What Is NMT?
NMT (Neural Machine Translation) is a machine translation approach where a neural network learns to translate a whole sentence, drawing on its full context, instead of assembling a translation from individual words and phrases using statistical lookup tables. NMT replaced earlier statistical machine translation (SMT) in the mid-2010s and became the foundation for every modern translation system, including translation through large language models.
Why NMT Beats Statistical Translation
Statistical systems translated text in chunks, picking the most likely translation for each piece from frequency tables and then stitching the pieces together — which often produced grammatically inconsistent results across chunk boundaries. NMT trains end-to-end on sentence pairs and considers the whole sentence's context when choosing each word of the translation, producing noticeably more natural and coherent output.
From NMT to LLM Translation
Modern large language models used for translation are, in essence, an extension of NMT ideas at a much larger scale: the same core principle (attending to the whole sentence's context via the transformer architecture), trained on an incomparably larger volume of text and able to consider context beyond a single sentence — the whole document, and sometimes a user's instructions about style and terminology.
How NMT Quality Gets Measured
NMT systems are evaluated with the same metrics used for translation in general: historically BLEU, and today primarily COMET and LLM-judge approaches. Comparing different NMT systems on the same test set is the standard way to pick a translation engine for a specific language pair.
Limits of NMT
NMT systems, including LLMs, can sound confident and fluent even when they translate incorrectly — unlike older systems, which more often produced noticeably broken but less misleading output. That's one reason automated quality scoring (QE) and human review remain necessary even as NMT's overall quality improves.
FAQ
Are NMT and LLM translation the same thing?
Not quite: NMT is the broader term, covering any neural translation system, including specialized models trained only for translation. LLM translation is a specific case of NMT that uses a general-purpose model capable of much more than translation.
Why did NMT replace statistical translation?
Because it translates more coherently and naturally: statistical systems struggled to keep chunks consistent with each other, while NMT considers the whole sentence's context at once.
Which metric best captures NMT quality?
There's no single best metric — the current approach combines COMET with LLM-judge scoring rather than relying on one number.
Does NMT need specialized infrastructure?
Yes, training and running NMT models takes significant compute, which is why most companies use ready-made models through an API rather than training their own.
Can NMT translate rare language pairs as well as common ones?
Generally no — NMT quality depends heavily on how much training data exists for a given language pair, and rare pairs get noticeably lower-quality translation.