DeepL: the new machine translation tool that is stealing Google Translator’s limelight
Advances made in artificial intelligence and their application to the field of machine translation never cease to amaze us. Although we tend to dwell on the errors produced by machine translation, the truth is that the results obtained are getting better and better over time. Initially, the software used simply substituted one word for another, without taking the context into account. This system was only valid to search for vocabulary or very short sentences, without considering syntactic structures or grammar. A real leap forward was taken in 2014, when the University of Montreal opened up the world of deep learning to the field of machine translation. Deep learning uses advances in neuroscience and artificial neural networks along with algorithms to improve the different fields of artificial intelligence. To date, nobody has been able to compete against the multidisciplinary giant Google, whose system is based on a multilingual corpus that is updated on a daily basis.
On 29 August 2017, a European company, DeepL, revealed a new machine translation tool. Although its name is unknown, it is based on the famous online dictionary, Linguee. Linguee provides users with several translations found on the Internet, taken from a corpus of documents published on the Web. It also enables users to consult the source on which the translation was found, ensuring greater confidence. The huge database pertaining to this translation search engine, updated using reliable sources such as the European Union, serves as the basis for the training and learning of the new machine translator. The results are spectacular in many different fields: technical documentation, press articles, sports analyses, etc.
What is behind such good results?
For strategic reasons, the company does not want to reveal its secret, but it stands to reason that its success is related to the two factors that we will list below:
Firstly, it can be understood that one of the keys for the success of a machine translator is the quality of the translations with which it is updated. Besides which, this philosophy is the strategy followed and the key to the success of the parent company, Linguee. Before introducing any text into their database, it must obtain a minimum score in their algorithm, first validated by a human. Compared to the amount of data manipulated by the giant Google which, incidentally, has over one million servers and data centres throughout the world, DeepL prefers quality.
Secondly, the application of the new paradigm based on neural networks or deep learning has enabled considerable progress. It is highly likely that the programmers behind this machine translator have used, optimised and personalised a learning paradigm based on one of the existing neural networks.
Even so, it is still easy to deceive a machine translator. Their capacity to translate long sentences, literary or commercial texts or any type of text that implies cultural knowledge is still extremely limited.
Josh Gambin holds a 5-year degree in Biology from the University of Valencia (Spain) and a 4-year degree in Translation and Interpreting from the University of Granada (Spain). He has worked as a freelance translator, in-house translator, desktop publisher and project manager. From 2002, he is a founding member of AbroadLlink and currently works as Marketing and Sales Manager.