Knowledge and Reasoning in Practical Dialogue Systems

We describe Scusi?, a spoken language interpretationmechanism designed to be part of a robot-mounted di-alogue system. Scusi?’s interpretation process mapsspoken utterances to text, which in turn is parsed andthen converted to conceptual graphs. In order to sup-port robust and flexible performance of the dialoguemodule, Scusi? maintains multiple options at eachstage of the interpretation process, and uses maximumposterior probability to rank the (partial) interpreta-tions produced at each stage. The time and space re-quirements of maintaining multiple options are han-dled by means of an anytime search algorithm. Ourevaluation focuses on the impact of the speech rec-ognizer and the search algorithm on Scusi?’s perfor-mance.