Final Project Submission for the Natural Language Processing at McGill University!

Group: Dorian Desblancs, Jonah Nimijean, Jacob Sanz-Robinson

Abstract:

Emotion classification in music is a common problem with numerous applications. With the rise of streaming services, creating effective song recommendation systems is of interest to companies that wish to satisfy users by suggesting music which elicits the emotional qualities they desire. Classifying users’ mood preferences allows advertisers, users, and streaming services to customize and optimize content and advertising strategies. Traditionally, the focus of papers in the field consists of classifying audio data using signal processing techniques. Comparatively fewer papers exist on sentiment classification applied to lyrics to help identify the underlying emotional quality, even though they often outperform signal processing-based methods. In this paper we present a simple Convolutional Neural Network (CNN) which uses pre-trained GloVe embeddings to classify emotions in lyrics. Our novel approach to this problem surpassed the literature baselines on the MoodyLyrics dataset, achieving 90.8% accuracy on the multiclass version, and 92.5% accuracy on the binary version.