Merging Statistical Feature via Adaptive Gate for Improved Text Classification

Currently, text classification studies mainly focus on training classifiers by using textual input only, or enhancing semantic features introducing external knowledge (e.g., hand-craft lexicons and domain knowledge). In contrast, some intrinsic statistical of the corpus, like word frequency distribution over labels, are not well exploited. Compared with knowledge, deterministic naturally compatible corresponding tasks. this paper, we propose an Adaptive Gate Network (AGN) to consolidate representation selectively. particular, AGN encodes through a variational component merges information via well-designed valve mechanism. The adapts flow into classifier according confidence in decision making, which can facilitate robust address overfitting caused features. Extensive experiments datasets various scales show that, incorporating information, improve performance CNN, RNN, Transformer, Bert based models effectively. also indicate robustness against adversarial attacks manipulating information.