GeoNarrative: Transforming geospatial textual data into narrative storytelling visualization
Authors
Ruixian Ma (Massachusetts Institute of Technology), Biru Cao (Independent Researcher), Winston Yap (Massachusetts Institute of Technology), Jiaxin Gao (Independent Researcher), Wenzhe Peng (Independent Researcher), Takehiko Nagakura (Massachusetts Institute of Technology), Fábio Duarte (Massachusetts Institute of Technology), Paolo Santi (Massachusetts Institute of Technology), Carlo Ratti (Massachusetts Institute of Technology)
Presentation
- Session
- Story time
- Time
- Thursday, Nov 12, 08:12 – 08:24 (US/Eastern) · session 08:00 – 09:30
- Location
- Hall America north
Keywords
Narrative visualization, geospatial visualization, storytelling, large language models, automatic visualization generation, data provenance.
Abstract
Geographic narrative visualization integrates maps, charts, and explanatory text into cohesive spatial stories and plays an important role in journalism, urban planning, and public communication. However, creating these narratives still requires collaboration among GIS specialists, data analysts, and web developers, creating a high barrier to adoption. Existing visualization tools rely on static dashboards or rigid templates with steep learning curves, while large language models (LLMs) alone struggle to produce spatially grounded, visually coherent storytelling pages. To bridge this gap, we introduce GeoNarrative, an LLM-augmented system that transforms natural-language queries into multi-module narrative pages composed of text, maps, and charts. GeoNarrative includes three components: (i) an agent-based scheduling module for intent understanding and spatial data retrieval, (ii) a CLIP-based layout retriever that selects professionally inspired page templates, and (iii) a generative visualization engine that populates content while preserving narrative coherence. A built-in provenance mechanism lets users trace every data element to its origin, whether a named authoritative source or the generating model, supporting transparency and trust. We evaluate GeoNarrative through a controlled comparative experiment (N=16), an expert usability study (N=8), and a component-level ablation study. The results indicate that GeoNarrative produces narrative pages with stronger information flow and visual coherence than baseline methods, with large effect sizes, and that each architectural component contributes meaningfully to overall quality; given the modest sample sizes we regard these findings as promising rather than definitive. We also release an annotated dataset of 5,321 modules from 199 real-world narrative visualization pages to support future research.
For Practitioners
Data journalists and newsroom graphics teams; GIS analysts and urban planners; civic-tech and public-communication practitioners (e.g., in public health, transportation, and government agencies); data scientists and visualization designers who build geospatial storytelling or dashboard tools; and educators who teach data communication or geographic information science. Practitioners can use GeoNarrative's approach to go from a plain-language query to a publication-ready, multi-module geospatial narrative (maps, charts, and explanatory text) without assembling a full team of GIS specialists, developers, and designers, substantially lowering the time and skill barrier to producing spatial stories. Data journalists and public agencies can rapidly draft map-driven narratives for elections, mobility, or demographics and then refine them, while relying on the built-in data-provenance mechanism to trace every element back to an authoritative source or flag it as model-generated, which supports fact-checking and editorial trust. Tool builders can adopt the paper's intermediate-representation design, CLIP-based layout retrieval, and function-tag taxonomy as reusable components for their own AI-assisted authoring pipelines.