Presentation

Search Abstracts | Symposia | Slide Sessions | Poster Sessions | Poster Slams

Developmental Continuity in Neural Semantic Representations

Poster B35 in Poster Session B and Reception, Thursday, October 6, 6:30 - 8:30 pm EDT, Millennium Hall

Albert Kim¹, Tanya Levari³, Stefano Anzellotti², Joshua Hartshorne²; ¹University of Colorado, Boulder, ²Boston College, ³Harvard University

When we learn a word, how is this information represented in the mind and brain and how do such representations change with development? The advent of machine learning methods for interpreting brain data promises to revolutionize our search for the answers to this longstanding question. The current work addresses two major barriers to progress to date. First, functional neuroimaging with children is challenging, given children’s limited attention span and tendency to fidget. Second, most work to date has taken a data-driven approach that does not harness the rich theoretical and computational tradition of models of linguistic semantic knowledge. We present the first in a series of studies addressing these issues, making use of a promising new method for recording EEG in children during free listening to engaging stories (Brennan, et al. 2019; Levari & Snedeker, 2018) and applying existing, theory-driven, quantitative representations of word meaning. 27 children (ages 5-10, mean=7.5) and 21 adults listened to a 1594-word recorded excerpt from Matilda, by Roald Dahl (2003). EEG was recorded at 500hz using Brainvision’s Actichamp System with 32 active electrodes placed at International 10-20 System locations and on the left and right mastoids, and filtered to 0.1-40 Hz. Data were epoched from -200ms to 1000ms relative to onset of each critical word and baseline corrected using the pre-stimulus time window (-200–0ms). Ocular artifacts were removed through independent component analysis. Voltage movements greater than 100μV resulted in trial rejection. Analyses focused on ERPs for the 107 distinct verbs and 138 distinct nouns in the excerpt. For types with multiple tokens, we averaged ERPs over tokens. Data were analyzed using Representational Similarity Analysis (RSA; Kriegeskorte et al., 2008). RSA is essentially a correlation of correlations, measuring the degree to which stimuli that are similar under one metric are similar under another. Initial analyses established reliability: separate RSAs for verbs and nouns were significant and well above chance for both children and adults. Interestingly, RSAs comparing children to adults were nearly as strong as RSAs comparing children to other children or adults to other adults, suggesting little developmental change. We next replicated prior findings of systematic relationships between ERPs and neural network word embeddings such as fastText, which capture substantial amounts of lexical semantics and morphosyntax (Bojanowski et al., 2017; He et al., 2022). RSAs were significant for both nouns and verbs for both children and adults, though only between ⅓ and ½ the size of the between-subject RSAs, suggesting word embeddings capture only some of the systematic variability in ERPs. Finally, and critically, RSAs were nearly as strong when using pairwise distances in WordNet (a hierarchical ontology of meaning) or human pairwise similarity judgments (Small World of Words; De Deyne et al., 2019). However, we found no relationship between ERPs for verbs and pairwise similarity in terms of argument structure participation (as recorded in VerbNet; Kipper et al., 2006), despite prior evidence that the latter is highly correlated with semantics.

Topic Areas: Meaning: Lexical Semantics, Development