Contextual coherence in natural language processing

Abstract

Controlled and restricted dialogue systems are reliable enough to be deployed in various real world applications. The more conversational a dialogue system becomes, the more difficult and unreliable become recognition and processing. Numerous research projects are struggling to overcome the problems arising with more- or truly conversational dialogue system. We introduce a set of contextual coherence measurements that can improve the reliability of spoken dialogue systems, by including contextual knowledge at various stages in the natural language processing pipeline. We show that, situational knowledge can be successfully employed to resolve pragmatic ambiguities and that it can be coupled with ontological knowledge to resolve semantic ambiguities and to choose among competing automatic speech recognition hypotheses.