Exploring Collaboratively Annotated Data for Automatic Annotation
Jiafeng Guo, Xueqi Cheng, Huawei Shen, Shuo Bai · 2009
Social tagging, as a collaborative form of annotation pro-cess, has grown in popularity on the web due to its effec-tiveness in organizing and accessing web resources. This paper addresses the issue of automatic annotation, which aims to predict social tags for web resources automatically to help future navigation, filtering or search. We explore the collaboratively annotated data in social tagging services, in which collaborative annotations (i.e., combined social tags of many users) serve as a description of web resources from the crowds ’ point of views, and analyze its three important properties. Based on these properties, we propose an au-tomatic annotation approach by leveraging a probabilistic topic model, which captures the relationship between re-sources and annotations. The topic model is an extension of conventional LDA (Latent Dirichlet Allocation), referred as Word-Tag LDA, which effectively reflects the generative process of the collaboratively annotated data and models the conditional distribution of annotations given resources. Ex-periments are carried out on a real-world annotation data set sampled from del.icio.us. Results demonstrate that our ap-proach can significantly outperform the other baseline meth-ods in automatic annotation.