| In today’s era of big data,a large number of multi-domain policy texts appear in daily life.If these multi-domain policy text data can be effectively used,it can not only effectively help the people understand and maintain policies,but also assist the state in regulating governance methods,Improve governance efficiency.Because of this,many researchers have revolved around policy texts,hoping to apply the data more effectively.As one of the important tasks of natural language processing,text classification is the basis of many tasks.This paper uses the text classification technology of multi domain policy to support the construction project of multi domain policy knowledge map,and puts forward the corresponding model for multi domain policy text classification,so as to improve the accuracy of classification tasks.Aiming at multi domain policy texts,this paper proposes an efficient and accurate general policy text classification model BERT-RC.Based on the pre training model,this paper further improves the text representation and integrates more text context information into the text feature representation.Firstly,the model obtains excellent word vector representation through the pre training model,and then fully extracts the context features of the text through the two-way cyclic network.Then,the convolution neural network is used to strengthen the local text feature extraction,optimize the capture of text information and highlight the features of content.Finally,the results of text classification are obtained through these text features.In terms of policy text data,the F1 value of BERT-RC model is improved compared with the baseline model.The experiment shows that the BERT-RC model proposed in this paper can optimize the text feature representation and improve the classification accuracy of the model.The ablation experiment also proves the positive role of each module of the model.At the same time,aiming at the hierarchical multi text classification task,a fast and effective HFT-Trans model is proposed,which uses transformer and Bert word vector to optimize the text feature representation.and the hierarchical model structure is used to solve the problems of too little text training data and more detailed classification granularity in the lower level of hierarchical multi text classification,and the model parameters of the upper level are transferred to the model of the lower level,Make the lower model contain the classification information of the upper model,and use the data of the upper layer to promote the classification of the lower layer.At the same time,confrontation training is added to avoid the problem of over fitting of complex models on a small number of data sets,and the dynamic learning rate also accelerates the final convergence of the model.On this basis,integrated learning is finally added to verify that the network of integrated learning is indeed better than a single network,which can improve the generalization ability of the model.In the hierarchical multi text classification task scenario,the HFT-Trans model proposed in this paper has achieved good results.The experiments show that the HFT-Trans model can deal with the hierarchical multi text classification task well and improve the accuracy of the task. |
