Suffixes and plurals

Two Turkish mistakes only show once a placeholder is filled in: a suffix glued to a placeholder, and a plural noun after a number. The tr_TR rules file in polyglots-data carries a pattern for each, turned on. Each match reaches the AI review as a hint with its note, and the model decides in context; the pattern alone never rejects a translation.

They are patterns rather than built-in checks because what they look for is Turkish spelling: the apostrophe before a suffix, and the -lar/-ler endings. Another language with the same problem spells it differently, and can write its own pattern the same way (see Other languages).

Turning them on

polyglots rules fetch tr

downloads the tr_TR rules file with both patterns in it. To add them to a file of your own instead, open it with polyglots rules edit tr and put them under patterns:, keeping the two spaces of indentation:

patterns:
  - find: '(?<![\p{L}\p{M}\p{N}])(?:%(?:\d+\$)?[-+0#]?\d*(?:\.\d+)?[sdfu]|\{[\w-]+\})[''’]\p{L}+'
    level: hint
    note: A suffix on a placeholder cannot agree with every value; rephrase so it lands on a word, as in "%s adlı eklentide"
  - find: '%(?:\d+\$)?[du](?:</[a-z]+>)?\s+(?!(?:dolar|lari|taler|popüler|modüler|seküler|moleküler|selüler|vasküler|regüler|tübüler|angüler|oküler|kapiler|skaler|polar|bipolar|solar|kiler|keler|çalar|galeri|ileri)\p{L}{0,4}(?![\p{L}\p{M}]))\p{L}+(?:lar(?:ı(?:mız|nız)?|ım|ın)?(?:n?a|ya|n?[dt]an?|n?ın|n?ı|y?la|dır|ca|[dt]aki)?|ler(?:i(?:miz|niz)?|im|in)?(?:n?e|ye|n?[dt]en?|n?in|n?i|y?le|dir|ce|[dt]eki)?)(?![\p{L}\p{M}])'
    level: hint
    note: Turkish keeps the noun singular after a number, as in "%d dosya"

polyglots rules check tr confirms the file reads.

Turning a pattern on changes the rules, so the next review of every Turkish file starts over. To turn one off again, delete its lines or put a # before each.

A suffix glued to a placeholder

msgid  "Saved in %s"
msgstr "%s'de kaydedildi"

In Turkish the suffix follows the sounds of the word before it: Ankara’da, İzmir’de, Ordu’da. Whoever translated the string could not see what %s will hold, so 'de is right for some values and wrong for the rest. The same goes for %1$s'nin, {name}'i and %d'den.

The fix is to rephrase so the suffix lands on a word the translator can see:

Reported Rephrased
%s'de kaydedildi %s konumuna kaydedildi
%s'ın kopyası %s adlı öğenin kopyası
%d'den fazla %d öğeden fazla

What the pattern leaves alone:

  • A placeholder in quotes or in an attribute: '%s' silindi, href='%s'. Only an apostrophe followed straight by a letter counts.
  • A placeholder that closes a link around a word: in %1$sbu SSS%2$s'ye the suffix agrees with SSS, which the translator could see.

Some placeholders only ever hold one known value, and the translator comment says so. That is why this is a hint: the model reads the comment and decides.

A suffix that has the same form after every value, such as ken in %s'ken, can be skipped by adding (?!ken(?![\p{L}\p{M}])) straight after [''’]. Do not skip de, da, ın, in, deki or daki: they change with the value, which is exactly what the pattern is for.

A plural noun after a number

msgid        "%d file deleted"
msgid_plural "%d files deleted"
msgstr[1]    "%d dosyalar silindi"

Turkish keeps the noun singular after a number: 3 dosya, never 3 dosyalar. plural-count does not catch this, because Turkish has two plural forms and the entry has both. The fix is the singular: %d dosya silindi.

The pattern reads the word straight after %d or %u (and positional forms such as %1$d), past a closing inline tag, so <strong>%d</strong> dosyalar is caught. It reports the word when it ends in -lar or -ler, optionally followed by a possessive and a case ending in the matching vowel: dosyaları, dosyalardan, öğelerine. A word that only contains those letters without following the vowels, such as olarak or Cloudflare, is not read as a plural.

What it does not see:

  • %s. A count formatted with number_format_i18n() is passed as %s, but elsewhere %s is usually a name, and a name followed by a plural noun is correct Turkish. A pattern cannot tell the two apart.
  • A word in between: %d farklı dosyalar is not reported.

Both limits can miss a mistake; neither reports correct Turkish.

Words it skips

Some words end like a plural and are not one. The pattern skips the words in its (?!(?:dolar|…|ileri) part:

  • Currencies: dolar, lari, taler.
  • Adjectives: popüler, modüler, seküler, moleküler, selüler, vasküler, regüler, tübüler, angüler, oküler, kapiler, skaler, polar, bipolar, solar. %d popüler yazı is correct.
  • Singular nouns: kiler, keler, çalar, and galeri and ileri, whose ending reads as -ler plus a case ending.

To skip another word, add it to the list with a |, as in |ileri|spoiler). The word is then skipped with up to four more letters, so %d spoiler and %d spoilere pass. So does %d spoilerlar, which is a real plural: the pattern cannot tell a case ending from the plural. Add only a word whose plural you do not expect to see after a number.

Other languages

The same two mistakes exist in other languages, spelled differently:

  • A suffix that has to agree with an unknown value: Finnish writes it after a colon (%s:ssa), Hungarian after a hyphen (%s-ban), and Korean attaches the particle (%s을). For Finnish or Hungarian, change [''’] in the first pattern to : or -; Korean needs a pattern of its own.
  • A singular noun after a number: Hungarian, Persian, Georgian, Armenian and the Turkic languages keep it; Finnish and Estonian use a singular case. The endings in the second pattern are Turkish, so write your own.

A pattern that matches too much costs the model a look at each match, so start narrow. See Locale rules for the whole file.

This page is docs/suffixes-and-plurals.md in the repository.