PoliMeme: Exploring Offensive Meme Propagation in the Israel-Palestine Conflict

Abstract

Social media platforms have become powerful tools for communication, often harnessing memes to convey ideas, opinions, and at times misinformation. In the midst of global polarized issues, such as the Israel-Palestine conflict, memes can rapidly spread offensive content and speech, influencing public perception. We introduce PoliMeme, a novel multimodal dataset encompassing text-embedded memes relevant to the Israel-Palestine conflict, meticulously annotated for sentiment analysis (positive, negative, neutral) and offensive content detection, further subdivided into direction of offense (directed vs. undirected) and targeted entities (individuals, organizations, communities). We benchmark several state-of-the-art unimodal and multimodal models to evaluate their performance in classifying these memes. Our findings underscore the necessity of multimodal methods to fully capture the subtle interplay between textual cues and visual context. We further highlight challenges unique to politically charged and culturally sensitive content, emphasizing the importance of precise annotation and robust processing strategies. With the PoliMeme dataset and accompanying analysis, we present a timely contribution to advance the understanding of how offensive speech and sentiment are propagated through textembedded memes, while also contributing to safer digital spaces by informing content moderation practices. The dataset is available at https://github.com/PrathamLearnsToCode/PoliMeme