In plain words: A single neural system builds meaning graphs by piecing together parts of a sentence, instead of using a hand-designed parser for each graphbank. It came close to top accuracy across graphbanks, and with pretrained word meanings and joint training beat best scores on five.

Abstract

Most semantic parsers that map sentences to graph-based meaning representations are hand-designed for specific graphbanks. We present a compositional neural semantic parser which achieves, for the first time, competitive accuracies across a diverse range of graphbanks. Incorporating BERT embeddings and multi-task learning improves the accuracy further, setting new states of the art on DM, PAS, PSD, AMR 2015 and EDS.

Matthias Lindemann, Jonas Groschwitz, Alexander Koller
arXiv:1906.11746 · cs.CL · submitted Jun 27, 2019 · updated Jul 13, 2019
abstract · pdf · html · Accepted at ACL 2019