Research On Relation Extraction Technology Based On Neural Network

With the advent of the information age,more and more data and information fill our lives.Relation extraction is an important task in the knowledge graph.It can extract structured information from unstructured free text that can be stored in the knowledge base.This paper studies the relation extraction technology based on neural network,and proposes two effective methods for relation extraction based on neural network in view of the existing problems in existing research.First of all,in view of the problem that the distant supervision brings noise and affects the accuracy of the extraction results,a relation extraction method based on the joint coding of BiLSTM and the hole convolutional neural network is designed.This method performs denoising work from the perspective of joint extraction of features.First use the pre-trained word vector model to process the sentence,convert the word into a vector representation,as the input of the model.Then input it into BiLSTM to extract the dependency information of the sentence,and then use the hollow convolutional neural network to extract the semantic unit information of the sentence.On this basis,the convolutional neural network is used to fuse the obtained feature vectors,and then encode them to obtain feature vectors that contain both sentence-dependent information and semantic unit information,so that the feature vector contains more semantic information of the sentence and improves the extraction result.accuracy.At the same time,we use sentence-level selective attention mechanism to assign weights to sentences to further reduce the impact of noise caused by distant supervision.Finally,the effectiveness of the proposed method is verified by experiments.Secondly,aiming at the problem of overlapping triples in relation extraction,a sequence-to-sequence relation extraction method based on the fusion of convolutional neural network and Transformer is designed.The existing mainstream relation extraction models can only extract common types of triples,and there is less research on the problem of overlapping multiple triples in a sentence.This paper designs a sequence-to-sequence relation extraction model,integrates the convolutional neural network into the Transformer as the encoder of the sentence,extracts the features of the sentence,decodes it,and recognizes all possible subjects in the sentence and each subject Corresponding relations and objects,thereby obtaining triples of different overlapping types in the sentence.In the experimental part,we tested the designed model in two public data sets and proved the effectiveness of the model.