Inclusive Counterfactual Generation: Leveraging LLMs in Identifying Online Hate

Counterfactually augmented data has recently been proposed as a successful solution for socially situated NLP tasks such as hate speech detection. The chief component within the existing counterfactual data augmentation pipeline, however, involves manually flipping labels and making minimal content edits to training data. In a hate speech context, these forms of editing have been shown to still retain offensive hate speech content. Inspired by the recent success of large language models (LLMs), especially the development of ChatGPT, which have demonstrated improved language comprehension abilities, we propose an inclusivity-oriented approach to automatically generate counterfactually augmented data using LLMs. We show that hate speech detection models trained with LLM-produced counterfactually augmented data can outperform both state-of-the-art and human-based methods.