Enhancements on a Transition-Based Approach for AMR Parsing Using LSTM Networks

Roxana Pop, Anda Dregan, Florin Cristian Macicasan, Camelia Lemnaru, Rodica Potolea

Abstract: This work proposes two enhancements to a system of generating Meaning Representations (AMR) graphs from English textual data. We first enhance a transition-based approach with additional actions that aim to handle particularities in the structure of the AMR. We analyze actions to address multi-aligned nodes and non-projective word orders, and explore several algorithms for action sequence generation, which incorporate the newly proposed actions. Secondly, we explore strategies for tackling AMR re-entrant concepts, which represent co-references in the associated textual data. We choose to handle co-reference detection and resolution via specific pre-processing and post-processing operations.