Deep Reinforcement Learning for Conversational Systems

Deep reinforcement learning (RL) has seen tremendous successes in solving video and board games, by leveraging great representational power of deep-learning models in the RL framework. More applications are emerging in robotics as well as natural language processing. In this talk, we demonstrate how deep RL can be used to develop dialogue systems that can converse like humans and help users solve specific tasks. In particular, we report work on three related projects: end-to-end training with an external knowledge base, domain extension where efficient exploration is required, and composite (hierarchical) task-completion dialogues.

Authors

• Lihong Li

Keywords

No keywords are indexed for this paper.