The Proper Treatment of Verbal Idioms in German Discourse Representation Structure Parsing

Kilian Evang, Rafael Ehren, Laura Kallmeyer

Abstract: Existing datasets for semantic parsing lack adequate representations of potentially idiomatic expressions (PIEs), i.e., expressions consisting of two or more lexemes that can occur with either a literal or an idiomatic reading. As a result, we cannot test semantic parsers for their ability to correctly distinguish between the two cases, and to assign appropriate meaning representations. We address this situation by combining two semantically annotated resources to obtain a corpus of German sentences containing literal and idiomatic occurrences of PIEs, paired with meaning representations whose concepts and roles reflect the respective literal or idiomatic meaning. Experiments with a state-of-the-art semantic parser show that given appropriate training data, it can learn to predict the idiomatic meanings and improve performance also for literal readings, even though predicting the correct concepts in context remains challenging. We provide additional insights through evaluation on synthetic data.