Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Statistical regularities shape interactive neural representations of orthographic-semantic mappings
Poster Session B, Wednesday, September 30, 4:30 - 6:30 pm, Wangari Maathai
Rujun Duan1, Gangyi Feng1,2; 1Department of Linguistics and Modern Languages, The Chinese University of Hong Kong, 2Brain and Mind Institute, The Chinese University of Hong Kong
Learning to read requires the brain to discover how visual forms map to meaning. Connectionist accounts propose that this mapping is not a feedforward transformation from orthography to semantics but an interactive process in which form- and meaning-level representations become mutually constraining through experience. However, it remains unclear how such statistical regularities are neurally represented during learning and how orthographic and semantic regularities constrain one another to shape the mapping over time. We tested this question using a feedback-based visual word-learning paradigm with EEG in 57 adults. Participants learned 20 artificial logographic characters, each paired with a familiar visual object. Characters contained semantic radicals that varied in their mapping reliability. In the high-consistency condition (90%), the radical predicted the same semantic category for most items; in the low-consistency condition (20%), the radical provided weaker semantic-category information. During learning, participants selected which of two characters corresponded to a target picture and received corrective feedback. Behavioral results showed robust learning, recognition, and generalization of orthographic-semantic regularities. To characterize the neural representational dynamics underlying learning, we applied time-resolved representational similarity analysis (RSA) to EEG activity. We used three theoretically motivated representational dissimilarity matrices: an orthographic model based on shared radical structure, a semantic model derived from GPT ratings of the pictured meanings, and an orthographic-semantic mapping model capturing learned form-meaning correspondences. Orthographic, semantic, and mapping-related structures were detectable in neural activity from approximately 100 ms after stimulus onset, indicating rapid access to newly learned distributional structure. Critically, RSA with linear mixed-effects modeling revealed that orthographic and semantic representations interacted dynamically across time. In an early window around 110-150 ms, orthographic structure was expressed when semantic distance was low, suggesting rapid recruitment of visual-radical information when semantically similar meanings compete. In a middle window around 160-240 ms, orthographic structure was robust across semantic distances, consistent with a transient stage in which learned visual regularities are strongly represented. In a later window around 350-380 ms, orthographic effects emerged only when semantic distance was high and reduced when meanings were similar, suggesting that semantic convergence increasingly constrains or overrides residual orthographic distinctions. These findings provide time-resolved neural evidence that newly learned written forms are organized according to the statistical reliability of their mappings to meaning. Rather than supporting a strictly feedforward account in which orthographic analysis precedes semantic access, the results reveal an interactive representational architecture in which form and meaning are dynamically co-specified during learning and processing. This suggests that statistical regularities in orthographic-semantic mappings serve as a core organizing principle by which experience shapes the neural representations underlying word recognition.
Topic Areas: Meaning: Lexical Semantics, Reading