Evolving Multi-view Autoencoders for Text Classification
Tuan Cuong Ha, Xiaoying Gao · IEEE/WIC/ACM International Conference on Web Intelligence · 2021
Text classification has received more and more attention from researchers in recent years, due to its wide range of helpful applications in the real world. Its performance has significantly benefited from the rapid development of Deep Neural Networks (DNNs). However, designing the network architecture of DNNs is by no means an easy task and can increase operating costs, since it requires domain knowledge from both the deep learning and text classification areas, which is often not available. In addition, existing methods typically use one representation which is a single view of text documents, and hence, they may have difficulty capturing all the discriminative features for text classification. To resolve these issues, this paper proposes a novel method named Evolving Multi-view Autoencoder (EvoMAE), which can automatically design optimal architectures of Autoencoders to extract multi-view features for text classification. The multi-view features are obtained from (i) Word, (ii) Emotion, and (iii) Sentence embeddings. We also introduce a novel objective function that allows the training of Autoencoders and the classifier simultaneously. The proposed system is tested on two benchmark datasets, and the experimental results show that it achieves significantly better performance than the current state-of-the-art automatically-designed approaches.