We describe a connectionist model of the past tense that generates both regular and irregular past tense forms with good generalization. The model also exhibits frequency effects that have been taken as evidence for a past tense rule (Pinker, 1991) and consistency effects that are not predicted by rule-based accounts. Although not a complete account of the past tense, this work suggests that connectionist models may Capture generalizations about linguistic phenomena that rule-based accounts miss.

This research investigated the effects of prior knowledge on learning in psychologically-plausible connectionist networks. This issue was examined with respect to the benchmark orthography-to-phonology mapping task (Sejnowski & Rosenberg, 1986; Seidenberg & McClelland, 1989). Learning about the correspondences between orthography and phonology is a critical step in learning to read. Children (unlike the networks mentioned above) bring to this task extensive knowledge about the sound-structure of their language. We first describe a simple neural network that acquired some of this phonological knowledge. W e then summarize simulations showing that having this knowledge in place facilitates the acquisition of orthographicphonological correspondences, producing a higher level of asymptotic performance with fewer implausible errors and better nonword generalization. The results suggest that connectionist networks may provide closer approximations to human performance if they incorporate more realistic assumptions about relevant sorts of background knowledge.

The bases of developmental dyslexia were explored using connectionist models. The behavioral literature suggests that there are two dyslexic subtypes: "phonological" dyslexia involves impairments in phonological knowledge whereas in "surface " dyslexia phonological knowledge is apparently intact and the deficit may instead reflect a more general developmental delay. We examined possible computational bases for these impairments within connectionist models of the mapping from spelling to sound. Phonological dyslexia was simulated by reducing the capacity of the models to represent this type of information. The surface pattern was simulated by reducing the number of hidden units. Performance of the models captured the major behavioral phenomena that distinguish the two subtypes. Phonological impairment has a greater impact on generalization (reading nonwords such as NUST); the hidden unit limitation has a greater impact on learning exception words such as PINT. More severe impairments produce mixed cases in which both nonwords and exceptions are impaired. Thus, the simulations capture the effects of different types and degrees of impairment within a major component of the reading system.

Learning to read English requires learning the complex statis-tical dependencies between orthography and phonology. Pre-vious research has focused on how these statistics are learnedin neural network models provided with as much training asneeded. Children, however, are expected to acquire this knowl-edge in a few years of school with only limited instruction. Weexamined how these mappings can be learned efficiently, de-fined by tradeoffs between the number of words that are explic-itly trained and the number that are correct by generalization.A million models were trained, varying the sizes of randomly-selected training sets. For a target corpus of about 3000 words,training sets of 200–300 words were most efficient, producinggeneralization to as many as 1800 untrained words. Composi-tion of the 300 word training sets also greatly affected general-ization. The results suggest directions for designing curriculathat promote efficient learning of complex material.

Behavioral experiments have demonstrated that people encode knowledge of correlations among semantic prop- erties of entities and that this knowledge influences per- formance on semantic tasks (McRae, 1992; McRae, de Sa, & Seidenberg, 1993). Independently, in connectionist the- ory, it has been claimed that relationships among seman- tic properties may provide structure that is required for the relatively arbitrary mapping from word form to word meaning (Hinton&Shallice, 1991). W e explored these is- sues by implementing a modified Hopfield network (1982, 1984) to simulate the computation from word form to meaning. The model was used as a vehicle for developing explanations for the role played by correlated properties in determining short interval semantic priming effects and in determining the ease with which a property is verified as part of a concept. Simulations of the priming and property verification experiments of McRae (1992) are reported. It is concluded that correlations among properties encoded in conceptual memory play a key role in the dynamics of the computation of word meaning. Furthermore, a model in which property intercorrelations are central to forming basins of attraction corresponding to concepts may pro- vide important insights into lexical memory.

MacDonald, Pearlmutter, and Seidenberg (1993) propose an alternative to the dominant view in sentence processing that syntactic ambiguities are resolved by heuristics based on structural simplicity. MacDonald et al. argue that such ambiguities can be defined in terms of alternatives associated with information in individual lexical items, and thus that syntactic ambiguities can be resolved by lexical disambiguation mechanisms relying on access to the relative frequencies of alternatives and to biases created by contextual constraints. We present evidence from a computer simulation of the use of frequency-based and contextual constraints in the processing of the main verb/reduced relative syntactic ambiguity, showing that frequency and relatively limited contextual information from a sample of natural language can interact sufficiently to model basic results in the literature.

We routinely encounter speakers with different accents andspeaking styles. The speech perception literature offersexamples of disruption of comprehension for unfamiliarspeech and also of listeners’ rapid accommodation tounfamiliar accents. Much of this research uses a singlemeasure and/or focuses on isolated word perception. Weinvestigated listeners’ abilities to comprehend and shadowconnected speech spoken in a familiar or unfamiliar accent.We found increases in shadowing latencies andcomprehension errors in the Dissimilar Speech relative toSimilar Speech conditions—especially for relatively informalrather than more academic style speech. Additionally, therewas less accommodation over time to Dissimilar than SimilarSpeech. These results suggest that there are costs both in theimmediate timescale of processing speech (necessary forshadowing) and in the longer time scale of listeningcomprehension when accent and other speech quality is verydifferent from one’s own speech.

Despite anecdotal evidence of relative visuospatial processing strengths in individuals with reading disability (RD), only a few studies have assessed the presence or the extent of these putative strengths. The current study examined the cognitive and neural bases of visuospatial processing abilities in adolescents with RD relative to typically developing (TD) peers. Using both cognitive tasks and functional magnetic resonance imaging (fMRI) we contrasted printed word recognition with non-language visuospatial processing tasks. Behaviorally, lower reading skill was related to a visuospatial processing advantage (shorter latencies and equivalent accuracy) on a geometric figure processing task, similar to findings shown in two published studies. FMRI analyses revealed key group by task interactions in patterns of cortical and subcortical activation, particularly in frontostriatal networks, and in the distributions of right and left hemisphere activation on the two tasks. The results are discussed in terms of a possible neural tradeoff in visuospatial processing in RD.