CangjieToxi: A Chinese Offensive Language Detection Benchmark with Radical-Level Perturbations

ACL ARR 2025 February Submission8528 Authors

Abstract: In this paper, we introduce CangjieToxi, a novel benchmark dataset designed to address the challenges of detecting covert offensive language in Chinese social media. Existing detection systems are often ineffective against evasion techniques that manipulate character structure to bypass censorship. We focus on two key perturbation methods: character splitting and character substitution. Character splitting involves breaking down offensive words into visually similar but contextually distinct components, while character substitution replaces offensive characters with visually similar but non-offensive ones, thus concealing the original intent. Our dataset incorporates these techniques to create more complex forms of toxicity that are difficult for traditional models to detect. We conduct extensive experiments with state-of-the-art models, revealing their limitations in handling these perturbations and demonstrating the need for more robust systems. This work advances the field by providing a resource to improve the detection of cloaked offensive language and contributing to the development of censorship-resistant detection methods.

Paper Type: Long

Research Area: Resources and Evaluation

Research Area Keywords: benchmarking, language resources, lexical semantic change

Contribution Types: Data resources

Languages Studied: Chinese

Submission Number: 8528