Advancing music emotion recognition: large-scale dataset construction and evaluator impact analysis

This study introduces a large-scale, multilingual, and multi-label dataset for Music Emotion Recognition (MER), addressing key limitations in existing resources. Notably, it represents the largest MER dataset to date that incorporates Chinese, offering rich information such as lyrics and metadata. By leveraging user-generated playlist tags from music platforms, the dataset captures diverse emotional attributes using a probabilistic framework across 12 emotion categories, with both discrete and continuous representations. Extensive experiments with various baseline models validate the dataset’s effectiveness, particularly in recognizing complex emotional patterns. A detailed analysis further investigates the influence of annotator quantity and quality on data reliability, demonstrating that increasing annotator numbers and enhancing their expertise improve label consistency and mitigate noise. This dataset serves as a valuable resource for advancing MER research and supports practical applications, including music recommendation and emotion analysis.

Text 124024.pdf - Published Version Restricted to Repository staff only Download (2MB) Official URL or Download Paper: https://link.springer.com/article/10.1007/s00530-0...

Actions (login required)

View Item