This paper explores the fusion of BERT with various recurrent neural network (RNN) architectures, including simple RNN, LSTM, and GRU, for Twitter sentiment classification. For experimentation, we used a Kaggle Sentiment 140 dataset with positive and negative sentiments. Initially, three DL models (RNN, LSTM, and GRU) were applied for sentiment analysis and achieved accuracies of 86%, 91%, and 90%, respectively. By integrating BERT with these RNN variants, the proposed method enhanced the performance of sentiment analysis on Twitter data. The methodology involves tokenizing and encoding Twitter messages using the BERT tokenizer to obtain contextual embeddings. These embeddings are then fed into RNN-based classifiers, including Simple RNN, LSTM, and GRU, to capture sequential dependencies within the text. Later, a hybrid model with BERT and RNN variants was applied. The hybrid model combining BERT with LSTM achieves the highest accuracy at 93.5%, followed by BERT + GRU with 91% and BERT + RNN with 90%. These findings underscore the effectiveness of leveraging BERT in conjunction with LSTM and GRU architectures for Twitter sentiment classification, offering promising avenues for further research and real-world applications.