We propose a semantic kernel for Support Vector Machines (SVM) that takes advantage of higher-order relations between the words and between the documents. Conventional approach in text categorization systems is to represent documents as a "Bag of Words" (BOW) in which the relations between the words and their positions are lost. Additionally, traditional machine learning algorithms assume that instances, in our case documents, are independent and identically distributed. This approach simplifies the underlying models, but nevertheless it ignores the semantic connections between words as well as the semantic relations between documents that stem from the words. In this study, we improve the semantic knowledge capture capability of a previous work in , which is called chi-Sim Algorithm and use this method in the SVM as a semantic kernel. The proposed approach is evaluated on different benchmark textual datasets. Experiment results show that classification performance improves over the well-known traditional kernels used in the SVM such as the linear kernel (one of the state-of-the-art algorithms for text classification system), the polynomial kernel and the Radial Basis Function (RBF) kernel.