The mental lexicon is a well-known theoretical construct in psycholinguistic research but is relatively unexplored in large language models (LLMs). We prompted lexicon functions in ChatGPT and compared numeric responses to a sample of 390 words for a range of lexical variables, some derived from corpus analyses and some from human ratings. Responses for were correlated with corpus measures and human ratings, more so for GPT-4 versus GPT-3.5, and for some variables compared with others. This variability indicates that functions are not recalled from memorized training data but are instead soft-assembled from general-purpose representations. Further analyses showed that GPT-4 functions were sensitive to subtle changes in word list composition, and they more closely matched human ratings for variables and words with higher inter-rater reliability. LLMs offer a mental lexicon in which functions emerge through a process of soft-assembly directed by task demands and shaped by context.

Authors:

Christopher Kello: University of California Merced; Polyphony J Bruna: University of California Merced