AnCast++: Document-Level Evaluation of Graph-based Meaning Representations

Haibo Sun, Jayeol Chun, Nianwen Xue

Abstract: Uniform Meaning Representation (UMR) is a cross-lingual document-level graph-based representation that is based on Abstract Meaning Representation (AMR) but extends it to include document-level semantic annotations such as coreference, modal and temporal dependencies.With recent advancements in UMR annotation efforts, a reliable evaluation metric is essential for assessing annotation consistency and tracking progress in automatic parsing. In this paper, we present AnCast++, an aggregated metric that unifies the evaluation of four distinct sub-structures of UMR: (1) sentence-level graphs that represent word senses, named entities, semantic relations between events and their participants, aspectual attributes of events as well as person and number attributes of entities, (2) modal dependencies that represent the level of certainty that a source holds with respect to an event, (3) temporal dependencies between events and their reference times, and (4) coreference relations between entities and between events. In particular, we describe a unified method TC2 for evaluating temporal and coreference relations that captures their shared transitive properties, and present experimental results on English and Chinese UMR parsing based on UMR v1.0 corpus to demonstrate the reliability of our metric. The tool will be made publicly available on Github.