An Exemplar-based Approach to Unsupervised Parsing
Simone Dennis · eScholarship (California Digital Library) · 2005
We present an approach to syntactic processing based on the Syntagmatic Paradignmatic model (Dennis, in press) that assumes that the parse of a sentence can be viewed as a set of alignments with exemplars from memory.Alignment is achieved using a span-based version of the normalized edit distance measure (Marzal & Vidal, 1993), which is more appropriate for linguistic tasks.Span similarities used in the algorithm are derived using a version of the topics model (Griffiths & Steyvers, 2002) in which part-of-speech sequences are generated from their preceeding and postceeding word context.Approximate nearest neighbour exemplars are chosen using Locality Sensitive Hashing (Indyk & Motwani, 1998;Gionis, Indyk, & Motwani, 1999).Parses generated by the model are compared against gold standard parses from the Penn Treebank.The method provides state of the art precision and recall on this task and suggests that an unsupervised approach to parsing is feasible.Furthermore, the model is more directly comparable to exemplar-based accounts in other areas of cognition such memory and categorization than recursion-based approaches to syntax.