README.md

Problem Overview

Most of existing methods for hate speech detection adopt a supervised aproach that heavily depends on labeled datasets for training. The imbalance of datasets may be detrimental to performance of methods. We propose HateGAN, a deep generative reinforcement learning model to address the challenge of imbalance class by augementing datasets with hateful tweets.

This repository contains code for the implementation of HateGAN.

Dependencies:

• Python 3.5
• Pytorch 1.0.0

Data for Training HateGAN

Dataset	Label (Count)
WZ	hate (3,435) non-hate (9,767)
DT	hate (1,430) offensive (19,190) beither (4,163)
FOUNTA	hate (3,907) abusive (19,232) spam (13,840) normal (53,011)
HateLingo	hate (11,763)

Data for Hate Speech Detection

Dataset	Type	Label (Count)
WZ	Binary	hate (3,435) non-hate (9,767)
DT	Multi	hate (1,430) offensive (19,190) beither (4,163)

Getting Started

• Pretrain the toxic comment detection model, use code in Toxic Model.ipynb

• Train the generative model, use code in HateGAN:

  python main.py

• Evaluate the generated tweets, use code in Test-GAN-gen:

  python main.py

We have also reference the SeqGAN implementation: https://github.com/X-czh/SeqGAN-PyTorch.

Acknowledge

To cite:

@inproceedings{cao2020hategan,
  title={HateGAN: Adversarial Generative-Based Data Augmentation for Hate Speech Detection},
  author={Cao, Rui and Lee, Roy Ka-Wei},
  booktitle={Proceedings of the 28th International Conference on Computational Linguistics},
  pages={6327--6338},
  year={2020}
}