GeoVisA11y: An AI-based Geovisualization Question-Answering System for Screen-Reader Users
Chu Li, Rock Yuren Pang, Arnavi Chheda-Kothary, Ather Sharif, Henok Assalif, Jeffrey Heer, Jon E. Froehlich · 2026 · Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26) · doi:10.1145/3772318.3790334
Summary
This CHI 2026 paper tackles a long-neglected corner of accessible data visualization: geovisualizations — choropleth maps, dot density maps, and similar spatially encoded displays that pack demographic, environmental, and public-health information into visual layouts that screen-reader users cannot meaningfully access. While prior work has produced alt-text generators like AltGeoViz and keyword-matching QA tools like VoxLens, the authors argue these support "map reading" (reporting values) but not "map analysis" or "map interpretation" (recognizing spatial patterns, relationships, and geometric characteristics). The authors present GeoVisA11y, an open-source LLM-based question-answering system built on GPT-4o-mini that makes interactive geovisualizations accessible through natural-language interaction synchronized with a screen-reader-compatible map. The system's four-component pipeline — Input Classifier, Query Refiner, Scope Assessor, and Query Processor — classifies user input as map action or information query; resolves deictic references ("here", "this state") and pronoun ambiguity using current map focus and chat history; decides whether queries can be answered from the local DuckDB database or must fall back to GPT's general knowledge; and routes supported queries through five primary categories (analytical, geospatial, visual, contextual, map action) with 14 detailed query types. Geospatial pattern queries use Moran's I and Local Indicators of Spatial Association (LISA) to identify statistically significant clustering. Keyboard-based arrow-key navigation lets screen-reader users discretely step between cardinally adjacent states and counties, with status announcements on every focus change. Evaluation involved 12 participants (6 screen-reader users B1-B6, 6 sighted users S1-S6) completing two analytical tasks adapted from real data stories: distributing federal Digital Equity Planning Grant funding across states, and identifying regional heating-fuel patterns on choropleth and dot-density maps. Sessions lasted 90-120 minutes and produced 346 queries.
Key findings
GeoVisA11y supported 92% of the 346 questions asked, with 83.8% answered correctly. Query-type distribution converged across groups on three dominant categories: General Knowledge (25.8%), Pattern (22.6%), and Action (16.2%) queries. Usage diverged in illuminating ways: BLV users more often asked Retrieve (14.7% vs. 5.1%), Find Extremum (8.0% vs. 4.6%), Shape (3.3% vs. 0.5%), and Spatial Relationship (5.3% vs. 0%) queries — the latter two asked almost exclusively by BLV users to build their mental map of unfamiliar geography. Sighted users favored Aggregate (8.2% vs. 2.0%) and Sort (5.1% vs. 2.0%), leveraging their visual pattern recognition before querying for precise statistics. Both groups successfully identified similar patterns in the data — converging on Mississippi, Alabama, Louisiana, Arkansas, Georgia, West Virginia, and Texas for broadband funding — suggesting GeoVisA11y produced a shared understanding of geospatial patterns across visual abilities. On 7-point Likert ratings, BLV participants rated the chat component 7/7 and map analysis 5.5/7; sighted participants rated the map 7/7 and map interpretation 5/7. Error analysis of the 56 incorrect answers traced failures to Query Refiner (26.8%, mostly failing to catch focus state), Scope Assessor (21.4%, misrouting to GPT vs. local data), and requests for features beyond current scope (46.4%, e.g., creating new visualizations). Notable qualitative findings: BLV users showed strong trust in system responses but wanted source attribution and hyperlinks to verify; sighted users valued how natural-language access prevented misinterpretation of choropleth color scales; BLV users described map navigation as unprecedented autonomous spatial exploration, with one noting the experience contrasted sharply with passive tactile-graph interpretation from school.
Relevance
For accessibility practitioners, this paper is one of the first concrete demonstrations that LLM-based conversational interfaces can move data-visualization accessibility beyond alt text and static data tables into genuine analytical exploration — a gap the community has discussed for years but rarely closed. The clean decomposition of geovisualization tasks into reading, analysis, interpretation, and navigation provides a reusable evaluation framework for future accessible visualization work, and the query taxonomy (with real BLV query distributions) is a useful design input for anyone building accessible chart-QA tools. The paper also demonstrates the universal-design payoff: sighted users rated GeoVisA11y at median 6/7 for additional benefit over viewing the map alone, validating that accessibility features benefit broader populations. Limitations are significant: only 6 screen-reader users (all fully blind, no low-vision), reliance on GPT with documented hallucination risk, no explicit uncertainty communication between locally verified and LLM-generated answers, and a U.S.-only dataset. The authors are transparent about these, and the open-source release at github.com/makeabilitylab/geovisa11y makes the work tractable for extension to other domains.
Tags: geovisualization · data visualization · accessible visualization · screen readers · large language models · natural language interfaces · blind and low vision · choropleth maps · spatial analysis
Standards referenced: WCAG 2.1