No abstract

No abstract

Children learn their first language in a highly multimodal environment. This paper outlines a quantitative framework capturing children's typical language acquisition through multimodal conceptual features. Building on prior research from cognitive network science and distributional semantic theories from natural language processing, this work models toddlers' learning environment, between months 18 and 30, as either a multiplex lexical network capturing phonological/semantic/visual/sensorimotor and latent conceptual similarities, or as a collection of vectorial latent/sensorimotor/visual word embeddings. Each layer represents a set of information that toddlers might use to learn words over time. By comparing both attachment and acquisition, we reproduce past results about preferential acquisition capturing correlations with normative learning when using a semantic/phonological multiplex network. We extend this approach to show that: (i) preferential attachment can capture strong signals of normative word acquisition but only when visual and latent aspects of words are merged in a multiplex network with semantic/syntactic/phonological layers; (ii) preferential acquisition produces overall stronger signals in all other instances (in agreement with approaches); (iii) evidence for anti-correlations show the prevalence of word distinctiveness across early word learning strategies, as also identified in past approaches. We also explore cosine distance as a new attachment method for layers that are derived from embeddings and, as has been shown in prior work with multiplex networks, only when all layers are used do patterns emerge that correlate with normative word learning. Word embeddings and network structures provide analogous results, indicating how the combination of these structures for modelling strategies in word learning represents a viable and promising direction for future research.