Natural-language grammatical-correctness parsing (morphology + syntax, no semantics)

[konard]

Context

Issue #47 requires "natural languages (including actual parsing of their
grammar with no semantic checks ... grammatical correctness, syntax
correctness and so on - should be fully supported)". Current support is
identification, segmentation, normalization, and script annotations
(src/natural_language.rs) - nothing parses grammar, so correctness cannot be
checked. See requirements.md R-5 and
solution-plans.md S-8.

Research (competitors-natural-language.md):
Grammatical Framework + Resource Grammar Library is the only surveyed system
doing exactly this job (parse-or-reject, no semantics, covers the ten target
languages, LGPL/BSD); UD supplies the morphosyntax vocabulary and corpora;
Wikidata lexemes (CC0) and UniMorph (CC BY-SA) supply word forms; LanguageTool
(LGPL, proven portable to Rust by nlprule) supplies explainable negative
checks.

Scope (staged)

1. Adopt UD's UPOS/UFeats/deprel inventory as link-type vocabulary for
   morphosyntax links; import CoNLL-U fixtures (e.g. via rs-conllu) with
   per-treebank license provenance.
2. Word-level correctness: morphological lexica seeded from Wikidata lexeme
   Forms (CC0; UniMorph optional second source); unknown/ill-formed tokens
   become recoverable is_error links, reusing the parse-recovery contract.
3. Sentence-level correctness: integrate GF RGL grammars compiled to PGF
   (start with the gf-core crate; evaluate a native PMCFG reader over the
   links network as the long-term path); grammatical sentences parse clean,
   ungrammatical ones carry error links - verify_full_match() thereby
   answers "is this grammatical?".
4. Explainable negatives: port a starter set of LanguageTool-style pass/fail
   rule sentences per target language as fixtures (DELPH-IN "mal-rule" model
   for explanations).

Acceptance criteria

• For each of the ten NATURAL_LANGUAGE_TARGETS: at least one grammatical
  fixture parses with a clean verify_full_match() and one ungrammatical
  fixture surfaces error links - while both reconstruct byte-for-byte.
• Morphosyntax links use the UD-derived vocabulary and are queryable.
• All imported data carries license provenance in fixtures.
• Staging is allowed (language-by-language), but each landed language is
  fully gated; remaining ones are tracked in the roadmap.
• Changelog fragment added (bump: minor).

References

• Requirement: requirements.md R-5
• Solution: solution-plans.md S-8
• Part of We need to carefully compare all competitors in all scopes with out project, and make sure we have the richest features set #47; work lands on branch issue-47-76af108c0f24 (PR Finish issue #47 parity feature set #48).

---

Filed from docs/case-studies/issue-47/proposed-issues/08-natural-language-grammar-parsing.md. Part of the implementation plan for #47.