Machine Translation (MT)

Definition

Machine translation (MT) is the use of software to translate text from one language to another. Modern MT systems are based on neural machine translation (NMT) — deep learning models trained on billions of parallel sentence pairs. Unlike the older rule-based (RBMT) and statistical (SMT) approaches, NMT models learn contextual patterns across entire sentences, producing output that is far more fluent and natural.

Well-known MT engines include Google Translate (NMT since 2016), DeepL (transformer-based, praised for European language quality), and OpenAI GPT models, which combine translation capability with instruction-following and domain awareness. For high-traffic language pairs such as English→Spanish or English→French, modern MT output can be indistinguishable from human translation for many document types.

MT still has failure modes: low-resource languages (those with less training data), heavily idiomatic or cultural content, technical jargon in narrow domains, and ambiguous sentences where context beyond the document is required. This is why post-editing (human correction of MT output) remains a standard practice for publication-quality work.

Key Concepts

• Neural Machine Translation (NMT): The current dominant paradigm. Uses transformer neural networks trained on massive corpora of parallel texts. Captures long-range context within a sentence far better than statistical models.
• Domain adaptation: Fine-tuning or prompting an MT model on domain-specific terminology (legal, medical, engineering) to improve accuracy. Generic MT engines may mistranslate specialist jargon.
• MTPE (Machine Translation Post-Editing): A workflow where a human editor corrects MT output rather than translating from scratch. Typically 30–60% faster than full human translation at comparable quality when MT quality is high.

Related Terms