New reward system evaluates grammar corrections by final quality

Don't Count the Edits, Judge by the Outcome Alone: Reward-Based Evaluation for Grammatical Error Correction

Computation and LanguageArtificial Intelligence

Summary

Grammatical error correction tools usually check if a corrected sentence matches a reference version, which can unfairly judge other valid fixes. The authors created SURE, a new way to score corrections based on how well the output fixes the grammar, keeps the original meaning, and sounds natural, without relying too much on exact matches. SURE learns from examples that include both small fixes and complete rewrites and can identify where errors are fixed in the original sentence. Tests show SURE works well compared to existing methods, especially for more creative corrections.

What this means in practice

Authors

Hayeong Ryu, Sunhee Jo, Seunguk Yu, YoungBin Kim

Abstract

Grammatical error correction (GEC) evaluation has traditionally relied on reference or edit overlap, which can penalize valid rewrites that differ from gold corrections. Reference-free metrics reduce this dependence, but evaluating whether a fluent output is a valid correction of the source remains challenging. We propose SURE, a source-conditioned reward evaluator trained on within-source preferences spanning minimal-edit and rewrite-oriented corrections. SURE jointly learns an overall reward with criteria-level supervision for grammaticality, faithfulness, and fluency, together with span-level grounding for source-side error resolution. Experiments on SEEDA show that SURE performs competitively against strong baselines, with particular gains on rewrite-style corrections and more disentangled criteria-level diagnostics. Our code is available at https://github.com/hayeonggg/SURE.