Improving Search in Social Media Images with External Information

Abstract

The use of social media has increased considerably the recent years, andusers share a lot of their daily life in social media. Many of the users uploadimages to photo-sharing applications, and categorize their images withtextual tags. Users do not always use the best tags to describe the images,but add tags to get "likes" or use tags as a status update. For this reason,searching on tags are unpredictable, and does not necessary return the resultthe user expected.This thesis studies the impact of expanding queries in image searches withterms from knowledge bases, such as DBpedia. We study the methodsTF-IDF, Mutual Information and Chi-square to nd related candidates forquery expansion. The thesis reports on how we implemented and appliedthese methods in a query expansion setting. Our experiments show thatChi-square is the method that yields the best result with the best averageprecision, and was slightly better than a search without query expansion.TF-IDF gave the second best result with query expansion, and Mutual informationwas the method that gave the worst average precision. Queryexpansion with related terms is an exiting eld, and the information fromthis thesis gives a good indication that this is a eld that should be moreexplored in the future.