🗣 Yandex Teaches Neural Networks to Speak Naturally
Researchers at Yandex have developed a new method for evaluating machine translation quality called RATE (Refined Assessment for Translation Evaluation).
It is designed to improve models that translate texts accurately, but sound unnatural.
Existing systems often ignore stylistic errors. For example, “sorry, my bad” may be translated into an overly formal “I apologize, it is my fault” instead of the more natural “sorry, my mistake.”
The new metric evaluates translations across three user-critical criteria: accuracy of meaning, naturalness of the language and consistency with the original style.
RATE makes it possible to precisely identify where modern models fail, enabling translations that are both more accurate and more natural. The method can be applied to different types of text: in news it focuses on factual accuracy, while in social media posts it detects excessive formality. The system not only flags errors, but also assesses their severity.
World news