Abstract: Most approaches to classifying media content assume a fixed, closed vocabulary of labels. In contrast, we advocate machine learning approaches which take advantage of the millions of free-form tags obtainable via online crowd-sourcing platforms and social tagging websites. The use of such open vocabularies presents learning challenges due to typographical errors, synonymy, and a potentially unbounded set of tag labels. In this work, we present a new approach that organizes these noisy tags into well-behaved semantic classes using topic modeling, and learn to predict tags accurately using a mixture of topic classes. This method can utilize an arbitrary open vocabulary of tags, reduces training time by 94% compared to learning from these tags directly, and achieves comparable performance for classification and superior performance for retrieval. We also demonstrate that on open vocabulary tasks, human evaluations are essential for measuring the true performance of tag classifiers, which traditional evaluation methods will consistently underestimate. We focus on the domain of tagging music clips, and demonstrate our results using data collected with a human computation game called TagATune.