Revisiting the role of lyrics in music emotion recognition: A critical analysis of the semantic–perceived emotion gap and methodological challenges

Lyrics-based music emotion recognition (MER) faces a persistent performance gap: models that process full lyrics often underperform audio systems trained on brief excerpts. This review argues that the gap is largely methodological rather than an inherent limitation of lyrical semantics, and synthesizes evidence from 2017–2025 to identify three compounding causes. First, annotation practices routinely conflate semantic (text-only) and perceived (music-integrated) emotion, training lyrics models against targets that encode acoustic information the text cannot observe. Second, standard NLP pipelines treat lyrics as ordinary prose, discarding verse–chorus structure, narrative progression, and figurative and phonological devices that carry affective cues—resulting in representations that underestimate what lyrics could in principle convey. Third, prevailing fusion architectures assume uniform modality contributions across affect dimensions, despite systematic evidence of modality–dimension asymmetry; symmetric designs consequently inject noise that can degrade rather than improve prediction. Guided by these findings, we formulate three design principles: extract lyrics’ latent signals through domain-adapted, structure-aware encoders; evaluate under conditions that isolate the textual contribution; and integrate modalities with their dimension-specific strengths in mind. Reframing this performance gap as a design-alignment problem, this review provides a diagnostic framework and testable principles for advancing lyrics-centric MER research.

Text 124021.pdf - Published Version Restricted to Repository staff only Download (5MB) Official URL or Download Paper: https://www.sciencedirect.com/science/article/pii/...

Actions (login required)

View Item