Entity Extraction

Entity extraction, or named-entity recognition (NER), locates the mentions of entities in a text—people, organizations, places, dates, quantities—and classifies them into predefined categories. It combines two operations: delimiting where each mention begins and ends, and deciding what type it is.

What sets it apart from neighboring tasks

It should not be confused with two nearby tasks. Relation extraction goes a step further and looks for the links between already-detected entities. Entity linking connects each mention to a unique entry in a knowledge base, to distinguish, for example, “Paris” the city from “Paris” the person. Entity extraction detects and classifies; linking disambiguates.

From rules to language models

Methods have evolved in stages: rule-based systems and name lists; statistical sequence models such as CRFs; neural networks combining embeddings with BiLSTM-CRF; pre-trained models such as BERT; and, today, large language models guided by instructions and synthetic data. A nuance is in order: the arrival of BERT brought a strong improvement on standard benchmarks—around 92-93 F1 on the English CoNLL-2003 set—but it was incremental over the earlier contextual representations, not an “unprecedented” leap, and it depends heavily on domain and language.

Evaluation, challenges and privacy

Performance is measured with precision, recall and their harmonic mean, the F1 score, on reference sets such as CoNLL-2003 or OntoNotes. The open challenges are multilingualism and specialized domains—biomedical, legal—where the figures drop. And there is a privacy dimension: NER is the basis for detecting and redacting personal data, so a false negative leaves information exposed and a false positive redacts too much; moreover, detection can carry biases depending on the origin of the names.