Unveiling the reasons behind learners’ dropout from educational platforms: analyzing sentiment intensity throughout texts

Abstract Identifying learners’ emotions in real-time yields significant benefits, allowing teachers to better understand learning behavior and tackle obstacles, such as confusion and boredom, which can affect learner engagement and performance. E-learning platforms facilitate learning acquisition by taking advantage of behavioral data collected through collaborative tools. This study aimed to discern users’ emotions, whether positive, negative, or neutral, by analyzing their feedback on online courses. This approach is pivotal for predicting dropout risk and for tailoring educational content to identify emotions. However, extracting sentiments from feedback, particularly in longer texts, presents a greater challenge than extracting sentiments from shorter texts, and requires more nuanced analysis. While existing models often focus on short texts derived from social media and collaborative e-learning platforms, our study introduces a novel approach based on a Bayesian model that merges CNN and BiLSTM architectures to estimate uncertainty in sentiment analysis. This technique was compared with other non-Bayesian models such as BiGRU, BiLSTM, CNN, and CNN-BiLSTM, incorporating embeddings such as BERT, GloVe, Word2Vec, and FastText. The experimental findings revealed that our model outperformed the other models, achieving an accuracy of 97.12% for long texts and 96.53% for short texts with FastText. This high performance validates the effectiveness of our approach, showing a strong correlation between the sentiments expressed in the forums and the e-learning dropout rate, underlining its strategic importance in the educational field.