Teaching Tale Types to a Computer: A First Experiment with the Annotated Folktales Collection

Computational motif detection in folk narratives is an unresolved problem, partly because motifs are formally fluid, and because test collections to teach machine learning algorithms are not generally available or big enough to yield robust predictions for expert confirmation. As a result, standard tale typology based on texts as motif strings renders its computational reproduction an automatic classification exercise. In this brief communication, to report work in progress we use the Support Vector Machine algorithm on the ten best populated classes of the Annotated Folktales test collection, to predict text membership in their internationally accepted categories. The classification result was evaluated using recall, precision, and F-1 scores. The F-1 score was in the range 0.8-1.0 for all the selected tale types except for type 275 (The Race between Two Animals), which, although its recall rate was 1.0, suffered from a low precision.

Automated summary

Computational motif detection in folk narratives is a challenging task due to the fluid nature of motifs and the lack of adequate training data. This study uses the Support Vector Machine algorithm on a test collection of annotated folktales to predict text membership in different categories. The results show high F-1 scores for most tale types, except for type 275, which has a low precision rate despite a perfect recall rate.