Analyzing How BERT Performs Entity Matching

State-of-the-art Entity Matching (EM) approaches rely on transformer architectures, such as BERT, for generating highly contextualized embeddings of terms. The embeddings are then used to predict whether pairs of entity descriptions refer to the same real-world entity. BERT-based EM models demonstrated to be effective, but act as black-boxes for the users, who have limited insight into the motivations behind their decisions. In this paper, we perform a multi-facet analysis of the components of pre-trained and fine-tuned BERT architectures applied to an EM task. The main findings resulting from our extensive experimental evaluation are (1) the fine-tuning process applied to the EM task mainly modifies the last layers of the BERT components, but in a different way on tokens belonging to descriptions of matching / non-matching entities; (2) the special structure of the EM datasets, where records are pairs of entity descriptions is recognized by BERT; (3) the pair-wise semantic similarity of tokens is not a key knowledge exploited by BERT-based EM models.