The Quarterly Journal of Experimental Psychology Section A ISSN: 0272-4987 (Print) 1464-0740 (Online) Journal homepage: http://www.tandfonline.com/loi/pqja20 The Representation of Sublexical OrthographicPhonologic Correspondences: Evidence from Phonological Dyslexia Mary F. Lesch & Randi C. Martin To cite this article: Mary F. Lesch & Randi C. Martin (1998) The Representation of Sublexical Orthographic-Phonologic Correspondences: Evidence from Phonological Dyslexia, The Quarterly Journal of Experimental Psychology Section A, 51:4, 905-938, DOI: 10.1080/713755790 To link to this article: http://dx.doi.org/10.1080/713755790 Published online: 22 Oct 2010. Submit your article to this journal Article views: 36 View related articles Full Terms & Conditions of access and use can be found at http://www.tandfonline.com/action/journalInformation?journalCode=pqja20 Download by: [University of Florida] Date: 05 November 2015, At: 18:09 T HE QUARTERLY JOURNAL OF EXPERIMENTAL PS YCHOLOGY, 1998, 51A (4), 905± 938 T h e Rep resen tation of Su blexical Orth og rap h ic± Ph on ologic Corresp on d en ces: Ev id en ce from Ph on ological Dyslexia Mary F. Lesch and Randi C. Martin Downloaded by [University of Florida] at 18:09 05 November 2015 Rice University, Houston, Texas, U.S.A. Although there is considerable evidence that grapheme and body units are involved in assembling phonology from print, there is little evidence supporting the involvement of syllabic representations. We provide evidence on this point from a phonological dyslexic patient (ML) who, as a result of brain damage, is relatively unable to read nonwords. ML was found to be able to perform tasks assumed to re¯ ect processes involved in assembled phonology (i.e. segmentation, orthographic± phonologic conversion, and blending) when the units involved were syllables, but demonstrated considerable dif® culty when they were onset, body, or phoneme units. Additionally, both ML and matched controls were much better able to ® nd words in an anagrams task (Treiman & Chafetz, 1987) when they resulted from the combination of segments corresponding to syllables than when they did not. It is suggested that the relationship between print and sound is represented at multiple levels (including the syllable) (Shallice, Warrington, & McCarthy, 1983) and that ML’ s nonword reading impairment is the result of disruption of representations below the level of the syllable. Traditional dual-route models of reading (e.g. Coltheart, 1978) suggest that there are two means of obtaining the phonological representation of a written word: (a) by ``addressing’ ’ a word’ s lexical entry, which speci® es a whole-word phonological representation (the lexical route) or (b) by ``assembling’ ’ a phonological representation on the basis of spelling± to± sound correspondences (the sublexical route). T he lexical route is assumed to be necessary in order to read (out loud) words like HAVE that form exceptions to spelling± to± sound correspondence rules (i.e. words with similar spellings are pronounced differentlyÐ WAVE , GAVE, RAVE, S AVE ). Evidence for a lexical route has frequently come in the form of null effects of assembled phonology (see Van Orden et al., 1992 for a discussion of this point) and from patients with acquired reading disorders. In support of a sublexical route to phonology, many researchers have observed a ``regularity effect’ ’ Ð the ® nding Requests for reprints should be sent to Mary F. Lesch, Rice University, Psychology DepartmentÐ MS 25, 6100 Main S treet, Houston, T X 77005± 1892. E-mail: lesch@ rice.edu T his research was supporte d by NIH Grant DC00218 to Rice University. We wish to thank ML for his participation in this project and Trisha Burton, Jennifer Davidow, Sarah Doss, and Anna Mandelbaum for their help at various stages. We would also like to thank Alexander Pollatsek, David Howard , Karalyn Patterson, and Helen Moss for their comments on earlier versions of this manuscript . q 1998 T he Experimental Psycholog y S ociety Downloaded by [University of Florida] at 18:09 05 November 2015 906 LESCH AND MARTIN that regular words (words with predictable grapheme± to± phoneme correspondences) are named more quickly than irregular words (Baron & Strawson, 1976; Gough & Cosky, 1977 ; Stanovich & Bauer, 1978). However, more recent research (Andrews, 1982; Backman, Bruck, Hebert, & Seidenberg, 1984; Seidenberg, 1985; Seidenberg, Waters, Barnes, & Tanenhaus, 1984; Waters, S eidenberg, & Bruck, 1984) indicates that this effect may be restricted to the processing of low-frequency words. S imilarly, Glushko’ s (1979) ® nding of an effect of ``consistency’ ’ Ð the consistency with which orthography is mapped to phonology at the level of the word body (the vowel and succeeding consonants)Ð has also been found to be restricted to low-frequency words (Seidenberg et al., 1984). It has been argued that the restriction of the regularity effect to the processing of low-frequency words indicates that the sublexical route serves as a slower ``back-up’ ’ to the faster lexical route (e.g. Seidenberg, 1985). T he occurrence of two distinct syndromesÐ surface dyslexia and phonological dyslexiaÐ suggests a functional separation of the lexical and sublexical routes. In surface dyslexia, an effect of spelling± to± sound regularity on reading (such that regular words are read more easily than irregular words) and the occurrence of regularization errors (reading irregular words as though they were regular) suggest that the lexical route has been damaged and that reading is forced to proceed on the basis of a relatively intact sublexical route. In acquired phonological dyslexia, the patients’ relatively spared ability to read words, coupled with their relative inability to read nonwords, suggests that, in this syndrome, the sublexical route is damaged whereas the lexical route is relatively spared. While most researchers seem to agree that there are both sublexical and lexical routes to phonology, there is considerable disagreement concerning the nature of these two routes. Some theorists have suggested an indirect lexical route that proceeds through semantics (e.g. Coltheart, Curtis, Atkins, & Haller, 1993; Morton & Patterson, 1980; Patterson & Shewell, 1987; Plaut, McClelland, Seidenberg, & Patterson, 1996). A semantic lexical route to phonology is suggested by the reading performance of deep dyslexic patients (see Marshall & Newcombe, 1973; S affran & Marin, 1977). T he inability of deep dyslexics to read nonwords indicates impaired assembled phonology. T herefore, their word reading cannot rely on grapheme± to± phoneme conversion and must rely on a lexical route to phonology. Evidence that this lexical route proceeds through semantics comes in the form of word class effects (better reading of concrete than of abstract words) and the occurrence of semantic errors. Evidence for a more direct route from the written word to the spoken word comes from the study of patients who can read irregular words that they do not understand (see Coltheart, Masterson, Byng, Prior, & Riddoch, 1983; Funnell, 1983 ; Schwartz, Saffran, & Marin, 1980). However, Hillis and Caramazza (1991) have questioned the necessity of postulating a direct lexical route, as it may be the case that the reading of these patients is mediated by the summation of partial semantic information and partial phonological information (resulting from attempts to apply grapheme± to± phoneme correspondence rules). Some theorists (e.g. Morton & Patterson, 1980; Patterson & Shewell, 1987; see also Coltheart et al., 1993) propose both semantic and direct lexical routes. T here is also considerable debate as to how the sublexical route works and, hence, what types of item it is capable of processing. For example, Coltheart and colleagues (Coltheart et al., 1993) propose a three-route model in which the sublexical route is capable of Downloaded by [University of Florida] at 18:09 05 November 2015 THE REPRESENTATION OF SUBLEXICAL CORRESPONDENCES 907 producing the correct pronunciation for regular words whereas exception words must be read through the lexical route. In their formulation, it is not clear how much a semantic route contributes to skilled oral reading. T he class of models proposed by S eidenberg and colleagues (e.g. Plaut et al., 1996; Seidenberg & McClelland, 1989), on the other hand, has two routes: (a) a sublexical route that processes both regular and exception words and (b) a semantic route that makes a strong contribution to reading (particularly for low-frequency exception words). In a third class of models (e.g. Shallice, Warrington, & McCarthy, 1983) phonological representations are assembled on the basis of stored orthographic± phonologic correspondences ranging from the level of the individual grapheme up to the level of the whole word (see Taft, 1991, for a similar model that does not incorporate whole word orthographic± phonologic representations). T hus, sublexical and lexical phonology are subsumed under one mechanism. Such a multiple-levels model has been implemented by Norris (1994). T he following discussion of the sublexical route will be organized around two issues: (a) the levels of orthographic± phonologic representations involved in reading and (b) the processes involved in assembling phonology from print. Some researchers argue that the conversion from orthography to phonology proceeds at the level of individual graphemes and phonemes (e.g. Coltheart, 1978; Coltheart et al., 1993) whereas others suggest the involvement of other levels of correspondence (e.g. Patterson & Morton, 1985; Shallice et al., 1983). In the current investigation we present a case study of a phonological dyslexic patient, which sheds some light on how sublexical orthographic± phonologic correspondences are represented and their role in reading. While traditional dual-route theories (e.g. Coltheart, 1978) suggest that the mapping from orthography to phonology proceeds at the level of individual graphemes and phonemes, there is evidence that other levels of correspondence may be relevant to the assembly process. T he observation of a consistency effect suggests that the word-body may be important. Treiman (e.g. Treiman and Zukowski, 1988) has also obtained evidence for the importance of the body along with the onset in the processing of written words. T he onset and body are subsyllabic units. In monosyllabic words, the onset is the initial consonant or consonant cluster whereas the body is the vowel and all following consonants. T hus, in the word BLACK , BL forms the onset of the syllable and ACK forms the body. Treiman and Zukowski found that the pronunciation assigned to a nonword was in¯ uenced more by its body than by its onset or vowel grapheme alone. T hat is, they found that subjects were more likely to assign the phoneme / E/ to the grapheme ``ie’ ’ when it occurred in CHIEND (which shares its body with the real word FRIEND) than when it occurred in FRIET H or CHIET H. Treiman (Treiman & Chafetz, 1987) has also observed that words are found more easily in an anagrams task when the units to be combined correspond to onset and body units (e.g. it is easier to ® nd the word BLACK within the segments [ BL OFT ACK FR] than within the segments [ BLA FT CK FRO]). Additional evidence for the importance of onset and body units in the processing of written words comes from the demonstration that these sublexical units prime naming of target words containing these units. For example, Bowey (1990) found that primes corresponding to the onset or body of a following target word (e.g. BR followed by BRAND or AIL followed by HAIL ) speeded naming times relative to when the primes did not correspond to these units (e.g. BI followed by BIRCH or RAY followed by PRAY). Downloaded by [University of Florida] at 18:09 05 November 2015 908 LESCH AND MARTIN Whereas Treiman and colleagues have made a strong case for the importance of subsyllabic (onset± body) units in the processing of written words, a much less consistent picture emerges when syllabic units are considered (see Seidenberg, 1989). When subjects are asked to make same± different decisions on words matched in length (number of letters), some researchers have observed an effect of number of syllables on latencies to make ``different’ ’ responses (e.g. Klapp, 1971), whereas others have observed an effect only on latencies to make ``same’ ’ responses (e.g. Taylor, Miller, & Juola, 1977). Still others have failed to observe an effect on either type of response (e.g. Barron & Pittinger, 1974). A similar inconsistency results from the use of naming latency, with some researchers observing an effect of number of syllables (Butler & Hains, 1979; Klapp, Anderson, & Berrian, 1973) and some not (Frederiksen & Kroll, 1976; Richardson, 1976). Another method that has been used to examine the role of syllabic units in the processing of written words involves the brief presentation of written words whose letters are coloured. Prinzmetal, Treiman, and Rho (1986) found that illusory conjunctions (when a letter is erroneously reported as being the colour of another letter in the word) occur more often among letters within the same syllable than among letters that cross a syllable boundary. T his effect was observed for words containing orthographically marked boundaries (where syllable boundaries are marked by letter combinations that are illegal within syllables, e.g. NV in ANVIL ). Seidenberg (1987), however, observed a syllable effect with words whose syllable boundary was marked by a ``bigram trough’ ’ (a bigram with a lower frequency than that preceding or following it), but found that orthographically similar pairs with different syllable structure (e.g. WAIVE± NAIVE ), behave similarly in this paradigm. S eidenberg (1987, 1989) suggests that the critical factor in the observation of a syllable effect is the presence of a bigram trough at the syllable boundary. However, as pointed out by Rapp (1992), Seidenberg (1987) did not examine words lacking a bigram trough at the syllable boundary. Using stimuli with orthographically marked syllabic boundaries, Rapp observed a signi® cant effect of syllable structure on both trough-present and trough-absent stimuli indicating that the presence of a bigram trough is not critical to the observation of a syllable effect. More recently, Carreiras, Alvarez, and deVega (1993) observed an effect of the frequency of syllables (in S panish) on naming and lexical decision times, which they argued could not be accounted for in terms of orthographic redundancy. T he view of reading that emerges from our discussion thus far is that there are at least two different processes by which a phonological representation may be obtained from a written word (lexical vs. sublexical phonology) and that multiple orthographic± phonologic representations (grapheme± phoneme, onset, body, and syllabic units) may be involved in assembling a pronunciation on the basis of spelling± to± sound correspondences. Additional evidence regarding the operation of the sublexical route comes from neuropsychological case studies (e.g. Beauvois & Derouesne, 1979; Derouesne & Beauvois, 1985 ; Funnell, 1983; Goodall & Phillips, 1995). According to Beauvois and Derouesne (1979; Derouesne & Beauvois, 1985) there are three distinct stages upon which the sublexical route depends: (a) a grapheme stage that parses the written stimulus; (b) print± to± sound conv ersion that operates on the units from Stage a; and (c) a phonemic stage in which the sounds obtained in Stage b are assembled into the phonological representation for the written stimulus (see Figure 1 for an illustration of Beauvois and Derouesne’s THE REPRESENTATION OF SUBLEXICAL CORRESPONDENCES 909 Downloaded by [University of Florida] at 18:09 05 November 2015 stages of assembled phonology. However, we have assumed the involvement of multiple levels of orthography and phonology even though Beauvois and Derouesne only consider graphemes and phonemes). It is argued that disruption to any of these processes results in a pattern of reading performance characteristic of phonological dyslexia. T here is some support for such an analysis (see Goodall & Phillips, 1995). For example, Newcombe and Marshall’ s (1985) patient MS seems to suffer an impairment to graphemic analysis (S tage a), whereas Funnell’s (1983) patient WB suffers an impairment in the assignment of phonemes to graphemes (Stage b), and Bub, Black, Howell, and Kertesz’ s (1987) patient FIG. 1 . Beauvois and Derouesne ’s (1979) stages of assembled phonolog y illustrated with multiple levels of orthographi c and phonological representations. Downloaded by [University of Florida] at 18:09 05 November 2015 910 LESCH AND MARTIN MV, as well as the patients studied by Patterson and Marcel (1992), seems to have an impairment in phonological blending (Stage c). More recently, Goodall and Phillips (1995) concluded that their patient AN has de® cits in print± to± sound conversion and phonological blending, but spared segmentation abilities (graphemic and phonological). In a further investigation of WB’ s impairment, Patterson and Marcel (1992) found that 1 this patient was unable to perform phonological blending with nonword targets. Berndt, Haendiges, Mitchum, and Wayland (1996) reported 11 patients who all showed evidence of impairment to all three of the components of nonword reading. Additionally, for some patients, they observed an in¯ uence of segment size on performance. T hese studies suggest that nonword reading impairments result when any of the stages of nonword reading have been disrupted. An issue that is not addressed in Beauvois and Derouesne’s analysis is how the conversion from print to sound takes place. T here has been relatively little neuropsychological research addressing the issue of the nature of the orthographic units and phonological units involved in reading. An exception in this regard is the work of S hallice et al. (1983; see also Berndt et al., 1996; Funnell, 1983). Based on work with a surface dyslexic patient (HT R), S hallice et al. (1983; see also Shallice & McCarthy, 1985; S hallice & Warrington, 1980) suggested that correspondences between orthography and phonology are represented at multiple levels. T heir data suggested that HT R’ s surface dyslexia re¯ ected a reliance upon smaller units of orthographic± phonologic correspondence (e.g. grapheme± phoneme, subsyllabic units) in the face of disruption to larger units of correspondence (e.g. word or morpheme units). More speci® cally, they observed an effect of degree of regularity on HT R’ s reading performance such that an ``irregular’ ’ word (where ``regularity’ ’ is de® ned in terms of the vowel) was more likely to be read correctly if that irregular pronunciation were the most frequent one for the subsyllabic CV or VC than if the most frequent pronunciation were the regular one. An effect of degree of regularity is not predicted by models that posit that assembled phonology proceeds solely on the basis of grapheme± to± phoneme correspondences. T hus, there is evidence both from studies with normal subjects and studies with patients with acquired dyslexias that reading is in¯ uenced by the relationship between multiple levels of orthography and phonology. Shallice et al. (1983) provide an interpretation of surface dyslexia in terms of a selective disruption of word or morpheme units. It is of interest to determine whether a similar interpretation (i.e. in terms of the disruption of representational units) is possible for phonological dyslexia. With that goal in mind, we investigated the word and nonword reading abilities of an acquired phonological dyslexic, ML. T his investigation was initially spurred by the observation that ML’s nonword reading, although very poor, was not entirely abolished. Preliminary testing suggested the possibility that ML’s nonword reading is mediated by syllabic representations (see later). We were therefore interested in whether the pattern of disruption of the sublexical route could be attributed to a disruption of one or more of the processes hypothesized by Beauvois and Derouesne (i.e. 1 T he pattern of performanc e of some patients on phonological tasks has led to the suggestion that phonological dyslexia is a ``symptom’’ of a more generalized phonological de® cit (see Farah, Stowe, & Levinson, 1996; Friedman, 1996; Patterson, Suzuki, & Wydell, 1996; S asanuma, Ito, Patterson, & Ito, 1996). However, it appears that not all patients can be accommodated within this account. Downloaded by [University of Florida] at 18:09 05 November 2015 THE REPRESENTATION OF SUBLEXICAL CORRESPONDENCES 911 orthographic segmentation, conversion of orthographic units into phonological units, and blending of phonological units) (1979; Derouesne & Beauvois, 1985), or whether it might be better explained in terms of the disruption of different units of orthographic± phonologic correspondence (those below the level of the syllable). Testing of ML’ s nonword reading was organized by administering tasks designed to tap processes at the three stages hypothesized by Beauvois and Derouesne. At each stage, processing of different sizes of orthographic± phonologic units was assessed (see Figure 1). To the extent that ML’s performance is found to vary with the task, but not with the level of representation involved, there would be evidence that ML’s nonword reading impairment is due to a disruption of the processes involved in assembling phonology from print. On the other hand, if ML’ s performance is found to vary with the level of representation involved, but not with the task, then there would be evidence that his impairment is due to the disruption of orthographic± phonologic representations. We were particularly interested in whether ML’ s performance on the various tasks would reveal syllabic effects as there was some indication in initial tests of his nonword reading that these units might be important for him. As indicated earlier, the evidence is inconsistent regarding a role for syllabic representations in reading. T herefore, the observation of such an effect on the word and nonword reading of this phonological dyslexic patient would be an important ® nding. PATIENT DESCRIPTION ML is a 57-year-old white male who suffered a left hemisphere cerebral vascular accident in May, 1990. CT scan revealed an infarction involving the left frontal and parietal operculum. Atrophy in the left temporal operculum was noted, as was mild diffuse atrophy. ML had completed two years of college study and prior to his injury had been employed as a draughtsman. ML exhibits mild agrammatism and word-® nding dif® culties in his spontaneous speech. ML’s word reading is mostly intact as re¯ ected by his performance on word lists obtained from Roeltgen, Cordell, and Sevush (1983) and from the Psycholinguistic Assessments of L anguage Processing in Aphasia (PAL PA) (Kay, Lesser, & Coltheart, 1992). On the Roeltgen et al. word lists ML obtained 100% (40/ 40) correct on nouns, 75% (30/ 40) correct on function words, and 98% (44/ 45) and 96% (43/ 45) correct on regular and irregular words, respectively (matched on imageability, length, word frequency, and word class) (see Table 1). ML’ s errors on function words were generally function word substitutions. On the PALPA test of grammatical class reading, ML obtained 100% (20/ 20) correct for nouns, adjectives, and verbs, and 85% (17/ 20) correct for function words. T hus, although ML reads most word classes at a very high level, he appears to have a slight impairment in reading function words. A dif® culty with function word reading is commonly associated with phonological dyslexia (see Beauvois & Derouesne, 1979; Friedman, 1995; Patterson, 1982) suggesting the possibility that nonwords and function words are read through the same process (Patterson, 1982). However, arguing against this notion are reports of patients who have severe nonword reading de® cits, but good word reading even for function words (Caramazza, Miceli, Silveri, & Laudanna, 1985; Funnell, 1983). We will not comment on the possible causes of ML’s 912 LESCH AND MARTIN TABLE 1 Word Reading Downloaded by [University of Florida] at 18:09 05 November 2015 Score a Word Type Roeltgen PALPA Nouns Adjectives Verbs Function Regular Irregular 100 100 100 100 85 a 75 98 96 a In percentages. Nonword Reading PALPA N onwords Example Score 3 letter 4 letter 5 letter 6 letter cug boak snite dringe 2/ 6 2/ 6 4/ 6 0/ 6 Roeltgen N onwords Simple Score Complex Score 2 phoneme 3 phoneme 4 phoneme 5± 7 phonemes ep nud scod epzim 6/ 10 4/ 10 4/ 10 2/ 4 aub shev tharp endurf 5/ 10 4/ 10 2/ 10 0/ 4 dif® culty in reading function words beyond indicating that it does not merely re¯ ect an effect of imageability on his reading performanceÐ he reads high- and low-imageability nouns (Kay et al., 1992) at about the same level (39/ 40 and 37/ 40 for high and low imageability nouns, respectively) (see Bartha & Martin, in preparation, for other evidence on this issue). ML’s nonword reading was assessed using a set obtained from the PALPA (Kay et al., 1992 ) in which the nonwords were monsyllabic and varied on number of letters. He was also tested on nonword lists obtained from Roeltgen et al. (1983). T his set varied on number of phonemes and complexity (i.e. the size of graphemes involvedÐ one letter vs. two letters). In sharp contrast to his high level of performance on word reading (almost perfect performance on all classes of words except function words), ML only obtained 38% correct across the two sets of nonwords with a slight indication of poorer performance with longer nonwords than with shorter nonwords (see Table 1). Good word reading and relatively poor nonword reading are the de® ning features of phonological dyslexia. We will return to a more detailed discussion of ML’s nonword reading later. ML performs well on tests of single word processing as evidenced by his performance on tasks obtained from the Philadelphia Comprehension Battery (S affran, Schwartz, Downloaded by [University of Florida] at 18:09 05 November 2015 THE REPRESENTATION OF SUBLEXICAL CORRESPONDENCES 913 Linebarger, Martin, & Bochetto, 1989): auditory lexical decision (99% correct), auditory discrimination (97% correct), and naming (98% correct) (see Martin & Lesch, 1996). ML suffers from a short-term memory (S T M) impairment, which is re¯ ected in his memory span of 2.33 items for word lists and 2.20 items for nonword lists (see Martin & Lesch, 1996, for more detailed information). T hus, ML’ s recall bene® ts little from the semantic information contained within words. However, ML’ s span performance does reveal standard effects of phonological similarity and mode of presentation (auditory vs. visual) indicating that, although still impaired, he has some ability to retain phonological information. Elsewhere (see Martin & Lesch, 1996) we have interpreted ML’ s ST M de® cit in terms of separate (but interactive) capacities for retaining semantic and phonological information: Although ML’ s ability to retain phonological information is relatively intact, recall suffers from the lack of semantic support. Consistent with this interpretation, ML’ s errors in recall are generally phonologically related items indicating that some phonological information has been retained but that insuf® cient semantic information is available to select the appropriate target from a pool of phonologically similar competitors. Additionally, recall of early list items is affected by imageability. In other formulations, Patterson and colleagues (Graham, Hodges, & Patterson, 1994; Patterson & Hodges, 1992 ; Patterson, Graham, & Hodges, 1994; Patterson, Suzuki, & Wydell, 1996) have suggested that the loss of semantic support results in phonological representations becoming ``unglued’ ’. Most of the testing reported here occurred between S eptember 1993 and May 1997. EXPERIMENTAL TASKS Orthographic Segmentation In order to translate print to sound, it is ® rst necessary to parse the written stimulus into the units relevant to the translation process (see Figure 1). Although Beauvois and Derouesne (1979) term this the grapheme stage, the research reviewed earlier suggests the importance of orthographic units larger than the grapheme, such as the onset, body, and syllable. T herefore, we investigated ML’ s ability to segment written stimuli into these units. Syllables. ML was asked to mark syllable breaks in 31 written words varying from 2 to 4 syllables in length (e.g. future, orchestra, cemetery). For all of the syllable tasks reported in this paper, syllable boundaries were determined by using the syllabi® cation provided by the dictionary. ML performed at a high level (see control performance reported later) in segmenting the words (77% items correct; 80% segments correct). Of his errors, 4/ 7 involved two-syllable words ending with ``-er’ ’ (e.g. blund/ er). Onset± body Units. We also assessed ML’ s ability to segment one-syllable written words into onset and body units (e.g. cat, beef, phone). We asked ML to think about words that rhyme and to focus on that part of the word that is shared among rhyming words (the word-body, e.g. bed, red, shed). He was instructed to place a mark separating that portion of the word from the rest of the word and was given several 914 LESCH AND MARTIN examples. ML found this task very dif® cult and obtained only 3/ 15 (20% ) items correct. Erroneous segmentations included bla/ ck and be/ ef. Downloaded by [University of Florida] at 18:09 05 November 2015 G raphemes. Finally, ML was given an orthographic segmentation task in which he was asked to mark graphemes (single letters or groups of letters corresponding to an individual phoneme) in words (e.g. shy, boot, coat). It was explained to ML that sometimes two written letters combine to produce one sound, and he was given several examples to demonstrate this. However, he performed quite poorly on this task, obtaining only 8/ 15 (53% ) items correct. When performance was scored for number of graphemes marked correctly, he obtained 27/ 48 (56% ) graphemes correct. Erroneous segmentations included st/ e/ ak and co/ at. Performance of Control Subjects on Orthographic Segmentation Control subjects of about the same age and education level as ML were tested on all the tasks administered to ML. As can be seen in T able 2, ML’s pattern of performance differed from the pattern shown by all six control subjects in that he showed great dif® culty with onset± body and grapheme units, whereas control subjects showed little dif® culty with these units. Although ML’ s performance with syllable units was clearly within the range of the controls (see Table 2), he performed well outside the range for onset± body and grapheme units. It should be noted that ML and several of the control subjects performed at well below-perfect levels on the syllable task. T his is undoubtedly partly due to the fact that the orthography does not always carry explicit cues to syllable boundaries. In fact, some sounds straddle the syllable boundary and can be considered ambisyllabic (i.e. / r/ in TABLE 2 A Comparison of the Performance of ML and Control Subjects on the Orthographic Segmentatio n Tasks Percentage Correct Syllable Onset± body Grapheme Control Subjects Items Segments Items Segments Items Segments PS PH MB MW TP HF 100 81 90 84 61 71 100 76 93 88 72 76 100 100 100 100 100 80 100 100 100 100 100 80 100 87 87 100 93 100 100 88 94 100 96 100 Mean Range 81 61± 100 84 72± 100 97 80± 100 97 80± 100 95 87± 100 96 88± 100 ML 77 80 20 20 53 56 THE REPRESENTATION OF SUBLEXICAL CORRESPONDENCES 915 ``sparrow’’ ). Additionally, there are different theories of syllabi® cation. For the current study, we used the syllabi® cation provided by the dictionary. However, some researchers (e.g. Treiman, personal communication) are investigating how individuals syllabify different types of stimuli. Downloaded by [University of Florida] at 18:09 05 November 2015 Phonological Segmentation In the second stage of assembled phonology (print± to± sound conversion) proposed by Beauvois and Derouesne (1979), the units from the grapheme stage are converted into their corresponding sounds (see Figure 1). In learning to read, we learn to associate written units with spoken units. T his process would seem to entail the ability to extract the relevant phonological units from spoken words (because the conversion process proceeds over units smaller than the word). T hus, in order to establish initially the phonological representations involved in reading, the beginning reader would need to have some awareness of the structure of spoken language. Morais, Cary, Alegria, and Bertelson (1979) have put forward an interactionist view of the relationship between phonological awareness and alphabetic literacy. According to this view, phonemic awareness is both a cause and a consequence of alphabetic literacy. T here is evidence that speech segmentation skills develop as a result of learning to readÐ ex-illiterates (those with a rudimentary ability to read and write) demonstrate an ability to analyse speech into phonemes (they perform at a high level in phoneme deletion and addition tasks), whereas illiterates do not (see Morais, Bertelson, Cary, & Alegria, 1986; Morais et al., 1979). T hus, learning the mappings between spoken language and written language seems to increase phonological awareness. However, whether ``phonological awareness’ ’ is important in skilled reading is not so clear. T hat is, once one has established mappings between orthographic and phonological units the ability to blend the phonological outputs would be critical, whereas the ability to segment spoken words might not. Nonetheless, we thought it of interest to determine whether ML’s ability to extract phonological units from spoken words would depend crucially upon the nature of the unit involved (as it had in orthographic segmentation). Whether ML can segment speech into phonemes, onset± body units, and syllables may have important implications for whether he can re-learn sublexical mappings from orthography to phonology. Syllables. ML was given auditory words (the two- to four-syllable words used in the orthographic segmentation task) and was asked to produce their constituent syllables. His performance was perfect. Onset± body Units. Additionally, ML was asked to produce onset and body units from real words (e.g. ``bl’ ’ and ``ack’ ’ from ``black’’ ). S ince it is not possible to produce some consonant clusters in isolation, responses that included a reduced vowel were scored as correct. He obtained 0/ 15 items correctÐ for no individual item did he accurately produce the onset or the body. Errors included failures to include any segmentation (slip ® ``slip’’ ) and incorrect segmentations (e.g. ¯ irt ® / f­ / -/ | ­ rt/ ; sock ® / s ­ / . . . / s OO/ . . . / sO/ - / k/ ). Downloaded by [University of Florida] at 18:09 05 November 2015 916 LESCH AND MARTIN Phonemes. ML was asked to produce the phonemes contained within one-syllable words consisting of three phonemes (e.g. ``cat’’ ). As was the case with the onset units, some phonemes cannot be produced in isolation (e.g. stop consonants). T herefore, responses that included a reduced vowel were scored as correct. ML never succeeded in producing all the phonemes contained within a single word (0/ 15 items correct). However, he showed some ability to produce the initial phoneme (4/ 15), but quite often he produced a syllable that sometimes corresponded to the initial consonant and vowel (3/ 15), but sometimes did not (e.g. ``mit’ ’ ® / m E/ ) (1/ 15). ML correctly produced 9/ 45 of the phonemes. Usually, his responses were a combination of individual phonemes, syllables, and letter names (e.g. rig ® / r ­ / . . . / k­ / . . . ``a’’ . . . ``g’ ’ ; hen ® / t­ / . . . / h En/ . . . ``n’ ’ ). In light of ML’s severely restricted ST M capacity, it seemed possible that his poor performance in producing constituent phonemes could be partly due to the memory load entailed in retaining the words while sounding them out. In an attempt to decrease the memory load, we asked ML to produce the constituent phonemes in written words (the same three-phoneme, one-syllable words used with auditory presentation). As ML has dif® culty with grapheme segmentation, it was not clear whether written presentation would aid performance. However, ML is able to read written words quite well and this should enable him to ``refresh’ ’ the whole-word phonology while he is attempting segmentation. Again, however, he never succeeded in producing all the phonemes contained within a single word (0/ 15). ML correctly produced 15/ 45 of the phonemesÐ however, these phonemes were scattered among letter names, productions of the whole word, and combinations of phonemes. Performance of Control Subjects on Phonological Segmentation As can be seen in Table 3, ML’ s poor performance on onset± body and phoneme units is in stark contrast to the excellent performance of the control subjects: Although ML’s performance fell within the range of controls on syllable segmentation, his performance was at ¯ oor on onset± body units and phoneme units, whereas the performance of the controls was at ceiling. Summary of Segmentation Results ML performs well when his task is to segment visually or auditorily presented stimuli into constituent syllables. However, when asked to segment stimuli into graphemes/ phonemes or into onset± body units, ML demonstrates considerable dif® culty in performing the task. A comparison of ML’ s performance with that of control subjects indicates that he performs at a normal level when the units involved are syllables, but performs at well below normal levels when the units are graphemes/ phonemes or onset± body units. THE REPRESENTATION OF SUBLEXICAL CORRESPONDENCES 917 TABLE 3 A Comparison of the Performance of ML and Control Subjects on the Phonologica l Segmentatio n Tasks Percentage Correct Downloaded by [University of Florida] at 18:09 05 November 2015 Syllable Onset± body Phoneme Control Subjects Items Segments Items Segments Items Segments PS PH MB MW TP HF 100 100 100 100 100 94 100 100 100 100 100 99 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 Mean Range 99 94± 100 99 99± 100 100 100 100 ML 100 100 0 N/ A 100 N/ A 0 0 20 CONVERSION OF ORTHOGRAPHIC SEGMENTS INTO PHONOLOGICAL SEGMENTS According to Beauvois and Derouesne (1979), assembled phonology requires segmentation of orthographic stimuli into the relevant units, the conv ersion of these orthographic units into their corresponding phonological units, and the blending of these phonological units (see Figure 1). T hus far, the data we have reported from ML suggest that he is able to segment both orthographic and phonological representations when the relevant unit is the syllable. However, performance suffers considerably when other units are involved. What about ML’ s ability to convert orthographic segments into phonological segments? ML’s ability to convert from orthography to phonology was examined at the levels of the syllable, onset± body, and letter units. Nonword Reading On an initial testing of ML’ s nonword reading, ML performed very poorly. As reviewed earlier (see Table 1), ML obtained only 38% correct across two sets of nonwords. On the PALPA nonwords (Kay et al., 1992) 11/ 16 errors were lexicalizations, whereas 20/ 41 errors on the Roeltgen et al. (1983) nonwords were lexicalizations. ML’s tendency to produce real words in response to nonword stimuli suggests a strong lexical in¯ uence on his nonword reading attempts. It is interesting to note that occasionally, when ML was able to give a correct response, the nonword corresponded to the initial syllable of a real word, e.g. ``rab’ ’ (rabbit), ``mur’ ’ (murder). T his observation raised the possibility that ML may be able to retrieve sublexical phonological representations when they correspond to syllable units in words. 918 LESCH AND MARTIN Downloaded by [University of Florida] at 18:09 05 November 2015 Reading of Nonwords Corresponding to Sublexical Units of Words Syllables. To test the possibility that ML is able to retrieve phonological representations for syllabic units when they correspond to a syllable from a word, we asked ML to read nonwords corresponding to the initial segment of real words. Of these items, 34 corresponded to the initial syllable of real words (syllable nonwords) (e.g. ``¯ ut’ ’ , which forms the ® rst syllable of ``¯ utter’ ’ ), whereas the other 34 formed the initial segment of real words but never corresponded to an initial syllable of a real word (control nonwords) (e.g. ``furb’ ’ from ``furbish’ ’). T he two types of nonword were approximately matched in terms of bigram frequency (Mayzner & Tresselt, 1965), neighbourhood size (N ), and summed frequency (Francis & KucÏ era, 1982) of all words containing the segment in any position (this was done by searching the American Heritage Electronic Dictionary, 1992). T hese stimulus characteristics are presented in Table 4. ML was tested twice on these stimuli (mixed presentation of syllable and control items) in sessions separated by at least one week. As can be seen in Table 5, ML’ s level of performance was similar across the two sessions (S ession 1: 51% correct; Session 2: 47% correct), and he read the syllable nonwords at a much higher level (overall: 71% correct) than the control nonwords (overall: 28% correct). T he difference in performance (D ) was signi® cant for Session 2 1, D = 39% (24/ 34 vs. 11/ 34), x = 9.95, p < .01, as well as for Session 2, D = 47% (24/ 2 34 vs. 8/ 34), x = 15.11, p < .01. Lexicalizations were the most common error type: 5/ 10 and 6/ 10 errors for syllable stimuli for S essions 1 and 2, respectively, and 14/ 23 and 20/ 26 errors for the control stimuli. To investigate further ML’s reading of syllabic units from words, we constructed a second set of stimuli that corresponded to medial or word-® nal segments (e.g. syllable: ``vorce’ ’ from ``divorce’ ’ ; control: ``droom’ ’ from ``bedroom’ ’). T hese nonwords were also matched on bigram frequency, N, and summed frequency of items containing the segment (see Table 4). T here were 27 syllable nonwords and 27 control nonwords. ML was tested twice on these stimuli (mixed presentation of syllable and control items). As can be seen in Table 5, ML’ s level of performance was again similar across the two test sessions (though a bit lower in session 2), and his level of performance was much higher for the syllable than for the control nonwords. T he difference in performance was signi® cant for TABLE 4 Stimulus Characteristics of Syllable and Control Nonwords Used in Reading and Repetition Bigram F req N Summed Freq M SD M SD M SD Word-Initial syllable control 37.83 40.66 39.78 41.58 3.79 5.09 2.95 2.99 56.32 60.71 115.70 129.43 Non-initial syllable control 40.49 43.89 28.50 35.97 4.44 4.74 3.29 3.69 104.19 101.00 186.05 178.67 THE REPRESENTATION OF SUBLEXICAL CORRESPONDENCES 919 TABLE 5 ML’s Performance in Reading and Repetition of Nonwords of Different Types Downloaded by [University of Florida] at 18:09 05 November 2015 Percentage Correct Read 1 Read 2 Repeat 1 Repeat 2 Word-initial syllable control 71 32 71 24 74 65 82 79 Non-initial syllable control 81 30 67 19 67 85 74 89 Body 16 24 84 89 Onset 9 23 77 91 Session 1, D = 51% (22/ 27/ vs. 8/ 27), x = 14.70, p < .01, as well as for Session 2, D = 2 48% (18/ 27 vs. 5/ 27), x = 12.80, p < .01. For Sessions 1 and 2 there were 4/ 5 and 8/ 9 lexicalization errors for the syllable stimuli and 15/ 19 and 14/ 22 for the control stimuli. Level of performance was similar for the word-initial and nonword-initial stimuli. 2 Onset± body U nits. ML was also asked to read nonwords corresponding to onset or body units in real words (e.g. ``bl’ ’ and ``ack’ ’ from the word ``black’ ’ ). He was tested twice on these stimuli in test sessions at least one week apart. Although ML’s nonword reading improved when the nonwords to be read corresponded to syllables within real words, ML’s nonword reading did not seem to bene® t from having the nonwords correspond to onset units (S ession 1: 2/ 22 correct; Session 2: 5/ 22 correct) or body units (Session 1: 4/ 25 correct; Session 2: 6/ 25 correct). Error responses for the onset stimuli were generally words beginning with the consonant cluster (e.g. tr ® tree; gr ® grow), but were occasionally the consonant cluster with a vowel attached (e.g. cl ® ``clo’’ ; pl ® ``ploo’ ’ ). T his suggests that ML may have been using a lexical strategy to read these items. T hat is, he may have been able to come up with a word beginning with the given onset, which he then attempted to segment down to the appropriate unit. ML’s de® cits in the segmentation tasks reported above suggest (consistent with his performance on the reading of onset units) that this would not be a very successful strategy). For the body stimuli, there was a variety of error types. some examples of errors include: orse ® ``orst’ ’; atch ® ``attach’ ’ ; outh ® ``use’ ’ . . . ``u’ ’ . . . ``u’ ’ . . . ``out’ ’ ; urt ® ``hurt’ ’; udge ® ``u’ ’ . . . ``crunch’’ ; eath ® ``eat’ ’ . Processing of Letters. ML does fairly well in naming individual letters, obtaining 23/ 26 correct. When given letter sounds, ML was able to provide the corresponding letter name on 15/ 22 trials. When provided with the written letter, ML produced an appropriate sound for only 9/ 26 letters. It is interesting to note that ML continued to have dif® culty even when it was suggested that he think of a word beginning with the given letter and then segment off the initial sound (see Funnell, 1983). 920 LESCH AND MARTIN Downloaded by [University of Florida] at 18:09 05 November 2015 Performance of Control Subjects on Nonword Reading We expected that control subjects would make few errors in unspeeded nonword reading. T herefore, we recorded reaction times to name the nonwords that ML had been asked to read. Subjects saw nonwords one at a time in the middle of a computer screen and were asked to name the nonword as quickly and as accurately as possible. Presentation was blocked by type of nonword (as was the case for ML ). T he mean error rate was 3% . T he reaction time data are presented in Table 6 along with the percent errors for ML. T he data were subjected to a one-way analysis of variance with nonword type as a withinsubject factor. T here was no main effect of nonword type; F(5, 20) = 1.98, p = .13. Reaction times appeared to be longest in the onset condition. T his is not surprising, given that we are not normally asked to read consonant clusters. It can be seen that ML had the most dif® culty with the onset unitsÐ the units that produced the longest naming times for controls. Otherwise, there was little correspondence between ML’s performance and that of the controlsÐ ML was signi® cantly worse at reading control nonwords than syllable nonwords, whereas for the control subjects, there was little difference between non-initial syllables and their controls (reaction times were 10 msec longer to controls) and the difference was in the opposite direction for the initial syllables and controls (21msec difference). However, these differences did not approach signi® cance. Additionally, it appeared that body units had the shortest latencies. If latency can be assumed to re¯ ect dif® culty of processing, then this result would be inconsistent with the idea that ML suffers from a global decrement to the sublexical reading process in which it might be expected that the most dif® cult tasks would be the most impaired. Conversion without Output In order to assess whether ML’s nonword reading de® cit is related to a dif® culty with production, we assessed ML’s ability to convert from orthography to phonology in a task that does not require output. In this task ML was presented with each item that he had TABLE 6 Performance of Control Subjects on Nonword Reading a Reaction Time I± Syll Control N I± Syll Control Body Onset FM PH MB MW PS 675 578 667 700 510 688 542 650 656 488 764 546 662 660 463 781 572 652 676 464 570 492 647 679 485 758 604 651 669 549 Mean Range 626 510± 700 605 488± 688 619 463± 764 629 464± 781 575 485± 679 646 549± 758 29 72 26 76 80 84 b ML Errors a In msec. b In percentages. Downloaded by [University of Florida] at 18:09 05 November 2015 THE REPRESENTATION OF SUBLEXICAL CORRESPONDENCES 921 been asked to read earlier (syllable vs. control, onset, and body) and was given a pronunciation that either did or did not correspond to the nonword. Incorrect pronunciations differed from the correct pronunciation by one phoneme, and the position of the difference was varied. Each item was paired with an incorrect and a correct pronunciation. ML was asked to judge whether or not a pronunciation was correct. If ML’s reading impairment is related to a dif® culty with production, then he should perform at a relatively high level on this task. Contrary to this prediction, ML performed quite poorly. Items were scored as correct if he rejected the incorrect pronunciation and accepted the correct pronunciation. ML obtained 7/ 25 correct for the body units and 7/ 22 correct for the onsets. He obtained 7/ 27 correct for non-initial syllables against 3/ 27 correct for their controls, and 14/ 34 correct for the initial syllables against 4/ 34 correct for their controls. T hus, the pattern of performance on this judgement task was similar to the pattern on the reading tasks. T he similar pattern of performance in a task not requiring output suggests that ML’s poor nonword reading cannot be accounted for in terms of an output problem. ML’s low level of performance on this task (which stemmed from his willingness to accept incorrect pronunciations) suggests that there may be a high level of ``noise’ ’ in the reading system. Summary: Orthographic± Phonologic Conversion ML performs poorly when asked to produce the sound of a written letter or to read nonwords sharing onset- or body-status with real word segments. However, ML shows a clear advantage in reading nonwords when they correspond to a syllable within a real word relative to when they form the initial segment of real words but do not have syllable status within a real word. T his difference emerged even though the two types of stimuli (syllable vs. control) were matched on several variables. T he results suggest that at least one sublexical unitÐ the syllableÐ plays a role in ML’ s nonword reading. An account in terms of an output problem seems unlikely given the similar pattern of performance on the conversion without output task. Repetition of Nonwords ML’s greater accuracy in reading single-syllable nonwords that correspond to syllables within real words compared to those that do not correspond to syllables within real words might possibly be related to his short-term memory de® cit. T hat is, in attempting to produce a nonword, it is possible that he has better ability to hold onto the phonological representation of syllables in real words because these syllables activate lexical representations to some extent and therefore feedback from the lexical level could help to maintain these units in short-term memory (see Martin & S affran, 1992a, 1992b). Also, units corresponding to syllables of real words might be more easily articulated because of prior practice in producing these syllables. To address these issues, ML was asked to repeat the control and the syllable nonwords. If short-term memory is the source of the difference in reading the two types of nonwords, he should show the same pattern in repetition. However, it should be noted that repetition is an easier task than is reading because the phonological speci® cation of the nonword is providedÐ that is, the phonological system is more strongly activated by spoken input than by written input (see Patterson Downloaded by [University of Florida] at 18:09 05 November 2015 922 LESCH AND MARTIN et al., 1996). For this reason alone, ML’s repetition performance may not mirror his reading performance. However, if ML’s repetition performance is not at ceiling, and his reading dif® culties are due to articulatory or short-term memory problems, then there should be some relation between his reading and repetition performance. ML was tested twice on repetition of the nonwords that he had been asked to read. As can be seen in Table 5, ML’ s repetition performance was not at ceiling. However, there appears to be little relation between reading performance and repetition performance: For the wordinitial syllable and control nonwords, performance was slightly better for syllables than for control nonwords (over session: syllable 78% vs. control 72% ). However, for the nonword-initial stimuli, the difference was in the other direction (over session: syllable 70% vs. control 87% ). Additionally, repetition performance was very high for those items causing ML the most dif® culty in reading: the onset and body units. T he difference in patterns of performance across reading and repetition suggests that ML’s nonword reading pattern should not be attributed to short-term memory or articulatory factors. All ® ve control subjects repeated all the word-initial syllable and control nonwords correctly and three control subjects repeated all the nonword-initial syllable and control words correctly. One subject made one lexicalization error on a syllable item, and another made one error on a syllable item and two lexicalization errors on control items. None of the control subjects made any errors on repetition of onset and body stimuli. ML’s performance on the tasks we have presented thus far is consistent with the suggestion that his nonword reading impairment is due to a disruption to certain levels of sublexical representation and not to a disruption to processes underlying assembled phonologyÐ ML is able to segment both orthographic and phonological representations when the relevant unit is the syllable, but performs quite poorly when the relevant units are onsets and bodies or phonemes. Furthermore, ML’ s nonword reading does not bene® t from having the nonwords correspond to onset or body units from real words, but does bene® t from having them correspond to syllables within real words. Phonological Blending (Auditory Presentation) We have evidence that the processes involved in the ® rst two stages of assembled phonology are relatively intact in ML: He is able to segment orthographic and phonological representations, and he is able to convert orthographic units to phonological units. However, his performance depends upon the size of the units involvedÐ performance is substantially better with syllable units than with smaller units. In the third stage suggested by Beauvois and Derouesne (1979), the phonemic stage, the sounds obtained in Stage 2 are assembled into the phonological representation for the written stimulus (see Figure 1). We therefore assessed ML’ s ability to blend different phonological units to produce words. We presented ML with auditory words segmented into individual phonemes, syllables, or onset± body units. We explained to ML that he would hear words being sounded out (sound by sound, syllable by syllable, or split into onset± body units) and that he was to report the word resulting from the blending of these units. He was given several practice trials and was told the appropriate responses for these trials in an attempt to ensure that he understood the task. THE REPRESENTATION OF SUBLEXICAL CORRESPONDENCES 923 Downloaded by [University of Florida] at 18:09 05 November 2015 Syllables. ML performed at a high level when the segments to be blended were syllables (19/ 20 correct). His one error on the syllable-blending trials was making the 2 response ``celery’’ to the syllables contained within ``salary’ ’ . Onset± body units. ML was also given 49 words with a CCVCC structure (e.g. BLACK). T hese words were divided at their onset± body boundary (``bl’ ’ - ``ack’ ’ ) and ML was asked to blend the segments to produce a word. ML obtained 20/ 49 correct. Of the erroneous responses, 7/ 29 were failures to respond. Other error types included two-word responses (2/ 29) (e.g. ``gr’ ’ - ``unt’ ’ ® ``great aunt’ ’ ; ``pl’ ’ - ``uck’ ’ ® ``pull-up’’ ) and multisyllabic responses (11/ 29) (e.g. ``pr’ ’ - ``ess’’ ® ``purchase’’ ) that generally bore some phonological relationship to the target item. In order to investigate further the nature of ML’ s dif® culties with this phonological blending task, a second version of the task was administered that would better allow for the determination of whether ML’s performance varies with type of onset. One possibility is that ML has dif® culty blending onsets that require schwa-deletion before blending (i.e. stops). In this task all stimuli had a CVC or CVCC structure (rather than CCVCC) as there are fewer constraints on phonemes that can occur in word-initial position than on those that can occur as the second consonant in a word-initial consonant cluster. T his would allow us to vary more systematically the type of onset involved. We used a set of 15 bodies in combination with six different types of onset (nasal, stop, affricate, fricative, liquid, and glide) resulting in a total of 90 onset± body pairs. T he different onset± body pairings were presented in pseudo-random fashion. T he stop, affricate, and glide onsets all require schwa-deletion before blending, whereas the nasal, fricative, and liquid onsets do not. T herefore, if it is schwa-deletion that is causing ML’s dif® culty, then he should perform at a much higher level with the nasal, fricative, and liquid onsets than with onsets that require schwa-deletion. ML obtained 15/ 90 (17% ) correct: 10/ 45 correct for onsets requiring schwa-deletion and 5/ 45 correct for onsets not requiring schwa-deletion. T herefore, we obtained no evidence that ML’s poor performance in onset± body blending is related to a dif® culty with schwa-deletion. T he majority (53/ 75) of ML’ s erroneous responses consisted of failures to respond. For 8/ 75 errors ML produced a multisyllabic or multi-word response (this type of response occurred for all types of onset). Occasionally, error responses shared an onset (wives ® walk) or a body (meek ® week). T he discrepancy in performance between the two tasks (Task 1: 20/ 49; Task 2: 15/ 90) merits some comment. In Task 1, because of the structure of the targets, the second consonant of the consonant cluster was repeated over many trials (i.e. there is a limited set of consonants that can occur in that position). T his could have bene® ted ML’ s performance in Task 1 relative to Task 2. Another possibility is that ML’ s perception or memory is more stable for consonant clusters than for individual phonemes. Again, this would result in a bene® t for Task 1 performance relative to T ask 2 performance. 2 Although we have classi® ed ``celery’ ’ as an erroneou s respons e to the syllables contained within ``salary’ ’, it may be the case that the response was actually correct and re¯ ects differences between the examiner’s New England accent and ML’s Texan accent. 924 LESCH AND MARTIN Phonemes. ML failed to make a single correct response on the phoneme-blending trialsÐ indeed, he was very unwilling to make any response at all. His complete inability to perform this task is evident by his response on one trial when he was pressed to respondÐ to the phonemes contained within ``tree’ ’ , he responded ``rough’’ . Downloaded by [University of Florida] at 18:09 05 November 2015 Performance of Control Subjects on Phonological Blending (Auditory Presentation) As can be seen in Table 7, ML performs at a near-normal level when asked to blend syllables, but performs well below the range of the control subjects when asked to blend onsets and bodies or phonemes. No individual control subject showed a pattern similar to that of ML. Orthographic/Phonological Blending (Visual Presentation) In order to provide a test of ML’ s ability to blend orthographic/ phonological units, we decided to test ML on the anagrams task introduced by Treiman and Chafetz (1987). In this task, the subject sees four word fragments (e.g. BL OFT ACK FR) and is asked if any two fragments can be joined to form a word. Although the units to be blended in this task are presented in the visual modality, we would argue that performance of the task may rely upon the manipulation of phonological units. Treiman and Chafetz found that subjects were quicker to ® nd the words when the segments corresponded to onset and body units than when they did not correspond to these units, which suggests that onset and body units are functional in visual word recognition (see also Bowey, 1990; Treiman & TABLE 7 A Comparison of the Performance of ML and Control Subjects on the Phonologica l Blending (with auditory presentation ) Tasks Percentage of Items Correct Control Subjects Syllable Onset-body Phoneme PS PH MB MW TP HF 100 100 100 100 100 100 100 98 100 100 96 100 100 85 80 100 85 95 Mean Range 100 N/ A 99 96± 100 91 80± 100 ML 95 41 0 Downloaded by [University of Florida] at 18:09 05 November 2015 THE REPRESENTATION OF SUBLEXICAL CORRESPONDENCES 925 Zukowski, 1988). We were interested in determining whether there would be a similar bene® t to performance when the units to be combined correspond to syllables within real words. Data from age-matched controls would be relevant to the debate as to whether syllabic units are generally (i.e. in non-brain-damaged populations) functional in reading. T he use of this task with ML would provide an opportunity to observe an effect of sublexical structure on the processing of real words and would thereby provide a test of the hypothesis that syllabic units are functional in ML’ s visual word recognition processes. Compared to the task of blending auditorily presented phonological segments, we view the anagrams task as more directly relevant to the issue of the role of sublexical phonology in reading. If ML’ s performance bene® ts from having the fragments to be combined correspond to syllable units, then we would have more evidence that the syllable effects we have observed in reading with this patient are not speci® c to the processing of nonwords. If control subjects show a similar advantage, we would have evidence that the syllable effect demonstrated by ML does not re¯ ect the pathological operation of a disrupted sublexical route. Although the primary focus in this experiment is on the processing of the syllabic stimuli, onset± body trials were also included. ANAGRAMS TASK Method Subjects ML and eight control subjects of about the same age and education level served as subjects in this experiment. Stimuli and Design Each subject received 288 trials. On each trial the subject was presented with four word fragments and was asked whether the joining of any two of these fragments resulted in a word. Trials requiring a ``yes’ ’ response were of two types: experimental and control. On experimental trials, syllable stimuli (two-syllable words) were split at the syllable boundary, whereas onset-body stimuli (monosyllabic words with a CCVCC structure) were split at the onset± body boundary. T he syllable stimuli had a mean word frequency (Francis & KucÏ era, 1982) of 53 (SD = 114.96), whereas the onset± body stimuli had a mean word frequency of 18.29 (SD = 46.24). On control trials, words were split so as to violate syllable or onset± body boundaries (see Table 8 for an example of the stimulus materials). Although two of the four fragments could be joined to form a word (target fragments), neither of the other two fragments (distractor fragments) could be combined with any other fragment to form a word. Order and position of the fragments was varied and was matched across the two stimulus types. Distractor fragments were formed by changing one or two letters in a real word and splitting them in a manner similar to the way in which the target word was split. Words were present on 192 of the trials: 48 trials in each of the four conditions outlined above. T he other 96 trials required a ``no’ ’ response and were modelled closely on the ``yes’ ’ trials. ``No’ ’ trials were formed by taking an equal number of two-syllable words and monosyllabic words with a CCVCC structure and changing one or two letters to form a nonword. T hese nonwords were then split at the syllable or onset± body boundary. 926 LESCH AND MARTIN TABLE 8 Sample Stimuli for the Anagrams Task Trial Type Stimulus Type Syllable Downloaded by [University of Florida] at 18:09 05 November 2015 Onset± Body Experimental Control (FABRIC ) (CASKET ) (CYNIC) FAB TEL GIS RIC FA ST EL GI BRIC T OB PIS CAS KET ST OB PI CA SKET LE IC CYN SAK KL E NIC CY SA (CRANK ) (BLACK ) (T RICK) CR S K BRO ANK CRA OSK BR NK ONT GR BL ACK NT GRO BLA CK ALK ICK TR FR LK CK T RI FRA Procedure Stimulus presentation was controlled by a Macintosh IIcx computer. On each trial subjects were presented with four fragments and were asked whether any two could be joined to form a word. T hey were instructed to make their decision as quickly and as accurately as possible and to press one of two keys to indicate a ``yes’ ’ or a ``no’ ’ response. T hey were further instructed that if their responses were ``yes’’ , they should indicate the word to the experimenter, who then recorded it. T he experiment was run in two separate sessions; items that appeared in the experimental condition in the ® rst session appeared in the control condition in the second session (and vice versa). T he two sessions were separated by at least one day. Results and DiscussionÐ Error Data Both the error data and the reaction time data reported here are for the trials containing 3 target words. Table 9 shows the frequency of errors (failures to ® nd the target word) as a function of stimulus condition. In scoring errors, both the initial ``yes/ no’ ’ response and the correctness of the reported word were taken into account. Mean reaction times are also reported (see later). Due to the yoked nature of the design, an error to one member of an experimental± control pair (e.g. BL ACK vs. BLA CK ) necessitated the removal of the other member from the reaction time analysis (this resulted in the loss of 87% of ML’ s reaction time data). T hus, the error data tended to be more reliable. T herefore, we will focus on these data. 2 3 2 (stimulus type: onset± body vs. syllable; trial type: control vs. experimental) analyses of variance over items and planned comparisons were conducted on the 4 error data for each subject. As can be seen in Table 9, seven out of eight subjects showed a signi® cant advantage on trials where words were split at the syllable boundary relative to when the split violated the syllable boundary. Whereas seven of the eight control subjects 3 False- positive response s (i.e. reporting a word where there was none) were extremely rare, comprising fewer than 2% of response s on ``no’ ’ trials. 4 We have chosen to use ANOVA to analyse our error data on the grounds that it has been demonstrate d that violations of the normality assumption appear to have little effect on the level of signi® cance of the F- test (e.g. Glasshouse , Peckham, & Sanders, 1972; Winer, Brown, & Michels, 1991). As reported by Glass et al. (1972), data from Lunney (1970) suggest the robustnes s of the F-statistic even when the dependent variable is dichotomous (see also Cochran, 1950). 927 THE REPRESENTATION OF SUBLEXICAL CORRESPONDENCES TABLE 9 a Mean Number of Errors on the Anagrams Task as a Function of Stimulus Type and Trial Type Syllable Downloaded by [University of Florida] at 18:09 05 November 2015 Control Subjects Onset± Body Control Experimental Difference Control Experimental Difference MB JC BC PH SM LS PS MW 18 14 21 35 29 24 8 8 3 4 18 18 16 9 2 1 15* 10* 3 17* 13* 15* 6* 7* 10 2 15 14 19 16 2 2 2 5 11 8 15 8 1 1 Mean Range 20 8± 35 9 1± 18 11* 3± 17 10 2± 15 6 1± 15 2 3± 8 ML 46 24 22* 36 26 10* 8* 2 3 4 6 4 8 1 1 4* a Out of a possible 48. * Denotes signi® cance at p < .05. showed some indication of an advantage on trials where words were split at the onset± body boundary, this difference reached signi® cance for only one individual control subject. T he S timulus Type 3 Condition interaction was signi® cant for subjects JC and PH, re¯ ecting a larger advantage of the experimental trials over the control trials for syllable stimuli than for onset± body stimuli. T he interaction approached signi® cance for two other subjects: MW and PS (p < .06). (However, see later for a discussion of problems associated with making this comparison.) T he error data from the control subjects were also analysed at a group level using a 2 3 2 analysis of variance with the within-subject factors stimulus type (syllable vs. onset± body) and trial type (control vs. experimental). T here was a main effect of stimulus type such that more errors were made to syllable stimuli (mean = 14.5) than to onset± body stimuli (mean = 8), F(1, 7) = 18.59, p < .005. T his difference suggests that the somewhat higher frequency of the syllable stimuli did not aid performance relative to the onset± body stimuli. T here was also a main effect of trial type re¯ ecting higher error rates on control trials than on experimental trials (see Table 9), F(1, 7) = 26.61, p < .005. T he difference between the control and experimental trials for the syllable stimuli was signi® cant, F(1, 7) = 36.46, p < .005, as was the difference for the onset± body stimuli, F(1, 7) = 7.22, p < .05. Finally, there was a Stimulus Type 3 Trial Type interaction, F(1, 7), p < .005, indicating a larger bene® t of the experimental over the control trials when the units to be combined were syllables than when they were onset± body units. T hus, these data replicate those of Treiman and Chafetz (1987) in that we observed a signi® cant advantage when units to be combined formed onset± body units than when they did not. However, the current results go beyond those of Treiman and Chafetz in that they also indicate an advantage when units to be combined form syllables within Downloaded by [University of Florida] at 18:09 05 November 2015 928 LESCH AND MARTIN words. Interestingly, the bene® t appeared to be greater for syllable units than for onset± body units. However, there is some dif® culty in interpreting this difference as there was a much higher level of performance on the control onset± body trials than on the control syllable trials. T here are at least two factors that could have contributed to a higher level of performance on the onset± body trials: (a) As there is a relatively limited set of wordinitial and word-® nal consonant clusters, these trials contained units that were repeated within the experiment, and (b) there are phonotactic constraints on which consonant clusters can occur in word-inital or word-® nal positions (for example, ``bl’ ’ never occurs in word-® nal position, whereas ``ck’ ’ never occurs in word-initial position). T herefore, for the onset± body trials it may have been easier to determine the ordering of units. Although these factors may have contributed to a higher level of performance on onset± body trials and may compromise a comparison of onset± body with syllable performance, they do not compromise the main ® nding that the control subjects showed a large syllable effect. ML demonstrated a pattern similar to that of the control subjects (see Table 9): He was more likely to ® nd the target word when the units to be combined were syllables or onsets and bodies than when they were not. For the syllable stimuli, the advantage for the experimental over the control condition was signi® cant, t(94) = 6.13, p < .01, as was the advantage for the onset± body stimuli, t(94) = 2.51, p < .05. T he difference between these differences was signi® cant, F(1, 94) = 6.56, p < .01, indicating that ML was more likely to ® nd the target word when the units to be combined were syllables than when they were onset and body units. T hus, as was the case with the segmentation and nonword reading tasks, ML performs at a higher level with syllabic units than with units that do not correspond to syllables within words. Furthermore, the control subjects showed a very similar pattern. However, the advantage for the control subjects (mean difference = 7) was not as large as that for ML (12-item advantage). Furthermore, as can be seen from Table 9, the control subjects performed substantially better than ML on this task. Results and Discussion± Reaction Time Data T he reaction time data are presented in T able 10. T he pattern of results essentially replicated that of the error data, but was less reliable: For the syllable stimuli, seven out of eight control subjects were faster to ® nd the target word when the units to be combined were syllables than when they were not. However, this difference only reached signi® cance for three of these subjects (see Table 10). For the onset± body stimuli, six out of eight control subjects were faster to ® nd the target word when the units to be combined were onset± body units than when they were not, although this difference only reached signi® cance for two of these subjects (see T able 10). T he interaction between stimulus type and trial type failed to reach signi® cance for any individual control subject. Although we did not analyse ML’s reaction time data (due to the large number of missing observations), we computed his mean reaction time (over condition) on correct ``yes’ ’ responses to be about 17 sec. T his indicates that ML was much slower in this task than were the control subjects (see Table 10). T he reaction time data from the control subjects were also analysed at a group level using a 2 3 2 analysis of variance with the within-subject factors stimulus type (syllable vs. onset± body) and trial type (control vs. experimental). T here was no main effect of 929 THE REPRESENTATION OF SUBLEXICAL CORRESPONDENCES TABLE 10 a Mean Reaction Times on the Anagrams Task as a Function of Stimulus Type and Trial Type Downloaded by [University of Florida] at 18:09 05 November 2015 Syllable Onset± Body Control Subjects Control Experimental Difference Control Experimental Difference MB JC BC PH SM LS PS MW 11.81 5.81 7.56 3.48 3.98 4.36 3.78 4.28 9.27 4.62 6.24 3.87 3.34 2.75 2.79 3.08 2.54 1.19 1.32 2 0.39 0.64 1.61* 0.99* 1.20* 11.58 4.11 6.97 4.66 2.93 4.56 3.16 3.38 7.58 3.26 5.29 4.70 3.15 3.18 2.82 3.23 4.00* 0.85 1.68 2 0.04 2 0.22 1.38* 0.34 0.15 Mean 5.63 4.50 1.14* 5.17 4.15 1.02 a In sec. * Denotes signi® cance at p < .05. stimulus type, F(1, 7) = 2.14, p = .19, but there was a main effect of trial type such that subjects were 1.08 sec faster to ® nd the target word on experimental trials than on control trials, F(1, 7) = 8.16, p < .05. Finally, the Stimulus Type 3 Trial Type interaction failed to reach signi® cance, F < 1. Summary of Blending Results In the phonological blending task with auditory presentation, ML was completely unable to blend phonemes, yet had little dif® culty blending syllables. ML showed some ability to blend onset and body units, though he performed quite poorly. Control subjects demonstrated little dif® culty in performing any of these tasks. In the anagrams task, ML’ s performance showed a bene® t of having the units to be combined correspond to onset± body units and syllable units. T hus, as was the case with the segmentation and reading tasks, ML’s performance on blending tasks also demonstrated a bene® t of syllable status. T he pattern of results across the control subjects was quite similar to that shown by ML. T herefore, we have obtained evidence that syllabic representations are more generally involved in visual word recognitionÐ that is, the syllable effect does not appear to represent the pathological operation of a disrupted sublexical route. However, although ML’s pattern of performance was similar to that of the control subjects, it should be noted that his overall level of performance never fell within the range of the control subjects. One possible explanation for ML’s lower accuracy is that the anagrams task stressed ML’ s short-term memory (ST M) capacity more so than that of the controls. In order to be able to perform this task, it is necessary to be able to retain each segment in memory while it is paired with each of the other segments. As we have indicated previously, ML suffers from a severe ST M impairment. Both his word and his nonword span are between two and three items with auditory presentation (see Martin & Lesch, 1996). Furthermore, while ML’s phonological retention 930 LESCH AND MARTIN appears to be better than his semantic retention, it is clearly not normal. T hus, holding onto multiple segments (segments that are meaningless prior to blending), would be expected to be dif® cult for ML. T his impairment may also explain ML’s extremely long reaction times in this taskÐ if ML has dif® culty retaining the segments, he may need to rely on the printed stimulus to refresh the phonological trace of the stimulus. T his would be assumed to be a time-consuming and dif® cult process given ML’s impairment in nonword reading. Downloaded by [University of Florida] at 18:09 05 November 2015 GENERAL DISCUSSION We investigated a phonological dyslexic’s ability to segment written and auditory stimuli, to read word and nonword stimuli, and to blend auditorily and visually presented sublexical segments to produce a word. ML’ s pattern of performance on these tasks suggests that his nonword reading impairment is not due to a disruption of the processes subserving sublexical phonology per se, but rather to disruption of some of the representations upon which these processes operate. ML showed dif® culty in processing phonemes across all tasks (segmentation, orthographic± phonological conversion, and blending). We observed consistent bene® ts of syllable status across all tasks and observed relatively little bene® t of onset or body status. A comparison of ML’s performance with that of matched control subjects argues against the hypothesis that ML’ s dif® culties re¯ ect a global decrement, with the most dif® cult tasks the most impaired. Furthermore, it does not seem to be the case that the ``word-likeness’’ of the units can explain ML’s pattern of performance. ML performs better with the syllabic units than with bodies. It’s not clear how body units could be argued to be less ``word-like’ ’ than syllable units, especially as they would tend to overlap with a larger portion of a real word. Also, performance was better for syllable than for control nonwords even though these items were matched on variables that should re¯ ect ``word-likeness’ ’ . Based on our results with ML, and the observation of onset, body, and syllable effects in normals, we attribute ML’s de® cit in nonword reading to the disruption of orthographic and phonological representations at levels below the word and the syllable (i.e. onset, body, and phoneme). However, it should be noted that phonemic and onset± body representations do not seem to have been completely abolished as ML does show a limited ability to process these units. Additionally, ML’ s processing of syllables, although good, was not perfect. Possibly, ML’ s processing of syllabic representations suffers somewhat from the lack of support from lower level representations. We attribute ML’ s poor nonword reading to a disruption of representations below the level of the syllable. However, the possibility that a blending impairment forms the basis of ML’ s dif® culties in nonword reading merits further comment. Blending impairments are commonly associated with phonological dyslexia (see, e.g. Berndt et al., 1996; Patterson & Marcel, 1992). Furthermore, the pattern demonstrated by ML on blending tasks is somewhat similar to other patients reported in the literature with syllable blending being less affected than phoneme blending (see Berndt et al., 1996). However, our arguments regarding ML’ s reading impairment are not based solely on the blending results. Across all tasks, ML’ s performance bene® ts from having the units correspond to syllables within real words. How would a blending impairment explain ML’ s differential performance in reading the syllable and control nonwords? Both types of items correspond to segments of Downloaded by [University of Florida] at 18:09 05 November 2015 THE REPRESENTATION OF SUBLEXICAL CORRESPONDENCES 931 real words and both form syllables (although the control nonwords did not map onto syllables within real words). Additionally, an explanation in terms of a blending impairment presumes that conversion from print to sound is intact. However, ML showed considerable dif® culty in producing the sounds corresponding to individual letters. It would also be dif® cult to explain this pattern in terms of a more generalized phonological de® cit (see Farah, S towe, & Levinson, 1996; Friedman, 1996; Patterson et al. 1996; Sasanuma, Ito, Patterson, & Ito, 1996). One dif® culty with our interpretation regarding preserved syllabic representations for ML is that, as reviewed earlier, the results of studies examining syllabic effects in written word processing have been extremely inconsistentÐ some studies indicate an effect whereas many others do not. In the current study we have obtained strong evidence from both our patient and from control subjects for the importance of syllabic representations in reading. What is the basis of this inconsistency? One possibility is that, in a fully functioning system (with intact representations at all levels), syllable effects are washed out by effects of other levels of representation. T his might provide an explanation for why we observed a syllable effect for ML, but we also observed a syllable effect with control subjects. Interestingly, Balota and Ferraro (1993) have observed an increased incidence of regularization errors in the speeded naming performance of healthy older adults and individuals with senile dementia of the Alzheimer type. Although Balota and Ferraro suggest that the increase in regularization errors is the consequence of a breakdown in the inhibition of output from the assembled route, it would be of interest to ascertain whether these results could be accommodated by a shift, with age, in the units which dominate the assembly process (e.g. a shift from larger to smaller units). T hat is, perhaps the conversion process characteristic of the older adults in the present study is dominated by different units from those of the younger adults that generally serve as subjects in psychological experiments. A less interesting explanation for the inconsistency of syllable effects is that studies may employ different de® nitions of the syllableÐ different linguistic theories locate syllable boundaries differently (see Seidenberg, 1989). Our results are consistent with the approach taken by Shallice et al. (1983; see also Marcel, 1980). In their model, stored spelling± to± sound correspondences exist for various sizes of orthographic units including graphemes, consonant clusters, subsyllabic units, syllables, and morphemes. According to this view, the visual word-form system (see Warrington & Shallice, 1980) detects letter groups of various sizes and then transmits this information to corresponding units in the phonological system. Spelling± to± sound conversion is assumed to take place at a number of levels in parallel, and although higher levels are assumed to be available more rapidly, they are also assumed to be more sensitive to neurological degeneration. T hus, Shallice et al. (1983; S hallice & McCarthy, 1985) propose that, as a result of brain damage, larger units of correspondence should drop out before lower levels. T heir case study of the surface dyslexic patient HT R supports this claim. However, we are suggesting that the pattern of performance that we have observed in our phonological dyslexic patient, ML, is due to the selective impairment of levels of orthographic± phonological representation below the level of the syllable (with intact syllabic and whole word representations). T herefore, it would seem that larger units are not necessarily the ® rst to be disrupted as a result of brain damage and that a multiple-levels model can provide an account for both surface dyslexia and (at least some instances of) phonological dyslexia. Downloaded by [University of Florida] at 18:09 05 November 2015 932 LESCH AND MARTIN In Norris’ (1994) implementation of Shallice et al.’ s (1983) model, different levels of spelling± to± sound correspondence (e.g. phoneme, body, word) are combined in an interactive activation network to produce a pronunciation of the input. In this model there are six levels of orthographic input units: initial consonant clusters, vowels, ® nal consonant clusters, initial cluster plus vowel, vowel plus ® nal cluster (body), and words/ syllables (the syllable level corresponds to the word level when only monosyllabic words are used). T here are three sets of output nodes: initial consonant/ cluster, the vowel, and the ® nal consonant/ cluster. As a result of competition among all possible pronunciations at all levels, a single unambiguous pronunciation of the input is produced. Parameters were set by ® tting the model to the data of T araban and McClelland (1987) and Waters and Seidenberg (1985). In simulations of eight experiments (from Andrews, 1989; Brown, 1987 ; Glushko, 1979; Jared, McRae, & S eidenberg, 1990; Kay & Bishop, 1987; Seidenberg et al., 1984) the model accurately predicts effects of word frequency, regularity, and consistency. Presumably, it would be relatively straightforward to simulate ML’s performanceÐ nodes corresponding to subsyllabic representations (which are explicitly represented) would have to be disrupted. T he approach taken by Seidenberg and McClelland (1989) is similar to that of Norris in the sense that there are no distinct mechanisms corresponding to sublexical and lexical phonology. However, the two approaches differ in how they represent orthographic± phonological correspondences: Norris’ (1994) model represents these correspondences ``locally’ ’ Ð that is, there are indiv idual nodes that represent different units (e.g. graphemes, morphemes, words). Seidenberg and McClelland’ s (1989) model, on the other hand, represents the relationship between orthography and phonology in a more distributed fashion. (See Besner, in press, for a discussion of single, dual, and three-route models of reading and for a discussion of localist versus distributed representations.) In their model, the mapping from orthography to phonology is accomplished via a layer of ``hidden’ ’ units between orthographic input and phonological output layers. During learning, sets of nodes become dedicated to certain representations. T he hidden units are necessary because with a two-layer network that maps graphemes to phonemes there is no set of weights that will work for all the training patterns. T he hidden units serve to ``rerepresent’ ’ the input patterns by distributing activation over sets of units with the result that similarity across input patterns is decreased (see Plaut et al., 1996). It is dif® cult to see how these hidden units could be damaged to produce ML’ s pattern of performance. In these models there is no special status afforded to speci® c levels of representationÐ the network merely encodes statistical regularities of the language. ML’ s nonword reading bene® ted when the units corresponded to syllables within real words relative to when the units formed segments of real words but did not map onto a syllable within real words. T his bene® t was observed even though the two types of stimuli were matched on a number of variables to which distributed models would be expected to be sensitive: summed frequency of words containing the segment, number of neighbours, and bigram frequency. We suggest that the representations upon which the assembled route operates should not be thought of as ``non-lexical’ ’ but rather as ``sub-lexical’ ’ Ð that is, derived from lexical items. We propose that an intact reading system contains multiple levels of representation for words that it has previously encountered. When a new word is encountered, Downloaded by [University of Florida] at 18:09 05 November 2015 THE REPRESENTATION OF SUBLEXICAL CORRESPONDENCES 933 it is parsed and mapped onto every level of representation for which there is a corresponding unit. For ML , whole word and syllabic representations appear to be relatively intact, hence his relatively good reading of words and relatively good reading of nonwords that correspond to syllables within real words. However, when the nonwords to be read are letter sequences that occur within a real word, but do not form a syllable within a real word, ML performs very poorly because lower level mappings (i.e. grapheme± to± phoneme) are not available. It is interesting to note that more recent implementations of the Seidenberg and McClelland-type model (Plaut et al., 1996), have moved towards incorporating more localist representations. T he Plaut et al. model incorporates onset, vowel, and coda (word-® nal consonant cluster) representations instead of representing phonology through the use of Wickelfeatures (context-sensitive triplets) as in the Seidenberg and McClelland (1989) model. However, the work of converting orthography to phonology is done by the hidden units. T he hidden units are necessary in order to decrease similarity across input patterns. However, an alternative strategy suggested by Plaut et al. is for the modeller to specify different levels of input unit (as does Norris, 1994). T he current results suggest that this may be the better approach and that the syllable should be represented among those units. As argued earlier, the Norris and Plaut et al. approaches differ primarily in terms of the representation of the units involved in orthographic± phonological conversion. In the approach taken by Norris (see also S hallice et al., 1983), the different levels (i.e. phoneme, onset, body, syllable, word) are represented in such a way that they could be independently affected by brain damage. In the Seidenberg and McClelland (1989) and Plaut et al. approach, the relation between orthography and phonology is represented in a much more distributed fashionÐ there is no clear separation between different levels of representation. T hus, models incorporating localist representations provide a more ready account of how it is that certain levels of representation can be disrupted while others remain relatively intact (see Patterson & Hodges, 1992). A separate issue is whether either class of model can account for syllable effects. To date, both types of model have been restricted to the processing of monosyllabic words. T hus, whether or not either class of model can accommodate our results must remain an open question until models are constructed to deal with multisyllabic words and until the appropriate simulations are run. REFERENCES American Heritage Electronic Dictionary. (1992). Houghton Mif¯ in Co. Andrews, S. (1982). Phonological recoding: Is the regularity effect consistent? Memory & Cognition, 10, 565± 575. Andrews, S. (1989). Frequency and neighborhood effects on lexical access: Activation or search. Journal of Experimental Psychology: Learning, Memory, & Cognition, 15, 802± 814. Backman, J., Bruck, M., Hebert, M., & S eidenberg, M. (1984). Acquisition and use of spelling-soun d correspondences in reading. Journal of Experimental Child Psychology, 38, 114± 133. Balota, D.A., & Ferraro, F.R. (1993). A dissociation of frequency and regularity effects in pronunciation performance across young adults, older adults, and individual s with Senile Dementia of the Alzheimer Type. Journal of Memory and Language, 32, 573± 592. Downloaded by [University of Florida] at 18:09 05 November 2015 934 LESCH AND MARTIN Baron, J., & Strawson, C. (1976). Use of orthographic and word-speci® c knowledge in reading words aloud. Journal of Experimental Psychology: Human Perception & Performance, 2, 386± 393. Barron, R.W., & Pittinger, J.B. (1974). T he effect of orthographic structure and lexical meaning on same± different judgments. Quarterly Journal of Experimental Psychology, 26, 566± 581. Bartha, M.C., & Martin, R.C. (In preparation). A third route for reading: Implications from a case of phonological dyslexia. Beauvois, M.F., & Derouesne, J. (1979). Phonological alexia: T hree dissociations. Journal of Neurology, Neurosurgery, and Psychiatry, 42, 1115± 1125. Berndt, R., Haendiges, A., Mitchum , C., & Wayland, S. (1996). An investigation of nonlexical reading impairments. Cognitiv e Neuropsychology, 13, 763± 801. Besner, D. (in press). Basic processes in reading: Multiple routines in localist and connectionis t models. To appear in P.A. McMulluen and R.M. Klein (Eds.), Converging methods for understanding reading and dyslexia. Cambridge, MA: MIT Press. Bowey, J.A. (1990). Orthographic onsets and rimes are functional units in reading. Memory & Cognition, 18, 419± 427. Brown, G.D.A. (1987). Resolving inconsistency: A computational model of word naming. Journal of Memory & Language, 26, 1± 23. Bub, D., Black, S., Howell, J., & Kertesz, A. (1987). Speech output processes and reading. In M. Coltheart, G. S artori, and R. Job (Eds.), The cognitiv e neuropsychology of language. Hove, UK: Lawrence Erlbaum Associates Ltd. Butler, B., & Hains, S. (1979). Individual differences in word recognition latency. Memory & Cognition, 7, 68± 76. Caramazza, A., Miceli, G., S ilveri, M.C., & Laudanna, A. (1985). Reading mechanisms and the organisation of the lexicon: Evidence from acquired dyslexia. Cognitiv e N europsychology, 2, 81± 114. Carreiras, M., Alvarez, C., & de- Vega, M. (1993). S yllable frequency and visual word recognition in Spanish. Journal of Memory and Language, 32, 766± 780. Cochran, W.G. (1950). T he comparison of percentages in matched samples. Biometrika, 37, 256± 266. Coltheart, M. (1978). Lexical access in simple reading tasks. In G. Underwood (Ed.), Strategies of information processing (pp. 151± 216). San Diego, CA: Academic Press. Coltheart, M., Curtis, B., Atkins, P., & Haller, M. (1993). Models of reading aloud: Dual-route and parallel-distributed - processing approaches. Psychological Review, 100, 589± 608. Coltheart, M., Masterson, J., Byng, S., Prior, M., & Riddoch, J. (1983). Surface dyslexia. Quarterly Journal of Experimental Psychology, 35A, 469± 495. Derouesne, J., & Beauvois, M.F. (1985). T he ``phonemic’ ’ stage in the nonlexical reading process: Evidence from a case of phonological alexia. In K. Patterson, J.C. Marshall, & M. Coltheart (Eds.), Surface dyslexia (pp. 399± 457). Hove, UK: Lawrence Erlbaum Associates Ltd. Farah, M.J., S towe, R.M., & Levinson, K.L. (1996). Phonological dyslexia: Loss of a reading-speci® c component of the cognitive architecture? Cognitive N europsychology, 13, 849± 868. Francis, W.N., & KucÏ era, H. (1982). Frequency analysis of English useage: Lexicon and grammar. Boston: Houghton Mif¯ in. Frederiksen, J.R., & Kroll, J.F. (1976). Spelling and sound: Approache s to the internal lexicon. Journal of Experimental Psychology: Human Perception and Performance, 2, 361± 379. Friedman, R.B. (1995). Two types of phonological alexia. Cortex, 31, 397± 403. Friedman, R.B. (1996). Phonological text alexia: Poor pseudoword reading plus dif® culty reading functors and af® xes in text. Cognitive N europsychology, 13, 869± 885. Funnell, E. (1983). Phonological processes in reading: New evidence from acquired dyslexia. British Journal of Psychology, 35A, 111± 138. Glass, G.V., Peckham, P.D., S anders, J.R. (1972). Consequences of failure to meet assumptions underlying the ® xed effects analysis of variance and covariance. Review of Educational Research, 42, 237± 288. Glushko, R. (1979). T he organization and activation of orthographic knowledge in reading aloud. Journal of Experimental Psychology: Human Perception & Performance, 5, 674± 691. Goodall, W.C., & Phillips, W.A. (1995). T hree routes from print to sound: Evidence from a case of acquired dyslexia. Cognitive Neuropsychology, 12, 113± 147. Downloaded by [University of Florida] at 18:09 05 November 2015 THE REPRESENTATION OF SUBLEXICAL CORRESPONDENCES 935 Gough, P.B., & Cosky, M.J. (1977). One second of reading again. In N.J. Castellan, D.B. Pisoni, & G.R. Potts (Eds.), Cognitive theory 2. Hillsdale, NJ: Lawrence Erlbaum Associates Ltd. Graham, K., Hodges, J., & Patterson, K. (1994). T he relationship between comprehension and oral reading in progressive ¯ uent aphasia. Neuropsychologia, 32, 299± 316. Hillis, A.E., & Caramazza, A. (1991). Mechanisms for accessing lexical representations for output: Evidence from a category- speci® c semantic de® cit. Brain and Language, 40, 106± 144. Jared, D., McRae, K., & S eidenberg, M.S. (1990). T he basis of consistency effects in word naming. Journal of Memory and Language, 29, 687± 715. Kay, J., & Bishop, D. (1987). Anatomical differences between nose, palm, and foot, or, the body in question: Further dissection of the processes of sub- lexical spelling - sound translation. In M. Coltheart (Ed.), Attention and performance X II: The psychology of reading. Hove, UK: Lawrence Erlbaum Associates Ltd. Kay, J., Lesser, R., & Coltheart, M. (1992). Psycholinguistic Assessments of Language Processing in Aphasia: Reading and Spelling. Hove: Lawrence Erlbaum Associates Ltd. Klapp, S.T. (1971). Implicit speech inferred from response latencies in same± different decision. Journal of Experimental Psychology, 91, 262± 267. Klapp, S.T., Anderson, W.G., & Berrian, R.W. (1973). Implicit speech in reading reconsidered. Journal of Experimental Psychology, 100, 368± 374. Lunney, G.H. (1970). Using analysis of variance with a dichotomou s dependent variable: An empirical study. Journal of Educational Measurement, 7, 263± 269. Marcel, A. (1980). S urface dyslexia and beginning reading: A revised hypothesis of pronunciation of print and its impairments. In M. Coltheart, K. Patterson, and J.C. Marshall (Eds.), Deep Dyslexia. London: Routledge & Kegan Paul. Marshall, M.C., & Newcombe, F. (1973). Patterns of paralexia: A psycholinguisti c approach. Journal of Psycholinguistic Research, 2, 175± 199. Martin, N., & S affran, E. (1992a). A computational account of deep dysphasia: Evidence from a single case study. Brain and Language, 43, 240± 274. Martin, N., & S affran, E. (1992b, October). Temporal effects on repetition: Ev idence for common processes mediating lexical retrieval and verbal STM. Paper presented at the Academy of Aphasia, Toronto. Martin, R.C., & Lesch, M.F. (1996). Associations and dissociations between language impairment and list recall: Implications for models of ST M. In S. Gathercole (Ed.), Models of short-term memory. Hove: Lawrence Erlbaum Associates Ltd. Mayzner, M.S., & Tresselt, M.E. (1965). Tables of single letter and digram frequency counts for various word-length and letter position combinations. Psychonomics Monograph Supplements, 1, 13± 32. Morais, J., Bertelson, P., Cary, L., & Alegria, J. (1986). Literacy training and speech segmentation. Cognition, 24, 45± 64. Morais, J., Cary, L., Alegria, J., & Bertelson, P. (1979). Does awareness of speech as a sequence of phones arise simultaneously? Cognition, 7, 323± 331. Morton, J., & Patterson, K.E. (1980). A new attempt at an interpretation , or, an attempt at a new interpretation . In M. Coltheart, K.E. Patterson, & J.C. Marshall (Eds.), Deep dyslexia. London: Routledge & Kegan Paul. Newcombe, F., & Marshall, J.C. (1985). Reading and writing by letter sounds. In K.E. Patterson, J.C. Marshall, & M. Coltheart (Eds.), Surface dyslexia: N europsychological and cognitive studies of phonological reading. Hove, UK: Lawrence Erlbaum Associates Ltd. Norris, D. (1994). A quantitativ e multiple-levels model of reading aloud. Journal of Experimental Psychology: Human Perception and Performance, 20, 1212± 1232. Patterson, K., Graham, N., & Hodges, J. (1994). T he impact of semantic memory loss on phonological representations. Journal of Cognitive N euroscience, 6. 57± 69. Patterson, K., & Hodges, J. (1992). Deterioration of word meaning: Implications for reading. N europsychologia, 12, 1025± 1040. Patterson, K., S uzuki, T., & Wydell, T. (1996). Interpreting a case of Japanese phonological alexia: T he key is in phonolog y. Cognitive Neuropsychology, 13, 803± 822. Downloaded by [University of Florida] at 18:09 05 November 2015 936 LESCH AND MARTIN Patterson, K.E. (1982). T he relation between reading and phonological coding: Further neuropsychological observations. In A.W. Ellis (Ed.), Normality and pathology in cognitive functions. London: Academic Press. Patterson, K.E., & Marcel, A. (1992). Phonological ALEXIA or PHONOLOGICAL alexia? In J. Alegria, D. Holender, J. Junca de Morais, & M. Radeau (Eds.), Analytic approaches to human cognition. Amsterdam, T he Netherlands: Elsevier. Patterson, K.E., & Morton, J. (1985). From orthography to phonology: An attempt at an old interpretation. In K.E. Patterson, J.C. Marshall, & M. Coltheart (Eds.), Surface dyslexia: Neuropyschological and cognitive studies of phonological reading. Hove, UK: Lawrence Erlbaum Associates Ltd. Patterson, K.E., & S hewell, C. (1987). S peak and spell: Dissociations and word-class effects. In M. Coltheart, G. S artori, & R. Job (Eds.), The cognitive neuropsychology of language. Hove, UK: Lawrence Erlbaum Associates Ltd. Plaut, D.C., McClelland, J.L., S eidenberg, M.S., & Patterson, K.P. (1996). Understanding normal and impaired word reading: Computational principles in quasi-regular domains. Psychological Review, 103, 56± 115. Prinzmetal, W., Treiman, R., & Rho, S.H. (1986). How to see a reading unit. Journal of Memory and Language, 25, 461± 475. Rapp, B.C. (1992). T he nature of sublexical orthographic organization: T he bigram trough hypothesis examined. Journal of Memory and Language, 31, 33± 53. Richardson, J.T.E. (1976). T he effects of stimulus attributes upon latency of word recognition. British Journal of Psychology, 67, 315± 325. Roeltgen, D.P., Cordell, C., & Sevush, S. (1983, Octobe r 31). A battery of linguistic analysis for writing and reading. International Neuropsychological Society Bulletin. Saffran, E., & Marin, O.S.M . (1977). Reading without phonology : Evidence from aphasia. Quarterly Journal of Experimental Psychology, 29, 515± 525. Saffran, E.M., S chwartz, M.F., Linebarger, M., Martin, N., & Bochetto, P. (1989). The Philadelphia Comprehension Battery. Unpublished. Sasanuma, S., Ito, H., Patterson, K., & Ito, T. (1996). Phonological alexia in Japanese: A case study. Cognitive Neuropsychology, 13, 823± 848. Schwartz, M.F., Saffran, E.M., & Marin, O.S.M . (1980). Fractionating the reading process in dementia. In M. Coltheart, K.E. Patterson, & J.C. Marshall (Eds.), Deep dylsexia. London: Routledge & Kegan Paul. Seidenberg, M., & McClelland, J.L. (1989). A distributed, developmental model of visual word recognition and naming. Psychological Review, 96, 523± 568. Seidenberg, M.S. (1985). T he time-course of information activation utilization in visual word recognition. In D. Besner, T.G. Waller, & G.E. MacKinnon (Eds.), Reading research: Advances in theory and practice, Vol. 5. New York: Academic Press. Seidenberg, M.S. (1987). Sublexical structures in visual word recognition: Access units or orthographic redundancy? In M. Coltheart (Ed.), Attention and performance X II (pp. 281± 298). Hove, UK: Lawrence Erlbaum Associates Ltd. Seidenberg, M.S. (1989). Reading complex words. In G. Carlson & M. Tanenhau s (Eds.), Linguistic structure in language processing (pp. 53± 105). Dordrecht: Kluwer. S eidenberg, M.S., Waters, G.S., Barnes, M.A., and Tanenhaus, M.K. (1984). When does irregular spelling or pronunciation in¯ uence word recognition? Journal of Verbal Learning and Verbal Behavior, 23, 383± 404. Shallice, T. & McCarthy, R. (1985). Phonological reading: From patterns of impairment to possible procedures. In K.E. Patterson, J.C. Marshall, and M. Coltheart (Eds.), Surface dyslexia: N europsychological and cognitive studies of phonological reading (pp. 361± 397). Hove, UK: Lawrence Erlbaum Associates Ltd. Shallice, T., & Warrington, E.K. (1980). S ingle and multiple component central dyslexic syndromes. In M. Coltheart, K.E. Patterson, & J.C. Marshall (Eds.), Deep dyslexia. London: Routledge & Kegan Paul. Shallice, T., Warrington, E.K., & McCarthy, R. (1983). Reading withou t semantics. Quarterly Journal of Experimental Psychology, 35A, 111± 138. Downloaded by [University of Florida] at 18:09 05 November 2015 THE REPRESENTATION OF SUBLEXICAL CORRESPONDENCES 937 Stanovich, K.E., & Bauer, D.W. (1978). Experiments on the spelling± to± sound regularity effect in word recognition. Memory & Cognition, 6, 410± 415. Taft, M. (1991). Reading and the mental lexicon (pp. 59± 92). Hove: Lawrence Erlbaum Associates Ltd. Taraban, R., & McClelland, J.L. (1987). Conspiracy effects in word pronunciation. Journal of Memory and Language, 26, 608± 631. Taylor, G.A., Miller, T.J., & Juola, J.F. (1977). Isolating visual units in the perception of words and nonwords. Perception and Psychophysics, 21, 377± 386. Treiman, R., & Chafetz, J. (1987). Are there onset- and rime-like units in written words? In M. Coltheart (Ed.), Attention and performance X II (pp. 281± 298). Hove, UK: Lawrence Erlbaum Associates Ltd. Treiman, R., & Zukowski, A. (1988). Units in reading and spelling. Journal of Memory and Language, 27, 466± 477. Van Orden, G.C., S tone, G.O., Garlington, K.L., Markson, L .R., Pinnt, G.S., S imonfy, C.M., & Brichetto, T. (1992). ’ ’ Assembled’ ’ phonolog y and reading: A case study in how theoretical perspective shapes empirical investigation . In R. Frost & L. Katz (Eds.), Orthography, phonology, morphology, and meaning. New York: North Holland. Warrington, E.K., & S hallice, T. (1980). Word form dyslexia. Brain, 103, 99± 112. Waters, G.S., & Seidenberg, M.S. (1985). S pelling± sound effects in reading: T ime course and decision criteria. Memory & Cognition, 13, 557± 572. Waters, G.S., S eidenberg, M.S., & Bruck, M. (1984). Children’ s and adults’ use of spelling± sound information in three reading tasks. Memory & Cognition, 12, 292± 305. Winer, B.J., Brown, D., & Michels, K.M. (1991). Statistical principles in experimental design. New York: McGraw-Hill. Original manuscript receiv ed 30 June 1997 Accepted revision receiv ed 26 N ovember 1997 938 LESCH AND MARTIN APPENDIX Downloaded by [University of Florida] at 18:09 05 November 2015 Syllable and Control Nonwords Employed in the Reading Tasks Word-initial N onwords Non- initial N onwords Syllable Control Syllable Control clois cres chol chis giz boun pud brin griz frus meth cem bril clem frac chal skel treb cush chig grov chev splut slov cler crev grum clav slen ¯ ut breth vod bris shim brist ® ct spee fost lup mout mep nort nymp misp dist fav ¯ ig dimp ® lt chen amph dext balc banj barm belf kerch pral calc drud bulg rept anth furb freeb soj canv hois sorb bort mous vert brate serk ceps muth vorce voke crete fect rette plode dren tigue brupt boose phoid droit plex ¯ ect blem dral vage nence pult oist hage rame ront stent tomp trab krat ¯ ess terg droom blic welve spass spen troke kless twise troom scous trut scove roup dway scue slead slin