Adapting Text-based Dialogue State Tracker for Spoken Dialogues

Although there have been remarkable advances in dialogue systems through thedialogue systems technology competition (DSTC), it remains one of the keychallenges to building a robust task-oriented dialogue system with a speechinterface. Most of the progress has been made for text-based dialogue systemssince there are abundant datasets with written corpora while those with spokendialogues are very scarce. However, as can be seen from voice assistant systemssuch as Siri and Alexa, it is of practical importance to transfer the successto spoken dialogues. In this paper, we describe our engineering effort inbuilding a highly successful model that participated in the speech-awaredialogue systems technology challenge track in DSTC11. Our model consists ofthree major modules: (1) automatic speech recognition error correction tobridge the gap between the spoken and the text utterances, (2) text-baseddialogue system (D3ST) for estimating the slots and values using slotdescriptions, and (3) post-processing for recovering the error of the estimatedslot value. Our experiments show that it is important to use an explicitautomatic speech recognition error correction module, post-processing, and dataaugmentation to adapt a text-based dialogue state tracker for spoken dialoguecorpora.