Toward automatic role identification in unannotated folk tales

Abstract

This paper presents an approach for automatically identifying high-level narrative structure information, particularly character roles, from unannotated folk tales. We introduce a new representation called action matrices to encode Propp's narrative theory on character role and their "sphere of action." We tested our approach in a fully automated system (Voz) using a corpus of 10 folk tales. Our experimental evaluation shows that action matrices capture useful information for role identification, provides insight into the error introduced by individual steps, and identifies the current bottlenecks.