User talk:TJones (WMF)/Notes/N.O.R.M. MVP Design Notes

Couple comments/questions

Thanks for putting this together!

Depending on your layout capitalized accented letters in french can by typed without a dead-key: AltGr-Shift-0 -> À (fr-oss layout on linux). I don't have a windows machine at hand to see if there are similar behavior but looking at a couple forums I often see users a bit frustrated by how difficult it is to write such letters. I think it's probably safe to ignore them.

Also, with autocomplete, less precise mappings, like not-quite-the-right-keyboard mapping or a sketchy transliteration, have a chance at a fuzzy match. cqstle gets good suggestions on English and Russian Wikipedias (though bqnqnq and qzkzqrd are too much to hope for).

Do you mean here that NORM may generate imprecise candidates that still work because of autocomplete fuzziness or that implementing some mappings might not be necessary because they're already solved by the fuzzy algorithm? If the former, how often do you anticipate fuzziness to help with hazardous mappings? Mainly asking because if this is frequent we might be less inclined to present the generated suggestions in a DYM context, Did you mean cqstle? might not be ideal.

I think it makes sense to start with auto complete but just wanted to point out that users found some discrepancies quite misleading when it comes to different behaviors regarding auto-complete & DYM, for instance complaining that searching for Rissell's paradox you get a proper autocomplete suggestion to Russell's paradox but none of our current DYM implementations manage to fix it. So even though we start with only autocomplete I hope we can keep an eye on what it'll look like when we implement the DYM version.

Regarding computational cost (related to DYM if/when we tackle it), last time we discussed it I might have been a bit conservative mentioning that we only allow one "costly" call. It is a bit misleading, it's true that we don't want to do several sequential costly calls to the backend while applying the fallbacks, but assuming that NORM suggestions are relatively cheap to compute (implemented in PHP) we could stack several query to check if there are results with this candidate in the main search query, which should hopefully not impact the latency too much. DCausse (WMF) (talk) 16:38, 5 August 2025 (UTC)

Thanks for the feedback!

I added a disclaimer about which "French" keyboard I used. The details don't matter too much since it's just an example pointing out that dead keys exist and they can make keyboard mapping less clear.

Do you mean here that NORM may generate imprecise candidates that still work because of autocomplete fuzziness...

Yeah, that was the idea. If someone types with RU/AZERTY wrong keyboard, and we use the RU/QWERTY mapping to map it to cqstle, then autocomplete fuzziness could still save the query. There are other cases where it could go astray, I'm sure.

I don't have strong intuitions about how often various keyboard combinations are used. I looked for RU/QWERTY and never found anything that looked like RU/AZERTY. But I probably never would have noticed something like RU/DVORAK because QWERTY & DVORAK are so different. My weak guess is that one or two keyboards dominate, but we'll have to figure that out.

My solution to Did you mean cqstle is to check that our suggestions get >0 results. The current DYM could use that, too. TJones (WMF) (talk) 16:42, 12 August 2025 (UTC)