Modelling typing disfluencies as finite mixture process

Abstract

To writing anything on a keyboard at all requires us to know first what to type, then to activate motor programmes for finger movements, and execute these. An interruption in the information flow at any of these stages leads to disfluencies. To capture this combination of fluent typing and typing hesitations, researchers calculate different measures from keystroke-latency data—such as mean inter-keystroke interval and pause frequencies. There are two fundamental problems with this: first, summary statistics ignore important information in the data and frequently result in biased estimates; second, pauses and pause-related measures are defined using threshold values which are, in principle, arbitrary. We implemented a series of Bayesian models that aimed to address both issues while providing reliable estimates for individual typing speed and statistically inferred process disfluencies. We tested these models on a random sample of 250 copy-task recordings. Our results illustrate that we can model copy typing as a mixture process of fluent and disfluent key transitions. We conclude that mixture models (1) map onto the information cascade that generate keystrokes, and (2) provide a principled approach to detect disfluencies in keyboard typing.