CaMMT Dataset Card

CaMMT is a human-curated benchmark dataset for evaluating multimodal machine translation systems on culturally-relevant content. The dataset contains over 5,800 image-caption triples across 19 languages and 23 regions, with parallel captions in English and regional languages, specifically designed to assess how visual context impacts translation of culturally-specific items.

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("villacu/cammt")

# Load a specific split if available
dataset = load_dataset("villacu/cammt", split="ar_egy")

Dataset Details

Dataset Description

CAMMT addresses the challenge of translating cultural content by investigating whether images can serve as cultural context in multimodal translation. The dataset is built upon the CVQA (Culturally-diverse multilingual Visual Question Answering) dataset, transforming question-answer pairs into declarative caption statements. Each entry includes parallel captions in English and regional languages, with special attention to Culturally-Specific Items (CSIs) and their translation strategies.

The dataset includes both conserved translations (preserving original cultural terms) and substituted translations (using familiar equivalents) for items containing CSIs, along with native speaker preferences for translation strategies.

• Curated by: MBZUAI and collaborating institutions across the globe.
• Language(s) (NLP): 19 languages across 23 regions (Amharic, Arabic, Bengali, Bulgarian, Chinese, Filipino, Igbo, Indonesian, Japanese, Korean, Malay, Marathi, Oromo, Portuguese, Russian, Spanish (4 regional variants), Swahili, Tamil, Urdu (2 regional variants))
• License: Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)

Dataset Sources

• Paper: CAMMT: Benchmarking Culturally Aware Multimodal Machine Translation

Dataset Structure

The dataset contains 5,817 main entries plus an additional 1,550 entries with conserved and substituted CSI translations. Each entry includes:

• ID: Unique identifier from the original CVQA dataset
• regional: Caption in the regional language
• English: Parallel caption in English
• Conserved_translation: English translation preserving the original CSI (if applicable)
• Substituted_translation: English translation using a familiar equivalent for the CSI (if applicable)
• Category: Classification of cultural relevance:
• "not culturally-relevant sentence"
• "non-CSI" (culturally relevant but no specific CSI)
• "CSI- has possible translation" (CSI with cultural equivalent)
• "CSI-forced translation" (CSI without direct equivalent)
• Preferred_translation: Native speaker preference between conserved or substituted translation (if applicable)

The dataset spans 23 regions with varying numbers of samples per region (ranging from 200 to 369 samples).

Citation

BibTeX:

@misc{villacueva2025cammtbenchmarkingculturallyaware,
      title={CaMMT: Benchmarking Culturally Aware Multimodal Machine Translation}, 
      author={Emilio Villa-Cueva and Sholpan Bolatzhanova and Diana Turmakhan and Kareem Elzeky and Henok Biadglign Ademtew and Alham Fikri Aji and Israel Abebe Azime and Jinheon Baek and Frederico Belcavello and Fermin Cristobal and Jan Christian Blaise Cruz and Mary Dabre and Raj Dabre and Toqeer Ehsan and Naome A Etori and Fauzan Farooqui and Jiahui Geng and Guido Ivetta and Thanmay Jayakumar and Soyeong Jeong and Zheng Wei Lim and Aishik Mandal and Sofia Martinelli and Mihail Minkov Mihaylov and Daniil Orel and Aniket Pramanick and Sukannya Purkayastha and Israfel Salazar and Haiyue Song and Tiago Timponi Torrent and Debela Desalegn Yadeta and Injy Hamed and Atnafu Lambebo Tonja and Thamar Solorio},
      year={2025},
      eprint={2505.24456},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2505.24456}, 
}