In this paper, we describe a new approach to classification of text documents based on the minimization of system entropy, i.e., the overall uncertainty associated with the joint distribution of words and labels in the collection. The classification algorithm assigns a class label to a new document in such a way that its insertion into the system results in the maximum decrease (or least increase) system entropy. We provide insights into the minimum system entropy criterion, and establish connections to traditional naive Bayes approaches. Experimental results indicate that the algorithm performs well in terms of classification accuracy. It is less sensitive to feature selection and more scalable when compared with SVM.