Yandex has introduced a new version of its neural network-based photo translation technology, which now better understands texts from images in many languages. The technology will be useful, for example, for quickly translating menus in restaurants while traveling or for working with technical documentation in foreign languages. Thanks to the update, translations become more accurate, taking into account the context and style of the original text, as well as visually resembling the original design.
The YandexGPT family of models now understands wordplay and complex phrases, which helps to translate advertising slogans and headlines, avoiding literal errors. This has significantly improved both simple translations (for example, the composition of cosmetic products) and complex ones (scientific articles, instructions). The new technology is already available in the "Translator" and "Browser" applications, and will soon appear in the Smart Camera.
To improve the accuracy of translation, Yandex algorithms remove the original text from the image and replace it with the translated text, preserving the font, color, and other visual parameters so that the text is natural and easy to read. Improved visualization makes the translation even more convenient to read than the original, thanks to contrast and correct word recognition.
The developers trained the YandexGPT model using many pairs of original and translated texts, which improved the quality of translation and avoided typical errors. Thanks to the use of the distillation method, Yandex was able to reduce computational costs while maintaining high quality of request processing.
Read materials on the topic:
Проверка орфографии, анализ данных: «Яндекс» запустил нейросеть YandexGPT 3 Lite
Упростит работу с контентом: в «Яндекс Браузере» появился текстовый нейроредактор