Research On Song Sentiment Classification Method For Chinese Lyrics

Songs are an integral part of people’s lives.In recent years,the Internet has developed rapidly and songs have grown tremendously in the number of distribution,but with it comes the problem of retrieving and managing the huge amount of online songs.Songs are the carriers of people’s emotions.Traditional song emotion classification is mostly based on audio,but the features contained in audio are complicated and difficult to extract.Lyrics are the soul of a song,and they have a simpler structure than audio.Aiming at Chinese lyric texts,a lyric sentiment classification model based on Shuffle Net and sentiment dictionary is proposed,which can effectively classify the sentiment of songs.The main work is as follows.(1)A sentiment dictionary for Chinese lyrics based on the cyclic sentiment model is constructed.Firstly,the sentiment root table is constructed based on the cyclic sentiment model;Then the sentiment root table is expanded twice using the Hownet sentiment dictionary and the Chinese lyrics data set to construct the sentiment dictionary of Chinese lyrics;Finally the Chinese lyrics sentiment dictionary,degree adverb dictionary and negation dictionary were used to calculate the sentiment score of the lyrics by the designed calculation method.(2)A Shuffle Net-based sentiment classification model for lyrics is proposed.Firstly,we obtain the Chinese lyrics dataset by packet crawling and crawler;Then we pre-process the Chinese lyrics dataset,including redundant information removal,deactivation,word separation and word embedding;Then we build Shuffle Net network using Shuffle Net basic unit;Finally,we input the preprocessed data to Shuffle Net network,and then we get the classification target by fully connected layer.(3)A Sentiment Classification Model for Lyrics Based on Shuffle Net and Sentiment Dictionary is constructed.After obtaining the sentiment analysis results based on Shuffle Net and sentiment dictionary respectively,the final classification results were obtained by decision fusion using D-S evidence theory;Finally,three experiments are compared in terms of the availability of sentiment dictionaries,different decision fusion methods and replacement of different models,and finally the feasibility and accuracy of the model proposed in this paper are demonstrated.