Code for Simulating Bandit Learning from User Feedback for Extractive Question Answering.
Please contact the first author if you have any questions.
Brief intro for each file:
- train.py: training script
- model.py: model implementaton
- util*.py: codes for evaluation
You can download MRQA datasets from MRQA official repo: training data and in-domain development data.
You can download small sets of supervised examples for initial training from Splinter repo: follow the instruction under "Downloading Few-Shot MRQA Splits".
We suggest you to create a data folder and save all data files there.
-
This project is developed in Python 3.9.5. Using Conda to set up a virtual environment is recommended.
-
Install the required dependencies.
pip install -r requirements.txt -
Install PyTorch from http://pytorch.org/.
You can run the following command to start an online simulation experiment with wandb logging:
python train.py --notes 'your own notes for this experiment if needed' --wandb --do_train --do_eval --model SpanBERT/spanbert-base-cased --seed 46 --train_file ??? --dev_file ??? --max_seq_length 512 --doc_stride 128 --eval_metric f1 --num_train_epochs 1 --eval_per_epoch 8 --output_dir .simulation --initialize_model_from_checkpoint ??? --train_batch_size 80 --eval_batch_size 20 --gradient_accumulation_steps 4 --scheduler constant --algo 'R' --turn_off_dropout --argmax_simulation
??? means the path to file needed by the argument. Please read the argparse code at the bottom of train.py to understand what arguments you could further configure.
To obtain a model initially trained on some supervised data, you are welcome to 1) use training scripts in SpanBERT repo on SQuAD and MRQA datasets, or 2) configure train.py in this repo.
@inproceedings{Gao2022:banditqa-simulation,
title = {Simulating Bandit Learning from User Feedback for Extractive Question Answering},
author = {Gao, Ge and
Choi, Eunsol and
Artzi, Yoav},
booktitle = {Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics},
month = may,
year = {2022},
address = {Dublin, Ireland},
publisher = {Association for Computational Linguistics},
url = {https://aclanthology.org/2022.acl-long.355},
pages = {5167--5179}
}