Long-text Entity Relation Extraction Based On Graph Neural Network

With the rapid development of Internet technology,people send and receive massive information on the Internet every day.However,information has exposed in an explosive growth rate,which seriously affects people's ability to identify and filter out the important information that they need,and thus the efficiency of people's retrieving and understanding information has been reduced.Therefore,how to lead the machine to help people locate and extract that key information from the massive data is of great significance to the development of today society.And thus,information extraction technology has become a hot research topic.As an important subfield of information extraction,entity relation extraction aims to automatically identify potential semantic relationships between target entities based on their related unstructured data.Most of previous methods focus on extract entity relation from the context of short texts.But in the real world,texts are often formed in the long-length way,such as novels,news,diaries,etc.Those methods based on short-text modeling are difficult to handle long texts.To address these issues,this study investigates the characteristics of long texts,and introduces the technology of graph neural network,which transforms texts into graphs for modeling data.This study takes the advantages of graph neural network models to automatically identify the semantic information of words which are closely related to the target entity pair,and devise a framework combining a variational autoencoder and a label-enhanced attention mechanism to extract entity relation.It utilizes the variational autoencoder to project entity information and contextual word information into a unified multivariate Gaussian distribution space.Besides,an alternative measurement is designed to quantify the semantic similarity between the couple words and the targeted entity pair,which is used as the edge weight between words in the graph.Then,A multiple Convolutional graph neural network model is applied to extract the semantic features related to the targeted entity pair.In addition,this study introduces a graph pooling network incorporating the label-enhanced attention mechanism,to capture those entity relation indicators that is correlated to the targeted entity relation,thereby improving performance of the whole model.The experimental results on the public dialogue dataset(Dialogue RE)show that the proposed method effectively improves the performance compared to other benchmark methods.The F1 values of the validation dataset and test dataset are increased by 1.7% and 2.1%,respectively.Considering the large number of entities in long texts and their complex connections,this study proposes a method which joints extraction of multiple entity pairs base on a heterogeneous graph network and incorporates an adaptive classification framework base on the relation-imbalanced data distribution.The method mainly constructs a global heterogeneous graph neural network encoding module,a local entity encoding module,and a context-based entity encoding module,so as to mine the semantic representation of entities and the semantic relations between entities from multiple perspectives.Furthermore,this study focuses on the problems of unbalanced categories distribution cause by too many none relation type in entity relation extraction tasks,and designs a learning framework that combines threshold adaptation and unbalances classification adjustment,which effectively promotes the model to distinguish the difference between targeted entity relation type and none relation type.The experimental results on the Wikipedia document data set(Doc RED)show that compared with other benchmark models,the F1 value and Ign F1 value of the test set are improved by 1.19% and 1.22% respectively.