Elamur Groupantsmova-parse-dispatch
The groupant beside tools/mova-parse/src/lib.rs in the Mova repository: what the code computes, what it never does, how to go back from its outputs, its invariants, cost and checks.
Facts
Implementation material. A Mova code model describes this code at one source revision: partial models, finite checks and their stated scope. It is not a reviewed general exposition, and its labels are shown as written, never raised.
Components
| component | label | where |
|---|---|---|
| Approximant | partial | Approximant |
| Ast | partial | syntax.json |
| Cost | symbolic source derived | Cost |
| Descriptant | present | Descriptant |
| Invariant | partial | Invariant |
| Tests | executed bounded checks | Tests |
Obernants
The typed relations around the forward model: the behaviour complement (what the code never does), the reverse reconstruction (every input compatible with an output) and specialised views.
| relation | type | label | where |
|---|---|---|---|
| Behavior complement | behavior complement | partial source derived model | Behavior complement |
| Reverse reconstruction | reverse reconstruction | partial source derived model | Reverse reconstruction |
| Matrix composition | matrix composition | partial source derived model | Matrix composition |
Model text
Descriptant
The two-level dispatcher: sentences of a line; level 1 (langseg::Segmenter with the Latin table) gives language
spans; the pieces of the sentence's language are parsed together by its model, each longer switched piece by its own
model with its root on the matrix root (dep, CSAttach=root), short switched pieces inside the matrix parse marked
CSParse=matrix; a language without a model gets unannotated tokens with Parse=None. The CoNLL-U writer adds
# lang, # dialect, # style, # parse, # spans.
sentence_as / parse_as parse as one language the reader chose: one piece, level 1 not asked (the parse service's
language selector). The style tagger is Send + Sync, so a dispatcher is shared between threads.
Approximant
parse(text) = concat_lines concat_sentences sentence(s); sentence(s) = compose(matrix(annotate_lang(tokens of pieces with l = lang or short)), {annotate_l(tokens of piece) | l ≠ lang, long}); lang = argmax_l letters(l).
Invariant
Every token comes from exactly one piece; pieces cover the sentence (first starts at 0, last ends at its length,
neighbours touch); no token of a language without a model gets a head; each annotated unit has its own tree, and only
a switched unit's root is re-attached; a French piece of min_span words or more never reaches the English model.
Cost
Per sentence: level-1 O(n·K) plus each model's annotation on its unit (matrix once, each switched unit once).
Behavior complement
It does not parse the attachment between a switched span and its matrix; it does not merge tokenizers; it does not decide Hebrew-script varieties (no Hebrew-script model loaded) nor parse them; it does not fill the style without a detector.
Reverse reconstruction
The # spans line and # text recover the pieces; Lang=/CSParse/CSAttach recover the routing of each token; a
# parse = none sentence recovers only its tokens.
Matrix composition
Removing switched spans from the matrix unit keeps the matrix clause contiguous for its parser; the cost is a lost
argument (the quote) in the matrix tree, which the placeholder dep edge marks.
Tests
sentences_and_pieces, no_model_is_reported_not_parsed, a_forced_language_skips_level_one (with level 1 itself as
the control); a two-run identity check on 60 Global Voices lines
(2026-11-06).
Provenance
This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/tools/mova-parse/src/lib.rs.groupant/