Short Text Classification Based on Cross-Connected GRU Kernel Mapping Support Vector Machine

Abstract

Support vector machine (SVM) has achieved excellent results in short text classification. However, its performance is limited in the kernel function. This paper presents a short text classification method based on Cross-connected GRU Kernel Mapping Support Vector Machine (C-GRUKMSVM), to further improve the accuracy of short text classification. The method consists of a feature mapping module and a classification module. The feature mapping module first represents the text as a word vector using the glove method, and then explicitly maps the low-dimensional word vector to a high-dimensional space using a three-layer cross-connected GRU; the classification module uses a soft-margin support vector machine for classification. Experimental results on five publicly available short text datasets show that C-GRUKMSVM achieves better text classification performance than convolutional networks, support vector machines and Naïve Bayes. Additionally, different cross-connected methods, recurrent units and recurrent structures have an impact on the performance of C-GRUKMSVM.