Aksi Marsovich
Intergalactic expert
At the YaC 2020 conference, Yandex announced that a new ranking algorithm, YATI (Yet Another Transformer with Improvements), has appeared . This name can be translated as follows: "Yet Another Transformer with Improvements".
The new neural network architecture for ranking is 100% active phone number list the most significant change to appear in search in the last 10 years. The search engine has become smarter. How does it work? We will tell you about it.
The evolution of search
YATI analyzes the texts of search queries and sites displayed by these queries. YATI's work is more perfect than that of its predecessors (Palekh and Korolev).
Palekh. It was created to compare search query texts and page titles. It was trained on positive and negative examples. It received them from accumulated statistics.
Search engine algorithms cannot read texts, so they identify the correspondence of queries to headings by comparing numbers. To make searching for semantic correspondences easier, page headings were transformed into groups, each of which had three hundred numbers. Thus, each document scanned by Palekh received its place in a hypothetical space of 300 dimensions.
This method of processing was called a semantic vector. Its advantage is that the algorithm learned to establish a match between a request and a page even when they did not have a single word in common.
Palekh is a heavy algorithm. It was usually used at the very last stages of ranking, when the 150 best pages among the total number had already been filtered.
Korolev. He worked on the same principle as Palekh, transforming web page texts into semantic vectors. Korolev began to perform these heavy operations not in real time, but in advance, at the indexing stage. A person wrote a query, and the algorithm compared the query vector with the page vectors already known to it. Such preliminary calculations greatly reduced the time it took to issue results.
In addition, Korolev began to compare the vectors of new queries with those queries for which the best answer had already been generated. If the vectors were close, then the results were similar.
