Discriminative clustering of text documents

Vector-space and distributional methods for text document clustering are discussed. Discriminative clustering, a recently proposed method, uses external data to find task-relevant characteristics of the documents, yet the clustering is defined even with no external data. We introduce a distributional version of discriminative clustering that represents text documents as probability distributions. The methods are tested in the task of clustering scientific document abstracts, and the ability of the methods to predict an independent topical classification of the abstracts is compared. The discriminative methods found topically more meaningful clusters than the vector space and distributional clustering models.