Extracting Phrase Patterns with Minimum Redundancy for Unsupervised Speaker Role Classification

This paper addresses the problem of learning phrase patterns for unsupervised speaker role classification. Phrase patterns are automatically extracted from large corpora, and redundant patterns are removed via a graph pruning algorithm. In experiments on English and Mandarin talk shows, the use of phrase patterns results in an increase of role classification accuracy over n-gram lexical features, and more compact phrase pattern lists are obtained due to the redundancy removal.