Sentiment Classification of Fine-Tuned mBERT for Myanmar Cosmetic Domain

Nowadays, sentiment analysis is the crucial task for the feedback of customer review on social media, business organization, and very popular between the market researcher. It becomes an important in making decision for business owners because they want to know the feedback of their products. Sentiment analysis classifies the polarities into positive, negative, or neutral classes. Recent advances in deep learning, particularly transformer-based models, have significantly enhanced the performance of sentiment analysis across various languages. This paper explores sentiment classification using the Multilingual Bidirectional Encoder Representations from Transformers (mBERT) on the user review comment that are collected from YouTube channels, Facebook and Tiktok applications written in Myanmar language. Different from BERT transformer, mBERT support 104 languages including Myanmar language. By leveraging the pre-trained mBERT architecture, the study demonstrates significant improvements in classifying Myanmar text into sentiment classification task. Experimental results show the effectiveness of fine-tuning for low-resource languages like Myanmar. The model achieves state-of-the-art performance on a manually annotated dataset by comparing the accuracy on both BERT and mBERT on customized cosmetic dataset.