What Can Language Models Tell Us About Human Cognition?

Abstract

Language models are a rapidly developing field of artificial intelligence with enormous potential to improve our understanding of human cognition. However, many popular language models are cognitively implausible on multiple fronts. For language models to offer plausible insights into human cognitive processing, they should implement a transparent and cognitively plausible learning mechanism, train on a quantity of text that is achievable in a human’s lifetime of language exposure, and not assume to represent all of word meaning. When care is taken to create plausible language models within these constraints, they can be a powerful tool in uncovering the nature and scope of how language shapes semantic knowledge. The distributional relationships between words, which humans represent in memory as linguistic distributional knowledge, allow people to represent and process semantic information flexibly, robustly, and efficiently.