AutoCAD

Code and datasets for our paper "AutoCAD: Automatically Generating Counterfactuals for Mitigating Shortcut Learning"

Updates

[Oct 13 2022] Code released

1. Environment Setup

• python >= 3.8
• pip install -r requirements.txt

2. Data Preprocessing

• Run unzip data.zip

3. Run the Code

Before running the code, please change the WORKING_DIR in the script according to your own path.

(1) Train a Classifier

bash scripts/train.sh

(2) Identify Rationales

• get gradient-based saliency scores: bash scripts/saliency.sh

• get rationale-corrupted data: python tools/highlight.py

(3) Train a Counterfactual Generator

• train a counterfactual generator on the rationale-corrupted data: bash scripts/train_generator.sh

(4) Generate Counterfactuals

• generate counterfactuals on the label-flipped rationale-corrupted data: bash scripts/generate.sh

(5) Consistency Filtering

• predict the label of the generated counterfactuals with the fine-tuned classifier: bash scripts/predict.sh