Exploring the Design Space of Unsupervised Blocking with Pre-trained Language Models in Entity Resolution.

Entity resolution (ER) finds records that refer to the same entities in the real world. Blocking is an important task in ER, filtering out unnecessary comparisons and speeding up ER. Blocking is usually an unsupervised task. In this paper, we develop an unsupervised blocking framework based on pre-trained language models (B-PLM). B-PLM exploits the powerful linguistic expressiveness of the pre-trained language models. A design space for B-PLM contains two steps. (1) The Record Embedding step generates record embeddings with pre-trained language models like BERT and Sentence-BERT. (2) The Block Generation step generates blocks with clustering algorithms and similarity search methods. We explore multiple combinations in above two dimensions of B-PLM. We evaluate B-PLM on six datasets (Structured + dirty, and Textual). The B-PLM is superior to previous deep learning methods in textual and dirty datasets. We perform sufficient experiments to compare and analyze different combinations of record embedding and block generation. Finally, we recommend some good combinations in B-PLM.