Research on the construction of text classification model based on deep neural network

This paper presents a novel text categorization model that integrates multi-channel feature extraction and attention mechanisms. The architecture combines convolutional neural networks (CNNs) and bidirectional LSTM (Bi-LSTM) in a parallel structure to effectively capture both local n-gram patterns and global contextual dependencies. Pre-trained word embeddings are employed in the embedding layer to initialize semantic representations, which are further refined through a self-attention mechanism for enhanced semantic relevance. The model is built upon and extends CNN and BERT structures, aiming to improve semantic expressiveness and classification accuracy. Experiments conducted on the THUCNews dataset demonstrate that the proposed model achieves an accuracy of $97.1 \%$, outperforming BERT ($96.2 \%$) and TF-IDF+SVM ($86.2 \%$), thereby validating its superior accuracy and robustness.