named-entity recognition
/ NAYMD EN-tih-tee rek-ug-NISH-un /
Named-entity recognition, or NER, is the task of reading text and highlighting the proper names — the specific people, places, organizations, dates, and amounts — and saying what kind each one is. Run it on "Apple opened a store in Berlin on Tuesday" and it tags "Apple" as an organization, "Berlin" as a location, and "Tuesday" as a date. Think of it as the machine taking a highlighter to a document and color-coding the who, where, and when.
The tricky part is that names are slippery. "Apple" might be a company or a fruit; "Washington" might be a person, a city, or a state; "Jordan" could be a country or a basketball player. Only the surrounding words decide. Because each word's label depends on its neighbors, NER is handled as sequence labeling — the model reads the sentence left to right (and right to left) and assigns each token a tag like "start of an organization" or "not an entity." It learns these patterns from text that humans have painstakingly annotated.
NER quietly powers a lot of everyday software: it pulls names and dates out of resumes, flags companies in financial news, links your search query to the right database entry, and feeds the first step of question-answering systems. Its honest limits are worth noting — it stumbles on names it has never seen, on informal or noisy text, and especially on languages with little annotated data. And it only finds the names; understanding what is said about them is a separate, harder job.
"Tim Cook met Angela Merkel in Berlin last March." NER output: Tim Cook = PERSON, Angela Merkel = PERSON, Berlin = LOCATION, last March = DATE. The verbs and small words are left untagged.
NER highlights the who, where, and when — leaving the rest alone.
NER finds and types names; it does not understand the claims made about them. The same string ("Apple," "Washington") can be different entity types, so context, not a name list, drives the decision.