Research And Application Of Distantly Supervised Relation Extraction Using Bag-Based Graph Neural Network

The rapid development of information technology leads to the explosive growth of Internet data.Information extraction technology could obtain the key from redundant information.Relation extraction,one of the fundamental topics in the community of information extraction,aims to identify the relation between entities.It is a critical step in knowledge graph construction.Traditional supervised relation extraction requires a largescale high-quality human-labeled data.Different from it,distantly supervised relation extraction could obtain a large-scale training data by heuristically labeling plain text with a knowledge base.This method could save human resources and time consumption,and can be adapted to different fields based on different knowledge graphs.Therefore,distantly supervised relation extraction could be widely applicated.However,the labeling way of distant supervision causes wrong labeling and long-tail problem.To alleviate above problems,we propose a hierarchical graph convolutional networks based on bag structure for distantly supervised relation extraction.The main research and contribution of this paper are as follows:1)A Local to Global Graph Convolutional Network framework,i.e.,L2G-GCN,is proposed for distantly supervised relation extraction.The framework improves the performance of distantly supervised relation extraction by first locally learning syntactic knowledge in a single instance,and then globally aggregating semantic correlation information among instances within the bag structure.Specifically,to strengthen the model’s ability of handling long texts,the dependency parsing is incorporated to construct a word-level graph convolutional network(Word-GCN).To alleviate the noise problem,a sentence-level graph convolutional network(Sen-GCN)with self-attention mechanism based on bag structure is proposed.To solve the long-tail problem,a regularization term of mutual information maximization is incorporated to constrain the model weights.2)A Heterogeneous Graph Convolutional Network based on BERT,i.e.,BH-GCN,is proposed for distantly supervised relation extraction.In order to enhance the guiding role of entity information for relation extraction,we add the processing of entity on the basis of L2G-GCN framework and propose BH-GCN.Specifically,we construct a heterogeneous graph with instances and entities as nodes,and the graph updates the nodes by self-attention mechanism.Meanwhile,according to different edge types,the model could interrelate the nodes selectively.Besides,we design a gated mechanism for the fusion of entity and instance information,which enhances the bag representations.In addition,in order to introduce common-sense knowledge,syntactic and semantic information,we adopt a pre-trained model as an encoder.And the effectiveness of the pre-trained model is verified for distantly supervised relation extraction.3)Design and implement a relation extraction system based on distantly supervised.The system contains user management module,data management module,relation extraction module and web design module.And the system supports user-defined numbers of input texts,relation extraction in bag-level,and visualization of the related results.