Paper by Bob Frank in Language Acquisition

Bob Frank, together with co-authors Don Mathis and Bill Badecker, has just published a paper in the journal Language Acquisition. The paper, entitled The Acquisition of Anaphora by Simple Recurrent Networks, explores the degree to which statistical learning of the sort embodied in a certain class of connectionist networks can extract the abstract regularities that govern the interpretation of pronouns and reflexives.  They demonstrate that although these networks induce structural representations that are sufficient for the task, they do so in a way that attends to probabilistically viable, but grammatically irrelevant, distinctions. As a result, the networks exhibit certain systematic failures in generalizing their knowledge. The paper can be found here.