Definition and scope of named entities

Summary

In named-entity recognition (NER), the term “named entity” restricts the task to entities that are consistently referred to by one or more strings (words or phrases). This is closely related to the philosophical idea of rigid designators, but in practice NER is applied to many names and referents that are not strictly rigid. NER typically focuses on proper names and other referential expressions (e.g., certain species/substances), while excluding pronouns and descriptive phrases that identify referents by properties, as well as names for kinds of things rather than individuals. The scope of NER is often operationalized by splitting the problem into two parts: (1) detection of entity name spans and (2) classification of those spans into an entity type using an ontology (e.g., person, organization, location). Detection is commonly treated as a segmentation/chunking problem where entity mentions are contiguous, non-nested token spans (e.g., “Bank of America” as one span). Classification may also include temporal and numerical expressions (years, money, percentages), though the “named entity” definition becomes looser for practical reasons because some such expressions can be ambiguous in scope (e.g., “June” without a specified year).

Key Points

• 1

  “Named entity” refers to referents that are expressed by stable strings, but the definition is not strict in real-world NER practice.

• 2

  NER is commonly scoped as two subproblems: detecting contiguous, non-nested mention spans and classifying them into predefined entity types using an ontology.

• 3

  Temporal and numerical expressions may be included as named entities, even though their referential scope can be ambiguous, leading to a practical (loosened) definition.