Year: 2026
Type: article
Abstract: Despite the advances promoted by the Web Content Accessibility Guidelines (WCAG), images on the Web still frequently lack adequate alternative text, limiting access for people with visual impairments. This paper investigates the use of Multimodal Large Language Models (MLLM) for the automatic generation of image descriptions in journalistic contexts, using domain-specialist alt text authored by journalists as the reference baseline. We examine the impact of contextual information extracted from the page’s HTML and the effect of different word-length constraints on the generated descriptions. A case study was conducted with 73 images collected from the Brazilian news portal O Globo , for which descriptions were generated by two MLLM (Gemini and GPT) under multiple configurations. The generated descriptions were evaluated using BERTScore, CLIPScore, and RefCLIPScore, which measure alignment with both domain-specific alt text and the images’ visual content. The results show that imposing word limits increases semantic overlap with reference texts, whereas including contextual information has a limited, model-dependent effect. Overall, the descriptions produced by MLLM achieved similarity levels comparable to domain-specialist alternatives under textual and hybrid metrics, suggesting that MLLM constitute a promising baseline for large-scale deployment of alternative texts.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $3,190
Awards: G2526581899, G26019921, G6280542792
OpenAlex ID: W7221513620