Right now has far as i can see, edits are being classified as vandalism or good faith. Seems interesting to add a new category, testing, for those edits made by newbies who aren't vandalisms, but bad edits made in good faith. Applied to bots, would allow a less intimidating message, since many of them specially when it's a mistake made in good faith loose some of the interest. Despite that, this classification would be theoretically more accurate, with situations that fall in a gray area between vandalism and good faith, and in most cases heuristics could easily give a hand to improve labeling and scoring.
Description
Description
| Status | Subtype | Assigned | Task | ||
|---|---|---|---|---|---|
| Declined | None | T108304 3 categories labels system (multi-class classification) | |||
| Resolved | Halfak | T108679 Train/deploy "damaging" and "goodfaith" models from data collected through (finished) "edit quality" campaigns |
Event Timeline
Comment Actions
Actually, since each edit was labeled according to two orthogonal aspects, we already have four possible classes for each edit:
- A good faith edit, which unfortunately damages the page (inexperienced user trying to help)
- A good faith edit, which improves the page (or at least does not makes it worse)
- A bad faith edit, which damages the page (aka vandalism)
- A bad faith edit, which improves the page (impossible?! non-vandal by accident? labeling error?)
- See also comments on https://en.wikipedia.org/wiki/Wikipedia_talk:Labels/Edit_quality#Feedback
Therefore, I think this feature request is already implemented: you just need to combine the two categories, and look for edits with high probability of being goodfaith, and low probability of being damaging to find out which are the good contributions. Similarly for bad edits made in good faith.
Comment Actions
+1 I'll be training models of "good-faith" and "damaging" first. We can reconsider a 3 class model if that proves insufficient.
Comment Actions
It looks like the "goodfaith" and "damaging" models fit this use-case. Please re-open if that proves insufficient.