Establish cross-linguistic and cross-domain transfer of taxonomy instructions and batching

Determine whether the ERRANT-based taxonomy instructions and batched inference strategy developed for English learner essays transfer effectively to morphologically richer languages, lower-resource settings, professional editing, and fluency-oriented grammatical error correction benchmarks such as JFLEG.

Background

The experiments evaluate English learner essays using minimal-edit grammatical error correction benchmarks, principally BEA-2019 and CoNLL-2014. The taxonomy is based on 25 ERRANT error categories, while batching is studied as a mechanism for reducing overcorrection under a precision-weighted objective.

The applicability of these design choices beyond this setting is unresolved. In particular, richer morphology, limited training resources, professional editing contexts, and fluency-oriented benchmarks may alter both the usefulness of the fixed taxonomy and the behavioral effects of packing multiple sentences into a single model context.

References

It remains unclear how our taxonomy instructions and batching transfer to morphologically richer languages, lower-resource settings, professional editing, or fluency-oriented benchmarks like JFLEG \citep{napoles2017jfleg}.

Larger Context Window, Fewer Overcorrections: Optimizing Prompts and Batching for Minimal-Edit Grammatical Error Correction  (2609.10810 - Karpo et al., 9 Sep 2026) in Section ‘Limitations’, item 4