Abstract by George Spendlove
Inferring the Linguistic and Semantic Formats of Hierarchically Structured Creative Artifacts
High-quality human-created artifacts are distinguished by cohesion and semantic richness that can be difficult for computational systems to emulate effectively. Certain classes of artifacts feature relationships between their constituent elements that naturally form a hierarchical structure which underpins the artifact's meaning. A tractable method for programmatically extracting such artifacts' structures would allow a computational system to better model those meaningful relationships. We formalize such a method, framed in hierarchical Bayesian program learning (HBPL), by modeling an artifact's structure as a factorization of a probability distribution over possible words conditioned according to the structure's hierarchy. We present HIEROS, a creative system that extracts linguistic and semantic formats from human-written six-word stories and uses them to create novel stories. We describe the system in detail, present and evaluate stories it has written, and discuss how our format model aids the system's creativity.