Cost-level integration of statistical and rule-based dialog managers

Statistical dialog managers can potentially make more robust decisions than their rule-based counterparts, because they can account for uncertainties due to errors in speech recognition and natural language understanding. In practice, however, statistical dialog managers can be difficult to use, as they may require a large number of parameters to be inferred from limited data. Consequently, hand-crafted rule based systems are still effective for practical use. This paper proposes a method to integrate an existing rule-based dialog manager with a statistical dialog manager based on Bayes decision theory, by incorporating the rule-based dialog manager into the cost function of the statistical dialog manager. The cost function has two parts: an efficiency cost that penalizes inefficient actions, as in conventional statistical dialog approaches, and a regularization cost that slightly penalizes system actions that differ from those that would be chosen by the rule-based system. Our experiments, which use a destination-setting task in an automobile dialog scenario, demonstrate that the integrated system produces system actions that are similar to those of an existing rule-based dialog manager but enable task completion using fewer turns than the rule-based system.