Research On Entity Disambiguation Method Of Knowledge Graph Based On Deep Learning

Entity Disambiguation technology is a critical technique for constructing and expanding knowledge graphs,and an important supporting technology for semantic retrieval systems,question answering systems,recommendation systems,etc.The goal of Entity Disambiguation is to map the mentions mentioned in the text to the corresponding entities in a given knowledge base or knowledge graph.However,there are several drawbacks in the current Entity Disambiguation technology: 1)Existing methods mostly express the contextual features of entities in a sentence form.But the sentence lacks the relationship information of the candidate entity,accordingly,making the contextual information expressed in this form inaccurate.2)The existing methods don't make full use of the semantic correlation between the entities corresponding to the mentions in the same text,making the disambiguation ineffective when there is less entity context information.For solving the above problems,this thesis uses the knowledge graph as the link base and proposes a local entity disambiguation method based on deep learning and a global entity disambiguation method based on deep learning.The main content of this thesis is as follows:(1)Embedded expression.The entity disambiguation method based on deep learning requires that the input of its deep network is a real-valued vector.Therefore,it is necessary to transform the related information of mentions in the text data and of entities in the knowledge graph into the corresponding embedding expression vector.(2)Due to the inaccuracy of semantic information of entities and mentions in existing methods,this thesis proposes a local entity disambiguation method based on deep learning.The method combines the mentions and mentions' contextual information and entities and entities' local feature graph.The working process of the method is as follows: To begin with,the attention mechanism-based convolutional neural network is used to learn the semantic representation of the mentions' related information in the text.Then,the graph attention network is used to learn the semantic representation of the related information of entities in the knowledge graph.Finally,the similarity between the mentions and entities is calculated,and then combined with the entity popularity to train the model.In the prediction,the candidate entity with the highest ranking is taken as the target entity.(3)Aiming at the problem that the local entity disambiguation does not make full use of the semantic relevance of mentions in a text,a global entity disambiguation method based on deep learning is proposed on the basis of the local entity disambiguation method.The global entity disambiguation method combines local information and global information.The working process of the method is as follows: First,the global feature map of the candidate entity is inputted into the improved graph attention network to obtain the global information score of the candidate entity.And then the global information score and the local information score are combined to train the model.In the prediction,the candidate entity with the highest ranking is also taken as the target entity.The local entity disambiguation and global entity disambiguation models are trained on the AIDA-train dataset and tested on AIDA-B and other datasets.The test results show that the two models proposed in the thesis are better than the comparison method in the disambiguation effect,and the global entity disambiguation model is better than the local entity disambiguation effect.