A constrained graph algebra for semantic parsing with AMRs

Abstract

When learning grammars that map from sentences to abstract meaning representations (AMRs), one faces the challenge that an AMR can be described in a huge number of different ways using traditional graph algebras. We introduce a new algebra for building graphs from smaller parts, using linguistically motivated operations for combining a head with a complement or a modifier. Using this algebra, we can reduce the number of analyses per AMR graph dramatically; at the same time, we show that challenging linguistic constructions can still be handled correctly.