We develop a probabilistic latent-variable model to discover semantic frames---types of events and their participants---from corpora. We present a dirichlet-multinomial model in which frames are latent categories