Multi-classification of Patent Applications with Winnow

The Winnow family of learning algorithms can cope well with large numbers of features and is tolerant to variations in document length, which makes it suitable for classifying large collections of large documents, like patent applications. Both the large size of the documents and the large number of available training documents for each class make this classification task qualitatively different from the classification of short documents (newspaper articles or medical abstracts) with few training examples, as exemplified by the TREC evaluations. This note describes recent experiments with Winnow on two large corpora of patent applications, supplied by the European Patent Office (EPO). It is found that the multi-classification of patent applications is much less accurate than the mono-classification of similar documents. We describe a potential pitfall in multi-classification and show ways to improve the accuracy. We argue that the inherently larger noisiness of multi-class labeling is the reason that multi-classification is harder than mono-classification.