Explainable semantic question answering system for cultural heritage: a framework based on a dual-layer knowledge graph and self-RAG-inspired adaptive retrieval—a case study of Liao Dynasty tomb mural costume

In the domain of complex cultural heritage, the profound confluence of physical forms and cultural semantics poses a primary challenge for knowledge retrieval and reliable interpretation, a problem vividly epitomized by the multi-layered structures and multicultural characteristics of Liao Dynasty costumes. Existing methods based on general RAG or single structured querying struggle to effectively bridge the gap between physical morphology and cultural semantics. This often results in high retrieval noise, semantic drift, and a lack of interpretability. To address this, we propose a domain-adaptive explainable semantic question-answering framework for Liao Dynasty costume, grounded in a dual-layer ontology and adaptive retrieval augmentation. First, a dual-layer knowledge graph encompassing “morphological structure layer” and “cultural semantics layer” is constructed, mapping visual elements from 80 Liao Dynasty tomb murals into computable semantic structures. Second, domain-adaptive vectorization is applied to the graph, alongside a Self-RAG-inspired adaptive retrieval mechanism. Through a two-stage decision process involving “retrieval triggering and relevance filtering,” the framework reduces the injection of weakly relevant cultural contexts. Finally, we construct a generative framework that visualizes evidence chains. Experimental results demonstrate that the framework achieves a Recall@10 of 0.88 for cultural-. semantic queries, and reduces the Top-10 noise ratio to 13.0%. Furthermore, within the expert-curated test set, it shows. higher average scores than traditional RAG and Cypher query baselines in subjective evaluation metrics such as factual accuracy and evidence traceability. This study provides a domain-adaptive reference framework for knowledge association and trustworthy interpretation in complex cultural heritage scenarios.