Tracking Linguistic Shifts in Russian Propaganda: A Diachronic Analysis Using KLD

Language, being a powerful tool, can be manipulated by malicious actors to shape public opinion. In this study, we examine the diachronic evolution of Russian propaganda about the Russo-Ukrainian War. Since linguistic shifts are shaped by social, political, and cultural factors, analyzing linguistic change in propaganda helps advance media literacy and enhance disinformation detection. This is evident in the evolving nature of fake news (Adriani, 2019), with Russian propaganda serving as a key example used to justify the invasion and mobilize public support (Solopova et al., 2023).

We use Kullback-Leibler Divergence (KLD; Kullback & Leibler, 1951) as our method, which offers an interpretable way to identify distinctive linguistic features. In previous work, we have demonstrated that KLD is suitable for detecting divergences in propaganda strategies between state-controlled and social media covering the Russo-Ukrainian War (Vestel & Degaetano-Ortlieb, 2025). This study extends prior work by applying KLD to edits in a Russian Wikipedia Fork (RWFork; Trokhymovych et al., 2025), thus adding another text type and a diachronic dimension to the analysis.

RWFork, created in June 2023, is a copy of Russian Wikipedia revised to comply with Russia’s legislation (Cohen, 2023). We use Trokhymovych et al.'s (2025) dataset, containing edits from 1.9M page titles between May and September 2023; most of the changes involve knowledge manipulation linked to the 2022 Russian invasion of Ukraine. KLD is employed to identify contrasting language between Russian Wikipedia and RWFork. Specifically, KLD quantifies the divergence between two probability distributions by measuring how many more bits of information are needed to represent one using the other, highlighting distinctive features (e.g., words) contributing to linguistic differences.

We hypothesize that differences between Russian Wikipedia and RWFork resemble those between state and social media. In particular, direct terms like war, distinctive for social media, are often replaced by euphemisms in state media (Vestel & Degaetano-Ortlieb, 2025); similarly, Trokhymovych et al. (2025) have shown that the words war and invasion are removed from RWFork. Moreover, RWFork replaces names of occupied Ukrainian regions with Kremlin-aligned terms, which is also characteristic of state media (Trokhymovych et al., 2025; Vestel & Degaetano-Ortlieb, 2025). However, since Wikipedia is stylistically different from both media types, we also expect to gain insights into the specificities of Russian propaganda in this genre.