Abstract
The paper uses NLP&CC2013 public data sets of emotion assessment tasks as the corpus. On account of these issues, such as the grammar of text is not standard, and the corpus is unbalanced distributing. Smote algorithm is put forward to over-sampling training data to get balanced corpus and CNN model is used to improve the sentence semantic differential. The optimal results of the paper improve 16.75% and 15.84% respectively compared with looseAP and strictAP of the optimal evaluation results in NLP&CC, increased by 11.15% and 11.55% respectively compared with the basic multi-label classifier ML-KNN. In addition, the paper also compares the performance of unbalanced and balanced training data based on CNN, confirming that the balanced training set can improve classification accuracy, at the same time through the change of the iterations, observe the semantic synthesis rule of CNN model.