Hierarchical text categorization using coding matrices

We discuss the task of ontology population as a machine learning problem with a large hierarchy of classes. Since many machine learning methods are designed primarily for two-class problems, it is desirable to transform the multiclass classification problem into several two-class problems. Coding matrices are a unifying formalism for describing such transformations. We present an approach for constructing coding matrices in a greedy way, with a focus on achieving good performance with a tractable number of two-class classification models.