Using symbolic prominence to help design feature subsets for topic classification and clustering of natural human-human conversations

In this work, we use the output of a symbolic prominence classifier rather than acoustic cues of prominence, to improve the tasks of clustering and classification of spontaneous conversations to topics. In our experiments, we combine the output of a prominence classifier with lexical feature selection and combination methods to build improved feature subsets. Evaluated for the task of topic classification on a subset of Switchboard-I, the combination method offered a 11% relative reduction of classification error compared to using lexical-only feature selection methods; similar gains are reported for clustering.