Application of the Bidirectional Long Short Term Memory with Comparison of Word2Vec, GloVe, and FastText for Emotion Classification in Song Lyrics

Abstract

Along with the development of today's world, many music player software and platforms offer the flexibility of listening to music anywhere and anytime. Song lyrics can help focus listeners' attention on certain emotions, therefore it is important to classify emotions in song lyrics for recommendation systems. This study aims to test the accuracy of the Bi-LSTM deep learning method, identify and compare the effect of word embedding, and examine the impact of parameters on accuracy in the classification of emotions in song lyrics. This study compares the impact of the algorithm on two datasets, namely the Music Lyrics Kaggle dataset of 3890 data with 8 emotional categories and the MoodyLyrics dataset of 2000 data with 4 emotional categories. The main indicators used in the test consist of embedding weights, embedding dimension, learning rate, and epoch. Results of the experiments show an accuracy of 62% on the MoodyLyrics dataset, while the Music Lyrics Kaggle dataset produced an accuracy of 42%. The test results also show that the GloVe word embedding model gives the most optimal results compared to Word2Vec and FastText.