Research On Emotion Classification Of Pop Music App Reviews

As the online social media becomes popular and develops in a high speed,it has promoted the emergence of a music social platform that integrates music players and Internet social networking platforms.In addition to being the original music player-type tool platforms,the mainstream music social platforms also add the functions of music comments,interaction between musicians and users,matching of music hobbies and friends,etc.Music is the main content of music social platforms,and music itself can easily guide the emotion resonance of audience,so the communication content can be more full of emotion than ordinary communication platforms,in which the value of emotion analysis is also embodied.In this thesis,the Hou Yi collector was adopted to crawl 10,705 pieces of com ment data for singer Li Ronghao’s song Getting Achievements in Youth from June 2020 to June 2021 on the comment area of NetEase Cloud platform.Based on 10,601 pieces of data after removing duplicate comments,the LDA model,machine le arning method and Snownlp library were used to do some text analysis and sentim ent analysis.The following is the main research work and conclusions:(1)The LDA model of text mining technology was adopted to make the theme extraction and word cloud visualization of song reviews.According to the research results,it indicates that the themes mainly focus on "encouragement","love relationship","hard working" and "college entrance examination",and the emotions are generally more positive.(2)The TF-IDF was adopted to extract features from text,the positive and negative emotional labels were manually marked,and the traditional machine learning classifiers were adopted for classification.After screening the optimal model and optimal parameters,the logistic regression model was used for modeling,the final accuracy rate reached 65.14%,which achieved the general model fitting effect.(3)The Snownlp was used to make sentiment scoring for the comments,and the t hreshold was set as 0.5 for positive and negative sentiment classification.The accur acy,precision,recall,F1 value,confusion matrix,ROC curve and AUC area were considered to carry out performance evaluation,so the more excellent model could be chosen.The accuracy rate reached about 76.28%,which increased about ten per cent.The Snownlp based on the self-built corpus can effectively distinguish positive and negative emotions.Then,the topic extraction was conducted in the positive an d negative sentiment corpus,and some positive parts could be found in the negativ e corpus.According to the research conclusions of this thesis,it can help creators and platforms better understand listeners’ emotional tendencies and comment topics for songs.To a certain extent,it can also guide creators’ creative direction and platform’s song recommendations.