Laura G. Miller
Allen L. Gorin
AT&T Bell Labs., 600 Mountain Ave., Rm. 2D-463, Murray Hill, NJ 07974
In this research on adaptive language acquisition, connectionist networks are explored that map input messages into machine actions. In this work, the capability of such networks to learn associations between messages and meaningful responses to them as the set of machine actions increases in size and complexity is investigated. Specifically, how to reflect task structure in a network architecture is considered in order to provide improved generalization capability. A method for constructing networks from component subnetworks, namely a product network, provides improved generalization by factoring the associations between words and action through an intermediate layer of semantic primitives. A two-dimensional product network was evaluated in a 1000-action data retrieval system, the object of which is to answer questions about 20 attributes of the 50 states of the US. The system was tested by 13 subjects over a 2-week period, during which over 1000 natural language dialogues were recorded. The experiment was conducted using typed input with unconstrained vocabulary and syntax. During the course of performing its task, the system acquired over 500 words and retained 92% of what it learned. The results of this experiment demonstrate the scalability of previous results obtained on smaller, less complex tasks [Gorin et al., Comput. Speech Lang. 5, 101--132 (1991)] [Miller and Gorin, Proc. ICASSP 201--204 (1992)].