Using Harmonic Grammar in performance theory: Explaining phonological production errors

Paul Smolensky (Johns Hopkins Cognitive Science)
Matt Goldrick (Northwestern Linguistics)
Don Mathis (Johns Hopkins Cognitive Science)

Abstract

General properties of phonological error patterns in tongue-twister production tasks can be modeled and explained by embedding phonological Harmonic Grammars in the connections of a connectionist network with distributed representations, stochastic Harmony optimization, and a distributed winner-take-all dynamics that yields outputs that are discrete phonological representations. The existence of gradient 'traces' of the correct segment in the phonetic realization of incorrect segments' can also be accounted for.