Casting a Spell: Identification and Ranking of Actors in Folktales

We present a system to extract ranked lists of actors from fairytales ordered by importance. This task requires more than a straightforward application of generic methods such as Named Entity Recognition. We show that by focusing on two specific linguistic constructions that reflect the intentionality of a subject, direct and indirect speech, we obtain a high-precision method to extract the cast of a story. The system we propose contains a new method based on the dispersion of terms to rank the different cast members on a scale of importance to the story. 1 Introducing the Problem Today do I bake, tomorrow I brew, The day after that the queen’s child comes in; And oh! I am glad that nobody knew That the name I am called is Rumpelstiltskin! So says the song of the little man in Brother Grimm’s Rumpelstiltskin with which he accidentally reveals his name to the queen. Advances in the task of Named Entity Recognition (NER) make it possible for computer systems to recognize his name as well. NER seeks to locate and extract atomic textual units into a predefined set of categories (e.g. PERSON, LOCATION and ORGANIZATION). Now that we can recognize Rumpelstiltskin as being the name of a person, how do we know that he is a character or actor in the story? Are all named entities within the category of PERSON in a story automatically actors? How about characters that have no proper name in a story, but are referred to by nominals or noun phrases? How do we know that they are part of the cast of a story? Most NER systems are developed for the domain of non-fiction, including newspapers, manuals and so forth. Applications of NER in literary texts or folktales, however, have been discussed only marginally (see [13] for an exception).