In this paper we investigate whether it is possible to create a computational approach that allows us to distinguish topical tags (i. e. , talking about the topic of a resource) and non-topical tags (i. e. , describing aspects of a resource that are not related to its topic) in folksonomies , in a way that correlates with humans. Towards this goal , we collected 21M tags (1. 2M unique terms) from Delicious and we developed an unsupervised statistical algorithm that classifies such tags by applying a word space model adapted to the folksonomy space. Our algorithm analyses the co-occurrence network of tags to a target tag and exploits graph-based metrics for their classification. We validated its outcomes against a reference classification made by humans on a limited number of terms in three separate tests. The analysis of the outcomes of our algorithm shows , in some cases , a consistent disagreement among humans and between humans and our algorithm about what constitutes a topical tag , and suggests the rise of a new category of overly generic tags (i. e. , umbrella tags) .