Toward Self-Learning End-to-End Dialog Systems

End-to-end task-oriented dialog systems often suffer from out-of-distribution (OOD) inputs after being deployed in dynamic, changing, and open environments. In this work, we propose SL-Agent, a self-learning framework that combines supervised learning, reinforcement learning, and machine teaching for building end-to-end dialog systems in a more realistic changing environment setting. SL-Agent consists of a dialog model and a pre-trained reward model to judge the quality of a system response. SL-Agent enables dialog agents to automatically adapt to environments with user behavior changes by learning from human-bot interactions via reinforcement learning, with the incorporated pre-trained reward model. We validate SL-Agent in four different dialog domains. Experimental results show the effectiveness of SL-Agent for automatically adapting to changing environments using both automatic and human evaluations. Furthermore, experiments on a challenging domain extension setting demonstrate that SL-Agent can effectively adapt to new tasks using limited human corrections provided via machine teaching. We will release code, data, and pre-trained models for further research.