Towards a Better Semantic Role Labeling of Complex Predicates

We propose a way to automatically improve the annotation of verbal complex predicates in PropBank which until now has been treating language mostly in a compositional manner. In order to minimize the manual re-annotation effort, we build on the recently introduced concept of aliasing complex predicates to existing PropBank rolesets which encompass the same meaning and argument structure. We suggest to find aliases automatically by applying a multilingual distributional model that uses the translations of simple and complex predicates as features. Furthermore, we set up an annotation effort to obtain a frequency balanced, realistic test set for this task. Our method reaches an accuracy of 44% on this test set and 72% for the more frequent test items in a lenient evaluation, which is not far from the upper bounds from human annotation.