Archival fragmented texts pose considerable challenges for knowledge extraction owing to semantic deficiency, contextual discontinuity, and entity recognition ambiguity, all arising from document deterioration and incomplete digitization. This paper proposes an integrated framework that combines BERT-XL with multi-dimensional knowledge graphs to tackle these challenges in archival text processing. The framework employs a synergistic fusion mechanism incorporating cross-modal attention and gated integration to enable bidirectional information flow between neural text representations and structured graph knowledge. We design knowledge completion algorithms that draw on multi-dimensional graph structures encompassing temporal, hierarchical, and spatial dimensions to infer missing entities and attributes. Additionally, we develop multi-hop relation reasoning methods that traverse heterogeneous graph dimensions to discover implicit connections across document boundaries. Experiments conducted on historical archival corpora from three provincial archives demonstrate that the proposed approach achieves substantial improvements over baseline methods, with Hits@1 of 0.578 and MRR of 0.648 for knowledge completion, and F1 score of 0.769 for relation reasoning. The results confirm that integrating contextual language models with multi-dimensional knowledge representations effectively compensates for information scarcity in fragmented archival contexts.
Zhenghan Li (Mon,) studied this question.