corto2corto/agora-ngram

Daily word and word-pair counts (unigrams and bigrams) from 37 French online news outlets, up to 30 June 2026. One SQLite database per outlet.

Fréquences quotidiennes des mots et des paires de mots dans 37 médias de la presse française en ligne, jusqu'au 30/06/2026 inclus. Une base SQLite par média.

See the curves directly: agoragram.fr

"covid", "confinement" and "vaccin" in Le Monde, July 2019 – December 2022, on agoragram.fr.

The data is the 37 SQLite databases in bases/. The tables shown at the top of this page are only a preview (see below).

The data: one database per outlet

One SQLite database per outlet, in bases/<media>.db. Words are stored once in token; counts refer to them by id (w1, w2). The same id means the same word in all 37 databases.

Table	Columns	Content
token	id, word	words and their id
unigram	w1, date, n	occurrences of word w1 on day date (YYYYMMDD)
bigram	w1, w2, date, n	occurrences of the pair w1 w2 on day date
total_unigram	date, total	total number of words published that day
total_bigram	date, total	total number of pairs published that day

The frequency of a word on a given day is n ÷ total at the same date. unigram and bigram are indexed by word (w1, then w2, then date): looking up a word is immediate, reading a whole day means scanning the table.

Preview tables (preview/)

These three tables are not the raw data: they are small extracts computed from the databases, to show what they contain and what can be done with them. They are the tables displayed by the Dataset Viewer at the top of this page.

Table	Content
media	one row per outlet: first and last day, number of days, number of words, database size
words_by_month	one row per outlet and month since 2010: words published that month (words_published) and the number of occurrences of ten words or word pairs (climat, covid, ukraine, inflation, retraites, immigration, dissolution, pouvoir d'achat, intelligence artificielle, réchauffement climatique); frequency = occurrences ÷ words_published
one_day	every word published by Mediapart on 8 March 2024, with its number of occurrences and its id in the token table

For instance, in words_by_month, "covid" goes in Le Monde from 0 occurrences in 2019 to 3,334 in March 2020; "climat" doubles in November 2015 (COP21). In one_day, "femmes" appears 79 times on 8 March 2024, against 3 the day before.

Download a single outlet

The whole dataset is 121 GB; you can take one database only (sizes in the media table):

hf download corto2corto/agora-ngram bases/mediapart.db --repo-type dataset --local-dir .

Read a database in Python

Daily frequency of "climat" in Mediapart:

import sqlite3
import pandas as pd

con = sqlite3.connect("bases/mediapart.db")
df = pd.read_sql("""
    SELECT u.date, 1e6 * u.n / t.total AS per_million
    FROM unigram u
    JOIN token w ON w.id = u.w1
    JOIN total_unigram t ON t.date = u.date
    WHERE w.word = 'climat'
    ORDER BY u.date
""", con)

Same thing in SQL for a word pair ("pouvoir d'achat"):

SELECT b.date, 1.0 * b.n / t.total AS frequency
FROM bigram b
JOIN token w1 ON w1.id = b.w1
JOIN token w2 ON w2.id = b.w2
JOIN total_bigram t ON t.date = b.date
WHERE w1.word = 'pouvoir' AND w2.word = 'd''achat'
ORDER BY b.date;

Words are lower-cased, without punctuation; apostrophes stay attached to the following word ("d'achat", "l'europe"). The articles themselves are not distributed.

License

CC BY 4.0 (Creative Commons Attribution 4.0 International).

Citation

If you use this dataset, please cite it:

@misc{agoragram2026,
  author       = {Echikr, Elias},
  title        = {Agoragram: daily word frequencies in the French online press},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/corto2corto/agora-ngram},
  note         = {Website: \url{https://agoragram.fr}}
}

Agoragram extends Gallicagram to online news; please also cite:

@article{decourson2023gallicagram,
  author  = {de Courson, Beno{\^i}t and Azoulay, Benjamin and de Courson, Clara
             and Vanni, Laurent and Brunet, {\'E}tienne},
  title   = {Gallicagram : les archives de presse sous les rotatives
             de la statistique textuelle},
  journal = {Corpus},
  number  = {24},
  year    = {2023},
  doi     = {10.4000/corpus.7944}
}