Research On Key Technologies Of Knowledge Graph Reasoning Based On Relational Path

Knowledge graphs,one of the latest ways of representing knowledge,enable users to better organize,manage and understand the vast amount of information on the Internet.Their importance is growing with the booming development of artificial intelligence technologies.Knowledge graph reasoning techniques are used to reason from existing knowledge to new,unknown knowledge.They incorporate the relational paths between entities to perform related predictive tasks,supporting applications such as semantic reasoning,natural language understanding and information retrieval.The related technologies have shown promising applications in various fields such as intelligent healthcare,financial security and cyberspace security.However,based on a systematic study of current work,this thesis finds that there are still some practical problems to be solved,mainly including broken reasoning chains in sparse knowledge graphs,short reasoning lengths in large knowledge graphs,and insufficient reasoning performance in low-latency application requirements.From the perspective of relational path reasoning,this thesis investigates the above three key problems from three aspects: dynamic path complementation,multi-agents hierarchical reasoning,and relational path generative reasoning,and provides an outlook on their application prospects.The thesis mainly completes the following work:1.An overview of current approaches on knowledge graph reasoning is summarized from the perspective of reasoning interpretability.During a long period of research and practice,researchers have successfully summarized different techniques for knowledge graph reasoning and tried to review the reasoning models from different perspectives(e.g.distributed representation,graph embedding perspective).However,there is a lack of relevant summaries and comparisons on the interpretability of reasoning models.In this thesis,we present the core ideas and evolution of different knowledge graph reasoning techniques,analyze their features and performance,compare the advantages and disadvantages of different techniques,and give a comprehensive evaluation of their efficiency and interpretability.2.A knowledge graph reasoning method based on the dynamic path completion strategy is proposed,which realizes the dynamic completion of local paths with global information and improves the ability to conduct reasoning in sparse knowledge graphs.Many domain knowledge graphs are sparse.When knowledge reasoning is performed in the corresponding graphs,the sparse graph environment causes many key reasoning paths to be missing,thus breaking the reasoning chain.In this thesis,based on existing research,we propose SparKGR,a knowledge graph reasoning method based on the dynamic path completion strategy.Spar KGR applies reinforcement learning to model the multi-hop reasoning as a sequential decision problem,and trains the reinforcement learning agent through feedback and interaction.When the reasoning paths are missing in the sparse knowledge graphs,the reinforcement learning agent dynamically completes the paths based on the current state and historical path information of the entities,thus expanding the reasoning paths and improving the model’s ability to reason in the sparse knowledge graphs.3.A hierarchical knowledge graph reasoning method based on multi-agent reinforcement learning is proposed,which achieves hierarchical reasoning of complex reasoning tasks and improve the ability to perform long-distance reasoning.In the process of path-based knowledge graph reasoning,a large number of paths need to be explored and the search space of reasoning models is generally comparable large,making it difficult to perform longdistance reasoning tasks.In this thesis,based on existing research,we propose a multi-agents reinforcement learning reasoning method RMLH,which divides the reasoning process into independent abstract relational subtasks by means of multi-agent reinforcement learning,and then performs fine-grained path exploration for each subtask to achieve a significant reduction in the search space.This hierarchical reasoning approach is more in line with the human divide and conquer way of thinking,and the reasoning results are more interpretable.4.A generative knowledge graph reasoning method based on a sequence model is proposed,which achieves parallel learning of relational path reasoning sequences and generatively outputs the next reasoning action,improving the efficiency and robustness of the model performance.Current path exploration-based methods generally need search a large path space during reasoning,causing the model take a long time to converge and the reasoning thus to be slow.At the same time,the absence of key reasoning paths can further affect the efficiency.In this thesis,building on existing research,we propose a generative knowledge graph reasoning method DT4KGR.Unlike previous approaches based on path exploration,this method applies the Decision Transformer model to simulate the interaction process of reinforcement learning,and learns the rewards,states,and action inputs in the historical reasoning sequence in parallel.The reasoning model generatively outputs the next reasoning action based on the learning of rewards,states,and action inputs in the historical reasoning sequence,which accelerates the efficiency and improves the robustness of the model,enabling faster and more efficient knowledge reasoning performance.