Why coffee and tea mean more than their definitions

Coffee and tea have similar properties but evoke completely different atmospheres. This paper proposes Scene Abstraction, which represents words not just by what they are but by the situations they inhabit: the events, entities, settings, emotions, and atmospheres surrounding them. Using few-shot LLM prompting, they built COCA-Scenes (520 labeled examples) and showed their structured scenes align with human interpretation 11.8 percentage points better than standard text embeddings and outperform existing semantic databases.