Dutch Translation Problem

Question

Dutch Translation Problem

Dogu Tumerdem 0

Hello,

I am using Translation service to translate some user inputs into several languages, including Dutch (NL).

On code side I used Azure.AI.Translation.Text v2.0.0 but Experienced some problems. Then I tried direct rest call the result was wrong too.

Doing the following call:

curl -X 'POST' \
  '{{azureTranslatorUrl}}/translate?api-version={{apiVersion}}' \
  -H 'Ocp-Apim-Subscription-Key: {{subscriptionKey}}' \
  -H 'Ocp-Apim-Subscription-Region: {{subscriptionRegion}}' \
  -H 'Content-Type: application/json' \
  -d '{
  "inputs": [
    {
      "Text": "Senior Collector",
      "language": "en",
      "targets": [
        { "Language": "nl" },
        { "Language": "tr" }
      ]
    }
  ]
}'

The Result is:

{
  "value": [
    {
      "translations": [
        {
          "language": "nl",
          "text": "Senior Collector",
          "sourceCharacters": 16
        },
        {
          "language": "tr",
          "text": "Kıdemli Koleksiyoncu",
          "sourceCharacters": 16
        }
      ]
    }
  ]
}

Switching off automatic language detections fails too. "Cherry" is returned as "Cherry" but "Banana" translated in correct way.

Am I missing something?

Pavan Kumar Purilla 5 Reputation points Microsoft External Staff Moderator

2026-07-23T11:26:27.9166667+00:00

Hi Dogu Tumerdem,
Thanks for reaching out to Microsoft Q&A.

Based on the information provided, the REST API request appears to be correctly formatted, and the service is returning a successful response. The fact that some terms, such as “Senior Collector” and “Cherry,” are returned unchanged while others are translated suggests that the translation model may be preserving certain source terms in the target language based on context.

To investigate this further, could you please share a few more examples of words or phrases that show the same behavior, along with the expected Dutch translations? It would also be helpful to know whether this issue occurs only with isolated words or job titles, or whether it also happens with complete sentences and larger text samples.

As a troubleshooting step, please try the same terms within a full sentence and compare the results. If the behavior remains inconsistent or does not match the expected translations, we may need to review the specific inputs and outputs in more detail to determine whether this is expected model behavior or a service issue.
Pavan Kumar Purilla 5 Reputation points Microsoft External Staff Moderator

2026-07-27T01:58:23.8233333+00:00

Hi Dogu Tumerdem,
Did you get any chance to review the above response. Do let me know if you have any further queries.

1 answer

Your answer

Pavan Kumar Purilla 5 Reputation points Microsoft External Staff Moderator

2026-07-23T11:26:27.9166667+00:00

Hi Dogu Tumerdem,
Thanks for reaching out to Microsoft Q&A.

Based on the information provided, the REST API request appears to be correctly formatted, and the service is returning a successful response. The fact that some terms, such as “Senior Collector” and “Cherry,” are returned unchanged while others are translated suggests that the translation model may be preserving certain source terms in the target language based on context.

To investigate this further, could you please share a few more examples of words or phrases that show the same behavior, along with the expected Dutch translations? It would also be helpful to know whether this issue occurs only with isolated words or job titles, or whether it also happens with complete sentences and larger text samples.

As a troubleshooting step, please try the same terms within a full sentence and compare the results. If the behavior remains inconsistent or does not match the expected translations, we may need to review the specific inputs and outputs in more detail to determine whether this is expected model behavior or a service issue.
Pavan Kumar Purilla 5 Reputation points Microsoft External Staff Moderator

2026-07-27T01:58:23.8233333+00:00

Hi Dogu Tumerdem,
Did you get any chance to review the above response. Do let me know if you have any further queries.

Answer 1

Nothing is wrong with your call — the API is working as designed. What you're hitting is model behaviour, not a bug, and it's specific to how Dutch treats English loanwords.

Dutch business language borrows English job titles wholesale. "Senior Developer", "Account Manager", "Senior Collector" — these appear verbatim in Dutch job postings and CVs constantly. The NMT model has learned from parallel corpora that the most natural Dutch rendering of an English job title is often... the English title itself. Some other languages like Turkish doesn't have this borrowing pattern, so you get "Kıdemli Koleksiyoncu". So the model isn't failing to translate; it's making a (defensible) judgment that the "translation" into Dutch is the English term.

Why "Cherry" but not "Banana"

Same mechanism plus capitalization. Capitalized "Cherry" in isolation looks like a proper noun (a person's name, a brand), and MT engines preserve proper nouns. "Banana" is rarely a name, so it translates. Try these and you'll likely see the difference:

"cherry" → kers

"Cherry" → Cherry

"I ate a cherry." → Ik heb een kers gegeten.

You have few options:

Add context or lowercase the input: Sending "senior collector" or a full sentence ("She works as a senior collector in the debt recovery team") will usually flip the behaviour. If your inputs are UI labels or job titles, wrapping them in a carrier sentence and extracting the phrase afterwards is a common (if ugly) trick.
Dynamic dictionary: If you know the translation you want, you can force it inline (works in v3.0 with textType=html semantics preserved):
```
   It is a <mstrans:dictionary translation="senior incassomedewerker">senior collector</mstrans:dictionary>.
```
Try the LLM model option: Since you're already on the new preview API (the inputs/targets request shape), you can set the target's model/deployment to the LLM-based translation (e.g. GPT-based) instead of the default NMT. LLM translation handles short ambiguous fragments and "should this loanword be translated?" decisions differently, and you can also pass a reference or adaptive examples to steer it. Worth an A/B test on your problem terms.