Attention-Enhanced Large Language Model-Guided Semantic Fusion for Representation Learning in Historical Cultural Knowledge Graphs

Historical-cultural collections contain heterogeneous descriptions that are difficult to compare as isolated records. This protocol organizes these descriptions into evidence-supported knowledge graphs and combines structural and textual representations for relation-learning tasks. The workflow comprises corpus selection and provenance recording, domain ontology design, text normalization, candidate entity detection, schema-constrained large language model (LLM) extraction, evidence checking, entity alignment, and graph export. Generated descriptions and local subgraph summaries are constrained by source evidence and encoded with the BGE-m3 text embedding model. A relation-aware graph attention encoder provides structural representations, which are projected into a shared space and combined with semantic representations through a relation-conditioned gate. Link prediction and relation classification provide the principal quantitative evaluations. The Chinese historical-cultural collection illustrates the protocol with 6,744 normalized entities and 46,284 retained triples from 10,998 text entries. The fusion model achieves a mean reciprocal rank of 0.624, Hits@10 of 0.828, and a relation-classification macro-F1 of 0.865. Baseline and ablation comparisons indicate complementary contributions from structural and semantic information. Application to another collection involves language-appropriate preprocessing, a domain-specific ontology, documented alignment decisions, and evaluation that controls shared-source information across data partitions.