Aramaic in Mova: what it would need

Status (2026-10-17): Mova has no Aramaic data parsed. There is no Aramaic treebank in treebanks/ except the tiny Neo-Aramaic aii (UD_Assyrian-AS, 57 sentences, 453 tokens, CC BY-SA 4.0), which is a modern language and not one of ours. The Hebrew side is further along: tools/macula converts the MACULA trees of the Torah into hbo dependency trees (125 807 words, 2026-10-17; macula).

This file is the plan for all five varieties. Each variety's own Aramaic in Mova: what it would need adds what is specific to it.

1. The codes

arc, arc-x-targum, tmr, jpa, arc-x-zohar — fixed in the corpus registry texts.tsv (column lang, with other languages of a mixed file in mixed) and explained in Aramaic of the Jewish texts — overview and Biblical Aramaic. In CoNLL-U the code goes in the document's # lang = line, or per sentence when a document mixes languages (# lang = tmr / # lang = hbo-x-mishnaic), and per word in MISC (Lang=hbo) for code-switches inside a sentence.

2. What exists on disk

All under the Scientist's raw store, $RESEARCH_RAW (outside the repository). Counts of whitespace-separated words were made on 2026-10-17 by flattening the Sefaria JSON text arrays with jq and stripping HTML tags; they include the Hebrew parts of mixed texts.

Dictionaries (dictionaries):

3. Tokenisation

Aramaic words carry clitics on both sides, like Hebrew, and the texts come in three spellings (pointed, unpointed, plene). What the tokenizer must do:

1. Proclitics. ו "and", ד "of, which", ל "to; object", ב "in", כ "like", מ "from" (מִן reduced), and in JBA ק (קָא) and the demonstrative ה. They stack: וּדְלָא "and which not", דְּבֵי "of the house of", וְכִדְנָה. In MACULA each is a word of its own; Mova's macula already follows that (macula). In BA the relative is still a separate word (דִּי); from the targums on it is the prefix ד־, and every word starting with ד is ambiguous (דְּהַב "gold" vs ד־ + הַב). A dictionary check against Jastrow headwords is the cheap first filter.
2. Enclitics. Pronoun suffixes on nouns, prepositions and verbs; in JBA also enclitic subject pronouns on participles (קָאָמֵינָא = קָא + אָמֵר + אֲנָא "I am saying") and the copula ניהו. These should be split into syntactic words, as UD does for Arabic and Hebrew.
3. The determined ending ־א / ־ה is not split: it is inflection (a feature, Definite=Def in UD terms), not a word. MACULA splits it as an article (Td); the converter should fold it back into the noun and set the feature. This is the one place where MACULA's segmentation should not be copied for Aramaic.
4. Ketiv/Qere. The Qere goes into the tree; the Ketiv into MISC (Ketiv=…).
5. Abbreviations (JBA and the Zohar): ת"ש = תָּא שְׁמַע, ת"ח = תָּא חֲזֵי, א"ר = אָמַר רַבִּי, ר"ש = רַבִּי שִׁמְעוֹן, קב"ה = קוּדְשָׁא בְּרִיךְ הוּא. A table of abbreviations per variety expands them before parsing; the abbreviated form stays in MISC.
6. Pointing. Strip points and accents for matching, keep them in FORM. The order of combining marks differs between sources; compare by consonants (as macula does for Hebrew lemmas).

4. Trees

4.1 The first Aramaic trees: Daniel and Ezra from MACULA

MACULA already has syntax trees for the Aramaic chapters, with lang = A on each word. Extending tools/macula with a language split would give the first arc treebank: about 7 500 morph units, a few hundred sentences, CC BY 4.0. What is needed:

• read lang per word and split each verse into its Hebrew and Aramaic parts (Dan 2:4 changes language mid-verse, Biblical Aramaic texts, annotated §4.1);
• write arc trees to their own file (DAN.arc.conllu, EZR.arc.conllu, JER.arc.conllu, GEN.arc.conllu);
• fold the determined ending into the noun (§3);
• map MACULA's Aramaic stems (peal, pael, haphel, aphel, shaphel, peil, hithpeel, hithpaal, hophal, ithpeel, ithpaal, hishtaphel) to a feature. The Hebrew converter uses HebBinyan; for Aramaic a neutral Stem= (or AramStem=) is clearer, with the same values in lower case as the manual (Stem=Peal);
• states: Definite=Ind (absolute), Definite=Cons (construct), Definite=Def (determined), as UD uses for Hebrew and Arabic;
• keep the same head rules as for Hebrew; add one for the pronominal copula (הוּא, הִמּוֹ as cop) and one for ל־ marking a direct object (case on an obj, not obl).

There is no gold UD Aramaic treebank to measure against. The check is therefore: (a) the tree validator (one head, one root, no cycle; a negative control must show red), (b) two runs give the same bytes, (c) a hand-checked sample of 50 sentences, annotated independently, before any number is claimed.

4.2 Targums by projection from Hebrew

The targums translate the Hebrew Bible verse by verse, and Onkelos often word by word (Translation technique and syntax of the targums). MACULA gives trees for every Hebrew verse. So:

1. align each Onkelos verse with its Hebrew verse (same reference: free);
2. align words inside the verse: by the regular correspondences (Hebrew אֶת ↔ Aramaic יָת, Hebrew ה־ ↔ Aramaic ־א, Hebrew בָּרָא ↔ Aramaic בְּרָא), by a Hebrew–Aramaic dictionary built from Jastrow's and BDB's cross-references, and by position;
3. project the Hebrew tree onto the aligned Aramaic words, and let a small rule set fix the known divergences (Aramaic additions such as מֵימְרָא דַּייָ, קֳדָם, determined state, the ית object).

This is the cheapest way to tens of thousands of Aramaic sentences with trees. Its accuracy must be measured on a hand-checked sample before it is trusted; projection errors are expected where the targum paraphrases (poetry, Genesis 49, Deuteronomy 32–33) and should be flagged by alignment confidence.

4.3 The Talmuds and the Zohar

No treebank exists for JBA, JPA or Zoharic Aramaic. The path: a tagger and parser trained on the arc and projected arc-x-targum trees, used as a first guess, with the vocabulary of Jastrow; then corrections and self-training. The Bavli's dialogue syntax (short questions, quotations, technical terms, The technical language of the sugya) is far from the Bible's narrative, so expect a large drop; measure it on a hand-annotated sample, never on the training material.

5. Language identification

Every corpus here mixes Hebrew and Aramaic: the Bavli (Mishnah and baraitot in Hebrew), the Yerushalmi, Daniel 2:4, the Zohar (Midrash ha-Ne'elam is mostly Hebrew), the commentaries (Hebrew with Aramaic quotations). The registry's lang and mixed are set by Sefaria category, not by sentence.

A per-sentence identifier is the first tool to build. Features that need no training (the comparison table in The Aramaic family and the Jewish varieties §5):

Score each sentence by counts of cue words (a deterministic, explainable classifier: it can say which cues decided), and measure on hand-labelled sentences from each corpus, including the hard cases (Hebrew baraitot quoted in Aramaic sugyot; Aramaic phrases inside Hebrew commentaries). A classifier that cannot be shown wrong on a planted Hebrew sentence in an Aramaic paragraph is not tested.

Built: tools/langseg uses this table as named cues beside learned word and affix features, splits mixed runs with a Viterbi pass, and explains each segment; its README has the measurements (2026-10-18: macro 83.0% over the eight Hebrew-script varieties on held-out tractates, chapters and verses; 81.0% on a hand audit of Talmud sentences).

6. The lexicon

• One Aramaic–English dictionary for all five varieties: Jastrow 1903 (public domain, 32 512 entries, already as JSONL). It covers the targums, both Talmuds and the midrash, and its entries cite sources (Targ., Y., B.), so each sense can be tagged by variety. BDB-Aram (863 entries) covers the Bible.
• Lemmas. Jastrow's headwords are pointed and sometimes numbered (קַמָּא, חֲמִי, כְּדוּן); key them by consonants, as macula dict does for Hebrew.
• Hebrew–Aramaic links. The regular correspondences (Script and phonology §4) and the cross-references in Jastrow and BDB ("BH", "h. text") give links between Hebrew and Aramaic lemmas. These are what the projection of §4.2 needs, and what a reader of Hebrew needs to learn Aramaic.
• Variety tags on words. The comparison words (חזי / חמי, השתא / כדון, איכא / אית) are few, but very frequent: they belong in a small hand-made table, not in a learned model.

7. Licences, in one place

Following Data and model licences:

• Build and ship: public-domain texts (WLC, Onkelos and Jonathan in Sefaria's PD versions, Pseudo-Jonathan, the Fragment targum, Zohar Ḥadash, Tiqqunei ha-Zohar vocalised) and dictionaries (Jastrow, BDB, Levy, Gesenius); CC BY data (MACULA, the Guggenheimer Yerushalmi) with attribution.
• As little as possible: CC BY-SA (the Wikisource Bavli, the Yemenite Taj Onkelos): test, measure, learn; not carried into a release.
• Tests only: CC BY-NC (William Davidson Talmud, Metsudah Onkelos).
• Not used: licence unknown (Venice and Mechon-Mamre Yerushalmi, Neofiti, the main Zohar's modern editions, the Sulam).

The ancient texts themselves are public domain; the restriction is on the digital editions. A PD source for the main text of the Zohar and the Bavli is therefore on the "wanted" list (Sources).

8. A plan, in order

Nothing in this plan is started (2026-10-17). Steps 4–6 are new directions and need the operator's word before they are run.

The Elamur team

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