Modeling L2 writing processes beyond pause thresholds: A window-level hidden Markov approach to keystroke-logging data

Abstract

Keystroke logging has made second language (L2) writing visible as a sequence of pauses, production, cursor movement, and editing, yet much evidence is still analyzed through fixed pause cut-offs or isolated process measures. This study examined whether a window-level hidden Markov model (HMM) could provide an interpretable local segmentation of Inputlog data without treating pauses or deletions as direct cognitive evidence. The corpus contained 100 computer-based argumentative writing sessions from undergraduate L2 writers. Each session was converted into 1-second windows using event-derived timing, text-growth, caret-movement, and Revision Matrix edit-action features, while linguistic boundary location was retained separately for subsequent boundary analyses. A theory-specified three-state solution estimated planning-like, formulation-like, and revision/editing-like behavioral states. Inferential models tested boundary-conditioned state distribution, production in post-onset windows, and edit episodes after transitions into revision/editing-like states. Planning-like occupancy and dwell increased before words and sentences, with the clearest dwell contrasts at sentence and paragraph boundaries. Formulation-like onsets following planning-like spells showed a small, specification-sensitive adjusted association with post-onset fluency and clean rate. Transitions into revision/editing-like states were often followed by Revision Matrix-confirmed edit episodes, although above-word edits were uncommon. The findings support a cautious state-based description of local L2 writing sequences.