Machine Translation

When and how to use machine translation?

Machine translations have been with us for many years. However, it is only recently that they have become more widely known, mainly due to the improving quality of the translated text.
Machine translation has moved to the next stage of development - the earlier statistical (SMT - Statistical Machine Translation) models are being replaced by neural ones (NMT - Neural Machine Translation).
Simply put, machine translators have started to learn. Not only from the data we provide, but also from it’s own mistakes.

But this seemingly small step is a big step for the final quality of the translated text.

However, it is still true that there is no substitute for human translation - so this does not mean the end of translators or translation agencies. It's just that the translation work itself is changing, or rather the services associated with translation.

Aspena uses two main machine translators for customers interested in machine translation - DeepL and Google Cloud Translate
Each is suitable for something different: DeepL offers a limited number of languages compared to Google and is slower, but the resulting translation is often far better.
Other compilers we work with include KantanMT, Omniscien, Amazon MT and Tilde MT.

Now we come to the basic question - when and how to use machine translation.

As mentioned before, human translation cannot be replaced by machine translation. However, there are cases where the use of machine translation may be appropriate - for example, for internal familiarisation with a text in another language or for testing a multilingual environment when writing software or an application.

Let's look at a basic translation. It is defined as a transfer from one natural language to another, while preserving the meaning. 

After translation by a human translator, you will have a text whose meaning you can understand without any knowledge of the source language.

When using machine translation, you will have a text that can be simplistically described as follows:

  • much of it you will understand (how much depends on the translator used, the language combination, the source text, the translator’s level of training, etc.),
  • part of the text will be more likely to be understandable to someone who knows the source language (guessing how the machine "meant" it when it translated the text),
  • part of the text may not make any sense.

Not so long ago, the incomprehensible (even nonsensical) parts after machine translation came to 10, 20 or even 30 percent of the total text volume. This is changing rapidly, and it is thanks to the use of neural networks and deep learning.

However, it is dangerous to rely on machine-translated documents to ensure 100% clarity.

Machine translations without further editing are not suitable for publication. This is only ensured by the next step involving a human translator; that is Machine Translation Post-Editing (MTPE).

Aspena offers two post-editing options:

Light Post-Editing - the text is reviewed by a human translator, the aim is to ensure only the most necessary corrections to the text so that the result as a whole is understandable to the end user - it does not deal with stylistics, it does not address preferential changes. Its purpose is not to provide a translation that is comparable to a human translation. It is largely used for urgent internal translations.

Full Post-Editing - a deeper revision of the text to achieve a quality comparable to human translation. This may mean that the translator translates some parts again and ignores the machine-translated text. 

From our point of view, machine translation is one of the services in our portfolio - it is not a substitute for traditional translation.

It is important to note that not all documents are suitable for machine translation. These are mainly documents that contain creative texts or specific styles of expression (irony, jokes, poetic turns of phrase, etc.)

On the other hand, technical, stereotyped, standardized texts are suitable.

It is worth noting that even a text that is originally unsuitable for machine translation can become a well-translated text. It depends on the right combination of machine translator and human translator.

An equally important thing to consider is whether the customer should produce the machine translation itself or outsource it to a translation agency.

DeepL and Google also offer translations by copying texts into a web interface and then copying them back into the document. This can be laborious and tedious and the formatting is not preserved.
Unlike Google, DeepL also offers translations of entire files (currently only MS Word and PowerPoint), but the number and size of documents is limited (3 documents per month). Even in the case of paid versions, the number of documents and volume is limited.
For the record, Google offers translations of basic documents in their Google Docs web app.
DeepL also offers a desktop application for Windows and Mac which simplifies the work, but still has its limits.

But what should I do if I don’t want to be limited by the size and type of the source file?
When I need to translate web pages, items in software code, database files or complex graphic printouts created in inDesign?
And even more specifically - what if I need to translate, for example, an e-shop that has several million words?

This is where Aspena enters the stage as a translation agency.
For customers who require machine translation, this is just one step in the processing process of our standard workflow, we are not limited by file size, number of documents or source format.

In addition to standard formats such as MS Word, Excel, PowerPoint, etc., we can provide translations of the following formats:

  • xml, resx, po, json, csv files....
  • web html and php files....
  • graphic documents created in inDesign, Illustrator, FrameMaker, AutoCAD....
  • txt, inf, ini, reg text files....
  • and others

A great advantage is the connection between machine translators and our CAT (Computer Assisted Translation) tools. This is a software environment for processing translations using translation memories and terminology dictionaries that are individual for each customer.
But beware of the common misconception about what a CAT tool actually is - it is not machine translation, it is just an environment in which professional translators work. 

The combination of a CAT tool and a machine translator is always more profitable for our regular customers than a simple translation via the web interface or the DeepL application, Google or other engines (translators).
Translation memories and terminology dictionaries ensure better consistency and more accurate translations.

We are also able to train these translators on the basis of customer data (usually using previously produced translations). Currently, DeepL does not allow training, but Google and other tools do.

Now all that remains to do is to write to us at and we will be happy to advise you on the most suitable way to process your translations, either using machine translation or the standard way with a human translator.

It is very likely that any variant of machine translation will involve working with a human translator - either Light or Full Post-Editing, or a combination of linguistic, technical and pre-press proofreading. There are many possibilities and combinations, it always depends on what our customers expect and need.



Author: Břetislav Svoboda, Head of Technical and Sales Development