"STATE ToxiCN"

"STATE ToxiCN is a Chinese fine-grained toxicity component parsing dataset introduced by Bai et al. (2025), designed to support structured analysis of harmful expressions in social media. It provides span-level annotations for target\u2013argument\u2013hatefulness\u2013group quadruples, where each instance explicitly identifies the attacked target, the corresponding argument span, whether the expression is hateful, and the associated social group category. Unlike traditional sentence-level toxicity datasets, STATE ToxiCN emphasizes structured and interpretable representations of abusive content, enabling models to capture both span boundaries and cross-component dependencies. In addition, the dataset includes a large proportion of implicit and non-standard expressions, such as slang, sarcasm, and homophonic substitutions, making it particularly suitable for evaluating model robustness under realistic and challenging linguistic conditions."