Grammar grading that shows its work: a shared evidence graph ties every Chinese-language mark to a rule certificate
Synopsis
Grammar assessment for international Chinese education is recast as proof-carrying constrained inference: retrieved rules, error spans, minimal repairs and ordinal grades are coupled through one latent evidence graph, and every score-affecting judgment must carry a rule-level support certificate. On public corpora including MuCGEC, PC2G reaches an error-detection F0.5 of 71.4 against 67.6 for the strongest listed retrieval baseline, CRAG-style, while five-level grade MAE falls from 0.57 to 0.42 and evidence faithfulness reaches 91.8%. Those grades are measured against a silver rubric derived from reference edits, not independent teacher ratings.
Interpretation
Automatic Chinese grammar assessment no longer stops at flagging a malformed sentence; it must identify the construction under test, separate acceptable variants from true violations, assign a rubric-consistent grade, and justify the mark with inspectable pedagogical evidence. Public Chinese grammatical error corpora typically label error spans and corrected sentences rather than curriculum-specific mastery grades, leaving a gap between sentence repair and defensible educational judgment. MuCGEC, FCGEC, NaSGEC and CGED-2020 are all listed as lacking teacher grade annotations, so the five-level evaluation is defined by a silver rubric built from reference edits and fixed construct weights.
PC2G treats retrieval, error diagnosis, minimal correction and ordinal grading as one structured decision rather than a sequence of separately optimized text generators. Conventional retrieve-then-generate architectures give access to relevant rules but do not guarantee that those rules license an individual diagnostic or grading decision. In the listed experimental configurations, PC2G reports detection F0.5 of 71.4 and correction F0.5 of 57.8 against 67.6 and 53.0 for CRAG-style retrieval; grade MAE is 0.42 versus 0.57, QWK 0.87 versus 0.80, and evidence faithfulness 91.8% versus 82.1%.
Each diagnostic claim is represented by a rule-span-edit support path, the score is solved under rubric and evidence constraints, and feedback is generated only from accepted certificates. Evidence deletion, conflicting-rule substitution and learner-response paraphrases serve as training signals that separate evidence necessary for a grade from material that merely resembles the input. Removing the typed proof graph lowers detection F0.5 from 71.4 to 68.3 and proof validity from 91.8% to 78.0%; removing exception edges lowers proof validity to 83.1%.
When irrelevant rules are injected into retrieval, proof-constrained grading keeps error detection comparatively stable. Conventional retrieval-augmented methods are vulnerable to long-context dilution and noisy or irrelevant documents. With 50% irrelevant retrieved rules inserted, PC2G reports detection F0.5 of 66.3 against 50.9 for vanilla RAG; at 80% automatic grading coverage, PC2G reports a silver-label MAE of 0.29.
Perspective
The reported evaluation covers the MuCGEC and CGED-2020 learner corpora, with FCGEC and NaSGEC used as native-speaker domain-shift analyses, and the grade results correspond to a silver rubric derived from reference edits. It is aimed at researchers and engineers working on retrieval-augmented generation, language assessment and educational measurement.
The reported metrics are point estimates, and raw outputs, exact model checkpoints, multi-seed distributions and paired uncertainty intervals are not provided, so the comparisons are descriptive rather than claims of statistical significance. Grade results concern a rule-derived silver rubric, not independent teacher ratings or learner-level mastery; counterfactual operators rely on validated meaning-preservation assumptions, and proof validity depends on the quality and currency of the grammar knowledge base.