Automatic Document Classification - A thorough Evaluation of various Methods

(Automatic) document classification is generally defined as content-based assignment of one or more predefined categories to documents. Usually, machine learning, statistical pattern recognition, or neural network approaches are used to construct classifiers automatically. In this paper we thoroughly evaluate a wide variety of these methods on a document classification task for German text. We evaluate different feature construction and selection methods and various classifiers. Our main results are: (1) feature selection is necessary not only to reduce learning and classification time, but also to avoid overfitting (even for Support Vector Machines); (2) surprisingly, our morphological analysis does not improve classification quality compared to a letter 5-gram approach; (3) Support Vector Machines are significantly better than all other classification methods.