Computational analysis of Wikipedia editorial activity during the 2024 European Parliament election

This record contains the complete, reproducible pipeline behind the computational analysis of Wikipedia editorial activity around the 2024 European Parliament election, together with the analysis-ready corpus it produces and the full set of figures it generates. It supersedes version 1, which contained only the English policy-shortcut allowlist and a README.

Scope: 31 Wikipedia language editions, the 24 official languages of the European Union plus Catalan, Basque, Galician, Russian, Turkish, Arabic and Chinese, covering the pages of MEP candidates, national and European parties, and political groups. The observation window opens with the announcement of the election (23 May 2023), spans the campaign (6 February - 9 June 2024) and closes with the formation of the new mandate (1 December 2024).

The analysis measures article and talk-page revisions, editors and editor activity tiers, reverts and vandalism, attention (pageviews), the vocabulary of the debate, and, the central question, how editors invoke Wikipedia's own content and conduct policies and guidelines when they moderate.

Project:

This work is part of DEM-Debate: Building an Enabling Environment for DEMocratic Debate, an 18-month participatory research project (September 2024 – February 2026) run by Wikimedia Europe, the University of Amsterdam (Institute for Information Law, IViR) and Eurecat, Centre Tecnològic de Catalunya, together with the Wikimedia communities in Europe. The project combines a legal and a computational analysis of the fact-checking and content-moderation practices Wikipedia used during the 2024 European Parliament election, and asks whether they can inform future EU policy on election disinformation. This deposit is the computational strand of that work, carried out at Eurecat.

Contents:

 dem-debate-code-v2.zip. The pipeline: an eight-stage extraction and preprocessing library (demdebate/), one analysis module per figure group (analysis/), 60 tests, thin notebooks, and the exact pinned environment (pyproject.toml, uv.lock, Python 3.12, managed with uv). It also ships the small derived data needed for a quick start: policy-shortcut allowlist snapshots for 12 editions (the material published in version 1), an ISO stopword list, page-to-country and page-to-European-party maps, and page-level gender data from Wikidata.

figures-*.zip contains every figure produced by the analysis, one archive per module (revisions, users, vandalism, attention, words, policies, moderation, expanded moderation), each covering all 31 editions.

The remaining files are the corpus, deposited in full so that every stage boundary is archived rather than only the analysis-ready layer: the master page table before and after the Wikidata gender enrichment; article and talk-page revisions with word-level diffs, both as extracted and with the cross-language page identifiers added; editor metadata; the merged, labelled and cleaned per-language datasets the figures are computed from; and the policy-citation occurrences and counts, together with the shortcut allowlists generated for the 19 editions that had no published snapshot.

Two pairs are deliberately redundant, wikipedia_pages_extraction_w_gender.csv adds a single gender column to wikipedia_pages_extraction.csv, and revisions_by_lang_2 adds lang and local_id to revisions_by_lang, so that the output of each stage can be inspected exactly as the next stage received it.

Reproducing from scratch:

The code archive is self-contained. Install with uv sync --locked, check with make tests, and run the whole workflow with make reproduce. Stages 1–6 query the public Wikimedia APIs: over the full language set they take days, not hours, and download several GB. The pipeline is deliberately serial, in line with the Wikimedia robot policy; please identify your run by setting DEMDEBATE_USER_AGENT to a string containing a contact address.

Provenance. Because extraction reads the live APIs, a fresh run reflects Wikipedia as it stands at run time and cannot recreate a past state. The corpus deposited here is the August 2026 snapshot, and it is the one the deposited figures were computed from, which is why the data is archived alongside the code rather than left to be regenerated.

Licensing. this deposit may not be used for commercial purposes Three licenses apply, one per kind of material. Source code: PolyForm Noncommercial 1.0.0. Figures, README and documentation: CC BY-NC 4.0. Data derived from Wikipedia: CC BY-SA 4.0.

The data is the one part that is not noncommercial, and that is not a choice this deposit makes: Wikipedia content is published under CC BY-SA 4.0, whose ShareAlike condition does not allow a NonCommercial restriction to be added to derivative works. Those files keep the upstream license, with attribution owed to the Wikimedia projects and their contributors. Wikidata statements are CC0. The code uses PolyForm Noncommercial rather than a Creative Commons license because Creative Commons advises against applying its licenses to software.

Funding. DEM-Debate is supported by the European Media and Information Fund (EMIF), managed by the Calouste Gulbenkian Foundation in partnership with the European University Institute.

Contact: david.abella@eurecat.org, david.laniado@eurecat.org