A Comparative Study of Deep Learning Models for Sentiment Analysis in Online Learning Environments

Abstract

This study presents a comparative analysis of various deep learning models for sentiment analysis on student feedback collected from Coursera. The models evaluated include traditional architectures such as Feedforward Neural Networks (FNN) and Recurrent Neural Networks (RNN), advanced sequential models like BiLSTM and BiGRU, hybrid architectures such as CNN-LSTM and attention-based variants, as well as transformer-based models like DistilBERT and RoBERTa. Performance was assessed using several metrics, including Accuracy, Precision, Recall, F1-Score, AUC-ROC, MCC, and confusion matrices. Results indicate that transformer-based models, particularly RoBERTa and DistilBERT, consistently outperform others across most metrics. Among traditional and hybrid models, CNN-LSTM showed competitive results. However, all models struggled to some extent with accurately classifying neutral sentiments, suggesting that further improvements in feature representation and contextual modeling are needed. These findings underscore the importance of model selection and fine-tuning in educational sentiment analysis to better understand and respond to learner feedback.