Run a terminology check on a Turkmen file and watch the tool have a small breakdown. It flagged 40 term violations in a 3,000-word oil-and-gas job last week. I looked through all of them. Zero were real. Every single one was the same word wearing a different suffix.
That's the thing nobody selling QA software wants to say out loud: most of these checks were designed for English, German, maybe Spanish. Languages where a noun mostly stays put. You approve "valve" in the termbase, and "valve" shows up in the text, and everyone's happy. Turkmen doesn't work like that, and neither does Turkish, Finnish, Hungarian, Kazakh, or half the languages an agency actually ships into.
What the checker sees vs. what's there
Turkmen is agglutinative. You take a stem and stack meaning onto it — case, possession, number, all glued on the end. So the word for "fuel," ýangyç, becomes ýangyjy, ýangyçda, ýangyçdan, ýangyçlar, ýangyjyň, depending on what the sentence is doing. Same term. Six forms on a good day. The consonant even shifts — that ç softens to a j — so it's not a clean suffix the tool can strip.
The termbase has one entry: ýangyç. The QA engine does a substring match, sees ýangyjy, and screams inconsistency. Or worse — it does an exact-match check, decides the approved term is "missing," and now I've got a "terminology not used" flag on a segment where I used it perfectly, three times, in three grammatical cases.
Multiply that across a legal contract with forty defined terms, each declining through the sentence, and the QA report is a wall of red. A project manager who doesn't read Turkmen sees that wall and assumes I dumped a machine draft and left. The score tanks. I've had to write explanations for QA reports where my actual error count was two typos and the other 58 flags were the tool not understanding that words change shape.
The checks that lie, and the ones that don't
After years of this I've sorted the QA categories into ones I trust for Turkmen and ones I mute on sight.
Trust: tag consistency, number transfer, untranslated segments, double spaces, empty target. These are language-agnostic. If a tag from the source didn't land in the target, that's real, and it doesn't care about morphology. Number checks catch the genuinely dangerous stuff — a torque spec, a pressure value, a contract sum. That's where the money is anyway.
Mute or verify manually: terminology consistency, forbidden terms, and anything doing target-side spell or grammar checks. There's no serious Turkmen spellchecker built into Trados or memoQ or Phrase. What passes for one either flags every second word or nothing at all. Both are useless, so I turn it off and read.
Number-plus-unit is its own trap. In Turkmen the unit often takes a suffix too, and the number can carry a case ending — 3-den, 5-de. The check that verifies "5" appears in the target sees 5-de and either flags it as altered or, if it's lenient, misses a genuine typo hiding in the suffix. I don't let it decide. I scan the numbers by eye on anything technical.
Stop fighting the tool, feed it better
The honest fix isn't a plugin. It's how you build the termbase.
If your CAT tool supports fuzzy or stem matching on terms, turn it on — memoQ and a couple of others let you set match leniency, and even a crude prefix match cuts the false positives hard because Turkmen suffixes come off the end, not the front. It's imperfect (that consonant softening still trips it), but it drops the noise from 40 flags to maybe 8.
Better, if the client insists on strict terminology QA, I ask for the termbase to store the stem plus the most common inflected forms as approved variants. Yes, it's more work up front. But that work is a one-time cost, and it means the QA report becomes something a PM can actually read and believe. A clean report with three real flags is worth more to everyone than a dirty one with sixty fake ones and three real ones buried inside.
And here's the part I'll argue with anyone about: the QA score on an agglutinative language, run with default European settings, is not a measure of my work. It's a measure of the tool's ignorance. When an agency ranks vendors by raw QA pass rate across all languages equally, they're quietly punishing everyone who works into Turkish, Finnish, Kazakh, Turkmen — the people whose languages break the checker by design. Nobody meant to build that bias. It's baked into the defaults, and the defaults never get touched.
So I do two things. I configure the profile myself before I start, not after the flags pile up. And when a report comes back looking ugly, I annotate it — flag by flag — so the human on the other end sees that ýangyjy and ýangyç are the same word doing its job. It takes ten minutes. It's the difference between looking sloppy and looking like someone who knows exactly what the machine can't see.