Automatic Document Tagging in Social Semantic Digital Library

The emergence of Web 2.0 has created a lot of annotation and personalization information about web resources. Extracting and utilizing these information to enhance the quality of services is a key target of modern digital libraries. In this paper, we present a novel Automatic Document Tagging (ADT) approach for digital libraries. In our approach, the ADT problem is formulated as a variant of multi-class classification problem. But differently, the training data for ADT is collected from the user's historic tags and only partially labeled. The incompleteness of the training data makes the training a more challenging problem. To overcome this problem, an efficient randomized online training algorithm (RPL) is proposed. RPL algorithm has two phases: (i) random exploitation and (ii) classifier update. The experimental results from both synthetic and real-word data demonstrate the effectiveness.