---

Implementation of our EMNLP Findings 2023 paper: Learning Semantic Role Labeling from Compatible Label Sequences

@inproceedings{li-etal-2023-learning-semantic,
    title = "Learning Semantic Role Labeling from Compatible Label Sequences",
    author = "Li, Tao  and
      Kazeminejad, Ghazaleh  and
      Brown, Susan  and
      Srikumar, Vivek  and
      Palmer, Martha",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2023",
    year = "2023",
    url = "https://aclanthology.org/2023.findings-emnlp.1041",
    doi = "10.18653/v1/2023.findings-emnlp.1041",
}

---

---

Preprocessing

1. First install dependencies.

pip install -r requirements.txt

2. Data extraction and processing

scripts/semlink/semlink2conll.sh --ptb [PATH_TO_TREEBANK3_PARSED_MRG_DIR] \
--roleset both \
--senses vn \
--brown data/datasets/conll-brown-release/prop.txt

scripts/semlink/semlink2conll_brown.sh --ptb [PATH_TO_TREEBANK3_PARSED_MRG_DIR] \
--roleset both \
--senses vn \
--brown data/datasets/conll-brown-release/prop.txt

Specify the PATH_TO_TREEBANK3_PARSED_MRG_DIR to the right directory.

3. Further extraction and processing

Follow iwcs2021 to extract the joint set of VN and PB. Let PATH_TO_IWCS2021_REPO point to the cloned repository in your system.

Then follow structured_tuning_2021 to process CoNLL 05 dataset. We will need those to continue the extraction. Let CONLL05_DIR points to the processed directory

CONLL_PATH=PATH_TO_IWCS2021_REPO/semlink1.1
python3 process_iwcs2021.py ${CONLL_PATH}/vn.both-valid.txt \
	${CONLL_PATH}/dev.txt \
	${CONLL_PATH}/dev.props.gold.txt \
 	${CONLL_PATH}/dev.propid.txt \
 	${CONLL_PATH}/dev.fileid.txt \
 	${CONLL_PATH}/dev.origfileid.txt

python3 process_iwcs2021.py ${CONLL_PATH}/vn.both-train.txt \
	${CONLL_PATH}/train.txt \
	${CONLL_PATH}/train.props.gold.txt \
 	${CONLL_PATH}/train.propid.txt \
 	${CONLL_PATH}/train.fileid.txt \
 	${CONLL_PATH}/train.origfileid.txt

python3 process_iwcs2021.py ${CONLL_PATH}/vn.both-test-wsj.txt \
	${CONLL_PATH}/test1.txt \
	${CONLL_PATH}/test1.props.gold.txt \
 	${CONLL_PATH}/test1.propid.txt \
 	${CONLL_PATH}/test1.fileid.txt \
 	${CONLL_PATH}/test1.origfileid.txt

python3 process_iwcs2021.py ${CONLL_PATH}/vn.both-test-brown.txt \
	${CONLL_PATH}/test2.txt \
	${CONLL_PATH}/test2.props.gold.txt \
 	${CONLL_PATH}/test2.propid.txt \
 	${CONLL_PATH}/test2.fileid.txt \
 	${CONLL_PATH}/test2.origfileid.txt

4. Final processing to batch up examples

python3 -u -m preprocess.extract_iwcs2021 | tee extract.iwcs2021.log.txt
python3 -u -m preprocess.preprocess_iwcs2021 --output iwcs_wp | tee preprocess.iwcs2021.log.txt
python3 -u -m preprocess.extract_extra_srl
python3 -u -m preprocess.preprocess_extra_srl \
	--arg_role_map iwcs_wp.arg_role.txt \
	--indexer iwcs_wp \
	--output iwcs_wp.extra.conll05


python3 -u -m preprocess.extract_extra_srl --data ./conll05.test.wsj.txt --filter "" --output iwcs.entire.conll05.test1.vn_srl.txt
python3 -u -m preprocess.extract_extra_srl --data ./conll05.test.brown.txt --filter "" --output iwcs.entire.conll05.test2.vn_srl.txt
python3 -u -m preprocess.preprocess_extra_srl \
	--data iwcs.entire.conll05.test2.vn_srl.txt \
	--arg_role_map iwcs_wp.arg_role.txt \
	--indexer iwcs_wp.extra.conll05 \
	--aug_indexer 0 \
	--output iwcs_wp.entire.conll05.test2


python3 -u -m preprocess.preprocess_iwcs2021_completion --output iwcs_completion
python3 -u -m preprocess.preprocess_extra_srl_completion \
	--arg_role_map iwcs_completion.arg_role.txt \
	--indexer iwcs_completion \
	--output iwcs_completion.extra.conll05

Training

1. Train a joint model for the 1st round

for SEED in 1 2 3; do
EP=20
DROP=0.5
LR=0.00003
DATA=iwcs_wp
SIZE=base
ENC=bert_wp
LAMBD=1,1
MODEL=./models/multi_${DATA}_${SIZE}_${ENC//_}_ep${EP}_lr${LR//.}_drop${DROP//.}_multiw${LAMBD//,}_seed${SEED}
python3 -u train.py --gpuid [GPUID] --dir [DATA] \
	--train_data ${DATA}.train.hdf5 --val_data ${DATA}.val.hdf5 \
	--train_res ${DATA}.train.orig_tok_grouped.txt --val_res ${DATA}.val.orig_tok_grouped.txt \
	--data_flag vn_srl \
	--vn_class_dict ${DATA}.vn_class.dict \
	--vn_label_dict ${DATA}.vn_label.dict --srl_label_dict ${DATA}.srl_label.dict \
	--roleset_dict ${DATA}.roleset_label.dict --roleset_suffix_dict ${DATA}.roleset_suffix.dict \
	--arg_role_map ${DATA}.arg_role.txt \
	--enc $ENC --cls multi --loss multi_crf --multi_w ${LAMBD} \
	--optim adamw_fp16 --epochs $EP --learning_rate $LR --dropout $DROP \
	--bert_type roberta-${SIZE} --hidden_size 768 \
	--percent 1 --val_percent 1 \
	--seed $SEED --conll_output $MODEL --save_file $MODEL | tee ${MODEL}.txt
done

Specify the [GPUID] and [DATA].

2. Train for the 2nd round

for SEED in 1 2 3; do
EP=5
DROP=0.5
LR=0.00001
DATA=iwcs_wp
SIZE=base
ENC=bert_wp
LAMBD=1,1
LOAD=./models/multi_${DATA}_${SIZE}_${ENC//_}_ep20_lr000003_drop${DROP//.}_multiw${LAMBD//,}_seed${SEED}
MODEL=./models/multi2_${DATA}_${SIZE}_${ENC//_}_ep${EP}_lr${LR//.}_drop${DROP//.}_multiw${LAMBD//,}_seed${SEED}
python3 -u train.py --gpuid [GPUID] --dir [DATA] \
	--load_file $LOAD \
	--train_data ${DATA}.train.hdf5 --val_data ${DATA}.val.hdf5 \
	--train_res ${DATA}.train.orig_tok_grouped.txt --val_res ${DATA}.val.orig_tok_grouped.txt \
	--data_flag vn_srl \
	--vn_class_dict ${DATA}.vn_class.dict \
	--vn_label_dict ${DATA}.vn_label.dict --srl_label_dict ${DATA}.srl_label.dict \
	--roleset_dict ${DATA}.roleset_label.dict --roleset_suffix_dict ${DATA}.roleset_suffix.dict \
	--arg_role_map ${DATA}.arg_role.txt \
	--enc $ENC --cls multi --loss multi_crf --multi_w ${LAMBD} \
	--optim adamw_fp16 --epochs $EP --learning_rate $LR --dropout $DROP \
	--bert_type roberta-${SIZE} --hidden_size 768 \
	--percent 1 --val_percent 1 \
	--seed $SEED --conll_output $MODEL --save_file $MODEL | tee ${MODEL}.txt
done

Specify the [GPUID] and [DATA].

To evaluate the model, run:

for SEED in 1 2 3; do
for SPLIT in test1 test2; do
EP=5
DROP=0.5
LR=0.00001
SIZE=base
ENC=bert_wp
DATA=iwcs_wp
LAMBD=1,1
EVAL_SRL=1
MODEL=./models/multi2_${DATA}_${SIZE}_${ENC//_}_ep${EP}_lr${LR//.}_drop${DROP//.}_multiw${LAMBD//,}_seed${SEED}
python3 -u eval.py --gpuid [GPUID] --dir [DATA] \
	--data ${DATA}.${SPLIT}.hdf5 \
	--res ${DATA}.${SPLIT}.orig_tok_grouped.txt \
	--data_flag vn_srl \
	--vn_class_dict ${DATA}.vn_class.dict \
	--vn_label_dict ${DATA}.vn_label.dict --srl_label_dict ${DATA}.srl_label.dict \
	--roleset_dict ${DATA}.roleset_label.dict --roleset_suffix_dict ${DATA}.roleset_suffix.dict \
	--arg_role_map ${DATA}.arg_role.txt \
	--enc $ENC --cls multi --loss multi_crf --multi_w ${LAMBD} --eval_with_srl ${EVAL_SRL} \
	--bert_type roberta-${SIZE} \
	--seed $SEED --conll_output ${MODEL} --log pretty --load_file $MODEL | tee ${MODEL}.${SPLIT}.txt
done
done

Specify the [GPUID] and [DATA].