Research On Multi-hop Reasoning Of Knowledge Graph Based On Reinforcement Learning

The utilization of knowledge graphs in finance,transportation,medical care,and consumption has been greatly expanded due to the emergence of knowledge graph-related technologies.However,the existing knowledge graph data collected by machines or humans usually has a lot of missing and abnormal knowledge.Data completion and correction.Most of the existing knowledge graph reasoning methods are based on deep neural networks and representation learning methods for reasoning at a fixed number of steps.These methods have a fixed number of reasoning steps and it is difficult to obtain the specific path of the reasoning process,and their interpretability is poor.The reasoning method based on reinforcement learning can obtain the path information in the reasoning process,and the meta-learning method has better performance on sparse data sets..By utilizing the meta-learning method for dynamic prediction and dynamic completion in the multi-hop reasoning process of the knowledge graph based on reinforcement learning,this thesis dynamically resolves the sparse problem of the knowledge graph reasoning process by completing the action space of the current entity in real time.And retain the path information in the reasoning process of the agent,so as to solve the problems existing in the above reasoning methods.Therefore,the research work of this thesis has extremely strong practical significance.The main work of this thesis is as follows:(1)A dynamic reward function construction method based on representation learning is proposed.This method builds a new reward function based on a pre-trained representation learning model,and the agent learns and optimizes the reasoning process according to dynamic rewards and judgment conditions.Compared with the existing knowledge graph multi-hop inference model,this method can infer the result within a non-fixed number of steps,and it is verified by experiments that the method has improved the inference performance.(2)This thesis proposes a multi-hop reasoning model of reinforcement learning knowledge graph,combining meta-learning with it.To address the sparse problem of the knowledge graph,pre-training is first done with meta-learning,then the potential action space in the current state is dynamically predicted,and finally the accessory action is dynamically added to the action space.This thesis affirms that the model surpasses the current knowledge graph multi-hop reasoning model in five sparse knowledge graph datasets,and further confirms that it is effective in tackling the sparse issue through ablation experiments.(3)Design and implement a knowledge graph multi-hop inference platform based on reinforcement learning in combination with practical applications.Through this platform,users can upload data to train and test the knowledge graph reasoning model,and use the trained model for knowledge graph completion Heavy task.