A novel kernel for text classification based on semantic and statistical information

Abstract

In text categorization, a document is usually represented by a vector space model which can accomplish the classification task, but the model cannot deal with Chinese synonyms and polysemy phenomenon. This paper presents a novel approach which takes into account both the semantic and statistical information to improve the accuracy of text classification. The proposed approach computes semantic information based on HowNet and statistical information based on a kernel function with class-based weighting. According to our experimental results, the proposed approach could achieve state-of-the-art or competitive results as compared with traditional approaches such as the k-Nearest Neighbor (KNN), the Naive Bayes and deep learning models like convolutional networks.