CCUB Dataset: Cross-Cultural Understanding Benchmark

Overview

The Cross-Cultural Understanding Benchmark (CCUB) is a culturally curated dataset of text-image pairs designed to evaluate and improve cultural representation in text-to-image synthesis models. Each image and caption is validated by people with personal connections to the respective cultures to ensure authentic representation and avoid stereotypes.

Paper: Towards Equitable Representation in Text-to-Image Synthesis Models with the Cross-Cultural Understanding Benchmark (CCUB) Dataset
Published: AAAI Workshop on Creative AI Across Modalities, 2023
Authors: Zhixuan Liu, Youeun Shin, Beverley-Claire Okogwu, Youngsik Yun, Lia Coleman, Peter Schaldenbrand, Jihie Kim, Jean Oh

For fine-tuning applications using this dataset, see our follow-up work: SCoFT: Self-Contrastive Fine-Tuning for Equitable Image Generation

Dataset Structure

Data Schema

Field	Type	Description
country_code	String	3-letter country/region code (e.g., "KOR")
category_code	String	Cultural category (ARCH, CLOTH, CITIES, FOOD, etc.)
url	String	Direct URL to the image
caption	String	Culturally accurate description
created_at	Timestamp	ISO 8601 formatted date

Coverage

• Countries: 5 cultures including Korea, China, Nigeria, and others
• Categories: 9 categories (Architecture, Clothing, Cities, Food, Art, Festivals, Nature, Daily Life, etc.)
• Size: ~140 text-image pairs per culture
• Total: Approximately 700-1000 curated pairs

Example Entry

country_code: KOR
category_code: FOOD
url: https://upload.wikimedia.org/wikipedia/commons/2/22/Tteok-bokki_3.jpg
caption: a plate of rice cake tteokbokki covered in spicy sauce on a table
created_at: 2024-07-01T12:02:12Z

Usage

Accessing the Dataset

The dataset is organized in multiple sheets (one per country) in a single Google Sheets document.

Google Sheets URL: https://docs.google.com/spreadsheets/d/1dbSXgwCZuu9zpDZyUQ-pKDZGRYhzeaXCR5geRYNUwEA/

Download Methods

Option 1: Download All Sheets as ZIP

1. Open the Google Sheets link
2. Go to File → Download → Comma Separated Values (.csv)
3. This downloads all sheets as separate CSV files

Option 2: Download Individual Sheets

Each country has its own sheet with a unique gid:

import pandas as pd

# Sheet URLs for each country (replace gid values)
sheets = {
    'KOR': 'https://docs.google.com/spreadsheets/d/1dbSXgwCZuu9zpDZyUQ-pKDZGRYhzeaXCR5geRYNUwEA/export?format=csv&gid=1345336179',
    'CHN': 'https://docs.google.com/spreadsheets/d/1dbSXgwCZuu9zpDZyUQ-pKDZGRYhzeaXCR5geRYNUwEA/export?format=csv&gid=567535443',
    # Add other countries with their gid values
}

# Load a specific country
df_korea = pd.read_csv(sheets['KOR'])
df_china = pd.read_csv(sheets['CHN'])

# Or load all countries
all_data = []
for country, url in sheets.items():
    df = pd.read_csv(url)
    all_data.append(df)

df_all = pd.concat(all_data, ignore_index=True)

Applications

• Text-to-Image Model Evaluation: Benchmark cultural understanding and bias
• Fine-Tuning: Improve cultural representation in generative models (see SCoFT)
• Cultural Bias Research: Study representation in vision-language models
• Dataset Bias Analysis: Compare against web-scraped datasets

Dataset Curation

All images and captions are curated and validated by cultural insiders:

1. Selection: Images chosen by people with personal connections to each culture
2. Captioning: Native people (cultural experts) give caption the images
3. Validation: Cultural experts manually correct misinterpretations and errors
4. Quality Control: Focus on authentic representation, containing both traditional and modern representation, avoiding stereotypes

Citation

@misc{liu2023equitablerepresentationtexttoimagesynthesis,
      title={Towards Equitable Representation in Text-to-Image Synthesis Models with the Cross-Cultural Understanding Benchmark (CCUB) Dataset}, 
      author={Zhixuan Liu and Youeun Shin and Beverley-Claire Okogwu and Youngsik Yun and Lia Coleman and Peter Schaldenbrand and Jihie Kim and Jean Oh},
      year={2023},
      eprint={2301.12073},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2301.12073}, 
}

@inproceedings{liu2024scoft,
  title={SCoFT: Self-contrastive fine-tuning for equitable image generation},
  author={Liu, Zhixuan and Schaldenbrand, Peter and Okogwu, Beverley-Claire and Peng, Wenxuan and Yun, Youngsik and Hundt, Andrew and Kim, Jihie and Oh, Jean},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={10822--10832},
  year={2024}
}

Links:

• Paper: arXiv:2301.12073
• Dataset: Google Sheets
• Fine-tuning Method: SCoFT (arXiv:2401.00805)

Team

Carnegie Mellon University: Zhixuan Liu, Beverley-Claire Okogwu, Lia Coleman, Peter Schaldenbrand, Jean Oh
Dongguk University: Youeun Shin, Youngsik Yun, Jihie Kim

Contact

For questions or collaborations: github.com/cmubig/CCUB

License

Verify individual image licenses before use. Dataset compilation follows academic research guidelines.