Topic representation: A novel method of tag recommendation for text

With tags widely used in organizing and searching contents in massive data era, how to automatically generate appropriate tags of resource for users became a hot issue on social networks research. Tag recommendation for text resource can be modeled as a keyword extraction problem, hence topic modeling such as LDA which extracts latent semantic topics from text is suitable for tag recommendation. However, latent topics are too coarse-grained to describe resource. Meanwhile, LDA trains corpus globally without considering context information. Besides, topics generated are difficult to be quantifiably represented. These problems lead to the poor quality of tag recommendation in topic model based method. In this paper, we propose topic representation method, which introduces embedding semantic representation into topic model. Our results of evaluation on real social networks show that the proposed method improves the quality of tag recommendation for Chinese text resource, when comparing with traditional LDA-based method, which demonstrates the effectiveness of modifying topic modeling.