The Design And Implementation Of Knowledge Graph-Oriented Entity Linking Method

With the rapid development of information technology such as the mobile Internet,massive amounts of unorganized text data have been generated on the network,which contains a large number of entities and knowledge.Due to the complexity and variety of natural language,it is difficult for computers to understand and utilize this knowledge.At the same time,in order to effectively organize and manage large-scale knowledge in network data,knowledge graphs emerge as the times require.Knowledge graphs use graph-structured data models to describe objective fact concepts and their relationships.Knowledge graph-oriented entity linking builds a bridge between sequential text and structured knowledge,and is the technical support for knowledge graph expansion and its downstream applications.In recent years,it has gradually become a research hotspot in academia and industry.Therefore,this thesis conducts an in-depth exploration of knowledge graph-oriented entity linking.Existing methods do not make full use of entity referent context and entity features,ignoring the knowledge graph topology and decisionmaking dependencies between entities,especially for short texts with sparse features,the entity linking accuracy is poor.To solve the above problems,this thesis proposes a two-stage entity linking framework SR-EL that fuses relational features to enhance existing methods based on deep learning.On the one hand,the framework can extract deep interaction features between referent context and entity relations,thereby improving the spatial consistency between referent context and entity embedding.On the other hand,by virtue of long-term dependencies among multiple entities,the framework accumulates knowledge of previously linked entities as dynamic context to assist the linking decisions of related entities and further enhance the linking performance of the model.Further,in order to alleviate the model popularity bias problem of existing methods based on pre-trained models,this thesis proposes an adversarial attack-based entity linking optimization method PA-EL.By dynamically decoupling the popularity information in the entity representation,the method enables the link model to make equal link decisions for entities with different popularity,thereby reducing the influence of the model from the pretraining corpus distribution.Experiments on public benchmark datasets show that,compared with existing methods,the proposed method SR-EL has improved the entity link accuracy,and the method PA-EL proposed in this thesis has achieved a significant improvement in the accuracy of tail entity links.Furthermore,this thesis implements a knowledge graph-oriented entity linking system and applies the proposed method to the system.By visualizing the results of entity link analysis,the system enables users to intuitively observe part of the reasoning basis of the algorithm,which provides a certain degree of interpretability for the algorithm.