An information extraction technique that locates spans of text referring to entities (such as persons, locations, organizations, dates, monetary amounts) and assigns each span a category label; implemented with rule‑based patterns, statistical models or machine learning, and producing annotated spans that may require further disambiguation or linking to canonical identifiers.

Principle

NER performs surface‑level classification: it identifies textual mentions and assigns types using lexical, syntactic and contextual cues, but it does not by itself resolve identity ambiguities (entity linking) or coreference across mentions without additional processing.

Demonstration

Illustrative scenario → Processing article abstracts: the system tags 'Marie Curie' as PERSON, 'Paris' as LOCATION and 'Institut du Radium' as ORGANIZATION. Action → tagged spans are used to index documents and generate candidate links to authority records; ambiguous cases are flagged for disambiguation. Consequence → improved faceted search and metadata extraction, with known limits where labels are ambiguous or missing.

Misapplication

Treating NER labels as authoritative identities (for example, assuming every 'Washington' tagged as LOCATION refers to the same geographic entity). The semantic error is conflating surface classification with disambiguated entity identity.

Consequence

NER enables structured indexing, entity‑centric search and downstream analytics; misclassifications or unhandled ambiguities skew statistics, reduce retrieval precision and can mislead automated workflows.

Reversal

In domain‑specific corpora (biomedical, legal, historical), generic NER models often fail because entity categories and surface forms differ; successful application therefore requires schema alignment, domain adaptation or custom annotation. Additionally, languages with limited resources or complex tokenization need adapted models.

Boundary

Clearly within: identifying and labeling mentions in plain text with coarse entity types. Boundary case: nested or overlapping entities where annotation schemas differ. Clearly outside: entity linking/disambiguation and coreference resolution, which are separate but complementary tasks.

Semantic Tension

Precision ↔ Recall — stricter models reduce false positives but miss legitimate mentions; looser models increase coverage at the cost of more spurious labels.

Synthesis

NER is a surface‑level labeling step that makes entity mentions machine‑actionable; to produce reliable, canonical entity representations it should be combined with disambiguation, linking and provenance‑aware validation.