Main experiment of "STGN: an Implicit Regularization Method for Learning with Noisy Labels in Natural Language Processing" (EMNLP 2022) by Tingting Wu, Xiao Ding, Minji Tang, Hao Zhang, Bing Qin, Ting Liu.

Note:

• To reproduce the paper results, you can run the stable version 'v5.0' on tesla_v100-sxm2-16gb. However, the noise on SST is not in strictly uniform distribution.
• We will fix data with uniform distribution and adjust code in dev branch.

Experiment on SST, NoisyNER and wikiHow:

• SST:

  • our code

• NoisyNER:

  • our code
  • original work

• wikiHow:

  • our code
  • original work

• BERT bert-base-uncased, batch_size=32, epochs=10, Adam(lr=1e-5)
  • tesla_v100-sxm2-16gb 0.2h/run; for GMMP: 0.8h/run.

• cmd_args_sst.py The command arguments.
• tm_train_hy_nruns.py The entry and main code for BERT. Read paramters from json file, run experiments and log results.
• data/
  • sst_dataset.py Dataloader and Models for BERT.
  • sst/ SST Dataset for 5-class classification.
• common/ Useful codes.
• choose_params/ Best params for BERT.

You need to make sure the output folder exists.

• with sbatch:

mkdir -p ../sst-bert-output/output
sbatch base.sh

• without sbatch:

mkdir -p ../sst-bert-output/output
bash base.sh >../sst-bert-output/output/base.txt 2>../sst-bert-output/output/base.err

You can also use run it as a nohupping backgrounded job.

For BERT:

#!/bin/bash
            
#SBATCH -J SST_base
#SBATCH -p compute
#SBATCH -N 1
#SBATCH --gres gpu:tesla_v100-sxm2-16gb:1
#SBATCH -t 10:00:00
#SBATCH --mem 20240
#SBATCH -e ../sst-bert-output/output/SST_base.err
#SBATCH -o ../sst-bert-output/output/SST_base.txt

source ~/.bashrc
conda activate base

dataset=SST
noise_rate=0.0
method=base
i=0

for noise_rate in 0.0 0.2 0.4 0.6
do
  for i in 0 
  do
    echo "${i}"
    python tm_train_hy_nruns.py \
    --dataset $dataset \
    --noise_rate $noise_rate \
    --seed $i \
    --exp_name ../sst-bert-output/nrun/$dataset-$method/nr$noise_rate/seed$i \
    --params_path choose_params/$dataset/$method/best_params$noise_rate.json
  done
done

You can change arguments for different experiments.

• dataset
  • We provide ['SST', 'QQP', 'MNLI']
  • For ['QQP','MNLI'], We provide experimental parameters for ['SLN', 'STGN'] under 40% uniform label noise.
• method
  • You can choose ['base', 'GCE', 'GNMO', 'GNMP', 'SLN', 'STGN', 'STGN_GCE']
• noise_rate
  • For 'base', you can choose [0.0, 0.1, 0.2, 0.4, 0.6].
  • For other methods, you can choose [0.1, 0.2, 0.4, 0.6].
• seed (i)
  • You may try many different seeds to analyse the method performance, since the seeds make a difference on the results(peak test acc). For example, you can choose [0, 1, 2, 3, 4].

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

@inproceedings{wu-etal-2022-stgn,
    title = "{STGN}: an Implicit Regularization Method for Learning with Noisy Labels in Natural Language Processing",
    author = "Wu, Tingting  and
      Ding, Xiao  and
      Tang, Minji  and
      Zhang, Hao  and
      Qin, Bing  and
      Liu, Ting",
    editor = "Goldberg, Yoav  and
      Kozareva, Zornitsa  and
      Zhang, Yue",
    booktitle = "Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing",
    month = dec,
    year = "2022",
    address = "Abu Dhabi, United Arab Emirates",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.emnlp-main.515",
    doi = "10.18653/v1/2022.emnlp-main.515",
    pages = "7587--7598",
}