Cross-Platform Chinese Offensive Comment Detection via Dual-Threshold Hard Example Mining

A new machine-learning method aims to solve a persistent problem in Chinese-language content moderation: models trained on one social media platform often fail when deployed on another. The technique, described in a paper submitted to arXiv on 26 June 2026, uses a dual-threshold strategy to mine difficult examples and adapt a detection model across four major platforms without requiring large volumes of new labeled data. The research targets cross-platform performance degradation in offensive comment detection for Chinese social media [1]. The authors first fine-tuned a clean-Chinese-base RoBERTa model on the COLD dataset to create a binary classification baseline [2]. They then built a three-class, fine-labeled test set spanning posts from Weibo, Xiaohongshu, Tieba, and Zhihu [2]. Domain distances between the source training data and each target platform were quantified using Jaccard similarity and Proxy-A Distance, systematically revealing where and how the baseline model's performance broke down under domain shift [2]. To close that gap, the team introduced a dual-threshold hard example mining strategy. The process filters high-confidence and low-confidence error-prone samples from unlabeled corpora based on the model's prediction confidence scores [2]. Only a small set of these filtered examples is then manually labeled. The model undergoes a secondary fine-tuning phase under implicit contexts using this compact, hand-labeled set, an approach the authors describe as low-cost cross-platform domain adaptation [2]. The paper reports significant performance gains for the optimized model across all four platforms, though specific metrics were not detailed in the abstract [2]. The work arrives as Chinese platforms continue to scale their automated moderation systems, where maintaining accuracy across linguistically and culturally distinct user bases remains a technical challenge. The preprint was posted on arXiv under the Computation and Language category and is supported by arXivLabs, a framework that allows collaborators to develop and share new features on the repository [1].