Refining Word-Based Grammatical Error Annotation for L2 Korean

Korean grammatical error correction (K-GEC) presents a structural mismatch between word-based evaluation and the morpheme-level locus of many learner errors. Postpositions and verbal endings are bound to lexical hosts, but they encode grammatical relations that must be represented in correction and evaluation. This paper refines word-based grammatical error annotation for L2 Korean by addressing three connected problems in existing resources: surface target realization, Korean-specific edit annotation, and single-reference evaluation. We reconstruct target sentences from the National Institute of Korean Language (NIKL) L2 corpus under morphologically constrained realization rules and convert its morpheme-level annotations into word-level m2 edits. We then define a Korean ERRANT-style annotation scheme that preserves the MRU core while distinguishing functional morpheme errors, spelling errors, word boundary errors, and word order errors. We also augment the KoLLA corpus with an additional reference correction, yielding a multi-reference evaluation setting for Korean GEC. Empirical validation shows that the refined NIKL targets yield lower perplexity, the converted m2 files achieve higher agreement with source-target edit representations, and the refined resources improve KoBART-based correction under the same model setting. Multi-reference KoLLA evaluation further reduces the penalty imposed on valid corrections that diverge from a single reference, especially for neural and prompted GEC systems. These results show that Korean GEC evaluation depends not only on correction models, but also on reference data and edit annotations that reflect Korean morphology, spacing, and correction variability.

Many Korean learner errors sit inside the eojeol, in postpositions and verbal endings bound to a lexical host, while GEC is trained and scored on surface words. The paper addresses three connected problems:

1. Target reconstruction (NIKL). The NIKL L2 corpus gives corrections only as morpheme sequences, so target sentences must be rebuilt. Contraction and irregular-conjugation rules are applied only to predicates and only before endings that license the alternation (avoiding errors such as 아빠에게 → *아빠게); the few lemmas with ambiguous conjugation classes (e.g. 묻다 ‘bury’ vs. ‘ask’, 22 occurrences) are checked by hand.
2. Korean ERRANT-style edits. Morpheme-level annotations are converted into word-level m2 edits that keep ERRANT's Missing / Replacement / Unnecessary core but expose bound morphology, e.g. R:NOUN -> NOUN+ADP for a missing particle, with ADP for postpositions and PART for verbal endings. Further subtypes mark spelling (R:SPELL), word order (R:WO), and word boundary (R:WB) errors, assigned by a deterministic classifier.
3. Multi-reference KoLLA. The 100-essay KoLLA learner corpus receives a second, independently produced reference correction that follows a minimal sentence-level correction principle, alongside the original, more context-enriched layer.

Example (Figure 3). The prior conversion records the particle fixes but leaves the predicate uncorrected; the refined m2 file also corrects the predicate and marks 때 as unnecessary:

S 그래서 온돌은 사용하기 전에 힘들고 편리하지 않지만 사용 후에 온도 나왔을 때 편하다 .

(a) prior:
A 3 4|||PARTICLE|||전에는|||REQUIRED|||-NONE-|||0
A 9 10|||PARTICLE|||온도가|||REQUIRED|||-NONE-|||0
target: *그래서 온돌은 사용하기 전에는 힘들고 편리하지 않지만 사용 후에 온도가 나왔을 때 편하다 .

(b) refined:
A 3 4|||R:NOUN+ADP -> NOUN+ADP|||전에는|||REQUIRED|||-NONE-|||0
A 9 10|||R:NOUN -> NOUN+ADP|||온도가|||REQUIRED|||-NONE-|||0
A 10 11|||R:VERB -> VERB|||올라가면|||REQUIRED|||-NONE-|||0
A 11 12|||U||||||REQUIRED|||-NONE-|||0
target: 그래서 온돌은 사용하기 전에는 힘들고 편리하지 않지만 사용 후에 온도가 올라가면 편하다 .
('So, the ondol is difficult and not convenient before use, but it is comfortable when the temperature rises after use.')