LLM-Driven CIDOC-CRM Knowledge Graph Extraction from Museum Archives

Museum archives contain vast amounts of cultural heritage information, but accessing, integrating and analyzing this data effectively remains challenging. Knowledge graphs (KGs), particularly those structured according to ontologies such as the CIDOC Conceptual Reference Model (CIDOC-CRM), offer a powerful solution through semantic representation. However, traditional KG construction is labor-intensive. Recent advancements in Large Language Models (LLMs) present an opportunity for automated knowledge extraction. This paper investigates the use of LLMs for constructing CIDOC-CRM knowledge graphs from museum archives, employing a single-stage in-context learning approach. We present a case study using the online archives of the Museum Nasional in Jakarta, Indonesia. The constructed knowledge graphs are evaluated using both intrinsic (ontology conformance) and extrinsic (Knowledge Base Question Answering - KBQA) metrics. Evaluation results show that larger models (e.g., DeepSeek R1) are capable of high ontology conformance (up to 98.73%) while smaller models struggle to reach 50% conformance. While all generated KGs improved KBQA performance over a no-context baseline, the ontology-guided approach showed limited and mixed impact on KBQA accuracy compared to the ontology-free graphs in our setup. This research underscores the potential of LLMs for generating ontology-conformant KGs within the cultural heritage domain and highlights the need for more sophisticated evaluation benchmarks to assess CIDOC-CRM KG generation methods.