E-government Text Classification Algorithm Research Based On Deep Learning

In recent years,with the rapid development of the Internet and information and communication technology,people have become accustomed to dealing with various businesses online.It has led to an explosive growth trend of massive text data on the Internet,which has brought much pressure on government platforms that rely on manpower to handle government affairs and provide services.As one of the feasible ways to accelerate work order approval and file classification,text classification technologies have gradually received wide recognition and attention from academia and industry.At present,in the field of egovernment,there are very few applications of text classification algorithms.Poor robustness of models is the main problem faced by existing models in practical applications.On the one hand,the reason for this phenomenon comes from problems of data,which are mainly manifested in: the noise rate of the text itself and its labels is high;there are a huge amount of text categories in text data and the distributions of samples among all categories are extremely unbalanced.On the other hand,the current models have some defects such as insufficient correlation mining between text labels and so on.In order to cope with the problems above,the paper takes the 12345 government service hotline as an application scenario to design and implement an e-government text classification algorithm based on deep learning.First of all,in order to solve the problems of high noise rate among labels and the unbalanced distribution of text samples among categories,the paper proposes an e-government text classification algorithm based on label correction and attribute perception.The algorithm uses the actual government work order text data to design the similarity matrix of labels,and build the intelligent evaluation and correction of labels module through the threshold setting,which realizes the reduction of the noise rate of text and labels without losing any data.And then the work order dispatching process is modeled as an e-government text classification problem,and the attribute-aware text classification module is designed to obtain more discriminative and representative text features to complete the single-label text classification task.Combining the two modules above,a two-stage e-government text classification algorithm framework is constructed to complete the task of intelligently dispatching government work orders.The experimental results indicate that compared with the direct employment of the Text CNN(Text Convolutional Neural Network,Text CNN)model,the accuracy of the algorithm on the two sets of e-government text data is improved by 5.23%and 3.91%,respectively.Next,to mine the global semantic features in the text and the correlation features between the labels,the paper proposes a multi-attribute prediction algorithm for the text from government affairs based on the bi-directional temporal convolutional and attention unit.The algorithm takes the pre-training language models and bidirectional temporal convolutional unit as the encoder.Meanwhile,it takes the long and short-term memory unit as the decoder,and the attention layer is used to generate intermediate semantic states.Finally,the encoder,the attention layer and the decoder jointly build a sequence to sequence and multi-label text classification network to obtain global and diverse semantic features for classification,thereby completing the multi-attribute prediction task of e-government text and improving the availability and generality of the intelligent government decision-making prediction framework.The experimental results show that compared with the baseline model,the algorithm achieves an improvement of about 1% ～ 2% on many classification indicators.The e-government text classification algorithm based on deep learning which is researched and implemented in the paper can be widely used in government affairs scenarios such as work order approval,policy and regulation release and document and web page classification.Also,it can be extended to related fields such as smart judicial judgment.