Image retrieval methods mainly depend on image tags. Nonetheless, images are uploaded to WEB with noisy image tags or not a single tag at all. On the other hand, generating tags for every image on the WEB is a time consuming job. In this paper, we present a novel method to improve image retrieval problem using additional text corpora which resides in WEB. Proposed method works in two modes, each of which tries to solve following problems separately; 1) cleaning irrelevant image tags from a noisy image tag list, 2) generating a tag list for an untagged image. In this manner, we designed two models which operates on bag of visual words namely Statistical Matching Model and Variance Model. Test results show that proposed methods are very successful in terms of both image tag cleaning and tag generation.