A Dual-Attention Hierarchical Recurrent Neural Network for Dialogue Act Classification

Proceedings of the 23rd Conference on Computational Natural Language Learning, pages 383–392 Hong Kong, China, November 3-4, 2019. c⃝2019 Association for Computational Linguistics 383 A Dual-Attention Hierarchical Recurrent Neural Network for Dialogue Act Classification Ruizhe Li♠, Chenghua Lin♥, Matthew Collinson♠, Xiao Li♠ and Guanyi Chen♣ ♠Department of Computing Science, University of Aberdeen, UK {r02rl17, matthew.collinson, x.li}@abdn.ac.uk ♥Department of Computer Science, University of Sheffield, UK c.lin@sheffield.ac.uk ♣Department of Information and Computing Sciences, Utrecht University, The Netherlands g.chen@uu.nl Abstract Recognising dialogue acts (DA) is important for many natural language processing tasks such as dialogue generation and intention recognition. In this paper, we propose a dual- attention hierarchical recurrent neural network for DA classification. Our model is partially inspired by the observation that conversational utterances are normally associated with both a DA and a topic, where the former captures the social act and the latter describes the sub- ject matter. However, such a dependency be- tween DAs and topics has not been utilised by most existing systems for DA classifica- tion. With a novel dual task-specific atten- tion mechanism, our model is able, for utter- ances, to capture information about both DAs and topics, as well as information about the interactions between them. Experimental re- sults show that by modelling topic as an auxil- iary task, our model can significantly improve DA classification, yielding better or compara- ble performance to the state-of-the-art method on three public datasets. 1 Introduction Dialogue Acts (DA) are semantic labels of utter- ances, which are crucial to understanding com- munication: much of a speaker’s intent is ex- pressed, explicitly or implicitly, via social actions (e.g., questions or requests) associated with utter- ances (Searle, 1969). Recognising DA labels is important for many natural language processing tasks. For instance, in dialogue systems, know- ing the DA label of an utterance supports its in- terpretation as well as the generation of an appro- priate response (Searle, 1969; Chen et al., 2018). In the security domain, being able to detect inten- tion in conversational texts can effectively support the recognition of sensitive information exchanged in emails or other communication channels, which is critical to timely security intervention (Verma et al., 2012). A wide range of techniques have been inves- tigated for DA classification. Early works on DA classification are mostly based on general machine learning techniques, framing the prob- lem either as multi-class classification (e.g., us- ing SVMs (Liu, 2006) and dynamic Bayesian net- works (Dielmann and Renals, 2008)) or a struc- tured prediction task (e.g., using Conditional Ran- dom Fields (Kim et al., 2010; Chen et al., 2018; Raheja and Tetreault, 2019, CRF)). Recent stud- ies to the problem of DA classification have seen an increasing uptake of deep learning techniques, where promising results have been obtained. Deep learning approaches typically model the depen- dency between adjacent utterances (Ji et al., 2016; Lee and Dernoncourt, 2016). Some researchers further account for dependencies among both con- secutive utterances and consecutive DAs, i.e., both are considered factors that influence natu- ral dialogue (Kumar …

Content truncated to 500 words