Sentiment analysis has become an essential tool for understanding public opinion, with widespread applications in business, politics, and social studies. This paper proposes an advanced multi-class sentiment classification model by combining pre-trained BERT embeddings with recurrent neural network (RNN) variants, including LSTM and GRU, along with an attention mechanism to enhance performance in identifying positive, negative, neutral, and mixed sentiments. The integration of BERT allows the model to capture rich contextual information, while the attention mechanism enables the model to focus on the most relevant parts of the input text, improving both classification accuracy and interpretability. Six experiments were conducted to assess the effectiveness of the attention mechanism, testing models with and without attention across RNN, LSTM, and GRU architectures. The results showed significant improvements with the attention-enhanced models, with the BERT + LSTM + Attention model achieving 93.5% accuracy, 92.5% recall, and 93% F1-score, outperforming traditional approaches. Experimental validation on the Kaggle Sentiment 140 dataset demonstrates the robustness and scalability of the proposed model, offering a highly effective solution for multi-class sentiment classification tasks. These findings suggest potential applications in areas such as brand monitoring, public opinion analysis, and social media analytics.