Advancing Grammar Error Correction in informal digital communication: a hybrid CNN-BART approach with adaptive noise modeling

Grammar Error Correction (GEC) has emerged as a critical natural language processing task with applications in educational technology, automated proofreading, and digital communication enhancement. Although current transformer-based methods have shown remarkable results in formal textual settings, they struggle to handle the linguistic diversity and situational ambiguity of informal digital communication, including purposeful abbreviations, code-switching, and intentional grammatical violations. This study presents a hybrid CNN-BART model that effectively merges local pattern recognition with transformer-based contextual understanding for robust text error correction. The proposed system integrates curriculum-based noise injection, multi-scale convolutional encoding, and a comprehensive feature selection pipeline to address the challenges of informal text processing. Experimental validation on three benchmark datasets demonstrates consistent absolute F0.5 improvements of $$+6.5$$ points over Pillars of GEC on CoNLL-2014, $$+10.5$$ points over T5 on ErAConD, and $$+2.4$$ points over Pillars of GEC on BEA-2019, representing statistically significant advances ( $$p < 0.05$$ , bootstrap resampling, 95% CI). This work offers a practical framework for automated writing assistance in both academic and industry contexts.