Quantifying the Emotional Landscape of Music with Three Dimensions

Kirtana Sunil Phatnani, Hemant A. Patil

Abstract: Music is a powerful art form that fosters a deep connection between the listener and the sound. Yet, sentiment analysis alone is limited in capturing the breadth and depth of emotions conveyed in songs, especially as individuals’ perceptions and interpretations of music vary widely. Our goal is to offer a more immersive and meaningful experience for listeners by harnessing the emotional contagion elicited by each song and gauging it through a multifaceted lens that considers identity, setting, and sentiment metrics. By analyzing the lyrics of songs that garnered similar Vader sentiment scores, we demonstrate that our innovative approach not only captures the essence of each composition but also uncovers nuanced differences in sentiment that escape traditional sentiment analysis. The divergence between our methodology integrating first-person sentiment shows variation of sentiment scores from –0.32 to 0.65 across the 10 songs having sentiment of approximately 0.99 (based on the traditional sentiment analysis method). Expanding the dataset to 47 positive songs and 48 negative songs from the Moody Lyrics dataset [3], we observe a variance of approximately 0.22 in positive and 0.19 in the negative songs (as compared to 0 in the traditional sentiment analysis), underscoring the remarkable intricacies that our approach can reveal. We also propose optimizing song recommendation using Reinforcement Learning (RL) utilizing these dimensions as states, choice of music as actions and accurate choice as reward. We propose this analysis can help dive deeper into the potential of emotions in music impacting the society as a whole.

External IDs:dblp:conf/specom/PhatnaniP23