NuBerea/distributional-lexicon

Distributional Sense Lexicon — Greek (archaic → Byzantine) + Latin + Hebrew

A corpus-driven sense lexicon of ancient Greek, Latin, and Biblical Hebrew that represents word meaning directly as the distribution of attested usage — graded sense membership, frequency mass, collocational profile, and diachronic drift — scaffolded on the Louw–Nida sense taxonomy. The Greek arm spans the full diachronic range from archaic/classical literature through the New Testament, Septuagint, Hellenistic-Jewish writings, and the patristic corpus into the Byzantine era; the Latin arm applies the same method to the Vulgate and patristic Latin; the Hebrew arm covers Biblical Hebrew on the Masoretic spine with Dead Sea Scrolls, Samaritan Pentateuch, and Aleppo witnesses. The three languages are joined on a shared Louw–Nida pivot for cross-language sense alignment.

The lexicon is calibrated and honest about uncertainty: fine intra-domain sense distinctions are hard for any current method, and the shipped tables carry per-occurrence confidence and abstention rather than false precision. Only openly licensed lexica ship in full; copyrighted reference works are represented by coverage facts and "see s.v." pointers only, never by their prose.

Full specification is gated. The complete dataset card — per-config schemas, join contracts, counts, methodology, and build provenance — lives in SPEC.md, available to approved users behind the repository gate. This public card carries the concept summary only.

Configs. This repo carries 82 live HF dataset configs across the greek_* (28), latin_* (24), hebrew_* (24), and unprefixed crosslang_* (3) namespaces plus 3 unprefixed shared-layer configs (concept_instances, pl_tll_pointer_bridge, sch_lemma_lexicon_bridge); see configs: above for the config-name list (SPEC.md carries the full per-config schema, row counts, and license notes).

License and attribution

Released under CC BY 4.0.

Key upstream sources and their licenses:

• MACULA Greek & MACULA Hebrew (Clear Bible) — https://github.com/Clear-Bible — CC BY 4.0 (gold Louw–Nida senses; attribution required)
• STEPBible data (Tyndale House) — https://github.com/STEPBible/STEPBible-Data — CC BY 4.0
• LSJ, Abbott-Smith, Lewis & Short, Forcellini, Du Cange, BDB, Jastrow — public domain lexica
• SDBH (UBS Semantic Dictionary of Biblical Hebrew) — CC BY-SA 4.0 © United Bible Societies — its content ships only in the share-alike companion repo NuBerea/sdbh, never in this repo
• Copyrighted lexica (BDAG, Lampe, HALOT, TLL) — no content reproduced here; coverage facts and pointers only

EU sui-generis database-right consideration applies on redistribution.

Lampe/TLL pointer-facts union (lexicon_sidecar_bridge_all)

14,244,039 rows in three rails: lampe 3,080; tll 2,580,908; pg_lampe 11,660,051. Each component retains its original source license strings under license, tll__license or pg_lampe__license. Coverage flags, counts, headword keys and citation-template pointers ship; Lampe/TLL prose does not ship. Original member tables and their attribution remain intact. PG region-lemma anchors preserve volume_id, page, region_idx and anchor; no verse projection is asserted. TLL lemma grain and the Greek headword grain remain unchanged.

Native analytics union (analytics_all)

analytics_all adds 423,831 additive native rows in seven facets. Existing member configs remain published. Its row count includes the same member content and is not a count of unique corpus items. The physical file has 90 fields, lexical match_grain = segment, and osisRef is NULL for every row. It provides computed lexical usage, sense attribution and coverage records; it does not provide Bible verse lookup or printed dictionary adjudication. frac_axis states each native denominator; fraction axes are not comparable across language/source arms.

Every member field, native type, value, order and multiplicity is preserved through the reversible COLUMN_MAP.json. Conflicting names receive arm prefixes. Native match_grain is retained as native_match_grain where present. PROOFS/FULL_BUILD_VERIFICATION.json and PROOFS/ORCHESTRATION_VERIFICATION.json record full native reconstruction and physical output checks.

Licensing and prose boundaries stay component-specific. Public-domain BDB text retains its existing source/revision/licence provenance. TLL, BDAG, Lampe and other restricted estates retain their existing fact/pointer/coverage boundaries; no restricted dictionary prose or omitted lexical reference fold is introduced. README metadata is not a grant for all component estates. Prior attribution, share-alike source homes, exclusions and paper holds remain binding.

Facet	Source config	Rows	Immutable source revision
clusters	clusters_all	1,309	d2d80aa05b2d8fa1473954fd1632c3048881c996
hallmarks	cluster_hallmarks_all	3	d2d80aa05b2d8fa1473954fd1632c3048881c996
lemma_collocations	lemma_collocations_all	28,608	d2d80aa05b2d8fa1473954fd1632c3048881c996
sense_collocations	sense_collocations_all	66,049	d2d80aa05b2d8fa1473954fd1632c3048881c996
sense_cell	sense_cell_all	82,902	d2d80aa05b2d8fa1473954fd1632c3048881c996
coverage_apparatus	lexicon_coverage_apparatus_all	53,116	d2d80aa05b2d8fa1473954fd1632c3048881c996
dist_by_source	sense_dist_by_source_all	191,844	d2d80aa05b2d8fa1473954fd1632c3048881c996

lemma_resolver v2: word index

lemma_resolver now has two row kinds in the same table (row_kind). english_term rows are the 10,334,076 rows of the earlier release, unchanged in all 27 earlier columns. word_index rows (1,463,368) list every lemma of the Septuagint (449,219 rows), the SBLGNT New Testament (110,006), the Hebrew OT (344,632) and the Vulgate (559,511) per verse, so a word can be found by lemma, stem, transliteration or Strong's number without an English term. Eight columns are new: row_kind, lemma_norm, form_norm, translit, translit_norm, link_method, link_source, evidence_class. The two files are train-term-00000.parquet and train-word-00000.parquet. Peshitta and Targum have no lemma in their source tables and get no word-index rows. Licence is per component: each row carries its own k (arm of origin); LXX strong_id on word-index rows is a numeric key from STEPBible tbesg (CC BY 4.0, STEPBible / Tyndale House); no tbesg text ships.