Identifying Motifs in Folktales using Topic Models

With the undertake of various folktale digitalization initiatives, the need for computational aids to explore these collections is increasing. In this paper we compare Labeled LDA (L-LDA) to a simple retrieval model on the task of identifying motifs in folktales. We show that both methods are well able to successfully discriminate between relevant and irrelevant motifs. L-LDA represents motifs as distributions over words. In a second experiment we compare the quality of these distributions to those of a simple baseline that ranks words using a TF·IDF weighting scheme. We show that both models produce representations that match relatively well to a manually constructed motif classification system used in folktale research. Finally we show that unlike L-LDA, this simple baseline is capable of representing abstract motifs as generalizations over more specific motifs.