Image tag clarity: in search of visual-representative tags for social images

Tags associated with images in various social media shar-ing web sites are valuable information source for superior image retrieval experiences. Due to the nature of tagging, many tags associated with images are not visually descrip-tive. In this paper, we propose Normalized Image Tag Clar-ity (nitc) to evaluate the effectiveness of a tag in describing the visual content of its annotated images. It is measured by computing the zero-mean normalized distance between the tag language model estimated from the images annotated by the tag and the collection language model. The visual-representative tags that are commonly used to annotate vi-sually similar images are given high tag clarity scores. Eval-uated on a large real-world dataset containing more than 269K images and their associated tags, we show that nitc score can effectively identify the visual-representative tags from all tags contributed by users. We also demonstrate through experiments that most popular tags are indeed vi-sually representative