Research And System Design Of Technologies For Mailboxes Based On Deep Learning

In modern society,email is an important way for people to communicate with each other.How to quickly know the information content of the huge amount of emails to be processed in mailboxes and pinpoint important emails has become the core problem to restrict the speed of information acquisition and improve the efficiency of information exchange.To address this problem,this thesis proposes to develop a new email assistant system,which realizes that email contents are displayed in the system inbox mailing list in the form of summary,helping users to greatly improve the speed and accuracy of processing emails;emails are distributed in different areas of the mailing list according to their importance.The main research contents of the thesis are.(1)Email data preprocessing: based on the publicly available Enron Email Dataset dataset,we use Deep L software to translate and correct morphemes to obtain Chinese email dataset;manually mark the email importance level and email summary;use jieba word splitting software to segment sentences in emails into words and filter deactivated words;use Word2 Vec model to generate word vectors.Prepare the dataset for deep learning.(2)Email text summary: use RNTN(Recursive Neural Tensor Network)model to extract sentence features to generate sentence vectors,which are used as the basis for calculating similarity scores by TextRank model,which is better than the original word frequency calculation of similarity;use different sentence positions to assign more weights to sentences at the beginning of the sentence,and use TF-IDF(term frequency-inverse document frequency)algorithm,using word frequency and inverse document frequency information to build a keyword word list,adjusting sentence scores,increasing the dimensional considerations other than similarity,and further improving the accuracy of summary sentence extraction.The test results show that the improved TextRank model achieves values of 0.3842,0.1526 and 0.3631 in Rouge-1,Rougle-2 and Rouge-L,which are higher than the other five models;the sentence with the highest final score is selected as the first summary sentence,and the sentence with the highest score in the dataset where the similarity of the same summary sentence is below the threshold is selected as the second summary sentences to express the email content more comprehensively and precisely.(3)Email classification: the sentence vector generated by RNTN is used as input,and the BiGRU(Bi-directional Gate Recurrent Unit)model is used to extract text features;the Attention mechanism is introduced to improve the feature grasping ability;the random setting method of the Query value in the Attention mechanism is modified,and the The Query value in the Attention mechanism is modified,and the Query value is initialized with the sentence vector of the text summary to improve the text feature extraction ability of the whole neural network.The test results show that the classification accuracy of this ATT-RNTN-BiGRU model reaches 83.7%,which is better than the Bi LSTM(Bi-directional Long Short-Term Memory),BiGRU and other neural network models.Meanwhile,Text CNN(Text Convolutional Neural Network)model is used to extract and classify text features of short emails to make up for the shortage of BiGRU network model.(4)Mailbox assistant system: vue framework and springboot framework are used to build the mailbox assistant system;the system provides mail summary information and importance classification in the inbox mail list.The usage shows that the summary accuracy rate reaches 82.5%,the classification accuracy rate reaches 85.6%,and the efficiency of customer mail processing is improved by more than 40%.