Abstract: In this study, we use various RNN architectures namely, RNN, Bi-LSTM, and GRU — alongside BERT to analyze sentiment across university departments. Our aim is a comparative analysis of these models in sentiment classification within education. We collected and pre-processed textual data from multiple departments for balanced training and validation. Results showed that traditional RNNs achieved 90% accuracy, Bi-LSTM 93%, and GRU 89%. BERT, leveraging its Transformer architecture, outperformed with 94% accuracy. These findings highlight the superiority of BERT in capturing complex language patterns for sentiment analysis. This study underscores the potential of advanced neural network architectures to gain insights into departmental sentiments, informing policy decisions and educational strategies. Aligning with sustainable development goals in education, we aim to use AI models to develop effective, inclusive, and responsive educational strategies, enhancing quality and accessibility.