Propaganda detection in textual modality on telegram by means of transformers

DOI: 10.31673/2412-9070.2025.061206

Abstract

The state propaganda machine of the Russian Federation, supported by multi-billion annual budgets, strategically employs the Telegram platform as the primary channel for creating and disse minating propagandistic content. This activity significantly affects the perception of the war in Ukraine, the international information environment, and within Russia itself. Since Telegram already has one billion active users, the platform plays a crucial role in shaping public narratives and information flows. Moreover, the modern development of large language models makes the dissemination of propaganda more automated and rapid. This paper explores deep transformer-based neural networks, their role, and effectiveness in detecting Russian propaganda, using an algorithmically created dataset of Telegram channels that reflects the digital footprint of Russian propaganda during the full-scale invasion period. An effective transformer-based model is proposed, post-trained, and evaluated on this dataset.
The results of the study demonstrated that certain transformer-based models outperform other multilingual models, including other transfromers. These models also show higher effectiveness compared to classical and ensemble baselines. Moreover, the developed approach to processing and constructing efficient Telegram datasets—through an asynchronous algorithm with well-developed systematic approach, the randomized sampling and adherence to data construction principles and configurations—has proven to make machine learning algorithms highly effective, achieving accuracy levels between 80% and 96%.
Future research should focus on expanding multimodal aspects (text combined with images and metadata), exploring cross-linguistic transformer models, and developing online monitoring systems for tracking propaganda campaigns in social networks.

Keywords: deep neural networks; transformers; computational propaganda; social media; Telegram.