ollabalala/MOUSE

Abstract

MOUSE is a novel diagnostic benchmark designed to evaluate the capability of Large Language Models (LLMs) in decoding and translating "Chouxiang" language—a highly sophisticated Chinese internet subculture. Unlike standard slang, Chouxiang language utilizes multi-layered adversarial obfuscation, combining phonetic ciphers (homophones), semiotic metaphors (emojis), and semantic shifts to bypass automated moderation and establish group identity.

Ethical Disclaimer

This study focuses on Chinese Chouxiang Language, a form of online subcultural language with complex pragmatic functions. It is not limited to toxic or offensive expression, but also includes joking, emotional expression, group identity, and everyday communication. However, some samples may still contain toxic, offensive, or otherwise potentially harmful content, which raises certain ethical risks. We construct Mouse to support scientific research and to improve the understanding of subcultural language in NLP, rather than to encourage, spread, or amplify harmful expression.

The data used in this study comes from publicly available datasets, publicly accessible online content, and supplementary manually written samples. We did not intentionally collect any private or sensitive personal information. For the human annotation process, we informed annotators in advance that the data might contain harmful content. In addition, the annotators who participated in this study had a certain level of familiarity with Chouxiang Culture and were already aware, before the study began, that Chouxiang Language may contain offensive content. Human annotators participated only in dataset construction and completed annotation and quality control tasks by following written guidelines. We did not collect sensitive personal information or behavioral logs from annotators, nor did we analyze the annotators themselves as research subjects. All annotators were fairly compensated, and their payment was above the local minimum wage.

Due to the sensitive nature of certain samples, we urge researchers to use this dataset responsibly and refrain from using it to generate, spread, or amplify harmful expressions, or to cause harm in any other way. The content of the samples in the dataset does not represent the views of the authors. Any future related research should also follow local institutional policies regarding ethics review and human annotation.

QuickLink

Arxiv link: https://arxiv.org/abs/2604.15841

Code link: https://github.com/csdq777/Mouse

✍️ Citation

If you use this dataset, please cite our work:

@article{lin2026exploring,
  title={Exploring the Capability Boundaries of LLMs in Mastering of Chinese Chouxiang Language},
  author={Lin, Dianqing and Lan, Tian and Zhu, Jiali and Li, Jiang and Chen, Wei and Liu, Xu and Su, Xiangdong and Hou, Hongxu and Gao, Guanglai and others},
  journal={arXiv preprint arXiv:2604.15841},
  year={2026}
}