Abstract Analyzing language development from a net

1AnalyzinglanguagedevelopmentfromanetworkapproachJinyunKEYaoYAOEnglishLanguageInstituteDepartmentofLinguisticsUniversityofMichigan,AnnArborUniversityofCalifornia,[e-mail]@berkeley.eduAbstractInthispaperweproposesomenewmeasuresoflanguagedevelopmentusingnetworkanalyses,whichisinspiredbytherecentsurgeofinterestsinnetworkstudiesofmanyreal-worldsystems.Children’sandcare-takers’speechdatafromalongitudinalstudyarerepresentedasaseriesofnetworks,wordformsbeingtakenasnodesandcollocationofwordsaslinks.Measuresonthepropertiesofthenetworks,suchassize,connectivity,hubandauthorityanalyses,etc.,allowustomakequantitativecomparisonsoastorevealdifferentpathsofdevelopment.Forexample,theasynchronyofdevelopmentinnetworksizeandaveragedegreesuggeststhatchildrencannotbesimplyclassifiedasearlytalkersorlatetalkersbyoneortwomeasures.Childrenfollowdifferentpathsinamulti-dimensionalspace.Theymaydevelopfasterinonedimensionbutslowerinanotherdimension.Thenetworkapproachrequireslittlepreprocessingofwordsandanalysesonsentencestructures,andthecharacteristicsofwordsandtheirusageemergefromthenetworkandareindependentofanygrammaticalpresumptions.Weshowthatthechangeofthetwoarticlestheandaintheirrolesasimportantnodesinthenetworkreflectstheprogressofchildren’ssyntacticdevelopment:thetwoarticlesoftenstartinchildren’snetworksashubsandlatershifttoauthorities,whiletheyareauthoritiesconstantlyintheadult’snetworks.Thenetworkanalysesprovideanewapproachtostudylanguagedevelopment,andatthesametimelanguagedevelopmentalsopresentsarichareafornetworktheoriestoexplore.1.IntroductionChildrenacquiretheirlanguageindifferentways.Variouskindsofmeasurehavebeenusedtocomparethetrajectoriesofdevelopment.Amongthem,vocabularysizeandMeanLengthofUtterance(MLU)aretwobasiconesusedfrequentlyforevaluatingtherateofearlydevelopment.Twotypicalstylesoflearnershavebeenidentified:earlyvs.latetalkers(Bates&Goodman,1997).Earlytalkerstendtohavealargevocabularysizeandlongsentences(i.e.alargeMLU)earlierthanotherchildren,andlatetalkerslagbehindtheaverage.Childrenalsovaryinthedistributionoftypesofwordsintheirearlyvocabularyandtheorderofacquisition.Forexample,somechildrenacquiremorenounsforobjectsattheearlystage,whilesomeusemoreformulaicexpressions,suchas“Lemmesee”,“Don’tdodat”,intheirspeech.Thiscontrastistermedasadichotomybetween“referential”and“expressive”style(Nelson,1973).Differenttypesoflearningstyleshavealsobeenidentifiedinwordlearningandphonologicaldevelopment.(Batesetal.,1988;Shore,1995).Researchhasshownsignificantdifferencesinthedevelopmentinmorphologyandsyntaxaswell.Forexample,childrendifferintheratesandroutesofacquisitionofdifferentgrammaticalmorphemes,suchasthepluralisticfunctionof“-s”inEnglish(Brown1973),aswellasvariousgrammaticalconstructions,suchasauxiliaries(Stromswold,1990),questions(Stromswold,1995),andsoon.Veryoftenthesemeasuresofdifferencesarebasedontextanalysesofcorporacollectedfromspontaneousspeech.Itisoftenpresumedthattheappearanceofconstructioninchildren’sspeechdataindicatestheexistenceoftheknowledgeinthechildrencorrespondingtothegrammarsintheadults.However,suchpresumptionneedtobetakenwithcaution,as2childrenusecertainconstructionsoftenasformulaicexpressions,insteadofknowingtheconstructionsaredecomposableasadultsdo(Peters,1977;Wray,2002).Inthispaper,weproposeanovelapproachtoanalyzethelanguagedevelopmentfromanetworkperspective.Wetakethechildren’slongitudinalspeechdatatoconstructlexicalnetworksatdifferentstages,andcomparethesenetworkswithseveralmeasures.Also,wecomparenetworksbetweendifferentchildren,andnetworksbetweenchildrenandtheircaretakers.Throughthecomparison,children’sdevelopmentandindividualdifferencescanbedemonstratedquantitativelybymeasureswithlittlelanguage-specificassumptions.Networkshavebeenextensivelystudiedwithintheareaofgraphtheoryinmathematics.Randomnetworksusedtobethemainobjectofresearchandwereoftenassumedasthedefaultmodelforreal-worldnetworks.However,recentstudiesonseveralreal-worldnetworks,suchascollaborationnetworks,the(Watts&Strogatz,1998;Barabási&Albert,1999).Ithasbeenfoundthatmanycomplexnetworksintherealworldarelike“asmallworld”(Watts&Strogatz,1998):regardlessofthelargesizeofthenetwork,anytwonodesinthenetworkcanbeconnectedthroughonlyasmallnumberofintermediatenodes,anddirectlyconnectednodesoftensharecommonneighbors(resultinginahigh“clustercoefficient”ofthenetwork).Meanwhile,studieshaveshownthatinmanynetworks,thereareanumberofnodeshavinganextremelylargenumberofconnectionswhilemostnodesonlyhaveahandful;theexistenceoftheso-called“hubs”makethenetworkappeartobescale-free(Barabási&Albert,1999).Thediscoveryofthesefeatureshastriggeredanewsurgeofnetworkresearchinrecentyears(somegeneralaccountsofthedevelopmentofthefieldcanbefoundinBarabási,2002andBuchanan,2002).Thereisanincreasinginterestinnetworksinabroadrangeofdisciplines.Ithasbeenshownthatnetworksofvariousnaturessharethesmall-worldand/orscale-freeproperties,whichimpliestheexistenceofsomeuniversalprincip