Generative models of disfluency

Abstract:

This thesis describes a generative model for representing disfluent phenomena in human speech. This model makes use of observed syntactic structure present in disfluent speech, and uses a right-corner transform on syntax trees to model this structure in a very natural way. Specifically, the phenomenon of speech repair is modeled by explicitly modeling the partially completed constituent structure, building up phrasal structure as it is read in, and allowing incomplete constituents as part of the final structure of an utterance.