RuyuanWan/CodedLang

Dataset

We introduce CodedLang, a dataset of 7,744 Chinese Google Maps reviews, including 900 reviews with span-level annotations of coded language.

Taxonomy

We define 7 categories of coded language:

• Ambiguous Homophones
• Non-Lexical Homophones
• Phonetic Substitution
• Emoji Substitution
• Orthographic Substitution
• Cross-Lingual Phonetic Encoding
• Cipher

Data Sources

Our dataset is constructed based on two large-scale Google Maps review datasets:

• Google Local Reviews (Li et al., 2022, Yan et al., 2023): 666 million reviews
• Google Restaurant Reviews (Yan et al., 2023): 1.77 million reviews

Annotations

Each sample includes:

• id: Index of the review
• original_review: Original Chinese review text
• translated_review: English translation of the original review
  • Note: 309 reviews are missing translations from the raw Google Reviews dataset and are kept as-is
• rating: Review rating (1–5 stars)
• char_mask_review: Character-level masked version of the review
• span_mask_review: Span-level masked version of the review
• decode_review: Decoded version of the review, where coded language is replaced with the intended meaning.
• coded_lang_class: Taxonomy label(s) for coded language classes
• code_span: Text span of coded expressions
• coded_language: Binary label
  • 1: contains coded language
  • 0: non-coded review

---

Citation

If you find this work useful for your research, please cite our paper:

@inproceedings{wan2026newspaper,
  title={"Newspaper Eat" Means "Not Tasty": A Taxonomy and Benchmark for Coded Languages in Real-World Chinese Online Reviews},
  author={Wan, Ruyuan and Li, Changye and Huang, Ting-Hao'Kenneth'},
  booktitle={The 64th Annual Meeting of the Association for Computational Linguistics},
  year={2026}
}