Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Developing by predicting: neural representations of naturalistic linguistic prediction mature with age and support language development
Poster Session C, Saturday, September 13, 11:00 am - 12:30 pm, Field House
Xinyi Tang1, Enyu Liu1, Xi Yu1; 1State Key Laboratory of Cognitive Neuroscience and Learning, Beijing Normal University
Introduction: The human brain is classically conceptualized as a prediction machine. In the domain of language, convergence between the adult brain and large language models (LLM) highlights that prediction is a core mechanism supporting language comprehension, with LLMs that are better at predicting upcoming words also more accurately capturing human brain responses to natural language. Behavioral studies further indicate that children rely on predictive coding abilities to acquire linguistic knowledge, such as words and sentence structure. However, the neural mechanisms underlying predictive coding in the developing brain, whether domain-general or domain-specific, in supporting language acquisition, remain largely unknown. The current study explored these questions by combining the predictive power of LLMs with a large-scale pediatric neuroimaging dataset. Methods: This study included 1,383 participants aged 5-22 years from the Healthy Brain Network dataset. All participants had usable fMRI data collected while watching a 10-minute clip from the movie Despicable Me. GPT-3 model was used to calculate the log probability of each spoken word given its preceding context in the clip. These probabilities were then converted into surprisal values, quantifying linguistic predictions during movie-watching. A voxel-wise encoding model using banded ridged regression was implemented to estimate the linear mapping from movie features onto brain activity. Model features included GPT-3 surprisal, lexical frequency, emotional valance, presence of words, bodies, and faces, and low-level audiovisual features. Model performance was evaluated on held-out test sample through 5-fold cross-validation by correlating the actual and model-predicted time series, using either the full model or individual features. One-sample t-tests at both voxel and regions of interest (ROI) levels were used to identify regions that significantly predicted by GPT-3 surprisal. Support vector regression (SVR) with 10-fold cross-validation then tested whether encoding performance in these regions could predict age and language development, controlling for motion, sex, scan location, and handedness, and significance was determined after permutation (n=1000). Results: (1) Both univariate and ROI-level analyses revealed that regions within a bilateral extended language network, including inferior frontal gyri (IFG), middle frontal gyri, superior temporal gyri and middle temporal gyri, were significantly predicted by GPT-3 surprisal feature (ts > 9.05, ps < 1e-19). (2) SVR model showed significant associations between encoding performance of GPT-3 surprisal and chronological age in the left-hemisphere language network only (left: r = 0.21, p < 0.001, mean squared error (MSE) = 12.73, p < 0.001; right: r = 0.11, p < 0.001, MSE = 16.08, p = 0.661). (3) When additionally controlling for age and nonverbal IQ, SVR revealed that only encoding performance of GPT-3 surprisal in the orbital part of right IFG significantly predicted individual differences in language development (r = 0.11, p < 0.001, MSE = 19.49, p = 0.001). Conclusion: This study demonstrates that linguistic predictions generated during naturalistic movie-watching are selectively encoded in the language network, highlighting the domain specificity of predictive language processing. Moreover, neural representations of linguistic predictions predict both age and individual difference in language skills, further emphasizing the role of predictive coding in supporting language development.
Topic Areas: Language Development/Acquisition,