[Purpose/significance] The existing text representation and classification algorithms are combined, and a combination mode of low complexity and less training time is selected to construct an optimization model for the classification of emotional texts of commodity reviews. [Method/process] Firstly, this paper took the Keras API as an application environment, input Word2vec word vector into Embedding embedded layer. Then, based on sentence word index sequence, three kinds of commodity comment text representation were realized by controlling the trainable parameter. Finally，in this paper, different text representations were matched with different classification algorithms, differences in classification effects were analyzed, and the better combination of algorithms was established. [Result/conclusion] The text representation method which is continued training by Inputting Word2vec Word Vector into Embedding embedded Layer, combined with the TextCNN algorithm establishes the classification model. It performs better on the product review test set. Its accuracy and ROC curve area AUC values are 94.02% and 0.9827, respectively. The application shows that the classification model can better realize the emotional classification of commodity reviews and has better classification generalization ability.