"Yandex" has learned to more accurately respond to complex queries

The new algorithm "Yandex" converts the search request, and several sites on the semantic units or so-called semantic vectors. Once a user enters a query, the search engine compares the vectors with the title and the content of potentially suitable sites. Previous algorithm "Palekh" could only compare the semantic vectors of the request and headers, and "Korolev" scans web pages entirely.

"Yandex" explains the algorithm "Korolev" as an example of "War and Peace". Previous version can search only to find a book by title or keyword associated with the content: the names of the characters, chapter titles, and so on. A new algorithm reads like a novel and fully understands the meaning of the narrative.

According to the developers, the main advantage of understanding the meaning of queries that the search language becomes more human. The user can not search for a movie by name, but just briefly describe the plot. For example: "a film about the space in which the hero was talking to her daughter through the clock." The algorithm realizes that there is a film "Interstellar", where the hero is Matthew McConaughey was in space and used the clock to communicate with her daughter. And this film will be shown in the search results.

To search for understanding as the meaning of the query relates to the content of a web page, you need to use a large-scale neural network. For her training will need billions of examples, why "Yandex" users to use their training algorithms.

If a person is moved from the search results to the site and stayed on it, most likely, the algorithm worked properly. The neural network is trained on a bad example: if the site is immediately shut down or just missed, the system will remember. So that in the future the search should work better and better.