Named Entity Recognition

Named Entity Recognition finds and classifies people, organizations, locations, dates, money, and other typed spans.

In "Apple released the iPhone in San Francisco on January 9, 2007," NER tags Apple as an organization, iPhone as a product, San Francisco as a location, and January 9, 2007 as a date. The usual framing is sequence labeling: each token gets a tag for entity type and span. BIO marks Beginnings, Inside continuations, and Outside tokens.

Older systems used conditional random fields (CRFs) over hand-built features. Modern NER fine-tunes transformers for token classification and scores high on standard types.

Hard cases include nested entities (an organization name that contains a location), ambiguity (Apple the company versus the fruit), and domain names such as drugs and genes. Cross-lingual NER copies patterns into languages with less labeled data. NER feeds knowledge graphs, summarization, search, and compliance monitoring.

BIO tags turn "San Francisco" into B-LOC I-LOC. A transformer token classifier emits those tags, then a parser stitches spans. Nested NER needs extra structure because one span can sit inside another. Ambiguity needs context: Apple next to iPhone is a company. Drug and gene names rarely appear in general pretraining, so domain models or dictionaries still help.

Cross-lingual transfer reuses the same tag scheme in a language with fewer labels. Downstream, those spans populate knowledge graphs, filters for compliance, and better search facets. Named entity recognition tags people, places, and organizations. CoNLL-2003 is the long-running benchmark.