The traditional text classification method faces great challenges when processing with the multi-round dialogue including several intents. In this paper, we propose an intent classification model based on comparative learning and an attention mechanism. The text is divided into long and short categories and encoded into the Transformer. Then, the word embedding matrix is perturbed to generate adversarial samples, and the positive sample pairs are then compared for loss. Finally, the positive sample pair is input into the multi-round inference module, and the inference features are obtained by learning the semantic clues in the whole scene through multi-round dialogue. Experiments on two datasets exhibit that the proposed method achieved good performance.