Proceedings of the Annual Meeting of the Cognitive Science Society

As children learn language, they organize their knowledge in lexical-semantic networks. Comparing pre-existing methods of accessing underlying networks, we examine developmental verbal fluency in 5-8-year-olds (N=37, mean age=80.27 mos) across two different prompt types — taxonomic (animals and foods) vs. (location-based) thematic prompts (zoo and grocery store) — using two different graph-theoretic estimation strategies: random-walk modeling (e.g., U-INVITE) and GloVe word embeddings. We observed several consistencies: taxonomic prompts elicited more words than thematic prompts (linear mixed-effects model: t(35) = 3.16, p<0.05); networks expanded with age (U-INVITE, t(35) = 4.26, p<0.05; ESN, t(35) = 4.53, p<0.05); and structures spread out (versus clustering densely). However, key differences emerged. Random-walk networks uncovered different highest-degree (most densely clustered) words depending on the prompt type (e.g., "dog" for animals, "monkey" for zoo). By contrast, networks based on word embeddings identified networks with very similar highest-degree words for animals and zoo. Hence, alternative assumptions informing method choices may result in distinct network estimates, with consequences for how we map the growth of lexical knowledge.