From collection to publication,
six operations, every day.
One processing chain reads the corpus: it collects, annotates sentence by sentence, extracts the people, organisations and places named, groups converging coverage into events, and publishes the result in the live observatory. It runs on today’s news and on nearly fifty years of archives alike.
Our method, in a nutshell
Our methodological steps are presented below in the order they were carried out. The first two ran once, when the models were built; the other four run every day, over breaking coverage and over an archive that reaches back to 1978.
Where the result lands
Every annotated article joins the live observatory as soon as it clears the chain, feeding the daily journal, the media pulse and the provincial map. The real-time processing view shows the chain at work, article by article.
What our method has produced, 1978 to today
Every article is cut into two-sentence units before the models read it.
The reading grain: each label is set on one of these units.
Events, frames, quoted voices, emotions: the grid applied to every unit.
Measured on a thousand hand-coded sentences never seen in training.
What 0.866 means
We set aside one thousand sentences, hand-coded by an expert and never shown to the models during training. On those sentences the models find the right code roughly 87 times out of 100, all categories taken together. That is an average: some categories are recognised almost every time, others far less often. Each category therefore carries its own reliability tier, A, B or C, published with it. The Validation chapter gives them one by one, along with how closely the two annotators who coded those same sentences agreed without seeing each other’s work (0.894).
From labelled sentences to events, waves and voices
Each reading crosses the labelled sentences across the whole corpus. Several outlets converging on the same facts become a dated, typed event; a frame climbing above its historical baseline and staying there becomes a cascade; a person’s measured presence, crossed with the type of voice they carry, becomes a figure of the day; and the arguments about the science itself become a balance sheet, side by side. The What we extract chapter opens the animated demonstration for each of them.
Detected events
Thousands of annotated sentences, condensed into the real news event they describe, typed and summarised in both languages.
See the method Detection method · duration and intensityMedia cascades
When coverage of one frame swells far past its usual level and stays there, the wave is measured over its whole duration.
See the method Detection method · sentence-level rankingDominant voices and figures
Who carries the climate story of the day: the scientist, the decision-maker, the campaigner, ranked sentence by sentence and confirmed by our model.
See the method Reading method · argument by argumentScience under scrutiny
When climate science itself is contested, defended or debated: annotations flag the sentences, then the model reads the article and weighs the arguments one by one.
See the methodThe technical paper and the code
The paper, published in Scientific Data (Nature Portfolio, 2026), sets out how the corpus is built, how the models were trained and how their reading is checked against human annotators. The public repository holds the code that applies them.