AFIENet: Feature Interaction Enhancement Network for Adaptive Text

To address the difficulties existing text categorization models face in capturing global text semantics and local details, we propose an Adaptive Feature Interactive Enhancement Network (AFIENet). This network uses two branches to model the text globally and locally. The adaptive segmentation module in the local network can dynamically split the text and capture key phrases. The global network grasps the overall central semantics. After obtaining the results from the two branches, an interaction valve is designed to evaluate the confidence of the global features and selectively fuse them with the local features effectively. Finally, the interactively enhanced features are re-input into the classifier to improve text classification performance. Experiments verify that AFIENet can effectively improve the performance of basic networks like TextCNN, RNN, and Transformer with the introduction of fewer parameters. The accuracy reaches up to 97.55% on the THUCNews dataset, 6.53 percentage points higher than the original model. The average accuracy improvement is 2.11 percentage points compared to the original model when tested on multiple datasets. Comparable results to MacBERT are obtained when using static word vectors, reflecting the wide applicability of the model.