Interpreting Structured Narrative Representations in Pretrained Language Models

Abstract

This thesis investigates whether unstructured embedding spaces—such as those learned by large language models (LLMs)—implicitly encode the structured knowledge required for narrative understanding. Storytelling is central to human cognition and social interaction, and by probing how AI systems process narrative structure, we gain insight into both the capacities of these models and the contours of human intelligence itself. The work proceeds along two primary lines of inquiry. First, I broadly survey the research landscape, revisiting foundational claims about distributed representations in language models and asking whether modern Transformer architectures might natively support narrative understanding tasks. I describe key limitations that suggest current architectures fall short of representing the deeper structures that underpin narrative meaning. Second, I ask whether LLMs encode structured patterns characteristic of human storytelling. Through a series of controlled experiments focusing on those structured patterns which are primarily evidenced and documented in classical texts like the Bible, I examine whether these models (1) detect internal rhetorical structures such as chiasmus, (2) recognise cross-textual parallels, (3) surface passages grounded in shared narrative frames rather than surface form, and (4) make analogical inferences about narratives based on deep conceptual structure rather than lexical overlap. Together, these investigations offer a comprehensive evaluation of whether the unstructured embedding spaces learned by LLMs can support structured, hierarchical, and analogical reasoning. The results shed light on the representational limits of current language models and point toward the need for architectures that integrate narrative structure more explicitly—advancing both cognitive modelling and downstream tasks that rely on robust, explainable meaning representations.