Proceedings of the Annual Meeting of the Cognitive Science Society

Young children often stretch terms to novel objects when theylack the proper adult words—a phenomenon known as overex-tension. Psychologists have proposed that overextension relieson the formation of a chain complex, such that new objectsmay be linked to existing referents of a word based on a diverseset of relations including taxonomic, analogical, and predicate-based knowledge. We build on these ideas by proposing a com-putational framework that creates chain complexes by multi-modal fusion of resources from linguistics, deep learning net-works, and psychological experiments. We test our models ina communicative scenario that simulates linguistic productionand comprehension between a child and a caretaker. Our re-sults show that the multimodal semantic space accounts forsubstantial variation in children’s overextension in the liter-ature, and our framework predicts overextension strategies.This work provides a formal approach to characterizing lin-guistic creativity of word sense extension in early childhood.