DeepL releases next-generation translation AI, surpassing GPT-4 in translation performance

DeepL, a company that provides highly accurate translation tools, has announced the implementation of a next-generation language model with improved translation performance. DeepL's next-generation language model is said to be able to translate with higher accuracy than Google Translate and GPT-4.

DeepL implements next-generation LLM model that outperforms ChatGPT-4, Google, and Microsoft in translation

https://www.deepl.com/ja/blog/next-gen-language-model

DeepL is a web app that uses AI to translate texts with high accuracy, and can translate between a variety of languages, including Japanese and English. Below is the result of translating the opening of the Declaration of Independence into Japanese using DeepL. You can see that the translation is done with a fairly high degree of accuracy.

As mentioned above, DeepL is equipped with a language model that is capable of high-quality translation, but it has announced that it will implement a next-generation language model that will improve translation performance even further than the existing language model.

DeepL's next-generation language model uses 'more than seven years of unique data for translation and content creation' as training data and is tuned for use in language fields such as translation and proofreading. At the time of writing, the next-generation language model supports translation of Japanese, English, German, and Chinese (Simplified), and more languages are planned to be supported in the future.

DeepL conducted a blind test comparing the translation quality of the 'conventional language model' and the 'next-generation language model' with professional translators. The results showed that the next-generation language model was 1.7 times better in 'Japanese-English mutual translation', 1.7 times better in 'Chinese (Simplified)-English mutual translation', and 1.4 times better in 'German-English mutual translation' than the conventional language model.

In addition, a blind test was conducted to compare the translation quality of 'DeepL's next-generation language model,' 'Google Translate,' 'GPT-4,' and 'Microsoft Translate.' The results showed that DeepL's next-generation language model was rated as having 'better translation quality than other products' 1.3 times more frequently than Google Translate, 1.7 times more frequently than GPT-4, and 2.3 times more frequently than Microsoft Translate. Furthermore, it was revealed that to achieve the same translation quality as DeepL's next-generation language model using other products, Google Translate requires twice as much editing work, and GPT-4 requires three times as much editing work.

DeepL's next-generation language model has already been implemented for paid DeepL Pro plan subscribers, and it is possible to switch to the next-generation language model with the touch of a button. The monthly fee for the DeepL Pro plan is 1,150 yen (tax included) for the Starter plan, 3,750 yen (tax included) for the Advanced plan, and 7,500 yen (tax included) for the Ultimate plan.