strategic control in word reading: evidence from speeded...

To appearin theJournalof ExperimentalPsychology:Learning,MemoryandCognition

Strategic Controlin WordReading:EvidenceFromSpeededRespondingin theTempoNamingTask

ChristopherT. KelloDepartmentof Psychology

Centerfor theNeuralBasisof CognitionCarnegie Mellon University

David C. PlautDepartmentsof PsychologyandComputerScience

Centerfor theNeuralBasisof CognitionCarnegie Mellon University

To investigatemechanismsof strategic control over responseinitiation in word reading,weintroducethe temponamingtask. Relative to baselineperformancein the standardnamingtask,subjectswereinducedto respondwith fasterlatencies,shorterdurations,andlower levelsof accuracy by instructingthemto time responseinitiation with anexperimentallycontrolledtempo. The effects of printed frequency and spelling-soundconsistency on latencieswereattenuatedin thetemponamingtask,comparedto standardnaming.Thenumberof spelling-sounderrorsremainedconstantwith fastertempos,while thenumberof word, nonword, andarticulatoryerrorsincreased.We interprettheseresultsas inconsistentwith a time criterionmechanismof controlover responsetiming in thetemponamingtask.Instead,weinvoke inputgainasamechanismof controloverprocessingspeedthroughoutthewordreadingsystem.Wesketchhow input gain could accountfor the temponamingresults,aswell assomestimulusblockingresultsthathavebeenusedin thedebatebetweentherouteemphasisandtimecriterionhypothesesof strategic controlin word reading.

It takes roughly 400–600ms for a skilled readerto begin the pronunciationof a single, clearly printedword. This ball park rangecomesfrom a long historyof speededword namingstudiesin which subjectshadbeenaskedto pronouncea printedword “as quickly andaccuratelyaspossible”(or someinstructionsto that ef-fect). The speededword namingtaskhasbeenusedtoexaminea wide variety of theoreticalissues,including:processesthatmaporthographyto phonology(Glushko,1979;Seidenberg, Waters,Barnes,& Tanenhaus,1984),organizationof the lexicon (Forster& Chambers,1973;Frederiksen& Kroll, 1976),semantic,phonological,andorthographicpriming (Forster& Davis, 1991; Tabossi& Laghi, 1992;Taraban& McClelland,1987),sentenceand discourseprocesses(Hess,Foss,& Carroll, 1995;Trueswell,Tanenhaus,& Kello, 1993), readingimpair-ments(Patterson& Behrmann,1997;Stanovich, Siegel,& Gottardo,1997), and readingacquisition(Lemoine,Levy, & Hutchinson,1993; Manis, 1985). In eachoftheseareasof research,a primary sourceof data hascomefrom latenciesin namingtasks.Therefore,under-standingtheprocessesresponsiblefor the initiation of a

We would like to thankFernandaFerriera,StephenLupker,Max Coltheart,JayMcClelland,Alan Kawamoto,oneanony-mous reviewer, and the PDP researchgroup for their help-ful commentsand suggestions.We would also like to thankBen Payne for his help in collecting datafor Experiment3.This researchwassupportedby NIMH grantsMH47566andMH19102. Pleasedirect all correspondenceto: ChristopherKello, HouseEar Institute,2100WestThird St.,Los Angeles,CA 90057.

namingresponseis of generalimportancefor interpret-ing namingdataacrossresearchdomains.

Thestandardmodeof thinking aboutthe initiation ofa naming responseis as follows. A representationofpronunciationis built up over time, basedon theresultsof processingat oneor moreotherlevelsof representa-tion (e.g., lexical, semantic,orthographic,andsyntacticknowledge; Coltheart,Curtis, Atkins, & Haller, 1993;Kawamoto,1988;Plaut,McClelland,Seidenberg,& Pat-terson,1996). When the pronunciationis resolved ac-cordingto somecriterionof completeness,theresponseis initiated. Often, the exact natureof the criterion isleft unexplained,but a commonassumptionis thata re-sponseis initiated assoonasan entirepronunciationiscompleteby somecriterion (but seeKawamoto,Kello,Jones,& Bame,1998). For example,activation or sat-urationthresholdshave beenused(e.g.,Coltheartet al.,1993;Plautetal., 1996).

Onereasonwhy issuesof responsegenerationareof-tenneglectedis whatBock (1996)hastermedthe“mindin the mouth” assumption.Shearguedthat researchersoften implicitly assumethat articulationprovidesa rel-atively direct reflectionof cognitive processing,but thelink from cognitionto behavior is mediated.For exam-ple, with regard to an activation thresholdof pronun-ciation readiness,different subjects,or even the samesubjectsacrosstrials, may set the thresholdat differ-ent levels as a function of trading speedfor accuracy(Colombo& Tabossi,1992;Lupker, Taylor, & Pexman,1997; Stanovich & Pachella,1976; Strayer& Kramer,1994;Treisman& Williams, 1991). The fact that nam-ing instructionsareusuallyambiguousasto emphasisonspeedversusaccuracy increasesthe likelihood of vari-

1

2 KELLO AND PLAUT

ability�

in thresholdplacement.Thresholdvariability is oftennot consideredanissue,

in partbecausethemorecentralprocessesdriving activa-tion of a pronunciationarethoughtto berelatively unaf-fectedby shiftsin responsegenerationthresholds.Morerecently, however, researchershave arguedthat gainingabetterunderstandingof responsegenerationin namingis importantfor interpretingnamingdataanddevelopingtheoriesof the underlyingcognitive processes(Balota,Boland,& Shields,1989;Jared,1997;Kawamotoet al.,1998; Kello & Kawamoto, 1998; Lupker, Brown, &Colombo,1997).

One theoreticalissuethat hasbeeninformed by re-searchfocusedon responsegenerationis that of strate-gic control over processingroutesin generatinga pro-nunciationfrom print (Lupker et al., 1997;Jared,1997;Monsell, Patterson,Graham,Hughes,& Milroy, 1992;Rastle& Coltheart,1999). Many researchershave pro-posedthat subjectscan strategically emphasizeor de-emphasizeoneof two availableprocessingroutesbasedon taskdemands(the routeemphasishypothesis;Herd-man, 1992; Herdman,LeFevre, & Greenham,1996;Monsell et al., 1992; Paap& Noel, 1991; Plaut et al.,1996). The route emphasishypothesisdoesnot distin-guishwhethera given processingroute is actuallyem-phasizedor de-emphasized.Rather, the hypothesisde-scribesany situationin which the processingof oneorboth routesis changedsuchthat one is privilegedovertheother. For example,if thestimuli in a word namingtaskconsistedof nothingbut irregularwords(e.g.,SURE,PINT, ROUSE, etc.),1 it would behoove subjectsto de-emphasizethe sub-lexical routebecausethis routemayprovide incorrectinformationon the irregular spelling-soundcorrespondences.

Monsell et al. (1992) testedthe route emphasishy-pothesisby dividing stimuli in a word namingtaskintopureand mixed blocks. The pureblockscontainedei-therall pseudowordsor all irregularwords(of mixedfre-quency in Experiment1, andseparatedby frequency inExperiment2). The mixedblockscontainedboth pseu-dowordsandirregularwords.Monsellandhiscolleaguesfoundthatirregularwordsweregenerallynamedfasterinpureversusmixedblocks,andthey interpretedthisasev-idencethatsubjectsde-emphasizedthesub-lexical routein pureblocksof irregularwords,presumablyto reduceinterferencefrom sub-lexical processing.However, onepuzzlingaspectof theirresultswasthatthepureblockla-tency advantagewasonly reliablefor blocksof high fre-quency (HF),but notlow frequency (LF), irregularwords(for similar results,seeAndrews, 1982; Frederiksen&Kroll, 1976).

Lupker et al. (1997)andJared(1997)revisited thesefindingsand arguedfor an alternative to the route em-phasisaccount.They first notedthat if onedefines“de-emphasis”asslowedprocessingtimes(of thesub-lexicalroutein thiscase,andregardlessof changesin variance),thenLF irregularwordsshouldhave anequalor greateradvantagein the pureblock comparedto HF irregulars.

This is becauseprocessingtimes to pseudowordsmustoverlapmorewith LF comparedto HF words,providedthat the meanof the sub-lexical routeprocessingtimesis greaterthanthat of the lexical route(assuggestedbyprevious findings; for example,wordsarenamedfasterthan nonwords; Forster& Chambers,1973). By con-trast,studieshave founda greaterpureblock advantagefor HF irregular words. Lupker et al. (1997) rerantheMonsell et al. (1992)blocking experiment(with minorvariations),andthey replicatedthepureblock advantagefor HF irregulars. Moreover, they found a statisticallyreliable pure block disadvantage for the LF irregulars.Lupkeretal. (1997)ranasecondexperimentto provideafurthertestof therouteemphasisaccount,in whichall ofthe stimuli containedregular spelling-soundcorrespon-dences.In this case,thesub-lexical routeshouldremainactive in both the pureandmixedblocks,andthereforeno blocking effect shouldbe found. Onceagain, theyfounda pureblock advantagefor HF words(now regu-lar), but unlike their first experiment,they founda pureblock advantagefor theLF wordsaswell. Jared(1997)foundanalogousresults,exceptthatshecomparedblocksmixed with pseudowordsversusblocksmixed with LFinconsistentwords. In summary, the resultsfrom Jared(1997) and Experiments1 and 2 from Lupker et al.(1997)werenotpredictedby therouteemphasishypoth-esis.

To explain their results, Lupker et al. (1997) re-categorizedthestimuli asfastor slow, basedonthemeanlatency for eachstimulustype in the pureblocks. Thepseudowordsand LF irregularswere slow, and the HFregularsandirregularswerefast(LF regularswerein themiddle). The patternof resultscould thenbe describedasfollows: whenever fastandslow stimuli weremixed,responselatenciesincreasedfor the faststimuli, but de-creasedfor theslow stimuli, relativeto whenthosestim-uli were in pure blocks. This insight lead Lupker andhiscolleaguesto proposethattheblockingmanipulationpromptedsubjectsto adjusta time criterion to initiatenamingresponses.Thegeneralideais thatsubjectscan,to somedegree,seta time deadlinerelative to stimulusonset(Ollman & Billington, 1972). If the pronuncia-tion is not fully activatedby thattime (i.e.,anactivationthresholdis alsoin place),thentheresponsemaybeini-tiatedbasedonwhateverrepresentationof pronunciationis availableat that time (alsoseeMeyer, Osman,Irwin,& Kounios,1988).

In orderto maintaina certainlevel of accuracy whilerespondingquickly, subjectsadjust the time criterionbasedon the difficulty of the stimuli presentedduringthe experiment. A pureblock of faststimuli allows fora quicker criterion than a pure block of slow stimuli.Whenfastandslow stimuli aremixed,subjectssetamid-dling time criterion: thus,fewer HF but moreLF words

1 Coltheartet al. (1993) definesan irregular word as onethat has one or more irregular grapheme-to-phonemecorre-spondences(GPC),asdeterminedby a setof correspondencerules(seealsoVenezky, 1970).

STRATEGIC CONTROL IN NAMING 3

are� hurried.This hypothesisembodiesaspeed–accuracytradeoff (Pachella& Pew, 1968; Wickelgren,1977),soit predictsan increasein errorsto slow stimuli in mixedblocks.2This is in factwhatLupker et al. (1997)found.

Motivationfor CurrentStudy

The studyby Lupker et al. (1997) raisestwo issuesin thecurrentcontext. First, in its simplestform, a timecriterionmeansthata responseis initiatedataparticularpoint in time, neitherbeforenor after that point (asidefrom randomfluctuations).Of course,thiscannotbethecasebecausesucha criterion would predict no effectsof stimulusprocessingon reactiontimes. One way toamendthetimecriterionhypothesisis to combineit withtwo activation criteria, a minimum and maximum: theminimum must be reachedto initiation pronunciation,a pronunciationis always initiated when the maximumis reached,and the time criterion operatesbetweenthetwo. Thismorecomplicatedversionof thetimecriterionhypothesiswould seem,in principle, to accountfor theblockingresultsreferredto above. However, it would benecessaryto specifyhow theactivationcriteriaaresetinorderto draw any clearpredictionsfrom thishypothesis,andto ourknowledge,thishasnot beenaddressed.

The secondissueraisedby the Lupker et al. (1997)studyis that the time criterion hypothesiscouldnot, onits own, accountfor all of their findings. In particu-lar, Lupker et al. (1997) estimatedLF irregular wordsand pseudowords to be of comparablespeed(i.e., dif-ficulty) basedon meanlatenciesto thesestimuli in thepure block conditions. In this case,the time criterionhypothesispredictsno blocking effect whencomparingpureblocksof eachtype with a mixed block of LF ir-regularsand pseudowords. However, the resultsfromExperiment1 showeda pureblock disadvantagefor LFirregularsanda pureblock advantagefor pseudowords.Lupker andhis colleaguesproposedan additionallexi-cal checkingstrategy thatsubjectsinvokedonly (but notalways)whenwordswerepresentin thestimulusblock.ThefactthatLupker andhiscolleaguesneededto invokeanadditionalmechanismraisesthe questionof whetheramoreparsimoniousalternativeto thetime criterionhy-pothesiscouldbeproposed(for similar issuesin decisionresponsetasks,seeRuthruff, 1996).

We believe that the time criterion hypothesisis wor-thy of investigationfor two main reasons:1) it providesa novel explanationfor stimulusblocking effects,but amoreexplicit mechanismneedsto be proposed,and2)it canpotentiallybe usedto addressthe time courseofphonologicalprocessingin word reading. In light ofthesereasons,we settwo goalsfor thecurrentstudy: 1)to formulatea moreexplicit mechanismof control overresponsetiming, and2) to formulateahypothesisof howpressurefor speedrelatesto the time courseof process-ing. Thefirst goalwassetin theserviceof investigatingthetimecriterionhypothesis,andthesecondgoalwassetto explore a specificpredictionmadeby currentmodelsof word reading.In thenext section,westepthroughthe

logic behindthis predictionandprovide somecomputa-tional supportfor it. We thenpresentthreeword nam-ing experimentsexaminingcontrolover theinitiation ofa speedednamingresponse.In theGeneralDiscussion,we proposeamechanismof controlover processingthatcouldaccountfor thecurrentfindings,aswell aseffectsof stimulusblocking like thosefound by Lupker et al.(1997).

Implicationof aTimeCriterionfor Modelsof Word

Reading

Thehypothesisof a timecriterionsuggeststhat,to theextent thatanexperimentercanmanipulatethesubject’stime criterion, onecould investigatethe time courseofprocessingin a fairly direct manner. If the shifting ofa time criterion is one type of speed/accuracy tradeoff(Pachella& Pew, 1968), then setting it earlier in timeshouldcausean increasein namingerrors(as it did inLupker et al., 1997). If subjectscouldshift thecriterionvery early in time, a very high error rateshouldensue.Speecherrorshave served as a primary sourceof evi-dencefor developingandtestingmodelsof speechpro-duction(Dell, 1986;Dell & Reich,1981;Levelt, 1989),andonecouldusethesameapproachtowardthestudyofwordreading.In thecurrentcontext, fasterrorresponsescould serve as a window into the early time courseofprocessing.

Currentmodelsof word readinghaveanexplicit timecourseof processingfrom stimulus onset to responsegeneration,but predictionsconcerningthe trajectoryofprocessinghave only been testedindirectly. For ex-ample, Kawamoto and Kitzis (1991) showed that theinteractive-activation model of word recognition(Mc-Clelland& Rumelhart,1981;Rumelhart& McClelland,1982), as well as a distributedmodel of lexical mem-ory (Kawamoto,1988),bothmake a specificpredictionconcerningthetime courseof phonologicalactivationinword reading.Whenprocessingan irregular word suchas PINT, the modelsshowed a strong influenceof theregular, incorrectpronunciation

��(to rhyme with

MINT) earlyin thetimecourseof processing(i.e.,aregu-larizationerror).As activationsettledto afixedstate,themodelsshowedthatthecorrectphonemeusuallyquashedactivationof theincorrectphoneme,but only laterin pro-cessing.Thisgeneralhypothesiswassupportedin anam-ing experimentthey conducted:the meanlatency of 16regularizedresponsesto thewordPINT was601mswhilethe meanlatency of 46 correct,irregularpronunciationswas711ms.

Thepredictionmadeby KawamotoandKitzis (1991)is also a more generalpropertyof most existing con-nectionistmodelsof word reading. In particular, if dis-

2 The complementarypredictionfor faststimuli (i.e., moreerrorsin thepureblock) couldnot beverified becauseperfor-mancewas at ceiling for thosestimuli in both the pure andmixedblocks.

4 KELLO AND PLAUT

tributed� representationsof orthography, phonologyandsemanticsall interact with eachother, then orthogra-phy can generatea phonologicalcodethroughtwo in-teractingbut distinct pathways: a direct orthography-to-phonologymapping (i.e., the non-semanticroute),and an indirect orthography-to-semantics-to-phonologypathway (i.e., thesemanticroute). We shall referto thisgeneralclassof modelasthetriangle framework of wordreading(Harm, 1998; Kawamoto, 1988; Plaut et al.,1996; Seidenberg & McClelland, 1989; Orden,1991).3

In this framework, thesub-lexical resonance(i.e., corre-lationalstructure)betweenorthographyandphonologyisstrongerthanthatbetweenphonologyandsemanticsbe-causethereis morestructurein theformermapping(Or-den,1991; Van Orden& Goldinger, 1994). Therefore,for an item that containsoneor moreexceptional(i.e.,rare)spelling-soundcorrespondences,the non-semanticroutemay generatethe morecommoncorrespondencesearly in processing.For theseitems,the semanticroutehelpsto overridetheinfluenceof sub-lexical knowledge,andthis tendsto occur later in processing.The modelof word readingpresentedby Zorzi, Houghton,andBut-terworth (1998) is alsolikely to make the samepredic-tion, given its cleardistinction betweenassembledandretrievedphonologies.

To providesomeevidencethatearlyregularizationer-rorsare,in fact,a characteristicof currentconnectionistmodelsof word reading,we examinedthe time courseof phonologicalrepresentationsin the attractormodelof word readingpresentedby Plautet al. (1996). As aroughapproximation,we applieda simpletime criterionto themodelby haltingits processingatsuccessivelyear-lier pointsin time,andcategorizingeachresponseinto 4possiblecategories: correctresponse,regularizationer-ror, word error, andmiscellaneouserror. The detailsofthis simulationarereportedbelow. The resultswereasexpected: the total numberof errors increasedas pro-cessingwashaltedat earlierpointsin processing,whichincludedanincreasein thenumberof regularizations.

Wecanalsoconsiderthetime courseof phonologyinthedual-routeframework. Intuitively, onemight expectthata dual-routeimplementation,suchasthedual-routecascade(DRC) model (Coltheartet al., 1993), wouldpredict no increasein the proportion of regularizationerrorsduring the early stagesof processing.This runscounterto thetriangleframework’sprediction.Thedual-routepredictionseemsto arisebecausethe lexical routeprocessingtimesarehypothesizedto be faster, on aver-age,thantherule routeprocessingtimes.Therefore,onewould expectthatword errors,but not regularizationer-rors,would increasein proportionduring theearliercy-clesof processing.On the otherhand,the rule route inthe DRC model processesthe input string from left toright over time,andtheirregulargraphemeof atestwordis usuallythesecondor third in positionfrom left to right(in amonosyllabicEnglishword,thevowel ismostlikelyto beirregular;Berndt,Reggia,& Mitchum,1987).If therule routehasenoughtime to outputat leastthefirst few

phonemesbeforethelexical routecansignificantlyinflu-encethe computationof phonology, thenregularizationerrorsmay occurasoften as,or moreoften than,worderrorsin theearlycyclesof processing.

Similar to ouranalysisof thePlautetal. (1996)attrac-tor model,weexaminedthetimecourseof processinginthe DRC model.4 The detailsarereportedbelow, but tosummarize,theresultsweresimilar to thoseof theattrac-tor model:thenumberof regularizationerrors,aswell asothererror types,increasedas responseswere taken atearlierpointsin processing.

SimulationMethodsandResults

Weransimulationswith boththePlautetal. (1996)at-tractormodel(Simulation3 in thatarticle)anda currentversionof theDRC model(Coltheart,personalcommu-nication). Both simulationswererun with the teststim-uli from Experiment2 of thecurrentstudy. To examinethetimecourseof phonology, baselinelatenciesfor eachmodel undernormal processingconditionswere deter-mined.Themodelswerethentestedagainwith thestim-uli from Experiment2, but processingwas haltedat anumberof differentpoints in time prior to the baselinelatency for eachmodel. The phonologicalrepresenta-tionsactive at thesehaltingpointswerecategorizedinto1 of 4 possiblecategories: correct,word error, regular-izationerror, andmiscellaneouserror. A word errorwasaphonologicaloutputthatcorrespondedto aword in themodel’s training corpus,but wasnot the target. A reg-ularizationerror was a phonologicaloutput that corre-spondedto the GPCrulesasdefinedby Coltheartet al.(1993),but wasnot thetarget(theseerrorswerepossibleonly for the irregularstimuli). Miscellaneouserrorsin-cludedall otherphonologicaloutputsthatdid notreachacriterionof correctness(definedfor eachsimulationbe-low). Miscellaneouserrorswerenot separatedinto non-wordandarticulatoryerrorsbecausein thesemodels,thisdistinctioncanonly bedrawn by anarbitrarythreshold.Therefore,an increaseor decreasein miscellaneouser-rorscanbeassumedto correspondto an increaseor de-creasein bothnonword andarticulatoryerrors. The at-tractormodelwasusedfrom Plautet al. (1996)becauseit is oneof the few publishedinstantiationsof the trian-gle framework thathasanexplicit time course(i.e.,usescontinuoustime units).

The Plaut et al. (1996) Attractor Model. The meanlatency undernormalprocessingconditionsfor thestim-

3 The term triangle framework is usedto refer to this classof connectionistword reading models becausethey imple-ment lexical processingin terms of interactionsamongdis-tributedrepresentationsof orthography, semantics,andphonol-ogy, whicharetypically drawn at thecornersof a triangle.Theterm is not intendedto apply to othermodels(e.g.,Coltheartetal., 1993)which,coincidentally, mayalsobedepictedin theshapeof a triangle.

4 We thankMax Coltheartfor makingavailable the outputof theDRC modelover processingcyclesfor ourstimuli.


T able1

Error countsfor theattractormodelandtheDRCmodel,categorizedbyerror typeandhalting time.

AttractorModel0.8 0.7 0.6 0.5 Total

Word 6 7 9 11 33(26.1) (22.6) (23.7) (21.2) (23.4)

Reg. 10 14 18 19 61(43.5) (45.2) (47.4) (36.5) (43.1)

Misc. 7 10 11 22 50(30.4) (32.3) (28.9) (42.3) (33.5)

Total 23 31 38 52 144

DRCModel65 60 55 50 Total

Word 7 11 9 11 15(9.3) (12.8) (11.3) (13.3) (11.7)

Reg. 10 13 15 17 55(13.3) (15.1) (15.5) (15.0) (14.7)

Misc. 58 62 71 81 272(77.3) (72.1) (73.2) (71.7) (73.6)

Total 75 86 97 113 371

uli from Experiment2 was1.85units of time. The er-ror breakdown for eachhaltingtime is shown in Table1.All threeerrortypes(word, regularization,andmiscella-neous)werefound to increaseasresponsesweregener-atedatearlierpointsin time.

The DRC Model (Coltheart et al., 1993). The meannumberof processingcycles undernormal processingconditionsfor the stimuli from Experiment2 was 98.Undernormalconditions,processingwascompletewhentheactivationof oneor morephonemesin eachposition-specificpool crosseda thresholdof 0.43. The model’sresponseswere determinedby taking the most activephonemeat eachposition (including null phonemes,ifthesewere the most active). The error breakdown foreachhalting time is shown in Table 1. All three er-ror types(word, regularization,andmiscellaneous)werefound to increaseasresponsesweregeneratedat earlierpointsin time.

SimulationDiscussion

Thetwosimulationsproducedcomparableresults:thenumberof regularizationerrors,as well as other errortypes,increasedasresponsesweretakenatearlierpointsin processing.Themaindifferencein errorpatternsbe-tweenthe two simulationswasthat the attractormodelproducedalargeproportionof regularizationerrorsover-all, but the DRC modelproduceda large proportionofmiscellaneouserrors. Most of the miscellaneouserrorsin theDRCmodelwereafailureto activatetherightmostphoneme(s)to criterion.Wedonotdraw any conclusionsbasedon thisdifferencein thesimulationsbecauseit de-pendson thesettingof criteria thatcouldbechangedin

futuresimulations.Thesesimulationsshow that currentmodelsof word

reading can make explicit predictionsabout the timecourseof processing.The experimentsreportedin thecurrentstudy were meantto, in part, explore the timecourseof processingin word reading.

Experiment1

Our initial researchquestionwastwo-fold. First,howpreciselycan subjectscontrol their timing of responseinitiation? A demonstrationof their ability to controlre-sponsetiming (or lack thereof)wouldbepotentiallyuse-ful in formulatinga morespecificmechanismof controlover responsetiming thangivenby Lupker et al. (1997)andJared(1997). Second,cansubjectsinitiate their re-sponsessubstantiallyfasterthanthey do in the standardnamingtask?If so,theerrorscouldhelpformulateanac-countof controlover responsetiming asit relatesto thetime courseof processingin theword readingsystem.

To addressthesequestions,we developed a novelmethodologycalledtemponaming. Prior to the presen-tationof a letterstring,subjectsarepresentedwith a se-riesof evenlyspacedauditorybeeps,accompaniedby theincrementalremoval of visual flankerson the computerscreen.Theletterstringis presenteduponthefinal beep,andthetaskis to pronouncetheletterstringsuchthattheresponseis initiatedsimultaneouslywith thesubsequentbeep(which is not actuallypresented).Therateof beeppresentation(i.e., tempo)canbe increasedor decreasedto requiresubjectsto respondmore quickly or slowly.Temponamingis similar to deadlinenaming(Colombo& Tabossi,1992; Stanovich & Bauer, 1978), in whichsubjectsare simply told to respondmore quickly if agiven responseis slower thana presetdeadline.A ver-sionof thedeadlineparadigmanalogousto temponam-ing would instructsubjectsto gono fasterthanthedead-line, aswell as no slower. However, temponamingisdistinct in two importantrespects.First, temponaminggivesan explicit and precisecue (the beepsand visualflankers)for whento initiate eachresponse.Second,thesubjectreceivesquantitativefeedbackoneverytrial indi-catingtheamount(in hundredthsof asecond)anddirec-tion (fastor slow) thattheresponsewasoff tempo.Sub-jectsareinstructedto adjustthetiming of their responsessuchthattheir feedbackis ascloseto zeroaspossibleonevery trial, evenat theexpenseof accuracy.

If subjectshave a mechanismakin to a time criterionat their disposal,thenthey shouldplaceit with a fixedrelationto tempo(to thebestof their ability). Studiesinfingertappinghaveshown thatbehavior canbeentrainedto a rhythmic cue (Kurganskii, 1994; Mates,Radil, &Poppel,1992), althoughthere is significant error andvariability within and acrosssubjects(Yamada,1995).One strategy that subjectscould adopt to perform thetemponamingtaskis to entrainan“internalmetronome”to thetempo,andthensynchronizethehypothesizedtimecriterion with the rate of the internal metronome.Theway in which subjectsusethe tempois a researchques-

6 KELLO AND PLAUT

tion in itself, andwe shalladdressthis questionto someextent. However, theprimarygoalof creatingthetemponamingtaskwas to examinethe mechanismof controlover responsetiming (independentof its relationto theperceptionandprocessingof tempo),aswell asthetimecourseof phonologicalprocessing.

With regardsto amechanismof controlover responsetiming, oneextremehypothesisis thatsubjectscanbaseresponseinitiation exclusively on somecue other thanthetargetstimulus(i.e., the tempoin thecurrentstudy).We shall refer to this as the cue-driven hypothesisofcontrol over responseinitiation. It might seemthat thedelayednaming task is a good test of this hypothesisbecauseit is cue driven (do not initiate a responsetillthecueis presented).Not surprisingly, researchershavefound that stimuluseffectsaregenerallyreducedin thedelayednamingtask (e.g., Balota & Chumbley, 1985;McRae,Jared,& Seidenberg, 1990;Savage,Bradley, &Forster, 1990), and in somecases,are eliminatedalto-gether(McRaeetal., 1990).Thepersistenceof stimuluseffectsin somedelayednamingexperimentsmight seemlike evidenceagainstthe cue-driven hypothesisof con-trol (i.e., responseswere presumablydriven by factorsother thanthe cue). However, delayednamingis not asufficient testbecauseit is not a purelycue-driventask:subjectshave the freedomto respondanytimeafter thecue.Thetemponamingtaskispurelycue-drivenbecausesubjectsareinstructedto initiate a responsein time withthe tempo,no soonernor later. If subjectscanobey thetempoabsolutely, thenstimulusfactorsshouldhave noeffecton responsetiming.

To testthis extremehypothesis,we chosestimuli thatvariedalongdimensionsknown to affect latenciesin thestandardnamingtask: printed frequency and spelling-soundconsistency (Jared,McRae,& Seidenberg, 1990;Seidenberg et al., 1984; Taraban& McClelland,1987;Waters& Seidenberg, 1985,; seetheStimulussectionofExperiment1 for detailsonouruseof consistency). If thecue-drivenhypothesisis correct,thenwe shouldfind noeffectof frequency or consistency on responselatencies,evenif thetempois setsuchthatsubjectsareinducedtorespondasfastor fasterthantheir averagelatency in astandardnamingtask.

Wemanipulatedspelling-soundconsistency for asec-ond purposeaswell: to examinethe influenceof sub-lexical spelling-soundcorrespondencesasa function ofresponsetiming. As explainedandsupportedabove, thetriangleand dual-routeframeworks both predict an in-creasein thenumberof regularizationerrorsasprocess-ing is haltedat earlierpointsin time. If the temponam-ing taskdoesindeedtapinto earlierpointsin processing,thenthesemodelspredictan increasein the numberofregularizationerrorsto exceptionwords. To investigatethis issue,temposwere set to be as fast or fasterthaneachsubject’s baselinenaming latency, as determinedby an initial block of standardnamingtrials. The onlyguidewe hadto determinehow muchfasterthanbase-line subjectsshouldbe inducedto respondwasa study

by ColomboandTabossi(1992).Usingaresponsedead-line, they inducedsubjectsto respondmorethan60 msfasterthan baselinewithout any significantincreaseinerror rate. We wantedto induceerrors,so we set themaximumtempoto induceresponsesconsiderablyfasterthan60 msbelow baseline(150msmaximum). We ex-ploredarangeof temposfasterthanbaselinebecausewedid notknow how well subjectscouldperformthetask.

We neededto consideroneauxiliary issuein creatingthe temponamingtask. How do the acousticpropertiesof an initial phonemeaffect subjects’ability to time agiven responsewith the tempo? It hasbeenknown forsometime that such propertieswill affect naming la-tenciesas traditionally measured(Sherak,1982; Stern-berg, Wright, Knoll, & Monsell,1980).Kawamotoetal.(1998)showed that evenwhenproblemswith the voicekey arealleviated,acousticenergy from responseswithplosive stopsasthe initial phoneme(e.g.,

�� )

will begin muchlater (i.e., about60–100ms) thancom-parable responseswith non-plosive initial phonemes.Timing in thetemponamingexperimentswasmeasuredacousticallyonline andgiven asfeedback,andsubjectswere instructedto respondwith the bestpossibletim-ing asmeasuredby their feedback.If subjectstime re-sponseinitiation basedonly on articulatorycommands,thentheacoustictiming (i.e.,feedback)wouldbeconsis-tentlyslow for plosivecomparedwith non-plosiveinitialphonemes.Alternately, subjectsmight be able to timetheir responsesbasedon the onsetof acousticenergy.If this doesnot dependon the type of acousticenergy(i.e., periodic versusnon-periodic,as in voiced versusunvoiced phonemes),then the type of initial phonemewould have no effect on timing. As a third alternative,onemight find differencesbasedon the type of acous-tic energy (elaboratedin the Resultssectionof Experi-ment1). This issuewasnot centralto our line of investi-gation,sowe simply includeda mix of initial phonemesin thetestwordsandwithin blocks(butcontrolledfor ini-tial phonemeacrosslevelsof theindependentvariables).Wementionit herebecauseit will be importantin inter-pretingcertainaspectsof the findingsin the currentsetof experiments.

Methods

Subjects. A total of 33 subjectsparticipatedin theex-perimentaspart of a requirementfor an undergraduatepsychologycourse. Subjectsreportedbeingnative En-glishspeakerswith normalor correctedvision.

Stimuli. The test stimuli in the tempo naming taskwere composedof 52 high-frequency exception(HFE)words,52 low-frequency exception(LFE) words,and52low-frequency consistentwords(LFC). An additional13of eachstimulustype werealsochosenfor thestandardnamingportion. For eachstimulustype, 13 of the 52wordschosenfor temponamingwere also includedinstandardnamingfor a total of 26 wordsof eachtype instandardnaming.All testwordsweremonosyllabic.The


T able2

Meansof control variablesfor the standard and temponamingportionsof Experiment1, categorizedby stimu-lus type

StandardNamingHFE LFE LFC

WordFreq 999.0 4.2 2.8BigramFreq 9492 5589 4561# Letters 4.2 4.5 4.2# Phonemes 3.4 3.7 3.7

TempoNamingHFE LFE LFC

WordFreq 617.3 8.2 6.2BigramFreq 8072 5107 5023# Letters 4.3 4.6 4.3# Phonemes 3.3 3.6 3.6

factorscontrolledfor were initial phoneme,numberoflettersand phonemes(HFE words had higher summedpositionalbigramfrequenciesthanLFE andLFC words).Words were chosenin triplets (one of eachtype), andeachtriplet wasmatchedonthecontrolfactorsascloselyaspossible(LFE–LFC pairswerealsomatchedon fre-quency, andHFE–LFEpairswerealsomatchedonbodyconsistency). The meanvaluesfor the independentandcontrolfactorsasafunctionof stimulustypearegiveninTable2, andthetripletsaregivenin AppendixA. Printedfrequency was estimatedwith the Kuceraand Francis(1967)norms.

Spelling-soundconsistency is a generalconceptthatcapturesastatisticalrelationshipbetweensub-lexical or-thographicunitsandtheir correspondingpronunciations(Plautetal.,1996):thedistributionof differentpronunci-ationsfor a givenorthographicunit (which canbemea-suredby token or type; Jaredet al., 1990). The con-sistency of a particularpronunciationfor a givenortho-graphicunit increasesasthe numberof alternative pro-nunciationsdecreases.To quantify the consistency ofmonosyllabicwords,researchersusuallyconsidera sin-gle orthographicunit (i.e., vowel plus any final conso-nants,or body), eventhoughtheconceptappliesto otherunitsaswell.

To createaslarge a pool of exceptionwordsaspos-sible, we usedthe conceptof consistency in its generalform, ratherthanits useasa label for bodyconsistency.We categorizedstimuli as consistentor exceptionalasa function of the numberof alternative pronunciationsfor all orthographicunits greaterthan or equal to thegrapheme,andlessthantheword, in size.

In particular, the pronunciationof a contiguousor-thographic unit was exceptional if it comprisedlessthan50%of all theposition-specificpronunciationtypes(basedonthepositionsonset,vowel, andcoda),summedacrossall monosyllabicEnglishwords.For example,theI, the IN, andthe INT in PINT areall exceptional(e.g.,comparewith TICK, BIN, andHINT). As anotherexam-

ple, the I and IN, but not IND, in KIND areexceptional(comparewith BIND, FIND, MIND, etc.).

A word wasdefinedasexceptionalif it containedoneor moreexceptionalorthographicunits. A word wasde-fined as consistentif all orthographicunits mappedtotheir mostcommonpronunciation.For example,HOOKis exceptionalbecausethegraphemeOO usuallymapstothe long vowel

��. By the samelogic, SPOOK is also

exceptionalbecausetheorthographicbody OOK usuallymapsto theshortvowel

��.

Consistency is differentfrom GPCregularity in twoimportantrespects.First, the conceptof irregularity isbasedongraphemes,whereastheconceptof consistencyappliesto multiple levelsof orthographicstructure.Sec-ond, irregularity is basedon discrete,all-or-nonecrite-ria (i.e., rules),whereasconsistency is basedon a con-tinuousmeasureof thestatisticaldistributionof pronun-ciations. Despitethesedifferences,85% of our excep-tion words werealso irregular by GPC rules. In addi-tion, even thoughour definition of consistency is moreinclusivethanthebodydefinitionof consistency, 85%ofour exceptionwordswerebodyexceptional.Irregularityandconsistency have beenthefocusof researchin otherstudies(e.g.,Glushko, 1979;Jaredetal., 1990),but theirdifferencesare not important for the issueof strategiccontrolover responseinitiation.

The standardnamingblocks included2 filler wordsat the beginning of eachblock, and the tempo nam-ing blocks included156 fillers mixed throughoutthe 4blocks;10 at the beginning of eachblock, and29 inter-spersedthroughouteachblock. Filler wordsweremono-andbisyllabic,andrangedin frequency andconsistency.Standardnamingconsistedof onepracticeblockandtwotest blocks, and temponamingconsistedof 1 practiceblock and4 testblocks. Teststimuli wereevenly mixedandbalancedacrossblocks,andcounter-balancedacrosssubjects.Theorderof trials within blockswasrandom-ized for eachsubject,underthe constraintthat 2 fillersbeganeachstandardnamingblock, and10 fillers beganeachtemponamingblock. Standardnamingalwayspre-cededtemponaming,andthepracticeblocksbeganeachportionof the experiment.The orderof testblockswascounterbalancedacrosssubjects(in aLatin squaredesignfor the tempoblocks). An equalportion of eachstimu-lus typeappearedin eachtestblock within standardandtemponaming,andfillers weredivided equallyamongtestblocksaswell. Thestandardnamingandtemponam-ing practiceblocksconsistedof 10and40fillers, respec-tively.

Procedure. Subjectssatin front of a 17-inchmonitor,approximately2 feetaway, andworeanAudio Technicaheadsetcardioidmicrophone.The microphonewaspo-sitionedapproximately1 inch away and2 inchesdownfrom the subject’s mouth, and it was plugged into aSoundblaster16-bit soundcard. Subjectswere givenwritten instructionsfor the standardnaming task andasked to read them silently. Following this, the ex-

8 KELLO AND PLAUT

perimenter� summarizedthe instructions,andany ques-tionswereanswered.Subjectswereinstructedthat theywould seewords presentedin isolation on the monitor(the width of eachletter subtendedapproximately1.2degreesof visual angle),andthat their taskwasto pro-nounceeachword out loud asquickly andaccuratelyaspossible.Thelevel of recordingwascalibratedfor eachsubject. Following calibration,the subjectran throughthepracticetrials with theexperimenterpresentto makesurethat the subjectperformedthe taskcorrectly. Thepracticeblock was followed by 2 continuousblocksoftesttrials.

Each standardnaming trial began with a READY?prompt centeredon the monitor. The subjectpressedthe spacebar to begin the trial, uponwhich the promptwas replacedwith a fixation point. The fixation pointremainedfor 500ms,andwasreplacedby asingleword.The word remainedon the screentill a vocal responsewas detected. All stimuli (for temponamingas well)were presentedin lowercase,in a large, distinct font(similar in appearanceto timesnew Roman)to minimizeletterconfusions.Thetimefrom wordonsetto thebegin-ningof thenext trial wasafixed1500ms(i.e.,thescreenwasblankfor any remainingtime aftertheresponsewasdetected). This was necessarybecauseeachvocal re-sponsewas digitized and storedon the hard drive us-ing theRunword softwarepackage(Kello & Kawamoto,1998).

Immediatelyafter eachsubjectcompletedthe stan-dardnamingportionof theexperiment,themeanlatencyof all testtrials wascalculated(excludingany responsesfasterthan200 ms, but including any errorsthe subjectmay have made). Naminglatency wascalculatedusingan acousticanalysisalgorithm describedin Kello andKawamoto(1998). In brief, the algorithm is sensitiveto increasesin amplitude(e.g.,to detectvoicing)aswellasfrequency of acousticenergy (e.g.,to detectfrication).Eachsubject’smeanlatency in thestandardnamingtaskwasset as the baselinetempofor the upcomingtemponamingblocks. The four test blocks for eachsubjectwereassignedfour different tempos: baseline,and50,100,and150msfasterthanbaseline(B-0, B-50,B-100,andB-150, respectively). As notedabove, the orderoftestblockswascounterbalancedin aLatin squaredesign.Stimuli wererotatedacrosssubjectssuchthateachsub-jectsaw eachtestwordoncein thetemponamingblocks(1/4 of the test words appearedin standardnamingaswell), andeachtestword appearedin every tempoandeveryblock orderacrosssubjects.

After completingthestandardnamingblocks,thesub-jectwasgivenwritten instructionsfor thetemponamingtask. After readingthemsilently, the instructionsweresummarizedandany questionswereanswered.A para-phrasingof the instructionsis as follows (alsoseeFig-ure1):

Eachtrial will begin with a promptfol-lowedby thepresentationof 5pairsof visualflankers. Then,5 beepswill beplayedsuc-

Ready? Space Bar

Blank Screen 500 ms

>>>>> <<<<< Tempo, 20 ms

>>>> <<<< Tempo, 20 ms

>>> <<< Tempo, 20 ms

>> << Tempo, 20 ms

> pint < Tempo, 20 ms

beep

pint 1300 ms −− Tempo

Your response was # Space Bar

Visual Auditory Respon se DurationStimulus

Subject

beep

beep

beep

beep

“ pint”

Subject

Figure 1. Diagramof the courseof eventsfor a single trialin thetemponamingtask.The“ �! ” symbolsareflankersin-dicatingthepositionof the targetstimulus.Tempois the timeinterval betweeneachbeep,determinedby thetempoconditionandthesubject’sbaseline.“Subject” indicatesthatthedurationis subject-dependent.

cessively in asteadyrhythm,andthepairsofflankerswill disappearoneby onewith eachbeep.Uponpresentationof thefifth beep,aword will appearin betweenthelastpair offlankers.Try to nametheword suchthatthebeginningof yourresponseis timedwith thesixth beep.However, no sixth beepwill beplayed; your responseshouldbegin wherethesixthbeepwouldhavebeen.Youwill getfeedbackafter completingyour responsetotell youhow well-timedit wasto thetempo.Thefeedbackis in theform of anumber, themorepositiveit is, thesloweryourresponse,the more negative it is, the faster. A per-fectly timed responseproducesa feedbackof zero. Your primary task is to namethewordontempo,regardlessof makingerrors.In thepracticeblock will seea mix of bothrelatively fastandslow tempos,but thenyouwill run throughfour testblocks,andeachone will be set at a different, but uniform,tempo.

After instruction, subjectsran through the practiceblock, which representeda randomlyordered,but bal-anceddistributionof thefour tempos.Theexperimenterstayedwith the subjectthrougha numberof trials to besurethetaskwasunderstood,andto give any additionalinstructionif necessary. Eachtemponamingtrial pro-ceededas follows: A READY? prompt was presentedin the centerof the screen,and the subjectpressedthespacebar to begin. Therewas a 500 ms delaywith ablank screenafter pressingthe spacebar, followed bythepresentationof a pairedsetof flankerssimultaneouswith a brief tone20 ms in duration. The flankersweresequentiallyerasedfrom the outsideinward, in time in-tervalsequalto thetempofor thattrial. Eachtime a pair


of" flankerswaserased,thetonewassimultaneouslypre-sented. Interval durationswere roundedup so that theremoval of eachflanker pair, aswell asthepresentationof eachbeep,could be synchronizedwith the video re-freshrate(14 ms roundup at most). Uponpresentationof thefifth beep(andremoval of thefourth flanker pair),thetargetword waspresented,centeredbetweenthelastflanker pair. Recordingfrom the microphonewasiniti-ated200msbeforethenext (sixth)interval,andlastedfor1500ms. On thesixth interval, the lastflanker pair wasremoved,andthewordremainedonthescreenfor thedu-rationof recording.Thewordwasthenreplacedwith themessageYOUR RESPONSE WAS #, where# equaledtheamountof time in hundredthsof a secondthat responselatency differedfrom the sixth interval. This wascom-putedby subtracting200msfrom thecalculatedonsetofacousticenergy relative to the onsetof recording. Thenumberwas a positive or negative integer correspond-ing to the responseoffset from tempo. The feedbackremaineduntil the subjectpressedthe spacebar, whichbroughtup the READY? promptfor thenext trial. Sub-jectswereexplicitly askedto take a shortbreakafterthefirst two testblocks,andthey weredebriefedaftercom-pletingall 4 testblocks.

Results– Standard Naming

Throughoutthe experimentalresultssectionsin thisstudy, statisticsover thestandardnamingmeansandfre-quency counts(for errortypeanalyses)arereportedfirst,alongwith the magnitudesof any relevant effects. Thetemponamingresultsarereportedafterwards,alongwithgraphsincludingdatafrom boththestandardandtemponamingresults.Thedataarepresentedin this format tofacilitate direct comparisonof the standardand temponamingmeans.Exceptfor errortypeanalyses,all statis-tics wereanalysesof variance(ANOVAs), andall anal-yseswere conductedwith subjectsas the randomfac-tor, aswell asitems(denotedasFs andFi , respectively,whenpresentingF values).Finally, all reportedmeansinthecurrentstudyweresubjectmeansunlessstatedother-wise.

Data Removal. Datafrom 1 subjectwereremoveddueto equipmentfailure, anddatafrom 1 item (chic) wereremoved due to an excessof errors(78%). Responseswere removed if naminglatency was lessthan200 msor greaterthan1200ms. Responseswerecodedfor er-rors(blind to theblock they appearedin), removedfromall otheranalyses,andanalyzedseparately. In any caseswheremultiple responseswere given on a single trial,only thefirst onewasconsideredfor errorcategorization(but all suchresponseswereconsiderederrorsof sometype). Stuttersthatwerefollowedby a fluentbut incor-rectresponsewerecategorizedby theincorrectresponse,ratherthan as articulatoryerrors. The error categorieswereasfollows (exampleswere taken from the corpusof errorsgeneratedin Experiments1 and2):

# Word errors were responsesthat formed a word,but did not matchthe target pronunciation.In all casesthroughoutthis study, word errorswerephonologicallyand/ororthographicallysimilar to their targets(i.e., theydiffered in no more than 2 phonemesfrom the tar-get). For example: PINT $ PINE, HITCH $ PITCH, andGLARE $ GLAD.# Legitimate Alternative Reading of Components(LARC) errors were responsesto exceptionwords thatfollowedanalternatepronunciationof their exceptionalorthographicunit, anddid not form a word. Strain,Pat-terson,Graham,andHodges(1998)usedthetermtohaveessentiallythe samemeaning. For example: PINT torhymewith MINT, MOW to rhymewith NOW, andNOWto rhymewith MOW.# RegularizationandNon-regularizationerrors weretwo different types of LARC errors. Regulariza-tions were those that followed GPC rules, and non-regularizationsweretheremainder. Dividing theLARCcategoryin thiswaymaybeimportantfor relatingtemponamingresultsto theDRC modelof word reading.# Mixederrors wereLARCs thatalsoformedawordother than the target. For example, GREAT $ GREET,GHOUL $ GOAL, andPLAID $ PLAYED.# Nonword errorswerefluentpronunciationsthatdidnot form a word or a regularization. For example,GLOVE $ GUV, SHOE $ SHOPE, andTUNT $ TURT.# Articulatory errors includedall non-fluentpronun-ciations,in particular, stuttersandgarbledor incompre-hensibleresponses.

Naming Latency Analyses. Themaineffectof stimu-lus typewassignificantby subjectsanditems, Fs(2,62)% 34, p & .001, Fi (2,74) % 5.4, p & .01. Therewasa non-significant7 ms decreasein meanlatency fromblock 1 to 2, Fs(1,31) % 1.1, p ' .2, Fi (1,74) % 2.88,p& .1, andtherewasno reliableinteractionof block andstimulustype, Fs(2,62) & 1, Fs(1,49) & 1. Pairwisecomparisonsrevealedthat a 26 ms advantageof HFEover LFE words(hereafterreferredto asa frequency ef-fect) wassignificantby subjectsanditems, Fs(1,31) %47, p & .001, Fi (1,49) % 11.8, p & .001. However, a 7msadvantageof LFC overLFE words(hereafterreferredto asa consistency effect) wasonly reliableby subjects,Fs(1,31) % 5.1, p & .05, Fi (1,49) % 1.3, p ' .2.

Error Analyses. Themaineffectof stimulustypewassignificant, Fs(2,62) % 44, p & .001, Fi (2,74) % 6.5, p& .01,but therewasnomaineffectof blocking, Fs(2,62)% 2.7, p ' .1, Fi (1,74) % 2.3, p ' .1. The interactionof block with stimuluswassignificant, Fs(2,62) % 6.0,p & .01, Fi (2,74) % 5.4, p & .01. Post-hocanalysesshowed that whencollapsedacrossfrequency, the errorrateto exceptionwordsreliably decreasedfrom thefirstto secondblock, Fs(1,31) % 8.7,p & .01, Fi (1,49) % 7.6,p & .01,but marginally increasedfor regularconsistentwords Fs(1,31) % 3.5, p & .05, Fi (1,25) % 2.8, p &.1. Plannedcomparisonsshowed that LFE wordswere

10 KELLO AND PLAUT

reliably( 9.1%moreerrorpronethanLFC words(consis-tency effect), Fs(1,31) % 48, p & .001, Fi (1,49) % 6.7,p & .05, and 9.9% more error pronethan HFE words(frequency effect), Fs(1,31) % 59, p & .001, Fi (1,49) %7.5, p & .01.

To provide moredetail concerningerrors,frequencycountswere analyzedas a function of block and errortype.Sincethedependentmeasureis a frequency count,chi-squareanalyseswereperformedonthe2 ) 5 contin-gency tableformedby block anderror type.5 Collapsedacrossstimulustype,errorcountswerenot reliably dif-ferentthantheir expectedvaluesbasedon row andcol-umn meanscalculatedacrosslevels of block and errortype χ2(4) % 6.17,p ' .15.

Results– TempoNaming

Data Removal. The subject removed from standardnaminganalyseswasalsoremoved from temponaminganalyses.Errorswerecodedin thesameway asin stan-dardnaming(i.e., blind to block and thereforetempo),and were removed from all other analysesand treatedseparately(seebelow). Then, responsesthat were lessthan 175 ms or greaterthan 1000 ms from the sixthtempointerval were removed (recordingbegan200 msbeforethesixth tempo).

Latency Analyses. Figure2 graphsmeannamingla-tencies(i.e., time from stimulusonset)asa function ofstimulustype and tempo(including the meanlatenciesfrom the standardnamingtask; note that the statisticspresentedheredo not includestandardnamingdata,seeabove for those). The main effect of stimulustype wassignificantonly by subjects, Fs(2,62) % 6.0, p & .05,Fi (2,153)% 1.6,p ' .2,whereasthemaineffectof tempowasreliablein bothanalyses,Fs(3,93) % 214,p & .001,Fi (3,459) % 245,p & .001.Theinteractiondid not reachsignificance, Fs(6,186) % 1.6, p ' .1, Fi (6,459) & 1.Plannedcomparisonsshowedthatthe5 msfrequency ef-fectwasreliableonly bysubjectsFs(1,31) % 7.2,p & .05but not items Fi (1,102) % 2.3, p ' .1, whereasthe 0.6msdifferencein latenciesto LFE versusLFC wordswasnotsignificant Fs(1,31) & 1 and Fi (1,102) & 1. Plannedcomparisonsof the tempomanipulationconfirmedthateachsuccessively fasterlevel of tempocausedresponsesto bereliably fasterthanthepreviouslevel: from B-0 toB-50, Fs(1,31) % 130, p & .001, Fi (1,153) % 104, p& .001, from B-50 to B-100 Fs(1,31) % 73, p & .001,Fi (1,153) % 46, p & .001, and from B-100 to B-150,Fs(1,31) % 85, p & .001, Fi (1,153) % 54, p & .001.

Theseinitial testssuggestthattheinfluenceof stimu-lus typeon latenciesis smallerin thetemponamingtaskcomparedto standardnaming(a 26 msfrequency effectand7 ms consistency effect in standardnaming,versus5 ms and0.6 ms in temponaming,respectively). Onepossiblereasonfor this attenuationof stimuluseffectsisthattheoverallvariability in latenciesdecreasedin temponaming. This might be expected,given that we askedsubjectsto respondatparticulartimeintervals.However,

350

400

450

500

550

Hi Freq Exc Lo Freq Exc Lo Freq Con

Nam

ing

Late

ncy

(ms)

Std 0 -50 -100 -150

Figure 2. Meanlatenciesfrom thestandardandtemponam-ing portionsof Experiment1 asafunctionof stimulustypeandtempo. Thedashedlines separatethe standardnamingmeansfrom thetempomeans.

inspectionof the standarderror barsin Figure2 showsthat the within-cell variability was comparableacrosstasks(6.2 ms for standardnaming, 7.9 ms for temponaming,within-cell standarderrorsaroundthe subjectmean).An analysisof variancewith taskastheindepen-dentvariableandstandarderrorasthedependentvariableshowed this differenceto be non-significant, Fs(1,31)& 1. Therefore,in termsof analysesof variance,thebetween-conditionvariancedecreasedin temponaming,but not thewithin-conditionvariance.

However, thereare threeconcernswith drawing theconclusionthat between-conditionvariability decreasedin temponaming: 1) standardnamingalwaysprecededtemponaming,2) 25%of thetemponamingstimuli alsoappearedin thestandardnamingblocks,and3) only halfof thestandardnamingstimuli appearedin temponam-ing (the otherhalf wasnot explicitly controlledagainstthe temponamingstimuli). Thesethreeconcernswereaddressedasfollows: 1) analysesof varianceon temponaminglatencieswereconductedwith therepeatedstim-uli removed, and 2) the interactionof block order andstimulustype wasexaminedto testfor a practiceeffectwithin thetemponamingtask,and3) thestandardnam-ing latencieswere re-analyzedwith only thosestimulithatappearedin temponaming.

The tempo analyseswith repeatedstimuli removedwereessentiallyidenticalto theanalysesreportedabove.Most relevantly, the main effect of stimulus type wasagain reliable only by subjects Fs(2,62) % 4.3, p &.05, Fi (2,114) & 1. The pairwisecomparisonsshoweda 6 ms frequency effect that did not reachsignificance,Fs(1,31) % 2.4, p ' .1, Fi (1,76) % 1.1, p ' .2, anda3 ms non-significantdisadvantagefor LFC wordscom-paredto LFE words Fs(1,31) % 1.6, p ' .2, Fi (1,76)

5 In all of our chi-squareanalyses,theobservationsarenotindependent,andmaythereforeby positively biased.


350

400

450

500

550


Nam

ing

Late

ncy

(ms)

Block 1 Block 2 Block 3 Block 4

Figure 3. Meanlatenciesfrom the temponamingportion ofExperiment1 asa functionof stimulustypeandblock.

& 1. If practicehadreducedtheeffect of stimulustypeon latency, thenonewould expectthis effect to increasewhenrepeatedstimuli areremoved; if anything, the ef-fect decreasedslightly (albeit at leastin part dueto re-ducedpower). Theanalysesof block orderandstimulustype revealedno discerniblemain effect of block order,Fs(3,93) & 1, Fi (3,459) & 1, nor interactionof blockorderwith stimulustype, Fs(6,186) & 1, Fi (6,459) &1. If therewasa practiceeffect from standardto temponaming,onemightexpectthiseffect to continuethroughtheblocksof temponaming.

To illustratethe lack of a practiceeffect on latencies,Figure3 graphsnaminglatency asafunctionof blockor-derandstimulustype.Finally, there-analysisof standardlatenciesincluding only temponamingstimuli showedthe samepatternof effectsas the original analysis,butwith lesspower andthereforefewer significantcompar-isons. Relevant to the comparisonswith temponamingresults,themain effect of stimulustype wasreliablebysubjects,Fs(2,62) % 19, p & .001,but marginally signif-icantby items Fi (2,36) % 2.7,p & .08. Plannedcompar-isonsshowed that the 23 ms frequency effect (cf. a 26ms effect with all stimuli) wasreliable, Fs(1,31) % 31,p & .001, Fi (1,24) % 4.8, p & .05,but the5 msconsis-tency effect (cf. a 7 ms effect with all stimuli) wasnot,Fs(1,31) & 1, Fi (1,24) & 1. This final analysissuggeststhat the largereffect of stimulustype in standardversustemponamingwasnot dueto item differences.

Timing Analyses. Naming latencies can also begraphedasoffsetsfrom perfecttempo. In otherwords,subjectswere instructedto begin their responseexactlyon the sixth tempointerval, so onecangraphtheir tim-ing accuracy. Analysesof varianceon naminglatencyindicatedthat increasesin tempocausedlarge, reliabledecreasesin naminglatency. Analysesof varianceontiming will indicatewhetherresponseonsetswerereli-ably different from tempo. Figure4 graphsmeantim-

-30

-15

0

15

30

45

60


Tim

ing

(ms)

0 -50 -100 -150

Figure4. Meantiming accuracy from Experiment1 asafunc-tion of stimulustypeandtempo.

ing offsetsasa functionof stimulustypeandtempo.Asin the latency analyses,themaineffect of stimulustypewasonly reliableby subjects6 Fs(2,62) % 6.0, p & .01,Fi (2,153) % 1.3, p ' .2, themaineffectof tempowasre-liable Fs(3,93) % 119, p & .001, Fi (3,459) % 240, p &.001,andthe interactionwasnot significant, Fs(6,186)% 1.6, p ' .1, Fi (6,459) & 1. Plannedcomparisonson stimulustype werenot conductedbecausetheseareequivalentto theanalogouscomparisonswith namingla-tency asthedependentmeasure.Plannedcomparisonstotestwhethertiming wasprogressively delayedastempoincreasedrevealedthatresponseswerein factfurtherde-layedfromtempoateachincrease:B-0 to B-50, Fs(1,31)% 27, p & .001, Fi (1,153) % 33, p & .001,B-50 to B-100, Fs(1,31) % 70, p & .001, Fi (1,153) % 70, p & .001,B-100 to B-150, Fs(1,31) % 62, p & .001, Fi (1,153)% 126, p & .001. To characterizethe failure of tempoto perfectlydrive responseinitiation, a linear regressionline wasfit to the timing meansat eachlevel of tempo,averagedacrossstimulustype; the slopewas0.43 (i.e.,1 * Sr t, whereSr t is the slopeof the regressionline forlatency). If thetempomanipulationwasperfectin deter-mining responseinitiation, theslopewouldbe0.

One unexpectedfinding in the timing analyseswasthatresponseswere,onaverage,fasterthantempoin theB-0 condition. If tempoconditionsweremixed withinblocks,thisfindingcouldhavearisenfromhysteresisof aresponsecriterion(i.e.,a relatively slow tempotrial pre-cededby a fastonemight have a tendency to be overlyfast; Lupker et al., 1997). However, sincetempowas

6 In principle, significancelevels for latency and timinganalysesof stimulus effects (not tempo or blocking effects)shouldbe thesame.However, they differ slightly becausela-tency is relative to the onsetof recording,whereastiming isrelative to an estimateof the tempointerval. Theseestimatesarenot alwaysaligneddueto small variationsin the onsetofrecording,andsmall variationsin the timing of the tempoin-terval dueto alignmentwith themonitorrefreshrate.

12 KELLO AND PLAUT

-40

-20

0

20

40

60

80

Plos-Unvoc Nonplos-Unvoc Plos-Voc Nonplos-Voc

Tim

ing

(ms)

0 -50 -100 -150

Figure 5. Mean timing with tempofrom Experiment1 asa function of tempoand the articulatorycharacterof the ini-tial phoneme(plosiveversusnon-plosiveandvoicedversusun-voiced).

blockedandcounterbalanced,andeachblockbeganwith10practicetrials, this explanationcannotbecorrect.Wereasonedthatanadequateexplanationwould dependontheacousticcharacteristicsof responseonset,aswell asan answerto the questionof what articulatory/acousticmarker subjectstry to time with the tempo. We inves-tigatedthis issueby examining timing asa function oftempoandinitial phonemetype(graphedin Figure5).

We categorizedinitial phonemetypebasedon acous-tic characteristicsthat areknown to affect the measure-mentof responselatencies(Kello & Kawamoto,1998):voicing (voiced or unvoiced) and plosivity (plosive ornon-plosive). Examplewords with an initial phonemein eachof thefour categoriesare:voicedplosives(BED,DEAL , GATE), voicednon-plosives(VET, ZOO, RED, un-voiced plosives (PET, TEA , KITE), and unvoiced non-plosives(FAT, SEA, THIN). Figure5 showsthatunvoicedinitial phonemes,especiallynon-plosive ones,werefastin the B-0 condition, whereasvoiced initial phonemeswerecloselytimed to tempoin the B-0 condition. Thisresult suggeststhat subjects,despitethe numericfeed-back, timed their responseswith the onsetof voicing(which we argue and statistically supportbelow). Toanswerthe questionof why responseswere fasterthantempoin theB-0 condition,recall that the latency algo-rithm usedin this study is sensitive to high amplitudeandhigh frequency acousticenergy. The onsetof peri-odicenergy is laterin responsesbeginningwith unvoicedcomparedto voicedinitial phonemes,relative to theon-setof any typeof acousticenergy. To accuratelytimetheonsetof voicing, measuredlatenciesfor unvoicedinitialphonemesmustbefast.

To argue for the hypothesisthat subjectstime re-sponseswith the onsetof voicing, we shall distinguishit from the alternatehypothesesthat timing was basedon theonsetof any acousticenergy (whichwasthebasis

of feedback),the onsetof articulation,or the onsetofthe vowel. If timing wasbasedon the onsetof acous-tic energy, thereshouldbe no effect of initial phonemetype. However, the timing of plosive initial words(i.e.,��+� �� ,� �� -��

;7Nitem% 63) was14 msslower thannon-

plosive initial words(Nitem% 93), Fs(1,31) % 33, p &

.001, Fi (1,154) % 16, p & .001, andthe interactionofplosivity andtempowasmarginally significantby itemsFs(3,93) % 1.5,p ' .2, Fi (3,462) % 2.5,p & .06.Theef-fect of plosivity diminishedslightly astempoincreased(18 ms effect at B-0, 17 ms at B-50, 10 ms at B-100,and11 ms at B-150). The main effect of plosivity (aswell asothereffectsof initial phonemereportedbelow)rules out the acousticenergy hypothesis,and we shallreturn to the interactioneffect in the sectionon dura-tion analyses.Next, if timing wasbasedon theonsetofarticulation,thenresponsesbeginning with non-plosivephonemesshouldbe relatively well-timed(to theextentthat subjectscan keep up with the tempo), and thosewith plosive initial phonemesshouldbe slow by com-parison. This is becausethe onsetof articulationmorecloselycorrespondsto the onsetof acousticenergy fornon-plosivecomparedto plosive-initial responses(Kello& Kawamoto,1998).However, plosive-initial responseswereon tempoin the B-0 condition(timing %/. 2 ms),whereasnon-plosive-initial responseswere fast (timing%0. 18 ms), Fs(1,31) % 10.8,p & .05, Fi (1,153) % 32,p & .001. Finally, if timing wasbasedon the onsetofthe vowel,8then thereshouldbe no effect of voicing ofthe initial phonemesincethe vowel presumablybeginsat roughly thesamepoint in comparableresponseswithvoicedvs. unvoicedinitial phonemes.A post-hocsplitof thenon-plosive-initial wordsby voicing on theinitialphonemerevealedthat responsesto voicedstimuli were23 msslower thanunvoicedstimuli Fs(1,31) % 70, p &.001, Fi (1,91) % 32, p & .001.This effectof voicing in-teractedwith temposuchthat,aswith plosivity, theeffectdiminishedastempoincreased,Fs(3,93) % 2.4, p & .07,Fi (3,273) % 6.6, p & .001;wereturnto thiseffectbelow.Furthermore,themeantiming of voiced,non-plosiveini-tial wordsin theB-0 conditionwas . 3 ms,comparedto. 35 ms for comparableunvoiced words. The voicingonsethypothesisfor the basisof timing in temponam-ing explainsboththevoicingandplosivity effectsfound,but thevowel onsethypothesiscannotexplain theeffectof voicing. Therefore,thedataprovide evidencefor thevoicingonsethypothesis.

Naming Duration Analyses. Researchershaveshownthat cognitive processesaffect not only the onsetof anamingresponse,its articulatorydurationaswell (Balota

7 The phoneme1321 (e.g., in the first phonemein CHIP), istechnicallyanaffricate,but we treatedit asaplosivebecauseithasplosive-like characteristics.

8 Previous researchon the perceptualcenterof syllables(Hoequist,1983;Marcus,1981)suggeststhatvowel onset(orat leasta correlatethereof)may play a role in synchronizingrepeatedsyllableswith ametronome.


et4 al., 1989;Kawamotoet al., 1998;Kawamoto,Kello,Higareda,& Vu, 1999).For instance,Kawamotoandhiscolleaguesshowed initial phonemedurationeffects bycontrastingthe size of consistency effects in responsesbeginning with plosive versusnon-plosive phonemes.They arguedthat a larger consistency effect on latencyfor plosive versusnon-plosive initial stimuli (control-ling for confounds)is evidencethat consistency of thevowel affects the duration of the initial consonant(s).The logic wasbasedon the premise(mentionedabove)that the acousticonsetof non-plosive-initial responsescloselycorrespondto theactualresponseonset,whereasthe acousticonsetof plosive-initial responsesconflatesresponseonsetwith durationof theinitial phoneme.

If we apply the samelogic to the timing analysesconductedabove, then the interactionof stimulustype(with both voicing andplosivity) andtempoin the tim-ing analysesabove suggeststhat, at the least, namingdurationsfor unvoiced, non-plosive initial stimuli de-creasedastempoincreased.For example,if thedurationof the initial phonemecausedthe plosivity effect, thentheweakeningof this effectasa functionof tempoindi-catesa decreasein initial phonemeduration. Similarly,if durationof the entire namingresponsewas reducedfor both plosive andnon-plosive-initial words, thenthesameinteractionof tempoand plosivity would be pre-dicted(analogousargumentscouldbemadefor voicing).Wetestedthesealternatehypothesesbymeasuringwholewordnamingdurations(i.e., time from onsetto offsetofacousticenergy) asafunctionof initial phonemevoicingandplosivity, in conjunctionwith tempo.9Thepredictionswereasfollows. If tempoaffectedonly initial phonemeduration,thenthereshouldbenoeffectof tempoonnam-ing durationsfor plosive-initial responses.Sincea du-rationeffect on the initial phonemein plosive-initial re-sponseswill mostlyalterthesilentgapin acousticenergycausedby pressurebuild-up for the plosive release,thedurationeffectwill bereflectedin naminglatency ratherthan acousticduration. By contrast,if tempoaffectedwhole word durations,thenall responseswith any typeof initial phonemeshouldshow aneffectof tempo.

The results unambiguouslyshowed that tempo af-fectedwhole word durations. Figure 6 graphsnamingdurationsas a function of tempo and initial phonemetype. Namingdurationswerecalculatedusingthesamealgorithmfor detectingtheacousticonsetof a response,except that the algorithm was run backwardsfrom theend of responserecording(Kello & Kawamoto,1998,; durationsless than 50 ms or greaterthan 1000 mswere removed from the analyses). Naming durationsof plosive-initial responsesshowed a reliable effect oftempo, Fs(3,93) % 4.0, p & .01, Fi (3,303) % 6.9, p& .001. Themaineffect of tempoon namingdurations,collapsingacrossinitial-phonemetype,wasalsoreliable,Fs(3,93) % 7.4, p & .001, Fi (3,582) % 21, p & .001.This effect did not significantly interact with voicing,Fs(3,93) & 1, Fi (3,582) % 1.4, p ' .2, but it did in-teractmarginally with plosivity, Fs(3,93) % 1.6, p ' .2,

275

300

325

350

375

400


Nam

ing

Dur

atio

n (m

s)

0 -50 -100 -150

Figure 6. Mean acousticnamingdurationsfrom the temponamingtask in Experiment1, as a function of stimulustypeandtempo.

Fi (3,582)% 3.3,p & .05.Qualitatively,namingdurationsof non-plosive-initial responsesshowedastrongereffectof tempothanplosive-initial responses(which nonethe-lessshoweda reliableeffect of tempo;seeabove). Thisinteractionsimply indicatesthat initial phonemedura-tions contributedto the effect of tempoon whole worddurations,and that this contribution wasattenuatedforplosive-initial responsesbecauseplosivephonemeshavemuchlessacousticextentthannon-plosivephonemes.

Error Analyses. Errors were categorized as in thestandardnaminganalyses.Figure7 graphsmeanerrorrateasa functionof stimulustypeandtempo.Themaineffect of stimulustypewasreliable, Fs(2,62) % 38, p &.001, Fi (2,153) % 9.4, p & .001,aswasthemaineffectof tempo, Fs(2,93) % 11, p & .001, Fi (3,459) % 10,p & .001.As in thelatency analysis,theinteractionwasagainnon-significant,Fs(6,186)% 1.5,p ' .1, Fi (6,459)% 1.4, p ' .2. Plannedcomparisonsfor stimulustypeshowed that LFE wordswerereliably moreerror pronethanHFE words, Fs(1,31) % 55, p & .001, Fi (1,102) %13, p & .001,andlikewise for LFE wordscomparedtoLFC words,Fs(1,31) % 43, p & .001, Fi (1,102) % 9.4, p& .001.Errorrateresultsfromplannedcomparisonsoverlevelsof tempodifferedin part from thosefound in thelatency analyses.Whereasthe eachlevel of tempowasreliably fasterthanthe previous,only the increasefromB-100 to B-150showeda reliableincreasein error rate:from B-0 to B-50, Fs(1,31) & 1, Fi (1,153) & 1, fromB-50 to B-100, Fs(1,31) % 3.2, p ' .05, Fi (1,153) %2.7, p ' .1, andfrom B-100to B-150, Fs(1,31) % 10.2,p & .01, Fi (1,153) % 8.2, p & .01.

9 Thedurationsof mostinitial phonemesaredifficult tomea-sure(Kawamotoet al., 1998;Kello & Kawamoto,1998),andwedid notcarefullychoosestimuli with easilymeasuredinitialphonemedurations.We thereforechosewholeword durations

14 KELLO AND PLAUT

0

0.1

0.2

0.3

0.4


Err

or R

ate

Std 0 -50 -100 -150

Figure 7. Mean error rates(proportionof errors) from thestandardandtemponamingportionsof Experiment1 asafunc-tion of stimulustypeandtempo.Thedashedlinesseparatethestandardnamingmeansfrom thetempomeans.

Errorswerefurtheranalyzedby calculatingchi-squarestatisticson a frequency table of tempoby error type,presentedin Table3. Theoverall frequency countsweresignificantly different than their expectedvaluesbasedon row and column means χ2(12) % 37, p & .001.Our hypothesesconcernedhow the countsof differenterrortypeswould vary astempoincreased.TheMantel-Haenszel(M-H) chi-squaretestfor trend(Cody& Smith,1997)is especiallysuitedfor contingency tablesin whichone or both variableshave a specificordering of lev-els(i.e., it collapsestheindependentcontributionsto de-greesof freedomfor individual levelsof eachvariable).TheM-H chi-squaretestfor trendaskswhetherthecellcountsin eachof theN trendlevelsareincreasingor de-creasingin a uniformly linearfashion.Thenull hypoth-esisin the M-H testis that cell countsacrosstrendlev-elsarenot changinglinearly in proportionto eachother.This is exactly thequestionwewantedto askof ourdata:is theproportionof LARC errorsdecreasingsignificantlyasa functionof tempo(thetrendvariable),relativeto thenumberof othererrors? The M-H testapproachedsig-nificancefor theoverall frequency counts, χ2(1) % 2.9,p & .1.

Motivationfor analyzingerrortypesspecificallycamefrom predictionsconcerningLARC errors. To addressthesemoreclosely, LARC andmixederrors(which areLARCs themselves) were pooled together, and thesewerecomparedagainstword errors: the M-H testwasnow significant, χ2(1) % 10.4, p & .001, as was thechi-square,χ2(9) % 37, p & .001. M-H analyseswerealsoperformedonthesubsetof dataincludingonlyword,LARC, andmixederrors(with regularizationsandmixederrorscollapsed);theM-H testwassignificant, χ2(1) %5.2, p & .05, but the chi-squaretestwasnot, χ2(6) %5.8, p ' .1. Basedon thepatternof columnpercentsinthe frequency table, the chi-squareresultsindicatethat

thefrequency of LARC andmixederrorsremainedcon-stantwhile that of othererror typesincreasedastempoincreased. If we considerthe mixed errorsas mostlyLARCsthatcoincidentallyform words,thenwecancon-clude that the proportionof LARC errorsdecreasedastempoincreased.

We also divided LARC errors into true regular-ization errors (those following GPC rules) and non-regularizationerrors.Dependingonhow thetemponam-ing task relatesto the time courseof processing,theDRC model and the attractormodel of word readingmay make different predictionsconcerningtheseerrortypes. In particular, the DRC model predictsno non-regularizationerrorsabove chance(Coltheart,personalcommunication).By contrast,the attractormodel pre-dicts mostly regularizationerrors, but also somenon-regularizationerrors. As it turned out, a substantialnumberof non-regularizationerrorsdid occur(25%),al-thoughthemajority of LARC errorswereregularizationerrors(75%). The proportionof regularizationto non-regularizationerrorsdid not seemto changea functionof tempo(thecell countsweretoo small to performchi-squarestatistics).In orderof increasingtempo,thereg-ularizationcountswere15, 14, 16, and13, andthenon-regularizationcountswere5, 5, 4, and5. Theoccurrenceof non-regularizationerrorsmay beproblematicfor theDRC model,but this interpretationis dependenton howthetemponamingtaskis operationalizedin themodel.

Discussion

Theresultsof Experiment1canbesummarizedasfol-lows. Themanipulationof frequency andconsistency inthe standardnamingtask basically replicatedthe find-ingsof previousstudies.Responsesto HFE wordswerefasterthan thoseto LFE words, and responsesto LFCwordswerefasterthanthoseto LFE words(reliablebysubjectsonly). Error ratesalsoshowedthis pattern,butmorereliably. The temponamingtaskwaseffective ininducingprogressively faster, moreerrorproneresponses(responseswere94 ms fasterand7% lessaccurate,onaverage,than baselinein the fastesttempo condition).This resultindicatesthat,asLupker et al. (1997)noted,subjectsmustusea fairly conservative criterion (what-ever themechanism)to respondin thestandardspeedednamingtask(otherwiseonewouldexpecttempoto affectspeedlessand accuracy more, relative to the observedeffects).

Theeffectof stimulustypeon naminglatency dimin-ished in the tempo namingtask comparedto standardnaming,and this did not seemto be due to a practiceeffect (asindicatedby blockinganalyses),or achangeinitems(asindicatedby analysesof itemsubsets).Further-more,therewasno indicationof aninteractionof stimu-lus typewith tempoin thelatency analyses.An analysisof stimuli by plosivity andvoicingof theinitial phonemeindicatedthat subjectsattemptedto time the onsetof

asthedependentmeasureto examine.


T able3

Frequencycountsof errors in the standard and temponamingtasksin Experiment1, categorizedby error typeandtempo(for thetemponamingtask).

Std 0 -50 -100 -150 TotalLARC 25 20 19 20 18 77

(22.7) (19.8) (16.7) (10.8) (16.4)Word 38 33 43 54 60 190

(37.5) (44.8) (45.0) (36.1) (40.4)Mixed 31 20 14 17 17 68

(22.7) (14.6) (14.2) (10.2) (14.5)Nonword 31 10 8 21 39 78

(11.4) (8.3) (17.5) (23.5) (16.6)Articulatory 39 5 12 8 32 57

(5.7) (12.5) (6.7) (19.3) (12.1)Total 164 88 96 120 166 470

Note: Frequency countswere drawn from 2496 test responsesfor standardnaming (78 per subject),and 1248pertempoblock (39persubjectperblock).

voicing with the tempo. In addition,analysesof nam-ing durationshowedthatastempoincreased,durationoftheentirenamingresponsedecreased.Unlikenamingla-tency, thepatternof errorrateeffectswasessentiallythesamebetweenstandardandtemponaming:errorratestoHFE andLFC wordswerebothlower thanthoseto LFEwords. To complementthe latency results,error ratesincreasedwith tempo,indicatingaspeed/accuracy trade-off (albeit theonly reliableincreasewasfrom theB-100to B-150 condition). An analysisof the error typesasa function of temposhowed that while word errorsandothererror typesincreasedin numberwith increasesintempo,thenumberof LARC errorsremainedconstant.

The findingsfrom Experiment1 indicatethe follow-ing in termsof the two main researchagendasstatedattheoutset.First,subjectsarequitegoodat timing theini-tiation of a namingresponse,asevidencedby thestrongeffect of tempoin even in the fastestcondition. How-ever, the reducedbut enduringfrequency effect in thetemponamingtaskarguesagainstthe strongcue-drivenresponsehypothesis:subjectswereunableor unwillingto initiate a responsebasedon the cuealone. Note thatthelatency dataareprobablyconsistentwith thetimecri-terion hypothesis,despitethe evidenceagainstthe cue-driven hypothesis(dependingon how one handlestheaccompanying activationcriteria). Additionally, the la-tency dataareconsistentwith aweakcue-drivenhypoth-esisin which,on sometrials, responsesarebasedsolelyon the tempo,but on others,responsesarebasedsolelyonstimulusprocessing.

However, neitherof thesehypothesesaccountfor thedecreasein namingdurationsastempoincreased.Thesehypothesessimply donotaddressresponseexecution,ofwhich namingdurationis a crudemeasure.Onemightarguethatdurationeffectsfall outsidethescopeof thesehypotheses.However, understandingthe factorsunder-lying responsedurationwill shedlight on theprocessesunderlyingresponseinitiation becausethey are, in fact,both integral parts of the generationof pronunciation.

Therefore,we believe that an integratedaccountof ef-fectson responseinitiation aswell asexecutionis desir-able. In theGeneralDiscussion,we considerwhat typeof mechanismmightaccountfor thesedata.

The secondimmediateresearchquestionwas, cansubjectsbedrivento respondsubstantiallyfasterthaninthe standardnamingtask? The answeris clearly yes;in fact, becauselatenciesreliably decreasedby 27 msfrom the B-100 to B-150 conditions,we suspectthatsubjectscould be driven to respondcorrectly at evenshorterlatencies. The point of driving responsesto befast was generatenamingerrorsas a window into thetime courseof phonologicalprocessing.Theproportionof LARC errors(which were mostly regularizationer-rors) significantlydecreasedastempoincreased,duetoan increasein theoccurrenceof othererror types(worderrorsmostnotably). If responsesin the temponamingtaskreflectedearlierstatesof phonologyin the normalcourseof processing,thenthe decreasein proportionofLARCs would seemto be problematicfor both the tri-angleframework andthedual-routeframework of wordreading.As presentedearlier, simulationswith thePlautet al. (1996)modelandtheDRCmodel(Coltheartetal.,1993)showedanincreasein regularizationerrorsaspro-cessingin the modelswashaltedat successively earlierpoints in time. However, theseresultscannotbe usedasevidencefor or againstthesemodelsof word readingwithout specifyinghow the temponamingtask affectsprocessing. As a simplification, we useda strict timecriterion in thesimulations,but thepersistenceof a fre-quency effecton latenciesindicatesthatastrict time cri-terionis incorrect.Wereturnto this issuein theGeneralDiscussion.

Experiment2

Experiment2 wasintendedto replicateandextendtheresultsof Experiment1, andto testa routeemphasisac-countof the error patternin the temponamingportion

16 KELLO AND PLAUT

of Experiment5 1. The temponamingtask is novel, soit would be useful to know if the resultsfrom Experi-ment1 canbereplicatedwith anextendedsetof stimulianda secondgroupof subjects.More importantly, how-ever, we needto explain the decreasein proportionofLARC errorswith increasedtempos.Thusfar, we haveattributed this effect purely to the increasein pressurefor speed. However, the effect could have arisenfromstrategic factorsbasedon stimuluscomposition.Recallthat one motivation for this study camefrom a debateconcerningstrategic effectsin word naming. The routeemphasisaccountproposedthat subjectsemphasizeorde-emphasizeoneof theroutesbasedon compositionofthestimuluslist (Monsellet al., 1992),whereasthetimecriterion accountproposedthat subjectsadjusta crite-rion to initiatepronunciation(Jared,1997;Lupker et al.,1997).

In Experiment1, subjectsmay have de-emphasizedthesub-lexical routebecausecloseto half of all stimuli inExperiment1 containedexceptionalspelling-soundcor-respondences,andprocessingin thespelling-soundroutewill tendto interferewith thepronunciationof exceptionwords.Thefactthaterrorsto LFE wordsdecreasedfromthe first to secondblock in the standardnamingtask isconsistentwith theideathatsubjectsde-emphasizedthespelling-soundroute as they becamefamiliar with thestimuluscomposition.To testthis account,we includedpseudowordsasstimuli in Experiment2. Following thelogic laid out by Monsell et al. (1992), pseudowordsshouldinhibit de-emphasisof the spelling-soundroutesinceit is generallyrequiredto for correctpseudowordperformance.If thedecreasein proportionof LARC er-rorsacrosstempoin Experiment1 wasdueto astrategicde-emphasisof the spelling-soundroute,thenthe effectshouldbe diminishedwhenpseudowordsare included.If, however, the rate of LARC errorscontinuesto de-creasewith increasedtempos,a routeemphasisexplana-tion wouldbediscredited.

Methods

Subjects. Thirty-four subjectsparticipatedin the ex-perimentaspart of a requirementfor an undergraduatepsychologycourse. Subjectsreportedbeingnative En-glishspeakerswith normalor correctedvision.

Stimuli. All test stimuli from Experiment1 were in-cluded. In addition, 52 pseudowords were createdbyshuffling theonsetsandbodiesof theLFC words(listedin Appendix A). All 52 pseudowords appearedin thetemponamingtask, but noneof theseappearedin thestandardnaming task; an additional 26 pseudowordswerecreatedto appearexclusively in thestandardnam-ing blocks.Carewastakento avoid pseudohomophones(e.g., BRANE). Standardnaming and tempo namingblocksof trialswerecreatedin thesamewayasin Exper-iment1. An equalportionof eachstimulustypeappearedin eachtestblockwithin standardandtemponaming,andfillers were divided equally amongtest blocksaswell.

Therewere 108 fillers in the temponaming task, and1/4 of thesewere pseudowords. The standardnamingand temponamingpracticeblocksconsistedof 10 and40 fillers, respectively, and 1/4 of both practiceblockswere pseudowords. Therefore,1/4 of all stimuli werepseudowords.

Procedure. Theprocedurewasthesameasthat in Ex-periment1, except subjectswere told that someletterstringswere not legal English words. As with words,they wereto nametheseletter stringsasquickly andasaccuratelyaspossible.


Data Removal. Data from 2 subjectswere removeddue to equipmentfailure, and data from 1 item (chic)was removed as in Experiment1. Pseudoword errorswerecategorizedaswordswerein Experiment1, exceptregularizationand mixed were not possible(all pseu-doword bodiescontainedconsistentspelling-soundcor-respondences).Furtherdataremoval wascarriedout asin Experiment1.

Naming Latency Analyses. Therewasa reliablemaineffect of stimulus type, Fs(3,93) % 25, p & .001,Fi (3,99) % 11.1, p & .001,but not block, Fs(1,31) & 1,Fi (1,99) & 1. Therewasno reliableinteractionof blockandstimulustype, Fs(3,93) & 1, Fi (3,99) & 1. Pairwisecomparisonsrevealeda reliable18 msfrequency effect,Fs(1,31) % 17, p & .001, Fi (1,49) % 7.2, p & .01, anda reliable15 msconsistency effect, Fs(1,31) % 15, p &.001, Fi (1,49) % 5.1, p & .05. The overall meanlaten-ciesof wordscomparedto pseudowordswasalsotested(a lexicality effect); latenciesto pseudowords were 35msslower thanto wordsoverall, andthis differencewassignificant, Fs(1,31) % 36, p & .001, Fi (1,101) % 24, p& .001.

Error Analyses. Errorswerecategorizedandanalyzedasin Experiment1. Therewasa reliablemaineffect ofstimulus type, Fs(3,93) % 31, p & .001, Fi (3,99) %10.4, p & .001, but no main effect of block, Fs(1,31)& 1, Fi (1,99) & 1. Theinteractionof block with stimu-lus wassignificantby subjectsFs(3,93) % 2.6, p & .05,but not by items Fi (3,99) % 1.7, p ' .1. Although thelackof a fully significantinteractionprohibitedpost-hocanalyses,thecell meansclearlyshow a differentpatternof resultscomparedto Experiment1: whereasas LFEerrorsdecreasedfrom block 1 to 2 in Experiment1, theyremainedconstantin Experiment2, andpseudoword er-rorsdecreasedfrom block 1 to 2. Plannedcomparisonsshowedareliable12.8%consistency effectonerrorrates,Fs(1,31) % 59, p & .001, Fi (1,49) % 15, p & .001,anda reliable12.9%frequency effect, Fs(1,31) % 70, p &.001, Fi (1,49) % 14, p & .001. Finally, therewasnoreliablelexicality effect, Fs(1,31) & 1, Fi (1,101) & 1.


350

400

450

500

550

Hi Freq Exc Lo Freq Exc Lo Freq Con Pseudo

Nam

ing

Late

ncy

(ms)

Std 0 -50 -100 -150

Figure 8. Meanlatenciesfrom thestandardandtemponam-ing portionsof Experiment2 asafunctionof stimulustypeandtempo. The dashedlinesseparatethe standardnamingmeansfrom thetempomeans.

Frequency countsof error typesby block were alsoanalyzed. Cell countswere not significantly differentthan their expectedvaluesbasedon row and columnmeanscalculatedacrosslevels of block anderror type,χ2(4) % 4.7, p ' .2. To comparewith Experiment1,countswerealsotallied with thepseudowordsremoved,andthesetoo did not differ reliably from their expectedvalues,χ2(4) % 3.4, p ' .5.


Data Removal. The procedurefor dataremoval wasthesameasthatin Experiment1. Latency Analyses.

Latency Analyses. Figure8 graphsmeannamingla-tenciesas a function of stimulus type and tempo (in-cludingthestandardnamingmeans).Themaineffectofstimulustypewassignificant,Fs(3,93) % 15.5,p & .001,Fi (3,204) % 3.3,p & .05,aswasthemaineffectof tempo,Fs(3,93) % 439, p & .001, Fi (3,612) % 246, p & .001.The interactiondid not reachsignificance, Fs(9,279) %1.5,p ' .1, Fi (9,612) & 1. Plannedcomparisonsshowedthatthe9 msmaineffectof frequency effectwasreliableby subjects, Fs(1,31) % 16.3, p & .001,but not items,Fi (1,102) % 2.4, p ' .1, aswasthe6 msmain effect ofconsistency effect, Fs(1,31) % 9.7, p & .01, Fi (1,102)% 1.8, p ' .1. The 9 ms main effect of lexicality wasreliableaswell, Fs(1,31) % 24, p & .001, Fi (1,206) %7.0, p & .01. Plannedcomparisonsof thetempomanip-ulation confirmedthat eachsuccessively fasterlevel oftempocausedresponsesto bereliably fasterthanthepre-viouslevel: from B-0 to B-50, Fs(1,31) % 223,p & .001,Fi (1,207) % 94, p & .001,from B-50 to B-100, Fs(1,31)% 234, p & .001, Fi (1,207) % 120, p & .001,andfromB-100to B-150, Fs(1,31) % 87, p & .001, Fi (1,207) %40, p & .001.

As in Experiment1, theinfluenceof stimulustypeonlatencieswas smaller in the tempo naming task com-paredto standardnaming(an 18 ms frequency effect,15 ms consistency effect, and 35 ms lexicality effect,versus9 ms, 6 ms, and 9 ms effects in tempo nam-ing, respectively). Also replicatingExperiment1, thewithin-conditionvariability wasnot significantlydiffer-ent betweenthe standardand temponamingtasks(9.5ms versus12.6ms respectively, within-cell standarder-rorsaroundthesubjectmean),Fs(1,31) & 1.

The conclusionthat stimuluseffectswereattenuatedin temponamingis supportedby thefollowing analyses(asin Experiment1). First,weremovedrepeatedstimuli(andall pseudowordssincenoneof thesewererepeated)from thetemponaminganalyses,andtheeffectsof stim-ulus type were essentiallyunchanged,albeit therewasa lossof power by items(all effectsweresignificantbysubjectsbut not items): themaineffectof stimulustype,Fs(2,62) % 8.2, p & .01, Fi (2,114) % 1.3, p ' .2, an 8ms frequency effect, Fs(1,31) % 9.1, p & .01, Fi (1,76)% 1.3, p ' .2, anda 9 ms consistency effect, Fs(1,31)% 12.4, p & .01, Fi (1,76) % 2.2, p ' .1. Second,theanalysesof block orderandstimulustypeonceagainre-vealedno discernibleeffect of block order, Fs(3,93) &1, Fi (3,612) % 2.1, p ' .1. Third andfinally, we an-alyzedstandardnaminglatenciesincluding only temponamingstimuli (therebyexcludingall pseudowords),andthemaineffectof stimulustypewasstill reliableby sub-jects, Fs(2,62) % 5.7,p & .01,but notby items, Fi (2,36)% 1.4, p ' .2. Plannedcomparisonsshowedthat the16msfrequency effect (cf. an18 mseffect with all stimuliincluded)wasreliableby subjectsonly, Fs(1,31) % 8.8,p & .01, Fi (1,24) % 2.9,p & .1,aswasthe12msconsis-tency effect(cf. a15mseffectwith all stimuli), Fs(1,31)% 7.5, p & .01, Fi (1,24) % 1.7, p ' .2. Takentogether,thesethreeanalysesindicatethat thedecreasedstimuluseffect on latenciesin the temponamingtaskwasduetothetaskitself, ratherthanpracticeor stimulusselection.

Timing Analyses. Figure9 graphsmeantiming offsetsasa function of stimulustype andtempo. As in the la-tency analyses,themaineffectof stimulustypewasreli-able, Fs(3,93) % 15.5, p & .001, Fi (3,204) % 2.7, p &.05, aswasthe main effect of tempo, Fs(3,93) % 86, p& .001, Fi (3,612) % 165, p & .001. The interactionofstimulustype andtempowasnot significant, Fs(9,279)% 1.5, p ' .1, Fi (9,612) % 1.3, p ' .2. Plannedcompar-isonsshowed that, as in Experiment1, responseswereincreasinglydelayedfrom tempoat eachstep:B-0 to B-50, Fs(1,31) % 28, p & .001, Fi (1,207) % 32, p & .001,B-50 to B-100, Fs(1,31) % 10.0, p & .01. Fi (1,207) %24, p & .001,andB-100 to B-150, Fs(1,31) % 85, p &.001, Fi (1,207) % 120,p & .001.

The analysesof timing as a function of initialphonemein Experiment1 supportedthehypothesisthatsubjectstimedtheirresponsesbasedontheonsetof voic-ing. Themainpieceof evidencewasthatresponseswithvoiced non-plosive-initial phonemeswere timed more

18 KELLO AND PLAUT

-30

-15

0

15

30

45

60


Tim

ing

(ms)

0 -50 -100 -150

Figure 9. Meantiming with tempofrom Experiment2 asafunctionof stimulustypeandtempo.

accuratelythan thosewith unvoiced ones,which weretoo fast in the B-0 condition. Timing in Experiment2wasalso analyzedasa function of voicing and tempo:Responsesto voicednon-plosive-initial responseswere28 ms slower (cf. 23 ms effect in Experiment1) thanthe unvoicedcounterparts, Fs(1,31) % 104, p & .001,Fi (1,123) % 54, p & .001,andthis effect interactedwithtempoin the sameway asin Experiment1, Fs(3,93) %3.3, p & .05, Fi (1,123) % 3.5, p & .05. In particular,the voicing effect steadilydecreasedfrom a 38 ms dif-ferencein the B-0 condition,to an 18 ms in the B-150condition. Finally, timing of non-plosive-initial voicedresponseswas3 ms in the B-0 condition, comparedto. 35msfor theunvoicedresponses.Theseanalysescor-roboratethosefrom Experiment1 in showing that sub-jectstimedtheir responsesbasedontheonsetof voicing,ratherthanothercandidatessuchastheonsetof acousticenergy or the vowel. They alsoshow indirect evidenceof adecreasein namingdurationastempoincreased.

Naming Duration Analyses. Namingdurationanaly-sesin Experiment1 showed that an increasein tempocauseda decreasein whole word namingduration forall initial phonemetypes. Namingdurationsin Exper-iment 2 replicatedthe pattern found in Experiment1(shown in Figure10: durations,collapsedacrossinitialphonemetype,steadilyandreliably decreasedfrom B-0to B-150, Fs(3,93) % 10.3,p & .001, Fi (3,612) % 20, p& .001. Thesamepatternof effectsheldwhenanalyseswererestrictedto voicedresponses,Fs(3,93) % 4.0, p &.01, Fi (3,177) % 6.7, p & .001,aswell asplosive-initialresponses,Fs(3,93) % 7.3, p & .001, Fi (3,243) % 11.5,p & .001. The fact that the namingdurationeffect heldfor bothof theabove initial phonemetypesindicatesthattherimeportionof theresponsedecreasedin duration,aswell astheonset.

6 7 89 : :9 6 89 8 :9 7 8; : :


Nam

ing

Dur

atio

n (m

s)

: < 8 : < = : : < = 8 :

Figure 10. Meannamingdurationsfrom the temponamingportion of Experiment2 as a function of stimulus type andtempo.

Error Analyses. Figure11graphsmeanerrorrateasafunctionof stimulustypeandtempo(includingstandardnamingmeans). The error rate resultsreplicatedthoseof Experiment1. The main effect of stimulustype wasreliable, Fs(3,93) % 55, p & .001, Fi (3,204) % 14.9,p & .001,aswasthe main effect of tempo, Fs(3,93) %13.9, p & .001, Fi (3,612) % 17.0, p & .001. The inter-actionof stimulustype and tempowasnon-significant,Fs(9,279) & 1, Fi (9,612) % 1.3, p ' .2. Pairwisecom-parisonsshowed that 7.2% increasein erredresponsesto LFE over HFE words was reliable, Fs(1,31) % 35,p & .001, Fi (1,102) % 12.6, p & .01, aswasthe 7.4%increasein LFE over HFE errors, Fs(1,31) % 60, p &.001, Fi (1,102) % 15.5, p & .001. To test the extentto which error ratesincreasedwith tempoas latenciesdecreased(i.e., a speed-accuracy tradeoff), error ratesbetweenadjacentlevels of tempowere compared.Al-thougherror rateincreasedwith eachincreasein tempo,only thechangefrom B-50 to B-100wassignificant(cf.only thechangefrom B-100to B-150wassignificantinExperiment1): fromB-0 to B-50, Fs(1,31) % 1.9,p ' .1,Fi (1,207) % 2.3, p ' .1, from B-50 to B-100, Fs(1,31)% 11.9, p & .01, Fi (1,207) % 14.2, p & .001,andfromB-100to B-150, Fs(1,31) & 1, Fi (1,207) & 1.

Errorswerefurtheranalyzedbycalculatingchi-squarestatisticson the frequency table of error counts, bro-ken down by tempoby error type (shown in Table 4).With LARCs and mixed errors treatedseparately, theoverall frequency countswerenot significantlydifferentthantheir expectedvaluesin a chi-squaretest,basedonrow andcolumnmeans, χ2(12) % 17.8, p ' .1. How-ever, whenthesetwo errortypeswerecombined,thechi-squaretestwassignificant, χ2(9) % 16.7, p & .05. Thepurposeof analyzingerror typesherewas to comparethemwith the analogousanalysesin Experiment1. Inparticular, we wantedto know whethertheproportionofLARC errorsfell astempoincreased(asthey did in Ex-


0

0.1

0.2

0.3

0.4


Err

or R

ate

Std 0 -50 -100 -150

Figure 11. Meanerror rates(proportionof errors)from thestandardandtemponamingportionsof Experiment2 asafunc-tion of stimulustype and tempo. The dashedlines separatestandardnamingmeansfrom thetempomeans.

periment1), so we conductedthe M-H chi-squaretest.TheM-H testwassignificantwhencomparingword er-rors,LARCs, andmixederrorsfor all stimuli, χ2(1) %5.6, p & .05(LARCsandmixederrorsseparated),χ2(1)% 8.3, p & .01 (combined). The M-H was significantalsowhenpseudowordswereexcludedfrom the analy-sisto comparewith thecorrespondingresultsof Experi-ment1, χ2(1) % 4.4, p & .05 (separated),χ2(1) % 6.2,p & .05 (combined).Finally, asin Experiment1, most,but not all, LARC errorswereregularizations.In addi-tion, theproportionof regularizationsdid not changeasa function of tempo. In orderof increasingtempo,theregularizationcountswere24, 23, 21, and20, and thenon-regularizationcountswere4, 3, 10,and5.

Discussion

Overall, the resultsof Experiment2 replicatedthoseof Experiment1. Tempoagainhada stronginfluenceonresponselatenciesanderrorrates(i.e.,it inducedsubjectsto incrementallytradeaccuracy for speed),but subjectsdid not fully keeppacewith thetempoat thefasterrates.Timing analysesshowedagainthatsubjectsattemptedtotimetheir responsesbasedontheonsetof voicing,ratherthanthe onsetof articulationor of the vowel. The fre-quency effect was diminished(but reliable) from stan-dard to tempo naming, as was the consistency effect.Naming durationanalysesagainshowed that durationsdecreasedwith increasedtempo.

Two differencesbetweentheresultsof Experiments1and2 werethat1) theconsistency effect wasmorereli-ablein both thestandardandtemponamingportionsofExperiment2, and2) theaccuracy of namingLFE wordsdroppedfrom thefirst to secondblock of standardnam-ing in Experiment2, whereasthe accuracy of namingpseudowordsincreased;errorproportionsdid notsignifi-cantlychangein thestandardnamingblocksfrom Exper-

iment 1. Theseresultsseemto indicatethat assubjectssaw moreandmorepseudowordsduringthefirst blockofstandardnaming,they emphasizedthe sub-lexical route(and/orde-emphasizedthelexical route)to facilitatepro-cessing.On the otherhand,the routeemphasishypoth-esisalso predictsthat LARC errorsshouldincreaseassubjectstradespeedfor accuracy. To the contrary, theproportionof LARCs droppedwith increasesin tempo,replicatingExperiment1.

Taken together, the resultsfrom Experiments1 and2 provide moreevidenceagainsttherouteemphasishy-pothesisasanexplanationof theerrorpatternsfoundinthe temponamingtask,but the resultsweremixed. InExperiment3, we conducteda further testof the routeemphasisexplanation.

Experiment3

If stimuluscompositioncancausesubjectsto prefer-entiallyemphasizethesub-lexical route(aswashintedatin the patternof error ratesin the standardnamingpor-tion of Experiment2), then perhapsthe strongestwayto encouragesub-lexical emphasiswould be to presentsubjectswith a stimulus block consistingof all non-homophonicpseudowords. Theseitemsshouldrely al-most exclusively on the sub-lexical processingroute.Theproblemwith usingblocksof all pseudowordsis thattherewould be no opportunityto observe LARC errors(i.e., for pseudowordswith inconsistentbodies,all alter-native pronunciationsshouldbe consideredlegitimate).The proportionof LARC errorsserved asa measureofemphasisplacedon thesub-lexical routein Experiments1 and2. A stimulusblock of all pseudowordscanstillbeuseful,however, becausewe caninsteadlook for evi-denceof de-emphasisof thelexical route.

In particular, if word errors to pseudoword targetsarise,at leastin part,from processingin thelexical route,thensucherrorsshoulddecreasein proportionwhenthelexical routeis de-emphasized,comparedto thepropor-tion of word errorsfoundin Experiments1 and2. Alter-nately, if thelexical routeisnotde-emphasizedin ablockof all pseudowords,thenthe rateof word errorsshouldbe approximatelyequalto that found in Experiments1and2. Wetestedthesetwo possibilitiesin Experiment3.

Methods

Subjects. Thirty-six subjectsparticipatedin theexper-imentaspartof a requirementfor anundergraduatepsy-chologycourse.Subjectsreportedbeingnative Englishspeakerswith normalor correctedvision.

Stimuli. Out of the 52 pseudowords from Experi-ment2 (all createdby mixing LFC onsetsandbodies),48 wereincludedin Experiment3. Theonsetsandbod-ies of an additional48 LFE wordsand 48 HFE wordsfrom Experiments1 and2 weremixed to createan ad-ditional 96 pseudowords, for a total of 144 test pseu-dowordsin Experiment3 (all testpseudowordsaregiven

20 KELLO AND PLAUT

Table4 4Frequencycountsof errors in the standard and temponamingtasksin Experiment2, categorizedby error typeandtempo(for thetemponamingtask).

Std 0 -50 -100 -150 TotalLARC 25 28 26 31 25 110

(17.7) (14.9) (12.8) (9.0) (12.9)Word 31 63 70 108 128 369

(39.9) (40.2) (44.4) (46.0) (43.3)Mixed 14 16 18 17 21 72

(10.1) (10.3) (7.0) (7.6) (8.4)Nonword 48 39 39 54 57 189

(24.7) (22.4) (22.2) (20.5) (22.2)Articulatory 17 12 21 33 47 113

(7.6) (12.1) (13.6) (16.9) (13.3)Total 135 158 174 243 278 853

Note: Frequency counts were drawn from 3328 test responsesfor standardnaming (104 per subject), and1664pertempoblock (52 persubjectperblock).

in AppendixB). An additional50pseudowordswerecre-atedfor the standardnamingtaskin orderto measureabaselinenaminglatency for eachsubject.Standardnam-ing and temponamingblocksof trials were createdinthesameway asin Experiments1 and2. An equalpor-tion of eachpseudoword type (i.e., thosecreatedfromLFC, LFE, or HFE words)appearedin eachtestblockwithin temponaming,and fillers were divided equallyamongtestblocksaswell. Therewere40 pseudowordfillers in thetemponamingtaskand2 pseudowordfillersin the standardnamingtask. The standardnamingandtemponamingpracticeblocks consistedof 10 and 40pseudowordfillers, respectively.

Procedure. Theprocedurewasthesameasthatin Ex-periment1, except subjectswere told that noneof theletterstringswould make legal Englishwords. Subjectswere instructedto nametheseletter stringsas quicklyandasaccuratelyaspossible.


Data Removal. Data from 4 subjectswere removeddueto equipmentfailure. Furtherdataremoval wascar-riedout asin Experiments1 and2.

Naming Latency Analyses. Meanlatencieswereana-lyzedasa functionof block. A 15 ms increasein meanlatency from block 1 (538ms) to block 2 (553ms)wasnotsignificant, Fs(1,31) % 1.8, p ' .1, Fi (1,49) & 1.

Error Analyses. Errorswerecategorizedandanalyzedasin Experiments1 and2. For pseudowordswith incon-sistentword bodies,all legitimatespelling-soundcorre-spondenceswerecodedascorrect. A 2.1% increaseinerror rate from block 1 (4.2%) to block 2 (6.3%) wasnot significant, Fs(1,31) % 2.5, p ' .1, Fi (1,49) & 1.Countswere not reliably different than their expected

valuesbasedon meanscalculatedacrosslevelsof blockanderrortype, χ2(2) & 1.


Data Removal. The procedurefor dataremoval wasthesameasin Experiments1 and2 (therewasnooppor-tunity for LARC errors).

Latency Analyses. Figure12 graphsmeannamingla-tenciesfrom the standardandtemponamingtasksasafunctionof tempoandstimulustype. Tempoonceagainhad a stronginfluenceon responselatencies, Fs(3,93)% 345, p & .001, Fi (3,423) % 166, p & .001, but themain effect of stimulus type was significantly reducedcomparedto Experiments1 and 2, Fs(2,62) % 3.4, p& .05, Fi (2,141) & 1. The interactionof stimulustypeandtempowasnotsignificant, Fs(6,186) & 1, Fi (6,423)& 1. Plannedcomparisonsshowed that the “pseudo-frequency” effect (i.e., latenciesof pseudowordscreatedfrom HFE versusLFE words)wasnot reliable, Fs(1,31)% 2.8, p ' .1, Fi (1,94) & 1, but thepseudo-consistencyeffectwasreliableby subjects,Fs(1,31) % 6.4, p & .05,Fi (1,94) & 1. Plannedcomparisonsof the tempoma-nipulationconfirmedthat eachsuccessively fasterlevelof tempocausedresponsesto be reliably fasterthantheprevious level: from B-0 to B-50, Fs(1,31) % 187, p& .001, Fi (1,141) % 61, p & .001, from B-50 to B-100, Fs(1,31) % 128, p & .001, Fi (1,141) % 46, p& .001, and from B-100 to B-150, Fs(1,31) % 150, p& .001, Fi (1,141) % 71, p & .001. The small, par-tially reliable pseudo-consistency effects was presum-ably duetheambiguousspelling-soundcorrespondencesthat thesestringscontained(e.g.,BOST canrhymewithCOST or MOST; Glushko, 1979; Seidenberg, Plaut,Pe-tersen,McClelland,& McRae,1994;Taraban& McClel-land,1987).


350

400

450

500

550

Pseudo Hi Exc Pseudo Lo Exc Pseudo Lo Con

Nam

ing

Late

ncy

(ms)

Std 0 -50 -100 -150

Figure12. Meanlatenciesfrom thestandardandtemponam-ing portionsof Experiment3 asafunctionof stimulustypeandtempo.Thedashedlinesseparatestandardnamingmeansfromthetempomeans.

-30

-15

0

15

30

45

60


Tim

ing

(ms)

0 -50 -100 -150

Figure 13. Meantimingswith tempofrom Experiment3 asafunctionof stimulustypeandtempo.

Timing Analyses. Figure13 graphsmeantiming off-setsas a function of stimulus type and tempo. As inthe latency analyses,the main effect of stimulus typewasonly reliableby subjects, Fs(3,62) % 3.4, p & .05,Fi (2,141) & 1. In addition,themaineffectof tempowassignificantby subjectsand items, Fs(3,93) % 26, p &.001, Fi (3,423) % 53, p & .001,andtheinteractionwasnot significant, Fs(6,186) & 1, Fi (6,423) & 1. Plannedcomparisonsshowedthat,asin Experiments1 and2, re-sponseswere increasinglydelayedfrom tempoat eachstep:B-0 to B-50, Fs(1,31) % 18, p & .001, Fi (1,141) %25, p & .001,B-50 to B-100, Fs(1,31) % 16.0,p & .01,Fi (1,141) % 20, p & .001,andB-100to B-150, Fs(1,31)% 2.9, p ' .1, Fi (1,141) % 5.9, p & .05.

Timing analyseswereagainconductedasa functionof initial phonemecharacteristicsto test the hypothesisthatsubjectsattemptto timetheir responsesbasedonthe

275

300

325

350

375

400


Nam

ing

Dur

atio

n (m

s)

0 -50 -100 -150

Figure 14. Meannamingdurationsin thetemponamingpor-tion of Experiment3 asa functionof stimulustypeandtempo.

onsetof voicing. As in Experiments1and2, responsestovoiced,non-plosive stimuli werereliably slower (32 msdifference)thanthoseto unvoiced,non-plosive stimuli,Fs(1,31) % 68, p & .001, Fi (1,85) % 44, p & .001,andthis effect interactedwith tempo,albeit reliableonly bysubjects, Fs(3,93) % 3.3, p & .05, Fi (3,255) % 2.0, p' .1. In particular, the voicedeffect steadilydecreasedfrom a40msdifferencein theB-0 condition,to an24msin the B-150 condition. Finally, timing of voiced,non-plosive itemswas0 ms in the B-0 condition,comparedto . 40msfor unvoiced,non-plosiveresponses.

Naming Duration Analyses. Namingdurationanaly-seswereagainconductedto show thattheentireresponsedecreasedin durationastempoincreased,ratherthanjustthe initial phoneme.The resultsreplicatedthoseof Ex-periments1 and2 (shown in Figure14: Namingdura-tions,collapsingacrossinitial phonemetype,steadilyde-creasedfrom B-0 to B-150, Fs(3,93) % 5.3, p & .01,Fi (3,423) % 16.8, p & .001. The samepatternof ef-fectsheldwhenanalyseswererestrictedto voiced,non-plosivepseudowords, Fs(3,93) % 2.8,p & .05, Fi (3,117)% 6.3, p & .001,aswell asplosive-initial pseudowords,Fs(3,93) % 4.0, p & .01, Fi (3,165) % 5.6, p & .01. Thefactthatthewholeresponsedurationeffectheldfor bothof theabove initial phonemetypesindicatesthattherimeportionof theresponsedecreasedin duration,aswell astheonset.

Error Analyses. Figure15graphsmeanerrorratesforthe standardand tempoportionsof Experiment3 as afunctionof stimulustypeandtempo.Errorratesagainin-creasedastempoincreased,indicatinga speed/accuracytradeoff, Fs(3,93) % 5.4, p & .01, Fi (3,423) % 8.8, p &.001.However, unlikenaminglatencies,errorrateswereonly marginally affectedby stimulustype, Fs(2,62) %2.7, p & .1, Fi (2,141) & 1. The interactionof stimulus

22 KELLO AND PLAUT

0

0.1

0.2

0.3

0.4


Err

or R

ate

Std 0 -50 -100 -150

Figure 15. Mean error ratesfrom the standardand temponamingportionsof Experiment3 asafunctionof stimulustypeandtempo.

type and tempowas also not significant, Fs(6,186) %1.3, p ' .2, Fi (6,423) % 1.4, p ' .2. Error ratesgener-ally increasedwith eachincreasein tempo,but not eachincreasewasreliable: from B-0 to B-50 wasnot signifi-cant, Fs(1,31) & 1, Fi (1,141) & 1, from B-50 to B-100wasreliableby itemsbut notsubjects,Fs(1,31) % 1.9, p' .1, Fi (1,141)% 3.9,p & .05,andfrom B-100to B-150wasfully reliable, Fs(1,31) % 5.9, p & .05, Fi (1,141) %4.2, p & .05.

Errorcountsasa functionof errortypeandtempoarepresentedin Table5. To thecontraryof therouteempha-sishypothesis,theproportionof word errorswasgreaterin thecurrentexperiment(53%)thanin theprevioustwo(40% in Experiment1 and43% in Experiment2). Fur-thermore,the rate of increasein word errorsas tempoincreasedwasapproximatelyequalto that for nonworderrors,as was the casein Experiments1 and 2. TheM-H test for trend was significant, χ2(1) % 4.8, p &.05,but inspectionof thecolumnpercentsshowsthatthiswasdueto agreaterrateof increasein articulatoryerrorsacrosstempo,comparedto word andnonword errors.

Discussion

The results of Experiment3 were consistentwiththoseof Experiments1 and 2. Timing analysesagainsuggestedthatsubjectstime their responseswith theon-setof voicing in the temponamingtask. Furthermore,thenumberof word,nonword, andarticulatoryerrorsasa functionof tempowerecomparableto thoseof Exper-iments1 and2: they all steadilyincreasedastempowasincreased.If the routeemphasishypothesiswascorrectin accountingfor the constantnumberof LARC errorsacrosstemposin Experiments1 and2, thenoneshouldalsoexpecta decreasein the numberof word errorsinExperiment3, which includednothingbut pseudowords.The resultsfailed to supportthis prediction. Therearetwoassumptionsto benotedbeforeconcludingthatroute

emphasiswasnot the causeof error patternsin Experi-ments1 through3.

First, Tabossiand Laghi (1992)have suggestedthatin English,subjectsmay beunableto de-emphasizethelexical route. If this is thecase,thenExperiment3 doesnot bearon the route emphasisexplanation. However,even if the lexical route cannotbe de-emphasized,onewould not expectan increasein the influenceof lexicalknowledgein Experiment3, asa functionof tempo.Wedid, in fact, find a significantincreasein word errorsasa function of tempo,so the studyby TabossiandLaghi(1992)wouldnotseemto beanissue.This logic restsonthe interpretationof word errorsasarisingfrom lexicalknowledge,but they mayhave occurredby chance(e.g.,faultyvisualor articulatoryprocessing).

We addressedthis ambiguity by estimatingthe rateof chanceword errors in Experiment3. Determiningchanceerror ratesis a difficult problemsinceonemusttake into considerationthe specificstimuli and condi-tionsof theobservederrors(e.g.,neighborhooddensity),the typesof errorsthat arepossible(e.g.,only phoneti-cally legalerrorsareallowed),aswell asany errorbiasesthatmaybeindependentof a lexical bias(e.g.,in speechproduction,initial consonantsaremorelikely to produceerrorsthanfinal ones;Dell, 1988;Schwartz& Goffman,1995).Weempiricallydeterminedachanceerrorrateforthestimuli in Experiment3, usingthemostconservativemethodthatwecoulddevisegivenour intentions.

We first estimatedthe frequency of occurrenceof alargenumberof phonologicalerrortypes(e.g.,position-dependentphonemesubstitutions,deletions,insertions,andtranspositions)basedon the observed errorsin Ex-periments1 through3. We thenappliedeachphonolog-ical error type (when applicable),weightedby the fre-quency estimates,to eachtest item, therebygeneratinga set of possibleerror instances. Our estimateof thechancefrequency of word errorswas the proportionoferrorsthat formedlegal Englishwordsout of the setofempirically determinedpossibleerrors. This methodisconservative becausethe estimatedrateof eachphono-logical error type is determinedbasedon all observederrors(i.e., including word errors);theseestimateswillthereforeincludeany real lexical biasesthatmaycorre-latewith phonologicalbiases(e.g.,initial consonantsub-stitutionsmay causemoreword errorsthanfinal conso-nantsubstitutions).Thechanceestimateof theword er-ror ratefor Experiment3 was28%.10Theobservedrateofword errorswas53%,which wassignificantlydifferentthanchance,t(27) % 8.1,p & .001by subjects;t(120) %6.9,p & .001by items).

10 As a comparison,Garrett (1976) and Dell and Reich(1981) estimatedthe chanceoccurrenceof word errors inspeechproduction (i.e., within phrasesand sentences),andtheirestimateswere33%and40%,respectively. Notethatdif-ferent type of errorscan and do occur in speechproductionversuswordnaming(e.g.,transpositionsbetweenwords;“barndoor” goesto “darn bore”), so oneshouldnot necessarilyex-pecttheseestimatesto match.


T able5

Frequencycountsof errors in the standard and temponamingtasksin Experiment3, categorizedby error typeandtempo(for thetemponamingtask).

Std 0 -50 -100 -150 TotalWord 38 46 41 51 68 206

(63.0) (51.3) (48.1) (51.5) (52.7)Nonword 39 23 33 45 43 144

(31.5) (41.3) (42.5) (32.6) (36.8)Articulatory 8 4 6 10 21 41

(5.5) (7.5) (9.4) (15.9) (10.5)Total 85 73 80 106 132 391

Note: Frequency countswere drawn from 1600 test responsesfor standardnaming (50 per subject),and 1152pertempoblock (48persubjectperblock).

In summary, theassumptionsthatde-emphasisof thelexical routeis possiblein English,andthat lexical pro-cessescontributed to the observed rate of word errors,bothseemjustified.Therefore,theproportionof worder-rorsobservedin Experiment3standsasindirectevidenceagainstthe routeemphasisexplanationof the patternofLARC errorsfoundin Experiments1 and2.

GeneralDiscussion

We set two goalsat the beginning of this study: 1)to formulatea moreexplicit mechanismof control overresponsetiming, and2) to formulateahypothesisof howpressurefor speedrelatesto the time courseof process-ing. The resultsfrom Experiments1 through3 canbesummarizedasfollows:# Subjectswerelargelyableto entrainresponseiniti-ationto anexternaltempo.# Stimuluseffectson latencieswereattenuatedin thetemponaming task, comparedto the standardnamingtask.# Thetempomanipulationinducednaminglatenciesthatweresubstantiallyfasterandsignificantlymoreerrorpronethanin thestandardnamingtask.# Namingdurationsdecreasedastempoincreased.# Therateof LARC errorsdid not increasewith othererror typesastempoincreased.This resultwasnot at-tributableto ade-emphasisof thesub-lexical route.# Subjectstimedtheir responseswith theonsetvoic-ing. This finding is orthogonalto the issuesat hand,sowe donot considerit further.

The temponaming resultsshowed clearly that sub-jects have somemechanismof responsetiming avail-able to them. Moreover, the mechanismthat controlsresponseinitiation seemsto be tightly coupledwith re-sponseexecution;thetaskwasto timeresponseinitiationwith the tempo(andfeedbackwasbasedon this alone),yet responsedurationsshortenedalongwith latencies.Inconjunctionwith amodelof word reading,onemight beableto formulateatimecriterionmechanismthatcanac-countfor thelatency anderrorresultsof thetemponam-ing experiments.However, a time criterionalonehasnointrinsic implicationsfor responsedurations,soit cannot

accountfor durationresults.Weproposeanalternateexplanationthat is motivated

by theevidencefor a couplingof responseinitiation andresponseexecution.Justasthetempoinducedcompres-sion of the responsetrajectory, we hypothesizethat italsoinducedcompressionof theprocessingtrajectoryinthewordreadingsystem.Strategic controlover responsetiming doesnot manipulatethe stoppageof the normalcourseof processing,it changesthecourseof processingitself suchthata responsecanbeinitiatedat thedesiredpoint in time. This propertyof acceleratedprocessingcan be instantiatedin a connectionistnetwork by ma-nipulating the input gain to processingunits (Cohen&Servan-Schreiber, 1992; Kello, Plaut, & MacWhinney,in press;Nowlan, 1988). Input gain is a multiplicativescalingfactoron the net input to processingunits (it isequivalent to the inverseof temperaturein BoltzmannmachinesAckley, Hinton, & Sejnowski, 1985). The ef-fectof gainonunit outputsdependsontheiroutputfunc-tions. Considerthe commonlogistic function in a pro-cessingunit, aj , that updatesits output in continuoustime,

a> t ?j % 1

1 * exp @ . x> t ?j γ A (1)

wherex> t ?j is the net input to unit aj at time t, and γis the input gain. For the logistic, input gain serves tosharpen(for largevaluesof γ) or flatten(for smallvaluesof γ) theeffect thata changein thenet input hason theoutput. For high valuesof gain, small changesin thenet input to a unit canbe sufficient to move the outputbetween0 and1. For low valuesof gain, largechangesin thenetinputarenecessaryto haveacomparableeffectonaunit’soutput.

Raisingtheinputgainonunitsin anetwork canaccel-erateprocessingbecausenetinputscanchangeunit out-putsin asmallernumberof timesteps.With anon-linearactivation function and interactionsamongprocessingunits,themodulationof gaincanreversetheasymptoticstatesof units. Dependingon how unitsstatesareinter-preted,thiscancorrespondto aqualitativechangein net-work behavior. Kello et al. (in press)have demonstrated

24 KELLO AND PLAUT

both of thesebasiceffects in an abstract,connectionistmodelof informationprocessing.

In additionto the couplingof responselatenciesanddurations,how might the manipulationof input gain ina connectionistmodel of word readingaccountfor theotherfindingsfrom thetemponamingexperiments?Thetwo mostrelevantresultsto considerare1) theattenua-tion of stimuluseffectson latencies,and2) thepatternoferrors.Thefirst fallsnaturallyoutof thefactthatacceler-atedprocessesarecompressedin time. Otherfactorsbe-ing equal,asoverall processingtimesareshortened,anydifferencesin processingtimesacrossconditionsshouldalsobeshortened.

With regardsto the error results,thereare two find-ings to accountfor: 1) the overall increasein error ratewith increasedtempos,and2) thecorrespondingincreasein the rateof all error typesexceptLARC errors. First,Kello etal. (in press)haveshown thathigh levelsof gaincancausea generalloss in accuracy. This canoccur ifincreasedgain amplifiesany noise in processing,or iftherelativetiming of unit outputtrajectoriesis disturbed.

With regardsto the secondfinding, let us considerthe rate of occurrenceof eacherror type asa functionof tempo. In all three experiments,the rate of word,nonword,andarticulatoryerrorsall increasedwith fastertempos,while the rateof LARC errors(in Experiments1 and2) remainedconstant.We assumethat input gainis bemanipulatedacrossall processingunits in a modelof word reading. This is a logical extensionof our hy-pothesisthatinternalprocessesarecoupledto motorpro-cesses.Given this, it is straightforward to explain theincreasein nonword andarticulatoryerrors. A primarysourceof articulatoryerrorsis likely to bewithin thepro-cessesthat map phonologicalrepresentationsto motorcommands.As gainis increasedwithin theseprocesses,the rate of articulatoryerrorsshould increase. Analo-gously, a primary sourceof nonword errorsis likely tobewithin phonologicalprocesses(e.g.,phonotacticcon-straints),so as gain is increasedwithin phonology, therateof nonword errorsshouldincrease.

By this logic, a primary sourceof word and LARCerrorswould be a dysfunctionin the mappingfrom or-thographyto phonology(althoughphonologicalandor-thographic/visualprocessesmay also contribute). Theabove-chanceincreasein word errors,coupledwith theconstantrate of LARC errors, suggeststhat increasedtemposcauseda proportionalamplificationof the influ-enceof lexical knowledgeversussub-lexical knowledge.How might an increasein the level of gain in a connec-tionist modelof word readingcausesuchan effect? Inthe triangleframework, lexical knowledgehasits influ-enceprimarily throughsemantics,whereassub-lexicalknowledgeis storedprimarily in theweightsbetweenor-thographyandphonology. Therefore,to accountfor theconstantrateof LARC errors,an increasein input gainwould have to emphasizethe contribution of semanticsover sub-lexical knowledge.

We hypothesizethat gain would have this effect dueto thearbitrarynatureof therelationshipbetweenseman-ticsandphonology, relativeto thesystematicrelationshipbetweenorthographyandphonology. Overall, networkshaveatendency to mapsimilar inputsto similaroutputs.This propertyfacilitatesthe mappingfrom orthographyto phonologydue to the systematicityin their relation-ship, andtherefore,the weightsbetweenthesetwo lev-elsof representationdo not have to grow very largedur-ing learning. By contrast,the non-systematicrelation-ship betweensemanticsand phonologyrequireslargerweightsto overcomethe biasfor similar inputs to pro-ducesimilar outputs. Input gain is multiplicative withrespectto weightmagnitude,but thecontributionsfromdifferentincomingweightsareadditivewith respectto areceiving unit’s net input. Therefore,anincreasein gainwill amplify theinfluenceof largerweightsover smallerweights. Given that weightsfrom semanticsto phonol-ogyarelargerthanthosefrom orthographyto phonology,onemightexpectanincreasein inputgainto amplify lex-ical knowledgeover sub-lexical knowledge.

It remainsto be seenwhether, in a full-scalemodelof word reading,the manipulationof input gain canbedemonstratedto have the propertiesarguedfor above.Preliminarysimulationshave beenencouraging(Kello& Plaut, 1998), and a full treatmentof theseissuesisthetopic of ongoingwork (Kello & Plaut,1999).

StimulusBlockingandInputGain

Part of the motivation for the temponamingexper-imentswas the proposalof a time criterion to accountfor someeffectsof stimulusblocking found by Lupkeret al. (1997)andJared(1997). We have arguedthat it isdifficult to accountfor the temponamingresultswith atimecriterion,andwehaveproposedthegainhypothesisasanalternative. To whatextentcanthegainhypothesisaccountfor stimulusblockingeffects?

As a mechanismto accountfor stimulusblockingef-fects, input gain is actually in the samespirit asa timecriterion.Bothaccountsaredistinguishedfrom therouteemphasishypothesisin thatthey eacharguefor strategiccontrol over responseinitiation ratherthanthe process-ing of individual routes. Furthermore,both hypothesesproposethata responsecriterion is shiftedasa functionof stimulusdifficulty; asstimuli in ablockbecomemoredifficult, on average,the responsecriterion will shift toa more conservative setting. As explainedabove, thekey differencebetweenthesetwo hypothesesis the ex-act mechanismunderlyingthe responsecriterion. Thetimecriterionhaltsthenormaltrajectoryof processingataparticularpoint in time,whereasinput gainacceleratesor deceleratesthetrajectoryof processing.

This differentiatesthetwo hypothesesin how predic-tionsaremadebasedonlatency datain stimulusblockingexperiments.Lupker et al. (1997)andJared(1997)usemeanlatenciesin the pureblocksconditionsasa directmeasureof therelativepositioningof thetime criterion,andthey usethismeasureto predictmeanlatenciesin the


mix� edblocks. For example,Lupker et al. (1997)foundthat themeanlatency of a pureblock of high frequencyconsistent(HFC) words was fasterthan that of a pureblock of pseudowords. They interpretthis differenceinmeanlatenciesasindicatingthat the time criterion wassetmoreconservatively in pseudowordblock thanin theHFC block. Thetime criterionhypothesisstipulatesthata middling time criterionshouldbesetin a mixedblockof bothpseudowordsandHFCwords.Therefore,thehy-pothesispredictsanincreasein HFC latenciesanda de-creasein pseudoword latenciesin the mixedblock; thisis whatthey found.

The gain hypothesiscan make the sameprediction,but one must interpret the pure block latenciesin thecontext of a theoryof themappingfrom orthographytophonology. This is becauselatenciesare a function ofgain in conjunctionwith stimulus type, and the effectof stimulustype on latenciesis determined(at least inpart) by one’s theoryof how orthographyis mappedtophonology. Strictly speaking,this is alsotrueof thetimecriterionhypothesis;asdiscussedin theIntroduction,la-tenciescannotbe a function of the time criterion alonebecausethis would predict no stimuluseffects. How-ever, thecontributionof thetime criterionto latenciesissomewhatindependentof themappingfrom orthographyto phonology, whereasgainexertsits influenceon laten-ciesvia this mapping.Thecurrentstudydoesnot makeclaimsabouthow orthographyis mappedto phonology(i.e., we are not presentinga theory of word readinghere),so we cannotquantitatively determinehow wellthe gain hypothesiscan accountfor the relevant stim-ulus blocking effects. However, qualitatively speaking,gainis verysimilar to a timecriterionasamechanismofstrategic control,soit will haveasimilar setof problemsandbenefits.

OtherUsesof InputGain

Input gainhasbeenusedasasa mechanismof strate-gic control in other studiesas well. In one line of re-search,gain hasbeenusedas a mechanismof controlover the influenceof contextual information on stimu-lus processing(Cohen& Servan-Schreiber, 1992). Inthat study, a connectionistmodel of Stroop phenom-enawaspresentedin which the role of two processingunits wasto provide taskinformation(context, i.e., oneunit representedthe color namingtask,the otherrepre-sentedtheword namingtask).Theinput gainof thetaskunits (mathematicallyequivalent to the gain parameterusedin the currentstudy)wasmanipulatedto simulatethe hypothesizedrole of the neurotransmitterdopaminein pre-frontalcortex. A large body of neurophysiolog-ical evidencehas indicatedthat dopaminemay modu-late the gain of postsynapticinput summationin PFC(aswell asotherareas;seeCohen& Servan-Schreiber,1992),andthey view thefunctionof thepre-frontalcor-tex asthemaintenanceof taskandsituationcontext. Fur-thermore,other studieshave shown that the regulationof dopamineis impaired in schizophrenics(Cohen&

Servan-Schreiber, 1992). Therefore,the modulationofgain in their modelserved to simulationnormalversusschizophrenicperformancein Stroopandrelatedtasks.

In astudymoresimilar to thecurrentone,Kello etal.(in press)conductedtwo Stroopcolor namingexperi-mentsto examinethetemporalrelationshipbetweencog-nitiveprocessingandovert articulation.They foundthatassubjectswerepressuredto respondfaster(by the in-troductionof a deadline),the effect of Stroopinterfer-encebled over from naminglatenciesto namingdura-tions.Justasin thecurrentstudy, inputgainwasinvokedasa mechanismto strategically control the speedof re-sponseinitiation. Theauthorspresentedanabstract,con-nectionistmodel of information processingwhich cap-turedthebasiceffectof thedeadlinein theirexperimentsby manipulatingtheinput gainto processingunitsin thenetwork. Relevantto thetemponamingexperiments,anincreasein gaincausedboth responselatenciesanddu-rationsto shortenin the network. The efficacy of inputgain asa mechanismof control over responsespeedinthestudyby Kello et al. (in press)is evidencethat inputgainmayaccountfor thetemporesults,andmaydevelopinto amoregeneraltheoryof cognitivecontrol.

ThecurrentstudyandtheKello et al. (in press)studyinvokedgainasa mechanismof controlover processingspeed,whereasthe CohenandServan-Schreiber(1992)studyinvoked gain to modulatethe influenceof certaininformation on stimulusprocessing. Theseare differ-entinterpretationsof thespecificrole thatgainplays,butthey bothtreatgainasamechanismof strategic control.

Conclusions

We introducedthe temponamingtask to investigatemechanismsof responsetiming, and to provide a newempiricalwindow into the time courseof phonologicalprocessingin word reading. The resultsof threetemponamingexperimentswereinterpretedasevidenceagainsttheuseof a time criterionasthemechanismof responsetiming in temponaming.With regardsto thetimecourseof phonologicalprocessing,fast temposcausedan in-creasein all error typesexceptLARC errors. This re-sult was interpretedas evidencethat the influenceofsub-lexical knowledgeonthemappingfrom orthographyto phonologyis reducedunderpressurefor speededre-sponding.We proposedthat input gain,asa mechanismof controloverprocessingspeedin thewordreadingsys-tem, can potentially accountfor the temponamingre-sults. Whetherinput gain canprovide a generaltheoryof strategic control over processingis a topic for futurestudies.

References

Ackley, D. H., Hinton, G. E., & Sejnowski, T. J. (1985). Alearningalgorithmfor BoltzmannMachines.CognitiveSci-ence, 9(2), 147-169.

Andrews, S. (1982). Phonologicalrecoding:Is the regularityeffectconsistent?MemoryandCognition, 10, 565-575.

26 KELLO AND PLAUT

Balota,B D. A., Boland,J. E., & Shields,L. W. (1989). Prim-ing in pronunciation:Beyondpatternrecognitionandonsetlatency. Journalof MemoryandLanguage, 28(1), 14-36.

Balota,D. A., & Chumbley, J. I. (1985). The locusof word-frequency effectsin thepronunciationtask: Lexical accessand/orproduction?Journal of MemoryandLanguage, 24,89-106.

Berndt,R. S.,Reggia,J.A., & Mitchum,C. C. (1987).Empiri-cally derivedprobabilitiesfor grapheme-to-phonemecorre-spondencesin English. BehaviorResearch Method,Instru-ments,& Computers, 19, 1-9.

Bock,K. (1996).Languageproduction:Methodsandmethod-ologies.PsychonomicBulletin andReview, 3, 395-421.

Cody, R. P., & Smith, J. K. (1997). Applied statisticsandthe SASprogramminglanguage. UpperSaddleRiver, NJ:Prentice-Hall.

Cohen,J.D., & Servan-Schreiber, D. (1992). Context, cortex,anddopamine:A connectionistapproachto behavior andbiology in schizophrenia.Psychological Review, 99(1), 45-77.

Colombo,L., & Tabossi,P. (1992).Strategiesandstressassign-ment:Evidencefrom a shallow orthography. In R. Frost&L. Katz (Eds.),Orthography, phonology, morphology, andmeaning(p. 319-399).Amsterdam:North-Holland.

Coltheart,M., Curtis,B., Atkins, P., & Haller, M. (1993).Mod-els of readingaloud: Dual-routeand parallel-distributed-processingapproaches.Psychological Review, 100(4), 589-608.

Dell, G. S. (1986). A spreading-activation theoryof retrievalin sentenceproduction.Psychological Review, 93(3), 283-321.

Dell, G. S. (1988).Theretrieval of phonologicalformsin pro-duction: Testsof predictionsfrom a connectionistmodel.Journal of MemoryandLanguage, 27, 124-142.

Dell, G. S.,& Reich,P. A. (1981). Stagesin sentenceproduc-tion: An analysisof speecherror data. Journal of VerbalLearningandVerbal Behaviour, 20, 611-629.

Forster, K. I., & Chambers,S. (1973).Lexical accessandnam-ing time. Journalof VerbalLearningandVerbalBehaviour,12, 627-635.

Forster, K. I., & Davis, C. (1991). Thedensityconstraintonform-primingin thenamingtask: Interferenceeffectsfroma masked prime. Journal of Memoryand Language, 30,1-25.

Frederiksen,J. R., & Kroll, J. F. (1976). Spellingandsound:Approachesto theinternallexicon.Journalof ExperimentalPsychology: HumanPerceptionand Performance, 2, 361-379.

Garrett,M. F. (1976).Syntacticprocessesin sentenceproduc-tion. In R. J.Wales& E. Walker (Eds.),New approachestolanguage mechanisms.Amsterdam:North-Holland.

Glushko, R. J. (1979). Theorganizationandactivation of or-thographicknowledgein readingaloud. Journal of Exper-imentalPsychology: HumanPerceptionand Performance,5(4), 674-691.

Harm, M. W. (1998). Division of labor in a computationalmodelof visual word recognition. Unpublisheddoctoraldissertation,Departmentof ComputerScience,Universityof SouthernCalifornia,Los Angeles,CA.

Herdman,C. M. (1992). Attentionalresourcedemandsof vi-sualwordrecognitionin namingandlexicaldecisions.Jour-

nal of ExperimentalPsychology: HumanPerceptionandPerformance, 18, 460-470.

Herdman, C. M., LeFevre, J.-A., & Greenham, S. L.(1996). Base-word frequency andpseudohomophonenam-ing. Quarterly Journal of ExperimentalPsychology: Hu-manExperimentalPsychology, 49A(4), 1044-1061.

Hess,D. J., Foss, D. J., & Carroll, P. (1995). Effects ofglobal and local context on lexical processingduring lan-guagecomprehension.Journal of ExperimentalPsychol-ogy: General, 124(1), 62-82.

Hoequist,C. E. (1983). The perceptualcenterand rhythmcategories.Language andSpeech, 26, 367-376.

Jared,D. (1997).Evidencethatstrategy effectsin wordnamingreflectchangesin outputtiming ratherthanchangesin pro-cessingroute.Journalof ExperimentalPsychology: Learn-ing, Memory, andCognition, 23, 1424-1438.

Jared,D., McRae,K., & Seidenberg, M. S. (1990). Thebasisof consistency effectsin word naming. Journal of MemoryandLanguage, 29, 687-715.

Kawamoto,A. (1988). Distributedrepresentationsof ambigu-ouswordsandtheir resolutionin a connectionistnetwork.In S. L. Small, G. W. Cottrell, & M. K. Tanenhaus(Eds.),Lexical ambiguityresolution:Perspectivesfrompsycholin-guistics,neuropsychology, and artificial intelligence. SanMateo,CA: MorganKaufmann.

Kawamoto,A. H., Kello, C., Jones,R., & Bame,K. (1998).Initial phonemeversuswholewordcriterionto initiatepro-nunciation:Evidencebasedon responselatency andinitialphonemeduration. Journal of ExperimentalPsychology:Learning, Memory, andCognition, 24, 862-885.

Kawamoto,A. H., Kello, C. T., Higareda,I., & Vu, J. (1999).Parallelprocessingandinitial phonemecriterionin namingwords: Evidencefrom frequency effectson onsetandrimeduration. Journal of ExperimentalPsychology: Learning,Memory, andCognition, 25(2), 362-381.

Kawamoto,A. H., & Kitzis, S. N. (1991). Time courseofregular andirregular pronunciations.ConnectionScience,3, 207-217.

Kello, C. T., & Kawamoto,A. H. (1998). Runword: An IBM-PCsoftwarepackagefor thecollectionandacousticanalysisof speedednamingresponses.BehaviorResearch Methods,InstrumentsandComputers, 30, 371-383.

Kello, C. T., & Plaut,D. C. (1998). Modulationof lexical ef-fectsin word readingwithoutstrategic controlof pathways:Evidencefrom temponaming. In Proceedingsof the 39thannualmeetingof thepsychonomicsociety. Dallas,TX.

Kello, C. T., & Plaut,D. C. (1999). Strategic control by gainmanipulationin a computationalmodelof word readinginthetemponamingtask.

Kello, C. T., Plaut,D. P., & MacWhinney, B. (in press).Thetask-dependenceof stagedversuscascadedprocessing:Anempiricalandcomputationalstudyof stroopinterferenceinspeech production.

Kucera,H., & Francis,W. N. (1967). Computationalanalysisof present-dayAmericanEnglish. Providence,RI: BrownUniversityPress.

Kurganskii,A. V. (1994). Quantitative analysisof periodicaltime-patternedfingertapping.HumanPhysiology, 20, 117-120.

Lemoine,H. E., Levy, B. A., & Hutchinson,A. (1993). In-creasingthenamingspeedof poorreaders:Representations


formedacrossrepetitions. Journal of ExperimentalChildPsychology, 55, 297-328.

Levelt, W. J.M. (1989).Speaking:Fromintentionto articula-tion. Cambridge,MA: MIT Press.

Lupker, S. J., Brown, P., & Colombo, L. (1997). Strate-gic control in a namingtask: Changingroutesor changingdeadlines?Journalof ExperimentalPsychology: Learning,Memory, andCognition, 23, 570-590.

Lupker, S.J.,Taylor, T. E., & Pexman,P. M. (1997).Strategiccontrol of a time criterion in naming: New evidenceandeffects. In Proceedingsof the 38th annualmeetingof thepsychonomicsociety. Philadelphia,PA.

Manis,F. R. (1985).Acquisitionof wordidentificationskills innormalanddisabledreaders.Journal of EducationalPsy-chology, 77, 78-90.

Marcus,S. M. (1981). Acousticdeterminantsof perceptualcenter(p-center)location. Perceptionand Psychophysics,52, 691-704.

Mates,J.,Radil,T., & Poppel,E. (1992).Cooperative tapping:Time control underdifferentfeedbackconditions. Percep-tion andPsychophysics, 52, 691-704.

McClelland, J. L., & Rumelhart,D. E. (1981). An interac-tive activationmodelof context effectsin letterperception:Part 1. An accountof basicfindings.Psychological Review,88(5), 375-407.

McRae,K., Jared,D., & Seidenberg, M. S. (1990).Ontherolesof frequency andlexical accessin wordnaming.JournalofMemoryandLanguage, 29, 43-65.

Meyer, D. E., Osman,A. M., Irwin, D. E., & Kounios, J.(1988).Thedynamicsof cognitionandaction:Mentalpro-cessesinferred from speed-accuracy decompostion. Psy-chological Review, 95, 183-237.

Monsell,S.,Patterson,K., Graham,A., Hughes,C. H., & Mil-roy, R. (1992).Lexical andsublexical translationof spellingto sound:Strategic anticipationof lexical status.JournalofExperimentalPsychology: Learning, Memory, and Cogni-tion, 18(3), 452-467.

Nowlan,S. J. (1988). Gainvariationin recurrenterrorpropa-gationnetworks. Complex Systems, 2, 305-320.

Ollman,R. T., & Billington, M. J. (1972).Thedeadlinemodelfor simple reactiontimes. Cognitive Psychology, 3, 311-336.

Orden,G. C. V. (1991). Phonologicalmediationis fundamen-tal in reading.In D. Besner& G. Humphreys (Eds.),Basicprocessesin reading: Visual word recognition (p. 77-103).Hillsdale,NJ: Erlbaum.

Paap,K. R., & Noel,R. W. (1991).Dual routemodelsof printto sound:Still a goodhorserace. Psychological Research,53, 13-24.

Pachella,R. G.,& Pew, R. W. (1968).Speed-accuracy tradeoffin reactiontime: Effect of discretecriteriontimes. Journalof ExperimentalPsychology, 76(1), 19-24.

Patterson,K., & Behrmann,M. (1997).Frequency andconsis-tency effectsin a puresurfacedyslexic patient. Journal ofExperimentalPsychology: HumanPerceptionand Perfor-mance, 23(4), 1217-1231.

Plaut,D. C.,McClelland,J.L., Seidenberg, M. S.,& Patterson,K. (1996).Understandingnormalandimpairedword read-ing: Computationalprinciples in quasi-regular domains.Psychological Review, 103, 56-115.

Rastle,K., & Coltheart,M. (1999).Serialandstrategic effectsin readingaloud.Journalof ExperimentalPsychology: Hu-manPerceptionandPerformance, 25(2), 482-503.

Rumelhart,D. E., & McClelland, J. L. (1982). An interac-tive activationmodelof context effectsin letterperception:Part 2. The contextual enhancementeffect andsometestsandextensionsof themodel.Psychological Review, 89, 60-94.

Ruthruff, E. (1996). A testof the deadlinemodel for speed-accuracy tradeoffs. PerceptionandPsychophysics, 58, 56-64.

Savage,G. R., Bradley, D. C., & Forster, K. I. (1990). Wordfrequency andthe pronunciationtask: The contribution ofarticulatoryfluency. Language andCognitiveProcesses, 5,203-236.

Schwartz, R. G., & Goffman, L. (1995). Metrical patternsof wordsandproductionaccuracy. Journal of Speech andHearingResearch, 96, 523-568.

Seidenberg, M. S.,& McClelland,J.L. (1989). A distributed,developmentalmodelof wordrecognitionandnaming.Psy-chological Review, 96, 523-568.

Seidenberg, M. S., Plaut,D. C., Petersen,A. S., McClelland,J. L., & McRae,K. (1994). Nonword pronunciationandmodelsof word recognition.Journal of ExperimentalPsy-chology: HumanPerceptionandPerformance, 20(6), 1177-1196.

Seidenberg, M. S.,Waters,G. S.,Barnes,M. A., & Tanenhaus,M. K. (1984). Whendoesirregularspellingor pronuncia-tion influenceword recognition? Journal of Verbal Learn-ing andVerbal Behaviour, 23, 383-404.

Sherak,R. (1982). A real-timesoftwarevoicekey andanap-plication.BehaviorResearch MethodsandInstrumentation,14(2), 124-127.

Stanovich, K. E., & Bauer, D. W. (1978). Experimentsonthespelling-to-soundregularity effect in word recognition.MemoryandCognition, 6, 410-415.

Stanovich, K. E., & Pachella,R. G. (1976). The effect ofstimulusprobability on the speedandaccuracy of namingalphanumericstimuli. Bulletin of thePsychonomicSociety,8, 281-284.

Stanovich, K. E., Siegel, L. S., & Gottardo,A. (1997). Con-verging evidencefor phonologicalandsurfacesubtypesofreadingdisability. Journal of EducationalPsychology, 89,114-127.

Sternberg, S.,Wright,C.E.,Knoll, R.L., & Monsell,S.(1980).Motor programsin rapid speech:Additional evidence. InR.A. Cole(Ed.),Perceptionandproductionof fluentspeech(p. 507-534).Hillsdale,NJ: Erlbaum.

Strain,E., Patterson,K., Graham,N., & Hodges,J.R. (1998).Word readingin alzheimer’s disease:Cross-sectionalandlongitudinalanalysesof responsetime andaccuracy data.Neuropsychologia, 36, 155-171.

Strayer, D. L., & Kramer, A. F. (1994). Strategiesandauto-maticity: II. Dynamicaspectsof strategy adjustment.Jour-nal of ExperimentalPsychology: Learning, Memory, andCognition, 20, 342-365.

Tabossi,P., & Laghi, L. (1992). Semanticpriming in thepro-nunciationof wordsin two writing systems:ItalianandEn-glish. MemoryandCognition, 20(3), 303-313.

Taraban,R., & McClelland,J. L. (1987). Conspiracy effectsin wordrecognition.Journalof MemoryandLanguage, 26,608-631.

28 KELLO AND PLAUT

Treisman,A., & Williams, T. C. (1991). A theoryof crite-ria settingwith an applicationto sequentialdependencies.Psychological Review, 91, 68-111.

Trueswell,J.C., Tanenhaus,M. K., & Kello, C. (1993). Verb-specificconstraintsin sentenceprocessing:Separatingef-fects of lexical preferencefrom garden-paths.Journal ofExperimentalPsychology: Learning, Memory, and Cogni-tion, 19, 528-553.

VanOrden,G. C.,& Goldinger, S.D. (1994).Interdependenceof form andfunction in cognitive systemsexplainspercep-tion of printedwords.Journalof ExperimentalPsychology:HumanPerceptionandPerformance, 20(6), 1269.

Venezky, R. L. (1970). Thestructure of Englishorthography.TheHague:Mouton.

Waters,G. S.,& Seidenberg, M. S. (1985).Spelling-soundef-fectsin reading:Timecourseanddecisioncriteria.MemoryandCognition, 13, 557-572.

Wickelgren,W. (1977).Speed-accuracy tradeoff andinforma-tion processingdynamics.ActaPsychologica, 41, 67-85.

Yamada,N. (1995). Natureof variability in rhythmicalmove-ment.HumanMovementScience, 14, 371-384.

Zorzi,M., Houghton,G.,& Butterworth,B. (1998).Two routesor one in readingaloud? A connectionist“dual-process”model. Journal of ExperimentalPsychology: HumanPer-ceptionandPerformance, 24, 1131-1161.


Table6 AppendixA – StimulifromExperiments1 and2StandardNaming TempoNaming

HFE LFE LFC Pseudo HFE LFE LFC Pseudobreak breast brunt brish blood blown bloke blispboth bush bum bax both bush bum biftfriend frost fret frope break breast brunt brotefull flange flask flade broad plaid plod plakegone guise grope gret come caste cask cethave hearth hark hoam cost comb coal cipelow limb lisp lask dead dough dole destput pear pest pag death deaf desk detchsource suede swish swench done drought dank ditwant wand wit weke door doll dolt doanwhat wad wax wub foot flood flake flainwhere wool wade wibe four flown float flillword womb wane wuff friend frost fret fraskare ere eke ean full flange flask floddog cache crag crelp give glove glum glanehis swath swoon swunt gone guise grope grolehour rouge rune roon good gong gob gadenone mule mend mest great grown groan grigonce wharf whelp whark gross ghoul goon gaxthere chic shrub shrane have hearth hark haitwarm wolf wilt wune head hind hitch hamewatch rouse roam rilt heard hook hump hinkwhom vise vibe vit key cough cuff coomwhose farce fluff flisp learn leapt letch leskworth swamp stench stend lost loath lobe lokeyour wan wean wum love lose loom luff

low limb lisp lummonth mourn munch meepmost mould mole mellmove mow moan mickpoor pint pine pankput pear pest pitesaid swap swig swasksays spook spurt spoleshall shoe shame shumpshow shove shell shunchsome sew sole surtson steak stain stinesource suede swish swobestood suave swill swopethrough threat thrift thrishtouch tomb tote tunttruth trough trite tritchtwo tread trait troatwant wand wit warkwar wart wink woalwere worm wick woanwhat wad wax wobwhere wool wade woleword womb wane woltwork wasp wipe woonwould warp weep wum

Note: Pseudowords only appearedin Experiment2. “Chic” was removed from the analysesdue to an inordinatenumberof errors.

30 KELLO AND PLAUT

Table7 AppendixB – StimulifromExperiment3

Std. Tempo Std. TempoHFE LFE LFC HFE LFE LFC

swant bloor blad blift losh mave mough murtwug bood bost bope loup mork pleaf pakeblan breat broul brole mape pead pove pankblig brone cearth cark meeb pone pown shainboog cays cong cobe morp pood sarp shigbrear diend cood dest murp sall shart spoancarg dord deak disp pleam seard shint stellcowe dource deast dit poot shouch sook suntdait dut duave flane sak sood sourn swaxdier fost flaste fletch shap sove spown swolediz foth flomb fline shink stost swand taitdorg fough flove fritch soop sull thromb throkedup frome frap gite soor throot torm trobflet gaid geat glick spow tove trind troteflote gead ghough goom sunth trour trould wamefrak goor glead grum swode twey woll waskgabe grat goath hade throob wearn wook woatgeave grive grear het tring woad wose wolegleap har haid hink trouch wome wough woongrap heak heapt hod wape wonth wought wumgream huth huede lask woap woss wown wumpgrud ko lange lish wom wost wush wunchheab leath lew lolt woophib lon lomb luff woskhuf lould masp meep woupleab mant mool mipe wunt

strategic control in word reading: evidence from speeded...

Documents