# CULTURE-MT: Beyond Literal Translation โ€” Evaluating Cultural Effectiveness in Social Media UGC [![Leaderboard](https://img.shields.io/badge/๐Ÿ†-Leaderboard-yellow)](https://huggingface.co/spaces/Wulinjuan/CULTURE-MT) [![License](https://img.shields.io/badge/License-CC%20BY%204.0-green)](https://creativecommons.org/licenses/by/4.0/) CULTURE-MT is a benchmark for evaluating **CUL**tural **T**ransmission and **U**GC-specific emotion **RE**sonance in Chinese-to-English social media translation. It consists of 1,002 user-generated notes (UGC) spanning 14 content domains, presented at ICML 2026. ๐Ÿ“„ **Paper:** [Beyond Literal Translation: Evaluating Cultural Effectiveness in Social Media UGC](https://arxiv.org/abs/2605.25626) --- ## ๐ŸŒŸ Why CULTURE-MT? Standard machine translation metrics (BLEU, ChrF, COMET) fail to capture whether a translation truly resonates with target-language users. CULTURE-MT introduces **cultural effectiveness** as a new evaluation criterion, covering: - **Expressive accuracy** โ€” semantic fidelity, emotional tone, proper noun handling, and unit/measurement accuracy - **Cultural adaptability** โ€” culture-loaded term handling, overall cultural fluency, and addressing/politeness adaptation --- ## ๐Ÿ“Š Dataset | Split | Notes | Domains | Note Types | |-------|-------|---------|------------| | Benchmark | 1,002 | 14 | 4 (General, Express, Symbol, Hybrid) | ### Content Domains Pets ยท Travel ยท Food ยท Crafts ยท Painting ยท Home Decoration ยท Outdoor ยท Sports ยท Fitness & Weight Loss ยท Technology & Gadgets ยท Cars ยท Games ยท Movies & TV ยท Celebrity News ### Note Types | Type | Description | |------|-------------| | **General** | Informal UGC with few culture-loaded symbols or distinctive styles | | **Express** | Strong rhetorical/expressive style (e.g., "planting grass", hyperbole, rhetorical questions) | | **Symbol** | High density of internet cultural symbols (slang, memes, platform-specific jargon) | | **Hybrid** | Both rich cultural symbols and distinctive expressive style | --- ## ๐Ÿ“ Evaluation Translations are evaluated by **JUDGER**, a fine-tuned Qwen3-32B model trained on 30K expert- and LLM-annotated samples. It achieves 86.03% accuracy and Cohen's ฮบ = 0.72 against human expert judgments. ### Scoring Rubric (0โ€“3 scale) | Score | Meaning | |-------|---------| | **0** | Severe meaning loss or distortion; target readers cannot grasp the original intent or emotion | | **1** | Main idea barely understandable; critical cultural errors, poor adaptation | | **2** | Main information conveyed accurately; reasonable emotional/contextual expression | | **3** | Precise, natural, culturally fluent; fully conveys all information and emotion for English social media readers | Scores 0โ€“1 are treated as **culturally ineffective**; scores 2โ€“3 as **culturally effective**. --- ## ๐Ÿ† Leaderboard Submit your translations and get evaluated automatically by our trained **JUDGER** model at: ๐Ÿ‘‰ https://huggingface.co/spaces/Wulinjuan/CULTURE-MT > **๐Ÿ“Œ Note on Leaderboard vs. Paper Results** > Scores reported in the ICML 2026 paper were produced with a single JUDGER inference pass (temperature = 0.6, top-p = 0.95). Since stochastic decoding introduces minor variation across runs, leaderboard scores are computed as the **average of four independent inference passes** under identical settings to ensure fairness and reproducibility. As a result, leaderboard scores may differ slightly from those reported in the paper. ### Top Results (as of 2026-05-27) | Rank | Model | Ineff. โ†“ | Eff. โ†‘ | 0 โ†“ | 1 โ†“ | 2 | 3 | Avg Score | |------|-------|----------|--------|-----|-----|---|---|-----------| | 1 | GPT-5 | 8.08% | 91.92% | 0.90% | 7.19% | 51.56% | 40.35% | 2.31 | | 2 | Gemini 3 Pro | 8.95% | 91.05% | 0.20% | 8.75% | 51.97% | 39.09% | 2.30 | | 3 | CULTURE-MT-baseline-32B | 12.08% | 87.92% | 0.73% | 11.35% | 56.56% | 31.37% | 2.19 | | 4 | CULTURE-MT-baseline-8B | 14.34% | 85.66% | 1.27% | 13.07% | 57.42% | 28.24% | 2.13 | | 5 | GLM4.6 | 15.94% | 84.06% | 3.12% | 12.81% | 57.39% | 26.68% | 2.08 | | 6 | DeepSeek V3.2 | 17.73% | 82.27% | 2.59% | 15.14% | 56.82% | 25.45% | 2.05 | | 7 | Qwen3-235B-A22B | 29.27% | 70.73% | 12.31% | 16.97% | 52.86% | 17.87% | 1.76 | | 8 | Seed-X-PPO | 30.58% | 69.42% | 3.43% | 27.25% | 62.23% | 7.19% | 1.73 | --- ## ๐Ÿ“ค How to Submit Submissions are made directly through the leaderboard interface at: ๐Ÿ‘‰ https://huggingface.co/spaces/Wulinjuan/CULTURE-MT ### Step 1 โ€” Upload your translations Upload a `submission.jsonl` file. Each line should be a JSON object with the `id` of the source note and your `translation`: ```json {"id": "0001", "translation": "This place is amazing. I'll definitely come back!"} {"id": "0002", "translation": "This is way too ridiculous."} ``` ### Step 2 โ€” Fill in model information Complete the submission form with your model details (Model Name, Organization / Team, Base Model, Method, and a brief description). No additional files are required. ### Step 3 โ€” Get your results Aggregated scores will appear on the leaderboard automatically after evaluation. If you need detailed per-sample evaluation results, please contact us by email at **wulinjuan525@zju.edu.cn**. --- ## ๐Ÿ“– Citation If you use CULTURE-MT in your research, please cite: ```bibtex @misc{wu2026literaltranslationevaluatingcultural, title={Beyond Literal Translation: Evaluating Cultural Effectiveness in Social Media UGC}, author={Linjuan Wu and Ruiqi Zhang and Xinze Lyu and Ye Guo and Daoxin Zhang and Zhe Xu and Yao Hu and Yixin Cao and Yongliang Shen and Weiming Lu}, year={2026}, eprint={2605.25626}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2605.25626}, } ``` --- ## ๐Ÿ“ฌ Contact - Linjuan Wu: wulinjuan525@zju.edu.cn This benchmark was developed with support from Zhejiang University, Fudan University, and Xiaohongshu Inc.