Emotion Classification of Song Lyrics using Bidirectional LSTM Method with GloVe Word Representation Weighting

Abstract

The rapid change of the music market from analog to digital has caused a rapid increase in the amount of music that is spread throughout the world as well because music is easier to make and sell. The amount of music available has changed the way people find music, one of which is based on the emotion of the song. The existence of music emotion recognition and recommendation helps music listeners find songs in accordance with their emotions. Therefore, the classification of emotions is needed to determine the emotions of a song. The emotional classification of a song is largely based on feature extraction and learning from the available data sets. Various learning algorithms have been used to classify song emotions and produce different accuracy. In this study, the Bidirectional Long-short Term Memory (Bi-LSTM) deep learning method with weighting words using GloVe is used to classify the song's emotions using the lyrics of the song. The result shows that the Bi-LSTM model with dropout layer and activity regularization can produce an accuracy of 91.08%. Dropout, activity regularization and learning rate decay parameters can reduce the difference between training loss and validation loss by 0.15.