Covariance Matrix Adaptation Evolution Strategy for Convolutional Neural Network in Text Classification

Text classification has become relevant in recent years because of its usefulness in supporting different text mining solutions. Neural networks for this purpose have benefited from the creation of word embedding for learning semantics among words in a corpus. However, artificial neural network training by conventional methods present several theoretical and computational limitations. In this work, we develop a hybrid training method that combines gradient-based methods and Covariance Matrix Adaptation Evolution Strategy to train Convolutional Neural Network for the text classification task. For this, the training process is divided into two stages taking advantage of the speed of the gradient-based methods for learning the parameters of the convolutional filters and the application of the Covariance Matrix Adaptation Evolution Strategy for learning the weights of the fully connected layer. Our proposal was evaluated using a Spanish dataset for text classification, taken off the EcuRed Cuban Encyclopedia, divided into five classes. The proposed method increases the accuracy significantly of the convolutional network applied to the text classification.