An unsupervised multi-view contrastive learning framework with attention-based reranking strategy for entity...

Area of Science:

• Artificial Intelligence
• Data Science
• Knowledge Representation

Background:

• Entity alignment is vital for integrating diverse knowledge graphs.
• Unsupervised methods are increasingly important due to limited pre-aligned data.
• Existing unsupervised methods lack sufficient guidance for complex entity matching.

Purpose of the Study:

• To develop an effective unsupervised entity alignment framework.
• To address the challenge of aligning entities with similar names and structures.
• To improve the accuracy of entity alignment in real-world scenarios.

Main Methods:

• Proposed AR-Align: an unsupervised multi-view contrastive learning framework.
• Utilized two data augmentation methods for complementary neighborhood and attribute views.
• Implemented an attention-based reranking strategy for hard-to-align entities.

Main Results:

• AR-Align demonstrated superior performance compared to state-of-the-art methods.
• Outperformed both supervised and unsupervised approaches on benchmark datasets.
• Effectively reduced the semantic gap between entity views via contrastive learning.

Conclusions:

• AR-Align offers a robust solution for unsupervised entity alignment.
• The framework successfully handles entities with similar names and structures.
• Achieved significant improvements in entity alignment accuracy.