Check out the Report.pdf
!!
This is and extension of the code used to run the experiments in Deep Subjecthood: Higher Order Grammatical Features in Multilingual BERT.
This has been altered to use English BERT and tag verbs by their type.
run_one_experiment.py
runs one experiment: training a classifier to classify subjects from objects in mBERT embeddings in one training language, and testing this classifier on one test language. Classifiers and mBERT embeddings are cached, to reuse for other experiments with the same train/test language.
make_script_to_run_all_experiments.py
makes a script to run run_one_experiment.py
for every pair of languages in source_langs.txt
and sink_langs.txt
To reproduce the results:
- Change the paths in
source_langs.txt
andsink_langs.txt
, so that they match the location of the Universal Dependency Treebanks on your machine - Run
make_script_to_run_all_experiments.py
to create the scripts that will run the experiments. This will create a scriptrun_batch_0.sh
(or more if you're running for multiple seeds - Run
run_batch_0.sh