Simulating the Noun-Verb Asymmetry in the Productivity of Children’s Speech

Abstract

Several authors propose that children may acquire syntactic categories on the basis of co-occurrence statistics of words in the input. This paper assesses the relative merits of two such accounts by assessing the type and amount of productive language that results from computing co-occurrence statistics over conjoint and independent preceding and following contexts. This is achieved through the implementation of these methods in MOSAIC, a computational model of syntax acquisition that produces utterances that can be directly compared to child speech, and has a developmental component (i.e. produces increasingly long utterances). It is shown that the computation of co-occurrence statistics over conjoint contexts or frames results in a pattern of productive speech that more closely resembles that displayed by language learning children. The simulation of the developmental patterning of children’s productive speech furthermore suggests two refinements to this basic mechanism: inclusion of utterance boundaries, and the weighting of frames for their lexical content.