How does machine translation actually work now?
Through neural networks trained on enormous quantities of translated text, which replaced the earlier statistical and rule-based approaches entirely and produced a step change in quality.
The three generations:
Rule-based. Hand-written grammatical rules and bilingual dictionaries. Predictable and brittle, requiring enormous expert effort per language pair, and poor at anything idiomatic.
Statistical (SMT). From the 1990s. Learn from parallel corpora — the same documents in two languages, with EU and Canadian parliamentary proceedings being famously useful sources. The system learns which phrases correspond and how to order them probabilistically. Better, and characteristically produced locally plausible output that fell apart across a sentence.
Neural (NMT). From around 2016, and now standard. A network encodes the whole source sentence into a numerical representation and generates the target from it, rather than assembling phrase by phrase. The transformer architecture, with its attention mechanism, allows the model to weigh every part of the input when producing each output word — which is what solved long-range dependencies, the persistent weakness of everything before it.
Why output improved so much: fluency, because the model generates natural sequences; consistency across a sentence; and handling of grammatical agreement at a distance.
What it still gets wrong:
Hallucination of fluent nonsense. NMT output reads confidently even when wrong, which is more dangerous than obviously broken output because errors are harder to spot.
Low-resource languages. Quality depends heavily on training data volume, so languages with little parallel text are served poorly — which reproduces existing inequalities.
Context beyond the sentence. Most systems translate sentence by sentence, losing pronoun reference, formality level and terminology consistency across a document.
Gender and formality, where the source is ambiguous and the target requires a choice, producing well-documented stereotyped defaults.
Domain-specific and legal terminology, where a plausible near-synonym is wrong.
Named entities and rare words, which are frequently mangled.
Where it is genuinely used professionally: as machine translation post-editing, where a human corrects the output — now standard practice, and a different job from translating.