A Path Value-Aware Reinforcement Learning Method for Knowledge Graph Question Answering

Reinforcement learning-based multi-hop knowledge graph question answering is modeled as path-based reasoning, in which immediate rewards for intermediate actions in the path are critical for alleviating aimless exploration. Existing works consider the time-step distance between the entities in intermediate actions and the final answers to calculate immediate rewards, which reward incorrect intermediate actions on the false positive path that accidentally connects the correct answer entity. In this paper, we propose a Path Value-aware reinforcement learning method for knowledge graph Question Answering (PVQA), it regards the question as guidance of the reasoning path to overcome the challenge. Specifically, firstly, we regard the question as guidance and compare the similarity between the question and the reasoning path to quantify the path value. Secondly, to ensure the accuracy of similarity in incomplete knowledge graphs, we embed the semantic and structural features into the representation of the question and path. Finally, we design a joint reward function by combining the feasibility of the answer with the path value. Extensive experiments demonstrate that our proposal outperforms the state-of-the-art.