PROBABILISTIC MODEL FOR RESPONSE GENERATION WITH APPLICATION OF BAYESIAN NETWOKRS

Abstract

This paper presents Bayesian Networks for probabilistic response generation in a mixed-initiative, transaction dialogue system. The probabilistic approach is to compute the probability of each server response based on the training corpus, and returns to user the response with the highest probability of answering the semantic input from user utterance. Apart from the dialogue acts as provided by the DAMSL-annotated corpus, we proposed two additional types of random variables for the Bayesian Networks, which are the Surface Linguistic Features and the Dialogue Context Features. The BN representation of the dialogue corpus is hoped to allow the system to reason about the dialogue utterances available so far and predict the next system response that will satisfy the user utterance both linguistically and contextually.