Genre Ontology Learning: Comparing Curated with Crowd-Sourced Ontologies

The Semantic Web has made it possible to automatically find meaningful connections between musical pieces which can be used to infer their degree of similarity. Similarity in turn, can be used by recommender systems driving music discovery or playlist generation. One useful facet of knowledge for this purpose are fine-grained genres and their inter-relationships. In this paper we present a method for learning genre ontologies from crowd-sourced genre labels, exploiting genre co-occurrence rates. Using both lexical and conceptual similarity measures, we show that the quality of such learned ontologies is comparable with manually created ones. In the process, we document properties of current reference genre ontologies, in particular a high degree of disconnectivity. Further, motivated by shortcomings of the established taxonomic precision measure, we define a novel measure for highly disconnected ontologies.