Constraint-Based Entity Matching

Entity matching is the problem of deciding if two given mentions in the data, such as "Helen Hunt" and "H. M. Hunt", refer to the same real-world entity. Numerous solutions have been developed, but they have not considered in depth the problem of exploiting integrity constraints that frequently exist in the domains. Examples of such constraints include "a mention with age two cannot match a mention with salary 200K" and "if two paper citations match, then their authors are likely to match in the same order". In this paper we describe a probabilistic solution to entity matching that exploits such constraints to improve matching accuracy. At the heart of the solution is a generative model that takes into account the constraints during the generation process, and provides well-defined interpretations of the constraints. We describe a novel combination of EM and relaxation labeling algorithms that efficiently learns the model, thereby matching mentions in an unsupervised way, without the need for annotated training data. Experiments on several real-world domains show that our solution can exploit constraints to significantly improve matching accuracy, by 3-12% F-1, and that the solution scales up to large data sets.