Machine translation at a glance
The use of artificial intelligence in the field of translation has recently shown many developments, now making it possible to translate text in a matter of seconds.
Artificial Intelligence is making life easier for people all over the world, from smart phones to self-driving cars. But the use of this new technology is not limited to phones or cars: it has recently been applied to translation as well. Applications like Google Translate, which millions of people use every day, are made with a type of Artificial Intelligence called neural machine translation (NMT). This technology can translate large bodies of text in just a few seconds with a high degree of accuracy. It has even been said that texts translated with this technology almost seem like they were translated by humans.
What are the different types of machine translation?
In the past, computer translations were rule-based, meaning that the computer calculated a translation based on a set of rules created by linguists. The computer would transform one language into another based on its grammar rules, and a dictionary for vocabulary. Nevertheless, many mistakes would slip through the cracks, as it is almost impossible to include all grammar rules and their exceptions in a computer programme. This is why statistical machine translation was developed shortly thereafter. This type of computer translation uses a system that learns how to translate from a set of texts. A large body of text is aligned with its translation into another language, and the computer analyses this to learn the translation rules. After going through a lot of sentences and their translations, the computer can use what it has learned to translate a new sentence. This system works rather well, but in the last few years it has been overtaken. The revolution of neural machine translation has made it possible to translate a lot of text in a matter of seconds.
How does neural machine translation (NMT) work?
Neural machine translation works especially well for languages with a different word order or word segmentation, and it can understand the most complex languages. The technology was modelled after the human brain, which works by delivering signals to neurons. The idea behind neural machine translation is that it learns how to translate in the same way the human brain works. Essentially, the system is trained with large networks of neurons to predict the next word in a sentence in the translation, based on the context of the original sentence. In each step of the translation, the neural network transforms the text from one language into another, based on a complex formula. After the training of the network, it can autonomically translate texts. It is difficult to pinpoint exactly how it manages to produce such good translations, because the concept is a “black box”. It’s impossible to see how exactly the process of neural machine translation unfolds, but it works, nonetheless.
Neural machine translation works best for repetitive and simple texts, at the moment. When it comes to creative texts, such as literature or poetry, neural machine translation is not able to translate the style of these types of text. Stylistic elements, such as phrases, humour and sarcasm are still a big challenge for NMT. For use in areas like marketing such as slogan translations for example, neural machine translation does not yet perform too well either. Structured and repetitive technical texts, on the other hand, can be translated rather accurately with neural machine translation.
Will manual intervention still be needed in the future?
Even if neural machine translations already seem to be flawless, errors still slip through. By means of post-editing, human translators can correct these faults, which normally range from spelling and syntax mistakes to different errors, such as incorrect translations or a lack of meaning, mistranslations etc.
After this process the final product is clear and understandable for the target public. To learn more about the challenges and implications of post-editing, read our upcoming post on this topic.
Can data protection still be guaranteed with data processed by MT?
Data protection is a key aspect to be taken into account when translating. However, when free online machine translation applications, such as DeepL, Reverso, Google Translate, are used, data will be uploaded and stored in the cloud and made available to the public. Furthermore, any data fed into the system could be also used as training data for developing underlying algorithms or even sold to other entities. So sensitive data is no longer protected.
Therefore, many clients specifically forbid the use and input of data into machine translation tools freely available on the Internet and require that their language services providers abide by these rules. Many would, instead, prefer to invest in proprietary machine translation solutions that guarantee maximum data protection.
NMT entails other risks. Our upcoming post will address these in detail. Stay tuned.
What does the future hold for translators?
Currently, texts processed with neural machine translation applications still need human intervention. Whether they will get better, or even replace the work of humans completely, remains still uncertain. But we know for sure: technology is evolving and so are neural machine translation programmes. That’s why it’s important to always pay attention to what technology has in store for us, so that we as service providers can evolve constantly and follow the actual market.
- FIT – The voice of associations of translators, terminologists and interpreters around the world – FIT Position Paper on Post-Editing: Position and Discussion Papers – FIT (fit-ift.org), 2021
- Manhart, K. (28.01.2011). Die Masse macht’s. Statische Verfahren. PC-Magazin. 13.09.2021
- Gelmetti, T. Ein Rennen gegen die Digitalisierung: Verdrängt künstliche Intelligenz menschliche Übersetzer?, 11.10.2021