Title: CNN-based text multi-classifier using filters initialised by N-gram vector
Authors: Yan Xiang; Ying Xu; Zhengtao Yu; Dangguo Shao; Hongbin Wang; Yantuan Xian
Addresses: Department of Information Engineering and Automation, Kunming University of Science and Technology, No. 727, Jingming South Road, Kunming, Yunnan Province, China ' Department of Information Engineering and Automation, Kunming University of Science and Technology, No. 727, Jingming South Road, Kunming, Yunnan Province, China ' Department of Information Engineering and Automation, Kunming University of Science and Technology, No. 727, Jingming South Road, Kunming, Yunnan Province, China ' Department of Information Engineering and Automation, Kunming University of Science and Technology, No. 727, Jingming South Road, Kunming, Yunnan Province, China ' Department of Information Engineering and Automation, Kunming University of Science and Technology, No. 727, Jingming South Road, Kunming, Yunnan Province, China ' Department of Information Engineering and Automation, Kunming University of Science and Technology, No. 727, Jingming South Road, Kunming, Yunnan Province, China
Abstract: Text classification based on convolutional neural networks (CNN) has got more attention recently. This paper presents an improved CNN-based text multi-classifier. First, word vector training is performed on the corpus to be classified. Then, the most important N-grams for a particular category are selected and clustered into different groups. Finally the centroid vectors of different groups are used to initialise the centre weights of filters. Initialisation weights enable CNN to extract N-gram features more effectively and ultimately improve text classification results. Multi-classification experiments using multiple advanced models were performed on different data sets. Experiments show that the proposed model is more accurate and stable than other baseline models.
Keywords: convolutional neural networks; text classification; N-gram; word embedding; clustering; filter; word vector.
DOI: 10.1504/IJICT.2019.103202
International Journal of Information and Communication Technology, 2019 Vol.15 No.4, pp.419 - 430
Received: 13 Aug 2018
Accepted: 15 Nov 2018
Published online: 22 Oct 2019 *