State-of-the-art Entity Matching (EM) approaches rely on transformer architectures, such as BERT , for generating highly contex-tualized 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.

State-of-the-art approaches model Entity Matching (EM) as a binary classification problem, where Machine (ML) or Deep Learning (DL) based techniques are applied to evaluate if descriptions of pairs of entities refer to the same real-world instance. Despite these approaches have experimentally demonstrated to achieve high effectiveness, their adoption in real scenarios is limited by the lack of interpretability of their behavior.

The state of the art approaches for performing Entity Matching (EM) rely on machine & deep learning models for inferring pairs of matching / non-matching entities.Although the experimental evaluations demonstrate that these approaches are effective, their adoption in real scenarios is limited by the fact that they are difficult to interpret.Explainable AI systems have been recently proposed for complementing deep learning approaches.Their application to the scenario offered by EM is still new and requires to address the specificity of this task, characterized by particular dataset schemas, describing a pair of entities, and imbalanced classes.This paper introduces Landmark Explanation, a generic and extensible framework that extends the capabilities of a post-hoc perturbation-based explainer over the EM scenario.Landmark Explanation generates perturbations that take advantage of the particular schemas of the EM datasets, thus generating explanations more accurate and more interesting for the users than the ones generated by competing approaches.