Multi-Label Sentiment Analysis on 100 Languages With Dynamic Weighting for Label Imbalance

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Machine Learning

Background:

• Cross-lingual sentiment analysis is crucial for market research, politics, and social sciences.
• Existing methods often use static weighting, failing to adapt to class imbalance.
• Plutchik's wheel of emotions provides a framework for multi-label sentiment classification.

Purpose of the Study:

• To develop an advanced cross-lingual sentiment analysis framework for multi-label classification.
• To introduce a dynamic weighting method for improved class contribution balancing.
• To adapt focal loss and derive optimal thresholds for enhanced performance.

Main Methods:

• Implemented a multi-label sentiment analysis framework.
• Introduced a novel dynamic weighting method to balance class contributions.
• Adapted focal loss for the multi-label setting and derived optimal class-specific thresholds.

Main Results:

• Achieved state-of-the-art performance in seven of nine metrics across three languages.
• Outperformed common baselines and SemEval competition methods.
• The single model demonstrated effectiveness in cross-lingual sentiment analysis.

Conclusions:

• The proposed dynamic weighting and focal loss adaptation significantly improve multi-label cross-lingual sentiment analysis.
• The method offers a robust and efficient solution for sentiment analysis across 100 languages.
• Publicly released code facilitates further research and development in the field.