Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Word Frequency Shapes Prediction-Error Updating in Natural Speech Comprehension
Poster Session C, Thursday, October 1, 10:45 am - 12:45 pm, Wangari Maathai
Jiajie Zou1,2, Peter Donhauser3, David Poeppel4; 1Zhejiang University, 2Max Planck Institute for Biological Cybernetics, 3Ernst Strüngmann Institute, Frankfurt, Germany, 4New York University
Language is a complex, dynamic sequence, yet the human brain processes it effortlessly. A prominent account of this capacity is predictive processing: the language system uses statistical regularities to anticipate upcoming words and encodes only the prediction error when the next word appears. Surprisal, -logP(w|context), is widely used to quantify this prediction error and has been proposed as a causal bottleneck underlying predictive representations. Consistent with this view, many studies report robust behavioral and neural responses to surprisal, often dissociable from other linguistic variables such as frequency. We hypothesize that extensive language exposure induces frequency-dependent efficiencies that shape prediction error updating. For words matched in surprisal, higher-frequency words may elicit reduced surprisal responses because they rely on lower-cost representations (efficient representation), and/or earlier surprisal responses because lexical access and associated error signals propagate more rapidly (efficient propagation). We test these hypotheses in natural speech processing using four large-scale neural datasets (~150,000 words) spanning ECoG and MEG recordings of English- and Mandarin-language narratives. Across datasets, we replicate a classical N400-like response to surprisal. Critically, lexical frequency modulates this response: low-frequency words show amplified surprisal response, and the lowest-frequency words additionally exhibit delayed peak latencies. This joint pattern supports the coexistence of efficient representation and efficient propagation mechanisms. Together, these findings indicate frequency-based efficiencies in prediction-error updating, suggesting that surprisal alone does not fully account for predictive processing. The convergent evidence across languages and neural modalities highlights the value of large, naturalistic datasets for testing mechanistic theories of language processing.
Topic Areas: Speech Perception, Meaning: Lexical Semantics