Basic argument structure is adult-like by 30 months: An evaluation of more than 700,000 child utterances

Children rapidly acquire a wide range of linguistic skills, from perceptual to pragmatic over the first few years of life. Syntax, the ability to construct and infer meaning from word order, blossoms around 2 years of age. Within syntax, some of the most widely studied features are the development of subjects and objects. However, computational identification of subjects and objects in natural language is not yet a solved problem. In this paper, we use a public machine-learning model to identify internal and external arguments from child speech, and show that it functions with high precision and recall (>0.92), and furthermore, tracks very closely with results from the same model applied to adult speech. We demonstrate this model is suitable for application to child speech, and show growth curves highlighting the rates of maximum growth for internal, external, and tertiary arguments in nearly 1 million child utterances.