danielquillanroxas/etymograph-unified

Etymograph Unified

A unified etymological knowledge graph built from five publicly available datasets. Contains 22.7 million relations across 6.6 million words in 5,529 languages, covering cognate pairs, derivation chains, borrowing paths, and more.

The goal was to bring together the best open etymology data into a single, queryable SQLite database with a normalized schema — one you can point a graph search at without wrangling five different formats.

What's in the database

Two core tables:

words — 6,577,941 entries

Column	Type	Description
id	INTEGER	Primary key
term	TEXT	The word itself
lang	TEXT	ISO 639-3 language code
lang_name	TEXT	Full language name (where available)
lang_family	TEXT	Language family
gloss	TEXT	Short definition (where available)

Unique constraint on (term, lang). Full-text search index on term.

relations — 22,758,746 entries

Column	Type	Description
id	INTEGER	Primary key
source_word_id	INTEGER	Foreign key → words
target_word_id	INTEGER	Foreign key → words
relation_type	TEXT	One of: cognate, derived_from, inherited_from, borrowed_from, compound_of
confidence	REAL	0–1 confidence score
source_dataset	TEXT	Which dataset this relation came from

Indexed on source_word_id, target_word_id, relation_type, and a compound index on (target_word_id, relation_type).

Relation breakdown

Type	Count	Description
cognate	11,620,617	Words in different languages sharing a common ancestor
derived_from	9,825,071	Word X evolved from word Y (e.g. English water from Middle English water)
inherited_from	654,763	Direct inheritance within a language family
borrowed_from	338,095	Loanwords across language boundaries
compound_of	320,200	Word formed by combining two or more roots

Source datasets

All relation types from the original datasets were normalized into the five types above during the pipeline.

Dataset	Relations	What it covers	Reference
CogNet v2.0	8,117,981	Cognate pairs across 338 languages, automatically extracted from wordnets	Batsuren et al., 2019. CogNet: A Large-Scale Cognate Database. ACL.
Etymological Wordnet	6,031,431	Etymological relationships extracted from Wiktionary	de Melo, 2014. Etymological Wordnet: Tracing the History of Words. LREC.
Etymology Atlas	4,167,968	Wiktionary-derived etymology with typed relations (borrowed, inherited, etc.)	Wikimedia / community contributors
etymology-db	3,797,383	Structured Wiktionary etymology parses with granular relation types	Roher, 2020. github.com/droher/etymology-db
EtymDB 2.1	643,983	Curated etymological links with inheritance, borrowing, and derivation	Fourrier & Sagot, 2020. Methodological Choices in EtymDB 2.1. LT4HALA Workshop.

How it was built

A Python pipeline downloads the raw data from each source, parses each format (TSV, CSV, Parquet, JSON), normalizes language codes to ISO 639-3, normalizes relation types to the five canonical types, deduplicates words by (term, lang), and bulk-loads everything into SQLite with appropriate indices.

Language codes from ISO 639-1 (two-letter) are mapped to 639-3 (three-letter) where applicable. Non-standard codes from individual datasets are mapped as well.

Usage

import sqlite3

db = sqlite3.connect("etymograph.db")
db.row_factory = sqlite3.Row

# Find all words matching "water"
words = db.execute("SELECT * FROM words WHERE term = 'water'").fetchall()

# Get the derivation chain for English "water"
word_id = db.execute("SELECT id FROM words WHERE term = 'water' AND lang = 'eng'").fetchone()["id"]
chain = db.execute("""
    SELECT w1.term as src, w1.lang as src_lang, w2.term as tgt, w2.lang as tgt_lang, r.relation_type
    FROM relations r
    JOIN words w1 ON w1.id = r.source_word_id
    JOIN words w2 ON w2.id = r.target_word_id
    WHERE r.source_word_id = ? AND r.relation_type = 'derived_from'
""", (word_id,)).fetchall()

# Find cognates of a word across languages
cognates = db.execute("""
    SELECT w2.term, w2.lang
    FROM relations r
    JOIN words w2 ON w2.id = r.target_word_id
    WHERE r.source_word_id = ? AND r.relation_type = 'cognate'
""", (word_id,)).fetchall()

File

• etymograph.db — SQLite database, ~3.5 GB

Intended use

Research in historical linguistics, computational etymology, NLP tasks involving word origins, language family analysis, building etymology explorers, or anyone who finds word histories interesting.

Limitations

• Coverage is uneven across languages. European languages (especially English, Latin, French, German) are heavily represented. Many languages have sparse or no data.
• Confidence scores vary by source. CogNet relations are automatically extracted and may contain noise. Wiktionary-derived datasets inherit the quality (and occasional errors) of community edits.
• Some datasets overlap — the same relation may appear from multiple sources. No cross-dataset deduplication is performed at the relation level (though words are deduplicated by term+lang).
• Historical and reconstructed languages use varying conventions for transcription.

License

The unified database is released under CC BY-SA 4.0, consistent with the most restrictive license among the source datasets (Wiktionary-derived data is CC BY-SA).

Individual source datasets have their own licenses:

• CogNet: MIT
• Etymological Wordnet: CC BY-SA 3.0
• Etymology Atlas: CC BY-SA (Wiktionary)
• etymology-db: MIT
• EtymDB 2.1: CC BY-NC-SA 4.0

Citation

If you use this dataset, please cite the underlying sources:

@inproceedings{batsuren2019cognet,
  title={CogNet: A Large-Scale Cognate Database},
  author={Batsuren, Khuyagbaatar and Bella, G{\'a}bor and Giunchiglia, Fausto},
  booktitle={Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics},
  year={2019}
}

@inproceedings{demelo2014etymwn,
  title={Etymological Wordnet: Tracing the History of Words},
  author={de Melo, Gerard},
  booktitle={Proceedings of the 9th Language Resources and Evaluation Conference (LREC)},
  year={2014}
}

@inproceedings{fourrier2020etymdb,
  title={Methodological Choices in the Construction of EtymDB 2.1},
  author={Fourrier, Cl{\'e}mentine and Sagot, Beno{\^i}t},
  booktitle={LT4HALA Workshop at LREC},
  year={2020}
}