Analyzing Sentiments in eLearning: A Comparative Study of Bangla and Romanized Bangla Text Using Transformers

In the modern world, learning is becoming increasingly critical due to rapid technological breakthroughs, which highlight the need for continuous skill development in both the personal and professional spheres. As a result, eLearning is a cutting-edge approach to education that delivers lessons, courses, and instructional materials remotely via digital technology and the Internet. It makes learning more flexible and accessible by enabling users to interact with teachers online and access classes or other content. Sentiment analysis is an eLearning technique that evaluates user opinions, typically via written feedback, to improve the overall quality of instruction in a course. Sentiment analysis for e-learning feedback has been extensively studied in several languages, except Bangla and Romanized Bangla. The three datasets produced were one for Bangla, one for Romanized Bangla, and one for a combination of Bangla and Romanized. Three datasets contained 3178 Bangla, 3090 Romanized Bangla, and 6268 Bangla and Romanized Bangla texts. The feedback has been divided into three categories: positive, negative, and neutral. The validation of the datasets was conducted using Krippendorff’s alpha and Cohen’s kappa metrics, ensuring the reliability and consistency of the dataset annotations. Several techniques were used to train the preprocessed datasets, including transformers, deep learning, machine learning, ensemble learning, and hybrid approaches. Transformer-based algorithms, such as XLM-RoBERTa, outperformed the others in terms of accuracy, achieving the highest values of 89.46% and 85.81% for the Bangla and Combined datasets. At 89.59%, ANN demonstrated exceptional performance on the Romanized Bangla dataset.