Wiktionary Structured

Wiktionary parsed into normalized, queryable tables. The raw wiktextract output is a single deeply-nested JSON object per entry; this dataset flattens it into seven relational tables joined on a shared key, so you can pull every IPA transcription, every usage example, or every synonym pair without writing a JSON parser.

Covers 4,868 languages and 10,687,358 entries from the English Wiktionary dump of 2026-07-06.

Quickstart

from datasets import load_dataset
 
# All senses, all languages
senses = load_dataset("jake-anto/wiktionary", "senses", split="train")
 
# English only
senses_en = load_dataset("jake-anto/wiktionary", "senses_en", split="train")

Or query it in place with DuckDB, without downloading anything:

SELECT word, gloss
FROM 'hf://datasets/jake-anto/wiktionary/senses/part=en/*.parquet'
WHERE list_contains(tags, 'archaic')
LIMIT 20;

Dataset Structure

Every table has an entry_id. Tables with sense-level rows also have a sense_idx. Together these join everything.

Config	Rows	One row is
entries	10687358	a word × language × part-of-speech × etymology
senses	12913686	one definition
examples	1517484	one usage example or quotation
translations	3550704	one translation into one language
forms	99094092	one inflected or alternative form
sounds	13201049	one pronunciation (IPA, audio, hyphenation)
linkages	14072346	one synonym/antonym/hypernym/… relation

entries table

Column	Type	Notes
entry_id	int64	join key
word	string	surface form as it appears in Wiktionary
word_norm	string	NFC-normalized, lowercased (great for lookups)
word_fold	string	additionally accent-stripped (NULL outside Latin/Greek/Cyrillic)
lang_code	string	Wiktionary language code (mostly from ISO 639-1 and ISO 639-3, a few from ISO 639-2, and some exceptional ones)
lang	string	language name
pos	string	noun, verb, adj, … (see all possible ones)
etymology_number	string	present when a word has several numbered etymologies
etymology_text	string	etymology section as cleaned prose
categories, topics	list<string>	non-disambiguated
n_senses, n_translations, n_forms	int32	counts, to filter without joining

senses table

Column	Type	Notes
entry_id, sense_idx	int64	composite key
gloss	string	the primary definition
glosses, raw_glosses	list<string>	raw_glosses keeps the parenthesized qualifiers
tags	list<string>	list of qualifiers and tags, like archaic, colloquial, plural, …
topics, categories	list<string>	sense-disambiguated
form_of, alt_of	string	set when this sense is an inflection or variant of another word
n_examples	int32

sounds table

ipa, enpr, audio (file name), ogg_url, mp3_url, audio_ipa, plus homophones, hyphenation and tags (contains context information, like dialect) as lists.

Audio files themselves are not included, you can bulk them download from kaikki.org. Please don't scrape them individually from Wikimedia.

linkages table

All nine relation types in one table with a relation column: synonym, antonym, hypernym, hyponym, holonym, meronym, derived, related, coordinate_term. source_word → target_word.

holonym and meronym are sparse. Wiktionary does not encode them systematically.

Partitioning

Files are partitioned by language: senses/part=en/…, senses/part=de/…, and so on.

Wiktionary has codes for thousands of languages, and most have a handful of entries. Languages below 50,000 entries are pooled into part=misc rather than given a directory each (4,900 directories are a lot!). lang_code remains a real column inside every file, so filtering always works:

-- Works for any language, partitioned or not
SELECT * FROM 'hf://datasets/jake-anto/wiktionary/senses/**/*.parquet'
WHERE lang_code = 'ml';

ds = load_dataset("jake-anto/wiktionary", "senses",
                  split="train", streaming=True)
ml = ds.filter(lambda r: r["lang_code"] == "ml")

Named per-language configs exist for the largest languages only. Everything else uses the pattern above.

Bias, Risks, and Limitations

• 148 byte-identical duplicates were removed.
• word_fold is NULL outside Latin, Greek and Cyrillic. Accent-stripping is meaningless for Han, Devanagari and Arabic script.
• Coverage is deeply uneven. English dominates with a whopping 1,481,699 entries. Many languages have fewer than a hundred entries.
• Quality is Wiktionary's quality.

Licensing

Wiktionary content is dual-licensed CC BY-SA 4.0 and GFDL. This dataset is a derivative work and is released under CC BY-SA 4.0.

Citation

BibTeX:

@misc{ Jake2026 ,
  title  = {Wiktionary Structured},
  author = { Anto, Jake },
  year   = { 2026 },
  url    = {https://huggingface.co/datasets/jake-anto/wiktionary},
  note   = {Derived from en Wiktionary via wiktextract. Dump 2026-07-06.}
}

Dataset Card Authors

• Jake Anto