DIGITAL REPRESENTATIONS OF TELEGRAM CHANNELS

Abstract

The subject of research of this article is digital representations of textual information resources on the example of Telegram channels. The purpose of the work is to determine the optimal method of forming digital representations of Telegram channels for further thematic classification. The following tasks are solved in the article:definition of approaches to the formation of the input vector; determination of the stages of text data processing for the digital representation of the Telegram channel; creation of a dataset of digital representations of Telegram channels; dataset marking for solving the classification problem; determination of hyperparameters of optimal classification models. The following results were obtained: a dataset of digital representations of Telegram channels formed on the basis of the SBERT network was created using three approaches: aggregation of publication vectors, concatenation of keywords using the TF-IDF method, and a combination of the first two approaches; it was determined that the approach of concatenation of keywords using the TF-IDF method and the combination of the first two approaches to the formation of digital representations of Telegram channels based on text publications is the most effective for further classification by topic; the optimal hyperparameters of the thematic classification models are determined in terms of accuracy: Logistic Regressio and deep learning neural networks. A promising direction of further research is the evaluation of the application of the proposed digital representations to clustering and search tasks.