The full-scale conflict between the Russian Federation and Ukraine generated an unprecedented amount of news articles and social media data reflecting opposing ideologies and narratives. These polarized campaigns have led to mutual accusations of misinformation and fake news, shaping an atmosphere of confusion and mistrust for readers worldwide. This study analyses how the media affected and mirrored public opinion during the first month of the war using news articles and Telegram news channels in Ukrainian, Russian, Romanian, French and English. We propose and compare two methods of multilingual automated pro-Kremlin propaganda identification, based on Transformers and linguistic features. We analyse the advantages and disadvantages of both methods, their adaptability to new genres and languages, and ethical considerations of their usage for content moderation. With this work, we aim to lay the foundation for further development of moderation tools tailored to the current conflict.
We implement a binary classification using the following models for input vectors consisting of 41 handcrafted linguistic features and 116 keywords (normalized by the length of the text in tokens): decision tree, linear regression, support vector machine (SVM) and neural networks, using stratified 5‑fold cross-validation (10% for test and 90% for training). For comparison with learned features, we extract embeddings using a multilingual BERT model  and train a linear model using them.We performed 3 sets of experiments contrasting the handcrafted and learned features: Experiment 1. Training models on Russian, Ukrainian, Romanian and English newspaper articles, and evaluating them on the test sets of these languages (1.1) and on French newspaper articles (1.2). We add the French newspapers to benchmark the multilingualism of our models. We choose French because it is in the same language family as Romanian.Experiment 2. Training models on Russian, Ukrainian, Romanian, English and French newspaper articles, and validating them on the test set (2.1). Additionally, we use this model to test the Russian and Ukrainian Telegram data (2.2.). Here the goal is to investigate whether this model will perform well out-of-the-box for the Telegram articles, which are 10 to 20 times shorter. See an example of the genre-related difference in distributions in Fig. 1.Experiment 3. Training models on the combined newspaper and Telegram data and applying them to the test set. Here we verify whether adding the Telegram data to the training set can improve generalization power, although data distributions differ.The full paper Download the paper