Asakiri/langdex-dictionaries

LangDex Dictionaries

Digitized public-domain and openly-licensed scanned dictionaries for under-resourced languages, extracted from the Internet Archive and structured into a uniform schema. A companion to the LangDex dataset, kept separate so that its mixed per-dictionary licensing never complicates LangDex's clean CC BY-SA core.

Why a separate repo

Each source dictionary carries its own license (public domain, CC-BY, CC-BY-SA, or CC-BY-NC). Mixing those with LangDex's uniformly CC BY-SA, commercial-friendly core would be confusing and risks license contamination. Here, every artifact is scoped and licensed per-dictionary, and nothing is commingled with LangDex.

The license of each dictionary is the license of its source. Check the scope and rights columns in the catalog and the per-file provenance before use. In particular, scope = 'nc' content is NonCommercial and must not be used commercially.

Contents

• catalog.parquet — index of scanned dictionaries on archive.org for under-resourced languages, resolved to Glottolog codes and joined against LangDex coverage. Columns include source_id, title, year, glottocode, language, family, licenseurl, scope (open / nc / locked), rights, langdex_lexemes.
• entries/<source_id>.parquet — a digitized dictionary (source of truth): glottocode, gloss_lang, glosses[], forms[], pos, ipa, confidence, extract_model, source_id.
• ocr/<source_id>.parquet — layout-preserving OCR per page (dots.ocr), so entries can be re-extracted without re-OCR.

Scope tiers

scope	meaning	usable
open	public domain / CC-BY / CC-BY-SA	yes (incl. commercial)
nc	CC-BY-NC(-SA)	non-commercial only, opt-in
locked	in-copyright / no-derivatives / unknown	no — reference only

Relationship to LangDex

The target-language forms in each entry become lexemes linked to the shared LangDex concept hub (the pivot-free translation graph). For open dictionaries this connection can be materialized into LangDex; nc dictionaries stay here as a standalone, non-commercial layer.