AI link prediction helps screen readers understand "Read more" buttons

A new registered report from Kishan Rakesh and Shiyi Wei (University of Texas at Arlington) tackles a persistent web accessibility problem: screen readers often present elements like "Read more" or "Click here" that are technically valid under W3C specs but meaningless in isolation. The surrounding context exists on the page but isn't programmatically linked to the target. The study proposes treating the accessibility tree as a graph and using link prediction—a technique common in social networks—to recover those contextual associations.

In a pilot on 5 author-annotated websites, four machine learning models (MLP, GCN, GAT, and SEAL) were compared against heuristic baselines based on spatial and DOM proximity. The best model (GAT) achieved Hit@10 of 0.85, while baselines maxed at 0.30. The full study will expand to 35 websites stratified from the Tranco top-million list, with three independent annotators per page (105 annotations total). Each page will be represented as an accessibility tree graph augmented with spatial and semantic features from DOM and CSS. Models will be evaluated under leave-one-site-out cross-validation using Hit@K and Mean Reciprocal Rank. If successful, this approach could automatically generate meaningful labels for screen readers, dramatically improving navigation for blind and visually impaired users.