Entity Linking Based On Graph And Deep Learning

Nowadays,with the rapid development of Internet,improving the text understanding ability of computers is of great value to all aspects of Natural Language Processing application.Entity linking is the task of helping to improve text understanding ability.There are two main steps in entity linking tasks.First,we identify all the ambiguous entity named mention,and then link mention to the appropriate location in the knowledge base.In this paper,two kinds of entity linking algorithms are proposed on the basis of relevant research in the domain of entity links,and on the basis of existing research methods at home and abroad,including named entity relation model based on graph(GECM)and an entity linking algorithm combined with graph and deep learning.In the relational degree model of named entity based on graph,the contribution of this paper is to put forward a new method of constructing graph model,can make full use of the existing Wikipedia knowledge,Linking the link relationship between candidate entities and the theme of the candidate entities to improve the accuracy of the global correlation of the entity.And put forward a new algorithm of Semantic Relevancy between mention and candidate entity,calculating the semantic similarity between the wiki category of the candidate entity and the wiki concept in the extracted ambiguous text,and further extract the semantic relationship between the mention.and candidate entity.On this basis,we incorporate a variety of features into the graph model,and design a model of relational degree of named entity based on graph.Experimental results show that the proposed graph based named entity correlation model is more effective than the traditional graph based approach.And we propose an entity linking algorithm combining graph and deep learning.The Mention-context model is designed by using BiLSTM and the Entity model is designed by using CNN.The two models can get the semantic features of mention from the context,and get the semantic features of entity through the description of candidate entities' category labels and Wiki concepts.The relationship between the mention and the candidate entities is predicted by combining the features of the candidate entities in the graph model and using the multi-layer neural network.Experimental results show that the proposed method combining depth learning and graph model is better than using graph based entity link algorithm.