L1-Aware Grammatical Error Correction for ESL Writing: Performance–Disparity Trade-Offs

Abstract

Grammatical error correction (GEC) systems for learner English are usually trained as general-purpose models and rarely account for the writer's first language (L1), although L1 influences error patterns and may affect correction quality. To address this, this paper investigates whether lightweight L1 conditioning improves GEC for ESL writing. Specifically, using the FCE learner corpus, we focus on essays written by learners whose L1 is Chinese, Russian, or Spanish and compare BERT-based GECToR models in four settings: baseline, continued training, L1-aware input, and L1-aware training on a downsampled L1-balanced training set. For evaluation, we use ERRANT to compute precision, recall, and the precision-weighted F-score (F0.5) on both the original and a balanced test set. Our findings show that the L1-aware model obtains the best overall performance, whereas the balanced model yields the smallest performance gap across learner groups but lower overall precision and F0.5. Thus, the results suggest a performance–disparity trade-off in this setting: lightweight L1 conditioning improves average correction quality, while balanced training reduces cross-group score dispersion at the cost of overall F0.5.