Using Large Language Models for Understanding Narrative Discourse
Andrew Piper, Sunyam Bagga · 2024
In this study, we explore the application of large language models (LLMs) to analyze narrative discourse within the framework established by the field of narratology.We develop a set of elementary narrative features derived from prior theoretical work that focus on core dimensions of narrative, including time, setting, and perspective.Through experiments with GPT-4 and fine-tuned open-source models like Llama3, we demonstrate the models' ability to annotate narrative passages with reasonable levels of agreement with human annotators.Leveraging a dataset of human-annotated passages spanning 18 distinct narrative and non-narrative genres, our work provides empirical support for the deictic theory of narrative communication.This theory posits that a fundamental function of storytelling is the focalization of attention on distant human experiences to facilitate social coordination.We conclude with a discussion of the possibilities for LLM-driven narrative discourse understanding.