Dual channel semantic enhancement-based convolutional neural networks model for text classification

Text classification is an essential research aspect in the field of natural language processing. The shortcomings of insufficient capture of long-range semantic information and poor model generalization in text classification tasks, to surmount the above-mentioned problems, a two-channel convolutional module with weighted attentional utility and a semantic enhancement module for textual features are introduced. In this paper, we proposed a Dual-channel Semantic enhancement-based Convolutional Neural Network text classification model (DcSeCNN). First, Atlas convolution is introduced to collaborate with TextCNN convolution for forming dual-channel convolution and acquiring local and global semantic information of sentence-level text. Second, to better exploit the textual information flow, a weighted average attention mechanism module is used to enhance the features of the two-channel weighting. Finally, the original text vectors are semantically augmented with the original discourse and fused into a two-channel augmented features map, respectively, tuned by the learning rate attenuation coefficient λ and dropout parameters to constitute a DcSeCNN model. Extensive experimental comparative analyses with seven baseline models on six data sets, including MR, R8, R52, TREC, IMDBR, and THUCNews, have shown the advantages of the scheme in terms of enhanced semantic message extraction, improved classification and model generality.