Knowledge graphs (KGs) are often incomplete. Knowledge graph reasoning (KGR) seeks to infer new facts from the current information present in the KG. One popular approach to KGR is multi-hop reasoning within a Reinforcement learning (RL) framework, especially policy-based method REINFORCE. However, numerous previous RL-based models rely on conventional policy networks, which struggle with large state space and high-dimensional input and will lead to inaccurate results. In order to address this challenge, we propose a novel policy network with self-correction mechanism called Self-Rectified Policy Network (SRPN). Unlike conventional policy networks, SRPN rectifies the original predictions by making targeted modifications based on the current input, thereby enhancing the accuracy of the reasoning process. To evaluate SRPN’s effectiveness, we perform comparative experiments using three large datasets. According to the experimental results, SRPN outperforms most metrics compared to the originals. Furthermore, we explore key factors such as hyperparameters and reasoning path length to optimize the network’s performance.
Zhu et al. (Wed,) studied this question.