Dual-input deep learning–based approach for propaganda narratives detection in a low-resource language: a case study in Lithuanian

Propaganda narratives play a central role in disseminating disinformation and are tailored to each targeted country’s context and native language. This highlights the pressing need for automated systems capable of detecting and analyzing such narratives, particularly in low-resource languages, where natural language processing tools and data resources remain limited. In this study, we introduce the first supervised machine learning system for identifying pro-Kremlin propaganda narratives in Lithuanian, a low-resource language spoken in a country targeted by Kremlin disinformation. To our knowledge, this represents the first such approach not only for Lithuanian but also for other languages in Russia’s broader neighborhood. Our method employs a novel dual-input architecture that significantly outperforms single-input baselines across all analyzed narratives and even surpasses the performance of ChatGPT-5. Although our study focuses on Lithuanian, our findings and methodology are applicable to other low-resource languages, offering practical guidelines for extending propaganda-narrative detection globally.