The D3CNN text classification model based on multi-scale textual feature maps

Abstract Owing to the inherent limitations of convolutional neural networks (CNNs), text classification has historically fallen short of the performance achieved by 2D CNNs in computer vision. Furthermore, the intrinsic complexity and variability of textual data make shallow classification models insufficient for capturing deeper and more abstract semantic representations. To overcome these challenges, this study adopts a constrained 2D CNN coupled with a vision-inspired attention mechanism to extract informative textual features, thereby constructing a novel representation termed a multi-scale textual feature map. Building on the demonstrated effectiveness of ResNet and SENet in deep architectures, we further designed a competitive attention mechanism specifically tailored for multi-scale textual feature maps and introduced the Deep 2D CNN (D3CNN). Experimental results indicate that D3CNN yields substantial performance improvements over baseline models across five benchmark datasets, approaching the accuracy of state-of-the-art Transformer-based architectures while maintaining a favorable balance between performance and computational efficiency.