A novel fast and accurate supervised learning algorithm isproposed as a general text classification algorithm for linearly separateddata. The strategy of the algorithm takes advantage of the training errorsto successively refine an initial classifier. Experimental evaluation of theproposed algorithm on standard text collections, show that results com-pared favorably to those from state of the art algorithms such as SVMs. Experiments conducted on the datasets provided in the framework of theECDL/PKDD 2008 Challenge for Spam Detection in Social Bookmark-ing Systems, demonstrate the effectiveness of the proposed algorithm.