We study ethnic bias and how it varies across languages by analyzing and mitigating ethnic bias in monolingual BERT for English, German, Spanish, Korean, Turkish, and Chinese. We compare our proposed methods with monolingual BERT and show that these methods effectively alleviate the ethnic bias.

This dataset labels the aggression and hate targets of Korean news comments. This dataset discriminates offensive and not offensive for each comment as binary. Also, it finds hate span, the basis of classification, and classifies it into target-term, predicate-term, and no-span hate speech.