Can developmental disorders reveal the component parts of the human language faculty?

Differential profiles of language impairments in genetic developmental disorders have been argued to reveal the component parts of the language system and perhaps even the genetic specification of those components. Focusing predominantly on a comparison between Williams syndrome and Specific Language Impairment, we argue that the detailed level of behavioural fractionations observed in these disorders goes beyond the possible contribution of genes and implicates the developmental process as a key contributor to the cognitive outcome. Processes of compensation and interaction across development make highly specific developmental deficits unlikely, and in line with this view, even in SLI (a paradigmatic example of a supposedly selective deficit) the actual level of specificity remains controversial. We consider the challenge of characterising the (atypical) developmental process from the perspectives of brain development, cognitive development, and computational modelling. Failure to take up this challenge leaves many current explanations of developmental deficits at best ill-specified and at worst implausible. Thomas & K-S paper – page 3 Introduction Selective impairments in adult neuropsychological patients arguably reveal the component parts of the human language faculty. Indeed, individuals who have develo ped normally but sustain brain damage in adulthood often present with conditions in which a single aspect of language is differentially impaired. For example, patients with agrammatism produce well-formed lexical items, with appropriate pragmatics, but which are strung together without appropriate morphosyntactic markers. In contrast, other patients present with normal grammar and pragmatics but with severe word-finding difficulties. Yet others produce normal language in terms of structure and semantics, but display inappropriate pragmatics such as poor assessment of the situations in which certain registers of language output are fitting. From this, one might conclude that different parts of the human brain are specialised in adulthood for the different component parts of language. In the absence of an obvious account of how such a specialised modular structure might have developed, one temptation might be to conclude that the coarse structure of language is already pre-specified and under genetic control, with segregated modules to deal independently with the lexicon, the morphology, the syntax and the pragmatics. However, although the adult brain may be specialised and localised for different aspects of language sufficient to generate selective impairments, to address the question of the possible innate specification of these different components, one cannot focus solely on adult neuropsychological patients. This is because it cannot be assumed that infants start life with the same brain structure as adults. Ontogenetic development may play a Thomas & K-S paper – page 4 significant role in giving rise to adult specialisation and localisation of function. It could be the case that higher-level modules such as morphology, syntax, the lexicon and pragmatics are an emergent product of development rather than its starting state (Karmiloff-Smith, 1992). But what if developmental disorders of genetic origin point to the same fractionations of language as in cases of adult neuropsychology? This might argue in favour of the innate specification of the different components of our language faculty. So, do functional aspects of language come apart in developmental disorders? And if we do find such differential disability across atypically developing language system, what inferences may be draw about the development of the normal language system? Cross-syndrome comparisons have indeed identified apparent dissociations in the development of components of language. For example, in a comparison of Down syndrome (DS), Williams syndrome (WS), autism and Fragile X (FraX), Fowler (1998) described dissociations between phonology, lexical semantics, morphosyntax and pragmatics. These disorders illustrate that different aetiologies can have dramatically different linguistic profiles. General cognition is not a reliable indicator of language function in children with learning disabilities (although this is a correlation that we must take with some caution, since the holistic notion of IQ or mental age (MA) is less valid in disorders with uneven cognitive profiles). While language acquisition typically lags behind mental-age level expectations in children with learning disabilities, Fowler noted that disorders such as Williams syndrome and hydrocephalus with associated myelomeningocele appear superficially to represent exceptions to this pattern. From her Thomas & K-S paper – page 5 comparison, Fowler concluded that pragmatics and lexical semantics are more closely tied to MA than phonology and morphosyntax. Tager-Flusberg and Sullivan (1997) carried out a similar comparison of the four disorders, but this time seeking possible asynchronies in the early development of semantic, grammatical, and pragmatic aspects of language. These authors also noted disparities in areas such as vocal development, social communicative development, gesture, lexical development, phonological development, early grammar and pragmatics. Although Fowler (1998) and Tager-Flusberg and Sullivan (1997) identified differences in the language development of DS, WS, autism and FraX, they also noted similarities across the disorders. For example in early development, there were consistent patterns of errors displayed in speech articulation; and in morphosyntax, although some disorders stopped short of mastery, the order of acquisition of syntactic structures appeared similar. These similarities prompted the same conclusions in the two reviews: that the common patterns reflected constraints in the underlying brain mechanisms of motor articulation and syntax acquisition, respectively, and that, as summarised by Fowler, the ‘non-deviant development is consistent with a model of language acquisition that is heavily constrained by the brain that is acquiring the language’ (1998, p. 309). Such a position downplays the possibility that commonalities in development could also arise from the structure of the shared problem domain to which all individuals are exposed, a possibility to which we return later. More importantly for current purposes, some theorists have taken the combination of differential patterns of atypical language Thomas & K-S paper – page 6 acquisition and the genetic basis of these disorders to draw stronger inferences about links between genes and components of the language system: ‘...Overall, the genetic double dissociation is striking... The genes of one group of children [Specific Language Impairment] impair their grammar while sparing their intelligence; the genes of another group of children [Williams syndrome] impair their intelligence while sparing their grammar.’ (Pinker, 1999, p. 262). ‘In 1998 [researchers] linked the [KE] disorder to a small segment of chromosome 7, which they labelled SPCH1. Now...Lai et al. [Nature October 2001] have narrowed the disorder down to a specific gene, FOXP2...The discovery of the gene implicated in speech and language is amongst the first fruits of the Human Genome Project for the cognitive sciences. Just as the 1990s are remembered as the decade of the brain and the dawn of cognitive neuroscience, the first decade of the twentyfirst century may well be thought of as the decade of the gene and the dawn of cognitive genetics.’ (Pinker, 2001, p. 465). Our interest in this paper is to explore the kinds of links that might be appropriate between a cognitive and a genetic level of description, sufficient to motivate the notion of a ‘cognitive genetics’. Note that cognitive genetics as a discipline would appear to make stronger claims than ‘behavioural genetics’ about the relationship between particular genes and particular cognitive structures. Behavioural genetics is concerned with Thomas & K-S paper – page 7 correlations in behavioural scores between individuals with different levels of relatedness (that is, different probabilistic proportions of shared genes). If it is to be a separate discipline, cognitive genetics would seem to imply links between individual genes (or sets of genes) and individual cognitive structures. For example, Pinker indicates the sort of theoretical account his cognitive genetics would license, again in the context of Williams syndrome. Here, a single gene is postulated to contribute to cognitive structures for spatial reasoning but not language or face processing: ‘Presumably LIM-kinase1 [one of the 25 genes deleted from one copy of chromosome 7] plays an important role in the development of the neural networks used in spatial reasoning, possibly in the parietal lobes. The other missing genes, perhaps, are necessary for the development of other parts and processes of the brain, though not for language or face perception.’ (Pinker, 1999, p. 260-261). We do not believe that many researchers hold the view that individual genes uniquely specify individual cognitive structures in the adult, although quotes such as those above can sometimes be read in that light (particularly when they use terminology such as ‘spared’ and ‘impaired’ modules to describe the endstate of genetic developmental disorders). However, Pinker’s comparison of Williams syndrome and Specific Language Impairment (SLI) provides a useful basis for discussion of the relevant issues. The structure of this article unfolds as follows. First, we argue that to be plausible, links between genes and adult cognitive structure are contingent on a certain level of Thomas & K-S paper – page 8 granularity of fractionation in the language system (for example, a fractionation between the lexicon and the grammar). We then review evidence from Williams syndrome demonstrating that fractionation actually occurs on a much finer level than is plausible for the expression of genetic effects. Second, we argue that links between genes and differentially impaired cognitive structures in adults with developmental disorders rely on accurate descriptions of the specificity of cognitive deficits in these adults. We then review evidence from SLI indicating a current lack of agreement on specificity of outcome. In the second half of the article, we argue that that an essential component of an explanation of the detailed fractionations that we do see in developmental disorders is the developmental process itself (Karmiloff-Smith, 1998). Of course, characterising the developmental process is a significant challenge. Here, we address this challenge in four ways. First, we review the mechanisms the brain offers for genes to specify particular cognitive structures. Second, we discuss two features that any explanation of the atypical developmental process must incorporate at the cognitive level: interactivity and compensation. Third, in the face of the challenge of development, some theories simply ignore its contribution (see critical discussions in Karmiloff-Smith, 1992, 1998; Thomas & Karmiloff-Smith, 2002a). We demonstrate that even a rudimentary developmental process applied to such theories finds them wanting as coherent endstate descriptions of an atypical language system. Finally, we discuss one possible method to aid characterisation of both typical and atypical developmental processes, that of computational modelling. Thomas & K-S paper – page 9 Developmental fractionations of the language system: Williams syndrome and Specific Language Impairment Intuitively, the search for links between genes and components of the language system rests on the assumption that genes target a particular level of cognitive granularity. Thus, previously we saw research that pointed to developmental dissociations between phonology, lexical-semantics, morphosyntax and pragmatics across different genetic disorders. One might perhaps imagine that genes could reduce the efficiency of memory systems in the developing brain or motor articulation systems. However, one wouldn’t perhaps expect genes to target the acquisition of semantic categories like vehicles but not tools, or to target the ability to form present tense verb inflections but not past tense inflections. The idea of granularity remains somewhat intuitive because the specificity of outcome is itself up for debate. But let us bear in mind the intuition of a plausible level of cognitive granularity for genetic fractionations as we consider the example of language development in Williams syndrome. Williams syndrome The genetic disorder, Williams syndrome (WS), involves the deletion of some 25 genes on one of the copies of chromosome 7 (see Donnai & Karmiloff-Smith, 2000, for full details of the syndrome). It has been hailed as an example of the so-called sparing of the language faculty in general or of certain component parts of language in particular. For example, Clahsen and Almazan (1998) argue that grammar develops normally but the lexicon develops atypically and sub-optimally. Individuals with WS usually present with IQs in the 50-60s range, with exceedingly poor spatial and numerical cognition. Yet, Thomas & K-S paper – page 10 despite this, their language seems surprisingly proficient. Indeed, although there is an initial delay in language development, by adolescence and adulthood many individuals display large vocabularies that co-exist with relatively good scores on standardised grammatical tests. Could it then be that language (or grammar) develops independently of intelligence, under the control of a different set of genes, as Pinker (1999) claims? It is worth recalling (as argued elsewhere, Karmiloff-Smith, 1998) that IQ scores rather than MA can give a misleading picture of the situation. Most individuals with WS who possess high vocabularies and proficient syntax also have an MA of around 7-9 years, the time at which all normal children have developed rather sophisticated language. That is, language can be made to look impressive given IQ, but is unremarkable given MA. Nevertheless, language emerges late in development in WS which, given their extreme attention to sounds and faces early in development, seems at first blush to be inexplicable. Is the initial delay in getting WS language off the ground merely the late maturation of a set of language-specific genes, or are there more complex developmental reasons? Studies from at least four different laboratories in the UK, USA and Italy have shown that WS infants and toddlers are surprisingly late in their language development (Mervis & Bertrand, 1997; Paterson et al., 1999; Singer-Harris, Bellugi, Bates, Jones & Rossen, 1997; Vicari, Brizzolara, Carlesimo, Pezzini, & Volterra, 1996; Volterra, Capirci, Pezzini, Sabbadini, & Vicari, 1996). Why is this? Several factors interact. Our studies of infants’ segmentation of the speech stream show that those with WS are some 10-20 Thomas & K-S paper – page 11 months behind their typical controls in segmenting words out of the speech stream (Nazzi, Paterson & Karmiloff-Smith, 2002). This difficulty will contribute to the delay. Furthermore, despite their abilities with dyadic interaction, infants and toddlers with WS are surprisingly atypical in triadic interaction and in their understanding of the referential function of pointing (Mervis & Bertrand, 1997; Laing et al., 2002) – one of the ways in which children normally learn new words. In addition, while they behave like controls in mapping similarities between perceptual features of objects, toddlers with WS are significantly poorer than controls at using words to map identity of object categories (Nazzi & Karmiloff-Smith, 2002). What about older children who by then have quite well developed language? Was their language simply delayed and then subsequently followed a normal developmental trajectory? This appears not to be the explanation. Brain imaging data hints that underlying structures may be atypical. Mills, Neville and their collaborators showed that children with WS have atypical hemispheric specialisation with respect to the difference between open and closed class words (Mills, Alvarez, St.George, Appelbaum, Bellugi, & Neville, 2000), suggesting a lack of the normal progressive specialisation and localisation of brain function in WS compared to controls. Behavioural studies have also revealed subtle differences in cognitive processing. For example, Karmiloff-Smith, Tyler, Voice, Sims, Udwin, Howlin, and Davies (1998) found that when individuals with WS monitored a sentence for a target word, performance was disrupted by syntactic violations except when those violations involved lexically based information (e.g., subcategory constraints such as transitive/intransitive). This led the authors to propose Thomas & K-S paper – page 12 that in WS, there is a deficit in integrating different sources of linguistic information in real-time processing. In some respects, the developmental trajectory appears normal in WS. Thus, Mervis, Morris, Bertrand, and Robinson (1999) noted that, while the syntactic abilities of children with WS (39 children from 2 years 6 months to 12 years) were considerably delayed, syntactic complexity was nonetheless appropriate for the mean length of utterance (MLU). This contrasts with DS, autism and FraX, where syntactic complexity turned out to be less than would be expected at MLUs over 3. This result prompted Mervis et al. to claim that WS is the first syndrome in which the normal relation between utterance length and complexity has been demonstrated. Might this constitute evidence of a normal developmental trajectory within an isolated domain-specific module for grammar (despite the atypical patterns of brain activation)? The answer again appears to be no. On closer inspection, there turn out to be inconsistencies in the pattern of grammatical development in WS. First, there are more errors in morphology (verb tense agreement, personal pronouns, grammatical gender – Karmiloff-Smith et al., 1997; Volterrra et al., 1996) than in syntax. Second, although syntactic performance is often broadly in line with MA-controls (Zukowski, 2001), within syntax itself WS reveals fractionated development that is appropriate neither with respect to chronological age nor mental age (Grant, Valian, & Karmiloff-Smith, 2002; Mervis et al., 1999). For example, Mervis et al. (1999) reported that while the syntactic complexity scores of children with WS were significantly higher than would have been Thomas & K-S paper – page 13 expected on the basis of spatial constructive ability, they were nevertheless significantly lower than would have been expected on the basis of receptive vocabulary ability, verbal ability, or auditory short-term memory. Across a large sample of 77 individuals between 5 and 52 years, Mervis et al. (1999) reported that performance on the Test of Receptive Grammar (Bishop, 1983) was poor for complex constructions. Only 18% of the participants (22% of the adults) passed the test block that assessed relative clauses and only 5% (9% of the adults) passed the block assessing embedded sentences. This fine level of fractionation within grammar acquisition brings us back to the wider issue of granularity. The dissociations we find occur within domains to a degree of specificity of language structure that seems beyond the reach of anything like targeted gene expression (Thomas, in press a). The deep level of fractionation is a pattern that that reappears in other areas of WS language. Thus pragmatics, less advanced in WS than grammar, also exhibits within-domain fractionation. There is relatively good performance in social sensitivity (e.g., making dyadic eye contact, sensitivity to non-verbal cues) but problems in areas such as greeting behaviours, topic maintenance, and question answering (Semel & Rosner, 2003). In lexical-semantics, a relative strength in category concepts (e.g., animals vs. clothing) contrasts with problems understanding semantic relational concepts such as spatial-temporal terms. Even within category concepts, recent evidence has indicated differential naming problems across categories (Temple et al., 2002; Thomas et al., 2004). It seems unlikely that genetic events are uniquely to blame for each of these fractionations. Thomas & K-S paper – page 14 Outside the domain of language, the fractionation proceeds apace (see Semel & Rosner, 2003, for a review). Although sociability is a strength in WS (Bellugi et al., 2000), within sociability there is a fractionation between friendliness / success with adults, and the disinterest or ineptness shown when interacting with peers. There is a fractionation between the sensitivity of individuals with WS towards others’ emotions and the difficulty they often exhibit in respecting the private space of peers. Within the domain of memory, there are fractionations between relative skill in phonological working memory tasks (e.g., in digit span) but poor performance in visuo-spatial memory tasks (e.g., Corsi span). Within phonological memory itself, there is a fractionation between a strength in learning words but not in learning to read phonologically similar words (Laing et al., 2001) nor in repeating non-words (Grant et al., 1997). There is a strength in remembering semantically salient items like poems, stories, and songs over long periods, but not in learning or retaining facts over a few minutes (Semel & Rosner, 2003). To these we may add the domain of numeracy, where children with WS reveal a weakness in global quantity judgements, but mental-age appropriate learning of the count sequence (Ansari et al., 2003). And there is a highly salient dissociation between weaknesses in some visuo-perceptual skills (e.g., deciding which of two lines is longer) and a strength in recognising faces, which has been the focus of much recent research (see KarmiloffSmith et al., in press; Tager-Flusberg et al., 2003, for discussion). In short, what started out as a neat theoretical example of a disorder where language develops normally (and uniformly), in the face of general (and uniform) cognitive impairment, turns out to be a disorder suffused by cross-domain and within-domain Thomas & K-S paper – page 15 fractionations. Language ability in WS is clearly unusual when contrasted with other disorders. However, the fractionations are so detailed that a genetic explanation is unlikely to account for much of the variation. Moreover, evidence from the brain level is suggestive that, in terms of localisation and specialisation of language function, development has not proceeded normally in WS. Let us now turn to Specific Language Impairment (SLI), proposed as a neat theoretical example in which language is impaired alongside normal intelligence. SLI is a disorder that, by its very definition, promises a tidier developmental fractionation. Specific Language Impairment To establish potential links between genes and differentially impaired cognitive structures in developmental disorders, it is important to establish the specificity of the cognitive deficits. This will set bounds on the locus of the putative genetic effects on cognition. For Pinker’s “genetic double dissociation” (1999), Specific Language Impairment presents the opposite case to WS. That SLI has a genetic component is clear. It runs in families, particularly in males, and twin studies have shown a strong genetic component to the disorder (Bishop, 1992) even though molecular genetics has yet to identify all the genes that may contribute to the outcome. In 1998 and more specifically in 2001, the scientific world became very excited about the discovery of what came to be known as “the gene for speech and language”. A British Thomas & K-S paper – page 16 family, the now well-known KE family, had been identified in whom an allelic variation in the FOXP2 gene in some family members gave rise to serious impairments in speech and language, whereas family members without this allelic variation developed language normally (Pinker, 2001; Vargha-Khadem et al., 1998). Is this a gene that is novel to the human genome and could explain the onset of language and its component parts in the human species? Does such a gene have a unique and specific effect on speech and language in humans? Some researchers have claimed that this might indeed be the case (Pääbo, 1999). Once more, however, a closer look at the phenotypic outcome in the affected KE family members highlights the need for a more complex explanation. To begin with, the deficits are not specific to language, nor even to speech output. The dysfunctions in the affected family members not only involve oral-facial movements, but also particular aspects of the perception of rhythm as well as the production of rhythmic movements of the hands (Alcock, 1995; Watkins, Dronkers, & Vargha-Khadem, 2002). How these impact over developmental time in the language outcome has still to be clarified. Moreover, at the behavioural level it is far from clear what we can ascribe to the action of this gene in terms of its contribution to human language. When a genetic mutation causes dysfunction of a particular behaviour, this does not mean that intactness of that same gene causes the proper functioning of the behaviour. (Numerous analogies make this point obvious. For example, if the carburettor of a car is not functioning properly, the car will not run. But it is not the carburettor that explains how the car runs in normal circumstances.) More importantly, it is highly unlikely that a single gene or even specific set of genes will Thomas & K-S paper – page 17 explain the development of human language. In the vast majority of cases, genes involve many-to-many mappings, not one-to-one mappings. The genes that affect the outcome of language structures are likely to influence other brain structures as well. Evidence from the brain level supports this view. Detailed research on the KE family has revealed widespread structural and functional brain differences in affected family members, beyond those areas of the brain typically associated with language function in normal adults (e.g., Watkins et al., 2002). This should not come as a surprise. There is clear precedent that single gene disorders can produce phenotypic outcomes with multiple impairments. Fragile X is one such example. In this disorder, a single mutated gene produces widespread alterations because the gene in question is deeply involved in synaptogenesis across the whole developing system (Scerif et al., 2004). FraX is associated with the silencing of a single gene, FMR1, whose gene product (FMRP) is normally involved in mechanisms of experience-dependent plasticity throughout the brain (Churchill et al., 2002, Greenough et al., 2001). Although the deficit to this domain-general feature is indeed associated with generalised delay, FraX also exhibits an uneven cognitive profile in the adult phenotype. It is characterised by relative strengths in vocabulary, long-term memory and holistic information processing, but relative weaknesses in visuospatial cognition, attention, short-term memory and sequential information processing (Cornish, Munir, & Cross, 1999, 2001; Freund & Reiss, 1991). The uneven cognitive profile results from a complex interaction of FMRP with other proteins across development, presumably triggering a series of imbalances that have cascading effects on other elements of the developmental pathway Thomas & K-S paper – page 18 at differing times through ontogeny (Scerif et al., 2004; see Scerif, 2003, for a discussion). Thus, a brain-wide general change at the cellular level may have differential, seemingly domain-specific outcomes, via interactions across developmental time (see Karmiloff-Smith & Thomas, 2003, for discussion). Discussion of cellular level differences in developmental disorders may seem remote from language outcome. How does one link synaptogenesis to language development? The point here is that genes are even more remote from language development, yet cognitive genetics is premised on establishing just such links between genetic mutations and language impairments. Bates has frequently stressed (e.g., Bates, 1997) that there are numerous ways in which language can end up being impaired. These include genetic mutations in many parts of the genome, as well as social and other causes. It is therefore premature at the least to imagine that a single gene or specific set of genes will be our best bet for explaining language impairments. Indeed, the shorthand of “the gene for language” is a particularly dangerous one (see discussion in Karmiloff-Smith, 1998). Bates’ subtle position on this debate is of particular relevance (Bates, personal communication, September 2002). Like us, she in no way denied that genes play a crucial role in human development in general and in language in particular. Rather, the multiple functions that each gene may have, including genetic contributions to language over evolutionary time, need to be considered. Amongst these Bates identified: (a) genetic alterations that gave us better fine motor control (like FOXP2); (b) genetic alterations that gave us better perceptual Thomas & K-S paper – page 19 abilities; (c) genetic alterations that permitted a more direct mapping from perception to production (cross-modal perception which is essential for imitation, a major tool of cultural transmission); (d) genetic alterations that made us faster information processors; (e) genetic alterations that led to the particular social make-up that makes us want to imitate each other and think about what other people are thinking. In Bates’ view, none of these genes will end up being specific to speech/language, and yet all of them will be important for the emergence of speech, language, culture and technology in our species. Such an interactive, emergentist position is far removed from the notion of a specific set of genes solely for language. Despite these cautionary remarks, some researchers remain convinced that SLI is the key to unveiling the genetic determination of different component parts of the human language faculty, and maintain that components of grammar will turn out to be domainspecific and ‘genetically controlled’ (e.g., van der Lely, 1997, 1999). For these researchers, the KE family was a false dawn but the sky in the East continues to brighten. There are many children who fall under a behavioural definition of SLI, that is, a developmental disorder of language (or, more precisely several different developmental disorders of language with different causes, Bishop 1997) found in the absence of frank neurological damage, hearing deficits, severe environmental deprivation, or learning disability (Bishop, 1997; Leonard, 1998). A careful behavioural screen of these children may reveal cases where the deficits are restricted to certain aspects of language, and perhaps even very rare cases where the deficits are restricted to specific aspects of syntactic structure (Tomblin & Pandich, 1999; van der Lely, 1999). The heritability of Thomas & K-S paper – page 20 general SLI provides the promissory note that the causes of the behavioural deficits in childhood are genetic, with the precise gene or genes to be revealed at a later date. However, behaviourally defined case studies necessarily confound the contribution of genotype, individual variability, and a particular history of interaction with the environment. They therefore provide at best ambiguous evidence of the specific contribution of genes to language structures. The single-gene version of SLI has turned out not to produce language-specific effects, but the behaviourally defined version also remains ambiguous in terms of its cognitive specificity. Ullman and Pierpont (in press) have identified three currently competing classes of theory on the cognitive-level explanation of behaviourally defined SLI. One class posits deficits in language-specific structures involved in the rule-governed movements or combinations of words into complex structures. According to different versions, children may be impaired in establishing structural relationships such as agreement or specifier head-relations; or they may lack rules for linguistic features; or they may be stuck in a period of language development where marking of tense is taken to be optional; or they may be solely impaired on non-local dependency relations; or they may have problems with more general language functions such as learning implicit rules. The second class of theory views behavioural SLI as caused by a non-linguistic processing deficit that happens to particularly impact on language. Versions include claims for reduced processing rate or capacity limitations on cognitive processing; an information-processing deficit that particularly affects phonology; and a low-level perceptual or temporal processing deficit. The third class of theory is espoused by Ullman Thomas & K-S paper – page 21 and Pierpont themselves (in press), and argues that language exploits a more general duality in the cognitive system between declarative and procedural memory. Vocabulary is claimed to rely more on the declarative system and grammar more on the procedural system. SLI is then taken to be a developmental disorder of the procedural system, with the linguistic profile a result of the atypical development of the procedural system along with the attempts of the declarative system to compensate (see also van der Lely & Ullman, 2001). Ullman and Pierpont argue that such an account can explain behavioural deficits sometimes observed outside the domain of language, since in their view such deficits are all in skills that rely on the procedural memory system (Ullman & Pierpont, in press). Needless to say, each of the three classes of theory provides a different target for gene expression, if behavioural SLI is indeed to reveal the effects of genes on language structures during development. The ambiguity regarding the specificity of the deficits has to be clarified before behavioural SLI could serve as a “genetic double dissociation” of grammar and general cognition (Pinker, 1999), but current research leads us to doubt that it will ever fall neatly into this framework. The role of development The key difference between adult acquired aphasia and language deficits in developmental disorders is the process of development. For developmental disorders, a central feature of explanations of the behavioural profile will be the way that language structures are acquired over time and the internal and external constraints that shape this Thomas & K-S paper – page 22 process. It is the developmental process acting under atypical constraints that will account for the fine level of fractionation observed in disorders such as WS (KarmiloffSmith, 1998). Specifying the nature of the developmental process, however, is a significant challenge. We begin by reviewing the mechanisms that the brain offers for genes to specify particular cognitive structures. The brain level of description We firmly believe that the appropriate level of explanation for language deficits is cognitive. Nevertheless, we also believe in consistency between levels of description is vital, so that a cognitive level theory should not invoke developmental mechanisms that cannot be implemented in the available repertoire of mechanisms of brain development. The idea that uneven language profiles in genetic disorders can be explained by isolated, atypically developing functional brain systems does not fit well with what is currently known about how genes control brain development. Pennington (2001) summarises three broad classes of genetic control. These are effects: (1) on brain size, in terms of altering the number of neurons or synapses; (2) on neuronal migration, sometimes in a regionally specific fashion; and (3) on neurotransmission, either by changing levels of neurotransmitter or the binding properties of receptor proteins. In addition, the timing of gene expression contributes a crucial aspect of the emergent organisation of the functional structure (Elman, Bates, Johnson, Karmiloff-Smith, Parisi, Plunkett, 1996). According to current knowledge, genetic effects do not appear to operate in a regionspecific fashion over the areas of cerebral cortex that eventually underlie higher cognitive Thomas & K-S paper – page 23 processes (Kingsbury & Finlay, 2001). Regional specialisation is achieved by diffuse gradients of gene expression, along with activity-dependent processes (although the primary sensory cortices and the limbic system are to some extent exceptions to this characterisation; see Kingsbury & Finlay, 2001, for discussion). The final organisation of the cortex depends very much on the way in which the cortex has been activated from birth. There are, in short, no current candidate genes that could impair in an isolated way the development of a syntactic structure without other (perhaps more subtle) differences in brain development. In line with the idea that developmental disorders do not involve region-specific structural atypicalities in cortex (despite some apparently quite specific cognitive outcomes), post-mortem studies of genetic disorders, and subsequently a growing body of work in structural brain imaging, have revealed widespread anomalies in gross and fine anatomy of the brains of individuals with developmental disorders. Gross anatomical differences in both the relative and absolute size of large-scale structures can be found in disorders such as WS (Bellugi et al., 1999), DS (Nadel, 1999), and FraX (Reiss et al., 1995). A similar picture emerges in terms of brain structures at a finer scale. Kaufmann and Moser (2000) list a range of neocortical cytoarchitectonic and dendritic abnormalities such as laminar disturbance, increased neuronal packing density, reduced dendritic length, and spine dysgenesis that have been found across a range of disorders, including DS, FraX, WS, neurofibromatosis, Patau syndrome, tuberous sclerosis, phenylketonuria, Rett syndrome, and Rubinstein-Taybi syndrome. It is therefore hard to imagine that SLI, for example, will turn out to present a very different picture. Indeed, Bishop and Thomas & K-S paper – page 24 colleagues currently argue for a generalised neuro-immaturity in SLI early on, which may be undetectable in later development (see, also, similar discussion in Karmiloff-Smith, 1998). Genes clearly influence the computational repertoire of the initial 6-layer structure of cerebral cortex and the broad pattern of inputs and outputs. However, differentiated, specialised processing structures seem to be contingent on the patterns of activity induced in the cortex by interaction with the environment. Given the presence of widespread brain differences in many developmental disorders and the probability that (adult) modules emerge as a product of development (Johnson, 2001; Karmiloff-Smith, 1992), it is clear that explaining uneven language profiles in the adult phenotype of developmental disorders will be a complex endeavour. It appears likely that a final account of developmental deficits will need to begin by identifying differences in low-level neurocomputational properties, perhaps in numbers of neurons and their thresholds, local or global connectivity, and activity-dependent changes in these parameters (KarmiloffSmith, 1998; Oliver et al., 2000). The perturbations that these initial differences cause on the subsequent developmental trajectories of emerging systems must then be mapped out more precisely, taking into account atypical interactions, both internally between developing components and externally with the environment. In terms of specificity of cause and outcome, our understanding of the relationship between neurocomputational parameters and cognitive performance is at present very limited. For example, it might be that a computational property is anomalous throughout Thomas & K-S paper – page 25 the brain but only impacts on those cognitive domains that particularly rely upon it during development. Or it might be that its impact is crucial at one time in development when learning of a particular domain is predominant, and inconsequential if it occurs at other times (Karmiloff-Smith, 1998), since timing is of the essence in development. Or it might be that the cytoarchitectonic properties that specify regions of cortex are disrupted by diffuse gene expression gradients in such a way that computational anomalies are topologically restricted despite wider structural differences across the brain. The latter possibility might support a more restricted scope for the cognitive domains impacted during development. Such issues obviously remain to be worked out. From the brain level of description, then, we may conclude that little in the repertoire of developmental brain mechanisms as currently understood seems able to target specific high-level components of the adult language system (let alone restricted aspects of linguistic structure) while allowing others to develop normally. We now turn to consider the cognitive level, and identify two features that any explanation of the atypical developmental process should incorporate. The cognitive level of description At a pure cognitive level, what must a developmental theory look like that explains the uneven language profile found in some developmental disorders? We believe that it must emphasise at least two (linked) characteristics. These are interactivity and compensation. Thomas & K-S paper – page 26 Another is developmental timing which we alluded to above and is dealt with in more detail elsewhere (Elman et al., 1996). Several authors have argued that early language development is characterised by interactions between multiple sources of information (e.g., Bishop, 1997; Chiat, 2001; Karmiloff-Smith, 1997, 1998; McDonald, 1997). For example, Chiat (2001) maintains that language acquisition should be construed as a mapping task between sound and meaning, through which the words and sentence structures of a language are established. To achieve this mapping, multiple sets of information are exploited. When semantics is ambiguous, phonology can be used to bootstrap the extraction of meaning. When phonology is ambiguous, semantics can be used to bootstrap the extraction of wordsound information. Together, phonological and semantic information help bootstrap the acquisition of morpho-syntax. In a developmental disorder where there are indications of differential deficits across the components of the language system, any explanation of behavioural impairments must incorporate the altered pattern of interactions (and their timing) between the different information sources across development. Chiat (2001) carried out this very exercise for behavioural SLI. She came down in favour of an account that considers the language deficits as arising from impaired phonological processing and the consequent disruption of the interactions inherent in the mapping process. Such impairments may exist early in development, and yet fail to be measurable 1 We discuss the brain level and the cognitive level separately, under the view that these levels of description are mutually constraining. However, a causal theory needs to remain at a single level of description, in the sense that neural events do not cause cognitive events but are cognitive events. We Thomas & K-S paper – page 27 in the mature system (Karmiloff-Smith, 1998). In other words, the failure to find, say, a phonological deficit in an adult with SLI cannot be assumed to mean that a phonological deficit did not exist in infancy and impact on development early on. The enabling condition in early development is simply no longer evident in overt behaviour in later development. While the absence of phonological deficits in adults may comprise the falsifiability of the causal theory, the theory is eminently testable using longitudinal studies in children with SLI. The second characteristic that any theory of atypical development must incorporate is compensation. This is illustrated by a triangular comparison of adult aphasics, healthy children following early focal brain damage, and children with developmental disorders (see Karmiloff-Smith & Thomas, 2003; Thomas, 2003): (1) Following focal brain damage to their left hemispheres, adults can show persistent selective deficits in their language abilities. However, (2) following similar damage, healthy children usually demonstrate recovery from initial aphasic symptoms to later perform within the normal range (see Bates & Roe, 2001, for a review). Presumably, the greater effective brain plasticity of the child brain has permitted compensation and reorganisation of function. As a consequence, when we (3) compare adults who had focal lesions as children with adults who have developmental disorders of language, we find significant deficits only in the latter. Pointing to the presence of deficits in a disorder is somewhat tautological, but the third comparison does raise the question that if genetic developmental disorders of language are to be characterised by initial deficits to language-relevant structures, why discuss the issue of levels of description and causal models in developmental disorders elsewhere (see Karmiloff-Smith & Thomas, 2003; Mareschal et al., forthcoming). Thomas & K-S paper – page 28 has compensation-to-recovery not occurred as it does in the children with early focal lesions? The answer is that compensation in the developmental disorder probably has occurred, but the constraints of the system are insufficient to allow performance to develop to a level within the normal range (Mareschal et al., forthcoming; Thomas, 2003). This must be true for behaviourally defined disorders, because any child that had successfully compensated for their initial deficit would not be diagnosed as having a disorder. There are two implications of including interactivity and compensation into theories of atypical cognitive development. They become most stark in the context of those theories that seek to explain developmental language impairments in terms of the architecture of the normal language system, with selective components of the system being underor over-developed. The implications can be summarised by two questions: (1) If a deficit arises from initial damage to a selective component, why hasn’t this impairment been smeared across other components through the interactions that occur between components during development? (2) If a deficit arises from initial damage to a selective component, why haven’t other components in the system managed to compensate for this deficit and so attenuate the impairment across development? In many cases, answers to these two questions are hard to formulate because the precise nature of the developmental processes involved in normal language acquisition, let alone atypical language acquisition, remain ill specified. In the next two sections, we employ two worked examples to illustrate this point. Thomas & K-S paper – page 29 Specifying the developmental process: Example 1 In the face of the challenge of characterising the (atypical) developmental process, some theories respond by simply ignoring its contribution. Such theoretical approaches are often accompanied by an empirical approach that makes development ‘disappear’ by using ‘age-matched’ or ‘ability-matched’ controls. When the results of these studies are discussed, they focus on whether the atypical group differs from controls or not. It is only a small step to re-describe any significant differences in behavioural data as reflecting a process that is ‘intact’ in the control group and ‘impaired’ in the atypical group. The onus to construct a developmental account of the structures involved has vanished (see Karmiloff-Smith et al., in press, for discussion of the importance of building task-specific developmental trajectories in evaluating developmental deficits). It is instructive to take a static model of a developmental deficit and attempt to add a developmental process to this model. The following example comes from work on Williams syndrome. Clahsen and Almazan (1998) reported evidence from 4 children with WS in English past tense formation, whereby these children exhibited worse performance on inflecting irregular past tenses than regular past tenses. This effect has proved hard to replicate in larger samples of individuals with WS (see, Thomas et al., 2001). But for our purposes, it is the nature of the explanation that is of interest. Clahsen and Almazan (1998) proposed that children with WS had a specific deficit in their language system, in which they experienced problems in accessing the ‘sub-nodes’ of lexical entries but not in accessing the nodes themselves. This account falls within a theory of inflectional Thomas & K-S paper – page 30 morphology that proposes qualitatively separate mechanisms for producing regular and irregular forms (Pinker, 1999). The theory is still controversial (see Thomas & KarmiloffSmith, 2003, for discussion) but we will accept it as correct for the purposes of this example. Regular forms are inflected by a rule-mechanism (for English verbs, “add –ed to the verb stem”) whereas irregular inflections are stored as individual entries in the lexicon. In this theory, lexical representations have hierarchical structure, whereby the past tense form of an irregular verb is stored as a sub-entry of the lemma for the verb’s stem. For WS, Clahsen and Almazan suggest that normal access to nodes permits regular verb stems to be operated upon by the rule, but access to sub-nodes further down the hierarchy is restricted, impairing irregular inflection (see also Temple, Almazan, & Sherwood, 2002). The important point here is that the specification of the inflectional mechanism in this account is non-developmental. It is a specification of the normal adult system or, if one assumes a smaller lexicon, a static picture of the child language system. Let us try and turn this into a developmental account, by proposing a normal development process by which these structures could have been put in place. The developmental account requires at least three components. First, there is a mechanism for learning rules of inflection, able to spot the relationship between present and past tense forms (see Pinker’s ‘epiphany’ mechanism, 1999; and Marcus et al., 1992, for ideas on the information that such a rule learning mechanism might exploit). Second, there is a mechanism for storing lexical entries. This establishes ‘nodes’ for individual word forms, along with a specification of their meaning, so that DRINK might be registered as a verb with semantic features Thomas & K-S paper – page 31 corresponding to ‘the consumption of liquid’. Third, there is a mechanism for attaching sub-entries to these nodes, so that when the child hears the word form ‘drank’ in the context of ‘consuming liquid in the past’, DRANK is established as a sub-entry of DRINK, as a specification of its past tense form. Similarly, in time, DRUNK may be established as a sub-node specifying the past participle. With a normal developmental process in place (and our tongues tied as to the psychological plausibility of these learning mechanisms!), let us return to the claims regarding WS. Immediately there is an ambiguity, because Clahsen and Almazan (1998) characterise the WS behavioural impairment as stemming from a problem with accessing sub-nodes, as if those sub-nodes were already present. This implies no problem with learning sub-nodes, but one with accessing information that has already been learned. Presumably one could establish this distinction empirically by showing that retrieval is inconsistent rather than non-existent, although Clahsen and Almazan report no data of this nature. Alternatively, one could interpret the claim in terms of an impairment to the mechanism for learning sub-nodes, so that insufficiently robust representations are put in place following exposure to irregular past tense forms. Either version of the account would fit within the framework of a disorder in which some components of a normal system develop normally (the first and second mechanisms) whilst others develop atypically (the third, sub-node mechanism). 2 This, of course, holds only if one puts to one side the recognised fact that language development is delayed in WS. For the sake of argument, let us say that the first and second mechanisms are developing normally but slowly, although developmental delay cannot be simply negated as irrelevant (see discussion in Karmiloff-Smith, Scerif & Ansari, 2003) Thomas & K-S paper – page 32 Now, let us return to one of our previous questions: If a behavioural deficit arises from initial damage to a selective component, why haven’t other components in the system managed to compensate for this deficit and so attenuate the impairment across development? Applied to this example, if children with WS are struggling to learn ‘drank’ as a sub-node of the entry ‘drink’ (or access the information once learnt), why can’t they exploit either of their other normally functioning learning mechanisms to achieve normal-looking behaviour? Why can’t they exploit the first mechanism to learn ‘DRINK => DRANK’ as a mini past-tense rule? Why can’t they exploit the second mechanism to establish DRANK as a lexical node (rather than a faulty sub-node) that has the semantic specification ‘consuming liquid in the past’? To show a behavioural impairment with irregular inflection (if that is indeed what some individuals with WS show), such compensation cannot be available. Why not? We have attempted to build a possible developmental process for the acquisition of inflections within the words-and-rule theory, but clearly there must be additional constraints that we have missed preventing compensation. Our attempt was crucial because the Clahsen and Almazan proposal did not include any hypotheses about the developmental process. If there are constraints that our version of the developmental processes lacks with respect to compensation, it is surely contingent on the authors of the proposal to provide them. The proposal is at best incomplete, at worst implausible, without these additional developmental constraints. Thomas & K-S paper – page 33 Our conclusion from this example is that if questions of compensation (or lack thereof) are to be addressed, it is vital first to attempt a specification of the developmental process. In the above example, the initial plausibility of the developmental deficit rides on a non-developmental characterisation of the relevant language structures, but it is far from obvious that it would remain plausible if upgraded into a developmental explanation. Either way, one can only explore these issues if one attempts to specify a developmental process. Specifying the developmental process: Example 2 Clahsen and Almazan are not unique in under-specifying developmental mechanisms. This is a challenge that has long faced developmental psychologists in general. One possible response to the under-specification has been to implement computational models of the developmental process, in order to evaluate the impact of using different types of learning algorithm, different types of initially constrained representations, or different training environments on the subsequent success of acquiring cognitive abilities (see Thomas & Karmiloff-Smith, 2002c, for a review). Building computational models necessarily involves simplification and restriction often to single domains such as inflectional morphology, vocabulary acquisition, or parsing. But it does provide a level of computational precision lacking in most other forms of theory. One of the computational architectures most applied to developmental problems has been that of connectionist networks. These have been used in league with a variety of theoretical commitments, from their use as a reasonably theory-neutral tool to explore the Thomas & K-S paper – page 34 information available in certain learning environments, to servicing the much stronger claim that associative mechanisms are sufficient to explain language acquisition. In this example, we appeal to the role these models can play in expanding the set of candidate inferences that one can draw about underlying cognitive structure based on certain patterns of surface behaviour (Thomas & Karmiloff-Smith, 2002b). Recently, connectionist models have been increasingly applied to developmental disorders. For example, in our own work, we have explored the implications of damaging a learning system in its initial state (analogous to a developmental disorder) compared to damaging a system in its trained state (analogous to an adult acquired deficit) (Thomas & Karmiloff-Smith, 2002a). The results demonstrated that some types of damage hurt the system much more in the ‘adult’ state (e.g., severing network connections) while others hurt the system much more in the ‘infant’ state (e.g., adding noise to processing). The adult system can tolerate noise because it already has an accurate representation of the knowledge, but loss of network structure leads to a decrement in performance since connections contain established knowledge. By contrast, the infant system can tolerate loss of connections because it can reorganise remaining resources to acquire the knowledge, but is impaired by noisy processing since this blurs the knowledge to be acquired. Empirical evidence supports the importance of a good representation of the input during language acquisition. When McDonald (1997) analysed the conditions for successful and unsuccessful language acquisition across a range of typical and atypical populations (including late L2 learners, individuals with Down syndrome, Williams Thomas & K-S paper – page 35 syndrome and SLI), the results indicated that good representations of speech sounds were key in predicting the successful acquisition of a language, including its syntax. In other work, we have applied connectionist models to a much more detailed, datadriven consideration of one domain and one developmental disorder, the acquisition of past tense formation in Williams syndrome (Thomas & Karmiloff-Smith, 2003). The latter model provides a framework in which to evaluate a recent proposal regarding the cause of language deficits, once more in the domain of English past tense formation but this time in Specific Language Impairment (Ullman & Pierpont, in press). Importantly, Ullman and Pierpont’s proposal includes a process of compensation in explaining the final behavioural impairment. As we described earlier, Ullman and Pierpont (in press) have put forward a theory of SLI contingent on the differential involvement of two memory systems in normal language acquisition. In Ullman and Pierpont’s theory, there is a distinction between procedural memory (for fast, sequential, automatic processing) and declarative memory (for slower, parallel, conscious processing). The acquisition of grammar is taken to rely more on the former and the acquisition of the lexicon more on the latter. SLI, with its primary behavioural impairments in grammar, is then construed as a developmental disorder of the procedural system. Importantly, Ullman and Pierpont’s account explains the SLI profile as including the attempts of the declarative system to compensate for the developmental shortcomings of the procedural system. Van der Lely and Ullman’s (2001) English past tense data are illustrative, here. Children with SLI show low levels of Thomas & K-S paper – page 36 inflection for both regular and irregular verbs (10-20% correct), and similarly low levels of extension of the regular rule to novel stems. Since regulars are normally inflected more accurately than irregulars, this amounts to a greater deficit for regular verbs – viewed as a kind of fractionation. Van der Lely and Ullman’s explanation of this pattern of behaviour again relies on a linguistic theory that distinguishes separate mechanisms for acquiring regular and irregular verbs. Regulars are learned by a rule-implementing mechanism (part of the procedural system), irregulars by an associative memory (part of the declarative memory). According to van der Lely and Ullman, the children with SLI are unable to learn the regular rule with their procedural system, and the few regulars and irregulars that are correctly inflected reflect the compensatory action of the declarative system. The idea that regulars are now inflected by an associative memory instead of a rule mechanism is supported by the presence of increased frequency effects in regular inflection in the SLI group compared to typically developing children – frequency effects are taken to be the hallmark of domain-general associative memory. It is important to be clear about the chain of inference in this case, because it clearly illustrates how researchers move from behavioural evidence to deducing structural fractionations of the language system. The relatively greater impairment of regular inflections, along with the increased frequency effects in residual regular inflection are taken as evidence that in SLI, there has been a startstate deficit to a domain-specific computational structure responsible for learning regular past tense forms. Thomas & K-S paper – page 37 Ullman and Pierpont are to be lauded for their attempts to be more specific about the developmental process in explaining the behavioural data in a developmental disorder and, in particular, for including compensation in their account. Their theory may turn out to be the correct one. However, there is a difficulty. Computational models of atypical development have indicated that intuiting how compensation ‘will probably work’ can be a very hit-and-miss affair. This turns out to be the case in past tense acquisition when we look at an implemented model. The computational model of past tense acquisition that we recently explored combines lexical-semantic information about a verb with phonological information about the verb’s stem to generate its past tense form (Thomas & Karmiloff-Smith, 2003; see Lavric et al. 2001 for discussion of this architecture). As an outcome of the developmental process, the network comes to rely differentially on the two sources of information for driving each type of inflection. In particular, it relies more heavily on lexical-semantic information for driving irregular inflections, so that in the trained model, a lesion to lexical-semantics differentially impaired irregulars (see also Joanisse & Seidenberg, 1999). The model employs a ‘three-layer’ architecture, where a layer of internal processing units intercedes between input and output layers. This layer is a common representational resource involved in processing regular, irregular, and novel inflections. Recently, we demonstrated that manipulating the discriminability of the activation function of the processing units in the internal layer and the output layer of the initial, untrained network led to a system that exhibited a developmental disorder (Thomas, in Thomas & K-S paper – page 38 press b). This startstate manipulation roughly had the effect of making computations fuzzier. It reduced the ability of the system to make sharp categorisations, so that it required much more training to produce very different outputs patterns for similar input patterns. When the disordered network was ‘aged-matched’ to a normally developing past tense network, it exhibited low levels of both regular and irregular inflection along with poor regularisation of novel stems. In other words, the disordered network gave an approximate fit to the SLI data presented by van der Lely and Ullman (2001). Moreover, in the model just as in the empirical data, regular verbs now exhibited an elevated frequency effect. Subsequent analysis of the network revealed that regular inflection was being driven more strongly by lexical-semantic input than in the normal network. In effect, the system was treating regulars in the same way as irregulars, as if all verbs were exceptions to be generated via support from the lexicon. On the face of it, this model would appear to parallel van der Lely and Ullman’s explanation of their SLI data: residual regular inflection reflects the action of the declarative memory system storing word-specific information. Regulars and irregulars are treated in the same way in the disordered network, with equivalent reliance of lexicalsemantics and equivalent sized frequency effects. Crucially, however, the startstate manipulation to the connectionist network was not to a domain-specific processing structure affecting only regulars, as assumed by Ullman and Pierpont and van der Lely. Instead, the manipulation targeted a general processing resource used to inflect both regular and irregular verbs. Because the particular computational property altered was one upon which regular verbs differentially relied, the result was a deflection of the Thomas & K-S paper – page 39 developmental trajectory such that in terms of the relative size of deficits, there was an apparent fractionation between regular and irregular verbs. In effect, the startstate computational manipulation altered a property that was domainrelevant to regular inflection rather than domain-specific (Karmiloff-Smith, 1998). Specifically, an essential characteristic of the regular rule is to treat all items within a category in the same way. For this, the system requires the ability to form sharp category boundaries. Reduced discriminability of the processing units caused delayed learning to all verbs but particularly impaired the network in forming the sharp categories necessary to learn and generalise regular inflections. These initial alterations to the common computational resource had the effect of altering the balance of the information sources on which the network relied to generate past tense forms. Phonological regularities were downplayed, while word-specific information was emphasised. The atypical constraints of the learning system served to alter the interaction between phonological and semantic sources of knowledge during development of this morpho-syntactic ability. These computational simulations do not demonstrate that Ullman and Pierpont’s procedural-declarative theory of SLI is wrong. What they demonstrate is that the inferences made by these authors are not the only ones legitimised by their behavioural data. Inferences drawn from developmental behavioural deficits to affected underlying structures are entirely contingent on a precise specification of developmental process. Thomas & K-S paper – page 40 Crucially for the general argument, in this example details of the developmental process fully determine whether a behavioural dissociation should be taken as evidence for an initial deficit to domain-specific processing structure (and therefore, a fractionation of the language system); or should be taken as an initial deficit to a general processing resource that results in a seeming domain-specific outcome (and no structural fractionation). Without specification of the developmental process, we do not know whether there are domain-specific effects to be explained by gene expression or modality-specific effects or domain-general effects. Links between genes and language can never be answered without considering the details of the process of development. The importance of the problem domain Up till now we have focused on differences between normal and developmentally disordered systems. Lastly, we return to consider the possible implications of behavioural similarities between the patterns of language development exhibited by different typical and atypical populations. Recall, one explanation is that these similarities reflect internal constraints of the language development system. On theoretical grounds, however, similarities between typical and atypical development may have another explanation. It is possible that the range of behaviours that individuals can exhibit in language development is constrained to some extent by the physical and social environment in which the individual’s cognitive system is embedded. That is, behaviours normal or otherwise are in part constrained by the structure of the problem domain to which the cognitive system is exposed, whatever its underlying architecture. It is a serious and Thomas & K-S paper – page 41 unresolved issue of the extent to which cognitive architecture is visible in the behavioural changes and error patterns exhibited across development. The simplest illustration of this would be a cognitive domain that had an easy part and a hard part. A wide range of learning systems would naturally acquire the easy part before the hard part. Consequently, a developmental fractionation here would tell us little about the actual learning system involved. Computer simulations can again provide a means to probe this question further. When we exposed a variety of associative architectures to the past tense domain, there was great variation across the developmental profiles (Mareschal et al., forthcoming). Nevertheless, the systems also exhibited similarities in their profiles: regular acquisition was usually in advance of irregular acquisition; generalisation of the regular rule was usually weaker to novel stems that rhymed with irregulars than to those that did not. These patterns were a result of the common problem domain to which the systems were exposed (see Thomas & Redington, 2004, for a similar exercise in modelling the atypical processing of syntax). The developmental commonalities across ‘disordered’ networks would not, in this instance, be strongly supportive of claims that language acquisition is primarily constrained by the ‘brain’ that is acquiring the language, without consideration of the properties of the problem domain. Yet, as we saw in the introduction, commonalities in the developmental trajectories across different development disorders have been used to draw just such a conclusion (see Fowler, 1998; Tager-Flusberg & Sullivan, 1997). Thomas & K-S paper – page 42 Conclusion So, can developmental disorders reveal the component parts of the human language faculty? In a sense, we are left with a puzzle to which we can only begin to sketch a solution. There does appear to be fractionation between the different aspects of the language system in developmental disorders. In Williams syndrome, such fractionation occurs at a very fine-grained level. In Specific Language Impairment, there were initially claims for a neat fractionation, but increasingly it appears that deficits are differential, simply with lesser impairments in non-linguistic domains. How, then, are we to explain the level of fractionation that we encounter within the development of disordered language systems? We have argued that the observed fractionations are the consequences of the processes of ontogenetic development acting on a neonatal brain that has been constructed with (perhaps subtly) altered initial neurocomputational biases. Elsewhere, we have referred to this theoretical framework as ‘neuroconstructivism’ (Karmiloff-Smith, 1998; KarmiloffSmith & Thomas, 2003). On the one hand, it is unlikely that genetic effects during brain development in neurogenetic disorders are uniform across the entire brain. On the other hand, they are also unlikely to be highly region specific. Differential effects are probably graded rather than targeting certain circuits, particularly with regard to higher cortical functions identified in adults (Kingsbury & Finlay, 2001). We still await further data here. For example, Reiss et al. (2004) recently reported that for a sample of adults with WS, there were decreases compared to controls in grey matter Thomas & K-S paper – page 43 volumes and densities in sub-cortical and cortical regions comprising the human visualspatial system, which Reiss et al. argued were associated with visuospatial impairments in the disorder. They also reported increases compared to controls in grey matter volumes and densities in several areas including the amygdala, orbital and medial prefrontal cortices, anterior cingulate, insular cortex, and superior temporal gyrus, areas involved in emotion and face processing. Reiss et al. argued these could be associated with ‘enhanced emotionality and face processing’. However, emotion and face processing are rarely at chronological age level in WS, and have been argued to be atypical in the disorder. Moreover, the level of specificity of the brain differences is still some way short of the level of (frequently within-domain) fractionation we saw in the behavioural evidence. In short, in developmental disorders, it is likely that the granularity of genetic differences in cortex is at a coarser level than that of cognitive modules or parts thereof. The final highly differentiated cognitive profile in disorders such as WS is due to the result of complex processes of development, attenuating or exaggerating (perhaps initially subtle) neurocomputational differences. At the brain level, the usual emergence of an interactive network of neural systems may be perturbed by several factors: by the differing effect of the atypical computational biases on the ability of various areas to process the signal with which they are provided by the initial large scale input-output connectivity of the brain; by anomalies in the emergence of specialised circuits through pruning or competition; by subtle differences in timing due to maturational delay; by compensatory changes during interactions between different brain regions; and by the atypical subjective environment to which the individual with the disorder is exposed (see Thomas & K-S paper – page 44 Mareschal et al., forthcoming, for discussion). At the cognitive level, for the domain of language the interaction between separate sources of information, phonological, semantic, morphological, syntactic, and pragmatic may be altered following initial problems in one or more domains, leading to an uneven profile. However, patterns of strengths and weaknesses must also be viewed through the lens of what is hard and what is easy for all learners in language acquisition. For developmental disorders, then, the outcome at the cognitive level is a granularity of subsequent behavioural fractionations likely to be considerably finer than cognitive-level modules. In our own neuroconstructivist framework, we have a view of what we expect developmental processes to look like. However, our main argument here is that explanations of developmental deficits depend on having a developmental account of some kind, and in many cases even this minimal requirement is absent. Particular methodological approaches emphasise development – the use of longitudinal studies, the construction of task-specific developmental trajectories, the tracing of childhood and adult deficits back to their precursors in infancy, the use developmental computational models to simulate behavioural data. But a developmental perspective can also be applied to data gained from more traditional methodologies. Apparent fractionations of the language system in developmental disorders can tell us about the constraints that shape the development of the language, and even how genes may influence those constraints. Merely stating that disorders have a genetic component Thomas & K-S paper – page 45 tells us nothing about how genes are expressed. So, this story is really only just beginningand, in our view, the developmental processes itself will eventually lie at its heart. ReferencesAlcock, K.J. (1995). Motor aphasia A comparative study. Unpublished doctoral thesis,University of Oxford.Ansari, D., Donlan, C., Thomas, M. S. C., Ewing, S., Peen, T., & Karmiloff-Smith, A.(2003). What makes counting count? Verbal and visuo-spatial contributions totypical and atypical number development. Journal of Experimental ChildPsychology, 85, 50-62.Bates, E. (1997). Origins of language disorders: A comparative approach. In D. Thal & J.Reilly, (Eds.), Special issue on Origins of Communication Disorders, DevelopmentalNeuropsychology, 13(3), 447-476.Bates, E., & Roe, K. (2001). Language development in children with unilateral braininjury. In C. A. Nelson & M. Luciana (Eds.), Handbook of Developmental CognitiveNeuroscience (p. 281-307). Cambridge, Mass: MIT Press.Bellugi, U., Mills, D., Jernigan, T., Hickok, G., & Galaburda, A. (1999). Linkingcognition, brain structure, and brain function in Williams syndrome. In H. Tager-Flusberg (Ed.), Neurodevelopmental disorders (pp. 111-136). Cambridge, Mass.:MIT Press.Bellugi, U., Lichtenberger, L., Jones, W., Lai, Z., & St. George, M. (2000). Theneurocognitive profile of Williams syndrome: A complex pattern of strengths andweaknesses. Journal of Cognitive Neuroscience 12(1), 1-29. Thomas & K-S paper – page 46 Bishop, D.V.M. (1983). The Test for Reception of Grammar. Age and CognitivePerformance Research Centre, University of ManchesterBishop, D. (1992). The underlying nature of specific language impairment. Journal ofChild Psychology and Psychiatry, 33 (1), 3-66.Bishop, D.V.M. (1997). Uncommon understanding: development and disorders oflanguage comprehension in children. Hove: Psychology Press.Chiat, S. (2001). Mapping theories of developmental language impairment: Premises,predictions and evidence. Language and Cognitive Processes, 16, 113-142.Churchill, J. D., Grossman, A. W., Irwin, S. A., Galvez, R., Klintsova, A. Y., Weiler, I.J., & Greenough, W. T. (2002). A Converging-Methods Approach to Fragile XSyndrome. Developmental Psychobiology, 40, 323-328.Clahsen, H. & Almazan, M. (1998). Syntax and morphology in Williams Syndrome.Cognition, 68, 167-198.Cornish, K.M., Munir, F., & Cross, G. (1999). Spatial cognition in males with Fragile-Xsyndrome: evidence for a neuropsychological phenotype. Cortex, 35, 263-271.Cornish, K.M., Munir, F., & Cross, G. (2001). Differential impact of the FMR-1 fullmutation on memory and attention functioning: a neuropsychological perspective.Journal of Cognitive Neuroscience, 13, 144-151.Donnai, D., & Karmiloff-Smith, A. (2000) Williams Syndrome: from genotype throughto the cognitive phenotype. American Journal of Medical Genetics, 97, 164-71.Elman, J. L., Bates, E. A., Johnson, M. H., Karmiloff-Smith, A., Parisi, D., & Plunkett,K. (1996). Rethinking innateness: A connectionist perspective on development.Cambridge, Mass: MIT Press. Thomas & K-S paper – page 47 Fowler, A. (1998). Language in mental retardation: Associations with and dissociationsfrom general cognition. In J. A. Burack, R. M. Hodapp, and E. Zigler, Handbook ofMental Retardation and Development (p. 290-333). Cambridge: CambridgeUniversity Press.Freund, L., & Reiss, A. L. (1991). Cognitive profiles associated with the fragile Xsyndrome in males and females. American Journal of Medical Genetics, 38, 542-547.Grant, J., Karmiloff-Smith, A., Gathercole, S., Paterson, S., Howlin, P., Davies, M., &Udwin, O. (1997) Verbal Short-term Memory and its relation to LanguageAcquisition in Williams Syndrome. Cognitive Neuropsychiatry, 2, 2, 81-99.Grant, J., Valian, V., & Karmiloff-Smith, A. (2002). A study of relative clauses inWilliams syndrome. Journal of Child Language, 29, 403-416.Greenough, W.T., Klintsova, A. Y., Irwin, S. A., Galvez, R., Bates, K. E., & Weiler, I. J.(2001). Synaptic regulation of protein synthesis and the fragile X protein.Proceedings of the National Academy of Sciences, 98, 7101-7106.Joanisse, M. F., & Seidenberg, M. S. (1999). Impairments in verb morphology followingbrain injury: A connectionist model. Proceedings of the National Academy ofScience USA, 96, 7592-7597.Johnson, M. H. (2001). Functional brain development in humans. Nature ReviewsNeuroscience, 2, 475-483.Karmiloff-Smith, A. (1992). Beyond Modularity: A Developmental Perspective onCognitive Science. Cambridge, Mass.: MIT Press/Bradford Books. Thomas & K-S paper – page 48 Karmiloff-Smith, A. (1997). Crucial differences between developmental cognitiveneuroscience and adult neuropsychology. Developmental Neuropsychology, 13(4),513-524.Karmiloff-Smith, A. (1998). Development itself is the key to understandingdevelopmental disorders. Trends in Cognitive Sciences, 2(10), 389-398.Karmiloff-Smith, A., Grant, J., Berthoud, I., Davies, M., Howlin, P. & Udwin, O. (1997)Language and Williams Syndrome: How Intact is "Intact"? Child Development, 68,246-262.Karmiloff-Smith, A. Scerif, G. & Ansari, D. (2003) Double dissociations indevelopmental disorders? Theoretically misconceived, empirically dubious.Cortex, 39, 161-163.Karmiloff-Smith, A. & Thomas, M. S. C. (2003). What can developmental disorders tellus about the neurocomputational constraints that shape development? The case ofWilliams syndrome. Development and Psychopathology, 15, 969-990.Karmiloff-Smith, A., Thomas, M. S. C., Annaz, D., Humphreys, K., Ewing, S., Grice, S.,Brace, N., Van Duuren, M., Pike, G., & Campbell, R. (in press). Exploring theWilliams Syndrome face processing debate: The importance of buildingdevelopmental trajectories. Journal of Child Psychology and Psychiatry and AlliedDisciplines.Karmiloff-Smith, A., Tyler, L. K., Voice, K., Sims, K., Udwin, O., Howlins, P., andDavies, M. (1998). Linguistic dissociations in Williams syndrome: evaluatingreceptive syntax in on-line and off-line tasks. Neuropsychologia, 36, 343-351. Thomas & K-S paper – page 49 Kaufmann, W. E., & Moser, H. W. (2000). Dendritic anomalies in disorders associatedwith mental retardation. Cerebral Cortex, 10, 981-991.Kingsbury, M. A. & Finlay, B. L. (2001). The cortex in multidimensional space: wheredo cortical areas come from? Developmental Science, 4, 125-156.Laing, E., Hulme, C., Karmiloff-Smith, A. & Grant, J. (2001). Learning to read inWilliams syndrome: Looking beneath the surface of atypical development. Journalof Child Psychology and Psychiatry. 42 (6), 729-739.Laing, E., Butterworth, G., Ansari, D., Gsödl, M., Laing, E., Barnham, Z., Lakusta, L.,Tyler, L.K., Grice, S., Paterson, S., & Karmiloff-Smith, A. (2002). Atypicallinguistic and socio-communicative development in toddlers with Williamssyndrome. Developmental Science, 5 (2), 233-246.Lavric, A., Pizzagalli, D., Forstmeier, S., & Rippon, G., (2001). Mapping dissociations inverb morphology. Trends in Cognitive Sciences, 5, 301-308.Leonard, L. B. (1998). Children with Specific Language Impairment. Cambridge, MA:MIT Press.Marcus, G., Pinker, S., Ullman, M., Hollander, J., Rosen, T. & Xu, F. (1992).Overregularisation in language acquisition. Monographs of the Society for Researchin Child Development, 57 (Serial No. 228).Mareschal, D., Johnson, M., Sirios, S., Spratling, M., Thomas, M. S. C., & Westermann,G. (forthcoming). Neuroconstructivism: How the brain constructs cognition. Oxford:Oxford University Press.McDonald, J. L. (1997). Language acquisition: The acquisition of linguistic structure innormal and special populations. Annual Review of Psychology, 48:215-241. Thomas & K-S paper – page 50 Mervis, C. B., & Bertrand, J. (1997). Developmental relations between cognition andlanguage: Evidence from Williams syndrome. In L. B. Adamson & M. A. Romski(Eds.), Research on communication and language disorders: Contributions totheories of language development (pp. 75–106). New York: Brookes.Mervis, C. B., Morris, C. A., Bertrand, J., and Robinson, B. F. (1999). WilliamSyndrome: Findings from an integrated program of research. In H. Tager-Flusberg(ed.), Neurodevelopmental disorders. MIT Press: Cambridge, Mass.Mills, D.L., Alvarez, T.D., St.George, M., Appelbaum, L.G, Bellugi, U. & Neville, H.(2000). Electrophysiological studies of face processing in Williams syndrome.Journal of Cognitive Neuroscience, 12, 47-64.Nadel, L. (1999). Down syndrome in cognitive neuroscience perspective. In H. Tager-Flusberg (Ed.), Neurodevelopmental disorders (pp. 197-222). Cambridge, Mass.:MIT Press.Nazzi, T. & Karmiloff-Smith, A. (2002). Early categorization abilities in young childrenwith Williams syndrome. NeuroReport, 13, 1259-1262.Nazzi, T., Paterson, S., & Karmiloff-Smith, A. (2002). Early word segmentation byinfants and toddlers with Williams syndrome. Infancy, 4, 251-271.Oliver, A., Johnson, M. H., Karmiloff-Smith, A., & Pennington, B. (2000). Deviations inthe emergence of representations: A neuroconstructivist framework for analysingdevelopmental disorders. Developmental Science, 3, 1-23.Pääbo, S. (1999). Human Evolution. Trends in Genetics, 15(12): M13-M16.Paterson, S.J., Brown, J. H. , Gsödl, M. K. , Johnson, M. H. & Karmiloff-Smith, A.(1999). Cognitive Modularity and Genetic Disorders. Science, 286, 2355-2358. Thomas & K-S paper – page 51 Pennington, B. F. (2001). Genetic methods. In C. A. Nelson & M. Luciana (Eds.),Handbook of Developmental Cognitive Neuroscience (pp. 149-158). Cambridge,Mass: MIT Press.Pinker, S. (1999). Words and rules. London: Weidenfeld & Nicolson.Pinker, S. (2001). Talk of genetics and vice-versa. Nature, 413, 465-466.Reiss, A., L., Abrams, M. T., Greenlaw, R., Freund, L., & Denckla, M. B. (1995).Neurodevelopmental effects of the FMR-1 full mutation in human. Nature Medicine,1, 159-167.Reiss, A. L., Eckert, M. A., Rose, F. E., Karchemskiy, A., Kesler, S., Chang, M.,Reynolds, M. F., Kwon, H., & Galaburda, A. (2004). An experiment of nature: Brainanatomy parallels cognition and behaviour in Williams syndrome. Journal ofNeuroscience, 24(21), 5009-5015.Scerif, G. (2003). Infant and toddler precursors of attentional difficulties in fragile Xsyndrome: A neurodevelopmental perspective. Unpublished Ph.D. Dissertation,University of London.Scerif, G., Cornish, K., Wilding, J., Driver, J., & Karmiloff-Smith, A. (2004). Visualsearch in typically developing toddlers and toddlers with fragile X and Williamssyndrome. Developmental Science.Semel, E., & Rosner, S. R. (2003). Understanding Williams syndrome: Behavioralpatterns and interventions. Lawrence Erlbaum Associates, Publishers: Mahwah, NewJersey.Singer Harris, N. G., Bellugi, U., Bates, E., Jones, W., & Rossen, M. (1997). Contrasting Thomas & K-S paper – page 52 profiles of language development in children with Williams and Down syndromes.Developmental Neuropsychology, 13, 345-370.Tager-Flusberg, H., Plesa-Skwerer, D., Faja, S., & Joseph, R., M. (2003). People withWilliams syndrome process faces holistically. Cognition, 89, 11–24.Tager-Flusberg, H. & Sullivan, K. (1997). Early language development in children withmental retardation. In Ê. J. Burack, R. Hodapp, & E. Zigler (Eds.), Handbook ofdevelopment and retardation (pp. 208-239). New York: Cambridge University Press.Temple, C., Almazan, M., & Sherwood, S. (2002). Lexical skills in Williams syndrome: acognitive neuropsychological analysis. Journal of Neurolinguistics, 15(6), 463-495.Thomas, M. S. C. (2003). Limits on plasticity. Journal of Cognition and Development,4(1), 95-121.Thomas, M. S. C. (in press a). Williams syndrome: Fractionations all the way down?:Commentary on Semel and Rosner. Cortex.Thomas, M. S. C. (in press b). Characterising compensation. Cortex.Thomas, M. S. C., Dockrell, J. E., Messer, D., Parmigiani, C., Ansari, D., & Karmiloff-Smith, A. (2003). Naming in Williams syndrome. Manuscript under revision.Thomas, M. S. C., Grant, J., Gsödl, M., Laing, E., Barham, Z., Lakusta, L., Tyler, L. K.,Grice, S., Paterson, S. & Karmiloff-Smith, A. (2001). Past tense formation inWilliams syndrome. Language and Cognitive Processes, 16, 143-176.Thomas, M. S. C. & Karmiloff-Smith, A. (2002a). Are developmental disorders likecases of adult brain damage? Implications from connectionist modelling.Behavioural and Brain Sciences, 25(6), 727-780. Thomas & K-S paper – page 53 Thomas, M. S. C. & Karmiloff-Smith, A. (2002b). Residual Normality: Friend or Foe?Behavioural and Brain Sciences, 25(6), 772-780.Thomas, M. S. C. & Karmiloff-Smith, A. (2002c). Modelling typical and atypicalcognitive development. In U. Goswami (Ed.), Handbook of Childhood Development(pp. 575-599). Blackwells Publishers.Thomas, M. S. C. & Karmiloff-Smith, A. (2003). Modelling language acquisition inatypical phenotypes. Psychological Review, 110(4), 647-682.Thomas, M. S. C. & Redington, M. (2004). Modelling atypical syntax processing. Toappear in: Proceedings of the COLING-2004 Workshop: Psycho-computationalmodel of human language acquisition. Geneva, Switzerland, 28 August 2004.Tomblin, J. B., & Pandich, J. (1999). What can we learn from children with exceptionallanguage development? Lessons from children with specific language impairment.Trends in Cognitive Science, 3,283-285.Ullman, M. T. & Pierpont, E. I. (in press). Specific Language Impairment is not specificto language: The Procedural Deficit Hypothesis. Cortex.van der Lely, H. K. J. (1997). Language and cognitive development in a grammatical SLIboy: Modularity and innateness. Journal of Neurolinguistics, 10, 75-107.van der Lely, H. K. J. (1999). Learning from Grammatical SLI: a response to J. Tomblinand J.B. Pandich. Trends in Cognitive Sciences, 3, 8, 286-288.van der Lely, H. K. J. & Ullman, M. T. (2001). Past tense morphology in speciallylanguage impaired and normally developing children. Language and CognitiveProcesses, 16, 177-217. Thomas & K-S paper – page 54 Vargha-Khadem, F., Watkins, K. E., Price, C. J., Ashburner, J., Alcock, K. J., Connelly,A., Frackowiak, R. S. J., Friston, K. J., Pembrey, M. E., Mishkin, M., Gadian, D. G.,& Passingham, R. E. (1998). Neural basis of an inherited speech and languagedisorder. Proceedings of the National Academy of Sciences of the USA (95), 12695–12700.Vicari, S., Brizzolara, D., Carlesimo, G., Pezzini, G., & Volterra, V. (1996). Memoryabilities in children with Williams syndrome. Cortex, 32, 503-514.Volterra, V., Capirci, O., Pezzini, G., Sabbadini, L., & Vicari, S. (1996). Linguisticabilities in Italian children with Williams syndrome. Cortex, 32, 663-677.Watkins, K. E., Dronkers, N. F., & Vargha-Khadem, F. (2002). Behavioural analysis ofan inherited speech and language disorder: comparison with acquired aphasia. Brain,125(3), 452-464.Watkins, K. E., Vargha-Khadem, F., Ashburner, J., Passingham, R. E., Connelly, A.,Friston, K. J., Frackowiak, R. S. J., Mishkin, M., & Gadian, D. G. (2002). MRIanalysis of an inherited speech and language disorder: structural brain abnormalities.Brain, 125(3), 465-478.Zukowski, A. (2001). Uncovering grammatical competence in children with Williamssyndrome. Unpublished doctoral thesis. Boston University. Thomas & K-S paper – page 55 Acknowledgements:This work was supported by Medical Research Council Career Establishment GrantG0300188 to MT and by MRC Programme Grant No. xxxxx and PPHealthcare GrantNo. xxxxx to AK-S. (will complete numbers at proof)