The papers in this post describe research on user-generated tags. Most papers have multiple authors, although I've credited by name only the author(s) who presented the paper. Entire papers are available in the conference proceedings; let me know if you'd like to see one.
No Bull, No Spin: a Comparison of Tags with Other Forms of User Metadata
Catherine C. Marshall
The researcher's goal was to answer the question, "Do folksonomies work?" She found Flickr pictures and their tags a good way to study this question, because many unrelated people post very similar pictures that are tagged and otherwise identified in all sorts of ways. She began work by finding as many photos of the same object as possible using a variety of search methods. She searched Flickr for pictures of a specific mosaic in Milan (photo by stephen (MixedMediaExpressions.com)) and recorded user metadata about 603 photos of the mosaic. She recorded user metadata from the following fields: title, photostream, written description, tags, and geotags. Her conclusions about this research are pretty interesting. People seemed to use tags differently from the ways they use other forms of metadata. Their tags tended to focus on a very general definition of place, e.g.: "Milan" rather than "Galleria Vittorio Emanuele II". Since geotags are available, one might users to rely on them as much more accurate place descriptors. Users tended to focus on the story of the mosaic in captions, but incorporated aspects of the artifact, the story, and the place into their titles. Finally, users' tags were more likely to express personal context, e.g.: tagging a photo with the name of the person the user traveled with. Based on these findings, the researcher concluded that captions and titles form a better basis for retrieval and description from public collections. Implications include replacing current tagging methods with geotagging, since current use of tags seems to focus on place.
How Do You Feel about "Dancing Queen"? Deriving Mood & Theme Annotations from User Tags
Kerstin Bischoff, et al
The researchers collected all text associated with the song "Dancing Queen" on Last.FM, and found 4 general types: moods, themes, genres, and styles. The research started with 2 hypotheses: that theme and mood tags would enhance the search capacity (findability) for music, and that song lyrics may be an indicator of appropriate machine-assigned theme or mood tags. After gathering and analyzing the data, the researchers used algorithms to attempt to replicate the data in a recommender service. They found that the relevance of recommendations depended on the type. Genres were easiest to recommend, while styles were very difficult because they are so fine-grained. Mood recommendations were mixed, working best when based on 6 basic human emotions. Theme recommendations were also mixed, since some themes are very subjective. They evaluated the recommender service by creating a Facebook app and examining a couple of basic questions: How well did the service perform compared with human recommendations? How well did average users agree with experts? Their evaluation found that the recommender service showed improved performance over users in matching recommended tags. This research has several practical applications, including the ability to recommend tags to users while they're tagging a piece of music, indexing tags to aid discovery, and creating mood- or theme-based playlists.