Dialog Act classification using acoustic and discourse information of MapTask Data

Abstract

In this work, we analyze both acoustic and discourse information for Dialog Act (DA) classification of HCRC MapTask dataset. We extract several different acoustic features and exploit these features in a Hidden Markov Model (HMM) to classify acoustic information. For discourse feature extraction, we propose a novel parts-of-speech (POS) tagging technique that effectively reduces the dimensionality of discourse features manyfold. To classify discourse information, we exploit two classifiers such as a HMM and a Support Vector Machine (SVM) respectively. We further obtain classifier fusion between HMM and SVM to improve discourse classification. Finally, we perform an efficient decision-level classifier fusion for both acoustic and discourse information to classify twelve different DAs in HCRC MapTask data. We obtain accuracy of rate 65.2% (58.06% with cross validation) and 55.4% (51.08% with cross validation) DA classification using acoustic and discourse information respectively. Furthermore, we obtain combined accuracy of 68.6% (61.02% with cross validation) for DA classification. These accuracy rates of DA classification are comparable to previously reported results for the same HCRC MapTask dataset. In terms of average Precision and Recall, we obtain accuracy of 74.89% and 69.83% (without cross validation) respectively. Therefore, we obtain much better precision and recall rate for most of the classified DAs when compared to existing works on the same dataset.