• Technology. Innovation. Business.

    Knowledge to be shared

Domain Adaptive NMT Engines: a customized machine translation

2019-18-09  |  

In recent 10-15 years, we have put almost the entire workflow routine on computers. The same situation is observed in technical translation industry. The functions of machine translators are improved and expanded every year. They currently translate more precisely; and their main feature is an instant data processing. Although, non-comprehension of the context still remains a serious problem. Domain-Adaptive Engines are tools that help to solve this issue.

What’s about context?

It is sometimes absent in the original text. Moreover, when translating technical content, the meaning of many seemingly common words and phrases differs from the basic vocabulary. For instance, «A naked conductor runs under the tram.» Google Translate without context translates this short sentence into Russian in general terms: «Голый кондуктор бежит под трамваем.» A piquant situation that has nothing to do with reality; the sentence actually means the following: «Оголённый проводник проходит под вагоном», is quite a common technical text.

This blooper is caused by the absence of the context in the initial phrase. Google Translate inserts words of common vocabulary by default. This is might be right if you are not engaged in specialized themes (construction, IT, medicine, etc.). With such input data, you usually need a translation using dictionaries of your field, and for this the domain-adaptive engines have been developed.

What is Domain-Adaptive Engine and how it works?

This is the core of the machine translation system – the engine, which uses databases (cases) on a specific topic. You can install such a system on your website, load the databases corresponding to the industry in which your company involved, and you will have your own customized online translator. Technically, the system works the same as Google Translate, but it will use the vocabulary of your industry as a priority.

The domain-adaptive engines are widely used for technical translation in various fields:

  • Medicine;
  • Oil and Gas;
  • Engineering Surveys, etc.

If you constantly work with texts in other languages, for example, in order to communicate with partners from other countries or sell your products globally, launching your own domain-adaptive translation system will greatly simplify the work of your translators and save your resources.

Why are domain-adaptive engines better than regular online translation?

If we talk about a specific industry, the difference is significant. In 2018, Intento conducted an experiment: they analyzed the accuracy of the domain-adaptive and standard medical technical translation from English into German.

The following 6 domain-adaptive engines were involved in the research:

  • Google Cloud AutoML Translation;
  • IBM Cloud Language Translator v3;
  • Microsoft Custom Translate v3;
  • Enterprise AI ModernMT;
  • Tilde Custom MT;
  • Globalese Custom NMT.

These engines were compared with their «base» versions, as well as with products of other companies: Baidu (China), Amazon (USA), Systran (USA), Yandex (Russia), etc. As a result, it turned out that systems with connected specialized cases are 5-20% more efficient than standard ones. Follow the link for more information about the research results and testing methods.

Thus, machine technical translation becomes faster and more accurate, requires less human attention. For the enterprise, this means that the cost of translation decreases. Of course, these systems are still not perfect but they are instantly developing. Nevertheless, the future of the professional technical translation may well turn out to be precisely for domain-adaptive engines.

In Russia, this topic was discussed in detail on 25 May 2019 in Moscow, at the Seventh Hyperbaton from Yandex.  The recording of the broadcast is available here, and the related article on the domain-adaptive engines can be  downloaded here.

Request a translation