A Multi-Task Transformer Ensemble for Explainable English Language Learning: Unifying Grammatical Correction, Tense Prediction, and CEFR Proficiency Grading
Hồ Quốc Dũng, Lien Thi Quynh Le, Le Hung, Vu Tran · Sakarya University Journal of Computer and Information Sciences · 2026
Not cached yet. This record was fetched live for this request and is stored after the lookup. Reopening the page should serve the cached copy.
Sakarya University Journal of Computer and Information Sciences · OpenAlex · article · 2026
Abstract
This paper presents a multi-task transformer ensemble that unifies Grammatical Error Correction, Tense Prediction, and CEFR-level Proficiency Grading within a single explainable framework for English language learning. The system integrates fine-tuned transformer models—Flan-T5, BERT, RoBERTa, and DistilBERT—through a confidence-weighted ensemble to generate grammatical corrections, verify tense consistency, and predict learner proficiency levels. The framework was trained and validated on large-scale linguistic datasets, including BEA-2019, C4, PTB, BNC, EFCAMDAT, and CLC, to improve robustness across writing domains.
Experimental results show that the Flan-T5 GEC module achieved BLEU = 32.4 with a 36.8% reduction in training time, while the RoBERTa tense classifier reached F1 = 0.99. For CEFR grading, the model achieved κ = 0.90 , indicating strong aggregate agreement with human ratings, although performance at the most advanced proficiency level remained more challenging. Compared with non-integrated module outputs, the ensemble further improved overall linguistic accuracy and interpretability, with statistically significant gains in key aggregate metrics (p < 0.05).
These findings suggest a hierarchical dependency among grammar, tense, and proficiency, while provid- ing an interpretable AI mechanism for delivering personalized and pedagogically meaningful feedback.
Open the resolved locator · 10.35377/saucis...1830952
Metadata licence: OpenAlex metadata (CC0)