Research On Knowledge Graph Reasoning Method Based On Attention Mechanism And Rule Injection

Knowledge graph reasoning aims to use the existing knowledge in the knowledge graph to mine the hidden errors or missing knowledge through the reasoning model automatically and efficiently,so as to alleviate the problems of sparseness and incompleteness in the knowledge graph,and better for the supported downstream applications such as personalized recommendation,intelligent question answering and natural language understanding.Due to its wide application scenarios and important research value,knowledge graph reasoning has attracted more and more researchers’ attention.In recent years,the path mining knowledge graph reasoning model based on reinforcement learning has made great progress.It regards the reasoning process as a Markov decision process,and gradually selects the most reasonable action through the continuous interaction between the agent and the knowledge graph environment,and uses the final reasoning path as the explanation of the reasoning process,which has high reasoning performance and interpretability.But the existing reinforcement learning based knowledge reasoning model underestimates the importance of adjacent candidate entity information in the reasoning process,which makes it difficult for the agent to choose similar actions composed by the1-N/N-N complex relations and prone to erroneous reasoning results.In addition,there is a wrong reasoning path problem in the reasoning process,i.e.,low-quality reasoning paths are difficult to explain the reasoning results.To solve the above problems,this thesis proposes an attention and rule injection based knowledge graph reasoning method.The main research contents are as follows:(1)Aiming at the entity selection problem caused by 1-N/N-N complex relations,an attention and reward shaping based reinforcement learning knowledge graph reasoning method is proposed,which uses the adjacent entity information and semantic information to alleviate the entity selection problem.Firstly,an attention mechanism is introduced to extract relevant hidden features from the adjacent entities.Then,an adversarial trained convolutional neural network is used to distinguish the rationality of reasoning paths through semantic features and feedback corresponding rewards.To mitigate the problem of sparse rewards caused by the terminal reward function,a potential-based reward shaping function is introduced,which considers the potential gap between agent states as the reward and without any pre-training.Compared with some embedding based reasoning models and interpretable reasoning models,the experimental results show that the reasoning performance of the proposed model is better than baselines on the WN18 RR dataset.Finally,the effectiveness of each component in the model is verified by the ablation experiment,and some reasoning cases are used to demonstrate the role of the proposed model in the entity selection problem and the existence of the wrong reasoning path problem in the reasoning process of the model.(2)Aiming at the wrong reasoning path problem,imitation learning is introduced on the basis of the model proposed above,and a rule-injection based generative adversarial imitation learning knowledge reasoning method is proposed,which performs imitation learning from semantic and logical rules to improve the quality of reasoning paths.Logical rules contain rich factual logic that can be used for reasoning.Thus,this thesis combines the logic rule and generative adversarial imitation learning to imitate and learn the reasoning strategy in the expert demonstration path from the semantic and rule levels.Firstly,a method for extracting high-quality expert demonstration paths is designed in the model.Then,the path semantic discriminator and the logic rule discriminator are designed respectively to discriminate the generated paths and feedback the corresponding adaptive rewards for imitation learning.Experimental results show that,compared with the reinforcement learning reasoning model proposed above,the reasoning performance of proposed generative adversarial imitation learning reasoning model is similar in the WN18 RR dataset,and improved in the NELL-995 dataset.Finally,the effectiveness of each component in the model is verified by the ablation experiment,and some reasoning cases show that the proposed model can alleviate the wrong reasoning path problem and improve the quality of the reasoning path.This thesis takes interpretable path mining based knowledge graph reasoning methods as the research goal.The main contributions are to propose corresponding solutions for the entity selection problem and the wrong reasoning path problem in the reinforcement learning based knowledge graph reasoning model,and verify the effectiveness of the proposed model through extensive experiments.