Named Entity Recognition

Definition

Named entity recognition finds mentions of real-world things in text, such as people, organisations, places and products, and labels their type. It is the first step toward counting what a text is about; linking each mention to one specific identifier is a separate and harder step.

In plain terms

A recogniser reads a sentence and marks spans such as “Acme Corp.” as ORGANISATION. Systems are usually built from labelled examples and judged per entity type with Precision and Recall. Typical errors are wrong boundaries (“Acme” or “Acme Corp.”), wrong types, and names that can mean several things.

Recognition says what kind of thing a span is. Entity linking decides which one, for example which of several places called Paris. Systems that count mentions of a company need both, plus a list of its aliases.

Example

“Jordan said Washington would meet Acme Corp. in Paris.”

Recognition would label four entities; only linking turns them into countable, unambiguous things.

Common misunderstandings

• “NER tells you which entity a name refers to.” It tells you the type; linking resolves the specific one.
• “A list of names is enough.” Lists miss new names and cannot tell a company from a person with the same name.
• “Accuracy carries over between domains.” A model trained on news will often miss product names or medical terms.