A Reinforcement Learning-based Knowledge Graph Reasoning Verification Method

To address more complex problems and enhance the effectiveness of reasoning paths, multi-hop reasoning methods utilizing reinforcement learning have gained significant attention recently. Multi-hop reasoning aims to predict missing relationships or identify the correct answer entity by exploring multiple triples in a knowledge graph. However, existing methods often overlook the semantic information of intermediate reasoning paths, which could enhance answer accuracy, resulting in incorrect paths. Additionally, in complex scenarios, the action search space grows exponentially with more reasoning hops, drastically reducing task efficiency. This paper proposes a Reinforcement Path Reasoning Verification (RPRV) method that integrates reinforcement learning techniques. By implementing a hop limitation strategy and incorporating additional path reward signals into the reward function, this method addresses incorrect reasoning paths and improves search efficiency. These adjustments ensure that the model remains close to the initial entity during reasoning, enhancing both answer accuracy and path interpretability. Experimental results demonstrate that the model performs well on three publicly available benchmark datasets, significantly improving knowledge graph reasoning efficiency and prediction accuracy through the path validator.