MixER: linear interpolation of latent space for entity resolution

Abstract Entity resolution, accurately identifying various representations of the same real-world entities, is a crucial part data integration systems. While existing learning-based models can achieve good performance, are extremely dependent on quantity and quality training data. In this paper, MixER model proposed to alleviate these problems. The utilizes our newly designed augmentation method called EMix. EMix map discrete entity records continuous latent space variables (e.g., probability distributions) then linearly interpolate in generate many augmented samples. matching further optimized based strengthen its generalization capability. achieves significant strengths sensitivity experiments when below 50. robustness experiments, presents an absolute performance advantage label noise exceeds 20%. addition, ablation demonstrate that developed effectively improve ability model. overall experimental results prove exhibited excellent over current state-of-the-art methods.