Extracting a PP Attachment Data Set from a German Dependency Treebank Using Topological Fields

PP-attachment has traditionally been tackled as a binary classification task where a preposition is attached to the immediately preceding noun or to the main verb. In this paper, we provide an analysis of PP-attachment in German to show that the assumption that prepositions have only two head candidates does not hold. We propose a realistic PP-attachment data set, in which each preposition has multiple head candidates. The data set is extracted automatically from a dependency treebank with topological field annotations. Finally, we show that the task of PP-attachment is substantially more difficult with this realistic data set than with a binary classification data set.