Page MenuHomePhabricator

Add a link: Deploy prioritization of underlinked articles feature
Open, In Progress, MediumPublic

Description

The work in T301096: Add a link: prioritize suggestions of underlinked articles allows us to configure a underlinkedWeight property for the link recommendation task type. By default, articles to be used for link recommendation tasks are selected at random; when underlinkedWeight is set, articles with few links (compared to the number of words in the article) are selected with a higher probability. (The exact formaly is given at T301096#8388234.)

We want to progressively roll out the new functionality to different wikis. Per @KStoller-WMF's comment in T301096#8387455, we should first deploy to pilot wikis, wait for / gather feedback, then rollout to all wikis.

In case of problems, the old functionality can be restored simply by removing the underlinkedWeight key from the Growth configuration page (MediaWiki:NewcomerTasks.json).

Acceptance Criteria

  • Prioritization of underlinked articles is enabled on testwiki and the four pilot wikis (arwiki, bnwiki, cswiki, eswiki).
  • Feedback is gathered from the pilot wikis (and no issues are identified).
  • Prioritization of underlinked articles is enabled by default.

Event Timeline

kostajh triaged this task as Medium priority.Nov 14 2022, 1:04 PM

We can do this after the train deploy on Thursday (or before, but it will have no effect until then).

Tgr changed the task status from Open to In Progress.Nov 18 2022, 7:08 AM
Tgr claimed this task.

Note that this happened today, not in November as suggested in the previous comment.

The formula used is (1-(links/words))^4, with the assumption that there are always more words than links. This did not work: words are counted in wikitext, while links are counted via MediaWiki link tracking which includes links from templates. So on short articles with navigation templates, links/words is a large number, and since we are using an even power for smoothing, we end up with a very large weight, resulting in not particularly useful suggestions.

Also for word count we use the text.word_count field, which is apparently some kind of cleaned text with no templates, tables etc so articles like this one with lots of links in tables cause the same issue.

Disabled for now, will retry after making sure links/words can't make the weight go over 1.

Interesting, thanks for explanation and quickly disabling!

Re-enabled after a fix, seems to be working well. (I did not update the configuration on arwiki as link recommendations are disabled there anyway.)