Addressing class-imbalance challenges in cross-lingual aspect-based sentiment analysis: Dynamic weighted loss and anti-decoupling

Nankai Lin, Meiyu Zeng, Xingming Liao, Weizhong Liu, Aimin Yang, Dong Zhou

Abstract: Highlights•Equi-XABSA mitigates sample imbalance and languages’ representation disparity.•The anti-decoupling strategy enhances semantic transfer between languages.•The dynamic weighted loss modulates the model’s attention across categories.•Our method surpasses existing models and resolves XABSA imbalance.