Data Augmentation for Lyrics Emotion Estimation

Lyrics emotion estimation can allow us to realise song retrieval systems and song recommendation systems which are based on not only text retrieval nor melody matching but also emotions in lyrics or transitions of emotions within lyrics of a whole song. This requires lyrics emotion corpora of phrase. However, it is difficult to build large scale lyrics emotion corpora because emotions are labelled manually. In this paper, we propose a method to augment lyrics emotion corpora. As a result, we augmented a corpus consisting of 366 phrases into a larger corpus consisting of 2145 phrases. We also evaluate the proposed method using 2 convolutional neural networks trained on original corpus and augmented corpus respectively. We define the target emotion classes as Joy, Love, Anger, Sorrow and Anxiety. Mean accuracy of the model trained on the augmented corpus was 75.9% whilst the model trained on the original corpus performed 70.7%.

© Academy of Sciences Malaysia