Entity Linking Research Based On Semantic Representation And Graph Regularization

Named entity linking is an important task in the field of natural language processing. As an important part of natural language, Named Entity has the characteristics of ambiguity, diversity and lack of standard, which has brought great trouble. Entity linking technology can link the mention of entity to the corresponding entity concept in the knowledge base precisely, solve problems caused by the above characteristics.On the basis of studying the related research of entity linking, the following research was carried out in this paper.(1) Construction of multi-source knowledge base. After referencing a variety of encyclopedia resources, we established knowledge bases of synonym and ambiguous words to support the generation of candidate entities; and established knowledge bases of entity popularity, entity relations, context entities and context words to support the disambiguation of candidate entities, and finally determined the exact target entity.(2) Learning entity semantic representation. Based on the DNN technology, we constructed an entity semantic relationship model, selected the related entities, entity types, relation types and entity descriptions as features, utilized encyclopedia entity linking results as training data, learned entity semantic representation.(3) Entity disambiguation based on graph regularization. According to the context of mention, we constructed the entity relationship diagram, initialized the weights of vertices and edges, obtained a stable score of every vertice by optimizing the loss function and selected the entity with highest score as the target entity. In the end, experiments were carried out on the entity link data sets of NLPCC2013 and NLPCC2014 respectively. Results showed that the accuracy of entity linking based on semantic representation and graph regularization, which reached more than 90%, was better than the traditional methods.