Reinforcement learning with dynamic completion for answering multi-hop questions over incomplete knowledge graph

Hai Cui, Tao Peng, Ridong Han, Beibei Zhu, Haijia Bi, Lu Liu

Abstract: Highlights•The sparse KG is dynamically augmented with additional unseen actions.•This paper extends the line of RL-based methods for better interpretability.•The process of action space completion is separated into two sub-steps.•The relation pruner filters out noisy and inappropriate candidate relations.•The beam search-based action selection strategy maintains multiple partial paths.