Entity Linking

How it works

Entity linking runs in three steps. Named entity recognition finds the spans that name something (people, places, organizations). Candidate generation pulls the possible knowledge-base matches for each span. Disambiguation picks the right one using context. The classic example is the word “Jordan.” In “Jordan closed its border,” the system should resolve to the country; in “Jordan hit the game-winner,” it should resolve to the basketball player. Same string, different entities.

The target is a stable identifier, not just a label. Google references entities with Knowledge Graph IDs like kg:/m/0dl567 (Taylor Swift), and Wikidata uses Q-numbers. Linking a mention to that ID is what turns loose text into a structured fact a machine can reuse. When the mention points to a Wikipedia or Wikidata entry, the task is often called wikification.

This matters because search engines and AI assistants do not read your page as a bag of keywords. They read it as entities and the relationships between them. Entity linking is how a system decides that “the agency” on your about page is the same organization named in a review three sites over, or that your founder is a distinct person from a same-named author. Getting resolved correctly, and consistently, is what earns a knowledge panel, feeds semantic search, and lets question-answering systems cite you by name instead of guessing.