Can a perceptual processing deficit explain the impairment of inflectional morphology in developmental dysphasia? A computational investigation.

Developmental dysphasia (also referred to as specific language impairment) is a developmental language disorder in which children display delayed or abnormal language development but have normal non-verbal intelligence and no gross perceptual or neurological disorders. Several investigators have hypothesized that the underlying cause of developmental dysphasia is a subtle perceptual processing deficit. But there have also been criticisms of the perceptual processing account (Gopnik & Crago, 1991) and several alternative theories have been put forth (Gopnik & Crago, 1991; Clahsen, 1989). Furthermore, it has recently been argued that the behavior of developmental dysphasics provides strong support for dual mechanism accounts of morphology acquisition and processing (Pinker & Prince, 1991; Gopnik & Crago, 1991). This paper contributes to the debate by studying the effects of a perceptual processing deficit on a model that simulates the acquisition of morphology. A neural network model that is capable of learning a system of semantics to phonology mappings analogous to part of the system of English inflectional morphology is presented. When the phonological input to the model selectively distorted in a manner consistent with phonological processing deficits of developmental dysphasics, the model's performance simulates dysphasic children s performance. When compared to the undamaged model, the dysphasic" model displays slower overall learning, a higher rate of errors on suffixation, differential success on the third person singular, progressive -ing and oed suffixes, and greater difficulty with regular than irregular past tenses. The model also addresses arguments that have been put forth against the perceptual processing theories, including the assertion that the dysphasics ' greater diffir.ulty with morphemic than, with non-morphemic phonemes undermines the validity of a perceptual processing account. Gopnik & Crago (1991) argue that a perceptual processing account must predict that dysphasics will have equal difficulty producing identical phonemes regardless of morphological structure. For example, they should have equal difficulty with the -s in nose and the -s in i..ees. The model demonstrates that this claim is false and that differential performaJ, ';~ on identical phonemes can arise from the effects of competition between the members of an inflectional paradigm. The model demonstrates that many of the symptoms of dysphasic speech can be produced by distorting the phonological input to a unitary learning system. The model's performance provides support for perceptual processing accounts of developmental dysphasia and, like Marchman (1993), calls into question the claim that dysphasics' behavior can be used to support dual mechanism accounts of morphology acquisition. Introduction: Developmental dysphasia, or specific language impairment (SLI), is a developmental language disorder in which children display delayed or abnormal language development but have normal non-verbal intelligence and no gross perceptual, behavioral, or neurological deficits. Dysphasics form a relatively heterogeneous population and there may in fact be a variety of related disorders all grouped under the rubric of SLI. Nonetheless, many children with SLI have a similar linguistic prof1le: " ... a mild to moderate deficit in a range of language areas and a more serious deficit in the use of morphology." (Leonard et ai, 1992, p. 1077) There have been a variety of theories put forth to explain SLI. One group of theories posits a perceptual processing deficit as the ultimate cause of the children s linguistic difficulties. This hypothesis is based on a large body of work investigating the perceptual processing abilities of children with SLI. series of investigations carried out by Paula Tallal and her colleagues (Talial & Piercy, 1973a, 1973b, 1974, 1975; Tallal & Stark, 1981) have shown that dysphasic children suffer from a subtle perceptual processing deficit. This deficit affects their ability to perceive and discriminate rapidly changing sounds. For example, children with SLI are significantly worse than normals at discriminating stop consonant-vowel pairs such as baor fricative-vowel pairs such as sa-sha. A perceptual deficit of this kind could plausibly lead to the kinds of linguistic problems evidenced in children with SLI. Cross-linguistic studies of English, Italian and Hebrew speaking dysphasics conducted by Leonard and his colleagues (Leonard et ai, 1988; Leonard, 1989, Leonard, 1993) have added further support for the perceptual processing hypothesis. A comparison of English and Italian dysphasics is particularly instructive. Most of English grammatical morphology is represented by relatively low salience markers (e. , word final non-syllabic consonants and unstressed syllables). Consequently, English dysphasics have problems with much of their language s grammatical morphology. But most Italian morphological markers are more salient than the English equivalents. In keeping with the perceptual processing hypothesis, Italian children perform much better than English dysphasics (and at levels similar to normal Italian children) on most morphological markers. But the Italian dysphasics are impaired on some morphological markers; those which are low in phonetic substance. A particularly interesting contrast is the fact that the Italian dysphasics perform much better with the vowel-final feminine articles fa and una then they do with the consonant-final masculine equivalents if and un. Leonard has hypothesized that the SLI children s difficulties with low salience items may be the consequence of a reduced processing capacity. Recently, Gopnik & Crago (1991) have proposed that the underlying cause of SLI is neither a perceptual processing deficit nor a more general cognitive problem, but a specifically linguistic deficit. Gopnik & Crago adopt Pinker (Pinker & Prince, 1991; Marcus et ai, 1992), dual mechanism theory of morphology acquisition and processing. According to this theory, the computational demands required for the acquisition and production of regularly and irregularly inflected forms are so different that two separate and qualitatively distinct systems have evolved to handle regular and irregular forms. The regular forms are produced by a rule-based system which functions by appending oed any verb stem that is not marked as an irregular. The second system is an associative memory system. Through frequent exposure it can learn the irregulars on a case by case basis. Since these systems are thought to be distinct neurally as well as functionally, we can expect that they may dissociate under certain conditions. Gopnik & Crago claim that SLI is a result of a genetic deficit which affects the part of the language system responsible for regular inflectional morphology. Because of the defect in the regular rule-based system, children with SLI cannot learn to form morphological paradigms and manipulate grammatical features the way that normals do. Therefore, these children must rely solely on the associative memory system, Since the memory system is performing a task it was not designed for, learning of regular morphology is dramatically impoverished, In this paper, we contribute to the debate by comparing the behavior of two neural networks as they attempt to learn a system of semantic to phonology mappings. The two networks are identical in all respects. but one: the phonological input to the "SLI network" is degraded in a manner consistent with the hypothesized processing deficit of dysphasic children. Models: The model discussed in this paper is a neural network model that learns to relate semantic and phonological representations. The model was developed to simulate morphology acquisition and processing in normals (Hoeffner, 1992). The architecture of the model is shown in Figure 1. The portion of the figure included in the rectangular box is the implemented model. Architecture: There are three layers, a semantic layer, a hidden layer and a phonological layer. There are connections between layers, and, for the hidden and phonological layers, there are also bi-directional weights connecting the units within a layer. The bi-directional weights allow the network to perform an iterative computation, passing information back and forth through the network as it gradually settles to a stable state. This type of network is known as an attractor network. There are several reasons for using this type of network: its ability to learn arbitrary mappings, the flexibility regarding input/output relations, the ability to vary input strength (soft-clamping), and the use of settling time as an analog of reaction time (see Hoeffner, 1992, for further explanation). We envision the current model as being embedded in a larger set of systems (see Figure 1). We assume that a perceptual system is responsible for feeding phonological representations into the morphology learning network. If this system is impaired, then systematically distorted representations will be fed into the phonological layer of the morphology learning system. Corpus: The training corpus is based on data from the Marcus et al (1992) monograph. It consists of all the monosyllabic regular past tense verbs produced by Adam , Eve, Sarah, Abe, or the adults . conversing with them (Table All in Marcus et al) and all the irregular past tense verbs produced by the four children (Tables AS-A8). There are a total of 1925 verb forms in the corpus, arranged in 385 paradigms. Each paradigm has five members: a zero marked form (jump), a past tense (jumped), third person singular (jumps), progressive (jumping), and a past participle (jumped). The frequencies of each verb are taken from Kucera and Francis (1967), Phonological representation: The phonology is represented by a feature/slot scheme. There are 8 slots: 3 prevocalic, one vocalic, 3 post vocalic and one syllabic suffix slot. Each co