Narrative Prose Generation

Charles B. Callaway, James C. Lester

Abstract:

For computational systems to play a more central role in education and entertainment, they must be able to engage users in ways that currently can only be achieved by one-on-one interaction with another person. By dynamically creating compelling narratives that take into account the context and requests of users, computers can stimulate children to learn and enable adults to participate in new forms of interactive media. An important aspect of this interaction is narrative prose generation, the process of automatically generating narrative prose that closely mirrors the process that human authors use to create extended prose, such as that found in short stories and novels.
The objective of this research is to design, implement, and empirically evaluate a comprehensive computational model of narrative prose generation (NPG) that can create natural language stories for educational and entertainment environments. Given a fictional world, its characters, and a sequence of events, a narrative prose generator should in realtime create full-length short stories or fairy tales in rich natural language.
An effective NPG model should exhibit the following characteristics: (1) it must create narrative descriptions that provide scene descriptions, generate character dialogue, and relate events that take place in the fictional world; (2) it must provide computational mechanisms for creating narratives with multiple points of view, character dialogue, and scene descriptions; (3) it must support methods for generating stories whose linguistic and syntactic complexity is customizable; (4) it should operate effectively in a multilingual environment; (5) all of these elements must function in realtime in order to keep the reader engaged.
We have designed the AUTHOR architecture that describes the process by which a narrative request from a user is used to create an original, customized narrative experience that is comparable in quality to stories written by human authors. In addition, we have created STORY BOOK, a narrative prose generator based on the A UTHOR architecture that produces narratives in the Little Red Riding Hood domain. To write stories, STORYBOOK takes a story plan consisting of the actors, scenes, props and temporally ordered events as input from a narrative planner. It then uses a sentence planner to create sentential descriptions, revises them, and then employs a unification-based functional systemic grammar to create narratives in the language of choice. STORYBOOK offers features not found in other text generation systems, including prose that is multiple pages in length, and smoothly integrates a number of aspects of text generators that have previously been studied in isolation.
We have conducted a series of evaluations that demonstrates that our method of narrative prose generation is superior to previous methods and that STORYBOOK is capable of producing a wide variety of narrative styles depending on the initial specifications. Finally, we show that AUTHOR is a well-defined and modularized architecture through a series of ablation experiments that demonstrate the degradation of prose quality as its different architectural components are disabled.