Thus, it became possible to ask the search engine complex questions in a conversational speech format and receive correct answers.
A significant advantage of Korolev was also the ability to apply it to a significantly larger number of pages without affecting the time it takes to return results for a russia company email list query. Palekh was a relatively heavy algorithm and was used exclusively at the later stages of ranking, to approximately the 150 best pages from a list filtered by the old rules.
About transformers
Palekh and Korolev allowed Yandex not just to find matches, but to understand the essence of the issue, significantly improved the ranking process, but still did not cope with it perfectly. Only since the introduction of YATI, the factors of meaning began to exceed the factors of occurrences by mLF phrases.
Before we begin to talk in more detail about YATI, it is necessary to separately explain what transformers are.
In simple terms, transformers in this case are called super-large and super-complex neural networks that can easily cope with a variety of tasks in the field of natural language processing, be it translation or text creation.
Behind this lies enormous computing power. And rapidly growing. Thus, before the use of transformers, the neural network used in Yandex was trained on only one Tesla v100 graphics accelerator. Such training took no more than one hour. But training a transformer neural network on such an accelerator would take about 10 years. Therefore, the introduction of new technologies required the use of about a hundred similar accelerators with fast data transfer between each other. For this, Yandex built a special cluster designed for computing, with distributed training inside it.