Abstract

Building spoken dialog systems is difficult because speech recognition errors are common and user’s behavior is unpredictable, which introduces uncertainty in the current state of the conversation. At AT&T, we have been applying partially observable Markov decision processes (POMDPs) to building these systems. We model the uncertainty in the dialog state explicitly as a Bayesian network and apply machine learning techniques to determine what the system should say or do.

In this talk, I’ll review the overall approach of applying statistical techniques and then describe two recent advances: first, because the system must operate in real-time, efficient Bayesian inference is crucial, yet the set of possible dialog states is enormous. To solve this, I’ll present a technique which uses a particle filter to perform approximate inference in real-time. Second, to choose actions, ideally we would like to combine the robustness of machine optimization with the expertise of human designers. To tackle this, I’ll present a method which unifies human expertise with automatic optimization.

To illustrate these techniques, I’ll provide examples of two dialog systems: a voice dialer, and a troubleshooting system that helps users restore connectivity on a failed DSL connection. Graphical displays illustrate the operation of the techniques, and quantitative results show that applying statistical techniques outperforms the traditional method of building systems by hand.