SemioMeme: A Symbolic–Subsymbolic Knowledge Graph Dataset for Multimodal Meme Analysis

Abstract

Internet memes present a challenge for computational analysis as their meaning derives from cultural context external to their observable features; however, visually similar memes carry distinct cultural meanings, whilst semantically related memes may share no perceptual similarity resulting in a decoupling of format and meaning. To support analysis requiring both, we present SemioMeme, a knowledge graph providing symbolic representations of meme concepts with their cultural connections alongside subsymbolic vision and text embeddings connected via a dedicated property between both. This supports hybrid queries that can surface cultural associations accruing through graph proximity, often invisible to similarity search or explicit labelling alone. The resource, including source data and code, is made openly available and covers 16,707 meme concepts, 507K meme instances with multimodal embeddings, and 7.2M RDF triples.