Sber calls them models. The Rewriter model selects other words when rewriting the text.

But tries not to distort the meaning, and the Summator model.

Creates a condensed presentation of the text, preserving the main points.

The benefits of using

models are obvious. The first one creates unique texts, which will interest everyone who needs them.

from bloggers and sellers to small editorial offices engaged in rewriting. The second one ideally increases the chances that people will read the text.

to the end, be it a review of a large scientific work or a news digest for the week.

It is clear that before putting the models in open access for beta testing, cambodia phone number data the SberDevices developers trained them by forcing them to process a large number of materials of different lengths, styles and genres. In the development of artificial intelligence, working with natural language is one of the most important areas.

We at SberDevices are actively working on the development of NLP technologies and strive to create neural

networks that can take on routine work and help users save time and resources. Denis Filippov CEO SberDevices, Vice President for Digital Surfaces Salute Sberbank Testing “Rewriter” To evaluate this tool, we used the text uniqueness check service, and fed TexTerra Daily news,

which had already been

indexed by search engines, into the Sber model. Thus, we could judge the depth of the rewrite by the degree of uniqueness of the new text. We left the quality assessment to a professional editor. The downside of the beta-testing version of “Rewriter” is that the text fed into it needs to be freed from paragraphs, turning it into a single “canvas”.

Otherwise, the model returns an error and does not product owners see their role as creating a good product rewrite the text. But having returned an error, the model gives the original text with recognized paragraphs. All that remains is to copy it and feed it into the model again – and then the AI ​​no longer returns an error.

The big advantage of “Rewriter” is that it gives several versions of the rewritten text (pieces of which can be substituted into each other), and at the beginning it shows the “Best version” (“prediction_best”).

The uniqueness of the text selected by the neural network fans data  was 23.64%. As for the quality, as in the case of living authors, editing is needed. Here is a part of one of the texts rewritten by the neural network. More details on.