Creating Taxonomies And Training Data For Document Categorization

A method comprising generating sets of features representing a plurality of categories from a taxonomy of categories, said step of generating including the steps of:

• selecting a set of categories which both make sense to a human and are minimally overlapping with one another;
• choosing a set of training documents for each particular category which both typify documents for said particular category and are only representative of said particular category;
• and
• determining a set of features for each particular category which best distinguishes said each particular category from all other categories in said plurality of categories.

A method comprising generating sets of features representing a plurality of categories from a taxonomy of categories, said step of generating including the steps of:

• selecting at least one set of potential categories for a categorization system;
• selecting training data into a set of training data for each of said potential categories;
• winnowing down the training data from each of said sets of training data into a smaller set of training data by applying at least one set of winnowing criteria;
• grouping the smaller set of training data from several related categories into a supercategory using at least one supercategory formation criterion;
• comparing the training data grouped in said supercategory and reducing overlap among categories within said supercategory so as to produce a modified group of training data;
• and
• extracting a set of differentiating features from the modified group of training data.