Research On Knowledge Graph Reasoning Based On Multi-information Representation Learning

With the rapid development of intelligent information service applications,knowledge graph gradually becomes the main form of Internet data knowledge service and is widely used and developed in the intelligent field.As there are a large number of implied relationships between entities in most of the current open knowledge graphs that have not been explored,the problem of incompleteness restricts the effectiveness of many downstream applications,and Knowledge Graph Completion becomes an important problem that needs urgent research.In contrast,knowledge graph inference,with fact prediction and relationship inference as the task,is the most important method for Knowledge Graph Completion and the key technology to support the downstream applications of knowledge graph.While early inference mainly relied on rules and logic,inference methods based on knowledge graph representation learning,driven by deep learning techniques,have become the main approach for knowledge graph inference research by virtue of their advantages in computational and storage efficiency.This thesis addresses the shortcomings of existing methods in structural information modeling and auxiliary information utilization,investigates how to effectively mine and utilize two types of information,namely,neighborhood structure and text description,and proposes an embedding and inference model with hierarchical perception of neighborhood information and text semantic enhancement,which effectively improves the effectiveness of knowledge graph inference.The main contributions of this thesis are as follows.(1)A representation learning and inference model based on hierarchical perception of neighborhood information is proposed.The model is based on the Encoder-Decoder framework,which differentiates attention weights by two levels of relations and entities to achieve finer-grained node embedding representation.effectively improves the accuracy and interpretability of inference prediction.(2)A representation learning and inference model based on semantic enhancement of text description is proposed.The framework uses pre-trained language models to encode textual input of entities and relations,and fine-tunes and trains them in stages by using KGE models,and verifies the applicability and effectiveness of the framework.(3)A joint inference model based on structure and text multi-information representation learning is proposed.An inference model combining two coding modules,neighborhood information and text description,is proposed to learn the triad structure information while complementing the knowledge representation learning with node neighborhood information and text description information together to achieve joint structure and textbased inference.