Task-oriented Dialogue System Based on Reinforcement Learning
Meina Song, Zhongfu Chen, Peiqing Niu, E Haihong · 2019
In this paper, we propose a task-oriented dialogue system based on reinforcement learning. The overall system is composed of three parts: natural language understanding (NLU), dialogue management (DM) and natural language generation (NLG). And our model can interact with the database in real time and acquire effective information from it. For the DM part, reinforcement learning is applied. Specially, we adopt an improved double deep Q-learning (DQN) strategy. In that case, the DM agent can resist the environmental noise considerably. Besides, we put forward a joint model for NLU module, and the experiments on ATIS and Snips datasets have proved the effectiveness of the joint model. For the overall system, the experiments are conducted on a public movie-ticket booking dataset. The experimental results indicate that the proposed model has outperformed the traditional rule-based multi-turn dialogue system both on simulated and real users. Besides, the double-DQN agent has better performance for both objective and subject evaluation, which demonstrates the effectiveness and superiority of our proposed model.