Investigating the applications and limitations of GPT-4 in conceptual metaphor analysis in political discourse: An experimental approach

This study used the GPT-4 model to analyze conceptual metaphors in the political discourse of the 2024 U.S. presidential election candidate TV debates and evaluate the AI's language processing capabilities and limitations. The experimental results showed a 13.56% error rate and 86.44% precision in the overall analysis. For the partial correct items, which accounted for 46.61% of the total correct results, the error rates were as follows: TD error rate 17.80%, SD error rate 9.32%, TD&SD error rate 8.47%, and mixed metaphor error rate 11.02%. In addition, the model missed mixed metaphors, showing the limitations of metaphor detection. The study shows that while AI is highly accurate at analyzing conceptual metaphors, it does not fully understand complex figurative language due to differences in contextual analysis and human cognitive understanding. The results provide academic insights into AI's language processing capabilities and suggest directions for further research.