Path-Based Knowledge Graph Completion Combining Reinforcement Learning with Soft Rules

Knowledge graphs are useful resources for numerous applications, but they are often greatly incomplete. A popular method of knowledge graph completion is path-based reasoning that is to infer new relations by connecting other existing paths at a pair of entities. But this traditional method does not consider the reality of paths. In this paper, we propose a model that combines the reinforcement learning (RL) framework with soft rules to learn reasoning path. In our model, we adjust the partially observed Markov decision process and extract the soft rules with different confidence levels from datasets. In contrast to prior work, we modify the reward function to make RL-agent incline to choose the paths which conform to soft rules, and set the probability of reasoning paths. Meanwhile, we make some restrictions in order to extract soft rules with high confidence levels. We analyze the complexity of our algorithm and use an instance to evaluate the correctness of our model.