Unsupervised Induction of Compositional Classes for English Adjective-Noun Pairs

The paper examines how adjectival modification classes can be detected by applying unsupervised methods on adjective-noun co-occurrence data. It evaluates a k-means baseline, two graphical models, and a recently introduced bidirectional clustering algorithm against HeiPLAS, a manually annotated gold standard for hidden modificational classes. The paper shows that the bidirectional clustering algorithm performs best on this task, and discusses how the results of the unsupervised approaches can be employed for building a frame-based inventory of adjectival modification.