Entity resolution

Entity resolution is the task of deciding whether records from one or more sources denote the same real-world thing. Names and addresses are evidence, not identity itself: they can change, be shared or be recorded incorrectly. The task includes both linking true matches and keeping distinct entities apart. It also requires defining the level of identity, such as a legal organization rather than its parent group.

In practice, records are compared using identifiers, names, locations, dates and other contextual attributes. Candidate comparisons may be restricted for efficiency, then reviewed against explicit match criteria. A match decision can be uncertain or time-dependent. Merging records should preserve the original source records and the reason for the identity decision so later evidence can challenge it.

For an organization acting on a record, a false merge can contaminate every associated claim, while a missed match can split the history of one subject. Record linkage is a closely related statistical formulation; entity disambiguation emphasizes selecting the intended referent in context; a canonical identifier is a stable handle for the resolved identity.

Pairwise precision and recall are the usual measures, and they can mislead. Because most pairs are non-matches, accuracy is meaningless; because clusters are built transitively, a single wrong link can merge two large groups, which pairwise counts understate. A fuller evaluation looks at cluster-level measures, at the effect on the downstream decision, and at the errors that matter most to the use. Merging two different companies has different consequences from failing to merge two records of one company, and the threshold should reflect that asymmetry. The article on entity resolution evaluation beyond pairwise F1 covers these measures in detail.

Resolution decisions should be reversible and explainable. A merge that cannot be undone, or that records only that two records were joined and not why, converts an uncertain judgment into a permanent fact. The article on merging or relating organisation records argues for keeping the decision as a link with a stated confidence and basis, and merging for presentation only, so that a correction does not require reconstructing history.

Resolution is the broader goal, of which record linkage is the pairwise core and entity disambiguation is the version that links a mention in text to an entry in a reference list. Its output supports golden records and master data management, and its quality limits every later step of data fusion, since an identity error produces conflicts that are not real.