Technical editor: Ming Fei from Beijing
SegmentFault has he reported the official account number: SegmentFault
Technology giants Google, Microsoft and Facebook are all applying machine learning lessons to translation, but a small company called deep has surpassed them and raised the standard in this field. The speed of its translation tools is no less than those of its large-scale competitors, but it is more accurate and detailed than any translation tool we have tried.
After several experiments, we all think that the translation of deepl is generally better than that of Google translate and Bing. Google Translate often looks for a very straightforward translator, missing some subtle differences and idioms (or translating these idioms incorrectly), while deepl often provides a more natural translation, closer to the translation of a trained translator.
Deepl evolved from linguiee
Deepl was born out of the equally excellent linguiee, a translation tool that has existed for many years. Although it is very popular, it has not reached the level of Google translation – after all, the latter has great advantages in brand and status. Geeon frahling, co-founder of linguee, once worked for Google research, but left Google in 2007 to start a new business, linguee.
The team has been working on machine learning for many years, working on tasks adjacent to core translation, but it was only last year that they began to seriously study a new system and company, both of which will be named deep.
Frahling mentioned that the time is ripe: “we have set up a neural translation network, which contains most of the latest developments, and we have added our own ideas to it. “
A huge database composed of more than 1 billion translations and queries, together with the method of landing translation by searching similar fragments on the Internet, has laid a solid foundation for the training of the new model. They also put together what they claim to be the world’s 23rd most powerful supercomputer in Iceland.
Deepl’s translation service uses convolutional neural networks built on the linguee database and another proprietary method that has not been published, involving attention mechanism. Deep GmbH has a 5 petaflops floating-point performance machine for training and production of its translation services.
Developments announced by universities, research institutions and competitors of linguiee show that convolutional neural network is the direction of development, rather than the recurrent neural network that the company has been using. Now it is not really the place to study the difference between CNNs and rnnns, so it must be said that for the accurate translation of long and complex Related words, as long as you can control its weaknesses, the former is a better choice.
CNN, for example, is roughly a sentence that can solve one word at a time. But when, as often happens, a word at the end of a sentence determines how the word at the beginning of a sentence should be formed. Reading the whole sentence, only to find that the first word selected by the network is wrong, and then start again based on this knowledge is very wasteful. Therefore, deep and other people in the field of machine learning have applied the “attention mechanism” to monitor this potential stumbling block and solve them before CNN moves on to the next word or phrase.
Neither deep Pro nor the free deepl translator is allowed to be used to translate “text containing any kind of personal data.” unlike the free version, deep Pro claims that it does not store translated text. See their privacy for more.
Interested students can have a try, I believe deep will become your new productivity tool to help you translate. We also welcome to share more details about the technology behind deep.