This paper introduces SR-HNM, a novel framework for Multi-Modal Entity Alignment (MMEA) that addresses false negatives in hard negative mining and sparse discriminative signals in embedding space regions. SR-HNM uses density-based clustering to identify sparse decision boundary regions and applies hardness-aware weighting to reduce false negative contamination and enhance discriminative power.
Key findings
SR-HNM identifies sparse decision boundary regions through density-based clustering.
Applies principled hardness-aware weighting to reduce false negative contamination.
Derives theoretical bounds on the false negative rate and convergence guarantees.
Achieves state-of-the-art performance, improving Hits@1 by 3.2%–5.7% over existing methods.
Limitations & open questions
The paper does not discuss the scalability of SR-HNM to very large datasets.
The robustness to modality noise is tested under specific conditions, which may not cover all real-world scenarios.