The shape of stories. Or how AI sees a story
Can we plot a story by its embeddings
When I saw Kurt Vonnegut’s video on shape of stories, I was fascinated.
Then I started asking myself. How can we create a visualization of a story. Now, because of LLMs we have the tools at hand.
We can take the embeddings of the sentences, and then string them together and create a visualization.
An embedding is a representation of a sentence in the latent space. So when we go from one sentence to the next, we are in effect doing a space walk in the latent space.
But obviously we cannot visualize this walk in the 784 or more dimension space of the embeddings. We humans are poor at anything above 3 dimensions :)
So enter, UMAP. Use UMAP and project onto 2d space to get the final shape of a story.
It really is magical to see these stories take a walk in the latent space. I love how Redi Riding hood looks a little like a wolf and Rapunzel clearly takes a big jump from one of the space to another :)
Hi Nutanc,
I love the creative approach you’ve taken to visualize the “shape” of stories using embeddings and UMAP! It’s a fascinating way to blend narrative and AI, and your references to Kurt Vonnegut’s theory really highlight the potential of this approach.
I’d be curious to know—how did you select UMAP over other dimensionality reduction techniques like t-SNE? And do you see this visualization method being used for more complex, longer stories, or perhaps even other narrative structures like movies or episodic series?
Looking forward to seeing more explorations like this!
Best,
Jenni
Amazing.
Wow