Empirically Sampling Universal Dependencies
Natalie Schluter, Żeljko Agić · IT University Of Copenhagen (IT University of Copenhagen) · 2017
Universal Dependencies incur a high cost in computation for unbiased system development. We propose a 100% empirically chosen small subset of UD languages for efficient parsing system development. The technique used is based on measurements of model capacity globally. We show that the diversity of the resulting representative language set is superior to the requirements-based procedure.