A connectionist model of sentence comprehension and production

Abstract

Linguists have historically favored symbolic, rule-based models to explain the human language faculty. Such models typically possess clear explanatory clarity and are well able to handle the apparently recursive structure of natural sentences. In fact, they tend to be too powerful in this respect and ad hoc external constraints and manipulations are often necessary to limit their performance to better match humans' imperfect abilities. Furthermore, due to their reliance on simple rules and structured knowledge, symbolic methods do not lend themselves to learning. A focus on symbolic models and an idealized notion of language processing has contributed to the widespread belief that language is primarily innately specified. But over the past twenty years there has been increasing interest in connectionist models of language processing. Initially, these were relatively straight-forward implementations of symbolic methods using simple processing units. However, as the representatio...