Reducing the number of labeled examples required to learn accurate prediction models is an important problem in structured output prediction. In this paper we propose a new transductive structural SVM algorithm that learns by incorporating prior knowledge constraints on unlabeled data. Our formulation supports different types of prior knowledge constraints, and can be trained efficiently. Experiments on two citation and advertisement segmentation tasks show that our transductive structural SVM can learn effectively from unlabeled data, achieving similar prediction accuracies when compared against other state-of-art algorithms.