https://github.com/datngu/nli-artifacts
NLI artifacts mitigation
https://github.com/datngu/nli-artifacts
Last synced: 2 months ago
JSON representation
NLI artifacts mitigation
- Host: GitHub
- URL: https://github.com/datngu/nli-artifacts
- Owner: datngu
- Created: 2023-12-07T13:13:59.000Z (over 1 year ago)
- Default Branch: main
- Last Pushed: 2023-12-07T13:14:42.000Z (over 1 year ago)
- Last Synced: 2025-01-21T20:48:47.426Z (4 months ago)
- Language: Jupyter Notebook
- Size: 18.6 MB
- Stars: 1
- Watchers: 1
- Forks: 1
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# Dissecting vocabulary biases in Natural Language Inference datasets through Statistical testing approach and Automated data augmentation for artifact mitigation
## Getting Started
You'll need Python >= 3.6 to run the code in this repo.
First, clone the repository:
`git clone [email protected]:datngu/nli-artifacts.git`
and follow the README.md file in the __run__ directory to install the base version.
You then need to install: __nlpaug__ package from: https://github.com/makcedward/nlpaug
Please check its homepage to install it properly.
## Obtain the baseline dataset
Please follow steps by steps in the notebook of __data_writer.ipynb__ to write the original dataset and generate the hypothesis-only data.
## Statistical testing
Implemtation of the poposed statistical test can be done by running the __statistical_test.ipynb__.
Related figures and data are automatically generated.
### Note:
Please make sure that related packages are installed.
## Obtain augmented data
We prepare customized python and bash scripts in the subdirectory of aug to generate the augmented data.
Please remember to download the related models/databases if you like to use sysnomyn and embbeding augmentation.
Links and instructions to download these data are provided in the python script.
## Training and optimizing the models
In the __run__ subdirectory, we provided python scripts for training and optimize the ELECTRA small model.
We also provided slurm scripts that we used to run the experiments in the __run_experiments__ directory.