Reproduce “MEAformer: Multi-modal Entity Alignment Transformer for Meta Modality Hybrid” 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
Zhuo Chen, Jiaoyan Chen, Wen Zhang, Lingbing Guo, Yin Fang, Yufeng Huang, Yichi Zhang, Yuxia Geng and 3 more
MEAformer introduces a Transformer-based framework for Multi-modal Entity Alignment (MMEA) that dynamically predicts and fuses modality contributions for each individual entity, moving beyond static, graph-level fusion. This adaptive approach achieved new state-of-the-art results on various MMEA benchmarks, including average gains of 5-6% on Hits@1 for bilingual DBP15K datasets.
Key insights
- Allowing entity-level dynamic fusion captures nuanced characteristics and varying information quality of individual entities, leading to more robust multi-modal representations.
- A Transformer-based cross-modal attention mechanism can effectively infer adaptive modality importance for each entity, moving beyond static, one-size-fits-all fusion.
- Integrating modal-adaptive contrastive learning and hard entity replay boosts performance and stability across various training regimes, including low-resource and unsupervised settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.