Hypergraph models of the mental lexicon capture greater information than pairwise networks for predicting language learning

Abstract

Human memory is a complex system that works in associative ways: Reading a cue word can lead to the recollection of associated concepts. The network structure of memory recall patterns has been shown to contain insights about a wide variety of cognitive phenomena, including language acquisition. However, most current network approaches use pairwise connections, i.e. links between only two words at a time. This ignores the possibility that more than two concept representations might be simultaneously associated in memory. We overcome this modelling limitation by introducing cognitive hypergraphs as models of human memory. We model memory recall patterns through word associations from the Small World of Words project for N=6003 concepts (Study 1) and for N=497 concepts (Study 2). In each study we represent word associations as either a pairwise network or a hypergraph. By combining psycholinguistic norms and network centrality measures with machine learning, we quantitatively investigate whether there is any benefit to using the hypergraph model over a pairwise network in predicting test-based age of acquisition norms in children up to age 9 years (Study 1) or normative learning in toddlers up to age 30 months (Study 2, based on CHILDES data). We show that cognitive hypergraphs capture more information than pairwise networks from the same data: Cognitive hypergraphs are considerably more powerful than pairwise networks at predicting age of acquisition trends in toddlers, children and teenagers. Our studies showcase how novel approaches merging artificial intelligence and higher-order interactions can help us understand cognitive development.