Narrative Understanding with Knowledge Graphs

Narratives help humans make sense of experience, explain events, and establish shared understanding. Implementing such capabilities in AI can make systems more human-centric and aligned with how people reason, communicate, and explain. Beyond textual storytelling, narratives support a broad range of cognitive functions such as question answering, hypothesis generation, and discourse analysis, by linking new observations to background knowledge. While large language models have demonstrated remarkable proficiency in generating fluent text, they face limitations such as hallucinations or verifiability. Knowledge-based approaches, particularly knowledge graphs, offer interpretability and structured reasoning. This thesis combines the strengths of both traditions to support narrative construction across heterogeneous domains, investigating how knowledge graphs enable narrative-like functions such as explanation, hypothesis generation, and discourse analysis. Towards representing narrative-like structures as knowledge graphs, we first define narrative requirements and review ontological models for computational implementation. Practical guidelines are proposed for choosing suitable structured formats, which we apply to three use cases: historical explanation, analysis of social media discourse on inequality, generation of hypotheses in the field of social science. Despite differences in data formats and domain assumptions, these cases share a unified high-level approach. We split the construction of narrative-centric knowledge graphs into retrieving relevant content and transforming it into a knowledge graph. We introduce a semantically guided retrieval method for identifying event-relevant subgraphs and apply domain-adapted strategies to build knowledge graphs across the three use cases. We evaluate the effectiveness of structured narrative representations through quantitative and qualitative methods. In the historical domain, semantically guided retrieval improves content relevance, while overly complex narrative elements can hinder performance. Prompts enriched with event-centric knowledge graph improve factual accuracy in question answering without compromising succinctness. In the social science domain, AI-generated hypotheses, though not always surpassing human-generated ones, provide diverse options that complement and enhance the hypothesis generation process. Our findings show that structured narrative representations can support both automated and human-centered reasoning tasks. The three use cases demonstrate broad applicability and adaptability across different domains. The methods presented here advance AI systems that engage more meaningfully with human reasoning, enabling narrative-aware tools to assist in answering questions and analyzing complex phenomena.