Topic and Stance Detection in Memes

This thesis tries to test whether the topics and stances present in political memes can be related to the persuasion techniques they use to propagandize their message. The data is based on the SemEval-2024 Task 4 corpus though augmented in dierent ways, and the thesis presents a persuasion technique detection backbone, a multimodal topic pipeline reduced to a stance-ready taxonomy, and a stance detection system, then measures topic/stance × technique associations and tests each signal inside the classier. The strongest topic-technique associations are found in expected trivial pairs: Reductio ad Hitlerum recurs in Nazi and fascist topics, Flag-Waving recurs in patriotic and military topics, and Appeal to Emotions in war topics. Some stance targets of the dataset are universally favourable, and some universally opposed. Stance splits technique use along a clean praise-attack axis with favourable memes having Flag-Waving and Glittering Generalities while opposing memes having Smears. Yet, these correlations are narrow, and injecting topic or stance labels does not improve detection for the tested framework, since the meme's text already carries what the labels tell the model.