Partially Observable Markov Decision Processes with Continuous Observations for Dialogue Management

This work shows how a spoken dialogue system can be represented as a Partially Observable Markov Decision Process (POMDP) with composite observations consisting of discrete elements representing dialogue acts and continuous components representing confidence scores. Using a testbed simulated dialogue management problem and recently developed optimisation techniques, we demonstrate that this continuous POMDP can outperform traditional approaches in which confidence score is tracked discretely. Further, we present a method for automatically improving handcrafted dialogue managers by incorporating POMDP belief state monitoring, including confidence score information. Experiments on the test-bed system show significant improvements for several example handcrafted dialogue managers across a range of operating conditions.