Abstract for Robert Frank's talk

Syntax Acquisition/Computational
Robert Frank (Johns Hopkins University)
April 4, 2006 (Tuesday)
6:30 PM - 8:00 PM; Room 7102, The CUNY Graduate Center

The past decade has witnessed a resurgence of attention to the role of statistical induction in human language learning. One line of such work argues that proper attention to statistical patterns overcomes traditional "poverty of the stimulus" arguments, and thereby obviates the need for innate grammatical knowledge. In this talk, I will survey the current state of this debate, considering what kinds of information has and has not been shown to be extractable by statistical techniques. I will focus attention on the error-driven training of artificial neural networks, specifically Elman's Simple Recurrent Networks (SRNs), as this constitutes a particularly flexible and powerful technique for statistical induction, and is the one that has achieved what appear to be the most impressive results to date. I will then report on a line of experiments (conducted in collaboration with Don Mathis and Bill Badecker) that explore the capacity of SRNs to extract generalizations concerning the referential dependencies in reflexive and pronominal anaphora. This task requires a more refined sensitivity to grammatical structure than those studied previously. We find that the statistical nature of the training yields the SRN to a solution that is successful when measured quantitatively, but which diverges in certain key respects from the target generalization. We also find variation in the network's ability to extend its grammatical generalizations to novel structures depending on the conditions of the training, but no ability at all to generalize to novel lexical items.