Page MenuHomePhabricator

Tone check: Improve handling of quoted content
Open, Needs TriagePublic

Description

We have some basic handling to ignore quoted content for tone check (and some other checks).

Currently this checks

  • If the content is in a <blockquote>
  • If the first chracter is the matched range is inside quotes. This is done by counting the quotes either side of that offest.

This misses cases where most of the text is quoted, e.g. paragraph 2 here: https://en.wikipedia.org/w/index.php?title=Henry_Hudson_(artist)&oldid=1347919439#A_Rake_Revisited

Possible solutions:

  1. Let ignoreQuotedContent set a % of how much content should be quoted before the whole paragraph is ignored, e.g. if 80% of the paragraph is inside quotes, ignore the check.
  2. Mask quoted content before passing it to the model, e.g.

He was described as "the greatest artist ever" by his contemporaries, and "an inspiration" by his partner.

V

He was described as "----" by his contemporaries, and "----" by his partner.

Event Timeline

Change #1287836 had a related patch set uploaded (by Esanders; author: Esanders):

[mediawiki/extensions/VisualEditor@master] [POC] Remove quoted text before running ToneCheck

https://gerrit.wikimedia.org/r/1287836

@Esanders to be doubly sure I'm understanding...

Would it be accurate for me to think this fix:

  • Will make it so Tone Check and the Tone Suggestions we surface in Suggestion Mode will be less likely to activate on quotes
  • Will have no effect on the the Revise Tone suggestions currently being surfaced in the newcomer homepage?

Will make it so Tone Check and the Tone Suggestions we surface in Suggestion Mode will be less likely to activate on quotes

Yes.

Will have no effect on the the Revise Tone suggestions currently being surfaced in the newcomer homepage?

Yes. Those are calculated independently, something our server side suggestions system could resolve.

We might need to coordinate with the model team to see how that "----" approach would work. It could impact the analysis, since they won't have trained on any content like that.