How do you feel about "dancing queen"?: deriving mood & theme annotations from user tags

Web 2.0 enables information sharing, collaboration among users and most notably supports active participation and creativity of the users. As a result, a huge amount of manually created metadata describing all kinds of resources is now available. Such semantically rich user generated annotations are especially valuable for digital libraries covering multimedia resources such as music, where these metadata enable retrieval relying not only on content-based (low level) features, but also on the textual descriptions represented by tags. However, if we analyze the annotations users generate for music tracks, we find them heavily biased towards genre. Previous work investigating the types of user provided annotations for music tracks showed that the types of tags which would be really beneficial for supporting retrieval - usage (theme) and opinion (mood) tags - are often neglected by users in the annotation rocess. In this paper we address exactly this problem: in order to support users in tagging and to fill these gaps in the tag space, we develop algorithms for recommending mood and theme annotations. Our methods exploit the available user annotations, the lyrics of music tracks, as well as combinations of both. We also compare the results for our recommended mood / theme annotations against genre and style recommendations - a much easier and already studied task. Besides evaluating against an expert (AllMusic.com) ground truth, we evaluate the quality of our recommended tags through a Facebook-based user study. Our results are very promising both in comparison to experts as well as users and provide interesting insights into possible extensions for music tagging systems to support music search.