Abstract

In this paper we present a proposal to develop conversational agents that avoids the effort of manually defining the dialog strategy for the agent and also takes into account the benefits of using current standards. In our proposal the dialog manager is trained by means of a POMDP-based methodology using a labeled dialog corpus automatically acquired using a user modeling technique. The statistical dialog model automatically selects the next system response. Thus, system developers only need to define a set of files, each including a system prompt and the associated grammar to recognize user responses. We have applied this technique to develop a conversational agent in VoiceXML that provides information for planning a trip.