Prompt-GAN–Customisable Hate Speech and Extremist Datasets via Radicalised Neural Language Models
Jarod Govers, Philip Feldman, Aaron Dant, Panos Patros · 2023
Online hate speech and violent extremism knows no borders, no political boundaries, no remorse. Researchers face an uphill battle to collect hate speech data in volumes and topical diversity suitable for training state-of-the-art content-moderation systems. Neural language models ushered in a new era of synthetic data generation in use across various businesses, all despite calls for research to protect against unintended toxic output. We present a method for radicalising pre-trained neural language models to identify real hate speech and highlight the risks of AI which could undermine our trust in social media. We present Prompt-GAN, a prompt-tuning adversarial approach with three achievements. Namely, we demonstrate prompt-tuning’s ability to generate realistic types of hate and non-hate speech which mimics political extremist discourse. Prompt-GAN’s architecture offers a twofold reduction in memory and runtime requirements compared to fine-tuning. Prompt-GAN improves hate speech classification F1-scores by up to 10.1% and sets a new record in neural language simulation compared to the current state-of-the-art across three benchmark social media datasets.