BZU-AUM@ImageEval2025: An Arabic Image Captioning Dataset for Conflict Narratives with Human Annotation
Year: 2025
Type: conference-paper
Abstract: This paper presents a new Arabic image captioning dataset created for the ImageEval 2025 Shared Task.The dataset focuses on images related to conflict, resistance, and everyday life under occupation.Each image is paired with a Modern Standard Arabic caption of 40-70 words that describes what is shown and adds cultural or emotional context.To help annotators write rich and consistent captions, we used prompt-based guidelines, including stepby-step reasoning and writing from specific roles such as journalists or humanitarian observers.This method produced captions that are both descriptive and meaningful.The dataset fills an important gap in Arabic resources, especially for sensitive and historically significant topics.It can be used to train and evaluate Arabic vision language models, test multilingual AI systems, and support applications in journalism, education, and cultural preservation.
Institutions: Birzeit University, American University of Madaba
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4416037667