Inter-class relationships in text classification

Text classification is an active research area motivated by many real-world applications. Even so, research formulations and prototypes often make assumptions that are not suitable for deployment. For example, in many real applications, the set of class labels keeps evolving, continual user feedback must be integrated into the classifier, and test documents may come from a population statistically different from the training distribution. The main aim of our work is to build solutions for these problems using the idea of exploiting inter-class relationships. We learn noisy, approximate, and probabilistic mappings between related classes across label-sets in a semi-supervised framework we call cross-training. We exploit the notion of confusion between closely related classes, study its effect on label hierarchies, and present an algorithm for scaling up training of multi-class classifiers. We design discriminative, multi-label classifiers that are robust in the face of significant overlap, in terms of word distributions, between related classes. In many real applications, the set of labels is not predefined but must be constructed from vague specifications and a study of the corpus. Moreover, the label-set has to keep evolving as the corpus changes. We propose an algorithm that supports such temporal evolution by detecting classes in unseen data not defined during training. Our algorithm detects such classes using new notions of coverage of label-sets, support and confidence in a classification setting, and abstractions to represent documents. To enable continual interactive learning and to incorporate human input, we present a framework for active learning that combines terms and documents in a symmetric manner, reducing cognitive burden on the trainer. We conclude by proposing a new architecture for next-generation text classification platforms that embodies the ideas and contributions in this dissertation. To summarize, our work fills in conspicuous gaps between research prototypes and industry requirements, by exploiting one central idea: class labels are mutable variates just like words, documents and their assigned labels.