Preprint submitted to Information Processing & Management Tags are a convenient way to label resources on the web. An interesting question is whether one can determine the semantic meaning of tags in the absence of some predefined formal structure like a thesaurus. Many authors have used the usage data for tags to find their emergent semantics. Here, we argue that the semantics of tags can be captured by comparing the contexts in which tags appear. We give an approach to operationalizing this idea by defining what we call paradigmatic similarity: computing co-occurrence distributions of tags with tags in the same context, and comparing tags using information theoretic similarity measures of these distributions, mostly the Jensen-Shannon divergence. In experiments with three different tagged data collections we study its behavior and compare it to other distance measures. For some tasks, like terminology mapping or clustering, the paradigmatic similarity seems to give better results than similarity measures based on the co-occurrence of the documents or other resources that the tags are associated to. We argue that paradigmatic similarity, is superior to other distance measures, if agreement on topics (as opposed to style, register or language etc.), is the most important criterion, and the main differences between the tagged elements in the data set correspond to different topics