Computational model for syntactic development: Identifying how children learn to generalize nouns and verbs for different languages

Abstract

By three years of age, children are supposed to start learning to understand syntactic structures, and at around five years of age, they are reported to be able to infer a syntactic category, such as a noun or a verb, for a novel word. Finding the syntactic cue enables them to infer a target directed by a novel word in visual stimuli. The study also found that their inference performances depended on their native languages. In this article, we propose a model to explain how children learn to generalize novel nouns and verbs in the Japanese, English, and Chinese languages. We use a Bayesian hidden Markov model (BHMM) to learn syntactic categories represented as hidden states in a BHMM. Here, an increase in the number of hidden states indicates the children's syntactic development. A model with a larger number of hidden states is able to infer a clearer syntactic category of a novel word, resulting in the correct choice of a category for the visual target. Syntactic categories that depend on input languages are acquired by BHMMs, and therefore result in different performances among the languages. We entered English-, Japanese-, or Chinese-corpus into the model and examined how the model inferred a correct target indicated by a novel word through the acquired syntactic categories. The results showed that the performances by our model are very similar to the children's performances. Further analysis of representations of hidden states clarified that the model acquires syntactic categories reflecting orders of words in English, suffixes in Japanese, and adverbs in Chinese.