Dialogue-act tagging using smart feature selection; results on multiple corpora

This paper presents an overview of our on-going work on dialogue-act classification. Results are presented on the ICSI, switchboard, and on a selection of the AMI corpus, setting a baseline for forthcoming research. For these corpora the best accuracy scores obtained are 89.27%, 65.68% and 59.76%, respectively. We introduce a smart compression technique for feature selection and compare the performance from a subset of the AMI transcriptions with AMI-ASR output for the same subset.