User simulations for online adaptation and knowledge-alignment in troubleshooting dialogue systems
Srinivasan Janarthanam, Oliver Lemon · 2008
We study the problem of alignment be-tween dialogue participants, using the prac-tical example of “troubleshooting ” dialogue systems. Recent work on troubleshooting concerns automated spoken dialogue sys-tems which support users who need to re-pair their internet connection. We address the problem that different users have differ-ent types of knowledge of problem do-mains, so that automated dialogue systems need to adapt online to the different know-ledge of these users as it encounters them. We approach this problem using policy learning in a Markov Decision Process (MDP). In contrast to related work we pro-pose a new user model which incorporates the different conceptual knowledge of dif-ferent users, together with an environment simulation. We show that this model allows us to learn dialogue policies that automati-cally adapt online to new users, and that these policies are significantly better than threshold-based adaptive hand-coded poli-cies for this problem. 1