The New York Times Annotated Corpus contains over 1.5 million of manually tagged articles, and could become a useful source for evaluation of algorithms for documents clustering. Since documents were labeled over twenty years, it is argued that classication may contains errors due to possible dissent between experts, and the necessity to add tags over time. This paper presents an approach to improve classication quality by using assigned tags as a starting point. It is assumed that tags can be described by a set of features. These features are selected based on the value of mutual information between the tag and stems from documents with it. An algorithm for reassigning tags in case the document does not contain features of its labels is presented. Experiments were performed on about ninety thousand articles published by the New York Times in 2005. Results of applying the algorithm to the collection are discussed.