Research On Department Matching And Automatic Transferring Method Of Online Government-civilian Interaction Mailbox

With the improvement of the government openness and the advancement of the social democratization,the demand of public participation in politics is increasingly strong.However,due to the time and distance constraints,it takes a lot of manpower and material resources to express demands and make suggestions to the government offline,while the online government-civilian interaction mailbox provides a more convenient and efficient channel for the public.When there is a large number of letters from the public,the manual classification method is laborious,while the automatic transfer has strong advantages.To solve this problem,this paper studies and designs the automatic transfer method of online government-civilian interaction mailbox based on text classification.This method uses algorithm to learn the text characteristics of different departments’ letters,and then matches the new letters with the characteristics of the departments learned before and automatically transfers them to the right departments.The deep pre-training language model is the most superior and cutting-edge application model in the field of natural language processing.This paper crawls the public letters in online government-civilian interaction mailbox of several cities and builds a deep learning environment.Then optimize the BERT model on the dataset,and complete the text classification experiment.At the same time,this paper selects several traditional machine learning algorithms for comparison.The experimental results show that the accuracy and F1-score of the logistic regression algorithm in the datasets of three city are the best among the eight traditional machine learning algorithms.However,even compared with the logistic regression algorithm,the BERT pre-training language model has obvious advantages and its classification accuracy in the datasets of Beijing,Changsha and Shenzhen reaches91.53%,85.02% and 90.12%.When analyzing the classification results of BERT model,it is found that the departments with poor classification accuracy usually cross functions with other departments;the sample size also has some influence on the experimental results: the more the samples are,the better the classification effect is,which is consistent with the conventional cognition that the deep language model has better performance on large-scale datasets;the analysis of the TOP N accuracy of the model shows that the accuracy of TOP 2 has been greatly improved compared with TOP 1accuracy.The TOP 2 accuracy in three cities are 97.03%,93.14% and 95.71%respectively.The error rate of TOP 5 is very small,which is lower than 3%.And the TOP 5 accuracy of Beijing’s dataset even reaches 99.46%.Based on the analysis of experimental results,this paper designs the scheme of automatic transfer of online government-civilian interaction mailbox from four steps:data collection and preprocessing,model training,model application and model update.The flow chart of the scheme based on the BERT model is given.At the same time,this paper gives the suggestions of automatic transfer method in practical application based on the analysis results of the experimental results:(1)Standardize the department labels in the data collection and preprocessing stage;(2)In the model training stage,use the letters of online government-civilian interaction mailbox to pre-train the BERT model and then apply it to the downstream classification task;(3)In the application stage,set a threshold to transfer letters to multiple departments;(4)Enrich and strengthen the features of department with few letters.The automatic transfer method of online government-civilian interaction mailbox designed in this paper can optimize the current operation mechanism of the interaction mailbox,which is of great significance to improve the efficiency of online governmentcivilian interaction and reduce the government human and administrative costs.