When you paste a paragraph into a translation app and get fluent output in seconds, that is usually NMT. The model reads the whole source sentence, builds an internal representation of its meaning, and generates the translation word by word. Earlier systems translated short phrases separately and stitched them together, which often produced choppy results.
NMT took off around 2014 with sequence-to-sequence models, and attention mechanisms were first popularized to improve it. Google moved Google Translate to a neural system in 2016. The transformer, now the basis of large language models, was introduced in a translation paper.
Why it matters
NMT made translation far more fluent, but fluency is not the same as accuracy. Automated scores such as the BLEU score and human review are used to check whether the meaning survived.
On the podcast
In episode 20, Smartling’s Olga Beregovaya compares translation approaches and says “Quite often, neural machine translation would do much better” at 6:54 on accuracy in the cases she describes. That is her professional assessment. Galileo sponsored episode 20.
In episode 54, Dan Klein notes that with older machine translation, output that wasn’t fluent “tended to correlate with it being not very faithful,” so users had a warning sign at 1:05:26. His point is that systems which always sound right remove that signal. Galileo was the presenting sponsor of episode 54; the sponsor read is at 2:53.
Related terms
BLEU score is a common automatic translation metric. Natural language processing is the broader field.
Hear it from the guest
“Quite often, neural machine translation would do much better”
“So if you used a machine translation system and the output was not very fluent, that tended to correlate with it being not very faithful.”
Quotes lightly edited to remove filler words.
Sources
- Sutskever, Vinyals and Le: Sequence to Sequence Learning with Neural Networks — Shows an end-to-end neural network that maps a source sentence to a target sentence, applied to English-to-French translation; checked October 24, 2026.
- Bahdanau, Cho and Bengio: Neural Machine Translation by Jointly Learning to Align and Translate — Describes neural machine translation as a single network trained end to end and adds attention so the model can focus on relevant source words; checked October 24, 2026.
- Wu et al. (Google): Google's Neural Machine Translation System — Describes the neural system Google built for production translation and reports large error reductions compared with its phrase-based system; checked October 24, 2026.
