Abstract
<p>Word predictability is a central determinant of eye movements during reading, withmultiple theories aiming to explain how reading proficiency and one’s background as afirst (L1) vs. second (L2) language reader relate to individual differences inpredictability effects. However, findings from individual-differences studies and L1/L2comparisons have been mixed, with interpretation complicated by limited reliability,varied methods for estimating predictability, and difficulty separating L1/L2 status fromproficiency. Here, we leverage large-scale data from the Multilingual Eye-movementCorpus (N = 1201) to revisit these questions. We estimated readers’ predictabilityeffects using Bayesian hierarchical models, with word predictability quantified usingboth human cloze and language-model surprisal. We then examined the reliability ofresulting estimates and their relation to reading proficiency and L1/L2 status. Afterestablishing which measures are reliable, we show that stronger predictability effectscorrelate with lower proficiency indexed by slower reading rate, with strongercorrelations among L1 than L2 readers. We also find that once controlling forproficiency, L1 readers show larger predictability effects than L2 readers. Thesefindings suggest that predictability effects reflect multiple sources of variance, includingthe overlap between objective and subjective predictability, and one’s reliance oncontextual support given their knowledge of the writing system.</p>