Semantic Update as a Predictor of Reading Time: Moving Beyond Word-Level Surprisal

This study explores the role of semantic prediction error in shaping reading behavior, as reflected in reading time during naturalistic reading. We operationalize semantic prediction error using semantic update (SU), a measure derived from the Sentence Gestalt (SG) model, which simulates incremental predictive comprehension at the level of sentence meaning. While SU has been previously linked to the N400 ERP component - an index of semantic processing - here we test its behavioral relevance by examining how it predicts total word-fixation durations in an eye-tracking corpus. To contextualize the role of semantic prediction error, we compare SU with lexical prediction error, quantified as next-word surprisal estimated from a language model (LM) trained on the same corpus as the SG model. Our results show that SU significantly contributes to explaining total word-fixation durations, even after accounting for LM-based surprisal and lower-level lexical factors. We also find that this predictive relationship holds primarily for content words, but not for function words, which parallels findings for lexical surprisal. These findings support the view that semantic prediction error plays a unique and measurable role in guiding real-time reading behavior, complementary to lexical prediction.