Extracting Phrase Patterns with Minimum Redundancy for Unsupervised Speaker Role Classification

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 717–720, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Extracting Phrase Patterns with Minimum Redundancy for Unsupervised Speaker Role Classification Bin Zhang, Brian Hutchinson, Wei Wu and Mari Ostendorf∗ University of Washington, Seattle, WA 98125 Abstract This paper addresses the problem of learning phrase patterns for unsupervised speaker role classification. Phrase patterns are automati- cally extracted from large corpora, and redun- dant patterns are removed via a graph prun- ing algorithm. In experiments on English and Mandarin talk shows, the use of phrase pat- terns results in an increase of role classifi- cation accuracy over n-gram lexical features, and more compact phrase pattern lists are ob- tained due to the redundancy removal. 1 Introduction The identification of speaker roles is fundamental to the analysis of social content and information re- liability in conversational speech. Previous work has primarily used supervised learning in automatic role classification. Barzilay et al. (2000) described a speaker role classification system for English broad- cast news (BN), where the speakers were catego- rized into three types: anchor, journalist, and guest. The authors used supervised learning to discover n- gram signature phrases for speaker introduction and structural features such as duration, achieving an ac- curacy of 80% on ASR derived transcripts. Liu et al. (2006) studied speaker role classification on TDT-4 Mandarin BN data. Hidden Markov and maximum entropy models were used to label the sequence of speaker turns with the roles anchor, reporter, and other, based on n-gram features, which yielded 80% classification accuracy on human transcripts. Hutchinson et al. (2010) extend previous work to the case of unsupervised learning, with the goal of portability across languages. That work explored ∗ This research was funded by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA). All statements of fact, opinion or conclusions contained herein are those of the authors and should not be construed as representing the official views or policies of IARPA, the ODNI or the U.S. Government. speaker role classification using structural and n- gram features on talk show (or broadcast conversa- tion (BC)) data. In this paper, we address a limita- tion of n-grams as features by proposing a method for learning phrases with gaps, which is particularly important for conversational speech, since there are more disfluencies that can cause failure of n-gram matching. In addition, we want to avoid topic words (e.g., proper nouns) in order to model speaker roles rather than topics. For example, for identifying the host, the phrase pattern “We’ll be back with * in a minute” is more general than the n-grams “We’ll be back with John Smith in a minute.” To prevent these problems with n-grams, one must limit the length of learned n-grams, making them less discriminative. Phrase patterns have been used in other NLP ap- plications such as (Sun et al., 2007). To remove the redundancies in the automatically extracted phrase …

Content truncated to 500 words