On many photo-sharing social media, e.g., Instagram, a user posting a
photo can add tags, i.e., words describing it. Tags are used for
keyword-based image search. Some tags, however, describe not image
contents but some metadata, e.g., camera names. We propose a method
of determining which tag is more likely to describe the associated
image contents. We determine it based on pairwise comparisons of
tags. Given a pair of tags A and B, we compare three sets of images:
(1) those associated with A, (2) those associated with B, and (3)
those associated with both A and B. If (3) is more similar to (1)
than to (2), A is more likely to describe image contents. To compute
similarity between two image sets, we use the similarity of their
associated tags, instead of their visual similarity. Our experiment
with Instagram data shows that our method is effective.