Research And Implementation Of Entity Linking Algorithm For Web Encyclopedia And Knowledge Graph

With the rapid development of big data,massive amounts of natural language text data have appeared on the Internet.Natural language text data contains a large amount of knowledge,whose organization is not fixed.So,it's difficult to accurately mine this kind of knowledge with traditional methods.Knowledge base is a tool for organizing,managing and utilizing knowledge in a certain way,which stores a large number of named entities and related knowledge,mainly represented by the network encyclopedia and knowledge graph.Natural language texts often contain a large number of proper nouns,which play a key role in understanding sentences.These proper nouns are called entity mentions.Entity linking is the process of finding the knowledge base entities corresponding to entity mentions.Through entity linking,a large amount of knowledge in natural language text can be fully mined,and downstream applications such as semantic search and intelligent question answering can be well supported.According to the different types of knowledge bases,entity linking tasks can be divided into entity linking for web encyclopedias and entity linking for knowledge graphs.Encyclopedia-oriented entity linking is the process of linking entity mentions in natural language texts to the corresponding entities of the encyclopedia.Knowledge-graph-oriented entity linking is the process of linking entity mentions in natural language texts to the corresponding knowledge graph entities.Aiming at the shortcomings of existing research,this thesis has conducted in-depth research on two different entity linking tasks,and achieved the following research results:(1)In encyclopedia-oriented entity linking,the existing papers ignore the few-shot problem of some mentions,resulting in insufficient training of deep learning models in these methods,which affects the accuracy of these methods.This paper proposes to divide the entity link task into a large number of sub-tasks according to different entity mentions,and then proposes a meta-learning-based entity linking method(Meta-EL)for encyclopedia-oriented entity linking,which uses meta-learning algorithm to solve the few-shot problem of some mentions.In this method,in order to solve the class imbalance problem of some sub-tasks,this paper proposes an adaptive coefficient mechanism based on the set encoder to dynamically adjust the training process of different tasks.The results of experimental evaluation on multiple real data sets show that Meta-EL has significantly improved the F1-score compared with the existing work.(2)In knowledge-graph-oriented entity linking,existing papers ignore the semantic dependencies contained in the context of entity mention.Also,exisiting papers does not fully mine the semantic information of the knowledge graph entities,which affects the accuracy of these methods.This paper proposes a knowledge-graph-oriented entity linking algorithm DPGAT that combines semantic dependency parsing and graph attention networks.Aiming at the syntactic features in the context of entity mention,the DPGAT algorithm use semantic dependency parsing to construct a dependency tree,and then uses graph convolutional networks to encode the semantic dependencies among them.For the semantic information of the knowledge graph entity,DPGAT introduces attention mechanisms in the representation learning of the knowledge graph entity,and selectively uses the semantic information in the knowledge graph.The experimental results on multiple real data sets show that,compared with the existing work,DPGAT significantly improves the F1-score of the knowledge-graph-oriented entity linking.