Techniques for utterance disambiguation in a human-computer dialogue system

Abstract

Disambiguating an utterance occurring in a dialogue context is a complex task, which requires input from many different sources of information -- some syntactic, some semantic, and some pragmatic.

The central question addressed by this thesis is how to integrate data sources for utterance disambiguation within a bilingual human-computer dialogue system. First, a simple scheme is proposed for classifying disambiguation data sources; then this scheme is used to develop a method for combining data sources in a principled manner. Next, several actual sources of disambiguation data are explored; each is fitted into the previously described implementation framework. In particular, a probabilistic grammar is developed and augmented using novel techniques to increase its performance with respect to the local dialogue context.

In a dialogue system, ambiguities which cannot be resolved automatically can be clarified by asking the user what was meant. This thesis also presents a model of clarification subdialogues which is integrated within the utterance disambiguation framework. This is followed by a brief treatment of how user errors may be accommodated, and how this process can also be fitted -- conceptually and in implementation -- into the previously described disambiguation framework.

Finally, I describe the details of implementing these techniques within an existing dialogue system, and give examples demonstrating their effectiveness.