12. Mova and Modern Hebrew: where it stands, what it would need

Mova does not parse Modern Hebrew yet. This chapter says what Mova already has that touches Hebrew, what is missing, and a plan in steps that keeps to Mova's rules: deterministic Rust, measured before claimed, licences respected, a large model only as a teacher. It is a design note, not a commitment: a new language is a new direction and needs the operator's word before work starts (AGENTS.md — Mova).

← README · ← 11 UD analysis · sources →

12.1 What exists today (2026-10-17)

So the pieces that would carry a Hebrew pipeline exist in general form (CoNLL-U with MWTs, a language-agnostic GPU parser, a UD dialect with converters, a rule-based Hebrew reader for the Bible), but none of them reads Modern Hebrew text.

12.2 What Hebrew asks of a pipeline that English does not

From chapters 02, 04, 05 and 11:

1. Segmentation before tagging. One token in three is a multiword token; prefixes ו ה ש ב כ ל מ and pronoun suffixes must be split, and the split is ambiguous in unpointed text (בצלם has five readings). English has nothing comparable beyond don't.
2. The covert article. בבית is "in a house" or "in the house"; the letters do not say. The answer lives in a feature (Definite=Def on the preposition, IAHLT scheme) and is decided by context.
3. No vowels. Most words have several readings: ספר séfer "book", sapar "barber", safar "counted", siper "told", sfar "frontier". Lemma and POS depend on context. Pointing (niqqud) appears only in poetry, children's books and dictionaries.
4. Spelling variation. Full spelling (ktiv male) is standard but variable: היתה/הייתה "was", אמא/אימא "mother". A lexicon must hold the variants or normalise.
5. Root-and-pattern morphology. A verb's lemma (the 3MS past) and its binyan predict voice and valency (pi'el active, pu'al its passive; hif'il causative; hitpa'el reflexive or reciprocal). This is exactly the kind of structure a frame builder needs (who did what to whom), and it is invisible in the letters unless analysed.
6. Gender and number agreement everywhere — on verbs, adjectives, numerals, pronouns. Agreement is a strong cue for attachment; a parser should use it.
7. Verbless clauses. Present-tense "be" is absent or a pronoun (הוא, זה). The predicate heads the clause; a frame reader must make "X is Y" frames without a verb.
8. Existentials and possession (יש/אין לי) instead of "have": in world terms, a holding expressed by an existential with a dative.
9. Right-to-left text, mixed with digits and Latin. Logical order in files, bidirectional display; never reverse strings. Gershayim ״ and geresh ׳ are often typed as " and '.

12.3 Licences of the Hebrew data (for building vs testing)

Following Data and model licences and the operator's rules (NC and research-only: tests only; CC BY-SA: as little as possible, never carried or modified into a release):

The paradigms of Hebrew are facts of the language: Mova can write its own generator of inflection (binyan × person × tense, weak-root rules) under its own licence and check it against UniMorph, Hspell and Wiktionary without copying them. That is the cleanest road to a lexicon that can ship.

12.4 A plan in steps

Each step has its measure and its negative control. Nothing here is started without the operator's word.

Step 0 — data in the Mova dialect

• A Hebrew page in the dialect: base = the IAHLT scheme (concatenative MWTs, Definite=Def on the preposition, no inserted words), because it is the scheme of the CC BY-SA treebanks, it is simpler, and it matches Arabic. Keep compound:smixut as a subtype in the Mova dialect (it is more informative than compound and convertible without loss: § 11.11 rule 4).
• Converters both ways, as rules in the convert language: IAHLT ⇄ old HTB (deterministic, Zeldes et al. 2022), so HTB serves as a second test set.
• Seeds in dialects/mova/seeds/ for each Hebrew decision, with the evidence.
• A FEATS registry for Hebrew (ud-registry-he.tsv), including HebBinyan, HebExistential, Definite=Cons, Prefix, ExtPos.
• Measure: round trip IAHLT → old → IAHLT is byte-identical on all three IAHLT files. Negative control: a converter with one rule removed must fail the round trip.

Step 1 — a segmenter

• Character-level: for each character, "does a new word start here?" (Zeldes 2018), plus the covert-article decision on ב/ל/כ. Deterministic features: the character window, prefix and suffix letter sets, a lexicon lookup of the candidate pieces.
• Lexicon from our own paradigm generator (§ 12.3), checked against UniMorph/Hspell.
• Trained on IAHLTwiki + IAHLTknesset train; decided on two-fold cross-validation on train; IAHLT test read once per release.
• Target to beat: MWT F1 of Stanza (92.9 on Wiki) on the way to HebPipe (98.8) (Zeldes et al. 2022: Table 4). Negative control: a segmenter that never splits must score far lower (it gets every MWT wrong: about one token in three).

Step 2 — tagger and parser

• gparse trained on gold Hebrew words (it is language-agnostic); then end to end on the segmenter's output.
• The en perceptron tagger/parser generalised to a second language would need its lexicon and feature templates made per-language; agreement features (gender, number, person) should be template inputs.
• Witnesses for the agreement loop (tools/treeagree): Stanza and Trankit Hebrew, UDPipe 2 — as tests only, under their data licences.
• Measure: LAS on IAHLTwiki dev (cv), HTB test as an out-of-domain measure. Run twice; the two runs must be identical.

Step 3 — Hebrew into frames

• prag rules for Hebrew trees:
  • verbless clause with cop or no verb → a state frame (X is Y);
  • יש/אין + ל → possession (a holding in world), with polarity;
  • binyan → voice and valency: nif'al/pu'al/huf'al → passive or unaccusative frame, hif'il → causer + causee, hitpa'el → reflexive or reciprocal;
  • את + definite noun → patient;
  • pro-drop: 1st/2nd person subject from the verb's person feature;
  • negation לא / אין / אינו / אל.
• A bilingual concept table Hebrew lemma → Mova concept (the English lexicon), built from our own seeds and checked against Wiktionary glosses.
• Measure: the same as macula prag-compare for the Bible: agreement of the Hebrew frames with the frames of an English translation of the same sentences, on a parallel set we may test on.

Step 4 — answering

world and the readers are language-independent once frames exist. A Hebrew question (מי נתן לדני את הספר? "who gave Dani the book?") becomes a frame query like an English one. A large model may teach (check silver trees, propose seeds) but never answers.

12.5 What we would not do

• Train on HTB or ship anything derived from it (NC).
• Copy Wikibooks, Wiktionary or grammar tables into the repo.
• Let a teacher model segment or parse at test time.
• Normalise spelling silently: every normalisation is a rule with a reason, and the original form is kept.

12.6 Open questions (parked for the operator)

1. Is Modern Hebrew a language Mova should take on now, after English, or only after the abstract annotation (The Mova dialect: a UD metadialect, treebanks, and converters) is ready?
2. Hspell is AGPL: may it be used to build (e.g. as a lexicon source), or only to test?
3. Should the Mova dialect keep compound:smixut (HTB, PostRab, UD guidelines) or follow IAHLT's plain compound?

References

• Mova repo: AGENTS.md — Mova, Architecture, The Mova dialect: a UD metadialect, treebanks, and converters, English parsers — local collection, Data and model licences, en/src/conllu.rs, en/src/gram.rs, gparse — the SLM parser trained by gradient descent on the GPU, gparse/src/data.rs, macula, treeagree — agreed UD trees (Haiku + the best open parsers), treebanks/he/.
• Zeldes, Amir. 2018. A Characterwise Windowed Approach to Hebrew Morphological Segmentation. SIGMORPHON 2018. aclanthology.org/W18-5811 our card
• Zeldes, Amir, Nick Howell, Noam Ordan and Yifat Ben Moshe. 2022. A Second Wave of UD Hebrew Treebanking and Cross-Domain Parsing. EMNLP 2022. aclanthology.org/2022.emnlp-main.292 our card
• Tsarfaty, Reut, Shoval Sadde, Stav Klein and Amit Seker. 2019. What's Wrong with Hebrew NLP? And How to Make it Right. EMNLP-IJCNLP 2019 demos. aclanthology.org/D19-3044 our card
• Licences: Hspell (AGPL-3.0) hspell.ivrix.org.il; UniMorph Hebrew (CC BY-SA 3.0) github.com/unimorph/heb; Dicta datasets huggingface.co/dicta-il; UD treebanks per their LICENSE.txt.

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/languages/he/12-mova/