Jump to content

Wikipedia:Wikipedia Signpost/2024-07-04/In focus

From Wikipedia, the free encyclopedia
File:Discord stream of ruwiki's vandalism detection system (cropped).png
MBH
CC BY-SA 4.0
0
0
300
In focus

How the Russian Wikipedia keeps it clean despite having just a couple dozen administrators

The Russian Wikipedia (ruwiki) historically has fewer administrators per active user or per article than many other large Wikipedias; the number of active users is ten times smaller than the English Wikipedia, and 2 times smaller than the German or French. Currently there are 63 administrators on ruwiki (not counting adminbots), and many of them are not very active; at the same time, ruwiki is in the top three most-visited Wikipedias (after English, and sometimes after Japanese or Spanish[1]). Voters in ruwiki's RfAs tend to be very critical of candidates — which makes it impossible for many experienced, active and well-known users to be elected.

In recent years, the Russo–Ukrainian war and political repression in Russia have made it more dangerous to be a member of the wiki community, and even more so to be an admin there; in the last two years (since summer 2022) only two have been elected. Since the start of the war, more than ten administrators (about 15% of the corps) either lost their status through inactivity, abandoned it due to fears for their own safety, or joined one of the pro-Kremlin forks; one was killed in action in the Ukrainian army, and one was designated a "foreign agent" by the Russian government.

In the face of this long-standing severe shortage, the ruwiki community has developed a few mechanisms to spread some of the burden of project maintenance: automation with bots (including some that use machine learning), and unbundling.

Unbundling

Several semi-admin user groups have been created, to grant individual rights to users without requiring a formal request for adminship. Right now, ruwiki has several dozen users who can perform closures at deletion discussions, delete pages, edit protected pages, edit in the MediaWiki namespace, change site-wide scripts and styles, block vandals and protect pages — but they aren't admins. They do this by using admin permissions from their usergroups, and by using userscripts connected to an adminbot.

In 2009, ruwiki gained a new user group — "summarizers" — who were given the authority to sum up the results of AfDs. Initially, their powers were very limited, and they didn't even have the technical permission to delete; there was a bot to handle deletion-by-request. Since 2010, they've had the permission to actually delete pages, and since 2012 their powers in AfD have been nearly equal to those of administrators. The number of such users is slightly less than the number of admins, and 60% of AfD decisions have been made by these users. Initially, many users supported a strict hierarchy of users and usergroups, and accordingly objected to the creation of these unbundled usergroups. But over the course of ruwiki's history, the decision to create such groups has been vindicated.

Another group, added in 2016, are "engineers". These are technical specialists, who need to edit protected pages — a stronger analogue of the template-editor flag on enwiki — since technically-competent users usually can't pass ruwiki's RfA due to a lack of social skills. Since the establishment of this usergroup, almost all edits on protected templates, MediaWiki pages, sitewide scripts, and sitewide stylesheets have been made by engineers and not admins. When the WMF moved interface administrator (intadmin) permissions to a separate usergroup, it was mostly engineers and not admins who became interface administrators. In fact, this WMF decision caused a huge conflict between the engineers and the old-school administrators – the latter argued that non-admins shouldn't be given higher rights than administrators, while the former proposed that (unlike engineers, who had already proven their competence) administrators be required to pass a JavaScript/CSS proficiency exam to become intadmins. Nowadays, almost all edits to sitewide scripts, styles and system messages are made by one user, who isn't an admin, but is an engineer and intadmin.

Bots

Cosplayer of Reimu Hakurei
Reimu, with spell cards, prepared to do battle.

In 2017, the first machine learning-powered vandalism patrolling bot was introduced on ruwiki: Reimu Hakurei. Unlike enwiki's ClueBot NG, this bot didn't use its own detection system, but used ORES — a model trained and operated by the WMF, used for highlighting suspicious edits in watchlists and recent changes. Like ClueBot, Reimu reverted edits and sent messages to the users who made them, explaining its actions and where to appeal a revert. Edits with ambiguous ORES scores were left alone, but listed at a special page for further analysis, which showed up in the watchlists of experienced users. Reimu also made automatic reports about users who made several suspicious edits to the ruwiki's analogue of WP:AIV. This bot rolled back tons of vandalism, including ideologically driven edits, and whitewashing on the articles of Russian political officials; for this it received mentions in the media.[2][3]

Screenshot
With this adminbot script, a user can be blocked for vandalism, spam, inappropriate or promotional username. If the script is used by admin, he could block user, hide his contribs and delete pages, created by him, with a single click. Non-admin can also cancel a block imposed by him.

Since the second half of the 2010s, ruwiki's checkuser developed two adminbots, which perform many tasks, from blocking open proxies and IP/users whose edits were repeatedly rolled back, to protecting articles from (automatically detected) edit wars and vandal raids. In 2023, he created a userscript and bot allowing trusted non-admins to block IP addresses and new users for clearly-disruptive edits and to apply protection to actively-vandalized articles. Currently, this script and bot are used by about ten trusted users.

In 2022, after the start of the war, another user created a bot for detection of specifically anti-Ukrainian vandalism: this one streamed suspicious edits to a special channel on ruwiki's Discord server. In 2024, Reimu's author improved it by adding detection based on scores from the newer language-agnostic and multilingual revert risk models on the LiftWing platform, AbuseFilter-generated edit tags and text patterns (so it absorbed the anti-Ukrainian bot) and a feature to post suspicious edits to the ruwiki's Discord server. After that, he decided to shut down the automatic-rollback feature, because he wasn't satisfied with the false-positive rate. From the fall of 2017 to the spring of 2024, Reimu made 120,000 rollbacks. The checkuser mentioned above created his own automatic-rollback bot, which uses both ORES and a self-written detection system; the bot owner hopes that the new bot will have fewer false positives.

Automation

Screenshot of a Discord message.
Vandal edit in Discord stream: the word "deputies" has been replaced with "faggots", spelled with a mixture of Cyrillic and Latin letters to avoid AbuseFilter rules. The red button does a rollback, the blue one allows to select more detailed or custom revert reason, and the green one just deletes the post from the stream. A post contains links to the article's history, the edit's diff, and the editor's contributions.
TKTK
Suspicious edits stream in watchlist, with a rollback button

Meanwhile, authors of Reimu and anti-anti-Ukrainian bot have implemented a mechanism to revert problematic edits directly from the Discord server, by clicking the buttons under the post in the Discord channel. An edit can be rolled back with a standard reason, or with one of 12 more-detailed reasons (for example, “No reliable sources” or “Replacing the transcription without page move or move request”), or a manually entered reason. The bot deletes processed edits (rolled back or approved) from the channel, so the channel contains only edits that have not yet been processed. More than 1500 edits have been rolled back using this tool since its establishment 2 months ago. The same edits are posted to a page onwiki, with an excerpt of their text in the edit summary — that's often enough to recognize vandalism on its own), and edit can be rolled back by pressing a link just in the comment without opening a diff (the link leads to a Toolforge-hosted tool that does rollback).

This bot also works on the Ukrainian Wikipedia, posting suspicious edits on-wiki and to Discord (for certain reasons — to ruwiki's Discord too, not ukwiki's — but there are several experienced ukwiki users in ruwiki's Discord). This makes more than 100 rollbacks per month on ukwiki.

This story shows how, thanks to bot owners and semi-admin usergroups, a very small group of people numbering only several dozen active users can effectively maintain ruwiki's two million articles, ensuring the functioning and reliability of one of the top three most-visited Wikipedias.

  1. ^ See statistics at pageviews.wmcloud.org.
  2. ^ "Did the Ministry of Internal Affairs try to remove data about Kolokoltsev's "offence" from Wikipedia?" (in Russian). REGNUM News Agency. 2018-12-10.
  3. ^ "The Great Wikipedia Edit War". iStories. 2024-04-15.