Unsupervised broadcast conversation speaker role labeling

We present an approach to unsupervised speaker role labeling in talk show data that makes use of two complementary sets of features: structural features that encode the participation patterns of speakers, and lexical features, which capture characteristic phrases. Techniques for using multiple clusterings are explored, leading to more robust results. Experiments on English and Mandarin talk shows yield performance similar to that reported for broadcast news using supervised learning.