Evaluating large-language models for dimensional music emotion prediction from social media discourse
Published in ICNLSP, 2022
The task of music emotion recognition has been of interest in the music information retrieval domain. We investigate the use of online social media discussions as a potential input for music emotion prediction models. Using this commentary, we evaulate the performance of two pre-trained large language models in predicting the valence and arousal of a piece of music. We query three social media platforms to build a corpus of conversations surrounding 2,402 songs. We achieve modest Pearson’s correlations of 0.62 and 0.72 to valence and arousal targets respectively. These results demonstrate that there may be a connection between the sentiment expressed in the online discourse around a song and a listener’s affective response to said song.
Recommended citation: P. J. Donnelly and A. Beery, “Evaluating Large-Language Models for Dimensional Music Emotion Prediction from Social Media Discourse”