UF Gators Win AmericasNLP 2026 Cultural Image Captioning Task
The University of Florida Gators clinched victory at the AmericasNLP 2026 shared task focused on generating captions for cultural images in Indigenous languages. They utilized a two-step method that first crafted a caption in Spanish with Qwen2.5-VL, then translated it into the target language using retrieval-augmented many-shot prompting with Gemini 2.5 Flash. Their results showed remarkable improvements: 164.1% for Bribri, 131.7% for Guaraní, and 122.6% for Orizaba Nahuatl, compared to the baseline on the development set. In the test set, both Bribri and Orizaba Nahuatl saw gains of over 150%. The team emphasized that retrieval methods vary significantly by language, and synthetic data augmentation contributed around 28 chrF++ to Guaraní's results. They finished second out of five in human assessments.
Key facts
- University of Florida Gators won AmericasNLP 2026 shared task
- Task: cultural image captioning for Indigenous languages
- Pipeline: Qwen2.5-VL for Spanish intermediate caption, then Gemini 2.5 Flash for target caption
- Dev set improvements: 164.1% (Bribri), 131.7% (Guaraní), 122.6% (Orizaba Nahuatl)
- Test set improvements >150% for Bribri and Orizaba Nahuatl
- Retrieval beneficial only for large, in-domain corpora
- Synthetic data augmentation contributed ~28 chrF++ for Guaraní
- Placed second out of five finalists in human evaluations
Entities
Institutions
- University of Florida Gators
- AmericasNLP
Locations
- United States