Towards a Neuro-Symbolic Architecture for Trustworthy Cultural Heritage Question Answering

Although contemporary Large Language Models (LLMs) are highly effective in producing coherent conversational responses in various natural languages, their application in cultural heritage domains is frequently limited by problems related to trustworthiness, explainability, and knowledge grounding due to incomplete information and potential hallucinations. Linked Open Data (LOD) provides structured cultural heritage knowledge, but LOD alone is typically insufficient for visitor-style questions, because narrative details such as designers, alternative names and contextual descriptions are rarely represented in structured form. This gap motivates enrichment methodologies that extend LOD-based knowledge graphs (KGs) with narrative facts while preserving provenance. This paper presents the preliminary architecture and prototype of a reliable question-answering (QA) AI assistant focused on the cultural heritage of the City of Dubrovnik. This work investigates an ontology-driven methodology, in multiple stages, that extends LOD-based knowledge by extracting narrative attributes from representative domain documents using foundational LLMs. The extracted facts are combined with Wikidata entities into a specialized ontology optimized for typical visitor queries. The system employs a hybrid storage architecture that integrates AllegroGraph for RDF triple management and Neo4j for efficient property graph traversals. By linking natural language queries to explicit RDF triples, it is demonstrated that efficient LLM generation can be systematically connected to verifiable graph nodes representing factual information about the real world. Preliminary findings indicate that this architecture effectively reduces hallucination thus ensuring that conversational responses are explicitly grounded in their source entities.