Learning Lexical Alignment Policies for Generating Referring Expressions for Spoken Dialogue Systems

We address the problem that different users have different lexical knowledge about problem domains, so that automated dialogue systems need to adapt their generation choices online to the users’ domain knowledge as it encounters them. We approach this problem using policy learning in Markov Decision Processes (MDP). In contrast to related work we propose a new statistical user model which incorporates the lexical knowledge of different users. We evaluate this user model by showing that it allows us to learn dialogue policies that automatically adapt their choice of referring expressions online to different users, and that these policies are significantly better than adaptive hand-coded policies for this problem. The learned policies are consistently between 2 and 8 turns shorter than a range of different hand-coded but adaptive baseline lexical alignment policies.