NepAES: exploring the promise of automated essay scoring for Nepali essays

Sweta Poudel, Kritesh Rauniyar, Ashish Acharya, Junaid Rashid, Surabhi Adhikari, Usman Naseem, Surendrabikram Thapa

Abstract: Natural language processing (NLP) has been extensively studied and developed for the purpose of automated essay scoring (AES). This field of research has attracted significant attention and has been explored across multiple languages and machine-learning models. Researchers, over the years, have dedicated significant resources to enhance the accuracy and dependability of AES systems. Multiple studies have shown that by utilizing advanced NLP approaches, AES models can attain performance levels that are equivalent to those of human evaluators. This has been accomplished by continuously improving algorithms and using extensive datasets for training, enabling these models to gain a deeper understanding of the contents of the essay and evaluate the subtleties of written language. However, these systems have primarily been developed for English and other high-resource languages. Nepali, which is a low-resource language based on the Devanagari script, remains unexplored in the context of AES due to its complex script formation and low research effort. In this article, we prepare a large translated dataset using machine translation algorithms and evaluate the efficiency of various machine learning and deep learning models in Nepali AES using scores like the Quadratic Weighted Kappa (QWK) score. For the classical machine learning (ML) approach, we used a feature-based method. Meanwhile, for state-of-the-art transformer-based models, we fine-tuned the models based on the transformer architecture. Our findings demonstrate that the effectiveness of AES systems is greatly influenced by the quality of translations, as the accuracy and precision of the translation process have a direct impact on the overall performance of the AES models. By comparing various models using the QWK score, we have demonstrated that fine-tuned transformer architectures perform quite similar to the traditional feature-based ML method. Our research efforts are a step further in enabling deep learning and artificial intelligence (AI) access to the Nepali-speaking community. The dataset is available at https://github.com/rkritesh210/NepAES.

External IDs:dblp:journals/peerj-cs/PoudelRARANT25