Automatic Assessment and Grading of Chinese Grammar Knowledge for International Chinese Education Using Retrieval-Augmented Generation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Accurate assessment of Chinese grammar in international Chinese education requires more than detecting a malformed sentence: a system must identify the linguistic construction being tested, distinguish acceptable variants from true violations, assign a rubric-consistent grade, and justify feedback using inspectable pedagogical evidence. While retrieval-augmented generation (RAG) provides access to external grammar resources, conventional retrieve-then-generate architectures do not guarantee that retrieved rules license individual diagnostic or grading decisions. We formulate grammar assessment as proof-carrying constrained inference, and propose PC$^2$G (Proof-Carrying Counterfactual Grading), a model in which rule retrieval, error-span inference, minimal repair, ordinal scoring, and uncertainty-aware feedback are coupled by a shared latent evidence graph. Candidate evaluations must carry rule-level support certificates and satisfy counterfactual necessity, paraphrase consistency, rubric monotonicity, and abstention constraints. A variational objective combines evidence posterior learning with structured prediction and a distributionally robust counterfactual term; a conformal audit layer is designed for deployment-time risk control when calibration assumptions hold. We detail an evaluation protocol using publicly released Chinese learner and native-speaker grammatical error corpora, clarify which tasks their gold annotations actually support, and propose a separate human-grade annotation extension. The evaluation compares retrieval-assisted baselines on Chinese grammatical error corpora, and analyzes graded proxy agreement, evidence faithfulness, retrieval noise, counterfactual stability, and selective prediction.