Analyzing algorithms and solutions for automatic generation of news article leads in social networks by using artificial intelligence

Abstract

The article highlights the approaches to automatic abstracting the articles. When publishing articles on social networks, the editors of news portals need to create a short abstract of each article spending a minimum of time. Prompt and simultaneous placement of the publications on all registered resources is facilitated by automatic generation of leads. There is proposed to apply the artificial intelligence algorithms trained on corpora of the Russian texts. There are three approaches to text abstracting for the automated formation of article leads: extractive, abstract, and combined. There is carried out comparative analysis of the methods of extractive and abstract approaches in the frames of solving the problem by using neural network models of machine learning. Different stages of extractive abstracting are analyzed using both simple and more complex methods of LexRank, TextRank and on top of Deep Learning. The compared abstract models were selected as the most suitable ones for abstracting the news articles on top of the BERT model. More complex generating texts process the data in parallel, which speeds up processing, but requires training on large corpora of news documents. When using the abstract models Pointer General Network and MBART the information processing time is reduced and work efficiency increases.