Gender Antagonism Detection in Chinese Social Media: Multi-Feature Fusion with LLM-Assisted Label Cleaning

Gender-antagonistic discourse pervades Chinese social media, yet automated detection remains challenging owing to informal register, creative character substitution, and noisy human annotations. This study presents CMLG, a lightweight classifier that fuses four embedding streams---CKBERT contextual representations, BGE dense semantics, Pinyin phonetic vectors, and Wubi structural vectors---through learnable additive attention before a Bi-GRU encoder. Because label noise in the SWSR corpus proves to be the dominant performance bottleneck, this study constructs a cleaning pipeline in which three heterogeneous LLMs independently re-annotate every sample, with disagreements resolved by majority vote. Controlled experiments across multiple feature configurations, dataset versions, and classifiers yield two principal findings: (1) label cleaning substantially raises macro F1 substantially points for all methods, confirming annotation quality as the primary constraint; and (2) Pinyin and Wubi features contribute nothing under noisy labels yet become the top-performing configuration once labels are cleaned, indicating that fine-grained orthographic cues are learnable only when the training signal is sufficiently reliable. BERT fine-tuning with cleaned labels achieves the highest absolute score, but CMLG's modular design uniquely enables per-feature contribution analysis that monolithic fine-tuning cannot support. These findings underscore the critical role of annotation quality in enabling fine-grained linguistic features for Chinese text classification and demonstrate a practical, model-agnostic pipeline for improving corpus reliability.