Covariance Pooling Layer for Text Classification
Mai Ibrahim, Ayman Shaawat, Marwan Torki · Procedia Computer Science · 2021
In this paper, we address the text categorization problem using a variant of traditional Convolutional Neural Networks (CNN). Our novel approach computes covariance of CNN features extracted for a given document. That is we add a covariance pooling layer on top of the convolutional features. The covariance pooling layer has more discrimination ability for classification tasks over regular max pooling layers that are used widely with CNN. This combination of CNN features and covariance pooling layers shows better classification accuracy against the traditional max pooling when tested on standard evaluation benchmarks.