A Labelled Corpus for Prepositional Phrase Attachment

This paper describes a labelled corpus intended for training learning algorithms to attach prepositional phrases (PPs). Taken from the PTB2, we believe it is the largest available resource for this purpose, especially as it contains many patterns in which PPs occur ambiguously (nearly all previous research has focused on just one pattern) and we present some results for the five most common patterns. Moreover, the corpus contains some features that, to our knowledge, have not been used before for attaching PPs.