System for News Summarizing Using a Neural Network Algorithm

This paper presents implementation of a system for summarizing news using a neural network algorithm. Relevance of the development of this system is due to the need for users to receive news in a concise form, while maintaining objectivity and accuracy of information transfer. Multilingual mBART Large CC2S model was chosen as the initial model. The main source of data for further training of the model was the gazeta dataset, taken from the Hugging Face project, supplemented by a collected data set from various news sources. To determine the qualities of the trained model, ROUGE and BLEU metrics were used. When developing the summarization algorithm, Python programming language with deep learning libraries PyTorch and Fairseq were used. The system works through the telegram bot interface. Telegram bot provides the ability to generate an annotation for the required news and produce the result in form of a message. The result of this work is a system that allows to receive a news summary in form of a short message.