Using Reinforcement Learning for Dialogue Act Classification in Task-oriented Conversation Systems

Abstract

With rapid development and comprehensive application of artificial intelligence (AI), dialogue system, question-answering system and chatting robot go into people's daily lives. Therefore, DA classification plays an important role because it helps the system handle different types of user motivations. For this reason, most of dialogue systems, especially task-oriented dialogue systems, create a DA classification module in order to make the whole framework more vivid and simpler to develop and usually it does determine the overall quality of the output result. Previous work on DA classification focused on how to improve the performance of neural network models or deep learning models by modifying or mixing the model framework, most commonly the inner layers. However, little attention was paid on investigating the effect of reinforcement learning (RL) on DA classification. We simply use convolutional neural networks (CNN) as our baseline system and conduct our training progress using reinforcement learning. Compared to traditional methods and machine learning methods, our method firstly exploits reinforcement learning for DA classification and we conduct certain experiments to demonstrate the significant improvement of classification results in task-oriented dialogue systems. We create two Chinese-language industrial datasets by collecting data from task-oriented conversation systems in restaurant-ordering and cooking assistant and set up live study experiments to prove that our method not only improves the performance of DA classification, but also makes the overall quality of dialogue system better.