A Web Navigation Tool for the Blind
Mary Zajicek, Chris Powell, Chris Reeves · 1998 · Proceedings of the Third International ACM Conference on Assistive Technologies (Assets '98) · doi:10.1145/274497.274534
Summary
This short paper from Oxford Brookes University and the Royal National Institute for the Blind (RNIB) presents BrookesTalk, a prototype speech-output web browser designed to help blind users make rapid decisions about whether a web page is useful to them. The authors identified a gap in existing screen readers and accessible browsers like JAWS and pwWebSpeak: while these tools allowed users to navigate pages and read text, they became unwieldy for the quick information-seeking and triage tasks that characterise real web browsing. BrookesTalk addressed this by applying information retrieval techniques to generate multiple complementary summaries of each web page. Users could access these via a virtual toolbar driven by function keys, choosing from: a list of headings, a list of links, automatically extracted keywords, a page summary (including word count, number of headings, and number of links), abridged text generated from key trigrams (three-word phrases), bookmarks, and a scratchpad for saving interesting sentences. The keyword extraction used established information retrieval methods, while the abridged text feature identified high-scoring trigrams and assembled the sentences containing them into a summary averaging 20% of the original text length.
Key findings
A preliminary evaluation with 20 subjects compared three page summary representations — headings, links/anchors, and keywords — across six web pages. Headings scored highest (58.09% of available marks), followed closely by keywords (55.28%), with links scoring lowest (45.24%). The difference between headings and keywords was not statistically significant, validating the keyword approach as a useful complement especially when page headings are missing, uninformative, or represented as images without alt text. Testing at RNIB with blind workers revealed several practical insights: users tended to rely heavily on one summary function rather than switching between views as the designers intended; one user unexpectedly built his mental model of pages primarily by navigating links (90% of the time), orienting himself by understanding what the page linked to rather than what it contained. The automated summary function received the most criticism — trigram analysis frequently selected the wrong phrases as significant, and important headings were often omitted. Users suggested weighting trigrams that appear in page titles or headings. RNIB users valued the tool for managing their daily information workload and noted that the HTML-to-speech translation methods could be applied to other formatted documents.
Relevance
This paper is an early example of applying natural language processing and information retrieval techniques to accessibility — an approach that has since become mainstream with AI-powered accessibility tools. The core problem it addresses remains highly relevant: blind web users still need efficient ways to triage pages and decide quickly whether content is worth reading in full. The finding that users develop their own unexpected strategies for page orientation (like the user who navigated primarily by links) is a valuable reminder that designers cannot fully predict how assistive tools will be used in practice. The concept of multiple complementary "views" of a page — structural (headings), relational (links), and content-based (keywords, summary) — anticipates features that modern screen readers and reading tools still struggle to provide well. For practitioners, the paper reinforces the importance of meaningful headings and well-labelled links, since these are the primary orientation tools blind users depend on.
Tags: web accessibility · blindness · information retrieval · screen readers · text summarization · nonvisual web access · web browsers