Paul A. Crook and Oliver Lemon."Accurate Probability Estimation of Hypothesised User Acts for POMDP Approaches to Dialogue Management"

Post date: Jan 19, 2011 12:16:48 PM

Published in Proceedings of the 12^th Annual Research Colloquium of the special-interest group for computational linguistics in the UK and Ireland (CLUKI 2009), Dublin, April 2009.

Abstract:

Current Partially Observable Markov Decision Process (POMDP) Reinforcement Learning (RL) approaches to dialogue management require accurate estimates of the probability of each hypothesised user semantic act given the input user's utterance. In previous work, the probabilities for each hypothesised user act have been approximated using the Automatic Speech Recogniser's (ASR's) confidence score for each item in an N-best list. In this paper we examine this approximation and propose a better one. Our results suggest that this approximation does not in general hold. The validity of the assumption also depends on the choice of "confidence score" measure. Our results provide independent support for a previous result which showed that confusion network-based inference evidence (IE) scoring should be preferred for POMDP dialogue management. We also propose a simple remapping of IE scores which should improved probability estimates and show how our analysis can be used to build data driven Automatic Speech Recogniser - Spoken Language Understanding (ASR-SLU) models for simulated user training.

A copy of the paper (pdf) and bibtex entry are provided below. You can also view the slides from the conference presentation.