Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Sensorimotor Dimensions of Word Meaning During Continuous Speech: A Word-level sEEG Encoding Study
Poster Session D, Thursday, October 1, 4:30 - 6:30 pm, Wangari Maathai
This poster is part of the Sandbox Series.
Sarah Saneei1, Pierre Mégevand2,3,4, Laurent Spinelli4, Valentina Borghesani1, Nizar Michaud3,5, Timothée Proix5; 1Faculty of Psychology and Educational Sciences, University of Geneva, Switzerland, 2Department of Clinical Neuroscience, Faculty of Medicine, University of Geneva, 3Department of Fundamental Neuroscience, Faculty of Medicine, University of Geneva, 4Division of Neurology, Geneva University Hospital, Geneva, Switzerland, 5Institute of Neuroinformatics, University of Zürich and ETH Zürich, Switzerland
How does the brain encode the sensorimotor properties of word meaning during continuous speech comprehension? Embodied theories of language propose that word meaning is grounded in perceptual and motor experience [Barsalou, 1999], yet most neural evidence comes from controlled single-word paradigms [Pulvermüller, 2013], with few exceptions using continuous speech [Huth et al., 2016]. Here, we use stereoelectroencephalography (sEEG) to test whether sensorimotor word properties predict neural responses during continuous speech listening, and which specific and interpretable semantic dimensions drive those responses. Neural data consist of word-locked sEEG responses from 4 patients - to date - with drug-resistant focal epilepsy implanted with 135-145 electrodes during their clinical stay for pre-surgical evaluation. Participants listened to 20-24 stories . The short stories (texts: 170 - 300 words, audios: 54-100 seconds) are designed to vary in their sensorimotor content, with each story weighted toward particular sensorimotor dimensions, making them well-suited for dissociating the neural contribution of different modalities and effectors (e.g., one story narrates a sportive event, another is about a person creating objects with different textures, the other one describing a scenery, etc…). Embodiment is operationalized via the Lancaster Sensorimotor Norms [LSN, Lynott et al., 2020] which is an 11-dimensional vector spanning six perceptual modalities (touch, hearing, smell, taste, vision, interoception) and five action effectors (mouth/throat, hand/arm, foot/leg, head, torso), which are normed across nearly 40,000 words. Here, rather than treating embodiment as a single scalar, we ask which sensorimotor dimension(s) explain neural variance, at which latencies relative to word onset, and in which cortical regions. We hypothesize that different semantic dimensions will recruit spatially and temporally distinct neural responses, consistent with the view that semantic representations are distributed across modality-specific cortical regions. We fit time-lagged linear encoding models electrode-wise with regularization, comparing LSN dimensions on unique variance explained in held-out data. This allows us to map which sensorimotor properties of words best account for neural activity across the electrode array and across time. As more data becomes available, we plan to situate embodiment within a broader comparison of linguistic feature spaces, including surprisal [Frank et al., 2015], semantic distance [Rabovsky et al., 2018; Reilly et al., 2025], concreteness [Brysbaert et al., 2014] and emotional valence [Cato et al., 2004], to examine which dimensions of word meaning most strongly shape neural responses during continuous speech.
Topic Areas: Computational Approaches, Meaning: Lexical Semantics