Modeling Metaphors: Analyzing Fine-Tuning Effects in ChatGPT
Cassia Reddig
Department of Computer Science
Faculty Supervisor: Dragutin Petkovic
Fine-tuning enables large language models to specialize in specific domains, yet its effects on higher-level linguistic features like metaphor remain underexplored. This study investigates how fine-tuning alters metaphor frequency, consistency, and framing in model outputs. Leveraging Conceptual Metaphor Theory as an analytical framework, we compare pre- and post-fine-tuning responses to identify systematic shifts in language patterns. The results are expected to contribute to a deeper understanding of model behavior, with implications for interpretability, domain adaptation, and ethical deployment of AI models.