Research On Semi-supervised Text Classification Method Based On Deep Learning

On the labeled datasets with correct labeling and sufficient quantity,supervised classification models can often achieve better classification results.However,there is limited labeled data in practical application,the collection of training data is very cumbersome,and the work of labeling data is time-consuming,laborious and expensive.So in the case of lack of labeled data,how to use limited labeled data and a large amount of unlabeled data to train models with stronger generalization ability is a crucial problem.Therefore,this paper uses a large amount of cheap unlabeled data and a small amount of previously labeled data,combined with the teacher-student model and semi-supervised learning,and based on the consistency training framework to carry out the research on the semi-supervised deep learning method for text classification.In view of the above problems,this paper carries out relevant research.The main research contents are as follows:（1）Based on the wordMixup method,this paper proposes the u-wordMixup method for data augmentation of unlabeled data,and a semi-supervised deep learning model based on u-wordMixup（SD-uwM）by combining the consistent training framework and Mean Teacher model.①The u-wordMixup data augmentation method augments the unlabeled data,constructs the consistency hypothesis,and aims to reduce the unsupervised consistency loss to constrain the quality of the augmented unlabeled training data,which can reduce the over fitting of the model;②Combined with supervised cross entropy loss and unsupervised consistency loss,a new objective loss function is constructed,and the mean teacher method is used for consistency training of semi-supervised deep learning to improve the generalization ability of the model.On the datasets of AGNews,THUCNews and 20 Newsgroups,TextCNN or LSTM network is selected.Compared with the classical models,the SD-uwM model can not only improve the classification accuracy by 6.5%～14%,but also significantly improve the time performance.（2）Based on the framework of integrated learning and teacher-student model,using labeled data and unlabeled data,combined with consistency training framework,a Semisupervised deep learning model based on Multi-Teacher networks（Semi-MT）is proposed.①The model uses labeled data to train Student’s network parameters,and uses unlabeled data to generate comprehensive pseudo-label by fusing Multi-Teacher networks.B ack propagation guides Student’s network parameters training and learning,and improves the performance and stability of Student’s model;②Through the guidance and training of Student’s network parameters through Teacher models with different network structures,the purpose of integrating the advantages of Multi-Teacher models into the Student model is achieved,and the generalization ability of the Student model is improved,so as to improve the performance of the Semi-supervised model.On AGNews,THUCNews and 20 Newsgroups datasets,Semi-MT is compared with SemiBoost,Ranking Distillation and Mean Teacher models.It can be seen that the classification accuracy of Semi-MT model can be improved by 1%～9.9%,and the time performance is also significantly improved.