Research On Document-Level Relationship Extraction Method Based On Deep Learning

Document-level relation extraction is a natural language processing technology,which aims to automatically extract the relationship between entities from large-scale text.Different from traditional sentence based relation extraction,document-level relation extraction needs to consider the context and Semantic information of the whole document.In the document-level relationship extraction method based on deep learning,the document-level relationship extraction model based on graph convolutional network is mainly used.In the document-level relationship extraction model based on graph convolutional networks,there is a problem of how to deeply mine cross sentence information: on the one hand,how to obtain document-level article codes containing contextual entity information,and on the other hand,how to use article structure to infer entity relationships.In response to the above issues,this article has conducted the following research:1.We have designed a document-level encoder that can mine contextual entity information.In traditional document-level encoders,a new attention mechanism called inter sentence attention mechanism has been proposed to better mine contextual information.By focusing on the information in the previous sentence while also paying attention to all the information in the previous text,this principle enables the final expression of the article to contain relatively rich cross sentence information,improve the correlation ability between cross sentences,and further explore the information between sentences.This document-level encoder is suitable for graph convolutional neural networks,and the inter sentence attention mechanism is also suitable for traditional recurrent neural networks.2.A new dynamic graph convolutional neural network model is proposed.In existing document-level relationship extraction models based on graph convolutional networks,graph construction methods mostly use static graphs.In the document-level relationship extraction model of a few dynamic graphs,the edges between nodes are obtained through node representation,so all elements in the adjacency matrix used for graph neural networks are changed.In the new dynamic graph convolutional neural network model,a new composition method is used,which can deeply explore the relationships between certain nodes while preserving the structural information of the original article.A new dynamic edge generation strategy has been designed,similar in principle to article similarity calculation.By comparing the expressions between two nodes,it is determined whether there is a relationship between the two nodes.Enhance the reasoning ability between entities across sentences through a combination of dynamic and static edge composition and dynamic graph update strategies.In summary,this article conducted multiple sets of experiments on three publicly available datasets: Doc RED,CDR,and GDA.Through comparative experiments and ablation experiments,as well as specific analysis of two dimensions in a case,the above work has been validated to effectively mine cross sentence information.