Visit our company website at

Monday, May 15, 2017

Just how bad is neural machine translation?

Hearing that Google's neural machine translation quality has been over-hyped sounds strangely familiar. It brought to mind the claims made about a fantastic new approach called "statistical machine translation" made by dozens of players over the last decade.

Not to say that machine translation of whatever stripe does not have its place. Every day we are seeing the creative use of these tools. However, developers cannot seem to resist making the claim that their results are "as good – or almost as good – as human translation." Such claims in the past have proven to be nonsense--and are usually the verbalization of fantasies of developers and those who had to justify the amount of money they threw at the process.

For the translation community, the fact that translation done by a computer is still not much better is good reason to relax a bit. These latest claims will always get a lot of attention when initially made because, such a development, if true, can impact the lives and livelihoods of many linguists all over the world. They fulfill the dream of the large localization companies that spend their days trying to eliminate translators from the translation process. However, professional translators see their work as a creative art form, not something that can replaced by a machine.

Showing graphs of how good machine translation results are is like telling people how beautiful your sister is, but not letting anyone see her. Show us the actual results of the machine translation output without any tinkering or "adjustments" afterwards, and we can then see for ourselves how good it really is. And start with marketing texts for fashion products into Japanese!

In the meantime, don't waste our time. We are busy providing top quality real translation work that our customers can rely on for accuracy. This means they have confidence it will be fit for the purpose for their target audience... And we do it within a reasonable budget and bring the projects in on schedule... every time. And when it is done, the customers can ask a real live linguist why they made a choice of terminology, or style, or tone, or any one of the other facets of final translation work that must be taken into account during the process.

Can neural machine translation really take any text and instantly spew out an "almost as good" translation? Even work by top-notch linguists is often rejected by other linguists. It is likely that the claims of "almost as good" translation still won't stand up in the real world where translation has to be created for real people to read and understand.