NuBerea/distributional-lexicon

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

An in-house, corpus-driven sense model of Greek that represents word meaning directly as the distribution of attested usage — graded sense membership, frequency mass, collocational profile, diachronic drift, and emergent macro-structure — scaffolded on Louw–Nida gold senses (MACULA). v2 (2026-06): the full 7-corpus diachronic span — NT + LXX + Philo/Josephus + Pseudepigrapha + patristic + classical (Perseus∪First1K) — 13,836,557 sense-resolved occurrences across BCE → VI+ CE. BDAG is a check surface it agrees with, never a source.

Reproducibility

Every config is a deterministic CPU function of the HF source corpora + the irreducible primitives, regenerated by scripts/build_greek.py (python scripts/build_greek.py --out data --verify; Latin by scripts/build_latin.py, Hebrew by scripts/hebrew/build_hebrew.py + its build_hebrew_* sibling builders). The irreducible inputs live in NuBerea/distributional-lexicon-primitives: the GPU-trained sense readouts, curated rule tables, and the frozen token_membership L0 chain — now the full 7-corpus, 13.84M-occurrence projection (GPU + hardware-float-sensitive, so it ships there as a primitive with its scripts, not reproduced in HF). The patristic L0 carries a per-occurrence token_index (segment-level source position), enabling token-precise quote attribution downstream.

Configs — naming convention (greek_ / latin_ / crosslang_)

The unified j-repo is three parallel language namespaces plus a shared layer: greek_* (28 configs, this card's focus), latin_* (24 configs — same method + schema on Vulgate + patristic Latin), hebrew_* (15 configs — Biblical Hebrew MT/WLC + DSS/SP/Aleppo witnesses), and unprefixed crosslang_* for Greek↔Latin↔Hebrew alignment on the shared Louw–Nida pivot. The SDBH share-alike sidecars live in the companion repo NuBerea/sdbh (cc-by-sa-4.0 © UBS). A base name denotes the same deliverable in any language — greek_sense_cell ↔ latin_sense_cell ↔ hebrew_sense_cell. Primitives mirror it: greek/, latin/, and hebrew/ subtrees in distributional-lexicon-primitives. A base-name slot that's empty on one side simply isn't built there yet (parity is the naming + tree, not forced content).

Column nomenclature. Latin and Hebrew share one canonical column vocabulary for the same concepts — sense_cluster, silver_ln (primary silver Louw–Nida), mean_conf (mean per-occurrence confidence), n_occ, layer, silver_scope, consensus_k/tuned_k/delta_pre. Greek deliberately uses a distinct soft-mass vocabulary (ln_code gold-LN, soft_mass/hard_mass/conf_mass, soft_frac, per_source/per_time): its sense model is graded-distributional rather than hard-cluster, so the columns differ by data model, not by naming drift. Per-lexicon and per-source columns differ by design — each arm keys its own lexica (Greek LSJ/Abbott-Smith/BDAG, Latin L&S/Forcellini/Du Cange/TLL, Hebrew BDB/Jastrow/SDBH/HALOT) and its own corpus sources. At the occurrence grain the L0 primitive uses silver_ln_top/confidence (per-occurrence argmax LN + confidence), which roll up to silver_ln/mean_conf in the cell-grain configs. (Hebrew's silver_ln_top→silver_ln + confidence→mean_conf were harmonized to this canon 2026-06-25.)

base config (greek_*, mirrored as latin_*)	grain	what
gold_cells	(lemma × LN)	NT gold sense inventory from MACULA (Greek only)
sense_cell, sense_cell_enriched	(lemma × LN)	graded soft-mass + per-source/per-era + exemplars + own-gloss; the live Greek sense_cell is a backward-compatible serving superset with reference prose and source provenance
unit_cell_mass	(work × cell)	per-work sense-usage mass (clustering input)
sense_collocations, lemma_collocations	cell / lemma	PMI collocation profiles
clusters, cluster_hallmarks	work	emergent macro-units (classical → patristic)
sense_dist_by_era, sense_dist_by_source	cell × era/corpus	normalized sense distributions (BCE→VI+)
drift_ranking	lemma	diachronic sense-drift (Jensen-Shannon)
drift_quote_purified	lemma	drift with patristic scripture-quotations removed at token grain (js_all vs js_ownvoice + delta)
source_distance	corpus × corpus	per-word sense agreement
function_freq, prep_entry, particle_entry, grammatical_entry	—	function-word entries (SNACS / PDTB / grammatical)
token_apparatus	occurrence (greek 13.84M)	per-occurrence preferred sense + ranked alternatives + per-witness support + calibrated confidence + era + token_index; the critical-apparatus core
lemma_apparatus, apparatus_crosswalk, witness_genealogy	lemma × witness	LSJ + Abbott-Smith (open) sense crosswalk + classical→Koine genealogy
apparatus_crosswalk_sidecar, lampe_sidecar	lemma	copyrighted-lexicon sidecars (parity with Latin TLL) — BDAG: cover+sense-count+ARI/delta+'see BDAG s.v.'+license; Lampe: cover+'see Lampe s.v.'+license (no count, v0.1 parse). Coverage/metric/pointer ONLY, never prose.
lexica_prose, lexica_prose_as_sidecar	lemma	legacy modeled-inventory projection of LSJ + Abbott-Smith prose; not a complete LSJ inventory and not the live serving surface
inventory_tuned	(lemma × LN)	S4 consensus-tuned LN granularity (merge view; fine codes kept as substructure)
quote_projection	occurrence (greek 170,839)	verbatim-quotation occurrences carrying cited NT/OT gold LN as provenance + inherited↔distributional divergence + token_index
crosslang_sense_alignment (unprefixed)	LN cell	Greek↔Latin↔Hebrew lemma alignment on the shared LN pivot (aligned = n_langs ≥ 2; 1,906 codes carry Hebrew, 637 full three-way). SDBH crossing via an ln_code join to NuBerea/sdbh::sdbh_to_ln, not embedded here

Language-specific extras: Latin adds latin_lexica_apparatus (per-lexicon coverage), latin_lexica_concordance (sense-count comparison), latin_sense_apparatus, and the Vulgate↔Hebrew text-critical modules latin_module_h/v; Greek has its own greek_token_apparatus / greek_gold_cells / greek_witness_genealogy. Each engine reserves the full base-name taxonomy and fills its own slots.

Live Greek lexical front door

As of 2026-08-02, greek_sense_cell remains at 17,780 rows and preserves its original 13 analytical columns byte-for-byte. The serving config is a backward-compatible 36-column superset that additionally carries:

• the generated sense_gloss, collocates, and collocate_scope;
• reference_sources and reference_prose_status;
• full, untruncated LSJ prose with entry id, headword, match method, source label/repository/revision, and license;
• full Abbott-Smith prose with Strong identifiers, form, gloss, match method, source label/repository/revision, and license.

Abbott-Smith attachment is source-firewalled: a TBESG row is eligible only when its meaning ends in the explicit (AS) marker. The serving build is reproducible with scripts/enrich_greek_sense_cell_prose.py. Reference prose is witness evidence alongside the corpus-derived sense cell; it is not treated as the model's own output.

Selected findings (honest — see SI / validation artifacts)

• Calibration is strong (the headline that holds): per-occurrence confidence tracks model uncertainty, conf ⇔ softvec-entropy r = −0.904. Low-confidence occurrences are genuinely ambiguous.
• "Low-confidence ⇔ lexicon-disagreement" — supported at the OCCURRENCE grain, not the lemma grain. On held-out NT pilots, low CV-confidence coincides with BDAG↔anchor sense-crossing (Spearman −0.24, p<1e-4). It does not hold at lemma grain — an aggregate checker_finer correlation we initially measured (+0.43) was a lemma-frequency confound; the real relationship is per-occurrence. Stated precisely so it can't be over-read.
• The encoder accuracy ceiling is domain-bound, not size-bound. Larger general multilingual encoders (XLM-R-large 560M −0.09, mE5-large −0.06) are worse than the 125M ancient-Greek-pretrained GreBERTa on fine WSD; ensembling, syntax features, and a BEM bi-encoder are flat. Fine intra-domain senses cap ~0.52 macro-F1 — at/near where the WSD literature, the lexica, and LLM probes all plateau (human ITA 67–80%).
• Diachronic drift across the full span (drift_ranking): Χριστός drifts hardest (classical literal "anointed" → patristic theological "Christ", JS 0.371); Μαρία rises; σάρξ moralizes.
• Emergent macro-units cut across corpus taxonomy (clusters): classical-mythology vocabulary clusters apart from Christian-psalter vocabulary; NT, LXX+Pseudepigrapha, and Philo+patristic group by register.
• Consensus-tuned inventory (inventory_tuned): a light merge of over-split LN cells raises held-out apparatus↔BDAG agreement +0.021 ARI (Pareto published); aggressive merging over-collapses.
• Quote-projection = a coverage layer, NOT accuracy (quote_projection): 64,471 occurrences of 718 otherwise-abstaining sparse-tail lemmas gain a sense by inheriting the cited NT/OT gold LN — as a provenance fact + low-confidence witness, never asserted as the author's sense; 822 author reinterpretations are flagged as inherited↔distributional divergence, exposed not resolved.
• BDAG agreement (local benchmark, never shipped): our LN divisions are a consistent refinement of BDAG — adjusted Rand 0.42, homogeneity 0.63 over 521 co-attested polysemous lemmas.

Limitations (stated up front)

• WSD accuracy on fine intra-domain senses is ceilinged ~0.52 and not liftable by scale/ensemble/syntax (measured). The product is a calibrated-partial apparatus — confidence + abstention carry uncertainty; it never claims false precision.
• The NT-gold sparse tail (~1,297 polysemous lemmas with <5 gold/cell) is sparse in the gold itself and structurally unresolvable by exemplar WSD; quote-projection adds coverage there, not resolved senses.
• cluster_hallmarks retains the v1 build pending a refresh. Latin diachronic configs have known fixes pending — see distributional-latin/DIACHRONIC_FIXES.md (silver-scope + floruit-binned drift); do not read latin_drift_ranking as floruit-binned yet.

Provenance, license, IP firewall

• License: CC BY 4.0. MACULA CC BY 4.0 attribution required (gold LN); per-corpus licenses for LXX / secondary-sources / pseudepigrapha / patristics / classical-greek / intertextuality / Hebrew sources (MACULA-Hebrew · StepBible · BDB · Jastrow · ot-apparatus). SDBH is CC BY-SA 4.0 © UBS and ships only in the companion repo NuBerea/sdbh.
• IP firewall: output senses are opaque LN cells + own glosses. No BDAG or Lampe content — their verse→sense maps / PDFs / sense-counts are local-only; the public token_apparatus ships only open lexica (LSJ, Abbott-Smith) signals + the aggregate checker_finer agreement flag; BDAG agreement rates are cited, never the content. The copyrighted-lexicon sidecars (apparatus_crosswalk_sidecar=BDAG, lampe_sidecar=Lampe) ship coverage + sense-count(BDAG) + 'see … s.v.' pointer + license ONLY (parity with Latin's TLL sidecar) — never prose/glosses; Lampe omits the count (v0.1 parse unreliable). quote_projection carries inherited LN as flagged provenance, never truth.
• Hebrew IP firewall: SDBH (UBS, cc-by-sa-4.0) semantic domains ship only in NuBerea/sdbh and are never an anchor here (share-alike containment — never_anchor_guard enforced); HALOT (Brill, ©) ships as cover + sense-count + 'see HALOT s.v.' pointer + license only via hebrew_apparatus_crosswalk_sidecar, its source isolated in NuBerea/quarantine-works. PD BDB + Jastrow ship full prose.
• EU sui-generis database-right consideration applies on redistribution.
• Known gap: patr shard_00028 ~1.5% of patr occurrences skipped (corrupt upstream); refreshes on re-upload.

Latin (latin_* configs) — unified multilingual j (D67)

The same method + schema, applied to Vulgate + patristic Latin, published as latin_* configs alongside the Greek, rebuilt by scripts/build_latin.py from the Latin primitives subtree joined onto the PD-lexica witness repos. The Latin tail is filled by cross-lingual gold projection (Latin inherits Greek NT gold LN via the Vulgate Strong's bridge) — a weak gold-anchored silver (confidence

• layer carry the uncertainty). ⚠ Diachronic-correctness fixes are pending (DIACHRONIC_FIXES.md): silver is currently stamped per-lemma incl. classical (anachronism), and latin_drift_ranking is a 2-bucket proxy not yet floruit-binned. PD lexica (Lewis & Short · Du Cange · Forcellini) ship from their T0 repos; TLL = coverage fact only (BY); no copyrighted prose ships.

Hebrew (hebrew_* configs) — unified multilingual j (D67)

The same method + schema, applied to Biblical Hebrew (Masoretic MT/WLC spine) with a Hebrew encoder (alephbertgimmel, MLM-adapted) + per-lemma KMeans L0, published as hebrew_* configs alongside Greek and Latin and rebuilt by the scripts/hebrew/ builders. 15 configs fill the parallel base-name taxonomy — hebrew_sense_cell (+ _enriched), hebrew_lemma_apparatus, hebrew_inventory_tuned, hebrew_lexica_apparatus / hebrew_lexica_prose / hebrew_lexica_concordance, hebrew_sense_apparatus, hebrew_lemma_collocations / hebrew_sense_collocations, hebrew_unit_cell_mass, hebrew_clusters / hebrew_cluster_hallmarks — plus two Hebrew-distinctive members:

• hebrew_module_v — the multi-witness Vorlage divergence registry (Latin module_v analog): where a witness implies a Hebrew Vorlage divergent from MT/WLC. witness_kind splits hebrew_witness (DSS / Samaritan Pentateuch / Aleppo — genuine Hebrew-language content) from crosslang_vorlage (LXX / Vulgate — pointers only; their sense data lives in the greek_* / latin_* arms + the lxx-wlc / vulgate bridges, never duplicated here). Consolidated from NuBerea/ot-apparatus.
• hebrew_apparatus_crosswalk_sidecar — HALOT (Brill) coverage sidecar: cover + sense-count + 'see HALOT s.v.' pointer + license ONLY (parity with Greek BDAG / Latin TLL sidecars). No prose, no page numbers; the HALOT source is isolated in NuBerea/quarantine-works.

Lexica witnesses. PD BDB (NuBerea/hebrew-lexicon)

• Jastrow ship full prose in hebrew_lexica_prose. SDBH (UBS Semantic Dictionary of Biblical Hebrew) is a gold/witness signal only: its semantic domains ship exclusively in the share-alike companion NuBerea/sdbh (sdbh_lemma_to_domains, sdbh_to_ln, sdbh_domains; cc-by-sa-4.0 © UBS) and are never an anchor in this CC-BY-4.0 repo — anchoring on an SDBH domain would propagate share-alike to the whole lexicon. The cross-language hub reaches SDBH only by an ln_code join to NuBerea/sdbh::sdbh_to_ln.

Cross-language. Hebrew joins Greek + Latin on the shared Louw–Nida pivot in crosslang_sense_alignment (per-LN lemma lists; aligned = n_langs ≥ 2): 1,906 LN codes carry Hebrew, 637 are full three-way (Greek + Latin + Hebrew). Textual cross-links run through lxx-wlc-bridge (Hebrew↔Greek) and vulgate-bridge (Hebrew↔Latin), consolidated in hebrew_module_v.

Granularity. Verse/occurrence grain is preserved — the L0 token_membership ships occurrence-level over the WLC spine (hebrew/ subtree of distributional-lexicon-primitives); every apparatus/edge row carries match_grain + an osisRef / headword anchor.