Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Using OPM-MEG and Contextual Surprisal to Investigate Naturalistic Language Processing: A Protocol for Psychosis Research
Poster Session F, Friday, October 2, 2:45 - 4:45 pm, Wangari Maathai
This poster is part of the Sandbox Series.
Tobias Schweiger1, Victoria R. Poulton1, Lars Meyer2, Peter Krueger3, Joachim Gross4, Peter J. Uhlhaas1; 1Charité Universitätsmedizin Berlin, 2Johannes Gutenberg-Universität Mainz, 3PTB Physikalisch-Technische Bundesanstalt Institute Berlin, 4Universität Münster
Language comprehension relies on the ability to generate predictions about upcoming input and to integrate incoming words with prior context. Disruptions in predictive processing have been observed in psychiatric disorders such as schizophrenia, where signals of semantic processing and contextual integration, such as the N400, are often reduced or delayed (Kumar & Debruille, 2004; Mohammad & DeLisi, 2013; Kiang & Gerritsen, 2019). However, it remains unclear how sentence-level and broader discourse context are dynamically integrated during naturalistic language comprehension in schizophrenia. In this study, we will use word surprisal values reflecting local (sentence-level) and global (discourse-level) contextual information to investigate how hierarchical (local vs global) predictive information is weighted during naturalistic spoken language processing. We aim to first establish our protocol in the general population before extending it to clinical cohorts. Participants (healthy controls) will listen to a continuous audiobook in German while undergoing whole-head Optically Pumped Magnetometer Magnetoencephalography (OPM-MEG), a novel MEG technology (Brookes et al., 2022). Using the large language model GPT, we will estimate two types of surprisal for each target noun: local surprisal based on preceding words within the same sentence; and global surprisal, defined by words within the five preceding sentences. These will serve as predictors of mean neural response amplitudes (ERFs) time-locked to target nouns. Nouns were selected as target words because they carry high semantic content and are strongly modulated by contextual predictions, making them well suited to capture both highly predictable and highly surprising cases. We hypothesize that both local and global surprisal will predict ERF amplitude, with higher surprisal associated with greater amplitude. Within a single-step model of language comprehension, local and global context are assumed to be immediately integrated, with their relative influence modulated over time (Nieuwland & Berkum, 2006). Therefore, neural responses time-locked to noun onset will be used to extract mean neural amplitudes from an early (200–300 ms) and a later N400-like time window (300–500 ms). Analyses will use linear mixed-effects regression models (Nieuwland et al., 2019; Poulton & Nieuwland, 2022; Eisenhauer et al., 2022). Separate models will be estimated for each time window to assess contextual effects. The first models will include both surprisal predictors as fixed effects and their interaction, and will then be compared to an additive model without interaction. We expect global context to exert a relatively stronger effect in the later window, reflecting a shift in contextual weighting. An interaction between local and global surprisal in the early time window would indicate that early predictive processing is jointly sensitive to both levels of surprisal, such that the effect of one depends on the other. An interaction in the late time window would indicate that local and global surprisal jointly shape later stages of processing, reflecting continued integration across hierarchical levels during contextual updating. At the conference, we aim to show preliminary results (N ≈ 15) for the Sandbox Series. This project will be at a critical stage for feedback prior to large-scale data collection with schizophrenia patients.
Topic Areas: Meaning: Discourse and Pragmatics, Computational Approaches