Towards the Unsupervised Acquisition of Implicit Semantic Roles

This paper describes a novel approach to find evidence for implicit semantic roles. Our data-driven models generalize over large amounts of explicit annotations only, in order to acquire information about implicit roles. We establish a generic background knowledge base of probablistic predicate-role co-occurrences in an unsupervised manner, and estimate thresholds which trigger the prediction of a missing role. Our approach outperforms the stateof-the-art in terms of recognition rate and offers a more flexible alternative to rulebased solutions which rely on costly, language and domain-specific lexica.