Semi-supervised latent Dirichlet allocation for multi-label text classification

Abstract

This paper proposes a semi-supervised latent Dirichlet allocation (ssLDA) method, which differs from the existing supervised topic models for multi-label classification in mainly two aspects. Firstly both labeled and unlabeled learning data are used in ssLDA to train a model, which is very important for reducing the cost by manually labeling, especially when obtaining a fully labeled dataset is difficult. Secondly ssLDA provides a more flexible training scheme that allows two ways of labeling assignment while existing topic model-based methods usually focus on either of them: (1) a document-level assignment of labels to a document; (2) imposing word-level correspondences between words and labels within a document. Our experiment results indicate that ssLDA gains an advantage over other methods in implementation flexibility and can outperform others in terms of multi-label classification performance. © 2013 Springer-Verlag.