Combining Acoustic Confidence Scores with Deep Semantic Analysis for Clarification Dialogues

This paper describes a technique to include acoustic confidence scores as returned by automated speech recognisers in generic semantic representations. The method we propose requires only minimal changes to an existing grammar used for speech applications. Special attention is paid to the treatment of multi-word lexemes and combining several (N-best) speech recognition results into one semantic representation. The approach has been implemented and tested using the Nuance speech recognition software and a chart parser, in the formalism of underspecified discourse representations. The potential relevance of confidence scores in rich semantic representations is illustrated by generating more flexible clarification questions in dialogue systems.