GeoVisA11y: An AI-based Geovisualization Question-Answering System for Screen-Reader Users

Paper Title

GeoVisA11y: An AI-based Geovisualization Question-Answering System for Screen-Reader Users

Publication Info

• Topic area: Accessible geovisualization using AI for screen-reader users.
• Keywords: geovisualization, accessibility, screen-reader, large language models, question-answering, spatial analysis, natural language interaction, inclusive design, geospatial data, visualization accessibility.

Background and Problem

• Problem / challenge: Geovisualizations are largely inaccessible to screen-reader users, with existing solutions limited to descriptive features like alt text or data tables, which fail to support deeper spatial analysis and interpretation.
• Significance: Making geovisualizations accessible is essential for enabling screen-reader users to independently analyze spatial data, which is critical in fields like urban planning, public health, and disaster response.
• Motivation and related work: Prior systems like AltGeoViz and VoxLens provide basic descriptive capabilities but lack support for complex spatial queries and interactive exploration. Existing QA systems for geovisualizations are limited to simple keyword-based queries. This paper addresses the gap by enabling advanced geospatial analysis and natural language interaction.

Solution

• Proposed approach: GeoVisA11y, an AI-based question-answering system that integrates geostatistical analysis with large language models (LLMs) to make geovisualizations accessible through natural language interaction.
• Novelty:
  1. Development of an open-source geovisualization system for screen-reader users.
  2. Empirical insights into query and navigation differences between blind/low-vision (BLV) and sighted users.
  3. Creation of a dataset of geospatial queries to inform future research on accessible visualization tools.
• Procedure and key techniques:
  • A structured QA pipeline with four components: Input Classifier, Query Refiner, Scope Assessor, and Query Processor.
  • Integration of geostatistical methods like Moran’s I and LISA for spatial pattern detection.
  • Bidirectional synchronization between a map interface and a chat interface for seamless interaction.
  • Iterative co-design process with blind participants to refine system features.

Results

• Concrete findings:
  • GeoVisA11y answered 92% of 346 user queries correctly, with 83.8% accuracy overall.
  • BLV participants rated the system highly for map navigation (median 6.5/7) and interpretation (median 6/7).
  • Sighted participants rated the system highly for map analysis (median 7/7) and reading (median 6/7).
• Advantage over baselines:
  • Supports complex spatial queries (e.g., patterns, relationships) beyond the capabilities of prior systems like VoxLens.
  • Enables both BLV and sighted users to identify similar spatial patterns, fostering shared understanding.
• Experiments / evaluation:
  • User study with 12 participants (6 BLV, 6 sighted) performing two analytical tasks: funding allocation based on underserved populations and identifying regional heating fuel patterns.
  • Quantitative and qualitative evaluation of system performance, user engagement, and query strategies.
• Limitations and future work:
  • Challenges in handling ambiguous queries and asymmetric navigation.
  • Need for more transparent reasoning, source attribution, and dynamic guided prompting.
  • Future work includes enhancing navigation, supporting direct manipulation, and scaling to diverse datasets and user needs.

Summary

GeoVisA11y is an AI-powered geovisualization QA system designed to make spatial data accessible to screen-reader users while also benefiting sighted users. By combining geostatistical analysis with LLM-based natural language interaction, the system enables users to read, analyze, interpret, and navigate geovisualizations. A user study demonstrated its effectiveness in supporting diverse query types and fostering shared understanding of spatial patterns. Future work will focus on improving navigation, transparency, and scalability to broader datasets and user groups.