Reproduce “SWSR: A Chinese Dataset and Lexicon for Online Sexism Detection” on OpenResearch

About this paper

Aiqi Jiang, Xiaohan Yang, Yang Liu, Arkaitz Zubiaga

This research introduces SWSR, the first dedicated Chinese dataset for online sexism detection, and SexHateLex, a corresponding lexicon of 3,016 abusive terms. It establishes benchmark performance using state-of-the-art models, showing transformer-based architectures achieve up to 0.780 macro F1 for sexism identification, with slight improvements from lexical features.

Key insights

• Transformer-based models, particularly RoBERTa, consistently outperformed other model categories in detecting Chinese sexism across all tasks, achieving the highest F1 scores.
• Character-level models demonstrated superior performance over word-level models in Chinese, suggesting their importance due to the language's lack of explicit word boundaries.
• Integrating the SexHateLex lexicon provided consistent, albeit modest (0.5-1% F1 increase), performance improvements across models, validating its utility for Chinese sexism detection.
• Models struggled significantly with implicit sexism, sarcasm, and context-dependent abusive terms, indicating a need for better contextual understanding and integration of external knowledge.

Summary by alphaXiv.