UD 2.18 — the standard
Release 2.18 of 15.05.2026: 353 treebanks, 193 languages. English has 14 treebanks in it, the main one is EWT. This is the base of the mova dialect (The Mova dialect: a UD metadialect, treebanks, and converters): "UD 2.18 with EWT conventions".
What counts as the standard. Three layers, and they do not always agree:
- Guidelines:
markdown-source/from ud-documentation-v2.18.tgz. Universal —_u-*,u/; English —_en/. - The validator and its data: ud-tools-v2.18.tgz,
validate.py,udtools0.2.7,data/*.json. For English they are collected inen/data/ud-registry-en.tsv: 251 "UPOS × feature" pairs, 58 relations, 16 auxiliaries. The registry matches the 2.18 releasedata/*.jsonline for line. - EWT 2.18 data: UD_English-EWT, 16,622 sentences.
Where the _en text diverges from the data, mova takes the EWT data if they match the universal guideline. Below is each such discrepancy separately.
Discrepancies between the _en guidelines and the EWT 2.18 data
Numbers are train+dev+test. Each seed has counter rules, so one command reproduces the numbers:
target/release/en expert-check dialects/ud-2.18/seeds \
$MOVA_DATA/en/ud/ud-treebanks-v2.18/UD_English-EWT/en_ewt-ud-{train,dev,test}.conllu
| seed | the guideline says | EWT 2.18 | why |
|---|---|---|---|
| iobj-sole | iobj only with obj/ccomp; with xcomp — obj | 795 iobj: 243 without obj/ccomp, of them 93 with xcomp | 2.12 amendment "Sole iobj" |
| iobj-to-phrase | to me — nmod | to phrase on a verb: obl 1285, nmod 0 | v1 leftover |
| imperative-do | VB with an auxiliary — Inf | with imperative do: Mood=Imp|Fin 98, Inf 8 | EWT annotates by sentence type |
| subjunctive | subjunctive → Inf / Ind | Mood=Sub 57: VB 49, were 8; ≥5 subjunctives remained Inf | Mood.md not updated; Tense.md already mentions the subjunctive |
| modal-mood | every finite form has Mood | MD with Fin without Mood — 4049; validator warning | open question, #1118 |
| depictives | acl (:612) and advmod (:878) | ADJ depictive advcl 22, acl 0, advmod 0 | 2.10 amendment |
| clausal-predicate | the copula is the head, the clause is ccomp | nsubj:outer/csubj:outer 257, old analysis 0 | 2.10 amendment |
| wh-pronouns | TODO: Case and the lemma who for whom; Poss for whose | whom with lemma whom 23; Case 0 of 447; whose Poss=Yes 13 of 13 | the decision was made after 2.18 |
| discourse-adverbs | only non-adverbial markers | ADV with discourse — 9 (in GUM — 78) | boundary by meaning, not by word class |
| v1-labels | examples with dobj, nsubjpass, neg, mwe… (36 lines) | 0 | the text has not been updated since v1 |
Conclusion: mostly the _en text is out of date, while the EWT data follow later universal amendments. Genuine inconsistencies in the data themselves:
- 8 imperatives with
Inf; - ≥5 subjunctives with
Inf; - 9 adverbs with
discourse; - the lexical
iobj/objsplit withxcomp: permit and recommend have both variants.
Decisions for mova are in dialects/mova/.
Path notation. 2.18:docs/… — markdown-source/… from the 2.18 documentation archive. 2.18:tools/… — the 2.18 tools archive. docs/… without a prefix — the post-2.18 snapshot (docs). For all _en/ files except index.md and pos/PRON.md, line numbers are the same in both. #N — an issue in issues; it holds only issue bodies, without comments.
This page in Markdown → https://md.mova.elamur.ai/free/en/docs/dialects/ud-2.18/