Named Entity Recognition
Extract people, places, and organizations from French text with spaCy and CamemBERT: BIO format, subword alignment, entity-level evaluation.
Classifying a document means assigning it a label. Recognizing its named entities means extracting structure from it: who, where, when, which organization, how much. You move from one prediction per document to one prediction per word, and that difference changes everything — the data format, the model architecture, and above all how you measure success.
The use cases are immediately concrete. A firm receives ten thousand letters and wants to know which companies are mentioned in them. A press office wants to track the public figures being quoted. A legal department has to pseudonymize case files before sharing them, which means finding every personal name — a recall of 95% here means that one identity in twenty leaks.
French adds its own difficulties. Nobiliary particles and prepositions are part of names ("Charles de Gaulle", "Jeanne d'Arc"), articles are part of some place names ("Le Havre", "La Rochelle"), institution names are long and partly lowercase ("ministère de l'Économie et des Finances"), and capitalization — the model's strongest signal — disappears entirely in text that is typed quickly or automatically transcribed.