Driving story generation with learnable character models

Abstract

Stories are a ubiquitous part of the human experience and an essential component of human intelligence. Stories serve countless functions, working in tandem with our language facilities to enable communication, problem solving, imagination, and more. Stories are eminently cross disciplinary, fueling how people think about everything from chemical reactions to novels, social encounters to politics. If we are to develop a comprehensive computational system mirroring human intelligence, the system must be able both to understand stories and to go beyond that, using stories to empower how it thinks about the world. On a high level, a key component of story understanding is story generation. In human thought, generation enables us to imagine, solve problems, and make decisions. Therefore, to take computational story understanding to the next level, this research makes a number of contributions centered on story generation. In particular, this thesis describes a novel story generation system capable of creating new and interesting stories by learning from a corpus of human-written stories. The design of robust learnable character models facilitated this generation system. The character models are capable of handling characters that span a number of genres including romantic comedies, fairy tales, warfare, and Shakespearean tragedies. The work also includes methods for the automatic learning of characters traits in both an unsupervised way using topic modeling and a semi-supervised way using a new method of alignment-based trait learning. Through my research these algorithms were tested using the Genesis system for story understanding. The results from the research illustrate how the robust methods are able to handle a variety of situations, including internally conflicted characters, characterizations that span genres, and conflicting character goals. Critically, this research demonstrates how story generation is important to story understanding.