Dialogue Act Classification with Context-Aware Self-Attention

Proceedings of NAACL-HLT 2019, pages 3727–3733 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 3727 Dialogue Act Classification with Context-Aware Self-Attention Vipul Raheja Joel Tetreault Grammarly firstname.lastname@grammarly.com Abstract Recent work in Dialogue Act classification has treated the task as a sequence labeling problem using hierarchical deep neural net- works. We build on this prior work by lever- aging the effectiveness of a context-aware self- attention mechanism coupled with a hierarchi- cal recurrent neural network. We conduct ex- tensive evaluations on standard Dialogue Act classification datasets and show significant im- provement over state-of-the-art results on the Switchboard Dialogue Act (SwDA) Corpus. We also investigate the impact of different utterance-level representation learning meth- ods and show that our method is effective at capturing utterance-level semantic text repre- sentations while maintaining high accuracy. 1 Introduction Dialogue Acts (DAs) are the functions of ut- terances in dialogue-based interaction (Austin, 1975). A DA represents the meaning of an utter- ance at the level of illocutionary force, and hence, constitutes the basic unit of linguistic communica- tion (Searle, 1969). DA classification is an impor- tant task in Natural Language Understanding, with applications in question answering, conversational agents, speech recognition, etc. Examples of DAs can be found in Table 1. Here we have a conver- sation of 7 utterances between two speakers. Each utterance has a corresponding label such as Ques- tion or Backchannel. Early work in this field made use of statis- tical machine learning methods and approached the task as either a structured prediction or text classification problem (Stolcke et al., 2000; Ang et al., 2005; Zimmermann, 2009; Surendran and Levow, 2006). Many recent studies have pro- posed deep learning models for the DA classifi- cation task with promising results (Lee and Der- noncourt, 2016; Khanpour et al., 2016; Ortega and Speaker Utterance DA label A Okay. Other A Um, what did you do this weekend? Question B Well, uh, pretty much spent most of my time in the yard. Statement B [Throat Clearing] Non Verbal A Uh-Huh. Backchannel A What do you have planned for your yard? Question B Well, we’re in the process of, revitalizing it. Statement Table 1: A snippet of a conversation sample from the SwDA Corpus. Each utterance has a corresponding di- alogue act label. Vu, 2017). However, most of these approaches treat the task as a text classification problem, treat- ing each utterance in isolation, rendering them un- able to leverage the conversation-level contextual dependence among utterances. Knowing the text and/or the DA labels of the previous utterances can assist in predicting the current DA state. For in- stance, in Table 1, the Answer or Statement dialog acts often follow Question type utterances. This work draws from recent advances in NLP such as self-attention, hierarchical deep learning models, and contextual dependencies to produce a dialogue act classification model that is effective across multiple domains. Specifically, we propose a hierarchical deep neural network to model dif- ferent levels of utterance and dialogue act seman- tics, …

Content truncated to 500 words