neural machine translation deepl

At the end of August DeepL Translator, a new online translator has been launched, which is expected to beat the greatest rival, Google Translate over time. It might seem to be a bold statement, as the translation services provided by the web giant are currently available in over 100 languages, while DeepL only supports translation between seven languages (English, German, French, Spanish, Italian, Dutch and Polish), in a total of 42 combinations. Development of new languages such as Mandarin, Japanese and Russian is in the process but there is no doubt that the new translator can no way compete with similar services already on the market, at least not in terms of the number of languages provided.

On the service itself, only limited information is available most of which is provided in the press information published on the deepl.com website. According to this, DeepL outperformed the competition in a blind test, making it “the world’s best translation machine”. Four online translators, DeepL Translator, Google Translate, Microsoft Translator, and Facebook were fed 100 sentences to translate from English into German, French and Spanish and from these three languages into English. Professional translators were then asked to assess the translations, without knowing which system delivered which target text. In fact, the new online translator achieved way better results than the competition: Translators preferred DeepL’s translations by a factor of 3:1 compared to the second contestant, Google Translate. It may sound surprising as most users still prefer Google Translate above all other online machine translation services as it is considered to be the most powerful and effective choice. Now, let’s take a closer look on the newest competitor.

DeepL Translator is the product of the German tech company, DeepL that is specialised in developing products and services based on artificial intelligence. If the company name still does not ring a bell, their former name may sound familiar: in 2017 the company changed its name from Linguee to better reflect their objective of creating neural network-based products. Linguee.com is a popular web-based dictionary and translator service launched by the company back in 2009. This will serve as the basis in the development process of DeepL Translator. Let’s see how.

DeepL Translator is based on the concept of artificial neural networks. This technology may be familiar to most readers, as it was introduced in the world of artificial intelligence in 2014. In the translation industry, it represents a new paradigm under the name neural machine translation (NMT). DeepL’s greatest competitor, Google has started to use this cutting edge technology in the translation engines of Google Translate in certain language pairs, and Microsoft has also incorporated NMT in their speech-based translation systems such as Skype Translator. The neural network technology is basically an artificial neural network that imitates the operation of human neural cells. These cells, and thus the neural network is capable of learning including deep learning just as human brain does.

Gereon Frahling, founder and CEO of DeepL, who by the way also used to work for Google stated that “by arranging the neurons and their connections differently, we have enabled our networks to map natural language more comprehensively than any other neural network to date.” According to the developers, the performance of a neural network highly depends on the quality of the material it trains on. The neural networks of DeepL Translator train on billions of high-quality translated sentences provided by the search engine Linguee which gives DeepL an unassailable advantage.

Linguee uses special web crawlers to search the Internet for bilingual texts including various European Union and patent documents translated by human professionals. It currently offers access to over 100 million multilingual websites and this number is continuously growing. The bilingual texts identified undergo automatic quality evaluation by a special algorithm and are stored on Linguee’s servers in bilingual form. The online platform allows the users to provide feedback and rate translations manually, thus further improving their quality.

The system of DeepL runs on the 23rd largest supercomputer of the world which was built in Iceland. The country was an ideal location to use renewable energy in a cost-effective way so that this computer allows DeepL to translate 1,000,000 words in under a second.

Journalists at the French newspaper Le Monde tested the newly introduced web translator that is claimed to be “three times more powerful” than Google Translate: DeepL was compared to other translation services provided by the servers of Google, Bing, Yandex and Baidu by having them translated various types of text including a poem by Emily Dickinson, a paragraph from the user manual of PlayStation 4, an archaeology-related newspaper article, a ministerial speech written in first person and a report on a sports match from English into French. Although the test was far from exhaustive, DeepL clearly outperformed the competitors in translating the poem, the newspaper article and the first-person speech. In most types of texts the results given by Yandex and Baidu were no way near to the quality of translation provided by the American giants and their German challenger. The testers concluded that the results demonstrate that the progress in online translation is real. However, it is always easy to trick an automatic translator, and even if the sentences are correctly translated, the reader can still sense the ‘coldness’ of the machine. The performance of online translation services is still far from that of human translators.

Sources:

deepl.com

LeMonde.fr