Title: LSTM with compensation method for text classification

Authors: Wei Huang; Mengyu Liu; Wenqian Shang; Haibin Zhu; Weiguo Lin; Chunjie Zhang

Addresses: Division of Scientific Research, Communication University of China, Beijing, China ' School of Computer Science and Cyberspace Security, Communication University of China, Beijing, China ' School of Computer Science and Cyberspace Security, Communication University of China, Beijing, China ' Department of Computer Science, Nipissing University, North Bay, Ontario, Canada ' School of Computer Science and Cyberspace Security, Communication University of China, Beijing, China ' School of Computer Science and Cyberspace Security, Communication University of China, Beijing, China

Abstract: As a foundational task, text classification is widely used in the field of Natural Language Processing (NLP). In the recent research on text classification, neural network-based methods have produced promising results. Nevertheless, most previous works ignore the fact that information may be lost or misinterpreted after the calculation of the neural network. In the research of this paper, we avoid such problems by using historical information, such as the original information of a text and the output information of the hidden layers, then performing text classification. This proposed method is called Long Short-Term Memory with compensation, or simply, LSTM-Com, which dynamically selects the important historical information as compensation for the neural network. In the classification experiment, the improved algorithm showed excellent performance compared to the baseline.

Keywords: text classification; LSTM; neural networks; compensation mechanism.

DOI: 10.1504/IJWMC.2021.114139

International Journal of Wireless and Mobile Computing, 2021 Vol.20 No.2, pp.159 - 167

Received: 28 Sep 2020
Accepted: 23 Nov 2020
Published online: 09 Apr 2021 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article