INFORMATION TECHNOLOGY FOR IDENTIFYING PROPAGANDA IN TIKTOK COMMENTS BASED ON NLP AND DEEP LEARNING
Ukrainian Journal of Information Technology · 2026 · 8 (1) · 16-31
DOI: 10.23939/ujit2026.01.016 ↗Kurz gefasst
Fachartikel in der Zeitschrift Ukrainian Journal of Information Technology, 2026. Verfasst von V. A. Vysotska und 2 weiteren Autor:innen. Die Arbeit nennt 27 Literaturangaben.
Abstract (Originalsprache)
The article investigates the current scientific problem of automated detection of propaganda influence in short text comments of users of the TikTok social network, which operates in conditions of hybrid warfare and intensive disinformation campaigns. A hybrid model for detecting propaganda content has been developed, which integrates deep contextual representations of the text (transformer-based contextual representations) based on the BERT architecture with an additional vector of semiotic and structural features (number of emojis, repetition of symbols, use of caps lock). The final classification block of the hybrid model is implemented as a two-layer fully connected perceptron (MLP), which takes as input the concatenation of the [CLS]-token vector of the transformer model (768 dimensions) and the vector of additional semiotic features (10–20 dimensions). A methodological approach is proposed that combines classification with a teacher and elements of cluster analysis (k-means clustering) for effective work in conditions of limited marked-up data corpora (low-resource languages) for the Ukrainian and Russian languages. It has been found that the integration of additional lexical and graphic markers allows the model to better identify hidden ideological narratives, which often do not contain direct aggression. It was found that the implementation of the hybrid architecture provides a statistically significant improvement in quality metrics: the F1-score increased by 3–5% compared to the basic transformer models, and the area under the curve (ROC-AUC) value increased by 0.04–0.06. Specific differences in the functioning of the model for different language segments were identified: for the Ukrainian-speaking corpus, higher accuracy (Precision) was recorded, while for the Russian-speaking one, higher completeness of detection (Recall) of propaganda messages. The effectiveness of the developed system was studied on real control examples (use cases), which confirmed the model’s ability to detect up to 80% of propaganda content in thematic videos about Crimea and other socio-political issues. The information space monitoring tools were improved through the implementation of an application interface in the form of a Telegram bot, which provides operational analysis of TikTok comments in real time. Prospects for further research lie in expanding multimodal analysis for simultaneous processing of text, video, and audio tracks of a social network.
Zitieren
APA
Vysotska, V. A., Chyrun, L. V., & Romanchuk, R. V. (2026). INFORMATION TECHNOLOGY FOR IDENTIFYING PROPAGANDA IN TIKTOK COMMENTS BASED ON NLP AND DEEP LEARNING. Ukrainian Journal of Information Technology, 8(1), 16-31. https://doi.org/10.23939/ujit2026.01.016
Vancouver
Vysotska VA, Chyrun LV, Romanchuk RV. INFORMATION TECHNOLOGY FOR IDENTIFYING PROPAGANDA IN TIKTOK COMMENTS BASED ON NLP AND DEEP LEARNING. Ukrainian Journal of Information Technology. 2026;8(1):16-31. doi:10.23939/ujit2026.01.016
BibTeX
@article{Vysotska2026,
author = {Vysotska, V. A. and Chyrun, L. V. and Romanchuk, R. V.},
title = {INFORMATION TECHNOLOGY FOR IDENTIFYING PROPAGANDA IN TIKTOK COMMENTS BASED ON NLP AND DEEP LEARNING},
journal = {Ukrainian Journal of Information Technology},
year = {2026},
volume = {8},
number = {1},
pages = {16--31},
doi = {10.23939/ujit2026.01.016}
}
EUVIMED