Dialogue Management based on Multi-domain Corpus

Abstract

Dialogue Management (DM) is a key issue in Spoken Dialogue System. Most of the existing data-driven DM schemes train the dialogue policy for some specific domain (or vertical domain), only using the dialogue corpus in this domain, which might suffer from the scarcity of dialogue corpus in some domains. In this paper, we divide Dialogue Act (DA), as semantic representation of utterance, into DA type and slot parameter, where the former one is domain-independent and the latter one is domain-specific. Firstly, based on multiple-domain dialogue corpus, the DA type prediction model is trained via Recurrent Neutral Networks (RNN). Moreover, DA type decision problem is modeled as a multi-order POMDP, and transformed to be a one-order MDP with continuous states, which is solved by Natural Actor Critic (NAC) algorithm and applicable for every domain. Furthermore, a slot parameter selection scheme is designed to generate a complete machine DA according to the features of specific domain, which yields the Multi-domain Corpus based Dialogue Management (MCDM) scheme. Finally, extensive experimental results illustrate the performance improvement of the MCDM scheme, compared with the existing schemes.