Implement text analysis solutions 4.1 Apply language model text analysis

A support team translates tickets with an LLM deployment through Azure Translator (API…

Study guide skill: Build solutions that translate text by using Azure Translator in Foundry Tools or LLM-powered translation flows

Question B5-013

A support team translates tickets with an LLM deployment through Azure Translator (API version 2026-06-06). Product terminology changes every few weeks, and the team has a small set of approved source and target sentence pairs. Translations must follow that terminology without training or deploying a new model, and updates must take effect within minutes. What should you use?

Select one answer.

  • A Adaptive custom translation: build an adaptive dataset from the sentence pairs and pass its adaptiveDatasetId in the translate request's target ✓ Correct answer
  • B Custom Translator: train and deploy a custom NMT model on the sentence pairs
  • C Set the target's tone to formal
  • D Set allowFallback to false on the target

Answer: A

Adaptive custom translation improves LLM translation output with a compact set of reference sentence pairs: the service builds a dataset in minutes and, at inference, retrieves similar segments per request to guide terminology, context, and style, with no training or separate deployment. You apply it through the adaptiveDatasetId target parameter of the 2026-06-06 Translate API. Custom Translator trains a dedicated NMT model on large datasets and needs retraining and redeployment for updates, tone only selects formal, informal, or neutral output, and allowFallback only controls what happens when the model doesn't support a language pair.

Deep dive

How it works

Adaptive custom translation (AdaptCT) adapts LLM translation output with a compact set of reference sentence pairs, without training a model:

  1. Upload prealigned source-target segment pairs. The page allows 5 to 10,000 pairs, each up to 250 characters.
  2. The service builds an adaptive dataset in minutes.
  3. In the Translator 2026-06-06 Translate API, pass the dataset's adaptiveDatasetId on the target. At inference, the service uses few-shot retrieval: it pulls similar segments from the dataset for each request to guide terminology, context and style.

There's no offline training and no separate deployment. Updates to the dataset take effect within minutes, which is what a terminology list that changes every few weeks needs.

AdaptCT is the "small data, fast change" choice. When you have large parallel datasets and want a dedicated, stable model, Custom Translator is the alternative.

The other options, and when you'd use them

  • Custom Translator NMT model: It creates a dedicated model through end-to-end training and production deployment. Training can take up to 48 hours, and updates need retraining and redeployment, which breaks both constraints.
  • Tone formal: tone only sets the desired tone of the translation (formal, informal or neutral). It doesn't carry terminology.
  • allowFallback false: allowFallback is about what happens when the chosen model doesn't support a language pair, when the service may fall back to a general system. It has nothing to do with terminology.

Exam tip

LLM translation + a few approved sentence pairs + changes in minutes + no training = adaptive custom translation via adaptiveDatasetId. Large data + dedicated model = Custom Translator.

References

Last reviewed:

Ready to test yourself? The quiz asks this and every other AI-103 question, with instant feedback. Practice this in a quiz →

All Text analysis questions · AI-103 study guide

Thanks for your feedback!