“How Novelists Use Generative Language Models: An Exploratory User Study”, Alex Calderwood, Vivian Qiu, Katy Ilonka Gero, Lydia B. Chilton2020 (, , ; backlinks)⁠:

Generative language models are garnering interest as creative tools. We present a user study to explore how fiction writers use generative language models during their writing process.

We had 4 professional novelists complete various writing tasks while having access to a generative language model that either finishes their sentence or generates the next paragraph of text [‘Talk To Transformer’ vs ‘Write With Transformer’ GPT-2 services].

We report the primary ways that novelists interact with these models, including: to generate ideas for describing scenes and characters, to create antagonistic suggestions that force them to hone their descriptive language, and as a constraint tool for challenging their writing practice.

We identify 6 criteria for evaluating creative writing assistants, and propose design guidelines for future co-writing tools…We propose a series of evaluation questions, which could be answered computationally, to guide system design:

  1. Does a suggestion match the tense of the preceding text?

  2. Does a suggestion introduce new characters or objects, or does it reference preceding ones?

  3. Are new characters or objects coherent given the context?

  4. Does a suggestion include description?

  5. Does a suggestion include action?

  6. Given a single request, how diverse are the suggestions?

[Keywords: co-creativity, natural language processing, user interface, writing tools, user-study]