Intelligent Information Management / Web Engineering

Knowledge graphs are fundamental in retrieval-augmented generation systems for knowledge graph question answering. AI assistants supported by such systems can produce answers that are more fact-grounded and less prone to hallucinations, since their responses are based on structured knowledge. While substantial research has focused on improving the retrieval of information from knowledge graphs, for example through better SPARQL query generation, comparatively less attention has been given to assessing the validity of the claims made by the assistant itself.

This thesis examines approaches to claim verification on knowledge graphs. More specifically, it addresses the following problem: given a claim, represented as the hypothesis, and a knowledge graph represented by RDF triples, serving as the premises, to what extent is the hypothesis entailed by the premises? In this context, the hypothesis corresponds to a claim generated by an AI assistant, while the premises correspond to the subgraph retrieved and used as the basis for the assistant’s answer.

To address this problem, the thesis first introduces the topic and outlines its motivation and problem setting. It then presents a state-of-the-art analysis and a requirement analysis, identifying existing approaches and deriving the requirements for a suitable solution. Based on these findings, a concept is developed, and its feasibility is demonstrated through a prototypical implementation. Finally, the effectiveness of the prototype, 6specifically the performance of its entailment checking, is assessed through a meaningful evaluation.