Multi-class Emotion Mining Of Epidemic Public Opinions

Since December 2019,patients infected with COVID-19 have appeared in China one after another,followed by an increasing amount of information about its outbreak on microblogs and other platforms.Through sentiment mining of microblog data related to COVID-19,a large amount of valuable sentiment information can be extracted,thus enhancing the real-time monitoring of online public opinions,which is an important method for assisting relevant organizations and departments in public opinion analysis through statistical tools.The current sentiment analysis method for public opinions is mainly textbased two-class sentiment mining,but binary sentiment mining suffers from the problem that sentiment polarity is too homogeneous,which is gradually not applicable to complex texts.To mine text sentiment more accurately,multi-class sentiment mining of public opinions on the epidemic is explored in this paper from a tri-class perspective and the visualization of epidemic data,the expansion of sentiment lexicon in the domain of epidemic as well as machine learning methods are investigated respectively.The main work and innovations of this paper are as follows.(1)Descriptive mining is conducted in this paper for multi-class data of opinions on epidemic.Data statistics and visualization analysis are conducted in this paper to explore the distribution of text length and the proportion of each class of sentiment tendency.The word cloud maps under emotional states of different classes are drawn and analyzed to understand the overall emotional state of the masses on the microblogging platforms and grasp the trend of online public opinions.The experimental results show that people’s overall sentiment towards the COVID-19 epidemic is rational and positive,and the topics of concern and discussion mainly include its prevention and control,curing patients and helping Wuhan.(2)A multi-class sentiment lexicon construction method is proposed in this paper.Although emotions are multi-class in practical situations,emotion words are only classified into two classes in general in most current emotion dictionaries.To address the current situation that there is almost no research on sentiment dictionaries in the field of multi-class epidemics,a multi-class sentiment dictionary construction method is proposed in this paper by introducing the PMI calculation of neutral words.Through the method,the problem of sparsity of the SO-PMI algorithm in neutral word discrimination is solved,successfully expanding the sentiment lexicon in the domain of epidemic,and text sentiment mining is completed on this basis.The experimental results show that based on the binary data,compared with the original dictionary with an accuracy of 52.79%,the accuracy of the expanded sentiment dictionary in the domain of epidemic reaches 77.21%,with an improvement of 24.42% and a growth rate of 46.26%;based on the multi-class data,compared with the original dictionary with an accuracy of 48.38%,the accuracy of the expanded dictionary in this paper is 55.23%,with an improvement of 6.85% and a growth rate of 14.16%,proving the effectiveness of the method.However,due to the limitations of the dictionary itself,the sentiment words outside it cannot be accurately labeled,and there is still room for the improvement of the effect of sentiment mining,and better sentiment mining methods need to be found.(3)The feasibility of machine learning methods in multi-class sentiment mining is explored in this paper.In order to further improve the performance of sentiment mining,this paper tries to use machine learning methods for multi-class sentiment mining of data.Through sentiment mining experiments,we compare sentiment dictionaries,traditional machine learning(plain Bayesian model,random forest model,logistic regression model and support vector machine model)and pre-trained models in deep learning(ALBERT model,called "A Little Bert",which is based on the BERT pre-training model with compression optimization)in sentiment mining.The experimental results show that the machine learning model outperforms the sentiment dictionary in this experimental dataset.The ALBERT model,with its superior semantic information capturing ability,achieves an accuracy of 95.65% in two-class data and an accuracy of 81.48% in multi-class data,which has better results.To sum up,multi-class emotion mining for public opinions on epidemic is conducted in this paper and the public emotional states are objectively demonstrated,which has strong application value for guiding public sentiment and creating a favorable environment for public opinions.