Integrating BERT-XL with multi-dimensional knowledge graphs for knowledge completion and relation reasoning in...

Area of Science:

• Digital Humanities
• Information Science
• Computer Science

Background:

• Archival fragmented texts present significant knowledge extraction challenges due to deterioration and digitization issues.
• Semantic deficiency, contextual discontinuity, and entity recognition ambiguity hinder effective information retrieval from historical documents.

Purpose of the Study:

• To propose an integrated framework for enhanced knowledge extraction from fragmented archival texts.
• To address challenges of semantic deficiency, contextual discontinuity, and entity recognition ambiguity in archival data.

Main Methods:

• The framework integrates BERT-XL (Bidirectional Encoder Representations from Transformers-XL) with multi-dimensional knowledge graphs.
• It employs a synergistic fusion mechanism with cross-modal attention and gated integration for bidirectional information flow.
• Knowledge completion algorithms utilize temporal, hierarchical, and spatial dimensions for inferring missing entities and attributes.

Main Results:

• The proposed approach significantly improves knowledge completion (Hits@1: 0.578, MRR: 0.648) and relation reasoning (F1: 0.769) compared to baseline methods.
• Multi-hop relation reasoning traverses heterogeneous graph dimensions to uncover implicit connections across document boundaries.
• Experiments on historical archival corpora demonstrate the framework's effectiveness.

Conclusions:

• Integrating contextual language models with multi-dimensional knowledge representations effectively compensates for information scarcity in fragmented archival contexts.
• The framework offers a robust solution for knowledge extraction from challenging archival materials.
• This research advances the field of digital humanities and archival information science.