“…More specifically, these authors train models such as logistic regression or conditional random fields on a corpus of human eye-tracking data, and then predict fixation time, skipping rate, and other eye-movement measures on an unseen test set. Unlike our goals here, the literature in this tradition (e.g., Bestgen, 2021;Hara, Kano, & Aizawa, 2012;Hollenstein et al, 2021;Matthies & Søgaard, 2013;Nilsson & Nivre, 2009 does not primarily aim to construct explanatory cognitive models, and the use of supervised training (i.e., models learn their behavior from a pre-existing training set of eye-movement data) is not psychologically realistic, as outlined in the introduction. However, machine learning-based models typically achieve good prediction accuracy, which makes them suitable for comparison with cognitive models.…”