NU_Internship team at ImageEval 2025: From Zero-Shot to Ensembles: Enhancing Grounded Arabic Image Captioning

Résumé

Arabic image captioning remains underexplored in vision-language research due to limited resources and the linguistic complexity of Arabic.In the ImageEval 2025 Shared Task, we evaluated three models, AIN, BLIP-Arabic-Flickr-8k, and Qwen 2.5, across zero-shot, fine-tuning, retrieval-augmented, and ensemble setups.Our official submission, fine-tuned BLIP with retrieval augmentation, ranked 5th overall based on both cosine similarity and LLM-as-a-judge scores.Post-submission experiments showed that ensemble captioning yielded the strongest captions across metrics.These findings demonstrate that even modest fine-tuning combined with retrieval augmentation can substantially improve Arabic captioning quality, which is significant in light of the limited resources for the language.

Citer ce document

Accès au document

Texte intégral en lecture en ligne, réservé aux abonnés SPHAERO et aux membres de l'institution. Se connecter

Auteur(s)

Statistiques

Consultations : 1

Téléchargements : 0