Investigating the Portability of Corpus-Derived Cue Phrases for Dialogue Act Classification

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 977–984 Manchester, August 2008 Investigating the Portability of Corpus-Derived Cue Phrases for Dialogue Act Classification Nick Webb and Ting Liu ILS Institute University at Albany, SUNY Albany, NY, USA {nwebb|tl7612}@albany.edu Abstract We present recent work in the area of Cross-Domain Dialogue Act tagging. Our experiments investigate the use of a sim- ple dialogue act classifier based on purely intra-utterance features - principally in- volving word n-gram cue phrases. We ap- ply automatically extracted cues from one corpus to a new annotated data set, to de- termine the portability and generality of the cues we learn. We show that our auto- matically acquired cues are general enough to serve as a cross-domain classification mechanism. 1 Introduction A number of researchers (Hirschberg and Litman, 1993; Grosz and Sidner, 1986) speak of cue or key phrases in utterances that can serve as useful indi- cators of discourse structure. We have previously investigated the use of such cue phrases to predict dialogue acts or DAs (functional tags which rep- resent the communicative intentions behind each user utterance) (Webb et al., 2005a). We devel- oped an approach, in common with the work of Samuel et al. (1999), where word n-grams that might serve as cue phrases are automatically de- tected in a corpus and we have previously reported the results of experiments evaluating this approach on the SWITCHBOARD corpus, where our results rival the best reported over that data (Stolcke et al., 2000), although our method adopts a significantly less complex algorithm. c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. An interesting by-product of our approach is the ranked list of cue phrases derived from the source corpus. Visual inspection of these cues reveals that, as one might expect, there is a high degree of correlation between phrases such as “can you” and the DA <yes/no question>, “where is” and “who is” with the DA <wh-question> and “right” or “ok” with DA <agree/accept>. These cues appear to be of a general nature, unrelated to the source domain or application. Therefore, despite being automatically acquired from one domain specific corpus, these cues should be equally applicable to new corpora, from a different domain and it is this hypothesis we test. This paper presents our work on dialogue act classification using cues automat- ically extracted from a corpus from one domain, and applying these cues directly as a classifier over a new corpus from a different domain. The material is presented as follows: Previous work with dialogue act modelling is outlined in Section 2. An overview of the corpora used for the experiments we report can be seen in Section 3. A brief overview of our classification method is given in Section 4. Our experiments evaluating the cue-based dialogue act classifier tagging new, out-of-domain data are given in Section 5. Finally we end with some discussion and an outline of in- tended further work. 2 Related …

Content truncated to 500 words