An Empirical Study On The Classification Of Petition Problems Based On BERT

With the diversification of petition channels,the deepening of petition work,the intensity of petition work,the timeliness of petition work is higher than before,and various new problems such as weak ability,insufficient number and professionalism of petition staff have now become key factors restricting the development of petition work towards modernization.With the deepening of the research on text classification,deep learning has become the mainstream method of text classification,and this research is based on the text of the petition and the text classification experiment of the petition problem through deep learning.It is a topic of great significance to provide a valuable reference for petition personnel to classify and separate petitions by multi-label classification of petition categories in the text of petitions,improve the efficiency of the separate handling of petition staff,and lay the foundation for the automatic classification of petition issues.Based on systematically summarizing the research results on the classification of petition problems at home and abroad,combined with the basic theory,the difference between multi-classification and multi-label classification,the correlation between petition problems and government functional departments are clarified,and a multi-classification model based on petition problems is constructed by using text classification methods such as adversarial training and BERT pre-training model,and empirical analysis is carried out.The main research contents of the paper are as follows:According to the classification of letters and visits,86 labels are constructed according to the government functions and contents of letters and visits.Based on the letters and visits of a city in a province in 2022,the available experimental data are obtained through data collection,data cleaning and data annotation.BERT pre-training model is used to conduct in-depth text mining of letters and visits.The vector representation of the text information in the letters and visits problem and the establishment of a reliable empirical model to classify the letters and visits problem.The experimental data show that the accuracy rate of the model is 88.49% in the multi-classification of letters and visits.Secondly,this paper optimizes BERT pre-training model.By integrating confrontation training,perturbation is added to word embedding to improve the robustness of the model to cope with confrontation samples and enhance the generalization ability of the model.In this way,the pre-training model can mine and classify letters and calls more effectively,thus improving the accuracy of the classification of letters and calls.The experiment shows that the performance of BERT pre-training model integrated with confrontation training has been significantly improved,with an accuracy of 91.03%,which is 2.54% higher than that of petition problem classification based on BERT pre-training model.