Interworlds: studying metaphor using neurosymbolic AI

This thesis advocates for a pattern-centric account of meaning, where shared knowledge patterns are used to guide and constrain large language models (LLMs), and evaluates it on the case of metaphor. In fact, metaphor is a core mechanism of human meaning-making across expert and lay discourse. At the same time, computational systems have long struggled with it, making it an appropriate domain of study. We first investigate ontology engineering as a more controlled setting for understanding to what extent can LLM-based methods assist ontology engineering tasks. Concretely, we introduce a suite of resources and methods for human-in-the-loop ontology engineering, including the Ontogenia generation framework and studies on automated ontology evaluation. To ensure robustness, we also propose CoLLM, a three-test framework for LLM consistency. Taken together, these studies show that heterogeneous, task-specific datasets are essential and that LLM support is helpful yet unstable: results drift across versions, and automated suggestions can bias experts. These outcomes argue for human-in-the-loop oversight, explicit reliability tracking, and pattern-guided constraints, which motivates our shift to computational metaphor processing, where these design choices are crucial. For this second strand, we introduce two resources, the Medical Metaphor Corpus (MMC), and the Balanced Conceptual Metaphor Testing Dataset (BCMTD), and we formalize Conceptual Blending mechanisms for metaphorical combination in a Blending Ontology. Building on this ontology, we implement a Logic Augmented Generation (LAG) neurosymbolic pipeline to represent, recognize, and analyze metaphors. Empirically, LLM-only pipelines perform strongly on clear cases yet systematically struggle on borderline metaphors; the proposed LAG-based approach outperforms baselines, improves robustness and interpretability, and preserves the dynamism of generative methods. This work has implications for clinical communication systems that must generate and interpret patient metaphors, educational tools, and any AI application where capturing or producing graded, context-sensitive meaning is essential.