Entity Linking and Knowledge Graph Construction › Limitations and Impact

A linked identifier lets an application join a text mention to records elsewhere. A wrong identifier can make the same join carry the wrong attributes into search, question answering, or analytics. Aggregate accuracy alone does not show which entities or contexts suffer those errors. Inspect errors by entity frequency, alias ambiguity, domain, and KB coverage before choosing an acceptance policy.

Coverage can be uneven. Few recorded aliases or a sparse description can leave a linker with less useful evidence, depending on the mention and the other available records. The source corpus's most common reading can also compete with a less common domain-specific reading. For a catalog of small suppliers, for example, evaluate those suppliers and their aliases directly rather than extrapolating a newswire score. Description-based models offer one approach to rare and unseen entities, as in Botha et al. (2020); their availability does not guarantee accurate transfer.

Prior shift is a related risk. An empirical $P(e \mid m)$ summarizes the linked examples used to estimate it. A different deployment distribution can change which reading is most common. If representative linked data are available, compare re-estimated priors and retune the ranking weights on a separate development set. If they are unavailable, measure uncertainty about transfer instead of assuming the old prior describes the new corpus.

Evaluation scope can be mistaken for generality. Belonging to the target KB is expected for an in-KB test entity and is not itself contamination. Report separately whether test documents, supervised entity labels, aliases, or domains overlap training. State what is unseen in a zero-shot experiment. For candidate-constrained ranking on fixed gold in-KB mentions, candidate recall bounds attainable accuracy; ranking and detection need their own measurements.

Missing provenance makes repair harder. Without the supporting source version and span, an investigator may be unable to distinguish a linking error from a relation-extraction error. Other retained logs might still permit repair, so rebuilding is not inevitable. Decide what text, model versions, decisions, and access controls to retain, and account for their storage and privacy costs.

Changing assertions require an update policy. An employment relation may cease to hold even though its old source remains unchanged. Retain applicable validity bounds and recording history when queries need them; mark unknown bounds explicitly. A newer source is not automatically a more accurate one. Re-extraction, correction, and retention policies should preserve the distinction between an assertion being superseded and a source having been wrong.

Entity-aware retrieval can use linked IDs to retrieve records about a referent, while retaining the text that justifies a result. This connects to Retrieval-Augmented Generation. The outline for Knowledge, Memory, Editing, and Unlearning also lists external knowledge and temporal updates. The next chapter, Semantic Parsing and Meaning Representations, considers how to compose a structured interpretation of a request. A structured query can then use linked entities as its arguments, provided the system checks both the interpretation and the links.