| Computational narrative focuses on methods to algorithmically analyze, model, and generate narratives. Most current work in story generation, drama management or even literature analysis relies on manually authoring domain knowledge in some specific formal representation language, which is expensive to generate.;In this dissertation we explore how to automatically extract narrative information from unannotated natural language text, how to evaluate the extraction process, how to improve the extraction process, and how to use the extracted information in story generation applications. As our application domain, we use Vladimir Propp's narrative theory and the corresponding Russian and Slavic folktales as our corpus. Our hypothesis is that incorporating narrative-level domain knowledge (i.e., Proppian theory) to core natural language processing (NLP) and information extraction can improve the performance of tasks (such as coreference resolution), and the extracted narrative information. We devised a non-linear information extraction pipeline framework which we implemented in Voz, our narrative information extraction system. Finally, we studied how to map the output of Voz to an intermediate computational narrative model and use it as input for an existing story generation system, thus further connecting existing work in NLP and computational narrative. As far as we know, it is the first end-to-end computational narrative system that can automatically process a corpus of unannotated natural language stories, extract explicit domain knowledge from them, and use it to generate new stories. Our user study results show that specific error introduced during the information extraction process can be mitigated downstream and have virtually no effect on the perceived quality of the generated stories compared to generating stories using handcrafted domain knowledge. |
