Leveraging Large Language Models for Alt-Text evaluation in e-commerce: a data-driven study

Generative artificial intelligence opens new opportunities for accessibility validation. An interesting case is the assessment of alternative text (alt-text) for images. This paper investigates the use of a large language model (LLM) to analyse alt-texts in real e-commerce websites, a domain in which images play an important role and have specific requirements. We present a novel data-driven method and an associated tool that employs tailored prompting strategies to incorporate contextual information when generating and evaluating image descriptions. The approach also supports systematic comparison between human-authored and LLM-generated alt-text. We conducted a user study (N = 16) involving 494 assessments of 157 images, and their corresponding alt-texts extracted from real e-commerce websites. The results show that the proposed solution can provide valid results, supporting its possible integration into existing accessibility validation workflows and authoring tools.