From speaker identification to affective analysis: A multi-step system for analyzing children's stories

Abstract

We propose a multi-step system for the analysis of children's stories that is intended to be part of a larger text-to-speechbased storytelling system. A hybrid approach is adopted, where pattern-based and statistical methods are used along with utilization of external knowledge sources. This system performs the following story analysis tasks: Identification of characters in each story; attribution of quotes to specific story characters; identification of character age, gender and other salient personality attributes; and finally, affective analysis of the quoted material. The different types of analyses were evaluated using several datasets. For the quote attribution, as well as for the gender and age estimation, substantial improvement over baseline was realized, whereas results for personality attribute estimation and valence estimation are more modest.