Jewish Palestinian Aramaic in Mova: what it would need
General plan: Aramaic in Mova: what it would need. Status (2026-10-17): no JPA text is parsed or tagged.
1. The data
Counted 2026-10-17 in the Sefaria export (registry texts.tsv); words are whitespace tokens after stripping HTML; Hebrew and Aramaic together.
| text | edition | licence | files | words | use |
|---|---|---|---|---|---|
| Jerusalem Talmud (39 tractates) | Guggenheimer (De Gruyter 1999–2015), vocalised | CC BY | 39 | 811 526 | build and test, with attribution |
| Jerusalem Talmud | Venice 1523; Mechon-Mamre | unknown → private | not used | ||
| Pseudo-Jonathan (Torah) | Sefaria "Targum Jonathan on Genesis" … | public domain | 5 | build and test (mixed language) | |
| Fragment targum | Sefaria "Targum Jerusalem on Torah" | public domain | 1 | build and test | |
| Pseudo-Jonathan + Fragment targum together | public domain | 6 | 122 656 | ||
| Targum Neofiti | Vatican MS (Sefaria file) | unknown → private; the file holds only 137 words | 1 | 137 | not used; a full PD text is wanted |
| Palestinian midrashim (Genesis Rabbah etc.) | various Sefaria versions | some PD, some unknown | mostly Hebrew; Aramaic stories inside |
The Guggenheimer Yerushalmi is the most valuable open Aramaic text Mova has: 39 tractates, vocalised, CC BY (attribution only). It is the one large Talmudic corpus that can be used to build.
2. What is special about JPA for a model
- Very little parsed or annotated material exists, and none of it is open. Sokoloff's dictionary and Fassberg's grammar are in copyright.
- The text is mixed with Hebrew more than the Bavli; Aramaic sentences are often short links between Hebrew statements (The lexicon of Jewish Palestinian Aramaic §4).
- The printed Yerushalmi was Babylonianised by copyists; a model trained on it will see some JBA forms (לימא 3×, האי 6×). These are not errors of the model; they should be counted and reported.
3. Telling JPA from JBA: a measured cue set
The counts in Jewish Palestinian Aramaic (jpa) §2 and The lexicon of Jewish Palestinian Aramaic §1 give a deterministic classifier with explainable cues:
| JPA cue | Y | B | JBA cue | Y | B |
|---|---|---|---|---|---|
| כְּדוֹן | 419 | 0 | הַשְׁתָּא | 1 | 974 |
| לוֹן | 731 | 0 | לְהוּ | 9 | 4 024 |
| הָדָא | 2 117 | 23 | הַאי | 6 | 2 246 |
| הֵיךְ | 218 | 0 | הֵיכִי | — | 1 399 |
| אוּף | 188 | 0 | נַמִי | — | 6 858 |
| בְּגִין | 250 | 2 | קָא | 0 | 2 828 |
| סַגִּין | 61 | 0 | טוּבָא | 0 | 191 |
| אִין תֵּימַר | 382 | — | אִיכָּא | 6 | 2 488 |
(Y = Guggenheimer Yerushalmi, CC BY; B = Wikisource Bavli, CC BY-SA, measurement only; unpointed tokens; 2026-10-17.)
Proposed test (not yet run): score each paragraph by (JPA cues − JBA cues); measure on held-out paragraphs of both Talmuds, and on the hard cases — Palestinian traditions quoted in the Bavli, and Babylonian forms in the Yerushalmi. The negative control: the same classifier on Hebrew paragraphs (Mishnah) must say "neither", not guess.
The same cue table extends to the targums: Onkelos has neither set (it has כְּעַן, אֲרֵי, יָת), Pseudo-Jonathan has the JPA set partly (The lexicon of the targums §5).
4. Tokenisation and trees
- Proclitics and suffixes as in Aramaic in Mova: what it would need §3.
- לוֹן = לְ + ־הוֹן contracted: one token in writing, split as preposition + pronoun.
- The ethical dative (נִיזוּל וְנִיחוּת לִי, נָחַת לֵיהּ) — a ל־ phrase co-referent with the subject — needs one convention (proposal:
obl:ethicalorexpl:dat), applied everywhere (The verb in Jewish Palestinian Aramaic §6). - חַד / חֲדָא as an indefinite article:
det(withPronType=IndorNumType=Cardkept in MISC), notnummod, when it introduces a new referent. - לֵית + pronoun + participle: the existential is the negation (
advmodwithPolarity=Neg), the pronoun the subject, the participle the head. - Vocalisation: the Guggenheimer pointing is modern and consistent; keep it in FORM, match by consonants.
5. What JPA can teach the SLM
- Short stories with clear events (y. Bava Metzia 2:5, y. Berakhot 2:8): agent, goal, purchase, return, moral — material for the event frames of the school, from an open (CC BY) source.
- Dialect identification as an explainable task: the cue table above is a model that can say why it decided.
- Midrashic expansion in the Palestinian targums (Gen 4:8): a natural contrast between a literal (Onkelos) and an interpretive (Palestinian) rendering of the same verse.
The Elamur team
This page in Markdown → https://md.mova.elamur.ai/free/en/docs/languages/jpa/mova/