Machine translation converts text from one language to another. Neural models, and now large language models, have made it fluent enough for many everyday uses.
How It Works
Modern translation uses transformer models trained on large collections of parallel text (the same content in two languages). General-purpose LLMs can also translate well, and can follow instructions about tone, terminology and format.
Strengths
- Fast, cheap and available for many language pairs.
- Good for understanding the gist of foreign-language content.
- Useful for first drafts that a human translator then edits.
Weaknesses
- Quality drops for less-resourced languages and specialised domains.
- Terminology inconsistency across a document.
- Mistakes with idioms, names, numbers, units and cultural context.
- Fluent output can hide errors, because it reads naturally.
Improving Results
- Provide context: audience, purpose and domain.
- Supply a glossary of required terms and names not to translate.
- Translate complete documents or sections rather than isolated sentences.
- Keep source text clear and unambiguous.
When Human Review Is Essential
Legal, medical, safety-critical and contractual content, published marketing, and anything where an error has real consequences.
Evaluate Quality
Have fluent reviewers rate a sample, and track corrections. Automated metrics exist but correlate imperfectly with human judgement.