Controlling listening-oriented dialogue using partially observable Markov decision processes

This paper investigates how to automatically create a dialogue control component of a listening agent to reduce the current high cost of manually creating such components. We collected a large number of listening-oriented dialogues with their user satisfaction ratings and used them to create a dialogue control component using partially observable Markov decision processes (POMDPs), which can learn a policy to satisfy users by automatically finding a reasonable reward function. A comparison between our POMDP-based component and other similarly motivated systems using human subjects revealed that POMDPs can satisfactorily produce a dialogue control component that can achieve reasonable subjective assessment.