The University of Arizona

Events & News

Cog Sci Brown Bag Seminar

DateFriday, February 12, 2010
Time12:00 pm
LocationGS 906
SpeakerCanceled-Paul Smolensky, Ph.D.
TitleKrieger-Eisenhower Professor of Cognitive Science
AffiliationDepartment of Cognitive Science - Johns Hopkins University

Relating Linguistic Competence and Psycholinguistic Performance in Phonological Encoding through Neural Network Computation

ABSTRACT: While activation-based models capture important aspects of psycholinguistic performance, the standard simple implementations through local connectionist (“neural”) networks are unable to incorporate other important aspects of performance which are related to linguistic competence. Linguistic theory argues that competence requires representations with rich discrete structure – such as hierarchically related constituents – and discrete knowledge about the well-formedness of these structures – grammar. Such discrete structure and knowledge is lacking in the representations and processing of simple local connectionist networks. This talk will present early work on phonological encoding that proposes:
• theoretical principles which integrate activation-based approaches to linguistic performance with a recent approach to linguistic competence, Optimality Theory (Prince & Smolensky 1993/2004), and
• a connectionist implementation of these principles.

The network realization combines general formal techniques for:
• realizing discrete symbolic structures in distributed activation patterns within the network;
• realizing Optimality – Theoretic grammatical constraints on these structures in the connections of the network: the grammatical well-formedness of a linguistic structure is realized as the connectionist well-formedness ('Harmony') of the corresponding activation pattern;
• optimizing over these constraints to generate grammatically well-formed representations by spreading activation through the network connections; and
• selecting a coherent response that corresponds to a single discrete structure (not a mixture of structures).

The principles of this theory and applications to extremely simple cases will allow us to examine the theory’s ability to explain – and ground in processing mechanisms – some very general relations between grammatical well-formedness and performance.

Partially supported by a Chaire Internationale de Recherche Blaise Pascal de l’Etat et de la Région d’Ile-de-France. Co-authored with Matthew Goldrick and Donald Mathis.