Entity Resolution

Abstract

References to real-world entities are often ambiguous, more commonly across data sources but frequently within a single data source as well. Ambiguities occur due to multiple reasons, such as incorrect data entry, or multiple possible representations of the entities. Given such a collection of ambiguous entity references, the goal of entity resolution is to discover the unique set of underlying entities, and map each reference to its corresponding entity. Resolving such entity ambiguities is necessary for removing redundancy and also for accurate entity-level analysis. This is a common problem that comes up in many different applications and has been studied in different branches of computer science. As evidences for entity resolution, traditional approaches consider pair-wise similarity between references, and many sophisticated similarity measures have been proposed to compare attributes of references. The simplest solution classifies reference pairs with similarity above a threshold as referring to the same entity. More sophisticated solutions use a probabilistic framework for reasoning with the pair-wise probabilities. Recently proposed relational approaches for entity resolution make use of relationships between references when available as additional evidences. Instead of reasoning independently for each pair of references, these approaches reason collectively over related pair-wise decisions over references. One line of work within the relational family uses supervised or unsupervised probabilistic learning using probabilistic graphical models, while another uses more scalable greedy techniques for merging references in a hyper-graph. Beyond improving entity resolution accuracy, such relational approaches yield additional knowledge in the form of relationships between the underlying entities.