Generating Reports, Fiction, and Text That Sounds Good

Abstract

Natural language generation (NLG) came into its own in the 1970s, using AI based on symbolic methods. Automated reporting is one thread of NLG work; another is research in storytelling and fiction generation. The former sort of systems “textualize” underlying data; the latter generate plots and/or produce a narrative discourse based on plots. I argue that these two types of research are very relevant to each other. Writing fiction often involves imagining an underlying textual actual world in which characters undertake actions, then narrating what happens in this world, just as automated reporters narrate based on real-world data. In a literary sense, today’s large language models (LLMs) are very different from both automated reporters and storytelling systems. They let language play out from probability distributions over sequences of words. To understand non-LLM approaches, I survey both automated reporters (which have presented news about weather, seismology, sports, finance, and elections) and storytelling systems (which have narrated invented events, giving us insight into narrative and cognition). These two sorts of systems, and LLMs, have important differences, but can also inform each other.