Probabilistic, Structure-Aware Algorithms for Improved Variety, Accuracy, and Coverage of AMR Alignments

We present algorithms for aligning components of Abstract Meaning\nRepresentation (AMR) graphs to spans in English sentences. We leverage\nunsupervised learning in combination with heuristics, taking the best of both\nworlds from previous AMR aligners. Our unsupervised models, however, are more\nsensitive to graph substructures, without requiring a separate syntactic parse.\nOur approach covers a wider variety of AMR substructures than previously\nconsidered, achieves higher coverage of nodes and edges, and does so with\nhigher accuracy. We will release our LEAMR datasets and aligner for use in\nresearch on AMR parsing, generation, and evaluation.\n