Research On Entity Relation Extraction Based On Graph Convolutional Network

With the rapid development of the Internet,information on the Web has also exploded,this information contains a large number of knowledge,so how to obtain the information we need from this unstructured text data has become the focus of researchers' attention.The purpose of entity-relation extraction is to accurately extract the entities and their semantic relations that people need from the massive amount of text on the Internet,and these entities and relations have a wide range of applications in many sub-tasks in natural language processing.At present,the mainstream method of relation extraction is based on deep learning.They use recurrent neural network or convolutional neural network as feature extractor to automatically learn text features.Compared with traditional relation extraction methods,they save complex feature engineering and greatly improve the ability of relation extraction.However,there are still two problems: 1.the syntactic structure information is not fully utilized in the feature representation;2.there are entity overlapping relations in some sentences,and many models do not consider the extraction of these relations.The main contributions made in this paper to investigate the above two problems are as follows:(1)To make fuller use of the information in the dependency tree,a graph convolutional relation extraction model based on dependency graph is proposed,which captures the syntactic structure information in the dependency tree by using the Graph Sage algorithm,then makes the relation classification more focused on entity nodes by adding entity labeling auxiliary tasks,and finally uses the BERT aggregation function for node information aggregation,incorporating the word node aggregation features in the semantic representation of words.Experimentally,the model in this chapter is compared with several advanced relation extraction models,and the results show that in TACRED dataset DGGCN achieves an F1 value of 67.1%,which is ahead of the other comparison models.(2)In order to extract entity overlapping relations better,a graph convolutional relation extraction model based on trigger word mechanism is proposed.The method uses a bidirectional graph convolutional network to model the dependency tree so that the node features are enriched with syntactic structure information.In order to extract the triplets more completely,the relation extraction is carried out in two stages: the first stage predicts the relations of all entity pairs;the second stage constructs an entityrelation weighted graph under each relations,further considers the connection between entities and relations,finally combine the relation trigger words and entity pair features to predict the relations again,integrating the results of the two stages thus improving the extraction ability of overlapping relations.Through comparative experiments,the effectiveness of TWGCN in entity overlapping relation extraction is verified,and the influence of relation trigger words on the extraction results is also verified.(3)A relation extraction system based on graph neural network is constructed,which provides entity recognition,relation extraction,dependency parsing,model training,model configuration and other functions.Django is used to develop the web application background,Bootstrap is used to make the foreground page,and Docker is used to package the whole service environment.The performance test shows that the system has certain concurrent access ability.