Music Recommendation Method for Time-Series Emotions from Lyrics Using Valence-Arousal-Dominance Model

Abstract

In this paper, we represent a music recommendation system based on the lyrics data of music content. A system was built to recommend songs with a similar transition of an impression of lyrics. Our research consists of the lyrics metadata creation phase and the recommendation phase. In the lyrics metadata creation phase, as the Valence-Arousal-Dominance (VAD) model can show our human emotions in three independent dimensions, we extracted time-series emotional data from lyrics as numerical data. We repeated this process to create a dataset of lyrics with VAD score. We first extracted the VAD score the same way as in the first phase in the recommendation phase. To recommend songs that have a similar transition of the impression of lyrics, we used DTW to calculate the similarity of the songs. We implemented a recommendation system for songs with similar lyrical impressions through these two phases.