Computational analysis of narrative structures

Abstract

This thesis explores the computational analysis of high-level narrative structures in literary texts, bridging narratological theory with natural language processing techniques. Narratives are examined at two core levels: the story (histoire), encompassing temporal-causal event sequences and narrative arcs, and the narrative discourse (récit), involving elements such as scenes, levels, anachronisms and focalisation. The research addresses key challenges in computational narratology, including the scarcity of standardised annotated data sets and fragmented computational approaches that hinder large-scale analysis. A literature review maps existing terminology, methodologies, and data sets, revealing inconsistencies in definitions for terms like "plot", "scene", and "narrative level". To identify narrative arcs, two systems using latent dirichlet allocation (LDA) are developed and evaluated against human annotations on whodunit short stories, demonstrating feasible detection of single and multiple transitions. For narrative discourse, modular annotation guidelines are proposed and refined by addressing challenges in the annotation process. These guidelines enable the creation of a unified data set combining existing annotations with guidelines specifically designed for the mystery genre data. Fine-tuned RoBERTa (Robustly optimised BERT pretraining approach) models, trained on this data set, achieve promising results for automated discourse annotation, with the combined model yielding an average F1 score of 68.60 and high ROC-AUC (94.01), outperforming existing baselines. Other contributions include novel arc detection systems, standardised guidelines, and a publicly available data set advancing reproducible computational literary studies. Limitations, such as data set size and genre-specific focus, suggest future expansions to diverse genres and advanced architectures for enhanced scalability and interpretability.