Source-linked AI summary
Beyond Initial Creation: Two Decades of Community-based Chart Revisions from Wikimedia Commons
Jasmine Y. Shih, Chase Stokes, Melanie Tory, Marti Hearst
TL;DR
Community editing of chart images is poorly understood, despite open communities’ role in maintaining public-facing visualizations. The paper builds a labeled dataset of Wikimedia Commons chart revisions and analyzes patterns by contributor count, editing cadence, and file format. Its analyses find differences in edit types and timing, and tighter-scoped revisions for SVG charts than raster charts.
Problem
Community editing of chart images is little studied, and synthetic chart-edit datasets do not capture how charts evolve in unstructured open environments.
Source: §1 · p. 1Method
The paper constructs a longitudinal Wikimedia Commons chart-revision dataset, labels edits with an LLM pipeline, and analyzes patterns across contributors, cadence, and file format.
Source: §1 · p. 1, §3.2 · p. 2, §4 · p. 3Results
Takeaways & Limitations
The dataset and initial analysis characterize how an open community revises visualization content and design for public communication.
Source: §5 · p. 4The sample includes charts present in Wikipedia articles at collection time, misses charts since removed, and omits some chart attributes that the LLMs could not reliably label.
Source: §5 · p. 4
Abstract
from arXiv · showhide
Open online communities like Wikipedia offer a rich space for understanding the organic evolution of public-facing artifacts, but little is known about how these communities collectively edit chart images. We investigate the phenomenon of "community editing" of visualizations by exploring chart revision patterns on Wikimedia Commons. We introduce a novel dataset of 7,786 longitudinal chart revisions across 1,664 images, labeled with edit categories using an LLM pipeline. Initial data analysis uncovered several novel findings. Multi-contributor revision is associated with more edits to mark styling and text content compared to single-contributor revision. Temporally, mark and text edits occur in more "bursty" cadences than data edits. In addition, while SVG charts attract more unique contributors and revisions than raster charts, their edit histories involve more tightly scoped and less diverse changes. This dataset enables a wide range of analyses of chart change over time.
1 INTRODUCTION
Open communities provide a setting for studying how public-facing visualizations are maintained, but community editing of chart images is poorly understood. The paper introduces a labeled Wikimedia Commons revision dataset and analyzes editing patterns across contributors, time, and file formats.
1 INTRODUCTION
- Community editing of chart images remains understudied, unlike Wikipedia text editing; synthetic chart-edit datasets do not capture revision in unstructured open environments.The paper frames public chart revisions as a way to study how communities maintain visual content and design over time.Source: §1 · p. 1
- The paper contributes 7,786 longitudinal revisions across 1,664 Wikimedia Commons chart images, labeled with edit categories using an LLM pipeline.These are public-facing revisions on a platform without strict chart ownership.Source: §1 · p. 1
- Multi-contributor charts show more aesthetic edits and fewer structural modifications; mark and text edits are burstier than data edits.SVG charts attract more unique contributors, while their revisions are more tightly scoped than raster-chart revisions.Source: §1 · p. 1
- The dataset and initial analyses offer a basis for studying how open-community members maintain and iterate on visualization content and design.Source: §1 · p. 2
2 RELATED WORK
Prior work studies Wikipedia text editing, collaborative visualization tools, and the costs and benefits of collective editing, but less often the revisions of public-facing chart artifacts. This paper situates its categories in visualization critiques and examines chart changes themselves.
2 RELATED WORK
- Wikimedia visual-asset curation has received sparse study, and revision dynamics have not been applied to data-visualization images.Existing editing research has largely focused on Wikipedia article text, while Wikimedia Commons contains both community-authored and externally imported media.Source: §2 · p. 2
- Collaborative visualization research mainly supports exploratory analysis or design discussion, rather than studying public-facing visual artifacts as revision outcomes.This study shifts focus from collaboration processes to chart images made for public communication.Source: §2 · p. 2
- Collective editing can improve accuracy and design but also incur coordination friction; raster formats can be harder to revise directly than vector graphics.Prior research also finds that revision timing varies with peer interaction.Source: §2 · p. 2
- The paper’s edit categories operationalize prior findings on data relevance, visual encoding, wording, context, and clutter in chart revisions.These considerations come from research on visualization workshops and online critiques.Source: §2 · p. 2
3 DATASET
The paper assembles a longitudinal Wikimedia Commons dataset of chart revisions, labels observed changes with an LLM pipeline, and summarizes the dataset’s composition and edit coverage.
3 DATASET
- The dataset contains 7,786 revisions across 1,664 chart images created between 2004 and 2026, with LLM-labeled edits and chart metadata.Source: §3 · p. 2
3.1 Chart Revision History Collection
- The authors traversed Wikimedia Commons’ Charts category tree to identify candidates, then filtered chart images and collected revision metadata.The search extended six category levels and yielded 82,516 images from 1,659 categories; revisions include timestamps, contributors, and comments.Source: §3.1 · p. 2
3.2 Edit Category Labeling with LLMs
- Eight edit categories use action labels for observed changes; the codebook retained categories with reliable LLM detection and omitted several difficult-to-label attributes.The three-stage pipeline consolidated generated edit descriptions into category labels; on 80 revisions, LLM-human agreement was κ = 0.85.Source: §3.2 · p. 2
3.3 Dataset Overview
- Line charts are most common (53%), followed by bar (14%), area (13%), pie (7%), and other (13%); line and area charts are more prevalent than in prior open-web analyses.Charts average 4.7 revisions (median 3); 20% of revisions fit no labeled category, while 95% of charts have at least one labeled revision. 58% of charts have multiple contributors, 66% are community-authored, and 53% are SVG.Source: §3.3 · pp. 2–3
4 RESEARCH QUESTIONS AND RESULTS
The analyses compare editing patterns by contributor count, edit cadence, and file format. They find differences in the kinds and timing of edits, alongside higher activity but narrower edit scope in SVG charts.
RQ1 Results
- Multi-contributor charts show higher likelihoods of Mark Color (OR = 1.32) and Text Content (OR = 1.60) edits, but lower likelihood of Mark Type changes (OR = 0.45).Data-edit likelihood also differs by chart origin: lower for community-authored charts (OR = 0.54) and higher for externally sourced charts (OR = 1.48).Source: §4 · p. 3
RQ2 Results
- Mark edits (β = 0.09) and Text edits (β = 0.10) occur in more irregular bursts than Data edits; data-edit cadence does not significantly differ by chart origin (p = 0.65).Multi-contributor charts also exhibit higher burstiness than single-contributor charts (β = 0.2, p < .001).Source: §4 · pp. 3–4
RQ3 Results
- SVG charts attract more contributors (IRR = 1.15) and revisions (IRR = 1.13) than raster charts, but have narrower edit-category breadth within revisions (β = −0.46) and across histories (IRR = 0.87).The higher activity is associated with format rather than chart age.Source: §4 · p. 4
- In more recent years, revisions (β = 0.04, p < .001) were slightly more likely to be in SVG than raster formats, suggesting a steady community migration toward vector graphics.Source: §4 · p. 4
5 DISCUSSION
The discussion interprets community chart revisions as reflecting both shared editing patterns and distinctive text and format behaviors, while noting limits in dataset coverage and edit labels. It outlines analyses that could extend these findings.
5 DISCUSSION
- Text is highly edited, and content-versus-presence differences suggest communities refine what text expresses more than whether it appears.Such edits may neutralize biased annotations or translate labels for multilingual audiences.Source: §5 · p. 4
- The findings largely align with prior work on community editing and visualization design, while text editing and narrower SVG edit breadth are identified as distinctive patterns.The authors suggest these SVG patterns may partly reflect how direct DOM changes differ from re-rendering raster charts.Source: §5 · p. 4
- SVG revisions span fewer edit categories within revisions and across histories; direct DOM edits versus raster re-rendering may contribute, but user intent and tool effects remain unresolved.Raster re-rendering can bundle changes to default colors or text labels; the authors call for further research to disentangle intent from tool-driven artifacts.Source: §5 · p. 4
- The sample may not represent all Wikipedia charts because collection used Wikimedia Commons categories and excluded charts removed from articles; some images or edit labels may also be erroneous.The LLM-based categories also omit aspects affecting data-ink ratio, including non-data-mark dimensions, grid lines, tick marks, and graphical annotations.Source: §5 · p. 4
- Future work will expand edit categories, quantify revision degree, model full edit sequences and reversions, examine discussion pages and category co-occurrences, and compare language versions.The authors aim to investigate contributor intent and cultural differences in visualization revisions.Source: §5 · p. 4
- The dataset and initial analysis characterize public-communication chart revision, and the authors propose that future research can use these findings to design community-driven chart-editing tools.The reported findings include differences between multi- and single-contributor charts and tighter editing scope in SVG than raster charts.Source: §5 · p. 4
FIGURE CREDITS
Figure credits identify the creators and licenses for charts shown in Figure 1 and identify the Figure 2 chart as public domain.
FIGURE CREDITS
- Figure 1 credits identify charts (a) and (b) as by Quillaja and SCZMGE, with Public Domain and CC BY 4.0 licenses, respectively.The chart by SCZMGE is licensed CC BY 4.0.Source: FIGURE CREDITS · p. 4
- Figure 1 credits identify charts (c), (d), and (e) as by Nikolai Twin, ArcCan, and Borvan53, each CC BY-SA 4.0.Source: FIGURE CREDITS · p. 4
- Figure 2 credits the chart to Morpheus7377 and labels it Public Domain.Source: FIGURE CREDITS · p. 4