Year: 2026
Type: conference-paper
Abstract: Standard surprisal is typically computed from the linear text prefix, but human reading is nonlinear and memory-constrained: readers skip words, regress, and do not retain prior context perfectly.We propose a formulation of surprisal conditioned on a reader-specific accessible information state given by the scanpath history and memory dynamics, rather than by the written prefix alone.Prior context is treated as only probabilistically accessible at each fixation, allowing predictability to depend on both non-linear exposure and forgetting.We evaluate the approach on eye-tracking corpora using held-out log-likelihood over standard durationbased reading measures.Across model variants, conditioning on accessible information states improves predictive fit over standard surprisal baselines.These results suggest that predictability in human reading is better characterized relative to the reader's evolving accessible information state than to the written prefix alone.
Authors: Michael Mooney, Edmond S. L. Ho
Institution: University of Glasgow
Language: English
Subfield: Developmental and Educational Psychology
Field: Psychology
Domain: Social Sciences
Open Access status: hybrid
OpenAlex ID: W7166879898