Predicting individual vocabulary learning: The importance of approximating toddlers' linguistic environment.

Abstract

Using network representations of the lexicon has expanded our understanding of vocabulary growth processes and vocabulary structure during early development. These models of vocabulary development have used multiple types of sources to create lexical representations. More recently, Weber and Colunga (2022) demonstrated that predictions of early vocabulary norms can be improved by using network representations based on a corpus incorporating language a young child might typically hear. The present work goes a step further by evaluating the accuracy of network representations for predicting individual children's word learning that are based on embeddings that are readily available or embeddings gathered from the same child language corpus. We predicted the specific words that individual children add to their vocabulary over time, using a longitudinal data set of 86 monolingual English-speaking toddler's changing vocabulary from 18 to 30 months of age. The toddler-based network predicted word learning more accurately than the off-the-shelf network. Further, there was an advantage for prediction methods that took into account the individual child's particular network structure rather than overall network connectivity. These results highlight the importance of tailoring representational and processing choices to the population of interest. (PsycInfo Database Record (c) 2025 APA, all rights reserved).