Exploiting Machine-Transcribed Dialog Corpus to Improve Multiple Dialog States Tracking Methods

摘要

This paper proposes the use of unsuper-vised approaches to improve components of partition-based belief tracking systems. The proposed method adopts a dynamic Bayesian network to learn the user action model directly from a machine-transcribed dialog corpus. It also addresses confidence score calibration to improve the observation model in a unsuper-vised manner using dialog-level grounding information. To verify the effectiveness of the proposed method, we applied it to the Let's Go domain (Raux et al., 2005). Overall system performance for several comparative models were measured. The results show that the proposed method can learn an effective user action model without human intervention. In addition, the calibrated confidence score was verified by demonstrating the positive influence on the user action model learning process and on overall system performance.