Reproduce “IBMEA: Exploring Variational Information Bottleneck for Multi-modal Entity Alignment” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Taoyu Su, Jiawei Sheng, Shicheng Wang, Xinghua Zhang, Hongbo Xu, Tingwen Liu
IBMEA (Information Bottleneck for Multi-modal Entity Alignment) introduces a variational framework that leverages the Information Bottleneck principle to address misleading information in multi-modal entity alignment. The method consistently outperforms state-of-the-art baselines by 5-10% in Hits@1 and MRR across various datasets, demonstrating particular robustness in low-resource and noisy data environments.
Key insights
- The Information Bottleneck principle provides a principled way to identify and suppress alignment-irrelevant information in multi-modal data.
- Variational encoders enable the learning of probabilistic representations that can be regularized to retain only information predictive of entity alignment.
- Regularizing both modal-specific and modal-hybrid representations through contrastive learning improves robustness, especially in low-resource and noisy conditions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.