A Computational model for child inferences of word meanings via syntactic categories for different ages and languages

Abstract

Children exploit their morphosyntactic knowledge in order to infer the meanings of words. A recent behavioral study has reported developmental changes in word learning from three to five years of age, with respect to a child's native language. To understand the computational basis of this phenomenon, we propose a model based on a hidden Markov model (HMM). The HMM acquires syntactic categories of given words as its hidden states, which are associated with observed features. Then, the model infers the syntactic category of a new word, which facilitates the selection of an appropriate visual feature. We hypothesize that using this model with different numbers of categories can replicate the manner in which children of different ages learn words. We perform simulation experiments in three native language environments (English, Japanese, and Chinese), which demonstrate that the model produces similar performances as the children in each environment. Allowing a larger number of categories means that the model can acquire a sufficient number of obvious categories, which results in the successful inference of visual features for novel words. In addition, cross-linguistic differences originating from the acquisition of language-specific syntactic categories are identified, i.e., the syntactic categories learned from English and Chinese corpora are relatively reliant on word orders, whereas the Japanese-Trained model exploits morphological cues to infer the syntactic categories.