Research On Campus Network Public Opinion Based On Deep Learning

The development of social media has brought about the proliferation of false rumors,and unrealistic rumors manipulate public opinion and mislead public judgment,and affect social stability.Students,who are more self-aware but less able to distinguish true from false information,are more likely to be influenced by rumors and thus generate a lot of negative public opinion and emotions.In order to maintain students’ psychological health and build a harmonious campus environment,scientific screening of public opinion information is needed first,and artificial intelligence is trying to play this role nowadays.In order to solve the bulk collection of campus public opinion data and realize automatic discrimination of public opinion sentiment tendencies,this dissertation investigates the task of sentiment analysis of public opinion data in campus networks with the help of deep learning methods,and the main research contents are as follows:1)To solve the problem of real-time collection of massive campus opinion data and ensure the authenticity and timeliness of the sources of opinion monitoring data,this dissertation firstly implements the batch collection of campus opinion data with the help of a crawler script written by Scrapy framework and cleans the data by means of regular matching.The pre-processing process includes word separation,deactivation and text representation of opinion data using Word2 vec and BERT pre-training models.2)The training of opinion analysis models often requires a large amount of data to complete,and in order to improve the discriminative ability of opinion sentiment,the existing public dataset of university opinion,Senti,needs to be further enhanced.this dissertation designs a Transformer-based generative adversarial network（T-GAN）model to expand the unevenly distributed dataset by generating sentence-level text.the T-GAN is based on the traditional GAN The T-GAN is improved on the basis of the traditional GAN,in which the generator introduces the structure of a variational self-encoder,which extracts the text features as the input of the generator;the Transformer-based discriminator can better capture the longterm dependencies between words and generate high-quality samples.3)To address the automated analysis of massive campus public opinion data,advanced deep learning techniques are used to achieve accurate mining of campus information content,make sentiment tendency judgments for student opinion,and help schools formulate relevant policies with targeting.In this dissertation,a hybrid embedded multi-scale convolutional neural network（HWE-MCNN）model is designed to extract features from the collected public opinion data using convolutional structures,and the model uses a hybrid embedded text representation based on the word vector text representation and multi-scale convolutional structures,combining words and lexicality as another channel to fuse with the word vector convolutional network.4)The system is designed and implemented to automatically collect public opinion data on campus networks,analyze public opinion sentiment tendencies in batch and extract public opinion themes.The system uses crawler scripts to collect data,determines opinion sentiment types by HWE-MCNN model,and Text Rank model extracts themes of opinion data and performs statistics and analysis.The system provides timely correction of negative public opinion impacts on campus and provides data to rely on for school management and building a good campus environment.To address the problem of overfitting arising from uneven data distribution,the T-GAN model combines the advantages of variational self-encoder and traditional GAN,and the experimental results show that the classification accuracy is improved by 18.61% and 1.67% in the SVM and Ro BERTa models using enhanced dataset training.To address the problem that a single text representation cannot accurately express the semantics,the HWE-MCNN model extracts multiple text representation features after fusing them on the basis of the traditional single-channel model and adds an attention mechanism to amplify the key features.The accuracy is improved by 1.66%.