Research On Document-level Relation Extraction With Graph Convolutional Networks

With the development of communication technology,all kinds of data are growing explosively.News,newspapers and other unstructured long texts contain a lot of valuable information.How to extract this information automatically has become an urgent problem.As one of the key tasks of information extraction,document-level relation extraction aims to extract intra-sentence and inter-sentence relations among many entities in long text.In this thesis,we do the following research on the task of document level relation extraction.(1)Document-level relation extraction based on contextual semanticsExploring the method of document-level context semantic extraction and its impact on performance of document-level relation extraction.Firstly,the entity interaction graph is constructed with the entity as the node.Then,based on three graph convolutional networks,the context semantics are encoded,the document-level information related to entities is captured,and the relation is extracted.Experiments on Doc RED dataset show that,based on entity interaction graph and graph convolutional network,context semantic information in long text can be effectively captured to model complex interaction relationship between entities.Using multi-head attention mechanism and dense connection technology can further improve the performance of document-level relation extraction model.(2)Document-level relation extraction based on global context enhancementSentences,where the entity pair located,contain local information related to the entity pair,while the whole document contains global information of the entity pair.This thesis proposes a document-level relation extraction method based on global context enhancement.First,the entity interaction graph is constructed.Then,the attention mechanism and gating mechanism are used to fuse all the sentence information of the entity to obtain the entity-aware edge representation.Furthermore,the hierarchical graph convolutional reasoning networks are used to capture rich local and global information of entities.Finally,the entity representation is obtained by fusing the two parts of information,which is used to infer document-level relation of entities.Experiments on Doc RED dataset show that entity-aware edge representation can integrate entity related context information into node representation,and hierarchical graph convolutional reasoning networks can capture rich local and global information to obtain high-quality document-level relation reasoning model.(3)Document-level relation extraction with reasoning pathPaths between entity pairs contain rich information for relation reasoning,which can provide explicit reasoning guidance for document-level relation extraction.Firstly,paths between entities in the entity interaction graph are obtained,and multiple paths of node pairs are constructed based on node representation,edge representation and other information.Then,based on the attention mechanism,the path representations of entity pairs are obtained by fusing multiple paths,which helps to infer the entity relation.Experiments on Doc RED dataset show that reasoning path representation learning method based on attention mechanism can capture relevant reasoning information of entity pairs.Based on entity representation,reasoning path representation is introduced to improve the performance of document-level relation extraction.The research in this thesis has achieved advanced performance in document-level relation extraction task.Entity interaction graph can effectively model the complex interaction relations between entities.Global context enhanced relation extraction method can capture abundant local and global information to obtain high-quality document-level relation reasoning model.Reasoning path representations of entity pairs are introduced into the relation extraction method with reasoning paths,which further improves the performance of document-level relation extraction.