Memes play a central role in modern online communication, especially in politically charged contexts where visual and textual cues jointly convey persuasive messages. This work focuses on Ukrainian wartime memes and leverages vision‑language models (VLMs) to detect the underlying persuasion techniques. We introduce the UkrMeme dataset, containing 200 organically created Ukrainian war memes annotated with specific persuasion tactics. To assess cross‑lingual transfer, we also release PTM‑UA, a Ukrainian translation of a subset of the Propaganda 950 benchmark that preserves the original visual content. Experiments adopt a zero‑shot prompting setup, evaluating VLM performance on image‑only, text‑only, and multimodal inputs using hierarchical F1 scores. Multimodal inputs consistently improve recall of persuasion techniques. We further compare two explanation methods: Layer‑wise Relevance Propagation (LRP) and model‑generated token attributions. Semantic analysis of the explanations reveals that LRP tends to highlight visual evidence (e.g., colors, composition), whereas token attributions focus on textual cues (e.g., keywords, sentiment words). These insights inform the design of more transparent and trustworthy persuasion detection systems.
Review: The UkrMeme dataset fills a gap in Ukrainian‑language meme research, and the experimental framework balances zero‑shot capability with interpretability analysis, showcasing VLMs' potential for cross‑modal, cross‑lingual persuasion detection. Future work could enlarge the dataset and explore finer‑grained persuasion strategy classifications.