In the classification task of the fairly specialized chemical domain texts, it is hard to wholly characterize the spe-cialized phrases and some other phrases in the domain just by way of relying on the current word vector representation due to the specialized and complex diversity of the texts, which leads to the low accuracy of the classification task for chemical domain texts. To tackle these problems, the text classification model incorporating multi-granularity dy-namic semantic representations was proposed. Firstly, the adversarial perturbation was introduced into the word embedding layer of the model to enhance the ability of dynamic word vectors to represent the semantics. Then the word vector weights were redistributed by a multi-headed attention mechanism to obtain a better textual represen-tation of key semantic information. Finally, text representations of different granularities were extracted through the proposed multi-scale residual shrinkage deep pyramidal convolutional neural network (MSRS-DPCNN) and hybrid attention capsule bidirectional LSTM (HAC-BiLSTM) network model, fusing them and classifying the result of the fusion. Compared with existing models, the experimental results show that the proposed model achieves an F1-score of up to 84.62% on the chemical domain text dataset when using different word vector representations, an im-provement of 0.38 ~ 5.58 percentage points; The model also had better performance when it was evaluated for model generalization performance on the publicly available Chinese dataset THUCNews and the Tan Songbo hotel review dataset ChnSentiCorp.