Context-aware Natural Language Generation for Spoken Dialogue Systems

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2032–2041, Osaka, Japan, December 11-17 2016. Context-aware Natural Language Generation for Spoken Dialogue Systems Hao Zhou, Minlie Huang, Xiaoyan Zhu State Key Laboratory of Intelligent Technology and Systems, National Laboratory for Information Science and Technology, Dept. of Computer Science and Technology, Tsinghua University, Beijing 100084, PR China tuxchow@gmail.com , aihuang@tsinghua.edu.cn , zxy-dcs@tsinghua.edu.cn Abstract Natural language generation (NLG) is an important component of question answering(QA) sys- tems which has a significant impact on system quality. Most tranditional QA systems based on templates or rules tend to generate rigid and stylised responses without the natural variation of hu- man language. Furthermore, such methods need an amount of work to generate the templates or rules. To address this problem, we propose a Context-Aware LSTM model for NLG. The model is completely driven by data without manual designed templates or rules. In addition, the con- text information, including the question to be answered, semantic values to be addressed in the response, and the dialogue act type during interaction, are well approached in the neural network model, which enables the model to produce variant and informative responses. The quantitative evaluation and human evaluation show that CA-LSTM obtains state-of-the-art performance. 1 Introduction Natural language generation (NLG), the task of generating natural language from a knowledge base or a logical form representation, is an important component of dialogue or question answering system. NLG can be treated as a single-turn dialogue generation. Traditional approaches to NLG problem are mostly rule-based or template-based (Bateman and Henschel, 1999; Busemann and Horacek, 2002). However, these methods tend to generate rigid and stylised language without the natural variation of human language. In addition, they need a heavy workload to design the templates or rules. Recently due to the growth of artificial neural networks and the increase of labeled data available on the Internet, data-driven approaches are developed to attack the NLG problem (Ritter et al., 2011; Shang et al., 2015). Shang et al. (2015) and Serban et al. (2015) apply the RNN-based general encoder-decoder framework to the open-domain dialogue response generation task. Although their model can generate the relevant and variant responses according to the input text in a statistical manner, the quality and content of responses depend on the quality and quantum of the training corpus. Wen et al. (2015) propose a task- oriented NLG model that can generate the responses providing the correct answers given the dialogue act (for instance, confirm or request some information), including the answer information. However the context information, such as the input question and dialogue act, is ignored. Yin et al. (2016) propose a neural network model that can generate answers to simple factoid questions based on a knowledge base. But a large error rate is observed due to the complex architecture introduced. In this paper, we deal with the NLG problem in this setting: given a question, the corresponding dialogue act, and the semantic slots to be addressed in the …

Content truncated to 500 words