Music Emotion Recognition Using Hierarchical Contrastive Learning

The Music Emotion Recognition (MER) task has garnered significant attention from both academic and industrial fields due to its versatility in various fields, such as music recommendation systems and psychotherapy. Due to advancements in language model performance, there has been a rise in approaches that address the MER tasks using only lyric data. Despite the empirical evidence suggesting that valence is a more reliable indicator of emotion disambiguation than arousal, previous studies have often overlooked this relationship. To address this research gap in the literature and enhance performance, we propose considering valence and arousal within a hierarchical framework. To preserve this hierarchical structure in the embedding space, we employed Hierarchical Contrastive Learning to enhance the MER performance while preserving the hierarchical organization. The experimental results show that this approach improves the classification performance and leads to a more balanced model. Furthermore, it is noteworthy that the model’s performance improvement rate is faster when compared to the application of cross-entropy loss.