A Bit of History

In 2017, we tested several engines for neural machine translation from English to Russian language and came to the conclusion not to use machine translation in our company. The reason is that the sentence structure in Russian looks the same as in English after machine translation. A proofreader recognizes and corrects loan-translations during the first 30–60 minutes, but then a human brain starts taking it for granted. Translation quality drops noticeably.

Delight and Apathy at the Same Time

When I was writing this article, I decided to update my knowledge and tested the DeepL neural machine translation engine. It appeared on the market in May 2017 and at that time it was technically impossible to connect it to our CAT tool. We translated the contract from English into Russian. The result caused delight and apathy at the same time.

On the one hand, I was pleased with the decent quality of the translation without the ubiquitous English calque, with the correct terminology and stylistics of the Russian language. Of course, one should read and proofread the translation, correct tags, few terms, and calque, which occurs sometimes. However, in general, the translation was of a good quality.

On the other hand, I was attacked by apathy. In the next 5-10 years, translators will become editors and proofreaders for machine translation. Later, machines may replace 50-80% of humans. That is the reality. Artificial intelligence has been improving every day.

Which Text Genres are Easy to Translate for Machines Today?

The simpler the sentence structure is, the better the machine translation quality is. It means, nowadays, common-purpose documentation can and should be translated by neural machine translation engines. These are technical and legal documents. These two fields are the world market leaders in terms of the number of translations. There are huge translation memories made by humans. Translation memory is a database that helps to train neural networks. I believe that the next fields will be medical and marketing ones, followed by all other. There are many complicated speech patterns in fiction and, likely, machines won’t be able to translate such texts properly, although it is a matter of time.

The Conclusion

We are going to implement neural machine translation networks to translate legal and technical documents in 2020. Of course, firstly, we will run a few dozen tests and adapt the quality control guidelines to the new realities. Unfortunately, most clients now are completely against machine translation as it is difficult for those not being a specialist to understand the difference between a classical machine translation and neural machine translation engines. But let us remember the implementation of computer-assisted translation systems (CAT tools) and their benefits for clients. In the beginning, they were considered as “Google translation” and translators were asked to translate texts by hand, but today the use of CAT tools, translation memories and glossaries has become a sine qua non. The same is going to happen to neural machine translation systems, in other words, the translation of documents by artificial intelligence.

By the way, DeepL has a free online translator. Comparing to Google Translate, there are fewer languages, but the level of translation is much better. It is perfect for translating business correspondence and any other texts. I recommend it.

Write comments, click the “Like” button, and share the article with friends. I tried to make it interesting.