Short text classification method with dual channel hypergraph convolution networks

Abstract

In some fields such as e-commerce and social media platforms and sentiment analysis, efficient short text classification is crucial to enable users to locate pertinent information effectively. Along with the increasing number of short texts, classifying short texts with brief contents and sparse features has become a major research topic in recent years. Towards this end, a short text classification method based on a dual channel hypergraph convolutional network is proposed to flexibly capture the complex higher-order relationships among short texts and words. Specifically, our method firstly models the pre-processed short text data into short text hypergraph and short text association graph; secondly, two different short text feature representations are learned via a dual channel hypergraph convolutional network and fused by an attention network to enhance the short text embedding; at last, a classification model is adopted to perform short text classification. Extensive experimental results indicate that the method has superior short text classification effect and stability compared with the existing model, which has better performance among comparable short text classification models.