Multi-Label Sentiment Analysis on 100 Languages with Dynamic Weighting for Label Imbalance
This work introduces a sentiment analysis framework for 100 languages using XLM-RoBERTa that handles multi-label classification. It uses dynamic weighting to balance class contributions during training and adapts focal loss from single-label tasks. The method achieves state-of-the-art results across three languages on the SemEval competition, demonstrating superior performance in handling label imbalance.