Leah S. Larkey
The classification of U.S. patents poses some special problems due to the enormous size of the corpus, the size and complex hierarchical structure of the classification system, and the size and structure of patent documents. The representation of the complex structure of documents has not received a great deal of previous attention, but we have found it to be an important factor in our work. We are exploring ways to use this structure and the hierarchical relations among patent subclasses to facilitate the classification of patents. Our approach is to derive a vector of terms and phrases from the most important parts of the patent to represent each document. We use both k-nearest-neighbor classifiers and Bayesian classifiers. The k-nearest-neighbor classifier allows us to represent the document structure using the query operators in the Inquery information retrieval system. The Bayesian classifiers can use the hierarchical relations among patent subclasses to select closely related negative examples to train more discriminating classifiers.