Improving Accessibility to Statistical Graphs: The iGraph-Lite System
Leo Ferres, Petro Verkhogliad, Gitte Lindgaard, Louis Boucher, Antoine Chretien, Martin Lachance · 2007 · Proceedings of the 9th International ACM SIGACCESS Conference on Computers and Accessibility (Assets '07) · doi:10.1145/1296843.1296857
Summary
This paper from Carleton University and Statistics Canada presents iGraph-Lite, a system that makes statistical graphs accessible to blind and visually impaired people by generating natural language descriptions and providing an interactive navigation tool for exploring graph data. The system was developed to address the inaccessibility of graphs published in "The Daily," Statistics Canada's main daily publication containing statistical information about Canadian society and economy. The architecture is inspired by Pinker's cognition theory of how humans process graphs and comprises several distinct processing stages: a Microsoft Excel plug-in extracts graph data into a Visual Description (V_desc) — an acyclic graph encoding the physical properties of the graph including titles, axes, data series, values, and their relationships. Enrichment algorithms then identify patterns such as trends, reporting patterns, and statistical features (maxima, minima, increases, decreases). These are stored in a Knowledge Base (KB), and a natural language generation module using XSLT-based template transformations produces text descriptions. The system provides two access modes: longdesc-style static descriptions that summarise the graph's key information, and the iGraph Navigation Tool (iGNT) — a screen reader-inspired keyboard interface allowing users to navigate point-by-point through graph data, exploring at whatever granularity they choose.
Key findings
The iGNT navigation tool uses arrow commands to move through data points and a skip command to jump multiple points, with a "where" command providing current context (position, previous/next points, beginning/end of line). The natural language generation uses approximately 100 templates in its repository, with short texts and longer coherent descriptions available. The event-based template architecture makes it simple for non-programmers to create and modify descriptions. Pattern-matching algorithms identify reporting patterns (e.g., quarterly data with Roman numerals), trends (trend-line vs. no trend-line, one vs. multiple series), and automatically detect features like the visual "hump" in a 2001 graph that neither the website text nor the generated summary mentioned — demonstrating the navigation tool's value in revealing information invisible in text summaries. Early evaluation with 5 blind users and 2 sighted Statistics Canada analysts showed encouraging results, with users mostly answering questions correctly about graphs. Key design insights included that chronological order in descriptions was important but the title should come first for context, and starting with graph type (line, bar) before describing axes was preferred. The sub-language used was tailored for blind users, though multilingual support (English and French, per Canadian mandate) was straightforward to add.
Relevance
iGraph-Lite represents a significant approach to graph accessibility that combines automatic description generation with interactive exploration — addressing the fundamental limitation that static alt text can never fully capture all the information in a complex statistical graph. The two-mode access strategy (summary descriptions plus interactive navigation) is now recognised as best practice for data visualisation accessibility. For practitioners, the architecture demonstrates that effective graph accessibility does not require image recognition — by extracting data from the source application (Excel) rather than analysing rendered images, the system accesses the actual data values and relationships. The template-based NLG approach, while simpler than deep generation, proved sufficient and has the advantage of being easily extensible and localisable. The finding that navigation tools can reveal information that both the website text and the automated summary miss — like the 2001 data anomaly — powerfully argues that accessible graphs should support exploration, not just description.
Tags: data visualization accessibility · graph accessibility · natural language generation · screen readers · blindness · statistics · non-visual interaction · information accessibility