Learnable Convolutional Attention Network for Unsupervised Knowledge Graph Entity Alignment.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Machine Learning

Background:

• Supervised entity alignment (EA) relies heavily on costly labeled data, limiting practical applications.
• Existing unsupervised EA methods struggle with high complexity or balancing effectiveness and practicality.
• Lack of labeled data presents a significant performance bottleneck in current EA tasks.

Purpose of the Study:

• To propose a novel, efficient, and practical unsupervised entity alignment method.
• To overcome the limitations of existing unsupervised EA approaches, including complexity and performance.
• To improve the accuracy and scalability of entity alignment in scenarios with limited labeled data.

Main Methods:

• Introduced LCA-UEA, a learnable convolutional attention network for unsupervised entity alignment.
• Employed convolution operations before attention to capture structural information and prevent redundant data.
• Developed a relation structure reconstruction method and a consistency-based similarity function to enhance alignment.

Main Results:

• LCA-UEA demonstrated superior performance across diverse datasets (cross-lingual and monolingual).
• The proposed method significantly improved alignment accuracy, outperforming 25 existing supervised and unsupervised methods.
• Achieved a 6.4% improvement in Hits@1 over the best baseline in the best-case scenario.

Conclusions:

• LCA-UEA offers an effective and scalable solution for unsupervised entity alignment.
• The network's design efficiently processes structural information and filters noise, enhancing usability.
• Experimental results validate the superiority of LCA-UEA in improving alignment accuracy.