As an experienced editor reviewing Add a Link edits, I want suggestions that are reverted to be automatically removed, so that newcomers are less likely to encounter and accept known bad suggestions.
Problem:
The Add a Link system currently learns from explicit newcomer feedback. When a newcomer declines a suggestion, that signal is used to prevent the suggestion from being shown again.
However, if a newcomer accepts a suggestion and publishes the edit, the system treats that as a successful outcome. If the edit is later reverted by another editor because the link was incorrect, that revert is not currently incorporated into suggestion quality controls. As a result, the same incorrect suggestion may continue to be surfaced to other newcomers, creating repeated cleanup work for experienced editors.
Community Feedback:
Feedback from enwiki:
I have noticed multiple instances where the link suggestion tool resulted in adding a link that does not fit the context, specifically a similarly named organization from a different era.
Three times I have had to revert the addition of Eastern Canada Amateur Hockey Association (ECAHA) to Cecil Duncan:
- Cecil Duncan: Difference between revisions
- Cecil Duncan: Difference between revisions
- Cecil Duncan: Difference between revisions
It has also happened in other articles:
The ECAHA existed from 1906 to 1909. It is not the same organization as the similarly named association from the 1940s. I'm not sure what the best solution is, but it is frustrating to repeatedly revert these edits despite the good intentions of new editors.
Solution
TBD
- Suppress suggestions associated with reverted edits: Track whether an Add a Link edit is subsequently reverted. If a suggestion is reverted, prevent that exact suggestion from being surfaced again.
- Incorporate revert signals into model training: Use reverted Add a Link edits as negative training examples when generating future link suggestions. Suggestions that are frequently reverted would reduce the model's confidence in similar suggestions, while accepted and retained suggestions would serve as positive training examples.