Considering multiple options when interpreting spoken utterances

摘要

We describe Scusi?, a spoken language interpretation mechanism designed to be part of a robot-mounted dialogue system. Scusi?’s interpretation process maps spoken utterances to text, which in turn is parsed and then converted to conceptual graphs. In order to support robust and flexible performance of the dialogue module, Scusi? maintains multiple options at each stage of the interpretation process, and uses maximum posterior probability to rank the (partial) interpretations produced at each stage. The time and space requirements of maintaining multiple options are handled by means of an anytime search algorithm. Our evaluation focuses on the impact of the speech recognizer and the search algorithm on Scusi?’s performance.