On Fri, Feb 17, 2017 at 8:33 AM, Niklas Laxström <niklas.laxstrom@gmail.com> wrote:
2017-02-16 23:38 GMT+02:00 Stryn <strynwiki@gmail.com>:
> just a few questions regarding translation memory:

Hi,

in short, it is a known issue (T101236).

To alleviate performance issues, the algorithm of the translation
memory was changed as follows: load N items in the source language
that match most closely to the string in translation. Then load the
translations for them (if any exist). If we do not get more results,
repeat the same for M more items. Currently N = 100 and M = 500.

It so seems that if there are many untranslated strings, we often fail
to choose the ones which would have translations. In addition I think
we have some sort of matching and ranking problem as well, since top
600 matches should definitely contain the previous translations.

It is easy to rule out an updating problem, since I can find those
translations with translation search [1]. And this search [2] shows
that there are many untranslated ones.

[1] https://meta.wikimedia.org/w/index.php?title=Special:SearchTranslations&filter=translated&language=fi&query=You+can+join+the+next+meeting+with+the+VisualEditor+team
[2] https://meta.wikimedia.org/w/index.php?title=Special:SearchTranslations&filter=untranslated&language=fi&query=You+can+join+the+next+meeting+with+the+VisualEditor+team

You are probably interested when this will be fixed. I don't know the
answer for that. It is a non-trivial amount of a work, and it requires
some expertise in algorithms suitable for this purpose. Comments (in
here and Phabricator) which explain how this issue affects you and
wastes you time and how extensive the issue is are welcome and helpful
to show that it is an important issue.

OK, I have commented on https://phabricator.wikimedia.org/T101236 trying to explain how this issue affects us. Any individual translators who feel affected by this, feel free to add something to the ticket (or here, if you don't want to use Phabricator).

//Johan Jönsson
--