Counter hate speech detection in Youtubeconversations (IPMU2024)

Responding with counter-speech is one way to stop the spread of hate speech. In this work, models for identifying counterspeech are presented, based on annotated YouTube video comments from people speaking Portuguese for the first time. Using the order in which responses to comments are made, we create a corpus of comment pairings in which a target is classified as neutral, hate speech, or counter-speech in relation to a context that is related to a previous comment. The authors experiment with both multilingual and Portuguese pre-trained models, computing models by optimizing pre-trained models based on Transformers using such a corpus. Both the experimental setup and corpus design of the technique are based on previous work for English, and we get comparable results in Portuguese.

https://ipmu2024.inesc-id.pt/files/paper_1154.pdf