Efficient Dialogue Using a Probabilistic Nested User Model

We describe a set of dialogue simulation experiments, in which a probabilistic nested user model is employed in deciding between speech acts for a collaborative planning task, finding that a gain in utility can be obtained by using a probabilistic rather than a logical model. Given a set of ordinary dialogue plan rules, our system generates a gametree representation of the dialogue, using chance nodes to represent uncertain preconditions in the plan. Then, the game-tree is evaluated with respect to a given user model state.