At the moment, we are using LTR on large wikis, where we have enough data to train models. By increasing data retention to > 90 days (which requires ensuring that data is anonymized), we could apply the same techniques to more wikis. LTR is likely to increase the quality of search, and to continue to increase it over time.
Description
Description
Status | Subtype | Assigned | Task | ||
---|---|---|---|---|---|
Open | None | T235857 Learning to Rank (LTR) applied to additional languages and projects to improve ranking (needs experimentation, might not work at all) | |||
Resolved | • JFishback_WMF | T235858 Increase of training data retention (>90 days) is validated with Legal / Privacy | |||
Open | None | T235859 Any new data retention requirements are implemented | |||
Resolved | EBernhardson | T238703 Investigate applying PDGD to our datasets |
Event Timeline
Comment Actions
@Gehel: Does Q2 refer to WMF's Financial Year 2019-20? If so, please say so for the sake of transparency and not to collide with other years. Also see https://www.mediawiki.org/wiki/Phabricator/Project_management#Use_plain_language,_define_actions_and_expected_results - thanks!