This Small Business Innovation Research Phase project will implement and test the feasibility of a new algorithm for statistical text categorization. This algorithm combines the ease and effectiveness of learning from examples, while incorporating task-specific constraints that currently require ad hoc rules. The project will evaluate the efficiency and effectiveness of these algorithms. Alternative optimization algorithms and alternative approximations to intractable quantities will be benchmarked. Categorization accuracy will be evaluated on public text categorization data sets and on data from operational text categorization users. Ornarose, Inc. will develop and license software libraries including this algorithm to software vendors in a variety of industries. Vendors for whom text categorization is not a core competency increasingly wish to support categorization in software that works with text data. Market niches where text categorization under task-specific constraints is compelling include knowledge management, news alerting, email/web filtering, and data mining. Discounted licenses will be provided for academic institutions and scholarly publication of nonproprietary results on publicly available text categorization data sets is planned
