https://github.com/bouchardi/BERT_for_GAP-coreference

BERT finetuning for GAP unbiased pronoun resolution
https://github.com/bouchardi/BERT_for_GAP-coreference

Last synced: about 1 month ago
JSON representation

BERT finetuning for GAP unbiased pronoun resolution

• Host: GitHub
• URL: https://github.com/bouchardi/BERT_for_GAP-coreference
• Owner: bouchardi
• Created: 2019-04-01T19:09:16.000Z (over 5 years ago)
• Default Branch: master
• Last Pushed: 2019-04-22T07:13:30.000Z (over 5 years ago)
• Last Synced: 2024-08-11T16:09:17.125Z (4 months ago)
• Language: Python
• Homepage:
• Size: 122 KB
• Stars: 5
• Watchers: 0
• Forks: 1
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

• awesome-bert - isabellebouchard/BERT_for_GAP-coreference

README

        
# BERT_for_GAP-coreference

This project was realised in the context of the INF8225 AI course. In this project,

we aim to reduce gender bias in pronoun resolution by creating a coreference 

resolver that performs well on a gender-balanced pronoun dataset, the Gendered 

Ambiguous Pronouns (GAP) dataset. We leverage BERT's strong pre-training tasks on 

large unsupervised datasets and transfer these contextual representations to the fine-tuning stage. The fine-tuning stage was trained in a SWAG-like manner on the GAP supervised dataset.

We have submitted our best performing model to the [Gendered Pronoun Resolution](https://www.kaggle.com/c/gendered-pronoun-resolution/) Kaggle competition. 

## Setting up

```

git clone --recursive [email protected]:isabellebouchard/BERT_for_GAP-coreference.git

```

Make sure the submodules are properly initialized. 

## First steps

To run the code, first install [Docker](https://docs.docker.com/install/) to be able

to build and run a docker container with all the proper dependencies installed

```

docker build -t IMAGE_NAME .

nvidia-docker run --rm -it -v /path/to/your/code/:/project IMAGE_NAME

```

If you don't have access to GPU, change `nvidia-docker` for `docker`. It is 

highly recommended to run the training on (multiple) GPUs.

Once inside the container you should be able run the training script:

```

python run_GAP.py --data_dir gap-coreference \

                  --bert_model bert-base-cased \

                  --output_dir results \

```

This will run the training script and save checkpoints of the best model in the 

output directory.