Hebrew Morphosyntax

Status: Current Last updated: 2026-09-07 07:04 EDT

Hebrew-specific handling in batchalign3’s morphosyntax pipeline.

Language Code

ASR

Hebrew uses a fine-tuned Whisper model for the HuggingFace engine:

The ivrit-ai/whisper-large-v3 model is trained on Hebrew conversational speech and significantly outperforms generic Whisper on Hebrew audio.

RTL Punctuation

Hebrew text may contain Arabic-script punctuation from mixed content. The ASR post-processing pipeline normalizes RTL punctuation to ASCII:

This normalization runs for all languages, not just Hebrew, it ensures CHAT files contain only ASCII punctuation terminators regardless of source script.

Morphosyntax Features

Hebrew has two language-specific UD features that batchalign3 maps to CHAT %mor suffixes: HebBinyan and HebExistential.

HebBinyan (Verb Conjugation Pattern)

Hebrew verbs belong to one of seven binyanim (conjugation patterns): PAAL, NIFAL, PIEL, PUAL, HIFIL, HUFAL, HITPAEL.

Stanza’s Hebrew model outputs the HebBinyan feature on verbs. batchalign3 converts it to a lowercase suffix in %mor:

UD features: HebBinyan=PAAL|Number=Sing|Person=3|Tense=Past|VerbForm=Fin
%mor suffix: -paal&3S&PAST

The binyan is lowercased in the suffix: PAAL → paal, HIFIL → hifil.

HebExistential

The Hebrew existential (יש/אין, “there is”/“there isn’t”) gets a special feature in Stanza:

UD features: HebExistential=True|VerbForm=Fin
%mor suffix: -true

The value is lowercased: True → true.

Feature Format

The full verb suffix format (shared across all languages):

-VerbForm-Aspect-Mood-Tense-Polarity-Polite-HebBinyan-HebExistential-NumberPerson-irr

Hebrew-specific features slot into their dedicated positions. The -irr suffix (English irregular verbs) is not applied to Hebrew, it is gated to English only.

Implementation

The feature extraction is language-agnostic in implementation, the code in features.rs checks for HebBinyan and HebExistential in any language’s feature set, but only Stanza’s Hebrew model actually produces these features:

// crates/batchalign-transform/src/morphosyntax/features.rs (lines 43, 46)
if let Some(v) = feats.get("HebBinyan") {
    parts.push(v.to_lowercase());
}
if let Some(v) = feats.get("HebExistential") {
    parts.push(v.to_lowercase());
}

MWT

Hebrew uses the MWT processor. Stanza’s Hebrew model handles Hebrew contractions (preposition + article combinations like בַּ → ב + ה).

Number Expansion

Hebrew does not have a dedicated number expansion table in num2lang.json. Digit strings in Hebrew ASR output pass through unexpanded. This is a known gap, Hebrew numbers in CHAT output will appear as digits rather than Hebrew word forms (אחת, שתיים, שלוש, etc.).

No Other Language-Specific Workarounds

Unlike English, French, Japanese, Italian, Portuguese, and Dutch, Hebrew has no Stanza workarounds in batchalign3. The HebBinyan and HebExistential feature mapping is standard UD feature processing, not a bug workaround.

If systematic Stanza errors are discovered for Hebrew, a crates/batchalign-transform/src/morphosyntax/lang_he.rs file should be created following the pattern of existing language files (lang_en.rs, lang_fr.rs, lang_it.rs, lang_ja.rs).

Source Files

Test Coverage

This page last changed: 2026-09-09 (commit 8f4ed0f2). The whole book last changed: 2026-09-16 (commit 34d249d8).