A Multimodal Fusion and Hybrid Label Framework for Sentiment Analysis of Online Learning Platform

This study presents a hybrid sentiment analysis framework for understanding student feedback on educational platforms, using over 33,000 Udemy course reviews. Unlike prior work that relied on either rating-based or text-based sentiment labeling, this research integrates both numerical star ratings and textual review content through a multimodal fusion strategy. Five distinct labeling methods—rating-based, lexicon-based (TextBlob, VADER), and transformer-based (BERT, XLM-RoBERTa)—were combined using a consensus-based hybrid labeling approach to improve label reliability and reduce bias. Multiple machine learning classifiers were trained on feature representations including TF-IDF, Count Vectorization, Word2Vec, and GloVe. The Hybrid All labeling strategy demonstrated more balanced class distribution and robust classification performance. After hyperparameter tuning, the Extra Trees and XGBoost models achieved notable improvements, with accuracies reaching up to 83.9\% and F1 scores above 78\%. The study shows that hybrid labeling with tuned models can improve sentiment classification in MOOC platforms, offering insights for more accurate and reliable learner sentiment interpretation.