Meaning in Acquisition Semantic Structure, Lexical Organization, and Crosslinguistic Variation

2��

9MeaninginAcquisition

Semantic Structure, Lexical Organization, and

Crosslinguistic Variation

PINGLIThe Pennsylvania State University

introDuction

I nthischapter,Iexaminethreeproblems:theacquisitionoftenseandaspect,theacquisitionofcryptotypes,andthedevelopmentoflexicalstructure.Thecentralissueinallthreeproblemsishowthechilddiscoverswordmeanings

inalexicalsystem,andwhatmechanismsareatworkintheprocessofthisdiscov-ery.Ineachcase,theproblemdomaininvolvesthelearningofsemanticstructuresthatareimportantfortheuseofgrammaticalmorphologyorlexicalcategories.Inexaminingthesethreecases,Iarguethatstructuredsemanticrepresentationsofthelexiconcanemergeasanaturaloutcomeofthemeaning-formandthemean-ing-meaningmappingsinlanguageacquisition.

Twomajorperspectivesinformtheexaminationofchildren’sdevelopmentofthese structures (and have been key assumptions underlying Bowerman’s ownwork).First, onecannot lookat theacquisitionof agiven setof forms in isola-tion.Inthestudyoftheacquisitionofmorphology,forexample,itisimportanttoexaminenotonlythemorphologicaldevicesperse,butalsothewords(particularlyverbs)withwhichthesedevicesareused(Bowerman,1982,1983).Forexample,theacquisitionoftheprefixun-needstobeconsideredcloselywiththeverbsthatcantakeun-.InWhorf’s(1956)view,un-marksacryptotype,acovertsemanticcategorythatcanonlybeimplicitlydefinedbythedevicesitgoeswith.Thus,theunderstandingofun-inchildlanguagecannotbecompletewithoutaconsideration

routes to Language2��

of thecryptotypemeanings that arecharacteristicof theun-ableverbs.By thesametoken,theacquisitionoftense-aspectsuffixessuchas-ingand-edneedstobeconsideredtogetherwiththetypesofverbs(atelic,telic,result,etc.)towhichthesuffixesareattached.Intheempiricalliteraturethereisampleevidencethatchildrenpayattentiontotheinherentmeaningsoftheverbswhentheylearntousetense-aspectsuffixes(Bloom,Lifter,&Hafitz,1980;Brown,1973;Slobin,1985).InLiandBowerman(1998)andLiandShirai(2000),weexaminedspecificallythe interactionbetweengrammatical aspect (expressedbymorphologicalmark-ers)andlexicalaspect(expressedininherentmeaningsofverbs)inbothfirstandsecondlanguageacquisition,inasystematicattempttoconnectmorphologyandlexicalsemantics.

AsecondimportantperspectivechampionedbyBowermanisthatinformulat-ingandtestingtheoriesoflanguageacquisition,itisimportanttoexamineacquisi-tiondatafromnotonlyonelanguage(usuallyEnglishbydefault),butalsofromotherlanguages.Crosslinguisticevidenceiscrucialinmanycasesbecauseofthelanguage-specificpropertiesassociatedwithindividuallanguages.Often,agivenhypothesisappearstobeperfectforonelanguage,butitmayturnouttobeinac-curateorincompleteuponacloselookatdatafromotherlanguages.Forexample,Englishusestheprepositionsonandintoexpressspatial locationsofobjectsinattachmentandfittingsituations.InDutch,theorientationofattachmentmatters,suchthatverticalversushorizontalattachmentsinvolvetheuseofdifferentprepo-sitions;inKorean,thedegreeoffitmatters,suchthattight-fitversusloose-supportsituationsinvolvetheuseofdifferentlocativeverbs.Thus,howEnglish-speakingchildrenlearnonandinwillnotbesufficientlyinformativeastohowDutchandKoreanchildren learnthecorrespondingdevices in their languages.Acrosslin-guisticperspectivecanbecrucial in illuminatingtheextent towhich language-specificinputmayplayaroleinshapingchildren’searlygrammar(e.g.,Bowerman,1985,1989,1996).

This chapter summarizes our research that aims at connecting morphologyand lexical semantics and at connecting acquisition theories and crosslinguisticdata.Thediscussion isorganizedby the three topics thatImentioned,namely,theacquisitionoftenseandaspect,theacquisitionofcryptotypes,andthedevel-opment of lexical structure. Through careful analysis of crosslinguistic data, aswellasconnectionistmodelingofacquisition,Ishowthatthesemanticstructureunderlyingeachofthesedomainsemergesfromtheassociationsbetweenformsandmeaningspresentintheinput.

the acquisition of tense anD asPect

Empirical Observations

Theexpressionoftimeisoneofthecentralconceptualdomainsoflanguage,andtheacquisitionof theability to talkabout timethroughtheuseof tense-aspectmarkers isoneof theearliest tasks in languageacquisition. In the last30yearsresearchershavedevotedmuchattention to theacquisitionof tenseandaspect

meaning in acquisition 2��

invariouslanguages(forgeneralreviews,seeLi&Shirai,2000;Shirai,Slobin&Weist,1998;Slobin,1985;Weist,1986).Anearlyobservationofchildren’sacquisi-tionoftense-aspectmarkerscamefromRogerBrown(1973).Browndocumentedtwointerestingpatterns:First,theearliestgrammaticaldeviceinchildren’sspeech,theprogressiveaspectmarker-ing,appearsvirtuallyalwaystobeusedcorrectly.In particular, children never use -ing incorrectly with state verbs; for example,theydonotproduceovergeneralizationslikeknowingorwanting.Second,Eng-lish-speakingchildrenfirstusepast-tense formswithonly a small, semanticallycoherentsetofverbs,includingdropped,slipped,crashed,andbroke,verbsthatindicateevents thathappen instantaneouslybut lead toclearendresults.SomeyearsafterBrown’sobservations,Bloom,Lifter,andHafitz(1980)providedfur-ther evidence that confirmed Brown’s analyses (especially the second observa-tion). They found that the inflections used by young English-speaking childrencorrelatedwiththesemantictypesofverbs:-ingoccurredalmostexclusivelywithverbssuchasplay,ride,andwrite(durative,nonresultative),whereaspast-tenseformsoccurredpredominantlywithverbssuchasfind,fall,andbreak(punctual,resultative).Together,BrownandBloometal.’sdatasuggestedapictureofearly“undergeneralization”intheacquisitionofinflectionalmorphology:Ratherthanusingtense-aspectmarkerswithalltypesofverbs,asadultsdo,childrenusethemmorerestrictively.Obviously,thismorerestrictiveuseofgrammaticalmorphemesisassociatedwiththeinherentsemanticdifferencesoftheverbstowhichthepro-gressiveandpast-tenseformsareattached.

Thestrongassociationsbetweentense-aspectmorphemesandverbsemanticsobserved by Brown and Bloom et al. have generated a considerable amount ofresearch.Manystudieshaveexaminedtheacquisitionoftenseandaspectinotherlanguages,includingChinese(Erbaugh,1982;Li,1990;Li&Bowerman,1998),French (Bronckart & Sinclair, 1973), Italian (Antinucci & Miller, 1976), Polish(Weist, Wysocka, Witkowska-Stadnik, Buczowska, & Konieczna, 1984), Turkish(Aksu,1978;Aksu-Koç&Slobin,1985),andotherlanguages(seeLi&Shirai,2000forareviewofrelevantcrosslinguisticdata).Ingeneral,datafromthesestudiesarerobustandconsistentwithBrownandBloometal.’sobservations,indicatingthattheuseofimperfectiveorprogressiveaspectmorphologyinchildren’sspeechisfirstassociatedwithatelic,activityverbs,whereasthatofperfectiveaspect/pasttensemorphemesisassociatedwithtelic,resultativeverbs.1

Subsequentdiscussionsontheseempiricaldatahavestimulatedintensedebatesonhowchildrenacquirelexicalsemanticsandgrammaticalmorphologyandhavemotivatedtheoreticalaccountsofchildren’ssemanticandmorphologicaldevelop-mentingeneral(seeLi&Shirai,2000,foranoverview).HereIdiscusstwomajorcontrastingproposalsandthenattempttoexplainthedatafromacrosslinguisticdevelopmentalperspective.

1 Notallacquisitionresearchersagreethattheseassociationpatternsexistcrosslinguistically.Forexample,WeistandhiscolleaguesarguedagainsttheproposalofBloometal.(1980).TheyshowedthatPolishchildrenareabletounderstandandproducethebasiccontrastbetweenperfectiveandimperfectiveaspectasearlyas2;6(Weistetal.,1984).SeeLiandShirai(2000,pp.40–47)foradiscussionoftherelevantdebate.

routes to Language2�0

Contrasting Perspectives

The major divide in perspectives on the acquisition of tense and aspect fallsbetween formalist-nativist and functionalist-cognitivist approaches. The formal-ist perspective is strongly associated with a nativist view in the tense-aspectacquisitionliterature.Initsstrongestform,Bickerton(1981,1984)proposedthelanguagebioprogramhypothesis,accordingtowhichspecificsemanticcategoriesand concepts are biologically preprogrammed for the human language learner,andthesecategoriesorconceptswillnaturallyunfoldintheprocessoflanguageacquisition.Because thecategoriesandconceptsarehard-wiredaheadof time,thechildsimplyneedstodiscoverhowtheyareinstantiatedinspecificformsinthelanguagetobelearned.Twoimportantinnatedistinctionsinthedomainoftenseandaspectarebetweenstateandprocessandbetweenpunctualandnonpunctualcategories.Giventheinnatenatureofthesedistinctions,accordingtoBickerton,earlyoninlanguagedevelopmentstateswillbemarkeddifferentlyfromprocesses,andpunctualsituationswillbemarkeddifferentlyfromnonpunctualsituations,probablybytheuseofdifferenttense-aspectmarkers.2

BickertonfirstsupportedhishypothesiswithevidencefromCreolegrammars,arguing that in theabsenceof relevant input (pidgins, thepredecessorsofCre-oles, do not have tense-aspect markers), first-generation Creole speakers inventtense-aspectsystemstomarkthebioprogrammeddistinctions.Drawinginaddi-tion on child language data, he argued that children first use the tense-aspectmarkers of their language to mark the distinctions between state and processand between punctual and nonpunctual. For example, Bickerton used Brown’s(1973)observationinsupportofaninnatespecificationofthedistinctionbetweenprocess and state: young English-speaking children never overgeneralize theprogressivemarker -ing to stateverbsbecause theyare sensitive to thebiopro-grammedstate–processdistinction.Similarly,Bickertonarguedfortheexistenceofthepunctual–nonpunctualdistinctioninthebioprogram,onthebasisofinter-preting theobservationsbyAntinucciandMiller (1976),BronckartandSinclair(1973), and Aksu-Koç and Slobin (1985; originally Slobin & Aksu, 1980) thatyoungchildrenuse thepastorperfectivemorphemes tomarkpunctualevents.Thus,althoughBickerton’slanguagebioprogramhypothesisoriginatedfromstud-iesofCreolelanguages,manyofitsargumentsrestoninterpretationsofdatainchildren’s acquisitionof tenseandaspect.Bickertonconsideredearly child lan-guagesandCreolestobeidealcasesforobservingthebioprogram,becausetheinnatedistinctionsof thebioprogramare realized in these caseswithoutbeing

2 A somewhatdifferent,but relatedview is advocatedbySlobin (1985).Slobinproposed thebasicchildgrammar,whichcontainsaprestructured“semanticspace”withuniversalseman-ticnotionsorcategories.Thesesemanticcategoriescanactstrongly toattract themorpho-logicalmappingfromtheinputlanguage.ResultandProcessaretwosuchsemanticcategoriesthathavetodowithchildren’sacquisitionoftense-aspectmorphology.However,becausetheissueofinnatenessislessfundamentaltothebasicchildgrammarthanitistothelanguagebioprogramhypothesis,andbecauseSlobin’s(1997)recentreformulationisconsistentwithacognitive–functionalview,Idonotconsiderthebasicchildgrammarasaformalist–nativisttheoryinthisdebate.

meaning in acquisition 2�1

contaminatedbyexternalfactorssubjecttoculturalevolution(e.g.,individuallin-guisticvariationsthatareidiosyncratic).

Thelanguagebioprogramhypothesisattemptedtoexplainlanguageacquisitionbyappealingtoinnatelydeterminedsemanticdistinctions.Acontrastingperspec-tiveisthefunctionalist-cognitivistapproachtotense-aspectacquisition,whichhashadmanyvariantsintheliterature.Oneearlyexplanationfortherestrictedtense-aspectusesinchildlanguagedrewonthechild’spurportedlyinsufficientcognitiveability(Antinucci&Miller,1976;Bronckart&Sinclair,1973).Otherinvestigatorsturnedtothe“inputhypothesis”toexplaintheearlyassociationsbetweengram-maticalmorphologyandlexicalaspect(e.g.,Stephany,1981forModernGreek;Li,1990forchildMandarin;Shirai,1991forEnglish).Stillothersusedthe“aspecthypothesis”orthe“prototypehypothesis”toexplainsimilarpatternsinL2learn-ing(Shirai,1991;Shirai&Andersen,1995).Morerecently,LiandShirai(2000)presentedanintegratedviewofthefunctionalapproach,drawingideasfromboththeprototypehypothesisandconnectionistnetworks.Theyarguedthat inbothL1andL2,thelearner’searlyassociationsbetweenlexicalmeaningsofverbsandgrammatical morphemes do not indicate innate specifications of semantic cat-egories.Rather,theseassociationsreflectthelearner’ssensitivityto(andrecogni-tion of) the statistical properties of the linguistic input, and in turn, statisticalpropertiesintheinputmayreflectinherentconstraintsonlinguisticcommunica-tionandeventcharacteristics.Thesemanticcategoriesbelievedtobeinnatecanemergenaturallyfromthelearningofthelexicalcharacteristicsofverbsincontext.Finally,LiandShiraiarguedthattheassociationsbetweengrammaticalmorphol-ogyandlexicalaspectareprobabilisticandnotabsolute,countertowhatnativistproposalswouldassume.Dependingonthestructureof thetarget language, insomecases,learnersretreatfromtheprobabilisticassociationsanddevelopmoreflexiblepatternsofuse;inothercases,theyholdontotheseassociationsevenastheyacquireadultpatternsoflanguageuse.

Resolving the Conflict Crosslinguistically

Ithasbecomeincreasinglyclearoverthelastfewyearsthatastrongnativistpro-posal like Bickerton’s cannot account for the large body of crosslinguistic data.Instead, a functional, input-based, probabilistic learning mechanism seems tobemostcompatiblewiththewaychildrenapproachtheproblemoftense-aspectacquisition.EvenBickerton(1999)himselfhassignificantlyweakenedthestrongpredictionsoftheoriginallanguagebioprogramhypothesis,assigninganincreasedrole to input languagepatterns inacquisition.Buthowdoescrosslinguisticevi-dencehelpinresolvingtheconflict?

LiandBowerman(1998)presentedcrosslinguisticdataontheacquisitionofaspect inChinese. In threeexperimentswe showed that youngchildren learn-ing Mandarin Chinese displayed the same type of strong associations betweengrammaticalmorphologyandlexicalaspectasinEnglishandotherlanguages.Inparticular,childrencomprehendedtheprogressivemarkerzaibetterwithatelic,activityverbsthanwithtelicverbs,and,conversely,theperfectivemarker-lebet-terwithtelicverbsthanwithatelicverbs.Childrenalsoproducedtheimperfective


aspectmarkerszai and -nemostlywithatelic verbs and rarelywith telic verbs,whereastheyproducedtheperfectivemarker-lemorefrequentlywithtelicverbsthanwithatelicverbs.

WhiletheChinesedatawouldseemtobenosurpriseascomparedwithdatafromEnglishandotherlanguages,therewereatleasttwoimportantfeaturesthatmakethemspecial.

1.Unlike English and other Indo-European languages, Mandarin Chi-nese has a special set of state verbs, the posture verbs, that cannottake the progressive marker zai (e.g., zhan ‘stand,’ zuo ‘sit,’ and tang‘lie’). Recall that Brown (1973) observed that children donot overgen-eralize -ing to stateverbs inEnglish (postureverbsarenot stateverbsin English as they are in Chinese), and Bickerton interpreted this asindicative of children’s innate sensitivity to the state-process distinc-tion. In child Mandarin, however, children do overgeneralize the pro-gressivemarkerzaitoposturestateverbsasrevealedinourproductionexperiment,andsucherrorsarealsoobservedinchildren’sspontaneousspeech.Twolessonscanbedrawnfromthisfinding.

a. First,thestate-processdistinctionisnotauniversalsemanticprimi-tiveasdictatedbythelanguagebioprogramhypothesis,anddifferentlanguages may define the distinction differently. For example, pos-tureverbsarestateverbsinChinesebutnotinEnglish(inChinese*ta zai zuo isungrammaticalwhile inEnglishhe is sitting there isperfectlyacceptable).TheEnglish-speakingchildcanuse-ingwithsitandstand to indicate thedynamicaspectof theaction, treatingsittingandstandingasevents,whiletheChinese-speakingchilduseszaiwithzuo andzhan inChinese, treating themas states.Comrie(1976) discussed this phenomenon in connection with perceptionverbs(e.g.,see, hear):Englishtreatsperceptionverbsasstativeandthese verbs consequently do not accept progressive marking, whilePortuguesetreatsthemasdynamicsotheycannaturallyacceptpro-gressivemarking.

b. Even if the state-process distinction is neatly defined by language,childrenmaynotobservethedistinctionintheiruseoftense-aspectmarkers.ThisisconsistentwithShirai’s(1994)analysisthatshowsthateveninEnglish,childrendooccasionallygeneralize-ingtostateverbsdependingonthetypeofinputtheyreceivefrommaternalspeech.

2.AsecondimportantfeatureisthatChinesehasaspecialsetofresultativeverb compounds (RVCs), such as ti-dao ‘kick-down,’ qie-kai ‘cut-open,’orreng-diao‘throwaway,’whicharetelicverbsthatencodeaclearendresult(Klein,Li,&Hendriks,2000;Smith,1997;Tai,1984).Theseverbsaccept only perfective marking, not progressive marking in Chinese,unlikeresultativeverbsinEnglish,whichcaneasilytaketheprogressive -ing (andsotheresultative -ingcombinationsarefoundinchildEnglish).Inour studychildrenrarelyused theprogressiveaspectwith theRVCverbs,showingthattheyrespectedtheincompatibilitybetweenaspectual


imperfectivityandlexicalresultativityfromearlyon.Interestingly,theseRVCverbscorrespondtoasetofverbsthat“nameeventsofsuchbriefdurationthat theevent isalmostcertaintohaveendedbeforeonecanspeak” (Brown, 1973, p. 334)—i.e., verbs with which young English-speakingchildrenonlyuse thepast-tense forms.Bickertonwouldhavetaken theexclusiveperfectivemarkingofRVCsasan indicationof thepunctual–nonpunctualdistinction,justashedidwiththeEnglishdata.Butthecrucialevidencefromsemelfactiveverbs(Smith,1997;e.g.,tiao‘jump,’qiao‘knock,’ti‘kick,’whichhavethepunctualitybutnottheresul-tativityfeature)indicatedthatitwasnotpunctualitybutresultativitythatchildrenpayattentionto(Li&Bowerman,1998):Withthesesemelfactiveverbs,Chinese-speakingchildrendousetheprogressivemarkerzai.

Thelanguage-specificpropertieswithRVCverbsinChinesenotonlyinfluence the specificpatternsof children’s acquisitionof aspectmark-ers with these verbs, but also speak to several general crosslinguisticdifferences.

a. First,becausetheadultlanguagehasaconstraintonthecombinationofRVCswiththeprogressivemarker,thepictureoflexicalaspectandgrammaticalmorphologyappearsmuchmoreabsolutethanprobabi-listic,ascomparedwiththatinotherlanguages.

b. Second, as compared with children learning other languages, Chi-nese-speaking children display no developmental transition fromprototypical associations (e.g., result verbs with perfective aspect)to nonprototypical associations (e.g., result verbs with imperfectiveaspect). The prototype hypothesis predicts such transitions as thechild’slinguisticexperienceenriches(Shirai&Andersen,1995).Now,giventhattheprototypicalassociationbetweenRVCsandperfectivemarking is preserved in the adult language, there is no reason forchildren toretreat fromthisassociationandmovetononprototypi-calones,astheywouldinotherlanguageswheretheassociationsaremoreflexible.

c. Finally,thecombinatorialconstraintsinadultChinesemightreflectageneralconstrainton linguisticcommunicationandeventcharac-teristics.TheprototypicalassociationbetweenRVCsandperfectiveaspect reflectsoneof themostnatural combinationsbetweenverbsemantics and grammatical morphology. In Comrie’s (1976) view,certainaspectmorphemescombinemostnaturallywithcertainverbtypesbutnotothers(the“naturalnessofcombination”principle).AsBrown(1973)hadalsopointedout,eventsdenotedbyverbslikedrop, fall,and crash(eventsexpressedbyRVCsinChinese)occurinstanta-neously;anycommentonthemwillhaveoccurredaftertheirending.Thus,itisonlynaturaltodescribetheseeventswithperfectiveaspect(i.e.,tocombine-lebutnotzaiwithRVCsinChinese).


The above crosslinguistic analyses, along with results from studies of otherlanguagesindicatethatchildrenarehighlysensitivetolanguage-specificproper-tiesoftheinput,andarecapableofextractingsystematicpatternsfromtheinput(seeLi&Bowerman,1998;Li&Shirai,2000fordetaileddiscussions).Giventhis,wedonotneedtopresuppose,asnativistsdo,thatcertainsemanticcategoriesareinnatelyspecifiedandbroughttobearonthe languageacquisitiontask.Rather,semantic categories can emerge from the learning of the statistical regularitiesintheinputlanguage.Butwhatcapacityallowsthechildtocarryoutthepatternextractioninlearning?Putsimply,ifinputisimportanttolanguageacquisition,inwhichwaydoesitplayacausalrole?Ireferredearliertoafunctional,input-based,andprobabilisticlearningmechanismthatcouldberesponsiblefortheacquisitiontask.Inthenextsection,Idiscusshowsuchalearningmechanismcouldworkintermsoftheoperationsofconnectionistnetworks.

the acquisition of a cryPtotyPe

Whorf’s Cryptotypes

Inoneoftheclassicpapersofearlycognitivelinguistics,Whorf(1956)presentedthe following puzzle. In English, the reversative prefix un- can be used pro-ductivelywithmanyverbstoindicatethereversalofanaction,asin uncoil, uncover, undress, unfasten, unfold, unlock,or untie (themeaningofreversalcanalsobeexpressedbyotherprefixessuchasdis-orde-inEnglish).However,manyseem-inglyparallelformsarenotallowed,suchas*unbury, *unfill, *ungrip, *unhang, *unpress, *unspill, or *unsqueeze. Whyisun-prefixationallowedwithsomeverbsbutnotothers?

Whorf’spuzzlewasdeeperthanthissimplediscrepancy.Henotedthatun-isaproductivedevice inEnglishmorphology,andthatdespitethedifficulties lin-guistshaveincharacterizingitsuse,nativespeakersdohaveanintuitivefeelforwhich verbs canbeprefixedwithun- andwhich cannot.Hepresented the fol-lowingthoughtexperiment:Ifanewverbflimmickiscoinedtomean“totieatincan to something,” thennative speakersarewilling toaccept thesentence“Heunflimmickedthedog”asexpressingthereversalofthe“flimmicking”action; ifflimmickmeans“totakeapart,”thentheywillnotaccept“Heunflimmickedthepuzzle”asdescribingtheactofputtingapuzzlebacktogether.Theconstrainedproductivityofun-promptedWhorftoconjecturethatthereissomeunderlyingorcovertsemanticcategory,acryptotype,thatgovernstheproductiveuseofun-.AccordingtoWhorf,cryptotypesonlymaketheirpresenceknownbytherestric-tionstheyplaceonthepossiblecombinationsofovertforms.Whentheovertprefixun-iscombinedwiththeovertverbtie,thereisacovertcryptotypethatlicensesthecombinationuntie.This samecryptotypealsoblocksacombinationsuchas*unmove.

ToWhorf,thedeeppuzzlewasthatwhiletheuseoftheprefixun-isproduc-tive, thecryptotypethatgoverns itsproductivity isunclear:“Wehavenosingleword in the language which can give us a proper clue to this meaning or into


whichwecancompressthismeaning;hencethemeaningissubtle,intangible,asistypicalofcryptotypicmeanings” (Whorf,1956,p.71).Herewehaveacase forwhich languageuse isconditionedinaprincipledway,buttheprinciplesthem-selvesarenotclearlysubjecttolinguisticanalysis.AtsomepointWhorfdidpro-posethattherewas“acovering,enclosing,andsurface-attachingmeaning”thatcouldbe thebasisof thecryptotype forun-.But thisdefinitionwasstill ratherelusive, as it was not clear whether we should view this as a single unit, threeseparatemeanings,oraclusterofrelatedmeanings.Subsequentanalysesalsosug-gestedtheexistenceofotherimportantaspectsintheuseofun-(Clark,Carpenter,&Just,1995;Marchand,1969)—forexample,thatun-takeschangeofstateverbs,andthattheseverbsinvolveadirectobject(sointransitiveverbssuchas*unswim,*unplay,and*unsnoreareill-formed).

Cryptotypes in Child Language

Whorf’sdiscussionshowsclearlyhowacryptotypeisimportanttotheuseofun-inadultEnglish.Bowermanwasthefirsttopointoutthatthenotionofacryptotypemightalsoplayanimportantroleinchildlanguageacquisition.

AccordingtoBowerman(1982,1983,1988),children’sacquisitionofun-tendsto follow a U-shaped pattern, a pattern found in other areas of morphologicalacquisition as well, such as the acquisition of the English past tense. Childreninitiallyproduceun-verbsinappropriatecontexts,treatingun-anditsbaseverbasanunanalyzedwhole.Thisinitialstageofrotecontrolisanalogoustothechild’ssayingwentwithoutrealizingthatitisthepast-tenseformofgo.

Productivity of un- comes at the next stage, when children realize that un-is independentof theverb in indicating the reversalof anaction.This stage intheacquisitionofun-beginsataroundage3.Atthisstage,childrenstarttopro-duce overgeneralizations in spontaneous speech such as *unarrange, *unbreak, *unblow, *unbury, *unget, *unhang, *unhate, *unopen, *unpress, *unspill, *unsqueeze, or *untake (Bowerman,1982,1983).SuchovergeneralizationshavealsobeendocumentedbyClarketal.(1995)inbothexperimentalandnaturalisticdatawithchildrenfromages3to5, andwerefoundintheCHILDESdatabase(Li&MacWhinney,1996).Duringthisperiod,childrenalsomakecertain“over-marking”errors.Forexample,thechildmightsay“unopen”butreallyonlymeantosayopen,or“unloosen”tomeanloosen.Insuchcases,thebaseformsopenandloosenhaveareversativemeaningthattriggerstheattachmentoftheprefix,evenwhentheactionofthebasemeaningisnotactuallybeingreversed.Theseerrorsareanalogoustoredundantpast-tensemarkingasin*camedandredundantpluralmarkingasin*feets.Finally,atathirdstage,overgeneralizationandovermarkingerrorsbothdisappear.

Acritical factor that leads tochildren’sovergeneralizationofun-at thesec-ondstageofthisU-shapedlearning,accordingtoBowerman(1982),isthatchil-drenhavesomehowdiscoveredtheinherentmeaningcommontotheverbsthattakeun-.Inotherwords,theyhavedevelopedwhatWhorfcalledthenative“in-tuitivefeel”forEnglishverbswithrespecttowhethertheyareun-able:Theyhaveacquiredtherepresentationfortheun-cryptotype.Examiningspeecherrorsfrom


alongitudinaldataset,Bowermanfurthersuggestedtwopossiblerolesforacrypto-type to influence the learning of un-: (1) “Generalization via cryptotype”: Thecryptotypetriggersmorphologicalproductivityandleadstoovergeneralizations.Thisoccursbecause,oncechildrenhaveidentifiedthecryptotype,theywillover-generalizeun-toallverbsthatfitthecryptotype,irrespectiveofwhethertheadultlanguageactuallyallowsun-inthesecases(e.g.,squeezefitsinthecryptotypejustasclenchdoes,sosay“unsqueeze”).(2)“Recoveryviacryptotype”:Thecryptotypehelpsthechildtoovercomeovergeneralizationsmadeatanearlierstage,iftheseovergeneralizationsinvolveverbsthatfalloutsidethecryptotype(e.g.,hatedoesnotfitinthecryptotypicmeaning,sostopsaying“unhate”).

While Bowerman correctly identified the important role of cryptotypes inchildlanguageacquisition,oneissueremainsunclear.AccordingtoWhorf,cryp-totypesarecovertsemanticcategoriesthatareelusive,subtle,andintangible,andlinguistshaveahardtimetopindowntheirprecisemeanings.Howthencouldthechildextractthecryptotypeanduseitasabasisformorphologicalgeneralizationorrecovery,iftheun-cryptotypeisintangibleeventolinguistslikeWhorf?

Theanswers to thisquestionhavesignificant implications for the issuecon-cerningwhetherlanguageacquisitionisarule-basedprocessorastatisticallearn-ingprocess(Pinker,1991;Pinker&Prince,1988;Rumelhart&McClelland,1986;Seidenberg,1997).UsingtheacquisitionoftheEnglishpasttenseasanexample,researchershavedebatedwhethertheacquisitionprocessshouldbecharacterizedbydualmechanisms(aninternalizedlinguisticruleforregularsandanassocia-tivelearningprocessforirregulars)orbyasinglemechanism(connectionistlearn-ingwithdistributedknowledgerepresentationandadaptiveconnectionweights).Cryptotypesprovideanothertestcaseforthisdebate.Ifthelearningofun-anditsgoverningcryptotype is aprocessof ruleextraction (category identification),thentheovergeneralizationerrorswithun-arerule-governed,inthesamewayasaretheovergeneralizationerrorswith-ed.However,ifthelearningofun-andthecryptotypeisaconnectioniststatisticalprocess,thentheovergeneralizationerrorsareduetothesystem’scomputationofrelevantsemanticfeatures,lexicalforms,andprefixationpatternsintheform-meaningmappingprocess.

A Connectionist Model of the un- Cryptotype

Connectionistnetworksaredynamiclearningsystemsthatexploretheregulari-tiesintheinput-outputmappingprocessesthroughtheadjustmentofconnectionweightsandtheactivationofprocessingunits.Toanswertheabovequestions,Li(1993)andLiandMacWhinney(1996)builtaconnectionistmodel to learnthereversativecryptotypeassociatedwiththeuseofun-.Ourmodelwasastandardfeedforwardnetworkconsistingofthreelayersofprocessingunits(input,output,and hidden units). The network was trained with the backpropagation learningalgorithm(Rumelhart,Hinton,&Williams,1986).Inoursimulations,weusedasinputtoournetwork49verbsthatcantakeun-,19verbsthatcantakethecompet-ingprefixdis-,and92randomlyselectedverbsthatcantakeneitherprefix.Eachverbwasrepresentedbyasemanticpattern(avector)thatconsistedof20semanticfeaturesthatwereselectedinanattempttocapturebasiclinguisticandfunctional


propertiesinherentinthesemanticrangeoftheseverbs(seeLi,1993fordetails).Thetaskofthenetworkwastotakethesemanticvectorsofverbsasinputandmapthemontodifferentprefixationpatternsintheoutput:un-,dis-,orzero.

Toanalyzehowournetworkdevelopedinternalrepresentations,weusedthehierarchicalclusteranalysis(Elman,1990)toprobeintotheactivationofthehid-denunits at variouspoints in timeduring thenetwork’s learning.Thenetworkreceivedinputverbsonebyoneanddeterminedifeachverbshouldbemappedtoun-,dis-, or noprefix. Each time the network received some feedback as towhetherthemappingwascorrect.Afterlearning,theaveragedrepresentationsoftheverbsatthehidden-unitlevelwereclustered.Figure9.1presentsasnapshotofthenetwork’shidden-unitrepresentationswhenthenetworkhadlearned50verbscumulatively.Thisgraphshows twogeneralclusters:one for theun-verbs,andtheotherforthezeroverbs,verbsthatcannotbeprefixedwithun-ordis-.Ourinterpretationoftheseclustersisthatthenetworkhasacquiredadistinctrepre-sentation for theun-verbsby identifying the relevant semantic features sharedbytheseverbs,andthisrepresentationcorrespondsmostclosely toWhorf’sun-cryptotype.Forexample,mostoftheverbsintheun-clustersharethemeaningofbindingor locking:bind, chain, fasten, hitch, hook, latch,etc.Notallmean-ingsrelevanttothecryptotypeareidentifiedatthisearlystageinFigure9.1.Forexample,theverbsravelandcoilwerecorrectlycategorizedintotheun-cluster,buttheverbrollwasincorrectlyclassifiedintothezero-verbcluster.

Notethatournetworkreceivednodiscretelabelofthesemanticcategoryasso-ciatedwithun-(thelabelsinFigure9.1weretheresimplytoindicatewhichprefixtheverbissupposedtotake),norwasthereasinglecategoricalfeaturethattellswhichverbshouldtakewhichprefix(henceWhorf’sproblem).Allthatthenetworkreceivedwasthesemanticfeatureinformationdistributedacrossinputpatterns.Over time,however, thenetworkwasable to identify the regularities thatholdbetweendistributedsemanticpatternsandpatternsofprefixation,anddevelopedastructuredrepresentationthatcorrespondedtoWhorf’scryptotype.Thecrypto-typerepresentationsinthenetworkthusemergedasafunctionofthenetwork’slearningoftheassociationbetweenformandmeaning,notasapropertythatwasgivenadhoctothenetworkbythemodeler.

Our simulation resultsprovide support forBowerman’s (1982)hypothesizedroleofcryptotypesininducingovergeneralizations.InBowerman’sdata,mostoftheerrorsfellwithintherealmofWhorf’scryptotype(e.g.,squeezeissimilartoclench,sosqueezecanalsotakeun-).InClarketal.’s(1995)data,thechild’sinno-vativeusesofun-alsorespectedthecryptotypefromthebeginning:Theymatchedthesemanticcharacteristicsofthecryptotypeevenwhentheconventionalmean-ingsoftheverbintheadultlanguagedidnot.Forexample,*unbuildwasusedtodescribetheactionofdetachingLego-blocks,*undisappearwasusedtodescribethereleasingofthechild’sthumbsfrominsidehisfists.3Oursimulationsalsoindi-catehowthenetworkgeneralizedun-onthebasisofthecryptotype.InFigure9.1,

3 Diarynotesofmydaughter’s speechalso includesimilaruses: “unbuild thesnowman”wasusedtorefertothedetachmentofdecorativepiecesfromthesnowman,and“untape”torefertotheremovalofscotchtapefromapieceofpaper(ageofchildwas6;9).


thenetworkincludedbothholdandmount(whichshouldnottakeun-)intheun-category.Theseverbswereincludedapparentlybecauseoftheirsemanticsimilar-itywithmembersofthecryptotype(e.g.,bind, chain, fasten, hitch, hook, latch).Examiningholdandmountinthenetworkoutputpatterns,wefoundthatun-wasovergeneralizedtotheseverbsattheendofthe50-wordstage.Similarerrorspro-ducedbythenetworkincluded *unbury, *uncapture, *unfill, *unfreeze, *ungrip, *unhold, *unloosen, *unmelt, *unpeel,*unplant, *unpress, *unsplit, *unsqueeze, *unstrip, *untack, and *untighten. In contrast to these cases, thenetworkpro-ducedfewerrorsthatconstituteflagrantviolationsofthecryptotype;hencetherewasnobasisforthemodeltoverifyBowerman’s(1982)secondhypothesisthatthe

reach ZERO allow ZERO charge DIS

use ZERO show ZERO look ZERO

work ZERO go ZERO

start ZERO like ZERO

hear ZERO talk ZERO

see ZERO say ZERO tell ZERO ask ZERO

give ZERO run ZERO walk ZERO

take ZERO get ZERO come ZERO

help ZERO wait ZERO

believe ZERO call ZERO

keep ZERO roll UN

stop ZERO turn ZERO

learn ZERO make ZERO

put ZERO chain UN

braid UN latch UN fasten UN bind UN

hitch UN cork UN plug UN

coil UN lace UN

mount DIS hook UN wind UN

hold ZERO ravel UN

connect DIS arrange DIS

Figure �.1 Ahierarchicalclusteranalysisofthenetwork’shidden-unitrepresentationsatthe50-wordstage.ThesimilarityfunctionisdeterminedbytheEuclideandistancesoftheitemsontheclustertree.Thecapitalizedmarkeraftereachverbindicatestheprefixationpatternoftheverb,butthenetworkdidnotreceivetheselabelsduringtraining.


cryptotypecanserve toeliminateovergeneralizations.Thus, inoursimulations,overgeneralizationswenthand-in-handwith thenetwork’s representationof thecryptotype. This is clearly another case (in addition to tense-aspect acquisitiondiscussedearlier)wheretheunderstandingofmorphologicalbehaviorrequirestheexaminationofthesemanticstructureofthewordswithwhichthegrammaticalmorphemesareused.

Implications of the Connectionist Model

A connectionist perspective as described above provides us with a natural wayofcapturingWhorf’sinsightsintocryptotypesaswellasaformalmechanismtomimictheiracquisition.Inourview,therecanbeseveral“mini-cryptotypes,”eachofwhichrepresentssomeunderlyingsemanticfeaturesthatworktogetherasinter-active“gangs”(McClelland&Rumelhart,1981).Forexample,“enclosing”verbs,suchascoil, curl, fold, reel, roll, screw, twist, and wind,allseemtoshareamean-ingofcircularmovement.Similarly,“attaching”verbs,suchasclasp, fasten, hook, link, plug, and tie,allinvolvehandmovements.Stillanotherclusterofverbssuchascover, dress, mask, pack, veil, and wrapformsthe“covering”mini-cryptotype.Thesemini-cryptotypesormini-gangsinteractcollaborativelytosupporttheform-ationofthelargercryptotypethatlicensestheuseofun-intermsofsummedacti-vationinaconnectionistnetwork.

Notethatthemini-gangscollaborateratherthancompetebecausetheirmem-bersarecloselyrelatedbytheoverlapofsemanticfeatures.Forexample,theverbscrew inunscrew maybeviewedashavingbothameaningofcircularmovementandameaningofbindingorlocking,zipinunzipmaybeviewedassharingboththe“binding/locking”meaningandthe“covering”meaning,andbothscrewandzip involvehandmovements.Moreover,a featuremayalsovary in thestrengthwithwhich it is represented indifferentverbs.Forexample,circularmovementisanessentialpartofthemeaningofscrew,butlesssoforwrap(onecanwrapasmallballwithatissuepaperwithoutturningaroundeithertheobjectorthewrap-pingpaper).Inthisway,themeaningoftheun-cryptotypeconstitutesacomplexsemanticnetwork,inwhichverbscandifferin(1)howmanyfeaturesarerelevanttoeachverb,(2)howstronglyeachfeatureisactivatedintherepresentation,and(3)howstronglyfeaturesoverlapwitheachotheracrosscategorymembership(alltrue with the input to our network). It is these complex relationships that giverisetothemeaningofthecryptotype.Itisalsothesecomplexrelationshipsthatgavetroubletotraditionalrule-basedlinguisticanalyses(henceWhorf’sstatementregardingtheelusiveandintangiblecharacterofitssemanticcontent).

Whilethesecomplexstructuralpropertiesrenderasymbolicanalysislessef-fectiveifnotimpossible,theyareaccessibletonativeintuition,accordingtoWhorf.Nativeintuitionsareclearlyimplicitrepresentationsofthecomplexsemanticrela-tionships among verbs and morphological markers, and connectionist networksprovide a formalmechanism to capture these intuitions throughweightedcon-nections,distributedrepresentations,andstatisticallearning.Inoursimulations,thenetworkwasabletoexploreandidentifytheserelationshipsthroughtheinput-outputmappingprocess.Thenetworkcomputedthecombinatorialconstraintson


theco-occurrencesoftheprefixun-andthedistributedsemanticfeaturesofverbs.Theresultofthisprocesswasthatnewrepresentationsthatdevelopedatthehid-denlayerofthenetworkdifferedfromeachotherinthenumberoffeaturestheysharedandinthestrengthswithwhichthefeatureswereactivated,asrevealedinFigure9.1.

Our network’s behavior suggests that children, in learning to use the re-versativeprefixun-,mayalsoabstractthesemanticregularitiesfromtheun-verbsthroughcombinatoryrestrictionsthattheprefixplacesontheseverbs.Inthisper-spective,children’slearningofun-isnotthelearningofasymbolicrulefortheuseoftheprefixwithaclassofverbs,butrathertheaccumulationoftheconnectionstrength that holds between a particular prefix and a set of weighted semanticfeaturesinverbs.Thelearnergroupstogetherthoseverbsthatsharethelargestnumberoffeaturesandtakethesameprefixationpatterns.Overtime,theverbsgradually form coherent classes, with respect to both meaning and prefixation.Thislearningprocesscanbestbedescribedasastatisticalprocedureinwhichthechildimplicitlytalliesandregistersthefrequenciesofco-occurrencesofdistrib-utedsemanticfeatures,lexicalitems,andmorphologicalmarkers.Notsurprisingly,thesameprocesswouldapplyequallywelltotheacquisitionoflexicalaspectcat-egoriesandtense-aspectmorphology(seeLi&Shirai,2000forananalysis;Zhao&Li,inpress,forarecentconnectionistformalization).

Theformationofacryptotypeasinthecaseofun-isnotanisolatedlinguisticphenomenon.Itcanbeobservedinmanydomainsinwhichtheproblemisprimar-ilysemanticallymotivated.Forexample,theuseofclassifiersisoneofthehardestproblemsforsecondlanguagelearnersofChinese,aswellasamajorchallengeforlinguisticdescription(cf.Chao,1968;Lakoff,1987).EachnouninChinesehastobeprecededbyaclassifierthatcategorizestheobjectofthenounintermsofitsshape,orientation,dimension,texture,countability,andanimacy.Theappropriateuseofclassifiersbynativespeakersislargelyautomatic,yetitisdifficultforlin-guiststocomeupwithacleardescriptionofrulesthatgoverntheiruse(Erbaugh,2006).Wecanprobablyassumethatnativespeakershaveacquiredarepresenta-tion that iscryptotype-like, inwhichmultiple semantic featuresconnected inanetworkjointlysupporttheuseofclassifiers.

Inshort,ourconnectionistmodelprovidessignificantinsightsintotheunder-standing of Whorf’s cryptotype—in particular, the understanding of complexstructuralrelationshipsinlexicalsemanticsandtheroleofastructuredsemanticrepresentation in theovergeneralizationofmorphology in language acquisition.Ourmodeldemonstrateshowcryptotyperepresentationscanemergeinconnec-tionistnetworksasanaturalresultofthemeaning-formmappingprocesses.Thesefindingshelpustobetterunderstandtheprocessesunderlyingimportantphenom-enasuchasU-shapedbehaviorinlanguageacquisition.


LexicaL organization in DeveLoPment

Lexical Categories and Lexical Organization

Thepreviousdiscussionexaminedtwomorphologicalcases,tense-aspectsuffixesandreversativeprefixes,andinbothcasesourfocushasbeenontheemergenceof thecorresponding lexical semanticcategoriesassociatedwith theusesof themorphology.Butlexicalcategoriesdevelopovertimeinchildren,andthestruc-turalrelationshipsbetweenwordscanchangeaslearningprogresses.Inaseminalpaperontheacquisitionofwordmeaning,Bowerman(1978)discussedtheissueofsemanticorganizationinchildlanguage(theessenceoftheissuewasreflectedclearlyinthetitle:“SystematizingSemanticKnowledge:ChangesoverTimeintheChild’sOrganizationofWordMeaning”).Bowerman’sdiscussionfocusedonreor-ganization(seealsoBowerman,1982):Reorganizationinthechild’smentallexiconoccurs when word pairs that are apparently not initially recognized as seman-ticallyrelatedmoveclosertogether inmeaning.Thereorganization isoftensig-naledbyspeecherrors,wheresemanticallysimilarwordscompeteforselectioninproduction in particular speech contexts—for example, substitutions of put forgive(“putmethebread”)orfallfordrop(“Ifalledit”).

Suchareorganizationviewhasimportantimplicationsforcurrenttheoriesofthe mental lexicon. It suggests that we need to take a developmental, dynamicperspectiveonthelexicalsystem(withrespecttobothsemanticandgrammaticalpropertiesofthelexicon).Apopulartrendincognitiveneurosciencetodayistheattempt to localize various “lexiconmodules” in thebrain.Usingneuroimagingtechniques,researchershaveidentifiedspecificareasinthebrainthatrespondtodifferentlexicalcategories,suchasnounsandverbs,concretewordsandabstractwords, content words and function words, and words for animals, persons, andtools (e.g., Caramazza & Hillis, 1991; Damasio, Grabowski, Tranel, Hichwa, &Damasio,1996,Pulvermueller,1999).Theunderlyinghypothesisisthatdifferentlinguisticcategoriesaresubservedbydifferentneuralsubstrates,thussupportingthemodularityofthemind/brainhypothesis(Fodor,1983).Neuroscienceresearchinthisdirection,alongwithitscompaniontheoryofmodularityinpsychologyandphilosophy, echoes a long historical tradition in brain localization (Bates, 1999;Gardner,1986;Uttal,2001).Afundamentalproblemwiththisapproach,however,isthatitignoresthefactthatlexicalmodules,iftheyexist,neednotbeinthebrainfromthebeginning.Byassigninganundueweighttoastaticstructurethatis“inthere,”itfailstoaddresstheoriginoftherepresentationoflinguisticcategoriesinthebrain.

Taking a developmental, dynamic perspective on this issue, research in mylabattemptsto(1)identifycrosslinguisticdifferencesinlexicalorganizationand,(2)capturestructured, localized,representationsof lexicalcategoriesasa func-tionoflearninganddevelopment.First,toidentifycrosslinguisticdifferences,inarecentneuroimagingstudywefoundthatMandarinspeakersshownodistinctneuralresponsestonounsandverbsinChinese(boththefrontalandthetemporalregionswereactivatedbybothnounsandverbs),incontrasttofindingsthatdis-tinctcorticalregionsareinvolvedwithnounsversusverbsinotherlanguages(Li,


Jin,&Tan,2004).Second,assumingthatthesedistinctlexicalmodulesdoexistinotherlanguages,wehaveattemptedtodescribehowtheycouldemergefromthelearner’sorganizationandreorganizationinresponsetocharacteristicsofthelearningenvironment, inbothmonolingualandbilingualcontexts (seeHernan-dez,Li,&MacWhinney2005, forareview;Chanetal., inpress for theneuralrepresentationofnounsandverbsinbilinguals).Tothisend,wehavedevelopedDevLex,aself-organizingconnectionistmodelofthedevelopmentofthelexicon.Ourmodelisdesignedtoachievetwogoalsatthesametime:tomodeltheacquisi-tionoflexicalorganizationasadevelopmentalprocessinchildlanguage,andtoexaminetheemergenceofcategoricalrepresentationsinournetworkasapossibleexplanationofneuralrepresentations.

The DevLex Model

Currentconnectionistmodelsoflanguageacquisitionhavefocusedontheexam-inationofphonologicalpatternsratherthanmeaningstructureofwords,ontheuseofartificiallygeneratedinputratherthanrealisticlinguisticdata,andontheuseofsupervisedlearningalgorithmsratherthanunsupervisedlearning(seeLi,2003fordiscussion).Theselimitationsledustoconsideramodelthatdealswiththe acquisition of semantics, with exposure to realistic child-directed parentalinput,andinself-organizingneuralnetworkswithunsupervisedlearning.Ourpri-maryconcernhasbeenthedevelopmentofapsycholinguisticallyplausiblemodelthatcanhandlerealisticlinguisticdatainthedomainoflexicalacquisition.

Like most previous connectionist models of language acquisition, the un-modelIpresentedearlierwasbasedontheback-propagationlearningalgorithm.Althoughsignificantprogresshasbeenmadewithmodelsbasedonback-propaga-tion,suchmodelshaveseriouslimitationswithregardtotheirneuralandpsycho-logicalplausibilityasmodelsofhumanlearning.Inparticular,“back-propagationnetworks”areknowntosufferfromcatastrophicforgetting(inabilitytorememberoldinformationwithnewlearning),fromscalability(inabilitytohandlerealistic,large-scale problems), and above all, from error-driven learning, which adjustsweightsaccordingtotheerrorsignalsfromthediscrepancybetweendesiredandactual outputs. Some of these problems become most transparent in the con-textoflanguageacquisition.Forexample,itwouldtakeastrongargumentifoneclaimedthatthefeedbackprocessusedinback-propagationresemblesprocessesofchildlanguagelearning.Childrendonotreceiveconstantfeedbackaboutwhatis incorrect in their speech, nor do they get the kind of error corrections on aword-by-wordbasisthatisprovidedtothenetwork(cf.the“nonegativeevidenceproblem”inlanguageacquisition;Baker,1979;Bowerman,1988).Instead,muchoflanguageacquisitioninthenaturalsetting,especiallytheorganizationofthemen-tallexicon,isaself-organizingprocessthatproceedswithoutexplicitteaching.

TheDevLexmodel(Li,Farkas,&MacWhinney,2004;Li,Zhao,&MacWhin-ney, 2007) is a type of self-organizing neural network. In contrast to networkswithback-propagationlearning,self-organizingnetworksdonotrequirethepres-enceofanexplicit teaching signal; learning is achievedentirelyby the system’sself-organization in response to the input. Self-organization in these networks


typicallyoccursinatwo-dimensionalmap—aself-organizingmap(SOM;Koho-nen,1982,2001).Eachprocessingunitinthenetworkisalocationonthemapthatcanuniquelyrepresentoneorseveralinputpatterns.Atthebeginningoflearning,aninputpatternrandomlyactivatesasetofunitsthatsurroundthebestmatchingunit(the“winner”).Oncetheseunitsbecomeactiveinresponsetoagiveninput,theweightsofthewinnerandthoseofitsneighboringunitsareadjustedsuchthattheybecomemoresimilartotheinput,andtheseunitswillthereforerespondtothe sameor similar inputsmore strongly thenext time.Thisprocess continuesuntilalltheinputscanelicitspecificresponsepatternsinthenetwork.Asaresultofthisself-organizingprocess,thenetworkgraduallydevelopsconcentratedareasofunitsonthemap(the“activitybubbles”)thatcaptureinputsimilarities,andthestatisticalstructuresimplicitinthehigh-dimensionalspaceoftheinputarepre-servedonthe2-Dspaceinthemap.

Theself-organizingprocessand its representationhaveclear implications forlanguageacquisition:Theformationofactivitybubblesmaycapturecriticalpro-cessesfortheemergenceofsemanticcategoriesinchildren’sacquisitionofthelex-icon.Inparticular,thenetworkorganizesinformationfirstinlargeareasofthemapandgraduallyzerosinontosmallerareas;thiszeroing-inisaprocessfromdiffusepatternstofocusedones,asafunctionofthenetwork’scontinuousadaptationtotheinputstructure.Thisprocessallowsustomodeltheemergenceoflinguisticcatego-riesasagradualprocessoflexicaldevelopment.Italsohasthepotentialofexplain-inglanguagedisordersthatresultfromthebreakdownoffocusedactivationortheinabilitytoformfocusedrepresentations(Miikkulainen,1997;Spitzer,1999).

Figure 9.2 presents a diagrammatic sketch of the DevLex model (for tech-nicaldetails,seeFarkas&Li,2002a;Li,Farkas,&MacWhinney,2004;seealsoLi,Zhao,&MacWhinney,2007forDevLex-II,anextensionoftheoriginalDev-Lexmodel).Itconsistsofalexical(phonological)mapthatprocessesphonologicalinformationofwords(PMAP),andagrowingsemanticmapthatprocessesseman-ticinformation(GSM).AnimportantfeatureofthemodelisthattheGSMnet-workcanautomaticallyextractsemanticandgrammaticalfeaturesofeachwordbycomputingthetransitionalprobabilitiesofco-occurringwordsinspeechcontext(Farkas&Li,2001).ThesizeofGSMcanalsogrowalongwithagrowinglexicon

self–organization

Hebbian learning

self–organization

word form

word meaning

PMAP

phonological map

GSMsemantic map

Figure �.2 DevLex:Aself-organizingneuralnetworkmodelofthedevelopmentofthelexicon.


inincrementalvocabularylearning(Farkas&Li,2002b;Lietal.,2004).GSMisconnectedtoPMAPviaHebbianlearning(Hebb,1949),accordingtowhichtheassociativestrengthbetweentwounitsisincreasediftheunitsarebothactiveatthesametime.Upontrainingofthenetwork,aphonologicalrepresentationofthewordispresentedtothenetwork,andsimultaneously,thesemanticrepresentationofthesamewordisalsopresentedtothenetwork.Throughself-organizationthenetworkformsanactivityonthephonologicalmapinresponsetothephonologicalinput,andanactivityonthesemanticmapinresponsetothesemanticinput.Atthesametime,throughHebbianlearningthenetworkformsassociationsbetweenthetwomapsforalltheactiveunitsthatrespondtotheinput.ThecombinationofHebbian learningwith self-organization in thiswaycanaccount for thepro-cessofhowthelearnerestablishesrelationshipsbetweensemanticfeatures,lexicalforms,andmorphologicalmarkers,onthebasisofhowoftentheyco-occurandhowstronglytheyarecoactivatedintherepresentation.

DevLexbasedontheabovecharacteristics(1)allowsustotrackthedevelop-mentofthelexiconclearlyasanemergentpropertyinthenetwork’sself-organiza-tion(fromdiffusetofocusedpatternsorfromincompletetocompleteassociativelinks),(2)allowsustomodelone-to-manyormany-to-manyassociationsbetweenformsandmeaningsinthedevelopmentofthelexiconandmorphology,and(3)providesuswithasetofbiologicallyplausibleandcomputationallyrelevantprin-ciples to study language acquisition—biologically plausible because the humancerebralcortexcanbeconsideredasessentiallyaself-organizingmap(ormultiplemaps)thatcompressesinformationona2-Dspace(Kohonen,2001;Spitzer,1999),andcomputationallyrelevantbecauselanguageacquisitioninthenaturalsetting(especiallyorganizationandreorganizationofthelexicon)islargelyaself-organiz-ingprocess(MacWhinney,1998,2001).Inwhatfollows,Ifocusonthemodel’sfirstproperty,thatis,howitisabletomodelthedevelopmentoflexicalorganization.Otheraspectsofthemodel,includingsimulationsoftheacquisitionofcryptotypesandtense-aspectmarkersandanaccountofthevocabularyspurt,canbefoundinLi(2003),Lietal.(2004),andLietal.(2007).

Lexical Organization in Development

BecauseDevLexwasdesignedtomodelarealistic lexicon,weusedtwosetsofchild language corpora as the basis of our modeling: the vocabulary from theMacArthur-Bates Communicative Development Inventories (the CDI; Dale &Fenson,1996)andtheparentalspeechfromtheCHILDESdatabase(MacWhin-ney,2000).FromCDI’sToddlerList(680words)weextracted500words,exclud-inghomographs,wordphrases,andonomatopoeias in theoriginal list.The500wordsweresortedaccordingtotheirorderofacquisition,determinedbytheCDIlexicalnormsatthe30thmonth.IntheCDI,earlywordscanbedividedintofourmajor categories: (1) nouns, including animals, body, clothing, food, household,outside,people,rooms,toys,andvehicles;(2)verbs;(3)adjectives;and(4)closed-classwords,includingauxiliaryverbs,connectingwords,prepositions,pronouns,quantifiers,andquestionwords.


To represent these 500 words as input to DevLex, we first produced theCHILDES parental corpus, which contains the speech transcripts from child-directedadult speech in theCHILDESdatabase (Li,Burgess,&Lund,2000).Next we presented the sentences (word by word) in the parental corpus to thegrowingsemanticmap(GSM).TheGSMthencomputedthelexicalco-occurrencestatistics for each of the 500 words in terms of the transitional probabilities ofsuccessivewords,andusedthesestatisticsasvaluesinavectortorepresentwordmeaning.Onemightwonderhowmuchtheseco-occurrencestatisticscancapturethesemanticaswellasthegrammaticalinformationofwords,butaseriesofprevi-ousexperimentsindicatethevalidityofthismethodinrepresentingthemeaningofwords(Li,Burgess,&Lind,2000;Li,Farkas,&MacWhinney,2004;Farkas&Li,2001,2002a,2000b).Tomodellexicalorganizationovertime,wedividedthe500wordsinto10growthstages,eachcomprising50words(cumulatively).Thus,lexicalrepresentationsmayvary(becomeenriched)fromstagetostage,asmoreandmorewordsareaddedintothetargetlexicon.

Figure 9.3 presents snapshots of the GSM at four different stages of learn-inginthenetwork.Thesesnapshotsillustratetheprocessoflexicalorganizationand reorganization in the network, as a result of the growing lexicon and thedevelopmentofenrichedlexicalrepresentationsovertime.Inparticular,onecanseehow themajorcategoriesofnouns,verbs, adjectives, andclosed-class itemsstart to form coherent classes. At the beginning of learning (stage 1), due to astrongbiasinfavorofnounsintheCDIvocabulary(andperhapsinchildEnglishin general), nouns spread all over the map. A few verbs present in the lexiconarescatteredbuthavenotformedanycompactclusters.Aslearningprogresses,moreverbs,adjectives,andclosed-classwordsenterthevocabulary,andthenounareastartstogivewaytowordsinothercategories.Thedevelopingcategoriesarealsoclearlyreflectedintheformationofsmallerclustersoverseveralareasonthemap—forexample,verbsatstage2.Moreimportantforourdiscussionhere,theseresultsnotonlyshowthedevelopmentofmajorgrammaticalcategories,butalsotheemergenceofsemanticcategorieswithinthemajorcategories.Mostnotice-ably,withintheboundaryofnounsanumberofcompactclustersemergetowardtheendoflearning.Forexample,wordsrepresentinganimals,people,householditems,food,andbodyparts(theCDIsemanticsubcategories)arecorrespondinglyclusteredwithinthenearestneighborhoodsonthemap.Bycontrast,thecluster-ingof semanticcategories for thesewords ismuch lessclearat theearly stages(i.e.,wordsthatbelongtothesamecategoryarespreadfartherapartonthemap).Theseresults illustrateclearlyhowlexicalorganizationmaychangeanddevelopovertime,withrespecttobothgrammaticalandsemanticfeaturesofwords.

Implications of the Model

Asmentionedearlier,DevLexisdesignedtoachievetwogoals:tomodeltheacqui-sitionoflexicalorganizationasadevelopmentalprocessinchildlanguageandtoexaminetheemergenceofcategoricalrepresentationsinthebrain.

Withrespecttothefirstgoal,ourhypothesisisthatthereareinherentstatistical(e.g.,distributional)differencesbetweennouns,verbs,adjectives,andclosed-class


wordsinEnglishandthatthenetworkcanidentifythesedifferencesthroughtheanalysisoflexicalco-occurrences.Traditionallinguisticanalyses(e.g.,structural-ism)alreadyhadampleevidenceonthedistributionaldifferencesbetweenlexicalcategories(deSaussure,1916).Thequestionforacquisitionistsishowthechildcanmakeuseofthesestatisticaldifferencesin learningthefunctionsofwords,andhowthislearningcanresultindifferentorganizationalpatternsovertime.

TheDexLexmodelshowsthedevelopmentalpathwaystolexicalrepresenta-tionandorganization,inthatearlyonthecategorymembershipsareratherdiffuseanddistributed,andlateronwithlearning,theybecomemorefocusedandlocal-ized.ThisdevelopmentfromdiffusetofocusedpatternsisconsistentwithrecentfindingsbySchlaggaretal.(2002)showingthatchildrenandadultsdisplaydiffer-entpatternsofneuralactivitiesinlanguageprocessing.Forchildren,theactivationpatternisdiffuseandunfocused,whereasforadults,theactivationpatternismore

Figure �.� SnapshotsofGSMattheendofdifferentstagesofdevelopment:stage1(50words—upperleft),3(150words—upperright),5(250words—lowerleft)and10(all500words—lower right). The sequence of images illustrates the nature of changes underly-ingthedevelopmentalprocess,asaresultofadevelopingvocabularyandthechanging/enrichedwordrepresentations.GSMclearlyseparatesthefourmajorcategories(andthesemanticsubcategorieswithineachcategory)towardthefinalstage.Becauseofthelargenumberofwordsinvolvedineachmap,theindividualwordsarenotlegibleinthisfigure.(ReprintedfromTrends in Cognitive Sciences, Vol.9,Hernandez,A.,Li,P.,&MacWhin-ney, B: The emergence of computing modules in bilingualism, pp. 220–225, Copyright(2005),withpermissionfromElsevier.)


focused,anddedicated to specificcortical regions.Note that thisearly-diffuse-late-focusedpatternisnotanartifactofthemodulationofthenetwork’strainingparameters—criticalparameterssuchaslearningrateandnetworksizewerekeptconstant across stages in our simulations. Our simulation results also match upwith Bowerman’s (1978) analyses of the organization and reorganization of thelexiconindevelopment:Manywordsthatarefarapartonthemapatanearlierstagebecomegroupedtogetheratalaterstage.Thistypeofreorganization,withinandacrosscategories,clearlyservestostructurelexicaldomainsasawhole(Carey,1978). In addition, according to Bowerman (1978), such reorganization is oftensignaledbyspeecherrors,wheresemanticallysimilarwordscompeteforselectioninproduction.InaseparatestudywhereproductionwasmodeledinDevLex,wefoundthattheorganizationalstructureofthecategorieshasasignificantimpactonthenumberofnamingerrorsaswellaswordconfusionsinthenetwork(Farkas&Li,2002b).

Withrespecttothesecondgoal,ourmodelshowsthatadevelopmentalcon-nectionistperspectivecanyieldsignificantinsightsintothenatureandoriginofcategoricalrepresentationinthebrain.Cognitiveneuroscientistshaveidentifiedvarious“braincenters”oflanguagefornouns,verbs,tools,fruits,animals,andsoon.Animportantassumptioninmanyoftheirstudiesisthatthebrainisahighlymodularized system, with different cognitive functions localized to differentcerebralregions,perhapsfromthebeginning(theclassical“modularityofmind”hypothesis;Fodor,1983).TheabilityofDevLextorepresentmajorlinguisticcat-egoriesdistinctivelywithoutdistinct representationalmodulesattests further totheprocessof“emergentorganization”throughwhichlocalizedrepresentationsormodulescanariseasafunctionofthedevelopmentalprocessesduringontogenesis,confirmingElizabethBates’motto“modulesaremade,notborn”(seeHernandezetal.,2005foradiscussion).Ourmodelshowshowtheorganizationoflocalmapscangiverisetoemergentcategoriesacrossstagesoflearning,withsubstantialearlyplasticityandearlycompetitionfortheorganizationandreorganizationofcategorymembers.Inourmodel,theGSMhasnopoolofunitsdedicatedtoanyspecificcategoryattheoutsetof learning,butas itprogressesthroughlearning,certaingroupsofunitsstart todevelopsensitivity toonly thesamekindsofwordsthatformcoherentcategories.

Althoughthereisanobviousdifferencebetween“neuronal”activitiesinourmodel and neural activities in the brain, DevLex can illustrate the functional mechanisms for the emergence of categorical representations through learninganddevelopment.InthespiritofBowerman’sresearchphilosophy,onecanoftenfind alternative explanations to a strong brain localizer’s account. For example,Farah(1994)pointedoutthatthewell-knowndeficitwithsomepatientsatprocess-ingclosed-classwordsmightbedue to speech stresspatterns thatdifferentiateclosed-classwordsfromopen-classwords—thatis,patientshavetroubledealingwith certain words because of the unstressed pattern and short duration, as ifdamageoccursonly to theclosed-classwords.Similarly,Shi (2006)pointedoutthatnewborninfantsareabletocategoricallydiscriminatelexicalcontentwordsfromgrammaticalfunctionwordsonthebasisofmultipleacousticandphonologi-calcues.FunctionalmechanismsofferedbyconnectionistmodelssuchasDevLex


areconsistentwithsuchexplanations,butareoftenatoddswithbrainlocalizationaccountsincurrentcognitiveneuroscience.

concLusions

InthischapterIhavediscussedtheacquisitionofwordmeaninginthreedomains:tense-aspect acquisition, cryptotype formation, and lexical organization. I haveattemptedtoshowthatthesemanticstructureofthelexicon,whetheritbeovertorcovert,canemergefromtheinteractionbetweenthelearnerandthelinguisticinput in the form-meaning mapping process, and we need no a priori assump-tionsabouttheinnatestatusorthesymbolicnatureofsemanticcategoriesinthelearner’srepresentationalsystem.

Thesearchformeaningischaracteristicofchildren’searlylexicalandmorpho-logical development. It is also characteristic of Bowerman’s research emphasis.ThethreeexamplesthatIhavediscussedhereillustratetheimportanceofseman-tic learning, the roleof linguistic input, and the roleof crosslinguisticdata, allofwhichhavebeencarefullyexaminedbyBowerman.Inmydiscussion,Ihavehighlightedhowconnectionistnetworkscanhelpusunderstand theacquisitionof lexicalsemanticrepresentations—inparticular,howsemanticrepresentationsmayemergethroughthestatisticalanalysisof the linguistic input.Thus,beforeattributingsemanticrepresentationsinchildrenas“pre-linguistic”orinnate,itisbettertofirstconsidermechanismsoflearningandthelearningenvironmentaspotentialsourcesofsolution(consistentwithBowerman’sapproachin“consideringthealternativesfirst”).Ourdiscussionsshowthatthelinguisticinputcontainsrichinformationthatthechildcanexploit intheacquisitionofthe lexicon,andthatmodular lexical categories that have often been considered innate (e.g., Bicker-ton,1981)mayemergefromthelearningofstatisticalpropertiesinlanguageuse.Clearly,thedata-rich,highlyconsistent,andstatisticallyregularinputinthechild’slearningenvironmentisatoddswiththe“povertyofstimulus”hypothesisthatthechildisexposedtoerror-laden,random,andinconsistentinput.Theviewthatchil-drenarestatisticallearnersforlanguageisgainingpopularityinrecentyearssincethepublicationofanumberofimportantvolumesandarticles(e.g.,Elmanetal.,1996;MacWhinney,1998;Saffran,Aslin,&Newport,1996).

Currentdebatesincognitivescienceandcognitiveneurosciencerevolvearoundtheissueofthenatureoflinguisticrepresentation.Modelsbasedonclassicallin-guistic theories construe linguistic representations in terms of rules in symbolsystems.Achildissaidtohaveinternalizedageneralruleinhermentalrepresen-tation,“adding-ed tomakethepasttense,”atsomestageoflanguageacquisition.Thiskindofdescriptionseemsintuitivelyclear,andtherulesoffergreatdescrip-tivepower.However,connectionistmodelsprovidealternativeexplanationstothisperspective, explanations that place strong emphasis on the statistical learningprocessesthatleadtorule-likebehaviors.Thesemodelsareespeciallysuitedforsolvingproblemswithwhichtraditionalsymbolicanalyseshavedifficulty,suchasthecryptotypeproblemdiscussedherethatwasoncethought“subtle”and“intan-gible” by Whorf. The distributed representations and adaptive weights used in


connectionistmodelsprovidemechanismstocapturethecomplexsemanticrela-tionshipsamongwordsandbetweenwordsandtheirmorphologicalmarkers.

Insum,myconclusionisthatstructuredsemanticrepresentationscanemergethrough statistical computations of the various constraints among lexical items,semanticfeatures,andmorphologicalmarkers,andthedevelopmentandorgan-izationoftherepresentationsareduetobasicprobabilisticproceduresofthesortembodiedinconnectionistnetworksinthelearningofform-to-formandform-to-meaningmappings.Weonlyneedtomakeafewsimpleassumptionsforthistypeofprobabilisticprocedures towork foryoungchildren—forexample,aworkingmemory that can hold items in sequence and the ability to track distributionalregularities(e.g.,co-occurrences)inthesequence.Suchabilities,asrecentstudiesofstatisticallearningininfantshaverevealed,seemtobereadilyavailabletotheyoungchildataveryearlystage(Saffran,Aslin,&Newport,1996;Saffran,New-port,Aslin,Tunick,&Barrueco,1997).

acKnoWLeDgments

Preparation of this article was supported by grants from the National ScienceFoundation(#BCS-9975249and#BCS-0131829).IwouldliketothankElizabethBates,MelissaBowerman,BrianMacWhinney,andRistoMiikkulainenfortheircommentsandinsightsontheideaspresentedhere.

references

Aksu,A.(1978).Aspectandmodalityinthechild’sacquisitionoftheTurkishpasttense.Unpublisheddoctoraldissertation,DepartmentofPsychology,UniversityofCalifor-nia,Berkeley.

Aksu-Koç, A., & Slobin, D. (1985). The acquisition of Turkish. In D. Slobin (Ed.), The crosslinguistic study of language acquisition (Vol. 1, pp. 839–878). Hillsdale, NJ:Erlbaum.

Antinucci, F., & Miller, R. (1976). How children talk about what happened. Journal of Child Language, 3,167–189.

Baker, C. (1979). Syntactic theory and the projection problem. Linguistic Inquiry, 10,533–581.

Bates,E.(1999).Plasticity,localizationandlanguagedevelopment.InS.Broman&J.M.Fletcher (Eds.), The changing nervous system: Neurobehavioral consequences of early brain disorders(pp.214–253).Newyork:OxfordUniversityPress.

Bickerton,D.(1981).Roots of language.AnnArbor,MI:Karoma.Bickerton, D. (1984). The language bioprogram hypothesis. Behavioral and Brain Sci-

ences, 7,173–188.Bickerton,D.(1999).Creolelanguages,thelanguagebioprogramhypothesis,andlanguage

acquisition. In W. C. Ritchie & T. K. Bhatia (Eds.), Handbook of child language acquisition. SanDiego,CA:AcademicPress.

Bloom,L.,Lifter,K.,&Hafitz,J.(1980).Semanticsofverbsandthedevelopmentofverbinflectioninchildlanguage.Language, 56,386–412.

Bowerman,M.(1978).Systematizingsemanticknowledge:Changesovertimeinthechild’sorganizationofwordmeaning.Child Development, 49,977–987.


Bowerman,M. (1982).Reorganizationalprocesses in lexical and syntacticdevelopment.InE.Wanner&L.Gleitman(Eds.),Language acquisition: The state of the art(pp.319–346).Cambridge,UK: CambridgeUniversityPress.

Bowerman,M.(1983).Hiddenmeanings:Theroleofcovertconceptualstructuresinchil-dren’sdevelopmentoflanguage.InD.Rogers&J.Sloboda(Eds.),The acquisition of symbolic skills(pp.445–470). Newyork:Plenum.

Bowerman,M.(1985).Whatshapeschildren’sgrammars?InD.Slobin(Ed.),The crosslinguistic study of language acquisition (Vol. 2, pp. 1257–1319). Hillsdale, NJ:Erlbaum.

Bowerman,M.(1988).The“nonegativeevidence”problem:Howdochildrenavoidcon-structinganoverlygeneralgrammar?InJ.Hawkins(Ed.),Explaining language uni-versals (pp.73–101). Newyork:Blackwell.

Bowerman,M.(1989).Learningasemanticsystem:Whatroledocognitivepredispositionsplay? InM.Rice&R.L.Schiefelbusch (Eds.),The teachability of language (pp.133–169).Baltimore,MD:PaulH.Brookes.

Bowerman, M. (1996). Learning how to structure space for language: A crosslinguisticperspective.InP.Bloom,M.Peterson,L.Nadel,&M.Garrett(Eds.),Language and space(pp.385–436).Cambridge,MA:MITPress.

Bronckart,J.,&Sinclair,H.(1973).Time,tenseandaspect.Cognition, 2,107–130.Brown,R.(1973).A first language.Cambridge,MA:HarvardUniversityPress.Caramazza,A.,&Hillis,A.E.(1991).Lexicalorganizationofnounsandverbsinthebrain.

Nature, 349,788–790.Carey,S.(1978).Thechildaswordlearner.InM.Halle,G.Miller,&J.Bresnan(Eds.),

Linguistic theory and psychological reality (pp. 264–293). Cambridge, MA: MITPress.

Chan,A..Luke,K.,Li,P.,Li,G.,yip,V.,Weekes,B.etal.(inpress).Neuralcorrelatesofnounsandverbsinearlybilinguals.Annals of the New York Academy of Sciences.

Chao, y.-R. (1968). A grammar of spoken Chinese. Berkeley: University of CaliforniaPress.

Clark,E.,Carpenter,K.,&Deutsch,W.(1995).Referencestatesandreversals:Undoingactionswithverbs.Journal of Child Language, 22,633–662.

Comrie,B.(1976).Aspect: An introduction to the study of verbal aspect and related prob-lems.Cambridge,UK:CambridgeUniversityPress.

Dale,P.S.,&Fenson,L.(1996).Lexicaldevelopmentnormsforyoungchildren.Behav-ior Research Methods, Instruments, & Computers, 28,125–127(InformationaboutCDIisavailableathttp://www.sci.sdsu.edu/cdi/#lexnorms).

Damasio,H.,Grabowski,T.,Tranel,D.,Hichwa,R.,&Damasio,A.(1996).Aneuralbasisforlexicalretrieval.Nature, 380,499–505.

Elman,J.(1990).Findingstructureintime.Cognitive Science, 14,179–211.Elman,J.,Bates,A.,Johnson,A.,Karmiloff-Smith,A.,Parisi,D.,&Plunkett,K.(1996).

Rethinking innateness: A connectionist perspective on development. Cambridge,MA:MITPress.

Erbaugh,M. (1982).Coming to order: Natural selection and the origin of syntax in the Mandarin speaking child.Unpublisheddoctoraldissertation,UniversityofCalifor-nia,Berkeley.

Erbaugh,M. (2006).Chinese classifiers:Theiruse andacquisition. InP.Li,L.Tan,E.Bates,&O.Tzeng(Eds.),Handbook of East Asian psycholinguistics: Vol.1. Chinese(pp.39–51).Cambridge,UK:CambridgeUniversityPress.

Farah,M.(1994).Neuropsychologicalinferencewithaninteractivebrain:Acritiqueofthe“locality”assumption.Behavioral and Brain Sciences, 17,43–104.


Farkas, I.,&Li,P. (2001).Aself-organizingneuralnetworkmodelof theacquisitionofword meaning. In E. M. Altmann, A. Cleeremans, C. D. Schunn, & W. D. Gray(Eds.),Proceedings of the Fourth International Conference on Cognitive Modeling(pp.67–72).Mahwah,NJ:Erlbaum.

Farkas,I.,&Li,P.(2002a).Modelingthedevelopmentoflexiconwithagrowingself-or-ganizingmap.InH.J.Caulfieldetal.(Eds.),Proceedings of the 6th Joint Conference on Information Sciences(pp.553–556).JCIS/AssociationforIntelligentMachinery,Inc.

Farkas,I.,&Li,P.(2002b).DevLex:Aself-organizingneuralnetworkofthedevelopmentof lexicon. InProceedings of the International Conference on Neural Information Processing.

Fodor,J.(1983).The modularity of mind. Cambridge,MA:MITPressGardner,H.(1986).The mind’s new science:A history of the cognitive revolution. New

york:BasicBooks.Hebb,D.(1949).The organization of behavior: A neuropsychological theory.Newyork,

Newyork:Wiley.Hernandez,A.,Li,P.,&MacWhinney,B.(2005).Theemergenceofcompetingmodulesin

bilingualism.Trends in Cognitive Sciences, 9, 220–225.Klein,W.,Li,P.,&Hendriks,H.(2000).AspectandassertioninMandarinChinese.Nat-

ural Language and Linguistic Theory,18,723–770.Kohonen,T.(1982).Self-organizedformationoftopologicallycorrectfeaturemaps.Bio-

logical Cybernetics, 43,59–69.Kohonen,T.(2001).The self-organizing maps(3rded.).Berlin:Springer.Lakoff, G. (1987). Women, fire, and dangerous things. Chicago: University of Chicago

Press.Li,P.(1990).Aspect and aktionsart in child Mandarin.Unpublisheddoctoraldissertation,

LeidenUniversity,theNetherlands.Li, P. (1993). Cryptotypes, form-meaning mappings, and overgeneralizations. In E. V.

Clark(Ed.),The Proceedings of the 24th Child Language Research Forum (pp.162–178). Stanford, CA: Center for the Study of Language and Information, StanfordUniversity.

Li,P.(2003).Languageacquisitioninaself-organisingneuralnetworkmodel.InP.Quin-lan (Ed.),Connectionist models of development: Developmental processes in real and artificial neural networks (pp. 115–149). Hove, UK & New york: PsychologyPress.

Li,P.,&Bowerman,M.(1998).TheacquisitionoflexicalandgrammaticalaspectinChi-nese.First Language, 18,311–350.

Li,P.,Burgess,C.,&Lund,K.(2000).Theacquisitionofwordmeaningthroughgloballexicalco-occurrences.InE.Clark(Ed.),Proceedings of the Thirtieth Stanford Child Language Research Forum (pp.167–178).Newyork:CambridgeUniversityPress.

Li,P.,Farkas,I.,&MacWhinney,B.(2004).Earlylexicaldevelopmentinaself-organizingneuralnetwork.Neural Networks,17,1345–1362.

Li,P.,Jin,Z.,&Tan,L.(2004).NeuralrepresentationsofnounsandverbsinChinese:AnfMRIstudy.NeuroImage,21,1533–1541.

Li,P.,&MacWhinney,B.(1996).Cryptotype,overgeneralization,andcompetition:Acon-nectionistmodelofthelearningofEnglishreversiveprefixes.Connection Science, 8,3–30.

Li,P.,&Shirai,y.(2000).The acquisition of lexical and grammatical aspect.Berlin&Newyork:MoutondeGruyter.

Li,P.,Zhao,X.,&MacWhinney,B. (2007).Dynamic self-organizationandearly lexicaldevelopmentinchildren.Cognitive Science, 31, 531–612.


MacWhinney,B.(1998).Modelsoftheemergenceoflanguage.Annual Review of Psychol-ogy,49,199–227.

MacWhinney,B. (2000).The CHILDES project: Tools for analyzing talk.Hillsdale,NJ:Erlbaum.

MacWhinney,B. (2001).Lexicalist connectionism. InP.Broeder& J.M.Murre (Eds.),Models of language acquisition: Inductive and deductive approaches (pp. 9–32).Oxford:OxfordUniversityPress.

Marchand,H.(1969).The categories and types of present-day English word-formation: A synchronic-diachronic approach. Munich:C.H.Beck’scheVerlagsbuchhandlung.

McClelland,J.,&Rumelhart,D.(1981).Aninteractiveactivationmodelofcontexteffectsinletterperception:Part1.Anaccountofthebasicfindings.Psychological Review, 88,375–402.

Miikkulainen,R.(1997).Dyslexicandcategory-specificaphasicimpairmentsinaself-orga-nizingfeaturemapmodelofthelexicon.Brain and Language, 59,334–366.

Pinker,S.(1991).Rulesoflanguage.Science, 253,530–535.Pinker,S.,&Prince,A.(1988).Onlanguageandconnectionism:Analysisofaparalleldis-

tributedprocessingmodeloflanguageacquisition.Cognition, 28,73–193.Pulvermueller,F.(1999).Wordsinthebrain’s language.Behavioral and Brain Sciences,

22,253–336.Rumelhart,D.,Hinton,G.,&Williams,R.(1986).Learninginternalrepresentationsby

error propagation. In D. Rumelhart, J. McClelland, & the PDP Research Group(Eds.),Parallel distributed processing: Explorations in the microstructures of cognition (Vol.1,pp.318–362).Cambridge,MA:MITPress.

Rumelhart,D.&McClelland,J.(1986).OnlearningthepasttensesofEnglishverbs.InJ.McClelland,D.Rumelhart,&thePDPresearchgroup(Eds.),Parallel distributed processing: Explorations in the microstructure of cognition (Vol. 2,pp. 216–271).Cambridge,MA:MITPress.

Saffran, J.,Aslin,R.,&Newport,E. (1996).Statistical learningby8-month-old infants.Science, 274,1926–1928.

Saffran,J.,Newport,E.,Aslin,R.,Tunick,R.,&Barrueco,S.(1997).Incidentallanguagelearning:Listening(andlearning)outofthecornerofyourear.Psychological Sci-ence, 8,101–105.

Saussure,F.de.(1916).Cours de linguistique générale.Paris:Payot.(Englishtranslation: A course in general linguistics.Newyork:PhilosophicalLibrary;Chinesetranslation:Putong Yuyanxue Daolun.Beijing:PekingUniversityPress).

Schlaggar,B.,Brown,T.,Lugar,H.,Visscher,K.,Miezin,F.,&Petersen,S.(2002).Func-tional neuroanatomical differences between adults and school-age children in theprocessingofsinglewords.Science, 296,1476–1479.

Seidenberg,M.(1997).Languageacquisitionanduse:Learningandapplyingprobabilisticconstraints.Science, 275,1599–1603.

Shi,R.(2006).Basicsyntacticcategoriesinearlylanguagedevelopment.InP.Li,L.Tan,E.Bates,&O.Tzeng(Eds.),Handbook of East Asian psycholinguistics:Vol.1. Chinese(pp.90–102).Cambridge,UK:CambridgeUniversityPress.

Shirai,y.(1991).Primacy of aspect in language acquisition: Simplified input and proto-type.Unpublisheddoctoraldissertation,UniversityofCalifornia,LosAngeles.

Shirai,y.(1994).Ontheovergeneralizationofprogressivemarkingonstativeverbs:Biopro-gramorinput?First Language, 14,67–82.

Shirai,y.,&Andersen,R.(1995).Theacquisitionoftense-aspectmorphology:Aprototypeaccount.Language, 71,743–762.

Shirai, y., Slobin, D., & Weist, R. (1998). Introduction: The acquisition of tense/aspectmorphology.First Language, 18, 245–253.


Slobin,D.(1985).Crosslinguisticevidenceforthelanguage-makingcapacity.InD.Slobin(Ed.),The crosslinguistic study of language acquisition (Vol.2,pp.1158–1249).Hills-dale,NJ:Erlbaum.

Smith,C.(1997).Theparameter of aspect (2nded).Dordrecht:Kluwer.Spitzer, M. (1999). The mind within the net: Models of learning, thinking, and acting.

Cambridge,MA:MITPress.Stephany,U.(1981).VerbalgrammarinmodernGreekearlychildlanguage.InP.S.Dale

&D.Ingram(Eds.),Child language: An international perspective. Baltimore,MD:UniversityParkPress.

Tai,J. (1984).Verbsandtimes inChinese:Vendler’s fourcategories.InPapers from the parasession on lexical semantics of the Chicago Linguistic Society(pp.289–296).

Uttal,W.(2001).The new phrenology: The limits of localizing cognitive processes in the brain.Cambridge,MA:MITPress.

Weist,R.(1986).Tenseandaspect:Temporalsystemsinchildlanguage.InP.Fletcher&M. Garman (Eds.), Language acquisition. Cambridge, UK: Cambridge UniversityPress.

Weist,R.,Wysocka,H.,Witkowska-Stadnik,K.,Buczowska,E.,&Konieczna,E.(1984).Thedefectivetensehypothesis:OntheemergenceoftenseandaspectinchildPolish.Journal of Child Language, 11,347–374.

Whorf, B. (1956). Thinking in primitive communities. In J.B.Carroll (Ed.), Language, thought, and reality.Cambridge,MA:MITPress.(Originalworkpublished1936)

Zhao,X.,&Li,P.(inpress).Theacquisitionoflexicalandgrammaticalaspectinadevelop-mentallexiconmodel.Linguistics.

Meaning in Acquisition Semantic Structure, Lexical Organization, and Crosslinguistic Variation

Documents

Transcript of Meaning in Acquisition Semantic Structure, Lexical Organization, and Crosslinguistic Variation