Improving Non-Visual Web Access Using Context
Jalal Mahmud, Yevgen Borodin, Dipanjan Das, I.V. Ramakrishnan · 2006 · Proceedings of the 8th International ACM SIGACCESS Conference on Computers and Accessibility (Assets '06) · doi:10.1145/1168987.1169051
Summary
This paper presents CSurf, a context-directed non-visual web browsing system that significantly reduces browsing time for blind users by intelligently rearranging page content based on navigational context. The core insight is that when a sighted user clicks a link, they can visually scan the destination page and quickly locate relevant content while ignoring banners, navigation bars, and other peripheral elements. Blind screen reader users, by contrast, must listen through content sequentially, and even with shortcut keys, they often wade through substantial irrelevant material before reaching what they need. CSurf addresses this by capturing the context around a link on the source page when a user follows it, then using that context to identify and prioritise relevant content on the destination page. The system architecture comprises five components: an Interface Manager using VoiceXML dialogs for keyboard-based interaction, a Browser Object built on Mozilla with JREX Java API, a Frame Tree Processor that extracts and partitions the page's rendering tree, a Context Analyzer that performs the core context identification and matching algorithms, and a Dialog Generator that converts the reordered frame tree into VoiceXML output.
Key findings
The context analysis algorithm operates in five phases: Context Identification extracts text around the clicked link and stores keywords in a multiset after removing function words; Context Ranking orders these words by proximity to the link; Context Matching compares the keywords against text in the destination page's frame tree leaves, assigning weights; Block Ranking propagates weights up the tree to identify semantically related content blocks; and Block Rearrangement reorders the frame tree so the highest-weighted block is presented first. Evaluation across 24 websites in four content domains showed browsing time reductions compared to JAWS of 63.7% for news sites, 57.9% for books, 45.2% for consumer electronics, and 54.4% for office supplies — even when JAWS users were allowed to use shortcut keys. The algorithm achieved over 80% accuracy in identifying relevant information across all domains, peaking at 88% for news sites, which tend to have better structural and geometric organisation. The technique works best on well-structured, information-driven, dynamic websites such as news engines, encyclopaedias, and online stores.
Relevance
This paper demonstrates that intelligent content processing can dramatically improve the screen reader browsing experience beyond what shortcut keys and skip links alone can achieve. The 45-64% reduction in browsing time across domains is substantial and addresses a core frustration of non-visual web access. The context-based approach — using what the user was looking at on the source page to predict what they want on the destination page — is an early example of intent-aware assistive technology that anticipates modern AI-driven accessibility tools. The geometric partitioning method is notable for being fully automated and domain-independent, unlike rule-based or domain-specific approaches. For practitioners, the research reinforces that well-structured, semantically organised HTML benefits not just basic screen reader navigation but also enables more sophisticated automated accessibility enhancements. The work laid groundwork for subsequent research on intelligent screen readers and content adaptation at Stony Brook University.
Tags: screen readers · non-visual access · web navigation · context analysis · web page segmentation · content rearrangement · blind users · browsing efficiency