Hybrid Approach in Emotionally Classification of Sinhala Music using NLP and ML

Abstract

"The historical musicology of Sri Lankan music binds with Sri Lankan's rich culture. The roots of Sri Lankan music began with Folk music around 1400 AD. Sri Lankan music considerably influenced Buddhist traditions with the development of Kolam and puppetry/ Nurthi music. Music is considered to be a universal language in the modern era. Whatever the language, music is considered mixed with emotions where listeners would stick to listen. Sinhala music tends to represent emotional representation to its listeners based on the melody and lyrics. Russell's circumplex model has been used in this research to classify songs into Happy / Calm and Sad emotions. The study uses JAudio to extract music information from the songs. Machine Learning techniques like Random Forest, SMO, Naive Bayes, Decision Tree and logistic regression algorithms were used in doing the classification. NLP was conducted using different word embedding techniques like TFIDF, Word2Vec and using pre-trained BERT models to get the classification. SMO performed well with all extracted features from JAudio, with an accuracy of 75.92%. Reliefbased attribute selection increased the SMO accuracy to 79.01%. Correlation-based attribute selection had 76.54% accuracy with SMO. The BERT model trained on top of xlm-roberta-based gained the highest accuracy of 61.22% in categorizing emotional labels using the lyrical data. The best performance was recorded with SMO with Relief-based attribute selection for music information with Machine Learning. The best NLP technique was the BERT model trained using transformers on top of the xlm-roberta-based model."