Research On Entity And Relationship Joint Extraction Based On Graph Structure

With the rapid development of internet technology,a large number of unstructured texts have emerged.How to extract useful information from these massive unstructured text data has become a hot and difficult research topic in many fields.Therefore,based on the above background,information extraction technology has emerged.Named entity recognition and relation extraction are two main branches of information extraction,with the aim of extracting entities and relationships from unstructured text data.Entity relationship extraction provides important basic data for many fields,such as question answering robots,intelligent customer service,search engines,and so on.Although entity relationship extraction has achieved certain results through continuous research,there are still obvious problems,such as interaction separation,entity nesting,and relationship overlap.Therefore,based on the existing problems in entity relationship extraction,this article proposes two joint extraction models based on graph structure.The main research content and innovation are as follows:1.Aiming at the problem that the existing entity relationship model rarely considers the entity information and the relationship information between entities before extracting entities and relationships,which may lead to the problem that most of the extracted entities and relationships cannot form valid triples,this paper proposes a method based on different Entity-Relation Extraction Model(HD-RE)for Composition and Dual Decoding.First,the head entity is extracted first.Specifically,in order to solve the entity nesting problem and enhance the prior knowledge of tail entity and relationship extraction,we use the form-filling method to parse out the head entity in the output word embedding of the Bert pre-training model.Then,in order to enhance the semantic representation of words and relations,we model the words,predefined relations and entities in the text sequence as nodes on the graph,and realize the information transfer and information fusion between nodes through the graph attention network model.Finally,the fused features are fed into a binary tagger to parse out tail entities under specific relations.The experimental results show that the F1 value of the HD-RE model is improved by 2.83% and 5.54% compared with the baseline model on the two public datasets of NYT and Web NLG,respectively.2.Aiming at the problem of information loss caused by entity nesting and pruning operations in entity relationship extraction,this paper proposes a joint entity relationship extraction model(SGNet)based on soft pruning and joint decoding.First,in order to obtain the word vector representation of rich context information,we use the Bert pre-trained model as the word embedding layer;second,in order to solve the problem of information loss,this paper proposes a graph module that can extract local and non-local features of graph nodes.and load it into the feature extraction layer.Specifically,we utilize a Gaussian graph generator and a multi-head attention mechanism to construct a fully connected graph,then adopt GCN and a densely connected layer to realize node information transfer and information aggregation;finally,we use a global pointer decoder to convert triplets to quintuples Group extraction to address overlapping relations and entity nesting.Experiments show that the F1 value of the SGNet model is 3.69% and 3.98% higher than the baseline model on the two public datasets of NYT and Web NLG,respectively.And compared with mainstream models,it has better performance.3.The above research results are applied to the Chinese medical text dataset to extract the entity relationship in the Chinese medical text,so as to promote the practical application of this method in relevant fields.