Using machine learning analysis to interpret the relationship between music emotion and lyric features

Area of Science:

• Music Psychology
• Computational Linguistics
• Affective Computing

Background:

• Music combines melody and lyrics to convey emotions, with research focusing more on musical structure than lyrical content.
• Understanding the role of lyrical features in music emotion recognition (MER) is crucial for a comprehensive approach.

Purpose of the Study:

• To investigate the impact of linguistic features extracted using Linguistic Inquiry and Word Count (LIWC) on the perceived arousal and valence of music.
• To assess the contribution of lyric features to the performance of music emotion recognition models.

Main Methods:

• Extracted lyric features from 2,372 Chinese songs using LIWC technology.
• Performed correlation analysis between lyric features and perceived music arousal/valence.
• Developed random forest regression models for MER using audio and lyric features, analyzing feature importance.

Main Results:

• Perceived music arousal correlated positively with lyric word count and negatively with past-oriented and insight words.
• Perceived music valence showed a negative correlation with the proportion of negative emotion words in lyrics.
• Lyric features significantly improved valence recognition models but were largely ineffective for arousal recognition models.
• Audio features were decisive for both arousal and valence recognition, while specific lyric features (sadness, positive emotions, tentativeness) were important for valence prediction.

Conclusions:

• Lyrical content, particularly word choice related to emotions, significantly influences the perceived valence of music.
• While audio features are primary for MER, lyric features offer valuable, albeit specific, contributions, especially for valence prediction.
• Future MER models can benefit from integrating carefully selected lyrical features to enhance emotional prediction accuracy.