Ambiguity representation and resolution in spoken dialogue systems

Spoken natural language often contains ambiguities that must be addressed by a spoken dialogue system. In this work, we present the internal semantic representation and resolution strategy of a dialogue system designed to understand ambiguous input. These mechanisms are domain independent; task-specific knowledge is represented in parameterizable data structures. Speech input is processed through the speech recognizer, parser, interpreter, context tracker, pragmatic analyzer and pragmatic scorer. The context tracker combines dialogue context and parser output to yield raw attribute-value (AV) pairs from which candidate values are derived. The pragmatic analyzer adjusts the confidence associated with each AV candidate based on system intent, e.g., implicit confirmation and user input. Pragmatic confidence scores are introduced to measure the dialogue managers confidence for each AV: MYCIN-like scoring is used to merge multiple information sources. Pragmatic analysis and scoring is combined with explicit error correction capabilities to achieve efficient ambiguity resolution. The proposed strategies greatly improve dialogue interaction, eliminating about half of the errors in dialogues from a travel reservation task.