Mohamed Elghazali presents an in-depth developer guide to Adaptive Custom Translation (AdaptCT) in Microsoft Foundry Tools, highlighting streamlined translation customization with Azure-based AI, hand-on API usage, and practical workflow tips.

Translation Customization: A Developer’s Guide to Adaptive Custom Translation

Introduction

Translation is more than word substitution—it’s core to global communication. Businesses demand high-accuracy, real-time multilingual solutions that capture domain-specific language. Mohamed Elghazali introduces Adaptive Custom Translation (AdaptCT) from Microsoft Foundry Tools as a transformative approach to these challenges, leveraging advanced large language models (LLMs) like GPT-4o and smart dataset indexing.

Challenges with Traditional Custom Translation

  • Requires large curated datasets (10,000+ pairs)
  • Slow, manual retraining and deployment
  • Not agile enough for dynamic, frequently updated content (support, docs, etc.)

Adaptive Custom Translation (AdaptCT): Key Advantages

  • Uses existing LLMs (e.g., GPT-4o, GPT-4o-mini)
  • Customizes with small, focused datasets (from as few as 5 sentence pairs)
  • Employs smart indexing instead of retraining
  • Updates in minutes, not days
  • Ideal for low-volume, frequently updated or niche domains

How AdaptCT Works

  1. Dataset Indexing: Upload aligned sentence pairs (e.g., English–French) using supported file types (TMX or TSV).
  2. Index Creation: Generate an index for the uploaded dataset.
  3. On-the-Fly Customization: When translating, include the dataset index ID. The LLM applies few-shot learning during inference, returning domain-adapted translations.
  4. No Retraining: New content or terminology is handled by updating the index and uploading revised datasets—no model training downtime.

Example Workflow: Setting Up AdaptCT

Step 1: Create a Project

curl -X POST "https://<your-resource>.cognitiveservices.azure.com/translator/customtranslator/api/texttranslator/v1.0/workspaces" \
  -H "Ocp-Apim-Subscription-Key: <your-key>" \
  -H "Content-Type: application/json" \
  -d '{ "name": "my-translation-project", "subscription": { "billingRegionCode": "<billing-region>", "subscriptionKey": "<your-key>" } }'

Step 2: Upload Translation Pairs

For example, upload a TMX or TSV file with aligned sentences.

curl -X POST "https://<your-resource>.cognitiveservices.azure.com/translator/customtranslator/api/texttranslator/v1.0/documents/import?workspaceId=<workspace-id>" \
  -H "Authorization: Bearer <token>" \
  -F "DocumentDetails=[{\"DocumentName\":\"product-terms\",\"DocumentType\":\"Adaptive\",\"FileDetails\":[{\"Name\":\"translations.tmx\",\"LanguageCode\":\"en\",\"OverwriteIfExists\":true}]}]" \
  -F "FILES=@translations.tmx"

Step 3: Create the Index

curl -X POST "https://<your-resource>.cognitiveservices.azure.com/translator/customtranslator/api/texttranslator/v1.0/index?workspaceId=<workspace-id>" \
  -H "Content-Type: application/json" \
  -d '{ "documentIds": ["123456"], "IndexName": "product-index", "SourceLanguage": "en", "TargetLanguage": "fr" }'

Step 4: Translate Using the Index

  • In Microsoft Foundry, configure Text Translation with Adaptive customization.
  • Select/deploy GPT-4o or GPT-4o-mini.
  • Use the dataset index for instant, domain-specific translations (English ↔ French supported).
  • Alternatively, use the Text Translation API and pass your index as a parameter.

Supported Data Formats

  • TMX (Translation Memory eXchange) Example:
<?xml version="1.0" encoding="UTF-8"?>
<tmx version="1.4">
  <body>
    <tu>
      <tuv xml:lang="en"><seg>Click the submit button</seg></tuv>
      <tuv xml:lang="fr"><seg>Cliquez sur le bouton soumettre</seg></tuv>
    </tu>
  </body>
</tmx>
  • TSV (Tab-Separated Values) Example:
en	fr
Click the submit button	Cliquez sur le bouton soumettre
Upload your file	Téléchargez votre fichier

Decision Matrix: AdaptCT vs. Custom Translator

Scenario Choice
Low-volume, fast-changing content AdaptCT
Need updates within minutes AdaptCT
< 10K translation pairs AdaptCT
Support, chat, dynamic content AdaptCT
High-volume, consistent content Custom Translator
Strict terminology (legal, medical) Custom Translator
> 10K translation pairs Custom Translator
Rarely changing content Custom Translator

Best Practices

  • Start Small: Launch with 50–100 key sentence pairs; iterate as needed.
  • Domain Organization: Use indexes per domain/topic (e.g., support, docs).
  • Focus on Quality: Prioritize clean, impactful sentence pairs.
  • Versioning: Track changes and context across updates.
  • Monitor Results: Use feedback and metrics to refine datasets.
  • Alignment: Ensure high-quality alignment and context in input files.

When to Avoid Pitfalls

  1. Don’t bulk-upload many low-value pairs “just in case.”
  2. Keep TMX/TSV pairs well-aligned and context-rich.
  3. Provide full sentence context over isolated words.
  4. Use blind-testsets to validate translation quality on updates.

Benefits for Developers and Businesses

  • Dramatically faster delivery of multilingual features
  • Lower translation costs
  • Highly agile and maintainable translation pipelines
  • Reliable, continually-updated terminology

Additional Resources

Adaptive Custom Translation enables developers to deliver smarter, faster, and context-aware translations for global applications using Microsoft’s Azure AI services.

This post appeared first on “Microsoft AI Foundry Blog”. Read the entire article here