MTAGEC: Multi-Task Arabic Grammatical Error Correction as a sequence generation with synthetic data

Automatic grammatical error correction (GEC) is challenging due to limited annotated data and complex linguistic features, specifically in Arabic. Existing GEC systems focus on detecting and correcting grammatical errors, overlooking the integration of explanations with the correction process. To bridge this limitation, we propose Multi-Task Arabic Grammatical Error Correction (MTAGEC), a framework that unifies grammatical error correction, error-type classification, evidence extraction, and error explanation into a single sequence-generation task, yielding both accurate corrections and human-readable explanations. To address data scarcity, we propose the Arabic Explainable Error Generator (AEEG) to construct ExplAGEC—a large-scale synthetic dataset comprising 21.8 million sentence pairs annotated with 25 diverse error types, including morphological, syntactic, and orthographic errors, along with corresponding explanatory annotations. Experiments conducted with Large Language Models for Arabic, including AraT5, AraBART, AraBERT, and ALLaM, show that training on ExplAGEC boosts correction quality on the QALB-2014/2015 and ExplAGEC benchmarks. Our best-performing MTAGEC variant based on AraT5 achieves an F0.5 score of 80.02 on QALB-2014 and 81.73 on QALB-2015, while simultaneously attaining 69.6% explanation accuracy on the ExplAGEC test, based on token-level alignment between model-generated evidence spans and gold-standard rationales. Multi-task GEC learning enhanced the correction and explanation quality, demonstrating that large-scale synthetic data and unified modelling enable interpretable Arabic GEC. The code and data are available at: https://github.com/Zainabobied/MTAGEC.