While social media has many virtues, it has undoubtedly also facilitated the flow of hate speech towards minority groups. Researchers from Carnegie Mellon have been working on an AI-based system to give voice to those who ordinarily lack one.
The system analyzes thousands of comments left online to better identify those who defend or sympathize with the disenfranchized communities. They hope that this will then enable human moderators to better showcase these more positive comments, and therefore change the narrative.
Such interventions are often crucial as minority groups lack the voice to stand up for themselves, whether due to a lack internet access or language proficiency. Indeed, so many minority groups have far more pressing issues than replying to comments online. Nonetheless, such comments do help to shape the debate, which in turn influences policy making on issues that do affect them.
Shaping the debate
The researchers began the project by searching through over 250,000 comments left on YouTube videos relating to the Rohingya refugee crisis. They have also used the same approach on around a million comments left on videos associated with the Pulwama terrorist attack in Kashmir.
The model is so effective that it can even work with short pieces of text riddled with spelling and grammar mistakes in a range of South Asian countries, even with their multiple languages and dialects. They did this by using embeddings that reveal distinct language clusters that help the system to identify language effectively.
The analysis revealed that around 10% of comments on YouTube videos are positive, but these can often get drowned out by the mass of negative commentary. The system was able to accurately detect such comments with around 88% accuracy, which the team believe will significantly reduce the work required of human moderators.
Social media is a genie that has been firmly out of the bottle, but it’s pleasing to see technology being used to tame the unhelpful aspects of the platforms and ensure that discrimination doesn’t fester.
