ChinaHeritaQA: A Culturally-Grounded Visual Question Answering Dataset for World Heritage Sites in China
Researchers have introduced ChinaHeritaQA, a new dataset designed to test the cultural reasoning capabilities of vision-language models (VLMs). The dataset includes over 2,000 images of Chinese World Heritage sites, paired with more than 14,000 bilingual questions covering various cognitive dimensions. Initial evaluations show that while current top VLMs perform well on visual recognition tasks, they struggle with deeper cultural and historical understanding, indicating a gap in their ability to process culturally grounded information. AI
IMPACT This dataset highlights current limitations in AI's cultural and historical understanding, potentially guiding future research in culturally aware multimodal learning.