• required:
  • python == 3.7
  • cudatoolkit == 10.0.130
  • tensorflow 1.x >=1.13
  • pytorch == 1.2
  • pytorch-transformers
  • bert-serving
  • bert-serving-client
  • numpy == 1.19.5
  • psutil
  • matplotlib
  • scipy

conll 2005 and conll2012 with formation as follow example sentence:

#100002
1	Ms.	_	_	NNP	0	_	_	3:(1,2)-A0	_
2	Haag	_	_	NNP	0	_	_	_	_
3	plays	_	_	VBZ	1	_	_	_	play
4	Elianti	_	_	NNP	0	_	_	3:(4,4)-A1	_
5	.	_	_	.	0	_	_	_	_

each line has ten columns,

• column 5: POS;
• column 6: predicate indicator;
• column 9: arguments;
• column 10: predicate sense.

arguments formation: take "3:(1,2)-A0" as an example, it means predicate with id 3 has an argument span contain tokens from id 1 to id 2 and argument label is 'A0'

example configs are in './configs'. config names with 'argument_detection' are for pruning networks and other configs are for SRL network.

To use elmo embedding, we need to extract token features from elmo model and store it into an HDF5 file as a cache file. First download option "elmo_2x4096_512_2048cnn_2xhighway_5.5B_options.json" and weight "elmo_2x4096_512_2048cnn_2xhighway_5.5B_weights.hdf5" and then use './bilm-tf/usage_cached.py' to extract the cache file and place it in the fold './bilm-tf'

Please download the bert-large uncased checkpoint, and modify the model path in the config file.

Please download the pretrained pytorch RoBERTa-large checkpoint and follow the instruction of "https://github.com/vickyzayats/roberta_tf_ckpt" to transfer it into tensorflow version and place it in './roberta/tf_ckpt/large'.

python3 main.py train SrlParserNetwork  --noscreen --config_file config/conll12/gpred/roberta/firstorder/SRL12.cfg

python main.py --save_dir saves/conll12/gpred/roberta/firstorder/SRL12/SrlParserNetwork run --output_dir TestResult/conll12/gpred/roberta/firstorder/SRL12/SrlParserNetwork data/conll12/test.en.srl.conllu

After test step, we can calculate F1 with official script using example commands as follow With gold predicate:

python scripts/srl_eval.py data/srlconll-1.1/conll12.test.gold TestResult/conll12/gpred/roberta/firstorder/SRL12/SrlParserNetwork/test.en.srl.conllu True

With predict predicate:

python scripts/srl_eval.py data/srlconll-1.1/conll12.test.gold TestResult/conll12/ppred/roberta/firstorder/SRL12/SrlParserNetwork/test.en.srl.conllu False

For conll05 dataset, please replace 'conll12.test.gold' with 'test.wsj.gold' or 'test.brown.gold'.