Interpreting deep learning models for entity resolution: an experience report using LIME

Entity Resolution (ER) seeks to understand which records refer to the same entity (e.g., matching products sold on multiple websites). The sheer number of ways humans represent and misrepresent information about real-world entities makes ER a challenging problem. Deep Learning (DL) has provided impressive results in the field of natural language processing, thus recent works started exploring DL approaches to the ER problem, with encouraging results. However, we are still far from understanding why and when these approaches work in the ER setting. We are developing a methodology, Mojito, to produce explainable interpretations of the output of DL models for the ER task. Our methodology is based on LIME, a popular tool for producing prediction explanations for generic classification tasks. In this paper we report our first experiences in interpreting recent DL models for the ER task. Our results demonstrate the importance of explanations in the DL space, and suggest that, when assessing performance of DL algorithms for ER, accuracy alone may not be sufficient to demonstrate generality and reproducibility in a production environment.