Large Linguistic Models: Analyzing theoretical linguistic abilities of LLMs

The performance of large language models (LLMs) has recently improved to the point where the models can generate valid and coherent meta-linguistic analyses of data. This paper illustrates a vast potential for analyses of the meta-linguistic abilities of large language models. LLMs are primarily trained on language data in the form of text; analyzing their meta-linguistic abilities is informative both for our understanding of the general capabilities of LLMs as well as for models of linguistics. In this paper, we propose several types of experiments and prompt designs that allow us to analyze the ability of GPT-4 to generate meta-linguistic analyses. We focus on three linguistics subfields with formalisms that allow for a detailed analysis of GPT-4's theoretical capabilities: theoretical syntax, phonology, and semantics. We identify types of experiments, provide general guidelines, discuss limitations, and offer future directions for this research program.