Nonlinear dynamics in the resolution of lexical ambiguity: A parallel distributed processing account.

Abstract A connectionist account of lexical ambiguity resolution is presented in which the nonlinear dynamics that arise during processing is emphasized. In the implementation, the spelling, pronunciation, part of speech, and meaning of both ambiguous and unambiguous words are represented as a distributed pattern of activity over a set of simple processing units in a fully recurrent network. After the network is trained on this lexicon using an error-correction algorithm, the performance of the network is assessed by presenting just part of a lexical entry (e.g., the spelling). The number of processing cycles to activate all the units representing the spelling, the pronunciation, or the meaning to their minimal or maximal activation are used as indices of lexical decision times, naming times, or reading times. respectively. Consistent with empirical results, the simulations demonstrate the effect of frequency and context on the processing of unambiguous words, as well as the effect of relative dominance and context on the time course of the activation of meanings of ambiguous words. Advantages of a distributed representation are also discussed.