SADIe: exposing implicit information to improve accessibility
Darren Lunn · 2007 · Proceedings of the 9th International ACM SIGACCESS Conference on Computers and Accessibility (Assets '07) · doi:10.1145/1296843.1296904
Summary
Darren Lunn's Assets '07 demonstration paper introduces SADIe (Structural-Semantics for Accessibility and Device Independence), a transcoding system from the University of Manchester that uses Semantic Web ontologies to expose the implicit visual structure of web pages to screen reader users. The core insight is that most web pages carry substantial information in their CSS rendering — which element is a menu, which is a decoration, which is a search result — that assistive technology cannot access because it lives in the presentation layer rather than in the markup. SADIe addresses this with a two-layer ontology: an upper ontology of abstract role concepts (menu, decoration, header, search result) and a site-specific lower ontology that annotates the CSS classes of a particular site with their role. Once the lower ontology is in place, the system applies transcoding operations — 'Defluff' removes low-information decoration, 'Reorder' moves important regions to the top of the page, and 'Menu' surfaces navigation ahead of content. Because annotation is done at the CSS-class level rather than per-page, SADIe scales more efficiently than purely manual annotation systems while producing more accurate transcoding than purely heuristic systems. Evaluation used the ABA Reversal single-subject method with one visually impaired participant completing twenty fact-finding tasks across a mix of transcoded and unmodified pages drawn from a top-100-websites list. The evaluator and participant were both blinded to which pages had been transcoded, and differences were assessed using a randomization test — a non-parametric technique appropriate for small-n single-subject designs.
Key findings
Fact-finding tasks on unmodified pages took an average of 61.8 seconds, versus 21.1 seconds on SADIe-transcoded pages — a mean time saving of about 40 seconds per task, or roughly two-thirds faster. The randomization test gave p = 0.039, statistically significant even for a single-subject design. Two qualitative observations are arguably more important than the headline number. First, SADIe helped most on *unfamiliar* page types: the participant had never used a blog before and found the unmodified blog task frustrating, but completed the transcoded blog task relatively easily. Second, transcoding *familiar* pages can cause transient confusion — on modified search-result pages, the participant had to re-establish orientation using cues such as finding the word 'cached' before recognising the results area. This suggests the benefit of exposing implicit structure is real but depends on stable, predictable signposts (explicit headings like 'Web Results') rather than just simplification. The paper also notes the interesting search-strategy work done by the participant (looking for 'cached', 'headline') as a window into how screen reader users navigate pages in practice.
Relevance
SADIe is an early, tidy worked example of the accessibility-through-semantics agenda that WAI-ARIA later formalised: annotate the structural meaning of page regions so that assistive technology has something to work with beyond the bare HTML. The CSS-class-keyed annotation strategy anticipates modern selectors used by accessibility overlays and the component-library-level accessibility hooks used in React/Vue/Svelte today. For practitioners running accessibility-remediation or transcoding services, the paper's finding that simplification helps most on unfamiliar content and can confuse users on familiar content is directly useful: it argues for predictability and consistent landmarks over heavy-handed page rewriting. Limitations are severe by modern standards — n = 1 participant, 20 tasks, no longitudinal use, no comparison against screen-reader keyboard shortcuts or existing ARIA. The methodology section's use of randomization testing for single-subject designs is itself worth noting; it remains underused in accessibility research, where tiny participant samples are the norm. Recent systems with similar aims include server-side accessibility transcoders and overlays based on LLM-inferred DOM semantics, both of which repeat many of the design tensions SADIe identified.
Tags: web accessibility · transcoding · semantic web · ontology · CSS · screen readers · visual impairment · accessibility evaluation · randomisation test · single-subject design