In J. Reggia,R. Berndt,& E. Ruppin(Eds.),Neural modeling of cognitive andbrain disorders (pp.177–199).New York: World Scientific,1996.

CONNECTIONS AND DISCONNECTIONS:A CONNECTIONIST ACCOUNT OF SURFACE DYSLEXIA

KARALYN PATTERSONMedical Research Council Applied Psychology Unit

15 Chaucer Road, Cambridge CB2 2EF, UK, karalyn@mrc-apu.cam.ac.uk

DAVID C. PLAUTJAMES L. MCCLELLAND

Department of Psychology, Carnegie Mellon University,and the Center for the Neural Basis of Cognition, Pittsburgh, USA

MARK S.SEIDENBERGNeuroscience Program, University of Southern California, Los Angeles, USA

MARLENE BEHRMANNDepartment of Psychology, Carnegie Mellon University, Pittsburgh, USA

JOHNR. HODGESDepartment of Neurology, University of Cambridge Clinical School, Cambridge, UK

Theacquiredreadingdisorderof surfacedyslexia, in which lower-frequency wordswith atypi-calspelling-sound correspondences(e.g.,PINT) becomehighly vulnerableto error, is presentedin a framework basedon interactionbetweendistributed representationsin a triangle of or-thographic,phonological,andsemanticdomains.The framework suggeststhat low-frequencyexceptionwordsareratherinefficiently processedin termsof orthographic-phonological con-straints,becausethesewordsareneithersufficiently commonto have muchimpacton learningin the networknor sufficiently consistentwith thepronunciations of their orthographicneigh-bors to benefitfrom sharedstructure. For thesewords, then,the interactionbetweenphono-logical and semanticrepresentationsmay be especiallyimportantfor settling on the correctpronunciation. It is thereforeviewed asno coincidentalassociationthat all reportedpatientswith markedsurfacedyslexia have alsobeenprofoundlyanomic,suggestingreducedsemantic-phonologicalactivation.Thechaptersummarizesthesimulationof surfacedyslexia in thecom-putationalmodelof readingdevelopedby Plaut,McClelland,Seidenberg, andPatterson(1996),andpresentsnew datafrom threesurfacealexic patients.Thegradedconsistency effectsin thepatients’readingperformancearemorecompatiblewith the distributed connectionistframe-work thanwith dual-routemodelsmaintainingastrictdichotomybetweenregularandexceptionwords.

1 Background

Oneof the main issuesin readingresearchconcernsthe proceduresthat readersuseto computethe pronunciationof written words. According to the dominantview,two differenttypesof letterstringin Englishdemandtwo qualitatively differentpro-cedures:the ability of readersto pronounceletter stringsthat they have never seen

177

before(for example,a nonsenseword suchasMAVE) requiresa procedurebasedonrulesfor translatinggraphemesto phonemes;andtheability to pronouncefamiliar“ir -regular” wordsthatbreakstandardgrapheme-phonemecorrespondencerules(suchasHAVE) requiresa procedurein which theword’s orthographiclexical entryactivatesits whole-word pronunciation.Considerableevidence—fromnormaladult readers,from adultswith acquireddisordersof reading,andfrom childrenwith developmen-tal dyslexia—hasbeeninterpretedassupportfor this dual-routemodel,mainly byColtheartandhis colleagues(Castles& Coltheart,1993;Coltheart,1985;Coltheart,Curtis, Atkins & Haller, 1993; Coltheart& Rastle,1994)but by many other read-ing researchersaswell (e.g.,Baluch& Besner, 1991;Funnell,1983;Paap& Noel,1991). Dual-routetheory, instantiatedin computationalmodelsbothby Coltheartetal. (1993)andby Reggia,BerndtandD’Autrechy (1994),is sufficiently well knownandaccessiblein theliteratureto obviatetheneedfor a full descriptionhere.Simplyput, its basicpremiseis thatnomodelof readingwill succeedin explainingtheknowndataunlessit incorporatesseparatelexical andsub-lexical mechanismsfor translatinganorthographicstringinto apronunciation(hereinafterO � P).

Despitetheprominenceandsuccessof dual-routetheory, it hashadits criticsandalternatives. For example,a numberof readingresearchershave arguedthat regularandexceptionwordscorrespondto pointson a consistency continuumratherthanadichotomy(Glushko,1979;Seidenberg,Waters,Barnes& Tanenhaus,1984;Shallice,Warrington& McCarthy, 1983). Othershave claimedthatnonword reading,insteadof requiringa separaterule-based“non-lexical” system,could be accomplishedbyextractingandpooling knowledgefrom lexical representationsfor structurallysimi-lar words(Henderson,1982;Humphreys& Evett,1985;Kay & Marcel,1981).In thelastdecade,theproposalthatasingleO � Pmechanismis in factcapableof capturingboth thegeneralizationsandtheexceptionsin spelling-soundrelationshipshasbeendevelopedin computationalmodelsof readingaloud,first by Sejnowski andRosen-berg (1987)andsubsequentlyby Seidenberg andMcClelland(1989).TheSeidenbergand McClelland model, asa major theoreticalstatementaboutthe acquisitionandskilled performanceof single-word reading,attractedconsiderableattention,muchpositive but somecritical. For example,dual-routetheorists(Besner, Twilley, Mc-Cann& Seergobin,1990;Coltheartet al., 1993)contestedtheclaim thatSeidenbergandMcClellandhaddemonstratedthe adequacy of a singlemechanismfor readingboth exceptionwordsand nonwords, becausethe original simulationachieved no-tably lesssuccessthanmosthumanreadersdo in generalizingits knowledgeto thepronunciationof nonwords.

In themostrecentphaseof thisdebate,Plaut,McClelland,SeidenbergandPatter-son(1996)presentedfour new simulationsof theO � P computationin English. Asoneprincipal developmenton their predecessor(Seidenberg & McClelland,1989),thenetworksin thePlautetal.modelemployedorthographicandphonologicalrepre-

178

sentationsdesignedto capturemoresuccessfullythesimilaritiesin orthographicandphonologicalspace.This new designof representationsenabledthemodelto attainaccuracy in nonword readingwell within therangeof realadultreaders.Threeof thenetworkshada feedforwardarchitecture;thefourth wasanattractornetwork involv-ing interactivity amongthephonologicaloutputunitsandbetweenthephonologicalunits and the hidden-unitlayer. Two of the simulationswere trainedusing actual(Kucera& Francis,1967)frequenciesof the3000wordsin thecorpus,while theoth-ershadsomedegreeof frequency compression,eithermoreor lesssevere(logarith-mic andsquare-root,respectively). Detailedanalysesof theresultsfrom thesevarioussimulationscanbe found in the original article; for presentpurposes,the importantsummaryis the following. Given(a) orthographicandphonologicalrepresentationsthat effectively capturespelling-soundconsistencies,makingthe network appropri-atelysensitiveto therangeof consistenciesin thetrainingvocabulary, and(b) a train-ing regimebasedon realor approximateword frequencies,makingtheperformanceof the network appropriatelysensitive to the impactof word frequency, a networkwith a single O � P procedurecan reproducethe patternof accuracy andresponsetimesin namingregularwords,exceptionwords,andnonwordsthat is characteristicof realadultreaders.

Theseresultsfrom thefully trainednetworksof Plautet al. (1996)establishthata singlemechanism,in additionto learningto pronouncebothfamiliar regularwordsand exceptions,cangeneralizeto novel words. Criticismsof Seidenberg andMc-Clelland’s (1989)modelof reading,however, did not focusexclusively on its normalreadingperformance.Early attemptsto simulateone prominentform of acquiredreadingdisorderby damagingtheSeidenberg andMcClellandnetwork hadbeenac-knowledgedevenby its authorsasprovocativebut insufficient(Patterson,Seidenberg& McClelland, 1989); and Coltheartet al. (1993) arguedthat separatelexical andnon-lexical routesareessentialto accountnot only for the correctpronunciationofboth nonwordsandexceptionwordsby normal readersbut also for the patternsofperformanceobservedin neurologicallyacquireddisordersof word naming.Anothermajorcomponentof thework byPlautetal. (1996),therefore,wasaddressedto theis-sueof whetherandhow suchdisorders,in particularacquiredsurfacedyslexia, mightfind an explanationin a model that dispenseswith separatelexical andnon-lexicalprocedures.

2 Surface Dyslexia

Surfacedyslexia is one of the main forms of readingdisorderobserved when thepreviously competentreadingability of an adult is disruptedby brain injury or dis-ease(Shallice& Warrington,1980). This disorderwasgiven its nameby MarshallandNewcombe(1973)to convey the ideathat, whensurfacedyslexic patientsread

179

a word aloud incorrectly, their errorstypically reflect the “surface” structureof theword; the syndromewascontrastedwith deepdyslexia, in which errorswerecon-struedasreflectingtheword’s “deep” structure.For thewritten word PINT, the typ-ical surfacedyslexic’s errorwould be /pInt/ (i.e., pronouncedlike MINT andindeedevery word with the spellingpattern-INT in English, except for PINT), whereasadeepdyslexic patientmisreadingpint would belikely to respond“quart” or possibly“beer” (at leastin Britain, wherebeerin pubsis still served in pints). In describingthetypical readingerrorsof surfacedyslexia—oftencalledregularization errors,be-causethe irregular word pint is pronouncedlike its regular neighborsMINT, LINT,PRINT, etc.—MarshallandNewcombe(1973)hadidentifiedoneof themostsalientcharacteristicsof thedisorder. Furtherresearch(Behrmann& Bub,1992;Bub,Can-celliere& Kertesz,1985;McCarthy& Warrington,1986;andShallice,Warrington& McCarthy, 1983)establishedthat, in its purestform, acquiredsurfacedyslexia ischaracterizedby readingperformanceon regularwordsandnonwordsthat is withinnormal limits of both accuracy and speed,and a deficit on irregular words that isstronglymodulatedby word frequency.

Theaccountof surfacedyslexia in Coltheart’sDRC(Dual-RouteCascaded)modelis asfollows: thenon-lexical grapheme-phonemeroute,whichcancorrectlycomputeO � P for regular wordsandnonwords, is intact; the lexical route,which is neces-saryfor correctpronunciationsof exceptionwords,is damagedin a mannerthatstillenablessuccesson a high-frequency exceptionword like HAVE but fails on a lesscommonword like PINT, forcing the patientto respondwith the non-lexical route’soutput for this word, i.e., the regularizationerror /pInt/. As with the mainstreamdual-routeaccountof normal reading,this interpretationhasbeenthe leading,butnot quitetheonly, bid onthesurfacedyslexic table.Accordingto MarshallandNew-combe(1973,1980)—and,with minorvariations,to Hillis andCaramazza(1991)andHoward andFranklin (1988)—anadequatetheoryof O � P requirestwo routes,butnot preciselythe sametwo asproposedby Coltheart(1985;Coltheartet al., 1993).In theseconceptions,thelexical routeis a lexical semanticprocedure;thusa writtenword canbetranslatedto a phonologicalcodeeitherby sub-lexical correspondencesor by activationof the word’s meaningfollowedby the processesnormallyused,inobjectnamingandspontaneousspeech,to activatephonologyfrom meaning.Surfacedyslexic readingis thoughtto arisefrom a combinationof intact sub-lexical proce-dureanddamagedlexical-semanticroute. The accountof surfacedyslexia offeredby Plautet al. (1996)differs somewhat from all of the above; aswill be seenin amoment,however, in onecrucialrespectit is moreakin to thesealternativedual-routeproposalsthanto Coltheart’sview.

Beforeweexplainourposition,weshouldsaythatit is still evolving. Wepresentour somewhat preliminaryaccountherein the following spirit: McClelland(in hisfinal discussionat themeetingthatengenderedthis book)emphasizedthat,although

180

Orthography Phonology

MAKE�

/mAk/�

Context

Meaning

Figure1: The “triangle” framework for single-wordprocessingfrom Seidenberg andMcClelland(1989)andPlaut,McClelland,Seidenberg andPatterson(1996).

all currentmodelsareboundto bewrong,they mayneverthelessbegropingtheirwaytowardssomeimportant,andevencorrect,underlyingprinciples.

3 An Account of Surface Dyslexia in the Plaut et al. (1996)Framework

In thefinal empiricalsectionof theirpaper, Plautetal. turnedtheirattentionto surfacedyslexia. While stoppingshortof agreeingwith Coltheart’s claim that damageto asingleprocedurenevercould,in principle,accountfor surfacedyslexia, they acknowl-edgedthat the dramaticpatternof puresurfacedyslexia (i.e., normalreadingaloudof regularwordsandnonwordscoupledwith a severe,frequency-sensitivedeficit onexceptionwords)seemsunlikely to arisefrom damageto the kind of single,directO � P computationdevelopedthusfar. “Lesions” to the network sufficiently severeto reproducethe appropriatedegreeof impairmenton exceptionwordsalsodisruptthemodel’s performanceon regularandnonwords. ThePlautet al. model,however,like theSeidenberg andMcClellandmodelbeforeit, andalsoliketheviewsof single-word processingproposedby Bullinaria (in press),KawamotoandZemblige(1992),andVanOrdenandGoldinger(1994),hasthebroader“triangle” framework sketchedin Figure1. Therefore,despitethedemonstrationthata fully trainedmodelof directO � Ptranslationis capableof learningto producecorrectpronunciationsfor all sortsof letterstrings,it maybethatthetypicalhumanreaderdoesnotsolvetheO � Pprob-lem in preciselythis way. In particular, thepossibility remainsthatsomeaspectsof

181

skilled written word pronunciationmight rely on accessto semanticrepresentationsof words(S). This influencecould occuron the basisof eitherO � S� P activationor O � P� Sactivationandinteraction;in eithercase,theimportantcomponentis theS� P link. Theaccountdevelopedin Plautet al. is thatcommunicationfrom mean-ing to phonologyis particularlycritical for processingof wordsthatareonly weaklylearnedby directO� P, namelylow-frequency wordswith atypicalspelling-to-soundcorrespondences,andthat puresurfacedyslexia is thusattributableto reducedacti-vation from S� P. This is not a new idea: althoughthe natureof the direct O� Ptranslationin Plautet al. differs from that in other theoriesof the readingprocess,recall thatHillis andCaramazza(1991),HowardandFranklin (1988),andMarshallandNewcombe(1973,1980)all implicatedwordmeaningin theiraccountsof surfacedyslexia.

Beforepresentingthe simulationwork from Plautet al. (1996)designedto ex-plorethishypothesis,weshallbriefly summarizeexistingevidenceabouthumanread-ing performance,both normalandabnormal,which supportsthe ideathat semanticrepresentationsof wordsshouldbeconsideredgermaneto theprocessof translatingprint to pronunciation.

(1) In the usual context for which readingskills are mobilized—text reading—few peoplewould doubtthat the pronunciationof written wordsmustbeopentosemanticinfluences.For example,readerscorrectlypronounceheterophonichomo-graphs(suchasWIND, LEAD andBASS) whenthey arereadingtext aloud.Assumingthat thehumanO � P directcomputationhaslearnedto activateboth legitimatepro-nunciationsof suchwords(like the networks in Plautet al., which weretrainedonboth), the pronunciationappropriateto a particularcontext (e.g.,of thenounWIND,thatblows,or theverbWIND, thatoneusedto doto watches)canbeselectedvia S� Pactivation.

(2) In a studyof accuracy andresponsetimes(RTs) for single-word namingbynormaladult readers,Strain,PattersonandSeidenberg (1995)manipulateda seman-tic variable,imageability, in their selectionof stimuluswords. Hypothesizingthatasignificantimpactof this variableon word namingshouldbe observablemainly forwordsratherweaklysupportedby thedirectO � P procedure,they alsoincludedthevariablesof word frequency andregularity. As predicted,a disadvantagein bothac-curacy andRT for wordswith low-imageabilityratingswasobtainedprimarily forlower-frequency wordswith atypicalspelling-soundcorrespondences.This doesnotof coursemeanthat, for wordsthat areeithercommonlyencounteredor that fit themostcommonspelling-soundpatterns,thereis noautomaticactivationof wordmean-ing; for suchwords,however, O� Pactivationon its own maybesufficientto achieverapidandstablephonologicalrepresentations.This computationis lesseffective forlow-frequency inconsistentwords; and then communicationfrom S� P (which isstrongerfor imageablewordswith richer semanticrepresentations)may detectably

182

assistin settlingonapatternof phonologicalactivation.

(3) If surfacedyslexia is attributableto somedisruptionin communicationbe-tweensemanticandphonologicalrepresentations,thenpatientswith acquiredsurfacedyslexia shouldbe anomic,sincenamingof objectsor conceptsrelieson the S� Plink. As far aswe know, thereareno exceptionsto this association;that is, all pub-lishedcasesof patientswith acquired“pure” surfacereadinghave alsohada promi-nentanomia.

(4) It appearsthat the associationbetweensurfacedyslexia and anomiamaysometimeseven respectcategory specificity: DRB (Franklin, Howard & Patterson,1995),who wasonly measurablyanomicfor abstractwordsandconcepts,wasalsosignificantlyimpairedin readingexceptionwordsonly if they hadabstractmeanings.

(5) Theentailmentor predictionin this accountis only for a disruptionin com-municationfrom S� P. It shouldnotmatter, in principle,whetherthisdifficulty arisesfrom degradedsemanticrepresentationsor from a reducedcapacityfor activatingphonologyfrom meaning. Although the majority of reportedcasesof puresurfacedyslexia have hada profoundimpairmentto semanticmemoryperse(e.g.,Breedin,Saffran & Coslett,1994;Bub et al., 1985;Funnell, in press;McCarthy& Warring-ton, 1986;Parkin, 1993;Patterson& Hodges,1992),therearealsosurfacedyslexicpatientswhosedeficits in namingandreadinghave beenassignedto the S� P link(Graham,Patterson& Hodges,1995;Watt,Jokel & Behrmann,in press).

(6) In studiesof patientswhosesurfacedyslexia is attributedto degradedseman-tic memory, threeseparategroupsof investigatorshave reportedasignificantconcor-dancebetweenreadingaloudandcomprehension:that is, with irregular words,theitemsthatarenamedcorrectly(asopposedto regularized)tendsignificantlyto bethesameonesonwhich thepatientsucceedsin acomprehensiontestsuchmatchingspo-ken wordsto pictures(Funnell,in press;Graham,Hodges& Patterson,1994;Hillis& Caramazza,1991).

Finally, it is importantto notethatsurfacealexia maysometimesbemaskedbyadditionaldeficits,especiallyat the level of speechproduction. Oneperspicaciousquestionaboutour predictedassociationbetweenimpairedword comprehensionandsurfacedyslexia hasbeen:shouldonenot thenfind a commonassociationbetweensurfacedyslexiaandWernicke’saphasia,alanguagedisorderstandardlyinterpretedasacomprehensiondeficit?Ourfirst responseto thisquestionis thatspeechproductionin Wernicke’s aphasiais often so disturbedanddistortedby phonologicalproblemsthatit mightbehardto observeanadvantagein oralreadingfor regularoverexceptionwords. Our secondresponseis that at leastoneWernicke’s aphasicpatientwithouta profoundphonologicaldeficit hasshown thefrequency-by-regularity interactioninreadingaccuracy thatischaracteristicof surfacealexia (Behrmann,unpublisheddata).

183

0�

200�

400�

600�

800�

1000�

1200�

1400�

1600�

1800�

2000�

Training Epoch�0.0

1.0

2.0

3.0

4.0

5.0E

xter

nal I

nput

to P

hone

mes

�

Contribution of Semantic Pathway

High Frequency (1222/million)�Low Frequency (20/million)

�

Figure2: An illustration of the gradualadditionof notionalsemanticinput to the phonemeunitsduringO P training in the Plautet al (1996)simulationof surfacedyslexia; shown herefor the Tarabanand

McClelland(1987)high-andlow-frequency words.

4 Simulation of Surface Dyslexia by Plaut et al. (1996)

Theinstantiationby Plautetal. (1996)of thewayin whichwordmeaningmight influ-enceO � Ptranslationdid notattemptto provideagenuinerepresentationof meaning;rather, thecontributionof meaningwasapproximatedby providing anextrasourceofinput to thephonemeunits to pushthemtoward their correctactivations.Thesimu-lation useda feedforwardnetwork with square-rootcompressionof word frequency,plus a small weightdecayfactor; this biasesthe network to keepweightssmall andhastheeffectof preventingoverlearning.Themajorchangefrom theprevioussimu-lationsperformedwithoutany semanticcomponentwasthat,over thecourseof O� Ptraining, the additionalsourceof input to phonemeunits (notional semantics)wasgraduallyintroduced.The basisfor this gradedprocedurewasthe assumptionthat,becauseO � P mappingsin analphabeticorthography(evena quasi-regularonelikeEnglish) are much more systematicthan O � S mappings,beginning readerslearnthe former morerapidly thanthe latter. Furthermore,a larger additionalinput wasprovided for high- thanfor low-frequency wordson the assumptionthat real O� Slearningis strongerfor frequentlyencounteredwords. The gradualandfrequency-modulatedadditionalinput to phonemeunits is illustratedin Figure2, for a subsetof the 2998monosyllabicwords in the training corpus,namelythe high- andlow-frequency wordsusedin experimentsby TarabanandMcClelland(1987).

Performanceduringtrainingis illustratedin Figure3 for thesetsof TarabanandMcClelland(1987)wordsandGlushko (1979)nonwords.At earlystagesof training(epochs50–100),adequacy of word pronunciations(percentcorrect,i.e.,whetherthe

184

100 200 300 2000Training Epoch�

0

10

20

30

40

50

60

70

80

90

100P

erce

nt C

orre

ct

Training with Semantics�

HF Reg LF Reg

�HF Exc

LF Exc

�Nonwords

�

Figure 3: Performanceof the Plaut et al (1996) O P network trainedwith additional input (notionalsemantics)to the phonemeunits, at variousstagesof training, on the Tarabanand McClelland (1987)

high-andlow-frequency regularandexceptionwordsandtheGlushko(1979)nonwords.

correctphonemeunit at eachsegmentof the monosyllable—onset,vowel, coda—isthe mostactive unit) clearly variesasa functionof both frequency andconsistency;andperformanceon itemswhich benefitfrom neitherfrequency nor consistency (LFExc words in Figure 3) is still not perfectat epoch150. By epoch200, all wordpronunciationsare correct,and nonwords (which receive no additional“semantic”input) have achieveda level of 95% correctpronunciations.Although this networkyields frequency andregularity effectscharacteristicof networkstrainedwithout no-tional semantics,thenetwith additionalinputnaturallylearnsfasterthansimulationswith only O � Pinput. A nettrainedwith theidenticallearningparametersandinitialrandomweights,but with no semanticcontribution, reachedasymptoteon percentcorrectat aroundepoch500ratherthan200.

From epoch200 to 2000, when responsesto the different word setscan nolongerbe distinguishedby the percentcorrectmeasure(seeFigure3), a significantfrequency-by-consistency interactionin the model’s performanceis still observablein thecross-entropy errorscore,a measureof thediscrepancy betweenthepatternofactivationover thephonemeunitsgeneratedby thenetwork andtheprecisetargetpat-ternof activation.Ontheassumptionthatoutputpatternsapproximatingmorecloselyto “perfect” will supportmorerapidresponses,thiscontinuingsensitivity to frequencyandconsistency in themodel’s fully trainedperformancecanbeseenasananalogueof thehumanskilledreader’sresponsetimesto namehigh-andlow-frequency regular

185

1.6� 1.4� 1.2� 1.0� 0.8� 0.6� 0.4� 0.2� 0.0�Strength of Semantics (Epoch 2000)�

0

10

20

30

40

50

60

70

80

90

100P

erce

nt C

orre

ct

HF Reg�LF Reg

�HF Exc

�LF Exc

�Regularizations

�Nonwords

�

Figure 4: Performanceof the Plaut et al (1996) O P network trainedwith additional input (notionalsemantics)to thephonemeunits for 2000epochs,whenthestrengthof thesemanticinput is graduallyre-duced;resultsareshown for percentcorrecton theTarabanandMcClelland(1987)wordsandtheGlushko

(1979)words,andfor percentageof errorsto Excwordsthatareregularizations.

andexceptionwords.The resultsparticularlygermaneto modelingsurfacedyslexia arepresentedin

Figure4, which shows the performanceof the network that hadbeentrainedwithnotionalsemanticsfor 2000epochs,whenthe strengthof this additionalsourceofactivation to the output units is graduallyreduced. Although surfacedyslexia canresult from an abruptbrain insult like a cerebrovascularaccidentor a headinjury,thegreatmajority of reportedcasesof puresurfacedyslexia have sufferedfrom pro-gressivebraindisease:eithersemanticdementia(e.g.,Hodges,Graham& Patterson,1995),in whichtheprimaryimpairmentis apparentlyin conceptualknowledgeitself,or aprogressiveaphasiawhichseemsto affectmainly thelink betweensemanticandphonologicalrepresentations(e.g.,Grahamet al., 1995;Wattet al., in press).In bothcases,weassumethatthereisagradualreductionin theactivationof phonologybyse-mantics,andthesimulationwasintendedto mimic thisphenomenonby post-trainingwithdrawal of the additionalinput to the phonemeunits, leaving only orthographicinput to phonology.

Although Figure4 representspost-trainingperformance,what it revealsis theeffect of thenotionalsemanticinput during training. Becausethis secondsourceofinput pushesthe activationof eachoutputunit in the appropriatedirection—upif itshouldbe on for the target word, down if it shouldbe off—the gradualincreaseinsemanticinput in thetrainingphasereducestheamountof error in thenetwork’s re-

186

sponses,which in turnreducespressurefor furtherlearningin theO � Pcomputation.As errordecreases,theweightsbecomesmaller(recall that this netwasalsotrainedwith a tendency for weight decay). Larger weightsareespeciallyimportantfor ex-ceptionwords,which mustcompeteagainsttheconspiracy of many smallerweightssupportingtypical spelling-soundcorrespondences.The impactof theadditionalse-manticinput duringtraining is thereforeanactualdeclinein O� P competence,firstfor LF Excwordsandeventually, to somedegree,evenfor HF Excwords.Thiseffectis not observablein the model’s accuracy during training becausethe incrementingsemanticinput keepsimproving the net’s performance.When,in the simulationofprogressive brain disease,this additionalinput is graduallywithdrawn, the underly-ing competenceof O � Pknowledgeis revealed.As a functionof decreasingstrengthof semanticinput, accuracy of pronunciation(a) remainshigh for Reg wordsof anyfrequency andfor nonwords,(b) declinessteadilyon LF Exc words,and(c) alsode-clines on HF Exc words, thoughthe vulnerability of thesecommonlyencounteredwordsis slower to emergeandalwayslessdramatic.

There is one further curve in Figure 4, somewhat difficult to resolve visuallyfrom thoseindicatingessentiallyceiling accuracy on regular wordsandnonwords:thepointsrepresentedby asterisksindicatethenatureof thenetwork’serrorresponsesto exceptionwords. Essentiallyall of theerrorswereregularizationsof the PINT

�/pInt/ variety. In thepurestcasesof reportedhumansurfacedyslexia (e.g.,Behrmann& Bub, 1992; McCarthy& Warrington,1986),onealso observesvirtually perfectreadingaloudof regular wordsandnonwordsandalmostexclusively regularizationerrorsto exceptionwords.

Althoughthisisapreliminaryandcertainlyincompleteaccountof surfacedyslexia,thesuccessof thesimulationis considerable.Thevaluesof accuracy for thedifferentword classesat variouspointsalongthe abscissain Figure4 correspondreasonablywell to thoseobserved for genuinecasesof surfacealexia. For example,on similar(thoughnot identical)setsof monosyllabicwords,PBandKT (two of theprogressivepatientsreportedby Patterson& Hodges,1992,Table2, p. 1030)both scored90–100%correcton high- andlow-frequency regular words. PB’s scoresfor the high-andlow-frequency exceptionwordswere86%and48%,respectively, which is verycloseto thenetwork’s performancewith “strengthof semantics”reducedto 0.5; andKT’s high- andlow-frequency exceptionscoreswere50%and8%, respectively: nottoo far off themodel’s performanceat thepoint whereall putative semanticsupporthasbeenwithdrawn,thoughthepatient’sLF Excvalueis clearlypoorerthanthenet’s.Critically, KT’scomprehensionlosswasmuchmoreseverethanPB’sasmeasuredbyother testssuchword-picturematching. We acknowledgethat our framework (andperhapsthoseof othersaswell) is a long way from an understandingof individualdifferencesin bothnormalandimpairedreadingskill. In particular, the ideasdevel-opedherehave recentlybeenchallengedby reportsof a few patientswith deficits

187

on semantictestswho arewithin thenormalrangeof performancein readingexcep-tion words(Cipolotti & Warrington,1995; Lambon-Ralph,Ellis & Franklin, 1995;Raymer& Berndt,1994).Somepossibleinterpretationsof thisdissociation,which isapparentlyincompatiblewith our prediction,arediscussedin Plautet al. (1996).

Weconcludefrom thissimulationwork that,althoughthestoryis perhapssome-whatdifferentfrom andmorecomplicatedthanthatoriginally envisioned,patternsofbothnormalanddisorderedreadingcanbeunderstoodin theframework proposedbySeidenberg andMcClelland(depictedin Figure1). Some,but not all, of thecritiqueof thisapproachby Coltheartetal. (1993)wasapposite.Contraryto theircritique,theinitial simulationsof Plautetal. (1996)demonstratethatasinglemechanismfor acti-vatingphonologyfrom orthographyis capableof acquiringknowledgethatsupportsbothpronouncingexceptionsandgeneralizingto novel forms. In line with their cri-tique,however, lesionsto suchnetworksdonot reproducethepatternof puresurfacealexia. We thereforeacknowledgethat another“pathway” or, aswe have character-izedit, anothersourceof input to phonologyappearsto benecessaryto modelsurfacealexia. Perhapsan even betterterm is the one employed by Kawamotoand Zem-blige (1992): sourceof constraint.They arguethattheir experimentalresults(whichconcerntheprecisetime-courseof normalskilled pronunciationof heterophonicho-mographslike bassandlead)canbeexplainedby a distributedperspective,but onlyif it providestwo constraintson pronunciation,orthographicandsemantic.On thebasisof aquitedifferentsetof humanandnetwork results,wehave reachedthesameconclusion.

Despitethis new apparentsimilarity betweendual-routeand triangle perspec-tives,someimportantdifferencesremain. For onething, andthis is the reasonthatwe preferthephrase“sourceof input or constraint,” we do not think in termsof twowholly separatepathwaysor routes.UnliketheDRCmodelof Coltheartetal. (1993),we assumethatthesameorthographicandphonologicalrepresentationssupportbothword andnonword reading. Although it hasnot beenfully implemented,the trian-gle modelof Figure1 is meantto representa genuinelyinteractive, recurrentsystemratherthanthecombinationof two separatepathways,O � PandO � S� P. Secondly,it mustbeemphasizedthatthetwo sourcesof constraintin thisapproachdonotdividetheEnglishlanguagevocabularyneatlyinto rule-obeyingandrule-infringingwordsasin thetwo routesof theDRCmodel.Theorthographicsourceof constraintonpronun-ciation,evenwith additionalsemanticinput,embodiesconsiderableknowledgeabouttheinconsistenciesof O � Prelationships,especiallyderivedfrom its experiencewiththe morecommonexceptionpatterns.This is why both the modelandthe patientsstill pronouncea fair numberof exceptionwordscorrectlyeven whenall semanticsupporthasbeeneliminated.Furthermore,aswe shallarguebelow, performanceonwordswith intermediatedegreesof consistency—irregularby strict “rule,” but withsupportfrom othersimilarly structuredwords—findsa morenaturalexplanationin

188

thetriangleframework thanin dual-routetheories.

5 Further Observations on Surface Dyslexic Reading

In the final sectionof this chapter, we introducesomefurther datafrom our recentstudiesof surfacedyslexic patientsthatseemespeciallycompatiblewith thekind offramework presentedhere.

5.1 Sub-Regularities

Table1 presentstheperformanceof threesurfacealexic patientsa on24monosyllabicexceptionwordswhich have atypicalpronunciationsof the body/ rime—wordslikePINT. The wordsin SetA wereselectedon a particularbasis,to be explainedin amoment,which led us to expectthat surfacealexic patientsmight attainanunusualdegreeof successin namingtheseparticularexceptionwords;theexceptionwordsinSetB werethenchosen,from a larger list of wordsnamedby all of thesepatients,to matchthosein seta ascloselyaspossibleon two othermeasuresknown to affectperformance:

(i) KuceraandFrancis(1967)writtenword frequency. Whereit wasnot possibleto find aSetB wordwith aprecisefrequency matchto ana item,thebiaswastowardsselectingamorefrequentwordfor B, towork againstourpredictionof greatersuccessin SetA. Thisbiasis reflectedin themeanfrequenciesfor thetwosets,whichareclosebut favor SetB.

(ii) Theratio of regular:exceptionpronunciationswithin thesetof monosyllabicwordssharingthatparticularorthographicbody. Word8a,for example,is PINT; of themonosyllabicEnglishwordsendingin -INT, 12 have regular pronunciations(HINT,LINT, PRINT, etc);PINT is theonly exception.For each(unique)bodyrepresentedinTable1, theregularexemplarsoutnumbertheexceptionexemplar(s),in mostcasesbyasubstantialnumber, asreflectedin themeanson thismeasureshown in theTable.

Although the numbersof items for this contrastare small, all three patientsshowed a reliableor nearlyreliableadvantagefor SetA over B: MP, 9/12 vs. 4/12,χ2 � 1� =4.2,p=.04;PB,10/12vs.4/12,χ2 � 1� =6.2,p=.01;AM, 11/12vs.7/12,χ2 � 1� =3.6,p=.059.Combiningthethreepatients’datayieldsahighly reliablecontrast,30/36vs.15/36,χ2 � 1� =13.3,p � .001. No word in SetB wasnamedcorrectlyby all patients,whereas7/12SetA wordsweregivencorrectpronunciationsby all three. Virtuallyall of the errorsby all threepatientson bothsetsof wordswereregularizations:allthreenamedPINT to rhymewith “hint” andGROSS to rhymewith “moss”; two of the

aDescriptionsof the generalcharacteristicsof the threepatientsin Table1 canbe found as follows:for MP: BehrmannandBub (1992),Bub, Black, HampsonandKertesz(1988)andBub, CancelliereandKertesz(1985);for PB: PattersonandHodges(1992);for AM: HodgesandPatterson(in press).

189

Table 1: Two setsof 12 exceptionwords,showing—for eachword—theKucera-Francisfrequency, thenumberof exemplarsin theBody Neighborhoodwith Reg:Excpronunciations, andthenumberof correctpronunciations (out of 3) for threesurfacealexic patients(MP, PB,AM); thelower partof theTableshowsperformanceon thesamesetsof wordsby thePlautet al. (1996)networkwith variousdegreesof reduced

“semantic”input (asin Figure4).

SetA SetBWord Freq Reg:Exc +/3 Word Freq Reg:Exc +/3

1a 464 9:1 3 1b 437 9:1 12a 37 15:3 3 2b 88 14:2 23a 3 10:3 3 3b 14 9:3 24a 4 4:1 3 4b 11 5:1 25a 760 4:1 3 5b 391 10:2 26a 11 3:1 3 6b 16 2:1 27a 67 3:2 3 7b 58 8:3 18a 5 12:1 2 8b 13 12:1 09a 81 9:1 2 9b 66 9:1 0

10a 2 5:1 2 10b 84 5:1 211a 94 2:1 2 11b 730 4:2 012a 4 4:1 1 12b 9 6:4 1

Mean 127.7 6.7:1.4 0.83 159.8 7.8:1.8 0.42

Network’sproportioncorrectsemanticstrength= 0.6 1.00 0.67

= 0.4 0.92 0.50= 0.2 0.83 0.42= 0.0 0.50 0.25

190

threepronouncedPUT to rhymewith “hut” andLOSE to rhymewith “nose”; andsoon.

WhenthePlautet al. (1996)network (whereadditional“semantic”input to thephonemeunitsduringtrainingis subsequentlyreducedin strength,asin Figure4 de-scribedabove) is testedon the same24 words,it yields a similar advantagefor SetA over B, asshown in the lower part of Table1. As it happens,whenthe strengthof additionalinput is at0.2,thenetwork’sperformanceexactlymatchesthepatients’:83%correcton SetA, 42%correcton SetB. All of thenetwork’s errorswereregu-larizations.

Whatis it aboutSetA wordsthatmakesthemrelatively (thoughclearlynot alto-gether)immuneto regularization,by bothpatientsandnetwork? Theanswer, in ourview, is that thesewordsenjoy a kind of sub-regularity, basednot on the body/rimebut on thecombinationof theinitial consonantandvowel. Almost all (10/12)of thewordsin setabegin eitherWA- or WO-, andtheremainingtwo (SWAMP andQUART)have the samecharacter.b Leaving asidewordssuchas WAKE or WOKE wherethepronunciationof thevowel is signalledby thefinal -E, a substantialmajority of WA-wordsarepronouncednot in accordancewith the usualpronunciationof the vowelandbody (as in CASH or CART) but ratherlike WASH or WART. Likewise, a greatmajority of WO- wordsarepronouncednot with thevowel in NORTH but ratherlikeWORTH. As discussedby Seidenberg (1992),in quasi-regularsystemslike spelling-soundcorrespondenceandpast-tenseverbformationin English,anumberof patternsthat do not follow the most generalrule are nonethelesscharacterizedby this sortof sharedirregularity, thusforming a sub-regularity. Spelling-soundknowledgeap-parentlyreflectsthis sub-regularity; asa result,comparedto otherexceptionwordswith similar familiarity levels andbody neighborhoods,“W-words” dependlessonthe additionalsourceof constrainton pronunciationandsoarelessvulnerableto itsremoval.

Dual-routemodelslike thatof Coltheartet al. (1993)can,of course,accountfortherelativeinvulnerabilityof W-words,but only by complicatingtherulesystem.

bIn view of demonstrationsthat at leastnormalreaders’pronunciationsof wordsandnonwordsmaybesubjectto priming or biasingeffectsfrom othersimilar itemsin the list context (Kay & Marcel,1981;Seidenberg et al., 1984), it shouldbe notedthat the patientswerenot askedto readthe 12 W-wordsasa block; theseitemswereembeddedin, andwell distributed throughout,a muchlarger list of wordsandnonwords(total N � 198) . Furthermore,asthis list containedboth regularwordsandnonwordswith thesamebodiesastheW-words(e.g.,FORK andLORK aswell asWORK, FARM andDARM aswell asWARM;wordssharingbodiesaswell asonsetswerewell separatedthroughouttheset),any biasingeffect from thepronunciation of otherW-wordsshouldhave beenoffsetby effectsof theseitemswith thesamebodybutdiscrepantpronunciations.

191

5.2 The Fate of Regular Words in Progressive Surface Dyslexia

Over the pastfive yearsor so, variousauthorsof this chapterhave carriedout de-tailed investigationsof arounda dozenpatientswith acquiredsurfacedyslexia. Allof thesecaseshave hadeitheramoderate-to-severeimpairmentof semanticmemory(e.g.,Behrmann& Bub,1992;Patterson& Hodges,1992)or at leasta profoundim-pairmentin activationof phonologyby semantics,asrevealedfor exampleby severeanomia(e.g.,Graham,Patterson& Hodges,1995;Watt, Jokel & Behrmann,1996).All but oneof the caseshassufferedfrom a neuro-degenerative diseasec character-ized behaviorally assemanticdementiaor progressive aphasia(seeHodges,Patter-son,Oxbury & Funnell,1992,for furtherdescription);theoneexceptionis MP, whosustainedamajorheadinjury resultingin unusuallyfocaldamageto theleft temporallobe (Behrmann& Bub, 1992;Patterson& Behrmann,submitted).With a numberof the progressive patients,we have beenable to performlongitudinalassessmentsof readingperformance,someof which arestill in progress.On initial assessment,virtually all of the patientscould be describedas having a patternof pure surfacedyslexia, in thesensethattheir accuracy of word namingwasnotablyoutsidenormallimits for exceptionwordsbut within normalrangefor regularwords. This is illus-tratedin Figure5 for PB andFM, two of the casesfrom the PattersonandHodges(1992)study. Thesedataarefrom the list of wordsemployed in that study, whichconsistsof 126 pairsof monosyllabicregular andexceptionwords(e.g., PINE andPINT, BLACK and BLOOD, etc) matchedfor frequency, lengthandinitial phoneme,in threefrequency bands.Figure5 alsoshows the two patients’performanceon thesamelist approximatelythree(FM) or four (PB) yearslater. Although thereis stilla highly reliableadvantagefor regularover exceptionwordsfor bothpatientsat thisstage,regularwordsnow yield botherrorsandadegreeof frequency sensitivity (withaslight reversalfor FM betweenmedium-andlow-frequency regularwords).

What is thenatureof theseemerging errorson regularwords,andhow aretheyto be explained? Over the courseof the 3–4 yearperiod,PB and FM wereaskedto namethis list of wordsa total of five andeight times,respectively. Table2 pro-videsa classificationof theentiresetof errorsmadeby eachpatient,on bothregularand exceptionwords, into two broadcategories. Taking the less interestingcate-gory of “Other” errorsfirst, bothPB andFM—like virtually all patientswith surfacealexia (seefor exampleColtheartet al., 1983),andindeedlike virtually all patientswith any kind of readingdisorder—make a certainnumberof errorswherethe re-sponsebearsa relationshipto the target word that is neither“surface” nor “deep”

cFor readersinterestedin theunderlyingpathologyof theseconditions:threeof the patientsreportedin PattersonandHodges(1992)have cometo post-mortemanalysis.Two hadPick’s disease.The thirdhadAlzheimerpathologybut in ahighly atypicaldistribution: theprofoundfocal left temporalatrophyandsevereneuronallossin this regionwasmorecharacteristicof semanticdementiadueto Pick’s diseasethanof AD.

192

High� Medium� Low�Frequency�

0.0

0.2

0.4

0.6

0.8

1.0

Pro

port

ion

Cor

rect

PB (June 1990)

High� Medium� Low�Frequency�

0.0

0.2

0.4

0.6

0.8

1.0

Pro

port

ion

Cor

rect

PB (April 1994)

Regular�Exception

�

High� Medium� Low�Frequency�

0.0

0.2

0.4

0.6

0.8

1.0

Pro

port

ion

Cor

rect

FM (April 1991)

High� Medium� Low�Frequency�

0.0

0.2

0.4

0.6

0.8

1.0

Pro

port

ion

Cor

rect

FM (May 1994) Regular

�Exception

�

(a) (b)

Figure5: Readingperformanceby two progressive surfacedyslexic patients,PBandFM, (a)at their initialassessment,on thePattersonandHodges(1992)high-,medium-andlow-frequency regularandexception

words;and(b) about4 (PB)and3 (FM) yearslater.

193

Table2: Errordatafrom multipleadministrations of thesurfacelist to PBandFM, showing eachpatient’stotal numbersof errorsandthe classificationof theseinto two broadclasses:(1) LARC errors(shortforLegitimateAlternative Readingof Components),e.g.,PINT (anexceptionword)pronouncedto rhymewith“mint” andHOOT (a regularword)pronouncedto rhymewith “foot”; (2) Othererrors(variouslydescribed

by otherauthorsasvisual,orthographic,phonological), e.g.,ONCE ”ounce.”

Times Error WordTypePatient Tested Type Regular Exception

PB 5Total 71/630 291/630LARC 34 246Other 37 45

FM 8Total 129/1008 337/1008LARC 41 243Other 88 94

but simply resemblesit orthographicallyand/orphonologically. FM in particularhasbecomemoreproneto “Other” errorsasher readingdisorderhasworsened;for ex-ample,shenow makesa numberof theletterconfusionerrors,mainly betweenvisu-ally similar letters(e.g.,PRAY

� ”bray”), thataremorecharacteristicof purethanofsurfacealexia. Most, thoughnot all, of the“Other” errorsaresubstitutionsof anor-thographically/phonologicallysimilar realword for thetarget,asin ONCE

� “ounce,”THROAT

� “trout,” etc.BothPB andFM producedroughlyequalnumbersof “Other”errorsto regularandexceptionwords. This is aswe would expect,providedthat theregularandexceptionwordsarereasonablywell matchedin orthographicandphono-logical characteristics(suchassimilarity to otherwordsin their neighborhoods),anddiffer only in thepredictabilityof therelationshipbetweenspellingandsound.

The error category germaneto the presentdiscussionis what we have dubbedLARC errors(a term first usedby Patterson,Suzuki, Wydell & Sasanuma,1995),shortfor LegitimateAlternativeReadingof Components.LARC meansthat the in-correctresponsereflectsalegitimatepronunciationof eachcomponentof theword,inthe sensethat theorthographiccomponenttakesthatpronunciationin otherEnglishwords. The“Other” responseTHROAT

� ”trout” cannotbeclassedasa LARC errorbecausethereareno Englishwords(at leastin the dialectof Englishspoken by PBandFM), in which TH is pronounced/t/, andtherearealsono wordsin which OA ispronounced/au/asin “trout.” ThequintessentialLARC error is a regularizationlikePINT

� /pInt/; but thereareothertypesof sucherrorsaswell. First of all, exceptionwordscananddo yield LARC errorswhich arenot pureregularizations.For exam-ple,theregularizationof BLOOD wouldrhymewith “food,” but PBpronouncedit like

194

“good”; theregularizationof SWEAT would rhymewith “treat” but hepronounceditlike “great”; andsoon. Secondly, of even greaterinterest,regular wordsalsoyieldLARC errors:PBnamedHOOT to rhymewith “foot,” YEAST like“breast,” HEAR like“bear”; FM pronouncedBROWN with thebody/rimeof “blown,” HEAT like “threat,”COST like “post”; andsoon. Not surprisingly, sincethevery definitionof anexcep-tion word is thatat leastoneof its componentshasadifferent,legitimate,andindeedmorecommon,pronunciation,both PB andFM madefar moreLARC errorsto ex-ceptionwordsthanto regularwords. The importantobservation is thatwhile errorsto regularandexceptionwordsmaydiffer in quantity, they donotdiffer in natureandthereforedo not requireadifferentkind of account.

Many of theregularwordsyieldingLARC errorsare,of course,thetypeof wordknown (sinceGlushko, 1979)as regular inconsistent. A regular inconsistentword(like HOOT) takes the pronunciationthat is most typical of its body neighborhood;but oneor more wordswith the samebody have a conflicting pronunciation(e.g.,FOOT). Accordingto theframework andsimulationsin Plautetal. (1996),suchregu-lar inconsistentwords—ratherakin to theW-words—representanintermediatecasebetweenwordswith completelypredictablecomponentsand true exceptions. Therelatively infrequentbut illuminating readingerrorsof surfacedyslexic patientstothesewords strike us asanothersignificantmatchbetweenthe triangle model andreal data. As with the decreasedvulnerability of W-words in surfacedyslexia, theslightly increasedvulnerabilityof regularinconsistentwordsis a directpredictionofthe triangleframework. No doubt thesedatacanalsobe given an explanationin adual-routeframework. In theDRCmodel,this wouldpresumablyinvolve interactionbetweenthephonemesystem(which is activatedby theGPCrulesystem,andshouldsupportcorrectreadingof a regularword like hoot)andthephonologicaloutputlex-icon. If presentationof HOOT partiallyactivatesthelexical representationsfor ortho-graphicallysimilarwordslikeFOOT andSOOT, therule-basedpronunciationof HOOT

might occasionallysuccumbto this influenceandbeir-regularized.As suggestedbySasanuma,Itoh, PattersonandItoh (in press),to theextentthatsuchinteractiveinflu-encesbetweenlexical andnon-lexical systemsprovide amajorexplanatoryprinciplein theDRCmodel,thedifferencesbetweenthetwo approachesbecomelesscritical.

Weconcludewith anobservationwhich, in thecontext of thisbook,mayconsti-tutepreachingto theconverted:thatbothcomponentsof thebook’s topicareprovingimportantin the effort to understandthe humanbrain andits capabilities.Not onlymustwebuild computationalmodelswhosepredictionscanbetested;thedataagainstwhich themodelsaretestedmustincludedisorderedaswell asnormalfunctioning.

195

Acknowledgments

Thework reportedherewassupportedin partby grantsfrom theNIMH (USA) to J.L.McClellandandcolleagues,andfrom theMRC (UK) to J.R. Hodgesandcolleagues.Wearegratefulto NaidaGrahamandEamonStrainfor researchassistance.

References

Baluch, B., & Besner, D. (1991).Visual word recognition: Evidencefor strategiccontrolof lexical andnonlexical routinesin oral reading.Journal of ExperimentalPsychology: Learning, Memory and Cognition, 17, 644–652.

Behrmann,M., & Bub, D. (1992).Surfacedyslexia and dysgraphia:Dual routes,singlelexicon.Cognitive Neuropsychology, 9, 209–251.

Besner, D., Twilley, L., McCann,R. S., & Seergobin,K. (1990).On theconnectionbetweenconnectionismanddata:Are a few wordsnecessary?Psychological Re-view, 97, 432–446.

Breedin,S.D., Saffran,E. M., & Coslett,H. B. (1994).Reversalof theconcretenesseffect in a patientwith semanticdementia.Cognitive Neuropsychology, 11, 617–660.

Bub, D., Black, S., Hampson,E andKertesz,A. (1988).Semanticencodingof pic-turesandwords:Someneuropsychologicalobservations.Cognitive Neuropsychol-ogy, 5, 27–66.

Bub, D., Cancelliere,A., & Kertesz,A. (1985).Whole-word andanalytictranslationof spellingto soundin anonsemanticreader. In K. Patterson,J.C. Marshall& M.Coltheart(Eds.),Surface dyslexia (pp.15–34).London:Erlbaum.

Bullinaria,J.A. (in press).Connectionistmodelsof reading:Incorporatingsemantics.In Proceedings of the 18th Annual Conference of the Cognitive Science Society.Hillsdale,NJ:Erlbaum.

Castles,A., & Coltheart,M. (1993).Varietiesof developmentaldyslexia. Cognition,47, 149–180.

Cipolotti, L., & Warrington,E. K. (1995).Semanticmemoryandreadingabilities: Acasereport.Journal of the International Neuropsychological Society, 1, 104–110.

Coltheart,M. (1985).Cognitive neuropsychologyand the study of reading.In M.I. PosnerandO. S. M. Marin (Eds.),Attention and performance XI (pp. 3- 37).Hillsdale,NJ:Erlbaum.

Coltheart,M., Curtis,B., Atkins, P., & Haller, M. (1993).Modelsof readingaloud:Dual-routeandparallel-distributed-processingapproaches.Psychological Review,100, 589–608.

Coltheart,M., Masterson,J., Byng, S., Prior, M., & Riddoch,J. (1983). Surfacedyslexia. Quarterly Journal of Experimental Psychology, 35A, 469–495.

196

Coltheart,M., & Rastle,K. (1994).Serial processingin readingaloud: Evidencefor dual-routemodelsof reading.Journal of Experimental Psychology: HumanPerception and Performance, 20, 1197–1211.

Franklin,S., Howard,D., & Patterson,K. (1995).Abstractword anomia.CognitiveNeuropsychology, 12, 549–566.

Funnell,E. (1983).Phonologicalprocessingin reading:New evidencefrom acquireddyslexia. British Journal of Psychology, 74, 159–180.

Funnell,E. (in press).Responsebiasesin oral reading:An accountof theco- occur-renceof surfacedyslexia andsemanticdementia.Quarterly Journal of Experimen-tal Psychology.

Glushko, R. J. (1979).The organizationandactivation of orthographicknowledgein readingaloud.Journal of Experimental Psychology: Human Perception andPerformance, 5, 674–691.

Graham,K. S.,Hodges,J.R.,& Patterson,K. (1994).Therelationshipbetweencom-prehensionandoral readingin progressive fluent aphasia.Neuropsychologia, 32,299–316.

Graham,K. S., Patterson,K., & Hodges,J. R. (1995).Progressive pureanomia:In-sufficientactivationof phonologyby meaning.Neurocase, 1, 25–38.

Henderson,L. (1982).OrthographyandWordRecognitionin Reading.London:Aca-demicPress.

Hillis, A. E., & Caramazza,A. (1991).Mechanismsfor accessinglexical represen-tationsfor output: Evidencefrom a category specificsemanticdeficit. Brain andLanguage,40,106–144.

Hodges,J. R., Graham,N., & Patterson,K. (1995).Chartingthe progressionin se-manticdementia:Implicationsfor theorganisationof semanticmemory. Memory,3, 463–495.

Hodges,J.R.,& Patterson,K. (in press).Non-fluentprogressiveaphasiaandsemanticdementia:A comparative neuropsychologicalstudy. Journal of the InternationalNeuropsychological Society.

Hodges,J. R., Patterson,K., Oxbury, S., & Funnell,E. (1992).Semanticdementia:Progressivefluentaphasiawith temporallobeatrophy. Brain, 115, 1783–1806.

Howard, D., & Frankln,S. (1988).Missing the meaning? Cambridge,Mass: MITPress.

Humphreys,G.W., & Evett,L. J.(1985).Are thereindependentlexicalandnonlexicalroutesin word processing?An evaluationof the dual-routetheory of reading.Behavioral and Brain Sciences, 8, 689–740.

Kawamoto,A. H., & Zemblige,J. H. (1992).Pronunciationof homographs.Journalof Memory and Language, 31, 349–374.

Kay, J., & Marcel, A. J. (1981).One process,not two, in readingaloud: Lexicalanalogiesdo thework of nonlexical rules.Quarterly Journal of Experimental Psy-

197

chology, 33A, 397–414.Kucera,H., & Francis,W. N. (1967).Computational analysis of present-day Ameri-

can English. Providence,RI: Brown UniversityPress.Lambon-Ralph,M., Ellis, A. W., & Franklin,S.(1996).Semanticlosswithoutsurface

dyslexia. Neurocase, 1, 363–369.Marshall,J. C., & Newcombe,F. (1973).Patternsof paralexia: A psycholinguistic

approach.Journal of Psycholinguistic Research, 2, 175–199.Newcombe,F., & Marshall,J.C. (1980).Transcodingandlexical stabilizationin deep

dyslexia. In M. Coltheart,K. Patterson& J.C. Marshall(Eds.),Deep dyslexia (pp.176–188).London:Routledge.

McCarthy, R., & Warrington,E. K. (1986).Phonologicalreading: Phenomenaandparadoxes.Cortex, 22, 359–380.

Paap,K. R., & Noel, R. W. (1991).Dual routemodelsof print to sound:still a goodhorserace.Psycholinguistic Research, 53, 13–24.

Parkin,A. J. (1993).Progressiveaphasiawithout dementiadueto focal left temporo-frontal hypometabolism—Aclinical andcognitiveanalysis.Brain and Language,44, 201–220.

Patterson,K., & Behrmann,M. (1996).Frequency andconsistency effectsin a puresurfacedyslexic patient.Submittedto Journal of Experimental Psychology: Hu-man Perception and Performance.

Patterson,K., & Hodges,J. R. (1992).Deteriorationof word meaning:Implicationsfor reading.Neuropsychologia, 30, 1025–1040.

Patterson,K., Seidenberg, M. S., & McClelland,J. L. (1989).Connectionsanddis-connections:Acquired dyslexia in a computationalmodel of reading.In R. G.M. Morris (Ed.),Parallel distributed processing: Implications for psychology andneuroscience (pp.131–181).Oxford: OUP.

Patterson,K., Suzuki,T., Wydell, T., & Sasanuma,S.(1995).Progressiveaphasiaandsurfacealexia in Japanese.Neurocase,1, 155–165.

Plaut,D. C., McClelland,J. L., Seidenberg, M. S., & Patterson,K. (1996).Under-standingnormalandimpairedword reading: Computationalprinciplesin quasi-regulardomains.Psychological Review, 103, 56–115.

Raymer, A. M., & Berndt, R. S. (1994).Models of word reading: EvidencefromAlzheimer’sdisease.Brain and Language, 47, 479–482.

Reggia, J. A., Berndt,R. S., & D’Autrechy, C. L. (1994).Connectionistmodelsinneuropsychology. In F. Boller& J.Grafman(Eds.),Handbook of Neuropsychology,Vol. 9 (pp297–33).NorthHolland,Elsevier.

Sasanuma,S., Itoh, H., Patterson,K., & Itoh, T. (in press).Phonologicalalexia inJapanese:A casestudy. Cognitive Neuropsychology.

Seidenberg, M. S. (1992).Connectionismwithout tears.In S. Davis (Ed.), Connec-tionism: Advances in theory and practice (pp 84–137).Oxford: OUP.

198

Seidenberg, M. S., & McClelland,J. L. (1989).A distributed,developmentalmodelof word recognitionandnaming.Psychological Review, 96, 523–568.

Seidenberg, M. S.,Waters,G. S.,Barnes,M. A., & Tanenhaus,M. K. (1984).Whendoesirregular spellingor pronunciationinfluenceword recognition? Journal ofVerbal Learning and Verbal Behavior, 23, 383–404.

Sejnowski,T. J.,& Rosenberg,C.R. (1987).Parallelnetworksthatlearnto pronounceEnglishtext. Complex Systems, 1, 145–168.

Shallice, T., & Warrington,E. K. (1980). Single and multiple componentcentraldyslexic syndromes.In M. Coltheart,K. Patterson& J. C. Marshall(Eds.),Deepdyslexia (pp.119–145).London:Routledge.

Shallice,T., Warrington,E. K., & McCarthy, R.(1983).Readingwithout semantics.Quarterly Journal of Experimental Psychology, 35A, 111–138.

Strain,E.,Patterson,K., & Seidenberg, M. S.(1995).Semanticeffectsin singlewordnaming.Journal of Experimental Psychology: Learning, Memory and Cognition,21, 1140–1154.

Taraban,R., & McClelland, J. L. (1987).Conspiracy effects in word recognition.Journal of Memory and Language, 26, 608–631.

VanOrden,G. C., & Goldinger, S. D. (1994).Interdependenceof form andfunctionin cognitivesystemsexplainsperceptionof printedwords.Journal of ExperimentalPsychology: Human Perception and Performance, 20, 1269–1291.

Watt, S., Jokel, R., & Behrmann,M. (in press).Surfacedyslexia in non-fluentpro-gressiveaphasia.Brain and Language.

199