Errors in Proofreading Evidence for the Ude of Word Shape in Word Recognition

8/3/2019 Errors in Proofreading Evidence for the Ude of Word Shape in Word Recognition

1/8

Memory& Cognition1983, Vol.ll(1), 16-23Errors inproofreading: Evidence for theuse ofword shape inword recognition

ANDREW F. MONK and CHARLES HULMEUniversity 01York, Heslington, York YOl 5DD, England

Subjects were given a passage to read for meaning while at the same time checking offany misspellings. Criticalwords in the text were mutilated by either deleting or substitutingletters. In half of these mutilations, an ascender in the word was removed, resulting in alarge change in the overall shape of the word. In the first experiment, it was found thatmutilations involving the substitution of letters were more likely to be noticed than mutilations involving deletions. For both types of mutilation, alterations to word shape were moreoften noticed than alterations that preserved word shape. When word shape cues were eliminated by using mixed-ease stimuli in Experiment 2, the shape effect was abolished. Theresults areinterpretedas evidence fora supraletter feature correspondingto word shape.Since the last century, it has been known that skilled

readers may identify words more efficiently than nonwords or individualletters (e.g., Cattell, 1886). Possiblythe best known demonstration of such an effect is thatby Reicher (1969). Here the subject is presented with aword, nonword, or letter followed by a pattern maskand is given a forced choice between two letters thatcould have occurred in a specified position. Even whenthe letters are selected so that in the word conditionboth would spell words, performance is reliably betterin the word than in the nonword or letter conditions. Anadvantage of words over nonwords is also found instudies of simultaneous matehing in which responselatency is the dependent measure (Eichelman, 1970).These various "word-superiority effects" (for a review,see Baron, 1978) have often been interpreted asmeaningthat word recognition involves the use of word shapeinformation rather than simply depending upon therecognition of the individual letters ofwhich the word iscomposed (e.g., McClelland, 1977). According to such aview, features associated with groups of letters (supraletter features) such as the overall shape of the wordmay be extracted prior to, or in parallel with, information about its component letters. Information aboutsupraletter features may then facilitate the processingof letter identity information.

A more parsimonious view might be to hold thatword recognition depends entirely upon informationabout the identity of its component letters. Sophisticated guessing models assume that supraletter featuresare not extracted duringword recognition. For example,Massaro (1973, 1975) has argued that many existingdemonstrations of word-superiority effects can beaccounted for in terms of a failure to control for redundancy operating to facilitate the early stages of visualThe first author's research is supported by agrant from theMedical Research Council.

16

information processing. Letters in words are simplymore predictable than letters in nonwords or out ofcontext, and so partial information may be used tocorrectly identify letters in words when time or information is limited. An alternative explanation for wordsuperiority effects that also does not necessitate the ideaof supraletter features is in terms of information loss.Information from mechanisms responsible for wordidentification may simply be more durable and lesssusceptible to disruption than information from letterrecognition processes (Johnston & McClelland, 1973,1980).

Aresolution of the possible importance of supraletter features is vital to the development of adequatemodels of word recognition. The most common formof supraletter feature that has been considered is wordshape, which may be defined as the pattern of ascendersand descenders in a lowercase representation of theword. In studies of word-superiority effects, a fewattempts have been made to manipulate word shape,with mixed results. Ifword shape is a supraletter feature,then it will be of no use to nonwords, andhence, destroying word shape should affect word stimuli more thannonword stimuli. Such a manipulation can be achievedby printing alternate letters in uppercase; the results ofusing these mixed-case stimuli are then comparedwith the results obtained using all lowercase stimuli.McClelland (1976) found no evidence of the predictedinteraction in an experiment in which subjects had toidentify letters presented in words as opposed to nonwords. Adams (1979) obtained a similar result using atachistoscopic recognition paradigm. However, Bruder(1978), using a simultaneous matehing task, did obtainsuch an interaction, although it was not apparent in allher experiments. Bruder claims that her effect dependson word frequency and word length.There are, however, various sorts of evidence fromother experimental tasks that are consistent with readers'

Copyright 1983 Psychonomic Society, Inc.


2/8

PROOFREADING 17

EXPERIMENT 1

Table 1Mean Pronounceability Rating (and Standard Deviation)for the Four Different Misspelling Conditions(Maximum 4)

letters. The letter changed might be an ascender, inwhich case the result was a large change in word shape(so, for example, "Iatest" becomes "lacest" in thesubstitution condition or "laest" in the deletion condition). In other cases,when the letter changed wasnot anascender or descender, the word shape of the misspellingis reasonably similar to the original word (e.g., "latest"becomes "latect" or "latet"). One might expect changesthat result in a largedifference in word shape to be morenoticeable than those that do not. In the substitutionconditions, this could possibly be explained as a letterlevel effect (e.g., changing t to c ismore noticeable thanchanging s to c), but in the deletion conditions, this isnot so. Here there is an equivalent degree of sirnilarityin terms of letter identity information to the originalword, as all the letters in the misspelling "belong" inthe word. Any effects of shape change found in thedeletion conditions cannot be attributed to greatersimilarity at the letter level.

MethodSubjects. The subjects in the proofreading experiment were60 undergraduates . In addit ion, there were another 17 undergraduates who ra ted the misspellings for pronounceability.Materials and Design. The passage used in this exper imentwas taken from a wornen' s magazine. It was 2,093 words longand dea lt with the subject of making stained glass windows.Initially, 60 base words were selected and used to generate240 misspellings. These misspellings were then given to a groupof subjects to rate on a 4-point scale from 4 (very easy to pronounce) to 1 (very difficult). Using these ratings, 40 base wordswere selected, avoiding base words whose delet ion condi tionmisspellings were judged to be very much more pronounceablethan their subst itut ion condition misspellings. The resul t is aset of misspellings balanced for pronounceability (see Table 1).Each base word had two critica1 let ters , one of each type,which were consonants and which were not the initial or lastletter in the word. The first type was an ascender, and thesecond type was neither an ascender nor a descender. The nonascenders were selected from the set c, m, n, r, s, v, w; theascenders were from the set b, d, f, h, k, I, t (in I7 of the 40base words used in the proofreading experiment, the nonascendercame before the other critica1 l ett er in the word). The basewords were selected to contain suitable critical letters, asdescribed above, as well as to satisfy the following constraints:They should be between four and eight let ters long, they shouldnot be names, proper nouns, or the first word in a sentence,and they should not be in the same senten ce as another baseword.

.66.59SD

2.972.99

DifferentMean

Shape

SD.47.97

Same

3.142.87

MeanDeletionSubstitution

utilizing information about word shape. Rayner (1975)conducted aseries of elegant experiments in whichsubjects read prose passages presented on a CRT displaywhile their eye movements were monitored. In thissituation, alterations to specified words in the text canbe made just as the subject is moving his eyes to fixatethem. In all cases, the word is presented in its correctform when fixated, but while it is in peripheral visionit can be mutilated in a variety of ways. Substitutingletters of the same relative height, which preserves theoverall shape of the word, produces less disruption ofreading performance, as indicated by fixation duration,than substituting letters of different height, whichchanges the shape of the word. Rayner and his co1-leagues have now replicated this effect using a varietyof experimental techniques (Ehrlich & Rayner, 1981;Rayner, 1978; Rayner, McConkie, & Ehrlich, 1980).In studies of the above type, changes in word shapeare produced by changing letters, and thus, the identityof the word itself is changed. It is possible by typographical manipulations to alter word shape whilepreserving word and letter identity. Using uppercaseprint simply removesword shape information, and it hasbeen suggested that this is responsible for the reliabledecreases in reading speed found for uppercase material(Tinker, 1965). Using case-alternating script actuallydistorts the shape of a word, and again, this slows reading down (Fisher, 1975). F. Smith (1969) makes theimportant point that both these manipulations confounda change in the available discriminable features (shapechanges)with a change in familiarity. He ran an experiment that suggests that it is the change in availableinformation that is crucial for the effects of case alternation. By using different-sized type faces, he constructed alternating-case passages that maintain appropriate height information. He found that reading timeon such passages was normal, whereas alternating-casepassages composed of a single type face were readsignificantly more slowly than normal text.These experiments, among others, indicate thatchanges in word shape do affect recognition, but theymay be explicable without recourse to explanationsinvolving word shape or other supraletter features. Asimpler explanation in terms of letter-level effects maybe possible. When words are mutilated by substitutingletters, as in Rayner's experiments, the degree of similarity of the mutilated word to the originalisconfoundedwith the degree of similarity of the changed letter to itsoriginal. Bouma (1971) has shown that the relativeheight of letters is an important cue to their identification. Similarly, removing relative height informationby using alternating-case texts may have its effect at theletter level.In an attempt to overcome these problems in demonstrating the importance of word shape, we decided touse a proofreading task in which critical words in thetext are mutilated by either deleting or substituting


3/8

18 MONKANDHULME

Table 2Detransformed MeanArcsine Hit Rates for the FourMisspellingConditions in Experiment I

Results and DiscussionThe mean reading time was 470 sec which representsan average reading rate of 267 words/min, This iswithinthe normal range of reading speeds and indicates thatthe addition of the proofreading task has not dramaticaily changed the way the readers process the text.Similarly, the mean number of comprehension questionsanswered correctly was 4.1 out of 6. This indicates thatsubjects understood the passage adequately, given thateach question had three alternative answers, and on thebasis of guessing, subjects would be expected to answertwo of the six questions correctly.The detection of misspellings was evaluated in thefollowing way. An analysis ofvariance was performed onthe proportion of hits scored by each subject in each ofthe misspelling conditions (proportions out of 10).Before the analysis was performed, these proport ionswere transformed using the arcsine transform recommended by Winer (1971), and consequently, the resultsreported in Table 2 are detransformed mean arcsineproportions. The four misspelling conditions are gener-

Deleting the critical letter that was a nonascender resultedin the same-shape deletion condition misspelling. Deleting thecritical le tter that was an ascender resulted in the differentshape deletion condition misspelling. In the substitution conditions a letter was chosen from the set c, m, n, r, s, v, w, a, e, 0 ,u and substituted for the criticalletter, creating the substitutionsame- and different-shape conditions in an analogous fashion.The substituted letters were selected to maximize the pronounceability of the misspellings. In no case did any of themisspellings generated constitute an English word. The complete set of misspellings used in the proofreading experiment ispresented in the appendix.Using these materials, four versions of the text were prepared. There were 10 exemplars of each of the four types ofmisspelling in each version. This assignment of misspellings toversions was random, with the further constraint that each of the160 misspellings occur only once in the complete set of fourversions. Fifteen subjects were assignedto each of four groups atrandom, and each group had one of the different versions of thepassage. The resulting design thus confounds group with theMaterials (base word) by Conditions interaction.Procedure. All the subjects were tested simultaneously in asingle session. They were told to read the passage, which wasprinted with conventional capitalization, for meaning as r a ~ i d l Yas possible and that they would be asked to answer some s t r a I ~ h t -forward comprehension questions at the end. Each subjectrecorded the time at which he finished reading and then went onto answer six multiple-choice comprehension questions. Theproofreading task was introduced as a subsidiary task. Subjectswere to circle any misspellings they noticed while they werereading.

DeletionSubstitutionSame.554.759

ShapeDifferent.615.818

ated by two two-level variables: shape (same or different) and type of mutilation (deletion or substitution).Version (represented by groups) is a random materialsvariable, and so the appropriate error term to evaluatethe shape and type of mutilat ion main effects are theShape by Version interaction and the Type of Mutilat ion by Version interaction, respectively. Similarly,the appropriate error term for evaluating the Shape byType of Mutilation interaction is the Shape by Type ofMutilation by Version interaction.The most obvious effect in the means presented inTable 2 is type of mutilation. Deletions were considerably less likely to be noticed than substitutions [F(1,3)=37.841, p< .01]. The shape main effect was alsosignificant [F(1 ,3) = 11.284, P < .05] , bu t there was no

Shape by Type of Mutilation interaction [F(1 ,3) < 1,n.s.]. It should be noted that the analysis performedtakes account of variability across subjects and materials,and because of the small number of degrees of freedom,provides a conservative test of significance. An analysisthat takes account of subject variance only, that is, onethat takes version to be a fixed effect, gives the samepattern of results [mutilation type: F(1,56) =93.158,p< .01; shape effect: F(1 ,56) = 11.695, p< .01; interaction: F(1,56) < 1, n.s.] .This fmding, that misspellings that maintain word

shape are less noticeable than those that do not and thatthis shape effect is just as strang in the deletion conditions as it is in the subst itut ion conditions, cannot beexplained as a letter-level effect. In the deletion conditions, there is no substituted foreign letter that may bemore or less confusable with the original let ter; all theletters in the deletion condition misspellings belong inthe target word. These results demonstrate that, in thesecircumstances at least, word shape has an effect onreading performance and that the effect is a wordlevel effect; that is to say, word shape has been identified as a supraletter feature.The fmding of a shape effect in praofreading isconsistent with the results of Haber and Schindler(1981). Haber and Schindler, however, considered onlysubstitutions, and hence, their results might be explicable in terms of letter-level effects. Asmentioned in theintraduction, previous studies that have attempted todetermine whether word shape is a letter-level or a wordlevel effect have come to differing conclusions. Bruder(1978) has suggested that a possible reason for thesediscrepancies is that the effects may depend on frequency and word length (see also Healy, 1980).In the light of these ideas, correlations were computed between the logarithm of the base word's Kueraand Francis (1967) frequency and the shape effect foreach base word. The lat ter was computed as the difference between the transformed hit rate in the differentshape condition and the transformed hit rate in the sameshape condition. There was no evidence that the shapeeffect depends on word frequency or length (seeTable 3).


4/8

Table 3Correlations With Properties of the BaseWordsWord NumberFrequency Length Ascenders

Deletion ConditionsShape Effect .04 .01 .05

Substitution ConditionsShape Effect .11 .06 .11Total Hits -.51* .07 -.10

Note-N = 40 in oll cases. Hit rates were transformed using thearcsine transformation recommended by Winer (1971). Wordfrequency was transfomed by calculating its logarithm.*p < .05.Also, given that we obtained a shape effect with ourstimuli, it is difficult to argue that our stimuli are sornehow atypical. Each of the studies mentioned above useda different experimental task, and thus each differedfrom the others in a number of ways. Our task is different again, and arguably much more realistic than any ofthe others. It is not clear what the critical differencesare between our task, noticing spelling errors whilereading for meaning, and these other experimentaltasks. It must be admitted, however, that our failure tofmd an effect of frequency or length on the size of theword shape effect is not conclusive. Our base wordsinclude no very high-frequency words such as commonfunction words and only one word shorter than fiveletters long. It could weIl be that any effects of thesevariables are not linear, and it would be desirable infuture studies to consider larger rangesof frequency andlength (Haber & Schindler, 1981, present some datarelevant to this point).The other fmding, that misspellings created bydeleting one letter are noticed 1ess often than misspellings created by substituting one letter, indicates that,unlike word shape, word length is not an importantfeature of a word. In some ways, this result is notsurprising. The visual system is much more effective indetecting patterns such as word shape than detectingabsolute quantities such as word length. Furthermore,with proportionally spaced type, it may be difficult toestimate the number of letters in a word from its physical length (our stimuli were not proportionally spaced,but most print is). Nevertheless, this fmding has certainimplications for models of word recognition, which wewill consider later.Table 3 presents a number of correlations computedbetween hit rates and properties of the base words usedto create misspellings. Correlations were computed withthe shape effect, as defmed above, and with the total(transformed) hit rate for that base word. It is notablethat none of the corre1ations with the shape effectwas significant. Number of ascenders was included to seeif the shape of the word prior to mutilation was a criticalvariable. It is possible that deleting the only letter withan ascender in a word will result in a more dramatic

PROOFREADING 19change than deleting one of two or three letters withascenders. There were 14 one-ascender words, 14 twoascender words, and 12 three-ascender words. Neitherthe correlations nor a subsequent examination of meansshowed any effect of this variable. Turning to the correlations with total number of hits, there is a strongnegative correlation with frequency. Errors in highfrequency words are harder to spot than errors in lowfrequency words. This result is consistent with a varietyof models, such as Morton's (1970) logogen model, inwhich word frequency has the effect of reducing thecriterial evidence necessary forword recognition.The correlations computed above used the baseword as their sampling unit (N = 40). To look at theeffects of pronounceability and orthographic regularity,correlations were computed using the misspellings assampling units. One word, "vocation," has eight letters,and so it was not possible to compute the regularity datawith the norms used in the substitution conditions.Hence, N =158 for all the following correlations.Number of hits showed no correlation with pronounceability (r = -.070, n.s.). Likewise, there was no evidenceof a correlation with mean single-letter positionalfrequency (r = -.083, n.s.) (Mayzner & Tresselt, 1965),mean digram positional frequency (r = .012, n.s.), ormean trigram positional frequency (r = .053, n.s.)(Mayzner, Tresselt, & Wolin, 1965). These correlationsprovide no evidence for effects of pronounceability ororthographic regularity on this task. While Corcoran(1966) claimed to show effects of pronounciation in aletter-cancellation task, this interpretation has sincebeen rejected (Frith, 1979; Healy, 1976; P. T. Smith &Groat, 1979). The lack of any effect of orthographicregularity is consistent with the fmdings from severalother studies (Henderson & Chard, 1980) and shouldperhaps make us question the adequacy of our presentmeasures of orthographic regularity.

EXPERIMENT 2

Experiment 1 showed that misspellings that maintainoverall word shape are less noticeable than those that donot, even when the misspellingsare created by deletinga single letter. This shape effect carmot be explained asdue to systematic differences in the pronounceabilityof the misspellings used or to differences in the orthographic regularity of the misspellings. However, thepossibility still exists that there are other differencesthat are not related to word shape. Nonvisual effects ofthis kind are not difficult to think of. For example,while we can rule out absolute pronounceability, it isconceivable that there may be systematic differencesbetween same- and different-shape misspellings in thesimilarity of their sound when pronounced to the soundof their base word. An alternative, quite arbitrary explanation for our results would be to postulate that lettersthat make up the set of ascenders are somehow given


5/8

20 MONK AND HULME

Shape

DISCUSSION

Table 4Detransformed Mean Arcsine Hit Rates for the FourMisspelling Conditions in Experiment 2

Since the absence of a shape e ffec t is essentially anull result, a joint analysis was performed on the datafrom Experiments 1 and 2, in which the factors wereexperiment, shape, and type of mutilat ion. A significant Experiment by Shape interaction [F(1 ,3) = 11.830,r-< .05] confirmed that the effects of word shape weresignificantly reduced by using a mixed-case text, as wehad predicted. This abolition of the effect ofword shapeis a crucial result, since it indicates that the effect ofword shape found in Exper iment I cannot be explainedin t erms of nonvisual differences between the differenttypes ofmisspelling.

.517

.680.529.678

Same DifferentDeletionSubstitution

In Exper iment 1, we found that misspellings thatalter the shape of a word were more often noticed thanthose that preserved word shape. This is an effect ofward shape that does no t seem explicable in terms ofletter-level effects. In Experiment 2, we obtained moredirect evidence for this idea by distorting word shapethrough the use of mixed-case passages. When this wasdone, the effects of word shape observed in Experiment 1 were eliminated. The other effect in our resultsthat is worthy of comment is the very large differencein detection rates between misspellings produced bydeleting or substitut ing letters, We will now considerthe implications of these results for theories of wordrecognition.As outlined in the introduction, a central theoreticalissue in studies of word recognition has been the necessityof postula ting the use of suprale tter features such asword shape, as opposed to simpler models in which allthe information used for word recognit ion comes fromletter recognition processes. Recently, Johnston andMcClelland (1980) and McClelland and Rumelhart(1981) have proposed models that predict the occurrence of word-superiority effects without the use ofsupraletter features. Thus, our demonstration of a supral ette r feature corresponding to word shape is incompatible with these models in their current form. To makethem compatible, it is necessary to postulate supraletterfeature detectors that work in parallel to the sub letterfeature detectors already postula ted by these investigators and wh ich output directly to the word-levelprocesses, that is, bypassing the letter-level processes.As noted earlier, previous studies that have attempted

Results and DiscussionThe mean reading time was 670 sec, which represents a speed of 187 words/rnin. As one might expect,this is slower than the reading speeds were in Experi

ment 1, bu t it is still within the bounds of normal reading speeds. Similarly, although the overall numberof misspellings noticed was slightly lower than that inExperiment 1, hit rates in the two experiments wereclearly comparable. The mean comprehension score was4.0 ou t of 6, which is essentially identical to the scorein Experiment 1. There is no reason to believe that usingmixed-case text induced a dramatically different readingstrategy.The results for the detection ofmisspellings are givenin Table 4. It is clear that the shape effect was completelyabolished by using mixed-case text. There was no shapemain effect [F(1,3) < 1, n.s.] and no Shape by Type ofMuti lat ion interact ion [F(1 ,3) < 1, n.s.]. The type ofmutilation main effect was still apparent [F(l ,3) =12.364, p< .05] ; deletion misspellings were less oftennoticed than substitution misspellings. Since this effectis no t related to word shape, there is no reason for usto expect this effect to be different with mixed-casestimuli.

MethodSubjects. The subjects were 60 school sixth-formers between17 and 18 years old who had applied to the University of Yorkto study psychology.Materials and Design. The materials used were identical tothose used in Experiment 1, except that they had been processedby a program that randomly changed 50% of the lowercaseletters to uppercase. Random case changes were not used for themisspellings, however. The deletion rnisspellings were changed sothat the remaining crit icai letter and its immediate neighborswere lowercase and a11 the remaining letters were uppercase.The substitution misspe11ings were changed in the same way, bu tin addition, the substituted letter was also in lowercase. Thus,the base word "latest" becomes "Late'T," "Laes'I'," "LatecT,"and "Laces 'I'. " When case is changed in this way, the relativeheight of the critical let te rs is still apparen t, thus giving themaximum chance for letter-level effects to emerge. The procedurc in Experiment 2 was exact1y the same as that in Experiment 1.

more weight in the analysis. Thus, delet ing a "k" or an"I," for example, might be more important than deletingan "m " or an "n."Experiment 2 was designed to rule ou t any explana

t ion based on systematic differences in nonvisual properties of the different types of misspelling. In thisexperiment, the material the subject read was printed inmixed case. The presence of uppercase let ters makesoverall word shape inappropriate without changing theletters in the word. Thus, if the apparent shape effectsobserved in Experiment I were due to nonvisual properties of the stimuli, they should be unaffected by usingmixed-case text. If, on the other hand, as we maintain,the shape effect is due to a supraletter effect due toword shape, the use ofmixed-case stimuli will abolish it.


6/8

to demonst ra te a supra let te r feature cor responding toword shape have produced confl icting results (Adams,1979; Bruder, 1978; McClelland, 1976). These studieshave used tachistoscopic presentation of single words,and there are many complex diffe rences between themand the present study that make direct comparisonsbetween the various results impossible. l t is not at allclear, therefore , why we have found evidence for wordshape as a supraletter feature while some previousstudies have not. The delineation of conditions underwhich word shape effects are found is an important issuefor further research. However, the present experimentalsituation is arguably more closely comparable to normalreading than are tachistoscopic studies involving isolatedwords. It seems reasonable to suppose, therefore, thatthe word shape effect we have identified is operativeduring normal reading.

A natural way of explaining the greater frequencywith which substit utions are noti ced as compared todeletions is in terms of contradictory information. Inthe case of substitutions, a foreign letter is present ,whereas in deletions, it is only that some informationnormally present in the word is absent. l t seems plausible to argue that the system responsible for wordrecogni tion should be more toleran t of missing information than of contradictory information that is potent ially more informative. A similar effect has been notedby Healy (1981) in detecting substitution errors. Shefound in a proofreading task that for let ters that havethe same envelope, readers make more errors when asubst ituted let ter is missing one of the features of thecorrect l ett er (as when an "e" is replaced with a "c")than when the subst ituted letter has an addit ional featur e (as when a "c " is replaced by an "e").

The fact that this effect is still apparent with themixed-case text used in Experiment 2 suggests thatthese absent or contradictory e lements may be let te rsor graphemes rather than subletter or supraletter features. In other words , this result means that we cantentatively locate the effect at the point at wh ich arepresentation of the stimulus as a sequence of lettersor candidate let ters is matched with a representation ofthe word as a sequence of let ters . This in turn impliesa ra ther poor knowledge of the posit ion of the lettersidentified, as the deletion condition misspellings have adifferent length to the base words they are mistakenfor. However, it is common to assurne that letter identityinformation is posit ion specific (Bouwhuis & Bourna,1979; McClelland & Rumelhar t, 1981) . Such a systemcould not match astimulus consisting of n - 1 letterswith a word consisting of n letters. Thus , like thc wordshape effect, this finding poses problems for currentmodels of word recognition.

An alternative way of characterizing word and lettershape is in terms of spatial frequency. The shape of aletter or word corresponds to the low spatial f requencyinformation in its printed form. Thus, if one blurs a

PROOFREADING 21word so that only low spatial frequency information ispresent, what one sees is the overall shape of the word.The visual details correspond to the high spatial frequency information. There are, however, certain seriousproblems with the idea that letters and words are recognized via their Fourier spectra (see Navon, 1977 ; alsoCoffin, 1978).A less well specified bu t possibly useful distinctionis that made by Broadbent between global and localprocesses (Broadbent, 1977; Broadbent & Broadbent ,1977). According to this view, there are two phases inthe process of word recognition that interact with eachother in a cyclical fashion. The first phase is passive andutilizes information about gross word shape; this phaseoperates to reduce the range of words to a set that isthen tested in the second active phase. The existence ofword shape effects is clearly compatible with models ofthis type. Further , one may speculate that under theconditions of our experiments, syntactic and semanticconst ra in ts may lead readers to place a greate r relianceon the first phase , especially at points in the text atwhich those const ra in ts are high. There are data fromeye movement studies that are consistent with thispossibility. Ehrlich and Rayner (1981) found that subjects had lower probabilities of fixating target words inhigh-constraint passages and that when they werefixated, they were fixated for less time than in lowconstraint passages.One other theoretical issue in reading that our resultsaddress is the role of phonological coding. In Experiment I, a correlation was computed between the pronounceability of misspellings and their detectability,and essentially no relationship was found. This suggeststhat in the presen t si tua tion of reading a simple prosepassage quick ly for meaning, subjects did not utilizea phonological code to any great extent. Clearly, thisdoes not mean that readers never make use of suchcoding strategies, and indeed, there is evidence fromother proofreading tasks that subjects may do so (Cohen,1980). However, to the extent that our task reflectsnormal reading, our result suggests that the use of aphonological code is not of great import ance to theskilled reader. Our results seem well explained in termsof visual factors , together with the large effect of wordfrequency that we found, and this is consistent withthe claims of Healy (1981) that visual similarity is themain determinant of performance in a proofreading task.

As a final point, we believe that the results of thisstudy, together with others, indicate that proofreadingtasks in which the subject is also reading for meaning area useful additional means for exploring the cognitiveprocesses in fluent reading. They have the advantage ofbeing relatively simple techniques to use. Most important they complement the many experimental tasks thatuse single words as stimuli by offering the opportunityto manipulate word recognition processes when thewords in question are placed in a realistic context. Our


7/8

22 MONK AND HULMEfmding of a clear effect of word shape when such aresult has not been obtained in tachistoscopic studieswith single-word stimuli suggests that this may make aconsiderable difference to the conclusions drawn.

REFERENCESADAM8,M.J. Modelsof word recognition, CognitivePsychology,1979,11,133-176.BARON, J. Theword superiority effect: Perceptuallearning fromreading, In W. K. Estes (Ed.), Handbock of leaming andcognitive processes (V01. 6). Linguistic functions in cognitivetheory, Hillsdale,N.J: Erlbaum, 1978.BOUMA, H. Visual recognition of isolated lower case letters,Vision Research, 1971,11, 4 ~ 9 - 4 7 4 .BOUWHUI8, D., & BOUMA, H. Visual recognition of three letterwordsas derived from the recognitionof the constituent letters,Perception clPsychophysics, 1979,15, 12-22.BROADBENT, D. E. The hidden preattentive process, AmericanPsychologist,1977,31,109-118.BROADBENT, D. E., & BROADBENT, M. H. P. General shapeandlocal detail in word perception. In S. Dornie (Ed.), Attentionandperformance VI.Hillsdale,N.J: Erlbaum, 1977.BROADBENT, D. E., & BROADBENT, M. H. P. Priming and thepassive/activemodel of word recognition. In R. S. Nickenon(Ed.), Attentionandperformance VIII.Hillsdale, N.J: Erlbaum,1980.BRUDER, G. A. Role of visual similarityin the word-superiorityeffects obtained with the simultaneous matehing task. Journalof Experimental Psychology: Human Perception and Performance, 1978,4, 88-100.CATTELL, J. The time taken up by cerebral operations. Mind,1886,11,220-242.COFFIN, S. Spectral frequency analysis of block letters does notpredict experimental confusions. Perception cl Psychophysk,1978,13,69-74.COHEN, G. Reading and searching for spellingerrors, In U. Frith(Ed.), Cognitive processes in spelling,London: AcademicPress,1980.CORCORAN, D. W. J. An acoustic factor in letter cancellation,Nature, 1966,110, 6S8.EHRLICH, S., & RAYNER, K. Contextual effects on word perception and eye movements during reading, Journal of VerbalLearning and Verbal Behavior, 1981,10, 641-655.EICHELMAN, W.H. Familiarityeffects in thesimultaneousmatching task. Journal ofExperimental Psychology, 1970,16, 27S282.FI8HER, D. F. Readlng and visual search,MemorycI'Cognition,1975,3,188-196.FRITH, U. Reading by eye and wrltlng by ear, In P. A. Kolers,M. Wrolstead, & H. Bouma (Eds.), Processing of visible language (Vol. 1).NewYork: Plenum, 1979.HABER, R. N., & SCHINDLER, R. M. Error in proofreading:Evldence for syntactic control of letter processing. JournalofExperimentalPsychology: HumanPerception andPerformance,1981, 7, m - ~ 7 9 .HEALY, A. F. Detection erron on the word the: Evidence forreading unit! larger than letten. Journal of Experimental Psychology: HumanPerception andPerformance, 1976, 1, 2 3 ~ - 2 4 2 .HEALY, A. F. Proofreadingerron on the word the: New evidenceon reading units. Journal of Experimental Psychology: HumanPerception andPerformance, 1980,6,4S-47.HEALY, A. F. The effectsof visuaI similarity on proofreading formisspellings.Memory cl Cognition, 1981, 9, 4S3-460.HENDERSON, L., & CHARD, J. The readers implicitknowledge oforthographiestructure. In U. Frith (Ed.), Cognitive processes inspe//ing. London:AcademicPress, 1980.JOHN8TON, J. C., & McCLELLAND, J. L. Visual factors in wordperception. Perception cl Psychophysics, 1973, 14,36S-370.

JOHN8TON, J. C., & McCLELLAND, J. L. Experimental tests of ahlerarchlcal model of word identification. Journal of VerbalLearning and Verbal Behavior, 1980, 19, ~ 0 3 - S 2 4 .KU(':ERA, H., & FRANCI8,W. N.Computational analysis ofpresentdoyAmerican Bnglish. Providence, R.I: BrownUniversity Press,1967.MA88ARO, D. M. Perception of letters words and nonwords.Journal ofExperimental Psychology, 1973, 100, 3 4 9 - 3 ~ 3 .MA88ARO, D.M. Experimental psychology and information processing. NewYork: R a n d M c N a l l y , I 9 7 ~ .MAYZNER, M. S., & TRE88ELT, M. E. Tablesof singleletter anddigramcounts for various word-length and letter position combinations. Psychonomic Monograph Supplements, 1 9 6 ~ , I,13-32.MAYZNER, M. S., TRE88ELT, M. E., & WOLIN, B. R. Tables oftrigram frequency counts for various word-length and letterposition combinations. Psychonomic Monograph Supplements,

1 9 6 ~ , I, 33-78.McCLELLAND, J. L. Prelirninary letter identification in the perception of words and nonwords. Journal ofExperimental Psychol01O': HumanPerception and Performance, 1976,1, 80-91.McCLELLAND, J. L. Letter and configuration information inwordidentification. Journalof Verbal Learning cl Verbal Behavior,1977,16, 137-1S0.McCLELLAND, J. L.,&RUMELHART, D.An interactive activationmodelof contexteffectsin letter perception: Part 1.An accountof basic findings. Psychological Review, 1981, 88, 357-407.MORTON, J. A functional model for memory. In D. A. Norman(Ed.), Modelsfor humanmemory, NewYork: AcademicPress,1970.NAvoN, D. Forestbefore trees: The precedence of global featuresin visual perception, Cognitive Psychology, 1977, 9, 3 ~ 3 - 3 8 3 .RAYNER, K. The perceptual span and peripheral cues in reading.CognitivePsychology, 1975, 7, 6 ~ - 8 1 .RAYNER, K. Fovealand parafoveal cuesin reading, In J. Requin(Ed.), Attention andperformance VII.Hillsdale, N.J: Erlbaum,1978.RAYNER, K., McCONKIE, G., & EHRLICH, S. Eye movementsand integrating information across fixations. Journal ofexperimentalPsychology: HumanPerception andPerformance, 1980,4, ~ 2 9 - ~ 4 4 .REICHER, G.M. Perceptual recognitionas a function of meaningfulness of stimulus material. Journal of Experimental Psychology, 1969,111, 2 7 ~ - 2 8 0 .SMITH, F. Familiarity of configuration vs discrlminability offeatures in the visual identification of words. Psychonomtcsctence, 1969,14, 261.SMITH, P. T., & GROAT, A. Spellingpatterns, letter cancellationand the processing of text, In P. A. Kolers, M. Wrolstead, &H. Bouma (Bds.), Processing of visible language (Vol. 1).NewYork:Plenum, 1979.TINKER, M. A. Bases for effective reading. Minneapolis: Universityof Minnesota Press, 1 9 6 ~ .WINER, B. J. Statistical principles in experimental design.NewYork:McOraw-HiII, 1971.

APPENDIXBASEWORDS ANDMUTILATIONSbenches beches bences besches benceesrarity raity rariy raeity rarirymodern moden moern modean momemnumber nuber n\lmer nurber nummerdecide deide decie devide decivestrips stips srips stoips soripstaking takig taing takirg tavinguntil util uni! urti! unrilhardest hadest harest handest harmestclassle clasie eassie clarsie erassle


8/8

PROOFREADING 23created ceated ereaed eseated ereawed ehureh ehueh eureh ehuneh erurehpurpIes puples purpes pumples purpmes making makig maing makirg maringdepends depeds depens deperds depenns fourth fouth fourh founth fourehonly oly on y osly onsy before befoe beore befome bemorehands hads hans hamds hanes setting settig seting settirg sertingfemales feales femaes fevales femaves alive alie aive alise asivesmiled siled smied seiled smieed ehanged ehaged eanged ehaeged erangedthink thik tink thirk trink eraft eaft erat eoaft erastbetween beteen beween betreen berween afford affod aford affond anfordfeeling feelig feeing feelirg feering awaited aaited awaied amaited awaimedlatest latet laest lateet lacest trieky triky triey trinky trieeysueked suked sueed surked suered plastie platic pastie plantic prasticreekons reekos reeons reekoms reemons glowing gloing gowing gloning gnowingvoeation voation voeaion vosation voeasioninstalI intall insall inntall insrall (Reeeived for publieation June 9,1982;worked woked wored wonked worned revision aeeepted September 27,1982.)

Documents

Errors in Proofreading Evidence for the Ude of Word Shape in Word Recognition