Research On Key Technologies Of Knowledge Graph Reasoning Based On Deep Reinforcement Learning

Knowledge graph is an efficient semantic network structure,which has been widely used in the fields of natural language understanding,recommendation system,intelligent question answering,medical education,etc.A large amount of triplet information in knowledge graph has not been discovered,leading to problems such as structural deficiencies and semantic confusion in knowledge graph,which further seriously affects the performance of upstream tasks.How to excavate the potential connections between entities and discover the path rules in knowledge graph is an urgent problem in knowledge graph field.To solve the above problems,thesis studies the knowledge reasoning method based on deep reinforcement learning,aiming to complete the knowledge graph through reinforcement learning technology.Firstly,thesis conducts a comprehensive study on related research techniques of knowledge graph reasoning,proposes a path reasoning model that integrates representation learning and rule learning,and further analyzes the shortcomings of heuristic rewards in reinforcement learning,proposes a knowledge graph reward reasoning model based on inverse reinforcement learning,and extends the path reasoning model to uncertain knowledge graph,based on this,an uncertain knowledge graph reasoning method based on causal inference is proposed.The main research work of thesis is as follows:(1)A path reasoning model combining representation learning and rule learning is proposed.In order to solve the problems of sparse reward and large action space of reinforcement learning model in knowledge graph reasoning task,a reward function optimized by representation learning and rule learning is proposed,and an agent architecture including relational policy network and entity policy network is designed,and a variety of optimization mechanisms are introduced.Finally,the effectiveness of the proposed model is proved by experimental results.(2)A knowledge graph reward reasoning model based on inverse reinforcement learning is proposed.In order to solve the problems of low reliability and inaccurate rewards caused by heuristic rewards in the path reasoning model,an adaptive rule reward learning mechanism based on agent trajectory updating and an iterative hit reward learning mechanism based on agent strategy were designed from the aspects of rule reward and hit reward,and a reward dropout mechanism was designed.Finally,the effectiveness of the proposed model is proved by experimental results.(3)An uncertain knowledge graph inference model based on causal inference is proposed.The conventional path inference model is not suitable for the uncertain knowledge graph.Therefore,an inference agent is designed to capture the path uncertainty and solve the uncertainty ambiguity caused by multiple paths through the causal inference framework.