Distant Supervision For Relation Extraction Based On Deep Learning

With the development of the information age,the scale of the text data stored and used by people on the Internet has increased dramatically.For discover the target information from massive data,information extraction(IE)technology emerges as the times require.As an important research topic in the field of information extraction,the fundamental purpose of research on relation extraction is to discover the semantic relation between entities from semi-structured text or unstructured text,so that people can quickly understand hidden information in complex texts.Therefore this work is of great significance and worthy of researching.This thesis mainly studies the implementation of the relation extraction task on the distant supervision data set.Proposed a new method of relation extraction based on deep learning technology and distant supervision.According to the features of the distant supervision data and the existing research on distant supervision-relation extraction,this thesis carried out distant-supervision relation extraction from relation features extraction and suppressing the impact of noise data.In relation features extraction,the deep learning model usually only contains shallow convolutional neural network structure,and the designed model extract the shallow semantic information,therefore this thesis focus the network depth,make the model to learn more relation features via introducing the Deep Residual Learning(ResNet)to extract more abstract relation features,and designs the multi-layer network model experiments to explore the better model depth;in suppressing the impact of noise data,existing research focuses on introducing information such as syntactic structure and requires a large amount of expert knowledge,but from the perspective of the importance of sentence information to the target relation,Attention Mechanism is introduced at the pooling level to capture more feature information which we want to reduce the impact of noise information,The results were verified by the Attention Pooling model experiments.Finally,a new model structure is proposed based on the advantages of the two algorithms.In summary,this thesis designed a new deep learning model ARCNN based on combining the Attention Mechanism and the Deep Residual Learning.Compared with several classical distant-supervision relation extraction models,the results show that Our model can effectively mitigate the noise impact of distant supervision data,learn more abstract relation features,and discover entity semantic relations in complex text more accurately.