Adjudication of coreference annotations via finding optimal repairs of equivalence relations

Abstract

We describe encodings for merging multiple coreference annotations into a single annotation, subject to hard constraints (consistency) and optimization criteria (minimal divergence from annotators) using Answer Set Programming (ASP). This task requires guessing an equivalence relation with a large number of elements. We report on experiments with real-world instances based on the METU-Sabanci Turkish Treebank using the ASP tools Gringo, Clasp, and Wasp. We also describe a patch for Gringo which facilitates fine-grained instantiation analysis and use it to analyze experimental results.