Interpretable Computational Approaches to Propaganda: Detecting Linguistic Divergence in Russian Online Discourse

 The Russo-Ukrainian War has intensified the need to understand disinformation and its societal impacts. This study investigates language variation in Russia’s propagandistic narratives about the war, focusing on how language is manipulated to serve ideological goals. We apply computational methods to detect linguistic shifts and propaganda strategies across diverse media types and Wikipedia versions.

Previous work showed that state and social media used different propaganda frames (Alyukov et al., 2025; Vestel & Degaetano-Ortlieb, 2025). While normalization (understating the war’s impact on everyday life in Russia) was prevalent in state media, which aims to demobilize the population, disinformation (portraying news from Ukraine and the West as fake), a mobilizational approach, was more characteristic of social media. Wikipedia can also enable knowledge manipulation, as in the Russian Wikipedia Fork (RWFork; Trokhymovych et al., 2025), created in June 2023 by revising the Russian Wikipedia (RW) to comply with Russia’s legislation (Cohen, 2023).

 

We use Kullback-Leibler Divergence (KLD; Kullback & Leibler, 1951) as our method, which provides an interpretable approach to quantifying divergence between two probability distributions, highlighting distinctive linguistic features (e.g., words). We apply KLD on two datasets: the Wartime Media Monitor corpus (Alyukov et al., 2023) to compare the state and social media, and a collection of edits in RWFork (Trokhymovych et al., 2025) to contrast RW with RWFork. Some examples of the most distinctive words for each media type and Wikipedia version can be viewed in Table 1.

 

 

Table 1. Examples of the most distinctive words across media types and Wikipedia versions (higher KLD values indicate higher distinctiveness).

Our media analysis reveals that Russia-occupied Ukrainian territories (such as DPR and LPR, which stand for the self-proclaimed “Donetsk People’s Republic” and “Luhansk People’s Republic”) and references to the sham referendums on joining Russia conducted in these territories (e.g., referendum and voting) are more distinctive for state media, reflecting a territorial control narrative consistent with a normalization frame. Social media shows a higher contribution of direct war-related terminology (war, soldier, to fight, etc.), indicative of a mobilizational approach. In contrast, euphemisms for war (such as special military operation) are more distinctive for state media, aligning with efforts to downplay the invasion. Additionally, words like propaganda and truth exhibit high KLD values on social media, pointing to the disinformation frame. Therefore, our study shows a clear distinction between mobilization on social media and demobilization on state media, confirming previous research (Alyukov et al., 2025; Vestel & Degaetano-Ortlieb, 2025).

The KLD study on Wikipedia shows that divergences between RW and RWFork resemble those between social and state media. In particular, direct war-related terminology, such as invasion and war, is distinctive for RW and substituted with vague expressions in RWFork. Similarly, words like occupation and annexation are replaced with euphemisms such as inclusion and entry. Moreover, the recognition of Ukraine’s statehood in RW is absent from RWFork, with words like independence and sovereignty being removed. Finally, RWFork clearly recognizes the so-called DPR and LPR, since their names are among the most distinctive words for this version.

References

Alyukov, M., Kunilovskaya, M., & Semenov, A. (2023). Wartime Media Monitor (WarMM-2022): A Study of Information Manipulation on Russian Social Media during the Russia-Ukraine War. In S. Degaetano-Ortlieb, A. Kazantseva, N. Reiter, & S. Szpakowicz (Eds.), Proceedings of the 7th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature (pp. 152–161). Association for Computational Linguistics. https://doi.org/10.18653/v1/2023.latechclfl-1.17

Alyukov, M., Kunilovskaya, M., & Semenov, A. (2025). Confuse and Normalise: Authoritarian Propaganda in a High-Choice Media Environment and Russia’s Invasion of Ukraine. In P. Goode (Ed.), Russian propaganda today: Challenges, effectiveness, and resistance (p. in print). University of Michigan press, University of Manchester Press.

Cohen, N. (2023, July 12). Russian Wikipedia’s Top Editor Leaves to Launch a Putin-Friendly Clone. Bloomberg.Com. https://www.bloomberg.com/news/articles/2023-07-12/russian-wikipedia-editor-leaves-to-launch-a-putin-friendly-clone

Kullback, S., & Leibler, R. A. (1951). On Information and Sufficiency. The Annals of Mathematical Statistics, 22(1), Article 1. https://doi.org/10.1214/aoms/1177729694

Trokhymovych, M., Kosovan, O., Forrester, N., Aragón, P., Saez-Trumper, D., & Baeza-Yates, R. (2025). Characterizing Knowledge Manipulation in a Russian Wikipedia Fork. Proceedings of the International AAAI Conference on Web and Social Media, 19, 1924–1936. https://doi.org/10.1609/icwsm.v19i1.35910

Vestel, A., & Degaetano-Ortlieb, S. (2025). From War to Special Military Operation: Interpretable Detection of Linguistic Propaganda Framing in Russian Media. Workshop Proceedings of the 19th International AAAI Conference on Web and Social Media, 2025, 50. https://doi.org/10.36190/2025.50