Folktales, embedded: testing the ATU index with unsupervised clustering
Embedded 1,500+ folktales and clustered them to see whether the Aarne-Thompson-Uther tale types hold up. Presented as a poster at the American Folklore Society annual meeting.
Folklorists classify tales with the Aarne-Thompson-Uther (ATU) index: Cinderella is type 510A, the Frog Prince is 440, and so on. Those labels were assigned by hand over a century, which raises a question. Do the categories match how the stories actually resemble each other?
Method
I embedded more than 1,500 folktales with a text embedding model, ran unsupervised clustering on the embeddings, and compared the resulting clusters to the tales’ ATU labels. Where a tale’s nearest neighbors carry a different label than its own, that is a candidate for a missed or ambiguous classification.
Findings
Many tales sit between clusters and plausibly fit more than one type, which a single-label scheme cannot express. The analysis also surfaced tales whose assigned type looked like a poor fit next to their semantic neighbors.
I presented the work as a poster, A computational investigation of missed Aarne-Thompson-Uther labels using unsupervised semantic clustering, at the 2024 American Folklore Society annual meeting.
Illustrations: Arthur Rackham, public domain.
