A Novel Method Using Local Feature to Enhance GCN for Text Classification

Abstract

Text classification is a classical and basic task of natural language processing (NLP). In recent years, machine learning has been widely used for text classification. However traditional machine learning methods depend heavily on high quality feature engineering. Recently, deep learning methods have contributed to improving text classification performance. Graph convolution network (GCN) has been proved to be able to capture spatial feature of documents. However, the ability of GCN to capture sentence local features and context information is limited. In this paper, we propose a novel method using local features to enhance GCN for text classification. The combination methods based on Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) (such as Bi-LSTM, C-LSTM and ServeNet) are used to capture local features to enrich feature information, and a weight value is used to adjust the intensity of enhancement. We conducted extensive experiments on 5 benchmark datasets (WSDataset, Ohsumed, R52, R8, 20NG), proving the proposed method significantly outperforms the baseline deep learning methods.