Title: Building Historical Knowledge Graphs Using BERT and Graph Neural Networks

Abstract: The integration of digital humanities methodologies with large-scale historical data analysis has facilitated the transformation of vast quantities of traditional historical texts into structured knowledge graphs. This study presents a comprehensive architectural framework that integrates Bidirectional Encoder Representations from Transformers (BERT) with Graph Neural Networks (GNN) to identify entities and relationships within diverse historical documents. The proposed approach systematically addresses challenges inherent in historical texts, such as linguistic ambiguity, context-dependent references, and the absence of standardized grammatical conventions. Additionally, the research introduces a novel image retrieval system utilizing FastRQNet alongside the pre-trained vision-language model Vilt-qaformer+RoBInet, adhering to the guidelines established in this work. The experimental validation leverages an extensive dataset comprising municipal records, parliamentary papers, and historical correspondence. Comparative analyses demonstrate that the hybrid BERT-GNN model outperforms traditional rule-based methods and other widely used deep-learning baselines in terms of Precision, Recall, and F1-score, as detailed in Table 2. This architecture exhibits robust accuracy and comprehensiveness in managing complex nested structures and implicit references during knowledge graph construction. These findings indicate that merging relational graph learning algorithms with context-aware semantic representation techniques enables the automatic extraction of historical data, thereby enriching the collective knowledge repository.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC