LyricEmotionNet for robust emotion recognition with hybrid CapsNet-memory network architecture

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Music Information Retrieval

Background:

• Accurate lyric emotion analysis is vital for personalized music recommendations on streaming platforms.
• Existing methods struggle with local emotional features, long-range dependencies, and incomplete song data.

Purpose of the Study:

• To propose LyricEmotionNet, a hybrid deep learning model combining CapsNet and Memory Networks.
• To enhance lyric emotion analysis by effectively processing local features and long-range dependencies.
• To improve the robustness of emotion analysis models when dealing with incomplete textual information.

Main Methods:

• Developed a hybrid deep learning architecture, LyricEmotionNet, integrating CapsNet for local feature extraction and Memory Networks for long-range dependency modeling.
• Utilized CapsNet to precisely capture local emotional nuances within lyrics.
• Employed Memory Networks to effectively model emotional dependencies across extended lyrical sequences.

Main Results:

• Achieved a classification accuracy of 94.29% on a dataset of 660 songs across six emotion categories.
• Maintained a performance of 90.20% accuracy even with missing song data, outperforming existing methods.
• Demonstrated significant advantages in accuracy and robustness through comparative experiments and ablation studies.

Conclusions:

• LyricEmotionNet offers a robust solution for lyric emotion analysis, particularly with incomplete data.
• The model provides new technical insights for music emotion analysis and personalized recommendation systems.
• Findings offer valuable reference for studies involving incomplete textual data analysis.