Dialogue Act Classification in Domain-Independent Conversations Using a Deep Recurrent Neural Network

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2012–2021, Osaka, Japan, December 11-17 2016. Dialogue Act Classification in Domain-Independent Conversations Using a Deep Recurrent Neural Network Hamed Khanpour, Nishitha Guntakandla, and Rodney Nielsen University of North Texas HiLT Lab {hamedkhanpour, nishithaguntakandla}@my.unt.edu rodney.nielsen@unt.edu Abstract In this study, we applied a deep LSTM structure to classify dialogue acts (DAs) in open-domain conversations. We found that the word embeddings parameters, dropout regularization, decay rate and number of layers are the parameters that have the largest effect on the final system accu- racy. Using the findings of these experiments, we trained a deep LSTM network that outperforms the state-of-the-art on the Switchboard corpus by 3.11%, and MRDA by 2.2%. 1 Introduction Dialogue Act (DA) classification plays a key role in dialogue interpretation, especially in spontaneous conversation analysis. Dialogue acts are defined as the meaning of each utterance at the illocutionary force level (Austin, 1975). Many applications benefit from the use of automatic dialogue act classi- fication such as dialogue systems, machine translation, Automatic Speech Recognition (ASR), topic identification, and talking avatars (Kr´al and Cerisara, 2012). Due to the complexity of DA classification, most researchers prefer to focus on the task-oriented systems such as restaurant, hotel, or flight, etc. reservation systems. Almost all standard approaches to classification have been applied in DA classification, from Bayesian Networks (BN) and Hidden Markov Models (HMM) to feed forward Neural Networks, Decision Trees (DT), Support Vector Machines (SVM) and rule-based approaches. Recently, the advancement of research in deep learning has led to performance upheavals in many Natural Language Processing (NLP) tasks, even leading Manning (2016) to refer to the phenomenon as a neural network ”tsunami”. One of the main benefits of using deep learning approaches is that they are not as reliant on handcrafted features; instead, they manufacture features automatically from each word (Turian et al., 2010), sentence (Lee and Dernoncourt, 2016; Kim, 2014), or even long texts (Collobert et al., 2011; Mikolov et al., 2013; Pennington et al., 2014). Inspired by the performance of recent studies utilizing deep learning for improving DA classification in domain-independent conversations (Ji et al., 2016; Lee and Dernoncourt, 2016; Kalchbrenner and Blunsom, 2013), we propose a model based on a recurrent neural network, LSTM, that benefits from deep layers of networks and pre-trained word embeddings derived from Wikipedia articles. 2 Related Work Prior work has defined general sets of DAs for domain-independent dialogues that are commonly used in almost all research on DA classification (Jurafsky et al., 1997; Dhillon et al., 2004). The task of DA classification (sometimes called DA identification) is to attribute one member of a predefined DA to each given utterance. Therefore, DA classification is sometimes treated as short-text classification. Similar to many other traditional text classification methods, five sources of information have been used for DA classification tasks: lexical information, syntax, semantics, prosody, and dialogue history. Among all This work is licensed under a Creative Commons Attribution 4.0 International License. License details: http:// …

Content truncated to 500 words