Over the past three decades, phonological theory has advanced in many areas, but it has changed little in its foundational assumptions about how computational processes can serve as a basis for the theory. This volume suggests that it may be worthwhile to reconsider some of those assumptions. Is there an order to the rules in a phonological derivation? What kinds of links other than derivations are possible between the level of mental representation and the level of speech sounds? Since phonological representations are so much more sophisticated today than they were a few decads ago, do we need any phonological rules at all? In this provocative book, leading linguists and computer scientists consider the challenges that computational innovations pose to current rule-based phonological theories and speculate about the advantages of phonological models based on artificial neural networks and other computer designs.
The authors offer new conceptions of phonological theory for the 1990s, the most radical of which proposes that phonological processes cannot be characterized by rules at all, but arise from the dynamics of a system of phonological representations in a high-dimensional vector space of the sort that a neural network embodies. This new view of phonology is becoming increasingly attractive to linguists and others in the cognitive sciences because it answers some difficult questions about learning while drawing on recent results in philosophy, psychology, artificial intelligence, and neuroscience. The contributors are John A. Goldsmith, Larry M. Hyman, George Lakoff, K. P. Mohanan, David S. Touretzky, and Deirdre W. Wheeler.