Does entity linking always use Wikipedia?

No. Wikipedia is a common target, but a specialist or internal knowledge base can also be used.

Can two differently written mentions link to one record?

Yes, if they refer to the same represented entity.

Does a high model score prove the link is correct?

No. Scores require interpretation and evaluation, and the correct candidate may have been omitted before ranking.

AI Glossary / Natural language and speech / Intermediate

Entity linking

Definition

Entity linking is the task of connecting a mention in text to the appropriate entity record in a knowledge base. It resolves which particular person, organization, place, or other represented entity a mention refers to, usually returning a stable identifier rather than only a name or category.

What it means

Names are often ambiguous. “Mercury” might refer to a planet, an element, or an organization, depending on context and the knowledge base being used. Different names can also refer to one entity, such as a company’s full name and an accepted abbreviation.
Entity linking turns a textual reference into a more specific connection. That connection can help software retrieve descriptions or combine information across documents.
The task depends on the target inventory. A correct link in one database may have no equivalent in another. A mention can also refer to a real entity that the knowledge base does not contain, so forcing every mention onto an existing record can introduce false information.

How it works

A common pipeline identifies a mention, retrieves plausible candidate records, and compares those candidates with the mention’s context. Candidate retrieval may use names, aliases, descriptions, or learned representations.
Disambiguation then chooses among the candidates using relevant evidence. Nearby words, entity types, document context, and relationships between mentions can help, but their usefulness varies by application.
Some systems include a NIL outcome for mentions without a suitable knowledge-base entity. This is different from silently choosing the nearest name. Low confidence and genuine absence also deserve separate consideration: the correct entity might exist even when the model cannot identify it reliably.
Implementations may combine these stages, but the underlying detection, coverage, and selection problems remain useful to distinguish.

Three practical examples

A research index links an author mention to a specific author record, helping separate people who share a name.
A business-news system connects company mentions with an internal organization directory before aggregating coverage.
A museum catalog connects place names in descriptions with geographic records while preserving cases that need curator review.

Think of it this way

Think of a librarian matching a name written on a note to the correct catalog card. The spelling helps, but surrounding details may decide the match. The analogy has a limit: an automated linker can misread context, and the necessary card may be missing from its catalog.

Worked example

Imagine a fictional knowledge base with three records: M1, Mercury the planet; M2, mercury the chemical element; and M3, Mercury Logistics, an invented transport company.
The sentence “Mercury Logistics opened a new distribution depot” contains a company mention. Suppose the candidate retrieval stage returns M1 and M2 but accidentally omits M3.
Even an excellent ranker cannot select M3 from that two-record candidate list. Choosing M1 because it has the highest remaining score would produce an incorrect identity. This is a candidate-coverage failure, not simply a poor final ranking.
Now suppose retrieval includes all three records. The full company name and depot context support M3. The output should retain M3’s identifier and the original mention so a reviewer can trace the decision.
If the sentence instead names an invented “Mercury Gardens Cooperative” absent from the database, the system should not equate it with M3 merely because both names begin with Mercury. Under a task that supports out-of-database mentions, NIL is a relevant outcome.

Illustrative case study

In a fictional project, Larch Research builds a searchable collection of supplier reports. Several unrelated organizations share short names, and older reports use abbreviations.
The team tests entity linking against a maintained supplier directory. Reviewers label ambiguous mentions and distinguish missing directory records from incorrect links to existing suppliers.
The interface displays the linked record alongside the original passage. When evidence is insufficient, an unresolved state invites review rather than presenting a guess as an established supplier identity. This is an illustrative workflow, not a claim of measured accuracy or commercial results.

Common misconceptions and mistakes

Entity linking is not the same as named entity recognition. Recognition can identify an organization span without determining which organization it is.
Coreference resolution is also distinct: it groups references to the same entity in discourse, whereas linking connects a reference to an external inventory. The tasks can support each other.
A matching name is not proof of matching identity. Conversely, different names do not necessarily imply different entities.
Finally, NIL should not automatically be interpreted as “this entity does not exist.” It concerns the target knowledge base and the task’s annotation conventions, which may distinguish several reasons for leaving a mention unlinked.

Business use

Entity linking can improve document navigation, organization-level analysis, catalog integration, and knowledge retrieval. Its value depends on how much the application benefits from reliable identity connections.
Evaluation should inspect mention detection, candidate coverage, correct-link performance, and handling of missing entities. A single overall score can hide errors concentrated among uncommon names, languages, or newly added organizations.
For consequential workflows, retain provenance and allow corrections to the link without changing the source text. Updating the knowledge base may also change what can be linked, so evaluation results should identify the inventory version used.

Check your understanding

A system correctly marks “Mercury Logistics” as an organization but links it to the planet Mercury. Which parts succeeded and failed?

Answer: The organization mention was recognized, but the identity link was wrong. Correct span detection and entity type do not guarantee correct entity linking.

Related terms

Named entity recognition
Knowledge base
Candidate generation
Coreference resolution
entity disambiguation
entity resolution
NIL prediction
knowledge graph

Learn next

Learn named entity recognition to understand mention boundaries, then examine candidate retrieval and ranking separately. Compare entity linking with coreference resolution using one short document. Include an entity missing from the target directory so the exercise tests coverage as well as matching.

Sources

Jurafsky and Martin, Speech and Language Processing, entity-linking discussion: https://web.stanford.edu/~jurafsky/slp3/24.pdf. Ling, Singh, and Weld, “Design Challenges for Entity Linking” (2015), author abstract: https://aclanthology.org/Q15-1023/. Zhu and colleagues, “Learn to Not Link: Exploring NIL Prediction in Entity Linking” (2023), author abstract: https://aclanthology.org/2023.findings-acl.690/. Independent review consulted the two primary author abstracts for pipeline components and NIL taxonomy; no full-paper review is implied by those abstract citations.

Previous: Entity integrity Next: Entity-Relationship Model