The Mova dialect of Universal Dependencies (`dialects/mova`)

mova is Mova's own internal variant of Universal Dependencies. Treebanks and annotation schemes that differ in conventions — UD releases, the customs of individual treebanks, SUD, Stanford Dependencies, classic Penn Treebank conversions — are converted into mova; the English tagger/parser en is trained on it; and its output is converted out of mova into any target (for example EWT 2.18 or the latest UD). Quality is measured against each treebank's own gold standard through the reverse converter.

What mova consists of:

Base: UD 2.18 with English Web Treebank (EWT) conventions (the feature/relation registry is en/data/ud-registry-en.tsv).
Layers on top of the base:
- ud-next — guideline changes made after the 2.18 release that are expected in 2.19 (dialects/ud-next/);
- selected borrowings from alternative dialects;
- own normalizations: det:poss → nmod:poss; :tmod/:npmod → :unmarked; verb FEATS (Number, Person) taken from the subject.
Selection principle: prefer the more consistent and more informative analysis, provided it can be converted to the standard without loss. A lossy scheme is only ever an export target, never the base.

How converters work

Conversion rules are ```convert blocks in convert.md files under treebanks/ and dialects/ (and occasionally inside a seed). They use the same matching language as the English grammar seeds (match, require, unless; see en/seeds), plus from/to (source and target dialect) and set actions that change a node's relation, UPOS, XPOS, lemma or features (feats+=X=V, feats-=X, lemma=@form, feats+=X=@v to copy a feature from another node):

rule: ewt.imperative-do-main
what: main VB with imperative "do" gets Mood=Imp|VerbForm=Fin
from: ewt
to: mova
match: v[xpos=VB, feats.VerbForm=Inf]; d[lemma=do, rel=aux, head=v, feats.Mood=Imp]
set: v[feats+=VerbForm=Fin; feats+=Mood=Imp]
source: EWT 2.18

Application is deterministic:

rules for a from → to pair run in file order (paths alphabetically) and block order within a file;
for each rule, all bindings are found on the current state of the sentence (all require clauses hold, no unless clause holds); action values are read from that same state, and then all actions are applied together;
two actions that set different values on the same field of the same word are an error (each FEATS feature and each MISC key counts as a separate field);
after all rules, a tree gate checks that heads are in range, there is exactly one root, root is used only on it, and there are no cycles.

Lines the actions do not touch are carried over byte for byte: comments, multiword tokens, empty nodes, DEPS, MISC, and features outside the known feature set. The engine is run as en convert; en convert-check reports what a rule set changes on a corpus.

The seeds in this folder

Each seed in Seeds: dialects/mova records one mova convention where UD 2.18 guidelines, post-2.18 changes and EWT data could point in different directions. Each entry below states what mova does, its scope, examples, and how it relates to standard export. Status values: adopted (part of mova), pending (mova keeps current EWT practice until UD settles the question), not applicable.

Only three conventions change mova annotation relative to EWT 2.18: case and lemma on who-pronouns (wh-case), adverbs moved out of discourse (discourse-adverbs), and imperative main verbs that EWT marks Inf (imperative-do). The rest confirm that mova follows EWT data where the English guideline text is out of date.

clausal-predicate

dialects/mova/seeds.md#clausal-predicate

Status. Adopted. Gist. When the predicate of a copular sentence is a clause, mova follows the UD 2.10 amendment and EWT: the predicate of the inner clause is the head, be is cop, and the subject of the outer clause is nsubj:outer (or csubj:outer for a clausal outer subject). The older analysis with the copula as head and the clause as ccomp is not used. Conditions and exceptions. The copula stays cop in all sentences, with no "transitive copula" exception. Be with ccomp remains only for quotative be like and constructions such as there are hints … that. Export to UD 2.18 is identical; the older analysis can be derived deterministically from nsubj:outer + cop if a conversion to UD v1 or Stanford Dependencies needs it. Examples. The problem is that these sentences are difficult: head difficult, is — cop, problem — nsubj:outer, sentences — nsubj. Sources. UD 2.18 docs/changes.md (Multiple subjects); UD docs/_en/specific-syntax.md (older analysis); checked by en.verbal.outer-subject-needs-cop and en.verbal.cop-verb-head-outer (en/seeds/verbal).

depictives

dialects/mova/seeds.md#depictives

Status. Adopted. Gist. An optional depictive adjective (a secondary predicate describing a participant's state during the event) is advcl of the verb, as in the UD 2.10 amendment, _en/dep/advcl.md and EWT data. The older acl and advmod analyses are not used. Conditions and exceptions. The depictive's own subject (e.g. she), where visible, is expressed in Enhanced UD, not in the basic tree. Examples in _en/specific-syntax.md involving unable and assured are annotated differently in EWT (parataxis and verbal advcl) and should not be used as depictive examples. Export to UD 2.18 and ud-next is identical. Examples. She entered the room sad. — came back dead — came into office obsessed with Iraq. Sources. UD 2.18 docs/changes.md (Optional depictives); UD docs/_en/dep/advcl.md; UD docs/_en/specific-syntax.md.

discourse-adverbs

dialects/mova/seeds.md#discourse-adverbs

Status. Adopted. Gist. discourse is reserved for interjections, discourse particles, list numbering and emoticons, following the post-2.18 definition. An adverb is never discourse but advmod; a prepositional phrase is never discourse but obl. Conditions and exceptions. discourse remains for INTJ (oh, well, like, um, yes), item numbering (NUM: 1., (a)), emoticons (SYM), thanks, ps. The criterion is the word class, visible without interpreting pragmatics. Conversion EWT/GUM → mova uses rule mova.ud-next.discourse-adv (Converter mova ⇄ ud-next); mova → 2.18/ud-next needs no change, since the 2.18 guideline already excluded adverbs. On a round trip, the few EWT/GUM tokens annotated discourse on adverbs differ from the gold file; reverse-conversion reports should list them on a separate line. DEPS still carry N:discourse and need a DEPS action or an EUD rebuild. Examples. advmod: though, so, also, maybe, btw; obl: in other words; discourse: oh, well, um, yes, 1., (a). Sources. UD ud-next _u-dep/discourse.md; UD 2.18 discourse guideline ("non-adverbial discourse markers"); gates udnext.discourse-not-adv, udnext.discourse-not-pp (error).

feat-layer-ids

dialects/mova/seeds.md#feat-layer-ids

Status. Adopted. Gist. Any layered feature identifiers that mova introduces (for example, for agreement with several participants in other languages) are named with lowercase Latin letters only, [a-z]+, without digits. Conditions and exceptions. The post-2.18 guideline requires this, although the validator still accepts digits; digits are expected to become an error. Nothing changes for English, which has no feature layers. Examples. None in the seed (non-English illustration: an Old Georgian layer renamed sauf2 → dsauf). Sources. UD docs/_u-overview/feat-layers.md; UD validator utils.py.

features-context

dialects/mova/seeds.md#features-context

Status. Principle adopted; Exponence pending. Gist. Feature values are taken from context, but where context and word form conflict, the form wins (post-2.18 clarification no. 18). mova keeps its "FEATS from the subject" normalization with an explicit limit: agreement never overrides an unambiguous form. VBZ is always Number=Sing|Person=3; VBP is never Sing|3; was is always Sing. Conditions and exceptions. With notional agreement (the team are, a number of people are) the verb gets Plur|3 from its form rather than a value the standard does not allow. VBD and VBP take Number/Person from the subject, and you/it get Case by position, as in EWT. The Exponence MISC attribute is not used until UD standardizes its values. Export to 2.18 is unchanged. Examples. the team are → Plur|3 on are; a number of people are. Sources. UD docs/changes.md (Morphosyntactic features, clarification no. 18); UD issue #1233; gates udnext.vbz-3sg, udnext.vbp-not-3sg (applied to en output and silver data).

imperative-do

dialects/mova/seeds.md#imperative-do

Status. Adopted. Gist. In an imperative with do, both do and the main VB verb have Mood=Imp|VerbForm=Fin — following EWT practice rather than the letter of _en/feat/VerbForm.md (which would give the main verb Inf). So Go! and Don't go! are annotated the same way on the lexical verb. Conditions and exceptions. The minority of EWT tokens with VerbForm=Inf in this construction are aligned by the EWT → mova rule ewt.imperative-do-main (shown in the introduction); there is no reverse rule, so these tokens appear as a separate line in reverse-conversion reports. The literal VerbForm.md form is derived deterministically: a VB with aux do carrying Mood=Imp → VerbForm=Inf, Mood removed. Don't with Mood=Ind in imperatives is an EWT error. Examples. Don't worry. — Do not hesitate. — DO NOT GO HERE. — Don't cling. — don't let that fool you. Sources. UD docs/_en/feat/VerbForm.md; EWT 2.18 data; English seed Imperative mood (en.verbal.imperative-do-main-verb; en.verbal.vb-aux-infinitive excludes imperative do).

iobj-sole

dialects/mova/seeds.md#iobj-sole

Status. Adopted. Gist. Following the UD 2.12 "Sole iobj" amendment and EWT (rather than the older text of _en/dep/iobj.md and _en/specific-syntax.md), iobj may be the only internal argument and may co-occur with xcomp when the verb takes an addressee or recipient. Conditions and exceptions. iobj + xcomp: ask/tell/allow/convince/urge/persuade/cause/trust/teach/remind/warn X to do. obj + xcomp: let, make, get, have, keep, find, want, help. The boundary between the classes is set by the lexicon (Verbs that take an indirect object, VerbNet), not by examples; permit and recommend, which have both analyses in EWT, follow the gold data until the lexicon rule decides. Export to EWT 2.18 is identical; iobj → obj (for the literal _en text) is deterministic. Examples. remind me — ask Bush — ask X to do (iobj) vs make X do (obj). Sources. UD 2.18 docs/changes.md (Sole iobj); UD docs/_en/dep/iobj.md; UD docs/_en/specific-syntax.md; VerbNet.

modal-mood

dialects/mova/seeds.md#modal-mood

Status. Pending (UD issues #1118, #1155). Gist. Modals (MD) keep the EWT convention: VerbForm=Fin without Mood, Number or Person. Mood=Pot/Nec/Cnd is not added, and neither is Mood=Ind. Conditions and exceptions. Mood=Pot/Nec/Cnd would make export invalid, since the English feature registry allows only Imp|Ind|Sub; Mood=Ind would wrongly assert indicative for might, would. The modal meaning is still available from the lemma: the auxiliary registry gives the function of each of the 16 auxiliary lemmas, and a clause-level "possibility/necessity" feature can be derived from it. The validator's verbform-fin-without-mood warning on these tokens is expected. Revisit when #1118 closes or _en/feat/Mood.md changes. Examples. None in the seed beyond the modal lemmas (can — possibility, must — necessity, would — conditional). Sources. UD docs/_u-feat/Mood.md; UD auxiliary data (data.json); UD issues #1118, #1155, #1233.

no-effect

dialects/mova/seeds.md#no-effect

Status. Accepted as is / not applicable. Gist. A list of post-2.18 changes that do not affect English mova annotation: general documentation of expl:pass and expl:pv; validator data; the English treebank list; other languages and the website; UD v1 labels in specific-syntax.md. Conditions and exceptions. The Perl script conllu_convert_uposf_to_xpos.pl is not part of the conversion pipeline (pipeline mechanics are in Rust). Examples taken from specific-syntax.md must have v1 labels replaced (dobj → obj, etc.). For a future language-independent layer, expl:pass/expl:pv (a subtype shared by three language groups), flat:redup and expl:rel are noted. Validator data are to be compared again before 2.19. Examples. None. Sources. Linked seeds in Seeds: dialects/ud-next and Seeds: dialects/ud-2.18.

spoken-subtypes

dialects/mova/seeds.md#spoken-subtypes

Status. Pending. Gist. The spoken-language subtypes documented for French after 2.18 — discourse:filler, discourse:tag, conj:reform — are not used in mova for now. Conditions and exceptions. They would be more informative (a filler um is not an interjection oh; a tag question is not ordinary parataxis) and could be exported losslessly by dropping the subtype (discourse:filler → discourse). They wait until the UniDive spoken-language working group's guidelines are merged into UD, or until spoken English treebanks are taken up; they are not in the English registry, so export would have to strip them. Examples. um (filler) vs oh (interjection). Sources. UD ud-next French spoken subtypes; UD issues #1290, #1280, #1273, #1289.

subjunctive

dialects/mova/seeds.md#subjunctive

Status. Adopted. Gist. The subjunctive is marked Mood=Sub, as in EWT, GUM and ParTUT, rather than following the note in _en/feat/Mood.md that maps it to Inf/Ind. Present subjunctive (a bare VB without an auxiliary, with its own subject, after suggest/insist/recommend that, it is vital that, etc.): Mood=Sub|Tense=Pres|VerbForm=Fin, person and number from the subject. Past subjunctive were with a singular subject: Mood=Sub|Tense=Past. Conditions and exceptions. Examples that EWT left as Inf are not corrected automatically; rule ud218.subjunctive-as-inf only suggests candidates for a correction pass. The literal Mood.md form is derived deterministically (VB Sub → VerbForm=Inf with Mood, Tense, Number, Person removed; VBD Sub → Mood=Ind), but the reverse is not deterministic, so Sub must live on the mova side, not the export side. Export to EWT 2.18 is identical. Examples. suggest that he see a doctor — it is vital that she be present — if I were — were I to. Sources. UD docs/_en/feat/Mood.md, docs/_en/feat/Tense.md; checked by en.verbal.subjunctive-present-feats (Subjunctive mood).

to-phrase-obl

dialects/mova/seeds.md#to-phrase-obl

Status. Adopted. Gist. A to-phrase attached to a verb (e.g. a recipient: gave it to me) is obl, as in EWT and the universal obl/nmod guidelines. The nmod reading in _en/dep/iobj.md is a UD v1 leftover. Conditions and exceptions. Examples taken from _en/dep/iobj.md into the English seeds annotate to me as obl. The validator checks the nmod/obl boundary (obl-should-be-nmod). Export is identical. Examples. to me (obl). Sources. UD docs/_en/dep/iobj.md; UD docs/changes.md (no. 17); universal obl, nmod pages.

wh-case

dialects/mova/seeds.md#wh-case

Status. Adopted (from ud-next). Gist. Whom has lemma who (whomever → whoever). Who/whoever get Case=Nom or Case=Acc according to their syntactic function; whom/whomever are always Acc, since the form wins. Whose with a noun (nmod:poss) is Case=Gen|Poss=Yes; independent whose is Poss=Yes. Conditions and exceptions. Case is taken from the base function of who; in a few free relatives and predicative uses it comes from the position inside the subordinate clause (edge cases described in Converter mova ⇄ ud-next, rules mova.ud-next.who-*). Export to 2.18 uses deterministic reverse rules (ud-next.who-case-drop, ud-next.whose-case-drop, ud-next.whom-lemma): remove Case and restore the lemma from the form; an EWT round trip is identical. The English registry already allows PRON Case=Nom|Acc|Gen and Poss=Yes. Examples. the man who I saw (Acc) vs the man who saw me (Nom). Sources. UD ud-next _en/pos/PRON.md; UD issue #517; gates udnext.whom-who-acc, udnext.who-case, udnext.whose-dependent (error); hints udnext.who-object-acc, udnext.who-subject-nom, udnext.whose-independent (warn); en.morph.whom-acc (Interrogative and relative pronouns: who, whom, whose, which, what, that).

xpos-not-empty

dialects/mova/seeds.md#xpos-not-empty

Status. Adopted. Gist. XPOS is never an empty string: an unknown XPOS is written as _. This is enforced by the CoNLL-U writer for all export targets, not only ud-next (validator 0.2.8, empty-string-in-xpos, which 2.19 treats as a level-2 error). Conditions and exceptions. Mainly relevant for future languages with free-string XPOS; in en, XPOS is always a Penn Treebank tag. The seed rule language cannot see raw columns, so the gate lives in the writer code, with a negative control: a line with empty XPOS must produce an error. _ is also valid in 2.18, so there is no loss. Examples. None. Sources. UD validator 0.2.8 (empty-string-in-xpos).

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/repo-docs/seeds/dialect-mova/

📘 The Mova dialect of Universal Dependencies (dialects/mova)