greek-inflexion-eee

A fork of jtauber/greek-inflexion packaged as an installable Python library (greek-inflexion-eee) for use in the Ελληνικά Εκπαιδευτικά Εργαλεία (EEE) — Greek Language Educational Tools ecosystem.

The original library generates and analyzes Ancient Greek inflectional paradigms (verbs, nouns, adjectives) with precise accentuation. See README-greek-inflexion.md for the upstream documentation.

What this fork adds

• Installable package — src/ layout; data files bundled via importlib.resources so the library works correctly when installed (not just run from the source directory).

• Factory functions — ready-to-use entry points:

  from greek_inflexion_eee import (
      load_default, load_noun_default, load_adj_default,
      load_verb_lexicons, load_noun_lexicons, load_adj_lexicons,
  )
  
  gi = load_default()                          # verb inflection (Pratt lexicon)
  gi = load_noun_default()                     # noun inflection (Pratt lexicon)
  gi = load_adj_default()                      # adjective inflection (Pratt lexicon)
  gi = load_verb_lexicons("homer")                  # verb inflection — Homeric corpus
  gi = load_verb_lexicons(["homer", "lxx"])         # merge two corpora
  gi = load_verb_lexicons(["pratt", "/my.yaml"])    # Pratt + custom file
  gi = load_noun_lexicons("homer")             # noun inflection — Homeric corpus
  gi = load_noun_lexicons(["homer"])           # merge noun lexicons
  gi = load_adj_lexicons("pratt")              # adjective inflection — Pratt lexicon
  

• Bundled corpus lexicons — named lexicons for verbs and nouns:

  Verbs (load_verb_lexicons):

  Name	Verbs	Source	Period / dialect
  "pratt"	20	Pratt textbook	teaching
  "dik"	10	Dik textbook	teaching
  "ltrg"	34	LTRG textbook	teaching
  "lsj"	9	hand-authored	Classical Attic, V–IV c. BCE
  "homer"	2335	Homeric corpus	Epic/Ionic, ~800 BCE
  "lxx"	1905	Septuagint	Biblical κοινή, ~250–100 BCE
  "morphgnt"	1848	New Testament	κοινή, ~1st c. CE
  "morpheus"	46	Morpheus-confirmed attested forms	Epic/Homeric (mixed)
  "byzantine"	61	hand-curated from Sophocles' Lexicon (1887)	Byzantine, ~4th–15th c. CE

  Combined unique coverage: ~5050 verbs. Custom YAML files (same format) are also accepted as absolute paths.

  Nouns (load_noun_lexicons) — always includes Pratt as base:

  Name	Nouns	Source
  "pratt"	26	Pratt textbook paradigm nouns
  "homer"	15	Homeric Odyssey/Iliad vocabulary
  "lsj"	18	Classical Attic (Perseus/LSJ-verified)
  "morpheus"	62	Morpheus-confirmed attested forms, Epic/Homeric (mixed)

  Adjectives (load_adj_lexicons) — always includes Pratt as base:

  Name	Source
  "pratt"	Pratt textbook paradigm adjectives

  Course-specific vocabulary lexicons (formerly "odyssey_morpheus", "palaestra_morpheus") now live as course-local files in created_with_eee, not here — see "Course-specific lexicons moved out" below.

• Perseids Morpheus — the "morpheus" lexicon is sourced from or verified against Perseids Morpheus, a free, public, no-auth HTTP API that analyzes a single Ancient Greek surface form and returns every morphological reading it can find — lemma, part of speech, full grammatical features, and a dial field naming the attested dialect(s) for that specific reading (e.g. "epic", "Homeric Ionic", "Attic"). It analyzes; it does not generate (no "give me all forms of X"). Built originally by the Perseus Project for the Homeric corpus specifically, so Epic/Ionic coverage is a genuine strength, not an afterthought — but always check dial before trusting a reading for a specific register; see tools/morpheus/README.md for the querying/ auditing scripts and two real bugs they caught.

• Morpheus-confirmed attested-form lexicon ("morpheus") — unlike every other bundled lexicon, every entry is a forms: block: a verbatim attested surface form, not a stems: entry generated on demand. Built for lemmas the stem-based lexicons can't handle cleanly — athematic -μι verbs, contract verbs, compounds, deponents, non-2nd-declension nouns, oxytone nouns, and irregular/ suppletive nouns (Ζεύς) — by collecting attested (form, UD-feats) pairs from the UD_Ancient_Greek-Perseus and UD_Ancient_Greek-PROIEL treebanks and independently re-confirming each one against the Perseids Morpheus analyzer (matching lemma + tense/aspect + voice, or case + gender). Since forms: bypasses stem lookup entirely (generate() checks it before any stem-based generation), there's no stem-extraction/re-inflection risk for these irregulars — and an explicit override always wins over a stem-generated guess, which is what makes it safe to merge alongside a stem-based lexicon like "homer" for the same lemma. See ancient_greek_backend_eee's own README for the companion fix that makes .paradigm() (the full-table view) render this data correctly — restricting a noun's enumerated genders to what it actually has, since the fix and this lexicon were built together to solve the same problem (Ζεύς-style irregulars) from two ends.

• Byzantine lexicon ("byzantine") — a hand-curated forms:-only lexicon documenting several well-known Byzantine-period verb morphology shifts, mostly variants of an analogical -αν/-σαν/-ασι ending spreading into slots classical Greek marked differently (e.g. ἔγνωκαν replacing classical ἐγνώκᾱσι(ν); ἐποιοῦσαν replacing ἐποίουν; ἐδώκασι(ν) replacing ἔδωκαν/ἔδοσαν) — the direct ancestor of Modern Greek's uniform past-tense endings. See the lexicon file's own header for the complete pattern-by-pattern breakdown and sourcing/verification discipline for each. 61 lemmas, sourced from Sophocles' Greek Lexicon of the Roman and Byzantine Periods (1887, public domain) — specifically its Introduction's own systematic survey of this phenomenon, not scattered dictionary entries. The two entries whose citation is an NT verse (γιγνώσκω, ὁράω) are additionally cross-verified against the Westcott-Hort/Nestle 1904 critical text; the rest cite patristic/Byzantine authors (Barnabas, Hippolytus, the Sibylline Oracles, Malalas, Theophanes, ...) instantiating the same independently-documented phenomenon (see Wikipedia's Medieval Greek article). Unlike morpheus, this is not a systematic sweep of a corpus — TLG and LBG (the two best-fitting Byzantine lexicons) were investigated and ruled out as sources, since both explicitly prohibit bulk/programmatic extraction; see the lexicon file's own header for the full sourcing story and what was deliberately excluded (ambiguous mood readings, illegible OCR, forms already reachable via another lexicon's own stem-based generation).

  Use "byzantine" merged with a Koine/Attic base, not standalone. Sophocles' Introduction documents specific, optional deviations from an already-known classical paradigm (e.g. "3rd plural sometimes ends in -αν instead of -ασι"), not a self-contained stemming engine — there's no principal-parts information here, only citations for individual already- inflected cells. Standalone (load_verb_lexicons("byzantine")), the lexicon therefore only covers its own 61 lemmas with 1-2 cells each. Merged with lxx/morphgnt/lsj as the base, it inherits their combined ~3000+ lemma paradigm coverage and the byzantine override silently wins wherever Sophocles documents a divergence, falling through cleanly to the Koine/Attic-generated form everywhere else — which is also the linguistically accurate picture: most Byzantine literary Greek genuinely is Koine/Attic morphology, cell for cell; the exceptions layer is exactly where (and only where) it actually diverges.

  gi = load_verb_lexicons(["lxx", "morphgnt", "pratt", "ltrg", "lsj", "byzantine"])
  gi.generate("γιγνώσκω", "XAI.3P")   # {'ἔγνωκαν': [...]} (byzantine override)
  gi.generate("πάσχω", "AAI.3P")      # {'ἔπαθον': [...]}  (plain Koine, no override)
  

• GreekInflexion.known_lemmas() (2026-09-24) — the set of lemmas an instance's lexicon actually defines, through stems: or forms:. Use it instead of gi.lexicon.lemma_to_stems.keys(): upstream Lexicon.find_stems() reads that defaultdict with [], so every generate() call for a lemma the lexicon lacks leaves an empty entry behind for it.

• Byzantine lexicon: ἔρχομαι -οσαν gap closed (2026-08-14) — the -οσαν pattern's own entry for ἔρχομαι (ἤλθοσαν, AAI.3P) was in Sophocles' source list from the start of that pattern's mining pass but was missed; added now. Still 61 lemmas (no new lemma key — ἔρχομαι already had an AAO.3P entry from a different pattern).

• Course-specific lexicons moved out (2026-07-31) — "odyssey_morpheus" and "palaestra_morpheus" (forms:-only lexicons covering gaps in the created_with_eee Odyssey and Palaestra course vocabularies, verified against the Perseids Morpheus analyzer) were removed as named lexicons here and now live as course-local files directly in created_with_eee (see each course's own AGENTS.md). Both courses are still under active development — bundling their data as a named package lexicon had forced a full version bump + PyPI republish every time one more lesson needed a few more gap-mined words (7 releases in 16 days for Odyssey alone). Course-specific data goes back into the package only as a deliberate, one-time consolidation once a course is actually finished — not incrementally per lesson. Pass an absolute file path in lexicons=[...] instead of a registered name to use them (same mechanism as any other custom YAML lexicon file, see "Absolute file paths" above).

• Unrecognized lexicon names now raise, and load_lexicons is renamed (2026-07-31) — an unrecognized bare (non-absolute-path) name passed to load_verb_lexicons/load_noun_lexicons/load_adj_lexicons/load_pron_lexicons now raises ValueError naming the bad lexicon, instead of the previous split behavior: the noun/adj/pronoun loaders silently dropped an unknown name with no signal anything went wrong, while the verb loader (worse) fell through to opening it as a literal package-resource filename and crashed with a confusing raw FileNotFoundError. Both were real problems in practice, not theoretical — removing "odyssey_morpheus" above is exactly the kind of change the silent-skip path was built to hide. New known_verb_lexicons() / known_noun_lexicons() / known_adj_lexicons() / known_pron_lexicons() return each POS's registered name set, for a caller (e.g. ancient_greek_backend_eee, which shares one general-purpose lexicons=[...] list across every part of speech) to filter a shared list down to what's actually valid for a given POS before calling its loader — a bare name irrelevant to one POS is not a mistake in general, only when it's not registered anywhere. Also: load_lexicons() (verbs) is renamed to load_verb_lexicons(), matching load_noun_lexicons/load_adj_lexicons/ load_pron_lexicons — it predated that naming convention (added when verbs were the library's only part of speech) and was never retroactively renamed until now.

• Adjective morphology — adj_stemming.yaml + pratt_adjs_lexicon.yaml covering 2-1-2 uncontracted/contracted, two-termination, 3-1-3 participial and υ-stem, 3-3 σ-stem, and comparative adjectives. 3-3 σ-stem two-termination adjectives (e.g. ἀληθής) generate masculine oblique forms directly now — previously only the feminine cells were reachable via noun_stemming.yaml's existing masculine sigma-stem rule, which is correctly scoped to a different, genuinely masc-only class (contracted proper names like Περικλῆς) and left untouched; the two-termination adjective case gets its own rules instead. Verified via a full sweep (real Perseids Morpheus lookups, not guessing) that only μείζων (comparative) and ἐυπλόκαμος were genuinely missing a cell for their shared gender's data; added those directly. Every other adjective with an apparent masc/fem gap turned out to be a regular 3-termination adjective (distinct forms exist, just not yet in the bundled lexicons) or a genuinely gender-restricted fixed epithet — not fixed here, since mirroring the other gender for those would be grammatically wrong. Separately, δίκαιος's feminine paradigm had 5 wrong cells: NSF/GSF/ASF/APF came out accented on the antepenult (δίκαια etc.) instead of the correct penult (δικαία), and GPF got a wrong baked-in circumflex (δικαιῶν instead of δικαίων) — both are the stemming engine's inability to know a bare, undiacritized alpha's true (here, long) vowel length. Fixed via targeted forms: overrides for just those 5 cells (same pattern as ταχύς's entry in the same file); the unaffected DSF/NPF/DPF and all masc/neut cells stay on the regular mechanism.

• Improved accent engine — correct handling of nominal accent overrides for neuter forms; _fix_nominal_oxytone converts acute on inflected long ultima to circumflex while preserving native long ultimas (e.g. βασιλεύς). The pedagogical vowel-length macron some lexicon stems carry (e.g. λύω's long υ, needed so participles/infinitives/imperatives get circumflex where Greek requires a long vowel) is now stripped from final output once accent computation has used it — previously it leaked into acute-accented forms as a stray combining mark (λύω → λῡ́ω).

Installation

pip install greek-inflexion-eee

Development version (latest, from Codeberg):

pip install "greek-inflexion-eee @ git+https://codeberg.org/EEE-project/greek-inflexion-eee.git"

Or from source:

pip install -e .

Quick start

from greek_inflexion_eee import (
    load_default, load_noun_default, load_adj_default,
    load_verb_lexicons, load_noun_lexicons, load_adj_lexicons,
)

# Verbs — Pratt lexicon (20 verbs, teaching vocabulary)
gi = load_default()
gi.generate("λύω", "AAN")               # {'λῦσαι': [...]}

# Verbs — corpus lexicon by name
gi = load_verb_lexicons("homer")
gi.generate("λέγω", "PAD.2S")           # {'λέγε'}
gi.generate("ἀκούω", "PAD.2S")          # {'ἄκουε'}

# Verbs — merge corpora
gi = load_verb_lexicons(["homer", "lxx"])
gi.generate("παύω", "PAD.2P")           # {'παύετε'}

# Nouns — Pratt paradigm words only
gi = load_noun_default()
gi.generate("θεός", "NSM")              # {'θεός': [...]}
gi.generate("βασιλεύς", "NSM")          # {'βασιλεύς': [...]}

# Nouns — Homeric vocabulary (Pratt + Homer merged)
gi = load_noun_lexicons("homer")
gi.generate("θάνατος", "NSM")           # {'θάνατος': [...]}
gi.generate("μάχη", "GSF")              # {'μάχης': [...]}
gi.generate("βοῦς", "GPM")              # {'βοῶν': [...]}

# Nouns — Morpheus-confirmed attested forms (real Epic spellings, not generated)
gi = load_noun_lexicons("morpheus")
gi.generate("Ζεύς", "GSM")              # {'Διός': [...]} (suppletive, not a stem+ending)
gi.generate("θεός", "GSM")              # {'θεοῖο': [...]} (Epic genitive, not Attic -ου)

# Verbs — Byzantine-period attested divergence (merge alongside a Koine lexicon)
gi = load_verb_lexicons(["morphgnt", "byzantine"])
gi.generate("γιγνώσκω", "XAI.3P")       # {'ἔγνωκαν': [...]} (not ἐγνώκᾱσι(ν))

# Adjectives
gi = load_adj_default()
gi.generate("ἀγαθός", "NSM")            # {'ἀγαθός': [...]}
gi.generate("ἀγαθός", "NSF")            # {'ἀγαθή': [...]}
gi.generate("ἀγαθός", "NSN")            # {'ἀγαθόν': [...]}

Origin

Forked from jtauber/greek-inflexion by James Tauber. The upstream library and its data (stemming rules, Pratt/Dik/Homer lexica) are the work of James Tauber and contributors; see AUTHORS and LICENSE.

This fork is maintained by Ελληνικά Εκπαιδευτικά Εργαλεία (EEE) — Greek Language Educational Tools.