Social Media Implicit Hate Speech Dataset

It is derived from the TOXICLOAKCN dataset, which itself is constructed by applying perturbations—such as homophone substitution, emoji replacement, and Pinyin augmentation—to original sentences from datasets like TOXICN. These perturbations aim to simulate common strategies used on social media to evade content moderation.

In this work, we randomly sampled 1,000 instances from this dataset to assess the generalization capability of large language models in identifying implicit hate speech.