I liked the analogy I picked up somewhere (I don't remember where, maybe I synthesized it from some of Eliezer's writings), that words are handles, pointers to associations[0] - both factual and emotional ones - and poetry, specifically, is an art of using words to pull those emotional associations in a very calculated way, effectively executing code in your readers' brain to achieve the desired effect.

--

[0] - Today we'd say "tokens stand for their embeddings in the latent space". Funny how those old articles about pointers to areas in concept-space, fuzzy boundaries, associations, etc. suddenly map pretty much perfectly, 1:1, to how LLMs and other generative transformer models work.