Narrative constraint systems: how the networked computer can tell stories
Kristian J. Hammond, Laurence Birnbaum, Patrick Summerhays McNally · 2013
The field of automated content generation attempts to build software systems that generate varied but coherent content that is compelling for an audience, a very hard problem. Many systems have employed template structures as a strategy to achieve coherent output. Some of the problems commonly faced by these systems are repetitive output and a limited range of topics and language. These problems stem from overly restrictive templates and access to data that is static. This dissertation describes a series of systems that use template structures and syntactic selection mechanisms on large-scale dynamically changing information from the internet to overcome these limitations. Many examples of output from these systems are presented and discussed. A successful narrative structure must maintain coherence and yet not be entirely predictable, i.e., it must provide for just the right amount of variation in output. Not enough variation, and the result is not much different from a static narrative. Too much, and the reader or viewer will not be able to draw meaningful relationships among the content items. The purpose of a constraint in a content generation system is to act as a proxy for some relationship needed by a larger template structure. If the constraint is too loose, the needed relationship may not be reflected in the retrieved content due to noise; the coherence of the relationship is lost. However, if the constraint is too specific, the desired relationship may not be represented fully; the variance and richness of the relationship is lost. Despite the simplicity of the retrieval mechanisms employed, these systems can appear to leverage knowledge that they do not explicitly model but that is implicit in the content that is retrieved and synthesized into output, like slang vocabulary and topical events.