Automated Detection of Writing Phases: An Unsupervised Learning Validation of the Hayes & Flower Cognitive Writing Model

The 1980 Hayes and Flower model established that writing is comprised of three distinct cognitive processes: Planning, Translating, and Revising. This model has been foundational to writing studies, yet has lacked large-scale computational validation. This paper presents a novel methodology to address this gap. Using 135,178 minute-by-minute observations from 4,618 sessions in the KLiCKe keystroke logging corpus, we first used the Elbow Method to empirically confirm that the optimal number of behavioral clusters is three. We then applied K-Means clustering to partition the data. The resulting three clusters mapped precisely onto the theoretical phases, and provide empirical percentages of time spent in each phase in argumentative writing: Planning (43.5% of writing time), Translating (46.3%), and Revising (10.2%). Post-clustering analysis confirmed that the clusters matched their expected behavioral signatures; for example, the Planning cluster was characterized by extensive pause time (94.7% of the interval), while the Translating cluster was dominated by word production. To our knowledge, this study represents the first large-scale, unsupervised machine learning validation of the Hayes and Flower model, and provides empirical support for a foundational idea in writing theory, and provides a scalable, nonintrusive method for future writing process analysis.