Yes! A duplicate! Sorry, resolving ...

@egardner can this be moved out of code review now?

Or ... better still - use a dismax of a field and its plain version when creating the query

Any progress on this?

in the user testing Growth did in Dec 2020, we asked users to rate the different pieces of information for usefulness and the source (labeled "Suggestion reason" in the test UI) was one of the most highly rated pieces of information by users, particularly when the source was that it was used in the same article in another language Wikipedia as this was easy to understand. This appears to be reflected in the Android MVP data, where ratings were 80% for when this was the suggestion reason shown

In the future, we'd like to be able to filter by source type, e.g. only show Wikidata-based recommendations (which tend to be significantly more reliable). This is not needed for the first iteration, but since otherwise the search index data would probably not need to be reloaded between iterations, it might be easier to deal with it now.

So the easier-to-follow process for doing search tuning is

1. https://github.com/cormacparle/media-search-signal-test#how-to-create-a-dataset-in-ranklib-format-quick-version
2. https://github.com/cormacparle/media-search-signal-test#3-train-a-logistic-regression-model-using-ranklib

FWIW here's the code used to train/test the logistic regression model
https://github.com/cormacparle/media-search-signal-test/blob/master/logreg.py

tl;dr: gathering more labeled data does not look like it will measurably improve the precision of our results, so there's no point in making a big effort to do it

Thank you!

@ArielGlenn Can we close this now?

@egardner is this something the design systems team is already planning?

blocked by https://phabricator.wikimedia.org/T280368

Blocked by https://phabricator.wikimedia.org/T280368

This will get done as part of T280368

So when querying the search api, you guys need to set srqiprofile =empty in the request, then the confidence score can be worked out as descirbed

Hmm or maybe not. maintenance/createFileListfromCategoriesAndTemplates.php doesn't handle sub-categories

After some poking around I see that step 1 can be accomplished using the scripts

@Vlad.shapik can this be resolved now?

Turns out the problem was the clock on my phone had drifted! Working fine now, thanks @Urbanecm

Fix deployed, working now

We could do. Should I make a ticket?

I think we could do another cycle of tuning search results incorporating the data from the image-recommendation test, and then graph the data again and see where we are

We can, but it means going through all unillustrated articles and grabbing their scores, so it's going to need a ticket in itself. I'll make one

We have 2485 rated recommendations using mediasearch, from 984 search terms

Talked to @LSobanski and we don't think we need extra storage atm

@John_Cummings - structured data on a commons File page is for describing the file. For example:

• what an image depicts
• the copyright licence associated with the file
• who created the file

Comparison of results from the analysis tool before and after this change

Here's a csv of the results

Yeah, that's expected behaviour for now - https://phabricator.wikimedia.org/T279072 has been raised to modify the behaviour of the dropdown so that it's more like desktop

aaaand done

... so we just need a few more ratings to get to our target in ar, cs and ceb