ECIMP

Event Causality Identification via Derivative Prompt Joint Learning

论文代码

Requirements

• beautifulsoup4==4.10.0
• numpy==1.18.5
• scikit_learn==1.0.2
• torch==1.10.1+cu113
• tqdm==4.47.0
• transformers==4.14.1
• typing_extensions==4.1.1

运行代码

使用了Timebank和Eventstory两种数据集，且Timebank十折交叉验证 Eventstory五折交叉验证

Roberta+linear baseline

当使用Timebank数据集时:

nohup python -u run_mlm_base.py --main_device=0 --device_ids=0 --data_path=data/timebank.json --mlm_name_or_path=../roberta-base --split_n=10 --gradient_accumulate=1 > log/mlm_base_timebank.log &

使用Eventstory时:

nohup python -u run_mlm_base.py --main_device=0 --device_ids=0,1 --data_path=data/eventstory.json --mlm_name_or_path=../roberta-base --split_n=5 --gradient_accumulate=1 > log/mlm_base_eventstory.log &

以此类推,具体参数配置见run_*.py文件

single prompt

Timebank

nohup python -u run_prompt_tuning.py --main_device=0 --device_ids=0,1 --mlm_name_or_path=../roberta-base --data_path=data/timebank.json --split_n=10 --gradient_accumulate=1 > log/prompt_tuning_timebank.log &

Event Causality Identification via Derivative Prompt Joint Learning

如果是要运行中文数据请修改run_ecimp.py的第92行

Timebank

nohup python -u run_ecimp.py --data_path data/timebank.json --split_n=10 --use_event_prompt=1 --use_signal_prompt=1 --use_sep_gate=1 --use_mask1_gate=1 --reuse=1 --batch_size=6 --use_linear=1 --device_ids=1 --main_device=1 --gradient_accumulate=3 --learning_rate=0.0001 --mlm_name_or_path=../roberta-base > log/ecimp_timebank_full.log &