Marathi Grammar Error Correction Using Fine-Tuned IndicBART for Low-Resource Language Processing

MarathiGEC, a transformer-based Marathi Grammar Error Correction framework based on IndicBART, is introduced in this paper. MarathiGEC framework helps in addressing the increasing demand for automation of grammar correction systems for Marathi, an Indian language that is morphologically complex and low-resource. Parallel sentences in the form of grammatically wrong and correct Marathi sentences were collected and employed for the fine-tuning process of the pre-trained IndicBART model. Using grammar correction as a sequence-to-sequence text generation problem helps in learning contextual and grammatical connections between grammatically wrong and correct sentences.   MarathiGEC framework is capable of correcting several kinds of grammatical errors including gender agreement, verb conjugation, tense, spelling, number agreement, and sentence structure errors. Contextual learning ability of the IndicBART model results in generating fluent and grammatically correct sentences without altering the semantics of the sentence. Besides, this study provides useful linguistic materials and sets a platform for further research in the area of Marathi Grammar Error Correction. Both the prepared data set and the trained model may be used in a variety of tasks, such as education websites, writing assistants, and content creation applications, as well as other related fields. The future work is planned to be directed at the data set expansion, taking into account the dialect differences, improvement of the model in regard to difficult sentence structures, and implementation of the framework as a web/mobile application. The proposed MarathiGEC framework is an important stage in development of Marathi NLP.

[1]   Arora, A., Kumar, V., and Ailani, N. (2019). An overview of correcting grammatical errors. 3536–3540 in International Journal of Scientific & Technology Research, 8(11).   [2]  Sonawane, S., Choudhury, M., & Mahata, D. (2020). Hindi grammar correction using inflectional error generation. 242–251 in Proceedings of the 17th International Conference on Natural Language Processing (ICON).   [3]    Bhattacharyya, P., and Sharma, V. (2025). Hi-GEC: A standard for correcting grammatical errors in Hindi. Proceedings of the Association for Computational Linguistics' (ACL) 2025 Annual Conference.   [4]   Joshi, A., Patil, M., and Magdum, P. (2023). MahaNLP is an open-source Marathi natural language processing library that uses models based on MahaBERT. 112–121 in Proceedings of the 2023 Conference on Computational Linguistics.   [5]   Gawali, M., Joshi, A., and Kulkarni, S. (2005). A Marathi spellchecker based on morphology. 67–74 in Journal of Indian Linguistics and Technology, 4(2).     [6]    R. Joshi (2022). L3Cube-MahaCorpus and MahaBERT: Resources, Marathi BERT Language Models, and Marathi Monolingual Corpus. arXiv preprint arXiv:2202.01159. [7]        K. Kumarasinghe, G. Dias, and I. Herath(2021) "SinMorphy: A Morphological Analyzer for the Sinhala Language," in 2021 Moratuwa Engineering Research Conference (MERCon), Moratuwa, Sri Lanka, 2021, pp. 681-686, doi: 10.1109/MERCON52712.2021.9525636.   [8]     Ujjwal Sharma and Pushpak Bhattacharyya. 2025. Hi-GEC: Hindi Grammar Error Correction in Low Resource Scenario. In Proceedings of the 31st International Conference on Computational Linguistics, pages 6063–6075.   [9]     Raj, J. (2024). Development of a grammatical error correction tool for the low-resource Gujarati language. Bachelor’s Thesis, Innopolis University. DOI: 10.13140/RG.2.2.21773.76003   [10]    K. B. Khandale and C. N. Mahender(2020), “Natural Language Processing based Rule Based Discourse Analysis of Marathi Text,” in 2020 International Conference on Electronics and Sustainable Communication Systems (ICESC), Coimbatore, India, Jul. 2–4, pp. 356–362, doi:10.1109/ICESC48915.2020.9155653.