CodedLang coded-language detector

Paste a Chinese review. The model estimates whether it hides meaning with homophones, orthographic tricks, phonetic spellings or ciphers. It runs entirely in your browser; nothing you type is sent anywhere.

Try an example

From the dataset's held-out split. Some contain disguised profanity.

Know its limits

• Chinese only. Thai, English or other languages give meaningless scores.
• It mostly recognises codes it saw in training. On codes it had never seen, it caught about half.
• It reads surface forms, not context: 辣鸡汉堡 (spicy-chicken burger) is flagged because 辣鸡 is also slang for 垃圾 "garbage".
• Scores are nearly always close to 0 or 1 and are not calibrated probabilities.
• This page uses an 8-bit compressed copy of the model. On the 763-review held-out test it scored F1 0.891 vs 0.885 for the full model (1 review decided differently).
• Research demo from a master's thesis. Do not use it to judge or moderate real people.

Model: nopp23/codedlang-coded-language-detector (fine-tuned hfl/chinese-roberta-wwm-ext) · Data: CodedLang (Wan, Li & Huang, ACL 2026, CC BY 4.0).