Machine Translation
Translating between languages automatically — the task that invented modern NLP, where fluency arrived long before reliability.
Reviewed July 12, 2026Stable
When not to use it
- For anything you'd sign. Fluent errors are invisible if you can't read the source. Contracts, medical, legal — human, with review.
- On low-resource languages, unchecked. Quality tracks corpus size, and most of the world's languages have small corpora.
- Sentence by sentence, for a document. Pronouns lose referents, terminology drifts, register wanders.
- When the text is culturally loaded. Idiom, humour, register, connotation. Those don't map, and the system will produce something fluent regardless.
Reach for something else instead
- Post-editing — machine draft, human fix. Faster than from scratch, and how professional translation now works.
- Human translation — for anything published or consequential.
- Controlled source language — write the original to be translatable: short sentences, no idiom, consistent terms.
- Terminology-constrained MT — force specific term translations. Underused and effective for technical content.
Read more on the blog
- AI in translation: the field that retired its own metricTranslation has fifty years of formal evaluation practice. In 2022 its own shared task published under the title "Stop using BLEU", retiring the metric the field had run on for twenty years.
- Where the attention mechanism actually came fromAttention was a fix for a specific engineering failure in 2014, three years before the paper that made it famous. It was not designed as a theory of cognition, and the name was applied afterwards by analogy.
- What is natural language processing (NLP)?Every search, translation, voice assistant, and chatbot runs on natural language processing, the field of getting computers to work with human language. Its defining modern story is a quiet revolution: the dozens of separate, task-specific methods that made up NLP for decades collapsed into a single general approach, the large language model.
Further reading
- Bahdanau, Cho & Bengio (2015), Neural Machine Translation by Jointly Learning to Align and Translate — attention, invented for translation, and the ancestor of the transformer.
- Papineni et al. (2002), BLEU: a Method for Automatic Evaluation of Machine Translation — the metric everyone knows is inadequate and still uses.
- Läubli, Sennrich & Volk (2018), Has Machine Translation Achieved Human Parity? A Case for Document-level Evaluation — how the parity claims dissolved under better evaluation.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Trusting fluency. A wrong translation reads exactly as well as a right one.
- Judging quality on BLEU. It rewards word overlap and penalises legitimate paraphrase.
- Assuming quality transfers across language pairs. It tracks corpus size, and that varies enormously.
- Translating documents sentence-by-sentence and losing everything that spans sentences.
- Not noticing that gender and formality were invented for you, usually along stereotype lines.
At a glance
FieldApplied AI
Historical roleinvented attention, hence transformers
Signature failurefluent when wrong
MetricBLEU, known inadequate
Quality trackscorpus size, not technology
Right shape for productionpost-editing
DifficultyBeginner
Often compared with
Machine translation vs. post-editing — raw output is for gisting; machine draft plus human fix is for publishing. The second is how professional translation actually works now.
Where this sits
A destination. 26 concepts lead here, and nothing in the corpus depends on it.
12Levelsteps in
26Needs firstconcepts
0Opens upnothing further
1Areastays here
Computed from the prerequisite graph, not assigned. How this works