Encoding Imagism? Measuring Literary Imageability, Visuality and Concreteness via Multimodal Word Embeddings

This paper addresses the challenge of measuring "imageability" in literary texts — a concept from psycholinguistics that describes how words evoke sensory experiences. Imageability is connected to literature's immersive quality, but current methods face limitations due to vague definitions, poor lexical coverage, and difficulty scaling to sentence-level (/literary) analysis. To tackle this, we propose a data-driven approach using a multimodal model, alongside traditional word embeddings, to quantify imageability more effectively. Through three experiments: 1) a word-level analysis, 2) sentence-level comparison, and 3) a case study contrasting imagist and love poems — we test whether embeddings created with a multimodal model capture imageability, and the related features concreteness and visuality. We assess the extent to which multimodal embeddings capture imageability in literary texts, while considering compositionality and the multimodal nature of literary imagery.