Machine Learning for Accessible Web Navigation
Tlamelo Makati · 2022 · Proceedings of the 19th International Web for All Conference (W4A) · doi:10.1145/3493612.3520463
Summary
This extended abstract explores how machine learning techniques, particularly reinforcement learning, can be adapted to improve web navigation accessibility in alignment with WCAG Guideline 2.4 (Navigable). The author observes that ML techniques have already been applied to website navigation for goal-directed tasks like answering questions or completing workflows such as booking flights, but accessibility has not been their primary focus. These approaches typically model websites as state spaces represented by directed graphs, where actions like clicking links, filling forms, and pressing buttons move an agent between states. Reinforcement learning uses reward signals to learn optimal navigation policies through these large state spaces. The paper asks whether these same techniques can be repurposed specifically to improve navigability from an accessibility perspective. Proposed applications include automatically generating meaningful anchor text for links that lack descriptive labels, optimizing pathways to website functionality, and improving search and query processes. The author also raises the intriguing question of whether accessibility features themselves — headings, anchor text, alternative text — could be incorporated into reward calculations, making the ML agent actively leverage and value good accessibility practices. The paper presents a case study using DocTTTTTQuery, a sequence-to-sequence model (built on T5) that generates potential queries from document content, effectively expanding documents with predicted user questions to improve search retrieval.
Key findings
The case study applied DocTTTTTQuery to the Technological University Dublin website, generating predicted queries from page content. The system produced queries of varying quality, demonstrating both the potential and current limitations of the approach. The generated queries could be matched against historical user queries from web analytics to validate whether the technique improves navigation efficiency. The author identifies several promising research directions: using SEO metadata and information science indexing techniques to enrich state representations for better reward calculation; incorporating accessibility-specific features (headings, alt text, link text) into reinforcement learning reward functions; applying ML classification to identify priority areas on web pages; and using W3C Personalization Semantics to enhance content presentation based on ML-identified page structure. The work is still in early stages, with plans to filter low-quality generated queries, incorporate richer HTML structural information beyond plain text, and compare generated queries against actual user search behavior.
Relevance
This research opens an innovative direction for accessibility practitioners and researchers by framing web accessibility improvement as a machine learning optimization problem. The idea of incorporating accessibility features into reinforcement learning reward functions is particularly compelling — it suggests a future where ML agents could automatically identify and remediate navigation barriers by treating good accessibility as inherently rewarding. For practitioners, the document expansion technique (DocTTTTTQuery) offers a concrete near-term application: enhancing website search by predicting what users are likely to ask, which directly supports WCAG 2.4 success criteria around findability and navigation. The work is at an early stage with limited empirical results, and the case study is preliminary. However, the conceptual framework connecting reinforcement learning state-space exploration to accessibility barrier detection is valuable. The author also plans to apply findings to projects with people with intellectual disabilities using co-design methods, which would ground this technical work in real user needs.
Tags: machine learning · web accessibility · reinforcement learning · web navigation · query optimization · WCAG · document expansion · information retrieval
Standards referenced: WCAG 2.1 · EN 301 549 · Personalization Semantics 1.0