Deterministic and Trustworthy LLM-Driven Semantic Knowledge Graph Construction

LLMs offer a powerful basis for automated knowledge graph construction, but their outputs are often non-deterministic and structurally unreliable, limiting their direct suitability for trustworthy semantic applications. This paper presents DOPKG-SHACL, a hybrid approach that combines deterministic Internationalized Resource Identifier (IRI) minting, canonicalisation, ontology alignment, provenance-aware materialisation, and Shapes Constraint Language-based (SHACL) validation. We evaluate the approach through three experiments. The first forms a baseline that involves repeated-run analysis with gpt-4o-mini Large Language Model (LLM) under greedy and stochastic decoding to show that raw extraction varies substantially across runs, with instability increasing as temperature rises, particularly for relations. The second involves a three-part experiment that compared the impact of raw and minimally cleaned data on (DOPKG-SHACL) to show that it produces structurally richer, provenance-aware and validation conformant graphs, achieving complete ontology mapping and full conformance with the implemented SHACL constraints. Third, an ablation study demonstrates that deterministic IRI minting, ontology alignment, canonicalisation, and SHACL validation each make important contributions to DOPKG-SHACL’s quality, with deterministic IRI minting and ontology alignment having the strongest overall effect. We argue that determinism is not achieved by the LLM alone, it emerges from a hybrid approach that constrains, canonicalises, materialises, and validates extraction outputs.