Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Surprisal outperforms psycholinguistic norms as a predictor of reading times across the adult lifespan
Poster Session E, Friday, October 2, 11:00 am - 1:00 pm, Wangari Maathai
Sandra Martin1,2, Christopher Conrad3, Merle Schuckart4, Jonas Obleser4, Gesa Hartwigsen1,2; 1Max Planck Institute for Human Cognitive and Brain Sciences, 2Leipzig University, 3University of Vienna, 4University of Lübeck
Recognizing a word during reading incurs processing costs that reflect both stored lexical knowledge and the unfolding sentence context, captured in reading times. At the lexical level, properties such as familiarity, concreteness, and age of acquisition have long been shown to modulate processing effort (Rayner, 1998; Brysbaert et al., 2018), while information-theoretic measures derived from large language models (LLM) have gained recent prominence as predictors of reading times (Shain et al., 2024). Lexical surprisal captures integration difficulty, while contextual entropy reflects the uncertainty a reader holds about the next word. Both surprisal and entropy have been linked to increases in reading time (Schuckart, Martin, et al., 2026; Clark et al., 2025), yet their relative contributions compared to classic psycholinguistic norms, and how these contributions shift across the adult lifespan, have not been directly compared. The present study addresses this gap by systematically comparing the unique variance accounted for by LLM-generated psycholinguistic ratings, surprisal, entropy, and frequency in a lifespan sample of native German readers. 175 adults (M = 44.9, age range 18–85 years) participated in a self-paced reading task in which texts were presented one word at a time and comprehension questions followed each text (Schuckart, Martin, et al., 2026). Psycholinguistic ratings for familiarity, concreteness, valence, arousal, and age of acquisition were generated using GPT-4o-mini (OpenAI) following the procedure of Conde et al. (2026). Context-dependent surprisal and entropy were computed based on the two preceding words using GPT-2. Zipf frequency for each word was computed using the wordfreq library. Linear mixed-effects models were fitted to log-transformed reading times with a baseline model including age, recording location, word length, and block and trial number as fixed effects, and text and participant ID as random intercepts. Psycholinguistic predictors were added to the baseline model alongside surprisal, entropy, and frequency. Hierarchical variance partitioning was then applied to derive the unique contribution of each predictor (individual marginal R²), accounting for collinearity among predictors. Among the five GPT-generated psycholinguistic ratings, familiarity and arousal accounted for the greatest unique variance in word reading times. When all predictors, including surprisal, entropy, and frequency, were considered jointly, context-dependent surprisal explained the greatest unique variance in reading time, outperforming all psycholinguistic ratings as well as entropy and frequency. Interactions with age revealed that the effects of most psycholinguistic ratings on reading time increase with age. The strongest age-related interaction was found for surprisal, followed by familiarity and frequency. These findings highlight the central role of context-dependent surprisal in word-by-word reading, consistent with predictive processing accounts of language comprehension. Our results demonstrate that surprisal outperforms traditional psycholinguistic measures such as frequency and familiarity as a predictor of reading-associated processing costs. The observed strengthening of surprisal and psycholinguistic effects with age suggests that older readers increasingly rely on contextual predictability during sentence processing, potentially reflecting a greater dependence on top-down predictions. Our results also prove LLM-generated psycholinguistic norms sufficiently valid to detect processing effects, supporting their utility as a scalable tool for future language research.
Topic Areas: Reading, Methods