What language models can tell us about learning adjectives

Abstract

It has been argued that language models (LMs) inform our knowledge of language acquisition. While LMs are claimed to replicate aspects of grammatical knowledge, it remains unclear how this translates to acquisition directly. We ask if a language model trained specifically on child-directed speech (CDS) is able to capture grammatical knowledge of adjectives. Ultimately, our results reveal that what the model is “learning” is how adjectives are distributed in CDS, and not the grammatical properties of different adjective classes. While highlighting the ability of LMs to learn distributional information, these findings suggest that LMs alone cannot explain how children generalize beyond their input.