Learning Frames from Text with an Unsupervised Latent Variable Model

We develop a probabilistic latent-variable model to discover semantic frames—types of events or relations and their participants—from corpora. Our key contribution is a model in which (1) frames are latent categories that explain the linking of verb-subject-object triples in a given document context; and (2) cross-cutting semantic word classes are learned, shared across frames. We also introduce an evaluation methodology that compares to FrameNet, interpreting the learned model as an alternative frame lexicon.