Social network “Vkontakte” have trained the neural network to formulate headlines for news. The team applied research worked so well that to perform the task model in both Russian and English. The creators trained on texts of RIA “news” and The New York Times.
The process of creating a heading is as follows: the model handles the text and formulates a proposal from fragments of words. Cases and declination remain correct, it turns out a coherent sentence.
In order to evaluate the quality of work, the team conducted a small study. The respondents were shown the news and two headers to it: in 45% of cases the machine was better than the original, and 15% – even better.
However, the creators did not soar in the clouds, and a sober look at things. According to them, the headlines created by the neural network, while does not always turn out the same quality as written by a man.
“We continue to improve the model and will achieve equal results,” – said the designer-researcher “Vkontakte” Daniil Gavrilov.
Earlier it was reported that users of the social network has offered to shelter dogs instead of watching memes.