A probabilistic model for understanding composite spoken descriptions

Abstract

We describe a probabilistic reference disambiguation mechanism developed for a spoken dialogue system mounted on an autonomous robotic agent. Our mechanism receives as input referring expressions containing intrinsic features of individual concepts (lexical item, size and colour) and features involving more than one concept (ownership and location). It then performs probabilistic comparisons between the given features and features of objects in the domain, yielding a ranked list of candidate referents. Our evaluation shows high reference resolution accuracy across a range of spoken referring expressions. © 2008 Springer Berlin Heidelberg.