Quantifying Relational Exploration in Cultural Heritage Knowledge Graphs with LLMs: A Neuro-Symbolic Approach for Enhanced Knowledge Discovery

This paper introduces a neuro-symbolic approach for relational exploration in cultural heritage knowledge graphs, exploiting Large Language Models (LLMs) for explanation generation and a mathematically grounded model to quantify the interestingness of relationships. We demonstrate the importance of the proposed interestingness measure through a quantitative analysis, highlighting its significant impact on system performance, particularly in terms of precision, recall, and F1-score. Utilizing the Wikidata Cultural Heritage Linked Open Data (WCH-LOD) dataset, our approach achieves a precision of 0.70, recall of 0.68, and an F1-score of 0.69, outperforming both graph-based (precision: 0.28, recall: 0.25, F1-score: 0.26) and knowledge-based (precision: 0.45, recall: 0.42, F1-score: 0.43) baselines. Furthermore, the proposed LLM-powered explanations exhibit better quality, as evidenced by higher BLEU (0.52), ROUGE-L (0.58), and METEOR (0.63) scores compared to baseline approaches. We further demonstrate a strong correlation (0.65) between the interestingness measure and the quality of generated explanations, validating its ability to guide the system towards more relevant discoveries. This system offers more effective exploration by achieving more diverse and human-interpretable relationship explanations compared to purely knowledge-based and graph-based methods, contributing to the knowledge-based systems field by providing a personalized and adaptable relational exploration framework.