Research On Text Sentiment Analysis Of Weibo COVID-19 Topic Based On Deep Learning

With the continuous development of the times,the Internet has become more and more important in human life,and has become a necessity in daily life.At the same time,with the rise of software such as Weibo and TikTok,the Internet began to enter the era of self-media.This progress has also prompted netizens to express a large number of views and opinions on things freely,and public opinion analysis came into being.Public opinion analysis can quickly determine the people's emotional tendency towards some major emergencies,objectively reflect the social public opinion orientation,and help the state and relevant departments to quickly understand public opinion,so as to achieve supervision and accurate guidance,thus maintaining social stability.Since the Year of the Rat,the focus of the whole society has always been Covid-19.In the face of the menacing epidemic,the citizens can not communicate offline,so they rely more on the Internet to express their concerns about the current situation of the epidemic or the improvement of the epidemic.Alas,especially in Weibo,comments spread faster.Once there is a slight recurrence of the epidemic,it will inevitably arouse great social concern and even panic.Therefore,how to grasp public opinion in a timely and accurate manner and how to correctly guide public opinion to avoid panic has become a vital issue and a new test in the context of the new crown pneumonia epidemic.Based on the above issues,this article conducts research on the microblog network public opinion related to COVID-19.Through gradual optimization,three different models were obtained.The main research work consists of the following parts.1.Preprocessing of comment texts related to COVID-19 in Weibo.The text uses web crawling technology to crawl 100,000 pieces of new coronary pneumonia-related comment information from August to September 2020 on Sina Weibo,and filter it through topic models,similar algorithms,and search keywords to remove low correlation with new coronary pneumonia or irrelevant noise data;After manual tagging,the data is preprocessed by using word segmentation technology and statistical methods.2.According to the characteristics of CNN and BiLSTM in text processing,the model for text sentiment analysis is constructed by combining them.The advantage of CNN is to extract complex features,and the advantage of the BiLSTM model is that it can better control the whole sentence by retaining historical information.In order to determine the optimal parameters under the new model,a multi-parameter comparison experiment is carried out in this paper.Compared with the traditional model,the classification performance is improved.3.According to the characteristics of CNN and BiGRU in text processing,a new model is constructed by combining them.The basic structure of the GRU model contains only two kinds of gating settings,while the LSTM model contains three kinds of gating settings.In contrast,the GRU has one less gating setting,so the model structure is simpler.Therefore,under the same classification performance,GRU model is better.The optimal parameters of the new model are determined by multi-parameter comparison experiment,and the classification performance is improved compared with traditional model.At the same time,compared with the C-BiLSTM model,it is found that the classification performance is similar,but the structure is simpler and the application time shorter.4.The shortcomings of the improved C-BiLSTM model.Its disadvantage is that it can not effectively solve the problem of distribution weight of feature vectors.Simply changing the parameters can no longer be optimized,so the self-attention mechanism is introduced,and a c-The BiLSTM-SA model is used to analyze text emotion.This model pays attention to the difference in importance of each word in a sentence relative to the whole sentence,and gives different weights to different words.At last,the improved C-BiLSTM-SA model is verified through the self-built data set of this paper,which proves that it has a better classification effect.Compared with other representative papers,the experimental results show that the model has better performance.