Transformer-based Denoising Adversarial Variational Entity Resolution

Entity resolution (ER), precisely identifying different representations of the same real-world entities, is critical for data integration. The ER question has been studied many years, and methods have proposed to solve it. Although deep learning achieved good performance in tasks, there are some challenges regarding manual labeling model transfer. This paper proposes a novel model, Transformer-based Denoising Adversarial Variational Resolution (TdavER). For entity embedding, we develop an unsupervised embedding based on denoising autoencoders pre-trained language models, which takes corrupted input as training motivate encoder generate rather stable robust high-quality representations. Furthermore, propose feature transformation adversarial variational ease constraints from data. converts low-level embeddings high-level probability distributions, not constrained by source contain similarity features. To better implement transformation, adopt networks optimize autoencoder’s process help it learn correct posterior distribution. Extensive experiments confirms that our TdavER comparable with current state-of-the-art its transferable.