Targumic Aramaic in Mova: what it would need

The general plan for all five varieties is in Aramaic in Mova: what it would need. This file adds what is specific to the targums. Status (2026-10-17): nothing targumic is parsed.

1. The data

Counted on 2026-10-17 from the Sefaria export in sefaria (registry texts.tsv; verses = entries of the text arrays; words = whitespace tokens after stripping HTML).

Every verse has a Hebrew original with the same reference, and for the Torah the Hebrew already has trees (tools/macula, hbo, 2026-10-17). That makes the targum the best-placed Aramaic corpus for Mova.

2. Alignment and projection

The plan of Aramaic in Mova: what it would need §4.2, in detail for Onkelos:

1. Verse alignment: by reference. Check that the verse division of the Sefaria Onkelos matches the WLC for every chapter (5 846 Onkelos verses against 5 846 Hebrew verses in the Torah, if they match one to one — to be verified per chapter, since the Masoretic and printed-targum divisions sometimes differ).
2. Word alignment inside the verse, in three passes, all deterministic:
   • anchors: the pairs of The lexicon of the targums §2.3 (אֶת ↔ יָת, כִּי ↔ אֲרֵי, אֲשֶׁר ↔ דִּי, הִנֵּה ↔ הָא, לֵאמֹר ↔ לְמֵימַר), proper names (identical consonants), יְהוָה ↔ יְיָ;
   • lexicon: the Hebrew → Aramaic table of The lexicon of the targums §2 extended from Jastrow (whose entries cite "h. text" and the Hebrew word a targum word renders) and from the regular sound correspondences (Script and phonology §4);
   • position: the remaining words in order, between anchors (monotonic alignment).
3. Projection: each Aramaic word takes the head and relation of its Hebrew counterpart; Aramaic words with no counterpart (קֳדָם, מֵימְרָא דַּ־, added explanations) get rules: קֳדָם + noun is an obl of the verb; מֵימְרָא is the noun the Hebrew subject's relation goes to, with יְיָ as its nmod.
4. Confidence: the share of aligned words per verse. Verses below a threshold (poetry, expansions; Translation technique and syntax of the targums §5) are flagged, not projected.

How to measure. No Aramaic gold exists. Annotate a sample by hand (for example 100 verses spread over the five books, including Gen 49 and Deut 32–33), then compare. The rule: decisions on half the sample, the other half read once, as with SVAMP. A negative control: projection from a shuffled verse (wrong Hebrew) must give a much worse score.

3. Tokenisation specifics

• יְיָ is a token of its own (lemma = the divine name, the same as Hebrew יהוה); in unpointed text it is the string יי, also inside words: דַּיְיָ = ד־ + יְיָ, וַיְיָ = ו־ + יְיָ, לַייָ.
• The object marker יָת takes suffixes (יָתֵהּ, יָתְהוֹן, יָתְכוֹן); split as יָת + pronoun, as UD does for Hebrew אוֹתוֹ.
• Editorial notes in parentheses inside the printed text (Gen 28:20 "(נ''י לְחֵם)", Exod 19:4 "(נ״א וְאַטְּלִית)") are removed from the sentence and kept in MISC.
• ד־ and דִּי: both occur; the separate דִּי is a word, the prefix is split.

4. What the targums can teach the SLM (Mova's small language model)

1. A bilingual dictionary for free. Word alignment of Onkelos gives a Hebrew–Aramaic glossary with frequencies, verse examples and an English column from the JPS 1917 translation (already in the corpus).
2. Paraphrase detection. The departures of Translation technique and syntax of the targums §3–5 are a labelled set of "translation versus interpretation", decidable by explainable cues (added words, replaced metaphors).
3. Pragmatics. The same event told with and without anthropomorphism is a natural test of event structure: who acts (God / the Memra), who perceives (sees / "it is revealed before"), what changes (repents / "turns back in his Memra"). The school's pragmatic frames can be checked against both texts.
4. A bridge to the Talmud. The targumic vocabulary is the core of all Jewish Aramaic; the differences from JBA and JPA are few and listed (The lexicon of the targums §5).

5. Licences

The Sefaria Onkelos, Jonathan and Writings targums are public domain and can be used to build. The Taj edition (CC BY-SA) is the best vocalised text but must stay out of what is built and shipped; it may be used to test and measure. The Metsudah edition (CC BY-NC) is for tests only. Sperber's critical edition (Brill) is in copyright and not on disk. See Aramaic in Mova: what it would need §7.

6. Open questions

• Do the Sefaria Onkelos verse divisions match the WLC everywhere? (To check before alignment.)
• How much of the Writings targums is JPA-like? The sentence classifier of Aramaic in Mova: what it would need §5 should be run on them and the result reported per book.
• Is the Melamed 1921 Song of Songs (six Yemenite MSS) a better witness than the Mikraot Gedolot text for training? Compare the two token by token.

The Elamur team

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/languages/arc-x-targum/mova/