When Headers Are Not There: Design and User Evaluation of an Automatic Topicalisation and Labelling Tool to Aid the Exploration of Web Documents by Blind Users
Jorge Sassaki Resende Silva, André Pimenta Freire, Paula Christina Figueira Cardoso · 2022 · Proceedings of the 19th International Web for All Conference (W4A 2022) · doi:10.1145/3493612.3520470
Summary
This paper addresses a fundamental problem for blind screen reader users: what happens when web documents lack proper heading markup, eliminating one of the primary navigation strategies available to them. The researchers designed and evaluated a tool that uses Natural Language Processing to automatically generate headings for unstructured web text. The system works in two stages: first, a topic segmentation algorithm based on SBERT (Sentence-BERT) processes the document by computing sentence embeddings, building a similarity matrix, and using Reynar's maximisation algorithm to detect topic boundaries where semantic similarity drops between adjacent sentences. Second, a topic labelling algorithm generates keyword-based headers for each segment by ranking words by frequency, grouping synonyms to avoid redundancy, and filtering by part-of-speech to retain only meaningful nouns, verbs, and adjectives. The tool was developed for Brazilian Portuguese texts using a multilingual BERT model. The researchers evaluated both the algorithm's performance against established benchmarks (C99, Reynar, TextTiling) and its practical impact through a user study with eight blind and partially-sighted screen reader users (ages 22-68, with 10-20 years of screen reader experience). Participants completed information-seeking tasks on four texts in two counterbalanced scenarios — with and without automatically generated headers — rating cognitive load after each task and providing qualitative feedback in post-test interviews.
Key findings
The segmentation algorithm outperformed previous approaches for longer documents (9-11 sentences per topic), which was the target use case, though it had higher error rates for shorter topics compared to the established C99 algorithm. Processing time scaled linearly with sentence count, taking about 10 seconds for a 1131-word text. In the user study, while differences in cognitive load (W=16, N=8, p=0.1) and task completion time (W=29, N=8, p=0.79) were not statistically significant with this small sample, descriptive data showed promising trends — mean cognitive load scores were lower with headers (average 3.8 vs 6.5 on a 10-point scale for some participants) and mean task times were shorter. Qualitative findings were compelling: all eight participants said they would use such a tool if available. Four of eight reported that headers facilitated task completion. Six of eight could infer topic content from the generated labels. Critically, even participants who did not directly use header navigation reported that topic segmentation itself made the text less exhausting to navigate, as it created landmarks and "checkpoints" for relocating information on a second read. One participant described the experience as "running his eyes through the text" — recreating a form of text skimming that is normally unavailable to screen reader users. Participants also offered constructive feedback: labels sometimes felt like disconnected word groups rather than meaningful phrases, and some suggested improvements like colour highlighting, zoom integration, and symbol definition lookups.
Relevance
This research tackles a pervasive real-world accessibility gap. Surveys show that heading navigation is the preferred strategy for 67.7% of screen reader users worldwide, yet many web pages either lack headings entirely or use them incorrectly for visual styling rather than document structure. The tool demonstrates that NLP can partially compensate for poor authoring practices, offering a client-side remediation strategy that does not depend on web developers fixing their markup. For accessibility practitioners, the study reinforces how critical proper heading structure is — the qualitative data vividly illustrates the cognitive and time costs blind users pay when headings are absent. The finding that even imperfect auto-generated headings provide meaningful navigation benefits suggests a pragmatic path forward: deploying approximate solutions now while continuing to push for proper semantic authoring. The study's limitations (small sample, remote testing during COVID, texts shorter than ideal, processing speed too slow for real-time plugin use) point to clear next steps for making this approach practical as a browser extension or screen reader plugin.
Tags: screen readers · headings · natural language processing · text segmentation · blind users · web accessibility · information seeking · cognitive load