Skip to content

Latest commit

 

History

History
94 lines (70 loc) · 2.6 KB

README.md

File metadata and controls

94 lines (70 loc) · 2.6 KB

Domain-agnostic Question-Answering with Adversarial Training

Code for our paper "Domain-agnostic Question-Answering with Adversarial Training" which is accepted by EMNLP-IJCNLP 2019 MRQA Workshop.

Model Architecture

Data Preparation

Download the original data

  • Download the data by running shell file.
  • Then run the code. Preprocessed train data will be created before training (It will takes quite a long time)
$ cd data
$ ./download_data.sh

(Optional) Download the pickled data (for fast data loading)

  • Download the pickled data from this link.

  • Unzip the zipfile on the root directory.

.
├── ...
├── pickled_data_bert-base-uncased_False
│   ├── HotpotQA.pkl
│   ├── NaturalQuestions.pkl
│   ├── NewsQA.pkl
│   ├── SQuAD.pkl
│   ├── SearchQA.pkl
│   └── TriviaQA.pkl
└── ...
  • Arguments should be same as below if you use pickled data. If you want to change one of these two arguments.
parser.add_argument("--bert_model", default="bert-base-uncased", type=str, help="Bert model")
parser.add_argument("--skip_no_ans", action='store_true', default=False, help="whether to exclude no answer example")

Requirements

Please install the following library requirements specified in the requirements.txt first.

torch==1.1.0
pytorch-pretrained-bert>=0.6.2
json-lines>=0.5.0

Model Training & Validation

$ python3 main.py \
         --epochs 2 \
         --batch_size 64 \
         --lr 3e-5 \
         --do_lower_case \
         --use_cuda \
         --do_valid \
         --adv \
         --dis_lambda 0.01
  • If you are using uncased bert model, give the option --do_lower_case.
  • If you want to do validation, give the option --do_valid.

Reference

@inproceedings{lee-etal-2019-domain,
    title={Domain-agnostic Question-Answering with Adversarial Training},
    author={Seanie Lee and Donggyu Kim and Jangwon Park},
    booktitle={Proceedings of the 2nd Workshop on Machine Reading for Question Answering},
    publisher={Association for Computational Linguistics},
    year={2019},
    url={https://www.aclweb.org/anthology/D19-5826},
}

Contributors