Traditional machine learning and deep learning models ignore the strength of emotional feature words when processing emotion classification tasks. As a result, the emotional semantic relationship is poor, resulting in low accuracy of emotion classification. To this end, this paper proposes an improved BiLSTM-CNN+Attention sentiment classification algorithm with a sentiment dictionary. Firstly, the algorithm optimizes the weights of feature words by fusion of sentiment dictionaries. Secondly, it uses a convolutional neural network (CNN) to extract local features and uses a bidirectional long and short-term memory network (BiLSTM) to extract contextual semantic features and long-distance dependencies efficiently. Then, this model combines the attention mechanism to add to the emotion features. Finally, the Softmax classifier realizes the text emotion prediction. Experimental results show that the sentiment classification algorithm proposed in this paper dramatically improves precision, recall, and F-measure. Compared with TextCNN, BiLSTM, LSTM, CNN, and random forest models, the F-measure of this method is increased by 2.35%, 3.63%, 4.36%, 2.72%, and 6.35%, respectively. In short, the proposed method can fully integrate the weights of emotional feature words, use contextual semantic features, and improve the performance of emotional classification. Therefore, this method has specific educational value and application prospects.