AMR Parsing beyond English: An Experiment on Bulgarian, French, Hungarian and Ukrainian

Abstract

Under the assumption that the meaning of a sentence should be unchanged when it is translated into another language, recent work has developed on cross-lingual semantic parsing in an effort to extend the access to semantic resources beyond English. In this paper, we develop the automatic production of Abstract Meaning Representations (AMR), a graph-based semantic formalism, for four languages – Bulgarian, French, Hungarian and Ukrainian. We achieve high-performance on French and Hungarian, and execute, to our knowledge, the first semantic parsing of Bulgarian and Ukrainian on translations of the AMR3.0 corpus (Knight et al., 2020). Furthermore, we perform a complementary experiment on a novel parallel corpus of gold AMR annotations of the first chapter of "The Adventures of Pinocchio" in Bulgarian and Ukrainian. The experiment reveals that, despite their above-average performance, the models’ performance decreases when probed on texts outside of the domain of the training data.

Resources

Details