Comparison of Learning Approaches to Dialogue Management in Conversational Systems

Abstract

Dialogue systems have an increasingly higher number of applications and so their development is raising interest both in academic and industrial setting. Dialogue management is a key aspect for the development of these systems, as it is in charge of the decision making processes and the identification of the most appropriate responses to the user inputs. In this paper we compare different statistical techniques to train dialogue managers using generally available corpora at the disposal for the scientific community. Our results show that the use of generative models and in particular an intent classifier with Neural Networks and Seq2Seq using GRU cells attain the best accuracy for dialogue management.