COGNITIVE NEUROPSY CHOLOGY, 2000, 17 (6), 517–546 THE SELECTIVE IMPAIRMENT OF THE PHONOLOGICAL OUTPUT BUFFER Tim Shallice University College London, UK and SISSA-ISAS, Trieste, Italy Raffaella I. Rumiati SISSA -ISAS, Trieste, Italy Antonella Zadini Ospedale Maggiore, Trieste, Italy A single case study is presented of a patient, LT, with a reproduction conduction aphasic pattern of performance on word reproduction tasks; thus he made substitutions, insertions, deletions, and transpositions in reading, writing, and repetition of words and nonwords, as well as in sentence production, and in spoken and written picture and action naming. Further analyses revealed that there was no effect of semantic or syntactic structure on performance, and that reading was slightly better than repetition and writing. Finally, the observed effects of lexicality, length, and word frequency were similar to those found in other phonological output buffer patients. Overall, the pattern observed fits the characteristics typical of phonological output buffer patients, as characterised by Caramazza, Miceli, and Villa (1986). We discuss the implications of these results for understanding the role of the output phonological buffer in neuropsychological and computational models of writing, reading, and repetition. From the perspective of LT’s performance, the hypothesis suggested by Caramazza et al. (1986), and that of Hartley and Houghton (1996), that word production in reading and repetition uses an additional route to access articulatory or phoneme-level representations from the phonological output lexicon, is unnecessary; instead, word-nonword differences in other patients can be attributed to resource demand differences between the two types of stimuli. LT’s preserved span fit with the assumption that two phonological buffers exist, one for input and the other for output. Results from a word repetition experiment, in which word syllable structure was manipulated, are in conflict with one further noncentral aspect of the Hartley and Houghton’s model, which otherwise fits the results well. INTRODUCTION The existence of a phonological output buffer holding the extended phonological representation of an intended utterance has long been a well-established concept in studies of the short-term memory performance of normal subjects (e.g., Ellis, 1979; Monsell, 1984; Morton, 1969; Sperling, 1967). Indeed, early versions of the working memory model had presupposed that such a speech output Requests for reprints should be addressed to Tim Shallice, Institute of Cognitive Neuroscience, University College London, Gower Street, London WCIE 6BT, UK, (Email: [email protected]) or Raffaella Rumiati, Cognitive Neuroscience Sector, Scuola Internazionale Superiore di Studi Avanzati (SISSA), Via Beirut 2-4, Trieste, Italy (Email: [email protected]). We would like to thank Randi Martin and two referees for their very helpful comments on an earlier version of the paper and LT for his kind collaboration. We are also grateful to Claudio Luzzatti for his comments on the patient’s CT scan and to Emanuela Bricolo for computer assistance. The research was assisted by support from grants EC PSS* 0859 and MURST. Ó 2000 Psychology Press Ltd http://www.tandf.co.uk/journals/pp/02643294.html 517 SHALLICE, RUMIATI, ZADINI buffer was the auditory-verbal short-term store linked to the articulatory loop slave system (Baddeley & Hitch, 1974). Mechanisms that hold phonological information for the order of a clause have also been postulated in research with a more linguistic orientation focusing on speech errors (e.g., Fromkin, 1973). These have included a buffer and related hypothetical mechanisms (scan copier, checkoff monitor, etc.) in Shattuck-Hufnagel’s (1979) account or the phonological assembly system in Butterworth’s theorising (1980). The representations being retained are generally assumed to be at the positional level on Garrett’s (1980) model (Buckingham, 1992; but see Pate, Saffran, & Martin, 1987). From a neuropsychological perspective, a disorder of a phonological output buffer has been held responsible for conduction aphasia of the “reproduction” variety and contrasted with conduction aphasia of the repetition variety, which was linked to damage to an input auditory-verbal short-term store (Shallice & Warrington, 1977). On this approach the reproduction and repetition varieties of conduction aphasia are empirically distinguishable according to whether the patient’s difficulty in repeating an utterance occurs principally for individual longer, less familiar words—the “reproduction” variety—or for strings of short familiar words (span)—the “repetition” variety. The “reproduction” variety then corresponds to the classical form of the conduction aphasia syndrome, in which patients are fluent but make literal paraphasic errors both in spontaneous speech and in the reproduction of single words (e.g., Dubois et al., 1964). The errors of this form are characterised in terms of substitutions, insertions, deletions, and moves of phonemes, often resulting in a “doublet” (Buckingham, 1992; Lecours & Lhermitte, 1969), and occasionally a full exchange. The first detailed analysis of the impairment of a patient who was characterised as having an output phonological buffer disorder was that of patient IGR by Caramazza, Miceli, and Villa (1986). Their paper was of considerable methodological as well as empirical and theoretical interest. In general conduction aphasics perform more poorly in nonword than word production but the performance is 518 COGNITIVE NEUROPSY CHOLOGY , 2000, 17 (6) qualitatively equivalent (see Caplan & Waters, 1990). IGR, however, showed the conduction aphasia pattern, but only for nonwords. He did not produce paraphasias in spontaneous speech or in the reproduction of single long words. Instead a difficulty was apparent in the reproduction of pronounceable nonwords. Thus 61% of IGR’s errors could be explained in terms of a single substitution, insertion, deletion of a phoneme, or a transposition of two phonemes. The others were either combinations of the more basic error types or “fragments.” Moreover, 82% of the errors fell in the same phonological category, defined by manner of articulation. The methodological and theoretical importance of the paper was, however, derived from the relation between these findings and those in two other tasks—the reading aloud and the writing of pronounceable nonwords. The characteristics described earlier for the reproduction of nonwords were also basically found in these other two tasks: The single letter/phoneme error type comprised 83% of the reading errors and 75% of the writing errors. The similarity in the pattern of errors across tasks also extended to three other characteristics. Thus the relative rates of single substitution (S), single insertion (I), and single deletion (D) and transposition (T) errors are remarkably similar: repetition 81% (S), 11% (I), 3% (D), 5%(T); reading 81% (S), 13% (I), 6% (D) 0% (T); writing: 68% (S), 13%(I), 17% (D), 2% (T).) Moreover, the errors occur in a roughly comparable fashion across different letter positions within the nonwords and the errors were nearly all on consonants for all three tasks. The key methodological innovation of the paper was that this common pattern of characteristics in the errors produced by a patient was used to support the assumption of the existence of a common system or common set of subsystems utilised in the three tasks. Although inferences from a common quantitative pattern of impairment across different basic tasks was considered in principle appropriate evidence for supporting the assumption that the same system or set of systems was used in performance of the tasks, unlike mere association of deficits (see Shallice, 1979a), no systematic and detailed application of the principle were available THE PHONOLOGICAL OUTPUT BUFFER until this and a related group of papers were produced in the mid-1980s by Caramazza, Miceli, and their colleagues (see also Caramazza & McCloskey, 1985; Caramazza, Miceli, Silveri, & Laudanna, 1985; Caramazza, Miceli, Villa, & Romani, 1987; see also for metatheoretical polemic McCloskey & Caramazza, 1991) and by Caplan, Vanier, and Baker (1986) on reproduction conduction aphasia. Moreover, the procedure clearly demonstrated the power of error analysis, which had previously been held to be unreliable for discovering theories about the functional organisation of the normal system when contrasted with dissociations (see, e.g., Shallice, 1979a). Theoretically, the findings of Caramazza et al. (1986) were basically explained by assuming that the phonological output buffer was utilised for a variety of purposes in speech production, repetition, reading aloud, and writing. In these respects this neuropsychological evidence was fully consistent with theoretical claims about the operation of a phonological buffer drawn from other sources of evidence (e.g., Ellis, 1980; Morton, 1969). Minor differences between the three basic tasks could be easily explained by particular characteristics of one of them. Thus writing and repetition show major effects of the length of the nonwords but reading does not; this is presumably because reading does not require all of a nonword to be maintained in the buffer before production begins (see Coltheart, Curtis, Atkins, & Haller, 1993; Làdavas, Shallice, & Zanella, 1997). However, Caramazza et al. (1986) made an additional novel theoretical claim, which puts their position in conflict with accounts of “reproduction” conduction aphasia where that is seen as a disorder of the phonological buffer. They argued that for all operations that required spoken output, the use of the phonological buffer was critical in only one respect for the ecologically most important output: spoken words. The critical aspect was word order, it being held that the processing of single words and even words in short sentences could be achieved by the use of a route from the phonological output lexicon to the “lexicon-articulatory system,” which bypassed the phonological output buffer. They made this assumption to explain the fact that IGR was 100% correct at repeating words, 98% correct at reading them, and 97% correct at writing them. The syndrome of phonological agraphia (Shallice, 1981) indicates that word writing can be carried out independently of sound-to-spelling translation, which would require the use of the representation stored in the phonological output buffer. However, Caramazza et al. argue that a second route which did not utilise the phonological output buffer was necessary to explain the intact repeating and reading aloud of words by IGR. Detailed investigations on two similar patients from a related perspective are known to us (a third patient, that of Romani, 1992, also appears to be similar). Bub, Black, Howell, and Kertesz (1987) carried out a detailed analysis of the reading and repetition impairments of a phonological alexic patient, MV. The patient differed from IGR in being severely agraphic, so no analogous study of writing could be carried out. Again, errors in reading or repeating a nonword were very similar in nature. They tended to differ from the target by a single distinctive feature with substitutions being much more frequent than omissions or additions of a phoneme. Words were virtually perfectly read and repeated except for the repetition of three syllable words. Bub et al. (1987) came to a different theoretical conclusion to Caramazza et al. (1986), arguing that the contents of the response buffer of MV were limited in capacity and subject to abnormally rapid decay. They argue: “… incompletely specified traces … trigger a higher-level interpretative mechanism which evaluates the response options. Since the potential candidates at this level of representation are morphemes that provide the best possible fit to the description in the response buffer, the output selected will be a word that is a close phonological approximation to the target” (p. 95). Thus, they argue, word performance will be better than performance with nonwords. Bisiacchi, Cipolotti, and Denes (1989) reported a patient, RR, who had a very similar pattern of errors in reading, writing, and repeating nonwords to IGR, even down to the relative ratios of transpositions, insertions, deletions, and transpositions. RR, like IGR, had a greatly reduced auditoryverbal span and Bisiacchi et al. raise the question of COGNITIVE NEUROPSYCHOLOGY , 2000, 17 (6) 519 SHALLICE, RUMIATI, ZADINI whether the deficits in phonological nonword processing and span arise due to an impairment of different functional systems, a common system, or partially overlapping systems. They do, however, point out the existence of short-term memory patients, namely patients with a much reduced span without problems in word comprehension or production, who can read nonwords, e.g., JB (Shallice & Butterworth, 1977), and PV (Vallar & Baddeley, 1984) (Vallar, personal communication). However, reading is the least demanding of the three basic tasks as far as the use of the phonological output buffer is concerned, as PV might have read syllableby-syllable, using the phonological reading route. Both Bub et al. (1987) and Bisiacchi et al. (1989) essentially adopt the same overall interpretation of the syndrome as Caramazza et al. (1986), namely an impairment to an (output) phonological buffer. However, neither Caramazza et al. nor Bisiacchi et al. label their patients as conduction aphasic and indeed their patients do not show phonemic paraphasias in spontaneous speech. However, many characteristics of the reproduction of nonwords in patients IGR and RR had previously been described in word reproduction in conduction aphasic patients (e.g., Blumstein, 1973; Lecours & Lhermitte, 1969; Pate et al., 1987). One possibility, to be considered further in the Discussion, is therefore that the impairment of IGR, MV, and RR was merely a mild one and if it had been quantitatively more severe but qualitatively identical, then it would have also affected word reproduction and spontaneous speech; their patients would have presented as typical reproduction conduction aphasics. The argument of Caramazza et al. that word repetition does not involve the phonological output buffer would be contaminated by what has been called a “resource artefact” (Shallice, 1979a). A second alternative possibility that none of the studies considered is provided by the hypothesis of McCarthy and Warrington (1984) to explain the impaired repetition performance of two conduction aphasic patients ORF and RAN. The hypothesis was that in their patients there was a disconnection of input orthographic and phonological processors, including the auditory-verbal short-term store, the (input) phonological buffer in the terminology of 520 COGNITIVE NEUROPSY CHOLOGY , 2000, 17 (6) Baddeley (1986), from output phonological processes including the phonological output buffer. On this view, the phonological output buffer would receive inadequate input from the phonological input system unless the information were routed via the semantic system; indeed, in ORF and RAN forcing the patient to use a semantic processing strategy improved word reproduction. Such a hypothesis would account for the findings obtained on IGR, MV, and RR as information about visually or auditory presented words could be routed through the semantic system to the output system, but this would not be possible for nonwords, which lack a semantic representation. The disconnection hypothesis did not raise any novel possibilities about the organisation of the normal language system. The impairments of IGR and RR have recently become of key theoretical interest because they have provided major neuropsychological support for a model of phoneme selection and retention in speech developed by Hartley and Houghton (1996) in the context of a novel theory of the serial ordering of speech (see Figure 1). The starting point of their theory is a class of neural network models of serial order known as “competitive queuing” (CQ) models (Houghton, 1990). CQ models are activationbased models in which order is represented by associations between localist nodes representing item (in this case phoneme) identity and other nodes that contain some abstract representation of an item’s positions, called context nodes. Item selection for each position is effected through lateral inhibitory competition between activation within item nodes. The simplest form of context nodes is a pair, one representing distance from the beginning of the sequence and the other distance from the end. The activation of each changes over time in a complementary fashion in an initial (Hebbian) learning phase and in the same way when the sequence is retrieved. Models of this type have been applied to auditory-verbal short-term memory phenomena (e.g., Burgess, 1995; Burgess & Hitch, 1992) and to the sequencing of letters in spelling (e.g., Houghton, Glasspool, & Shallice, 1994; Shallice, Glasspool, & Houghton, 1996). In the Hartley and Houghton THE PHONOLOGICAL OUTPUT BUFFER Temporally Correlated Internal Context ll Novel Phonological Content Pathway l l l l l l l l l l Forms l ll l l l l l l l l l l l Structural Pathway ll ll ll ll ll Familiar Phonological l l l l l ll l ll Forms Template lll lll ll l Phonemes Figure 1. The basic components of the Hartley-Houghton (1996) model of speech production and short-term memory. The phonological form representations contain two nodes, one corresponding to the onset and the other to the rhyme. The template representation has five nodes, each corresponding to the position of a phoneme in the syllable. The phoneme representation has a node for each phoneme in the language. The shaded area represents structures not included in the implementation. The dashed lines represent pathways made up of temporary weights, the solid lines represent connections which do not decay over time (from Hartley & Houghton, 1996, p.22). quencing of phonemes in word production the context nodes are more complex. Indeed, they have two levels. A syllable template is created that is intended to approximate to the structure of a putatively universal generalised syllable and is based around the concept of sonority. The syllable template consist of five nodes, two for the initial consonant cluster, the third for the vowel, and two for the coda, only one of which is activated for each input phoneme. Above this set of phoneme nodes are a string of four syllable position nodes, with again only one being activated at a time. Again, the same sequence of syl- lable and phoneme position nodes occurs in the learning phase and when the information is retrieved. A further complexity of the Hartley and Houghton model is that there are held to be two sets of connections between the syllable representations and the phoneme identity and context representations. They explain the performance of aphasic patients by increasing the weight-decay parameter within the model.1 Hartley and Houghton situate this specific model in a broader model, which makes two further commitments, neither in our view derived from the 1 An unfortunate aspect of the Hartley and Houghton simulation is that the impairment is simulated by an increase in the general level of noise throughout the network. In our view it is preferable to increase the noise in a particular layer, as done in the Houghton, Glasspool, and Shallice simulation. COGNITIVE NEUROPSYCHOLOGY , 2000, 17 (6) 521 SHALLICE, RUMIATI, ZADINI basic logic of the approach. The first is that no distinction at the level of short-term retention of phonological information is made between the input and the output phonological systems; their model in this respect is in the same spirit as another connectionist model of repetition, that of Martin and Saffran (1992). The second, which has some relation to the position of Caramazza et al. (1986), is that a partially separate pathway where the connections are permanent can carry out single word production. These permanent connections can be used in isolation. Hartley and Houghton do not provide an independent justification for this last assumption and ironically a contrasting assumption was made in a related model of spelling (Houghton et al., 1994). Given this second assumption, they naturally based their model neuropsychologically not on reproduction conduction aphasia in general but on IGR and RR. They use parameters previously employed in simulations of findings on normal subjects and with only one arbitrary parameter change—of the half-life of weight-decay, which is reduced from 5.0 s to 3.5 s—a good quantitative fit is obtained to the performance of IGR and RR. The incidence rates of certain error types, which will be discussed in more detail later (single deletions, single insertions, single substitutions, multiple substitutions, substitution and insertions, substitutions and deletions and transpositions), are reproduced accurately. The largest discrepancy is over multiple substitution errors, which the simulation predicts to occur for 0.67% of responses, but IGR produces them on 3.52% of trials. However, this is clearly acceptable within-syndrome variability as RR produces 0/130 such responses in repetition but makes 2/130 in reading and 6/130 in writing. To our knowledge no quantitative predictions of comparable accuracy exist in neuropsychology. We present in this paper a fourth patient, LT, who would be classified as a reproduction conduction aphasia as his disorder involves both word and nonword processing. Our principal aim is to show that there are no good grounds for assuming that he has a different functional syndrome (in the sense of Plaut & Shallice, 1993) from IGR, MV, and RR. 522 COGNITIVE NEUROPSY CHOLOGY , 2000, 17 (6) We then consider the performance of LT in the context of those of IGR, MV, and RR with respect to four theoretical issues. For the first three issues we use the concept of the “phonological output buffer.” By this we refer to the mechanisms by which information is retained in any form of sublexical phonological code, within the speech production system over short intervals of time (i.e., less than 10 seconds, say). We do not intend the concept to refer to any specific architecture such as a store with a fixed number of slots. In the fourth issue we are more specific about how the system retains such information. 1. How does LT’s performance speak to the hypothesis of Caramazza et al. of a separate output route for word production that bypasses the phonological output buffer? 2. Is the disorder, at least in LT, best captured by an impairment of the phonological output buffer or by disconnection between input and output phonological systems? 3. How does LT’s performance relate to the issue of the separability or identity of the phonological input and output buffers? 4. How does LT’s performance relate to the predictions of the Hartley and Houghton model? CASE HISTORY A right-handed Italian male patient, LT, born in 1932, who had 8 years of schooling was studied. During his working life he had held several jobs such as waiter, barman, and telegraph operator on ships. For a number of years before retiring, he worked as a secretary in a transport company. LT was admitted to the hospital of his hometown on March 1, 1995, with weakness to the right side of the body and global aphasia. A CT scan (23/10/96) revealed two hypodense areas in the left hemisphere: an older superior fronto-parietal area and a more recent perisylvian area (see Figures 2a and 2b). Broca’s area is spared, but Wernicke’s area, the inferior parietal lobule, the lateral area 7, and probably also the insula are affected. After having been discharged from hospital, LT began to attend a THE PHONOLOGICAL OUTPUT BUFFER Figure 2. Two horizontal sections: (a) z = + 4.0; (b) z = + 24.0; showing the lesion that affects the left superior temporal and inferior parietal lobes. speech and physiotherapy programme in the Rehabilitation Unit of the Ospedale Maggiore in Trieste. The present study was carried out between November 1995 and April 1997, during which time LT was always collaborative and aware of his difficulties. During the testing period his general neuropsychological condition remained unchanged. The patient was well oriented in time and space. He scored 23 (50th percentile) on Raven’s Coloured Progressive Matrices. However, his performance on the WAIS was clearly below that level. His verbal IQ was 88, and nonverbal IQ 89, with subtests ranging from above average (Digit Span) to borderline defective (Comprehension, Digit Symbol, and Picture Arrangement). Of importance is that his digit span was 6 forwards and 3 backwards. A general neuropsychological assessment revealed no evident agnosic problems. He was at ceiling on a shortened version of the Efron Test (6/ 6) and the Object Decision task (10/10) and on matching pictures on the basis of their meaning (5/ 5), and identity (5/5). A test of utilisation apraxia on verbal instructions (16/20, 80%) indicated a mild deficit only with the left limb. However, when he was asked to imitate gestures made by the examiner, left limb performance improved (19/20, 95%). LT was not tested with his right limb because he had a motor difficulty. No buccofacial (20, normal mean = 19.8, see Spinnler & Tognoni, 1987) or constructional apraxia (14, normal mean = 12.5, see Spinnler & Tognoni) was present. On the Corsi block tapping test he showed a span of 6 forwards (control mean = 5, see Spinnler & Tognoni), and 5 backwards. COGNITIVE NEUROPSYCHOLOGY , 2000, 17 (6) 523 SHALLICE, RUMIATI, ZADINI LANGUAGE EVALUATION LT’s language abilities were evaluated using BADA (Batteria per l’Analisi dei Deficit Afasici), devised by Miceli, Burani, and Laudanna (1991). The patient’s spontaneous speech displayed a somewhat impoverished use of morphological and syntactic structures and showed the classical conduction aphasic pattern of fluent but paraphasic speech with the individual phonemes normally produced. The simple argument structure subjectverb-object was often present in his utterances and the prosody was appropriate. Comprehension of sentence material was also mildly impaired. In a task in which he was asked to match a sentence said by the examiner to one of two pictures, LT scored 54/60 (90%) and he scored similarly when judging whether a spoken sentence was grammatically correct (41/48, 85%). He also showed a deficit on a written version of the sentence comprehension task (23/30, 77%) and on the written version of the grammatical judgement task (20/ 24, 83%). Finally, on a shortened version of the Token Test (De Renzi & Faglioni, 1978) he obtained 16/36.2 However in tasks in which input processing was required of only single lexical items or smaller units, he always scored at above 90%. Thus LT scored 76/ 80 (95%) in auditory lexical decision (with nonwords being made from words by changing one phoneme) and his comprehension of single words was virtually normal both for auditory and written inputs. When asked to match a spoken word to one of two pictures, when the distractor was phonologically similar to the target (e.g., treccia [plait] and freccia, [arrow]), he made only three errors (37/40, 92.5%, controls matched for age and education scored 38 or 39, mean = 38.6). In an analogous version in which the word to be matched was written, he again made only three errors (37/40, 92.5%). In 2 an additional task, LT was asked to say whether two phonemes spoken by the experimenter were the same or different; he was able to discriminate correctly 56/60 (93.3%) pairs of phonemes differing on a single feature (e.g., ta/da). However, when production was required his performance was much worse. Thus LT’s naming performance was impaired both with line drawings of objects (14/30, 47%) and of actions (4/28, 14%) as stimuli. When asked to give a written response rather than a spoken one, LT scored 16/22 with objects (73%), and 11/22 with actions (50%). Lastly, the patient named correctly 9/16 (56%) definitions of objects described by the examiner. No dysarthria or difficulty in individual phoneme production was noted. However, he showed the reproduction conduction aphasic pattern of substitutions, deletions, insertions, and transpositions of phonemes within words.3 A similar pattern occurred in reading aloud, writing, and repetition. He read aloud promptly all letters of the Italian alphabet presented in random order (21/21, 100%) and performed at a similar level when required to write letters to dictation. Moreover, he read 61/63 (97%) syllables with a CV structures (e.g., ba, ta, etc.) correctly. However, his reading, writing, and repetition of real words were affected. He correctly repeated only 10 words out of 45 (22%), read 56 out of 92 (61%), and wrote 23 out of 46 (50%). The ability to perform a delayed copy of words was almost spared (9/10). Finally, he spelt aloud correctly 16/20 (80%) words (including nouns, verbs, adverbs, and prepositions). LT also had difficulties in repeating, reading, and writing nonwords (19/36, 28%; 24/42, 57%; and 8/25, 32%, respectively). However, the delayed copy of nonwords (5/6, 83%) was again almost spared. The errors of word and nonword production were again of a reproduction conduction aphasic type, as will be discussed later. It is conceivable that an inability to effectively rehearse material resulting from his speech production problem may be contributing to his mild impairment on some comprehension tasks. This possibility was not investigated further. 3 It is conceivable that his output problems may have led to an apparent input problem through his recycling information inappropriately. For simplicity we have ignored this possibility, which if it were true would merely simplify the argument presented in the paper. 524 COGNITIVE NEUROPSY CHOLOGY , 2000, 17 (6) THE PHONOLOGICAL OUTPUT BUFFER Comment LT principally had problems in two domains—syntactic processing and in word production. His problems in word production were, however, of a specific sort. He made substitutions, insertions, deletions, and transpositions in reading, writing, and repetition of words, nonwords, and sentences. These problems were also found in spoken and written picture and action naming, where his difficulties arose from production rather than agnosic problems. In the following section we examine in greater detail the nature of LT’s pattern of errors in writing, reading, and repetition and also in object naming. His syntactic processing problems will be treated as an additional deficit not relevant to the theoretical purpose of the paper. EXPERIMENTAL STUDIES OF WORD REPRODUCTION Study 1: Basic analyses of repetition, reading, and writing LT was asked to repeat, read, and write (using his right hand) words and nonwords. There were two sets of words (set A, N =203, and set B, N=104) and Table 1. Percentages of correct responses made by LT in repetition, reading, and writing words (set A and set B) and nonwords (set C) Repetition Reading Writing Words Set A (N = 203) Test 1 Test 2 Total 65.0% 74.0% 69.5% 77.0% 78.0% 77.5% 48.0% 63.0% 55.5% Set B (N = 104) Test 1 Test 2 Total 50.0% 55.0% 52.5% 80.0% 73.0% 76.5% 60.0% 55.0% 57.5% Nonwords Set C (N = 192) Test 1 Test 2 Total 27.0% 32.0% 29.5% 44.0% 46.0% 45.0% 24.5% 32.0% 28.0% one of nonwords (set C, N=192). Since set A included mostly short words, of which most were nouns, we also administered a set B that contained three- and four-syllable verbs and adjectives. Nonwords were partly taken from Caramazza et al. (1986), and partly made by substituting one letter in a real word. All sets (A, B, C) were presented in blocked fashion in two different sessions to allow a consistency analysis of his responses to be made. Each session required 2 months to be completed and the two sessions were 1 month apart. Tasks (repetition, reading, and writing) were counterbalanced for each set of stimuli. A hierarchical fully saturated loglinear analysis, with tests of partial associations, was applied to the data. The factors were task (repetition, reading, and writing), test (first and second sessions), set (A, B, and C), and correct/error. An appropriate model of the data does not require three-way and higher interactions (c 2 (16) = 23.8, p > 0.05). The model did require significant effects on correct/error of task c 2 (2) = 99.1, p < .001, set, c 2 (2) = 320.7, p < .001, and test, c 2 (1) = 7.3, p < .01, indicating some improvement with practice. Further analyses Words (sets A and B) were categorised on several dimensions: (1) concreteness; (2) part of speech (noun, non-noun, and verb); (3) frequency, using a cut-off of 25 occurrences per million in the CNR dictionary of frequencies (Dizionamo di Frequenze, 1990), and (4) number of syllables (ranging from 1 to 4). These variables are considered in turn. Concreteness LT’s repetition, reading, and writing responses were analysed according to whether the referent of the target word was concrete or abstract. The analysis was carried out on two subsets of pairs of concrete and abstract words drawn from sets A and B respectively. The subsets were made by matching to each member of the smaller sets (either of abstract or of concrete words) one from the larger set of words (either concrete or abstract) according to the following criteria applied in order: COGNITIVE NEUROPSYCHOLOGY , 2000, 17 (6) 525 SHALLICE, RUMIATI, ZADINI 1. if a member of the larger set is chosen then it cannot be chosen again; 2. the pair is the same part of speech (adverbs were not differentiated from verbs); 3. the pair is from the same frequency band, with all non-nouns with frequency values above 534 per million eliminated; 4. the pair has the same length in syllables; 5. when there is more than one candidate match on the first four criteria, then the nearest in the alphabet is chosen. A, we compared nouns with other parts of speech (non-nouns); in the second, selected from set B, verbs were compared with adjectives. An analogous procedure was used to the previous analysis, with the selection from larger sample sizes using the same criteria as before omitting criterion (2). The results are shown in Table 3. There was no significant effect of part of speech for any of the three tasks. Frequency and length LT’s performance with words was analysed according to their written frequency values and length. Words were divided into high frequency and low frequency sets. Word length was analysed in terms of number of syllables (two, three, and four). Tables 4 and 5 show the numbers of words that LT wrote, read, or repeated correctly, differentiated in terms of word frequency and word length. A hierarchical fully saturated loglinear analysis, with tests of par- As shown in Table 2, there is no effect of concreteness on any of the three tasks for either of the word sets. Part of speech In this section we examine whether LT’s difficulties in repetition, reading, and writing are influenced by part of speech. An analysis was carried out on two subsets of words. In the first subset, drawn from set Table 2. LT’s performance on repetition, reading, and writing using matched subsets of stimuli Concrete Abstract Set A ————————————— Repetition Reading Writing Set B ————————————— Repetition Reading Writing 19/36 (53%) 24/36 (67%) 16/36 (44%) 17/36 (47%) 26/36 (72%) 24/36 (67%) 18/36 (50%) 19/36 (53%) 26/36 (72%) 25/36 (69%) 18/36 (50%) 15/36 (42%) Samples from set A include an equal number of two-, three-, and four-syllable concrete and abstract words. Samples from set B contain an equal number of three- and four-syllable concrete and abstract words. In both samples words are matched for part of speech, frequency, and word length. Table 3. LT’s performance on repetition, reading, and writing as a function of grammatical categories using matched subsets of words (set A and set B, test 1 plus test 2) Subset Part of Speech Syllables Repetition Reading Writing Set A Nouns 2, 3, & 4 Non-nouns 2, 3, & 4 33/52 (63%) 28/52 (54%) 38/52 (73%) 35/52 (67%) 30/52 (58%) 24/52 (46%) Verbs 3&4 Adjectives 3&4 24/60 (40%) 31/60 (52%) 43/60 (72%) 47/60 (78%) 36/60 (60%) 32/60 (53%) Set B Five monosyllabic words were not included. 526 COGNITIVE NEUROPSY CHOLOGY , 2000, 17 (6) THE PHONOLOGICAL OUTPUT BUFFER Table 4. LT’s performance on set A and set B, administered twice Words Length Set A Set B Frequency Repetition Reading Writing 2 High 3&4 High 2 Low 3&4 Low 101/124 (81%) 33/46 (72%) 93/134 (69%) 42/90 (47%) 107/124 (86%) 30/46 (65%) 110/134 (82%) 59/90 (66%) 90/124 (73%) 18/46 (39%) 66/134 (49%) 38/90 (42%) 3 High 4 High 3 Low 4 Low 36/52 (63%) 30/54 (56%) 21/46 (46%) 22/56 (39%) 47/52 (90%) 44/54 (81%) 34/46 (74%) 34/56 (61%) 40/52 (77%) 39/54 (72%) 17/46 (37%) 23/56 (41%) Five monosyllabic words and one without a frequency value were not included. Table 5. The number of words from sets A and B and of nonwords (set C) (all administered twice) that LT repeated, read and wrote correctly Length 2 (N = 258) 3 (N = 210) 4 (N = 134) Total (N = 602) Words ————————————– Repetition Reading Writing 194/258 (75%) 124/210 (59%) 60/134 (45%) 378/602 (63%) 217/258 156/258 (84%) (65%) 160/210 105/210 (76%) (50%) 88/134 70/134 (66%) (52%) 465/602 331/602 (77%) (55%) Length 2 (N=136) Nonwords ————————————— Repetition Reading Writing 55/136 (40%) 3 (N=194) 52/194 (27%) 4 (N=52) 4/52 (8%) Total (N=382) 111/382 (29%) 87/136 (64%) 74/194 (38%) 10/52 (19%) 171/382 (45%) 57/136 (42%) 47/194 (24%) 4/52 (8%) 108/382 (28%) Five monosyllabic words and one without frequency value (robot) were not included. Length is measured in terms of syllables. tial associations, was applied to the data from sets A and B. The factors were task (repetition, reading, and writing), frequency (high or low),4 length (two, three, and four syllables), test (first and second sessions), and correct/error. An appropriate model of the data does not require three-way and higher interactions, c 2 (45) = 34.46, p > .05. The model did require significant effects on correct/error of frequency, c 2 (1) = 59.1, p < .0001, length c 2 (1) = 46.9, p < .0001, task c 2 (2) = 73.8, p < .0001, and test, c 2 (1) = 5.2, p < .05. Both word frequency and length had a strong effect on probability of correct performance and the two variables did not interact with other variables in the most appropriate model. Table 5 shows that for all Task x Length combinations performance with 4 For set A, high-frequency words had a mean of 533 (SD 1479), whereas low-frequency words had a mean of 7.9 (SD 6.8). For set B, high-frequency words had a mean of 89.6 (SD 62.2), whereas low-frequency words had a mean of 10.3 (SD 6.9). COGNITIVE NEUROPSYCHOLOGY , 2000, 17 (6) 527 SHALLICE, RUMIATI, ZADINI words are much better than with nonwords. Finally, most of LT’s responses were themselves nonwords: in repetition 66%, in reading 75%, and in writing 71%. Summary The main findings from previous sections are as follows. First, there was no effect of semantic or syntactic structure on performance. Second, LT showed a somewhat superior performance with reading over repetition and writing. Third, the observed effect of lexicality, length, and word frequency were similar to those found in other phonological output buffer patients. Overall, the pattern observed fits the characteristics typical of phonological buffer patients, as specified by Caramazza et al. (1986). By contrast the reproduction conduction aphasic studied by Pate et al. (1987) did not show a word frequency effect; it seems probable that this patient’s impairment was at a more peripheral stage of processing. Error Analysis In the present section a more fine-grained classification analysis is presented. The following classification was used. Single errors, generally involving one phoneme (or letter) (either consonant or vowel), or geminate are categorised as one of the following: 1. Deletion (D). 2. Insertion (I). 3. Substitution (S). 4. Move (an error in which a phoneme/letter moves from its original position within the word). 5. Transposition (T) (an error in which phonemes/letters exchange places within the word). 6. Stress (an error in which a word is read or repeated with the accent on the incorrect syllable). We called double errors any combination of two such errors involving two phonemes/letters (either consonants or vowels), irrespective of their positions within the word. Errors affecting more than two phonemes of a word are called: 528 COGNITIVE NEUROPSY CHOLOGY , 2000, 17 (6) 1. Fragment: an error in which more than one syllable is omitted. 2. Complex: a response that includes more than two errors. 3. Omission: a failure to give any response. However, omissions and fragments have been ignored because they comprised less than 1% of all responses. A quantitative comparison between the magnitude of the errors in the different types of task is shown in Table 6. It is clear that quantitatively repetition and writing are remarkably similar with reading errors being generally quantitatively somewhat smaller, which fits with the relative success rates in the three tasks overall. A more detailed analysis is shown in Table 7. Two conclusions may be drawn from it. First there is a remarkable quantitative and qualitative similarity in error types both between nonwords and words and, to a lesser extent, across tasks. The slight reduction in insertion and deletion errors for reading can again be explained by the fact that reading can be carried out on smaller units than repetition or writing. This is confirmed by the analysis of error position within the word given in Table 8. For reading there is no change with the syllable position, but for all eight repetition and writing comparisons except one there is a consistent increase in error rate through the word. A final analysis concerned the phonological relation between target phoneme and error phoneme in the case of substitutions. The sub-tables for repetition, reading, and writing of words (Table 9 a, b, c) and nonwords (Table 9 d, e, f) again show a related pattern with a significant tendency for phonologically similar errors to occur in all cases but one: rf = .56, c (1) = 11.89, p < .001 for repetition of words; rf = .87, Fisher Exact Test p < .005 for reading words; rf = .61, c (1) = 15.47, p < .001 for writing words; rf = .54, c (1) = 19.88, p < .001 for repetition of nonwords; rf = .46, Fisher Exact Test p < .005 for reading nonwords; rf = .16, c (1) = 1.56, p > .25 for writing nonwords; in all cases comparing stop consonants against nasals plus liquids; considering only errors where the stimulus and responses were stops, nasals, and liquids (so as to retain a sufficiently large cell size). This fits with an explanation in terms of THE PHONOLOGICAL OUTPUT BUFFER Table 6. Percentages of single, double letter, and complex errors (including fragments and omissions) out of the total errors that LT made in repeating, reading, and writing words (set A and set B) and nonwords (set C) Words Set A (N = 406) Tests 1 & 2 Total Errors Set B (N = 208) Tests 1 & 2 Error Type Repetition Reading Writing Single Double Complex 54.0% 25.0% 21.0% 30.5% 62.5% 26.5% 11.0% 22.5% 51.0% 24.5% 24.5% 44.5% Single Double Complex 45.0% 25.5% 29.5% 47.5% 66.0% 20.0% 14.0% 23.5% 56.0% 20.0% 24.0% 42.5% Single Double Complex 44.0% 27.0% 32.0% 70.5% 59.0% 26.0% 15.0% 55.0% 34.5% 28.0% 37.5% 72.0% Total Errors Nonwords Set C (N = 384) Tests 1 & 2 Total Errors Table 7. Percentages of individual types of errors as well as their combinations, out of the total number of single and double errors, that LT made in repeating, reading, and writing words (set A and set B) and nonwords (set C) Error Type Repetition ————————— Words Nonwords Reading ————————— Words Nonwords Writing ————————— Words Nonwords Single Errors Substitutions Deletions Insertions Transpositions Moves Stress 46% 7% 8% 4% 1% 1% 45% 4% 5% 5% 1% – 57% 3% 2% 2% 1% 7% 53% 7% 5% 3% 1% 1% 46% 6% 10% 5% 1% 1% 38% 6% 9% 2% – – Double Errors Multiple Substitutions Multiple Deletions Multiple Insertions Multiple Transpositions Multiple Moves 16% 8% 8% 2% 1% 21% 7% 5% 5% 2% 14% 5% 4% 4% 1% 20% 5% 2% 4% 1% 16% 5% 6% 4% 1% 24% 9% 5% 6% 2% Total 168 184 124 182 201 171 For double errors, the error is counted as half in each of the basic types. Thus a double error Substitution counts one half as a Multiple substitution, one half as a Multiple insertion. + Insertion COGNITIVE NEUROPSYCHOLOGY , 2000, 17 (6) 529 SHALLICE, RUMIATI, ZADINI Table 8. Number of words (set A and set B administered twice) and nonwords (set C, administered twice) that LT failed to repeat, read and write as a function of error position within the word Length Type 2 3 Syllable Repetition Reading Writing Words I Words II Nonwords I Nonwords II Words I Words II Words III Nonwords I Nonwords II Nonwords III 12/258 (5%) 32/258 (12%) 7/136 (5%) 41/136 (30%) 21/210 (10%) 8/210 (4%) 20/210 (10%) 12/194 (6%) 12/194 (6%) 41/194 (21%) 14/258 (5%) 13/258 (5%) 22/136 (16%) 15/136 (11%) 12/210 (6%) 11/210 (5%) 11/210 (5%) 37/194 (19%) 25/194 (13%) 21/194 (11%) 19/258 (7%) 40/258 (16%) 21/136 (15%) 23/136 (17%) 12/210 (6%) 20/210 (10%) 25/210 (12%) 8/194 (4%) 10/194 (5%) 43/194 (22%) This table contains only errors (single or double) affecting only one syllable in the word. Table 9. Substitution errors made by LT in repeating, reading, and writing both words (a, b, c) and nonwords (d, e, f) in terms of the phonemic category of the stimulus and response: Stimuli are plotted by row, responses by column Words Nonwords a. Repetition Stops Nasals Liquids Stops 11 2 2 d. Repetition Nasals 1 7 4 Liquids 3 3 5 b. Reading Stops Stops Nasals Liquids 530 Stops 36 4 2 Nasals 4 5 5 Liquids 5 7 1 Nasals 26 – 1 Liquids 5 1 – 7 5 2 Stops 22 6 6 Nasals 4 3 1 Liquids 6 5 2 e. Reading Nasals 5 1 – Liquids – 4 – 1 5 – c. Writing Stops Nasals Liquids Stops Nasals Liquids Stops Stops Nasals Liquids f. Writing Stops 13 3 3 Nasals – 5 5 Liquids 2 6 4 COGNITIVE NEUROPSY CHOLOGY , 2000, 17 (6) Stops Nasals Liquids THE PHONOLOGICAL OUTPUT BUFFER loss of information in the phonological buffer for repetition and writing, with a reduced effect by use of a different strategy in reading. Experiment 2: Picture naming compared with repetition In this experiment we compared LT’s ability to name pictures to confrontation with his ability to repeat the corresponding names. LT was asked to name 34 line drawings and to repeat the corresponding 34 names, counterbalancing the 2 tasks in an ABBA design. Pictures and words were presented one at a time. On the picture naming task LT scored 19/34 (56%) and on the repetition task 27/34 (79%). In picture naming he made nine single phoneme errors (seven substitutions, one insertion, and one transposition) and three double phoneme errors (substitution plus substitution, substitution plus insertion, and transposition plus substitution), and three complex errors. In the repetition task LT made four single errors (three substitutions, one insertion), two double errors (two insertions plus substitutions), and one complex error. The types of errors LT made in naming pictures are qualitatively similar to those made in repetition. Summary The pattern of errors produced by LT is remarkably similar across reading, writing, and repetition. The basic errors are substitutions, deletions, insertions, transpositions, and phoneme/letter moves. They occur at a similar rate across tasks and in single error, double error, and complex error combinations, with slight differences in reading which are explicable in terms of a more serial/ sequential mode of processing being applied. In phoneme substitutions, the response phoneme tends to be phonologically similar to the target. Overall the pattern is strikingly similar to that obtained in the output phonological buffer patients described in the Introduction. PHONOLOGICAL MANIPULATION AND SHORT-TERM MEMORY Given the relevance for the present study of demonstrating that LT’s errors in repetition are not secondary to a phonological short-term memory deficit, we further investigated his ability to temporarily hold phonological information in a series of digit and word span tasks. Experiment 3: Phonological manipulation Bisiacchi et al. (1989) argued that their patient’s poor reading of nonwords might conceivably have been caused by an inability to manipulate phonemes. They tested this ability by asking him to count sounds in spoken words and nonwords. Since in Italian there is a very high correspondence between graphemes (number of letters) and phonemes (number of sounds), this task could well be carried out using orthographic strategies. They took this possibility into account by comparing performance on items in which the number of sounds corresponds to the number of letters with performance on items in which it does not (e.g., in Italian, ch and sc followed by i or e, gl followed by i, and gn, can correspond to only one phoneme). If a patient were to use orthographic strategies for counting phonemes, then performance should be better with the former set of items. LT was asked to count the phonemes in 30 words and 30 nonwords. Half the words and half the nonwords were characterised by a one-to-one correspondence between letters and sounds. He counted phonemes correctly for 15/15 words and 14/15 nonwords. For the remaining words and nonwords there were more letters than sounds. He counted phonemes correctly for 13/15 of both words and nonwords. Overall, LT’s ability to manipulate phonemes was spared. No significant differences between words and nonwords, or between the same or different number of sound to letter correspondences were observed. Experiment 4: Digit span test In the neuropsychological evaluation LT’s digit span appeared normal. However, very few trials COGNITIVE NEUROPSYCHOLOGY , 2000, 17 (6) 531 SHALLICE, RUMIATI, ZADINI were used. In order to provide further evidence on the integrity of his phonological short-term memory, we ran an additional experiment where 10 trials for each span length, ranging from 4 to 8, were used. The stimuli were presented auditorily by the experimenter and the patient was required to repeat them aloud in the same order in which he heard them. The experiment was replicated 2 months later. As shown in Table 10, LT’s scores confirmed earlier results. To be counted as correct, the digit had to be accurately produced. LT’s performance on this test was not dissimilar to that of five control subjects without history of brain damage and matched for age and education (mean age = 68 years, range 63–74). He made slightly more errors at short span lengths, which is hardly surprising given his speech production problem, but he is insignificantly above the mean at lengths six and seven (it should be noted that Italian digits have two syllables and so span is slightly less than in English). Experiment 5: Probe recognition using words and nonwords To assess the efficiency of auditory-verbal short term memory in LT when speech production problems were eliminated, we carried out a probe recognition task. In this task LT was required to recognise whether a word or a nonword had been previously said by the experimenter. We ran three versions of this task. In the first version, LT had to decide on every trial whether a target word, or a Table 10. The performance of LT and controls on a digit span test No. of Digits 4 5 6 7 8 532 LT —————————– 1st Session 2ndSession 9/10 (90%) 6/10 (60%) 6/10 (60%) 2/10 (20%) 0/5 7/10 (70%) 6/10 (60%) 6/10 (60%) 2/10 (20%) 0/5 Controls ————————— Mean Range 9.6 9–10 8.2 7–10 5.0 2–7 1.4 0–3 0.0. – COGNITIVE NEUROPSY CHOLOGY , 2000, 17 (6) nonword, was one of four just said by the examiner, at the rate of one each second. In the second version we used six words (or nonwords), and eight in the third. When the probe word occurs in the list, its position within the set of the words (or nonwords) to be recognised was varied. LT performed well on this task with both words and nonwords, irrespective of the number of alternatives among which he had to recognise the target (see Table 11). His performance on this test was similar to that of five control subjects without history of brain damage and matched for age and education (mean age = 68 years, range 57–74). Experiment 6: Span and syllable structure of words and nonwords On the Hartley and Houghton (1996) model an improvement in span should be obtained if all words have the same syllabic structure. In order to assess whether his performance was affected by syllable structure, LT was required to repeat triplets of words. The words used had three different types of syllable structure: CCVCV (from now on called (SS1), CVCCV (SS2), and VCCCV (SS3). All the words had the same length (five letters). Words were presented at the rate of one per second. The experiment, which involved two blocks, was run twice. In each block, two kinds of triplets were used. Half of the triplets (N=12) comprised words (N=36) having the same syllable structure (three triplets with SS1, three with SS2, and three with SS3). These triplets of words were called “identical structure.” For the other half, each triplet (N=12) contained words (N=36) each of which had a different syllable structure (SS1, SS2, and SS3). These triplets of words were called “mixed structure.” In the second block, the design and procedure were the same as in the first block. Words contained in the identical triplets were taken from mixed triplets in the first block, while those used in the mixed triplets were taken from the identical triplets of the first block. The presentation of sets of identical and mixed triplets of words was alternated within each block. The first block began with the identical triplets and the second with the mixed ones. Using the same design and procedure, a second experiment THE PHONOLOGICAL OUTPUT BUFFER Table 11. Results obtained by LT on a probe recognition test together with the mean and range of five control subjects Length 4 6 8 Words ———————————— Controls ———————– LT Mean Range Nonwords ———————————— Controls ———————– LT Mean Range 22/24 (92%) 19/24 (79%) 22/24 (92%) 21/24 (92%) 20/24 (83%) 18/24 (75%) 23.0 22–24 21.4 20–23 21.2 20–23 22.4 21–24 21.6 20–23 20.0 18–22 was carried out in which words were replaced by nonwords. Nonwords were obtained by substituting only a single phoneme in each of the words used in the previous experiment. As Table 12a shows, no differences in performance were found between the identical and the mixed conditions for either words or nonwords. A detailed analysis of different error types showed them to be virtually identical both quantitatively and qualitatively across conditions, with no significant differences between them. It is clear from these results that span performance is not affected by repetition of the structure of the syllable involved, in conflict with predictions derived from the model put forward by Hartley and Houghton (1996). Twelve control subjects, matched for age and education to LT, performed clearly better than LT on Experiment 6. However, the prediction derived from the Hartley and Houghton model that repetition of mixed syllable nonwords should be more taxed than repetition of identical syllable nonwords was not supported (see Table 12b). Table 12. Performance on repeating three words of “identical” or “mixed” structure (a) The mean of LT’s correct responses across two testing sessions involving both word and nonword stimuli DISCUSSION Words Correct ————————— Responses Identical Mixed LT is a patient with a number of aphasic problems. However, his difficulties in single word production tasks can be accounted for by a specific deficit within the production system. Thus his performance on single word tasks that did not involve production was far better than if production was required, and when the ability to retain input phonological information was stressed, performance was normal (see Experiment 5). More critically, his error production was quantitatively and qualitatively equivalent, in all conditions, whether the input was presented auditorily (repetition, writing to dictation) or visually (reading, object naming). The one minor exception is that reading performance was quantitatively slightly better for the second and later syllables in a word than in the other conditions (see below), but there was still no quali- 3 2 1 0 2.0 8.0 10.0 4.0 4.5 6.0 10.5 3.0 Nonwords ————————— Identical Mixed 0.0 1.0 11.5 11.5 0.0 2.0 12.5 9.5 (b) The mean number of correct responses involving words and nonwords of 12 control subjects (mean age = 72.8 years, range of education 3–8 years) Words Correct ————————— Responses Identical Mixed 3 2 1 0 17.7 4.9 1.4 0.0 18.2 4.4 1.2 0.2 Nonwords ————————— Identical Mixed 4.3 8.8 7.6 3.4 4.1 7.8 6.9 5.2 Basic pattern of performance COGNITIVE NEUROPSYCHOLOGY , 2000, 17 (6) 533 SHALLICE, RUMIATI, ZADINI tative difference. Thus, there is no major influence on the pattern of errors from any impairment to a particular input system. Turning to the nature of the production difficulties, a number of properties were listed in the Introduction which Caramazza et al. (1986) showed to hold for the nonword production of IGR and which they argued could result from an impairment to the phonological buffer. All were also shown by RR (Bisiacchi et al., 1989) and also by MV (Bub et al., 1987) (except for writing). These properties are all also shown in LT in nonword production tasks and in word production tasks too. 5. For all three basic tasks the predominant error types are single substitution errors and multiple errors containing a substitution. Table 7 also shows that this property holds for LT. 6. The ratio of consonant-vowel errors on substitutions is comparable across all three basic tasks. For LT, 62% of all single substitution errors for nonword repetition were consonants, 57% were for reading nonwords, and 59% were for writing them. While the rate is rather lower than for IGR (repetition 81%, reading 81%, writing 68%), the critical point that there is a close similarity across the three basic tasks. 1. Repeating, reading and writing all produce many errors which have a common pattern. Tables 6, 7, and 9 clearly indicate that this is also true for LT. 2. There is a marked effect of length in all basic tasks. This is also found in LT (see discussion of Table 4). The somewhat better quantitative level of performance on reading fits with Caramazza et al.’s findings, which they interpret in terms of a “part-by-part processing strategies,” which in turn fit with ideas that the phonological reading route can operate on less than a whole-word entry (Coltheart et al., 1993; Làdavas et al., 1997). This is further supported by the relative decrease in substitution errors in reading compared with repetition, and by the reading not showing an increase in errors for later syllables in the word (see Table 8). 3. Error phonemes tend to bear a phonological relation to the target phoneme in all basic tasks. Table 9 shows that this is also true for LT in all three basic tasks. 4. Errors are predominantly phoneme/letter substitutions, insertions, additions, and transpositions or pairs of such errors. Evidence that this is so for LT is presented in Tables 6 and 7. While more complex and omission errors are made by LT than by IGR and RR, at least 60% of his errors in all conditions fit with the prototypic types. The similarity in pattern of performance both across patients and across tasks raises two issues. The first is whether the patients have the same functional syndrome in the sense of Plaut and Shallice (1993), 5 that is whether the nature of the impairment of the functional architecture is qualitatively equivalent for each. Such a hypothesis can never be proved. However, given the controversial nature of the concept of syndrome even in the above sense (see Caramazza, 1986; Shallice, 1988, 1991, for discussion), it is worth considering whether the qualitative similarities extend to quantitative ones. Figure 3 shows the probability of an error involving only a single letter/phoneme plotted against the probability of a correct response.6 In fact, while LT’s score must be to the left of the other patients as he has a generally poorer performance, there is no a priori reason why it should lie on the same curve as those of the other patients. It can be seen that LT’s scores fit well to the same curve as the other patients’. The quantitative similarity can be extended to the rate of substitution errors. Figure 4 shows the probability of a single error being a substitution also plotted against the probability of a correct response. There is a gentle decline with reduction in percentage correct. However, the critical point is that the values of all patients fit the pattern well. Thus LT’s pattern of performance fits with those of the other 5 It should be underlined that a functional syndrome may refer to only part of a patient’s overall impaired performance—see Beauvois et al. (1978) for discussion. 6 For LT as for RR, the only other patient who provides useful data, the word values lie somewhat below the nonword curve. An explanation for this can be given from the attractor-basin account developed later. 534 COGNITIVE NEUROPSY CHOLOGY , 2000, 17 (6) THE PHONOLOGICAL OUTPUT BUFFER Figure 3. Percentage of single phoneme/letter errors made by the four patients on nonwords: LT; IGR (Caramazza et al., 1986); MV (Bub et al., 1987); RR (Bisiacchi et al., 1989); plotted against their corresponding correct rate. For all patients the values for repetition, reading and writing are plotted separately (except for patient MV—writing). The best fitting quadratic curve is given passing through (0,0). patients quantitatively as well as qualitatively. It may provisionally be assumed that their functional syndrome is the same. The second conclusion concerns the relation in the pattern of performance across the three tasks within each patient. LT, like the other three patients, shows not only qualitative similarities across tasks, but also quantitative ones (see Tables 13 and 14). There are some differences between patients as to which of repetition and reading they find easiest. In addition, writing is considerably worse than other tasks in MV and RR. Moreover there appear to be differences between the three other patients, for instance in the rates of insertions by comparisons with deletions. However, for IGR, the prototypical output phonological buffer patient, and for LT, the differences in pattern across tasks in Tables 13 and 14 are really minor. By an analogous logic to that used for similarity of performance across patients, the similarity in pattern of performance across tasks provides evidence that we are not dealing merely with associated deficits. Therefore following the theoretical inference procedure of Caramazza et al. (1986), this implies that a common (set of) component(s) (X) is used in the performance of the three basic tasks—at least in patients like LT—and that they are impaired in LT, or that they are operating normally in LT but damage to another more powerful procedure has left him reliant on them. This second possibility seems implausible, except possibly for writing, where a lexical writing route (see Shallice, 1981) exists for English patients; its status for Italian patients is less clear. COGNITIVE NEUROPSYCHOLOGY , 2000, 17 (6) 535 SHALLICE, RUMIATI, ZADINI Figure 4. Percentage of single errors for nonwords that are substitutions for each of the four patients. For all patients the values for repetition, reading, and writing (if more than 10 single errors) are plotted separately. Table 13. The percentages of single or double or complex errors made by LT and by three other patients: IGR (Caramazza et al., 1986), MV (Bub et al., 1987), and RR (Bisiacchi et al., 1989) for nonword and word stimuli Error Rates Nonwords Single Double Complex % correct Words Single Double Complex % correct Repetition ——————————————— IGR MV RR LT Reading ——————————————— IGR MV RR LT Writing ——————————— IGR RR LT 61 18 21 76 69 31 – 53 95 5 – 87 44 27 32 30 83 17 – 85 71 29 100 ? ? ? 90 – 100 – 98 50 25 25 61 * * * 98 42 83 17 – 63 59 26 15 45 77 18 5 71 58 35 7 40 35 28 37 28 ? ? ? 91 * * * 97 64 24 12 77 76 24 – 97 36 55 9 86 53 23 24 56 * = too few observations to be useful; ? = the values are not given by the authors. 536 COGNITIVE NEUROPSY CHOLOGY , 2000, 17 (6) THE PHONOLOGICAL OUTPUT BUFFER Table 14. The percentages of different types of single errors—involving only one phoneme/letter—made in repeating, reading, and writing nonwords and words by LT and by three other patients: IGR (Caramazza et al., 1986), MV (Bub et al., 1987), and RR (Bisiacchi et al., 1989) Single Errors Nonwords Substitution Insertion/Add. Deletion Transposition Total Words Substitution Insertion/Add. Deletion Transposition Total Repetition —————————————— IGR MV RR LT 81 11 3 5 (63) * 80 8 12 – (114) 80 – 13 7 (15) * Reading —————————————— IGR MV RR LT 75 8 7 9 (110) 81 13 6 – (70) 71 12 10 6 (108) * 62 12 23 3 (50) * 74 6 21 – (34) 79 7 10 4 (122) 88 4 5 1 (81) Writing ——————————– IGR RR LT 68 13 17 2 (114) 83 7 10 – (41) 68 17 11 4 (94) 91 9 75 16 10 8 (122) – – (11) Only conditions where 10 or more errors were made are included. * = too few observations to be useful. As argued by Caramazza et al., the obvious locus for a component common to the three basic tasks is the phonological output buffer. Thus it is in the appropriate part of the speech production system; it is “below” the level of semantic processing and also of any output phonological lexicon, as the findings in LT are similar for words and nonwords. Moreover it is “above” any phonetic processing level and the stage at which the writing process diverges from speech production. Error phonemes tend to bear a phonological relation to their targets. Could the results be explained by a difficulty in generating a phonetic code from a phonological code, independent of short-term memory aspects? These are three characteristics of LT’s impairment that argue against such an account. One is the strong effect of word length. A second related effect is the way that most errors occur in the last syllable in repetition and writing. This fits with a simple trace decay or interference hypothesis. Interestingly, the pattern differs from that typically observed in graphemic output buffer patients where a bow-shaped serial position curve is generally found for errors (e.g., Caramazza et al., 1987; Shallice et al. 1996, but see Badecker, Hillis, & Caramazza, 1990, for an exception) but is more similar to that in short term memory patients (Saffran & Martin, 1990). Finally there is the existence of transposition errors, which are well above the level expected from the base rate of multiple substitutions. However, as will be apparent when we discuss the CQ model, in using the term “output buffer” we do not want to argue for a separable unit with a fixed set of slots. At this point it is useful to consider the disorder in relation to two other putative buffer disorders— the short-term memory syndrome (e.g., Vallar & Shallice, 1990; Warrington & Shallice, 1969) and the graphemic buffer syndrome (Caramazza et al., 1987; Shallice et al., 1996). Thus, characteristic 4 is the complementary characteristic to that occurring for letters (not phonemes) in the graphemic buffer syndrome. Characteristic 5, however, although not being in any sense a difficulty for the phonological buffer hypothesis, is not simply explained by it. Thus, the 15 values for substitutions errors in Table 14 lie between 62% and 91% of all errors. In graphemic buffer patients, for instance, the rate of substitution errors is considerably lower and the rate of such substitution errors lies between 30% and 50%. In Table 15 the maximum rate of transposition errors is 9%. In the graphemic buffer patients reviewed by Shallice et al. (1996), with the exception of one patient with a very mild impairment, all patients made more than 15% transposiCOGNITIVE NEUROPSYCHOLOGY , 2000, 17 (6) 537 SHALLICE, RUMIATI, ZADINI Table 15. A comparison of the rate of single and double phoneme errors in repeating multisyllabic nonwords of LT, IGR, and RR (16 errors only) with that produced by Hartley and Houghton’s (1996) simulation, in which 2–6 syllable nonwords are presented to their model with its time decay parameter reduced in half-life Single Phoneme Errors Substitution Deletion Insertion Transposition Multiple Phoneme Errors Double substitutio n Substitution & Deletion Substitution & Insertion Other double transpositions/insertions Expected on Simulation LT IGR RR 55 21 1 0.3 47 4 4 6 62 2 8 4 75 13 0 6 6 8 9 0 16 4 4 13 18 1 2 1 0 6 0 0 Each column gives the percentages of the overall number of single and double phoneme errors. tion errors. There is thus almost no overlap in these rates between the two syndromes. In STM patients too, substitutions would also be considerably lower and deletions higher. We will return to this issue later. The separability of input and output phonological processes and buffers The findings obtained with LT are relevant to two issues concerned with the relation between input and output phonological processes. Classically, within cognitive neuropsychology the two processing systems and related buffers were held to be different (e.g., Howard, 1993; Howard & Franklin, 1990; Patterson & Morton, 1980; Shallice & Warrington, 1977; Monsell, 1987). However, this view has been increasingly criticised (e.g., Allport, 1984; Campbell, 1990; Martin & Saffran, 1992). One type of support for the separability of input and output phonological processing systems comes from descriptions of patients whose impairment is best described by a disconnection of the two (see Shallice, 1988, Chapter 7 for discussion). One possible interpretation of the pattern of performance observed in IGR and RR is that it results from a disconnection of input and output phonological systems, in other words it corresponds to a classical conduction aphasia (Lichtheim, 1885; McCarthy 538 COGNITIVE NEUROPSY CHOLOGY , 2000, 17 (6) & Warrington, 1984). To account for the overall pattern of reading nonwords one would need to assume that the disconnection also affects the phonological reading route; anatomically such a combination of deficits seems quite possible (see McCarthy & Warrington, 1984; Paulesu, Frith, & Frackowiak, 1993). However, this is not a plausible explanation for LT’s difficulties as he shows the same error pattern in object naming (see Experiment 2), where the phonological output buffer must be accessed from the phonological output lexicon and the semantic system and not from the phonological input buffer, as in repetition of nonwords. Thus the disconnection account is not a strong candidate. In another respect LT’s overall performance strongly supports the existence of both input and output phonological buffers. The pattern of performance shown in certain output phonological buffer patients described earlier has been used to support the position that “the phonological buffer system occupies space within the short-term phonological store” (Bisiacchi et al., 1989). Thus Bisiacchi et al. point out that their patient RR had an auditoryverbal short-term memory impairment. Her digit span was only 3; IGR and MV also had reduced spans. So did the patient GC of Romani (1992), who had an apparently similar spontaneous speech difficulty. However, Romani convincingly shows THE PHONOLOGICAL OUTPUT BUFFER that if GC does not have to produce speech, as for instance in probe tasks, his performance is normal. She therefore argues that GC’s pattern of impairments supports the two-store position. Moreover, a reproduction conduction aphasic of Wilshire and McCarthy (1996) had a relatively normal span performance. This is also the case for LT, whose phonological output buffer deficit was considerably more severe than that of any of the other patients discussed and yet he performed normally on a word probe span task and had a digit span of 6.25, a normal level in Italian speakers. The most critical result is his performance above the mean level for longer strings, as for shorter strings, where normal subjects are at ceiling, the unreliability of an impaired output system will almost inevitably lead to below-normal performance. When considered in the context of patients such as JB (Shallice & Butterworth, 1977; Warrington, Logue, & Pratt, 1971) and PV (Vallar & Baddeley, 1984), who have phonological store impairments and yet can speak normally and can also reproduce single words virtually normally, one has a classic double dissociation. Given the argument presented earlier that LT has an impairment of a phonological store, this supports the assumption that there are (at least) two quite distinct forms of syndrome within the conduction aphasia complex (Shallice & Warrington, 1977)7 and that these correspond to impairments of two quite different stores. In this respect the investigation supports the conclusions drawn from the patients of Romani (1992) and Wilshire and McCarthy (1996). The results do not fit with theories that essentially collapse the two stores into one (e.g., Martin & Saffran, 1992). If LT has a phonological output buffer impairment, why does this not lead to a lower digit span? The presupposition is that the phonological input buffer has much greater capacity than the phonological output buffer (see Shallice, 1975, 1979b) and is no more affected by interference or decay. Then if the relative vulnerability of items in the end and middle positions is roughly comparable across the two stores, when information is lost from the 7 input store it will no longer be represented in the output store. Digits, being very high-frequency items, can in general be output satisfactorily or rehearsed using the damaged phonological output buffer. However this does not apply to words, which leads to a reduced span for such material in LT even though his digit span is intact. Thus the sole role of the output buffer in span is to act as a conduit for the rehearsal or production of each word in turn. One surprising aspect of this theoretical conclusion is the localisation of the lesion. The lesion spares Broca’s area but involves parts of both the supramarginal gyrus (Brodmann area 40) and Wernicke’s area. As the best estimate of the location of the (input) phonological buffer is left hemisphere Brodmann area 40 (Paulesu, et al., 1993), it is possible that we are seeing an example of minor re-localisation of a function into the region adjacent to the lesion. The hypothesis of a route bypassing the phonological output buffer Caramazza et al. (1986), Bub et al. (1987), and Bisiacchi et al. (1989) all interpreted their patients as having a problem essentially limited to nonwords. This led Caramazza et al. to use the syndrome to support a novel architectural proposal. They suggested that a transmission route existed from the phonological output lexicon to a “lexicalarticulatory representation system” that could bypass the phonological output buffer in the production of known words. This hypothesis led to an apparent contradiction, because in normal sentence production the phonological buffer appears to be involved because of the existence of phoneme exchange errors (e.g., Garrett, 1980). Yet IGR did not make the types of errors he made in nonword reproduction when repeating short sentences and a complex ad-hoc explanation had to be produced for the lack of such errors. However, Bub et al., (1987), starting from similar findings, do not make the assumption of a separate route for lexical information. Instead they A third type has also been discussed by McCarthy and Warrington (1984). COGNITIVE NEUROPSYCHOLOGY , 2000, 17 (6) 539 SHALLICE, RUMIATI, ZADINI argue that “lexical levels of representation could be used to compensate for the impoverished quality of traces in the buffer” (p. 105). From the perspective of LT’s performance the Caramazza et al. hypothesis of a bypass procedure is unnecessary. As Table 7 indicates, LT shows just the same quantitative and qualitative pattern of different error types in repeating, reading, and writing words as he does for nonwords. This strongly suggests that whatever impairment affects the processing for nonwords also affects that of words. In addition, LT shows a considerably higher basic level of performance with words than nonwords. Thus the obvious explanation for LT’s pattern of performance is that the two types of stimuli differ in the quantitative amount of the damaged resource they require for adequate production (Shallice, 1979a), which relates to the position advanced by Bub et al. of word representations being less vulnerable than those of nonwords. Would such an explanation also work for IGR, MV, and RR?8 First, their nonword results lie roughly on the same overall functions as those for LT for two variables relating to task performance (see Figures 3 and 4) but are higher on the curves than the results of LT. Thus placing the patients on a single dimension of impairment to an unidimensional resource fits the gross aspects of their performance well. Moreover, from Tables 13 and 14 it can also be seen that their performance with nonword stimuli is qualitatively similar to that of LT with words. One small difference is that single phoneme errors are somewhat less likely in all three patients at a given level of performance for words than for nonwords. This will be discussed later. However, overall it is clear that for all patients their performance is also qualitatively similar for words and nonwords. Most critically, the absolute difference between word and nonword performance for LT is 28–34% across the three basic tasks. The performance of IGR and RR for nonwords on three of the six basic tasks lies between 76% and 87%, so the fact that their performance on words for these tasks is virtually 100% is hardly surprising when considering the resource hypothesis. For the other three tasks, their 8 improvement in performance for words by comparison with nonwords is 34% (RR—reading); 26% (IGR—writing); and 46% (RR—writing), while MV shows 37% and 49% improvements, much the same magnitude on average (38%) as the superiority for words over nonwords shown by LT (32%). Without a quantitative theory of the relation between resource loss and performance, it would be premature to treat this rough equality as strong evidence for the resource position. However it shows that it would be premature to draw any theoretical conclusion from the qualitative dissociation shown by IGR, RR, and MV. The use of the concept resource, moreover, undermines the claim of the philosopher of science, Glymour (1994), that the concept has no utility in the drawing of inferences from neuropsychological evidence to the organisation of normal cognition. We will return later to the issue of how the resource requirements placed on a supposedly nonlexical output buffer can be affected by the lexical status of the stimulus. However, it should be noted that a similar qualitative and quantitative relation between performance on nonword and word stimuli in the graphemic buffer domain has been generally accepted as indicating that both types of stimuli are processed using the same buffer (e.g., Caramazza et al., 1987). Thus there seems to be no neuropsychological justification to propose a route bypassing the phonological output buffer. Expectations based on “Competitive Queuing Theories” The Hartley-Houghton theory is a hybrid between a connectionist model and a symbolic one. It is a localist model trained by Hebbian learning, but also containing much internal structure so as to realise the innate specificity of the speech production system. Such a model transcends the barren ideological debates between supposedly innate structuralist positions and supposedly tabula rasa connectionist ones. It may be seen as a detailed model of how phonological information is temporarily retained in No attempt was made by Hartley and Houghton to fit MVs results, but they do not seem likely to present a problem for the theory. 540 COGNITIVE NEUROPSY CHOLOGY , 2000, 17 (6) THE PHONOLOGICAL OUTPUT BUFFER the speech production system; in other words it can be viewed as one type of implementation of the phonological buffer model, but one where part of the concept of “buffer” disappears. It does not retain the sense of copying temporary representations into a separate store. Indeed, from another perspective, it could be viewed as a more detailed implementation, with respect to its phoneme ordering and short-term storage aspect, of a model such as that of Dell (1986) or with somewhat more licence as to how a phonemic code is generated from higherlevel one. The other great advantage of the HartleyHoughton theory, as mentioned in the Introduction, is its quantitatively accurate predictions of the rates of the different error types in patients IGR and RR, in particular the high rate of substitution errors. How does the theory fare with respect to LT’s errors? Given our previous analysis one would expect to use a different weight-decay parameter from that used to fit the results of IGR and RR. Table 15 compares LT’s performance with the model’s predictions when one uses the same parameter values as in Houghton and Hartley’s fit to IGR and RR’s data. In fact the basic single substitution rate is fitted as well for LT by this parameter as for IGR and RR. However, the addition of LT’s result underlines certain deviations between the patients and the model. The very low predictions for insertions and, in particular, for transpositions suggest that the way the architectural constraints, which are a key part of the model, are realised may make it a little too inflexible as far as variations in phoneme order are concerned. Whether the model would perform better with different weight-decay parameters remains a challenge to the theorist. In any case we assume that with minor alterations the model should be able to deal with all aspects of the findings. One other aspect of LT’s behaviour that needs to be considered with respect to the model is his performance in word reproduction. It is argued by Hartley and Houghton that: “A global increase in the rate of decay in the temporary weights of the network will cause more phonological misorderings for unfamiliar materials,” which fits the findings well. However, they also say: “Familiar materials (words) will not be similarly affected, because the connections in this pathway are permanent. However a general reduction in memory span is to be expected because the connections between the internal content units and syllable units representing both words and nonwords are temporary ones.” This fits well with findings both from IGR and RR in the sense that they both have reduced spans and their performance with word stimuli appears to be qualitatively different from their nonword performance. However, neither of these predictions made by Hartley and Houghton fits with the word production results of LT. In our view the model needs to be revised by separating phonological input buffer processes from phonological output buffer ones. This will mean that under some circumstances memory span performance per se is not relevant to the model. A possible advantage of the competitive queuing approach, however, arises when one compares “phonological buffer” patients such as those discussed in this paper with “graphemic buffer” patients. The two types of patients perform qualitatively very similarly in their respective domains. However, there are interesting quantitative differences. Thus, as discussed earlier, for every phonological buffer patient who performed in the 20–90% correct range for repetition, reading, and writing, over 60% of the single phoneme errors were substitutions (see Table 13). However, for graphemic buffer patients the equivalent rate of substitutions of single letters is 30–55% (see Shallice et al., 1996). Why do these differences occur? One possible explanation, based on the competitive queuing (CQ) model, comes from the assumption of parallel activation of the set of phoneme/letter nodes involved in producing the word and of the internal word structure. The Hartley-Houghton model of spoken word production—a hybrid between a symbolic and a connectionist model—incorporates a complex structure having both between- and within-syllable constraints. This has the effect of “fixing” the ordinal position of a phoneme in the word fairly efficiently—deletions, insertions, and, in particular, transpositions becoming relatively infrequent. For writing, on the Houghton et al., (1994) model, position context constraints are limCOGNITIVE NEUROPSYCHOLOGY , 2000, 17 (6) 541 SHALLICE, RUMIATI, ZADINI ited to representations of each letter’s distance from the word’s beginning and distance from the word’s end; this is a development of the assumptions involved in the typing model of Rumelhart and Norman (1982). The “weaker” structure presumed for the writing process allows more opportunities for errors where the ordinal position of a letter is not maintained, so substitution errors become, relatively speaking, a smaller proportion of all single element errors. However, the existence of the end distance node present in the spelling simulations lead to a bow-shaped pattern of errors in the writing model. Whether other aspects of the graphemic buffer syndrome fit with the idea that it has a less complex internal structure than the phonological output buffer remains a matter of debate. That letter gemination is realised by a separate process is generally agreed (Caramazza & Miceli, 1990; Jonsdottir, Shallice, & Wise, 1996; Miceli, Benvegnu, & Caramazza, 1995; Tainturier & Caramazza, 1996). Moreover, there are strong arguments for the existence of vowel/consonant distinctions (Cubelli, 1991; McCloskey, Badecker, GoodmanSchulman, & Aliminosa, 1994; Miceli et al., 1995). However, the critical issue is whether syllabic organisation is realised though a specific orthographic symbolic structure; here there is a clear conflict of interpretations of empirical evidence (see Jonsdottir et al., 1996, for discussion and Caramazza & Miceli, 1990; McCloskey et al., 1994 for a conflicting perspective). The word-nonword qualitative differences The key empirical results of the present paper are, however, those that support the idea that the production of nonwords and words involves the same mechanisms at the level of phonological buffer processes, and that apparent dissociations between them can be explained in terms of resource differences in the processing of nonword and word stimuli. It should be noted that the quantitative differences in performance comparing word and nonword stimuli is much less in the output buffer case than in, say, phonological alexia (see, e.g., RG, 542 COGNITIVE NEUROPSY CHOLOGY , 2000, 17 (6) Beauvois & Derouesné, 1979; GRN, Shallice & Warrington, 1980; AM, Patterson, 1982). Hartley and Houghton presuppose that there are two types of input to the phoneme nodes, which carry the content information: one for “novel phonological forms” and one for familiar phonological forms, which can be used in isolation from each other. If one accepts the conclusions of the present paper, that the lexical/nonlexical differences at the “phonological buffer” level involve resource rather than structural differences, then how should the lexical/ nonlexical difference be modelled? An alternative to the localist approach adopted by Houghton et al. (1994), however, is to assume that the lexical representation is provided by an attractor structure, as in reading models (e.g., Hinton & Shallice, 1991; Plaut & Shallice, 1993). Then, the existence of learned attractors in the language processing system at a particular level will allow words and nonwords to use the same processing route and yet to show performance differences following lesions. As an analysis by Plaut and Shallice (1993, see p. 411) shows, the existence of attractors “protects” words, but not nonwords, from some of the impairing effects of a lesion, given that the lesion precedes or is at the stage in the processing system which has attractor characteristics. If one assumes that the depth of the attractor is influenced by word frequency, which is natural given that increased word frequency is realised on a model by an increase in training trials (see, e.g., Plaut et al. 1996), then the word frequency effect found would be predicted. Its virtual absence in IGR, RR, and MV would arise through the much smaller degree of impairment in their cases and their being at ceiling. Localist connectionist networks do not naturally have attractor characteristics. However, within the writing domain, Glasspool, Shallice, and Cipolotti (1999) have investigated a multi-level distributed net which has attractor properties realised at the output (letter) level representation, in an analogous way to that of the output phoneme level of Plaut and Shallice (1993). However the Glasspool et al. net attempts to realise a number of characteristics of the localist competitive queuing models by having input units that represent the distance of letter posi- THE PHONOLOGICAL OUTPUT BUFFER tions relative to key points (beginning, end) and by having processes analogous to the suppression of activation following selection of a letter. In the writing domain such a model shows somewhat similar “breakdown” characteristics to that found in the graphemic buffer patients and by the localist competitive queuing model. Moreover, the differences in the behaviour of the model between how it processes word and nonword stimuli are quantitative rather than qualitative. A similar model could in principle be developed as a distributed version of the Hartley-Houghton model and would be expected to show similar properties to that model with respect to the effect of lesions except that the effect would apply to a reduced extent for words and for nonwords. This would therefore fit the pattern shown in LT. Conclusions In this paper we have considered the performance of a patient who, in his nonword performance, was similar to patients who have been characterised as having an “output phonological buffer disorder” in that the effects were qualitatively equivalent for repetition, reading, and writing, with a quantitatively reduced impairment in reading. Moreover, in the pattern of errors produced, the patient essentially resembled these output phonological buffer patients. However, in LT the properties of word processing were qualitatively and in many respects quantitatively equivalent to those for nonword processing. It was argued that the dissociations between the two in the output phonological buffer domain were better explained by “resource” differences between the processing of the two types of stimuli than by qualitative differences in the processing procedures. It was speculated that the overall pattern could be produced by a model that combined an attractor architecture and distributed representations with aspects of the competitive queuing architectures developed by Houghton (1990). However, it remains to be seen whether such a model could be effectively implemented in the speech output domain. Manuscript received 24 July 1998 Revised manuscript received 4 August 1999 Revised manuscript accepted 13 August 1999 REFERENCES Allport, D.A. (1984). Speech production and comprehension: One lexicon or two? In W. Prinz & A.F. Sanders (Eds.), Cognition and motor processes (pp. 209–228). Berlin/Heidelberg: Springer-Verlag. Baddeley, A.D. (1986). Working memory. Oxford: Clarendon Press. Baddeley, A.D., & Hitch, G. (1974). Working memory. In G.H. Bower (Ed.), The psychology of learning and motivation. Advances in research and theory (pp. 47– 89). New York: Academic Press. Badecker, W., Hillis, A., & Caramazza, A. (1990). Lexical morphology and its role in the writing process: Evidence from a case of acquired dysgraphia: Cognition, 35, 205–243. Beauvois, M.F., & Derouesné, J. (1979). Phonological alexia: Three dissociations. Journal of Neurology, Neurosurgery, and Psychiatry, 42, 1115–1124. Beauvois, M.F., & Derouesné, J. (1981). Lexical or orthographic agraphia. Brain, 104, 21–49. Béland, R., Capland, D., & Nespoulous, J.L. (1990). The role of abstract phonological representations in word production: Evidence from phonemic paraphasias. Journal of Neurolinguistics, 5, 125–164. Bisiacchi, P., Cipolotti, L., & Denes, G. (1989). Impairment in processing meaningless verbal material in several modalities: The relationship between shortterm memory and phonological skills. Quarterly Journal of Experimental Psychology, 41A, 293–319. Blumstein, S.E. (1973). Phonological investigation of aphasic speech. Den Haag, The Netherlands: Mouton. Bub, D., Black, S., Howell, J., & Kertesz, A. (1987). Speech output processes and reading. In M. Coltheart, G. Sartori, & R. Job (Eds.), The cognitive neuropsychology of language. Hove, UK: Lawrence Erlbaum Associates Ltd. Buckingham, H. (1992). The mechanisms of phonemic paraphasia. Clinical Linguistics and Phonetics, 6, 41– 63. Burgess, N. (1995). A solvable connectionist model of immediate recall of ordered lists. In G. Tesauro, D. Touretzky, & J. Alspector (Eds.), Neural information processing systems 7. San Mateo, CA: Morgan Kaufmann. COGNITIVE NEUROPSYCHOLOGY , 2000, 17 (6) 543 SHALLICE, RUMIATI, ZADINI Burgess, N., & Hitch, J.G. (1992). Towards a network model of the articulatory loop. Journal of Memory and Language, 31, 429–460. Butterworth, B. (1980). Peggy Babcock’s relatives. In G.E. Stelmach & J. Requin (Eds.), Tutorials in motor behaviour. Amsterdam: North-Holland. Butterworth, B. (1992). Disorders of phonological encoding. Cognition, 42, 261–286. Campbell, R. (1990). Lipreading, neuropsychology, and immediate memory. In G. Vallar & T. Shallice (Eds.) Neuropsychological impairments of short-term memory (pp. 268–286). Cambridge: Cambridge University Press. Caplan, D., Vanier, M., & Baker, C. (1986). A case study of reproduction conduction aphasia: I. Word production. Cognitive Neuropsychology, 3, 99–128. Caplan, D., & Waters, G.S. (1990). Short-term memory and verbal comprehension: A critical review of the neuropsychological literature. In G. Vallar & T. Shallice (Eds.), Neuropsychological impairments of short-term memory (pp. 337–389). Cambridge: Cambridge University Press. Caramazza, A. (1986). On drawing inferences about the structure of normal cognitive systems from the analysis of pattern of impaired performance. The case for single-patient stadies. Brain and Cognition, 5, 41– 66. Caramazza, A., & McCloskey, M. (1985) Number system processing: Evidence from dyscalculia. In N. Cohen, M. Schwartz, & M. Moscovitch (Eds.), Advances in cognitive neuropsychology. New York: Guilford Press. Caramazza, A., & Miceli, G. (1990). The structure of graphemic representation. Cognition, 29, 59–85. Caramazza, A., Miceli, G., Silveri, C., & Laudanna, A. (1985). Reading mechanisms and the organisation of the lexicon: Evidence from acquired dyslexia. Cognitive Neuropsychology, 2, 81–114. Caramazza, A., Miceli, G., & Villa, G. (1986). The role of the (output) phonological buffer in reading, writing and repetition. Cognitive Neuropsychology, 3, 37–76. Caramazza, A., Miceli, G., Villa, G., & Romani, C. (1987). The role of the graphemic buffer in spelling: Evidence from a case of acquired dysgraphia. Cognition, 26, 59–85. Coltheart, M., Curtis, B., Atkins, P., & Haller, M. (1993). Models of reading aloud: Dual-route and parallel distributed processing approach. Psychological Review, 100, 589–608. 544 COGNITIVE NEUROPSY CHOLOGY , 2000, 17 (6) Cubelli, R. (1991). A selective deficit for writing vowels in acquired dysgraphia. Nature, 353, 258–260. Dell, G.S. (1986). A spreading activation theory of retrieval in speech production. Psychological Review, 93, 283–321. De Renzi, E., & Faglioni, P. (1978). Normative data and screening power of a shortened version of the Token Test. Cortex, 14, 41–49. Dizionario di frequenze (1990). Rome: CNR. Dubois, J., Hécaen, H., Angelergues, R., Maufras du Chatelier, A., & Marcie, P. (1964). Etude neurolinguistique de l’aphasie de conduction. Neuropsychologia, 2, 9–44. Ellis, A.W. (1979). Speech production and short-term memory. In J. Morton & J.C. Marshall (Eds.), Psycholinguistic series: Structures and processes (pp. 157– 187). London: Paul Elek. Ellis, A.W. (1980). Errors in speech and short-term memory: The effects of phonemic similarity and syllable position. Journal of Verbal Learning and Verbal Behaviour, 19, 624–634. Fromkin, V.A. (1973). Speech errors as linguistic evidence. The Hague: Mouton. Garrett, M.F. (1980). Levels of processes in sentence production. In B. Butterworth (Ed.), Language production, Vol. I, Speech and talk (pp. 177–220). New York: Academic Press. Garrett, M.F. (1984). The organisation of processing structure for language production: Application to aphasia research. In D. Caplan, A.R. Lecours, & A. Smith (Eds.), Biological perspectives on language. Cambridge, MA: MIT Press. Glasspool, D.W., Shallice, T., & Cipolotti, L. (1999). Neuropsychologically plausible sequence generation. In D. Heinke, G.W. Humphreys, & A. Olsen (Eds.), Connectionist models in cognitive neuroscience. London: Springer. Glymour, C. (1994). On the methods of cognitive neuropsychology. British Journal of Philosophy of Science, 45, 815–835. Hartley, T., & Houghton, G. (1996). A linguistically constrained model of short-term memory for nonwords. Journal of Memory and Language, 46, 1–31. Hinton, G.A., & Shallice, T. (1991). Lesioning an attractor network: Investigations of acquired dyslexia. Psychological Review, 98, 74–95. Houghton, G. (1990). The problem of serial order: A neural model of sequence learning and recall. In R. Dale, C. Mellish, & M. Zock (Eds.), Current research in natural language generation. London: Academic Press. THE PHONOLOGICAL OUTPUT BUFFER Houghton, G., Glasspool, D., & Shallice, T. (1994). Spelling and serial recall: Insights from a competitive queuing model. In G.D.A. Brown & N.C. Ellis (Eds.) Handbook of spelling: Theory, process and intervention (pp. 365–404). Chichester, UK: John Wiley. Howard, D., & Franklin, S. (1990). Memory without rehearsal. In G. Vallar & T. Shallice (Eds.) Neuropsychological impairments of short-term memory (pp. 287–318). Cambridge: Cambridge University Press. Jonsdottir, M., Shallice, T., & Wise, R. (1996). Language-specific differences in graphemic buffer disorder. Cognition, 59, 169–197. Kohn, S.E., & Smith, K.L. (1990). Between-word speech errors. Cognitive Neuropsychology, 7, 133–156. Làdavas, E., Shallice, T., & Zanella, M.T. (1997). Preserved semantic access in neglect dyslexia. Neuropsychologia, 35, 257–270. Lecours, A.R., & Lhermitte, F. (1969). Phonemic paraphasias: Linguistic structures and tentative hypotheses. Cortex, 5, 193–228. Levelt, W.J.M. (1989). Speaking. From intention to articulation. Cambridge, MA: MIT Press. Lichtheim, L. (1885). On aphasia. Brain, 7, 433–484. Martin, N., & Saffran, E.M. (1992). A computational account of deep dysphasia: Evidence from a single case study. Brain and Language, 43, 240–274. McCarthy, R., & Warrington, E.K. (1984). A two-route model for speech production. Evidence from aphasia. Brain, 107, 463–485. McCloskey, M., Badecker, W., Goodman-Schulman, R.A., & Aliminosa, D. (1994). The structure of graphemic representation in spelling: Evidence from a case of acquired dysgraphia. Cognitive Neuropsychology, 11, 341–392. McCloskey, M., & Caramazza, A. (1991). On crude data and impoverished theory. Behavioural and Brain Sciences, 14, 453–455. Miceli, G., Benvegnu, B., & Caramazza, A. (1995). Selective deficit in processing double letters. Cortex, 31, 161–171. Miceli, G., Burani, C., & Laudanna, A. (1991). Batteria per l’Analisi dei Deficit Afasici. Milano: Associazione per lo Sviluppo delle Ricerche Neuropsicologiche. Miceli, G., Silveri, C., & Caramazza, A. (1985). Cognitive analysis of a case of pure dysgraphia. Brain and Language, 25, 187–212. Monsell, S. (1984). Components of working memory underlying verbal skills: A “distributed capacities” view. In H. Bouma & D.G. Bouwhuis (Eds.), Attention and performance X. Control of language processes (pp. 327–350). Hove, UK: Lawrence Erlbaum Associates Ltd. Morton, J. (1969). Interaction of information in word recognition. Psychological Review, 76, 165–178. Pate, D.S., Saffran, E., & Martin, N. (1987). Specifying the nature of the production impairment in a conduction aphasic: A case study. Language and Cognitive Processes, 2, 43–84. Patterson, K.E. (1982). The relation between reading and psychological coding: Further neuropsychological observations. In A.W. Ellis (Ed.), Normality and pathology in cognitive functions. London: Academic Press. Patterson, K.E., & Morton, J. (1980). “Little words—no! In M. Coltheart, K.E. Patterson, & J.C. Marshall (Eds.), Deep dyslexia. London: Routledge. Paulesu, E., Frith, C.D., & Frackowiak, R.S.J. (1993). The neural component of working memory. Nature, 362, 342–345. Plaut, D.C., McClelland, J.L., Seidenberg, M.S., & Patterson, K.E. (1996). Understanding normal and impaired word reading: Computational principles in quasi-regular domains. Psychological Review, 103, 56– 115. Plaut, D., & Shallice, T. (1993). Deep dyslexia: A case study of connectionist neuropsychology. Cognitive Neuropsychology, 10, 377–500. Romani, C. (1992). Are there distinct input and output buffers? Evidence from an aphasic patient with an impaired output buffer. Language and Cognitive Processes, 7, 131–162. Rumelhart, D.E., & Norman, D. (1982). Simulating a skilled typist: A study of skilled motor performance. Cognitive Science, 6, 1–36. Saffran, E., & Martin, N. (1990). Short-term memory impairment and sentence processing: A case study. In G. Vallar & T. Shallice (Eds.), Neuropsychological impairments of short-term memory (pp. 428–447). Cambridge: Cambridge University Press. Shallice, T. (1975). On the contents of primary memory. In P.M.A. Rabbitt & S. Dornic (Eds.), Attention and performance (Vol. 5). London: Academic Press. Shallice, T. (1979). Case-study approach in neuropsychology. Journal of Clinical Neuropsychology, 1, 183–211. Shallice, T. (1981). Phonological agraphia and the lexical route in writing. Brain, 104, 413–429. Shallice, T. (1988). From neuropsychology to mental structure. Cambridge: Cambridge University Press. COGNITIVE NEUROPSYCHOLOGY , 2000, 17 (6) 545 SHALLICE, RUMIATI, ZADINI Shallice, T. (1991). How neuropsychology informs an understanding of normal function: Author’s response to commentaries. Behavioral and Brain Sciences, 14, 457–470. Shallice, T., & Butterworth, B. (1977). Short-term memory impairment and spontaneous speech. Neuropsychologia, 15, 729–735. Shallice, T., Glasspool, D.W., & Houghton, G. (1996). Can neuropsychological evidence inform connectionist modelling? Analyses of spelling. Language and Cognitive Processes, 10, 195–225. Shallice, T., & Warrington, E.K. (1977). Auditory-verbal short-term memory impairment and conduction aphasia. Brain and Language, 4, 479–491. Shallice, T., & Warrington, E.K. (1980). Single and multiple component central dyslexic syndromes. In M. Coltheart, K.E. Patterson, & J.C. Marshall (Eds.), Deep dyslexia. London: Routledge. Shattuck-Hufnagel, S. (1979). Speech errors as evidence for serial-ordering mechanism in sentence production. In W.E. Cooper & E.C.T. Walker (Eds.), Sentence processing: Psycholinguistic studies presented to Merrill Garrett. Hillsdale, NJ: Lawrence Erlbaum Associates Ltd. 546 COGNITIVE NEUROPSY CHOLOGY , 2000, 17 (6) Sperling, G. (1967). Successive approximations to a model for short-term memory. Acta Psychologica, 27, 285–292. Spinnler, H., & Tognoni, G. (1987). Standardizzazione e taratura italiana di test neuropsicologici. The Italian Journal of Neurological Sciences, 8, 1–120. Tainturier, M.J., & Caramazza, A. (1996). The status of double letters in graphemic representation. Journal of Memory and Language, 35, 57–73. Vallar, G., & Baddeley, A.D. (1984). Phonological short-term store, phonological processing and sentence comprehension: A neuropsychological case study. Cognitive Neuropsychology, 1, 121–142. Vallar, G., & Shallice, T. (Eds.) (1990). Neuropsychological impairments of short-term memory. New York: Cambridge University Press. Warrington, E.K., Logue, V., & Pratt, R.T.C. (1971). The anatomical localisation of selective impairment of auditory-verbal short-term memory. Neuropsychologia, 9, 377–387. Warrington, E.K., & Shallice, T. (1969). The selective impairment of auditory-verbal short-term memory. Brain, 92, 885–896. Wilshire, C.E., & McCarthy, R. (1996). Experimental investigations of an impairment in phonological encoding. Cognitive Neuropsychology, 13, 1059–1098.
© Copyright 2026 Paperzz