In short. Entity linking uses a textual mention and its context to select an identifiable entity in a knowledge base.

Definition: Entity linking resolves a textual mention, using its surrounding context, to an identifiable entry in a knowledge base. A knowledge base is a collection of entities with information about them; BLINK uses Wikipedia entries as its target entities.

Key facts

The BLINK paper, Scalable Zero-shot Entity Linking with Dense Entity Retrieval, published in November 2020, defines linking output as pairs of mentions and knowledge-base entities.
As of 17 October 2026, the BLINK repository documents an interactive workflow that uses Flair for named entity recognition before linking mentions to Wikipedia.
Google's organization structured data documentation, updated 8 September 2026, says that adding organization information to a home page can help Google distinguish that organization in search results.

How does BLINK resolve a mention?

The BLINK paper separates linking into candidate retrieval and candidate ranking. Candidates are possible knowledge-base entries for the mention. The surrounding words help the model choose between entries with overlapping names.

A bi-encoder converts the mention context and each entity description independently into numerical representations called embeddings. Retrieval selects nearby entity representations. A cross-encoder then reads the mention context together with each retrieved entity's text and reranks the candidates.

The paper's illustration uses “My kids really enjoyed a ride in the Jaguar!” and compares a vehicle brand with a roller coaster. This example shows why the mention's spelling alone leaves an identity choice unresolved.

In the paper's zero-shot setting, the entities used at test time are absent from training, and their textual descriptions supply entity information. Its evaluation assumes each mention has a corresponding entity in the knowledge base; handling mentions outside that base is left to future work.

What does the linking output identify?

Named entity recognition finds mentions in text. Linking assigns those mentions to entries: the BLINK README explicitly describes recognition followed by linking in its interactive mode. That distinction matters when checking an output: detecting a name and selecting its identity are separate operations.

The documented interactive command accepts a setting that displays each entity's Wikipedia URL:

python blink/main_dense.py -i --show_url

The same README explains that the displayed numeric identifier refers to an entity's position in the downloaded entity.jsonl catalogue. That catalogue contains Wikipedia URLs, titles and text. The URL therefore identifies the selected Wikipedia entry, while the displayed identifier belongs to BLINK's catalogue ordering.

What can a website supply about its identity?

Structured data is a standardized format for describing page content. Google's organization guidance gives a practical application of entity disambiguation: home-page organization markup can help distinguish a business from other organizations in search results.

The documentation says that url, the organization's website address, helps Google identify the organization uniquely. It defines legalName as the registered legal name when that differs from name. Its sameAs property accepts links to other websites with additional organization information, including social media or review profiles.

For a business, these fields offer concrete identity information to check against its published details. Google's guidance concerns its organization markup; BLINK documents a research linking system with Wikipedia as its target knowledge base.

FAQ

How can you inspect the entity BLINK selects?

The BLINK README documents --show_url to display the Wikipedia URL for each entity. You can inspect that page to check the selected identity.

Does zero-shot linking require an entity description?

In the BLINK paper's setup, each entity is defined through a short textual description. Zero-shot evaluation separates the training entities from the test entities.

What does Google's organization `sameAs` field contain?

Google's organization documentation describes a URL on another website with additional information about the organization. A business profile on a social media or review site is an example.

Sources

Check how a page reads to answer engines. Open the GEO score analyzer