Research On Document Oriented Entity Linking Method

Entity linking aims to link entity mentions in text to corresponding entities in knowledge base,which helps semantic understanding of text and is widely used in information extraction tasks.As one of the basic topics in natural language processing,entity linking actually involves disambiguation of entities in text,and has significance in both application and research.Entity linking methods based on deep neural network have made great progress,but the accuracy needs to be improved further.This thesis focuses on entity representation learning and global linking algorithm.The research results are summarized as follows.(1)This thesis proposes an entity representation learning method integrating finegrained type information.First,the entities in the corpus are annotated with fine-grained type tags by existing annotation tool and then used as prediction objects for pre-training the word vectors in skip-gram model.Second,the entity representation is learned with max-margin algorithm by making it close to the vectors of surrounding words and away from others.The validation results applied to two typical entity linking models show that the average F1 score of linking accuracy on five public test sets are improved by 0.82%and 0.42% respectively.(2)This thesis designs and implements a global entity linking model based on knowledge graph.Since there is relevance between entities in the same document,we consider to use large-scale entity triples in knowledge graph for constructing associations between candidate entities.Therefore,we propose a global entity linking model based on knowledge graph.We first design an entity global representation learning module based on graph network model,where nodes and edges represent candidate entities in the same document and entity associations from knowledge graph,respectively.And we introduce graph attention network to update each entity representation with features from other nodes.Secondly,we design a global mention representation learning module based on convolutional neural network which extracts context features and fuses other mentions’ features with the current mention representation.Finally,we design a linking decision method based on the similarity between each mention and its candidate entities.The evaluation results on five open test sets show 0.39% improvement of the average F1 score,with three of the test sets exceeding the previous work.