This paper investigates Long Short-Term Memory (LSTM) neural network models for sentiment analysis using a dataset of 107,018 reviews sourced from student feedback on the Coursera platform. The study emphasizes the role of sentiment analysis in understanding student opinions, focusing on optimizing model performance to effectively classify negative, neutral, and positive sentiments. Evaluations are based on recall rates across sentiment classes, highlighting the effectiveness of the model trained with class weights as the top performer. Despite challenges in accurately classifying nuanced sentiments and managing frequently used words across different classes, the research underscores the significance of addressing class imbalance and refining model parameters for enhanced sentiment analysis accuracy in educational contexts. Future research directions include further refinement of model architectures and exploration of advanced techniques to mitigate classification complexities in large-scale, imbalanced datasets from educational platforms.