This paper proposes a novel multiclass classification method and exhibits its advantage in the domain of text categorization with a large label space and, most importantly, when some of the labels were not observed in the training data. The key insight is the introduction of intermediate aspect variables that encode properties of the labels. Aspect variables serve as a joint representation for observed and unobserved labels. This way the classification problem can be viewed as a structure learning problem with natural constraints on assignments to the aspect variables. We solve the problem as a constrained optimization problem over multiple learners and show significant improvement in classifying short sentences into a large label space of categories, including previously unobserved categories.