Vietnamese sentiment analysis on social media based on BERT architecture

Abstract

This study conducts a comparative evaluation of machine learning models for Vietnamese sentiment analysis, including ViBERT4News, PhoBERT, XLM-RoBERTa, ViBERT, SVM, and LSTM. The dataset comprises 10,169 text samples collected from Facebook related to students at the Ho Chi Minh City Campus of the University of Transport and Communications (UTC2), labeled into three classes: positive, negative, and neutral. Results show that BERT-based models achieve superior performance compared to traditional methods, with ViBERT4News achieving the highest accuracy (89.25%), followed by PhoBERT (88.76%). The study not only demonstrates the effectiveness of pre-trained language models in processing emotions on Vietnamese social media data but also suggests potential applications in platforms for analyzing and visualizing student feedback, supporting academic monitoring and improving educational quality.