Semi-supervised Learning of Utterances Using Hidden Vector State Language Model

Spoken dialogue system has an uncertain parameter during the speech recognition which controls its performance that vary for the different users as well as for the same user during multiple repetitions of even the same dialogue. This paper discusses how recognition errors in the users utterances can be handled by making use of semi-supervised learning techniques over the hidden vector state (HVS) model. The HVS Model is an extension of basic Markov model in which the context is encoded in each state as a vector. The state transitions in the HVS are factored into a stack shift operation similar to the push-down automaton. HVS-Model being a statistical model requires lot of labeled training data which is practically difficult. In this paper we present how classification and expectation-maximization semi-supervised learning approaches can be trained on both labeled and unlabelled corpora for handling the uncertainty by the user as well as the recognition errors by speech recognition system. The experimental results show that the proposed framework using the HVS model can improve the performance of the dialogue management of the spoken dialogue system when compared with the baseline model.