/Financial_Event_Ranking

Primary LanguagePython

Financial Event Ranking

This is the implementation of our SIGIR 2021 paper: Hybrid Learning to Rank for Financial Event Ranking.

Requirements

numpy==1.19.2 torch==1.4.0 tqdm==4.49.0 transformers==3.2.0

To create an environment with Conda:

git clone https://github.com/fulifeng/Financial_Event_Ranking.git
cd Financial_Event_Ranking
conda create -n financial_event_ranking
conda activate financial_event_ranking
pip install -r requirements.txt

Dataset

Please send an email to ___ if you need a copy of our data.

Training and Testing

Ret

  • Training
cd Ret
python train_ret.py --lr 1e-05 --maxsteps 5000 --warmup_steps 100 --bsize 128 --accum 2 --do_eval_steps 100 --print_log_steps 100 --model_save_dir ./model_metal_ret
  • Testing(just set --maxstep=0)
python train_ret.py --maxsteps=0 --model_to_test ./model_metal_ret_LR1e-05_BSIZE128 --result_save_file metal_ret_test.json

Cla_M

  • Training
cd Cla_M
python train_clam.py --lr 1e-05 --maxsteps 5000 --warmup_steps 100 --bsize 128 --accum 2 --do_eval_steps 100 --print_log_steps 100 --model_save_dir ./model_metal_clam
  • Testing(with --maxsteps=0)
python train_clam.py --maxsteps=0 --model_to_test ./model_metal_clam_LR1e-05_BSIZE128 --result_save_file metal_clam_test.json

HNB_CNN

  • First obtain the result from Ret and Cla_M(with --augment)
cd Ret
python train_ret.py --maxsteps=0 --model_to_test ./model_metal_ret_LR1e-05_BSIZE128 --result_save_file metal_ret_hybrid_test.json --augment
cd ../Cla_M
python train_clam.py --maxsteps=0 --model_to_test ./model_metal_clam_LR1e-05_BSIZE128 --result_save_file metal_clam_hybrid_test.json --augment
cd ..
  • Training
cd hybrid_method
python train_h.py --lr 1e-04 --maxsteps 50000 --bsize 32 --accum 2 --do_eval_steps 100 --print_log_steps 100 --ret_result_file ../Ret/model_metal_ret_LR1e-05_BSIZE128/metal_ret_hybrid_test.json --clam_result_file ../Cla_M/model_metal_clam_LR1e-05_BSIZE128/metal_clam_hybrid_test.json --rerank_num 100 --model_save_dir ./model_metal_hybrid
  • Testing
python train_h.py --maxsteps 0 --model_to_test ./model_metal_hybrid_LR0.0001_BSIZE32 --result_save_file metal_hybrid_test.json --ret_result_file ../Ret/model_metal_ret_LR1e-05_BSIZE128/metal_ret_hybrid_test.json --clam_result_file ../Cla_M/model_metal_clam_LR1e-05_BSIZE128/metal_clam_hybrid_test.json --rerank_num 100