Procedures for Learning Syntactic Categories: A Model and Test with Artificial Grammars
Dorrit O. Billman · Deep Blue (University of Michigan) · 1983
A model of syntax acquisition must include powerful learning procedures if it is to account for the complexity of the systems learned and the circumstances of acquisition. One way of increasing the power of a learning system is to capitalize on information as it is learned to improve future learning. The research reported here investigates one method of using new knowledge in learning syntactic categories. The Internal Feedback Model claims that the basis for learning syntactic categories is learning the intercorrelated set of properties shared by members of the same category. To make learning more efficient for a highly structured domain such as language, the Internal Feedback Model proposes a focused sampling plan for encoding input regularities. This plan changes the allocation of attention to focus on features proving important. As learners discover that a particular feature is important in one rule, they will sample it more heavily, facilitating the discovery of interrelated rules. Two experiments tested the prediction of facilitation for a rule in the context of an interrelated rule set versus in isolation. The experiments used artificial grammars learned by college students. Experiments varied the correlational structure available for inducing lexical classes analogous to declension or classifier systems in natural languages. Subjects watched interacting, novel shapes on a CRT, described by sentences of the invented language. No external feedback was provided. As predicted, learning a given rule among a cluster of interrelated rules was easier than learning the identical rule presented in isolation. Additional exploratory work with more complex grammars using phrase structures is reported, and broader, more speculative interpretations of learning in this task and its relation to natural acquisition are included.