CulturalGround: Multilingual Visual Question Answering for Cultural Knowledge

CulturalGround is a large-scale Visual Question Answering (VQA) dataset developed by neulab and presented as an EMNLP 2025 Oral paper. It provides multimodal data designed to ground large language models in cultural knowledge across dozens of global regions and languages. The dataset was released in late 2025 to address cultural gaps in multilingual multimodal reasoning.

Use Cases

• Benchmarking cultural reasoning in multimodal LLMs using visual question-answer pairs
• Evaluating cross-lingual transfer in VQA tasks across languages like Greek, Hindi, and Arabic
• Fine-tuning models to recognize culture-specific artifacts or traditions from images

Strengths

• Covers over 15 distinct languages including Amharic, Bengali, Igbo, and Persian
• Peer-reviewed EMNLP 2025 Oral publication status
• Multimodal design specifically targeting cultural knowledge rather than general VQA

Limitations

• Specific record counts and file sizes are not explicitly detailed in the metadata
• Potential for uneven distribution of cultural representation across the 40+ flagged regions

Provenance

Source

neulab, EMNLP 2025 Oral paper

Collection Method

Pipeline-based generation and curation for cultural grounding

Freshness

Last updated October 2025, reflecting current research in multimodal cultural grounding.

Geography

Global (40+ countries indicated by flags including Japan, Ukraine, Nigeria, and Brazil)