Reproduce “Conflict-Aware Pseudo Labeling via Optimal Transport for Entity Alignment” on OpenResearch

About this paper

Qijie Ding, Daokun Zhang, Jie Yin

Conflict-aware Pseudo Labeling via Optimal Transport (CPL-OT) is an entity alignment model that leverages an Optimal Transport mechanism to mitigate alignment conflicts during iterative pseudo-labeling, ensuring one-to-one mappings between entities across knowledge graphs. It consistently achieves superior accuracy across multiple benchmark datasets, demonstrating robust performance even without initial prior alignment seeds.

Key insights

• Explicitly addressing alignment conflicts during the pseudo-labeling phase through Optimal Transport significantly improves the quality and reliability of augmented training data, leading to enhanced overall entity alignment performance.
• Integrating global-level relation aggregation with local-level entity aggregation, supported by a highway gate, produces more robust and discriminative entity embeddings less susceptible to over-smoothing.
• The use of a rectified embedding distance, which combines feature-based similarity with relational neighborhood matching, provides a more accurate basis for identifying potential alignment candidates before conflict resolution.

Summary by alphaXiv.