Using Context Information for Dialog Act Classification in DNN Framework

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2170–2178 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Using Context Information for Dialog Act Classification in DNN Framework Yang Liu and Kun Han and Zhao Tan and Yun Lei Applied Machine Learning, Facebook Abstract Previous work on dialog act (DA) classifi- cation has investigated different methods, such as hidden Markov models, maximum entropy, conditional random fields, graph- ical models, and support vector machines. A few recent studies explored using deep learning neural networks for DA classifi- cation, however, it is not clear yet what is the best method for using dialog context or DA sequential information, and how much gain it brings. This paper proposes several ways of using context informa- tion for DA classification, all in the deep learning framework. The baseline system classifies each utterance using the con- volutional neural networks (CNN). Our proposed methods include using hierar- chical models (recurrent neural networks (RNN) or CNN) for DA sequence tagging where the bottom layer takes the sentence CNN representation as input, concatenat- ing predictions from the previous utter- ances with the CNN vector for classifica- tion, and performing sequence decoding based on the predictions from the sentence CNN model. We conduct thorough ex- periments and comparisons on the Switch- board corpus, demonstrate that incorporat- ing context information significantly im- proves DA classification, and show that we achieve new state-of-the-art performance for this task. 1 Introduction Dialog act (DA) represents a function of a speaker’s utterance in either human-to-human or human-to-computer conversations. Correct identi- fication of DAs is important for understanding hu- man conversations, as well as for developing intel- ligent human-to-computer dialog systems (either written or spoken dialogs). For example, recogniz- ing DAs can help identify questions and answers in meetings, customer service, online forum, etc. Many machine learning techniques have been in- vestigated and shown reasonable performance for DA classification, for example, (Ang et al., 2005; Ji and Bilmes, 2005; Kalchbrenner and Blunsom, 2013; Ribeiro et al., 2015), just to name a few. Intuitively we would expect that leveraging dia- log context can help classify the current utterance. For example, if the previous sentence is a question, then there is a high probability that the current sen- tence is a response to that question. Such con- text information has been explored in some pre- vious methods, for example, hidden Markov mod- els (HMM), conditional random fields (CRF), dy- namic Bayesian networks (DBN). Given the re- cent success of the deep learning framework in various language processing tasks, in this work we also employ neural networks for DA classifica- tion. In fact, such models have been used in some recent studies for DA classification, e.g., (Rojas- Barahona et al., 2016; Kalchbrenner and Blunsom, 2013; Zhou et al., 2015); however, previous work has not thoroughly evaluated the use of context in- formation for this task, and there is still a lack of good understanding about how we can use context information and how …

Content truncated to 500 words