Relational dialogue

Abstract

Conversational agents are increasingly common in everyday life. Dialogue with these agents is often limited to the task at hand, and is not focused on conversation as a shared social experience. Previous work has demonstrated that strengthening the user-agent relationship increases the agent's efficacy, and leads to a more enjoyable user experience. I present a relationship-driven dialogue system that aims to strengthen and expand the relationship between the agent and user. The system uses a knowledge graph to represent relevant information about the world and the agent's and user's preferences. When choosing a response, a novel probabilistic approach, called MRF-Chat, models the mutual knowledge of the agent and the user, as well as the contextual relevance of concepts in candidate responses. In human evaluations, the system was considered significantly more collaborative, engaging, and trusted by human partners in a semi-structured interaction on food preferences.