DeepL expands its offering and uses large language model for translations
DeepL has presented a large language model for its translation service. To do this, DeepL worked together with linguists.
![Finger points to brain, finger of a robot too. Many letters float in the background.](https://heise.cloudimg.io/width/610/q85.png-lossy-85.webp-lossy-85.foil1/_www-heise-de_/imgs/18/4/6/3/5/5/9/0/shutterstock_2301264223-5a6cf97c36b6e082.jpeg)
(Image: Ole.CNX/Shutterstock.com)
The Cologne-based company DeepL has introduced a new Large Language Model (LLM) for its translation service. According to DeepL, the model is specifically tailored to language and has been trained using proprietary data collected over seven years. Thousands of language experts have also "tutored" the model.
In blind tests with language experts, translations from DeepL were preferred 1.3 times more often than those from Google Translate and 1.7 times more often than those from ChatGPT-4, the company reports. The number of manual corrections required after an automatic translation was twice as high with Google and three times as high with ChatGPT as with DeepL.
More natural translations, fewer errors
Since 2017, DeepL has been offering translation and writing solutions for companies that are based on specially developed, language-specific AI models. These are designed to deliver more accurate translations and reduce the risk of misinformation and hallucinations.
In addition, further blind tests with linguists have shown a 1.7-fold improvement in translation quality for the language combinations English-Japanese and "English-Simplified Chinese" and a 1.4-fold improvement for English-German compared to the previous model.
Translations with the new model are now available for "DeepL Pro" customers for English, Japanese, German and Simplified Chinese, with more languages to follow. Users can activate the new LLM in the web translator by selecting the "Next-Gen model".
Customers of DeepL Pro are "protected by enterprise-level security and compliance standards (ISO 27001 certification, GDPR/SOC 2 Type 2 compliance)". Their translations are never used to train the models.
"Just the beginning"
CEO and founder Jarek Kutylowski sees the new LLM as just the beginning for DeepL's language AI solutions for companies. The aim is to push the boundaries of translation and writing quality, efficiency and adaptability and offer customers industry-leading technology to overcome language barriers. In May, DeepL received venture capital of 300 million US dollars. The company was valued at 2 billion US dollars.
(mack)