Soft Target-Enhanced Matching Framework for Deep Entity Matching

Deep Entity Matching (EM) is one of the core research topics in data integration. Typical existing works construct EM models by training deep neural networks (DNNs) based on samples with onehot labels. However, these sharp supervision signals labels harm generalization models, causing them to overfit and perform badly unseen datasets. To solve this problem, we first propose that challenge a well-generalized model lies achieving compromise between fitting imposing regularization, i.e., bias-variance tradeoff. Then, novel Soft Target-EnhAnced (Steam) framework, which exploits automatically generated soft targets as label-wise regularizers constrain training. Specifically, Steam regards trained previous iteration virtual teacher takes its softened output extra regularizer train current iteration. As such, effectively calibrates obtained model, tradeoff without any additional computational cost. We conduct extensive experiments over open datasets results show our proposed outperforms state-of-the-art approaches terms effectiveness label efficiency.