🗣️ What every language gets

Mova takes on a language with the same package, so that every language reaches the same footing:

1. A grammar manual in languages/<code>/ (a variety as <code>-x-<variety>/ or as its own chapters), written in our own words from public-domain grammars and open research. Its examples come from the corpora Mova holds. Every chapter ends with how Mova handles the point and where it stops. sources.md lists the sources; works under share-alike or non-commercial terms are cited and paraphrased, never copied. The model is hbo.
2. Papers on the language's linguistics and NLP, each with its card on Elamur Research.
3. Open annotated corpora: treebanks, coreference, frames, named entities, discourse, with their licences (Data and model licences). Non-commercial and research-only data is for tests and measurements only.
4. Wiktionary, raw: the language's own Wiktionary and the English Wiktionary's entries for it. It is CC BY-SA, so it is for learning and tests only.
5. Open parsers as witnesses: Stanza, Trankit, UDPipe, spaCy and the language's own (on the GPU or not), set up as in Open parsers: witnesses and tests.
6. A baseline of Mova's own parser on the language's UD treebanks (en xpos-fill, then en parse-eval: train → dev), with LAS for its own tags and for gold tags, dated.
7. A place in the multilingual parser (tools/mova-parse): a level-1 lexicon for langseg (Latin script: langseg latin-lexicon from open train splits) and a level-2 model file <code>.bin (en train on the train split after en xpos-fill). Text in several languages is then parsed span by span, each by its own model, in one CoNLL-U document with # lang per sentence.

Baselines (UD 2.18, train → dev, 2026-10-22)

French GSD again, 2026-11-06, with the model the Elephant's French pass uses: LAS 87.09 (our tags) / 89.78 (gold tags), UAS 91.12 / 92.59, on the parser's relation set (French subtypes read as their base); 78.80 with the French subtypes.

The Ukrainian rows were measured on 2026-11-06 with the same command pair (en xpos-fill, en parse-eval; the model is trained in memory and not kept). IU is CC BY-NC-SA 4.0, so its figures are a measurement only; ParlaMint is CC BY-SA 4.0. The neural witnesses for Ukrainian are in Open parsers: witnesses and tests.

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