Towards Robustifying NLI Models Against Lexical Dataset Biases
This is the official repo for the following paper
- Towards Robustifying NLI Models Against Lexical Dataset Biases, Xiang Zhou and Mohit Bansal, ACL 2020 (arxiv)
This code require Python 3.4 and TensorFlow 1.12.0
All the datasets (train/eval) can be downloaded at here. For detailed description of the datasets, please check the README in the downloaded file.
- Download the datasets and put it under the
- Download the GloVe embeddings and put it under the
Example scripts for BoW Sub-Model Orthogonality with HEX
- First train the baseline BiLSTM model by running
- Train the debiased model by running
The HEX implementation is adapted from https://github.com/HaohanWang/HEX.
The evaluation scripts is at
evaluation.py. When running evaluation, first change the
TESTING_DATASETS in the file. Then run
python evaluation.py scripts/TRAININGSCRIPT. This script will automatically generate and runs the testing scripts with respect to your training script.
More codes, model checkpoints and documentations will come soon.