Parsing/stanfordnlp

Background

The StanfordNLP pipeline makes available two multilingual parsing systems with a Python interface, one implemented natively on top of PyTorch, the other a server–client interface to the venerable Stanford CoreNLP software (which runs in Java).

module purge; module load nlpl-stanfordnlp
python3 -c 'import stanfordnlp; \
  foo = stanfordnlp.Pipeline(lang = "en")("Kim wanted to be heard."); \
  foo.sentences[0].print_dependencies();'

It is possible to combine this module with the NLPL installation of CoreNLP and use its server interface to offload processing requests to CoreNLP. In a multi-user environment, it may in principle be necessary to pick a different (non-privileged) port for the server, as each port can only be used by one process at any point in time (see the initialization in test.py):

module purge; module load nlpl-corenlp nlpl-stanfordnlp
python3 /projects/nlpl/software/stanfordnlp/test.py

Installation on Abel

module purge; module load python3/3.70
/projects/nlpl/operation/python/initialize --version 0.1.1 stanfordnlpl

Next, it appears we need to manually patch the default location for model files in .../site-packages/stanfordnlp/utils/resources.py, to point to a shared directory for all users (viz. /projects/nlpl/software/stanfordnlp/0.1.1/resources/). To download all available pre-trained models (for the complete UD 2.x set of treebanks):

module purge; module load nlpl-stanfordnlp
yes | python3 /projects/nlpl/software/stanfordnlp/0.1.1/download.py