Sequence structure in children's speech reveals non-linear development of relations between word categories

Why do children learn some words earlier than others? Can children's speech patterns reveal how their evolving models of language determine what they learn? This study presents a systemic analysis of children's speech using low-dimensional embeddings to examine how the contextual knowledge reflected in their utterances reorganizes as linguistic experience increases. We analyzed age-stratified samples from the CHILDES database (18-36 months: n = 1,693,641 tokens; 3-6 years: n = 1,750,007; 5-12 years: n = 1,721,828) and adult speech from the SUBS2VEC subtitle corpus (n = 1,742,885). Our results suggest that the order and position of words in sequences produced by children from different age groups reflect changes in the way they represent categories of words. Rather than being ungrammatical, children's utterances appear to be structured by temporary grammars that optimize the distribution of information in sequences. The results point to shifts in how words are organized in semantic space, reflecting the gradual alignment of lexical categories during learning; this restructuring appears to draw on functionally ambiguous (multipurpose) categories in English. These findings are somewhat counterintuitive, as they suggest that not knowing the exact meaning of words can facilitate both learning and communication.