|Abstract Meaning Representation (AMR) parsing attempts to extract a structured representation of a sentence's meaning. This paper enhances an existing processing pipeline for AMR parsing, inspired by state-of-the-art solutions in dependency parsing. It enhances the existing Concept Identification module by using Pointer-Generator Networks. A further considerable improvement of this module is brought by the use of embeddings. An alternative approach is provided through Transformers, an architecture which needs large data-sets to accurately predict concepts. For predicting the relations between concepts, the proposed pipeline combines the two Heads Selection and now trainable Arcs Labelling tasks into a joint Relation Identification module, which enhances the overall performance of edge prediction. The improvements made to this AMR parser have resulted in a completely trainable model that can be improved further with end-to-end training.|
*** Title, author list and abstract as seen in the Camera-Ready version of the paper that was provided to Conference Committee. Small changes that may have occurred during processing by Springer may not appear in this window.