Does a correct link prove the document is accurate?

No. It identifies a relevant concept or entity. Verifying factual assertions requires separate evidence and review.

What if there is no suitable Wikipedia page?

A system should be able to leave the mention unlinked or mark it as lacking a target under its policy, rather than inventing a correspondence.

AI Glossary / Natural language and speech / Intermediate

Wikification

Definition

Definition: Wikification is the process of identifying relevant terms or mentions in a text and linking them to the appropriate Wikipedia articles. It combines deciding what should receive a link with resolving which article matches the intended meaning in context.

What it means

What it means: A word such as Java can refer to a programming language, an island or something else. Adding the first matching page is not enough. A useful link connects the particular occurrence to the right concept and provides relevant background for the reader.
The original Wikify! research emphasized selecting important words and phrases, including technical concepts and named entities. Later usage often overlaps with entity linking to Wikipedia. These scopes are related but not identical: a system that links every supplied entity mention may not also decide which concepts deserve links in a readable document.

How it works

How it works: A pipeline first detects candidate spans or receives them from another component. It then retrieves possible Wikipedia targets, evaluates their compatibility with the local context and sometimes the rest of the document, and selects an appropriate target or abstains.
Candidate selection and disambiguation create separate failure points. The right article cannot be chosen if it never enters the candidate set. A correctly disambiguated span can still be unhelpful if the concept is trivial or excessive linking distracts the reader. A practical system also records the article identity or version and handles redirects, absent targets and uncertain cases according to a stated policy.

Three practical examples

Three practical examples:
1. An educational article links a technical phrase to an explanatory encyclopedia entry. Editors check both that the link refers to the intended concept and that the destination is suitable background for the audience.
2. A search index associates a mention of Python in a programming tutorial with the programming-language concept, while a wildlife article requires a different interpretation.
3. An enterprise document mentions a private internal project called Cedar. A wikification system should not force a link to a tree article merely because the spelling matches. The appropriate result may be no Wikipedia link, or a separate internal knowledge-base link handled by another system.

Think of it this way

Think of it: Wikification resembles an editor adding carefully chosen cross-references in a textbook. The editor must decide which passages benefit from explanation and which reference actually fits. The analogy is limited because an automated system infers those decisions from data and scoring rules, and a plausible-looking reference can still be wrong.

Worked example

Worked example: Consider three invented snippets with a marked mention. A: “Java compiled the application,” with Java intended as shorthand for the programming technology. B: “The expedition crossed Java,” referring to the island. C: “Our internal Java project closes Friday,” where Java is a private codename with no intended Wikipedia target.
Suppose the agreed annotation policy links A to the programming-language article, B to the island article and leaves C unlinked. A system that always selects the programming-language page produces three links, only one correct. Its link precision is 1/3. Among the two mentions that should be linked, it supplies one correct target, so recall is 1/2 under this defined policy.
A system that links only A correctly has precision 1/1 and recall 1/2. These figures illustrate the coverage-versus-precision trade-off, not a complete quality judgment. The task must define whether span selection, target identity and abstention are scored together or separately.

Illustrative case study

Illustrative case study: Fictional Elm Learning adds explanatory links to introductory technology lessons. The first prototype links every repeated technical term and sometimes chooses a popular article that does not match the lesson's meaning.
Editors introduce a limited linking policy, review ambiguous targets and retain uncertain mentions without links. They also separate suggested background reading from evidence supporting factual claims in the lesson. The team evaluates the annotated pages for both target correctness and reading usefulness. This invented case demonstrates editorial integration; it does not claim that automatic linking guarantees comprehension or that Wikipedia links validate the surrounding text.

Common misconceptions and mistakes

Common misconceptions and mistakes: Wikification is not converting text into wiki markup or publishing it on Wikipedia. It is not merely named-entity recognition, because recognizing a span does not choose its target article. It is not ordinary dictionary sense tagging, although contextual disambiguation is shared. A link to an article is not evidence that every nearby claim is true. Nor should every unfamiliar phrase be forced into an available page. Missing coverage and uncertainty are legitimate outcomes that a system should represent.

Business use

Business use: Wikification can support learning materials, document navigation and concept-based search. Define the target encyclopedia edition, linking policy and acceptable confidence level. Review ambiguous or consequential links before publication, and keep enough provenance to revisit targets when content changes. For internal concepts, use an appropriate internal knowledge base rather than pretending public encyclopedia coverage is complete. Measure end-to-end link correctness as well as the separate quality of candidate detection and target selection.

Check your understanding

Check your understanding: A wildlife article contains Python, and a linker selects the programming-language page because it is the most frequent candidate in its training data. Has the spelling match solved the task?

Answer: No. The intended referent depends on context. Candidate popularity can be useful evidence, but it cannot replace contextual disambiguation.

Related terms

Learn next

choose an ambiguous name and write three contexts with different intended targets, including one without a suitable page. Separate span detection from target selection, then explain when abstaining is preferable to a confident but unsupported link.

Sources

Sources: Mihalcea and Csomai, Wikify! Linking Documents to Encyclopedic Knowledge, 2007, institutional publication record and abstract: https://digital.library.unt.edu/ark:/67531/metadc31001/ . Csomai and Mihalcea, Linking Educational Materials to Encyclopedic Knowledge, 2007, primary description of two-stage wikification: https://web.eecs.umich.edu/~mihalcea/papers/csomai.aied07.pdf . The Java scoring example and Elm Learning case are original. Historical research claims are not presented as current universal performance guarantees.

Previous: Wike's law of low odd primes Next: Wilcoxon signed-rank test