Self-Attention-Based Convolutional Neural Networks for Sentence Classification

Abstract

Sentence classification is a challenging task. The research on convolutional neural networks combined with the attention mechanism for sentence classification is not yet complete, especially the performance of multi-classification tasks needs to be improved. In this paper, we propose a self-attention-based convolutional neural network (SACNN) for sentence classification, which consists of two self-attention layers and a convolutional neural network. We conducted multiple experiments on seven benchmark datasets. Experimental results show that the proposed model can achieve up to 0.4%-1.4% higher accuracy than other CNN-based models, and outperform other CNN-based models on five out of seven tasks.