Masterarbeit

Modern research infrastructures are challenged by isolated data silos that restrict knowledge exchange and interoperability between university systems. The Across Alliance addresses this fragmentation by developing a Federated Knowledge Graph (KG), but the ontology underlying the Across KG has not yet been systematically aligned with globally recognized standard vocabularies, leaving it semantically isolated from European and national research-data ecosystems. Although manual alignment may be feasible for a single alliance, it requires substantial domain expertise and does not scale to the many university alliances that could benefit from it, while established automated ontology-matching systems rely primarily on lexical and structural features and may miss deeper semantic correspondences.

Recent advances in Large Language Models (LLMs) offer new possibilities for ontology matching, if challenges such as hallucinations, non-determinism, and inconsistent outputs are explicitly addressed. This thesis investigates a domain-specific approach in which candidate mappings between the Across ontology and selected standard vocabularies are generated by multiple open-weight LLMs using prompts specialised for the university-alliance research domain and benchmarked against the established systems AgreementMakerLight (AML) and LogMap as well as the recent LLM-based toolkit OntoAligner. The reference alignment is constructed independently by the author from the source ontology and the target vocabularies and validated by domain experts, using formal Semantic Web constructs (e.g., OWL, RDFS, and SKOS). Each system is then evaluated separately against this independently constructed, expert-validated reference alignment.

The primary objective of this master’s thesis is to empirically evaluate the suitability of domain-specific LLM-based ontology matching for aligning a research-infrastructure ontology with multiple heterogeneous standard vocabularies– including the Basic Formal Ontology (BFO), OpenAIRE, and the Kerndatensatz Forschung (KDSF)– benchmarked against the current state of the art using precision, recall, and F1-score, complemented by a qualitative error analysis. As an applied outcome, the same procedure establishes semantic interoperability for the existing Across KG by aligning its ontology with these standards. The final deliverables are empirical evidence on the performance and limitations of LLM-based ontology matching in the research-data domain and the aligned Across ontology, together with documentation of the reference-alignment methodology to support reproducibility and independent assessment.