Addressing the issue that the combination of emojis and text data may alter the original semantics, and the mechanism of their interaction with text data has not been fully explored. For this reason, textual sentiment classification incorporating dual emoji attention mechanisms is proposed in the paper. First, a BERT pre-training model is used to obtain the dynamic word vector representation of text; then a CNN-BiGRU dual-channel model is constructed to extract local and global features respectively; after that, the Emoji2vec model is used to obtain the emoji vector representation and construct a dual emoji attention mechanism, which strengthens the key information of the combination of emoji and text from the level of local and global emoji attention mechanisms respectively; then the output feature vectors are fused to classify emotions. In order to verify the effectiveness of the proposed model, contrast and ablation experiments were set up. The Emoji-phone and EmojifyData datasets were used for sentiment classification training, and the findings indicate that the model in this article outperforms the more recent RoBERTa-3xBiGRU model by 0.0176 and 0.0166, respectively.