Best Online Translation Service
The DeepL translator is an online service of DeepL GmbH in Cologne for machine translation, which went online on 28 August 2017. Hence DeepL, in our opinion, surpassed the translation quality of competitors such as Google and Microsoft in studies.
With the radical change in the approach to the development of multilingual translation systems, machine-made translations have gained considerably in quality. Rule-based, as well as statistics-based systems, got overtaken by the new paradigm of neural networks with deep learning. Thanks to their comparatively better results, moreover, these are now considered the reference technology par excellence.
In September 2016, Google announced that the newly developed “Google’s Neural Machine Translation System ” delivers better translation results across all language combinations. Not quite a year later, on 28 August 2017, DeepL went online. The new translation service astounded the media world in particular with translations that stood out considerably from Google Translate in terms of quality and seemed almost perfect at first glance. Since then, interest in DeepL and machine conversion increased. Moreover, DeepL gets used actively, especially in the free version with limited text length, and enjoys great popularity.
One million words < 1 second
Hence, DeepL relies on an artificial neural network running on a supercomputer in Iceland. It has enough power to translate a million words in less than a second.
The fact that at least the translation into German at DeepL could work better than with the competition from Silicon Valley, at least that is what various tests suggest.
DeepL vs Google Translate
Original text (source: Wikipedia DeepL):
Der Dienst verwendet Convolutional Neural Networks, die mit der Datenbank von Linguee trainiert wurden. Die Übersetzung wird mittels eines Supercomputers erzeugt, der 5,1 Petaflops erreicht und mit günstiger Wasserkraft auf Island betrieben wird. Convolutional Neural Networks sind für lange zusammenhängende Wortfolgen grundsätzlich etwas besser geeignet, wurden von der Konkurrenz jedoch aufgrund ihrer Schwächen bislang zugunsten von rekurrenten neuronalen Netzen nicht verwendet. Die Schwächen werden bei DeepL mit teilweise öffentlich bekannten zusätzlichen Kunstgriffen aufgefangen.
Auch Word-Dokumente im .docx-Format sowie PowerPoint-Präsentationen (.pptx) werden übersetzt. Fußnoten, Formatierungen und eingebettete Bilder bleiben dabei erhalten.
Übersetzt werden die Sprachen Deutsch, Englisch, Französisch, Spanisch, Italienisch, Niederländisch, Polnisch, Portugiesisch und Russisch jeweils in beide Richtungen.
The service uses Convolutional Neural Networks trained with the Linguee database. The translation is generated by a supercomputer that reaches 5.1 petaflops and is powered by cheap hydroelectric power in Iceland. Convolutional Neural Networks are in principle somewhat better suited for long, coherent word sequences, but have not been used by competitors so far in favour of recurrent neural networks due to their weaknesses. The weaknesses are compensated for at DeepL with additional tricks, some of which are publicly known.
Word documents in .docx format and PowerPoint presentations (.pptx) are also translated. Footnotes, formatting and embedded images are preserved.
The languages translated are English, German, French, Spanish, Italian, Dutch, Polish, Portuguese and Russian in both directions.
The service uses convolutional neural networks that get trained with the Linguee database. The translation generation is utilising a supercomputer. These computers reach 5.1 petaflops and operate with cheap hydropower in Iceland. Convolutional neural networks are generally a little more suitable for long coherent word sequences. Still, they have so far not gotten used by the competition due to their weaknesses in favour of recurrent neural networks. The failings at DeepL get partially compensated with publicly known additional tricks.
Word documents in .docx format as well as PowerPoint presentations (.pptx) are also translated. Footnotes, formatting and embedded images get preserved.
With “DeepL” the following languages German, English, French, Spanish, Italian, Dutch, Polish, Portuguese and Russian translate in both directions.
The test result clearly shows that DeepL Pro can deliver acceptable quality for understanding general language texts at this stage. However, it often happens that the translation produced by DeepL Pro sounds linguistically correct. The translated text seems to make sense in terms of content but does not correspond to the content of the source text. When it comes to translating specialist documents, DeepL Pro reaches its limits.
Hence, we use that solution for this website and are very satisfied with it. However, we found that DeepL works better if you have longer texts. Presumably, DeepL understands the subject matter better and also chooses appropriate technical terms. Yet, translations of single words via Google Translate is still a good option.
Finally, DeepL is, as you might expect, not wholly error-free. Therefore we recommend the small detour via grammarly.com, especially for translations into English.