Six words for one number
Every tone label is wording over a number: the directed-tone score, from −1 (hostile) to +1 (favourable). The six bands are absolute, anchored at zero, so a word means the same thing on every page. On a single deputy's page this number is the same reliability-adjusted score the boards rank on (see How members are ranked); a parliamentary group, with thousands of mentions, is shown as the plain average.
Cold is hostile, warm is favourable. We avoid “hot”, which flips the meaning (“under fire” versus “glowing”). A chip's tooltip always shows the exact number and band — the word never replaces the data.
Two honest questions
The band answers one question — is coverage of this person hostile? A second phrase answers another — are they treated worse than their peers? — by comparing them to the all-member baseline (-0.18), since the press runs negative toward nearly everyone. Both are true; they simply measure different things.
Tone is read in two steps
Tone is read in two steps, so a score reflects tone toward the member, not the mood of the news. Step 1, role: is the member the one acting or speaking, the target of the action, or just mentioned? Step 2, tone given that role: positive, neutral, or negative. This is why the per-member "How the press casts them" panel matters: being a target reads far more negative than speaking.
It matters because a single score blends roles. Across the whole corpus the same position reads very differently by role — being the object of a story is far harsher than being quoted acting in one:
In the one-line summary a role is named only when this member is cast in it significantly more than the typical member (z ≥ 1.64). “Mentioned”, about 74% of all coverage, is the default and is never headlined.
0.657 (CI 0.65–0.67, ordinal)
0.642 (nominal)
Five frames, each with a reliability tier
Frames are the five generic news frames of Semetko & Valkenburg (2000). Each carries its own reliability score — agreement on the majority label — which decides how far to trust it.
Negativity lives in conflict & morality stories
Net tone when a frame is present vs absent.
How loud, and how steady
Visibility bands come from a member's coverage-volume percentile among covered members. Coverage concentrated in a single month is a burst, not standing presence, so it is marked down one band and tagged.
A one-off spike tag flags coverage that nearly all landed in one month; a quiet-since tag flags a member with no recent headlines. 209 of 577 seated deputies never appear at all — absence of coverage is itself data, not missing data.
Six leaderboards, gated for robustness
A member needs at least 30 headline mentions to make any board, enough that the average can't be a fluke (worst-case error ≤ 0.18). Deputies are then ranked on a reliability-adjusted score — an empirical-Bayes shrunk tone that pulls a thinly-covered member toward the all-member average, so a board isn't topped by one or two near-unanimous mentions; well-covered deputies barely move, and the raw average and mention count sit alongside. The same shrinkage governs the targeted and conflict boards. Parliamentary groups are ranked on the plain pair-weighted average instead: each carries thousands of mentions, so shrinkage would shift them under 0.03 and would only blur the differences between groups.
The Most divisive board adds a second gate: only outlets that mention the member at least 16 times count, so a three-mention cell can't manufacture a fake spread (error ≤ 0.25, about one tone band).
The left–right gap
An outlet's L−R gap is its tone toward the left minus its tone toward the right. Positive means it covers the left more warmly — harder on the right; negative is harder on the left. Across outlets this gap tracks ideology closely (ρ = −0.79 across 25 outlets) — the paper's signature result.
On the bias map the morality axis still separates outlets (ρ = +0.57) even though that frame is exploratory per headline: the per-headline noise averages out over hundreds of headlines.
Where an outlet's ideology score comes from
The left–right position of each outlet is FrIdéo, a published outlet-ideology scale for the French press built by fusing nine independent evidence families — ownership, broadcast guest balance, surveys, the inter-outlet hyperlink network, third-party bias ratings, stated editorial orientation, encyclopedic profiling, a distancing lexicon, and Ad Fontes. It is a standardised score with bootstrap confidence intervals, and it places all 29 outlets in this corpus.
Crucially, no FrIdéo evidence family uses the sentiment or framing labels shown on this site, so the correlation above is a measurement against an outside quantity rather than a scale checked against itself. Outlets whose credible interval spans zero are marked, because the scale itself does not treat their side as settled.
This site runs on a rolling corpus that extends past the paper's frozen window, so its figures track the paper's method but will not match its published numbers digit for digit. On the paper's window the same correlation is ρ = −0.83 across 25 outlets.
What moves week to week
As fresh headlines land, the homepage flags movers — newly in the spotlight, sharp tone swings — and each profile shows week-over-week deltas, with the header marked en direct. This layer is indicative: at roughly 80 headlines a week it is too thin to be a research result, and stays a live demo signal rather than a paper finding.
How the sentence is built
Each profile's sentence stitches the same evidence in a fixed order — the visibility word, the tone band, the relative phrase, then a role clause when it clears the significance test. None of it is hand-written: change a number upstream and the sentence changes with it.
What to keep in mind
- Labels are automated. Tone and frames are assigned by a language model reading each headline, so every score is an indicative measurement, not a hand-checked editorial judgement.
- Headlines only. We read the headline, not the article body. The headline is what most readers actually see.
- Frames vary in reliability. Economic (validated) and conflict (usable) carry the weight. Morality and human-interest score low mostly because they are rare, so there is little to measure, not because they are random. Responsibility is the genuine weak spot, with coders disagreeing on about a third of headlines. Exploratory frames are faded (~) wherever they appear.
- Who's missing. 209 seated deputies never appear at all. Absence of coverage is data in itself, not missing data.
Based on the paper
Every measurement on this site operationalises the method introduced in:
Baromètre Médiatique de l’Assemblée: How French News Outlets Cover Deputies of the Seventeenth National Assembly — Headline-Level Directed Sentiment and Framing at Scale
Amr Sobhy · Le French News Lab
Deputy details from Civix, built on the données ouvertes de l’Assemblée nationale.