With the rapid development of the internet and social media, network public opinion has become an essential channel for expressing public opinions and sentiments. Governments and enterprises need to quickly understand public sentiment when responding to emergencies, policy releases, and social hotspots, in order to adjust response strategies promptly. This paper proposes a hybrid model based on RoBERTa, BiLSTM, and multi-head attention mechanisms (RBMA), aiming to enhance the accuracy and efficiency of sentiment analysis in network public opinion. The model utilizes the RoBERTa pre-trained model to capture the deep semantic features of text, incorporates BiLSTM to learn semantic dependencies, and applies multi-head attention to focus on critical sentiment information, thus handling complex emotional transitions more precisely. The study conducts model construction and comparison research and validates the model’s effectiveness in practical applications through case studies. Experimental results demonstrate that the RBMA model significantly outperforms traditional models in terms of accuracy, precision, recall, and F1-score, particularly excelling in handling sudden public opinion events and the spread of negative sentiments. By applying this model, public opinion managers can monitor sentiment dynamics more promptly and effectively, reducing the potential risks of negative sentiment for society and organizations, and improving the overall response efficiency in public opinion management.