Implement text analysis solutions 4.1 Apply language model text analysis
A manufacturer's manuals use industry terms that standard machine translation keeps…
Study guide skill: Customize language model outputs for domain tasks, such as compliance summarization and domain extraction
Question B1-040
A manufacturer's manuals use industry terms that standard machine translation keeps getting wrong. The company has years of previously translated manuals. What should it do?
Select one answer.
Answer: B
Custom Translator builds custom neural machine translation systems from parallel documents that capture your business and industry terminology, and it can also train from dictionary data. Document translation preserves formatting but uses the standard model, and Translator has no temperature setting.
Deep dive
How it works
Custom Translator is a feature of Azure Translator for building customized neural machine translation (NMT) systems. Its core idea fits the scenario: you build translation systems from parallel documents, meaning previously translated content, so they learn the terminology of your own business and industry. The docs note that previously translated documents capture domain-specific terminology and style better than a standard translation system does.
Practical details:
- You can upload many formats, including PDF, HTML, XLIFF, TMX, TXT, DOCX and XLSX.
- Data can be parallel at the document level. If you have the same content in separate documents per language, Custom Translator can automatically match sentences across them.
- If you don't have a training set, you can train a model with dictionary data only.
- You call your custom model from existing apps through the Translator Text API v3, and systems can be built and published to and from English.
The other options, and when you'd use them
- Switch to document translation: Document translation translates whole documents while preserving their original structure and format. It solves layout and file handling, not terminology. (It can apply a custom translation model, which still means building one first.)
- Lower the temperature of the Translator API: Temperature is a sampling setting for generative models that controls how random the output is. Wrong industry terms are a vocabulary problem, which training on your parallel documents addresses.
- PII redaction before translating: PII detection identifies, classifies and redacts sensitive data. It protects sensitive data rather than fixing translation terminology.
Exam tip
"Domain terminology" plus "existing human translations" means Custom Translator trained on parallel documents. "Keep the file format" means document translation.
References
Last reviewed: