Sentiment analysis on student feedback using BERT combined with multi-channel CNN-GRU architecture

Abstract

Student feedback is a valuable data source for enhancing teaching quality and improving learner satisfaction. Numerous studies have conducted sentiment analysis on this data, yielding notable results. However, research in the Vietnamese language still faces significant limitations, including a limited number of published studies, challenges related to the target of sentiment, and data issues such as imbalance that pose difficulties for application. This study proposes a model that combines BERT with a multi-channel architecture consisting of CNN and GRU. By leveraging the strengths of each network, the performance of sentiment analysis on Vietnamese student feedback is expected to improve. The model focuses on classification tasks (topic and sentiment polarity) and supporting specific satisfaction measurements. Additionally, the model's ability to handle data imbalance is emphasized to utilize available datasets, saving time and finance effectively. Experiments on the UIT-VSFC dataset show performance improvements in Macro F1-Score compared to recent studies, with an increase of 0,01 in the topic classification task and 0,0051 in the sentiment polarity task. The study’s result will be a useful solution for educational institutions, which can be applied to improve teaching, reputation management, and learner support and be a motivation for expanding future research.