Cross-domain classification using generalized domain acts

Andrew N. Pargellis, Alexandros Potamianos

Abstract: Cross-domain classification for speech understanding is an interesting research problem because of the need for portable solutions in the design for spoken dialogue systems. In this paper, a two-tier classifier is proposed for speech understanding. The first tier consists of domain independent dialogue acts while the second tier consists of application actions that are domain specific. A maximum likelihood and a minimum classification error formulation are proposed for the first tier of the classifier, i.e., for dialogue act classification. The performance of the classifier is investigated for three application domains. Cross-domain classification error is two to four times higher than in-domain classification error. A 10-15% reduction in cross-domain classification error rate is achieved by adding generic domain independent training data for each dialogue act and by mapping words to semantic concepts.