Disentangling Instructive Information from Ranked Multiple Candidates for Multi-Document Scientific Summarization
This is the Pytorch implementation for [Disentangling Instructive Information from Ranked Multiple Candidates for Multi-Document Scientific Summarization], accepted by SIGIR 2024.
- Python == 3.6.13
- torch == 1.6.0
- transformers == 4.18.0
- pyrouge == 0.1.3
-
Create folder
trained_model,result,logunder the root directory. -
Download Multi-Xscience Dataset from here. Download TAD and TAS2 Dataset from Paper
-
Generate summary candidates for each instance of the datasets
-
Obtain the ROUGE ranking results of these summary candidates.
export PYTHONPATH=.
python train.py --mode train --cuda --data_dir <path-to-training-dataset-folder> --batch_size 1 --seed 666 --train_steps 26000 --warmup_steps 4000 --save_checkpoint_steps 2000 --report_every 1 --visible_gpus 0 --gpu_ranks 0 --world_size 1 --accum_count 8 --dec_dropout 0.2 --enc_dropout 0.1 --model_path ./trained_model/train_promptctr --log_file ./log/train_source.txt --inter_layers 6,7 --inter_heads 6 --dec_hidden_size 768 --hier --doc_max_timesteps 50 --use_bert true --prop 3 --num_workers 5 --lr 0.001 --enc_layers 6 --dec_layers 6 --use_nucleus_sampling false --label_smoothing 0.1 export PYTHONPATH=.
python train.py --mode test --cuda --data_dir <path-to-test-dataset-folder> --batch_size 8 --valid_batch_size 8 --seed 666 --visible_gpus 0 --gpu_ranks 0 --dec_dropout 0.2 --enc_dropout 0.1 --lr 0.2 --label_smoothing 0.1 --log_file ./log/log_full_test_wordlevel_copy001.txt --inter_layers 6,7 --inter_heads 6 --dec_hidden_size 768 --doc_max_timesteps 50 --use_bert true --report_rouge --alpha 0.4 --max_length 200 --result_path ./resultmx/prompt_ctr_ --prop 3 --test_all false --sep_optim true --use_nucleus_sampling false --min_length 120 --no_repeat_ngram_size 2 --test_from <path-to-saved-reranker-checkpoint> --bce True export PYTHONPATH=.
python train.py --mode train --cuda --data_dir <path-to-training-dataset-folder> --batch_size 2 --seed 666 --train_steps 80000 --save_checkpoint_steps 4000 --report_every 1 --visible_gpus 0 --gpu_ranks 0 --world_size 1 --accum_count 2 --dec_dropout 0.1 --enc_dropout 0.1 --model_path ./trained_model/train_mx_vaeR1 --log_file ./log/train_source.txt --inter_layers 6,7 --inter_heads 8 --hier --doc_max_timesteps 50 --use_bert false --prop 3 --sep_optim false --num_workers 5 --warmup_steps 8000 --lr 0.005 --enc_layers 6 --dec_layers 6 --use_nucleus_sampling false --label_smoothing 0.1 --candidate_type 0shot --loss_kl 0.001 --loss_bow 0.01 --kl_annealing_steps 100000 --cand_num 3 --rank_type Rouge1 export PYTHONPATH=.
python train.py --mode test --cuda --data_dir <path-to-test-dataset-folder> --batch_size 8 --valid_batch_size 8 --seed 666 --visible_gpus 0 --gpu_ranks 0 --dec_dropout 0.1 --enc_dropout 0.1 --lr 0.2 --label_smoothing 0.0 --log_file ./log/log_full_test_wordlevel_copy001.txt --inter_layers 6,7 --inter_heads 8 --doc_max_timesteps 50 --use_bert false --report_rouge --alpha 0.4 --max_length 200 --result_path ./result/model_ --prop 3 --test_all false --sep_optim false --use_bert false --use_nucleus_sampling false --min_length 100 --no_repeat_ngram_size 2 --test_from <path-to-saved-summarization-model-checkpoint> --predrank_path <path-to-predicted-test-ranking-result> --cand_num 3 --enc_layers 6 --dec_layers 6 --rank_type Piratio --use_z2 true@inproceedings{wang2024disentangling,
title={Disentangling Instructive Information from Ranked Multiple Candidates for Multi-Document Scientific Summarization},
author={Wang, Pancheng and Li, Shasha and Tang, Jintao and Wang, Ting},
booktitle={Proceedings of the 47th international ACM SIGIR conference on research and development in information retrieval},
year={2024}
}