Phonetic Diversity, Statistical Learning, and Acqu

PhoneticDiversity,StatisticalLearning,andAcquisitionofPhonology*JanetB.PierrehumbertNorthwesternUniversity1IntroductionInfantsshowevidenceofphoneticcategorizationandofperceptualparsingofthespeechstreambeforetheylearntospeak,beforetheyhavelargevocabularies,andpossiblybeforetheyevenunderstandthatwordsarereferential.Somepartsofthespeechprocessingsystemareinitiatedearly.However,thesystemtakesalongtimetodevelop,notachievingadultlevelsevenattheageof12,accordingtosomerecentresultsbyHazanandBarrett(2000).Inadults,itachievesastoundinglevelsofspeed,accuracyandrobustnessinparsingcomplex,language-specific,phoneticpatterns.Inthisarticle,Ipresentsomeideasabouthowthesystemisinitiatedandsubsequentlyrefined.Theseideasarebasedonanintegrationoftheresearchliteratureinlin-guisticphonetics,psycholinguistics,andphonologicalacquisition.J.Pierrehumbert115AbstractInlearningtoperceiveandproducespeech,childrenmastercomplexlanguage-specificpatterns.Dauntinglanguage-specificvariationisfoundbothinthesegmentaldomainandinthedomainofprosodyandintonation.Thisarticlereviewsthechallengesposedbyresultsinphonetictypologyandsociolinguisticsforthetheoryoflanguageacquisition.Itarguesthatcategoriesareinitiatedbottom-upfromstatisticalmodesinuseofthephoneticspace,andsketcheshowexemplartheorycanbeusedtomodeltheupdatingofcategoriesoncetheyareinitiated.Italsoarguesthatbottom-upinitiationofcategoriesissuccessfulthankstotheperception-productionloopoperatinginthespeechcommunity.Thebehaviorofthisloopmeansthatthesuperficialstatisticalpropertiesofspeechavailabletotheinfantindirectlyreflectthecontrastivenessanddiscriminabilityofcategoriesintheadultgrammar.Thearticlealsoarguesthatthedevelopingsystemisrefinedusinginternalfeedbackfromtypestatisticsoverthelexicon,oncethelexiconiswell-developed.Theappli-cationoftypestatisticstoasysteminitiatedwithsurfacestatisticsdoesnotcauseafundamentalreorganizationofthesystem.Instead,itexploitsconfluencesacrosslevelsofrepresentationwhichcharacterizehumanlanguageandmakebootstrappingpossible.KeywordscategorizationexemplarsphonotacticsprosodystatisticallearningLANGUAGEANDSPEECH,2003,46(2–3),115–154115*Addressforcorrespondence:JanetPierrehumbert,LaboratoiredeSciencesCognitivesetPsycholinguistique,EcoleNormaleSupérieure,46,Rued’Ulm,75005Paris,France.e-mail:[e-mail]‘LanguageandSpeech’is©KingstonPressLtd.2003Asbackground,Iassumethattheultimatetargetofphonologicalacquisitionisacognitivearchitecturewithmultiplelevelsofrepresentation.Theselevelsmin-imallyinclude:(1)Parametricphonetics:Aquantitativemapoftheacousticandarticulatoryspace,onwhichproximityinmultipledimensionscanbedefined.(2)Phoneticencoding:Low-levelcategorizationofthephoneticspace.(3)Thelexicon:Lexicalrepresentationsofword-forms,whichprovidealocusforassociationbetweenformandmeaning.(4)Thephonologicalgrammar:Generalconstraintsonword-formsinthelexiconsuchasconstraintsonmetricalstructureorsegmentalsequencing.(5)Morphophonologicalcorrespondences:Phonologicalrelationshipsamongstmorpho-logicallyrelatedwordswhicharenotindependentlypredictablefromconstraintsonwordform.Consider,forexample,thelexicalizedcompound,theverbblindfold(topreventsomeonefromseeingbyfixingaclothovertheeyes).Theparametricphoneticspaceprovidesawaytorepresentthetimecourseofspectraland/orarticulatoryparametersonanyindividualoccasionofthewordbeinguttered.Inspeechperception,itrep-resentstheperceptualcaptureofthespeechwhichmakesitpossibleforthespeechtobesubmittedtocognitiveprocessingofanykind.Inspeechproduction,itrepresentsamotorplanwithappropriatespecificationintimeandspaceofmotorgestures.Thefactscapturedbythephoneticencodingsystemwouldincludethedifferencebetweentheclear/l/intheinitial/bl/clusterandmorevocalic/l/inthefinal/ld/cluster,aswellasthecontrastbetweentheobligatoryreleaseoftheinitial/b/andtheoptionalreleaseofthemedialandfinal/d/s.Thesearelanguageparticularsub-phonemicdetailswhichnonethelesshavetheirroleinproductionorperceptionoftheword.Thephonologicalgrammarassignsaphonologicalwordboundaryinthemiddleofthewordblindfold,despiteitssemanticopacity.Twocrucialfactorsinthephono-logicalparsearethemedial/df/clusterandthesuperheavyfirstsyllable,whichcontainsadiphthongplustwoconsonants.Thesefeaturesarerareorimpossibleintheabsenceofawordboundary.Lastly,knowledgeofmorphophonologicalcorre-spondencesyieldsthepredictionthattheneologismblindfoldeewouldexhibitashiftoftheprimarystresstothelastsyllable,justasinexamine,examinee.Incontrast,thestresswouldremainonthestemintheneologismblindfolder,asinemploy,employer.Therearesystematiclogicaldependencesamongsttheselevels,andthesedepen-dencesmustbebothexploitedandcreatedduringlanguageacquisition.Attheperipheryofthesystem,encodingthespeechsignaldependsoncapturingitinthefirstplace.Developmentofthelexicondependsontheexistenceofasystemforencodinglexicalitems.Generalizationsaboutword-formsdependonknowingasuf-ficientnumberofwords.Knowledgeofmorphophonologicalrelationslikewisedependsonhavingasufficientvocabulary,andasufficientknowledgeofsyntacticandsemanticrelatio