Addressing the issue of incomplete semantic extraction from long texts and the uncontrollable increase in computational complexity due to simply extending the token sequence length of models, this paper proposes an enhanced model, SubBERT(SubsamplingBERT for long text), that combines BERT with data and convolutional downsampling. Initially, low-semantic characters are eliminated during the data subsampling phase, leaving more useful information within a limited space. Subsequently, a convolutional encoding layer is introduced in the encoder stack, which condenses the initial 512-dimensional high-dimensional vector representation down to a 128-dimensional vector representation. This compression allows the model to handle four times the input scope under the same computational load. Finally, classification results are obtained through a fully connected layer. Experimental research based on the THUCNews dataset demonstrates that the enhanced model (SubBERT) proposed in this paper outperforms other models in terms of precision, recall, and F1 score, providing an effective solution for long-text classification tasks and finding a balance between accuracy and efficiency.