Learning to Balance Grounding Rationales for Dialogue Systems

Joshua Gordon1,  Rebecca J. Passonneau1,  Susan L. Epstein2
1Columbia University, 2Hunter College and The Graduate Center of the City University of New York

Abstract

This paper reports on an experiment that investigates clarification subdialogues in intentionally noisy speech recognition. The architecture learns weights for mixtures of grounding strategies from examples provided by a human wizard embedded in the system. Results indicate that the architecture learns to eliminate misunderstandings reliably despite high word error rate.