Research And Application Of Social Conflict And Dispute Event Classification Based On Deep Learning

With the increase of social conflicts and disputes,mediators need to deal with more and more events every day.At present,the event categories need to be summarized by mediators themselves.The workload is heavy and classification errors often occur.A simple,objective and efficient automatic text classification method is needed to replace manual classification.In recent years,the improvement of computer hardware performance has led to the vigorous development of deep learning,which is far better than traditional machine learning in text classification.In this paper,we use deep learning to classify social conflicts and disputes based on judicial texts to multiple categories and single labels.Based on the deep learning models such as CNN and RNN,we focus on the methods of extracting important features and retaining long-distance context dependent information,and propose HN-ATT+LDA+Text CNN text classification model,which is applied to social conflict sensing and control big data system for event classification.The specific contents are as follows:(1)Important feature extraction based on LDA weighted convolution.To solve the problem that CNN can't judge whether a word is important or not when it is applied to text classification,the LDA topic model is used to introduce the weight information on "word topic" to guide the convolution process of CNN.The LDA-conv+Text CNN model designed based on this method has a classification accuracy of 75.00%,which is superior to the Text CNN model constructed by other important feature extraction methods.(2)Research on context dependence extraction based on multi-level attention inheritance.To solve the problem that CNN can't extract the global context dependency when it is applied to text classification,the multi-layer attention model with increased attention allocation at paragraph level is used,and the word level coding layer is used to inherit the attention on each layer to solve the alignment problem.The classification accuracy of the HN-ATT-inherit+Text CNN model designed based on this method is 82.90%,which is better than the Text CNN model constructed by other context-dependent extraction methods.(3)HN-ATT+LDA+Text CNN combined model is proposed.Aiming at the two shortcomings of CNN applied to text classification,combined with the important feature extraction method and context-dependent extraction method proposed in this article,the combined model of HN-ATT+LDA+Text CNN is proposed,and its classification accuracy rate is 84.60%,higher than other researchers.(4)Application of HN-ATT+LDA+Text CNN combined model.In the business data management module of the social contradiction sensing and control big data system,an automatic text classification component based on the HN-ATT+LDA+Text CNN combined model is designed for recommendation of event classification.After testing,the classification accuracy of the automatic text classification component At about 85%,the response time is less than 1.3 seconds,which can meet the actual application needs,and can continuously improve the classification performance through feedback.