Skip to content

Latest commit

 

History

History
64 lines (52 loc) · 3.39 KB

README.md

File metadata and controls

64 lines (52 loc) · 3.39 KB
Raw

Few Clean Instances Help Denoising Distant Supervision

This repository contains the PyTorch code for the paper

Yufang Liu*, Ziyin Huang*, Yijun Wang, Changzhi Sun, Man Lan, Yuanbin Wu, Xiaofeng Mou and Ding Wang. Few Clean Instances Help Denoising Distant Supervision. Coling, 2022.

A model-agnostic denoise method for distant supervision relation extraction. We propose a new criterion for clean instance selection based on influence functions. It collects sample-level evidence for recognizing good instancess (which is more informative than loss-level evidence). We also propose a teacher-student mechanism for controlling purity of intermediate results when bootstrapping the clean set.

Setup

Environment: One or more multi-GPU node(s) with the following software/libraries installed:

Data preprocessing: We split the dataset to create training data for each relation classification, please read the paper for more details. We provide a simple dataset business_person_company(bpc) for example.

Running

This repository contains the code for 4 different criteria, each directory contains whole code for one criterion. Here we show the steps to run the code for Assumption 2 with Teacher-Student style update (ByAssump2_TS).

  1. To accelerate the IF calculation process, we split the dataset, run the same code on the different parts and collect the results together. So before running the code, specify specific gpu for each process. For step 1, if you have 4 gpus, and you want to run 2 process on each gpu, set device info in run/run3.step1.shas:

    Devices=("0" "0" "1" "1" "2" "2" "3" "3")

    and set the div number in src/framework/MeanTeacher.py/IF_calc_step1 as 8(gpu_num * process):

     a = subprocess.Popen(["./run3_step1.sh", input_file, '%d'%(self.line_counter(file=input_file)//8+1), '%d'%batch_size,
                                        model_path, '%d'%scale, '%f'%damp, '%s'%self.config_file], shell=False)

    same as for step 2, set device info in run/run3.step1.sh and div number in src/framework/MeanTeacher.py/IF_calc_step2.

  2. Running the code for relation business_person_company as

    CUDA_VISIBLE_DEVICES=0,1,2,3 python run.py --config_file ../configs/default_bpc.cfg

    check the shell script run3_step1.sh and run3_step2.sh have the execute permission, if not, permission denied error will be reported. Simple solution is chmod +x ${file_name} for these two files.

Output

After running one relation classification, a workspace directory for this relation will be generated. It contains all the running statistics, including the selected training data for each iteration, student model, teacher model, and etc.

Citation

If you find this code useful in your research, please cite:

@inproceedings{IFDSRE,
  title={Few Clean Instances Help Denoising Distant Supervision},
  author={Yufang Liu, Ziyin Huang, Yijun Wang, Changzhi Sun, Man Lan, Yuanbin Wu, Xiaofeng Mou and Ding Wang},
  booktitle={The 29th International Conference on Computational Linguistics},
  year={2022}
}