.png){kind=icon}
.png){kind=icon}
Return
Unsupervised temporal knowledge graph entity alignment via BERT and structure adaptation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2026.131860.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
With the widespread application of knowledge graphs (KGs) in artificial intelligence, temporal knowledge graphs (TKGs) have attracted increasing attention for their ability to capture dynamic interactions among entities. Accurate entity alignment (EA) across TKGs is crucial for effective knowledge integration and reasoning. However, most existing EA approaches still heavily rely on pre-aligned seed pairs, which are expensive to obtain and limit the scalability of models in practical applications. Moreover, directly incorporating structural features across heterogeneous TKGs may introduce noise rather than informative alignment signals. To address these challenges, we propose BSAUEA (Unsupervised Temporal Knowledge Graph Entity Alignment via BERT and Structure Adaptation), which combines BERT-based contextual embeddings with a structure adaptation mechanism to selectively utilize structural information. In addition, BSAUEA introduces an unsupervised seed generation module that leverages entity name semantics and temporal co-occurrence patterns to automatically identify high-confidence alignment pairs. Compared to existing methods, BSAUEA significantly improves robustness and scalability in unsupervised temporal knowledge graph alignment. Experiments show Hits\@1 of 97.5% and 98.3% on homogeneous datasets (DICEWS, YAGO-WIKI), outperforming baselines by 1.5% and 0.2%, respectively; and 70.9% and 82.7% on heterogeneous datasets (ICEWS-WIKI, ICEWS-YAGO), approaching supervised state-of-the-art. The unsupervised version maintains high accuracy without aligned seeds, and its seed generation module is compatible with other methods, demonstrating strong potential for practical applications.
Keywords:
BERT
structure adaptation
entity alignment
temporal knowledge graphs
unsupervised learning
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W
Organization
No organization information available