A Multifactorial Analysis of Syntactic Variation Particle Movement Revisited

#Swets&Zeitlinger

AMultifactorialAnalysisofSyntacticVariation:ParticleMovementRevisitedÃ

StefanT.Gries

SouthernDenmarkUniversity,Denmark

ABSTRACT

ThepresentpaperinvestigatesthewordorderalternationofEnglishtransitivephrasalverbssuchas,e.g.,topickupthebookversustopickthebookup.Itbuildsontraditionalmono-factorialanalyses,butarguesthatpreviouslyusedmethodsofanalysisaregrosslyinadequatetodescribe,explainandpredictthewordorderchoicebynativespeakers.Ahypothesisintegratingvirtuallyallrelevantvariableseverpostulatedisproposedandinvestigatedfromamultifactorialperspective(usingGLM,lineardiscriminantanalysisandCART).Asaresult,morethan84%ofnativespeakers'choicescanbepredicted.Furtherimplications(linguisticandmetho-dological)arediscussed.

INTRODUCTION

Anotoriouslydif®cultproblemforsyntacticresearchistheexistenceofsyntacticvariation,i.e.closelyrelatedsyntacticvariantswithtruth-condition-allyequivalentmeanings.ExamplesinEnglishincludethewell-knownwordorderalternationsDativeMovement,PrepositionStrandingandParticleMovementin(1),(2)and(3)respectively.1(1)

(a)John[VPgave[NPthebook][PPto[NPBill]]].(b)John[VPgave[NPBill][NPthebook]].

*Addresscorrespondenceto:StefanT.Gries,SouthernDenmarkUniversity,Institutfor

Â150,DK-00Sùnderborg.ErhvervssprogligInformatikogKommunikation,GrundtvigsAlle

E-mail:stgries@sb.sdu.dk1Thegrammaticalnotationisnotcommittedtoanyparticulargrammaticalframeworkandservesexpositoryreasonsonly.Likewise,thechoiceofterminologyintermsofmovementprocessesisnotmeanttotrulyimplyanysuchprocessesÐitmerelyre¯ectsthatthesephenomenahavemostfrequentlybeendealtwithwithinthetransformational-generativeparadigm.

34(2)(3)

S.T.GRIES

(a)Whodidyou[VPseeBill[PPwithti]]?(b)[PPWithwhom]ididyou[seeBillti]?(a)Johnpickedup[NPthebook].(b)Johnpicked[NPthebook]up.

Severalinterrelatedquestionsarisewithrespecttotheseconstructionalalternations:

󰀏󰀏󰀏

howdothetwoconstructionalvariantsofeachpairdifferfromeachother?whyandtowhatextentdodifferentvariablesin¯uencethesubconsciouschoiceofconstructionbynativespeakersinanaturalsetting?

howcannativespeakerchoicesofconstructionsinaparticulardiscoursesituationbepredicted?Morespeci®cally,whichtechniquesaremostsuitableforthepredictionofnativespeakers'choicesgiventhecomplexityofnaturaldiscoursesettings?

Thisstudyinvestigatesthesequestionsforthelastoftheabove-mentionedwordorderalternationsoftransitivephrasalverbs;theconstructionwherenoelementintervenesbetweenverbandparticlewillbereferredtoasconstruc-tionÐtheconstructionwhereadirectobjectNPintervenesbetweenverbandparticleisreferredtoasconstruction1:2(4)

(a)Johnpickedupthebook.(b)Johnpickedthebookup.

construction0construction1Section2isconcernedwithabriefsummaryofpreviousanalysesofParticleMovement.Atthesametime,severalmethodologicallymotivatedpointsofcritiqueareraised.Section3outlinesthemethodsbymeansofwhichtheabovethreeobjectivesarepursued.Section4dealswiththeresultsofthestudy:monofactorialaswellasmultifactorialresultswillbepresentedinsomedetail.Finally,Section5concludesbysituatingthestudywithinabroaderpsycholinguisticframeworkandbybrie¯ydiscussingfurtherimplicationsgoingbeyondtheimmediatescopeofParticleMovement.

ÐÐÐÐÐÐÐÐÐÐÐÐÐÐ2Thereasonforthisseeminglyarbitraryandcounter-intuitivelabellingwillbeaddressedlater;asamnemonichelp,considertheindextonamethenumberofcontituentsinterveningbetweenverbandparticle.

AMULTIFACTORIALANALYSISOFSYNTACTICVARIATION

35

PREVIOUSANALYSES

Variablesthatpurportedlygovernthealternation

Thepositionofparticlesintransitivephrasalverbshasbeeninvestigatedtimeandagainwithinthelast100years.Theapproachescomefromwidelydi-verginglinguisticschoolsofthoughtsuchasChomskyantransformational-generativegrammar(cf.,e.g.,Fraser,1974,1976;DenDikken,1992,1995;Rohrbacher,1994,tonamebutafew),traditionalgrammarians(Sweet,12;Jespersen,1928;Kruisinga&Erades,1953),cognitivegrammar(Yeagle,1983),discourse-functionalorientedapproaches(Chen,1986),psycholinguis-tically-orientedapproaches(Hawkins,1994)etc.Overthetime,variousvariableshavebeenproposedinordertoaccountforboththeoptimalstructuralcon®gurationofthetwoconstructionsandthequestionofwhichconstructionischosenbynativespeakers.Table1providesanoverviewover

Table1.

Variablesthatallegedlygovernthealternation.

Variable

Valueforconstruction1Valueforconstruction0LongDOLongDOComplex

LengthoftheDOinwords(LengthW)LengthoftheDOinsyllables(LengthS)ComplexityoftheDO(Complex)NP-TypeoftheDO:semi-pronominal

(Type)

Inde®niteDetermineroftheDO(Det)NoPreviousmentionoftheDO(Lm)Low3ÐÐÐTimesofprecedingmentionoftheDO(Topm)ÐÐÐÐ\"High3ÐÐDistancetolastmentionoftheDO(Dtlm/ActPC)ÐÐÐÐ\"High3ÐÐÐÐÐÐNewsValueoftheDOÐÐÐÐÐÐÐÐÐÐÐ\"Yes(Contrastive)StressoftheDOYesSubsequentmentionoftheDO(NM)High3ÐÐÐTimesofsubsequentmentionoftheDO(Tosm)ÐÐÐÐ\"How3ÐDistancetonextmentionoftheDO(Dtnm/ClusSC)ÐÐÐÐ\"

OverallfrequencyoftheDO(OM)followingdirectionaladverbial(PP)

YesPrepofthefollowingPPisidenticaltothe

particle(Part󰂈Prep)

Register

Idiomatic3ÐÐÐÐÐMeaningoftheVP(Idiomaticity)ÐÐÐÐÐÐÐ\"Low3ÐÐÐÐÐCognitiveEntrenchmentoftheDOÐÐÐÐÐÐÐ\"InanimateAnimacyoftheDO(Animacy)AbstractConcretenessoftheDO(Concreteness)

pronominalde®nite

yeshighlowlownolowhighyes

literalhighanimateconcrete

36

S.T.GRIES

themultitudeofvariablesproposedsofar.Thecentralcolumnnamesthevariableproposedwhereastheleftandrightcolumnnamethevalues/levelsofeachvariableassociatedwithaparticularpreferenceforaconstruction.Commentsandpointsofcritique

Thislistofvariablesmayseemquiteimpressiveat®rst.Itisespeciallyinterestingtonotethatata®rstsuper®cialglancequitesimplewordorderalternationisinfactin¯uencedbyvariablesfrommanydifferentsub-branchesoflinguistics:phonology,syntax,semantics,pragmaticsandothervariables.Unfortunately,however,therearealsoseveralshortcomingsthathavehin-deredprogressconsiderably.

Firstofall,mostvariablesarebasedonintrospectiveanalysis(i.e.,acceptabilityjudgements)andnon-authenticexamplesentences.Whiletherearesomelinguisticframeworkswhichconsiderthistobearewardingwayofgatheringdata,Iwouldcontendthat(i)acceptabilityjudgementsveryoftendonotnecessarilyconstituteobjective,reliableandvaliddata(cf.,e.g.,Labov,

Ètze,1996);(ii)itisquestionablethatananalysissolelybasedon1975;Schu

dreamt-upsentencescaninfactobtainrepresentativeresults;and(iii)`noonehaseverpresentedevenahintofevidencethatanypartofthehuman'slinguisticcompetenceistheabilitytoevaluatesentencesproducedarti®cially,outofcontext'(Prince,1991,p.80).

Second,mostanalysesonlyconsidermonofactorialresults(i.e.,theeffectonevariablehasonthealternationinisolation)althoughforthespeakerallvariablesaregivensimultaneously.Forinstance,Fraser(1974)arguedthatverbswithoutinitialstresspreferconstruction1,offeringthefollowingsentencesaswhatheclaimstobesupportingevidence.(5)(6)(7)

(a)?Iwillinsultbacktheman.(b)Iwillinsultthemanback.

(a)?Weconvertedovertheheatingtosteam.(b)Weconvertedtheheatingovertosteam.(a)?Theyattachedupthetagonthewall.(b)Theyattachedthetaguponthewall.

Butwhatisproblematicaboutthisapproach?Isthisnotanexampleofoneofthemosttraditionalandwell-establishedmethodsinlinguistics,namelytheminimal-pairtest?Theproblemliesinthefactthattheexamplesdonot

AMULTIFACTORIALANALYSISOFSYNTACTICVARIATION

37

warrantthisclaimatall:thepreferenceforconstruction1intheseexamples(ifthereisoneatall,recallthescepticismexpressedaboveconcerningsuchisolatedacceptabilityjudgements)neednotberelatedtoFraser'sclaimatallandmightaswellderivefromthefactthatshortandsimpledirectobjectsalreadyfavourconstruction1,asdode®nitedeterminersandliteralVPmean-ings(cf.Table1).Ifwegeneralizefromthisphenomenontootheranalyses(whichwecando:nearlyallpreviousanalysesaremonofactorial)we®ndthat,giventhecomplexityof20orsointeractingvariables,wecannotrelyonmonofactorialanalysestodescribeParticleMovementadequately.

Finally,itisgenerallyacceptedthat,normally,sciencetriestodescribe,explainandpredictphenomena.WithParticleMovement(andmanyothercasesofsyntacticvariation),however,themostrigoroustestofone'stheory,namelytheactualpredictionofspeakers'behaviour,hasneverbeenattempted.Everyanalysishasaimedatdescribingparticleplacementatleasttosomeextent;someanalyseshaveaimedatexplainingparticleplacement,butthereareonlyfewanalysesaimingatsubsumingthevariablesunderacommon(setof)factor(s);noanalysishasaimedatpredictingparticleplacementinnaturaldiscoursesituations.Thisisofcourseaconsequenceofthepreviouslymentionedshortcomings:If,forinstance,oneisnotabletoquantifytheimportanceoftheindividuallyproposedvariables,thentradi-tionalaccountswouldalreadyfailtopredictconstructionalchoiceswhenonlytwovariableshavecon¯ictingpreferences.ConsiderJohnpickedupabook.Theshortdirectobjectprefersconstruction1whereastheinde®nitedeterminerprefersconstruction0.Evidently,withoutawaytoweighindividualvariables'preferences,traditionalaccountscannotevenpredictspeakers'preferencesinthesimplecaseswhereonlytwovariablesareconcerned,whichiswhysofarnobodyhasmanagedtopredictspeakers'choicessimultaneouslyaccountingformorethanadozenvariables.

Asisevidentfromthethreeabove-mentionedresearchquestions(cf.section1)Iintendtoovercometheseshortcomings.ThefollowingsectionisconcernedwiththemethodsIusetothatend.

METHODS

TheProcessingHypothesis(PH)

Inordertoexplainwhyspeakerschoosetheconstructiontheydo,Iproposethefollowinghypothesis:themultitudeofvariables(mostofwhichare

38

S.T.GRIES

concernedwiththedirectobjectNP)thatseemstoberelatedtoParticleMovementcanallberelatedtotheprocessingeffortoftheutterance.3Myideaofthenotionofprocessingeffortisafairlybroadone:itencompassesnotonlypurelysyntacticdeterminants,butalsofactorsfromotherlinguisticlevels.Morespeci®cally,Iassumethatvirtuallyalllevelsoflinguisticdescriptionmentionedabovecancontributetoprocessingeffort:

󰀏

󰀏

󰀏

phonologicallyindicatedprocessingcost:contrastivestressonalinguisticexpressionincreasestheamountofprocessingeffortbecausethespeakerfocuses(whichisitselfnotaneffortlesstask)thehearer'sattentiononthereferentofthecontrastively-stressedexpression;

morphosyntacticallydeterminedprocessingcost:thelongerandthemorecomplexthedirectobjectNPis,themoreeffort(andworkingmemory)isneededtoprocesstheutterancecorrectly;

semanticallyconditionedprocessingcost:ifthemeaningoftheVPisidiomatic,thenthemeaningofthewholeofthetransitivephrasalverbisnotcomputablefromthemeaningoftheindividualpartssothatthepartsoftheidiomaticphrasalverbsrelymoreononeanotherthanwithtotallyliteralphrasalverbs(whichmostlydesignatecausedmotion).Following,sayinHawkins(2000),wemayassumethatthereisatendencytominimizewhathereferstoaslexicaldependencydomains(LPDs),i.e.(slightlysimpli®ed)thedistancebetweenexpressions(attimesmutually)dependentononeanotherfortheirinterpretation.Withidiomaticphrasalverbs(e.g.,toekeout),thesemanticdependencybetweenverbandparticleishigherthanthedependencyforliteralverbs(e.g.,tobringback),sowewouldaccordinglyexpectconstruction0,minimizingthedistancebetweenthecomponentpartsofthephrasalverb,whereaswithliteralphrasalverbs,nopreferenceforaconstructionistobeexpectedbecausethelowdegreeofinterdependencedoesnotrequireaparticularlysmalldistancebetweenthecomponentpartsand,thus,licensesbothwordorders.Inconnectionwiththat,concreteDOsaremorelikelytocorrelatewithaliteralinterpretationoftheverb-particleconstructionsincethesereferentscanundergothecausedmotionthatverb-particleconstructionscommonlydenoteÐabstractDOs,ontheotherhand,

ÐÐÐÐÐÐÐÐÐÐÐÐÐÐ3ThequestionarisesastowhetherIrefertotheprocessingeffortofthespeakerorthehearer.Myownfocusisonthespeaker'sprocessingeffortÐwithParticleMovement,however,we®ndthatwhatmakesprocessingef®cientforspeakersisalsobene®cialtohearers.Thus,nostrictdifferentiationbetweenthetwointerlocutorsisnecessaryhere.

AMULTIFACTORIALANALYSISOFSYNTACTICVARIATION

39

󰀏

giverisetolessliteralinterpretations(tobringbackpeaceisnotacaseofliteralcausedmotion).Thus,concreteness/abstractnessoftheDO'sreferentscorrelateswiththeliteralness/idiomaticityoftheverb-particleconstructionsandyieldsidenticalpredictions.

discourse-functionallydeterminedprocessingcost:ifthereferentofthedirectobjectNPisdiscourse-givenorcanbeinferredfromtheprecedingcontext,thenitsactivationandproductionincurslessprocessingcostthantheactivationandproductionofsomediscourse-neworevencompletely

Ân,1992).Amorphosyntacticphenomenonunknownreferent(cf.,e.g.,Givo

stronglycorrelatingwithdegreeofgivennessofNPreferentsisthechoiceofdeterminer.Itiswidelyacknowledgedthatinde®nitedeterminersaretypicallyusedfornewreferentswhilede®nitedeterminersaremoreoftenfoundwithgivenreferents.

Since(i)speakersstrivetocommunicatewhatevertheyintendtocommunicatewithaslittleeffortaspossibleand(ii)construction0isinherentlyeasiertoprocess(cf.Hawkins,1994;Rohdenburg,1996),theywilltendtouseconstruction0insituationswheretheprocessingeffortassociatedwiththeutteranceisalreadyhigh.Inotherwords,iftheVPdoesnotrequirealotofprocessingeffort(duetoitslimitedlength,thedegreeofactivationoftheDO'sreferent,etc.)thenconstruction1ischosenÐiftheVPrequiresalotofprocessingeffort(duetotheprocessingcostfortheDO'sreferent)thenconstruction0ischoseninordertofacilitatecommunicationbyminimizingthestructurallydeterminedprocessingeffort.Note,however,thatthishypoth-esisimpliesthatsomeofthevariablesmentionedabovewillnotberelevantforthechoiceofconstructionsincethereisnoreasonwhyvariablesconcernedwiththediscoursefollowingtheverb-particleconstructionshouldplayarolejustasthereisnoreasonwhytheanimacyofthedirectobject'sreferentshouldbeimportant.Finally,contrarytoapreviousanalysis(cf.Gries,1999),thevariableofentrenchmentisnotconsideredtoberelevantsincethevariablesthat,takentogether,constitutetheentrenchmenthierarchyusedpreviouslyareinvestigatedhereseparatelyandthusmuchmoreprecisely(cf.Gries,2000foramoredetailedstatisticalanalysisofthesevariables).

Thedata

Inordernottorelyonmade-upsentencesandtheir(attimes)doubtfulacceptabilityjudgements,Iadvocatetheuseofnaturally-occurringdata.Ihave,therefore,compiledasampleof403utteranceswithverb-particle

40

Table2.

S.T.GRIES

DatafromtheBritishNationalCorpus.

Construction0Construction113376209

Rowtotals

200203403

SpokenWritten

Columntotals67127194

constructionsfromtheBritishNationalCorpus.Theverb-particleconstruc-tionschosenmainlyconsistofcombinationsofthemostfrequentverbsandparticlesenteringintotransitivephrasalverbs.4Table2showsthedistributionofmydata.

Toeachofthesesentencesthe10precedingandsubsequentclauses(withoutfalsestartsordiscoursemarkerssuchasYouknoworImean)wereadded.Then,eachsentencewasinvestigatedwithrespecttothevariableslistedaboveinTable1.5Thetableresultingfromtheseprocesseswasthebasisfortheanalysistofollow.

Statisticaltechniques

Firstofall,foreachvariableamonofactorialcorrelationwascomputed.Dependingonthemeasurementscaleoftheindependentvariables,thecoef®cientsgiveninTable3weredetermined.6Itshallbenoted,however,thatthemonofactorialcorrelationsareonlydeterminedinordertotestpreviousmonofactorialanalysesempirically,mostofwhichhavenotbeenempiricallytestedbefore.Theprimarypurposeofthispaper,i.e.,thepredictionofspeakers'choices,canonlybeachievedwithmoreelaboratetechniques.Themultifactorialtechniquesthatwillbeusedarethe

ÐÐÐÐÐÐÐÐÐÐÐÐÐÐ

Ihavedeterminedthemostfrequentverbsandparticlesintransitivephrasalverbsonthebasisofmyowncollectionof1,357transitivephrasalverbsfromseveraldictionaries.5Thedegreesofcomplexityandidiomaticityweremeasuredonordinalscales:simpleNPs,intermediateNPs(NPswithmodi®cationbyadjectivesand/orgenitives)andcomplexNPs(containingembeddedclauses)forcomplexityandsimple/literal,metaphorical/®gurativeandidiomatic/opaqueVPs.6Therankingisroughlyaccordingtothesizeofthecorrelation(butcf.below).Whilethetwovaluesoffandldonotalwayscoincide,thesizesofallcorrelationcoef®cients(oncewiththefcoef®cient,oncewithl)correlatehighlysigni®cantly(g󰂈0.85;z󰂈6.923;p`0X001ÃÃÃ),whichiswhytheseminorrankingdifferenceswillnotbedealtwith.

4AMULTIFACTORIALANALYSISOFSYNTACTICVARIATION

41

Table3.Monofactorialcorrelationscomputedforthecorpusdata.

Correlationcoef®cient

Phi/Cramer'sIandLambda

g(equallingKendall'stwithcorrectionforties)Pearsonproduct±momentcorreation

Measurementscaleoftheindependentvariable

Nominal/categoricalOrdinalInterval

generallinearmodel(GLM),lineardiscriminantanalysis(LDA)andclassi-®cationandregressiontrees(CART).

RESULTS

Monofactorialresults

Table4summarizesthemonofactorialcorrelationcoef®cientsofeveryindependentvariable;fornominalandordinalvariables,thecorrelationsbetweenindividuallevelsandthechoiceofconstructionarealsoprovided.7ThePHisstronglysupportedby

󰀏

󰀏

thosevariables/valuesthatindicatealowdegreeofprocessingcost(e.g.,literalVPswithpronominalDOsorshortlexicalDOswithde®nitedeterminerswheretheDOhasbeenmentionedbeforefrequently)favourconstruction1;

thosevariables/valuesthatindicateahighdegreeofprocessingcost(e.g.,idiomaticVPswithdiscourse-newlexicalreferentsoflongDONPswithinde®nitedeterminers).

Onthewhole,themorphosyntacticvariablesaremostin¯uential,semanticanddiscourse-functionalvariablesarelesspowerfulindeterminingthechoiceofconstruction.However,threecommentsarenecessary.First,giventhe

ÐÐÐÐÐÐÐÐÐÐÐÐÐÐ

Therankingisroughlyaccordingtothesizeofthecorrelation(butcf.below).Whilethetwovaluesoffandldonotalwayscoincide,thesizesofallcorrelationcoef®cients(oncewiththefcoef®cient,oncewithl)correlatehighlysigni®cantly(g󰂈0.85;z󰂈6.923;p`0X001ÃÃÃ),whichiswhytheseminorrankingdifferenceswillnotbedealtwith.

742

Table4.

S.T.GRIES

Monofactorialcorrelationsbetweenvariablesandthechoiceofconstruction.

Correlationcoef®cientg󰂈À0.85***g󰂈À0.6***

f󰂈0.522***(l󰂈0.49)f󰂈0.492***(l󰂈0.366)rpbis󰂈À0.481***

f󰂈0.47***(l󰂈0.366)f󰂈0.468***(l󰂈0.32)f󰂈0.455***(l󰂈0.412)rpbis󰂈0.452***rpbis󰂈0.429***rpbis󰂈À0.423***rpbis󰂈0.414***

f󰂈0.411***(l󰂈0.387)rpbis󰂈0.357***

f󰂈0.339***(l󰂈0.314)f󰂈À0.328***(l󰂈0.253)f󰂈0.319***(l󰂈0.206)f󰂈0.314***(l󰂈0.268)f󰂈0.291***(l󰂈0.263)f󰂈À0.288***(l󰂈0.206)f󰂈0.284***(l󰂈0.16)f󰂈0.232***(l󰂈0.191)f󰂈À0.193***(l󰂈0.077)rpbis󰂈0.191***

f󰂈0.166***(l󰂈0.057)rpbis󰂈0.142**

f󰂈0.104*(l󰂈0.072)rpbis󰂈0.1*

f󰂈0.092***(l󰂈0)

f󰂈À0.047ns(l󰂈0.016)f󰂈0.023ns(l󰂈0)f󰂈À0.018ns(l󰂈0)f󰂈0.003ns(l󰂈0)

Variable/Variable:Value

ComplexityoftheDOIdiomaticityoftheVPComplex:simpleNPNPTypeoftheDO

LengthofthedirectobjectinsyllablesType:lexicalNPType:pronominalNP

Complex:intermediateNP

DistancetolastmentionoftheDO

CohesivenessoftheDOtotheprecedingdiscourseLengthoftheDOinwords

TimesofprecedingmentionoftheDOLastmentionoftheDOOverallmentionoftheDOConcretenessoftheDOIdiomaticity:idiomaticVPDetermineroftheDOIdiomaticity:literalVPRegister

DET.inde®nitedeterminer

DirectionaladverbialfollowingtheDODET.nodeterminerComplex:complexNP

TimesofsubsequentmentionoftheDOAnimacyoftheDO

CohesivenessoftheDOtothesubsequentdiscourseNextmentionoftheDO

DistancetonextmentionoftheDOType:semi-pronominalNPIdiomaticity:metaphoricalNPType:propername

DET:de®nitedeterminer

ParticleequalstheprepositionofthefollowingPP

mathematicallydifferentwaysofcalculatingthesecoef®cients,itisnotpossibletosimplycomparethevariables'powerbysimplycomparingtheabsolutevaluesofthecorrelationcoef®cients.Second,somevariablesthatwerepredictednottocorrelatesigni®cantlywithchoiceofconstructiondoinfactdisplayasigni®cantcorrelationsofurtherinvestigationiscalledfor.

AMULTIFACTORIALANALYSISOFSYNTACTICVARIATION

43

Lastly,aswasarguedabove,amonofactorialperspective(i)doesnotdojusticetothecomplexityofthephenomenonand(ii)de®esanycognitivelyrealisticaccountofthealternation.

Multifactorialresults:GLMandLinearDiscriminantAnalysis(LDA)ThemultiplecorrelationbetweenallvariablesincludedbythePHandthechoiceofconstructionasdeterminedbytheGLMishighlysigni®cant:r󰂈0X786;F71Y331󰂈7X512;p`0X001ÃÃÃ.Giventhelargenumberofinter-correlationsbetweenthepredictorvariables,however,thiscorrelationcoef®-cientneedstobeadjustedforshrinkageusingWherry'sformula;radjusted󰂈0X732.Still,thisvalueisstillquitehighandstronglysupportsthePH.Itdoessoespeciallywhenweconsiderthefollowingtwopoints:

󰀏

󰀏

RobtainedonthebasisofthePHisevenlargerthanRobtainedwhenweincludeallvariablesmentionedinTable1(namely,Radjusted󰂈0X718;F126Y276󰂈4X4;p`0X001ÃÃÃ),whichshowsthatthevariablesIhavechosentoeliminateonlyaddrandomnoisetotheanalysisanyway;

multiplecorrelationsthatareobtainedinotherbehaviouralsciencesareoftenmuchsmallersotheaccountofvarianceaccountedforinthepresentstudyiscomparativelylarge.

Butwhataboutthepredictivepowerofmyhypothesis?AnLDAshowsthatthevariablesincludedinthePHmakeitpossibletopredictwhichconstructionaspeakerwillchooseinaparticulardiscoursesituation.Thediscriminantfunctionishighlysigni®cant(canonicalr󰂈0X73;w2󰂈297X37;dXfX󰂈20Yp󰂈0ÃÃÃ).Moreover,thediscriminantfunctioncanclassify86.1%oftheconstructionalchoiceswithinthesample.However,itismoreimportanttoalsocross-validatethisresultinordertoavoidcircularityofreasoningbyusingcasesfortheirown`prediction'.Twomeasureswerethereforetakentoimprovetheanalysis:

(a)theleave-one-outmethodforcross-validation,yieldingaprediction

accuracyof84.1%;aresultthatisvirtuallyimpossibletoobtainbypurechance,accordingtotheexactbinomialtest,thechancefor339correcthitsin403trialsapproacheszero;

(b)thesplit-sampletechnique,whereIdividedthecorpusdataintoa

learningsampleandapredictionsampleinordertoderiveadiscriminantfunctionfromthelearningsamplewhichwassubsequentlyappliedtothepredictionsample.Inordernottobeaccusedofapossiblybiasedchoice

44

Table5.

S.T.GRIES

Cross-validatedpredictionaccuracyofLDAforsplitsamples.

Predictionsample53writtensentences53spokensentences26spokensentencesand27writtensentences

Correctpredictionsforpredictionsample

84.9%;

󰂅pbinomialtest%1X184Â10À7ÃÃÃ󰂆.2%;

󰂅pbinomialtest%0X027Ã󰂆75.5%;

󰂅pbinomialtest%1X343Â10À4ÃÃÃ󰂆74.8%

Learningsample200spokensentencesand150writtensentences150spokensentencesand200writtensentences174spokensentencesand176writtensentencesAverage

ofsamples,thiswasdonethreetimeswithrandomlychosensentencesfromthedifferentregisters.Table5givesanoverviewoftheresults.Obviously,theresultsarequiterobust:thepredictionaccuraciesareallsigni®cant,asdeterminedbytestingthecorrecthitrateagainsttheoneexpectedbypurechanceusingtheexactbinomialtest.8Thebestpredictionresultsareachievedforwrittendata,theworstfororaldata,whichistobeexpected,giventhemorespontaneousandinteractivenatureofnaturaldiscourseasopposedtoplannedwriting.

Letusnowtryto®ndoutwhichvariablesareresponsibleforthegooddiscriminationbetweenthetwoconstructions.ThepreviousresultswereconcernedwithanLDAwhere,fortheoreticalreasons,onlythevariablesofthePHwereincluded.Butwealsoneedto®ndoutwhetheritisempiricallyplausibletoexcludesomevariablesfromfurtherconsideration,beitonlytosupporttheresultsobtainedbytheGLM.Thus,asecondLDAwascompu-tedwhereallvariableswhereincluded;Table6summarizesitsresults:thehighertheabsolutevalueofafactorloadingforavariablethemoreimportantitisforthechoiceofconstructionintheonlycognitivelyrealisticanalysisofthesituation,namelywhenallofthevariablesareconsideredsimultaneously.

ÐÐÐÐÐÐÐÐÐÐÐÐÐÐ

Onemightwonderwhythesplit-sampletechniqueyieldsslightlyworseresultsthantheleave-one-outmethod.Thisisduetothefactthatthelearningsamplesfortheleave-one-outmethodcontain52sentencesmorethanthoseofthesplit-sampletechnique.

8AMULTIFACTORIALANALYSISOFSYNTACTICVARIATION

45

Table6.Variable

Factorloadingsofthediscriminantanalysis.

Factorloading

0.5220.4980.4790.4470.3250.2810.1840.0440.0160.0060.002À0.021À0.086À0.094À0.098À0.135À0.157À0.183À0.223À0.278À0.309À0.337À0.358À0.422À0.427À0.445À0.474À0.496À0.573

KindofvariableMorphosyntactic

ChoiceofconstructionHighvariablevaluesAconstruction0LowvariablevaluesAconstruction1Duetothelowfactorloadings(`0X22)thesevariablesdonotdiscriminatewellbetweenthetwoconstructions

LengthSType:lexical

Complex:intermediateLengthW

Idiom:idiomaticDET:inde®niteComplex:complexIdiom:metaphoricalDET:de®niteDis¯uencyPart󰂈Prep

Type:propernameType:semipronominalClUSSCNM

COHSCAnimacyTOSM

DET:nodeterminerPP

Idiom:literalConcreteOMLMTOPMCOHPCACTPC

Type:pronominalComplex:simple

Semantic

MorphosyntacticMorphosyntacticSemantic

MorphosyntacticOtherOther

MorphosyntacticDiscourse-functional(subsequentcontext)Semantic

Discourse-functional(s.c.)

MorphosyntacticOtherSemantic

Discourse-functionalDiscourse-functional(precedingcontext)Morphosyntactic

HighvariablevaluesAconstruction1LowvariablevaluesAconstruction0Obviously,thePHisagainstronglysupported.Notonlydowe®ndthatallvariablesincludedinthePHcorrelatewiththechoiceofconstructionaspredictedÐinthemultifactorialanalysis,weseethatthevariablesconcernedwiththesubsequentcontextareindeedirrelevant,aswaspredictedbythePH.Onthewhole,we®ndthefollowingrankingofstrengthofvariablegroups:

46

S.T.GRIES

discourse-functionalvariables(precedingcontext),syntacticvariables,semanticvariablesandothervariables.9Multifactorialresults:Classi®cationandRegressionTrees(CART)

WhiletheresultsoftheLDAarequiteconvincing,thereisoneobjectionconcerningtheapplicationofanLDAthatmightberaised.ItisconcernedwiththestandardassumptionthatanLDArequiresamultivariatenormaldistributionofthedata,andonemight(correctly)claimthatitisdoubtfulthatmydatadoindeedmeetthisdemandand,thus,thattheaboveresultsaretobetakenwithagrainofsalt.However,thereareseveralargumentssupportingtheaboveanalysis,resultsandinterpretationeventhoughthedistributionalassumptionsofLDAsarenotmet.

First,whilemanyresearcherstendtoemphasizetheimportanceofdistributionalassumptions(suchasnormality,homogeneityofvariancesandthelike),anumberofscholarsarguethat,inpractice,theseassumptionsarenotasessentialastheymightseemonapurelymathematicalbasis(cf.Wineretal.,1991,p.5).Second,ithasevenbeenclaimedthatthereisnotestthatreliablyidenti®esmultivariatenormaldistributions(cf.Bortz,1999,p.435).Third,thedifferencebetweenLDAandCARTisofcoursenotjustastatistical/mathematicaloneÐrather,thereisalsoaconceptualdifference:whileanLDAincludesallvariablessimultaneouslyinthecalculationtocomputeapredictionforoneofthetwoconstructionalchoices,thetreeresultingfromCARTanalysesincludesvariablessequentially.Foranativespeakerhowever,IbelievethatthemodelunderlyingLDAismorerealistic.Itisintuitivelymoreplausibletoassumethatallthevariables'values/levelsIhavediscussedaresomehowsetatthepointoftimethespeakerchoosesthewordorderratherthanthatthevalues/levelsareincludedonebyoneinasequentialfashion.Moreover,whilethereisstillconsiderabledebatewhetherpsycholinguistictheoriesofspeechproductionshouldincorporateparallelorserialmodelsofprocessing,I,followingBerg(1998),considerparallelprocessingtheoriesmorerewardingfrombothatheoreticalandapracticalpointofview.Ihavedecided,forthesereasons,topredictnativespeakers'choiceswithanLDAwhich,asopposedtoCART,comesclosesttopredictingchoicesonthebasisofasimultaneous/parallelinclusionoftherelevantdata.

ÐÐÐÐÐÐÐÐÐÐÐÐÐÐ9Thiswasdeterminedbycalculating(i)theAMsoftheabsolutevaluesofthefactorloadingsforeachvariablegroupand(ii)themediansoftheranksofallvariablesinagroupwhenthevaiablesareorderedaccordingtotheirfactorloadings.Bothresultswereidentical.

AMULTIFACTORIALANALYSISOFSYNTACTICVARIATION

47

Table7.Parameter

ParametersandsettingsoftheCARTanalysis.

Setting

CART-styleexhaustivesearchforunivariatesplitsFACT-styledirectstopping:fractionofobjects󰂈0.5identical:construction0:p%0X5

construction1:p%0X5

Ginimeasure

MethodStoprule

PriorprobabilitiesGoodness-of-®tindex

Nevertheless,itmightverywellbethecasethatthesereasonsdonotsatisfytrulymathematically-orientedresearchers.Ihave,therefore,alsoanalyzedmydatausingtheCARTmoduleofStatistica5.5;thealgorithmsusedthereinarebasedonCARTbyBreimanetal.(1984),whereCARTandQUESTalgorithmsareusedtoclassifyandpredictdataintheabsenceofdistributionalassumptions.MyCARTanalysisofthedatawasbasedontheparametersandsettingsgiveninTable7.

Theresultoftheanalysiscanbesummarizedasfollows:outofall403sentences,349(86.6%)wereclassi®edcorrectlywhile(13.4%)wereclassi®edwrongly,againaresultthatisextremelyunlikelytobeobtainedrandomly.Again,however,wemustalsodeterminethepredictionaccuracybyacross-validationtechnique.Sincetheleave-oneoutmethodforCARTisnotavailableinStatistica5.5,Iusedonlythesplit-sampletechniqueanalogoustotheLDA;thesamplesandtheresultsarelistedinTable8.

Admittedly,thecross-validatedpredictionaccuracyofCARTisnotashighastheLDAresults,but,apartfromthepredictionsamplefororaldataalone,

Table8.

Cross-validatedpredictionaccuracyofCARTforsplitsamples.

Predictionsample53writtensentences53spokensentences26spokensentencesand27writtensentences

Correctpredictionsforpredictionsample

81.1%;

󰂅pbinomialtest%2X775Â10À6ÃÃÃ󰂆56.6%;

󰂅pbinomialtest%0X205ns󰂆77.4%;

󰂅pbinomialtest%4X086Â10À5ÃÃÃ󰂆71.7%

Learningsample200spokensentencesand150writtensentences150spokensentencesand200writtensentences174spokensentencesand176writtensentencesAverage

theyarestillwaybetterthanwhatmightbeexpectedbypurechance.Moreover,thereisareasonfortheseminordifferences.Giventheaboveparametersettings,theCARTtechniquedoesnotutilizeallvariablesforthepredictionofachoiceofconstructionbutonlythemostimportantonesasdeterminedbytheanalysis.Thus,forcaseswherevariablesofanoverallminorimportancearedecisive,falsepredictionsaremorelikely.

Asfarastheimportanceoftheindividualvariablesisconcerned,theoverallpicturedoesnotdifferstronglyfromtheresultsoftheLDA.Theoverallran-kingofthevariablegroupsfromCARTisidenticaltothatoftheLDA;forthesakeofcompleteness,Figure1showstheresultsfortheindividualvariables.

SUMMARY

Foreachrelevantvariableeverinvestigated,itwasshownhowitcontributestoparticleplacementinisolation.Moreover,itwasshownhowallofthesevariablestogetheryieldapreferenceforaconstructioninparticulardiscoursesituations.Itisnowpossibletopredictwithquiteahighpredictionaccuracywhataspeakerwillsayifthediscoursesituationhe/sheisengaginginisknowntotheanalyst.

Ahypothesiswasproposedandsupportedthatincludesallrelevantvariablesandthatcorrectlypredictedsomevariablesnottoberelevant.It

AMULTIFACTORIALANALYSISOFSYNTACTICVARIATION

49

couldbeshownthatacognitivelyrealisticapproachtolanguageusagemadeitpossibletosummarizeandextendthepreviousknowledgeonparticleplacement.

Onamethodologicallevel,wehaveseenthattheanalysisofsyntacticvariationcanbene®tconsiderablyfromtheuseofmultifactorialtechniquesjustastheanalysisofregistervariationhaspro®tedfromBiber's(1988)groundbreakingwork.Personally,Iwouldgoasfarastosaythatonlybysuchtechniquescanwestarttoreallydetecthithertounknownpatternsthatarenotalreadyknownfromearlytraditionalgrammarians'works(aswas,unfortu-nately,thecasewithmanyworksonParticleMovement).Wehavealsoseenthatdifferentmultifactorialtechniques,althoughquitedifferentfromoneanotherwithrespecttotheirdistributionalassumptions,yieldcomparableresults.BothLDAandCARTachieveconvincingclassi®cationandpredictionaccuracies,whichalsostronglysupportthePH.Moreover,theindividualvariables'importanceratingsofthetwoproceduresarestrikinglysimilar,so,atleastforthecaseathand,thedifferentmathematicalrequirementsofbothkindsofanalysesdonotseemtoplayavitalrole.

Lastly,itwasatleastbrie¯yhintedatthewealthofinformationthatcanbeobtainedonthelinguisticdataandthewayspeakerspresumablyorganizetheirknowledgeinordertoarriveatconstructionalchoices.ThisisnottosaythatspeakersactuallyperformLDAorCARTanalyses,butitismeanttoimplythatwecanlearnsomethingabouttheimportanceof(groupsof)variablesintheprocessofonlineproductionandanymodeloflanguageproductionshouldbeabletoaccommodatethesefactsincognitively/psychologicallyrealways.Onepossiblemodelthatcanaccommodatethe®ndingsreportedabovenaturallyistheCompetitionmodelbyBatesandMacWhinney(1982,19),wheredifferentvariables'cuestrengthscompeteinordertogettheirpreferencesrecognized.Moreover,thepresent®ndingscanbereadilyinte-gratedintoactivationmodelswherevariableweightings(beitintermsoffactorloadingsorimportancevalues)correspondtoassociationstrengthsorsimilarconcepts.Inthisrespect,theinvestigationoffurthercasesofsyntacticvariationcanprobablyshedlightonthenatureofactivationnetworks.

REFERENCES

Bates,E.,&MacWhinney,B.(1982).Functionalistapproachestogrammar.InE.Wanner&

L.R.Gleitman(Eds.),LanguageAcquisition:TheStateoftheArt.Cambridge:CambridgeUniversityPress.

50

S.T.GRIES

Bates,E.,&MacWhinney,B.(19).Functionalismandthecompetitionmodel.InE.Bates&

B.MacWhinney(Eds.),Thecrosslinguisticstudyofsentenceprocessing.Cambridge:CambridgeUniversityPress.

Berg,T.(1998).Linguisticstructureandchange:Anexplanationfromlanguageprocessing.

Oxford:ClarendonPress.

Biber,D.(1988).Variationacrossspeechandwriting.Cambridge:CambridgeUniversityPress.

ÈrSozialwissenschaftler.5.Ausgabe.Berlin,Heidelberg,NewYork:Bortz,J.(1999).Statistikfu

Springer.

Breiman,L.,Friedman,J.H.,Olshen,R.A.,&Stone,C.J.(1984).Classi®cationandregression

trees.Monterey,CA:Wadsworth&Brooks/ColeAdvancedBooks&Software.

Chen,P.(1986).DiscourseandparticlemovementinEnglish.StudiesinLanguage,10,79±95.DenKikken,M.(1992).Particles.HollandInstituteofLinguisticsDissertations.TheHague:

HollandAcad.Graphics.

DenDikken,M.(1995).Particles:Onthesyntaxofverb-particle,triadic,andcausative

constructions.Oxford:OxfordUniversityPress.

Fraser,B.(1974).ThephrasalverbinEnglish,byDwightBolinger.Language,50,568±575.Fraser,B.(1976).Theverb-particlecombinationinEnglish.NewYork:AcademicPress.

Ân,T.(1992).Thegrammarofreferentialcoherenceasmentalprocessinginstructions.Givo

Linguistics,30,5±55.

Gries,S.T.(1999).Particleplacement:Acognitiveandfunctionalaccount.Cognitive

Linguistics,10,105±146.

Gries,S.T.(2000).Towardsmultifactorialanalysesofsyntacticvariation:Thecaseofparticle

placement.DoctoralDissertation,UniversityofHamburg,FacultyofLanguageSciences.

Hawkins,J.A.(1994).Aperformancetheoryoforderandconstituency.Cambridge:Cambridge

UniversityPress.

Hawkins,J.A.(2000).Therelativeorderingofprepositionalphrasesinenglish:Goingbeyond

manner±place±time.LanguageVariationandChange,11,231±266.

Jespersen,O.(1928).AmodernEnglishgrammaronhistoricalprinciples.London:George

AllenandUnwinLtd.

Kruisinga,E.,&Erades,P.A.(1953).AnEnglishGrammar.Vol.I.Groningen:P.Noordhoff.Labov,W.(1975).Empiricalfoundationsoflinguistictheory.InR.Austerlitz(Ed.),TheScope

ofAmericanLinguistics(pp.77±133).Lisse:ThePeterdeRidderPress.

Prince,E.F.(1991).Onfunctionalexplanationinlingisticsandtheoriginsoflanguage.

LanguageandCommunication,11,79±82.

Rohdenburg,G.(1996).Cognitivecomplexityandincreasedgrammaticalexplicitnessin

English.CognitiveLinguistics,7,149±182.

Rohrbacher,B.(1994).Englishmainverbsmovenever.ThePennReviewofLinguistics,18,

145±159.

Ètze,C.T.(1996).Theempiricalbaseoflinguistics:GrammaticalityjudgementsandSchu

linguisticmethodology.Chicago:UniversityofChicagoPress.Sweet,H.A.(12).AnewEnglishgrammar.Oxford:ClarendonPress.

Winer,B.J.,Brown,D.R.,&Michels,K.M.(1991).Statisticalprinciplesinexperimental

design.3rded.NewYork:McGraw-Hill.

Yeagle,R.(1983).ThesyntaxandsemanticsofEnglishverb-particleconstructionswithoff:A

spacegrammaranalysis.UnpublishedM.A.Thesis,SouthernIllinoisUniversityatCarbondale.