Neural machine translation
Neural machine translation (NMT) is machine translation built on neural networks. It is the current mainstream approach, producing markedly more fluent and context-aware output than the phrase-based statistical methods it replaced, though fluent output can still be confidently wrong.
How it works
Neural machine translation uses a neural network trained on large parallel corpora to translate whole sentences in context, rather than stitching together separately translated phrases. It represents meaning as numbers, then generates the target sentence token by token, weighing the whole source as it goes.
This context-awareness is why NMT reads so much more naturally than the statistical methods before it. The same fluency is a trap: output can be grammatically perfect and factually wrong, with nothing on the surface to signal the error.
How SourceTarget uses it
SourceTarget uses neural engines as the drafting step in its post-editing workflow, choosing the one that performs best for the content and language direction. Their fluency does not change the rule that a human reviews anything published, because fluent output can still be wrong.
Neural machine translation compared with Statistical machine translation
| Neural machine translation | Statistical machine translation | |
|---|---|---|
| Approach | One neural network, whole sentences in context | Separate models stitching phrases together |
| Fluency | High, natural word order | Often stilted |
| Weakness | Fluent but can be confidently wrong | Visibly rough, easier to distrust |
| Status | Current mainstream | Largely superseded |