Creating Taxonomies And Training Data For Document Categorization

摘要

Methods, apparatus and systems are provided to generate from a set of training documents a set of training data and a set of features for a taxonomy of categories. In this generated taxonomy the degree of feature overlap among categories is minimized in order to optimize use with a machine-based categorizer. However, the categories still make sense to a human because a human makes the decisions regarding category definitions. In an example embodiment, for each category, a plurality of training documents selected using Web search engines is generated, the documents winnowed to produce a more refined set of training documents, and a set of features highly differentiating for that category within a set of categories (a supercategory) extracted. This set of training documents or differentiating features is used as input to a categorizer, which determines for a plurality of test documents the plurality of categories to which they best belong.