Abstract
With the rapid development of music streaming platforms, accurate understanding of lyric emotions has become crucial for enhancing personalized services in music recommendation systems. However, existing methods show significant limitations in processing local emotional features and long-range dependencies, particularly performing poorly when dealing with incomplete song information. This paper proposes LyricEmotionNet, a hybrid deep learning architecture based on CapsNet and Memory Networks, to address the challenges of local feature extraction and long-range dependency modeling in lyric emotion analysis tasks. The model achieves precise capture of local emotional features through CapsNet while utilizing Memory Networks to process long-sequence emotional dependencies, achieving a classification accuracy of 94.29% on a dataset comprising 660 songs across six emotion categories. Moreover, the model maintains a performance level of 90.20% in scenarios with missing data, significantly outperforming existing methods. Through systematic comparative experiments and ablation studies, we validate the model's advantages in terms of accuracy and robustness. The research findings provide new technical insights for music emotion analysis and personalized recommendation systems, while offering valuable reference for studies dealing with incomplete textual information.