ERI-AIC arabic image captioning dataset: development and performance evaluation

Abstract

Abstract This paper presents ERI-AIC, an Arabic image captioning dataset designed to better reflect Arab cultural contexts and the informational needs of visually impaired individuals. To the best of our knowledge, ERI-AIC is the first dataset to provide original, human-written Arabic captions created from scratch rather than translated from English, while also being guided by a user-study-based understanding of the image description needs of visually impaired users. The dataset was constructed in two main stages. First, we conducted a questionnaire study to identify the interests, preferences, and essential description elements required by people with visual impairments. Second, we developed ERI-AIC by collecting 4,000 images and 12,000 original Arabic captions, with three captions per image, written directly in Arabic to describe culturally relevant scenes and accessibility-oriented details. To evaluate the usefulness of ERI-AIC, we benchmarked baseline captioning models and compared their performance with prior Arabic image captioning resources. Within the reported experimental setting, models trained on ERI-AIC generally achieved higher captioning scores than the corresponding models trained on Arabic Flickr8k and Arabic MS-COCO, with the Transformer model achieving the highest ERI-AIC performance (BLEU-1: 65.1, BLEU-4: 21.4, METEOR: 0.43, CIDEr: 0.38, ROUGE: 0.39). Although the Encoder-Decoder architecture reports unusually high scores on Arabic MS-COCO, this advantage does not transfer to ERI-AIC, suggesting that performance differences are influenced by dataset characteristics, caption style, and model architecture. Overall, the findings indicate that ERI-AIC can serve as a useful training and evaluation resource for culturally grounded and accessibility-oriented Arabic image captioning, while broader claims about generalization require further systematic cross-dataset quantitative evaluation.