Modeling user behavior online for disambiguating user input in a spoken dialogue system

A spoken dialogue system (SDS) interacts with its user in a spoken natural language. It interprets user speech input and responds to the user. User speech in a spoken natural language may be ambiguous. A challenge in building an SDS is dealing with ambiguity. Without good abilities for disambiguation, an SDS can hardly have meaningful and smooth dialogues with its user in practical applications. The existing techniques for disambiguation are mainly based on statistical knowledge about language use. In practical situations, such knowledge alone is inadequate. In our research, we develop a new disambiguation technique, which is based on application of knowledge about user activity behavior, in addition to knowledge about language use. The technique is named MUBOD, standing for modeling user behavior online for disambiguation. The core component of MUBOD is an online reinforcement learning algorithm that is used to learn the knowledge and apply the knowledge for disambiguation. In this paper, we describe the technique and its implementation, and present and analyze some initial experimental results.

Graphical abstractFigure optionsDownload full-size imageDownload as PowerPoint slideHighlights► We apply knowledge about user behavior to disambiguate user speech input. ► The knowledge is represented in a user behavior model. ► We use reinforcement learning to acquire, update, and apply the knowledge. ► The knowledge can be automatically updated online. ► Experimental results show that disambiguation has been significantly improved.