You can't judge the Vietnamese yourself, so don't try. You can still control quality in three ways. Have an independent Vietnamese reviewer check a sample against a written error scale. Run automated checks that need no Vietnamese. Test the text in context with a few real Vietnamese users.
Key points
- Judge the process, not the prose: a second qualified reviewer, a written scale and a sample.
- Many serious faults show up without reading Vietnamese: missing accent marks, broken placeholders, wrong numbers, clipped text.
- Back-translation is a weak check. It catches meaning errors and misses poor fluency and wrong tone.
- An AI model is a useful second opinion for flagging likely errors, never a verdict.
- The reviewer must not be the translator.
- Finish with a short test with native users doing real tasks.
What does "good" mean for a Vietnamese translation?
A translation can be accurate and still sound stiff or not fit on a button.
The MQM framework (Multidimensional Quality Metrics) is a community standard for sorting translation errors into types. Its current error typology has seven top-level dimensions: Terminology, Accuracy, Linguistic conventions, Style, Locale conventions, Audience appropriateness, and Design and markup. The framework is maintained by the MQM Council (MQM error typology).
In plain terms, here is what a reviewer looks at:
- Accuracy: the meaning matches the source. Nothing is added, dropped or distorted.
- Fluency: grammar, spelling and punctuation are correct, and it reads like Vietnamese rather than English in disguise.
- Terminology: your product terms are translated the same way every time, and match your glossary.
- Tone and forms of address: Vietnamese has no single neutral "you". Pronouns shift with age, gender and relationship, and a wrong choice can sound rude or oddly distant. A banking app and a gaming app need different choices. Your style guide should settle this; see my Vietnamese localization style guide.
- Locale formats: dates, numbers and currency follow Vietnamese conventions.
- Fit in the UI: the text fits the space and the markup still works.
What can I check myself without speaking Vietnamese?
Quite a lot, with only the source text, the translated file and a careful eye. Localization tools such as Crowdin and Phrase TMS run several of them automatically.
- Diacritics (accent marks) are present. Vietnamese uses marks for tones (á, à, ả, ã, ạ) and for letters (ă, â, ê, ô, ơ, ư, đ). It is written "Tiếng Việt", not "Tieng Viet". Text stripped of these marks is a red flag. It usually means a bad export, a wrong encoding or a tool that dropped the marks, and it makes words ambiguous. Also look for garbled characters or empty boxes, which point to encoding or font problems. URLs and usernames are exceptions.
- Numbers match the source. Prices, quantities, versions and phone numbers should be the same digits. Crowdin flags numbers that differ from or are missing from the source, and Phrase TMS has separate "missing numbers" and "extra numbers" checks.
- Number, date and currency formats look Vietnamese. The Unicode CLDR locale data for Vietnamese uses a comma as the decimal separator and a dot for grouping, and places the currency symbol after the number with a space (CLDR numbers data). Dates use day before month (CLDR dates data). So 1,234.50 in English should appear as 1.234,50. The dong (VND) has 0 decimals, so a price looks like "1.234.000 ₫" (CLDR currency data).
- Placeholders and tags are intact. Strings such as
{name},%sor<b>…</b>must survive unchanged. A missing or altered placeholder can break a sentence or crash a screen. Both Crowdin and Phrase TMS check this. - Text fits. Open the screens or take screenshots and look for cut-off words, overlapping text and buttons that wrap badly. Phrase TMS's length check defaults to flagging a target above 130% of the source length; use a limit that suits your layout.
- Glossary terms are used. Search the file for your approved terms. Crowdin and Phrase TMS can flag a string that doesn't use the glossary term, or that uses a forbidden one.
- No English left behind. Search for strings identical to the source. Phrase TMS has a "target identical to source" check. Brand names and technical terms may stay in English on purpose; the rest is likely a miss.
Is back-translation a good way to check?
Back-translation means asking someone to translate the Vietnamese back into English and comparing it with your original. It is cheap and feels reassuring, but it is weak. It can catch a dropped sentence, a reversed meaning or a changed number. It misses what matters most to Vietnamese readers: a stiff, literal translation can back-translate perfectly, and a wrong pronoun or too-casual tone leaves no trace in English.
My view: use back-translation to spot-check meaning on a handful of high-risk strings, such as legal text, safety text and payment screens. Don't use it to approve fluency or tone.
Can I use AI as a second opinion?
Yes, with limits. Research has shown that large language models can mark likely error spans in a translation and rate their severity in MQM-style categories (GEMBA-MQM, Kocmi and Federmann, 2023). That makes an LLM a reasonable reviewer's assistant.
How I would use it:
- Give it the source, the translation and your glossary. Ask it to list possible errors by type and quote the words involved.
- Ask for doubts, not a score. "Is this good?" gets a confident, unhelpful yes.
- Run two different models and look at where they disagree.
It can miss errors and invent errors that aren't there, and it may not know your audience or preferred form of address. Treat each flag as a question for the human reviewer.
Confidentiality: don't paste unreleased product text, customer data or contract wording into a public AI tool. Check your provider's data terms first, or use an approved enterprise setup.
For the wider picture of how machine output and human review fit together, see my notes on human-in-the-loop translation for Vietnamese localization.
How does an independent LQA review work?
LQA (language quality assurance) is a review of the translation by a qualified linguist, scored against an agreed error scale. It is the one check here that tests the Vietnamese itself.
The reviewer is not the translator. The ISO 17100 translation-services standard (2015) makes revision by a second person, separate from the translator, a required step (ISO 17100). A translator checking their own work will miss their own habits.
Review a sample, not everything. Pick a sample that covers your content types (UI strings, marketing text, help articles, legal text) and your riskiest screens, and agree its size with the reviewer. If the sample fails, send the whole batch back, not just the sampled strings.
Use a scorecard with severity levels. Each error gets a type (from a list such as MQM's) and a severity. One example is Smartling's default MQM scheme, which weights Critical errors at 25, Major at 5, Minor at 1 and Neutral at 0, with a configurable pass threshold (Smartling help: What is MQM). Those weights are one vendor's defaults. What matters is writing the scale down before the review, so a "fail" doesn't depend on opinion.
Ask for comments you can act on. Each flagged error should show the source, the translation, the error type, the severity and a suggested fix. For a practical list, see my Vietnamese localization QA checklist for AI translation. If you need to find a reviewer, where to find Vietnamese translators covers that.
Test with Vietnamese users too. If the text is user-facing, a reviewer checks it against rules, and users show you whether the product works for them. A small session is enough. Ask a few native Vietnamese speakers who match your real audience to do real tasks: sign up, find a setting, complete a purchase. Watch where they hesitate. Ask what each unclear label means in their own words, and note tone that feels too formal or too familiar. Treat what you hear as direction, not as statistics.
What are the red flags in a vendor's delivery?
- No named reviewer, or the translator reviewed their own work.
- No glossary or style guide was used or shared with you.
- Missing accent marks, broken characters, altered placeholders, or terms that vary from screen to screen.
- English text is left in without explanation.
- The vendor won't let an independent reviewer see a sample.
- A quality score comes without who scored it, on what scale, or on how much text.
- Nobody can say whether machine output was used or how it was checked.
What should the full check look like?
| Check | Who does it | Needs Vietnamese? |
|---|---|---|
| Diacritics (accent marks) present, no garbled characters | You or QA tool | No |
| Numbers match the source | QA tool | No |
| Number, date and currency formats | You and developer | No |
| Placeholders and tags unchanged | QA tool | No |
| Text fits in screenshots | You or tester | No |
| Glossary terms used | QA tool | No |
| No unexplained English left in | QA tool | No |
| LLM flags on a sample | You, then reviewer | No, but the reviewer judges each flag |
| Back-translation of high-risk strings | Another linguist | Yes (bilingual person) |
| Independent LQA on a sample, scored | Vietnamese linguist, not the translator | Yes |
| Test with native users | Native users | Yes |
If you want an independent check, I review English–Vietnamese localization against a written scale and send findings you can act on. See what I do, or get in touch.
FAQ
Can Google Translate or ChatGPT tell me if my Vietnamese is correct?
ChatGPT or another AI model can flag likely problems, but it can't give a reliable yes. Google Translate can only translate the text back into English, which is a weak check. Use them as a second opinion for a human reviewer, and don't paste confidential text into public tools.
How much of the translation should be reviewed?
Usually a sample, covering each content type and your riskiest screens. Agree the size with the reviewer. If it shows serious errors, ask for a wider review or a re-do.
Why is Vietnamese without accent marks a problem?
Accent marks change the word, so without them many words become ambiguous and the text looks careless. It also often signals a technical fault, such as a bad export, that can affect other strings.
Can the translator review their own work?
Not as the only check. ISO 17100 requires revision by a person other than the translator.
What is MQM?
MQM (Multidimensional Quality Metrics) sorts translation errors into types and rates each by severity, giving reviewer and client a shared scale.
Sources
- MQM error typology, checked 2026-11-19
- Smartling help: What is MQM, checked 2026-11-19
- Crowdin: QA checks, checked 2026-11-19
- Phrase TMS: QA checks, checked 2026-11-19
- Unicode CLDR: Vietnamese number data, checked 2026-11-19
- Unicode CLDR: Vietnamese date data, checked 2026-11-19
- GEMBA-MQM: Detecting Translation Quality Error Spans with GPT-4, checked 2026-11-19
- ISO 17100:2015 Translation services, checked 2026-11-19
- Unicode CLDR: currency data (VND decimals), checked 2026-11-19