Dual-Threshold Mining Boosts Chinese Offensive Comment Detection Across Platforms.

Detecting offensive comments on Chinese social media platforms presents a significant challenge, particularly when models need to perform consistently across different platforms. Existing methods often experience performance degradation due to the inherent differences in language use and context across platforms like Weibo, Xiaohongshu, Tieba, and Zhihu. To tackle this, a new dual-threshold hard example mining strategy has been introduced. This method begins by fine-tuning a base RoBERTa model on a Chinese offensive language dataset to establish a baseline. It then systematically quantifies domain distances and identifies performance bottlenecks under domain shift using a newly constructed, fine-labeled test set spanning multiple platforms. The core of the approach involves filtering high- and low-confidence error-prone samples from unlabeled data based on prediction confidence. The model is then re-fine-tuned using only a small set of manually labeled hard examples, enabling low-cost adaptation across platforms. Experiments demonstrate substantial performance improvements on all four tested platforms, making it a more robust solution for content moderation.