Recent emergence and flourishment of social media services (SMS) have led to drastic changes in the way people interact. When Internet users share their feelings and opinions about some specific topic online, the massive amount of discussion forms a real-time public opinion pool. Given this phenomenon, employing sentiment analysis to measure public opinions from online social communities has become a hot research issue in these years. To achieve higher classification accuracy in Chinese sentiment analysis, this research proposes the LMAEB-CNN model, which combines Bi-LSTM and CNN with multi-head attention mechanism. The proposed model not only solves the over-fitting problem, but also promotes the accuracy of emotional polarity classification. To carry out experiments, we used datasets collected from four famous Chinese SMS platforms, including Plurk, PTT, Dcard, and Mobile01, and the LMAEB-CNN model demonstrated higher accuracy than other methods, including SVM, CNN, LSTM, and AEB-CNN.