Prasta Nepali: A Transformer Based Approach for Automated Nepali Grammar (Byakaran) Error Detection and Correction

Grammatical error checking is a crucial research area in Natural Language Processing (NLP), helping to refine text and assist language learners. High-resource languages such as English have made substantial advancements while low-resource languages, including Nepali, continue to lag due to the scarcity of annotated datasets, linguistic tools, and computational resources. Few studies have focused on error detection and correction in Nepali grammar (Byakaran, the Nepali term for grammar), with the majority of works focusing on rule-based techniques rather than modern deep learning methods. This has limited its applicability to simple Nepali sentences with constrained sentence structures. This research introduces Prasta Nepali, a system that uses transformer-based deep learning models to tackle these issues. An encoder-decoder framework is used to compute the contextual relationships, based on which the model identifies and corrects grammatical errors in text. Tests against stacked LSTMs and Bi-LSTMs with attention show the Transformer model works better achieving 90.45% accuracy in training and 92.15% in validation. Also, Bilingual Evaluation Understudy (BLEU) scores of 0.9037, 0.8170, 0.7838, and 0.7694 for 1-gram to 4-gram matches between corrected and ground-truth sentences compute the similarity score, proving how well the system corrects the erroneous sentence. The findings show that transformer models can push forward grammar checking in Nepali giving useful insights and opening doors to future NLP applications.