So here’s a quick overview of NMT. If you’ve not heard of it then it might be a good idea to get a little more familiar with it as this is likely to come up in conversations quite a lot over the coming months.
A handy little overview can be found here
And for what it’s worth here’s my take on it.
Neural Machine Translation has been reported to be so good that native speakers can barely tell the difference between that and human translation… Well that’s the headline we’re reading all over the web right now. Fortunately for our industry this is, once again, a very murky window into reality.
If you ask native speakers, as they have done in this study, to review a completed translated text in their own language they will award it merit upon how easy it is to read, both grammatically and contextually. However without having visibility and being able to understand the source text this is a flat and singular viewpoint 2D if you’d like.
Upon deeper investigation this is where the serious flaw of this new art of machine translation lies. In fact it is more dangerous than you can imagine. Let me elaborate.
The translated text is highly inaccurate when compared to the source text, so much so that regular Google MT translated text is significantly more accurate. As we all know the accuracy of Google MT translated material is already decreasing through their own admission of “Garbage in, Garbage out”. So the main issue here with NMT is, unlike regular Statistical Machine Translation, that the way it is written it is conceived to be true; even though it can be a wholly inaccurate account of the source material. Think misleading information from the gutter press. You take a single statement out of context and build a story around it that sounds believable but in reality is devoid of almost all truth.
As it stands right now, this is where NMT sits. It is in the long line of attempts to automate a human process that really cannot be done, particularly if accuracy and accountability are your end goals.
Yes there are very legitimate and viable uses of MT and a number of companies have them. By using a hybrid of human and machine it can eliminate a huge amount of unnecessary overheads, but here’s the heads up. Please be wary of the claims to NMT; it isn’t there yet and for all the good intentions it is not likely to be for some considerable time.