Learning neural networks for text classification by exploiting label relations

Text classification has been a fundamental problem in the realm of Natural Language Processing (NLP), and a variety of approaches have been proposed with the development of deep learning. Despite recent progress, most existing approaches deal with the problem of multi-class text classification in a flat way, assuming that text labels are semantically independent. This assumption doesn't always hold in realistic settings, since there usually exist hierarchical or within-layer dependencies in the latent label space, especially when we consider larger label sets. In this paper, we propose a label clustering algorithm to exploit the underlying structure of label relations, and express the stacked concept relationships in the form of a two-layer label space. Next, we present two different neural network structures to capture inter-layer and intra-layer label relations. The first model HSNN organizes a group of local classifiers in a hierarchical way according to the exploited label space, while the second model LSNN takes advantages of text representations in different granularity levels and the bidirectional inferences with recurrent connections to make predictions. Finally, we evaluate our methods on three benchmark datasets. The results empirically demonstrate that both models are capable of leveraging the exploited label relations to improve text classification performance.