Skip to content

Latest commit

 

History

History
77 lines (67 loc) · 2.72 KB

README.md

File metadata and controls

77 lines (67 loc) · 2.72 KB

THiFLY Research at SemEval-2023 Task 7

Codes of the SemEval2023 paper: THiFLY Research at SemEval-2023 Task 7: A Multi-granularity System for CTR-based Textual Entailment and Evidence Retrieval

Textual Entailment

Environment

  • torch==1.7.1
  • torch-scatter==2.0.5
  • transformers==4.18.0
  • tensorboardX==1.8
  • pytorch-pretrained-bert==0.6.2

Please download the necessary pre-trained models from huggingface.

Train & Evaluate

For inference models, you may directly run the command:

 cd textual_entailment
 FOLD=0
 python xxx.py --do_train --do_eval --fold $FOLD --output_dir xxx_${FOLD}

Here xxx can be {512_bi_bi_mul, 512_tf_bi_cl, 1024_tf_bi_mul, scifive}. The argument $FOLD can be varied from 1 to 10. To evaluate an model, you may run

  FOLD=0
  python xxx.py --do_eval --fold $FOLD --output_dir xxx_${FOLD} --load_dir xxx_${FOLD}/saved_model

For joint inference network, run

  python run_joint_inference.py --do_train --do_eval

and

  python run_joint_inference.py --do_eval --load_dir outputs_biolinkbert_joint_inference/saved_model

for training and evaluating.

Getting Results

Once you get all the results from the inference models and the joint inference network, run

python ensemble_avg.py

to ensemble the models and get the final results.

Evidence Retrieval

Environment

  • torch==1.7.1
  • tqdm==4.64.1
  • scikit-learn==1.0.2
  • transformers==4.24.0

Train & Evaluate

Please download the necessary pre-trained models from huggingface and Then save it to "biolinkbert/".

# For training models, you may directly run the command:
# biolinkbert_new_sentpooling_sentinter
python run_fold0.py --model biolinkbert_new_sentpooling_sentinter

# biolinkbert_new_sentpooling_sentinter_block
python run_fold0.py --model biolinkbert_new_sentpooling_sentinter_block

# biolinkbert_sent_tokenpooling_sentinter
python run_fold0.py --model biolinkbert_sent_tokenpooling_sentinter

Please note that for each model you should train on the fold0-9 dataset and save their top2 ckpt. You can directly run "run_fold0-9.py" to use the corresponding training set.

Getting Results

Once you get all the ckpt, you need to modify some variables in the test_ensemble_1.sh file (change to your ckpt path), and then run the script file to get intermediate results:

./test_ensemble_1.sh

In order to help you get test.json results easily, we have saved the intermediate results in the path "test_ensemble_score_5e_tokenmaxpooling_block_63/", you can use them directly

Once you have all the intermediate results and save it to "test_ensemble_score_5e_tokenmaxpooling_block_63", you can run:

python results_ensemble_ave_yuzhi.py

and you will get the final results here: SemEval/final_results/