Text classification is a fundamental task in natural language processing. This task is widely concerned and applied. However, previous methods mainly use traditional static word embedding, but static word embedding could not deal with the problem of polysemy. For this reason, we propose to utilize contextualized BERT word embedding to effectively encode the input sequence and then use the temporal convolutional module which simply computes a 1-D convolution to extract high-level features, finally, the max-pooling layer retains the most critical features for text classification. We conduct experiments on six commonly used large-scale text categorization datasets, including sentiment analysis, problem classification and topic classification tasks. Due to the limitation of BERT processing long text, we propose an effective truncation method. Experimental results show that our proposed method outperforms previous methods.