Unlearn dataset bias by fitting the residual

This repository (the drift branch) contains code for the paper

Unlearn dataset bias for natural language inference by fitting the residual. He He, Sheng Zha, and Haohan Wang. In proceedings of the DeepLo Workshop at EMNLP 2019. https://arxiv.org/abs/1908.10763

Dependencies

Python 3.6
MXNet 1.5.0, e.g., using cuda-10.0, pip install mxnet-cu100
GluonNLP 0.8.0

Install all Python packages: pip install -r requirements.txt

Data

SNLI, MNLI

mkdir -p data/glue_data
python scripts/download_glue_data.py --tasks MNLI --data_dir data/glue_data
python scripts/download_glue_data.py --tasks SNLI --data_dir data/glue_data

HANS

git clone https://github.com/tommccoy1/hans.git
mkdir -p data/glue_data/MNLI-hans
python scripts/hans_to_glue.py --hans-data hans/heuristics_evaluation_set.txt --outdir   data/glue_data/MNLI-hans
rm -rf hans

Datasets will be found in data/glue_data/{SNLI,MNLI,MNLI-hans}.

Code

Entry point: src/main.py.

Main options

Complete options are documented in src/options.py.

task: which dataset to load.
test-split: which split to use for validation (training) or evaluation (testing).
output-dir: directory to save models and artifacts. A UUID will be automatically generated to create a subdirectory under output-dir.
model-type: which model to use, e.g. bert, cbow etc.
max-num-examples: maximum number of examples to load.

NOTE: The code will automatically download files (pretraind models, embeddings etc.) through MXNet. These files will be saved in MXNET_HOME (default directory is ~/.mxnet), which can take up a large space. You might want to set MXNET_HOME to a different directory.

We use MNLI as the example training data below, but you can easily switch to SNLI by modifying the following options: --task SNLI --test-split dev.

We use a default batch size of 32 and the Adam optimizer.

Training NLI models by MLE (baselines)

BERT finetuning

MXNET_HOME=.mxnet MXNET_GPU_MEM_POOL_TYPE=Round GLUE_DIR=data/glue_data \
python -m src.main --task-name MNLI --test-split dev_matched \
--epochs 4 --lr 2e-5 --log-interval 5000 --output-dir output/MNLI \
--model-type bert

Decomposable Attention

MXNET_HOME=.mxnet MXNET_GPU_MEM_POOL_TYPE=Round GLUE_DIR=data/glue_data \
python -m src.main --task-name MNLI --test-split dev_matched \
--epochs 30 --lr 1e-4 --log-interval 5000 --output-dir output/MNLI \
--model-type da --hidden-size 300 --early-stop

ESIM

MXNET_HOME=.mxnet MXNET_GPU_MEM_POOL_TYPE=Round GLUE_DIR=data/glue_data \
python -m src.main --task-name MNLI --test-split dev_matched \
--epochs 30 --lr 1e-4 --log-interval 5000 --output-dir output/MNLI \
--dropout 0.5 --model-type esim --hidden-size 300 --early-stop --max-len 64

Training biased models

Hypothesis-only (finetuned BERT with the hypothesis sentence as the input)

MXNET_HOME=.mxnet MXNET_GPU_MEM_POOL_TYPE=Round GLUE_DIR=data/glue_data \
python -m src.main --task-name MNLI --test-split dev_matched \
--epochs 4 --lr 2e-5 --log-interval 5000 --output-dir output/MNLI \
--model-type bert --superficial hypothesis

CBOW

MXNET_HOME=.mxnet MXNET_GPU_MEM_POOL_TYPE=Round GLUE_DIR=data/glue_data \
python -m src.main --task-name MNLI --test-split dev_matched \
--epochs 40 --lr 1e-4 --log-interval 5000 --output-dir output/MNLI \
--model-type cbow --early-stop

Handcrafted features

MXNET_HOME=.mxnet MXNET_GPU_MEM_POOL_TYPE=Round GLUE_DIR=data/glue_data \
python -m src.main --task-name MNLI --test-split dev_matched \
--epochs 40 --lr 1e-4 --log-interval 5000 --output-dir output/MNLI \
--model-type cbow --early-stop --superficial handcrafted

Training NLI models by DRiFt (based on the biased models)

To train (debiased) models by the residual fitting algorithm, DRiFt, we need to have the biased models ready. Then, we reuse the MLE training script for BERT, DA, and ESIM, with the following updated/new options:

--additive-mode all: learning with an ensemble of the (fixed) biased model and the debiased model.
--additive <path>: path to directory of the biased model.
--epochs <num_epochs>: 8 for BERT; 40 for DA and ESIM.

Training NLI models by RM (based on the biased models)

RM simply removes examples predicted correctly by DRiFt. Similarly, we reuse the MLE training script for BERT, DA, and ESIM, with the following updated/new options:

--additive-mode last: equivalent to learning the debiased model by MLE.
--additive <path>: path to directory of the biased model.
--epochs <num_epochs>: 8 for BERT; 40 for DA and ESIM.
--remove: remove biased examples.

Training models on data with synthesized cheating features

--cheat <cheat_rate>: during training, set the cheating rate. At test time, set cheating rate to 0 such that examples will be prepended with random lables. A cheating rate of -1 means that no cheating feature is added, i.e. normal data.
--remove-cheat: remove cheated examples during training.

Evaluation on HANS

--test-split <HANS-split>: for HANS, the valid splits are lexical_overlap, subsequence, constituent.
--additive-mode last: only used the debiased model for testing. For models trained by MLE, this option has no effect.
--output-dir <path>: logs and predictions for each example (predictions.tsv) will be saved in path.

MXNET_HOME=.mxnet MXNET_GPU_MEM_POOL_TYPE=Round GLUE_DIR=data/glue_data \
python -m src.main --task-name HANS --eval-batch-size 128 --mode test \
--init-from <path> --output-dir eval/HANS --test-split <HANS-split> \
--additive-mode last

Utilities

scripts/summarize_results.py: summarize evaluation results from multiple runs and print in a tabular form.
scripts/error_analysis.py: simple classification error statistics based on predictions.tsv.

LinkToPast1990/debiased