Today was the Alice & Eve 2020 symposium at the University of Twente. I sent in an abstract for the poster session about last year’s paper about story generation, in which I discuss 60 projects from NaNoGenMo 2018.
The audience of Alice & Eve’s poster session was “people with a technical background”, so a relatively general audience. I tried to write the abstract in such a way that anyone can understand it. I’m pretty happy with how it turned out, so now you can read it here. :)
NaNoGenMo is an online programming challenge where participants try to create a novel generator, i.e. software that can autonomously compose a piece of fiction, in 30 days. At the end of the challenge, participants have to share their code and an output text of at least 50,000 words.
Although some generators trivially write the word ‘woof’ 50,000 times or modify one of the classics of world literature, a minority of participants creates a generator that outputs something resembling logical original prose. Given that coherent text generation is an open problem in the Natural Language Processing research field, this is no mean feat.
We analyzed the code and output of 60 projects of NaNoGenMo 2018 to see how participants generated a coherent 50,000 word story, hoping that this would provide the NLP community with useful design patterns. We found that participants used four different approaches for coherent text generation: hardcoding a high-level text structure in the generator, simulation and emergent narrative, framing the output in a clever way, and faking coherence.
Story generators can be used to write new works of fiction, but also to create readable, informative texts that might normally not be accessible for a general audience. Think of texts that describe complicated expert knowledge or quantative data, like medical information for patients in a hospital, governmental policy documents, scientific findings or financial reports of companies. Design patterns like those found in NaNoGenMo projects can help us build better story generators, so that we can automatically create coherent, clear and engaging narratives.
If this abstract has made you curious about my findings, or story generation in general, you can find the open-access paper here. And below is the accompanying poster that I showed during the poster sessions at ACL 2019 and Alice & Eve 2020:
Judith van Stegeren is a Dutch computer scientist. She is working as PhD candidate at the University of Twente, where she researches natural language generation for the video games industry. She occassionaly works as a consultant in data engineering for textual data.