Kartmaan/french-dictionary

French Dictionary

A ready-to-use offline French language dictionary derived from the French Wiktionary. Available in two formats to suit different use cases: SQLite for desktop applications and real-time querying, and Parquet for data science and machine learning pipelines.

Contains nearly 900,000 distinct word forms including conjugated verb forms, with structured definitions, usage examples, and rich linguistic metadata.

---

Acknowledgements

This dataset would not exist without the foundational work of Franck Sajous, CNRS research engineer and lecturer in the Language Sciences department at the University of Toulouse. He performed the complete parsing of the French Wiktionary into a structured XML resource, WiktionaryX, which this database is derived from.

The original XML resource is available at: 👉 http://redac.univ-tlse2.fr/lexiques/wiktionaryx.html

This dataset is published under CC BY-SA 4.0, in accordance with the license of the source material (French Wiktionary).

---

Available Formats

File	Format	Size	Best suited for
french_dict.db	SQLite 3	~280 MB	Desktop apps, real-time search, SQL queries
french_dict.parquet	Parquet (zstd)	~22 MB	Data science, ML pipelines, Pandas / PyArrow

Both files contain the same data. The format difference reflects a deliberate trade-off between query performance and compression efficiency.

---

Contents

Property	Value
Language	French
Total rows	1,256,143
Distinct word forms	895,090
Includes conjugated forms	✅ Yes
Definitions	✅ Yes
Gender	✅ Yes
Usage examples	✅ Yes
Register tags	✅ Yes (familier, vieux, littéraire…)
Semantic tags	✅ Yes (figuré, par extension…)
Domain tags	✅ Yes (musique, informatique…)
Etymologies	❌ Not included
Translations	❌ Not included

---

SQLite Format (french_dict.db)

Schema

CREATE TABLE mots (
    id          INTEGER PRIMARY KEY AUTOINCREMENT,
    forme       TEXT NOT NULL,       -- lowercase word form (search key)
    pos         TEXT,                -- part of speech
    definitions TEXT NOT NULL        -- JSON-serialized array of definitions
    gender TEXT DEFAULT NULL         -- gender ("m" - "f" - "e" - "NULL")
);

CREATE INDEX idx_forme ON mots(forme);

A word with multiple parts of speech (e.g. lire as both a verb and a noun) appears as multiple rows sharing the same forme. Always query with SELECT * FROM mots WHERE forme = ? to retrieve all senses.

Definition JSON structure

Each row's definitions field contains a JSON array:

[
  {
    "gloss": "Cérémonie ou prestation réservée à un nouvel arrivant...",
    "register": "Familier",
    "semantic": "Figuré",
    "domain": null,
    "exemples": ["Ils lui ont fait un accueil chaleureux."],
    "sous_definitions": [
      {
        "gloss": "Sous-définition optionnelle.",
        "register": null,
        "semantic": null,
        "domain": null,
        "exemples": []
      }
    ]
  }
]

Part-of-speech values

Code	Meaning
N	Noun (Nom)
V	Verb (Verbe)
ADJ	Adjective (Adjectif)
ADV	Adverb (Adverbe)
PRO	Pronoun (Pronom)
DET	Determiner (Déterminant)
PRE	Preposition (Préposition)
CON	Conjunction (Conjonction)
INT	Interjection

Usage example

import sqlite3, json

conn = sqlite3.connect("french_dict.db")
conn.row_factory = sqlite3.Row

# Exact search
rows = conn.execute(
    "SELECT pos, definitions FROM mots WHERE forme = ?", ("lire",)
).fetchall()
for row in rows:
    defs = json.loads(row["definitions"])
    print(f"[{row['pos']}] {defs[0]['gloss']}")

# Prefix-based suggestions
candidates = conn.execute(
    "SELECT DISTINCT forme FROM mots WHERE forme LIKE ? LIMIT 300",
    ("dict%",)
).fetchall()
print([r["forme"] for r in candidates])

---

Parquet Format (french_dict.parquet)

The Parquet file flattens the nested JSON structure into one row per definition, making each field directly accessible as a column. Sub-definitions are included as separate rows with a non-null sub_index.

Schema

Column	Type	Description
forme	string	Lowercase word form
pos	category	Part of speech (N, V, ADJ…)
gender	category	Gender of the noun (m, f, e, NULL)
def_index	int16	Definition index within the lexeme (1-based)
sub_index	Int16	Sub-definition index (null for top-level definitions)
gloss	string	Definition text
register	category	Usage register (Familier, Littéraire, vieux…)
semantic	category	Semantic annotation (Figuré, Par extension…)
domain	category	Subject domain (Musique, Informatique…)
examples	string	Usage examples joined by |
has_sub	bool	Whether this definition has sub-definitions

Categorical columns (pos, register, semantic, domain) use dictionary encoding — each unique value is stored once and referenced by an integer, which contributes significantly to the compression ratio (270 MB → 22 MB with zstd).

Usage example

import pandas as pd

# Load full dataset
df = pd.read_parquet("french_dict.parquet")

# Load specific columns only (columnar read — very fast)
df = pd.read_parquet(
    "french_dict.parquet",
    columns=["forme", "pos", "gloss", "register"]
)

# Filter nouns with a familiar register
nouns = df[(df["pos"] == "N") & (df["register"] == "Familier")]

# Load directly from Hugging Face
from datasets import load_dataset
ds = load_dataset("Kartmaan/french-dictionary")
df = ds["train"].to_pandas()

---

How This Dataset Was Built

Source

The raw data comes from WiktionaryX, a structured XML dump of the French Wiktionary produced by Franck Sajous at the CNRS / University of Toulouse.

Parsing pipeline

french_dict.xml (341 MB)
        │
        ▼  xml_to_json.py   — iterparse streaming, extracts gloss/examples/metadata
        │
french_dict.json (265 MB)
        │
        ▼  json_to_sqlite.py — batch inserts (5,000 rows/batch), VACUUM + ANALYZE
        │
french_dict.db (270 MB)  ✅
        │
        ▼  db_to_parquet.py  — flattens JSON definitions, zstd compression
        │
french_dict.parquet (22 MB)  ✅

Note : The addition of grammatical genders to the french_dict.db file was achieved through a data merging process with a French lexical database. This reference dataset was retrieved from lexique.org and originates from the OpenLexicon project, which is distributed under the CC BY-SA 4.0 LICENSE.

Key implementation choices:

• iterparse with elem.clear() after each <entry> to keep memory usage flat
• Attributes register, semantic, and domain preserved for rich downstream filtering
• B-tree index on forme for sub-millisecond exact lookups in SQLite
• Categorical encoding in Parquet for columns with low cardinality

---

Projects Using This Dataset

• Lexika-fr — French language desktop tools (dictionary, lexicon, quiz, analyzer) 👉 https://github.com/Kartmaan/lexika-fr

---

License

CC BY-SA 4.0 — derived from the French Wiktionary via WiktionaryX. You are free to use, share, and adapt this dataset for any purpose, provided you give appropriate credit and distribute any derivative works under the same license.

• Original source: French Wiktionary - https://fr.wiktionary.org
• WiktionaryX parser: Franck Sajous - http://redac.univ-tlse2.fr/lexiques/wiktionaryx.html
• OpenLexicon (for adding the gender of nouns) - https://openlexicon.fr/