Optimizing BERT for Nepali Text Classification: The Role of Stemming and Gradient Descent Optimizers

This study investigates the use of BERT for classifying Nepali news articles, addressing the specific challenges associated with Nepali as a low-resource language in natural language processing (NLP). While traditional text classification methods have proven effective for high-resource languages, they often fall short in capturing the contextual nuances necessary for accurate classification in Nepali. To address this gap, a pre-trained BERT model was fine-tuned on a balanced dataset of Nepali news articles sourced from various outlets. The study examined the effects of different preprocessing techniques, such as stemming, and optimization algorithms including Adam, AdamW, and Momentum, on classification performance. Experimental results demonstrate that the combination of stemming and the AdamW optimizer yielded the best performance, achieving a weighted accuracy of 93.67%, along with balanced macro precision, recall, and F1-scores of 0.94. These findings underscore the effectiveness of advanced optimization strategies-particularly AdamW-in enhancing model performance.