One article, followed
from ingestion to the observatory.
The same path every article takes: ingested with its metadata, read sentence by sentence, annotated and validated, compared with the rest of the corpus, enriched in both languages, then published.
The same six steps, four times a day
An article enters the corpus from our press archive databases, is read sentence by sentence, annotated, checked against human coding, compared with the rest of the corpus, summarised in both languages, then served to the observatory. Each step opens the chapter that documents it.
What the models attribute to one sentence
The bench isolates a single sentence so its annotations stay legible. In production the same pass covers every sentence of the article, before corpus comparison, event and cascade detection, bilingual enrichment and publication.
The classification bench — one sentence, its annotations. Four sentences from English-language coverage, with the categories and entities the models attributed to them. The four passes light up in the order the chain runs them: tokenising, embeddings, classifiers, entity extraction. The annotations shown are recorded rather than recomputed when the page opens; the word underlined in green is the one the entity extractor kept. Click a number to see another sentence.
The last step is public
Everything the six steps produce is served on the observatory: the day’s article thread and its summaries, the events and cascades detected, the pulse of the frames, the cards of the entities quoted. Chapter 04 lists the sixty-odd annotation categories; chapter 05 gives the validation figures, category by category.