Reproduce “Wordcraft: a Human-AI Collaborative Editor for Story Writing” on OpenResearch

About this paper

Andy Coenen, Luke Davis, Daphne Ippolito, Emily Reif, Ann Yuan

Google Research developed Wordcraft, a human-AI collaborative editor for story writing that leverages a dialog-oriented language model (Meena) and few-shot learning to assist with diverse tasks like continuation, infilling, and rewriting. The system demonstrated that conversational AI, when appropriately prompted, can provide flexible assistance across various stages of the creative process and offer more interpretable feedback than general-purpose language models.

Key insights

• Dialog-oriented language models, when combined with few-shot learning, offer a versatile and intuitive interface for human-AI collaborative writing and can provide interpretable feedback (e.g., admitting confusion).
• Few-shot learning is essential for enabling a single large language model to perform diverse and complex creative writing tasks, such as infilling, elaboration, and rewriting, without explicit retraining.
• A collaborative editor can effectively support multiple stages of the story writing process, from initial drafting to refining and stylistic modification.

Summary by alphaXiv.