A context aware approach for generating natural language attack.

This repository contains source code for the research work described in our AAAI 2021 paper:

A context aware approach for generating natural language attack

Requirements

• Pytorch >= 0.4
• Tensorflow >= 1.0
• Numpy
• Python >= 3.6
• Tensorflow 2.1.0
• TensorflowHub

Dependencies

• Download counter-fitted-vectors from here, unzip it and place the txt file in the main directory.
• Download pretrained BERT for each dataset here, unzip it and place all the folders it in BERT/results/ directory.
• Download top 50 synonym file from here, unzip it and place the txt file in the main directory.

How to Run:

• Run the following command to get classification results.

bash bert.sh

• Run the following command to get inference results.

python run_attack_nli.py

The results will be available in results directory.