Scaling up POMDPs for Dialog Management: The ``Summary POMDP'' Method

Partially observable Markov decision processes (POMDPs) have been shown to be a promising framework for dialog management in spoken dialog systems. However, to date, POMDPs have been limited to artificially small tasks. In this work, we present a novel method called a "summary POMDP" for scaling slot-filling POMDP-based dialog managers to cope with tasks of a realistic size. An example dialog problem incorporating a user model built from real dialog data is presented. A dialog manager is created using this method and evaluated using a second user model created from held-out dialog data. Results confirm that summary POMDP policies scale well, and also show that summary POMDP policies are reasonably robust to variations in user behavior