Grammatical Structure and Distributed Representations
Jeffrey L. Elman · 1992
Abstract Neural networks have a number of properties which make them attractive computational systems for modelling cognitive behaviour (McClelland, Rumelhart, and Hinton 1986; Rumelhart and McClelland 1986). One of the useful characteristics is their sensitivity to contextual factors, and another is their ability to seek a solution which satisfies multiple constraints. For example, work by Kawamoto and McClelland (1986) and by St. John and McClelland (in press) has shown how grammatical cues can be combined with context information to make inferences about events. Elman (1990) demonstrated that sequential context can provide the basis for adducing the category structure of internal representations of lexical items. Mikkulainen and Dyer (in press) showed how a similar effect could be achieved for external representations (i.e., forms which could then be used by other processing modules).