How useful are Enhanced Universal Dependencies for semantic interpretation?

Abstract

We discuss the role of enhanced Universal Dependencies (E-UD) in the task of deriving semantic predicate-argument structures from UD treebanks in a universal, non-language-specific way. We consider the usefulness of three kinds of E-UD annotation (controllers of xcomps, propagation of outgoing dependencies in coordinations, and coreference in relative clauses) and assess some heuristics for automatically adding such enhancements. We conclude that one large obstacle both for deriving predicate-argument structures from UD treebanks and for the automatic enhancement of basic UD treebanks is the fact that UD does not represent empty elements such as pro-dropped arguments, and we suggest that devoting effort to this would often provide a better return on investment than spending resources on improving or adding E-UD annotations.