What is neural machine translation? – FLaT FAQs

A type of "machine translation" called Phrase-Based Statistical Machine Translation (PBSMT) was commonly used up until the first half of the 2010s. PBSMT generates a translation by combining individual word strings, but often resulted in issues with overall fluency because sentences are composed by cutting and pasting these word strings.

FLaT uses Neural Machine Translation (NMT). Instead of working with individual word strings, NMT converts entire sentences into numeric data, and then generates a new translation from this numeric data. This can provide a more fluent translation result than PBSMT.

Features of neural machine translation

Important characteristics

It is difficult to tell why a given translation will go well.
-> It is not clear how to fix translation errors.
The translation may not cover for any excesses or deficiencies in the source text.
-> It may leave text untranslated, or generate excessive text.
It is sensitive to input differences, handles word inconsistencies poorly, and translation results may vary greatly depending on the presence or absence of punctuation (unreliable).
It is difficult to control and apply dictionaries.
-> Be sure to read "4. User dictionaries" thoroughly prior to using user dictionaries.

Refer to "Translation Tips" for more information.

Features of neural machine translation

Related articles