Speaker-change Aware CRF for Dialogue Act Classification

Proceedings of the 28th International Conference on Computational Linguistics, pages 450–464 Barcelona, Spain (Online), December 8-13, 2020 450 Speaker-change Aware CRF for Dialogue Act Classification Guokan Shang1,2, Antoine J.-P. Tixier1, Michalis Vazirgiannis1,3, Jean-Pierre Lorr´e2 1 ´Ecole Polytechnique, 2Linagora, 3AUEB Abstract Recent work in Dialogue Act (DA) classification approaches the task as a sequence labeling problem, using neural network models coupled with a Conditional Random Field (CRF) as the last layer. CRF models the conditional probability of the target DA label sequence given the input utterance sequence. However, the task involves another important input sequence, that of speakers, which is ignored by previous work. To address this limitation, this paper proposes a simple modification of the CRF layer that takes speaker-change into account. Experiments on the SwDA corpus show that our modified CRF layer outperforms the original one, with very wide margins for some DA labels. Further, visualizations demonstrate that our CRF layer can learn meaningful, sophisticated transition patterns between DA label pairs conditioned on speaker-change in an end-to-end way. Code is publicly available1. 1 Introduction A conversation can be seen as a sequence of utterances. The task of Dialogue Act (DA) classifica- tion aims at assigning to each utterance a DA label to represent its communicative intention. Dia- logue acts originate from the notion of illocutionary force (speaker’s intention in delivering an ut- terance) introduced back in the theory of Speech Act (Austin, 1962; Searle, 1969). DAs are as- signed based on a combination of syntactic, semantic, and pragmatic criteria (Stolcke et al., 2000). Change Speaker Utterance DA - B Of course I use, sd True A <laughter>. x True B credit cards. + False B I have a couple of credit cards sd True A Yeah. b True B and, uh, use them. + True A Uh-huh, b False A do you use them a lot? qy True B Oh, we try not to. ng Table 1: Fragment from SwDA conversation sw3332. Statement-non-opinion (sd), Non-verbal (x), Interruption (+), Acknowledge/Backchannel (b), Yes-No-Question (qy), Negative non-no an- swers (ng). As shown in Table 1, some examples of DAs include stating, questioning, answering, etc. The full set of DA labels is pre- defined. A number of annotation schemes have been devel- oped, varying from domain-specific, such as VERBMOBIL (Alexanderssony et al., 1997), to domain-independent, such as DAMSL (Allen and Core, 1997; Core and Allen, 1997) and DiAML2 (Bunt et al., 2010; Bunt et al., 2012). Automatically detecting DA labels is an essential step towards describing the discourse structure of conversation (Jurafsky et al., 1997). DAs are very useful annotations to a large variety of spoken language understanding tasks, such as utterance clustering (Shang et al., 2019), real-time infor- mation retrieval (Meladianos et al., 2017), conversational agents (Higashinaka et al., 2014; Ahmadvand et al., 2019), and summarization (Shang et al., 2018). It is difficult to predict the DA of a single utterance without having access to the other utterances in the context. For instance, for an utterance such as “Yeah”, it is hard …

Content truncated to 500 words