Research On Document-level Relation Extraction Technology

Information in daily life is mostly presented in the form of documents,and it is more practical to mine the relationship between entity pairs in documents.However,most of the existing relationship extraction methods are sentence-level,and it is difficult to capture the relationship between entities that are far apart in the text.In order to make up for this,document-level relationship extraction technology has emerged.This article intends to study the document-level relationship extraction technology from the following three aspects:(1)Research on document-level relation extraction method based on sequence.The sequence-based document-level relationship extraction method can be regarded as an extension of the sentence-level relationship extraction method.In this part,this article explores two specific models,namely a pre-trained model using remote supervision data sets and a Transformer-based coding model.And proposed improved models for these two models.(2)Research on document-level relation extraction method based on graph.The graph-based document-level relationship extraction method mainly converts the information of the document from the form of the sequence to the form of the graph,and then uses the graph convolutional neural network to complete the extraction of text information.This article also explores two specific models in this part,namely,the implicit reasoning model using graph convolutional neural network and the graph reasoning model based on the separate composition of the mentioned entities.This part of the research has laid a theoretical foundation for the optimal model of this article.(3)Research on the document-level relationship extraction method of fusion sequence and graph.This part of the research mainly integrates the document-level relationship extraction method based on sequence and the document-level relationship extraction method based on graph.The specific method is to use the well-performing modules in the sequence-based document-level relationship extraction model to add or replace part of the structure in the graph-based document-level relationship extraction model.In this part of the research,this paper proposes the optimal model,and compares the effects of this model with all models in this paper,analyzes the advantages and disadvantages of different models,and then draws the final conclusion.