Research On Document-level Entity Relation Extraction Method Based On Deep Learning

As a crucial sub-task of information extraction,relation extraction aims to extract semantic relationships between entities from text.In recent years,with the rapid development of knowledge graphs and the urgent demand for knowledge graphs in the era of big data,the importance of relationship extraction as one of the necessary tasks for constructing knowledge graphs has also been increasingly valued.Early research focused on extracting relationships between entities from a single sentences,namely sentence-level entity relation extraction.In recent years,with the booming development of big data and advances in deep learning,the development of document-level entity relation extraction has been promoted.Compared to sentence-level entity relation extraction,documentlevel entity relation extraction requires extracting relationships between entities from multiple different sentences and considering more entities,thus requiring stronger information gathering capabilities.In documents with complex contexts,the implicit meaning between entities also needs to be taken into account.Therefore,models need to acquire more accurate global information and have powerful reasoning abilities,making it more challenging and difficult to extract entity relations from documents,leading to slow progress in document-level entity relation extraction work.To address these challenges,this thesis systematically studies deep learning-based methods for document-level entity relation extraction.The main research content includes the following points:(1)In order to solve the problem of insufficient local and global information obtained by existing methods,this thesis proposes an entity-relation(E-R)graph-based document-level entity relation extraction method(E-R graph model).This method explores the interaction between entities by constructing a new entity relation graph(E-R graph)and using graph convolutional networks to capture and update the information between entities.Two types of nodes and three types of edges are explicitly defined in the E-R graph,and sufficient local information is obtained through the interactions between relevant nodes.Effective connections between entities are established through edges to collect sufficient global information to strengthen global reasoning.Experimental results show that the proposed model performs well on the large-scale publicly available document-level entity relation extraction dataset Doc RED,and outperforms the compared models.(2)In order to further enhance node representation and address the problem of redundant inputs,this thesis proposes an IN＿E-R＿AT model that integrates attention mechanism and improved input for document-level entity relation extraction.This method improves the input document on the basis of the E-R graph model,reducing negative impacts caused by information redundancy while ensuring sufficient key information to predict entity relationships.In addition,by fusing attention mechanism to assign different weights to adjacent entities based on the importance of entity information.It helps to obtain information features with higher relevance between entities and enhances node representation,thus achieving more accurate document-level entity-relation extraction.Experimental results on the Doc RED dataset show that the IN＿E-R＿AT model has significantly improved performance compared to the E-R graph model.