Learning Affective Responses to Music from Social Media Discourse

The ability to automatically estimate typical affective responses to music would enable the development of emotion-aware music recommendation systems. However, the lack of suitable datasets for this task has hindered attempts to design such systems. In this work, we introduce social media conversational data as a new feature space for music emotion recognition. We create a large dataset of social media musical discourse with over 11.8 million comments from Reddit and YouTube discussing 19,627 different songs. We fine-tune large language models on this conversational data in a two-target regression task to predict music valence and arousal annotations. We demonstrate a modest ability to estimate human annotated music emotion targets directly from social media comments. Our highest performing model achieves Pearson’s correlations of 0.80 and 0.79 for valence and arousal, respectively. These results imply that emotive qualities of a song may be inferred directly from social media conversations, without access to the audio or lyrics.