Metaphorical minds : an investigation of large language models’ ability to adequately generate and process metaphors

Abstract

Creativity in language is a qualitative feature that is claimed to be unique to humans (Chomsky, 2006). Current Large Language Models (LLM), such as ChatGPT, appear to have mastered the skill to use non-literal language, such as metaphors, and have therefore supposedly crossed the threshold between machines and humans (Mei et al., 2024) when it comes to mastering linguistic creativity. However, it is not clear how well LLMs can understand and produce novel non-literal language compared to humans, which is what this dissertation aims to investigate. This dissertation explores the role of creativity in human language, covering classical and cognitive approaches in metaphor research, with a focus on novel metaphors. It reviews key literature in both linguistic theory and natural language processing, and presents a qualitative analysis of human- and machine-produced paraphrases to showcase metaphor interpretation. Broader implications for scientific research are discussed, particularly in comparing human and machine capacities for metaphorical understanding.