Presentation
Search Abstracts | Symposia | Slide Sessions | Poster Sessions | Lightning Talks
Coordination of statistical and linguistic information during spoken language comprehension
There is a Poster PDF for this presentation, but you must be a current member or registered to attend SNL 2023 to view it. Please go to your Account Home page to register.
Poster A28 in Poster Session A, Tuesday, October 24, 10:15 am - 12:00 pm CEST, Espace Vieux-Port
This poster is part of the Sandbox Series.
Anna Mai1, Andrea E. Martin1,2; 1Max Planck Institute for Psycholinguistics, Nijmegen, Netherlands, 2Donders Center for Cognitive Neuroimaging, Nijmegen, Netherlands
In speech comprehension, while it is generally accepted that contextual probability can be bootstrapped for part of speech (POS) identification, the extent to which pre-existing language knowledge shapes how the brain responds to available statistical information is comparatively less explored. To determine how language knowledge may modulate the neural response to natural speech statistics, this study leverages maximum noise entropy (MNE) models (Kaardal et al. 2017) that will assess how statistical and linguistic features interact to explain magnetoencephalography (MEG) data recorded during natural speech listening. MNE models estimate the relationship between the stimulus and neural response as logistic functions of linear combinations of sets of stimulus features. First-order models assert that neural responses arise from the contributions of individual stimulus features (feature variance), and second-order models sum single feature contributions with contributions made by pairs of stimulus features (feature variance and covariance). Thus, by comparing the fits of first- and second-order models, this study will assess the manner in which relationships between statistical and POS features contribute to model goodness-of-fit beyond the contributions of individual features themselves, which has been the typical focus of previous literature. To date, MEG data have been recorded from 37 participants listening to short folktales in Dutch spoken by a native speaker, and these data have been time-aligned and annotated for phone, phone entropy, word, word frequency, word surprisal, and POS. Subsequently, broadband LFP (0.1-170Hz) and functional band power (δ: 2-4Hz, θ: 4-6Hz, α: 8-12Hz, β: 15-30Hz, ɣ: 30-50Hz) will be derived from the recordings. First- and second-order models will be fit for the five functional bands and broadband LFP of each participant for two types of stimulus feature sets: one set containing acoustic and statistical features only, and one set additionally including POS information. Feature-shuffled models will serve as controls. Fit models will be used to generate predicted neural responses for each sensor, and Pearson's r will be calculated and normalized with the Fisher Z-Transformation to assess the correlation of recorded vs. predicted responses across conditions. The bulk of these analyses should be complete by October 2023. We predict a significant interaction between the effect of model order (first vs. second) and the presence of POS information on MNE model prediction quality, driven by second-order models that include POS information. This would suggest that morphological and statistical information jointly account for neural activity that neither feature set does individually. Alternatively, if all second-order models equally outperform first-order models, it suggests that relationships between features are important to the brain but that POS information recapitulates statistical information. If all models perform identically, it suggests that the brain does not make use of covariance information for the feature types included in these models. Thus, regardless of its outcome, this study will add valuable structure to our understanding of how statistical and language-specific knowledge interact during language comprehension. In this way, this study advances knowledge of how speech comprehension emerges from the intricate coordination of statistical information with structured linguistic knowledge.
Topic Areas: Morphology, Computational Approaches