,asestimatedbytheFrancisandKuçera(1982)norms.ThisinsuredthatwordssuchasTHEwouldbepre-sentedmanytimesmoreoftenthanwordssuchasSIEVE.Thisprocedurediffersfromchildren’sexperience;inlearn-ingtoread,childrenstartwithasmallnumberofsimplewordsthatoccurwithhighfrequencyinspeech,andthesizeoftheirreadingvocabulariesexpandsovertime.Weusedthefrequency-weightedsamplingproceduremainlybecauseitiseasiertoimplementthanaprocedureinwhichthesizeofthetrainingvocabularygrowsovertime.Itisalsodifficulttoobtainreliableindependentinformationaboutwhenandhowoftenchildrenareexposedtodiffer-entwords,andthereislikelytobeconsiderablevariabilityacrosschildren.Inrecentworkwehavebeguntoinvesti-gatewhetherwaysofstructuringthetrainingregimehaveanimpactonnetworkbehavior.First,wehavetrainedsomeorthphonmodelsusingdatafromZeno(1995)concern-ingthefrequenciesofwordsinthetextsthatarereadbychildrenatdifferentgradelevelstodeterminewhichwordsarepresentedatdifferentpointsintrainingandhowoften.Wehavealsotrainedanorthphonmodelusingaproce-dureinwhichwordsareintroducedintheorderinwhichtheyoccurinchildren’sbasalreaders(Foorman,Perfetti,Seidenberg,Francis,&Harm,2001).Finally,wehaveex-aminedmorespecificwaysoforderingthewordsinthetrainingregimetodetermineifthereisasequencethatopti-mizesspeedoflearning(Harm,McCandliss,&Seidenberg,2003).Ingeneralthesedifferenttrainingregimesyieldper-formancethatdoesnotdiffergreatlyfromwhatwasob-tainedusingthefrequency-biasedsamplingprocedure.Be-causethewordsareallrepresentedinanalphabet,whatislearnedaboutoneitemcarriesovertootheritemswithwhichitsharesstructure;thisreducesthemodel’ssensi-tivitytoexactlywhenindividualwordsarepresented.Al-thoughwehadinitiallythoughtthatadheringmorecloselytothechild’sexperienceinlearningwordsovertimewouldimprovethemodel’sperformance,wehavenotobserved
16
HARM,SEIDENBERG
strongbeneficialorinterferingeffects.Harmetal.(2003)foundthatwhereasstructuringthetrainingcorpushaslit-tleimpactonnormalperformance,itdidimprovetheper-formanceofamodelthatwasgivenanimpairedcapacitytorepresentphonologicalinformation.Thus,theremaybewaysofoptimizingthetrainingsequenceforchildrenwithcognitiveorperceptualdeficitsthatinterferewithnormallearning;however,withinbroadlimits(seeSeidenberg&McClelland,1989;Plautetal.,1996,fordiscussion),differ-entsamplingproceduresyieldsimilarperformanceinnon-impairedmodels.
Insummary,thesamplingproceduredoesnotliterallycorrespondtothechild’sexperience.However,becauseofthesharedstructureamongwordsinanalphabeticwrit-ingsystem,themodelisnothighlysensitivetohowthetrainingtrialsareordered.Inrealitytheexactsequenceoftrainingtrialsandotherreading-relevantexperiencevariesacrosschildrenandwouldbeexpectedtoaffectwhenspe-cificwordsarelearnedbyanindividual.Inaddition,thesefactorsmayberelevanttodesigninginterventionsforchil-drenwhoarenotlearningtoreadnormally.However,theseissuesarenotcentraltothepresentresearch.
Wenowdescribetheproceduresusedtotrainthemodel,beginningwiththepre-literatespeaking/hearingmodelandcontinuingwiththefullreadingmodel.
3.PHASE1:THEPHONOLOGY
SEMANTICSMODEL
Webeganbyimplementingamodelofthecomputa-tionsbetweenphonologyandsemantics.Thisphasewasintendedtoapproximatetheknowledgeofpre-readers,whohaveacquiredasubstantialspoken-wordvocabulariesandknowaconsiderableamountaboutthephonologicalstruc-tureoftheirlanguageandaboutsemanticstructure(e.g.,thatitcontainsobjects,livingthings,animals,actions,andstates).Learningtoreadbuildsonthisexistingknowledge.Thephonologytosemanticscomputationisrelevanttohowpeoplecomprehendspeechandthesemanticstophonologycomputationtoproduction;however,thesetaskswerenotaddressedindetailinthepresentwork.
NetworkDynamics
Manypreviousmodelshaveemployedasimplefeedfor-wardarchitectureconsistingofasetofinputunits,asetofoutputunits,andasetofhiddenunitsmediatingbetweenthem.Oneachtrial,thejinputunitsujareclampedtosomedesiredvalue.Thehiddenunitscomputetheirval-uesbasedontheinputunitactivityandtheweightswthatmaptheinputunitstothehiddenunits.Eachhiddenunithiforeachoftheihiddenunitscomputesitsoutputvalueashif∑jwijuj,wherefisanonlinearsquashingfunction.Similarly,eachofthekoutputunitsokcomputesitsoutputbasedonthehiddenunitoutputs:okf∑iwkihi.Weights
areadjustedbypropagatingerrorbackwardsthroughthenetworkandmovingeachweightinadirectionthatmin-imizestheerror(thebackpropagationoferroralgorithm;Rumelhart,Hinton,&Williams,1986).
Suchnetworksadheretoaneuralmetaphortotheextentthattheprocessingofeachunitisdrivenbythelocalprop-agationofactivityalongweightedconnections,ratherthan,forexample,acentralprocessingexecutive.However,themetaphorstopsthere.Suchnetworksareexplicitlystate-less,thatis,therearenostatetransitionsinthenetwork;justthefinalcomputedstateinwhichactivityhaspropagatedthroughtheentiresystem.Thereisnotimecourseofacti-vation,noprocessingdynamics,andnosenseinwhichthecurrentstateofthenetworkmodifiesitssubsequentstates.
Recurrentnetworksutilizingbackpropagationoferrorthroughtime(hereafterBPTT;Williams&Peng,1990)ad-dresssomeoftheselimitations.Insuchnetworks,anotionoftimeisadded,suchthattheoutputofaunitattimetdependsnotontheactivityofunitsinapreviouslayer,asinfeedforwardnetworks,butonthatofallunitsataprevi-oustimeslice.Thiskindofnetworkisageneralizationofthefeedforwardnetwork,andallowsforrecurrent,orcyclicconnectivitypatterns.Theactivityofaunituiattimet,utiisdefinedasutif∑jwijutj1.Aunit’sactivityattimetthenistotallydeterminedbytheactivityofallunitscon-nectedtoitattimet1.Thesenetworksformdynamicalsystems,exhibitingeitherstablefixedpointsoroscillatingbehaviors.Further,activitywithinagroupofunitscanbuildupovertime,withtheunitsinfluencingeachother’sstates.
However,thetemporaldynamicsofsuchnetworksarestillquitesimple.Theyoperateinalockstepfashion,wheretheoutputoftheunitisthesquashedsumofitsinputre-gardlessofanythingelse.Theoutputofunits,then,tendsto“jump;”activitydoesnotrampupordowngraduallybutinsteadcanrespondinstantaneously.Hence,whilethenet-workwillexhibitglobaltemporaldynamics,eachindivid-ualunitstillhasaverysimpletimecourseofactivation.
Pearlmutter(1989,1995)formalizedawaytotrainnet-workswithmuchmoresubtletimecoursesofactivity.Con-tinuoustimenetworkssuchasthoseintroducedbyPearl-mutteraddunitdynamics:aunit’soutputrampsupgradu-allyasafunctionofitsinput,basedonaleakyintegratorequation:
σ∂oi
MULTICOMPONENTMODELOFREADING
17
activityoiandwhatitsactivityoughttobe(yi).Insimu-lations,thecontinuousdynamicsdefinedbyEquation1areapproximatedbydiscretesamples.Inthiscase,theoutputoftheunitattimetchangesbythedifferencebetweentheoutputattimet1anditsasymptoticoutputmultipliedbyσ.Totakeaconcreteexample,supposetheinputtoagivenunitwasstrongenoughtoasymptoticallydrivetheoutputto1.0,andtheunit’soutputisinitiallyzero,andoneuti-lizedaσ01.Onthefirstsample,theunit’soutputwouldmovefrom0.0to0.1(increasingbyσtimesthedifferencebetweenactualandasymptoticoutput).Onthenextsam-ple,itwouldmovefrom0.1to0.19(again,increasingbyσtimesthedifferencebetweenactualoutput0.1andasymp-toticoutput1.0).Onthethirdsample,itwouldincreaseto0.271(0190110019).Andsoon.
Pearlmuttergeneralizedthebackpropagationoferrorequationstoallowerrorgradientstobeintegratedupovertime,thewaythatactivityisintegratedupovertime.Thisallowsonetotrainsuchnetworkswiththefullpowerofthebackpropagationoferroralgorithm.
Plautetal.(1996)introducedasubtlebutimportantchangetothePearlmutterequations.ThePearlmutter(1989)formulationhadtheoutputofaunitrampingupovertimeinresponsetotheinstantaneoussquashedinputtothatunit.Plautmadetheoutputofaunittheinstantaneoussquashedvalueoftheinputtoaunit,andcausedtheinputtounitstorampupovertime.Formally,
σ
oi∂yi
fyi
bi(3)
18
HARM,SEIDENBERG
TrainingRepresentations
CorpusandThetrainingcorpusconsistedof6,103monosyllabicwords,consistingofallmonosyllabicwordsandtheirmostcommoninflection,forwhichsemanticandphonologicalrepresentationscouldbederived.Therewere497setsofhomophonescontaining1047words;447havingtwomem-bers,47having3(e.g.,THREW,THROUGH,THRU)and3setshavingfourmembers(e.g.,AIR,ERE,ERR,HEIR).Therewere39wordsinwhichasinglespellingwasas-sociatedwith2ormoremeanings(mainlywordssuchasSHEEP,FISHorHITwhosepluralorpasttensemorphologi-calinflectioninvolvesnochangefromthestem)10.
ThefrequencyofeachitemwascodedusingasquarerootcompressionoftheWallStreetJournalcorpus(Mar-cus,Santorini,&Marcinkiewicz,1993)accordingtotheformula
pi
10With
theexceptionofhomographssuchasWIND,eachword
inthecorpuswasassignedonepronunciation.WedidnotattempttocapturethedialecticalvariationinhowwordsarepronouncedinEnglish.Suchvariationmayhavealargeimpactonaword’spronunciationdifficulty,however.Forexample,POORrhymeswithTOURinsomedialectsandTOREinothers.ThusdifferentneighborhoodsarerelevanttoPOORdependingonhowitispro-nounced.Thisfactorwillaffectthefitofthemodeltobehavioraldata,particularlyifthereisamismatchbetweenthemodel’sdi-alect(roughly,SouthernCalifornian)andthedialectofsubjectstestedinotherregionsorcountries.
toencodethe6,103words.Therepresentationswererathersparse,withthenumberoffeaturesusedtoencodeawordrangingfrom1to37(mean7.6,standarddeviation4.3,me-dian7,outof1,989features).
EightphonemeslotswereusedtoencodetheCCCVC-CCCwords,withvowelcenteringtominimizethe“disper-sion”problem(seePlautetal.,1996).Asetof25phono-logicalfeatureswereusedtodescribeeachphoneme;thesewerederivedfromfeaturematricesinChomskyandHalle(1968),withminormodifications.Allfeatureswerebinary,takingvaluesof0or1.The25featuresperphonemeover8phonemeslotsyieldedatotalof200features.Thefea-turerepresentationsforphonologywereconsiderablymoredensethansemantics:overthewholetrainingset,theav-eragesemanticfeaturewason0.38%ofthetime,whiletheaveragephonologicalfeaturewason5.7%ofthetime.Wedidnotsetouttocreaterepresentationswiththisasymmetryinsparseness,butthisseemstoaccuratelyrepresentanim-portantdifferencebetweenthetwodomains.Thestructureofphonologicalspaceishighlyconstrainedbyarticulatoryandacousticfactors;thusthenumberofpossiblesegmentsissmallandtheycanbedescribedintermsofasmallnum-berofprimitives,creatingalargedegreeofoverlapbetweensegments.Semanticspaceislargerandmorevariable;thiscreateslessoverlap,onaverage,betweenthemeaningsofwordscomparedtotheirsounds.Itturnsoutthatthediffer-enceinsparsenessofsemanticsandphonologyisrelevanttoexplainingmaskingeffectsthatarediscussedbelow.
Architecture
Figure3depictsthemodelusedinthefirstphase.Thesemanticcomponentconsistedofthe1,989semanticfea-turesdescribedabove.Theseunitswereallconnectedto50unitsinthesemanticcleanupapparatus,whichprojectedbackontothesemanticfeatures.Thisarchitecture,whentrainedproperly,iscapableofforming“attractors”inse-manticspacewhichrepairnoisy,partialordegradedpat-ternsandtendtopullthestateofthesemanticunitsintoconsistentpatterns(Plaut&Shallice,1993).
Thephonologicalrepresentationconsistedofthe200phonologicalunits(8slotsof25unitseach),whichpro-jectedontoasetof50phonologicalcleanupunits.Thesecleanupunitsprojectbackontothephonologicalunits.Hereagainanattractornetworkcanbecreatedwhichwillrepairpartialordegradedphonologicalpatterns.HarmandSeidenberg(1999)examinedtheroleofthisattractor,anddamagetoit,inlearningorthographic-phonologicalcorre-spondences.
Thesemanticcomponentmappedontothephonologicalcomponentviaasetof500hiddenunits.Therewasfeed-backinbothdirections.Thenumber500waschosenfrompilotstudies;itisanumberlargeenoughtoperformthemappingwithoutbeingtoocomputationallyburdensome.
MULTICOMPONENTMODELOFREADING
19
SemanticsPhonologyFigure3.Thephonology-semanticsmodel.Duringthispreliteratephase,themodeldevelopedstructurewithinthesemanticandphonologicalcomponents,andlearnedthemappingsbetweenthem.
Training
Phase1involvedtrainingthemodelonthestructureofphonologyandsemantics,andonthemappingsbetweenthem.Themodelwastrainedonfourtasks:aphonologicaltask(10%ofthetrials),asemantictask(10%),aphonol-ogytosemanticstask(“comprehension;”40%)andase-manticstophonologytask(“production;”40%).Trainingonthefourtaskswasintermixed.Onceawordwasse-lectedfortraining,itwasassignedtooneofthefourtasks.Onlinelearningwasused,withwordsselectedfortrainingaccordingtotheirprobabilityofpresentation(Equation6).TomodelthecontinuoustimedynamicsdefinedbyEqua-tion4,adiscretetimeapproximationwasusedinwhichactualtimedefinedbytheintegralwasbrokendownintosmallerunits.Intrainingthenetwork,thenetworkwasrunfor4unitsofwholetime,modeledbyusing12samplesandanintegrationconstantof0.33.
PhonologicalTask
Thephonologicaltaskdevelopsthephonologicalat-tractorandisintendedtoapproximatethechild’sacqui-sitionofknowledgeaboutthestructureofspokenwords(seeJusczyk,1997).ThephonologicaltaskwassimilartothatusedinHarmandSeidenberg(1999),exceptthatitwasmodifiedslightlytoaccommodatecontinuoustimenetworks.Thephonologicalformofthetargetwordwasclampedonthephonologicalunitsfor2.66unitsoftime.Thenatargetsignalwasprovidedforthenext1.33unitsoftime,inwhichthenetworkwasrequiredtoretainthepho-nologicalpatternintheabsenceofexternalclamping.InHarmandSeidenberg(1999),autoconnectionswereusedtogivetheunitsatendencytoretaintheirvalue,butgrad-uallydecay.Toaccomplishthetask,thenetworkhadtolearnenoughofthestatisticalregularitiesoftherepresenta-tionstopreventthisdecay.Inthecurrentsimulations,theideaisthesame,butsincecontinuoustimeunitswereuti-lized,autoconnectionswerenotnecessarytoprovidetheunitswithatendencytograduallydecay;thiswaspartoftheunits’normalprocessingdynamics.
Onthephonologicaltask,onlytheweightsfromthephonologicalunitstothephonologicalcleanupsandbackweremodified.Figure4(a)showstheconnectionsinthemodelthatweretrainedinthistask.HarmandSeiden-berg(1999)foundthattrainingonthistaskallowedthenet-
worktoformattractorswhichallowedittoreliablyrepaircorruptedphonologicalpatternsandgaverisetootherin-terpretablebehavior(e.g.,categoricalperceptionofconso-nants,phonemerestorationeffects).Thus,thetaskcausesthemodeltoabsorbbasicinformationaboutthesoundstructureofEnglish.
SemanticTask
Thesetrialsweredevotedtotrainingthesemanticat-tractor.Thistaskwasconstructedtobeanalogoustothephonologicaltask:thepatternofsemanticunitscorrespond-ingtotheselectedwordwasclampedontotheunitsfor2.66unitsoftime,andthenetworkwasallowedtocycle.Thenthesemanticunitswereunclamped,andthenetwork’staskwastomaintaintheiractivityinthefaceofthetendencyoftheunits’activitytodecayfor1.33unitsoftime.Toac-complishthetask,thenetworkhadtolearnaboutthedistri-butionsofsemanticfeaturesacrosswords,specificallythecomplexcorrelationalstructurethattherepresentationsex-hibit.Encodingthesesystematicaspectsofsemanticstruc-tureallowedtheattractortomaintainpatternsinthefaceofdecay.Thistaskismoredifficultthanthephonologicaltaskbecausetherearemanymoresemanticunitsthanphonolog-icalunitsandthecorrelationsbetweenunitsaregenerallylower.TheconnectionsusedintrainingthistaskareshowninFigure4(b).
ProductionTask
Thistaskinvolvedtrainingthesemanticstophonologypathway(semphon).Itwaslooselybasedonthetaskofproducinganutterance,e.g.,naminganobjectorgenerat-ingfreespeech.Thetaskinvolvedtheproductionoftheap-propriatephonologicalformforawordgivenitssemanticrepresentation.
Onatrainingtrial,thesemanticpatternofawordwasclampedonthesemanticunitsforthefull4unitsoftimeandthetaskwastoproducethecorrectphonology.Theoutputofthephonologicalunitsforthefinal1.0unitsoftimewascomparedwiththetargetvalues;errorwasin-jectedaccordingtothestandardbackpropagationoferrorequations.TheconnectionsusedintrainingthistaskareshowninFigure4(c).Allweightswereupdated,exceptthoseleadingbackintosemantics(becausethevaluesofthesemanticunitswereclamped,noweightchangeswould
20
HARM,SEIDENBERG
Phonology(a). Phonological TaskSemantics(b). Semantic TaskSemantics(c). Comprehension TaskPhonologySemantics(d). Production TaskPhonologyFigure4.Thetasksusedintrainingthephonology-semanticsmodel.Seetextfordetaileddescriptions.
haveresulted).Notethattheweightsinthephonologicalattractorweretrainedaswellasthoseinvolvedinthecom-putationfromsemanticstophonology.
ScoringMethod
Thecomputedsemanticoutputwasconsideredcorrectifeachsemanticfeaturewhosetargetwas1.0wasactivatedtoatleast0.5,andeachfeaturewhosetargetwas0.0wasactivetolessthan0.5;thus,theoutputforeachfeaturehadtobeclosertothetargetthantoitsopposite.Thecom-putedphonologicaloutputwasassessedasfollows.Foreachslotinthephonologicaltemplate,theeuclideandis-tancebetweentherepresentationinthatslotandeachoftheveridicalsetofphonemeswascalculated.Iftheoutputineachslotwasclosesttoitscorrespondingtarget,theout-putwasconsideredcorrect,otherwiseitwasconsideredanerror.
ComprehensionTask
Thefinaltask,comprehension,wasthecomplementoftheproductiontask.TheconnectionsusedintrainingthistaskareshowninFigure4(d).Thephonologicalformofawordwasclampedonthephonologicalunitsforthefull4unitsoftime.Duringthefinal1.0unitsoftimetheoutputofthesemanticunitswascomparedwiththeirtargets.Thetaskwastoproducethesemanticpatternaccurately.
Insummary,themodelwastrainedfor700,000wordpresentations(approximately280,000production,280,000comprehension,70,000semanticand70,000phonolog-icaltrials).Alearningrateof0.2wasusedfor500,000wordpresentations,thenloweredto0.1fortheremaining200,000wordpresentations.Beginningwithahighlearn-ingrateandthenloweringitduringtrainingoftenresultsinfasterconvergencethaneithermaintainingahighlearn-ingrate(whichcanleadtonetworkoscillations),orstartingwithalowerone(whichcandramaticallyslowinitiallearn-ing).
ResultsofTraining
Figure5summarizesthemodel’saccuracyonthepro-duction(generatingphonologyfromsemantics)andcom-prehension(generatingsemanticsfromphonology)tasksoverthecourseoftraining.Attheendoftraining,themodelcorrectlygeneratedphonologicalcodesfor90%ofthewords,andcorrectlycomputedthesemanticsfor86%ofthewordsthatwerenothomophones.Althoughmodelper-formancecouldbeimprovedwithadditionaltraining,our
MULTICOMPONENTMODELOFREADING
21
goalwasnottoachieveperfectperformanceinthisphase,ontheviewthatthe5-yearoldbeginningreaderdoesnothaveperfectknowledgeofall6,000wordsinthecorpus.Thenonhomophonesonwhicherrorsweremadeweregen-erallylimitedtooneortwoincorrectsemanticfeatures(forexample,itrecognizedtheitemPRIMashavingfeaturessuchas[abstraction],[attribute]and[clean],butnot[R4],whichistherandomlygeneratedfeaturethatdistinguishesPRIMfromNEAT).Themodelwasthereforescoredasin-correctlycomputingthefullsemanticsofPRIM,byproduc-ingarepresentationthatisidenticaltoNEAT.
Forthe1125homophones,themodelproducedthecor-rectsemanticpattern26%ofthetime.Fortheotherho-mophones,themodelgenerallyproducedamixoffeaturesfromthealternativemeanings.Forexample,ALEwasin-terpretedas[beverage]atactivitylevel0.70,andastateofbeing(asinAIL)withthe[be]featureatactivitylevel0.61.Thisbehavioristypical;thenetwork’ssemanticunitsarenotdriventoextremevaluesforeitherinterpretation.Thisreflectstheinherentambiguityofthephonologicalform;thenetworkisonthefenceastowhichinterpretationiscor-rect.Suchwordsarenormallydisambiguatedbycontextualinformation.
SimulationthePhonology1:HomophonesinModel
Semantics
Themodelmakeserrorsinproducingthesemanticsformanyhomophonesbecausetheirphonologicalformsareas-sociatedwithmultiplemeanings.Weconductedadditionalanalysestoexaminehowsuchwordswereprocessed.
Method
Stimuli.The1125homophonesinthetrainingsetin-cludedpairssuchasBEAR,BAREandtriplessuchasPAIR,PARE,PEAR.Eachpairofhomophoneswascategorizedasfollows.Ifonewordhadaprobabilityofpresentationmorethan1.5timesthatoftheother,thehigherfrequencyitemwasconsidereddominant,andthelowerfrequencyonewasconsideredsubordinate.Iftheprobabilitiesdidnotdifferbythismuch,theyweretreatedasbalanced.Thisprocedureyielded404dominant,404subordinateand317balanceditems.
Procedure.Threepresentationconditionswereused.Inthe“nocontext”condition,thephonologicalformoftheitemwasclampedontothephonologicalfeaturesandthetrainednetworkprocessedtheitemasusual.Inthe“help-fulcontext”condition,thephonologicalformwasagainclamped,andthemostfrequentsemanticfeaturethatdis-tinguishesthewordfromitshomophonewasalsoclamped.Forexample,forthehomophonouspairBEAR,BARE,the[entity]featurewouldbeactivatedwhenBEARwaspre-sented,andthe[physical_property]featurewouldbeacti-vatedforBARE.Inthe“distractingcontext”condition,the
procedurewasthereverse;thesemanticfeaturefortheop-posingmemberofthehomophonepairwasactivated.Thecomputedsemanticrepresentationwascomparedtothetar-getrepresentationintermsofhits,misses,falsealarmsandcorrectrejectionsandd’wascomputed.Inconditionsinwhichasemanticfeaturewasclamped,thatfeaturewasex-cludedfromthed’calculation.
Results
Figure6summarizestheresults.Inthenocontextcon-ditiontherewasadropoffind’asafunctionoftypeofhomophone.Thisindicatesthatthemodeltendedtodefaulttothesemanticsofthedominant(higherfrequency)sense.Thehelpfulcontextyieldedimprovedperformanceinallconditions,withthebiggestgaininthesubordinatecondi-tion.Thusevenasmallamountofrelevantsemanticinfor-mationwassufficienttopushthesemanticattractortotheless-frequentmemberofahomophonepair.Finally,whenthesemanticcontextwasunhelpful,performancedeclinedrelativetothenocontextcondition,mostprominentlyforthedominanthomophones.Thisisbecausethedominanthomophonesenjoyafrequencyadvantageovertheirsubor-dinateiteminthenocontextcondition;theunhelpfulcon-textpullstherepresentationsfromdeepinthedominantin-terpretationandtowardthesubordinateone.
Thus,intheabsenceofbiasingcontextualinformationthemodelisbiasedtowardproducingthesemanticsofthehigherfrequencymemberofahomophonepair.Theeffectsduetotheadditionofasmallamountofbiasinginformationindicatethatthemodelhadformedattractorsforthealter-nativemeaningsofhomophones;thisinformationpushesthemodeltowardoneoftheattractors.Suchinformationistypicallyprovidedbythesyntactic,pragmatic,ordiscoursecontextsinwhichwordsoccur.
Althoughadetailedexplorationoftheuseofcontextinlexicalambiguityresolutionisbeyondthescopeofthiswork,thisbehaviorofthemodelispromising.Themodelshowssensitivitytofrequencydifferencesbetweenalterna-tivemeaningsofhomophones(examinedfurtherbelow),andalsosuggestsamechanismbywhichcontextualinfor-mationcanaffectthecomputationofmeaning.Itremainsforfutureresearchtoexaminethebehaviorofthemodelwithrespecttotheextensiveliteratureonlexicalambiguity(e.g.,Simpson,1994),particularlytheinteractionbetweenmeaningdominanceandcontextualconstraint(e.g.,Rayner&Duffy,1986;MacDonald,1993).
SimulationRegularities
2:Morphological
ThesemanticrepresentationincludedfeaturesthatareassociatedwithnumberandtenseinflectionsinEnglish.ThusthemodelwastrainedthatapluralformsuchasGOATSwasassociatedwiththesemanticfeaturesforGOATplusthepluralfeature;similarly,apasttenseformsuchas
22
HARM,SEIDENBERG
100
80
Percent Correct: PhonologyPercent Correct: SemanticsNon HomophonesHomophones100
80
6060
4040
2020
0
0K100K200K300K400K500K600K700K0
0K100K200K300K400K500K600K700K
IterationsIterations
Figure5.Developmentcurvesforthecomprehension(left)andproduction(right)tasks.Inthisandallotherfigures,iterationsreferstothenumberofrandomlyselectedtrainingtrials,measuredinthousands(k).Inthecomprehensiontask,themappingfromphonologytosemanticsisinherentlyambiguousforhomophonesandthereforethemodelperformsmorepoorly.
5.0
4.0
DominantBalancedSubordinate3.0
2.0
1.0
0.0
NoContextHelpfulContextDistractingContext
Figure6.Semanticcodesactivatedbyhomophones,measuredind’units.Intheabsenceofcontext,themodeltendstoproducethedominant(morefrequent)meaning.Relevant(“helpful”)contex-tualinformationcausesthemodeltoproducethecorrectmeaning,regardlessofdominance.Distractingcontextualinformation(i.e.,abitofinformationrelatedtoanalternativemeaningofthehomo-phone)wasmostharmfultothedominantmeanings,pullingtheiractivationstothelevelsofsubordinateandbalancedmeanings.
wasassociatedwiththesemanticfeaturesofBAKE
plusthepastfeature.TherewerealsowordssuchasBAKESwhosemostcommonusageintheFrancisandKuceracor-pusisasaverbwithathirdpersonsingularinflection.Therearestrongbutimperfectcorrelationsbetweenthesefeaturesandphonology,reflectingthequasiregularityofthemap-pings.Thepluralfeatureisusuallyassociatedwiththeplu-ralinflectionthatisspelledSandhasthreephonologicalallomorphs(asinLAKES,HANDS,BUSSES);howeverthereareirregularpluralssuchasMENandMICE.Conversely,therearewordsthathavethephonologicalformsofplu-ralsbutarenotplural;theseincludepluraliatantiasuchasPANTSandTIGHTS,andotherssuchasLENSandPONS.Thepasttensebehavessimilarly:thepasttensefeaturewasusuallyassociatedwithoneoftheallomorphsofthein-flectionspelledED;however,therearemanyirregularpasttensessuchasGAVEandformsthatsoundlikepasttensesbutarenot(e.g.,SCOLD,MELD).
Themodellearnedtoproducecorrectsemanticoutputforthewordsonwhichitwastrained;theadditionalques-tionweaddressedwaswhetherthisknowledgewasrepre-sentedinawaythatsupportedgeneralizationtonovel,un-trainedforms.GivenanonwordsuchasGOMESwouldthemodelproduceeitherthepluralorthirdpersonsingularse-manticfeature;givenanonwordsuchasBLAKEDwoulditactivatethepasttensefeature?
BAKED
Semantic d’Method
Stimuli.Thestimuliwerebasedon86nonwordsfromGlushko(1979).Onelistconsistedofpluralformsofthesenonwords(e.g.,GOMEGOMES).Fiveitemsforwhichthe
MULTICOMPONENTMODELOFREADING
23
resultingpluralwasbisyllabic(e.g.,COSECOSES)wereexcludedbecausethephonologicalrepresentationislim-itedtomonosyllables.Pasttenseswerealsogeneratedfromtheseitems,resultingin49monosyllabicstimuli.Thethirdlistconsistedoftheuninflectednonwordsthemselves.
Procedure.Thephonologicalformofeachnonwordwaspresentedtothetrainedmodel,whichprocessedthemusingthenormalparametersforintegrationcon-stantandnumberofsamples.Theactivitiesonthe[plu-ral],[third_person_singular]and[past_tense]featureswererecorded.ForstimulisuchasGOMES,boththepluralandthirdpersonsingulararevalidinterpretations.Asbefore,asemanticfeaturewasconsideredactiveifitsactivitylevelwas05,thatis,itwasclosertotheactivestateof1.0thantheinactivestateof0.0.
Results
Table1
MeaningsActivatedByInflectedWordsAndTheirStems,InPercent
Plural
3rdPluralandPastInflection
Person
3rdPerson
Tense
Note:Valuesdonotaddupto100%becausethemodelsometimesdidnotproduceanyoftheinflectionalfeatures.Table1summarizestheresults.For90%oftheitemssuchasGOMESthemodelactivatedeitherthepluralorthirdpersonsingularfeatureorboth.Thepasttensefeaturewasactivatedfor88%oftheitemssuchasGOMED.Unin-flecteditemssuchasGOMEactivatedthepluralfeatureon1.6%oftheitemsandthepastfeaturefornoitems.Oneoftheuninflecteditemshappenedtobethepseudohomo-phone(DERE)whichactivatedthe[plural]featurebecauseitphonologicallyoverlapswiththewordDEER.Ingeneralthemodelpickedupontheregularitiesconcerningthemappingbetweenthesefeaturesandtheirphonologicalrealizations.Themodel’slevelofperformanceisplausiblegiventhatthecorrelationsbetweenphonologyandthesefeaturesarenotperfect;themodeltreatsmostnonwordssuchasGOMESasinflectedbutnotallbecausesomewordswiththisendingarenotinflected.
ThemodelalsogeneratedsomeactivationofsemanticfeaturesinadditiontothemorphologicalfeaturesshowninTable1.However,thesefeaturestendtoberatherweaklyactivated,relativetothesemanticactivationthatwordspro-duce.Plaut(1997)usedameasurecalledstresstoquan-tifytheextenttowhichfeaturesweredriventoextremalvalues.Plaut’smethodwassymmetric:aunitthatwas
stronglydriventozeroprovidedthesamestressasonedrivenequallycloseto1.However,thisnetworkhassuchstrongnegativebiasesonsemanticfeatures(owingtotheirsparseness),includingsuchnegativestressresultstendstowashoutanyvariationinpositivestress.So,forthisdemon-strationweexaminedonlypositivestress;theextenttowhichunitsweredrivenon.Formally,forunitswhoseout-putwas05,stresswascomputedusingtheformulausedinPlaut(1997):
sj
ojlog2oj
1
ojlog21
oj
log205(7)
Wecomputedthemeanstressfortheinflectednonwordsandwords,andthestressvaluesforthethreemorphologicalfeaturesforthenonwords.Figure7showsthedistributionofstressvaluesforthewordsandnonwords,andforeithertheplural,thirdpersonorpasttensefeature,whicheverwasgreater.
60WordsepNonwordsTy50Nonword Morph hcaE40 fo sm30etI fo20 tnecre10P00.0
0.20.40.60.81.0
Stress Value
Figure7.Semanticstressvaluesforwords,nonwordsandnon-wordmorphologicalfeaturesfromSimulation2.
Thestressvaluesforthewordstendtobeconcentratedatthehigherendofthescale,whilethenonwordsaremuchweaker.Themeanstressforallsemanticfeaturesfornon-wordswas0.58,butthestressofmorphologicalfeaturesfortheseitemswasreliablyhigher,at0.84(F001).Additionally,thestressforwords1(mean:11094,p00.87)wasreliablyhigherthanforthenonwords(FOverall,themodelstronglyactivated117997,p0001).morpho-logicalfeaturesforinflectednonwords,andsemanticfea-turesforwords,buttheactivationofothersemanticfeaturesfornonwordswasfarlower.
TosummarizethePhase1results,themodellearnedtoaccuratelymapbetweenphonologyandsemanticsfor
24
HARM,SEIDENBERG
alargenumberofwords,subjecttolimitationsimposedbytheambiguitiesinherentinhomophonesandnonwordssuchasGOMES.Themodelencodedsomebasicaspectsoflexicalknowledgethatchildrenpossessbeforetheonsetofreadinginstruction.Wenowturntothesecondphase,inwhichthetaskoflearningtomaporthographicpatternsontophonologyandsemanticswasintroduced.
5.PHASE2:THEREADING
MODEL
Architecture
Figure8showsthearchitectureofthereadingmodel.ThetopsectionisthePhase1modeldescribedabove.Aslot-basedlocalistrepresentationwasusedtorepresentthespellingofaword,asinseveralpreviousmodels.Theor-thographicfeaturesweredefinedbycreatingtenslotsof26featurescorrespondingtothelettersofthealphabet.Theslotswerearrangedinavowelcenteredtemplate.Thefea-tureswerethenprunedbyremovingfeaturesinslotsthatneveroccurredinthetrainingset(forexample,onlythelettersC,P,SandToccurredthreepositionsbeforethevowel).Thisresultedin111orthographicunits.Onesetof500hiddenunitsmediatethemappingfromtheseor-thographicunitstosemantics,formingtheorthsempath-way.Similarly,asecondsetof100hiddenunitsmediatetheorthphonpathway.Thenumberofhiddenunitsintheorthphonpathwaywasthesameasinpreviousmodels.Moreunitswereusedintheorthsempathwaybecausethemappingismoredifficult.Varyingthenumberofhid-denunitsaffectsperformanceinwaysthatareinterpretableintermsofindividualdifferencesamongreaders(Seiden-berg&McClelland,1989),butwedidnotexaminethisfac-torinthepresentwork.ThearchitectureofthephonologysemanticscomponentwasidenticaltothatusedinthePhase1model.TheintegrationconstantandnumberofsamplesforthereadingmodelwerealsothesameasinthePhase1model.
SemanticsPhonologyOrthographyFigure8.Theimplementedreadingmodel.Thesemantics-phonologycomponentwastakenfromthemodeltrainedinPhase1.
Themodelalsoincludedasetofconnectionsmappingorthographicunitsdirectlyontophonologicalunitsandan-othersetmappingorthographicunitsontosemanticunits.Theformerwereaddedbecausetheytendtoimprovegen-eralization.Thelatterwereaddedchieflyforsymmetry.TheinclusionofthedirectconnectionsfromorthographytophonologywassuggestedbytheworkofZorzi,Houghton,andButterworth(1998),whoexploredaspellingtosoundmodelthatcontainedboththeseconnectionsandthemoreusualorthographyhiddenphonologyconnections.Theycharacterizedtheirmodelasadual-routemodel,withthedirectconnectionscorrespondingtoasublexicalrouteencodingregular,rule-governedmappings,andthehiddenunitpathwaycorrespondingtoalexicalroutenecessaryforexceptions.Whenonlydirectconnectionswereimplemen-ted,theirmodelperformedquitewellreadingnonwords(100%correct)andpoorlyonexceptions(14%correct).Theythenexaminedamodelcontainingbothdirectcon-nectionsandconnectionsmediatedbyhiddenunits.Whenthehiddenunit-mediatedpathwaywasselectivelyimpaired,performanceonregularwordswasspared(atornear100%correct)butexceptionswereimpaired(approximately45%correct;seeFigure12inZorzietal.,1998).Puttingthesetwopiecesofinformationtogether,themodelseemedtobeaconnectionistimplementationofthedual-routemodelwithseparatemechanismsforregular/rule-governedwordsandexceptions.
Inexploratorysimulationswefoundthatincludingdi-rectconnectionsbetweenorthographyandphonologyim-provedperformance,facilitatingthelearningofregularitiesthatsupportnonwordreading.Wethereforeincludedtheminthemodeldescribedbelow.However,wedisagreewiththefurtherclaimthatthehiddenunitanddirectconnectionpathwaysbecomehighlyspecializedforexceptionsvs.reg-ulars,respectively.Zorzietal.’sownmodeldoesnotex-hibitahighdegreeofspecializationandneitherhavethemodelsweimplemented.Thedirect-connectionspathwayintheirmodelreadregularsmuchbetterthanexceptions;however,thehiddenunit-mediatedpathwaydidnotreadex-ceptionswellatall.Zorzietal.(1998),Table7presentedthemodel’sperformancefortenrepresentativestimuli;themodelwiththedirectconnectionssevereddidnotproducethecorrectpronunciationsforanywords,regularorexcep-tion.Readingexceptionscorrectlyapparentlyrequiredin-putfrombothpathways.Thisisprobablybecauseexcep-tionwordssharestructurewithmanyregularwords(e.g.,HAVEoverlapswithHAT,HAS,HIM,HIVE,etc.);thedirectconnectionstendtoencodestrongregularitiessuchasthepronunciationofwordinitialH,whichoccursinbothregu-larandirregularforms.ThusthehiddenunitpathwayintheZorzietal.modelwasnotcomparabletothelexicalrouteintraditionaldualroutemodelsofnamingbecauseitdoesnotproducethecorrectpronunciationsforexceptionsbyitself.
OurmodeldoesnotdividethingsupasZorzietal.de-scribed,either.Wetestedthemodelonasetofexcep-MULTICOMPONENTMODELOFREADING
25
tionsfromPattersonandHodges(1992)andnonwordsfromGlushko(1979).Theintactmodelproducedthecorrectpronunciationsfor88.4%ofthenonwords,and99.2%oftheexceptions.Removingthehiddenunitsmediatingor-thographyandphonologyyielded74.4%accuracyonthenonwordsand40.3%ontheexceptions.Thus,performanceonnonwordswasmoreimpairedthanintheZorzietal.simulation,whereasperformanceonexceptionswaslessimpaired.Thehigherrateofaccuracyonexceptionwordsinourmodelderivesfromthefactthatthereisasemanticpathwaytophonology,incontrasttotheZorzietal.(1998)model.Thesemanticpathtakesresponsibilityformanyoftheexceptionwords,whichisunaffectedbyremovingthehiddenunitsbetweenorthographyandphonology.Thelowerrateofaccuracyonnonwordsindicatesthatthehid-denunit-mediatedpathwayencodedsomeregularthoughcomplexmappingsfromspellingtosound.Thiswasfacil-itatedbytheuseofadistributedphonologicalrepresenta-tionratherthanthelocalistoneemployedbyZorzietal.Insummary,thedirectconnectionsfacilitateperformanceandthereisnoapriorireasontoexcludethem;howevertheresultingmodeldoesnotorganizeitselfintothelexicalandsublexicalroutesintraditionaldual-routemodels.
TrainingRegime
TheweightsthatwereobtainedattheendofthePhase1modelwerefrozenandembeddedinthelargerreadingmodel.ThusonlytheconnectionsfromorthographytootherunitsweretrainedinPhase2.Freezingtheweightsisnotstrictlynecessary;earlierwork(Harm&Seidenberg,1997)usedaprocessofintermixinginwhichcomprehen-siontrialswereusedalongwithreadingtrials.Weightfreezinghasthesameeffectbutissimplerandlesscompu-tationallyburdensometoimplement.Intermixingiseffec-tiveandrealisticbutaddssubstantiallytonetworktrainingtime.
Itemswerepresentedtothenetworkaccordingtothesameonlinelearningschemeasbefore,withthesamefre-quencydistributions.Errorsignalswereprovidedforboththephonologicalandsemanticrepresentationsofaword.
Tocomputationallyinstantiatetheprinciplethatthereadingsystemisunderpressuretoperformrapidlyaswellasaccurately,errorwasinjectedintothesemanticandpho-nologicalrepresentationsearly,fromtimesamples2to12.Thenetworkthereforereceivedanerrorsignalnotonlyifitproducedincorrectsemanticorphonologicalcodes,butifitdidnotproducethemrapidly.
OverallResultsofTraining
Thenetworkwastrainedfor1.5millionwordpresenta-tions.Attheconclusionoftraining,thenetworkproducedthecorrectsemanticrepresentationsfor97.3%oftheitems.Fortheother2.7%ofthewords,itactivatedanaverageof1.6spuriousfeatures,andfailedtoactivateanaverage
of0.8features.Themodelproducedcorrectphonologicalrepresentationsfor99.2%ofthewords.Ontheremaining0.08%ofthewords,itproducedanaverageof1.1incor-rectphonemes.Figure9depictssemanticandphonologicalaccuracyoverthecourseoftraining.
10080tcerro60C tencr40eP20SemanticPhonological00K
250K500K750K1000K1250K1500K
Iteration
Figure9.Accuracyofsemanticandphonologicalrepresenta-tionsoverthecourseoftraining.
Thefocusofthisresearchisonbehavioralphenomenaconcerningtheactivationofmeaning.However,inordertoestablishcontinuitywithpreviousresearchontheactiva-tionofphonology,weexaminedthemodel’sperformanceonsomebenchmarkphenomena:theinteractionoffre-quencyandspelling-soundconsistency,nonwordgeneral-ization,andmorphologicalprocessing.
SimulationRegularity3:Interaction
FrequencybyOnewellestablishedphenomenoninreadingisthefre-quencybyregularityinteraction(Taraban&McClelland,1987;Seidenberg,Waters,Barnes,&Tanenhaus,1984).Thesestudiesexamined“exception”wordssuchasPINTandregularwordssuchasMUST.PINTisanexceptionbecauseINTshouldbepronouncedasinMINTandLINT.MUSTis“regular”insofarasallmonosyllabicwordsendingin-USTrhyme.Thetwofactorsinteract:lowerfrequencyexceptionstakelongertonamethanlowerfrequencyreg-ulars,butthetwotypesofhigherfrequencyitemsdonotdiffer.Theregularvs.exceptiondistinctionwasinheritedfromthedual-routemodel,whichdistinguishesbetweenwordspronouncedbyrule(regulars)andwordsthatvio-latetherules(exceptions).Ourmodelstreatspelling-soundcorrespondencesasacontinuum:spellingdifferwithre-specttothedegreeofconsistencyinthemappingbetween
26
HARM,SEIDENBERG
spellingandsound.“Rule-governed”formsand“excep-tions”representdifferentpointsonthiscontinuum;therearealsointermediatecasessuchasMINT,whichisrule-governedbutinconsistentbecauseoftheirregularneighborPINT;seeJaredetal.(1990)forasummaryofevidencethatdegreeofconsistencyaffectswordnaming.
DatafromTarabanandMcClelland(1987),Experiment1A(fromTable2,p.614)areplottedinFigure10(left).Theconditionsarelabeledasintheoriginalstudy.ThisresultandotherslikeitwerereplicatedbytheSeidenbergandMcClelland(1989)modelandanalyzedbyPlautetal.(1996),whoshowedhowtheinteractionoffrequencyandconsistencyarisesfromcomputationalpropertiesofsimpleconnectionistnetworks.
MethodandResults
ThewordsfromTarabanandMcClelland(1987),Ex-periment1A,wereused.Thereare96wordsinfourcondi-tionsthatresultedfromthecrossingfrequency(high,low)andregularity(regular,exception).
Eachitemwaspresentedtothetrainednetwork.Inprevioussimulationsofthiseffect(Seidenberg&McClel-land,1989;Plautetal.,1996)thedataconcernedthemeansummedsquarederrorforthephonologicalcode,whichwascomputedinasinglefeedforwardstep.Inthepresentmodel,theerrorcomputedattheendofprocessingwasessentiallyzeroforalmostallitems.Thisisbecausethemodelincorporatesaphonologicalattractor,whichtendstopullunitactivitiestotheirextremalvaluesovertime.Inor-dertomeasurethedifficultythenetworkhadreachingthesestates,werecordedtheintegraloftheerroroverthecourseofprocessingtheitem,fromtimestep4tothefinaltimestep,12(thesummationbeganwithtimestep4becauseittakes4samplesforinformationtoflowtophonologyfromorthographyviaallroutes).
ThemeansumsquarederrorisplottedinFigure10(right).Therewasamaineffectoffrequency(F192566,p0019),ofregularity(F192419,p0043),andamarginallyreliableinteractionofthetwo(F192362,p006).Aposthoctestrevealedaneffectofreg-ularityforthelowfrequencyitems(F05)butnosucheffectforhighfrequency1464items(F04,1p00).
Simulation4:NonwordReading
AnimportantissuethataroseregardingtheSeidenbergandMcClelland(1989)modelconcerneditsrelativelypoorabilitytogeneralizetonovelforms(Besner,Twilley,Mc-Cann,&Seergobin,1990;Coltheartetal.,1993),alimi-tationaddressedinsubsequentresearch(Seidenberg,Plaut,Petersen,McClelland,&McRae,1994;Plautetal.,1996;Harm&Seidenberg,1999).Itwasthereforeimportanttoevaluatethenewmodel’sbehavioronthistask.
Method
Themodelwastestedon86nonwordsfromGlushko(1979),Experiment1.Thislistconsistedof43non-wordsderivedfromconsistentneighborhoodsand43de-rivedfrominconsistentneighborhoods.80nonpseudoho-mophonenonwordsfromMcCannandBesner(1987)werealsotested.
Eachnonwordwaspresentedtothemodelandthecom-putedoutputwascomparedtothemostcommon(orinsomecasesthetwomostcommon)pronunciations.Forexam-ple,forthenonwordGROOK,either/
/(asinSPOOK)or//(asinCROOK)wereconsideredcorrect.Seidenbergetal.(1994)foundthatthetwomostcommonpronuncia-tionsaccountedforover90%ofsubjects’responsestoalargesetofnonwords.
Results
Themodelproducedcorrectpronunciationsfor93%ofthenonwordsderivedfromregularwordsand84%oftheonesderivedfromexceptionwords.CorrespondingresultsforthesubjectsintheGlushko(1979)studywere93.8%and78.3%,respectively.FortheMcCannandBesner(1987)stimuli,themodelscored83%correct,whilehu-mansubjectsaveraged88.6%.Themodelperformsslightlyworsethanpeople;thisismainlyduetothefactthattheexceptionnonwordsincludesomespellingpatternsthatdidnotoccurinthetrainingcorpus(e.g.,the-JEinJINJE),andhencecouldnotbefullyrepresentedintheorthographicunits.Thislimitationcouldbeovercomebyusinganon-slotbasedrepresentation(Plautetal.,1996),byexpandingthecorpustoincludemultisyllabicwordsthatcontainthespellingpatterns,orbymodelingadditionalstrategiesthatsubjectsmayuseinpronouncingdifficultnonwords(e.g.,pronounceJINJEbyreferencetoINJURE).
SimulationEffects
5:Imageability
Asnotedintheintroduction,manystudieshavedemon-stratedeffectsofphonologicalvariablesonthecomputa-tionofmeaning.Hereweconsiderthereciprocaleffect,inwhichsemanticpropertiesofwordsaffectnaming.Sucheffectshavebeenobservedinbraininjuredpatientswhoseabilitytocomputefromorthographytophonologyhasbeencompromised.Thustherearesemanticparaphasiasindeepdyslexia(Coltheart,Patterson,&Marshall,1980)andconcretenesseffectsinphonologicaldyslexia(Patter-son,Suzuki,&Wydell,1996).However,semanticeffectsonnaminghavealsobeenobservedinunimpairedreaders.ModelssuchasSeidenbergandMcClelland’s(1989)sug-gestthatmostmonosyllabicwordscanbereadusingtheorthphonpathway.Themodelperformedmostpoorlyonrelativelylowfrequencywordswithatypicalspellingsandpronunciationssuchasangstandbarre.Thusthemodel
MULTICOMPONENTMODELOFREADING
27
650
RegularException0.6
RegularException0.5
625
Sum Squared ErrorLowFrequency
HighFrequency
0.4
RT (ms)6000.3
0.2
575
0.1
5500.0
LowFrequencyHighFrequency
Figure10.Frequencybyregularityinteraction.DatafromTarabanandMcClelland(1987),experiment1A(left),andsimulationresults(right)oftheintegratedsumsquarederror(seetext).
suggestedthatcorrectlyreadingsuchwordsrequiresaddi-tionalinputfromorthsemphon(Plautetal.,1996).
Strain,Patterson,andSeidenberg(1995)testedthispre-dictionbyexaminingeffectsofimageability,asemanticvariable,onthenamingperformanceofskilledadultread-ers.Theirstimulifactoriallyvariedimageability,frequency,andspelling-soundregularity.Theprediction,then,wasthattherewouldbeaneffectofimageability(higherim-ageabilitywordsfasterthanlower)onlyforlowfrequencywordswithirregularspelling-soundcorrespondences.Themainresultsfromtheirstudy,showninFigure11(left),ex-hibitedthispattern.TheStrainetal.resultisimportantbecauserepresentsanonobviouspredictionconcerningtheinvolvementoforthphonseminnamingbasedonanal-ysesofthecapacitiesoforthphon.11Wethereforeexam-inedwhetherthepresentmodelwouldreplicatethiseffect.
Method
StimuliandProcedure.
ManyoftheitemsusedbyStrainetal.(1995)weremul-tisyllabicandcouldnotbeusedinthissimulation.Anewstimulussetexhibitingthesamepropertieswasthereforeconstructed.Wefirstperformedamediansplitofallitems
28
600
HARM,SEIDENBERG
Mean Naming Latency (ms)High ImageabilityLow Imageability1.0
High ImageabilityLow Imageability560
Sum Squared ErrorHFRLFRHFELFE5800.8
0.6
5400.4
5200.2
5000.0
HFRLFRHFELFEFigure11.DatafromStrainetal.(1995)(left)andSimulation5(right).Statisticallyreliableeffectsofimageabilitywereonlyobservedforlowerfrequencyexceptionwordsinbothexperimentandsimulation.Notethatthestimuliintheexperimentandsimulationwerenotidentical,asexplainedinthetext.HFR=highfrequencyregular;LFR=lowfrequencyregular;HFE=highfrequencyexception;LFE=lowfrequencyexception.
6.DIVISIONOFLABOR
Wenowconsiderthecentralissueaddressedinthisre-search,themodel’sdivisionoflaborinthecomputationofmeaninganditsrelationshiptohumanperformance.Wehaveseenthatthemodelwasabletocomputethemeaningsofwordsaccurately.Thequestionis,how:specifically,towhatextentisthecomputationofmeaningdrivenbytheorthsemvs.orthphonsemcomponents?Asafirststepwereportasimulationthatprovidesinformationabouthowrapidlyinputarrivesatthesemanticlayerfromdiffer-entsources.Wethenreportanalysesofhowthemodelper-formedwithoneortheotherpathwaydisabled(“lesioned”).
Simulation6:DynamicsoftheTrainedReadingModel
Thedynamicsofthereadingmodelarecomplex.Thetheoreticalmodelassumesthatactivationspreadsincontin-uoustime,muchlikeelectricityinacircuitorwaterpres-sureinaplumbingnetwork.Thus,inprinciple,activationtosemanticsarrivescontinuouslyfromallsourcesandbuildsovertime.Inpractice,adiscretetimeapproximationisre-quired.Timeissampledandthebehaviorofthenetworkisupdatedateachtimesample.Intrainingthenetwork,fourunitsofwholetimewereused,sampledovertwelvediscretetimeslices;henceeachsamplewas0.333unitsoftimeinduration.Thestrengthofactivationfromeachpath-wayvariesaccordingtofactorsthatwewillexploreintheremainderofthearticle.
Foreachdiscretesample,activityspreadsfromtheorthographicrepresentationstosemanticsandphonologyalongthedirectconnections,andtothehiddenunitsalongthosepathways(Figure8),causingtheactivityinthoseunitstobegintorise.Onsubsequentsamples,asunitsin-
creaseinactivity,theirinfluenceonsubsequentunitsin-creases.Astheinfluenceoforthographyonphonologyin-creases,thatinturninfluencessemantics,whichisalsoin-fluencedbyorthography.Asthesemanticandphonolog-icalrepresentationsbuildup,theyareinfluencedbytheirrespectiveattractors,andbegintoinfluenceeachotheraswell.Inthetheoreticalmodel,activationbuildsupthrough-outthenetworkcontinuously;inpracticeitisacloseap-proximationtocontinuously.Activationofthesemanticrepresentationaccumulatesfrombothpathwaysinthisfash-ion.However,therateatwhichactivitybuildsupalongthevariouspathwaysisafunctionoftherepresentationalca-pacityofthosepathways,andhowtunedtoaspectsofthestimulithosepathwayshavebecome.
Thepurposeofthissimulationwastoexaminethetimecourseofactivationalongdifferentpathways.Thedataconcerntheactivationofsemanticsfromorthography(fromboththedirectandhiddenunit-mediatedpathways),theac-tivationofphonologyfromorthography(againfrombothdirectandhidden-unitmediatedpathways),theactivationofsemanticsfromphonology,andtheactivationofseman-ticsfromthecleanupunits.
Method
StimuliandProcedure.Allwordsinthetrainingsetwerepresentedtothetrainedreadingmodel.Toassessthetimecourseofactivityatamorefinegrain,thenetworkwasrunfor4unitsofwholetime,asintraining,butdis-cretizedover48samples,ratherthan12,givinganintegra-tionconstantof0.083.Thetotalinputtotargetphonolog-icalunitsfromtheorthphonpathwassummedateachsample.Similarly,thetotalinputtotargetsemanticunitsfromorthsem,phonsemandfromthesemanticcleanupunitstosemanticswasmeasuredateachsample.
MULTICOMPONENTMODELOFREADING
29
Results
AsindicatedinFigure12,theinputtosemanticsfromorthographyandtheinputtophonologyfromorthographyriseatverysimilarrates,withtheorthphonhavingasomewhathigherasymptote.Interestingly,thecontributiontosemanticsfromphonologyrisesatamuchslowerrate.Activationfromphonologytosemanticscannotbeginun-tilsignificantactivationbuildsuponthephonologicalunitsfromorthography.Hence,thephonsemlineinFigure12risesatarateproportionalnottotheconstantinputfromor-thography(unlikeorthsemandorthphon),butratheratarateproportionaltotheactivityinphonology,indicatedbytheorthphonline.Hence,whileorthographydirectlyac-tivatessemanticsandphonologyrapidly,thecontributiontosemanticsviaorthphonsemlagsbehind.Thecleanupunitsaretheweakestsourceofinputtosemantics;theirac-tivityisdrivenbyactivityinsemanticsitself,andislimitedbytheverysparsenatureofthesemanticrepresentations.
Figure12demonstratestwoofthekeypropertiesofthismodel.First,theactivationofsemanticinformationisdrivenbyinputfrommultiplesources;thereisnoonepath-waythatisdoingallofthework.Second,thestrengthofthatinputvariesaccordingtopropertiesofthepathways.Inthefully-trainedmodelactivationarrivesmorerapidlyfromorthsemthanorthphonsem.Itisequallyimportanttonote,however,thatovermosttimestepsthereissignif-icantinputtosemanticsfrombothpathways.Moreover,thisanalysisignorestheinteractivitybetweensemanticsandphonologythatoccursintheintactmodel.Asortho-graphicinformationbeginsactivatingsemantics,thatinturnactivatesphonologyviathesemphonpathway,whichinturncanfurtheractivatesemanticsviathephonsempath-way.Thispropertyalsocontributestotheinvolvementofbothpathwaysintheactivationofmeaning.Finally,thecontributionsfromthedifferentpathwaysaremodulatedbyword-specificpropertiessuchasfrequencyandhomophony,asdescribedbelow.
Figure13showshowindividualfeaturesforatypicalitem,BOOT,aresactivatedovertimebytheorthsemandorthphonsempathways,andthetotalofthetwo.The[object],[artifact],[covering]and[footwear]featuresareshown.Formostfeatures,theorthsempathwaydomi-natedthecomputation.However,forthe[artifact]feature,theorthphonsempathwayprovidedgreaterinputto-wardtheendofprocessing.Forallfourtargetfeatures,bothpathwaysareprovidingpositiveinput,thusthesumoftheircontributionisgreaterthaneithereitherpathways’contributionalone.
SimulationtheDivision7:DevelopmentofLabor
of
Wenextpresentaseriesofsimulationsthatprovidefurtherinformationaboutthedivisionoflaborusingale-sioningmethodology.Thefirstofthesesimulationsexam-
12.0Orth->SemOrth->Phon10.0Phon->SemCleanup->Semtupn8.0I tiUn6.0 aneM4.02.00.00.0
0.51.01.52.02.53.03.54.0
Time
Figure12.Inputtophonologicalandsemanticunitsovertime.Activationrisesmostrapidlyforthephonologicalandseman-ticunits,whichareclosesttotheorthographicinput;howeverthephonologicalunitsreachhigherasymptoticlevels,indicatingsomewhatbetterlearningofthismapping.Activationofsemanticunitsfromphonologyoccursmoreslowlybecausethephonolog-icalunitsmustfirstbeactivatedsufficientlybyorthography.
inedthemodel’saccuracyincomputingsemanticsoverthecourseoftrainingunderthreetestingconditions:thein-tactmodel,themodelwithinputfromorthsemdisabled(i.e.,withthedirectandhiddenunit-mediatedorthsemconnectionsdisabled),andthemodelwithinputfromorthphonsemdisabled.Themodelwastestedonallitemsinthetrainingcorpusonceevery10,000trialswitheachconfigurationofthemodel.Thustheintactmodelwastrainedthroughout,buttestedatregularintervalsinthe3waysdescribedabove.12
Themodelwastestedonallwordsinthetrainingcor-puswithperformancescoredasdescribedpreviously.TheresultsaresummarizedinFigure14.Theaccuracyoftheintactmodelrisesrapidly,thenflattensout,growing
30
HARM,SEIDENBERG
20
12
Input to Artifact UnitInput to Object Unit16
Orth->SemOrth->Phon->SemSum of Two Pathways20.016.0Orth->SemOrth->Phon->SemSum of Two Pathways12.088.044.000.0
1.02.03.04.0
0.00.0
1.02.03.04.0
Time
20
Time
2016
Input to Footwear UnitInput to Covering UnitOrth->SemOrth->Phon->SemSum of Two Pathways16Orth->SemOrth->Phon->SemSum of Two Pathways1212884400.0
1.02.03.04.0
00.0
1.02.03.04.0
TimeTime
Figure13.Sourcesoftheactivationofindividualsemanticfeaturesforatypicalword,BOOT.Inthisandsubsequentfigures,orth=orthography,phon=phonologyandsem=semantics.All4typesoffeaturesreceivesignificantinputfrombothdirect(orthsem)andphonological(orthphonsem)pathways,thustheactivationsummedoverthetwosourcesofinputisgreaterthanforeitherpathwayinisolation.
moreslowlyfortheremainderofthetrainingperiod.Ini-tially,theaccuracyoftheintactmodelandthemodelwithonlyorthphonsemparalleleachother,indicatingthatthelatterisdoingmostofthework.Quickly,however,theperformanceoftheintactmodelsurpassesthatofthephonology-onlymodel,whoseperformancereachesasymp-tote.Aftertheorthphonsempathwaypeaks,increasesintheaccuracyoftheintactmodelareduetoadditionallearningwithinorthsem.Notealsothatorthsemcon-tinuestoimproveevenafterlearningintheintactmodelhasslowed.
Figure14revealsanimportantresult.Earlyintrain-ing,thephonologicalpathwayisresponsiblemuchofthe
accuracyoftheintactmodel.Thisisbecauseorthphoniseasiertolearnthanorthsem,forreasonsdiscussedpre-viously.However,theorthsempathwaycontinuestode-velopfortworeasons.First,themodelcannotreadmanyhomophonescorrectlyviaorthphonsemduetotheirin-herentambiguity;second,evenwhenorthphonsemac-tivatesthecorrectsemanticsofaword,theorthsempath-waycontinuestodevelopbecauseofthepressuretorespondquickly.Theorthphonsempathwaymustcomputeanintermediaterepresentation(phonology)toactivateseman-tics;thislimitsitsspeed.Thuswhiletheorthsempath-wayismoredifficulttolearn,ithasthepotentialtoactivatesemanticsmorerapidlythanorthphonsem.Moreover,
MULTICOMPONENTMODELOFREADING
31
100
80
tcerro60
C tencr40
eP20
IntactOrth->SemOrth->Phon->Sem00K
250K500K750K1000K1250K1500K
Iteration
Figure14.Divisionoflaborassessedusinga“lesioning”method.Thedatareflecttheaccuracyofthecomputedsemanticrepresentationsintheintactmodel(inputfrombothpathways),andwitheithertheorthsemororthphonsemcomponentdisabled.Earlyintrainingtheintactmodelperformslittlebetterthantheisolatedphonologicalpathway.However,performanceinthephonologicalpathwayrapidlyreachesasymptote,whileper-formanceinorthsempathwaycontinuestoimprove.
Englishmonosyllablescontainfarmorehomophonesthanhomographs,andthustheorthsempathwayhasmuchlessintrinsicambiguitythanorthphonsem.
OnethingthatisnotclearfromFigure14iswhetherdifferentwordsarebeingreadbydifferentpathways.Itispossible,forexample,thatthemodelcouldpartitionthewordssuchthatsomearelargelyreadviaorthphonsemandothersbyorthsem.Wordscorrectlyreadbythein-tactnetworkwerecategorizedintofourdisjointsubgroups:thosethatrequirebothpathwaystoberead(cannotbereadbyeitherpathinisolation),thosethatcanbereadbyeitherpathway,thosethatcanbereadbyorthsembutnotorthphonsem,andthosethatcanbereadbyorthphonsembutnotorthsem.Figure15showsthisbreakdownoverthecourseofdevelopment.
Asexpected,thereisaninitialburstofwordsthatcanbereadonlybythephonologicalpathway.Thisadvantagebe-ginstofalloffby500Ktrainingtrials,atwhichpointmorewordscanbereadbyeitherroute.Interestingly,atthatpointabout15%oftheitemscanonlybereadbytheorthsemroute.Thisnumbergrowstoabout22%,whereitflattensout.Asymptotically,abouthalfofthewordsareredundant;theycanbereadaccuratelybyeitherroute.Fairlylowper-centagesofitemscanbereadonlybytheconjoinedcooper-ationofbothroutes,byorthphonsem,orbyorthsem.
Thisbehaviorofthemodelisconsistentwithacentralfindinginthereadingacquisitionliterature:theimportance
ofphonologicalinformationintheearlystagesoflearn-ingtoread(Adams,1990;Bradley&Bryant,1983;Liber-man&Shankweiler,1985).Thesysteminitiallyaffordsbothorthsemandorthphonsempossibilities.Devel-opmentwithinthetwosubsystemsisdeterminedbytheirinherentcomputationalproperties:orthphonarecorre-lated,phonsemisknown,andorthsemisdifficulttoacquirebutultimatelyfastertocompute.Thesystem(andbyhypothesisthechild)doesnotchooseaninitialstrategyorswitchstrategiesasskillisacquired;ratheritrespondstothetaskitisassigned:computingthemeaningofthewordquicklyandaccurately,subjecttointrinsiccomputa-tionalconstraints,yieldingtheobserveddivisionoflabor.Themodelalsosuggeststhatthedivisionoflaborgradu-allyshiftsasskillisacquired,withtheorthsempathwaybecomingincreasinglyefficientovertime.
Theseresultsneedtobeinterpretedcarefully,however.TheanalysisinFigure14providesinformationaboutthecapacitiesofeachcomponentofthesystem.Itisclear,forexample,thattheorthphonsemcomponentdevelopsmorerapidlythanorthsem.However,aswehavenoted,intheintactmodelsemanticsreceivesactivationfrombothpartsofthesystem.Thewordsinthe“byeitherpath”con-ditionmakethispointmostclearly.Thefactthattheycanbereadbyeitherpathinisolationmeansthatbothpathswillbestronglyactivatingsemanticsintheintactmodel.Similarly,therearewordsthatcanonlybecorrectlyreadbyorthseminisolation,butitwouldbeincorrecttoinferthatthesewordsonlyreceiveactivationviathispathwayintheintactmodel.Belowwepresentadditionalanalysesbearingonthispoint.
Simulation8:SpeedEffects
Thepressuretoactivatesemanticsrapidlyisanimpor-tantpropertyofthemodel;itiswhatforcestheorthsempathwaytocontinuetodevelopevenforwordscorrectlyrecognizedbytheorthphonsempathway.Inthissimu-lationweexaminedhowtheintactmodelandthetwopathsinisolationcompareintermsofhowrapidlysemanticsisactivated.
Asbefore,allwordsinthetrainingsetweretested.Thetimecourseofsemanticactivationwasassessedasfollows.Thenetworkwasrunfor4unitsoftime,asbefore,butagainafinerdiscretizationwasusedtomorepreciselymeasuretime.Inthissimulation,the4unitsoftimewerediscretizedover48samples,givinganintegrationconstantof0.083.Anitemwasassumedtoberecognizedwhenallsemanticfeatureshadsettled;thatis,theiractivationvaluesdidnotchangebymorethan0.05for0.5unitsoftime(6samples).Settlingtimeswerecomputedforallcorrectitemsandav-eraged.Thismeasurewastakenatvariouspointsindevel-opmentasintheprevioussimulation.
TheresultsareshowninFigure16.Becausethenet-workwaspressuredtoactivatesemanticsquicklyaswellas
32
HARM,SEIDENBERG
100By Either PathOnly By Orth->Phon->SemOnly By Orth->Sem80Only By Bothtcerro60C tencr40eP2000K
250K500K750K1000K1250K1500K
Iteration
Figure15.Accuracyofeachcomponentofthemodelincom-putingthesemanticpatternsforwords.Earlyintrainingcorrectoutputismainlyproducedbythephonologicalpathway,reflect-ingmorerapidlearningwithinorthphonthanorthsem.Thisisconsistentwiththepredominanceofphonologicalrecodinginchildren’searlyreading.Withadditionaltraining,however,thelargestclassconsistsofwordsforwhichbothpathwaysproducecorrectoutput(“ByEitherPath”).Therelativelysmallclassofwordsthatrequireinputfrombothpathways(“OnlyByBoth”)primarilyconsistsofthesubordinatemeaningsofhomophones.Theseanalysesprovideinformationaboutwhathasbeenlearnedineachpathway;however,evenifawordcannotbereadbyagivenpathwayinisolation,itmaycontributesignificantpartialac-tivationintheintactmodel.Infactalmostallwordsreceivesomeactivationfrombothpathways.
accurately,latenciescontinuedtodecreaseevenafteraccu-racywashigh.
Asnotedintheprevioussection,anumberofwordscanbereadbyeitherpathwayinisolation.Thisfactmasksasubtlebutimportantpointthatisrevealedbythelatencyanalyses:theeffectofthetwocomponentsworkingto-getherisdifferentfromtheeffectofeachinisolation.Thespeedoftheorthphonsempatheventuallyflattensout;itsmaximumislimitedbythefactthatitmustcomputeareasonablystablephonologicalrepresentationtobeginactivatingsemantics.Thereisnosuchlimitationontheorthsempathway,whichcontinuestoimproveovertime.Asaresult,theoverallspeedofthenetworkalsoimproveswithtraining.Importantly,thespeedofthenetworkwithbothcomponentsoperatingisfasterthanthespeedofeithercomponentinisolation.Thisarisesbecauseoftheprocess-ingdynamicsofthemodel;asshowninFigure2,therateatwhichaunit’sactivityincreasesisafunctionofthestrengthofitsinputactivation.Thus,thenetworkachievedgreaterefficiencyusingbothcomponents.Thispropertystandsin
4.0Intact Model3.5Orth->Sem)Orth->Phon->Semem3.0iT el2.5oWh2.0( cy1.5entaL1.00.5
0.00K
250K500K750K1000K1250K1500K
Iteration
Figure16.Semanticlatencies(definedinthetext)forwordsprocessedbyindividualpathwaysandbybothtogether,overthecourseoftraining.Themainfindingisthatthetwopathwaysact-ingtogetherproduceoutputmorerapidlythaneitherinisolation.Unitsaremeasuresofwholetime,asdefinedbyEquation4.
contrasttothe“horserace”modelofPaapandNoel(1991),inwhichthelatencytorecognizeawordischieflydeter-minedbywhichoftwoindependentroutesfinishesfaster.
SimulationReduced9:ReadingWithFeedback
Phonological
AsshowninSimulation7,theorthphonsempath-waydevelopsmorerapidlythanorthsem.Inthissimu-lationweexploredtheeffectofreducingthephonologicalfeedbackthenetworkreceived,whichforcedthemodelrelymoreontheorthsempathway.
Method
Materials.Allitemsinthetrainingsetwereused.Procedure.Thereadingmodeldescribedabovewasre-trainedwithachangeinprocedure:feedbackabouttheac-curacyofcomputedphonologicalcodeswasprovidedononly1%ofthetrainingtrials,whereasfeedbackaboutse-manticswasprovidedonalltrials,asbefore.Thesameorthphonsemmodelwasusedandthemodelwasagaintrainedfor1.5milliontrials.
ResultsandDiscussion
Atasymptote,thenormalmodelcomputedthecorrectsemanticsfor97.3%oftheitemsinthetrainingset;themodelwithreducedfeedbackonphonologywascorrecton91.8%oftheitems.Thereducedphonology(RP)model
MULTICOMPONENTMODELOFREADING
33
10080
tcerro60
C tencr40
eP20
NormalRP IntactRP O->SRP O->P->S00K
250K500K750K1000K1250K1500K
Iteration
Figure17.Accuracybypathwayforthenormalmodelandforthemodelwithreducedphonologicalfeedback(RP),overthecourseoftraining.TheRPmodellearnsmoreslowly,withthebiggestdecrementintheorthphonsempathway.
alsotookmuchlongertoreachthislowerlevelofasymp-toticperformance.Figure17showstheaccuracyofthenor-malmodel,theintactRPmodel,andthecomponentpath-waysoftheRPmodel.TheRPmodelexhibitedlessre-lianceontheorthphonsempathwaythroughouttrain-ingcomparedtothenormalmodel,andagreaterrelianceontheorthsempathway.Throughoutdevelopment,theRPmodel’sperformancelaggedbehindtheintactmodel.
Figure18showsthelatenciesofthemodelsoverthecourseofdevelopment.Themeanlatencyoncorrectitemsforthenormalmodelwas0.82unitsoftime,whilethemeanlatencyoncorrectitemsforthereducedphonologicalfeedbacksimulationwas1.08unitsoftime.Thiseffectofsimulationcondition,measuredoveritemswhichwerecor-rectinbothsimulations,wasreliable(F001).
155211825,p0Theasymptoticdifferencesinlatencyandaccuracybe-tweenthereducedphonologymodelandthenormalmodelarenotverylarge.However,therewerepronouncedde-velopmentaldifferences.Reducingfeedbackonthesoundsofwordformssignificantlyreducedtherateatwhichthemeaningsofwordscanbelearned,andthespeedatwhichthiscomputationcanbeperformed.
Thissimulationmakestwopoints.First,itprovidesfurthersupportfortheobservationthatthemodelper-formsmostefficiently(intermsofspeed,accuracy,andrateoflearning)usinginputfrombothcomponents.Second,thesimulationhassomesuggestiveimplicationsregardingmethodsforteachingreading.Oneofthemaincontro-versiesinreadingeducationconcernswhetherinstructionshouldemphasizethecorrespondencesbetweenthespoken
4.0Normal3.5RP Intact)RP O->Sem3.0RP O->P->SiT el2.5oWh2.0( cy1.5entaL1.00.50.00K
250K500K750K1000K1250K1500K
Iteration
Figure18.Latenciesinthenormalandreducedphonologicalfeedback(RP)models,overthecourseoftraining.TheRPmodelcomputessemanticcodesmoreslowly,withthebiggestdecrementagainintheorthphonsempathway.
andwrittenformsoflanguageornot.“Wholelanguage”methodstendtodiscouragethistypeofinstruction,focus-inginsteadondevelopingefficientproceduresfromcom-putingmeaningsdirectlyfromprint.Thepresentsimulationsuggeststhatfailingtoprovidefeedbackaboutspelling–soundrelationsmaymakethetaskoflearningtocomputemeaningsmoredifficult.Thesimulationcanonlybetakenassuggestivebecausewehavenotexaminedallofthefac-torsthatcanplayaroleinlearningtoreadwords;wholelanguagemethods,forexample,oftenemphasizetheuseoflinguisticandnonlinguistictextualinformationandguess-ingstrategiesinplaceofphonologicalrecoding.Moreover,thereductioninphonologicalfeedbackinthesimulationwassevereandsorepresentsanextremecase.Otherfactorsbeingequal,however,feedbackaboutboththemeaningsandsoundsofwrittenwordswillyieldmorerapidacquisi-tionandbetterperformancethanmeaningalone.
Wenowexamineseverallexicalfactorsthathavebeenwidelystudiedinbehavioralexperimentswhichinfluencethedivisionoflabor.
DivisionSimulationofLabor10:ModulationbyFrequency
ofOneissueraisedbybehavioralstudiesiswhethertherelativecontributionsofthedifferentpathwaysdependonwordfrequency,withmoreinputfromorthsemforhigherfrequencywords.Thissimulationexaminedhowfrequencyaffecteddivisionoflaborinthemodel.
34HARM,SEIDENBERG
Method
Stimuli.Itemsfortestingwereselectedasfollows.Thetrainingsetitemsweresortedaccordingtofrequency,and500itemsfromthetoponethirdwereselectedrandomly;thesewerethehighfrequencyitemsusedfortesting.An-other500itemswereselectedrandomlyfromthebottomthird;thesewerethelowfrequencyitemsusedfortest-ing.Thisyieldedaverystrongfrequencymanipulation(t10162915,p0001)wherethemeanhighfre-quencyitemhadaprobabilityofpresentationof0.42andthemeanlowfrequencyitemhadaprobabilityofpresenta-tionof0.05.
Procedure.Thenetworkwastestedoneachitemattheconclusionoftraining,andaccuracyoverthesemanticunitswasrecordedforthemodelwithnoorthsempathway,andthemodelwithnoorthphonsempathway.
ResultsandDiscussion
Figure19summarizesthemainresultsatasymptote.Asexpected,highfrequencyitemswerereadmoreaccu-ratelythanlow;however,frequencyinteractedwithpath-way.Forhighfrequencyitems,theorthsempathwayper-formedmoreaccuratelythantheorthphonsempath-way.Forlowfrequencyitemsthedifferencewasmuchsmaller.Consideringthehighandlowfrequencyitemsovertheorthsemandorthphonsempathways,theinterac-tionwasreliable(χ2594,df1,p0015).Theaccu-raciesfortheintactmodelwere99%forthehighfrequencyitemsand95%forthelow.
Recallthatthemodelispressuredtoproducetheseman-ticsofthewordasrapidlyaspossible,bycreatingerrorforeachsampleoftimethatthemodelhasnotyetsettledtothecorrectsemanticrepresentationforthatword.Overthecourseoftraining,thiserroraffectsthenetworkweights;hencethenetworkispressuredtoreducetherunningerroroverallwordsitistrainedon.Wordsarepresentedprob-abilistically;anerroronafrequentwordsuchasTHEaf-fectsthenetworkmuchmorethanerrorgeneratedonpre-sentationofamuchlowerfrequencywordsuchasYULE.Minimizingthetotalerroristhereforebestaccomplishedbyprimarilyoptimizingthehighfrequencyitemsoverthelowfrequencyitems.Thus,whileallitemsarepressuredtobereadasquicklyaspossible,themorerapidorthsempathwayreceivesgreaterpressurefromthehighfrequencyitemsthanthelowerfrequencyones.
Totakeanexample,thefrequencyofpresentationofthewordTHEistwentytimesthatofBRIM,meaningthattheerrorduetoslownessinprocessingitemsistwentytimesgreaterforTHEthanBRIM.Thus,allotherthingsbeingequal,thenetworkresourcesallocatedtorapidlyprocessingTHEwillfaroutpacethoseallocatedtoprocessingBRIM.ThisbehaviorofthemodelstronglycontradictsSmith’s(1973)conjecturesabouttheefficiencyofdifferentdecod-100
80
tcerro60
C tencr40
eP20
Orth->SemOrth->Phon->Sem0
HighLowFrequency
Frequency
Figure19.Divisionoflaborinthecomputationofsemantics:effectsofwordfrequency.Thedataareforeachpathwayinisola-tion.Forhigherfrequencywords,theorthsempathwayismoreaccurate;forlowerfrequencywords,bothpathwaysareequallyaccurate.
ingstrategies.Smitharguedthatreadingisaccomplishedtoorapidlytoaccommodatephonologicalrecoding.How-ever,Zipf’slaw(Zipf,1935)statesthatthereisaconstantrelationshipbetweenthenumberofwordsatagivenfre-quencyrangeandthesquareofthatfrequencyrange;i.e.thefrequencyhistogramforanylanguagefollowsacurveykx2,forsomeconstantk.Onlythemosthighlyfre-quentitemstendtoviolatethisrelationship.Whatthismeansisthatthereisaverysmallnumberofwordsthatoccurveryfrequently,andaverylargenumberofwordsthataremuchmoreinfrequent.Evenifstrongrelianceonorthsemislimitedtothesehighestfrequencywords,theyaccountforalargeproportionofthetokensapersonreads.
Figure20showsthelatencies(calculatedasdescribedpreviously)fortheseitems,bypathandfrequency.AsinFigure16,theintactmodelisfasterthaneithertheorthsemororthphonsempathsalone.Thehighfrequencyitemsaremorerapidbytheorthsempath,whereasthelowfrequencyitemsarecomputedaboutequallyfastbyboth.Theinteractionoffrequencyandpathway(orthsemversusorthphonsem)wasreli-able(F19987347,p0001).Theintactmodelalsoshowedamaineffectoffrequencyinitslatencies(F1998252,p0001).Wematchedtheitemsusedinthistestwithitemsfromalargescalestudyofread-ingtimes(Seidenberg&Waters,1989).Therewere351highfrequencyitemsand122lowfrequencyitemspresentinbothlists;inthelatenciesreportedbySeidenbergandWaters(1989)theseitemsshowedastrongfrequencyef-fect(F1471
296,p0001).ThesubsetofitemsinMULTICOMPONENTMODELOFREADING
35
4.0
Orth->SemOrth->Phon->SemIntact Model3.0
cyent2.0
aL1.0
0.0
HighLowFrequencyFrequency
Figure20.Effectsoffrequencyonlatenciestocomputeseman-ticsbyindividualpathwaysandintheintactmodel.Latenciesareinunitsofwholetime.
bothlistsalsoshowastrongfrequencyeffectinthemodel(F14712101,p0001).
Theintactmodelalsoshowedamaineffectoffrequency
initslatencies(F1998
252,p0001).Matchingthesetestitemswiththoseusedinalargescalecollectionofreadingtimes(Seidenberg&Waters,1989)revealedthatsubjectsalsoexhibitareliableadvantageforthesehighfre-quencyitems(p0001).
SimulationFrequency11:andInteractionConsistency
ofTheaboveanalysisconsideredtheeffectsoffrequencyonthedivisionoflaborincomputingmeaning.Wenextex-aminedwhethertheseeffectsaremodulatedbyanotherlex-icalfactor,spelling-soundconsistency(Seidenberg&Mc-Clelland,1989).Consistencyaffectsthedifficultyofcom-putingphonologicalcodes,especiallyforlowerfrequencywords(Seidenbergetal.,1984;Taraban&McClelland,1987).Thisfactorshouldthereforeslowtheactivationofsemanticsviaorthphonsem,creatinggreaterdepen-denceonorthsem(seealsoStrainetal.,1995).
Method
StimuliandProcedure.Wordsinthetrainingsetwerecategorizedaccordingtotheirconsistency,whichasinpre-viousstudies(Jaredetal.,1990)wasdefinedintermsoforthographicrimes(e.g.,-INTinMINT).13Allitemswith
biggestimpactinourmodels(seeJaredetal.,1990).Statisticalregularitiesinvolvingotherpartsofwordscanalsoaffectperfor-mance,butnotasstrongly.
36
HARM,SEIDENBERG
100
80
Percent Correct60
40
20
0
Orth->SemOrth->Phon->SemHFHFLFLFConsistentInconsistentConsistentInconsistent
Figure21.Divisionoflaborinthecomputationofsemantics:Effectsoffrequencyandspelling-soundconsistencyineachpath-way.
ofpathway,frequencyandconsistencywasnotreliable.Therewasareliableinteractionoffrequencyandpathway(F189617306,p0001),andconsistencyandpath-way(F1896409,p0001).
Consideringtheintactmodel,therewasaninterac-tionbetweenfrequencyandconsistency(F189644,p0035);theeffectofconsistencywasnotreliableforhighfrequencyitems(0.756versus0.730;F144817,p0184),butwasforthelowfrequencyitems(1.08ver-926,p0002).Thisisparticu-sus0.96;F1448
larlyimportant,becausenumerousstudieshaveshownthatinstandardwordrecognitiontasks,consistencyeffectsarenotfoundforhighfrequencyitems.14Inspectionofthelesionedmodelssuggestswhythismaybethecase.Forthehighfrequencyitems,theorthphonsempathwayhasreliablylowerlatenciesforconsistentitemsthanin-consistent(2.22versus2.49;F1448176,p0001),whiletheorthsempathwayhasreliablylowerlatenciesforinconsistentitemsthanconsistent(1.47versus1.69;F1448133,p0001).Thisillustratestwoimpor-MULTICOMPONENTMODELOFREADING
37
ofthefullytrainedmodel:intact,severedorthphon,andseveredorthsem.Thecomputedsemanticandphonolog-icalcodeswererecordedandtheactivationoftheplural,pasttense,andthirdpersonsingularfeatureswereexam-ined.
Results
AsindicatedinTable2,theintactmodelproducedaplausibleinflectionfor82%ofthepluralinflecteditems,and100%ofthepasttenseitems.Interestingly,theorthsempathwayinisolationwasalmostasaccurateastheintactmodel,whereastheisolatedorthphonsempathwaywaslessaccurate.Thedataindicatethattheorthsempathwayencodedthefactthat-EDand-Sareassociatedwithparticularsemanticfeatures.Thus,therewaslearningofsublexicalquasiregularitieswithintheorthsemcomponent.
Themodelwasmoreaccurateindeterminingtheinflec-tionofpasttensenonwordsthanplurals.Thisisbecauseinthetrainingset,-EDattheendofthecodaisverystronglypredictiveofpasttense;noitemswithacodaendingin-EDarenotpasttense.However,manyitemsendin-Sthatarenotsemanticallypluralorthirdpersonsingular(BUS,PLUS,NEWS).Hence,-EDisamuchbettermorphologicalcuethan-S;thisisreflectedinthemodel’sperformance.
Table2
MorphologicalEffectsinReadingModelSimulation
Plural
3rdPerson
PastTense
DivisionofLabor:Summary
Wehavedescribedamodelinwhichdirect-visualandphonologically-mediatedpathwaysjointlydeterminethesemanticsofwords.Therelativecontributionsofthetwopathwaysareinfluencedbyfactorsincludingtheskilllevelofthemodel,andlexicalpropertiessuchasfrequencyandspelling-soundconsistency.Inthenexttwosectionsweex-aminethemodel’sperformanceinprocessinghomophones
andpseudohomophones,stimulithathaveplayedanimpor-tantroleintheorizingabouttheroleofphonologicalinfor-mationinreading.
7.HOMOPHONES
Asnotedearlier,spellingandphonologyarehighlycor-relatedinEnglishbecausetheorthographyisalphabetic;incontrast,thecorrespondencesbetweenspellingandmean-ingaremorearbitrary,althoughastheprevioussimulationshowed,theorthsempathwaycanlearnmorphologicalregularitiessuchasnumberandtensemorphology.Wehaveseenhowthesecharacteristicsofthemappingsaffectthedevelopmentoftheorthsemandorthphonsempathways.Homophonespresentanimportanttestcasebe-causeorthphonsemcomputationisambiguous;ROSEandROWSactivatethesamephonologicalcode,whichisas-sociatedwithtwodistinctmeanings.Inthissectionwefirstcharacterizehowhomophonesareprocessedinthemodelandthenpresentsimulationsofthreerepresentativebehav-ioralstudies(VanOrden,1987;Jared&Seidenberg,1991;Lesch&Pollatsek,1993).
Simulation13:Homophones
Thedivisionoflaborforhomophonesoverthecourseoftrainingwasexaminedusingthelesioningmethodol-ogy.Effectsoftherelativefrequenciesofthealternativesensesofthehomophones,afactorthatpreviousstudieshaveshownaffectsperformance(Rayner&Duffy,1986;Simpson,1994)werealsoassessed.
Method
StimuliandProcedure.Therewere497pairsofhomo-phonesinthetrainingset.Allhomophoneswhoseproba-bilityofpresentationwasatleast1.5timesgreaterthantheothermemberofitspairwerecategorizedasdominantandthealternativeassubordinate.Allotherpairswerecodedasbeingapproximatelybalancedinfrequency.Thisyielded324highfrequencyhomophones,324low,and346bal-ancedones.Thedivisionoflaboranalysisusedinprevioussimulationswasrepeatedusingtheseitems.Themethodofpresentingitemstothenetworkandlesioningpathwayswasidenticaltotheprecedingsimulations.
ResultsandDiscussion
Theresultsforallhomophones,collapsedacrossfre-quency,areshowninFigure23.Thedatareflectthefactthattheorthphonsempathwayhasalimitedcapac-itytoreadhomophonesbecauseoftheirinherentambigu-ity,whereastheorthsempathwayisonlylimitedbytheamountoftraining.Figure24presentstheeffectsofrela-tivefrequencyintheintactmodelandthetwoisolatedpath-ways.Atasymptotetheintactmodelisabletocomputethecorrectmeaningsforalmostallhomophonesregardlessof
38
HARM,SEIDENBERG
relativefrequency;thedominantitemsareacquiredfirst,withthebalancedandsubordinatehomophoneslearnedmoreslowly.Theorthsempathway(topright)learnsthedominanthomophonesmoreslowlythantheintactmodelbutasymptotesatnearlythesameaccuracylevel.Thispath-wayperformsmuchlesswellthantheintactmodelonbal-ancedandsubordinatehomophones.Theorthphonsempathway(bottomcenter)canreadsomeofthedominantho-mophones,fewerofthebalancedonesandalmostnoneofthesubordinates.Allofthehomophonesareinherentlyam-biguous,butthispathwaygetssomedominantandbalanceditemscorrectbecauseitdefaultstoonemeaningthatturnsouttobecorrect.
ThedatainFigure24showthattheintactmodelper-formsbetterthaneitheroftheisolatedpathwaysforalltypesofhomophonesthroughouttraining.Thisfindingpro-videsadditionalevidencethatthetwopathwaysjointlyde-terminemeaningintheintactmodel.Themostdirectev-idenceisprovidedbythebalancedandsubordinateitems,forwhichtheintactmodel’saccuracyisgreaterthanthesumoftheaccuraciesofthetwoindependentpathways.ThisresultisalsoseeninTable3,whichsummarizestheresultsattheendoftraining.Thedominantitemsdonotshowthiseffectbecausethemodeldoessowellonthem;theisolatedorthsemgetsmostofthemcorrect,andorthphonsemalsogetsaover30%ofthem.Thus,fordominant,higherfrequencyhomophones,bothpath-wayscontributebecausetheybecometunedtotheseitems(orthsemmoresothanorthphonsem),whereasforbalancedandsubordinatehomophonesbothpathwayscon-tributebecausetheyarejointlyneededtocomputeseman-ticsaccurately.
Table3
AsymptoticPerformanceOnHomophones:PercentCorrectModel
Dominant
Balanced
Subordinate
Note:O=Orthography,S=Semantics,P=Phonol-ogy.
Attheendoftraining,allhomophoneswerereadmoreaccuratelybythesemanticpaththanthephonologicalone.Infact,essentiallynoneofthehomophonescouldbereadonlybyorthphonsemandnotbyorthsem.Thisisnotsurprisinggiventhattheorthphonsempathwayisfundamentallyambiguousforhomophones.Interestingly,orthphonsempathwaywasalmosttotallyunabletoreadthesubordinatehomophones(e.g.EWESversusUSE);thebulkofthesubordinatehomophonescouldbereadei-therbytheorthsempathinisolation,orthetwopaths
IntactOnly By BothOrth->SemOrth->Phon->Sem100tc80errCo60 tenc40reP2000K
250K500K750K1000K1250K1500K
Iteration
Figure23.Divisionoflaborincomputingthesemanticsofhomophonesoverthecourseoftraining.Themodellearnstoproducethecorrectmeaningsusinginformationfrombothpath-ways.Mostcanbecomputedcorrectlyonlyusinginputfrombothpathways;asmallandnearlyfixedproportioncanbereadbyorthphonsemalone(thesearedominant,highfrequencymeanings).
together.Theorthphonsempathwaywasmuchmoresuccessfulinreadingthebalancedmembersandstillmoresuccessfulatreadingthedominantmembers.Thereasontheorthphonsempathwaywasmoresuccessfulatread-ingthedominantmemberofahomophonepairthanthesubordinatememberisinpartbecausethephonsempath-waywasbetteratreadingsuchitems.
Figure23alsorevealsaninterestingdevelopmentalef-fect.Theconditionlabeled“onlybyboth”consistsofitemswhichcouldnotbereadbyeitherpathwayinisolation,butcouldbereadbytheconjoinedeffortsofthetwopathways.Thisisofparticularlyinterest,inthattheorthsempath-waywasnotabletoreadtheseitemsbyitself,butcouldpro-videenoughinformationtodisambiguatethephonologicalformoftheword.RecallSimulation1,inwhichasmallamountofsemanticcontexthadadramaticeffectontheabilityofthenetworktodisambiguatehomophonouspho-nologicalpatterns.Thesharpinitialriseinthe“onlybyboth”conditionearlyintrainingshowsthattheorthsempathwaywasnotprovidingenoughinformationtoproducethecorrectsemanticsbyitself,butwasprovidingenoughtodisambiguatemanyhomophones.
Forallthreetypesofhomophones,thisconditionreachedapeakintheearlystagesoftrainingandthendroppedoffasthemodelcontinuedtodevelop.Theorthsempathwaybecamebetterabletoreadhomophonesinisolationastrainingprogressed.Thebroadimplicationofthissimulationisthattheextenttowhichhomophonesre-MULTICOMPONENTMODELOFREADING
39
100100
8080
Percent Correct60Percent CorrectDominantBalancedSubordinate250K500K750K1000K1250K1500K
60
4040
2020
00K00K
DominantBalancedSubordinate250K500K750K1000K1250K1500K
Iteration
100
Iteration
DominantBalancedSubordinate80
Percent Correct60
40
20
00K
250K500K750K1000K1250K1500K
Iteration
Figure24.Homophoneaccuracyoverthecourseoftraining:intactmodel(topleft),byorth(bottomcenter).
sem(topright),andbyorth
phon
sem
quireinputfromorthphonsem,orthsemorbothde-pendsontherelativedominanceofthehomophone,andtheoveralldegreeofreadingskill.
Thesemanticfeatured’wascomputedforthethreeclassesofhomophonesforthethreesimulationconditions(intact,byphonologyandbysemantics)forthefullytrainedmodel.Foreachitemtobepresented,thesemanticrepre-sentationwasrecordedandcomparedwiththetargetrep-resentation.Hits,misses,falsealarmsandcorrectrejec-tionswereusedtocomputethed’.Additionally,foreachhomophonepair,thed’forthegeneratedsemanticsandthetargetsfortheothermemberofthepairwasalsocom-puted.So,forexample,forthehomophonepairEWES/USE,EWESisasubordinatemember;whenitwaspresentedto
thenetworkthesemanticrepresentationitproducedwascomparedtothetargetsforEWES,andUSE.Thesetwod’valuesareshowninTable4.Interestingly,thereissomeinformationavailabletothesemanticsysteminallcondi-tions;thed’isneverzero.Thereliabilityandcomplete-nessofthisinformationiswhatvariesaccordingtopathwayandrelativefrequencyofthehomophone.Further,forthesubordinatehomophonesbeingreadbyorthphonsem,thed’fortheopposingmemberofthehomophonepairishigher.ThisindicatesthatthepresentationofEWESre-sultsinmoreUSE-likeinformationbeinggeneratedalongtheorthphonsempath.
40
HARM,SEIDENBERG
Table4
SemanticFeatured’forHomophones
Dominant
IntactOSOPS
Balanced7.35.42.0
2.01.62.0
Subordinate7.35.41.8
1.31.22.2
UNDEF6.42.21.21.31.8
MULTICOMPONENTMODELOFREADING
41
1.00
Mean Semantic ActivationSimilarDifferent0.30
SimilarDifferentMean False Positives0.80
0.25
0.20
0.60
0.15
0.40
0.10
0.20
0.05
0.00
ShortLong0.00
ShortLongDuration of StimuliDuration of Stimuli
Figure25.Experiment1and2fromVanOrden(1987)(left),andsimulationresults(right).DatafromVanOrden(1987)showthedifferenceinfalsepositivesbetweenfoilsandcontrols.Datafromthemodelshowsthedifferenceinsemanticactivationbetweenfoilsandcontrols.
Table5
StimuliUsedInExperiment14
SimilarlySpelledItemsbeachbeechbenchcreekcreakcheekteamteemtermseamseemslamreinrainranpeakpeekpeckmeatmeetmeltbowlbollboil†arcarkare†pollpolepaleLessSimilarItemsdoedoughdoubtnoseknowssnobssuitesweetsheetmaidmademaimnunnonenoonlutelootlostroserowsrobsweightwaitwrit†neighnaybay†hawkhockbock
[geologicalformation]
[brook][unit][joint]
[implement]
[indefinitequantity][foodstuff][vessel][container][analyze]
[animal][organ]
[musicalcomposition][lifeform][lifeform][material][rise]
[physicalproperty][horse]
[haspartwing]
15IntheInteractiveActivationmodelofMcClellandandRumel-
hart(1981),unitscorrespondingtosegmentsoflettersactivatedlocalistletterrepresentations,whichinturnactivatedwordrepre-sentations.Theweightshadbeenchosensothatwhenallsegmentsofaletterpositionwereactivatedtheletternodesweresuppressed.Wehavenotimplementedalettersegmentrepresentation,butas-sume,followingMcClellandandRumelhart(1981),thattheeffectofapatternmaskistoobliterateactivityinletterrepresentations.Thereisagrainissueinsofarasthemodelsometimesmakesmorespecificpredictionsthancanbeobservedinbehavioralstudies(seeSimulation16below,forexample).
42
HARM,SEIDENBERG
50.0Semantic ErrorPhono ErrorSum Squared Error40.030.020.010.00.0012345678
Time
Figure26.Sumsquarederrorofsemanticandphonologicalrep-resentationswhenorthographicinputismaskedattime2.0.
ResultsandDiscussion
Figure25(right)showstheresults.16Intheshortpresentationcondition,therewasnoreliableeffectofvi-sualsimilarity.Inthelongpresentationcondition,there
508,wasareliableeffectofvisualsimilarity(F118
p0037).Thus,theVanOrden(1987)resultsappearinamodelthatincorporatesverydifferentmechanismsconcerningtheactivationofmeaning.Thepresentmodelhasnoexplicitspellingcheckmechanism;rather,thecor-rectmeaningsofhomophonesarecomputedonthebasisofinputfrombothorthsemandorthphonsempath-ways.17
Toseewhytheseresultsobtain,considerthedatainFig-ure26.Thedataarethesumsquarederrorforboththese-manticandphonologicalrepresentationsmeasuredateach
MULTICOMPONENTMODELOFREADING
43
tosemanticactivationandhomophonedisambiguation.Theimplicationofthesefindingsconcerningtheinterpretationofmaskingexperimentsshouldbeclear:itcannotbeas-sumedthatwhatoccursinthemaskedconditionalsooc-curswhentheinputisnotmasked.Thus,theapparentpri-macyoforthphonsemobservedintheseexperimentsisinpartduetotheuseofanexperimentaltechniquethatdifferentiallydisruptsprocessingwithinorthsemvs.orthphonsem.Wereturntothisissuebelowinconnec-tionwithsimulationsofanotherstudyusingthemaskingprocedure.
SimulationSeidenberg15:Jaredand(Homophones)
(1991)WenowturntothestudybyJaredandSeidenberg(1991)thatprovidedevidenceconcerningtheeffectsofho-mophonefrequencyonfalsepositives.AsinVanOrden(1987),subjectsperformedasemanticdecisiontask(e.g.,isitanobject?)andtargetitemswereeitherexemplars(MEAT),ahomophonousfoil(MEET),oraspellingcontrol(MEAN).Wordswerenotmaskedbutratherpresenteduntilthesubjectresponded.Thehomophonefoilsvariedintermsoftheirfrequencies(highvs.low)andthefrequenciesofthematchedexemplar(high,low)inafactorialdesign.Theprincipaldataconcernthenumberoffalsepositivesineachfoilconditioncomparedtothoseonspellingcontrols.
25
LF ExemplarrHF Exemplarro20
rE tne15
creP n10
i ecne5
reffiD0
-5
HF FoilLF FoilFigure27.TheJaredandSeidenberg(1991)homophoneresults.Falsepositivesoccuronlywhenatargetisalowfrequencyfoil,andtherelevantexemplarisalsolowinfrequency.
Figure27showstheneteffects(percentfalsepositivesinafoilconditionminusthespellingcontrolcondition).Theonlyconditioninwhichpresentationofthefoilyieldedasignificantnumberoffalsepositiveswastheoneinwhich
boththehomophonefoilanditscorrespondinghomophoneexemplararelowinfrequency.Highfrequencyfoilsandlowfrequencyfoilswithhighfrequencyexemplarsdidnotyieldstatisticallyreliablefalsepositiveeffects.
Theseresultsareabitpuzzling.Itiseasytoseefromthesimulationspresentedpreviouslywhylowfrequencyfoils,butnothighfrequencyones,wouldproducefalsepos-itives.Highfrequencyitemsaremorelikelytobenefitfromthedirectorthsemroutethanlowfrequencyones;theorthsemrouteisnot“fooled”byhomophonythewaytheorthphonsemrouteis.Asmoreorthographicinforma-tionisavailabletothesemanticsystem,theprobabilityofafalsepositiveforahomophonedecreases.Whatispuzzlingistheeffectoftheexemplarfrequencyonthetendencyofhomophonefoilstoproducefalsepositives.WhyshouldthefrequencyofMEATmodulatetheprobabilityofafalsepos-itiveforMEET?JaredandSeidenberg(1991)werenotabletoprovideadefinitiveanswer,insteademphasizingthelackofafalsepositiveeffectforhighfrequencywordswhichseemedtocontradictthestrongpositionthatorthsemdoesnotinfluencetheinitialcomputationofmeaning.Ifthefalsepositiveeffectistakenasevidenceforphonologically-activatedaccessofmeaning,thentheabsenceoftheeffectinsomeconditionsimpliedthatmeaningwasnotaccessedviaphonology.WeconductedareplicationoftheJaredandSeidenberg(1991)studyusingthemodelwiththegoalofclarifyingtheseeffects.
Method
Stimuli.Stimuliwereselectedasfollows.Allitemsinthetrainingsetweredividedintothecategoriesofobject,livingthing,orother,basedonthepresenceorabsenceofthesemanticfeatures[object]and[life_form].Itemsthatareobjectsorlivingthingswerecandidatesforexemplars.Itemsthatarenotobjectswerecandidatestobeafoilorspellingcontrolforobjectexemplars.Thosewhicharenotlivingthingswerecandidatestobefoilsorspellingcontrolsforlivingthingexemplars.
Foreachcandidateexemplar,wedeterminediftheitemhadacorrespondinghomophonefoil.Tocreatespellingcontrolsweidentifiedanitemwiththesamenumberoflet-tersastheexemplar,thesameinitialletter,andaspellingthatdifferedbyatmostoneletterfromtheexemplar.Allfoilsandexemplarswithaprobabilityofpresentation021werecodedashighfrequency;thosewithaprobability
005werecodedaslow.Table6showsasamplesetofitems;atotalof397foilsandmatchedcontrolsresulted.
Procedure.JaredandSeidenberg’sprocedurewassim-ulatedbypresentingthefoilsandspellingcontrolstotheintactmodel,andobservingtheactivationontheseman-ticfeaturefortheexemplar.Forexample,ifCAUGHTwaspresentedtothemodel,the[object]featurewouldbemoni-tored,becausetheexemplar(COT)isanobject.Activityfortheinappropriatesemanticfeatureforthefoilwasrecorded,
44
HARM,SEIDENBERG
Table6
SamplestimuliforJaredandSeidenberg(1991)replication
alescotroadsonLFLFHFHFailscaughtrodesunLFHFLFHFaidstaughtbodebun
MULTICOMPONENTMODELOFREADING
45
420420
Feature Input-4-6-8-10-12-14-160.0
1.0
2.0
3.0
4.0
Feature Input-2-2-4-6-8-10-12-14-16
0.0
Orth->SemPho->SemOrth->SemPho->Sem1.0
2.0
3.0
4.0
Time
(a) HF Foil, HF Exemplar
420420
Time
(b) LF Foil, HF Exemplar
Feature Input-4-6-8-10-12-14-160.0
1.0
2.0
Feature InputOrth->SemPho->Sem3.0
4.0
-2-2-4-6-8-10-12-14-16
0.0
1.0
2.0
Orth->SemPho->Sem3.0
4.0
Time
(c) HF Foil, LF Exemplar
Time
(d) LF Foil, LF Exemplar
Figure29.Inputtodistractorsemanticfeatureforfourfoilconditions:(a)Highfrequencyfoil,highfrequencyexemplar.(b)Lowfrequencyfoil,highfrequencyexemplar.(c)Highfrequencyfoil,lowfrequencyexemplar.(d)Lowfrequencyfoil,lowfrequencyexemplar.
semanticfeatures.
2.Lowfrequencyfoil,highfrequencyexemplar(Fig-ure29b).Inthiscondition,theorthphonsempathwaywasalsoactivatingtheinappropriatesemanticfeature;morestronglythaninFigure29a.ThisisconsistentwiththedatafromTable4,inwhichtheorthphonsempathwayproducedthesemanticsofadominanthomophone(theex-emplar,inthiscase)muchmoresothanthesubordinatehomophone(here,thefoil).Asabove,whenthefoilwaspresented,theorthsempathwayhadtoextinguishthisin-appropriateactivation,andhenceastrongnegativeinputtotheinappropriatesemanticfeaturedeveloped.Thisresultedinnoreliablefalsepositivesinthiscondition.
3.Highfrequencyfoil,lowfrequencyexemplar(Fig-ure29c).Here,bothpathwayswereinhibitingtheinap-propriatesemanticfeature.Theorthphonsempathwaydidsobecausethesemanticsofthedominanthomophone(here,thefoil)wereactivated,andthesemanticsofthesub-ordinatehomophonesuppressed.Thus,therewasverylit-tleerrorproducedbyorthphonsemfortheorthsempathwaytocorrect.However,thefoilwashighinfre-quency,andconsistentontheresultsofSimulation10,theorthsemdevelopedtheabilitytoquicklyrecognizetheitemandhencesuppressinappropriatesemanticinforma-tion.
4.Lowfrequencyfoil,lowfrequencyexemplar(Fig-46
HARM,SEIDENBERG
ure29d).Thiswastheconditionthatproducedreliablefalsepositives,bothintheempiricalstudybyJaredandSeidenberg(1991),andinthissimulation.Here,theho-mophonesarebalanced,andlowinfrequency,thereforetheorthphonsempathwayproducesratherambivalentac-tivationoftheexemplar’ssemanticfeature,particularlyattheendofprocessing.18Whenthefoilwasprocessedbytheorthsempathway,itdidnothavetosuppressstrongerroneousresponsesgeneratedbyorthphonsem,asincaseswheretheexemplarishighinfrequency.Thefoilwasfurtherwaslowinfrequencyitself,andhencetheabilityoftheorthsempathwaytoprocessitwaslimitedrelativetohighfrequencyfoils.Hence,spuriousfalsepositivesre-sulted.
TheresultsofthissimulationprovideareconciliationoftheviewsofVanOrdenandcolleaguesandJaredandSeidenberg(1991).ConsistentwithVanOrdenetal.’sin-terpretation(andcontrarytoJared&Seidenberg,1991),theorthphonsempathwayproducessomesemanticactiva-tionforhighfrequencyhomophones.However,consistentwithJaredandSeidenberg(andcontrarytoVanOrdenetal.),highfrequencyfoilssuppressinappropriateactivationoftheirpairedhomophoneviatheorthsemrouteinpar-allelwiththeprocessingoftheorthphonsempathway,ratherthanasaresultofapost-lexicalspellingcheckop-eration.ThisnovelaccountoftheJaredandSeidenberg(1991)studyarisesfromcorecomputationalprinciplesofthemodel:cooperativecomputationtoreduceerror,andthepressureforthemodeltorespondrapidly.
Simulation16:Leschand
Pollatsek(1993)
Importantadditionalevidenceconcerningtheroleofphonologyinwordreadinghasbeenobtainedfromstud-iesadifferentmethodology,semanticpriming.LeschandPollatsek(1993)createdtripletsofwordsconsistingofanexemplarsuchasTOAD,ahomophonesuchasTOWED,andatargetthatissemanticallyrelatedtotheexemplarsuchasFROG.Subjectswerepresentedwithaprimethatwasei-thertheexemplar,thehomophone,oranunrelatedcontrol,andthenthetarget,whichwasnamedaloud,withnaminglatencythedependentmeasure.Thestudyemployedtwopresentationconditions:short(primepresentedfor50msthenpatternmaskedfor200ms)andlong(primepresentedfor200msthenmaskedfor50ms).ThecriticalquestionwaswhetherhomophonessuchasTOWEDwouldprimetar-getssuchasFROG.ThedataaresummarizedinFigure30.Intheshortcondition,bothexemplars(suchasTOAD)andhomophones(suchasTOWED)yieldingsignificantpriming
MULTICOMPONENTMODELOFREADING
47
Method
Stimuli.Alistofhomophonicwordpairswascreatedalgorithmicallybyscanningthetrainingcorpusforwordswithdifferentspellingsbutidenticalphonologicalrepresen-tations.Asecondlistofsemanticassociateswascreatedbyscanningthesemanticrepresentationsofalluninflectedwordsandfindingallpairswherethesemanticrepresen-tationsdifferedbynomorethanonefeature.Fromthesetwolistswefoundasetoftripletsconsistingofanexem-plar,ahomophoneandatargetsemantically-relatedtotheexemplar.Acontrolitemwasselectedforeachtripletthatdifferedfromthehomophonebyatmosttwoletters.Boththehomophoneandthecontrolitemhadtodifferfromthetargetbyatleast8semanticfeatures.Afurtherconstraintwasimposedsuchthatforapproximatelyhalfoftheitems,theexemplarhadtobehigherinprobabilityofpresenta-tiontothemodelthanthebalancedhomophonebyatleastafactoroftwo;fortheotherhalf,thehomophonehadtodominatebyatleastafactoroftwo.Thehomophonesinbothsetswerematchedontheiroverallmeansemanticdif-ferencefromthetarget.Asetof53quadruplesresulted,consistingofanexemplar,homophone,controlandtarget(e.g.,CREEK,CREAK,BLEAK,STREAM).Therewere28biasingtheexemplarand25biasingthehomophone.Procedure.Tosimulatetheshortprimingcondition,primeswerepresentedfor2unitsoftime.Thenthemodelwasallowedtocontinueprocessingforanadditional5unitsoftime.Inthelongcondition,theprimewaspresentedfor5unitsoftimeandthemodelcontinuedprocessingforanadditional2unitsoftime.Overthecourseofprocess-ing,thesemanticandphonologicalerrorforallprimeswasrecorded,aswellastheirsemanticdistancefromthetargetitem.Attheendof7unitsoftimethestateoftheseman-ticunitswasalsorecorded.Weassumedthattheamountofprimingwouldbeafunctiontheamountofsemanticoverlapbetweentheprimeandtarget,asshowninpreviousstudiesbyMcRaeetal.(1997)andPlautandBooth(2000).
Results
Figure31showsthesemanticdistanceat7unitsoftimeasafunctionofprimetypeandduration.TheresultsreplicateLeschandPollatsek’sfindingthatboththeexem-plarandhomophoneproducedprimingattheshortdura-tioncomparedtoanunrelatedcontrol;inthelongdurationconditiontherewasstrongprimingfortheonlyfortheex-emplar.Thispatternisreflectedinasignificantinterac-tionbetweenprimetypeandduration(F001).Thereisasmallresidualpriming231210effectforthe9,p0homophoneinthelongdurationconditioninthesimulation,aneffectsizethatwouldbedifficulttodetectinabehavioralexperiment.
30
tExemplareHomophonergaT25
Control mro20
F ecan15
tsiD c10
itanm5
eS0
ShortLongFigure31.SimulationofLeschandPollatsek(1993).AsinFigure30,bothExemplarandHomophoneprimesproducedsig-nificantprimingattheshortdelay,whereasatthelongerdelaytheexemplarproduceslargerfacilitationthanthehomophone.
30Homophone Phono ErrorHomophone Sem Error24Exemplar Phono ErrorroExemplar Sem ErrorrrE d18erauqS 12mSu600.0
1.02.03.04.05.06.07.0
Time
Figure32.Semantic(Sem)andphonological(Phono)errorforhomophonesandexemplarsintheshortprimecondition.
48HARM,SEIDENBERG
Discussion
Tounderstandwhytheseeffectsobtain,considerthedatainFigure32,whicharethesumsquarederrorofthemodel’sphonologicalandsemanticrepresentationoverthecourseofprimepresentationfortheshortcondition.AswasshowninthesimulationoftheVanOrden(1987)data,phonologyismuchmoreresilienttotheeffectofthemaskthansemantics.Whenthevisualinputismasked,phonol-ogytendstoremainatalowerror,whilethesemanticrep-resentationdriftsfromthatassociatedwiththeorthographicform.
Figure33showstheaveragedistancefromthemodel’ssemanticrepresentationfromthetargetovertime,forallthreeprimetypes.Fortheshortprimecondition,therep-resentationoftheexemplarandthehomophonebegintoconverge.Atthepointatwhichthetargetwouldbepre-sented,botharemuchcloser,insemanticspace,tothetargetthanthecontrols.Forthelongcondition,thevisualstimulidrivestheexemplarclosetothetargetandthehomophoneandcontrolaway(andtowardstheirownsemanticrepre-sentation).Whenthevisualstimulusisremoved,theho-mophone(butnotthecontrol)beginstobeinfluencedbythephonologicalbutnotvisualinformation,anditdriftstowardsthatofthetarget.However,theISIisshorterinthelongcondition,andthusitdoesnothaveasmuchtimetomovenearertothetarget.ThusthemaineffectfoundbyLeschandPollatsek(1993)occurs:homophonesprimemuchmoreeffectivelyatshortpresentationdurationsandlongISIthanthereverse.Thiseffectinthemodelisnotduetoaninitialactivationofphonologyandasubsequentspellingcheck,butratherreflectsthedifferentialeffectofthemaskonsemanticandphonologicalinformation.
RelativeHomophones
Frequenciesof
Asdescribedearlier,thecomputationofmeaningforhomophonesalongthephonsempathwayissensitivetotherelativefrequenciesofthehomophones.Suchresultsareconsistentwithotherstudiesmanipulatingthisfactor.Thephonsempathwaywillactivatethesemanticsofadom-inanthomophonemoststrongly,asubordinateonemostweakly,andabalancedonetoanintermediatedegree.Themodelthereforepredictsaneffectofdominanceonthede-greeofhomophoneprimingwithshortstimuluspresenta-tion.Forexample,ifanexemplar/homophone/targettripleconsistedofastronglydominantexemplar(e.g.,USE/EWES->MAKE),wewouldexpecttheauditoryformofthesubor-dinatehomophoneEWEStostronglyactivatesemanticsforMAKE/USE,andhenceconsiderableprimingwouldoccur.Similarly,foratriplewherethehomophonewasstronglydominant(e.g.,EWES/USE->SHEEP)wewouldnotex-pectthehomophoneUSEtoactivateSHEEPsemanticsverystrongly,andhencemuchless(thoughperhapsmorethanzero)primingshouldoccur.Wereanalyzedthesimulation
Table7
PrimingEffectsinMsec,fromLukatelaandTurvey1994),Experiments5and6Short:RelativetoTOLD
Unsupportive115
Supportive6
4
Long:RelativetoTOLD
Unsupportive136
Supportive13
3
outputbygroupingthestimuliintothetwosets:supportivetrials,wheretheexemplaristhedominantmemberofthehomophonepair(andthustheauditoryformofboththeex-emplarandhomophonesupportthemeaningofthetarget,e.g.USE/EWESprimingMAKE),andunsupportive,wherethehomophoneisthedominantmemberandthustheaudi-toryformoftheexemplarandhomophonedonotsupportthemeaningofthetarget(e.g.,EWES/USEprimingSHEEP).
Figure34showstheresultsfromFigure33,brokendownbysupportiveness.Whenthestimulusispresent,thesemanticrepresentationforthesupportivehomophonemovestowardsthatofthetargetmorerapidlythantheun-supportiveone.Similarly,whenthestimulusisremoved,thesemanticsfortheunsupportiveexemplarmovesawayfromthetargetmorerapidlythanthesupportivecase.Cru-cially,eventheunsupportivehomophonesareclosertothetargetthanthematchedcontrolswhenthevisualinputisremoved,eventhoughtheyareequidistantwhenthevisualpatternispresent.
Aspredicted,therewasareliableeffectofsupportive-nessonsemanticdistancefromthetargetattime7intheshortprimecondition(F115352,p003)andthelongprimecondition(F1153179,p0001).
Althoughtheseresultsareconsistentwithotherstudiesmanipulatingtherelativefrequenciesofhomophones,therearetwoprominentfailurestoobserveeffectsofrelativefre-quencyinthehomophoneprocessingliterature:LeschandPollatsek(1993)andLukatelaandTurvey(1994b).
LeschandPollatsek(1993)didnotmanipulatetherel-ativefrequenciesofhomophonepairs,butreportedapost-hoctest.Thestimuliweredividedintotwosublists:oneinwhichtheexemplarwashigherinfrequencythanitspairedhomophone(whatweterma“supportive”condition),andoneinwhichthehomophonewashigherinfrequency(“un-supportive”).Theydidnotfindareliableeffectofsublist(“supportiveness”)onprimingeffects.Theythereforecon-MULTICOMPONENTMODELOFREADING
49
30Semantic Distance From Target24Semantic Distance From TargetExemplarHomophoneControl30
24
ExemplarHomophoneControl1818
1212
66
00.0
1.02.03.04.05.06.07.0
00.0
1.02.03.04.05.06.07.0
TimeTime
Figure33.Semanticdistancefromtargetitem,asafunctionofprimeconditionandprimeduration.Forbriefprimes(inputmaskedattime2.0,left),homophonesbecomedrawntowardthesemanticrepresentationoftheexemplar.AtlongerdurationsandshorterISI(maskattime6.0,right),thereislesstimeforthesemanticrepresentationofthehomophonetobecomeinfluencedbythesoundpattern.
30Semantic Distance From Target24Semantic Distance From Target18Supportive ExemplarSupportive HomophoneSupportive ControlUnsupportive ExemplarUnsupportive HomophoneUnsupportive Control30
24
18
Supportive ExemplarSupportive HomophoneSupportive ControlUnsupportive ExemplarUnsupportive HomophoneUnsupportive Control1212
66
00.0
1.02.03.04.05.06.07.0
00.0
1.02.03.04.05.06.07.0
TimeTime
Figure34.Closenessinsemanticspacetothetargetforprimetypesoverthecourseofprocessing,bysupportiveness.Shortprimedurationisshownontheleft,longprimedurationontheright.
50
HARM,SEIDENBERG
cludedthatbothhighandlowfrequencyhomophonesareprocessedviaphonology,incontrasttotheJaredandSei-denberg(1991)results.Thedifferingresultsappeartoberelatedtodifferencesinthesizeofthefrequencymanipula-tionsinthetwostudies.
InspectionoftheindividualitemsfromLeschandPol-latsek(1993)indicatedthatthemediandifferencebetweentheirhighandlowfrequencymatcheditemswas29.Therewere8paireditemsoutof32forwhichthefrequencydif-ferencewas10.ThesenumbersshouldbeconsideredinlightoftheknowninsensitivityoftheKuçeraandFrancis(1967)normsatthelowerendofthefrequencydistribution(Gernsbacher,1984).
Incontrast,theJaredandSeidenberg(1991)highandlowfrequencypaireditemshadamedianfrequencydiffer-enceof50,andnoitemshadadifference
10.Thus,thedifferencebetweenconditionswaslargerandmoreconsis-tentacrossitems.Itisnotsurprisingthatthefrequencyma-nipulationwasstrongerintheJaredandSeidenberg(1991)study;itwasbuiltintothedesignofthestudyratherthantestedposthoc.Inshort,theLeschandPollatsekmaterialsexhibitedsmaller,lessconsistentdifferencesbetweenhighandlowfrequencyitemsandthehigherfrequencywordswererelativelylowinfrequency,therangeinwhichtheKuceraandFrancisnormsarelessreliable.ThefailuretoobtainafrequencyeffectinthisstudycomparedtoJaredandSeidenberg’sseemslikelytoberelatedtotheseproper-tiesofthestimuli.
LukatelaandTurvey(1994b)presentedseveraladdi-tionalstudiesusingtheprimingmethodology.LikeLeschandPollatsek(1993),theyfoundprimingbyexemplarandhomophoneatshortdurations,butnotatlong.Theyalsomanipulatedwhatwehavetermedsupportiveness,andfoundnoeffect,leadingtotheconclusionthataccesstomeaningisinitiallyphonologicalregardlessofhomophonefrequency.Therearethreeproblemswiththesestudiesthatcloudtheinterpretationoftheresults.First,LukatelaandTurvey(1994b)thedataaresomewhatambiguous,becausethepatternofresultsdiffersdependingonwhichoftwocon-trolconditions19isusedtoassessthemagnitudeofprimingeffects.Second,thereisaproblemwiththestimuliinthe
visualcontrols(e.g.,TOLD)areusedasthebaselineforcalculatingprimingeffects;inthiscasealleffectsarewithin3-6msec.20We
calculatedthemeanbigramfrequenciesforthecontrols
(derivedfromanon-lineversionoftheAmericanHeritageDic-tionary).Thesummedbigramfrequenciesoftheexemplar,ho-mophoneandspellingcontrolprimeswerehigherthanfortheunrelatedcontrols(33,368versus24838,FThesedifferencesappearedwhenconsidering150219just4,p0001).ex-emplarsandtheircontrols(33693versus25549summedfre-quency,F116661,p0014),thehomophones(34855ver-sus25253,F116667,p001),andthespellingcontrols
(31556versus23713,F1166
65,p0012).Simplyput,PLASMisnotaverygoodcontrolforTOWED.
MULTICOMPONENTMODELOFREADING
51
capacitytocontributesignificantlytotheprocessingofho-mophonesandotherwords.Thesimulationsofbehavioralstudiesareconsistentwiththeconclusionthepeoplepro-cessthesewordsinasimilarmanner.
Theothermajorfindingfromthesimulationscon-cernedtheeffectsofmaskingonthecourseoflexicalprocessing.Thesimulationssuggestthatmaskinghasdifferenteffectsonprocessingwithintheorthsemandorthphonsempathways;itmainlyeliminatesnormalinputfromorthsem.Underthemaskedcondition,sub-jectscanonlyprocesshomophonesviaorthphonsem,yieldingasignificantnumberoffalsepositiveresponsesonthesemanticdecisiontask.However,itdoesnotfollowfromthisdemonstrationthatorthsemalsomakesnocon-tributionwithnormalstimuluspresentation.
Finally,thesimulationoftheJaredandSeidenberg(1991)studyprovidedfurtherevidenceconcerningthede-pendenciesbetweenthetwomainpathwaysingeneratingbehavior.Thesimulationcapturedthemainfeaturesofthehumandata,buttheprocessesthatgaverisetotheseeffectsweredifferentthaneitherVanOrdenetal.orJaredandSeidenberghadsurmised.Wefoundtheseresultssurpris-ingbutalsosoberinginsofarastheysuggestthatbehavioraldatacanbeconsistentwithunanticipatedunderlyingmech-anismsthatareonlyrecognizedbyusingacomputationalmodel.
8.PSEUDOHOMOPHONES
Thefinalphenomenatobeaddressedconcernthepro-cessingofpseudohomophonessuchasSUTE.Thesestim-ulihavebeenwidelystudiedbecauseoftheleveragetheyprovidewithrespecttodiagnosingtheuseofphonologicalinformationinreading.Pseudohomophonesarenovelstim-ulithathappentosoundlikeactualwords.Asubjectwillnothaveencounteredsuchstimulibefore;hencetheywillnothaveformedanyassociationsbetweentheirspellingsandspecificmeanings.Afalsepositiveresponseonatrialsuchas“Isitanarticleofclothing:SUTE”wouldresultifthesubjectphonologicallyrecodedthestimuluswhichactivatedthemeaningassociatedwithSUIT.Thefactthatthesubjectdoesnotknowinadvancewhetherthetargetisawordorpseudohomophoneimpliesthatphonologicalrecodingoccursinreadingwordsaswell.
Ourmodelisconsistentwiththeobservationthatpseu-dohomophonescanactivatesemanticsviaphonology;ingeneral,anorthographicpatternsuchasSUTEactivatesaphonologicalcodethatisverysimilartothatproducedbySUIT,whichinturnactivatesSUIT-semantics,providingthebasisforafalsepositive.However,themodelprovidesad-ditionalinformationthatraisesquestionsaboutpseudoho-mophoneprocessinganditsrelationtonormalreading.Thestandardviewthatfalsepositivesforpseudohomophonesareduetophonologically-mediatedactivationofsemanticsassumesthattheycannotactivatemeaningdirectlyfrom
orthography.Thisassumptionisworthexaminingmoreclosely.Somepseudohomophonesoverlapconsiderablywiththewordsfromwhichtheyarederived,forexampleBOXXorGHOAST.Aswehaveseen,inthemodelmanyfamiliarwordsactivatesemanticinformationdirectlyfromorthography.Althoughsubjectswillnothavelearnedtoas-sociateameaningwithanovelpatternsuchasBOXX,itmayoverlapsufficientlywithBOXtoproducesignificantseman-ticactivation.Ifthisiscorrect,falsepositivesforsuchstim-uliwouldnotnecessarilyimplicatephonologicalrecoding.Simulation17addressesthispossibility.
Arelatedissueconcernshowsubjectscorrectlyrejectpseudohomophonesonmosttrials.IthasbeenassumedthatdecidingthatSUTEisnotanarticleofclothingrequiresaspellingcheck–assessingthesemanticpatterncomputedviaorthphonsemagainsttheinputorthographicpat-tern(VanOrdenetal.,1988).Thisprocesswasalsoas-sumedtoapplytohomophonessuchasBEAR.Aswehaveseen,homophonescanbedisambiguatedviaorthseminthemodel,suggestingthatthespellingcheckisnotre-quired.Somepseudohomophones(e.g.,oneslikeBOXX)mayalsoactivatesemanticsfromorthography,butunlikehomophones,thiswouldonlyincreasethelikelihoodofafalsepositiveresponse.Onepossibilityisthat,unlikeho-mophones,pseudohomophonesdorequireaspellingcheck.Theremaybeotherbasesformakingthisdecision,how-ever.Forexample,pseudohomophonescoulddiffersystem-aticallyfromwordsintermsofthequalityofthephonolog-icalorsemanticcodestheyactivate.Likelexicaldecision(decidingifastimulusisawordornot),semanticdecision(decidingifastimulusisamemberofadesignatedcate-gory)isajudgmenttaskinwhichsubjectsmustestablishreliablecriteriaformakingaccurateresponses(Balota&Chumbley,1984;Seidenberg,Waters,Sanders,&Langer,1984).Simulation18examinedhowpseudohomophonesareprocessedinthemodelinordertoaddressthesepossi-bilities.
SimulationPseudohomophones17:ReadingofOrthby
Sem
Thissimulationaddressedwhetherpseudohomophonesactivatesemanticsviatheorthsempathway.Thedistri-butionofwordsinthespaceofpossibleorthographicpat-ternsisnonrandom:forexample,therearedenseclustersofwords(e.g.,onescontaining-AT)andtherearewordsthathavenocloseneighbors(so-calledstrangeorhermitwordssuchasYACHT),aswellasintermediatecases.The“receptivefields”ofunitsintheorthsempathwaywillvaryinresponsetothesedistributionalfacts.WordssuchasCAThavesomanycloseneighborsthattheweightsmustbenarrowlytunedtothatparticularwordorerrorswillre-sult.Incontrast,awordlikeGHOSThasfewneighbors,andsothenetworkcanhaveabroaderattractorforthatwordwithoutgeneratingerrors.Thisanalysispredictsthattwo
52
HARM,SEIDENBERG
Table8
SampleitemsusedinSimulation17Neighborhood
High
Low
Note.Pseu=pseudohomophone.Firstmemberineachpairisthepseudohomophone;secondmem-beristhecorrespondingword.
factorsshouldjointlyinfluencewhattheorthsempath-wayactivatesforapseudohomophone:thesimilarityofthepseudohomophonetothebaseword,andtheneighborhoodsofthebasewordandpseudohomophone.Forexample,thepseudohomophoneKATisunlikelytoproducesemanticac-tivationforCATviaorthsembecauseboththewordandpseudohomophonearefromverydenseorthographicneigh-borhoods;iftheunitsthatdetectCATwereinsensitivetothefirstletter,forexample,theywoulddrawfalsepositivesfromHAT,RAT,MATandsoon.PseudohomophonessuchasGHOAST,however,mayactivateGHOST-likesemantics;GHOSThasfewneighborsandsothecorrectsemanticsmaybeactivatedevenwithpartialinformationabouttheinput.Ineffect,thereceptivefieldforGHOSTmayincludeapseu-dohomophonesuchasGHOAST,whereasthereceptivefieldforCATdoesnotincludeKAT.Thepredictionthenisthattheabilityoftheorthsempathwaytoactivatesemanticsforpseudohomophoneswillbejointlydeterminedbyneigh-borhooddensityandclosenesstothebaseword.
Method
StimuliandProcedure.Asetofword-pseudohomophonepairswasgeneratedbyalgorithmicallyidentifyingonsetsandrimesthathavemultiplepossiblespellings,andthencreatingpseudohomophonesthathavethesamepronunciationasacorrespondingword.Theseitemsweresplitalongthreedimensions:visualsimilarityofthepseudohomophoneandcorrespondingword,wordneighborhooddensityandpseudohomophoneneighborhooddensity.Wordswereconsideredvisuallysimilartotheirpseudohomophoneiftheydifferedbyoneletter,anddissimilarotherwise.NeighborhooddensitywasassessedusingtheColtheartN(Coltheartetal.,1977)measure(whichequalsthenumberofwordsthatcanbederivedfromaletterstringbychangingoneletteratatime).DenseneighborhoodsweredefinedasN10,andsparseasN1.889pairsweregenerated.Table8showsasampleoftypicalitemsintheeightconditions,withtheirpairedhomophonousword.
Theorthphonpathwayinthetrainednetworkwasdis-connectedinordertoexaminethecapacityoftheorthsempathwaytoactivatesemantics.Pseudohomophoneswerepresentedtothenetworkinthestandardway,andforeachtrialtheresultingsemanticfeatureswererecordedandcom-paredtothetargetsforthepairedword.Forexample,forthepseudohomophoneTOSE,thesemantictargetsforthehomophonouswordTOESwerecomparedwiththeseman-ticoutputfortheinputpseudohomophoneTOSE.Asbefore,weconsideredasemanticfeaturetobeonifitsvaluewasabove0.5,andoffotherwise.Thed’wasthencomputedbasedonthethehits,misses,falsealarmsandcorrectrejec-tionswithrespecttotheactivatedsemanticfeaturescom-paredwiththeveridicalsemanticrepresentationofthetar-getword.Forexample,ifthewordTOEScontainedthesemanticfeatures[digit,extremity,body-part,foot,entity],andthepseudohomophoneTOSEactivated[digit,extrem-ity,entity]andalso[animal],thentherewouldbe3hits,2misses,onefalsealarm(from[animal]),andcorrectrejec-tionsforallothersemanticfeatures.
ResultsandDiscussion
Figure35showstheresults.Intheanalysisofvariance,thereweremaineffectsofvisualsimilarityofthepseudoho-mophonetothebaseword(F1887568870,p0001),wordneighborhooddensity(F162,p0013)andpseudohomophoneneighborhooddensity(Fp0001).Thethree-wayinteractionofthese1887107,factorswasalsosignificant,(F1881585,p0015).Pseu-dohomophonesthatwerevisuallydissimilartotheirsourcewordsdidnotactivatethesourcewords’semantics;hencethed’saresmall.Pseudohomophonesthatwerevisuallysimilartotheirsourcewordsactivatedsemanticpatternsthatstronglyoverlappedwiththesourcewords’semantics,yieldinglarged’s.However,thelattereffectwasmodulatedbyneighborhooddensity.Ifboththepseudohomophoneandsourcewordwerefromdenseneighborhoods,thed’wasverysmall.Thus,thefactthatapseudohomophonesuchasKARisvisuallysimilartothehomophonouswordCARhadlittleimpactbecauseitisalsoclosetomanyotherwords.Wheneitherthewordorpseudohomophonewerefromsparseneighborhoods,thesemanticactivationeffectwasmuchstronger.
Themodelsuggeststhatsomepseudohomophonesacti-vatesemanticinformationdirectlyfromorthography.Thesefindingsarerelevanttopreviousbehavioralstudiesofpseu-dohomophones,whichincludeditemsthatproducedse-manticactivationviaorthographyinthemodel.ManyofthepseudohomophonesintheVanOrdenetal.(1988)study,forexample,werevisuallysimilartothesourcewords.Inad-dition,thepseudohomophonesandvisualcontrolsdifferedintermsofrelevantneighborhoodcharacteristics.VanOr-denetal.(1988)carefullyequatedthevisualsimilarityofthecontrolnonwordsandthepseudohomophonestotheex-emplarusingameasureoforthographicdistance.How-MULTICOMPONENTMODELOFREADING
53
5.0
Visually SimilarVisually DissimilarSemantic Feature d’4.0
3.0
2.0
1.0
0.0
Dense Word,Dense PseuNeighborhoodDense Word,Sparse PseuNeighborhoodSparse Word,Dense PseuNeighborhoodSparse Word,Sparse PseuNeighborhood
Figure35.Effectsofvisualsimilarity,wordneighborhooddensityandpseudohomophoneneighbor-hooddensityforthecomputationofsemanticfeaturesalongorthsem.PSEU=pseudohomophone.
ever,5ofthe10pseudohomophonesusedintheirfirstexperimentwerecreatedbychangingthespellingofthevowelofthesourceword(e.g.,SHEAP/SHEEP)whilere-tainingtheonsetandcoda,whereas1ofthe10controlnon-wordsinvolvedthisminimalchange.Manyofthecontrolnonwordswerealsocloserorthographicallytootherwordsthantotheexemplar(e.g.,PARRIT,thecontrolforCAR-ROT/KARRETisvisuallyclosertoPARROT;HERT,thecon-trolforHEAT/HEETiscloserto(andhomophonouswith)HURT).Thenetresultisthatthestimulivariedintermsoftheneighborhoodpropertiesthataffectedsemanticactiva-tionviaorthographyinthissimulation.
Wetestedthepseudohomophones,thematchedbasewords,andmatchedcontrolitemsusedbyVanOrdenetal.(1988)(excludingoneset,containingKARRIT,becauseitwasbisyllabic)andcomputedthesemanticactivationsfortheseitemsintheintactmodel,themodelwiththeorthphonsempathwaydeleted,andwiththeorthsempathwaydeleted.Thed’oftheresultingsemanticrepresen-tationtotheveridicalsemanticrepresentationsofthebasewordswascalculatedasbefore.Table9showstheresults.
Table9
Semanticd’forVanOrden,JohnstonandHale(1988)StimuliModel
Word
Pseudohom
Control
Note.UNDEF=undefinedd’,astherewerenomissesorfalsealarms.Pseudo-hom=pseudohomophone.
tivationthantheorthphonsempathwayalone.Aswithwords,meaningisjointlydeterminedbyinputfrombothpathways.Thus,themodelisconsistentwithVanOrdenetal.’sconclusionthatthesestimuliactivatemeaningviaphonologicalrecoding,butsuggeststhatsemanticsisalsopartiallyactivatedviaorthsem,contributingtotheoccur-renceoffalsepositiveresponses.
Consistentwiththeaboveobservationsregardingdif-ferencesbetweenthepseudohomophonestimuliandnon-wordcontrols,themodelproducesmoreaccuratesemanticrepresentationsofthetargetwordsviatheorthsempath-waythanthecontrolnonwords,althoughthisdifferenceisrathersmall(ad’of2.0versus1.7).Thedisparitybetweenpseudohomophonesandcontrolnonwordsismuchgreaterfortheorthphonsempathway,indicatingthatthebulkoftheactivationofsemanticsisdoneviathephonologicalpathway,aswouldbeexpected.Itshouldbenoted,how-ever,theintactmodelproducesevenstrongersemanticac-Simulation18:JaredandSeidenberg(1991)(Pseudohomophones)
Inthefinalsimulationsweusedthemodeltoexam-inepossiblebasesforsubjects’decisionsthatpseudohomo-phonesarenotwords.Pseudohomophonesactivateseman-ticsviaphonologyand,aswehavejustseen,somemayactivatesemanticsdirectlyfromorthographyaswell.Inthesimulationstobereportedweexaminedthepatternsofacti-vationproducedbypseudohomophonesandaskedwhethertheydifferedsystematicallyfromthoseproducedbywords,
54
HARM,SEIDENBERG
providingabasisforcorrectrejections.
ForthissimulationweusedthestimulifromJaredandSeidenberg(1991).Theirstudiesincludedbothhomo-phones(theresultsofwhichwerediscussedabove),andpseudohomophones.Aswiththehomophones,forthepseu-dohomophonestheymanipulatedthefrequencyoftheho-mophonousexemplar(e.g.,HF:DAWGDOG,LF:CAUD
COD).Onlythepseudohomophonesoflowfrequencyexemplars(e.g.,CAUD)producedastatisticallyreliablenumberoffalsepositives.
TheaccountoftheJaredandSeidenberg(1991)homo-phonedatapresentedearliersuggestedthattheorthsempathwayprovideddisambiguatinginformationthatallowedsubjectstoavoidfalsepositivesonmosttrials.Itisnotclearwhetherthisaccountcanalsoaccommodatethepseu-dohomophoneresults.Thestimuliintheirexperimentwerevisuallydissimilarpseudohomophoneswhich,wehaveob-served,donotproduceverymuchactivityalongorthseminthemodelandsoitwouldnotprovidethedisambiguatinginformation.VanOrdenetal.suggestedthatsubjectsuseaspellingcheck.Thesimulationexaminedwhetherpseu-dohomophonesprovideanyotherbasisformakingcorrectdecisions.
Method
StimuliandProcedure.Thestimuliwereconstructedalgorithmically,byselectingsetsoffourwordsconsistingoftwopairsofwordswhichrhymebuthavedifferentor-thographicrimes.Pseudohomophoneswerealgorithmicallygeneratedbyswappingtheorthographicwordrimes(e.g.,WAX,CRACKSWACKS,CRAX).Nonwordsweregen-eratedbychangingtheonsets.Thismethodgeneratedalargesetofwords,nonwordsandpseudohomophonesinwhicheachsetof12itemswasperfectlymatchedfordis-tributionofonsetsandrimes;seeTable10foranexam-ple.Asetof158pseudohomophonesresulted;28derivedfromhighfrequencyexemplarsand130fromlowfrequencyexemplars.21Here,asinallsetsof12,theonsets(e.g.,F,CH,DandCL)appearonceandonlyonceineachofthethreecolumns,asdotherimes(ACT,ACKED,IDEandIED).Thevisualsimilaritybetweenpseudohomophonesandtheiryokedwordswasgenerallylow.
Rowswhosewordexemplarwere[objects]or[livingthings]wereextracted,andtheseweresplitintogroupswithahighfrequencyexemplarandalowfrequencyone,inthesamemannerasintheprevioussection.ThepresentationprocedurewasidenticaltothesimulationoftheJaredandSeidenberg(1991)homophoneconditions;theintactmodelwasused.Asbefore,wetrackedtheactivationlevelsofthedistinguishingsemanticfeaturesfortheobjectandlivingthingconcepts.
WordPseudohomophoneNonword
Table11
ReplicationofJaredandSeidenberg(1991)Pseudo-homophoneExperimentCondition
LFExemp.
HFExemp.
Results
Table11presentssummarydataconcerningseman-ticactivityforthepseudohomophonesoflowandhighfrequencyexemplars.Pseudohomophonesproducedhighamountsofactivationonthecriticalsemanticfeatures,muchmorethanseeninthesimulationsofVanOrden(1987)orthewordeffectsinJaredandSeidenberg(1991)describedpreviously.Thisdegreeofsemanticactivationisconsistentwithproducingalargerfalsepositiveratethanobservedinthebehavioralstudy.Further,theeffectisintheoppositedirection:thepseudohomophonesofHFex-emplarsproducedreliablymorefalsepositivesthantheLFones(F1156148,p0001),whereastheyproducefewerfalsepositives.
Theseresultsfollowfrompropertiesofthemodelwehavediscussedpreviously.Theorthsempathwaydoesnotgeneratesignificantactivationforpseudohomophonesandnonwordsthatareloners(i.e.,apseudohomophonethatisvisuallydissimilartoitssourcewordornonpseudohomo-phonethathasfewneighbors).Theonlysourceofsemanticactivationisviaphonsem,viawhichpseudohomophonesreliablyactivatesemanticfeaturesofthesourceword.Fur-ther,giventhatthephonsemcomponentisfrequencysen-sitive,pseudohomophonesofhighfrequencyexemplarsac-tivatesemanticsmorestronglythanlowfrequencyexem-plars.Somethingelseisclearlyneeded,however,toac-countforthefactthatsubjects’falsepositiveratesaretyp-icallylow,withpseudohomophonesofhighfrequencyex-emplarsgeneratingfewerfalsepositivesthanthoseoflowfrequencyexemplars.
Onepossibilityisthatthereareothersourcesofin-formationrelevanttomakingthedecisionavailablewithintheexistingmodel.Asmentionedearlier,Plaut(1997)
MULTICOMPONENTMODELOFREADING
55
usedatermcalledstress(seeEquation7)tomeasurehowstronglyunitswereactivated.Plaut(1997)foundthatwordstendedtoproducehigherstressthannonwords,andthiswaspositedasabasisformakinglexicaldecisions.WefoundinSimulation2thatwordsproducedgreaterstressthannon-words.Therefore,wefollowedthemethodusedinSimula-tion2andcomputedthestressforitemsinthissimulation.
Unfortunately,likethesemanticactivation,thesemanticstressmeasureshowedtheoppositepatternfromthebehav-ioraldata(Table11).ThepseudohomophonesofHFexem-plarsproducedhigherstress,whichmeansthereislessofareasontorejecttheitemasanonword.However,pseudoho-mophonesderivedfromhigherfrequencywordsareeasierforsubjectstorejectaswordsthanonesderivedfromlowerfrequencywords(Jared&Seidenberg,1991).Thepseu-dohomophonesofHFexemplarsproducehigherstressforthesamereasonthattheyproducemoreactivationoftheinappropriatesemanticfeature:thephonsempathwayisfrequencysensitive,andhighfrequencyphonologicalformscanmorepowerfullyactivatesemantics.
Inthepresentcontext,theimportantquestioniswhetherthemodelwehavedescribediscompatiblewiththefactsabouthowsubjectsprocesspseudohomophones.Ourgen-eralviewisthatmakingsemanticdecision(isitamemberofacategory?),likemakingalexicaldecision(isitaword?)isajudgmenttaskofconsiderablecomplexity(seeSeiden-berg,1985,fordiscussion).Thetaskdemandsthatthesub-jectestablishcriteriaforreliablymakingaccuratedecisions.Themodeltellsussomethingaboutthekindsofinforma-tionthatbecomeavailablewhenwordsandnonwordsareprocessed.Thisinformationisthenusedinperformingvar-ioustasks,suchassimplycomputingthemeaningofaletterstring,namingitaloudormakingsemanticorlexicaldeci-sionsaboutit.Taskssuchaslexicalandsemanticdecisioninvolveadditionalprocessesrelatedtomakingsuchjudg-ments.Weknowthatwordsandpseudowordsproducedif-ferentactivationpatternsinthemodel.Forexample,SUITisamorefamiliarspellingpatternthanSUTE,whichcouldbedetectediforthographywere,likephonologyandsemanticsintheimplementedmodel,treatedasanattractorsystem.Similarly,SUTEandSUITdonotproduceidenticalseman-ticpatterns.Howthesedifferencestranslateintodecisioncriteriarequiresatheoryofhowsuchtasksareperformedwhichisbeyondthescopeofthecurrentwork.
Ofcourse,thereisanotherpossibility:aspellingcheck.Althoughthespellingcheckprocedureisnotnecessaryfordisambiguatinghomophones(asdiscussedabove),itmayberequiredforpseudohomophonesandotherveryword-likenonwords.Thismakesintuitivesense:forfamiliar,learnedwords,theorth-sempathwayprovidesdisambiguat-inginformation;fornovel,unlearnedwords,orth-sempro-videsnousefulinformationandsothemodel/readermustchecktoseewhetherameaningisassociatedwithapar-ticularspelling,i.e.,generatetheorthographiccodefromsemantics.Themodelwehavebeendiscussingcannotper-formthiscomputation;forsimplicitywedidnotimplementthesemanticstoorthographyconnections,whichwouldhaveaddedsignificantlytothealreadyconsiderabletimerequiredtotrainthemodel.SeidenbergandMcClelland(1989)conductedpreliminaryresearchalongtheselines,however.Theirmodeloftheorthphoncomputationin-cludedafeedbackloopfromtheorthographicinputtoitselfviaanintermediatesetofhiddenunits.Thispermittedthecalculationofthediscrepancybetweentheveridicalinputpatternandtheonethatwasrecreatedontheinputunitsviathisfeedbackloop.Thisscorereflectedhowword-likealetterstringwasrelativetotheentiretrainingcor-pusandprovidedabasisformakingsomeword-nonworddecisions.Thefollowingsimulationextendsthisideabyconsideringthediscrepancybetweentheorthographicinputandonecomputedbymeansofthesemorthpathway.Weimplementedasimpleformofthesemanticstoorthographycomputationinordertodetermineifitwouldprovideasuf-ficientbasisfordetectingthatpseudohomophonesarenotwords,assuggestedbyVanOrdenetal.(1988).
Simulation19:Jaredand(PseudohomophonesSeidenberg(1991)
Revisited)
Method
Thebasicmethodinvolvedaddingasemanticstoor-thographypathwaytothemodel,illustratedinFigure36.Aftertrainingofthiscomponentwascomplete,thespellingcheckprocedurewasoperationalizedasfollows.Apseudo-homophonewaspresentedasinput,andsemanticswasacti-vatedviatheintactorthsemandorthphonsempath-ways.Inaddition,theactivatedsemanticpatternwasusedtocomputeanorthographicrepresentationviasemorth.Thespellingcheckwasbasedontheorthographicpatterncomputedonthebackwardpassfromsemantics.
Thismethodofimplementingthesemantics-orthographycomputationisasimplificationinsofarasitusesaduplicatesetoforthographicunitsandthenacomparisonbetweenthemtodeterminehowwordlikealetterstringis.Asnotedabove,thiswasdoneforcomputationalfeasibility.Ideallythesemanticunitswouldhavefeedbackconnectionstothesameorthographicunitsusedtoinputaword.Thespellingcheckwouldthenbeperformedbydetermininghowwellthemodelrecreatestheinputpatternthroughthefeedbackconnections.SeidenbergandMcClelland(1989)implementedthisprocedureintheirmuchsimplermodelandusedittocomputewhattheytermedanorthographicerrorscorewhichprovidedanindexofhowworldlikealetterstringis.SeidenbergandMcClelland(1989)providedevidencethatthiscomputationoforthographicfamiliarityplaysaroleinmakinglexicaldecisions.Thepresentmodelimplementedthesameideausingasomewhatsimplertechnique,necessitybythecomplexityoftrainingthemuchlargermodel.
56
HARM,SEIDENBERG
SemanticsPhonologyOrthography2OrthographyFigure36.Therevisedmodel,withsem
orthpathwayimple-mented.
StimuliandProcedure.Thestimuliforthisexperimentwerethesameasintheprevioussimulation.Thesemorthcomponentwastrainedinthesamefashionastheothersim-ulations.Itwastrainedfor800,000wordpresentationsus-ingtheentiretrainingcorpus,atwhichpointtraininghadasymptotedat99%accuracy.Thesemorthmodelwasthenattachedtotheexistingmodelasshowninthefigure.
Wemeasuredthreevariables:thedisparitybetweentheorthographicinputandtheorthographicrepresentationre-createdfromsemantics,thestressonthosere-createdor-thographicrepresentations,andthesemanticstress.Thespellingcheckwasoperationalizedasacomparisonbe-tweentheinputorthographicpatternandthepatternrecom-putedonthebackwardpassfromsemantics.Iftheinputisacorrectly-spelledwordinthemodel’svocabulary,thetwopatternswillcloselymatch.Iftheinputisnotthecor-rectspellingofaword,therewillbeadiscrepancybetweenthetwoorthographiccodes.Thus,SUTEwillactivatethesemanticsofSUITviaorthphonsem,butthissemanticpatternwillactivatethespellingSUITviasemorth.Thedecisiontorejectthestimuluswilldependonthedegreeofdiscrepancyandthemodel’sconfidenceabouttheword’sspellingpattern,whichwasreflectedinthestressmeasureovertheorthographicunits.
ResultsandDiscussion
Table12depictstheresultsforthewords,pseudo-homophonesandnonwords.Theeffectofexemplarfre-quencywasreliableforthepseudohomophonesfortheor-thographicstressmeasure(F115633,p0001),thesemanticstressmeasure(F115641,p0042),andtheorthographicdistancemeasure(F01).Asbefore,thesemanticstressmeasure11566produced5,p0ef-fectsintheoppositedirectionastheempiricaldata:higherstressforthehighfrequencyitems,whichwouldmakeitmoredifficulttorejectsuchitemsasnonwords.
However,theorthographicstressmeasureandtheor-thographicdistancemeasureeachpatternedinthecorrectdirection.Thiswasbecausethesemanticrepresentations
wereactivatedmoreweaklybythephonologicalformofthelowfrequencyexemplars,andhencere-createdamorenoisyorthographicrepresentation,resultingingreateror-thographicdistance,andagreaterbasisforrejectingtheitemasnotbeingaword.Further,theorthographicstressforthepseudohomophonesoflowfrequencyexemplarswaslowerthanforthoseofhighfrequencyexemplars.Thus,thenetworkhadgreaterevidenceforrejectingthepseudoho-mophonesofHFexemplars,onthebasisoferrorandconfi-denceinspelling,thantheLFexemplars.
Tosummarize,thesimulationsindicatethatsomepseu-dohomophones(onesthatarevisuallysimilartotheirho-mophonouswordandfromorthographicallysparseneigh-borhoods)activatesemanticinformationviaorthsem.Vi-suallydissimilarpseudohomophonesyieldlittleactivationalongthispathway.Theeffectofexemplarfrequencyob-servedbyJaredandSeidenberg(1991)canbeaccountedforbyincludingfeedbackfromthesemanticsystemtotheorthographiccomponent.ThisisasimpleversionofthespellingcheckproposedbyVanOrdenandcolleagues(e.g.VanOrdenetal.,1990),butwithouttheadditionalassump-tionthatorthographydoesnotactivatesemanticsdirectly.
Asnotedearlier,aformalsimulationoflexicaldeci-sionisbeyondthescopeofthiswork.Suchasimulationwouldincludeadetailedaccountoftheprocessesinvolvedinmakingboth“yes”and“no”decisions,andwouldhavetoaccountforamassofpublishedresultsshowingthatlexicaldecisionresultsareaffectedbyexperiment-specificfactorsthataffectsubjects’responsestrategies(seeSei-denberg,1995,fordiscussion).However,Table12showssomeofthesourcesofinformationthatcouldplausiblybeusedinthelexicaldecisiontask.Thesemanticstressmeasurediffersverystronglybetweenwordsandnonwords(F1157681,p0001),asdoestheorthographicdis-tance(F1157752,p0001).Figure37showsthedistributionofstressvaluesforthewords,nonwordsandpseudohomophonesinthisexperiment.Orthographicstressisalessstrongdiscriminator,butstilldiffersreliablyforwordsandnonwords(F1157129,p0001).Thesevariablesaremostlikelynottheonlyonesthatcouldbeinvolvedinperforminglexicaldecision(forinstance,onecouldplausiblyjudgethatXPMKisnotawordsimplybynotingthatitdoesnotcontainavowel).Nonetheless,thesevariablesproduceresultsthatprovideabasisonwhichlex-icaldecisionscouldbemade.
9.GENERALDISCUSSION
Asnotedattheoutset,therehasbeenconsiderablede-bateconcerningthemechanismsinvolvedincomputingthemeaningsofwordsfromprint.Althoughpositionsontheissuevary,mostdiscussionshavepresupposedthatthereareindependentdirect-visualandphonologically-mediatedpathwaysandthatforanygivenword,oneofthesemech-anismsprovidesaccesstomeaning.Sometheoriesas-
MULTICOMPONENTMODELOFREADING
57
Table12
SecondReplicationofJaredandSeidenberg(1991)PseudohomophoneExperiment.Measure
HF
LF
HF
LF
HF
LF
Note.Standarddeviationsareshowninparenthesis.
100
WordPseuNWs80
metI fo60
noitro40
porP20
00.0
0.20.40.60.81.0
Stress Level
Figure37.ThedistributionofstressvaluesforitemsusedinSimulation19.Pseu=pseudohomophones,NW=nonwords.
sumethedirectrouteisused,somethatphonological-mediationisdominant,andothersthatbothroutesareusedbutfordifferentwordsorwritingsystems.TheVanOr-denetal.(1990)articlewasadepartureinsofarasitem-phasizedtheinteractivitybetweendifferentcomponentsofthesystem;however,theyalsoidentifiedfactorsthatwerethoughttocauseprocessingtoproceedprimarilyviaorthphonsem,withadditionalfeedbackfromseman-ticstoorthographyforhomophones.
Ourviewisthattheexistenceofthedirect-visualandphonologically-mediatedpathwaysneedstobeconsideredseparatelyfromcomputationalpropertiessuchasthekindsofrepresentationstheyoperateover,howtheyarelearned,andwhethertheyareindependent.Everymodelofwordreadingwillhavetoincorporatesomeversionofthetwoproceduresbecausetheyarelicensedbythenatureoftheorthographic,phonological,andsemanticcodesandthere-lationshipsbetweenthem.Ourmodeldiffersfromprevi-ousaccountsinacriticalway:meaningsaredetermined
bybothpathwayssimultaneously.Themodelalsodiffersfromotherproposalswithrespecttohowthemechanismswork.Forexample,thetraditionalideathatthevisualpath-wayinvolvesactivatingatomicentriesinamentallexicondiffersinessentialwaysfromtheideathatapatternofac-tivationdevelopsoverthesemanticunitsbasedonortho-graphicinput.Similarly,mostprevioustheorieshaveas-sumedthatphonologicalcodesaregeneratedbyapplyinggrapheme-phonemecorrespondencerules,butourmodelinvolvesastatisticallearningprocedure.Thesedifferencesbetweentheoriesmatter;theyrepresentdifferentclaimsabouthowknowledgeisrepresented,acquired,andusedandultimatelyhowitisrepresentedinthebrain.
Wehavedescribedthebasicoperationofthemodelindetailandshownthatitisconsistentwithvarioustypesofbehavioraldata.Thedynamicsoftheimplementedmodelarecomplexbuttheprinciplesthatgovernitsbehavioraremuchsimpler.Webuiltandtrainedamodelconsistentwiththetheoreticalframeworkoutlinedintheintroduction,whichincludesexplicitclaimsaboutthenatureofthewordreadingproblemandhowthetaskisperformedbyhumans.Thebehaviorsofthemodelthatwehavedescribedfol-lowedasempiricalconsequences.Wethenobservedthatthemodelwasconsistentwithvariousbehavioralphenom-ena.Themodelalsoprovidednovelinsightsaboutmanyphenomena,insofarastheyarisefromsomewhatdifferentmechanismsthanhadbeenproposedinothertheories.
Inconcludingthisarticlewewillsummarizetheessen-tialpropertiesofthemodelanddiscussissuesthatneedtobeaddressedinfutureresearch,includinglimitationsofthecurrentimplementation.
SummaryofProperties
theModel’sBasic
1.Activationofmeaningfrommultiplesources
Theactivationofsemanticsbuildsupovertimebasedoncontinuousinputfromallavailablesources,principallyorthsemandorthphonsem,butalsothesemanticcleanupcircuit.Thischaracteristicderivesfromthearchi-tectureofthemodel,particularlythefactthatitsettlesinto
58
HARM,SEIDENBERG
adistributedsemanticpatternovertimeratherthaninstan-taneouslyaccessingastoreddefinition-likemeaning.Inconnectionistterminology,thecomputationofmeaningisaconstraintsatisfactionproblem:thecomputedmeaningisthatwhichsatisfiesthemultipleconstraintsrepresentedbytheweightsonconnectionsbetweenunitsindifferentpartsofthenetwork.2.Courseoflearning
Learningoccurswithinboththeorthsemandorthphonsemcomponentsthroughoutthecourseoftraining,asindicatedbythelesionexperiments(Figures14-15).Withsufficienttrainingbothpathwaysbecomehighlyaccurateformanywordsandthusbothmakesignificantcontributionsintheintactmodel.
3.Precedenceofphonologicalpathwayinacqui-sition
Becauseofthenatureofthemappings,theorthsemcomponenttakeslongertodevelopthanorthphonsemandsophonologicalmediationassumesprimacyearlyon,asinbeginningreaders.
4.Developmentoftheorth
sempathway
Theorthsempathwayhasanadvantageoverorthphonsembecauseinthelattercasesemanticacti-vationcannotoccuruntilthepatternoverthephonologicalunitshasbecomesufficientlyclear.Theorthsempath-wayhasanintrinsicspeedadvantagebecauseitinvolvesfewerintermediatesteps.Thisproperty,takenwithatrain-ingprocedurethatemphasizesproducingsemanticpatternsrapidlyaswellasaccurately,leadstocontinueddevelop-mentinorthsemevenwhenorthphonsemproducescorrectoutput.Theotherfactorthatpromoteslearningintheorthsempathwayisitsroleindisambiguatingthemanyhomophonesinthelanguage.
5.Capacityoftheorth
sempathway
Althoughithasanintrinsicspeedadvantage,theorthsempathwaytakeslongertolearn,whichlimitsitsroleinitially.Thecapacityoforthsemisalsolimitedbythefactthatsomeoftheworkisdonebyorthphonsemandthesemanticcleanupapparatus.Thus,theorthsempathwayisnotforcedtodelivertheentiresemanticpatternforawordbyitselfwithinthefirstfewtimesteps.Inread-ingalowfrequencyhomophonesuchasEWES,forexam-ple,orthsemonlyhastoactivatesufficientinformationtosuppresstheincorrectmeaningbeingactivatedthroughorthphonsem.
6.Co-operativecomputationofmeaning
Giventhedynamicsofthesystemandthecomputa-tionalpropertiesofthecomponents,thenetresultisthat
semanticsreceivessignificantinputfrombothorthsemandorthphonsemforalmostallwords.Moreover,themodelwithbothpathwaysintactcomputesmeaningsmoreefficientlythanthepathsdoindependently.Thedivisionoflaborbetweenthetwoisaffectedbylexicalpropertiesincludingfrequencyandspelling-soundconsistencyaswellastheamountoftraining.
7.Processingofhomophones
Undernormalpresentationconditions,homophonesaredisambiguatedthroughtheuseofbothorthsemandorthphonsem.Theisolatedorthphonsempathwaycanproducecorrectpatternsforhigherfrequency,dominanthomophones.Intheintactmodel,however,orthsemalsodeliversrelevantactivationquickly,particularlyforhigherfrequencywords.Theroleoforthsemisshapedbythefactthattheorthphonsempathwaycannotaccuratelycomputebothmeaningsofahomophonepair.Thelat-terpathwayeventuallybecomesmoretunedtothehigherfrequencymemberofapairbecauseitistrainedmoreof-ten;however,orthsemalsoprocessesthesewordseffec-tivelyandsocontributessignificantly.TheanalysisinFig-ure29showsthattheorthsempathwaybecomesveryef-fectiveatsuppressingfeaturesassociatedwiththealterna-tivemeaningthatareactivatedthroughphonology.
8.UseofOrth
SemandSem
Orth
Intheimplementedmodel,homophonesaredisam-biguatedusinginformationfromorthsemratherthanaspellingcheck(semorth).Thisaspectofthemodeldemonstratesthatthereisnocomputationalreasonwhyorthsemcannotcontributetosemanticactivation,andthemodel’sbehaviorindisambiguatinghomophoneswascon-sistentwiththatseeninhumansubjects.Althoughwedidnotincludeitinthisimplementation,thereisnoreasontoprohibitfeedbackfromsemanticstoorthographywhichmayalsoplayaroleinhumanperformance.Thecontribu-tionfromorthographytosemanticsismoredirect,however,andthuscanbeutilizedmorerapidly.
9.Effectsofmasking
ThesimulationssuggestthatthefalsepositiveresponsesobservedinstudiessuchasVanOrdenetal.(1988)arisebecausethenormalinputfromorthsemisterminatedbypresentationofamask.Thiscontrastswiththestandardin-terpretationthatthemaskremovestheorthographicpatternusedinmakingapost-accessspellingcheck.Maskinghaslessofaneffectonactivationwithinorthphonsem;thephonologicalsystemisahighlystructuredattractorwhichallowspatterncompletiontooccurevenintheabsenceofcontinuedorthographicinput.Althoughthesemanticsys-temisalsoanattractor,itismoresparseandthereforehighlydependentoninputfromothersources(eitheror-thographyorphonology).Theprimingeffectsobservedin
MULTICOMPONENTMODELOFREADING
59
studiessuchasLeschandPollatsek’sariseinasimilarman-ner.
FutureDirections
Themodelwehavedescribedisapartialrealizationofabroadertheory.Theimplementationalstepwasnottrivial;itinvolvedsignificantchallengesconcerningthephonologicalandsemanticrepresentations,trainingbothcomponentsofthemodelsimultaneously,analyzingthemodel’sbehavior,andrelatingittobehavioralevidence.Althoughthemodelhasconsiderablescope,therearemanyotherphenomenathatcanbeexploredusingthisversionofit.Ourdecisiontolimitthediscussionofthemodeltotheresultspresentedabovewasmotivatedbypracticalconsiderations(theneedtokeepthearticletoamanageablelength;thedesiretogetthetheoreticalframeworkintotheliteraturesothatotherscouldutilizeit),ratherthanhavingexhaustedtherangeofphenomenatowhichthemodelcanbeapplied.Belowwedescribesomeoftheissuesthatcanbepursuedusingtheexistingmodel.However,themodelcanalsobeseenasinstantiatingacomputationalframeworkortoolkitforgen-eratingandtestinghypothesesaboutmanyaspectsofread-ing,byvaryinghowitisconfiguredandtrained.Suchex-plorationsmayshedlightonadditionalreadingphenomenaandalsohelpinidentifyinglimitationsoftheframeworkandthecurrentimplementation,whichcanbeaddressedinfuturemodels.Wetakethisexploratoryfunctionofthemodeltobeasimportantasshowingthatthisparticularim-plementationcanaccountforadditionalfacts.Belowwebrieflysummarizesomeoftheprominentdirectionsforfu-tureresearch.
1.Robustnessoftheimplementation
Thegeneralformofthemodelwascloselytiedtothe-oreticalconcernsbutmanydetailsoftheimplementationwerenot.Implementingthemodelrequiresmakingdeci-sionsaboutdetailssuchasthenumberofhiddenunitsinapathway,thesettingoftheparameterthatdetermineshowrapidlyactivationrampsup,andthewaywordsaresampledduringtraining.Itwillbenecessarytodeterminewhethertheseaspectsoftheimplementationcontributeinsignificantwaystoitsbehavior,whichcanbedonebycomparingvari-antsofthebasicmodels.Wethinkthemodel’sbehaviorislikelytoberobustbecauseofthewayitwasdeveloped,whichdidnotinvolvetryingalargenumberofpossibilitiesandthenfindingtheonesthatproducedthebestresults.Wemadeimplementationaldecisionsbasedonpreviousexpe-rienceandourunderstandingofnetworkbehaviorandthenobservedtheconsequences.ThisstronglycontrastswiththeapproachofColtheartetal.(2001),whosemethodol-ogyexplicitlyinvolvesfittingmodelstodataratherthanderivingresultsfrommoregeneralprinciples.Somepa-rametersofourmodelareexpectedtoaffectperformancebutintheoretically-interpretableways.Forexample,Sei-
denbergandMcClelland(1989)foundthatreducingthenumberofhiddenunitsintheorthphonpathwayaffectedtheirmodel’scapacitytolesscommonspelling-soundmap-pings;thisparametermayberelatedtoindividualdiffer-encesamongreaders.Otherparametersthatwerechosenforpragmaticreasons(e.g.,tokeepnetworkrunningtimewithinthelimitssetbyourcomputers)canalsobevaried(e.g.,usingfastercomputers).Thesekindsofparametersshouldnothavealargeimpactoncoreaspectsofthemodel(e.g.,thefactthatmeaningsarejointlydeterminedbyin-putfrombothpathways),butthisneedstobedeterminedempirically.
2.Generatingandtestingnewpredictions
Onequestionoftenraisedinconnectionwithsimula-tionmodelsiswhetheritispossibletogobeyondmerelyaccountingfortheresultsofexistingstudiestogeneratingtestablenovelpredictions.Thisquestionisofparticularconcernwithrespecttomodelsthataredevelopedbyfit-tingparticularbehavioraldata(Seidenberg,Zevin,&Harm,2002),butourmodelwasnotdevelopedinthisway,aswehaveemphasizedthroughoutthisarticle.Twoquestionsdoneedtobeaddressed,however:(a)doesourmodelaccountforphenomenaothertheoneswehavedescribed?(b)doesthemodelgeneratenovelpredictionsthatcanbetestedinnewbehavioralexperiments?
Themodelisadevicethatgeneratesphonologicalandsemanticcodesforwords.Theresearcherthengenerateshypotheses(basedonhumanormodelperformance)andteststhembyrunningappropriatesimulationandbehav-ioralexperiments.Ourexperiencewithpreviousmodels(Seidenberg&McClelland,1989;Plautetal.,1996;Harm&Seidenberg,1999)isthatresearchershavethoughtofmanyhypothesesthatcanbetestedusingourmodels.Thuswehaveprovidedmodel-generateddatawhichhavebeenusedinstudiessuchasSpielerandBalota(1997),Jared(1997),Treiman,Kessler,andBick(inpress)andothers.Thecurrentmodelgeneratesmanypredictionsthatcanbetestedimmediately;forexample,basedonthemodel’sper-formancewecoulddesignanexperimentthatwouldbeanadvanceontheVanOrdenparadigminsofarasitmadespe-cificpredictionsaboutwhichhomophonesorpseudohomo-phonesactivatesemanticsandthusarelikelytogeneratefalsepositives.Thesemanticrepresentationsinthemodelprovideabasisforgeneratingpredictionsabouthowseman-ticstructureaffectsperformanceontaskssuchassemanticpriming,categorydecision,similarityjudgment,andmanyothers.McRaeetal.(1997)showedthatthemagnitudeofsemanticprimingeffectscouldbepredictedbymeasuresoffeaturaloverlapbetweenprimeandtarget;ourmodelcanalsobeusedtogeneratespredictionsaboutthemag-nitudeandtimecourseofsucheffects,usingmaskedandunmaskedstimuli.
Amuchbroaderrangeofphenomenacouldbead-dressedbyextendingthemodeltoincorporateanexplicit
60
HARM,SEIDENBERG
theorylinkingmeasuresofnetworkperformancetore-sponselatencies(seebelow).Finally,themodelmakessomepredictionsthatareveryexplicitbutchallengingtotestusingexistingmethodologies.Thissituation,inwhichatheorymakespredictionsthatawaitthedevelopmentofmethodsfortestingthem,isnotuncommoninmanysci-ences.Forexample,themodelmapsoutthetimecourseofactivationalongdifferentpathways,butthisisdifficulttoassessinabehavioralstudy.Toillustratetheproblem,therearemethodsfordetectingtheuseofphonologicalin-formationinactivatingmeaningbutnotacomparablydirectmethodfordetectingwhenmeaninghasbeenactivateddi-rectlyfromprint.Afalsepositivefor“isitaflower?ROWS”providesstrongevidenceforphonologically-mediatedacti-vationofmeaning,buttheabsenceofafalsepositivecan-notbetakenasevidencethatphonologicalmediationdidnotoccur(itcouldbethatphonologicalmediationoccurredbutthesubjectwasabletoavoidafalsepositiveusingotherinformation,e.g.,orthsem).Itmaybethatneuroimagingtechniqueswillsoonbeabletoprovideevidenceaboutthetimecourseofprocessinginbrainregionsthatunderliedi-rectandphonologically-mediatedmechanisms,particularlyonessuchasMEGthatyielddynamicratherthanstaticin-formation.Couplingthemodelwithsuchtechniqueswouldfacilitatetestingthemodelandalsofacilitateinterpretingsuchneuroimagingdata.
3.Otherphenomena
Ourfocushasbeenonissuesconcerningthedivisionoflaborinthecomputationofmeaningbutthemodelcanbeusedtoaddressadditionalissues.
Divisionoflaborinpronunciation.Issuesconcerningthepronunciationofwordsandnonwordshavebeenthefo-cusofconsiderablepreviousmodelingresearchwithinthetriangleframeworkandinColtheartetal.’s(1993,2001)DRCmodel.Oneissueiswhetherthemodelwehaveproposedcanaccountforthenamingphenomena(e.g.,frequencyandconsistencyeffects)thathavebeenthefo-cusofongoingdebateabouttheadequacyofthetwoap-proaches.Asecondissueconcernstheroleofsemanticinformationinnamingaloud.Wehaveextensivelydis-cussedhowtheorthsemandorthphonsempathwaysjointlydeterminemeanings.Thecomplementaryissuewithrespecttopronunciationconcernsthecontributionsoftheorthphonandorthsemphonpathwaysinpronuncia-tion.Thecomputationofphonologyisconstrainedbythesameprinciplesthatwehavediscussedwithrespecttothecomputationofmeaning.Thephonologicalcodeforawordwillbejointlydeterminedbyinputfrombothpathways;howevertheresultingdivisionoflabormayhaveadiffer-entcharacterthanwehaveobservedforthecomputationofmeaning.Inthecaseofmeaning,bothpathwayscontributesignificantly;thetradeoffsbetweenthetwopathwayswithrespecttocomputationalefficiencymeanthatneitherdom-inatesinskilledperformance.The“direct”pathwayhasanadvantagebecauseitinvolvesfewerstepbutadisadvantagebecausethemappingislargelyarbitrary.Inthecomputationofphonology,however,the“direct”pathwayalsoinvolvesthemoreconsistentmapping;henceitshoulddominatetoaconsiderabledegree.Thereissomeevidencethatsemanticinformationplaysaroleinnamingforsometypesofwords,particularlyonesforwhichthecomputationfromorthog-raphytophonologyisverydifficult(e.g.,becausetheyin-volvehighlyatypicalspelling-soundmappings;Strainetal.,1995),buttheseeffectsmayberelativelyrare,atleastinEnglish.
Otherwritingsystems.Oneofthemainfactorsthatde-terminedthedivisionoflaborinthepresentmodelwasthenatureofthemappingbetweenorthographyandphonology,whichisquasiregular(Seidenberg&McClelland,1989).Otheralphabeticwritingsystems(suchastheonesforItal-ian,Spanish,andSerbo-Croatian)adheremorecloselytotheprinciplethatindividuallettersorcombinationsoflet-terscorrespondtoasinglephoneme(Seidenberg,1992b;Hung&Tzeng,1981).ThemodelwastrainedonEnglishbutwithminorchangesintheinputrepresentationandthedevelopmentofsuitabletrainingcorpora,itcanbetrainedonotherwritingsystems.Themodelcouldthenbeusedtomakecross-orthographypredictionsandsimulateresultsofbehavioralstudies.
Howthedivisionoflaborisachievedindifferentwrit-ingsystemsislikelytobeacomplexissueinvolvingin-teractionsamongseveralpropertiesofthewritingsystemsandthelanguagestheyrepresent.Todate,mostdiscus-sionhasfocusedononedesignfeature,orthographicdepth,i.e.,theconsistencyofthemappingbetweengraphemesandphonemes.Otherfactorsbeingequalthisfactorwillcer-tainlyaffectthedivisionoflaborbetweenvisualandphono-logicalpathways.However,theeffectsofnumerousotherfactorsneedtobeconsidered.ConsiderthedualCyrillicandRomanwritingsystemsforSerbo-Croatian,whichhavebeenextensivelystudied(e.g.,Lukatela,Turvey,Feldman,Carello,&Katz,1989).BothalphabetsareshallowandthereforelackminimalpairssuchasMINT-PINTinEnglish.However,thesewritingsystemsdonotrepresentsyllabicstress,andSerbo-Croatianhasmanyminimalpairscon-sistingofwordswiththesamespellingbutdifferentpro-nunciationsandmeanings,duetodifferencesinstressorintonationcontour.Thus,LUKhastwodistinctmeanings(arch,onion)dependingonwhetherthevowelisshortandrisingorlongandfalling.Thus,theSerbianandCroatianorthographiesexhibitconsiderableambiguityinthemap-pingbetweenspellingandsounddespitebeing“shallow”atthelevelofgraphemesandphonemes.Moreover,theseam-biguitiesalsoexistinthemappingfromspellingtomean-ing.Resolvingtheambiguitiesmaythereforerequireus-ingcontextualinformation(asrequiredforEnglishhomo-graphssuchasWINDandnoun-verbalternationssuchas
MULTICOMPONENTMODELOFREADING
61
CONtrastvs.conTRAST).Similarly,Hebrewisashalloworthographywhenitsvowelsarerepresented,buttypicallytheyarenot.Removingthevowelsshiftstheorthographyto“deep,”againcreatingdependenceoncontextualinforma-tionforambiguityresolution.Althoughwedrawdiagramsofourmodelingframeworkwithcontextunits,wehavenotexploredtheiruse.Contextseemsparticularlyrelevant,however,tounderstandingambiguitiesthatariseinwritingsystemsforreasonsotherthantransparencyofgrapheme-phonemecorrespondences.
Althoughmostresearchhasfocusedonalphabeticwrit-ingsystems,thereisconsiderabledataconcerningthenon-alphabeticwritingsystemsforChineseandJapanese.AnimportantrecentcorpusanalysisofChinese(Shu,Chen,Anderson,Wu,&Xuan,2003)showedthatalargeper-centageofChinesewordsconsistofphonologicalandse-manticcomponentsthatjointlyprovidecuestotheword’smeaning.Thus,thevisualandphonologicalprocessesinthemodelarerealizedbycomponentsofthewordsthem-selves.ReadingChinesewordsisaclassicconstraintsatis-factionproblem:whereasthecomponentsinisolationmaybeambiguous,theconjunctionofthecomponentsishighlyconstraining.Shuetal.’sanalysessuggestthatChinesehasmuchincommonwithEnglishwithrespecttonatureofthemappingsbetweenthewritten,spoken,andsemanticcodesforwords,thefactthatirregularmappingstendtooccurinhigherfrequencywords,theexistenceofquasireg-ularneighborhoodsofrelatedwords,andsoon.Thesefactssuggestthattheremaybemoresimilaritiesintheprocess-ingofEnglishandthenonalphabeticChinesewritingsys-temthanbetweenEnglishandashallowalphabeticwritingsystem,butthisremainstobeexploredindetail.Itwouldnotrequiremajortechnicalinnovationtobeabletorepre-sentChinesecharactersasthe“orthographic”inputinourmodel.Withasuitabletrainingcorpus,themodelcouldthenbeusedtoexaminewherethestatisticalregularitiesinthewritingsystemlie,howthedifferentcomponentsofwordsjointlydeterminemeaning,andhowtheresultingdi-visionoflaborcomparestothatforEnglishandotherwringsystems.
Acquisition.Mostofthefindingsdiscussedinthisarti-cleconcernskilledperformance.Readingacquisitionwasconsideredonlywithrespecttocomputationalpropertiesthatyieldinitialdominanceoftheorthphonsempath-way.Inon-goingresearchweareexaminingdevelopmen-talissuesinmoredetail.Onegoalistoutilizeatrainingregimethatadheresmorecloselytothechild’sclassroomexperience.Inlearningtoread,childrenareinitiallyex-posedtoasmallvocabularyofwordswhichexpandsovertime.Instructionalprogramsstructurethisexperienceindifferentways.Ourmodelsuseafrequency-biasedsam-plingprocedurewhichdoesnotbuildasmuchstructureintothesequenceoflearningevents.Incurrentworkweareex-amininghowperformanceisaffectedbydifferentwaysof
structuringthissequence(Foormanetal.,2001),especiallywhethertherearewaystooptimizeefficiencyoflearning.Arelatedissueconcernsthenatureofthefeedbackprovidedtothechildormodelinthecourseoflearning.Weusedanidealizedprocedureinwhichthemodelwasprovidedwithfeedbackaboutthecorrectsemanticandphonolog-icalcodesforwords.Childrenreceivemorevariablefeed-back;explicitfeedbackfromateacherorlistenerissome-timesprovidedbutmoreoftenchildrenprovidetheirownfeedback(e.g.,bylisteningtowhattheyhavesaid;byus-ingbackgroundknowledgeorillustrationstoinferintendedmeaningsofwords).Thisfeedbackcanbepartialorevenincorrect.Ourgeneralviewisthatthelearningthatoccursundertheseconditionsfollowsthesameprincipleswehaveexploredbutmaybelessefficient.Ontheotherhand,chil-drenreceiveadditionalinstructionthatfocusesonpartsofwords(e.g.,thepronunciationsoflettersorrimes),whichcanalsobeincorporatedinthetrainingregimeandmayim-proveefficiency.Ingeneralthemodelprovidesapowerfultoolforexaminingassumptionsabouthowtoteachwordreading.
Acquireddyslexia.Dataconcerningthepartiallossofreadingabilityfollowingbraininjuryhaveprovidedim-portantevidenceconcerningbasicmechanismsinreadingandtheirbrainbases.Differenttypesofacquireddyslexiahavebeenaddressedusingconnectionistmodelsofspecificcomponentsofthetriangle(seeHinton&Shallice,1991;Plaut&Shallice,1993forapplicationstodeepdyslexia;Patterson,Seidenberg,&McClelland,1989andPlautetal.,1996forsurfacedyslexia,andHarm&Seidenberg,2001forphonologicaldyslexia).Itwouldbeaclearadvancetodeterminewhetherallofthesetypesofacquireddyslexiacanbehandledwithinasingle,unifiedmodel.
4.Extensionstotheexistingmodel
Therangeofphenomenathemodelcanaddressislim-itedbyvariousaspectsoftheimplementation.Atthetimewebegantheresearchitseemedimportanttolimititsscopesomewhatinordertomakeprogressinunderstandingbasiccomputationalmechanismsandinassessingthepotentialrelevanceoftheframeworktodivisionoflaborquestions.Givenwhathasbeenlearnedfromthepresentwork,aswellasadditionalinsightsaboutcomputationalmechanismsthathavebeenachievedsincewebeganseveralyearsago,itshouldbepossibletoaddressmanyoftheselimitationsinnext-generationmodels.
Orthographicrepresentation.Whereaswehavespentconsiderableeffortexaminingthenatureofsemanticandphonologicalrepresentations,particularlytheroleofat-tractorstructuresinprocessing,thenatureoforthographicknowledgehasnotbeenaddressedtothesamedegree.Clearlythereareissuesaboutletterrecognitionthatcanbeinvestigatedinthebroadercontextoftheoriesofvisualpro-
62
HARM,SEIDENBERG
cessingandobjectrecognition.Ofmoreimmediatecon-cerntousisthefactthatourmodels(datingbacktoSei-denberg&McClelland,1989)donotadequatelyaddressissuesaboutthestructureoforthographicknowledgeitself,i.e.,orthographicredundancy.Oneobviousstepwouldbetoimplementorthographyinamanneranalogoustowhathasbeendonewithsemanticsandphonology,usingdis-tributedrepresentationsoforthographicfeaturesandanat-tractorstructurecapableofencodingcomplexdependen-ciesamongletters.Wewouldexpectthiscomponentofthemodeltoexhibitpropertiesassociatedwiththe“visualwordform”area,aleftinferiortemporalregion(thefusiformgyrus)involvedintheprocessingofletterstrings(e.g.,Polk&Farah,2002).
Modelingresponselatencies.Thereareunresolvedis-suesaboutthemodelingofresponselatenciesinconnec-tionistandothertypesofcomputationalmodels.Ourmod-elscomputesemanticorphonologicalpatterns;therehavetobeadditionalassumptionsthatlinkthebehaviorofthemodeltotheperformanceoftaskssuchasnamingorse-manticdecisionandtotheresponsemeasuresthatarecol-lected(e.g.,namingordecisionlatenciesanderrors).Wehavenotasyetattemptedtomodelresponsetimesinarigorousway.Inpreviousresearchwefoundthatgeneralmeasuresofthemodel’sperformance(e.g.,meansummedsquarederror,“settlingtimes”)relatedcloselytogeneralmeasuresofhumanperformance(e.g.,meanlatenciesbycondition).Thesemeasuresdolesswellataccountingformoredetailedaspectsofperformancesuchasresponsela-tenciesforindividualwords(Spieler&Balota,1997).Therelativelypoorerfitatthismorerefinedgrainreflectslimi-tationsofboththemodelsandthehumandata,whichcon-tainconsiderablemeasurementerror(Seidenberg&Plaut,1998).Nonethelessitisclearthatmuchmorecouldbedoneintermsofmodelingresponselatencies.Settlingtimesareeasytocalculate(theysimplyreflectwhenactivationstopschangingsignificantlyinanattractornet)andtheycapturesomeaspectsofrelativedifficultybuttheyneedtobere-placedbyameasurewithbettertheoreticalmotivation.Set-tlingtimesreflecthowlongittakesthemodeltocompleteapattern,whereasmanytasksthatsubjectsperformcanbeinitiatedbeforetheprocessingoftheentirestimulushasbeencompleted.Naminglatencies,forexample,reflectthetimetoinitiateaspokenresponse,whichmayoccurwellbeforethesubjecthascompiledanarticulatorymo-torprogramfortheentireword(Kawamoto,Kello,Jones,&Barne,1998;Kawamoto,Kello,Higareda,&Vu,1999).Thus,whatisneededinthemodelisameasurerelatedtohowlongittakesforenoughofthepronunciationtohavebeencomputedtoinitiatearesponse,nottheamountoftimeittakestheentirepatterntosettle.Settlingtimesfortheonsetphoneme(s)oronsetandvowelmayprovideacloseraccountofnaminglatencies.Thesameissueariseswithrespecttoperformingtasksthatinvolvemeaning.Asub-
jectmaybeabletodecidethatSUITisanobjectandnotalivingthingwellbeforetheentiresemanticpatternhasbeencomputed.Inthiscase,thesettlingtimesforfeaturesthatidentifySUITasanobjectmayprovideabetterfittodecisionlatencies.Theseareunresolvedissues,however.Multisyllabicwords.Themodelwaslimitedtomono-syllabicwords,asinpreviousresearch(e.g.,Seidenberg&McClelland,1989).Multisyllabicwordsintroducemanyadditionalissues,e.g.,concerningtheassignmentofsyllabicstressinpronunciationandthedevelopmentofmorphologicalrepresentations(Seidenberg&Gonnerman,2000).Expandingthescopeofthemodeltoincludemul-tisyllabicwordswillentailalargermodelthattakeslongertotrainandgeneratesmorecomplexbehavior.Thelaborinvolvedindeveloping,training,andtestingamodelofthisscopeisconsiderableofcourse.Leavingthispracticalissueaside,themainobstacleistheoreticalnotcomputational.Howaremultisyllabicwordsread?Complexwordscouldbeprocessedaswholes(asinourcurrentmodel)orinparts(asseemstooccurwhenwordsarefixatedmorethanonce;Rayner,1998).Thepartscouldbesyllablesormorphemesorclumpsofadjacentlettersthatsometimescrossstructuralboundaries.Theseissueshavenotbeenresolvedbybehav-ioralresearch.Iftherewerebetterinformationabouthowcomplexwordsareprocessed,itcouldbeusedtoguidethedevelopmentofamodel.However,considerableadditionalworkisneededhereonbothcomputationalandbehavioralfronts.
Connectionstothebrain.Ourmodelwasbasedoncomputationalandbehavioralconsiderations;itmakesuseofsomedesignprinciplesthoughttoreflectgeneralproper-tiesofhowthebrainlearns,processes,andrepresentsinfor-mationbutisnotcloselytiedtofactsaboutthebrain.Intheperiodsincewebeganthisresearch,agrowingbodyofin-formationaboutlexicalprocessing,particularlyinreading,hasemergedfromtheuseofneuroimagingmethodologies.Giventhespecificityofthecomputationaltheoryandtheincreasingspecificityofneuroimagingmethodologiescon-cerningbothbraincircuitryandthetimecourseofprocess-ing,itshouldbepossibletoestablishcloserlinksbetweenthetwo.Threetypesofquestionscanbeaddressed:
First,arebasicpropertiesofthemodelconsistentwithevidenceconcerninghowreadingisaccomplishedbythebrain?Althoughwecannotyetcloselylinkthemodeltothebrain,therearesomeencouragingpreliminaryresults.BasedonfMRIstudiesofnormalanddyslexicindividu-als,Pughetal.(2000)havearguedthattherearetwoma-jorcircuitsinvolvedinnormalreading.One,termedthedorsalparietotemporalsystem,involvestheangulargyrus,supramarginalgyrus,andposteriorportionsofthesuperiortemporalgyrus.Theothercircuit,termedtheventraloccip-itotemporalsystem,involvesportionsofmiddletemporalgyrusandmiddleoccipitalgyrus.Pughetal.noteseveraldifferencesbetweenthetwosystems:thedorsalsystemde-
MULTICOMPONENTMODELOFREADING
63
velopsearlierinreadingacquisitionthantheventralsystem;thedorsalsystemismorestronglyimplicatedinphonolog-icalprocessing;andthedorsalsystemoperatesmoreslowlyinskilledreaders.Therearesomestrikingcorrespondencesbetweenthepropertiesofthesetwosystemsandthema-jorcomponentsofourmodel.Thedorsalsystemseemstocapturecharacteristicsoftheorthphonsemcomponentofthemodel:itdevelopsmorerapidlyandisresponsibleforphonologicalcodingbutultimatelyactivatessemanticsmoreslowly.Theventralsystem,liketheorthsempath-wayinthemodel,developsmoreslowly,isnotassociatedwithphonologicalprocessing,andultimatelyactivatesse-manticsmoreefficiently.Thus,thereareisomorphismsbe-tweenthebraincircuitsandmodelatleastatagenerallevel.Thesesuggestiveresultsraisemanyquestionsthatcanbeaddressedinfutureresearch.Wedonotknowifthetwocir-cuitsthatPughetal.haveidentifysolvethewordreadingprobleminthesamewayasourmodel.Forexample,inourmodel,thetwopathwayscooperativelyactivatesemantics;thePughetal.datadonotaddressthisissueandsoarealsoconsistentwithanindependentpathwaysaccount.
Asecondtypeofquestionis,howcanneuroimagingdatabeincorporatedinthemodelstomakethemmorebio-logicallyrealistic?Asanexample,thereisagrowingbodyofevidenceconcerningthebrain’srepresentationofdiffer-enttypesofsemanticinformation(seeMartin,2002,forre-view).Thereisconsiderableevidenceconcerningtherep-resentationofdifferentsemanticcategories(e.g.,animals,tools,bodyparts)anddifferenttypesofsemanticinforma-tion(e.g.,sensory,motoric,affective,factual,etc.).Theprinciplesgoverningtheorganizationofsemanticmemoryinthebrain,includingmanyofthebasictopographicfact,arestillunknown.Stillitisclearthatsemanticmemoryisnottheunorderedvectorofunitsinourmodel.Itisarea-sonablegoalforafuturegenerationmodeltoincorporateinformationabouttheorganizationofsemanticrepresenta-tionsasitbecomesavailable.Weexpectfuturemodelstoincorporateanincreasingnumberofsuchneurobiologicalconstraints.
Finally,thethirdtypeofquestioniswhetherourmod-elscaninformtheinvestigationofthebrainbasesofread-ing(andotheraspectsofcognition)usingneuroimaging.Aswehavealreadysuggested,themodelmakesspecificpredictionsaboutthetimecourseofprocessingfordiffer-enttypesofwords,whichsuggestsanimportantdirectionforneuroimagingtechniques,suchasMEG,whichcanpro-videtimecourseinformation.Similarly,understandinghowreadingisaccomplishedinthecomputationalmodelmayhelpininterpretingtheresultsofneuroimagingstudies,forexamplebysuggestingwhatfunctionsdifferentcircuitsareperforming.Thiswouldtakesuchstudiesbeyondlocaliza-tionquestionstoissuesabouthowthebrainaccomplishesatasksuchasreading.
Thus,weenvisionaproductivefeedbackloopbetweenmodeldevelopmentandneuroimaging,whereeachcan
constraintheotherandultimatelyconvergeonaninte-gratedcomputational-neurobiologicalmodelthatcapturesfactsaboutovertbehavior.
Conclusions
Wehavedescribedageneraltheoryofthecomputationofmeaningfromprintbasedonmotivatedprinciples,andpresentedanimplementedmodelthatinstantiatesthetheoryandrelateswelltobehavioraldata.Toourknowledgethisisthefirstlarge-scaleimplementedmodelthataddresseshowmeaningsarecomputedinamulticomponentprocess-ingsystem.Theresultsofthisworkarequitepromising,andsuggestawiderangeoffuturedirectionsforbehavioral,neuroimaging,andmodelingresearchonreading.
Inimplementingthemodelweattemptedtoaddresssomecontroversiesaboutbasicmechanismsinreadingatamoreexplicitcomputationallevelthaninprevioustheo-rizing.Themodelisnotlikelytobecorrectineveryde-tailandofcoursethegoalistoreplaceitwithsomethingbetter.Themodelservesanimportantfunctionbyraisingthebarintermsofthetheoreticalandmechanisticlevelsatwhichthesebehavioralphenomenacanbeengaged,andbyclarifyingtheinferencesthatcanbevalidlydrawnfromthebehavioralstudiesthathaveprovidedthemaindatatobeexplained.
Themodelwasconstructedfromtheoreticalcompo-nentssuchasdistributedrepresentationsandstatisticallearningproceduresthataregeneralratherthanspecifictoreadingandhavealreadybeenappliedtoabroadrangeofphenomena.Thenovelaspectsofthemodelconcerntheemergenceofthedivisionoflaborinamulticomponentsys-tem,aconceptthatisalsobeginningtobeappliedinotherdomains(Gordon&Dell,2003).Thus,thewaythatpeopleachieveanefficientsolutiontothecomputationofmean-ingproblemmayexemplifyhowmanycomplextasksaremastered.
References
Adams,M.(1990).Beginningtoread.Cambridge,MA:MITPress.
Andersen,R.(1999).Multimodalintegrationfortherepresenta-tionofspaceintheposteriorparietalcortex.InN.Burgess&K.Jeffery(Eds.),Thehippocampalandparietalfoundationsofspatialcognition(p.90-103).NewYork,NY:OxfordUni-versityPress.
Anderson,S.(1988).Morphologicaltheory.InF.Newmeyer(Ed.),Linguistics:TheCambridgesurvey,volume1:Linguis-tictheory:Foundations(p.146-191).Cambridge,England:CambridgeUniversityPress.
Balota,D.(1990).Theroleofmeaninginwordrecognition.InD.A.Balota,G.B.Floresd’Arcais,&K.Rayner(Eds.),Comprehensionprocessesinreading(p.9-32).Hillsdale,NJ:LawrenceErlbaum.
Balota,D.A.,&Chumbley,J.I.(1984).Arelexicaldecisionsagoodmeasureoflexicalaccess?Theroleofwordfrequency
64
HARM,SEIDENBERG
intheneglecteddecisionstage.JournalofExperimentalPsy-chology:HumanPerceptionandPerformance,10,340-357.Baron,J.(1973).Phonemicstagenotnecessaryforreading.Quar-terlyJournalofExperimentalPsychology,25,241-246.
Baron,J.,&Strawson,C.(1976).Useoforthographicandword-specificknowledgeinreadingwordsaloud.JournalofExper-imentalPsychology:HumanPerceptionandPerformance,4,207-214.Barto,A.G.(1985).Learningbystatisticalcooperationofself-interestedneuron-likecomputingelements.HumanNeurobi-ology,4,229-256.Bertelson,P.,&deGelder,B.(1989).Learningaboutreadingfromilliterates.InA.M.Galaburda(Ed.),Fromreadingtoneurons(p.1-23).Cambridge,MA:MITPress.
Besner,D.,Twilley,L.,McCann,R.,&Seergobin,K.(1990).Ontheconnectionbetweenconnectionismanddata:Areafewwordsnecessary?PsychologicalReview,97,432-446.Bishop,C.(1995).Trainingwithnoiseisequivalenttotikhonovregularization.NeuralComputation,7(1),108-116.Bradley,L.,&Bryant,P.(1983).Categorizingsoundsandlearningtoread-Acausalconnection.Nature,301,419-421.Browman,C.,&Goldstein,L.(1990).Representationandreal-ity:Physicalsystemsandphonologicalstructure.JournalofPhonetics,18,411-424.Bullinaria,J.(1996).Connectionistmodelsofreading:Incorpo-ratingsemantics.InProceedingsofthefirsteuropeanwork-shoponcognitivemodeling(p.224-229).Berlin:TechnischeUniversitatBerlin.Caplan,D.(Ed.).(1992).Language:Structure,processing,anddisorders.Cambridge,MA:MITPress.Carey,S.(1978).Thechildasword-learner.InM.Halle,J.Bres-nan,&G.Miller(Eds.),Linguistictheoryandpsychologicalreality.Cambridge,MIT:MITPress.Carr,T.H.,&Pollatsek,A.(1985).Recognizingprintedwords:Alookatcurrentmodels.InD.Besner,T.G.Waller,&G.E.MacKinnon(Eds.),Readingresearch:Advancesintheoryandpractice,Vol.5.NewYork,NY:AcademicPress.Chomsky,N.,&Halle,M.(1968).ThesoundpatternofEnglish.NewYork:Harper&Row.Cleermans,A.(1997).Principlesforimplicitlearning.InD.Berry(Ed.),Howimplicitisimplicitlearning?Oxford,UK:OxfordUniversityPress.Coltheart,M.(1978).Lexicalaccessinsimplereadingtasks.InG.Underwood(Ed.),Strategiesofinformationprocessing(p.151-216).NewYork:AcademicPress.
Coltheart,M.(1981).TheMRCPsycholinguisticDatabase.Quar-terlyJournalofExperimentalPsychology,33A,497-505.
Coltheart,M.(2000).Dualroutesfromprinttospeechanddualroutesfromprinttomeaning:Sometheoreticalissues.InA.Kennedy,R.Radach,D.Heller,&J.Pynte(Eds.),Read-ingasaperceptualprocess(p.475-490).Amsterdam:Oxford:Elsevier.
Coltheart,M.,Curtis,B.,Atkins,P.,&Haller,M.(1993).Modelsofreadingaloud:Dual-routeandparallel-distributed-processingapproaches.PsychologicalReview,100(4),589-608.Coltheart,M.,Davelaar,E.,Jonasson,K.,&Besner,D.(1977).Accesstotheinternallexicon.InS.Dornic(Ed.),Attention&performanceVI.Hillsdale,NJ:Erlbaum.
Coltheart,M.,Patterson,K.E.,&Marshall,J.C.(Eds.).(1980).Deepdyslexia.London:Routledge&KeganPaul.
Coltheart,M.,Rastle,K.,Perry,C.,Langdon,R.,&Ziegler,J.(2001).Drc:Adualroutecascadedmodelofvisualwordrecognitionandreadingaloud.PsychologicalReview,108(1),204-256.
Crowder,R.(1982).Thepsychologyofreading.NewYork,NY:OxfordUniversityPress.
Dell,G.S.(1986).Aspreading-activationtheoryofretrievalinsentenceproduction.PsychologicalReview,93(3),283-321.Ellis,A.W.,&Monaghan,J.(2002).ReplytoStrain,PattersonandSeidenberg(2002).JournalofExperimentalPsychology:Learning,MemoryandCognition,28(1),215-220.
Flesch,R.(1955).WhyJohnnycan’tread.NewYork:Harper&Brothers.
Foorman,B.R.,Perfetti,C.,Seidenberg,M.,Francis,D.,&Harm,M.(2001).Whatkindoftextisadecodabletext?Andwhatkindoftextisanauthentictext?(TalkpresentedattheAn-nualMeetingoftheAmericanEducationResearchAssocia-tion,Seattle,WA,April2001)
Forster,K.I.(1994).Computationalmodelingandelementaryprocessanalysisinvisualwordrecognition.Specialsection:Modelingvisualwordrecognition.JournalofExperimen-talPsychology:HumanPerceptionandPerformance,20(6),1292-1310.
Francis,W.N.,&Kuçera,H.(1982).FrequencyanalysisofEn-glishusage.Boston:Houghton-Mifflin.
Frost,R.(1998).Towardastrongphonologicaltheoryofvisualwordrecognition:Trueissuesandfalsetrials.PsychologicalBulletin,123(1),71-99.
Frost,R.,Katz,L.,&Bentin,S.(1987).Strategiesforvisualwordrecognitionandorthographicdepth:Amultilingualcompari-son.JournalofExperimentalPsychology:HumanPerceptionandPerformance,13,104-115.
Gainotti,G.(2000).Whatthelocusofbrainlesiontellsusaboutthenatureofthecognitivedeficitunderlyingcategory-specificdisorders:Areview.Cortex,36,539-559.
Gaskell,M.G.,&Marslen-Wilson,W.D.(1997).Integratingformandmeaning:Adistributedmodelofspeechperception.LanguageandCognitiveProcesses,12(5-6),613-656.
Gathercole,S.,&Baddeley,A.(1993).Phonologicalworkingmemory:Acriticalbuildingblockforreadingdevelopmentandvocabularyacquisition.EuropeanJournalofPsychologyofEducation,8(3),259-272.
Gernsbacher,M.A.(1984).Resolving20yearsofinconsistentinteractionsbetweenlexicalfamiliarityandorthography,con-creteness,andpolysemy.JournalofExperimentalPsychology:General,113(2),256-281.
Glushko,R.J.(1979).Theorganizationandactivationofortho-graphicknowledgeinreadingaloud.JournalofExperimentalPsychology:HumanPerceptionandPerformance,5(4),674-691.
Gordon,J.,&Dell,G.(2003).Learningtodividethelabor:Anac-countofdeficitsinlightandheavyverbproduction.CognitiveScience,27,1-40.
MULTICOMPONENTMODELOFREADING
65
Grainger,J.,&Jacobs,A.M.(1996).Orthographicprocessinginvisualwordrecognition:Amultipleread-outmodel.Psycho-logicalReview,103(3),518-565.
Harm,M.W.(1998).Divisionoflaborinacomputationalmodelofvisualwordrecognition.Unpublisheddoctoraldissertation,UniversityofSouthernCalifornia,LosAngeles,CA.
Harm,M.W.(2002).Buildinglargescaledistributedsemanticfeaturesetswithwordnet(Tech.Rep.No.PDP.CNS.02.01).115MI,4400FifthAvenue,PittsburghPA15213:CarnegieMellonUniversity,CNBC.
Harm,M.W.,McCandliss,B.D.,&Seidenberg,M.S.(2003).Modelingthesuccessesandfailuresofinterventionsfordis-abledreaders.ScientificStudiesofReading,7,155-182.
Harm,M.W.,&Seidenberg,M.S.(1997).Theroleofphonologyinreading:Aconnectionistinvestigation.(Paperpresentedatthe1997ComputationalPsycholinguisticsConference,Berke-ley,CA)
Harm,M.W.,&Seidenberg,M.S.(1999).Phonology,readingacquisition,anddyslexia:Insightsfromconnectionistmodels.PsychologicalReview,106(3),491-528.
Harm,M.W.,&Seidenberg,M.S.(2001).Arethereorthographicimpairmentsinphonologicaldyslexia?CognitiveNeuropsy-chology,18(1),71-92.
Hebb,D.O.(1949).Theorganizationofbehavior.NewYork:JohnWiley&Sons.
Henderson,L.(1982).Orthographyandwordrecognitioninread-ing.London:AcademicPress.
Hetherington,P.,&Seidenberg,M.S.(1989).Isthere“catas-trophicinterference”inconnectionistnetworks?InProceed-ingsofthe11thannualconferenceofthecognitivescienceso-ciety(p.26-33).Hillsdale,NJ:LawrenceErlbaum.
Hinton,G.E.,&Shallice,T.(1991).Lesioninganattractornet-work:Investigationsofacquireddyslexia.PsychologicalRe-view,98(1),74-95.
Hung,D.,&Tzeng,O.(1981).Orthographicvariationsandvisualinformation-processing.PsychologicalBulletin,90(3),377-414.
Ishai,A.,Ungerleider,L.,Martin,A.,&Haxby,J.(2000).Therepresentationofobjectsinthehumanoccipitalandtemporalcortex.JournalofCognitiveNeuroscience,12(Suppl2),35-51.Jared,D.(1997).Spelling-soundconsistencyaffectsthenamingofhigh-frequencywords.JournalofMemoryandLanguage,36(4),505-529.
Jared,D.,McRae,K.,&Seidenberg,M.S.(1990).Thebasisofconsistencyeffectsinwordnaming.JournalofMemoryandLanguage,29(6),687-715.
Jared,D.,&Seidenberg,M.S.(1991).Doeswordidentifica-tionproceedfromspellingtosoundtomeaning?JournalofExperimentalPsychology:General,120(4),358-394.
Joanisse,M.F.,&Seidenberg,M.S.(1999).Impairmentsinverbmorphologyafterbraininjury:Aconnectionistmodel.Pro-ceedingsoftheNationalAcademyofScience,96(13),7592-7597.
Jorm,A.F.,&Share,D.L.(1983).Phonologicalrecodingandreadingacquisition.AppliedPsycholinguistics,4,103-147.Jusczyk,P.W.(1997).Thediscoveryofspokenlanguage.Cam-bridge,MA:MITPress.Kawamoto,A.,Kello,C.,Higareda,I.,&Vu,J.(1999).Paral-lelprocessingandinitialphonemecriterioninnamingwords:Evidencefromfrequencyeffectsononsetandrimeduration.JournalofExperimentalPsychology:Learning,MemoryandCognition,25(2),362-381.
Kawamoto,A.H.,Kello,C.T.,Jones,R.,&Barne,K.(1998).Initialphonemeversuswhole-wordcriteriontoinitiatepronun-ciation:Evidencebasedonresponselatencyandinitialpho-nemeduration.JournalofExperimentalPsychology:Learn-ing,MemoryandCognition,24(4),862-885.
Kelly,M.H.(1992).Usingsoundtosolvesyntacticproblems:Theroleofphonologyingrammaticalcategoryassignments.PsychologicalReview,99(2),349-364.
Kuçera,H.,&Francis,W.N.(1967).Computationalanalysisofpresent-dayAmericanEnglish.Providence,RI:BrownUniver-sityPress.
LaBerge,D.L.,&Samuels,J.(1974).Towardatheoryofau-tomaticwordprocessinginreading.CognitivePsychology,6,293-323.
Landauer,T.K.,&Dumais,S.T.(1997).AsolutiontoPlato’sproblem:Thelatentsemanticanalysistheoryofacquisition,induction,andrepresentationofknowledge.PsychologicalRe-view,104(2),211-240.
Lesch,M.F.,&Pollatsek,A.(1993).Automaticaccessofseman-ticinformationbyphonologicalcodesinvisualwordrecogni-tion.JournalofExperimentalPsychology:Learning,MemoryandCognition,19(2),285-294.
Liberman,I.Y.,&Shankweiler,D.(1985).Phonologyandtheproblemsoflearningtoreadandwrite.RemedialandSpecialEducation,6(6),8-17.
Liberman,I.Y.,Shankweiler,D.,&Liberman,A.M.(1989).Thealphabeticprincipleandlearningtoread.InD.Shankweiler&I.Y.Liberman(Eds.),Phonologyandreadingdisability:Solv-ingthereadingpuzzle(p.1-33).AnnArbor,MI:UniversityofMichiganPress.
Locke,J.L.(1995).Developmentofthecapacityforspokenlan-guage.InP.Fletcher&B.MacWhinney(Eds.),Thehandbookofchildlanguage(p.278-302).Oxford:Blackwell.
Lukatela,G.,Turvey,M.,Feldman,L.,Carello,C.,&Katz,L.(1989).Alphabeticpriminginbi-alphabeticwordperception.JournalofMemoryandLanguage,28,237-254.
Lukatela,G.,&Turvey,M.T.(1994a).Visuallexicalaccessisinitiallyphonological:2.Evidencefromphonologicalprimingbyhomophonesandpseudohomophones.JournalofExperi-mentalPsychology:General,123(4),331-353.
Lukatela,G.,&Turvey,M.T.(1994b).Visuallexicalaccessisinitiallyphonological:I.Evidencefromassociativeprimingbywords,homophones,andpseudohomophones.JournalofExperimentalPsychology:General,123(2),107-128.
Lundberg,I.,Olofsson,A.,&Wall,S.(1980).Readingandspellingskillsinthefirstschoolyearspredictedfromphone-micawarenessskillsinkindergarten.ScandinavianJournalofPsychology,21,159-173.
MacDonald,M.C.(1993).Theinteractionoflexicalandsyntacticambiguity.JournalofMemoryandLanguage,32,692-715.Marchand,H.(1969).Thecategoriesandtypesofpresent-dayEn-glishword-formation:Asynchronic-diachronicapproach(2nded.).Munchen:Beck.
66
HARM,SEIDENBERG
Marcus,M.,Santorini,B.,&Marcinkiewicz,M.A.(1993).Build-ingalargeannotatedcorpusofEnglish:ThePennTreebank.ComputationalLinguistics,19,313-330.
Marshall,J.C.,&Newcombe,F.(1973).Patternsofparalexia:Apsycholinguisticapproach.JournalofPsycholinguisticRe-search,2,175-199.
Martin,A.(2002).Functionalneuroimagingofsemanticmemory.InR.Cabeza&A.Kingstone(Eds.),Handbookoffunctionalneuroimagingofcognition(p.153-186).Cambridge,MA:MITPress.
McCann,R.S.,&Besner,D.(1987).Readingpseudohomo-phones:Implicationsformodelsofpronunciationassemblyandthelocusofword-frequencyeffectsinnaming.JournalofExperimentalPsychology:HumanPerceptionandPerfor-mance,13(1),14-24.
McClelland,J.L.,McNaughton,B.L.,&O’Reilly,R.C.(1995).Whytherearecomplementarylearningsystemsinthehip-pocampusandneocortex:Insightsfromthesuccessesandfail-uresofconnectionistmodelsoflearningandmemory.Psycho-logicalReview,102,419-457.
McClelland,J.L.,&Rumelhart,D.E.(1981).Aninteractiveacti-vationmodelofcontexteffectsinletterperception:Part1.Anaccountofbasicfindings.PsychologicalReview,88,375-407.McCloskey,M.,&Cohen,N.J.(1989).Catastrophicinterferenceinconnectionistnetworks:Thesequentiallearningproblem.InG.H.Bower(Ed.),Thepsychologyoflearningandmotivation(Vol.23).NewYork,NY:AcademicPress.
McCusker,L.,Hillinger,M.,&Bias,R.(1981).Phonologicalrecodingandreading.PsychologicalBulletin,89(2),217-245.McLeod,P.,Plunkett,K.,&Rolls,E.T.(1998).Introductiontoconnectionistmodellingofcognitiveprocesses.Oxford,UK:OxfordUniversityPress.
McRae,K.,&Boisvert,S.(1998).Automaticsemanticsimilar-itypriming.JournalofExperimentalPsychology:Learning,MemoryandCognition,24(3),558-572.
McRae,K.,deSa,V.R.,&Seidenberg,M.S.(1997).Onthenatureandscopeoffeaturalrepresentationsofwordmeaning.JournalofExperimentalPsychology:General,126(2),99-130.Meyer,D.E.,Schvaneveldt,R.W.,&Ruddy,M.G.(1974).Func-tionsofgraphemicandphonemiccodesinvisualwordrecog-nition.MemoryandCognition,2,309-321.
Miller,G.A.(1990).WordNet:Anon-linelexicaldatabase.In-ternationalJournalofLexicography,3,235-312.
Morton,J.(1969).Theinteractionofinformationinwordrecog-nition.PsychologicalReview,76,165-178.NationalReadingPanel.(2000).Teachingchildrentoread:Anevidence-basedassessmentofthescientificre-searchliteratureonreadinganditsimplicationsforread-inginstruction.(NationalInstituteofChildHealthandHumanDevelopment,Washington,D.C.Availblefromwww.nationalreadingpanel.org)
O’Reilly,R.C.,&Munakata,Y.(2000).Computationalexplo-rationsincognitiveneuroscience:Understandingthemindbysimulatingthebrain.Cambridge,MA:MITPress.
Paap,K.,Newsome,S.,McDonald,J.,&Schvanevelt,R.W.(1982).Anactivationverificationmodelforletterandwordrecogntion-theword-superiorityeffect.PsychologicalRe-view,89(5),573-594.Paap,K.R.,&Noel,R.W.(1991).Dualroutemodelsofprinttosound:Stillagoodhorserace.PsychologicalResearch,53,13-24.
Patterson,K.,&Hodges,J.R.(1992).Deteriorationofwordmeaning:Implicationsforreading.Neuropsychologia,30(12),1025-1040.
Patterson,K.,LambonRalph,M.,Bird,H.,Hodges,J.,&Mc-Clelland,J.(2000).Normalandimpairedprocessinginquasi-regulardomainsoflanguage:ThecaseofEnglishpast-tenseverbs.InProceedingsofthe6thinternationalconferenceonspokenlanguageprocessing(Vol.2,p.15-19).Beijing,China.Patterson,K.,LambonRalph,M.A.,Hodges,J.R.,&McClel-land,J.L.(2001).Deficitsinirregularpast-tenseverbmor-phologyassociatedwithdegradedsemanticknowledge.Neu-ropsychologia,39,709-724.
Patterson,K.,Seidenberg,M.,&McClelland,J.(1989).Dyslexiainadistributed,developmentalmodelofwordrecognition.InR.Morris(Ed.),Paralleldistributedprocessing.OxfordUni-versityPress.
Patterson,K.,Suzuki,T.,&Wydell,T.N.(1996).InterpretingacaseofJapanesephonologicalalexia:Thekeyisinphonology.CognitiveNeuropsychology,13,803-822.
Patterson,K.E.,Marshall,J.C.,&Coltheart,M.(Eds.).(1985).Surfacedyslexia:Neuropsychologicalandcognitivestudiesofphonologicalreading.London:Erlbaum.
Pearlmutter,B.A.(1989).Learningstatespacetrajectoriesinre-currentneuralnetworks.NeuralComputation,1(2),263-269.Pearlmutter,B.A.(1995).Gradientcalculationsfordynamicrecurrentneuralnetworks:Asurvey.IEEETransactionsonNeuralNetworks,6(5),1212-1228.
Perfetti,C.,&McCutchen,D.(1983).Speechprocessesinread-ing.InN.Lass(Ed.),Advancesinspeechandlanguage.NewYork:AcademicPress.
Perfetti,C.A.,&Bell,L.(1991).Phonemicactivationduringthefirst40msofwordidentification:Evidencefrombackwardmaskingandpriming.JournalofMemoryandLanguage,30,473-485.
Perfetti,C.A.,Bell,L.,&Delaney,S.(1988).Automaticpho-neticactivationinsilentwordreading:Evidencefromback-wardmasking.JournalofMemoryandLanguage,27,59-70.Pinker,S.(2000).Wordsandrules:Theingredientsoflanguage.NewYork,NY:HarperCollins.
Plaut,D.C.(1997).Structureandfunctioninthelexicalsystem:Insightsfromdistributedmodelsofwordreadingandlexicaldecision.LanguageandCognitiveProcesses,12,765-805.Plaut,D.C.,&Booth,J.R.(2000).Individualanddevelopmentaldifferencesinsemanticpriming:Empiricalandcomputationalsupportforasingle-mechanismaccountoflexicalprocessing.PsychologicalReview,107(4),786-823.
Plaut,D.C.,&Kello,C.T.(1999).Theinterplayofspeechcomprehensionandproductioninphonologicaldevelopment:Aforwardmodelingapproach.InB.MacWhinney(Ed.),Theemergenceoflanguage(p.381-415).Mahweh,NJ:Erlbaum.Plaut,D.C.,McClelland,J.L.,Seidenberg,M.,&Patterson,K.E.(1996).Understandingnormalandimpairedwordreading:Computationalprinciplesinquasi-regulardomains.Psycho-logicalReview,103(1),56-115.
MULTICOMPONENTMODELOFREADING
67
Plaut,D.C.,&Shallice,T.(1991).Effectsofwordabstractnessinaconnectionistmodelofdeepdyslexia.InProceedingsofthethirteenthannualconferenceofthecognitivesciencesociety(p.73-78).Hillsdale,NJ:LawrenceErlbaum.
Plaut,D.C.,&Shallice,T.(1993).Deepdyslexia:Acasestudyofconnectionistneuropsychology.CognitiveNeuropsychology,10(5),377-500.
Polk,T.A.,&Farah,M.(2002).FunctionalfMRIevidenceforanabstract,non-perceptualword-formarea.JournalofExper-imentalPsychology:General,131,65-72.
Pugh,K.,Mencl,W.,Jenner,A.,Katz,L.,Lee,J.,Shaywitz,S.,&Shaywitz,B.(2000).Functionalneuroimagingstudiesofreadingandreadingdisability(developmentaldyslexia).Men-talRetardationandDevelopmentalDisabilitiesReview,6(3),207-213.
Rayner,K.(1998).Eyemovementsnreadingandinformationpro-cessing:20yearsofresearch.Psychologicalbulletin,124(3),372-422.
Rayner,K.,&Duffy,S.A.(1986).Lexicalcomplexityandfix-ationtimesinreading:Effectsofwordfrequency,verbcom-plexity,andlexicalambiguity.Memory&Cognition,14(3),191-201.
Rayner,K.,Foorman,B.,Perfetti,C.,Pesetsky,D.,&Seidenberg,M.(2001).Howpsychologicalscienceinformstheteachingofreading.PsychologicalScienceinthePublicInterest,2(2),31-74.
Rayner,K.,&Pollatsek,A.(1989).Thepsychologyofreading.EnglewoodCliffs,NJ:Prentice-Hall.
Rolls,E.,Critchley,H.,&Treves,A.(1996).Representationofolfactoryinformationintheprimateorbitofrontalcortex.Jour-nalofNeurophysiology,75(5),1982-1996.
Rubenstein,H.,Lewis,S.S.,&Rubenstein,M.A.(1971).Evi-denceforphonemicrecodinginvisualwordrecognition.Jour-nalofVerbalLearningandVerbalBehavior,10(6),645-657.Rumelhart,D.,Hinton,G.,&Williams,R.(1986).Learningin-ternalrepresentationsbyerrorpropagation.InD.Rumelhart&J.McClelland(Eds.),Paralleldistributedprocessing,vol.1.Cambridge,MA:MITPress.
Rumelhart,D.E.,McClelland,J.L.,&PDPResearchGroupthe(Eds.).(1986).Paralleldistributedprocessing:Explorationsinthemicrostructureofcognition.Volume1:Foundations.Cam-bridge,MA:MITPress.
Saffran,J.R.,Aslin,R.N.,&Newport,E.L.(1996).Statisticallearningby8-month-oldinfants.Science,274(5294).
Seidenberg,M.,&MacDonald,M.(1999).Aprobabilisticcon-straintsapproachtolanguageacquisitionandprocessing.Cog-nitiveScience,23(4),569-588.
Seidenberg,M.,Zevin,J.,&Harm,M.(2002).DRCdoesn’treadcorrectly.(PaperpresentedattheannualmeetingofthePsy-chonomicSocietyofAmerica,KansasCity,November2002)Seidenberg,M.S.(1985).Thetimecourseofinformationactiva-tionandutilizationinvisualwordrecognition.InD.Besner,T.G.Waller,&E.M.MacKinnon(Eds.),Readingresearch:Advancesintheoryandpractice(p.199-252).NewYork:Aca-demicPress.
Seidenberg,M.S.(1987).Sublexicalstructuresinvisualwordrecognition:Accessunitsororthographicredundancy.In
M.Coltheart(Ed.),AttentionandPerformanceXII:Thepsy-chologyofreading.Hillsdale,NJ:LawrenceErlbaum.
Seidenberg,M.S.(1992a).Beyondorthographicdepth:Equi-tabledivisionoflabor.InR.Frost&L.Katz(Eds.),Orthogra-phy,phonology,morphologyandmeaning.Amsterdam:North-Holland.Seidenberg,M.S.(1992b).Dyslexiainacomputationalmodelofwordrecognitioninreading.InP.Gough,L.Ehri,&R.Treiman(Eds.),Readingacquisition(p.243-274).Hills-dale,NJ:Erlbaum.
Seidenberg,M.S.(1993).Connectionistmodelsandcognitivetheory.PsychologicalScience,4(4),228-235.
Seidenberg,M.S.(1995).Visualwordrecognition:Anoverview.InP.Eimas&J.L.Miller(Eds.),Handbookofperceptionandcognition:Language.NewYork:AcademicPress.Seidenberg,M.S.,&Gonnerman,L.M.(2000).Explainingderivationalmorphologyastheconvergenceofcodes.TrendsinCognitiveScience,4,353-361.Seidenberg,M.S.,&McClelland,J.L.(1989).Adistributed,developmentalmodelofwordrecognitionandnaming.Psy-chologicalReview,96(4),523-568.
Seidenberg,M.S.,&Plaut,D.C.(1998).Evaluatingwordreadingmodelsattheitemlevel:Matchingthegrainoftheoryanddata.PsychologicalScience,9(3),234-237.Seidenberg,M.S.,Plaut,D.C.,Petersen,A.S.,McClelland,J.L.,&McRae,K.(1994).Nonwordpronunciationandmodelsofwordrecognition.JournalofExperimentalPsychology:Hu-manPerceptionandPerformance,20(6),1177-1196.
Seidenberg,M.S.,&Tanenhaus,M.K.(1979).Orthographiceffectsonrhymemonitoring.JournalofExperimentalPsy-chology:HumanLearningandMemory,5(6),546-554.Seidenberg,M.S.,&Waters,G.S.(1989).Wordrecognitionandnaming:Amegastudy[Abstract30].BulletinofthePsycho-nomicSociety,27,489.Seidenberg,M.S.,Waters,G.S.,Barnes,M.A.,&Tanenhaus,M.K.(1984).Whendoesirregularspellingorpronunciationinfluencewordrecognition?JournalofVerbalLearningandVerbalBehavior,23(3),383-404.Seidenberg,M.S.,Waters,G.S.,Sanders,M.,&Langer,P.(1984).Pre-andpostlexicallociofcontextualeffectsonwordrecognition.MemoryandCognition,12(4),315-328.Shu,H.,Chen,X.,Anderson,R.C.,Wu,N.,&Xuan,Y.(2003).PropertiesofschoolChinese:Implicationsforlearningtoread.ChildDevelopment,74,27-47.
Simpson,G.B.(1994).Contextandtheprocessingofambiguouswords.InM.A.Gernsbacher(Ed.),Handbookofpsycholin-guistics(p.359-374).SanDiego,CA:AcademicPress.Smith,F.(1971).Understandingreading.NewYork:Holt,Rine-hartandWinston.Smith,F.(1973a).Thepoliticsofignorance.InR.Winklejohann(Ed.),Thepoliticsofreading:Point-Counterpoint.Newark,Delaware:IRA/ERIC.Smith,F.(1973b).Psycholinguisticsandreading.NewYork:Holt,RinehartandWinston.Snowling,M.J.(1991).Developmentalreadingdisorders.JournalofChildPsychologyandPsychiatry,32,49-77.
68
HARM,SEIDENBERG
Spencer,A.(Ed.).(1991).Morphologicaltheory.London:Black-well.
Spieler,D.H.,&Balota,D.A.(1997).Bringingcomputationalmodelsofwordnamingdowntotheitemlevel.PsychologicalScience,8(6),411-416.
Strain,E.,&Herdman,C.M.(1999).Imageabilityeffectsinwordnaming:Anindividualdifferencesanalysis.CanadianJournalofExperimentalPsychology,53(4),347-359.
Strain,E.,Patterson,K.,&Seidenberg,M.S.(1995).Semanticeffectsinsingle-wordnaming.JournalofExperimentalPsy-chology:Learning,MemoryandCognition,21,1140-1154.Sutton,R.S.(1988).Learningtopredictbythemethodoftempo-raldiferences.MachineLearning,3,9-44.
Swinney,D.(1979).Lexicalaccessduringsentencecomprehen-sion:Reconsiderationofcontexteffects.JournalofVerbalLearningandVerbalBehavior,18,645-660.
Tanenhaus,M.,Leiman,J.,&Seidenberg,M.(1979).Evidenceformultiplestagesintheprocessingofambiguouswordsinsyntacticcontexts.JournalofVerbalLearningandVerbalBe-havior,18,427-440.
Taraban,R.,&McClelland,J.L.(1987).Conspiracyeffectsinwordpronunciation.JournalofMemory&Language,26(6),608-631.
Treiman,R.,Kessler,B.,&Bick,S.(inpress).Influenceofconso-nantalcontextonthepronunciationofvowels:Acomparisonofhumanreadersandcomputationalmodels.(Inpress,Cog-nition)
VanOrden,G.C.(1987).AROWSisaROSE:Spelling,soundandreading.MemoryandCognition,15(3),181-198.
VanOrden,G.C.,Johnston,J.C.,&Hale,B.L.(1988).Wordidentificationinreadingproceedsfromthespellingtosoundtomeaning.JournalofExperimentalPsychology:Language,MemoryandCognition,14,371-386.
VanOrden,G.C.,Pennington,B.F.,&Stone,G.O.(1990).Wordidentificationinreadingandthepromiseofsubsymbolicpsycholinguistics.PsychologicalReview,97(4),488-522.Vihman,M.M.(1996).Phonologicaldevelopment:Theoriginsoflanguageinthechild.Oxford:Blackwell.
Wagner,R.K.,&Torgesen,J.K.(1987).Thenatureofphonolog-icalprocessinganditscausalroleintheacquisitionofreadingskills.PsychologicalBulletin,101,192-212.
Williams,R.J.,&Peng,J.(1990).Anefficientgradient-basedalgorithmforon-linetrainingofrecurrentnetworktrajectories.NeuralComputation,2,490-501.
Zeno,S.(Ed.).(1995).Theeducator’swordfrequencyguide.Brewster,NJ:TouchstoneAppliedScienceAssociates.
Zevin,J.D.,&Seidenberg,M.S.(2002).Ageofacquisitioneffectsinreadingandothertasks.JournalofMemoryandLanguage,47,1-29.
Zipf,G.K.(1935).Thepsycho-biologyoflanguage:Anintroduc-tiontodynamicphilology.Boston,MA:HoughtonMifflin.Zorzi,M.,Houghton,G.,&Butterworth,B.(1998).Tworoutesoroneinreadingaloud?Aconnectionistdual-processmodel.JournalofExperimentalPsychology:HumanPerceptionandPerformance,24(4),1131-1161.
因篇幅问题不能全部显示，请点此查看更多更全内容