Exploiting prosodic features for dialog act tagging in a discriminative modeling framework

Cue-based automatic dialog act tagging uses lexical, syntactic and prosodic knowledge in the identification of dialog acts. In this paper, we propose a discriminative framework for automatic dialog act tagging using maximum entropy modeling. We propose two schemes for integrating prosody in our modeling framework: (i) Syntaxbased categorical prosody prediction from an automatic prosody labeler, (ii) A novel method to model continuous acoustic-prosodic observation sequence as a discrete sequence through the means of quantization. The proposed prosodic feature integration results in a relative improvement of 11.8% over using lexical and syntactic features alone on the Switchboard-DAMSL corpus. The performance of using the lexical, syntactic and prosodic features results in an dialog act tagging accuracy of 84.1%, close to the human agreement of 84%.