Massive Open Online Courses (MOOCs) are online courses that are created to be open and accessible to a large number of learners from all around the world. MOOCs evaluation plays an important role in preserving the quality and effectiveness of the online course. Sentiment analysis is the process of finding a person’s opinion on a particular product or service using a small piece of text. MOOCs evaluation using sentiment analysis helps the MOOC providers and instructors to enhance the overall learning experience for students but understanding contextual information, and working with limited labeled dataset are major challenges. To address these problems, this paper proposes a hybrid architecture that combines XLNet model with long short-term memory (LSTM) where XLNet is a pretrained transfer learning model that leverages preexisting knowledge and extracts potential features with limited dataset, LSTM is a deep neural network that performs sentiment classification and thereby both together achieves faster convergence and improves the performance of MOOCs evaluation. Performance of the proposed XLNet-LSTM model is also compared with other hybrid models that are built by combining different variants of BERT (BERT, ALBERT, RoBERTa) with LSTM and CNN. Experimental results show that the proposed hybrid model outperforms all other hybrid models and gives state-of-the-art results for MOOCs evaluation task.