← CHI 2026 map

CHI '26 · Best paper · full-paper review · confidence high

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

DOI PDF Program page Page JSON

GeoVisA11y is a notably strong CHI contribution because it turns map accessibility from passive description into active analysis. The system contribution is concrete, the evaluation is well matched to the claims, and the paper is unusually clear about both user benefits and present limits such as unsupported query types, backend gaps, and pipeline errors.

Video Figure

Axes Lens

Rare contribution shape, typical evidence profile. The point here is not a score. It is to show what kind of claim the paper makes, and whether the evidence pattern is unusual or baseline in this 268 -review set.

Review Summary

GeoVisA11y is impressive because it addresses a longstanding accessibility gap with a system that is more ambitious than typical alt-text or table-based accommodations. The central contribution is not just that blind and low-vision users can access map content, but that they can ask analytical, geospatial, visual, and contextual questions through a conversational interface tied to an interactive map. That matters because prior accessible geovisualization approaches often stopped at description, whereas this paper explicitly targets interpretation and analysis. The technical design is also meaningful rather than superficial: the authors combine a screen-reader-compatible UI with a custom QA pipeline and integrate geostatistical analysis, including Moran’s I and LISA, so some answers are grounded in structured spatial computation rather than only free-form language generation. The validation is appropriately matched to those claims. The study includes 12 participants across blind/low-vision and sighted groups, analyzes 346 questions, and reports 83.8% correct answers, while also examining how query patterns differ between groups. That empirical component strengthens the paper beyond a pure system demo because it shows both feasibility and behavioral differences that matter for design. Just as importantly, the paper does not hide its limits. It states that some questions were outside the supported taxonomy, some required calculations were not implemented, and some failures came from query refinement, scope assessment, hallucinations, ambiguity, and server errors. Those limitations narrow the deployment story, but they also make the contribution more credible. Overall, this is a strong artifact-and-evidence paper that advances accessible geovisual analytics in a concrete, reusable direction while clearly marking the boundaries of what the current prototype can and cannot do.

What Changed

Canon before

Before this work, accessible geovisualization support for blind and low-vision users was largely framed as descriptive access: alt text, data tables, and limited QA interfaces could help with map reading but not with richer spatial interpretation or exploratory analysis. Existing geovisualization QA systems also tended to rely on keyword matching or simpler rule-based handling, leaving complex spatial reasoning, contextual querying, and fluid navigation under-supported.

Departure from common sense

The paper departs from the common assumption that screen-reader users can only receive simplified descriptions of maps by showing that they can engage in complex spatial interpretation, analysis, and navigation through a natural-language geovisualization QA system. It also challenges the idea that blind and sighted users will interact with such systems in the same way, documenting distinct query and navigation patterns across groups.

Actual novelty

GeoVisA11y’s main novelty is an accessible geovisualization system that combines a screen-reader-compatible interactive map and chat interface with a custom QA pipeline integrating geostatistical analysis and LLM-based summaries. The system supports analytical, geospatial, visual, and contextual queries, including spatial pattern and outlier reasoning via Moran’s I and LISA, and it contributes empirical evidence about how blind/low-vision and sighted users differ in their querying behavior.

Evidence

The paper grounds its claims in both system description and user-study evidence. It specifies the accessible interface and QA pipeline, including geostatistical processing for spatial pattern questions. Validation comes from a mixed-methods study with 12 participants, 346 total questions, and reported answer accuracy of 83.8%, alongside analyses of query distributions and differences between blind/low-vision and sighted users. The paper also explicitly reports unsupported query types, unimplemented backend calculations, and pipeline failure modes.

Limits

Method limits

The evaluation is based on a relatively small study of 12 participants and task scenarios centered on two specific data-analysis activities. The system also does not answer all supported-looking questions reliably: some failures came from query refinement, scope assessment, misclassification, and GPT-related errors, while some requested calculations were not implemented in the backend.

Deployment limits

Deployment is currently bounded by U.S.-centric data preparation and infrastructure choices. The implementation uses 48 contiguous U.S. states and 3,222 counties with ACS data, and depends on a stack including MapboxJS, Flask, DuckDB, Google Cloud Platform, MongoDB, and Whisper-based speech input. The paper also notes missing support for visualization creation and recoloring requests.

Boundary conditions

The contribution is best understood as an accessible QA system for preprocessed geovisualizations where natural-language questions can be grounded in structured geographic data, map state, and statistical routines. Its strongest evidence concerns screen-reader and sighted users working on U.S.-focused tasks with the provided prototype; broader generalization to other geographies, datasets, or richer authoring/manipulation workflows is not established.

Position in field

This paper pushes accessible visualization research beyond descriptive accommodations toward interactive analytic access. Within CHI, it stands out by combining accessibility design, conversational interaction, and structured spatial computation, showing that accessible geovisualization can support substantive analysis rather than merely translating visual content into static summaries.

Abstract

Geovisualizations are powerful tools for communicating spatial information, but are inaccessible to screen-reader users. To address this limitation, we present GeoVisA11y, an LLM-based question-answering system that makes geovisualizations accessible through natural language interaction. The system supports map reading, analysis, interpretation and navigation by handling analytical, geospatial, visual and contextual queries. Through user studies with six screen-reader users and six sighted participants, we demonstrate that GeoVisA11y effectively bridges accessibility gaps while revealing distinct interaction patterns between user groups. We contribute: (1) an open-source, accessible geovisualization system, (2) empirical findings on query and navigation differences, and (3) a dataset of geospatial queries to inform future research on accessible data visualization.