"Yandex Translator" was trained using its own YandexGPT neural network. The service has become better at navigating professional vocabulary, recognizing idioms, and understanding context. This was reported by the press service of the "Yandex" company.
A model from the YandexGPT family was used to generate reference training examples. The service's system training was divided into several stages.
The model analyzed many texts in English and Russian during the pre-training stage. Next came the turn of supervised finetuning (SFT) of the language model for translation tasks. SFT is a method used in machine learning to improve the performance of a pre-trained model. First, the model is trained on a large dataset, and then it is fine-tuned on a smaller, specific dataset. This allows the model to retain the general knowledge gained from the large dataset while adapting to the specific characteristics of the smaller dataset.
In conclusion, experts assessed the quality of YandexGPT's responses, sorting them from best to worst.
The Side by Side method (1:1 comparison of two objects) was used to evaluate the work. Pairs of long texts in Russian and English were translated using old and new technologies. The new one coped with the task 57% better.
Now, the service developers promise that it will be better at translating texts in narrowly focused topics and identifying relationships both within and between sentences.
It is worth recalling that "Yandex Translator" has been operating since 2018, translating texts from 100 languages. The service can also translate inscriptions from pictures and photographs.
Earlier, www1.ru reported that "Yandex" launched the YandexGPT 3 Lite neural network.