Abstract: NLU models often exploit biased features to achieve high dataset-specific performance without properly learning the intended task. Recently proposed debiasing methods are shown to be effective in mitigating this tendency. However, these methods rely on a major assumption that the type of biased features is known a-priori, which limits their application to many NLU tasks and datasets. In this work, we present the first step to bridge this gap by introducing a self-debiasing framework that prevents models from mainly utilizing biases without knowing them in advance. The proposed framework is general and complementary to the existing debiasing methods. We show that the proposed framework allows these existing methods to retain the improvement on the challenge datasets (i.e., sets of examples designed to expose models’ reliance to biases) without specifically targeting certain biases. Furthermore, the evaluation suggests that applying the framework results in improved overall robustness.

The repository contains the code to reproduce our work in debiasing NLU models without prior information on biases. We provide 3 runs of experiment that are shown in our paper:

1. Debias MNLI model from syntactic bias and evaluate on HANS as the out-of-distribution data using example reweighting.
2. Debias MNLI model from syntactic bias and evaluate on HANS as the out-of-distribution data using product of expert.
3. Debias MNLI model from syntactic bias and evaluate on HANS as the out-of-distribution data using confidence regularization.

The code requires python >= 3.6 and pytorch >= 1.1.0.

Additional required dependencies can be found in requirements.txt. Install all requirements by running:

pip install -r requirements.txt

Our experiments use MNLI dataset version provided by GLUE benchmark. Download the file from here, and unzip under the directory ./dataset The dataset directory should be structured as the following:

└── dataset 
    └── MNLI
        ├── train.tsv
        ├── dev_matched.tsv
        ├── dev_mismatched.tsv
        ├── dev_mismatched.tsv

For each evaluation setting, use the --mode arguments to set the appropriate loss function. Choose the annealed version of the loss function for reproducing the annealed results.

To reproduce our result on MNLI ⮕ HANS, run the following:

cd src/
CUDA_VISIBLE_DEVICES=9 python train_distill_bert.py \
  --output_dir ../experiments_self_debias_mnli_seed111/bert_reweighted_sampled2K_teacher_seed111_annealed_1to08 \
  --do_train --do_eval --mode reweight_by_teacher_annealed \
  --custom_teacher ../teacher_preds/mnli_trained_on_sample2K_seed111.json --seed 111 --which_bias hans

To obtain predictions of the shallow models, we train the same model architecture on the fraction of the dataset. For MNLI we subsample 2000 examples and train the model for 5 epochs. For obtaining shallow models of other datasets please see the appendix of our paper. The shallow model can be obtained with the command below:

cd src/
CUDA_VISIBLE_DEVICES=9 python train_distill_bert.py \
 --output_dir ../experiments_shallow_mnli/bert_base_sampled2K_seed111 \
 --do_train --do_eval --do_eval_on_train --mode none\
 --seed 111 --which_bias hans --debug --num_train_epochs 5 --debug_num 2000

Once the training and the evaluation on train set is done, copy the probability json files in the output directory to ../teacher_preds/mnli_trained_on_sample2K_seed111.json.

Results on the MNLI ⮕ HANS setting without annealing: