Refining text-based dialogue state tracker for spoken dialogue systems

Abstract

In recent years, there has been remarkable progress in dialogue systems, as evidenced by various competitions and datasets. Despite this progress, building robust task-oriented dialogue systems that incorporate speech interfaces remains a significant challenge. A major problem is that the direct use of spoken dialogue as input for text-based dialogue systems can lead to a significant performance degradation. This is mainly due to differences in phrasing, grammar, and issues such as disfluencies and background noise. However, as evidenced by voice assistant systems such as Siri and Alexa, it is of practical importance to transfer this success to spoken dialogues. In this paper, we present our model, which demonstrated strong performance in the Speech-Aware Dialogue Systems technology challenge track at Dialog System Technology Challenge (DSTC). Our model consists of three major modules: (1) Automatic Speech Recognition (ASR) error correction, which bridges the gap between spoken and text utterances. (2) A text-based dialogue system (D3ST) that estimates slots and values using slot descriptions. (3) Named entity post-processing to correct any errors in the estimated slot values. Additionally, we conducted an ablation study to compare the individual performance of each module. Our experiments show the importance of incorporating an explicit ASR error correction module, post-processing for named entity, and data augmentation to effectively adapt a text-based dialogue state tracker to spoken dialogue corpora.