learning-dynamics

If running this code in the Oscar (SLURM), first run ./scripts/env_setup.sh to generate the virtual environment. Running source ./scripts/load_env.sh loads necessary modules and the virtual environment.

Dataset Creation/Experiment Running

Ontonotes

1. Download ontonotes-release-5.0 from LDC
2. Put ontonotes into 2012 CoNLL task format using *_gold_skel template and scripts with steps here
3. Extract phrase_start, phrase_end, and ner using the data_gen.read_onto_notes_format

Universal Dependencies

Penn Treebank 3.0

To run experiments, please specify parameters and make changes in the *.yaml files in configs/ and then run the python script with the dataset name. To run batches of experiments, please modifyl pipeline.sh or slurm_pipeline.sh if you have a slurm environment.

# Sets up folders.
setup.sh 

# Create dataset.
sbatch configs/datasets/pipeline.sh
sbatch configs/experiments/pipeline.sh

Model	CoNLL2003 NER	CoNLL2003 CPOS	CoNLL2003 FPOS	OntoNotes NER	en_ewt-ud DEP	en_ewt-ud CPOS	en_ewt-ud FPOS	ptb_3 DEPTH	ptb_3 DISTANCE
bert-base-cased (Layer 7)	0.96	0.87	0.88	0.96	0.85	0.95	.95	(0.8776, 0.8585)	(0.8015, 0.8342)