Ontology-Based Knowledge Organization and Multimodal Information Integration for Chinese Zisha Teapot Heritage

Chinese Zisha teapot heritage information is dispersed across catalogs, scholarly literature, and museum collections, presenting challenges for consistent knowledge organization, automated recognition, and structured information access. This study presents a provenance-aware framework combining ontology-based knowledge organization, multimodal teapot-type recognition, and graph-backed question answering. A domain ontology and Neo4j knowledge graph were constructed to organize catalog-derived information, with the initial graph containing 607 core teapot entities. For six catalog-derived operational categories, 543 image–text pairs were evaluated using Chinese-CLIP representations, unimodal baselines, feature concatenation, equal-weight fusion, and bidirectional cross-attention under grouped cross-validation and text-field ablation. Equal-weight fusion achieved a mean Macro-F1 of 84.41%±0.78% under author-based grouping, while cross-attention showed no statistically significant advantage over simpler fusion methods. Additionally, 543 archived recognition predictions were incorporated into the extended knowledge graph as separate, provenance-linked assertions. On a locked 120-question benchmark, the structured question-answering system achieved 85.00% accuracy, which was reproduced through direct Neo4j execution. A partial alignment with CIDOC CRM was further examined through RDF validation and limited Omeka S metadata exchange. The results demonstrate the feasibility of traceable knowledge integration and structured retrieval for Zisha teapot heritage, while highlighting limitations concerning dataset scale, external generalization, and semantic interoperability.