Grapheme-to-lexeme feedback in the spelling …web.jhu.edu/cogsci/templates/images/rapp/mccloskeymacarusorapp2006.pdfGrapheme-to-lexeme feedback in the spelling system: Evidence from

Grapheme-to-lexeme feedback in the spellingsystem: Evidence from a dysgraphic patient

Michael McCloskeyJohns Hopkins University, Baltimore, USA

Paul MacarusoCommunity College of Rhode Island, Warwick, and Haskins Laboratories, New Haven, USA

Brenda RappJohns Hopkins University, Baltimore, USA

This article presents an argument for grapheme-to-lexeme feedback in the cognitive spellingsystem, based on the impaired spelling performance of dysgraphic patient CM. The argumentrelates two features of CM’s spelling. First, letters from prior spelling responses intrude into sub-sequent responses at rates far greater than expected by chance. This letter persistence effect arisesat a level of abstract grapheme representations, and apparently results from abnormal persistence ofactivation. Second, CM makes many formal lexical errors (e.g., carpet ! compute). Analysesrevealed that a large proportion of these errors are “true” lexical errors originating in lexical selec-tion, rather than “chance” lexical errors that happen by chance to take the form of words.Additional analyses demonstrated that CM’s true lexical errors exhibit the letter persistenceeffect. We argue that this finding can be understood only within a functional architecture inwhich activation from the grapheme level feeds back to the lexeme level, thereby influencinglexical selection.

INTRODUCTION

Like other forms of language processing, writtenword production implicates multiple levels ofrepresentation, including semantic, orthographiclexeme, grapheme, and allograph levels. In thisarticle we explore how the levels interact: Doesprocessing proceed in a strictly feedforwardmanner, from earlier to later levels of represen-tation? Or are there feedback as well as feed-forward connections between levels, so thatlater levels can influence processing at earlierlevels? We describe the performance of CM,

a brain-damaged patient with an acquired spellingdeficit, arguing from his error pattern thatthe cognitive system for written word produc-tion includes feedback connections from gra-pheme representations to orthographic lexemerepresentations.

Figure 1 depicts a general theory of thecognitive mechanisms implicated in written wordproduction (see Tainturier & Rapp, 2001, for arecent overview of this theoretical frameworkand relevant evidence; for a somewhat differentperspective, see Graham, Patterson, & Hodges,1997, 2000). The major assumptions of the

Correspondence concerning this article should be addressed to Michael McCloskey, Department of Cognitive Science, Krieger

Hall, Johns Hopkins University, Baltimore, Maryland 21218, USA (Email: [email protected]).

This study was supported in part by NIH grant NS22201 and NIMH grant R29MH55758. We thank Donna Aliminosa for

testing patient CM, and Marie-Josephe Tainturier for helpful comments on the manuscript.

COGNITIVE NEUROPSYCHOLOGY, 2006, 23 (2), 278–307

278 # 2006 Psychology Press Ltd

http://www.psypress.com/cogneuropsychology DOI:10.1080/02643290442000518

theory may be described in the context of writingwords to dictation. When a familiar word isdictated, its phonological lexeme is activated in aphonological lexicon. Activation of the lexemeleads to activation of a lexical-semantic represen-tation within the semantic system, and thesemantic representation in turn activates an ortho-graphic lexeme in an orthographic lexicon. (Sometheorists have proposed that orthographic lexemesmay also be activated directly from phonologicallexemes; see, e.g., Patterson, 1986; Roeltgen,Rothi, & Heilman, 1986; Romani, Olson, Ward,& Ercolani, 2002; but see Hillis & Caramazza,1991b.)

From the orthographic lexeme, activationspreads to representations of the word’s constitu-ent graphemes. These abstract letter-identityrepresentations provide the basis for activating

allograph (letter-shape) representations appro-priate for the desired form of written output(e.g., lower-case script, upper-case print), leadingultimately to production of writing movements.In the case of oral spelling, the abstract graphemerepresentations serve to activate letter-namerepresentations.

When an unfamiliar word or pseudoword(e.g.,/floup/) is dictated, a somewhat differentprocess is required, because no appropriate lexemeor semantic representations are available to beactivated. According to the theory, graphemescorresponding to a plausible spelling (e.g., flopeor floap) are activated by means of a sublexicalphoneme–grapheme conversion process thatexploits sound-to-spelling correspondences. Asin the case of familiar word stimuli, the activatedgraphemes then activate letter shape or lettername representations.

Spelling processes may, of course, also beinvoked by nonphonological stimuli. Forexample, when lexical-semantic representationsare activated by pictures, objects, or thoughts thelexical item may be realised through written ororal spelling by activating orthographic lexemes,graphemes, and allographs just as in writing todictation.

Although there are many issues regardingorthographic representation and processing thatwe have not discussed, the functional architecturewe have sketched is sufficiently detailed to serve asa framework for the work reported in this article.

Bidirectional processing between segmentsand words

The present study concerns the connectivity, andhence the processing interactions, between ortho-graphic lexeme and grapheme levels of represen-tation. One possibility, illustrated in Figure 2a(and Figure 1), is that the flow of informationbetween levels is strictly feedforward, such thatsemantic representations activate orthographiclexemes that, in turn, activate their constituentgraphemes. An alternative, depicted inFigure 2b, is that in addition to the forward flowof activation, information from the grapheme

Figure 1. Schematic depiction of the cognitive spellingmechanisms.

GRAPHEME-TO-LEXEME FEEDBACK IN THE SPELLING SYSTEM

COGNITIVE NEUROPSYCHOLOGY, 2006, 23 (2) 279

level influences lexical selection through thefeeding-back of activation from grapheme tolexeme representations prior to lexical selection.

The question of unidirectional versus bidirec-tional connectivity between lexeme and segmentlevels has received little attention in studies ofwritten word production (although see Romaniet al., 2002, for a brief discussion). However, theissue has been intensively debated in research onspoken production. Evidence from the spokenproduction errors of both neurologically intactand neurologically impaired individuals has beenput forward in support of phoneme-to-lexemefeedback. Several researchers have argued for feed-back on the basis of results suggesting that form-based errors take the form of real words moreoften than expected by chance (the lexical biaseffect; see, e.g., Baars, Motley, & MacKay,1975; Best, 1996; Dell, 1986; Dell & Reich,1980; Martin, Dell, Saffran, & Schwartz, 1994;Stemberger, 1985). Analyses demonstrating asignificant influence of phonology on semanticerrors have also been taken as evidence forphoneme-to-lexeme feedback (Dell, 1986; Dell& Reich, 1981; Harley, 1984; Martin, Gagnon,Schwartz, Dell, & Saffran, 1996). AlthoughLevelt and colleagues (Levelt, Roelofs, & Meyer,1999), as well as Garrett (1976), have claimedthat the data can be interpreted within a strictlyfeedforward architecture, Dell and colleagues

(Dell, Schwartz, Martin, Saffran, & Gagnon,1997), Rapp and Goldrick (2000, 2003;Goldrick & Rapp, 2002) have argued that the evi-dence implies at least some influence of phoneme-level activation on the lexical selection process. Inthis paper we argue that the pattern of spellingerrors produced by CM reveals an effect of gra-pheme-level activation on lexical selection, andhence provides strong evidence for feedbackfrom grapheme to orthographic lexemerepresentations.

CASE HISTORY

CM is a right-handed man with a PhD inelectrical engineering. He worked as a universityprofessor until suffering a stroke in September1986, at age 59. CT showed extensive corticaland subcortical damage in the distribution of theleft middle cerebral artery. A premorbid writingsample, consisting of handwritten lecture notes,confirmed that CM’s spelling was normal priorto his stroke. All 430 words in the sample werespelled correctly, including many long low-frequency words (e.g., sinusoidal).

General language assessment

Boston Diagnostic Aphasia Examination (BDAE;Goodglass & Kaplan, 1983)CM performed well on the auditory comprehen-sion tasks. He made no errors in word discrimi-nation and body part identification, and scored14/15 in following spoken commands and 10/12in responding to questions involving complexideational material. He also performed well inreading comprehension, obtaining high scores insymbol discrimination, word recognition, andword/picture matching. His comprehension scorefor reading sentences and paragraphs was 7/10.

In spoken language production CM showedimpairment. He was able to recite automatisedsequences and repeat single words, but scoredonly 2/16 in repeating phrases. The speech path-ologist conducting the examination rated hisspeech at 3 (of a possible 7) for melodic line, and

(a) (b)

Figure 2. (a) A strictly feedforward architecture. (b) Afunctional architecture with feedback between graphemeand orthographic lexeme levels.

MCCLOSKEY, MACARUSO, RAPP

280 COGNITIVE NEUROPSYCHOLOGY, 2006, 23 (2)

4 for articulatory agility. Phrase length was twowords. Confrontation and responsive namingwere poor, and CM generated only 1 animalname in 90 seconds. He read aloud 3 of 10words and no sentences.

In written language production the examinerrated CM at 5 (of 5) in writing mechanics (sameas premorbid ability, with allowances made foruse of nonpreferred hand). CM could write thealphabet in sequence, and had success writingnumerals, letters, and short high-frequencywords to dictation. Spelling to dictation waspoor for long words and sentences. CM alsoperformed poorly in written picture naming, andproduced only 3 relevant words in a writtennarrative (rating 1 of a possible 5).

Peabody Picture Vocabulary Test-Revised (Dunn &Dunn, 1981)In this test words were presented aurally one at atime, and CM indicated which of four picturesmatched the word. On each of two adminis-trations his score—147/175 and 161/175—fellwithin the average range.

Word–picture matchingOn each trial in this task a single picture was firstpresented; a word was then presented aurally, andCM indicated whether the word matched thepicture. Stimulus pictures were 144 line drawingsof concrete objects (e.g., a stool). Each picture waspresented three times, once paired with the correctname (stool), once with a semantic foil (bench), andonce with a phonological foil (stamp). CMresponded correctly on 98% (425/432) of the trials.

Spelling assessment

CM’s written spelling was assessed with theJohns Hopkins University Dysgraphia Battery(Goodman & Caramazza, 1985). The batteryprobes spelling of dictated words and nonwords,as well as copy transcoding and written picturenaming. Several factors (e.g., word length, concre-teness) are varied systematically across the set of

stimulus items. Normal control subjects performall of the tasks with very low error rates.

Writing words to dictationAs shown in Table 1, CM was only 39%correct (128/326) in writing words to dictation.His accuracy was higher for high-frequencywords (44%) than for low-frequency words(33%), although this difference was only margin-ally significant, x2 (1, N ¼ 292) ¼ 3.70, .05 ,

p , .10. No significant effects were observed

Table 1. CM’s performance on the Johns HopkinsDysgraphia Battery

Task No. of stimuli No. correct % correct

Writing to dictation

Word/Nonword status

Words 326 128 39

Nonwords 34 4 12

Word frequency

High 146 64 44

Low 146 48 33

Grammatical word class

Nouns 28 10 36

Verbs 28 8 29

Adjectives 28 4 14

Functors 20 11 55

Concreteness

Concrete 21 8 38

Abstract 21 6 29

Phoneme–grapheme conversion

Probability

High 30 15 50

Low 80 42 53

Word length

4 letters 14 6 43

5 letters 14 3 21

6 letters 14 6 43

7 letters 14 5 36

8 letters 14 3 21

Written picture naming 51 13 25

Copy transcoding

Direct

Words 84 84 100

Nonwords 40 40 100

Delayed

Words 84 59 70

Nonwords 40 23 58



for grammatical word class, concreteness, orphoneme–grapheme conversion probability (thelikelihood of generating a correct spellingthrough sublexical phonology-to-orthographyconversion processes). The Dysgraphia Batteryresults also showed no reliable effect of wordlength. However, in subsequent testing with alarger set of stimuli that manipulated wordlength while controlling word frequency, CM’sspelling accuracy was significantly higher for 3–4letter words (69%) than for 5–8 letter words(50%), x2 (1, N ¼ 276) ¼ 10.16, p , .01.

One fourth of CM’s errors on the DysgraphiaBattery (49/198) were lexical errors, in which heproduced the spelling of a word other than thetarget word (e.g., solid ! solar, method !mother). In the vast majority of these errorsCM’s response bore a clear (although not alwaysclose) orthographic relation to the target word(e.g., cross ! cough, vivid ! given). Only 3 ofthe lexical errors (6%) could be considered seman-tic or morphological (e.g., weave ! weaver).

Three fourths of CM’s errors (149/198) werenonlexical (e.g.,valley ! valice, kitchen ! kithen).As in the case of the lexical errors, CM’s responseand the target word were nearly always orthogra-phically related, although not always closely(e.g., merge ! mourer). Only 8 of the 141 errors(4%) could be considered phonologically plausible(e.g., open ! opon), suggesting that CM wasnot spelling through application of sublexicalphonology-to-orthography conversion processes.

To characterise more specifically the ortho-graphic relatedness between target and errorresponses, the lexical and nonlexical errors weresubmitted to a computer program that attemptedto interpret each error as a letter substitution,deletion, transposition, insertion, or some com-bination of these error types. (For a descriptionof the program, see McCloskey, Badecker,Goodman-Schulman, & Aliminosa, 1994.) Anerror was assigned an interpretation if it could beexplained by one of the individual error types, orby a combination of two individual errors (e.g.,crisp ! criosy, interpreted as an o insertion plus ap-to-y substitution). Errors attributable to del-etions or insertions of any number of letters were

also interpreted as such (e.g., palace ! pa, inter-preted as a deletion of l, a, c, and e). If an errorcould not be interpreted in any of these ways,the program classified it as uninterpreted, ongrounds that it was too complex for a meaningfulinterpretation to be assigned.

For the nonlexical errors the program placed52% (77/149) in the uninterpreted category,reflecting the fact that CM’s spellings were oftensubstantially discrepant from the correct spellings(e.g., future ! furance, brisk ! brouches). Acrossthe 72 interpreted nonlexical errors, CM made59 letter substitutions (e.g., ruin ! roin), 22letter insertions (e.g., myth ! mytch), 16 letterdeletions (e.g., journal ! joural), and 9 lettertranspositions (e.g., fluid ! fliud). (The totaladds to more than 72 because some incorrectresponses included multiple errors, as innuisance ! nuestance.)

The pattern was similar for the lexical errors.Sixty-five per cent (32/49) were too complex tobe interpreted (e.g., bring ! better). Among theinterpreted errors, letter substitutions (18 errors)and insertions (8) were the most common errortypes, occurring either alone (e.g., head ! heat)or in combination (e.g., motel ! mother). Letterdeletions (3) and transpositions (2) were lesscommon.

Writing nonwords to dictationIn writing nonwords to dictation CM was only12% correct (4/34), a level of performance signifi-cantly below the 39% correct he achieved forwords, x2 (1, N ¼ 360) ¼ 8.88, p , .01. Nearlyall of the errors (27/29) were nonlexical(e.g.,/tibl/! trerer).

Written picture namingCM was 25% correct (13/51) in the writtenpicture naming task from the DysgraphiaBattery. On 16 trials he produced no response.Among the remaining 22 errors 8 were lexical(e.g., tent ! ticket) and 14 were nonlexical(e.g., orange ! origage). Almost all of the errors(20/22) were too complex for the error analysisprogram to interpret (e.g., flag ! hacket).



Typing to dictationThe Dysgraphia Battery includes oral spellingtasks, but CM’s spoken language production wastoo impaired for these tasks to be feasible, andtyping to dictation tasks were administered as analternative. A 92-word list was dictated twice (inseparate test sessions), and CM spelled each wordby typing on a computer keyboard. The list wasalso presented in three other sessions for writingto dictation. CM’s performance was extremelysimilar in the typing and writing tasks. He was61% correct (112/184) in typing and 59%(167/276) correct in writing, x2(1) , 1. Lexicalerrors made up 36% of the total errors in typing(e.g., diamond ! diameter), and 30% of theerrors in writing (e.g., pocket ! rock). The error-classifying program placed 35% of the typingerrors (e.g., perfume ! pofuer) and 39% of thewriting errors (e.g., perfume ! roefer) in theuninterpreted category, and among errorsthat were assigned an interpretation, letter sub-stitutions were the most common error type inboth typing (e.g., pool ! rool) and writing (e.g.,woman ! womar).

Copy transcodingIn direct copy transcoding CM copied stimuli thatremained in view, translating from upper case tolower case, and vice versa. For delayed copy trans-coding CM looked at the stimulus for as long ashe wished, after which it was covered and hewrote it from memory (again changing case). Inthe direct copy task he was 100% correct for bothwords (84/84) and nonwords (40/40). In delayedcopying, however, his performance was significantlyworse: 70% correct (59/84) for words and 58%(23/40) for nonwords, x2 (1, N ¼ 168) ¼ 27.06,p , .01, and x2 (1, N ¼ 80) ¼ 19.12, p , .01,respectively. Most of his errors (40/42) were non-lexical (e.g., rather ! rathen; talent ! talbet), andappeared similar to the errors observed in thewriting to dictation tasks.

Intact and impaired processes

CM performed well on tasks requiring compre-hension of spoken words, including the BDAE

auditory comprehension tasks, the PeabodyPicture Vocabulary Test, and the word–picturematching test. Further, he was able to repeatwords successfully in the context of writing todictation. In one writing-to-dictation task CMwas asked to repeat the stimulus word bothbefore and after writing it. On 115 of 120 trials(96%) both repetitions were correct (yet CMmisspelled 61 of the 120 words). These resultssuggest that CM was capable of generatingphonological and (in the case of words) semanticrepresentations from dictated stimuli, and hencethat his spelling deficit resulted from impairmentat some later stage(s) of the spelling process.

CM’s perfect performance on the direct copytranscoding task provides evidence that allo-graphic conversion and more peripheral grapho-motor processes were intact. Also, his spellingperformance—in accuracy and error types—wasremarkably similar in writing and typing to dicta-tion, suggesting that his spelling deficit affectedthe processing of abstract lexeme and/or gra-pheme representations prior to the applicationof processes that convert the abstract graphemerepresentations to the allographic or keystrokerepresentations required for writing and typing,respectively.

CM’s spelling errors were predominantlyphonologically implausible letter substitutions,insertions, deletions, and more complex mis-spellings. This error pattern, as well as the wordlength effect and CM’s impaired performancein delayed copy transcoding, suggest a deficitaffecting the processing of abstract graphemerepresentations at the level of the graphemicbuffer (i.e., the Graphemes level in Figures 1 and2). Below we offer a more specific characterisationof the deficit.

The marginal effect of word frequency, andCM’s substantial number of lexical errors (e.g.,danger ! dampen), suggest that the lexeme levelof representation may also be implicated in hisspelling impairment. Finally, CM’s very poornonword spelling, the phonological implausibilityof his spelling errors for words, and the findingthat phoneme–grapheme conversion probabilitydid not affect his spelling accuracy, suggest



severe impairment of sublexical phonology-to-orthography conversion processes.

Intruded letters and the letterpersistence effect

Most of CM’s erroneous spelling responsescontained what we will call intruded letters—that is, letters that did not match letters fromthe correct spelling of the stimulus word. Forexample, in the substitution error head ! heatthe t is an intruded letter, as is the i in the insertionerror spend ! speind. We also count as intrudedletters any “extra” occurrences in the response ofletters that appear in the correct spelling, such asthe second t in turkey ! turget.

Letters that occurred as intrusions in a responsewere often present in one or more of the severalimmediately preceding responses. Consider, forexample, the error value ! valod, in which theerroneous response contained two intrudedletters, o and d. Table 2 illustrates the trial onwhich this error occurred (labelled E) and thefive immediately preceding trials (E-1, E-2, . . . ,E-5). As can be seen from the table, the lettersintruded on trial E were present in many of theimmediately preceding responses.

This pattern in CM’s errors suggested that theletters occurring as intrusions in a response wereoften in some sense coming from precedingresponses. Perhaps, in particular, grapheme repre-sentations activated during production of aresponse sometimes persisted abnormally in anactivated state, rather than being deactivatedupon completion of the response. If this werethe case, then on any given trial one or more gra-pheme representations from previous trials mightbe abnormally active, leading potentially to letter

intrusions. For example, the o and d intrusions inthe value ! valod error may have occurredbecause o and d grapheme representations wereactivated on preceding trials, and remainedactive beyond the point at which the stimulusvalue was presented.1

The present study explored the apparent letterpersistence effect and its implications morethoroughly. In the following discussion we firstdemonstrate that the persistence effect is real byshowing that the intruded letters in CM’sresponses were present in preceding responses farmore often than expected by chance. Next, wepresent several results in support of the claimthat CM’s letter persistence effect arises at anabstract graphemic level of representation. Withthese initial conclusions as a foundation we thenuse the letter persistence effect to argue that thecognitive spelling system has an interactive archi-tecture that incorporates grapheme-to-lexemefeedback.

EXPERIMENTAL STUDY

The letter persistence effect was explored primar-ily through writing-to-dictation tasks. Over a

Table 2. CM’s error value ! valod and the stimuli andresponses from the five immediately preceding trials

Trial Stimulus Response

E-5 drama drama

E-4 avert ovent

E-3 provide provide

E-2 open opon

E-1 leopard leopord

E value valod

1 Cohen and Dehaene (1998) have suggested that perseverations may arise from normal persistence of activation, when a level of

representation receives weaker-than-normal input from other levels. From this perspective, CM’s letter intrusions might be inter-

preted by assuming not that grapheme activation persists to an abnormal extent, but rather that signals from the lexeme level to the

grapheme level are abnormally weak, with the result that grapheme activation persisting normally from prior responses has a greater-

than-normal influence on grapheme selection for the current response. However, the distinction between abnormally strong gra-

pheme persistence and abnormally weak input from the lexeme level is not relevant for our purposes in this paper. The relevant

assumption, made by both interpretations, is that persisting grapheme activation from prior responses is abnormally strong relative

to input from the lexeme level. For convenience, we will therefore continue to describe CM’s deficit as one of abnormally persisting

grapheme activation.



5-year period, during which CM’s performanceremained stable, he wrote 3797 words to dictation(including the words from the Dysgraphia Batteryand the other spelling tasks discussed above).Across the entire stimulus set his accuracy was55% (2101/3797).

On 9 of the 1696 error trials CM produced noresponse. A further 86 errors were whole-responseperseverations, in which CM’s response matchedan entire response or stimulus from earlier in thecurrent list (e.g., wagon ! watch, occurring 4trials after he had spelled watch correctly). Theremaining 1601 errors were nonperseverativemisspellings, and these errors made up thecorpus for subsequent analyses.2 Within thiscorpus, 538 (34%) of the errors were lexical (e.g.,chin ! chance) and 1063 (66%) were nonlexical(e.g., shatter ! stach).

The 1601 errors collectively included 3174intruded letters. Among the 1483 errorsthat included letter intrusions, 574 contained 1intruded letter (e.g., sort ! soft), 432 errors had2 intruded letters (e.g., between ! brethen), andthe remaining 477 had 3 or more intrusions(e.g., bulb ! bumber).

Above-chance letter persistence

To determine whether CM’s spelling errorsshowed a real letter persistence effect, we assessedwhether intruded letters were present in precedingresponses more often than expected by chance. Foreach occurrence of a letter intrusion we first scoredwhether the intruded letter was present in CM’sresponse on each of the five immediately preced-ing trials. (Errors occurring in the first five trialsof each test list were excluded from the analysis,because for these errors there were fewer than

five preceding trials.) For instance, in the case ofthe value ! valod error for the trial designatedE in Table 2, CM’s response on each of the fivepreceding trials (E-1 through E-5) was scoredfor the presence of an o and a d (the lettersintruded on trial E). On trial E-1(leopard ! leopord) and E-3 (provide ! provide)both o and d were present in the response,whereas on the remaining trials only one of thetwo intruded letters was present: o on trials E-2(open ! opon) and E-4 (avert ! ovent), and don trial E-5 (drama ! drama).

A computer program carried out this tabulationfor all of the letter intrusions, with the resultsshown in the upper curve of Figure 3. The .453proportion shown for trial E-1 reflects thefinding that 1342 of 2960 intruded letters(45.3%) were present in CM’s response on theimmediately preceding trial (E-1).3 Similarly,the .344 proportion shown for trial E-5 indicatesthat 34.4% of the intruded letters were presentin CM’s response on the trial five back fromthat on which the intrusion occurred. It isevident from the figure that the likelihood of anintruded letter being present in a precedingresponse decreased monotonically with thenumber of trials intervening between the letterintrusion and the preceding response.

The computer program also estimated thelikelihood of an intruded letter being present ina preceding response by chance. The rationalewas as follows: If the occurrence of a letter intru-sion is completely unrelated to the presence of theintruded letter in the immediately precedingresponses, then the intruded letter should be justas likely to be present in responses that do notimmediately precede the intrusion. For example,if the d intrusion on trial E in Table 2

2 Even though the letters in whole-word perseverative responses are by definition present in preceding responses, these letters

were excluded from tabulations of letter intrusions in letter persistence analyses, because the processes leading to whole-response

perseverations may not be the same as those leading to persistence of individual letters. This conservative decision avoids the possi-

bility of inflating estimates of the magnitude of the letter persistence effect by including errors that may not involve persistence of

individual letters.3 The number of intrusion errors in this analysis was 2960 rather than 3174 (the total number of intrusions) largely because errors

occurring in the first five trials of each test list were excluded, as described above. A small number of additional errors were excluded

because for these errors no control items were available for the analysis described below, which estimated the probability of an

intruded letter being present by chance in preceding responses.



(value ! valod ) had nothing to do with whetherds were present in the responses on trials E-1through E-5, then ds should be just as likely tobe present in responses that did not immediatelyprecede the valod response—for example, in theresponse on trial E-20 or Eþ20. Put anotherway, tabulating the likelihood of a d beingpresent in responses not proximal to trial Eprovides an estimate of the likelihood of a dbeing present by chance on the trials immediatelypreceding trial E.

The chance estimates were obtained by tabula-ting, for each letter intrusion, the presence andabsence of the intruded letter in responses ontrials not in the range E-5 through Eþ5. Forexample, given a 70-word list and an intrusionoccurring on trial 50, chance estimates wereobtained for the intrusion by consideringresponses on trials 1–44, and 56–70. We referto the trials used for generating chance estimatesas the control trials. (Trials E-5 through Eþ5were excluded on grounds that the presence ofthe intruded letter on these trials could havebeen nonaccidentally related to the occurrence ofthe letter intrusion. The presence of the intrudedletter in responses on trials E-1 through E-5

might have influenced the occurrence of the intru-sion; and the presence of the intruded letter ontrial E could have led to its occurrence in responseson trials Eþ1 through Eþ5.)

The specific procedure for generating chanceestimates may be illustrated by referring onceagain to the d-intrusion in value ! valod. Thelikelihood of a d being present by chance on trialE-1 was assessed by first identifying all control-trial responses of the same length as the responseon trial E-1. Given that CM’s response on theE-1 trial (leopord ) had 7 letters, all 7-lettercontrol responses were identified. The analysisprogram then randomly selected one of thesecontrol responses, and determined whether ornot it contained a d. The same operations werethen performed for trials E-2 through E-5. Forexample, the response on trial E-2 (opon) was4 letters in length, and so the analysis identi-fied all 4-letter control responses, randomlyselected one of these, and determined whether itcontained a d.

This procedure was carried out for all ofCM’s letter intrusions. Thus, for each of the2960 E-1 responses, a control response of thesame length was selected and checked forthe intruded letter; and similarly for each ofthe E-2, E-3, E-4, and E-5 responses. Foreach of the five sets of control responses the pro-portion that contained the intruded letter wascalculated. For example, on one run of theanalysis 924 of the 2960 control responses fortrial E-1 contained the intruded letter, for aproportion of .312. This value represents anestimate of the proportion of E-1 responsesthat would be expected by chance to containthe intruded letter.

The entire process of generating estimatedchance proportions for trials E-1 through E-5was carried out repeatedly. Because the controlresponses entering into the estimates were selectedrandomly from the control-trial responses withthe appropriate length, each repetition of theanalysis produced a new sample of controlresponses, and hence a new estimated chanceproportion. A total of 1,000,000 repetitions werecarried out, yielding a distribution of estimated

52.0

03.0

53.0

04.0

54.0

05.0

1-E2-E3-E4-E5-E

rorrE ot evitaleR lairT

Pro

porti

on o

f Res

pons

es C

onta

inin

g In

trude

d Le

tter lautcA

ecnahC yb detcepxE

Figure 3. The proportion of trials in which a letterintruded on trial E was present in the response on each ofthe five immediately preceding trials (E-1 through E-5),and the proportion expected by chance.



chance proportions for each of the trials E-1through E-5.4

The distribution for trial E-5 is presented as anexample in Figure 4. The mean of the distributionwas .305, and this value was taken as the pro-portion of E-5 trials expected by chance toinclude the intruded letter. The distribution alsoprovided a basis for assessing the statisticalreliability of the difference between the observedproportion of E-5 responses that contained theintruded letter (.344) and the proportion expectedby chance (.305). As the figure illustrates, none ofthe estimated chance proportions obtained in1,000,000 repetitions of the analysis was as highas the observed proportion of .344 (shown as thetall vertical bar on the far right side of thefigure), indicating that the probability of obtainingthe observed proportion by chance was less thanone in a million (p , .000001).

The estimated chance proportion (i.e., themean of the frequency distribution) for each ofthe trials E-1 through E-5 is shown in the lowercurve of Figure 3. For each of these trials theobserved proportion of responses containing theintruded letter was reliably higher than theproportion expected by chance (all p , .000001).Accordingly, we conclude that letters occurringas intrusions in CM’s responses were present inthe several immediately preceding responses atrates considerably greater than chance, and hencethat his spelling performance shows a letterpersistence effect.

In discussing the results of subsequent analyseswe report the magnitude of the persistence effectsin terms of the difference between observed andchance proportions, rather than presenting theobserved and chance values separately. Forexample, in the analysis of all intrusions, themagnitude of the persistence effect for trial E-1was calculated as .453 (observed proportion)minus .306 (proportion expected by chance),

yielding .147. Figure 5 plots the magnitude ofthe persistence effect for trials E-1 through E-5.

For some analyses we will report the magnitudeof the persistence effect as a single value, calcu-lated as the average of the effects for trials E-1through E-5. (For these analyses we generateda distribution of expected average chanceproportions, for use in significance testing.) Inthe analysis of all intrusions the average magnitudeof the persistence effect was .084 (p , .000001).This value reflects the fact that when resultswere averaged over trials E-1 through E-5, theobserved proportion of responses containingthe intruded letter was .384, whereas the averageproportion expected by chance was .300.

Our method for determining that CM’sintruded letters were present in precedingresponses more often than expected by chance isvery similar in rationale to a method describedby Cohen and Dehaene (1998) for evaluatingwhether perseverations occur at above-chancerates. However, our specific procedures forestimating rates expected by chance are somewhatdifferent from those of Cohen and Dehaene,the most notable difference being that our cri-teria for selection of control responses are morestringent and therefore may yield more accurateestimates of chance rates.

Functional locus of the letterpersistence effect

In this section we present several results andanalyses aimed at clarifying the level(s) of repre-sentation giving rise to CM’s letter persistenceeffect.

Orthographic or phonological persistenceIn the preceding discussion we have assumedthat the persisting representations were grapheme

4 The mean number of available control items for a source item was 20.3. On each repetition of the analysis for each source

position (E-1, E-2, . . .) a control item was sampled for each of the approximately 3000 intrusion errors. As a consequence the

number of possible control item samples was astronomically large: An average of approximately 20 options were available for

each of approximately 3000 choices. (An attempt to calculate the exact number of possible samples in Excel failed due to overflow

errors when the value reached approximately 4 � 10307.) Hence, in carrying out 1,000,000 repetitions of the analysis we were not

simply generating the same control-item samples repeatedly.



representations activated in the course of produ-cing responses. However, an alternative possibilityis that the persistence effect stems from abnormalpersistence of activation among phoneme rep-resentations that became active when stimuluswords were dictated. For example, the d-intrusionin the value ! valod error (see Table 2) may haveoccurred not because a representation of theletter d remained active after the productionof the response on the preceding trial(leopard ! leopord ), but rather because a repre-sentation of the phoneme /d/ remained abnor-mally active, leading eventually to activationof the grapheme d via lexical or sublexicalprocesses.

To explore this possibility we presented CMwith two lists for writing to dictation. Each listcontained 24 sets of three consecutive words. Ineach three-word set the first word was a prime

0

00001

00002

00003

00004

00005

00006

543.0043.0533.0033.0523.0023.0513.0013.0503.0003.0592.0092.0582.0082.0572.0Estimated Chance Proportion

Freq

uenc

y in

One

Mill

ion

Ana

lysi

s R

uns

Figure 4. Frequency distribution for proportion of E-5 trials expected by chance to contain the intruded letter. Vertical bar atfar right indicates the observed proportion.

50.0-

00.0

50.0

01.0

51.0

02.0

1-E2-E3-E4-E5-E


Effe

ct S

ize

Figure 5. Letter persistence effect (difference betweenactual and chance proportions of preceding responsescontaining the intruded letter) for each of the five trialspreceding the letter intrusion.



that contained the phoneme /f/, and the next twowords were targets that did not contain /f/. Inthe Prime-with-F list the /f/ in each primeword was spelled with the letter f (e.g., afraid ),whereas in the Prime-without-F list the /f/ wasspelled with other letters (e.g., sphere). Thetarget words were identical in both lists. Forexample, as illustrated in Table 3, the Prime-with-F list contained the prime-target-targetsequence afraid-direct-sudden, whereas the Prime-without-F list contained sphere-direct-sudden.Prime words were matched in frequency andlength across the two lists.

The question of interest was how often theletter f would occur as an intruded letter inCM’s responses to the target words. If the persist-ence effect were phonologically based, we wouldexpect f intrusions to occur equally often inthe Prime-with-F and Prime-without-F lists,because the prime words in both lists containedthe phoneme /f/. In contrast, if the persistenceeffect were due to persisting activation of gra-pheme representations, we would expect fintrusions to be more frequent in the Prime-with-F list than in the Prime-without-F list.

The results were unequivocal. For the Prime-with-F list CM’s response to 12 of the 48 targetwords contained the letter f, but for thePrime-without-F list CM produced no fintrusions in spelling the target words, x2

(1, N ¼ 96) ¼ 11.52, p , .001. Table 3 illustratesthis pattern of results, showing that CM intrudedfs into his spellings for the target words direct andsudden when the prime word was afraid, but notwhen the prime was sphere. We conclude thatCM’s persistence effect arises from abnormally

persisting activation of grapheme and notphoneme representations.

Stimulus- or response-based persistenceFurther support for the interpretation ofpersisting grapheme representations came fromanalyses carried out over the full set of letterintrusions. Two predictions of the interpretationwere tested: First, letters appearing in thecorrect spelling of a stimulus word but not inCM’s response should show little or no tendencyto intrude into subsequent responses. Forexample, in the error dial ! dian the correctspelling of the stimulus word contains theletter l, but this letter did not appear in CM’sresponse. If we assume that CM failed toproduce the l because an internal representationof this letter did not become sufficiently acti-vated, then we would not expect persisting acti-vation of the letter representation to be strongenough to cause l intrusions in subsequentresponses.

The second prediction was that lettersappearing in CM’s response but not in thecorrect spelling should intrude into subsequentresponses at above-chance rates. For example,the presence of the n in the dial ! dian responseimplies that a representation of the letter n becameactivated during the production of the response.Persisting activation of this letter representationcould then lead to n intrusions on subsequenttrials.

Note that on a persisting-phonemes interpre-tation we might expect the opposite pattern ofresults. Given that CM was able to generateaccurate phonological representations of dictatedstimulus words, the persisting phonemes shouldbe those present in the stimulus words, and asa consequence persistence effects should bestimulus-based.

Two analyses were carried out to evaluatethe contrasting predictions. The stimulus-not-response analysis tabulated, for each letter intru-sion, whether the intruded letter was presenton each of the five preceding trials among lettersthat appeared in the correct spelling of the

Table 3. Example of a trial sequence from the Prime-with-Fand Prime-without-F lists

Trial

Prime-with-F list Prime-without-F list

sequence Stimulus Response Stimulus Response

Prime afraid aflersuil sphere sphere

Target direct diffent direct diriction

Target sudden suffent sudden suddion



stimulus word but not in CM’s response. Considerthe following sequence of trials:


E-1 chap chorlonE thick thirk

In making the tabulations for the r intrusion ontrial E, the analysis considered for trial E-1whether an r was present in the set of lettersfa, pg. Chance probabilities were determined bycalculating the likelihood of the intruded letterbeing present on control trials among letterspresent in the correct spelling of the stimulusword but not in CM’s response.

The response-not-stimulus analysis was thesame, except that for trials E-1 through E-5,and the corresponding control trials, the lettersconsidered were those that appeared in CM’sresponse but not in the correct spelling of thestimulus word. Thus, for the r-intrusion in trialE above the analysis considered for trial E-1(chap ! chorlon) whether an r was present in theset of letters fo, r, l, ng.

The results, presented in Figure 6, were entirelyas predicted by the interpretation of persist-ing letter representations. The Stimulus-not-Response analysis revealed no hint of a letterpersistence effect. In the Response-not-Stimulusanalysis, however, the likelihood of an intrudedletter being present among letters that appearedin CM’s response but not in the correct spellingwas reliably greater than chance for each of thetrials E-1 through E-5 (all p , .0001).5

Visual feedback as the cause of thepersistence effect?We have assumed that the letter representationssubject to abnormally persisting activation were

localised within the cognitive spelling system, andwere activated by processes intrinsic to the pro-duction of spelling responses. In this section weconsider a potential alternative interpretation—that visual feedback occurring as CM looked atwhat he was writing activated visual (or moreabstract) letter representations specific to reading,and these representations persisted abnormally inan activated state. Perhaps, for example, whenCM looked at his written response leopord,reading-specific representations for d, o, and otherletters became activated, and remained activatedduring subsequent spelling trials (even thougheach written response was covered before thenext trial began). The reading-specific letter rep-resentations may then have directly or indirectly(e.g., via lexeme representations) activated letterrepresentations implicated in spelling, leading toletter intrusions on the subsequent trials (e.g.,value ! valod ).

To evaluate this possibility we presented CMwith two 70-word lists for writing to dictation

20.0-

00.0

20.0

40.0

60.0

1-E2-E3-E4-E5-E


Effe

ct S

ize

sulumitS-ton-esnopseR

esnopseR-ton-sulumitS

Figure 6. Letter persistence effect for the Response-not-Stimulus and Stimulus-not-Response analyses.

5 The magnitude of the persistence effect (i.e., the difference between observed and chance rates) was somewhat smaller in the

Response-not-Stimulus analysis than in the original analysis that considered all letters in CM’s responses. This result is exactly as

expected, given that some of CM’s letter intrusions were presumably caused by persisting activation of representations for letters that

were present on preceding trials in both the correct spelling and in CM’s response. For these intrusions the Response-not-Stimulus

analysis would not have found the intruded letter among letters present in the response but not in the stimulus on the preceding trial.

As a result, we would expect the persistence effect to be smaller in the Response-not-Stimulus analysis than in the main analysis

considering all letters in preceding responses.



under each of two conditions. In the eyes opencondition the normal writing-to-dictation pro-cedures were followed: CM was allowed to lookat what he was writing (and each response wascovered before the next stimulus was dictated).In the eyes closed condition, however, CM kepthis eyes closed throughout administration of thestimulus lists, so that he received no visual feed-back from his written responses. If the letterpersistence effect were due to activation ofreading-specific letter representations as CMlooked at what he was writing, then little or nopersistence effect should be observed in the eyesclosed condition.

CM’s spelling accuracy was comparable in thetwo conditions: 41% (57/140) in the eyes closedcondition, and 50% (70/140) in the eyes opencondition, x2 (1, N ¼ 280) ¼ 2.44, p . .05.Table 4 presents the magnitude of the letter per-sistence effect for trials E-1 through E-5 in eachof the two conditions. A highly reliable persistenceeffect was obtained in the eyes closed condition(p , .0001). If anything, the effect was somewhatlarger in this condition than in the eyes opencondition. We conclude that the letter persistenceeffect was not caused by visual feedback occurringas CM looked at what he was writing.

Persistence of abstract graphemic or moreperipheral representationsThe results presented thus far indicate that CM’spersistence effect arises from abnormally persistingactivation of letter representations in the cognitive

spelling system. In this section we considerwhether the persisting letter representations wereabstract grapheme representations—that is,abstract representations of letter identity—ormore peripheral representations, such as represen-tations of allograph shapes (e.g., the shape of alower-case cursive g) or even motor plans forproduction of writing movements.

The issue was explored with a 60-trial writing-to-dictation task in which CM wrote his responsein upper case (e.g., BEAN ) on odd-numberedtrials, and lower case (e.g., fat) on even-numberedtrials. The rationale was as follows: If the letterpersistence effect arises at an abstract graphemiclevel of representation, then intruded lettersshould be present at above-chance rates in theresponses on all of the immediately precedingtrials (E-1 through E-5), independent of whetherthe preceding response was written in the samecase as the intruded letter, or in the alter-native case. Consider the following sequence oftrials:


E-5 door DOORE-4 fad lageE-3 toil TOLEE-2 doom domeE-1 dim DIMOTE mob mot

For the lower-case t intrusion on trial E, we wouldexpect the letter t to be present with above-chancelikelihood not only in preceding lower-caseresponses (trials E-2 and E-4), but also in preced-ing upper-case responses (trials E-1, E-3, andE-5). Put another way, production of upper-caseTs as well as lower-case ts on trials in the E-1through E-5 range should influence the occur-rence of lower-case t intrusions on trial E.

On the other hand, if the persistence effectarises at a level of representation more peripheralthan that of abstract graphemes (e.g., a level ofshape-specific allograph representations) thenintruded letters should more often be present onpreceding trials with responses in the same case,

Table 4. Size of the letter persistence effect in the eyes openand eyes closed conditions

Condition

Trial Eyes closed Eyes open

E-5 –.044 –.016

E-4 .048 .013

E-3 .014 .061

E-2 .142 .006

E-1 .153 .080



than on trials with responses in the alternativecase. For the intruded lower-case t inmob ! mot, we would expect the letter t to bepresent with above-chance likelihood on preced-ing lower-case trials (E-2, E-4), but not (or to alesser extent) on preceding upper-case trials(E-1, E-3, E-5).

The results are presented in Table 5. As usual,the letter persistence effect was largest for trialE-1, even though the E-1 response (e.g.,DIMOT ) was written in a different case to theE response (e.g., mot). The persistence effect wasreliable for trial E-1 (p , .01), as well as for thethree different-case trials (E-1, E-3, E-5) takentogether (p , .01). Thus, letters written in onecase intruded at above-chance rates into sub-sequent responses written in the alternative case.Furthermore, the data do not give any indicationthat the persistence effect was diminished fordifferent-case trials relative to same-case trials.These results strongly suggest that the letterpersistence effect has its locus in a level of rep-resentation shared by upper- and lower-casewriting, and hence that the letter representationsgiving rise to the effect are abstract graphemerepresentations.

Another relevant finding is that CM showedthe letter persistence effect not only in writing,but also in typing: For the typing to dictationtask described above (see pp. 10-11) the meaneffect size across trials E-1 through E-5—differ-ence between observed and chance likelihoods ofan intruded letter being present in a precedingresponse—was .085 (p , .000001), virtually iden-tical to the effect of .084 observed in the analysisof CM’s writing to dictation performance. This

result suggests that the letter persistence effectarises at a level of letter representation sharedby writing and typing, and hence that thepersisting representations are abstract graphemerepresentations.

Having presented evidence that the letterpersistence effect is real, and that the effect arisesat an abstract graphemic level of representation,we next use the persistence effect as part of thebasis for arguing that the cognitive spellingsystem has an interactive architecture.

LEXICAL ERRORS, THEPERSISTENCE EFFECT,AND INTERACTIVITY

As we have noted, 34% of CM’s misspellings(538/1601) were lexical errors (e.g.,dignify ! define,arm ! amber, carpet ! compute). Such errorscould arise systematically, as a result of activatingthe wrong word representation at the lexemelevel. For example, the dignify ! define errormay have occurred when the incorrect lexemeDEFINE became activated, leading to activationof the grapheme sequence D-E-F-I-N-E at thegrapheme level. However, lexical errors couldalso arise by chance, for reasons having nothingto do with the fact that the incorrect spellinghappens to be a word—that is, without thelexeme representation of the incorrect wordplaying any role in the activation of the incorrectgraphemes. For example, the error bean ! beadmight have occurred when the grapheme D wasactivated instead of the grapheme N, withoutthe BEAD lexeme representation playing anyrole in the activation of that incorrect grapheme.We will refer to lexical errors resulting at leastin part from activation of an incorrect lexemerepresentation as true lexical errors, and to lexicalerrors occurring wholly for other reasons aschance lexical errors.

The distinction between true and chancelexical errors is important for an argument wedevelop in the following discussion. Specifically,we argue that CM’s lexical errors, in conjunctionwith the letter persistence effect, motivate the

Table 5. Letter persistence effect in the alternating-casewriting-to-dictation task

Trial Case relative to trial E Size of persistence effect

E-5 Different .011

E-4 Same .096

E-3 Different .069

E-2 Same .060

E-1 Different .138



postulation of grapheme-to-lexeme feedback inthe spelling system. The argument has threemajor steps:

1. At least some of CM’s lexical errors are truelexical errors.

2. The true lexical errors show the letter persist-ence effect.

3. The presence of the persistence effect in thetrue lexical errors implies grapheme-to-lexeme feedback.

Establishing the first two steps of the argu-ment, and demonstrating that the third stepthen follows, are far from trivial exercises, andwill require substantial discussion.

Step 1: Some observed lexical errors aretrue lexical errors

The first task is to demonstrate that at least someof CM’s lexical errors are true lexical errors, result-ing from activation of incorrect word represen-tations at the lexeme level. To establish thispoint we need to show that CM made lexicalerrors more often than expected by chance, sothat at least some of the errors must be truelexical errors.

How, though, can we estimate the rate atwhich lexical errors should occur by chance, inorder to show that CM’s rate of lexical errorsexceeds the chance rate? If we consider CM’sentire corpus of errors, the problem appearsintractable. Given that CM’s erroneous spellingswere often grossly different from the correctspellings (e.g., umbrella ! ublerrer, string ! slint,rope ! shop), it is entirely unclear how we couldestimate the rate at which his misspellingsshould happen by chance to come out as words.

Faced with this difficulty, we decided to limitour analyses to a subset of CM’s misspellings forwhich estimation of the chance rate of lexicalerrors appeared tractable (although not simple).In particular, we limited the analysis to errorstaking the form of letter substitutions. Morespecifically still, we considered those errors inwhich the incorrect spelling differed from thecorrect spelling only in that one or two letters of

the correct spelling were replaced by incorrectletters; examples are shown in Table 6. The aimof the analysis was to determine whether lexicalerrors occurred more often than expectedby chance among the one- and two-lettersubstitutions.

Restricting the analysis to one- and two-lettersubstitution errors is a very conservative strategy,because these errors are certainly among thosemost likely to come out as words by chance. Forexample, bell ! ball could presumably occur forreasons having nothing to do with the fact thatball is a word, far more readily thancarpet ! compute could occur for reasons unre-lated to the fact that compute is a word. In limitingour analysis to the simple substitution errors, weare therefore stacking the deck against ourselves,making it difficult to show that lexical errorsoccurred more often than expected by chance.(Romani et al., 2002, took a similar approach inassessing whether lexical errors occurred at anabove-chance rate in their patient DW.)

Among CM’s total of 1601 errors, 469 tookthe form of one- or two-letter substitutions(including both lexical errors such as bell ! ball,and nonlexical errors such as mate ! mape).Twenty of these errors were excluded from theanalysis because they occurred too early in astimulus list to be considered in persistenceeffect computations (required for a later step inthe argument for grapheme-to-lexeme feedback).Of the remaining 449 errors, 191 were lexicalerrors and 258 were nonlexical errors. The ques-tion we attempted to answer was whether theobserved proportion of lexical errors (.43) was

Table 6. Examples of CM’s one- and two-letter substi-tution errors

Lexical errors Nonlexical errors

Stimulus Response Stimulus Response

sold cold jeep jeek

square squire hint rint

log bag fled flug

blind blank water waton



reliably greater than the proportion expectedby chance.

The methods we developed for answering thisquestion were similar, although not identical, toprocedures previously used by other researchersfor assessing whether lexical errors occurredmore often than expected by chance in spokenproduction (e.g., Best, 1996; Martin et al., 1996)and writing (e.g., Romani et al., 2002). Our strat-egy was to consider what letter substitutions mighthave occurred by chance in each of the 449 errors,and then to determine from this set of candidatesubstitutions what proportion of lexical errorswould be expected by chance. For example, inthe error take ! make an m was substituted forthe initial t, yielding a lexical error. However,other letters could conceivably have been sub-stituted for the t, some of which would haveyielded lexical errors (e.g., bake), and others ofwhich would not (e.g., pake).

For the take ! make error we decided (throughmethods described below) that in addition tom—the letter that actually occurred as a substi-tution—the letters b, c, d, f, n, and p could alsopotentially have been substituted for the t, sothat the potential substitution errors were bake,cake, dake, fake, make, nake, and pake. Four ofthese seven potential errors are words; accordingly,we estimated the probability that a substitutionfor the letter t would produce a word by chanceas 4/7, or .57.

Obviously, the validity of such estimates hingeson the method for deciding what letters to con-sider as possible substitutions. Clearly it will notdo simply to assume that any letter of the alphabetcould have been substituted. For example, someletters (e.g., low-frequency letters, such as q or z)may be very unlikely to occur as substitutionerrors.

For each actual substitution error we decided toconsider as candidate letters for substitution anyletter that occurred in the responses to the fiveimmediately preceding stimuli, plus any letterthat occurred as a substitution error in theresponse under consideration.

Consider, for example, the error list ! lift, inwhich an f was substituted for the s. This error,

and the stimuli and responses from the fivepreceding trials, are shown below:


E-5 beef beefE-4 felt feltE-3 mood moofE-2 dirt dirtE-1 wall wallE list lift

The preceding responses contain the letters a, b, d,e, f, i, l, m, o, r, t, and w. These letters were taken asthe preliminary candidates for substitution errors.(The letter f would have been considered a candi-date substitution even if it had not occurred in thefive preceding responses, because it was the letteractually substituted.)

This method of selecting candidate sub-stitutions directly takes into account the possibilityof letter persistence at the grapheme level; theletters most likely to occur as persistence errorsare treated as candidates for substitution. Theselection of letters occurring in precedingresponses also takes into account letter frequencyin CM’s responses—letters that CM rarely pro-duced are unlikely to be treated as substitutioncandidates.

The selection method may not sample all ofthe letters that might have occurred by chance assubstitution errors. However, what is importantis not that all possible substitutions be considered,but rather that the candidates considered berepresentative of the full set with respect to thelikelihood of producing a word. As far as we cansee, the method of taking candidates from pre-vious responses is unlikely to introduce systematicbiases.

After generating the preliminary set ofcandidate letters for substitutions, we imposedan additional constraint, corresponding to aregularity apparent in CM’s errors. When CMmade substitution errors, he nearly always substi-tuted consonants for consonants, and vowels forvowels (e.g., s for c in clash ! slash, and o for uin drum ! drom). (This phenomenon has also



been observed in other dysgraphic patients; see,e.g., Caramazza & Miceli, 1990; McCloskeyet al., 1994.) Therefore, whatever the processesleading to CM’s substitution errors, these pro-cesses did not freely substitute vowels for conso-nants, or vice versa. Accordingly, we consideredas candidate letters for substitution only thosehaving the same consonant-vowel status as theletter CM actually substituted. For example, inthe list ! lift error, the letter CM substituted( f ) is a consonant, and therefore only consonantswere allowed in the candidate substitution set.Eliminating the vowels from the initial set fa, b,d, e, f, i, l, m, o, r, t, wg yields fb, d, f, l, m, r, tg.

After developing the candidate letter set, wegenerated candidate substitution errors by insert-ing each candidate letter into the word in placeof the letter CM actually substituted. Thus,for the list ! lift error, with candidate letter set{b, d, f, l, m, r, t} we generated the candidatesubstitution errors libt, lidt, lift, lilt, limt, lirt,and litt.

For errors in which two different letters weresubstituted, we developed a candidate letter setfor each substituted letter, and generated a candi-date substitution error for all combinations ofletter candidates. In the case of the errordesk ! task, for example, the set of candidateletters for substitution into the d’s position wasfb, c, g, h, n, r, s, tg whereas for the e’s positionthe candidate set was fa, i, o, ug. As a consequence,there were 32 candidate substitution errors, eachrepresenting 1 of the combinations of the 8 candi-dates for the d’s position and the 4 candidates forthe e’s position: bask, bisk, bosk, . . . , tosk, tusk.

Many of the candidate errors generated in thisway were orthographically unacceptable lettersequences, such as libt or tlash. However, CM’sactual errors, whether words or nonwords, werealmost always orthographically acceptable sequen-ces of letters, such as toble or dorn. This resultsuggested that orthographically unacceptable seq-uences should not be considered potential errors.

Accordingly, we eliminated such sequences fromthe lists of candidate errors.6 For the list ! lifterror, libt, lidt, limt, and litt were eliminated,leaving lift, lilt, lirt. Across the entire set of 449one- and two-letter substitution errors, 3431 ofthe 8303 candidate substitution errors (41%)were eliminated as orthographically unacceptable.

Imposing an orthographic acceptability con-straint is a conservative measure, in that the con-straint works against our attempts to demonstratethat CM’s proportion of lexical errors is greaterthan expected by chance. Orthographically unaccep-table letter sequences are, of course, quite unlikely tobe words, and eliminating these sequences thereforeincreases the proportion of words in the set ofpossible substitution errors.

The procedures described thus far generated,for each of CM’s substitution errors, a set oferrors (including the actual error) that couldplausibly have occurred via processes that wereentirely insensitive to the word vs. nonwordstatus of letter strings (such as processes wherebyincorrect grapheme representations become acti-vated by means other than signals from incorrectlexeme representations).

For each of CM’s substitution errors, we thencalculated the proportion of candidate errors thatwere English words. For the list ! lift error,two of the three potential errors—lift andlilt—were words, whereas the third—lirt—wasnot. The proportion of words in the candidateset was therefore .67.

The likelihood of obtaining a word by chancewas estimated in this way for each of the 449substitution errors made by CM, including notonly the lexical errors like list ! lift, but also thenonlexical errors like half ! halp. (For this lattererror, 2 of the 6 candidate errors—hall, halt—werewords, yielding .33 as the estimated probability ofobtaining a word by chance in a letter substitutionat word-final position.) Across all 449 substitutionerrors, the mean likelihood of obtaining a word bychance was .226.

6 Orthographic acceptability judgments were made by a computer program that assessed whether each candidate substitution

error could be parsed into a sequence of one or more orthographically acceptable syllables. Acceptable syllables were those with

an orthographically acceptable nucleus, and (optionally) an acceptable onset and coda.



Estimating the number of lexical errorsexpected by chanceUsing the chance probabilities, we conductedMonte Carlo simulations to estimate the totalnumber of lexical errors expected by chance inthe set of 449 substitution errors. The simulationprogram assumed that all of the errors occurredvia processes insensitive to the lexical status ofletter strings. On each run the program decidedprobabilistically, for each of the 449 errors,whether the error had come out by chance as aword or a nonword. Consider, for example, thehalf ! halp error. As described above, we esti-mated the probability of a word being producedby chance through a word-final letter substitutionto be .33. Accordingly, on each run of the simu-lation program this error had a .33 probability ofcoming out as a lexical error, and a .67 probabilityof coming out as a nonlexical error. (The determi-nation was made each time by generating arandom number between 0 and 1, and countingthe error as lexical if the generated number wasless than or equal to .33.)

On each run the simulation program carriedout this process for each of the 449 errors, andthen tallied the number of errors that came outby chance as lexical errors. Because of therandom element in each lexical/nonlexical deter-mination, the total number of simulated lexicalerrors was different on different runs of the simu-lation. By running the simulation multiple timeswe generated a distribution of numbers of lexicalerrors expected by chance. More specifically, weran the simulation one million times, and con-structed from these runs the frequency distri-bution presented in Figure 7. The mean of thedistribution was 101.7, indicating that whereasthe actual number of lexical errors was 191, themean number expected by chance in the set of449 substitutions was about 102. Ninety-fivepercent of the simulated lexical error frequenciesfell below 115, indicating that the probability ofobtaining 115 or more lexical errors by chancewas less than .05; 99% of the simulated valueswere below 120, indicating a probability of lessthan .01 of obtaining 120 or more lexical errors

0

00001

00002

00003

00004

00005

00006

591091581081571071561061551051541041531031521021511011501001590958085707

ecnahC yb detcepxE srorrE lacixeL fo rebmuN

Freq

uenc

y in

One

Mill

ion

Sim

ulat

ion

Run

s

Figure 7. Frequency distribution for number of lexical errors expected by chance among 449 one- and two-lettersubstitution errors.



by chance. The highest number of simulatedchance lexical errors obtained in one million runsof the simulation was 140, indicating that theobserved number of 191 had a probability (far)lower than one in a million of occurring by chance.

This result provides a basis for concluding thatamong CM’s one- and two-letter substitutionerrors, the proportion of lexical errors was reliablyhigher than expected by chance, and therefore thatat least some of the lexical errors were true lexicalerrors, resulting from activation of the wronglexical representation at the lexeme level.

Estimating the number of true lexical errorsWe can use the results of the preceding analysisnot only to conclude that the observed lexicalerrors include some true lexical errors, but alsoto estimate the number of true lexical errors. Itmight appear that we could simply take 191, thenumber of observed lexical errors, and subtract102, our best estimate of the number of lexicalerrors expected by chance, to yield 89 as the esti-mated number of true lexical errors. However,this procedure would be erroneous. The analysisthat yielded 102 as the expected number ofchance lexical errors was aimed at determininghow many lexical errors would be expected ifthe error corpus included no true lexical errors.Accordingly, the analysis assumed that all of the449 substitution errors were grapheme-levelerrors (i.e., errors not resulting from activation ofan incorrect lexeme representation), and hencethat each of the errors had an opportunity tocome out as a lexical error by chance.

For present purposes we need to estimate thenumber of chance lexical errors in light of theconclusion that some of the 449 substitutionerrors were true lexical errors. By definition, truelexical errors do not have the opportunity tocome out as lexical errors by chance, and this com-plication must be considered.

Fortunately, the problem can be overcome bytaking from the preceding analysis not an esti-mated number of chance lexical errors, but ratheran estimated percentage. The analysis indicatedthat if all 449 substitution errors had been

grapheme-level errors, then about 102, or 23%,would have come out as lexical errors by chance.

According to this estimate, grapheme-levelerrors should by chance produce lexical errorsabout 23% of the time, and nonlexical errorsabout 77% of the time. The observed number ofnonlexical errors—258—should therefore rep-resent about 77% of the total number ofgrapheme-level errors in the set of 449 substi-tutions. By means of simple algebra (258 ¼ .77g,where g is the total number of grapheme-levelerrors) we can estimate the total number ofgrapheme-level errors to be about 335, withabout 77 of these (335 total–258 nonlexical)being chance lexical errors.

Given that the total number of lexical errorswas 191, we can then estimate the number oftrue lexical errors as 191–77, or 114. Accordingto this estimate, about 60% of the observedlexical errors (114/191) were true lexical errors.

In these computations we used the meannumber of lexical errors expected by chance(102) to estimate the proportion of grapheme-level errors that should happen by chance to bewords. We can perform the same computationsusing the number of lexical errors at the 99thpercentile of the chance frequency distribution:120. From this number we obtain .27 (120/449)as the estimated proportion of grapheme errorsthat yield words by chance. This value can betaken as the highest proportion of lexical errorsthat could be reasonably expected to occur bychance in the set of 449 substitution errors.Using the .27 value, we estimate that the observednumber of nonlexical errors (258) represents about73% of the total number of grapheme level errors,which yields 353 as the estimated total number ofgrapheme-level errors, and 95 (353 – 258) as thenumber of chance lexical errors. The estimatednumber of true lexical errors then becomes 96(191 – 95), and the estimated proportion of truelexical errors in the set of observed lexical errorsis .50 (96/191). This value can be taken as theminimum proportion of true lexical errors in theset of observed lexical errors. Therefore, ouranalyses suggest that at the very least, 50% of theobserved lexical errors were true lexical errors,



and our best estimate is that about 60% of theobserved lexical errors were true lexical errors.

This discussion concludes the first step of ourthree-step argument: At least some of CM’slexical errors are true lexical errors.

Step 2: True lexical errors show thepersistence effect

In this section we argue that CM’s true lexicalerrors show the letter persistence effect. Thesolid line in Figure 8 shows the results of a persist-ence effect analysis applied to the 191 lexical errorswithin the set of 449 one- and two-letter substi-tution errors. As is evident from the figure, thelexical errors showed a strong persistence effect(p , .000001).

However, this result alone is not sufficient toestablish that true lexical errors exhibit the persist-ence effect. Although the arguments presented inthe previous section demonstrated that at leastsome of the observed lexical errors are truelexical errors, the set of observed lexical errorsmay (and almost certainly does) include somechance lexical errors as well. If the chance lexicalerrors show a persistence effect, then the full setof observed lexical errors might show the effect

even if there is no persistence effect among thetrue lexical errors.

To demonstrate that the true lexical errorsshow a persistence effect we must show that thepersistence effect for the set of observed lexicalerrors is too large to be due entirely to whateverchance lexical errors are included in the set.Think of the 449 substitution errors as a mix of(1) true lexical errors resulting from activation ofan incorrect word representation at the graphemelevel, and (2) grapheme-level errors in which oneor two incorrect graphemes were activated at thegrapheme level for reasons other than activationof the wrong lexeme representation. The observedlexical errors will then be a mixture of true lexicalerrors, and grapheme-level errors that happenedby chance to produce words (i.e., chance lexicalerrors). The observed nonlexical errors will thenbe the grapheme-level errors that happened bychance not to produce words. Presumably theletter persistence effect should be the same forgrapheme-level errors that did and did nothappen to produce words—that is, the persistenceeffect should be the same for the chance lexicalerrors and the nonlexical errors. Therefore, iftrue lexical errors do not show a persistenceeffect, the persistence effect for observed lexicalerrors should be weaker than the effect for theobserved non-lexical errors. For example, if theset of observed lexical errors is made up of halftrue lexical errors (which show no persistenceeffect) and half chance lexical errors (which showthe same persistence effect as the nonlexicalerrors), then the persistence effect for the observedlexical errors should be half as large as that for thenonlexical errors.

According to the arguments developed in theprevious section, our best estimate is that about60% of the observed lexical errors were truelexical errors, and the remaining 40% were chancelexical errors. Accordingly, if the true lexical errorsshow no persistence effect, we should expect themagnitude of the persistence effect for theobserved lexical errors to be only 40% of that forthe nonlexical errors. Our estimate of the maxi-mum proportion of observed lexical errors thatcould be chance lexical errors was 50%, so at the

50.0-

00.0

50.0

01.0

51.0

02.0

1-E2-E3-E4-E5-E


Effe

ct S

ize

srorrE lacixeL

srorrE lacixeL-noN

elbatubirttA tceffE mumixaMsrorrE lacixeL ecnahC ot

Figure 8. Letter persistence effect for the lexical and non-lexical one- and two-letter substitution errors. Also shownis maximum effect size expected for the lexical errors if thepersistence effect were due entirely to chance lexical errors.



very most the persistence effect for observed lexicalerrors should be half that for nonlexical errors.

As shown in Figure 8, however, the persistenceeffect was at least as large for the observed lexicalerrors (.094) as for the nonlexical errors (.083).The figure also shows the persistence effectexpected for the observed lexical errors if(1) half of these errors were chance lexical errors(the maximum proportion according to ourestimates), and (2) only the chance lexical errorscontributed to the persistence effect, with truelexical errors making no contribution. Obviously,the obtained persistence effect for lexical errorswas much too large to be attributed solely to thechance lexical errors. We conclude that CM’strue lexical errors show the letter persistence effect.

This conclusion constitutes the second step inour three-step argument:

1. At least some of CM’s lexical errors are truelexical errors.

2. The true lexical errors show the letter persist-ence effect.

3. The presence of the persistence effect inthe true lexical errors implies grapheme-to-lexeme feedback.

Despite stacking the deck against ourselves invarious ways—for example, by limiting the analy-sis to CM’s one- and two-letter substitutionerrors, surely among those most likely to havecome out as words by chance—we were able tomake a strong case for the first two steps of thethree-step argument. Nevertheless, the case wedeveloped rests upon a substantial number ofassumptions, and a complicated chain of reason-ing. In the following section we strengthen thecase by offering an independent motivation forsteps 1 and 2. Based on a different subset ofCM’s errors, the alternative foundation is lessformal, but equally convincing.

An alternative foundation for steps 1 and 2

Whereas one- and two-letter substitutions couldoccasionally have produced lexical errors bychance, many of CM’s lexical errors were muchmore complex (e.g., carpet ! compute, grass !

craft, feather ! fielder). These complex lexicalerrors are, we suggest, very unlikely to haveoccurred by chance—it seems extremely impro-bable that letter substitutions, deletions, insertions,and transpositions occurring at a post-lexical levelwithout the influence of incorrect lexemerepresentations would by coincidence producesuch errors. For example, to interpret thecarpet ! compute error as a chance lexical errorone would have to assume that three letter sub-stitutions (a ! o, r ! m, e ! u) and a letterinsertion (adding e at the end of the word) justhappened by coincidence to produce a letterstring corresponding to an English word.

We assert, therefore, that CM’s complex lexicalerrors—or, at the very least, the vast majority ofthese errors—are true lexical errors, reflecting acti-vation of incorrect lexeme representations.Therefore, if the complex lexical errors show arobust persistence effect, this result would consti-tute another piece of evidence that CM’s truelexical errors showed the persistence effect.

We defined complex lexical errors as those thecomputer program for classifying errors (see p. 282)considered too complex to interpret—that is, theerrors placed by the program in the uninterpretedcategory. One hundred and eighty-nine of CM’slexical errors (35%) fell into this category.Table 7 presents some examples.

Presumably, the vast majority of CM’s complexlexical errors are true lexical errors. Therefore, iftrue lexical errors do not show a letter persistenceeffect, little or no persistence effect should beseen in the complex lexical errors. In contrast, asubstantial persistence effect would make astrong case that true lexical errors show the effect.

Table 7. Examples of CM’scomplex lexical errors

Stimulus Response

shook floor

stem chasten

machine improve

bulk brush

grew glove

few feature



Figure 9 presents the results of the persistenceanalysis for the complex lexical errors. It isevident from the figure that these errors showeda robust persistence effect (p , .000001). Likethe results from the one- and two-letter substi-tution errors, the findings for the complex lexicalerrors argue strongly that CM’s true lexicalerrors show the letter persistence effect.

Step 3: The persistence effect for true lexicalerrors implies grapheme-to-lexeme feedback

The finding of a letter persistence effect in CM’strue lexical errors cannot be explained by apurely feedforward process in which activationflows from the lexeme to the grapheme level, butdoes not flow back from the grapheme to thelexeme level. To explain the result it is necessaryto postulate grapheme-to-lexeme feedback.Consider the following two-trial sequence:


E-1 bench benchE arm amber

Amber, the incorrect response on trial E, includestwo intruded letters (b and e), both of whichwere present in the immediately precedingresponse (bench).

We might try to explain the error by assumingthat only the correct lexeme—ARM—wasactivated, but that persisting activation of the Band E graphemes at the grapheme level led toproduction of amber. However, errors occurringin this way are chance lexical errors; what we aretrying to explain is the occurrence of the letterpersistence effect for true lexical errors, in whichthe wrong lexeme representation is activated.

Assume, then, that the arm ! amber error wasa true lexical error, resulting at least in part fromactivation of the AMBER representation at thelexeme level. What needs to be explained is why,given that the lexeme representation for an incor-rect word was activated, the word was one thatshared letters with the preceding response bench.Why, for example, was the AMBER lexeme acti-vated, and not, say, the lexeme ARMOR orARMY or RAMP? We cannot assume that theincorrect word (e.g., AMBER) shared letterswith preceding responses simply by chance—theletter persistence effect is the nonchance occurrencein a response of letters appearing in precedingresponses.

The question, then, is why incorrect lexemesfor words (e.g., amber) that systematically sharedletters with preceding responses (e.g., bench)came to be activated. In a purely feedforwardprocess, lexemes are activated from semanticand/or phonological information. Such a processcould lead to semantic or phonological errors.However, arm and amber are not semanticallyrelated; more generally, CM’s lexical errors rarelyshowed semantic similarity between stimulus andresponse. Phonological errors could lead to anapparent letter persistence effect—persisting acti-vation of phoneme representations from previoustrials could lead to activation of lexemes forwords sharing phonemes (and therefore letters)with previous responses, leading to a letter persis-tence effect. However, we have already shown thatCM’s letter persistence effect is not phonologicallybased (see pp. 287–289).

To account for the letter persistence effect intrue lexical errors, we need to assume that persist-ing activation of letter representations at thegrapheme level influences the activation of word

50.0-

00.0

50.0

01.0

51.0

02.0

1-E2-E3-E4-E5-E


Effe

ct S

ize

Figure 9. Letter persistence effect for the complex lexicalerrors.



representations at the lexeme level. For example,in the case of the arm ! amber error we need toassume that persisting activation of the B and Egraphemes (activated during generation of thepreceding response bench) led to some activationof the AMBER representation at the lexeme level,and that this activation contributed to the selectionof the wrong representation at the lexeme levelduring the processing of the stimulus arm.

The most straightforward way for activation atthe grapheme level to affect activation on thelexeme level is to assume that connectionsbetween lexeme and grapheme nodes are bidirec-tional, such that activation spreads not only“forward” from lexeme nodes to the correspondinggrapheme nodes, but also “backward” from gra-pheme nodes to lexeme nodes for words contain-ing the graphemes (see Figure 2b). According tothis hypothesis, activation of the BENCHlexeme activates the B, E, N, C, and H graphemenodes, and activation then feeds back from thesenodes to the lexemes for all words containingthese letters (including, among many others,AMBER). Then, when the stimulus “arm” isdictated, activated semantic and/or phonologicalnodes send activation to the ARM lexeme,which in turn activates A, R, and M nodes atthe grapheme level. These nodes in turn feedback activation to lexemes for words containingthe letters a, r, or m. By this means the activationof the AMBER lexeme is increased and thislexeme, which is also receiving activation fromthe persisting grapheme nodes B and E, maybecome more activated than the correct lexemeARM (which receives no support from thepersisting B, E, N, C, and H graphemes). Thiscomputational architecture, in which activationspreads forward and then backward over multipleiterations, is an instance of what has been referredto as an attractor system (e.g., Plaut & Shallice,1993).7 In such a system the interconnectedlevels constrain one another as the system settles

into a stable state. In the case of CM, graphemesthat remained active from preceding trials feedactivation back to the lexeme level, contributingto selection of an incorrect lexeme, and leadingin turn to the activation of that lexeme’s corre-sponding graphemes.

Two potential alternative interpretations meritbrief consideration. First, one might suggest thatthe letter persistence effect for true lexical errorsreflects external rather than internal grapheme-to-lexeme feedback. In particular, one mightimagine that visual feedback occurring when CMlooked at what he was writing (e.g., bench) activatedletter representations (e.g., b, e, etc.), which inturn activated lexeme representations for wordssharing those letters (e.g., amber). However, wehave shown that the letter persistence effect doesnot result from visual feedback (see pp. 290–291).

Second, it is conceivable that our results could beexplained by some form of attractor architecture thatdoes not involve feedback from a grapheme to alexeme level. This category of potential explanationscannot be ruled out as a class. However, specificproposals could be formulated and evaluated. Herewe briefly consider one example. It might besuggested that the persistence effect for true lexicalerrors resulted from connections among ortho-graphic lexeme nodes. Specifically, one might positexcitatory connections between lexemes based onorthographic similarity, such that activationspreads from an activated lexeme representation tothe representations for orthographically similarwords. For example, activation of BENCH wouldresult in some activation of the AMBER lexeme,due to the shared letters b and e. Activation ofARM on the following trial would also lead tosome activation of AMBER, which, with persistingactivation from the previous trial, might leadAMBER to become more activated than ARMitself. This interpretation could account for a letterpersistence effect in true lexical errors; however, theassumption of orthographically based excitatory

7 Although the notion of an attractor network is often associated with the assumption of distributed representations, this

association is by no means necessary. Systems with either local or distributed representations for lexemes and graphemes can

have an attractor architecture. Although our figures show lexemes and graphemes as individual nodes, our hypothesis is neutral

with respect to whether the representations are local or distributed.



connections among lexeme nodes is problematic forseveral reasons. Most notably, positing excitatorylexeme-to-lexeme connections is implausible inlight of processing considerations. At the ortho-graphic lexeme level the goal of processing is toarrive at a state in which the correct lexeme node isstrongly activated and all other lexeme nodesare inactivated (or at most weakly activated).Excitatory connections between lexeme nodeswould not contribute in any obvious way to thisgoal, and in fact would probably work against it bypromoting activation of incorrect lexemes. Indeed,computational models often assume inhibitory con-nections among the nodes in a set when the goal ofprocessing is to activate only one of the nodes. Inthese models the goal state is achieved through aniterative “winner-takes-all” process in which asingle node gradually emerges as the winner, inpart by inhibiting the competing nodes (e.g.,McClelland & Rumelhart, 1981). For example, thecomputational model of spelling proposed recentlyby Houghton and Zorzi (2003) assumes inhibitoryconnections and competitive interactions amongorthographic lexeme nodes (although this aspect ofthe model is not implemented in their simulations).

Whereas excitatory lexeme-to-lexeme connec-tions would probably be counterproductive, excita-tory grapheme-to-lexeme feedback could facilitateselection of the correct lexeme. Although feedbackconnections would send activation not only to thecorrect lexeme but also to lexemes for orthogra-phically similar words, the correct lexeme wouldreceive the strongest activation, contributing tothe “victory” of this lexeme in a competitive acti-vation process. (See the General Discussion forelaboration of this point.)

On the basis of the preceding evidence andarguments, we conclude that the letter persistenceeffect for CM’s true lexical errors constitutesevidence of grapheme-to-lexeme feedback in thespelling system.

GENERAL DISCUSSION

Through extensive analyses of the impairedspelling performance exhibited by patient CM we

have developed an argument for grapheme-to-lexeme feedback in the cognitive spelling system.Two features of CM’s spelling were central tothe argument. First, letters from prior spellingresponses intruded into subsequent responses atrates far greater than expected by chance. We pre-sented evidence that this letter persistence effectresulted from abnormal persistence of activationat the level of grapheme representations. Second,CM made many formal lexical errors (e.g.,arm ! amber). We demonstrated that a largeproportion of these errors were “true” lexicalerrors originating in lexical selection, rather than“chance” lexical errors that happened by chanceto take the form of words. Crucially, analysesrevealed that CM’s true lexical errors exhibitedthe letter persistence effect. This finding, weargued, can be understood only within an architec-ture in which activation from the grapheme levelfeeds back to the lexeme level, in this way exertingan influence on lexical selection.

Other evidence for grapheme-to-lexemefeedback?

As we noted in the Introduction, little attentionhas previously been directed toward questionsconcerning feedback in the cognitive spellingsystem. Folk, Rapp, and Goldrick (2002) dis-cussed several findings that were consistent withgrapheme-to-lexeme feedback, but concludedthat none constituted clear-cut evidence for feed-back. Feedback is also discussed, albeit very briefly,in a recent paper by Romani and colleagues(2002). These authors studied a dysgraphicpatient (DW) whose spelling errors consistedpredominantly of words formally related to thetarget (e.g., table ! cable). The principal aim ofthe study was to investigate alternative hypothesesregarding the source of formal lexical errors and,in this regard, Romani et al. concluded that DWsuffered from “a lexical impairment that makesselection between competing representationsdifficult (possibly because of reduced inhibition-increased activation)” (p. 329).

In interpreting their results Romani et al. alsoposited grapheme-to-lexeme feedback, suggesting



that “DW’s errors arise because of confusionsamong a cohort of lexical neighbors activatedtop-down from a phonological input and bottom-up from shared letters” (p. 325, emphasis added).However, feedback was not a focus of their inves-tigation, and the evidence presented was notstrong. Analysing 252 nonmorphological lexicalerrors, Romani and colleagues found that mostof the errors were both phonologically and ortho-graphically similar to the target word. However,29 of the errors were more phonologically thanorthographically similar to their targets, and 15were more orthographically than phonologicallysimilar. Romani et al. (2002) interpreted theformer subset of errors as evidence for top-downactivation of lexical representations from phono-logy, and the latter as evidence for bottom-upactivation of lexical representations via feedbackfrom letter representations.

However, these conclusions are not entirelywell founded. In particular, the fact that someerrors were more orthographically than phonolo-gically related to the target does not imply thatorthographic influences (such as grapheme-to-lexeme feedback) played any role in the genesisof the errors. The phonological neighbours of atarget word will vary in their orthographic simi-larity to the target, and occasionally the ortho-graphic similarity may just happen to be greaterthan the phonological similarity. For example,steak and steal are phonological neighbours byvirtue of sharing the first two phonemes, but actu-ally have greater orthographic than phonologicaloverlap. Therefore, even if lexical representationswere activated in a purely top-down fashionfrom phonology (and semantics), with no feedbackfrom grapheme representations, activation of thetarget’s phonological neighbours could lead toselection of a word with greater orthographicthan phonological overlap with the target. Thus,in DW’s error curbs ! rubs (cited by Romaniet al. as an example of greater orthographic thanphonological target-error overlap), rubs may havebeen activated solely because of its phonologicalsimilarity to curbs. Hence, the finding that DWoccasionally made errors of this sort does not con-stitute evidence for grapheme-to-lexeme feedback.

(See Folk et al., 2002, for a method of evaluatingorthographic and phonological influences onformal lexical errors.) In contrast, the evidencefrom CM specifically supports feedback bydemonstrating the influence of persisting gra-pheme activation on lexical selection.

Feedback and graphemic buffer deficits

From our assumptions about grapheme-to-lexemefeedback, and the role of the feedback in CM’slexical errors, a question arises about interpre-tation of results from other dysgraphic patients:Does our position imply that any patient with adeficit affecting the grapheme level should makesubstantial numbers of formal lexical errors?Consider in particular patients characterised ashaving graphemic buffer deficits, who frequentlyproduce letter substitution and insertion errors,presumably reflecting activation of incorrectletter representations at the grapheme level(e.g., Caramazza & Miceli, 1990; Caramazza,Miceli, Villa, & Romani, 1987). In positinggrapheme-to-lexeme feedback, are we committedto predicting that such patients should showhigh rates of formal lexical errors, due to feedbackfrom erroneously activated graphemes to incorrectlexemes? If so, the prediction represents a signi-ficant problem for the grapheme-to-lexeme feed-back hypothesis, because results from severalwell-studied graphemic buffer patients show verylow rates of lexical errors. For example, Caramazzaet al. (1987) reported that lexical errors accountedfor only 16 of patient LB’s 305 errors in spellingwords, and most if not all of these errors couldhave occurred by chance. (See also McCloskey,et al., 1994.)

Fortunately, the grapheme-to-lexeme feedbackhypothesis is not inconsistent with the resultsfrom graphemic buffer patients. In patients withgraphemic buffer deficits grapheme-level acti-vation may be abnormally weak—buffer patientsare typically assumed to have difficulty activating,or maintaining activation of, grapheme represen-tations. If grapheme-level activation were abnor-mally weak, then grapheme-to-lexeme feedbackwould also be weak, and might therefore have



little effect on selection of an orthographic lexeme.For CM, in contrast, we have assumed thatgrapheme-level activation is abnormally strong,because activation persists abnormally over time.

Another potentially relevant observation is thatfeedback from erroneously activated graphemesmay be more likely to cause errors in selecting ortho-graphic lexemes when there is some impairmentat the lexeme level, such that activation at thislevel is abnormally weak or noisy. Thus, patientswith selective graphemic buffer deficits may makefew lexical errors not only because of weak activationat the grapheme level, but also because processingat the lexical level is intact. (As we noted in anearlier section, CM may, in contrast, have mildimpairment at the orthographic lexeme level.)

Potential functions of feedback

Our arguments and evidence for grapheme-to-lexeme feedback lead naturally to questionsabout the computational functions of such feed-back. Generally speaking, feedback can serve tworelated functions. The first is to bring additionalconstraints to bear on a decision or process—inthis case the lexical selection process. The secondfunction is to strengthen, stabilise, or bias certainrepresentations or responses—in this case, thetarget lexeme and grapheme representations, andhence the correct spelling response.

The lexical selection process chooses lexemes toexpress meanings that an individual has in mind.Lexical selection should therefore be constrainedprimarily by semantic input. However, the lexicalselection process might also usefully be informedby information at subsequent representationallevels (see, e.g., Dell et al., 1997; Rapp &Goldrick, 2000). In the case of spelling, gra-pheme-to-lexeme feedback allows lexical selectionto be influenced by any factor that can exert aneffect on grapheme activation. For example, ifwe posit a functional architecture in which bothlexical and sublexical processes activate the sameset of graphemes (as in Figure 1; see alsoHoughton & Zorzi, 2003), grapheme activationwill be determined in part by the output of thesublexical phoneme–grapheme conversion system

(in tasks such as spelling to dictation). Throughgrapheme-to-lexeme feedback, lexical selectionwould then be influenced by the plausible spellingsof the sounds of the target word (Folk et al.,2002), providing a mechanism for the “sum-mation” of lexical and sublexical informationproposed by Hillis and Caramazza (1991a). Onemight wonder if the contribution of the sublexicalphoneme–grapheme conversion system would bedetrimental in the case of irregular words.Although this question would certainly benefitfrom computational analysis, it is important tonote (as pointed out by Houghton & Zorzi,2003) that the majority of phoneme–graphemecorrespondences, even in irregular words, arehighly predictable. As a result, input from thesublexical system should primarily serve tostrengthen a target word relative to its competitors(see Folk et al., 2002, for an empirical test ofthis claim).

In addition to providing multiple constraintson selection, feedback provides a means bywhich particular representations or responses canbe strengthened, stabilised, or favoured. Dell(1986) pointed out that cyclic feedforward andfeedback processing can create positive feedbackloops between lexical representations and theirsegments—that is, a lexeme activates the corres-ponding graphemes, which further activate thelexeme via feedback, increasing the feedforwardactivation sent to the graphemes, and so forth.Under normal circumstances, positive feedbackloops would favour the target lexeme and itscorresponding graphemes, stabilising the correctrepresentations and biasing the system towardthe correct response. (See Dell & Gordon, 2003,for computational evidence regarding some ofthe specific benefits of feedback.)

These suggestions about potential functionsof grapheme-to-lexeme feedback do not con-stitute, and are not put forward as, definitivelogical arguments for the necessity of feedbackin the cognitive spelling system. (If such argu-ments could be made, we would not needthe empirical evidence from CM.) Rather, ourpoint is simply that the conclusions aboutgrapheme-to-lexeme feedback we drew from



analyses of CM’s performance are not implausiblein light of potential computational functions offeedback.

Feedback as a mechanism for“summation” effects

We conclude by elaborating the above-mentionedpoint that feedback provides a mechanism forthe interaction or summation of lexical and sub-lexical information in lexical selection. A numberof researchers (e.g., Hillis & Caramazza, 1991a,1995; Hillis, Rapp, & Caramazza, 1999; Miceli,Benvegnu, Capasso, & Caramazza, 1997; Miceli,Capasso, & Caramazza, 1994) have proposedthat selection of a lexical representation isconstrained not only by activation from semanticrepresentations, but also by the output ofsublexical phonology-to-orthography conversionprocesses (in the case of spelling to dictation)and orthography-to-phonology conversion pro-cesses (in the case of reading). For example,Hillis and colleagues (1999) tested a dysgraphicpatient (RCM) in the first week after she suffereda stroke, and again 2 weeks later. Results from thefirst phase of testing pointed to deficits affectingthe orthographic output lexicon and the sublexicalphonology–orthography conversion process. Atthis time RCM exhibited high rates of semanticerrors in spelling to dictation. However, by thetime of the second testing phase RCM’s sublexicalprocessing had improved, and her rate of semanticerrors was far lower.

Hillis et al. (1999) interpreted this pattern ofresults as evidence for summation of lexical andsublexical information in lexical selection.According to their account, dictation of a stimulusword (e.g., “knife”) activated a lexical-semanticrepresentation, leading in turn to activation of theorthographic lexemes for multiple semantically-related words (e.g., knife, fork, spoon, plate).Because RCM’s orthographic output lexicon wasdamaged, a nontarget lexeme (e.g., fork) wassometimes more strongly activated than the targetlexeme (knife). In the first testing phase no addi-tional constraints were brought to bear on lexicalselection, and as a consequence semantic errors

were frequent (e.g., knife ! fork). However, bythe time of the second testing phase RCM’sorthography–phonology conversion process wasfunctioning to some extent. When a word was dic-tated, this sublexical process generated a potentialspelling (e.g., nife), and this spelling represen-tation somehow influenced the activation oflexeme representations in the orthographicoutput lexicon. Even when the sublexically gene-rated spelling was not entirely correct (as in thecase of nife), it usually favoured the target withinthe pool of lexemes activated from semanticinput (e.g., knife, fork, spoon, plate), therebyreducing the incidence of semantic errors.

Although these results and others have beenpresented in support of the summation hypothesis,specific mechanisms for summation have notusually been discussed. We suggest that feedbackprovides such a mechanism (see also Folk et al.,2002). In the case of spelling, grapheme-to-lexeme feedback constitutes a plausible mecha-nism for summation effects: Sublexical processesactivate grapheme representations, which feedactivation back to the orthographic lexemes, inthis way influencing lexical selection in favour ofthe target word. Similarly, phoneme-to-lexemefeedback could conceivably account forsummation effects in reading.

Manuscript received 24 November 2003

Revised manuscript received 14 January 2005

Revised manuscript accepted 25 January 2005

PrEview proof published online 28 June 2005

REFERENCES

Baars, B. J., Motley, J. T., & MacKay, D. (1975).Output editing for lexical status from artificiallyelicited slips of the tongue. Journal of Verbal

Learning and Verbal Behavior, 14, 382–391.Best, W. (1996). When racquets are baskets but baskets

are biscuits, where do the words come from? A singlecase study of formal paraphasic errors in aphasia.Cognitive Neuropsychology, 13, 443–480.

Caramazza, A., & Miceli, G. (1990). The structure ofgraphemic representations. Cognition, 37, 243–297.

Caramazza, A., Miceli, G., Villa, G., & Romani, C.(1987). The role of the graphemic buffer in spelling:



Evidence from a case of acquired dysgraphia.Cognition, 26, 59–85.

Cohen, L., & Dehaene, S. (1998). Competitionbetween past and present: Assessment and inter-pretation of verbal perseverations. Brain, 121,1641–1659.

Dell, G. S. (1986). A spreading-activation theory ofretrieval in sentence production. Psychological

Review, 93, 283–321.Dell, G. S., & Gordon, J. K. (2003). Neighbors in the

lexicon: Friends or foes? In N. O. Schiller &A. S. Meyer (Eds.), Phonetics and phonology in

language comprehension and production: Differences

and similarities. New York: Mouton de Gruyter.Dell, G. S., & Reich, P. A. (1980). Toward a unified

model of slips of the tongue. In V. A. Fromkin(Ed.), Errors in linguistic performance: Slips of the

tongue, ear, pen, and hand (pp. 473–486).New York: Academic Press.

Dell, G. S., & Reich, P. A. (1981). Stages in sentenceproduction: An analysis of speech error data.Journal of Verbal Learning and Verbal Behavior, 20,611–629.

Dell, G. S., Schwartz, M. F., Martin, N., Saffran, E. M.,& Gagnon, D. A. (1997). Lexical access in aphasicand nonaphasic speakers. Psychological Review, 104,801–838.

Dunn, L. M., & Dunn, L. M. (1981). Peabody PictureVocabulary Test-Revised. Circle Pines, MN:American Guidance Service.

Folk, J., Rapp, B., & Goldrick, M. (2002). Interactionof lexical and sublexical information in spelling:What’s the point? Cognitive Neuropsychology, 19,653–671.

Garrett, M. F. (1976). Syntactic processes in sentenceproduction. In R. J. Wales & E. Walker (Eds.),New approaches to language mechanisms (pp.231–255). Amsterdam: North Holland.

Goldrick, M., & Rapp, B. (2002). A restricted inter-action account (RIA) of spoken word production:The best of both worlds. Aphasiology, 16, 20–55.

Goodglass, H., & Kaplan, E. (1983). The assessment ofaphasia and related disorders. Philadelphia: Lea &Febiger.

Goodman, R. A., & Caramazza, A. (1985). The JohnsHopkins University Dysgraphia Battery. Baltimore,MD: Johns Hopkins University.

Graham, N. L., Patterson, K., & Hodges, J. R.(1997). Progressive dysgraphia: Co-occurrenceof central and peripheral impairments. Cognitive

Neuropsychology, 14, 975–1005.

Graham, N. L., Patterson, K., & Hodges, J. R. (2000).The impact of semantic memory impairment onspelling: Evidence from semantic dementia.Neuropsychologia, 38, 143–163.

Harley, T. A. (1984). A critique of top-down indepen-dent levels models of speech production: Evidencefrom non-plan-internal speech errors. Cognitive

Science, 8, 191–219.Hillis, A. E., & Caramazza, A. (1991a). Category-

specific naming and comprehension impairment: Adouble dissociation. Brain, 114, 2081–2094.

Hillis, A. E., & Caramazza, A. (1991b). Mechanismsfor accessing lexical representations for output:Evidence from a category-specific semantic deficit.Brain and Language, 40, 106–144.

Hillis, A. E., & Caramazza, A. (1995). Convergingevidence for the interaction of semantic and sub-lexical phonological information in accessinglexical representations for spoken output. CognitiveNeuropsychology, 12, 187–227.

Hillis, A. E., Rapp, B. C., & Caramazza, A. (1999).When a rose is a rose in speech but a tulip inwriting. Cortex, 35, 337–356.

Houghton, G., & Zorzi, M. (2003). Normal andimpaired spelling in a connectionist dual-routearchitecture. Cognitive Neuropsychology, 20,115–162.

Levelt, W. J. M., Roelofs, A., & Meyer, A. S. (1999). Atheory of lexical access in speech production.Behavioral and Brain Sciences, 22, 1–75.

Martin, N., Dell, G. S., Saffran, E. M., & Schwartz, M.(1994). Origins of paraphasias in deepdysphasia: Testing the consequences of a decayimpairment to an interactive spreading activationmodel of lexical retrieval. Brain and Language, 47,609–660.

Martin, N., Gagnon, D. A., Schwartz, M. F., Dell, G. S.,& Saffran, E. M. (1996). Phonological facilitation ofsemantic errors in normal and aphasic speakers.Language and Cognitive Processes, 11, 257–282.

McClelland, J. L., & Rumelhart, D. E. (1981). Aninteractive activation model of context effects inletter perception: Part 1. An account of basic find-ings. Psychological Review, 86, 287–330.

McCloskey, M., Badecker, W., Goodman-Schulman, R.,& Aliminosa, D. (1994). The structure of graphemicrepresentations in spelling: Evidence from a case ofacquired dysgraphia. Cognitive Neuropsychology, 11,341–392.

Miceli, G., Benvegnu, B., Capasso, R., & Caramazza, A.(1997). The independence of phonological and



orthographic lexical forms: Evidence from aphasia.Cognitive Neuropsychology, 14, 35–69.

Miceli, G., Capasso, R., & Caramazza, A. (1994). Theinteraction of lexical and sublexical processes inreading, writing, and repetition. Neuropsychologia,32, 317–333.

Patterson, K. (1986). Lexical but nonsemantic spelling?Cognitive Neuropsychology, 18, 729–748.

Plaut, D. C., & Shallice, T. (1993). Deep dyslexia: Acase study of connectionist neuropsychology.Cognitive Neuropsychology, 10, 377–500.

Rapp, B., & Goldrick, M. (2000). Discreteness andinteractivity in spoken word production.Psychological Review, 107, 460–499.

Rapp, B., & Goldrick, M. (2003). Feedback by anyother name is still interactivity: A reply to Roelofs’comment on Rapp and Goldrick (2000).Psychological Review.

Roeltgen, D. P., Rothi, L. G., & Heilman, K. M.(1986). Linguistic semantic agraphia: A dissociationof the lexical spelling system from semantics. Brainand Language, 27, 257–280.

Romani, C., Olson, A., Ward, J., & Ercolani, M. G.(2002). Formal lexical paragraphias in a singlecase study: How “masterpiece” can become “mister-pieman” and “curiosity” “suretoy”. Brain and

Language, 83, 300–334.Stemberger, J. P. (1985). An interactive activation model

of language production. In A. W. Ellis (Ed.), Progressin the psychology of language (pp. 143–186). Hove,UK: Lawrence Erlbaum Associates Ltd.

Tainturier, M.-J., & Rapp, B. (2001). The spellingprocess. In B. Rapp (Ed.), The handbook of cognitiveneuropsychology: What deficits reveal about the human

mind (pp. 263–289). Philadelphia: PsychologyPress.



Documents

Grapheme-to-lexeme feedback in the spelling …web.jhu.edu/cogsci/templates/images/rapp/mccloskeymacarusorapp2006.pdfGrapheme-to-lexeme feedback in the spelling system: Evidence from