Metaphor Processing in Political Discourse : A Comparative Evaluation of Conceptual Mapping in GPT-4 and Gemini

초록

This study investigates how two state-of-the-art large language models (LLMs), GPT-4 and Gemini, process metaphors in the cognitively and rhetorically demanding context of the 2024 U.S. presidential debate. Using candidate-only utterances (N = 1,188), each model labeled sentences as metaphorical or non-metaphorical, enabling paired analysis. McNemar’s exact test showed that Gemini marked significantly more sentences as metaphorical than GPT-4 (p < .001), revealing a bias toward over-detection. Yet once metaphors were identified, both models exhibited similar distributions of conceptual mappings—full accuracy, partial mismatches, and full errors—with no significant differences. From a cognitive linguistics perspective, these findings reaffirm metaphor as a central mechanism of thought, grounded in embodied experience, while highlighting the persistent gap between human conceptual integration and statistical modeling. Methodologically, the study underscores the value of evaluating LLMs in authentic political discourse, advancing dialogue between cognitive linguistics and NLP on the boundaries of computational metaphor processing.

키워드