Research On Coreference Resolution Oriented To Knowledge Graph

In the last few years,what has come with this is the huge amount of data on the web,that has existed every corner of our lives,it has become more and more difficult to find the data you need on the browsers.The emergence of knowledge graphs has solved this problem.The knowledge graph is a knowledge base that expresses the concepts and entities in the objective world and the relationships between them in the form of graphs.Building a knowledge graph is mainly divided into four parts,including knowledge extraction,knowledge fusion(entity alignment),data model construction,and knowledge reasoning.As a vital part of entity link,coreference resolution plays an important role in constructing knowledge graph.Coreference resolution is a technology that can merge different depictions in the knowledge graph.Now coreferential resolution methods include rule-based,machine learning,and deep learning-based methods.There are some problems with coreference resolution methods.First of all,most traditional coreference resolution models make use of the grammar,syntactic structure or classification and clustering methods of the sentence itself,and do not involve the semantics in the article;secondly,the traditional coreference resolution models cannot take the global knowledge and global characteristics of the context into account in the document;and most of the common coreference resolution models only can be applied to a certain field,and the generalization ability is not strong.In addition,this paper also tries to apply various deep learning models to solve the problem of coreference resolution.Coreference resolution is a technology that can merge different descriptions in the knowledge graph.Traditional coreference resolution methods include rule-based,machine learning,global optimization,and knowledge-based methods.Most traditional coreference resolution methods take advantage of the sentence's proper grammar,syntactic structure,classification and clustering methods.It involves the semantics and prior knowledge in the article.In order to solve above problems,this paper combined with the deep learning theory and natural language processing technology,adopted based on end-to-end total digestion method,not only to maximize the use of the semantic knowledge and prior knowledge in the article,solve the problem of global knowledge and the shortage of global features,and improved the generalization ability of total refers to dissolve,make the model effective,The model is optimized from the aspects of features and computational cost respectively.The model fully considers the local features and global features in the document,and constructs a fine and accurate coreferential chain.The effects of feature addition and computational performance optimization on the model were discussed through parameter discussion and multi-group experiments.Then,through changing the model structure,this paper proposes an end-to-end coreference resolution model based on Transformer,which substitutes the encoding part of the end-to-end coreference resolution model span with Transformer for the experiment.The experiment proves that this model is compared with the baseline model and the more classic current coreference resolution model.Both efficiency and performance are improved.