Beyond LLMs: small multimodal models for semantic accessibility validation of alternative texts

Generative artificial intelligence is opening up new possibilities in Web accessibility authoring and validation. Multimodal language models can address semantic analyses of Web content that have not been provided so far. Indeed, most available accessibility validation tools are limited to syntactic analysis, focusing mainly on whether required elements are present. Some work has been put forward to explore the possibilities of large language models. However, such models raise several issues regarding cost, privacy, and sustainability. To address such issues, we have conducted a study investigating the feasibility of small, open-weighted models for the semantic analysis of a common accessibility issue: whether the alternative textual descriptions of images provide sufficient equivalent information (Success Criterion 1.1.1 Non-text Content). In the paper, we present how we have fine-tuned the small language model, and the results of a user test involving 30 experts in accessibility validation, in which they had to rate the results of the assessment performed by a large language model (GPT–4o), a small visual-language model and perform the same task themselves over 645 images. The results show that visual-language models can provide useful support for this type of task, and a small, appropriately trained language model tends to achieve even better results than a large one. These results show that novel paths for authoring and validating Web content accessibility can be enabled by small language models.