SCR: A completion-then-reasoning framework for multi-hop question answering over incomplete knowledge graph

Abstract

Reinforcement learning has become the widely adopted technique for multi-hop knowledge graph question answering task thanks to its excellent interpretability in reasoning process. However, it is severely affected by the incompleteness of knowledge graphs and the sparse rewards caused by weak supervision. In this paper, we propose a completion-then-reasoning framework, called SCR, to address these two issues. For the incompleteness of knowledge graphs, we first extract a subgraph from the given knowledge graph for a given question, and use the knowledge graph embedding model to predict and complete missing triples, followed by reinforcement learning for answer reasoning on the completed subgraph. To alleviate the sparse rewards in reinforcement learning, we introduce a semantic reward based on the semantic similarity between original question and full relational path, enabling the model to receive partial rewards for partially correct paths instead of a zero reward. Detailed experiments on PQ, PQL, MetaQA, and WebQSP datasets demonstrate that our SCR model effectively improves the performance of multi-hop knowledge graph question answering task. Particularly, under sparse KG setting, SCR model outperforms baselines by a large margin, highlighting the effectiveness of completion-then-reasoning framework in mitigating the incompleteness of knowledge graphs.