Backoff Model Training using Partially Observed Data: Application to Dialog Act Tagging

Proceedings of the Human Language Technology Conference of the North American Chapter of the ACL, pages 280–287, New York, June 2006. c⃝2006 Association for Computational Linguistics Backoff Model Training using Partially Observed Data: Application to Dialog Act Tagging Gang Ji and Jeff Bilmes Department of Electrical Engineering University of Washington Seattle, WA 98105-2500 {gang,bilmes}@ee.washington.edu Abstract Dialog act (DA) tags are useful for many applications in natural language process- ing and automatic speech recognition. In this work, we introduce hidden backoff models (HBMs) where a large generalized backoff model is trained, using an embed- ded expectation-maximization (EM) pro- cedure, on data that is partially observed. We use HBMs as word models condi- tioned on both DAs and (hidden) DA- segments. Experimental results on the ICSI meeting recorder dialog act corpus show that our procedure can strictly in- crease likelihood on training data and can effectively reduce errors on test data. In the best case, test error can be reduced by 6.1% relative to our baseline, an improve- ment on previously reported models that also use prosody. We also compare with our own prosody-based model, and show that our HBM is competitive even without the use of prosody. We have not yet suc- ceeded, however, in combining the bene- fits of both prosody and the HBM. 1 Introduction Discourse patterns in natural conversations and meetings are well known to provide interesting and useful information about human conversational be- havior. They thus attract research from many differ- ent and beneficial perspectives. Dialog acts (DAs) (Searle, 1969), which reflect the functions that ut- terances serve in a discourse, are one type of such patterns. Detecting and understanding dialog act patterns can provide benefit to systems such as au- tomatic speech recognition (ASR) (Stolcke et al., 1998), machine dialog translation (Lee et al., 1998), and general natural language processing (NLP) (Ju- rafsky et al., 1997b; He and Young, 2003). DA pat- tern recognition is an instance of “tagging.” Many different techniques have been quite successful in this endeavor, including hidden Markov models (Ju- rafsky et al., 1997a; Stolcke et al., 1998), seman- tic classification trees and polygrams (Mast et al., 1996), maximum entropy models (Ang et al., 2005), and other language models (Reithinger et al., 1996; Reithinger and Klesen, 1997). Like other tagging tasks, DA recognition can also be achieved using conditional random fields (Lafferty et al., 2001; Sut- ton et al., 2004) and general discriminative model- ing on structured outputs (Bartlett et al., 2004). In many sequential data analysis tasks (speech, lan- guage, or DNA sequence analysis), standard dy- namic Bayesian networks (DBNs) (Murphy, 2002) have shown great flexibility and are widely used. In (Ji and Bilmes, 2005), for example, an analysis of DA tagging using DBNs is performed, where the models avoid label bias by structural changes and avoid data sparseness by using a generalized back- off procedures (Bilmes and Kirchhoff, 2003). Most DA classification procedures assume that within a sentence of a particular fixed DA type, there is a fixed word distribution …

Content truncated to 500 words