Research On Flickr Photo Tag Recommending Based On User Comments

In the Web2.0 era, user generated content has become the main source of information to many websites such as Flickr. In the www.flickr.com website, each user can share his photos by uploading and browse others’ at the same time.Tagging system is an important approach to manage the photos in Flickr. Users can also browse all the tagged photos in Flickr by clicking the tag.However, there is a problem in Flickr that many photos hava little tag because only the uploader can mark tags to this photo. Meanwhile, when the user browsers the photos he interested in, he may make comments to express his understanding of the photo. Every user has his own ideas when seeing the same photo. So it becomes meaningful to conduct research on Flickr photo tag recommending based on user comments. Thus, the main work is as follows in this paper :We analyze the tag and user comments feature to solve five problems including "how do users tag", "what do users tag", "number of user comments to each photo", "number of words to each user comment" and "what kind of user comments". And we transform finding method of recommending Flickr photo tag based on user comments to the problem of traditional keywords automatic tagging.In the phase of generating candidate tag, we proposed two strategies, among which one is based on words, the other is based on phrases. Candidate tag based on words generation process consists of five natural language processing modules including word segmentation, stopwords removing, stemming, spell check and correction as well as part of speech tagging. Candidate tag based on phrases generation process consists of shallow parsing, word semantic calculation, phrase semantic calculation, hierarchical clustering and extension from cluster center. The experimental results show that the phrases based method is superior to the method based on words in tag recommenation. But the phrases based method has the problem of high computation complexity.In the phase of sorting and recommending tag, we proposed an algorithm considering the location of the candidate tag, statistical information and semantic similarity. This algorithm improves the iterative formula in Text Rank by joining TFIDF features and refining edge weights calculation in the word graph model. The experimental results show that the recommending result by using our algorithm on the data set of this research is better than the existing algorithms, such as TF-IDF, LDA and Text Rank.