Using Knowledge Graphs to Track Claim Evolution in Particle Physics Literature: An Ontology-First Approach

Abstract

Within the expanding landscape of scientific publications, researchers lack the tools necessary to track how scientific claims evolve across the literature. They also lack tools to understand, at a semantic level, how different authors build on each other’s work. Existing systems, such as Connected Papers and Research Rabbit, attempt to address this challenge using citation graphs; however, these graphs primarily capture the topological citation network rather than the underlying semantic relationships between papers. This thesis introduces a two-tier ontology architecture in OWL that combines categorical classification of papers with semantic comparison across citation edges. We focus specifically on particle physics literature. We selected this corpus because particle physics topics evolve in ways that produce especially rich and dynamic relationships across papers, and because the field provides a strong foundation for ontology design as a relatively closed community with shared terminology. The framework operates in two stages. In Step 1, the model extracts Purpose, Mechanism, and Evaluation facets from research paper abstracts using large language models, aligning them with upper-ontology classifications designed to serve as a more categorical classification step for each query. In Step 2, it classifies citation edges into eight relationship types, from S0 to S3, based on facet overlap, and encodes the results in an extensible 43-class ontology with logical axioms that enable automated reasoning. This model was applied to a corpus of 59 dark matter papers on hidden photons and axions, yielding a total of 415 citation edges. Human validation of 40 edges showed an 85% agreement with the pipeline. We also compared two different LLMs, Claude Sonnet and Haiku, to assess how results changed across models. Sonnet achieved 85% accuracy, while Haiku achieved 75%, with errors driven by vocabulary blindness and semantic overreach, respectively.