Using graphical models for PP attachment

PP attachment has attracted considerable interest the last two decades. The standard set-up has been to extract quadruples of verbs, direct objects, prepositions and prepositional complements from the Wall Street Journal and classify them as either low or high attachments. State-of-the-art results almost equal human performance in the standard set-up. Recently, however, Atterer and Schütze (2007) has questioned this methodology. In this paper, we show that state-of-theart results can be achieved by simpler means than what has previously been shown, using graphical models, but also that state-of-theart parsers perform insignificantly worse than state-of-the-art PP attachment classifiers. This questions the usefulness of previous studies of PP attachment, even if the methodology in these studies is sound.