Dozat & Manning (2017; ICLR) describe the basics of what in the Following will be called the Dozat Parser. In 2017 and 2018, it has repeatedly achieved state-of-the-art performance across languages and types of (syntactic) dependency representations. Dozat & Manning (2018; ACL) generalized the technique for general graphs and report state-of-the-art results on the standard semantic dependency parsing (SDP) benchmarks. Owing to its outstanding performance and applicability to both syntactic dependency trees and semantic dependency graphs, the Dozat Parser will likely be a relevant tool for a range of users and experiments. However, the software is neither packaged nor documented in a way that makes it easy to train and apply. This page is intended to provide a collection of ‘recipes’ for different use cases, maintained collectively by the NLPL community.
module load nlpl-dozat python3 /projects/nlpl/software/dozat/201812/src/main.py \ --save_dir ~/tmp \ train ParserNetwork --force --noscreen \ --config_file /projects/nlpl/software/dozat/201812/src/config/ewt.cfg