Inferring Propp’s Functions from Semantically Annotated Text

Vladimir Propp’s Morphology of the Folktale is a seminal work in folkloristics and a compelling subject of computational study. I demonstrate a technique for learning Propp’s functions from semantically annotated text. Fifteen folktales from Propp’s corpus were annotated for semantic roles, co-reference, temporal structure, event sentiment, and dramatis personae. I derived a set of merge rules from descriptions given by Propp. These rules, when coupled with a modified version of the model merging learning framework, reproduce Propp’s functions well. Three important function groups—namely A/a (villainy/lack), H/I (struggle and victory), and W (reward)—are identified with high accuracies. This is the first demonstration of a computational system learning a real theory of narrative structure.