Detecting hate speech and offensive language is essential for maintaining a
safe and respectful digital environment. This study examines the limitations of
state-of-the-art large language models (LLMs) in identifying offensive content
within systematically perturbed data, with a focus on Chinese, a language
particularly susceptible to such perturbations. We introduce
\textsf{ToxiCloakCN}, an enhanced dataset derived from ToxiCN, augmented with
homophonic substitutions and emoji transformations, to test the robustness of
LLMs against these cloaking perturbations. Our findings reveal that existing
models significantly underperform in detecting offensive content when these
perturbations are applied. We provide an in-depth analysis of how different
types of offensive content are affected by these perturbations and explore the
alignment between human and model explanations of offensiveness. Our work
highlights the urgent need for more advanced techniques in offensive language
detection to combat the evolving tactics used to evade detection mechanisms.

通过改进的数据集在汉语中测试大型语言模型（LLMs）对于侮辱性内容的检测，研究表明现有模型在存在这些干扰时检测侮辱性内容的性能明显不足，强调了对进阶技术的迫切需求以应对逃避检测机制的不断演变的策略。

ToxiCloakCN: 评估中文攻击性语言识别模型的鲁棒性

ToxiCloakCN: Evaluating Robustness of Offensive Language Detection in  Chinese with Cloaking Perturbations

This paper addresses the problem of detecting the offensive and abusive
content in Facebook comments, where we focus on the Algerian dialectal Arabic
which is one of under-resourced languages. The latter has a variety of dialects
mixed with different languages (i.e. Berber, French and English). In addition,
we deal with texts written in both Arabic and Roman scripts (i.e. Arabizi). Due
to the scarcity of works on the same language, we have built a new corpus
regrouping more than 8.7k texts manually annotated as normal, abusive and
offensive. We have conducted a series of experiments using the state-of-the-art
classifiers of text categorisation, namely: BiLSTM, CNN, FastText, SVM and NB.
The results showed acceptable performances, but the problem requires further
investigation on linguistic features to increase the identification accuracy.

本研究旨在解决 Facebook 评论中的攻击性和辱骂性内容检测问题，针对阿尔及利亚方言阿拉伯语进行探讨，使用 BiLSTM，CNN，FastText，SVM 和 NB 等文本分类器进行多项实验，并建立了一个包含 8.7k 手动标注为正常，虐待和攻击性的文本的新语料库。结果表明，分类器的性能尚可，但需要进一步研究语言特征以提高识别精度。