How Shall a Machine Call a Thing?

Federico Torrielli, Amon Rapp, Luigi Di Caro

Abstract: This paper aims to investigate the feasibility of utilising Large Language Models (LLMs) and Latent Diffusion Models (LDMs) for automatically categorising word basicness and concreteness, i.e. two well-known aspects of language having significant relevance on tasks such as text simplification. To achieve this, we propose two distinct approaches: i) a generative Transformer-based LLM, and ii) a image+text multi-modal pipeline, referred to as stableKnowledge, which utilises a LDM to map terms to the image level. The evaluation results indicate that while the LLM approach is particularly well-suited for recognising word basicness, stableKnowledge outperforms the former when the task shifts to measuring concreteness.