Adaptive feature interaction enhancement network for text classification.

Area of Science:

• Natural Language Processing
• Machine Learning

Background:

• Text classification models struggle to capture both global semantics and local details.
• Existing methods often fail to effectively integrate diverse textual features.

Purpose of the Study:

• To propose a novel network, AFIENet, for enhancing text classification.
• To improve the integration of global and local text features for better classification accuracy.

Main Methods:

• Developed a two-branch network (global and local) for text modeling.
• Implemented an adaptive segmentation module for capturing key phrases locally.
• Designed an interaction gate for selective fusion of global and local features.

Main Results:

• AFIENet significantly improved performance of backbone networks like TextCNN, RNN, and Transformer.
• Achieved an average accuracy of 3.82% and F1-score of 3.88% improvement with the Transformer backbone.
• Demonstrated comparable results to MacBERT, highlighting method applicability.

Conclusions:

• AFIENet effectively enhances text classification by interactively fusing global and local features.
• The proposed method offers improved performance with fewer parameters.
• AFIENet shows broad applicability across different backbone architectures.