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

TL;DR

This work presents GeoVisA11y, an LLM-based question-answering system that makes geovisualizations accessible through natural language interaction, and contributes an open-source, accessible geovisualization system and empirical findings on query and navigation differences.

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 12 screen-reader users and 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.

Figures (9)

• Figure 1: GeoVisA11y UI overview. (A) Interactive map component. (B) Chat component where users can ask questions and receive answers through text or voice input. (C) Welcome section with an overview of the visualization, example questions, and shortcut key instructions. (D) Status indicator that announces changes to screen-reader users, e.g., "Map interaction enabled. Focused on Kansas." or "Chat interaction enabled. Type a question to ask GeoVisA11y." (E) Legend of the current interactive map.
• Figure 2: System pipeline diagram. All four components use GPT-4o-mini via few-shot prompting. (1) The Input Classifier separates map action commands from information queries. (2) The Query Refiner resolves ambiguities (e.g., replacing "here" with "Ohio", "it" with "population density") by ingesting current map focus and chat history. (3) The Scope Assessor determines if the query can be answered from the local database. (4) The Query Processor classifies the query and triggers relevant operations, then updates the Chat and/or Map interface.
• Figure 3: Pipeline workflow for spatial pattern queries. When users ask "Is there a pattern on the map?", the system runs global Moran's I analysis to detect significant patterns, then performs LISA cluster analysis if patterns exist. LISA outputs are summarized by GPT with representative examples, while the map displays colored outlines for cluster types. Color highlights in the answer text indicate which part of the pipeline produced that information: the red highlight on "clustered pattern" corresponds to the Moran's I result, while orange, yellow, and green highlights correspond to specific cluster types (HH, LL, LH) identified by the LISA analysis.
• Figure 4: Interactive map keyboard navigation. (A) Ctrl + M activates the map interface. (B) TAB focuses on an initial state, then arrow keys navigate to adjacent states in cardinal directions. (C) Example navigation from Kansas to Nebraska using the up arrow. (D) + and - keys zoom between county and state levels, shown here focused on Cuming County, Nebraska.
• Figure 5: AI chat key interactions. (A) Ctrl + M activates the chat if previously focused on the map. (B) Upon receiving a query, system repeats the user query followed by a status indicator. (C) The three example questions can be refreshed through Ctrl + I. (D) Asking "What else can you do?" can see/hear all the available query types.