"Newspaper Eat" Means "Not Tasty": A Taxonomy and Benchmark for Coded Language in Real-World Chinese Online Reviews

This repository provides datasets and code for the ACL 2026 long paper on Coded Language.

"Newspaper Eat" Means "Not Tasty": A Taxonomy and Benchmark for Coded Language in Real-World Chinese Online Reviews
Ruyuan Wan, Changye Li, Ting-Hao 'Kenneth' Huang
ACL 2026

Overview

Example of creative coded language in a Google Maps restaurant review.
Top: the original user review in Chinese, as posted on Google Maps.
Left: the machine-generated English translation produced by Google Translate.
Right: the inferred underlying meaning obtained by decoding phonetic substitutions.

CodedLang Dataset

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

🤗 Hugging Face Dataset: https://huggingface.co/datasets/RuyuanWan/CodedLang

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

Coded Language Dictionary

We provide a coded language dictionary derived from human annotations:

Dictionary Fields

Each entry includes:

• code_span: Coded expression
• decode: Decoded form
• coded_lang_class: Corresponding coded language category

For phonetic-based coded language, we additionally provide:

• mid_homophone: Intermediate homophone form (if applicable)
• code_pinyin: Pinyin of the coded expression
• decode_pinyin: Pinyin of the decoded expression
• code_ipa: IPA representation of the coded expression
• decode_ipa: IPA representation of the decoded expression

Benchmark Results

We evaluate language models on taxonomy-aware coded language classification.

Multi-label coded language classification performance across categories by DeepSeek-V3.2. The highest (lowest) scores in each column are highlighted in bold (underlined).

Results show substantial performance variation across coded language categories, cross-lingual phonetic encoding remaining particularly challenging for current language models.

For full benchmark results and analyses, please refer to the paper.

---

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}
}