Leveraging VLLMs for Visual Clustering: Image-to-text mapping shows increased semantic capabilities and interpretability

Automated image categorization is vital for computational social science, particularly considering the rise of visual content on social media, as it helps the identification of emerging visual narratives in online debates. However, the methods currently used in the field to represent images numerically are unable to fully capture their connotative meaning and do not produce interpretable clusters. In response to these challenges, we evaluate an approach based on the automated generation of intermediate textual descriptions of the input images with respect to the connotative semantic validity of the generated clusters and their interpretability. We show that both aspects are improved over the currently typical clustering approach based on convolutional neural networks.